Professional Documents
Culture Documents
~ IN MATHEMATICS
Functional Analysis
s. Kesavan
I1,Jr,i1@HINDUSTAN
UillJ UBOOK AGENCY
TEXTS AND READINGS
IN MATHEMATICS 52
Functional Analysis
Texts and Readings in Mathematics
Advisory Editor
C. S. Seshadri, Chennai Mathematical Institute, Chennai.
Managing Editor
Rajendra Bhatia, Indian Statistical Institute, New Delhi.
Editors
R. B. Bapat, Indian Statistical Institute, New Delhi.
V. S. Borkar, Tata Inst. of Fundamental Research, Mumbai.
Probal Chaudhuri, Indian Statistical Institute, Kolkata.
V. S. Sunder, Inst. of Mathematical Sciences, Chennai.
M. Vanninathan, TIFR Centre, Bangalore.
Functional Analysis
s. Kesavan
Institute of Mathematical Sciences
Chennai
fl:[gUgl HINDUSTAN
U U2J UBOOKAGENCY
Published in India by
email: info@hindbook.com
http://www.hindbook.com
All export rights for this edition vest exclusively with Hindustan Book
Agency (India). Unauthorized export is a violation of Copyright Law
and is subject to legal action.
The present book grew out of notes prepared by myself while lectur-
ing to graduate students at the Tata Institute of Fundamental Research
(Bangalore Centre) and the Institute of Mathematical Sciences, Chen-
nai. The material presented in this book is standard and is ideally suited
for a course which can be followed by masters students who have cov-
ered the necessary prerequisites mentioned earlier. While covering all
the standard material, I have also tried to illustrate the use of various
theorems via examples taken from differential equations and the calcu-
lus of variations, either through brief sections or through the exercises.
In fact, this book is well suited for students who would like to pursue a
research career in the applications of mathematics. In particular, famil-
iarity with the material presented in this book will facilitate studying
my earlier book, published nearly two decades ago, 'Topics in Func-
tional Analysis and Applications' (Wiley Eastern, now called New Age
International), which serves as a functional analytic introduction to the
theory of partial differential equations.
The first chapter of the present book gives a rapid revision of linear
algebra, topology and measure theory. Important definitions, examples
and results are recalled and no proofs are given. At the end of each
section, the reader is referred to a standard text on that topic. This
chapter has been included only for reference purposes and it is not in-
tended that it be covered in a course based on this book.
vi
Chapter 6 introduces the Lebesgue spaces and also presents the the-
ory of one of the simplest classes of Sobolev spaces.
Occasionally, some hints for the solution are provided. It is hoped that
the students will benefit by solving them.
Chennai, s. Kesavan
May 2008.
Notations
Certain general conventions followed throughout the text regarding
notations are described below. All other specific notations are explained
as and when they appear in the text.
• Sets (including vector spaces and their subspaces) and also linear
transformations between vector spaces are denoted by upper case
latin letters.
• To distinguish between the scalar zero and the null (or zero) vector,
the latter is denoted by the zero in boldface, i. e. o. This is also
used to denote the zero linear transformation and the the zero
linear functional.
1 Preliminaries 1
1.1 Linear Spaces . . . . . . 1
1.2 Topological Spaces . . . 11
1.3 Measure and Integration 18
3 Hahn-Banach Theorems 69
3.1 Analytic Versions . . . 69
3.2 Geometric Versions .. 77
3.3 Vector Valued Integration 81
3.4 An Application to Optimization Theory 85
3.5 Exercises 92
6 LP Spaces 162
6.1 Basic properties . · 162
6.2 Duals of LP Spaces · 168
6.3 The Spaces LP(n) .170
6.4 The Spaces W1,P(a, b) .178
6.5 Exercises ...... · 188
Bibliography 265
Index 266
Chapter 1
Preliminaries
x+y = y+x;
(ii) (associativity) for all x, y and Z E V, we have
x + (y + z) = (x + y) + z;
(iii) there exists a unique vector 0 E V, called the zero or the null
vector, such that, for every x E V,
x+o = x;
x + (-x) = o.
2 1 Preliminaries
Definition 1.1.3 Let V be a vector space and let Xl, ... , Xn be vectors
in V. A linear combination of these vectors is any vector of the form
alxl + ... +anxn , where the ai, 1 :S i :S n are scalars. A linear relation
between these vectors is an equation of the form
alXl + ... + anXn = 0.•
1.1 Linear Spaces 3
Example 1.1.2 The space ]RN (respectively, eN) has a basis which
is defined as follows. Let 1 ::; i ::; N. Let ei be the vector whose i-th
component is unity and all other components are zero. Then {el' ... , eN}
is a basis for]RN (respectively, eN) and is called the standard basis . •
where the ai and bi are real numbers and x is the variable. Let a E R
Assume, without loss of generality, that m ::; n. Define
n
(p + q)(x) = I)ai + bdxi
i=O
v= WI + ... + Wn
where Wi E Wi for each 1 :::; i :::; n. The spaces are said to be independent
if an element in the span is zero if, and only if each, Wi = O. In particular,
it follows that, if the Wi are independent, then for all 1 ::; i, j ::; n such
that i f. j, we have Wi n Wj = {O}. Further, every element in the span
will have a unique decomposition into vectors from the spaces Wi'
Definition 1.1.7 Let V be a vector space and let Wi, 1 ::; i ::; n be
subspaces. Then, V is said to be the direct sum of the Wi if the spaces
Wi are independent and their span is the space V. In this case we write
Definition 1.1.8 Let V and W be vector spaces (over the same base
field). A linear transformation, or linear operator, is a mapping
T : V -> W such that for all x and y E V and for all scalars a and (3,
we have
T(ax + (3y) = aT(x) + (3T(y).
1.1 Linear Spaces 5
If W is the base field (which is a vector space over itself), then a linear
transformation from V into W is called a linear functional on V. •
The coefficients (tij) in the above relation form a matrix with m rows
and n columns. Such a matrix is referred to as an m x n matrix. The
j-th column of the matrix represents the coefficients in the expansion
of T(vj) in terms of the basis { ~1 of W. Of course, if we change
the bases for V and W, the same linear transformation will be given by
another matrix. In particular, let dim(V) = n and let T : V - t V be a
linear operator. Let T be represented by the n x n matrix (also known
as a square matrix of order n) T = (tij) with respect to a given basis. If
we change the basis, then T will be represented by another n x n matrix
T = (tij) and the two will be connected by a relation of the form:
6 1 Preliminaries
where P is called the change of basis matrix and represents the linear
transformation which maps one basis to another. The matrix p-l rep-
resents the inverse of this change of basis m p~n and is the inverse
matrix of P. In this case, the matrices T and T are said to be simi-
lar. The identity matrix I represents the identity mapping x 1-+ x for
all x E V for any fixed basis of V. For a given basis, if T : V ~ V is
invertible, then the matrix representing T- 1 will be the inverse of the
matrix representing T.
A square matrix is said to be diagonal if all its off-diagonal entries
are zero. A n x n square matrix A = (aij) is said to be upper trian-
gular (repectively, lower triangular) if aij = 0 for all 1 :S j < i :S n
(repectively, aij = 0 for all 1 :S i < j :S n). It can be shown that every
matrix is similar to an upper triangular matrix. A matrix is said to be
diagonalizable if it is similar to a diagonal matrix.
Given an m x n matrix and two vector spaces of dimensions nand m
respectively along with a basis for each of them, the matrix can be used,
as in relation (1.1.1), to define a linear transformation beteween these
two spaces. Thus, there is a one-to-one correspondence between ma-
trices and linear transformations between vector spaces of appropriate
dimension, once the bases are fixed.
Definition 1.1.11 If T = (tij) is an m x n matrix, then the n x m
matrix T' = (tji), formed by interchanging the rows and the columns of
the matrix T, is called the transpose of the matrix T. If T = (tij) is
an m x n matrix with complex entries, then the n x m matrix T* = (tTj )
where tTj = tji (the bar denoting complex conjugation), is called the
adjoint of the matrix T . •
If x and y E ]Rn (respectively, en), then y'x (respectively, y*x) repre-
sents the 'usual' scalar product of vectors in ]Rn (respectively, en) given
by L~~l XiYi (respectively, L~=l x/fh). If the scalar product is zero, we
say that the vectors are orthogonal to each other and write x ~ y. If
W is a subspace and x is a vector orthogonal to all vectors in W, we
write x ~ W.
Definition 1.1.12 Let T be an m x n matrix. Then its row rank is
defined as the number of linearly independent row vectors of the matrix
and the column rank is the number of independent column vectors of
the matrix. •
The column rank is none other than the rank of the linear transfor-
mation defined by T and the row rank is the rank of the transformation
1.1 Linear Spaces 7
Proposition 1.1.2 For any matrix, the row and column ranks are equal
and the common value is called the rank of the matrix. •
Corollary 1.1.1 Ann x n matrix is invertible if, and only if, its nullity
is zero or, equivalently, its rank is n. Equivalently, a linear operator on
a finite dimensional space is one-to-one if, and only if, it is onto . •
n
Definition 1.1.14 Let T be an n x matrix with complex entries and
T* its adjoint. The matrix is said to be normal if
TT* = T*T.
TT* = T*T = I.
x*Tx ~ O.
Remark 1.1.2 A hermitian matrix is equal to its adjoint and the inverse
of a unitary matrix is its adjoint. A matrix T, with real entries, which
is equal to its transpose is called symmetric and one whose inverse is
its own transpose is called orthogonal. In case the matrix T has real
entries, we can still define positive semi-definiteness (or positive definite-
ness) by considering real column vectors x in the above definition. •
8 1 Preliminaries
det(ST) = det(S).det(T).
In particular, T is invertible if, and only if, det(T) =I- 0 and
det(T- 1 ) = (det(T))-l.
(iii) If T is an n x n matrix, then
det(T') = det(T) . •
Tx = b.
This is because the corresponding linear transformation is invertible and
hence onto (which gives the existence of the solution) and one-to-one
(which gives the uniqueness of the solution). If T is singular, this is no
longer the case and we have the following result.
1.1 Linear Spaces 9
Tx = b
has no solution or has an infinite number of solutions. The latter pos-
sibility occurs if, and only if
b'u = 0
Tu = AU.
if i = j,
if i 1= j.
1.2 Topological Spaces 11
Ai RT( Vi)
RT(X)
maxxE\!;
mind Vi-I RT(X)
minwcc n , dim(W)=i maxxEW RT(X).
In particular,
For more details on linear spaces, the reader is referred to, for in-
stance, Artin [1].
It is clear from the above definition that X and 0 are both open and
closed. Further, a finite union and an arbitrary intersection of closed
12 1 Preliminaries
sets is closed. It then follows that given any set A eX, there is a
smallest closed set containing it. This is called the closure of the set
A and is usually denoted by A. If A = X, we say that A is dense in
X. Similarly, given any set A eX, there is a largest open set contained
in A. This set is called the interior of A and is denoted by A o . A set
A c X is said to be nowhere dense if (A)O =.0.
If {X, T} is a topological space and if Y eX, then Y inherits a
natural topology from that of X. The open sets are those of the form
Un Y, where U is open in X.
The set B(x; r) described above is called the (open) ball centred at x and
of radius r. It is a simple exercise to check that open balls themselves
are open sets. It is also immediate to see that this topology is Hausdorff.
On]R or e, we have the 'usual' metric defined by
d(x, y) = Ix - YI·
The topology induced by this metric will be called the 'usual' topology
on ]R or e, as the case may be. Similarly, on ]RN (or eN), we have
the 'usual' Euclidean distance which defines a metric on that space: if
x = (x, ... , XN) and Y = (YI, ... , YN) are vectors in]RN (respectively, eN),
then
Definition 1.2.7 Let {Xi, 'Ii}, i = 1,2, be two topological spaces and
let f : Xl -+ X 2 be a given function. We say that f is continuous
if f-l(U) is an open set in Xl for every open set U in X 2 • If f is
a bijection such that both f and f- l are continuous, then f is said to
be a homeomorphism and the two topological spaces are said to be
homeomorphic to each other. •
This is called the distance of the point x from the set E. The following
proposition is easy to prove.
Proposition 1.2.3 Let {X, d} be a metric space and let E eX. Then
(i) for all x and y E X, we have
E = {x E X I d(x, E) = o}.
Definition 1.2.8 Let {X,d} be a metric space. A sequence {x n } in X
is said to be Cauchy if, for every E > 0, there exists a positive integer
N such that
for every k 2: N, l 2: N . •
From the above definition it follows that a subbase for the weak
topology generated by the li is the collection of all sets of the form
fi-l(U) where U is an arbitrary open set in Xi and the index i ranges
over the indexing set I. A typical neighbourhood of a point x E X will
therefore be a finite intersection of sets of the form f i- 1 (Ui) where Ui is
a neighbourhood of Ji(x) in Xi.
Thus, sets of the form IIi EI Ui , where Ui = Xi for all i -I- io (an
arbitrary element of I) and Uio is open in X io ' form a subbase for the
product topology. A base for the topology is the collection of all sets
of the form IIiEIUi where Ui = Xi for all but a finite number of indices
and, for those indices, Ui is an open set in Xi.
X C 7~ { B(Xi; E) .•
Proposition 1.2.6 Let {X, d} be a metric space. The following state-
ments are equivalent:
(i) X is compact.
(ii) X is sequentially compact.
(iii) Every infinite subset of X has a limit point.
(iv) X is complete and totally bounded. •
In this section, we will recall basic facts and results of Lebesgue's theory
of measure and integration.
p(U~l i) = :LJ-l(Ai ).
i=l
The triple (X, S, J-l) is called a measure space. •
Definition 1.3.4 Let (X,S,J-l) be a measure space. If p(X) < 00, then
J-l is said to be a finite measure. If X can be covered by a countable
union of measurable sets, each with finite measure, then J-l is said to be
a a-finite measure . •
Example 1.3.1 Let X be any non-empty set and let S be the collection
of all subsets of X, which is obviously a a-algebra. If A c X, define
J-l(A) to be the number of elements in A, if A is a finite set and to be 00,
otherwise. This defines a measure and is called the counting measure
onX . •
if Xo E A
J-l(A) = { 1,
0, otherwise.
This defines a measure on X which is called the Dirac measure con-
centrated at the point Xo. •
The most important example of a measure is the Lebesgue mea-
sure defined on lRN.
Consider the real line lR endowed with its usual topology. Then, it is
possible to construct a a-algebra M on lR which contains all the Borel
sets and to define a measure p such that for any interval of the form
[a, b) (or [a, b], (a, b], (a, b)), where a, bE lR, we have
a + E = {a + x Ix E E} E M
20 1 Preliminaries
XE
( ) _{I,
X - 0,
if x E E,
if x (j. E.
s(x) = I>l:iXAi(X),
i=l
and
lim sn(x) f(x)
n-->oo
for all x E X . •
in place of Jx f dJL.
Example 1.3.4 If X is any non-empty set and if <5xo is the Dirac measure
concentrated at Xo EX, then
f(xo)
We now list some of the main results from the theory of the Lebesgue
integral, which behaves very nicely with respect to limit processes.
Theorem 1.3.1 (Monotone Convergence Theorem) Let (X, S, JL)
be a measure space. let {fn} be an increasing sequence of non-negative
measurable functions on X. Let f n (x) ---+ f (x) for all x EX. Then f is
measurable and
lim { fn dJL ---+ { f dJL . • (1.3.1)
n->oojx jx
Theorem 1.3.2 (Fatou's Lemma) Let (X, S, JL) be a measure space.
Let {fn} be a sequence of non-negative measurable functions on X. Then
lim
n----+oo ixr Ifn - fl dfJ- = O.
Qx = {y E Y I (x, y) E Q}.
QY = {x E X I (x, y) E Q}.
Theorem 1.3.4 Let (X, S, 1-") and (Y, T, A) be a-finite measure spaces.
Let Q E S x T. For x E X and y E Y, define
We can now use the above theorem to define the measure on the
product of two a-finite measure spaces.
Definition 1.3.8 Let (X, S, 1-") and (Y, T, A) be two a -finite measure
spaces. The product measure I-" x A on X x Y is defined by
for every Q E S x T. •
( 1.3.3)
verifies
i'P dl-" < 00,
Remark 1.3.2 The relation (1.3.4) may be written in the form of iter-
ated integrals as follows:
and the common value is equal to IxxY f d(J-L x ,X) whenever Fubini's
theorem is applicable. •
For further details, the reader is referred to Halmos [2], Royden [5]
or Rudin [7J.
Chapter 2
are continuous, each product space being endowed with the appropriate
product topology using the given topology of V and the usual topology on
the scalar field IF (= lR or C) . •
x- z = (x - y) + (y - z).
(Thus, the norm of a vector is its distance from the origin and is a
generalization of the notion of the length of a vector, as we know it, in
Euclidean space.)
Thus, V is endowed with a metric topology. In this topology, a
sequence {x n } converges to x in V if, and only if
2.2 Examples
Example 2.2.2 Let 1:::; p < 00. For x = (Xl,X2, ... ,XN) E lR N , define
We now proceed to prove the triangle inequality for each of the norms
11.llp for1 :::; p < 00.
Definition 2.2.1 Let 1 :::; p :::; 00. If p = 1, set p* = 00 and vice-versa.
Otherwise, let 1 < p* < 00 be such that
1 1
- +- = 1 (2.2.1)
p p*
The number p* defined thus is called the conjugate exponent of p . •
Lemma 2.2.1 Let 1 < p < 00. Let p* be its conjugate exponent. Then,
if a and b are non-negative real numbers, we have
(2.2.2)
2.2 Examples 29
f (t) = k (t - 1) - t k +1
for some k E (0,1). Then f'(t) = k(l - t k - 1 ) 2:: 0 since k < 1. Thus, f
is an increasing function on [1, (0) and, since f(l) = 0, we immediately
deduce that
t k :s; k( t - 1) + 1 (2.2.3)
for t 2:: 1 and 0 < k < 1.
Now, if a or b is zero, then (2.2.2) is obviously true. So let us assume
that a 2:: b > O.
The inequality (2.2.2) now follows by setting t = alb and k = lip in
(2.2.3) and using the definition of p* . •
Lemma 2.2.2 (HOlder's inequality) Let 1 < p < 00. Let p* be its
conjugate exponent. Then, for x, Y E ]RN,
N
L IXiYil :s; IlxllpIIYllp·· (2.2.4)
i=l
It is easy to verify that this also defines a norm on ]RN. Again con-
vergence and the Cauchy criterion hold if and only if they hold compo-
nentwise and so ]RN is a Banach space for this norm as well. Again all
these assertions hold for eN as well. It is immediate to see that Holder's
inequality is true when p = 1 as well. •
2.2 Examples 31
Remark 2.2.3 The spaces ffi.N (or C N when the base field is C) with the
norm 11.llp, where 1 ~ p ~ 00 are usually denoted by C!; in the literature .
•
Remark 2.2.4 The notation 11.1100 for the norm defined in Example 2.2.3
can be 'justified' as follows. Let x E ffi.N. Assume that the maximum
for IXil is attained for a single index, say, io. Then,
when p -+ 00 . •
We also define
We will prove the triangle inequality for 1I.llp (which will also simulta-
neously show that Cp is closed under vector addition and hence that it
is a vector space). Since properties (i) and (ii) of Definition 2.1.2 are
obvious, it will follow that II_lip defines a norm on Cp .
32 2 N ormed Linear Spaces
r
Let x and Y E Pp . Then, for any positive integer N, we have
L IXi + Yil
00
which shows that x + Y E Pp and also proves the triangle inequality for
1I·llp· •
Remark 2.2.5 The triangle inequality
L Ix~m) -
00
L IXi IP
00
Passing to the limit as l ----+ 00, we get that for any m ~ N and for any
k,
k
L Ix~m) - xil P :S E.
i=l
i.e. the space of all bounded real (or complex) sequences. This is clearly
a vector space under componentwise addition and scalar multiplication.
Define
Ilxll oo = sup IXil.
l::;i<oo
34 2 JVorr,ned Linear Spaces
Example 2.2.6 Let e[O, 1] denote the set of all continuous real valued
functions on the closed interval [0, 1]. This becomes a vector space under
the operations of addition and scalar multiplications defined pointwise,
i.e.
(f + g)(x) = f(x) + g(x) and (af)(x) = af(x)
for f and 9 E e[O, 1], a E ffi. and for x E [0,1]. Define
This is well defined since [0,1] is compact and so every continuous func-
tion is bounded and attains its maximum. The verification that this
defines a norm on e[O, 1] is routine and is left to the reader.
Let Un} be a Cauchy sequence in e[O, 1]. This implies that for every
c > 0, there exists a positive integer N such that, for all x E [0,1] and
for all n ~ Nand m ~ N, we have
(2.2.7)
Thus the pointwise sequences Un(x)} are all Cauchy and hence conver-
gent. Define
f(x) = lim fn(x).
n->oo
'Ne will show that the function f thus defined is in e[o, 1] and that
Ilfn - fll ~ 0. This will show that C[0,1], with the given norm, is a
Banach space.
2.2 Examples 35
°
for all n 2: IV and for all x E [0,1]. Fix a point Xo E [0,1]. Since fN is
continuous, there exists 8 > such that for all Ixo - yl < 8, we have
Thus, if Ixo - yl < 8, we get, using the above inequality and also the
inequality (2.2.8),
This proves that f is continuous and from (2.2.8), we see that Ilfn - fll ----+
0. •
If x ""' x' and y ""' y', then, clearly, x + y ""' x' + y' and ax ""' ax', since
W is a linear subspace of V. Thus, addition and scalar multiplication
are well defined. Thus the quotient space becomes a vector space. On
this, we define
Ilx + Wllvjw = inf Ilx + wll·
wEW
In other words, the 'norm' defined above is the infimum of the norms of
all the elements in the coset and so, clearly, it is well defined.
36 2 N ormed Linear Spaces
Now choose Yk E x nk + W such that IIYk - Yk+lll < 1/2k. Then the
sequence {yd is Cauchy (why?) and so, since V is complete, Yk -+ Y in
V. Thus
II (xnk + W) + W)llvjw
- (y :::; IIYk - yll -+ O.
Thus, the Cauchy sequence {x n + W}
has a convergent subsequence
{xnk + W} and so the Cauchy sequence itself must be convergent and
converge to the same limit (why?). Hence V/W is complete. •
where 11.11 V and 11·11 w denote the respective norms in the spaces V and
W.
(iv) If B = {x E V I IIxliv :S I} is the (closed) unit ball in V, then T(B)
is a bounded set in W.
2
IIT(x)lIv :S Jllxllv
(iii) {:} (iv) By virtue of (2.3.1), it follows that IIT(x)llw ~ K for all
x E B Thus, T(B) is bounded in W. Conversely, if T(B) is bounded in
W, there exists a K > 0 such that IIT(x)llw ~ K for all x E B. Now, if
0-# x E V is arbitrary, set y = x/llxli. Then IIT(y)llw ~ K from which
(2.3.1) follows, by linearity. •
IITII = 0: = f3 = , .•
for all x E V. •
II00TII = 100111TII·
The zero element of £(V, W) is the trivial map which maps every element
of V into the null vector of W. The element - T is defined by ( - T) (x) =
-T(x). Thus, £(V, W) is a vector space; in fact, it is a normed linear
space for the norm of a continuous linear transformation defined above.
Proof: Let {Tn} be a Cauchy sequence in .c(V, W). Then, given c > 0,
we can find a positive integer N such that, for all m and n ;::: N , we
have
Let x E V. Then,
Clearly, the map x t-t T(x) is linear. We will show that it is continuous
and that II Tn - Til ~ O. This will complete the proof.
Since the sequence {Tn} is Cauchy, it is bounded, i.e. there exists
M > 0 such that, for all positive integers n, we have IITnl1 ::; M. Now,
since for any x E V, we have IITn(x)lIw ::; IITnllllxliv ::; Mllxllv, it
follows, on passing to the limit as n ~ 00, that, for all x E V,
for all x E B. This shows that, for n ;::: N, we have IITn - Til::; c, i.e.
Tn ~ T in .c(V, W). This completes the proof. •
In particular, since the scalar field is a Banach space over itself (cf.
Example 2.2.1), the set of all continuous linear functionals .c(V,JR) (or,
.c(V, C), as the case may be) is always a Banach space.
Definition 2.3.4 Let V be a normed linear space. The space of all
continuous linear functionals on V is a Banach space and is called the
dual space of V. It is denoted by V* . •
2.3 Continuous Linear Transformations 41
(2.3.4)
11111 = 1.
standard basis of]RN (cf. Example 1.1.2). If x = (Xl, ... , XN) E ]RN,
then X = I:[:l Xiei· Then T(x) = I:[:l xiT(ei). Thus,
N
IIT(x)IIIRM < KL IXil Kllxlh
i=l
where
Example 2.3.2 Let al, ... ,aN be scalars. For X = (XI, ... ,XN) E ]RN,
define
N
f(x) = L aixi·
i=l
( Xl X2 Xi )
T(x) = 1'2'''''i'''' .
Then, since
Li 2 1Yil 2 < 00 . •
i=l
L XiYi·
00
fy(x)
i=l
2.3 Continuous Linear Transformations 43
We will see later that, when 1 :::; P < 00, all continuous linear functionals
on £p occur only in this way and that the last inequality is, in fact, an
equality. •
Example 2.3.5 Consider an infinite matrix (aij )0=1. This can be used
to define a linear mapping on £p as follows. Let x = (Xi) E £p. Define a
sequence A(x) by
00
A(X)i = L aijXj.
j=1
Showing that A defines a continuous linear map of £p into itself is usually
a non-trivial problem and some examples are given in the exercises at
the end of this chapter. We now give an example (due to Schur).
Assume that aij ~ 0 for all i and j. Assume further there exists a
sequence {Pi} of positive real numbers and {3 > 0 and, > 0 such that
00
IA( ) 12 aiil x il 2
",,00
L..ti=l X i < ",,00
L.Ji=1 '"'(P2 L.Jj=1
_",,00
Pi
",,00 Ix _1 2 ",,00
= r L...Jj==l ~ L.-i=l aijPi
< ",,00 Ix _12
'"'( L.Jj=1 ~f3Pj
= f3'"'( Ilx II§
which establishes the claim.
An interesting particular case is that of the Hilbert matrix. Set
1
a-- -
2J - i +j +1
roo dx
< Jo (x+J+!)y'X
= 2 roo dt
Jo t 2 +J+!
P
IITII ~ p-l
Indeed,
2.3 Continuous Linear Transformations 45
( 1 _ 2:!E..)
p-1
~
n
+ (n-1)p p-1
p-1 an an-1
~) ~n + (n
( 1 p-1 p-11)p(~)7(
n .n-1 )~.
( 1 - 2:!E..)
p-1 ~n + (n-1)p
p-1 (E.=
p ~n +!d?.
p n-1 )
N N
~ ~ :s; p~l~ ~ llxnl :s; P~l ~ ~
(N )7(N~lxnlP )~
Dividing both sides by (2:;;=1 c1n)«p-1)/p), which is strictly positive for
non-zero x, we get
which establishes our claim. (In fact, Hardy, Littlewood and Polya also
show that IITII = p/(p - 1).) •
46 2 o~ Linear Spaces
Further, given E > 0, there exists 8> a such that, whenever lSI -S21 < 8,
we have
IK(sl' t) - K(S2' t)1 < E
whenever lSI - 821 < 8. This shows that T(f) is a continuous function.
The mapping T being clearly linear, it thus defines a linear operator on
e[o, 1]. Further,
IT(f)(S)1 ::; Kllfl18 ::; Kllfll·
Thus T is a continuous linear operator on e[O, 1] and IITII ::; K. •
Note: When dealing with normed linear spaces, the word isomorphism
is understood in the topological sense: not only is it an isomorphism in
the usual algberaic sense i. e. it is linear and is a bijection, but it also
implies that both the mapping and its inverse are continuous.
Definition 2.3.7 Two norms defined on the same vector space are said
to be equivalent if the topologies induced by these two norms coincide .
•
Proposition 2.3.4 Let V be a vector space and let 11.11(1) and 11.11(2) be
two norms defined on it. The two norms are equivalent if, and only if,
there exist two constants Gl > a and C2 > a such that, for all x E V,
we have
Proof: The topologies induced by the two norms coincide if, and only
if, the identity mapping
Example 2.3.9 Let x = (Xl, ... , XN) E]RN (or eN). Then, clearly,
matter to check that the topology induced by 11.1100 is none other than
the product topology on ]RN (respectively CP) when]R (respectively q
is given its usual topology for every component . •
Step 1. Let {el, ... , eN} be the standard basis of .ef. Fix a basis
{Vl, ... ,VN} for V. Define T:.ef ----t V by setting T(ei) = Vi for all
1 :S i :S N and then extending T linearly to all of .ef. Clearly T is a
bijection and an identical argument as in Example 2.3.1 shows that T
is continuous.
IIfl11 = 10 1
If(t)1 dt.
It is simple to check that this defines a norm on C[O, 1]. Consider the
sequence {fn} defined by
1- nx, °
for < - x <
for .1n < x
- .1
<
n
0, - - 1.
Clearly, Ilfnll = 1 for all n (where 11.11 denotes the usual 'sup' norm)
while Ilfnlh = ~ fn(t) dt = 1/2n which tends to zero as n tends to in-
finity. Thus it is clear that these two norms cannot be equivalent. Thus
in infinite dimensional spaces, two norms are not, in general, equivalent .
•
Since all norms on ~N (respectively, eN) generate the same topology,
it is now clear that any matrix generates a continuous linear transforma-
tion, whatever the norm on that space may be. Thus, if T is an N x N
matrix and if 11.11 is a norm on ~N (respectively, eN), we can define
ITII
I sup IITxll (2.3.5)
= iixii::;1, #0 Ilxll
or, via any of the other equivalent formulations as in Proposition 2.3.2.
Since the unit ball (and hence the unit sphere) is compact, the 'sup'
50 2 lVorTined Linear Spaces
above is, in fact, a 'max'. If T and S are matrices of order lV, then TS
represents the composition of the corresponding linear transformations
and so we also have
IITSII ::; IITII·IISII· (2.3.6)
Let MN denote the set of all matrices of order lV with entries from the
corresponding field. This itself (under the operations of matrix addition
and scalar mUltiplication) is a vector space (of dimension N 2 ). Any
norm on this space which satisfies (2.3.6) is called a matrix norm. If
such a norm were induced by a vector norm on ]RN (respectively, (CN)
via (2.3.5), then we always have
IIIII 1
N N
II II~ ::; L Itikl2 L ISkjl2 = II II~II II~
i,k=l k,j=l
which shows that II· liE is indeed a matrix norm. This is also known as
the Hilbert-Schmidt nOrTin.
2.3 Continuous Linear Transformations 51
However, notice that this norm is not induced by any vector norm
when N ~ 2. Indeed to see this, observe that lilliE = VN # 1. •
8
8 :S IIv-wll :S -1-·
-E
B = {x E V IlIxll ~ I},
is compact.
52 2 N armed Linear Spaces
(2.3.7)
where B(Xi' 1/2) = {x E V Illx-xili < 1/2} is the open ball centered at
Xi and of radius 1/2. Set W = span{xl' ... , x n }. We claim that W = V
and this will prove that the dimension of V is less than, or equal to,
n and so V has to be finite dimensional. Assume the contrary. Then
W will be a proper and closed (since it is finite dimensional) subspace
of V. Now, by the preceding lemma of Riesz, we have the existence of
u E V such that Ilull = 1 (and so u E B) and such that d(u, W) ~ 2/3.
In particular, it follows that u E B is such that Ilu - xiii ~ 2/3 for all
1 ::; i ::; n which contradicts (2.3.7). This completes the proof. •
converges to x* . •
which can be made arbitrarily small for large nand m since the geometric
series 2:%"=1 ck is convergent for 0 < c < 1. Thus {x n } is a Cauchy
sequence and, since X is complete, it converges to some x* EX. The
continuity of F and the definition of the Xn now imply that x* = F(x*).
If there were two distinct fixed points of F, say, x and y, then, we
have that
o < d(x, y) = d(F(x), F(y)) :s: cd(x, y)
which is a contradiction, since 0 < c < 1. This completes the proof. •
~; = f(x, y)
y(xo) = Yo
has a unique solution in the interval (xo - h, xo + h).
Proof: Since R is compact and f and U are continuous on R, there
exist K > 0 and M > 0 such that
for all (x, y) E R. Then, by the mean value theorem, it follows that, for
all (x, yd and (x, Y2) in R, we have
F(g)(x) = Yo + r
Jxo
f(t, g(t)) dt
and thus, F(x*) is also a fixed point for Fn. By the uniqueness of the
fixed point, it follows that
F(x*) = x*
and so F has a fixed point, viz. x*.
On the other hand, any fixed point y of F is also a fixed point of F n
since
F(y) y.
where
~ max IK(s, t)l·
[O,1]x[O,1]
Hence, we get
Again,
where 0: and (3 are two given real numbers. Set u(s) = x"(s). Then,
x'(s) = (3 + 18
u(t) dt.
Again,
x(s) 0: + J; x'(t) dt
0: + (3s + J; ~ U(T) dT dt
0: + (3s + J; J: U(T) dt dT
0: + (3s + J; U(T)(S - T) dT.
2.5 Exercises
Note: In all the function spaces which occur below, it is assumed that
vector addition and scalar multiplication are defined pointwise.
1 1.
Ilfllp (folf(t)IP dt) P
58 2 N ormed Linear Spaces
where 1 ::; p < 00. Show that this defines a norm on C[O, 1J.
2.2 Show that the space C[O, 1J with the norm 11.111 defined in the previ-
ous exercise is not complete by producing a Cauchy sequence which is
not convergent.
2.3 Let f E C(JR). The support of f is the closure of the set of points
where f does not vanish. Let Cc(JR) denote the space of all continuous
real valued functions on JR whose support is a compact subset of JR. Show
that it is a normed linear space with the 'sup-norm' and that it is not
complete.
2.4 Let Ca(JR) denote the space of all continuous real valued functions
on JR which vanish at infinity, i.e if f E Ca(JR), then, given any c > 0,
there exists a compact subset K c JR such that
If(x)1 < c
for all x E JR\K. Show that Ca(JR) is a Banach space with the 'sup-
norm'. Show also that the space Cc(JR) defined in the previous exercise
is dense in Ca(JR).
2.5 Let Cl[O, 1J denote the space of all continuous real valued functions
on [0, 1J which are continuously differentiable on (0,1) and whose deriva-
tives can be continuously extended to [0, 1J. For fECI [0, 1]'define
Ilfll = m[ax]{lf(t)I,I!'(t)l}
tE a,l
2.6 Let fECI [0, 1J and let f' denote its derivative. Define
(1
1
1
Ilfllt = (If(t)1 2 + 1!'(tW)dt) 2
Show that 11.111 defines a norm on Cl[O, 1J. If we define
2.5 Exercises 59
2.7 Let
v = {J E Cl[O, 1]1 f(O) = a}.
Show that 1.11 defines a norm on V.
x = C([O, 1]; V)
to be the space of all continuous functions from [0,1] into the space V.
Define, for f EX,
Show that 1I.lIx is well defined and that it defines a norm on X. Show
also that, under this norm, X is a Banach space.
2.10 Let Cl[O, 1] be endowed with the norm as in Exercise 2.5 above. Let
qO,I] be endowed with the usual 'sup-norm'. Show that T : Cl[O, 1] ~
qO,I] defined by T(f) = l' is a continuous linear transformation and
that IITII = 1.
2.11 Let C[O, 1] be endowed with its usual norm. For f E qo, 1], define
1:
2.12 Consider the space CcOR) defined in Exercise 2.3 above. For E
Cc(lR), define
<p(f) = f(t) dt.
60 2 N ormed Linear Spaces
Show that <p is well defined and that it is a linear functional on this
space. Is it continuous?
2.13 Let {tdi=l be given points in the closed interval [0,1]. Let {wi}i=l
be given real numbers. Let f E C[O, 1]. Define
n
<p(f) = Lwd(td.
i=l
Show that <p defines a continuous linear functional on qo, 1] and that
n
11<p11 = L IWil·
i=l
2.14 Let Mn denote the set of all n x n matrices with complex entries.
Let 11.llp,n denote the matrix norm induced by the vector norm 11.llp on
en, for 1 ::; p::; 00. If A = (aij) E M n , show that
IIAI12,n = y'p(A*A)
where p(T) denotes the spectral radius of a matrix T. (Hint: Use Propo-
sition 1.1.8). If A is a normal matrix, show that IIAI12,n = p(A).
2.16 With the notations introduced above, show that, for any matrix
A E M n , we have
Show that
(a) cond(A) 2: 1 for any invertible matrix A E Mn;
(b) cond(o:A) = cond(A) for any invertible matrix A and for any scalar
0: # 0;
(c) for any invertible and normal matrix A,
2.19 For what class of matrices does cond2,n attain its minimum value?
where
2::;,j=1 laij 12
(Y =
2Idet(A)1
2.21 Let Mn be endowed with the topology generated by any matrix
norm. Let GLn(CC) denote the set of all invertible matrices in Mn. Show
that GLn(tC) is an open and dense set. Is it connected?
2.22 (a) Let Vn be the subset of all n x n matrices with distinct eigen-
values. Show that Vn is dense in Mn (endowed with any matrix norm).
(b) Prove the Cayley-Hamilton theorem for any diagonalizable matrix:'Every
n x n matrix satisfies its characteristic equation'.
(c) Deduce the Cayley-Hamilton theorem for all n x n matrices.
. 1
mf II A - B 112 n = 1 .
B is singular ' IIA - 112,n
62 2 lVorr,ned Linear Spaces
2.24 Show that the set of all orthogonal matrices in the space of all
n x n real matrices (endowed with any norm topology) is compact.
2.25 Let 1 ::; p < q ::; 00. Show that lp C lq and that, for all x E lp,
LaijXj.
j=l
ao 0 0
a1 ao 0
a2 a1 ao
a3 a2 a1
2.5 Exercises 63
Let A be the linear map defined on £2 by this matrix (as in the preceding
exercise). Show that A E £(£2) and that
00
IIAII ~ L lakl·
k=O
is convergent and the limit of the sequence {Sn} is called the sum of the
series. If 2:~1 IIAkl1 < 00, we say that the series 2:~1 Ak is absolutely
convergent. Show that any absolutely convergent series in convergent.
2.29 Let V be a Banach space. If A E £(V) is such that IIAII < 1, show
that the series
00
1+ LAk
k=1
2.30 (a) Let V be a Banach space and let A E £(V). Show that the
series
00 Ak
I+Lkf
k=1
where 0: and ware real numbers. Show that, for any t E JR,
exp (t A) -_ e
ad [coswt
. - sinwt ] .
smwt coswt
2.31 Let V be a Banach space. Show that g, the set of invertible linear
operators in £(V) is an open subset of £(V) (endowed with its usual
norm topology).
where x = (X1,X2, ... ) E £2. Show that T and S define continuous linear
operators on £2 and that ST = I while TS =1= I. (Thus, T and S, which
are called the right and left shift operators respectively, are not invert-
ible) .
(b) If A is a continuous linear operator on £2 such that IIA - Til < 1,
show that A is also not invertible. Deduce that, in general, g, defined
in Exercise 2.31 above, is not dense in £(V), if V is infinite dimensional.
(Compare this with the finite dimensional case, cf. Exercise 2.21).
2.33 Let P denote the space of all polynomials in one variable with
real coefficients. Let pEP and let p = 2:~1 aixi, where ai E JR for
1 :S i :S n. Define
n
Show that 11.111 and 11.11 00 define norms on P and that they are not equiv-
alent.
2.36 Let V and Y be normed linear spaces and let W be a dense subspace
of V. Let A E £(W, Y). Show that there exists a unique continuous ex-
tension A E £(V, Y) (i.e. Alw = A) and that IIAllc(V,y) = IIAllc(w,Y)'
. IIJ(u
11m + h) - J(u) - T(h)11
O.
h-+O Ilhll
66 2 lVorTined Linear Spaces
(Equivalently,
exists. (We call the limit the Gateau derivative of J at u along w.)
If J is Fhkhet differentiable at u E U, show that it is Gateau differ-
entiable at u along any vector w E V and that the corresponding Gateau
derivative is given by J'(u)w.
2.40 Let V and W be normed linear spaces and let A E £(V, W). Let
Wo E W be given. define J : V ---t W by J(u) = A(u) + wo0 Show that
J is differentiable at every u E V and that J' (u) = A.
2.43 (Chain Rule) Let V, Wand Z be normed linear spaces and let
j : V ~ Wand 9 : W ~ Z be mappings such that f is differentiable at
a point v E V and 9 is differentiable at j (v) = w E W. Show that the
map 9 0 j : V ~ Z is differentiable at v E V and that
tu + (1 - t)v E K.
J'(u)(v - u) ~ a
for every v E K.
(b) Let K = V. If J attains its minimum at u E V and if J is differen-
tiable at u, show that J' (u) = o.
J'(u)(v - u) > 0
J(u) = minJ(v).
vEK
of V) at u.
Ax = b.
This system may not have a solution since the number of equations
exceeds the number of unknowns. A least squares approximate solution
is a vector Xo E ~n such that
A*Axo = A*b
Hahn-Banach Theorems
g(x) ~ p(x)
f(x) ~ p(x)
for all x E V.
Proof: Step 1. Let P denote the collection of all pairs (Y, h), where Y
is a subspace of V containing Wand h : Y ~ lR a linear map which is
an extension of 9 and which is also such that
h(x) ~ p(x)
70 3 Hahn-Banach Theorems
or, equivalently,
f(x) + 0: ~ p(x + xo) (3.1.2)
for all x E Z. Similarly, by considering t < 0, we deduce that
f(x) - 0: ~ p(x - xo) (3.1.3)
f(x) + f(y) = f(x + y) :s: p(x + y) < p(x + xo) + p(y - xo)
or, equivalently,
Ilfllv* = Ilgllw*.
Proof: Set p(x) = Ilgllw*llxll. Then p verifies (3.1.1). Also g(x) :s: p(x)
for all x E W. Thus, there exists a linear extension of g, viz. f : V ~ lR
such that, for all x E V, f(x) :s: Ilgllw*lIxll. This implies that f is con-
tinuous and that Ilfllv* :s: Ilgllw*. But, since f = 9 on W, it follows that
we do have equality of the norms of f and g. This completes the proof. •
We will now prove the same result for complex vector spaces.
which shows that h(x) = -g(ix). Now, let f(x) = eit9 lf(x)1 where
() E [0, 27T). Then,
since the left extreme of the above relation is real. Thus If (x) I :S: II 9 II II x I
which implies that Ilfll :S: Ilgll. On the other hand,
Ilfllv* = Ilgllw*.
Proof: Let 9 = h(x) - ih(ix) where h is the real part of g. We consider
Vasa real normed linear space by st i ti~ ourselves to scalar multi-
plication by reals only. Then, !here exists h : V -+ lR. which is a linear
extension of h and such that Ilhllv* = Ilhllw*. Now set
for all x E V. Then, clearly, f (x + y) = f (x) + f (y) and, for real scalars
a, f(ax) = af(x). Now,
Proof: Clearly, If(x)1 :S Ilfll Ilxll :S Ilxll when Ilfll :S 1. On the other
hand, by the preceding corollary, there exists f E V* such that Ilfll = 1
and f(x) = IIxll when x is non-zero. Thus the result is established for
non-zero vectors and is trivially true for the null vector. •
Jx(f) = f(x)
Since V**, being a dual space, is always complete, the notion of re-
flexivity makes sense only for Banach spaces. By applying Corollary
3.1.2 to V*, it is readily seen that the supremum in (3.1.6) is attained
for reflexive Banach spaces. A deep result due to R. C. James is that
the converse is also true: if V is a Banach space such that the supremum
is attained in (3.1.6) for all f E V*, then V is reflexive. We will study
reflexive spaces in greater detail in Chapter 5.
74 3 Hahn-Banach Theorems
Example 3.1.1 Let 1 < p < 00. Let p* be the conjugate exponent of
p (cf. Definition 2.2.1). Let Y E Cp •. We already saw that (cf. Example
2.3.4) the linear functional fy defined on Cp by
2:
00
fy(x) = X iYi
i=l
where ei is the sequence whose i-th entry is unity and all other entries
are zero. Set f = (h).
Let n be any positive integer. Define
Since n was arbitrary, we deduce that f E Cp • and that Ilfllp· ::; Ilfll·
For any X = (Xi) E Cp , we have 2:~1 Xiei ----+ X in Cp and so, by the
continuity of f, it follows that
2:
00
f(x) X di
i=l
3.1 A nalytic Versions 75
It is easy to see that using these identifications of the dual spaces, the
canonical isometry from £p into £;*
= £p is nothing but the identity map,
which is onto. Thus, the spaces £P' for 1 < p < 00 are all reflexive . •
Remark 3.1.2 Though in the above example, we have not dealt with
the case of real and complex sequence spaces separately, it is customary
to identify the dual of the complex sequence space £p with £p' via the
following relation: if y E £p" then we define fy E by £;
00
fy(x) = L XiYi·
i=l
for all x = (Xi) E £1, where fi = f(ej). Further, setting f = (Ii), we have
Ilfll oo = Ilfll·
76 3 Hahn-Banach Theorems
for all x = (xd E £00 is continuous and Ilfyll = Ilylh. But there exist
continuous linear functionals on £00 which do not arise this way. So the
identity mapping of £1 which is still the canonical imbedding of £1 into
£i*, is not surjective. Thus the space £1 is not reflexive.
To see this, let G be the subspace of all convergent sequences in £00.
For x = (Xi) E G, define
g(x) = .lim Xi.
2--"00
Then 9 : G -+ lR. is linear and Ig(x)1 ~ Ilxll oo . Thus, 9 is continuous as
well and so, by the Hahn-Banach theorem, can be extended to a con-
tinuous linear functional f on £00' preserving the norm. We claim that
this continuous linear functional cannot be obtained from an element of
£1 by the above outlined procedure.
Assume the contrary and let Y = (Yi) E £1 be such that f = fy.
Consider the sequence {x(n)} in £00 given by
x(n) = {O, 0, ... , 0,1,1,1, ... }
where the 1's start from the n-th entry. Then Ilx(n) 1100 = 1 and x(n) E G.
We have 00
i=n
which is impossible since Y E £1 implies that
00
i=n
as n -+ 00 . •
H = {x E V I f(x) = a},
denoted by [f = a], where f is a non-zero linear functional on V . •
f(xo) - a
t
f(xo) - f(XI)·
Then 0< t < 1 and, if Xt = tXI +(l-t)xo, then f(xt) = a. But B(xo; 2r)
is a convex set and so Xt E B(xo; 2r), which is a contradiction.)
Thus, for any z E V such that Ilzll ::; 1, we get f(xo + rz) ::; a or,
equivalently,
f(z) ::; a - f(xo) .
r
Thus the image of the unit ball is bounded and so f is continuous. •
p(x) °
= inf{a > I a-Ix E C}.
78 3 Hahn-Banach Theorems
Proof: Since 0 E C and C is open, there exists an open ball B(O; 2r),
centered at 0 and of radius 2r, contained in C. Now, if x E V, we have
rx/llxli E C and so, by definition, p(x) S ~ Ilxll which proves (3.2.1).
Let x E C. Since C is open, and since 0 E C, there exists c > 0 such
that (1 + c)x E C. Thus, p(x) S (1 + c)-l < 1. Conversely, let x E V
such that p(x) < 1. Then, there exists 0 < t < 1 such that tx
E C.
Then, as C is convex, we also have ttx + (1 - t)O E C, i.e. x E C. This
proves (3.2.2).
If 0: > 0, it is easy to see that p( o:x) = o:p( x). This is the first
relation in (3.1.1). Now, let x and y E V. Let c > O. Then
1 1
--,--,---x E C and () Y E C.
p(x) + c P Y +c
Set
t = p(x) + c
p(x) + p(y) + 2c
so that 0 < t < 1. Then, as C is convex,
1 1 1
t p (x)+c x +(I-t)p(y)+c Y = p(x)+p(y)+2c(x+ Y )EC
f(x) 5: 0: 5: f(y)
C = UyEB(A-{y}),
we see immediately that C is both open and convex. Since A and Bare
disjoint, it also follows that 0 ¢ C. Hence, by the preceding proposition,
80 3 Hahn-Banach Theorems
there exists f E V* such that f(z) < 0 for all z E C = A-B. In other
words, f(x) < f(y) for all x E A and y E B. Choose 0: E lR such that
Proof: Let TJ > O. Then, if B(O; TJ) is the open ball of radius TJ > 0
centered at 0, then A+B(O; TJ) and B+B(O; TJ) are non-empty, open and
convex. Further, if TJ > 0 is sufficiently small, the two sets are disjoint
as well. If not, there exists a sequence TJn ---t 0 and Xn E A, Yn E B such
that IIx n - Ynll ~ 2TJn . Since B is compact, there exists a subsequence
Ynk which converges to Y E B. This implies then that x nk ---t Y and,
since A is closed, YEA, i. e. yEA n B, which is a contradiction.
Thus, we can choose TJ > 0 such that A + B(O; TJ) and B + B(O; TJ)
are disjoint. Then, by the preceding theorem, there exists f E V* and
0: E lR such that for all x E A, y E Band Zl and Z2 in the closed unit
ball, we have
where ~i E [Xi-I, Xi] for 1 ::; i ::; n, and define the integral (if it exists)
as a suitable limit of such sums. Assume that such a limit exists and
denote it by y E V. Let f E V*. Then, by the continuity and linearity
of f, it will follow that f (y) will be the limit of the Riemann sums of
the form
n
Z)Xi - xi )f( p(~i)).
i=I
But since f 0 <p : [0,1] --? lR. is continuous, the above limit of Riemann
sums is none other than
(3.3.1)
Definition 3.3.1 Let V be a real normed linear space and let <p : [0,1] --?
Proof: Since [0,1] is compact, the set H which is the closure (in V) of
the set H which is the convex hull of r.p([0, 1]) (i.e. the smallest convex
set containing r.p([0, 1])), is compact.
Let L be an arbitrary finite collection of continuous linear functionals
on V. Define
EL = {Y E HI f(y) = 10 1
f(r.p(t)) dt for all f E L}.
It is immediate to see that EL is a closed set.
where
mi = 10 1
Ji(r.p(t)) dt.
it follows from the previous step that the class of closed sets
for every f E V*. Thus y = fol 'P(t) dt. This completes the proof. •
Proposition 3.3.2 Let V be a real normed linear space and let 'P
[0,1] --+ V be continuous. Then
(3.3.2)
Thus,
lb 'P(t) dt = (b - a) 10 1
'lj;(t) dt.
3.4 An Application to Optimization Theory 85
Again the result of Proposition 3.3.2 remains valid in this case as well.
Assume now that cp : [0, (Xl) ----t V is continuous. Assume further that
the limit
ior <p(t) dt
A
lim
A--+OO
exists, i. e., for any sequence An ----t (Xl (as n ----t (Xl), we have that the
limit
An
lim
n--+oo
lo0 cp(t) dt
roo cp(t) dt
io
= lim
A--+OO
r cp(t) dt
io
A
Lemma 3.4.1 Let Vi, 1:S; i :s; n be elements in a normed linear space
V. Define
Proof: Step 1. Clearly, C is convex since, for any 0 < t < 1, we have
n n n
and so, if Ai 2: 0 and f.,li 2: 0 for all i, we also have tAi + (1 - t)f.,li 2: o.
Also if x E C, it is obvious that AX E C for any A 2: O. Thus, C is a
convex cone.
w = span{v1,···, vn },
Step 3. Assume now that the Vi are linearly dependent. Then there
exists a linear relation between them and so we can find scalars CYi such
that L:~=1 CYiVi = 0 and such that the set
t = min {- Ai} 2: O.
tEJ CYi
Then, for all 1 ::; i ::; n, we have Ai + tCYi 2: 0 and at least one of them
must vanish. Then
where I = {1, 2,· .. ,n}. Each set in the union described on the right-
hand side is a cone but generated by fewer elements from V. Iterating
this procedure, we can ultimately write C as the finite union of cones
3.4 An Application to Optimization Theory 87
Proof: Let
Remark 3.4.2 It is clear that C(u) is a cone (cf. Definition 3.4.1) since
W E C(u) implies that AW E C(u) as well, for any A > 0 (with the
same associated sequence {Uk})' This is a cone (which is not necessarily
convex) with its vertex at the origin. Its translate
u+C(u) = {u+wiWEC(u)},
i5 n
+ IIWn II - Ilwll
( Wn W )
''In = ken) ---t 0
Of particular interest is the case when the functionals 'Pi are affine linear,
z. e. there exist fi E V* and vectors di E V for 1 ~ i ~ m such that
(3.4.2)
where
J(u) = {i 11 ~ i ~ m, 'Pi(U) = O}.
Proof: Notice that (cf. Exercise 2.40) P~( ) = Ii- If i E J(u), then
°
'Pi attains its maximum over U at u. Then, by Proposition 3.4.2,
P~( )( ) ~ for all W E C(u).
°
Conversely, if w =1= 0 satisfies fi (w) ~ for all i E J (u), set Uk =
u + EkW where {Ed is a sequence of positive reals converging to zero.
° °
then Uk =1= u and Uk ---+ u. If i ¢ J (u), then 'Pi ('U) < 0, and so, by
continuity, 'Pi(Uk) < for large enough k. If i E J(u), then 'P(u) =
and so
o (3.4.4)
Proof: By Propositions 3.4.2 and 3.4.3, we have that for all w such
that P~ (u)w :S 0, i E I (u), we have J' (u)w 2: O. Thus, by the Farkas-
Minkowski lemma, there exist Ai (u) 2: 0 for i E I (u) such that
J' (u ) = - L Ai (U ) P~ (u ) .
iEI(u)
Setting Ai(U) = 0 for all i ~ I(u), we get (3.4.4). This completes the
proof. •
The above theorem can be generalized to cases when the 'Pi are not
affine. In this situation, in addition to differentiability at u, we need to
assume another technical condition of 'admissibility' on the constraints
at u. In particular, when the constraints 'Pi, 1 :S i :S m are all convex,
the admissibility condition reads as follows:
• either, all the 'Pi are affine and the set U given by (3.4.1) ·is non-
empty;
• or, there exists an element v* E V such that 'Pi (v*) :S 0 for all
1 :S i :S m and 'Pi (v*) < 0 whenever 'Pi is not affine linear.
3.5 Exercises
Show that the spaces £1 and £00 are not strictly convex, while £2 IS
strictly convex.
3.4 Use Corollary 3.1.1 to prove Corollary 3.2.1. (Hint: consider the
quotient space V jW.)
V = WEBZ.
3.7 (a) Let CO denote the space of all real sequences which converge to
zero, equipped with the norm 11.1100. Prove that
(£~) = ff.
3.5 Exercises 93
d(j, ~) = Ilfl w Il w *
where fl w is the restriction of f to Wand
d(j, ~) = inf
gEW.l
Ilf - gllv*.
(Z ~)~ =
-
Z.
94 3 Hahn-Banach Theorems
3.10 Let r.p and 'ljJ be continuous mappings from [0,1] into a real normed
linear space V. For arbitrary scalars a and {3, show that
3.11 Let V and W be normed linear spaces and let A E £(V, W). Let
r.p: [0,1] -----> V be a continuous mapping. Show that
3.12 Let r.p : [a, b] -----> V be a continuous mapping and let c E (a, b).
lb l lb
Show that
r.p(t) dt =
c
r.p(t) dt + r.p(t) dt.
I a
b
r.p(t) dt = Ib+h 'ljJ(t) dt
a+h
3.14 Let r.p : [0,1] -----> V be a continuous mapping and let t E [0,1). Show
that
lim -h
~O
11t+h
t
r.p(s) ds = r.p(t).
3.15 Let V be a normed linear space and let T E £(V) such that IITII ::;
°
1. Let >. > and let x E V. Define
R(x) = 1 00
e-AtT(x) dt.
1
IIRII < ~.
Define
Ilullx = supe-ktllu(t)llv.
t20
Show that X is a Banach space with this norm.
(b) Let f : V --) V be a mapping. Assume that there exists L > 0 such
that
Ilf(u) - f(v)llv ::; Lllu - vllv
for all u and v E V. For u E X, define
where Uo E V is a fixed vector. Show that F(u) E X and that, for any
u and v EX, we have
L
IIF(u) - F(v)llx ::; "k 11u - vllx.
(c) Deduce that, for any Uo E V, there exists a unique u E C([O, (0); V)
such that
u(t) = Uo + fat f(u(s)) ds
3.17 Let V be any vector space and let fo, h, ... , f n be linear functionals
on V. Let Ker(fi) denote the kernel of h 0::; i ::; n. Assume that
Show that there exist scalars ai, 1::; i ::; n such that
N
fo = I:adi.
i=l
(Hint: Consider the image of the map A as defined in Section 3.3 and
apply Corollary 3.2.1 to its image.)
which is a non-empty closed set. Show that this collection of closed sets
has finite intersection property.)
Chapter 4
is also dense.
Proof: Let W be any non-empty open set in X. We need to show that
the intersection nn Vn has a point in W.
Since VI is dense, it follows that W n VI t- 0. Thus, we can find a
point Xl in this intersection (which is also an open set). Hence, there
exists rl > 0 such that the open ball B(XI; rd, with centre at Xl and
radius rl, is contained in W n VI. By shrinking r if necessary, we may
also assume that the closure of this ball B(XI; rd is also a subset of
W n Vi and that 0 < rl < l.
If n ?: 2, assume that we have chosen Xn-l and rn-l suitably. The
denseness of Vn shows that Vn n B (Xn-l; r n-l) I- 0 and so we can choose
Xn and rn such that 0 < rn < l/n and
Now, for i > n, it follows that Xi E B(xn; rn) and so x E B(xn; rn)
for all n and so x E Vn for all n. Also x E B(X1; r1) and so x E W as
well. This completes the proof. •
Then each Wn is also dense and open, since X has no isolated points.
But n~=l Wn = 0, which contradicts Baire's theorem. Hence the result .
•
4.2 Principle of Uniform Boundedness
Let
Vn = {xEVI'P(x»n}.
Since each Ti is continuous and since the norm is a continuous function,
it is easy to see that Vn is open for each n.
Assume now that there exists N such that VN fails to be dense in V.
Then, there exists Xo E V and r > 0 such that x + Xo rt VN if Ilxll < r.
(In other words, there is an open ball B(xo; r), centered at Xo and of
radi us r, which does not intersect V N .) This implies that 'P (x + xo) ::; N
for all such x and so, for all i E I,
4N
IITil1 ::; - r
and so the first alternative holds with M = 4N/r.
The other possibility is that each Vn is dense, and so, V being com-
plete, by Baire's theorem, nn Vn is a dense G8 and for each x E nn Vn ,
we have that 'P(x) = 00. This completes the proof. •
T(x) = n-->CXl
lim Tn(x).
for each x E V. •
Corollary 4.2.3 Let V be a Banach space and let B c V be a subset.
Assume that
f(B) = {f(x) I x E B}
is a bounded subset of the scalar field for each f E V*. Then B is a
bounded subset of V.
4.3 Application to Fourier Series 101
IIJxllv** = Ilxllv.
Taking B as the indexing set and V* as the Banach space, it follows
from the Banach-Steinhaus theorem that IIJxl1 is uniformly bounded in
V**, which is the same as saying that B is bounded in V. •
Let f : [-7f, 7f] ---t lR be an integrable function. We can write its formal
Fourier series (in exponential form) as follows:
2:
00
where
j(n) -1 j7f
f(s) exp( -ins) ds (4.3.1)
27f -7f
are the Fourier coefficients of f. The first question that springs to the
mind is "in what sense does the Fourier series of a function represent
the function?" In particular, does the Fourier series of a continuous 27f-
periodic function f converge to f (t) at every point t E [-7f, 7f l? This is
relevant since each term in the Fourier series is a continuous 27f-periodic
function. Unfortunately, the answer is 'No!'. It was Dirichlet who first
established (around 1829, nearly seven decades after a lengthy contro-
versy began in Europe - about the validity of representing a function
in terms of sines and cosines - and raged through the latter half of the
eighteenth century) the sufficient conditions for the Fourier series of a
function to converge to its value at a point. This was later strengthened
by Jordan. In fact the study of the validity of Fourier expansions led to
102 4 Baire's Theorem and Applications
Dm(t) = L exp(int).
n=-m
The function Dm is called the Dirichlet kernel. If we multiply Dm(t) suc-
cessively by exp(it/2) and exp( -it/2) and subtract, we see immediately
that
(4.3.3)
1 71"
-71" IDn(t)1 dt 2: 4 Jo
r Isin(nt+ ~)tl
dt
(n+~)71" Isin tl
= 4l - - dt
o t
k=l (k-l)7I"
Isin tl
--dt
k1r
8 1
-2:-
n
7r k
k=l
where sn(f) is the n-th partial sum of the Fourier series of f. Then cPn
is a continuous linear functional on V and
1 171"
IlcPnll = 27r -71" IDn(t)1 dt. (4.3.4)
and so
1 171"
IlcPnll ::; 27r -71" IDn(t)1 dt.
Now, let En = {t E [-7r,7rll Dn(t) ~ o} . Define
A+B {x + y I x E A, y E B}.
4.4 The Open Mapping and Closed Graph Theorems 105
Proof: Step 1. We will first show that there exists a constant c > 0
such that
Bw(O; 2c) c T(Bv(O; 1)). (4.4.2)
Set Xn = nT(Bv(O; 1)). Then each Xn is a closed set. Since T is linear
and surjective, it follows immediately that W = U~=l X n . Hence, W
being complete, it follows from Baire's theorem (cf. Remark 4.1.2) that
there exists n such that Xn has non-empty interior. Hence by change of
scale, it follows that
(T(Bv(O; l))t 1= 0.
Thus, there exists Yo E Wand c > 0 such that
Bw(Yo;4c) c T(Bv(O; 1)).
to 2y). Set E = c/2 and let Zl E V be such that IlzIilv < 1/2 and
IIT(Zl) - yllw < c/2.
We can iterate this procedure. By another application of (4.4.2) ( to
4(T(zI) - y)) we can find Z2 E V such that
1 c
IIz211v < 4' IIT(Zl + Z2) - yllw < 4·
Thus, we can find, by repeated use of (4.4.2), a sequence {zn} in V such
that
1 c
Ilznllv < 2n ' IIT(Zl + ... + zn) - Yllw < 2n ·
Then, it follows that the sequence {Zl + ... + zn} is Cauchy in V, and,
since V is complete, it will converge to an element Z E V such that
Ilzllv < 1 and we will also have T(z) = y. This completes the proof. •
Theorem 4.4.1 (Open Mapping Theorem) Let V and W be Ba-
nach spaces and let T E £(V, W) be surjective. Then T is an open map.
Proof: We need to show that T maps open sets in V onto open sets
in W. Let G be an open set in V. Let y E T(G). Then, there exists
x E G such that y = T(x). Since G is open, there exists r > 0 such that
x + Bv(O; r) c G. Hence, y + T(Bv(O; r)) c T(G). But by the previous
proposition, there exists s > 0 such that Bw(O;s) c T(Bv(O;r)) and
so y + Bw(O; s) c T(G) which means that T(G) is open . •
Corollary 4.4.1 Let V and W be Banach spaces and let T E £(V, W)
be a bijection. Then T is an isomorphism.
Proof: Since T is a continuous bijection, in particular it is onto and
so, by the open mapping theorem, it is an open map, which means that
T- 1 is also continuous. Thus T is an isomorphism. •
Corollary 4.4.2 Let V be a Banach space with respect to two norms,
11.111 and 11·112. Assume that there exists c > 0 such that
Remark 4.4.1 The closed graph theorem gives a convenient way for
verifying the continuity of linear maps between Banach spaces V and
W. We just have to verify that if Xn ---t x in V and if T(x n ) ---t y in W,
then y = T(x) . •
4.5 Annihilators
In the remaining sections of this chapter, we will apply the open mapping
theorem to obtain results about the range and the kernel of linear maps.
In order to do this, we need to introduce an important notion.
Definition 4.5.1 Let V be a Banach space and let W be a subspace of
V. Let Z be a subspace of V*. The annihilator of W is the subspace
of V* given by
W (4.5.1)
and
(4.5.2)
4.5 Annihilators 109
(cf. Exercises 3.8 and 3.9). Also, if G and Hare subspaces of V such
that G c H, then H..l c G..l.
and
(4.5.4)
This is a Banach space since G and H are closed. Consider the map
G x H --> G + H given by (x, y) t----t X + y. This map is clearly continuous,
linear and onto. Since G + H is a closed subspace of the Banach space
V, it is also a Banach space and so, by the open mapping theorem, there
exists a constant c> 0 such that if z E G + H with Ilzll < c, then there
exists (x , y) E G x H such that z = x + y and such that Ilxll + Ilyll < 1.
Given any z E G + H, by considering, for instance, the vector 211zl1 z we
see immediately that z can be written as z = x + y with x E G, y E H
such that
2
Ilxll ::; -llzll,
c
Ilyll < ~ll
c
ll
which proves the result with C = 2/c . •
a- b = a' + b', Ila'll ::; clla - bll, 11b'11 < clla - bll·
Now, a - a' = b' + bEG n H. Thus,
(iv)
Proof: Since an annihilator is always closed, the implication (iv) => (ii)
is trivial. The equivalence (i) {:} (iii) is an immediate consequence of
(4.5.6). To complete the proof, we need to show that (i) => (iv) and that
(ii) => (i), which we now proceed to do.
Step 2. (ii) => (i) . For any f E V*, we have that (cf. Exercise (3.8)(b))
By Corollary 4.5.2, there exists a constant e > 0 such that, for every
f E V*,
d(f,G.l n H.l) :::; e [d(f,G.l) + d(f,H.l)]
since we are assuming that G.l + H.l is closed in V*. In other words,
we have, for every f E V*,
(4.5.10)
where Bc(O; 1) denotes the (open) unit ball in G and so on. If not, let
Xo E G + H with Ilxoll < lie such that
1
T (BE(O; 1)) =:J CBF(O; 1).
Now, this implies that (cf. Step 2 of the proof of Proposition 4.4.1)
Ladi(X) = 0
i=l
for every x E V. But this implies that 'Lt=l adi = 0 in V*, which con-
tradicts the linear independence of the Ud. Thus, we can find vectors
{e1,··· ,ed} in V such that
. {I,
0,
if i = j,
if i 1= j.
It is now easy to verify that {e1,···, ed} are linearly independent and
that their span (which has dimension d) is a complement to ~ . •
Remark 4.6.3 We will see later (cf. Chapter 7) that in a Hilbert space,
every closed subspace is complemented. In fact, a deep result of Lin-
denstrauss and Tzafriri states that any Banach space which is not iso-
morphic to a Hilbert space will always have uncomplemented closed
subspaces. Since Hilbert spaces are always reflexive, it follows that, in
particular, non-reflexive spaces will always have uncomplemented closed
subspaces. For example de Vito has shown that the space Co (cf. Exer-
cise 3.7) is uncomplemented in £00 . •
4.6 Complemented Subspaces 115
T(x') = 211~l y.
Hence, if we set x = tllyllwx', we get that T(x) = y and that
2
Ilxllv < C Ilyilw.
Thus,
2
IIS(y)llv = IIP(x)llv ::; 1IPllllxliv ::; CllPllllyllw
which proves the continuity of S. Thus S is a right inverse for T . •
116 4 Baire's Theorem and Applications
A : D(A) cV -+ W.
Example 4.7.1 Let V = W = c[a, 1]. Let D(A) = Cl[a, 1]. Define
A : D(A) c V -+ V by A(u) = u' where u' stands for the derivative of
the function u. Clearly, A is densely defined and N(A) is the subspace
of all constant functions. It follows from the fundamental theorem of
4.7 Unbounded Operators, Adjoints 117
z = {v E W*
there exists C > 0 such that for all x E D(A)
I I < v, A(x) >w',w I :s; Cllxllv }.
(4.7.2)
Note that the constant C mentioned above depends on v. Clearly, Z is
a subspace of W*. If v E Z, define, for x E D(A),
Thus, the adjoint is defined for densely defined linear operators. N<r
tice that there is no reason for A * to be densely defined. We have
the following important duality relationship: for all u E D(A) and all
v E D(A*), we have
where x = (x!,···, xn) is in the base space and Y = (YI,···, Yn) IS III
the dual space. If A : en --+ em is a linear transformation which is
represented by the m x n matrix A, then we have, for all y E em and
for all x E en,
from which it follows that the matrix representing A * is none other than
the adjoint matrix A * (cf. Definition 1.1.11).
In the same way, if the matrix A represents a linear transformation
A from ]Rn into ]Rm, the adjoint of this transformation will be repre-
sented by the transpose matrix A'. •
.:J : W* x V* -+ V* x W*
by
.:J(v,j) = (-f,v).
Proposition 4.1.2 Let V and W be Banach spaces and let A: D(A) c
V -+ W be a densely defined linear operator. Let.:J be as defined above.
Then
.:J(G(A*)) = ( ( ))~.
(iii) =} (iv). Since G(A*) is always closed, the result again follows from
the closed graph theorem.
which proves (4.7.8). The remaining two relations are proved from these
two and applying (4.5.1) and (4.5.2) . •
(iii) =? (i). If N(A*) = {O}, and R(A*) is closed, it follows from the
preceding theorem that R(A) = N(A*).l = W. Thus A is onto.
is closed.
{O} implies that C.lnH.l = {O} and R(A*) closed implies that C.l+H.l
is closed. Hence, by Proposition 4.5.2, there exists a constant C > 0
such that for every z E C.l + H.l = R(A *) x W*, there exist a E C.l
and b E H.l such that z = a + b, and lIallv*xw* ~ Cllzllv*xw* and
Ilbllv*xw* ~ Cllzllv*xw*. Further, this decomposition is unique, since
C.l n H.l = {O}. Let v E D(A*). Set z = (A*(v),O) E R(A*) x W* =
C.l + H.l . Then a = (A*(v), -v) E C.l and b = (0, v) E H.l and
a + b = z. The inequality(4.7.11) now follows immediately. •
Remark 4.7.3 A similar result can be stated and proved for the sur-
jectivity of A* . •
then A = A * (check!) and A is clearly one-one, but not onto (cf. Exam-
ple 2.3.3) . •
Example 4.7.4 Let V and W be real Banach spaces and assume that
W is reflexive. Let
a(.,.) : V X W --t lR
be a bilinear form such that:
(i) a(.,.) is continuous, i.e. there exists M > 0 such that, for all v E V
and w E W, we have
la(v, w)1 II II .
sup
vEV, v~O
Ilvllv 2: a w W,
(iii) there exists a constant (3 > 0 such that, for all v E V, we have
la(v, w)1
sup
wEW, ~o
II W II W 2: (3llv I v·
Then, given f E V* and 9 E W*, there exist unique elements Vo E V
and Wo E W such that
To see this, define A: V --t W* by < A(v),w >w*,w= a(v,w) for all
w E W. By the continuity of the bilinear form, A is a well defined and
bounded linear operator. Since W is reflexive, we have A * : W --t V*,
which is also a bounded linear operator (cf. Proposition 4.7.3) and it is
easy to see that < A*(w),v >v*,v= a(v,w). By (ii), we have
IIA*(w)llv* 2: allwllw.
IIA(v)llw* 2: (3llvliv
whence A is one-one as well. In the same way, A * is one-one and onto
as well. Thus, there exist unique solutions to the equations
4.8 Exercises
4.1 Show that a Banach space cannot have a basis whose elements form
a countable set. Deduce that the space P of all polynomials in one vari-
able cannot be complete for any norm.
4.2 (a) Let {an} be a sequence of real numbers such that for any given
real sequence {xn} such that Xn ---> 0 as n ---> 00, the series
for all f E V, where {x~ ~=o are given points in [0,1] and { ~ ~=o
are real numbers (called weights). Let
'P(J) = 11 f(t)dt
for f E V.
(a) Show that 'Pn(J) ---> 'P(J) for every f E V, as n ---> 00, if, and only if,
the following conditions are verified:
(i) 'Pn(Jj) ---> 'P(Jj) as n --4 00, for every integer j 20, where fj(t) = t j ;
(ii)
(c) (Trapezoidal rule) Set Pn = n and x~ = min for 0 ::; m ::; n. Let
wn = {~ if m i= 0, n
m 2~ if m = 0 or n.
4.4 Let V be a Banach space and let {8 (t) k::o be a family of continuous
linear operators on V. Assume that the following conditions hold:
(i) 8(0) = I, the identity operator on V.
(ii) For all tl ~ 0 and t2 ~ 0, we have
118(t)11 ::; M.
for all t ~ o.
(c) The semi group is said to be exponentially stable if we can find ]v! >0
4.8 Exercises 127
and w < 0 such that the preceding inequality is true. Show that a C()-
semigroup {S (t) h2:o is exponentially stable if, and only if, there exists
a to> 0 such that IIS(to)11 < l.
(d) Prove that for every x E V, fixed, the mapping t 1---+ S(t)x is contin-
uous from the interval [0, (0) into V.
(e) Prove that
lim -h
hiO
11 t
t +h
S(T)(X) dT S(t)(x)
4.7 Let V and W be Banach spaces and let {fihEI (where I is an in-
dexing set) be a collection of continuous linear functionals on W which
separates points in W. Let T : V -+ W be a linear map. If Ii 0 T is
continuous for each i E I, show that T E £(V, W).
4.9 Let V and W be Banach spaces. let T E £(V, W). We say that
S E £(W, V) is a left inverse of T if SoT = Iv, where Iv is the iden-
tity operator on V. Show that T has a left inverse if, and only if, T is
injective and R(A) is closed and complemented in W.
IIT(v)llw 2: cllvllv.
4.12 Let Cp be the set of all real sequences such that all but a finite
number of terms are zero provided with the norrnll.lloo. Define T: cp --t
Cp by
T(x) = (Xl, ~2, ... , ~, ... )
and
AU) = f'
4.8 Exercises 129
and that
A (lot S(T)(X) dT) = S(t)(x) - x.
°
V. Let A : D(A) C V ----7 V be its infinitesimal generator.
(a) Let x E D(A). Show that S(t)(x) E D(A) for all t > and that
d
dt (S(t)(x)) = A(S(t)(x)) = S(t)(A(x)).
(b) If x E D(A) and if °:s; t2 < tl, show that
S(tl)(X) - S(t2)(X) = i h
t2
S(T)(A(x)) dT =
iti
t2
A(S(T)(X)) dT.
R(A)(X) = 10 00
e-,xtS(t)(x) dt
(cf. Exercise 3.15). If x E D(A), show that R(A)(X) E D(A) for all
A> 0 and that
R(A)(A(x)) = A(R(A)(X)).
(b) Show that for all A > 0,
(AI - A)(R(A)(X)) x
for all x E V and that
R(A)((AI - A)(x)) = x
of f(y). Then f-l(U) and f-l(V) are disjoint weakly open neighbour-
hoods of x and y respectively. This completes the proof. •
(iii) This follows from (i) and the Banach-Steinhaus theorem (cf. Corol-
lary 4.2.2 applied to the sequence {JXn } in V**).
(iv) We have
Example 5.1.1 Consider £2 the space of all square summable real se-
quences. We can identify £2 with £2 (cf. Example 3.1.1). Consider the
134 5 Weak and Weak * Topologies
Proof: Since the weak topology is coarser than the norm topology,
every weakly open set is also open in the norm topology. We thus have
to prove the converse. Let U be open in the norm topology and let
Xo E U. There exists r > 0 such that the open ball B(xo; r) c U. Let
dim(V) = n and let {VI, ... , vn } be a basis for V such that, without loss
of generality, Ilvill = 1 for all 1 :::; i :::; n. If x E V, then x = L:~=1 XiVi
and define fi to be the i-th coordinate projection, i.e. fi(X) = Xi. Then
n n
i=1 i=1
Define
w = {X E V I Ifi(X - xo)1 < ~, 1~ i ~ n} .
Then W is a weakly open neighbourhood of Xo and it is clear from the
above computations that W C B(xo;r) C U. Thus U is open in the
weak topology as well and this completes the proof. •
Thus, in a finite dimensional space the weak and norm open (respec-
tively, closed) sets are the same. However, in infinite dimensional spaces,
the weak topology is strictly coarser than the norm topology. We will
presently see examples of norm closed (respectively, open) sets which are
not closed (respectively, open) in the weak topology (cf. Examples 5.1.2
and 5.1.3 below). However, for convex sets, the situation is different.
5.1 The Weak Topology 135
U = {xEVlf(x)<a}
{x E X I f(x)::; a}
is closed in X . •
°
subsequence x nk such that f(x nk )::; a+c for all k. Since f- 1 ((-00,a+
cD is closed, it follows that f(x) ::; a + c and since c > was arbitrarily
chosen, (5.1.1) follows.
Corollary 5.1.1 Let V be a Banach space and let 'P : V ----7 JR be convex
and lower semi-continuous (with respect to the norm topology). Then
'P is also lower semi-continuous with respect to the weak topology. In
particular, the map x I--t Ilxll, being continuous, is also lower semi-
continuous with respect to the weak topology and, if Xn - ' x in V, we
have
Ilxll ::; liminfllxnll· (5.1.2)
n--->oo
Proof: For every a E JR, the set 'P- 1 (( -00, aD is closed (in the norm
topology) and is convex. Hence it is weakly closed. This completes the
136 5 Weak and Weak* Topologies
proof. •
where E > 0 and fi E V* for 1 :::; i :::; n. Consider the map A : V ---+ ~n
defined by
A(x) = (h (x), .. · ,fn(x)).
This map cannot be injective (otherwise, we will have dim(V) :::; n, which
is a contradiction). Thus, there exists Yo i- 0, such that fi(yo) = 0 for
all 1 :::; i :::; n. Then Xo + tyo E U for all t E R Set g(t) = IIxo + tyoli.
Then g(O) < 1 while g(t) ---+ +00 as t ---+ +00. Hence, there exists to
such that g(to) = 1. Thus, Xo + toYo E Un S. We have thus proved that
every weakly open neighbourhood of every point in the open unit ball,
D, intersects the unit sphere, S. Hence the closed unit ball, B, must
lie in the weak closure of S. But B being closed (in the norm topol-
ogy) and convex, is itself weakly closed. Thus the weak closure of the
unit sphere, S, is the closed unit ball, B. Thus, S is not weakly closed . •
for very positive integer i. Assume, if possible, that {xn} does not
converge to zero in norm. Then, there exist E > 0 such that, for infinitely
many n,
00
Llx~1 2 E.
k=l
Lj=l
Ixtl < :
L IX~kl < ;
j=mk+ 1
if xt = 0,
otherwise.
By varying k over all positive integers, yj will be defined for all positive
integers j. Clearly I/ylloo = 1. Also
00 mk-l 00
4E
L (X~kyj -Ixtl) < 2L IX~kl + 2 L IX~kl < -
j=1 5
j=1 j=mk+1
Thus,
00
4E E
LX~kyj > E-- -
5 5
j=1
which contradicts the weak convergence of {x nk } to zero. Hence the
result . •
Remark 5.1.1 The space £1 being infinite dimensional, the weak and
norm topologies are different. Nevertheless, we see from the preced-
ing proposition that the convergent sequences for these topologies are
the same. While two metric spaces which have the same convergent
sequences are equivalent (i. e. their topologies are the same), two topo-
logical spaces with the same convergent sequences need not be the same.
This illustrates the inadequacy of considering just sequences in a general
topological space. We also conclude that the weak topology on £1 is not
metrizable. •
Now let j E W*. Let U be open in lR (or C). Then j-l(U) is weakly
open in W, by definition. If joT is weakly continuous, we also have
that T- 1 (1-1 (U) is weakly open in V. But by the definition of the weak
topology, every weakly open set in W is the union of finite intersections
of sets of the form j-l(U), where U is open in lR (or C) and j E W*.
Thus, it follows from the above that the inverse image of every weakly
open set in W is weakly open in V, i. e. T is weakly continuous . •
Let V be a Banach space. Then its dual space, V*, has its natural norm
topology. It also is endowed with its weak topology, viz. the coarsest
topology such that all the elements of V** are continuous. We now
define an even coarser topology on V*.
Definition 5.2.1 The weak* topology on V* is the coarsest topology
such that the functionals {Jx I x E V} are all continuous, where J : x I-t
Jx is the canonical imbedding oj V into V** . •
Clearly, the weak* topology is coarser than the weak topology on
V*. Thus if S, Wand W* denote the norm, weak and weak* topologies,
respectively, on V*, we have
W* eWe S.
140 5 Weak and Weak* Topologies
are both weak* open sets and are clearly disjoint and contain JI and 12
respectively. This completes the proof. •
Proof: Let D be the open unit ball in lR (or C, in case of complex Ba-
nach spaces). Since cp is weak* continuous, there exists a weak* neigh-
bourhood of the origin in V*, say, U, such that cp(U) c D. Assume
that
1
I'PU)I < k
and so 'PU) = O. It then follows (cf. Exercise 3.17) that there exist
scalars ai for 1 ~ i ~ n such that
n
'P = :2:= aiJXi·
i=l
H = {f E V* j f(x) = a}
Proof: For simplicity, we will assume that the base field is R Since H
is a weak* closed hyperplane, it is closed in the norm topology as well
and so (cf. Proposition 3.2.1) there exists 'P E V** such that
H = {f E V* 1 'PU) = a}
or rp(J) > a for all 1 E U. Assume the former (the proof in the latter
case will be similar). Let W = U - Uo} = U - 10 lIE U}. Then,
W = {g E V* I 9 + 10 E U}
for all 9 E W. Since we can always find 10 E He such that la-rp(Jo) I < 'r},
for any 'r} > 0, and since W is weak* open, it follows that rp is weak*
continuous at the origin, and so, by linearity, weak* continuous every-
where. Then, by the preceding proposition, rp = J x for some x E V.
This completes the proof. •
x = IIxEv[-llxll,llxll]
5.2 The Weak* Topology 143
with the usual product topology inherited from R This space is clearly
compact since each bounded and closed interval in lR is compact. Let
f E B*. Then, for each x E V, we have f(x) E [-llxll, Ilxlll. Thus, the
map f t----+ <p(f) = (f(X))XEV is a bijection from B* onto its image in X.
If B is endowed with the topology induced by the weak* topology of V*,
then the definitions of this topology and the product topology on X tell
us that <p is a homeomorphism. We thus just need to show that <p(B*)
is closed in X, which will prove <p(B*), and hence B*, to be compact.
Let (fx)xEV E <p(B*). Define, for x E V, f(x) = fx. The proof will
be complete if we can show that f is linear; since If(x)1 ~ Ilxll for all
x E V, it will then follow that f E B*, i.e. <p(B*) = <p(B*).
Let E > 0. Then, given x and y E V, we can find 9 E B* such that
E E E
Ig(x) - f(x)1 < 3' Ig(y) - f(y)1 < 3' Ig(x + y) - f(x + y)1 < 3
Thus,
If(x + y) - f(x) - f(y)1 < E
for all 1 ~ i ~ n.
(ii) For all scalars f3i, 1 ~ i ~ n, we have
n
"f3'a'
~ ~ ~ <
-
i=l
from which (ii) follows, since we can choose E > 0 to be arbitrarily small.
(ii) * (i). Let a = (al,"', an) E ]Rn (the proof when the scalar field
is C is similar). Define A : V ----> ]Rn by A(x) = (h(x),"', In(x)). We
then need to show that a E A(B) where B is the closed unit ball in V. If
not, by the Hahn-Banach theorem, we can find scalars A and f31, ... ,f3n
such that, for every x E B,
n n
L a if3i > A > Lf3ili(x)
i=l i=l
Proof: Since B** is weak* compact, it is weak* closed. Let <Po E B**.
Consider a weak* open neighbourhood of <Po of the form
Notation: Given a Banach space V, we will denote the closed unit balls
in V, V* and V** by B, B* and B**, respectively.
Proof: It is easy to see that the weak topology on W is none other than
the topology induced on W by the weak topology of V (cf. Exercise 5.1).
Since V is reflexive, it follows that B is weakly compact. The unit ball
in W is none other than W n B. But W being a closed subspace, it is
weakly closed and since B is weakly compact, it follows that W n B is
weakly compact as well. Thus, it follows that W is reflexive. •
Proof: Let c.p E W** which vanishes on D(A*). It suffices to show that
c.p = O. Since W is reflexive, there exists yEW such that
for all u E D(A) and such that < v, y >w*,w# 0, by virtue of the
Hahn-Banach theorem (d. Corollary 3.2.1). It follows from (5.3.1) that
°
v E D(A*) which then implies that < v, Y >w*,w= which is a contra-
diction . •
Proof: It suffices to show that the graphs G(A) and G(A**) are the
same. Recall that if we define J : W* x V* ---+ V* x W* by J(v, f) =
(-f, v), we then have J(G(A*)) = ( )~ (d. Proposition 4.7.2). Then
( ( )~)~ = ( ( ( ))~ c V x W. Thus, ( ( )~)~ consists of all
(v, w) E V x W such that
for all <p E D(A*). This is equivalent to saying that v E D(A**) and
that w = A**(v). Thus,
Proof: Let {In} be a countable dense set in V*. Choose {xn} in V such
that
1
Ilxnll = 1 and In(xn) > 2111nll.
Assume, for simplicity, that the base field is R Let W be the linear
subspace generated by the sequence {xn} and let Wo be the set of all
finite linear combinations of the {xn} with rational coefficients. Then
Wo is countable and it is dense in W. So it suffices to show that W is
dense in V. Let I E V* which vanishes on W. We need to show then
that I vanishes on all of V (i. e. I is identically zero). Let E > O. Then,
there exists 1m such that 111-Imll < E, by the density of the {In}. Since
I(xn) = 0 for all n, we have
1
2111mll < Im(xm) (fm - J)(xm) < 111m - III·
Thus,
IIIII ~ 111m - III + Illmll < 3E
from which it follows that I = 0 since E was arbitrarily chosen. •
Ii = {O, ~f II{I ~ 1,
2, If Iii I < 1.
Then I E foo and III - Inlloo ~ 1 for all n. Thus {In} cannot be dense
in f oo .•
Corollary 5.4.1 Let V be a Banach space. Then V is both separable
and reflexive if, and only ii, V* is both separable and reflexive.
Proof: If V* is both separable and reflexive, then so is V. Conversely, if
V is separable and reflexive, so is J(V) = V**, where J is the canonical
imbedding of V into V**. Thus, it now follows that V* is separable and
reflexive . •
5.4 Separable Spaces 149
(5.4.1)
It is easy to check that d(.,.) is well defined and that it defines a metric
on B*. Let U be a weak* open neighbourhood of fo E B* of the form
1 1 r
2:
00
2n 2k < "4
n=k+l
150 5 Weak and Weak * Topologies
If f E Uk' then
Thus, Uk c Bd(fO; r) which shows that every open set in the metric
topology is also weak* open. Thus, the weak* and the metric topologies
on B* are the same.
is then countable and the set E of all finite rational linear combinations of
the elements of D is a countable set which will be dense in the subspace
generated by D. If f E V* is such that f(x) = 0 for all x in the subspace
generated by D, then clearly, f E Un for each n. Thus,
lim
k --+00
< el, enk >£=, £1 = o.
which converges to zero. Thus {en} weak* converges to zero. Thus the
weak and weak* convergent sequences in £1 are not the same (while its
norm and weakly convergent sequences are the same). •
Remark 5.4.1 The converse of the above theorem is also true and it is
a deep result due to Eberlein and Smulian: if every bounded sequence
admits a weakly convergent subsequence in a Banach space, then the
space is reflexive. •
We know that the unit ball in a normed linear space is convex. How-
ever, the nature of the boundary of this ball depends on the norm. For
instance, in ]R2, with the euclidean metric (i.e. ]R2 = ~), the unit ball is
a very symmetric object which 'bulges uniformly' in all directions. On
the other hand, if we consider ]R2 as CT or as ~, then the unit ball will
be, the rhombus bounded by the lines (±xI) + (±X2) = 1 or the unit
square, respectively. In both these cases, the boundary has a lot of 'flat'
portions. Uniform convexity makes precise the notion of the boundary
'bulging uniformly' in all directions. This is a condition describing the
'geometry' of the norm, but has an important 'analytic' consequence,
which will be the main theorem of this section. It also has important
consequences in the calculus of variations, which we will see in the next
section.
it follows that
Example 5.5.1 The spaces €l{ and ~ are not uniformly convex. In
fact, they are not even strictly convex. •
Example 5.5.2 The space €!j is uniformly convex. Let x and y E €!j.
Then it is easy to verify that
(5.5.1)
If IIxl12 :::; 1, IIyl12 :::; 1 and Ilx - Yl12 > E, with E sufficiently small, we see
that
1
11 2(x + y) 1122 < 1- 4
E2
= (1 - 8) 2
where
8=I-V 1- E
: .•
(5.5.2)
Define
U = {~E V** I I(~ - 'P)(J) I < 8/2}.
Then U is a weak* open neighbourhood of 'P in V**. Since J(B) is
weak* dense in B** (cf. Proposition 5.2.4), it follows that there exists
x E B such that J(x) E U.
"Assume now that IIJ(x) - 'Pllv** > c. In other words, 'P tf. J(x) +
cB**. Since EB** is weak* compact (by the Banach-Alaoglu theorem),
it is weak* closed and so is its translation by J(x). Thus, there exists
a weak* open neighbourhood UI of'P such that for all ~ E UI , we still
have II~ - J(x)llv** > E. Again, as before, there exists Xl E B such that
J(XI) E un UI by the weak* density of J(B) in B**. Thus,
< 1
{)
I'P(J) - f(x)1
I'P(J) - f(XI)1 < 2"
and so
Then Xn - t x in V.
liminf
n->CXl
Ilxnll ~ Ilxll. (5.5.4)
Thus we have
K= {x E K I rp(x) :s Ao}.
Since rp is lower semi-continuous and convex, it follows that K is closed
and convex. Further, by (5.6.1), it follows that K is bounded as well.
Thus, K is weakly closed (d. Proposition 5.1.4). Now, let {xn} be
a minimizing sequence in K for rp, i.e., rp(x n ) ~ infxEK rp(x). Then,
since the sequence is bounded and since V is reflexive, it has a weakly
convergent ~ s n , say, {x nk }, onv in~ weakly to some x E V.
But since K is weakly closed, we have x E K. Further, by the lower
semi-continuity and convexity of rp, it follows that rp is also weakly lower
semi-continuous (d. Corollary 5.1.1) and so
in[ rp(y).
yEK
Thus,
rp(x) = migrp(y).
yEK
rp(x) = minrp(y).
yEK
Proof: The functional z f--t Ilx - zll is clearly coercive, convex and
weakly lower semi-continuous. Thus the existence of y follows from the
preceding proposition. Assume that V is uniformly convex. Assume
that there exist Yi E K, i = 1,2 such that
Ilx - ylh 11 - al + 11 - bl
> 1- lal + 1- Ibl
2- (Ial + Ibl)
> 1.
However, if a + b = 1, a 2: 0, b 2: 0, then, for all such points Y = (a, b),
we have
Ilx - ylll = 1- a + 1- b = 1.
Thus we have uncountably many Y which satisfy (5.6.3) . •
5.7 Exercises
5.4 Let V be a Banach space. Show that V with its weak topology and
V* with its weak* topology are both locally convex topological vector
spaces (cf. Remark 3.2.2).
5.5 Use the preceding exercise and Remark 3.2.2 to show that if V is a
Banach space and if W is a subspace of V*, then the weak* closure of
W is (W l..)l... (Compare this with Exercise 3.9.)
5.7 Let 1 < p < 00. Let Xn = (xh) and x = (xj) be elements of €p. Show
that Xn -' x in €p if, and only if, xh - t xj for every positive integer j.
T(x) = (~~).
Show that T E C(CO,€1)' If B is the closed unit ball in CO, show that
160 5 Weak and Weak * Topologies
(c) Show that K does not admit an element with minimal norm.
(d) Deduce that C[O, 1] is not reflexive.
for all y E V.
(a) For any closed convex subset K c V, show that there exists x E K
such that
J(x) = min J(y). (5.7.2)
yEK
for every y E K.
(c) Show that the solution x E K of (5.7.3) (and hence, that of (5.7.2))
is unique.
(d) If K is a closed convex cone (cf. Definition 3.4.1), show that the
solution x E K of (5.7.3) is characterized by
for all y E K.
(e) If K = V, show that the solution x of (57.3) is the solution of
(5.7.1).
Chapter 6
LP Spaces
(L Ifl
1
(6.1.3)
6.1 Basic Properties 163
If(x)g(x)1 :S If(x)I·llglloo
for almost every x E X and then (6.1.3) follows on integrating this
inequality over X.
Let us now assume that 1 < p < 00 so that 1 < p* < 00 as well. The
relation (6.1.3) is trivially true if Ilfllp (respectively, Ilgllp') equals zero,
for then f (respectively, g) will be equal to zero almost everywhere. So
we assume further that Ilfllp i- 0 and that Ilgllp' i- O. Then (cf. Lemma
2.2.1)
1 1 •
If(x)g(x)1 :S -If(x)I P + -lg(x)IP
p p*
for all x E X. Assume now that Ilfllp = Ilgllp' = 1. Then, integrating
the above inequality over X, we get
For the general case, apply the preceding result to the functions f Illfllp
and g/llgilp' to get (6.1.3) . •
Thus,
..P...-
Dividing both sides by Ilf + gil$' and using, once again, the definition
of p*, we get (6.1.4). The cases where p = 1 and p = 00 follow trivially
from the inequality
While we may often talk of ' £P - functions' we must keep in mind that
we are really talking about equivalence classes of functions and that we
6.1 Basic Properties 165
Example 6.1.1 Let X = {1, 2,· .. ,n}. Let S be the collection of all sub-
sets of X and let J.L be the counting measure (cf. Example 1.3.1). Then
a measurable function can be identified with an n-tuple (aI, a2,··· ,an).
In this case LP(J.L) = e;. Notice that in this example, equality almost
everywhere is the same as equality everywhere and so every equivalence
class is a singleton. •
Example 6.1.2 Let X = N, the set of all natural numbers and let S
be the collection of all subsets of X. Let J.L be the counting measure. In
this case, functions are identified with real sequences and LP(J.L) = ep .
Again, in this example, equivalence classes are just singletons . •
Proposition 6.1.3 Let (X,S,J.L) be a finite measure space, i.e. J.L(X) <
00. Then
Proof: The result is trivial if p = 00. Let 1 :S q <p < 00 and let
166 6 LP Spaces
where
Example 6.1.4 Nothing can be said about the reverse inclusions. For
example, if f(x) = 1/ y'X, then f E L1(0, 1) but f tj L2(0, 1). However,
we know that (cf. Exercise 2.25) £p C £q for all 1 :S p < q :S 00 . •
1
Ilfnk - fnk+lllp :S 2k ·
Set
n
gn(x) L Ifnk+Jx) - fn,,'(x) I
k=l
and
00
It follows that gn(x) ----+ g(x) and, by the monotone convergence theorem
(cf. Theorem 1.3.1), we see that Ilgllp :S 1. In particular, g(x) < 00
almost everywhere. Further, if k 2 I 2 2, we have
Ifnk(X) - fnl(X)1 < Ifnk(X) - fnk-l (x)1 + ... + Ifnl+l (x) - fnl(X)1
< g(x) - gZ-l(X).
Thus, it follows that, for almost every x EX, {fnk (x)} is a Cauchy
sequence and converges almost everywhere to a finite limit f(x) and
that, for such x,
If(x) - fnk (x)1 :S g(x)
for k 2 2. Set f 0 elsewhere, which is a set of measure zero. It
then follows that f is p- integrable. Further, Ifnk (x) - f (x)IP ----+ 0 almost
everywhere and is bounded by Ig(x)IP which is integrable. Hence, by the
dominated convergence theorem (cf. Theorem 1.3.3), we deduce that
Ilfnk - flip ----+ O. Thus we have that
in V(p,).
Case 2. p = 00. Let {fn} be Cauchy in LOO(p,). Then, for each k, there
exists a positive integer Nk such that
for all m, n 2 N k . Thus, there exists a set Ek of measure zero, such that
1
If(x) - fn(x)1 :S k·
Hence, it follows that f is essentially bounded and that fn ----+ f in LOO(p,).
Tg(f) = Ix fg dJ.L
Thus, the map g 1---+ Tg is an isometry from V* (J.L) into LP(J.L)*. Hence
its image is closed. It is enough now to show that the image is dense.
Step 2. We now show that V(J.L) is reflexive for all 1 < p < 00. This
has already been proved for 2 ::; p < 00. Thus LP(J.L)* is also reflexive
for such p and so is every closed subspace of this dual space. Thus, by
170 6 V Spaces
ixfgdJ.L = o.
Remark 6.2.1 We have seen earlier that ti = too. In the same way, it
is true that for O"-finite measure spaces, we have L 1 (J.L)* = Loo(J.L). How-
ever, the proof of this result relies on very measure theoretic arguments
and we shall omit it. Nevertheless, in the next section, we will prove it
for a very important class of L1 spaces . •
Proposition 6.3.1 Let S be the set of all simple functions which vanish
outside a set of finite measure. Then S is dense in V(n) for 1 S p < 00.
6.3 The Spaces V(n) 171
In Itpn - flP dx -t 0
Remark 6.3.1 In fact it can be shown that the space of infinitely dif-
ferentiable functions with compact support contained in 0 is dense in
V(O) for 1 :S p < 00. For this we need to develop the theory of convolu-
tion of functions (cf. Theorem 6.3.3 below). For details, see Kesavan [3] .
•
Corollary 6.3.1 Let 0 C jRN be an open set. Let 1 :S p < 00. Then,
V(O) is separable.
Proof: Let x E O. Let r r(x) > 0 be chosen such that the ball
B(x; r) c O. Define
I, if Y E B(x;r)
Xx(y) {
0, otherwise.
6.3 The Spaces LP (0,) 173
Set
Ux = {f E LOO(n) I Ilf - Xxlloo < 1/2}.
Then, for each x E 0" Ux is a non-empty open subset of LOO(n). Further,
if x i- y, then Ux n Uy = 0. For, if f belonged to their intersection,
then on one hand, since f E Uy, we have that If(z)1 < 1/2 in a small
neighbourhood of x. On the other hand, since f belongs to Ux, it follows
that If(z)1 > 1/2 in a small neighbourhood of x, which is a contradiction.
Now let E = {fn} be any countable set in LOO(n). If such a set were
dense, then En Ux i- 0 for each x E n. However, any fn can belong
to at most one such open set Ux since the sets Ux are pairwise disjoint.
This is a contradiction since the number of open sets Ux is uncountable.
Thus, no countable set in LOO(n) can be dense . •
In fg dx = 0
Let
Kl = {xEnlh(x)2:c}
K2 = {x En I h(x) :s; -c}.
in
r hh dx = r
in\K
hh dx + r hh dx,
iK
whence, in view of (6.3.1), we have
r Ihl dx
iK
=
iK
r hh dx :S E + r
in\K
Ihhl dx < E + r
in\K
Ihl dx.
Proof: Step 1. There exists wE L2(!1) such that w(x) ~ EK > 0 for all
x E K for every compact subset K of!1. Indeed, define w(x) = an > 0
on the set
En = {x E !1 I n :S Ix I < n + I},
where Ixl denotes the Euclidean norm of the vector x E ]RN. Now choose
the constants an such that
00
where IEnl denotes the (Lebesgue) measure of the set En. Then w has
the required properties.
Step 2. Let <P E L1(n)*. Consider the mapping f f--+ <p(wf) from L2(n)
into R Clearly, this defines a linear functional which, by Holder's in-
equality, is also continuous. Thus, by the Riesz representation theorem
(cf. Theorem 6.2.1) applied to the case p = 2, there exists v E L2(n)
In
such that
<p(wf) = fv dx
A = {x E 0. Ilu(x)1 > C}
is of measure zero.
Assume the contrary for some such C > 11<pII.Then, there exists a
subset B of A of finite and positive measure. Consider the function
l'u,w dx S "<p'll w dx
C l w dx s 1'<p'll w dx
Step 4. Thus we now have u E DXl(O) with Iluli co :S 11<p11 such that, for
all f E L2(0),
<p(wf) = in fuw dx.
Since Cc(O) is dense in L1(0), the above relation also holds for all g E
L1(0). Further, it follows that
Step 5. Thus, for every <p E L1(0)*, we have u E LCO(O) such that
11<p11 = Ilull oo and such that (6.3.3) holds for all g E L1(0). Such a u is
unique as well. Indeed if we have two essentially bounded functions U1
l
and U2 such that
g(U1 - U2) dx = 0
for all g E L1(0), then it is in particular true for all 9 E Cc(O) and, since
essentially bounded functions are locally integrable, it follows that (cf.
Proposition 6.3.3) U1 -U2 = 0 almost everywhere, i.e. U1 = U2 in Loo(O).
(6.3.4)
Now choose h E Cc(n\ {O}). Then, for sufficiently large k, we have that
l/nkh dx = 0
(since the two functions in the integrand will then have disjoint sup-
ports) and so, it follows that, for all such h, we have In
fh dx = O. By
Proposition 6.3.3, it then follows that f(x) = 0 almost everywhere in
0, \ {O} and so f (x) = 0 almost everywhere on 0, as well. On the other
hand, if we choose h(x) = 1 for all x E 0, in (6.3.4), we get In
f dx = 1,
which is a contradiction. Thus, L1 (0,) is not reflexive . •
Corollary 6.3.2 Let 0, C ]RN be an open set. Then, LOO(n) is not
reflexive.
Proof: Since L1(n)* ~ vx>(n), the result follows immediately from the
preceding proposition (cf. Corollary 5.3.3) . •
To sum up, we have that V(n)* ~ LP' (0,) for 1 'S p < 00. The
spaces V(n) are separable for 1 'S p < 00 and reflexive for 1 < p < 00.
The space LOO(n) is neither separable nor reflexive.
We conclude by proving an important inequality.
Theorem 6.3.3 (young's Inequality) Let 1 < p < 00. Let f E
L1(]RN) and let g E LP(]RN). Then the map
I = r r
~ j ~f
If(y)g(x - y)h(x)1 dy dx.
178 6 LP Spaces
ilR
r f(y)g(x - y)h(x) dy
N
exists for almost all x E ]RN. Let us choose h E y* (]RN) such that
h(x) i- 0 for all x E ]RN. For example, we can choose h(x) = exp( -lxI 2 ).
Thus, it follows that the integral
ilR
r f(y)g(x - y) dy
N
exists for almost all x E ]RN and so the convolution f * g is well defined.
Further, by the above computation it follows that the map
Remark 6.3.3 The result of Theorem 6.3.3 is valid for the case p = 1
as well. The proof of this fact is left as an exercise (cf. Exercise 6.16) .
•
6.4 The Spaces W1,P(a, b)
Lemma 6.4.1 Let f E V(a, b). Assume that there exists g E V(a, b)
such that, for all rp E V( a, b), we have
(6.4.1)
Definition 6.4.1 Let (a, b) c lR be a finite interval and let 1 ::; p < 00.
The Sobolev space W 1,P( a, b) is 9iven by
Further, we define
1
Ilfl11,p = (1Ifll~ + Il ll~)p •
It is a routine verification to see that 11.111,p defines a norm on W 1 ,P(a, b)
and this is left to the reader. Thus, W 1 ,P(a, b) is a normed linear space.
Example 6.4.1 Let f E e 1 [a,bJ. Clearly f E LP(a, b). If f' denotes the
derivative of f, then f' E era, bJ and so f' E V(a, b) as well. Further,
if <p E V(a, b), then since <p(a) = <p(b) = 0, we have, by integration by
parts,
lb frp' dx = -l b
!,rp dx.
for all n. Passing to the limit as n - t 00, we deduce that the pair (f, g)
satisfies (6.4.1). Thus f E W1,P(a, b) and f' = g. Further, it follows that
fn - t fin W1,P(a, b). This completes the proof. •
Proof: Since the space V(a, b) is reflexive if 1 < p < 00, so is the space
(V(a, b))2 (why?). Similarly, (V(a, b))2 is separable if 1 :S p < 00.
Now, the space W1,P(a, b) is isometric to a subspace of (LP(a, b))2 via
the mapping f f---t (f, f'). Since W1,P(a, b) is complete, the image is a
closed subspace of (LP (a, b))2 and so it inherits the reflexivity and sep-
arability properties from that space. This completes the proof. •
Lemma 6.4.2 Let r.p E V(a, b). Then, there exists 7jJ E V(a, b) such
that 7jJ' = r.p if, and only if,
lab r.p(t) dt O.
6.4 The Spaces W1,P(a, b) 181
Proof: Assume that cp = 'Ij;' for some 'Ij; E V(a, b). Then, since 'Ij;(a) =
'Ij;(b) = 0, it follows that
Conversely, let J:
cp(t) dt = O. Let the support of cp be contained in
[c, dJ c (a, b). Now, define
Clearly 'Ij; is infinitely differentiable since 'Ij;' = cpo Further, 'Ij; vanishes
on the interval (a, c) and, by hypothesis, on the mterval (d, b) as well.
Thus the support of'lj; is also contained in [c, dJ and so 'Ij; E V( a, b). This
completes the proof. •
Corollary 6.4.1 Let f E £P(a, b) where 1 ::; p < 00. Assume that
Then J: ¢>(t) dt = 0 and so ¢> = 'Ij;' for some 'Ij; E V(a, b). Thus,
J: f ¢> dt = 0 which yields
1b (J - c) cp dt 0
182 6 LP Spaces
for all 'P E V(a,b), whence, by Proposition 6.3.3, it follows that f(x) = c
almost everywhere in (a, b). This completes the proof. •
Remark 6.4.1 If 'PI were another function in V(a, b) such that J: 'PI (t)dt
= 1, then since J: ('Po - 'PI) dt = 0, it follows that 'Po - 'PI = 1/J' for some
1/J
E V( a, b). Therefore, by hypothesis, J: f ('Po - 'PI) dt = O. Thus, the
constant c defined in the above proof does not depend on the choice of
the function 'Po whose integral is unity. •
Let us denote by Coo [a, bJ the space of all functions which are in-
finitely differentiable in the open interval (a, b) and such that the func-
tions and all their derivatives possess continuous extensions to [a, bJ.
Proposition 6.4.3 Let 1 :s: p < 00. Then COO [a, b] is dense in WI,P(a, b).
Proof: It is clear that if f E Coo [a, b], then f E WI,P( a, b) and its
distributional derivative is just its classical derivative. Now, let f E
WI,P(a, b). Since f' E V(a, b), choose 'Pn E V(a, b) such that 'Pn --t f' in
l
V(a, b). Define
x
1/Jn(x) = 'Pn(t) dt.
we have
n
L If(xD - f(Xi)1 < E• •
i=l
Clearly, any absolutely continuous function is uniformly continuous.
It can also be shown (cf. Royden [5]) that an absolutely continuous
function is differentiable almost everywhere and that its derivative is an
integrable function. The following two results are very important (cf.
Royden [5]).
f(x) = f(a) + l x
f'(t) dt . •
1'(X) = l x
f'(t) dt.
Since f' E LP (a, b) and since p 2: 1, it follows that f' is integrable on (a, b)
and so u is ail absolutely continuous function. Further, since integration
by parts is valid for absolutely continunus functions, it follows that for
all cp E V(a, b), we have
(6.4.2)
and
f(x) = f(a) + l x
f'(t) dt. (6.4.3)
> ~
2
whence it would follow that
Further, let x, y E [a, bj. Assume, without loss of generality, that x ::; y.
Then
Definition 6.4.3 Let (a, b) C R be a finite interval and let 1 ::; p < 00.
The closure of V(a, b) in W1,P(a, b) is denoted ~ P( , b) . •
f(x) = l x
f'(t) dt
and so
ib 'Pn dt -+ O.
7/Jn = 'Pn - (i b
'Pn dt) 'Po,
6.4 The Spaces W1,P(a, b) 187
f(x) = l x
f'(t) dt.
Clearly,
1
Ifh,p < Ilflh,p < [(b - a)P + 1] p IfI1,p'
188 6 IJ> Spaces
Sobolev spaces can also be defined when p = 00. The definition can
also be extended to cover functions defined on arbitrary open sets 0 C
jRN. It is also possible to define 'higher order distributional derivatives'
and define Sobolev spaces Wm,P(O), mEN, based on these derivatives.
All these spaces have properties similar to those proved in this section,
with or without additional hypotheses. For a detailed study of Sobolev
spaces, see Kesavan [3]. See also the Exercises 6.24-6.26 below.
6.5 Exercises
6.1 Let (X, S, JL) be a measure space. let 1 :S p, q, r < 00 such that
1 1 1
- + -q
p
= -.
r
6.2 Let (X, S, JL) be a measure space and let 1 :S p < 00. Define, for
t > 0,
Show that
Ilfll~ = p fooo tp-1hf(t) dt.
(Hint: Write hf as an integral over a subset of X and apply Fubini's
theorem (cf. Theorem 1.3.5).)
6.3 (a) Let (X, S, JL) be a measure space. Let In, 9n'!, 9 be measurable
functions such that In -. I and 9n -. 9 almost everywhere in X. Assume
further that IIn(x)1 :S 9n(X) for all x E X and that
6.6 Let (X, S, f.L) be a measure space and let 1 < p < 00. Let f :
X x X ----t JR be such that for almost every y EX, the section fY (cf.
Definition 1.3.7) is p-integrable. Define, for x E X,
6.7 Let 9 E Cc(JR). Define c.p(g) = g(O). Then c.p can be extended to a
continuous linear functional on L 00 (JR). Show that there does not exist
f E Ll(JR) such that
c.p(g) = l gf dx
for all 9 E Loo(JR). (This gives another proof that Loo(JR) is not reflexive.)
190 6 LP Spaces
6.9 Let
fn = X[n,rz+11'
the characteristic function of the closed interval [n, n + 1] for n E N (cf.
Definition 1.3.5). Then {fn} is a bounded sequence in L1(1R.). Show that
it does not have a weakly convergent subsequence. (In view of Theorem
5.4.2, this p;ives another proof that Ll (IR.) is not reflexive.)
11 xnf(x) dx = 0.
Show that f : ;: : 0.
6.11 (Hardy's inequality) Let f E LP(O, 00), where 1 < p < 00. Define
g(x) = - 11
x 0
X
f(t) dt
lim -
c---.oo C
1i 0
c
h(t) dt = 0.
lim
n---.oo
lb
a
f(t) cos nt dt = lim
n---.oo
lb
a
f(t) sin nt dt = 0.
6.14 (a) Let (a,b) C (0,00) be any finite interval. Let fn(t) = cosnt
and let gn(t) = sinnt. Show that fn ----' 0 and gn ----' 0 in £P(a, b) for any
1 ::; p < 00.
(b) What is the weak limit of hn in £P(a, b) for 1 ::; p < 00 where
hn(t) = cos 2 nt?
° °
(a) above converges over a set E whose measure is strictly positive, then
an - ? and bn - ? as n - ? 00. (Hint: Use the amplitude-phase form of
the series.)
192 6 LP Spaces
f*g = r
J{N
f(y)g(x-y)dy
is well defined for almost all x E ]RN. Show also that f * g E L1 (]RN) and
that
Ilf * gill : : : Ilfllt Ilglh·
°
6.17 Let {Pc}c>o be a family of Coo functions in ]RN such that for each
E > 0, we have that Pc(x) 2 for all x E ]RN, the support of Pc is
contained in the closed ball with centre at the origin and radius E, and
r
JlR N
Pc(x) dx = l.
(a) Let 'P E Cc(]RN). Show that Pc * 'P converges uniformly to 'P on ]RN
as E -> 0.
(b) Deduce that, if U E V(]RN), then Pc * u converges to u in V(]RN)
as E -> 0.
6.18 Let {fn} be a bounded sequence in LP(a, b), where (a, b) is an open
interval in ]R and 1 ::::: p ::::: 00. Show that fn ~ f in V(a, b) when
1 ::::: p < 00 and fn ~ f in LOO(a, b) if, and only if, for every 'P E V(a, b),
lb lb
we have
fn'P dx -> f'P dx.
6.19 Let f : [O,IJ -> ]R be a continuous function such that f(O) = f(I).
Define the sequence Un} as follows. Let fn(x) = f(nx) on [0, ~ and
extend this periodically to each subinterval k~l, ~ for 2 ::::: k ::::: n. Let
m = Jo1 f(t) dt. Show that fn ~ fin V(O, 1) for 1 ::::: p < 00 and that
fn ~ fin LOO(O, 1), where f(t) = m for all t E [0,1].
6.21 Let a < c < b in JR. Let f : [a, b] -+ JR be continuous. Assume that
f E W1,P(a, c) and that f E W1,P(b, c). Show that f E W1,P(a, b).
Show that f tf. W1,P( -1,1) for 1 :::; p < 00. (Thus, continuity is essential
in the previous exercise.)
lb f(t) dt = 0.
°: :;
(b) Let ( be a Coo function on JR with compact support contained in
[-2,2] such that «(x) :::; 1 for all x E JR and such that ( == 1 on
[-1,1]. Define (m(x) = «(x/m) for all x E R Show that if u E W 1,P(IR)
for 1 :::; p < 00, then (mu -+ U in W1,P(JR) as m -+ 00.
Hilbert Spaces
Hilbert spaces form a special class of Banach spaces with the geometric
notion of orthogonality of vectors, or more generally, the notion of an
angle between vectors, built into them.
Consider the space JR2. If x = (Xl, X2) and Y = (YI, Y2) are vectors
in JR2, then we define the scalar product of these vectors by
Remark 7.1.1 The linearity with respect to the second variable is,
clearly, a consequence of conditions (i) and (ii) above . •
196 7 Hilbert Spaces
Remark 7.1.2 In case the base field is e, then the inner product is a
sesquilinear form. If x and Y E V, we have
Example 7.1.1 Consider the space ]R.n. For x = (Xl,··· ,xn ) and y =
(YI,··· ,Yn), we define
n
(x,y) = LXiYi.
i=l
This defines an inner product and the norm associated to it is the norm
11.112. Thus f~ is a Hilbert space. In the case of en, the inner product is
given by
n
(x, y) L XiYi·
i=l
Example 7.1.2 Consider the space £2. For x and Y E £2, define
ex)
(X,y) = LXiYi
i=l
where x = (Xi) and Y = (Yi) are real sequences. Again, if the base field
is e, then we define
00
(x,y) = LXiYi.
i=l
(f, g) = L fg dJ-l
in the real case; if the field is <C then the middle term on the right will
be replaced by 2Re(x, y), where Re z denotes the real part of a complex
number z. Writing a similar expression for Ilx - Yl12 and adding the two,
we get
(7.1.2)
(7.1.2) is a Hilbert space, i.e. the norm comes from an inner product .
•
Example 7.1.5 The space C[-l, 1] cannot be made into a Hilbert space.
To see this, consider the functions
Proof: The proof is of the uniform convexity follows from the par-
allelogram identity (7.1.2) exactly as described in Example 5.5.2; the
reflexivity now follows from Theorem 5.5.1. •
Equality occurs in this inequality if, and only if, x and yare scalar
multiples of each other.
fy(x) = (x, y)
Remark 7.1.4 We will see in the next section that all continuous linear
functionals on a Hilbert space occur in this manner. •
Corollary 7.1.2 Let H be a Hilbert space and let Xn ----'- x and Yn ----+ Y
in H. Then
and
Thus,
and the result now follows from the Cauchy-Schwarz inequality being
applied to the term on the right-hand side . •
for every y E M.
for all y EM. Since we also have -y EM, we get the reverse inequality
as well and this proves (7.1.6). It now follows from (7.1.61 that PM is
a linear map and it is continuous by the preceding proposition. This
completes the proof. •
Proof: Set
Earlier, we saw that every vector in a Hilbert space gave rise to a con-
tinuous linear functional. The main result of this section is to show that
all continuous linear functionals arise in this way.
that there exists x E H such that fy(x) = 0 for all y E H. This implies
that (x, y) = 0 for all y E H. In particular,
IIxl12 = (x, x) = 0
which shows that x = 0, i.e. 'P is zero. This shows that iI>(H) is dense
in H* and the proof is complete. •
It is easy to see that this defines an inner product which gives rise to
the usual norm on H*. Thus, H* also becomes a Hilbert space in its
own right. In the same way, H** also becomes a Hilbert space. Now,
we have two natural mappings from H into H**. The first is the usual
canonical imbedding x f-+ J(x). The second is the mapping x f-+ iJx'
i. e. the composition of the Riesz map H ---> H* and that of H* ---> H**.
We will show that these are the same. The latter map, by the Riesz
representation theorem, is onto and so J will be onto, giving another
proof of the reflexivity of a Hilbert space, provided we prove the Riesz
representation theorem independently, as suggested in Remark 7.2.1. To
see that the maps are the same, observe that if f = fy E H*, then
Remark 7.2.4 In the case of real Hilbert spaces, while we can identify
a Hilbert space with its dual, we have to be careful in doing so and
204 7 Hilbert Spaces
for every v E V.
Let us now identify H* with H via the Riesz representation theorem.
Let f E H. Then the map v 1----+ (v, f) H defines a continuous linear
functional on V since
IIT(f)ILC(H,v*) :S CllfllH.
If T(f) = 0, then (v, f) = 0 for all v E V and so, by density, we have
f = 0. Thus T is one-one as well. Finally, we claim that the image of T
° °
is dense in V*. Indeed, if'P E V** vanishes on T(H), then, by reflexivity,
there exists v E V such that T(f)(v) = for all f E H i.e. (v, f) =
for all f E H. Since V c H, it follows that (v, v) = 0, i.e. v = 0, which
means that 'P is idertically zero, which establishes the claim.
Thus we have the following scheme:
V c H ~ H* c V*
where both the inclusions are dense. It would now be clearly absurd
for us to identify V with V* as well. Thus we cannot simultaneously
identify V and H with their respective duals. The space H in this case
is called the pivot space and is identified with its dual, whereas the other
spaces, though they are also Hilbert spaces, will not be identified with
their respective duals. This situation typically arises when we have a
parametrized family of Hilbert spaces as in the case of the Sobolev spaces
(cf. Kesavan [3]). In particular, we can set V = HJ(a, b) (cf. Example
7.1.4) and H = L2(a, b). We identify L2(a, b) with its dual while we
denote the dual of HJ(a,b) by H- 1 (a,b) and we have the inclusions
7.2 Dual of a Hilbert Space 205
Remark 7.2.5 In the case of real Hilbert spaces, since Hand H* can
be identified via the Riesz isometry, the map A * is just the adjoint in
the sense of Definition 4.7.2 . •
II All = sup sup I(A(x), y)1 sup sup l(x,A*(y))1 :::; IIA*II·
IIxl1911yll9 IIxl1911yl19
206 7 Hilbert Spaces
Similarly,
IIA*II = sup sup I(x, A*(y))1 = sup sup I(A(x), y)1 < IIAII·
iiyii9iixli9 iiyii9iixii9
This proves (i). Again, if x and y E H, then
This proves (ii). The other relations follow trivially from (7.2.2) . •
All the results of Section 4.7, in particular, Proposition 4.7.3 and The-
orem 4.7.1, are true . •
for all x and y E H. Assume further that a(.,.) is H-elliptic (or, coercive;
cf. Exercise 5.16). Let 0: > 0 such that
for all x E H.
a(x, y) = y' Ax
where y' is the transpose of the column vector y. Then a(.,.) defines
a continuous bilinear form on £~. If A is symmetric, then the bilinear
form is symmetric as well. If A is positive definite, then the bilinear
form is coercive . •
(A(u), v) = a(u, v)
for all v E H. Clearly, the map u f--t A(u) is linear. Further, by (7.3.1)
and (7.3.2) we have
(-pA(x)+pf+x-x,y-x) < 0
for all y E K. In other words (cf. Theorem 7.1.2),
a(x, y) = (f, y)
for every y E H.
Proof: Applying the preceding theorem with K = H, there exists a
unique x E H satisfying (7.3.3). Replacing y by y + x, we get
a(x,y) ~ (f,y)
for every y E H. Since -y E H as well, we also get the reverse inequal-
ity. Hence the result . •
and
Remark 7.4.1 If we use the Kronecker symbol, viz. 8ij which equals
unity if i = j and equals zero if i ¥: j, then the above relations can be
written as
Next, consider the vector X2 - (X2' el)el. This vector cannot vanish
since Xl and X2 are linearly independent and el is a scalar multiple of
Xl. Thus, we can define
It is now immediate to check that IleI II = IIe211 = 1 and that (el' e2) = O.
Further, e2 is a linear combination of Xl and X2, since el is just a scalar
multiple of Xl.
We can now proceed inductively. Assume that we have constructed
the vectors el, ... , ek, for 1 :S k :S n - 1. We then define
ek+1 = II
xk+l
_ 2: i=l
k\ .) .11 [Xk+l - I)Xk+l,ei)eij.
xk+l,e l el i=l
212 7 Hilbert Spaces
It is now easy to verify that the set {el,··· , en} verifies the conditions
mentioned in the statement of the proposition. •
In the exercises at the end of this chapter, we will see important ex-
amples of the Gram-Schmidt orthogonalization process leading to vari-
ous well known special functions of mathematical physics.
Remark 7.4.3 Consider the space IR n = £2. For any 1 :'S j :'S n, the
sets {Xl, ... , Xj} and {el' ... , ej} span the same subspace. Thus, we can
write
j
Xj = L rijei·
i=l
Let A be the matrix whose columns are the Xj, and Q the matrix whose
columns are the ej. Let R be the matrix whose entries are the rij. For
any j, we have that rij = 0 if i > j. Thus, R is an upper triangular
matrix. Further, we see that
A =QR.
l
roIIJ(t) - p(t)12 dt = min r IJ(t) - q(tW dt
in qEPn io
where P n is the space of all polynomials (in one variable) of degree less
than or equal to n. In other words, we are looking for the projection of
J onto the subspace of polynomials of degree less than , or equal to, n
7.4 Orthonormal Sets 213
in the space L2(0, 1). We know that such a P exists uniquely and that
it is characterized by (cf. Corollary 7.1.3)
(p, q) = (f, q)
Ax = f
a··
tJ
= (p.J, p.)
t
= 1e+
0
1 . .
J dt =
1
---'
i + j + 1'
Now,
n n
(p,qd = L (Jj(qj, qi) L {Jjbji {Ji
j=O j=O
and so
(Ji = fal f(t)qi(t) dt.
Thus, without solving any linear system, we can directly compute t~
least squares approximation. •
q2(t) = t 2 - v'3
v'2
(1 1
-1
t 3 dt) v'3 t __1_
v'2 v'2
(1 1
-1
t 2 dt) ~
v'2
= t2 _ ~
3
Then, a straight forward computation yields that IIq2112 = 2v'2/3J5.
Thus, for all t E [-1,1], we get
and so on.
An easier way of computing these polynomials will be seen in the
exercises at the end of this chapter. •
7.4 Orthonormal Sets 215
Proof: We know that Ilx - L~=l (x, ei)eil1 2 2: O. Expanding this, and
using the orthonormality of the set, we get (7.4.1) immediately. Further,
Proof: Define
L l(x,ei)1 2.
iEI
choose a numbering el, e2,··· ,en,··· for the elements in the orthonormal
set whose inner product with x is non-zero. Then we define the above
sum to be
00
L(X, ei)ei
iEI
If m > n, then
111
IIYm - Ynl1 =
2 L I(x, ei)1 2
i=n+l
using the orthonormality of the set. But the sum on the right-hand side
can be made arbitrarily small for large nand m since it is part of the tail
7.4 Orthonormal Sets 217
L
00
llYn - yll < E, Ily~ - y'll < E and I(x, edl 2 < E2.
i=N+I
{el,"',e n } C { ~, , ~ .
Then, the difference y~ - Yn will consist of a finite number of terms of
the form (x, ei)ei where all the i concerned are greater than n(2:: N).
Hence, it follows that
L
00
Thus,
n
lim ~(x, j) j.
n +oo~
j=l
x - I:(x, ei)ei
iEI
Proof: Given any chain (with respect to the partial ordering induced
by set inclusion on orthonormal sets), the union of its members gives
an upper bound. Hence, by Zorn's lemma, there exists a maximal or-
thonormal set. •
(iv) If x E H, then
IIxl12 = I: I(x, edl 2 (7.4.5)
iEI
Proof: (i) :::::} (ii). Assume that the orthonormal set is complete and
that (x, ed = 0 for all i E I. If x -# 0, then the set
is also an orthonormal set strictly larger than the given set which con-
tradicts the maximality of the given set.
7.4 Orthonormal Sets 219
(iv) =} (i). If the given set were not complete, then there exists e E H
such that Ilell = 1 and (e, ei) = 0 for every i E I. But then, this will
contradict (7.4.5) (applied to the vector e) . •
Corollary 1.4.1 Let H be a Hilbert space and let {ei liE I} be an
orthonormal set. It is complete if, and only if, the subspace of all (finite)
linear combinations of the ei is dense in H.
Proof: If the orthonormal set is complete, then by the preceding theo-
rem, every element of H is the limit of finite linear combinations of the
ei by (7.4.4) and so the subspace spanned by the ei is dense in H.
Conversely, if the subspace spanned by the ei is dense in H, then if
x E H is such that it is orthogonal to all the elements of this subspace,
then x = o. In particular, if (x, ei) = 0 for all i E I, then, clearly, x
is orthogonal to the subspace spanned by the ei and so it must vanish.
Thus, statement (ii) of the preceding theorem is satisfied and so the
orthonormal set is complete. •
Corollary 1.4.2 Let H be a Hilbert space and let {el' e2, ... , en ... } be
a sequence in H which is also an orthonormal basis for H. Then en ----'" o.
Proof: Let x E H. Then, by (7.4.5), it follows that (x, en) -+ 0 as
n -+ 00. Then, by the Riesz representation theorem, it follows that
en ----'" 0 . •
for all n i= m . •
220 7 Hilbert Spaces
Example 7.4.5 (Fourier series) Consider the space L2( -7r, 7r). The
set
where
for t E (-7r, 7r), forms an orthonormal set. By Theorem 6.3.1, the space
of continuous functions with compact support contained in (-7r, 7r) is
dense in L2( -7r, 7r). Such functions vanish at the end points of the in-
terval [-7r,7r] and so they are 27r-periodic on the interval [-7r,7r]. Con-
sider the space spanned by the orthonormal set mentioned above. By
an application of the Stone-Weierstrass theorem (cf. Rudin [6]), it fol-
lows that this space is dense in the space of all 27r-periodic continuous
functions with respect to the sup norm, which is nothing but the norm
11.1100. Since the interval (-7r,7r) has finite measure, this implies that
this space is dense with respect to the norm 11.112 as well (cf. Proposition
6.1.3). This shows that the space spanned by this orthonormal set is
dense in L2( -7r, 7r) as well and so, by Corollary 7.4.1, it follows that it
is a complete orthonormal set in L2( -7r, 7r).
7.4 Orthonormal Sets 221
This is nothing but the classical Fourier series of a function f and the
°
an, n 2 and bn , n 2 1 are the usual Fourier coefficients of f. The
above series expansion means that the partial sums of the Fourier series
converge to f in the 11.112 norm. In other words if
N
fN(t) ~o + l:)a n cos nt + bn sin nt),
n=l
for t E (-Jr, Jr), then
Example 7.4.6 (Fourier sine series) Consider the space L2(0, Jr).
Consider the set
. This is an orthonormal set in L2(0, Jr) as one can easily verify. Let
f E L2(0, Jr) be orthogonal to every element of this set. Extend f as an
222 7 Hilbert Spaces
odd function to all the interval (-7r, 7r). Thus, we set f(x) = - f( -x) if
I: I:
x E [-7r,0). Since f is an odd function, it follows that
L b sin nt
00
f(t) n
n=l
where
bn = -217r f(t) sin nt dt.
7r 0
This is called the Fourier sine series of the function f. •
x = L(x, en)en
n=l
as its Fourier expansion and the coefficients (x, en) are called its Fourier
coefficients.
In the next chapter we will see how orthonormal bases occur very
naturally in Hilbert spaces.
7.5 Exercises
7.1 Let V be a real Banach space and assume that the parallelogram
identity holds in V. Define
Show that this defines an inner product which induces the given norm
and hence that V is a Hilbert space.
7.2 Let V be a complex Banach space and assume that the parallelogram
identity holds in V. Define
Show that this defines an inner product which induces the given norm
and hence that V is a Hilbert space.
y = >..g + z
cp(y) = (y, x)
a(u,v) = (f,v)
°
where sgn(t) equals 1 when t 2: and equals -1 when t < 0.
(b) Show that { n(t) ~o is orthonormal in L2(0, 1) but that it is not
complete.
7.9 Let (a, b) C lR be a finite interval and let {cPn }nEN be an orthonormal
basis for L2(a, b). Define
These are the Legendre polynomials. Show that Pn(t) consists only of
even powers of t when n is even, and of only odd powers of t, when n is
226 7 Hilbert Spaces
odd.
(b) Show that Po(t) == 1, PI(t) = t and that, for n;::: 1,
and let H be the space of all equivalence classes (with respect to equality
almost everywhere) of functions in ii. Define the inner-product
2 ~ 2
Hn(x) = (-lte x _ e - x .
dxn
7.13 Let f, g E L2 (-1f, 1f) and let their Fourier series be given by
1j7r
;: -7r f(t)g(t) dt
7.5 Exercises 227
7.16 (a) Compute the Fourier sine series and the Fourier cosine series
of the function f(t) = ton [0,7r].
(b) Evaluate:
7.17 (a) Let f E L2( -7r, 7r). Let its Fourier series be given by
00
7.18 Let f E H6( -7r, 7r). Show that if its Fourier series expansion is
given by
00
ao ~ .
f(t) = 2 + L.,...(an cos nt + bn sm nt),
n=l
then the Fourier series expansion of f' is given by
Remark 7.5.1 Comparing the results of the above exercises with the
comments made in Remark 4.8.1, we deduce that an operator A: D(A) C
H ~ H will be the infinitesimal generator of a co-semigroup of contrac-
tions if, and only if, it is maximal dissipative. Unlike Banach spaces,
where the Hille-Yosida theorem involves verification of infinitely many
conditions, one for each A > 0, this is much easier to verify in Hilbert
spaces. The dissipativity is usually very easy to check. Further, it
is enough to verify that the equation (AI - A)x = f has a solution
x E D(A) for every f E H just for one fixed A > O.
7.23 Let Hi, i = 1,2,3 be Hilbert spaces with norms II· Iii, i = 1,2,3
respectively. Let T : D(T) C HI ~ H2 and S : D(S) C H2 ~ H3 be
closed and densely defined linear transformations. Assume that R(T) c
N(S). Assume further, that there exists a constant C > 0 such that,
7.5 Exercises 229
T*(x) = T*(Px)
1.24 Let II be a Hilbert space and let GL(H) be the set of all invertible
continuous linear operators on H. Then GL(H) is a group with respect
to the binary operation defined via composition of operators. Consider
the unit circle Sl C ]R2 with its usual topology inherited from ]R2. Repre-
senting a point g E Sl as (cos 0, sinO) where 0 E [0,211"), we have that Sl
is a group under the operation defined via (0 1 , ( 2 ) I----t (0 1 + (2)mod(211").
A representation of Sl is a group homomorphism 7i' : Sl -+ GL(H)
which is also continuous. For simplicity, we will denote the image of
g = (cosO, sinO) by 7i'(O).
°
(a) Show that every representation is uniformly bounded, i.e. there
exists a constant C > such that
117i'(O) II < C
Show that < .,. > defines a new inner product on H whose induced
norm is equivalent to the original norm on H.
(c) With respect to the inner-product < .,. > on H, show that 7i'(O)
is a unitary operator on H for every 0 E [0, 211"). (We say that every
representation of Sl is equivalent to a unitary representation.)
1.25 Let V = ef
and let {An} be a sequence of N x N matrices such
that An = ~ for each n. Assume further that, for each v E V, we have
230 7 Hilbert Spaces
that the sequence {(Anv, v)} decreases to zero as n - t 00. Show that
there exists a matrix A = A * such that (A v, v) 2: 0 for all v E V and
such that An -) A in £(V).
Chapter 8
Compact Operators
Definition 8.1.1 Let V and W be Banach spaces and let T E £(V, W).
If dim(R(T)) is finite, then we say that T is of finite rank. If the image
of every bounded set in V is relatively compact in W, we say that T is
compact . •
Example 8.1.5 Let i : C1 [0, 1J ---+ C[O, 1J be the canonical inclusion map.
Let us denote the usual norm in C1 [0, 1J (cf. Exercise 2.5) by 11.111,00 and
the usual norm in C[O, 1J by 11.1100. Consider the set of all f E C1[0, 1J
such that 111111,00 ~ c. Then, clearly, Ilflloo ~ c as well. Further, if x
and yare points in [0, 1], we have, by the mean value theorem,
It follows from this that the set of all functions considered here is uni-
formly bounded and equicontinuous in C[O, 1J. Thus, by Ascoli's theo-
rem, the set is relatively compact and so the inclusion map is compact . •
K, sup IK(x,y)l.
(x,Y)E[O, 1] x [o,lJ
T(f)(x) = 10 1
K(x, y)f(y) dy.
Then, it is easy to see that T(f) is continuous and that IIT(f) 1100 ~
K,llflloo for all f E C[O,lJ and so T is a continuous linear operator on
C[O,lJ. Thus, if we look at the set of all f E C[O, 1J such that Ilflloo ~ c,
it follows that the set of all T(f) is uniformly bounded. Further, if x
and yare points in [0, 1], we have
°
Now, since K is uniformly continuous (since [O,lJ x [O,lJ is compact),
given E > 0, there exists 8 > such that, for all t E [0,1]' we have
whenever Ix-YI < 5. This shows that the set of all T(f) where Ilflloo :S c
is equicontinuous and so T is a compact operator. •
Example 8.1.8 Let f E £2(a, b) be given. The map v r---t f01 vf dx is ob-
viously a continuous linear functional on HJ(O, 1). Since, by Poincare's
11
inequality,
(v, wh = v'w' dx
defines an inner product on this space (cf. Example 7.1.4), it follows from
the Riesz representation theorem that there exists a unique U E HJ
(a, b)
11 11
such that
u'v' dx = fv dx
for all v E HJ(O, 1). Further, setting v = U in the above relation and
using Poincare's inequality (cf. Theorem 6.4.6), we get
a(v,w) = 11 o:(x)v'(x)w'(x) dx
for v and w in HJ(O, 1). Then, it is easy to check that a(.,.) defines a
continuous, symmetric and elliptic bilinear form on HJ
(0, 1). Thus, if
234 8 Compact Operators
a(u, v) = 10 1
fv dx
for all v E HJ(O, 1), by virtue of the Lax-Milgram lemma (cf. Corollary
7.3.1). Set u = G(f). Again, it is easy to check that
10 1
-(a(x)u'(x))'v(x) dx = 10 1
f(x)v(x) dx
fin (0, 1)
O.
The case of Example 8.1.8 corresponds to the case when a == 1 and so
the differential operator is - £.x
in the above boundary value problem.
The operator G mapping the data f to the solution u is sometimes called
the Green's operator. •
A(f)(t) = a(t)f(t).
8.1 Basic Properties 235
It is easy to verify that A E £(L2(0, 1)). Let a(to) =I- 0 for some to E
(0,1). Then, there exists a compact interval J = [to - 0:, to + 0:] such
that
1
la(t)1 ~ 2Ia(to)1 > °
for all t E J. Let {fn } be an orthonormal basis for L2(J). Define
fn(t) = {In(t), if t E J
0, if t E [0, l]\J.
Thus the sequence {A(fn )} does not have a convergent subsequence and
so A is not compact. •
Thus, T is compact . •
°
Proof: Let B be the closed unit ball in V. Let K = T(B), which is
compact, by hypothesis. Let E > and let fi E W for i E I, a finite
indexing set, such that
K C UiEIBW(fi; E).
K(f)(t) = 10 1
k(t, s)f(s) ds.
k
i,j=1
If we set
n
kn(t,s) = L (k,4>ij)4>ij(t,S),
i,j=1
then Ilk - kn l1 2 ---+ 0. Consequently, if we define
Kn(f)(t) = 10 1 kn(t,s)f(s) ds
for f E L2(0, 1), we have that IIKn - KII :S Ilkn - kl12 ---+ 0. Clearly, the
image of K n is contained in the span of {<PI,' .. , <Pn} and so K n is of
finite rank. Hence K is a compact operator on L2(0, 1). (The operator
K is called the Hilbert-Schmidt operator induced by k.) •
Proposition 8.1.3 Let V and W be Banach spaces and let T E £(V, W).
Then T is compact if, and only if, Tot : W* ---+ V* is compact.
Proof: Assume that T is compact. Set K = T(Bv), where Bv is the
closed unit ball in V. Then, K is compact and Ilwll :S IITII for all wE K.
Now, let {v n } be a sequence in Bw*. Consider the sequence {l.{Jn} defined
by I.{Jn (w) = Vn (w) for w E K. Then, this is a sequence of continuous
functions on the compact metric space K. We havell.{Jn(w)1 :S liT II for all
wE K. Further Il.{Jn(wd-l.{Jn(W2)1 :S IlwI-W211 and so the sequence {l.{Jn}
is uniformly bounded and equicontinuous in C(K). By Ascoli's theorem,
it has a uniformly convergent subsequence {I.{Jnlj. Thus, in particular
can be made arbitrarily small for large k and l, i.e. the sequence
{T*(vnkn is a Cauchy sequence and is hence convergent. This shows
that T* is compact.
Now assume that T* is compact. Then, by the preceding arguments,
it follows that T** : V** --+ W** is compact. Let Jv : V --+ V** and
Jw : W --+ W** be the canonical imbeddings. Let x E Bv. Then,
for any v E W*, we have T**(Jv(x))(v) = Jv(x)(T*(v)) = T*(v)(x) =
v(T(x)) = Jw(T(x))(v). Thus T**(Jv(Bv)) = Jw(T(Bv)). Conse-
quently, Jw (T( Bv)) is relatively compact in W** and since Jw is an
isometric isomorphism of W onto Jw(W), it follows that T(Bv) is rel-
atively compact in W. Thus, T is compact and the proof is complete .
•
8.2 Riesz-FredhOlm Theory
Proof: (a) Let Bl be the closed unit ball in N(I - T). If x E Bl, then
Ilxll :::;
1 and, further, x = T(x). Thus, Bl c T(B), where B is the
closed unit ball in V. Since T is compact, it follows that Bl is compact
and so N(I - T) is finite dimensional (cf. Proposition 2.3.6).
In = (Un - v n ) - T(u n - vn )·
We now claim that {un - Vn } is a bounded sequence in V. If not, for a
subsequence, we have Ilu nk - v nk II --+ 00. Set
1
wnk = II u nk ._ v nk II (unk - v nk )·
Thus, on one hand z E N(I - T), while on the other hand, the above
considerations imply that d(z,N(1 - T)) = 1, which is a contradiction.
This establishes the claim.
Since {un - V n } is a bounded sequence, and since T is compact, we
have that, for a subsequence, T(u nk - vnk ) --+ 9 E V. Thus u nk - vnk --+
1+ 9 = h, say. It then follows that T(h) = T(f) + T(g) = g. This
implies that
h-T(h) = I
i.e. R(1 - T) is closed. Then, by virtue of Theorem 4.7.1, we deduce
that
R(I - T) = N(I - T*)-L.
(c) Assume that N (1 - T) = {O}. Assume further that VI = R( 1 - T) =1=
V. Since VI is a closed subspace of V, it is also a Banach space. Further,
if x E V, we have
Thus,
T( Urn - un) = Urn - w
where w E Vrn + l . Thus, if m i= n, by construction,
1
IIT(urn) - T(un)11 ~ 2".
Hence, {T( un)} cannot have a convergent subsequence while {un} is
bounded, contradicting the fact that T is compact. Thus, it follows that
R(I - T) = V.
Conversely, if R(1 - T) = V, then N(I - T*) = R(I - T).L = {O}.
Since T* is also compact, it now follows that R(I - T*) = V*. Then,
again, N(1 - T) = R(I - T*).L = {O}.
S = T+AoP'
8.2 Riesz-Fredholm Theory 241
o = (u - T(u)) + A(P(u))
where the first term belongs to R(I - T) while the second belongs to
its complement, W. Thus u - T(u) = 0 and so u E N(I - T). Thus,
P(u) = u. Further, 0 = A(P(u)) = A(u) and so u = 0 since A is
injective. Thus N(I - S) = {O} and so R(I - S) = V.
Now choose I E W\R(A). Let u E V be such that u - S(u) = I,
i.e. (u - T(u)) - A(P(u)) = I. Since both I and A(P(u)) belong to
W, which is the complement of R(I - T), it follows that u - T(u) = 0
which implies that I = -A(P(u)) E R(A), a contradiction. Thus,
d* < d.
Similarly,
Remark 8.2.1 The conclusion (c) holds for all operators on finite di-
mensional spaces and is referred to in short as 'one-one if, and only if,
onto'. This property carries over to compact perturbations of the iden-
tity in infinite dimensional Banach spaces. In general it is not true for
a general operator on a Banach space. If x = (Xi) E £2 and we define
T E £(£2) by
T(x) = (O,Xl,X2,"'),
then, clearly T is one-one but not onto. •
for every v E V.
A(v, v) ~ o:llv"~.
8.2 Riesz-Fredholm Theory 243
z - (3G(z) = j, (8.2.3)
where z = j + (3u. Thus, (8.2.2) has a unique solution for given j if,
and only if, the same is true for (8.2.3).
Now, I - (3G is a compact perturbation of the identity and so, it is
surjective if, and only if, it is injective. Assume that (I - (3G)w = O.
Thus, w = G((3w), and thus, w E V and, for every v E V, we have
Example 8.2.1 Let V = H1(0, 1) and H = L 2 (0, 1). Let k > o. Set
1
a(u, v) = 10 u'(x)v'(x) dx + k[u(l)v(l) + u(O)v(O)],
is the norm in HI (0, 1). Thus, all the hypotheses of the preceding propo-
sition are satisfied and so, given f E L2 (0, 1), we have a unique solution
u E Hl(O, 1) satisfying
In this section, we will assume that all Banach spaces are complex.
then it follows that T - S is also invertible. We deduce from this that the
set of all invertible operators is open in £(V). In particular, if A E p(T),
then A + <5 E p(T) as well, if <5 is sufficiently small. Thus, it follows that
p(T) is an open subset of C and so the spectrum is a closed subset of C.
Now assume that IAI > IITII. Then
T - AI = -A(I - A-IT)
<p(A) = f(T(A))
246 8 Compact Operators
and, indeed, S(x) = AX. Further, S(el) = O. Thus, it follows that every
A such that IAI < 1 is an eigenvalue of S and so is in the spectrum of S.
Since the spectrum is closed, it now follows that
• O"(T) = {a},
• O"(T) \ {a} is finite,
Proof: (a) We have already seen (cf. Example 8.1.4) that a compact
operator is not invertible. Thus a E O"(T).
(b) Let A i= a. Now T-AI = -A(1 _A-IT) and the Riesz-F'redholm the-
ory applies. Thus, if N (I - A-IT) = {O}, then T - A-I I is onto as well.
Then, by the open mapping theorem, it is also invertible. Thus, if A i= a
is in the spectrum of T, it follows that N(T - AI) = N(1 - A-IT) i= {O}
and that it is finite dimensional (cf. Theorem 8.2.1).
n
Un+l LajUj.
j=I
8.3 Spectrum of a Compact Operator 249
Then,
n n
L O jA n+1 U j LCXjAjUj.
j=l j=l
Since the set {U1,'" ,Un} is linearly independent and since Aj =1= An+1
for all such j, we deduce that CXj = 0 for 1 ::; j ::; n, whence Un+1 = 0,
which is a contradiction. This completes the induction and proves the
linear independence of the set of eigenvectors {un}.
Set Vn = span {U1, ... , un}. Then we have an increasing sequence of
subspaces such that Vn is a proper subspace of Vn+1 for all n. These
subspaces are all finite dimensional and hence, closed. By Riesz' lemma
we can find Vn E Vn such that Ilvnll = 1 and d(vn, Vn-1) ~ 1/2 for all
n ~ 2.
Let An - t A =1= O. Then tL v n } is a bounded sequence and so, since
T is compact, the sequence
Thus,
which contradicts the fact that the sequence {(l/A n )T(vn )} has a con-
vergent subsequence. This proves that A = O.
250 8 Compact Operators
(d) In view of (c) we deduce that, for all positive integrs n, the set
where x = (Xl, X2, ... ), and where {an} is a sequence of complex num-
bers. We saw that if an ---+ 0, then T is compact. Conversely, if T
is compact, it follows that an ---+ 0. Indeed, since an are all in the
spectrum, by Theorem 8.3.1 (c), we deduce that an ---+ 0. •
IIMu -
1
T(u)11 ~ C(Mu - T(u), U)2
Let {un} be a sequence in H such that Ilunll = 1 for all n and such
that (T(u n ), un) - t M as n - t 00. Then, it follows that MU n -T(u n ) - t
O. This proves that M E aCT). (If not, then M! - T would be invertible
and we would have
for n E N.
The spaces En, n 2::': 0 are all mutually orthogonal. To see this, if
u E En and v E Em, where n i=- m, we have
An(U,V) = (T(u), v) = (u,T(v)) = Am(U,V)
254 8 Compact Operators
Fl.. c N(T) c F
~. . Fl.. = {O}.
Now, if Eo is a non-zero subspace, it is either finite dimensional or
is an infinite dimensional separable subspace of H and so it has an or-
thonormal basis. Each En, n 2: 1, is finite dimensional and admits an
orthonormal basis. The union of all these bases forms an orthonormal
basis of H, since F is dense in H (cf. Corollary 7.4.1) . •
(T(v),v)
m xv~o; ..v~ l JiViT2""
(T(v),v)
minVCH; dim(V)=m-l m xv~o; vl.V JiViT2"".
Proof: Since ~ is an eigenvector corresponding to the eigenvalue .~,
and since Il ~1I = 1, it follows immediately that
(8.4.1)
and so
T(v) = I:>.t(v,ut)ut + I:>.;;(v,u;;)u;;.
k n
(T(v), v)
>.+ = max (8.4.2)
m v~O;vl. m l IIvll 2
Finally, let V be a (m - 1)-dimensional subspace of H. Then, there
exists a non-zero vector u in the m-dimensional space v;t such that
u ~ V (why?). Thus
(T(v), v)
sup
V#Ojvl.V IIvl1 2
In fact, since T is compact, the 'sup' above is a 'max' (why?). Further,
in view of (8.4.2), we deduce that
G('xu) = u.
(G(u), u)
All = /11 = max
~O; EL2(O,1) Il ll~
Notice that the equation (8.4.4) is the 'weak formulation' (cf. Re-
mark 8.1.2) of the following two-point boundary value problem:
_ -sL
dx
(o:du)
dx AU in (0,1)
u(O) = u(l) 0.
Remark 8.4.3 Notice that, in the preceding example, all the eigenvec-
tors will, in fact, belong to HJ(O, 1). Using arguments very similar to
those in the proof of Theorem 8.4.2, it can also be shown that if the
eigenvalues {An} are numbered in increasing order of magnitude (tak-
ing into account the geometric multiplicities), then they can be given
the following variational characterization in HJ(O, 1). Let {w n } be the
orthonormal basis of eigenvectors (in L2(0, 1)) with Wn corresponding
to An for each n. Let Vn be the finite dimensional space spanned by
{WI,"" W n }. Set
a(v,v)
R(v)
TvTIr
for v E HJ(O, 1), v i= O. Then
minv~o; ~ n l R(v)
In particular, we have
Al = min R(v).
v;fo; vEHJ(O,I)
~ AU in (0,1)
u(O) u(l) = O.
A simple computation yields that the only possible solutions to this
equation are given by
2 . f01 l'\7vl 2 dx
Al = 7r = mIn "-"-'1--'---
40; VEHJ(O,l) fo Ivl 2 dx
Thus, for every v E HJ(O, 1), we deduce that
8.5 Exercises
8.2 (a) Let V and W be Banach spaces and let V be reflexive. Let T E
£(V, W). Assume that whenever Xn -->. X in V, we have T(x n ) -+ T(x)
in W. Show that T is compact.
(b) Show, by means of an example, that the conclusion is not true, in
general, if we drop the assumption that V is reflexive.
8.3 (a) Let V be a reflexive Banach space and let T E £(V,RI). Show
that T is compact.
(b) Let W be a reflexive Banach space and let T E £(CO, W). Show that
T is compact.
8.5 Let V be a Banach space and let T E £(H) be compact. Show that
if 8 E £(V), then 8(1 - T) = 1 if, and only if, (1 - T)8 = 1. In this
case, deduce that 1 - (1 - T)-I is compact.
IITII = max
IlxlI=I
XEH;
IIT(x)ll·
8.9 (a) Let W be a Banach space. Assume that there exists a sequence
of operators of finite rank, {Pn }, in £(W) such that Pn(Y) ~ Y for all
YEW. Show that, if V is any Banach space, and if T E £(V, W) is
compact, then T is the limit (in £(V, W)) of operators of finite rank.
(b) Deduce that, if V is any Banach space, and if T E £(V,£p), where
1 :S: p < 00, is compact, then T is the limit of operators of finite rank.
8.11 Let L 2 (0, 1) be the space of complex valued square integrable func-
tions. Let K be the Hilbert-Schmidt operator (cf. Example 8.1.11)
defined on this space by the function ik where
I if s < t
k(t,s) = { -1 if s;: t.
A - 2 kEZ
k - (2k + 1)7r'
with corresponding eigenvector t ~ exp(i(2k + l)7rt).
8.12 Consider the operator K E £(L2(0, 1)) defined by
8.15 (a) Let V be a Banach space and let T E £(V) be compact. Let
A 1= O. Show that the results of the previous exercise are valid when
I - T is replaced by AI - T. (If A is an eigenvalue of T, then N(AI - T)
is called the eigenspace associated to A while the space Nl obtained in
section (b) of the preceding exercise is called the generalized eigenspace
associated to A and is denoted N (A). The corresponding complement Fl
is denoted F(A).
(b) Show that if A and fJ, are disctinct non-zero eigenvalues of T, then
N(fJ,) C F(A).
8.16 (a) Let H be a Hilbert space and let Sand T E £(H). Let A =1= 0 be
an eigenvalue of ST with eigenvector u. Show that T(u) is an eigenvector
ofTS.
(b) Show that the non-zero eigenvalues of ST and T S are the same and
that
T(N(ST - AI)) = N((TS - AI))
262 8 Compact Operators
8.18 Let H be a Hilbert space and let T E £(H) be compact and self-
adjoint. Show that either IITII or -IITII is an eigenvalue of T.
8.20 Let V and H be real Hilbert spaces with inner products (., .)v and
(., ')H respectively; let the corresponding norms be denoted by 11.llv and
II.IIH. Assume that V c H, the inclusion being continuous and compact.
Assume, further, that V is a dense subspace of H. Let a (., .) : V X V ~ lR
be a continuous, symmetric and coercive bilinear form.
(a) Let f E H. Show that there exists a unique vector u E V such that
for every v E V.
(b) Define G(f) = u. Show that G E £(H) and that it is self-adjoint
and compact. Show, further, that
for every u E H, u 1= o.
(c) Show that the image of G is dense in V.
(d) Show that there exists a sequence of positive real numbers {An} and
an orthonormal basis {un} of H such that each Un E V and
for every v E V.
(e) Set Vn = (l/A)u n . Show that {v n } forms an orthonormal basis of
V (for the inner product induced on V by the bilinear form a(., .)).
(f) Show that
, . a(v,v)
"'I = mIn 2 .
vEV; #0 IlvllH
8.5 Exercises 263
(g) If U E V is such that Il ll~ = 1 and a(u, u) = A1, then show that
a(u,v) = A1(U,V)H
for every v E V.
(h) Let Vn be the space spanned by {u 1, u2, ... , un}. Set
R( ) = a(v,v)
v Ilvll~
for v E V, v I- O. Let
°
for every v E V.
(a) Show that there exists a constant C > such that
J'(O)(h) = (p + 0, h)H
where p E V satisfies
for every v E V.
(e) Deduce that if 0* minimizes J over K, then
266
267
rank, 5 finite, 19
locally convex space, 81 Lebesgue, 19
locally integrable, 173 product, 24
lower semi-continuous, 135 regular, 20
space, 19
matrix, 5 translation invariant, 19
adjoint, 6 metric space, 12
characteristic polynomial, 9 Minkowski functional, 78
condition number, 61 multiplicity
definite, 7 algebraic, 9, 261
determinant, 8 geometric, 9, 246
diagonalizable, 6, 11
eigenspace, 9 neighbourhood, 11
eigenvalue, 9 norm, 26,38
eigenvector, 9 norm topology, 27
hermitian, 7 normed linear space, 27
lower triangular, 6 completion, 65
non-singular, 8 now here dense set, 12
norm, 50 null space, 5
normal, 7
nullity, 7 open sets, 11
rank, 7 orthogonal vectors, 6
self-adjoint, 7 orthonormal
singular, 8 basis, 218
spectral radius, 10 set, 210
spectrum, 10
parallellogram identity, 197
trace, 10
Parseval's identity, 218
transpose, 6
path connected, 17
unitary, 7
product measure, 24
upper triangular, 6
product topology, 15
measurable
projection, 200
function, 18
orthogonal, 201
Lebesgue, 20
sets, 18 quotient space, 35
space, 18
measure, 19 Rayleigh quotient, 10
O"-finite, 19 reflexive space, 73, 145
complete, 19 resolvent, 244
counting, 19 right inverse, 115
Dirac, 19 Rodrigues' formula, 226
269