You are on page 1of 10

Three theorems about bounded linear operators

P. B. Kronheimer, for Math 114, with thanks to my own teachers


November 15, 2010

1. Introduction

These notes present the proofs of a standard collection of basic theorems about operators
on Banach spaces. These are the uniform boundedness theorem (also known as the BanachSteinhaus theorem), the open mapping theorem and the closed graph theorem. All of these
results rest one basic result, the Baire category theorem, with which we begin.

2. Baires category theorem

Before coming to Baires theorem, we need a simple lemma about complete metric
spaces.
Lemma 2.1. Let (X , d) be a complete metric space, and let {Sn }nN be a collection of closed sets

that are nested:


S1 S2 .
Writing
n = diam(Sn )
= sup{ d(x , y) | x , y Sn }

we suppose that n 0 as n . Then the intersection S = n=1 Sn consists of exactly one


point.
Proof. For each n, let x n be a chosen element of Sn . For all m , m 0 n, we have
d(x m , x m0 ) n , because both points belong to Sn . So our chosen sequence is Cauchy.

Because X is complete, there is a limit point x. Each Sn is closed, so x belongs to Sn for


all n; so x S. The condition on the diameters also ensures that there cannot be two
distinct points in S.

2.

Baires category theorem

The simplest statement of the Baire category theorem is the following.


Theorem 2.2. Let X be a compete metric space, and for each n N let G n be an open, dense
subset of X . Let G = G n . Then G is also dense in X .

Remark. G need not be open.


Proof. To show that G is dense is to show that G has non-empty intersection with every
open set E X . So let E be open. Since G 1 is dense, G 1 E is non-empty. It is also
open, so there is an open metric ball S1 contained in G 1 E. Let 1 be the diameter of
S 1 . Now G 2 S 1 is also non-empty and open, so it contains an open ball S 2 . We can
arrange that the diameter, 2 , of the ball S 2 satisfies 2 1 /2. We can also arrange, by
decreasing 2 , that the closed ball S2 is contained in S 1 G 2 .

In this way, for all n, we find balls Sn whose diameters decrease to zero, such that
Sn G 1 G n and Sn+1 Sn . By the lemma, the intersection of the closed balls
(and therefore of Sn ) is non-empty. A point of this intersection lies in E and in all G n .
So G E is indeed non-empty, as we wished to show.
The Baire category theorem is often restated as follows. A set K in a metric space X is
nowhere dense if any of the following equivalent conditions hold:
(a) for all balls B in X , the intersection K B is not dense in B;
(b) the closure of K does not contain any ball;
(c) the complement of the closure of K is (open and) dense in X .
A subset M X is meager if it is a countable union of nowhere dense sets:
M = K1 K2 . . .
Note that a countable union of meager sets is meager. The Baire theorem is often stated
like this:
Theorem 2.3. The complement of a meager set in a complete metric space X is dense. In particular,

a meager set cannot comprise all of X .


Proof. If M is meager, then it is a countable union of nowhere dense sets K n . Consider

the larger set M formed as the union of their closures:


M = K 1 K 2 . . . .

The complement of M is the intersection


X \ M = G1 G2 . . .
where G n = X \ K n . The definition of nowhere dense means precisely that each G n
is dense. These sets are also open; so by the previous statement of the Baire theorem,
their intersection is dense. That is, X \ M is dense. The larger set X \ M is therefore
dense too.
The complement of a meager set is comeager. The theorem says that comeager sets are
dense. A countable intersection of comeager sets is comeager.
Example. In Rn , a hyperplane is a nowhere dense subset. A countable union of hyper-

planes is a meager set. The theorem tells us that a countable union of hyperplanes cannot
fill all of Rn . The complement of the union of hyperplanes is comeager.

3. Uniform boundedness principle


Theorem 3.1. Let X be a Banach space and Y a normed space. Let E L(X , Y ) be any set
of bounded linear operators. Suppose that for every x X , there exists M x 0 such that, for all
T E,
kT xk M x .

Then there exists M 0 such that, for all T E,


kT k M .
Remark. The conclusion of the theorem states that the operators T in E are uniformly
bounded.
Proof. For each x X , let M x be chosen as in the statement of the theorem. For each
n, let K n X be the closed set defined by

K n = { x : kT xk n , T E }.
If n M x , then x belongs to K n ; so the union of all these K n is the whole of X . So not
all of the K n can be nowhere dense, by the Baire theorem in its second form above. So
we can find a particular n so that the closed set K n actually contains a closed ball:
0;  ) Kn .
B(x

3.

Uniform boundedness principle

0 ;  ) and therefore belongs to


If x X , then x 0 +  x/kxk belongs to the closed ball B(x
the closed set K n . So, for all x and all T E,





T x 0 +  x n .

kxk

With an application of the triangle inequality, this yields (for all x and T E)
kT xk

1


1


So
kT k

n + kT x 0 k kxk


n + M x 0 kxk
1


n + M x0

for all T in E.
We can sharpen up the statement of the theorem a little. The last part of the proof
shows that if there is any n such that K n contains a ball, then the family of operators
E L(X , Y ) is uniformly bounded. To state this the other way round: if the family of
operators E is not uniformly bounded, then every K n is nowhere dense. This leads to
the following version:
Corollary 3.2. Suppose E L(X , Y ) is not uniformly bounded: that is,

sup{ kT k : T E } = .
Then there exists x X such that
sup{ kT xk : T E } = .
Indeed, the set of points x X with this property is comeager.
Proof. The first part of this corollary is just a restatement of the original theorem. For

the second part, we noted above that if E is not uniformly bounded, then the closed set
K n = { x : kT xk n , T E }.
is nowhere dense, for all n. The union K = n K n is therefore meager. If x belongs to
the comeager set X \ K , then
sup{ kT xk : T E } = .
This proves the corollary.

Application: non-convergence of Fourier series. Let C( T ) denote the Banach space of continuous periodic functions f : R C with period 1. For each f C( T ), let sn f denote

as before the nth partial sum of its Fourier series. In class, we indicated the construction
of a function f for which the partial sums (sn f )(0) at x = 0 diverged as n . Using
the uniform boundedness principle, we can establish the existence of such an f , and
show further that functions f with this apparently pathological behavior are comeager
in C( T ).
Proposition 3.3. There is a comeager set of functions f C( T ) for which the partial sums of the
Fourier series at x = 0, (sn f )(0), diverge

sup |(sn f )(0)| = .


n

Proof. For each n 0, let Tn L(C( T ), C ) be the linear functional

Tn ( f ) = (sn f )(0).
Each Tn is a bounded linear functional, but the collection of functions
E = {Tn : n N }
is not uniformly bounded. Indeed, as we saw in class, by taking fn to be a continuous
approximation to the
sign(Dn )
where Dn is the Dirichlet kernel, we can see that
Z

kTn k

|Dn |
0

which diverges like log n. By the above corollary to the uniform boundedness theorem,
there exists a comeager set, Z, of functions f C( T ) such that
sup |Tn f | =
n

for all f Z. This statement is precisely the conclusion of the Proposition.


The same argument applies, of course, to the values of sn f at any point x [0, 1], not
just x = 0. Because a countable intersection of comeager sets is comeager, we deduce:
Corollary 3.4. Given any countable set of points {x k }k N in [0, 1], we can find a continuous
periodic function f C( T ) whose Fourier series is divergent at every x k . Indeed, the functions f

with this property are a comeager set.

3.

Uniform boundedness principle

So, at any given countable set of points, convergence of a Fourier series is the exception,
not the rule. This should be contrasted with the corollary to Carlesons very difficult
theorem, which tells us that for any continuous periodic function f , the partial sums
sn f (x) converge for almost all x.

Application: boundedness of limit operators. Here is another application of the uniform

boundedness theorem.
Proposition 3.5. Let X be a Banach space and Y a normed space. Let Tn , n N, be a sequence
of bounded linear operators X Y , and suppose that for every x X the sequence {Tn x}nN is
convergent in Y . Then there exists a bounded linear operator T : X Y such that

lim Tn x = T x

for x.
Proof. Define T x, for all x, as the limit of the sequence Tn x. It is straightforward

to see that T is a linear map from X to Y , and the only thing that is in question is
whether T is bounded. The hypotheses imply that, for all x, the sequence kTn xk,
n N, is bounded. So the uniform boundedness theorem implies that the norms kTn k
are uniformly bounded by some constant M . That is,
kTn xk M kxk
for all n and x. Taking the limit as n we see that kT xk M kxk also; so T is
bounded.
A linear map X Y between normed spaces is bounded if and only if it is continuous.
So we can rephrase the above proposition suggestively by saying: if X is a Banach
space and we have a sequence of continuous linear maps Tn : X Y which converge
pointwise on X to a map T : X Y , then the limit T is also continuous. In short, a
pointwise limit of continuous maps is continuous, as long as we are talking about linear
maps on a Banach space. With this phrasing, we can contrast the theorem with what we
know about continuous functions R R, for example, where we can find a sequence
of continuous functions which converge pointwise to a function that is discontinuous.
Of course, such examples are not comprised of linear functions.

4. The open mapping theorem


Theorem 4.1. Let X and Y be Banach spaces and T L(X , Y ) a surjective, bounded linear
operator. Then the map T is open: that is, if G X is open, then T (G ) is an open subset of Y .
Proof. Put

U = { x X : kxk < }
V = { y Y : kyk <  }.

We proceed in steps.
(i) TU1 contains some V . Because

X=

Un ,

we certainly have
Y=

TUn .

So by the Baire category theorem, there exists m such that TUm contains an open ball.
Hence TU1 contains an open ball, and even TU1/2 contains an open ball: so TU1/2
contains
V = { y : ky y0 k <  }.
But then
V = V y 0
TU1/2 TU1/2
TU1 .
(ii) TU1 TU2 . Let y TU1 . Then there exists x 1 U1 such that

y T x 1 V /2 TU1/2 .
Similarly there exists x 2 U1/2 such that
y T x 1 T x 2 V /4 TU1/4 ;
and for each n there exists x n U1/2n such that
y T x 1 T x n V /2n TU1/2n .
P

But 1 kx j k < 2, so the series whose terms are the x j converges to some x X with
kxk < 2; that is, x U2 . And T x = y because T is continuous.

4.

The open mapping theorem

(iii) Completion of the proof. Now TU2 contains some V , so TU contains V/2 . So if

U is any neighborhood of 0 in X , then TU contains some V 0 . Now given any open


set G X and any x G, the set G x is a neighborhood of 0; so TG T x contains
some V 0 . That is, we can find an open ball (of radius  0 ) around T x contained in TG.
This shows that TG is open.
The open mapping theorem is often put to work via the following corollary:
Corollary 4.2. Let T : X Y be a bounded linear operator that is one-to-one and onto. Then
the inverse linear map S : Y X is also bounded.
Proof. The open mapping theorem applies to T because T is surjective. So T maps

open sets to open sets. The inverse linear map S has the property that S 1 (G) is open
whenever G is open (because S 1 is T ). This property means that S is continuous, so S
is bounded.
Alternatively, we can phrase the proof as follows. The open mapping theorem implies
that, if U1 is the unit ball in X , then TU1 contains V , so TU1/ contains V1 . Looking
at then inverse map, we see
SV1 U1/ .
In other words, if kyk < 1, the kSyk < 1/ . So S is bounded and its norm is at most
1/ .
Remark. A bounded linear map T : X Y which has an inverse map that is also bounded
is an isomorphism of Banach spaces. This is a much less restrictive condition than saying
that X and Y are isometrically isomorphic. For example, two finite-dimensional Banach
spaces are isomorphic if they have the same dimension (because, as discussed in a problem
set, any two norms on Rn or Cn are equivalent). But it is not the case that any two finitedimensional Banach spaces of the same dimension are isometric. For example, R2 with
the taxi-cab norm and R2 with the Euclidean norm are not isometrically isomorphic.
As an illustration of the theorem, consider the Banach spaces L 1 ( T ) and c 0 ( Z ) and the
linear map
^ : L 1(T) c0(Z)
given by taking the Fourier series:
^ f = f (n)


n Z

We can ask, is ^ onto? That is: is it the case that every sequence a c 0 ( Z ) is the
sequence of Fourier coefficients of some integrable periodic function? The answer is no,
and the open mapping theorem allows us to deduce this on rather general grounds, as

follows. We already know that ^ injective: this again is the statement that if f (n) = 0
for all n, then f = 0. We also know that ^ is bounded: this follows from the fact that
supn | f (n)| k f kL 1 (T ) . If ^ were surjective, then its inverse map would be bounded, by
the corollary to the open mapping theorem. That is to say, there would be a constant M
such that
k f kL 1 (T ) M sup k f (n)k
n

for all f . However, we can see that there is no such M by looking at the function f =
Dm , the mth Dirichlet kernel, which has supn k f (n)k = 1 while (as we have remarked
before) k f kL 1 (T ) diverges like log(m).
Without mentioning a specific example like the Dirichlet kernel, we can finish the above
argument by playing the endgame differently. If ^ were surjective, then the open mapping theorem would be telling us that it is an isomorphism, so L 1 ( T ) and c 0 ( Z ) would be
isomorphic Banach spaces. This cannot be the case, however, because the duals of these
two Banach spaces are not isomorphic: the dual of c 0 ( Z ) is separable (it is l 1 ( Z )) but the
dual of L 1 ( T ) is not separable (it is L ( T )). Of course, these assertions (particularly the
second one) are actually harder to understand than the relatively straightforward fact that
the norm of the Dirichlet kernel increases like log(m). But this line of argument does
illustrate what we can get from thinking about Banach spaces in the large.

5. The closed graph theorem

If X and Y are Banach spaces, then X Y is also a Banach space when equipped with
the norm
k(x , y)k = kxk + kyk.
Concretely, a sequence (x n , yn ), n N, converges to (x , y) X Y in this norm if and
only if x n x and yn y. The sequence {(x n , yn )}nN is Cauchy if and only to the
sequences {x n }nN and {yn }nN are both Cauchy.
Now let T : X Y a linear map (not necessarily bounded). The graph of T is the linear
subspace
(T ) = { (x , T (x)) : x X }
X Y.
We say that T has closed graph if (T ) is a closed linear subspace of X Y . In the spirit of
the concrete description of convergence described just above, the closed graph condition
can be stated as follows: if x n is a convergent sequence converging to x X , and T x n
converges to y Y , then the closed graph condition requires that y = T x.

10

5.

The closed graph theorem

The following is the closed graph theorem:


Theorem 5.1. If X and Y are Banach spaces and T : X Y is a linear map with closed graph,

then T is bounded.
Proof. Suppose the graph (T ) is closed. As a closed subspace of the Banach space
X Y , the space (T ) is itself a Banach space with the inherited norm, k(x , T x)k =
kx| + kT xk. The linear map
: (T ) X

given by projection onto the first factor,


(x , T x) 7 x
is certainly bounded, because it decreases the norm. It is also one-to-one and onto. So
by the open mapping theorem, it follows that the inverse map is bounded. The inverse
map is the map x 7 (x , T x), and its boundedness means that there exists a constant M
with
kxk + kT xk M kxk
for all x. This implies also that kT xk M kxk. In other words, T is bounded.
Remark. The converse of the theorem holds too, but is straightforward: if T is bounded,
then the graph is closed.
The idea of the product Banach space X Y appears in another application of the open
mapping theorem, as follows. If Z is a Banach space and X , Y are closed subspaces
such that Z = X Y as vector spaces (i.e X Y = {0} and X + Y = Z), then we can
compare Z with the Banach space X Y equipped with the above norm. We have a
natural map
X Y Z (x , y) 7 x + y
which is one-to-one and onto. The triangle inequality tells us it is onto. It follows
that it is an isomorphism of Banach spaces. So Z is isomorphic to X Y (though not
necessarily isometrically isomorphic).

You might also like