Topological Vector Spaces
Topological Vector Spaces
Topological Vector Spaces
3.1 Definitions
Banach spaces, and more generally normed spaces, are endowed with two structures:
a linear structure and a notion of limits, i.e., a topology. Many useful spaces
are Banach spaces, and indeed, we saw many examples of those. In certain cases,
however, one deals with vector spaces with a notion of convergence that is not
normable. Perhaps the best example is the space of C0∞ (K) functions that are
compactly supported inside some open domain K (this space is the basis for the
theory of distributions). Another such example is the space of continuous functions
C(W) defined on some open set W ⊂ Rn . In such cases, we need a more general
construct that embodies both the vector space structure and a topology, without
having a norm. In certain cases, the topology is not only not normable, but even not
metrizable (there are situations in which it is metrizable but not normable).
We start by recalling basic definition in topological spaces:
A topological space is the most basic concept of a set endowed with a notion of
neighborhood.
Nx = {U ∈ t � x ∈ U}.
182 Topological vector spaces
Bases are useful because many properties of topologies can be reduced to statements
about a base generating that topology, and because many topologies are most easily
defined in terms of a base that generates them.
� Example 3.1 The open balls
form a base for a topology in a metric space. A set is open if and only if it is the
union of open balls. �
Definition 3.5 — Local base. Let (S,t) be a topological space and let x ∈ S. A
collection gx ⊂ Nx is called a local base at x (*/&8/ 2*2") if every neighborhood
of x contains an element of gx .
It is easy to see that the union of all local bases is a base for the topology. Indeed, let
A ∈ t. Let x ∈ A. Since A is a neighborhood of x, then by the definition of the local
base, there exists a Ux ∈ gx , such that Ux ⊂ A. Clearly,
A = � Ux ,
x∈A
Definition 3.6 — Closure. Let (S,t) be a topological space and let E ⊂ S. The
closure of E, denoted E, is the set of all points all of whose neighborhoods
intersect E:
E = {x ∈ S � ∀U ∈ Nx , U ∩ E ≠ �}.
3.1 Definitions 183
Definition 3.7 — Induced topology. Let (S,t) be a topological space and let
E ⊂ S. Then t ∩ E is called the induced topology on E.
So we have two notions: that of a vector space and that of a topological space. We
now blend the two together.
¿ For every point x ∈ V , the singleton {x} is a closed set (namely, {x}c ∈ t).
¡ The vector space operations are continuous with respect to t.
V ×V → V (x,y) � x + y
F ×V → V (a,x) � ax
x+y
x y
That is, every neighborhood of zero contains E after being blown up sufficiently.
As will be shown later, the topological notion of boundedness may not coincide with
the metric notion of boundedness.
∀a ∈ F , �a� ≤ 1 aE ⊂ E.
The notion of being balanced is purely algebraic. When we talk about balanced
neighborhoods we connect the algebraic concept to the topological concept. Note
that in C the only balanced sets are discs and the whole of C.
Proof. Both Ta and Ma are bijections by the vector space axioms. Their inverses
are T−a and M1�a . All are continuous by the very definition of a topological vector
3.1 Definitions 185
space. �
Na = a + N0 .
Exercise 3.1 Consider the vector space R endowed with the topology t gener-
ated by the base
B = {[a,b) � a < b}.
Show that (R,t) is not a topological vector space. �
A topological vector space can be quite abstract. All we know is that there is a
vector space structure and a topology that is compatible with it. We have to start
make our way from these very elementary concepts.
∀W ∈ N0 ∃U ∈ N0 U +U ⊂ W.
sym
such that
U +U ⊂ W.
W
V2
V1
3.2 Separation theorems 187
∀W ∈ N0 ∃V ∈ N0 V +V +V +V ⊂ W.
sym
such that
(K +V ) ∩ (C +V ) = �.
In other words, there exist disjoint open sets that contain K and C.
C
K
Define
V = Vx1 ∩ ⋅⋅⋅ ∩Vxn .
Then, for every i,
and a forteriori,
Comment 3.3 Since K +V and C +V are both open and mutually disjoint, it follows
that also
K +V does not intersect C +V.
Indeed, if x ∈ C +V then there exists U ∈ Nx such that
U ⊂ C +V.
y ∈ (K +V ) ∩ (C +V ),
which is a contradiction.
Corollary 3.6 Let B be a local base at zero for a topological vector space X .
Then every member in B contains the closure of some other member in B. That
is,
∀U ∈ B ∃W ∈ B such that W ⊂ U.
V ∩ (U c +V ) = �.
It follows that
V ⊂ (U c +V )c ⊂ U.
By the definition of a local base there exists a neighborhood W ∈ B such that
W ⊂ V ⊂ (U c +V )c ⊂ U.
3.2 Separation theorems 189
Since (U c +V )c is closed,
W ⊂ (U c +V )c ⊂ U.
�
Proof. Let x,y ∈ X , x ≠ y. Both sets {x} and {y} are closed (by definition) and
compact, hence X is Hausdorff by Theorem 3.5. �
The following proposition establishes relations between sets, their closure, and their
interior.
¿ For A ⊂ X ,
A = � (A +V ).
V ∈N0
That is, the closure of a set is the intersection of all the open neighborhoods
of that set.
¡ A + B ⊂ A + B.
¬ If Y ⊂ X is a linear subspace, then so is Y .
√ If C is convex so is C.
ƒ If C is convex so is C○ .
≈ For every B ⊂ X : if B is balanced so is B.
∆ For every B ⊂ X : if B is balanced and 0 ∈ B○ then B○ is balanced.
« If E ⊂ X is bounded then so is E.
Proof.
x ∈� � (A +V ).
V ∈N0
190 Topological vector spaces
V +W ⊂ U.
and
x ∈V and y ∈W implies x + y ∈ V +W ⊂ U.
In other words, every neighborhood of a + b ∈ A + B intersects A + B, which
implies that a + b ∈ A + B, and therefore
A + B ⊂ A + B.
Y +Y ⊂ Y and ∀a ∈ F , aY ⊂ Y .
aY ⊂ Y .
√ Convexity is a purely algebraic property, but closures and interiors are topo-
logical concepts. The convexity of C implies that for all t ∈ [0,1]:
tC + (1 −t)C ⊂ C.
ƒ Suppose once again that C is convex. Let x,y ∈ C○ . This means that there exist
neighborhoods U,V ∈ N0 , such that
x +U ⊂ C and y +V ⊂ C.
Since C is convex:
t(x +U) + (1 −t)(y +V ) = (tx + (1 −t)y) +U + (1 −t)V ⊂ C,
which proves that tx + (1 −t)y ∈ C○ , namely C○ is convex.
≈ Since multiplication by a (non-zero) scalar is a homeomorphism,
aB = aB.
If B is balanced, then for �a� ≤ 1,
aB = aB ⊂ B,
hence B is balanced.
∆ Again, for every 0 < �a� ≤ 1,
aB○ = (aB)○ ⊂ B○ .
Since for a = 0, aB○ = {0}, we must require that 0 ∈ B○ for the latter to be
balanced.
« Let V be a neighborhood of zero. By Corollary 3.6 there exists a neighborhood
W of zero such that W ⊂ V . Since E is bounded, E ⊂ tW ⊂ tW ⊂ tV for
sufficiently large t. It follows that for large enough t,
E ⊂ tW ⊂ tV,
which proves that E is bounded.
Proof.
and by convexity,
�b �A = �b �A + (1 − �b �){0} ⊂ A.
Since A contains the origin, A○ is balanced; it is also convex.
Corollary 3.10 Every topological vector space has a local base whose elements
are balanced (a balanced local base).
Corollary 3.11 Every locally convex topological vector space has a local base
whose elements are convex and balanced (a balanced convex local base).
{dnV � n ∈ N}
Proof.
¿ Let x ∈ X and consider the sequence x�rn . This sequence converge to zero by
the continuity of the scalar multiplication F × X → X . Thus, for sufficiently
large n,
x�rn ∈ V i.e. x ∈ rnV.
¡ Let K ⊂ X be compact. We need to prove that it is bounded, namely, that for
every V ∈ N0 ,
K ⊂ tV for sufficiently large t.
Let V ∈ N0 be given. By Theorem 3.9 there exists a W ∈ N0bal such that W ⊂ V .
By the previous item
∞
K ⊂ � (nW ).
n=1
Since K is compact,
K K nk
K ⊂ � (nkW ) = nK � � W � ⊂ nK W,
k=1 k=1 nK
where in the last step we used the fact that W is balanced. Thus,
nK
∀t > nK K ⊂ nK W = t � W � ⊂ tW,⊂ tW ⊂ tV,
t
which proves that K is bounded.
¬ Let U ∈ N0 . Since V is bounded,
V ⊂ tU for sufficiently large t,
or,
t −1V ⊂ U. for sufficiently large t.
Thus, dnV ⊂ U for sufficiently large n.
�
194 Topological vector spaces
Exercise 3.3 Let (X ,t) be a topological vector space. Show that a set E ⊂ X
is bounded if and only if every countable subset of E is bounded. �
We already know what are linear maps between vector spaces. The following
theorem is an obvious consequence of the translational invariance of the space:
Lemma 3.14 Linear maps between topological vector spaces map balanced sets
into balanced sets.
In Banach spaces continuous and bounded linear maps are the same. We need a
notion of bounded map to obtain a generalization for topological vector spaces.
The following theorem concerns only scalar-valued linear maps, i.e., linear function-
als. Please note that since F is a topological vector space, there is a well-defined
notion of continuity for functionals. In particular, we may define:
Definition 3.16 Let (X ,t) be a topological vector space. Its dual X ∗ is the
space of continuous linear functionals over X .
¿ f is continuous.
¡ ker f is closed.
¬ ker f is not dense in X .
√ f maps some V ∈ N0 into a bounded set in F .
Proof.
¿ Suppose that f is continuous, then since {0} ∈ F is closed then f −1 (0) = ker f
is closed.
¡ Suppose that ker f is closed. Since we assume that f ≠ 0, it must be that
ker f ≠ X , hence ker f can’t be dense in X .
¬ Suppose that ker f is not dense in X . That is, its complement has a non-empty
interior. There exists an x ∈ X and a V ∈ N0 , such that
(x +V ) ∩ ker f = �. (3.1)
Comment 3.4 If f maps some V ∈ N0 into a bounded set then f is bounded. Let
W ⊂ X be bounded. Then for large enough t:
W ⊂ tV,
hence
sup� f (W )� ≤ sup� f (tV )� = t sup� f (V )� < ∞.
As for Banach spaces, finite-dimensional topological vector spaces are simpler than
infinite-dimensional ones.
u j = T (e j ) j = 1,...,n.
Proof.
K = T (S)
S
T −1 (V ∩ Y ) V ∩Y
n
F Y
E = T −1 (V ∩ Y )
is disjoint from S.
By Lemma 3.14, E which is the linear image of a balanced set is also balanced,
hence it is connected. It follows that T −1 maps V ∩ Y into BF n . T −1 is an
n-tuple of linear functionals on Y . Since it maps a bounded neighborhood of
zero in Y into a bounded set in F , it is continuous by Theorem 3.15, hence
T is a homeomorphism.
198 Topological vector spaces
Proposition 3.19 Every locally compact topological vector space has finite di-
mension.
Proof. Let X be a locally compact topological vector space: it has some neighbor-
hood V whose closure V is compact. By Theorem 3.12 V is bounded and hence also
V . Moreover,
B = {V �2n ∶ n ∈ N}
is a local base for X .
Let y ∈ V . By definition y − 12 V intersects V , i.e., y ∈ V + 12 V , from which we infer
that
V ⊂ V + 12 V = � (x + 12 V ).
x∈V
1
It is easy to show that x ∈ V and x ∈ Y implies that x ∈ V ∩ Y .
3.5 Metrization 199
and hence
V ⊂ Y + 12 V ⊂ Y + Y + 14 V = Y + 14 V.
We may continue similarly to show that V ⊂ Y + 18 V , and so on, namely
∞
V ⊂ � (Y + 21n V ).
n=1
Corollary 3.20 Every locally bounded topological vector space that has the
Heine-Borel property has finite dimension.
3.5 Metrization
What does it take for a topological vector space to be metrizable? Suppose there is
a metric d compatible with the topology t. Thus, all open sets are unions of open
balls, and in particular, the countable collection of balls B(0,1�n) forms a local
base at the origin. It turns out that the existence of a countable local base is also
sufficient for metrizability.
Theorem 3.21 Let X be topological vector space that has a countable local base.
Then there is a metric d on X such that:
Proof. Let B ′ be a countable local base. By Theorem 3.9 and Lemma 3.3 there
exists a countable local base B = {Vn } whose members are all balanced, and further-
more,
Vn+1 +Vn+1 +Vn+1 +Vn+1 ⊂ Vn .
(If X is locally convex, the local base can be chosen to include convex sets.) This
implies that for all n and k:
The sum in this definition is always finite. For example, j(1.2) = X and j(0.75) =
V1 +V2 . By the property of the base {Vn },
� N2 cn � N2
j � = � cnVn ⊂ VN1 −1 .
�n=N1 2n � n=N1
and define
d(x,y) = f (y − x).
We will show that d is indeed a metric on X that satisfies all required properties.
This will rely on the following fact:
0 ∈ j(r).
f (x + y) ≤ f (x) + f (y).
Proof. Let x,y ∈ X be given. Note that the range of f is [0,1], hence we can limit
ourselves to the case where the right hand side is less than 1. Fix e > 0. There are
r,s ∈ D such that
f (x) < r f (y) < s and r + s < f (x) + f (y) + e.
Because {j(r)} is fully ordered, this implies that x ∈ j(r), y ∈ j(s), hence
x + y ∈ j(r) + j(s) ⊂ j(r + s).
Thus,
f (x + y) ≤ r + s < f (x) + f (y) + e,
which holds for every e > 0. �
¿ f (x) = f (−x).
¡ f (0) = 0.
¬ f (x) > 0 for x ≠ 0.
Proof. Since the j(r) are unions of balanced sets they are balanced, from which
follows that f (x) = f (−x). f (0) = 0 because 0 ∈ j(r) for all r ∈ D. Finally, if x ≠ 0, it
does no belong to some Vn (by separation) i.e., to some j(s), and since the {j(r)}
form an ordered set, it does not belong to j(r) for all r < s, from which follows that
f (x) > 0. �
Back to the main proof. The properties of f imply that d(x,y) = f (x − y) is a metric
on X . It is symmetric, vanishes if and only if x = y, it is translationally invariant,
and
d(x,y) = f (x − y) = f (x − z − (y − z))
≤ f (x − z) + f (y − z)
= d(x,z) + d(z,y).
We next want to show that this metric is compatible with the topology. Consider the
d-open balls,
B(0,d ) = {x ∶ d(x,0) < d } = {x ∶ f (x) < d } = � j(r).
r<d
3.6 Cauchy sequences 203
(We used the fact that f (x) < t implies that x ∈ j(t).) In particular, if d < 1�2n , then
B(0,d ) ⊂ Vn , which proves that the open balls, B(0,1�2n ), form a local base.
The open balls are balanced because each j(r) is balanced and the union of balanced
set is balanced. If the Vn are convex then j(r) are convex.
�
We know from second year undergrad: a sequence (xn ) in a metric space (X ,d) is
called a Cauchy sequence if
This definition is independent of the local base since if this is true for one local base
it is true for any other local base.
Proposition 3.27 Let (X ,d1 ) and (Y ,d2 ) be metric spaces, where (X ,d1 ) is
complete. Let E ⊂ X1 be a closed set and suppose that
f ∶E →Y
Proof. Take y ∈ f (E). Then there exists a sequence f (xn ) ∈ f (E) such that f (xn ) → y
in Y . It follows that (xn ) is a Cauchy sequence, and since E is closed, the limit x is
in E. By the continuity of f ,
y = lim f (xn ) = f (x) ∈ f (E).
n→∞
�
Proof. Recall that being an F-space means that the topology on Y is consistent with
a complete translationally invariant metric.
Let d be such an invariant metric, and set
B1�n = {y ∈ Y ∶ d(y,0) < 1�n},
which are open balls in Y .
Let Un ∈ N0 be neighborhoods of zero in X such that Un ∩ Y = B1�n (such neigh-
borhoods exist because the open sets in Y are intersections of open sets in X with
Y ). Let then Vn ∈ N0 such that Vn+1 ⊂ Vn and Vn +Vn ⊂ Un .
sym
FnW = (x +W ∩Vn ) ∩ Y .
By the exact same same argument the intersection of the Y closure of the sets FnW
contains exactly one point. But since FnW ⊂ En , this point has to be y0 . Thus, there
exists a unique point y0 , such that for every U,W ∈ N0 in X and every n,
(y0 +U) ∩ (x +W ) ≠ �.
Lemma 3.29
d(nx,0) ≤ nd(x,0).
Proof. The first part is obvious by successive applications of the triangle inequality,
n
d(nx,0) ≤ � d(kx,(k − 1)x) ≤ nd(x,0).
k=1
For the second, we note that since d(xn ,0) → 0, there exists a diverging sequence of
positive integers nk , such that
1
d(xk ,0) ≤ 2 ,
nk
from which we get that
1
d(nk xk ,0) ≤ nk d(xk ,0) ≤ → 0.
nk
�
206 Topological vector spaces
We defined bounded sets in topological vector spaces. When the space is metrizable,
one could also define a set to be bounded if its diameter is finite. The two definitions
are not equivalent. For example, the metric that was defined in Theorem 3.21
satisfies d(x,y) ≤ 1, so that even X is a bounded set. We will always refer to the
topological notion of boundedness.
We already know that compact sets are bounded (Theorem 3.12), and that the closure
of a bounded set is bounded (Proposition 3.8).
Here is a simple exercise:
Proof. Let (xn ) be a Cauchy sequence. Let W ∈ N0bal , and let V ∈ N0bal satisfy
V +V ⊂ W.
By the definition of a Cauchy sequence, there exists an N such that for all m,n ≥ N,
xn − xm ∈ V,
and in particular,
∀n > N xn ∈ xN +V.
Set s > 1 such that xN ∈ sV (we know that such an S exists), then for all n > N,
xn ∈ sV +V ⊂ sV + sV ⊂ sW.
Since for balanced sets sW ⊂ tW for s < t, and since every neighborhood of zero
contains a balanced neighborhood, this proves that the sequence is indeed bounded.
�
Corollary 3.32 The only bounded subspace of a topological vector space is {0}.
an E ⊂ (ant)V ⊂ V.
We are now in measure to establish the relations between continuous and bounded
linear transformations. As we shall see, the relation is more involved than in normed
spaces, where the two notions coincide.
¿ T is continuous.
¡ T is bounded.
¬ If xn → 0 then {T (xn ) � n ∈ N} is bounded.
√ If xn → 0 then T (xn ) → 0.
Then,
¿ �⇒ ¡ �⇒ ¬.
If, moreover, X is metrizable, then
¬ �⇒ √ �⇒ ¿,
V ⊂ tU.
T (V ) ⊂ T (tU) = t T (U) ⊂ t W,
E = {xn � n ∈ N}
is bounded, then
T (E) = {T (xn ) � n ∈ N}
is bounded.
¬→√. Suppose now that X is metrizable and let d be a compatible and invariant
metric. Suppose that xn → 0 and {T (xn ) � n ∈ N} is bounded. By the second part of
Lemma 3.29 (and here we need metrizability), there exists a sequence an → ∞, such
that an xn → 0. Then,
1
T (xn ) = T (an xn ),
an ��� � � � � �� � � � � � � �
bounded
Comment 3.5 The only difference between a norm and a seminorm is that the
latter does not separate between points; knowing that p(y − x) = 0 does not guarantee
that x = y. While a single seminorm does not separate between points, a family of
seminorms might do so.
Recall the definition of absorbing sets (;3-&" %7&"8). In fact, we proved that
every neighborhood of zero is absorbing. For convex absorbing sets K (in a vector
space—no topology required) we defined the Minkowski functionals,
We proved that the Minkowski functionals are sub-additive, and pK (ax) = a pK (x)
for a > 0. Also, pK (x) ≤ 1 for x ∈ K and pK (x) < 1 for internal points (+&; ;&$&81)
of K.
There turns out to be a strong relation between Minkowski functionals and semi-
norms: every seminorm is a Minkowski functional of some balanced, convex,
absorbing set. This follows from the following propositions:
¿ p is symmetric.
¡ p(0) = 0.
¬ �p(x) − p(y)� ≤ p(x − y).
√ p(x) ≥ 0.
ƒ ker p is a linear subspace.
≈ The set B = {x ∶ p(x) < 1} is convex, balanced and absorbing.
∆ p = pB .
p(x) ≤ p(y) + p(x − y) and p(y) ≤ p(x) + p(y − x) = p(x) + p(x − y).
ƒ If x,y ∈ ker p:
≈ If p(x) < 1 then for every �a� ≤ 1, p(ax) = �a�p(x) < 1, which proves that B is
balanced. If x,y ∈ B then for every 0 ≤ t ≤ 1
hence B is convex. For every x ∈ X and s > p(x), p(x�s) < 1, i.e., x ∈ sB, so
that B is absorbing.
∆ Finally, consider
pB (x) = inf{s > 0 ∶ x�s ∈ B} = inf{s > 0 ∶ p(x�s) < 1} = inf{s > 0 ∶ p(x) < s}.
If p(x) < r then pB (x) < r, i.e., pB ≤ p. If pB (x) < r, then there exists an
s < pB (x) such that p(x) = s, i.e., p ≤ pB .
And now for the reverse direction, which we have already shown:
¿ pK (x + y) ≤ pK (x) + pK (y).
¡ pK (tx) = t pK (x) for all t > 0.
¬ If K is also balanced then pK (ax) = �a� pK (x).
To conclude:
Proposition 3.38 Let B be a convex and balanced local base in a (locally con-
vex!) topological vector space X . Then,
Comment 3.6 This is essentially what local convexity gives—the fact that the
Minkowski functional over that base are both continuous and separating.
Proof.
¿ Let x ∈ V . Since V is open, there is a t < 1 such that x�t ∈ V , which implies that
pV (x) = inf{s > 0 ∶ x�s ∈ V } < 1
i.e.,
V ⊂ {x � pV (x) < 1}.
Conversely, if x ∈� V , then by convexity x�t ∈ V only for t > 1, i.e., pV (x) ≥ 1.
¡ We already know that the functions pV are seminorms. We will first show that
this family of seminorms is separating. Let x ≠ 0. Let V ∈ B be such that x ∈� V
(separation). Then pV (x) ≥ 1 > 0. The continuity of the pV ’s follows from the
inequality
�pV (x) − pV (0)� ≤ pV (x).
Recall that {V �n ∶ n ∈ N} is a local base. If x ∈ V �n then
1 1
�pV (x) − pV (0)� ≤ pV (x) = pV (nx) < .
n n
That is, for all e > 0 there exists a neighborhood U ∈ N0 such that
pV (U) ⊂ [0,e).
And why do we care about separating families of seminorms? The answer is below:
a separating family of seminorms allows us to define a topology on a vector space.
In fact, very often a vector space is given without any topology, and the way to
topologize it by means of a separating family of seminorms.
¿ Every p ∈ P is continuous.
Proof. We defined
where all the intersections are finite. Clearly, B is a collection of sets that contain
the origin and that are closed under finite intersections.
Set A ⊂ X be be open if it is a union of translates of members of B. That is, A is
open if and only if it can be represented in the form
A= � (x + B(x)),
x∈X ,B(x)∈B
where the union can be empty. We denote the collection of all such set by t. t is a
translation-invariant topology on X because:
¿ � ∈ t.
¡ X ∈ t (obvious as well).
¬ t is closed under arbitrary unions.
√ t is closed under finite intersection as
� (x + B(x)) ∩ � (y + B(y)) = � � (x + B(x)) ∪ (y + B(y))
x∈I,B(x)∈BI y∈J,B(y)∈BJ x∈I,B(x)∈BI y∈J,B(y)∈BJ
=????
∃p ∈ P p(x) > 0.
We next show that addition is continuous. Let U ∈ N0 . Then, by the property of the
local base, there exist
V (p1 ,n1 ) ∩ ⋅⋅⋅ ∩V (pk ,nk ) ⊂ U.
Let
W = V (p1 ,2n1 ) ∩ ⋅⋅⋅ ∩V (pk ,2nk ),
then since seminorms are subadditive,
V (p1 ,2n1 ) +V (p1 ,2n1 ) = {x � p1 (x) < 1�2n} + {x � p1 (x) < 1�2n}
= {x + y � p1 (x) < 1�2n, p1 (y) < 1�2n}
⊂ {x + y � p1 (x + y) < 1�n} = V (p1 ,n1 ),
from which we conclude that
W +W ⊂ U.
This proves that addition is continuous! A similar argument shows that scalar
multiplication is continuous as well.
X is a locally convex topological vector space. Every p ∈ P is continuous at 0
because for every e > 0 set n > 1�e and
p(V (p,n)) ⊂ B(0,e).
It is continuous everywhere by
�p(x) − p(y)� ≤ p(x − y).
Comments 3.1
¿ Let (X,t) be a locally convex topological vector space, and let B be a convex
balanced local base. This local base defines a separating family of seminorms
on X , which by the previous theorem induces a topology t ′ on X . Is this
topology the same as t? The answer is positive.
¡ If P = (pn ) is countable separating family of seminorms, then the resulting
topology has a countable local base, and by Theorem 3.21 it is metrizable.
One can show that the following is a compatible metric:
cn pn (x − y)
d(x,y) = max
1 + pn (x − y)
,
n
where (cn ) is any sequence of positive numbers that decays to zero (it is easy
to see that the maximum is indeed attained). Clearly, d(x,x) = 0. Also, since
the pn ’s are separating d(x,y) > 0 for x ≠ y. Symmetry, as well as translational
invariance are obvious. Finally, the triangle inequality follows form the fact
that every pn is subadditive, and that a ≤ b + c implies that
a b c
≤ +
1+a 1+b 1+c
.
And finally:
Theorem 3.40 (X ,t) is normable if and only if there exists a convex bounded
neighborhood.
3.9 Examples
Let W an open set in Rn . We consider the space C(W) of all continuous functions.
Note that the sup-norm does not work here. There exist unbounded continuous
functions on open sets.
Every open set W in Rn can be written as
∞
W = � Kn ,
n=1
where Kn � Kn+1 , where the Kn are compact, and � stands for compactly embedded,
i.e., Kn is a compact set in the interior of Kn+1 . We topologize C(W) with the
separating family of seminorms,
(These are clearly seminorms, and they are separating because for every f ≠ 0 there
exists an n such that f �Kn ≠ 0.)
Since the pn ’s are monotonically increasing,
M n M n
� � V (pk ,m) = � � { f ∶ pk ( f ) < 1�m} = V (pn ,M),
m=1 k=1 m=1 k=1
which means that the V (pn ,M) form a convex local base for C(W). In fact, V (pn ,M)
contains a neighborhood obtained by taking n,M to be the greatest of the two, from
which follows that
Vn = { f ∶ pn ( f ) < 1�n}
is a convex local base for C(W), and the pn ’s are continuous in this topology.
We can thus endow this topological space with a compatible metric, for example,
2−n pn ( f − g)
d( f ,g) = max
1 + pn ( f − g)
.
n
We will now show that this space is complete. Recall that if a topological vector
space has a compatible metric with respect to which it is complete, then it is called
an F-space. If, moreover, the space is locally convex, then it is called a Fréchet
space. Thus, C(W) is a Fréchet space.
Let then ( fn ) be a Cauchy sequence. This means that for every e > 0 there exists an
N, such that for every m,n > N,
2−k pk ( fn − fm )
max < e,
k 1 + pk ( f n − f m )
216 Topological vector spaces
and a fortiori,
2−k pk ( fn − fm )
(∀k ∈ N) < e,
1 + pk ( f n − f m )
which means that ( fn ) is a Cauchy sequence in each Kk (endowed with the sup-
norm), and hence converges uniformly to a function f .
Given e let K be such 2−K < e, then
2−k pk ( fn − f )
max < e,
k>K 1 + pk ( fn − f )
2−k pk ( fn − f )
max < e,
k≤K 1 + pk ( fn − f )
{sup{� f (x)� ∶ x ∈ Kn } ∶ f ∈ E}
Vk = { f ∶ pk ( f ) < 1�k},
hence,
sup{� f (x)� ∶ x ∈ Kn , f ∈ E} ≥ sup{� f �Kn ∶ � f �Kk < 1�k}.
The right hand side can be made as large as we please for n > k, i.e., no set is
bounded, and hence the space is not normable.
We endow C∞ (R) with a topology, turning it into a Fréchet space, such that for
every compact set K, DK is a closed subspace. As above, let (Kn ) be an increasing
sequence of compact sets covering R. We define the seminorms:
3.10 Completeness
This section is devoted to the notion of completeness. Most of the content of the
second chapter of Rudin coincides with material learned along this course. We will
only highlight differences that result from the more general context.
218 Topological vector spaces
Theorem 3.41 Let X be a Banach space and let Y be a normed space. Let
{Ta ∶ a ∈ A} be a collection of bounded linear maps, such that
Then,
sup �Ta � < ∞.
a
We will now see how this generalizes for topological vector spaces. First, a defini-
tion:
We have shown that continuity implies boundedness. We will now show that
equicontinuity implies uniform boundedness:
(∀a ∈ A) Ta (E) ⊂ F.
F = � Ta (E).
a∈A
(∀a ∈ A) Ta (V ) ⊂ U.
3.10 Completeness 219
Then,
(∀a ∈ A)(∀t > s) Ta (E) ⊂ Ta (tV ) = t Ta (V ) ⊂ tU,
hence F ⊂ tU, which proves that F is bounded. �
If B is of the second category (i.e., not a countable union of nowhere dense sets),
then B = X and the {Ta } are equicontinuous.
Comment 3.8 OK, let’s remind ourselves what is this category stuff. A set is called
nowhere dense (-*-$) if its closure has an empty interior. An example of a nowhere
dense set in R is the set
{1�n ∶ n ∈ N}.
The complement of the closure of a nowhere dense set is dense.
A set is called of the first category if it is a countable union of nowhere dense sets.
An example of a set of the first category is the set of functions in C[0,1] that are
differentiable at some point. A set that is not of the first category is of the second
category.
Baire’s theorem states that in a complete metric set, if the space can be covered by
a countable union of closed set, then one of them contains an open set. (Equivalently,
Baire’s theorem states that the intersection of every countable collection of dense
open sets is dense.)
Let En be a countable collection of nowhere dense sets. Then, (En )c is open and
dense, and by Baire’s theorem,
∞
� (En ) ≠ �.
c
n=1
Taking complements,
∞
� (En ) ≠ X .
n=1
That is, a complete metric space is not a countable union of nowhere dense sets—it
is of the second category.
220 Topological vector spaces
U +U ⊂ W,
and set
E = � Ta−1 (U) = {x ∈ X ∶ ∀a ∈ A,Ta (x) ∈ U}.
a∈A
If B is of the second category, then at least one of the nE is of the second category3 ,
i.e., E itself is of the second category. A set of the second category has an interior
point, say x, hence E − x contains a neighborhood V of zero. Then for every a ∈ A,
Ta (V ) ⊂ Ta (x) − Ta (E) ⊂ U −U ⊂ W,
{Ta (x) ∶ a ∈ A}
is bounded, i.e., B = X . �
{Ta (x) ∶ a ∈ A}
How does this last corollary translate in the case of a Banach space? If X is a
Banach space and Y is a normed space, then if
then the {Ta } are equicontinuous, which in turn implies that they are uniformly
bounded. That is, for every there exists an M such that the (bounded) unit ball in X
is mapped into a bounded set in Y ,
and hence,
sup �Ta � < ∞,
a∈A
We felt with the Hahn-Banach theorem in Chapter 2, but recall that most of the
content had nothing to do with Banach spaces. Let us first refresh our memory.
The theorem that we called the Hahn-Banach theorem was the following:
� f ,⋅� ≤ p�Y .
There there exists a linear functional F on V such that F�Y = f and �F,⋅� ≤ p.
and �F� = � f �.
Definition 3.23 Let X be a topological vector space. Its dual, X ∗ , is the vector
space of continuous linear functionals on X .
Comment 3.9 As for now, the dual space is only a vector space, but not a topologi-
cal vector space.
Note that this theorem refers to an internal point, which unlike an interior point is
not a topological concept; it is a vector space concept.
In contrast, the following corollary applies to normed spaces:
Corollary 3.48 Let X be a normed space and let K1 ,K2 be disjoint convex sets
that at least one of them has an interior point. Then there exists a bounded linear
functional f ≠ 0 that separates K1 and K2 .
Id ∶ (S,t1 ) → (S,t2 )
is an open mapping and
Id ∶ (S,t2 ) → (S,t1 )
is continuous.
Comment 3.12 This means that one cannot weaken a compact Hausdorff topology
without losing the Hausdorff property. This also means that one cannot strengthen a
compact Hausdorff space without losing compactness.
Let S be a set and let F be a family of mappings from S into topological spaces:
F = { fa ∶ S → Ya � a ∈ I}.
Let t be the topology generated by the subbase
{ fa−1 (V ) � a ∈ I, V ∈ tYa }.
Then t is the weakest topology on S for which all the fa are continuous maps (it
is the intersection of all topologies having this property). It is called the weak
topology induced by F , or the F -topology of S.
224 Topological vector spaces
Comment 3.13 Here separates between point means that x ≠ y implies that f (x −
y) ≠ f (0). In the linear case it reduces to f (x) ≠ f (y).
f (x − y) ≠ f (0).
Because Y is Hausdorff there exists U ∈ N f (x−y) and W ∈ N f (0) that are disjoint. By
definition f −1 (U) and f −1 (W ), which are neighborhoods of x −y and 0, are F -open.
The sets
x + f −1 (U) and y + f −1 (W )
are disjoint and open. �
Proof. Since (S,t) is compact and the fn are continuous then they are bounded.
Thus, we can normalize them such that � fn � ≤ 1. Define:
∞
� fn (x) − fn (y)�
d(x,y) = � .
n=1 2n
td ⊂ t.
By Proposition 3.51, since t is compact and td is Hausdorff (like any metric space)
then t = td . �
3.12 Weak topologies 225
Lemma 3.54 Let f1 ,..., fn and f be linear functionals on a vector space V (no
topology). Let
N = ker f1 ∩ ⋅⋅⋅ ∩ ker fn .
Then the following are equivalent:
i.e.,
f ∈ Span{ fk � k = 1,...,n}.
¬ N ⊂ ker f .
then f vanishes on N.
Finally, suppose that ¬ holds. Define T ∶ V → F n ,
If is a linear map. If T (x) = T (y) then y − x ∈ N and f (x) = f (y). Define a map
s ∶ T (V ) → F
by
s ( f1 (x),..., fn (x)) = f (x).
s can be extended as a linear functional to all of F n , i.e., there are scalars a1 ,...,an
such that
n
s (u1 ,...,un ) = � ak uk ,
k=1
226 Topological vector spaces
and in particular
n
f (x) = s ( f1 (x),..., fn (x)) = � ak fk (x).
k=1
Theorem 3.55 Let V be a vector space (no topology) and let V ′ be a separating
vector space of linear functionals on V . Denote by t ′ the V ′ -topology on V .
Then (V ,t ′ ) is a locally convex topological vector space whose dual space is V ′ .
{ f −1 (A) � f ∈ V ′ , A open in F }.
and f is linear.
The topology t ′ is generated by the local subbase,
These sets are balanced and convex (by linearity) hence every finite intersection of
them (i.e. every element in the local base) is balanced and convex. If follows that
(V ,t ′ ) is a locally convex topological space.
We next show that (V ,t ′ ) is a topological vector space. Note that
1 1
V ( f ,r) + V ( f ,r) = { 12 x + 12 y � � f (x)� < r, � f (y)� < r} ⊂ V ( f ,r).
2 2
It follows that every U in the base satisfies such an inclusion as well, hence addition
is continuous. Scalar multiplication is continuous by a similar argument.
It remains to show that V ′ is the dual space of (V ,t ′ ). Every f ∈ V ′ is t ′ -continuous
as for all e > 0
� f (V ( f ,e))� = � f ({x � � f (x)� < e})� ⊂ [0,e).
3.12 Weak topologies 227
� f (U)� ⊂ [0,1).
for which
sup � f (x)� < 1.
x∈U
In particular,
ker f ⊂ ker f1 ∩ ⋅⋅⋅ ∩ ker fm .
It follows from that last lemma that f ∈ Span{ f1 ,..., fm }. �
Definition 3.25 Let (X ,t) be a topological vector space whose dual X ∗ (the
vector space of continuous linear functionals) separates points. The X ∗ -topology
on X is called the weak topology (it is the weakest topology with respect every
t-continuous linear functional is continuous).
Note that since every f ∈ X ∗ is t continuous and since tw is the weakest topology
with this property, it follows that
tw ⊂ t,
justifying the name of weak topology.
The next proposition shows that weak convergence is consistent with what we know:
∀f ∈ X ∗ f (xn ) → 0.
Proof. Weak convergence means that for every tw -neighborhood V of zero, the
sequence is eventually in V . Since every tw -neighborhood of zero contains a set of
the form
{x � � f j (x)� < r j , j = 1,...,n},
if follows that f (xn ) → 0 for all f ∈ X ∗ guarantees weak convergence.
Conversely, if xn ⇀ 0, then the sequence is eventually in every tw -neighborhood of
zero, and in particular, for all f ∈ X ∗ and e > 0, the sequence is eventually in
hence f (xn ) → 0. �
∀f ∈ X ∗ f is a bounded functional on E.
3.12 Weak topologies 229
Proof. E is weakly bounded iff (by definition) for all tw -neighborhoods V of zero
E ⊂ tV
for sufficiently large t. This means that for every set of the form
Comment 3.15 Recall that a topological vector space is metrizable if and only if it
has a countable local base. Local boundedness implies the existence of a countable
local base. Not being locally bounded does not imply a lack of metrizability.
Set as before
N = ker f1 ∩ ⋅⋅⋅ ∩ ker fn .
The map
x � ( f1 (x),..., fn (x))
is a map X → F n with null space N. Hence,
dimX ≤ n + dimN,
i.e., dimN = ∞. Since N ⊂ V it means that every element in the base contains an
infinite-dimensional subspace. �
Proof. The proof of this theorem relies on a version of the Hahn-Banach separation
theorem: if A,B are convex disjoint non-empty sets in a locally convex topological
vector space, A is compact and B is closed, then there exists an f ∈ X ∗ such that
Let x0 ∈� E. It follows from the Hahn-Banach separation theorem that there exists an
f ∈ X ∗ and a g ∈ R such that
The set
{x � Re f (x) < g}
x ∈� E w ,
namely E w ⊂ E. �
Comment 3.16 This last theorem states that if E is a convex set in a locally convex
topological vector space and there is a sequence xn ∈ E that weakly converges to x
(which is not necessarily in E), then there is also a sequence yn ∈ E that originally-
converges to x. We proved such a theorem for Hilbert spaces (using the Banach-Saks
theorem).
Corollary 3.63 For convex subsets of locally convex topological vector spaces:
Let (X ,t) be a topological vector space. The dual space X ∗ does not come with
an a priori topology. In Banach spaces, the dual space has a natural operator norm,
and we proved that the dual space endowed with that norm is also a Banach space
(in fact, we proved that the dual of a normed space is a Banach space). But for
general topological vector spaces, we don’t have as for now a topology.
Recall the natural inclusion from X to linear functionals on X ∗ ,
i ∶ x → Fx ,
where for f ∈ X ∗ :
Fx ( f ) = f (x).
The family of functionals {i(x) � x ∈ X } (which we can’t call continuous because
there is no topology on X ∗ ) separates points on X ∗ as if
i(x)( f ) = i(x)(g)
� Example 3.2 Recall that c∗0 = `1 and `∗1 = `∞ . Weak convergence of a sequence
(xn )k ⊂ `1 to zero (with `1 viewed as a topological vector space) means that
∞
∀yk ∈ `∞ lim �(xn )k yk = 0.
n→∞
k=1
Clearly, weak convergence implies weak-* convergence (but not the opposite). In
reflexive spaces the two notions of convergence are identical. �
232 Topological vector spaces
The following central theorem states a compactness property of the weak-* topology.
It was proved in 1932 by Banach for separable spaces and in 1940 by Alaoglu in the
general case. (Leonidas Alaoglu (1914–1981) was a Greek mathematician.)
with the product topology tP (the weakest topology with respect to which the
projections px ∶ P → Dx are continuous). Each Dx is compact, and by Tychonoff’s
theorem (which relies on the axiom of choice), P is compact. Every element in P
can be identified with a function f ∶ X → F (not necessarily linear) satisfying
� f (x)� ≤ g(x).
K ⊂ X ∗ ∩ P.
The set K can therefore be assigned two topologies: the weak-* topology t∗ induced
from X ∗ and the product topology tP induced from P.
W1 = { f ∈ X ∗ � � f (x j ) − f0 (x j )� < d , j = 1,...,n},
and
W2 = { f ∈ P � � f (x j ) − f0 (x j )� < d , j = 1,...,n}.
Sets of the form W1 and W2 are local bases for t∗ and tP . Since
W1 ∩ K = W2 ∩ K
{ f ∈ P � � f (x) − f0 (x)� < e, � f (y) − f0 (y)� < e, � f (ax + b y) − f0 (ax + b y)� < e}.
By definition of the closure there is an element f ∈ K that belongs to this set, i.e., it
is a linear functional satisfying
� f (x) − f0 (x)� < e � f (y) − f0 (y)� < e and �a f (x) + b f (y) − f0 (ax + b y)� < e.
Since this holds for all e > 0, we conclude that � f0 (x)� ≤ 1 for all x ∈ V , i.e., f0 ∈ K. �
The Banach-Alaoglu theorem holds in a very general setting. The following theorem
holds in the case where the space is separable:
234 Topological vector spaces
Comment 3.17 The claim is not that X ∗ endowed with the weak-* topology is
metrizable. For example, this is not true in infinite-dimensional Banach spaces.
Proof. Take
{xn � n ∈ N}
a dense set in X and Fn = i(xn ). By definition of the weak-* topology on X ∗ , the
functional Fn are weak-* continuous. Also, if for every n,
Fn ( f ) = Fn (g),
i.e.,
f (xn ) = g(xn ),
then f = g (continuous functionals that coincide on a dense set).
Thus, {Fn � n ∈ N} is a countable family of continuous functionals that separate
points in X ∗ . It follows by Proposition 3.53 that K is metrizable. �
for all x ∈ X .
Definition 3.26 Let V be a vector space and E ⊂ V . The convex hull (9&/8) of
E, denoted Conv(E), is the intersection of all the convex sets that contains E
(equivalently, it is the set of convex combinations of elements in E).
Definition 3.27 Let V be a vector space and E ⊂ V . The closed convex hull of
E, denoted Conv(E) is the closure of the convex hull of E.
Theorem 3.69
Proof.
Consider the set K = f (S,A). It is the continuous image of compact sets and
it is therefore compact. Moreover,
K ⊃ Conv(A1 ∪ ⋅⋅⋅ ∪ An ).
236 Topological vector spaces
It is easy to show that K is convex, and since it includes all the Ak ’s it must in
fact be equal to Conv(A1 ∪ ⋅⋅⋅ ∪ An ).
¡ Let U ∈ N0 . Because X is locally convex there exists a convex neighborhood
V such that
V +V ⊂ U.
Since E is totally bounded there exists a finite set F such that
E ⊂ F +V ⊂ Conv(F) +V.
By the first item Conv(F) is compact, therefore there exists a finite set F ′
such that
Conv(F) = F ′ +V,
i.e.,
Conv(E) ⊂ F ′ +V +V ⊂ F ′ +U,
which proves that Conv(E) is totally bounded.
¬ In every metric space the closure of a totally bounded set is totally bounded,
and if the space is complete it is compact. Since K is compact, then it is
totally bounded. By the previous item Conv(K) is totally bounded and hence
its closure is compact.
√ Let S ⊂ Rn+1 be the convex simplex. One can show that Conv(K) is the image
of the continuous map S × K n+1 :
n+1
(s,x1 ,...,xn ) � � sk xk ,
k=1
Extreme points (0&7*8 ;&$&81) are extreme sets that consist of one point. The set
of all extreme points of K is denoted by Ext(K).
3.13 The Krein-Milman theorem 237
Comment 3.18 Note that extreme sets and extreme points are a pure vector space
concept.
Comment 3.19 Every set is an extreme set of itself.
� Example 3.3 The extreme points of a convex polygon are its vertices. The
extreme points of a circle are the entire circle. �
Comment 3.20 David Milman (1912–1982) was a Soviet and then an Israeli
Mathematician. He is the father of Vitali Milman, a mathematician at Tel-Aviv
University, and grandfather of Emanuel Milman, a mathematician at the Technion.
Comment 3.21 Locally convex topological vector spaces have the property that
X ∗ separates points.
Proof. Let
P = {compact extreme subsets of K}.
This set is not empty because K ∈ P.
P is closed under non-empty intersections: For every P ′ ⊂ P and non-empty
S= � A
A∈P ′
tx + (1 −t)y ∈ S,
where x,y ∈ K and t ∈ (0,1) implies that x,y ∈ A for every A ∈ P ′ , i.e., x,y ∈ S. I.e.,
every non-empty intersection of compact extreme sets of K is a compact extreme set
of K.
Let S ∈ P and f ∈ X ∗ . Define
S f = {x ∈ S � Re f (x) = µ},
where
µ = max Re f (x).
x∈S
Since S is compact, S f is non-empty; in fact, it is compact. Suppose
z = tx + (1 −t)y ∈ S f .
238 Topological vector spaces
PS = {A ∈ P � A ⊆ S}.
M = � A.
A∈W
This intersection has the finite intersection property hence its it is not empty. We
know that M ∈ PS . By the maximality of W, there is no proper subspace of M that
belongs to PS .
Let f ∈ X ∗ . Since M f ⊂ M and M f is non-empty we conclude that M = M f , i,e, Re f
is constant on M. Multiply by ı and we conclude that f is constant on M. Since X ∗
separates point, it follows that M is a singleton, i.e., an extreme point. Thus, for
every compact extreme set S of K,
S ∩ Ext(K) ≠ �.
Since K is convex
Conv(Ext(K)) ⊂ K.
Since it is also compact,
Conv(Ext(K)) ⊂ K.
We need to show that this is in fact an equality.
Suppose, by contradiction that
Conv(Ext(K)) ∩ K f = �,