Operator Theory and Harmonic Analysis: David Albrecht, Xuan Duong and Alan M Intosh

Download as pdf or txt
Download as pdf or txt
You are on page 1of 60

OPERATOR THEORY AND HARMONIC ANALYSIS

DAVID ALBRECHT, XUAN DUONG and ALAN McINTOSH

Lectures presented at the Workshop in Analysis and


Geometry, C.M.A., A.N.U., Canberra, Jan. { Feb. 1995
Contents

1. Spectral Theory of Bounded Operators


(A) Spectra and resolvents of bounded operators on Banach spaces
(B) Holomorphic functional calculi of bounded operators
2. Spectral Theory of Unbounded Operators
(C) Spectra and resolvents of closed operators in Banach spaces
(D) Holomorphic functional calculi of operators of type S!+
3. Quadratic Estimates
(E) Quadratic norms of operators of type S!+ in Hilbert spaces
(F) Boundedness of holomorphic functional calculi
4. Operators with Bounded Holomorphic Functional Calculi
(G) Accretive operators
(H) Operators of type S! and spectral projections
5. Singular Integrals
(I) Convolutions and the functional calculus of ?i dxd
(J) The Hilbert transform and Hardy spaces
6. Calderon{Zygmund Theory
(K) Maximal functions and the Calderon{Zygmund decomposition
(L) Singular integral operators
7. Functional Calculi of Elliptic Operators
(M) Heat kernel bounds
(N) Bounded H1 functional calculi in Lp spaces
8. Singular Integrals on Lipschitz Curves
(O) Convolutions and the functional calculus of ?i dzd j
(P) The Cauchy integral on a Lipschitz curve and Hardy spaces
1991 Mathematics Subject Classi cation. 42-01, 42B20, 47-01, 47A10, 47A60.
The authors were supported by the Australian Research Council as well as by Macquarie University.
Typeset by AMS-TEX
1

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

Parts of the rst four lectures are based on notes of previous lectures of Alan

McIntosh, which were taken, edited, typed and re ned by Ian Doust and Elizabeth
Mans eld, whose willing assistance he gratefully acknowledges.

It is assumed that the reader has a basic knowledge of metric spaces, topology,
measure theory, and the theory of bounded linear operators on Banach and Hilbert
spaces. Suitable references for this material are the books \Real and Complex Analysis"
by W. Rudin, \Real Analysis" by H.L. Royden, \Introduction to Topology and Modern
Analysis" by G.F. Simmons, \Functional Analysis" by F. Riesz and B. Sz.-Nagy, and
\Linear Operators, Part I, General Theory" by N. Dunford and J.T. Schwartz. Later,
we shall also expect some knowledge of Fourier theory and partial di erential equations.
Lecture 1. Spectral Theory of Bounded Operators

Much of the material in the rst two lectures is presented in greater detail in the
books \Perturbation Theory for Linear Operators" by T. Kato and \Spectral Theory"
by E.R. Lorch, as well the book by Dunford and Schwartz just mentioned.
In these two lectures, X denotes a non{trivial complex Banach space, while H
denotes a non{trivial complex Hilbert space.

(A) Spectra and resolvents of bounded operators on Banach spaces.


Let T be a bounded operator on X . The norm of T is

k T k = sup f k Tu k : u 2 X ; k u k = 1 g :
The Banach algebra of all bounded operators on X is denoted by L(X ) . As with
matrices,  2 C is called an eigenvalue of T if there exists a non{zero vector u 2 X
such that Tu = u . The resolvent set (T ) of T is the set of all  2 C for which
(T ? I ) is a one{one mapping and (T ? I )?1 2 L(X ) . The spectrum (T ) of T is
the complement of (T ) . Clearly every eigenvalue of T lies in (T ) .

In a nite dimensional space, every one{one operator is an isomorphism, and (T )


is precisely the nite set of eigenvalues of T .
For  2 (T ) , the resolvent operator RT () 2 L(X ) is de ned by

RT () = (I ? T )?1 :


These operators satisfy the resolvent equation

RT () ? RT () = ( ? ) RT () RT ()


for all  ,  2 (T ) . Thus, in particular, RT () and RT () commute.

OPERATOR THEORY AND HARMONIC ANALYSIS

Example. Let lp denote the Banach space of all sequences

x = (x1 ; x2 ; : : : ; xn ; : : : )
of complex numbers, with nite norm
1

kxk =

n=1

jxn

jp

=p

!1

Given a sequence d = (d1; d2 ; : : : ) such that supn jdnj < 1 , de ne the operator
D = diag(d) on lp by (Dx)n = dnxn for every n . Then D 2 L(lp ) and (D) is the
closure of the set fd1; d2; : : : g in C . Moreover


RD () = diag  ? d ;  ? d ; : : :
1
2

for all  2 (D) .


We can treat in nite series of vectors (or operators) in the same way that we treat
in nite
series of complex numbers. For example, we
say that an in nite series of operaP
P
tors
T
is
absolutely
convergent
if
the
series
k
Tn k Pis convergent, in which case
n
P
P
Tn is necessarily convergent in L(X ) and k Tn k  k Tn k .
Also several of the de nitions for real and complex functions can be extended to
vector valued (or operator valued) functions u(t) de ned for a real or complex variable
t and taking values in X (or L(X ) ). For instance limt!a u(t) = v means that for
every " > 0 there exists  > 0 such that k u(t) ? v k < " whenever 0 < jt ? aj <  .
A function u is continuous in a region E if limt!a u(t) = u(a) for every a 2 E . We
de ne u0(t) to be the derivative of u(t) whenever the limit

u(t + h) ? u(t)
(
t
)
=
lim
u0 (t) = du
h!0
dt
h
exists. For a piecewise Rcontinuous function u de ned on a recti able curve , we de ne
the Riemann
integral u(t) dt to be the limit in X (or in L(X ) ) of the appropriate
P
sums j ( (tj ) ? (tj?1 ))u(tj ) .
Moreover, several of the formulae such as
Z

d ( u(t) + v(t)) = u0 (t) + v0(t)


dt
Z
Z
( u(t) + v(t)) dt = u(t) dt + v(t) dt
Z


u(t) dt 

k u(t) k jdtj

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

which are valid for real valued functions, also hold for vector valued and operator valued
functions. Also, for holomorphic functions (i.e. functions which are de ned and di erentiable everywhere on an open subset of the complex plane), Cauchy's integral theorem
holds and there is a Taylor expansion about every point in that set. If a function is
bounded and holomorphic on the entire complex plane, then Liouville's theorem holds,
meaning that such a function is constant.
Whenever k T k < 1 then (on de ning T 0 = I )
(I ? T )?1 =

n=0

Tn

and the series is absolutely summable. A consequence of this is the fact that, if  2 (T )
and
j ? j < k RT () k?1 ;
then
1
X
?
1
RT ( ) = (I ? ( ?  )RT ()) RT () = ( ?  )nRT ()n+1 :
n=0

This shows that (T ) is an open subset of C , and that RT ( ) is holomorphic in 


with the above Taylor expansion about each point  2 (T ) . Hence

dn R ( ) = (?1)n n! R ( )n+1 :
T
d n T

Theorem A. Suppose that T is a bounded operator on X . Then (T ) is a non{


empty compact subset of

Proof. In view of the above remarks, it suces to show that (T ) is bounded and
non{empty. Take any j j > k T k . Then

RT ( ) =  ?1(I ?  ?1T )?1 =

n=0

 ?n?1T n

and

k RT ( ) k  j j ?1k T k :
Therefore (T ) is bounded and RT ( ) ! 0 as  ! 1 . If (T ) were empty, then

RT ( ) would be a bounded entire function on C which tends to zero at in nity, and


therefore, by Liouville's theorem, RT ( ) = 0 for every  2 C . However, this is
impossible as the inverse of an operator on a non{trivial Banach space cannot be zero.
We conclude that (T ) is a non{empty subset of C . 

OPERATOR THEORY AND HARMONIC ANALYSIS

(B) Holomorphic functional calculi of bounded operators.


Let T be a bounded operator on X . Let
be an open subset of C which
contains (T ) , and let H (
) denote the space of all complex valued holomorphic
functions de ned on
.
Let us use the word contour to mean a nite collection of oriented smooth closed
curves, and let us say that a contour envelopes (T ) in
if is contained in

n (T ) and

1 Z d = 1 ; if 2 (T )
2i  ?
0 ; if 62
:
For each function f 2 H (
) , de ne the operator f (T ) 2 L(X ) by
Z
Z
1
1
?
1
f (T ) = 2i f ( )(I ? T ) d = 2i f ( )RT ( ) d

where is a contour which envelopes (T ) in


.
Since envelopes (T ) in
, and the integrand is a holomorphic operator valued
function on
n (T ) , it follows that f (T ) 2 L(X ) , and by Cauchy's theorem, that the
de nition of f (T ) is independent of the particular choice of .
The terminology f (T ) is natural, in view of the following result, which can be
found in Chapter VII of the book already referred to by Dunford and Schwartz. In the
Notes and Remarks at the end of that Chapter, there is an account of the historical
development of this theory. People involved include E.H. Moore, D. Hilbert, F. Riesz,
C. Neumann, I.M. Gelfand, N. Dunford and A.E. Taylor.

Theorem B. Let T 2 L(X ) , and suppose that


is an open set which contains (T ) .
Then the mapping from H (
) to L(X ) which maps f to f (T ) satis es the following
properties:
(1) if f , g 2 H (
) and 2 C then f (T ) + g(T ) = (f + g)(T ) ;
(2) if f , g 2 H (
) then f (T )g(T ) = (fg)(T ) ;
P
P
(3) if p( ) = nk=0 ck  k then p(T ) = nk=0 ck T k ;
(4) if 2 (T ) and R ( ) = ( ? )?1 then

R (T ) = (T ? I )?1 = ?RT ( ) ;
(5) if ff g is a net in H (
) which converges uniformly on compact subsets of

to f 2 H (
) then f (T ) = lim f (T ) ;
(6) if f 2 H (
) then f ((T )) = (f (T )) .

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

Proof.
(1) This follows from the linearity of the integral.
(2) Since the de nitions of f (T ) and g(T ) are independent of the contours which
envelope (T ) in
, we are free to chose suitable ones for our purposes. Let
U be an open subset of
which contains (T ) , and whose closure U is a
compact subset of
. Let ? and be contours such that ? envelopes (T )
in U , and envelopes U in
. Then, by the resolvent equation, we have
Z
Z
1
1
f (T )g(T ) = 2i f ( )RT ( ) d 2i g(w)RT (w) dw
Z Z
?
1
= (2i)2
f ( )g(w)RT ( )RT (w) d dw
?
Z
Z
1
= (2i)2 f ( )RT ( ) wg(?w) dw d

?
Z
Z
1
? (2i)2 g(w)RT (w) wf (? ) d dw
?

Z
= 0 + 21i f (w)g(w)RT (w) dw
?
= (fg)(T ) :

(We can interchange the order of integration because the integrals are absolutely
convergent.)
(3) It follows from (1) and (2) that it suces to prove this result for the functions
1 and Id de ned by 1( ) = 1 and Id( ) =  for all  2 C . Let be a
contour which envelopes the closed disc of radius k T k , and hence (T ) , in C .
Then
Z
1
1(T ) = 2i (I ? T )?1 d

Z
1 Tn
X
1
d
= 2i
n=0  n+1
Z
1
X
1
n
=
T 2i  nd+1

n=0
= I
since

2i if n = 0
0 otherwise.

Using the same argument, we see that
Z

d =
 n+1

Z
Z
1
X
1
1
?
1
n
T 2i dn = T :
Id(T ) = 2i  (I ? T ) d =


n=0

OPERATOR THEORY AND HARMONIC ANALYSIS

(4) Let 2 (T ) . Then, by the above properties, we obtain:


(T ? I )R (T ) = 1(T ) = I
R (T )(T ? I ) = 1(T ) = I:
Therefore R (T ) = (T ? I )?1 .
(5) Since has nite length and
Z


(f ( ) ? f ( ))RT ( ) d

 max
f jf ( ) ? f ( )j k RT ( ) k g length( ) ;
 2

it follows that f (T ) = lim f (T ) .


(6) Suppose f 2 H (
) . Let  2 (T ) . For  2
, let

f ( ) ? f () when  6= 
g( ) = :  ? 
f 0 ()
when  =  :
8
<

Then g 2 H (
) and f (T ) ? f ()I = g(T )(T ? I ) . If f () 2 (f (T )) then
f (T ) ? f ()I has a bounded inverse, and hence so does (T ? I ) , which contradicts the assumption that  2 (T ) . Therefore f () 2 (f (T )) .
Now let  2 (f (T )) . If  62 f ((T )) then h( ) = (f ( ) ? )?1 is
holomorphic on a neighbourhood of (T ) , say
0 . Applying the above results
to H (
0 ) , we get h(T )(f (T ) ? I ) = I , which contradicts the assumption that
 2 (f (T )) . Therefore  2 f ((T )) . 
We say that T has a bounded H (
) functional calculus (\bounded" because

k f (T ) k  c k f k1 for all f 2 H (
) ). There is a straightforward consequence concerning spectral decompositions.

Suppose that T is an operator whose spectrum (T ) is a pairwise disjoint union


of non{empty compact sets

(T ) = 1 [ 2 [    [ N :
For each k , let
k be an open subset which contains k , chosen so that the
k 's are
pairwise disjoint, let
= [k
k , and de ne the holomorphic functions k 2 H (
) by

 k ( ) =

1 if  2
k
0 otherwise:

Then  k 2 =  k ;  k  j = 0 if k 6= j , and

N  = 1.
k=1 k

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

Thus the operators Pk de ned by Pk = k (T ) form a family of spectral projections


associated with T , meaning that they satisfy

Pk = Pk ; Pk Pj = 0 if k 6= j ; and
2

k=1

Pk = I

as well as Pk T = TPk .
It is easy to infer from this, that the range Xk of Pk is a closed subspace of X
which is invariant under T , that X = k Xk , and that, if Tk denotes the restriction of
T to Xk , then Tk 2 L(Xk ) and (Tk ) = k . Thus we have a spectral decomposition
of X associated with T .
In the case when X is nite dimensional and the sets k consist of distinct eigenvalues, then this decomposition of X into Xk can be used to obtain the Jordan canonical
form of T , as is shown, for example, in Kato's book.
Exercise. For the operator D in the Example in Section A, show that f (D) =
diag(f (d1 ); f (d2 ); : : : ) . In this case, k f (D) k = sup2(D) jf ()j  k f k1 . Actually this operator D has a bounded F functional calculus, where F is the space of
all bounded complex valued functions de ned on fdng .
Exercise. Prove that the functional calculus de ned above is unique in the following
sense. There is no other mapping from H (
) to L(X ) which satis es properties (1),
(2), (5), and either (3) or (4) of Theorem B.
Exercise. For each t 2 R , let gt denote the holomorphic function gt( ) = e?t . De ne
e?tT 2 L(X ) by e?tT = gt(T ) . Use the holomorphic functional calculus of T and the
properties of the exponential functions to obtain whatever properties you can concerning
the operators e?tT . Show that the unique C 1 function u : R ! X which satis es

du (t) = ?Tu(t) ; u(0) = v


dt
is given by u(t) = e?tT v .

OPERATOR THEORY AND HARMONIC ANALYSIS

Lecture 2. Spectral Theory of Unbounded Operators

Material in Section D and subsequent sections is based on my paper [Mc] on H1


functional calculi, and on subsequent papers with A. Yagi [McY], T. Qian [McQ], M.
Cowling, I. Doust and A. Yagi [CDMcY], and P. Auscher and A. Nahmod [AMcN], all
of whom I take this opportunity to thank.

(C) Spectra and resolvents of closed operators in Banach spaces.


Recall that X denotes a non{trivial complex Banach space. By an operator in
X we mean a linear mapping T : D(T ) ! X , where the domain D(T ) is a linear
subspace of X . (Note that we do not require that subspaces are closed.) The range of
T is denoted by R(T ) . The norm of T is
k T k = supf k Tu k : u 2 D(T ); k u k = 1 g
so that 0  k T k  1: We say that T is bounded if k T k < 1 , and that T is bounded
on X if it is bounded and D(T ) = X . The algebra of all bounded operators on X is
denoted by L(X ) . We call T densely{de ned if D(T ) is dense in X , and closed if its
graph, G (T ) = f(u; Tu) : u 2 D(T )g is a closed subspace of X  X . The space of all
closed operators is denoted C (X ) .
A complex valued function f de ned on X is called a conjugate linear functional
 (v) , for every , 2 C and u , v 2 X . The
on X if f ( u + v) = f (u) + f
complex Banach space X  of all conjugate linear functionals on X , under the norm
k f k = supf jf (u)j : k u k = 1 g ;
is called the adjoint space of X .
If T is a densely{de ned operator in C (X ) , then its adjoint operator T  2 C (X )
is de ned as follows. The domain of T  consists of all g 2 X  such that g(Tu) = f (u)
for every u 2 D(T ) and some f 2 X  . Since T is assumed to be densely{de ned, the
f 2 X  is uniquely de ned by g , and so we de ne T  g = f .
In the case that X is a Hilbert space, we can identify X  with X by taking
f (x) = hf; xi . If T is a densely{de ned operator in X such that T = T  , then T is
called a self{adjoint operator in X .
When new operators are constructed from old, the domains are taken to be the
largest for which the constructions make sense. For example, if S and T are linear
operators, then S + T and ST are the linear operators de ned by
(S + T )u = Su + Tu ; u 2 D(S + T ) = D(S ) \ D(T ) ;
(ST )u = S (Tu) ; u 2 D(ST ) = f u 2 D(T ) : Tu 2 D(S )g ;

10

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

and, if T is one{one, then

D(T ?1 ) = R(T ) :

Care needs to be taken with the domains of operators. We write S  T if D(S ) 


D(T ) and Su = Tu for all u 2 D(S ) . So S = T if and only if S  T and T  S .
Exercise. Show that

and, if S is one{one,

S+T
(S + T ) + U
S+0
0S
S ?S
(ST )U
S (T + U )
(S + T )U

= T +S ;
= S + (T + U ) ;
= S;
 S0 = 0 ;
 0;
= S (TU ) ;
 ST + SU ;
= SU + TU ;

S ?1S  I and SS ?1  I :
Sometimes S (T + U ) 6= ST + SU , as the example given by T = ?U = I and D(S ) 6= X
shows. If S and T are one{one, then so is ST , and (ST )?1 = T ?1 S ?1 .

Exercise. If B 2 L(X ) and T 2 C (X ) , then the following operators are closed: B ,


TB , B?1T (if B is one{one) and T ?1 (if T is one{one).

The resolvent set (T ) of T is the set of all  2 C for which (T ? I ) is one{
one and (T ? I )?1 2 L(X ) . The spectrum (T ) of T is the complement of (T ) ,
together with 1 if T 2= L(X ) .
For T 2 C (X ) and  2 (T ) , de ne the resolvent operator RT () 2 L(X ) by
RT () = (I ? T )?1 :

Lemma C. Suppose T is a closed operator in X and  2 (T ) . Let R( ) =


( ? )?1 for every  2 C , and let R(1) = 0 . Then ?R((T )) = (RT ()) .
Proof. Suppose 1 2 (T ) . Then T 62 L(X ) and hence T ? I 62 L(X ) . Therefore

R(1) = 0 2 (RT ()) :


Next, let  2 (T ) \ C and assume that ?( ? )?1 = ( ? )?1 2 (RT ()) . Then

?1
1
1
P =  ?  RT () RT () ?  ?  I
2 L(X ) :

OPERATOR THEORY AND HARMONIC ANALYSIS

11

Therefore P (T ? I )  I and (T ? I )P = I , so that  2 (T ) , which contradicts


the assumption that  2 (T ) . Therefore ?R((T ))  (RT ()) .
Now, suppose  2 (RT ()) and assume that ? 62 R((T )) . Then  2 C and
 ? 1= 2 (T ) . Let
Q = ? 1 (I + 1 RT ( ? 1=)) 2 L(X ) :
Then Q(RT () ? I ) = (RT () ? I )Q = I . So  2 (RT ()) . However, this is a
contradiction, and hence (RT ())  ?R((T )) . 
In the above, we have extended the complex plane

to the Riemann sphere C 1 =

[ f1g . The topology on the Riemann sphere consists of the usual neighbourhoods
in C together with the neighbourhoods of 1 which are the sets of the form U [ f1g ,
where C n U is a compact subset of C . A function f de ned in a neighbourhood of
1 is holomorphic at 1 if f (1= ) is holomorphic at  = 0 .
Let T be a closed operator in X and let  2 (T ) . Then the function R de ned
in the last lemma, is a homeomorphism on C 1 and is holomorphic on C 1 n fg . It

therefore follows from the last lemma that (T ) is a non{empty compact subset of
C1 .

We now develop the holomorphic functional calculus for closed operators with non{
empty resolvent set, following the treatment of Dunford and Schwartz.
Let
be an open subset of C 1 which contains (T ) , but does not contain  2 C ,
and let H (
) be the space of holomorphic functions de ned on
. Then R induces
a bijection between H (
) and H (R(
)) , where f 2 H (
) is mapped to f  R? 1 .
Using this bijection we de ne
f (T ) = (f  R? 1 )(?RT ()) ;
for every f 2 H (
) . The following result is then a consequence of Theorem B.

Theorem C. Let T be a closed operator in X with non{empty resolvent set, and


suppose that
is a proper open subset of C 1 which contains (T ) . Then the mapping
from H (
) to L(X ) which maps f to f (T ) satis es the following properties:
(1) if f , g 2 H (
) and 2 C then f (T ) + g(T ) = (f + g)(T ) ;
(2) if f , g 2 H (
) then f (T )g(T ) = (fg)(T ) ;
(3) 1(T ) = I ;
(4) if 2 (T ) and R ( ) = ( ? )?1 then
R (T ) = (T ? I )?1 = ?RT ( ) ;
(5) if ff g is a net in H (
) which converges uniformly on compact subsets of

to f 2 H (
) then f (T ) = lim f (T ) ;
(6) if f 2 H (
) then f ((T )) = (f (T )) .

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

12

Exercise. State and prove a theorem concerning the uniqueness of this holomorphic
functional calculus.
Exercise. Prove A.E. Taylor's formula
Z
1
f (T ) = f (1)I + 2i f ( )(I ? T )?1 d


in the case when 1 2 (T ) and  is a contour which envelopes (T ) in


in some
appropriate sense.

(D) Holomorphic functional calculi of operators of type S!+ .


For 0  ! <  <  , de ne the closed and open sectors in the complex plane

S!+ = f  2 C : j arg  j  !g [ f 0g
S0+ = f  2 C :  6= 0; j arg  j < g:
We employ the following subspaces of the space H (S0+) of all holomorphic functions on S0+ .
H1 (S0+) = f f 2 H (S0+) : k f k1 < 1 g ;
where k f k1 = supf jf ( )j :  2 S0+ g;
(S0+) = f 2 H (S0+) : 9s > 0; j ( )j  C j js(1 + j j2s)?1 g
and
so that

F (S0+) = f f 2 H (S0+) : 9s > 0; jf ( )j  C (j j?s + j js) g ;


(S0+)  H1 (S0+)  F (S0+)  H (S0+) :

Let 0  ! <  . An operator T 2 C (X ) is said to be of type S!+ if (T )  S!+


and, for each  > ! , there exists C such that

(T ? I )?1  Cj j?1 ;  62 S+ :


Example 1. Suppose that T is a self{adjoint operator in a Hilbert space H and that

hTu ; ui  0 for every u 2 X . Then T is an operator of type S0+ .

Example 2. More generally, for 0  !  2 , suppose that T is an ! -accretive operator in a Hilbert space H , by which we mean that (T )  S!+ and hTu ; ui 2

OPERATOR THEORY AND HARMONIC ANALYSIS

13

S!+ for all u 2 D(T ) . Then it is easy to show that k RT ( ) k  (dist(; S!+))?1 for
all  2= S!+ , so that T is an operator of type S!+ . We remark that ! -accretive
operators necessarily have dense domain, and are self{adjoint if and only if ! = 0 .
Example 3. Let 0 < ! < 2 . A family of bounded operators fS (z) : z 2 S!0 +g on X
is called a holomorphic semigroup if
(1) S (z)S (w) = S (z + w) for all z , w 2 S!0 + ;
(2) S (z) is holomorphic on S!0 + .
Let T be the generator of fS (z)g , namely the operator in X such that

S (t)u ? u
Tu = lim
t#0
t
with domain consisting of all u 2 X for which the limit exists. Then ?T is an operator
of type S(  ?!)+ in X .
2

Example 4. De ne

?k
T =  2 0 2?1 k
k2N
in the Hilbert space H = PC 2 = f(u1 ; u2; : : : ) : uk 2
((u1 ; u2; : : : ) ; (v1 ; v2 ; : : : )) = (uk ; vk ) . Then, for  < 0 ,


(T ? I )?1 =

C2

g with inner product

(2?k ? )?1 ?(2?k ? )?2 ;



0
(2?k ? )?1
k2N


so (T ? I )?1  supk 2?k ?  ?2  jj?2 . Thus T is not of type S!+ for any ! .

Suppose that T is a one{one operator of type S!+ with dense domain and dense
range in X . Every such operator has a holomorphic functional calculus which is consistent with the usual de nition of polynomials of an operator. By this we mean that for
each  > ! and for each f 2 F (S0+) , there corresponds a densely{de ned operator
f (T ) 2 C (X ) . Further, if f; g 2 F (S0+) and 2 C , then

f (T )u + g(T )u = ( f + g)(T )u
g(T )f (T )u = (gf )(T )u
for every u 2 D(f (T )) for which one side (and hence the other side) of each equation
is de ned.
In particular, if

2 (S0+) , then
Z
1
(T ) = 2i (T ? I )?1 ( ) d

14

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

where is the unbounded contour f = rei : r  0g parametrised clockwise around


S!+ , and ! <  <  . Clearly, this integral is absolutely convergent in L(X ) , and it
is straightforward to show, using Cauchy's theorem, that the de nition is independent
of the choice of  2 (!; ) , and that if is holomorphic on a neighbourhood of S!+ ,
then this de nition is consistent with the previous one. Moreover ( (T )) = ((T )) .
One way to construct this functional calculus is to rst de ne these operators (T )
by this integral when 2 (S0+) , and to then de ne f (T ) when f 2 F (S0+) as
follows.
Suppose f 2 F (S0+) . Then for some c and k , jf ( )j  c(j jk + j j?k ) for every
 2 S0+ . Let

k+1

( ) = (1 +  )2
:
Then , f 2 (S0+) and (T ) is one{one. So (f )(T ) is a bounded operator on
X , and (T )?1 is a closed operator in X . De ne f (T ) 2 C (X ) by

f (T ) = ( (T ))?1 (f )(T ) :
One must then check that the properties of the functional calculus are satis ed.
This functional calculus is the unique one in which the following Convergence
Lemma holds. It is one of the most important properties which is satis ed by this
functional calculus.

Theorem D (The Convergence Lemma). Let 0  ! <    . Let T be an


operator of type S!+ which is one{one with dense domain and range. Let ff g be
a uniformly bounded net in H1 (S0+) , which converges to f 2 H1(S0+) uniformly
on compact subsets of S0+ , such that ff (T )g is a uniformly bounded net in L(X ) .
Then f (T ) 2 L(X ) , f (T )u ! f (T )u for all u 2 X , and k f (T ) k  sup k f (T ) k .
Proof. Let ( ) =  (1 +  )?2 and ( ) = ( )(f ( ) ? f ( )) . Then, for each ,
2 (S0+) and

jf ( ) ? f ( )j jd j
1 + j j2

 const.
r (k f k1 + k f k1) + const. ? sup fjf ( ) ? f ( )jg

k (T ) k  const.

r 1 j jr

for all r > 1 . Since the net ff g is uniformly bounded and converges on compact
subsets of S!0+ , it follows that k (T ) k ! 0 . Now take any u 2 X . Then

k f (T ) (T )u ? f (T ) (T )u k = k (T )u k  k (T ) k k u k ! 0 :

OPERATOR THEORY AND HARMONIC ANALYSIS

15

Therefore

k f (T ) (T )u k  (sup k f (T ) k) k (T )u k :

Since (T ) has dense range it follows that k f (T ) k  sup k f (T ) k and that
f (T )u ! f (T )u for every u 2 X . 
The Convergence Lemma allows us to give immediate proofs of many of the formulae
used in studying such entities as powers of T and semigroups associated with T .
Let us consider the powers of a one{one operator T of type S!+ with dense domain
and dense range. (The density assumptions can often be removed.) For 2 C , let
f 2 F (S0+) be given by f ( ) =  , where  > ! . Then de ne T 2 C (X ) by

T = f (T ) :
From the functional calculus of T , formulae such as
T T u = T + u ; u 2 D(T ) \ D(T + ) ;
follow immediately. From the Convergence Lemma, it is also easy to prove standard
formulae for these operators such as the following. If 0 < s < 1 , then
Z R
1
s
T u = Rlim
t?s(I + tT )?1 Tu dt ; u 2 D(T ) ;
!1
!0 

where

t?s(1 + t)?1 dt :

We can develop the theory of the associated holomorphic semigroups in a similar


manner. Suppose T is a one{one operator with dense domain and dense range which
is of type S!+ , where 0  ! < 2 . For each z 2 S(0 ?!)+ let
2

gz ( ) = e?z :
Then gz 2 F (S 0 +) and the operators
2

e?zT = gz (T )
form a holomorphic semigroup.
Exercise. State and prove a theorem concerning the uniqueness of this holomorphic
functional calculus.
Exercise. Show that, whenever f (T ) 2 L(X ) , then f (T  ) 2 L(X  ) and f (T ) =
f (T  ) , where f ( ) = f ( ) .

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

16

Lecture 3. Quadratic Estimates

In this lecture and the next, H denotes a non{trivial complex Hilbert space.
We are primarily interested in characterising and studying those one{one operators
T of type S!+ in H which have a bounded H1 functional calculus, meaning that for
all f 2 H1(S0+) (where  > ! ), f (T ) 2 L(H) and

k f (T ) k  c k f k1
for some constant c . In particular, we show that this is the case if and only if the
quadratic norm k u kT is equivalent to the original norm k u k on H .
In the Hilbert space case, every one{one operator T of type S!+ necessarily has
dense domain and dense range, so that the Convergence Lemma is always valid.
Further, D(T ) \ R(T ) is dense in H . If T is not one{one, then H = N (T )  H0
where N (T ) = fu 2 D(T ) : Tu = 0g , and H0 is the closure of R(T ) , so that we do
not lose generality by restricting our attention to one{one operators. (No orthogonality
is implied by the symbol  .) See [CDMcY] for a treatment of these results, and for
the situation in Banach spaces. References for the results which follow also include [Y],
[Mc], [McY], [McQ], and [AMcN]. There is related material in [deL].

(E) Quadratic norms of operators of type S!+ in Hilbert spaces.


Suppose, for 0  ! <  , that T is a one{one operator of type S!+ in H . Given
2 (S0+) and t > 0 , de ne t 2 (S0+) by t ( ) = (t ) . The operators
t (T ) 2 L(H) , de ned in the previous section, depend continuously on t , so the
integrals in the following results make sense.

Lemma E. Let T be a one{one operator of type S!+ in H . Let ,


where  > ! . Then there exists a constant c such that

2 (S0+)

k (f t )(T ) k  c k f k1

(i)

for all f 2 H1 (S0+) and all t > 0 , and


(ii)

8
<Z
:

1 Z

d
 (T ) t (T )g ( ) 

91
2
=

dt
t;  c

k g( ) k2

d


)1

OPERATOR THEORY AND HARMONIC ANALYSIS

17

for all continuous functions g from [ ; ] to H , and all 0 < < < 1 .
Proof. If = f = rei : r > 0g , ! <  <  , is an unbounded contour parametrised
clockwise around S!+ , then
Z
1
(f t )(T ) = 2i f ( ) (t )(T ? I )?1 d :

So, for some s > 0 ,

k (f t )(T ) k  const. k f k1
 c k f k1

jt js jd j
1 + jt j2s j j

which proves (i). To obtain (ii), rst verify that


Z


j js
jt js jd j


)(
T
)
( 
 const.
t
1 + j j2s 1 + jt j2s j j
const. (t= )s(1 + log(=t)) if 0 < t   < 1
 const. (=t)s(1 + log(t= )) if 0 <   t < 1 :
Therefore a variant of Schur's estimate becomes


1 Z


d
dt
 (T ) t (T )g ( )  t

1 (Z

)2

dt
t
0

(
)
Z
Z 1
Z


d
d



2
(T ) 
(T ) k g( ) k  dtt

 (T )
 (T )
t
t

0

)
)
(
Z 1
 (Z
Z


d

d
dt

2
k g( ) k 
 sup
( 
( 
t )(T )  sup
t )(T ) t

t

0

Z
2
 c k g( ) k2 d

as required. 

 (T ) t (T )

1
2

 (T ) t (T )

1
2

k g( ) k d

We are now in a position to de ne the quadratic norm k u kT; associated with a


one{one operator T of type S!+ in H and a non{zero function 2 (S0+) . It is
de ned by

Z 1
dt
2
k t (T )u k t
k u kT; =
1
2

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

18

on the space H0T; of all u 2 H for which the integral is nite.


Exercise. Show that there exist two more functions ;  2 (S0+) such that
Z

( )( ) ( ) d = 1 :

Prove, using the Convergence Lemma, that


d = u

(
T
)

lim
(
T
)
u
=
lim
(
T
)
(
T
)
u

;


!0
!0

!1
!1
Z

for all u 2 H , where ; 2 (S0+) is de ned by


; ( ) =

 ( ) ( )  ( ) d :

Exercise. Show that k u kT; is indeed a norm on H0T; .


Exercise. Suppose that fv g is a Cauchy net in H0T; in the topology associated with
k : kT; , and that v ! v 2 H in the original norm topology. Prove that v 2 H0T;
and that k v ? v kT; ! 0 .

Proposition E. Suppose that T is a one{one operator of type S!+ in H , and that

, are non{zero functions in (S0+) where  > ! . Then there exists a constant
c such that, for every f 2 H1(S0+) and every u 2 H0T; \ D(f (T )) ,

k f (T )u kT;  c k f k1 k u kT; :
Proof. Let u 2 H0T; \ D(f (T )) . Choose ;  and ; as in the exercise above. By
using the holomorphic functional calculus of T , and applying Lemma E, we have

; (T )f (T )u k T;

=
=
=

Z

8
<Z
:

t (T ) ; (T )f (T )u

1 Z
0

8
<Z
:

1 Z
0

dt
t

1

t (T ) (T )  (T )  (T )f (T )u

d


 (T ) t (T )(f  )(T )  (T )u d

91
2
=

dt
t;
91
2
=

dt
t;

OPERATOR THEORY AND HARMONIC ANALYSIS

 const.

k (f  )(T )  (T )u k2 d


k  (T )u k2 d
 c k f k1

 c k f k1 k u kT; :
Z

19

)1

by (ii)

)1

by (i)

Using this we see that f ; (T )f (T )ug is a Cauchy net in H0T; as ( ; ) !


(0; 1) . Moreover, by the rst of the above exercises, we see that it converges in H to
f (T )u . So, by the third exercise, f (T )u 2 H0T; , and satis es the required estimate. 
On considering the case when f  1 , in which case f (T ) = I , we see that the spaces
0
T; are independent of , and indeed of  , so can be denoted by HT , and that
the norms k u kT; are equivalent, meaning that, for each pair of non{zero functions
and , there exists a positive constant c such that c?1 k u kT;  k u kT;  c k u kT;
for all u 2 H0T . We write this as k u kT;  k u kT; . Henceforth we write k u kT in
place of any one of these equivalent norms.

H0

De ne the Banach space HT to be the completion of H0T under the norm k u kT .


Actually HT is a Hilbert space, because H0T is an inner product space under
Z 1
( t (T )u ; t (T )v ) dt :
(u ; v)T =
t
0
Exercise. Prove that R(T (I + T )?2 ) = D(T ) \ R(T )  D(T ) \ R(T )  H0T  H for
all 2 (0; 1) , and hence that H0T is dense in H . Also prove that R(T (I + T )?2 ) is
dense in HT .

We say that T satis es a quadratic estimate if H  HT and k u kT  c k u k for all


u 2 H , and that T satis es a reverse quadratic estimate if HT  H and k u k  c k u kT
for all u 2 HT . Our next aim is to show that T satis es a quadratic estimate if and
only if its adjoint T  satis es a reverse quadratic estimate, or indeed, if and only if its
dual with respect to any pairing does.
Two complex Banach spaces X and Y form a dual pair hX ; Yi if there is a
bilinear or a sesquilinear form h ; i on X  Y which satis es the following properties

jhu ; vij  c k u kX k v kY
k u kX  c1 sup jhkuv;kvij
v2Y
Y
jh
u
;
v
k v kY  c2 sup k u k ij
u2X
X

for all u 2 X and all v 2 Y


for all u 2 X
for all v 2 Y

20

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

for some constants c; c1 and c2 . (A sesquilinear form hu ; vi is a complex valued


function on X  Y which is linear in u and conjugate linear in v .)
Suppose hX ; Yi is a dual pair. A dual pair hT ; T 0 i of operators of type S!+
consists of operators T of type S!+ in X and T 0 of type S!+ in Y which satisfy
hTu ; vi = hu ; T 0 vi
for all u 2 D(T ) and all v 2 D(T 0 ) .
Exercise. Let T  be the adjoint of an operator T of type S!+ in a Hilbert space H
with inner product ( ; ) . Show that hT ; T  i is a dual pair of operators of type S!+
in hH ; Hi under the pairing hu ; vi = (u ; v) .
Exercise. If hT ; T 0 i is a dual pair of operators of type S!+ in hX ; Yi and if 2
(S0+) for some  > ! , then
h (T )u ; vi = hu ; (T 0 )vi
for all u 2 X and v 2 Y , where ( ) = ( ) when the pairing is sesquilinear,
and
= when the pairing is linear. Moreover there exists a constant c such that

(T 0 )  c k (T ) k for all
2 (S0+) .
Example. If is a contour in C and 1  p  1 , let Lp( ) be the Banach space of
equivalence classes of measurable complex valued functions u on for which

or

k u kp =

Z

ju(z)jp

jdzj

=p

1

<1 1p<1 ;

k u k1 = ess sup ju(z)j < 1 :


z2

For 1  p  1 and 1=p +1=q = 1 , hLp ( ) ; Lq ( )i is a dual pair of Banach spaces


under the bilinear pairing
Z
hu ; vi = u(z)v(z) dz :
To be speci c,

k u kp = sup jhu ; vij : v 2 Lq ( ) ; k v kq = 1 :

Suppose Tu = wu and T 0 v = wv for some measurable function w with essential range in S!+ . Then hT ; T 0 i forms a dual pair of operators of type S!+ in
hLp( ) ; Lq ( )i .
Exercise. Suppose hX ; Yi is a dual pair of Banach spaces, Z is a dense linear subspace
of Y , and f is a continuous function from a compact interval [ ; ] to X . Then
there exists a Borel function g from [ ; ] to Z such that k g(t) kY = 1 for all t and
k f (t) kX  2c1hf (t) ; g(t)i for all t .

OPERATOR THEORY AND HARMONIC ANALYSIS

21

Theorem E. Suppose hT ; T 0 i is a dual pair of one{one operators of type S!+ in a


dual pair of Hilbert spaces hH ; Ki . Then hHT ; KT 0 i is a dual pair of Hilbert spaces
under the induced pairing.

Proof. Choose 2 (S0+) such that 01 2(t)t?1 dt = 1 , and let k u kT = k u kT;


and k v kT 0 = k v kT 0; . It suces to prove the following bounds.
jhu; vij  c k u kT k v kT 0 for all u 2 H0T and all v 2 KT0 0
R

vij for all u 2 H0


k u kT  C1 sup jhk u;
T
v2KT 0 v kT 0
k v kT 0  C2 sup jhku;u kvij for all v 2 KT0 0 :
u2HT
T
0

On applying the Convergence Lemma, we see that if u 2 H0T and v 2 KT0 0 , then
Z 1
dt
2
hu ; vi = h
t (T )u t ; v i
0
Z 1
h t (T )u ; t(T 0 )vi dtt ;
=
0

so that jhu ; vij  c k u kT k v kT 0 .


Now take any u 2 H0T . Choose 0 < < < 1 such that
9 k u k2  Z k (T )u k2 dt :
t
T
10
t

By the previous exercise, there exists a Borel function g from [ ; ] to KT0 0 such that
k t (T )u k2  2c1h t (T )u ; g(t)i
and
k g(t) k = k t (T )u k
R
for all t . Let v = t (T 0 )g(t)t?1 dt . Then
9 k u k2  2c Z h (T )u ; g(t)i dt = 2c hu ; vi
t
1
1
T
10
t

and

k v kT 0 =

8
<Z
:

1 Z

 const.

 (T 0 ) t (T 0 )g (t)

k g k2

 const. k u kT :

dt
t

dt
t

91
2
=

d
;

)1

(by Lemma E)

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

22

Hence
for some constant C1 .

k u kT  C1 jhkuv;kvij
T0

Similarly, we can show that for any v 2 KT0 0 , there exists u 2 H0T such that

k v kT 0  C2 jhkuu; kvij :
T


Corollary E. Suppose hT ; T 0 i is a dual pair of one{one operators of type S!+ in a


dual pair of Hilbert spaces hH ; Ki . Then HT  H with k u k  c k u kT if and only
if K  KT 0 with k v kT 0  c0 k v k .
Proof. This follows immediately from the previous theorem. 

(F) Boundedness of holomorphic functional calculi.


We are now in a position to relate square function estimates to bounded H1 functional calculi.
If T is a one{one operator of type S!+ in H , and 0  ! <  <  , then we
say that T has a bounded H1(S0+) functional calculus if, for all f 2 H1 (S0+) ,
f (T ) 2 L(H) and
k f (T ) k  c k f k1 :
Exercise. If hT ; T 0 i is a dual pair of operators of type S!+ in a dual pair of Hilbert
spaces hH ; Ki , and  > ! , show that T has a bounded H1 (S0+) functional calculus
if and only if T 0 does, and that

hf (T )u ; vi = hu ; f (T 0 )vi
for all u 2 H and v 2 K , where f was de ned in a previous exercise.

Theorem F. Suppose T is a one{one operator of type S!+ in H . Then the following

statements are equivalent:

(a) T has a bounded H1(S0+) functional calculus for all  > ! ;


(b) T has a bounded H1(S0+) functional calculus for some  > ! ;

OPERATOR THEORY AND HARMONIC ANALYSIS

23

(c) k f (T ) k  c k f k1 for all f 2 (S0+) and some  > ! ;


(d) T is 2 L(H) for all s 2 R , and T is  c ejsj for some  > ! ;

(e) D(T ) = D(A ) and D(T  ) = D(B ) for some 2 (0; 1) , where A = (T  T )
and B = (TT  ) , with k A u k  k T u k and k B u k  k T  u k ;

1
2

1
2

(f) HT = H , and there exists c > 0 such that, for every u 2 H ,

c?1 k u kT  k u k  c k u kT :
Proof. Suppose that (f) holds. Then, for all  > ! , it is a consequence of Proposition E
that k f (T )u k  c k f k1 k u k for all u 2 H and all f 2 (S0+) . Thus k f (T ) k 
c k f k1 for all f 2 (S0+) .

To obtain (a), apply this estimate, the Convergence Lemma, and the fact that every
function f 2 H1(S0+) is the limit of a uniformly bounded sequence of functions fn 2
(S0+) in the sense of uniform convergence on compact subsets of S0+ . Statements
(b) and (c) follow.
So also does (d), because T is = fs (T ) where fs 2 H1 (S0+) is given by fs ( ) =
 is , and k fs k1 = ejsj .
We shall prove that (d) implies (e) in the next section. We refer the reader to the
literature, say [Y] or [Mc], for a proof that (e) implies (f).
A proof that (b) implies (f) which is more in the spirit of harmonic analysis goes as
follows. Let k be a sequence of functions in (S0+) such that j k ( )j  2?k j j (1 +
P
2
2?k j j)?2 and kk==1
?1 k ( ) = 1 . Then, for all u 2 H ,

k u k T  c1

kX
=1

k=?1

 c1 sup

j k j1

 c2 sup sup
j k j1 

(T )u k2

)1

k k (T )u

(proved similarly to Proposition E)

(proved below)

k k ( ) k u k (by (b))

 c2 sup j k ( )j k u k

 c3 k u k :
Also, T  has a bounded H1 (S0+) functional calculus, so k u kT   c4 k u k , and
hence, by Corollary E, k u k  c5 k u kT as well, thus proving (f). 

24

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

The second inequality above is well known. It is proved in the following way. Let
uk = k (T )u . Let rk denote the Rademacher functions
on [0; 1] . These are step
R1
functions such that rk (x) = 1 for all x 2 [0; 1] , and 0 rj (x)rk (x)dx = jk . Then
X

k uk

k2

=
=

j;k

rj (x)rk (x) dx uk ; uj )

1 X

sup

j k j1

rk (x)uk

k u k

dx

)2


as required.
A. Yagi rst proved the equivalence of (d), (e) and (f), (with k u kT de ned using
a particular choice of ) [Y]. Subsequently a full statement of this theorem was given
in [Mc], though various parts of it have been known for some time. Of course, quadratic
estimates have a long history in harmonic analysis, some aspects of which we shall touch
on later.
We conclude with the remark that, by virtue of the equivalence of (a) and (b),
the de nition which we made loosely at the very beginning of this lecture is perfectly
alright for the case of one{one operators of type S!+ in a Hilbert space. That is, we say
that T has a bounded H1 functional calculus provided T has a bounded H1 (S0+)
functional calculus for some, and hence all,  > ! .

OPERATOR THEORY AND HARMONIC ANALYSIS

25

Lecture 4. Operators with Bounded Holomorphic Functional Calculi

In this lecture we investigate some classes of operators T of type S!+ in a non{


trivial Hilbert space H with inner product ( ; ) , and also consider operators of type
S! , where S! = S!+ [ S!? .

(G) Accretive operators.


Let us rst consider self{adjoint operators. A self{adjoint operator S in H is
said to be positive if (Su ; u) > 0 for all non{zero u 2 D(S ) . Every positive self{
adjoint operator is a one{one operator of type S0+ , and has a bounded Borel functional
calculus, so it certainly has a bounded H1(S0+) functional calculus for all  > 0 .
Actually, the constant c is 1 , meaning that k f (S )u k  k f k1 for all f 2 H1 (S0+) .
Let us verify that statement (f) of Theorem F holds in this case. Let u 2 H . Then
Z 1

dt
2
k u kS =
k t (S ) u k t
0

Z 1
dt
( t (S ) t (S ) u ; u) t
=
0
 Z 1

dt
= (
t t (S ) u t ; u)
0
= kuk
1
2

1
2

1
2

nR

1 j (t)j2 dt o 12 . Therefore, in this special case, the quadratic norm is a


0
t

where  =
positive multiple of the original norm, and so HS = H .
An operator T in H is accretive if Re (Tu ; u)  0 for every u 2 D(T ) , and T is
maximal accretive if T is accretive and  2 (T ) whenever Re  < 0 . Every bounded
accretive operator on H is maximal accretive.
If T is maximal accretive and Re  < 0 , then k RT ( ) k  jRe  j?1 , from which it
follows that T is of type 2 .

Every one{one maximal accretive operator T has a bounded H1(S0+) functional


calculus for all  > 2 with constant c = 1 , meaning that k f (T )u k  k f k1 for all
f 2 H1(S0+) . This fact is essentially equivalent to von Neumann's inequality, namely
that k g(W ) k  1 for all operators W 2 L(H ) such that k W k  1 , and all functions
g which are holomorphic on a neighbourhood of the closed unit disc, and bounded by
1 on the disc. See Chapter XI of [RN] for a discussion of this inequality.
We present, however, a direct proof due to Edwin Franks.

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

26

Theorem G. Suppose that T is a one{one maximal accretive operator in H . Then


T has a bounded H1 functional calculus, and
k f (T ) k  k f k1 ;
for every f 2 H1 (S0+) and all  > 2 .

Proof. Let us rst suppose that T is bounded and strictly accretive, meaning that
Re (Tu ; u)  c k u k2
for all u 2 H and some c > 0 .

Choose  > 2 and f 2 (S0+) . Then, since (T  + I )?1 is holomorphic on a


neighbourhood of f  2 C : Re   0 g , we have
Z
1
f (T ) = 2i f ( ) (T ? I )?1 d
iR
Z
1
= 2i f ( ) ((T ? I )?1 + (T  + I )?1) d :
iR

Therefore

Z
1
f (T )u = 2i f ( ) (T  + I )?1 (T  + T )(T ? I )?1 u d
iR
for all u 2 H . Note that, because of our assumptions on T , the last integral is
absolutely convergent. On taking limits as usual, and applying the Convergence Lemma,
we nd that this formula holds for all f 2 H1 (S0+) and all u 2 H . In particular,
setting f = 1 ,
Z
1
u = 2i (T  + I )?1(T  + T )(T ? I )?1u d
iR
and, since the integrand is positive self{adjoint,

kuk =

1 Z 1 (T  + T ) (T ? iI )?1 u 2 d

2 ?1
1
2

1

Moreover, for all u; v 2 H ,




Z 1
1


?
1

?
1
j(f (T )u ; v)j = 2
f (i) ((T + iI ) (T + T )(T ? iI ) u ; v) d
?1


Z 1
2
1


?
1

 k f k1 2
u d
(T + T ) (T ? iI )
?1


1 Z 1 (T  + T ) (T ? iI )?1v 2 d

2 ?1
= k f k1 k u k k v k
1
2

1
2

1
2

1
2

OPERATOR THEORY AND HARMONIC ANALYSIS

27

so k f (T ) k  k f k1 as required.

Now let us consider the general case. Take any  > 0 and let T = ((T + I )?1 +
I )?1 . Then T is a bounded strictly accretive operator on H . Moreover, as  ! 0 ,
(T ? I )?1 converges in the norm topology of L(H) , to (T ? I )?1 . This convergence
is uniform on sets of the form f  2 S0+ : r  j j  R g when 0 < r < R < 1 . So,
for every 2 (S0+) , (T ) converges in the norm topology to (T ) as  ! 0 . It
therefore follows from the above result for bounded strictly accretive operators, that

k (T ) k  k k1
for every

2 (S0+) .

We have thus proved statement (c) of Theorem F with c = 1 for all  > 2 . The
result follows by applying the Convergence Lemma as in the proof of that theorem. 
Exercise. Extend this theorem to prove that k f (T ) k  k f k1 for functions such as
ft ( ) = e?t , t > 0 , which are bounded on the right half plane, but not on any bigger
sector. Hence prove that e?tT is a C0 semigroup with e?tT  1 for all t > 0 .

We remark that the formula for f (T ) used in proving Theorem G involves a kind
of Poisson integral, as opposed to the Cauchy integral we mostly use.
Exercise. Prove that every one{one maximal accretive operator T in H has a bounded
harmonic functional calculus (where this needs to be de ned). Hence prove that
k log T ? log T  k   .
Exercise. Let W and T be two bounded operators on H related by the formula
T = (I ? W )(I + W )?1 and hence by W = (I ? T )(I + T )?1 . (Assume that ?1 is not
in the spectrum of either operator.) Prove that k W k < 1 if and only if T is strictly
accretive.

Here is a proof of one implication. For all u 2 H ,

k u k2 ? k Wu k2 = Re ((I ? W )u ; (I + W )u)

= Re (T (I + W )u ; (I + W )u)
> 0:

Exercise. Derive von Neumann's inequality as a corollary of Theorem G.

In section D, we de ned, for 0  !  2 , T to be an ! -accretive operator in H


if (T )  S!+ and (Tu ; u) 2 S!+ for all u 2 D(T ) . Such an operator T is maximal
accretive, and so, if it is one{one, has an H1 functional calculus. It is also of type S!0+ ,

28

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

and thus, by Theorem F, for all  > ! there exists c such that k f (T ) k  c k f k1
for all f 2 H1(S0+) . We note however, that if  < 2 , then the constant c may be
larger than 1.
A third class of operators which occurs in applications are those given in the following exercise. These operators do not necessarily have bounded H1 functional calculi.
Exercise. Let T = V S where S is a positive self{adjoint operator in H , and V is a
bounded invertible ! -accretive operator. Prove that T is a one{oneP
operator of type
@ u (x) in
S!+ . An example of such an operator is given by Tu(x) = ?b(x) nj=1 @x
j
L2(Rn) , where b 2 L1(Rn; C ) with Re b(x)   > 0 for almost all x 2 Rn .
2

The theory of accretive operators was developed by T. Kato and others to study operators such as elliptic operators in divergence form with bounded measurable complex
coecients. These are de ned as follows.
Consider an elliptic sesquilinear form J de ned on V  V by

@u @v + a @u v + b u @v + auv g dx
J [u; v] = f ajk @x
k @x
j @x
k @xj
k
j

where
is an open subset of Rn , V is a closed linear subspace of the Sobolev space
H 1 (
) which contains Cc1(
) , ajk ; ak ; bj ; a 2 L1(
) and Re ajk (x)k j  j j2 for
all  = (j ) 2 C n and some  > 0 . Let L be the operator in L2(
) with largest
domain D(L)  V such that J [u; v] = (Lu; v) for all u 2 D(L) and all v 2 V . Then
L+I is a maximal accretive operator in L2 (
) for some positive number  . Therefore
L + I has a bounded H1 functional calculus, and satis es quadratic estimates.
The part of Theorem F which was actually the rst to be used in studying elliptic
operators, was the fact that certain operators satisfy statement (d), and hence (e), which
was used to determine that the domains of their fractional powers are Sobolev spaces.
The proof that (d) implies (e) is a consequence of the following result with B = I ,
by rst taking S = A , then interchanging the roles of S and T , and nally repeating
the process for the adjoint operators.
Z

Proposition G. Let S and T be one{one operators of type S!S + and S!T + in


Hilbert spaces H and K respectively, each having bounded imaginary powers (as in
(d) of Theorem F.). Let B be a bounded linear mapping from H to K such that
B(D(T ))  D(S ) and k SBu k  c1 k Tu k . Then, for every 2 (0; 1) , B(D(T )) 
D(S ) and k S Bu k  c k T u k where c = c1 k B k1? .
Proof. Suppose that c1 > 0 . It suces to show that, for any u 2 D(SBT ?1 ) , the
function


(z) = S z BT ?z u k B kz?1 c?1 z

OPERATOR THEORY AND HARMONIC ANALYSIS

29

satis es j(z)j  1 on the strip f  2 C : 0  Re   1 g . From the assumptions,


we have the bounds on the lines Re  = 0 and Re  = 1 . If ( ) ! 0 at 1 in
the strip, then j(z)j  1 by the maximum modulus principle. Otherwise, consider
n( ) = ( )e =n e?1=n . Since ( ) is bounded, n( ) ! 0 as  ! 1 in the strip.
Thus, jn( )j  1 everywhere in the strip since jnj  1 on the boundary. This is true
for all n , so j( )j  1 since e =ne?1=n ! 1 as n ! 1 . Therefore, as D(SBT ?1 )
is dense in D(S BT ? ) the result follows. 
2

(H) Operators of type S! and spectral projections.


For 0  ! <  < 2 , de ne S!? = ?S!+ and S0? = ?S0+ . Then de ne the
closed and open double sectors S! = S!? [ S!+ and S0 = S0? [ S0+ , and the function
spaces

(S0)  H1 (S0)  F (S0)  H (S0)


on them, exactly as before.

Let 0  ! < 2 . An operator T in H is said to be of type S! if (T )  S! and,


for each  > ! ,


(T ? I )?1  C j j?1 ;
 62 S :

Most of the results for one{one operators of type S!+ generalise directly to the
case when T is a one{one operator of type S! . Indeed the results presented in Lecture
3 remain valid, word for word, subject to the following minor modi cations. (i) The
contours have four rays, so as to enclose both sectors. (ii) The functions 2
(S0) used to de ne k u kT; cannot be identically zero on either sector. (iii) At every
occurrence, replace T (I + T )?2 by T (I + T 2)?1 , replace T by (T 2 ) =2 , and replace
T  by (T  2 ) =2 .
The main result is, once again, that T has an H1 functional calculus if and only
if HT = H , with equivalence of norms.
Exercise. Prove that, if T is a one{one operator of type S! in H , then T 2 is a one{
one operator of type S2!+ . Prove also that HT = HT . Thus T has a bounded H1
functional calculus if and only if T 2 does.
2

Let us consider the spectral projections associated with the parts of the spectrum
in each sector.
For some  > ! , de ne the holomorphic functions  + ;  ? ; sgn 2 H1 (S0) by
 +( ) = 1 if Re  > 0 and +( ) = 0 if Re  < 0 ,  ?( ) = 1 ?  +( ) , and
sgn( ) =  +( ) ?  ?( ) . Let E+ =  +(T ) and E? = ? (T ) 2 C (H) , so that,
by the identities of the functional calculus, D(E+ ) = D(E? ) , E+2 = E+ ; E? 2 =
E? ; E+E? = 0jD(E ) = E?E+ ; E+ + E? = I jD(E ) and E+ ? E? = sgn(T ) .
+

30

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

The operators E+ and E? form a pair of closed spectral projections corresponding


to the parts of the spectra in S!+ and S!? respectively, though in general they may
fail to be bounded operators on H . But, of course, if T has a bounded H1 functional
calculus, then E+; E? 2 L(H ) because  +;  ? 2 H1 (S0) .
We see therefore, that whenever HT = H , then E+ and E? are bounded spectral
projections on H , and that H = H+  H? , where H+ = R(E+ ) and H? = R(E? )
are the corresponding spectral subspaces. (The direct sum  has no connotations of
orthogonality.) This is an important reason for obtaining quadratic estimates.
Let us consider some special cases. If T is a one{one self{adjoint operator, then T
is of type S0 , and has a bounded Borel functional calculus, so it certainly has a bounded
H1 functional calculus, again with constants c = 1 . In this case k E k = 1 so the
decomposition H = H+  H? is an orthogonal one. We remark that there exist one{
one operators of type S0 which do not have a spectral decomposition H = H+  H?
[McY].
Another class of operators of type S! consists of the ones obtained as follows. Not
all such operators have a bounded H1 functional calculus.

Theorem H. Let T = V S where S is a one{one self{adjoint operator in H , and V

is a bounded invertible ! -accretive operator. Then T is a one{one operator of type


S! .
Proof. There exists a constant c such that, for  2= S! , and u 2 D(T ) ,

cjj(T ? I )ujj jjujj  jIm(V ?1(T ? I )u; u)j = jIm (V ?1u; u)j
 j(V ?1 u; u)j dist(; S! ) (because (V ?1 u; u) 2 S! ) :
Hence, for some C > 0 ,

jj(T ? I )ujj  C dist(; S! ) k u k :


The dual of (T ? I ) with respect to the pairing hu ; vi = (V ?1u; v) of H with
itself, is (T 0 ? I ) where T 0 = V S . Since T 0 has the same form as T , we also have

jj(T 0 ? I )ujj  C dist(; S! ) k u k = C dist(; S! ) k u k :


It follows from these two estimates that  2= (T ) and

jj(T ? I )?1jj  C ?1( dist(; S! ))?1 :


We conclude that T is of type S! . 

OPERATOR THEORY AND HARMONIC ANALYSIS

31

To conclude, let us introduce a speci c operator which has this form. In the nal
lecture, we shall prove that it does indeed have a bounded H1 functional calculus, and
we shall see the signi cance of the resulting decomposition H = H+  H? .
Let denote the Lipschitz curve in the complex plane which is parametrised by
a Lipschitz function g : R ! C such that g0 ; 1=g0 2 L1(R) and g0(x) 2 S!+ for all
x 2 R.
De ne the derivative of a Lipschitz function u on by

u0 (z) = lim u(z + hh) ? u(z) :


h!
z+
h20

Next use duality to de ne D to be the closed linear operator in Lp( ) with the largest
domain which satis es
hD u ; vi = hu ; iv0 i
for all Lipschitz functions v on with compact support. We are using the pairing

hu ; vi =

u(z)v(z) dz

de ned in Section E.
Then hD ; ?D i is a dual pair of one{one operators of type S! in hLp( ) ; Lq ( )i ,
1  p  1 , 1=p + 1=q = 1 .
If V denotes the isomorphism from L2 ( ) to L2(R) induced by the parametrization, (V u)(x) = u(g(x)) , then (V D u)(x) = b(x)(DV u)(x) , where b = 1=g0 , and
D = 1i dxd with domain D(D) = f u 2 L2(R) : Du 2 L2 (R) g . Now D is a one{one
self{adjoint operator, and the operator of multiplication by b is a bounded invertible
! {accretive operator in L2 (R), so, by Theorem H, the operator T = bD is a one{one
operator of type S! in L2(R) . It follows that the operator D = V ?1TV is a one{one
operator of type S! in L2( ) .
Exercise. Show that, once we have proved that the operator T has a bounded H1
functional calculus in L2(R) , then we can conclude that the operator D has a bounded
H1 functional calculus in L2( ) .

32

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

Lecture 5. Singular Integrals

Let us turn our attention to real harmonic analysis. Good references include the
books [St1] and [St2] of E.M. Stein.

(I) Convolutions and the functional calculus of ?i dxd .


Let us brie y consider the L2 theory of the gradient operator
D = (D1 ; D2 ; : : : ; Dn ) = ( 1i @x@ ; 1i @x@ ; : : : ; 1i @x@ ) :
1
2
n
For 1  p < 1 , let Lp(Rn) denote the Banach space of (equivalence classes of)
complex valued measurable functions u on Rn for which the norm

k u kp =

 Z

Rn

ju(x)jp

dx

1

< 1;

and let L1(Rn) be the Banach space of (equivalence classes of) measurable functions
u on Rn for which the norm

k u k1 = ess supn ju(x)j < 1 :


x2R

(Functions which are equal almost everywhere are identi ed in the usual way.)
For 1 < j < n , let Dj denote the operator Dj = 1i @x@ j in Lp(Rn) with domain

@u
n
D(Dj ) =
@xj 2 Lp(R )
where the derivative is taken in the sense of distributions.
It is well known that Dj is a closed operator. In the case when p = 2 ,
Fourier theory can be used to verify this, and to construct a functional calculus of
D = (D1 ; D2 ; : : : ; Dn ) . Here is a brief survey of the results from Fourier theory that
are needed.
The Fourier transform u^ = F (u) of a function u 2 L1(Rn) is de ned by


u 2 Lp(Rn) :

u^() = F (u) () =

?i hx ; i u(x) dx
e
n

OPERATOR THEORY AND HARMONIC ANALYSIS

33

for all  2 Rn . The function u^ is a bounded continuous function on Rn which satis es


k u^ k1  k u k1 for all u 2 L1(Rn) .
If u 2 L2(Rn) \ L1 (Rn) , then u^ 2 L2(Rn) , and Parseval's identity

k u^ k2 = (2)n=2 k u k2

holds for all such u , and so the Fourier transform extends to an isomorphism, also
called F , from L2(Rn) to L2 (Rn) .
If u 2 L2(Rn) , then Dj u 2 L2 (Rn) if and only if j u^() 2 L2 (Rn; d) , and
Z
@u
^
1
(Dj u) () = F ( i @x ) () = j e?i hx ; i u(x) dx = j u^() :
j
n
R

More generally, suppose that p is a polynomial in n variables. Then p(D)u 2 L2(Rn)


if and only if p()^u() 2 L2(Rn; d) , and
(p(D)u)^ () = p()^u() :
Exercise. Use Parseval's identity to show that Dj is a closed operator in L2(Rn) as
claimed above. More generally, show that if p is a polynomial in n variables, then the
operator p(D) with domain D(p(D)) = fu 2 L2(Rn) : p(D)u 2 L2(Rn)g is a closed
operator in L2(Rn) .

The joint spectrum of D = (D1 ; D2; : : : ; Dn) is n (D) = Rn . There are various
ways to de ne this which we need not go into here, but essentially what it means is that
Rn is the support of the functional calculus of D .
For any function f 2 L1(Rn) , there is a natural de nition of f (D) de ned via
the Fourier transform, namely
( f (D) u )^ () = f () u^() :
Then f (D) is a bounded operator on L2(Rn) with
k f (D)u k2 = (2)?n=2 k f u^ k2  k f k1 k u k2
for all u 2 L2 (Rn) . This, together with the facts (i) that the mapping from functions
f to operators f (D) is an algebra homomorphism, and (ii) that there is agreement
with the natural de nition for polynomials of several variables, means that we have a
bounded L1(Rn) functional calculus of D in L2(Rn) .
The agreement with the natural de nition of p(D) for polynomials p of D is in
the sense that, if f and pf 2 L1(Rn) , then
p(D) f (D) u = (pf )(D) u for all u 2 L2(Rn) and
f (D) p(D) u = (fp)(D) u for all u 2 D(p(D)) :

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

34

It is well known that there is a close connection between the functional calculus of
D and convolution operators. For example, if  2 L1(Rn) and f = ^ , then, for almost
all x 2 Rn ,
Z
(x ? y) u(y) dy :
f (D) u(x) =   u(x) =
Rn

When f is not the Fourier transform of an L1 function, it may still be possible to


represent f (D) as a singular convolution operator. For example, if rj () = ?ij = jj
for  2 Rn , then rj (D) = Rj , the j'th Riesz transform. Note that k Rj u k2  k u k2
for all u 2 L2(Rn) . It is well known that
Z
2
xj ? yj u(y) dy
Rj u(x) = rj (D)u(x) = "lim
!0 n
jx ? yjn+1
jx?yj>"
for all u 2 L2(Rn) and almost all x = (x1 ; x2 ; : : : ; xn ) 2 Rn . Here n is the n dimensional volume of the unit n-sphere fx 2 Rn+1 : jxj = 1g . We shall see in Section
L that, if 1 < p < 1 , then k Rj u kp  cp k u kp for some constant cp .
Exercise. Let

 =

be the Laplacian on Rn with domain

@2
2
j =1 @xj


2u
@
n
D() =
=
@xj @xk 2 L2(R ) for 1  j  m; 1  k  m :
(i) Prove
self{adjoint operator in L2(Rn) , and that, letting
p that ? is a positive
?
1
@
jDj = ? , then Rj = ? jDj @xj = ?i jDj?1 Dj .
(ii) Show that the semigroup St = exp(t) , t > 0 , generated by  is represented
by
Z
t

Stu(x) = e u(x) = n kt(x ? y)u(y) dy

W22(Rn)

u 2 L2(Rn) :

where

kt(x) = (4t)?n=2 e?jxj =4t


for all x 2 Rn . Use the fact that (St u)^() = e?tjj u^() .
2

In Section G, we showed that, for any one{one self{adjoint operator T in a Hilbert


space, and for any 2 (S0+) where  > 0 , the quadratic norm k u kT = k u kT; is
a multiple of the original norm. That is,

k u kT =

Z

= kuk

(T ) u k2

dt 
t

1
2

OPERATOR THEORY AND HARMONIC ANALYSIS

where  =

nR

35

1 j (t)j2 dt o 12 .
0
t

Let us now consider the operator T = ? in H = L2(Rn) . The quadratic norm


using ( ) = e? is

k u k? =

0
Z

0
(

Z

k t St

1Z
Rn

1Z
Rn

u k2

dt 
t

1
2

@ S u(x) 2 t dxdt
@t t

@U (x; t) 2 t dxdt

@t

)1

)1
2

where U (x; t) = Stu(x) is the solution of the heat equation @U


@t (x; t) = U (x; t) with
initial value U (x; 0) = u(x) , which decays as t ! 1 . This is in agreement with the
kind of quadratic estimate typically used in harmonic analysis. This norm can also be
expressed as
)
(
2
Z
Z 1Z


dt

t (x; y)u(y)dy dx t
k u k? =

n
n
R
0
R
where t (x; y) = t @t@ kt (x ? y) .
1
2

Exercise. Obtain other equivalent quadratic norms for ? by making other choices of
. For example, consider ( ) =  2e? .

The positive self{adjoint operator ? has a bounded H1 functional calculus in


L2(Rn) . In Lp(Rn) , ? is also of type S0+ , and has a bounded H1 (S0+) functional
calculus for all  > 0 . This is a consequence of the Marcinkiewicz multiplier theorem.
It also follows from the more general Theorem N to be proved later.

(J) The Hilbert transform and Hardy spaces.


Let us turn our attention to the one dimensional case n = 1 . A treatment of higher
dimensional analogues of this material using Cli ord analysis is discussed in [McI].
The operator D = 1i dxd with domain W21(R) = fu 2 L2(R) : Du 2 L2 (R)g is a
one-one self{adjoint operator in L2(R) with spectrum (D) = R .
The functional calculus introduced in the previous section is now just the usual
functional calculus of the self{adjoint operator D .
In particular, the closed operator jDj de ned by
(jDj u)^ () = jj u^()

36

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

also has domain W21(R) and

jDj = sgn(D) D where


sgn(D) = +(D) ? ?(D) = E+ ? E? :
Here E+ = +(D) and E? =  ?(D) are the spectral projections corresponding to
the positive and negative parts of the spectrum, as de ned in Section H. They give us
the orthogonal spectral decomposition H = H+  H? . Since H = L2(R) , we write
H = L2 (R) , so that L2(R) = L+2(R)  L?2 (R) , which is the Hardy decomposition of
L2(R) .
Note that sgn(D) is a bounded operator on L2 (R) satisfying (sgn(D))2 = I . It
is the identity on L+2(R) , and minus the identity on L?2 (R). In terms of convolutions,
it is represented by
Z
i
1 u(y) dy
sgn(D)u(x) = "lim
!0 
x?y
jx?yj>"

for all u 2 L2(R) and almost all x 2 R , being i times the Hilbert transform.
It is interesting to note that, since E = 21 (sgn(D) + I ) then
Z
1 u(y) dy + 1 u(x)
i
Eu(x) =  "lim
2
!0 2
x?y
jx?yj>"
i Z1

1
=  !
lim0 2
u(y) dy
?1 x + i ? y
for all u 2 L2(R) and almost all x 2 R .
Let us look at this in a di erent way using the semigroups generated by ? jDj .
Given u 2 L2 (R) and t > 0 , de ne u+(t) 2 L2 (R) by

u+(t) = e?tD E+u = e?tjDjE+u :


Then u+(t) has the properties
8
@u+ (t) + Du (t) = 0 t > 0
>
>
+
>
<
@t
lim
u (t) = E+u
>
t
!
0 +
>
>
:
lim u (t) = 0 :
t!1 +
Also, for t < 0 , the functions u?(t) 2 L2(R) de ned by
u?(t) = etD E?u = e?tjDjE?u

OPERATOR THEORY AND HARMONIC ANALYSIS

37

satisfy the properties


8
>
>
>
<
>
>
>
:

@u? (t) + Du (t) = 0 t < 0


?
@t
lim u (t) = E?u
t!0 ?
lim u (t) = 0 :
t!1 ?

Now these partial di erential equations are actually the Cauchy Riemann equations
if we identify (x; t) with x + it 2 C . That is, the functions U+ on C + = fx + it : t > 0g
and U? on C ? = fx + it : t < 0g de ned by

U+(x + it) = u+(t)(x) ;


U? (x + it) = u?(t)(x) ;

t>0 ; x2R
t<0 ; x2R:

are holomorphic
Exercise. Show that the holomorphic functions U+ and U? on the open half spaces
C + and C ? can be represented by
Z 1
1
i
u(y) dy
U (x + it) =  2
?1 x + it ? y

for all u 2 L2(R) . So, in agreement with the previous expressions for Eu(x) ,
lim U (x + it) =
t!0 
for almost all x 2 R , and
for all x 2 R .

E u(x)

lim U (x + it) = 0
t!1 

Thus the Hardy spaces L2 (R) consist of those functions u 2 L2(R) which can
be extended holomorphically to functions U on C  which decay at in nity.
Exercise. Give a characterisation of the quadratic norm k u kD in terms of these holomorphic extensions of u .

In later sections we shall generalise this material, and consider the Hardy decomposition of L2( ) , where is the graph of a Lipschitz function.

38

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

Lecture 6. Caldero n{Zygmund Theory

Let us consider some fundamental results concerning real or complex valued functions de ned on Rn . We follow the treatment given by E.M. Stein in his excellent
books [St1, St2], and recommend that the reader consult these books, as well as those
of Y. Meyer [M], M. Christ [Ch] and others, for a fuller treatment of these topics, including the historical background. Similar results also hold for functions de ned on a
space of homogeneous type. This is a measure space (
; ) which has a metric with
the property that (B(x; 2r))  c(B(x; r)) for all r > 0 and some constant c .
Here, and subsequently, B(x; r) denotes the open ball with centre x and radius
r . For a subset B of Rn , the notation jBj denotes its Lebesgue measure, and cB its
complement. In all our proofs, the constant c can change from line to line.

(K) Maximal functions and the Calderon{Zygmund decomposition.


(K1) Maximal functions.
We rst investigate the maximal functions which are closely related to the theory
of singular integrals.
For any locally integrable function f , and any r > 0 , the Hardy{Littlewood
maximal function Mf is de ned by
(Mf )(x) = supjB(x; r)j?1
r>0

B (x;r)

jf (y)jdy :

The main properties of the Hardy{Littlewood maximal functions are as follows.

Theorem K1. Let f be a complex valued function de ned on Rn .


(a) If f 2 Lp(Rn) , 1  p  1 , then the function Mf is nite almost everywhere.
(b) If f 2 L1(Rn) , then for every > 0

Z
c
jfx : (Mf )(x) > gj  n jf (y)jdy
R
where c is a constant which depends only on the dimension n .
(c) If f 2 Lp(Rn) , 1 < p < 1 , then Mf 2 Lp(Rn) and
k Mf kp  cp k f kp

OPERATOR THEORY AND HARMONIC ANALYSIS

39

where cp depends on p and the dimension n .


Property (b) is referred to as a weak type (1; 1) estimate.

Corollary K1. If f is a locally integrable function, then


Z

almost everywhere.

f (y)dy = f (x)
lim jB(x; r)j?1
r!0
B (x;r)

The proof of the corollary is left as an exercise. See [St1], Chapter 1.


Before proving the theorem, we rst state an improved version of the covering lemma
of Vitali. We say that a family of sets fB g covers a set E provided E  [ B .

Lemma K1. Let E be a measurable subset of

which is covered by a family fB g


of balls of bounded diameter. Then there exists a disjoint sequence (Bk ) of these balls
so that
X
jBk j  cjE j
Rn

where c is a positive constant which depends only on the dimension n ( c = 5?n will
do).
For a proof of the lemma, see [St1], Chapter 1.
Proof of Theorem K1. We shall prove the theorem by showing that the inequalities hold
even for the larger uncentred maximal function Mf which is de ned by
Z
1
(Mf )(x) = sup
jf (y)jdy ;
x2B jB j B

where the supremum is taken over all balls B containing x .


For > 0 , let E = fx 2 Rn : (Mf )(x) > g . The de nition of Mf implies
that for each x 2 E there exists a ball Bx which contains x so that
Z

Bx

jf (y)jdy > jBxj :

Thus jBxj < 1 k f k1 . The family fBx : x 2 E g covers E . Using the covering
lemma above, there exists a mutually disjoint family of balls Bk such that
1

jBk j  cjE j :

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

40

Therefore

k f k1 

[Bk

jf (y)jdy >

jBk j  cjE j

which proves (b).


We now prove (a) and (c) simultaneously.
The case p = 1 is obvious with cp = 1 . Hence we suppose that 1 < p < 1 .
De ne f1 (x) = f (x) if jf (x)j  =2 , and f1 (x) = 0 otherwise. Note that f1 2
L1(Rn) if f 2 Lp(Rn) . We also have (Mf )(x)  (M(f1 ))(x) + =2 , so

jE j 

jfx 2 Rn : (M(f1 ))(x) > =2gj

Z
2
c
2
c
 k f1 k1 =
jf jdx :
jf j> =2

We now set g = Mf and de ne the distribution function of g as ( ) = jfx : jg(x)j >


gj . Then
Z

Rn

(Mf )p dx =

pd( ) =

p?1( )d :

Note that the rst equality comes from the de nition of the distribution function, the
second equality from integration by parts.
Substituting the bound of jE j into the last integral, we obtain

k Mf kpp

p

"

p?1 2c

jf j> =2 jf (x)jdx d :

We evaluate the double integral by changing the order of integration, then integrating rst with respecth to i . Simple calculations show that this gives assertion (c) with
n 1=p
the constant cp = 2 p5?p1 . 

(K2) The Calderon{Zygmund decomposition.


In 1952, Calderon and Zygmund introduced a simple but powerful method which is
now known as the Calderon{Zygmund decomposition [CZ]. It plays an important role
in the real analysis of singular integrals. The idea is to split an integrable function into
its \small" and \large" parts, then analyse each part using di erent techniques.
We rst state a lemma using the covering idea of Whitney.

OPERATOR THEORY AND HARMONIC ANALYSIS

Lemma K2. Given a closed, nonempty subset E of

41

Rn ,

then its complement cE


is the union of a sequence of cubes Qk , whose sides are parallel to the axes, whose
interiors are mutually disjoint, and whose diameters are approximately proportional to
their distances from E . More explicitly:
(a) k Qk = cE ;
(b) the interiors of Qk and Qj are disjoint if k 6= j ;
S

(c) there exist positive constants c1 and c2 so that

c1d(Qk )  (Qk ; E )  c2d(Qk )


where d(Qk ) denotes the diameter of Qk and (Qk ; E ) is the distance from Qk to
E.
For the proof of this lemma, see [St1], Chapter 1.
The following theorem is referred to as the Calderon{Zygmund decomposition.

Theorem K2. Suppose f 2 L1(Rn) and > 0 . Then there exists a decomposition

of f ,

f = g+b = g+

so that

bk ;

(i) jg(x)j  c for almost all x 2 Rn ;


(ii) there exists a sequence of cubes Qk with mutually disjoint interiors so that the
support of each bk is contained in Qk and
Z

jbk (x)jdx  c jQk j and

(iii)

bk (x)dx = 0 ;

c
k jQk j  jf (x)jdx

where the constant c depends only on the dimension n .


Proof.

We use the uncentred maximal function M as in the proof of Theorem K1 and


de ne E = fx 2 Rn : (Mf )(x) > g . Then E is an open set and its complement is

42

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

nonempty. Using
the above covering lemma of Whitney, there exists a family of cubes
S
Qk so that k Qk = E and the interiors of the sets Qk are mutually disjoint.
De ne g(x) = f (x) for x 2 cE and

g(x) = jQk j?1


if x 2 Qk : It follows that f = g +

Qk

f (y)dy

bk , where

bk (x) = XQk f (x) ? jQk j?1

Qk

f (y)dy

and XQk denotes the characteristic function of Qk .


By Corollary K1, we have jf (x)j  for almost all x 2 cE . Let Bk be the
cube with the same centre as Qk but whose sides are expanded by the factor c2 in
Lemma K2. Then
Z
?
1
jf (x)jdx 
jBk j
because the cube Bk intersects cE .

Bk

Since jBk j = (c2)n jQk j , it follows that jg(x)j  c , hence (i) is proved.
It is straightforward from its de nition that bk is supported in Qk and has the
average value 0 . Also
Z

Rn

jbk jdx  2

Qk

jf (x)jdx  jBk j  c jQk j ;

thus (ii) is proved.


Property (iii) follows from the weak type (1; 1) estimate of the maximal function. 

(L) Singular Integral Operators.


In this section, we investigate singular integral operators T which are expressible
in the form
Z
(Tf )(x) =
K (x; y)f (y)dy
Rn

in some sense, where the kernel may be singular near x = y .


Because of this singularity, some care needs to be taken with the above expression.
Let us call T a Calderon{Zygmund singular integral operator with kernel K if
(1) the operator T is bounded on L2(Rn) with norm C2 ;

OPERATOR THEORY AND HARMONIC ANALYSIS

43

(2) the kernel K is a measurable function with the property that for each continuous
function f Rwith compact support, and for almost all x not in the support of f , then
(Tf )(x) = R n K (x; y)f (y)dy ;
(3) there exists a constant C > 1 and 0 <  1 so that
jK (x; y)j  jx ?Cyjn ;
j
x
?
x
1 j
jK (x; y) ? K (x1 ; y)j  C jx ? yjn+
for jx ? x1 j  jx ? yj=2 , and

jK (x; y) ? K (x; y1 )j  C jxjy??yyjn1j+
for jy ? y1 j  jx ? yj=2 .
Note that the above condition (3) implies the following condition (3').
( 30 ) There exist constants  , C so that
Z

c B (y;r)

jK (x; y) ? K (x; y1 )jdx  C 0

whenever y1 2 B(y; r) , for all y 2 Rn , and


Z

c B (x;r)

jK (x; y) ? K (x1 ; y)jdy  C 0

whenever x1 2 B(x; r) , for all x 2 Rn .


The main result is the following.

Theorem L. Under the above assumptions, for any p , 1 < p < 1 , the operator T

can be extended to a bounded operator on Lp(Rn) with the norm bound Cp depending
on p , the constants C2 , C and in assumptions (1) and (3) .
For p = 1 , the operator T is of weak type (1; 1) .
Proof. Let us prove that T is of weak type (1; 1) , using assumption (1) and the rst
estimate in (3'). The result then follows from the Marcinkiewicz interpolation theorem
for 1 < p < 2 , and then the standard duality argument for 2 < p < 1 .

For f 2 L1(Rn) L2(Rn) and > 0 , decompose f as f = g + b = g + k bk by


the Calderon{Zygmund decomposition of Theorem K2. Then
jfx : jTf (x)j > gj  jfx : jTg(x)j > =2gj + jfx : jTb(x)j > =2gj :
T

44

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

It is not dicult to check that g 2 L2(Rn) . Using the facts that T is bounded on
L2(Rn) and that jg(x)j  c , we obtain
jfx : jTg(x)j > =2gj  c1 ?2 k Tg k22  c2 ?2 k g k22  c 3 k f k1 :
On the other hand
Z
1
jfx : jTb(x)j > =2gj  jBk j + c jTbk (x)jdx
Bk
k
k
where Bk is the cube with the same centre yk as that of the cube Qk in the Calderon{
Zygmund decomposition but with sides multipliedPby the constant  in assumption (3').
Because of property (iii) of the decomposition, k jBk j  c4 ?1 k f k1 .
X

Since the average value of bk is 0, we have

Tbk (x) =

[K (x; y) ? K (x; yk )]bk (y)dy :

It then follows from the rst estimate in (3') and the estimates (ii), (iii) of the decomposition that
Z
X
jTbk (x)jdx  c5 k f k1 :
0
k

c Bk

On combining these estimates, we nd that

jfx : jTf (x)j > gj  c 4 k f k1

for all f 2 L1(Rn) L2(Rn) . We leave it to the reader to complete the proof by proving
this weak type (1; 1) estimate for all f 2 L1(Rn) . 
T

Remark. Condition ( 30 ) is known as the Hormander condition. It was already implicit


in [CZ] but Hormander is the rst who formulated it explicitly in [H].

Many important operators in harmonic analysis are Calderon{Zygmund operators,


the most important being the Hilbert transform and the Riesz transforms. Thus these
operators are bounded on Lp(R) and Lp(Rn) respectively when 1 < p < 1 . Actually
it was these operators that were studied rst.
The Riesz transforms play an important role in the analysis of partial di erential
equations. For example, it follows from the equation

@ 2 u = ?R R u
j k
@xj @xk

OPERATOR THEORY AND HARMONIC ANALYSIS

45

and the boundedness of the Riesz transforms, that the Laplacian


controls all second

@ u
order partial derivatives in the Lp norm. That is, @xj @xk  cp k u kp for all j
p
and k .
2

More consequences of Theorem L will be given in the forthcoming lectures, arising


from functional calculi of elliptic operators in Lecture 7, and from convolution singular
integrals on Lipschitz curves in Lecture 8.
Exercise L. Generalise Theorem L in the following way.

Let T be a bounded operator on L2 (Rn) . Suppose that Tk is a sequence of


bounded operators associated with kernels Kk in the sense of condition (2), such that
the kernels Kk satisfy condition (3)
with the constants C and independent of k .
T
n
Assume that for each f 2 L2 (R ) L1(Rn) there is a subsequence Tkj such that
(Tf )(x) = jlim
(T f )(x)
!1 kj
for almost all x 2 Rn .
Prove that T is of weak type (1; 1) .
Conclude that T can be extended to a bounded operator on Lp(Rn) with the norm
bound Cp depending on p , the constants C and , and the L2 norm of T .

46

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

Lecture 7. Functional Calculi of Elliptic Operators

(M) Heat kernel bounds.


Let
be a measure space equipped with a distance d and measure  . Assume
that T is an operator of type S!+ in L2(
) with ! < 2 . Then ?T generates a
holomorphic semigroup e?tT . If, for all t > 0 , these operators can be represented by
kernels kt (x; y) , then these kernels are called the heat kernels of T .
We say that the heat kernel of T satis es a Gaussian upper bound if

jkt (x; y)j  (x; y;Ct1=m ) e?c[d(x;y)m=t] = m?


1 (

1)

where c; C and m are positive constants, and (x; y;  ) denotes the maximum volume
of the two balls with centres x and y and with radius  .
In many applications, the heat kernel of such an operator also satis es, for some
0 <  1 , a Holder continuity estimate in the variable y :

1 (

1 (

1 )]
?c[d(x;y)m=t] = m?
e
jkt(x; y) ? kt(x; y1 )j  C(x;[dy;(y;t1y=m

)t
whenever d(y; y1 )  d(x; y)=2 , and a similar estimate in the variable x :
1)

jkt(x; y) ? kt(x1 ; y)j  C(x;[dy;(x;t1x=m1 )])t e?c[d(x;y)m=t] = m?


whenever d(x; x1 )  d(x; y)=2 .

1)

We remark that there do exist interesting operators which possess Gaussian bounds
on their heat kernels, but not the above Holder continuity estimates.
We now give a list of examples:
(1) In Chapter 5, we saw that the heat kernel of the Laplacian
n

2
@
 = @x2
k=1 k
X

on Rn is given explicitly by

kt (x; y) = (4t)?n=2 e?jx?yj =4t


2

OPERATOR THEORY AND HARMONIC ANALYSIS

47

for all t > 0 . In this case, the above formula for the heat kernel also extends to all
complex value t with Re t > 0 . This formula shows that the heat kernel has the
Gaussian bound with m = 2 and the corresponding upper bound on its derivatives
(with respect to space variables), hence the Holder continuity estimates hold for both
variables x and y .
(2) If
is a region in Rn with smooth boundary and T is the Laplacian on L2(
)
subject to Dirichlet boundary conditions, then it follows from the Feynman{Kac formula
that its heat kernel kt(x; y) is a positive, C 1 function which satis es
0 < kt (x; y)  (4t)?n=2 e?jx?yj =4t
2

for x 6= y and t > 0 .


The Gaussian bound for complex t with Re t > 0 can also be obtained by analytic
continuation. See, for example, [D]. The Holder continuity estimates are also true.
(3) Let V be a nonnegative function on Rn such that the closed quadratic form

Q(f ) =

j =1

k @j

f k2 +

1
2

is densely de ned. There exists a unique positive operator T such that (Tf ; f ) = Q(f )
and D(T ) = D(Q) . That operator T is called the Schrodinger operator with potential
V , usually written as T = ? + V (x) . The Trotter formula shows that the heat kernel
of T (for t > 0 ) is positive and satis es the Gaussian upper bound. However, unless
V satis es additional conditions, the heat kernel can be a discontinuous function of the
space variables and the Holder continuity estimates may fail to hold.
1
2

(4) Let A(x; D) be an elliptic operator of even order m de ned in a bounded domain

of Rn with smooth boundary, and let Bj (x; D) , j = 1; : : : ; m=2 , be operators of


order mj < m de ned on the boundary of
. We assume
2
(i) the system A(x; D); fBj (x; D)gm=
j =1 , as well as its formal adjoint system, are
regular on
in the sense of Agmon;

(ii) there is an angle  2 (0; 2 ) such that the system ei Dtm ? A(x; Dx ) ,
2
fBj (x; Dx )gm=
j =1 is an elliptic boundary value problem on
 (?1 < t < 1)) which
satis es the coerciveness condition for any     2 ?  .
Let T be the operator with domain

D(T ) = f u 2 W2m (
) : Bj (x; D)u = 0 on @
; j = 1; :::m=2 g
which is de ned by (Tu)(x) = A(x; D)u(x) for all u 2 D(T ) . It is known that the

operator de ned analogously by the formally constructed adjoint system coincides with
the adjoint of T .

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

48

Under these assumptions, T + c0I , where c0 is a suciently large positive constant,


generates a holomorphic semigroup which possess a heat kernel with Gaussian bound.
The Holder continuity estimates also hold. See [T].
(5) Let X1 ; :::; Xk be smooth vector elds on Rn satisfying the Hormander subelliptic
condition uniformly over Rn , which means the rank of the Lie algebra generated by
the vector elds is constant. Let d be the control distance corresponding to the vector
elds and  the Haar measure. Then the space (Rn; d; ) is of polynomial type with
the dimension at in nity n and the local dimension n0  n . Next let

T = ?

i=1

Xi Xi

denote the self{adjoint operator associated with the Dirichlet form

Q() =

i=1

k Xi  k22

on L2(Rn; d; ) with domain D(Q) = fu 2 L2(Rn; d; ) : Xiu 2 L2(Rn; d; ) for all ig .


The operator T then has a heat kernel which satis es a Gaussian bound with
m = 2 . (The small t bounds were rst derived by Sanchez{Calle and the large t
bounds were subsequently established by Kusuoka and Stroock). Corresponding bounds
on the \ rst order derivatives" are also true, hence the Holder continuity estimates hold.
See [SC], [R].
(6) Let
be a bounded open connected subset
of Rn . Let a = (aij ) be an nn matrix
P
with entries aij 2 L1(
) satisfying Re aij (x)i j   j j2 for all x 2
;  2 C and
some  > 0 , and de ne the quadratic form

Q() =

i;j =1

(@i ; aij @j )

where the domain D(Q) is the closure of Cc1(


) with respect to the norm (Q() +
k  k2) . The associated operator corresponds to the strongly elliptic operator
n @
X
@
T = ?
a
(
x
)
ij
@xj
i;j =1 @xi
1
2

with Dirichlet boundary conditions. This operator is ! accretive for some ! < 2 .

In the case the coecients aij are real, then the heat kernel of T satis es the
Gaussian bound (see [GW] for the case of symmetric coecients). The Holder continuity
estimates also hold (see [DMc]).
If
= Rn , the Gaussian bound on the heat kernel is still true in the case of real
coecients. See [A], [AMcT]. However, for complex coecients, there are examples to
show that the Gaussian bound may fail to hold in dimensions n  5 .

OPERATOR THEORY AND HARMONIC ANALYSIS

49

(N) Bounded H1 functional calculi in Lp spaces.


Let T be a di erential operator of type S!+ in Lp(
) . It depends on the operator
and the structure of the space that various methods can be used to establish a bounded
functional calculus for T . Here we try to give the reader a sketch of these methods.
(1) Fourier multiplier theory: Let T be a (higher order) elliptic partial di erential
operator with constant coecients on the Euclidean space Rn , or the half space Rn+ ,
with Dirichlet boundary conditions. Using the Fourier transform, we can establish
the bounds on the symbols of the resolvents. Then a bounded holomorphic functional
calculus of T on Lp spaces follows from the Mikhlin's multiplier theorem. This method
is restricted to the case of constant coecients. However, by combining it with a
perturbation argument, we can obtain a bounded H1 functional calculus when T is a
second order elliptic operator on Rn with Holder continuous coecients. See [PS] for
the case of f (T ) = T is; s 2 R .
(2) Pseudo{di erential operators: This technique can be used to establish the
bounded H1 functional calculus of a general elliptic operator with smooth coecients,
acting on q -tuples of functions on a compact manifold with smooth boundary. See [Se],
[Du]. In the case of Rn , by using a perturbation argument, the smoothness conditions
can be reduced to Holder continuous coecients. See [AHS].
(3) Transference method: This technique was introduced and used by Coifman,
Weiss and Rochberg in [CW] and [CRW] to prove various results, including a generalisation of von Neumann's inequality to operators in Lp spaces. It was also used to
establish the Lp boundedness of f (T ) when T is the generator of a positive contractive semigroup [CRW], a result which was improved by Cowling in [Co]. Duong applied
these ideas in his study of functional calculi of second order elliptic partial di erential operators [Du]. However, this method can only be used to obtain an H1 (S0+)
functional calculus for all p , 1 < p < 1 when  > 2 .
(4) Singular integral theory: This is a powerful technique and it is our aim in this
section to show how Calderon{Zygmund theory can be used to establish bounded H1
functional calculi on Lp spaces.
The main result is the following theorem.

Theorem N. Assume T is a one{one operator of type S!+ in L2(Rn) , 0  ! < =2 ,


which satis es the following conditions:

(1) T has a bounded H1 functional calculus in L2(Rn) for all f 2 H1 (S0 ) .


+

(2) There exists  > ! so that the kernel kz (x; y) of the holomorphic semigroup
e?zT associated with T satis es the Holder continuity estimate:

jkz (x; y) ? kz (x; y1 )j  C jy1 ? yj jzj?(n=m+ )exp(?[cjx ? yjm =jzj]1=m?1 )

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

50

for all z 2 S(0 ?)+ , and some 0 <  1 .


2

Then if f 2 H1(S0 ) , for any  >  , the operator f (T ) is of weak type (1; 1) .
For any p , 1 < p  2 , f (T ) can be extended to a bounded operator on Lp(Rn) with
norm at most c k f k1 . A similar result is true for 2 < p < 1 if the Holder continuity
estimate in condition (2) holds for variable x instead of y .
Proof. First consider functions f 2 (S0+) .

To obtain the estimates for f (T ) , we represent the operator f (T ) by using the


semigroup e?zT . As in Section D, the operator f (T ) on L2 is given by
Z
1
f (T ) = 2i (T ? I )?1f ()d ;

where the contour = + [ ? is given by +(t) = tei for t  0 and ? (t) = ?te?i
for t  0 .
For  2 + , substitute
(T ? I )?1 =

?+

ez e?zT dz

where the curve ?+ is de ned by ?+(t) = te(?)=2 for t  0 . We also have similar
expression for  2 ? .
We then change the order of integration and obtain

f (T ) =

?+

where

e?zT n+(z)dz +

??

e?zT n?(z)dz ;

Z
1
n = 2i ez f ()d ;


which implies the bound

jn(z)j  c k f k1 jzj?1

Consequently, the kernel Kf (x; y) is given by

Kf (x; y) =

?+

kz (x; y)n+ (z)dz +

??

kz (x; y)n? (z)dz :

Using this representation and the Holder continuity estimate of the heat kernel, elementary but tedious calculations show that we obtain
Z

c B (y;r)

jKf (x; y) ? Kf (x; y1 )jdx  C

OPERATOR THEORY AND HARMONIC ANALYSIS

51

and our theorem follows from Theorem L for f 2 (S0+) . Note that the constant C
in the above estimate does not depend on how fast the function f decays at 0 or 1 .
To extend the result to f 2 H1 (S0+) , take
1=k
fk () = f () (1 + )2=k :

By the Convergence Lemma (Theorem D), fk (T )u converges to f (T )u in the L2 norm


for all u 2 L2(Rn) . Hence there is a subsequence ki so that fki (T )u(x) converges to
f (T )u(x) almost everywhere. The theorem then follows from the above estimates for
f 2 (S0+) and Exercise L.
The case 2 < p < 1 follows from duality. 
Exercise. Let f (T )p denote the extension of f (T ) to Lp(Rn) . Show that f (T )p =
f (Tp ) for a suitably de ned operator Tp of type S!+ in Lp(Rn) . Thus Tp has a
bounded H1 (S0+) functional calculus in Lp(Rn) for all  > ! .

Notes.
(1) The theorem is still true when Rn is replaced by a space of homogeneous type.
(2) The operators in examples (1), (2), (3), (5) and (6) (with real, symmetric
coecients) of Section M are self{adjoint in L2 . The operators in (4) (under suitable
conditions) and (6) with complex coecients are maximal accretive. Hence they all have
bounded functional calculi in L2 . Therefore, for all the above operators, the Holder
continuity estimates imply the existence of bounded functional calculi in Lp .
(3) It was proved recently that the above theorem is still true if the condition of
Holder continuity estimates is replaced by an upper bound on the heat kernel which
satis es certain conditions, in which the Gaussian bound is a special case [DR].

52

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

Lecture 8. Singular Integrals on Lipschitz Curves

In this chapter we consider harmonic analysis on Lipschitz curves. Those who


are interested can nd a survey of higher dimensional analogues of this material using
Cli ord analysis in [McI].

(O) Convolutions and the functional calculus of ?i dzd j .


Let us recall the following material from the end of Section H.
Let denote the Lipschitz curve in the complex plane which is parametrised by a
Lipschitz function g : R ! C such that g0 ; 1=g0 2 L1(R) and g0(x) 2 S!+ for almost
all x 2 R .
De ne the derivative of a Lipschitz function u on by

u0 (z) = lim u(z + hh) ? u(z)


h!0
z+h2

for almost all z 2 . Next use duality to de ne D to be the closed linear operator in
Lp( ) with the largest domain which satis es

hD u ; vi = hu ; iv0 i
for all Lipschitz functions v on with compact support. We are using the pairing

hu ; vi =

u(z)v(z) dz

de ned in Section E.
Then hD ; ?D i is a dual pair of one{one operators of type S! in hLp( ) ; Lq ( )i ,
1  p  1 , 1=p + 1=q = 1 . This can be seen by giving an explicit formula for the
resolvent. See [McQ]. Alternatively, if p = q = 2 , we can proceed as follows.
If V denotes the isomorphism from L2 ( ) to L2(R) induced by the parametrization, (V u)(x) = u(g(x)) , then (V D u)(x) = b(x)(DV u)(x) , where b = 1=g0 , and
D = 1i dxd with domain D(D) = W21(R) = f u 2 L2(R) : Du 2 L2(R) g . Now D is
a one{one self{adjoint operator, and the operator of multiplication by b is a bounded
invertible ! {accretive operator in L2(R) , so by Theorem H, the operator T = bD is
a one{one operator of type S! in L2(R) . It follows that the operator D = V ?1 TV

OPERATOR THEORY AND HARMONIC ANALYSIS

53

is a one{one operator of type S! in L2( ) . It is not hard to verify that the dual of
D is ?D .
We have seen that every one{one operator of type S! has a holomorphic functional
calculus, meaning, in this case, that there is a natural way to de ne the closed operator
f (D ) for each holomorphic function f 2 F (S0) when  > ! . What we would like
to know is whether D has a bounded H1 functional calculus in Lp( ) , especially in
L2( ) . The situation is very di erent from the special case when = R and D = D
which was treated in Section J, because D is typically not self{adjoint, and there is
no identity of Parseval for functions in L2( ) .
Of course, if f 2 (S0) , then f (D ) 2 L(Lp( )) , and the agreement with the
natural de nition for polynomials p means that if pf 2 (S0) also, then

p(D ) f (D ) u = (pf )(D ) u for all u 2 L2( ) and


f (D ) p(D ) u = (pf )(D ) u whenever u 2 D(p(D )) :
Again there is a close connection between the functional calculus of D and convolution operators.
For example, suppose that  is a function ins L1 (R) with a holomorphic extension
cjzj
to S0 ,  > ! , which satis es j(z)j  jzj(1+
jzj s ) for some c and s > 0 , and let
f = ^ . Then f has a holomorphic extension, also called f , to S0 , and, whenever
! <  <  , then f 2 (S0) . Moreover
2

f (D ) u(z) =

(z ? w) u(w) dw

for all u 2 Lp( ) and almost all z 2 .


This is not too hard to verify, though not as simple as for the case = R treated
in Lecture 5, because we can no longer identify the Fourier transform of both sides with
f u^ . To see that the convolution formula for f (D ) is reasonable, let us just check the
formula p(D )f (D ) = (pf )(D ) when p() =  and 0 satis es
the same bound as
R
1
0
^
0
 , so that ( i  ) = pf 2 (S ) . Then D f (D )u(z) = D (z ? w)u(w) dw =
R
1 0
i  (z ? w)u(w) dw = (pf )(D )u(z ) for all z 2 . There is a full treatment of
holomorphic extensions of functions and their Fourier transforms in [McQ].
For these functions  , we shall use t for the function t(z) = 1t ( zt ) . This is so
that, if f = ^ , then ft ( ) = f (t ) = (t )^( ) .
We are now in a position to state our main theorem.

Theorem O. The operator D has an H1 functional calculus in H = L2( ) .

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

54

Equivalently, HD = H and k u kD  k u k for all u 2 H = L2( ) .


It suces to prove that k u kD  c k u k for
all u 2 H , for then, since D is dual
R
to ?D with respect to the pairing hu ; vi = u(z)v(z)dz , we can infer, on applying
Corollary E, that k u k  c k u k?D = c k u kD , and thus get equivalence. This is the
reason why we considered pairings on a Hilbert space other than the inner product.
To prove the quadratic estimate k u kD  c k u k , it suces to show that for some
 > !s and for some suciently nice function  2 H (S0 ) which satis es j(z)j 
C jzj
jzj(1+jzj s ) for some C and s > 0 , then there exists a constant c such that
2

1 Z

t( ? w)u(w) dw

dt
t

)1
2

 c k u k2

for all u 2 L2( ) .


There are by now several ways to prove such an estimate. One way is to apply the
following result of S. Semmes [SS]. The proof of this theorem requires more harmonic
analysis, such as properties of Carleson measures, than we have developed in these
lectures.

Proposition O. Let t be complex valued functions on Rn  Rn , t > 0 , such that,


for some C;  > 0 , and some function b 2 L1(Rn) satisfying Re b   ,
n+1

(1)

jt (x; y)j  (t + jCt


x ? yj)n+2

(2)

n jy ? y 0 j
jt (x; y) ? t (x; y0 )j  (t Ct
+ jx ? yj)n+2
Z

(3)


(
x;
y
)
b
(
y
)
dy
=
0
for
all
x
2
Rn :
t
n

Then there exists c such that


(

1 Z

Rn

t (; y)u(y) dy

dt
t

)1

 c k u k2

for all u 2 L2(Rn) .


This result provides us with a powerful tool for proving quadratic estimates. It is
intriguing in that the L2 estimates follow from the cancellation property of the kernels
with respect to any single function b 2 L1(Rn) .

OPERATOR THEORY AND HARMONIC ANALYSIS

55

Before proceeding, we remark that the quadratic norm for ? was expressed in
such a way at the end of Section I, with t (x; y) = t @t@ kt (x ? y) , a function which clearly
satis es (1), (2) and (3) with b = 1 .
Proof of Theorem O. Take, for example, (z) = (1+zz ) , and let t(z) = 1t ( zt ) . The
functions de ned by t (x; y) = t(g(x) ? g(y)) for almost all x; y 2 R satisfy (1), (2)
in Proposition O (with n = 1 ). They also satisfy (3) with b = g0 . Let us check this
when t = 1 . For all x 2 R ,
2 2

1 (x; y)b(y) dy =
=

(g(x) ? g(y))g0 (y) dy

(g(x) ? w) dw

= 21 dwd (1 + (g(x) ? w)2 )?1 dw



= 0:
Therefore, by Proposition O,
(

1 Z

t( ? w)u(w) dw

dt
t

)1
2

c

(
Z

1 Z 1

t (; y)u(g(y))g0 (y) dy

c k u(g())g0 () k


 c k u k2

dt
t

)1

for all u 2 L2( ) as required. 


Exercise. Suppose 1 < p < 1 and  > ! . Apply Theorem O and a variant of
Theorem N for operators of type S! to conclude that the operator D has an H1(S0)
functional calculus in Lp( ) .

(P) The Cauchy integral on a Lipschitz curve and Hardy spaces.


When f is not the Fourier transform of an L1 function, it may still be possible to
represent f (D ) as a singular convolution operator. In particular, we are interested in
the holomorphic functions + , ? , sgn 2 H1 (S0) de ned in Section H. Then

i
sgn(D )u(z) = "lim
!0 

jz?wj>"

1 u(w) dw
z?w

for all u 2 L2( ) and almost all z 2 , being the Cauchy singular integral operator on
.

56

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

Further, E =  (D ) = 21 (sgn(D ) + I ) , so


Z
1 u(w) dw + 1 u(w)
i
Eu(z) =  "lim
2
!0 2
z?w
jz?wj>"

Z
i
=  !
lim0 2 z + i1 ? w u(w) dw

for all u 2 L2( ) and almost all z 2 .
Here E+ =  +(D ) and E? =  ? (D ) are the spectral projections corresponding
to the parts of the spectrum (D ) in the right and in the left half planes, as de ned
in Section H. They give us the spectral decomposition H = H+  H? . Since H =
L2( ) , we write H = L2 ( ) , so that L2( ) = L+2( )  L?2 ( ) , which is the Hardy
decomposition of L2( ) .
Let us also consider the semigroups generated by ? jD j = ?sgn(D )D .
Given u 2 L2 ( ) and t > 0 , de ne u+(t) 2 L2( ) by
u+(t) = e?tD E+u = e?tjD j E+u :
Then u+(t) has the properties
8
@u+ (t) + D u (t) = 0 t > 0
>
>
+
>
<
@t
lim u (t) = E+u
>
t!0 +
>
>
:
lim u (t) = 0 :
t!1 +

Also, for t < 0 , the functions u?(t) 2 L2( ) de ned by


u? (t) = etD E?u = e?tjD jE? u
satisfy the properties
8
@u? (t) + D u (t) = 0 t < 0
>
>
?
>
<
@t
lim
u (t) = E?u
>
t
!
0 ?
>
>
:
lim u (t) = 0 :
t!1 ?
Now these partial di erential equations are actually the Cauchy Riemann equations
if we identify (z; t) with z + it 2 C . That is, the functions U+ on
+ = fz + it : t > 0g
and U? on
? = fz + it : t < 0g de ned by

U+ (z + it) = u+(t)(z) ;
U? (z + it) = u?(t)(z) ;

t>0; z2
t<0; z2

OPERATOR THEORY AND HARMONIC ANALYSIS

57

are holomorphic.
Exercise. Show that the holomorphic functions U+ and U? on the regions
+ and

? can be represented by
Z
i
U(z + it) =  2 z + it1 ? w u(w) dw

for all u 2 L2( ) . So


for almost all z 2 , and
for all z 2 .

lim U (z + it) =
t!0 

E u(z)

lim U (z + it) = 0
t!1 

Thus the Hardy spaces L2 ( ) consist of those functions u 2 L2( ) which can be
extended holomorphically to functions U on
 which decay at in nity.
Exercise. Give a characterisation of the quadratic norm k u kD in terms of these holomorphic extensions of u .

We saw in the exercise at the end of Section O that D has an H1 (S0) functional
calculus in Lp( ) for 1 < p < 1 , so the operators considered above are also bounded
on Lp( ) . Let us summarise these results as follows.

Theorem P. The Cauchy singular integral operator C = sgn(D ) is a bounded


operator on Lp( ) , 1 < p < 1 , as are the spectral projections E =  (D ) =
1
2 (C + I ) corresponding to the parts of  (D ) in each half plane. Thus Lp ( ) =
L+p ( )  L?p ( ) where Lp ( ) = E(Lp ( )) are the Hardy spaces consisting of those
functions u 2 Lp( ) which can be extended holomorphically to functions U on

which decay at in nity.

The proof of these theorems builds upon the work of Zygmund, Calderon, Carleson,
Stein, Fe erman, Meyer, Coifman and many other people. Calderon rst proved Theorem P in 1977, in the case when ! is small [C]. Subsequently, Coifman, McIntosh and
Meyer proved the boundedness for all such curves [CMcM].
The use of the Calderon rotation method leads to the boundedness of singular
double{layer potential operators on Lp(b
) , when b
is the boundary of a strongly
Lipschitz domain
 Rm+1 . This fact was used soon after by Verchota to solve the
Dirichlet and Neumann problems for harmonic functions with L2 boundary data on
such domains by using layer potentials [V]. (These problems had been solved previously
by Dahlberg [D1] and by Jerison and Kenig [JK] using other methods.)

58

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

The proof in [CMcM] involved complicated multilinear estimates. Subsequently,


other methods were developed which simpli ed and generalised these results. In particular, there was the T (b) theorem of McIntosh and Meyer [McM] and of David, Journe
and Semmes [DJS], as well as the method of Semmes which we presented [SS]. There
were wavelets [M]. And there was the paper by Coifman, Jones and Semmes [CJS] which
gave two elementary proofs of the boundedness of C , the rst reducing it to quadratic
estimates of Kenig in Hardy spaces, and the second using martingales. Actually, no
proofs are really elementary, because they all rely on the power of Calderon{Zygmund
theory and Carleson measures in some form.
During the same period, the functional calculus aspect of these results was developed, rst by Coifman and Meyer [CM1], and more fully by Tao Qian and myself [McQ].
Higher dimensional versions of this material were developed in [LMcS] and [LMcQ].
References
[A]

P. Auscher, Regularity theorems and heat kernels for elliptic operators, J. Math. London
Soc., to appear.
[AHS]
H. Amman, M. Hieber and G. Simonett, Bounded H 1 calculus for elliptic operators, Di .
Int. Eqns 7 (1994), 613{653.
[AMcN] P. Auscher, A. McIntosh and A. Nahmod, Holomorphic functional calculi of operators,
quadratic estimates and interpolation, submitted.
c
[AM T] P. Auscher, A. McIntosh and Ph. Tchamitchian, Heat kernels of second order complex
elliptic operators and applications, submitted.
[AT]
P. Auscher and Ph. Tchamitchian, Bases d'ondelettes sur les courbes corde{arc, noyau de
Cauchy et espaces de Hardy associes, Revista Matematica Iberoamericana 5 (1989), 139{
170.
[AT2]
P. Auscher et Ph. Tchamitchian, Conjecture de Kato sur des ouverts de R , Rev. Math.
Iberoamericana 8 (1992), 149{201.
[BL]
J. Bergh and J. Lofstrom, Interpolation Spaces. An Introduction, Springer{Verlag, 1976.
[C]
A.P. Calderon, Cauchy integrals on Lipschitz curves and related operators, Proc. Nat. Acad.
Sci. U.S.A. 74 (1977), 1324{1327.
[CDMcY] M. Cowling, I. Doust, A. McIntosh and A. Yagi, Banach space operators with a bounded
H 1 functional calculus, Journal of the Australian Math. Society, Series A, to appear.
[Ch]
M. Christ, Lectures on Singular Integral Operators, CBMS, 1991.
[CJS]
R.R. Coifman, P. Jones and S. Semmes, Two elementary proofs of the L2 boundedness
of Cauchy integrals on Lipschitz curves, Journal of the American Mathematical Society 2
(1989), 553{564.
[CM]
R.R. Coifman and Y. Meyer, Au-dela des operateurs pseudo{di erentiels, Asterisque 57,
Societe Mathematique de France, 1978.
[CM1]
R.R. Coifman and Y. Meyer, Fourier analysis of multilinear convolutions, Calderon's theorem, and analysis on Lipschitz curves, Lecture Notes in Mathematics, Springer-Verlag 779
(1980), 104{122.

OPERATOR THEORY AND HARMONIC ANALYSIS

59

[CMcM] R.R. Coifman, A. McIntosh and Y. Meyer, L'integrale de Cauchy de nit un operateur borne
sur L2 pour les courbes lipschitziennes, Annals of Mathematics 116 (1982), 361{387.
[Co]
M. Cowling, Harmonic analysis in semigroups, Annals of Mathematics 117 (1983), 267-283.
[CRW] R.R. Coifman, R. Rochberg and G. Weiss, Transference methods in analysis, CBMS regional
conference series in mathematics, 31, A.M.S., 1977.
[CW]
R.R. Coifman, and G. Weiss, Linear spaces and Approximation, P. Butzer and B. Sz. Nagy
eds., Birkhauser Verlag, 1978, pp. 53-67.
[CZ]
A.P. Calderon and A. Zygmund, On the existence of certain singular integrals, Acta Math.
88 (1952), 85-139.
[D]
E.B. Davies, Heat Kernels and Spectral Theory, Cambridge University Press, 1989.
[D1]
B.E.J. Dahlberg, Estimates of harmonic measure, Arch. Rat. Mech. Anal. 65 (1977), 275{
288.
[deL]
Ralph deLaubenfels, Existence Families, Functional Calculi, and Evolution, Lecture Notes
in Mathematics, 1570, Springer-Verlag, Berlin, 1994.
[DJS]
G. David, J.-L. Journe and S. Semmes, Operateurs de Calderon{Zygmund, fonctions para{
accretives et interpolation, Revista Matematica Iberoamericana 1 (1985), 1{57.
[DMc]
X. T. Duong and A. McIntosh, Functional calculi of second order elliptic partial di erential
operators with bounded measurable coecients, J. of Geometric Analysis, to appear.
[DR]
X. T. Duong and D. Robinson, Semigroup kernels, Poisson bounds and holomorphic functional calculus, Journal of Functional Analysis, to appear.
[DS]
N. Dunford and J.T. Schwartz, Linear Operators, Part I, General Theory, Interscience,
1958.
[Du]
X. T. Duong, H1 functional calculus of elliptic partial di erential operators in Lp spaces
(1990), PhD thesis, Macquarie University.
[GW]
M. Gruter and K. O. Widman, The Green function for uniformly elliptic equations,
Manuscripta Mathematica 37 (1982), 303{343.
[H]
L. Hormander, Estimates for translation invariant operators in Lp spaces, Acta Math.
104 (1960), 93-104.
[JK]
D.S. Jerison and C.E. Kenig, The Neumann problem on Lipschitz domains, Bulletin of the
American Mathematical Society 4 (1981), 203{207.
[K]
Tosio Kato, Perturbation Theory for Linear Operators, second edition, Springer{Verlag,
1976.
[Ke]
Carlos Kenig, Harmonic Analysis Techniques for Second Order Elliptic Boundary Value
Problems, C.B.M.S., Vol 83, American Mathematical Society, 1994.
[L]
E.R. Lorch, Spectral Theory, Oxford University Press, 1962.
c
[LM Q] C. Li, A. McIntosh and T. Qian, Cli ord algebras, Fourier transforms, and singular convolution operators on Lipschitz surfaces, Revista Matematica Iberoamericana, 10 (1994),
665{721.
[LMcS] C. Li, A. McIntosh and S. Semmes, Convolution singular integrals on Lipschitz surfaces,
Journal of the American Mathematical Society 5 (1992), 455{481.
[M]
Yves Meyer (vol. III with R.R. Coifman), Ondelettes et Operateurs, I, II, III, Hermann,
editeurs des sciences et des arts, 1990.
c
[M ]
Alan McIntosh, Operators which have an H 1 functional calculus, Miniconference on Operator Theory and Partial Di erential Equations, 1986, Proceedings of the Centre for Mathematical Analysis, ANU, Canberra, 14 (1986), 210{231.

DAVID ALBRECHT, XUAN DUONG AND ALAN McINTOSH

60
[McI]
[McM]
[McQ]
[McY]
[PS]
[R]
[RN]
[Ro]
[Ru]
[S]
[SC]
[Se]
[SS]
[St1]
[St2]
[T]
[V]
[Y]

Alan McIntosh, Cli ord algebras, Fourier theory, singular integrals, and harmonic functions
on Lipschitz domains, Cli ord Algebras in Analysis and Related Topics, edited by John
Ryan (1995), CRC Press.
Alan McIntosh and Yves Meyer, Algebres d'operateurs de nis par des integrales singulieres,
Comptes Rendus Acad. Sci., Paris, Ser. I, Math 301 (1985), 395{397.
A. McIntosh and T. Qian, Convolution singular integral operators on Lipschitz curves, Lecture Notes in Mathematics, Springer{Verlag, Proceedings of the Special Year on Harmonic
Analysis at Nankai Institute of Mathematics, Tianjin, China 1494 (1991), 142{162.
A. McIntosh and A. Yagi, Operators of type ! without a bounded H1 functional calculus,
Miniconference on Operators in Analysis, 1989, Proceedings of the Centre for Mathematical
Analysis, ANU, Canberra 24 (1989), 159{172.
J. Pruss and H. Sohr, Imaginary powers of elliptic second order di erential operators in
Lp spaces., Hiroshima Math. Journal 23 (1993), 161{192.
D.W. Robinson, Elliptic Operators and Lie Groups, Oxford University Press, 1991.
F. Riesz and B. Sz.-Nagy, Functional Analysis, translated by F. Boron, Ungar, 1955.
H.L. Royden, Real Analysis, third edition, Macmillan, 1988.
W. Rudin, Real and Complex Analysis, McGraw{Hill, 1966.
G.F. Simmons, Introduction to Topology and Modern Analysis, McGraw{Hill, 1963.
L. Salo -Coste, Analyse sur les groupes de Lie a croissance polynomiale, Arkiv for Mat. 28
(1990), 315{331.
R. Seeley, Norms and domains of the complex powers AzB , Amer. J. Math. 93 (1971),
299{309.
Stephen Semmes, Square function estimates and the T (b) theorem, Proceedings of the
American Mathematical Society 110 (1990), 721{726.
E.M. Stein, Singular Integrals and Di erentiability Properties of Functions, Princeton University Press, 1970.
E.M. Stein, Harmonic Analysis, Princeton University Press, 1993.
H. Tanabe, On Green's functions of elliptic and parabolic boundary value problems, Proc.
Japan Acad. 48 (1972), 709{711.
Greg Verchota, Layer potentials and regularity for the Dirichlet problem for Laplace's equation in Lipschitz domains, Journal of Functional Analysis 59 (1984), 572{611.
Atsushi Yagi, Coincidence entre des espaces d'interpolation et des domaines de puissances
fractionnaires d'operateurs, C.R. Sc. Paris, Serie A 299 (1984), 173{176.

School of Mathematics, Physics, Computing and Electronics, Macquarie University,


N.S.W. 2109, Australia. David Albrecht is now in the Computer Science Department,
Monash University, Clayton, Victoria 3168.

E-mail address : dwa@bruce.cs.monash.edu.au, duong@mpce.mq.edu.au, alan@mpce.mq.edu.au

You might also like