3 The Principle of Uniform Boundedness

Many results of functional analysis rest on the following proposition.

Proposition 3.1 (Baire)

Let (X, d) be a complete metric space, let (Un )n∈N be a sequence of open subsets of X, let
Un be dense in X for all n ∈ N. Then
\
D= Un (3.1)
n∈N

is dense in X.

Proof: It suffices to show: if V ⊂ X is open and not empty, then V ∩ D 6= ∅. Let V be

open, V 6= ∅. We choose x1 ∈ U1 ∩ V (this is possible as U1 is dense in X) and an ε1 > 0
such that
x1 ∈ K(x1 ; ε1 ) ⊂ U1 ∩ V . (3.2)
Assume that (x1 , ε1 ), . . . , (xn−1 , εn−1 ) are defined already. We choose

xn ∈ Un ∩ B(xn−1 ; εn−1 ) (possible since Un is dense in X) (3.3)

and εn > 0 such that

1
xn ∈ K(xn ; εn ) ⊂ Un ∩ B(xn−1 ; εn−1 ) , εn ≤ εn−1 . (3.4)
2
Then

d(xn+1 , xn ) ≤ εn ≤ 2−n+1 ε1 , d(xm , xn ) ≤ 2 · 2−n+1 ε1 for all m > n.

Therefore, (xn )n∈N is a Cauchy sequence. Let x = limn→∞ xn . We have K(xn ; εn ) ⊂

K(xn−1 ; εn−1 ) for all n ∈ N and therefore xm ∈ K(xn ; εn ) for all m > n. Since K(xn ; εn )
is closed, it follows that x ∈ K(xn ; εn ) ⊂ V for all n. Consequently,
\ \
x∈ K(xn ; εn ) ⊂ Un ∩ V = D ∩ V .
n∈N n∈N

Definition 3.2 Let (X, d) be a metric space, M ⊂ X. M is called nowhere dense in

X, if int (M ) = ∅ (that is, the closure of M has no interior points). 2

A closed set A ⊂ X is nowhere dense in X if and only if its complement X \ A is open

and dense in X.

Corollary 3.3 Let (X, d) be a complete metric space, let (An )n∈N be a sequence of closed
and nowhere dense subsets of X. Then
[
X\ An (3.5)
n∈N

is dense in X.

32
Proof: We apply Proposition 3.1, setting Un = X \ An . 2
Proposition 3.1 and Corollary 3.3 imply in particular that D = ∩n Un resp. X \ ∪n An are
nonempty. Thus, they provide a way to prove existence of certain objects. For example
one can show in this manner (see Werner p.139) that the set of all continuous but nowhere
differentiable functions forms a dense subset of (C([a, b]), k · k∞ ).
We present another existence result.

Corollary 3.4 Let (X, d) be a complete metric space, let (An )n∈N be a sequence of closed
subsets of X such that [
X= An . (3.6)
n∈N

Then there exists a k ∈ N such that int (Ak ) 6= ∅.

Proof: This is a direct consequence of Corollary 3.3. 2

From Corollary 3.4 one can conclude (exercise) for example that there does not exist a
Banach space which has a countably infinite vector space basis.
Subsets of X of the form
[
Mn , Mn nowhere dense in X,
n∈N

are also called sets of first category in X. Subsets of X which are not of first category
are called of second category in X. When one uses this terminology, a complete metric
space is always of second category (in itself).
The statement of the following proposition is called the principle of uniform boundedness.

Proposition 3.5 (Banach-Steinhaus)

Let X be a Banach space, Y a normed space, let T ⊂ L(X; Y ). Assume that

sup kT xk < ∞ , for all x ∈ X. (3.7)

T ∈T

Then
sup kT k < ∞ . (3.8)
T ∈T

Thus, if a family of linear and continuous operators is pointwise bounded, it is uniformly

bounded.
Proof: For n ∈ N we define

An = {x : x ∈ X, sup kT xk ≤ n} . (3.9)
T ∈T

An is closed, since for fT (x) := kT xk we have

\
An = fT−1 ([0, n]) .
T ∈T

33
By assumption (3.7) we have [
X= An .
n∈N

Due to Corollary 3.4 we can choose a k ∈ N such that int (Ak ) 6= ∅. Let x0 ∈ Ak , ε > 0
with K(x0 ; ε) ⊂ Ak . Let now x ∈ X be arbitrary, x 6= 0, then
x
z = x0 + ε ∈ K(x0 ; ε) ⊂ Ak .
kxk
In addition, let T ∈ T be arbitrary. Then
 
kxk kxk kxk
kT xk = T (z − x0 ) = kT z − T x0 k ≤ 2k ,
ε ε ε
and therefore
2k
kT k ≤ , for all T ∈ T .
ε
2

Definition 3.6 (Open Mapping) Let (X, dX ), (Y, dY ) be metric spaces. A mapping
T : X → Y is called open if T (U ) is open in Y for every open subset U of X. 2

When the mapping T is open, the image of closed subset of X is not necessarily closed in
Y ; not even when T is linear (Example: exercise.)

Proposition 3.7 (Open mapping theorem)

Let X, Y be Banach spaces, T : X → Y linear and continuous. Then there are equivalent:

(i) T is open.

(ii) There exists δ > 0 such that B(0; δ) ⊂ T (B(0; 1)).

(iii) T is surjective.

Proof: “(i)⇒(ii)”: obvious.

“(ii)⇒(i)”: Let U be open in X, let y ∈ T (U ). We choose x ∈ U such that T x = y, and
ε > 0 such that B(x; ε) ⊂ U . Then

T (U ) ⊃ T (B(x; ε)) = T x + T (B(0; ε)) ⊃ y + B(0; δε) ,

where δ > 0 is chosen according to (ii). Then y ∈ int (T (U )).

“(ii)⇒(iii)”: As T is linear, it follows from (ii) that B(0; rδ) ⊂ T (B(0; r)) for all r > 0.
“(iii)⇒(ii)”: Since T is surjective, we have
[
Y = T (B(0; n)) .
n∈N

According to Corollary 3.4 we choose a k ∈ N such that int (T (B(0; k))) 6= ∅. We set
V = T (B(0; k)). Let now y ∈ V , ε > 0 such that

B(y; ε) ⊂ V .

34
Since V is symmetric (that is, z ∈ V ⇒ −z ∈ V ), we see that B(−y; ε) ⊂ V . Since V is
convex, it follows that
B(0; ε) ⊂ V = T (B(0; k)) . (3.10)
It now suffices to show that
B(0; ε) ⊂ T (B(0; 3k)) , (3.11)
because then (ii) is satisfied for δ = ε/3k. By (3.10), for every y ∈ B(0; ε) we find an
x ∈ B(0; k) such that ky − T xk < ε/2, so

2(y − T x) ∈ B(0; ε) . (3.12)

Let now y0 ∈ B(0; ε) be given. We construct sequences (xj ) in X, (yj ) in Y by choosing

x0 , y1 , x1 , y2 , x2 , . . . such that

xj ∈ B(0; k) , yj+1 = 2(yj − T xj ) , yj+1 ∈ B(0; ε) , j ∈ N. (3.13)

It follows that
2−(j+1) yj+1 = 2−j yj − T (2−j xj ) , j ∈ N,
therefore !
m
X
T 2−j xj = y0 − 2−(m+1) ym+1 → y0 (3.14)
j=0

f”ur m → ∞. Since m
X
2−j kxj k ≤ 2k , (3.15)
j=0

2−j xj is absolutely convergent, hence convergent (as X is a Banach space).

P
the series j
Let ∞
X
x= 2−j xj .
j=0

Now (3.14) implies that T x = y0 , and (3.15) implies that kxk ≤ 2k < 3k. Thus (3.11) is
proved. 2

Corollary 3.8 Let X, Y be Banach spaces, T : X → Y linear, continuous and bijective.

Then T −1 : Y → X is linear and continuous.

Proof: That T −1 is linear, is a result of linear algebra. It follows from Proposition 3.7
that T is open. Therefore, for every open U ⊂ X it follows that (T −1 )−1 (U ) = T (U ) is
open. Thus, T −1 is continuous. 2
Let us consider the situation when a vector space X is endowed with two different norms
k · k1 and k · k2 so that
kxk1 ≤ Ckxk2 , for all x ∈ X, (3.16)
holds with a constant C which does not depend on x. This means that k · k1 is weaker
than k · k2 in the sense that convergence w.r.t. k · k2 implies convergence w.r.t. k · k1 , but
not the other way round. (3.16) means that

id : (X, k · k2 ) → (X, k · k1 )

35
is continuous. We now conclude from Corollary 3.8 that
id : (X, k · k1 ) → (X, k · k2 )
is continuous, too, if (X, k · k1 ) and (X, k · k2 ) both are Banach spaces; in this case it
follows that both norms are equivalent, and
(X, k · k1 ) ' (X, k · k2 ) .
If (3.16) holds, but the two norms are not equivalent, then at least one of the spaces
(X, k · k1 ) and (X, k · k2 ) is not complete. Example: X = C([a, b]). C([a, b]) is complete
for k · k2 = k · k∞ , and for the L1 -Norm we have
kxk1 ≤ (b − a)kxk∞ ,
but these two norms are not equivalent (there exist sequences (xn )n∈N with kxn k1 → 0,
but kxn k∞ 9 0). Therefore (C([a, b]); k · k1 ) is not complete.

Corollary 3.9 Let X, Y be Banach spaces, T : X → Y linear and continuous, assume

that T (X) = Y and T (X) 6= Y . Then there exists y ∈ Y such that kxn k → ∞ holds for
every sequence (xn )n∈N in X which satisfies limn→∞ T xn = y.

Proof: Exercise. 2
In the situation of Corollary 3.9, problems arise if one wants to solve T x = y for a given
y in a “stable” manner. The corollary implies that there exists a sequence (yn ) in Y of
approximations yn of y, namely yn → y, for which kxkn → ∞ holds for the exact solutions
xn of the equation T x = yn . We might want to use (xn ) as a sequence of approximate
solutions of T x = y; but they are worthless for large values of n.
The graph of a mapping T : X → Y is defined as
graph (T ) = {(x, T x) : x ∈ X} . (3.17)

Proposition 3.10 (Closed graph theorem)

Let X, Y be Banach spaces, let T : X → Y be linear. Then there are equivalent:

(i) graph (T ) is a closed subset of X × Y .

(ii) T is continuous.

Proof: “(ii)⇒(i)”: Let (xn , yn ) be a sequence in graph (T ) such that (xn , yn ) → (x, y) ∈
X × Y ; then yn = T xn , xn → x and T xn → T x, thus y = T x.
“(i)⇒(ii)”: As T is linear, graph (T ) is a closed subspace of X × Y and therefore a Banach
space. The projection PX |graph (T ) : graph (T ) → X is linear, continuous and bijective.
By Proposition 3.8, it has a linear and continuous inverse Q : X → graph (T ). It follos
that T = PY ◦ Q is continuous. 2
This result is useful when one wants to prove that a given T is continuous; it says that it
suffices to prove that if xn → x and T xn → y, then T x = y. (Otherwise one has to prove
first that if xn → x, then T xn converges to some y.)

36