Banach Spaces

J Muscat 2005-12-23
(A revised and expanded version of these notes are now published by
Springer.)

Banach Spaces

Definition A normed vector space X is a vector space over R or C with

a function called the norm k.k : X R such that,
kx + yk 6 kxk + kyk,
kxk = ||kxk,
kxk > 0, kxk = 0 x = 0.

1.0.1

Easy Consequences
kx yk > kxk kyk,

k xk = kxk,

kx1 + . . . + xn k 6 kx1 k + . . . + kxn k.

1.0.2

Examples

1. The set of real numbers with the norm taken to be the absolute value.
P
2 1/2
2. RN or CN with the norm defined by
kxk
=
(
. There are
2
i |xi | )
P
other
however, kxk
= i |xi | or kxk = max
|. Thus
 possibilities

 1
i |xi

3
3
3
k
k = 3 + 4 = 7, k
k = 9 + 16 = 5, k
k =
4 1
4 2
4
max(3, 4) = 4.
3. Sequences: sequences can be added and multiplied by scalars, and form
a vector space. But there are different ways of taking the length of a
sequence:
P
1
The
space

=
{
(a
)
:
n
n |an | 6 } with norm defined by k(an )k1 =
P
n |an |.
1

2 = { (an ) :

|an |2 6 } with norm defined by k(an )k2 =

pP

|an |2 .

= { (an ) : |an | 6 C } with norm defined by k(an )k = supn |an |.

4. Functions: functions also form a vector space, and different norms can
be defined for them as with sequences
R
The space LR1 = { f : A R : A |f (x)| dx < }, with norm defined
by kf kL1 = A |f (x)|.
R
The space L2 = { f : A R : A |f (x)|2 dx < }, with norm defined
qR
by kf kL2 =
|f (x)|2 dx.
A
The space L = { f : A R : f is measurable, and |f (x)| 6 C a.e. },
with norm defined by kf kL = supx a.e. |f (x)| (i.e. the smallest c such
that |f (x)| 6 c a.e.
The space C = { f : [a, b] R : f is continuous } with kf kC =
kf kL = maxx[a,b] |f (x)|.

Proposition 1.1 A vector space is normed if and only if it has a

translation invariant and scalar-homogeneous distance,
d(x, y) = kx yk.
Proof: Given a norm, the function d(x, y) = kx yk is translation invariant in the sense that d(x + a, y + a) = kx + a y ak = d(x, y), scalar
homogeneous in the sense that d(x, y) = ||d(x, y). Conversely if d(x, y)
is translation invariant, then d(x, y) = d(x y, y y) = d(x y, 0) and we
can define kxk = d(x, 0) so that d(x, y) = kx yk. We will now show that
the axioms for the norm correspond precisely to the axioms for a distance.
d satisfies the triangle inequality the norm k.k does.
d(x, y) 6 d(x, z) + d(z, y)
corresponds to kx yk 6 kx zk + kz yk. Similarly d(x, y) = d(y, x)
corresponds to kx yk = ky xk; d(x, y) > 0 is kx yk > 0, while d(x, y) =
0 x = y becomes kx yk = 0 x y = 0.
The scale-homogeneity of the metric supplies the final axiom for the norm.

This invariance under translations and scaling has the following easy consequences
2

Proposition 1.2
Br (x) = rB1 (x/r) = x + Br (0) = x + rB1 (0)
Br (x) = { y : d(x, y) 6 r }
Proof: The first statement is the set equivalent of
d(x, y) < r d(x/r, y/r) < 1 d(0, y x) < r d(0, (y x)/r) < 1
Let a be a limit point of Br (x), so that we can find y Br (x) such that
ky ak < , which implies that ka xk 6 ka yk + ky xk < + r. Hence
a { y : d(x, y) 6 r } as required.
Conversely, suppose d(x, y) 6 r; let a = x + (y x) with 1 /r <
|lambda| < 1. Then ka xk = ||ky xk < r, and ka yk = |1
|kx yk < .


1.0.3

Note

Normed vector spaces are therefore metric spaces, as well as vector spaces.
So we can apply ideas related to both, in particular open/closed sets, limit
points, convergence of sequences, completeness, continuity, compactness, as
well as linear subspaces, linear independence and spanning sets, linear transformations, kernels etc.
Since, at the end of the day, we are interested in issues like convergence
and so on, we say that two norms are equivalent when they induce the same
open sets.
For example, the norms k.k1 , k.k2 and k.k are all equivalent on RN . (But
note that they are not equivalent for sequences or functions! In particular, a
sequence of functions may converge in L1 but not in L or vice-versa.)
1.0.4

Banach Spaces

When the induced metric is complete, the normed vector space is called a
Banach space.
So, a closed linear subspace of a Banach space is itself a Banach space.

1.0.5

Example

Not every norm is complete of course. For example, suppose we take the
vector space of continuous functions f : [0, 1] R with the norm kf k =
R1
|f (x)|. This is indeed a norm but it is not complete, for consider the
0
sequence of continuous functions fn (x) = n(x 21 )[ 1 , 1 +1/n] + ( 1 +1/n,1] ; it is
2 2
2
R1
1
Cauchy since kfn fm k = 0 |fn fm | = 2 |1/n 1/m| 0 as n, m ,
but it does not converge in C[0,R1], because suppose it does i.e. kfn f k 0
1
as n . This means that 0 |fn (x) f (x)| dx 0, which implies that
R1
R1
|f (x)| dx 0, and 1 +1/n |1 f (x)| dx 0, so that f (x) = 0 on [0, 21 )
0
2

and f (x) = 1 on ( 12 , 1], implying it is discontinuous.

1.1

Metric and Vector Properties

Proposition 1.3 The maps x 7 x + y, x 7 x and x 7 kxk are continuous.

Proof: Addition by a vector is continuous,
kx1 x2 k < k(x1 + y) (x2 + y)k = kx1 x2 k < .
Scalar multiplication is continuous,
kx1 x2 k < kx1 x2 k = ||kx1 x2 k < ||.
The norm is continuous,
kx1 x2 k < |kx1 k kx2 k| 6 kx1 x2 k < .
Corollary
lim (xn + yn ) = lim xn + lim yn ,

lim xn = lim xn .

.
Proposition 1.4 If M is a linear subspace of X, then so is M
Proof: Let x, y be limit points of M. Then we can find a sequence xn x
. Similarly
and yn y; but xn + yn M, so x + y = limn (xn + yn ) M

x = limn (xn ) M .

For example, the set of polynomials forms a linear subspace of C[0, 1]; its
closure in this space is a closed linear subspace (and hence complete).
4

1.1.1

Series

P
A series is a sequence of vectors obtained by adding, i.e. N
n=0 xn . A series
PN
therefore converges when kP n=0 xn xk 0 as N . It is said to
converge absolutely when n kxn k converges in R.

We can show that a series is absolutely convergent by using the usual

tests for real series, for example, the comparison, ratio or root tests.

Proposition 1.5 For a Banach space, absolutely convergent series

converge.
P
PN
Proof: Suppose that n kxn k converges. Let yN =
n=0 xn , so that
PM
PM
kyM yN k = k n=N +1 xn k 6 n=N +1 kxn k 0. Hence yP
N is a Cauchy
sequence in X, and so converges to y, say. This means that n xn = y.
It is also true (see the exercises) that if a normed vector space is such that
all its absolutely convergent series converge, then the space is also complete,
i.e. a Banach space.
1.1.2

Quotient Spaces

If M is a closed linear subspace, then we can factor it out and form the space
X/M with addition, scalar multiplication and norm defined by
(x + M) + (y + M) = (x + y) + M,
(x + M) = x + M,
kx + Mk = d(x, M) = inf vM kx vk.
That the addition and scalar multiplication satisfy the axioms of a vector
space is a triviality; let us show that we do indeed get a norm:
k(x + M) + (y + M)k = kx + y + Mk = inf vM kx + y vk
= inf v1 ,v2 M kx + y v1 v2 k
6 inf v1 M kx v1 k + inf v2 M ky v2 k
= kx + Mk + ky + Mk
k(x + M)k = kx + Mk
= inf vM kx vk
= inf uM kx uk
= inf uM ||kx uk
= ||kx + Mk
5

kx + Mk = inf kx vk > 0.
vM

= M x + M = 0 + M.
kx + Mk = 0 inf kx vk = 0 x M
vM

Proposition 1.6 If X is complete, and M is a closed linear subspace,

then X/M is also complete.
P
Proof: let xn + M be an absolutely convergent series i.e.
n kxn + Mk
n
converges. Now n > 0, vn M, kxn vn k 6 kxn + Mk
P + 1/2 , so the left
hand side converges by comparison with the right. So n (xn vn ) = x since
X is complete, and thus
k

N
X
n=1

(xn + M) (x + M)k = k

1.1.3

N
X
n=1

xn x + Mk 6 k

N
X
n=1

(xn vn ) xk 0

Exercises

1. Show that the norms defined for RN and CN are indeed norms.
2. Show that the norms defined for 1 and are indeed norms.
1
3. Show that the norm
R defined for L is indeed a norm, except that it
may happen that A |f (x)| dx = 0 without f = 0; however from the
theory of Lebesgue integration, in this case we get f = 0a.e., so that
the failure of this axiom is not drastic. In fact we can identify those
functions that are equal almost everywhere into equivalence classes and
work with these, but the notation would be too pedantic to be useful.

4. Suppose that k.k1 and k.k2 are two norms on a vector space X, such
that there are positive constants c, d,
kxk1 6 ckxk2 ,

kxk2 6 dkxk1 .

Show that every ball of norm 1 has a smaller ball of norm 2 with the
same centre, and vice-versa. Deduce that the two norms have the same
open sets. Show that the converse is also true.
5. Show that in RN , the three norms k.k1 , k.k2 and k.k defined previously
are all equivalent. What do the unit balls in each norm look like?
6. Show that if n and xn x then n xn x. In Corollary 1.4,
strictly speaking, we can only immediately deduce that limn (xn +
y) = x + y; give a proper proof of the corollary.
6

P
7. Show that if a series n xn converges, then xn 0. (Hint: convergent
sequences are Cauchy)
8. * Let X be a normed vector space for which every absolutely convergent
series converges. Show that X is complete as follows: let xn be a
Cauchy sequence in X, and let yn := xn+1 xnP
; show that there is a
r
subsequence ynr such that kynr k 6 1/2 and so r kynr k 6 1. Deduce
P
that N
r=0 ynr converges, and hence xn .

Continuous Linear Operators

Definition An operator is a linear continuous transformation T : X Y

between normed vector spaces,
T (x + y) = T x + T y,

T (x) = T x.

In particular, a functional is a continuous linear map : X C.

The set of operators is denoted by B(X, Y ). The set of functionals is
called the dual space.
R
P
Examples of functionals on sequences are n n an , on functions k(x)f (x) dx.
In general linear maps are not continuous. For example, let X be the
vector space of P
finite real sequences with the 1 norm. Consider the functional (x) =
n nan (check linearity!), and the sequence of sequences
x1 = (1, 0, . . .), x2 = (0, 21 , 0, . . .), x3 = (0, 0, 1/3, . . .), . . .. Then kxr k1 =
P
1
n |an | = 1/n 0 as n , so that xn 0, yet (xn ) = n n = 1 6 0.
Another important example is differentiation: although the map is linear
(say on the vector space of polynomials) it is not continuous eg D cos(nx) =
n sin(nx) so kD cos(nx)kC = n cannot be bounded by a multiple of k cos(nx)kC =
1.
A simple test for continuity of a linear transformation is the following
bounded property,
Proposition 2.1 A linear transformation is continuous if, and only
if,
c > 0, x X, kT xkY 6 ckxkX
Proof: If T is linear and continuous, then
> 0, kyk < kT yk < 1
In particular, putting y = 21 x/kxk, we get kyk = 21 < , so that kT xk =
|2kxk/|kT yk < 2 kxk.
Conversely, suppose T is bounded, then
kT x T yk = kT (x y)k 6 ckx yk.
So for any > 0, let = /c and
kx yk < kT x T yk <
8

Proposition 2.2 The kernel of an operator is a closed linear subspace.

Proof: If T x = 0 and T y = 0 then T (x + y) = T x + T y = 0 and
T (x) = T x = 0, so that the kernel is a linear subspace. Let x be a limit
point of the kernel i.e. there is a sequence xn x with T xn = 0. Then
T x = T (limn xn ) = limn T xn = 0, and x is in the kernel.

Definition An isomorphism between normed vector spaces is a map
T : X Y such that both T and T 1 are linear and continuous (ie T is a
linear homeomorphism).
An isometry between normed vector spaces is a bijective linear map
T : X Y such that kT xk = kxk

It follows, of course, that isometries are isomorphisms (since continuity

of T follows from the bounded property, and kT 1 xk = kT T 1 xk = kxk as
well).
Theorem 2.3 Every N-dimensional normed vector space is isomorphic to CN , and so is complete.
!
x
1
Proof: Let T : CN X be the standard map T ..
:= x1 e1 +. . . xN eN
.xN
where e1 , . . . , eN form a basis (wolog of unit vectors) for X. That T is linear
and bijective is trivial to show.
T is continuous since
kT xkX = kx1 e1 + . . . xN eN k

6p
|x1 | + . . . + |xN |
P
= q( i |xi |)2
P
=
ij |xi ||xj |
q P
6 12 ij |xi |2 + |xj |2
pP
2
= N
i |xi |
= Nkxk2

Moreover T 1 is continuous, since let f (x) := kT xk, which is a composi

tion of two continuous functions. Now the sphere S = { u : kuk2 = 1 } is a

compact set (since it is closed and bounded in CN ), so f S is also compact.
Suppose 0 f S; this would mean that 0 = kT uk for some u with kuk2 = 1.

But we get that T u = 0, and T is invertible, so u = 0, a contradiction. Hence

0 6 f S, so 0 is in the exterior of f S (since f S is closed), so it is contained
9

in a ball c < 0 < c completely outside f S. This means that c 6 kT uk,

for any unit vector u. Hence in general, ckxk2 6 kT xk for any vector x. By

exercise, we get that T 1 is also continuous.


Theorem 2.4 The Open Mapping Theorem
Every onto continuous linear map T : X Y between Banach
spaces, maps open sets to open sets.
Proof: Let U be an open subset of X, and let T x T U, so that x
B (x) U. We will show that T B1 (0) contains a ball Br (0), from which
follows that
T x Br (T x) = T x + Br (0) T x + T B1 (0) = T B (x) T U
proving that T U is an open set.
Let F = T B1 (0), and suppose that it does not contain any open balls
(i.e. it has no interior points). It follows that F also contains no open balls
for any > 0. Since F is bounded by kT k it is possible to find a ball B1 of
radius less than 1 such that B 1 F = . In general, suppose we have a ball
Bn of radius at most 1/n such that B n nF = . Then Bn 6 (n + 1)F ,
and so Bn (n + 1)F is a non-empty open set, which must contain an open
ball. By reducing its radius if necessary, we get a ball Bn+1 of radius at most
1/(n + 1) and such that Bn+1 (n + 1)F = . In this way, we get a nested
sequence of balls Bn+1 Bn with diminishing radii. Their centers yn form
a Cauchy sequence since kyn ym k 6 max(1/n, 1/m), and so converge to
some y Y . Now, for any m > n, ym Bn , so y Bn for any n.
However y = T x for some x X since T is onto. Choose n > kxk,
to get x nB1 (0), and so y = T x nT B1 (0) nF , which contradicts
y Bn (nF ) . Thus F must contain an open ball B8r (a) = a + B8r (0); by
translating by a F , and scaling, we get Br (0) 14 F = T B1/4 (0).
Claim: Br (0) T B1 (0). Let y Br (0) T B1/4 (0), so that there must
be an x1 B1/4 (0) such that ky T x1 k < r/2; that is, kx1 k < 1/4 and
y T x1 Br/2 (0) T B1/8 (0). But this implies that there is an x2 B1/8 (0)
such that ky T x1 T x2 k < r/4. We therefore get a sequence xn such that
kxn k <

1
2n+1

r
2n
P

ky T (x1 + . . . + xn )k <

P
1
We conclude that x := n xn converges, with kxk 6 n=1 2n+1
= 12 , and
that y = T x. Thus x B1 (0) and y = T x T B1 (0), proving the claim.

10

Corollary If T : X Y is in addition also 1-1, then T 1 is

continuous and X is isomorphic to Y .
Theorem 2.5 If Y is complete, then B(X, Y ) is a Banach space,
with the norm defined by
kT k = sup
x6=0

kT xkY
.
kxkX

In particular X is a Banach space, with norm defined by

kk = sup
x6=0

|(x)|
.
kxk

Proof: That B(X, Y ) is a vector space is a triviality. The norm is welldefined in the sense that if T is an operator, then kT xk 6 ckxk and kT k 6 c.
Thus,
kT xk 6 kT kkxk.
kS + T k = supx6=0

kSx+T xk
kxk

kT k = supx6=0

6 supx6=0 kSxk
+ supx6=0
kxk
= kSk + kT k

kT xk
kxk

kT xk
kxk

= ||kT k

kT k = 0 kT xk = 0, x = 0,
T =0

The resulting normed vector space is complete. Let Tn be a Cauchy

sequence in B(X, Y ), that is, kTn Tm k 0 as n, m . So
kTn (x) Tm (x)k 6 kTn Tm kkxk 0
This means that Tn (x) is a Cauchy sequence in Y , which is complete, so that
Tn (x) T (x) as n for each x X. We now show that T is linear and

11

continuous:

T (x + y) = limn Tn (x + y)
= limn (Tn (x) + Tn (y))
= T (x) + T (y)
T (x) = limn Tn (x)
= limn Tn (x)
= T (x)
kT (x)k 6 kTn (x)k + kT (x) Tn (x)k
6 kTn kkxk + kxk
6 ckxk

since Tn is a Cauchy sequence and so is bounded.

Finally Tn T since k(Tn T )xk 6 kTn Tm kkxk + kTm x T xk 6
kxk + kxk, so that kTn T k 6 2 and kTn T k 0 as n .

Proposition 2.6 If S and T are compatible operators, then so is ST ,
with kST k 6 kSkkT k.
Proof: That ST is linear is obvious, ST (x+y) = S(T x+T y) = ST x+ST y
and ST (x) = S(T x) = ST x. Also,
kST xk 6 kSkkT xk 6 kSkkT kkxk
and taking the supremum on both sides gives kST k 6 kSkkT k.
Corollary B(X) := B(X, X) is closed under multiplication (called
a Banach algebra).
2.0.4

Examples

The right-shift operator R : 1 1 is defined by R(x0 , x1 , . . .) = (0, x0 , x1 , . . .).

It is obviously linear, and also bounded since
kRxk = k(0, x0 , x1 , . . .)k1 =

X
n=0

|xn | = kxk1

The multiplication operator T : 1 1 defined by T (x0 , x1 , . . .) =

(0 x0 , 1 x1 , . . .), where (n ) is a bounded sequence, is also linear, and bounded
since

X
X
kT xk =
|n xn | 6 M
|xn | = Mkxk
n=0

n=0

12

The Fourier transform on a function f L1 [0, 1] is defined to be the

sequence of numbers (f) where
f(n) = F f (n) =

e2inx f (x) dx
0

Similarly the Fourier transform on a function f L1 (R) is defined to be the

function
Z

f () = F f () =
e2ix f (x) dx
R

It is easy to show that the Fourier transform is linear; moreover it is

continuous with F : L1 [0, 1] and F : L1 (R) L (R), since
Z 1
Z 1
2inx
kfk = sup |
e
f (x) dx| 6
|f (x)| dx = kf kL1 [0,1]
n

and similarly for the second case. In fact, one can show further that F :
L1 [0, 1] c0 and F : L1 (R) C0 (R).
R
More generally, any transformation T f () = k(x, )f (x)dx where k(x, )
is a bounded function, is an operator.

2.1

The Dual Space

Functionals are simply row vectors when X = RN ; thus X is isomorphic to

RN and is generated by a dual basis.
P
Examples: in 1 , weRhave already seen that n n xn is a functional when
n are bounded. Also, f (x)dx is a functional on L1 . Another example, for
C, is given by 0 (f ) = f (0), which has norm k0 k = 1.
There is a certain duality between X and its dual space X that we will
explore below. The connection between the two is the following: Definition
The annihilator of a set of vectors A is the set of functionals A = {
X : (x) = 0, x A }.
2.1.1

The Hahn-Banach Theorem

Theorem 2.7 Let Y be a subspace of a normed vector space, X.

Then every functional Y can be extended to X , that is

= kk .
(y)
= (y), y Y and kk
X
Y
Proof: Let us try to extend from a functional on Y to a functional on
+ v) = (y) + c
Y + hvi for any vector v. The only possibility is to let (y
13


(where c would be equal to (v)).
Whatever c, this gives a linear map (easy
to check). For certain values of c, this is also continuous:
+ v)| = |(y) + c|
|(y
= |||(y/) + c|
We thus want |(y/) + c| 6 kkky/ + ck, which is equivalent to finding
a c such that
(y) kkky + vk 6 c 6 (y) + ky + vk,

y Y.

Is this possible? Yes because |(y1 y2 )| 6 kkky1 y2 k 6 kk(ky1 + vk+

ky2 + vk), so we get (when is real-valued)
(y1 ) (y2) 6 kkky1 + vk + kkky2 + vk
i.e.
(y2 ) kkky2 + vk 6 (y1 ) + kkky1 + vk,

y1 , y2

Hence there must be a constant c separating the two sides of the inequality.
6 kk; in fact, equality holds because for y Y , the supremum of
Thus kk

|(y)|/kyk = |(y)|/kyk is kk. A different proof (see reference book) shows

that this is also the case when is complex-valued.
If X can be generated by a countable number of vectors vn (X is called
separable) then we can keep on extending until is a functional on X. But
even if X needs an uncountable number of generating vectors, then can apply
Zorns lemma to conclude that the extension goes through to X (see book).

Corollary
x = 0 X , (x) = 0
Proof: Given x 6= 0, form the one-dimensional subspace Y = [[x]] = { x :
C }. Let (x) = kxk, then Y since
|(x)| = ||kxk = kxk
in fact kk = 1. So can be extended to a functional on all of X. Hence
there is a = X such that (x) = kxk =
6 0.
The converse statement is trivial to prove.


14

Corollary
kxk = sup
6=0

|(x)|
kk

6 kxk. We are left to show

Proof: Since |(x)| 6 kkkxk we get that |(x)|
kk
that the left hand expression approximates the right as closely as necessary.
By the theorem, there is a X with (x) = kxk and kk = 1, so for this
functional, |(x)|/kk = kxk.

2.1.2

Proposition 2.8 Every normed vector space X is embedded in its

double dual X . That is, there is a 1-1 linear isometry J : X X .
Proof: Given x X, let x : X C be the map x () = (x). Then
x X , ie.it is linear and bounded with kx k = kxk. Hence we can form
the map J : X X by Jx = x .
J is linear since for any X ,
(x + y) () = (x + y) = (x) + (y) = x () + y ().
(x) () = (x) = (x) = x ().
J is isometric (preserves the norm) since
kx ()k = |(x)| 6 kkkxk,
and there is a functional such that (x) = kxk and kk = 1, so that
kx k = sup
6=0

x ()
= kxk.
kk

That J is 1-1 follows from this isometry, since x = y (x y) =

0 k(x y) k = 0 kx yk = 0 x = y.

2.1.3

Completion of X

Given any normed vector space X, the double dual X is a Banach space.
is a closed linear subspace of X , and so is itself
Hence the closure JX
complete i.e. a Banach space. It is called the completion of X, usually

denoted X.
15

2.1.4

Weak convergence

We already know what Tn T means, namely, kTn T k 0. This of

course implies pointwise convergence ie x, Tn x T x since for each x,
kTn x T xk = k(Tn T )xk 6 kTn T kkxk 0; but the converse is false ie
it is possible to have pointwise convergence without Tn T . For this reason
Tn T is sometimes called convergence in norm.
Example: Let Tn : 1 1 be defined by Tn (xi ) = (0, . . . , 0, xn , xn+1 , . . .)
(well-defined, linear and continuous); we have pointwise convergence, Tn x

0 since for each x = (x

i ), kTn xk1 = k(0, . . . , 0, xn , xn+1 , . . .)k1 = |xn | +
P

|xn+1 | + . . . 0 since ni=1 |xi | converges. However Tn 6 0 since we can

always find sequences x = (0, . . . , 0, 1, 0, . . .) for which kTn xk = 1 = kxk so

that kTn k = 1 6 0.
There is yet another type of convergence, called weak convergence,
defined as follows Tn T x X, X , Tn x 0 as n .
Pointwise convergence implies weak convergence because Tn x T x and is
continuous. However the converse is again false (see exercise in Chapter 3).
Proposition 2.9 In finite dimensions, all these convergence types are
equivalent.
Proof: Let An A where An , A are matrices. This means that y (An
A)x 0 as n . In particular if we let y = ei , x = ej be basis vectors

component
then An,ij = ei An ej ei Aej = Aij , so that each
of the matrices


An convergesq
to the corresponding component in A. This then implies that
P
2
kAn Ak 6
ij |An,ij Aij | 0.


2.2

The Adjoint T

Definition The adjoint (or transpose) of a linear transformation T : X

Y is T : Y X defined by T (x) = (T x) for any Y .
It is well defined because

kT (x)k = k(T x)k 6 kkkT xk 6 kkkT kkxk

so that T X . It is also linear since T ( + )(x) = ( + )(T x) =
(T x) + (T x) = T (x) + T (x) and T ()(x) = (T x) = T (x).
If T is continuous and linear, then so is its adjoint, since by the above
inequalities, kT k 6 kT kkk; in fact kT k 6 kT k.
16

Proposition 2.10
(S + T ) = S + T , (T ) = T , kT k = kT k.
Proof: (S + T ) (x) = (Sx + T x) = (Sx) + (T x) = S (x) + T (x);
since this is true for all x and all , the result follows.
Similarly, (T ) (x) = (T x) = (T x) = T (x).
kT k = sup
x6=0

= sup sup
6=0 x6=0

|(T x)|
kT xk
= sup sup
kxk
x6=0 6=0 kxkkk

|(T x)|
|T (x)|
kT k
= sup sup
= sup
= kT k
kxkkk
kkkxk
kk
6=0 x6=0
6=0


Proposition 2.11
ker T = (im T ) ,

im T (ker T ) .

Proof: T = 0 x, T (x) = 0 x, (T x) = 0 im T = 0
(im T ) .
Let im T , ie = T . Then for any x ker T , we get that
(x) = T (x) = (T x) = 0.

Proposition 2.12 T has a continuous inverse T and T are onto.

Proof: T is onto T is 1-1, which implies that T is invertible and

continuous by the open mapping theorem.


2.2.1

Exercises

1. Let L : be defined by L(xn ) = (xn+1 ). Show that L is linear

and continuous.
2. For T invertible and linear, show that T 1 is continuous ckxkX 6
kT xkY , for some c > 0.
17

3. Show that if T and T 1 are operators, then kT 1 k > kT k1 .

4. Let T : 1 1 be defined by T (xn ) = (xn /n). Show that T is
linear and continuous, and that T 1 is linear but not continuous. Show
however that there are no sequences x such that T x = 0.
5. Show that (A + B) = A B , and A + B (A B) .
6. The Hahn-Banach theorem is trivial when Y is dense in X ie Y = X,

as follows: for x = limn yn , define (x)

:= limn (yn ). Show that

= kk.
is well-defined, linear and continuous with kk
R
7. Show that integration, f 7 f , is a functional in L1 by showing that
it is bounded.
1
8. Define the
R functional on the vector space of step functions in L , by
(f ) = f . Use the Hahn-Banach theorem to show that this extends
to a linear and continuous functional on L1 .
qP
2
9. * Show that, for X and Y finite dimensional, then kT k =
ij |Tij | .

10. A set is called weakly bounded when X , A is bounded. Show

that a set is weakly bounded if, and only if, it is bounded. (Hint: use
the fact that if x () 6 c , , then kx k 6 c - called the uniform
bounded principle)
11. In the embedding of X in X , show that T : X Y is an
extension of T : X Y in the sense that T x = (T x) .
12. * The corollary to the Hahn-Banach theorem, which states that there is
a functional X such that (x) 6= 0 can be strengthened as follows.
Let M be any closed linear subspace, and let x 6 M. Let Y = M [[x]],
and define (y) = 0 for y M and (x) = d(x, M). Show that is
a functional on Y , and hence that it can be extended to X, while still
having the properties that (x) 6= 0 and (y) = 0 for y M.

18

Sequence Spaces

3.1

The space 1

Proposition 3.1 1 is complete

1
P Proof: Let x1 + x2 + . . . be an absolutely summable
P series in , ie
kxn k1 , and so for each i, n |xni | converges
n kxn k1 = s. Thus |xni | 6 P
in R by P
comparison. Let yi = n xin .
y = n xn since
P
P
PN
ky N
i
n xn k1 = Pi=1 |yP
n xni |

=
|
i=1
n=N +1 xni |

But

K
X
i=1

as N .
y 1 since

M
X

n=N +1

xni | 6

M
K
X
X

i=1 n=N +1

|xni | 6

n=N +1

kxn k1 0

K
N
K X
N
N
X
X
X
X
|
xni | 6
|xni | 6
kxn k1 6 s
i=1

n=1

i=1 n=1

n=1


Proposition 3.2 Every functional on 1 is of the type (xn ) 7
where (yn ) ,

: 1 C defined by
Proof: Let y ; then the map y
X
(x) =
y
yn xn
n

is well-defined since
|
y (x)| 6

X
n

kyk|xn | 6 kyk kxk1

; moreover it is a functional in 1 since

X
X
X
(a + b) =
(a) + y
(b),
y
yn (an + bn ) =
yn an +
yn bn = y
n

19

yn xn

 (x) =
y

yn (xn ) =

yn xn =
y (x),

|
y (x)| 6 kyk kxk1

Note it follows that k

y k 6 kyk .

Every functional is of this type: let P

1 and let yn = (en ) where
en = (ni ) = (0, . . . , 0, 1, 0, . . .). Since x = n xn en , then
X
X
(x) = (
xn en ) =
xn yn
n

by linearity and continuity of ; also |yn | = |(en )| 6 kk so that y ,

with kyk 6 kk.
. The previous proofs have
Now let J : 1 be the map Jy = y
shown that this map is in fact onto 1 .
J is linear, since
X
X
X
vn xn = Ju(x)+Jv(x),
un xn +
x 1 , J(u+v)(x) =
(un +vn )xn =
n

so J(u + v) = Ju + Jv. Similarly

X
J(y)(x) =
(yn )xn = Jy(x),
n

and J(y) = Jy.

J is isometric since by the previous proofs we have k
y k 6 kyk 6 k
y k. It
follows that J is 1-1 as well.

3.2

The space

Proposition 3.3 is complete.

Proof: Let (xn ) be a Cauchy sequence in , ie kxn xm k 0 as
n, m .
For each i, we get that |xni xmi | 6 kxn xm k 0, so (xni xmi ) is a
Cauchy sequence in C, and so converges. Let yi = limi xni .
y since
|xni | 6 kxn k < C
since Cauchy sequences are always bounded.
xn y in since, by choosing m large enough depending on i,
1
|yi xni | 6 |yi xmi | + |xmi xni | 6 + kxm xn k 0,
n
as n , independently of i.

20

Proposition 3.4 The space of convergent sequences,

c = { (xn ) : lim xn = x, x }
n

is a complete subspace.
Proof: c since xn x implies that (xn ) is bounded. Moreover it is
easily shown to be linear.
c is closed in since let y: let y be a limit point of c ie we can find xn c
such that xn y. Now consider the limits of each of these sequences xn =
limi xn,i . These form a Cauchy sequence since |xn xm | = limi |xn,i
xm,i | 6 kxn xm k 0. So xn y for some y C.
Finally, we can show that limi yi = y since
|yi y| 6 |yi xn,i | + |xn,i xn | + |xn y|
6 3
for large enough n and i. Thus y c, and c is a closed set in .

Corollary The space of sequences with limit 0,

c0 = { (xn ) : lim xn = 0 }
n

is a complete subspace.
Proof: It is obvious that c0 c.
c0 is closed in c since if xn y with xn c0 , then by the above proof,
limi yi = 0.
Hence c0 is closed and so complete.

Note that is a complicated space not isomorphic to 1 .
Proposition 3.5
c0 1
P
Proof: (i) Every functional on c0 is of the type (xn ) 7 n yn xn with
(yn ) 1 . First, given (yn ) 1 , this is well-defined because |sumn yn xn | 6
kxk kyk1 ; secondly, this map is a functional because it is linear (as shown in
the previous proposition) and continuous (by the above inequality). Moreover

21

let be any functional on c0 , and let yn = (en ); then (x) = (

P
n yn xn by linearity and continuity of , and
N
X
n=1

|yn | =

N
X
n=1

(en ) = (

N
X

xn en ) =

en ) 6 kkk(1)k = kk

n=1

hence y 1 .
P
(ii)Let J : 1 c0 be defined by Jy(x) = n yn xn . Then J is linear
as in the previous proof, and itPis isometric since kJyk 6 kyk1 has already
been shown, and |Jy(ei )| = i |yi | (where yi = |yi |, so
X
kJyk >
|yi | = kyk
i

3.3

The space 2

Proposition 3.6 (i)

|

(ii) The function

s
X
n

xn yn | 6

X
n

|xn

|2

X
n

|yn |2

pP

|xn |2 is a norm, since

s
s
X
X
2
2
|xn + yn | 6
|xn | +
|yn |2
n

Proof: (i) It is easy to show that |ab| 6 (a2 + b2 )/2 for any real numbers
a, b. Hence,
X
X
X
X
X X
(
an bn )2 =
ai bi aj bj 6
(a2i b2j + a2j b2i )/2 =
a2i b2j =
a2i
b2j
n

i,j

i,j

i,j

It follows, that for complex numbers xn , yn ,

s
s
X
X
X
X
2
xn yn | 6
|xn ||yn | 6
|
|xn |
|yn |2
n

(ii)
X
n

|xn +yn |2 =

X
n

|xn |2 +|yn |2 +(
xn yn +xn yn ) 6 kxk2 +kyk2 +2kxkkyk = (kxk+kyk)2

22

Proposition 3.7 2 is complete.

Proof: Let (xn ) be an absolutely summable series, ie
Fix i, then
X
X
X
|
xni | 6
|xni | 6
kxn k2
n

kxk2 converges.

so can define yi = n xni .

Now consider
PM PN
PM PN
2
|xni |)2
i=1 |
n=A xni | 6 Pi=1 (P n=AP
N
N
= M
|
i=1
n=A q
m=A |xni ||xmiq
PN PN
PM
PM
2
6 n=A m=A
|xni |2
i=1
i=1 |xmi |
q
PM
P
2 2
=( N
i=1 |xni | )
n=A

So as M, N we get

X
X
X
|
xni |2 6 (
kxn k)2
i=1

n=A

n=A

Taking A = 1 shows that y 2 ; taking A = N + 1 shows that

in 2 .

Proposition 3.8

xn y


2 2
Proof: Exercise.

3.4

Function Spaces

In a similar fashion, but using theorems from integration, it can be shown

that L1 , L2 and L are Banach spaces, and that L1 L and L2 L2 .

The proofs are very similar, and use the analogous inequalities
kf gkL1 6 kf kL1 kgkL ,
kf gkL1 6 kf kL2 kgkL2 ,
called the Holders inequality, as well as the Monotone Convergence Theorem.

23

3.5

Exercises

1. Show that x =
and .

xn en , where en = (ni ), holds in the spaces 1 , 2 ,

2. Show that 2 2 by repeating the proofs for 1 and c0 .

3. Prove that kf gkL1 6 kf kL1 kgkL .
4. Show that the adjoint of thePFourier transform is a map from 1 to
L [0, 1] such that F (an ) = n an e2inx .

n
5. Let R be the right-shift operator on 1 . What is RP
x? Show that

n
n
R x 6 0, but that R 0. (Recall that (x) =
i yi xi for some

sequence (yi ) )

24

Banach Algebras

Definition A Banach algebra is a Banach space X, over C, with an associative multiplication of vectors with unity such that
(S + T )U = SU + T U,

S(T + U) = ST + SU,

(S)T = (ST ) = S(T ),

kST k 6 kSkkT k,

k1k = 1.

Proposition 4.1 Multiplication is continuous.

Proof: follows immediately from
Tn Sn T S = (Tn T )Sn + T (Sn S),
so that as Tn T and Sn S, then Tn Sn T S.

4.0.1

Examples

CN with pointwise multiplication.

C(K) the space of continuous functions on a compact set; this contains
the closed sub-algebra of holomorphic functions when K C.
B(X) for any Banach space X.
Definition The morphisms of Banach algebras are those continuous
maps which preserve the vector structure (linear) and the multiplication,
(ST ) = (S)(T )

In particular, the complex morphisms : X C.

Theorem 4.2 Every Banach algebra can be embedded as a closed
subalgebra of B(X).
Proof: Let La (x) = ax be left-multiplication by a. Then La B(X) since
multiplication is distributive and continuous.
La (x + y) = a(x + y) = ax + ay = La (x) + La (y),
25

La (x) = a(x) = La (x),

kLa (x)k = kaxk 6 kakkxk,
so that kLa k 6 kak. Now La+b = La + Lb , and Lab = La Lb , La = La ,
L1 = I are obvious.
La+b (x) = (a + b)x = ax + bx = La (x) + Lb (x),
La (x) = (a)x = La (x),
Lab (x) = (ab)x = a(bx) = La Lb (x),
L1 (x) = 1x = I(x),
kLa k = kak
since a = a1 = La 1, so kak = kLa 1k 6 kLa k. So, the space of such operators
is a sub-algebra of B(X), and the mapping L : X B(X) defined by
L : a 7 La is an isometric morphism of Banach algebras.
Moreover it is closed, since let Lan T in B(X); then Lan x = an x =
(Lan 1)x and so as n , T x = T 1x by continuity, i.e. T = LT 1 . Thus the
sub-algebra is complete.


4.1

Differentiation and Integration

Definition A function f between Banach algebras is said to be differentiable at T when there is a continuous linear map f (T ) such that for H in
a neighborhood of T ,
f (T + H) = f (T ) + f (T )(H) + o(H)
i.e. ko(H)k/kHk 0 as H 0.
A function is holomorphic when it is differentiable in the sense above
with H = z C, i.e.
f (T + z) = f (T ) + f (T )z + o(z)

Note - if f (T ) is not defined, but f (T + H) = A + B(H) + o(H), then can

redefine f at T to make it differentiable (T is called a removable singularity).
Example: any polynomial is differentiable;
26

Integration is also well-defined on paths in a Banach algebra via

Z
Z
F (z)dz := F (z(t))zds.

We
R shall assumeR the followingRfacts about Rintegration 1. it is linear, with
T F (z)dz = T F (z)dz, 2. k F (z)dzk 6 kF (z)kds.

Theorem 4.3 Cauchys Theorem

Let F be a function from C into a Banach algebra, which is
holomorphic in a bounded region U having a finite number of differentiable curves as boundary. Then,
I
F (z)dz = 0
Proof: At any holomorphic point, F (z + h) = F (z) + F (z)h + o(h), and
for z + h in a sufficiently small disk B (z), we have ko(h)k < . Thus, for
any triangle inside this disk we get
H
H
F
(w)dw
=
F (z + h)dh

H(z)
= F (z) + F (z)h + o(h)dh
H
so that k F (w)dwk 6 P erimeter() 6 6 2 .
Now any triangle can be partitioned into a number of triangles each
smaller than . Since the triangle is totally bounded, a single canH be found
that suffices for the whole triangle (using a Lebesgue net). Thus F (w)dw
over the whole triangle is the sum of integrals over the -triangles. The
number of these does not exceed (L/2)2 where L is the largest side of the
triangle, so that
I
k
F (w)dwk 6 6 2 (L/2)2 = c,

proving that the integral vanishes over triangles.

More generally, the curves can be approximated by many-sided polygons,
and the interior can be triangulated, proving the theorem.


Corollary Cauchy Residue Theorem

The integral over a closed path depends only on those areas
inside, where f is not holomorphic,
I
X
F (z)dz =
Residuei (F )
i

27

Proof: enclose the non-holomorphic areas by curves, to form one holomorphic region.
CorollaryR If F is holomorphic in a path-connected region, then
z
the integral a F (w)dw is well-defined, and is holomorphic.

4.2

Power Series

P
n
has a radius of convergence given by R =
Any power series
n=0 an z
1/n
lim inf n |an |
.
Lemma For any T , the sequence kT n k1/n converges to a number denoted
by r(T ).
Proof: It is clear that 0 6 kT n k1/n 6 kT k. Let r(T ) be the infimum
1/N
value of kT n k1/n . We can find an N such that r(T ) 6 kT N k
< r(T ) + .
Although it is not true that the sequence of numbers is decreasing, notice
1/qN
1/N
that kT qN k
6 kT N k . More generally, for any n, let n = qN + r with
0 6 r < N (by the remainder theorem), then we get 0 6 r/n < N/n 0 as
n so that
r(T ) 6 kT n k1/n 6 kT N k

q/n
1

6 kT N k N

kT kr/n

r
(1 n
)

1
N

kT kr/n

6 (kT N k + )(1 + )
6 r(T ) + (kT k + 3)

for n large enough.


P
n
Theorem 4.4 Let f (z) =
n=0 an z be a power series with
Pradiusn
of convergence R. Then if r(T ) < R, the series f (T ) := n=0 an T
converges absolutely, and diverges for r(T ) > R.
Proof: We are given that r(T ) < R, so for n > N large enough,
!n 
n
n 1/n
kT
k
r(T ) +
n
kan T k =
6
= n
|an |1/n
R

where < 1. Note that when RP

= , replace the denominator with any
positive real number. The series n kan T n k converges by comparison with
geometric series on the right.
28

Similarly, when r(T ) > R the series diverges since

!n 
n
n 1/n
r(T )
kT
k
n
kan T k =
>
>1
|an |1/n
R+
for an infinite number of n.

Note that power series are continuous in T .
Proposition 4.5 A power series is holomorphic within its radius of
convergence.
P
P
P
Proof: n an (T + z)n = n an T n + n nan T n1z + o(z), and the radius
of convergence of the coefficients nan is lim inf n |nan |1/n = R since n1/n 1
as n .

P
Proposition 4.6 (i) n an T n = 0 within a positive radius of convergence an = 0, n.
(ii) There is only one power series such that f (Tn ) = An with
Tn 0.
P
Proof: Let n an T n = 0 for T = Tn 0 not divisors of zero and not
nilpotents;
then a0 = 0 follows from Tn 0; suppose a0 , . . . , am1 = 0, then
P
0 = n an Tnn = Tnm (am + am+1 Tn + . . .), so that am = 0 by letting Tn 0.
Suppose fP
(Tn ) = An = g(Tn ) for two power series f, g. Let h = f g, so
that h(T ) = n an T n and 0 = h(Tn ), implying that h = 0.

For example, eT = 1 + T + T 2 /2! + . . . converges for any operator T ,
and satisfies keT k 6 ekT k . It can be shown (exercise) that when (and only
when) S and T commute, we get eS+T = eS eT , from which follows that eT is
invertible with inverse eT .
Similarly, for kT k < 1, we can define log(1 + T ) = T T 2 /2 + T 3 /3 + . . ..
Proposition 4.7 If kT k < 1 then
(1 T )1 = 1 + T + T 2 + . . .
is continuous, with k(1 T )1 k 6 (1 kT k)1 .

n
n
Proof:
when kT k <
P kTn k 6 kT k and the right-hand termsPare summable
n
1, so n kT k 6 1/(1 kT k). This implies that n T also converges (since

29

the Banach algebra is complete) to, say, S. Now (1 T )S = limn (1

T )(1 + T + . . . + T n ) = limn (1 T n+1 ) = 1 since kT n k 0 as n .
Similarly S(1 T ) = 1, proving that S is the inverse of (1 T ).

Corollary An affine map f (x) = a+T x with kT k < 1 has a unique
fixed point z = f (z).
Proof: x is a fixed point x = f (x) = a + T x a = (1 T )x x =
(1 T )1 a.

Of course, the result also follows from the Banach contraction mapping
theorem, since d(f (x), f (y)) = ka + T x a T yk 6 kT kd(x, y).

4.3

Group of Invertible Elements

The invertible elements form a group, which contains 1 and the surrounding
unit ball of elements, extended to the interior of the double cone { T :
6= 0, kT 1k < 1 }, as well as all the operators eT .
Proposition 4.8 The group of invertible elements is an open set, and
the map T 7 T 1 is continuous.
1

Proof: Let T be an invertible element, and let < kT 1 k . Then for

kHk < , we get (T +H)1 = (1+T 1 H)1 T 1 , which exists by the previous
theorem since k T 1 Hk 6 1. This shows that the group is an open set.
Moreover we have
(T + H)1 = T 1 T 1 HT 1 + T 1 HT 1 HT 1 + . . .
2

This shows that k(T + H)1 T 1 k 6 kHkkT 1k + kHk2 kT 1 k + . . . 6

2
kHkkT 1 k
which converges to 0 as H 0.
1kT 1 kkHk

The proof in fact shows that the map T 7 T 1 is differentiable; indeed
this shows that the group of invertibles is essentially what is called a Lie
group, a topic that has a vast literature devoted to it.
In particular note that for H = h1, we have (T +h)1 = T 1 hT 2 +. . .,
from which follows that the map z 7 (T z)1 is holomorphic whenever
z 6 (T ). Indeed, for z 6 (T ) we have
(T z + h)1 = (T z)1 h(T z)2 + O(|h|2)
30

so that

(T z + h)1 (T z)1
= (T z)2
h0
h
lim

Proposition 4.9 Let Rn be invertible elements that converge to an

element T on the boundary of the group of invertible elements.
Then
(i) kRn1 k > 1/kRn T k;
(ii) T is a generalized divisor of zero T , i.e. there are unit
elements Sn such that
T Sn 0, or Sn T 0, as n
Proof: Let Rn T be invertible elements converging to a boundary element T . Since the group of invertible elements is open, T cannot be invert1
ible, whereas Rn and its surrounding ball of radius kRn1 k are invertible.
1
Thus kRn T k > kRn1 k .
Let Sn = Rn1 /kRn1 k; then
kT Sn k = kT Rn1 k/kRn1 k
= k(T Rn )Rn1 + 1k/kRn1 k
6 2kT Rn k 0 as n .

The generalized divisors of zero are also called topological divisors of zero,
and are not invertible (exercise).
The set of non-invertible elements is connected (through the origin), and
may disconnect the group of invertible elements.
Proposition 4.10 The maximally connected subset of invertible elements containing 1, is an open normal subgroup generated by eT
for all T .
Proof: Let T be in the maximally connected subset containing 1, denoted
by G1 . Then the connected ball B (T ) around T is also in G1 (because
maximally connected), so that G1 is open. Moreover for any T , T G1 is also
connected because multiplication is continuous, so when T G1 we must get
T G1 G1 ; hence for any S, T G1 , we have T S G1 . Similarly, taking
the inverse is a continuous mapping which fixes 1 and maps the connected
set G1 to a connected set containing 1; thus T 1 G1 . Also, by the same
reasoning T 1 G1 T G1 , which makes G1 a normal subgroup.
31

The inverse of eT is eT from their definition as power series. Hence the

group generated by exponentials eT , . . . , eS consists of finite products of the
type eT . . . eS . Moreover the elements of the type eT form a connected
set since the Banach algebra is connected and the exponential is continuous.
Multiplication is also continuous so that the group generated by exponentials
is connected. As e0 = 1, it must be a subgroup of G1 .
Claim: This group is open and closed, and so must equal G1 . The elements near to 1 = e0 are all exponentials i.e. they all have a logarithm
(check! theorem 10.30 of Rudin?), and so any element T generated by the
exponentials also has a neighborhood inside this group. Its complement must
also be open, because it consists of open cosets of the group.


4.3.1

Exercises

1. Show that if ST = 1 = T R then S = R = T 1 .

2. Find Banach algebras of dimensions 1, 2, 3 (over C)
3. Show that for any morphism between Banach algebras, (1) = 1
and (T 1 ) = (T )1 . Deduce that for kT k = 1 and || < 1 then
k(T )k < 1 and hence that must be continuous.
P
P
4. Let f (z) = n an z n , and let |f |(z) = n |an |z n . Show that kf (T )k 6
|f |(kT k).
5. Show that (i) r(T ) = ||r(T ), (ii) r(ST ) = r(T S), (iii) r(T n )1/n =
r(T ), (iv) if ST = T S then r(T + S) 6 r(T ) + r(S) and r(T S) 6
r(T )r(S) (hint: for the addition take T = T0 with r(T0 ) = 1 and
similarly for S).
6. Show that when ST = T S, eS eT = eS+T . Deduce that the inverse of eT
is eT . Show further that the map t 7 etT is a group homomorphism
from R to the group of invertible elements, which is differentiable.
7. Show that k(1 T ) (1 T )1 k 6

kT k2
1kT k

(use (1T )1 = 1+T +. . .).

8. Show that the iteration xn+1 = a+T xn , starting from any x0 , converges
to the fixed point z = a + T z. (Hint: prove first that kxn+1 xn k 6
kT kkxn xn1 k, and then that (xn ) is a Cauchy sequence)

32

9. Suppose that Ax = b is a matrix equation, with A being practically a

diagonal matrix, in the sense that A = D+B, with D being the diagonal
of A, and kBk < kDk. Hence Dx = bBx and x = D 1 bD 1 Bx. Let
T x = D 1 b D 1 Bx and show that kT k < 1 so that there is a unique
fixed point. Hence describe a recursive algorithm (due to Jacobi) for
finding the solution of the equation.
10. Similarly suppose that Ax = b with A being practically a lower triangular matrix, in the sense that A = L + B where L is lower triangular
and kBk < kLk. The inverse of a triangular matrix, such as L, is
fairly easy to compute. Describe a recursive algorithm (Gauss-Siedel)
for finding the solution x.
Rb
11. Let T f (x) = a k(x, y)f (y)dy be a mapping between function spaces.
Show that if k(x, y) is bounded (ie in L [a, b]2 ), then T is an operator
mapping L1 [a, b] to itself and L [a, b] to itself. Show further that kT k 6
kkkL |b a| in both cases. Deduce that if |k(x,
P y)| < 1/|b a| then the
equation T f + g = f has a unique solution n T n g (find the kernel of
T n ).

4.4

Spectrum of T

Definition
as the set

The spectrum of an element T in a Banach algebra is defined

{ C : T is not invertible }

Proposition 4.11 The spectrum of T is a non-empty compact subset

of C bounded by r(T ).
Proof: If || > r(T ), then T = (1 T /) is invertible. So any
spectral values must satisfy || 6 r(T ).
Let 6 (T ), ie T is invertible. For any z C, k(T z) (T )k =
|z |, so that for z close enough to , we have T z also invertible.
This shows that is an exterior point of the spectrum, and hence that the
spectrum is closed.
Closed and bounded sets in C are compact.
Suppose (T ) = , so that (T z)1 is holomorphic everywhere. HThen
by Cauchys theorem, we get that the integral over any circle vanishes (T
z)1 dz = 0. But (T z)1 = z1 z12 T + . . . for |z| > kT k, so that for a

33

circle of radius larger than r(T ) we get

I
I
I
1
1
1
1
(T z) dz = 2 T + . . . dz =
dz = 2i 6= 0,
z z
z
so it must be the case that the spectrum does not vanish. (Alternative
proof: if the map 7 (T )1 is differentiable everywhere, then since
(T )1 0 as , by Liouvilles theorem we get (T )1 = S for
all , a contradiction)

Proposition 4.12 Spectral Radius Formula
The largest extent of the spectrum, called its spectral radius, is
1
r(T ) = limn kT n k n .

Proof: We have already shown that (T ) is bounded by r(T ). We need

to show that there are spectral values with magnitude r(T ).
Invert the spectrum to form the set S = { z 1 : z (T ) }. Thus
z 6 S (T z 1 ) is invertible, and so the largest radius of that disk in which
(T z 1 )1 exists is 1/r, where r is the spectral radius of T . Inside this disk,
(T 1/z)1 is holomorphic, and soPcan be written as a power series, which
1
turns out to be
= z n (zT )n . But the radius
P (T 1/z)
P of convergence
n
n
n
of the series
(T
z)
is
the
same
as
that
of
the
series
n
n kT kz , that
is lim inf n kT n k1/n = 1/r(T ). Thus we must have 1/r (T ) 6 1/r(T ), ie
r(T ) 6 r (T ).

Note that if is a boundary point of the spectrum, then T is at the
boundary of the group of invertible elements, and so is a topological divisor
of zero.
Theorem 4.13 Spectral Mapping Theorem
The spectrum of f (T ) is equal to the set { f () : (T ) }; ie
(f (T )) = f ((T ))

Proof: We start by considering f (x) = p(x) a polynomial. Let p(x) =

(xa)(xb) . . . where a, b, . . . are the roots, dependent on . Whether is a
spectral value of p(T ) or not depends on whether p(T ) = (T a)(T b) . . .
is invertible or not.

34

Suppose that 6 p((T )) i.e. p(z) 6= 0 for any z (T ). Hence any

z (T ) is not one of the roots a, b, . . . and so the factors (T a) are all
invertible. This shows that 6 (p(T )).
Conversely suppose that 6 (p(T )), so that p(T ) has an inverse
S. Then S(T a)(T b) . . . = 1 = (T a)(T b) . . . S; thus for any root
a, there are elements R and R such that (T a)R = 1 = R (T a). But
R = R (T a)R = R, so that any factor T a is invertible. Thus all the
roots of p(z) are not in the spectrum. But if = p(z) then z is a root,
and so = p(a) 6 p((T )).
P
P
PN
n
n
Now let f (T ) be a power series n an T n . Then ( N
a
T
)
=
n
n=0
n=0 (T )
converges separately to (f (T )) and f ((T )).

4.4.1

Exercises

1. Show that the resolvent C (T ) is open by proving that the map

7 T is continuous, and using the fact that the invertible elements
form an open set.
2. Suppose that p(T ) = 0 for some polynomial p. Show that (T ) consists
of roots of p.
3. Use the spectral radius formula to show that the spectral radii of T S
and ST are the same.
4. Use the spectral mapping theorem to show that if eT = 1 then T = 2iS
where (S) Z.
5. Let T be a nilpotent operator ie n : T n = 0; show that its spectrum
consists of just the origin.
6. Suppose that the only non-invertible element of a Banach algebra is 0
(called a division algebra). Using the fact that the spectrum is nonempty, show that T = for some 6= 0, and hence that the Banach
algebra is isomorphic to C.
7. Suppose that the only topological divisor of zero is 0. Show that the
spectrum of any element is a single point (consider its boundary) and
so each non-zero element is invertible. Deduce that the Banach algebra
is C.

4.5

Holomorphic Functions f (T )
35

Theorem 4.14 Taylor series If f : C C is holomorphic on a disk

Br (z0 ) that includes (T ) then f is a power series about z0 with
I
X
1
n
f (T ) =
an (T z0 ) =
f (w)(w T )1 dw.
2i
n
Laurent series If f is holomorphic on a disk Br (z0 ) except at
its center z0 6 (T ) and the disk includes (T ), then the (Laurent)
series converges

1
an (T z0 ) =
2i
n=
In both cases, an =

1
2i

f (w)(w T )1 dw.

f (w)(w T )n1 dw.

T z0 1
Proof: (i) (w T )1 = (w z0 + z0 T )1 = (w z0 )1 (1 wz
) =
0
1n
n
(T z0 ) since r(T z0 ) < R = |w z0 | by the spectral
n (w z0 )
mapping theorem.
H
H
P
1
1
n 1
Thus,
f
(w)(w

T
)
dw
=
(T

z
)
f (w)(w T )1n dw =
0
n
2i
2i
H
P
1
n
1n
f (w)(w T )
dw.
n an (T z0 ) , with an = 2i
By Cauchys theorem, we can take any curve surrounding the spectrum
of T .
(ii) Continuing the analysis of (i), we also get that on a small circle
around z0 , (wP
T )1 = (w z0 + z0 T )1 = (T z0 )1 (1 (w z0 )(T
1 1
z0 ) ) = n (w z0 )nH(T z0 )n1 . Thus taking
in a clockP the integraln1
1
1
wise direction,
we
get
f
(w)(w

T
)
dw
=
a
(T

z
)
where
0
n n
2i
H
1
n
an = 2i f (w)(w z0 ) dw.


Corollary If f : C C is holomorphic in a neighborhood of

a point, then it is a power series, and thus infinitely many times
differentiable, with
n!M
|f (n) (z0 )| 6 n
r
Proof: |f (n) (z0 )/n!| = |an | 6

1
kf kL (Br (z0 )) r n1 2r
2

Corollary (Fundamental theorem of Algebra)

Every polynomial p(z) has n roots.

36

= M/r n .


Proof: Suppose that p(z) has no roots, so that 1/p(z) is holomorphic.

Now p(z) , in fact for |z| large enough, |p(z)| > |z|n |an1 z n1 + . . . +
a0 | > |z|n /c. So
I
1
1
1
6
|w|1ds 6 c/Rn 0
|p(0)|
2
|p(w)|
as R , which is a contradiction.

Corollary (Liouville) if F (z) is holomorphic throughout C then

it must be unbounded or constant.
Proof: Suppose f : C C is holomorphic but bounded by M on C.
Then for any z, |f (z)| 6 M/R 0 as R , so f = 0.
More generally, if F (z) is holomorphic, then so is F : C C for any
functional . If it is bounded, then (F (z)) = (F (0)) for any , which
implies that F (z) = F (0) for all z.

Corollary Two holomorphic functions on C must agree on a
connected set if they agree on an interior disk.
Proof: let A be the set on which f = g locally i.e. z A f = g on
B (z). Then A 6= by hypothesis, A is open by definition, and A is closed,
since suppose z is a limit point, with zn z and zn A. Then f and g are
power series about z with f (zn ) = g(zn ); hence f = g throughout the radius
of convergence, and so z A. It follows that A is all of the connected set.

H
1
Taylors theorem implies that
for
any
polynomial
p(T
)
=
p(w)(w
2i
H
1
1
1
1
T )1 dw; also (T )
=
(w

)
(w

T
)
dw
for

6
(T
).
2i
H
1
1
The integral 2i f (w)(w T ) dw exists for any T for which f is holomorphic in a neighborhood of (T ) (since k(w T )1 k Ris continuous in w and
1
bounded since the path is compact; hence kf (T )k 6 2
|f (w)|k(w T )1 kds 6
M.
Definition For any function f : C C that is holomorphic in a
neighborhood of (T ),
I
1
f (T ) :=
f (w)(w T )1 dw
2i

37

Note that f is differentiable (and so continuous) at T (for H sufficiently

small, f (T H+ H) is defined since (T + H) (T ) + B ; also, f (T + H) =
1
f (T ) + 2i
f (w)(w T )1 H(w T )1 dw + o(H))
also discuss the integral about a part of (T ); in particular
HOne can 1
1
(w T ) dw is a projection map on T .
2i
Theorem 4.15

(f (T )) = f ((T ))
Proof: Let 6 f ((T )), so (f (z) )1 is holomorphic on the open set
U { z : f (z) = }, which contains (T ). Thus (f (T ) )1 exists, and
6 (f (T )).
()
Conversely, let (T ), and let F (z) := f (z)f
which is holomorphic
z
on (T ) and at . So since (z )F (z) = f (z) f () we get (T )F (T ) =
f (T ) f (). But T is not invertible, and so neither is f (T ) f () i.e.
f () (f (T )).

Examples: We can define log T whenever 0 6 (T ) and one can draw a
contour from 0 to outside kT k (so that one can enclose (T ) by a simple
closed path in which log(z) is holomorphic), but this definition is not unique.
Note that elog T = T .
Furthermore, can define T 1/n (again not unique) with (T 1/n )n = T .
4.5.1

Exercises

1. Show that (f + g)(T ) = f (T ) + g(T ), and that f g(T ) = f (g(T )). It

is also true, but more difficult to show that (f g)(T ) = f (T )g(T ).
2. Show that f (z1) = f (z)1 where z C and 1 is the identity element.
Hence f (T ) is truly an extension of the complex holomorphic function
f (z).

4.6

B(X)

T invertible means there is an operator T 1 B(X), that is the inverse is

continuous and linear. Now if an operator T B(X) is bijective, then T 1
will always be linear, but not necessarily continuous. So one must be careful
in interpreting the term invertible in B(X).
If is an eigenvalue, ie T x = x for some x 6= 0 (called an eigenvector)
then (T )x = 0, hence T is not 1-1, and so not invertible.
38

What if T is 1-1? If it is also onto, then by the open mapping theorem,

a bijective operator has a continuous inverse, and is not a spectral value.
So for spectral values for which T is 1-1, it cannot also be onto.
If T is 1-1, and its image is dense in X, ie im(T ) = X, then
is said to be part of the continuous spectrum. In particular we can find xn
such that (T )xn 0, ie T xn xn .
If T is 1-1, and its image is not even dense in X, then is said to be
part of the residual spectrum.
Proposition 4.16 The eigenvectors of distinct eigenvalues are linearly independent.
Proof:
PLet vi be eigenvectors associated with the distinct eigenvalues i .
Then if N
i=1 i vi = 0 implies
0=

N
X
i=1

i (T N )vi =

X
i

i (i N )vi =

N
1
X
i=1

i (i N )vi =

N
X

i vi

i=1

Thus by induction we get i = 0 ie i = 0 for i < N. Hence N vN = 0 which

implies N = 0 as well.


4.6.1

Example

1. Let L(xn ) = (xn+1 ) be the left-shift operator on 1 . The eigenvalues are

given by T x = x ie (xn+1 ) = (xn ), ie n, xn+1 = xn . This is a recurrence
n
relation, which can easily be solved
to give xP
n = x0 . For this to be an
P
eigenvector in 1 we need that n |xn | = |x0 | n ||n to be finite, and this
implies that || < 1. Conversely, for any such 6= 0, the sequence (n ) is an
eigenvector. Hence the spectrum consists of the unit closed ball, since the
norm is 1.
R1
2. Let T : L1 [0, 1] L [0, 1] be defined by T f (x) = 1x f (s)ds. Then
R1
T is trivially linear; it is continuous since kT f kL = supx | 1x f (s)ds| 6
kf kL1 . Thus kT k 6 1.
R1
For the eigenvalues, we get 1x f (s)ds = f (s). Differentiating twice
gives f (x) + 12 f (x) = 0 with boundary conditions f (0) = 0 = f (1). Thus
f (x) = B sin(x/) where = 2/k, k odd.
Proposition 4.17
(T ) = (T )

39

Proof: let 6 (T ), then one can easily check that T has the inverse
(T )1 , showing that 6 (T ).
Conversely, let 6 (T ), then by the previous proof 6 (T ). But
X is embedded in X , so that T equals T when restricted to X. However,
we need to show that its inverse can also be restricted to X. We can deduce
that T is 1-1, and we now show it is also onto. Suppose otherwise, so
we can find a non-zero functional in im(T ) = ker(T lambda); hence
T is not 1-1 contrary to our assumption that 6 (T ). Thus T
is 1-1 and onto, and hence its inverse equals the restriction of the inverse of
(T ) . Thus 6 (T ).


4.6.2

Exercises

1. Show that if is an eigenvalue of T then f () is an eigenvalue of f (T ),

with the same eigenvector.
2. Let T : 2 2 be defined by T (xn ) = (n xn ) where n are dense in
a compact set K. Show that its spectrum is K and that there is no
residual spectrum.
3. Let T : C[0, 1] C[0, 1] be defined by T f (x) = xf (x). Show that T is
an operator, find its norm and show that its spectrum consists of only
the residual part.
Rx
4. Let T : C[0, 1] C[0, 1] be defined by T f (x) = 0 f . Show that
Rx
n1
n
T n f = 0 (xy)
f (y)dy, and that kT n k 6 |x|
. Deduce, using the
(n1)!
n!
spectral radius formula, that its spectrum is just the origin.

C -Algebras

Definition A C -algebra is a Banach algebra with an involution map

: X X with the properties:
, (ST ) = T S , T = T, kT T k = kT k2
(T + S) = T + S , (T ) = T

40

Proposition 5.1 The involution map is continuous, and

kT k = kT k
Proof: We have kT k2 = kT T k 6 kT kkT k, so that kT k 6 kT k; furthermore, kT k 6 kT k = kT k.

Proposition 5.2
(T ) = (T )
Proof: If T is invertible, then so is T since T (T 1 ) = (T 1 T ) = 1 = 1,

and similarly (T 1 ) T = 1. Thus, if T is invertible, then so is T ,

and vice-versa.

It is a theorem that every C -algebra can be embedded into B(H) for
some Hilbert space H.

5.1

Normal and Self-adjoint elements

Definition
An element T is called normal when T T = T T , selfadjoint when T = T , and unitary when T = T 1 .
Every element can be written as T = A + iB with A and B self-adjoint.
5.1.1

Exercises

1. Show that T + T , T T and T T are self-adjoint for any T .

Note: For B(H), the group of invertible elements is connected, and generated by two exponentials. Proof: every T = RU where R is self-adjoint
and U unitary. Thus since logarithms exist, R = eS and U = eiH for selfadjoint operators S, H; in particular any element of the group of invertible
elements is of this form. Moreover the map t 7 etS eitH gives a connected
path starting from 1 and ending at T , so that the group is path-connected,
and so connected.


41