Mathematics 721

Fall 2010
The Uniform boundedness principle
The following theorem is known as the Theorem of Banach and Stein-
haus who published a proof based on the Baire category theorem.
Theorem. Let B be a Banach space and Y be a normed space and G
be a family of bounded operators from B to Y . Suppose that for every
f B there is a constant C(f) < so that
sup
TG
Tf
Y
C(f).
Then one also has
sup
TG
T < .
Here T T
BY
:= sup{Th
Y
: h
B
= 1}.
Textbooks nowadays base the proof on the Baire category theorem.
However the original proofs were essentially constructive. The proof
below is due to Hahn
1
, who gives credit for the idea to Lebesgue. Hahn
assumed that the operators T
n
are linear functionals (i.e. Y = R or
Y = C); this was later generalized. The method of proof is known as
the gliding hump method.
Proof. We argue by contradiction and assume that sup
TG
T = .
The goal now is to construct an example f B for which sup
TG
Tf
Y
=
, which achieves a contradiction.
We pick T
1
G with T
1
1 and f
1
B so that
T
1
f
1

Y
1 , f
1

B
1.
We construct sequences of operators {T
n
}
nN
with T
n
G and vectors
f
n
B so that for n 2
T
n
f
n

Y
n +
n1

k=1
C(f
k
)
f
n

B
2
n
min
1jn1
T
j

1
Indeed suppose that n 2 and the T
1
, . . . , T
n1
, f
1
, . . . , f
n1
are
already chosen. Since we assume sup
TG
T
BY
= we can pick an
1
Hans Hahn,

Uber Folgen linearer Operationen. Monatsh. Math. Phys. 32
(1922), no. 1
1
2
operator in G, labeled T
n
, and h
n
B with h
n

B
= 1 so that
T
n
h
n

Y
2
n+1
_
n +

n1
k=1
C(f
k
)
_
min
1jn1
T
j

1
.
The above statement then follows with f
n
= h
n
2
n
min
1jn1
T
j

1
.
Since by the induction hypothesis T
j
1 for j n 1 we obviously
get T
n
2
n+1
n 1.
By construction we have f
n
2
n
and by the completeness of B
we see that f := lim
N

N
n=1
f
n
exists in B.
We now check that
T
n
f
Y
n 1
which is a contradiction to the assumption sup
n
T
n
f sup
TG
Tf
Y

C(f) < .
T
n
is a bounded operator and therefore T
n
f =

k=1
T
n
f
k
with con-
vergence in Y . To prove a lower bound we show that in

k=1
T
n
f
k
the main contribution comes from the term T
n
f
n
(the hump) and we
estimate, by the triangle inequality,
T
n
f
Y
T
n
f
n

Y

n1

k=1
T
n
f
k

k=n+1
T
n
f
k

Y
.
For the last term we use the upper bounds for the functions f
k
, namely

k=n+1
T
n
f
k

Y
T
n

k=n+1
f
k

k=n+1
2
k
min
1jk1
T
j

1
T
n

T
n

1
T
n

k=n+1
2
k
= 2
n
< 1 .
By construction of T
n
T
n
f
n

Y

n1

k=1
T
n
f
k

Y
n +
n1

k=1
_
C(f
k
) T
n
f
k

n
(since C(f
k
) = sup
TG
Tf
k

Y
T
n
f
k

Y
). By the two previous dis-
plays,
T
n
f
Y
T
n
f
n

Y

n1

j=1
T
n
f
j

k=n+1
T
n
f
k

Y
n 1.

3
Example: Limiting methods
We are given an innite matrix A = (a
nk
)
n,k=1,2,...
with the property
that

k=1
a
nk
x
k
converges whenever lim
k
x
k
exists (this is certainly
true if

k=1
|a
nk
| < for all n N).
By assumption we may attach to every convergent sequence x
n
a
new sequence y = {y
n
} with y = Ax, i.e. y
n
=

k=1
a
nk
x
k
for n N.
Denition. The matrix A denes a limiting method (also known as
a summability method
2
) if the sequence y = Ax has limit L whenever
the sequence x has limit L. In other words lim
n
(Ax)
n
= lim
n
x
n
whenever the latter limit exists.
Question: Can one characterize those matrices A for which A denes
a limiting method?
Indeed one can; there are two obvious necessary conditions and an-
other not so obvious necessary condition, and these three conditions
will be also sucient. The necessary conditions are listed in the fol-
lowing denition.
Denition. The innite matrix A is called regular if the following
conditions are satised.
(i) lim
n
a
nk
= 0, for every k N.
(ii) lim
n

k=1
a
nk
= 1
(iii) sup
nN

k=1
|a
nk
| < .
The uniform condition (iii) is stronger than the niteness condition
on

k=1
|a
nk
| mentioned in parentheses above.
Example: The Cesaro means
y
n
=
x
1
+ + x
n
n
dene a regular method; here a
nk
= 1/n if 1 k n and a
nk
= 0 for
k > n.
Toeplitz theorem. A matrix A denes a limiting method if and
only if it is regular.
Proof: The interesting part of the proof is the necessity. We thus
assume that A denes a limiting method. Condition (i) seen to be
necessary by testing against the sequence
k
, with
k
k
= 1 and
k
n
= 0
if n = 0. Condition (ii) seen to be necessary by testing against the
sequence (1, 1, 1, ...).
2
this terminology makes sense if the x
n
are the partial sums of a series
4
To prove the necessity of (iii) we rst prove that

k=1
|a
nk
| < for
all n N.
Consider the normed space c of convergent sequences as a closed (and
thus complete) subspace of the complete space

(with the sup-norm).

n,M
(x) =
M

k=1
a
nk
x
k
.
Clearly |
n,M
(x)|

M
j=1
|a
nj
|x

. The constant

M
j=1
|a
nj
| is sharp
as one can test the inequality of the sequence with x
k
= sign (a
nk
) for
k M and x
k
= 0 for k > M. Thus

n,M
=
M

j=1
|a
nj
|.
Now x x c. Then by assumption on Athe sum

k=1
a
nk
x
k
converges
for every n. Thus for xed n we have
sup
M
|
n,M
(x)| C(n, x) < .
By the uniform boundedness principle we also get sup
M

n,M

C(n) < , i.e.

k
|a
nk
| < for every n N.
This means that for every n N

n
(x) =

k=1
a
nk
x
k
denes a bounded linear functional on c (indeed on

); moreover by
testing as above we see that

n
= sup
M
M

k=1
|a
nk
| =

k=1
|a
nk
|.
Now if A is a limiting method then the lim
n

n
(x) exists for every
x c and this limit is equal to lim
n
x
n
. In particular we have
sup
nN
|
n
(x)| C(x) <
for every x c. By the uniform boundedness principle we have sup
n

n
<
and hence sup
n

k=1
|a
nk
| < which is the necessary condition
(iii).
Proof of suciency. This is a standard argument in calculus, similar
to the one for the example of Cesaro means, and also similar to the
argument used for approximations of the identity.
5
We assume that the limiting method is regular and we have to show
that for any x c with lim
n
x
n
= L we also have lim
n

k=1
a
nk
x
k
=
L.
Thus let > 0 and we are going to verify that |

k=1
a
nk
x
k
L| <
for large n.
Let M = sup
n

k=1
|a
nk
|. We split

k=1
a
nk
x
k
L =

k=1
a
nk
(x
k
L) + L(1 +

k
a
nk
).
By condition (ii) there is N
1
() so that |L(1 +

k
a
nk
)| < /3 for
n N
1
().
Since limx
n
= L, there is N
2
() so that |x
k
L| < /(3M) for
k N
2
(). We then have for all n N

kN
2
()
a
nk
(x
k
L)

kN
2
()
|a
nk
||(x
k
L)|

3M

k=1
|a
nk
|

3
.
We still have to take care of the terms

1kN
2
()
a
nk
(x
k
L). But
for each k we have a
nk
0 as n by condition (i). Thus
also lim
n

1kN
2
()
a
nk
(x
k
L) = 0 and there is N
3
() so that
|

1kN
2
()
a
nk
(x
k
L)| < /3 for n N
3
(). We can combine the
three /3-estimates provided that n max
i=1,2,3
N
i
().
6
Example: Divergence of Fourier series.
We consider the partial sums of the Fourier series for a continuous
1-periodic function, given by
S
n
f(a) =
n

k=n

f
k
e
2ika
where

f
k
=
_
1
0
f(t)e
2ikt
dt .
Fix a. We show that there is a 1-periodic continuous function for
which the sequence S
n
f(a) does not converge
3
and is in fact unbounded.
Replacing f with f( a) we may assume a = 0.
Let
D
n
(x) =
n

k=n
e
2ikx
= 1 + 2
n

l=1
cos(2lx),
the Dirichlet kernel. After summing geometric series (or, alternatively,
applying trigonometric identities) one computes
D
n
(x) =
sin((2n + 1)x)
sin(x)
.
We consider the linear functionals
n
acting on the space of 1-periodic
continuous functions (with the sup-norm) by

n
(f) = S
n
f(0) =
_
1
0
D
n
(t)f(t)dt.
Clearly |S
n
f(0)|
_
1
0
|D
n
(t)|dtf

so that
n

_
1
0
|D
n
(t)|dt. In
fact, one can check that equality holds in the last estimate. Note that
for g
n
(t) = sign (D
n
(t)) equality is achieved but g
n
is not continuous.
However it is straightforward to construct a sequence of 1-periodic con-
tinuous functions g
n,j
so that |g
n.j
(x)| 1 for all x, g
n,j

= 1 and
g
n,j
converges to g
n
almost everywhere. Thus
_
1
0
D
n
(t)g
n,j
(t)dt
_
1
0
D
n
(t)g
n
(t)dt
and we conclude

n
=
_
1
0
|D
n
(t)|dt.
A straightforward calculation shows that
_
1
0
|D
n
(t)|dt c log n so
that {
n
} is unbounded. By the uniform boundedness principle
there exists (and by examining and slightly modifying Hahns proof
one can indeed construct) a continuous 1-periodic function f for which
sup
nN
|S
n
f(0)| = .
3
Such an example was rst given in 1876 by du Bois Reymond.