TOPOLOGICAL VECTOR SPACES

PRADIPTA BANDYOPADHYAY

1. Topological Vector Spaces

Let X be a linear space over R or C. We denote the scalar field by K.

Definition 1.1. A topological vector space (tvs for short) is a linear space X
(over K) together with a topology J on X such that the maps (x, y) → x+y
and (α, x) → αx are continuous from X × X → X and K × X → X
respectively, K having the usual Euclidean topology.

We will see examples as we go along.

Remark 1.2. Let X be a tvs. Then

(1) for fixed x ∈ X, the translation map y → x + y is a homeomor-
phism of X onto X and
6 0 ∈ K, the map x → αx is a homeomorphism of X
(2) for fixed α =
onto X.

Definition 1.3. A base for the topology at 0 is called a local base for the
topology J .

Theorem 1.4. Let X be a tvs and let F be a local base at 0. Then

(i) U, V ∈ F ⇒ there exists W ∈ F such that W ⊆ U ∩ V .
(ii) If U ∈ F, there exists V ∈ F such that V + V ⊆ U .
(iii) If U ∈ F, there exists V ∈ F such that αV ⊆ U for all α ∈ K
such that |α| ≤ 1.
(iv) Any U ∈ J is absorbing, i.e. if x ∈ X, there exists δ > 0 such
that ax ∈ U for all a such that |a| ≤ δ.
Conversely, let X be a linear space and let F be a non-empty family of
subsets of X which satisfy (i)–(iv), define a topology J by :

These notes are built upon an old set of notes by Professor AK Roy. The debt is
gratefully acknowledged.
1
2 BANDYOPADHYAY

A ⊆ X is open if, for each x ∈ A, there exists U ∈ F such that x + U ⊆ A.

Then J is a topology which makes (X, J ) a tvs with F as a local base.

Proof. Exercise. 

Some elementary facts concerning a tvs are summarized in

Lemma 1.5. Let X be a tvs. Then

(i) x + A = x + A.
(ii) A + B ⊇ A + B.
(iii) If U is open and A is any subset then A + U is open.
(iv) C, D compact ⇒ C + D compact.
(v) If A ⊆ X, A = ∩{(A + U ) : U neighborhood of 0}.
(vi) J is Hausdorff if and only if {0} is closed if and only if {0} =
∩{U : U ∈ F }, for any local base F.
(vii) If U is a neighborhood of 0, there exists a balanced neighborhood
V of 0 such that V ⊆ U .
(viii) Closure of a convex set is convex; closure of a subspace is a
subspace; closure of a balanced set is balanced.
(ix) If C is compact, U neighborhood of C then there exists a neigh-
borhood V of 0 such that C + V ⊆ U .
(x) C compact, F closed ⇒ C + F closed.
(xi) If U is a balanced neighborhood of 0 then int(U ) is balanced.
(xii) If U is any neighborhood of 0, U contains a closed balanced
neighborhood of 0. In other words, the closed balanced neighborhoods
form a local base (at 0).
(xiii) Every convex neighborhood of 0 contains a closed, balanced, con-
vex neighborhood of 0.

Proof. (x) Suppose x 6∈ C + F , i.e. (x − F ) ∩ C = ∅ or C ⊆ (x − F )c .

Now, (x − F )c is open and C compact, therefore by (ix), there exists a
neighborhood V of 0 such that C + V ⊆ (x − F )c , i.e. (C + V ) ∩ (x − F ) =
∅ ⇒ x 6∈ C + F + V ⇒ x 6∈ C + F by (v).
(xii) There exists a balanced neighborhood V of 0 such that V + V ⊆ U .
Now V is also balanced and V ⊆ V + V ⊆ U .
(xiii) Let U be a convex neighborhood of 0 and define V = ∩{αU : α ∈ K,
|α| = 1}. V is convex, balanced (easy to check!) and a neighborhood of 0 as
 TOPOLOGICAL VECTOR SPACES 3

U contains a balanced neighborhood of 0 by (xii). Now, 12 V ⊆ 12 V + 12 V ⊆ V ,

hence 12 V is a closed convex, balanced neighborhood of 0 contained in U . 

The class of tvs mostly used in analysis is given by

Definition 1.6. A tvs X is called locally convex if there is a local base at 0

whose members are convex.

The topology of a lctvs is precisely that generated by a family of seminorms.

Recall that

Definition 1.7. A function p : X → R is sublinear if p is subadditive and

positively homogeneous, i.e.,
(a) p(x + y) ≤ p(x) + p(y) for all x, y ∈ X, and
(b) p(αx) = αp(x) for all x ∈ X, and α ≥ 0
p is called a seminorm if for all x, y ∈ X and α ∈ K,
(a) p(x + y) ≤ p(x) + p(y), and
(b) p(αx) = |α|p(x).
If p(x) = 0 ⇒ x = 0, call p a norm.

We show now that a family of seminorms on a linear space generates a locally

convex topology in the following sense :

Theorem 1.8. Let {pi : i ∈ I} be a family of seminorms on the linear

space X. Let U be the class of all finite intersections of sets of the form
{x ∈ X : pi (x) < δi } where i ∈ I, δi > 0. Then U is a local base for a
topology J that makes X a locally convex tvs. This topology is the weakest
making all the pi continuous, and for a net {xα } ⊆ X, xα → x in J if and
only if pi (xα − x) → 0 for each i ∈ I.

Proof. We check that all the conditions of Theorem 1.4 are satisfied.
(i). U is clearly closed under finite intersections.
(ii). Suppose U = {x : pi (x) < δi , i = 1, . . . , n}. Let δ = mini δi /2 and
V = {x : pi (x) < δ, i = 1, . . . , n}. If y, z ∈ V then pi (y +z) ≤ pi (y)+pi (z) <
δ/2 + δ/2 = δ < δi ⇒ V + V ⊆ U .
(iii). Since pi (αx) = |α|pi (x), each U ∈ U is balanced.
(iv). If x ∈ X, pi (αx) = |α|pi (x) < δi if α is sufficiently small.
4 BANDYOPADHYAY

J is a locally convex topology as each U ∈ U is convex.

Each U = {x : pi (x) < δi , i = 1, . . . , n} is open as if x ∈ U , x + V ⊆ U
where V = {z ∈ X : pi (z) < minj [δj − pj (x)]}.
Now continuity of pi is equivalent to continuity at 0 as |pi (x) − pi (y)| ≤
pi (x − y) and the sets U ∈ U are open in any topology that makes each pi
continuous at 0. This proves that the given topology is the weakest, making
the pi ’s continuous.
Finally, xα → x ⇔ xα − x → 0 ⇔ pi (xα − x) → 0 for all i by the definition
of U. 

Remark 1.9. J is a Hausdorff topology if and only if the family {pi : i ∈ I}

is separating, i.e., given x 6= 0, there exists pi such that pi (x) 6= 0.

Example 1.10. (a) Let X be the vector space of all K-valued continuous
functions on a topological space Ω. For each compact set K ⊆ Ω, define
pK (f ) = supt∈K |f (t)|. The family {pK : K ⊆ Ω compact} gives a Hausdorff
vector topology in which convergence means uniform convergence on all
compact subsets of Ω. If K is restricted to finite subsets of Ω, we get the
topology of pointwise convergence. In general, if the sets K are restricted
to a class C of subsets of Ω, we obtain the topology of uniform convergence
on sets in C.
(b) Let X = C ∞ [a, b], the vector space of all infinitely differentiable (K-
valued) functions on the closed bounded interval [a, b]. For each n, define
pn (f ) = sup{|f (n) (t)| : t ∈ [a, b]} where f (n) is the n-th derivative of f . In
the topology defined by the pn , convergence means uniform convergence of
all derivatives.

We will now prove the converse of Theorem 1.8, that is, locally convex vector
topologies are generated by families of seminorms. But we first examine
convex sets in some detail.

Definition 1.11. A subset K ⊆ X is said to be radial at x if and only if

K contains a line segment through x in each direction, i.e. for every y ∈ X,
there exists δ > 0 such that x + λy ∈ K for all λ ∈ [0, δ] (sometimes x is
called an internal point of K). If K is convex and radial at 0 (equivalently,
K is absorbing), the Minkowski functional of K is defined as

pK (x) = inf{r > 0 : x ∈ rK}.

 TOPOLOGICAL VECTOR SPACES 5

Intuitively, pK (x) is the factor by which x must be shrunk in order to reach

the boundary of K.

Lemma 1.12. Let K be convex and radial at 0.

(a) pK is sublinear.
(b) {x ∈ X : pK (x) < 1} = {x ∈ K : K is radial at x} ⊆ K ⊆ {x :
pK (x) ≤ 1}.
(c) If K is balanced, pK is a seminorm.
(d) If X is a tvs and 0 ∈ K ◦ , the interior of K, then pK is continuous.
K = {pK ≤ 1}, K ◦ = {pK < 1}, hence {pK = 1} = ∂K, the
boundary of K.

Proof. (a). Let ε > 0 be given. There exists r > 0, s > 0 such that
r < pK (x) + ε/2, s < pK (y) + ε/2 and x/r, y/s ∈ K. Now,
x+y r x s y 
= + ∈K
r+s r+s r r+s s
by convexity, hence pK (x + y) ≤ (r + s) < pK (x) + pK (y) + ε.
(b). If pK (x) < 1 then x/r ∈ K for some r < 1 ⇒ x = r(x/r) + (1 − r)(0) ∈
K. If y ∈ X then pK (x + λy) ≤ pK (x) + λpK (y) < 1 if λ > 0 is sufficiently
small, hence {pK < 1} ⊆ {x ∈ K : K is radial at x}. Conversely, if K is
radial at x, then x + λx ∈ K for some λ > 0, hence pK (x + λx) ≤ 1 by
1
definition of pK ⇒ pK (x) ≤ 1+λ < 1. By definition of pK , K ⊆ {pK ≤ 1}.
a x
(c). If x/r ∈ K and a 6= 0 ∈ K, then |a| r ∈ K (as K is balanced) ⇒
pK (ax) ≤ |a|r. Thus, pK (ax) ≤ |a|pK (x) (taking infimum over r).
Taking x/a instead of x, we get pK (x) ≤ |a|pK (x/a). Putting b = 1/a, we
get pK (bx) ≥ |b|pK (x).
(d). 0 ∈ K ◦ ⇒ there exists neighborhood U of 0 with U ⊆ K. Let ε > 0
be given. If y ∈ εU then pK (y) = pK (εu) = εpK (u) ≤ ε (since x ∈ K ⇒
pK (x) ≤ 1) ⇒ pK is continuous at 0 ⇒ pK continuous everywhere.
pK continuous ⇒ {pK ≤ 1} is closed ⇒ K ⊆ {pK ≤ 1}. Suppose pK (x) ≤ 1.
If 0 < λ < 1, then pK (λx) = λpK (x) < 1 ⇒ λx ∈ K. If λ → 1 then λx → x,
hence x ∈ K ⇒ K ⊇ {pK ≤ 1}, hence K = {pK ≤ 1}.
pK continuous ⇒ {pK < 1} is open and hence ⊆ K ◦ . But if pK (x) = 1 then
xn = x/(1 − 1/n) ∈ / K as pK (xn ) = 1/(1 − 1/n) > 1 and K ⊆ {pK ≤ 1},
but xn → x, so x is a limit of points not in K, hence x 6∈ K ◦ . 
6 BANDYOPADHYAY

Theorem 1.13. If X is a locally convex tvs, then its topology is generated

by a family P of seminorms.

Proof. If X has a local base consisting of convex sets, it has a local base B
consisting of closed convex balanced neighborhoods of 0. For U ∈ B, the
Minkowski functional pU is a seminorm. Since U = {pu ≤ 1}, the family
{pU : U ∈ B} generates the topolgy of X. 

Definition 1.14. A set E in a tvs Y is said to be bounded if, for every

neighborhood U of 0 in Y , there exists t ∈ R+ , such that E ⊆ tU .

Theorem 1.15. Suppose X is locally convex, so its topology is generated by

a family P of seminorms. Then E ⊆ X is bounded if and only if each p ∈ P
is bounded on E.

Proof. Let E ⊆ X be a bounded set. Let p ∈ P. There exists k > 0 such

that E ⊆ k{p ≤ 1} ⇒ p(E) ≤ k.
Conversely, suppose each p ∈ P is bounded on E. Let U be a neighborhood
of 0. Then U ⊇ {p1 ≤ 1/n1 } ∩ · · · ∩ {pk ≤ 1/nk } for some p1 , . . . , pk ∈ P
and n1 , . . . , nk ∈ N. There exists numbers Mi > 0 such that pi (E) < Mi
(i = 1, . . . , k). Choose M > Mi ni (1 ≤ i ≤ k). If x ∈ E then pi (x/M ) <
Mi /M < 1/ni ⇒ x/M ∈ U ⇒ x ∈ M U . 

Example 1.16. Compact sets are bounded. In a Hausdorff tvs no subspace

other than {0} is bounded.

Remark 1.17. Suppose P = {pi : i = 1, 2, . . .} is a countable separating

family of seminorms on a linear space X generating a vector topology J .
Then there exists a translation-invariant metric compatible with J . Just
define
∞
X 1 pi (x − y)
d(x, y) = .
2n 1 + pi (x − y)
n=1

The only trouble is that the balls {x : d(0, x) ≤ r} need not be convex (they
are balanced though) as we see from the following example:

Example 1.18. Let s = {(xn )∞ n=1 : xn ∈ K for all n ≥ 1}, the space of all
scalar sequences. The topology of pointwise convergence is described by the
seminorms pk , (k ≥ 1), pk ((xn )) = |xk | and the metric is
X 1 |xn − yn |
d(x, y) = , x = (xn ), y = (yn ).
2n 1 + |xn − yn |
 TOPOLOGICAL VECTOR SPACES 7

The ball U = {x : d(0, x) ≤ 14 } is not convex, since (1, 0, 0, . . .),

3 1 3 1
(0, 1, 0, 0, . . .) ∈ U , but (1, 0, 0, . . .) + (0, 1, 0, . . .) = ( , , 0, 0, 0, . . .) ∈
/ U.
4 4 4 4
Can this be rescued? Indeed we have the following theorem whose proof is
omitted (See Rudin, Functional Analysis, Theorem 1.24).

Theorem 1.19. If (X, J ) is a (Hausdorff ) tvs, with a countable local base,

then there is a metric d on X such that
(a) d is compatible with the topology J ,
(b) the balls {x : d(0, x) ≤ r} are balanced
(c) d is translation-invariant: d(x+z, y +z) = d(x, y) for x, y, z ∈ X.
If, in addition, X is locally convex then d can be chosen so that
(d) all open balls {x : d(0, x) < r} are convex.

Remark 1.20. The notion of a Cauchy net for a tvs (X, J ) can be defined
without reference to any metric. Fix a local base F at 0 for the topology J .
A net {xα } is said to be (J -) Cauchy if, for any U ∈ F, there exists α0 such
that α ≥ α0 and β ≥ α0 ⇒ xα − xβ ∈ U . It is clear that different local bases
for J give rise to the same class of Cauchy nets. Now let (X, J ) be metrized
by an invariant metric d. As d is invariant and the d-balls centered at 0 from
a local base, we conclude that a sequence {xn } ⊆ X is a d-Cauchy sequence
if and only if it is a J -Cauchy sequence. Consequently any two invariant
metrics on X that are compatible with the topology have the same Cauchy
sequences and the same convergent sequences, viz. the J -convergent ones.

By far the most widely discussed locally convex spaces are those for which
the vector topologies are given by a single norm, the so-called normed spaces.

Theorem 1.21 (Kolmogoroff). A (Hausdorff ) tvs is normable if and only

if it has a bounded convex neighborhood of 0.

Proof. Let X be normed by k · k. Then the open unit ball {x : kxk < 1} is
a convex and bounded neighborhood of 0.
For the converse, let V be a bounded convex open neighborhood of 0. By
Lemma 1.5 (xiii), there exists neighborhood U of 0 such that U ⊆ V , U
convex, open and balanced. Obviously, U is also bounded. For x ∈ X,
define kxk = pU (x) where pU is the Minkowski functional of U .
Claim: {λU : λ > 0} form a local base for the topology J of X.
8 BANDYOPADHYAY

Let W ∈ J be a neighborhood of 0. U bounded ⇒ U ⊆ λ0 W for some

λ0 > 0 ⇒ λ10 U ⊆ W .
Now, if x 6= 0 then x 6∈ λU for some λ > 0 ⇒ kxk ≥ λ ⇒ k · k is actually a
norm on X. As U = {pU < 1} is open, it’s easy to see that {x : kxk < r} =
rU for all r > 0 and the norm topology coincides with the given one. 

Exercise 1. Show that s, the space of all scalar sequences, is not normable
where the topology on s is defined by the metric
∞
X 1 |xn |
d(0, {xn }) = .
2n 1 + |xn |
n=1

[Hint: Check that U = {x : d(0, x) < r} is not bounded by verifying that

pm (x) = |xm |, x = (xm ), is not bounded on U .]

Exercise 2. Let X be the space of analytic functions on the unit disc

D = {z ∈ C : |z| < 1} with the topology of uniform convergence on compact
subsets. Show that X is metrizable but not normable.
[Hint: Let pn (f ) = sup{|f (z)| : |z| ≤ 1 − n1 } and Un = {f : pn (f ) ≤ n1 }.
Then Un ’s form a local base for the topology, hence X is metrizable. To
show that no Un is bounded, fix N0 . Suppose there exists αn1 < ∞ such
that UN0 ⊆ αn1 Un1 for n1 > N0 . Fix z0 and zn1 such that 1 − N10 <
|z0 | < |zn1 | ≤ 1 − n11 . The functions fk (z) = ( zz0 )k converge uniformly
to 0 on {z : |z| ≤ 1 − N10 }, hence fk /αn1 ∈ Un1 for all k ≥ k0 . But
fk (zn1 ) 1 zn1 k
= ( ) → ∞ as k → ∞, contradiction].
αn1 αn1 z0

2. Quotient Spaces

Let X be a tvs and M ⊆ X a subspace. By definition, the quotient space

X/M consists of cosets x + M = [x] and the quotient map π : X → X/M
is defined by π(x) = x + M . We define a quotient topology on X/M by
stipulating that U ⊆ X/M is open if and only if π −1 (U ) is open in X. With
this definition, we have

Lemma 2.1. X/M is a tvs, π is a continuous and open map, and X/M is
Hausdorff if and only if M is closed.
 TOPOLOGICAL VECTOR SPACES 9

Proof. π is continuous by the definition of the quotient topology. Let V be

open in X. Then π(V ) is open in X/M ⇔ π −1 (π(V )) is open in X ⇔ V +M
is open in X which is true, hence π is an open map.
Now if W is a neighborhood of 0 in X/M . Then there exists a neighborhood
V of 0 in X such that V + V ⊆ π −1 (W ). Since π is an open map, π(V )
is a neighborhood of 0 in X/M and π(V ) + π(V ) ⊆ W . This proves the
continuity of addition in X/M .
Check similarly that K × X/M defined by (α, x + M ) → α(x + M ) = αx + M
is continuous. Hence the quotient topology is a vector topology.
Now X/M is Hausdorff if and only if [0] is closed in X/M if and only if
π −1 [0] = M is closed in X. 

Suppose that (X, J ) is a tvs which is metrized by an invariant metric d. Let

M ⊆ X be a subspace. If x + M, y + M ∈ X/M , define

¯
d([x], [y]) = d(x + M, y + M ) = d(x − y, M )

which is the distance of x − y to the subspace M . (The expression for d¯

shows that it is well-defined). d¯ is a psuedometric on X/M and it is a metric
if M is closed. In the latter event, d¯ is called the quotient metric on X/M .
If X is normed, this definition of d¯ specializes to the quotient norm on X/M :

¯ [x]) = d(x, M ) = inf{kx + mk : m ∈ M }.

kx + M k = d(0,

Theorem 2.2. Let M be a closed subspace of a metrizable tvs X. With the

quotient topology, X/M is a metrizable tvs. Indeed, if d is a translation-
invariant metric which defines the topology on X, the quotient metric d¯
induced by d induces the topology on X/M . If X is complete, so is X/M .

Proof. Let π : X → X/M be the quotient map. It is easy to see that if B is

a local base for X then {π(V ) : V ∈ B} is a local base for X/M . Check that

¯ [u]) < r}.

π({x : d(0, x) < r}) = {[u] : d(0,

It follows that d¯ is compatible with the quotient topology.

To show d complete ⇒ d¯ complete, let {[un ]} be a d-Cauchy
¯ sequence. We
¯
can choose a subsequence {[uni ]} such that d([uni ], [uni+1 ]) < 2−i .
10 BANDYOPADHYAY

Inductively choose xi ∈ X so that d(xi , xi+1 ) < 2−i and π(xi ) = [uni ]. Since
d is complete, xi → x ∈ X and as π is continuous, π(xi ) = [uni ] → π(x).
Hence {[un ]}, being Cauchy, must converge to π(x). 

Exercise 3. If M is a closed subspace of a metrizable tvs X and if both

X/M and M are complete then X is complete.

3. Duals of tvs

Theorem 3.1. Let X be a tvs and f : X → K be a linear functional on X,

f 6≡ 0. Following are equivalent:
(i) f is continuous
(ii) ker(f ) is closed
(iii) ker(f ) is not dense in X
(iv) there exists a neighborhood of 0 on which f is bounded
(v) the image under f of some non-empty open neighborhood of 0 is
a proper subset of K
If K = C, these are also equivalent to :
(vi) Ref is continuous.

Proof. (i) ⇒ (ii) and (iv) ⇒ (v) trivial. (ii) ⇒ (iii) clear as f 6≡ 0.
(iii) ⇒ (iv). Choose x ∈ X and a balanced neighborhood U of 0 such that
(x + U ) ∩ ker(f ) = ∅ ⇒ f (x) ∈
/ −f (U ) (which proves (v)). But f (U ) is
balanced as U is balanced and hence f (U ) is bounded [as a proper balanced
subset of K must be bounded] which proves (iv).
(v) ⇒ (i). Assume that f maps a balanced neighborhood U of 0 onto a
proper subset of K. Hence f (U ) is bounded. We have a k > 0 such that
f (U ) ⊆ {z ∈ K : |z| ≤ k}. Let ε > 0. Then f ( kε U ) ⊆ {z ∈ K : |z| ≤ ε} ⇒ f
is continuous at 0.
(vi) ⇔ (i) follows from the observation that f (x) = Ref (x) − iRef (ix). 

Remark 3.2. The last observation that f (x) = Ref (x)−iRef (ix) is useful.
We can consider a complex vector space X as a vector space over the R
by restricting scalar multiplication to R and the space XR thus obtained
is called the real restriction of X. The above shows that (X ∗ )R is (real)
linearly isomorphic to (XR )∗ under the map f → Ref .
 TOPOLOGICAL VECTOR SPACES 11

We want to know conditions which ensure that there exists non-zero contin-
uous linear functionals on a tvs.

Theorem 3.3. Let X be a tvs. The following are equivalent :

(i) there exists non-zero continuous linear functional on X.
(ii) there exists proper convex neighborhood of 0 in X.
(iii) there exists a non-zero continuous seminorm on X.

Proof. (i) ⇒ (ii). If f 6= 0 and f is continuous then U = {x : |f (x)| < 1} is

a proper convex neighborhood of 0 in X.
(ii) ⇒ (iii) If V is a convex neighborhood of 0 then V contains a convex
balanced neighborhood U of 0. We know that pU is continuous and that
U = {pu < 1}. As U 6= X, pu (x) ≥ 1 if x 6∈ U , hence pu 6≡ 0.
(iii) ⇒ (i). This is the Hahn-Banach theorem which we will prove in Sec-
tion 4. 

Remark 3.4. This result makes it clear why lctvs are important viz. these
have plenty of non-zero continuous linear functionals.

If X is a tvs then the continuous linear functionals on X form a linear space

in the usual way, denoted by X ∗ , and is called the dual or conjugate space
of X.

Exercise 4. Let X is a normed space and Y a Banach space then a linear

operator T : X → Y is continuous if and only if supkxk≤1 kT xk < ∞. If T
is continuous, let kT k = supkxk≤1 kT xk and denote by B(X, Y ) the space
of all continuous linear operators from X to Y . Show that k · k is a norm
on B(X, Y ) and that B(X, Y ) is complete in this norm, i.e. B(X, Y ) is a
Banach space. Specializing to Y = K, we get that X ∗ is always complete
even though X may not be.
P∞
Example 3.5. For 1 ≤ p < ∞, `p = {(xn ) : xn ∈ K, p
n=1 |xn | < ∞}
P∞
with norm k(xn )kp = ( n=1 |xn |p )1/p .
`∞ = {(xn ) : xn ∈ K, supn |xn | < ∞} with norm k(xn )k = supn |xn |.
c = {(xn ) ∈ `∞ : limn xn exists}.
c0 = {(xn ) ∈ `∞ : limn xn = 0}.
These are all Banach spaces and (`p )∗ = `q for 1 ≤ p < ∞, where p1 + 1q = 1,
c∗0 = `1 , c∗ = `1 . The continuous analogues of the `p -spaces are the Lp (µ)
12 BANDYOPADHYAY

spaces where (Ω, Σ, µ) is a σ-finite measure space.

Z
p
L (µ) = {f : Ω → K : f measurable and |f |p dµ < ∞},

with norm kf k = ( |f |p dµ)1/p , 1 ≤ p < ∞.

R

L∞ (µ) = {f : Ω → K : f measurable and essup|f | < ∞}

where essup f = inf{M > 0 : |f | ≤ M a.e. [µ]}.

One knows that (Lp )∗ = Lq for 1 < p < ∞, where p1 + 1q = 1. Also
(L1 )∗ = L∞ but (L∞ )∗ is usually much larger than L1 . More specifically,
Φ ∈ (Lp )∗ if and only if there exists unique g ∈ Lq , unique upto sets of
measure zero, such that Φ(f ) = Ω f ḡ dµ, f ∈ Lp and kΦk = kgkq , so the
R

correspondence between (Lρ )∗ and Lq is given by Φ ↔ g.

For a compact Hausdorff space X, C(X) is a Banach space normed by kf k =
supx∈X |f (x)| and C(X)∗ = M (X), the space of regular Borel measures on
X normed by kµk = supkf k≤1 |µ(f )| = |µ|(1), the total variation norm.
More specifically, Φ ∈ C(X)∗ if and only if there exists unique µ ∈ M (X)
R
such that Φ(f ) = f dµ, (f ∈ C(X)) and kΦk = kµk. If L is a positive
linear functional, i.e. L(f ) ≥ 0 whenever f ∈ C(X) and f ≥ 0 then µ is a
non-negative measure.

Exercise 5. When 0 < p < 1, Lp (dµ) is defined in the usual way but because
of the failure of Minkowski’s inequality for such values of p, one does not
get a norm. However, d(f, g) = |f − g|p dµ does define an invariant metric
R

on Lp (dµ). Show by using (ii) of Theorem 3.3, that (Lp [0, 1])∗ = {0}, with
µ as Lebesgue measure, in contrast to the case when 1 ≤ p < ∞.
Solution: Let Uε = {f : |f |p dµ < ε} is a basic neighborhood. Enough to
show that co(Uε ) is the whole space. Let f ∈ Lp (dµ) and let |f |p dµ = M .
R
Rx p
Since x 0 |f (y)| Rdµ(y) is a continuous function, there exists a point
x
x0 ∈ [0, 1] such that 0 0 |f |p dµ = M 2 . Subdivide [0, x0 ] and [x0 , 1] further
n
in this way to obtain 2 intervals In (n to be determined) in each of which
R p M P2n n
In |f | dµ = 2n . Let gi = f χIn . Then f = i=1 gi . Put hi = 2 gi . Then

2n Z Z
X 1 p M
f= n
hi and |hi | dµ = 2np |fi |p dµ = n(1−p) < ε
2 Ii 2
i=1

if n is sufficiently large. Hence hi ∈ Uε and f ∈ co(Uε ).

 TOPOLOGICAL VECTOR SPACES 13

4. The Hahn-Banach theorem

Theorem 4.1 (Hahn-Banach). Let X be a linear space (over R) and p a

sublinear map on X. Suppose Y ⊆ X is a subspace and f : Y → R a linear
functional with f ≤ p on Y . Then there exists an extension of f to a linear
functional fe on X such that fe ≤ p.

We first prove the following

Lemma 4.2. The theorem holds when Y has codimension one.

Proof. Suppose X = Y ⊕ Ra where a ∈ X \ Y . For a fixed k ∈ R, define

fe(y + λa) = f (y) + λk, y ∈ Y . Notice that fe ≤ p ⇐⇒

f (y) + λk ≤ p(y + λa) ∀λ ∈ R

⇐⇒ k ≤ p(u + a) − f (u) ∀ u ∈ Y (λ ≥ 0)
and k ≥ f (v) − p(v − a) ∀ v ∈ Y (λ < 0).

Thus fe ≤ p ⇐⇒

sup{f (v) − p(v − a) : v ∈ Y } ≤ k ≤ inf{p(u + a) − f (u) : u ∈ Y }

Now, if u, v ∈ Y ,

f (u) + f (v) = f (u + v) ≤ p(u + v) = p((u + a) + (v − a)) ≤ p(u + a) + p(v − a),

hence f (v) − p(v − a) ≤ p(u + a) − f (u). Therefore

(1) sup{f (v) − p(v − a) : v ∈ Y } ≤ inf{p(u + a) − f (u) : u ∈ Y }

and it follows that k, as required, exists. 

Proof of Theorem 4.1. Let G = {(V, g) : V ⊆ X a subspace, Y ⊆ V , g ∈ V 0 ,

g|Y = f , g ≤ p on V } where V 0 is the algebraic dual of V .
Partially order G by: (V1 , g1 ) ≺ (V2 , g2 ) if and only if V1 ⊆ V2 and g2 |V1 = g1 .
It is clear that any chain in G has an upper bound, hence G has a maximal
element (Y 0 , f 0 )  (Y, f ). If Y 0 6= X, we could obtain an element larger than
Y 0 (by one dimension) by Lemma 4.2. This completes the proof. 

Exercise 6. In the theorem, the extension fe is unique if and only if (1)

holds for all a ∈ X \ Y .
14 BANDYOPADHYAY

Exercise 7. If Y = {0} then fe is unique if and only if p is linear.

(Use (1) to get, in this case, −p(−a) = p(a).)

Theorem 4.3 (Complex form of the Hahn-Banach theorem). Let X be a

vector space over C, p a seminorm on X, Y ⊆ X a subspace and f : Y → C
a (complex) linear functional such that |f (x)| ≤ p(x). Then there exists an
extension fe : X → C also C-linear and |fe| ≤ p.

Proof. Let g = Ref , then g : Y → R is R-linear and g ≤ p. By Theorem 4.1,

there exists R-linear extension ge : X → R with ge ≤ p. Define fe : X → C
by fe(x) = ge(x) − ie g (ix). It is clear that fe is C-linear and that fe|Y = f .
Now ge ≤ p ⇒ |e g | ≤ p. Also, |fe| ≤ p, since if fe(x) = reiθ , r ≥ 0, then
fe(e−iθ x) = r ∈ R and hence,

r = |fe(x)| = fe(e−iθ x) = ge(e−iθ x) ≤ p(e−iθ x) = p(x).

5. Consequences

Corollary 5.1. Let X be a linear space and let A ⊆ X be a convex balanced

set which is radial at the origin. Let f be a linear functional on a subspace
M ⊆ X such that |f (y)| ≤ 1 for all y ∈ M ∩ A. Then there exists a linear
functional g on X such that

g|M = f, |g(x)| ≤ 1 ∀ x ∈ A.

Proof. Let y ∈ M , choose r > 0 such that y/r ∈ A (radiality of A) ⇒

|f (y/r)| ≤ 1 ⇒ |f (y)| ≤ r ⇒ |f (y)| ≤ p(y) where p = pA is the Minkowski
functional of A. p as we know is a seminorm. By HB Theorem, there exists
a linear functional g on X extending f and |g(x)| ≤ p(x) for all x ∈ X. If
x ∈ A, p(x) ≤ 1 and it follows that |g| ≤ 1 on A. 

Corollary 5.2. If (X, k · k) is a normed linear space and x0 ∈ X, there

exists a linear functional g on X such that kgk = 1 and g(x0 ) = kx0 k.

Proof. Let y = x0 /kx0 k and define f on Ky by f (αy) = α. Now, |f (αy)| =

|α| ≤ kαyk for all α ∈ K, so by HB Theorem, there exists g on X such that
g|Ky = f (⇒ g(x0 ) = kx0 k) and |g(x)| ≤ kxk for all x (⇒ kgk = 1). 

Remark 5.3. It follows that kxk = sup{|g(x)| : g ∈ X ∗ , kgk ≤ 1}.

 TOPOLOGICAL VECTOR SPACES 15

Theorem 5.4. Let M be a subspace of a lctvs X. Then any continuous

linear functional on M can be extended to a continuous linear functional on
X (when X is a normed linear space this can be accomplished in a norm-
preserving way).

Proof. Let f ∈ M ∗ . Then there exists a neighborhood V of 0 in X such

that |f | ≤ 1 on V ∩ M . Choose a convex balanced neighborhood U of 0
in X such that U ⊆ V and we have |f | ≤ 1 on U ∩ M . By Corollary 5.1
above, there exists linear functional g on X such that g|M = f , |g| ≤ 1 on
U ⇒ g ∈ X ∗. 

Corollary 5.5. If M is a closed subspace of a lctvs X and if x 6∈ M , there

exists f ∈ X ∗ such that f |M = 0 and f (x) 6= 0.

Proof. X/M is a lctvs and x + M 6= [0]. Find f on X/M by HB Theorem,

such that f ∈ (X/M )∗ and f (x + M ) 6= 0. Define g = f ◦ π ∈ X ∗ . Hence
g(x) = f (x + M ) 6= 0. 

Corollary 5.6. If X is locally convex and x, y ∈ X, x 6= y then there exists

f ∈ X ∗ such that f (x) 6= f (y), i.e. X ∗ separates points of X.

Theorem 5.7 (The Separation Theorems). Suppose A, B are disjoint, non-

empty convex sets in a tvs X.
(a) If A is open, there exists f ∈ X ∗ and λ ∈ R such that Re f (x) <
λ ≤ Re f (y) for all x ∈ A, y ∈ B.
(b) If A is compact, B closed and X is locally convex, then there exists
f ∈ X ∗ , λ1 , λ2 ∈ R such that Re f (x) < λ1 < λ2 < Re f (y) for all
x ∈ A, y ∈ B.

Remark 5.8. It is enough to prove these results for tvs over R.

Proof. (a). Fix a0 ∈ A, b0 ∈ B and let x0 = b0 − a0 . Put C = A − B + x0 .

Then C is a convex neighborhood of 0. Let p = pC be the Minkowski
functional of C. Then p is a sublinear functional. Now, A ∩ B = ∅ ⇒ x0 6∈
C ⇒ p(x0 ) ≥ 1. Define f (tx0 ) = t on the subspace M = Rx0 .
Now, if t ≥ 0, f (tx0 ) = t ≤ tp(x0 ) = p(tx0 ) and if t < 0, f (tx0 ) = t < 0 ≤
p(tx0 ) ⇒ f ≤ p on M .
16 BANDYOPADHYAY

By HB Theorem, extend f to a linear functional Λ on X so that Λ ≤ p. In

particular, Λ ≤ 1 on C ⇒ Λ(−x) ≥ −1 on −C ⇒ |Λ| ≤ 1 on C ∩ −C which
is neighborhood of 0 ∈ X ⇒ Λ ∈ X ∗ .
If a ∈ A, b ∈ B then Λ(a)−Λ(b)+1 = Λ(a−b+x0 ) ≤ p(a−b+x0 ) < 1 (since
C is open and a − b + x0 ∈ C) ⇒ Λ(a) < Λ(b). It follows that Λ(A) and
Λ(B) are disjoint convex sets in R with Λ(A) to the left of Λ(B). Also Λ(A)
is open since A is open and every non-constant continuous linear functional
on X is an open map. We get the result by letting λ be the right end point
of Λ(A).
(b). A compact. B closed ⇒ B − A closed ⇒ there exists convex neigh-
borhood U of 0 such that U ∩ (B − A) = ∅ (as A ∩ B = ∅, 0 6∈ B − A)
⇒ (U + A) ∩ B = ∅. By (a), there exists Λ ∈ X ∗ such that Λ(A + U ) and
Λ(B) are disjoint convex subsets of R with Λ(A + U ) open and to the left of
Λ(B). Since Λ(A) is a compact subset of Λ(A + U ), we get the result. 

Exercise 8. A closed convex set in a lctvs is the intersection of all the closed
half spaces containing it.

6. Complete metrizable tvs

Definition 6.1. A set A in a linear space X is said to be radial at the origin

if, for each x ∈ X, there exists rx > 0 such that 0 ≤ r ≤ rx ⇒ rx ∈ A. Some
authors call 0 a core point or internal point of A.

The basic result here is the following

Proposition 6.2. Let X be a complete metrizable tvs. Let A be a balanced,

closed set radial at 0. Then A + A is a neighborhood of 0. So, if A is convex
then A is a neighborhood of 0.

S∞
Proof. A radial at 0 and A balanced ⇒ X = n=1 nA.

A closed ⇒ nA closed. By Baire Category Theorem, some nA has non-

empty interior ⇒ A has non-empty interior ⇒ A + A (= A − A as A is
balanced) is a neighborhood of 0.
A convex ⇒ 21 A + 12 A ⊆ A. By applying the above argument to 12 A, we see
that A is a neighborhood of 0 if A is convex. 
 TOPOLOGICAL VECTOR SPACES 17

Theorem 6.3 (Equicontinuity Principle). Let X be a complete metrizable

tvs and Y a tvs. Suppose {Ti }i∈I is a family of continuous linear transforma-
tions from X into Y and suppose that, for each x ∈ X, the set {Ti x : i ∈ I}
is a bounded set in Y . Then limx→0 Ti x = 0 uniformly for i ∈ I.

Proof. Let V be a neighborhood of 0 in Y . It suffices to show that

T −1
i∈I Ti (V ) is a neighborhood of 0 in X. Choose a closed balanced neigh-
borhood U of 0 in Y such that U + U ⊆ V . Let A = i∈I Ti−1 (U ). Then
T

A is balanced (as Ti is linear for each i), closed (as each Ti is continuous
and U is closed) and radial at 0 because if x ∈ X, there exists rx > 0
such that rx (Ti x) ∈ U (for all i) by the boundedness of {Ti x : i ∈ I}
in Y . So by last Proposition, A + A is a neighborhood of 0 and clearly
A + A ⊆ i∈I Ti−1 (V ).
T


Remark 6.4. The above result is often called the Principle of Uniform
Boundedness when X, Y are, respectively, Banach and normed linear spaces.
Specifically, in this context, the theorem reads:
If {Ti }i∈I is a family of continuous linear transformations from X to Y such
that for each x ∈ X, sup{kTi xkY : i ∈ I} < ∞ then there exists M > 0 such
that kTi xkY ≤ M kxkX (for all i and x ∈ X), hence supi kTi k < ∞.

Theorem 6.5 (Open Mapping Theorem). Let X, Y be two complete metriz-

able TVS and let T be a linear continuous transformation from X onto Y .
Then T is an open mapping.

Proof. It suffices to prove that if V is a neighborhood of 0 in X then T (V )

is a neighborhood of 0 in Y . Let d be a invariant metric compatible with the
topology of X. Define Vn = {x : d(0, x) < r · 2−n }, (n = 0, 1, 2, . . .) where
r > 0 is so small that V0 ⊆ V . We will prove that some neighborhood W of
0 in Y satisfies W ⊆ T (V1 ) ⊆ T (V ). Since V1 ⊇ V2 + V2 and T is continuous,
we get
T (V1 ) ⊇ T (V2 ) + T (V2 ) ⊇ T (V2 ) + T (V2 ).
But T (V2 ) is closed, balanced and radial at 0 (the latter because V2 is radial
at 0 and T is onto), we have by our first proposition that T (V2 ) + T (V2 )
contains a neighborhood W of 0 in Y .
Now to prove that T (V1 ) ⊆ T (V ). Fix y1 ∈ T (V1 ). Assume n ≥ 1 and yn
has been chosen in T (Vn ). By what was just proved, T (Vn+1 ) contains a
neighborhood of 0 in Y . Hence [yn − T (Vn+1 )] ∩ T (Vn ) 6= ∅ ⇒ there exists
18 BANDYOPADHYAY

xn ∈ Vn such that T xn ∈ yn − T (Vn+1 ). Put yn+1 = yn − T xn ∈ T (Vn+1 )

and we can continue the construction. Since d(0, xn ) < 2rn , it is easy to see
P
that the sums x1 + . . . + xn form a Cauchy sequence and hence xn = x
exists with d(0, x) < r ⇒ x ∈ V0 ⊆ V . Moreover,
m
X m
X
T xn = (yn − yn+1 ) = y1 − ym+1
n=1 n=1

and as ym+1 → 0 as m → ∞ (this is because by the continuity of T ,

{T (Vn )}∞
n=1 forms a local base at 0 in Y ), we conclude that y1 = T x ∈ T (V ).
This completes the proof. 

Corollary 6.6. Any continuous one-to-one linear map of one complete

metrizable tvs onto another is a homeomorphism.

Exercise 9. Let X be a linear space which is complete under two norms

k · k1 and k · k2 and suppose there exists K > 0 such that kxk1 ≤ Kkxk2 for
all x ∈ X. Then there exists L > 0 such that Lkxk2 ≤ kxk1 ≤ Kkxk2 for
all x ∈ X, i.e. k · k1 and k · k1 are equivalent.

Theorem 6.7 (Closed Graph Theorem). Let X, Y be complete metrizable

tvs and let T be a linear transformation from X to Y . Then T is continuous
if and only if the graph GT = {(x, T x) : x ∈ X} ⊆ X × Y is closed in X × Y .

Proof. Clearly, T continuous ⇒ GT closed.

Suppose now that GT is closed. Consider X × Y with the product metric:

d((x, y), (x0 , y 0 )) = d1 (x, x0 ) + d2 (y, y 0 )

where d1 , d2 are the metrics inducing the topologies of X, Y respectively.

GT is a closed subspace of X × Y . Let P : GT → X be the projective map,
i.e. P {(x, T x)} = x. P is continuous linear, 1-1 and onto X. Hence by the
above Corollary, P is a homeomorphism, i.e. P −1 is continuous which says:
xn → x ⇒ T xn → T x, i.e. T is continuous. 

Exercise 10. Suppose Y is complete metrizable tvs such that the contin-
uous functionals on Y separate points of Y , i.e. given y1 6= y2 there exists
f ∈ Y ∗ such that f (y1 ) 6= f (y2 ). If X is a complete tvs then T : X → Y ,
(T assume to be linear) is continuous if and only if f ◦ T is continuous for
all f ∈ Y ∗ .
 TOPOLOGICAL VECTOR SPACES 19
P
Exercise 11. Prove that @ a sequence {λn } ⊆ C such that an converges
absolutely if and only if {λn an } is bounded.
[Hint: Assume (λn ) exists with λn 6= 0 for all n. Define T : `∞ → `1 by
T [(cn )] = {cn /λn }. Then kT k ≤ | λ1n | and
P P 1
| λn | < ∞ by the hypothesis.
Hence T is continuous, linear, one-one, onto ⇒ `∞ and `1 are homeomor-
phic which is impossible as `∞ is non-separable and `1 is separable in their
respective norm topologies.]

7. Weak and Weak* Topologies

Lemma 7.1. Let X be a linear space. Let F be a family of linear functionals

on X which is total, i.e. f (x) = 0 for all f ∈ F ⇒ x = 0 (equivalently: F
is separating). Let J be the weakest topology on X relative to which every
f ∈ F is continuous. Then (X, J ) is a lctvs and every continuous linear
functional on (X, J ) is a linear combination of functionals in F.

Proof. A topology on X will make each f ∈ F continuous if and only if

every set Uf,x,ε = {y ∈ X : |f (y) − f (x)| < ε} is open. The collection of all
such sets is translation-invariant and J is a translation invariant topology.
Thus it is sufficient to look at the topology at the origin where a local base
looks like {x : |fj (x)| < εj , j = 1, 2, . . . , n}. Such sets are convex, balanced
and absorbing. Therefore, by Theorem 1.4, (X, J ) is locally convex.
Suppose f is a linear functional on X which is J -continuous. Then there
exists a J -neighborhood of 0 on which f is bounded, i.e. there exists
f1 , . . . , fn ∈ F, ε > 0 and M > 0 such that
|fj (x)| < ε ∀j = 1, . . . , n ⇒ |f (x)| ≤ M.
Therefore, there exists c > 0 such that |f (x)| ≤ c max1≤j≤n |fj (x)| for any
x ∈ X. In particular, ker(f ) ⊇ nj=1 ker(fj ) ⇒ f = nk=1 αk fk for some
T P

α1 , . . . , αn ∈ K. 

Exercise 12. Prove the last statement in the above proof.

Remark 7.2. If F is a linear subspace (of the algebraic dual of X), it

follows from the lemma that the (continuous) dual of (X, J ) is F itself.

Definition 7.3. Suppose that (X, J ) a lctvs. We know that X ∗ separates

the points of X. Taking F = X ∗ , the lc topology Jw that we get in the above
lemma, is called the weak topology of X. Note that Jw ⊆ J . Moreover, a
20 BANDYOPADHYAY

typical neighborhood of 0 in Jw looks like {x ∈ X : |fk (x)| < ε, k =

1, . . . , n}, fk ∈ X ∗ .

Exercise 13. Let X be a lctvs and E ⊆ X is a convex set. Then the closure
of E in the weak and the original topology are the same.

Definition 7.4. Suppose that X is a lctvs For each x ∈ X, define the linear
functional x̂ on X ∗ by x̂(x∗ ) = x∗ (x). Note that {x̂ : x ∈ X} is a linear space
and that it separates the points of X ∗ . We are now in the situation described
by the lemma with X replaced by X ∗ and F replaced by {x̂ : x ∈ X}. The
corresponding topology is called the weak* topology on X ∗ . A basic w*-
neighborhood of 0 is then {x∗ : |x∗ (xk )| < ε, xk ∈ X, k = 1, . . . , n}.

The most important result concerning w*-topology is the following

Theorem 7.5 (Banach-Alaoglu). Let X be a lctvs and suppose that U is a

neighborhood of 0 in X. Then U ◦ = {f ∈ X ∗ : |f (x)| ≤ 1 for all x ∈ U } is
compact in the w*-topology on X ∗ .

Proof. First observe that if U ◦◦ is defined by {x ∈ X : |f (x)| ≤ 1 for all f ∈

U ◦ } then U ◦◦ = co(U ) and that U ◦ = (U ◦◦ )◦ (both are simple consequences
of the 2nd form of the separation theorem). Since U ◦◦ is convex, balanced
and closed, we may without loss of generality assume that the neighborhood
U is closed, convex and balanced to start with.
Let p ≡ pU be the associated Minkowski functional of U . Now, U = {x :
p(x) ≤ 1}, hence U ◦ = {f ∈ X ∗ : |f (x)| ≤ p(x) for all x ∈ X}. p being a
continuous seminorm, any linear functional g on X such that |g| ≤ p on X
is automatically continuous on X.
Q
For each x ∈ X, let Dx = {α ∈ C : |α| ≤ p(x)}. Look at Π = x∈X Dx . Π
is compact (Tychonoff). Identifying f with (f (x))x∈X , we see that U ◦ ⊆ Π.
Now Π is the set of all functions F : X → C such that |F (x)| ≤ p(x) for all
x and U ◦ = {F ∈ Π : F linear }. The topology on Π is the product topol-
ogy, i.e. the weakest topology defined by the coordinate projections. One
projection for each x ∈ X : x̂(F ) = F (x). Therefore, the weak∗ topology on
U ◦ is its relative topology in Π. But U ◦ is closed in Π:
\
U ◦ = {F ∈ Π : F linear} = \
ker[(αx + βy) − αb
x − βb
y ].
α∈C,β∈C
x,y∈X


 TOPOLOGICAL VECTOR SPACES 21

Corollary 7.6. Let X be a normed linear space. Then the unit ball of X ∗
is weak* compact.

Exercise 14. If X is a separable normed linear space. Show that the w*-
topology of the unit ball of X ∗ is metrizable.

8. Extremal points of compact convex sets

Definition 8.1. Let C be a convex set in a linear space X.

(a) A point x ∈ C is called an extreme point of C if x is not an interior
point of any line segment in C. That is, if y, z ∈ C, λ ∈ (0, 1) and
x = λy + (1 − λ)z, then x = y = z.
We will denote by ext(C) the set of extreme points of C.
(b) A convex set F ⊆ C is called a face of C if no point of F is an
internal point of a line segment whose endpoints are in C but not in
F . That is, x, y ∈ C, λ ∈ (0, 1) and λx + (1 − λ)y ∈ F ⇒ x, y ∈ F .
Thus, x ∈ ext(C) if and only if {x} is a singleton face of C.
If F has the above property, but is not convex, we call it an extremal set.

Example 8.2. Let K be a compact Hausdorff space and let P be the convex
set of regular probability measures on K, so P ⊆ X where X = C(K)∗ .
(i) The discrete measures form a face of P (a measure µ in P is
discrete if µ = ∞
P P
n=1 αn δxn , n αn = 1, αn ≥ 0, xn ∈ K).
(ii) The continuous measures (i.e., measures without any point
masses) from a face of P .
(iii) If m ∈ P , then {µ ∈ P : µ  m} is a face of P .
(iv) If m ∈ P , then {µ ∈ P : µ ⊥ m} is a face of P .
(v) The extreme points of P are the point masses {δx : x ∈ K} and
conversely.

Proof. Any δx is clearly extreme. If µ ∈ P and if the (topological)

support S(µ) of µ has two points x1 , x2 find an open U such that
µ|U ∩S(µ)
x1 ∈ U , x2 6∈ U . Now 0 < µ[U ∩ S(µ)] < 1. Let µ1 = ,
µ(U ∩ S(µ))
µ|S(µ)\U
µ2 = , then µ1 , µ2 ∈ P , µ1 6= m2 and
µ(S(µ) \ U )
µ = µ(U ∩ S(µ))µ1 + [1 − µ(U ∩ S(µ))]µ2 .


22 BANDYOPADHYAY

(vi) Using (v), show that the extreme points of the unit ball of
C(K)∗ = M (K) are exactly {λδx : λ ∈ C, |λ| = 1, x ∈ K} = B.
(vii) Consider a subspace A of C(K). Show that the extreme points
of the unit ball of A∗ are contained in B and find an example to
show that all points in B may not be extreme.
(The proof of (vii) follows easily from the Krein-Milman theorem)

Theorem 8.3 (Krein-Milman Theorem). Suppose X is a topological vector

space such that X ∗ separates points of X. If K is a compact convex set in
X, then K is the closed convex hull of its extreme points, i.e.,
K = co(ext(K)).

Proof. Step I : If K is a compact convex set in X, then ext(K) 6= ∅.

Let P be the collection of all closed faces of K. Clearly, K ∈ P.
Claim : If S ∈ P, f ∈ X ∗ and
Sf = {x ∈ S : f (x) = sup f (S)},
then Sf ∈ P.
To prove the claim, suppose z = λx + (1 − λ)y ∈ Sf , x, y ∈ K, λ ∈
(0, 1). Since z ∈ S and S ∈ P, we have x, y ∈ S. Hence f (x) ≤ sup f (S),
f (y) ≤ sup f (S). Since f (z) = λf (x) + (1 − λ)f (y) = sup f (S), we conclude
f (x) = f (y) = sup f (S), i.e., x, y ∈ Sf .
Order P by reverse inclusion, i.e., say K1 ≤ K2 if K2 ⊆ K1 . Let C be
chain in P. Since C is a collection of compact sets having finite intersection
property, the intersection M of all members of C is nonempty. It is easy
to see that M ∈ P. Thus, every chain in P has an upper bound. By
Zorn’s Lemma, P has maximal element, M0 . The maximality implies that
no proper subset of M0 belongs to P. It now follows from the claim that
every f ∈ X ∗ is constant on M0 . Since X ∗ separates points of X, M0 must
be a singleton. Therefore M0 is an extreme point of K.
Step II : If K is a compact convex set in X, then K = co(ext(K)).
Let H = co(ext(K)). If possible, let x0 ∈ K \ H. By Separation Theorem,
there is an f ∈ X ∗ such that f (x) < f (x0 ) for every x ∈ H. If Kf = {x ∈
K : f (x) = sup f (K)}, then, by the claim, Kf ∈ P and Kf ∩ H = ∅. By
Step I, ext(Kf ) 6= ∅. But since Kf is a face of K, ext(Kf ) ⊆ ext(K). This
contradiction completes the proof. 
 TOPOLOGICAL VECTOR SPACES 23

9. Integral Representations

Suppose K is a non-empty compact subset of a lctvs X and µ is a regular

probability measure on K. A point x ∈ X (if it exists!) is said to be
represented by µ if f (x) = µ(f ) = f dµ for all f ∈ X ∗ . Also, one says that
R

x is the resultant of µ and writes x = r(µ).

Proposition 9.1. Suppose C ⊆ X is compact and assume further that
K = co(C) is compact. If µ ∈ P (C) then there exists a unique point x ∈ X
which is represented by µ.

Proof. By K-M theorem, there exists a net µα of the form

nα  
(α)
X X (α) (α)
µα = ai δy(α) aαi ≥ 0, ai = 1, yi ∈ C .
i
i=1
P α (α) (α)
With µα → µ in the w*-topology (on M (C)). But ni=1 ai yi ∈ K and
P (β) (β)
since K is compact, there exists a subset ai yi , say, converging to a
P (β) (β)
point x ∈ K. Consequently, f ( ai yi ) → f (x) for all f ∈ X ∗ and thus
µ(f ) = limβ µβ (f ) = f (x) for all f ∈ X ∗ . The uniqueness of r(µ) is a direct
consequence of the fact that X ∗ separates the points of X. 
Proposition 9.2. Suppose C ⊆ X is compact and that K = co(C) is com-
pact as before. Then x ∈ K if and only if there exists µ ∈ M1+ (C), the
regular probability measures on C, which represents x.

Proof. Suppose µ ∈ M1+ (C). Then by the last result, r(µ) = x ∈ co(C).
Now, let x ∈ co(C). Then x is approximable by elements of the form
yα = ni=1
P α α α α P α
λi yi (αi ≥ 0, λi = 1, yiα ∈ C). Consider the corresponding
λα δyiα ∈ M1+ (C), r(µα ) = yα . There exists a subnet
P i
measures µα =
µβ → µ ∈ M1+ (C), hence for all f ∈ X ∗ , limβ f (yβ ) = limβ µβ (f ) = µ(f ).
But yβ → x (as yα → x), and we have µ(f ) = f (x) for all f ∈ X ∗ . 
Remark 9.3. It now follows from the above result that the following state-
ments are equivalent:
(a) If K ⊆ X is a compact convex set, then K = co(ext(K)).
(b) Each x ∈ K is the resultant of a µ ∈ M1+ (K) with µ supported
by ext(K).

Stat–Math Division, Indian Statistical Institute, 202, B. T. Road, Kolkata

700108, India, E-mail : [email protected]