BANACH SPACE THEORY

I this appendix we introduce basic concepts of Functional Analysis

To begin with we recall basic facts about metric spaces. No proofs are
provided for these.
Theorem [CANTOR]
If C
n
is a sequence of closed sets in a complete metric space with diameter
of C
n
0 as : and C
n+1
C
n
\: then there is a unique point that
belongs to each of the sets C
n
. Conversely if this properety holds in a metric
space (A, d) then (A, d) is complete.
Theorem
A set C in a metric space (A, d) is compact if and only if it is sequentially
compact (i.e. every sequence in C has a subsequence converging to some point
of C) if and only if it is complete and totally bounded. [C is called totally
bounded if for each c 0 we can nd a nite number of open balls of radius c
whose union conatins C].
Remark: compact sets are closed and bounded but the converse is false in a
general metric space. However, the converse is true in Eucliden spaces.
Theorem
If is a connected set in a metric space (A, d) then any set 1 with
1

is connected. Union of two connected sets is connected if they have a
point in common.
Theorem
If an open set in R
J
or C
J
is connected then it is polygonally connected.
Any convex set in R
J
or C
J
is connected.
NORMED LINEAR SPACES
By a norm on a vector space \ over 1(= R or C) we mean a function
r |r| from \ to [0, ) such that |r| = 0 = r = 0, |r +j| _ |r| + |j|
and |cr| = [c[ |r| whenever c 1 and r, j \ . A vector space with a
norm is called a normed linear space (nls). We provide a nls with the metric
d(r, j) = |r j|. If this makes \ a complete metric space then we call \ a
Banach space.
In the following examples the metric dened are treated as default metrics.
1. R
J
or C
J
is a Banach space with the usual norm (|r| =
q
[r
1
[
2
+[r
2
[
2
+... +[r
J
[
2
).
2. The space C[0, 1] of all continuous (real or complex functions) on [0, 1]
with the metric d(), q) = sup
0r1
[)(r) q(r)[ is a Banach space.
3. The space 1

(j) is a Banach space for any positive measure j and any

j [1, ).
4. The space |

of all sequences a
n
of real (or complex) numbers satisfying
the condition
1
X
n=1
[a
n
[

< with the metric d(a

n
, /
n
) = (
1
X
n=1
[a
n
/
n
[

)
1/
is a Banach space.
1
Remark: there are fundamental dierences between nite dimentional and
innite dimensional nlss. We rst investigate some of these dierences.
Theorem
The closed unit ball r : |r| _ 1 is compact i \ is nite dimensional.
We need the following lemma:
Lemma
A nite dimensional subspace of a nls is necessarily closed.
Proof of the lemma: let ' be a nite dimensional (f.d.) subspace of \
with basis r
1
, r
2
, ..., r
n
. Consider a seuqence r
()
=
n
X
I=1
c
I,
r
I
converging
to r. Let c

=
q
[c
1,
[
2
+[c
2,
[
2
+... +[c
n,
[
2
). If c

then it is clear
that
1
oj
r
()
0. However
1
oj
r
()
=
n
X
I=1
(c
I,
,c

)r
I
and the numbers (c
I,
,c

)
are bounded. Using the fact that bounded sequences in 1
J
have convergent
subsequences we get 0 =
X
c

for some c
1
, c
2
, ..., c
n
with
X

c
2
I

= 1. This
contradicts the fact that r
I
is a basis and hence c

9 . This implies that

a subsequence of the sequence (c
1,
, c
2,
, ..., c
.n
) in 1
J
is (bounded and hence
a further subsequence is) convergent. This implies that r =
n
X
I=1
c
I
r
I
for some
c
1
, c
2
, ..., c
n
. Hence ' is closed.
Proof of the theorem: assume that the closed unit ball is compact. Then
r : |r| _ 1
n
[
I=1
1(r
I
, 1,2) for some nite set r
1
, r
2
, ..., r
n
. Let ' be
the vector space spanned by this set. Then r : |r| _ 1 ' + 1(0, 1,2)
' +
1
2
r : |r| _ 1 ' +
1
2
(' +1(0, 1,2)) ' +1(0, 1,4), .... We nd that
r : |r| _ 1 ' + 1(0, 1,2
n
) for each :. Together with the above lemma
this shows that r : |r| _ 1 ' which implies that \ '. This proves one
part and the converse part follows by Bolzano-Weirstrass Theorem.
Theorem
Let ' be an d dimensional subspace of a nls \ and ) : ' 1
J
be a
vector space isomorphism. Then ) is a homeomorphism [i.e. ) and )
1
are
both continuous].
Proof: we show that any linear map from ' into a nls \ is continuous.
This would prove continuity of both ) and )
1
. Let r
I
be a basis for ' and
r
()
=
n
X
I=1
c
I,
r
I
converging to 0. If we prove that )(r
()
) 0 then it would
follow that ) is continuous at 0. However, by linearity of ) we can concludew
that it is continuous everywhere. It suces to show that c
I,
0 as , for
each i. This can be proved a way very similar to the proof of above lemma so
we omit the details.
Theorem
Every linear map from \ to 1 is continuous i \ is f.d..
2
Proof: suppose \ is innite dimensional. Then there exists a sequence r
n

of independent unit vectors in \. Let )(r

n
) = :\: and extend ) to a linear map
on \ by linearity. [We can extend r
n
to a basis and dene ) to be 0 on the
basis elements other than th r
0
n
:]. Then )(
1
p
n
r
n
) 90 though
1
p
n
r
n
0.
This proves one part of the theorem. The other part has already been proved.
For proving the basic theorems on Banach spaces we will have to exploit
completeness and we do that using the following:
Theorem (Baire Category Theorem)
A complete metric space cannot be expressed as the union of a countable
number of nowhere dense sets.
[A nowhere dense set is a set whose closure has no interior points].
Proof: If possible let A = '
n
where (A, d) is a complete metric space and

n
is a sequence of nowhere dense sets in A. There is no loss of generality in
assuming that each
n
is closed (since
0
n
: can be replaced by their closures).
Consider any open set \ in A. Let \
1
be a non-empty set with diameter < 1,2
such that

\
1
\ [We can take \
1
to be an open ball of small radius]. Since \
1
is not contained in
1
there is a point r
1
\
1

1
. Let \
2
be a non-empty open
set whose closure is contained in \
1

1
whose diameter does not exceed 1,4.
Then \
2

2
is a non empty open set so there is a non-empty open subset \
3
with diameter < 1,8 such that

\
3
\
2

2
. Procedding thus we get non-empty
open sets \
n
, : _ 1 such that diameter of \
n
is less than
1
2
n
and

\
n+1
\
n

n
.
By the thoerem of Cantor stated above there is a point r in

\
n
. This point
must belong to some
n
and since

\
n+1
\
n

n
it follows that r ,

\
n+1
for
this :. We have arrived at a contradiction and hence the theorem is proved.
Remark: above proof actullay shows that if
n
is a sequence of nowhere
dense sets in A then A '
n
is dense!. In fact if a nonempty open set \ does
not intersect A '
n
then we can construct the open sets \
n
in such a way
that

\
n
\\: in which case the point r constructed would be in
_
\. Then r
would be in some
n
and we get a contradiction as before.
As an interesting consequence of this theorem we show that Q cannot be
expressed as a countable intersection of open sets in R. [A countable intersection
of open sets is called a G
o
set]. Indeed if Q = \
n
then \
n
is dense for each
: because it contains RQ. This implies that
n
= (\
n
)
c
is nowhere dense and
since '
n
' r
1
, r
2
, ... = R if r
1
, r
2
, ... is an ennumeration of the rationals we
get a contradiction from Baire Category Theorem.
Linear maps between nlss and their continuity:
Theorem
Let A and 1 be nlss and T : A 1 be linear. Then T is continuous i
there is a constant C such that |T(r)| _ C |r| \r A. Further, the inmum
of all constant C satisfying this inequality is equal to sup|T(r)| : |r| _ 1.
Proof: if T is continuous and if there is no C then we can nd r
n
with
|r| _ 1 but |T(r
n
)| :|r
n
| . Clearly r
n
,= 0\:. Let j
n
=
1
p
n
rn
krnk
. It follows
3
that j
n
0 but T(j
n
) 90. Conversely if C exists as stated then T is obviously
continuous at 0 and continuity at other points follows by linearity.
Denition: |T| = sup|T(r)| : |r| _ 1.
Note that |T(r)| _ |T| |r|. We leave it as an exercise to the reader to
show that the vector space of all linear maps from A into 1 is itself a nls under
this norm!. This space is denoted by 1(A, 1 ).
If T and o are continuous linear maps between nlss such that T o is dened
than |T o| _ |T| |o| .
Theorem [Open Mapping Theorem]
If T is a continuous linear map from one Banach space A onto another
Banach space 1 then T is an open map i.e. T(\ ) is open in 1 for every open
set \ in A.
Corollary
If T is a bijective continuous linear map from one Banach space A onto
another Banach space 1 then the inverse of T is automatically continuous.
Proof of Open Mapping Theorem: let \ be open in A and j T(\ ). We can
write j = T() for some \. There exists r 0 such that + 1(0, r) \ .
If we can show that T(1(0, r)) contains some open ball 1(0, :) in 1 we can
conclude that j+1(0.:) T()+T(1(0, r)) T(\ ) proving that every point of
T(\ ) is an interior point. It suces therefore to show that T(1(0, 1)) contains
some open ball around 0 in 1. For this we rst prove that the closure 1 of
T(1(0, 1)) contains some open ball around 0. Since 1 =
1
[
n=1
T(1(0, :)) we see
from Baire Category Theorem that there is an integer : such that :1 contains
an interior point j
0
. We may assume that j
0
T(1(0, :)) [Why?]. Now 0 is
an interior point of :1 j
0
. Since T(1(0, :)) j
0
T(1(0, 2:)) we get
:1j
0
2:1. Hence 0 is an interior point of 2:1 which implies that 0 is an
interior point of 1. Now let . 1 with |.| < c where c is so small that 1(0, c)
1. By denition of 1 we see that r
1
with |r
1
| < 1 and |. T(r
1
)| < c,2.
Repeating the argument with . replaced by 2(. T(r
1
)) we get n
2
such that
|n
2
| < 1 and |2(. T(r
1
)) T(n
2
)| < c,2. Let r
2
=
1
2
n
2
so that |r
2
| <
1
2
and |(. T(r
1
)) T(r
2
)| < c,2
2
. We can now repeat the argument with 2(.
T(r
1
)) replaced by 2
2
[(. T(r
1
)) T(r
2
)] and so on. We get, by induction, a
sequence A
n
with |r
n
| <
1
2
n
and |. T(r
1
) T(r
2
) ... T(r
n
)| < c,2
n
.
Let r =
1
X
n=1
r
n
which is convergent series by completeness of A. By continuity
of T we get . = T(r). Since |r| < 1 we have proved that 1(0, c) T(1(0, 1)).
This completes the proof.
Theorem [TheClosed Graph Theorem]
Let T be a linear map from one Banach space A into another Banach space
1. Then T is continuous if its graph (r, Tr) : r A is a closed subset of A1
when A 1 is metrized by d((r
1
, j
1
), (r
2
, j
2
)) =
p
d
2
(r
1
, r
2
) +d
2
(j
1
, j
2
).
Remark: the converse is true and trivial to check. The metric on A 1
4
can be replaced by several other standard metrics. In fact, if j is any met-
ric on R
2
the we can replace d above by the metric 1((r
1
, j
1
), (r
2
, j
2
)) =
j(d(r
1
, r
2
), d(j
1
, j
2
)). We leave it as an exercise to the reader to show that
all norms on Euclidean spaces are equivalent in the sense that open sets for any
two of them are the same.
Proof: Dene a new norm on A by |r|
1
= |r|+|T(r)| . It is easy to verify
that this is a norm on A. We claim that this is a complete norm. To see this let
r
n
be a Cauchy sequence for this norm. Then r
n
is Cauchy in A with the
original norm, so it converges to some r in (A, ||). Also T(r
n
) is Cauchy in 1
so it converges to some j 1 . Now the sequence (r
n
, T(r
n
)) lies in the graph
of T and it converges to (r, j). The hypothesis now shows that j = T(r). But
then r
n
converges to r in (A, ||
1
). This proves the claim. We now consider
the identity map 1 of A as a map from the Banach space (A, ||
1
) into (A, ||).
It is obvious that this map is continuous. The Open Mapping Theorem now
shows that its inverse is continuous. This implies that there is a nite constant
C such that |r|
1
_ C |r| \r A. This gives us |T(r)| _ (C 1) |r| proving
that T is continuous.
Theorem [Uniform Boundedness Principle alias Banach-Steinhaus Theorem]
Let T
n
be a sequence of continuous linear maps from a Banach space A
into a nls 1 . If sup
n
|T
n
(r)| < for each r A then sup
n
|T
n
| < i.e.
sup
n
sup
krk1
|T
n
(r)| < .
Proof: we have A =
1
[
=1
r A : |T
n
(r)| _ \:. By Baire Category
Theorem we can nd an integer / and an open ball 1(r
0
, r) in A such that
1(r
0
, r) r A : |T
n
(r)| _ /\:. [We have used the obvious fact that the
set on the right side of this inclusion is a closed set]. But then |T
n
(r
0
)| _ /\:
and |r| _ 1 implies r
0
+
:
2
r 1(r
0
, r) which implies

T
n
(r
0
+
:
2
r)

_ /\:.
These two inequalities together yield |T
n
(r)| _
2
:
(/ + /)\: for all r in the
closed unit ball of A. This completes the proof.
This theorem is extremely useful. Our rst application is the following:
Theorem
Let T
n
be a sequence of continuous linear maps from a Banach space A
into a nls 1 and assume that T(r) = lim
n!1
T
n
(r) exists for each r in A. Then
the linear map T is necessarily continuous.
Proof: this follows immediately from Uniform Bouddness Principle; in fact,
|T| _ sup
n
|T
n
| < .
As another interesting application let us prove the following result:
if a
n
is a sequence of complex numbers such that
X
a
n
/
n
converges when-
ever
X
[/
n
[ < . Then a
n
is necessarily bounded.
For a proof dene T

: |
1
C by T

(/
n
) =

X
=1
a
I
/
I
. We claim that T

is
5
a continuous linear functional on |
1
with |T

| = sup
1
[a

[. The result follows

immediately from Uniform Bouddness Principle once the claim is established.
Of course, [T

(/
n
)[ _ sup
1
[a

[
1
X
=1
[/

[ which shows that T

is a continuous
linear functional and its norm does not exceed sup
1
[a

[. To prove that |T

| _
sup
1
[a

[ dene an element /

of |
1
as follows: /

= 0\, ,= i and /
I
=

oi
joij
if
a
I
,= 0, 1 if a
I
= 0, i being any xed integer between 1 and . Then T

(/
n
) =
[a
I
[ and
1
X
=1
[/

[ = 1. This proves that |T

| _ [a
I
[ for each i 1, 2, ..., .
The proof is now complete.
INNER PRODUCT AND HILBERT SPACES
An inner product on a vector space \ over 1 is a function <, from \ \
into 1 such that
i) < ar +/j, . = a < r, . +/ < j, .
ii) < r, j is the complex conjugate of < j, r
iii) < r, r _ 0 and < r, r = 0 if and only if r = 0.
If A is a nls and if its norm is related to an inner product <, by |r|
2
=<
r, r \r then we call A an inner product space. If the norm is also complete
then we call A a Hilbert space.
Examples:
1) |
2
is a Hilbert space with the inner product < a
n
, /
n
=
1
X
n=1
a
n

/
n
.
2) 1
2
(j) is a Hilbert space with the inner product < ), q =
Z
)

qdj.
Theorem
A nls A is an inner product space if and only if parallelogram law holds,
i.e.
|r +j|
2
+|r j|
2
= 2|r|
2
+|j|
2
\r, j A
Proof: Writing the norm in terms of the inner product one sees easily that
parallelogram law holds in an inner product space. Given a nls in which the
parallelogram law holds we dene
< r, j =
kr+k
2
krk
2
+Ikr+Ik
2
IkrIk
2
4
. It is clear that < r, r = |r|
2
.
The fact that <, is indeed an inner product is left as an exercise.
[Thus, a nls is an inner product space i every two dimensional subspace is
so!]
Denition: two vectors r and j in an inner product space are orthogonal if
< r, j = 0.
Theorem
Let C be a closed and convex set in a Hilbert space (H.S.) H. Then inf
r2c
|r|
is attained at exactly one point.
6
[A set C is convex if ar + (1 a)j C whenever r, j C and 0 _ a _ 1].
Proof: let r
n
C and |r
n
| c where c = inf
r2c
|r|. Then

rn+rm
2

_ c
and the parallogram law gives

rnrm
2

2
= 2

rn
2

2
+ 2

rm
2

rn+rm
2

2
_
2

rn
2

2
+ 2

rm
2

2
c
2

1
2
c
2
+
1
2
c
2
c
2
= 0. Thus r
n
is Cauchy. If
r = limr
n
then |r| = c. This proves existence. Uniqueness follows from the
fact that if the inmum is attained at r as well as at j then |r +j|
2
+|r j|
2
=
2|r|
2
+|j|
2
= 4c
2
and |(r +j),2| _ c so |r j|
2
_ 0.
This theorem is very useful. In particular it is used in the study of projec-
tions.
Theorem
Let H be a H.S. Let ' be a closed subspace of H. For any point r H any
point j ' the following are equivalent:
a) |r j| = inf
:21
|r .|
b) r j is orthogonal to ', i.e. it is orthogonal to every element of '.
Further, for any point r H, there is a unique point j ' such that above
equivalent conditions hold.
Proof: the last part is immediate from previous theorem with C = r +'.
To prove the equivalence of a) and b) note that if b) holds then, for any
. ', |r .|
2
= |(r j) + (j .)|
2
= |r j|
2
+ |j .|
2
(by b)) and
hence |r .|
2
_ |r j|
2
proving a). Conversely if a) holds then |r j|
2
_
|r (j .)|
2
= |(r j) +.|
2
and so 0 _ |.|
2
+2 Re < rj, . . Replacing
. by c. and dividing by c we get 0 _ c |.|
2
+ 2 Re < r j, . . Letting c 0
we get 0 _ Re < r j, . Replacing . by . we get the reverse inequality.
Thus Re < r j, . = 0. Replacing . by i. we get Im < r j, . = 0. This
proves b)
Denition: if ' is a closed subspace of a H.S. H and r H then the unique
point j of ' satisfying the equivalent conditions above is called the projection
of r on '. We get a map 1 : H H which takes any point r to its projection
j. The map 1 is called the projection of H on '.
Some obvious properties of 1 are: 1(r) = r if r ' and |1(r)| _ |r| \r.
[Since < r1(r), 1(r) = 0 we get |r|
2
= |r 1(r)|
2
+|1(r)|
2
_ |1(r)|
2
].
Theorem
The map 1 above is linear
Proof: let r
1
and r
2
H. If . ' then r
1
1(r
1
) and r
2
1(r
2
) are
orthogonal to . and so is their sum. Thus j = 1(r
1
)+1(r
2
) ' and r
1
+r
2
j
is orthogonal to '. This implies that 1(r
1
+ r
2
) = j = 1(r
1
) + 1(r
2
). An
easy verication shows that 1(ar) = a1(r) if a 1.
Notation: we write 1
1
for the map 1 above to emphasize its dependence
on '.
Thus 1
1
is a continuous linear map : H H with |1| _ 1.
Denition: the orthogonal complement '
?
of a subspace ' of H is dened
as the set of all vectors orthogonal to '.
'
?
is a closed linear subspace. If r H then r = 1(r) +(r 1(r)) where
1 = 1
1
. Note that 1(r) ' and r 1(r) '
?
. Thus any vector r is the
7
sum of a vector in ' and a vector in '
?
. This decomposition is unique, as the
reader can easily verify. This decomposition also shows that the projection of
r on '
?
is nothing but r 1(r). Thus, 1
1
? = 1 1 where 1 is the identity
operator.
Theorem [Hahn-Banach]
Let ' be a linear subspace of a H.S. H and ) : ' 1 be a continuous
linear map. Then there exists a continuos linear map 1 : H 1 such that
1(r) = )(r)\r ' and |1| = |)| .
[The theorem is true if the H.S. H is replaced by any nls. We will not prove
this result here. Any book on Functional Analysis will contain a proof].
Proof: just dene 1(r
1
+r
2
) = )(r
1
) if r
1
' and r
2
'
?
.
Theorem [Riesz]
Let H be a H.S. and ) : H 1 be a continuous linear map. Then there is
a unique vector n such that )(r) =< r, n \r H
Proof: let ' = r : )(r) = 0. ' is a closed subspace. If '
?
= 0
then H = ' + '
?
= ' so )(r) = 0\r. In this case we can take j = 0. If
'
?
,= 0 then we pick a vector j in '
?
such that |j| = 1. Now, raj ' if
a =
}(r)
}()
. Since j '
?
, < r aj, j = 0. Thus < r, j = a |j|
2
= a =
}(r)
}()
.
Thus, )(r) =< r, cj \r where c is the conjugate of )(j). Uniqueness part is
obvious.
ORTHONORMAL BASES
A set r
I
in H is called orthogonal if < r
I
r

= 0 for i ,= ,. It is called
orthonormal if it is orthogonal and |r
I
| = 1\i.
Theorem
Let r
I
be an orthonormal set in H. Then the following are equivalent:
1) r
I
be a maximal orthonormal set
2) < r, r
I
= 0\i implies r = 0
3) r =
X
< r, r
I
r
I
\r H
4) |r|
2
=
X
[< r, r
I
[
2
\r H
5) < r, j =
X
< r, r
I
< r
I
, j \r, j H
Before proving this theorem we have to give a meaning to a sum of the type
X
I21
c
I
where c
I
C for all i and 1 is any set, not necessarily countable.
We say that such a sum
X
I21
c
I
is convegent if i 1 : c
I
,= 0 is atmost
countable and the series
X
ci6=0
[c
I
[ < . This is so if and only if sup
J
X
I2J
[c
I
[ <
where 1 varies over all nite subsets of 1. [If the last condition holds then
i 1 : [c
I
[
1
n
is a nite set for each : and hence i 1 : c
I
,= 0 is atmost
countable. The equivalence of the two notions of summability is quite obvious
from this].
Proof of the theorem:
8
the following implications are all very easy: 3) = 5) = 4) = 1) = 2).
It reamins only to show that 2) = 3). To prove this we rst prove Bessels
inequality:
For any orhonormal set j
I

I21
and any j H,
X
[< j, j
I
[
2
_ |j|
2
.
To see this let 1 be a nite subset of 1 and ' be the closed subspace
generated by j
I

I2J
. [This is the closure of the linear span of the vectors j
I
].
Then |1
1
j| _ |j|. If we show that 1
1
j =
X
I2J
< j, j
I
j
I
then the proof
would be complete. Since
X
I2J
< j, j
I
j
I
' we only have to show that
j
X
I2J
< j, j
I
j
I
'
?
. But j
X
I2J
< j, j
I
j
I
l j

for each , 1 and we

are done.
We have proved Bessels inequality and hence r :< r, r
I
,= 0 is atmost
countable.
To show that 2) = 3) we rst show that the sum
X
< r, r
I
r
I
exists.
For any nite subset 1 of 1 we have

X
I2J
< r, r
I
r
I

2
=
X
I2J
[< r, r
I
[
2
and
since
X
I21
[< r, r
I
[
2
< and the space H is complete we see that the series
X
< r, r
I
r
I
is convergent. If its sum is j then < r j, r
I
= 0\i which
implies (by 2)) that r = j. This proves 3).
Exercise: show that the ve conditions above are also equivalent to
6) the vector space spanned by r
I
is dense in H.
Denition: an orthonormal set r
I
is called an orthonormal basis (ONB)
or a complete orthonormal set (COS) if the equivalent conditions of previous
theorem are satised.
Examples:
1) The usual bases in Euclidean spaces [c
1
= (1, 0, ..., 0), c
2
= (0, 1, 0, ..., 0)..., c
n
=
(0, ..., 0, 1) in 1
n
] are orthonormal bases
2) Let H = |
2
and c
1
= (1, 0, ...), c
2
= (0, 1, 0, ...), .... Then c
n
is an ONB
3) It can be shown that if j is
1
2t
time the Lebesgue measure on [, ] and
H = 1
2
(j) then the set c
Inr
: : Z is an ONB. An expansion of the type
r =
X
< r, r
I
r
I
with respect to this basis is called a Fourier series. The
theory of Fourier series deals with convergence of the series
X
< r, r
I
r
I
in dierent senses; instead of looking at convergence in the norm of 1
2
(j) we
can look at pointwise convergence, uniform convergence, almost everywhere
convergence etc.
When do we consider two nlss as identical? If A and 1 are two nlss and if
we have a bijective linear map T of A onto 1 such that |T(r)| = |r| \r A
we call T an isometric isomorphism and we say that the spaces A and 1 are
isometrically isomorphic. What about Hilbert spaces? Do we have to impose an
additional condition that T preserves inner products? Fortunately the formula
9
< r, j =
kr+k
2
krk
2
+Ikr+Ik
2
IkrIk
2
4
show that an isometric isomorphism
necessarily preserves inner products!
Dual spaces: let A be a nls. Let A

be the space of all continuos linear func-

tionals on A. If r

then teh norm of r

is dened by |r

| = sup[r

(r)[ :
|r| _ 1. A

is a Banach space (even if A is not complete!) and we call it the

dual of A.
Identication of duals of standard spaces:
If A = 1
J
then A

can be identied with T itself. Any element r

in the
dual is of the type r

(c
1
, c
2
, ..., c
J
) =
J
X
=1
c

for some (/
1
, /
2
, ..., /
J
) 1
J
and
we can use the map r

(/
1
, /
2
, ..., /
J
) to identify A

with A. The details are

left as an exercise.
More generally, the theorem of Riesz shows that the dual of any H.S. can be
identied with itself.
We now prove that if j is a nite measure on (, T) then the dual of 1

(j)
can be identied with 1
j
(j) where 1 < j < and =

1
.
Let (1

(j))

. Dene i(1) = (1
J
) \1 T. i is a complex valued
nitely additive function on T.
If 1
n
| O then 1
Jn
0 in 1

(j) and hence i(1

n
) 0. This property and
nite additivity together imply countable additivity as the reader can easily
verify. Thus i is a complex measure. Clearly, i << j. Ny RNT we can nd an
integrable function q such that i(1) =
Z
J
qdj\1 T. For any simple function
) we get ()) =
Z
)di =
Z
)qdj. We now show that q 1
j
(j). If ) is
as simple function then

Z
)qdj

_ || |)|

where |)|

is the norm of ) in
1

(j). Using Holders inequality we can easily extend this to all ) 1

(j).
Taking ) =

q [q[
q
p
1
we get
Z
[q[
j
dj _ || (
Z
[q[
j
dj)
1/
which simplies to
|q|
j
_ ||. We have proved the existence of a function q 1
j
(j) such that
()) =
Z
)qdj\) 1

(j) and |q|

j
_ ||. On the other hand, if q 1
j
(j)
then ()) =
Z
)qdj denes a continuous linear functional on 1

(j) with || _
|q|
j
as one can see easily from Holders inequality. Combining these facts we
get the desired identication of (1

(j))

with 1
j
(j).
Remark: the result is true if j is not nite. The general case is more in-
volved and a proof can be found in Hewitt and Strombergs Real And Abstract
Analysis.
The space 1
1
(j) : consider the class of all measurable functions ) such that
[)[ _ C a.e. for some nite constant C. [We say ) is essentially bounded in
this case]. Let 1
1
(j) be the set of equivalence classes of essentially boundd
functions under equality a.e.. Dene |)| to be the smallest constant C such
10
that [)[ _ C a.e.. Then 1
1
(j) becomes a Banach space. By an argument
similar to the above we can show that (1
1
(j))

can be identied with 1

1
(j).
Note that (|

= |
j
if 1 < j < and =

1
; also (|
1
)

= |
1
. These are
special cases of above results where j is taken to be the counting measure.
Theorem [Riesz representation Theorem]
Let A be a locally compact metric space and let 1 be the space of all
continuous maps from A into 1 with compact support. Make 1 into a nls
by the norm |)| = sup
r2
[)(r)[. Then the dual of 1 can be identied via the
equation ()) =
Z
)dj with the space of all complex regular Borel measures j
on A with the norm |j| = [j[ ().
Locally compact means each point has a neighbourhood whose closure is
compact. The theorem is also true for locally compact Hausdor spaces with
the same proof. The proof is long and somewhat tedious. First time readers
amy skip teh proof.
Proof: we rst consider a positive linear functional on C
c
(A), i.e. a
linear map: C
c
(A) R such that ()) _ 0 for any continuous function ) with
compact support having values in [0, ). Inthis case we will show that there is
a unique positive nite regular Borel measure j such that ()) =
Z
)dj\)
C
c
(A).
[Note that if ) C
c
(A) is real valued with |)| = sup
r2
[)(r)[ _ 1 then
1 _ ) _ 1 so () + 1) _ 0 and (1 )) _ 0. Hence [())[ _ 1. If ) is
complex valued then we can apply this argument to the real and imaginary parts
of ). Thus, any positive linear functional on C
c
(A) is necessarily a bounded
linear functional. We will show later that any continuous linear functional is a
linear combination of positive linear functionals].
Let us prove that we can have at the most one positive nite regular Borel
measure j such that ()) =
Z
)dj\) C
c
(A). If possible, let i be another
measure with the stated properties. Let C be a compact subset of A. If \ is
open and C \ then there is a continuous function ) : A [0, 1] such that
)(r) = 1\r C and )(r) = 0\r \
c
. We have j(C) =
Z
c
)dj _
Z
)dj =
Z
)di _
Z
\
c
1di = i(\
c
). Regularty of i gives j(C) _ i(C) and by symmetry
we get i(C) = j(C) for every compact set C. Using regularity again we get
j() = i() for any Borel set .
For the existence part we start with the deninition j(\ ) = sup()) : ) -
\ where ) - \ means ) C
c
(A), 0 _ ) _ 1 and support of ) is contained
in \ . Let j

() = infj(\ ) : \, \ open for any subset of A. Clearly,

j

(\ ) = j(\ ) if \ is open. We now dene a o eld L= as follows: L== 1

11
A : 1 C L=
0
whenever C is compact} where L=
0
= 1 A : j

(1) <
and j

(1) = supj(C) : C 1, 1 compact. We shall now verify that L=

is a o eld containing all Borel sets, j

is a regular measure on L= and that

()) =
Z
)dj\) C
c
(A).
From the denition we get j

() _ j

(1) if 1. We now prove that j

is
an outer measure. For this we have to prove countable subadditivity; let 1
n
A
for each :. We have to show that j

('1
n
) _
X
j

(1
n
). We begin by observing
that j

(\
1
' \
2
) _ j

(\
1
) + j

(\
2
) if \
1
, \
2
are open. Let q - \ = \
1
' \
2
.
Claim: /
1
, /
2
such that /
I
- \
I
(i = 1, 2) and /
1
(r) +/
2
(r) = 1 on support of
q.
Proof of the claim: let C = support of q. For each r C there is a neigh-
bourhood l
r
such that the closure of l
r
is a compact subset of one of the two
sets \
1
and \
2
(depending on whether r \
1
or r \
2
). Let l
r1
, l
r2
, ..., l
rm
cover C. Let H
1
be the union of the sets

l
ri
that are contained in \
1
and dene
H
2
similarly. There exists q
1
, q
2
C
c
(A) such that H
I
- q
I
- \
I
(i = 1, 2).
Here H
I
- q
I
means that q
2
C
c
(A), 0 _ q
I
_ 1 and q
I
= 1 on H
I
. [Exis-
tence of q
1
and q
2
is trivial in the case of metric spaces; in general it is the
Urysohns Lemma. We refer to Simmons book for a proof]. Now let /
1
= q
1
and /
2
= (1 q
1
)q
2
. At each ponit of C either q
1
= 1 or q
2
= 1 and this gives
/
1
+/
2
= 1. This proves the claim.
[Remark: /
1
, /
2
is a partition of unity subordinate to \
1
, \
2
. This ar-
gument works for any nite collection \
1
, \
2
, ..., \
n
of open sets whose union
contains a given compact set
_
C].
Now, /
I
q - \
I
and q = q/
1
+ q/
2
so (q) = (q/
1
) + (q/
2
) _ j

(\
1
) +
j

(\
2
). Since this holde whenever q - \ = \
1
' \
2
we get j

(\
1
' \
2
) _
j

(\
1
)+j

(\
2
). To prove that j

('1
n
) _
X
j

(1
n
) for any sets 1
1
, 1
2
, ... we
may assume that the right side is nite. Thus j

(1
n
) < \:. By denition,
for each : and c 0 we can nd open sets \
n
such that j

(1
n
) +c,2
n
j(\
n
)
and 1
n
\
n
. Let \ = '\
n
and let ) - \. Then, since ) has compact support,
) - \
1
' \
2
' ... ' \
|
for some /. Therefore, ()) _ j

(\
1
' \
2
' ... ' \
|
) _
j

(\
1
) +j

(\
2
) +... +j

(\
|
) < j

(1
1
) +j

(1
2
) +... +j

(1
|
) +c
_
X
j

(1
n
) + c. Taking supremum over ) we get j(\ ) _
X
j

(1
n
) + c
and this implies j

('1
n
) _
X
j

(1
n
). Our next aim is to show that every
compact set belongs to L=
0
and that j

(C) = inf()) : C - ) if C is compact.

For this let C - ) and 0 < c < 1. Let \
o
= r : )(r) c. Then C \
o
and cq _ ) whenever q - \
o
. Hence j

(C) _ j(\
o
) = sup(q) : q - \
o
_
1
o
()). Letting c 1 we get j

(C) _ ()). In particular j

(C) < . By
denition of L=
0
we get C L=
0
. Let c 0. By denition, j

(C) + c j(\ )
for some open set \ containing C. We can nd ) such that C - ) - \ . Then
()) _ j(\ ) < j

(C) + c. This and the inequality j

(C) _ ()) whenever C

- ) together yield the relation j

(C) = inf()) : C - ).
Next we show that for any open set \, j(\ ) = supj

(C) : C \, C
compact. This would show that every open set \ with j

(\ ) < belongs to
12
L=
0
.
13