Topological Vector Spaces

Maria Infusino
University of Konstanz

Winter Semester 2015/2016

Contents

1 Preliminaries 3
1.1 Topological spaces . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 The notion of topological space . . . . . . . . . . . . . . 3
1.1.2 Comparison of topologies . . . . . . . . . . . . . . . . . 6
1.1.3 Reminder of some simple topological concepts . . . . . . 8
1.1.4 Mappings between topological spaces . . . . . . . . . . . 11
1.1.5 Hausdorff spaces . . . . . . . . . . . . . . . . . . . . . . 13
1.2 Linear mappings between vector spaces . . . . . . . . . . . . . 14

2 Topological Vector Spaces 17

2.1 Definition and main properties of a topological vector space . . 17
2.2 Hausdorff topological vector spaces . . . . . . . . . . . . . . . . 24
2.3 Quotient topological vector spaces . . . . . . . . . . . . . . . . 25
2.4 Continuous linear mappings between t.v.s. . . . . . . . . . . . . 29
2.5 Completeness for t.v.s. . . . . . . . . . . . . . . . . . . . . . . . 31

1
3 Finite dimensional topological vector spaces 43
3.1 Finite dimensional Hausdorff t.v.s. . . . . . . . . . . . . . . . . 43
3.2 Connection between local compactness and finite dimensionality 46

4 Locally convex topological vector spaces 49

4.1 Definition by neighbourhoods . . . . . . . . . . . . . . . . . . . 49
4.2 Connection to seminorms . . . . . . . . . . . . . . . . . . . . . 54
4.3 Hausdorff locally convex t.v.s . . . . . . . . . . . . . . . . . . . 64
4.4 The finest locally convex topology . . . . . . . . . . . . . . . . 67
4.5 Direct limit topology on a countable dimensional t.v.s. . . . . . 69
4.6 Continuity of linear mappings on locally convex spaces . . . . . 71

5 The Hahn-Banach Theorem and its applications 73

5.1 The Hahn-Banach Theorem . . . . . . . . . . . . . . . . . . . . 73
5.2 Applications of Hahn-Banach theorem . . . . . . . . . . . . . . 77
5.2.1 Separation of convex subsets of a real t.v.s. . . . . . . . 78
5.2.2 Multivariate real moment problem . . . . . . . . . . . . 80
 Chapter 1

Preliminaries

1.1 Topological spaces

1.1.1 The notion of topological space
The topology on a set X is usually defined by specifying its open subsets of X.
However, in dealing with topological vector spaces, it is often more convenient
to define a topology by specifying what the neighbourhoods of each point are.

Definition 1.1.1. A topology τ on a set X is a family of subsets of X which

satisfies the following conditions:
(O1) the empty set ∅ and the whole X are both in τ
(O2) τ is closed under finite intersections
(O3) τ is closed under arbitrary unions
The pair (X, τ ) is called a topological space.

The sets O ∈ τ are called open sets of X and their complements C = X \O

are called closed sets of X. A subset of X may be neither closed nor open,
either closed or open, or both. A set that is both closed and open is called a
clopen set.

Definition 1.1.2. Let (X, τ ) be a topological space.

• A subfamily B of τ is called a basis if every open set can be written as
a union of sets in B.
• A subfamily X of τ is called a subbasis if the finite intersections of its
sets form a basis, i.e. every open set can be written as a union of finite
intersections of sets in X .

Therefore, a topology τ on X is completely determined by a basis or a

subbasis.
3
1. Preliminaries

Example 1.1.3. Let S be the collection of all semi-infinite intervals of the

real line of the forms (−∞, a) and (a, +∞), where a ∈ R. S is not a base for
any topology on R. To show this, suppose it were. Then, for example, (−∞, 1)
and (0, ∞) would be in the topology generated by S, being unions of a single
base element, and so their intersection (0, 1) would be by the axiom (O2) of
topology. But (0, 1) clearly cannot be written as a union of elements in S.
Proposition 1.1.4. Let X be a set and let B be a collection of subsets of X.
B is a basis for a topology τ on X iff the following hold:
1. B covers X, i.e. ∀ x ∈ X, ∃ B ∈ B s.t. x ∈ B.
2. If x ∈ B1 ∩B2 for some B1 , B2 ∈ B, then ∃ B3 ∈ B s.t. x ∈ B3 ⊆ B1 ∩B2 .
Proof. (Sheet 1, Exercise 1)
Definition 1.1.5. Let (X, τ ) be a topological space and x ∈ X. A subset U
of X is called a neighbourhood of x if it contains an open set containing the
point x, i.e. ∃ O ∈ τ s.t. x ∈ O ⊆ U . The family of all neighbourhoods of a
point x ∈ X is denoted by F(x).
In order to define a topology on a set by the family of neighbourhoods of
each of its points, it is convenient to introduce the notion of filter. Note that
the notion of filter is given on a set which does not need to carry any other
structure. Thus this notion is perfectly independent of the topology.
Definition 1.1.6. A filter on a set X is a family F of subsets of X which
fulfills the following conditions:
(F1) the empty set ∅ does not belong to F
(F2) F is closed under finite intersections
(F3) any subset of X containing a set in F belongs to F
Definition 1.1.7. A family B of subsets of X is called a basis of a filter F if
1. B ⊆ F
2. ∀A ∈ F, ∃ B ∈ B s.t. B ⊆ A
Examples 1.1.8.
a) The family G of all subsets of a set X containing a fixed non-empty sub-
set A is a filter and B = {A} is its base. G is called the principle filter
generated by A.
b) Given a topological space X and x ∈ X, the family F(x) is a filter.
c) Let S := {xn }n∈N be a sequence of points in a set X. Then the family
F := {A ⊂ X : |S \ A| < ∞} is a filter and it is known as the filter
associated to S. For each m ∈ N, set Sm := {xn ∈ S : n ≥ m}. Then
B := {Sm : m ∈ N} is a basis for F.
Proof. (Sheet 1, Exercise 2).

4
 1.1. Topological spaces

Theorem 1.1.9. Given a topological space X and a point x ∈ X, the filter of

neighbourhoods F(x) satisfies the following properties.
(N1) For any A ∈ F(x), x ∈ A.
(N2) For any A ∈ F(x), ∃ B ∈ F(x): ∀ y ∈ B, A ∈ F(y).
Viceversa, if for each point x in a set X we are given a filter Fx fulfilling the
properties (N1) and (N2) then there exists a unique topology τ s.t. for each
x ∈ X, Fx is the family of neighbourhoods of x, i.e. Fx ≡ F(x), ∀x ∈ X.
This means that a topology on a set is uniquely determined by the family
of neighbourhoods of each of its points.
Proof.
⇒ Let (X, τ ) be a topological space, x ∈ X and F(x) the filter of neighbour-
hoods of x. Then (N1) trivially holds by definition of neighbourhood of x. To
show (N2), let us take A ∈ F(x). Since A is a neighbourhood of x, there exists
B ∈ τ s.t. x ∈ B ⊆ A. Then clearly B ∈ F(x). Moreover, since for any y ∈ B
we have that y ∈ B ⊆ A and B is open, we can conclude that A ∈ F(y).
⇐ Assume that for any x ∈ X we have a filter Fx fulfilling (N1) and (N2).
Let us define τ := {O ⊆ X : if x ∈ O then O ∈ Fx }. Since each Fx is a filter,
τ is a topology. Indeed:
• ∅ ∈ τ by definition of τ . Also X ∈ τ , because for any x ∈ X and any
A ∈ Fx we clearly have X ⊇ A and so by (F3) X ∈ Fx .
• By (F2) we have that τ is closed under finite intersection.
• Let U be an arbitrary union of sets Ui ∈ τ and let x ∈ U . Then there
exists at least one i s.t. x ∈ Ui and so Ui ∈ Fx because Ui ∈ τ . But
U ⊇ Ui , then by (F3) we get that U ∈ Fx and so U ∈ τ .
It remains to show that τ on X is actually s.t. Fx ≡ F(x), ∀x ∈ X.
• Any U ∈ F(x) is a neighbourhood of x and so there exists O ∈ τ s.t.
x ∈ O ⊆ U . Then, by definition of τ , we have O ∈ Fx and so (F3)
implies that U ∈ Fx . Hence, F(x) ⊆ Fx .
• Let U ∈ Fx and set W := {y ∈ U : U ∈ Fy } ⊆ U . Since x ∈ U by
(N1), we also have x ∈ W . Moreover, if y ∈ W then by (N2) there exists
V ∈ Fy s.t. ∀z ∈ V we have U ∈ Fz . This means that z ∈ W and so
V ⊆ W . Then W ∈ Fy by (F3). Hence, we have showed that if y ∈ W
then W ∈ Fy , i.e. W ∈ τ . Summing up, we have just constructed an
open set W s.t. x ∈ W ⊆ U , i.e. U ∈ F(x), and so Fx ⊆ F(x).

Definition 1.1.10. Given a topological space X, a basis B(x) of the filter of

neighbourhoods F(x) of a point x ∈ X is called a base of neighbourhoods of x,
i.e. B(x) is a subcollection of F(x) s.t. every neighbourhood in F(x) contains

5
1. Preliminaries

one in B(x). The elements of B(x) are called basic neighbourhoods of x. If

a base of neighbourhoods is given for each point x ∈ X, we speak of base of
neighbourhoods of X.
Example 1.1.11. The open sets of a topological space other than the empty
set always form a base of neighbourhoods.
Theorem 1.1.12. Given a topological space X and a point x ∈ X, a base of
open neighbourhoods B(x) satisfies the following properties.
(B1) For any U ∈ B(x), x ∈ U .
(B2) For any U1 , U2 ∈ B(x), ∃ U3 ∈ B(x) s.t. U3 ⊆ U1 ∩ U2 .
(B3) If y ∈ U ∈ B(x), then ∃ W ∈ B(y) s.t. W ⊆ U .
Viceversa, if for each point x in a set X we are given a collection of subsets
Bx fulfilling the properties (B1), (B2) and (B3) then there exists a unique
topology τ s.t. for each x ∈ X, Bx is a base of neighbourhoods of x, i.e.
Bx ≡ B(x), ∀x ∈ X.
Proof. The proof easily follows by using Theorem 1.1.9.

The previous theorem gives a further way of introducing a topology on a

set. Indeed, starting from a base of neighbourhoods of X, we can define a
topology on X by setting that a set is open iff whenever it contains a point
it also contains a basic neighbourhood of the point. Thus a topology on a set
X is uniquely determined by a base of neighbourhoods of each of its points.

1.1.2 Comparison of topologies

Any set X may carry several different topologies. When we deal with topo-
logical vector spaces, we will very often encounter this situation of a set, in
fact a vector space, carrying several topologies (all compatible with the linear
structure, in a sense that is going to be specified soon). In this case, it is
convenient being able to compare topologies.
Definition 1.1.13. Let τ , τ 0 be two topologies on the same set X. We say
that τ is coarser (or weaker) than τ 0 , in symbols τ ⊆ τ 0 , if every subset of X
which is open for τ is also open for τ 0 , or equivalently, if every neighborhood
of a point in X w.r.t. τ is also a neighborhood of that same point in the
topology τ 0 . In this case τ 0 is said to be finer (or stronger) than τ 0 .
Denote by F(x) and F 0 (x) the filter of neighbourhoods of a point x ∈ X
w.r.t. τ and w.r.t. τ 0 , respectively. Then: τ is coarser than τ 0 iff for any
point x ∈ X we have F(x) ⊆ F 0 (x) (this means that every subset of X which
belongs to F(x) also belongs to F 0 (x)).

6
 1.1. Topological spaces

Two topologies τ and τ 0 on the same set X coincide when they give the
same open sets or the same closed sets or the same neighbourhoods of each
point; equivalently, when τ is both coarser and finer than τ 0 . Two basis of
neighbourhoods of a set are equivalent when they define the same topology.
Remark 1.1.14. Given two topologies on the same set, it may very well
happen that none is finer than the other. If it is possible to establish which
one is finer, then we say that the two topologies are comparable.
Example 1.1.15.
The cofinite topology τc on R, i.e. τc := {U ⊆ R : U = ∅ or R \ U is finite},
and the topology τi having {(−∞, a) : a ∈ R} as a basis are incomparable. In
fact, it is easy to see that τi = {(−∞, a) : a ∈ R} ∪ {∅, R} as these are the
unions of sets in the given basis. In particular, we have that R − {0} is in τc
but not τi . Moreover, we have that (−∞, 0) is in τi but not τc . Hence, τc and
τi are incomparable.
It is always possible to construct at least two topologies on every set X
by choosing the collection of open sets to be as large as possible or as small
as possible:
• the trivial topology: every point of X has only one neighbourhood which
is X itself. Equivalently, the only open subsets are ∅ and X. The only
possible basis for the trivial topology is {X}.
• the discrete topology: given any point x ∈ X, every subset of X contain-
ing x is a neighbourhood of x. Equivalently, every subset of X is open
(actually clopen). In particular, the singleton {x} is a neighbourhood
of x and actually is a basis of neighbourhoods of x. The collection of all
singletons is a basis for the discrete topology.
Note that the discrete topology on a set X is finer than any other topology on
X, while the trivial topology is coarser than all the others. Topologies on a set
form thus a partially ordered set, having a maximal and a minimal element,
respectively the discrete and the trivial topology.

A useful criterion to compare topologies on the same set is the following:

Theorem 1.1.16 (Hausdorff’s criterion).
For each x ∈ X, let B(x) a base of neighbourhoods of x for a topology τ on X
and B 0 (x) a base of neighbourhoods of x for a topology τ 0 on X.
τ ⊆ τ 0 iff ∀ x ∈ X, ∀ U ∈ B(x) ∃ V ∈ B 0 (x) s.t. x ∈ V ⊆ U .

7
1. Preliminaries

The Hausdorff criterion could be paraphrased by saying that smaller neigh-

borhoods make larger topologies. This is a very intuitive theorem, because
the smaller the neighbourhoods are the easier it is for a set to contain neigh-
bourhoods of all its points and so the more open sets there will be.
Proof.
⇒ Suppose τ ⊆ τ 0 . Fixed any point x ∈ X, let U ∈ B(x). Then, since U
is a neighbourhood of x in (X, τ ), there exists O ∈ τ s.t. x ∈ O ⊆ U . But
O ∈ τ implies by our assumption that O ∈ τ 0 , so U is also a neighbourhood
of x in (X, τ 0 ). Hence, by Definition 1.1.10 for B 0 (x), there exists V ∈ B 0 (x)
s.t. V ⊆ U .
⇐ Conversely, let W ∈ τ . Then for each x ∈ W , since B(x) is a base of
neighbourhoods w.r.t. τ , there exists U ∈ B(x) such that x ∈ U ⊆ W . Hence,
by assumption, there exists V ∈ B 0 (x) s.t. x ∈ V ⊆ U ⊆ W . Then W ∈ τ 0 .

1.1.3 Reminder of some simple topological concepts

Definition 1.1.17. Given a topological space (X, τ ) and a subset S of X, the
subset or induced topology on S is defined by τS := {S ∩ U | U ∈ τ }. That is,
a subset of S is open in the subset topology if and only if it is the intersection
of S with an open set in (X, τ ).
Alternatively, we can define the subspace topology for a subset S of X as the
coarsest topology for which the inclusion map ι : S ,→ X is continuous.
Note that (S, τs ) is a topological space in its own.
Definition 1.1.18. Given a collection of topological space (Xi , τi ), where i ∈ I
(I is an index set
Q possibly uncountable), the product topology on the Cartesian
product X := i∈I Xi is defined in the following Q way: a set U is open in X
iff it is an arbitrary union of sets of the form i∈I Ui , where each Ui ∈ τi and
Ui 6= Xi for only finitely many i.
Alternatively, we can define the product topology to be the coarsest topology
for which all the canonical projections pi : X → Xi are continuous.
Given a topological space X, we define:
Definition 1.1.19.
• The closure of a subset A ⊆ X is the smallest closed set containing A.
It will be denoted by Ā. Equivalently, Ā is the intersection of all closed
subsets of X containing A.
• The interior of a subset A ⊆ X is the largest open set contained in it.
It will be denoted by Å. Equivalently, Å is the union of all open subsets
of X contained in A.

8
 1.1. Topological spaces

Proposition 1.1.20. Given a top. space X and A ⊆ X, the following hold.

• A point x is a closure point of A, i.e. x ∈ Ā, if and only if each
neighborhood of x has a nonempty intersection with A.
• A point x is an interior point of A, i.e. x ∈ Å, if and only if there exists
a neighborhood of x which entirely lies in A.
• A is closed in X iff A = Ā.
• A is open in X iff A = Å.
Proof. (Sheet 2, Exercise 1)
Example 1.1.21. Let τ be the standard

euclidean topology on R. Consider
X := (R, τ ) and Y := (0, 1], τY , where τY is the topology induced by τ on
(0, 1]. The closure of (0, 21 ) in X is [0, 12 ], but its closure in Y is (0, 12 ].
Definition 1.1.22. Let A and B be two subsets of the same topological space X.
A is dense in B if B ⊆ Ā. In particular, A is said to be dense in X (or ev-
erywhere dense) if Ā = X.
Examples 1.1.23.
• Standard examples of sets everywhere dense in the real line R (with the
euclidean topology) are the set of rational numbers Q and the one of
irrational numbers R − Q.
• A set X is equipped with the discrete topology if and only if the whole
space X is the only dense set in itself.
If X has the discrete topology then every subset is equal to its own
closure (because every subset is closed), so the closure of a proper subset
is always proper. Conversely, if X is the only dense subset of itself, then
for every proper subset A its closure Ā is also a proper subset of X. Let
y ∈ X be arbitrary. Then to X \ {y} is a proper subset of X and so
it has to be equal to its own closure. Hence, {y} is open. Since y is
arbitrary, this means that X has the discrete topology.
• Every non-empty subset of a set X equipped with the trivial topology
is dense, and every topology for which every non-empty subset is dense
must be trivial.
If X has the trivial topology and A is any non-empty subset of X, then
the only closed subset of X containing A is X. Hence, Ā = X, i.e. A
is dense in X. Conversely, if X is endowed with a topology τ for which
every non-empty subset is dense, then the only non-empty subset of X
which is closed is X itself. Hence, ∅ and X are the only closed subsets
of τ . This means that X has the trivial topology.
Proposition 1.1.24. Let X be a topological space and A ⊂ X. A is dense in
X if and only if every nonempty open set in X contains a point of A.
9
1. Preliminaries

Proof. If A is dense in X, then by definition Ā = X. Let O be any nonempty

open subset in X. Then for any x ∈ O we have that x ∈ Ā and O ∈ F(x).
Therefore, by Proposition 1.1.20, we have that O ∩ A 6= ∅. Conversely, let
x ∈ X. By definition of neighbourhood, for any U ∈ F(x) there exists an
open subset O of X s.t. x ∈ O ⊆ U . Then U ∩ A 6= ∅ since O contains a point
of A by our assumption. Hence, by Proposition 1.1.20, we get x ∈ Ā and so
that A is dense in X.

Definition 1.1.25. A topological space X is said to be separable if there exists

a countable dense subset of X.
Example 1.1.26.
• R with the euclidean topology is separable.
• The space C([0, 1]) of all continuous functions from [0, 1] to R endowed
with the uniform topology is separable, since by the Weirstrass approxi-
mation theorem Q[x] = C([0, 1]).
Let us briefly consider now the notion of convergence.
First of all let us concern with filters. When do we say that a filter F on
a topological space X converges to a point x ∈ X? Intuitively, if F has to
converge to x, then the elements of F, which are subsets of X, have to get
somehow “smaller and smaller” about x, and the points of these subsets need
to get “nearer and nearer” to x. This can be made more precise by using
neighborhoods of x: we want to formally express the fact that, however small
a neighborhood of x is, it should contain some subset of X belonging to the
filter F and, consequently, all the elements of F which are contained in that
particular one. But in view of Axiom (F3), this means that the neighborhood
of x under consideration must itself belong to the filter F, since it must contain
some element of F.
Definition 1.1.27. Given a filter F in a topological space X, we say that it
converges to a point x ∈ X if every neighborhood of x belongs to F, in other
words if F is finer than the filter of neighborhoods of x.
We recall now the definition of convergence of a sequence to a point and
we see how it easily connects to the previous definition.
Definition 1.1.28. Given a sequence of points {xn }n∈N in a topological space
X, we say that it converges to a point x ∈ X if for any U ∈ F(x) there exists
N ∈ N such that xn ∈ U for all n ≥ N .
If we now consider the filter FS associated to the sequence S := {xn }n∈N ,
i.e. FS := {A ⊂ X : |S \ A| < ∞}, then it is easy to see that:

10
 1.1. Topological spaces

Proposition 1.1.29. Given a sequence of points S := {xn }n∈N in a topolog-

ical space X, S converges to a point x ∈ X if and only if the associated filter
FS converges to x.
Proof. Set for each m ∈ N, set Sm := {xn ∈ S : n ≥ m}. By Definition 1.1.28,
S converges to x iff ∀ U ∈ F(x), ∃N ∈ N : SN ⊆ U . As B := {Sm : m ∈ N}
is a basis for FS (see Problem Sheet 1, Exercise 2 c)) , we have that ∀ U ∈
F(x), ∃N ∈ N : SN ⊆ U is equivalent to say that F(x) ⊆ FS .

1.1.4 Mappings between topological spaces

Let (X, τX ) and (Y, τY ) be two topological spaces.
Definition 1.1.30. A map f : X → Y is continuous if the preimage of any
open set in Y is open in X, i.e. ∀ U ∈ τY , f −1 (U ) := {x ∈ X : f (x) ∈ Y } ∈ τX .
Equivalently, given any point x ∈ X and any N ∈ F(f (x)) in Y , the preimage
f −1 (N ) ∈ F(x) in X.
Examples 1.1.31.
• Any constant map f : X → Y is continuous.
Suppose that f (x) := y for all x ∈ X and some y ∈ Y . Let U ∈ τY . If
y ∈ U then f −1 (U ) = X and if y ∈
/ U then f −1 (U ) = ∅. Hence, in either
−1
case, f (U ) is open in τX .
• If g : X → Y is continuous, then the restriction of g to any subset S
of X is also continuous w.r.t. the subset topology induced on S by the
topology on X.
• Let X be a set endowed with the discrete topology, Y be a set endowed
with the trivial topology and Z be any topological space. Any maps f :
X → Z and g : Z → Y are continuous.
Definition 1.1.32. A mapping f : X → Y is open if the image of any open
set in X is open in Y , i.e. ∀ V ∈ τX , f (V ) := {f (x) : x ∈ X} ∈ τY . In the
same way, a closed mapping f : X → Y sends closed sets to closed sets.
Note that a map may be open, closed, both, or neither of them. Moreover,
open and closed maps are not necessarily continuous.
Example 1.1.33. If Y has the discrete topology (i.e. all subsets are open
and closed) then every function f : X → Y is both open and closed (but
not necessarily continuous). For example, if we take the standard euclidean
topology on R and the discrete topology on Z then the floor function R → Z
is open and closed, but not continuous. (Indeed, the preimage of the open set
{0} is [0, 1) ⊂ R, which is not open in the standard euclidean topology).

11
1. Preliminaries

If a continuous map f is one-to-one, f −1 does not need to be continuous.

Example 1.1.34.
Let us consider [0, 1) ⊂ R and S 1 ⊂ R2 endowed with the subspace topologies
given by the euclidean topology on R and on R2 , respectively. The map

f : [0, 1) → S1
t 7→ (cos 2πt, sin 2πt).

is bijective and continuous but f −1 is not continuous, since there are open
subsets of [0, 1) whose image under f is not open in S 1 . (For example, [0, 21 )
is open in [0, 1) but f ([0, 21 )) is not open in S 1 .)

Definition 1.1.35. A one-to-one map f from X onto Y is a homeomorphism

if and only if f and f −1 are both continuous. Equivalently, iff f and f −1 are
both open (closed). If such a mapping exists, X and Y are said to be two
homeomorphic topological spaces.

In other words an homeomorphism is a one-to-one mapping which sends

every open (resp. closed) set of X in an open (resp. closed) set of Y and
viceversa, i.e. an homeomorphism is both an open and closed map. Note that
the homeomorphism gives an equivalence relation on the class of all topological
spaces.

Examples 1.1.36. In these examples we consider any subset of Rn endowed

with the subset topology induced by the Euclidean topology on Rn .
1. Any open interval of R is homeomorphic to any other open interval of
R and also to R itself.
2. A circle and a square in R2 are homeomorphic.
3. The circle S 1 with a point removed is homeomorphic to R.

Let us consider now the case when a set X carries two different topologies
τ1 and τ2 . Then the following two properties are equivalent:
• the identity ι of X is continuous as a mapping from (X, τ1 ) and (X, τ2 )
• the topology τ1 is finer than the topology τ2 .
Therefore, ι is a homeomorphism if and only if the two topologies coincide.

Proof. Suppose that ι is continuous. Let U ∈ τ2 . Then ι−1 (U ) = U ∈ τ1 ,

hence U ∈ τ1 . Therefore, τ2 ⊆ τ1 . Conversely, assume that τ2 ⊆ τ1 and
take any U ∈ τ2 . Then U ∈ τ1 and by definition of identity we know that
ι−1 (U ) = U . Hence, ι−1 (U ) ∈ τ1 and therefore, ι is continuous.

12
 1.1. Topological spaces

Proposition 1.1.37. Continuous maps preserve the convergence of sequences.

That is, if f : X → Y is a continuous map between two topological spaces
(X, τX ) and (Y, τY ) and if {xn }n∈N is a sequence of points in X convergent to
a point x ∈ X then {f (xn )}n∈N converges to f (x) ∈ Y .

Proof. (Sheet 2, Exercise 4 b))

1.1.5 Hausdorff spaces

Definition 1.1.38. A topological space X is said to be Hausdorff (or sepa-
rated) if any two distinct points of X have neighbourhoods without common
points; or equivalently if:
(T2) two distinct points always lie in disjoint open sets.

In literature, the Hausdorff space are often called T2-spaces and the axiom
(T2) is said to be the separation axiom.

Proposition 1.1.39. In a Hausdorff space the intersection of all closed neigh-

bourhoods of a point contains the point alone. Hence, the singletons are closed.
Proof. Let us fix a point x ∈ X, where X is a Hausdorff space. Take y 6= x.
By definition, there exist a neighbourhood U (x) of x and a neighbourhood
V (y) of y s.t. U (x) ∩ V (y) = ∅. Therefore, y ∈
/ U (x).

Examples 1.1.40.
1. Any metric space is Hausdorff.
Indeed, for any x, y ∈ (X, d) with x 6= y just choose 0 < ε < 12 d(x, y)
and you get Bε (x) ∩ Bε (y) = ∅.
2. Any set endowed with the discrete topology is a Hausdorff space.
Indeed, any singleton is open in the discrete topology so for any two
distinct point x, y we have that {x} and {y} are disjoint and open.
3. The only Hausdorff topology on a finite set is the discrete topology.
In fact, since X is finite, any subset S of X is finite and so S is a
finite union of singletons. But since X is also Hausdorff, the previous
proposition implies that any singleton is closed. Hence, any subset S of
X is closed and so the topology on X has to be the discrete one.
4. An infinite set with the cofinite topology is not Hausdorff.
In fact, any two non-empty open subsets O1 , O2 in the cofinite topology
on X are complements of finite subsets. Therefore, their intersection
O1 ∩ O2 is still a complement of a finite subset, but X is infinite and so
O1 ∩ O2 6= ∅. Hence, X is not Hausdorff.

13
1. Preliminaries

1.2 Linear mappings between vector spaces

The basic notions from linear algebra are assumed to be well-known and so
they are not recalled here. However, we briefly give again the definition of
vector space and fix some general terminology for linear mappings between
vector spaces. In this section we are going to consider vector spaces over the
field K of real or complex numbers which is given the usual euclidean topology
defined by means of the modulus.
Definition 1.2.1. A set X with the two mappings:
X ×X → X
(x, y) 7→ x + y vector addition
K×X → X
(λ, x) 7→ λx scalar multiplication
is a vector space (or linear space) over K if the following axioms are satisfied:
(L1) 1. (x + y) + z = x + (y + z), ∀ x, y, z ∈ X (associativity of +)
2. x + y = y + x, ∀ x, y ∈ X (commutativity of +)
3. ∃ o ∈ X: x + o = x, ∀ x, ∈ X (neutral element for +)
4. ∀ x ∈ X, ∃! − x ∈ X s.t. x + (−x) = o (inverse element for +)
(L2) 1. λ(µx) = (λµ)x, ∀ x ∈ X, ∀ λ, µ ∈ K
(compatibility of scalar multiplication with field multiplication)
2. 1x = x ∀ x ∈ X (neutral element for scalar multiplication)
3. (λ + µ)x = λx + µx, ∀ x ∈ X, ∀ λ, µ ∈ K
(distributivity of scalar multiplication with respect to field addition)
4. λ(x + y) = λx + λy, ∀ x, y ∈ X, ∀ λ ∈ K
(distributivity of scalar multiplication wrt vector addition)
Definition 1.2.2.
Let X, Y be two vector space over K. A mapping f : X → Y is called lin-
ear mapping or homomorphism if f preserves the vector space structure, i.e.
f (λx + µy) = λf (x) + µf (y) ∀ x, y ∈ X, ∀ λ, µ ∈ K.
Definition 1.2.3.
• A linear mapping from X to itself is called endomorphism.
• A one-to-one linear mapping is called monomorphism. If S is a subspace
of X, the identity map is a monomorphism and it is called embedding.
• An onto (surjective) linear mapping is called epimorphism.
• A bijective (one-to-one and onto) linear mapping between two vector
spaces X and Y over K is called (algebraic) isomorphism. If such a
map exists, we say that X and Y are (algebraically) isomorphic X ∼
=Y.
• An isomorphism from X into itself is called automorphism.

14
 1.2. Linear mappings between vector spaces

It is easy to prove that: A linear mapping is one-to-one (injective) if and

only if f (x) = 0 implies x = 0.
Definition 1.2.4. A linear mapping from X → K is called linear functional
or linear form on X. The set of all linear functionals on X is called algebraic
dual and it is denoted by X ∗ .
Note that the dual space of a finite dimensional vector space X is isomor-
phic to X.

15
 Chapter 2

Topological Vector Spaces

2.1 Definition and properties of a topological vector space

In this section we are going to consider vector spaces over the field K of real
or complex numbers which is given the usual euclidean topology defined by
means of the modulus.
Definition 2.1.1. A vector space X over K is called a topological vector space
(t.v.s.) if X is provided with a topology τ which is compatible with the vector
space structure of X, i.e. τ makes the vector space operations both continuous.
More precisely, the condition in the definition of t.v.s. requires that:

X ×X → X
(x, y) 7→ x + y vector addition

K×X → X
(λ, x) 7→ λx scalar multiplication

are both continuous when we endow X with the topology τ , K with the eu-
clidean topology, X ×X and K×X with the correspondent product topologies.
Remark 2.1.2. If (X, τ ) is a t.v.s then it is clear from Definition 2.1.1 that
PN (n) (n) PN
k=1 λk xk → k=1 λk xk as n → ∞ w.r.t. τ if for each k = 1, . . . , N
(n)
as n → ∞ we have that λk → λk w.r.t. the euclidean topology on K and
(n)
xk → xk w.r.t. τ .
Let us discuss now some examples and counterexamples of t.v.s.
Examples 2.1.3.
a) Every vector space X over K endowed with the trivial topology is a t.v.s..

17
2. Topological Vector Spaces

b) Every normed vector space endowed with the topology given by the metric
induced by the norm is a t.v.s. (Sheet 2, Exercise 1 a)).
c) There are also examples of spaces whose topology cannot be induced by a
norm or a metric but that are t.v.s., e.g. the space of infinitely differentiable
functions, the spaces of test functions and the spaces of distributions (we
will see later in details their topologies).
In general, a metric vector space is not a t.v.s.. Indeed, there exist metrics
for which both the vector space operations of sum and product by scalars are
discontinuous (see Sheet 3, Exercise 1 c) for an example).
Proposition 2.1.4. Every vector space X over K endowed with the discrete
topology is not a t.v.s. unless X = {o}.
Proof. Assume by a contradiction that it is a t.v.s. and take o 6= x ∈ X. The
sequence αn = n1 in K converges to 0 in the euclidean topology. Therefore,
since the scalar multiplication is continuous, αn x → o, i.e. for any neighbour-
hood U of o in X there exists m ∈ N s.t. αn x ∈ U for all n ≥ m. In particular,
we can take U = {o} since it is itself open in the discrete topology. Hence,
αm x = o, which implies that x = o and so a contradiction.

Definition 2.1.5. Two t.v.s. X and Y over K are (topologically) isomorphic

if there exists a vector space isomorphism X → Y which is at the same time
a homeomorphism (i.e. bijective, linear, continuous and inverse continuous).
In analogy to Definition 1.2.3, let us collect here the corresponding termi-
nology for mappings between two t.v.s..
Definition 2.1.6. Let X and Y be two t.v.s. on K.
• A topological homomorphism from X to Y is a linear mapping which is
also continuous and open.
• A topological monomorphism from X to Y is an injective topological
homomorphism.
• A topological isomorphism from X to Y is a bijective topological homo-
morphism.
• A topological automorphism of X is a topological isomorphism from X
into itself.
Proposition 2.1.7. Given a t.v.s. X, we have that:
1. For any x0 ∈ X, the mapping x 7→ x + x0 ( translation by x0 ) is a
homeomorphism of X onto itself.
2. For any 0 6= λ ∈ K, the mapping x 7→ λx ( dilation by λ) is a topological
automorphism of X.

18
 2.1. Definition and main properties of a topological vector space

Proof. Both mappings are continuous by the very definition of t.v.s.. More-
over, they are bijections by the vector space axioms and their inverses x 7→
x − x0 and x 7→ λ1 x are also continuous. Note that the second map is also
linear so it is a topological automorphism.

Proposition 2.1.7–1 shows that the topology of a t.v.s. is always a transla-

tion invariant topology, i.e. all translations are homeomorphisms. Note that
the translation invariance of a topology τ on a vector space X is not sufficient
to conclude (X, τ ) is a t.v.s..
Example 2.1.8. If a metric d on a vector space X is translation invariant,
i.e. d(x + z, y + z) = d(x, y) for all x, y ∈ X (e.g. the metric induced by a
norm), then the topology induced by the metric is translation invariant and
the addition is always continuous. However, the multiplication by scalars does
not need to be necessarily continuous (take d to be the discrete metric, then the
topology generated by the metric is the discrete topology which is not compatible
with the scalar multiplication see Proposition 2.1.4).
The translation invariance of the topology of a t.v.s. means, roughly speak-
ing, that a t.v.s. X topologically looks about any point as it does about any
other point. More precisely:
Corollary 2.1.9. The filter F(x) of neighbourhoods of x ∈ X coincides with
the family of the sets O + x for all O ∈ F(o), where F(o) is the filter of
neighbourhoods of the origin o (i.e. the neutral element of the vector addition).
Proof. (Sheet 3, Exercise 2 a))
Thus the topology of a t.v.s. is completely determined by the filter of
neighbourhoods of any of its points, in particular by the filter of neighbour-
hoods of the origin o or, more frequently, by a base of neighbourhoods of the
origin o. Therefore, we need some criteria on a filter of a vector space X
which ensures that it is the filter of neighbourhoods of the origin w.r.t. some
topology compatible with the vector structure of X.
Theorem 2.1.10. A filter F of a vector space X over K is the filter of
neighbourhoods of the origin w.r.t. some topology compatible with the vector
structure of X if and only if
1. The origin belongs to every set U ∈ F
2. ∀U ∈ F, ∃V ∈ F s.t. V + V ⊂ U
3. ∀U ∈ F, ∀λ ∈ K with λ 6= 0 we have λU ∈ F
4. ∀U ∈ F, U is absorbing.
5. ∀U ∈ F, ∃V ∈ F s.t. V ⊂ U is balanced.

19
2. Topological Vector Spaces

Before proving the theorem, let us fix some definitions and notations:
Definition 2.1.11. Let U be a subset of a vector space X.
1. U is absorbing (or radial) if ∀x ∈ X ∃ρ > 0 s.t. ∀λ ∈ K with |λ| ≤ ρ we
have λx ∈ U . Roughly speaking, we may say that a subset is absorbing
if it can be made by dilation to swallow every point of the whole space.
2. U is balanced (or circled) if ∀x ∈ U , ∀λ ∈ K with |λ| ≤ 1 we have
λx ∈ U . Note that the line segment joining any point x of a balanced
set U to −x lies in U .
Clearly, o must belong to every absorbing or balanced set. The underlying
field can make a substantial difference. For example, if we consider the closed
interval [−1, 1] ⊂ R then this is a balanced subset of C as real vector space,
but if we take C as complex vector space then it is not balanced. Indeed, if
we take i ∈ C we get that i1 = i ∈ / [−1, 1].
Examples 2.1.12.
a) In a normed space the unit balls centered at the origin are absorbing and
balanced.
b) The unit ball B centered at ( 12 , 0) ∈ R2 is absorbing but not balanced in
the real vector space R2 endowed with the euclidean norm. Indeed, B is
a neighbourhood of the origin and so by Theorem 2.1.10-4 is absorbing.
However, B is not balanced because for example if we take x = (1, 0) ∈ B
and λ = −1 then λx ∈ / B.
c) In the real vector space R2 endowed with the euclidean topology, the subset
in Figure 2.1 is absorbing and the one in Figure 2.2 is balanced.

Figure 2.1: Absorbing Figure 2.2: Balanced

d) The polynomials R[x] are a balanced but not absorbing subset of the real
space C([0, 1], R) of continuous real valued functions on [0, 1]. Indeed, any
multiple of a polynomial is still a polynomial but not every continuous
function can be written as multiple of a polynomial.
e) The subset A := {(z1 , z2 ) ∈ C2 : |z1 | ≤ |z2 |} of the complex space C2 with
the euclidean topology is balanced but Å is not balanced.

20
 2.1. Definition and main properties of a topological vector space

Proposition 2.1.13.
a) If B is a balanced subset of a t.v.s. X then so is B̄.
b) If B is a balanced subset of a t.v.s. X and o ∈ B̊ then B̊ is balanced.
Proof. (Sheet 3, Exercise 2 b) c))

Proof. of Theorem 2.1.10.

Necessity part.
Suppose that X is a t.v.s. then we aim to show that the filter of neighbour-
hoods of the origin F satisfies the properties 1,2,3,4,5. Let U ∈ F.
1. obvious, since every set U ∈ F is a neighbourhood of the origin o.
2. Since by the definition of t.v.s. the addition (x, y) 7→ x+y is a continuous
mapping, the preimage of U under this map must be a neighbourhood of
(o, o) ∈ X × X. Therefore, it must contain a rectangular neighbourhood
W × W 0 where W, W 0 ∈ F. Taking V = W ∩ W 0 we get the conclusion,
i.e. V + V ⊂ U .
3. By Proposition 2.1.7, fixed an arbitrary 0 6= λ ∈ K, the map x 7→ λ−1 x
of X into itself is continuous. Therefore, the preimage of any neighbour-
hood U of the origin must be also such a neighbourhood. This preimage
is clearly λU , hence λU ∈ F.
4. Suppose by contradiction that U is not absorbing. Then there exists
y ∈ X such that ∀n ∈ N we have that n1 y ∈ / U . This contradicts the
1
convergence of n y → o as n → ∞ (because U ∈ F must contain infinitely
many terms of the sequence ( n1 y)n∈N .
5. Since by the definition of t.v.s. the scalar multiplication K × X → X,
(λ, x) 7→ λx is continuous, the preimage of U under this map must be a
neighbourhood of (0, o) ∈ K × X. Therefore, it contains a rectangular
neighbourhood N ×W where N is a neighbourhood of 0 in the euclidean
topology on K and W ∈ F. On the other hand, there exists ρ > 0 such
that Bρ (0) := {λ ∈ K : |λ| ≤ ρ} ⊆ N . Thus Bρ (0) × W is contained in
the preimage of U under the scalar multiplication, i.e. λW ⊂ U for all
λ ∈ K with |λ| ≤ ρ. Hence, the set V = ∪|λ|≤ρ λW ⊂ U . Now V ∈ F
since each λW ∈ F by 3 and V is clearly balanced (since for any x ∈ V
there exists λ ∈ K with |λ| ≤ ρ s.t. x ∈ λW and therefore for any α ∈ K
with |α| ≤ 1 we get αx ∈ αλW ⊂ V because |αλ| ≤ ρ).
Sufficiency part.
Suppose that the conditions 1,2,3,4,5 hold for a filter F of the vector space X.
We want to show that there exists a topology τ on X such that F is the filter
of neighbourhoods of the origin w.r.t. to τ and (X, τ ) is a t.v.s. according to
Definition 2.1.1.

21
2. Topological Vector Spaces

Let us define for any x ∈ X the filter F(x) := {U + x : U ∈ F}. It is

easy to see that F(x) fulfills the properties (N1) and (N2) of Theorem 1.1.9.
In fact, we have:
• By 1 we have that ∀U ∈ F, o ∈ U , then ∀U ∈ F, x = o + x ∈ U + x, i.e.
∀A ∈ F(x), x ∈ A.
• Let A ∈ F(x) then A = U + x for some U ∈ F. By 2, we have that
there exists V ∈ F s.t. V + V ⊂ U . Define B := V + x ∈ F(x) and take
any y ∈ B then we have V + y ⊂ V + B ⊂ V + V + x ⊂ U + x = A. But
V + y belongs to the filter F(y) and therefore so does A.
By Theorem 1.1.9, there exists a unique topology τ on X such that F(x) is
the filter of neighbourhoods of each point x ∈ X and so for which in particular
F is the filter of neighbourhoods of the origin.
It remains to prove that the vector addition and the scalar multiplication
in X are continuous w.r.t. to τ .
• The continuity of the addition easily follows from the property 2. Indeed,
let (x0 , y0 ) ∈ X × X and take a neighbourhood W of its image x0 + y0 .
Then W = U + x0 + y0 for some U ∈ F. By 2, there exists V ∈ F s.t.
V + V ⊂ U and so (V + x0 ) + (V + y0 ) ⊂ W . This implies that the
preimage of W under the addition contains (V + x0 ) × (V + y0 ) which
is a neighbourhood of (x0 , y0 ).
• To prove the continuity of the scalar multiplication, let (λ0 , x0 ) ∈ K × X
and take a neighbourhood U 0 of λ0 x0 . Then U 0 = U + λ0 x0 for some
U ∈ F. By 2 and 5, there exists W ∈ F s.t. W + W + W ⊂ U and W
is balanced. By 4, W is also absorbing so there exists ρ > 0 (w.l.o.g we
can take ρ ≤ 1) such that ∀λ ∈ K with |λ| ≤ ρ we have λx0 ∈ W .
Suppose λ0 = 0 then λ0 x0 = o and U 0 = U . Now
Im(Bρ (0) × (W + x0 )) = {λy + λx0 : λ ∈ Bρ (0), y ∈ W }.
As λ ∈ Bρ (0) and W is absorbing, λx0 ∈ W . Also since |λ| ≤ ρ ≤ 1
for all λ ∈ Bρ (0) and since W is balanced, we have λW ⊂ W . Thus
Im(Bρ (0)×(W +x0 )) ⊂ W +W ⊂ W +W +W ⊂ U and so the preimage
of U under the scalar multiplication contains Bρ (0) × (W + x0 ) which is
a neighbourhood of (0, x0 ).
Suppose λ0 6= 0 and take σ = min{ρ, |λ0 |}. Then Im((Bσ (0) + λ0 ) ×
(|λ0 |−1 W + x0 ))={λ|λ0 |−1 y +λx0 +λ0 |λ0 |−1 y + λx0 : λ∈Bσ (0), y∈W }.
As λ ∈ Bσ (0), σ ≤ ρ and W is absorbing, λx0 ∈ W . Also since ∀λ ∈
Bσ (0) the modulus of λ|λ0 |−1 and λ0 |λ0 |−1 are both ≤ 1 and since W
is balanced, we have λ|λ0 |−1 W, λ0 |λ0 |−1 W ⊂ W . Thus Im(Bσ (0) +
λ0 × (|λ0 |−1 W + x0 )) ⊂ W + W + W + λ0 x0 ⊂ U + λ0 x0 and so the
preimage of U + λ0 x0 under the scalar multiplication contains Bσ (0) +
λ0 × (|λ0 |−1 W + x0 ) which is a neighbourhood of (λ0 , x0 ).

22
 2.1. Definition and main properties of a topological vector space

It easily follows from previous theorem that:

Corollary 2.1.14.
a) Every t.v.s. has always a base of closed neighbourhoods of the origin.
b) Every t.v.s. has always a base of balanced absorbing neighbourhoods of the
origin. In particular, it has always a base of closed balanced absorbing
neighbourhoods of the origin.
c) Proper subspaces of a t.v.s. are never absorbing. In particular, if M is an
open subspace of a t.v.s. X then M = X.
Proof. (Sheet 3, Exercise 3)

Let us show some further useful properties of the t.v.s.:

Proposition 2.1.15.
1. Every linear subspace of a t.v.s. endowed with the correspondent subspace
topology is itself a t.v.s..
2. The closure H of a linear subspace H of a t.v.s. X is again a linear
subspace of X.
3. Let X, Y be two t.v.s. and f : X → Y a linear map. f is continuous if
and only if f is continuous at the origin o.
Proof.
1. This clearly follows by the fact that the addition and the multiplication
restricted to the subspace are just a composition of continuous maps
(recall that inclusion is continuous in the subspace topology c.f. Defini-
tion 1.1.17).
2. Let x0 , y0 ∈ H and let us take any U ∈ F(o). By Theorem 2.1.10-
2, there exists V ∈ F(o) s.t. V + V ⊂ U . Then, by definition of
closure points, there exist x, y ∈ H s.t. x ∈ V + x0 and y ∈ V + y0 .
Therefore, we have that x + y ∈ H (since H is a linear subspace) and
x + y ∈ (V + x0 ) + (V + y0 ) ⊂ U + x0 + y0 . Hence, x0 + y0 ∈ H. Similarly,
one can prove that if x0 ∈ H, λx0 ∈ H̄ for any λ ∈ K.
3. Assume that f is continuous at o ∈ X and fix any x 6= o in X. Let U be
an arbitrary neighbourhood of f (x) ∈ Y . By Corollary 2.1.9, we know
that U = f (x) + V where V is a neighbourhood of o ∈ Y . Since f is
linear we have that:

f −1 (U ) = f −1 (f (x) + V ) ⊃ x + f −1 (V ).

By the continuity at the origin of X, we know that f −1 (V ) is a neigh-

bourhood of o ∈ X and so x + f −1 (V ) is a neighbourhood of x ∈ X.

23
2. Topological Vector Spaces

2.2 Hausdorff topological vector spaces

For convenience let us recall here the definition of Hausdorff space.
Definition 2.2.1. A topological space X is said to be Hausdorff or (T2) if
any two distinct points of X have neighbourhoods without common points; or
equivalently if two distinct points always lie in disjoint open sets.
Note that in a Hausdorff space, any set consisting of a single point is
closed but there are topological spaces with the same property which are not
Hausdorff and we will see in this section that such spaces are not t.v.s..
Definition 2.2.2. A topological space X is said to be (T1) if, given two
distinct points of X, each lies in a neighborhood which does not contain the
other point; or equivalently if, for any two distinct points, each of them lies in
an open subset which does not contain the other point.
It is easy to see that in a topological space which is (T1) all singletons are
closed (Sheet 4, Exercise 2).
From the definition it is clear that (T2) implies (T1) but in general the
inverse does not hold (c.f. Examples 1.1.40-4 for an example of topological
space which is T1 but not T2). However, the following results shows that a
t.v.s is Hausdorff if and only if it is (T1).
Proposition 2.2.3. A t.v.s. X is Hausdorff iff

∀ o 6= x ∈ X, ∃ U ∈ F(o) s.t. x ∈
/ U. (2.1)

Proof.
(⇒) Let (X, τ ) be Hausdorff. Then there exist U ∈ F(o) and V ∈ F(x)
s.t. U ∩ V = ∅. This means in particular that x ∈/ U.
(⇐) Assume that (2.1) holds and let x, y ∈ X with x 6= y, i.e. x − y 6= o.
Then there exists U ∈ F(o) s.t. x − y ∈
/ U . By (2) and (5) of Theorem 2.1.10,
there exists V ∈ F(o) balanced and s.t. V + V ⊂ U . Since V is balanced
V = −V then we have V − V ⊂ U . Suppose now that (V + x) ∩ (V + y) 6= ∅,
then there exists z ∈ (V + x) ∩ (V + y), i.e. z = v + x = w + y for some
v, w ∈ V . Then x − y = w − v ∈ V − V ⊂ U and so x − y ∈ U which is a
contradiction. Hence, (V + x) ∩ (V + y) = ∅ and by Corollary 2.1.9 we know
that V + x ∈ F(x) and V + y ∈ F(y). Hence, X is (T2).

Note that since the topology of a t.v.s. is translation invariant then the
previous proposition guarantees that a t.v.s is Hausdorff iff it is (T1). As a
matter of fact, we have the following result:

24
 2.3. Quotient topological vector spaces

Corollary 2.2.4. For a t.v.s. X the following are equivalent:

a) X is Hausdorff.
b) the intersection of all neighbourhoods of the origin o is just {o}.
c) {0} is closed.
Note that in a t.v.s. {0} is closed is equivalent to say that all singletons
are closed (and so that the space is (T1)).
Proof.
a)⇒ b) Let X be a Hausdorff t.v.s. space. Clearly, {o} ⊆ ∩U ∈F (o) U . Now
if b) does not hold, then there exists x ∈ ∩U ∈F (o) U with x 6= o. But by the
previous theorem we know that (2.1) holds and so there exists V ∈ F(o) s.t.
x∈ / V and so x ∈ / ∩U ∈F (o) U which is a contradiction.
b)⇒ c) Assume that ∩U ∈F (o) U = {o}. If x ∈ {o} then ∀ Vx ∈ F(x) we have
Vx ∩ {o} 6= ∅, i.e. o ∈ Vx . By Corollary 2.1.9 we know that each Vx = U + x
with U ∈ F(o). Then o = u + x for some u ∈ U and so x = −u ∈ −U . This
means that x ∈ ∩U ∈F (o) (−U ). Since every dilation is an homeomorphism and
b) holds, we have that x ∈ ∩U ∈F (o) U = {0}. Hence, x = 0 and so {o} = {o},
i.e. {o} is closed.
c)⇒ a) Assume that X is not Hausdorff. Then by the previous proposition
(2.1) does not hold, i.e. there exists x 6= o s.t. x ∈ U for all U ∈ F(o). This
means that x ∈ ∩U ∈F (o) U ⊆ ∩U ∈F (o)closed U = {o} By c), {o} = {o} and so
x = 0 which is a contradiction.

Example 2.2.5. Every vector space with an infinite number of elements en-
dowed with the cofinite topology is not a tvs. It is clear that in such topological
space all singletons are closed (i.e. it is T1). Therefore, if it was a t.v.s. then
by the previous results it should be a Hausdorff space which is not true as
showed in Example 1.1.40.

2.3 Quotient topological vector spaces

Quotient topology
Let X be a topological space and ∼ be any equivalence relation on X. Then
the quotient set X/∼ is defined to be the set of all equivalence classes w.r.t.
to ∼. The map φ : X → X/∼ which assigns to each x ∈ X its equivalence
class φ(x) w.r.t. ∼ is called the canonical map or quotient map. Note that φ is
surjective. We may define a topology on X/ ∼ by setting that: a subset U of
X/∼ is open iff the preimage φ−1 (U ) is open in X. This is called the quotient
topology on X/ ∼. Then it is easy to verify (Sheet 4, Exercise 2) that:

25
2. Topological Vector Spaces

• the quotient map φ is continuous.

• the quotient topology on X/∼ is the finest topology on X/∼ s.t. φ is
continuous.
Note that the quotient map φ is not necessarily open or closed.
Example 2.3.1. Consider R with the standard topology given by the modulus
and define the following equivalence relation on R:

x ∼ y ⇔ (x = y ∨ {x, y} ⊂ Z).

Let R/ ∼ be the quotient set w.r.t ∼ and φ : R → R/ ∼ the correspondent

quotient map. Let us consider the quotient topology on R/∼. Then φ is not
an open map. In fact, if U is an open proper subset of R containing an integer,
then φ−1 (φ(U )) = U ∪ Z which is not open in R with the standard topology.
Hence, φ(U ) is not open in R/∼ with the quotient topology.
For an example of quotient map which is not closed see Example 2.3.3 in
the following.

Quotient vector space

Let X be a vector space and M a linear subspace of X. For two arbitrary
elements x, y ∈ X, we define x ∼M y iff x − y ∈ M . It is easy to see
that ∼M is an equivalence relation: it is reflexive, since x − x = 0 ∈ M
(every linear subspace contains the origin); it is symmetric, since x − y ∈ M
implies −(x − y) = y − x ∈ M (if a linear subspace contains an element, it
contains its inverse); it is transitive, since x − y ∈ M , y − z ∈ M implies
x − z = (x − y) + (y − z) ∈ M (when a linear subspace contains two vectors, it
also contains their sum). Then X/M is defined to be the quotient set X/∼M ,
i.e. the set of all equivalence classes for the relation ∼M described above. The
canonical (or quotient) map φ : X → X/M which assigns to each x ∈ X its
equivalence class φ(x) w.r.t. the relation ∼M is clearly surjective. Using the
fact that M is a linear subspace of X, it is easy to check that:
1. if x ∼M y, then ∀ λ ∈ K we have λx ∼M λy.
2. if x ∼M y, then ∀ z ∈ X we have x + z ∼M y + z.
These two properties guarantee that the following operations are well-defined
on X/M :
• vector addition: ∀ φ(x), φ(y) ∈ X/M , φ(x) + φ(y) := φ(x + y)
• scalar multiplication: ∀ λ ∈ K, ∀ φ(x) ∈ X/M , λφ(x) := φ(λx)
X/M with the two operations defined above is a vector space and therefore
it is often called quotient vector space. Then the quotient map φ is clearly
linear.

26
 2.3. Quotient topological vector spaces

Quotient topological vector space

Let X be now a t.v.s. and M a linear subspace of X. Consider the quotient
vector space X/M and the quotient map φ : X → X/M defined in Section 2.3.
Since X is a t.v.s, it is in particular a topological space, so we can consider
on X/M the quotient topology defined in Section 2.3. We already know that
in this topological setting φ is continuous but actually the structure of t.v.s.
on X guarantees also that it is open.
Proposition 2.3.2. For a linear subspace M of a t.v.s. X, the quotient map-
ping φ : X → X/M is open (i.e. carries open sets in X to open sets in X/M )
when X/M is endowed with the quotient topology.

Proof. Let V open in X. Then we have

φ−1 (φ(V )) = V + M = ∪m∈M (V + m)

Since X is a t.v.s, its topology is translation invariant and so V + m is open

for any m ∈ M . Hence, φ−1 (φ(V )) is open in X as union of open sets. By
definition, this means that φ(V ) is open in X/M endowed with the quotient
topology.

It is then clear that φ carries neighborhoods of a point in X into neighbor-

hoods of a point in X/M and viceversa. Hence, the neighborhoods of the
origin in X/M are direct images under φ of the neighborhoods of the origin
in X. In conclusion, when X is a t.v.s and M is a subspace of X, we can
rewrite the definition of quotient topology on X/M in terms of neighborhoods
as follows: the filter of neighborhoods of the origin of X/M is exactly the im-
age under φ of the filter of neighborhoods of the origin in X.

It is not true, in general (not even when X is a t.v.s. and M is a subspace

of X), that the quotient map is closed.
Example 2.3.3.
Consider R2 with the euclidean topology and the hyperbola H := {(x, y) ∈ R2 :
xy = 1}. If M is one of the coordinate axes, then R2 /M can be identified
with the other coordinate axis and the quotient map φ with the orthogonal
projection on it. All these identifications are also valid for the topologies. The
hyperbola H is closed in R2 but its image under φ is the complement of the
origin on a straight line which is open.

27
2. Topological Vector Spaces

Corollary 2.3.4. For a linear subspace M of a t.v.s. X, the quotient space

X/M endowed with the quotient topology is a t.v.s..
Proof.
For convenience, we denote here by A the vector addition in X/M and just by
+ the vector addition in X. Let W be a neighbourhood of the origin o in X/M .
We aim to prove that A−1 (W ) is a neighbourhood of (o, o) in X/M × X/M .
The continuity of the quotient map φ : X → X/M implies that φ−1 (W )
is a neighbourhood of the origin in X. Then, by Theorem 2.1.10-2 (we can
apply the theorem because X is a t.v.s.), there exists V neighbourhood of
the origin in X s.t. V + V ⊆ φ−1 (W ). Hence, by the linearity of φ, we get
A(φ(V ) × φ(V )) = φ(V + V ) ⊆ W , i.e. φ(V ) × φ(V ) ⊆ A−1 (W ). Since φ is
also open, φ(V ) is a neighbourhood of the origin o in X/M and so A−1 (W ) is
a neighbourhood of (o, o) in X/M × X/M .
A similar argument gives the continuity of the scalar multiplication.

Proposition 2.3.5. Let X be a t.v.s. and M a linear subspace of X. Consider

X/M endowed with the quotient topology. Then the two following properties
are equivalent:
a) M is closed
b) X/M is Hausdorff
Proof.
In view of Corollary 2.2.4, (b) is equivalent to say that the complement of the
origin in X/M is open w.r.t. the quotient topology. But the complement of
the origin in X/M is exactly the image under φ of the complement of M in X.
Since φ is an open continuous map, the image under φ of the complement of M
in X is open in X/M iff the complement of M in X is open, i.e. (a) holds.

Corollary 2.3.6. If X is a t.v.s., then X/{o} endowed with the quotient

topology is a Hausdorff t.v.s.. X/{o} is said to be the Hausdorff t.v.s. asso-
ciated with the t.v.s. X. When a t.v.s. X is Hausdorff, X and X/{o} are
topologically isomorphic.
Proof.
Since X is a t.v.s. and {o} is a linear subspace of X, {o} is a closed linear
subspace of X. Then, by Corollary 2.3.4 and Proposition 2.3.5, X/{o} is a
Hausdorff t.v.s.. If in addition X is Hausdorff, then Corollary 2.2.4 guaran-
tees that {o} = {o} in X. Therefore, the quotient map φ : X → X/{o} is
also injective because in this case Ker(φ) = {o}. Hence, φ is a topological
isomorphism (i.e. bijective, continuous, open, linear) between X and X/{o}
which is indeed X/{o}.

28
 2.4. Continuous linear mappings between t.v.s.

2.4 Continuous linear mappings between t.v.s.

Let X and Y be two vector spaces over K and f : X → Y a linear map. We
define the image of f , and denote it by Im(f ), as the subset of Y :

Im(f ) := {y ∈ Y : ∃ x ∈ X s.t. y = f (x)}.

We define the kernel of f , and denote it by Ker(f ), as the subset of X:

Ker(f ) := {x ∈ X : f (x) = 0}.

Both Im(f ) and Ker(f ) are linear subspaces of Y and X, respectively. We

have then the diagram:
f i
X Im(f ) Y

f¯
φ

X/Ker(f )

where i is the natural injection of Im(f ) into Y , i.e. the mapping which to
each element y of Im(f ) assigns that same element y regarded as an element
of Y ; φ is the canonical map of X onto its quotient X/Ker(f ). The mapping
f¯ is defined so as to make the diagram commutative, which means that:

∀ x ∈ X, f (x) = f¯(φ(x)).

Note that
• f¯ is well-defined.
Indeed, if φ(x) = φ(y), i.e. x − y ∈ Ker(f ), then f (x − y) = 0 that is
f (x) = f (y) and so f¯(φ(x)) = f¯(φ(y)).
• f¯ is linear.
This is an immediate consequence of the linearity of f and of the linear
structure of X/Ker(f ).
• f¯ is a one-to-one linear map of X/Ker(f ) onto Im(f ).
The onto property is evident from the definition of Im(f ) and of f¯.
As for the one-to-one property, note that f¯(φ(x)) = f¯(φ(y)) means by
definition that f (x) = f (y), i.e. f (x − y) = 0. This is equivalent, by
linearity of f , to say that x−y ∈ Ker(f ), which means that φ(x) = φ(y).

29
2. Topological Vector Spaces

The set of all linear maps (continuous or not) of a vector space X into another
vector space Y is denoted by L(X; Y ). Note that L(X; Y ) is a vector space
for the natural addition and multiplication by scalars of functions. Recall that
when Y = K, the space L(X; Y ) is denoted by X ∗ and it is called the algebraic
dual of X (see Definition 1.2.4).

Let us not turn to consider linear mapping between two t.v.s. X and Y .
Since they posses a topological structure, it is natural to study in this setting
continuous linear mappings.
Proposition 2.4.1. Let f : X → Y a linear map between two t.v.s. X and Y .
If Y is Hausdorff and f is continuous, then Ker(f ) is closed in X.
Proof.
Clearly, Ker(f ) = f −1 ({o}). Since Y is a Hausdorff t.v.s., {o} is closed in Y
and so, by the continuity of f , Ker(f ) is also closed in Y .

Note that Ker(f ) might be closed in X also when Y is not Hausdorff. For
instance, when f ≡ 0 or when f is injective and X is Hausdorff.
Proposition 2.4.2. Let f : X → Y a linear map between two t.v.s. X and Y .
The map f is continuous if and only if the map f¯ is continuous.
Proof.
Suppose f continuous and let U be an open subset in Im(f ). Then f −1 (U )
is open in X. By definition of f¯, we have f¯−1 (U ) = φ(f −1 (U )). Since the
quotient map φ : X → X/Ker(f ) is open, φ(f −1 (U )) is open in X/Ker(f ).
Hence, f¯−1 (U ) is open in X/Ker(f ) and so the map f¯ is continuous. Vicev-
ersa, suppose that f¯ is continuous. Since f = f¯ ◦ φ and φ is continuous, f is
also continuous as composition of continuous maps.

In general, the inverse of f¯, which is well defined on Im(f ) since f¯ is in-
jective, is not continuous. In other words, f¯ is not necessarily bi-continuous.

The set of all continuous linear maps of a t.v.s. X into another t.v.s. Y is
denoted by L(X; Y ) and it is a vector subspace of L(X; Y ). When Y = K, the
space L(X; Y ) is usually denoted by X 0 which is called the topological dual of
X, in order to underline the difference with X ∗ the algebraic dual of X. X 0
is a vector subspace of X ∗ and is exactly the vector space of all continuous
linear functionals, or continuous linear forms, on X. The vector spaces X 0 and
L(X; Y ) will play an important role in the forthcoming and will be equipped
with various topologies.

30
 2.5. Completeness for t.v.s.

2.5 Completeness for t.v.s.

This section aims to treat completeness for most general types of topological
vector spaces, beyond the traditional metric framework. As well as in the case
of metric spaces, we need to introduce the definition of a Cauchy sequence in
a t.v.s..
Definition 2.5.1. A sequence S := {xn }n∈N of points in a t.v.s. X is said to
be a Cauchy sequence if

∀ U ∈ F(o) in X, ∃ N ∈ N : xm − xn ∈ U, ∀m, n ≥ N. (2.2)

This definition agrees with the usual one if the topology of X is defined
by a translation-invariant metric d. Indeed, in this case, a basis of neigh-
bourhoods of the origin is given by all the open balls centered at the origin.
Therefore, {xn }n∈N is a Cauchy sequence in such (X, d) iff ∀ ε > 0, ∃ N ∈ N :
xm − xn ∈ Bε (o), ∀m, n ≥ N , i.e. d(xm , xn ) = d(xm − xn , o) < ε.

By using the subsequences Sm := {xn ∈ S : n ≥ m} of S, we can easily

rewrite (2.2) in the following way

∀ U ∈ F(o) in X, ∃ N ∈ N : SN − SN ⊂ U.

As we have already observed in Chapter 1, the collection B := {Sm : m ∈ N}

is a basis of the filter FS associated with the sequence S. This immediately
suggests what the definition of a Cauchy filter should be:
Definition 2.5.2. A filter F on a subset A of a t.v.s. X is said to be a Cauchy
filter if
∀ U ∈ F(o) in X, ∃ M ⊂ A : M ∈ F and M − M ⊂ U.
In order to better illustrate this definition, let us come back to our refer-
ence example of a t.v.s. X whose topology is defined by a translation-invariant
metric d. For any subset M of (X, d), recall that the diameter of M is defined
as diam(M ) := supx,y∈M d(x, y). Now if F is a Cauchy filter on X then, by
definition, for any ε > 0 there exists M ∈ F s.t. M − M ⊂ Bε (o) and this
simply means that diam(M ) ≤ ε. Therefore, Definition 2.5.2 can be rephrased
in this case as follows: a filter F on a subset A of such a metric t.v.s. X is a
Cauchy filter if it contains subsets of A of arbitrarily small diameter.

Going back to the general case, the following statement clearly holds.
Proposition 2.5.3. The filter associated with a Cauchy sequence in a t.v.s.
X is a Cauchy filter.

31
2. Topological Vector Spaces

Proposition 2.5.4.
Let X be a t.v.s.. Then the following properties hold:
a) The filter of neighborhoods of a point x ∈ X is a Cauchy filter on X.
b) A filter finer than a Cauchy filter is a Cauchy filter.
c) Every converging filter is a Cauchy filter.

Proof.
a) Let F(x) be the filter of neighborhoods of a point x ∈ X and let U ∈ F(o).
By Theorem 2.1.10, there exists V ∈ F(o) such that V − V ⊂ U and
so such that (V + x) − (V + x) ⊂ U . Since X is a t.v.s., we know that
F(x) = F(o) + x and so M := V + x ∈ F(x). Hence, we have proved that
for any U ∈ F(o) there exists M ∈ F(x) s.t. M − M ⊂ U , i.e. F(x) is a
Cauchy filter.
b) Let F and F 0 be two filters of subsets of X such that F is a Cauchy
filter and F ⊆ F 0 . Since F is a Cauchy filter, by Definition 2.5.2, for any
U ∈ F(o) there exists M ∈ F s.t. M − M ⊂ U . But F 0 is finer than F, so
M belongs also to F 0 . Hence, F 0 is obviously a Cauchy filter.
c) If a filter F converges to a point x ∈ X then F(x) ⊆ F (see Defini-
tion 1.1.27). By a), F(x) is a Cauchy filter and so b) implies that F itself
is a Cauchy filter.

The converse of c) is in general false, in other words not every Cauchy

filter converges.

Definition 2.5.5. A subset A of a t.v.s. X is said to be complete if every

Cauchy filter on A converges to a point x of A.

It is important to distinguish between completeness and sequentially com-

pleteness.

Definition 2.5.6. A subset A of a t.v.s. X is said to be sequentially complete

if any Cauchy sequence in A converges to a point in A.

It is easy to see that complete always implies sequentially complete. The

converse is in general false (see Example 2.5.9). We will encounter an impor-
tant class of t.v.s., the so-called metrizable spaces, for which the two notions
coincide.

Proposition 2.5.7. If a subset A of a t.v.s. X is complete then A is sequen-

tially complete.

32
 2.5. Completeness for t.v.s.

Proof.
Let S := {xn }n∈N a Cauchy sequence of points in A. Then Proposition 2.5.3
guarantees that the filter FS associated to S is a Cauchy filter in A. By the
completeness of A we get that there exists x ∈ A such that FS converges to x.
This is equivalent to say that the sequence S is convergent to x ∈ A (see
Proposition 1.1.29). Hence, A is sequentially complete.

Before showing an example of a subset of a t.v.s. which is sequentially

complete but not complete, let us introduce two useful properties about com-
pleteness in t.v.s..
Proposition 2.5.8.
a) In a Hausdorff t.v.s. X, any complete subset is closed.
b) In a complete t.v.s. X, any closed subset is complete.
Example 2.5.9.
Let X := Rd where d > ℵ0 endowed with the product topology given by
considering each copy of R equipped withQ the usual topology given by the mod-
ulus. For convenience we write X = i∈J R with |J| = d > ℵ0 . Note that X
is a Hausdorff t.v.s. as it is product of Hausdorff t.v.s.. Denote by H the sub-
set of X consisting of all vectors x = (xi )i∈J in X with only countably many
non-zero coordinates xi . Claim: H is sequentially complete but not complete.
Proof. of Claim. Let us first make some observations on H.
• H is strictly contained in X.
Indeed, any vector y ∈ X with all non-zero coordinates does not belong
to H because d > ℵ0 .
• H is dense in X.
In fact, let x = (xi )i∈J ∈ X and U a neighbourhood Q of x in X. Then,
by definition of product topology on X, there exist i∈J Ui ⊆ U s.t.
Ui ⊆ R neighbourhood of xi in R for all i ∈ J and Ui 6= R for all i ∈ I
where I ⊂ J with |I| < ∞. Take y := (yi )i∈J s.t. yi ∈ Ui for all i ∈ J
with yi 6= 0 for all i ∈ I and yi = 0 otherwise. Then clearly y ∈ U
but also y ∈ H because it has only finitely many non-zero coordinates.
Hence, U ∩ H 6= ∅ and so H = X.
Now suppose that H is complete, then by Proposition 2.5.8-a) we have that
H is closed. Therefore, by the density of H in X, it follows that H = H = X
which contradicts the first of the property above. Hence, H is not complete.
In the end, let us show that H is sequentially complete. Let (xn )n∈N a
(i)
Cauchy sequence of vectors xn = (xn )i∈J in H. Then for each i ∈ J we have
(i)
that the sequence of the i − th coordinates (xn )n∈N is a Cauchy sequence

33
2. Topological Vector Spaces

in R. By the completeness (i.e. the sequentially completeness) of R we have

(i)
that for each i ∈ J, the sequence (xn )n∈N converges to a point x(i) ∈ R. Set
x := (x(i) )i∈J . Then:
(i)
• x ∈ H, because for each n ∈ N only countably many xn 6= 0 and so
only countably many x(i) 6= 0.
• the sequence (xn )n∈N convergesQ to x in H. In fact, for any U neighbour-
hood of x in X there exist i∈J Ui ⊆ U s.t. Ui ∈ R neighbourhood of
xi in R for all i ∈ J and Ui 6= R for all i ∈ I where I ⊂ J with |I| < ∞.
(i)
Since for each i ∈ J, the sequence (xn )n∈N converges to x(i) in R, we
(i)
get that for each i ∈ J there exists Ni ∈ N s.t. xn ∈ Ui for all n ≥ Ni .
Take N := maxi∈I Ni (the max exists because I is finite). Then for each
(i)
i ∈ J we get xn ∈ Ui for all n ≥ N , i.e. xn ∈ U for all n ≥ N which
proves the convergence of (xn )n∈N to x.
Hence, we have showed that every Cauchy sequence in H is convergent.

In order to prove Proposition 2.5.8, we need two small lemmas regarding

convergence of filters in a topological space.
Lemma 2.5.10. Let F be a filter of a topological Hausdorff space X. If F
converges to x ∈ X and also to y ∈ X, then x = y.
Proof.
Suppose that x 6= y. Then, since X is Hausdorff, there exists V ∈ F(x) and
W ∈ F(y) such that V ∩ W = ∅. On the other hand, we know by assumption
that F → x and F → y that is F(x) ⊆ F and F(y) ⊆ F (see Definition 1.1.27).
Hence, V, W ∈ F. Since filters are closed under finite intersections, we get
that V ∩ W ∈ F and so ∅ ∈ F which contradicts the fact that F is a filter.

Lemma 2.5.11. Let A be a subset of a topological space X. Then x ∈ A if

and only if there exists a filter F of subsets of X such that A ∈ F and F
converges to x.
Proof.
Let x ∈ A, i.e. for any U ∈ F(x) in X we have U ∩ A 6= ∅. Set F :=
{F ⊆ X|U ∩ A ⊆ F for some U ∈ F(x)}. It is easy to see that F is a filter of
subsets of X. Therefore, for any U ∈ F(x), U ∩ A ∈ F and U ∩ A ⊆ U imply
that U ∈ F, i.e. F(x) ⊆ F. Hence, F → x.
Viceversa, suppose that F is a filter of X s.t. A ∈ F and F converges to
x. Let U ∈ F(x). Then U ∈ F since F(x) ⊆ F by definition of convergence.
Since also A ∈ F by assumption, we get U ∩ A ∈ F and so U ∩ A 6= ∅.

34
 2.5. Completeness for t.v.s.

Proof. of Proposition 2.5.8

a) Let A be a complete subset of a Hausdorff t.v.s. X and let x ∈ A. By
Lemma 2.5.11, x ∈ A implies that there exists a filter F of subsets of X
s.t. A ∈ F and F converges to x. Therefore, by Proposition 2.5.4-c), F is
a Cauchy filter. Consider now FA := {U ∈ F : U ⊆ A} ⊂ F. It is easy to
see that FA is a Cauchy filter on A and so the completeness of A ensures
that FA converges to a point y ∈ A. Hence, any nbhood V of y in A
belongs to FA and so to F. By definition of subset topology, this means
that for any nbhood U of y in X we have U ∩ A ∈ F and so U ∈ F (since
F is a filter). Then F converges to y. Since X is Hausdorff, Lemma 2.5.10
establishes the uniqueness of the limit point of F, i.e. x = y and so A = A.
b) Let A be a closed subset of a complete t.v.s. X and let FA be any Cauchy
filter on A. Take the filter F := {F ⊆ X|B ⊆ F for some B ∈ FA }. It is
clear that F contains A and is finer than the Cauchy filter FA . Therefore,
by Proposition 2.5.4-b), F is also a Cauchy filter. Then the completeness
of the t.v.s. X gives that F converges to a point x ∈ X, i.e. F(x) ⊆ F.
By Lemma 2.5.11, this implies that actually x ∈ A and, since A is closed,
that x ∈ A. Now any neighbourhood of x ∈ A in the subset topology is
of the form U ∩ A with U ∈ F(x). Since F(x) ⊆ F and A ∈ F, we have
U ∩ A ∈ F. Therefore, there exists B ∈ FA s.t. B ⊆ U ∩ A ⊂ A and so
U ∩ A ∈ FA . Hence, FA converges x ∈ A, i.e. A is complete.

When a t.v.s. is not complete, it makes sense to ask if it is possible to

embed it in a complete one. We are going to describe an abstract procedure
that allows to always associate to an arbitrary Hausdorff t.v.s. X a complete
Hausdorff t.v.s. X̂ called the completion of X. Before doing that, we need
to introduce uniformly continuous functions between t.v.s. and state some of
their fundamental properties.
Definition 2.5.12. Let X and Y be two t.v.s. and let A be a subset of X. A
mapping f : A → Y is said to be uniformly continuous if for every neighbor-
hood V of the origin in Y , there exists a neighborhood U of the origin in X
such that for all pairs of elements x1 , x2 ∈ A
x1 − x2 ∈ U ⇒ f (x1 ) − f (x2 ) ∈ V.

Proposition 2.5.13. Let X and Y be two t.v.s. and let A be a subset of X.

a) If f : A → Y is uniformly continuous, then the image under f of a Cauchy
filter on A is a Cauchy filter on Y .
b) If A is a linear subspace of X, then every continuous linear map from A
to Y is uniformly continuous.
Proof. (Sheet 6, Exercise 2)

35
2. Topological Vector Spaces

Theorem 2.5.14.
Let X and Y be two Hausdorff t.v.s., A a dense subset of X, and f : A → Y
a uniformly continuous mapping. If Y is complete the the following hold.
a) There exists a unique continuous mapping f¯ : X → Y which extends f ,
i.e. such that for all x ∈ A we have f¯(x) = f (x).
b) f¯ is uniformly continuous.
c) If we additionally assume that f is linear and A is a linear subspace of X,
then f¯ is linear.

Proof. (Sheet 6, Exercise 3)

Let us now state and prove the theorem on completion of a t.v.s..

Theorem 2.5.15.
Let X be a Haudorff t.v.s.. Then there exists a complete Hausdorff t.v.s. X̂
and a mapping i : X → X̂ with the following properties:
a) The mapping i is a topological monomorphism.
b) The image of X under i is dense in X̂.
c) For every complete Hausdorff t.v.s. Y and for every continuous linear map
f : X → Y , there is a continuous linear map fˆ : X̂ → Y such that the
following diagram is commutative:
f
X Y
i
fˆ
X̂

Furthermore:
I) Any other pair (X̂1 , i1 ), consisting of a complete Hausdorff t.v.s. X̂1
and of a mapping i1 : X → X̂1 such that properties (a) and (b) hold
substituting X̂ with X̂1 and i with i1 , is topologically isomorphic to (X̂, i).
This means that there is a topological isomorphism j of X̂ onto X̂1 such
that the following diagram is commutative:
i1
X X̂1
i
j

X̂

II) Given Y and f as in property (c), the continuous linear map fˆ is unique.

36
 2.5. Completeness for t.v.s.

Proof.
1) The set X̂
Define the following relation on the collection of all Cauchy filters (c.f.) on X:

F ∼R G ⇔ ∀ U nbhood of the origin in X, ∃ A ∈ F, ∃ B ∈ G s.t. A − B ⊂ U.

The relation (R) is actually an equivalence relation. In fact:

• reflexive: If F is a c.f. on X, then by Definition 2.5.2 we have that for
any U nbhood of the origin in X there exists A ∈ F s.t. A − A ⊂ U , i.e.
F ∼R F.
• symmetric: If F and G are c.f. on X s.t. F ∼R G, then by definition of
(R) we have that for any U nbhood of the origin in X there exist A ∈ F
and B ∈ G s.t. A − B ⊂ U . This implies that B − A ⊂ −U , which gives
G ∼R F considering that −U is a generic nbhood of the origin in X in
the same right as U .
• transitive: Let F, G, H be c.f. on X s.t. F ∼R G and G ∼R H. Take any
U nbhood of the origin in X, then Theorem 2.1.10 ensures that there
exists V nbhood of the origin in X s.t. V + V ⊂ U . By definition of
(R), there exists A ∈ F, B1 , B2 ∈ G and C ∈ H s.t. A − B1 ⊂ V and
B2 −C ⊂ V . This clearly implies A−(B1 ∩B2 ) ⊂ V and (B1 ∩B2 )−C ⊂
V . By adding we obtain

A − C ⊂ A − (B1 ∩ B2 ) + (B1 ∩ B2 ) − C ⊂ V + V ⊂ U.

We define X̂ as the quotient of the set of all c.f. on X w.r.t. the equivalence
relation (R). Hence, an element x̂ of X̂ is an equivalence class of c.f. on X
w.r.t. (R).

2) Operations on X̂
Multiplication by a scalar
Let 0 6= λ ∈ K and let x̂ be a generic element of X̂. For any F any repre-
sentative of x̂, we define λx̂ to be the equivalence class w.r.t. (R) of the filter
λF := {λA : A ∈ F}, i.e.

λx̂ := {G c.f. on X : G ∼R λF}.

It is easy to check that this definition does not depend on the choice of the
representative F of x̂ (see Sheet 7, Exercise 1).
When λ = 0, we have λx̂ = ô, where ô is the equivalence class w.r.t. (R)
of the filter of neighborhoods of the origin o in X (or, which is the same,
of the Cauchy filter consisting of all the subsets of X which contain o).

37
2. Topological Vector Spaces

Vector addition
Let x̂ and ŷ be two arbitrary elements of X̂, and F (resp. G) a representative
of x̂ (resp. ŷ). We define x̂ + ŷ to be the equivalence class w.r.t. (R) of the
the filter F + G := {C ⊆ X : A + B ⊆ C for some A ∈ F, B ∈ G}, i.e.

x̂ + ŷ := {H c.f. on X : H ∼R F + G}.

Note that this vector addition is well-defined because its definition does not
depend on the choice of the representative F of x̂ and G of ŷ (see Sheet 7,
Exercise 1).

3) Topology on X̂
Let U be an arbitrary nbhood of the origin in X. Define

Û := {x̂ ∈ X̂ : U ∈ F for some F ∈ x̂}. (2.3)

and consider the collection B̂ := {Û : U nbhood of the origin in X}. The filter
generated by B̂ fulfills all the properties in Theorem 2.1.10 (see Sheet 7, Ex-
ercise 2) and therefore, there exists a unique topology on X̂ compatible with
the vector space structure defined in Step 2 s.t. this is its filter of nbhoods of
the origin ô ∈ X̂. Clearly, B̂ is a basis of nbhoods of the origin ô ∈ X̂ w.r.t.
to such a topology.

4) X̂ is a Hausdorff t.v.s.
So far we have constructed a t.v.s. X̂. In this step, we aim to prove that X̂ is
also Hausdorff. By Proposition 2.2.3, it is enough to show that for any x̂ ∈ X̂
with ô 6= x̂ there exists a nbhood V̂ of the origin ô in X̂ s.t. x̂ ∈
/ V̂ .
Since ô 6= x̂, for any F ∈ x̂ and for any Fo ∈ ô we have F 6∼R Fo . Take
F0 := {E ⊆ X : o ∈ E}, then the fact that F 6∼R Fo means that there exists
U nbhood of the origin in X s.t. ∀ A ∈ F and ∀ Ao ∈ Fo we have A − Ao 6⊂ U .
In particular, {o} ∈ Fo and so ∀ A ∈ F we get A 6⊂ U , which simply means
that U ∈ / F. By Theorem 2.1.10 applied to the t.v.s. X, we can always find
another nbohood V of the origin in X s.t. V + V ⊂ U .
Claim: V does not belong to any representative of x̂. This means, in view
of the definition (2.3), that x̂ ∈
/ V̂ . Hence, as observed at the beginning, the
conclusion follows by Proposition 2.2.3.
Let us finally prove the claim. If F 0 is any representative of x̂, then F ∼R F 0 ,
i.e. ∃ A ∈ F and ∃ A0 ∈ F 0 s.t. A − A0 ⊂ V . Suppose that V ∈ F 0 then
A0 ∩ V ∈ F 0 and so A0 ∩ V 6= ∅. Therefore, we clearly have A − (A0 ∩ V ) ⊂ V
which implies
A ⊂ V + (A0 ∩ V ) ⊂ V + V ⊂ U.

38
 2.5. Completeness for t.v.s.

/ F0
Since A ∈ F, this proves that U ∈ F which is a contradiction. Then V ∈
0
for all F ∈ x̂ that is exactly our claim.

5) Existence of i : X → X̂
We define the image of a point x ∈ X under the mapping i : X → X̂ to be
the equivalence class w.r.t. (R) of the filter F(x) of neighborhoods of x in X,
i.e.
∀ x ∈ X, i(x) := {F c.f. on X : F ∼R F(x)}.

Note that the following properties hold.

Lemma 2.5.16.
a) Two c.f. filters on X converging to the same point are equivalent w.r.t. (R)
b) If two c.f. filters F and F 0 on X are s.t. F ∼R F 0 and F 0 converges to
x ∈ X then also F converges to x.

Proof. (Sheet 7, Exercise 3)

The previous lemma clearly proves that

i(x) ≡ {F c.f. on X : F → x)}.

6) i is a topological monomorphism (i.e. (a) holds)

i is injective
(see Sheet 7, Exercise 4).
i is a topological homomorphism
For the linearity of i see Sheet 7, Exercise 4. We aim to show here that i is
both open and continuous on X.
To prove that i is open, we need to show that for any nbhood U of the
origin in X the image i(U ) is a nbhood of the origin in i(X) endowed with
the subset topology induced by the topology on X̂. Therefore, it suffices to
show that for any nbhood U of the origin in X there exists U1 nbhood of the
origin in X s.t.
Uˆ1 ∩ i(X) ⊆ i(U ) (2.4)

where Uˆ1 is defined as in (2.3).

To show the continuity of i, we need to prove that for any nbhood V̂
of the origin in i(X) the preimage i−1 (V̂ ) is a nbhood of the origin in X.
Now any nbhood of the origin in i(X) is of the form Uˆ1 ∩ i(X) for some U1
nbhood of the origin in X. Therefore, it is enough to show that for any U1

39
2. Topological Vector Spaces

nbhood of the origin in X there exists another U nbhood of the origin in X

s.t. U ⊆ i−1 (Uˆ1 ∩ i(X)) i.e.

i(U ) ⊆ Uˆ1 ∩ i(X) (2.5)

In order to prove (2.4) and (2.5), we shall prove the following:

i(V̊ ) ⊆ V̂ ∩ i(X) ⊆ i(V ), ∀ V nbhood of the origin in X. (2.6)

Indeed, if (2.6) holds then the first inclusion immediately shows (2.5) (for any
U1 nbhood of the origin in X take U := U˚1 and apply the first inclusion of
(2.6) to V = U1 ). Moreover, (2.4) follows by combining the fact that for any
nbhood U of the origin in X exists another nbhood U1 of the origin in X s.t.
U1 = U1 ⊆ U (c.f. Sheet 3, Ex3-a)) together with the second inclusion of (2.6)
(applied to U1 ).
It remains to prove that (2.6) holds. Let V be any nbhood of the origin
in X, then for any x ∈ V̊ we clearly have that V is a nbhood of x, which
means that V belongs to a representative of i(x), i.e. i(x) ∈ V̂ . Hence,
i(V̊ ) ⊆ V̂ ∩ i(X). Now take ŷ ∈ V̂ ∩ i(X), i.e. ŷ = i(x) for some x ∈ X s.t.
i(x) ∈ V̂ . Then, by definition (2.3), we have that V ∈ F for some F ∈ i(x)
or in other words that V belongs to some filter F converging to x. Let W
be another nbhood of the origin in X then W + x is a nbhood of x in X and
so W + x ∈ F (since F → x). Hence, V ∩ (W + x) ∈ F which implies that
V ∩ (W + x) 6= ∅ i.e. x ∈ V . This means that ŷ = i(x) ∈ i(V ) which proves
V̂ ∩ i(X) ⊆ i(V ).

7) i(X) = X̂ (i.e. (b) holds)

Let xˆo ∈ X̂ and let N be any nbhood of xˆo in X̂. It suffices to consider the
neighborhoods N of the form Û + xˆ0 where Û is defined by (2.3) for some U
nbhood of the origin in X. We aim to prove that (Û + xˆo ) ∩ i(X) 6= ∅.
By Theorem 2.1.10, we know that for any U nbhood of the origin in X
there exists V nbhood of the origin in X s.t. V + V ⊂ U . Let Fo be a
representative of xˆ0 , then Fo is a c.f. on X and so there exists Ao ∈ Fo s.t.
Ao − Ao ⊂ V . Fix an element x ∈ Ao . Then we get:

(V + x) − Ao ⊂ V + Ao − Ao ⊂ V + V ⊂ U. (2.7)

Since V + x is a nbhood of x in X, V + x belongs to any Cauchy filter F con-

verging to x and so V + x ∈ F for any F ∈ i(x). Then (V + x) − Ao ∈ F − Fo
and so (2.7) gives U ∈ F − Fo i.e. i(x) − xˆo ∈ Û . Hence, we found that there
exists x ∈ X s.t. i(x) ∈ Û + xˆo which gives the conclusion.

40
 2.5. Completeness for t.v.s.

8) X̂ is complete
Let F̂ be a Cauchy filter on X̂. We aim to prove that there exists an element
x̂ ∈ X̂ s.t. F̂ → x̂.
Consider the filter

F̂ 0 := {Ĝ ⊂ X̂ : M̂ +Û ⊂ Ĝ for some M̂ ∈ F̂ and Û nbhood of the origin in X̂}.

Note that F̂ 0 ⊂ F̂ and F̂ 0 is also a Cauchy filter on X̂. In fact, since X̂ is a

t.v.s., for any Û nbhood of the origin in X̂ there exists V̂0 balanced nbhood of
the origin in X̂ s.t. V̂0 + V̂0 + V̂0 ⊂ U . Take V̂ := 31 V̂0 which is also a nbhood
of the origin in X̂, then

V̂ + V̂ − V̂ ⊂ V̂0 + V̂0 + V̂0 ⊂ U.

Since F̂ is a Cauchy filter, there exists M̂ ∈ F̂ s.t. M̂ − M̂ ⊂ V̂ . Then

(M̂ + V̂ ) − (M̂ + V̂ ) ⊂ V̂ + V̂ − V̂ ⊂ U

Now let us consider the family of subsets of i(X) given by

F 0 := {Â ∩ i(X) : Â ∈ F̂ 0 }.

It is possible to prove that F 0 is a filter on i(X) and actually a Cauchy filter

(see Sheet 7, Exercise 5). Moreover, since we proved that i is a topological
isomorphism between X and i(X), we have that i−1 (F 0 ) is a Cauchy filter on
X. Take
x̂ := {F c.f. on X : F ∼R i−1 (F 0 )}.
Then F̂ converges to x̂ (see Sheet 7, Exercise 5).

9) Proof of the universal property (i.e. (c) and (II))

We can now identify X with i(X) and so regard X as a dense linear subspace
of X̂. Since f : X → Y is continuous and linear by assumption, it is also
uniformly continuous by Proposition 2.5.13. Then applying Theorem 2.5.14
with X replaced by X̂ and A by X we get both the properties (c) and (II).

10) Uniqueness of X̂ up to isomorphism (proof of (I))

Since by assumption X̂1 is a complete Hausdorff t.v.s. and i1 : X → X̂1
is a topological monomorphism (in particular i1 is a continuous linear map-
ping), we have by (c) that there exists a unique continuous linear map iˆ1 s.t.
iˆ1 (i(x)) = i1 (x) for any x ∈ X. Let us define j := iˆ1 . On the other hand,
let us define f : i1 (X) → X̂ by f (i1 (x)) = i(x) for any x ∈ X. Since f is

41
2. Topological Vector Spaces

clearly linear and continuous and i1 (X) is a linear subspace of X̂, f is uni-
formly continuous and so by Theorem 2.5.14 we get that there exists a unique
fˆ : X̂1 → X̂ continuous and linear s.t. fˆ(i1 (x)) = f (i1 (x)) for any x ∈ X.
Using the density of i(X) in X̂, the density of i1 (X) in X̂1 and the continuity
of the mappings involved, it is easy to check that

fˆ(j(x̂)) = x̂ ∀ x̂ ∈ X̂

and that
j(fˆ(xˆ1 )) = xˆ1 ∀ , xˆ1 ∈ X̂1 .
This means that j and f are the inverse of each other and that both are
topological isomorphisms.

42
 Chapter 3

Finite dimensional topological vector spaces

3.1 Finite dimensional Hausdorff t.v.s.

Let X be a vector space over the field K of real or complex numbers. We know
from linear algebra that the (algebraic) dimension of X, denoted by dim(X),
is the cardinality of a basis of X. If dim(X) is finite, we say that X is finite
dimensional otherwise X is infinite dimensional. In this section we are going
to focus on finite dimensional vector spaces.
Let {e1 , . . . , ed } be a basis of X, i.e. dim(X) = d. Given any vector x ∈ X
there exist unique x1 , . . . , xd ∈ K s.t. x = x1 e1 +· · · xd ed . This can be precisely
expressed by saying that the mapping

Kd → X
(x1 , . . . , xd ) 7→ x1 e1 + · · · xd ed

is an algebraic isomorphism (i.e. linear and bijective) between X and Kd . In

other words: If X is a finite dimensional vector space then X is algebraically
isomorphic to Kdim(X) .
If now we give to X the t.v.s. structure and we consider K endowed with
the euclidean topology, then it is natural to ask if such an algebraic isomor-
phism is by any chance a topological one, i.e. if it preserves the t.v.s. structure.
The following theorem shows that if X is a finite dimensional Hausdorff t.v.s.
then the answer is yes: X is topologically isomorphic to Kdim(X) . It is worth
to observe that usually in applications we deal always with Hausdorff t.v.s.,
therefore it makes sense to mainly focus on them.
Theorem 3.1.1. Let X be a finite dimensional Hausdorff t.v.s. over K (where
K is endowed with the euclidean topology). Then:
a) X is topologically isomorphic to Kd , where d = dim(X).

43
3. Finite dimensional topological vector spaces

b) Every linear functional on X is continuous.

c) Every linear map of X into any t.v.s. Y is continuous.
Before proving the theorem let us recall some lemmas about the continuity
of linear functionals on t.v.s..
Lemma 3.1.2.
Let X be a t.v.s. over K and v ∈ X. Then the following mapping is continuous.

ϕv : K → X
ξ 7→ ξv.

Proof. For any ξ ∈ K, we have ϕv (ξ) = M (ψv (ξ)), where ψv : K → K × X

given by ψ(ξ) := (ξ, v) is clearly continuous by definition of product topology
and M : K × X → X is the scalar multiplication in the t.v.s. X which is
continuous by definition of t.v.s.. Hence, ϕv is continuous as composition of
continuous mappings.

Lemma 3.1.3. Let X be a t.v.s. over K and L a linear functional on X.

Assume L(x) 6= 0 for some x ∈ X. Then the following are equivalent:
a) L is continuous.
b) The null space Ker(L) is closed in X
c) Ker(L) is not dense in X.
d) L is bounded in some neighbourhood of the origin in X.

Proof. (see Sheet 4, Exercise 4)

Proof. of Theorem 3.1.1

Let {e1 , . . . , ed } be a basis of X and let us consider the mapping

ϕ: Kd → X
(x1 , . . . , xd ) 7→ x1 e1 + · · · xd ed .

As noted above, this is an algebraic isomorphism. Therefore, to conclude a)

it remains to prove that ϕ is also a homeomorphism.
Step 1: ϕ is continuous.
When d = 1, we simply have ϕ ≡ ϕe1 and so we are done by Lemma
3.1.2. When d > 1, for any (x1 , . . . , xd ) ∈ Kd we can write: ϕ(x1 , . . . , xd ) =
A(ϕe1 (x1 ), . . . , ϕed (xd )) = A((ϕe1 × · · · × ϕed )(x1 , . . . , xd )) where each ϕej is
defined as above and A : X × X → X is the vector addition in the t.v.s. X.
Hence, ϕ is continuous as composition of continuous mappings.

44
 3.1. Finite dimensional Hausdorff t.v.s.

Step 2: ϕ is open and b) holds.

We prove this step by induction on the dimension dim(X) of X.
For dim(X) = 1, it is easy to see that ϕ is open, i.e. that the inverse of ϕ:

ϕ−1 : X → K
x = ξe1 7→ ξ

is continuous. Indeed, we have that

Ker(ϕ−1 ) = {x ∈ X : ϕ−1 (x) = 0} = {ξe1 ∈ X : ξ = 0} = {o},

which is closed in X, since X is Hausdorff. Hence, by Lemma 3.1.3, ϕ−1 is
continuous. This implies that b) holds. In fact, if L is a non-identically zero
functional on X (when L ≡ 0, there is nothing to prove), then there exists
a o 6= x̃ ∈ X s.t. L(x̃) 6= 0. W.l.o.g. we can assume L(x̃) = 1. Now for
any x ∈ X, since dim(X) = 1, we have that x = ξ x̃ for some ξ ∈ K and so
L(x) = ξL(x̃) = ξ. Hence, L ≡ ϕ−1 which we proved to be continuous.
Suppose now that both a) and b) hold for dim(X) ≤ d−1. Let us first show
that b) holds when dim(X) = d. Let L be a non-identically zero functional
on X (when L ≡ 0, there is nothing to prove), then there exists a o 6= x̃ ∈ X
s.t. L(x̃) 6= 0. W.l.o.g. we can assume L(x̃) = 1. Note that for any x ∈ X the
element x − x̃L(x) ∈ Ker(L). Therefore, if we take the canonical mapping
φ : X → X/Ker(L) then φ(x) = φ(x̃L(x)) = L(x)φ(x̃) for any x ∈ X.
This means that X/Ker(L) = span{φ(x̃)} i.e. dim(X/Ker(L)) = 1. Hence,
dim(Ker(L)) = d − 1 and so by inductive assumption Ker(L) is topologically
isomorphic to Kd−1 . This implies that Ker(L) is a complete subspace of X.
Then, by Proposition 2.5.8-a), Ker(L) is closed in X and so by Lemma 3.1.3
we get L is continuous. By induction, we can cocnlude that b) holds for any
dimension d ∈ N.
This immediately implies that a) holds for any dimension d ∈ N. In fact,
we just need to show that for any dimension d ∈ N the mapping

ϕ−1 : X → Kd
Pd
x= j=1 xj ej 7 → (x1 , . . . , xd )
Pd −1
is continuous. Now for any x = j=1 xj ej ∈ X we can write ϕ (x) =
(L1 (x), . . . , Ld (x)), where for any j ∈ {1, . . . , d} we define Lj : X → K by
Lj (x) := xj ej . Since b) holds for any dimension, we know that each Lj is
continuous and so ϕ−1 is continuous.
Step 3: The statement c) holds.
Let f : X → Y be linear and {e1 , . . . , ed } be a basis ofPX. For any j ∈
{1, . . . , d} we define bj := f (ej ) ∈ Y . Hence, for any x = dj=1 xj ej ∈ X we

45
3. Finite dimensional topological vector spaces

have f (x) = f ( dj=1 xj ej ) = dj=1 xj bj . We can rewrite f as composition of

P P

continuous maps i.e. f (x) = A((ϕb1 × . . . × ϕbd )(ϕ−1 (x)) where:

• ϕ−1 is continuous by a)
• each ϕbj is continuous by Lemma 3.1.2
• A is the vector addition on X and so it is continuous since X is a t.v.s..
Hence, f is continuous.

Corollary 3.1.4 (Tychonoff theorem). Let d ∈ N. The only topology that

makes Kd a Hausdorff t.v.s. is the euclidean topology. Equivalently, on a
finite dimensional vector space there is a unique topology that makes it into a
Hausdorff t.v.s..

Proof. We already know that Kd endowed with the euclidean topology τe is

a Hausdorff t.v.s. of dimension d. Let us consider another topology τ on
Kd s.t. (Kd , τ ) is also Hausdorff t.v.s.. Then Theorem 3.1.1-a) ensures that
the identity map between (Kd , τe ) and (Kd , τ ) is a topological isomorphism.
Hence, as observed at the end of Section 1.1.4 p.12, we get that τ ≡ τe .

Corollary 3.1.5. Every finite dimensional Hausdorff t.v.s. is complete.

Proof. Let X be a Hausdorff t.v.s with dim(X) = d < ∞. Then, by Theo-

rem 3.1.1-a), X is topologically isomorphic to Kd endowed with the euclidean
topology. Since the latter is a complete Hausdorff t.v.s., so is X.

Corollary 3.1.6. Every finite dimensional linear subspace of a Hausdorff

t.v.s. is closed.

Proof. Let S be a linear subspace of a Hausdorff t.v.s. (X, τ ) and assume that
dim(S) = d < ∞. Then S endowed with the subspace topology induced by τ
is itself a Hausdorff t.v.s. (see Sheet 5, Exercise 2). Hence, by Corollary 3.1.5
S is complete and therefore closed by Proposition 2.5.8-a).

3.2 Connection between local compactness and finite

dimensionality
By the Heine-Borel property (a subset of Kd is closed and bounded iff it is com-
pact), Kd has a basis of compact neighborhoods of the origin (i.e. the closed
balls centered at the origin in Kd ). Thus the origin, and consequently every
point, of a finite dimensional Hausdorff t.v.s. has a basis of neighborhoods
consisting of compact subsets. This means that a finite dimensional Hausdorff

46
 3.2. Connection between local compactness and finite dimensionality

t.v.s. is always locally compact. Actually also the converse is true and gives
the following beautiful characterization of finite dimensional Hausdorff t.v.s
due to F. Riesz.
Theorem 3.2.1. A Hausdorff t.v.s. is locally compact if and only if it is finite
dimensional.
For convenience let us recall the notions of compactness and local com-
pactness for topological spaces before proving the theorem.
Definition 3.2.2. A topological space X is compact if every open covering
of X contains a finite subcovering. i.e. for any arbitrary collection {Ui }i∈I
of open subsets of X s.t. X ⊆ ∪i∈I Ui there exists a finite subset J of I s.t.
X ⊆ ∪i∈J Ui .
Definition 3.2.3. A topological space X is locally compact if every point of
X has a base of compact neighbourhoods.
Just a small side remark: It is possible to show that every compact Haus-
dorff space is also locally compact but there exist locally compact spaces that
are not compact such as:
• Kd with the euclidean topology
• any infinite set endowed with the discrete topology.
Indeed, any set X with the discrete topology is locally compact, because
for any x ∈ X a basis of neighbourhoods of x in the discrete topology is
given just {x} which is open and also compact. However, if X is infinite
then it is not compact. In fact, if we take the infinite open covering S of
X given by all the singletons of its points, then any finite subcovering
of S will not contain at least one point of X.

Proof. of Theorem 3.2.1

Let X be a locally compact Hausdorff t.v.s., and K a compact neighborhood
of o in X. As K is compact and as 12 K is a neighborhood of the origin (see
Theorem 2.1.10-3), there is a finite family of points x1 , . . . , xr ∈ X s.t.
r
[ 1
K⊆ (xi + K).
2
i=1

Let M := span{x1 , . . . , xr }. Then M is a linear subspace of X and dim(M ) <

∞ is finite, hence M is closed in X by Corollary 3.1.6. Therefore, the quotient
space X/M is Hausdorff t.v.s. by Proposition 2.3.5.
Let φ : X → X/M be the canonical mapping. As K ⊆ M + 21 K, we have
φ(K) ⊆ φ(M ) + φ( 12 K) = 12 φ(K), i.e. 2φ(K) ⊆ φ(K). By iterating we get

47
3. Finite dimensional topological vector spaces

φ(2n K) ⊆ S
φ(K) for any n ∈ N. As K is absorbing (see Theorem 2.1.10-5), we
have X = ∞ n
n=1 2 K. Thus

∞
[
X/M = φ(X) = φ(2n K) ⊆ φ(K).
n=1

Since φ is continuous and the continuous image of a compact set is compact, we

get that φ(K) is compact. Thus X/M is a Hausdorff t.v.s. which is compact.
We claim that X/M must be of zero dimension, i.e. reduced to one point.
This concludes the proof because it implies dim(X) = dim(M ) < ∞.
Let us prove the claim by contradiction. Suppose dim(X/M ) > 0 then
X/M contains a subset of the form Rx̄ for some ō 6= x̄ ∈ X/M . Since such a
subset is closed and X/M is compact, Rx̄ is also compact which is a contra-
diction.

48
 Chapter 4

Locally convex topological vector spaces

4.1 Definition by neighbourhoods

Let us start this section by briefly recalling some basic properties of convex
subsets of a vector space over K (where K is R or C).
Definition 4.1.1. A subset S of a vector space X over K is convex if, when-
ever S contains two points x and y, S also contains the segment of straight
line joining them, i.e.

∀ x, y ∈ S, ∀ α, β ≥ 0 s.t. α + β = 1, αx + βy ∈ S.

Figure 4.1: Convex set Figure 4.2: Not convex set

Examples 4.1.2.
a) The convex subsets of R are simply the intervals of R. Examples of convex
subsets of R2 are solid regular polygons. The Platonic solids are convex
subsets of R3 . Hyperplanes and halfspaces in Rn are convex.
b) Balls in a normed space are convex.
c) Consider a topological space X and the set C(X) of all real valued functions
defined and continuous on X. C(X) with the pointwise addition and scalar

49
4. Locally convex topological vector spaces

multiplication of functions is a vector space. Fixed g ∈ C(X), the subset

S := {f ∈ C(X) : f (x) ≥ g(x), ∀x ∈ X} is convex.
d) Consider the vector space R[x] of all polynomials in one variable with real
coefficients. Fixed n ∈ N and c ∈ R, the subset of all polynomials in R[x]
such that the coefficient of the term of degree n is equal to c is convex.
Proposition 4.1.3.
Let X be a vector space. The following properties hold.
• ∅ and X are convex.
• Arbitrary intersections of convex sets are convex sets.
• Unions of convex sets are generally not convex.
• The sum of two convex sets is convex.
• The image and the preimage of a convex set under a linear map is convex.
Definition 4.1.4. Let S be any subset of a vector space X. We define the
convex hull of S, denoted by conv(S), to be the set of all finite convex linear
combinations of elements of S, i.e.
( n n
)
X X
conv(S) := λi xi : xi ∈ S, λi ∈ [0, 1], λi = 1, n ∈ N .
i=1 i=1

Figure 4.3: The solid line is the border of the convex hull of the shaded set

Proposition 4.1.5.
Let S, T be arbitrary subsets of a vector space X. The following hold.
a) conv(S) is convex
b) S ⊆ conv(S)
c) A set is convex if and only if it is equal to its own convex hull.
d) If S ⊆ T then conv(S) ⊆ conv(T )
e) conv(conv(S)) = conv(S).
f ) conv(S + T ) = conv(S) + conv(T ).
g) The convex hull of S is the smallest convex set containing S, i.e. conv(S)
is the intersection of all convex sets containing S.
h) The convex hull of a balanced set is balanced
Proof. (Sheet 9, Exercise 1)

50
 4.1. Definition by neighbourhoods

Definition 4.1.6. A subset S of a vector space X over K is absolutely convex

if it is convex and balanced.
Let us come back now to topological vector space.
Proposition 4.1.7. The closure and the interior of convex sets in a t.v.s. are
convex sets.

Proof. Let S be a convex subset of a t.v.s. X. For any λ ∈ [0, 1], we define:
ϕλ : X × X → X
.
(x, y) 7→ λx + (1 − λ)y
Note that each ϕλ is continuous by the continuity of addition and scalar
multiplication in the t.v.s. X. Since S is convex, for any λ ∈ [0, 1] we have
that ϕλ (S × S) ⊆ S and so ϕλ (S × S) ⊆ S. The continuity of ϕλ guarantees
that ϕλ (S × S) ⊆ ϕλ (S × S). Hence, we can conclude that ϕλ (S × S) =
ϕλ (S × S) ⊆ S, i.e. S is convex.
To prove the convexity of the interior S̊, we must show that for any two
points x, y ∈ S̊ and for any λ ∈ [0, 1] the point z := ϕλ (x, y) ∈ S̊.
By definition of interior points of S, there exists a neighborhood U of the
origin in X such that x + U ⊆ S and y + U ⊆ S. Then, of course, the claim is
that z + U ⊆ S. This is indeed so, since for any element u ∈ U we can write
z + u in the following form:
z + u = λx + (1 − λ)y + λu + (1 − λ)u = λ(x + u) + (1 − λ)(y + u)
and since both vectors x + u and y + u belong to S, so does z + u. Hence,
z ∈ S̊ which proves S̊ is convex.

Definition 4.1.8. A subset T of a t.v.s. is called a barrel if T has the following

properties:
1. T is absorbing
2. T is absolutely convex
3. T is closed
Proposition 4.1.9. Every neighborhood of the origin in a t.v.s. is contained
in a neighborhood of the origin which is a barrel.

Proof.
Let U be a neighbourhood of the origin and define
!
[
T (U ) := conv λU .
λ∈K,|λ|≤1

51
4. Locally convex topological vector spaces

Clearly, U ⊆ T (U ). Therefore, T (U ) is a neighbourhood of the origin and so

it is absorbing by Theorem 2.1.10. By construction, T (U ) is also closed and
convex as closure of a convex set (see Proposition 4.1.7). To prove that T (U )
is a barrel it remains to show ! that it is balanced. It is easy to see that any
S
point z ∈ conv λ∈K,|λ|≤1 λU can be written as

z = tx + (1 − t)y

with 0 ≤ t ≤ 1, x ∈ λU and y ∈ µU , for some λ, µ ∈ K s.t. |λ| ≤ 1 and

|µ| ≤ 1. Then for any ξ ∈ K with |ξ| ≤ 1 we have:
!
[
ξz = t(ξx) + (1 − t)(ξy) ∈ conv λU
λ∈K,|λ|≤1

!
S
since |ξλ| ≤ 1 and |ξµ| ≤ 1. This proves that conv λ∈K,|λ|≤1 λU is bal-

anced 1 . Hence, by Sheet 3 Exercise 2b, its closure T (U ) is also balanced.

Corollary 4.1.10. Every neighborhood of the origin in a t.v.s. is contained

in a neighborhood of the origin which is absolutely convex.

Note that the converse of Proposition 4.1.9 does not hold in any t.v.s..
Indeed, not every neighborhood of the origin contains another one which is
a barrel. This means that not every t.v.s. has a basis of neighbourhood
consisting of barrels. However, this is true for any locally convex t.v.s.

Definition 4.1.11. A t.v.s. X is said to be locally convex (l.c.) if there is a

basis of neighborhoods of the origin in X consisting of convex sets.

Locally convex spaces are by far the most important class of t.v.s. and
we will present later on several examples of such t.v.s.. For the moment let
us focus on the properties of the filter of neighbourhoods of locally convex
spaces.

Proposition 4.1.12. A locally convex t.v.s. always has a basis of neighbour-

hoods of the origin consisting of open absorbing absolutely convex subsets.
1 S
One could also have directly observed that the set λ∈K,|λ|≤1 λU is balanced and used
Proposition 4.1.5-h to deduce that its convex hull is balanced.

52
 4.1. Definition by neighbourhoods

Proof.
Let N be a neighbourhood of the origin in X. Since X is locally convex, there
exists W convex neighbourhood of the origin in X s.t. W ⊆ N . Moreover, by
Theorem 2.1.10, there exists U balanced neighbourhood
S of the origin in X s.t.
U ⊆ W . The balancedness of U implies that U = λ∈K,|λ|≤1 λU . Then, using
that W is a convex set containing U , we get
!
[
O := conv λU = conv(U ) ⊆ W ⊆ N
λ∈K,|λ|≤1

and so O̊ ⊆ N . Hence, the conclusion holds because O̊ is clearly open and

convex and it is also balanced since o ∈ O̊ and O is balanced (as we have
already showed in the proof of Proposition 4.1.9).

Similarly, we easily get that

Proposition 4.1.13. A locally convex t.v.s. always has a basis of neighbour-
hoods of the origin consisting of barrels.

Proof.
Let N be a neighbourhood of the origin in X. We know that every t.v.s. has a
basis of closed neighbourhoods of the origin (see Sheet 3, Exercise 3a). Then
there exists V closed neighbourhood of the origin in X s.t. V ⊆ N . Since X is
locally convex, then there exists W convex neighbourhood of the origin in X
s.t. W ⊆ V . Moreover, by Theorem 2.1.10, there exists U balanced neighbour-
hood of the origin in X s.t. U ⊆ W . Summing up we have: U ⊆ W ⊆ V ⊆ N
for some U, W, V neighbourhoods of the origin s.t. U S balanced, W convex
and V closed. The balancedness of U implies that U = λ∈K,|λ|≤1 λU . Then,
using that W is a convex set containing U , we get
!
[
conv λU = conv(U ) ⊆ W
λ∈K,|λ|≤1

Passing to the closures and using that V we get

T (U ) = conv(U ) ⊆ W ⊆ V = V ⊆ N.

Hence, the conclusion holds because we have already showed in Proposi-

tion 4.1.9 that T (U ) is a barrel neighbourhood of the origin in X.

53
4. Locally convex topological vector spaces

We can then characterize the class of locally convex t.v.s in terms of ab-
sorbing absolutely convex neighbourhoods of the origin.
Theorem 4.1.14. If X is a l.c. t.v.s. then there exists a basis B of neigh-
bourhoods of the origin consisting of absorbing absolutely convex subsets s.t.
a) ∀U, V ∈ B, ∃W ∈ B s.t. W ⊆ U ∩ V
b) ∀U ∈ B, ∀ ρ > 0, ∃W ∈ B s.t. W ⊆ ρU
Conversely, if B is a collection of absorbing absolutely convex subsets of a vec-
tor space X s.t. a) and b) hold, then there exists a unique topology compatible
with the linear structure of X s.t. B is a basis of neighbourhoods of the origin
in X for this topology (which is necessarily locally convex).

Proof. (Sheet 9, Exercise 2)

In particular, the collection of all multiples ρU of an absorbing absolutely

convex subset U of a vector space X is a basis of neighborhoods of the origin
for a locally convex topology on X compatible with the linear structure (this
ceases to be true, in general, if we relax the conditions on U ).

4.2 Connection to seminorms

In applications it is often useful to define a locally convex space by means of a
system of seminorms. In this section we will investigate the relation between
locally convex t.v.s. and seminorms.
Definition 4.2.1. Let X be a vector space. A function p : X → R is called a
seminorm if it satisfies the following conditions:
1. p is subadditive: ∀x, y ∈ X, p(x + y) ≤ p(x) + p(y).
2. p is positively homogeneous: ∀x, y ∈ X, ∀λ ∈ K, p(λx) = |λ|p(x).
Definition 4.2.2. A seminorm p on a vector space X is a norm if p−1 ({0}) =
{o} (i.e. if p(x) = 0 implies x = o).
Proposition 4.2.3. Let p be a seminorm on a vector space X. Then the
following properties hold:
• p is symmetric.
• p(o) = 0.
• |p(x) − p(y)| ≤ p(x − y), ∀ x, y ∈ X.
• p(x) ≥ 0, ∀ x ∈ X.
• Ker(p) is a linear subspace.

54
 4.2. Connection to seminorms

Proof.
• The symmetry of p directly follows from the positive homogeneity of p.
Indeed, for any x ∈ X we have

p(−x) = p(−1 · x) = | − 1|p(x) = p(x).

• Using again the positive homogeneity of p we get that p(o) = p(0 · x) =

0 · p(x) = 0.
• For any x, y ∈ X, the subadditivity of p guarantees the following in-
equalities:

p(x) = p(x−y+y) ≤ p(x−y)+p(y) and p(y) = p(y−x+x) ≤ p(y−x)+p(x)

which establish the third property.

• The previous property directly gives the nonnegativity of p. In fact, for
any x ∈ X we get

0 ≤ |p(x) − p(o)| ≤ p(x − o) = p(x).

• Let x, y ∈ Ker(p) and α, β ∈ K. Then

p(αx + βy) ≤ |α|p(x) + |β|p(y) = 0

which implies p(αx + βy) = 0, i.e. αx + βy ∈ Ker(p).

Examples 4.2.4.
a) Suppose X = Rn and let M be a vector subspace of X. Set for any x ∈ X

pM (x) := inf kx − yk
y∈M

where k · k is the Euclidean norm, i.e. pM (x) is the distance from the point
x to M in the usual sense. If dim(M ) ≥ 1 then pM is a seminorm and not
a norm (M is exactly the kernel of pM ). When M = {o}, pM (·) = k · k.
b) Let X be a vector space on which is defined a nonnegative sesquilinear
Hermitian form B : X × X → K. Then the function
1
pB (x) := B(x, x) 2

is a seminorm. pB is a norm if and only if B is positive definite (i.e.

B(x, x) > 0, ∀x 6= o).

55
4. Locally convex topological vector spaces

c) Let C(R) be the vector space of all real valued continuous functions on the
real line. For any bounded interval [a, b] with a, b ∈ R and a < b, we define
for any f ∈ C(R):
p[a,b] (f ) := sup |f (t)|.
a≤t≤b

p[a,b] is a seminorm but is never a norm because it might be that f (t) = 0

for all t ∈ [a, b] (and so that p[a,b] (f ) = 0) but f 6≡ 0. Other seminorms are
the following ones:
Z b  p1
q(f ) := |f (0)| and qp (f ) := |f (t)|p dt for 1 ≤ p < ∞.
a

Note that if 0 < p < 1 then qp is not subadditive and so it is not a semi-
norm.

Seminorms on vector spaces are strongly related to a special kind of func-

tionals, i.e. Minkowski functionals. Let us investigate more in details such a
relation. Note that we are still in the realm of vector spaces with no topology!
Definition 4.2.5. Let X be a vector space and A a non-empty subset of X.
We define the Minkowski functional (or gauge) of A to be the mapping:

pA : X → R
x 7→ pA (x) := inf{λ > 0 : x ∈ λA}

(where pA (x) = ∞ if the set {λ > 0 : x ∈ λA} is empty).

It is then natural to ask whether there exists a class of subsets for which
the associated Minkowski functionals are actually seminorms. The answer
is positive for a class of subsets which we have already encountered in the
previous section, namely for absorbing absolutely convex subsets. Actually
we have even more as established in the following lemma.
Notation 4.2.6. Let X be a vector space and p a seminorm on X. The sets

Ůp = {x ∈ X : p(x) < 1} and Up = {x ∈ X : p(x) ≤ 1}.

are said to be, respectively, the closed and the open unit semiball of p.
Lemma 4.2.7. Let X be a vector space. If A is a non-empty subset of X which
is absorbing and absolutely convex, then the associated Minkowski functional
pA is a seminorm and ŮpA ⊆ A ⊆ UpA . Conversely, if q is a seminorm on X
then Ůq is an absorbing absolutely convex set and q = pŮq .

56
 4.2. Connection to seminorms

Proof. Let A be a non-empty subset of X which is absorbing and absolutely

convex and denote by pA the associated Minkowski functional. We want to
show that pA is a seminorm.
• First of all, note that pA (x) < ∞ for all x ∈ X because A is absorbing.
Indeed, by definition of absorbing set, for any x ∈ X there exists ρx > 0
s.t. for all λ ∈ K with |λ| ≤ ρx we have λx ∈ A and so the set {λ >
0 : x ∈ λA} is never empty i.e. pA has only finite nonnegative values.
Moreover, since o ∈ A, we also have that o ∈ λA for any λ ∈ K and so
pA (o) = inf{λ > 0 : o ∈ λA} = 0.
• The balancedness of A implies that pA is positively homogeneous. Since
we have already showed that pA (o) = 0 it remains to prove the positive
homogeneity of pA for non-null scalars. Since A is balanced we have that
for any x ∈ X and for any ξ, λ ∈ K with ξ 6= 0 the following holds:
λ
ξx ∈ λA if and only if x ∈ A. (4.1)
|ξ|
λ
Indeed, A balanced guarantees that ξA = |ξ|A and so x ∈ |ξ| A is equiv-
ξ
alent to ξx ∈ λ |ξ|= λA. Using (4.1), we get that for any x ∈ X and
A
for any ξ ∈ K with ξ 6= 0:

pA (ξx) = inf{λ > 0 : ξx ∈ λA}

 
λ
= inf λ > 0 : x ∈ A
|ξ|
 
λ λ
= inf |ξ| >0:x∈ A
|ξ| |ξ|
= |ξ| inf{µ > 0 : x ∈ µA} = |ξ|pA (x).

• The convexity of A ensures the subadditivity of pA . Take x, y ∈ X. By

definition of Minkowski functional, for every ε > 0 there exists λ, µ > 0
s.t.
λ ≤ pA (x) + ε and x ∈ λA
and
µ ≤ pA (y) + ε and y ∈ µA.
λ µ
Then, by the convexity of A, we obtain that λ+µ A + λ+µ A ⊆ A, i.e.
λA + µA ⊆ (λ + µ)A, and therefore x + y ∈ (λ + µ)A. Hence:

pA (x + y) = inf{δ > 0 : x + y ∈ δA} ≤ λ + µ ≤ pA (x) + pA (y) + 2ε

which proves the subadditivity of pA since ε is arbitrary.

57
4. Locally convex topological vector spaces

We can then conclude that pA is a seminorm. Furthermore, we have the

following inclusions:
ŮpA ⊆ A ⊆ UpA .
In fact, if x ∈ ŮpA then pA (x) < 1 and so there exists 0 ≤ λ < 1 s.t. x ∈ λA.
Since A is balanced, for such λ we have λA ⊆ A and therefore x ∈ A. On the
other hand, if x ∈ A then clearly 1 ∈ {λ > 0 : x ∈ λA} which gives pA (x) ≤ 1
and so x ∈ UpA .
Conversely, let us take any seminorm q on X. Let us first show that Ůq is
absorbing and absolutely convex and then that q coincides with the Minkowski
functional associated to Ůq .
• Ůq is absorbing.
Let x be any point in X. If q(x) = 0 then clearly x ∈ Ůq . If q(x) > 0,
1
we can take 0 < ρ < q(x) and then for any λ ∈ K s.t. |λ| ≤ ρ the
positive homogeneity of q implies that q(λx) = |λ|q(x) ≤ ρq(x) < 1, i.e.
λx ∈ Ůq .
• Ůq is balanced.
For any x ∈ Ůq and for any λ ∈ K with |λ| ≤ 1, again by the positive
homogeneity of q, we get: q(λx) = |λ|q(x) ≤ q(x) < 1 i.e. λx ∈ Ůq .
• Ůq is convex.
For any x, y ∈ Ůq and any t ∈ [0, 1], by both the properties of seminorm,
we have that q(tx + (1 − t)y) ≤ tq(x) + (1 − t)q(y) < t + 1 − t = 1 i.e.
tx + (1 − t)y ∈ Ůq .
Moreover, for any x ∈ X we easily see that
pŮq (x) = inf{λ > 0 : x ∈ λŮq } = inf{λ > 0 : q(x) < λ} = q(x).

We are now ready to see the connection between seminorms and locally
convex t.v.s..
Definition 4.2.8. Let X be a vector space and P := {pi }i∈I a family of
seminorms on X. The coarsest topology τP on X s.t. each pi is continuous is
said to be the topology induced or generated by the family of seminorms P.
Theorem 4.2.9. Let X be a vector space and P := {pi }i∈I a family of semi-
norms. Then the topology induced by the family P is the unique topology
making X into a locally convex t.v.s. and having as a basis of neighbourhoods
of the origin in X the following collection:
n o
B := {x ∈ X : pi1 (x) < ε, . . . , pin (x) < ε} : i1 , . . . , in ∈ I, n ∈ N, ε > 0,  ∈ R .

Viceversa, the topology of an arbitrary locally convex t.v.s. is always induced

by a family of seminorms (often called generating).

58
 4.2. Connection to seminorms

Proof. Let us first show that the collection B is a basis of neighbourhoods of

the origin for the unique topology τ making X into a locally convex t.v.s. by
using Theorem 4.1.14 and then let us prove that τ actually coincides with the
topology induced by the family P.
For any i ∈ I and any ε > 0, consider the set {x ∈ X : pi (x) < ε} = εŮpi .
This is absorbing and absolutely convex, since we have already showed above
that Ůpi fulfills such properties. Therefore, any element of B is an absorbing
absolutely convex subset of X as finite intersection of absorbing absolutely
convex sets. Moreover, both properties a) and b) of Theorem 4.1.14 are clearly
satisfied by B. Hence, Theorem 4.1.14 guarantees that there exists a unique
topology τ on X s.t. (X, τ ) is a locally convex t.v.s. and B is a basis of
neighbourhoods of the origin for τ .
Let us consider (X, τ ). Then for any i ∈ I, the seminorm pi is continuous,
because for any ε > 0 we have pi−1 ([0, ε[) = {x ∈ X : pi (x) < ε} ∈ B which
means that p−1 i ([0, ε[) is a neighbourhood of the origin in (X, τ ). Therefore,
the topology τP induced by the family P is by definition coarser than τ . On
the other hand, each pi is also continuous w.r.t. τP and so B ⊆ τP . But B is
a basis for τ , then necessarily τ is coarser than τP . Hence, τ ≡ τP .
Viceversa, let us assume that (X, τ ) is a locally convex t.v.s.. Then by
Theorem 4.1.14 there exists a basis N of neighbourhoods of the origin in X
consisting of absorbing absolutely convex sets s.t. the properties a) and b)
in Theorem 4.1.14 are fulfilled. W.l.o.g. we can assume that they are open.
Consider now the family S := {pN : N ∈ N }. By Lemma 4.2.7, we know that
each pN is a seminorm and that ŮpN ⊆ N . Let us show that for any N ∈ N
we have actually that N = ŮpN . Since any N ∈ N is open and the scalar
multiplication is continuous we have that for any x ∈ N there exists 0 < t < 1
s.t. x ∈ tN and so pN (x) ≤ t < 1, i.e. x ∈ ŮpN .
We want to show that the topology τS induced by the family S coincides
with original topology τ on X. We know from the first part of the proof how
to construct a basis for a topology induced by a family of seminorms. In fact,
a basis of neighbourhoods of the origin for τS is given by
( n )
\
B := {x ∈ X : pNi (x) < ε} : N1 , . . . , Nn ∈ N , n ∈ N, ε > 0,  ∈ R .
i=1
For any N ∈ N we have showed that N = ŮpN ∈ B so by Hausdorff criterion
τ ⊆ τS . Also for any B ∈ B we have B = ∩ni=1 εŮpNi = ∩ni=1 εNi for some
n ∈ N, N1 , . . . , Nn ∈ N and ε > 0. Then the property b) (of Theorem 4.1.14)
for N implies that for each i = 1, . . . , n there exists Vi ∈ N s.t. Vi ⊆ εNi and so
by the property a) of N we have that there exists V ∈ N s.t. V ⊆ ∩ni=1 Vi ⊆ B.
Hence, by Hausdorff criterion τS ⊆ τ .

59
4. Locally convex topological vector spaces

This result justifies why several authors define a locally convex space to
be a t.v.s whose topology is induced by a family of seminorms (which is now
evidently equivalent to Definition 4.1.11)
In the previous proofs we have used some interesting properties of semiballs
in a vector space. For convenience, we collect them here together with some
further ones which we will repeatedly use in the following.
Proposition 4.2.10. Let X be a vector space and p a seminorm on X. Then:
a) Ůp is absorbing and absolutely convex.
b) ∀ r > 0, rŮp = {x ∈ X : p(x) < r} = Ů 1 p .
r
c) ∀ x ∈ X, x + Ůp = {y ∈ X : p(y − x) < 1}.
d) If q is also a seminorm on X then: p ≤ q if and only if Ůq ⊆ Ůp .
e) If n ∈ N and s1 , . . . , sn are seminorms on X, then their maximum T
s defined
as s(x) := max si (x), ∀x ∈ X is also seminorm on X and Ůs = ni=1 Ůsi .
i=1,...,n
All the previous properties also hold for closed semballs.

Proof.
a) This was already proved as part of Lemma 4.2.7.
b) For any r > 0, we have
1
rŮp = {rx ∈ X : p(x) < 1} = {y ∈ X : p(y) < 1} = {y ∈ X : p(y) < r}.
| r
{z }
Ů 1 p
r
c) For any x ∈ X, we have

x + Ůp = {x + z ∈ X : p(z) < 1} = {y ∈ X : p(y − x) < 1}.

d) Suppose that p ≤ q and take any x ∈ Ůq . Then we have q(x) < 1 and
so p(x) ≤ q(x) < 1, i.e. x ∈ Ůp . Viceversa, suppose that Ůq ⊆ Ůp holds
and take any x ∈ X. We have that either q(x)
> 0 or q(x) = 0.εx In the
εx
first case, for any 0 < ε < 1 we get that q q(x) = ε < 1. Then q(x) ∈ Ůq
εx εx


which implies by our assumption that q(x) ∈ Ůp i.e. p q(x) < 1. Hence,
εp(x) < q(x) and so when ε → 1 we get p(x) ≤ q(x). If instead we are in
the second case that is when q(x) = 0, then we claim that also p(x) = 0.
x x


Indeed, if p(x) > 0 then q p(x) = 0 and so p(x) ∈ Ůq which implies by our
x
assumption that p(x) ∈ Ůp , i.e. p(x) < p(x) which is a contradiction.
e) It is easy to check, using basic properties of the maximum, that the subad-
ditivity and the positive homogeneity of each si imply the same properties
for s. In fact, for any x, y ∈ X and for any λ ∈ K we get:

60
 4.2. Connection to seminorms

• s(x + y) = max si (x + y) ≤ max (si (x) + si (y))

i=1,...,n i=1,...,n
≤ max si (x) + max si (y) = s(x) + s(y)
i=1,...,n i=1,...,n
• s(λx) = max si (λx) = |λ| max si (x) = |λ|s(x).
i=1,...,n i=1,...,n
Moreover, if x ∈ Ůs then max si (x) < 1 and so for all i = 1, . . . , n we
i=1,...,n
have si (x) < 1, i.e. x ∈ i=1 Ůsi . Conversely, if x ∈ ni=1 Ůsi then for all
Tn T
i = 1, . . . , n we have si (x) < 1. Since s(x) is the maximum over a finite
number of terms, it will be equal to sj (x) for some j ∈ {1, . . . , n} and
therefore s(x) = sj (x) < 1, i.e. x ∈ Ůs .

Proposition 4.2.11. Let X be a t.v.s. and p a seminorm on X. Then the

following conditions are equivalent:
a) the open unit semiball Ůp of p is an open set.
b) p is continuous at the origin.
c) the closed unit semiball Up of p is a barrel neighbourhood of the origin.
d) p is continuous at every point.

Proof.
a) ⇒ b) Suppose that Ůp is open in the topology on X. Then for any ε > 0
we have that p−1 ([0, ε[) = {x ∈ X : p(x) < ε} = εŮp is an open neighbourhood
of the origin in X. This is enough to conclude that p : X → R+ is continuous
at the origin.
b) ⇒ c) Suppose that p is continuous at the origin, then Up = p−1 ([0, 1]) is
a closed neighbourhood of the origin. Since Up is also absorbing and absolutely
convex by Proposition 4.2.10-a), Up is a barrel.
c) ⇒ d) Assume that c) holds and fix o 6= x ∈ X. Using Proposition 4.2.10
and Proposition 4.2.3, we get that for any ε > 0: p−1 ([−ε + p(x), p(x) + ε]) =
{y ∈ X : |p(y) − p(x)| ≤ ε} ⊇ {y ∈ X : p(y − x) ≤ ε} = x + εUp , which
is a closed neighbourhood of x since X is a t.v.s. and by the assumption c).
Hence, p is continuous at x.
d) ⇒ a) If p is continuous on X then a) holds because the preimage of an
open set under a continuous function is open and Ůp = p−1 ([0, 1[).

With such properties in our hands we are able to give a criterion to compare
two locally convex topologies using their generating families of seminorms.

61
4. Locally convex topological vector spaces

Theorem 4.2.12 (Comparison of l.c. topologies).

Let P = {pi }i∈I and Q = {qj }j∈J be two families of seminorms on the vector
space X inducing respectively the topologies τP and τQ , which both make X
into a locally convex t.v.s.. Then τP is finer than τQ (i.e. τQ ⊆ τP ) iff
∀q ∈ Q ∃n ∈ N, i1 , . . . , in ∈ I, C > 0 s.t. Cq(x) ≤ max pik (x), ∀x ∈ X.
k=1,...,n
(4.2)
Proof.
Let us first recall that, by Theorem 4.2.9, we have that
n\n o
BP := εŮpik : i1 , . . . , in ∈ I, n ∈ N, ε > 0, ε ∈ R
k=1

and
n
n\ o
BQ := εŮqjk : j1 , . . . , jn ∈ J, n ∈ N, ε > 0, ε ∈ R .
k=1
are respectively bases of neighbourhoods of the origin for τP and τQ .
By using Proposition 4.2.10, the condition (4.2) can be rewritten as
n
\
∀q ∈ Q, ∃n ∈ N, i1 , . . . , in ∈ I, C > 0 s.t. C Ůpik ⊆ Ůq .
k=1

which means that

∀q ∈ Q, ∃ Bq ∈ BP s.t. Bq ⊆ Ůq . (4.3)
since C nk=1 Ůpik ∈ BP .
T

Condition (4.3) means that for any q ∈ Q the set Ůq ∈ τP , which by
Proposition 4.2.11 is equivalent to say that q is continuous w.r.t. τP . By
definition of τQ , this gives that τQ ⊆ τP . 2

2
Alternate proof without using Prop 4.2.11. (Sheet 10, TmExercise 1)
Suppose that (4.2) holds and take any B 0 ∈ BQ , i.e. B 0 = k=1 εŮqjk for some
m ∈ N, j1 , . . . , jn ∈ J and 0 < ε ∈ R. Then, by using m times the condition (4.3),
we obtain that there exist B1 , . . . , Bm ∈ BP such that ∀k ∈ {1, . . . , m}, Bk ⊆ Ůqjk .
Tm Tm
Hence, k=1 Bi ⊆ k=1 Ůqjk . Multiplying by ε both sides of the inclusion, we get
m
B 0 ⊇ ε k=1 Bi ∈ BP and so, by Hausdorff criterion (see Theorem 1.1.16) τQ ⊆ τP .
T

Conversely, suppose that τP is finer than τQ and take any q ∈ Q. Since Ůq ∈ BQ ,
by Hausdorff criterion, we get that there exists B ∈ BP s.t. B ⊆ Ůq . Now such B will
Tn
be of the form B = k=1 εŮpik for some n ∈ N, i1 , . . . , in ∈ I and 0 < ε ∈ R. Then,
Proposition 4.2.10 gives that B = εŮ max pik ⊆ Ůq , i.e. Ů max pik ⊆ Ůεq which is
k=1,...,n k=1,...,n

equivalent to εq(x) ≤ max pik (x), ∀x ∈ X.

k=1,...,n

62
 4.2. Connection to seminorms

This theorem allows us to easily see that the topology induced by a family
of seminorms on a vector space does not change if we close the family under
taking the maximum of finitely many of its elements. Indeed, the following
result holds.

 Let P := {pi }i∈I be a family of seminorms on a vector

Proposition 4.2.13.
space X and Q := max pi : ∅ =
6 B ⊆ I with B finite . Then Q is a family
i∈B
of seminorms and τP = τQ , where τP and τQ denote the topology induced on
X by P and Q, respectively.

Proof.
First of all let us note that, by Proposition 4.2.10, Q is a family of seminorms.
On the one hand, since P ⊆ Q, by definition of induced topology we have
τP ⊆ τQ . On the other hand, for any q ∈ Q we have q = max pi for some
i∈B
∅ 6= B ⊆ I finite. Then (4.2) is fulfilled for n = |B| (where |B| denotes the
cardinality of the finite set B), i1 , . . . , in being the n elements of B and for
any 0 < C ≤ 1. Hence, by Theorem 4.2.12, τQ ⊆ τP .

This fact can be used to show the following very useful property of locally
convex t.v.s.
Proposition 4.2.14. The topology of a locally convex t.v.s. can be always
induced by a directed family of seminorms.
Definition 4.2.15. A family Q := {qj }j∈J of seminorms on a vector space
X is said to be directed if

∀ j1 , j2 ∈ J, ∃ j ∈ J, C > 0 s.t. Cqj (x) ≥ max{qj1 (x), qj2 (x)}, ∀ x ∈ X (4.4)

or equivalently by induction if

∀ n ∈ N, j1 , . . . , jn ∈ J, ∃ j ∈ J, C > 0 s.t. Cqj (x) ≥ max qjk (x), ∀ x ∈ X.

k=1,...,n

Proof. of Proposition 4.2.14

Let (X, τ ) be a locally convex t.v.s.. By Theorem 4.2.9, we have that there
exists a family of seminorms P := {pi }i∈I on X s.t. τ = τP . Let us define Q
as the collectionobtained by forming the maximum of finitely many elements
of P, i.e. Q := max pi : ∅ =6 B ⊆ I with B finite . By Proposition 4.2.13,
i∈B
Q is a family of seminorms and we have that τP = τQ . We claim that Q is
directed.

63
4. Locally convex topological vector spaces

Let q, q 0 ∈ Q, i.e. q := max pi and q 0 := max0 pi for some non-empty finite

i∈B i∈B
subsets B, B 0 of I. Let us define q 00 := max 0 pi . Then q 00 ∈ Q and for any
i∈B∪B
C ≥ 1 we have that (4.4) is satisfied, because we get that for any x ∈ X
 
Cq (x) = C max max pi (x), max0 pi (x) ≥ max{q(x), q 0 (x)}.
00
i∈B i∈B

Hence, Q is directed.

It is possible to show (Sheet 10, Exercise 2) that a basis of neighbourhoods

of the origin for the l.c. topology τQ induced by a directed family of seminorms
Q is given by:
Bd := {rŮq : q ∈ Q, r > 0}. (4.5)

4.3 Hausdorff locally convex t.v.s

In Section 2.2, we gave some characterization of Hausdorff t.v.s. which can
of course be applied to establish whether a locally convex t.v.s. is Hausdorff
or not. However, in this section we aim to provide necessary and sufficient
conditions bearing only on the family of seminorms generating a locally convex
topology for being a Hausdorff topology.

Definition 4.3.1.
A family of seminorms P := {pi }i∈I on a vector space X is said to be sepa-
rating if
∀x ∈ X \ {o}, ∃ i ∈ I s.t. pi (x) 6= 0. (4.6)

Note that the separation condition (4.6) is equivalent to

pi (x) = 0, ∀i ∈ I ⇒ x = o

which by using Proposition 4.2.10 can be rewritten as

\
cŮpi = {o},
i∈I,c>0

since pi (x) = 0 is equivalent to say that pi (x) < c, for all c > 0.

64
 4.3. Hausdorff locally convex t.v.s

Lemma 4.3.2. Let τP be the topology induced by a separating family of semi-

norms P := (pi )i∈I on a vector space X. Then τP is a Hausdorff topology.

Proof. Let x, y ∈ X be s.t. x 6= y. Since P is separating, ∃ i ∈ I s.t.

pi (x − y) 6= 0. Then ∃  > 0 s.t. pi (x − y) = 2. Let us define Vx := {u ∈ X |
pi (x − u) < } and Vy := {u ∈ X | pi (y − u) < }. By Proposition 4.2.10, we
get that Vx = x + εŮpi and Vy = y + εŮpi . Since Theorem 4.2.9 guarantees
that (X, τP ) is a t.v.s. where the set εŮpi is a neighbourhood of the origin, Vx
and Vy are neighbourhoods of x and y, respectively. They are clearly disjoint.
Indeed, if there would exist u ∈ Vx ∩ Vy then

pi (x − y) = pi (x − u + u − y) ≤ pi (x − u) + pi (u − y) < 2ε

which is a contradiction.

Proposition 4.3.3. A locally convex t.v.s. is Hausdorff if and only if its

topology can be induced by a separating family of seminorms.

Proof. Let (X, τ ) be a locally convex t.v.s.. Then we know that there always
exists a basis N of neighbourhoods of the origin in X consisting of open
absorbing absolutely convex sets. Moreover, in Theorem 4.2.9, we have showed
that τ = τP where P is the family of seminorms given by the Minkowski
functionals of sets in N , i.e. P := {pN : N ∈ N }, and also that for each
N ∈ N we have N = ŮpN .
Suppose that (X, τ ) is also Hausdorff. Then Proposition 2.2.3 ensures that
for any x ∈ X with x 6= o there exists a neighbourhood V of the origin in X
s.t. x ∈ / N 3 . Hence,
/ V . This implies that there exists at least N ∈ N s.t. x ∈
x∈ / N = ŮpN means that pN (x) ≥ 1 and so pN (x) 6= 0, i.e. P is separating.
Conversely, if τ is induced by a separating family of seminorms P, i.e.
τ = τP , then Lemma 4.3.2 ensures that X is Hausdorff.

Examples 4.3.4.

1. Every normed space is a Hausdorff locally convex space, since every norm
is a seminorm satisfying the separation property. Therefore, every Ba-
nach space is a complete Hausdorff locally convex space.
3
Since N is a basis of neighbourhoods of the origin, ∃ N ∈ N s.t. N ⊆ V . If x would
belong to all elements of the basis then in particular it would be x ∈ N and so also x ∈ V ,
contradiction.

65
4. Locally convex topological vector spaces

2. Every family of seminorms on a vector space containing a norm induces

a Hausdorff locally convex topology.
3. Given an open subset Ω of Rd with the euclidean topology, the space C(Ω)
of real valued continuous functions on Ω with the so-called topology of
uniform convergence on compact sets is a locally convex t.v.s.. This
topology is defined by the family P of all the seminorms on C(Ω) given
by
pK (f ) := max |f (x)|, ∀K ⊂ Ω compact .
x∈K
Moreover, (C(Ω), τP ) is Hausdorff, because the family P is clearly sepa-
rating. In fact, if pK (f ) = 0, ∀K compact subsets of Ω then in particular
p{x} (f ) = |f (x)| = 0 ∀x ∈ Ω, which implies f ≡ 0 on Ω.
More generally, for any X locally compact we have that C(X) with the
topology of uniform convergence on compact subsets of X is a locally
convex Hausdorff t.v.s.
To introduce two other examples of l.c. Hausdorff t.v.s. we need to recall
some standard general notations. Let N0 be the set of all non-negative integers.
For any x = (x1 , . . . , xd ) ∈ Rd and α = (α1 , . . . , αd ) ∈ Nd0 one defines xα :=
xα1 1 · · · xαd d . For any β ∈ Nd0 , the symbol Dβ denotes the partial derivative of
order |β| where |β| := di=1 βi , i.e.
P

∂ |β| ∂ β1 ∂ βd
Dβ := = ··· .
∂xβ1 1 · · · ∂xβd d ∂xβ1 1 ∂xβd d
Examples 4.3.5.
1. Let Ω ⊆ Rd open in the euclidean topology. For any k ∈ N0 , let C k (Ω) be
the set of all real valued k−times continuously differentiable functions
on Ω, i.e. all the derivatives of f of order ≤ k exist (at every point of Ω)
and are continuous functions in Ω. Clearly, when k = 0 we get the set
C(Ω) of all real valued continuous functions on Ω and when k = ∞ we
get the so-called set of all infinitely differentiable functions or smooth
functions on Ω. For any k ∈ N0 , C k (Ω) (with pointwise addition and
scalar multiplication) is a vector space over R. The topology given by
the following family of seminorms on C k (Ω):

pm,K (f ) := sup sup (Dβ f )(x) , ∀K ⊆ Ω compact, ∀m ∈ {0, 1, . . . , k},

β∈Nd x∈K
0
|β|≤m

makes C k (Ω) into a locally convex Hausdorff t.v.s.. (Note that when
k = ∞ we have m ∈ N0 .)

66
 4.4. The finest locally convex topology

2. The Schwartz space or space of rapidly decreasing functions on Rd is

defined as the set S(Rd ) of all real-valued functions which are defined
and infinitely differentiable on Rd and which have the additional property
(regulating their growth at infinity) that all their derivatives tend to zero
at infinity faster than any inverse power of x, i.e.
( )
S(Rd ) := f ∈ C ∞ (Rd ) : sup xα Dβ f (x) < ∞, ∀α, β ∈ Nd0 .
x∈Rd

(For example, any smooth function f with compact support in Rd is

in S(Rd ), since any derivative of f is continuous and supported on a
compact subset of Rd , so xα (Dβ f (x)) has a maximum in Rd by the
extreme value theorem.)
The Schwartz space S(Rd ) is a vector space over R and the topology
given by the family Q of seminorms on S(Rd ):

qα,β (f ) := sup xα Dβ f (x) , ∀α, β ∈ Nd0

x∈Rd

makes S(Rd ) into a locally convex Hausdorff t.v.s.. Indeed, the family is
clearly separating, because if qα,β (f ) = 0, ∀α, β ∈ Nd0 then in particular
qo,o (f ) = supx∈Rd |f (x)| = 0 ∀x ∈ Rd , which implies f ≡ 0 on Rd .
Note that S(Rd ) is a linear subspace of C ∞ (Rd ), but its topology τQ
on S(Rd ) is finer than the subspace topology induced on it by C ∞ (Rd ).
(Sheet 10, Exercise 1)

4.4 The finest locally convex topology

In the previous sections we have seen how to generate topologies on a vector
space which makes it into a locally convex t.v.s.. Among all of them, there is
the finest one (i.e. the one having the largest number of open sets) which is
usually called the finest locally convex topology on the given vector space.
Proposition 4.4.1. The finest locally convex topology on a vector space X is
the topology induced by the family of all seminorms on X and it is a Hausdorff
topology.
Proof.
Let us denote by S the family of all seminorms on the vector space X. By
Theorem 4.2.9, we know that the topology τS induced by S makes X into a
locally convex t.v.s. We claim that τS is the finest locally convex topology. In

67
4. Locally convex topological vector spaces

fact, if there was a finer locally convex topology τ (i.e. if τS ⊆ τ with (X, τ )
locally convex t.v.s.) then Theorem 4.2.9 would give that τ is also induced by
a family P of seminorms. But surely P ⊆ S and so τ = τP ⊆ τS by definition
of induced topology. Hence, τ = τS .
It remains to show that (X, τS ) is Hausdorff. By Lemma 4.3.2, it is enough
to prove that S is separating. Let x ∈ X \ {o} and let B be an algebraic basis
of the vector space X containing x. Define the linear functional L : X → R as
L(x) = 1 and L(y) = 0 for all y ∈ B \ {x}. Then it is easy to see that s := |L|
is a seminorm, so s ∈ S and s(x) 6= 0, which proves that S is separating.

An alternative way of describing the finest locally convex topology on a

vector space X without using the seminorms is the following:
Proposition 4.4.2. The collection of all absorbing absolutely convex sets of
a vector space X is a basis of neighbourhoods of the origin for the finest locally
convex topology on X.
Proof. Let τmax be the finest locally convex topology on X and A the collection
of all absorbing absolutely convex sets of X. By Theorem 4.1.14, we know
that every locally convex t.v.s. has a basis of neighbourhood of the origin
consisting of absorbing absolutely convex subsets of X. Then clearly the
basis of neighbourhoods of the origin Bmax of τmax is contained in A. Hence,
τmax ⊆ τ where τ denote the topology generated by A. On the other hand,
A fulfills all the properties required in Theorem 4.1.14 and so τ also makes
X into a locally convex t.v.s.. Hence, by definition of finest locally convex
topology, τ ⊆ τmax .

This result can be clearly proved also using the Proposition 4.4.1 and
the correspondence between Minkowski functionals and absorbing absolutely
convex subsets of X introduced in the Section 4.2.
Proposition 4.4.3. Every linear functional on a vector space X is continuous
w.r.t. the finest locally convex topology on X.
Proof. Let L : X → K be a linear functional on a vector space X. For any
ε > 0, we denote by Bε (0) the open ball in K of radius ε and center 0 ∈ K,
i.e. Bε (0) := {k ∈ K : |k| < ε}. Then we have that L−1 (Bε (0)) = {x ∈ X :
|L(x)| < ε}. It is easy to verify that the latter is an absorbing absolutely
convex subset of X and so, by Proposition 4.4.2, it is a neighbourhood of the
origin in the finest locally convex topology on X. Hence L is continuous at
the origin and so, by Proposition 2.1.15-3), L is continuous everywhere in X.

68
 4.5. Direct limit topology on a countable dimensional t.v.s.

4.5 Direct limit topology on a countable dimensional t.v.s.

In this section we are going to give an important example of finest locally
convex topology on an infinite dimensional vector space, namely the direct
limit topology on any countable dimensional vector space. For simplicity, we
are going to focus on R−vector spaces.
Definition 4.5.1. Let X be an infinite dimensional vector space whose di-
mension is countable. The direct limit topology (or finite topology) τf on X
is defined as follows:
U ⊆ X is open in τf iff U ∩ W is open in the euclidean topology on W ,
∀ W ⊂ X with dim(W ) < ∞.
Equivalently, if we fix a Hamel basis S {xn }n∈N of X and if for any n ∈ N we
set Xn := span{x1 , . . . , xn } s.t. X = ∞
i=1 Xi and X1 ⊆ . . . ⊆ Xn ⊆ . . ., then
U ⊆ X is open in τf iff U ∩ Xi is open in the euclidean topology on Vi for
every i ∈ N.
Theorem 4.5.2. Let X be an infinite dimensional vector space whose dimen-
sion is countable endowed with the finite topology τf . Then:
a) (X, τf ) is a Hausdorff locally convex t.v.s.
b) τf is the finest locally convex topology on X

Proof.
a) We leave to the reader the proof of the fact that τf is compatible with the
linear structure of X (Sheet 10, Exercise 3) and we focus instead on proving
that τf is a locally convex topology. To this aim we are going to show that
for any open neighbourhood U of the origin in (X, τf ) there exists an open
convex neighbourhood U 0 ⊆ U .
Let {xi }i∈N be an R-basis for X and set Xn := span{x1 , . . . , xn } for any
n ∈ N. We proceed (by induction on n ∈ N) to construct an increasing
sequence Cn ⊆ U ∩ Xn of convex subsets as follows:
• For n = 1: Since U ∩ X1 is openin X1 = Rx1 , we have that there exists
a1 ∈ R, a1 > 0 such that C1 := λ1 x1 | − a1 ≤ λ1 ≤ a1 ⊆ U ∩ X1 .
• Inductive assumption on n: We assume we have found a1 , . . . , an ∈ R+

such that Cn := λ1 x1 + . . . + λn xn | − ai ≤ λi ≤ ai ; i ∈ {1, . . . , n} ⊆
U ∩ Xn . Note that Cn is closed (in Xn , as well as) in Xn+1 .
• For n + 1: We claim ∃ an+1 > 0, an+1 ∈ R such that
Cn+1 := λ1 x1 +. . .+λn xn +λn+1 xn+1 |−ai ≤ λi ≤ ai ; i ∈ {1, . . . , n+
1} ⊆ U ∩ Xn+1 .
Proof of claim: If the claim does not hold, then ∀ N ∈ N ∃ xN ∈ Xn+1
s.t.
xN = λN N N
1 x1 + . . . λn xn + λn+1 xn+1

69
4. Locally convex topological vector spaces

1 1
with −ai ≤ λN
i ≤ ai for i ∈ {1, . . . , n}, − ≤ λNn+1 ≤ and xN ∈ / U.
N N
But xN has form xN = λN x + . . . + λN N N
n xn +λn+1 xn+1 , so {x }N ∈N
| 1 1 {z }
∈ Cn
is a bounded sequence in Xn+1 \U . Therefore, we can find a subsequence
{xNj }j∈N which is convergent as j → ∞ to x ∈ Cn ⊆ U (since Cn is
closed in Xn+1 and the n+1−th component of xNj tends to 0 as j → ∞).
Hence, the sequence {xNj }j∈N ⊆ Xn+1 \ U converges to x ∈ U but this
contradicts the fact that Xn+1 \ U is closed in Xn+1 . This establishes
the claim.
Now for any n ∈ N consider

Dn := λ1 x1 + . . . + λn xn | − ai < λi < ai ; i ∈ {1, . . . , n} ,

then Dn ⊂ Cn ⊆ U ∩ Xn is open and convex in Xn . Then U 0 := ∪n∈N Dn is

an open and convex neighbourhood of the origin in (X, τf ) and U 0 ⊆ U .

b) Let us finally show that τf is actually the finest locally convex topology
τmax on X. Since we have already showed that τf is a l.c. topology on X,
clearly we have τf ⊆ τmax by definition of finest l.c. topology on X.
Conversely, let us consider U ⊆ X open in τmax . We want to show that
U is open in τf , i.e. W ∩ U is open in the euclidean topology on W for
any finite dimensional subspace W of X. Now each W inherits τmax from
X. Let us denote by τmax W the subspace topology induced by (X, τmax ) on
W . By definition of subspace topology, we have that W ∩ U is open in τmax W .

Moreover, by Proposition 4.4.1, we know that (X, τmax ) is a Hausdorff t.v.s.

W ) is a finite dimensional Hausdorff t.v.s. (see by Proposition
and so (W, τmax
W
2.1.15-1). Therefore, τmax has to coincide with the euclidean topology by
Theorem 3.1.1 and, consequently, W ∩ U is open w.r.t. the euclidean topology
on W .

We actually already know a concrete example of countable dimensional

space with the finite topology:
Example 4.5.3. Let n ∈ N and x = (x1 , . . . , xn ). Denote by R[x] the space
of polynomials in the n variables x1 , . . . , xn with real coefficients and by

Rd [x] := {f ∈ R[x]| deg f ≤ d}, d ∈ N0 ,

then R[x] := ∞
S
d=0 Rd [x]. The finite topology τf on R[x] is then given by:
U ⊆ R[x] is open in τf iff ∀d ∈ N0 , U ∩ Rd [x] is open in Rd [x] with the
euclidean topology.

70
 4.6. Continuity of linear mappings on locally convex spaces

4.6 Continuity of linear mappings on locally convex spaces

Since locally convex spaces are a particular class of topological vector spaces,
the natural functions to be considered on this spaces are continuous linear
maps. In this section, we present a necessary and sufficient condition for
the continuity of a linear map between two l.c. spaces, bearing only on the
seminorms inducing the two topologies.
For simplicity, let us start with linear functionals on a l.c. space. Recall
that for us K = R or K = C endowed with the euclidean topology given by
the absolute value | · |. In this section, for any ε > 0 we denote by Bε (0) the
open ball in K of radius ε and center 0 ∈ K i.e. Bε (0) := {k ∈ K : |k| < ε}.
Proposition 4.6.1. Let τ be a locally convex topology on a vector space X
generated by a directed family Q of seminorms on X. Then L : X → K is a τ -
continuous linear functional iff there exists q ∈ Q such that L is q-continuous,
i.e.
∃q ∈ Q, ∃C > 0 s.t. |L(x)| ≤ Cq(x), ∀ x ∈ X. (4.7)

Proof.
Let us first observe that since X and K are both t.v.s. by Proposition 2.1.15-3)
the continuity of L is equivalent to its continuity at the origin. Therefore, it
is enough to prove the criterion for the continuity of L at the origin.
(⇒) Suppose that L is τ -continuous at the origin in X. Then we have that
L−1 (B1 (0)) = {x ∈ X : |L(x)| < 1} is an open neighbourhood of the origin in
(X, τ ). Since the family Q inducing τ is directed, a basis of neighbourhood
of the origin in (X, τ ) is given by Bd as in (4.5). Therefore, ∃B ∈ Bd s.t.
B ⊆ L−1 (B1 (0)), i.e.
∃ q ∈ Q, ∃ r > 0 s.t. rŮq ⊆ L−1 (B1 (0)). (4.8)
Then for any ε > 0 we get rεŮq ⊆ εL−1 (B1 (0)) = L−1 (B ε (0)). This proves
that L is q−continuous at the origin, because rεŮq is clearly an open neigh-
bourhood of the origin in X w.r.t. the topology generated by the single semi-
norm q.
(⇐) Suppose that there exists q ∈ Q s.t. L is q-continuous in X. Then,
since τ is the topology induced by the whole family Q which is finer than
the one induced by the single seminorm q, we clearly have that L is also
τ −continuous.
Note that by simply observing that |L| is a seminorm and by using Propo-
sition 4.2.10 we get that (4.7) is equivalent to (4.8) and so to the q-continuity
of L at the origin.

71
4. Locally convex topological vector spaces

By using this result together with Proposition 4.2.14 we get the following.
Corollary 4.6.2. Let τ be a locally convex topology on a vector space X
generated by a family P := {pi }i∈I of seminorms on X. Then L : X → K is
a τ -continuous linear

functional iff there exist n ∈ N, i1 , . . . , in ∈ I such that
L is max pik -continuous, i.e.
k=1,...,n

∃ n ∈ N, ∃ i1 , . . . , in ∈ I, ∃ C > 0 s.t. |L(x)| ≤ C max pik (x), ∀ x ∈ X.

k=1,...,n

The proof of Proposition 4.6.1 can be easily modified to get the following
more general criterion for the continuity of any linear map between two locally
convex spaces.
Theorem 4.6.3. Let X and Y be two locally convex t.v.s. whose topologies
are respectively generated by the families P and Q of seminorms on X. Then
f : X → Y linear is continuous iff

∀ q ∈ Q, ∃ n ∈ N, ∃ p1 , . . . , pn ∈ P, ∃ C > 0 : q(f (x)) ≤ C max pi (x), ∀ x ∈ X.

i=1,...,n

Proof. (Exercise)

72
 Chapter 5

The Hahn-Banach Theorem

and its applications

5.1 The Hahn-Banach Theorem

One of the most important results in the theory of t.v.s. is the Hahn-Banach
theorem (HBT). It is named for Hans Hahn and Stefan Banach who proved
this theorem independently in the late 1920s, dealing with the problem of
extending continuous linear functionals defined on a subspace of a seminormed
vector space to the whole space. We will see that actually this extension
problem can be reduced to the problem of separating by a closed hyperplane
a convex open set and an affine submanifold (the image by a translation of a
linear subspace) which do not intersect. Indeed, there are two main versions
of HBT showing respectively the analytic and the geometric side of this result.
Before stating these two versions of HBT, let us recall the notion of hy-
perplane in a vector space (we always consider vector spaces over the field
K which is either R or C). A hyperplane H in a vector space X over K is
a maximal proper linear subspace of X or, equivalently, a linear subspace of
codimension one, i.e. dimX/H = 1. Another equivalent formulation is that a
hyperplane is a set of the form ϕ−1 ({0}) for some linear functional ϕ : X → K
not identically zero. The translation by a non-null vector of a hyperplane will
be called affine hyperplane.
Theorem 5.1.1 (Analytic form of Hahn-Banach theorem ).
Let p be a seminorm on a vector space X over K, M a linear subspace of X,
and f a linear functional on M such that
|f (x)| ≤ p(x), ∀ x ∈ M. (5.1)
There exists a linear functional f˜ on X extending f , i.e. f˜(x) = f (x) ∀ x ∈ M ,
and such that
|f˜(x)| ≤ p(x), ∀ x ∈ X. (5.2)

73
5. The Hahn-Banach Theorem and its applications

Theorem 5.1.2 (Geometric form of Hahn-Banach theorem ).

Let X be a topological vector space over K, N a linear subspace of X, and Ω
a convex open subset of X such that N ∩ Ω = ∅. Then there exists a closed
hyperplane H of X such that

N ⊆H and H ∩ Ω = ∅. (5.3)

It should be remarked that the vector space X does not apparently carry
any topology in Theorem 5.1.1, but actually the datum of a seminorm on X
is equivalent to the datum of the topology induced by this seminorm. It is
then clear that the conditions (5.1) and (5.2) imply the p−continuity of the
functions f and f˜, respectively.
Let us also stress the fact that in Theorem 5.1.2 neither local convexity nor
the Hausdorff separation property are assumed on the t.v.s. X. Moreover, it is
easy to see that the geometric form of HBT could have been stated also in an
affine setting, namely starting with any affine submanifold N of X which does
not intersect the open convex subset Ω and getting a closed affine hyperplane
fulfilling (5.3).
We first prove Theorem 5.1.2 and then show how to derive from this the
analytic form Theorem 5.1.1.
Before starting the proof, let us fix one more definition. A cone C in a
vector space X over R is a subset of X which is closed under addition and
multiplication by positive scalars.

Proof. Theorem 5.1.2

We assume that Ω 6= ∅, otherwise there is nothing to prove.
1) Existence of a linear subspace H of X maximal for (5.3).
This first part of the proof is quite simple and consists in a straightforward
application of Zorns lemma. In fact, consider the family F of all the linear
subspaces L of X such that

N ⊆ L and L ∩ Ω = ∅. (5.4)

F is clearly non-empty since N belongs to it by assumption. If we take now a

totally ordered subfamily C of F (totally ordered for the inclusion relation ⊆),
then the union of all the linear subspaces belonging to C is a linear subspace of
X having the properties in (5.4). Hence, we can apply Zorn’s lemma applies
and conclude that there exists at least a maximal element H in F.
2) H is closed in X.
The fact that H and Ω do not intersect gives that H is contained in the

74
 5.1. The Hahn-Banach Theorem

complement of Ω in X. This implies that also its closure H does not intersect
Ω. Indeed, since Ω is open, we get
H ⊆ X \ Ω = X \ Ω.
Then H is a linear subspace (as closure of a linear subspace) of X, which is
disjoint from Ω and which contains H and so N , i.e. H ∈ F. Hence, as H is
maximal in F, it must coincide with its closure. Note that the fact that H
is closed guarantees that the quotient space X/H is a Hausdorff t.v.s. (see
Proposition 2.3.5).
3) H is an hyperplane
We want to show that H is a hyperplane, i.e. that dim(X/H) = 1. To this
aim we distinguish the two cases when K = R and when K = C.
3.1) Case K = R
Let φ : X → X/H be the canonical map. Since φ is an open linear mapping
(see Proposition 2.3.2), φ(Ω) is an open convex subset of X/H. Also we have
that φ(Ω) does not contain the origin ô of X/H. Indeed, if ô ∈ φ(Ω) holds,
then there would exist x ∈ Ω s.t. φ(x) = ô and so x ∈ H, which would
contradict the assumption H ∩ Ω = ∅. Let us set:
[
A= λφ(Ω).
λ>0

Then the subset A of X/H is open, convex and it is a cone which does not
contain the origin ô.
Let us observe that X/H has at least dimension 1. If dim(X/H) = 0 then
X/H = {ô} and so X = H which contradicts the fact that Ω does not intersect
H (recall that we assumed Ω is non-empty). Suppose that dim(X/H) ≥ 2,
then to get our conclusion it will suffice to show the following claims:
Claim 1: The boundary ∂A of A must contain at least one point x 6= ô.
Claim 2: The point −x cannot belong to A.
In fact, once Claim 1 is established, we have that x ∈/ A, because x ∈ ∂A and
A is open. This together with Claim 2 gives that both x and −x belong to the
complement of A in X/H and, therefore, so does the straight line L defined
by these two points. (If there was a point y ∈ L ∩ A then any positive multiple
of y would belong to L ∩ A, as A is a cone. Hence, for some λ > 0 we would
have x = λy ∈ L ∩ A, which contradicts the fact that x ∈ / A.) Then:
−1
• φ (L) is a linear subspace of X
• φ−1 (L) ∩ Ω = ∅, since L ∩ A = ∅
• φ−1 (L) ) H because ô = φ(H) ⊆ L but L 6= {ô} since x 6= ô is in L.
This contradicts the maximality of H and so dim(X/H) = 1.
To complete the proof of 3.1) let us show the two claims.

75
5. The Hahn-Banach Theorem and its applications

Proof. of Claim 1
Suppose that ∂A = {ô}. This means that A has empty boundary in the set
(X/H) \ {ô} and so that A is a connected component of (X/H) \ {ô}. Since
dim(X/H) ≥ 2, the space (X/H) \ {ô} is arc-connected and so it is itself a
connected space. Hence, A = (X/H) \ {ô} which contradicts the convexity of
A since (X/H) \ {ô} is clearly not convex.

Proof. of Claim 2
Suppose −x ∈ A. Then, as A is open, there is a neighborhood V of −x
entirely contained in A. This implies that −V is a neighborhood of x. Since
x is a boundary point of A, there exists y ∈ (−V ) ∩ A. But then −y ∈ V ⊂ A
and so, by the convexity of A, the whole line segment between y and −y is
contained in A, in particular ô, which contradicts the definition of A.

3.2) Case K = C
Although here we are consider the scalars to be the complex numbers, we
may view X as a vector space over the real numbers and it is obvious that its
topology, as originally given, is still compatible with its linear structure. By
step 3.1) above, we know that there exists a real hyperplane H0 of X which
contains N and does not intersect Ω. By a real hyperplane, we mean that
H0 is a linear subspace of X viewed as a vector space over the field of real
numbers such that dimR (X/H) = 1. √
Now it is easy to see that iN = N (here i = −1). Hence, setting
H := H0 ∩ iH0 , we have that N ⊆ H and H ∩ Ω = ∅. Then to complete the
proof it remains to show that this H is a complex hyperplane. It is obviously
a complex linear subspace of X and its real codimension is ≥ 1 and ≤ 2 (since
the intersection of two distinct hyperplanes is always a linear subspace with
codimension two). Hence, its complex codimension is equal to one.

Proof. Theorem 5.1.1

Let p be a seminorm on the vector space X, M a linear subspace of X, and f
a linear functional defined on M fulfilling (5.1). As already remarked before,
this means that f is continuous on M w.r.t. the topology induced by p on X
(which makes X a l.c. t.v.s.).
Consider the subset N := {x ∈ M : f (x) = 1}. Taking any vector
x0 ∈ N , it is easy to see that N − x0 = Ker(f ) (i.e. the kernel of f in
M ), which is a hyperplane of M and so a linear subspace of X. Therefore,
setting M0 := N − x0 , we have the following decomposition of M :
M = M0 ⊕ Kx0 ,

76
 5.2. Applications of Hahn-Banach theorem

where Kx0 is the one-dimensional linear subspace through x0 . In other words

∀ x ∈ M, ∃! λ ∈ K, y ∈ M0 : x = y + λx0 .

Then
∀ x ∈ M, f (x) = f (y) + λf (x0 ) = λf (x0 ) = λ,
which means that the values of f on M are completely determined by the ones
on N . Consider now the open unit semiball of p:

U := Ůp = {x ∈ X : p(x) < 1},

which we know being an open convex subset of X endowed with the topology
induced by p. Then N ∩ U = ∅ because if there was x ∈ N ∩ U then p(x) < 1
and f (x) = 1, which contradict (5.1).
By Theorem 5.1.2 (affine version), there exists a closed affine hyperplane
H of X with the property that N ⊆ H and H ∩ U = ∅. Then H − x0
is a hyperplane and so the kernel of a continuous linear functional f˜ on X
non-identically zero.
Arguing as before (consider here the decomposition X = (H − x0 ) ⊕ Kx0 ),
we can deduce that the values of f˜ on X are completely determined by the
ones on N and so on H (because for any h ∈ H we have h − x0 ∈ Ker(f˜) and
so f˜(h) − f˜(x0 ) = f˜(h − x0 ) = 0). Since f˜ 6≡ 0, we have that f˜(x0 ) 6= 0 and
w.l.o.g. we can assume f˜(x0 ) = 1 i.e. f˜ ≡ 1 on H. Therefore, for any x ∈ M
there exist unique λ ∈ K and y ∈ N − x0 ⊆ H − x0 s.t. x = y + λx0 , we get
that:
f˜(x) = λf˜(x0 ) = λ = λf (x0 ) = f (x),
i.e. f is the restriction of f˜ to M . Furthermore, the fact that H ∩U = ∅ means
that f˜(x)
 = 1 implies p(x) ≥ 1. Then for any y ∈ X s.t. f˜(y) 6= 0 we have
y y
that: f˜ f˜(y) = 1 and so that p f˜(y) ≥ 1 which implies that |f˜(y)| ≤ p(y).
The latter obviously holds for f˜(y) = 0. Hence, (5.2) is established.

5.2 Applications of Hahn-Banach theorem

The Hahn-Banach theorem is frequently applied in analysis, algebra and ge-
ometry, as will be seen in the forthcoming course. We will briefly indicate
in this section some applications of this theorem to problems of separation of
convex sets and to the multivariate moment problem. From now on we will
focus on t.v.s. over the field of real numbers.

77
5. The Hahn-Banach Theorem and its applications

5.2.1 Separation of convex subsets of a real t.v.s.

Let X t.v.s.over the field of real numbers and H be a closed affine hyperplane
of X. We say that two disjoint subsets A and B of X are separated by H
if A is contained in one of the two closed half-spaces determined by H and
B is contained in the other one. We can express this property in terms of
functionals. Indeed, since H = L−1 ({a}) for some L : X → R linear not
identically zero and some a ∈ R, we can write that A and B are separated by
H if and only if:

∃ a ∈ R s.t. L(A) ≥ a and L(B) ≤ a.

where for any S ⊆ X the notation L(S) ≤ a simply means ∀s ∈ S, L(s) ≤ a

(and analogously for ≥, <, >, =, 6=).
We say that A and B are strictly separated by H if at least one of the two
inequalities is strict. (Note that there are several definition in literature for
the strict separation but for us it will be just the one defined above) In the
present subsection we would like to investigate whether one can separate, or
strictly separate, two disjoint convex subsets of a real t.v.s..
Proposition 5.2.1. Let X be a t.v.s. over the real numbers and A, B two
disjoint convex subsets of X.
a) If A is open nonempty and B is nonempty, then there exists a closed affine
hyperplane H of X separating A and B, i.e. there exists a ∈ R and a func-
tional L : X → R linear not identically zero s.t. L(A) ≥ a and L(B) ≤ a.
b) If in addition B is open, the hyperplane H can be chosen so as to strictly
separate A and B, i.e. there exists a ∈ R and L : X → R linear not
identically zero s.t. L(A) ≥ a and L(B) < a.
c) If A is a cone and B is open, then a can be chosen to be zero, i.e. there
exists L : X → R linear not identically zero s.t. L(A) ≥ 0 and L(B) < 0.

Proof.

a) Consider the set A − B := {a − b : a ∈ A, b ∈ B}. Then: A − B is an open

subset of X as it is the union of the open sets A − y as y varies over B;
A − B is convex as it is the Minkowski sum of the convex sets A and −B;
and o ∈/ (A − B) because if this was the case then there would be at least a
point in the intersection of A and B which contradicts the assumption that
they are disjoint. By applying Theorem 5.1.2 to N = {o} and U = A − B
we have that there is a closed hyperplane H of X which does not intersect
A−B (and passes through the origin) or, which is equivalent, there exists a

78
 5.2. Applications of Hahn-Banach theorem

linear form f on X not identically zero such that f (A−B) 6= 0. Then there
exists a linear form L on X not identically zero such that L(A − B) > 0
(in the case f (A − B) < 0 just take L := −f ) i.e.

∀ x ∈ A, ∀ y ∈ B, L(x) > L(y). (5.5)

Since B 6= ∅ we have that a := inf x∈A L(x) > −∞. Then (5.5) implies that
L(B) ≤ a and we clearly have L(A) ≥ a.
b) Let now both A and B be open convex and nonempty disjoint subsets of
X. By part a) we have that there exists a ∈ R and L : X → R linear not
identically zero s.t. L(A) ≥ a and L(B) ≤ a. Suppose that there exists
b ∈ B s.t. L(b) = a. Since B is open, for any x ∈ X there exists ε > 0 s.t.
for all t ∈ [0, ε] we have b + tx ∈ B. Therefore, as L(B) ≤ a, we have that

L(b + tx) ≤ a, ∀ t ∈ [0, ε]. (5.6)

Now fix x ∈ X, consider the function f (t) := L(b + tx) for all t ∈ R whose
first derivative is clearly given by f 0 (t) = L(x) for all t ∈ R. Then (5.6)
means that t = 0 is a point of local minimum for f and so f 0 (0) = 0 i.e.
L(x) = 0. As x is an arbitrary point of x, we get L ≡ 0 on X which is a
contradiction. Hence, L(B) < a.
c) Let now A be a nonempty convex cone of X and B an open convex
nonempty subset of X s.t. A ∩ B = ∅. By part a) we have that there
exists a ∈ R and L : X → R linear not identically zero s.t. L(A) ≥ a and
L(B) ≤ a. Since A is a cone, for any t > 0 we have that tA ⊆ A and so
tL(A) = L(tA) ≥ a i.e. L(A) ≥ at . This implies that L(A) ≥ inf t>0 at = 0.
Moreover, part a) also gives that L(B) < L(A). Therefore, for any t > 0
and any x ∈ A, we have in particular L(B) < L(tx) = tL(x) and so
L(B) ≤ inf t>0 tL(x) = 0. Since B is also open, we can exactly proceed as
in part b) to get L(B) < 0.

Let us show now two interesting consequences of this result which we will
use in the following subsection.
Corollary 5.2.2. Let X be a vector space over R endowed with the finest
locally convex topology ϕ. If C is a nonempty closed cone in X and x0 ∈ X \C
then there exists a linear functional L : X → R non identically zero s.t.
L(C) ≥ 0 and L(x0 ) < 0.

Proof. As C is closed in (X, ϕ) and x0 ∈ X \ C, we have that X \ C is an open

neighbourhood of x0 . Then the local convexity of (X, ϕ) guarantees that there

79
5. The Hahn-Banach Theorem and its applications

exists an open convex neighbourhood V of x0 s.t. V ⊆ X \ C i.e. V ∩ C = ∅.

By Proposition 5.2.1-c), we have that there exists L : X → R linear not
identically zero s.t. L(C̊) ≥ 0 and L(V ) < 0, in particular L(x0 ) < 0.

Before giving the second corollary, let us introduce some notations. Given
a cone C in a t.v.s. (X, τ ) we define the first and the second dual of C w.r.t.
τ respectively as follows:

Cτ∨ := {` : X → R linear |` is τ − continuous and `(C) ≥ 0}

Cτ∨∨ := {x ∈ X |∀ ` ∈ Cτ∨ , `(x) ≥ 0}.

Corollary 5.2.3. Let X be a vector space over R endowed with the finest
ϕ
locally convex topology ϕ. If C is a nonempty cone in X, then C = Cϕ∨∨ .

Proof. Let us first observe that C ⊆ Cϕ∨∨ , because for any x ∈ C and any
` ∈ Cϕ∨ we have by definition of first dual of C that `(x) ≥ 0 and so that
ϕ ϕ
x ∈ Cϕ∨∨ . Then we get that C ⊆ Cϕ∨∨ . But Cϕ∨∨ is closed since Cϕ∨∨ =
∨ ϕ
⊆ Cϕ∨∨ .
T
∨ `([0, +∞)) and each ` ∈ Cϕ is ϕ-continuous. Hence, C
`∈Cϕ
ϕ
Conversely, suppose there exists x0 ∈ Cϕ∨∨ \ C . By Corollary 5.2.2, there
ϕ
exists a linear functional L : X → R non identically zero s.t. L(C ) ≥ 0 and
L(x0 ) < 0. As L(C) ≥ 0 and every linear functional is ϕ−continuous, we have
L ∈ Cϕ∨ . This together with the fact that L(x0 ) < 0 give x0 ∈
/ Cϕ∨∨ , which is
ϕ
a contradiction. Hence, C = Cϕ∨∨ .

5.2.2 Multivariate real moment problem

Let d ∈ N and let R[x] be the ring of polynomials with real coefficients and d
variables x := (x1 , . . . , xd ). Fixed a subset K of Rd , we denote by

P sd(K) := {p ∈ R[x] : p(x) ≥ 0, ∀x ∈ K}.

Definition 5.2.4 (Multivariate real K−moment problem).

Given a closed subset K of Rd and a linear functional L : R[x] → R, does
there exists a nonnegative finite Borel measure µ s.t.
Z
L(p) = p(x)µ(dx), ∀p ∈ R[x]
Rd

and supp(µ) ⊆ K (where supp(µ) denotes the support of the measure µ)?
If such a measure exists, we say that µ is a K-representing measure for L
and that it is a solution to the K−moment problem for L.

80
 5.2. Applications of Hahn-Banach theorem

A necessary condition for the existence of a solution to the K−moment

problem for the linear functional L is clearly that L is nonnegative on P sd(K).
In fact, if there exists a representing measure µ for L then for all p ∈ P sd(K)
we have Z Z
L(p) = p((x))µ(dx) = p((x))µ(dx) ≥ 0
Rd K
since µ is nonnegative and supported on K and p is nonnegative on K.
It is then natural to ask if the nonnegative of L on P sd(K) is also sufficient.
The answer is positive and it was established by Riesz in 1923 for d = 1 and
by Haviland for any d ≥ 2.
Theorem 5.2.5 (Riesz-Haviland Theorem). Let K be a closed subset of Rd
and L : R[x] → R be linear. L has a K−representing measure if and only if
L(P sd(K)) ≥ 0.
Note that this theorem provides a complete solution for the K− moment
problem but it is quite unpractical! In fact, it reduces the solvability of the
K−moment problem to the problem of classifying all polynomials which are
nonnegative on a prescribed closed subset K of Rd i.e. to characterize P sd(K).
This is actually a hard problem to be solved for general K and it is a core
question in real algebraic geometry. For example, P if we think of the
P case2
K = Rd then for d = 1 we know that P sd(K) = R[x]2 , where R[x]
denotes the set of squares of polynomials. However, for d ≥ 2 this equality
does not hold anymore as it was proved by Hilbert in 1888. It is now clear
that to make the conditions of the Riesz-Haviland theorem actually checkable
we need to be able to write/approximate a non-negative polynomial on K in
a way that makes its non-negativity apparent, i.e. as a sum of squares or as
an element of quadratic modules of R[x]. For a special class of closed subsets
of Rd we actually have such representations and we can get better conditions
than the one of Riesz-Haviland type to solve the K−moment problem.
Definition 5.2.6. Given a finite set of polynomials S := {g1 , . . . , gs }, we call
the basic closed semialgebraic set generated by S the following

KS := {x ∈ Rd : gi (x) ≥ 0, i = 1, . . . , s}.

Definition 5.2.7. A subset M of R[x] is said to be a quadratic module if

1 ∈ M , M + M ⊆ M and h2 M ⊆ M for any h ∈ R[x].
Note that each quadratic module is a cone in R[x].
Definition 5.2.8. A quadraticPd module M of R[x] is called Archimedean if
2
there exists N ∈ N s.t. N − ( i=1 xi ) ∈ M .

81
5. The Hahn-Banach Theorem and its applications

For S := {g1 , . . . , gs } finite subset of R[x], we define the quadratic module

generated by S to be:
( s )
X X
2
MS := σi gi : σi ∈ R[x] , i = 0, 1, . . . , s ,
i=0

where g0 := 1.
Remark 5.2.9. Note that MS ⊆ P sd(KS ) and MS is the smallest quadratic
module of R[x] containing S.
Consider now the finite topology on R[x] (see Definition 4.5.1 ) which
we have proved to be the finest locally convex topology on this space (see
Proposition 4.5.2) and which we therefore denote by ϕ. By Corollary 5.2.3,
we get that
ϕ
MS = (MS )∨∨ ϕ (5.7)
Moreover, the Putinar Positivstellesatz (1993), a milestone result in real al-
gebraic geometry, provides that if MS is Archimedean then
ϕ
P sd(KS ) ⊆ MS . (5.8)

Note that MS is Archimedean implies that KS is compact while the converse

is in general not true (see e.g. M. Marshall, Positive polynomials and sum of
squares, 2008).
Combining (5.7) and (5.8), we get the following result.
Proposition 5.2.10. Let S := {g1 , . . . , gs } be a finite subset of R[x] and
L : R[x] → R linear. Assume that MS is Archimedean. Then there exists a
KS -representing measure µ for L if and only if L(MS ) ≥ 0, i.e. L(h2 gi ) ≥ 0
for all h ∈ R[x] and for all i ∈ {1, . . . , s}.

Proof. Suppose that L(MS ) ≥ 0 and let us consider the finite topology ϕ
on R[x]. Then the linear functional L is ϕ-continuous and so L ∈ (MS )∨ϕ.
Moreover, as MS is assumed to be Archimedean we have
(5.8) ϕ (5.7)
P sd(KS ) ⊆ MS = (MS )∨∨
ϕ .

Since any p ∈ P sd(KS ) is also an element of (MS )∨∨

ϕ , we have that for any
∨
` ∈ (MS )ϕ , `(P sd(KS )) ≥ 0 and in particular L(P sd(KS )) ≥ 0. Hence, by
Riesz-Haviland theorem we get the existence of a KS -representing measure µ
for L.

82
 5.2. Applications of Hahn-Banach theorem

Conversely, suppose that the there exists a KS -representing measure µ

for L. Then for all p ∈ MS we have in particular that
Z
L(p) = p(x)µ(dx)
Rd

which is nonnegative as µ is a nonnegative measure supported on KS and

p ∈ MS ⊆ P sd(KS ).

From this result and its proof we understand that whenever we know that
ϕ
P sd(KS ) ⊆ MS , we need to check only that L(MS ) ≥ 0 to find out whether
there exists a solution for the KS −moment problem for L. Then it makes sense
to look for closure results of this kind in the case when MS is not Archimedean
and so we cannot apply the Putinar Positivstellesatz. Actually whenever we
τ
know that P sd(KS ) ⊆ MS where τ is a locally convex topology on R[x], the
condition L(MS ) ≥ 0 is necessary and sufficient for the existence of a solution
of the KS −moment problem for any τ −continuous functional on R[x] (see
M. Ghasemi, S. Kuhlmann, E. Samei, 2012). This relationship between the
closure of quadratic modules and the representability of functionals continuous
w.r.t. locally convex topologies started a new research line in the study of the
moment problem which is still bringing interesting results.

83