Weak topologies

David Lecomte
May 23, 2006

1 Preliminaries from general topology

In this section, we are given a set X, a collection of topological spaces (Yi )i∈I and a
collection of maps (fi )i∈I such that each fi maps X into Yi .
We wish to define a topology on X that makes all the fi ’s continuous. And we want to
do this in the cheapest way, that is: there should be no more open sets in X than required
for this purpose.
Obviously, all the fi−1 (Oi ), where Oi is an open set in Yi should be open in X. Then
finite intersections of those should also be open. And then any union of finite intersections
should be open. By this process, we have created as few open sets as required. Yet it is
not clear that the collection obtained is closed under finite intersections. It actually is, as
a consequence of the following lemma:

Lemma 1
Let X be a set and let O ⊂ P(X) be a collection of subsets of X, such that

• ∅ and X are in O;

• O is closed under finite intersections.

S
Then T = { O | O ⊂ O} is a topology on X.
O∈O

Proof: By definition, T contains X and ∅ since those were already in O. Furthermore,

T is closed under unions, again by definition. So all that’s left is to check that T is
closed under finite intersections.
Let A1 and A2 be two elements of T . Then there exist O1 and O2 , subsets of O, such
that
[ [
A1 = O and A2 = O
O∈O1 O∈O2

1
It is then easy to check by double inclusion that
[
A1 ∩ A2 = O1 ∩ O2
O1 ∈O1
O2 ∈O2

Letting O denote the collection {O1 ∩ O2 | O1 ∈ O1 O2 ∈ O2 }, which is a subset of O

since the latter is closed under finite intersections, we get
[
A1 ∩ A2 = O
O∈O

This set belongs to T . By induction, T is closed under finite intersections. 

Corollary 2
The collection of all unions of finite intersection of sets of the form fi−1 (Oi ) where i ∈ I
and Oi is an open set in Yi is a topology. It is
called the weak topology on X generated
by the (fi )i∈I ’s and we denote it by σ X, (fi )i∈I .
By definition, the functions (fi )i∈I are continuous for this topology.

Of course, a topology is useless if it is too complicated for us to deal with it. It turns out
there is a nice characterization of converging sequences, and continuous function from a
topological space into a weakly topologized space.

Theorem 3

Let (xn )n∈N be a sequence in X. It converges in the topology σ X, (fi )i∈I to some x ∈ X
if and only if
∀i ∈ I lim fi (xn ) = fi (x)
n→∞

Remember that in a topological space (X, T ),

• a sequence (xn )n∈N converges to x ∈ X if and only if

∀O ∈ T containing x ∃N ∈ N ∀n > N xn ∈ O

Notice also that if the topology T is not Hausdorff, there is no unicity for the
limit of a sequence. For example, if x and y are distinct points that cannot
be separated by open sets, the constant sequence equal to x will converge
both to x and y.

• a function f : X −→ Y, where Y is a topological space, is continuous at a

point x if and only if the preimage by f of any open set (in Y) containing
f (x) is in T .

2
 It is easy to show that in
such a case, for every sequence (xn )n∈N converging
to x, the sequence f (xn ) n∈N converges to f (x). But this property does not
characterize continuity at x in general!!! We need more assumptions on X, one
of them being for example that x has a countable basis of neighbourhoods.

So we see that there are bad habits, inherited from working with normed or
metric spaces all the time, that we have to lose. Here are more of them:

• If X is compact for the topology T , there is no reason to believe that any

sequence has a converging subsequence.
What can be shown is the following: if (xn )n∈N is a sequence in X compact,
there exists a point x such that any open set containing x contains infinitely
many terms of the sequence (xn )n∈N . However, this does not allow us to build
a subsequence of (xn )n∈N converging to x because there could be too many
open neighbourhoods of x.
However, if x has a countable basis of neighbourhoods, we can build such a
subsequence.

• If A is any subset of X, the closure of A cannot be characterized anymore as

the set of limit points of A. It is true that limit points of A are in A, but A
could contain strictly the set of limit points of A.
The only characterization of points x ∈ A is that every open set containing x
meets A.

Proof: Suppose first that the sequence converges in the weak topology
to some x ∈ X.
Since for every j ∈ I, the function fj is continuous for σ X, (fi )i∈I , we have
lim fj (xn ) = fj (x)
n→∞

Conversely, suppose that there exists x in X such that

∀i ∈ I lim fi (xn ) exists and equals fi (x)
n→∞

Let O be any open set containing x. By definition, there exist a finite subset J of I, and
open sets (Oj )j∈J such that Oj ⊂ Yj for all j ∈ J, such that
n
\
x∈ fi−1
j
(Oj )
j=1

which means that ∀j ∈ J fj (x) ∈ Oj

3


Given j ∈ J, we know that the sequence fj (xn ) n∈N
converges to fj (x). Then, since Oj
contains fj (x), there exists Nj ∈ N such that
∀n > Nj fj (xn ) ∈ Oj
Letting N = Max Nj , we have
j∈J

∀n > N ∀j ∈ J fj (xn ) ∈ Oj
\
In other words, ∀n > N xn ∈ fj−1 (Oj ) ⊂ O
j∈J


So (xn )n∈N converges to x for the topology σ X, (fi )i∈I . 

Theorem 4
Let (Z, T ) be a topological space,

and ϕ : Z −→ X be map. Then ϕ is continuous for the
topologies T and σ X, (fi )i∈I if and only if for every i ∈ I, fi ◦ ϕ is continuous.

Proof: Easy, try it. 

An example of a weak topology, aside from the ones we will present in those note, is
the topology of pointwise convergence. It is defined as follows: let A be any set and let
X be the set of functions A −→ R. For every a in A, define the function ϕa : X 7−→ R by
∀f ∈ X ϕa (f ) = f (a)


The topology of pointwise convergence is σ X, (fa )a∈A . In this topology, a sequence of
functions converges if and only if it converges pointwise, in view of Theorem 3. One can
show that this topology is not metrizable, this is the topic of a problem in the fall 2003
qualifying exam.

2 The topology σ(X, X⋆ )

In this section, X is a normed space. Unless stated otherwise, we do not assume that it
is complete.


Definition 5 The weak topology on X is the topology σ X, (f )f ∈X⋆ . For convenience, it
is simply denoted σ(X, X⋆ ).

A first thing we want to check is that the weak topology on X is Hausdorff, which will
guarantee us the unicity of limits.

Theorem 6
The topology σ(X, X⋆ ) is Hausdorff.

4
Proof: Let x and y be two distinct points in X. Since kx − yk > 0, there exists a positive
ǫ such that B(x, ǫ) does not contain y. Since B(x, ǫ) is convex open, we know that it can
be strictly separated from {y} by a hyperplane by the geometric form of the Hahn-Banach
theorem: there exist f ∈ X⋆ and α ∈ R such that
∀u ∈ B(x, ǫ) (f, u) < α < (f, y)

In particular, (f, x) < α < (f, y)

Therefore, x ∈ f −1 (−∞, α) and y ∈ f −1 (α, +∞)
Those two sets are weakly open, since they are preimages of open subsets of R by a linear
functional. And they are disjoint. Thus σ(X, X⋆ ) is Hausdorff. 
Then we prove a few easy facts comparing the weak topology and the norm (also called
strong) topology on X.

Proposition 7
1. The weak topology is weaker than the norm topology: every weakly open (resp.
closed) set is strongly open (resp. closed).

2. A sequence (xn )n∈N converges weakly to x if and only if

∀f ∈ X⋆ lim (f, xn ) = (f, x)

n→∞

3. A strongly converging sequence converges weakly.

4. If (xn )n∈N is a sequence in X converging weakly to x, then (xn )n∈N is bounded and

kxk 6 lim inf kxn k

n→∞

5. If (xn )n∈N is a sequence in X converging weakly to x and (fn )n∈N is a sequence in

X⋆ converging strongly to f , then

lim (fn , xn ) exists and equals (f, x)

n→∞

Proof: The first point is clear: the norm topology already makes all linear functionals
continuous. Since the weak topology is the weakest with this property, it is weaker than
the strong topolgy. So every weakly open set is strongly open, and by taking complements,
every weakly closed set is strongly closed.
The second point is just a restatement of Theorem 3 in the particular case of the
weak topology on X.

5
 The third point is clear as well: if (xn )n∈N converges strongly to x, then
   
∀f ∈ X⋆ (f, x) − (f, xn ) = (f, x − xn ) 6 kf k kx − xn k −−−−→ 0
n→∞

and therefore (xn )n∈N converges weakly to x by 2.

Let (xn )n∈N be a sequence in X converging weakly to x. We have already seen at least
once in quals that (xn )n∈N is bounded, as a consequence of the Banach-Steinhaus theorem
(see for example the spring 2002 exam). Also, if f is a bounded linear functional on X,
 
∀n ∈ N (f, xn ) 6 kf k kxn k
so by taking liminfs, we get
 
(f, x) 6 kf k lim inf kxn k
n→∞
 
(f, x)
Since kxk = Sup
f ∈X⋆ kf k
f 6=0

it follows that kxk 6 lim inf kxn k

n→∞

Finally, let (xn )n∈N be sequence in X converging weakly to x and (fn )n∈N be a sequence
in X⋆ converging strongly to f . We have for every integer n
     
(f, x) − (fn , xn ) = (f, x) − (f, xn ) + (f, xn ) − (fn , xn ) 6 (f, x − xn ) + kfn − f k kxn k
Since (xn )n∈N is bounded by 4, the righthandside tends to 0 and 5 is proved. 
The next step is to identify a basis of neighbourhoods for σ(X, X⋆ ).

Theorem 8
Let x0 ∈ X. A basis of neighbourhoods of x0 for the weak topology is given by the collection
of sets of the form
   
Wǫ,f1 ,...,fn = x ∈ X  ∀i ∈ {1, . . . , n} (fi , x) − (fi , x0 ) < ǫ

n ∈ N ǫ > 0 f1 , . . . , fn ∈ X⋆

Proof: Remember that a collection X of open sets containing x is a basis of neighbour-

hoods of x0 if and only if every open set containing x contains an element of X .
That every set Wǫ,f1 ,...,fn are weakly open is clear, since all f1 , . . . , fn are weakly
continuous. Furthermore, it contains x0 since
 
∀i ∈ {1, . . . , n} (fi , x0 ) − (fi , x0 ) = 0 < ǫ
Now, let O be any weakly open set containing x0 . By definition of the topology
σ(X, X⋆ ), it is a union of finite intersections of preimages of open sets in R by bounded

6
linear functionals. So there exist a finitely many bounded linear functionals f1 , . . . , fn
and open subsets O1 , . . . , On of R, such that
n
\
x0 ∈ fj−1 (Oj ) ⊂ O
j=1

Then for every j ∈ {1, . . . , n}, the real number (fj , x0 ) belongs to Oj . Since this set is
open, there is a positive ǫj such that


(fj , x0 ) − ǫj , (fj , x0 ) + ǫj ⊂ Oj

Let ǫ = Min ǫj > 0

16j6n

so that ∀j ∈ {1, . . . , n} (fj , x0 ) − ǫ, (fj , x0 ) + ǫ ⊂ Oj
Then, if x belongs to Wǫ,f1 ,...,fn , we have by definition of this set


∀j ∈ {1, . . . , n} (fj , x) ∈ (fj , x0 ) − ǫ, (fj , x0 ) + ǫ ⊂ Oj
n
\
which means that x∈ fj−1 (Oj ) ⊂ O
j=1

Thus Wǫ,f1 ,...,fn ⊂ O 

Now that we have two topologies on X, we may wonder if by any chance they coincide.
The answer is given by the following

Proposition 9
The weak topology and the strong topology on X coincide if and only if X is finite dimen-
sional.

Proof: Suppose first that X is finite dimensional. We already know that the weak
topology is included in the strong topology, through Proposition 7. All that is left to
show is the converse.
Since X is finite dimensional, it has a basis (e1 , . . . , en ). Any x ∈ X has a unique
decomposition along this basis, which means that
X n
n
∃!(x1 , . . . , xn ) ∈ R x= xi ei
i=1

Define then kxk∞ = Max |xi |

16i6n

It is easy to check that it is a norm on X.

7
 Remember that, as a consequence of the finite dimensionality of X, all norms are
equivalent and the strong topology on X is the topology defined by any norm. So if O is
any strong open set, it is in particular open for k k∞ . This means that for every x ∈ O,
there exists a positive ǫx such that
B∞ (x, ǫx ) ⊂ O
[
Therefore O= B∞ (x, ǫx )
x∈O

So if we show that any open ball is weakly open, we get that the strongly open set O is
weakly open, as union of weakly open sets.
Let x be any point in X and ǫ be any positive real number. Then
   
B∞ (x, ǫ) = y ∈ X  ky − xk∞ < ǫ = y ∈ X  ∀i ∈ {1, . . . , n} |yi − xi | < ǫ
But the functionals f1 , . . . , fn , defined by
X n
∀x = xj ej ∈ X (fi , x) = xi
j=1
⋆
are clearly in X . And we can then write
   
B∞ (x, ǫ) = y ∈ X  ∀i ∈ {1, . . . , n} (fi , y) − (fi , x) < ǫ
which proves that B∞ (x, ǫ) is weakly open, by Theorem 8.
So any strong open set is weakly open: the weak and strong topology on a finite
dimensional space coincide.
Now let’s suppose that X is infinite dimensional, and let’s show that the weak and
strong topologies do not coincide. Let S be the unit sphere in X:
 
S = x ∈ X  kxk = 1
Then S is strongly closed. But 0 belongs to the weak closure of S. Indeed, let O be any
weak neighbourhood of 0. By Theorem 8, there exist ǫ > 0 and f1 , . . . , fn in X⋆ such
that
  
W = x ∈ X  (fi , x) < ǫ ⊂ O

The map Φ : X −→ Rn

x 7−→ (f1 , x), . . . , fn , x)
is linear and
n
  \
Ker Φ = x ∈ X  (fi , x) = 0 ∀1 6 i 6 n = Ker fi
i=1

8
By the rank-nullity theorem,
Dim Ker Φ + Dim Im Φ = Dim X = ∞
Since Dim Im Φ 6 n
it follows that Ker Φ is infinite dimensional, and can certainly not be equal to {0}. So
there exists x 6= 0 such that
∀i ∈ {1, . . . , n} (fi , x) = 0
 
Then ∀λ ∈ R ∀i ∈ {1, . . . , n} (fi , λx) = 0 < ǫ

which proves that ∀λ ∈ R λx ∈ W ⊂ O

1
Now, taking λ = kxk shows that O intersect S. So any weakly open neighbourhood
of 0 intersects S: 0 is in the weak closure of S. Thus the weak and strong closure of S
are different. 
Notice how in the proof we showed that weakly open sets are very big: they
at least contain lines when the dimension of X is infinite.
Using exactly the same strategy, one can show that the weak closure of S
contains the closed unit ball B of X. And the next theorem will establish that B
is weakly closed. So we get that in an infinite dimensional normed space,
σ(X,X⋆ )
S =B

The next theorem answers the important question: if the weak topology is strictly
weaker than the strong topology (in infinite dimension), are there sets for which we can
guarantee that strongly closed implies weakly closed?

Theorem 10
Let C be a nonempty convex set in X. Then C is strongly closed if and only if it is weakly
closed.

Proof: Of course, by Proposition 7, if C is weakly closed, it is strongly closed. We are

more interested in the converse.
σ(X,X⋆ )
Suppose that C is strongly closed. We have C ⊂ C and we want to show the
converse inclusion. So let x be in the complement of C. By the geometric form of the
Hahn-Banach theorem, C (convex strongly closed) and {x} (convex strongly compact)
can be strictly separated by a hyperplane: there exist f ∈ X⋆ and α ∈ R such that
∀y ∈ C (f, y) < α < (f, x)


Thus the weakly open set f −1 (α, +∞) contains x and does not intersect C. So x is not
in the weak closure of C. Which achieves the proof. 
Now we worry about bounded linear maps.

9
Theorem 11
Let X and Y be two Banach spaces and let T be a linear map between X and Y. Then T
is continuous strong-strong if and only if it is continuous weak-weak.

Proof: The fact that continuous strong-stong implies continuous weak-weak is a conse-
quence of Theorem 4. The converse (or something very close) has already been studied
in the 2003 spring qualifying exam, and follows at once from the closed graph theorem.
And to finish this section, we wonder about whether the weak topology is metrizable.
We will first need a lemma from linear algebra:

Lemma 12
Let f1 , . . . , fn , f be linear functionals on a vector space X. Then f is a linear combination
of f1 , . . . , fn if and only if
\n
Ker fi ⊂ Ker f
i=1

Proof: If f is a linear combination of f1 , . . . , fn , we can write

Xn
f= αi fi α1 , . . . , αn ∈ R
i=1
n
T
and it clear from here that Ker fi ⊂ Ker f .
i=1
Conversely, suppose that this inclusion holds. Assume also, for now, that f1 , . . . , fn
are linearly independent. Define


∀x ∈ X Φ(x) = (f1 , x), . . . , (fn , x), (f, x)
Then (0, . . . , 0, 1) cannot be in the image of Φ, since if (fi , x) = 0 for 1 6 i 6 n, then
(f, x) = 0 as well. So Im Φ is a strict subspace of Rn+1 and there exists a nonzero linear
functional on Rn+1 that vanishes on Im Φ:
n
X
∃(α1 , . . . , αn , α) ∈ Rn+1 \ {0} ∀(x1 , . . . , xn , x) ∈ Im Φ αx + αi xi = 0
i=1

n
X
Then ∀x ∈ X α(f, x) + αi (fi , x) = 0
i=1

 n
X 
and ∀x ∈ X αf + αi fi , x = 0
i=1

n
X
which implies that αf + αi fi = 0
i=1

10
Now, α cannot be 0 since f1 , . . . , fn are linearly independent. Thus we can divide by α
and f is a linear combination of f1 , . . . , fn .
If we don’t assume that f1 , . . . , fn are linearly independent, up to renaming, we can
suppose that f1 , . . . , fp are linearly independent and that fp+1 , . . . , fn are linear combina-
tions of those. Then by the first part of the proposition,
p n
\ \
Ker fi = Ker fi ⊂ Ker f
i=1 i=1

From what we just proved, f is a linear combination of f1 , . . . , fp , which is fair enough.

Corollary 13
The weak topology on a normed space X is metrizable if and only if X is finite dimensional.

Proof: From Proposition 9, if X is finite dimensional, the weak topology is the norm
topology and therefore comes from a metric.
Conversely, suppose that there is a metric d on X, such that the induced topolgy
coincides with the weak topology. For every positive integer k, define
n  1o

Bk = x ∈ X  d(x, 0) <
k
Bk is open for d and thus is a weak open neighbourhood of 0. By Theorem 8, there exist
a finite collection Fk of bounded linear functionals and a positive ǫk such that
  
Wk = x ∈ X  (f, x) < ǫk ∀f ∈ Fk ⊂ Bk
[
Let F= Fk
k>1

which is a countable subset of X . We show that F spans X⋆ . This implies (see the spring
⋆

2004 qual) that X⋆ (complete) is finite dimensional. Thus X⋆⋆ is finite dimensional as
well, and therefore X follows the herd since it injects in X⋆⋆ .
So let g be any bounded linear functional on X and consider
  
W = x ∈ X  (g, x) < 1

T there exists a k big enough so that Bk ⊂ W and

This is a weak neighbourhood of 0. So
as a consequence Wk ⊂ W. Let x ∈ Ker f . Then
f ∈Fk
 
∀λ ∈ R ∀f ∈ Fk (f, λx) = 0 < ǫk

Hence ∀λ ∈ R λx ∈ Wk ⊂ W
 
which means that ∀λ ∈ R |λ| (g, x) < 1

11
Necessarily (g, x) = 0 that is x ∈ Ker g
According to Lemma 12, g ∈ Span Fk . Which achieves showing that
X⋆ = Span F 

3 The topology σ(X⋆, X)

Remember that any element of X can be seen as a bounded linear function on X⋆ , through
evaluation at x:
∀f ∈ X⋆ (x, f ) = (f, x)

Definition 14 The weak⋆ topology on X⋆ is the topology σ(X⋆ , (x)x∈X ). For convenience,
it is simply noted σ(X⋆ , X).

We rapidly check the usual properties of this new topology:

Theorem 15
The topology σ(X⋆ , X) is Hausdorff.

Proof: This is easier than Theorem 6, since it does not even involve the Hahn-Banach
theorem. We let f and g be distinct elements of X⋆ . Thus there exists x ∈ X such that
(f, x) 6= (g, x)
Assuming, for example, that (f, x) < (g, x), we can find a real number α such that
(f, x) < α < (g, x)



so that f ∈ x−1 (−∞, α) and g ∈ x−1 (α, +∞)
Those are two disjoint weak⋆ open sets that separate f and g. 

Proposition 16
1. The weak⋆ topology on X⋆ is weaker than the weak topology σ(X⋆ , X⋆⋆ ), itself weaker
than the norm topology.
2. A sequence (fn )n∈N in X⋆ is weak⋆ convergent to f if and only if
∀x ∈ X lim (fn , x) = (f, x)
n→∞

3. A strongly converging sequence in X⋆ is weak⋆ convergent.

4. If (fn )n∈N is weak⋆ convergent to f , then (fn )n∈N is bounded and
kf k 6 lim inf kfn k
n→∞

12
 5. If (fn )n∈N is a sequence in X⋆ , weak⋆ convergent to f , and if (xn )n∈N is a sequence
in X converging strongly to x, then

lim (fn , xn ) exists and equals (f, x)

n→∞

Proof: Same as Proposition 7. 

Theorem 17
Let f0 ∈ X⋆ . A basis of neighbourhoods of f0 for the weak⋆ topology is given by the
collection of sets of the form
   
Wǫ,x1,...,xn = f ∈ X⋆  ∀i ∈ {1, . . . , n} (f, xi ) − (f0 , xi ) < ǫ

n∈N ǫ > 0 x1 , . . . , xn ∈ X

Proof: Same as Theorem 8. 

Now, at this point, we might wonder why in the world someone would be that
obsessed with weakening topologies.
The basic answer is that, if there are less open sets, it is easier to extract finite
subcovers from open covers. So we are hoping to get more compact sets.
And indeed, the Banach-Alaoglu-Bourbaki theorem will show that the closed
unit ball in X⋆ actually is compact for the weaker σ(X⋆ , X). This is, for example,
the starting point of the (really neat) Gelfand theory of Banach algebras. See
Professor Katznelson’s book for that.
Aside from that very specific example, compactness is generally a good thing
because it allows us to show that things exist: minimums or maximums of contin-
uous functions, fixed points, and other stuff such as converging subsequences (even
though compactness alone is not enough for that).

Proposition 18
Let ϕ ∈ X⋆⋆ and suppose that ϕ is weak⋆ continuous. Then
∃x ∈ X ∀f ∈ X⋆ (ϕ, f ) = (f, x)

Proof: Since ϕ is weak⋆ continuous, the set

  
V = f ∈ X⋆  (ϕ, f ) < 1
is weak⋆ open and contains 0. According to Theorem 17, there exist x1 , . . . , xn in X and
a positive ǫ such that
  
W = f ∈ X⋆  (f, xi ) < ǫ ∀1 6 i 6 n ⊂ V

13
 n
T
If f ∈ Ker xi , then
i=1

∀λ ∈ R ∀1 6 i 6 n (λf, xi ) = 0 < ǫ

so that ∀λ ∈ R λf ∈ W ⊂ V
 
Thus ∀λ ∈ R |λ| (ϕ, f ) < 1

Necessarily, f ∈ Ker ϕ
According to Lemma 12, ϕ is a linear combination of x1 , . . . , xn . 

Proposition 19
Let H be a hyperplane in X⋆ and suppose that H is closed for the weak⋆ topology. Then
there exist x ∈ X and α ∈ R such that

H = f ∈ X⋆ | (f, x) = α

Proof: H is a hyperplane, that is there exist ξ ∈ X⋆⋆ and α ∈ R such that


H = f ∈ X⋆ | (ξ, f ) = α
By assumption, H is weak⋆ closed. So if f0 ∈ X⋆ \ H, there exists a weak⋆ open set
containing f0 that does not intersect H. So there are x1 , . . . , xn in X and ǫ > 0 such that
  
W∩H= ∅ where W = f ∈ X⋆ | (f, xi ) − (f0 , xi ) < ǫ for all i
We also have (ξ, f0 ) 6= α since f0 ∈
/ H so either (ξ, f0) < α or (ξ, f0 ) > α. Assume we are
in the first case.
Suppose that for some f in W, we have (ξ, f ) > α. Then


ϕ : t 7−→ ξ, tf + (1 − t)f0
is continuous on [0, 1], such that
ϕ(0) = (ξ, f0 ) < α and ϕ(1) = (ξ, f ) > α
So by the intermediate value theorem, there is f ′ in [f0 , f ] which belongs to H. But that
segment is contained in W (this set is convex, it is easy to check) so W intersects H: this
is absurd. Therefore,
∀f ∈ W (ξ, f ) < α

Now, consider W − f0 = {f
 − f0 | f ∈ W}  
= f − f0 | f ∈ X⋆ (f, xi ) − (f0 , xi ) < ǫ for all i
  
W − f0 = f ∈ X⋆ | (f, xi ) < ǫ for all i

14
This is a weak⋆ neighbourhood of 0. Also,
∀f ∈ W − f0 (ξ, f ) = (ξ, f + f0 ) − (ξ, f0 ) < α − (ξ, f0)
| {z }
∈W

Since −f belongs as well to W − f0 , we have

 
∀f ∈ W − f0 (ξ, f ) < α − (ξ, f0 )
If we had supposed that (ξ, f0 ) > α, then we would have obtained that
 
∀f ∈ W − f0 (ξ, f ) < (ξ, f0 ) − α
So we showed that, for any choice of f0 in X⋆ \ H, there exists a weak⋆ neighbourhood of

0 contained in the pre-image under ξ of the open interval − |(ξ, f0) − α|, |(ξ, f0) − α| .
It follows that ξ is weak⋆ continuous at 0 and by linearity, it is weak⋆ continuous.
Proposition 18 then tells us that ξ is actually in X. Which achieves the proof. 
Those two propositions are satisfying, in the sense that the weak⋆ topology allows us
to separate X from X⋆⋆ : the only linear functionals that are weak⋆ continuous are elements
of X. Similarly, the only weak⋆ closed hyperplanes are the ones induced by elements of X.
Thus, we see that if the space X is not reflexive, then σ(X⋆ , X) is strictly included in
σ(X⋆ , X⋆⋆ ): there are hyperplanes closed for the latter that are not closed for the former,
namely the ones induced by linear functionals on X⋆ that are not in X.
Finally, here comes the Banach-Alaoglu-Bourbaki theorem:

Theorem 20
The closed unit ball in X⋆ is weak⋆ compact.

Proof: We first check that BX⋆ is closed for the weak⋆ topology. Let f0 be in the weak⋆
closure of this set. Let ǫ be any positive real number. By definition of kf0 k, there exists
x ∈ X, with norm 1, such that
(f0 , x) > kf0 k − ǫ
The weak⋆ open neighbourhood
  
W = f ∈ X⋆ | (f, x) − (f0 , x) < ǫ
has to intersect BX⋆ non-trivially: there exists f ∈ X⋆ such that
 
kf k 6 1 and (f, x) − (f0 , x) < ǫ

In particular kf0 k − ǫ < (f0 , x) < (f, x) + ǫ 6 kf k kxk + ǫ 6 1 + ǫ

and it follows that kf0 k 6 1 + 2ǫ

This holds for every positive ǫ, so kf0 k 6 1 which proves our claim.

15
 Let Y be the space RX of all real-valued functions on X, together with the product
topology. That is, the weakest topology on Y that makes all the evaluations
ex : Y −→ R x∈X
ω 7−→ ex (ω) = ω(x)
continuous. Since every element of X⋆ is a function on X, we have an injection
J : X⋆ −→ Y
f 7−→ J(f )

where ∀x ∈ X J(f )(x) = (f, x)

We first check that J is continuous when X⋆ has the weak⋆ topology. According to
Theorem 4, this is the case by definition of the weak⋆ topology, since
∀x ∈ X ∀f ∈ X⋆ (ex ◦ J)(f ) = J(f )(x) = (f, x)
We also make sure that J−1 is continuous on J(X⋆ ). This is also a consequence of
Theorem 4 and the definition of the product topology:


∀x ∈ X ∀f ∈ J(X⋆ ) J−1 (f ), x = (f, x) = ex (f )
So J is a homeomorphism onto its image. Notice that
   
∀x ∈ X ∀f ∈ BX⋆ J(f )(x) = (f, x) 6 kf k kxk 6 kxk

Y 
so J B X⋆ ⊂ − kxk, kxk
x∈X

The righthandside is compact by Tychonoff’s theorem; the lefthandside is closed,

since J−1 is continuous and BX⋆ is weak⋆ closed. Therefore, J BX⋆ ) is compact. Again,
J−1 is continuous and the continuous image of a compact set is compact. So B X⋆
is compact for the weak⋆ topology. 
And now, we finally achieved something: the closed unit ball in X⋆ , which is certainly
not compact when X is infinite dimensional, has a topology that makes it compact. There
is also a neat immediate consequence of this theorem:

Corollary 21
If X is reflexive, the closed unit ball of X is weakly compact.

Proof: We identify X and X⋆⋆ , since X is reflexive. Then σ(X, X⋆ ) and σ(X⋆⋆ , X⋆ )
are the same topology on X. But in the latter, the closed unit ball of X is compact,
by Theorem 20. Simple as that. 
This is the “trivial” direction of Kakutani’s theorem, which actually asserts that the
converse is true: if the closed unit ball of X is weakly compact, then X is reflexive. This
will be proved in the next section.

16
4 Weak topologies, reflexivity and uniform convexity
4.1 Kakutani and consequences
Lemma 22
BX is weak⋆ dense in B X⋆⋆ .

Proof: We let ξ0 be in the unit ball in X⋆⋆ and we suppose that there is a weak⋆
neighbourhood of ξ0 that does not intersect BX : there are f1 , . . . , fn in X⋆ and a positive
ǫ such that
  
W ∩ BX = ∅ where W = ξ ∈ X⋆⋆ | (ξ, fi) − (ξ0 , fi ) < ǫ for all i
 
In other words, ∀x ∈ BX ∃i ∈ {1, . . . , n} (fi , x) − (ξ0 , fi ) > ǫ (1)
 
(f1 , x)
Φ(x) =  ... 
 
Define ∀x ∈ X
(fn , x)
and let’s use k k∞ on Rn . Then (1) tells us precisely that
 
(ξ0 , f1 )
α =  ... 
 
∀x ∈ BX Φ(x) − α
∞
>ǫ where
(ξ0 , fn )
In other words, α is not in the closure of the convex set Φ(BX ) so those can be separated
by a hyperplane in Rn : there exist real numbers β1 , . . . , βn , ω such that
Xn X n
∀x ∈ BX βi (fi , x) < ω < βi (ξ, fi )
i=1 i=i

Xn  X n  X n

This implies that βi fi < ω < ξ, βi fi 6 βi fi
i=1 i=1 i=1
and we have a contradiction. Therefore, W intersect BX non-trivially. 

Theorem 23 (Kakutani)
A Banach space X is reflexive if and only if BX is weakly compact.

Proof: One direction was shown in Corollary 21. So now suppose that B X is weakly
compact. Then it is compact in X⋆⋆ for the σ(X⋆⋆ , X⋆ ) topology. In particular, it is weak⋆
closed. By Lemma 22, its weak⋆ closure is BX⋆⋆ . Thus
BX = BX⋆⋆
and X is reflexive. 

17
 This theorem has very important consequences.

Corollary 24
Let X be a reflexive Banach space. Every closed subspace is reflexive.

Proof: Let Y be a closed subspace of X. There are, on Y, two topologies: the restriction
of the weak topology on X and the weak topology σ(Y, Y⋆ ). We check that those two
coincide.
Let y0 ∈ Y and consider an elementary σ(Y, Y⋆ )-neighbourhood of y0 :
  
W = y ∈ Y | (fi , y) − (fi , y0 ) < ǫ for all i n ∈ N f1 , . . . , fn ∈ Y⋆ ǫ > 0
By Hahn-Banach, f1 , . . . , fn can be extended to bounded linear functionals g1 , . . . , gn
on X. Thus, since the gi ’s coincide with the fi ’s on Y, we have:
     
W = y ∈ Y | (gi , y)−(gi, y0 ) < ǫ for all i = Y∩ x ∈ X | (gi , x)−(gi , y0 ) < ǫ for all i
so W is an open subset of Y for the restriction of the σ(X, X⋆ ) topology.
Conversely, let W be an elementary neighbourhood of y0 for the trace of the σ(X, X⋆ )
topology. There exist ǫ > 0 and g1 , . . . , gn in X⋆ such that
     
W = Y∩ x ∈ X | (gi , x)−(gi , y0 ) < ǫ for all i = y ∈ Y | (gi , y)−(gi, y0 ) < ǫ for all i
Since g1 , . . . , gn are bounded on X, their restrictions f1 , . . . , fn to Y are also bounded.
Thus
  
W= y∈Y | (fi , y) − (fi , y0 ) < ǫ for all i
is open for the weak topology on Y.
The closed unit ball BY is convex and strongly closed in X. Therefore, it is a weakly
closed subset of B X , which is compact by Kakutani’s theorem. Therefore, BY is compact
for σ(X, X⋆ ). Since that topology coincides on Y with σ(Y, Y⋆ ), it follows that BY is
weakly compact in Y. By Kakutani, Y is reflexive. 

Corollary 25
A Banach space X is reflexive if and only if X⋆ is reflexive.

Proof: Suppose that X is reflexive. Then X = X⋆⋆ and bounded linear functionals on
X⋆⋆ are the same as bounded linear functionals on X. In other words, X⋆⋆⋆ = X⋆ .
Conversely, let’s assume that X⋆ is reflexive. The weak⋆ topology and the weak topol-
ogy on X⋆⋆ then coincide. By Banach-Alaoglu-Bourbaki, X⋆⋆ is reflexive. But X is a
closed subspace of X⋆⋆ so by Corollary 24, X is reflexive. 

18
Corollary 26
Let X be a reflexive Banach space. Any closed convex bounded set is weakly compact.

Proof: If C is closed and convex, it is weakly closed. Since it is bounded, it is included

in B(0, R) for some R and this set is weakly compact by Kakutani. Closed subsets of
compact sets are compact, so C is weakly compact. 
This is all for the moment in general. More will come later, once we study the con-
nection between weak topologies and separability.

4.2 Uniformly convex Banach spaces

Let’s prove the Milmann-Pettis theorem about uniformly convex Banach spaces. Although
it is not a consequence of Kakutani, it uses Banach-Alaoglu-Bourbaki as well as Lemma
22 so it derives from all the work done so far.

Definition 27 A Banach space X is called uniformly convex if and only if

x + y

∀ǫ > 0 ∃δ > 0 ∀x, y ∈ X kxk = kyk = 1 > 1 − δ =⇒ kx − yk < ǫ
2
We first show that this definition is in fact equivalent to an apparently stronger state-
ment:
Lemma 28
Let X be a uniformly convex Banach space. Then
x + y

∀ǫ > 0 ∃δ > 0 ∀x, y ∈ B X > 1 − δ =⇒ kx − yk < ǫ
2

Be careful, this is ugly and does not result from a simple scaling argument. I have
not been able to find anything simpler...

Proof: Let ǫ be a positive real number. By uniform convexity of X, we are given a

positive α such that
x + y ǫ

∀x, y ∈ X kxk = kyk = 1 > 1 − α =⇒ kx − yk 6
2 2
α ǫ 
Let δ = Min ,
2 4
ǫ
so that 2δ 6 α and 2δ 6
2

19
Take x, y non-zero in X, both with norm less than 1 and such that
x + y

>1−δ
2
Suppose also, for example, that kyk 6 kxk. It will be useful to get also a lower bound for
kyk. This is done by noticing that
x+y
y =2× −x
2
x + y

so that kyk > 2 − kxk > 2 − 2δ − kxk
2
and kyk − kxk > 2 − 2δ − 2kxk > −2δ
We scale x and y to bring them on the unit sphere: let
x y
x0 = and y0 =
kxk kyk

x y x y y y
Then kx0 + y0 k =
+ = + − +
kxk kyk kxk kxk kxk kyk

x + y y y 2 − 2δ (kyk − kxk)y
> − − > −
kxk kxk kyk kxk kxk kyk
2 − 2δ kxk − kyk 2 − 2δ + kyk − kxk
> − =
kxk kxk kxk
2 − 4δ
kx0 + y0 k > > 2 − 4δ > 2 − 2α
kxk
x + y
0 0
Thus >1−α
2
ǫ
and from uniform convexity, kx0 − y0 k 6
2
Now we have to relate this to kx − yk. It is the same kind of mess as what we just did:

x y x y y y
kx0 − y0 k = − = −
kxk kyk kxk kxk kxk kyk + −

x − y kxk − kyk kx − yk 2δ
> kxk − > −

kxk kxk kxk
kx0 − y0 k > kx − yk − 2δ
ǫ ǫ
Therefore kx − yk 6 kx0 − y0 k + 2δ 6 + =ǫ 
2 2

20
Theorem 29 (Milmann-Pettis)
A uniformly convex Banach space is reflexive.

Proof: Let ǫ be a positive number, which provides us with a positive δ such that
x + y

∀x, y ∈ BX > 1 − δ =⇒ kx − yk 6 ǫ
2
Let ξ0 ∈ X⋆⋆ with kξ0k = 1 and α be any positive real number. There exists f ∈ X⋆ with
kf k = 1 such that
(ξ0 , f ) > 1 − δ

Define W = ξ ∈ X⋆⋆ | (ξ, f ) > 1 − δ
This is a weak⋆ neighbourhood of ξ0 and by Lemma 22, it intersects BX non-trivially:
there exists x0 ∈ X, with kx0 k 6 1, such that
(f, x0 ) > 1 − δ
Now, assume that kξ0 − x0 k > ǫ. This means that ξ0 is not in the ball B X⋆⋆ (x0 , ǫ).
This set is weak⋆ closed, as we proved in Theorem 20 so its complement is open. Thus,
BX⋆⋆ (x0 , ǫ)c ∩ W
is a weak⋆ neighbourhood of ξ0 . By Lemma 22, it contains a y0 ∈ X with ky0 k 6 1 and
y0 satisfies
kx0 − y0 k > ǫ and (f, y0 ) > 1 − δ
 x +y 
0 0
Therefore f, > 1−δ
2
x + y
0 0
which implies >1−δ
2
From uniform convexity, kx0 − y0 k 6 ǫ
and this is a contradiction. Thus kx0 − ξ0 k 6 ǫ, or in other words
∀ǫ > 0 B(ξ0 , ǫ) ∩ BX 6= ∅
So ξ0 is in the (strong) closure of BX , which is already closed: ξ0 is in fact in X. 

21
5 Weak topologies and separability
We finally investigate how the property of separability influences the weak topologies.
Let’s remember that

Definition 30 A normed space X is called separable if and only if it contains a countable

dense subset.
Theorem 31
Let X be a Banach space such that X⋆ is separable. Then X is separable.

Proof: Let (fn )n∈N be a countable dense subset of X⋆ . For every n, there exists xn in X
with norm 1, such that
kfn k
(fn , xn ) >
2
Let L0 be the Q-vector subspace of X spanned by (xn )n∈N , and L be the R-vector subspace
of X spanned by (xn )n∈N . Then L0 is countable, and dense in L. So if we show that L is
dense in X, we’re done.
Let f ∈ X⋆ , that vanishes on L. Since (fn )n∈N is dense in X⋆ , given a positive ǫ, there
exists n ∈ N such that kfn − f k < ǫ. Then
kfn k
< (fn , xn ) = (fn − f, xn ) < ǫ
2
and kf k 6 kfn − f k + kfn k 6 3ǫ
This is true for all positive ǫ, so f = 0 and L is dense. 

Note that there is no converse to this theorem, that is, if X is separable,

⋆
X has no reason to be separable as well.
Think, for example, of L1 (R) which is separable while its dual, L∞ (R) is not.

Corollary 32
A Banach space X is reflexive and separable if and only if X⋆ is reflexive and separable.

Proof: Assume that X is reflexive and separable. Then X⋆⋆ is reflexive and separable
since it is identified with X. By Corollary 25 and Theorem 31, X⋆ is reflexive and
separable.
Conversely, if X⋆ is reflexive and separable, Corollary 25 and Theorem 31 imply
that X is reflexive and separable. 

22
Theorem 33
A normed space X is separable if and only if the weak⋆ topology on B X⋆ is metrizable.

Notice that there is no contradiction at all with the fact, seen earlier, that a weak
topology is never metrizable on an infinite dimensional space. Indeed, if we remem-
ber how that was proved, we used the fact that weak open sets contained entire
lines. Obviously, if we restrict the topology to a bounded set such as the unit ball,
we cannot use that fact anymore.

Proof: Let X be a separable normed space, which means there is a countable dense subset
A in X. We define D to be A ∩ BX (0, 1); because D is countable, we can enumerate its
elements:

D = xn | n ∈ N
Finally, we define :
P |(f − g, xn )|
∀(f, g) ∈ S2 d(f, g) =
n∈N 2n
where S is the closed unit ball in X⋆ . We claim that d is a metric on S and that the
toplogy T it induces on S coincides with the weak topology σ(X⋆ , X). In order to do this,
there are a few things to check.
1 : D is dense in BX
Let x be in the unit ball of X and let O be an open set containing x. Then there is
some ǫ such that BX (x, ǫ) is included in O. And because BX is open, we can take ǫ smaller
so that BX (x, ǫ) is included in BX . Since A is dense, BX (x, ǫ) intersects A at some y.
And y is then in D. Which shows that D is dense in BX .
And we conclude using the fact that BX is dense in BX .
2 : d is a metric on S
We first check that d is properly defined, i.e. that it does not take the value ∞. This
is easy :
P |(f − g, xn )| P kf − gk kxn k P 1
∀(f, g) ∈ S2 n
6 n
62 n
n∈N 2 n∈N 2 n∈N 2

which is finite.
Triangular inequality and symmetry for d are trivial to check. The only thing remain-
ing is to see what happens if d(f, g) = 0. In this case, we get
∀x ∈ D (f − g, x) = 0

23
Now take any y in X with norm less than 1. From 1.1, there is some sequence (yn )n∈N
in D converging to y. So that :
(f − g, y) = (f − g, lim yn ) = lim (f − g, yn)
n→∞ n→∞ | {z }
=0

since f − g is continuous on X. Which shows that

∀y ∈ X kyk 6 1 =⇒ (f − g, y) = 0

Hence kf − gk = 0

and f =g

d is a metric on S.

3 : T ⊂ σ(X⋆, X)
Let f0 be in S and take an open set O in T containing f0 . Then we can find a positive
real number r such that :
Bd (f0 , r) = {f ∈ S | d(f0 , f ) < r} ⊂ O
1 r r
Now take an integer k such that k < and ǫ < . Define :
2 2 2

V = f ∈ S | |(f − f0 , xj )| < ǫ for j = 0, . . . , k
so that V is a weak⋆ open set in S containing f0 . Then
k
X |(f − f0 , xn )| X |(f − f0 , xn )| 1
∀f ∈ V d(f0 , f ) = n
+ n
6 ǫ + k
<r
n=0
2 n>k+1 |
2
{z } 2
1
6
2n−1

which shows V ⊂ Bd (f0 , r) ⊂ O

Thus for every f0 in S, every T -open neighbourhood of f0 contains a weak⋆ neigh-
bourhood of f0 .
4 : σ(X⋆, X) ⊂ T
Let f0 be in S and take O a weak⋆ open set in S containing f0 . Then O is a union of
elementary weak⋆ open sets containing f0 so that there exists ǫ > 0 and points y1 , . . . , yk
in X such that :

V = f ∈ S | |(f − f0 , yj )| < ǫ for j = 1, . . . , k ⊂ O

Now, take m = Max kyj k | j = 1, . . . , k so that
yj
∀j ∈ {1, . . . , k} ∈S
m

24
Given a positive η, and j ∈ {1, . . . , k}, there is some nj in N such that kxnj − yj /mk < η.
And we take a positive r such that
ǫ
∀j ∈ {1, . . . , k} m2nj r <
2
Then for all f in Bd (f0 , r), we have :
X |(f − f0 , xn )|
<r
n∈N
2 n

so that ∀j ∈ {1, . . . , k} |(f − f0 , xnj )| 6 2nj r

 yj 

Finally, ∀j ∈ {1, . . . , k} |(f − f0 , yj )| = m  f − f0 , 
m 
 yj 
6 m |(f − f0 , xnj )| +m  f − f0 , xnj − 
| {z } | {z m }
nj
62 r 62η
ǫ
6 + 2ηm
2
ǫ
Recall that η was arbitrary ; so just take it small enough so that 2ηm < and we get
2
f ∈ V. Hence
Bd (f0 , r) ⊂ V ⊂ O
So for every f0 in S and every weak⋆ open neighbourhood O of f0 in S, there is a
T -neighbourhood of f0 contained in O.
5 : Conclusion
Both topologies T and ⋆ − σ(X⋆ , X) coincide.
Now, suppose that S together with the weak⋆ topology is metrizable: there exists a
distance d on S such that the induced topology coincides with σ(X⋆ , X). Define
 1 n  1o
∀n ∈ N⋆ Bn = Bd 0, = f ∈ S  d(0, f ) <
n n
Each Bn is an weak⋆ open neighbourhood of 0, so there exist a positive ǫn and Fn ⊂ X,
finite, such that
  
Wn = f ∈ X⋆ | (f, x) < ǫn for all x ∈ Fn ⊂ Bn
[
Let F= Fn
n∈N⋆

This is a countable subset of X. Let’s show that it spans a dense subspace of X. Let f
be a bounded linear functional that takes the value 0 at each x ∈ F. Then
\ \
f∈ Wn ⊂ Bn = {0}
n∈N⋆ n∈N⋆

25
So f is 0 and as a consequence of the Hahn-Banach theorem, Span F is dense. This is a
separable set (take finite linear combinations of elements of F with rational coefficients).
Thus X is separable. 
This theorem has a dual equivalent, though more difficult to prove:

Theorem 34
Let X be a Banach space. Its dual is separable if and only if BX is metrizable for the weak
topology σ(X, X⋆ ).

We end our work with the two important consequences of everything done so far:

Theorem 35
Let X be a separable Banach space. Any bounded sequence in X⋆ has a weak⋆ converging
subsequence.

Proof: Let (fn )n∈N be any bounded sequence in X⋆ . There exists a positive real number
M such that
fn
∀n ∈ N gn = ∈ B X⋆
M
The sequence (gn )n∈N takes its values in the unit ball of X⋆ , which is weak⋆ compact by
Banach-Alaoglu-Bourbaki. And the weak⋆ topology on it is metrizable by Theorem 33.
Therefore, (gn )n∈N has a weak⋆ converging subsequence, as well as (fn )n∈N .

Theorem 36
Let X be a reflexive Banach space. Any bounded sequence has a weakly converging subse-
quence.

Proof: Let (xn )n∈N be a bounded sequence in X and let M be the closure in X of
Span (xn )n∈N . Then M is a closed subspace of X. By Corollary 24, M is reflexive.
By construction, M is also separable since the Q-vector subspace spanned by the (xn )n∈N
is dense.
By Theorem 34, there is a σ(M⋆⋆ , M⋆ ) converging subsequence of (xn )n∈N , with limit
x. But since M is reflexive, the topologies σ(M⋆⋆ , M⋆ ) and σ(M, M⋆ ) coincide. So (xn )n∈N
converges to x for the topology σ(M, M⋆ ).
Now, if f is any bounded linear functional on X, its restriction to M is of course
bounded on M. Therefore,
lim (f, xn ) = (f, x)
n→∞

So (xn )n∈N converges to x for σ(X, X⋆ ). 

26