Chapter I1

Elementary Spectral Theory

After this brief discussion of Operator Theory on finite-dimensional spaces,

we turn to infinite-dimensional ones. Until the end of the book, E will be a
complex, infinite-dimensional, separable Banach space. For basic notions about
norms, weak topologies (and so on) on Banach spaces, we refer the reader to
our book B.B. [l].

We denote by T a linear operator on E. We assume T to be continuous,

that is :
There exists a constant C such that, for all x in E :

As in Chapter I, we define the norm of T , by the formula :

This time, unlike in the finite-dimensional case, continuity does not follow
automatically from linearity. However, it is not possible t o exhibit explicitly
non-continuous operators or linear functionals on Banach spaces : to build
them, one needs to use the axiom of choice.

We observe that on a Banach space, the operator X is continuous if and

only if it is weakly continuous :

L e m m a . - Let T be an operator on a Banach space E. Then T is continuous

from ( E l11.11) into itself if and only if it is continuous from ( E , u ( E , E * ) )into
itself.

Proof. -- T h e fact that continuity for the norm implies continuity for the weak
topology can be found in B.B. 111, p. 37.
Conversely, let’s assume that T is continuous from ( E , v ( E , E * ) )into
itself. Let 8 be the closed unit ball of E , and let V be a neighborhood of 0
20 Chapter II

in ( E , u ( E ,E ' ) ) . Then T-'V is a neighborhood of 0 for u ( E , E * ) .But B is

weakly bounded, therefore there exists a X such t hat :

This implies th at XV 3 T(B), and shows that T ( 8 ) is weakly bounded. By

Mackey's theorem (see B.B. 111, p. 35) it is also bounded in norm, so T is
continuous in norm.

Continuity for u ( E ,E') can be characterized as follows : for every ( E E' ,

( o T E E'.

Of course, the previous Lemma holds for any operator between two Banach
spaces E and F .

These preliminaries being settled, we may a s k the following question : what

remains of the theory presented in Chapter I ? We remember that everything
started with th e fact t hat the operator T satisfied a polynomial equation : we
found a polynomial p such t hat p ( T ) = 0 .

Unfortunately, in infinite dimensional spaces, the construction breaks down

already a t this early stage. Not every operator satisfies a polynomial equation.

For instance, on the space

equipped with the norm II(a,)ll2 = (C,,a la,1z)1'2, the right shijt opcralor,
defined by :
S((an),tZ) (an-1)nEZ 9 (1)

does not satisfy any polynomial equation.

Indeed, let (e,),€Z be the canonical basis of 12(Z) : en is the sequence
having 1 a t the n t h place, 0 otherwise ( n E Z ) .
Then S e n = e,+1. Assume t hat a k S k c o = 0 . we get
1," co
N
akCk =0,
which implies th at all P k 's are 0.

T h e results of the previous chapter do not hold any longer, and we have
to introduce new notions.
 Elementary Spectral Thwry 21

1. T h e spectrum of an operator.

Definition 1.1 - The resolvent set p ( T ) is the set of X E C such that X - T

is invertible : that is, the inverse R(X,T) = (A - T)-' exists as a continuous
operator defined on the whole space E . This inverse is called the resoIucnt of
T. The operator being fixed, we just write R(X) instead of R(X,T) for the
resolvent.

The spectrum o(T) is the complement of the resolvent set : therefore, it

is the set of X ' s such that X - T is not invertible.

Proposition 1.2. The resolvent set p ( T ) is an open subset of the complex

~

plane, and the application X + R ( X ) , from p ( T ) into C(E), is an analytic

function in p ( T ) .

Remark : Analytic functions, in the elementary theory, usually take their

values in C , and not, as here, in a Banach space (namely L(E)). But this
makes no difference and the usual properties are valid with no changes.

Proof. - We fix X E p ( T ) , and let p E 6: , with 1p1 < l/IIR(X)ll. We will show
that X +p E p(T).
Set S ( P ) = z k >-O P k ( X - - T ) - ( ' + ' ) = C k >- O p k R ( X ) k + l *Since l I p ~ ( X ) l l<
1, this series converges in operator norm, and we have :

and since S ( p ) commutes with X + p - T ,it is its inverse. This proves that
p ( T ) is open, and also that R(X + p ) is analytic in some neighborhood of p = 0
(since it is expandable in power series).

We define d(A) = dist(X, a ( T ) )= inf { [ A - X'I ; A' E a ( T ) } .

Corollary 1.3. ~ IlR(X)ll 2 l/d(X), for all X E p ( T ) .

Proof. - If lpl < l/IIR(X)ll , we have seen that X +p E p ( T ) . So d ( X ) 2

1/llR(~)ll-

This Corollary implies obviously that IIR(X)ll + 00, when d(X) + 0 .

Therefore, p ( T ) is the domain of analyticity of R(X) : this function cannot be
extended analytically on a larger domain.
29 Chapter 11

Propoeition 1.4. - Thespectrum a ( T ) is non-empty, compact, and contained

in the disk { A ; 1x1 5 llTll}.
Proof. - We set g(X) = Ck,O -
T k / X k + ' .For 1x1 > IlTll, this series converges
in operator norm, and ( A - T ) g ( X ) = I , so g(X) = R ( X ) , for 1x1 > IlTll.
Assume t h a t the spectrum is empty. Then g(X) is analytic in the whole plane.
But it satisfies 11g(X)II + 0, when 1x1 + 00, so, by the maximum modulus
principle, we must have g(X) = 0 for all X f Q : , which is absurd. Thus,
a ( T ) is non-empty. T h e previous computation shows it is contained in the disk
{ A ; 1x1 5 llTll}. It is therefore compact.
We also note the analytic expansion :

and r(T) is called the spectral radius of T . The following Proposition gives a
way of computing r(T) and extends formula (2) :
Proposition 1.5. - We have :
a) r(T) = Iimk,, I I T ~ I ~ ' 5' ~ 1 1 ~ 1 1 ,
6) if 1x1 > r(T), R ( X ) = Ck>O , series converging in operator
T k / X k + ' this
norm.
Proof. - T h e radius of convergence of the Laurent series xk,OT k / X k f lis
-

1 = IimsupliT k /II l k .
k-m

So r(T) = I , All we have to show therefore is that r(T) .

Ilim infk,, /ITk 11 I l k .
For this, observe that T k- A k = (T- X ) p k ( T ), for some polynomial pk. There-
fore, if X E o(T),A' € a ( T k ) for
, every k >
0 ; so [ X I k <
/ITkll,and 1x1 <
liminfk,, ~ ~ T k ~This l k . the existence of the limit limk,,
~ l proves IITkII1lk,
and establishes our formula.
Proposition 1.6 (Resolvent Equation). - For X , p E p ( T ) , we have :

RIA) - R ( P ) = ( P - A ) . R ( X ) . R ( P ) .
Proof. ~ We have :

from which the result follows.

 Elementary Spectral Theory 23

We also recall :
Proposition 1.7. - If IlTli < 1, I - T is invertible.
Proof. - T h e series I + T + T 2+ ... is norm convergent, and provides the
required inverse.
Corollary 1.8. - The set of invertible operators is open in f (E).
Indeed, let T be invertible, and U ,with llUl1 < 1/1IT-'II. Then I + T - ' U
+
is invertible and also T ( I T-'U) = T U . +
If E' is the dual space of E (see B.B. [ l ] ,Part I, Chapter 2), we denote
by tT t h e transpose of the operator T . It acts from E' into itself, a n d is
defined by t h e equation :

' T ( ( ) z = ((Tz), for all z f E, ( f E' .

(see B.B. [ I ] , P a r t I, Chapter 2, for the elementary properties of this operator).

Proposition 1.9. .- o(T)= o ( ' T ) ,and, if X E p ( T ) , R(X, ' T ) = ' R ( A ,T ) .

Proof. - If X E p ( T ) , R(X,T)(X T ) = I , so :
~

(A - ' T ) ' R ( X , T )= I

and therefore X - ' T is invertible and its inverse is ' R ( X , T ) .

If H is a Hilbert space, equipped with the scalar product (., .), we define
the adjoint of T ,denoted by T' , by :

(T*z,y) = (z,Ty) , for all z , y E H.

If the field of scalars wasIR, we would have T' = ' T . However, for us the
field of scalars is (c , and we obtain :
Proposition 1.10. ~ We have : o ( T * )= conj(o(T)), where conj(a(T))is
t h e s e t { A ; X E a ( T ) ) ,and R(X,T)' = R ( X , T ' ) .
Proof. - If X E p ( T ) , R(X,T)(X- T ) = I , so :

R(X,T)'(X - T ' ) = I ,

which implies t h e result.

Further manipulations of the spectrum will be made later, when we deal
with functions of a n operator . We now turn to the description of several
subsets of o ( T ) .
 Chapter II

- up(T), point spectrum of T, is the set of eigenvalues of T ,

- o,(T), continuous spectrum of T , is the set :

a,(T) = {A E u(T); A-T is injective, non surjective, but has dense range }

- or(T), residual spectrum of T , is the set

ur(T) = { A E o ( T ) ; X - T is injective, b u t does not have dense range }

Obviously,
o(T) = o,(T) u G ( T )u + ( T )
As it is well-known (see B.B. [ I ] , part I, chap.2, p. 40),
a) T h e image of T is dense in E if and only if ' T is injective,
b) T is injective if and only if 'T has dense range in E' , for the topology
o(E',E).
Applying this t o X - T, we get :
Proposition 1.11. - We have :
a) +(T)= up('T),
b) u p ( T )c o , ( ' T ) , with equality if E is reflexive.
T h e point spectrum ap(T) may be empty. For instance, the right shift S
has no eigenvalue. Indeed, assume on the contrary that there is a X E (c and
a = (ak),,Z E 12 such that S a = Xu. Then, for all k E z,O k - 1 = X a k . But
llSallz = 11~112, so 1A1 = I , and this contradicts a k + 0 , when k - + 0 0 .
T h e same computation shows in this case that up ('T )is also empty. There-
fore, the spectrum reduces t o a c ( T ) .
T h e most important subset of a ( T ) is still to be defined :
- a , ( T ) , approximate point spectrum of T , is the set of all X E o ( T ) such
that there exists a sequence (z,),?~ of points in E l with llznll = 1, for all n ,
and T z , - Xz, -+ 0 , when n + 0 0 .
Such a sequence (z,),>o will be called a sequence of almost eigenucctors
for A .

Proposition 1.12. The set o , ( T ) is a closed subset of o ( T ) ,which contains

~

aa(T) (boundary of o ( T ) ) ,u p ( T )and a , ( T ) .

Proof. - To see that a,(T) is closed, we first observe t h a t X 4 o,(T) if and
only if there is a 6 > 0 such that, for all z E E ,
 Elementary Spectral Theory 25

+
Therefore, if IpJ< 6 , X g 4 o,(T), and o,(T) is closed.
Let's show t h a t the boundary of a ( T ) is contained in a,. Let XO E aa(T),
E > 0. Choose X 4 a(T),IX - Xol < ~ / 2 .By Corollary 1.3,

Il(X -T)-lIl 2 1 / w 2 2/E.

So, there is a y in E , llyll = 1 , such that :

II(X - T)-'ylI 2 2/. .

P u t z = (A - T)-'y/ll(X - T)-'yll. Then 1 1 ~ 1 1= 1 , and
II(X0 - T). - (A ~ T).ll < E/2,

and this proves our claim.

Finally, it's clear t h a t a,(T) c a , ( T ) . For a , ( T ) , we observe that if
X # a,(T), the inequality (3) implies t h a t X - T has closed range and is
injective. It cannot have dense range, since X E a(T).Therefore X 4 a , ( T ) .
T h e following Proposition describes the behaviour of a sequence of almost
eigenvectors ( z n ) , , > O , with respect to the weak topology.
Proposition 1.13. - Let T be an operator on a reflexive Banach space E ,
X # 0 in u,(T), and ( z ~ ) , , ?a~normalized sequence of almost eigenvectors for
A . Then, either X is not an eigenvalue, and (zn),,>o tends to 0 weakly, or X
is an eigenvalue, a n d every non-zero weak accumulation point of this sequence
is a eigenvector for A .
Proof. - Since E is reflexive, there is a subsequence z i converging weakly to
some point z in E . Then :
Tzk - Xzk -+ O in norm
Tzk -+ Tz weakly
and therefore :
Xzk + TZ weakly
But :
XzL + Xz weakly
So, since X # 0 , we obtain Tz = Xz.
If, for every subsequence z',. we find t = 0 , the sequence (z,,),,?~ con-
verges weakly to 0 (we are in a compact set : t h e unit ball for the weak topology,
so if every accumulation point is 0, t h e sequence will converge to 0), or, for
some subsequence x',, the point z is not 0. In this case, X is an eigenvalue,
and z a corresponding eigenvector.
 If we a r e in a Hilbert space a n d if X is not a n eigenvalue, we may assume
t h a t the sequence (zn),,20 is orthogonal. Indeed, in a Hilbert space, for any
normalized weakly null sequence (zn),,20,we can find a subsequence (z;) a n d
a n orthonormal sequence (yn),>o such that 1 1 ~ : - ynll + 0 (this is done t h e
following way : fix a n orthonormal basis of the Hilbert space. Take a sequence
( z n ) , , ~ oof vectors with finite support on this basis, such t h a t llzn - + 0,

and then take a subsequence of t h e zn 's, such t h a t t h e supports are disjoint.)

So the sequence ( Y , ) ~ > O will be also a sequence of almost eigenvectors.

T h e most important p a r t of the spectrum is a , ( T ) \ u p ( T ) ; the pieces

o , ( T ) a n d u p ( T )correspond to eigenvalues for T or 'T and may be empty.
We will see in Section 3 t h a t , when looking for Invariant Subspaces, one may
assume o(T)= a,(T).

As a n illustration of these notions, we now determine the spectrum of the

right shift S on 1 2 ( 2 ) .

Proposition 1.14. - The spectrum o f the right shift S on l z ( Z ) is the unit

circle C.

Proof. - Since llSll = 1, we have r(S) 5 1, and o(S) c 0.Also, X - S is

invertible if and only if S-'(X - S) is invertible, or AS-' - I , which is the
case if 1x1 < 1, by Proposition 1.7. This proves t h a t a(S) c C (and applies,
more generally, to any surjective isometry).
We have to show t h a t a ( S ) = c.
For y = Y k e k , we consider the
equation :
(ere - S)Z = y (4)

It can be written :
i0
e Zk-Zk-1 = yk

a n d implies :
e"zk - zo = y I + el'yz + . . . + e t ( k - 1 ) ' Yk

But if we choose for instance yk = e - ' ( k - l ) e / k , or, more generally, such t h a t :

this is impossible for z E 1 2 . Therefore, equation ( 4 ) cannot be solved for such

a y , and eie - S is not surjective. So a(S)= a,(S) = u c ( S )= C .

As we mentioned, t h e spectrum of a surjective isometry is contained in C .

For a non-surjective isometry, we have :
 Elementary Spectral Thwry 27

Proposition 1.15. - If T is a non-surjective isometry, a ( T ) = 0.

Proof. - For all 5 in E , all X 4 C , we have :

which implies that & ( T ) = C , by (3). Since T is not surjective, o(T)contains

0. These two facts together imply t h a t o ( T ) = 0.
2. T h e analvtic functional calculus.
As in Chapter I, let 7(T)be the set of functions f , with complex values,
defined and analytic on some neighborhood of o ( T ) (this neighborhood does
not need t o be connected, and may depend on the function). We simply say
"analytic on a ( T ) " .

Let f E 3(T), and U an open set, containing a ( T ) ,such that the bound-
ary r = aU consists in a finite number of rectifiable Jordan curves, oriented
in the positive sense. We amume that f is analytic on v.
We define :

This definition does not depend on the domain U with the required properties :
if we take another C J , R(X) is analytic between both boundaries, and we just
apply Cauchy's formula.

For instance, if r is a circle centered a t 0, with radius > r(T) , we have :

We now turn to elementary properties of this definition.

Propoeition 2.1. - For f , g E 7 ( T ) ,a,p E C , we have :
a) af + P!7 f 7 ( T ) and9 (af+ P g ) ( T ) = af(T) + 9

b) f.9 E 3 ( T ) , (f.q)(T) = f(T)g(T)>

c) if f(X) = C," akAk, this series converging in a neighborhood of a ( T ) ,

then f(T) = C," a k T k .

d) E 3 ( ' T ) , and / ( * T )= ' f ( T ) .

Proof. - a) is obvious. For b), f.g E 3 ( T ) is clear. Let Ul, U z , be two

neighborhoods of a ( T ) , with c U2, and f,g be analytic on az. Let
88 Chapter II

rl ,rz be the boundaries of U1 , UZ.Then :

But

and

So we get :

c) For E > 0, small enough, the series 1 ; akAk converges uniformly in the
disk { A E a ; . if Cr+= is thecircle { A E C ; 1x1 = r ( T ) + E } ,
1x1 5 r ( T ) - t ~ }So,
we have :

d) is a consequence of Proposition 1.9.

A similar proposition holds for adjoints in a Hilbert space :

 Elementary Spectral Thwry 29

Proposition 2 . 2 . - Let T be an operator on a Hilbert space H. For f E

3 ( T ) ,we de6ne : -
!'(A) = !(A) , A E a.
Then f f l E 3 ( T ' ) , and
f(T)' = fU(T*).
Proof. - Clearly, f ' is analytic on some neighborhood of o(T') = conj(u(T)).
Then from formula (1) follows :

--
1
- 2iA
= fn(T*).
L -
f(J)R(X,T')dA

We will now investigate how the spectrum is transformed by a function

of the operator. This question is of course of fundamental importance when-
ever a functional calculus is defined. In the present case, the result is quite
satisfactory, and is described by the next theorem :
Theorem 2.3 (Spectral Mapping Theorem). - If f E 3 ( T ),
f(4T))= o w ) ) ,
where f ( o ( T ) )= {f(X) ; X E o(T)}.

Proof. - Let X E o(T)and define g by

Then g E 3 ( T ) ,and satisfies :

(A - M E ) = !(A) - f(E)
Therefore :
(A - A X ) - f(T)
T)S(T)=
If f ( X ) - f ( T ) was invertible (that is f ( X ) 4 o ( f ( T ) ) )let
, A be its inverse.
Then A.g(T) is the inverse of X -T, so X 4 a ( T ). This contradicts our original
assumption, so /(A) E u ( f ( T ) )and
, f(u(T)C ) u(f(T)).
Conversely, let's assume that p I$ f ( u ( T ) ) . Then the function h ( [ ) =
r/(f(() - p) is in 3 ( T ) and satisfies h([)(f(E)- p) = 1. Therefore :
h(T)(f(T)
- P) = 1,
and p 4 o(f(T)).
This proves that f ( o ( T ) )3 u ( f ( T ) ) .
SO Chapter II

We now study the composition of functions.

Proposition 2.4. - Let f E 3 ( T ) , g E 3 ( f ( T ) ) and

, set h = g o f . Then
h E 3 ( T )9 and W )= g(f(T)).

Proof. - From the previous theorem, we deduce that h E 3 ( T ) . Let U be a

, of o ( T ) ,as previously, such that g is analytic on
neighborhood of o ( f ( T ) ) V
-
U ,and f is analytic on with v, f(v)
c U . Let's consider the operator :

If A 4 f ( a ( T ) ) ,( A ~ f(())-' is analytic on u ( T ) ,and A(A) = ( A - f)-'(T).

so :
(A - f ( T ) ) A ( X ) = I , and A = R ( X , / ( T ) ).

Therefore,

Proposition 2.5. - Let be a sequence o f functions, in 3 ( T ) , all

analytic on the same neighborhood V of a ( T ) . I f (f,,),,>o converges to f ,
uniformly on V , then fn(T) converges to f ( T ) in operator norm.

Proof. - Let U be a neighborhood of a ( T ) ,with c V , and r = a U . Then

f,, I uniformly on I?, and :
-+

Therefore, Iljn(T)- f(T)II -+ 0.

Definition. - A point XO E a ( T ) is said to be an isolated point of a ( T ) if

there exists a neighborhood U of A such that U n a ( T ) = {Ao).
An isolated point A 0 of o(T)is a pole of T if the function X -+ R ( A , T )
has a pole at Xo.
The order of AO is denoted by Y T ( A O ) and is called the order of the pole.
 Elementary Spectral Thwry 31

Theorem 2.6 (Minimal Equation). - Let f E 7(T).Then f(T) = 0 if

and only if f is identically 0 on a neighborhood of u(T),except maybe at a
finite number of poles X I , . . . ,Am , such that for all j = 1,. . .,rn, f has a zero
of order _> V T ( X , ) a t A,.
Proof. - Assume first that f has the described properties. Let C, be a small
circle around X i , positively oriented ( j = 1,. .. ,m ) . Then :

But, a t X j , R(A) has a pole of order v ~ ( X j ) ,and f has a zero of order

2 v ~ ( X j )So . f ( X ) R ( X ) is analytic on the disk limited by C j , and f ( T ) = 0.
Assume now that f(T) = 0. By Theorem 2.3, f ( a ( T ) )= 0 . Let U be a
neighborhood of a ( T ) on which f is analytic.
By a well-known theorem about the zeroes of an analytic function (see H.
Cartan [ I ] ) , f is identically zero on some neighborhood of any point of a ( T )
which is not isolated in a ( T ). Indeed, if X is such a point, if V is a disk around
A , with v c U , V contains a sequence of points of a ( T ) converging to A , and
f is identicalIy 0 on V .
Let now X E o(T), an isolated point : there is an open disk V c (I,with
a ( T ) n V = { A } . We assume that f is not identically 0 on V .
We have f(X) = 0 , and so f has a zero of order n a t A . So the function
g1(() = (A - ()"/f(() is analytic on V . Let e be a function identically equal
to 1 on V and to 0 on a neighborhood of o(T)\{A}.Let g = g1.c Then :

( A - T ) " c ( T ) = /(T)g(T)
= 0.

The Laurent series of R ( ( ) on V can be written :

€1

where

- ( A - T)m- e( T ),

where C is a circle around A , positively oriented.

SO A - m + l = -(A - T)me(T) = 0, for rn 3 a. Then (1) implies that
X is a pole of order 5 n , and the Theorem is proved.
32 Chapter II
We recall that an operator T’ is said to commute with T if TT‘ = TIT.
The set of operators which commute with T is called the commutant of T and
denoted by {T}’.
Proposition 2.7. - Let T’ be an operator which commutes with T and let
E T(T). Then T’ commutes with f ( T ) .

Proof. - By the definition of f(T), it’s enough t o show that T’ commutes

with R(A, T), for X E p(T). But this follows from the two equations :

(A - T)R(A)T’ = T’

( A -T)T’R(A) = T’,
since (A - T) is invertible, A E p ( T )

3. The Invariant SubsDace Problem.

Let T be an operator on a Banach space E. We say that a subspace F
of E is invariant by T if TF c F. Does there exist a closed linear subspace
F, invariant by T ,besides the trivial ones : F = E and F = ( 0 ) ?
(In old mathematical papers, the word “stable” was used instead of “in-
variant”. It s e e m better, but I take the modern terminology.)
This question arises naturally from the theory of eigenvectors in finite
dimensional spaces : if z is an eigenvector, of course F = {Az ; X E Q: } is such
a subspace. But we have seen an easy example of an operator on l z ( Z ) with no
eigenvalue. So, some other concept has to be substituted for it, and the concept
of non-trivial invariant subspace is the broadest and the most natural one (see,
however, the concept of Invariant Convez Subset, introduced in Exercise 10, at
the end of this Chapter).
A stronger concept is that of hyperinvariant subspace : this is a subspace
which is also invariant by all operators which commute with T.
If we fix a point x in E and consider the subspace :

. .}
F, = s i i { z , T z ,T’z,.

we obtain a closed subspace, invariant by T. Such a subspace will be called

elementary invariant subspace. If F, is the whole space, the point z is said to
be cyclic, it is non-cyclic otherwise. Of course, any Invariant Subspace contains
an elementary Invariant Subspace, and, in fact, is just the reunion of all the
elementary Invariant Subspace that it contains.
 Elementary Spectral Theory 33

A similar concept occurs for the Hyperinvariant Subspaces : instead of

just taking the iterates of T, we take all operators commuting with T ,and
define :
G, = W { A z ; A commutes with T )
Then G , is an elementary hyperinvariant subspace.
In the sequel, we will be looking for non-trivial Invariant Subspaces or
Hyperinvariant Subspaces. We will often omit the words "non-trivial", and ab-
breviate the four words in "Invariant Subspace" or "Hyperinvariant Subspace".
T h e Invariant Subspace problem is still open in Hilbert spaces. Its exis-
tence is the background for the whole book, and there is not a single chapter
which does not contain a part devoted to it.
Let's come back to spectra, and justify the claim we made a t the end
of Section 1. If a,(T) is not empty, T has eigenvectors, and so Invariant
Subspaces, as we already said. If a,(T) is not empty, the set :

F = Im(X-T)

is non-trivial, and is invariant under T. This proves our assertion : if one

is looking for Invariant Subspaces, one may assume o ( T ) = a,(T) \ u P ( T ) .
One can moreover, in a reflexive space, assume that any sequence of almost
eigenvectors converges weakly to 0 (Proposition 1.13).
If the spectrum u(T) is not connected, the analytic functional calculus
allows us to exhibit Invariant Subspaces :
Theorem 3.1 (F. Riesz). - If a ( T ) = U I U U ~ where
, 01 and a 2 are disjoint

closed sets, then T has Hyperinvariant Subspaces Fl and Fz ; moreover :

0 1 = O(TIF1) 02 = U(TlF2) 1

where T ( F denotes the restriction of T to the subspace F.

Proof. - Let Ul , Uz be disjoint open sets, containing u1 and a2 respectively.
Let e l = 1 in UI , 0 in U z , and e 2 = 0 in Ul , 1 in U 2 . Then e l and ez are
in F ( T ) ,and :
el( T ) + e z ( T ) = I

ei(T) = el(T) , C;(T)

= ez(T).
) projections, which commute with T. Their ranges
So e l (T ) and c ~ ( Tare
F1 and Fz are the required subspaces : They are closed, non-trivial, and
hyperinvariant, by Proposition 2.7.
$4 Chapter I1

To prove the second statement, we observe that, on Fl , T = T c l ( T ) . So

= g ( u ( T ) )where
u(Tel(T)) , g(X) = X c l ( A ) , by Theorem 2.3. But g ( a ( T ) )=
UI . This proves the Theorem.
Remark. - In the above Theorem, under the assumption that the spectrum
is not connected, we exhibit two functions c1 , c 2 , such that el (T)e2(T) = 0 ,
and none of the operators e l ( T ) or e 2 ( T ) is 0 . One should not think that
the assumption is, by any means, necessary for the conclusion. Indeed, for the
right shift S on 12(Z),if we define . I ( # ) = l ~ O , r ~ (a2(#)
0 ) l = llr,2rl(B)and
set e l = cial , c2 = eiaa, we have el(S) # 0 , e 2 ( S ) # 0 (as we will see in
Chapter VII), and el(S)ez(S)= 0. But in this case, e l ( S ) and e 2 ( S ) are
defined by other means than the Riesz’s functional calculus (see Chapter VII).
So the assumption of non-connectedness is linked with the tools we are using,
and not with the conclusion.
 Elementary Spectral Thwry 35

Exercises on Chapter 11.

Exercise 1. - Determine u p , u,, u, for the operator T from C [ O , l ) into
itself :

Same question if T is considered as an operator on 151 [0,1].

Exercise 2. - Determine up, u r , u, for the operator T from C[O,11 into
itself :
/(l) --* 9(t) = tN)
Same question if T is considered as an operator on 151[0,1].
Exercise 3. - Show that there exists a polynomial P such that P ( T ) = 0 if
and only if the spectrum of T consists in a finite number of eigenvalues.
Exercise 4. - Assume that I r' < r , and that f is analytic on the closed
disk D ( 0 , r ) . If moreover [!(A)] 5 R for every X with = r , show that :

Exercise 5. - Let T be an operator, C a set contained in p ( T ) . Set d =

d i s t ( C , u ( T ) ) .Show that there exists a K 2 1 such that :
IIR(X,T)II" 5 K / E " , n 2 0, X E C .
Exercise 6 : Upper semi-continuity of the spectrum. - If A, is a family of
sets, we define :
limsup A, = n k U , > k A k .
Let 4 be the function on L ( E ) : #(T)= o ( T ), with values into the subsets of
Q: . Prove that # is upper semi- continuous, :hat is :
if T , + T, limsupu(T,) c a(T).
[Hint : Take X E limsupa(T,). Show first that there exists a sequence Ank E
u(Tnk) such that A,, -P A , then use Cor. 1.8.1
Exercise 7 : T h e spectrum is not continuous. - Let sk be the weighted shift
on l z ( 2 ) defined by :
Ske, = if n f 0

1) Prove t h a t U ( S k ) is the unit circle (follow the proof of Proposition 1.13).

2) Prove t h a t the sequence s k converges t o the operator A defined by Ae, =
en+l (n # O ) , A ~ =oO .
3) Prove t h a t u ( A ) is the closed unit disk.
96 Chapter I1

Exerche 8 . - Assume that the spectrum of T does not intersect the open disk
D ( 0 , a ) . Show that :
a 5 l/limsuplIT-nII'/n,
and that, inside this disk, the resolvent is given by the formula :

Exercise 9. - Let T be an operator such that llTl\ 5 1 and 1 4 a(T).Show

that the sequence (C,"T k ) is~bounded.
Exercise 10. - Let T be an operator for which there exists M such that ;

n-I

Show that for every r , 0 < r < 1,

+ca

Exercise 11. - a) Let (an)n>O- be a sequence of real numbers such that there
exists a 6 > 0 and a C > 1 with :

a, 2 6 . C n , n20

Show that :
an 1
lim inf - 5 C '
n-- %+l

b) Let T be an invertible operator on a Banach space E . Assume that

there exists C > 1 such that for all z # 0 in E l there is a 6(z) > 0 with :

Show that, for every E > 0, there is a n such that :

c) Deduce that :
 Elementary Spectral Theory 97

Exercise 12. Let T be an operator on a reflexive Banach space E l with

~

IlTll = 1 and 1 E a ( T ) .
a) Show that, for any A , p such that 1 + X E p ( T ) , 1+ p E p ( T ) ,
XR(X + 1 ) ( 1 - p R ( p + 1)) = XR(p + 1 ) ( 1 - XR(X + 1))

b) Let X EIR, 0 < X 5 1 . Sho - that :

c) Fix z E E . Let (A,),, be a sequence converging to 0, with 0 < A, <

1, such that X,R(X, +
1)" converges weakly to some point y . Let p , 0 < p 5
1. Let P be the weak limit of some subsequence of p R ( p 1)X,R(X, + 1)z. +
Using a), show that z = y , Deduce that :

d) Deduce that T y = y (write T = - ( ( p + 1) - T )+ {p + 1)1).

e) Let

X = {u E E ; lim pR(p
P+O
+ 1)" exists for the norm}

Show that, for every X > 0 ,

pR(p + l)(z + XR(X + 1)z) + 0 , p + 0

(use a), exchanging the roles of X and p ) .

Deduce that
xtXR(Xt1)z E x
Deduce that z + y E X , and then that zE X.
f ) Deduce that, if 1 is not an eigenvalue of T , for all z E E ,

XR(X+ l ) z + 0 , X+O.

Exercise 13 : Closed Invariant Convex Subset of the Unit Ball. - We investi-

gate the following question : if T is a non-zero contraction (that is IlTll 5 1)
on a Banach space, does there exist a closed convex subset C of the unit ball
B E , with C # {0}, C # BE , and TC c C ? This problem is weaker than
the Invariant Subspace problem, but obviously related. Surprinsingly enough,
38 Chapter 11

it has never been considered. The answer is uyes", and is simple to obtain.
We give two proofs. The first one is constructive, and is given now. The second
one is posponed until Chapter XIV.
a) Solve the problem when T is an isometry, and when r(T) < 1.
b) Assume that T is not an isometry, and that r(T) = 1. Take a vector
20,11z0(1
= 1, such that :

Consider :
= { Y E BE ; IIT"Yll 5 llTnzoII , v n IN}
Show that C is convex, closed, invariant under T and not reduced to ( 0 ) .
c) Show that C is not the whole unit ball (consider a X, 1x1 = 1, in the
spectrum of T , and take a sequence ( z k ) k >-o of almost eigenvectors for A).
Observe that C is also different from any ME,for XE (I: , 1x1 = 1.

Notes and Comments.

Most results in Section 1 can be found in any book dealing with spectral
theory ; let's mention Dunford-Schwartz [l], vol. I, and H.R. Dowson [l],
among others. Proposition 1.14 was pointed out to us by V. Martel. The
analytic functional calculus presented in Section 2 is due t o F. Riesz. For both
sections, our presentation follows Dunford-Schwartz, with minor changes.
For weighted shifts and their spectra, an excellent reference is the survey
paper by A. Shields 111.
For exercises on Hilbeit spaces, the best source is the book by P. Halmos
[Z]. Exercises 6 and 7 come from there. Exercise 12 follows a nice work of R.
Emilion [ll. Exercise 13 is due t o the author.
 Elementary Spectral Theory 99

Complements on Chapter 11.

There are links between the growth of the sequence llT-nll, R + +oo
(assuming T to be invertible), and the growth of the resolvent R ( X , T ) ,when
X approaches the spectrum of T. As an example, we mention the following
result, due to A. Atzmon [I]:
Proposition. - Let T be an operator on a Banach space E , with o(T)c 2.
If:
a) There exists c > 0, a , O < a < 1, such that :

llTp"ll = O(ezp(cna)) , n -+ +m,

then :

withp=A 1-a
, d = 48ca1/3.
Conversely, if b) holds with some constants d > 0 , p > 0 , a) holds with
a= A,
c = 3d"iP.
Techniques connected with the growth of the resolvent will be used in Part
VI, Chapter I, in order to obtain invariant subspaces. Techniques connected
with the growth of the backwards iterates will be used in Part V, Chapter 111,
for the same purpose.