<br>

Chapter3

Hilbert Spaces
3.1 Orthogonality
Hilbert space theory is a blend of Euclidean geometry and modern analysis. The
inner product between vectors affords us the notion of orthogonality. This gives
us access to geometric ideas such as Pythagoras' Theorem (yes, really) and a best
approximation property which formalizes the idea of 'dropping a perpendicular'.
This property, together with the notion of an orthonormal basis, will lead us to
discover the Fourier series of an L' function. These ideas were originally studied
in the context of integral equations and are the historical roots of all of Functional
Analysis.

David Hilbert (1862-1943) is perhaps best known for the twenty-three

problems he posed in the International Congress of Mathematicians (1CM)
in Paris in 1900. These problems speak to the incredible range of knowledge
that Hilbert possessed and have been important signposts for mathematics
in the 20th century. The Paris address was sandwiched between his immense
work on axiomatizing geometry (published in 1899) and his work on integral
equations (1904-1910) which laid the foundations for modern analysis.
His mentorship of extraordinary students, including Hermann Weyl, Ernst
Zermelo, Max Dehn, and many others, has only enriched his mathematical
legacy.

71
<br>

72 Fwnctional Analysis

We begin with some definitions and notation. Throughout this section, we will
use the letter H to denote a Hilbert space and (,) to denote the inner produc
on H.

3.1.1 DEFINITION

) We say that two elements x,y E H are orthogonal if (x, y) = 0. If this

happens, we write x y. l
(ii) For two subsets A,
BCH, we write AI Bif xly for all x E A and y E B.

(ii) For any set

Ac H, write A:= {xE H:xly for all y
E A}. If A = {x},
then we simply write x
instead of fx}.

3.1.2 REMARK Let A be any subset of H.

(i) For each y E H, the linear functional p, :H K given by ,(x) := (,y)

is continuous (see Example 2.2.3) and

A= ker(gy).
yEA

Thus regardless of what A is, A is always closed and a subspace of H.

(ii) For any c H, it is easy to check that An Ac {0} and that AC(AL)t.
A

(iii) Note that (A) is always a closed subspace of H so one cannot expect
the equality A = (A-) unless A were also a closed subspace (see
Proposition 3.1.10).

The first collection of results do not require a proof. They simply follow from
the axioms of the inner product. Give it a try!

3.1:3 PROPOSITION Let x, y E H, then

() (Polarization ldentity): ||x +

yll= ||x|+2 Re(x,y + ly|2.
(ii) (Pythagoras' Theorem): If x 1 y, then |x + yll? = x|2 + ly||2.
(i) (Parallelogram law): ||x +
yl| + ||x-y|?= 2(||||2 + lyl?).
The parallelogram law gives us a quick way to determine which of our familiar
normed linear spaces are Hilbert spaces.
<br>

Hilbert Spaces 73

3.1.4 EXAMPLE p<o.

Fix 1

() Let = lP and set x := e1

E e
and y :=e-e2. Then |x+ylp
+
= |2e1|p=2
and |x yllp =|2e2| = 2 and

yl, = | 21/P :
if 1 <p< o,
|x=
1 : if p = 0,
Hence the parallelogram law holds if and only if 8 = 4 22/P, which
happens only when p = 2.
(ii) Let E = LP0, 1] and f(t) :=tand g(t) :=1-t.Then |f +
gllp = 1 and

(-p)77 : if 1
Kp< co,
|lf-sll, = ||fl, = lsllp
= 0. : if p

Once again the parallelogram law holds if and only if p = 2.

We have constructed these exanmples by showing that a space is not a Hilbert
space if the parallelogram law is violated. It is an interesting fact that this is the
only obstruction: A Banach space is a Hilbert space if and only if the norm satisfies
the parallelogram law. A proof of this lies buried in Proposition 3.1.3. Specifically,
one can recover inner product from the norm by rewriting the Polarization
the
Identity. Have a look at Exercise 3.5.
3.1.5 DEFINITION Let E be a vector space. A subset A c E is said to be convex if,
for any x, y E A, the set (x, y := {tx +(1- )y :0<t< 1} is contained in A.

There are two kinds of convex sets we will be most interested in: The closed (or
open) unit ball in a normed linear space E and any vector subspace ofE. Also, if A is
€
a convex set,
then A+x is also a convex set for any x E. Therefore when dealing
with (non-empty) convex sets,we may often assume that 0 E A by translating.
As discussed earlier, the next result is central to the subject. Notice the liberal
use of the parallelogram law in the proof.

3.1:6 THEOREM (BEST APPROXIMATION PROPERTY) Let A C H be a non-empty,

€
closed, convex set and x E H. Then there exists a unique vector xo A such that

||x- xo| = d(x, A) = inf{||x

– yll :ye A}.

This vector x is called the best approxination of æ in A.

<br>

Functional Analysis
74

||x-vll

FIGURE 3.1 Best approximation property

Proof. Since d(x, A) = d(0, A- x), replacing A by A-x (which is also convex),
we may assume without loss of generality that x =0.
(i) Existence: By definition there exists a sequence (V) CA such that d i=
d(0, A) = lim, l|yn||. We wish to prove that (y) is Cauchy. By the
parallelogram law,
Yn-Ym
2
Since A is €convex, (Vn +ym)/2 E A, so || (yn +ym) /2|| > d. For e > 0,
choose N N such that ly|2 < d²+e for all n> No. Then for n, > No,
m

we have
2
2e) - =e,
1
Yn-ym + d
2
<(2d
and so ||yn -ym|| << 2Ve for all n, m > No. Thus (yn) is Cauchy and hence
convergent in H. SinceA is closed, there exists xo E such that yn Xo. By
A

the continuity of the norm (Remark 2.1.2), d = lim,o y|| =

||xo|.
(ii) Uniqueness: Suppose xo, X1E A are such that ||xo| =|1|=d. Then (xo+
X1)/2 E and hence
A

1 +
ds;(xo+ x) <(xll ||1) sd
and so |(xo + x1)| =d. The parallelogram law then implies that

2
Therefore xo = X1.
<br>

Hilbert Spaces 75

The main application of the best approximation property is when the convex
set in guestion is a closed subspace. Here the best approximation to a vector is
geometrically visualized as the point obtained by 'dropping a perpendicular' from
the point onto the subspace.
a 17PROPOSITION Lef M < H
closed subspace and x
be a
e H. Then xoE Mis the
best approximationof x in M if and only if (x x0) L M. -
Proof.
o. u) =
If xo is the best approximation of x in M, then we wish to prove
that (x -
0
for all yE M. suffices to prove this when vl = 1, so fix y E Mwith
t
llull = 1 and let a := (x- Xo, y). Then z:= X0+ y E M, so
–
I|x xo||< ||x-z||= ||(*- x0) - «y|l
= ||x– xo||+ |layl|- 2 Re(x - X0,
«y)
= ||*-xo|l|?+ la|lly|2 - 2la|?
= ||x –
xol-
la2.
Hence la' = 0, which implies thatx- x0 L y.
Conversely, suppose (x- xo) M, then we wish to prove that lx xoll =
I -
d(x, M). In other words, wewish to prove that ||x– xo| < for all y E M. - yl

-
For any y€ M, we have (x0 y) E M, so (x- x0) L (x0 - y). By Pythagoras'
Theorem,
x0) + (xo - lxo – yll
|x- yl?=l(a- y)=|x-xo|+ >|l*- xo|2.

Therefore ||x – xo|| <||x-yl| for all y e M.

€
3.1.8 DEFINITION Let H be a closed subspace. For x H, let PM(x) E M
M<
denote the best approximation of x in M. In other words, Pu(*) is the unique vector
inM such that |x – Pu(*)|| = d(x,M). By Proposition 3.1.7, this isequivalent to
- €
requiring that x PM(x) M-. The map PM : H M is called the orthogonal
projection of H onto M (and we will often think of it as a map from H to itself and
denote it by P when the subspace is implicit).
That this projection map is linear is perhaps the most remarkable part of the
next result.
3.1.9 PROPosITION Let P :H M be the orthogonal projection onto a closed subspace
M<H. Then
(i) P is a linear transformation.
<br>

Functional Analysis
76

If
M
{0}, then |P = 1.
(ii) P is bounded and ||Pl| <1.
(iii) PoP=P.
(iv) ker(P) =
M
and Range(P) = M.

Proof.
() Let X1, X2 E H and «E K.Set z = X1 + ux and z P(I1) t aP(*2). We
wish to prove that P(z)
=
Z0. For any y E M,

(Z- Z0,y) = (*1– P(x1),y) +

a(x2-P(x2),y) = 0.
€
Hence z - Z0 E
M. Since Zo M, it follows from Proposition 3.1.7 that
P(z) = Zo
€
(i1) For any x H, x = (x– P(r)) + P(*) and (x– P(x)) L P(x). Therefore by
Pythagoras' Theorem,

||x|| = x- P(r)| + |P(x) || > |P(x)||".

Hence P is continuous and ||Pl < 1. Moreover, if M # {0}, then choose

a non-zero vector y E M. Then P(y) = y (by definition of the best
approximation) and thus||P)l| = ||y. This proves that | P|| =1.
€
(iii) Now if y E M then P(y) = y
(as above). If x M then y = P(x) E M, so
P(P(r)) = P(x).
(iv) If P (x) = 0, then x = X–P(x) M-. Converselv. C

M. Hence ker(P) E

€ €
if x
M

then 0 M satisfies the conditions of Proposition 3.1.7. By

uniqueness of the best approximation, it follows that P(x)
= 0. Hence
ker(P) = M. That Range(P) = M is evident from part (i) and the
definition.

3.1.10 PROPOSITION Let M < H be a

subspace, then (M) = M.

Proof. If x e M, then for anyy e M-, (x, y) = 0. Hence x e (M). Thus

MC(M)+. However, (M) is closed, so Mc(M)+.

Conversely, if x e (M), then let xXo = PM(x) denote the best approximation
of x in M. Then xo and X E
M.
M
x-
However, =
M
(by continuity M

of the inner product), so it follows that (x,x -xo) = 0 and (X0, x

xo) = 0. Thus -
x- xo2= (*- Xo, X - xo) = 0, whence x=X € M.
<br>

Hilbert Spaces 77

Thenext theorem is the first hint that the relationship between closed subspaces
may be recovered from the operator-theoretic
algebra of the corresponding
proiections. From now on, when S and T are two operators, 'ST' will be used to
denote the composition SoT.

3.1.11 PROPOSITION Let M < H be a closed subspace, and is a

P= PM. Since M

closed subspace, we may set Q= P. Then

(i) PQ= QP =0.

(i) P+Q=Iwhere I :

H H
denotes the identity map.

Proof.
(i) If x E H, then Q(x) E M. By Proposition 3.1.9, ker(P) = M
and hence
PQ(x) = 0. Since (M)= M (by Proposition 3.1.10), the fact that P =
0
follows by symmetry.
(ii) If x E H, then (x - P(x)) E M
and Q(x)
€
M-, so

x- P(x) - Q(r) e M.
€
However, x – Q(x) (M-) = M by Proposition 3.1.10 and P(x) E M,
so x- (x) - P(*) E M. Since MnM = {0}, it follows that x- P(x) -
€
Q() = 0. Hence, x = (P+ Q)(1). This is true for any x H and thus
(P+ Q) =1.

3.1.12 DEFINITION Given two Hilbert spaces (H,(',, )H,) and (H2,(',)H;),
there is a natural inner product on the direct product H x H2, given by

(*1,y1). (2, y2)) := (*,x2)H, + (*2, ya)H,:

Under this inner product, one can prove that H x H, is a Hilbert space (see
Exercise 3.12). This is called the Hilbert space direct sum of H and H, and is denoted
by H H2.

The next corollary is an extremely important one. Given a closed subspace of a

Hilbert space H, it says that we can decompose H as a direct sum in a very natural
way. This allows us to build up' the Hilbert space from closed subspaces and is
particularly useful when analyzing operators on Hilbert spaces. This is something
we willdo later in the book. Also, it is worth mentioning that such a theorem does
<br>

Fuunctional Analysis
78

not exist for arbitrary Banach spaces. We

will have more to say on this matter
in
Exercise 5.48 and the problems following it.

closed subspace. Then the map T

:

3.1.13 COROLLARY Let M<Hbe

a
MOM!
given by T(x, y) := x+yis an isometric isomorphism.
be the orthogonal projections onto
:
Proof. Let P H M and Q: H M M

and M-,respectively. Then by Proposition 3.1.11,

) (H) = P(H) + Q(H) = M+ M-.

H=(P+
Furthermore, MNMt = {0}.From this it is clear that T is a linear bijection. Finally,
€
the parallelogram law shows that for any x M and y e M,

|T(*,y)|²=||x+y|? =|l+ l= |l(*,y) RseM.

Thus T is isometric as well.

The proof of the next corollary is an easy consequence and is relegated to the
exercises.

3.1.14 COROLLARY If M <

H is any subspace, then M is dense in H if and only if
M
= {0}.

3.2 The Riesz Representation Theorem

For a finite dimensional vector space, the notion of a dual basis ensures that the
dual space is isomorphic to the original vector space. The analogous statement fails
quite spectacularly for arbitrary Banach spaces. In fact, the quest to understand
the dual space leads to one of the most important results in the subject, the
Hahn-Banach Theorem, which we will discuss in Chapter 4. For now, though,
we will contend with the case of a
Hilbert space. Here the Riesz Representation
Theorem completely answers the question: The dual of a Hilbert space is
isomorphic to itself.
Instead of relying on bases and their dual bases, the proof constructs a rather
natural isomorphism. This map (denoted by A below) is not only useful here, but
will also provide a template for answering similar questions about other Banact
Spaces.
<br>

Hilbert Spaces
79

32.1 DEFINITION If y E H, define p, H K by x + (x, y). Then p, € H* and

:

Ilo.ll = llyll (see Example 2.2.12). Hence we get a map A : H H*given by

A(y) := Py:

Note that A(ay) =

AA(y) and A(y1 +
y2) = A(y1)+ A(y2). In other words, A is a
conjugate-linear isometry.
3.2.2 THEOREM (RIESZ REPRESENTATION THEOREM (F. RIESZ AND M.FRÉCHET,
1907)) For any pE H* there exists a unique vector y e H such that

p(x) = (x, y)
for all xE H. Hence, A : H H* is a conjugate-linear isoinetric isomorphisin.

Proof. Fix 0 p€ H', then := ker(p) is a closed subspace

M
of H. Since p
#0,
M # H and thus # {O} by Corollary 3.1.14. Choose yo
M
E M such that
o(vo)=1. Now for any x H, if we set a := p(x), thenx- xyo E M. Hence
= (x-ayo,
0 yo) = (x, yo) - p(r) lyo||.
Note that yo # 0 since p(yo) = 1. Hence if y = yo/yol', then for
any x E H,
follows
o(x) = (x,y). This completes the proof of existence. As for uniqueness, this
from the fact that A is both linear and isometric (do check this!).
Note that the Riesz Representation Theorem fails for arbitrary inner product
spaces (see Exercise 3.18).
may be equipped with an inner
3.2.3 REMARK If H is a Hilbert space, then H*
product
(Py P:) := (z,y).
space.
Since H* is complete, it is a Hilbert
<br>

Functional Analysis
80

ofof the early pione of

Frigyes Riesz (1880-1956) was Onee
Functional analysis, and his contributions to the subject
include the Riesy
spaces for p 2, dual spaces
Representation Theorem, the study of the LP
this was done with
and the Riesz-Fischer Theorem (Theorem 2.3.9). Much of
a
view to connect the wOrk of Fréchet and Lebesgue (on real
functions) to
were
the work of Hilbert and Schmidt (on integral equations). These ideas
upon
later incorporated into the work of Banach and formed the foundation
Which the subject was built. His long career culminated in the 1952 book
an
Leçons d'analyse fonctionelle (co-authored with B.Sz-Nagy), which remains
important and lucid introduction to the subject even today.

3.2.4 PROPOSITION Let E be a normed linear space, F a Banach space and E < Ea
dense subspace. If To E B(Eo, F), then there exists a unique linear operator Te B(E, F)
Such that

T|E,= To.
Moreover, ||T|| = | To||, so T is called the
norn-preserving extension of To.
Proof. For any xE E, choose a sequence (x,) E such that x, x. Then (x,)
C

is Cauchy. Since To is bounded, it follows that (To(x,)) is Cauchy in F. Since F is

€
complete, there exists y F such that To(x,) y. Define T: E F by
T(x) := lim To(xn ).
noo
(i) We have to prove that T is well-defined: Suppose (zn) c E is another
sequence such that z, x, then ||To(z) -
To(x)|| < ||To|||Zn xl| 0. -
Hence, lim,,o To(2)= lim,o To(x,).
(ii) T is linear: If X, x and yn
Since To is linear, it follows that
y, then x, +yn x +y (see Remark 2.1.2).

T(*+ y) = lim = lim To(*n) + lim To(yn) = T() + T(y).

noo
To(xn
t yn)
no
Similarly, T(ax) = «T(x) for all a E
Kandxe E.

(iii) Tis bounded: If x,, x, then X x

and so
|T(*)| = lim ||To(*)|| <||To|| lim|x|| = ||Tol| ||x||.
1oo 11-o

Therefore T is bounded with ||T| < |Tol. However, T is an extension of lo

and so ||T||> Tol|| holds trivially.
Thus |T| =||To||.
<br>

Hilbert Spaces 81

(iv) Uniqueness: Let T,T2 B(E,F) be two bounded linear operators such that
E
TlE=To = T2\Eg: Then for any xE E, choose (X,) C Eo such
that xy x,
So that
Ti(r) = lim T(xn) =
lim To(,) = lim T2() = T(x).
110 1o0 11-0

Therefore T = T2.

3.2.5 COROLLARY Let E be a normed linear space and Eo < E be a dense subspace of E.
Then the map E* -
E, given by

is an isometric isomorphism of Banach spaces.

Proof. The map S : p

ple, is clearly well-defined and linear. Furthermore,
€
by

Proposition 3.2.4, for any ý E; there exists a unique element such that eE

Hence S is surjective. Also, since |lpll = |l, it follows that S is isometric and thus
injective.
The next result may not seem like much at the moment. After all it is a fairly
direct application of the Riesz Representation Theorem. What is remarkable is that
this result is true for arbitrary normed linear spaces and that fact is one of the
cornerstones of the subject.
3.2.6 CoROLLARY Let M < H be a subspace of H and let p :

M K be a bounded
linear functional. Then there exists p E H* such that

plM=and |l4|| = llpll.

We say that is a
norm-preserving extension of p. It may not be unique (see
Exercise 3.17).
€ M*.
Proof. Let p Then by Corollary 3.2.5 there exists po : M K linear such
Q
that o|M= and ||yol = l a
Since M < H, it is Hilbert space. By the Riesz
Representation Theorem there exists y E M such that po(x) = (x, y) for all x e M.
Now simply define
:H K by
p(x) := (x, y).
Then y is clearly an extension of yo and hernce of p.
Furthermore,|l| = |y =
loll =|lpll.
<br>

82 Functiona! Analysis

Aside. There is another theorem that goes by the namne of Ries

Representation Theorem'. That theorem is substantially deeper than this one
and describes bounded linear functionals on the space Cc(X) of continuous,
Compactly supported functions on a locally compact, Hausdorff space X. We
will havereason to discuss this theorem later in
the book (see Chapter 9). For
now, though, do keep this in mind in case you
wish to look the theorem up
online.

3.3 Orthonormal Bases

At the heart of analytic geometry lies the principle of cartesian
a vector in Euclidean space, one can coordinates. Given
a
assign to it tuple of numbers that uniquely
determines the vector. This simple idea acts as an
important visual aid, while
also providing numerous computational advantages. One of
the chief reasons to
develop Hilbert space theory was to study function spaces using
'coordinates'. We
describe this process in this section.
3.3.1DEFINITION Let H be a Hilbert space, and be a subset of H.
A

(i) A is said to be orthogonal if x ly for €

all distinct x, y A.
(ii) A is said to be orthonormal if it is orthogonal €
and ||x|| =1 for all x A.
(iii) A maximal orthonormal set is called an orthonormal
basis of H.
Warning: An orthonormal basis may not be a Hamel basis for H (see
Exercise 3.24).
3.3.2 LEMMA Every orthonormal set is linearly independent.

Proof. Let be an orthonormal set and {x1, X2,..., Xn}

A

Then for any CA satisfy L a{x; =0.

fixed je {1,2,...„n}, a, =(=1 a;Xi, xj) 0. Hence {x1,
is linearly independent.
X2,...„x}

3.3.3 LEMMA Let A H be an orthonormal set. Then the

following are equivalent:
(i) A
is an orthonormal basis ofH.
(ii) A
= {0}.

(ii) span(^) is dense in H.

<br>

Hilbert Spaces
83

Proof. Let A CHbe an orthonormal set.

(i) (ii): Suppose A is an orthonormal basis and
x E At # {0}, then choose
A
such that |x = 1. Then AU {x} is an orthonormal set.
This contradicts the maximality of A. Therefore
it must happen that
A= {0}.
(ii) (ii): Suppose A
={0}. Since span(A)C. At, it follows that
span(^)= {0}.By Corollary 3.1.14, span(A) is dense in H.
(iii) (i): Suppose span(^) = H,and let A'be an orthonormal set
that contains
A. We wish to prove that A' = A.Suppose not,
then there is a vector
xEA\A.
x
Since N' is
orthonormal, ||x|| = 1, and x
I A. This
implies
that L span(A). By continuity of the inner product
(Example 2.2.3),
it follows that

xl span(^).
Hence x l H and so x = 0. This contradicts the
assumption that
x|= 1. We must conclude that
A=A.
The next result follows directly as in Theorem 1.1.6 by
the use of Zorn's Lemma.
3.3.4 THEOREM If Ao His any orthonormal set, then there is an
C

orthonormal basis
ACH that contains Ao.
3.3.5 ExAMPLE
(i) Let H= (K", | 2). Then the standard basis is an orthonormal basis.
(ii) Let
H=.Then {e1, e2,..} is an orthonormal basis.
=
Proof. Clearly, A {ej, e2,..} is orthonormal. Furthermore, if x
() 1 A, then x, = (x, e) = 0 for all n and hence At = {0}.
= n €
(iii) Let H L'(-n, Tn] and K =C. For Z, define

en():=
V27
Then
=
|1:if 1 m,
(en,em) = iln-m)dt =
27t J-Tt |0: otherwise.
We willprove later (in Section 3.5) that {en :
n E Z} forms an orthonormal
basis.
<br>

Functional Analysis
84

familiar to you from linear algebra. The

be
The next theorem would below was used by Jørgen Gram
in
1883,
orthogonalization process described a function by a
finite
a problem of approximatin8
when he tried to resolve theorem
m

a given set of functions. The form of the

linear combination of Schmidt, who proved.it in 1907 in
familiar to us comes from Erhard
However, the theorem was known to Laplace as earl.
context of L² functions. time
was probably known to other mathematicians at that
as 1820 and
well [36].

ORTHOGONALIZATION (GRAM, 1883, AND

3.3.6 THEOREM (GRAM SCHMIDT
Define
SCHMIDT, 1907)) Let {x1, X2,.. Xn}
. , H be linearly independent.
{u1, 42,
... , un} inductively by u := X1 and

(3.1)
1=1

for j >
{uj, u2,..., un} is an orthogonal set and span({1l1,
2. Then
u2,
. ., Un}) =
span((*1, 12,...„Xn}).
n since this is clearly true if n
Proof. We proceed by induction on
1. If n 1, suppose {u1, u2,
...,
u,n-1} is an orthögonal set such that

span({u1, u2, ...,4-1}) span({x1, x2,...,xH-1}). Then if un is given by

€
Equation 3.1, then clearly n span(*1, X2,. ,Xn}) and u, 0
Since
. „ Xn} is linearly independent. Also, (un, u;) = 0 for all 1, so<

{1, 2.. ...,

{ui, u2,v.,u,} is orthogonal. In particular, {u1, 42, hn} is linearly independent
(by Lemma 3.3.2). Hence

span({u1, 12,...,un})=span({*1,*2,.xn})

since both spaces have the same dimension.

3.3.7 COROLLARY H is aHilbert space and {x1, X2,..} is a linearly independent

f
set, then there exists an orthonormal set ez,..., =
{e,e2,..}
such that span({e, e,})
span((*1, X2,...,xn }) for alln EN.
3.3.8 EXAMPLE Let H = L-1,1| and x, (t) = ". Then Xo, X1,...} is a linearly
independent subset of H. The Gram-Schmidt process gives us an orthogonal set
<br>

Hilbert Spaces
85

(Po, P,} which is

given explicitly as
= 1
P,(t) (-1)".
2"yi)
211n!

These polynomials are called Legendre polynomials. Since span({P,})

span({xH}), it follows by the Weierstrass Approximation
Theorem and
Proposition 2.3.12 that span({P, P2,...}) = H. Therefore if we divide by their
norms, the set

2P, :n=0,1,.
forms an orthonormal basis for H (see Exercise 3.27 for more on
these polynomials).
3.3.9 PROPOSITION Let {e1,e2,...,e,} be an
orthonormal set in H, and let M :=
span{e1, e2,,en. If P: H M denotes the orthogonal projection onto M, then

P(x) = (x,ex)ek
k=1
for any xEH.
Proof. Let x E Hand set x
:= L=1(x, ex)ek. Then E M, and (x-x0, e;) x =0 for
any <jsn.Hence
1 x Xo E -
M. Proposition 3.1.7 then implies that P(x) = xo
The previous result Would, of course, be familiar to you from the study of
finite dimensional inner product spaces. The next theorem, however, is only of
relevance in the infinite dimensional setting. This innocuous looking inequality
is fundamental to the theory and is the basis of all the theorems that follow. It was
originally proved for the trigonometric system (part (ii) of Example 3.3.5) by Bessel
in 1828 and was proved for orthonormal systems in L²{0, 1] by Erhard Schmidt in
his PhD dissertation of 1905.
3.3.10 THEOREM (BESSEL'S INEQUALITY) If {e1, ez,...} is an orthonormal set and
xe H, then r1l(*, en) |2 ||x||. s
€
Proof. For each n x, i=x- LL(x, e;)e;. Then x,
N,write l e;
for all 1<i<
n, s0 X,
I LLI(x, e;)e,. Thus by Pythagoras' Theorem,

Ila|| = lx, |? +
i=1 i=1
€
This is true for all n N,which implies the result.
<br>

Functional Analysis
86

complex numbers converge, thenthe n" term goes to zero as n

Ifa series of
an important consequence of Bessel's lnequality.
This observation gives us
Let {e,e2,.} be an
3.3.11 COROLLARY (RIEMANN-LEBESGUE LEMMA)
orthonormal set, and x lim,-ox, en)=0.
e
H. Then

The Riemann-Lebesgue Lemma is originally statement about the Fourier

a

coefficients of an 1
series. Specifically, a version of it states that the Fourier
as
function tend to zero. As we will see later, the term (x, en) may be thought of
the nh Fourier coefficient of x. Therefore the version of the Riemann-Lebesgue
Lemma stated here proves theoriginal version for L2 functions. We will discuss all
this later in thechapter and discuss the Fourier series of L'functions in Chapter 5,
3.3.12 COROLLARY Let A
a
be an orthonormal set
in H and x E
H. Then {e E A;
(x, e) 0} is countable set.
Proof. For each n E N, define A, := {eE A:|(*, e)|> 1/n}. If {ej, e2,...,eN C
An, then by Bessel's Inequality,

N 1 N
N
=
n2 n2
k=1 k=1

Thus A,n Must be finite and fe E A: (x,e) 0} =U;1Ay must be countable.

3.3.13 REMARK Let {x : a E J} be a (possibly uncountable) set in an normed

linear space E. We want to make sense of an expression of the form
(3.2)

If Jis countable, we can do this by enumerating J = N and treating Equation 32

as a series in the usual sense. However, if ] is uncountable, we
need to make some
changes. Define F:= {FCJ: Fis finite}. For each F EF, we define a partial sum

aEF
We say that the expression in Equation 3.2 exists if there is a vector s E E
with the
property that for all e > 0 there exists a finite set Fo E such that for any F
F €F
containing F, one has SF -s| <e.
In other words, F isa (directed) partially ordered set
under inclusion and {sp:
Fe F} forms a net. Our requirement is that this net be convergent in E. We will
not have much reason to discuss nets
in this book. The interested reader will find
more on the subject in Munkres [421.
<br>

Hilbert Spaces
87

Iaa T4COROLLARY Let A be an orthonornal set in H and x E H. Then

Lrea (x, e) |² < ||x||2.
Proof. Clearly, the sum is the same as
By Corollary 3.3.12, this set the sum over the set {e E A:(,e) #
0}.
is countable and so it reduces toa
Posult now follows from Bessel's countable sum. The
Inequality.
a3.15 LEMMA Let A be an orthonormal set
in H. Then for any x E H, the series

L(x,e)e
eEA
Converges in H.

Proof. By Corollary 3.3.12,

the set fe E
set by en :n A: (x,e) 0} is countable. Denote this
E N}. Then

eEA 1=1
Define Xn i= L1(X, e;)e; and it now suffices to prove
that (x) is Cauchy. By
Bessel's Inequality, Li1 (*, e;)2 < ||| <co. Hence if e > 0, there exists No € IN
such that N
(*, e;) <E. If n>
m> No, Pythagoras' Theorem tells us that

i=m+1 i=m+1 i=No

Hence (x,) is Cauchy and converges in H.
Thenext result is the main theorem of this section. What it says is
the following:
Given an orthonormal basis A of a Hilbert space H, any vector x E
H mnay be
assigned coordinates

*H (x, e)
)eeA:

Furthermore, the vector (and its norm) may be recovered from this tuple of scalars.
These ideas originated in the study of trigonometric series and thus bear the names
of Fourier and Parseval.
<br>

Functional Analysis
88

for each I, y E
H

of H. Then
16THEOREM Let A be anorthonormal basis

(i) Fourier Expansion: x= LeeA(x,e)e.

(i) (x,y) =Leea(X, (y,e).
e)

(ii) Parseval's identity: ||x2 =Se (x, e) |?.

€
Proof. Let x,y E H be fixed. Then by Corollary
3.3.12, the set {e A: (x,e)
set, we may assume tha
0}U{e EA: (y,e) 0} is countable. Replacing A by this
A= {e, :n€
N} is countable.
i) Write z := 1(x, e,)en, which exists by Lemma 3.3.15. By continuity and€
e;) = (I,e;) for any j N

linearity of the inner product, it follows that (z,

Hence x-z€ A.Since At ={O}, x = z.
e, )e, and y = -(y, emem. Then by the
(i1) By part (i), write x = L(,
continuity of the inner product in both variables, we see that

(x,y) = ) (*, en)en, (y, em)em).

1,M=1

Since en
I em if n # m, we get (x, y) = =(X, en) (y, eu).
(ii) Followsdirectly from part (ii).

3.4 Isomorphisms of Hilbert Spaces

Before wwe apply the results of the earlier section to the study of the function spaçe
L', we need to describe the Hilbert space that will be the target of the Fourier
transform. Then we will describe the rather self-evident notion of an isomorphism
between Hilbert spaces.
3.4.1 ILEMMA Any two orthonormal bases of a Hilbert space have the same cardinalitu.

This common number is called the dimension of the Hilbert space.

Proof. If H is inite dimensional, then this statement is fanmiliar from linear
algebra (see [24, Section I.8], for instance). Therefore we will assume
€ H is
infinite
dimensional. Let Aj and Ag be two orthonormal bases of H. If f Al, then there
e E A2
# := a
exists such that (f. e) 0. Furthermore, A,
fe A: (f.e) 0} is
<br>

Hilbert Spaces
89

countable set by Corollary 3.3.12. Since

eE^
we conclude that = A|.
|^i| |^2 x N| By symmetry, |A2 < |A1| holds as
well.
There is a clash of terminology here that you need to be aware of. The
am9ce dimension of H is not necessarily
Hilbert
the same as the dimension of H as a
vector space. For instance, dim() = o (countable infinity),
but its vector space
dimension is uncountable. Proceed with caution!
a
42 DEFINITION Let I be any set, and 1 <p<oo
(i) A function f:1 K
is said to be p-summable
if supp(f) :={i l:f) E

0} is countable and ;efi)P < o (where the series isdefined in the sense
of Remark 3.3.13).
(ii) Let lP(1) denote the set of all p-summable
functions on I. Then the
inequality
If +g|P s (2max{IfI, Is|}P< 2P[|f|P + \g|P]
shows that l() is a vector space. If f E P), we define
1/p
Ilfll, :=

and this satisfies Minkowski's Inequality since the verification only requires
a countable sum. Hence P(I) is a normed linear space. Furthermore, lP (I)
is a Banach space as before (Exercise 3.35).
(iii) Also, (I) has an inner product given by

(f.s)=Ef)g0).
iel

Once again this is well-defined since supp(f) Usupp(g) is countable. Hence

e(I) is a Hilbert space.
As mentioned before, these spaces lP (I) are themselves special cases of L?
spaces. Yousimply take the set I, equip it with the counting measure and voila!
3.4.3 LEMMA Let I be any set. Then dim(())=|I|:
<br>

90 Functional Analysis

Proof. For each ie I, define e; : I- K by

= |1 : ifi=j,
ei(j) ) óy-0|0 : ifi#j.
€
Then the set A := {e; : i e I) forms an orthonormal set in (1). If f Rn
satisfies f A, then for any i
€
I,

f(i) = (f, e;) = 0

and so f=0. Hence At = {0} and we conclude that A is an orthonormal basis for
((1). In particular, dim((1) = |A| = |.
The basis A constructed as above is called the standard orthonormal basis for
(0). We now examine the notion of isomorphisms of Hilbert spaces. Obviously,
any isomnorphism must be an isometry. Let us isolate one useful fact about
isometries which we will need later on, the proof of which we will leave to the
exercises.
3.4.4 LEMMA Let H and K be Hilbert spaces. A linear mnap T: H K is an isometry
if and only if (Tx, Ty)K = (, y)H for all x, y E H.
A simple example of an isometry that is not an isomorphism is the right shift
operator S : l2 given by
S(*1, 12,
x3,...) := (0, X1, X2, X3,
...).
Clearly, the only thing that prevents an isometry from being an isomorphism is
surjectivity.
3.4.5 DEFINITION An operator U : H K called a unitary operator
is u is
if
a
surjective isometry. H is said to be isomorphic to K if there is a unitary man
U:H- K.If such a map exists, we write H K.
Clearly, Ifl is arange also a unitary. Also, if U and Vare unitaries
unitary, then Uis
(with appropriate and domain), then UV is a unitary. Therefore the notion of
isomorphism is indeed an equivalence relation on the collection of Hilbert spaces.
The proof of the next result is more important than its statement. The
isomorphism described below is the 'Fourier transform' afforded by a fixed
orthonormal basis of the Hilbert space.
3.4.6 THEOREM Let H be a Hilbert space and A be an orthonormal basis of H. Then
HP(A).
<br>

Hilbert Spaces
91
€
Proof. For any x H, define £: A Kby
R(e) := (x, e).
By Corollary 3.3.12, supp(?)
Thuswe define U : H
ofthe inner product.
is countable
2(^) given by x .
and by Bessel's Inequality, & E (A).
+ Note that U is linear by the axioms
Furthermore, by Theorem 3.3.16, we see that
(u(x),u(y)) =x,e) (y,e) = (*y)
eEA
and so U is an isometry by Lemma 3.4.4. Finally, for
each f E A, f(e) = (f,e) = be,f
Therefore {u) :fE ^} is the standard orthonormal basis for (A) and hence
Range(U) ={0}. By Corollary 3.1.14, Range(U)is densein 2 (^). However, U
an isometry and so Range(u) is a is
complete subspace of (A). Thus Range(U) is
closed and U is surjective.
The next (rather surprising) fact makes the study of
Hilbert spaces line up
narticularly well with what we know of finite dimensional vector spaces.
In
particular, it implies that separable infinite dimensional Hilbert spaces are all
isomorphic to one another!
B4.7COROLLARY Two Hilbert spaces are isomorphic if and only if they have the same
dimension.
Proof. Suppose U an isomorphism and AH C H is any orthonormal
:HKis
basis for H, then U (AH) C K is an orthonormal set. Hence by Theorem 3.3.4, there
is an orthonormal basis AK of K such that U (AH) cAK. In particular,

dim(H) = |AH|= |u(AH)| < |Ak| = dim(K).

By symmetry, dim(K) dim (H) as well.
<

Conversely, suppose dim(H) = dim(K), let AH and AK denote two

orthonormal bases of H and K, respectively. Then |AHl = Ak| and hence there is
a
natural isomorphism 2(AH) e(AK). Now apply Theorem 3.4.6 to concude
that
H
K.
3.4.8 THEOREM A Hilbert space H is separable if and only if dim(H) is countable.
Proof. IfH
is separable andAC H is an orthonormal set, then for any pair of
distinct vectors e, f E A, we have
fl| = V2.
le-
a open sets in H. Since
Hence the balls {B(e; V2/2) e E A} form family of disjoint
:

A must be countable.
His separable, this family must be countable and hence
<br>

Functional Analysis
92

Conversely, if has countable orthonormal basis

H a := A
N},then
en :

Ko =Q or Ko
= x Q, according as
Q

span(^) is dense in H. Now set

and A = H. Therefore
R K=
or IK = C. Then A := span (A) is countable
separable.
spaces
3.49 COROLLARY Any two separable, infinite dimensional Hilbert are
isomorphic.

In particular, there is an isomorphism

L'-7, n] (z).
Our goal in the next section is to the describe one such isomorphism explicitly.

3.5 Fourier Series of L? Functions

Inwhat follows, we will perforce be working with complex-valued functions. In
other words, we will tacitly assume that K =C throughout this section. Back in
Example 3.3.5, we had defined an orthonormal set n E
:
{en Z} in L'(-7, n] by
1
en(t) = ceint,
V27
Now we show that this forms an orthonormal basis for L²(–T, T]. The proof is
relatively straightforward: We show that the span of this set is dense in C[-T, a]
and then appealto the fact that C[-, n] is dense in L²[-, 7).
For the first step of the proof, we need the following fact. lt isa consequence of
the Stone-Weierstrass Theorem, whose proof we have described in the appendix
(Section A.1).If you have not seen a proof of this theorem, do go ahead and take a
look at that section.The Stone-Weierstrass Theorem is an important one in modern
analysis andwill play a crucial role for us in this book.
3.5.1 PROPOSITION Let S' := {z E |z = 1} and let C(S') be equipped vith
C:
the supremum norm. Let F denote the space of all polynomials in z and ž, thought of as a
subspace of C(S' ). Then F is dense in C(S').

Fix E := {fE C[-t, n| : f(-n)= f(n)}. For n E Z, define e, E E as above

and set A:=span({ey}) CE. Note that

A= span({cos(nt), sin(nt) :n E
Z}).
Hence anelement of A iscalled a trigonometric polynonnial.
<br>

Hilbert Spaces 93

Nowif fE C|-TU, T
< o, then ||f
and 1<p < (27r)1/P|fo. Hence if a
subset of C[-, Tt
is dense with respect
to the supremum norm, then it is dense
with respect to any p-norms. We wish to prove that the set A defined above is
dense in C|–1, TT| With respect to | 2. In order to do this, we need one fact about
1, It is the quotient space of the interval –T, via the identification
-t
T
It.
Specificaly, the mapq:[-a,n s' given by

is the quotient map. Therefore a function F: S! Cis continuous if and only if

7] We use
Foq:l-n, C is continuous. this fact below.
Q5.2 PROPOSITION A
is dense in E with respect to the supremum norm. Therefore it is
dense with respect to || p
for all1<p<oo.
Proof. Define 0 : C(S') E by

O(f) (+) := f(et).

Then is linear and since the function q :(-r, n] s' is surjective, it follows that
|je(f)ll = |lflloo. Therefore 0 is isometric and hence injective.
Given f e E, define F: S! C
by F(e") := f(). since f(-n) = f(r), F is
well-defined and continuous (by the discussion preceding this proof). Hence FE
C(s') and (F) = f. Therefore is an isometric isomorphism.
If E C(S') denotes the identity function G(z) =z, we have 0(3)(t) = e and
9(7)(t) =e", Therefore if F is the subspace of C(s') defined in Proposition 3.5.1,
then e(F) =A. Since F is dense in C(S'), Ais densein E.
, for all oo.
3.5.3 LEMMA E is dense in LP|-n, T
with respect to || 1<p<
Proof. For n E N,define

- 7t
1 : if +1/n<t<TT-1/n,
fr(t) = 0 : if t e {-n, },
: otherwise.
linear
€ € -
Then ||fn – 1|p <. For any g C[-1,], note that fug E and ||f8 gllp <

lglco 0. Hence C[-n, n| C EP. Now the result follows from the fact that
C[-7, ] is dense in LP[-7, 7] (Proposition 2.3.12).
<br>

Functional Analysis
94

35 4THEOREM If
en(1) :=
V27
Z} forms an
orthonormal basis of L|-7, 7.
then the set {en
:
ne an orthonormal set. By
that it is
3.3.5, we know By Lemma 3.3.3.
Proof. By Example 3.5.3, span({en) = L-7, n|.
Proposition 3.5.2 and Lemma
is an orthonormal basis. time. And no.
course, what we have been after this whole
1his result is, of n, we are free to apply the result
basis for L2[-7,
that we have an orthonormal customary to normalize the Lebesgue
measure
follows, it is
of Section 3.3. In what measure 1. This is done by simply dividing the
on -7, n| So that the interval has Z} is
measure by 27t. In that case, we set (t) := ent and observe that
e,
e
:nE
now an orthonormal basis for L2(-7, (with this normnalized nmeasure).
T

is
e T]. For ne Z, the nth Fourier coefficient of f
3.5.5 THEOREM Let f L'|-n,
defined as

fln) := (f, en) = 27

1

J
f()e-in dt.
Thenwe have
(i) Fourier Expansion: f(t) ~ E-o
~ the convergence is
Note: It is traditional to use the symbol here indicate that
to

in the L? norm but not necessarily pointwise.

f is an
(ii) Fourier Series: The map U: L2-T, T
(Z) given by f
isomorphism of Hilbert spaces.
(iii) Parseval's identity: ||f5 =E-olf(n)|.
(iv) Riemann-Lebesgue Lemma:

f()en" dt = cos (nt)dt = lim sin(nt)dt =0

lim lim | f()
nto0J–TT „T t()
ntCoJ-T

(u) Riesz-Fischer Theorem: For any (cH) E

€
P(Z) there exists f e L'(-,n] such
that f(n) = cy for all n Z.
This concludes our discussion on Hilbert spaces and the Fourier series for the
moment. We will return to these ideas repeatedly as we go along. For now it is time
to consolidate our understanding with a few problems.
<br>

Hilbert Spaces
95

3.6 Exercises
Preliminaries
EXERCISE 3.1 Let (E, (') be an inner product space and let T :
E E be
an injective, linear map. Define a bilinear map (,):
(T(*),T(y)).
E Ex K
by (x,y)

(i) Show that (',) is an inner product on

E.
(ii) Let a, b, c E R with b² < 4ac and
ax + bxy +
consider the ellipse I := {(x, y) R?:
cy= 1}. Prove that there is an inner product
(,) on R such
that I = {vE R: (v,v) = 1}.
BXERCISE 3.2 For n E N, let Pu denote the space of all polynomials one
variable over K with degree atmost n. Let := {x1, X2,..., x*} be a finite in
A

K and define a bilinear map (,)A:Pu X P K

subset of
by
k
(f.s)a :=Ef(x)g(*).
i=1
Determine conditions on the set A under which this an
inner product on Pn.
3.6.1 DEFINITION A senti-inner product on a vector space E is a
function
(:, ): Ex E K
such that for all x,y, z E E, a, E K, B

(i) (ax + By, z) = a(x, z) + B(y,z),

(i) (x,y) = (y,x),
(iii) (x, x) >0.
Note: It is possible that (x, x) €
=0 for non-zero x E.
EXERCISE 3.3 (2) Let (,) be a semi-inner product on a vector space E.
(i) If x E
E is such that (x, x) = 0, then prove that (x,y)
=0 for all y e E.
Hint: Consider z
:= t(y, x)x -y, where tE R is arbitrary.
(ii) Prove the Cauchy-Schwarz Inequality: For any x,yE E, prove that
I(x, y)[? < (x,x)(y, y).
<br>

96 Functional Analysis

(ii) Letp:E R+ denote the function p(x) := x,x). Prove that p

isa
seminorm on E.
(iv) LetN:= {xE E: p(x) =
( 0} and H
:= E/N. Prove that the map
H× H K given by ()
(x+N,y +N) := (, y)

is a well-defined inner product on H.

Orthogonality
EXERCISE 3.4 (2) Prove Proposition 3.1.3.
EXERCISE 3.5 (2) Assume that (E, | |) is a a
Banach space such that the norm
satisfies the parallelogram law. If E is a real Banach space,
define
1
(*,y) i= (lx +
yl|'||x- yll).
If E is a complex Banach space, define

– i||x - iyll').
1

(x,y) := 4 (lx+yl-||x–yl|²+illx+iyl|
In either case, show that this formula defines an inner product on
E which induces
the given norm.
EXERCISE 3.6 Let H
be a Hilbert space and ACHbe any set.
(i) Prove that (A)= span(A).
(iü) If M, and M2 are two closed subspaces of H,
M
+ M.
then show that (M n M,) =

3.6.2 DEFINITION Let x1,X2,..., Xy} be a

finite set vectors in an inner product
of
space H. The Gram determinant
of these vectorsis defined as the determinant
(X1,X1) (x1, x2) ... (x1, X) |
(x2, X1) (2, X2)
... (x2, x)
G(x1, X2,...„
x,)=
(X, x1)(, x2) ...(x, x,)|
EXERCISE 3.7 For 1 < n, define
k< H := span{x1, X2,...x*}.
<br>

Hilbert Spaces
97

() Prove that G(1, N2, = G(r1,X2,...*n ).

Xp-1)d(xn, H,-1)
Hint: Let z = E
;X; denote the orthogonal projection of x,
Erom the last row of this matrix, subtract onto H-1.
, the first (n -1) rows multiplied
by 1, a2, an-1, respectively. This does not change the determinant. Now
dosomething similar with the columns.
(ii) Prove that G(*1, X2,...,X) >0 and that equality holds if and
vectors are linearly dependent. only if the

(iii) If {X1, X2, , Xn} are

linearly independent, then prove that

G(Xj, X2,,Xn)
d(Xn, H,,-1) = V
G(*1, *2,Xn-1)
The quantity VG(*1, X2,,Xn) represents the volume of the parallelotope
spanned by the vectors {x1, *2,...,Xn}.
BXERCISE 3.8 () Let M1,M2 < H be two closed subspaces such that M1 L M2,
and let P and P2 denote the orthogonal projections onto M and M2, respectively.
Prove that
(i) + M2 is a closed subspace of H.
M1

(ii) PP2 = PpP=0.

(ii) P + ½ is the orthogonal projection onto M1 + M2.

EXERCISE 3.9 Let H be Hilbert space over R. Showthat there is a Hilbert space
a

Kover and an R-linear map T H K satisfying the following properties:

C :

(a) (T(*), T(y)) = (x, y) for all x, y e H.

(b) For any z there exist unique x, y E H such that z = T(*) + iT(y).
E K
(c) If K is a complex Hilbert space and T: H K is an R-linear map satisfying
conditions as in parts (a) and (b), then there is a C-linear unitary U : K K
o
such that T=T.
U

This proves that the pair (K,T) is unique upto isomorphism. The Hilbert space K
is called the complexification of H.

Hint:Take K = H xH (as a set) and define addition and scalar multiplication over
Cappropriately.
<br>

Functional Analysis
98

Best Approxima.
The next two problems are meant to indicate that the spaces.
Theorenm (Theorem 3.1.6) can fail quite miserably in arbitrary Banach

EXERCISE 3.10 Let E:=Cí0, 1l, equipped with the supremum norn, and
ornr set

= 1}.
fof0)dt
A:=reE:sa-
(i) Prove that A is a closed, convex subset of E.
(ii) Provethat inf{ILf:fE A} =1.
(iii) Prove that there does not exist any f E A
such that |||=1.
EXERCISE 3.11 Let A and B be the subsets of given by

A= ) e
e:*=1
1=1
and

(i) Prove that A and Bare both closed, convex subsets of e'.
(i) Prove that inf{||xl :xE A} = inf{||xl|1 : xE B} = 1.

(iii) Prove that there exist infinitely many vectorsin A

with norm 1.

(iv) Prove that there does not exist any vector x E B with |||1 = 1.

EXERCISE 3.12 (2) For each n > 1, let (H,n,E) be a Hilbert space. Set

for each i > 1, and xl,

i=1
<o.
Define addition and scalar multiplication pointwise. Define an inner product on H
by

((xH), (y)) :=L(*iyi)H,

i=1

Show that this inner product is well-defined and that (H, (·, ) is a Hilbert space.
This Hilbert space is called the direct sumof the {H, : n> 1} and is denoted by
H,.
<br>

HilbertSpaces
99

OVERCISE 3.13 () Let (E, (,)E) be an inner product space

ompletion of E as a normed linear space (following and let E denote the
the notation of Exercise 2.29).
(i) Define a bilinear map on E by

((xn)+ N,(y) + N) =
lim (xi, yi)E.
Show that this limit exists and defines an
inner product on E.
(i1) Show that E, equipped with this inner product, is a
Hilbert space.
(iii) Define T: E E
by

T(*) = (x, x,
X,...)+ N.
Showthat T is an isometrywhose range
is dense in E.
(iv) Suppose K is a Hilbert space and
range, then show that there is a S: E K is an isometry with dense
unique unitarymap U: E K such that
UoT = S.This proves that the pair (E, T) unique upto
is isomorphism.
Asin Exercise 2.29, space
the E is called the Hilbert space completion
of E.
YERCISE 3.14 ()
For each n >1, let (H,,',)H.) be a
Hilbert space. Set
E:={(X,):x; E H; for eachi > 1, and there is NE N such that x; = 0 for all i>N}
Define addition and scalar multiplication pointwise.
Define an inner product onE
by

(xn),(yn)):=L(*i, yi)H,
i=1
Note that this is well-defined because it is a finite sum. Show
that the completion
of with respect to this inner product is the Hilbert space direct sum
E

H,,.
EXERCISE 3.15 (2) Prove Corollary 3.1.14.
EXERCISE 3.16 Let E denote the space of all functions f:-1,1| R such that
f(-1) = f(1) = 0, and f is infinitely differentiable on (-1,1) (the space E is
usually denoted by C(-1,1)). Note that we may think offunctions in E as defined
on all IR by extending them to be zero outside-1,1|.
of

(i) If p
is a polynomial such thatJe f()p(t)dt = 0 for all f E E, then prove
that p = 0.
<br>

Functional Analysis
100

2.3.12.
Follow the argument of Step (i)of Proposition
Hint:
Rm+1 by T) = (yo. Y1. ym),
E
whera
(ii) For
m
eNfixed, define T:

Prove that aa
surjective,
T
linear operator. In particular there exists
is
such that T(f) = (1,0, 0,... 0).
Hint: Use Corollary 3.1.14 on M = Range(T)<R".

(i) Conclude that for each

m E N there exists fE Esuch that
fe *q=9

for every polynomial= of degree <

q m and every e > 0 (here fe denotes
the function fe(x) e'f(x/e) and * denotes the convolution operation
u *v(x) := Sp u(y)v(*
-y)dy).
on IR" for n >
Note: The results of Exercise 3.16 also hold for functions defined
with appropriate changes to the spacesconcerned.

Riesz Representation Theorem

any subspace. Tho
EXERCISE 3.17 (4) Let H be a Hilbert space and M < H be
i
map H* M*given by olM is surjective by Corollary 3.2.6. Prove that it
Q

injective if and only if M = H.

= from
EXERCISE 3.18 (4) Let E Coo equipped with the inner product inherited
l. Define : E by K

(i) Prove that p is a bounded linear functional.

= (x, y) for all
(ii) Prove that there does not exist any y E such that p(x)
E

XEE.
= {0}. Is M dense in E?
If M= ker(), then prove that
M

(iii)
€
EXERCISE 3.19 Let E
be an inner product space such that for any E there
€
exists yE Esuch that q(r) = (r,y) for allx E.Prove that E is a Hilbert space.
<br>

Hilbert Spaces 101

Orthonormal Bases
EXERCISE 3.20 Prove Theorem 3.3.4.
EXERCISE 3.21 Let M be the plane M = {(x,y, z) E R: x+ y + 2z 0}.
Determine the orthogonal projection PM R' :
M (in other words, given a vector
E RS
(x, y, z) determine the vector PM(x,y, z) €
M).

EXERCISE 3.22Consider the Sobolev space W (0, 1] introduced in Exercise 2.25

(once again read ahead to Definition 4.1.8 before you attempt this problem). Let H
denote the subspace of W|0, 1] consisting of those functions f e W}j0,1] such that
= = 0.
f(0) f(1) Observe that the map

defines an inner product on H.

(i) Provethat H is a Hilbert space.
n
(i) For E Z\ {0}, define e, H by

p2int 1
en(t) :=
27T11

Prove that #0} forms an orthonormal basis for H.

{en :1
Hint: You will need the fact that {t
L'{0,1] (as in Theorem 3.5.4).
erint : ne
Z} spans a dense subspace of

EXERCISE 3.23 Let A be an orthonormal set in a Hilbert space H and let M =

span(A). If PM
:
H M denotes the orthogonal projection onto M, then prove
that

PM(x) = (,e)e
eEA

for all xE H,where the sum is defined as in Remark 3.3.13.

EXERCISE 3.24 (2) Let A be an orthonormal basis for a Hilbert space H.
€
(i) For any x H, prove that x can be expressed uniquely as x = Leen aee, where
a, E K for all eE A.
(ii) Conclude that A is a Hamel basis if and only if H is finite dimensional.
<br>

102 Functional Analysis

EXERCISE 3.25 () Let H be an infinite dimensional Hilbert space and

nEN} be a
countably infinite orthonormal set.
(i) Prove that the series x :=Lu=1 converges in H and that x ¢ span(A)
(ii) Conclude that H cannot have a countable Hamel basis.
EXERCISE 3.26 Let H := L20, 1]. Define
A := {1,
v2 cos(7zt), V2 cos(27rt),...} and
B={1, V2 sin(t), V2 sin(27tt), ...}.
Show that A
and B are both orthonormal bases
for H.
EXERCISE 3.27 (2) Let H := L2|-1,1]. For n >0, define the Legendre
polynomials P, E H by

1
P,(t) = (G)(-1)".
2" n!
i) Compute the first three Legendre polynomialsPo, and P2.
P
(ii) If x,(t) = ", then show that x,,
: n
L P, for any m < n. Conclude that the set
{P, 0} is orthogonal.
Hint:Use integration by parts
multiple times.
(ii) Show that the Legendre polynomials are
obtained by Gram-Schmidt
orthogonalization of the monomials xy :n> 0} in H.
(iv) Show that ||P, ll3 = 211-+]
(v) Show that P, satisfies
thedifferential equation
d d
di(-)P +
n(n+ 1)P, = 0.
(vi) Let M:=span{xo, X1, Xx} be the subspace of H
of degree < 2 and let consisting of polynomials
P:
projection. Let f E H be
H M denote the corresponding
orthogonal
€
the function f(t) := e'. Determine the best
approximation P) M.
EXERCISE 3.28
Define

i-ff:R-R:dx<a.
<br>

Hilbert Spaces 103

) Show that H is a real vector space and that (, ):H×H R

given by

(f.s):= ef()e()dx
isa semi-innerproduct on H.
(it) Let (H,() be the inner product space constructed as in Exercise 3.3. Prove
thatH is a Hilbert space.
Let H
3.29 Let be
EXERCISE H be the Hilbert space defined in Exercise 3.28. Define the
Hermite polynomials by

H,() = (-1)" e2
(1) Show that Ho(t) = 1and that

Hn+1()= 2tH,(1) – H,(1).

dt
Conclude that each H, is a polynomial of degree n.
(ii) Show that {H, : n> 0} is an orthogonal set in H.
Hint: Use integration by parts to evaluate

LH.t)()? dt.

(iii) Show that ||H, ||2 = 2"n!.

It takes some work, but one can show (with the help of the Fourier transform)
that the collection {H,
:

n>
0} spans a dense subspace of H. Thus after
normalization, one obtains an orthonormal basis of H. The Hermite polynomials
(like the Legendre polynomials from Exercise 3.27) are useful functions that are
studied by both physicists and mathematicians alike.
EXERCISE 3.30 Let la, b] R be a closed and bounded interval and let {en :

neN} be an orthonormal basis for L²(4, b]. Define fij€ L²(a, b]') by fij(s, t) =
e;(s)e;(1). Suppose f E L²((4, b|") is orthogonal to {fij i,je N).
:

) Set f.(t) := f(s,t). As in Example 2.2.17, fs e L'(a, b] for almost every s E

la, b]. Now show that
2
ds
j=1
<br>

Functional Analysis
104

(ii) Set g(s) := e(t)fs (t)dt and show that (f. fuj) =Ja 8;(s)e; (s)ds.
: 1,) E N} forme
(iii) Prove that f= 0. Conclude that the set ii an
orthonormal basis for L'([a, b]²).
Hint: Use the Fubini-Tonelli Theoremn liberally.
a
bounded : H K
be
Hilbert spaces and
T
()
EXERCISE 3.31 and Let H K be

linear operator of finite rank (in other words, Range(T) is finite dimensional)
Trove that there exist vectors X1, X2,...,X, E Hand /1,/2,..ynE K such that

T(2) = L(2, *i)yi

j=1

for all zE H.

Hint: Start with a Fourier expansion in Range(T).

EXERCISE 3.32 Let
functionalon H. Let
H
be a Hilbert space and
:= {ea
H :
K be a bounded linea.
I} be an orthonormal basis for H.
A
:aE
(i) H issuch that p(x) = (x,y) for all x E H, then determine the Fourie
Ify E
coefficients of y with respect to the basis A.
(ii) Use this to construct a proof of the Riesz representation that uses the Fourier
expansion.

Isomorphisms of Hilbert Spaces

EXERCISE 3.33 (2) Prove Lemma 3.4.4.

EXERCISE 3.34 Let H and K be Hilbert spaces and T : H K be a surjective

function such that (Tx, Ty) = (x, y) for all x,y E H. Prove that T is linear.
EXERCISE 3.35 (4) Let I be any set and let 1 <p < oo. Prove that P() is a Banach
space.
EXERCISE 3,36 (2) Let H be a separable Hilbert space. Prove that there a
is
sequence (P) C B(H) satisfying the following
conditions.
(a) For each x € H, lim,-co
P(x) =x.
(b) sup,eN Pa = 1.
<br>

Hilbert Spaces
105

Fourier Series L2
of Functions
EXERCISE 3.37 €
Let f
L4|-7, 7 be a function such that
Drove that the Fourier series of j converges lf(n)| < o.
uniformly (in the supremum norm).
EYERCISE 3.38 Let
f E L–T, T be a function whose Fourier series converges
1niformly (in the supremum norm). Prove
that the sum of the series must be f.
In order to attempt the next problem, you
will to know Lebesgue's
Endamental Theorem of Calculus (see Theorem 4.1.10 need
and the surrounding ideas).
EXERCISE 3.39 Let
f E C|-T, T be an absolutely continuous function,
derivative f. Prove that the Fourier series of f' takes with
the form
~
f'() inf(n)ent,
1=

In other words, the Fourier series of f' is obtained from

the Fourier series of f by
termwise differentiation.
EXERCISE 3.40 Let f E L'-, T
and let F:-, nCbe defined by
F(x)= f(t)dt
for somne a E -,
(note that this is well-defined because f E L'-t, T
by
Exercise 2.12). Then prove that there are constants a,BE C such
that the Fourier
series of F takes the form
–
F(1) at B+),gint
in
10
EXERCISE 3.41 Suppose fE C[-n, n] is a continuously differentiable function.
(i) Use Exercise 3.40to prove that -f (n)| < 0.

(ii) Conclude from Exercise 3.37 and Exercise 3.38 that the Fourier series of f
converges uniformlyto f.
question of uniform (indeed pointwise) convergence of the Fourier
Note: The
series will be taken up more seriously in Chapter 5. Specifically, have a look at
Remark 5.2.3 and the ensuing discussion.
EXERCISE 3.42 Consider f [-7,n] :
C given by f(x) = 2. By Exercise 3.41,
the Fourier series of f converges uniformly to f. Use this to compute the sum of
the series and
n
1
<br>

Functional Analysis
106

EXERCISE 3.43 Prove that the set

„
A := {1,cos(t), sin(t),... cos(kt), sin(kt), ..}

forms an orthonormal basis of L2[-7, n] when (-, 7| is equipped with th.

€
normalized Lebesgue measure. Conclude that everyf L-, can be
expressed as a series of the form
1
f(1) ~ 5a0 + a; cos(kt) + b sin(kt))
k=1

where the convergence is in the L² norm (as in Theorem 3.5.5).

Note:
The numbers ag and appearing above are called the Fourier cosine and
b
Fourier sine coefficients of f, respectively.
the

Reproducing Kernel Hilbert Spaces

EXERCISE 3.44 (%) Let
D:={zE C:z| < 1} and let H(D) denote the space of
all holomorphic functions f : D such that C

1 27 1/2
ILf
:=| sup f(re)Rde
\ocrei27Jo <o0.

This space is called the Hardy space.

() If fE H(D) is expressed as a Taylor series f(z) = €

ayz" (for z D).
0
then prove that

1/2

n=0
€
(i) Iff.g H(D) are expressed as = L=0 Ay2" and g(z) =
f(z) o
(for z E D),prove
that the formula bz

(f.g):= a,bn
n=0
is awell-defined inner
product on H (ID).
(ii) Prove that H(D) is isometrically
isomorphic to and hence a Hilbert
Space.
<br>

Hilbert Spaces 107

EXERCISE 3.45 Let H(D) dernote the Hardy space on the unit disc described in
Exercise 3.44. For a E D, let pa : H²(D) C denotethe linear functional

fH f(a).
() Prove that p. is bounded and that 1

|Pall a?'
(i) Determine a function kE H²(D) such that f(a) = (f, ka) for all
Use this function to prove that equality holds in part (i).
f€ H(D).
The phenomenon described above for H²(D) is a
special case of the following
notion.

36.3 DEFNITION Given a set X, we say that H is a reproducing kernel

Hilbert space on X if it satisfies the following conditions:
(a) H is vector space of K-valued functions on X,with pointwise operations.
a

(b) H is endowed with an inner product (,) that makes it a Hilbert space.
(c) For each y E X, the evaluation map f fu) defines a bounded linear
functional on H.
happens, then for each y E X, there is a unique function ky e H such that
If this
f(y) = (, ky) for all f€ H (by the Riesz Representation Theorem). The function
K:Xx X Kgiven by

K(x,y) := ky(x)

is called the reproducing kernel for H.

EXERCISE 3.46 Let H be a
reproducing kernel Hilbert space over a set X, with
reproducing kernelK.
(i) Prove that K(, x) = K(x,y) for all x, y E X.

(ii) For any y1 y2,yn X and W1, N2,n K, prove that

i K(yi y;)a;a; >0.
(x) := K(x,y) as above, then prove that span{ky
:
(ii) If k, E H is the function ky

yE X} is a dense subspace of H.
3.6.4 DEFINITION LetX be a set. A kernel function on X is a function K: XxX
IK satisfying the following conditions.
<br>

Functional Analysis
108

(a) K(y, x) = K(x, y) for allx,y E X and

ay E K, LL=1 K(yi,yi)aftj >0.
(b) For any y1, y2,..ynE X and aj, a2,...,
EXERCISE 3,47 Let Kbe a kernel function on set X. The
a purpose of this exercise
space having K as its reproducing
is to construct a reproducing kernel Hilbert
kernel.
on X, with pointwise
(i) Let F denote the vector €space of all K-valued functions
operations. For each y X, let ky F denote the function k,(x) := K(x, y).
Set W := spanfky : E X}. For two elements f
y = L
;ky, and g =
LL1B;kz, in W, define
11

(f.g) := L«A;K(zj,y:).
i=l j=1
Prove that this is a well-defined semi-inner product on W (see Exercise 3.3).
Note: In general, f and& may be representable as such linear combinations
in multiple ways. You need to prove that this inner product is independent
of how f and g are represented.
€
(ii) For any fE W and x X, provethat |f (x) |² sK(x, x) f. f). Conclude that
(:,) is an inner product on W.
Hint: Use the Cauchy-Schwarz Inequality from Exercise 3.3.
(iii) Let H be the completion of W with respect to this inner product (as in
Exercise 3.13). Prove that H may be realized as a subset of F.
(iv) Prove that H is a reproducing kernel Hilbert space over X, with reproducing
kernelK.
The purpose of the next exercise is to show that a reproducing kernel
Hilbert space is completely determined by its reproducing kernel. Together with
Exercise 3.47, we conclude that to each kernel function K, there is one and only one
reproducing kernel Hilbert space, which has K as its reproducing kernel.
EXERCISE 3.48 Let H; and H, be two reproducing kernel Hilbert spaces on a
common set X, with the same reproducing kernel K. For i = 1,2, set W; :=
span{ky€ H; :y E X} (with k, defined as above).
€
(1)
If fE W = W2 and x X, prove that the value f(x) is independent of
whether we regard it as an element of W or W2. Furthermore, prove that
I|fllH, = ||f,.
<br>

Hilbert Spaces 109

(ii) Prove that H = H).

Hint: Let f E H1, and (fn) be a sequence in W; such that fn
Prove that there exists g
f in H.
H, such that || f -lH 0 and conclude that
-
f=8.
€
(iii) Prove that |flH, =fH, for any f H.
BXERCISE 3.4
A
be a separable, infinite dimensional reproducing kernel
Let H

Hilbert space over a set X and let fe, n

E N} be an orthonormal basis for H.
:

Show that the reproducing kernelof H isgiven by the formula

K(x,y) = en(*)en(y).
n=1

EXERCISE 3.50 Let H be the Hilbert space described in Exercise 3.22.

€
(i) For each x [0, 1], prove that |f(r)) <
|lfll for all f E H. Conclude that
H is a reproducing kernel Hilbert space on (0, 1].
Hint: If f e H, then f(x) = o f'(t)dt.
(i) Show that the function K: (0,1] x [0,1] R given by

K(x, y) :=
J(1-y)x :if x <y,
l1-x)y : if x >y.
is the reproducing kernel of H (and with this you can show that the estimate
obtained in part (i) is 1ot optimal).

Additional Reading
. Given the long and storied history of the Fourier series, there is no shortage of
good books on the subject. However, my personal go-to book is this one.
Stein, Elias M. and Shakarchi, Rami. (2003). Fourier Analysis. Vol. 1. Princeton
Lectures in Analysis. An introduction. Princeton, NJ: Princeton University
Press. ISBN: 0-691-11384-X.
In the problems immediately after Exercise 3.45, we have developed (to an
extent) some basic ideas surrounding reproducing kernel Hilbert spaces. Their
theory is deep and touches a variety of allied fields such as complex analysis
and partial differential equations. While many articles exist on the subject (and
<br>

110 Functional Analysis

even anice book by Paulsen and Raghupathi [45), perhaps the best place to
start remains a paper by Aronszajn, written after the Second World War.
Aronszajn, Nachman (1950). Theory of reproducing kernels. Transactions
of the American Mathematical Society, 68, 337-404. ISSN: 0002-9947. DO:
10.2307/1990404.
<br>

Chapter 4

Dual Spaces

4.1 The Duals of LP Spaces

Given a normed linear space E, we have briefly encountered the dual space
E, consisting of bounded linear functionals on E. We know that it is a Banach
space (Theorem 2.3.13), but beyond that we do not know much. In the finite
dimensional case, E and E* are isomorphic as vector spaces. However, in the
infinite dimensional case, this need not be true. Moreover, even in the finite
dimensional case, an arbitrary isomorphism may not be isometric!
Now, ifE is a Hilbert space, then there is an isometric isomorphism EE* by
the Riesz Representation Theorem. This is not true for an arbitrary Banach space.
However, the core idea of that proof (specifically, the construction of the map A) is
applicable when studying the dual space of P or LP{a, b] for 1 < p< o. This is
the focus of this section, where we explicitly determine these dual spaces, giving
new meaning to Hölder's Inequality and revisiting some beautiful measure theory
along the way.
4.1.1 DEFINITION Throughout this section, fix 1 < p.q < oo, where q is the
conjugate exponent of p (in other words, + q = 1 if 1 < p < o, q = when o
p=1 and vice-versa).
(i) For each y e 9, define Q, : lP K by

Py((x})) := x,yn:
n=1

111
<br>

112 Functional Analysis

For 1 <p< o, p, is well-defined by Hölder'sInequality (Theorem 2.1.18)

and

Ifp=1 or p = o, the inequality is obvious. Hence @y E (eP)* and ||P s

lyllq. Furthermore, for any y, zE 9, and ae K, we have

Py+z = Py t Pz
and Pay = @y.
Therefore we obtain a linear operator A :
9 (eP)" given by

satisfying |A(y)|| <llyll, for allye 9.

(i) Similarly, for each g e L'(4, b], we define gg
:
LP [a, b] - Kby
Once again
operator
, f

is well-defined by Hölder's Inequality and we get a linear

A: L'|a, b] (L'{a, b])"

satisfying |A(s)|| < lsll, for allge LA|a, b].
The goalof this section is to understand when this map A is an isomorphism.

Aside. While it is true that l" andLP[a, b| are both special case of abstract
LP spaces, we treat these twocases separately to reinforce the idea and also
show the subtle differences between thesetwocases. It is true,though, that
you are about to see two separate proofs of the same fact.
€
We first introduce some useful notation: For x K, we write
x
sgn(x) := R: if #0,
: if x = 0.

so that sgn(x)x = x
for all x E K. For j E N, we will write e;
(0,0,...„0, 1,0, ...) (with 1 in the jh position) as in Example 1.1.7. Note that e; e ?
for all1<p<o0.
4.1.2 PROPosITION For 1 <
p< o, the map A
:4 (P)* is isometric.
<br>

Dual Spaces 113

Proof. In all cases, we know that | p,ll < vlla, So it suffices to prove
the reverse
inequality. We break the proof into three cases.
(i) If p =l andq = 0; Let y = (V1, y2,.) E (, Then
=
ly;l le,(e;)| s lle,lle;li = le,.
Hence, ||yllo < y as required.
(ii) Ifp= 0 and q = 1: Let y = (V1,/2,...)E e'. Then for each n E N, define
x"= (x) by
sgn(y;) : if 1 <jsn,
: otherwise.

Then xM E o and ||"o <1. Furthermore,

n

j=1 j=1

Hence |yll s ll@lI.

Let y = (y1, y2,...) e
. Then
(ii) If 1 < p< o: for each n e N, define
x"= (x) by
Jsgn(y;)lvjl-1 : if1 sjsn,
0 : otherwise.

Then xM E lP and

j=1 j=1 j=1

Also
1
1/p 1/p

-1

because qp- p plg-1) =q. Therefore we

concludethat
1/
< ll@yll.
j=1

This is true for all neN and so ||yll, s l 9yl:

<br>

114 Functional Analysis

now show that this map A is surjective, provided po. This assumptionie
We

crucial and we will return to it later in the chapter. For now recall that Coo is denss
in P for p o (Proposition 2.3.4) and it is this fact that gets used in the proof.

4.1.3 THEOREM If 1
<p<o, then the map A :
e0 (er)* is an isometrie
isomorphism.

Proof. By Proposition 4.1.2, it suffices to show that is surjective. So fix p E (P)*

€
and we wantto construct y l9 such that p = Py. If e; E P are defined as above,
we set

yi= (plei). e(ea),...).

(i) If p =1, then g = 0: For eachi e N,
|y:l = lo(ei)ls lpllleall1 = llpl.
€
Hence y o, and so Py ()". Now if x e Write x
(*1, X2,. .., Xn, 0,0,...), thern
Co0,

n
p(x) = x9(ei) = xy;=Py(*).
i=1 i=1

Now p and p, are both continuous functions that agree on Coo, which is
dense in ' (Proposition 2.3.4). It follows that p = Py.
(i) If1<p<o, then 1 <q<
o:
ForneN, write

y" i= (y1, y2,yn,0, 0,...).

Qyn E (e)*. For each n E

Then
N and x E ,
Then yM E Coo C l9, so we may consider the corresponding linear functional
let x" := (X1,X2, (xn,0,0,..).
. . .

j=1
(Z)=pr").
Hence for any x E er, we have
<br>

Dual Spaces 115

and therefore |pun||<lpl| for alln e N. Since A is an isometry, it follows

that

11 1/q

\i=1
for all n E N. We conclude that y e
, we 9 and that l|vlla < o
Now that
yE know that p, E (er)*. Furthermore, for any €
(X1, X2,..., Xn, 0, 0,...). Then
Co0, Write x

11

o(r) =Lxp(e;) = y(x)

1=1

and so = Py On Co0. Once again, since Coo is dense in ,

it follows that
= Py

Note that the same proofs work if we replace lP byE = as well, In

(K", || |)
fact, for any 1 < p< o, the map E* is an injective linear map.
A: (K",|)
Since both spaces have the same dimension, we conclude
that A is an isomorphism
(even if
p= co!).

4.1.4 COROLLARY Let 1<p<oo q

·
and let be the conjugate exponent of p. 1f E =
(K". || ||p), then E (K", || |L).
Of course, we know from linear algebra that E* K" as a vector space. The
point of this result is that we now know what the operator norm on E*
translates
tounder this isomorphism. We now turn tothe function space analogues of the ?
spaces and prove the same sort of result. Ifg is a function, we write sgn(g) for the
functiont H sgn(g()).
4.1.5 PROPOSITION For1<p< o, the map A : L9a,
(LPa, bl)* is isometric. b|

Proof. Fix g e L'a, b]. We know that g| < lglla, so it suffices to prove the
reverse inequality. Once again we break it into three cases.
(i) If p
=l and q= o:
For n E N, define

E, = {1E (a, b] Ig(t)| > ll9:| + 1/n}

:

and set f, = sgn(g)XE,. Then |f|1 = m(En) < oo and

les(fn)l = Ig()ldt > (llg:l +1/n) m(E,).

En
<br>

116 Functional Analysis

However, l@:(fu)| < l@|m(En) and so m(En) = 0 for each n e N

Therefore

m({te (a, b)
:
|g(t)|> ll@:l|}) =
m-(g-) =0.
as
Thus ||glS l@: required.
(ii) If p= 0
andg = 1: Let f = sgn(g). Then ||fo <l and

ls()]dt g:()= les()I s

= =
lisllh le:|llflleo:

Therefore |lsll1 < l@.

(ii) If 1 <p<oo: If we set
f= sgn(g)gl9-, then

|If(Pdt = | Ig(1)1(4-1)Pdt = (|isll, ) = Ig(1)1°dt = g: (f).

Ja

In particular, f e LP
(a,b] and ||fllp Also, (llgll,)9 <
=llsl. lP:llfp,
which implies that

lell, = (llgll, )y4-9/7 < lggll.

In order to determine when the map A is surjective in this case, we need to

revisit Lebesgue's Fundamental Theorem of Calculus. While we donot prove this
result (as it would take us too far off course), we will discuss the philosophy behind
it and pointyou to Royden (49]for further information on the subject. Consider the
classical Fundamental Theorem of Calculus, which consists oftwo statements:
(i) Let fe Ca, b] and define

=
F(*) f(1)dt.

Then F is differentiable on (a, b), and F= f.

(ii) If F: a, b] K is continuously differentiable on la, b\, then F is
€
Riemann-integrable and for all x [a, b],

F(x) =– F(a) +
/ F()dt.
<br>

Dual Spaces I17

Lebesgue's Fundamental Theorem of Calculus is an answer to the

corresponding questions for L' functions:
(i) Suppose f E L'a, b], then we may define

F() = f0)át.
Is F
differentiable, and, if so, is it true that F = f?
(ii) Suppose F : (a, b] C
is a is F
function, under what conditions
differentiable and F' e L'la, b]? Furthermore, if this is true, then does it
follow that
= F(a) +
F(x) F'()dt
a
€
holds for all x a, b]?
The answer to the first of these questions is the Lebesgue Diferentiation
Theorem (see [49, Section 5.3] for the proof).
e
4.1.6THEOREM (LEBESGUE'S DIFFERENTIATION THEOREM) Let f L'[a, b] and
define

F(*) = f()dt.
Then F is differentiable a.e. and F' = fa.e.

The next, rather commonly used fact must be known to you. The definition that
follows is the correct notion one needs to answer the second part of the question
raised earlier.
4.1.7 LEMMA Let fE L'a, b]. Then for any e > there exrists 6 > 0 such that for any
0

measurable set E C [a, b],

m(E) <
d f|<e.
4.1.8 DEFINITION A function F: (a,b] K is said to be absolutely continuous if
for any e >0, there exists > 0 such that for any finite collection {[a,, b;]:1<i<
n} of non-overlapping subintervals of la, b),

- F(b;) –
L(b; a;)
<8 F(a,)| <e.
i=1 =1
<br>

118 Functional Analysis

It is easy to see that an absolutely continuousfunction is necessarily continuoue

and there are examples of continuous functions that are not absolutely continuous
The most famous of these is the Cantor-Lebesgue function (see Wheeden and
Zygmund [61, Section 3.1).
4.1.9 EXAMPLE Let F: a, b
K be a function.
(i) If F is Lipschitz continuous, then F is absolutely continuous. This is because
if L > 0such that F(x) – F(y) < Lx-y| for all x,y E
[a, b], then for e > 0,
we may choose
&:=e/2L.
(ii) If F is continuously differentiable on (a, bl, then F is absolutely continuous.
This is because, by the Mean-Value Theorem, F is Lipschitz
continuous with
Lipschitz constant L := supxelab)
|F()):
(iii) If there exists f E L'|a, b] such that

F(x) =
then it follows from Lemma 4.1.7 that F is absolutely continuous.
Lebesgue's Fundamental Theorem of Calculus now completes
this circle of
ideas very succinctly (once again see Royden [49] for the details).
4.1.10 THEOREM (LEBESGUE'S } THEOREM OF CALCULUS) Let
F:la, bl K be an tsAMENTAL
absolutely continuous function. Then
(i) F isdifferentiable a.e. and F'E L'|a, b].
(ii) Furthermore, for almost every x a, b],

F(x) = F(a) +
F'()dt.
We are now in a position to use Lebesgue's theorem to
determine the dual space
of LPla, bl. To that end we need a lemma
that will serve us welI. The assumption
that p oin the following lemma is the first hint that the dual of L°[a, b] willbe
more difficult to determine than
that of the other LP spaces (just as it was in the
case of sequence spaces).
4.1.11 LEMMA Let 1
<p< oo and let g
E L'a, b]. Suppose that there exists
such that
M>0
rOs)a s MIflI,
for all bounded functions fe LP[a, b]. Then g E Lla, b) and
|gll, S M.
<br>

Dual Spaces 119

Proof. Once again we break theproof into two cases.

p n
(i) If =1:Let eNand consider
En = {x E (a, b): |g(r)| > M + 1/n}.
We wish to prove that m(E,) = 0. If we set let fy = sgn(g)xE,, then ||fl1 =
m(En) and

However,
=
fa()g(1)dt <M|f Mm(E,)

and therefore n €
(E,) =0. This is true for all n N,and hence

e (a,
b| |g(t)| > M)) = m
:
m({t =0.
n=1
Thus g EL°a, b] and ||g < M as required.
(ii) If 1 <p< o: Since g E L'(, b], it follows that |g(x)| < o a.e. Forn e
N,
define

= Js(*) : if lg(x)|
< n,
&n(x)
|0 : otherwise.
Then each g, is measurable, bounded and gn g pointwise a.e. If we set
fn i= sgn(gu)lgn |9-, then
b
=
lfl= I8n(t)|4-)Pdt
Ja
ISm(t)|"dt =|lsl18

In particular, f, e L'a, and |lp b] = ll8|1. Also, lg|9= fngn = fug

and hence

Therefore lglla S M for all ne N. Finally, by Fatou's Lemma,

flg()"44 s liminf |
|ls(t)|Pdt
lsn ()1°d4 < MI.

We conclude that ge L'[a, b] and ||gll, < M.

<br>

Functional Analysis

.
120

We are alnost ready to describe thedual space of LP|a, b] for 1 p< The
next lemma is the penultimate step in that process and a crucial one. It allows us to
on
construct an absolutely continuous function fromn bounded linear functional
a

LPla, b].
4.1.12 LEMMA Let 1
sp<o and e (L'(4, b|)". Define F: (a, b] - K by
=
F(r) p(x)
Then F is absolutely continuous.

Proof. We wish to show that for any e > 0 there exists 6 > 0 such that for any
finite collection {la;, b;] :1 <i< n} of non-overlapping subintervals of la, b],

LIF(bi) - F(a:)| <e.

1
L(b;- a) <i» =1

So assume is non-zero, fixe > 0 and let {la;, b;):1<i<n} be a finite collection
of non-overlapping intervals in (a, b]. Define

–
sgn(F(b;)
f= i=1 F(ai))x(o,b]

Then

i=1

and
1
- F(a;))o(xa,b) - F(a:)|
p()=Lsgn(F(b,) =IF(b:)
i=1

since q(X(b,ai) = F(0:)– F(«:). Therefore

lgllfll, = lpl(
–
Z\F(b;) F(a:)| < b;-a;|)/P.
i=1 =1

We conclude that S := e/2l||P works.

Recall thatfunction f: a, bl
a
K is said to be a step function if it is a finite
linear combination of characteristic functions of sub-intervals of a, b]. In particular,
a step function is a simple function. Let S denote the set of all step functions on
<br>

Dual Spaces 121

la,b]. Then it is clear that S c LPla, bl for any 1 <p < o.

a small piece oft
The next lemma 1
the puzzle we need to complete our proof and we leave it as an
exercise for the reader.

4.1.13 LEMMA Let S denote the set of allstep

functions on an interval a, b].
co, then S dense in LP{a, b. Furthermore, if f E L'la, bl is a bounded f1s pK
is
function, then
there is a sequence (fn) C S and M
all n
>0 such that f, f in LPa, and |f| S M jor
b]
EN.
We are finally in a position to prove the main
result of the section.

4.1.14 THEOREM (RIESZ REPRESENTATION THEOREM (F RIESZ,

1909)) If1 S
p<o, then the mapA : La, b (LPa, bl)* an
is isonetric isomorphism.
Proof. By Proposition 4.1.5, it suffices to prove that A is surjective. So €
fix p
(LPla, b|l)* and we want to construct g
L'a,b| such that p = pg. As you might
expect, we start with the function F: a, b
C given by

F(*) := (x(ax)

By Lemma 4.1.12, F is absolutely continuous. Hence by Lebesgue's Fundamental

Theorem of Calculus there exists g E L'(a, b] such that

oXax) F(x) = g(t)dt = X(e(1)g(4)dt.

Hence if fe LPa, b] is a step function, then by linearity it follows that

o()= f08)dt.

Now let SC L'a, b] denote the subspace of step functions. lIf f e LP [a, b] is a
bounded function, then by Lemma4.1.13 there is a sequence (f) E S and M>0
such that fn fin LP{a, b] and |fs M for all n e N. Replacing (f) by a
Subsequence if necessary, we may assume that fn f pointwise a.e. Hence fug
f8 pointwise a.e. and |fngl < Mg E L'a, b| for all n e N. By the Dominated
<br>

122 Functional Analysis

Convergence Theorem,

rOs()dt - lim / fa()g()dt = lim p(fa).

1-0Ja

Since ||fnllp ||fll, (by Remark 2.1.2), it follows that

lim lp(G)|< lim llplllll, =| plllfp

1o 1-00

Therefore the inequality

holds for all bounded measurable functions f E LPla, b]. By Lemma 4.1.11 we
conclude that g E L°(a, b] and gg s lpl
Now consider the bounded linear functional @, E (LP[a, b])* and observe that
p(f)= P (f) for all fe S. Since S is dense in LP(4, b], it follows that = g. This
concludes the proof.
The Riesz Representation Theorem explicitly identifies the dual space of L' [a, b)
) for 1 < p < o. However,
(and also in the course of the proof, we have
conspicuously left out the case of p = o0, In fact, the proof relied on various facts
(such as the density of step functions, the Dominated Convergence Theorem and
others) which only work for p < o. In order to resolve this question, we need to
now explain why the mapA: L'a, b (La, b|)* is not surjective (it is, of course,
isometricand therefore injective).
In order to do this, we need a recipe to construct bounded linear functionals
with prescribed conditions. The most general such theorem is the Hahn-Banach
Extension Theorem, which is what we turn to next.

4.2 The Hahn--Banach Extension Theorem

So now we know something about the duals of the P and LP spaces, but what
about a general normed linear space? If youthink about it, we do not even know
if the dual space is non-zero! Of course, ifE is finite dimensional, then E is
non-zero. However, for infinite dimensional spaces, the obvious 'dual basis' to a
given Hamelbasis may not yield bounded linear functionals (indeed, they will not
see Exercise 5.33). Therefore we need away toconstruct bounded linear functionals.
<br>

Dual Spaces 123

The Hahn-Banach Theorem is the most general way to do so. Its usefulness cannot
be overstated and it is one of the most important results in the subject.
Suppose E is a normed linear space and F subspace. Given a bounded
<Ea
linear functional:F K,we would like to construct a bounded linear functional
b:E Ksuch that

ple= 9.
In other words, would be a continuous extension of o. Furthermore, we
like ||
already seen
=
l9 Hence would be a
norm-preserving extension of .
would
We have
twosituations in which this is possible: If F = E, then there is a unique
continuousextension of p by Proposition 3.2.4; and if E is aHilbert space, then such
is a continuous, norm-preserving extension exists by Corollary 3.2.6.
However, in the gerneral case, we need a new idea toconstruct such an extension.
If codim(F) =1, then we can write

E= {x +
xe: xE F, a E K}
for some ee E\F.Hence we only need to define y(e) e K in such a way that
lp(z)| < llolll| (4.1)
for all z E E(this would automatically imply that = ||gl|). Note that the real
lll
issue here is not the existence of an extension. An extension may be constructed
by defining y(e) to be anything in K. However, it is preserving the inequality of
Equation 4.1 (and thereby ensuring that the extension is continuous) that is the
difficult part.
In order toensure that the result we obtain is widely applicable, it makes sense
to prove it in a little more generality than what is needed in this specific instance.
Recall from Definition 2.5.1 that a seminorm is a function p :E R+ on a vector
space that has all the requirements of a norm except that it need not be positive
definite. The next lemma is the most basic step in the theorem and the one that
needs the most attention. Since it needs the underlying field to have an order
structure, the proof only applies to real-valued linear functionals.
4.2.1 LEMMA Let E be a vector space 0ver IR and p: E R be a seminorm on E. Let
F<E beasubspace of E of codimension one and let p: + F R be a linear functional on
F satisfying

p(*) < p(x)

<br>

124 Functional Analysis

for all x EF. Then there exists a linear functional y :

E R such that yr= p and
p(x) < p(x)
for all x E E.
€
Proof. Since codim(F) = 1, there exists e E E\ F such that every z E can be
expressed in the form 2 = x+ ae for some (unique) x E Fand aE R. As mentioned
above,we need to determine a real number
t i= p(e)

that p(z) < p(z) may hold for all z

€
So E. In order to do so, we first assumne
that such an extension p exists. We use it to determine conditions that this real
number ' must satisfy. Having arrived at this necessary condition, we then proceed
to show that such a real number exists and complete the
proof by observing that
this necessary as
condition is, in fact, sufficient well.
(i) Suppose y: E Rexists such that yE= p and p(z) < p(z) for all z E E.
Then set t := y(e) so that y(x+ ae) = p(x) + at < p(x+ ae). Consider the
two cases:
(a) Ifa >0, then tmust satisfy the inequality

p(x+ ae) - p(x)

for all xEF. Hence
p(x + ae) - p(x)
t<th:= inf :xeF,a €
R>o

(b) If a<0, we
-a €
write B= >0. Then for any y F,

p(y- Be) = py) - Bt < p(y- Be).

Hence t must satisfy the inequality

P(y)-py- Be)
B
<br>

Dual Spaces 125

for all y€F. Therefore

Be)
!2h:=
t> t2
i= supP9)-p(y-
B
:yeF,fERso
However, this is only possible if t <t so let us verity this tact.
€
(ii) We want to show that for all x, y Fand all &, B>0, the inequality

p(y)-p(y - Be)
p(x+ ae) - p(x)
B

holds. Cross-multiplying,this amounts toproving that

p(ay)+ (Bx) s Bp(x + ae) + ap(y - Be).

Since p is a seminorm, this reduces to proving that

p(ay + Bx) p(Bx + aße) + p(ay - aße).

Now this last inequality follows from a calculation:

p(ay + Bx) = p(ay - aße t aße + Bx)

= -
p(ay aße) + (Bx + aße)
<p(ay - aße) + p(Bx + aße).
Therefore we conclude that t2 < tj as desired.
(iii) Having discussed the necessary condition in Step (i), wenow construct the
extension. Consider t and tz as above so that t2 S th. We choose tE R such
that t2 <<t :E by
R
and define

p(x+ ae) := p() + at.

Then this map ý is linear and satisfies the condition that ý(z) < p(z) for all
zE E.
Having completed the codimension one case, we need Zorn's Lemma to
complete the argument.
<br>

126 Functional Analysis

CASE (HAHN, 1927, AND

4.2.2 THEOREM (HAHN BANACH THEOREM: REAL
BANACH, 1929)) Let E be vector space over R and p: E
a
R be a seminorm o%

E. Let F<E be a subspace and p:F Ra linear functional such that

p(x) < p(x)
for allxEF. Then there exists a linear functional y : E R such that

and p(x)s p(*) for all x E E.

Proof. Ifn := codim (F) < o, then we repeat Lemma 4.2.1 inductively.
If not, then we appeal to Zorn's Lemma. Define F to be the set of all pairs
(W, pw), where W < E is a subspace containing F and pw : W R is a linear
functional on W such that

€
and ýw(x) < p(x) for all x e W. The set F is clearly non-empty since (F, g) F.
Define a partialorder < on F by setting (W1, w;) < (W2, w;) if and only if
W; C
Wz and w,w, = W,

It is easy to check that F now becomes a partially ordered set. We now verify that
Zorn's Lemma is applicable: Let C be a totally ordered subset of E. Define
Wo := U
W.
(W,yw)eC
Then since C is totally ordered, W is asubspace. Define yo :Wo Rby
po(x) := pw(x)
if x e W.
Then one can check that yo is well-defined (again since C is totally
ordered). Furthermore, yo(x) < p(x) holds for all x E Wo. Hence (Wo, o) E
an upper C.
F
and it is clearly bound for
Zorn's Lemma now tells us that F has a maximal element, which we denote by
(Fo, po). Now we claim that Fo = E and this is where we
need the previous lemma.
Suppose not, then there exists eE E\ Fo. Define Fj := span(Fo fe}). Then by
Lemma 4.2.1 we may extend po to a linear map P1 :
F R satisfying
P(r) <p(x)
€
for all x F. This would ean that (F1, p1) € F, which would contradict
the maximality of (Fo, po). Hence Fo = E, and y = Qo is the required linear
functional.
<br>

Dual Spaces 127

We now to prove the analogous theorem for complex vector spaces. In

wish

order to do this,we need a natural way to pass from complex linear functionals to
reallinear functionals and that is what the next lemma does for us.
4.2.3 LEMMA Let E be a complex vector space.

(i) If o
:E Ris an R-linear functional, then

@(x) := p(r) - ip(ix)

is a C-linear functional
(i) If : E Cis a C-linear functional, then qg:= Re(p) is a R-linear functional
and @ = y.
(iii) If p is a seminorm and p and @ are as in part (i), then the inequality

lp()|sp(x)
holds for all xe Eif and only if

l@(x)| <p(r)
holds for all x e E.

(iv) IfE is a normed linear space andp and areas above, then ||goll
=|l.
Proof. The proofs of the first two statements are left as an exercise.
(ii) Suppose that the inequality

l9()|< p(x)

holds for all x E. Then, clearly, the same inequality holds with .
Conversely, suppose

l9(x)| s p(r)
€ = 1 such that
holds for all x E, then fixxEE and choose À
EC with |A|
lp(x)| = Ao(x). Then

l9()|= @(Ax) =
Re(p(Ax)) = p(Àx) s p(Ax) = |A|p(x) = p(r).
Thus @(x)|< p(x) for all x E E.
<br>

128 Functional Analysis

(iv) Let p(x) := llpll||x||. Then |p(r)| < p(x) for all x e E. By part (ii), l@(x)| <
€ reverse inequality is obvious since
p(x) for all x E and so ||pl< l. The €
lp()| =| Re(@(x))| < l@(r)| for all x E.
The transition from the real case to the complex case is now easy.
4.2.4 THEOREM (HAHN BANACH THEOREM: GENERAL CASE) Let E be a normed
linear space and p:E R4 be a seminorm on E. Let F < E be a subspace and p :F K
a linear functional such that
-
lp(*)| < (x)
for allx E F. Then there exists a linear functional ý: E K such that ylr= and

l(x)| s p(x)
for all x€ E.

Proof.
(i) Suppose E is an R-vector space: By Theorem 4.2.2, there exists a linear
functional y: E R such that
ý(x) < p(x)
for all x€ E. Then

(x) = p(-*) <

p(-x) = p(r)
holds for all x E E,
and so p(x)|< p(x) holds.
(i) Suppose E is a C-vector space: Define f :F R by f = Re(p). Then f is
linear and

|f(x)| < p(x)

for all x e F. By part (i) there is a linear functional g : E R
such that
g|r=f
and lg(*)I s p(x) for all x E. Define :
E
C by y = as
in
Lemma 4.2.3. Then y is C-linear and satisfies
lp(x)| < p(r)
for all xE E. This completes the proof.
<br>

Dual Spaces 129

The first theorem extending linear functionals was proved by

F Riesz
(1910-11),who wanted an analogue of the Riesz-Fischer Theorem (part
()
of Theorem 3.5.5)
for L spaces when p 2. The next such result was due to
Helley (1912), who introduced the important step of enlarging
the original
subspace by one vector (Lemma 4.2.1). It was in the independent work of
Hahn (1927).and Banach (1929) that we see the emergence of
the complete
proof of the theorem for real Banach spaces. Both of them used Helly's idea
and added to it the use of transfinite induction (Zorn's Lemmna would not
exist until 1935).

The complex version of the theorem was proved by Bohnenblust and

Sobczyk (1938). They were also the first to refer to it as the Hahn-Banach
Theorem.

The next corollary is the way the Hahn-Banach Theorem gets used most of the
time. In fact, each corollary we state here is often taken to be a part of a suite
of results collectively termed 'the Hahn-Banach Theorem'. This slight abuse of
language is indicative of how ubiquitous the theorem is in the literature.
4.2.5 COROLLARY Let E be a norned linear space and
F<Ea subspace. Then for any
bounded linear functional E F* there exists pE E* such that

and ||| = |lpll.

Proof. Apply the Hahn-Banach Theorem with p(x) := |lplllx|
4.2.6 COROLLARY Let E be a normed linear space, {ej, e2,..., e}C E be a finite,
linearly independent set and {a1,a2, ...,
an} cK be arbitrary scalars. Then there exists
a bounded linear functional yp E E such that

y(ei) = aj
for all 1
<i<n.
Proof. Take F:= span({e1, e2,..,en}) and define g:F K
by

p(e;) = ai,
extended linearly to all of F. Since F is finite dimensional, is a bounded linear
functional on F (by Corollary 2.4.9). We may now apply Corollary 4.2.5 to get a
linear functional E E* satisfying the required conditions.
<br>

130 Functional Analysis

x
The next corollary is an interesting one. It states that for€ any vector in
normed linear space E, there is a bounded linear functional which satisfo E

lp(*)| = x| In other words,the norm of the vector can be realized by evaluation

by bounded linear functionals. In particular, it resolves one important question
we mentioned at the start of the section; we finally see that E" # {0} whenever
E {0}. That such a basic fact required a proof speaksvolumes for the subtleties
of the subject.

4.2.7 COROLLARY Let E bea normed linear space and x E E. Then

x= sup{|p(x)| : E E", lpll < 1).

Furthermore, this supremum is attained.

€
Proof. Set a i= sup{|p(x)| : p E', lpl| < 1}. Then, it is easy to see that & <
l||| Conversely, set F ;= span(r). Define p:F K by

Bx B||x||

Then pE F*; indeed, it is easy to check that |||| =1. By Corollary 4.2.5 there exists
yEE* such that y(x) = x|| and ||4|| =1. Hence |||| = as desired.
€
If E is a normed linear space and p E, then we know that

llpll= sup{|p(y)|:ye E, |lyll < 1}. (4.2)

Corollary 4.2.7 may be thought of as amirror image of this fact except that the norm
in Equation 4.2 may not be attained.
The next corollary, and the last of this section, states that the dual space
separates points of the underlying normed linear space. It a straightforward
is
consequence of Corollary 4.2.7.

4.2.8 COROLLARY If x, y
E E are two vectors such that p(x) =
v(y) for all ýE E,
then x =y.

We will revisit the Hahn-Banach Theorem in Chapter 6,

where we will explore
the deep relationship between a normed linear space and its dual space. We
will establish a geometric version of the theorem, which exploits the notion of
convexity. For now, though, let us give some useful consequences of the theorem.
<br>

Dual Spaces 131

4.3 Duals of Subspaces and Quotient Spaces

Let E be a normed linear space and F < E be a subspace. We wish to
understand the
relationship between the dual spaces E* and F*.For any bounded linear functional
o€ E", it is clear that its restriction olr is a bounded linear functional on F and that
ll olell < p This gives us a bounded linear map
R:E* F* given by

The Hahn-Banach Theorem merely says that this map is

surjective.
4.3.1 DEFINITION The space ker(R) := {p E E* : pl= 0} is called the
annihilator of F and is denoted by F. Note that F is a subspace of E* and that it
is closed.
€
If E isa Hiibert space, then for any E* there exists y E E such that
p(*) = (x, y)
€ E
for all x (by Theorem 3.2.2). Therefore glE= 0if and only if y eF (as defined
in Definition 3.1.1). This justifies the abuse of notation, although one should keep
in mind that when E is not a Hilbert space, F is asubset of E* (and not of E).

4.3.2 PROPOSITION Let F be asubspace of a normed linear space E. Then the map R :

E* F* given by

induces an isometric isomorphism R: E*/F F* given by

p+F H pr.
Proof. As mentioned above, F is a closed subspace of E*. Hence E*/F is a
normed linear space (by Proposition 2.5.2). That R exists follows from the First
Isomorphism Theorem. Furthermore, since R is surjective, R is an isomorphism of
vector spaces.
€ F we
It remains to show that R is isometric. So fix E E*. Then for any y
have plr= (p+ )E. Hence

|R(p)| < |lp+yll.

<br>

132 Functional Analysis

€
This is true for all
F
and therefore

|R(@ F)|| = |R(p)ll < lp+ F||.

+

For the reverse inequality, set n := olE. Then n E F* so there exists E E* such

that y|=) and = |yl bythe Hahn-Banach Theorem. Hence

= = |lyl| > ||p + F|
||R(+F)| ll
F= p+ F. We conclude that
+
But y -pE F so that y + ||R(p F)|> lo+
F|l, as desired.
Thenext theorem may be viewed as a strengthening of Corollary 4.2.8. There
we were able to separate points in a normed linear space by a bounded linear
functional. Here one separates a closed subspace from a point outside it by means
of a bounded linear functional.

4.3.3 PROPOSITION Let E be an normed linear space, F < Ea subspace, xoE E\Fand
Suppose that

d:= d(xo, F) > 0.

€
Then there exists p E* such that

(a) p(*) = 0 for all x E F,

(b) p(xo) =1and

(c) |l| =d-l.
Proof. Since d(xo, F) = d(xo, F) > 0, we may assume without loss of generality
that F is closed. If n : E E/F denotes the natural quotient map, then ||7(xo)||
= d.

Hence by Corollary 4.2.7 there exists p E (E/F)* such that ||p| =1and
p(x0 + F) = ||x0 + F|| = d.

Define Q:EK by

p:=do t.

Then E E
satisies conditions (a) and (b). Toverify condition (c), observe that
|7|< 1
and therefore
<
lp()lsd-llyllx(v)|| d'lvl
<br>

Dual Spaces 133

for all yE E. Hence

<d-'. However, llbll = 1 and hence
+ +
sup{|p(z F)|: ||2 F| =1}=1.
Therefore there exists a sequence (x, +
F) E
E/F such that ||x, + F =
1for all
nEN and
lim |y(x, + F) = 1.

Let yn E F such that ||Xn ty €

<1+1/n for each n N so
that
dl4(xn +
F)| = |p(In +
y)| < llellln +ynll S llpll(1+1/n).
Letting n o, we conclude that d-l< |lol|.
The next corollary is remarkably widely used in the subject and is another one
of those results that is simply subsumed under the title of "Hahn-Banach Theorem'.
4.3.4 COROLLARY Let E be an normed linear space, and F<E be a
subspace. Then

F= 0 ker(p).
pEFI

In other words, a vectorx€ E belongs to Fif andonly if p(x) = 0 for all e F.

Proof. Clearly,

W:= n ker(p)
pEF1

is a closed subspace containing F and hence Fc W. Conversely, if xo F, then

€
d(xo, F) >0. By Proposition 4.3.3 there exists p E* such that F C ker(@) and
0.
p(xo)

Hence xo W. We conclude that W C

Fas well.
above, the notion of the annihilator F arises most naturally in
As discussed

the context of Hilbert spaces. Along those lines, the next result may be thought
of as an analogue of Corollary 3.1.14. Its proof is an immediate consequence of
Corollary 4.3.4.
4.3.5 COROLLARY Let E be an normed linear space andF<Ea subspace. Then F is
dense in Eif andonly if F = {0}.
<br>

134 Functional Analysis

Now let F < be a closed subspace so that E/F is a normed linear space. We
E

wish to identify (E/F)* in terms of F and E. Note that 7t E E/F is continuous

:

Hence if is a bounded linear functional on E/F, then

p p o
7T:E K is bounded
and linear. Thisgives us map Q:(E/F)*
a E", given by

This map is clearly linear as well. The next theorem shows that is injective and
identifies its range.
4.3.6 THEOREM Let F be a closed subspace of a normed linear space E. Then there is an
isometric isomorphisnm Q: (E/F)* F.
Proof. Note that for any ye (E/F)",
Q(y)r=ýo n|= 0
e we have
since n|= 0. Hence Q() F, Since ||| <1,
(4.3)
Conversely, by Proposition 2.2.11 applied toE/F, we have
llpll =suplly(x+ F)|:x+ FE E/E, ||x + F|=1}.
Therefore there exists a sequence (x) C
E such that xy +
F|| =1 for all n eN
and

lim lý(3n + F)| = |lpll.

For each n
e N there exists y, E F such that|X, +yn|<1+1/n. Hence
|Q()(In tyn)|= lp(xn+ yn + F)| = p(x, + F)|.
Thisimplies that

||Q(1+1/1) > lp(*n +

F).
Passing to the limit, we conclude that ||() | Together with Equation 4.3,
it
followsthat ||Q()| =|l proving that Qis isometric.
now show that is surjective: If p E
F, then o€ E* and pl= 0. Hence
We

induces a map :
E/F -
Kgiven by
x
+F p(x).
<br>

Dual Spaces 135

This map is clearly well-defined and linear. Furthermore, since

€ or
is bounded,
it follows trom orollary 2.5.6 that p (E/F)*. By construction, Q(y) = P and
hence Q is surjective.

4.4 Separability and Reflexivity

Let us now return to an unresolved question from the first section of this chapter.
Recall that for each y = (yn) E e, the map , : poo
K, given by (*k)
1S a
i1 Xnyn bounded linear functional. Furthermore, the map
A:e (oo)*
given by A(y) := Py, is an isometry by Proposition 4.1.2. Similarly, we have an
isometry
A:L'(a,b] (L°(a, b|)*.
Earlier in the chapter, we had hinted that these maps may not be surjective. Armed
with the Hahn-Banach Theorem, the proof of this fact is now at hand. To drive the
point home, we give two proofs. We will first explain why these maps ought not
to be surjective (for abstract reasons) and then give examples of bounded linear
functionals that lie outside their ranges.
The firstof these proofsstems from the following result. Recall that a topological
space is said to be separable if it has a countable dense set.
4.4.1 PROPOSITION Let E be an normed linear space. If E* is separable, then E is
separable.

Proof. If E* isseparable, then it is easy to see that the unit sphere

E
Si={p E": l|| =1}
is separable.We choose a countable set {p1, P2,...}cS thatis dense in S.For each
neN, |l@,ll =1. Therefore there exist vectors xn E E such that ||*,| =1and
lPn(xn)| > 1/2.
Let F = span{x, : n> 1}. If we set Ko
=Qor QxQ according as K= RorC,
then
:
D;= spank, (Xn n>1}
is countable and dense in F. Hence F is separable. We wish to prove that F = E,
and in order to do that, we would like to appeal to Corollary 4.3.5. Suppose p E F
<br>

136 Functional Analysis

is non-zero bounded linear operator, then

a we may assume that l p|= 1. Since
o(x,)= 0 for all n e N, it follows that

l- Pll>\p(x») - pu()| > 1/2

€ €
for all n N. This contradicts theE fact that {@
:
n N} is dense in s
=
Thus F {0} and F is dense in by Corollary 4.3.5. We conclude that E is

separable.
44.2 COROLLARY The maps

A:e (eoo)* and

A:L'(4, b] (L°la, b|)*

are not surjective

Proof. If A were an isomorphism, then (oa)* would be isomorphic to !

However, isseparable so this would imply
Example 2.3.16. Similarly, (L°la,b|)* L'(a, b].
o
was separable, which contradicts

We now give 'concrete' examples of bounded linear functionals that lie outside
the range of A. That such functionals exist is, of course, a consequence of the
Hahn-Banach Theorem. The first example we construct (of a bounded linear
functional on loo) is quite a bit more general than what we need for our purposes.
However, this class of functionals is sufficiently important that they warrant such
attention.
Let S: poo loo denote the left-shift operator

S(x1, X2, x3, ..) = (x2, X3,...).

Let denote the vector space of all convergent sequences in K (see Exercise 2.27),
c

,
thought of as a closed subspace of Define c :
K by

p(x;)) := lim xn
This is well-defined,bounded and linear. Therefore by the Hahn-Banach Theorem,
we may extend this toa bounded linear functional on o, We wish to construct a
particular kind of extension so that we may have some additional properties. Such
a linear functional is referred to as a Banach limit.
<br>

Dual Spaces 137

44.3 THEOREM (BANACH

poo
LIMITS) There exists a bounded linear functional :

Ksatisfying the followingproperties:

(i) If x = (x}) E c, then y(x) = limy-yoo Xn.
:
(ii) IfS 0 denotes the left-shift operator, then p(x) = p(S(r))for allx e
l.
(i) ||l=1.
Proof. Define

F:={x- S(x):xE .

Then F is clearly a subspace of l°, If 1 := (1,1, 1,. .), then we claim that d(1, F) =
.

1. Firstly, observe that

d(1, F) < ||1|| = 1.

Now, if x = (X) E °o and z = (Z) := X-S(x) + 1, then z = Xn - Xn+1 +

1.
Therefore for any NeN, we have

N 1 N

N4 -1|
||z|| N|-*w41-1)
=1 |i=1
+12 1- N

Since |XN+1 – X1| < 2||x|| we may let N o

to conclude that ||zl> 1. Hence
d(1, F) >1 well and this proves the claim.
as
By Proposition 4.3.3 there is a bounded linear functional ý: oo K
such that
plr= p(1) =1and || p|| =1.
0, We y
verify that has the remaining twoproperties
as well.

(ii) x €
For any €
,x- S(x) =
F. Therefore y(x) p(S(x).
(i) Now if x = (x;}) E c, then set
a := lim xy.
10

>0 there exists N E N such that x, - a<e

€
Then for any for all n> N.
Set

y=(*N, XN41, XN+2,...) = SN (x),

<br>

138 Functional Analysis

where SN denotes the composition of S with itself N times. It follows from

property (b) that

(x) = (y).
However, by construction, ||y- a1|| <e. Hence
IP(x) - a| = ly()-a| = l(y-a1)| < ly -a1|| Se.
Thisis true for any e > 0 and therefore (x) = a.

We now return to our original problem of showing that the maps A :

(e)•
and A : L'a, b] L° (a, b] are not surjective.

4.4.4 EXAMPLE

(i) Let E
(oo)* be a Banach limit, as constructed in Theorem 4.4.3. We claim
that ý A(e'): Suppose there exists y = (y;) such that y = A(y) = Py, then

lim x;=
11=1

for all (x;) E c. In particular, for j e N fixed, consider e; e o0 (as defined in

Example 1.1.7). Then e; E Co C c so

y;= V(e;) = 0.

Thus y=0. But if this were true, it would imply that y = 0. However, #0
€
since y(1) =1.Therefore there is no y el such that p = A(y).
(ii) Consider C[a,b] C
L°[a, b) and define : Cla, b) K by

pf):=f(b).
Then p is a bounded linear functional so by the Hahn-Banach Theorem,
there exists a bounded linear functional ý : L° [a, bl K such that
p() = f(b)
for all f E Cla, b]. We claim that y €
A(L'|a,b|). Suppose there exists g
z,
L'a, b] such that y = then g would satisfy
f(0) = fe)g()d#
<br>

Dual Spaces 139

for allf E Cla, b]. For n e N, let fu E

cla, b) be given by

: if a <t<b-1/1,
fn(t) = 1 : if t= b,
linear : otherwise.

€
Then for any t a, b),

lim f.(t) =0.

1oo

€
Hence fn 0pointwise a.e. Since fr Lo [a, b] and g E L'a, b|, we may
apply Dominated Convergence Theorem toconclude that

lim y(f) = lim fn(t)g(t)dt =

0.
no 10oJa

Howeve, y(fn) = fn(b) = 1 for all n e N. This contradiction shows that

y Ps for any g e L' (a, b].

Having sorted out this question, we turn to the other property mentioned in the
title of thissection - reflexivity. The idea is to turn the tables on the dual space and
let the underlying normed linear space E act on E*. This action leads to a natural
inclusion of E into its double dual, the dual of E*. We wish to understand this map
and what it tells us about E.

4.4.5 DEFINITION Let E be an normed linear space.

(i) The double dualof E is the dual ofE* and is denoted by E'* := (E*)".
(ii) For each x E E, define î: E* Kby

î(@) :=p(x).
Note that î is a linear functional on E* and

sup{lp(x)| : pE E",|lpl =1} = |x||

=
||| (4.4)

E**.
by Corollary 4.2.7. Therefore f E
<br>

140 Functional Analysis

(iii) Define ]: E E** by

J(x) := .
Then J is a linear transformation, which is isometric by Equation 4.4.
(iv) Eis said to be reflexive if J is an isomorphism fromE to E**.
4.4:6 EXAMPE
(i) If E is finite dimensional, then it is reflexive.
Proof. IfE is finite dimensional, then
dim(E) = dim(E") = dim(E**).
Since J is injective, it must be surjective.
(ii) Every Hilbert space is reflexive.
Proof. Given the Riesz Representation Theorem, this proof is almost
tautological. Let H be a Hilbert space and define A : H H* by A(y) :
where

Py(x) := (x,y).
Then an conjugate-linear isomorphism of normed linear spaces.
A is In
particular, H* is a Hilbert space under the inner product

(Py. P:) := (z,y).

Hence if T e H**,then by the Riesz Representation Theorem applied to H

there exists p E H* such that

T() = (ý. ¢)
€
for allp E H".By the Riesz Representation Theorem there exists y H such
that = Py- For anyze H,taking y = @, in the above equation gives
=
T(P:) (Pz, Py) (y,z).
Now consider îe H** and observe that
99:)= Pz(y)= (y, z).
Therefore T = jand so J is surjective.
<br>

Dual Spaces 141

(iii) For 1
<p< oo,
lP is reflexive.
Proof. Let q e (1, oo) be such that 1/p + 1/g = 1and let A,: 9 (eP)*
be the map from Definition 4.1.1. Similarly, let : eP A
(e9)* denote the
corresponding map obtained by letting £P act on
". Then by Theorem 4.1.3,
both maps induce isomorphisms
Ap : (P)* and Ag : eP (09)".
Nowif TE (P)**, then T: (eP)* Kis bounded and linear. Hence :
To Ap
l9 K is bounded and linear. Since A, is surjective there exists xE P such
that

A,(x) = To Ay.
Hence for any y E 9, we have

*nyn = A,(x) (y) = T(A,(y)).

L
1=1

However, we observe that

R(Ap(y)) = (9) = gy
(x) = Xhyn

1=1

Since A, is surjective,we concludethat î= T. Therefore J :P (lP)** is

surjective.
(iv) A similar argument shows that LP|a, b] is reflexive, provided1 <p<oo.

Aside. When discussing reflexivity, one must be a little careful. There is an

example (due to James (28]) of a Banach space E that is isomorphic to E"* but
fails to be reflexive! The point is that reflexivity requires a specific map to be
an isomorphism. The mere existence of an isomorphism does not guarantee
reflexivity.

Once again we are left wondering about and e (and their continuous
analogues). In order to resolve this, we need to know some properties that reflexive
spaces ernjoy.
4.4.7 PROPOSITION Let E be a reflexive space. Then E is separable if and only if E* is
separable.
<br>

142 Functional Analysis

Proof. If E* is separable,then E is separable by Proposition 4.4.1. Conversely, if :

is separable and reflexive, then E* = (E')" is separable. Hence E* is separable b.
Proposition 4.4.1.

4.4.8 PROPOSITION Let E be a reflexive space andF<Ea closed subspace. Then F

i
reflexive.

Proof. We have a isometric maps JE : F F** and JE : E E**. Assuming I

is surjective, we want to show that IE is surjective. So if T E F**, then we wish
€
to show that there exists x F such that T = IE(x). Consider T : F* K as a
bounded linear functional and define S:E* K by

S(¢) :=T(p|r).

Then S is clearly a linear functional and

|S(@)| = |T(p|F)| s ||T||glr| < ||T|llgll.

Hence S is a bounded and therefore in E**, Since E is reflexive there exists x € E

Such that

S= JE().
We claim that xE Fand that T = JF(x). Suppose x ¢ F. Then by Proposition 4.3.3
there exists p E E* such that plr=0 and p(x) = 1. However, this would imply that

1= p(*)= Je(*)(P)=S(p) = T(p|r) = T(0) = 0.

This is a contradiction and therefore it must happen that x E F.Now for any p E€F,
we choose a Hahn-Banach extension E E of p. Then p(*) = p(*) since x F.
Therefore

T(p)= T(ý|e) = S(P) = JE(*)() = y(*) = p(r) = Jr(*) (p).

We conclude that J is surjective.

We are now in a position to prove the lack of reflexivity of the remaining familiar
spaces.
<br>

Dual Spaces 143

EXAMPILE
44.9
Any reflexive space is necessarilycomplete because the
(i)
dual space of any
normed linear space is complete. Therefore (Coo,l) is not reflexive for
any 1 Sp o
(see Proposition 2.3.4 and Exercise 2.26).
(ii) For the same reason, (Cla, b\, || |) is not reflexive, if 1 <p <
Proposition 2.3.12).
o (by
(ii) is separable, but its dual space loo is not separable (by
Example 2.3.16).
Therefore is not reflexive. Similarly, L'la, b] is not reflexive either.
(iv) Cla, b| (equiPped with the supremum norm) is not reflexive because Cla, b]"
is not separable (Exercise 4.31), while C[a, b] itself is
separable.
(v) L°a, b| is not reflexive since it has a closed subspace Cla, bl that is not
reflexive.
(vi) From Exercise 4.13, there is an isometric isomorphism e' (co)*. Hence

(co)** oo.

Now coo is dense in co (by Exercise 2.26),which tells us that co is separable

(by Remark 2.3.15). Since o is not separable, it follows that co is not
reflexive.
(vii) Since co is a closed subspace of °, it follows fromn Proposition 4.4.8 that loo
is not reflexive.

We must now point out an important gap in our knowledge. While we know
that Cla, b]* and (o )* are both non-separable, we do not as yet know what they
are. We willstudy these spaces in detail in Chapter 9. For now it suffices to know
that both these dual spaces are vector spaces of measures and each such measure
inducesalinear functional by integration.
With that mysterious comment, we move on to more tractable topics.

4.5 The Transpose of an Operator

Given a normed linear space E, we have now begun to understand the role of the
are
dual space E*. In many ways, it is the mirror image of E, and properties of E
reflected in those of E* and vice-versa. Now we will expand this relationship to
include linear operators. Given an operator T between normed linear
spaces, we
<br>

144 Functional Analysis

a natural dual operator between the corresponding dual spaces. Thi

will define
operator mirrors the properties of T and allows us to prove interesting results abo
T simply knowing its relationship with its transpose.
spaces, where th
This idea truly comes to fruition in the context of Hilbert
betwe
transpose of an operator is replaced by its adjoint. The deep relationship
an operator and its adjoint leads to some of the most striking results in the subiers
and willbe the focus of our attention later in the book. For now
our goal is mnodes:
We will define the transpose and give some useful examples.
5.1 DEHNITIÖN Let T : E F be a linear operator between two normed linear
space. The transpose of is the operator T: F*
T E* defined by the formula

T'(@)(*) :=p(T(«)).

F*, then po T:E

If p E -
K is a continuous
€ linear map since it is tha
composition of two such functions. Thus T(@) E*. Furthermore, if P. P2 E F,
then
T'(P1 + p2) = T'(p1) + T'(P2)
as can be seen by applying both sides to elements of E. Similarly, T' respects scalar
multiplication and we have confirmed that T is a linear map. In other words, our
definition makes sense and we may proceed without further ado.
4.5.2 EXAMPLE If E and F are both finite dimensional, we fix two ordered bases
AE = {ej, e2,..., en} and AF = {f1, f2,...fm} of E and F, respectively. Let A =
(a;.;)be the matrix associated to T in these bases. In other words, for each 1 <j<n,
M
T(e;) = a,jf;
i=1
= .. , @r}
and AF: = {y1, 2,... ,Ym} be the ordered dual bases to
.
Let Ag {p1. @2,
Ag and AF, respectively, and let B = (bij) be the associated matrix for the operator
T' in these bases. In other words, for each:1 < k< m,
T(Vk) = bexPe.
l=1

Then
bji = T(V:) (e;) = :(T(e;)) = La,j V:(fi) = a,j:
1=1

Therefore the matrix B is the transpose of the matrix A in the sense that you must
be familiar with from linear algebra.
<br>

Dual Spaces 145

The next example is a simple generalization of

this idea toinfinite dimensions.
45.3 EXAMPLE Let 1 <p< o
and 1 <r<oand let T E B(eP,e) be a bounded,
linear operator. Detine an infinite matrix
M= (a;;) by the formula
T(e;) = (a,j,
a2,j).
Under the identification A, :
9 =
(eP)* and A, : s
(e)* (where q and s
are the conjugate exponents to p
and r, respectively), we wish to determine the
corresponding matrix for the operator T', thought as a
linear map from to 1.
of

Tobe precise, we wish to understand

the map

A,oTo A, : 9.

For clarity, we write AE =

e1, e2,...}
for the standard basis in lP, and AF :=
{f, f2...} for the corresponding vectors in . We write Ag = {p1,
P2,.} C
(CP)* where

Pr((x;))= X.

Then AE: plays the role of the dual basis to Ag. Similarly, we define 'Ap :=
{1, V2,..}c ()* where n((y)) := ym. As before, we have

T'(P:)(e;)= y:(T(e;)) =La,:;) =

aij

,
i=1

= A,: we have
Therefore T(P:) L1 4j9; Composing with (eP)*

A,' oT (:) = (a,1,aj2, ...).

If we write {81,82, ...} for the standard basis of, then A, (8:) = ; so that

A,'oT' o A, (&:) = (a41 aj2,...).

In other words, A,'oT'o A, is associated to the matrix (a,), which is thetranspose

of the matrix M.
<br>

146 Functional Analysis

Let us now look at a continuous analogue of this example and

attempt
coordinate-free approach to the question of computing the transpose.
p< o
and a measurable function
simplicity, we fix 1 <
4.5.4 EXAMPLE For
K: (0, 1] x [0,1] such that
K

a := Sup
sup
|K(s, t) lds < oo, and ß := sE[0,1] | K(s, f) |dt < oo.

te[0,1] Jo

Then we let q denote the conjugate exponent of p. For any fE LP0, 1|, we have

IK(9,)||f()ldt=K(s, +)|/|K(s, t)|PIf()ldt

JO
1/p
sp'(K(s, )l|f(O)Pdt )

by Hölder's Inequality. Hence

ds < prlo K(S,

)||f()| Patds
/0 J0

Hence we may define an integral operator T :LP0, 1] LP{0,1] by

T()(S) := K(s, t) f(1)dt

and T is a bounded linear operator with |T|| <B/9al/P.

We now compute the transpose T', thought of as an operator on Lª(0,1], via the
identification A, : L9[0,1] (LP(0, 1])°. For anygE L0,1] and f E LP[0, 1], we
have
= (9)g (s)ds
T(A,(s))(f) A,(s)(T())=T)

-K(s, ) f()s(:)dtds
- K(s, t)g(s) f (t) dsdt
<br>

Dual Spaces 147

Lsy
the Fubini-Tonelli Theorem (which is applicable because
ji(0, 1]2) by Hölder's Inequality). Define S : L9|0, the integrand iS 1n
1] 1] L°0, by

S(8)() := [ K(,)s(s)ds.

Then, by the calculation above, S is a

bounded linear operator, with
Furthermore,
||S||< aP.
T'(Ap(g)) ()= () f(t)dt = A,(S(3))(f).
S)
Thisistrue for any f € L0,1] and ge L9(o, 1], so
we see that the transpose of T is an A,' oT'oA, = S.Once again
integral operator whose kernel is the 'transpose'
of the kernel K of T.

45.5 EXAMPLE Let E be a normed linear space

andF<Ebe a subspace of E. If
T:F E denotes the inclusion map, then
T': E* F* is the restriction map R
from Proposition 4.3.2.
Similarly, if F isclosed and : E E/Fis the quotient map,
t then 7t: (E/F)*
E* isthe map Q from Theorem 4.3.6.

4.5.6 PROPOSITION If TE B(E, F), then T' € B(F*, E*) = ||T||.

and |T'
€
Proof. For any F*, T'(p) is a
bounded linear functional on E. If xe E, then

|T'(P)(*)|= lp(T(*))|< lpll|T(x) || < .

ll@lllT|l
Thus|T (P)|l s |T||lp. This is true for any p e F and thus T' is a bounded linear
operator with |T"|| s ||T||. Conversely, if x e E,then by Corollary 4.2.7 there is a
€
bounded linear functional po F* such that pol = 1 and lpo (T(x))| = ||T(*) |.
Therefore

< =
||T(*)|| =|T'(9o)(*)| |T(po)|laxl| ||T'|llgol|llx|| ||T'llx|.
Thus ||T|| < ||T'| holds as well.

The proof of the next result is entirely straightforward and we leave it as an

exercise for the reader.
<br>

148 Functional Analysis

4.5.7 PROPOSITION LetE, F, andW be normed linear spaces.

(i) Let S and T E B(E, F) and a
e K. The (T +
S)'=T'+ S' and (aS)' = «s
=
(ii) IfS E
B(E,F) and T e B(F, W), then (TS)' S'T.
Given a subspace F of normed€ linear space E, we had defined its annihilato.
a

F in Definition 4.3.1 by F = fp E':plr= 0}. The next definition is meant to

be 'dual' to this.
a
4:5.8 DÈFINITION Let E be a.normed linear space and W be subspace of
E*

Define

w= {rE E:p(r) = 0 for all p

E W}.

In contrast with Definition 4.3.1, W is a subspace of E, not of E**. At this point

we must apologize for the abuse of notation. However, this is standard practice
sense
and it is probably for the best if yousimply confirm that this notation makes
in the case when E is a Hilbert space.
We end this section with'a simple fact relating an operator to its transpose. We
will see many more results like this as we go along, but this should give youa sense
for how these arguments are meant to go.
PROPOSITION Let T: E F be a bounded linear operator and let T' :F*
4.59 E'
be its transpose. Then
(i) ker(T') = Range(T)-.
(i) ker(T) = Range(T').
Proof. Part (1) is an easy exercisefor you to try (Exercise 4.40). For Part (i), we
prove containment both ways. If x€ ker(T), then for any E F", T(p) (x) =
p(T(*)) =0. Hence E Range(T'). Thus
x

ker(T) C Range(T)t.
Conversely, if xe Range(T')+, then for any q eF, T(@)(x) = (T(*)) =€ 0. Since
x ker(T)
F* separates points of F (by Corollary 4.2.8), this forces T(x) =0. Hence
and. this proves that Range(T') ker(T) as well.
C

This concludes our discussion on the dual space for the moment. We will return
to it in Chapter 6, armed with few more tools. Until then do try the problems
a

uses of the Hahn-Banach

given below and pay close attention to the variety of
Theorem.
<br>

Dual Spaces 149

4.6 Exercises
TheDuals of L
Spaces
EXERCISE 4.1 Let y = (Y1,y2,..) be a sequence in K. For each x = (In) E
Suppose that the series Lil Xiy; converges. Furthermore, suppose that the map
o: (0 Kgiven by

i=1

defines a bounded linear map on oo, Prove that y

e e and that yll1 = |pl|

BXERCISE 4.2 (2) Prove Lemma 4.1.13.

EXERCSE 4.3 < p<1, define LP[o, 1] exactly as in Definition 2.1.15. Define
For 0

a metric dp :
LP0,1] x LP[0, 1] R4 by -
dy(f.g) := If()-s()| Pdx.

The aim of this exercise is to show that the only linear functional on (LP(0, 1], d,)
that is continuous is the zero map (succinctly, LP[0, 1]* = {0}).
(i) Suppose p :
K is a non-zero, continuous linear functional. Let
LP0,1
LP0, 1] be such that pf)l> 1. Prove that there exists s E 0,1] such
fE
that

=
If(*))Pdx If)Pdx.
2 Jo

s € as above, set g1 := 2fx(0s] and g2 = 2fx(s.1]: Prove that

(ii) With [0, 1|
lp(8)l> 1 for
someie {1,2} and that d, (&, 0) = 2P-'d,(f, 0).
(ii) Use this idea to construct a sequence (fn) in LP{0, 1] such that

lim d,(fn, 0) =
1-0
0

and \p(f)l>1 for all nE N.

(iv) Conclude that LP[0,1]* = {0}.
<br>

150
Functional Analysis

The Hahn-Banach Extension Theoren

EXERCIŠE 4.4 Prove parts (i) and (ii) of Lemma 4.2.3.
EXERCISE 4.5 Let E := R²,eguipped with the supremum norm.
€ be given by p(x, y) =
:= E: y = 0} and let p:F
R
(i) Let F
{(x,y)
Prove that p has a unique norm-preserving extension to
E.

€ be given by o(*,y) =y
(ii) LetF := {(x, y) E: x = y} and let p: F p
R

are
Determine all norm-preserving extensions of to E. Observe that there
infinitely many such extensions.
a
EXERCISE 4.6 (2) Let E be a normned linear space and {e1, e2,...,en be
are bounded linear functionals
finite, linearly independent set. Prove that there
{P1 P2,..., @n} CE* such that
: ifi= j
Pi(e;) = 6i,j = :
0 if i #j
for all 1<i,j<n.
Note: The analogous statement does not hold for an infinite linearly independent
set- have a
look at Exercise 5.33.
EXERCISE 4.7 Let E and F be normed linear spaces with E non-zero. If B(E, F) is
a Banach space, then prove that F is complete.

EXERCISE 4.8 (4) Let E and W be normed linear spaces with W finite
dimensional. If F<Eis a subspace and T:F W is a bounded linear operator,
then prove that there is a bounded linear map S : E W such that

S|= T.

Give an example to show that this need not happen if W is infinite dimensional.
EXERCISE4.9 Let E
be a normed linear space and T:Eo be a bounded linear
map.
(i) Prove that there is a sequence (Pu) C E
such that
T(x) = (P(x), p2(x),..).

Furthernmore, prove that ||T|| = sup,EN llPll:

<br>

Dual Spaces
151

(1i) If F <E is a subspace and

that there is an extension T:F
poo a
is
bounded linear map, then prove
S: E poo
such that Slg= T and |S| ||7|:
SA61DEFINITION A normed
linear space E is said to be strictly convex
distinct x, yE Ewith ||x|| =yl we it, tor any
=1, have

<1.
(To understand
the terminology, interpret this as
statement about the unit ball.)
EXERaRCISE 44.10
(1) Prove that any Hilbert space
is strictlyconvex.
(ii) If1 <p<o,then show that l is strictly convex.
Hint:Read the proof
of Hölder's Inequality carefully.
(ii) Prove that and are not strictly convex (see
Exercise 2.4).
Note: Parts (ii) and (iii) are true for the function spaces LP a, b as well.
EXERCISE 4.1l(1AYLOR, 1939) Let E be an normed linear space such that E* is
strictly Convex. If F< E and E F, prove that there exists a unique y E E* such
p
that ý|r= and
lpll = llpll.
In other words, p has a unique norm-preserving extension.
EXERCISE 4.12 Let (E,|E) be a separable normed linear space.
(i) Prove that there is a countable collection {@1, P2,..} CE* such that
|xE = Sup p,(r)|
nEN
for each x e
E.

Hint: Take a countable dense subset in the unit sphere SE and apply
Corollary 4.2.7 to each element of it.
(ii) ForxE E, define

|xllo :=
1=1
2"

Prove that || lo is a norm on E and that it is strictly convex.

<br>

152 Functional Analysis

(iii) Prove that there is a strictly convex norm c on E, which is equivalent to

Note: It is a theorem due to Bourgain [5] that does not admit an equivalent
/co
strictly convexX norm. Therefore the conclusion of Exercise 4.12 does not hold
fo

non-separable spaces (see Exercise 4.22).

.For each y
EXERCISE 4.13 (2) Let E = (co, || |loc) and F = (yn) E F, define
Py :E
K
by

Py((x;)) := 11=1
Xnyn:

From Exercise 2.14, we know that @, E E* and that ||@yl= yl1. Hence we get a

map A:
F E*given by
y Py

which is linear and isometric. Prove that this map is an isomorphism.

EXERCISE 4.14 In Exercise 4.13, we have constructed an isometric isomorphism
' (co)*. By Theorem 4.1.3, we have an isometric isomorphism Aj (oo
: :
Ao
(e')*. Therefore we have an isometric isomorphism

If J: Co co
is the natural map from Definition 4.4.5, then show that the
composition I oJ : Co - is simply the natural inclusion map co o
€
EXERCISE 4.15 Let E be a normed linear space, F < Ebe a subspace and F

be a bounded linear functional on F. Let

Sp i= {ye E*:plE= p
and |lp|| = |lpll}.

In other words, Sp is the collection of all norm-preserving extensions of p. Prove

that S is a non-empty, convex, closed and bounded subset of E*. Give an example
to show that So may not be compact.

The Dual of Subspaces and Quotient Spaces

€
EXERCISE 4.16 Let E be a normed linear space such that for any p E' there
€
exists x E such that ||x|| =1 and |lo(x)| = ll. Let F <Ebe a proper closed
<br>

Dual Spaces 153

eubspace of E. I'rove that there exists a vector Xo E E

such that |xo| = 1 and
d(xo, F) =1.
In other words, if every element of E* attains itS norm on
the unit sphere of E,
hen the conclusion of Riesz' Lemma holds for t
=1. In particular, this applies to
reflexive spaces by Exercise4.28.
EXERCISE 4.17 Let E
be a
normed linear space and F
be a
subspace of E. For any
xE E, prove that
d(x, F) = sup{ly(r)|: E F, llpll <1}.

Separability and Reflexivity

ecssgtesss
EXERCISE 4.18 Let E be a
normed linear space and S be any subset of E. If E is
separable, then prove that S is separable.
Note: The analogous statement is true for metric spaces, but it is not true for
arbitrary topological spaces, so be careful!

EXERCiSE 4.19 (4)

(i) Follow the line of reasoning of Exercise4.13 to show that there is an isometric
isomorphism A :
c* (compare this result with Exercise 2.38).

(ii) Prove that c is not reflexive.

EXERCISE 4.20 Let E be a normed linear space and F < E
be a closed subspace
of E.
(1) IfE is separable, then prove that E/F is separable.
(ii) Give a (non-trivial) example where E/F is separable, but E is not.
EXERCISE 4.21 (2) Let E be a normed linear space and F<Ebe a closed subspace
of E. IfF and E/F are both separable, then prove that E is separable.
.
EXERCISE 4.22 Consider co as a closed subspace of
€
() Let S {(x) := (: x; E {0,1} for allkE N}. If x = (xn),y = (y) €S
are such that x; # yk for infinitely many values of k e N, then prove that
|x-y+ coll = 1 (see Exercise 2.42).

(i) Prove that /Co is not separable (this also follows from Exercise 4.21).
<br>

154 Functional Analysis

EXERCISE 4.28 If 1<p<o, provethat B(eP) = B(P, eP) is not separable.

Hint: Construct an isometric linear map T:o B(eP).

EXERCISE 4.24 For any normed linear space E, show that (E")** = (E**)",

EXERCISE 4.25 (4) IfE is a

reflexive normed linear space, then prove that E is
reflexive.

EXERGIE 4.26. (2) If E is a Banach space such that E* is reflexive, then prove that
E
is reflexive.

Hint: Use Exercise 4.25 and the fact that the natural map J:E - E** is isometric.

EXERCISE 4.27 1f E
is reflexive andF <E is a closed subspace, then prove that
E/F is reflexive.

Hint: Use Theorem 4.3.6 and Exercise 4.26.

EXERCISE 4.28 be a reflexive Banach space. For every p E E',prove

(2) Let E
that there exists x E E such that ||x|| = 1and p(x) = lpl|. Furthermore, if E is
assumed to be strictly convex (see Definition 4.6.1), then prove that this point x
must be unique.
EXERCISE 4.29 Give an example of a bounded linear functional E (e')* such
that l@(x)| < lllllx|| for all xel'.
EXERCISE 4.30 Use Exercise 4.13 to determine the form of every bounded linear
functional p E (Co)* which attains its norm on the closed unit ball.
EXERCISE 4.31 (2) Let E = (Cla, b), | |o) and fix t e (a, b]. Define :E
@; K
by p:(f):= f(). Prove that p: E E* and that ||@: =1. Use
the collection {o :tE
la, b]} to show that E* is not separable.

EXERCISE 4.32 Let E be a reflexive normed linear space and let f: [o, 1] Ebe a
€
continuousfunction. For each E",the function p o f: j0,1] K is continuous.
Therefore we may define T:E* by K

T(@) := of))dt.
<br>

Dual Spaces
155

) Prove that T is linear and

bounded. Concude that there exists a
vector xE E such that unique

o()=p(f()dt
for all p € E*.This vector is
denoted by

(ii) Prove that

Note:These results are true even if E is not reflexive.

However, the proof is much
longer (see [34, Section 3.3]).

The Transpose of an Operator

EXERCISE 4.33 (2) Prove Proposition 4.5.7.
EXERCISE 4.34 Let T: (' co map
be the T(x;)) := (y1,y2, ..), where

iM:
Yn i= Xm
1M=1

Observe that T is a bounded, linear operator. Under the identifications

of
Exercise 4.13 andTheorem 4.1.3, determine the transpose map
T':el lCo

EXERCISE 4.35 Let E be a

normed linear space and p E E". Compute p' as an
element in B(K, E").
EXERCISE 4.36 Let T: be the right-shift operator given by

T(x)) := (0,x1, X2, ...).

Under the identification o (l')*, determine the transpose map T':( oo,
<br>

156 Functional Analysis

EXERCISE 4.37 Let E be a normed linear space and W be a subspace of E*,

(i) Prove that W
is a closed subspace of E.
(ii) If F is subspace of E,then prove that (F) =F.
(iii) IfW is a subspace of E*, then show that (W) W.
EXERCISE 4.38

(i) IfE is reflexive and W is subspace of E*,then prove that (W) = W.

(ii) Let E = and let W = co, thought as a subspace of
of
(e)* a. Prova
that
W

(W.
EXERCISE 4.39 Let A :
L' (a, b] (L°la, b])* be the map given in Definition 4.1.1
Treating F:=Cla, b] as a closed subspace of
L°[a, b), prove that
A(L'|a, b])
nF= {O}.

This is another way of seeing that A is not surjective.

Hint: See the proof of Proposition 2.3.12.
EXERCISE 4.40 (2) Prove Part (i) of Proposition 4.5.9.
EXERCISE 4.41 If T: E F is a topological isomorphism, then prove that T' :

F* E is also a topological isomorphism and that

(T-ly = (T')-1.
EXERCISE 4.42 (2) Let T :E Fbe a linear operator and T" := (T)':E* F*
be the transpose of T". If JE : E E** and JF :F F** denote the canonical
inclusions, then prove that
T" o JE = J; oT.

Banach Limits and Amenable Groups

4.6.2 DEFINITION A Banach limit is a bounded linear functional ý o
:
K
satisfying the properties listed in Theorem 4.4.3). A sequence (x,) loo is said x=
to be almost convergent if there exists L E K such that y(x) = L for any Banach
limit p. If that happens, wwe write
Lim(x) = L.
<br>

Dual Spaces 157

EXERCISE 4,43
(i) Show that the sequence x = (1,0, 1,0,...) is almost convergent.
Hint: Consider u = x +
S(x), where S: poo oo denotes the left-shift
operator.
Let us now generalize this: A sequence x = (x,) is said to be periodic
if there exist natural numbers p
N and such that Xy+u=X, for all n
2N.
(ii) If x E is periodic, then prove x
that is almost convergent. In fact, if N and
p are as above,
then prove that
1
Lim(x) = =(XN + XN+1 +...+ XN+p-1).
EXERCISE 4.44 Consider ° as the real vector space of real-valued sequences
and
let y
: Rbe a Banach limit.
() If x = (XH) E ( is such that x, >0 for all n € N, then prove that p(x)>0.
Hint: IfO< X, <1 for all n € N, then |1– x
<1.
(ii) Ifx=(X,) E is a real-valued sequence, then provethat
lim inf x, < y(x)< lim sup x,.
1-o
11oo

Hint:Forthe first inequality, let a:= lim inf,-oXy. Then for any e> 0
there
exists NEN such that x, >a-e for all n > N.
4.6.3 REMARK The proof of Theorem 4.4.3 also applies to ( (Z) rather than
(o0= (o(N). In other words, there is a bounded linear functional y: ( (Z) K
satisfying all the properties given in Theorem 4.4.3 (note that the shift operator
S: o(Z) (Z) is defined as S((x;))n :=Xn+1).
EXERCISE 4.45 Consider (Z) as a real vector space of real-valued sequences
and let p : oo (Z) R be a Banach limit as described above. For Ac Z, let
XA E oZ) denote the sequence
1 : if n € A,
(XA) :=
|0 : otherwise.
Define :22 R by

(A) ;=pxA).
Prove the following properties of u:
<br>

158 Functional Analysis

() p is positive: u(A) >0for all A CZ.

(ii) p is finitely additive: If A, B C Z are disjoint, then u(AUB) = u(A) + u(B
(iii) 4 is normalized: u(Z) =1.
(iv) u is translation invariant: For each and m e Z, u(A+m) =
ACZ ulA)
(v) u is not countably additive.

46.4 DEFINITION A countable group G is said to be anmenable if there is a eo,

function u : 24 R+ such that
u
i) is positive, finitely additive and normalized (in the sense of Exercise 4.4s)
(ii) u is left-translation invariant: If A €
CG and g G, then (gA) = uA)
where gA := {gx:xE A}.

With a slight abuse of terminology, such set functions are also called finitely additie
1neasUres.

EXERCISE 4.46 Prove that a finite group is amenable.

EXERCISE 4.47 Let F

denote the free group on two generators {a, b}. Prove that
IP2 is not amenable.

€
Hint: For x {a, b, a-,b-}, let W, be the set of words in F2 starting with x.

EXERCISE 4.48 Prove that a subgroup of an amenable group is amenable.

Hint: Let # be a measure on the group G that witnesses amenability.

If H is a
subgroup of G, let M be a set consisting of exactly one representative from each
right coset of in G. Now define v :2H R by
H

v(A) :=p U
Am
mEM
<br>

Dual Spaces 159

Additional Reading
The result of Exercise 4.11 is a theorem
dueto Taylor from 1939.
Taylor, Angus Elis (1939). The extension of
linear functionals. Duke
Mathematical Journal, 5, 538-547.
ISSN:0012-7094.
The converse iS also true and was
proved in a one-page paper!
Foguel, Shaul R. (1958). On a theorem by A, E. Taylor.
Proceedings of the
American Mathematical Society, 325. ISSN: 0002-9939.
DOI: 10.2307/2033162.
The Taylor-Foguel theorem wasfurther explored by
Phelps, who proved that
a subspace has the property
that every bounded linear functional on it has a
unique extension if and only if its annihilator has a certain 'best approximation
property' (á la Theorem 3.1.6).
Phelps, Robert Ralph (1960). Uniqueness of Hahn-Banach extensions and
unique best approximation. Transactions of the American Mathematical Society,
95, 238-255. ISSN: 0002-9947. DOI: 10.2307/1993289.

• The space of almost convergent sequences (see Exercise 4.43) was

studied by
Lorentz.
Lorentz, George Gunter (1948). A contribution to the theory of divergent
sequences. Acta Mathematica, 80, 167-190. ISSN: 0001-5962. DOI: 10.1007/
BF02393648.
In this paper, he characterizes such sequences and uses that characterization
to show that many interesting classes of sequences are almost convergent - so
much so that the set of almost convergent sequences is not separable!
• In Exercise 4.45, we have proved that the integers (Z, +) form an amenable
group. These groups have been studied extensively because amenability is
reflected in a variety of ways in the operator algebras associated to the group
(we will see operator algebras later in the book). A good place to start learning
about these objects is this book by Runde.
Runde, Volker (2002). Lectures on Amenability. Vol. 1774. Lecture Notes in
Mathematics. Berlin: Springer-Verlag. ISBN: 3-540-42852-6. DOI: 10.1007/
b82937.