You are on page 1of 330

Function Spaces

Fourth Conference on Function Spaces


May 14-19,2002
Southern Illrnois University at Edwardsville

Krzysztof Jarosz
Editor
CONTEMPORARY
MATHEMATICS
328

Function Spaces
Fourth Conference on Function Spaces
May 14-19,2002
Southern Illinois University at Edwardsville

Krzysztof Jarosz
Editor

American Mathematical Society


Providence, Rhode Island
Editorial Board
Dennis DeTurck, managing editor
Andreas Blass Andy R. Magid Michael Vogelius
This volwne contains the proceedings of the Fourth Conference on Function Spaces,
held May 14-19, 2002, at Southern Illinois University at Edwardsville.

2000 Mathematics Subject Classification. Primary 32H02, 46E25, 46H05, 46JlO, 46J15,
46L07, 47AlO, 47B38, 47LlO, 54D05.

Library of Congress Cataloging-in-Publication Data


Conference on Function Space (4th: 2002 : Southern Illinois University at Edwardsville)
Function spaces : Fourth Conference on Function Spaces, May 14-19, 2002, Southern Illinois
University at Edwardsville / Krzysztof Jarosz, editor.
p. cm. - (Contemporary mathematics, ISSN 0271-4132 j 328)
Includes bibliographical references.
ISBN 0-8218-3269-7 (softcover : alk. paper)
1. Function spaces-Congresses. I. Jarosz, Krzysztof, 1953- II. Title. III. Contemporary
mathematics (American Mathematical Society) j v. 328

QA323.C66 2002
515'.73-dc21 2003045306

Copying and reprinting. Material in this book may be reproduced by any means for edu-
cational and scientific purposes without fee or permission with the exception of reproduction by
services that collect fees for delivery of documents and provided that the customary acknowledg-
ment of the source is given. This consent does not extend to other kinds of copying for general
distribution, for advertising or promotional purposes, or for resale. Requests for permission for
commercial use of material should be addressed to the Acquisitions Department, American Math-
ematical Society, 201 Charles Street, Providence, Rhode Island 02904-2294, USA. Requests can
also be made bye-mail to reprint-permissionalams. ~rg.
Excluded from these provisions is material in articles for which the author holds copyright. In
such cases, requests for permission to use or reprint should be addressed directly to the author(s).
(Copyright ownership is indicated in the notice in the lower right-hand corner of the first page of
each article.)
2003 by the American Mathematical Society. All rights reserved.
The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.
@) The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at http://www . ams. org/
10987654321 08 07 06 05 04 03
Contents

Preface v
Components of resolvent sets and local spectral theory
PIETRO AlENA AND FERNANDO VILLAFANE 1
The Fejer-Riesz inequality and the index of the shift
JOHN R. AKEROYD 15

A Cauchy-Green formula on the unit sphere in C 2


JOHN T. ANDERSON AND JOHN WERMER 21

On a-dual algebras
HUGO ARIZMENDI, ANGEL CARRILLO, AND LOURDES PALACIOS 31

A connected metric space that is not separably connected


RICHARD M. ARON AND MANUEL MAESTRE 39

Weighted Chebyshev centres and intersection properties of balls in Banach


spaces
PRADIPTA BANDYOPADHYAY AND S. DUTTA 43

The boundary of the unit ball in HI-type spaces


PAUL BENEKER AND JAN WIEGERINCK 59

Complete isometries - an illustration of noncommutative functional analysis


DAVID P. BLECHER AND DAMON M. HAY 85

Some recent trends and advances in certain lattice ordered algebras


KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI 99

An extension of a theorem of Wermer, Bernard, Sidney and Hatori to


algebras of functions on locally compact spaces
EGGERT BRIEM 135

Some mapping properties of p-summing operators with Hilbertian domain


QINGYING Bu 145

The unique decomposition property and the Banach-Stone theorem


AUDREY CURNOCK, JOHN HOWROYD, AND NGAI-CHING WONG 151

A survey of algebraic extensions of commutative, unital normed algebras


THOMAS DAWSON 157

iii
iv CONTENTS

Some more examples of subsets of Co and L1 [0, 1] failing the fixed point
property
P. N. DOWLING, C. J. LENNARD, AND B. TURETT 171
Homotopic composition operators on HOC) (Bn)
PAMELA GORKIN, RAYMOND MORTINI, AND DANIEL SUAREZ 177
Characterization of conditional expectation in terms of positive projections
J. J. GROBLER AND M. DE KOCK 189
The Krull nature of locally C* -algebras
MARINA HARALAMPIDOU 195
Characterizations and automatic linearity for ring homomorphisms on
algebras of functions
OSAMU HATORI, TAKASHI ISHII, TAKESHI MIURA, AND SIN-EI
TAKAHASI 201
Carleson embeddings for weighted Bergman spaces
HANS JARCHOW AND URS KOLLBRUNNER 217
Weak *-extreme points of injective tensor product spaces
KRZYSZTOF JAROSZ AND T. S. S. R. K. RAO 231
Determining sets and fixed points for holomorphic endomorphisms
KANG-TAE KIM AND STEVEN G. KRANTZ 239
Localization in the spectral theory of operators on Banach spaces
T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN 247
Abstract harmonic analysis, homological algebra, and operator spaces
VOLKER RUNDE 263
Relative tensor products for modules over von Neumann algebras
DAVID SHERMAN 275
Uniform algebras generated by unimodular functions
STUART J. SIDNEY 293
Analytic functions on compact groups and their applications to almost
periodic functions
THOMAS TONEV AND S. A. GRIGORYAN 299
Preface

The Fourth Conference on Function Spaces was held at Southern Illinois Uni-
versityat Edwardsville from May 14 to May 19, 2002. It was attended by over 100
participants from 25 countries. The lectures covered a broad range of topics, in-
cluding spaces and algebras of analytic functions of one and of many variables (and
operators on such spaces), LP-spaces, spaces of Banach-valued functions, isometries
of function spaces, geometry of Banach spaces, and other related subjects. The
main purpose of the conference was to bring together mathematicians interested
in various problems within the general area of function spaces and to allow a free
discussion and exchange of ideas with people working on exactly the same problems
as well as with people working on related questions. Hence, most of the lectures,
and therefore the papers in this volume, have been directed to non-experts. A num-
ber of articles contain an exposition of known results (known to experts) and open
problems; other articles contain new discoveries that are presented in a way that
should be accessible also to mathematicians working in different areas of function
spaces.
The conference was the fourth in a sequence of conferences on function spaces at
SlUE, with the first held in the spring of 1990, the second in the spring of 1994, and
the third one in the spring of 1998. The proceedings of the first two conferences
were published with Marcel Dekker in the Lecture Notes in Pure and Applied
Mathematics series (#136 and #172); the proceedings of the third conference were
published by the AMS in the Contemporary Mathematics series (#232).
The abstracts, the schedule of the talks, and other information, as well as
the pictures of the participants, are available on the conference Web page at
http://www.siue.edu/MATH/conference/.
The conference was sponsored by grants from Southern Illinois University and
from the National Science Foundation. The editor would like to thank everyone
who contributed to the proceedings: the authors, the referees, the sponsoring in-
stitutions, and the American Mathematical Society.
The editor would also like to express very special thanks to his wife, Dorota,
for her active professional help during all of the stages of the organization - without
her help the conference and the proceedings would not have been possible.

Krzysztof Jarosz
Contemporary Mathematics
Volume 328, 2003

Components of resolvent sets and local spectral theory

Pietro Aiena and Fernando Villafane

ABSTRACT. In this paper we shall study the components of various resolvent


sets associated with some spectra originating from Fredholm theory. In par-
ticular, we obtain a classification of these components by using, in the case of
operators of Kato type, the equivalences between the single valued extension
property at a point and some kernel-type and range type conditions established
in [2], [6], [3] and [5]. We also show that certain subspace valued mappings
coincide on the components of the Kato type resolvent and give a precise
description of the operators having empty Kato type spectrum.

1. Single valued extension property

Throughout this paper, T is assumed to be a bounded linear operator on a


complex Banach space X and L( X) will denote the algebra of all bounded linear
operators on L(X). If x E X, the local resolvent set of T at x E X, denoted by
PT(X), is defined as the union of all open subsets U of C such that there is an
analytic function f : U - X which satisfies the equation
(1.1) (>.1 - T)f(>..) = x for all >.. E U .
The local spectrum aT (x) of T at x is defined by aT (x) := C \ PT (x) and obviously
aT(x) ~ a(T), where a(T) denotes the spectrum of T. The operator T E L(X)
is said to have the single-valued extension property at >"0 E C ( SVEP at >"0 for
brevity) if, for every neighborhood U of >"0, the only analytic function f : U - X
which satisfies the equation (>.1 - T)f(>..) = 0 for all >.. E U is the function f == O.
The operator T E L(X) is said to have SVEP if T has SVEP at every point>" E C.
Clearly, if T has SVEP at >"0, then the analytic solution of (1.1) in a neighborhood
U of >"0 is uniquely determined.
The SVEP was first introduced by Dunford [8], [9] and has later received a
systematic treatment in Dunford-Schwartz [10]. The SVEP at a point was first
introduced by Finch [11] and successively investigated by several authors, see [20],
[28], [2], [6], [3] and [5] . The basic role of SVEP arises in local spectral theory,
since every operator which satisfies the so-called Bishop's property (13) enjoys this

1991 Mathematics Subject Classification. Primary 47AlO, 47A11. Secondary 47A53, 47A55.
Key words and phrases. Single valued extension property, semi-regular operators, Kato de-
composition property .
The research was supported by the International Cooperation Project between the University
of Palermo (Italy) and the University of Barquisimeto.

2003 American Mathematical Society


1
2 PIETRO AlENA AND FERNANDO VILLAFANE

property, see [18] for definition and results. Recall that an operator T E L(X)
on a Banach space X is said to be decomposable if for every open cover {UI , U2 }
of C there exist T-invariant closed linear subspaces Xl and X 2 of X for which
X = Xl + X 2 , a(T IXd ~ Ul and a(T IX2 ) ~ U2 The class of decomposable
operators contains, for instance, all normal operators, all spectral operators, all
operators with a non-analytic functional calculus and all compact operators, or
more generally all operators with a totally disconnected spectrum. Note that T
is decomposable if and only if T and its dual T* have property ((3), see Theorem
2.5.19 [18]. Consequently, if T is decomposable, then both T and T* have SVEP.
For an arbitrary subset F of C let XT(F) be the local spectral subspace associated
with F, defined by XT(F) := {x EX: aT(x) ~ F}. If F is a closed subset of C,
let XT(F) be the glocal spectral subspace associated with F, defined as the set of
all x E X for which there exists an analytic function f : C \ F ---+ X which satisfies
(>.J - T)f(A) = x for all A E C \ F. Clearly, XT(F) and XT(F) are (not necessarily
closed) linear subspaces of X with XT(F) ~ XT(F) for all closed sets F ~ C. Note
that, by Proposition 3.3.2 of [18], the identity XT(F) = XT(F) holds for all closed
sets F ~ C precisely when Thas SVEP, and this is the case if and only if X T (0) =
{O}, see Proposition 1.2.16 of [18]. The SVEP at a point Ao may be characterized
in a similar way: T has SVEP at Ao if and only if ker (AoI - T) n X T (0) = {O}, cf.
[1, Theorem 1.9].

Two important subspaces in Fredholm theory are the hyperrange of T, de-


fined by TOO(X) = n:'=l Tn(X), and the hyperkernel of T defined by NO(T) =
U:'=l ker Tn. Recall that the ascent of an operator T is the smallest non-negative
integer p := p(T) such that ker TP = ker TP+l. If such integer does not exist, we
put p(T) = 00. Obviously, if T has finite ascent p then No (T) = ker TP.
Analogously, the descent q := q(T) of an operator T is the smallest non-negative
integer q such that Tq(X) = Tq+l(X). If such integer does not exist we put
q(T) = 00. Also, if T has finite descent q then TOO(X) = Tq(X). It is well-known
that if p(T) and q(T) are both finite then p(T) = q(T), see [15, Proposition 38.3].
Furthermore, p(AoI - T) = q(AoI - T) < 00 if and only if Ao is a pole of the resol-
vent R(A, T) := (>.J - T)-l, [15, Proposition 50.23].

Associated with T E L(X) there is another linear subspace of X, the quasi-


nilpotent part of T defined as

Ho(T) := {x EX: lim IITnxll l / n = O}.


n-->oo

Evidently, NO(T) ~ Ho(T). Moreover, Ho(T) = X if and only if T is quasi-


nilpotent, i.e. a(T) = {O}, see [20, Theorem 1.5].
The following decomposition property studied by Mbekhta [20], [19], Mbekhta
and Ouahab [21], has its origin in the classical treatment of perturbation theory due
to Kato [17], who showed an important decomposition for semi-Fredholm operators:

DEFINITION 1.1. An operator T E L(X) is called semi-regular if T(X) is closed


and NO(T) ~ TOO(X). An operator T E L(X) is said to be of Kato type if there
exists a pair of T-invariant closed subspaces (M, N) such that X = M EB N, the
restriction T 1M is semi-regular and T IN is nilpotent. The pair (M, N) is called a
generalized Kato decomposition ( GKD, for brevity) for T.
COMPONENTS OF RESOLVENT SETS AND LOCAL SPECTRAL THEORY 3

If, additionally, in the definition above we assume that N is finite-dimensional


then T is said to be essentially semi-regular, see Rakocevic [25] and Miiller [24].
By Proposition 3.1.6 of [18], T is semi-regular if and only if T* is semi-regular.
Analogously, by Corollary 3.4 of [24], T is. essentially semi-regular if and only if
T* is essentially semi-regular. Moreover, if T is essentially semi-regular then Tn is
essentially semi-regular for all n E N, again by Corollary 3.4 of [24].
It should be noted that the range of an essentially semi-regular operator is
always closed. In fact, if (!vI, N) is a GKD for T with N finite-dimensional, then
T(X) is the direct sum of T(!vI), which is closed because T I !vI is semi-regular,
and of T(N), which is finite-dimensional.
Two very important class of essentially semi-regular operators is given by the
class of upper semi-Fredholm operators, defined by
~+(X) := {T E L(X) : dim ker T < 00, T(X) is closed}
and the class of all lower semi-Fredholm operators defined by
~_(X) := {T E L(X) : codim T(X) < oo},
see Kato [17, Theorem 4] or West [31]. The class of Fredholm operators is defined
by ~(X) := ~+(X)n~_(X). Note that a semi-Fredholm operator T is semi-regular
if and only if its jump j(T) is zero, see ([31, Proposition 2.2].
Moreover, if T is of Kato type, and in particular if T is semi-regular, then
TOO(X) is closed with T(TOO(X)) = TOO(X), see Theorem 2.3 and Theorem 2.4 of
[2], and TOO(X) coincides with the analytical core K(T) := {x E X: there exist a
constant c> 0 and a sequence of elements Xn E X such that Xo = x, TX n = Xn-l,
and Ilxnll ::; cnllxll for all n EN}. It should be noted that both subspaces K(T)
and Ho(T) admit a local spectral characterization. In fact,
(1.2) K(T) = XT(C \ {O}) = {x EX: 0 f/. O"T(X)},
see Mbekhta [20], Vrbova [30] and also Propositions 3.3.7 and 3.3.13 of [18], and
Ho(T) = XT({O}, see Propositions 3.3.7 and 3.3.13 of [18]. Therefore, if T has
SVEP, then Ho(T) = X T ( {O}).

In the following lemma by A.L and its proof we denote the annihilator of a
subset A ~ X, and by .L B the pre-annihilator of a subset B ~ X*.
LEMMA 1.2. For every T E L(X), the following statements hold
(i) Ho(T) ~.L K(T*) and K(T) ~.L Ho(T*).
(ii) If T is a Kato type operator and the pair (!vI, N) is a GKD for T then

K(T) = K(T I !vI) = K(T) n!vI,


and
(1.3) Ho(T) = Ho(T I !vI) EB Ho(T I N) = Ho(T I !vI) EB N.
(iii) If T is essentially semi-regular then

(1.4) Ho(T) = NO(T) =.L K(T*) and K(T) =.L Ho(T*),

(iv) 1fT is semi-regular then Ho(T) ~ K(T).


4 PIETRO AlENA AND FERNANDO VILLAFANE

PROOF. (i) See Proposition 4.1 of [5].


(ii) The proof of the equality K(T) = K(T I M) may be found in [1]. The
second equality in (1.3) is clear, since the nilpotency of TIN implies Ho(T I N) =
N. The inclusion Ho(T I M) + Ho(T I N) ;;; Ho(T) is evident. To show the
opposite inclusion, consider an arbitrary element x E Ho(T) and set x = y + z,
with y E M, zEN. Since TIN is quasi-nilpotent then N = Ho(T) ;;; Ho(T).
Therefore y = x - z E Ho(T) n M = Ho(T I M) and consequently Ho(T) ;;; Ho(T I
M) + Ho(T IN).
(iii) Assume that T is essentially semi-regular and hence also T* essentially
semi-regular. Then T*n is essentially semi-regular for all n E N, so that T*n(X*)
is closed for all n E N. From part (i) we know that NO(T) ;;; Ho(T) ;;;1. K(T*), so
that, to show the first two equalities of (1.4), we need only to prove the inclusion
1. K(T*) ;;; NO(T). For every T E L(X) and every n E N we have ker Tn ;;; NO(T)
and hence NO(T)1. ;;; ker Tn1. = T*n(X*), because the last subspace is closed for
-:--.------:-::::,-1.
all n E N. From this we easily obtain that NO(T) ;;; T*OO(X*) = K(T*), where
the last equality holds since T* is essentially semi-regular and hence of Kato type.
Consequently, 1. K(T*) ;;; NO(T). Thus the first two equalities of (1.4) are proved.
The equality K(T) =1. Ho(T*) is proved in a similar way.
(iv) The semi-regularity of T entails that NO(T) ;;; TOO(X) = K(T), so that,
by part (ii), Ho(T) = Noo(T) ;;; K(T) = K(T) , and this concludes the proof. 0
We have already observed that an operator T E L(X) has SVEP at >'0 precisely
when ker (>'01 - T) n K(>'ol - T) = {a}. From the inclusion
ker (>'01 - T) ;;; NO(>'ol - T) ;;; Ho(>'ol - T)
it then follows that the condition Ho(>'ol - T) n K(>'ol - T) = {a} implies that T
has SVEP at >'0' Example 2.5 of [5] shows that SVEP for T at a point does not
necessarily imply that Ho(>'ol - T) n K(>'ol - T) = {a}. In [2] it has been shown
that also the condition NO(>'ol - T) n (>'01 - T)OO(X) = {a} implies SVEP at >'0
for T. The next result shows that these implications are actually equivalences in
the case that >'01 - T is of Kato type.
THEOREM 1.3. If >'01 - T is of Kato type then the following properties are
equivalent:
(i) T has the SVEP at >'0;
(ii) Ho(>'ol - T) n K(>'ol - T) = {a};
(iii) Ho(>'ol - T) is closed;
(iv) >'01 - T has finite ascent;
(v) NO(>'ol - T) n (>'01 - T)OO(X) = {a}.
Furthermore, if >'01 - T E L(X) is essentially semi-regular, the assertions (i)-
(v) are equivalent to the following conditions:
(vi) Ho(>'ol - T) is finite-dimensional.
(vii) NO(>'ol - T*) + (>'01 - T*)(X*) is weak *-dense in X*;
(viii) Ho(>'ol - T*) + (>'01 - T*)(X*) is weak *-dense in X*;
(ix) Ho(>'ol - T*) + K(>'ol - T*) is weak *-dense in X*.
In this case >'01 - T E iP+(X).
COMPONENTS OF RESOLVENT SETS AND LOCAL SPECTRAL THEORY 5

PROOF. The equivalence of (i), (ii), (iii) and (iv) has been established in [3,
Theorem 2.6 and Corollary 2.7]. The equivalence (i) {:} (v) has been proved in
Theorem 2.6 of [2], see also Theorem 2.1 of the present paper.
Assume now that AoI - T is essentially semi-regular. We may assume that AO = O.
(i) {:} (vi) Obviously, if Ho(T) is finite-dimensional then Ho(T) is closed, so T
has SVEP at 0, by the equivalence (i) {:} (iii).
Conversely, if T has SVEP at 0 then also TIM has SVEP at 0, since the
local SVEP is inherited by the restrictions to closed invariant subspaces. The semi-
regularity of TIM then implies that TIM is injective, see Theorem 2.14 of [1]
and therefore NO(T) = {O}. By part (iii) of Lemma 1.2 we then conclude that
Ho(T I M) = NO(T 1M) = {O}. From part (ii) of Lemma 1.2 it follows that
Ho(T) = {O} EB N = N is finite-dimensional.
Finally, if T is essentially semi-regular then also Tn is essentially semi-regular
and therefore, the ranges Tn(x) are closed for all n E N. From Theorem 4.3 of
[5] it then follows that the condition (i) is equivalent to each one of the conditions
(vii), (viii) and (ix).
It remains to establish that (i) implies that AoI - T E ~+(X). Clearly, if
Ho(AoI - T) is finite- dimensional then also its subspace ker (AoI - T) is finite-
dimensional. Since (AoI -T)(X) is closed we then conclude that AoI -T E ~+(X).
o
The next Theorem 2.1 will show that, if AoI - T of Kato type, then Ho(AoI -
T) n K(AoI - T) = NO(AoI - T) n (AoI - T)OO(X).
The following characterizations of SVEP for the dual T* are dual, in a sense, to
those given in Theorem 1.3.
THEOREM 1.4. Suppose that AoI - T is of Kato type. Then the following
statements are equivalent:
(i) T* has SVEP at AO;
(ii) X = Ho(AoI - T) + K(AoI - T);
(iii) AoI - T has finite descent;
(iv) X = NO(AoI - T) + (AoI - T)OO(X);
(v) Ho(AoI - T) + K(AoI - T) is norm-dense in X;
(vi) NO(AoI - T) + (AoI - T)OO(X) is norm-dense in X;
Furthermore? if AoI - T E L(X) is a essentially semi-regular then the assertions
(i)-(vi) are equivalent to the following conditions:
(vii) K(AoI - T) is finite-codimensional;
(viii) No (AoI - T) + (AoI - T) (X) is norm-dense in X;
(ix) Ho(AoI - T) + (AoI - T)(X) is norm-dense in X.
In this case AoI - T E ~_(T).
PROOF. Also here we may assume that AO = 0 and T is of Kato type.
The equivalence of (i), (ii) and (iii) has been established in Theorem 2.9 of
[3]. The equivalence of (i) and (iv) has been proved in Theorem 2.9 of [2], see
also Theorem 2.1 of the present paper. Clearly, (ii) ::::} (v), (iv) ::::} (vi). The
implications (v) ::::} (i) and (vi) ::::} (i) have been proved in Corollary 4.2 of [5], so
that the statements (i)-(vi) are equivalent.
Now, assume that T is essentially semi-regular. Then Tn(x) is closed for
6 PIETRO AlENA AND FERNANDO VILLAFANE

all n E N, so that, by Theorem 4.3 of [5], the statements (i), (viii) and (ix) are
equivalent.
To conclude the proof note first that if (M, N) is a GKD for T then the pair
(Nl.,Ml.) is a GKD for T*. Now, if T* has SVEP at 0, then, as observed in the
proof Theorem 1.3, T* I Nl. is injective and therefore, see Lemma 2.8 of [2], TIM
is surjective. Therefore K(T) = K(T I M) = M is finite-codimensional, so that the
implication (i) => (vii) is proved.
Conversely, suppose that the analytical core K(T) is finite-co dimensional. From
K(T) = TOO(X) ~ Tn(x) we deduce that q(T) < 00, so that (vii) implies (iii), and
the proof of the equivalences is complete .
Finally, from the inclusion K(>"ol -T) ~ (>"01 -T)(X) we infer that, if K(>"o/-
T) finite-co dimensional, then also (>"01 - T)(X) is finite-codimensional, so that
>"01 - T E <L(X). D
Recall that T E L(X) is called bounded below if T is injective and has closed
range T(X). It is easily seen from definition of SVEP that, if the approximate point
spectrum
l7ap (T) := {>.. E C : >..I - T is not bounded below}
does not cluster at >"0, then T has SVEP at >"0 and, dually, if the surjectivity
spectrum
l7 su (T) = {>.. E C : >..I - T is not surjective}
does not cluster at >"0, then T* has SVEP at >"0'

The next result shows that for Kato type operators these implications may be
reversed.
THEOREM 1.5. If >"01 - T is of Kato type, then the following equivalences hold:

(i) T has the SVEP at >"0 precisely when l7ap(T) does not cluster at >"0, [6,
Theorem 2.2];
(ii) T* has the SVEP at >"0 precisely when l7su(T) does not cluster at >"0,' [6,
Theorem 2.5].

2. Components
In this section we shall take a closer look at the components of some resol-
vent sets associated with the various spectra originating from Fredholm theory. In
particular, we shall obtain a classification of these components, by using the con-
stancy of some mappings and the equivalences between the SVEP at a point and
the kernel-type and range type conditions, established in the previous section.
For an operator T E L(X), we consider the following parts of the ordinary
spectrum: the Kato spectrum
l7k(T) := {>.. E C : >..I - T is not semi-regular},
and the essential Kato spectrum
l7ke(T) := {>.. E C : >..I - T is not essentially semi-regular}.
Moreover, we define
l7kt(T) := {>.. E C : >..I - T is not of Kato type}.
COMPONENTS OF RESOLVENT SETS AND LOCAL SPECTRAL THEORY 7

It is known that the three sets O'k(T), O'kt(T) and O'ke(T) are closed, for the first
set see [18, Proposition 3.1.9], for the other two sets see [4, Corollary 1]. Moreover,
O'k(T) and O'ke(T) are nonempty, since the first spectrum contains the boundary of
O'(T), see [18, Proposition 3.1.6], while the second spectrum contains the boundary
of the Fredholm spectrum O'f(T) := p, E C : >.J - T ~ cJ>(X)}, see [24, Theorem
3.8]. Next we shall show that O'kt(T) is non-empty precisely when O'(T) is not a
finite set of poles.
Let Pk(T) := C \ O'k(T), Pkt(T) := C \ O'kt(T) and Pke(T) := C \ O'ke(T) be
the resolvents associated with these spectra. The sets Pk(T), Pkt(T) and Pke(T)
are open subsets of C, so they may be decomposed in connected disjoint open
non-empty components. Clearly,
(2.1)
Note that for every T E L(X) we have Pk(T) = Pk(T*) and Pke(T) = Pke(T*).
In [29] 6 Searcoid and West showed the constancy of the mappings

on the components of the semi-Fredholm resolvent Psf(T) := C \ O'sf(T), where


O'sf(T) is the semi-Fredholm spectrum defined by
O'sf(T) := {A E C: >.J - T ~ cJ>+(X) U cJ>_(X)}.
From the Kato decomposition for semi-Fredholm operators we easily obtain the
following inclusions
(2.3)
The work of 6 Searcoid and West [29] extended previous results established by
Homer [16], by Goldmann and Kraekovskii [13], [14], and by Saphar [26], which
have established the constancy of the functions

on a component of the semi-Fredholm resolvent Psf(T), except for the discrete


subset of points for which AI - T is not semi-regular.
In the same vein, Forster [12] showed that the mappings

are constant as A ranges through a component of the Kato resolvent Pk(T).


The constancy of these mappings has also been studied by Mbekhta and Oua-
hab [22], which showed the constancy of the mappings
(2.4) A ~ Ho(>.J - T) + K(>.J - T), A ~ Ho(>.J - T) n K(>.J - T)
on the components of Pkt(T). The next result shows that the mappings (2.2) and
(2.4) coincide, respectively, on the components of Pkt(T), so that the Mbekhta and
Ouahab result extends the previous result of 6 Searcoid and West.
THEOREM 2.1. Let >.J - T be of Kato type. Then
(i) NO(>.J - T) + (>.J - T)OO(X) = Ho(>.J - T) + K(>.J - T).
(ii) No(>.J - T) n (>.J - T)OO(X) = Ho(>.J - T) n K(>.J - T).
8 PIETRO AlENA AND FERNANDO VILLAFANE

PROOF. (i) Throughout this proof we may take>. = O. Let (M, N) be a GKD
for T such that (T I N)d = 0 for some integer dEN. By part (ii) of Lemma 1.2 we
know that K(T) = K(T I M) = K(T) n M. Moreover, by part (iv) of Lemma 1.2,
the semi-regularity of TIM implies that Ho(T I M) ~ K(T I M) = K(T). From
this we obtain
n K(T) = Ho(T) n (K(T) n M) = (Ho(T) n M) n K(T)
Ho(T)
= Ho(T I M) n K(T) = Ho(T 1M),

and therefore Ho(T) n K(T) = Ho(T 1M).


We claim that Ho(T) + K(T) = NEB K(T). From N ~ ker Td ~ Ho(T) we
obtain that NEB K(T) ~ Ho(T) + K(T). Conversely, from part (ii) of Lemma 1.2
we have
Ho(T) = NEB Ho(T I M) = NEB (Ho(T) n K(T)) ~ NEB K(T),
so that
Ho(T) + K(T) ~ (N EB K(T)) + K(T) ~ NEB K(T),
so our claim is proved.
Since K(T) = TOO(X) for every operator of Kato type, we obtain from the
inclusion N ~ ker Td ~ NO(T), that
Ho(T) + K(T) = NEB K(T) ~ NO(T) + TOO (X) ~ Ho(T) + K(T),
so the equality NO(T) + TOO (X) = Ho(T) + K(T) is proved.
(ii) Suppose again that>. = O. Let (M, N) be a GKD for T such that, for some
dEN, we have (T IN)d = O. Then ker Tn = ker (T IM)n for every natural n :2: d.
Since ker Tn ~ ker Tn+! for all n E N we then have
00 00

NO(T) = U ker Tn = U ker(T I M)n = NO(T 1M).


n2:d n2:d
The semi-regularity of TIM then implies, by part (iii) of Lemma 1.2, that
(2.5) NO(T) = NO(T I M) = Ho(T I M) = Ho(T) n M.
Next we show that the equality Ho(T) n M = Ho(T) n M holds. The inclusion
Ho(T) n M ~ Ho(T) n M is evident. Conversely, suppose that x E Ho(T) n M.
Then there is a sequence (xn) C Ho(T) such that Xn -+ x as n -+ 00. Let P be the
projection of X onto M along N. Then PX n -+ Px = x and PX n E Ho(T) n M.
Therefore x E Ho (T) n M.
Finally, from (2.5) and taking into account that K(T) n M = K(T) = TOO(X),
we then obtain
NO(T) n TOO (X) = Ho(T) n M n K(T) = Ho(T) n (M n K(T)) = Ho(T) n K(T),
so the proof is complete. D

From the constancy of the mappings>. -+ Ho(>.J - T)nK(>.J -T), or, which is
the same, of the mappings>. -+ Noo(>.J - T) n (>.J - T)OO(X), on the components
of Pkt(T) and the results established in the previous section we now obtain the
following classification.
COMPONENTS OF RESOLVENT SETS AND LOCAL SPECTRAL THEORY 9

THEOREM 2.2. LetT E L(X) and 0. a component ofpkt(T). Then the following
alternative holds:
either
(i) T has SVEP for every point ofn. In this case p(>.J - T) < 00 for all A E n.
Moreover, aap(T) does not have limit points in 0.; every point of 0., except possibly
for at most countably many isolated points, is not an eigenvalue of T.
or
(ii) T has SVEP at no point of n. In this case p(>.J - T) = 00 for all A E n.
Every point of 0. is an eigenvalue of T,
PROOF. (i) Suppose that T has SVEP at AO E n. Then, by Theorem 1.3,
Ho(>.J - T) is closed and
HO(AoI - T) n K(AoI - T) = HO(AoI - T) n K(AoI - T) = {O}.
Since the mapping A --+ Ho(>.J - T) n K(>.J - T) is constant on the component
0., then Ho(>.J - T) n K(>.J - T) = {O} for all A E 0. and this implies, again by
Theorem 1.3, that T has SVEP at every A E n. This is equivalent, also by Theorem
1.3, to saying that p(AI - T) < 00 for all A E n. Moreover, from Theorem 1.5,
aap(T) does not cluster in 0. and, consequently, every point of 0. is not an eigenvalue
of T, except a subset of 0. which consists of at most count ably many isolated points.

(ii) This is clear, again by Theorem 1.3. o


Recall that A E C is said to be a deficiency value for if >.J - T is not surjective.
THEOREM 2.3. Let T E L(X) and 0. a component of Pkt(T). Then the following
alternative holds:
either
(i) T* has SVEP for every point of n. In this case q(>.J - T) < 00 for all
A E n. Moreover, asu(T) does not have limit points in 0.; every point ofn, except
possibly for at most countably many isolated points, is not a deficiency value of T.
or
(ii) T* has the SVEP at no point of n. In this case q(>.J - T) = 00 for all
A E 0. and every A E 0. is a deficiency value of T.
PROOF. Proceed as in the proof of Theorem 2.2, combining the constancy on
the components of Pkt(T) of the mapping A E 0. --+ K(>.J - T) + Ho(>.J - T) (or,
equivalently, the constancy of the mapping A E 0. --+ NO(>.J -T) +(>.J -T)OO(X)),
with Theorem 1.4 and Theorem 1.5. 0
The previous results lead to a precise description of the operators whose Kato
type spectrum akt(T) is empty. Most of the results of the following theorem may be
found in Mbekhta [23] in the context of operators on Hilbert spaces. However, our
proofs, involving local spectral theory, are considerably simpler and are established
in the more general context of operators acting on Banach spaces.
Recall first that T E L(X) is algebraic if there exists a non-trivial polynomial h
such that h(T) = O.
THEOREM 2.4. For an operator T E L(X) the following statements are equiv-
alent:
(i) akt(T) is empty;
10 PIETRO AlENA AND FERNANDO VILLAFANE

(ii) AI - T has finite descent for every A E C;


(iii) AI - T has finite descent for every A E 8a(T), where 8a(T) is the topological
boundary of a(T);
(iv) a(T) is a finite set of poles of R(A, T);
(v) T is algebraic.
PROOF. (i) * (ii) Suppose that akt(T) = 0. Then Pkt(T) has an unique
component n = C and therefore, by Theorem 2.2, T has SVEP at every point of C,
since T has SVEP at the point ofthe resolvent p(T). On the other hand, if >.1 -T is
of Kato type, then also Al* - T* is of Kato type, see [4, Proposition 1). Therefore,
C = Pkt(T) = Pkt(T*) and consequently, by Theorem 2.3, also T* has SVEP. Since
>.1 - T is of Kato type, by Theorem 1.4, we then conclude that q(>.1 - T) < 00 for
every A E C.
(ii)* (iii) Obvious.
(iii) * (iv) Since T has SVEP at every A E 8a(T) then the condition q(>.1 -
T) < 00 entails that every A E 8a(T) is a pole of R(A, T), see Corollary 1 of [27),
and hence an isolated point of a(T). Clearly, this implies that a(T) = 8a(T), so
a(T) is a finite set of poles.
(iv) * (i) It suffices to prove that >.1 - T is of Kato type for all A E a(T).
Suppose that a(T) is a finite set of poles of R(A, T). If A E a(T), let P be the
spectral projection associated with the singleton {A}. Then X = M EB N, where
M := K(>.1 - T) = ker P and N := Ho(>.1 - T) = P(X), see the proof of
Theorem 1.6 of [19) or also Theorem 1 of [27). Since A is a pole of R(A, T),
by Proposition 50.2 of [15), >.1 - T has positive finite ascent and descent, and if
p := p(Aol - T) = q(>.1 - T), then N = ker (>.1 - T)P. From the classical Riesz
functional calculus we know that a(T I M) = a(T) \ {A}, [15, Theorem 49.1), so
that (>.1 - T) I M is bijective, while (>.1 - T I N)P = O. Therefore >.1 - T is of
Kato type for every A E C.
(iv) * (v) Assume that a(T) is a finite set of poles {Al, ,An}, where for every
i = 1, .. ,n with Pi we denote the order of Ai. Let h(A) := (AI - A)Pl ... (An - A)Pn.
Then, see Lemma 3.1.15 of [18),

h(T)(X) = n(Ail - T)Pi (X) = nK(Ai1 - T),


n

i=l
n

i=l

where the last equality follows since T has SVEP and Ail - T is of Kato type, see
Theorem 2.9 of [3). But the last intersection is {O}, since, by the local spectral
characterization of the analytical core (1.2), if x E K(Ai1 - T) n K(Ajl - T), with
Ai =1= Aj, then aT(x) ~ {Ai} n {Aj} = 0 and hence x = 0, since T has SVEP.
Therefore h(T) = O.
(v)* (i) As in the proof of (iv)* (i) it suffices to show that AI - T is of Kato
type for all A E a(T). Let h be a polynomial such that h(T) = O. From the spectral
mapping theorem we easily deduce that a(T) is a finite set {AI, ,An}. The
points AI, ,An are zeros of finite multiplicities of h, say k 1 , ,kn' respectively,
so that h(A) = (AI - A)kl ... (An - A)kn and hence
n
X = ker h(T) = $ker (Ail - T)ki,
i=l
COMPONENTS OF RESOLVENT SETS AND LOCAL SPECTRAL THEORY 11

see Lemma 3.1.15 of [18].


Now, suppose that A = Ai for some j and define
ho(A) := II(Ai - A)ki.
i"lj

We have
M := ker ho(T) = EB ker (Ail - T)ki
i"lj

and if N := ker (Ail - T)kj, then X = M EEl Nand M, N are invariant under
Ail - T. From the inclusion ker (Ail - T) ~ ker (Ail - T)k j = N, we infer that
the restriction of Ail - T on M is injective. It is easily seen that
(Ail - T)(ker (AJ - T)ki) = ker (AJ - T)ki, i =I- j,
so that (Ai 1- T)( M) = M. Hence the restriction of Ai 1- T on M is also surjective
and therefore bijective. Obviously, (Ajl - T) I N)k j = 0, so that Ajl - T is of Kato
type, as desired. 0
A bounded operator on a Banach space X is said to satisfy a polynomial growth

condition, if there exists a K > 0, a 8 > for which
IIexp(iAT)II ::; K(1 + IAleS) for all A E JR,
Examples of operators which satisfy a polynomial growth condition are her-
mitian operators on Hilbert spaces, nilpotent and projection operators, algebraic
operators with real spectra, see Barnes [1]. In Laursen and Neumann [18, Theo-
rem 1.5.19] it is shown that the class P(X) of operators which satisfy a polynomial
growth condition coincides with the class of all generalized scalar operators having
real spectra. As noted in Barnes [1], if T E P(X) and Aol - T has closed range
for some Ao E C then q(Aol - T) is finite. From Theorem 2.4 it follows that,
if T E P(X), then the condition (AI - T)(X) closed for all A E C implies that
O"kt(T) = 0. Other classes of operators for which O"kt(T) = 0 may be found in [23].
The classification of the components of Pes(T) may be easily obtained from
Theorem 2.2 and Theorem 2.3, once it has been observed, that the two sets Pes (T)
and Pkt(T) may be different only for a denumerable set, see for instance Corollary
1 of [4].
We now look at the components of PSf(T). Recall that for a semi-Fredholm
operator T E <P+(X) U <p_(X) we can consider the index defined by ind T :=
dim ker T - codim T(X). Clearly, from Theorem 2.2 and Theorem 2.3, T, as well
as T*, has SVEP either for every point or no point of a component n of PSf(T).
We can classify the components of Psf (T) as follows:
THEOREM 2.5. Let T E L(X) and n a component of PSf(T). For the BVEP,
the index, the ascent and the descent on 0 there are exactly the following four
possibilities:
(i) Both T and T* have BVEP at every point of n. In this case we have

ind (AI - T) = and p(AI - T) = q(AI - T) < 00 for every A E o. The eigenvalues
and deficiency values do not have a limit point in O. This case occurs exactly when
o intersects the resolvent p(T).
(ii) T has BVEP at the points of 0, while T* fails to have BVEP at the points
of o. In this case we have ind (AI - T) < 0, p(AI - T) < 00 and q(AI - T) = 00
12 PIETRO AlENA AND FERNANDO VILLAFANE

for every A E n.The eigenvalues do not have a limit point in n and every point of
n is a deficiency value.
(iii) T* has SVEP at the points of n, while T fails to have SVEP at the points
of n. In this case we have ind (AI - T) > 0, p(AI - T) = 00 and q(AI - T) < 00
for every A E n. The deficiency values do not have a limit point in n, while every
point of n is an eigenvalue.
(iv) Neither T or T* has SVEP at the points of n. In this case we have
p(AI - T) = q(AI - T) = 00 for every A E n. The index may assume every value
in Z; all the points of n are eigenvalues and deficiency values.
PROOF. The case (i) is clear from the results established in the previous section,
Theorem 2.2 and Theorem 2.3. The index ind (AI - T) = 0 by Proposition 38.6 of
[15J. In the case (ii) the condition p(AI - T) < 00 implies that AI - T has index less
or equal to 0, while the condition q(AI - T) = 00 excludes that ind (AI - T) = 0,
see Proposition 38.5 of [15J.
A similar argument shows in the case (iii) that ind (AI - T) > O.
The statements of (iv) are clear. 0

The following corollary establishes that a very simple classification of the com-
ponents of semi-Fredholm resolvent may be obtained in the case that T, or T* has
SVEP. Recall that the case that both T and T* have SVEP applies in particular
to the decomposable operators.
COROLLARY 2.6. Let T E L(X) and n any component of PSf(T). If T has
SVEP then only the case (i) and (ii) of Theorem 2.5 are possible, while if T* has
SVEP only the case (i) and (iii) are possible. Finally, if both T and T* have SVEP
then only the case (i) is possible.
In the next result we consider the components of Pk(T), which is the smallest
of the resolvent sets that we have considered.
THEOREM 2.7. Let T E L(X) and n any component of Pk(T). Then one of the
following possibilities occurs:
(i) Both T and T* have SVEP at every point of n. In this case we have
n ~ p(T).
(ii) T has SVEP at the points of n, while T* fails to have SVEP at every point
of n. In this case we have n n aap(T) = 0 and n ~ asu(T).
(iii) T* has SVEP at the points of n, while T* fails to have SVEP at the points
of n. In this case we have n n asu(T) = 0 and n ~ aap(T).
(iv) Neither T or T* have SVEP at the points of n. In this case we have
n ~ aap(T) n asu(T).
PROOF. (i) Let Ao E n. The subspaces M:= X and N := {O} give a GKD for
T and the subspaces ..L N = X and ..L M = {O} give a GKD for T*. As observed in the
proof of Theorem 1.3 and Theorem 1.4 if T has SVEP at Ao then Ho(AoI - T) =
N = {O} and if T* has SVEP at Ao then K(AoI - T) = M = X. Therefore
Ao E p(T).
(ii) In this case Ho(AI - T) = {O} and (AI - T)(X) is closed for every A E n.
If A ~ asu(T) then A E p(T) = p(T*) and this is impossible, since T* does not have
SVEP at A.
COMPONENTS OF RESOLVENT SETS AND LOCAL SPECTRAL THEORY 13

(iii) In this case K(M - T) = X for every A E n, so A f/. O'su(T). If A f/. O'ap(T)
then A E p(T) and this is impossible, since T does not have SVEP at the point A.
(iv) Use the same arguments as in part (ii) and (iii).
o
References
[IJ P. Aiena, o. Monsalve Operators which do not have the single valued extension property. J.
Math. Anal. Appl. 250, (2000),435-448.
[2J P. Aiena, O. Monsalve The single valued extension property and the generalized Kato de-
composition property. Acta Sci. Math. (Szeged) 67, (2001), 461-477.
[3J P. Aiena, M. L. Colasante, M. Gonzalez Operators which have a closed quasi-nilpotent part,
Proc. Amer. Math. Soc. 130, (2002), 2701-2710.
[4J P. Aiena, M. Mbekhta Characterization of some classes of operators by means of the Kato
decomposition. (1996), Boll. Un. Mat. It. 10-A, 609-21.
[5J P. Aiena, T. L. Miller, M. M. Neumann On a localized single valued extension property,
(2001), to appear on Proc. Royal Irish Acad.
[6J P. Aiena, E. Rosas The single valued extension property at the points of the approximate
point spectrum. to appear on J. Math. Anal. Appl.
[7J B. A. Barnes Operators which satisfy polynomial growth conditions. Pacific. J. Math. 138
(1989), 209-19.
[8J N. Dunford Spectral theory I. Resolution of the identity. Pacific J. Math. 2 (1952), 559-614.
[9J N. Dunford Spectral operators. Pacific J. Math. 4 (1954), 321-354.
[lOJ N. Dunford, J. T. Schwartz Linear operators, Part Ill. (1971), Wiley, New York.
[I1J J. K. Finch The single valued extension property on a Banach space Pacific J. Math. 58
(1975), 61-69.
[12J K. H. Forster Uber die Invarianz eineger Riiume, die zum Operator T - >'A gehoren. Arch.
Math. 17 (1966), 56-64.
[13J M. A. Goldman, S. N. Kraekovskii Invariance of certain subspaces associated with A - >'1.
Soviet Math. Doklady 5, (1964), 102-4.
[14J M. A. Goldman, S. N. Krackovskii Behaviour of the space of zero elements with finite-
dimensional salient on the zero kernel under perturbations of the operator. Soviet Nath.
Doklady 16, (1975), 370-3.
[15J H. Heuser Functional Analysis (1982), Marcel Dekker, New York.
[16J R. H. Homer Regular extensions and the solvability of operator equations. Proc. Amer.
Math. Soc. 12 (1961), 415-18.
[17J T. Kato Perturbation theory for nullity, deficiency and other quantities of linear operators.
J. Anal. Math. 6 (1958), 261-322.
[18J K. B. Laursen, M. M. Neumann Introduction to local spectral theory, Clarendon Press,
Oxford 2000.
[19J M. Mbekhta Sur I 'unicite de la decomposition de Kato generalisee. Acta Sci. Math. (Szeged)
54 (1990), 367-77.
[20J M. Mbekhta Sur la theorie spectrale locale et limite des nilpotents. Proc. Amer. Math. Soc.
110 (1990), 621-631.
[21J M. Mbekhta, A. Ouahab Operateur s-regulier dans un espace de Banach et theorie spectrale.
Acta Sci. Math. (Szeged) 59 (1994), 525-43.
[22J M. Mbekhta, A. Ouahab Perturbation des operateurs s-reguliers . Topics in operator theory,
operator algebras and applications, Timisoara (1994), Rom. Acad. Bucharest, 239-249.
[23J M. Mbekhta Ascent, descent et spectre essential quasi-Fredholm., Rendiconti Circ. Mat.
Palermo (2), 46, (1997), 175-196.
[24J V. Miiller On the regular spectrum., J. Operator Theory 31 (1994), 363-380.
[25J V. Rakocevic Generalized spectrum and commuting compact perturbation. Proc. Edinburgh
Math. Soc. 36 (2), (1993), 197-209.
[26J P. Saphar Contribution a l'etude des applications lineaires dans un espace de Banach. Bull.
Soc. Math. France 92 (1964), 363-84.
[27J C. Schmoeger On isolated points of the spectrum of a bounded operator. Proc. Amer. Math.
Soc. 117 (1993), 715-19.
14 PIETRO AlENA AND FERNANDO VILLAFANE

[28] C. Schmoeger (1995). Semi-Fredholm opemtors and local spectml theory., Demonstratio
Math. 4, 997-1004.
[29] M. 6 Searc6id , T. T. West Continuity of the genemlized kernel and mnge for semi-Fredholm
opemtors. Math. Proc. Camb. Phil. Soc. 105, (1989), 513-522.
[30] P. Vrbova On local spectml properties of opemtors in Banach spaces. Czechoslovak Math.
J. 23(98) (1973a), 483-92.
[31] T. T. West A Riesz-Schauder theorem for semi-Fredholm opemtors. Proc. Roy. Irish. Acad.
87 A, N.2, (1987), 137-146.

DIPARTMENTO DI MATEMATICA ED ApPLICAZIONI, VIALE DELLE SCIENZE, UNIVERSITA DI


PALERMO, 90128 PAI,ERMO, ITALY
E-mail address:paienafDmbox.unipa.it

DEPARTAMENTO DE :tViATEMATICAS, FACULTAD DE CIENCIAS, UNIVERSIDAD UCLA DE BAR-


QUISIMETO (VENEZUELA)
E-mail address:fvillafaluicm.ucla.edu.ve
Contemporary Mathematics
Volume 328, 2003

The Fejer-Riesz Inequality and the Index of the Shift

John R. Akeroyd

ABSTRACT. In this brief article we consider a result that can be characterized


as the converse of the Fejer-Riesz inequality. This result has bearing on theory
concerning the index of the shift.

Let I-" be a finite, positive Borel measure with support in {z : Izl ::; I} (~ := {z :
Izl < I}) and let p2(1-") denote the closure of the polynomials in L2(1-")' We assume
throughout that p2(1-") is irreducible (i.e., it contains no nontrivial characteristic
functions). From this it follows that:
a) I-"lalIJl m (normalized Lebesgue measure on 8~), and

b) for any w in ~, f f--+ f(w) defines a bounded linear functional for poly-
nomials f with respect to the L2(1-") norm, that is bounded independent
of w in any compact subset of~; cf. [15], Theorem 5.8. In other words,
~ = abpe(p2(1-")) - the collection of analytic bounded point evaluations
for P2(1-")'
In the case that 1-"(8~) > 0 and I-"llIJl is radially weighted area measure, there is
much in the literature concerning which weights have the property that p2(1-") is
irreducible; for instance, cf. [9] and [10]. Returning to our general setting, notice
that multiplication by the independent variable z is a bounded operator on p2 (1-").
We call this operator the shift (on p2 (1-")) and denote it by M z, suppressing reference
to 1-". Let Lat(Mz ) denote the collection of closed invariant subspaces for the shift
(on p2(1-"))' If {O} "I- M E Lat(Mz), then, since 0 E abpe(p2(1-")), zM is a closed
subspace of M and in fact dim(M e zM) 2: 1. In the case that 1-"(8~) = 0,
C. Apostol, H. Bercovici, C. Foias and C. Pearcy have shown that for any natural
number n, and for n = 00, there exists Min Lat(Mz ) such that dim(MezM) = n;
cf. [3], and for related work see [7]. This result is an indication of how very large
Lat(Mz ) is in the case that 1-"(8~) = O. In fact, it is large enough to "model"
the general invariant subspace problem for bounded operators on a Hilbert space;
again, cf. [3] and [7]. A classical example that falls under this heading (1-"(8~) = 0)
is the Bergman space L~(~), which equals p2(1-") when I-" = A - area measure on
~. At the other extreme, if I-" = m, then P2(1-") represents the Hardy space H2(~)
and so, by Benrling's Theorem, dim(M e zM) = 1 for all nontrivial members M

1991 Mathematics Subject Classification. Primary 47 A53, 47B20, 47B38; Secondary 30ElO,
46E15.

2003 American Mathematical Society


15
16 JOHN R. AKEROYD

of Lat(Mz ). It has been conjectured that for any measure I-" with mass on the
unit circle (i.e., 1-"(8lI))) > 0), the outcome mimics that of Hardy space case and
dim(M8 zM) = 1 whenever {OJ =I- M E Lat(Mz ); cf. [5]. There are a number of
results in the literature that support this conjecture. The first of these is found in a
paper ofR. Olin and J. Thomson (cf. [12]) who show that it holds whenever I-" that
has a so-called "outer hole" in its support. Subsequently (in [11]), L. Miller shows
that it also holds in the case that I-" = A + mi.,!' where 'Y is some nontrivial sub arc
of 8lI)). In [17], L. Yang extends this result of L. Miller to the case: I-" = A + mlE,
where E is any compact subset of 8lI)) of positive Lebesgue measure that satisfies
the Carleson condition
1
Ln m(In)log(m(In)) < 00;

{In} are the intervals that are complementary to E in 8lI)). And then (in [16])
J. Thomson and L. Yang obtain this extension of L. Yang in the more general
context of the shift on pt(I-"), for 1 < t < 00. The conjecture has recently been
established for any measure I-" for which there is a nontrivial subarc 'Y of 8lI)) such
that

1 10g ( 1;; )dm > -00,

with no special assumption made concerning 1-"10; cf. [2]. In [2], the author makes
use ofresults in an earlier paper (cf. [1]) that are intricately related to the seminal
work of R. Olin and J. Thomson in [12]. Specifically, in [1] the author defines what
it means for I-" to be strongly inscribed and shows that if I-" is such, then indeed
dim(M 8 zM) = 1 for each nontrivial member M of Lat(Mz ). To be explicit, I-" is
said to be strongly inscribed if there is a Jordan subregion W of lI)) with rectifiable
boundary (we let Ww denote harmonic measure on 8W for evaluation at some point
in W) with the properties:
i) ww(8lI))) > 0, and

ii) there is a nonnegative function h in LOO(ww) such that log(h) E L1 (ww)


and

law Ifl 2 w: ; f Ifl2dl-"


hdw

for all polynomials f.


Since 8W is rectifiable, (i) is equivalent to: m((8W)n(8lI)))) > O. And this definition
is not truly altered if we drop the requirement that W has rectifiable boundary,
because if W were any simply connected subregion of lI)) that satisfies (i) and (ii),
then we could find a Jordan subregion V of W, where V has rectifiable boundary
and V itself satisfies (i) and (ii); cf. [14], Proposition 6.23. It is still an open
question as to whether or not our general assumptions concerning 1-", along with
the hypothesis that 1-"(8lI))) > 0, together imply that I-" is strongly inscribed. In this
brief article we discuss what amounts to the converse of the Fejer-Riesz inequality
and find that this converse has close ties to the definition of "strongly inscribed" .
The Fejer-Riesz inequality, whose statement follows, falls under the general
heading of results concerning Carleson measures. For a proof, see [6],
Theorem 3.13.
THE FEJER-RIESZ INEQUALITY AND THE INDEX OF THE SHIFT 17

THEOREM 1 (Fejer-Riesz Inequality). If f E HP(][))) (0 < p < 00), then


ill If(tei'l')lPdt ::; 4121r If(eill)IPdO

for 0 ::; 'P < 27f. The constant ! is best possible.

With z fixed in ][)), ( f-+ Pz (() := 1~=~J~2 is the Poisson kernel on 8][)) for
evaluation at z. It is well-known that falf Pz(()dm(() = 1 independent of z, and
that if hE L1(m), then (by Fatou's Theorem)

( Pz(()h(()dm(() ~ h(~)
lalf
as z nontangentially approaches ~ for m-a.a. ~ in 8][)). One may consult [6] and
[8] as good references for these results. We begin with a rather straightforward
observation whose proof appears in [4]; see the proof of Lemma 3.1 in this reference.
LEMMA 2. Let rJ be a finite, positive Borel measure with support in ll} such that
rJ(8][))) = O. Then
1 - r2
lim 1m drJ(w) = 0
r--+1- If 11-r~wI2
for m-a.a. ~ in 8][)).
Our next result can be viewed as the converse of the Fejer-Riesz inequality.
THEOREM 3. Let v be a finite, positive Borel measure with support in ll} such
that vlalf m and][)) = abpe(p2(v)). For c > 0, let Bv(c) be the set of all ~ in 8][))
such that
11 If(t~Wdt J ::; c Ifl2dv

for all polynomials f. Then Bv (c) is a closed subset of 8][)) and :~ ;::: ~ (a. e. m) on
Bv(c). Furthermore, if E is a Lebesgue measurable subset of Bv(c) and m(E) > 0,
then XE p 2(v).

Proof. That Bv(c) is closed is an immediate consequence of the fact that any
polynomial is uniformly continuous on ll}. Now if 0 < r < 1 and I~I = 1, then
v'f=T2
g(w) := ----=-
1-r~w

is analytic in a region containing ll} and so, by Runge's Theorem, is the uniform
limit (on ll}) of polynomials. Therefore we can apply our hypothesis to get that

11 Ig(t~)12dt J ::; c Igl 2dv,

which yields:

l+r= 1o
1 1 - r2
(1
- rt
)2 dt ::;c.
JI 1 - r2
1- r~wl
2 dv (W),

for 0 < r < 1 and any ~ in Bp.(c). Letting r --+ 1 and applying Lemma 2, we
find that :~ ;::: ~ (a.e. m) on Bv(c). To finish the proof of this theorem, let
E be a Lebesgue measurable subset of Bv(c) such that m(E) > 0, and suppose
18 JOHN R. AKEROYD

that XE E P2(v); we look for a contradiction. Now h := (1 - XE) E p 2(v) (since


XE E p2(v)), and in fact, hg E P2(v) whenever 0 < r < 1 and ~ E E. Arguing as
before, with hg now in the place of g, we obtain: ~ ~ h :::. = 0 a.e. rn on E -
clearly a contradiction.D

It turns out that if f.L is strongly inscribed, then in fact there exists c > 0 such
that rn(BJ.L(c)) > o. En route to this result (Theorem 5, below), we make the
following observation.
PROPOSITION 4. Let f.L be a finite, positive Borel measure with support in iID
such that p2 (f.L) is irreducible. Then the following are equivalent.
1) f.L is strongly inscribed.
2) There is a Jordan subregion V ofH}, where 8V is rectifiable and rn((8V) n
(8H})) > 0, and there is a positive constant M, such that

lav Ifl 2dw v ~ M f Ifl2df.L

for all polynomials f.


Proof. We first assume (1), and so by definition there is a Jordan subregion W
of H} and a nonnegative function h in LOO(ww) that satisfy certain requirements.
One of these requirements, namely that log( h) EL I (ww ), guarantees the existence
of a bounded analytic function g in W such that g 0 'P is an outer function ('P is
a conformal mapping from H) onto W) and Igl has "boundary values" equal to h
(a.e. ww). Applying Proposition 2.2 of [1], we can find a Jordan subregion V of
W, where V has rectifiable boundary, wv(8H}) > 0 and Igl ~ e > 0 on V. So by
the subharmonicity of Ifl 21g1 in W, (2) holds, with M := ~. That (2) implies (1)
is immediate, and our proof is complete.D
THEOREM 5. Let f.L be a finite, positive Borel measure with support in iID such
that p2 (f.L) is irreducible. If f.L is strongly inscribed, then there exists c > 0 such
that rn(BJ.L(c)) > o.
Proof. Assuming that f.L is strongly inscribed, Proposition 4 provides a Jordan
subregion V of H} with the properties listed in (2). Let 'P be a conformal mapping
from H} one-to-one and onto V, and let 1/J = 'P -1. Since Wv (8H}) > 0, we can find
(cf. [14], Theorem 6.8 and Theorem 3.7) a closed subset E of (8V) n (8H}), where
each point in E is a point of tangency of 8V with 8H}, such that:
i) rn(E) > 0,
ii) for any ~ in E and any Stolz angle ~ whose closure is contained in VU{O,
there is a constant M > 1 such that

~ ~ 11/J'(z)l, 11/J(Z; =t~) I ~ M


for all Z in ~, and
iii) if ~ E E and 'Y is a smooth arc in V U {O having nontangential approach
in V to ~, then 1/J("() is smooth and has nontangential approach in H} to
1/J(~).
Now choose ~ in E. Since ~ is a point of tangency of 8V with 8H}, there exists s,
o < s < 1, such that t~ E V whenever s ~ t < 1. Let'Y be the smooth curve in
THE FEJER-RIESZ INEQUALITY AND THE INDEX OF THE SHIFT 19

ID>U{'l/J(~)} defined by ')'(t) = 'l/J(t~), s::; t::; 1. Then, by (i) - (iii), there are positive
constants Ck (k = 1,2,3) independent of to in [s, 1) such that

lengthb([to, l]))
=
It: W(t~)ldt
1 - b(to)1 1 -1'l/J(to~)1
I~ - to~1
< C1 . 1 - I'l/J(to~) I
I~ - to~1
<
C2 I'l/J(O - 'l/J(to~)1
< C3
From this it follows that arclength measure on ')'([s, 1)) is a Carleson measure for
H2(1D. And so there are positive constants Ck (k = 3,4) such that, for any poly-
nomial f,

11 If(t~)12dt = llU 0 cp)(wWlcp'(w)ldlwl

< C3 llU 0 cp)(wWdlwl

< C4 [ IU 0 cpWdm
laI)
C4 [ IfI 2ru.vv;
lav
by Harnack's inequality, we may assume that cp(O) is the point in V of evaluation
for wv. Once again recalling (2) (of Proposition 4), we can now find a positive
constant C5 such that

11 If(t~Wdt ::; C5 J Ifl2djj

for all polynomials f. Since ID> = abpe(p2(jj)), we may apply Lemma 2.6 of [13]
and find another positive constant C6 such that, for all polynomials f,

1If(t~Wdt J
s
::; C6 Ifl 2djj.

Consequently, ~ E B/-L(c) , for C := C5 + C6. Thus we have shown that E ~


U~=1 B/-L(n). Since m(E) > 0, we can assert that m(B/-L(n)) > 0 for some inte-
ger n, which completes the proof.D
QUESTION 6. Does the converse of Theorem 5 hold? That is, if jj is a finite,
positive Borel measure with support in iij such that p 2 (jj) is irreducible, and if
m(B/-L(c)) > 0 for some positive constant c, then is jj strongly inscribed? Indeed,
can we even assert that dim(M 8 zM) = 1 for each nontrivial, closed invariant
subspace M for the shift on p2(jj)?

We conclude this article with a rather anemic response to Question 6 that


supports an affirmative answer.
20 JOHN R. AKEROYD

REMARK 7. There are other more general forms of the Fejer-Riesz inequality,
where the integral on the left is taken over chords of the unit circle and not just over
diameters. A converse to this Fejer-Riesz inequality (for chords) can be established,
and involves integrals over segments that have nontangential approach in lI} to
certain points in alI}. Thus, analogues of B/-I(c) can be defined, where the integral
on the left is taken over segments in various Stolz angles. All of this leads to
a counterpart of Theorem 5, whose converse appears to be manageable. To this
author it seems most likely that if m(B/-I(c)) > 0, then there is a sizeable subset E
of B/-I(c) that is contained in these collections that are analogous to B/-I(c), and thus
the converse of Theorem 5 is likely a consequence of its counterpart in the context
of the Fejer-Riesz inequality for chords.

References
[1] J. Akeroyd, Another look at some index theorems for the shift, Indiana Univ. Math. J., 50
(2001),705-718.
[2] J. Akeroyd, A note concerning the index of the shift, Proc. Amer. Math. Soc., Vol. 130, No.
11 (2002), 3349-3354.
[3] C. Apostol, H. Bercovici, C. Foias, C. Pearcy, Invariant subspaces, dilation theory, and the
structure of the predual of a dual algebra, I, J. Functional Analysis, 63 (1985),369-404.
[4] A. Aleman, S. Richter, C. Sundberg, The majorization function and the index of invariant
subspaces in the Bergman spaces, J. Analyse Math., 86 (2002), 139-182.
[5] J. B. Conway, L. Yang, Some open problems in the theory of subnormal operators, Holomor-
phic spaces, Cambridge University Press, 33 (1998), 201-209.
[6] P. L. Duren, Theory of HP Spaces, Academic Press, New York, 1970.
[7] H. Hedenmalm, S. Richter, K. Seip, Interpolating sequences and invariant subspaces of given
index in the Bergman spaces, J. Reine Angew. Math., 477 (1996), 13-30.
[8] K. Hoffman, Banach Spaces of Analytic Functions, Prentice-Hall, Englewood Cliffs, N.J.,
1962.
[9] S. Hruscev, The problem of simultaneous approximation and removal of singularities of
Cauchy -type integrals, Trudy Mat. Inst. Steklov 130 (1978), 124-195; English transl., Proc.
Steklov Inst. Math. 130 (1979), no. 4, 133-203.
[10] T. L. Kriete, B. D. MacCluer, Mean-square approximation by polynomials on the unit disk,
Trans. Amer. Math. Soc., vol. 322, no. 1 (1990), 1-34.
[11] T. L. Miller, Some subnormal operators not in A2, J. Functional Analysis, 82 (1989), 296-302.
[12] R. F. Olin, J. E. Thomson, Some index theorems for subnormal operators, J. Operator Theory,
3 (1980), 115-142.
[13] R. F. Olin, L. Yang, A subnormal operator and its dual, Canad. J. Math., 48 (1996), 381-396.
[14] Ch. Pommerenke, Boundary Behaviour of Conformal Maps, Springer-Verlag, Berlin-
Heidelberg, 1992.
[15] J. E. Thomson, Approximation in the mean by polynomials, Ann. Math., 133 (1991), 477-507.
[16] J. E. Thomson, L. Yang, Invariant subspaces with the codimension one property in Lt(J.I),
Indiana Univ. Math. J., vol. 44, no. 4 (1995),1163-1173.
[17] L. Yang, Invariant subspaces of the Bergman space and some subnormal operators in Al\A2,
Mich. Math. J., 42 (1995), 301-310.

DEPARTMENT OF MATHEMATICS, UNIVERSITY OF ARKANSAS, FAYETTEVILLE, ARKANSAS 72701


E-mail address: jakeroydlDcomp.uark.edu
Contemporary Mathematics
Volume 328, 2003

A Cauchy-Green Formula on the Unit Sphere in C 2

John T. Anderson and John Wermer

ABSTRACT. In 1977 G. Henkin introduced an integral formula for solving


8bf = tLwhere tL is a measure, on the boundary of a smooth strictly con-
vex domain. This result is closely related to a "Cauchy-Green" formula on
the sphere (see Chen and Shaw [3]). We give a direct elementary proof of the
Cauchy-Green Theorem on the unit sphere and derive Henkin's solution of the
8b equation from this. We also give an application to an approximation result.

1. Introduction
Let n be a domain in the plane, with smooth boundary r. The classical
Cauchy-Green formula states that for any c/> E C 1 (0') and zEn,

(1.1 ) c/>( z) = _1
27l'i
r c/>( () d( _ _ lnr o~ d(
lr ( - Z
1
27l'i o( (-
1\ d(
Z

Note that the first term on the right of (1.1) is a holomorphic function 4> of Z in
the domain n. In fact, 4> extends continuously to 0', and hence defines an element
of the algebra A(O') consisting of functions holomorphic in n and continuous on 0'.
Of course, if c/> E A(O'), (1.1) reduces to the Cauchy integral formula and 4> = c/>.
The representation (1.1) has many applications in complex analysis. In the
theory of approximation of continuous functions on a compact set K c C by
rational functions with poles off K, one is led by considerations of duality to examine
measures supported on K. The Cauchy transform of such a measure J.l is defined
by

(1.2) p,(Z) = r
dJ.l(()
lK (-z
The integral defining [J, converges absolutely for almost all z E C. Using (1.1), one
can easily show that for any smooth compactly supported function C/>,

(1.3)
lK
rc/>(z) dJ.l(z) = ~ r
o~ p,(z) di 1\ dz
27l'zlc z
That is, [J, satisfies the equation
0[J,
(1.4) oi =-7l'J.l

1991 Mathematics Subject Classification. Primary 32A25, Secondary 32E30.

2003 American Mathematical Society


21
22 JOHN T. ANDERSON AND JOHN WERMER

in the sense of distributions, and hence defines a holomorphic function on e \ K.


The Cauchy transform is a key tool in rational approximation theory in the plane.
We have been motivated by problems of rational approximation for subsets of
the boundary S of the unit ball in e 2 . It is posible to do a kind of function theory
on S analogous to the theory of analytic functions in the plane. The operator a / az
is replaced by the tangential Cauchy-Riemann operator

(1.5) x = Z2-
a - Zl-.
a
aZl aZ2
x is well-defined 011 C l (S) and for any relatively open subset n of S, annihilates
the restrictions to n of functions holomorphic in a neighborhood of n in e 2 The
solutions to X = 0 on n are known as CR functions on n. A good general reference
for the theory of CR functions is the book [2].
One would like an analogue of the Cauchy transform for measures on S. Given
a measure JL on S, G. Henkin in 1977 [4] constructed a function KJ1.' summable
with respect to three-dimensional Hausdorff measure da on S, satisfying
- 2
(1.6) abKJ1. = -27r JL
in the sense of distributions, i.e.,

(1.7) [ (z) dJL(z) = ~ [ KJ1. X da(z)


is 27r is
for all smooth , provided that JL satisfies the necessary condition that Is
P dJL = 0
for all polynomials P. Note that (1.7) implies that KJ1. is a CR function (in the
sense of distributions) off the support of JL.
In attempting to use and understand Henkin's construction in the study of
rational approximation on subsets of S, we were led to the analogue of the Cauchy-
Green formula (1.1) that we present below. It plays the same role with respect
to Henkin's formula (1.6) as the classical Cauchy-Green formula on the plane does
to equation (1.4). The resulting formula, which is contained in our Theorems 2.1
and 3.1 below, is not new. It is given in a more general setting in Chen and Shaw
([3], see the remarks following Corollary 11.3.5) as a consequence of the theory of
Henkin for solving the 8b equation on the boundary of a strictly convex domain in
en. Our approach to establishing this Cauchy-Green formula on the sphere in e 2
is direct and elementary, and leads immediately to the property (1.6) of Henkin's
transform K J1.'
Let A(B) denote the algebra of functions holomorphic in the open unit ball
B of e 2 and continuous on its closure. We seek a kernel H((, z), defined for
((, z) E S x S, such that for all E Cl(S), there exists <I> E A(B) with

(1.8) (z) = <I>(z) + c Is H((, z) 8(() 1\ w(()

for all z E S, where w(() = d(l 1\ d(2, 8 = (a/azddz 1 + (a/az2)dz2, and c is a


universal constant. We call (1.8) a "Cauchy-Green formula for S". We will demand
that H have the following properties:
A CAUCHY-GREEN FORMULA ON THE UNIT SPHERE IN C 2 23

a: H((, z) is continuous on S x S \ {z = 0;
b: For all unitary transformations U of determinant 1, H (U (, U z) = H ((, z);
c: Is
jH((, el)j da(() < 00, where el = (1,0), and da is three-dimensional
Hausdorff measure 1 on S.
Properties (b) and (c) together with the unitary invariance of da imply that H
is uniformly summable with respect to da, Le., there exists a constant C so that

(1.9) is jH((, z)j da(() ::; C, Vz ES

They also imply that the integral

(1.10) K(z) == is H((, z) a(() /\ w(()

appearing in (1.8) is finite for all z E S, since a /\ w is absolutely continuous with


respect to da. A routine calculation gives
(1.11) a /\ w = 2(X) da
on S, where X is the operator in (1.5), for smooth . We can say more about K:
LEMMA 1.1. If H satisfies properties (a), (b) and (c), then K is continuous on
S.
PROOF. Fix z E S. For f > 0, put Sz) = S \ {jz - (j ::; f} and S~ =
S n {jz - (j ::; fl. Let

Kz) = r
ls.(z)
H((, z) a(() /\ w(()

Then K< is continuous on S, by property (a) of H. For all z E S, by (1.11),

jK(z) - K,(z)j = I r
ls~(z)
H((, z) a(() /\ W(()I ::; M r
ls~(z)
jH((, z)jda(()

where M is a constant independent of z and f. Let el = (1,0) and choose a


unitary transformation U of C 2 with Uel = Z; then U(S~(el)) = S~(z). Then using
property (b),

r
ls~(z)
jH((, z)j da(() = r
lS~(e,)
jH(Ury, Uel)j da(Ury) = r
lS:(e')
jH(ry, el)j da(ry)

Since Is jH(ry, edjda(ry) is finite by assumption (c),


lim
<-+0 1rS~(e,) jH(ry, el)jda(ry) = 0
It follows that K, -t K uniformly on S, and so K is continuous, as claimed. 0
We say that a measure /-L on S is orthogonal to polynomials if

(1.12) is Pd/-L = 0, V holomorphic polynomials P

Given any measure /-L on S, define

(1.13) K Il (() = is H((, z) d/-L(z), (E S

Ida is not normalized; oo(S) = 21T2.


24 JOHN T. ANDERSON AND JOHN WERMER

LEMMA 1.2. A kernel H((, z) satisfying (a), (b) and (c) satisfies (1.8) if and
only if for each measure J.L on 8 orthogonal to polynomials
(1.14) Is dJ.L = cis K,.. 8 A w
for all E C1(8).
PROOF. Suppose first that H((, z) satisfies (a), (b), (c) and (1.8). Let J.L be a
measure on 8 orthogonal to polynomials. Fix E C 1 (8). and let <fl E A(B) be as
in (1.8). Since polynomials are dense in A(B), Is
<fl dJ.L = O. Hence by (1.8),

Is (z)dJ.L(z) = Is (c Is H((,z) 8(() AW(()) dJ.L(z)

Is (c Is H((, Z)dJ.L(Z)) 8(() A w(()

cIs K,..(() 8(() A w(()

so that (1.14) holds. The application of Fubini's theorem is justified by (1.9).


Next, suppose that (1.14) holds, for H satisfying (a), (b) and (c). Choose a
measure J.L on 8 orthogonal to polynomials. Fix a function E C 1 (8), and define

<fl(z) = (z) - c Is H((, z) 8(() 1\ w(()

By Lemma 1.1, <fl is continuous on 8, and

Is <fl(z)dJ.L(z) Is (Z)dJL(Z) - c Is (Is H((, Z)dJ.L(Z)) 8 A w(()

= Is (z)dJ.L(z) - cis K,..(() 8(() A w(()


o
by (1.14). Since this holds for all J.L orthogonal to polynomials, <fl E A(B), and so
(1.8) follows. 0
In 1977, in [4] G. Henkin introduced the kernel
- -
( 1.15) H(( z) = (1 22 - (2 21 (, Z E8
,
11-<z,(>1 2 '
where <, > denotes the Hermitian inner product < z, ( >= Zl(l + Z2(2, and proved
the formula (1.14) using this kernel. It is easy to check that H satisfies properties
(a), (b) and (c) above. Formula (1.14) on 8 is actually very special case of a class
of general integral formulae on smooth convex domains established in [4]. In her
thesis [5], H.P. Lee gave an elementary proof of Henkin's formula for 8j the paper
[8] of Varopoulous also contains an exposition of Henkin's results on the sphere.
For applications of Henkin's formula to rational approximation, see the paper [6]
of Lee and Wermer.
In this paper, we shall
(1) give a direct proof of (1.8), using Henkin's kernel (1.15)j
(2) give a formula for <fl, in terms of j
(3) deduce an approximation result (Theorem 4.1) from (1.8).
A CAUCHY-GREEN FORMULA ON THE UNIT SPHERE IN C 2 25

1.1. Acknowledgment. The first author wishes to thank Joseph Cima for
helpful conversations on the results in section 3.

2. A Cauchy-Green Formula using Henkin's Kernel


With H as in (1.15) and E C 1 (S) as in section 1 put

K(z) = Is H((, z) 8(() 1\ w(()

For a E int(6), put r = Jl -lal 2 and denote by "fa the circle Z2 = rT, ITI = 1 in
the z2-plane.

LEMMA 2.1. Fix a E 6. For n = 0,1,2, ... we have, putting z = (a, Z2),

(2.1)

PROOF.

We denote the inner integral by J((). Multiplying both numerator and denom-
inator of the integrand by T, we get

Let

r(2
and T2 = 1 -a1
-(

Note that T1 T2 = 1. We have

Ir(21 2 -11 - a(11 2 = (1-laI 2)(1-1(112) -11 - a(11 2


1 - lal 2 - 1(11 2 + la1 211(112 - 1 - la1 21(112 + a(l + a(l
= -(laI 2 + 1(11 2 - a(l - a(t)
-Ia - (11 2
26 JOHN T. ANDERSON AND JOHN WERMER

Thus

(2.2)

Let Se be the part of S lying over the region


{I(I - al ~ f} n {I(II ::; I}
in the (I-plane. Let TE denote the boundary of SE' We claim that

(2.3) la K(Z)Z2 dZ 2 = - !~~ [l,4>(()I(()W(()]


To establish the claim, note that

1 'Ya
K(Z)Z2 dZ 2

lim [ 84> /\ w . I
<-+0 Js,
= lim [ d(4) wI)
e-+O Js,
since I is holomorphic on Se for f > O. By Stokes' Theorem, the latter integral
equals
- [ 4> wI
JT,
proving the claim.
Note that T, is the torus
(1 = a + fe ifJ , (2 = JI -1(II 2ei .p, 0::; 0, 'IjJ ::; 27l'.
A CAUCHY-GREEN FORMULA ON THE UNIT SPHERE IN C 2 27

On T< we have the following relations:


(() = (a, rei1/J) + O(f);

d(1 = ife i9 dB, d(~ = -ife- i9 dB;


d(2 = -(1 d(;. - (1 d(1 ei1/J + iJI -1(11 2ei 1/Jd'lj; = i re i1/Jd'lj; + O(f);
2JI-I(11 2
1 1
(1 - a fe i9
Using this information together with (2.2) and (2.3) we obtain

1K(z)z~
"fa
= lim [-27ri [ (()r2n+2 (
<--->0 Jr, 1- a(~1 )
n
+1 (( 1
1 - a) d(11\. d(2]
For fixed f, we rewrite the expression in brackets as
-27ri [ (a, rei1/J)rnein1/JidB I\. irei'IjJd'lj; + O(f)
Jr.

1K(z)z~
"fa
dZ2 =

This completes the proof of (2.1) and Lemma 2.1. o


Next, we define an operator T on C 1 (8) as follows:
(2.4) (T)(z) = 47r 2 (z) - K(z), for z E 8, E C 1(8)
Letting X denote the tangential Cauchy-Riemann operator on 8 as in section 1,
using (1.11) we can write

T = 47r 2 - is H((, z) (X)(()dcr(()

LEMMA 2.2. Fix E C 1 (8). Let L be a complex line in C2. Then the restriction
ofT() to L n 8 extends analytically to L n B.
PROOF. Lemma 2.1 gives us, for each a E int(h.), that

(2.5) 1(T)(z)z~
"fa
dZ 2 = 0, n = 0,1,2, ...

Note that "fa = La n 8, where La is the line {Z1 = a}. Then (2.5) implies
that T extends analytically to the disk La n B. Using the unitary invariance of
H,cr, and X, it is not hard to check that for all E C 1 (8),
(2.6) (T)oU=T(oU)
Fix a complex line L. Let N denote the complex line passing through the origin
which is orthogonal to L, and let zO denote the intersection point N n L. Write
L = {zO + (t I t E C} for some unit vector (. If U is a unitary transformation with
Ue2 = (, where e2 = (0,1) then U maps the line {Z2 = O} to N, and maps some
28 JOHN T. ANDERSON AND JOHN WERMER

point (a,O) to zOo Then U((a,O) + t(O, 1)) = zO + t(, for all t E C. So U maps
the line La to L and maps the disk La n B to L n B. By (2.6), T I Lns extends
analytically to the disk L n B if and only if (T) 0 U 1La ns extends to La n B. This
last is true by (2.5), as we have noted earlier, and so the proof is complete. 0
By Lemma 1.1, since H satisfies properties (a), (b) and (c) of section 1, K
and thus T are continuous on S. By Lemma 2.2, T has the "one-dimensional
extension property" as defined by Stout in [7], p. 105. A theorem of Agranovskii
and Val'skii [1] then gives that T lies in the ball algebra A(B). Putting cP = T(),
we have arrived at
THEOREM 2.3. Let E C 1 (S). Then there exists cP E A(B) such that

47r 2 (z) = cP(z) + is H((, z) 8(() /\ w(()

where H is Henkin's kernel

3. The Cauchy-Green formula and the Cauchy transform


In this section we identify the ball algebra function cP appearing in Theorem
2.3 as a certain principal value of the Cauchy transform of . The Cauchy kernel
for B is
1
C(z,() = (1- < z,( 2
For z E S we set
N. (z) = {( E S : I < (, z > I > 1 - f}
and we denote the boundary of N. (z) by r f (z).
THEOREM 3.1. Fix E C 1 (S). If cP is as in Theorem 2.3, then for z E S,

cP(z) = 2 lim f (()C(z, () du(()


.-+0 JS\N.(z)

REMARK 3.2. Since C(z,) rt. Ll(du), it is not immediate that the limit in
Theorem 3.1 exists.
PROOF. As in sections 1 and 2, set

K(z) = f H((, z) 8(() /\ w(() = lim f H((, z) 8(() /\ w(()


JS f-+O JS\N, (z)
For f > 0 fixed,
f H((, z) 8(() /\ w(() = f d[H((, z)(() /\ w(()]
JS\N,(z) JS\N.(z)

- f [8H((, z)]/\ (() /\ w(()


JS\N.(z)

= f H((, z) 8(() /\ w(()


Jr,(z)

-2 f (XH)((,z) (() du(()


JS\N.(z)
A CAUCHY-GREEN FORMULA ON THE UNIT SPHERE IN C 2 29

by Stokes' theorem, if r e(z) is oriented as the boundary of S \ Ne (z). We have also


used equation (1.11) from section 1. A computation shows (differentiation is in the
( variable)
(XH)((, z) = -C(z, ()
so that

K(z) = lim [
-+O
r
Jr,(z)
H((, z) </>(() 1\ w(() - 2 r
JS\N,(z)
C((, z) </>(()dO"(()]

Since
cI>(z) = 41r 2</>(z) - K(z)
by Theorem 2.3, the proof will be complete if we can show that

(3.1) lim
-+O
r
Jr,(z)
H((, z) </>(() 1\ w(() = 41r2</>(z)

To establish (3.1), choose a unitary map U with Ue1 = z. Then for fixed E > 0,
r
Jr,(z)
H((, z) </>(() 1\ w(() = r
Jr,(et}
H(TJ, ed (</> 0 U)(TJ) 1\ w(TJ)

The torus r e(e1) = {TJ : ITJ11 = 1 - E}, oriented as the boundary of S \ Ne(ed, is
parametrized by

where

Then on re(ed,

and

which gives

r
Jr,(z)
H((, z) </>(() 1\ w(()

(3.2)

where
r r
IIel ::; C Jo Jo
27r 27r r2
11 - (1 -.: )ei9112d(hd{;l2
for some C > o. An application of the Poisson integral formula shows that the first
integral in (3.2) converges to 41r2(</>oU)(e1) = 41r2</>(z) as E ---+ 0, while lime-+o Ie = O.
This completes the proof. D
30 JOHN T. ANDERSON AND JOHN WERMER

4. An Approximation Theorem
Fix E C 1 (S). The quantity
dist(, A(B)) = inf{ll - gil : g E A(B)}
where I . II is the uniform norm on S measures how closely can be approximated
by polynomials on S.
THEOREM 4.1. There exists C > 0 so that/or all E C1(S),
dist(, A(B)) ~ CIIXII
PROOF. Let IIHl11 denote the L1 - drr norm of Henkin's kernel H(, z) (which
is independent of z E S). By the representation in Theorem 2.3, there exists
<i> E A(B) so that for z E S,

14rr2(z) - <i>(z) I lis H(, z) 8() Aw(OI

211s H(, Z)(X)(Odrr(ol


< 211 H lh11X11
from which the result follows. o
References
[1] M.L. Agranovskii and R.E. Val'skii, Maximality of Invariant Algebras of Functions, Siberian
Math. J. 33 (1983),p. 227-250.
[2] A. Boggess, CR Manifolds and the Tangential Cauchy-Riemann Complex, CRC Press, 1991.
[3] S.-C. Chen and M.-C. Shaw, Partial Differential Equations in Several Complex Variables,
American Mathematical Society, 2001
[4] G. M. Henkin, The Lewy Equation and Analysis on Pseudoconvex Manifolds, Russian Math.
Surveys, 32:3 (1977); Uspehi Mat. Nauk 32:3 (1977),p. 57-118
[5] H. P. Lee, Orthogonal Measures for Subsets of the Boundary of the Ball in C 2 , Thesis, Brown
University, 1979.
[6] H. P. Lee and J. Wermer, Orthogonal Measures for Subsets of the Boundary of the Ball in
C2, in Recent Developments in Several Complex Variables, Princeton University Press, 1981,
pp. 277-289.
[7] E.L. Stout, The Boundary Values of Holomorphic Functions of Several Complex Variables,
Duke Math. J. 44, 1977,p. 105-108.
[8] N. Th. Varopoulos, BMO functions and the a-equation, Pac. J. Math. 71, no. 1 (1977). pp. 221-
273.

DEPARTMENT OF MATHEMATICS AND COMPUTER SCIENCE, COLLEGE OF THE HOLY CROSS,


WORCESTER, MA 01610-2395
E-mail address: andersonOradius.holycross .edu

DEPARTMENT OF MATHEMATICS, BROWN UNIVERSITY, PROVIDENCE, RI 02912


E-mail address: wermerlDmath. brown. edu
Contemporary Mathematics
Volume 328, 2003

On a-dual Algebras

Hugo Arizmendi, Angel Carrillo, and Lourdes Palacios

ABSTRACT. Let Ti(D) and TieD) be the algebras which consist of all holo-
morphic functions in the open unit disc D and in the closed unit disc 75,
respectively. These algebras, considered as algebras of sequences, are denoted
by A and B. Let A" and B" be the a-dual spaces of A and B, respectively;
we have that A" = Band B" = A, and these sequence spaces with their nor-
mal topologies are topological algebras. A similar treatment can be applied to
the algebra c of all entire functions and its a-dual space c" consisting of all
complex functions that are analytic in some neighborhood of the origin.
Here we examine a more general situation. If A(ap,n) is a matrix algebra,
we establish conditions under which the a-dual space A" (ap,n) is a topological
algebra relative to the normal topology. We also analyse some other important
examples of topological algebras with such properties.

1. Introduction
A sequence space), is a vector space of complex sequences x = (Xi)~O' The
vector space operations are the usual operations on the coordinates. ), can be
considered as a linear subspace of the space w of all complex sequences.
To each sequence space ), we assign another sequence space N", its a-dual.
),01. is the set of all complex sequences Y = (Yi)~O for which the scalar product
<Xl

yx = I: XiYi converges absolutely, for each x E ),.


i=O
The normal topology on ), is the topology determined by all the seminorms
defined by
<Xl

Ilxlly = 2: IXkYkl
k=O
where Y runs over all ),01..
An important class of sequence spaces are the echelon and co-echelon spaces
which have been studied by G. Kothe and O. Toeplitz. They are defined as follows:
Let (ap,n), p = 1,2, ... , n = 0,1, ... be an infinite matrix of non-negative
numbers such that
(1) 0:::; ap,n :::; ap+1,n
(2) For every n, there exists p such that ap,n > O.

2000 Mathematics Subject Classification. Primary 46; Secondary 30.


Key words and phmses. Topological Algebras, Normal Topology, Matrix Algebra.

2003 American Mathematical Society


31
32 HUGO ARIZMENDI, ANGEL CARRILLO, AND LOURDES PALACIOS

Let A = {(Xn) E W : E JXnJ ap,n < 00, p = 1,2, ... }. A is called an echelon space
n
and N" a co-echelon space. We note that A = fu!!ll(ap,n)'
p
In the following we denote the echelon space A by A(ap,n)' In [1] it is proved
that if A(ap,n) is an algebra under the convolution product, then this product is
jointly continuous and A(ap,n) is a topological algebra. As a matter of fact, it is a
metrizable locally convex complete algebra, i.e. aBo-algebra.
It can be seen that a necessary and sufficient condition under which A(ap,n) is
an algebra is the following:
(3) For each pEN, there exists q E N such that ap,n+m S aq,n aq,m;
for n, m = 0, 1,2, ....
If (ap,n) satisfies (1), (2) and (3), we call A(ap,n) a matrix algebra.
Let
00

A = {(an)~=o: if JzJ < 1, then L anz n converges in C}


n=O
and
00

B = {(bn)~=o: there exists a z E C, JzJ > 1, and L bnzn converges in C}.


n=O
These sequence spaces A and B are called analytic sequence spaces and they
are algebras under the usual linear operations and the convolution product. The
transformation
00

(Xk)k=O ~ L Xk zk
k=O
identifies those sequence algebras A and B with the function algebras H(D) and
H(D) consisting of all holomorphic functions in the open unit complex disc D and
in the closed unit complex disc D, respectively. Through this identification we shall
00
indistinctly write x = (Xk)k=O or x = E
Xkzk to refer to an element x of A or B.
k=O
By the same fact we can consider in A the compact-open topology that is originally
defined for H(D).
We note that A ~ A(r;), n = 0,1, ... , p = 0,1, ... , where (rp) is an increasing
sequence of positive numbers converging to 1, and the compact-open topology on
A can be given by the sequence of seminorms
00 00

JJxJJp = L JXnJr;, where x = LXnZnEA.


n=O n=O
In [5], O. Toeplitz studied the topological properties of the analytic sequences spaces
A and B and proved that AO: = Band BO: = A.
A similar treatment can be applied to the algebra c of all entire functions and
its a-dual space cO: consisting of all complex functions that are analytic in some
neighborhood of the origin.
In [2] it is proved that A and B are topological algebras when they are endowed
with the normal topology given by A and B, respectively. The same happens with
the algebra of all entire functions c and its a-dual space cO:.
ON ",-DUAL ALGEBRAS 33

If ap n > 0, n
,
= 0,1,2, ... , p = 1,2, .... then A"'(ap n) = limlOO(-1-)
, ap,n
as sets.
---t
p

The inductive limit topology is stronger than the normal topology in A"'(ap,n)'
Prom now on we are going to assume: ap,n > 0, n = 0,1,2, ... ,p = 1,2, ....
If A(ap,n) is nuclear, then A"'(ap,n) ~:::'n~lOO(a:.J ~ l~l1(a:.J as topological
p p

vector spaces.
Here we prove that l~l1(a:.J is an algebra if A(ap,n) satisfies:
p

(*) for each p there exist q > p and Mp such that ap,n ap,m :::; Mpaq,n+m (or
equivalentely _ 1 _ :::; Mp_1___1_) for all n,m.
aq,n+m a p . n ap,rn

And then the convolution product is jointly continuous. We also prove that if
(ap,n) does not satisfy (*), then l~l1(a:.J is not an algebra. Therefore l~l1(a:.J
p p

is an algebra if, and only if, it is a topological algebra. Thus, if A(ap,n) is nuclear,
then A"'(ap,n) is an algebra if, and only if, it satisfies (*). Therefore A(ap,n) and
A"'(ap,n) are topological algebras under the normal topology. This is a generaliza-
tion of the properties of (A, B) and (c, c"'). We also study some other important
examples of topological algebras with such properties.

2. Definitions and Notation


We recall some relevant definitions. Through this section we assume that X is
a commutative complex topological algebra with unit element. X is called a locally
convex algebra if it is also a locally convex space. In this case its topology can be
given by means of a family (11.11",)"'E;l of seminorms such that for each index Q E ~,
there is an index f3 E ~ such that

(2.1)

for all x, y EX.


If relation (2.1) can be replaced by

(2.2)

for all x, y EX, then we say that X is locally multiplicatively convex (shortly
m-convex) algebra.
X is called a Bo-algebra if it is a complete metrizable locally convex algebra. In
this case its topology can be given by means of a sequence (1I.lIn)~=1 of seminorms
satisfying

for n = 1,2, ... and for all x,y E X.


Let (a-y,k), "I E r, k = 0,1, ... , be an infinite matrix of positive real numbers.
Assume that for each "I E r there is a "I' E r such that

(2.3) a-y,k+l :::; a-y',k a-y',l


for all k, l = 0, 1, ....
34 HUGO ARIZMENDI, ANGEL CARRILLO, AND LOURDES PALACIOS

The matrix algebra A(a/"k) associated with the matrix (a/"k) is the algebra of
00
all formal complex power series x = L Xkzk such that, for each "I E r,
k=O
00

IIxli/' = L a/"k IXkl < 00.


k=O
By (2.3), A(a/"k) is a complete locally convex algebra under the usual linear op-
erations and the convolution product. If r = N, then A(a/"k) is a Bo-algebra. If
"I = "I' in (2.3), then A(a/"k) is an m-convex algebra.
We shall be mainly interested in a Bo matrix algebra A(ap,n)' We note that
00

A(>O'(ap,n) = {(zn)~=o 1 L IZnYnl < 00, V Y = (Yn)~=o in AO(ap,n)}'


n=O
We have the following:
REMARK 2.1. AOO(ap,n) = A(ap,n)'
PROOF. For each p, the row (ap,n) is an element of AO(ap,n)' Therefore, if
00
Z E AOO(ap,n), then L IZnap,nl < 00 and hence Z E A(ap,n)' o
n=O
3. A Matrix Algebra and its a-dual
PROPOSITION 3.1.
AO(apn
,
) = limloo(-I-)
-; ap,n
= {(Yn);:"=o 13 p: sup
n
IYn-1
Qp,n
-1
< oo} as sets.

PROOF. Let Y E liml oo


-;
(_1_);
ap,n
then sup IYn
n
_1_1
ap,n
< 00 for some p. If

M = sup
n
IYn-I-I,
ap,n
we have that IYn 1 :s M ap,n for each n. Therefore, for
00 00

x E A(ap,n)' L IXnYnl :s M L
Ixnap,nl = M Ilxlip and hence Y E AO(ap,n)'
n=O n=O
Conversely, let us suppose that there exists Y E AO(ap,n) such that
yrf.liml
-+
oo (_I_); that is, yrf.loo(_I_) for each p. Since for each p, sup IYn
ap,n Qp,n n ap,n
= 00, _1_1
I'

then there exists an increasing sequence (np) such that IYn p ap,np _1_1 > p2. Let us
0 ifn#np
define the sequence x = (xn);:"=o as follows: Xn = { _I_~ =n .
if n
ap,np P P
We claim that x E A(ap,n):
f: f:
00

For arbitrary q, IIxllq = n=O Ixnlaq,n = p=O 1~~laq,np


v,np P
"
= p=O aq,np..!..
L.J a p,np p2
q 00
"
L.J a
aq,np 1
p2'
+ "L.J aq,np 1
a p2"
< 00 as claimed
p=O p,n" p=q+l ",n"
f:
Nevertheless n=O IXnYnl = n=O I~~ p,n" P
f:
Yn,,1 = 00 since I~~pl
p,np Yn,,1 > 1 for
every n. This is a contradiction to the assumption that Y E AO(ap,n)' 0
PROPOSITION 3.2. The inductive limit topology is stronger than the normal
topology in AO(ap,n).
ON o-DUAL ALGEBRAS 35

PROOF. Let us recall that 11.11 is a continuous seminorm in limloo(-I-)


--+ av,n
p

if, and only if, for each p, there exists a constant Cp such that
Ilyll ~ Cp lIylll"" = Cps~p IYna:.n I for every Y E loo(a:.J
Let Y E 100 (_1_),
ap,n
then sup
n
1
IYn-
ap,n
-1 = AI < 00 and therefore IYnl-1-
ap,n
~ AI
00
for each n. Hence, for x E A(ap,n) we have that IIYllx = I: IXnYnl ~ <
n=O

n~o AI IXnl ap,n = AI n~o IXnl aq,n = AI IIxllp = s~p IYna:.n Illxll p = Ilxllp Ilyllloo' D

In [4] it is proved that an echelon space A(ap,n) is nuclear if, and only if, for
each p there exist q > p and a sequence u = (un);:::'=o E II such that ap,n = Un aq,n,
n = 0,1, ...
PROPOSITION 3.3. If A(ap,n) is nuclear, then AO(ap,n), endowed with the nor-
mal topology, is such that AO(ap,n) ~ limloo(-I-) ~ limll(-I-).
-+ ap,n -+ ap,n
p p

PROOF. Due to the nuclearity of A(ap,n) it is easy to see that


limll(-I-) ~ limloo(-I-). By the previous Proposition we know that the inductive
--+ ap,n -+ ap,n
p p

limit topology is stronger than the normal topology in AO(ap,n).


Let 1111 be a seminorm in limll(-I-).
-+ ap,n
Let (e n );:::'=1 be the sequence of the
p

canonical vectors and note that (e n );:::'=1 E 11(_1_)


ap,n
for each p. Therefore there
exists Cp > 0 such that lIenll ~ Cpllenill oo = CPa :. n for each p. We put
(Xn);:::'=1 = (1Ien ll);:::'=I' We will prove that (Xn);:::'=1 EA(ap,n)'
00 00 00

Indeed, IIxnllp I: Ilenllap,n I: lIenllunaq,n < I: Cqllenllqunaq,n =


n=1 n=1 n=1
00

Cq I: Un < 00
n=1
Therefore, for any Y E loo(a:.J, IlylI = 1L~1 ynenll ~ n~IIYnlllenll =
00
I: IYnllxnl = IIYllx < 00. SO II . II is continuous with respect to the normal topol-
n=1
ogy. D
Let A(ap,n) be a matrix algebra. We define the following condition:
DEFINITION 3.4. (*) for each p there exist q > P and AIp such that
ap,n ap,m ~ AIp aq,n+m for all n,m.
Let us recall that the absolutely convex hull of a subset V of a vector space is the
set r(V) = {I: .Aiai, I finite, ai EV, .Ai EC and I: l.Ail ~ I}. It is a straight for-
iEI iEI
ward matter to check that if V and Ware two subsets of an algebra, then the
property (x, Y E V ~ xy E W) implies the property (x, Y E r(V) ~ xy E r(W)).
PROPOSITION 3.5. If the matrix algebra A(ap,n) satisfies (*), then l~ll(a:.J
p

is a topological algebra under the convolution product.


36 HUGO ARIZMENDI, ANGEL CARRILLO, AND LOURDES PALACIOS

PROOF. Let x, Y E 11(_1_). From condition (*), let q > p and Mp be such that

.;," <; Mp.: ":_,, The:'il:YII, ~ "~O I(xy)" I .;," ~ I Ikt IXk!ln-kll;" <;

f
n=O k=O
t IXkYn-kl a:-
q,n
f (t
:s n=O k=O
Mp IXka:-k IIYn-k
". a l_k
p.n
I) =

Mp C~O IXnl a:,n) C~O IYnl a:. n) = Mp IIxlip IIYllp < 00.

So far, limll( a:-) is an algebra. From the previous proof it is easy to see that
--+ p.rt.
p

if x E II (a:.J
and Y E II 1 then xy E II (a..J, (a:.J
and IIxYllq ::; MpllxllrllYIIs,
where q and Mp are any two numbers that satisfy (*) for p = max(r, s).
Now let n be a neighborhood of the origin in limll(-I-);
--+ ap,n
then
"
n= r CQl V(O, IIIIp' ep)), where ep > 0 for all p ~ 1.
For each p ~ 1, let qp and Mp be two numbers that satisfy (*) for p and let us
take 0 < t5p < min ~-). (1,
If x E V(O, II.IIr,er) and Y E V(O, IIIIs,es), then IIxYllqp MpllxllrilYlls, where :s
p = max(r,s); therefore IIxyllq" < eq", thus xyEV(O, II.IIqp,eq,,)'
From the result stated before this proposition, it follows that,

x, Y E r CQl V(O, IIIIp' t5p)) ::::} xy r CQl V(O, IIIIp' ep)) ~ n.


E This shows that
the multiplication is continuous. 0
COROLLARY 3.6. If the matrix algebra A(ap,n) is nuclear and satisfies (*), then
AQ(ap,n) is a topological algebra under the normal topology.
PROPOSITION 3.7. If the matrix algebra A(ap,n) does not satisfy (*), then
limll(-I-)
-+ ap,n
is not an algebra.
p

PROOF. The hypothesis gives a Po with the property that for each q > Po
and Mq there exist m and n such that apo,n apo,m > !l-lq aq,n+m, or, equivalently,
_ 1 _ > M _1_ _ _ 1_.
aq,n+m q a pQ n apQ,m .

Fix an i > Po and choose Mi = 22i. There are positive integers mi, ni such
that 1 > Mi _1_ _ _1_.
ai,ni+mi apQ,ni a'PO1'1~i
00 00

Let x = L: anz n and Y = L: bnz n , with ani = 2ill)illpo' bmi = 2illz;';1I,,0


n=O n=O
00

for i ~ 1 and an = bn = 0 otherwise. Then IIxli po = ~ 2illz~illpo IIznili po = 1 and


.=1
00 00 00

IIyllpo = ~
.=1
2'lIz;'ql" IIzm; IIpo
0
= 1; but, since xy= ?:
J=O
anbmzi = L: anibm;zni+m;
.=1
n+m=j
00

we have IIxYllq ~ . L: 2i+illzni lI~o IIzmj IIp() II zn;+mj IIi = 00


.=q+l
Therefore the space limll(-I-)
-+ ap,n
can not be an algebra (see also Proposition
p

3.1). o
ON ",,-DUAL ALGEBRAS 37

COROLLARY 3.8. A(a pn


,
) satisfies (*) if, and only if, limll(-I-)
-+ ap,n
is a topological
p

algebra under the convolution product.

COROLLARY 3.9. If A(ap,n) is nuclear, then it satisfies (*) if, and only if,
A""(ap,n) is a topological algebra under the convolution product.

EXAMPLES 3.10.
(1) Let us consider the matrix algebra A(ap,n) where

ap,n = (n + l)P, n = 0, 1,2, ....


In this case, A(ap n) I
= {(xn) E w: lim (n
n-+oo
+ l)P IXnl = 0, p = 1,2, ... }.
A(ap,n) and A""(ap,n) are nuclear algebras and one is the a-dual of
the other. A(ap,n) is nuclear since for each p we can take q = 3p and

(un)~=o = ((n+I)q p) ~=o = (( (n;l)p ) 2) ~=o' which clearly is in ll,


and satisfies ap,n = unaq,n for all n.
A""(ap,n) is an algebra due to Corollary 3.6 since for each p we can
take q = 2p and Mp = 1 to satisfy ap,nap,m ::; Mpaq,n+m for all n, m.
(2) Let us consider the matrix algebra A(ap,n) where
1
a p,n -
- -
n 2 --
P , nEN.

In this case, A(ap,n) satisfies (*). For, if p is given, let q = p + 1 and


2- p - 1
Mp=2 .
Note that for arbitrary n , m 'nnt.~ < n+m = .1+..!.. < 2. Then ap,n ap,m =
m - nm n m - aq,n+m
1 1 2-p-1 2-p-1 2 p-l
;:;:z=P ~ _ (n+m) _ (n+m) _ (n+m) - < 2TP - 1 - M
; P 1 - (nm)2 P - (n2m2)2 P 1 - n2 m 2 - - p.
(n+m)
However, the algebra A(ap,n) is not nuclear. For, if p < q are arbitrary,
_1_ 2- P 2- q 2q-2P
~
aq,n = ~
2-q = (.1)
n - = (.1)
n ~ E II only if ~ > 1, which is
2P q
impossible.
(3) If w is the linear space of all complex sequences, we note that w = A(ap,n)
I ifn::;p ""
where ap,n = { 0 if n > p . Then A (ap,n) = I!m(N). Both of them are
nuclear topological algebras. Here Corollary 3.9 does not apply because
it is not true that ap,n > 0 for all indices p and n.

References
[IJ R. Arizmendi. "Matrix Algebras and m-convexity". Demostratio Mathematica, Vol. XVII, no.
3, 1984.
[2J R. Arizmendi; A. Carrillo; L. Palacios. "On continuous multiplicative Mappings on Ananlytic
Sequence Spaces". Commentatione Mathematicae, XXXIX (1999).
[3] G. Kothe. Topological Vector Spaces l. A series of Comprehensive Studies in Mathematics,
Springer-Verlag, 1966.
[4] A. Pietsch. Nuclear Locally Convex Spaces. Springer-Verlag, 1969.
[5] O.Toeplitz. Die linearen volkommenen Riiume der Funktionentheorie. Comment. Math. Relv.
23 (1949), 222-242.
38 HUGO ARIZMENDI, ANGEL CARRILLO, AND LOURDES PALACIOS

(Hugo Arizmendi, Angel Carrillo) INSTITUTO DE MATEMATICAS. UNIVERSIDAD NACIONAL


AUTON OM A DE MEXICO, APDO POSTAL 14455, MEXICO, D.F.
E-mail address:hugolDservidor.unam.mx

(Lourdes Palacios) DEPARTAMENTO DE MATEMATICAS, UNIVERSIDAD AlTTONOMA METROPOL-


ITANA. Av. SAN RAFAEL ATLIXCO 186, COL. VICENTINA, 07340 MEXICO, D.F.
E-mail address:pafalDxanum.uam.mx
Contemporary Mathematics
Volume 328, 2003

A Connected metric space that is not separably connected

Richard M. Aron and Manuel Maestre

ABSTRACT. We construct a subset of the unit ball of foo that is connected but
not separably connected.

A topological space X is said to be separably connected if for every two points X, Y E


X there exists a connected and separable subset C{x,y} such that x,y E C{x,y}'
The reader may be interested, perhaps even amazed, at the fact that this
concept arises naturally in economics. In fact, in their work on utility theory, J.
Candeal, C. Herves, and E. Induntin [B] ask whether there is a connected metric
space that is not separably connected. They observe that there are examples of
connected topological spaces which are not separably connected. However, no ex-
ample of a connected metric space which is not separably connected is given and,
in fact, this issue is explicitly raised in [A]. The purpose of this note is to give
a negative answer to this question. In fact, this problem was solved by R. Pol
over 25 years ago, who produced a somewhat different construction in [C]. The
construction presented here seems natural, in view of its similarity to the standard
construction of a non-measurable subset of JR.
Our example will be a subset of the Banach space (foo, 111100) of all sequences
(Xn)~=l in lK = JR or C, satisfying

II(xn)~=11l := sup{lxnl : n = 1, 2... } < 00,


Given U E Roo and T/ > 0, we denote by B(u, T/) := {v E Roo : IIv - ull ::; T/},
i.e., B( u, T/) is the closed ball with center u and radius T/. We will single out the
unit vector e := (1,0, ..... ) in Roo and the associated linear functional c.p : Roo ---+ lK,
c.p( (xn)) = x}, for all (xn) E f oo . Clearly, Ker c.p = ((xn) E Roo : c.p( (xn)) = Xl = o}.
Let .N denote the set consisting of all non-empty subsets S C {2, 3, .. , n, ... }, and
for each such S, let Us = (Xn)~=l E foo : Xn = 1 if n E Sand Xn = 0 if
n ~ S, n = 1,2 .... The set == {us : S E .N} is uncountable since .N is, any
element of has norm one and Ilu - vii = 1 for all u,v E , u I- v. Recall now

2000 Mathematics Subject Classification. Primary 54D05, 46B26.


Key words and phrases. Metric spaces non separably connected, Banach spaces.
The first author was partially supported by the Ministerio de Educacion y Cultura of Spain
(SAB1999-0214).
The second author was partially supported by MCYT and FEDER Project BFM2002-01423.

2003 American Mathematical Society


39
40 RICHARD M. ARON AND MANUEL MAESTRE

the following equivalence relation on [0, 1], which is used to prove the existence of
a non-Lebesgue measurable subset of [0,1]:
x '" y if and only if x - y E Q.
We denote the equivalence classes of [0,1] defined by this relation by {x : x E
[0, I]}. By applying the Axiom of Choice, we can choose one element from each

class, in particular taking E O. Let us denote the set so obtained by A.
LEMMA 1.1. The following properties hold:
(1) Card (A) = Card([O, 1]).
(2) Given x, yEA, x =1= y and r, SEQ, x + r =1= y + s.
(3) If a E [0,1]' then there exist a unique x E A and a unique r E Q with
a = x+r.
(4) Given a < b E JR and x E JR, the family {x +r : r E Q} n (a, b) is a dense
subset of [a, b].
The proof of the above properties is immediate. Let us remark that (4) can be
obtained by using the fact that for each x E JR, the map : JR --+ JR defined by
(t) := x + t is a homeomorphism and the fact that Q n (a - x, b - x) is always a
dense subset of [a - x, b - x].
Also it is very easy to construct a bijection R : A --+ , R(x) = ex. We define our
set C as C = C 1 U C2 , where
C1 = U re+ [O,eo]
rE[O,l]nQ

and
U
xEA\{O}.rEQ,x+rE[O,l]
(x + r)e + (0, ex].

For an intuitive, geometric idea of what C is, consider a type of "comb" set in the
plane given by {(s, t) : 0::; s ::; 1, s ~ Q,
Q,O>t~-I}.
: ;
t ::; I} U {(s, t) : 0::; s ::; 1, s E

We endow C with the metric induced by the norm of loc. Our goal in this note
is to prove the following:
THEOREM 1.2. C is a connected metric space that is not separably connected.
PROOF. By Lemma 1.1.(2), given x, YEA, x =1= y, r, SEQ and any u, v E
Ker rp we have that rp((x + r)e + u) = x + r =1= y + s = rp((y + s)e + v). Hence for
x E A \ {O} and r E Q with x + r E [0,1]' we will have that
(1.1) C \ {(x + r)e + (0, ex]} c + r)) U rp-l((x + r, +00)).
rp-l(( -00, X
Consider a connected subset DeC such that 0, e + eo E D. If there exist x E
A \ {O} and r E Q with x + r E (0,1) and D n ((x + r)e + (0, ex]) = 0, then, from

rp(O) = < x + r < 1 = rp(e + eo) and (1.1), we obtain that D is not connected.
Hence for all x E A \ {O} and r E Q with x + r E (0,1) we have that there exists
< Ax ::; 1 with
(x + r)e + Axex E D.
Thus, {ex}xEA\{O} C span{e, D}. However, the family {ex}xEA\{O} is uncountable,
and Ile x - ey II = 1 for all x, yEA \ {O}, x =1= y. As a consequence span {e, D} is not
A CONNECTED METRIC SPACE THAT IS NOT SEPARABLY CONNECTED 41

separable, which implies that D is non-separable.

Thus, it only remains to show that C is connected. If C were not connected, then
there would exist a separation of C by subsets U and V; that is, there would exist
open (and closed) sets U and V such that
U =I- 0 =I- V, U U V = C, UnV = 0.
To simplify the notation, given Z E [0,1] we denote by z the only x E A such that
z = X, and we denote by Iz = ze + (0, e z] if z fj. Q and I z = ze + [0, eo] if Z E Q.
Since Iz is connected we have that given Z E [0,1] if I z n U =I- 0 then Iz C U
(and the same with respect to V). As eo E U U V, we will assume that eo E U and
hence 10 C U.
Consider Z E Q n [0,1] such that I z C U. Since U is open, there exists TJ =
TJ(z) > a such that the closed ball B(ze,2TJ) n C c U and moreover such that
(z - TJ, z + TJ) C [0,1] if Z E (0,1) n Q. Now it is immediate that for every u E f(X),
Ilull = 1, every y E (z-TJ, z+TJ)n[O, 1] and every A, 0:$ A:$ TJ, lIye+Au-zell :$ 2TJ.
By Lemma 1.1.(3), there exist unique Xl E A and rl E Q with y = Xl +rl. If Xl = 0,
then ye + [0, TJeo] = rle + [0, TJeo] C B(ze, 2TJ) n C l c U and hence rle + [0, eo] C U.
Alternatively, if Xl E A\ {a}, then ye+(O, TJe X1 ] = (Xl +rl)e+(O, TJe X1 ] C B(ze, 2TJ)n
C2 c U, and again it follows that (Xl + rde + (0, eX!] C U. In either case, we see
that Iy C U for all y E (z - TJ,Z +TJ) n [0,1].
Let A:= {z E Qn [0,1] : Iz C U} and B:= {z E Qn [0,1] : Iz C V}.
Let K := UzEA(Z - TJ, Z + TJ) n [0,1] and L := UzEB(Z - TJ, z + TJ) n [0,1] where
TJ = TJ(z) is given above. K and L are open sets in [0,1] and Q n [0,1] c L U K.
If t is in k, the closure of K, then given n E N we can find Zn E A such that
(Zn - TJn, Zn + TJn) n (t - ~, t + ~) n [0, 1] =I- 0, where 1Jn = TJ(zn). Hence there exists
a < b such that (a, b) C (zn - TJn, Zn + TJn) n (t - ~, t + ~) n [0, 1]. By Lemma 1.1.(4)
there exists r n E Q such that t + r n E (a, b). Thus, the preceding paragraph shows
that I Hrn C U and as t + rn = f, we have (t + rn)e + ef E U. Since Irnl < ~ for all
nand U is closed in C we have that te + er E U, thus It cU. The same argument
applied to L implies that if S E L, then Is C V. Hence k and L are two closed
disjoint sets. But as Q n [0,1] c K U L c [0,1] we have [0,1] = k U L. Since [0,1]
is a connected set and a E A c K we have that k = [0,1] and hence C = U, a
contradiction. Therefore C is connected. 0
REMARK 1.3. This result extends to any non-separable Banach space. To see
this, consider a non-separable Banach space F. As above we can take e E F with
lIell = 1 and a continuous, linear functional <p : F ~ II{ with <p(e) = 1. Once
again, Ker <p is a non-separable Banach space. An immediate application of Zorn's
e
Lemma yields the existence of a i5 > a and an uncountable set c Ker <p such that
lIuli = 1 for all u E e and lIu - vII ~ i5 for all u, vEe, u =I- v. As above, we take
e
a map R: A ~ with uncountable range {ex := R(x) : X E A}, and we define
our set C by C = C l U C 2 , where
C l := U re + [0, eo] and C2 := U (X + r)e + (0, ex].
rEQn[O,l] x+rE[O,l]
xEA\{O}, rEQ
If C is endowed with the metric induced by the norm of F, then, by a proof similar
to the one given in Theorem 1.2, C is a connected metric space that is not separably
connected.
42 RICHARD M. ARON AND MANUEL MAESTRE

ACKNOWLEDGEMENT. This note evolved during visits the first author made
to the Departamento de Amilisis Matematico of the Universidad de Valencia, and
while the second author was a Visiting Professor in the Department of Mathematical
Sciences at Kent State University for the 1998-99 academic year. The authors
express their thanks to the Departments concerned for their hospitality. In addition,
the authors are very grateful to the Referee, whose careful reading of the original
version of this paper led to the present considerably simplified proof of Theorem
1.2.

References
[A] Balbas, A., Estevez, M., Herves, C., and Verdejo, A. Espacios separablemente conexos, Rev.
R.Acas. Cien. Exact. Fis. Nat. (Spain), 92, 1, (1998), 35-40.
[B] J. Candeal, C. Herves and E. Indurain, Some results on representation and extension of
preferences, Journal of Mathematical Economics 29 (1998), 75-81.
[C] R. Pol, Two examples of non-separable metrizable spaces, ColI. Math. 33, (1975),209-211.
DEPARTMENT OF MATHEl\IATICS, KENT STATE UNIVERSITY, KENT. OH 44242 USA
E-mail address: aronClmcs. kent. edu

DEPARTAMENTO DE ANALISIS MATEMATlCO, FACULTAD DE MATEMATICAS, UNIVERSIDAD DE


VALENCIA, 46100 BURJASOT (VALENCIA). SPAIN
E-mail address: manuel.maestreCluv. es
Contemporary Mathematics
Volume 328, 2003

Weighted Chebyshev Centres and


Intersection Properties of Balls in Banach Spaces

Pradipta Bandyopadhyay and S Dutta

ABSTRACT. Vesely has studied Banach spaces that admit weighted Chebyshev
centres for finite sets. Subsequently, Bandyopadhyay and Rae had shown, inter
alia, that Lt-preduals have this property. In this work, we investigate why
and to what extent are these results true and thereby explore when a more
general family of sets admit weighted Chebyshev centres. We extend and
improve upon some earlier results in this general set-up and relate them with
a modified notion of minimal points. Special cases when we consider the family
of all finite, or more interestingly, compact subsets lead to characterizations of
Lt-preduals. We also consider some stability results.

1. Introduction
Let X be a Banach space. We will denote by Bx[x,r] the closed ball ofradius
r > 0 around x EX. We will identify any element x E X with its canonical image
in X**. Our notations are otherwise standard. Any unexplained terminology can
be found in either [6] or [10).
In this paper we continue the study of Banach spaces that admit weighted
Chebyshev centres that began with [3).
DEFINITION 1.1. Let Y be a subspace of a Banach space X. For A ~ Y and
p:A ---? ~+, define

A,p(X) = sup{p(a)llx - all : a E A}


A point Xo E X is called a weighted Chebyshev centre of A in X for the weight p
if A,p attains its minimum at Xo.
When A is finite, Vesely [18) has shown that if X is a dual space, A admits
weighted Chebyshev centres in X for any weight p, that the infimum of A,p over
X and X** are the same, and
THEOREM 1.2. [18, Theorem 2.7) For a Banach space X and aI, a2, ... ,an E
X, the following are equivalent :
2000 Mathematics Subject Classification. Primary 4IA65, 46B20; Secondary 41A28, 46B25,
46E15, 46E30.
Key words and phrases. Weighted Chebyshev centres, minimal points, central subspaces,
I-complemented subspace, I PI,oo, Ll_preduals.

2003 American Mathematical Society


43
44 BANDYOPADHYAY AND DUTTA

(a) Ifrl, r2, ... , rn > 0 and ni=IBx " [ai, riJ =10, then ni=IBx[ai, riJ =10.
(b) {al,a2, ... ,an } admits weighted Chebyshev centres for all weights
rl,r2, ... ,rn > O.
(c) {aI, a2,' .. , an} admits f-centres for every continuous monotone coer-
cive f : IR+. -4 IR (see [18J for the definitions).
In this work, we investigate why and to what extent are these results true
and thereby explore when a more general family of sets admit weighted Chebyshev
centres. Extending the notion of central subspaces introduced in [3], we define an
A-C-subspace Y of a Banach space X with the centres of the balls coming from a
given family A of subsets of Y, the typical examples being those of finite, compact,
bounded or arbitrary sets. The first gives us the central subspace a la [3J and
the last one is related to the Finite Infinite Intersection Property (I Pj,oo) [8J. We
extend and improve upon some results of [3, 18J in this general set-up and relate
them with a modified notion of minimal points. We also improve upon one of the
main results of [4J on the structure of the set of minimal points of a compact set. As
in [3], special cases when we consider the family of all finite, or more interestingly,
compact subsets lead to characterizations of Ll-preduals. We also consider some
stability results.

2. General Results
We first extend VeselY's result in [18J on dual spaces from finite sets to all the
way upto bounded sets and also strengthens its conclusions. We need the following
notions.
DEFINITION 2.1. Let X be a Banach space and A ~ X.
(a) We define a partial ordering on X as follows : for Xl, X2 EX, we say
that Xl :SA X2 if IlxI - all :s
IIx2 - all for all a E A. We will denote
by mx (A) the set of points of X that are minimal with respect to the
ordering :SA and often refer to them as :SA-minimal points of X.
Note that :SA defines a partial order on any Banach space containing
A and we will use the same notation in all such cases.
(b) A function f : X -4 IR+ is said to be A-monotone if f(xd :s
f(x2)
whenever Xl :SA X2
(c) Let Y be a subspace of X and A ~ Y. Following [9], we say X E X is
a minimal point of A with respect to Y if for any y E Y, Y :SA X implies
y=x.
We denote the set of all minimal points of A with respect to Y in
X by Ay,x, Note Ay,x ;2 A. For A ~ X, the set Ax,x will be called
minimal points of A in X, and will be denoted simply by min A.
(d) For A ~ X bounded, the Chebyshev radius of A in X is defined by
r(A) = inf sup Ilx - all.
xEX aEA

THEOREM 2.2. (a) If A ~ X is bounded and X ~ A + r(A)B(X), then


there exists y E X such that y :SA x.
(b) If X = Z is a dual space and A is bounded, then every A-monotone
and w*-lower semicontinuous (henceforth, lsc) f : X -4 IR+ attains its
minimum. In particular, for every p, A,p attains its minimum.
CHEBYSHEV CENTRES AND INTERSECTION PROPERTIES OF BALLS 45

(c) If X = Z* is a dual space, for every Xo E X, there is a Xl E mx(A)


such that Xl ::;A Xo. In particular, the minimum in (b) is attained at a
point ofmx(A).
PROOF. (a). Let X ~ A + r(A)B(X). Then, there exists c > 0 such that
IIx - all > r(A) + c for all a E A. By definition of r(A), there exists y E X such
that sUPaEA Ily - all < r(A) + c. Clearly, y ::;A x.
(b). By (a), if X ~ A + r(A)B(X), there exists y E X such that y ::;A X,
and hence, f(y) ::; f(x). Thus, the infimum of f over X equals the infimum over
A + r(A)B(X). Moreover, since X is a dual space and f is w*-lsc, it attains its
minimum over any w*-compact set. Thus f actually attains its minimum over X
as well.
Since the norm on X is w*-lsc, so is cPA,p for every p.
(c). Consider {x EX: x ::;A xo}. Let {Xi} be a totally ordered subset. Let z
be a w*-limit point of Xi. Since the norm is w*-lsc, we have
liz - all ::; lim inf Ilxi - all = inf Ilxi - all for all a E A.
Thus the family {Xi} is ::;A-bounded below by z.
By Zorn's lemma, there is a Xl E mx(A) such that Xl ::;A Xo.
Now let Xo be a minimum for f. There is a Xl E mx(A) such that Xl ::;A Xo.
Clearly, f attains its minimum also at Xl. 0

REMARK 2.3. (a) It follows that for any bounded set A, minA C
A + r(A)B(X). This improves the estimates in [9] or [18].
(b) Apart from cPA,p, there are many examples of A-monotone and w*-
lsc f : X = Z* ---- lR+. One particular example that has been treated
extensively in [4] is the function cPf.,l defined by cPf.,l(x) = fA Ilx - aI1 2 dJ.L(a),
where J.L is a probability measure on a compact set A ~ X.
(c) Observe that though minimal points of A are ::;A-minimal, there is
some distinction between the two notions. The two notions coincide if X
is strictly convex. See Proposition 3.1 below.
Now, if A is a bounded subset of a Banach space X, then by Theorem 2.2, A
has a weighted Chebyshev centre in X**. But what about a weighted Chebyshev
centre in X?
When A is finite, Vesely [18] has shown that the infimum of cPA,p over X and
X** are the same, and A admits weighted Chebyshev centres in X for any weight
p if and only if X satisfies Theorem 1.2(a). We now show that both of these are
special cases of more general results. We need the following definition.
DEFINITION 2.4. Let Y be a subspace of a Banach space X. Let A be a family
of subsets of Y.
(a) We say that Y is an almost A-C-subspace of X if for every A E A,
X E X and c > 0, there exists y E Y such that

(1) Ily - all::; IIx - all + c forall a E A.


(b) We say that Y is an A-C-subspace of X if we can takec = 0 in (a).
(c) If A is a family of subsets of X, we say that X has the (almost) A-IP
if X is an (almost) A-C-subspace of X**.
46 BANDYOPADHYAY AND DUTTA

Some of the special families that we would like to give names to are :
(i) F = the family of all finite sets,
(ii) K = the family of all compact sets,
(iii) B = the family of all bounded sets,
(iv) 'P = the power set.
Since these families depend on the space in which they are considered, we will
use the notation F(X) etc. whenever there is a scope of confusion.
REMARK 2.5. (a) Note that F-C-subspaces were called central (C)
subspaces in [3], 'P-C-subspaces were called almost constrained (AC) sub-
spaces in [1, 2]. Also if X has the F-IP, it was said to belong to the class
(GC) in [18, 3], and the 'P-IP was called the Finite Infinite Intersection
Property (IPj,oo) in [7, 2].
(b) The definition of almost A-C-subspace is adapted from the definition
of almost central subspace defined in [17]. The exact analogue of the
definition in [17] would have, in place of condition (1),
sup Ily - all::; sup Ilx - all + c.
aEA aEA
Clearly, our condition is stronger. We observe below (see Proposition 2.7)
that this definition is more natural in our context.
(c) By the Principle of Local Reflexivity (henceforth, PLR), any Banach
space has the almost F-IP. More generally, if Y is a!l ideal in X (see
definition below), then Y is an almost F-C-subspace of X.
DEFINITION 2.6. A subspace Y of a Banach space X is said to be an ideal in
X if there is a norm 1 projection P on X* with ker(P) = y.l.
PROPOSITION 2.7. Let Y be a subspace of a Banach space X. Let A be a family
of bounded subsets of Y. Then the following are equivalent:
(a) Y is an almost A-C-subspace of X
(b) for all A E A and p : A ---t lR.+, if naEABx [a, p(a)] - 0, then for every
e > 0, naEABy[a, p(a) + c] - 0.
(c) for every bounded p, the infimum of A,p over X and Yare equal.
PROOF. Equivalence of (a) and (b) is immediate and does not need A to be
bounded.
(a) ::::} (c). Let Y be an almost A-C-subspace of X, A E A and p : A ---t lR.+
be bounded. Let M = supp(A). Let e > O. By definition, for x E X, there exists
y E Y such that
lIy - all::; Ilx - all + e for all a E A.
It follows that
p(a)lly - all ::; p(a)llx - all + p(a)e ::; p(a)llx - all + Me for all a E A.
and hence,
A,p(Y) ::; A,p(X) + Me.
Therefore,
inf A,p(Y) ::; inf A,p(X) + Me.
As e is arbitrary, the infimum of A,p over X and Yare equal.
CHEBYSHEV CENTRES AND INTERSECTION PROPERTIES OF BALLS 47

(c) =} (a). Let A E A, x E X and e > O. We need to show that there exists
y E Y such that
Ily - all sIIx - all + e for all a E A.
If x E Y, nothing to prove. Let x E X \ Y. Let N = sUPaEA IIx - all. Let
p(a) = l/llx - all. Since x r/: Y and A ~ Y, p is bounded. Then A.p(X) = 1,
and therefore, inf A,p(X) S 1. By assumption, inf A,p(Y) = inf A,p(X) S 1,
and so, there exists y E Y, such that A,p(y) s
1 + e/N. This implies Ily - all S
Ilx - all + ell x - all/N S Ilx - all + e for all a E A. 0
As noted before, by PLR, any Banach space has the almost F-IP. And therefore,
the result of [18J follows.
PROPOSITION 2.8. Let A and Al be two families of subsets of Y such that for
every A E A and e > 0, there exists Al E Al such that A ~ Al + cB(Y). If Y is
an almost AI-C-subspace of X, then Y is an almost A-C-subspace of X as well.
Consequently, any ideal is an almost K-C -subspace and any Banach space has
the almost K-IP. In particular, if A is a compact subset of X and p : A ---+ lR+ is
bounded, then the infimum of A,p over X and X** are the same.
PROOF. Let A E A and e > O. By hypothesis, there exist Al E Al such that
A ~ Al + eB(Y). Let x E X. Since Y is an almost AI-C-subspace of X, there
exists y E Y such that
lIy - alii S Ilx - alII + e/3 for all al E AI.
Now fix a E A. Then there exists al E Al such that lIa - alii < e/3. Then
Ily - all Ily - alii + lIa - alii S Ilx - alii + 2e/3
S
Ilx - all + Iia - alii + 2e/3 S IIx - all + e.
S
Therefore, Y is an almost A-C-subspace of X as well.
Since any Banach space has the almost F-IP, by the above, it has the almost
K-IP too. The rest of the result follows from Proposition 2.7. 0
EXAMPLE 2.9. Vesely [18J has shown that if A is infinite, the infimum of A,p
over X and X** may not be the same. His example is X = co, A = {en: n ;::: I}
is the canonical unit vector basis of Co and p == 1. Then inf A,p(X) = 1 and
inf A,p(X**) = 1/2. The example clearly also excludes countable, bounded, or,
taking Au {O}, even weakly compact sets. Thus Co fails the almost B-IP, almost
P-IP and if A is the family of countable or weakly compact sets, then Co fails the
almost A-IP too.
Stronger conclusions are possible for A-IP.
LEMMA 2.10. Let Y be a subspace of a Banach space X. For A ~ Y, the
following are equivalent :
(a) For every A-monotone f : A ---+ lR+ and x E X, there exists y E Y
such that f(y) S f(x).
( b) For every p : A ---+ lR+ and x EX, there exists y E Y such that
A,p(Y) S A,p(X).
(c) For every continuous p : A ---+ lR+ and x EX, there exists y E Y such
that A,p(Y) S A,p(X).
(d) For every bounded p : A ---+ lR+ and x EX, there exists y E Y such
that A,p(Y) S A,p(X).
48 BANDYOPADHYAY AND DUTTA

(e) Any family of closed balls centred at points of A that intersects in X


also intersects in Y.
(f) for any x EX, there exists y E Y such that y :SA x.
It follows that whenever any of the above conditions is satisfied, for every A-
monotone f : A --+ lR+, the infimum of f over X and Yare equal and if A
has a weighted Chebyshev centre in X, it has a weighted Chebyshev centre in Y.
PROOF. (a) => (b) => (c), (b) => (d) and (e) } (f) => (a) are obvious.
(c) or (d) => (f). As in the proof of Proposition 2.7, let p(a) = l/llx-all. Then
p is continuous and bounded and <PA,p(X) = 1. Thus, there exists y E Y such that
:s
<pA,p(y) 1. This implies lIy - all :s IIx - all for all a E A. 0
We now conclude the discussion so far by obtaining the extension of Theo-
rem 1.2.
THEOREM 2.11. For a Banach space X and a family A of bounded subsets
of X, the following are equivalent:
(a) X has the A-IP.
(b) For every A E A and every f : X** --+ lR+ that is A-monotone and
w*-lsc, the infimum of f over X** and X are equal and is attained at a
point of x.
(c) For every A E A and every p, the infimum of <PA,p over Xu and X
are equal and is attained at a point of X.
Moreover, the point in (b) or (c) can be chosen to be :SA-minimal.
We now study different aspects of A-C-subspaces.
DEFINITION 2.12. Let Y be a subspace of a Banach space X. Let A ~ Y. For
x E X and X* E B(X*), define
U(x,A,x*) = inf{x*(y) + IIx - yll : yEA}
L(x, A, x*) = sup{x*(y) -lix - yll : yEA}
The following lemma is in [1]. We include the proof for completeness.
LEMMA 2.13. Let Y be a subspace of a Banach space X and A ~ Y. For
:s
Xl, X2 EX, X2 :SA Xl if and only if for all x* E B(X*), U(X2, A, x*) U(XI, A, x*).
PROOF. If X2 :SA Xl, then for all x* E B(X*), x*(y) + IIx2 - yll :s
x*(y) +
IIxI -YII. And therefore, U(x2,A,x*):S U(xI,A,x*).
Conversely, suppose IIx2 - Yo II > IIxI - Yo II for some Yo E A. Then there
exists c > 0 such that IIx2 - Yo II - c ~ IIxI - Yoli. Choose x* E B(X*) such that
IIxI - Yo II :s
IIx2 - yolI- c < X*(X2 - Yo) - c/2. Thus U(XI, A, x*) :s
x*(yo) + IIxI -
Yoll < X*(X2) - c/2 < U(X2, A, x*). 0
REMARK 2.14. Instead of B(X*), it suffices to consider the unit ball of any
norming subspace of X*.
We compile in the following propositions several interesting facts about A-C-
subspaces and the A-IP.
PROPOSITION 2.15. Let Y be a subspace of a Banach space X. For a family A
of subsets of Y, the following are equivalent:
(a) Y is an A-C-subspace of X
CHEBYSHEV CENTRES AND INTERSECTION PROPERTIES OF BALLS 49

(b) for every x E X and A E A, there exists y E Y such that U(y, A, x*) ::;
U(x,A,x*) for every x* E B(X*).
(c) for any A E A, Ay,x ~ Y.
PROOF. This follows from Lemma 2.13 and the definition of Ay,x. 0
COROLLARY 2.16. X has the P-IP if and only if for every x** E X**, there
exists x E X such that x is dominated on B(X*) by the 1~pper envelop of x**
considered as a function on B(X*) equipped with the w*-topology.
PROOF. Observe that for any x E X, U(x,X,) == x on B(X*) and for x** E
X**, U(x**, X, x*) is the upper envelop of x** considered as a function on B(X*)
equipped with the w*-topology (see [8]). 0
PROPOSITION 2.17. (a) Let X be a Banach space and let Y be a sub-
space of X. Let A be a family of subsets of Y and let Al be a subfamily
of A. If Y is a A-C -subspace of X, then Y is a Al -C -subspace of X as
well. In particular, P-IP implies B-IP implies /C-IP implies F-IP.
(b) 1-complemented subspaces are A-C-subspaces for any A.
(c) Let Z ~ Y ~ X and let A be a family of subsets of Z. If Z is an
A-C-subspace of X, then Z is an A-C-subspace of Y. And, if Y is an
A-C-subspace of X, then the converse also holds.
PROOF. The proof follows the same line of argument as in [3, Proposition 2.2].
We omit the details. 0
PROPOSITION 2.18. For a family A of subsets of a Banach space X, the fol-
lowing are equivalent :
(a) X has the A-IP
(b) X is a A-C-subspace of some dual space.
(c) for all A E A and p : A ---+ lR+, ni=l Bx [ai, p(ai) + c] =f. 0 for all finite
subset {aI, a2, ... ,an} ~ A and for all c > 0 implies naEABx [a, p(a)] =f. 0.
In particular, any dual space has the A-IP for any A. Let S be any of the
families F, /C, B or P. The S-IP is inherited by S-C-subspaces, in particular, by
1-complemented subspaces.
PROOF. Clearly, (a) =} (b), while (c) =} (a) follows from the PLR.
(b) =} (c). Let X be an A-C-subspace of Z*. Consider the family {Bz. [a, p(a)+
c] : a E A, c > O} in Z*. Then, by the hypothesis, any finite subfamily intersects.
Hence, by w*-compactness, naEABz [a, p(a)] =f. 0. Since X is an A-C-subspace of
Z*, we have naEABx[a,p(a)] =f. 0. 0
The following result significantly improves [3, Proposition 2.8] and provides yet
another characterization of the A-IP.
PROPOSITION 2.19. Let Y be an almost F-C subspace of a Banach space X.
Let A be a family of subsets of Y. If Y has the A-IP, then Y is an A-C -subspace
of X. In particular, the conclusion holds when Y is an ideal in X.
PROOF. Let x EX, A E A. Since Y be an almost F-C subspace of X, for all
finite subset {aba2, ... ,an} ~ A and for all c > 0, nb:IBy[ai, Ilx.,... aill + c] =f. 0.
Since Y has the A-IP, by Proposition 2.18(c), naEABy[a, Ilx - alll =f. 0. 0
Since X is always an ideal in X**, the following corollary is immediate.
50 BANDYOPADHYAY AND DUTTA

COROLLARY 2.20. For a Banach space X and a family A of subsets of X, the


following are equivalent :
. (a) X has A-IP.
(b) X is an A-C-subspace of every superspace Z in which X embeds as an
almost F -C subspace.
(c) X is an A-C-subspace of every superspace Z in which X embeds as an
ideal.

3. Strict convexity and minimal points


PROPOSITION 3.1. If a Banach space X is strictly convex, then for every A ~
X, min A = mx(A).
PROOF. As we have already observed, min A ~ mx(A).
Let Xo E mx(A) and Xo ~ minA. Then there is an x E X such that x f= Xo
and x :S::A Xo. Since Xo E mx(A), we must have Ilx - all = Ilxo - all for all a E A.
Since X is strictly convex, II (x + xo)/2 - all < IIxo - all for all a. This contradicts
that Xo E mx(A). Hence Xo EminA. D
REMARK 3.2. If X is strictly convex, by a similar argument, for every Xo EX,
there is at most one Xl E mx(A) such that Xl :S::A Xo. Thus for a strictly convex
dual space, for every Xo E X*, there is a unique xi E mx (A) such that xi :S::A Xo'
PROPOSITION 3.3. Let X be strictly convex. Let A be a compact subset of X.
For each continuous p, A admits at most one weighted Chebyshev centre.
PROOF. Suppose A admits two distinct weighted Chebyshev centres Xo, Xl E
X. Then <PA,p(XO) = <PA,p(Xt) = r (say). Then for all a E A, we have XI. Xo E
Bx[a,r/p(a)]. By rotundity Z = (Xl +xo)/2 is in the interior of Bx[a,r/p(a)] for
all a. Thus, p(a)lIz - all < r, for all a. Since p is continuous, <PA,p(Z) < r, which
contradicts that minimum value is r. D
THEOREM 3.4. Let X be a Banach space such that
(i) X has the F-IP; and
(ii) for every compact set A ~ X, mx(A) is weakly compact.
Then X has the K-IP. Moreover, if X** is strictly convex, then the converse
also holds.
PROOF. Let X have the F-IP and for every compact set A ~ X, let mx(A)
be weakly compact. Observe that for any B ~ A, we have mx(B) ~ mx(A).
Let A ~ X be compact and let X** E X**. By Lemma 2.10, it suffices to show
that there is a Zo E X such that Ilzo - all :s:: Ilx** - all for all a E A.
Let {an} be a norm dense sequence in A. Take a sequence Ck --+ O. By com-
pactness of A, for each k, there is a nk such that A ~ U~k BX[an,ck]. Since X has
the F-IP, there exists Zk E nlk Bx [an' Ilx** - anll] and Zk E mx( {al, a2 ... ank }) ~
mx(A). Then IIzk-all :s:: Ilx**-all+2ck for all a EA. Now, by weak compactness of
m x (A), we have, by passing to a subsequence if necessary, Zk --+ Zo weakly for some
Zo EX. Since the norm is weakly Isc, we have Ilzo -all :s:: lim inf IIZk -all :s:: Ilx** -all
for all a E A.
Conversely, let X have the K-IP and X** be strictly convex. Let A ~ X be
compact. It is enough to show that any sequence {xn} ~ mx(A) has a weakly
convergent subsequence. Without loss of generality, we may assume that {xn} are
CHEBYSHEV CENTRES AND INTERSECTION PROPERTIES OF BALLS 51

all distinct. By Remark 2.3 (a), mx(A) ~ A + r(A)B(X) is bounded. Let x** be
a w*-cluster point of {x n } in Xu. It suffices to show that x** EX.
Suppose x** E X** \ X. Since X has the K-IP, there exists Xo E mx(A)
such that Ilxo - all ::; Ilx** - all for all a E A. Since X** is strictly convex,
lI(x** +xo)/2 - all < IIx** - all for all a E A. Since (x** +xo)/2 E X** \X, by K-IP
again, there exists Zo E mx(A) such that Ilzo -all::; II(x** +xo)/2-all < IIx** -all
for all a E A.
Since A is compact, there exists c > 0 such that Ilzo - all < IIx** - all - c for
all a E A. Observe that
Ilzo - all < Ilx** - all- c ::; liminf
n
IIx n - all- c for all a E A.

Therefore, for every a E A, there exists N(a) EN such that for all n ~ N(a),
IIzo - all < Ilxn - all- c. By compactness, there exists N E N such that IIzo - all <
IIx n - all - c/4 for all n ~ N and a E A. Thus, Zo ::;A Xn for all n ~ N. Since
Xn E mx(A) and X is strictly convex, Zo = Xn for all n ~ N. This contradiction
completes the proof. 0

REMARK 3.5. In proving sufficiency, one only needs that {Zk} has a subsequence
convergent in a topology in which the norm is lsc. The weakest such topology is
the ball topology, bx . So it follows that if X has the F-IP and for every compact
set A ~ X, mx(A) is bx-compact, then X has the K-IP. Is the converse true?
COROLLARY 3.6. [4, Corollary 1] Let X be a reflexive and strictly convex Ba-
nach space. Let A ~ X be a compact set. Then min(A) is weakly compact.

REMARK 3.7. Clearly, our proof is simpler than the original proof of [4].
If Z is a non-reflexive Banach space with Z*** strictly convex, then X = Z
is a non-reflexive Banach space with K-IP such that X** is strictly convex. Thus,
our result is also stronger than [4, Corollary 1].

4. L l -preduals and PI-spaces


Our next theorem extends [3, Theorem 7], exhibits a large class of Banach
spaces with the K-IP and produces a family of examples where the notions of F-
C-subspaces and K-C-subspaces are equivalent.
DEFINITION 4.1. (a) [12] A Banach space X is called an Ll-predual
if X* is isometrically isomorphic to 1(J-t) for some positive measure J-t.
(b) [11] A family {Bx[xi,ri]} of closed balls is said to have the weak
intersection property iffor all x* E B(X*) the family {Ba[x*(xd,ri]} has
nonempty intersection in JR.
THEOREM 4.2. For a Banach space X, the following are equivalent:
( a) X is a K -C -subspace of every superspace
(b) X is a K-C -subspace of every dual superspace
( c) X is a F -C -subspace of every superspace
(d) X is an almost F -C -subspace of every superspace
(e) X is a F-C-subspace of every dual superspace
(f) X is an almost F-C-subspace of every dual superspace
(g) X is an Ll-predual.
52 BANDYOPADHYAY AND DUTTA

PROOF. Observe that if X ~ Y ~ y** and X is a A-C-subspace of Y**,


then X is a A-C-subspace of Y. Thus (a) {::} (b) and (c) {::} (e). And clearly,
(a) ::::} (c) ::::} (d) ::::} (f).
(f) ::::} (g). Since the definition of almost central subspaces in [17J is weaker
than our definition of almost F-C-subspaces, this follows from [17, Theorem 1,
2::::} 3J
(g) ::::} (a). Suppose X is an LI-predual, and let X ~ Y. Let A ~ X be compact
with at least three points. Let Yo E Y. Then the family of balls {Bx[a, Ilyo - allJ :
a E A} have the weak intersection property. Since X is an l-predual and since the
centres of the balls are in a compact set, by [14, Proposition 4.4J, naEABx [a, Ilyo-
allJ - 0.
If A has two points, observe that two balls intersect if and only if the distance
between the centres is less than or equal to the sum of the radii, it is independent
of the ambient space. 0
COROLLARY 4.3. Every LI-predual has the /C-IP and hence also the F-IP.
PROPOSITION 4.4. Suppose X is an LI -predual space. Then for a subspace
Y ~ X, the following are equivalent :
(a) Y is an ideal in X
(b) Y is a /C-C-subspace of X
(c) Y is a F-C-subspace of X
(d) Y is an almost F-C-subspace of X
(e) Y itself is an LI -predual
PROOF. (e) ::::} (b) follows from Theorem 4.2 and (e) ::::} (a) follows from [16,
Proposition IJ. And clearly, (b) ::::} (c) ::::} (d) and (a) ::::} (d).
(d) ::::} (e). This again is an easy adaptation of the proof of [17, Theorem 1,
2 ::::} 3J. We omit the details. 0
The analog of Theorem 4.2 for 'P-C-subspaces involves 'PI-spaces.
DEFINITION 4.5. Recall that a Banach space is a 'PI-space if it is 1-
complemented in every superspace.
THEOREM 4.6. For a Banach space X, the following are equivalent:
(a) X is a 'PI -space
( b) X is i-complemented in every dual space that contains it
( c) X is a 'P -C -subspace of every superspace
(d) X is a 'P -C -subspace of every dual space that contains it
( e) X is isometric to C (K) for some extremally disconnected compact
Hausdorff space K.
PROOF. (a) {::} (b) and (c) {::} (d) follow as in the first paragraph of Theo-
rem 4.2. And clearly, (a) ::::} (c).
(d) ::::} (a). By Proposition 2.18 and Theorem 4.2, (d) implies X is an LI_
predual with 'P-IP. Recall that [12, Theorem 3.8J a Banach space X is a 'PI-space
if and only if every pairwise intersecting family of closed balls in X intersects. And
that X is a LI-predual if and only if X** is a 'PI-space.
Now given a pairwise intersecting family of closed balls in X, since X** is a
'PI-space, they intersect in X**. And since X has 'P-IP, they intersect in X too.
(a) {::} (e) is also observed in [12, Section 11J. 0
CHEBYSHEV CENTRES AND INTERSECTION PROPERTIES OF BALLS 53

PROPOSITION 4.7. Let A be a family of subsets of X such that F ~ A. Then,


the following are equivalent :
(a) X is an Ll -predual with A-IP
(b) X is an A-C-subspace of every superspace
(c) for every A E A, every pairwise intersecting family of closed balls in X
with centres in A intersects.
PROOF. (a) =? (b). Since X has the A-IP, it is an A-C-subspace of every
superspace in which it is an ideal (Proposition 2.20) and since X is an L1-predual,
it is an ideal in every superspace [16, Proposition 1]. Thus (b) follows.
(b) =? (a). Since F ~ A, this is immediate.
(a) =? (c). This is similar too the proof of Theorem 4.6 (d) =? (a).
(c) =? (a). If every finite family of pairwise intersecting closed balls in X
intersects, then X is an L1-predual. And that X has the A-IP follows from Propo-
sition 2.18 (c). 0

Let C(T, X) be the space of all X-valued bounded continuous functions on


a topological space T equipped with the sup norm. We now characterize when
C(T, X) is a real L1-predual. First we need the following lemma.
LEMMA 4.8. Suppose Y is a subspace of a Banach space X and Y is a
real L1-predual. Let A ~ Y be a compact set and r : A -7 lR,+ be such that
naEABx[a,r(a)] =I- 0. Let y E naEABy[a,r(a) + c] for some c > O. Then there
exists z E naEABy [a, r(a)] such that Ily - zll S c.
PROOF. Since naEABx [a, r(a)] =I- 0, and intersection of intervals is an interval,
for any y* E B(Y*), naEABJR[y*(a), r(a)] =I- 0 and is a closed interval. As y*(y) E
naEABJR[y*(a),r(a) + c] for any y* E B(Y*), the family {By[y,c].By[a,r(a)] :
a E A} is a weakly intersecting family of balls in Y. Since Y is a L1-predual,
By[y,C]nnaEABy[a,r(a)] =1-0. 0

PROPOSITION 4.9. A Banach space X is a real Ll -predual if and only if for


each paracompact space T, C(T, X) is a real L1-predual.
PROOF. Since X is I-complemented in C(T, X), hence a K-C-subspace, by
Proposition 4.4, if C(T, X) is an L1-predual, then so is X.
Conversely, suppose X is a real L1-predual. Let Z = C(T, X),
{!I, 12,, fn} ~ Z and rl, r2,, rn > 0 be such that the family {Bz[fi, ri] :
i = 1, ... , n} intersects weakly. Then for each t E T, the family {B x [fi (t), r i] : i =
1, ... , n} intersects weakly, and since X is a real L I-predual, they intersect in X.
Consider the multi-valued map F: T -7 X given by F(t) = nf=lBx[Ji(t),ri]. Note
for each t, F(t) is a nonempty closed convex subset of X.
CLAIM : F is lower semicontinuous, that is, for each U open in X, the set
V = {t E T: F(t) n U =I- 0} is open in T.
Let to E V. Let Xo E F(to) n U. Let c > 0 be such that lI:r - xoll < c implies
x E U. Let W be an open subset of to such that t E W implies 111;(t)- fi(to) II < c/2
for all i = 1, ... , n. We will show that W ~ V.
Let t E W. Then for any i = 1, ... , n, Ilxu - fi(t)11 S Ilxo - f.i(tO) II + 111;(to) -
fi(t)11 Sri + c/2. Therefore, Xo E nf=lBx[fi(t),rj + c/2]. By Lemma 4.8, there
exists z E F(t) = nf=lBx[fi(t), ri] such that Ilxo-zll S c/2 < c. Then z E F(t)nU,
and hence, t E V. This completes the proof of the claim.
54 BANDYOPADHYAY AND DUTTA

Now since T is paracompact, by l\Iichael's selection theorem, there exists 9 E Z


such that g(t) E F(t) for all t E T. It follows that 9 E ni'=l Bz[f;, ril, 0
REMARK 4.10. For T compact Hausdorff, this result follows from [13, Corollary
2, p 43]. But our proof is simpler.

5. Stability Results
In this section we consider some stability results. With a proof similar to [3,
Proposition 14], we first observl' that
PROPOSITION 5.1. K-IP is a separ-ably determined property. i.e .. if every sep-
amble subspace of a Banach space X have K-IP. then X also has K-IP.
DEFINITION 5.2. [10] A subspace Y of a Banach space X is called a semi-L-
summand if there exists a (nonlinear) projection P : X ----> Y such that
P(>..x + Py) >"P.r + Py, and
Ilxll IIPxl1 + Ilx - VI'II
for all x, y EX. >.. scalar.
In [3]. it was shown that semi-L snmmands are .1'-C-subspaces. Basically the
same proof actually shows that
PROPOSITION 5.3. A semi-L-summand is an A-C-subspace for any A.
Our next result concerns proximinal subspaces.
DEFINITION 5.4. A subspace Z of a Banach space X is called proximinal if for
every :r E X, there exists Zo E Z such that Ilx - zoll = d(J;, Z) = infzEz Ilx - zll
The map PZ(J:) = {zo E Z : Ilx - zoll = infzEz lI:r - zll} is called the metric
projection.
PROPOSITION 5.5. Let Z ~ Y ~ X, Z proximinal in X.
(a) Let A be a family of subsets of Y / Z. Let A' be a family of subsets of
Y s'ltch that for any :1; E X and A E A. there exists A' E A' such that for
any a + Z E A. {a + Pz(x - a)} n A' i=- 0. S-uppose Y is a A'-C -subspace
of X. Then Y/Z 'is a A-C-subspace of X/Z.
Let S be any of the families .1', l3 or P.
(b) IfY is a S(Y)-C-subspace of X. then Y/Z is a S(Y/Z)-C-8'ubspace of
X/Z.
(c) Suppose the metric projection has a continuous selection. Then. if Y
is a K(Y)-C-subspace of X, Y/Z is a K(Y/Z)-C-subspace of X/Z.
(d) Let Z ~ Y ~ X*, Z w*-closed in X*. 11' Y is a S(Y)-C-subspace
of X*, then Y/Z is a S(Y/Z)-C-subspace of X* /Z. and hence. has the
S(Y/Z)-IP.
(e) Let X have the S(X)-IP. Let l'd ~ X be a reflexive subspace. Then
X/M has the S(X/l'd)-IP.
PROOF. (a). Let A E A and x + Z E X/Z. Choose A' as above. Then,
for a + Z E A, there exists z E Pz(x - a) (depending on .1: and a) such that
a + zEA'. Since Y is a A' -C-subspace of X, there exists Yo E Y snch that
Ilyo - a - zll ~ Ilx - a - zll for all a + Z E A. Clearly then Ilyo - a + ZII ~
Ilyo - a - zll ~ 11:1; - a - zll = IIx - a + ZII
CHEBYSHEV CENTRES AND INTERSECTION PROPERTIES OF BALLS 55

If S is the family under consideration in (b) and (c) above and A = S(Y/Z),
then for any choice of A' as above, S(Y) ~ A'. Hence, (b) and (c) follows from
(a). For (d), we simply observe that any w*-closed subspace of a dual space is
proximinal. And (e) follows from (d). 0

As in [3, Corollary 4.6], we observe


PROPOSITION 5.6. Let Z ~ Y ~ X, Z proximinal in Y and Y is a semi-L-
summand in X. Then Y/Z is a P-C-subspace of X/Z.
Let us now consider the Co or fp sums.
THEOREM 5.7. Let r be an index set. For all 0: E r, let Ya: be a subspace of
Xa:. Let X and Y denote resp. the Co or fp (1 :::; p :::; 00) sum of Xa: 's and Ya: 'so
(a) For each 0: E r, let Aa: be a family of subsets ofYa: such that {O} E Aa:
and for any A E Aa:. there exists B E Aa: such that Au {O} ~ B.
Let A be a family of subsets of Y such that for any 0: E r, the 0:-
section of any A E A belongs to Aa:.
Then Y is an A-C-subspace of X if and only if for each 0: E r, Ya: is
an Aa:-C-subspace of Xa:.
Let S be any of the families F, 1C, B or P.
(b) Y is a S(Y)-C-subspace of X if and only if for any 0: E r, Ya: is a
S(Ya:)-C-subspace of Xa:.
(c) The S-IP is stable under fp-sums (1 :::; p:::; 00).
PROOF. (a). The proof is very similar that of to [3, Theorem 4.7]. We omit
the details.
(c). Xa: has S-IP if and only if Xa: is a S-C-subspace of some dual space Y';.
Now the fp-sum (1 :::; P :::; 00) of Y';'s is a dual space. 0
REMARK 5.8. The result for F-IP has already been noted by [18] with a much
different proof. The stability of the P-IP under f1-sums is noted in [15] again with
a different proof.
[18] also notes that F-IP is stable under co-sum. And Corollary 4.3 shows
that Co has the IC-IP. However, we do not know if the IC-IP is stable under Co-
sums. As for the B-IP or P-IP, we now show that co-sum of any infinite family of
Banach spaces lacks the B-IP, and therefore, also the P-IP. This is quite similar to
Example 2.9.
PROPOSITION 5.9. Let r be an infinite index set. For any family of Banach
spaces Xa:, 0: E r, X = fficoXa: lacks the B-IP.
PROOF. For each 0: E r, let Xa: be an unit vector in Xa: and define ea: E X by
if f3 = 0:
otherwise
Then the set A = {ea: : 0: E r} is bounded and the balls Bx--[ea:, 1/2] intersect at
the point (1/2xa:) E X**, but the balls Bx[ea:, 1/2] cannot intersect in X. 0

REMARK 5.10. As before, taking AU{O}, it follows that X lacks the A-IP even
for A = weakly compact sets.
Coming to function spaces, we note the following general result.
56 BANDYOPADHYAY AND DUTTA

PRDPOSITION 5.11. Let Y be a subspace of a Banach space X and A be a family


of subsets of Y.
(a) For any topological space T, if C(T, Y) is a A-C-subspace of C(T, X),
then Y is a A-C -subspace of X. Moreover, if C(T, X) has A-IP, X has
A-IP.
(b) Let (n, L-, fl) be a probability space. If for some 1 ~ P < 00, LP(fl, Y) is
aA-C-subspace of LP(fl, X), then Y is aA-C-subspace ofX. Moreover,
if LP(fl, X), has A-IP, then X has A-IP.
PROOF. For (a) and (b), let F(X) denote the corresponding space of functions
and identify X with the constant functions. In (a), point evaluation and in (b),
integral over n gives us a norm 1 projection from F(X) onto X. Thus X inherits
A-IP from F(X).
Now suppose F(Y) is an A-C-subspace of F(X). Let P : F(Y) -+ Y be the
above norm 1 projection. Let x E X and A E A. Then, there exists g E F(Y) such
that IIg-ali ~ IIx-all for all a E A. Let y = Pg. Then, lIy-all ~ IIg-all ~ IIx-all
for all a E A. 0
The following Proposition was proved in [3].
PROPOSITION 5.12. (a) Let X has Radon Nikodym Property and is 1-
complemented in Z* for some Banach space Z. Then for 1 < p < 00,
LP(IL,X) is 1-complemented in U(fl, Y)* (l/p+ l/q = 1), and hence has
the 'P-IP.
(b) Suppose X is separable and 1-complemented in X** by a projection P
that is w*-w universally measurable. Then for 1 ~ P < 00 LP(fl, X) is
1-complemented in Lq(fl,X*)* (l/p+ l/q = 1), and hence has the 'P-IP.
Since the B-IP or 'P-IP is inherited by I-complemented subspaces and Co lacks
the B-IP, the next result follows essentially from the arguments of [17].
PROPOSITION 5.13. (a) Let X be a Banach space containing Co and
let Y be any infinite dimensional Banach space. Then X 181" Y fails the
B-IP and 'P-IP.
(b) If C(K, X) has the B-IP, then either K is finite or X is finite dimen-
sional. C(K,X) has the 'P-IP if and only if either (i) K is finite and X
has the 'P-IP or (ii) X is finite dimensional and K is extremally discon-
nected.
(c) For any nonatomic measure space (n,L-,/L) and a Banach space X
containing co, 1 (fl, X) fails the B-IP.
In the next Proposition, we prove a partial converse of Proposition 5.11(a)
when Y is finite dimensional and K is compact and extremally disconnected.
PROPOSITION 5.14. LetS be any ofthefamiliesF, IC, B or'P. LetY be a finite
dimensional a S(Y)-C-subspace of a Banach space X. Then for any extremally
disconnected compact space K, C(K, Y) is a S(C(K, Y))-C-subspace of C(K, X).
PROOF. We argue similar to the proof of [3, Proposition 4.11]. Let K be
homeomorphically embedded in the Stone-Cech compactification .B(r) of a discrete
set r and let : .B(r) -+ K be a continuous retract. Let A E S(C(K, Y)) and
g E C(K, X). Note that since Y is finite dimensional, by the defining property
of .B(r), any Y-valued bounded function on r has a norm preserving extension
CHEBYSHEV CENTRES AND INTERSECTION PROPERTIES OF BALLS 57

in C(,B(r), Y). Thus C(,B(r), Y) can be identified with EBeoc(r) Y. Lift A to this
space. In view of Theorem 5.7, this space is S(Y)-C-subspace of EBooX. This latter
space contains C(,B(r) , X). Thus by composing the functions with 4>, we get a
f E C(K, Y) such that Ilf - hll ~ Ilg - hll for all h E A. Hence the result. 0

And now for a partial converse of Proposition 5.11(b).

THEOREM 5.15. Let Y be a separable subspace of x. IfY is a P-C-subspace


of X, then for any standard Borel space 0 and any a-finite measure jl, Lp(jl, Y) is
a P-C-subspace of Lp(jl,X).
PROOF. Let f E Lp(jl, X). Since Y is a P-C-subspace of X, for each x E X,
nyEyBy[y, IIx - yll] =1= 0.
Define a multi-valued map F : 0 ----+ Y, by

F(t) = { nyEy By[y, Ilf(t) - ylll if f(t) EX \ Y


{f(t)} if f(t) E Y

Let G = {(t, z) : z E F(t)} be the graph of F.


Claim: G is a measurable subset of 0 x Y.
To establish the claim, we show that GC is measurable. Since Y is separable,
let {Yn} be a countable dense set in Y. Observe that z ~ F(t) if and only if
either f(t) E Y and z =1= f(t) or f(t) E X \ Y and there exists Yn such that
liz - Ynll > IIf(t) - Ynll And hence,

G C = {f(t) E Y and z =1= f(t)} u U {f(t) EX \ Y and liz - Ynll > Ilf(t) - Ynll}
n;:::l

is a measurable set.
By von Neumann selection theorem, there is a measurable function 9 : 0 ----+ Y
such that (t,g(t)) E G for almost all tEO.
Observe that Ilg(t)11 ~ IIf(t)1I for almost all t. Hence 9 E Lp(jl, Y). Also for
any h E Lp(jl, Y) we have Ilg(t) - h(t)11 ~ Ilf(t) - h(t)1I for almost all t. Thus,
Ilg - hll p ~ IIf - hll p for all h E Lp(jl, Y). 0

QUESTION 5.16. Suppose Y is a separable lC-C-subspace of X. Let (0, ~,jl) be


a probability space. Is LP(jl, Y) a lC-C -subspace of LP(jl, X)?
REMARK 5.17. This question was answered in positive in [3] for F-C-subspaces
and we did it for P-C-subspaces. Both the proofs are applications of von Neumann
selection Theorem. The problem here is for a compact set A in LP(jl, Y) and wE 0
the set {f(w) : f E A} need not be compact in Y.

ACKNOWLEDGEMENTS. Partially supported by a DST-NSF grant no.


RP041/2000.
The first-named author availed this grant to visit Southern Illinois University
at Edwardsville, USA in May-June 2002 and attended the Fourth Conference on
Function Spaces, where he presented a talk based on this work. He would like to
thank Professor K. Jarosz for the warm hospitality and a wonderful conference.
We also thank the referee for suggestions that improved the paper.
58 BANDYOPADHYAY AND DUTTA

References
[1] P. Bandyopadhayay, S. Basu, S. Dutta and B. L. Lin Very nonconstmined subspaces of
Banach spaces, Preprint 2002.
[2] P. Bandyopadhayay and S. Dutta, Almost constmined subspaces of Banach spaces, Preprint
2002.
[3] Pradipta Bandyopadhyay and T. S. S. R. K. Rao, Centml subspaces of Banach spaces, J.
Approx. Theory, 103 (2000),206-222.
[4] B. Beuzamy and B. Maurey, Points minimaux et ensembles optimaux dans les espaces de
Banach, J. Functional Analysis, 24 (1977), 107-139.
[5] J. Diestel, Geometry of Banach Spaces, selected topics, Lecture notes in Mathematics,
Vo1.485, Springer-Verlag (1975).
[6] J. Diestel and J. J. Uhl, Jr., Vector' measures, Mathematical Surveys, No. 15, Amer. Math.
Soc., Providence, R. 1. (1977).
[7] G. Godefroy, Existence and uniqueness of isometric preduals: a survey, Banach space theory
(Iowa City, lA, 1987), 131-193, Contemp. Math., 85, Amer. Math. Soc., Providence, RI, 1989.
[8] G. Godefroy and N. J. Kalton, The ball topology and its applications, Banach space theory
(Iowa City, lA, 1987), 195-237, Contemp. Math., 85, Amer. Math. Soc., Providence, RI, 1989.
[9] G. Godini On minimal points, Comment. Math. Univ. Carotin., 21 (1980), 407-419.
[10] P. Harmand D. Werner and W. Werner, M-ideals in Banach spaces and Banach algebras,
Lecture Notes in Mathematics, 1547, Springer-Verlag, Berlin, 1993.
[11] O. Hustad, Intersection properties of balls in complex Banach spaces whose duals are L1
spaces, Acta Math., 132 (1974), 283-313.
[12] H. E. Lacey, Isometric theory of classical Banach spaces, Die Grundlehren der mathematis-
chen Wissenschaften, Band 208, Springer-Verlag, New York-Heidelberg, 1974.
[13] J. Lindenstrauss, Extension of compact opemtors, Mem. Amer. Math. Soc., No. 48, 1964.
[14] A. Lima, Complex Banach spaces whose duals are L1-spaces, Israel J. Math., 24 (1976),
59-72.
[15] T. S. S. R. K. Rao, Intersection properties of balls in tensor products of some Banach spaces
II, Indian J. Pure Appl. Math., 21 (1990),275-284.
[16] T. S. S. R. K. Rao, On ideals in Banach spaces, Rocky Mountain J. Math., 31 (2001),
595-609.
[17] T. S. S. R. K. Rao Chebyshev centers and centmble sets, Proc. Amer. Math. Soc., 130 (2002),
2593-2598.
[18] L. Vesely, Genemlized centers of finite sets in Banach spaces, Acta Math. Univ. Comen., 66
(1997),83-115.

(Pradipta Bandyopadhyay) STAT-MATH DIVISION, INDIAN STATISTICAL INSTITUTE, 203, B.


T. ROAD, KOLKATA 700 108, INDIA, E-mail: pradiptaClisical.ac.in

(S Dutta) STAT-MATH DIVISION, INDIAN STATISTICAL INSTITUTE, 203, B. T. ROAD, KOLKATA


700 108, INDIA, E-mail: sudipta.rGlisical.ac . in
Contemporary Mathematics
Volume 328, 2003

The boundary of the unit ball in HI-type spaces

Paul Beneker and Jan Wiegerinck

ABSTRACT. This is a survey on the extreme, exposed and strongly exposed


boundary points of the unit baH in various Hl-type spaces.

1. Introduction
Theorems like the Krein-Milman theorem and Phelps' theorem assert the exis-
tence of many extreme points in certain convex sets. However, to decide whether
a particular point in the boundary of a convex set is extreme, exposed or strongly
exposed is quite a different question. In this paper we survey what is known in the
very special situation of the unit ball of certain Hardy spaces. Of course we refer
to the literature for the majority of the proofs. But we have included some, and
sketched others, hoping that this will clarify the exposition.
After introducing extreme, exposed and strongly exposed points, we end the
introduction with definitions of the Hardy-spaces in which we study our problem.
These are: the standard Hl(]I))) of the disc ]I)) in e, HI(0.) of a domain of finite
connectivity 0. c e, the Bergman space Al(]I))) of the disc and Hl(lB n ) of the unit
ball in en.
In Section 2 we will study the boundary points of the unit ball of HI (]I))). Here
good function theoretic descriptions of extreme and strongly exposed points are
possible. Exposed points seem to escape such a description. The ball of HI (0.)
is the topic of Section 3. We will see that results for the disc in general do not
extend to this case. In Section 4 we will turn to the Bergman space. Now all
boundary points of the unit ball are exposed and many well-behaved functions, like
polynomials, are strongly exposed. However, there are boundary points that are
not strongly exposed. The final section is devoted to the little we know about the
higher dimensional situation.
The paper borrows heavily from the expository parts of the PhD-thesis of the
first author, [3].

1.1. Generalities. Let X be a Banach space with (only in this section) real
dual space X* and let A be a bounded closed subset of X. We recall some notions
that are connected with convexity properties at a boundary point of A.

Research of the first author was supported by the Dutch research organization NWO.

2003 American Mathematical Society


59
60 PAUL BENEKER AND JAN WIEGERINCK

DEFINITION 1.1. A point x E A is called extreme if x = tp + (1 - t)q for some


p, q E A and 0 < t < 1 implies that p = q = x.
A point x E A is called exposed if there exists L E X* such that Lx = 1 while
Ly < 1 for every yEA \ {x}.
The functional L is called an exposing functional for x.

For L E X* we set

(1.1) M(L, A) = sup{Lx : x E A}

and

(1.2) 1\1(A) = sup{llxll : x E A}.

If IILII = 1 and a > 0 we define a slice of A as

(1.3) S(L,a,A) = {x E A: Lx> 1\1(L, A) - a}.


Let B be the closed unit ball of X.

DEFINITION 1.2. A subset A of X is called dentable if for every c > 0 there is


a point x E A such that x is not in the norm-closed convex hull of A \ (x + cB).
Equivalently, A is dent able if for every c > 0 there is a slice S(L, a, A) of diameter
less than c.

DEFINITION 1.3. A point x E A c X is called a denting point if for every c > 0


there exists a slice S(L, a, A) of diameter less than c that contains x. A point
:1: E A c X is called a strongly exposed point if there exists L E X* such that for
every c > 0 there exists a > 0 such that the slice S(L, a, A) contains x and has
diameter less than c. In other words, there exists L E X* such that Lx = 1\1(L, A)
and if LX n ---+ 1\1(L, A) for {x n } C A, then Xn ---+ x in the norm of X.

With these definitions at hand we make some observations. A strongly exposed


point in A is exposed. Normalizing the functional L in the previous definition will
yield an exposing functional for x. An exposed point of A is extreme and an extreme
point of A belongs to the boundary of A.
A theorem of Phelps, [38], states that if every bounded subset of X is dentable,
then every bounded closed convex set is the convex hull of its strongly exposed
points. It is known that separable dual spaces have the property that every bounded
set is dentable, cf. [10], Chapter 6. However, neither Phelps' theorem nor its proof
gives us any information about, say, which of the exposed points of a convex set C
are strongly exposed points.
In the rest of the paper it will be more convenient to use an equivalent, more
analytic definition of strongly exposed point: An exposed point x of the set A C X
with exposing functional L is strongly exposed if and only if every sequence (xn)in
A with the property that LX n ---+ 1 converges strongly to x. Moreover, we will work
over C, so that the definition of exposed point becomes slightly adapted. A point
x E A is an exposed point of A, if there exists a (complex) functional L on X with
the following property: Lx = 1 and for all yEA, Y =1= x ~ Re L(y) < 1. (For
A E C, Re A denotes the real part of A.) Again we say that the functional L is an
exposing fUIlctional for x, or simply that L exposes the point x.
THE BOUNDARY OF THE UNIT BALL IN HI_TYPE SPACES 61

1.2. Hardy spaces. Next we briefly recall the definition and some properties
of Hardy spaces. Excellent introductions are [12, 16, 24], and for domains of finite
connectivity [13]. We will denote the space of holomorphic functions on a domain
D c en by H(D).
Let !Dl be the unit disc in e and 'JI' its boundary.
DEFINITION 1.4. Let 0 < p < 00. The Hardy space HP = HP(!Dl) consists of all
holomorphic functions I on !Dl for which
dO
1
2~
11111it-v:= sup II(re i9 W - < 00.
OCrCl 0 2w
The Hardy space HOC(!Dl) consists of the bounded functions in H(!Dl). It is a
closed subspace of LOC(!Dl) under the sup-norm 11.lIoc.
If IE HP(!Dl), then
(1.4)

exists for almost allO. The function 1* belongs to U('JI'). For alII::; p::; 00, II.IIHP
is a norm, IIIIIHP = 111*IILp and HP(!Dl) becomes a Banach space that is a subspace
of U('JI'). We will simply write 1IIIIp instead of 11111Hv.
Alternatively, the Hardy space HP(!Dl) (1::; P < 00) may be described by
HP(!Dl) = {f E H(!Dl) : IIIP admits a harmonic majorant}
= closure of holomorphic polynomials in P ('JI').
The value at 0 of the least harmonic majorant of IIIP,
uf(z) = inf{h(z) : h harmonic and h> IIIP}
equals the norm, 1IIIIp = u(O?/p. Equivalent norms are given by (uf(z)?/P, (z E
!Dl) .
These descriptions are useful for defining HP-type spaces on a domain n in e:
HP(n) = {f E H(n) : IIIP admits a harmonic majorant}.
Norms, all equivalent, can be defined via the least harmonic majorant uf by
(Uf(Z))l/P, (z En). If n has smooth boundary an then HP(n) is again a closed
subspace of U(an).
Similarly on the unit ball ~n of (n ~ 2):en
HP(~n) = closure of holomorphic polynomials in LP(a~n).

We end this section by defining the Bergman space AP(!Dl). Let 1 ::; p < 00. If
IE H(!Dl) is such that

1111I~p:= i'D[11(zW dxdy


W
< 00,
then we say that I belongs to the Bergman space AP(!Dl). In other words, AP(!Dl) =
H(!Dl) n U(!Dl). It is a closed subspace of U(!Dl).
One motivation for looking at the Bergman space Al(!Dl) is that
Al(!Dl) ~ {f E Hl(~2): I(Zl,Z2) = I(Zl'O)}.
Therefore B(Al(!Dl)) may serve as a testing ground for B(H1(!Dl)).
62 PAUL BENEKER AND JAN WIEGERINCK

2. The boundary of unit ball in H1(]])))


2.1. Extreme points. The subject started with a famous theorem of deLeeuw
and Rudin, [26], that characterizes the extreme points of B(H 1 (]])))). It will be
presented below.
Recall the inner-outer decomposition for f in H1(]]))), cf. [16]:
f(z) = I(z)U(z).
Here I E H(]]))) is an inner function, that is, I E HOC(]]))) with 1I*(ei9 )1 = 1 a.e.
and U E H(]]))) is an outer function, i.e. log IUI(z) = P[log IUI](z), where P is the
Poisson operator:

Pu z =
[ ]()
1 0
211" 1 - Izl2
1 - 2Re (ze- i9 ) + Izl2 u (e
t9
. d()
-.
) 27r
The inner factor thus contains all the zeros of f. In fact, every inner factor is
of the form
(2.1) I(z) = B(z)S(z),
with B a Blaschke product and S is the singular factor

(2.2)
where J.L is a positive, singular measure on 'JI' and - denotes harmonic conjugation.
EXAMPLES 2.1. Outer functions are zero free, and e.g. a holomorphic function
on]])) that is continuous and away from 0 on iij is outer. Taking a root and applying a
limit argument shows then that f E HI (]]))) that assumes values only in C \ (-00,0)
is outer too.
EXAMPLE 2.2. If I is an inner function then (1 + 1)2 is outer.
DeLeeuw and Rudin proved
THEOREM 2.3 ([26]). A function f is an extreme point of the unit ball of HI (]])))
if and only if f is an outer function and IIfl11 = 1.
PROOF. Suppose f is an outer function of unit norm and suppose g in H1 is
such that IIf + gill = Ilf - gill = 1. Let dJ.L be the probability measure If I on the g!
unit circle. Then, with k = gl f on 'JI', the relations Ilf + gill = Ilf - gill = 1 imply
that J;11" 11 + kl + 11 - kldJ.L = 2. Because for all z E C: 11 + zl + 11 - zl 2 2, with
equality if and only if z E [-1,1], we conclude that for almost dJ.L-every ~ E 'JI' :
g(~)1 f(O E [-1,1]. Observe that this inclusion also holds for almost d()-every ~ E 'JI',
because ~ = If I i:- 0 almost d()-everywhere on 'JI'. In particular, Igl :::; If I d()- a.e.
on 'JI'. By the fact that f is outer, then also for all zED: Ig(z)1 :::; If(z)l. Hence
gI f E HOC has real boundary values on 'JI'. The Poisson integral representation of
glf yields that glf is constant. Finally, because Ilf + gill = 1, this constant is
zero, i.e., g == 0, so f is extreme.
In the other direction, suppose f = I F is of unit norm, but I is a non-trivial
inner function. Because for every ei9 E 'JI' \ {1}:
(1 ei9 )2 .
---'-----:-.,,::--'--- = 2 2Re (e t9 ) = 2 2 cos( ()) > 0,
e t17
THE BOUNDARY OF THE UNIT BALL IN Hi_TYPE SPACES 63

the functions I and (1 I)2 have the same argument a.e. on'lI'. Let 9 = lt .F.
We have
dB dB
III gill =
h[2'" 1F1(1 Re (I)) 2w = 1 Re (h[2'" IFI' I ~ ).
Now if we replace I by AI throughout the preceding, where A E 'lI' is arbitrary, then
we obtain:
II/glll=IRe(A
o
2",
1 dB
IFII-).
2w
Therefore, we can choose A in such a way that Re (A J:'" IFI . I g!) = o. Conse-
quently, III gill = 1, but 9 "= 0, so I is not extreme.
D

EXAMPLE 2.4. Normalized polynomials without zeros in j[)) are extreme points.
2.2. Exposed points in B(Hl(j[)))), function theoretic methods. As far
as the authors know, there is no clearcut function theoretic characterization of
exposed points of BHl(j[)))). The necessary and sufficient condition of Helson,
[21] given below in Theorem 2.8 probably comes closest.
The following lemma identifies the exposing functionals.
LEMMA 2.5. II I is exposed in B(Hl(j[)))), then

Lf: 9 ~ J7
9 dB
UT2w
is the unique exposing functional.
PROOF. Suppose I is an exposed point of B(HI(j[)))) with exposing functional
L. By the Hahn-Banach theorem there exists a function cp E V"', Ilcplloo=l, which
represents the action of L:
[21< dB
L(g) = io
gcp 2w'
Because L(f) = 11/111 = 1, and because the function I has mass (almost) everywhere
on 'lI', there is only one cp that has the desired properties, namely cp = 711/1 =
exp( -i arg(f)). In particular we see that the exposing functional for I is unique
and of the above form. D
DEFINITION 2.6. A function I E HI (j[))) is called rigid if 9 E HI (j[))) and arg 9 =
arg I implies that 9 = cl for some c > O.
An HI-function is rigid if its argument on 'lI' determines the function up to a
multiplicative constant. By the previous lemma exposed points are precisely the
rigid functions of norm 1.
Not all outer functions are rigid as the following example demonstrates.
EXAMPLE 2.7. By the DeLeeuw-Rudin theorem (Theorem 2.3), no HI-function
with non-trivial inner factor is rigid. Inspection of the proof of the theorem explic-
itly gives another HI-function with the same argument: if I = I F with F outer,
then 9 = (1 + 1)2. F is an outer function, cf. Example 2.2 that has the same ar-
gument as f. This argument can be reversed: suppose I "= 0 is divisible in HI
by the outer function (1 + I)2, where I is any non-trivial inner function, then I
64 PAUL BENEKER AND JAN WIEGERINCK

and I I / (1 + I) 2 have the same argument, so I is not rigid. In other words, for all
non-trivial inner functions I we have
(2.3) If I is rigid, then 1/(1 + 1)2 rt HI
Nakazi [28, 29] conjectured that an outer function with fI(l + cz)2 rt HI
for every c E 'JI' must be rigid. E. Hayashi [17] gave a counter example. Later
Sarason [44, 45] conjectured that (2.3) characterizes rigid functions. Inoue [22]
subsequently gave an example of a non-rigid function I E HI with the property
that 1/(1 + 1)2 rt HI for all non-trivial inner functions I.
Helson showed that a variant of (2.3) indeed characterizes rigid functions;
THEOREM 2.8 ([21]). Let I be an outer lunction in HI. Then I is rigid il and
only il lor all inner lunctions p and q, not both constants, I/(p + q)2 rt HI.
The drawback of Helson's theorem is that his condition is difficult to check in
practice. More user friendly are the following results of Yabuta.
THEOREM 2.9. (1) II I E HI is invertible in HI, then I is rigid. [51]
(2) Suppose ther'e exists h E Hoo \ {O} such that Re(h!) 2: 0 a.e. on 'JI'. Then
I is rigid. [52]
The proof uses a result of Neuwirth and Newman [36], that is interesting in its
own right.
LEMMA 2.10 ([20],[36]). Every lunction I E H 1 / 2 that is positive a.e. on'JI' is
constant.
Nakazi [29] gives another proof of Yabuta's results.
EXAMPLE 2.11 ([23]). Normalized polynomials P without zeros on II)) and at
most simple zeros on 'JI' are exposed. Indeed, the argument of such a function makes
jumps of height 7r at the zeros in 'JI', hence by adding a smooth function <p to arg P,
we can assure that the sum is in (-7r /2, 7r /2). Hence multiplying P by the analytic
continuation of e-<i>+iv> yields a function with positive real part and the second of
Yabuta's criteria applies.
2.3. Strongly exposed points. Whereas it seems impossible to give a char-
acterization of exposed points that allows to check if a given function in aB(HI (II))))
is an exposed point, it is possible to characterize the strongly exposed points. This
is the content of work of Nakazi and, independently, the first author. Let us describe
strongly exposed points in terms of properties of the exposing functional.
THEOREM 2.12 ([2,4]). II alunction IE 8B(H 1 (1I)))) is strongly exposed, then
d( Tfr, HOO) < 1. (d = Loo-distance).
PROOF. Suppose that the LOO-distance of <p to HOO equals 1 and that I is
exposed. Let L' be the restriction of the I-exposing functional L(g) = J~1r g<p g!
to HJ = zHI. By the Hahn-Banach theorem, the operator norm of L' equals the
Loo-distance of <p to HOO ~ (HJ)1.. Thus we can find a sequence offunctions (fn) in
the unit ball of HJ such that L(fn) --+ IIL'II = 1. The functions In do not converge
to I in norm, because norm convergence implies convergence at 0, In(O) = 0 and
1(0) f= 0 because I is outer. Therefore I is not strongly exposed. D
The space HOO + G('JI') that appears in the next theorem, has been studied
extensively, see [42] and cf. [16]. It is a closed subalgebra of LOO('JI').
THE BOUNDARY OF THE UNIT BALL IN HI_TYPE SPACES 65

THEOREM 2.13 ([47]). Let IE B(Hl(]])))) be exposed. Ild( &!. Hoo+C(1I')) < 1,
then I is strongly exposed.
We will explain how the proof goes if 'P := 111/1 E Hoo + C(1I'). The exposing
functional L for I is defined by L(g) = fg 7r
g'Pg!. Suppose that (fn)f is a sequence
in the unit ball of Hl such that L(fn) --) 1 as n --) 00. Clearly this implies that
II In I 1 --> 1. We claim that there is a subsequence (fnk) that converges to I in HI.
By weak* compactness of the unit ball of (C(1I'))* the sequence of measures Ing!
has a weak* convergent subsequence Ink g! with limit dJ.L({}) , and IIJ.LII :::; 1. For all
n = 1,2,3, ... , we have Jo r 27r em
. (J
dJ.L({}) = O. Now the F. & M. Riesz theorem [16]
precisely states that J.L is of the form dJ.L( (}) = F g! for some F in the unit ball of HI .
Write 'P = 'Pl + 'P2 with 'Pl E H[J = zHoo and 'P2 E C. Because H[J annihilates
HI, the weak* convergence of Ink g! implies that

We conclude that F = I because I is exposed. Again, by weak* convergence Ink


converges pointwise to I. The following result of D.J. Newman ensures that the
convergence is in Hl with limit I. The argument shows that every subsequence of
(fn) has a subsubsequence that converges in HI to I, and therefore In --) I in HI.
PROPOSITION 2.14 ([35]). Let (fn)i be a sequence 01 functions in the unit ball
01 Hl. Suppose In converyes pointwise to I E Hl on ]])) and I In I 1 --> 11/111 as
n --> 00. Then In converges to I in HI-norm.
The method for the general case in [47] is to work in the maximal ideal space
of LOO(1I') and use the Helson-Lowdenschlager generalization of the F. & M. Riesz
theorem. Cf. [30] for a related approach.

2.4. Strong exposedness and Helson-Szego weights. In this section we


will explicitly describe the strongly exposed points in the unit ball of HI: they are
the outer functions induced by so-called Helson-Szego weights on the unit circle. We
will discuss three different (albeit related) roads to this result. The first one quickly
establishes that strongly exposed points are induced by Helson-Szego weights and
uses little background on exposed points and Hardy space theory. The second
one uses function theory to show that Helson-Szego weights give rise to strongly
exposed points and is essentially due to T. Nakazi [31]. The third approach (to
be discussed in Section 2.5) uses operator theory and the relation between exposed
points and Toeplitz operators [2] to prove that strongly exposed points and Helson-
Szego weights are "the same". These last two proofs use the characterization of
strong exposedness obtained in the previous section (Theorem 2.13). To put the re-
sults in perspective, we will mention some surprising properties of strongly exposed
points.

Let J.L be a finite (positive) measure on 1I'. Let P be the collection of polynomials
in z and let Q be the space of polynomials in z vanishing at the origin. On the unit
66 PAUL BENEKER AND JAN WIEGERINCK

circle P and Q consist of the trigonometric polynomials of the form Ln>o ane inlJ
and Ln<o ane inlJ , respectively. Finally, let P = p(/-L) be the number -

p:= sup I iPqd/-Ll,


where the supremum is taken over all pEP and q E Q restricted by Ilpll2(/J) ~ 1
and Ilqll2(/J) ~ 1. By Cauchy-Schwarz, 0 ~ P ~ 1. We see that the spaces P and
Q are orthogonal if and only if P = O. On the other hand, when the L 2 (/-L)-closures
P of P and Q of Q have a non-trivial intersection, then clearly p = 1. The size of
the number p is related to the question when the sum P + Q is closed in L 2 (/-L).
Namely, when p < 1, then for all p E P, q E Q in L 2 (/-L)-norm,
lip + ql12 ~ IIpl12 + IIqll2 - 2p Ilpll Ilqll ~ (1- p)(llpl12 + IlqIl2),
which implies that P + Q is closed in When p < 1 we say that P and Q
L 2 (/-L).
are at the positive angle cos-I(p) > O. When the spaces P and Q are at positive
angle, the projection
N N
P+(Lane inlJ ) = LaneinlJ ,
-N 0
which is densely defined on L 2 (/-L),
extends to a bounded operator on L 2 (/-L). Con-
versely, if the intersection of P and Q in L 2 (/-L) is trivial, then the definition of
P+ acting on trigonometric polynomials is well-defined and extends to a bounded
operator on L 2 (/-L) if and only if P and Q are at positive angle.

DEFINITION 2.15. We say that a function w ~ 0 on 'll' is a Helson-Szego weight


(on 'll') ifthere exist real valued u,v E Loo('ll') with Ilvll oo < I such that w = eU+V.
(Here v is the boundary function of the harmonic conjugate of the Poisson integral
of v to JIll.)

The following theorem of H. Helson & G. Szego elegantly describes all measures
/-L for which P and Q are at positive angle.
THEOREM 2.16 ([19]). The subspaces P and Q are at positive angle in L 2 (/-L)
if and only if the measure /-L is of the form d/-L = wdO, for some Helson-Szego weight
won'll'.
COROLLARY 2.17. If the function f is strongly exposed in the unit ball of HI,
then If I is a Helson-Szego weight on'll'.
PROOF. Assume f is an exposed point, such that If I is not a Helson-Szego
weight. We will show that f is not strongly exposed. Let /-L be the probability
measure If I g!.By the theorem of Helson and Szego the spaces P and Q are at
zero angle (p = 1). Thus we can find sequences (Pn) and (qn) in the L 2(/-L)-unit
balls of P and Q respectively, such that

(2.4) 1 0
211'
Pnqn d/-L --+ 1,

as n --+ 00. For n = 1,2, ... , let fn be the HI-function Pnqnf. These functions are
contained in the unit ball of HI:
THE BOUNDARY OF THE UNIT BALL IN Hl_TYPE SPACES 67

If we set <p = 7Ilfl (as usual), then the f-exposing functional is given by

L(g) = 10
211" dO
g<p 27r'
Now (2.4) expresses that L(fn) ~ 1 as n ~ 00, yet the functions fn do not
converge to f in HI because fn(O) = qn(O) = O. We conclude that f is not strongly
exposed. D
Already we can mention a surprising property of strongly exposed points.
COROLLARY 2.18 ([2]). If the function f is strongly exposed in the unit ball of
HI, then for all small c > 0, the functions f and 1/ f are contained in H1+ e .
The corollary together with Theorem 2.13 show that if f E B(HI(][})) satisfies
the distance condition in Theorem 2.13, then f is strongly exposed if and only if
1/ f E HI. (It is worth noting that there exist exposed points f E HI such that 1/ f
is in no HP, p > 0, [37, 44].) Since f and l/f are outer functions, the corollary is
immediate from the following lemma (which is actually an exercise in [16]).
LEMMA 2.19. If the function w is a Helson-Szego weight on 'JI', then for all
small c > 0, we have w E 1+ e ('JI') and l/w E L1+e('JI').
LEMMA 2.20. (Cf. [16], Lemma IV. 3. 3, p. 148.) If'lf; is a measurable real
function, then the LOO-distance of e-i.p to Hoo is less than 1 if and only if there
exist c > 0 and h E H oo such that
7r
(2.5) Ihl 2:: c and I'lf; + arghl $ "2 - c (mod 27r)
almost everywhere on 'JI'.
THEOREM 2.21 ([2, 31]). Let f be an extreme point of the unit ball of HI.
Then f is strongly exposed if and only if If I is a Helson-Szego weight.
PROOF. ([31]) By Corollary 2.17 we need only to prove the backward impli-
cation. Let us assume that the outer function f is such that If I is a Helson-Szego
weight on 'JI', say, If I = exp(u + v), where Ilvll oo < l We remark that the function
is exposed by Theorem 1 and Lemma 2.19. Next, again using the fact that f is
outer, we notice that
f(z) = eu(z)+iu(z) . ev(z)-iv(z),
because the right hand side is an outer function with the appropriate absolute
values on 'JI'. We set <p = 7Ilfl. Then <p = e-i.p with'lf; = u - v. Now the function
h(z) = exp(u(z) + iu(z)) is invertible in HOO and has argument u a.e. on 'JI'. We
observe that condition (2.5) in Lemma 2.20 is satisfied by the pair of functions
<p = e-i.p and h E HOO. We conclude that the Loo-distance of <p to Hoo is less than
one. Therefore, by Theorem 2.13, the function f is strongly exposed. D
Another way of formulating Theorem 2.21 is as follows:
COROLLARY 2.22. An extreme point f of the unit ball of HI is strongly exposed
if and only if f can be written as the product gh, where g is invertible in Hoo and
hE HI satisfies I arg hi $ ~ - c a.e. on 'JI', for some c > O.
Recall Corollary 2.18: if f is strongly exposed then for all small c > 0, the
functions p+e and 1/ f1+ e are contained in HI. This result can be sharpened
somewhat using Theorem 2.21:
68 PAUL BENEKER AND JAN WIEGERINCK

PROPOSITION 2.23 ([2]). Let f be a strongly exposed point in the unit ball of
HI. Then for all sufficiently small c > 0, the (normalized) functions ce f1+E: and
del f1+E: of unit norm are again strongly exposed in the unit ball of HI.
The Proposition explains the following examples from [47].
EXAMPLES 2.24. A polynomial is rigid in HI if and only if its zeros on the
unit circle are single zeros (and it is zero-free on llJ), obviously). Any normalized
polynomial P with at least one single zero on '][' is not strongly exposed however,
because 1/ P HI
Let f#(z) be the extreme point c(1 + z) log2(1 + z) in the unit ball of HI.
Because f# is outer and because 1/lf#1 E LI, we conclude that f# is exposed
(Theorem 2.9. 1). However, 1/lf#1 L1+E:, so f# is not strongly exposed.
2.5. Toeplitz operators and De Branges-Rovnyak spaces. Let P+ be
the orthogonal projection of L2('][') onto H2:
00 00
P+(Lane inll ) = Lane inll ,
-00 0

and let P_ be the orthogonal projection of L2 onto (H2).l. = H~:


00 -1
P_(Lane inll ) = Laneinll .
-00 -00

DEFINITION 2.25. Given a bounded function ' E LOC the Toeplitz operator T",
is the bounded map T", : H2 _ H2 given by
T",(f) = P+('f).
We say that ' is the defining function of the Toeplitz operator T",.

We see that the norm of the Toeplitz operator T", is at most 1I,1I00. It is not
difficult to show that the norm of T", is in fact equal to 1I'1I00, but we will not
need this result. Also, it is a routine exercise to verify that the adjoint of T", is the
Toeplitz operator Ttji. Clearly, if' E Hoo, then T",(f) = 'f. Combining these two
observations we have the following result:
LEMMA 2.26. If cp or' is contained in H oo , then T-qiT", = TVi",.

Given a function ' E Loo, the Hankel operator H", (with defining function ')
is the bounded operator
H",(f) = 'f - T",(f) = (I - P+)(,f) = P_('f)
from H2 into (H2).l.. By the same reasoning the norm of H", is at most 1I,1I00. If
two functions cp and ' in Loo differ by an element of H OO , then the associated Hankel
operators coincide. Hence the operator norm of H", is at most Loo-dist(',HOO).
The basic fact about Hankel operators, due to Z. Nehari, is that equality holds:
THEOREM 2.27 ([34]). The operator norm of H", equals the Loo -distance of '
to HOO.
THE BOUNDARY OF THE UNIT BALL IN HI_TYPE SPACES 69

The following result (and its corollaries) will be of great importance. We only
have as reference Sarason's notes [43].
THEOREM 2.28. (Devinatz-Rabindranathan) If'IjJ is unimodular, then T1fJ
is left-invertible if and only if Loo-dist('IjJ,Hoo) < 1.
A bounded linear operator L : X -+ Y is said to be left-Fredholm if the range
of X under L is closed and of finite codimension in Y. For left-Fredholm operators
one has the following result.
COROLLARY 2.29 ([11]). (Douglas-Sarason) If'IjJ is unimodular, then T1fJ is
left-Fredholm if and only if the Loo -distance of'IjJ to HOO + C is less than 1.
See [43], p. 119 ff. also for a proof of the next result
COROLLARY 2.30 ([9, 49]). (Devinatz-Widom) If'IjJ is unimodular, then
T1fJ is invertible if and only if'IjJ can be written as ei(uH), where u and v are real
functions in Loo such that IIvll oc < 'IT".
We will now explain the relation between Toeplitz operators and rigidity (ex-
posedness) of functions in HI.
LEMMA 2.31. Let f be an outer function in HI. Then f is rigid if and only if
the Toeplitz operator with defining function <p = 7/1fl is injective.
PROOF. Suppose f is rigid and assume G E H2 is contained in the kernel of
T<p. Let us write f = F2 for some outer function F E H2. Then <p = F I F. The fact
that T<p (G) = 0 means that there exists a function H E H~ such that FG I F = H.
Unless H is the zero function, the HI-function GH(= FIF 'IHI2) (which vanishes
at the origin) has the same argument as f a.e. Because f was rigid we conclude
that H = 0 and hence G = O. This means that T<p is injective.
For the converse, suppose that the outer function f = F2 is not rigid. Then
there is an outer function g = G 2 in HI that has the same argument as f but is
not a scalar multiple of f. Because T<p(F) = F(O) and T<p(G) = P+(G) = G(O), the
function F - F(O)GIG(O) is a non-trivial element of the kernel of T<p. 0
Next we will see how one relates Toeplitz operators to De Branges-Rovnyak
spaces. These spaces were introduced and studied by L. De Branges and J. Rovnyak
in [6, 7]. D. Sarason used the spaces to describe exposed functions in the unit ball of
HI; Sarason's book [46] and his earlier articles [44, 45] provide a detailed account
of this remarkable theory. Borrowing Sarason's notation, we will review his ideas
and show how to adapt them to give another approach to the (explicit) description
of the strongly exposed points that was discussed in the previous section. In our
situation the De Branges-Rovnyak spaces are Hilbert subspaces H of H2 with
another inner product, such that the inclusion map i : H -+ H2 is contractive.
For this reason such spaces are also called contractive spaces. We can realize such
spaces as operator ranges of bounded maps on H2.

Now then, let a be a function in HOO that is not constant. The linear space
M (a) is the set of all functions af, f E H2. We equip M (a) with the inner product
, {af, ag)M := (f, g)2.
This makes M(a) a Hilbert space, and the Toeplitz operator Ta : H2 -+ M(a)
unitary. Clearly, when lIall oo ::; 1, the inclusion i : M(a) -+ H2 is a contraction.
70 PAUL BENEKER AND JAN WIEGERINCK

Let b be a non-constant function in the unit ball of Hoo. We will construct the
space 1t(b), the so-called complementary space to M(b), in a similar fashion; the
narile reflects the fact that M(b) + 1t(b) = H2 (although the intersection M(a) n
1t(b) is trivial only when b is an inner function). Observe that I - TbTj) is a positive
contraction on H2, hence the operator (I - nTj)l/2 is well-defined on H2 and
contractive. As a linear space 1t(b) consists of all H 2-functions in the range of the
operator (I - TbTj)l/2 on H2. We also use this map to give 1t(b) a Hilbert space
structure. Namely, if lor 9 is orthogonal in H2 to ker(I -TbTj)l/2 = ker(I -TbTj),
we set
((I - TbTj)l/2 I, (I - TbTj)l/2g)b:= (/,g)2.
As a consequence (I -TbTj)?/2 is a co-isometry from H2 onto 1t(b) and the inclusion
map i : 1t(b) --+ H2 is another contraction. Moreover, we see that if Ilbll oo < 1,
then 1t(b) is all of H2 with an equivalent norm.

Given an outer function I E Hl (not a constant) Sarason constructs three


auxiliary holomorphic functions:
27T ifJ
(2.6) G(z) = { e, + z I/(eifJ)1 dO,
io e,fJ - z 211'

(2.7) ( ) _ 2v'f(Z)
a z - G(z) + l'
G(z) - 1
(2.8) b( z) = G ( z) + 1 '
where we may take any branch of -IT Observe that we can recover the function
I from a and b: 1= F2, with F = a/(1 - b) E H2. The function G is reasonably
well-behaved on JI): because ReG(z) ~ I/(z)1 ~ 0, G is contained in HP for all
o < p < 1. The real part of G is positive and majorizes Jill.
hence a and bare
contained in the unit ball of Hoo. The fact that Re G = III a.e. on 'II' gives us
that lal 2 + IW = 1 a.e. on 'II'. By a theorem in [26] the functions a and b are not
extreme in the unit ball of Hoo. It is known in such cases that the polynomials are
contained and dense in 1t(b) (cf. [46], IV-3).

We shall also encounter Toeplitz operators on H2 with unbounded defining


functions. For a function 't/J E L2, the Toeplitz operator T,p (with defining function
't/J) acts as follows. Take a function 9 E H2, then 't/Jg is an L l ('II')-function with
Fourier series E~oo cn['t/Jg]einfJ and we set
00

T,p(g) = L en ['t/Jg]zn E H(D).


n=O
For bounded 't/J this coincides with the Toeplitz operators as defined earlier. For
unbounded 't/J, the map T,p is a densely defined operator from H2 to H2 and is
continuous from H2 into the space of holomorphic functions on JI) with the topology
of uniform convergence on compact subsets. In this light it is perhaps surprising
that the product of the operators Tl-b and Tp (the former is ordinary multiplication
by 1 - b) is bounded on H2. This is a consequence of the following theorem of
Sarason. The proof uses that G = ~ is the Herglotz integral of an absolutely
continuous measure (cf. (2.6)); it may be found in [46], IV-13, p. 30.
THE BOUNDARY OF THE UNIT BALL IN HI_TYPE SPACES 71

THEOREM 2.32. The operator T1-bTF is unitary from H2 onto 1i(b).


Next one can show that
TI-bTFTF/F = TI-bTFF/F = TI-bTF = Ta
(compare with Lemma 2.26). Therefore T1-bTF maps the range ofTF/F in H2 onto
M(a) C 1i(b) (Theorem 2.32). Consequently, M(a) is dense in 1i(b) if and only if
the range ofTF/F on H2 is dense in H2, or equivalently ifTF/F = T;/F is injective
on H2. Combining this with Lemma 2.31 we find Sarason's result:
THEOREM 2.33 ([44]). Let I be an extreme point 01 the unit ball in HI. Then
I is exposed il and only il M(a) is dense in 1i(b).
This fundamental result has several striking consequences, one of them being
that if I is exposed, then so are the squares of the H 2-functions FA = a/(1 - )"b)
().. E 1l'), cf. [44, 46]
As a next step one may even replace).. by any inner function u. Sarason proves
that the square of Fu = a/(I- ub) is exposed in HI.
We now formulate a variant on Theorem 2.33 that deals with strongly exposed
points.
THEOREM 2.34 ([2]). Let I be an extreme point 01 the unit ball 01 HI. Then
I is strongly exposed il and only il M(a) = 1i(b).
PROOF. First observe that for any FE H2 \ {O}, the Toeplitz operator TF/F
is injective. Sarason's reasoning preceding Theorem 2.33 shows that (with 1= F2
as before)
(2.9) M(a) = 1i(b) if and only if TF/F maps H2 onto H2.
Hence,
(2.10) M(a) = 1i(b) :} TF/F is invertible :} TF/F is invertible.

Suppose I is strongly exposed. By Theorem 2.13, the LOO-distance of F / F to


H OO + C is less than 1. The Douglas-Sarason theorem (Corollary 2.29) implies that
the operator TF/F has closed range in H2. As we have seen its adjoint TF/F is
injective, hence the range of TF/F is dense in H2 too. Thus TF/F : H2 -+ H2 is
surjective. The operator is also injective by the rigidity of I = F2 (Lemma 2.31).
By (2.10), M(a) = 1i(b).
To establish the other half of the theorem suppose M(a) = 1i(b). We observe
that I is exposed by Theorem 2.33. By (2.10), the operator TF/F is invertible, in
particular, it is left-invertible. Theorem 2.28 implies that the LOO-distance of 711/1
to HOC is less than 1. We conclude that I is strongly exposed by Theorem 2.13. 0
There is an alternative proof of the implication "I strongly exposed =* 1/1 E
HI" that lies on the surface: if I is strongly exposed, then 1 E 1i(b) = M(a),
because 1i(b) contains all polynomials. Thus l/a E H2 and 1/1= (l-b)2/ a2 E HI.
lt should come as no surprise that l/a is actually in HP for slightly larger p > 2.
Indeed, equality of the spaces M (a) and 1i(b) occurs if and only if the operator Ta/a
is invertible and the non-extreme pair (a, b) satisfies the so-called Corona condition:
inf
zED
la(zW + Ib(zW > 0,
72 PAUL BENEKER AND JAN WIEGERINCK

see [44] and [46], IX-5, p. 66. We see in particular that for strongly exposed f,
a2 Illall~ is also strongly exposed (cf. (2.10)), and that lla E H2+, 1/(I-b) E Hi+
for sufficiently small E > O.
Let us go back to the distance condition used in Theorem 2.13 and assume f
is an extreme point such that the LOO-distance of 7/1fl to HOO + C is less than 1.
Starting with the Douglas-Sarason theorem (Corollary 2.29) and exploiting the full
strength of the statement that TF/F is (left-)Fredholm it follows that the operator
TF/F has closed range in H2 of finite codimension, say, N. Hence M(a) is closed
and of finite co dimension N in 1i(b). Using [20], Theorem 6 or [46], X-18, p. 77,
we conclude that f is of the form f = p2g, where 9 is strongly exposed in the unit
ball of Hi and p is a polynomial of degree N that has all its zeros on 11'. Conversely,
if f is of the above form, then

distallfl, H oo
+ C) = dist(z2N . gllgl, H oo + C) = dist(gllgl, H oo + C) < 1,
by the algebra structure of Hoo + C, but dist(7/lfl,H OO ) = 1 when N > O.
This calculation serves to illustrate that if for a given extreme point f the
distance of 7Ilfl to H oo + C is less than 1, the ("only") thing that can prevent f
from being exposed (and hence strongly exposed) is the divisibility of f in Hi by
functions of the form (1 - U)2 with u(z) = >.z (>. E 11') a particularly simple inner
function. Functions in Hi that lack this divisibility property are called strong outer
functions. We see that for any strong outer function that is not exposed, like Inoue's
example, [22], it is true that the distance of 7/1fl to H OO +C is 1. Alternatively, the
strongly exposed points are the (normalized) strong outer functions f that satisfy
distallfl, H oo + C) < 1.

3. The boundary of B(Hi(O))


Having studied the sets of exposed and strongly exposed points in the unit
ball of the classical Hardy space Hi of the unit disc, we will investigate how these
results hold up for the Hardy space Hi of a domain of finite connectivity ("finite
domain") 0 c C. Such domains are e.g. conformally equivalent to domains with a
smooth boundary consisting of finitely many components or to a disc from which
a finite number of disjoint slits are deleted, cf. [33]. We defined Hi (0) in Section
1.2. There are two important differences with the classical Hardy space that make
the analysis very different.
On domains of finite connectivity, Hi-functions may not allow a classical fac-
torization using Blaschke products, (singular) inner functions and outer functions.
Indeed, the argument principle prevents the existence of a holomorphic function f
on the annulus A = {I < Izl < 2}, such that If I = 1 a.e. on 8A and f has exactly
one zero on A. Secondly, and in a way related, there now exist extreme points
in the unit ball with (finitely many) zeros (Section 3.1). While such zero sets are
somewhat generic for extreme points, their location plays a surprisingly crucial role
in its being a (strongly) exposed point (Section 3.2).
As in Section 2.1 we call a function I E HOO(O) an inner function if 11*1 = 1
almost everywhere with respect to arc length (dO') on 80. Consequently, II(z)1 ::; 1
for all z E O. Let w(, z) denote harmonic measure on 80 at z. We call a function
F E Hi(O) an outer function if for every z in 0 (equivalently, for at least one
THE BOUNDARY OF THE UNIT BALL IN Hl_TYPE SPACES 73

zEn):
log IF(z)1 = ( log IF*(~)I dw(~, z).
lao.
An outer function is zero free on n.
Green's function G(z; zo) and the Blaschke factor Bzo(z) = -1:1
t:z:oz for ]I))
are related by G(z; zo) = log IBzo(z)l. This suggests the following definition.
DEFINITION 3.1. Let G(z; zo) be Green's function on a finite domain n with
pole at Zo and let G(z; zo) denote its (multiple valued) harmonic conjugate. Suppose
that Zl, Z2, . .. satisfy the Blaschke condition En G(z; Zn) < 00. Then the Blaschke
product with zeros at Zl, Z2, ... is the multiple valued function

(3.1) B(z) = exp( - L G(z; zn)"- iL G(z; Zn.


n n
DEFINITION 3.2. A multiple valued function F on n is modulus automorphic if
IF(z)1 is well-defined;
locally on n, IFI coincides with the absolute value of a holomorphic func-
tion.
If in addition, IF(z)IP admits a harmonic majorant on n (0 < p < oo), then we say
that F E MHP(n}. If IFI is bounded, we say that F E MHOO(n}. I E MHOO(n} is
called innerif 11*1 = 1 a.e on r. I is singular inner if I is in addition zero-free.
Blaschke products are modulus holomorphic inner functions. One can show
that if FE MHP(n}, then IFI has non-tangential limits (denoted IF*!) a.e. on an,
IF*I E p(an,da}, and unless F == 0, 10glF*1 E LI(an,da) and

(3.2) log IF*(z)l:::; ( log IF*(~}ldw(~, z), for all zEn.


lao.
If equality holds in (3.2) at all points zEn (equivalently, for at least one zEn),
we call F an outer function in MHI(n).
While a single valued inner-outer decomposition is impossible in HP(n), a de-
composition in modulus automorphic inner and outer factors is possible.
THEOREM 3.3 ([48]). Let f E HP(n) be not identically zero. Then there exist
a Blaschke product BE MHOO(n), a singular inner function 8 E MHOO(n}and an
outer function F E MHP (n) such that for all zEn
If(z}1 = IB(z)I18(z}IIF(z}l
This factorization is unique in the following sense: if B 1 , 8 1 and FI are a Blaschke
product, a singular inner function and an outer function on n for which If(z)1 =
IB1 (z}I181 (z}IIFI(z)l, then IBI = IBII, 181 = 181 1 and IFII = IFI
3.1. Extreme points in HI(n}. The question arises which functions are
extreme in the unit ball of Hl(n). After the deLeeuw-Rudin theorem 2.3, the
following result is elementary:
LEMMA 3.4. Iff E HI(n} is an outer function of unit norm, then f is extreme
in the unit ball of HI(n}.
Any attempt to copy the proof of the deLeeuw-Rudin theorem in the other
direction will break down. For suppose f = I F where I is a non-trivial inner
function (in MHOO(n)}. Then deLeeuw and Rudin look at the function 9 = (1+12)F
74 PAUL BENEKER AND JAN WIEGERINCK

and show that IIf gill = 1. However, unless I is a single valued inner function,
1 + 12 is not a well-defined modulus automorphic function. There is no remedy
for this problem, because as we have already mentioned, when m ~ 2 there exist
extreme points with a non-trivial inner part.
The following theorem of F. Forelli is crucial in understanding which inner
functions can appear in the inner-outer factorizations of an extreme points.
THEOREM 3.5 ([14]). Let f be an extreme point of the unit ball of H1(n). Then
the codimension of the H 1-closure of f . Hoo in H1 is at most T' when n is bounded
by m + 1 closed smooth curves.
All functions in the H 1-closure of f . HOO(n) inherit the zeros of f, hence
Forelli's theorem implies that an extreme point can only have a limited number of
zeros. In fact one has:
COROLLARY 3.6 ([14]). If f is an extreme point of the unit ball of H1 (n), then
the inner part of f is a finite Blaschke product with at most Tzeros.
Inspection of the proof of the deLeeuw-Rudin theorem gives us the following
criterion for extremity (where we identify H 1-functions with their boundary values):
LEMMA 3.7. Let f E H1(n) be of unit norm. Then f is not extreme if and only
if there exists a non-constant real function k E LOO(an) for which kf E H1(n).
Suppose f = I . F is of unit norm where I is a finite Blaschke product, and F
outer. Let us suppose that f is not an extreme point of the unit ball of H1(n). Then
let k be as in the lemma, and let 9 E H1(n) have boundary values kf. Because
F is an outer function, for all zEn: Ig(z)1 $ IIkll oo 1F(z)l. Hence, because also
III = 1 everywhere on an, the meromorphic function h = g/ f is bounded near an,
real-valued (a.e.) on an, and has its poles in the zeros of f (with corresponding
mUltiplicities). Conversely, if h is a meromorphic function on n with these three
properties, then with k := h on an and 9 := hf E H 1(n), we have 9 = kf on an,
so by the previous lemma, f is not an extreme point. We come to the following
definition.
DEFINITION 3.8. Let I be a finite Blaschke product. We say that I is an
extremal Blaschke product if there exists no meromorphic function h on n that is
bounded near an, real-valued on an and has its poles in the zeros of I, with no
greater multiplicity than the zeros of I.

The conclusions of the previous paragraphs may thus be summarized as follows:


PROPOSITION 3.9. Let the norm of f E H1(n) be 1. Then f is an extreme
point of the unit ball of H1(n) if and only if the inner part of f is an extremal
Blaschke product.
We wish to stress that Forelli's theorem also gives us an upper bound for the
number of zeros of an extremal Blaschke product on n. Also, by the previous
proposition we see that it is only the location of the zeros of a function in H1 (n)
and not so much the outer factor that decides whether or not the function is extreme
in the unit ball (after normalization).
The problem of determining the extreme points of H1 (n) has thus been
reduced to a problem on meromorphic functions on n with pre-described poles,
that is: a problem concerning meromorphic divisors on n.
THE BOUNDARY OF THE UNIT BALL IN Hl_TYPE SPACES 75

Let I be a finite Blaschke product with zeros Zl, Z2, ... , Zn repeated according
to multiplicity. Thus I has n zeros on O. Let 8 := 1'ZI +1'Z2+' -+1,zn be the divisor
on 0 associated with I. If 8' = L:zEf! d'(z) . Z is another divisor on 0 we say that
8' ~ 8 if at every Z E 0: 8'(z) ~ 8(z). The space of all meromorphic differentials w
on 0 that are real-valued on 00 and for which the associated divisor (w) satisfies
(w) ~ 8 is a real linear space of dimension MD(8). Using a theorem of H.L. Royden
[39], based on the Riemann-Roch theorem, T.W. Gamelin & M. Voichick proved
the following result:
THEOREM 3.10 ([15]). The Blaschke product I with zeros Zl, Z2, ... , Zn and
associated divisor 8 is extremal if and only if MD(8) + 2n = m.
In particular, using only the fact that the inner factor of an extreme point is a
finite Blaschke product (as shown by Forelli), Gamelin & Voichick also arrived at
Forelli's upper bound (W-) for the number of zeros of an extremal Blaschke product.
They proved that this upper bound is also sharp.
THEOREM 3.11 ([15]). The HI-closure of the set of extreme points in the unit
ball of HI(O) is the collection of all functions in HI(O) that have unit norm and
no more than W- zeros.
There is a special type of finite domains where Gamelin and Voichick described
the zero sets of extreme points explicitly, namely the so called real slit domains.
These are usually defined as the extended complex plane with a finite number of
intervals deleted. In our situation we prefer a (conformly equivalent) definition.
DEFINITION 3.12. We will call any domain n of the form JDl \ (It u ... U 1m),
where It, 12', . ,Im are disjoint, bounded and closed intervals in (-1, 1) a real slit
domain.
THEOREM 3.13 ([15]). Let n be a real slit domain and let Zl, Z2, ... , Zn be
points of n (not necessarily distinct). Then the Blaschke product with zero set
Zl, Z2, ... ,Zn is extremal if and only if:
n::; W- and
for all i, j: Zi =I- Zj. (In particular, none of the Zi is real.)
We omit most of the proof, and only observe that because the meromorphic
function Z~Zi + z~z; + l-~iz + l-kz is bounded and real-valued on an c JR., no
zeros of an extremal Blaschke product are conjugated.

3.2. Strong exposedness and the location of zeros. In this section we


investigate exposed and strongly exposed points in the unit ball of HI(O). We give
several examples and criteria for (strongly) exposed points. Also we show that non-
trivial properties of strongly exposed points in HI (JDl) (for example: Ll-invertibility
on the boundary) have no analogue for finite domains. Finally, we again look at
the zero sets of extreme points and the question of divisibility of extreme functions
by functions of the form (1 + U)2, where u is a non-constant inner function.
The Hahn-Banach theorem again gives that for f an exposed point of the unit
ball of HI(O) the exposing functional L for f is unique and given by;

L : 9 E HI 1-+ fan 9 I~I da.


76 PAUL BENEKER AND JAN WIEGERINCK

Hence, like H1(1I))), a function I in the boundary of the unit ball of H1(n) is exposed
if and only if it is rigid: apart from (positive) constant multiples of I there is no
H 1-function with the same argument a.e. (dO") on an.
The following criteria for rigidity of H 1 (1I)))-functions carryover to finite
domains word for word:
If IE H1 and 1/1 E Hl, then I is rigid (Theorem 2.9. 1).
If there is agE Hoo such that Re(fg) > 0 a.e. on an, then I is rigid
(Theorem 2.9. 2).
If u is a non-constant inner function such that 1/(1 + u)2 is in Hl, then
I is not rigid (or I == 0).
A priori the first two conditions can only be used to demonstrate rigidity of
outer functions. In both cases III cannot be too small near the boundary of 0.:
if I satisfies the second condition, then 1/1 E H1 -"'(an) for all E: > 0, so 1/1 is
"nearly" in H1. The first condition can be modified to allow for exposed points
with zeros on n.
PROPOSITION 3.14 ([4]). II I is extreme in H1(n) and 1/1/1 E L1(an) then I
is exposed.
Similar to Theorem 2.13 is:
THEOREM 3.15. Let I be a lunction in H1(n). Then I is strongly exposed in
the unit ball 01 H1 il and only il I is exposed and L oo-dist(7 /1/1, Hoo+C(an)) < 1.

Throughout the remainder of this section the domain R will be a real slit
domain with m slits that contains the origin.
Note that on 11'\ {i} the function (z+i)2 has the same argument as iz so (z+i)2
is not rigid in H1 (II))).
LEMMA 3.16 ([4]). For all m ~ 2, the normalized lunction I(z) = c(z + i)2 is
strongly exposed in the unit ball 01 H 1(R).
The proof is an amusing exercise in elementary function theory. One supposes
that 9 E H1 has the same argument as I a.e. on oR and sets h = g/ f. Schwarz's
reflection principle eventually shows that h extends to a rational function which
turns out to have no poles at all. Hence I is exposed and by Theorem 3.15 strongly
exposed.
REMARK 3.17. Similarly, if m > k + 1, then the normalized function hk(Z) =
c(z + i)2k is strongly exposed in the unit ball of H 1 (R). In particular we see that
there exist strongly exposed points in the unit ball of H1 (R) that are "small" on
the boundary: 1/lhkl rt L 1 / 2k (OR).

We recall that for I = I . F to be an extreme point the only requirement is


that the inner part I of I is an extremal Blaschke product - a generic zero set;
"most" of the properties of the function I then follow from its outer factor, i.e.,
the size of Ilion the boundary an. It is reasonable to ask whether exposedness
is also essentially a property of the outer factor. We make this question precise in
the following sense: if I E H1 is a rigid outer function, I is an extremal Blaschke
product on 0., and 9 is invertible in MHOO(n) and such that Ig E HOO(n) (a single
valued function), is the extreme point Ig 1/IIIg . 1111 also exposed? (Compare
THE BOUNDARY OF THE UNIT BALL IN Hl_TYPE SPACES 77

with the first example in Section 2.2.) Proposition 3.14 tells us the answer is yes if
11f E Ll(80). Theorem 3.18 below shows that in general the answer is no.
We mentioned in Section 2.2 Helson's criterium for exposedness of outer func-
tions, 2.8. It fails for finite domains:
THEOREM 3.18 ([4)). For m = 3, there exists ~ E R such that the function
f(z) = c(z - ~)(z + i)2
is extreme in the unit ball of Hl(R), but not exposed.
We will only describe a non-trivial HI (R)-function 9 with the same argument
as f a.e. on an, for suitable ~ E R \ lR and refer to [4J for details.
Suppose the three slits are the intervals [Xl,YlJ, [X2,Y2J, [X3,Y3J. Let k(z) be a
rational function with poles of order 2 in ~,~, 1/~, 1/~, and poles of order 4 in i
and zeros at the 12 points xtl, yt
l , and double zeros at 1, while k(oo) = -1.

Next, let q be the (well-defined) square root of k on R with q(oo) = -i.


Now with a suitable choice of ~ and an appropriate polynomial p of degree 8
that is positive on lR U '][', one can show that the function
g(z) = p(z) + q(z) 1
(z - ~)(z - i)2
(1- ~z)(l- ~z)
belongs to HOO(R) and has the same argument as f on 8R \ {-i}, which implies
that f is not rigid. Because the inner part of f is the extremal Blaschke product
with zero at ~, f is extreme, however.
By Lemma 3.16, (z + i)2 is rigid. We conclude that if Helson's criterion were
valid, for any two inner functions u, v not both constant on R, (z + i)2/(u(z) +
v(z))2 (j. Hl(R), hence also (z-~)(z+i)2/(u(z)+v(z))2 (j. Hl(R), a contradiction!
4. The Bergman space AI(JD))
Recall from Section 1.2 the definition of the Bergman space
AP(JD)) = H(JD)) n LP(JD))
In the Bergman space extreme and exposed points are extremely simple. Indeed,
every f E 8B(Al(JD))) is exposed, and (hence) extreme. The exposing functional is

Lf : gl---+ kgl~ldA(Z).
If Lfg = 1 for some 9 E 8B(Al(JD))), the argument of 9 would be equal to that of
f, (a.e) and this clearly implies f = g. However, to study strongly exposed points
we need some machinery.
4.1. Bergman projection and Bloch space. An important tool for the
study of strongly exposed points will be the Bergman projection; this is the orthog-
onal projection

It is given by an integral operator


[ f(w)
(4.1) Pcp(z) = (f, kz) = iD (1 _ zw)2 dA(w)
(4.2) = fUn+ 1) 1m f(w)'W'dA(w))zn.
n=O D
78 PAUL BENEKER AND JAN WIEGERINCK

Elementary properties are, see e.g. [18],


If cp has compact support in]l)), then Pcp is holomorphic on a neighborhood
of Jij.
If cp is Coo on Jij, then Pcp is Coo on Jij.
The effect of the Bergman projection on Loo(]I))) and on C(Jij) will be very important
for us. We need the Bloch spaces.
DEFINITION 4.1. The Bloch space B consists of all holomorphic functions f on
]I)) with the property that (1-lzI2)lf'(z)1 is bounded on]l)). Equipped with the norm
(4.3) IlfllB := If(OI + sup (1 -lzI 2)1f'(z)l,
zED

B becomes a Banach space.


The set of all functions f in B for which the expression (1-lzI2)1f'(z)1 --> 0 as
Izl --> 1 is a closed subspace of B, called the little Bloch space Bo.

Let Co denote the continuous functions on Jij that are zero on T. We have the
following theorem of R. Coifman, R. Rochberg and G. Weiss:
THEOREM 4.2 ([8]). The Bergman projection P maps Loo(.]I))) boundedlyonto
B. Furthermore, P maps both C(D) and Co boundedlyonto Bo.
The result proved in [8] is much more general than Theorem 4.2. As we state
it, the theorem may be found in [18], Theorem 1.12, with an elementary proof.
There we also find the following results
THEOREM 4.3. (1) The dual space of Al is the Bloch space B under the
following pairing:

(4.4) g E B : f E Al !--+lim ( fr(z)g(z)dA(z).


rTl JD
(2) The dual space of the little Bloch space Bo is the Bergman space A l under
the pairing:

f E Al : g E Bo 1--+ lim ( f(z)gr(z)dA(z).


rTl JD
REMARK 4.4. When one identifies (A1)* with the Bloch space in Theorem 1
B, the dual norm on B yields a norm that is equivalent with, but not equal to the
norm 11.1113 that we have previously defined on B. Hence, there exists a norm 11.11.
on the Bergman space A I that is equivalent to 11.111 and is such that the dual norm
of 9 E B = (AI)* equals IIgIiB. The strongly exposed points in the unit ball of Al
with the norm 11.11. have been described by C. Nara ([32]), who also showed that
up to isometrical isomorphisms, Al with the 11.11. norm is the unique pre-dual of B.
4.2. Strongly exposed points of AI(.]I))). The following theorem is a con-
sequence of Theorems 5.3 and 5.2, the fact that Al(]I))) may be identified with the
subspace of HI (lI2) consisting of functions that depend only on one variable, and
the fact that these functions are exposed in B(HI(lI2). We set

(4.5)
THE BOUNDARY OF THE UNIT BALL IN Hl-TYPE SPACES 79

THEOREM 4.5. Let I E Al be 01 unit norm. Then I is strongly exposed in


Ball(AI) il and only il the LOO-distance 01 filII to the space (AI)1. + C(D) is less
than one.
Henceforth we will simply write (AI)1. +C instead of (AI)1. + C{D).
The question now is: how can we estimate the distance in Loo of cp = 1/111 to
(AI)1. +C, where I is a given function in AI? (Clearly the distance cannot exceed
one. )
Let us first look at polynomials of a particularly simple form: I (z) = c( z - 0:) n ,
where c is normalizing. We will assume n 2: 1 because the constant functions are
strongly exposed by Theorem 4.5. We distinguish three cases in order of increasing
difficulty: 10:1 > 1, 10:1 < 1 and 10:1 = 1.
The case where 10:1 > 1 is very easy: 1/111 is continuous on D, so I is strongly
exposed. In fact, we may even take non-integer powers n and products of such
functions and we always obtain strongly exposed points after normalization.
When 10:1 < 1, let us write cp = 1/111 = 'l/Jo + 'l/JI, where 'l/JI is compactly
supported in 1D> and cp == 'l/JI on a neighborhood of 0:, while 'l/Jo is smooth on D. From
(4.1) we see that P'l/JI is holomorphic across the unit circle because 'l/JI is compactly
supported in 1D>. Next, because 'l/Jo is smooth on D, also P'l/Jo is smooth on D.
Hence Pcp is continuous on D. Now cp - Pcp is bounded, so cp = (cp - Pcp) + Pcp
is contained in (AI)1. + C. By Theorem 4.5 I is strongly exposed. Again, our
reasoning readily shows that the normalized product I of functions (z - O:i)n;, for
all ni and all O:i ft 'll', is strongly exposed.
Now suppose 10:1 = 1; we may take 0: = 1. Let us write In{z) = cn (1- z)n and
CPn = In/l/nl Introducing polar coordinates and applying Cauchy's theorem one
finds that the exposing functional L for h is given by

L(g) = [ g(z) ~1 - z~: dA{z) = [ Ig{z) (1 - z) dA{z) = Cog{O) + CIg'{O).


J"D 1- Z JD - z
But then there exists a polynomial P2 such that L(g) = fD gP2 dA. Therefore CP2 - P2
is contained in (A I) 1., hence CP2 E (A I) 1. + C so h is strongly exposed.
Similarly, for all even n, CPn is contained in (A I) 1. + C and In is strongly exposed
in AI. Again, we may introduce non-integer exponents. Let 1f3 = cf3(1 - z)f3,
where (3 > -2 to ensure that I f3 E A I; the constant Cf3 > 0 is normalizing. Set
cpf3 = 1f3/I/f3I
PROPOSITION 4.6 ([5]). For all (3 > -1, the Loo-distance 01 CPf3 to (AI)1. + C
is at most Isin( f327i") I In particular, lor all (3 > -1, (3 =I- 1,3,5, ... , the lunction I f3
is strongly exposed in the unit ball 01 A I.
The proof consists of observing that if 1(3 - 2nl < 1, then IIcpf3 - CP2nli00 =
IsinCf)1 < 1.
We will come back to odd exponents in Section 4.4 and end this section with
and example of a boundary points of B (A I (1D>)) that is not strongly exposed.
2
EXAMPLE 4.7 ([5]). The normalized function I(z) = (I-z)2~~g2(I--cz) is not
strongly exposed in the unit ball of A I. The functions I f3 tend pointwise to O.
However, limf3!-2 fD /{3"(j5 dA = 1.
80 PAUL BENEKER AND JAN WIEGERINCK

4.3. The space (AI).1 + C. We have observed that (AI).1 + C plays the
same role in Theorem 4.5 with respect to the Bergman space as (HI).1 + C('1I') =
Hoo +C('1I') with respect to the Hardy space HI(JI))) (Theorem 2.13). We mentioned
already in Section 2.3 that Hoo + C is closed in L oo . From this then it followed
relatively easily that Hoo + C('1I') is in fact an algebra, cf. [41], Theorem 6.5.5. How
far do these results extend to the space (A I ).1 + C? From [5] we quote
(1) Al (JI))).1 + c(il~) is a closed subspace of LOO(JI))). (It equals P-I(80 )!)
(2) AI(JI))).1 + C(ii)) is a C(ii))- module.
(3) A I (JI))).1 + C(ii)) is not an algebra.
(4) The space (A 1).1 + C is invariant under composition with holomorphic
automorphisms of JI)).
EXAMPLE 4.8. [5] As for (3), let J{3 = (1- z){3 and let 'P{3 = J{3/IJ{31 for {3 E R
Then 'P2 and 'P-4 E (AI).1, but 'P-2 is not contained in (AI).1 + C. The space
(A I ).1 + C is not an algebra.
The properties of (A 1).1 +C lead in a straightforward way to the following
proposition.
PROPOSITION 4.9. Let J be a strongly exposed point in AI. Then
(a) iJu is an automorphism oJJI)), then the normalized Junction FI = CI(fou)
is strongly exposed;
(b) iJ v E A(JI))) is zero-Jree on the circle, then the normalized Junction F2 =
C2!v is strongly exposed.
Furthermore, the functions J IIJI, FdlFII and F2/1F21 have the same Loo-distance
to (AI).1 + C.
4.4. Strong exposedness of (1 - z){3. We saw in Section 4.2 that the func-
tions J{3 = c{3(1 - z){3 are strongly exposed in the unit ball of Al for all {3 > -1
except possibly when {3 = 1,3,5, .... This was deduced from rather straightforward
estimates of the L 00 -distances of the functions 'P = ff3 I IJ{31 to the space (A 1).1 + C
(Proposition 4.6). In [5] a much sharper result is proved.
THEOREM 4.10. For all {3 ::::: 0, the Bloch distance oj the Junction P'P{3 to 8 0
4I sin({h)1
equals 1r {3+T .
SKETCH OF PROOF. It is convenient to rewrite 'P{3 as 'P{3(w) = (1- w){3/2/(1-
w){3/2. Using the series expansions for the Bergman kernel 1/(1- zw)2 (see (4.2)),
as well as for (1 - W){3/2, and 1/(1 - w){3/2, we evaluate the Bergman projection
P'P{3' One obtains P'P{3 = E~=o c{3,nzn, where
-(n + 1) sin(~) ~ r(m + ~)r(m + n - ~)
c{3,n = 271" ~ m!(m + n + I)! .
It is proved in [5] that for fixed (3 > 0:
~r(m+~)r(m+n-~)
(4.6) =:0 m!(m + n + I)!
4
= n2{3({3 + 2) (1 + 0(1)),

where the o(I)-term tends to zero as n -+ 00. This implies that


-2sin(~)
c{3,n = 7I"({3 + 2)n (1 + 0(1)),
THE BOUNDARY OF THE UNIT BALL IN HI-TYPE SPACES 81

where the o{I)-term vanishes as n --> 00. But then,

so the Bloch distance of PCP(3 to Bo is at least 41:~~~1I. On the other hand, for
large N,

I{ ~ c zn)'1 < ~ nlc 1.l z ln-1 < 21 sin{~)ll + 0(1)


!::tv (3,n - n~ (3,n - 7r(,B + 2) 1 - Izl '

where the o{I)-term tends to zero as N increases. Using the fact that the polyno-
mials are contained in Bo it follows that the Bloch distance of PCP(3 to Bo is at most
4I sin ()1
... ((3+2) . o
COROLLARY 4.11. Let d(cp(3, (Al)1. + C) denote the LOO-distance of CP(3 to
(Al)1. + C. Then for all,B 2: 0,

(4. 7) ~ 1sin( !?f) 1 < d( (A 1) 1. + C) < ~ 1sin( ~) 1 < ~


2 ,B + 2 - CP(3, - 7r ,B + 2 - 7r'
In particular, all f(3 are strongly exposed for ,B 2: O.
SKETCH OF PROOF. Let q: B --> B/Bo be the quotient map. By Theorem 4.2,
the map q 0 P : L OO --> B/Bo is continuous and surjective. The kernel of the map
q 0 P is the space (Al)1. + C, cpo Property (I) from Section 4.3. Hence the derived
map
P* : L oo / ((Al)1. + C) --> B /Bo
is bijective and bounded (by ~ as follows from the proof in [18] of Theorem 4.2).
This gives the lower bound for d(cp(3, (Al)1. + C), because IIP*cp(311 = ~ ISi~~y)l.
By the closed graph theorem, the inverse P* -1 of P* is also bounded. Actually,
one can show directly that II P* -111 ::; 1, which in turn yields the upper bound for
d{cp(3, (Al)1.+C). Supposing that FE B/Bo has norm 1, we will show that P*-I(F)
has norm at most 1 in L oo /((Al )1.+C). For any c > 0, we can find a representative
fEB of the coset F such that IIfl18 < 1 + c. In the proof of Theorem 4.2 in [18]
one finds that
'(w) = (I -IWJ2) . f'(w) ~ f'(O) E L oo
w
satisfies f(z) - P'{z) = f(O) + j'(O)z E Bo. Thus ' is a representative of P*-I(F)
in LOO. As a consequence

IIp*-l{F)IILOO/((AI)J.+C) ::; d(,, (Al)1. + C)


::; lim esssuPr<lwl<II,{w)1
r-->l
= lim sup 1(1 -lwI )f'(w)1 ::; IIfll8 < 1 + C.
2
Iwl-->l
o
The module structure of (Al)1. +C and the previous corollary lead to
82 PAUL BENEKER AND JAN WIEGERINCK

COROLLARY 4.12. Suppose that g E A(JI1 vanishes nowhere on'll'. Let Zl, Z2,
E 'll' be distinct and let {31, {32, ... ,(3n be real numbers greater than - 2. Then
.. , Zn.
the normalized Junction J(z) = cg(z) n~=l (1- ZZi),8i is strongly exposed in the unit
ball oj A I iJ and only iJ all Junctions J,8i = C,8i (1 - z ),8i are strongly exposed. In
particular, all choices oj {3i > -1 yield strongly exposed points and all normalized
polynomials are strongly exposed in the unit ball oj A I.
We end this section with a conjecture on the functions J,8 for -2 < {3 < 0, for
which strong exposedness is already implied by Proposition 4.6 when -1 < {3 < O.
In [5] it is shown that

~lsin(!?f)I<d((1l
'Tr {3+2 - .,-,8,
(AI)1.+C)
,
-2< (.J<O,
fJ

with equality in the limit as {3 ! -2. This leads us to surmise the following:

CONJECTURE 4.13. For all-2 < (3 < 0, d(cp,8,(AI)1. +C) = ~lsi~~y)J. In


particular, the Junctions J,8 are strongly exposed Jor all said (3.
5. Scattered results for B(HI (lRn ))
The geometry of B(HI(lR n )) is still quite out of reach. At one time it was
conjectured that all boundary points would be extreme, [41], but this conjecture
was refuted as soon as the existence of inner functions on lR n was proved, [1, 27].
Just as in Theorem 2.3 one finds that inner functions, or more generally products
with an inner factor cannot be extreme. However, it is well known that there
is no inner-outer decomposition possible for B(HI(lR n )), because the zero sets of
functions in various HP classes are essentially different in size. Rudin already
observed in [41] that a function J E 8B(H 1 (lR n )) which is continuous on an open
set in 8lRn must be extreme. The proof goes by considering the intersection of lR n
with appropriate complex lines and applying I-dimensional theory. The same idea
can be used to show that such functions are in fact exposed, [50].
EXAMPLE 5.1. Not all extreme points are exposed: c(1 + 1)2 is extreme but
has the same argument as I.
Example 4.7 also is an example of an exposed point in B(HI(lR2)) that is not
strongly exposed.
Let us write S = 8lRn , HI = HI(lR n ) and HJ for the functions in HI that
vanish at o. We define
(HI)1. = N E Loo(S) : Is {JjdA = 0 for allJ E HI}

Analogously we define (HJ)1.. Let d denote the Loo-distance on S.


Theorems 2.13 and 2.12 have a full analogue in B(H 1 ).
THEOREM 5.2 ([50]). Let J be an exposed point in B(H 1 ). Then I is strongly
exposed iJ and only il d(J 11/1, C(S) + (HI)1.) < 1.
The proof again uses methods from uniform algebra, but is more involved.
Noticing that (HJ)1.) is a fairly small subspace of (HI)1. +C(S), this theorem
can be strengthened in one direction.
THEOREM 5.3. Suppose that I is a strongly exposed point in B(H 1 ). Then
d(J IIJI, (HJ)1.) < 1.
THE BOUNDARY OF THE UNIT BALL IN Hl_TYPE SPACES 83

EXAMPLES 5.4. Homogeneous polynomials in 8B(H 1 ) are strongly exposed,


see [50J. Corollary 4.12 shows that polynomials of one variable in 8B(H 1 (B 2 )) are
strongly exposed.
It seems natural to try to generalize ideas of Section 4 to HI. The Bergman
projection should be replaced by the Szego projection P and the Bloch spaces 8
and 8 0 by BMOA, the space of analytic functions of bounded mean oscillation and
VMOA, the space of functions of vanishing mean oscillation (with respect to the
non-isotropic balls in S). It is known that P maps Loo(S) surjectively to BMOA
and C(S) surjectively to VMOA, see [25J for an overview of these kind of results.
So we again have a map P*: Loo(S)/H1)1. +C) -> BMOA/VMOA. However,
we have no idea how to estimate (p*)-l, nor how to estimate the norm of P*cp for
cp = p/ipi with p for example a polynomial. Answers to these questions would be
very interesting.

References
[1] A.B. Aleksandrov, A. B. The existence of inner functions in a ball, (Russian) Mat. Sb. (N.S.)
118(160) (1982), no. 2, 147-163,287.
(32A40 46J15)
[2] P. Beneker, Strongly exposed points, Helson-Szego weights and Toeplitz operotors, J. Int. Eq.
Oper. Th. 31 (1998), 299-306.
[3] P. Beneker, Strongly exposed points in unit balls of Banach spaces of holomorphic func-
tions, Academisch Proefschrift, Universiteit van Amsterdam, 2002. Electronically available
at http://remote.science.uva.nl/ ~beneker /thesis.ps
[4] P. Beneker, J. Wiegerinck, Exposedne'ss in Hardy spaces of domains of finite connectivity,
Indag. Math., N.S., 11 (4),2000,487-497.
[5] P. Beneker, J. Wiegerinck, Strongly exposed points in the ball of the Bergman space,
http://arXiv.org/abs/math.CV /0208234 to appear.
[6] L. de Branges, J. Rovnyak, Square Summable Power Series, Holt, Rinehart and Winston,
New York, 1966.
[7] L. de Branges, J. Rovnyak, Appendix on square summable power series, in "Perturbation
Theory and its Applications in Quantum Mechanics", John Wiley and Sons, New York (1966),
347-392.
[8] R.R. Coifman, R. Rochberg, G. Weiss, Factorization theorems for Hardy spaces in severol
variables, Ann. of Math. (2), 103 (1976), 611-635.
[9] A. Devinatz, Toeplitz operotors on H2 spaces, Trans. Amer. Math. Soc. 112 (1964), 304-317.
[10] J. Diestel, Geometry of Banach paces - Selected topics LNM 485, Springer, Berlin etc. 1975.
[11] R.G. Douglas, D.E. Sarason, Fredholm Toeplitz operotors, Proc. Amer. Math. Soc. 26 (1970),
117-120.
[12] P.L. Duren, Theory of HP-spaces, Academic Press, 1970.
[13] S.D. Fisher, Function Theory On Planar Domains, John Wiley & Sons, 1983.
[14] F. Forelli, Extreme points in Hl(R), Can. J. Math. 19 (1967), 312-320.
[15] T.W. Gamelin, M. Voichick, Extreme points in spaces of analytic functions, Can. J. Math.
20 (1968), 919-928.
[16] J. Garnett, Bounded holomorphic functions, Pure & Applied Mathematics, 96, Academic
Press, Inc., New York-London, 1981.
[17] E. Hayashi, The solution of extremal problems in Hl, Proc. Amer. Math. Soc. 93 (1985),
690-696.
[18] H. Hedenmalm, B. Korenblum, K. Zhu, Theory of Bergman Spaces, Springer Graduate Texts
in Mathematics 199, 2000.
[19] H. Helson, G. Szego, A problem in prediction theory, Ann. Mat. Pura Appl. 51 (1960),
107-138.
[20] H. Helson, D. Sarason, Past and future, Math. Scand. 21 (1967), 5-16.
[21] H. Helson, Large analytic functions II, Analysis and partial differential equations, Lecture
Notes in Pure and Appl. Math., 122, Dekker, New York (1990), 217-220.
84 PAUL BENEKER AND JAN WIEGERINCK

[22] J. Inoue, An example of a non-exposed extreme function in the unit ball of Hl, Proc. Edin-
burgh Math. Soc. 37 (1993), 47-51.
[23] J. Inoue; T. Nakazi, Polynomials of an inner function which are exposed points in Hl. Proc.
Amer. Math. Soc. 100 (1987), no. 3, 454-456.
[24] P. Koosis, Introduction to HP spaces, Cambridge University Press, 1980.
[25] S.G. Krantz, Geometric Analysis and FUnction Spaces CBMS Regional Conference Series in
Mathematics, 81. Amer. Math. Soc., Providence, RI, 1993.
[26] K. de Leeuw, W. Rudin, Extreme points and extremum problems in Hl, Pacific J. Math. 8
(1958), 467-485.
[27] E. Ll1lw,A construction of inner functions on the unit ball in CP. Invent. Math. 67 (1982),
no. 2, 223-229.
[28] T. Nakazi, Exposed points and extremal problems in Hl, J. Funet. Anal. 53 (1983), 224-230.
[29] T. Nakazi, Exposed points and extremal problems in Hl, II, T6hoku Math. J. 37 (1985),
265-269.
[30] T. Nakazi, Existence of solutions of extremal problems in Hl , Proc. Edinburgh Math. Soc.
(2) 34 (1991), no. 2, 99-112.
[31] T. Nakazi, personal communication.
[32] C. Nara, Uniqueness of the predual of the Bloch space and its strongly exposed points, Illinois
J. Math, 34 (1990), no. 1,98-107.
[33] Z. Nehari, Conformal Mapping, McGraw-Hill, 1952.
[34] Z. Nehari, On bounded bilinear forms, Ann. of Math. 65 (1957), 153-162.
[35] D.J. Newman, Pseudo-uniform convexity in HI, Proc. Amer. Math. Soc. 14 (1963), 676-679.
[36] J. Neuwirth, D.J. Newman, Positive Hl/2 functions are constant, Proc. Amer. Math. Soc.
18 (1987),958.
[37] A. Nicolau, The coefficients of Nevanlinna's pammetrization are not in HP, Proc. Amer.
Math. Soc 106 (1989), 115-117.
[38] R.R. Phelps, Dentability and extreme points in Banach spaces, J. Funet. Anal. 17 (1974),
78-90.
[39] H.L. Royden, The boundary values of analytic and harmonic functions, Math.Z. 78, (1962),
1-24.
[40] W. Rudin, Analytic functions of class H p , Trans. Amer. Math. Soc. 78 (1955), 46-66.
[41] W. Rudin, FUnction Theory in the Unit Ball of en, Grundlehren der Mathematischen Wis-
senschaften, 241, Springer-Verlag, New York-Berlin, 1980.
[42] D.E. Sarason, Algebras of functions on the unit circle, Bull. Amer. Math. Soc. 79 (1973),
286-299.
[43] D.E. Sarason, Function Theory on the Unit Circle, Virginia Polytechnic Institute and State
University, Blacksburg, Virginia, 1979.
[44] D.E. Sarason, Exposed points in Hl, I, Operator Theory: Advances and Applications 41
(1989), Birkhiiuser Verlag Basel, 485-496.
[45] D.E. Sarason, Exposed points in Hl, II, Operator Theory: Advances and Applications 48
(1990), Birkhiiuser Verlag Basel, 333-347.
[46] D.E. Sarason, Sub-Hardy Hilbert Spaces in the unit disc, Univ. of Arkansas Leeture Notes in
the Math. Sc. 10, Wiley-Interscience, 1995.
[47] D. Temme, J. Wiegerinck, Extremal properties of the unit ball of Hl, Indag. Mathern., N.S.,
3 (1) (1992), 119-127.
[48] M. Voichick, L. Zalcman, Inner and outer functions on Riemann surfaces, Proc. Amer. Math.
Soc. 16 (1965), 1200-1204.
[49] H. Widom, Inversion of Toeplitz matrices III, Notices Amer. Math. Soc. 7 (1960), 63.
[50] J. Wiegerinck, A chamcterization of strongly exposed points of the unit ball of Hl, Indag.
Mathern., N.S., 4 (4) (1993), 509-519.
[51] K. Yabuta, Unicity of the extremum problems in Hl(un), Proc. Amer. Math. Soc. 28 (1971),
181-184.
[52] K. Yabuta, Some uniqueness theorems for HP(un) functions, T6hoku Math. J. 24 (1972),
353-357.

FACULTY OF MATHEMATICS, UNIVERSITY OF AMSTERDAM, PLA:'ITAGE I\[UIDERGRACHT 24,


1018 TV, AMSTERDAM, THE NETHERLANDS
E-mail addre.~s:benekerClscience.uva.n1;janwiegClscience.uva.n1
Contemporary Mathematics
Volume 328, 2003

Complete isometries - an illustration of noncommutative


functional analysis

David P. Blecher and Damon M. Hay

ABSTRACT. This article, addressed to a general audience of functional ana-


lysts, is intended to be an illustration of a few basic principles from 'non-
commutative functional analysis', more specifically the new field of operator
spaces. In our illustration we show how the classical characterization of (pos-
sibly non-surjective) isometries between function algebras generalizes to oper-
ator algebras. We give some variants of this characterization, and a new proof
which has some advantages.

1. Introduction
The field of operator spaces provides a new bridge from the world of Banach
spaces and function spaces, to the world of spaces of operators on a Hilbert space.
For researchers in the new field, the philosophical starting point is the combination
of the following two obvious facts. Firstly, by the Hahn-Banach theorem any Banach
space X is canonically linearly isometric to a closed linear subspace of C(K), where
K is the compact space Ball(X*). Secondly, C(K) is a commutative C*-algebra.
Thus one defines a noncommutative Banach space, or operator space, to be a closed
linear subspace X of a possibly noncommutative C* -algebra A. This simplistic
idea becomes much more substantive with the addition of some additional metric
structure. The point is that if A is any C*-algebra, then the *-algebra Mn(A) of
nx 11, matrices with entries in A has a unique norm 11lln making it a C*-algebra (this
follows from the well known nnicity of C*-norms on a *-algebra). If X c A then
Mn(X) inherits this norm II lin, and more precisely we think of an operator space
as the pair (X, {11lln}n). We usually insist that maps between operator spaces are
completely bounded, where the adjective 'completely' means that we are applying
our maps to matrices too. Thus if T : X -4 Y, then T is completely contractive
if Tn is contractive for all n E N, where Tn is the map [Xij] 1--+ [T(x;j )]. Similarly
T is completely isometric if II[T(xij)]1I = I [Xij] I for all n E Nand [Xij] E 1I1n(X).
It is an easy exercise (using one of the common expressions for the operator norm
of a matrix in Mn = .Mn(C)) to prove that a linear map T : X -4 Y between

1991 Mat/;tematics Subject Classification. Primary 46L07, 46L05, 47L30; Secondary 46JlO.
This research was supported in part by a grant from the National Science Foundation..

2003 American Mathematical So(~iety


85
86 DAVID P. BLECHER AND DAMON M. HAY

subspaces of C(K) spaces is completely contractive if and only if it is contractive.


Consequently such a T is isometric if and only if it is completely isometric.
The identification of the term 'noncommutative Banach space' with 'operator
space' may be thought of as a relatively recent entry in the well known 'dictionary'
translating terms between the 'commutative' and 'noncommutative' worlds. We
spend a paragraph describing some other entries in this dictionary. Although t.hese
items are for the most part well known to the point of being tedious, it will be
helpful to collect them here for the dual purpose of establishing notation, and for
ease of reference later in the paper. The most. well known item is of course the fact
that the noncommutative version of a C(K) space is a unital C*-algebra B. The
noncommutative version of a unimodular function in C(K) is a unitary u E B (Le.
u*u = uu* = 1). The noncommutative version of a function algebra A C C(K)
'containing constant functions' is a closed subalgebra A of a C* -algebra B, with
18 E A. We call such A a unital operator algebra. For a unital subset S of a C*-
algebra B, we will take as a simple noncommutative version of the assertion'S C
C(K) separates points of K', the assertion 'the C*-subalgebra of B generated by S
(namely, the smallest C*-subalgebra of B containing S) equals B'. The analogue
of a closed subset E of a compact set K is a quotient B / I, where I is a closed two-
sided ideal in a unital C* -algebra B. More generally, unital *-homomorphisms 7r
between unital C* -algebra.'i are the noncommutative version of continuous functions
T between compact spaces. Indeed clearly any such T : Kl -> K2 gives rise to the
unital *-homomorphism C(K2 ) -> C(Kd of 'composition with T', and conversely
it is not much harder to see that any unital *-homomorphism C(K2 ) -> C(Kd
comes from a continuous T in this way. Moreover such 7r is 1-1 (resp. onto) if and
only if the corresponding T is onto (resp. 1-1). Thus the noncommutative version
of a homeomorphism between compact spaces is a (surjective 1-1) *-isomorphism
between unital C*-algebras. Coming back to 'noncommutative functional analysis',
it is convenient for some purposes (but admittedly not for others) to view 'complete
isometries' as the noncommutative version of isometries. It is very important in
what follows that a 1-1 *-homomorphism 7r : A - B between C*-algebras, is by a
simple and well known spectral theory argument, automatically an isometry, and
consequently (by the same principle applied to 7r n ), a complete isometry. Similarly,
a *-homomorphism 7r : A - B (which is not a priori assumed continuous) is
aut.omatically completely contractive, and has a closed range which is a C* -algebra
*-isomorphic to the C* -algebra quotient of A by the obvious two-sided ideal, namely
the kernel of the *-homomorphism.
The entries we have just described in this 'dictionary' are all easily justified by
well known theorems (for example Gelfand's characterization of commutative C*-
algebras). That is, if one applies the noncommutative definition in the commutative
world, one recovers exactly the classical object. Similarly one sometimes finds
oneself in the very nice 'ideal situation' where one can prove a theorem or establish
a theory in the noncommutative world (i.e. about operator spaces or operator
algebras), which when one applies the theorem/theory to objects which are Banach
spaces or function algebras, one recovers exactly the classical theorem/theory. An
illustration of this point is the Banach-Stone theorem. The following is a much
simpler form of Kadison's characterization of isometries between C*-algebras [17]:
COMPLETE ISOMETRIES - AN ILLUSTRATION 87

THEOREM 1.1. (Folklore) A surjective linear map T : A --+ B between unital


C* -algebras is a complete isometry if and only if T = U7r('), for a unitary u E B
and a *-isomorphism 7r : A --+ B.

PROOF. (Sketch.) The easy direction is essentially just the fact mentioned
earlier that 1-1 *-homomorphisms are completely isometric. The other direction
can be proved by first showing (as with Kadison's theorem) that T(I) is unitary,
so that without loss of generality T(I) = 1. The well known Stinespring theorem
has as a simple consequence the Kadison-Schwarz inequality T(a)*T(a) ~ T(a*a).
Applying this to T- 1 too yields T(a)*T(a) = T(a*a), and now the result follows
immediately from the 'polarization identity' a*b = ~ E~=o(a + ikb)*(a + ikb). 0

Note that if one takes A = C(Kd and B = C(K2 ) in Theorem 1.1, and
consults the 'dictionary' above, then one recovers exactly the classical Banach-Stone
theorem. Indeed as we remarked earlier, in this case complete isometries are the
same thing as isometries, unit aries are unimodular functions, and a *-isomorphism
is induced by a homeomorphism between the underlying compact spaces.
Indeed consider the following generalization of the Banach-Stone theorem:
THEOREM 1.2. [15,22, 1, 20] Let fl be compact and Hausdorff, and A a unital
function algebra. A linear contraction T : A --+ C(fl) is an isometry if and only if
there exists a closed subset E of fl, and two continuous functions "f : E --+ '][' and
r.p: E --+ 8A, with r.p surjective, such that for all y E E
T(f)(y) = "f(y)f(r.p(y))
Here 8A is the Shilov boundary of A (see Section 2). We have supposed that. T
maps into a 'selfadjoint function algebra' C(fl); however since any function algebra
is a unital subalgebra of a 'selfadjoint.' one, the theorem also applies to isometries
between unital function algebras. If A is a C(K) space too, then 8A = K and then
t.he theorem above is called Holsztynski's theorem. We refer the reader to [16] for
a survey of such variants on the classical Banach-Stone theorem.
Often the transition from the 'classical' to the 'noncommutative' involves the
introduction of much more algebra. Next we appeal to our dictionary above to give
an equivalent restatement of Theorem 1.2 in more algebraic language.
THEOREM 1.3. (Restatement of Theorem 1.2) Let A, B be unital function
algebras, with B selfadjoint. A linear contraction T : A --+ B is an isometry if and
only if
(A) there exists a closed ideal I of B, a unitary u in the quotient C*-algebra
B / I, and a unital 1-1 *-homomorphism 7r : A --+ B / I, such that qI (T( a)) =
u7r(a) for all a E A.
Here qI is the canonical quotient *-homomorphism B --+ B/I.
In light of Theorems 1.1 and 1.3 one would imagine that for any complete
isometry T : A --+ B between unital operator algebras, the condition (A) above
should hold verbatim. This would give a pretty noncommutative generalization of
Theorem 1.3. Indeed if Ran T is also a unital operator algebra, then this is true
(see ego B.l in [3]). However, it is quite easily seen that such a result cannot hold
generally. For example, let Mn = Mn(C); for any x E Mn of norm 1, the map
A 1---4 Ax is a complete isometry from C into Mn. Now Mn is simple (Le. has no
88 DAVID P. BLECHER AND DAMON M. HAY

nontrivial two-sided ideals), and so if the result above was valid then it follows
immediately that x = u. This is obviously not satisfactory.
To resolve the dilemma presented in the last paragraph, we have offered in [5]
several alternatives. For example, one may replace the quotient B / I by a quotient
of a certain *-subalgebra of B. The desired relation qJ(T(a)) = U7r(a) then requires
u to be a unitary in a certain C* -triple system (by which we mean a subspace X
of a C* -algebra A with X X* X c X). Or, one may replace the quotient B / I by
a quotient B / (J + J*), where J is a one-sided ideal of B. Such a quotient is not
an algebra, but is an 'operator system' (such spaces have been important in the
deep work of Kirchberg (see [18, 19] and references therein). Alternatively, one
may replace such quotients altogether, with certain subspaces of the second dual
B** defined in terms of certain orthogonal projections of 'topological significance'
(Le. correspond to characteristic functions of closed sets in K if B = C(K)) in
the second dual B** (which is a von Neumann algebra [25]). The key point of all
these arguments, and indeed a key approach to Banach-Stone theorems for linear
maps between function algebras, C* -algebras or operator algebra.-" is the basic
theory of C* -triple systems and triple morphisms, and the basic properties of the
noncommutative Shilov boundary or triple envelope of an operator space. These
important and beautiful ideas originate in the work of Arveson, Choi and Effros,
Hamana, Harris, Kadison, Kirchberg, Paulsen, Ruan, and others. Indeed our talk at
the conference spelled out these ideas and their connection with the Banach-Stone
theorem; and the background ideas are developed at length in a book the first author
is currently writing with Christian Le Merdy [7] (although we do not characterize
non-surjective complete isometries there). Moreover, a description of our work from
this perspective, together with many related results, may be found in [12]. Thus
we will content ourselves here with a survey of some related and interesting topics,
and with a new and self-contained proof of some characterizations of complete
isometries between unital operator algebras which do not appear elsewhere. This
proof has several advantages, for example the projections arising naturally with
this approach seem to be more useful for some purposes. Also it will allow us to
avoid any explicit mention of the theory of triple systems (although this is playing
a silent role nonetheless).
We also show how such noncommutative results are generalizations of the older
characterizations of into isometries between function algebras or C(K) spaces. We
thank A. Matheson for telling us about these results. In the final section we present
some evidence towards the claim that (general) isometries between operator al-
gebras are not the correct noncommutative generalization of isometries between
function algebras.
For the reader who wants to learn more operator space theory we have listed
some general texts in our bibliography.

2. The noncommutative Shilov boundary


At the present time the appropriate 'extreme point' theory is not sufficiently
developed to be extensively used in noncommutative functional analysis. Although
several major and beautiful pieces are now in place, this is perhaps one of the most
urgent needs in the subject. However there are good substitutes for 'extreme point'
arguments. One such is the noncommutative Shilov boundary of an operator space.
Recall that if X is a closed subspace of C(K) containing the identity fUIlction lK on
COMPLETE ISOMETRIES - AN ILLUSTRATION 89

K and separating points of K, then the classical Shilov boundary may be defined to
be the smallest closed subset E of K such that all functions J E X attain their norm,
or equivalently such that the restriction map J f-+ JIE on X is an isometry. This
boundary is often defined independently of K, for example if A is a unital function
algebra then we may define the Shilov boundary as we just did, but with K replaced
by the maximal ideal space of A. In fact we prefer to think of the classical Shilov
boundary of X as a pair (aX, i) consisting of an abstract compact Hausdorff space
ax, together with an isometry j : X -+ C(aX) such that j(IK) = lax and such
that j(X) separates points of ax, with the following universal property: For any
other pair (0, i) consisting of a compact Hausdorff space 0 and a complete isometry
i : X -+ C(O) which is unital (i.e. i(IK) = IA), and such that i(X) separates points
of 0, there exists a (necessarily unique) continuous injection r : ax -+ 0 such that
i(x)(r(w)) = j(x)(w) for all x E X, w E ax. Such a pair (aX, i) is easily seen to
be unique up to an appropriate homeomorphism. The fact that such ax exists is
the difficult part, and proofs may be found in books on function algebras (using
extreme point arguments).
Consulting our 'noncommutative dictionary' in Section 1, and thinking a little
about the various correspondences there, it will be seen that the noncommutative
version of this universal property above should read as follows. Or at any rate,
the following noncommutative statements, when applied to a unital subspace X c
C(K), will imply the universal property of the classical Shilov boundary discussed
above. Firstly, a unital operator space is a pair (X, e) consisting of an operator space
X with fixed element e EX, such that there exists a linear complete isometry Il,
from X into a unital C*-algebra C with Il,(e) = Ie. A 'noncommutative Shilov
boundary' would correspond to a pair (B,j) consisting of a unital C*-algebra B
and a complete isometry j : X -+ B with j(e) = IB, and whose range generates
B as a C* -algebra, with the following universal property: For any other pair (A, i)
consisting of a unital C* -algebra and a complete isometry i : X -+ A which is
unital (Le. i(e) = IA), and whose range generates A as a C* -algebra, there exists
a (necessarily unique, unital, and surjective) *-homomorphism IT : A -+ B such
that IT 0 i = j. Happily, this turns out to be true. The existence for any unital
operator space (X, e) of a pair (B,j) with the universal property above is of course
a theorem, which we call the Arveson-Hamana theorem [2, 13] (see [3] for complete
details). As is customary we write C;(X) for B or (B,j), this is the 'C*-envelope
of X'. It is essentially unique, by the universal property. If X = A is a unital
operator algebra (see Section 1 for the definition of this), then j above is forced to
be a homomorphism (to see this, choose an i which is a homomorphism, and use
the universal property). Thus A may be considered a unital subalgebra of C;(A).
If A is already a unital C* -algebra, then of course we can take C; (A) = A.
To help the reader get a little more comfortable with these concepts, we com-
pute the 'noncommutative Shilov boundary' in a few simple examples.

Example 1. Let Tn be the upper triangular n x n matrices. This is a unital


subspace of M n , and no proper *-subalgebra of Mn contains Tn. Let (B,j) be
the C* -envelope of Tn. By the universal property of the C* -envelope, there is a
surjective *-homomorphism IT : Mn -+ B such that IT(a) = j(a) for a E Tn. The
kernel of IT is a two-sided ideal of Mn. However Mn has no nontrivial two-sided
ideals. Hence IT is 1-1, and is consequently a *-isomorphism, and we can thus
identify Mn with B. Thus Mn is a C*-envelope of Tn.
90 DAVID P. BLECHER AND DAMON M. HAY

Example 2. Consider the linear subspace X of M3 with zeroes in the 1-3,


2-3, 2-1, 3-1 and 3-2 entries, and with arbitrary entries elsewhere except for the
3-3 entry, which is the average of the 1-1 and 2-2 entries. It is easy to see that
the C* -algebra generated by X inside M3 is M2 EB C. However this is not the
C*-envelope. Indeed the 3-3 entry here is redundant, since the norm of x E X is
the norm of the upper left 2 x 2 block of x. The canonical *-homomorphism from
M2 EB C onto Nh when restricted to X is a unital complete isometry from X onto
T2 (see Example 1). Thus if one takes the quotient of M2 EB C by the kernel of this
homomorphism, namely the ideal 02 EB C, then one obtains M 2 , which by Example
1 is the C* -envelope.
Indeed this is typical when calculating the C* -envelope of a unital subspace X
of Mn. The C*-algebra generated by X is a finite dimensional unital C*-algebra.
However such a C* -algebras is *-isomorphic to a finite direct sum B of full 'matrix
blocks' M nk Some of these blocks are redundant. That is, if p is the central
projection in B corresponding to the identity matrix of this block, then x t-+ x(IB-
p) is completely isometric. If one eliminates such blocks then the remaining direct
sum of blocks is the C* -envelope.
Example 3. Let B be a unital C*-algebra. Consider the unital subspace S(B)
of the C*-algebra M 2 (B) consisting of matrices

[~; :1]
for all x,y E B and >.,Ji, complex scalars. We claim that M2(B) is the C*-envelope
C of S(B), and we will prove this using a similar idea to Example 1 above. Namely,
first note that M 2(B) has no proper C*-subalgebra containing S(B), Thus by
the Arveson-Hamana theorem there exists a *-homomorphism 11' : M 2 (B) ---+ C
which possesses a property which we will not repeat, except to say that it cer-
tainly ensures that 11' applied to a matrix with zero entries except for a nonzero
entry in the 1-2 position, is nonzero. It suffices as in Example 1 to show that
Ker 11' = {a}. Suppose that 11'(x) = 0 for a 2 x 2 matrix x E M 2(B). Let Eij
be the four canonical basis matrices for M 2, thought of as inside M2(B). Then
11'(E1i XEj2 ) = 11'(Eli)11'{X)11'(Ej2) = 0 for i,j = 1,2. Thus by the fact mentioned
above about the 1-2 position, we must have EliXEj2 = O. Thus x = O.
In fact a variant of the C*-envelope or 'noncommutative Shilov boundary' can
be defined for any operator space X. This is the triple envelope of Hamana (see
[14]). This is explained in much greater detail in [3], together with many appli-
cations. For example it is intimately connected to the 'noncommutative .M-ideals'
recently introduced in [4]. This 'noncommutative Shilov boundary' is, as we men-
tioned in Section 1, a key tool for proving various Banach-Stone type theorems.
However in the present article we shall only need the variant described earlier in
this section.

3. Complete isometries between operator algebras


We begin this section with a collection of very well known and simple facts about
closed two-sided ideals] in a C* -algebra A, and about the quotient C* -algebra AI].
We have that ]1.1. is a weak* closed two-sided ideal in the von Neumann algebra
A**, and there exists a unique orthogonal projection e in the center of A** with
]1.1. = A**(1 - e). The projection 1 - e is called the support projection for I, and
COMPLETE ISOMETRIES - AN ILLUSTRATION 91

1 - e may be taken to be the weak* limit in A ** of any contractive approximate


identity for I. Thus it follows that A ** / I J..J.. ~ A ** e as C* -algebras. Therefore also
A/Ie (A/ 1)** ~ A** / IJ..J.. ~ A**e
as C*-algebras. Explicitly, the composition of all these identifications is a 1-1 *-
homomorphism taking an a + I in A/I, to ae = eae in A **. Here' is the canonical
embedding A -+ A** (which we will sometimes suppress mention of). Thus A/I
may be regarded as a C*-subalgebra of A**, or of the C*-algebra eA**e.
We next illustrate the main idea of our theorem with a simple special case.
(The following appeared as part of Corollary 3.2 in the original version of [5], with
the proof left as an exercise). Suppose that T : A -+ B is a complete isometry
between unital C*algebras, and suppose that T is unital too, that is T(I) = 1. Let
C be the C*-subalgebra of B generated by T(A). Applying the Arveson-Hamana
theorem 1 we obtain a surjective *-homomorphism () : C -+ A such that (}(T(a)) = a
for all a E A. If I is the kernel of the mapping (), then C / I is a unital C* -algebra
*-isomorphic to A. Indeed there is the canonical *-isomorphism 'Y : A -+ C / I
induced by (), taking a to T(a) + I. The next point is that C/I may be viewed as
we mentioned a few paragraphs back, as a C* -subalgebra of C** , and therefore also
of B**. Indeed if e is the central projection in C** mentioned there, then C / I may
be viewed as a C* -subalgebra of eC** e C eB** e C B**. In view of the last fact,
the map 'Y induces an 1-1 *-homomorphism 7r : A -+ B** taking an element a E A
to the element of B** which equals

(1) -- -- --
T(a)e = eT(a) eT(a)e

--
(these are equal because e is central in C**). Conversely, if T : A -+ B is a complete
contraction for which there exists a projection e E B** such that eT(a)e is a 1-1
*-homomorphism 7r, then for all a E A,

IIT(a)1I ~ Ile~ell = 117r(a)11 = Iiall


using the fact mentioned earlier that 1-1 *-homomorphisms are necessarily isomet-
ric. Thus T is an isometry, and a similar argument shows that it is a complete
isometry. Thus we have characterized unital complete isometries T : A -+ B.
If H is a Hilbert space on which we have represented the von Neumann algebra
B** as a weak* closed unital *-subalgebra, then B may be viewed also as a unital
C*-subalgebra of B(H), whose weak* closure in B(H) is (the copy of) B**. In
this case we shall say that B is represented on H universally. (The explanation
for this term is that the well-known 'universal representation' tru of a C*-algebra is
'universal' in our sense, and conversely if 7r is a representation which is 'universal'
in our sense then 7r(B)" is isomorphic to 7ru(B)" ~ B**. See [27] Section 1.)
If, further, e E B** is a projection for which (1) holds, then with respect to the
splitting H = eH EEl (1- e)H we may write

T(a) = [7ro() 0]
SO '
for all a E A. We will see that this is essentially true even if T(IA) -lIB:

1We remark in passing that one does not need the full strength of the Arveson-Hamana
theorem here, one may use the much simpler [8] Theorem 4.1.
92 DAVID P. BLECHER AND DAMON M. HAY

THEOREM 3.1. Let T : A --> B be a completely contractive linear map from a


unital operator algebra into a unital C* -algebra. Then the following are equivalent:
(i) T is a complete isometry,
(ii) There is a partial isometry u E B** with initial projection e E B**, and
a (completely isometric) 1-1 *-homomorphism 1r : C;(A) --> eB**e with
1r(1) = e, such that for all a E A

T(a)e = tL1r(a) and 1r(a) -


= u*T(a).
Moreover e may be taken to be a 'closed projection' (see [25] 3.11, and the
discussion towards the end of our proof).
(iii) If H is a Hilbert space on which B is represented universally, then there ex-
ist two closed subspaces E, P of the Hilbert space H, a 1-1 *-homomorphism
1r : C;(A) --> B(E) with 1r(1) = IE, and a unitary u : E --> P, such that

T(a)IE = u1r(a),

and T(a)IE.L C p.l., for all a E A. Here E.l. for example is the orthocom-
plement of E in H.
(iv) If H is as in (iii), then there exists two closed subspaces E, P of H, a
unital 1-1 *-homomorphism 1r : C;(A) --> B(E), a complete contraction
S : C; (A) --> B( E.l. , p.l.), and unitary operators U : E EEl p.l. --> Hand
V : H --> E EEl E.l., such that

T( ) - U [ 1r(a) 0 ] V
a - 0 S(a)

for all a E A.
(v) There is a left ideal J of B, a 1-1 *-homomorphism 1r from C;(A) into
a unital subspace of B I (J + J*) which is a C* -algebra, and a 'partial
isometry' u in B I J such that

qJ(T(a)) = u1r(a) &

for all a E A, where qJ is the canonical quotient map B --> BIJ.

Before we prove the theorem, we make several remarks. First, we have taken B
to be a C* -algebra; however since any unital operator algebra is a unital subalgebra
of a unital C* -algebra this is not a severe restriction. We also remark that there
are several other items that one might add to such a list of equivalent conditions.
See [5, 6]. Items (ii)-(iv), and the proof given below of their equivalence with (i),
are new. We acknowledge that we have benefitted from a suggestion that we use
the Paulsen system to prove the result. This approach is an obvious one to those
working in this area (Ruan and Hamana used a variant of it in their work in the
'80's on complete isometries and triple morphisms [28, 14]). However we had not
pushed through this approach in the original version of [5] because this method
does not give several of the results there as immediately. Statement (v) above has
been simply copied from [5, 6] without proof or explanation. We have listed it
here simply because Theorem 1.3 may be particularly easily derived from it as the
special case when A and B are commutative (see comments below). Note that (iii)
above resembles Theorem 1.2 superficially.
COMPLETE ISOMETRIES - AN ILLUSTRATION 93

PROOF. The fact that the other conditions all imply (i) is easy, following the
idea in the paragraph above the theorem, namely by using the fact that a 1-1
*-homomorphism is completely isometric.
In the remainder of the proof we suppose that T is a complete isometry. We
view A as a unital subalgebra of C;(A) as outlined in Section 3. We define a subset
S(B) of M 2(B) as in Example 3 in Section 2. Similarly define a subset S(T(A))
of S(B) using a similar formula (note that S(T(A)) has 1-2 entries taken from
T(A) and 2-1 entries taken from T(A)*). Similarly we define the subset S(A) of
the CO-algebra M2(C;(A)) (Le. S(A) has scalar diagonal entries and off diagonal
entries from A and A*). We write 1 EEl 0 for the matrix in S(A) with 1 as the 1-2
entry and zeroes elsewhere. Similarly for 0 EEl 1. We also use these expressions for
the analogous matrices in S(B). The map q, : S(A) -+ S(T(A)) c M 2(B) taking

[~! :1] ~ [T~)* ~~~)]


is well known to be a unital complete isometry (this is the well known Paulsen
lemma, see the proof of 7.1 in [23]). Let C be the C' -subalgebra of M2 (B) generated
by S(T(A)). The CO-envelope of S(A) is well known to be M2(C;(A)) (see Example
3 in Section 2 where we proved this in the case that A is already a C* -algebra, or
for example [3] Proposition 4.3 or [30]). Thus by the Arveson-Hamana theorem
we obtain a surjective *-homomorphism () : C -+ 1V/z(C;(A)) such that () 0 q, is
simply the canonical embedding of S(A) into M2(C;(A)). As in the special case
considered above the theorem, we let 10 be the kernel of the mapping (), then C /10
is a unital CO-algebra *-isomorphic to M2(C;(A)). Indeed there is the canonical
*-isomorphism "(: M2(C;(A)) -+ Clio induced by (), taking

[~! :1] ~ [T~:)* T~~)] + 10 ,

As in the simple case above the theorem, C /10 may be viewed as a C* -subalgebra of
PoC"Po, for a central projection Po E C** (namely, the complementary projection
to the support projection of 10)' Now PoC**Po C C** C M2(B)**, and it is well
known that M2(B)** ~ M2(B**) as CO-algebras. Thus we may think of C** as
a C* -subalgebra of .1112 (B**). Also, C'* contains C as a C* -subalgebra, and the
projections 1 EEl 0 and 0 EEl 1 in C correspond to the matching diagonal projections
1 EEl 0 and 0 EEl 1 in M2(B**). These last projections therefore commute with Po,
since Po is central in C**, which immediately implies that Po is a diagonal sum
fEEl e of two orthogonal projections e, f E B**. Thus we may write the C* -algebra
POM2(B**)po as the C*-subalgebra

[ fB** f f B**e ]
eB** f eB*'e
of M2(B**). We said above that Clio may be regarded as a C*-subalgebra of the
subalgebrapoM2(B**)po of M2(B*'). Thus the map "( induces a 1-1 *-homomorphism
III : M2(C;(A)) -+ .!I1z(B**). It is easy to check that 1lI(1 EEl 0) = fEEl 0 and
III (0 EEl 1) = 0 EEl e. Since III is a *-homomorphism it follows that III maps each of the
four corners of M 2(C;(A)) to the corresponding corner of POM2(B**)po C M2(B**).
We let R: C;(A) -+ fB**e be the restriction of III to the '1-2-corner'. Since III is
1-1, it follows that R is 1-1. If 7r is the restriction of III to the '2-2-corner', then 7r is
a *-homomorphism C;(A) -+ eB**e taking lA to e. Applying the *-homomorphism
94 DAVID P. BLECHER AND DAMON M. HAY

111 to the identity

[~ ~][~ ~]=[~ ~]
we obtain that u = R(I) is a partial isometry, with u*u = 71"(1) = e. Similarly
uu* = f. A similar argument shows that R(a) = R(I)7I"(a) for all a E C;(A). Thus
u* R(a) = u*u7l"(a) = 7I"(a) for all a E C;(A).

---- ----
Next, we observe that 111 takes the matrix z which is zero except for an a from
A in the 1-2-corner, to the matrix w = Pocp(z)Po. Since cp(z) E C** and Po is in
the center of that algebra, we also have w = cp(z)Po = Pocp(z). Also w viewed as
a matrix in M2(B**) has zero entries except in the .1-2-corner, which (by the last
sentence) equals
ff(;;) e = f(;;) e = ff(;;).
Also using these facts and a fact from the end of the last paragraph we have
--... ----* - ---*-
u*T() = R(I)*T() = (fT(I)e)*T() = eT(I) fT() = eT(I) T()e = u* R() = 71".
Thus
- - -
T()e = fT() = uu*T() = U7l"(')'
We have now also established most of (ii). One may deduce (iii) from (ii) by viewing
B c B** c B(H), and setting E = eH, and F = (uu*)H. We also need to use
facts from the proof above such as u*u = e. Clearly (iv) follows from (iii). As we
said above, we will not prove (v) here.
Claim: if e is the projection in (ii) above, then 1 - e is the support projection
for a closed ideal I of a unital *-subalgebra D of B. Equivalently (as stated at
the start of this section), there is a (positive increasing) contractive approximate
identity (bd for I, with bt ~ 1 - e in the weak* topology. This claim shows that
1 - e is an 'open projection' in B**, so that e is a closed projection, as will be
obvious to operator algebraists from [25] section 3.11 say. For our other readers we
note that for what comes later in our paper, one can replace the assertion about
closed projections in the statement of Theorem 3.1 (ii) with the statement in the
Claim above.
To prove the Claim, recall from our proof that Po = fEB e = Ie - Pi, where Pi
is the support projection for a closed ideal 10 of C. Thus Pi = (1 - f) EB (1 - e).
As stated at the start of Section 3, Pi is the weak* limit in C**, and hence also
in M 2(B**), of a contractive approximate identity (et) of 10, By the separate
weak* continuity of the product in a von Neumann algebra, it follows that the net
bt = (OEB l)et(OEB 1) has weak* limit (OEB I)Pl (OEB 1) = OEB (1- e). Viewing these as
expressions in B, the above says that bt ~ 1 - e weak* in B* *. View (0 EB 1) C (0 EB 1)
as a *-subalgebra D of B, and view (0 EB 1)10(0 EB 1) as a two sided ideal I in D. It
is easy to see that (bd is a contractive approximate identity of I. Thus it follows
that 1 - e is the support projection of the ideal I. 0
Some applications of results such as Theorem 3.1 may be found in [6].
Next we discuss briefly the relation between our noncommutative characteri-
zation of complete isometries (for example Theorem 3.1 above), and Theorem 1.3.
Our point is not to provide another proof for Theorem 1.3 - the best existing proof
is certainly short and elegant. Rather we simply wish to show that the noncommu-
tative result contains 1.3. Indeed Theorem 1.3 quite easily follows from Theorem
COMPLETE ISOMETRIES - AN ILLUSTRATION 95

3.1 (v). Since however we did not prove Theorem 3.1 (v), we give an alternate
proof.
COROLLARY 3.2. Let A, B be a unital function algebras, with B selfadjoint.
Then condition (ii) in Theorem 3.1 implies condition (A) in Theorem 1.3.
PROOF. By hypothesis, T()e = U1l"('), and u*u = e = 11"(1) so that u = u1l"(1) =
T(I)e. Thus eT(I)*T()e = u*U1l"(') = 11"(1)11"(') = 11", so that Ran 11" C eBe = Be
(note B** is commutative in this case). From [25] 3.11.10 for example, the 'closed
projection' e in B** corresponds to a closed ideal J in B whose support projection
is 1 - e. Alternatively, to avoid quoting facts from [25], we will also deduce this
from the 'Claim' towards the end of the proof of Theorem 3.1. If I is the ideal
in that Claim, let J be the closed ideal in B generated by I. Since J = Bl, the
contractive approximate identity of I is a right contractive approximate identity of
J. Thus J has support projection 1 - e too, by the first paragraph of Section 3
above.
By facts in the just quoted paragraph, we have a canonical unital 1-1 map
'11 : B / J ~ B** taking the equivalence class b + J of b E B to ebe. Indeed in this
commutative case we see by inspection that '11 is a *-homomorphism from the C*-
algebra B/J onto the C*-subalgebra M = eBe of B**. Define O(a) = '11-1(1I"(a)),
this is a 1-1 *-homomorphism A ~ B/J. Since 11"(1) = e, 0 is a unital map too.
Since uu* = u*u = e, u is unitary in M, and so 'Y = '11-1(U) is unitary in B/J.
Note also that T(a)e = '11(T(a) + J). Applying '11- 1 to the equation T()e = U1l"('),
we obtain qJ(T(a)) = 'Y O(a), that is, condition (A) in Theorem 1.3. D
If one attempts to use the ideas above to find a characterization analogous to
condition (A) from Theorem 1.3 but in the noncommutative case, it seems to us
that one is inevitably led to a condition such as (v) in Theorem 3.1.
We address a paragraph to experts, on generalizations of the proof of Theorem
3.1. Consider a complete isometry between possibly non-unital C* -algebras. Or
much more generally, suppose that T is a complete isometry from an operator space
X into a C* -triple system W. One may form the so called 'linking C* -algebra' of W,
with the identities of the 'left and right algebras of W' adjoined. Call this C'(W).
As in the proof of Theorem 3.1 we think of S(W) c .c'(W). Similarly, if Z is the
'triple envelope' of X (or if X = Z is already a C*-algebra or C*-triple system),
then we may consider S(X) C S(Z) c .c'(Z). As in the proof of Theorem 3.1 we
obtain firstly a unital complete isometry <I> : S(X) ~ S(T(X)) c .c'(Z), and then
a unital 1-1 *-homomorphism 11" : .c'(Z) ~ .c'(W)**. By looking at the 'corners' of
11" we obtain projections e, f in certain second dual von Neumann algebras, so that
fT( )e is (the restriction to X of a completely isometric) a 1-1 triple morphism into
W**. In fact we have precisely such a result in [5] (see Section 2 there), but the key
point is that the new proof gives different projections e, f, which are more useful
for some purposes.

4. Complete isometries versus isometries


Finally, as promised we discuss why we believe that in this setting of nonsurjec-
tive maps between C* -algebras say, general isometries are not the 'noncommutative
analogue' of isometries between function algebras. The point is simply this. In the
function algebra case we can say thanks to Holsztynski's theorem that the isome-
tries are essentially the maps composed of two disjoint pieces Rand S, where R is
96 DAVID P. BLECHER AND DAMON M. HAY

isometric and 'nice', and S is contractive and irrelevant. However at the present
time it looks to us unlikely that there ever will be such a result valid for general non-
surjective isometries between general C* -algebras. The chief evidence we present
for this assertion is the very nice complementary work of Chu and Wong [9] on
isometries (as opposed to complete isometries) T : A -+ B between CO-algebras.
They show that for such T there is a largest projection p E B** such that T()p
is some kind of Jordan triple morphism. This appears to be the correct 'struc-
ture theorem', or version of Kadison's theorem [11], for nonsurjective isometries.
However as they show, the 'nice piece' R = T()p is very often trivial (Le. zero),
and is thus certainly not isometric. Thus this approach is unlikely to ever yield a
characterization of isometries. A good example is A = M 2 , the smallest noncom-
mutative C* -algebra. Simply because A is a Banach space there exists, as in the
discussion in the first paragraph of our paper, a linear isometry of A into a C (K)
space. However it is easy to see that there is no nontrivial *-homomorphism or
Jordan homomorphism from A into a commutative C* -algebra. Such an isometry
is uninteresting, and this is perhaps because the interesting 'nice part' is zero. Thus
we imagine that the 'good noncommutative notions of isometry' are either complete
isometries or the closely related class of maps for which the piece T()p from [9] is
an isometry.
This leads to three questions. Firstly, can one independently characterize the
last mentioned class? Secondly, if T is a complete isometry, then is the projection p
in the last paragraph equal (or closely related) to our projection e above? Finally,
H. Pfitzner has remarked to us, there is already a gap between the isometry and
the 2-isometry cases (not only isometries and complete isometries). It would be
interesting if there were a characterization of 2-isometries.

References
[1] J. Araujo and J. J. Font, Linear isometries between subspaces of continuous functions, Trans.
Amer. Math. Soc. 349 (1997), 413-428.
[2] W. B. Arveson, Subalgebms ofC*-algebms, Acta Math. 123 (1969), 141-224; II, 128 (1972),
271-308.
[3] D. P. Blecher, The Shilov boundary of an opemtor space, and the chamcterization theorems,
J. Funct. An. 182 (2001),280-343.
[4] D. P. Blecher, E. G. Effros and V. Zarikian, One-sided M-Ideals and multipliers in opemtor
spaces. 1. To appear Pacific J. Math.
[5] D. P. Blecher and D. Hay, Complete isometries into C* -algebms,
http://front.math.ucdavis.edu/math.OA/0203182, Preprint (March '02).
[6] D. P. Blecher and L. E. Labuschagne, Logmodularity and isometries of opemtor algebms, To
appear, Trans. Amer. Math. Soc ..
[7] D. P. Blecher and C. Le Merdy, Opemtor algebms and their modules - an opemtor space
approach, To appear, Oxford Univ. Press.
[8] M. D. Choi and E.G. Effros, The completely positive lifting problem for C-algebms, Ann.
Math. 104 (1976), 585-609.
[9] C-H. Chu and N-C. Wong, Isometries between C -algebms, Preprint, to appear Revista
Matematica Iberoamericana.
[10] J. B. Conway, A Course in Opemtor Theory, AMS, Providence, 2000.
[11] E. G. Effros and Z. J. Ruan, Opemtor Spaces, Oxford University Press, Oxford (2000).
[12] R. J. Fleming and J. E. Jamison, Isometries on Banach spaces: function spaces, Book to
appear, CRC press.
[13] M. Hamana, Injective envelopes of opemtor systems, Pub!. R.I.M.S. Kyoto Univ. 15 (1979),
773-785.
COMPLETE ISOMETRIES - AN ILLUSTRATION 97

[14] M. Hamana, Triple envelopes and Silov boundaries of operator spaces, Math. J. Toyama
University 22 (1999), 77-93.
[15] W. Holsztynski, Continuous mappings induced by isometries of spaces of continuous func-
tions, Studia Math. 24 (1966), 133-136.
[16] K. Jarosz and V. Pathak, Isometries and small bound peturbations of function spaces, In
"Function Spaces", Lecture Notes in Pnre and Applied Math. Vol. 136, Marcel Dekker (1992).
[17] R. V. Kadison, Isometries of operator algebras, Ann. of Math. 54 (1951), 325-338.
[18] E. Kirchberg, On restricted peturbations in inverse images and a description of normalizer
algebras in C*-algebras, J. Funct. An. 129 (1995), 1-34.
[19] E. Kirchberg and S. Wassermann, C*-algebras generated by operator systems, J. Funct.
Analysis 155 (1998), 324-351.
[20] A. Matheson, Isometries into function algebras, To appear.
[21] M. Nagasawa, Isomorphisms between commutative Banach algebras with an application to
rings of analytic functions, Kodia Math. Sem. Rep. 11 (1959), 182-188.
[22] W. P. Novinger, Linear isometries of subspaces of continuous functions. Studia Math. 53
(1975), 273-276.
[23] V. I. Paulsen, Completely bounded maps and dilations, Pitman Research Notes in Math.,
Longman, London, 1986.
[24] V. I. Paulsen, Completely bounded maps and operator algebras, To appear Cambridge Uni-
versity Press.
[25] G. Pedersen, C*-algebras and their automorphism groups, Academic Press (1979).
[26] G. Pisier, Introduction to operator space theory, To appear Camb. Univ. Press.
[27] M. Rieffel, Morita equivalence for C*-algebras and W*-algebras, J. Pure Appl. Algebra 5
(1974), 51-96.
[28] Z. J. Ruan, Subspaces ofC*-algebras, Ph. D. thesis, U.C.L.A., 1987.
[29] E. L. Stout, The theory of uniform algebras, Bogden and Quigley (1971).
[30] C. Zhang, Representations of operator spaces, J. Oper. Th. 33 (1995), 327-351.

DEPARTMENT OF MATHEMATICS, UNIVERSITY OF HOUSTON, HOUSTON. TX 77204-3008


E-mail address, David P. Blecher: dblecher~ath. uh. edu
E-mail address, Damon Hay: dhayOmath.uh.edu
Contemporary Mathematics
Volume 328, 2003

Some Recent Trends and Advances in Certain Lattice


Ordered Algebras

Karim Boulabiar, Gerard Buskes, and Abdelmajid Triki

ABSTRACT. In this paper we give a survey, intended as both a supplement


as well as an update to a survey by Huijsmans [57], with results that have
been obtained in the last ten years on Archimedean lattice ordered algebras.
Special attention is paid to I-algebras, almost I-algebras and d-algebras and
problems that were posed in the survey by Huijsmans about these special
classes of lattice ordered algebras.

CONTENTS

l. Introduction
2. Definitions and elementary properties
3. i-algebra multiplications in C (X)
4. Multiplication by an element as an operator
5. Uniform completion and Dedekind completion
6. Powers in i-algebras
7. Functional Calculus on f-algebras
8. Relationships between i-algebra multiplications
9. Connection between algebra and Riesz homomorphisms
10. Positive derivations
11. Cauchy-Schwarz inequalities
12. Order biduals
13. Ideal theory
14. Representation of f-algebras
15. Linear biseparating maps on f-algebras
References

1991 Mathematics Subject Classification. 06F25, 13J25, 16W80, 46A40, 46840, 46842,
46E25, 47L07, 47847, 47865.
Key words and phrases. almost I-algebra, algebra homomorphism, I-algebra, d-algebra, lat-
tice ordered algebra, order ideal, orthomorphism, representation theory, Riesz homomorphism,
ring ideal, space of continuous functions, uniformly complete Riesz space.
The second named author gratefully acknowledges support from an Office of Naval Research
Grant with number N00014-01-1-0322. Part of this survey was written while the first named
author was visiting the University of Mississippi in the Spring of 2002.

2003 American Mathematical Society


99
~!3QtJLA8lAR, GERARD BUSKES, AND ABDELMAJID TRIKI

1. Introduction
ftehistory of lattice ordered vector spaces (so called Riesz spaces or vector
lattices) goes back to Riesz and the International Congress of Mathematicians in
Bologna in 1928. A study of the most important class of lattice ordered algebras
(but not their name), J-algebras, was initiated by Nakano in [75] for a-Dedekind
complete ordered vector space in 1951, subsequently in 1953 by Amemiya in [3],
and finally with its present definition and name in 1956 by Birkhoff and Pierce
in [22]. A precise date for the very first definition of lattice ordered algebras in
general is very hard to pinpoint, but originated at around the same time as the
previous three references. Indeed, in his review [44] of Birkhoff's 1950 address to
the International Congress of Mathematicians in Cambridge, Massachusetts, Fl.ink
observed that a general study of lattice ordered rings seems to be needed to study
what are now called averaging or Reynolds operators. A call for lattice ordered
rings also had gone out by Birkhoff himself in the form of a listed problem at the
end of his seminal 1942 paper [20] on lattice ordered groups. Thus lattice ordered
algebras and J-algebras seem to have multiple origins, including a study of averaging
operators, which themselves sprang forth from problems in fluid mechanics. An
appearance at about the same time of J-rings and J-algebras has not resulted in
an historical development on complete common ground for these objects. This
is not unlike the development of lattice ordered groups versus the development
of Riesz spaces. Where the latter have attracted attention from researchers in
analysis, the former have been more widely investigated by algebraists. A similar
divided attention from analysis versus algebra seems to underlie the connected but
somewhat separate tracks of lattice ordered rings versus lattice ordered algebras.
Though this separation of tracks is to some extent unavoidable, where each track
does have ground that is truly its own, some overlap in results does exist, resulting
in difficulties making accurate literature attributions in a survey like ours. We are
grateful to two referees for pointing at some references that were missing in our
manuscript, though we take full responsibility for possible remaining omissions in
the reference list. This survey places itself almost completely on the track of lattice
ordered algebras and our only apology for not linking algebra facts in a systematic
way to ring results is that all three of us authors were trained as analysts. There is a
natural back and forth between the two theories, in one direction by forgetting some
of the structure, and in the other by finding, so to speak adjointly, an enveloping
algebra. A nice survey on J-rings was written by Henriksen in 1995 (see [50]),
to which we refer the interested reader for linkage to some of what follows in this
survey. Historically, a lot of the credit for a revival of the theory of J-algebras
points to the highly motivating Arkansas Lecture Notes by Luxemburg [67] and the
1982 Ph.D. thesis of de Pagter [76], who systematically explored both the existing
literature as well as new directions. Another impetus to research in the area of
J-algebras derived from the desire of Zaanen, who in the late seventies started to
develop a program to prove many of the elementary results in the theory of Riesz
spaces without using representation theorems for vector lattices. This desire is
directly linked with a preference not to use the Axiom of Choice unnecessarily. The
present survey is intended as an update to the one by Huijsmans [57]. We hasten
to point out that we do not intend this survey to replace the one by Huijsmans,
but rather that we think of it as augmenting part of it. Since [57] appeared,
much progress has been made and several of the problems explicitly phrased in
LATTICE ORDERED ALGEBRAS 101

[57] have been solved. At the same time, some topics like ideal theory, connections
between Riesz homomorphisms and algebra homomorphisms, and representations
of f-algebras were absent in [57]. Thus an update as well as a supplement was
needed. However, some sections of [57] receive no attention at all in this survey. We
have not included important topics like the role of f-algebras in positive operator
theory (e.g., we do not even include results on the previously mentioned averaging
operators) and probability theory. We are rather focused on placing this update,
as much as possible, in the setting of spaces of functions, hoping to interest as
large an audience as possible via this approach. Moreover, we feel that the great
source of inspiration was and continues to be the beautiful book by Gillman and
Jerison [47], which certainly inspired and continues to inspire a large part of the
research in lattice ordered algebras and rings. Last but not least we focus on what
could be called distortions of f-algebras, in particular on Archimedean almost f-
algebras and d-algebras. Though these distortions have the potential to be seen as
aberrations by some, we believe they point the way to techniques that are needed
for the broader theory of lattice ordered algebras, as well as for illumination of
various aspects in the theory of f-algebras. Note that the distortions disappear if
the lattice ordered algebra under consideration has a multiplicative identity which is
a weak order unit. Indeed, such algebras are automatically f-algebras. It should be
mentioned that classes of lattice ordered algebras other than the ones that appear
in this survey have been studied. Notably, the papers [85], [86], [87], and [88] by
Steinberg discuss lattice ordered algebras in which every square is positive, a class
of algebras that includes all almost f-algebras. We also pay no attention at all to
non-Archimedean lattice ordered algebras. Finally, we point out that several results
in this survey rely heavily on the (relative) uniform topology on Riesz spaces. In
particular, since in Archimedean Riesz spaces uniform limits are unique, we shall
include the 'Archimedean' property in the definition of uniformly complete Riesz
spaces. A complete investigation of that topology can be found in Sections 16 and
63 of [69].
For terminology and concepts not explained or proved in this survey we refer
the reader to the standards books [2], [47], [69], [72], [92] and [93].

2. Definitions and elementary properties


A (real) Riesz space A is called a lattice ordered algebra (briefly, an i-algebra)
if A also is an algebra and the positive cone
A+ = {J E A : f ~ O}
is closed under multiplication, that is,
if f,g ~ 0 then fg ~ 0 (equivalently, if Ifgl :::; Ifllgl for all f,g E A).
We make the following blanket assumption: all Riesz spaces under consider-
ation in this paper are assumed to be Archimedean (however, the latter
blanket assumption has not stopped us to explicitly add the word Archimedean to
the list of conditions in various results below).
After B.irkhoff and Pierce (see [22, p. 55]), we define an i-algebra A to be an
f-algebra if for every f, 9 E A, the condition
f /\g = 0 implies Uh) /\g = (hi) /\g = 0 for all h E A+
102 KARIM BOULABIAR. GERARD BUSKES, AND ABDELMA.lID TRIKI

holds. We call the E-algebra A an almost f-algebra after Birkhoff in [21, Section 6]
if
f /\ g = 0 ill A implies fg = O.
An E-algebra A for which

f /\ 9 = 0 in A and h E A+ imply (.fh) /\ (gh) = (hf) /\ (hg) = 0

is called a d-algebra. The notion of d-algebra goes back to Kudlacek in [65]. Our
focus in this survey on E-algebras is almost exclusively on f -algebras, almost f-
algebras and d-algebras.
In this paragraph, we recall some properties of f-algebras. Using the Axiom
of Choice, Birkhoff and Pierce in [22, Theorem 13] proved that any f-algebra is
commutative. A constructive proof of this fact, due to Zaanen, can be found in [61,
Theorem 2.1] or [92, Theorem 140.10]. All squares in an f-algebra are positive.
Also, Ifgl = Ifllgl for all f,g in an f-algebra A. The multiplication by an element in
the f -algebra A is order continuous, i.e., if inf {fT : T} = 0 in A then inf {g fT : T} =
o for all 9 E A +. Phrased more generally, the multiplication 7rf by an element
f E A (7rf (g) = f 9 for all 9 E A) is an orthomorphism and all orthomorphisms
are order continuous. Recall that an orthomorphism on a Riesz space L is an order
bounded linear operator 7r such that 17r (f)I/\ Igl = 0 whenever Ifl/\ Igl = 0 in L
(the reader is referred to [2] or [92] for elementary properties of orthomorphisms).
There is another important relationship between orthomorphisms and f-algebras,
which we mention next. Indeed, let Orth (L) be the set of all orthomorphisms on
a Riesz space L. Under the operations and the ordering inherited from b (L), the
ordered algebra of all order bounded operators on L, and under composition as
multiplication, Orth (L) is an Archimedean f-algebra with the identity map h on
L as unit element. The details of the facts recalled above can all be found in [2],
[76] or [92].
Next we present some properties of almost f-algebras. Almost f-algebras,
like f-algebras, are commutative too. The latter fundamental property was first
established by Scheffold in [80, Theorem 2.1] for almost f-algebras that are Banach
lattices. Using both Scheffold's result and the Axiom of Choice, Basly and Triki
were the first to prove commutativity for arbitrary almost f-algebras [10, Thorme
1.1]. The first proof of the commutativity for almost f-algebras within Zermelo-
Fraenkel set theory was given by Bernau and Huijsmans in their paper [13, Theorem
2.15]. Recently, a shorter constructive proof was published in [34, Corollary 3] by
Buskes and van Rooij. Another property of f-algebras holds for almost f-algebras,
namely the positivity of squares. Also, if A is an almost f-algebra then f2 = Ifl2
for all f E A. However, contrary to the order continuity of the multiplication in f-
algebra..'l, the multiplication by a fixed element in an almost f-algebra is not always
order continuous as is shown in the following example.

EXAMPLE 2.1. Write A = C ([0,1]), the vector space of all real-valued contin-
uous functions on [0, 1]. With respect to the pointwise ordering (i. e., f :::; 9 in A if
an only if f (x) :::; 9 (x) for all x E [0,1]), A is an Archimedean Riesz space. Define
a multiplication in A by

{
f(x)g(x) (0 :::; x :::; 1/2) ;
(f. g)(x) = f (1/2) 9 (1/2) (1/2 :::; x :::; 1)
LATTICE ORDERED ALGEBRAS 103

lor all I, 9 E A. Then A is an almost I-algebm with respect to the multiplication


. For every natuml number n :::: 1, define the function In E A by
I (O :::; x :::; 1/2 - l/n) ;
f (:1") _ { n/2 - n:1: (1/2 - l/n :::; x :::; 1/2) ;
" ,- -n/2 + nx (1/2:::; x :::; 1/2 + l/n) ;
1 (1/2+1/n:::;x:::;1).
Then sup Un : n = 1,2, ... } exists in A and equals the function e defined bye (x) = 1
lor all x E [0,1]. On the other' hand,
(O :::; x :::; 1/2) ;
(1/2:::;x:::;1)
lor all n E {I, 2, ... }, and clearly the set {e. In : n = 1,2, ... } does not have a supre-
mum in A. We conclude that. is not order continuous.
For more information about elementary theory of almost f-algebras, the reader
is encouraged to consult [13], [23], [34], and [35].
At this point, we turn our attention to some properties of d-algebras. It follows
directly from the definition of d-algebras that an i-algebra A is a d-algebra if and
only if the multiplication map induced by any fixed element in A + is a Riesz (or
lattice) homomorphism. It follows that a necessary and sufficient condition for an
i-algebra A to be a d-algebra is that the identity If gl = 1/IIgi holds for all f, 9
in A. Contrary to (almost) I-algebras, d-algebras need not be commutative nor
have positive squares. Next we give an example of a non-commutative d-algebra in
which not all squares are positive.
EXAMPLE 2.2. Let in this example A be the algebra of real {2 x 2)-matrices of
the form

(~ g)
with the usual addition, scalar multiplication, matrix product and partial ordering.
It is not hard to see that A is an Archimedean d-algebra. But A is not commutative
and not all squares in A are positive. Indeed, if

p= (~ ~) and q = (~ ~)
then
pq =q and qp = O.
Moreover, the square

is not positive.
As for almost f-algebras, multiplication by a fixed element in a d-algebra is,
in general, not order continuous. Point in case is the almost I-algebra that we
considered in Example 2.1, which also is a d-algebra. Our main reference about
d-algebras is [13].
In what follows, we look at some of the connections between the three kinds of
i-algebras that we consider in this paper. It is immediate that any I-algebra is both,
an almost f-algebra and a d-algebra. Almost f-algebras need not be d-algebras as
we see in the next example.
104 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

EXAMPLE 2.3. Take A as in Example 2.1, and let () E A be the function


() (x) = { 1/2 - x (0::; x ::; 1/2) ;
x - 1/2 (1/2::; x ::; 1).
For f, 9 E A, define

(f
-9
)( ) _ {
x -
/,1-X f(s)g(s)ds
1/2
(0::; x ::; 1/2);
() (x) f (x) 9 (x) (1/2::; x ::; 1).
Then A is an almost f -algebra under the multiplication -. However, A is not a
d-algebra. Indeed, let e, f in A be defined by
e(x)=l forallxE[O,l]
and
f (x) ={ -4x + 1 (0::; x ::; 1/2) ;
4x-3 (1/2::;x::;1).

Observe that If - el (0) = 0 and (If I- e) (0) = /,1 1/2


If(s)1 ds -; O. Thus the property

If - gl = If I- Igl fails in A.
Since every almost f -algebra is commutative and has positive squares, Example
2.2 shows that d-algebras need not be almost f-algebras. However, if a d-algebra A
is commutative or has positive squares then A is automatically an almost f-algebra
[22, p. 60]. Summarizing part of the relations, we have the following diagram

f-algebra::::} commutative d-algebra ::::} almost f-algebra.


For more detail, see [22], [13] and [57].
The next lines deal with nilpotent element in i-algebras. The set of all nilpotent
elements in the i-algebra A is denoted by N (A). In other words,
N (A) = {f E A : r = 0 for some n = 1, 2, ... } .
Given a natural number p, we define
Np (A) = {f E A : fP = O} .
The i-algebra A is said to be semiprime if 0 is the only nilpotent element in A,
that is, if N (A) = {O} . If A is an f-algebra then the following equalities hold
N (A) = N2 (A) = {f E A : fg = 0 for all 9 E A}.
(see [76, Proposition 10.2] or [92, Theorem 142.5]). If A is an almost f-algebra
then
N (A) = N3 (A) = {J E A : fg2 = 0 for all 9 E A}
= {f E A : fgh = 0 for all g, hE A}
(see [51, Theorem 3.11]) and, as for f-algebras,
N2 (A) = {f E A : fg = 0 for all 9 E A}
(see [23, Lemma 5.3]). If A is a d-algebra then
N (A) = N3 (A) = {f E A : gfh = 0 for all 9 A}
(see [13, Theorem 5.5] or [29, Theorem 5]). However, and co ry to what
we documented for f-algebras and almost f-algebras, in d-algebl he equality
LATTICE ORDERED ALGEBRAS 105

N2 (A) = {J E A: fg = 0 for all 9 E A} does not necessarily hold as is illustrated


by the following example.
EXAMPLE 2.4. Take A as in Example 2.1 and define the multiplication. in A
by
(f. g)(x) = f (0) 9 (1) for all f, 9 E A.
Let f E A be the function defined by f (x) = 1 - x for all x E [0,1]. Clearly,
f. f = 0 but f. e i= 0 where e E A is defined bye (x) = 1 for all x E [0,1].
Finally, note that any f-algebra with multiplicative identity is semiprime and
any semiprime almost f-algebra or semiprime d-algebra is automatically an f-
algebra (see Section 1 in [13]). As a final comment we remark that an l-algebra
which has positive squares and has a multiplicative identity need not be an f-
algebra.

3. i-algebra multiplications in C (X)


Let C (X) be the set of all real-valued continuous functions on a compact
Hausdorff' topological space X. Under pointwise addition and scalar multiplication,
C (X) is a real vector space. Moreover, C (X) is an Archimedean Riesz space with
respect to the pointwise ordering (i.e., f ~ 9 in C (X) if and only if f (x) ~ 9 (x)
for all x E X). By defining the multiplication in C (X) pointwise as well (Le.,
(fg) (x) = f(x)g(x) for all f,g E C(X) and all x E X), the space C(X) is easily
seen to have the structure of an f-algebra with e as unit element, where e (x) = 1
for all x E X.
Now consider another associative multiplication. in C (X). The main topic
of this section is to produce necessary and sufficient conditions for C (X) to be an
f-algebra (respectively, an almost f-algebra, a d-algebra) with respect to this new
multiplication . The first theorem in this direction goes back to Conrad (see [38,
Theorem 2.2]), who obtained the following.
THEOREM 3.1. Let. be an associative multiplication in C (X). Then C (X) is
an f -algebra with respect to if and only if there exists a positive function w E C (X)
such that
(f.g)(x) =w(x)f(x)g(x)
for all f,g E C (X) and all x EX.
In fact, Conrad established the theorem above for any Archimedean f-ring with
unit element. The representation formula given in Theorem 3.1 above was obtained
in an alternative way by Scheff'old in [80, Korollar 1.4]. While Conrad's proof is
purely algebraic and order theoretic, the proof presented by Scheff'old relies on
analytic tools like the Riesz representation theorem. With the same analytic tools
Scheff'old also obtained the following representation theorem for almost f-algebra
multiplications in C (X) (see [80, Theorem 1.2]).
THEOREM 3.2. Let. be an associative multiplication in C (X). Then C (X) is
an almost f -algebra with respect to if and only if there exists a family (/Jx : x E X)
of positive measures such that

(f. g)(x) = Ix f(s) 9 (s) d/Jx (s)

for all f, 9 E C (X) and all x E x.


106 KARIM BOULABIAR, GERARD BUSKES. AND ABDELMAJID TRIKI

Since every commutative d-algebra is an almost I-algebra, the previous theorem


remains valid for commutative d-algebra multiplications in C (X) as well. Recently
in [25, Corollary 3.2], Boulabiar proved the following representation formula for
any (not necessarily commutative) d-algebra multiplication in C (X).
THEOREM 3.3. Let. be an associative multiplication in C (X). Then C (X) is
a d-algebra with respect to if and only if there exist
(i) a positive function wE C (X), and
(ii) functions h,k: X -+ X (continuous on coz(w) = {x EX: w(x)"! O})
such that
(J. g) (x) = w (x) I (h (x)) 9 (k (x))
for all I,g E C(X) and all x E X.
Notice that if C (X) is a d-algebra with respect to the multiplication. then
is commutative if and only if the functions hand k coincide on coz (w) (where
11" k and ware as in Theorem 3.3). The latter observation yields. in addition to
the formula cited in Theorem 3.2 above, another representation for commutative
d-algebra multiplications on C (X) .
More abstract versions of the results above will be given in Section 8 below.

4. Multiplication by an element as an operator


Let A be an (-algebra and recall that C b (A) denotes the ordered algebra of all
order bounded operators on A. For every f E A, we define the map 7rf on A by
7rf (g) = f 9 for all 9 E A. Clearly, 7rf is an order bounded operator on A for all
lEA. The map p : A -+ Cb (A) defined by p (J) = 7rf for all I E A is obviously
an algebra homomorphism, that is, p (J g) = p (J) p (g) for all I, 9 E A. Hence the
range p (A) of p is a subalgebra of Cb (A). In this section, we will see that if A is an
almost f-algebra then p (A) can canonically be equipped with an ordering, under
which p (A) is an Archimedean I-algebra. A corresponding result will also be given
for commutative d-algebras and f-algebras.
Let A be an almost I-algebra. Since
N2 (A) = {f E A : Ig = 0 for all 9 E A},
7rf= 7r9 if and only if f - 9 E N 2 (A). This allows us to define an ordering on p (A)
by putting
(0)
The ordering defined by (0) coincides with the ordering inherited from Cb (A),
namely, 7rf is positive with respect to (0) if and only if 7rf is a positive operator on
A. Under the usual addition and composition of operators, and with the ordering
defined by (0), p (A) is an Archimedean ordered subalgebra of Cb (A). In fact, we
have the following theorem (see Theorem 4.2 and Theorem 4.4 in [23]).
THEOREM 4.1. Let A be an Archimedean almost I -algebra. Then p (A) is an
Archimedean I -algebra with respect to the addition and composition of operators,
and the ordering inherited from Cb (A). The lattice operations in p (A) are given by
7r f V 7r9 = 7rfV 9' 7r f 1\ 7r9 = 7rf /1.9 for all f, 9 EA.
In particular,
(7rf)+ = 7rf+, (7rf)- = 7rf-' l7rfl = 7rlfl for all lEA.
LATTICE ORDERED ALGEBRAS 107

In other words, p defines a surjective Riesz homomorphism from A onto p (A).


Theorem 4.1 of course holds in commutative d-algebras. Moreover, let A be a
commutative d-algebra. For f E A and 9 E A + the equalities
17I'fl (g) = 71'tft (g) = Iflg = sup{lfllhl : Ihl ::; g}
= sup {Ifhl : Ihl ::;g} = sup {17I'f (h)1 : Ihl ::;g}
imply that 71'f has an absolute value in Lb (A), which coincides with its absolute
value in p (A). We collect the latter observations for commutative d-algebras.
THEOREM 4.2. Let A be an Archimedean commutative d-algebra. Then p (A)
is an Ar'chimedean f -algebra when equipped with the addition and composition of
operators, and the ordering inherited from Lb (A). Moreover, the absolute value 71'tft
of 71'f in p (A) coincides with the absolute value of 71' f in Lb (A) for all f E A. that
is,
17I'fl (g) = 71'tft (g) = sup {17I'f (h)1 : Ihl ::; g} for all 9 E A+.
We obtain the f-algebra case as a corollary.
COROLLARY 4.3. Let A be an Archimedean f-algebra. Then p (A) is an f-
subalgebra of the Archimedean f-algebra Orth (A) of all orthomoTphisms of A.
The fact that the range of p in Corollary 4.3 is an f -algebra was first proved in
[22, Corollary 3, p. 57] by Birkhoff and Pierce, while the fact that Orth (A) itself is
an f-algebra has been proved in [18] by Bigard and Keimel and in [39] by Conrad
and Diem. This topic was also discussed in great detail by de Pagter in his thesis
[76, Proposition 12.1]. Note that if A is an f-algebra then A is semiprime if and
only if p is one-to-one as a map from A into Orth (A). In this case, A and p (A)
are isomorphic a.'! f-algebras. Also, if A is an f-algebra then A has a multiplicative
identity if and only if the map p is one-to-one and onto as a map A ---> Orth (A),
and consequently A and Orth (A) are isomorphic as f-algebra.'!.

5. Uniform completion and Dedekind completion


Let A be an Archimedean i-algebra. The closure A1'U of A in its Dedekind
completion A8 with respect to the uniform topology is a uniformly complete Riesz
space. Using Quinn's Definition 2.12 in [79], A1'U is the uniform completion of A.
The following theorem was obtained by n'iki in [91].
THEOREM 5.1. Let A be an Archimedean f-algebra (respectively, almost f-
algebra, d-algebra, f -algebra). Then the multiplication in A extends uniquely to a
multiplication in A1'U such that A1'U is a uniformly complete i-algebra (respectively,
almost f -algebra, d-algebra, f -algebra) with respect to this extended multiplication.
Moreover, if A is semiprime (respectively, has a unit element e) then ATU is semi-
prime (respectively, has e as unit element).
We now turn our attention to the Dedekind completion of the i-algebra A.
Johnson in his paper [64] proved that if A is an f-algebra (or even an Archimedean
f-ring), then the multiplication in A extends uniquely to an f-algebra multiplication
in A8. The uniqueness of such an extended multiplication in A8 of course arises
from the order continuity of the multiplication in the f-algebra A. Alternative
proofs of this extension can be found in [76, pp. 66-67] and [59, p. 166].
108 KARIM BOULABIAR. GERARD BUSKES. AND ABDELMAJID TRIKI

THEOREM 5.2. Let A be an Archimedean f -algebra. Then the multiplication in


A extends uniquely to a multiplication in A" such that Ad 'is a Dedekind complete f-
algebra with respect to this extended multiplication. FUrthermore, if A is semiprime
(respectively, A has a unit element e) then A" is semiprime (respectively. has e as
unU element).
The corresponding results for (I-algebras in general, or almost f-algebras and
d-algebras in particular, is much harder because of the absence of order continuity
of the multiplication. Nonetheless, extensions of the multiplication to the Dedekind
completion often exist, though such extensions are no longer necessarily unique. For
almost f-algebras Buskes and van Rooij proved the following (see [35, Theorem 10]).

THEOREM 5.3. Let A be an Archimedean almost f-algebra. Then the multipli-


cation in A extends to a multiplication 'in A" such that A" is a Dedekind complete
almost f -algebra with respect to that extended m'ultiplication.

Using the previous result as a starting point, Boulabiar and Chil in [28, Corol-
lary 3] proved that from amongst the extensions provided, A" can be equipped with
a commutative d-algebra multiplication whenever A is a commutative d-algebra.
Then in [37, Theorem 7], Chil wa.c; able to drop the commutativity condition and
prove the following theorem.

THEOREM 5.4. Let A be an Archimedean d-algebra. Then the multiplication in


A extends to a multiplication in Ad such that A" is a Dedekind complete d-algebra
with respect to that extended multiplication.

In summary, all but one of the problems concerning Dedekind completions that
Huijsmans raised in his survey paper [57] have now been solved. The remaining
problem, though admittedly outside the scope of this survey, is the following.

PROBLEM 5.5. Let A be an Ar'chimedean (I-algebra. Does the multiplication in


A extend to a multiplication in A" so that A" is a Dedekind complete (I-algebra?

6. Powers in i-algebras
Let A be a uniformly complete i-algebra and let P E lR+ [Xl, ... , Xn] be a
homogeneous polynomial of degree a non zero natural number p. In their paper
[16]' Beukers and Huijsmans considered the following problem: does there exist
in A a 'p-th root' of P(iI, ... ,fn) for iI, ... ,fn in A+? They gave an affirmative
answer in the case where A is a semiprime f-algebra. More precisely, they prowd
the following theorem (see [16, Theorem 5]).

THEOREM 6.1. Let A be a uniformly complete semiprime f-algebra and let


P E lR+ [Xl, ... , Xn] be a homogeneo1Ls polynomial of degree a non zero natural
numbe1'p. Thenfo1' every fl, ... ,fn E A+ there exists a 'unique f E A+ such that
fP = P (iI, ... , fn).

As a consequence, one has the following corollary (see Corollary 6 in [16]).

COROLLARY 6.2. Let A be a uniformly complete f -algebra with unit element


and p E {I, 2, ... }. Then for each f E A +, there exists a unique 9 E A + such that
gl' = f.
LATTICE ORDERED ALGEBRAS 109

Note that the previous result was first proved for p = 2 in [17, Theorem 4.2
and Cororllary 4.3] by Beukers, Huijsmans and de Pagter. Also, it should be noted
that Theorem 6.1 above is proved alternatively by Buskes, de Pagter, and van Rooij
in [31, Corollary 4.11], a paper that deals with a more general functional calculus
on Riesz spaces and f-algebrar; to which we will return in the next section. The
problem corresponding to Theorem 6.1 for almost f-algebras was considered by
Boulabiar and follows next (see Theorem 3 in [24]).
THEOREM 6.3. Let A be a uniformly complete almost f -algebra and let P E
jR+ [Xl, ... , Xn] be a homogeneous polynomial of degree a natural number p. Then
for every II, ... , fn E A+ there exists a (not necessarily unique) f E A+ such that
fP = P (II, ... , fn)
Observe that, where roots are unique in semiprime f-algebras, this is no longer
always the case for almost f-algebras. We illustrate this with an example.
EXAMPLE 6.4. Let A = C ([-1,1]) be the uniformly complete Riesz space of all
real-valued continuous functions on [-1, 1] and define w E A by

W(x)={ O-x (-1::;x::;0);


(0::;x::;1).
For every f, g E A, we put
w(x)f(x)g(x) (-1::; x::; 0);
{
(f. g) (x) = lO/(S)g(S)dS (0::; x::; 1).

Then A is an almost f -algebra with respect to the multiplication.. Consider


h, g, Ct, {3 E A defined by
g(x) = Ixl h (x) = exp (x),
and
Ct (x) = Jx 2 + exp (2x) , and {3 (x) = X.X[O,I] (x) + Jx 2 + exp (2x)
for all x E [-1,1]' where X[O,I] (x) =1 if x E [0,1] and X[O,I] (x) =0 if x E [-1,0).
Then
Ct Ct = {3 {3 = 9 g + h h.
At this point, we define for each non zero natural number p,
Ap = {II ... fp : II, ... , fp E A}.

In what follows, we will investigate the order structure as well as the algebra struc-
ture of Ap (since Al = A, we suppose that p ~ 2). The sets A2 and A3 were first
considered in [35] by Buskes and van Rooij and then in great detail by Boulabiar in
[24] from which we summarize the results in the following theorem (see Theorems
4, 5, and 6 in [24]).
THEOREM 6.5. Let A be a uniformly complete almost f -algebra and let p ~ 3
be a natural number. Then Ap is a uniformly complete semiprime f -algebra under
the ordering and multiplication inherited from A. The positive cone At of Ap is
defined by
110 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

The lattice operations I\p and V p in Ap are given by


fP I\p gP = (f 1\ g)P and fP Vp gP = (f V g)P for all 0 ~ f,g E A,
and the absolute value 1.l p in Ap is defined by

If Pip = Ifl P for all f E A.

Contrary to Ap (p 2: 3), A2 need not be a Riesz space under the ordering


inherited from A as is proved by the next example.
EXAMPLE 6.6. Consider A = C ([0,1]) with the pO'intwise addition, scalar rrml-
tiplication and partial ordering. For f, 9 E A, define

(f _ g)(x) = { l x - 1/2 0
f(s)g(s)ds (1/2 < x ~ 1).
(0 ~ x ~ 1/2);

o
Then A is a uniformly complete almost f -algebra under the multiplication - and h
is an element of A2 if and only if h(x) = 0 for all x E [0,1/2] and the restriction of
h to [1/2,1] belong to C 1 ([1/2,1]). Hence A2 is not a Riesz space under the order
inherited from A.

The following example proves that though Ap (p 2: 3) is a Riesz space, in general


it is not a Riesz subspace of A.
EXAMPLE 6.7. Take A = C ([-1, 1]) with the pointwise addition, scalar multi-
plication and ordering, and define w EA by
(-l~x~O);
w (x) = { -x
0' (0 ~ x ~ 1) .

For f, 9 E A, define
w(x)f(x)g(x) (-l~x~O);
{
(f - g) (x) = 10/(S)g(S)ds (O~x~l).

Clearly, A is a uniformly complete almost f -algebra under the multiplication _.


Define 0 E A by
o (x) = 2x + 1 for all x E [-1, 1].
It follows that

10 - 0 - 01 (1) = 1/10 =110 - 0 - ob (1) = (101-101-101) (1) = 1/8.


If, however, A is a commutative d-algebra then some of the unpleasantness of
the preceding example disappears.
COROLLARY 6.8. Let A be a uniformly complete commutative d-algebra and
p 2: 2 be a natural number. Then Ap is a uniformly complete f -subalgebra of A. If
in addition p 2: 3 then Ap is semiprime.

In spite of the improvement in the conclusion of Corollary 6.8 over the con-
clusion for the more general situation of almost f-algebras, A2 still need not be
semiprime. This is illustrated in the next example.
LATTICE ORDERED ALGEBRAS 111

EXAMPLE 6.9. Let A be the coordinatewise ordered vector space R3 with the
multiplication defined by:

A, ~{( n; R}
Then A is a uniformly complete commutative d-algebm and

X,Y E

Dbuiou,ly, A, is an f-""balg,bm of A. FUrth,""ore, ( ~ ) ' ~ 0 and A, is not


semiprime.
For f -algebras we have the following corollary.
COROLLARY 6.10. Let A be a uniformly complete f -algebm and let p 2: 2 be a
natuml number. Then Ap is a uniformly complete semiprime f -subalgebm of A.
There is a universal way in which A 2 , or more generally Ap for any p 2: 2 can
be described. We provide the details of that description for A2 next (see [36]).
Let E and F be Riesz spaces. A bilinear map q> : E x E ~ F is called
orthosymmetric if whenever f I\g = 0 for f,g E E we have q>(f,g) = 0 (the notion
of orthosymmetric bilinear map was introduced by Buskes and van Rooij in [34]).
The bilinear map q> is a Riesz bimorphism if it is a Riesz homomorphism in each
variable separately (more about Riesz bimorphisms can be found in [29]). Let E
be a Riesz space. The pair (E8, 8) is called a square of E, if E8 is a Riesz space
and if
(1) 8: E x E ~ E8 is an orthosymmetric Riesz bimorphism, and
(2) for every Riesz space F, whenever q> : E x E ~ F is an orthosymmetric
Riesz bimorphism there exists a unique Riesz homomorphism q>8 : E8 ~
F such that q>8 08 = q>.
The existence and uniqueness of squares for any Riesz space follows easily from
the Riesz space tensor product as constructed by Fremlin in [43]. To understand the
structure of the square of a Riesz space is best not done via this tensor product. The
set A2 described above is often more helpful. The connection between semi prime
f-algebras and squares of uniformly complete Riesz spaces is described in the next
theorem. After reading that theorem the reader might feel like moving the lower
index 2 in A2 to an upper index.
THEOREM 6.11. Let E be a uniformly complete Riesz subspace of an Archimedean
semiprime f -algebm G whose multiplication is indicated by a period e. Put E2 :=
{x e y : x,y E E} as before. Then E2 is a Riesz subspace of G and (E2,e) is a
square of E.

7. Functional Calculus on f-algebras


The theorem that we presented in Section 6 on the existence of p-th roots of
homogeneous polynomials in f-algebras is a very special case of a rich functional
calculus on uniformly complete f-algebras. The idea behind functional calculus
112 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

for Riesz spaces in general is straightforward. For elementary functions on JRN


one ought to be able to simply substitute elements of the Riesz space into these
functions and get elements of the Riesz space as output. The idea of how to execute
this substitution of elements in sufficiently simple functions essentially goes back
to Yudin and in the form that we represent it to Lozanovsky [70]. The technical
problem surmounts to what the class of sufficiently simple functions really looks
like.
Let n EN. We denote by 1i(JRN) the Riesz space of all continuous functions
r.p : JRN ---+ JR for which
r.p(tx) = tr.p(x) for all x E JRN and all t ~ O.
Let E be a Riesz space, r.p E 1i(JRN) and iI, ... , fn E E. We say that
r.p(iI, ... '/n) exists in E
if there is an element 9 of E such that
w(g) = r.p(w(iI), ... ,w(fn))
for every real-valued Riesz homomorphism w on the Riesz subspace of E generated
by iI, ... , f n, g. For any given E, r.p and iI, ... ,/n there exists at most one 9 with
this property. This 9 is also indicated by
r.p(iI, ... , fn).
In this situation we have the following theorem (see Lozanovsky [70]).
THEOREM 7.1. Let E be a uniformly complete Riesz space and iI, ... , fn E E.
Then r.p(iI, ... , fn) exists fOT every r.p E 1i(JRN). The map
r.p(iI, ... , fn) (r.p E 1i(JRN))
r.p ---+

is a Riesz homomorphism from 1i(JRN) into E.


Remark. In a way, r.p(iI, ... , fn) is independent of E. Indeed, if D is any Riesz
subspace of E that is uniformly complete and contains iI, ... , fn then r.p(iI, ... , fn)
relative to D means the same as r.p(iI, ... , fn) relative to E. In particular, every
Riesz subspace of E that is uniformly complete and contains iI, ... , fn must also
contain r.p(fl, ... , fn).
By A(JR N) we denote the set of all continuous functions r.p : JRN ---+ JR that are
of polynomial growth and for which limt!o rlr.p(tx) exists uniformly on bounded
subsets of JRN (the latter condition is equivalent to the existence of a 't/J E 1i(JRN)
such that r.p(x) = 't/J(x) + 0(11 x II) (x ---+ 0)). Observe that A(JRN) is an f-algebra.
Let E be a semiprime f-algebra, r.p E A(JRN) and iI, ... , fn E E. We say that
r.p(iI, ... ,fn)exists in E
if there is agE E with
w(g) = r.p(w(iI), ... ,w(fn))
for every real-valued multiplicative Riesz homomorphism w defined on the f-subalgebra
of E generated by iI, ... , fn,g. There exists only one such g, which is then called
r.p(iI, ... , fn).
This definition is in accordance with the one we gave for 1i(JRN) if r.p E 1i(JRN).
For 1i(JRN) we have the following theorem (see [31, Theorem 4.10]).
LATTICE ORDERED ALGEBRAS 113

THEOREM 7.2. Let E be a uniformly complete semiprime f-algebra and let


ft, ... , fn E E. Then cp(ft, ... , fn) exists for every cp E A(~JII). The map
cp-+cp(ft, ... ,fn) (cpEA(~JII))
is a multiplicative Riesz homomorphism from A(~JII) into E.

8. Relationships between i-algebra multiplications


Let A be an i-algebra with multiplication denoted by juxtaposition, and assume
that A is equipped with another associative multiplication e. In the first theorem of
this section, we present a relationship between the two multiplications in A, under
the conditions that A is a unital f-algebra with respect to the initial multiplication
and an (almost) f-algebra with respect to the other multiplication e. For proofs,
see [38, Theorem 2.2] and [23, Theorem 5.2].
THEOREM 8.1. Let A be an Archimedean f -algebra with identity element e and
assume that A is furnished with another associative multiplication e. Then
(i) A is an f-algebra under e if and only if
feg=(eee)fg for allf,g E A, and
(ii) A is an almost f-algebra under e if and only 'if
f e9 = e e (I g) for all J, 9 E A.
The corresponding problem in the case where A is d-algebra with respect to e
is rather more difficult. Indeed, since then e need not be commutative, one cannot
write the product f e 9 as a function of the product fg (I, 9 E A). However, there
exists another way (involving f, 9 and the initial multiplication in A) to express the
product f e g. This is the subject of the next result. First recall that the maximal
ring of quotients Q (A) of the Archimedean f -algebra A with unit element e is again
an Archimedean J-algebra with the same e as multiplicative identity. Moreover, A
is an f-subalgebra of Q (A), a fact proved by Anderson in [4] (see also the recent
paper [71, Cororllary 2.7.1] by Martinez). For the definition of the maximal ring
of quotients of a ring, the reader can consult e.g. [66]. The proof of the following
theorem can be found in [25, Theorem 4.3].
THEOREM 8.2. Let A be an Archimedean f-algebra with 'identity element e and
let e be another associat'ive multiplication in A. Then A is a d-algebra with respect
to e if and only if there exist two algebra and Riesz homomorphisms cp and 'l/J from
A into its maximal ring of quotients Q (A) such that
f e 9 = (e e e) cp (I) 'l/J (g) for all f, 9 E A.
As mentioned at the end of Section 3, the two preceding theorems are abstract
versions of the corresponding results, given in that section, for the C (X)-case.
In the second part of this section, we are interested in A being a commutative d-
algebra with respect to the initial multiplication rather than an f-algebra with unit
element. However, we will impose the additional assumption that A is uniformly
complete. Uniform completeness is not needed for all our results but we will use
the set A2 and remind the reader of the special nature of that set under the extra
condition of uniform completeness (see Corollary 6.8). If there exists a positive
operator T OIl A2 such that
(T) f e9 = T (lg) for all f, 9 E A
114 KARIM BOULABIAR, GERARD BUSKES. AND ABDELMA.HD TRIKI

then e is an almost f-algebra multiplication and N2 (A) c N; (A), where


N; (A) = {f E A : f e f = O} .
In what follows, we show in detail what happens if we assume that A with the initial
multiplication is a (uniformly complete) commutative d-algebra, and the inclusion
N2 (A) c N; (A) holds. Under those circumstances, we then relate a necessary
and sufficient condition for the new multiplication to be an almost f-algebra, a
d-algebra or an f-algebra to the existence of some posit.ive operator T satisfying
the relation (T). The details follow in the next theorem, the proof of which can be
found in [23, Theorems 5.4 and 5.5].
THEOREM 8.3. Let A be a uniformly complete commutative d-algebra and as-
sume that A is an I!-algebra with respect to another associative multiplication e such
that p = 0 implies f e f = O. Then the following statements hold.
(i) A is an almost f -algebra undere if and only if ther-e exists a positive
operator T from A2 into A such that
feg=T(fg) forallf,gEA.
(ii) A is a commutative d-algebra under
e if and only if there exists a Riesz
homomorphism T from A2 into A such that
f e 9 = T (fg) for all f, 9 E A.
(iii) A is an f-algebra under e if and only if there exists an operator T from
A2 into A such that To 7rf E Ort.h (A) for all f E A+, where 7rf (g) = fg
for all 9 E A, and
f e9 =T (f g) for all f, 9 E A.
We remark that A2 in the previous theorem is a Riesz space (see Section 6).
Next, we produce an example which shows that in Theorem 8.3 above, the hy-
pothesis 'A is a commutative d-algebra' cannot be replaced by 'A is an almost
f-algebra'.
EXAMPLE 8.4. Take A = C ([-1,1]) with the usual operations and order and
define a, (3 E A by
-(4x+l) (-I~x~-1/4);
a(x)= { 0 (-1/4 ~ x ~ 1/4);
4x -1 (1/4 ~ x ~ 1)
and
(3 (x) = { ~4x + 1 (-1 ~x~ 1/4);
(1/4 ~ x ~ 1).
For f, 9 E A, define
a(x)f(x)g(x) (-1~x~1/4);
{ (1/4 ~ x
~ 3/4);
(f x g)(x) = J(X-3/4)
f(s)g(s)ds (3/4~x~1)
-(x-3/4)

and
(feg) (x) =f3(x)f(x)g(x)
LATTICE ORDERED ALGEBRAS 115

for all x E [-1,1]. Then A is an almost f-algebra (respectively, an f-algebra) with


respect to the multiplication x (respectively, .). It follows that
N; (A) = N 2x (A) = {f E A : f (x) = 0 for all x E [-1, I/4]}.
Consider <.p, w E A, defined by

"'(X)~{
-3x-I (-1 ::=;x::=; -1/4);
x (-I/4::=; x::=; 1/4);
-3x+I (i/4::=;x::=;I)
and

~
0 (-I::=;x::=;-I/4);
4x+ 1 (-I/4::=; x::=; 0);
oJ (x) { -4x+I (0::=;x::=;I/4);
0 (i/4::=;x::=;I).
Then
'Pxw=O and 'P. W =f:. O.
Therefore, there is no operator T satisfying the condition
f g = T (f x g) for all f, 9 E A.
We notice that the kind of relationships between two f-algebra multiplications
via an operator T as above were first studied in [35] by Buskes and van Rooij. In
particular part (i) of the theorem above has its origins in [35, Theorem 1], where
a general representation theorem for almost f-algebras (without any extra condi-
tions) was proved by first dividing out the radical and then showing the existence
of an operator T a.'l above on the square (see the end of Section 6 above) of the
resulting f-algebra, thus giving quantified credence to the almost part in the name
almost f-algebras. They also used this representation to discover that the Dedekind
completion of an almost f-algebra is an almost f-algebra and to prove (CSP) for
almost f-algebras. Finally, we should point out that the Buskes-van Rooij repre-
sentation theorem for almost f-algebras itself is, in a way, an abstract reformulation
of Theorem 3.2 by Scheffold that we discussed earlier.
We end this section with the following problem, left open in the above discus-
sion.
PROBLEM 8.5. Let A be a uniformly complete commutative d-algebra and as-
sume that A is a non comm'utative d-algebra with respect to another multiplication
. Does there exists a relationship between the two multiplications in A?

9. Connection between algebra and Riesz homomorphisms


Since in any f-algebra there is an order structure as well as an algebra structure,
it is natural to compare the operators that preserve the lattice operations, namely
the Riesz homomorphisms, with those that preserve the algebra structure, the alge-
bra homomorphisms. Such a comparison was initiated by Ellis [41], who considered
the problem for operators between spaces of continuous functions on compact Haus-
dorff spaces, and his central result is then one of equivalence: a Markov operator
(Le., a positive operator preserving the identity) between such spaces is a Riesz
homomorphism if and only if it is an algebra homomorphism. Some years. after El-
lis's paper was published, Hager and Robertson presented in [48] a more abstract
version of Ellis's theorem. More precisely, Hager and Robertson proved that any
116 KARIM BOULABIAR, GERARD BUSKES. AND ABDELMAJID TRIKI

Riesz homomorphism between two Archimedean f-algebras with unit elements that
preserves the identity is an algebra homomorphism. Later, van Putten established
the converse of the result by Hager and Robertson in his thesis [78]. The afore-
mentioned results were generalized by Huijsmans and de Pagter in [59, Theorem
5.4] as follows.
THEOREM 9.1. Let A be an Archimedean f-algebra with unit element e, B be
an Archimedean semiprime f-algebra. and T : A ---> B be a positive operator. Then
the following are equivalent
(i) T is an algebra homomorphism.
(ii) T is Riesz homomorphism with (Te)2 = Te.
Notice that the previous theorem generalizes all of the facts cited above. In the
same paper [59], Huijsmans and de Pagter proved that., if A is an Archimedean f-
algebra with unit element and B is an Archimedean semiprime f-algebra then any
order bounded algebra homomorphism from A into B is automatically a Riesz ho-
momorphism [59, Theorem 5.3]. Very recently in [91, Theorem 4.3], Triki obtained
this result in the more general setting of almost f-algebras.
THEOREM 9.2. Let A be an Archimedean almost f-algebra and let B be an
Archimedean semiprime f -algebra. Then any order bounded algebra homomorphism
from A into B is a Riesz homomorphism.
Since any conmmtative d-algebra is automatically an almost f-algebra, Theo-
rem 9.2 holds if one replaces 'A is an Archimedean almost f-algebra' by 'A is an
Archimedean commutative d-algebra'. However, in the recent. work [89], Toumi
proved that we have the same conclusion even if the d-algebra under consideration
is not commutative.
THEOREM 9.3. Let A be an arbitrary Archimedean d-algebra and let B be an
Archimedean semiprime f-algebra. Then any order bounded algebra homomorphism
from A into B is a Riesz homomorphism.
Even for f-algebras, the condition of order boundedness in Theorems 9.2 and
9.3 cannot be dropped as is shown in the following example.
EXAMPLE 9.4. Consider the set A of all real sequences u = {un} n>l for which
there exists a polynomial Pu E IR [X] and a natural number N such that Un =
Pu (n) for all n 2: N. Under the usual operations and partial ordering, A is an
Archimedean f-algebra (and thus an almost f-algebra and a d-algebra) and A is
not relatively uniformly complete. Define the algebra homomorphism T : A ---> IR by
T(u) = Pu (-1) (u E A). For every A E [l,+po[, we define u), = {U),.n}n~l E A by

{ 0An (n ~ A);
u)',n = (n> A) .
lfu = {n 2 } n>l then 0 ~ u), ~ u for all A E [1, +00[. Observe now that T (1L),) = -A.
Therefore T 1:s not order bounded and hence not a Riesz homomorphism.
In [59, Theorem 5.1], Huijsmans and de Pagter further illustrated the close
ties between Riesz homomorphisms and algebra homomorphisms on f-algebras as
follows. If A and Bare semiprime f-algebras and A is uniformly complete, then any
algebra homomorphism from A into B is a Riesz homomorphism. Their theorem
was generalized by Boulabiar in [27, Theorem 3] to the setting of almost f-algebras.
LATTICE ORDERED ALGEBRAS 117

THEOREM 9.5. Let A be a uniformly complete almost f -algebra and let B be an


Archimedean semiprime f-algebra. Then any algebra homomorphism from A into
B is a Riesz homomorphism.
Trivially, one can replace in Theorem 9.5 above 'almost f-algebra' by'commu-
tative d-algebra'. The case of d-algebras that are not necessarily commutative was
considered by Toumi in [89, Theorem 3].
THEOREM 9.6. Let A be an arbitrary uniformly complete d-algebra and let B
be an Archimedean semiprime f-algebra. Then any algebra homomorphism from A
into B is a Riesz homomorphism.
The condition that A is uniformly complete is not redundant in Theorems 9.5
and 9.6, even if A is an f-algebra. Indeed, A in Example 9.4 above is not uniformly
complete.
The following theorem by Triki in [91, Theorem 4.4] is an f-algebra version
of the well known Nagasawa's theorem (see [74, Theorem 1]). First, recall that
an operator T between two unital f-algebras A and B is said to be contractive if
IT (J)I ~ eB in B whenever If I ~ eA in A, where eA and eB are the unit elements
of A and B, respectively. If moreover T is bijective and the inverse T- 1 of T is also
contractive then we say that T is bicontractive.
THEOREM 9.7. Let A and B be Archimedean f-algebras with identity elements
eA and eB, respectively. For an order bounded bijection T : A ----t B such that
T (eA) = eB, the following are equivalent.
(i) T is an algebra homomorphism.
(ii) T is a Riesz homomorphism.
(iii) T is bicontractive.
The last result of this section again deals with operators between two unital f-
algebras that preserve identities, so called Markov operators. Denote by M (A, B)
the set of all Markov operators from an f-algebra A with unit element eA into an
f-algebra B with unit element eB, that is,
M (A,B) = {T: A ----t B: T linear positive with T(eA) = eB}.
Obviously, M (A, B) is a convex set. In [77, Theorem 2.1], Phelps proved that a
Markov operator T from C (X) into C (Y) (where X and Yare compact Hausdorff
topological spaces) is an algebra homomorphism if and only if T is an extremal
point in M (C (X) ,C (Y)). This connection to extreme points predates the paper
by Ellis and we refer the reader to Arens and Kelley [8] and A. and C. Ionescu
Thlcea [62]. Combining the aforementioned Phelps's theorem and Ellis's result
cited above we get that for T in M (C (X), C (Y)) the following are equivalent.
(i) T is a Riesz homomorphism.
(ii) T is an algebra homomorphism.
(iii) T is an extremal point in M (C (X) , C (Y)).
As a generalization, van Putten in his thesis obtained the following result [78,
Theorem 18.8].
THEORSM 9.8. Let A and B be Archimedean f-algebras with unit elements,
and let T E M (A, B). Then the following are equivalent.
(i) T is a Riesz homomorphism.
118 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

(ii) T is an algebra homomorphism,


(iii) T is an extremal point in M (A, B).
An elementary proof of van Putten's theorem above, due to Huijsmans and de
Pagter, can be found in [59, Theorem 5.7].

10. Positive derivations


We recall that an operator D on a commutative algebra A is said to be a
derivation if
D (1g) = f D (g) + gD(1) for all f,g EA.
In this section, we investigate positive derivations on f-algebras as well as on almost
f-algebras.
Positive derivations on f-algebras were first considered in great detail by Colville,
Davis and Keimel in [40]. Their main result provides the following necessary and
sufficient condition for a positive operator on an f-algebra to be a derivation (see
Theorem 5 in [40]), where we recall that if p is a non zero natural number, and A
is an algebra then
Ap = {11 ... fp : 11, ... , fp E A}.
THEOREM 10.1. Let A be an f-algebra and D be a positive operator on A. Then
D is a derivation if and only if
D (1) = 0 for all f E A2 and D (1)2 = 0 for all f E A.
Positive derivations on f-rings were considered by Henriksen and Smith in
[54]. It straightforwardly follows from Theorem 10.1 above that, if A is in addition
semiprime, then there exist no non trivial positive derivations on A.
We turn our attention now to positive derivations on almost f-algebras. The
result corresponding to Theorem 10.1 for almost f-algebras is the following (for a
proof, see [26, Theorem 3]).
THEOREM 10.2. Let A be an almost f-algebra and D be a positive derivation
on A. Then
D (1) = 0 for all f E A3 and D (1)3 = 0 for all f E A.
Contrary to f-algebras, Theorem 10.2 does not produce a characterization of
positive derivations on almost f-algebras. Also, the third power in Theorem 10.2
is the best possible. The next example illustrates these facts.
EXAMPLE 10.3. Consider A the Cartesian product IR x IR with coordinatewise
addition, scalar multiplication and ordering. Define the multiplication. on A by
(0,/1). (0',/3') = (0,00') for all 0,/3,0',/3' E R
Then A is an Archimedean almost f -algebra with respect to.. Observe that A3 =
{o}. Hence the identity map
fA satisfies

fA (1) = 0 for all f E A3 and fA (1) 3 = 0 for all f E A,


but fA is not a derivation on A. On the other hand, let D be the positive operator
defined on A by
D(o,/3) = (0,2/3) for all a, /3 E R
Clearly, D is a derivation on A. However,
D ((1, 0) (1,0)) = (0,2) =1= (0,0)
LATTICE ORDERED ALGEBRAS 119

and
D (1, 0) D (1, 0) = (0,1) =I- (0,0) .
Theorem 10.2 above of course holds for commutative d-algebras as well. More-
over, for commutative d-algebras the third power is the best possible too. Indeed,
the almost f-algebra considered in Example 10.3 is in fact a commutative d-algebra.

11. Cauchy-Schwarz inequalities


We say that an i-algebra A possesses the Cauchy-Schwarz property (abbrevi-
ated as (CSP)), if for every vector space V and every bilinear map 'l/J : V x V -> A
such that (i) 'l/J is symmetric, and (ii) 'l/J (I, f) E A+ for all f E V, we have
(CSI) 'l/J(I,g)2 S'l/J(I,f)'l/J(g,g) for all f,g E V
In [61, Corollary 3.5], Huijsmans and de Pagter proved that any Archimedean
semiprime f-algebra has (CSP). Later, Bernau and Huijsmans in [15, Theorem
2.6] generalized (CSP) to arbitrary Archimedean f-algebras.
THEOREM 11.1. Let A be an Archimedean f-algebra, V a vector space and
'l/J : V x V -> A a bilinear map such that (i) 'l/J is symmetric, and (ii) 'l/J (I, f) E A+
for all f E V . Then
'l/J (I, g)2 S 'l/J (I, f) 'l/J(g, g) for all f,g E V.
Some years after Theorem 11.1 was published, Buskes and van Rooij established
the corresponding inequality for Archimedean almost f-algebras and therefore for
Archimedean commutative d-algebras (see [34, Corollary 4]).
THEOREM 11.2. Let A be an Archimedean almost f-algebra, V a vector space
and'l/J : V x V -> A a bilinear map such that (i) 'l/J is symmetric, and (ii) 't/J (I, f) E
A + fOT all f E V . Then
'1/) (I, g)2 S .t/J (I, f) ,'(g, g) for all f, g E V.
An alternative proof of the previous inequality was given by Boulabiar in [23,
Theorem 3.9].
Not every d-algebra A has (CSP). An example illustrating this situation is the
following.
EXAMPLE 11.3. Let A be the set of all real-valued functions defined on [0,1]
equipped with the usual operations and order. Consider the multiplication. defined
in A by
(I )() - { f (0) g (1) (0 S x S 1/2) ;
.g x - f(l)g(l) (1/2SxS1)
for all f, g E A. Then A is an Archimedean non-commutative d-algebra with r-espect
to the multiplication .
At this point, let V be the f-algebra C ([0,1]) of all real-valued continuous func-
tions on [0, 1] , provided with the pointwise addition, multiplication, scalar multipli-
cation and partial ordering.
We define a positive operator T from V into A by T (I) (x) = f (1 - x) for all
f E V and x E [0,1].
Finally, let f E V such that f (0) = 2, f (1) = 1 and e E B such that e (x) = 1
for all x E [0,1]. Then
(T (Ie) T (Ie)) (0) = (T (I) T (I)) (0) =f (0) f (1) =2
120 KARIM BOULABIAR, GERARD BUSKES. AND ABDELMAJID TRIKI

and

Hence the inequality


T (fe). T (fe) :s T (J2) T (e 2)
does not hold in A. By defining '!/J(f,g) = T(fg) (f,g E V), we find a bilinear map
V x V --t A for which 1jJ (f, e)2 :s 1jJ (f, f) 1jJ(e, e) does not hold.
Though (CSP) does not hold for all non commutative d-algebras, Boulabiar
and Toumi proved, in essence, the following variant of (CSP) in such algebras (see
[29, Theorem 6]).
THEOREM 11.4. Let V be a vector space and let A be an Archimedean d-algebra.
Consider a bilinear map '!/J : V x V --t A such that (i) 1jJ is symmetric, and (ii)
1jJ (f, f) E A+ for all f E V. Then

11,b(f,g)21:s ~[(1jJ(f,f)1jJ(g,g))+('!/J(g,g)1jJ(f,f))]
and hence
11jJ(f,g)21:s (1jJ(f,f)'!/J(g,g)) V (l/J(g,g)'!/J(f,f))
for all f,g E V.

12. Order biduals


We refer the reader to [2] for terminology and notations not explained below.
For an Archimedean C-algebra A, the order dual is denoted by A~ and the
order bidual is denoted by A~~. Recall here that the order dual A~ of A is the
Dedekind complete Riesz space of all real-valued order bounded functionals on A.
A multiplication can be introduced in A~~ in three steps as follows: for all u, v E A,
f E A~ and cp, '!/J E A~~, we define f.u E A~, '!/J.f E A~ and cp.1jJ E A~~ by the
following equations
(1) (f:u) (v) = f (uv)
(2) ('!/J.f) (u) = '!/J (f.u)
(3) (cp.'!/J) (f) = cp ('!/J.f)
The multiplication defined by the equation (3) is called the Arens multiplication
in A~~ (see [6] and [7]). The order continuous order bidual (A~);: of A is the
projection band of all order continuous elements in A~~. Moreover, (A~);: is closed
under the Arens multiplication, that is, cp ..l/J E (A~);: whenever cp, '!/J E (A~);: . The
following theorem was established by Huijsmans and de Pagter (see [60, Theorem
4.1]).
THEOREM 12.1. Let A be an Archimedean C-algebra. Then A~~ (and hence
(A~);:) is a Dedekind complete C-algebra with respect to the Arens multiplication.
e
If the f-algebra A has, in addition, a unit element e then E (A~);: (defined
by e(f) = f (e) for all f E A~) is the unit element of A~~. Generally, A~~ need
not be commutative even if A is commutative. An example was provided by Arens
in [7]. However, commutativity does carryover from A to (A~);:, which was first
established for normed C-algebras by Scheffold in his paper [82] and generalized by
Grobler in [46, Theorem 4] to the more general setting of Archimedean C-algebras.
LATTICE ORDERED ALGEBRAS 121

THEOREM 12.2. Let A be an Archimedean i-algebra. Then the order continuous


(A~);: , provided with the Arens multiplication, is commutative.
orde1' bidual
The first results concerning the order bidual of f-algebras are due to Huijs-
mans and de Pagter (see [60]). They proved, among other results, that if A is an
f-algebra with unit element such that A~ separates the points of A (i.e., if u E A
and f (u) = 0 for all f E A~ then u = 0) then A~~ = (A~);:. They deduced
that A~~ then again is an f-algebra (of course, with unit element) with respect
to the Arens multiplication (note that if A~ separates the points of A then A is
Archimedean). Later Huijsmans was able to do away with the condition that A
have a unit element, and he showed that the order bidual of an f-algebra with sep-
arating order dual is also an f-algebra with respect to the Arens multiplication (see
[56, Theorem 2.8]). Then, without assuming any additional 'separation' condition,
Bernau and Huijsmans proved in [15] that the order bidual of an Archimedean f-
algebra, equipped with the Arens multiplication, is again an f-algebra (see Theorem
3.2 and Theorem 3.5 in [15]).
THEOREM 12.3. Let A be an Archimedean f-algebra. The order bidual A~~ of
A, equipped with the Arens multiplication is an f-algebra and hence 'is commutative.
In the paper [14], Bernau and Huijsmans also dealt with the case of almost
f-algebras.
THEOREM 12.4. Let A be an Archimedean almost f -algebra. Furnished with
the Arens multiplication, the order bidual A~~ of A is an almost f-algebra and
hence is commutative.
Combining Theorem 12.4 and the fact that commutative d-algebras are almost
f-algebras, Bernau and Huijsmans deduced the following result (see [14, Theorem
4.2]).
COROLLARY 12.5. Let A be an Archimedean commutative d-algebra. then the
order bidual A~~ of A is a commutative d-algebra with respect to the Arens multi-
plication.
It is interesting to observe that the square of the singular part of the bidual
(the orthogonal complement of the order continuous part) vanishes in the case of
an almost f-algebra. Contrary to the case of almost f-algebras, the corresponding
problem for non commutative d-algebras remains open and only a partial result has
been obtained, again by Bernau and Huijsmans in [14, Theorem 4.1]. Their result
is the following.
PROPOSITION 12.6. Let A be an Archimedean d-algebra. Then the order con-
tinuous order bidual (A~);: is a d-algebra with respect to the Arens multiplication.
As just observed, the question whether the full order bidual of a non com-
mutative d-algebra is a d-algebra with respect to the Arens multiplication is still
open.
PROBLEM 12.7. Let A be an Archimedean d-algebra. Is the order bidual A~~
of A a d-algebra with respect to the Arens multiplication?
At the end of this section, we notice for the sake of completeness that all of
the results above were obtained independently by Scheffold under the additional
. assumption of A being a Banach lattice (including the result on d-algebras) (see
[81] and [83]).
122 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

13. Ideal theory


In this section, we follow [69J in our terminology and notations. Recall that an
order ideal in a Riesz space L is a vector subspace I of L with the extra condition
III ::; Igl in Land gEl imply I E I. We call a ring ideal I in an i-algebra A an
i-ideal if I is simultaneously an order ideal in A. Ideal theory in lattice ordered
algebras comprises an investigation of ring ideals, order ideals and the connection
between them. In the first part of this section, we focus on the following problem:
(P 1) Under what conditions is any order ideal in an I -algebra a ring ideal?
In the second part we consider the converse problem:
(P2) Under what conditions is any ring ideal in an I -algebra an order -ideal?
The first problem was initiated by Henriksen in [49J, who studied the ideal
theory in I-rings by means of representations. In their work [58J on ideal theory in
I-algebras, Huijsmans and de Pagter proved the next result by a representation-free
approach (see [58, Proposition 3.1]).
PROPOSITION 13.1. Let A be an Archimedean I-algebra. Then any unilormly
closed order ideal is a ring ideal.
The problem (PI) for non uniformly closed order ideals was considered cxten-
sively by Basly and Triki in [l1J and [12J. In what follows, we present the major
results they obtained in that direction. First, we need to recall some prerequisites.
A ring ideal I in a commutative algebra A is said to be modular whenever there
exists h E A such that I - I h E I for all I E A. In particular, any ring ideal in
A is modular if A is in addition unital. From now on A is an I-algebra. We know
that the map
p: A ~ Orth(A)
I I---> 7rf

(where 7rf (g) = I g for all I, g E A) is an algebra and Riesz homomorphism (see
Section 4 above). An element I in A is said to be bounded if p (a) = 7ry. E Z (A),
where
Z(A) = {7r E Orth(A): 17r1::; >.IA for some real number A}
is the centre of A (Le., the principal order ideal generated by IA in Orth (A)). We
denote, after Triki (see [90]), the set of all bounded elements in A by Ab. Note
that rather than A b , Henrikscn and Johnson use A* in [53J after a similar usage in
Gillman and Jerison's book. Clearly, Ab is an I-subalgehra of A. The I-algebra A
is said to be bounded if A = Ab, that is, if p (A) is a subset of Z (A). In particular,
a unital I-algebra A is bounded if and only if its multiplicative identity is also a
strong order unit, and in this situation A and Z (A) are isomorphic as f-algebras.
Here we recall that the identity element in an Archimedean unital f-algebra is
a weak order unit (see [22, p. 60]). The equivalence of (i) and (ii) in the next
theorem is an immediate consequence of the definitions of an order ideal and a
bounded Archimedean I-algebra, while for the proof of the rest of tllis theorem, we
refer to Theorem 3 in [l1J and Theorem 4 in [12J.
THEOREM 13.2. Let A be an Archimedean I-algebra. Then the following are
equivalent.
(i) Every order ideal is a ring ideal.
(ii) A is bounded.
LATTICE ORDERED ALGEBRAS 123

If A is in addition uniforrn,ly complete then (i) above is equivalent to each of the


following.
(iii) Every maximal modular ring ideal in A is uniformly closed.
(iv) Every maximal modular ring ideal is the kernel of a Hiesz and algebra
homomorphism from A onto JR.
(v) Every maximal modular ring ideal is a maximal or'der ideal.
We turn our attention to semiprime f-algebras. Let L be a Riesz space. A norm
on a L is called a Riesz norm if IIfil ::; IIgil whenever If I ::; Igl in L. If a Riesz
11.11
norm on L exists then L is said to be a normed Riesz space. If the normed Riesz
space L is a Banach space as well, we say that L is a Banach lattice. We call the
Riesz norm 11.11 on L an AI-norm if IIf V gil = max {IIfil , IIgll} for all f,g E L. An
Ai-space is an At-normed Banach lattice. We can now state the following theorem
(see [11, Theorem 5] and [12, Corollary 5]).
THEOREM 13.3. Let A be an Archimedean semiprime f -algebra. The following
are equivalent.
(i) Every order ideal in A is a ring ideal in A.
(ii) A is a isomorphic as an f -algebra to a subalgebra of Z (A).
(iii) There exists an M -norm in A.
(iv) There exists a Riesz norm in A.
If, in addition, A is uniformly complete then each of (i), (ii), (iii) and (iv) above is
equivalent to
(v) Every maximal modular ring ideal in A is uniformly closed.
We note that Problem (PI) was extensively studied by Henriksen, Larson and
Smith [52] in the context of f-rings.
We move on to discuss Problem (P2). To this end, we need the notion of a
normal Riesz space. The Riesz space L is said to be normal if L = {f+} d+{f-} d for
allf E L, where {f+}d = {g E L: Igl.l\f+ = O} and {f_}d = {g E L: Igl.l\f- = O}.
For a completely regular topological space X, the Riesz space C (X) is normal if and
only if the sets P(f) = {x EX: f(x) > O} and N(f) = {x EX: f(x) < O} are
completely separated for every fEe (X). For spaces of the type C (X), Problem
(P2) has the following solution (see Theorem 14.24 in [47]).
THEOREM 13.4. Let X be a completely regular (Hausdorff) topological space.
Then the following are equivalent.
(i) Every ring ideal in C (X) is an order ideal.
(ii) Every finitely generated ring ideal in C (X) is a principal ring ideal (i.e.,
X is an F -space).
(iii) C (X) is normal.
Huijsmans and de Pagter considered (P2) in f-algebras. They only considered
the unital case and their central result is the following generalization of Theorem
13.4 above (see [58, Theorem 6.6] for the proof).
THEOREM 13.5. Consider the following conditions for an Archimedean f -algebra
A with unit element.
(i) Every ring ideal is an order ideal.
(ii) Every finitely generated ring ideal is a principal ring ideal.
124 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

(iii) A is normal.
Then (i) '* '*
(ii) (iii). If, in addition A is uniformly complete then (i) {:} (ii) {:}
(iii) .
Next we focus on the non unital case, which was considered by Triki in [90].
First we define the notion of a stable f-algebra. The f-algebra A is said to be
stable if 1f (f) E (f) (where (f) is the principal ring ideal generated by f in A)
for all 1f E Z (A). It is clear that A is stable if and only if 1f (1) c I for all ring
ideals I in A and all 1f E Z (A). We now are in a position to present the result
corresponding to Theorem 13.5 for semiprime f-algebras (see [90, Theorem 5.4]).
THEOREM 13.6. Consider the following conditions for an Archimedean scmi-
prime f -algebra A.
(i) Every ring ideal is an order ideal.
(ii) Every finitely generated ring ideal is a principal ring ideal.
(iii) A is stable and normal, and {f+} d or {f-} d is a modular ring ideal for
all f E A.
Then (i) '* '*
(ii) (iii). Furthermore, if A is in addition uniformly complete then
(i) :> (ii) {:} (iii).
We proceed to the non semiprime case (see Theorem 5.5 in [90]).
THEOREM 13.7. Let A be a non semiprime Archimedean f-algebra. Then the
following are equivalent.
(i) Every ring ideal is an orner zdeal.
(ii) There exist a semiprime f-algebra B such that
(a) every 7ing ideal in B is an order ideal, and
(b) {b} d is a modular ring ideal in B for every b E H, so that A is
isomorphic to the f -algebra B x JR endowed with the multiplication
defined by (f, 0:) (g, 13) = (fg,O) for all i, g E B; 0:, 13 E JR.
Finally, note that the only i-algebras that we have considered in this section
are i-algebras. Indeed, a Problem (PI) for more general i-algebras is futile since
an i-algebra in which every order ideal is a ring ideal automatically is an f-algebra
(see Page 144 in the classical book [42] or Proposition 1 in [12]). The situation
for Problem (P2) is less clear. Indeed, in matrix algebras and algebras of formal
power series in one variable, when ordered coordinatewise, every ring ideal is an
order ideal. Thus we phrase Problem (P2) for lattice ordered algebras in general.
PROBLEM 13.8. Study Problem (P2) for Archimedean i-algebras other than
i-algebras.

14. Representation of I-algebras


Let A be an f-algebra and recall that Ab denotes the i-subalgebra of all
bounded elements in A (see Section 13 above). Several properties are satisfied
by A if and only if they are satisfied by Ab and vice versa (see Sections 3,4 and
5 in [90]). Thus we can study some aspects of A via an investigation of Ab. For
instance, if we assume that every ring ideal in A is an order ideal then A b has the
same property and the converse holds if A in addition is uniformly complete (see
[84] by Steinberg). In particular, if every ring ideal in A is an order ideal then order
LATTICE ORDERED ALGEBRAS 125

ideals and ring ideals coincide in Ab (see Theorem 13.2). It turns out that under
the latter hypothesis, A b has a nice representation as a space of functions. This
kind of a representation then is precisely the topic of this section.
First assume that A has an identity element (Le., A is a <T?-algebra in the
Henriksen-Johnson terminology of [53]). The set of all maximal (-ideals in A will
be denoted by M (A). For any subset D of M (A), the kernel k (D) of Dis
k (D) = n {M : M E D}
(where it is understood that k (0) = A). The hull h (I) of an (-ideal I in A is
h(I) = {M E M (A): I c M}.
The subset D of M (A) is said to be closed if D = h (k (D)).
One thus defines the
hull-kernel (or Stone) topology on M (A). It turns out that M (A) with respect to
this hull-kernel topology is a compact Hausdorff space (see [53, Theorem 2.3]). Sup-
pose at this point that A is in addition uniformly complete (instead of 'uniformly
complete', Henriksen and Johnson in [53] use 'uniformly closed') and denote the
identity element of A bye. Since A is unital, Ab is precisely the principal order
ideal generated bye. Also recall that A and Orth (A) are isomorphic as f-algebra
under the given condition. As usual, the f-algebra of all real-valued continuous
functions on the compact Hausdorff space M (A) is denoted by C (M (A)). Hen-
riksen and Johnson in [53, 3.2, p. 84] proved, using the Stone-Weierstrass theorem,
the following variant of Stone's representation theorem.
THEOREM 14.1. Let A be a uniformly complete f-algebra with identity element
e. Then Ab and C (M (A)) are isomorphic as f-algebras. In particular, if e is a
strong order unit in A then A and C (M (A)) are isomorphic as f-algebras.

The latter result can of course also be obtained via Kakutani's representation
theorem. Indeed, being the principal order ideal in A generated bye, Ab is a
uniformly complete Riesz space with e as a strong order unit. It follows that Ab is
an M-space with respect to the M-norm 11.ll e defined by
IIflle = inf {A > 0 : If I :'S Ae} for all f E A

(see [72, Proposition 1.2.13]). Kakutani's representation theorem (see, for instance,
[72, Theorem 2.1.3]) guarantees the existence of a compact Hausdorff space 0
so that Ab and C (0) are isomorphic as f-algebras and isometric as M-spaces.
From an investigation of the cited proof of Kakutani's theorem, we see that 0 is
the set of all algebra homomorphisms from Ab onto JR, or, equivalently, the set
of all real valued Riesz homomorphisms that send e to 1. With respect to the
weak topology a ((Abf ,Ab), where (Ab)'" is the order dual of Ab, the set 0 is a
a ((Ab)'" ,Ab)-compact Hausdorff space. In view of Theorem 13.2, we observe that
o is homeomorphic to M (Ab) and then to M (A) since M (Ab) and M (A) are
also homeomorphic (see [53, Corollary 2.8]). We thus recover Theorem 13.1 above.
The reader should also compare the above with Gelfand theory for Banach algebras
[94].
We proceed to representation theorems for non unital f-algebras. For the
terminology and notations concerning the topological spaces under consideration,
we refer the reader to the book [47] by Gillman and Jerison. The following two
theorems deal with semiprime f-algebras (see Theorems 7.2 and 7.9 in [90]).
126 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

THEOREM 14,2, Let A be a uniformly complete semiprime f-algebra with a


weak order unit, Then the following are equivalent.
(i) Every ring ideal in A is an order ideal in A.
(ii) There exists an almost compact F -space 0 such that A b and Co (0) are
isomorphic as f -algebras.
For almost compact. F-spaces, we refer to [47, 6J].
THEOREM 14.3. Let A be a semiprime Dedekind a-complete f -algebra. Then
the following are equivalent.
(i) Every ring ideal in A is an order ideal in A.
(ii) There exist a basically disconnected locally compact Hausdorff space 0 1
and a basically disconnected almost compact Hausdorff space O2 so that
Ab is isomorphic as an f-algebra to one of the algebras
CK (Ot) ,Co (0 2 ) or CK (Od EB Co (0 2 ) .

Our next representation theorem represents non semiprime f-algebras (see [90,
Theorem 7.10]).
THEOREM 14.4. Let A be a uniformly complete non semiprime f-algebra A.
Then the following are equivalent.
(i) Every ring ideal in A is an order ideal in A.
(ii) There exists a locally compact F'-space 0 such that Ab is 'isomorphic as
an f -algebra to the Cartesian product Co (0) x JR, where the multiplication
in the f-algebra Co (0) x JR is defined by
(1,.x) (g, J.t) = (1g,O) for all f, g E Co (0) and.x, J.t E JR.
Note that the topological spaces 0 in Theorem 14.2 and in Theorem 14.4 are
constructed in the same way as follows: for the uniformly complete f-algebra A
(semiprime or not), the centre Z (A) is a uniformly complete f-algebra and its
multiplicative identity IA is a strong order unit. According to Theorem 14.1, Z (A)
is isomorphic as an f-algebra to the f-algebra C (M (Z (A))) of all real-valued
continuous functions on the compact Hausdorff topological space M (Z (A)) (the
set of all maximal f-ideals in Z (A)). It turns out that
0= u {coz (1) : f E C (M (Z (A)))).
In a similar way, the topological spaces 0 1 and O2 in Theorem 14.3 are constructed
from a certain uniformly complete f-subalgebra of A (for more information, we
refer to [90]).
We end this section with two comments. Observe that representation theorems
listed in this section apply to Ab and not to the whole f-algebra A. What one can
say about A itself is the following. If A is unital then A can be embedded as an
f-algebra into an algebra of extended functions on M (A) [53, Theorem 2.3], and if
A is semiprime then A can be considered as an f-subalgebra of M (Orth (A)) (recall
that if A is semiprime then A can be considered as an f-subalgebra of Orth (A)).
In addition, the representation theorems that we presented in this section are not
necessarily excessively restrictive due to the previously mentioned fact about trans-
ference of various properties of Ab to A.
Our final comment about representing lattice ordered algebras deals with a
matter of set theory. Zaanen started an ambitious program at around 1980, in-
tending to prove all available material in vector lattices in as far as possible in an
LATTICE ORDERED ALGEBRAS 127

elementary way, i.e., without using representation theorems. A commonly quoted


reason to adhere to Zaanen's program is the need to not use the Axiom of Choice
unnecessarily. It is therefore interesting to know that one can avoid the Axiom of
Choice and still use representation theorems, as long as the constructs that one has
in mind depend on say countably many elements of a Riesz space. For f-algebras
one of the most useful theorems in that direction is the following (combine Theorem
2.2 in [33] and Corollary 2.7 in [36]).
THEOREM 14.5. Let A be a semiprime f-algebra. If D is a countable subset of
A then the f -subalgebra generated by D in A can be r-epresented within Zermalo-
F'raenkel set theory as an f -subalgebra of the space of continuous functions on a
metric space.
It should be noted at that the connection between the Axiom of Choice and
representations of f-rings in terms of sub direct products of totally ordered algebras
was discussed in the works [42] by Feldman and Henriksen, [67] by Luxemburg,
and [9] by Banaschewski. Contrary to the possibility of locally representing count-
ably many elements of any semiprime f-algebra as continuous functions without
any choice, the global representation of f-algebras as sub direct products of totally
ordered algebras can not be obtained without appealing to some transcendent tool
from set theory. The latter kind of representation theorem is important for more
than historical reasons. First of all, many researchers define f-algebras as subdirect
products of totally ordered algebras. Secondly, Birkhoff and Pierce in their seminal
paper [22] observed that the Axiom of Choice seems to be involved if one wishes
to obtain (with the definition for f-algebras as in this survey) a representation the-
orem for f-algebras as a sub direct product of totally ordered algebras, which for
convenience we will now name the Birkhoff-Pierce Representation Theorem. The
three papers [42], [67], and [9] independently prove that the Boolean Prime Ideal
Theorem is both sufficient as well as needed for the Birkhoff-Pierce Representation
Theorem. Thus Stone's Representation Theorem for Boolean algebras is construc-
tively equivalent to the Birkhoff-Pierce Representation Theorem for f-algebras. In
turn, each of the latter representation theorems is constructively equivalent to the
Kakutani Representation Theorem for vector lattices with a strong order unit. In
about that same direction, we observe that it is still unknown whether the Boolean
Prime Ideal Theorem suffices for that other main representation tool for vector lat-
tices as vector sublattices of extended real valued continuous functions, the so-called
Maeda-Ogasawara Representation Theorem (see [32]).

15. Linear biseparating maps on I-algebras


A linear map T between two algebras A and B is said to be separating if
T (I) T (g) = 0 in B whenever f 9 = 0 in A. If in addition T is bijective and its
inverse T- 1 is separating as well then T is said to be biseparating. Clearly, if T is
one-to-one and onto then T is biseparating if and only if
f9= 0 in A <=> T (I) T (g) = 0 in B.
If A and B are assumed to be f-algebras with unit elements, then T is separating
if and only if Tis disjointness preserving, that is, If I II Igl = 0 in A implies IT (1)1 II
IT (g) I = 0 in B. This follows directly from the equivalence
fg = 0 <=> If I II Igl = 0,
128 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

which holds in any semiprime I-algebra. For the same reason, the linear map T
between two I -algebras with multiplicative identity is biseparating if and only if T is
ad-isomorphism (Le., T is bijective and both T and T- 1 are disjointness preserving)
The reader is encouraged to consult the beautiful memoir [1 J by Abramovich and
Kitover for the theory of disjointness preserving linear maps on Riesz spaces.
The study of when linear biseparating maps on algebras of real or complex
valued continuous functions are weighted isomorphisms started in 1990 with the
paper [63J by Jarosz and culminated in the work [5J by Araujo, Beckenstein and
Narici with the following result. Let C (X) and C (Y) be the algebras of real or
complex valued continuous functions on completely regular topological spaces X
and Y, respectively. If T is a linear biseparating map then there exist a non-
vanishing wEe (Y) and an homeomorphism h from the realcompactification vX
of X onto v Y, such that
T (f) (y) = w (y) I (h (y)) for all lEe (X) and y E Y.
Henriksen and Smith in [55J explored the aforementioned result by Araujo, Becken-
stein and Narici in the more general setting of unital I-algebras. They proved that
every positive linear biseparating map T between two unital I-algebras A and B
closed under inversion is a weighted isomorphism, that is, there exist an invertible
wEB and a Riesz isomorphism S : A ---> B which is simultaneously an algebra
isomorphism, such that
T (f) = wS (f) for all lEA.
Very recently in [30J, Boulabiar, Buskes, and Henriksen extended the latter result
to all order bounded linear biseparating maps on arbitrary (not necessarily closed
under inversion) unital I-algebras over the reals as well as over the complex numbers
(for the theory of complex I-algebras, we refer to [17]). The theorem they obtained
is the following.
THEOREM 15.1. Let A and B be (real or complex) I-algebras with unit elements,
and let T : A ---> B be an order bounded linear biseparating map. Then T is a
weighted isomorph'ism.
In the previously mentioned memoir by Abramovich and Kitover, we find the
following theorem. A d-isomorphism between two uniformly complete Riesz spaces
A and Jyf is automatically order bounded as soon as every universally a-complete
projection band in A is essentially one-dimensional (see [1, Corollary 15.3]). This
theorem is used in [30], under the same condit.ions on A, to show that every linear
biseparating map between two uniformly complete I-algebras A and B is a weighted
isomorphism. We point out that a complex I-algebra is by definition uniformly
complete. Thus the phrase 'uniformly complete unital I-algebra' is understood
to mean either a uniformly complete unital I -algebra over the reals, or simply a
unital I-algebra over the complex numbers. For t.he proof of the next theorem, see
Proposition 5.1 and Theorem 5.2 in [30J.
THEOREM 15.2. Let A and B be unilormly complete unital I -algebras and as-
sume that every universally a-complete projection band in A is essentially one-
dimensional. Then every linear biseparating map from A onto B is order bounded
and then a weighted 'isomorphism.
To obtain the result by Araujo, Beckenstein and Narici cited above as a con-
sequence of the preceding theorem, Boulabiar, Buskes and Henriksen proved t.hat
LATTICE ORDERED ALGEBRAS 129

every algebra of all scalar-valued continuous functions on a completely regular topo-


logical space has the property that every universally a-complete projection band is
essentially one-dimensional (see [30, Theorem 5.5]).
THEOREM 15.3. Let X be a completely regular topological space X. Then every
universally a-complete projection band in the Riesz space C (X) is essentially one-
dimensional.
Combining Theorems 15.2 and 15.3, we arrive at the next result.
COROLLARY 15.4. Let X and Y be completely regular topological spaces. Then
every biseparating linear map T : C (X) -> C (Y) is a weighted isomorphism. In
particular, C (X) and C (Y) are isomorphic as f -algebras if and only if there exists
a linear biseparating map from C (X) onto C (Y).
It is well known that if X and Yare completely regular topological spaces and
S is an isomorphism from C (X) into C (Y) then there exists an homeomorphism
h from vY into vX such that S (f) = f 0 h, where vX and vY denote the re-
alcompactifications of X and Y, respectively (see Section 10 in [47]). The latter
fact, together with Corollary 15.4 above, directly leads to the next corollary, which
was proved earlier in an alternative way by Araujo, Beckenstein and Narici in [5,
Proposition 3].
COROLLARY 15.5. Let X and Y be completely regular topological. Then for
every l'inear biseparating map T : C (X) -> C (Y) there exist a non-vanishing func-
tion wE C (Y) and an homeomorphism h : vY -> vX such that
T (f) (y) = w (y) f (h (y)) for all f E C(X) and y E Y.
It is shown in [47, Theorem 8.3] that two realcompact X and Yare homeomor-
phic if and only if C (X) and C (Y) are isomorphic as f-algebras. Another classical
result of rings of continuous functions theory is that if X is a completely regular
topological space then C (X) and C (vX) are isomorphic as f-algebras (see Remark
8 (a) in [47]). It follows immediately that if X and Y are two completely regular
topological spaces, and C (X) and C (Y) are isomorphic as f-algebras, then vX and
v Yare homeomorphic. The latter implies that, without further assumptions, the
conclusion that vX is homeomorphic to vY in Corollary 15.5 is best possible. Un-
der additional assumptions, however, X and Y may themselves be homeomorphic
as is shown in the next Corollary.
COROLLARY 15.6. Let X and Y be completely regular topological spaces and
assume either (i) or (ii) below.
(i) X and Yare realcompact.
(ii) The points of X, as well as those ofY, are Go-points.
If there exists a linear biseparating map from C (X) onto C (Y) then X and Yare
homeomorphic.
Under the condition (i), the result in Corollary 15.6 above follows straightfor-
wardly from Corollary 15.5 while under the condition (ii), it stems directly from
[73] by Misra.
Finally, it should be noted that the result of Theorem 15.2 above is false if the
universally a-complete projection bands in A fail to be essentially one-dimensional.
Indeed, Example 7.13 in [1] produces a linear biseparating map T on the Dedekind
130 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

complete (and then uniformly complete) unital i-algebra Lo ([0, 1]) of all equiva-
lence classes of measurable functions on [0,1)' which is not order bounded and thus
cannot be a weighted isomorphism.

References
[1] Abramovich, Y. A. and A. K. Kitover, Inverses of disjointness preserving operators, Memoirs
Amer. Math. Soc., 143 (2000), no 679.
[2] Aliprantis. C. D. and O. Burkinshaw. Positive Operators. Academic Press, Orlando, 1985.
[3] Amemiya, I., A general spectral theory in semi-ordered linear spaces, J. Fac. Sci. Hokkaido
Univ. Ser I, 12 (1953), 111-156.
[4] Anderson, F. W., Lattice-ordered rings of quotients, Canad. J. Math., 17 (1965), 434-448.
[5] Araujo, J., E. Beckenstein and L. Narici, Biseparating maps and homeomorphic real-
compactifications, J. Math. Ana. Appl., 12 (1995), 258-265.
[6] Arens. R, Operations induced in function classes. Monatshelte Math., 55 (1951), 1-19.
[7] Arens, R, The adjoint of bilinear operation, Proc. Amer. Math. Soc., 2 (1951), 839-848.
[8] Arens, R. and J. L. Kelley, Characterizations of the space of continuous functions over a
compact Hausdorff space, Trans. Amer. Math. Soc., 62 (1947), 499-502.
[9] Banaschewski, B., The Prime Ideal Theorem and representations of I-rings, Algebra Univ., 25
(1988), 384-387.
[10] Basly, M. and A. Triki, F F-algebres Achimediennes, Preprint.
[11] Basly, M. and A. Triki, I-algebras in which order ideals are ring ideals, Indag. Math. 50
(1988), 231-234.
[12] Basly, M. and A. Triki, On uniformly closed ideals in I-algebras, 2nd Conlerence, Functions
spaces, Marcel Dekker (1995), 29-33.
[13] Bernau, S. and Huijsmans, C. B., Almost I-algebras and d-algebras, Math. Proc. Cambridge.
Phil. Soc. 107 (1990), 287-308.
[14] Bernau, S. J. and C. B. Huijsmans, The order bidual of Almost I-algebras and d-algebras,
Trans. Amer. Math. Soc. 347 (1995), 4259-4274.
[15] Bernau, S. J. and C. B. Huijsmans, The Schwarz inequality in Archimedean I-algebras, Indag.
Math. 7 (1996), 137-148.
[16] Beukers, F. and C. B. Huijsmans, Calculus in I-algebras, J. Austral. Math. Soc. (Ser. A) 37
(1984), 110-116
[17] Beukers, F., C. B. Huijsmans and B. de Pagter, Unital embedding and complexification of
I-algebras, Math, Z., 183 (1983), 131-144.
[18] Bigard, A. and K. Keimel, Sur les endomorphismes conservant les polaires d'un groupe reticule
archimedien, Bull. Soc. Math. France, 97 (1969), 381-398.
[19] Bigard, A., K Keimel and S. Wolfenstein, Groupes et Anneaux Reticules. Lecture Notes in
Math. vo!' 608, Springer Verlag, 1977.
[20] Birkhoff, G., Lattice ordered groups, Annals of Math., 43 (1942), 298-331
[21] Birkhoff, G., Lattice Theory, 3rd Edition. Amer. Math. Soc. Colloq. Pub!. no. 25, American
Mathematical Society 1967.
[22] Birkhoff, G. and R S. Pierce. Lattice-ordered rings, An. Acad. Brasil. Cienc. 28 (1956),
41-69
[23] Boulabiar, K., A relationship between two almost I-algebra products, Algebra Univ., 43
(2000), 347-367
[24] Boulabiar, K, Products in almost I-algebras, Comment. Math. Univ. Carolinae, 41 (2000),
747-759.
[25] Boulabiar, K, A representation theorem for d-rings, Submitted.
[26] Boulabiar, K, Positive derivations on almost I-rings, To appear in Order.
[27] Boulabiar, K, On positive orthosymmetric bilinear maps, Submitted.
[28] Boulabiar, K and E. Chi!, On the structure of almost I-algebras, Demonstratio Math., 34
(2001), 749-760.
[29] Boulabiar, K and M. A. Toumi, Lattice bimorphisms on I-algebras, Algebra Univ. 48 (2002),
103-116.
[30] Boulabiar, K., Buskes, G. and Henriksen, M., A Generalization of a Theorem on Biseparating
Maps, To appear in J. Math. Ana. Appl.
LATTICE ORDERED ALGEBRAS 131

[31] Buskes, G., B. de Pagter and A. van Rooij, Functional calculus on Riesz spaces, Indag. Math.,
2 (1991),423-436.
[32] Buskes, G. and A. van Rooij, Riesz spaces and Ultrafilter Theorem, Compositio Math. 83
(1992), 311-327.
[33] Buskes, G. and A. van Rooij, Small Riesz spaces, Math. Proc. Cambridge Phil. Soc., 105
(1989), 523-536.
[34] Buskes, G. and A. van Rooij, Almost I-algebras: commutativity and Cauchy-Schwarz in-
equality, Positivity, 4 (2000), 227-231.
[35] Buskes, G. and A. van Rooij, Almost I-algebras: structure and the Dedekind completion,
Positivity, 4 (2000), 233-243.
[36] Buskes, G. and A. van Rooij, Squares of Riesz spaces, Rocky Mountain J. Math. 31 (2001),
45-56.
[37] Chil, E., Structure and Dedekind completion of d-algebras, Submitted.
[38] Conrad, P., The additive group of an I-ring. Canad. J. Math. 25 (1974), 1157-1168.
[39] Conrad, P. and J. Diem, The ring of polar preserving endomorphisms of an abelian lattice-
ordered group, Illinois J. Math., 17 (1971), 222-240.
[40] Colville, P., G. Davis and K. Keimel, Positive derivations on I-rings, J. Austral. Math. Soc.
23 (1977) 371-375.
[41] Ellis, A. J., Extreme positive operators, Quart. J. Math. Oxlord, 15 (1964), 342-344
[42] Feldman, D. and M. Henriksen, I-Rings, subdirect products of totally ordered rings, and the
Prime Ideal, Indag. Math., 50 (1998), 121-126.
[43] Fremlin, D. H., Tensor products of Archimedean vector lattices, Amer. J. Math. 94 (1972),
778-798.
[44] Frink, 0., MR 13,718c, Review of Birkhoff's 1950 address to the International Congress of
Mathematicians in Cambridge, Massachusetts
[45] Fuchs, L., Partially Ordered Algebraic Systems, Pergamon Press, Oxford-London-New York-
Paris, 1963.
[46] Grobler, J. J., Commutativity of the Arens product in lattice ordered algebras, Positivity, 3
(1999), 357-364.
[47] Gillman, L. and M. Jerison, Rings 01 Continuous Functions, Springer Verlag, Berlin-
Heidelberg-New York, 1976.
[48] Hager, A. W., and L. C. Robertson, Representing and ringifying a Riesz space, Symposia
Math., 21 (1977), 411-431.
[49] Henriksen, M., Semiprime ideals of I-rings, Symposia Math., 21 (1977), 401-409.
[50] Henriksen, M., A survey of I-rings and some of their generalizations, Ordered algebraic struc-
tures, Kluwer Acad. Publ., Dordrecht, (1997), 1-26.
[51] Henriksen, M., and J. R. Isbell, Lattice-ordered rings and functions spaces, Pacific J., 12
(1962), 533-565.
[52] Henriksen, M., S. Larson and F. A. Smith, When is every order ideal a ring ideal?, Comment.
Math. Univ. Carolinae, 32 (1991) 411-416.
[53] Henriksen, M. and D. G. Johnson, On the structure of a class of archimedean lattice-ordered
algebras, Fund. Math. 50 (1961), 73-94.
[54] Henriksen, M. and F. A. Smith, Some properties of positive derivations on I-rings, Ordered
fields and real algebraic geometry, Contemp. Math., 8, Amer. Math. Soc., (1982), 175-184.
[55] Henriksen, M. and F. A. Smith, A look at biseparating maps from an algebraic point of view,
Real Algebraic Geometry and Ordered Structures, Contemp. Math., 253, Amer. Math. Soc.,
(2000), 125-144.
[56] Huijsmans, C. B., The order bidual of lattice ordered algebras II, J. Operator Theory 22
(1989), 277-290
[57] Huijsmans, C. B., Lattice ordered algebras and I-algebras: A survey, in: Studies in Economic
Theory 2, Positive Operators, Springer Verlag, Berlin, 1991.
[58] Huijsmans, C. B. and B. de Pagter, Ideal theory in I-algebras, Trans. A mer. Math. Soc. 269
(1982), 225-245.
[59] Huijsmans, C. B. and B. de Pagter, Subalgebras and Riesz subspaces of an I-algebra, Proc.
London, Math. Soc., 48 (1984), 161-174.
[60] Huijsmans, C. B. and B. de Pagter, The order bidual of lattice ordered algebras, J. Funct.
Anal., 59 (1984), 41-64.
132 KARIM BOULABIAR, GERARD BUSKES, AND ABDELMAJID TRIKI

[61] Huijsmans, C. B. and B. de Pagter, Averaging operators and positive contractive projections,
J. Math. Ana. Appl., 113 (1986), 163-184.
[62] Ionesco Tulcea, A. and C. Ionesco Tulcea, On the lifting property I, J. Math. Anal. Appl., 3
(1961), 537-546.
[63] Jarosz, K., Automatic continuity of separating linear isomorphisms, Bull. Canadian Math.
Soc., 33 (1990), 139-144.
[64] Johnson, D. G., The completion of an archimedean I-ring, J. London Math. Soc, 40 (1965),
493-493.
[65] Kudhicek, V., On some types of R-rings. Sb. Vysoke. Ucen{ Tech. Bmo 1-2 (1962), 179-181.
[66] Lambek, J. Lectures on Rings and Modules, Blaisdell, 1966.
[67] Luxemburg, W. A. J., Some Aspects 01 the Theory 01 Riesz Spaces, Univ. Arkansas Lecture
Notes Math. 4, Fayetteville, 1979
[68] Luxemburg, W. A. J., A remark on a paper by D. Feldman and M. Henriksen concerning the
definition of I-rings, Nederl. Akad. Wetensch. Indag. Math., 50 (1998), 127-130.
[69] Luxemburg, W. A. J. and A. C. Zaanen, Riesz spaces I, North-Holland, Amsterdam, 1971.
[70] Lozanovsky, G., The functions of elements of vector lattices, Izv. Vyssh. Uchebn. Zaved.
Mat., 4 (1973), 45-54.
[71] Martinez, J. The maximal ring of quotient I-ring. Algebra Univ . 33 (1995), 355-369.
[72] Meyer-Nieberg, P., Banach Lattices, Springer Verlag, Berlin Heidelberg New York, 1991.
[73] Misra, P. R., On isomorphism theorems for C(X), Acta. Math. Acad. Sci. Hungar., 39 (1982),
179-180.
[74] Nagasawa, M., Isomorphisms between commutative Banach algebras with an application to
rings of analytic functions, Kodai Math. Semin. Rep., 11 (1959), 182-188.
[75] Nakano, H., Modern Spectral Theory, Tokyo Math. Book Series II, Maruzen, Tokyo, 1950.
[76] de Pagter, B., I-algebras and orthomorphisms (Thesis, Leiden, 1981).
[77] Phelps, R. R., Extremal operators and homomorphisms, Trans. Amer. Math. Soc., 108
(1963), 265-274.
[78] van Putten, B., Disjunctive linear operators and partial multiplication in Riesz spaces, (The-
sis, Wageningen, 1980).
[79] Quinn, J., Intermediate Riesz spaces, Pacific J. Math., 56 (1975), 225-263.
[80] Scheffold, E., FF-Banachverbandsalgebren, Math. Z., 177 (1981),193-205.
[81] Scheffold, E., Der Bidual von F-Banachverbandsalgebren, Acta Sci. Math., 55 (1991), 167-
179.
[82] Scheffold, E., Uber Bimorphismen und das Arens-Produkt bei kommutativen D-
Banachverbandsalgebren, Preprint.
[83] Scheffold, E., Uber den ordnungsstetigen Bidual von FF-Banachverbandsalgebren, Arch.
Math., 60 (1993), 473-477.
[84] Steinberg, S. A., Quotient rings of a class of lattice-ordered rings, Can. J. Math., 25 (1973),
627-645.
[85] Steinberg, Stuart A. On the unitability of a class of partially ordered rings that have squares
positive. J. Algebra 100 (1986), no. 2, 325-343.
[86] Steinberg, Stuart A. On lattice-ordered algebras that satisfy polynomial identities. Ordered
algebraic structures (Cincinnati, Ohio, 1982), 179-187, Lecture Notes in Pure and Appl. Math.,
99, Dekker, New York, 1985.
[87] Steinberg, Stuart A. Unital $l$-prime lattice-ordered rings with polynomial constraints are
domains. Trans. Amer. Math. Soc. 276 (1983), no. 1, 145-164.
[88] Steinberg, Stuart A. On lattice-ordered rings in which the square of every element is positive.
J. Austral. Math. Soc. Ser. A 22 (1976), no. 3, 362-370.
[89] Toumi, M. A., On some I-subalgebras of d-algebras, To appear in Math. Reports
[90] Triki, A., Stable I-algebras, Algebra Univ., 44 (2000) 65-86.
[91] Triki, A., On algebra homomorphisms in complex almost I-algebras, Comment. Math. Univ.
Carolinae, 43 (2002), 23-31.
[92] Zaanen, A. C., Riesz Spaces II, North Holland, Amsterdam-New York-Oxford, 1983.
[93] Zaanen, A. C., Introduction to Operator Theory in Riesz Spaces, Springer Verlag, Berlin,
1997.
[94] Zelazko, W., Banach algebras, Elsevier, Amsterdam-London-New York, 1973.
LATTICE ORDERED ALGEBRAS 133

DEPARTEMENT DU CYCLE AGREGATIF, INSTITUT PREPARATOIRE AUX ETUDES SCIENTIFIQUES


ET TECHNIQUES, UNIVERSITE DU 7 NOVEMBRE A CARTHAGE, BP 51, 2070-LA MARSA, TUNISIA
E-mail address:karim.boulabiarlipest.rnu.tn
DEPARTMENT OF MATHEMATICS, UNIVERSITY OF MISSISSIPPI, UNIVERSITY, MS-38677, USA
E-mail address: mmbuskeslsunset. olemiss. edu

DEPARTEMENT DE MATHEMATIQUES, FACULTE DES SCIENCES DE TUNIS, UNIVERSITE TUNIS EL


MANAR, 1060-TUNIS, TUNISIA
E-mail address:abdelmajid.trikilfst.rnu.tn
Contemporary Mathematics
Volume 328, 2003

An extension of a theorem of Wermer, Bernard, Sidney and


Hatori to algebras of functions on locally compact spaces

Eggert Briem

ABSTRACT. It follows from a theorem of J. Wermer that if A is a uniformly


closed algebra of continuous complex-valued functions vanishing at infinity
on a locally compact Hausdorff space X, the property that b2 E Re A for all
bE Re A implies A = Co(X). In other words, if the function h(t) = t 2 operates
by composition on ReA then A = Co(X). This result was generalized by O.
Hatori. He proved that in place of the function h(t) = t 2 one can put any
real-valued function defined in a neighbourhood of 0, thus extending a similar
result for the compact case. Here a simple alternative proof is given for the
locally compact case.

1. Introduction
Let X be a compact Hausdorff space and B a uniformly closed subspace of CIR(X),
the space of all continuous real-valued functions on X, which separates the points of
X and contains the constant functions. A version of the Stone-Weierstrass theorem
says that if b2 E B for all b E B then B = CIR(X). Clearly this result does not
hold if, instead of assuming that B is uniformly closed, one assumes that B is a
Banach space in some norm which dominates the sup-norm, as the example of any
non-trivial real Banach function algebra shows. However, if B is the real part of
a uniform algebra, a theorem of J. Wermer says that if b2 E B for all b E B then
B = CIR(X).
Let us say that a real-valued function h, defined on an interval I of the real line,
operates on B if hob E B whenever b E Band b maps X into I. Thus b2 E B for
all bE B means that h(t) = t 2 operates on B. The Stone-Weierstrass theorem W8.',
generalized by K. de Leeuw and Y. Katznelson (see [4], theorem 4.21), they showed
that h(t) = t 2 can be replaced by any continuous non-affine function (Le. a function
not ofthe type h(t) = o:t+!3) defined on an interval, and the theorem of J. Wermer
was similarily generalized by A. Bernard, S. Sidney and O. Hatori [1], [5] and [8].
Here one can even do without the continuity assumption, a function operating on
the real part of a uniform algebra of infinite dimension is automatically continuous.

1991 Mathematics Subject Classification. Primary 46JlOj Secondary 46E15.


Key words and phmses. uniform algebra, locally compact space, functional calculus.

2003 American Mathematical Society

135
136 EGGERT BRIEM

In the case where X is locally compact, the functional calculus for a uniformly
closed subspace of Co(X, JR), the space of all continuous real-valued functions van-
ishing at infinity on X, may be non-trivial (cf. [2]). What then about the real part of
a uniformly closed subalgebra of Co(X), the space of all continuous complex-valued
functions vanishing at infinity on X? Wermer's theorem, [10], clearly applies to
this situation. It turns out that the functional calculus for the real part of a uni-
formly closed subalgebra, which is not a C* -algebra, of Co(X) is trivial. This result,
which is due to O. Hatori, is to be found in [7]. To prove this one has to do more
than just adapt the proofs for the compact case to the locally compact situation.
The purpose of this paper is to give a simple alternative proof using antisymmetric
decomposition of uniform algebras.
2. Proofs and results
Let X be a locally compact Hausdorff space and let Xl denote the one-point
compactification of X, with Xoo denoting the point at infinity. The functions in
Co(X) have a natural extension to functions in C(Xl)' the space of all continuous
complex-valued functions on Xl. If A is a uniformly closed subalgebra of Co(X),
that separates the points of X and does not vanish identically at any point of X,
then Al = AEI1C is a uniformly closed subalgebra of C(Xd that separates the points
of Xl and contains the contant functions and A = {a E All a(x oo ) = O}. Thus,
we can assume that there is a uniformly closed subalgebra Al of C(Xd containing
the constant functions, where Xl is compact, and a point Xoo in Xl, such that A
is obtained by restricting the functions in the set {a E Al Ia(x co ) = O} to the set
X = Xl \ {xoo}.
We need some results from the theory of uniform algebras. A subset E of Xl is a
set of antisymmetry for Al if the only functions in Al that are real-valued on E are
constant functions. Maximal sets of antisymmetry are closed generalized peak sets
(intersections of peak sets), they form a partition of Xl and a continuous complex-
valued function f on Xl is in Al if and only if its restriction to each maximal
set of antisymmetry E is in the space AIlE of restrictions of functions in Al to
E. Thus Al(Xd = C(Xd if and only if each maximal set of antisymmetry is a
singleton. Also, AIlE is uniformly closed for each maximal set of antisymmetry,
there is actually for each a in AIlE an element a' in Al with a' = a on E and
II a' 1100=11 a llco,(E). (The latter norm is the supnorm w.r.t E). This material can
f. ex. be found in [3].
If Al is the disc algebra on the closed unit disc in the complex plane and A is
the algebra of functions vanishing at the origin, then any non-zero function in A
takes real values of opposite signs in any neighbourhood of O. Thus any real-valued
function, defined on an interval which is not a neighbourhood of 0, operates trivially
on Re A. We shall therefore always assume that the interval on which an operating
function is defined is a neighbourhood of O.
We can now state the main result of this note.
THEOREM 1. Let A be a uniformly closed subalgebra of Co(X) which separates
the points of X and does not vanish identically at any point in X and let Re A be
the space of real parts of functions in A. If Re A has a non-affine operating function
h: I ........ JR where I, an interval, is a neighbourhood of 0, then Re A = Co(X, JR) and
thus also A = Co(X).
ALGEBRAS ON LOCALLY COMPACT SPACES 137

The first step in the proof of Theorem 1 is to show that operating functions are
continuous. To prove this we need the following lemma.
LEMMA 1. Let E be a maximal set of antisymmetry for A1 and let b E Re A.
Then b(E) is either an interval or a singleton.
Proof. Suppose b(E) is not connected, and that b = Rea for some a in A 1. Then
we can find two disjoint closed rectangles Ro and R1 in the complex plane such
that a(E) is a subset of Ro u R1 but not a subset of anyone of the two rectangles.
By Runge's theorem there is sequence of polynomials (Pn) converging uniformly to
o on Ro and to 1 on R 1. Since AdE is uniformly closed, the sequence (Pn 0 a)
converges on E to an element of A1IE. But the sequence converges on E to a
function which takes only the values 0 and 1, contradicting the fact that E is a set
of antisymmetry. 0

LEMMA 2. Let E be a maximal set of antisymmetry for A 1 such that E\ {xoo } =I- 0.
Then there is an Xo E E such that for every to E I (resp. to in the interior of I),
there exists bo E ReA with bo(xo) = to and bo(X) C I (resp. bo(X) is a subset of
the interior of I).
Proof. If to = 0, the existence of the desired function is easily proved. We
consider the case where to > O. (The proof for the case to < 0 is similar.) Let
Xo be a generalized peak point in E, other than x oo , for A 1 Simple calculations
show that we can take 1 E A with I(xo) = 11/1100 = 1. Let c: be a positive number
such that (-2c:, to] C I (resp. (-2c:, to] is a subset of the interior of I), and R the
closed rectangle with corners -c: i and to i. Let <p be a conformal map from the
open unit disc onto the interior of R with <p(0) = O. By Carath6dory's theorem <p
can be extended to a homeomorphism from the closed unit disc onto R. We may
assume that <p(1) = to. By Runge's theorem <p can be uniformly approximated
on the closed unit disc by polynomials {Pn} with Pn(O) = 0, and thus Pn 0 f E A
converges uniformly to <p 0 I. Then Re <p 0 1 is the desired function. 0

PROPOSITION 1. Let h : I --+ ~ be an operating function for Re A. Then, unless


A = C o(X,IR), h is continuous.
Proof. If A =I- Co(X,~), there is a maximal set of antisymmetry E for A1 which
contains more than one point. Thus E \ {xoo} =I- 0. Let t be an arbitrary point in
I. Then by Lemma 2 there exists and x E E and bEReA such that b(x) = t and
b(X) C I. Since hob is in ReA, hob is continuous. It follows that h is continuous
at t. We conclude that h is continuous on its domain of definition.
If A = Co(X), h need only be continuous at 0, f. ex., if A = Co. 0

In the compact case, where A contains the constant functions, Re A is dense in


CR(X) (cf. Proposition 4 below). Here we obtain a similar result.
PROPOSITION 2. Let h : I --+ IR be an operating function for ReA which is
non-affine in every neighbourhood of an interior point to for I. Let Xo E X and
suppose that there is a function bo E ReA which maps X into the interior of I and
satisfies bo(xo) = to Then there is an open neighbourhood Xo of Xo such that every
f E Co (X, ~) which vanishes outside Xo can be approximated uniformly on X by
elements of ReA.
138 EGGERT BRIEM

Proof. We are assuming that X is locally compact and not compact so that
h(O) = o. We may assume that h is continuous by Proposition 1. Let us first look at
the case where h is a polynomial of degree at most three in some neighbourhood of 0
but not affine there. Then bed E Re A for all b, e, d E Re A and hence fb E cl(Re A)
for all b E cl(Re A) and all f in the algebra generated by Re A. Here cl(Re A)
denotes the uniform closure of Re A. It follows that
Co(X,JR) cl(ReA) ~ cl(ReA)
and hence ReA is dense in Co(X,JR).
Suppose first that h is affine in some neighbourhood ofO. Replacing h by h(t)-,t
we may assume that h = 0 in a neighbourhood of o. Put e = tC; l bo.
For bEReA, r,t E JR and '{) E CO'(JR) the function

J h 0 (rb o + tb - se)'{)(s)ds
is in cl(Re A) if r is close to 1, t is close to 0 and the support of'{) is contained in a
small neighbourhood of 0, so that the expression above is defined. We differentiate
twice w.r.t. 0, put t = 0 to obtain

b2 e- 2 J h 0 (rbo - se)'{)"(s)ds E cl(ReA),

for all bEReA. (If le(x)1 is small the expression is 0.) Put

d= J h 0 (rb o - se)'{)"(s)ds.

We have
d(xo) = J h 0 (rto - s)'{)"(s)ds = (h * '{))"(rto).
Since h is not affine in any neighbourhood of to we can choose '{) and r, where the
support of'{) is contained an arbitrarily small neighbourhood of 0 and r is arbitrarily
close to 1 such that d(xo) '" O. Let
Xo = {x E Xld(x) '" O}.
Since b2 e- 2 d E cl(Re A) for all b E Re A it follows that
b1 ,b2 e- 2 d E cl(ReA)
for all b1 , b2 E Re A. Put
M(ReA) = {J E Co(X,JR) If cl(ReA) ~ cl(ReA)}.
Above we have be- 2 d
E M(ReA) for bEReA. Since these functions separate the
points of Xo and since M(ReA) is an algebra we deduce that cl(ReA) contains
every f E Co(X,JR) which vanishes outside Xo.
Suppose now that h is not affine in any neighbourhood of o. For e E Re A let
B(e) = {u E cl(ReA) 13>' E JR+ s.t. lui::::; >'Iel}.
If b ERe A and b + ie E A then be and be(b2 - e2 ) belong to B(e).
Let also
M(e) = {g E Co(X, JR) Ig. cl(B(e)) ~ cl(B(e))},
a subalgebra of Co(X, JR). We note that since h is continuous h also operates on
cl(ReA).
ALGEBRAS ON LOCALLY COMPACT SPACES 139

For bl ERe A, for u E B(e), r, t E lR and <p E GOO(lR) the function

(1) J h 0 (rb l + tu - se)<p(s)ds

is in cl(Re A), if rand t are sufficiently small and the support of <p is contained in
a sufficiently small neighbourhood of 0, so that the function rbl + tu - se maps X
into the interval (0:, /3) if s E supp <po Let i = 1,2 or 3. We differentiate (1) i times
w.r.t. t, put t = 0 and find that

uie- i J ho (rb l - se)<p(il(s)ds E cl(ReA).

Put
d= J h 0 (rb l - se)<p(3) (s)ds.
With i = 2 and <p' instead of <p we get
u 2 e- 2 d E cl(Re A)
for all u E B(e) and hence
(2)
for all Ul, U2 E B(e). With i = 3 and u = be E B(e) in (1) we get
b3 d E cl(ReA),
and with i = 1, u = beW - e2 ) E B(e) and <p" in place of <p in (1) we get
b(b2 - e2 )d E cl(Re A)
and thus
be2 d E cl(ReA).
If we put U2 = be2 d in (2) we find that
ubd2 E cl(Re A)
for all u E B(e). Since u E B(e) implies ubd2 E B(e) we deduce that
bd 2 E M(e).
Let us determine how well the functions bd2 separate the points of the set
Xo = {x E X Ib(x)e(x) i- a}.
Let x, y E Xo. Suppose that does not separate x and y for every d. Since A is
bd 2
an algebra we can choose bl E ReA such that bl(x) = 0 and bl(y) = 1. Then
d2 (y) = (b(x)/b(y))d 2 (x).
The right hand side is independant of r so that

d(y) = J h 0 (r - se(y))<p(3) (s)ds,

a differentiable function of r, is constant in for r in some neighbourhood of O. It


follows that
J
h 0 (r - se(y))<p(4l(s)ds = 0
for all r in a neighbourhood of O. Since this holds for all <p with support in some
neighbourhood of 0, we deduce that h is a polynomial of degree at most three in
some neighbourhood of O. As we have already seen this implies that cl(ReA) is
140 EGGERT BRlEM

dense in Co(X,JR). Otherwise, the functions bd2 separate the points of X o, and
since these functions vanish outside X o, it follows that
{f E Co(X, JR) I I = 0 on Xo \ Xo} ~ M(c).
Thus by a simple calculation we have
{f E Co(X,JR) I I = 0 on Xo \ Xo} ~ cl(ReA).
o
To proceed to the proof of Theorem 1 we need a local version of the so called
Bernard's Lemma [1].
Let A be an infinite set and let C be a Banach space with norm I . II. Put
eOO(C) = {{c.d ICA E C and sup I CA 11< oo},
AEA
where {cA} denotes a function from A into C such that {cA}(a) = Co E C for each
a E A. Then eOO(C) is a Banach space in the norm
I {cA } 11= sup II CA I .
AEA
The space Re A is a Banach space in the norm
lib 11= inf{11 a 1100 Ia E A and b = Rea},
and clearly lib 1100::;11 b II. We thus have
eOO(ReA) ~ eOO(Co(X,JR)).
Bernard's lemma (see [1]) says that if eOO(ReA) is dense in eOO(Co(X,JR)) then
ReA = Co(X,JR).
To prove that eOO(ReA) is dense in eOO(Co(X,JR)), we represent eOO(Co(X,JR)) as
a space of functions on a compact Hausdorff space. We identify Co(X, JR) with a
subspace of CIR(Xd in the obvious manner.
Let A have the discrete topology and let (3(A x Xd) be the Stone-Cech compact-
ification of Ax Xl' The space eOO(C(xd) can be identified with C({3(A x Xd) in
a natural way
{fA}
FUll} --+
where F{!lI}b,x) = 1'Y(x) for each b,x) in A x Xl' Thus eOO(Co(X,JR)) is a
subspace of CIR ({3(A x Xl)) and we have the inclusions
eOO(ReA) ~ eOO(Co(X,JR)) ~ CIR({3(A x Xd).
The subsets of the Stone-Cech compactification may be complicated but the subsets
we are interested in have a simple description. For each I in C(Xl ) let {f} denote
the net {fA} where IA = I for each A
Let now Po be in {3(A x Xd. The map
I --+ {f }(Po)
is a homomorphism of C(X l ) and is thus given by point-evaluation at some point
Xo in Xl' Put
5: 0 = {p E (3(A x Xd I {f}(p) = I(xo) VI E CIR(Xl )}.
ALGEBRAS ON LOCALLY COMPACT SPACES 141

The sets x form a partition of {3(A x Xt} into closed sets. Their significance stems
from the following local version of Bernard's lemma, similar to the one given by O.
Hatori [6].

Bernard's lemma (local version) Let A denote the unit ball of Co(X,JR.), let x
be in X and suppose that the restriction space lOO(ReA)lx is dense in CIR(x). Then
there is a compact neighbourhood Kx of x such that ReAIKx = CIR(Kx )'
Proof. For each A in A let 1>.. = A. By assumption there is an element {b A } in
lOO(ReA) such that we have the inequality
I{fA} - {bA}1 < 1/2
on x and hence also on some open subset U of {3(A x Xt} containing X. By a simple
calculation we have
Xo = {p E (3(A x Xl) I I{f}(p) - f(xo)1 :=; ~ Vf E CIR(Xl )}.
Thus there is a function fo in CIR(Xt} such that the set
{p E {3(A x Xt} I I{fo}(p) - fo(x)1 :=; 1/2}
is a subset of U. Put
Kx = {y E Xd Ifo(Y) - fo(x)1 :=; 1/2},
a compact neighbourhood of x. Then A x Kx is contained in U so that
IfA - bA I :=; 1/2
on Kx for all A in A. Let M = SUPA II bA II, a finite quantity because (b A ) is in
lOO(Re A). Take any f in Co(X, JR.) with II f lloo:=; 1. By induction we construct a
sequence (b n ) of elements from ReA, with II bn II:=; M for all n, such that
n-l
12n( f - L 2- i bi ) - bnl < ~
i=O
on Kx. The function b = E:o 2- b is in Band b = f on Kx. 0
i i

PROPOSITION 3. Let Xo E X and suppose that there is an open neighbourhood


Xo of Xo such that every function in Co(X,JR.), which vanishes outside Xo can be
unifromly approximated on X by elements from ReA. Then lOO(ReA) sepamtes
the points of xo.
Proof. Suppose p,q are in xo. There is an element {fA} in lOO(Co(X,JR., where
each 1>.. vanishes outside X o, such that {fA}(P) = 0 and {fA}(q) = 1. By assumption
we can for each A E A find a function bA in ReA with II fA - bA lloo:=; ~. Thus
we have {bA}(p) :=; ~ and {bA}(q) ~ ~ since II{fA} - {bA}lIoo,{3(Axxt} :=; ~. Let
a A = bA + iC A be in A for each A. Now, the A-net {e a ... - 1} is in lOO(A) and it
separates p from q since {e an } does. Since
I{e a ... }(p)1 = {e b... }(p) :=; e l / 4 and I{e a ... }(q)1 = {e b... }(q) ~ e3 / 4 ,

we see that lOO(A) separates p from q and thus lOO(ReA) does so as well. 0
142 EGGERT BRIEM

We also need the aforementioned extension of the Stone-Weierstrass theorem due


to de Leeuw and Katznelson (see [4], Theorem 4.21). We can prove a version of
this result, Proposition 4, in the same way as the original one, the proof is omitted.
PROPOSITION 4. Let Y be a compact Hausdorff space and B a subspace ofCR(Y)
which separates the points of Y and contains the constant functions. Suppose that
B is also a normed space with the norm II liB which dominates the uniform norm.
Let h be a continuous function defined on an interval I. Suppose that h is non-
affine in every neighbourhood of an interior point to in 1. Suppose that there exists
a positive real number 8 > 0 with (to - 8, to + 8) ~ I such that h 0 (to + u) E B for
every u E B with Ilull < 8. Then B is uniformly dense in CIR(Y).
Proof of Theorem 1. We are going to use the local version of Bernard's Lemma.
We thus have to show that the conditions stated there are satisfied. For this we
use the result of de Leeuw and Katznelson above. The main obstacle is that we
can not conclude that if {b.>.} is in [00 (Re A) then h e {b.>.} is in [00 (Re A) although
the composite function is defined, i.e. we can not conclude that composition with h
maps bounded nets to bounded nets. To overcome this difficulty we use a method
of Sidney, [8], to obtain local boundedness for composition with h.
Suppose A =I- Co(X) and thus also Al =I- C(X 1 ). Let E be a maximal set of
antisymmetry for Al containing more than one point. Let further to be an interior
point of I such that h is not affine in any neighbourhood of to and let bo be a
function in Re A which maps X into the interior I for which to is an interior point
of bo(E). By Lemma 2 such a function exists.
We now choose E > 0 such that if b is in the E-ball, (ReA)., of ReA then ho(bo+b)
is defined and to is an interior point of (b o + b)(E). We write
(ReA). = U{b E (ReA). I I h 0 (bo + b) II::; n}.
n
The Baire Category Theorem shows that the closure of one of the sets on the right
hand side has an interior point and thus there is a function b1 E Re A and positive
numbers 8, lvI, with I b1 I +8 < E, such that
I he (bo +b 1 +c) II::; lvI
for c in a dense subset of the 8-ball of Re A. Let b = bo + b1 , and let [00 (Re A) be
as in the local version of Bernard's Lemma. We then have
h 0 {b + c.>.} E cl([OO(Re A)),
the uniform closure of [OO(ReA), if {c.>.} E [OO(ReA) and I {c.>.} 11< 8. Take Xo E E
such that b(xo) = to, an interior point of b(E).
We restrict to Xo and deduce that
h 0 {to + c.>.} E cl([OO(ReA)lxo),
for every {c.>.} E [OO(ReA)lxo whose quotient norm satisfies II {c.>.} 11< 8. Since {c}
is constant on xo for any c E ReA, the space [OO(ReA)lxo contains the constant
functions, it also separates the points of Xo by Proposition 2 and 3. Then Propo-
sition 4 shows that [OO(ReA)lxo is dense in CIR(xo) and thus, by the local version
of Bernard's Lemma, ReAIK = CIR(K) for some compact neighbourhood K of Xo.
The theorem of Sidney and Stout, [9], then shows that AIK = C(K).
By Proposition 2, there exists an open neighbourhood Xo of Xo such that every
f E Co(X, 1R) which vanishes outside Xo can be approximated uniformly on X by
ALGEBRAS ON LOCALLY COMPACT SPACES 143

elements of ReA. We may assume K c Xo. Let Kl be a compact neighbourhood


of Xo such that K 1 is in the interior of K. Since E is antisymmetric and contains
more than one point, EnK l contains more than one point. On the other hand, we
will show that
{f E Co(X, IR) II = 0 outside Kd CAl.
It will follow that En Kl = {xo}, which will be a contradiction proving A =
Co(X, IR). Let 1 E Co(X, IR) with 1 = 0 outside K 1 . A function u E Co(X, IR) with
lui::; 1 on X, u = 1 on K l , and u = 0 outside K can be uniformly approximated
by functions in ReA. Thus, for every positive integer n, there exists bn E ReA such
that 1 - ,& ::; bn ::; 1 on K 1 and bn ::; ,& outside K. Without loss of generality
we may assume that Ibnl ::; 1 on X. Take Cn E A with Recn = bn and put
an = (eCn-l)n. Then an E AI, e-1. ::; lanl ::; Ion Kl and lanl ::; e- n+1. outside K.
Since AllK = C(K) and thus AllKl = C(Kd, there exists a positive real number
M such that for every positive integer n there exists gn E Al such that gnan = 1
on Kl with IIgnlloo ::; M. Since AllK = C(K), there is a function af E Al with
af = 1 on K. Then afgnan E Al and by a simple calculation lIafgnan - 11100 --+ 0
as n --+ 00, that is, 1 E AI' 0
References
[lJ A. Bernard, Espaces des parties relies des Iments d'une algebre de Banach de fonctions, J.
Funct. Anal. 10 (1972), 387-409.
[2J E. Briem, Approximations from Subspaces of Co(X), J. Approx. Theory 112 (2001), 279-294.
[3J A. Browder, Introduction to Function Algebras, W. A. Benjamin, Inc. !969.
[4J R.B. Burckel, Characterizations of C(X) among its subalgebras (Lecture Notes in Pure and
Appl. Math. 6). Marcel Dekker, New York 1972.
[5J O. Hatori, Functions which operate on the real part of a uniform algebra, Proc. Amer. Math.
Soc. 83 (1981), 565-568.
[6J O. Hatori, Separation properties and operating functions on a space of continuous functions,
lntemat. J. Math. 4 (1993), 551-600.
[7J O. hatori, Range transformations on a Banach function algebra. IV, Proc. Amer. Math. Soc.
116 (1992), 149-156.
[8J S. J. Sidney, Functions which operate on the real part of a uniform algebra, Pac. J. Math.
80 (1979), 265-272.
[9J S. J. Sidney and E. L. Stout, A note on interpolation, Proc. Amer. Math. Soc. 19 (1968),
380-382.
[10J J. Wermer, The space of real parts of a function algebra, Pac. J. Math. 13 (1963), 1423-1426.

SCIENCE INSTITUTE, UNIVERSITY OF ICELAND. REYKJAVIK, ICELAND


E-mail address:briemillhi.is
Contemporary Mathematics
Volume 328, 2003

Some mapping properties of p-summing operators with


Hilbertian domain

Qingying Bu

ABSTRACT. We prove that if H is a Hilbert space, Y is a Banach space and


u : H ---+ Y is absolutely p-summing for some p :::: 1, then for any 1 < q < 00,
u takes absolutely q-summable sequences in H into members of iqY, the
projective tensor product of lq and Y.
Given a real or complex Banach space X and 1 ::; p < 00, we denote by
f~trong(x) and f;eak(X) the Banach spaces of sequences in X with norms
II(xn)nlle;trong(X) = 11(llxnll)nllep and II(xn)nlle;eak(X) = sUP",*eB x * II(x*xn)nllep,
respectively (cf. [4, pp. 32-36]). For 1 < p < 00, let fp(X) denote the space of
all (strongly p-summable) sequences in X such that E~=l Ix~(xn)1 < 00 for each
(X~)n E f't},eak(x*), normed by

II(xn)nlltp(X) = sup {I ~ x~(xn)1 : II(x~)nlle~~eak(X.) ::; 1} ,


where pI is the conjugate of p, i.e., lip + 11pl = 1. With this norm fp(X) is a
Banach space (cf. [1, 3]).
Note: In [2] it was shown that fp(X) is exactly fpX, the projective tensor
product of fp and X.
In this note we use this identification of fp(X) with fpX to deduce a surprising
mapping property of absolutely p-summing operators that have a Hilbert space do-
main. While the main result of this note can be derived from some by-now famous
results of Kwapien, it was discovered because of the identification of fp(X) with
fpX, moreover, this identification leads itself to a proof that is a clean and clear
application of Khinchin's inequality, Kahane's inequality, and Pietsch's Domination
theorem - all fundamental aspects of the theory of p-summing operators.
From the definitions, we have for 1 < p < 00, fp(X) ~ f;trong(x) ~ f;eak(x),
and 11lIe~..ak(X) ::; 1I'lIl~trong(X) ::; 11llep(x) Moreover, in case dimX = 00, all the
containments are proper.
For Banach spaces X and Y and a continuous linear operator 'U : X ---+ Y,
define ft. : XN ---+ yN by (xn)n 1---+ (uxn)n. Then ft. is a linear operator. Thanks

2000 Mathematics Subject Classification. 46B28.

2003 American Mathematical Society


145
146 QINGYING BU

to the Closed Graph Theorem, each of


it: f;eak(x) ----> f;eak(y); it: f~trong(x) ----> f~trong(y); it: fp(X) ----> fp(Y)
is a continuous operator with
Ilitll;:,eak(x)_e;:,eak(Y) = Ilitll~trong(x)_e~trong(y) = Ilitllep(x)-p(Y) = Ilull
We should mention here Khinchin's inequality (cf. [4, p. 10]) and Kahane's
inequality (cf. [4, p. 211]) each of which plays a critical role in this paper. Let rn(t)
denote the Rademacher functions (cf. [4, p. 10]), namely, rn : [O,IJ ----> ~, n E Pi!
defined by rn(t) := sign(sin2n7rt).

Khinchin's Inequality. For any 0 < p < 00, there are positive constants Ap, Bp
such that for any scalars a1, a2, ... ,an, we have

Kahane's Inequality. If 0 < p, q < 00, then there is a constant Kp,q > 0 for
which

([II t, r,('lx, I ,d') 1/, ,; K",' ([II t, r,('lx,lI' dt ) 1/,

regardless of the choice of a Banach space X and of finitely m.any vectors Xl, X2, ... , Xn
from.X.

In 1967, A. Pietsch [5J introduced p-summing operators between Banach spaces,


namely, a Banach space operator u : X ----> Y is called p-summing operator, 1 :::;
p < 00, if it takes f~eak(x) into f~trong(y). Let IIp(X, Y) denote the space of all p-
summing operators from a Banach space X to a Banach space Y. If u : X ----> Y is a
p-summing operator then, thanks to the Closed Graph Theorem, it : f~eak(x) ---->
f~trong(y) is continuous. So we define the p-summing norm 7rp(u) on IIp (X, Y) to
be 7rp (u) = Ilitll;:,eak(x)_~trong(y). With this norm IIp (X, Y) is a Banach space.
There is an equivalent definition of p-summing operators, namely, a Banach
space operator u : X ----> Y is p-summing if and only if there is a constant c > 0
such that for any XI,X2,'" ,Xn E X,

(1)

In this case,
7rp (u) = inf{c > 0: for all possible c in (I)}.
In 1973, J. S. Cohen [3J introduced strongly p-summing operators between
Banach spaces, namely, a Banach space operator u : X ----> Y is called strongly
p-summing operator, 1 < p < 00, if it takes f~trong(x) into fp(Y). Let Dp(X, Y)
denote the space of all strongly p-summing operators from a Banach space X to a
Banach space Y. If u : X ----> Y is a strongly p-summing operator then, thanks to
the Closed Graph Theorem, it : f~trong(x) ----> fp(Y) is continuous. So we define a
PROPERTIES OF p-SUMMING OPERATORS 147

strongly p-summing norm Dp(u) on Dp(X, Y) to be Dp(u) = lIulll~trong(X)-+lp(Y)'


With this norm Dp(X, Y) is a Banach space.
There is an equivalent definition of strongly p-summing operators, namely, a
Banach space operator U ': X ~ Y is strongly p-summing if and only if there is a
constant c > 0 such that for any Xll X2,'" , Xn E X, and any yi, Y2,'" , y~ E Y*,

(2)

In this case,
Dp(u) = inf{c > 0: for all possible c in (2)}.

Note: Actually, U E Dp(X, Y) means that u takes absolutely p-summable


sequences in X into members of ipY.

Main Theorem. Let 1 < p, q < 00; and let H be a Hilbert space and Y be a
Banach space. Then IIp(H, Y) ~ Dq(H, Y), i.e., if u : H ~ Y is absolutely
p-summing, then u takes absolutely q-summable sequences in H into members of
lqY.

PROOF. First consider H = i2 for n E N. Let u E IIp(l2' Y). By Pietsch's


Domination Theorem [4, p. 44], there is a regular probability measure J.L on Bl'.J:
such that for any x E i 2,

lip
( )
Iluxll ::; 1Tp (U)' Lt'.J: I(x, z)IP dJ.L(z) (3)

Now for Xl,X2,'" ,Xm Ei 2, and Yi'Y2"" ,Y;" E Y*, we have


n
Xk = ~Xk e
L.-J ,'&"', k = 1,2, ,m.
i=l

Then
m m n
L I(UXk, yZ)1 L I(Lxk,iuei, yZ)1
k=l i=1
k=l

< t. (t, lx,.,1 (t,1 2) ,/, , (He;, y;)I' ) II'

< t.IlX,II' 1" ' (1' 1t, r,(t)(ue" Yk) I" II"~
dt)
148 QINGYING BU

by Khinchin's inequality,

< L (t." (t.[1 t,


x."') II, . I" <It)
r,(t)(ue, , yk) 11,

L.II 1Ir,.... t.1 (t, Yk)I" 'I,


(x.);" (X) ( [ ",(t)""" dt)

< 1,.' II ",;(X) . ( [ (II t, II . II (Y;);"I1'~""(x.)) " <It) ""


(x,);" r,(t)"e,

~ L II (x.);"II,;'~(X) . II (Y;);"II,~,..,(X") . ([II t, I " 'I,


r,(t)ue, dt)

(4)
Now by Kahane's inequality and (3),

([II t, r,(t)"" II" dt) 'I,' "K,.,. ([II t, r,(t)ue,II' <It) ",
< (L'r I(t,
K,.,.~,(u) . ( [ <It) 'I,
r,(t)e;, z)I' dp(z) )

K,.,.~,( (L I(t,
q ([ I' r,(t)e;, z) dt) dP(Z)),,'

r
u) .

< K,.,. ~,( (L,. (B,. (t, I(e" I'r}


u) . z) dp(z)

~ K,.,. ~,(u) . (L" IIzlI'


B, . "K,.,. ~,(u).
dp(z) )'/' B, .

(5)
Combining (4) and (5),

So U E Dq(H, Y) with

1
Dq(u) :::; -A . Bp Kp,ql 7rp(u).
ql

Now consider a general Hilbert space H. Let Xl, X2, , Xm E H and


Yi,y2, ,y:n, E Y*. Then there is an n E N such that span{xk}r is isometrically
isomorphic to i 2. Let P be the orthogonal projection from H onto span{xk}r.
PROPERTIES OF p-SUMMING OPERATORS 149

Then by (6),
m m
L I(UPXk, Yk)1
k=1 k=1

< :
ql
. Bp' Kp,ql' ll"p(U) . II (PXk)i"lI egtron 9 (X) 11(Yk)i"lle~,.ak(X.)
:. Bp' Kp,ql' ll"p(U) 11(Xk)i"lIe~tr.on9(X) 11(Yk)i"lle~,eak(X.).
ql
This shows us that U E Dq(H, Y) and again
1
Dq(u) :::; A . Bp' Kp,ql' ll"p(u).
ql
The proof is complete. o
ACKNOWLEDGEMENT. I am grateful to my advisor, Professor Joe Diestel, for
his good suggestions for this paper.

References
1. H. Apiola, Duality between spaces of p-summable sequences, (p, q)-summing operators and
characterization of nuclearity, Math. Ann. 219 (1976), 53-64.
2. Q. Bu and J. Diestel, Observations about the projective tensor product of Banach spaces, I
- ipX, 1 < p < 00, Quaestiones Math. 24 (2001),519-533.
3. J. S. Cohen, Absolutely p-summing, p-nuclear operators, and their conjugates, Math. Ann.
201 (1973), 177-200.
4. J. Diestel, H. Jarchow, and A. Tonge, Absolutely Summing Operators, Cambridge Univ. Press,
Cambridge, 1995.
5. A. Pietsch, Absolut p-summierende abbildungen in normierten riiumen, Studia Math. 28
(1967), 333-353.

DEPARTMENT OF MATHEMATICS, UNIVERSITY OF MISSISSIPPI, UNIVERSITY, MS 38677


E-mail address: qbu(llolemiss. edu
Contemporary Ma.thema.tics
Volume 328, 2003

The Unique Decomposition Property and


the Banach-Stone Theorem

Audrey Curnock, John Howroyd, and Ngai-Ching Wong*

ABSTRACT. We show an affine version of the Banach-Stone theorem. Given


compact convex sets K and S with unique decomposition property, we show
that every surjective linear isometry T between the affine function spaces A(K)
and A(S) induces an affine homeomorphism between K and S. Furthermore,
T can be written as a weighted composition operator in this case.

1. Introduction
The celebrated Banach-Stone Theorem states that two compact Hausdorff spa-
ces X and Y are homeomorphic if and only if the corresponding real continu-
ous function spaces C(X) and C(Y) are linearly isometric (see, for example, [3,
Chapter 7]). As is well known, see for example [2, Theorem 1.4.9], C(X) can
be identified with the Banach space A(K) of real continuous affine functions on
K = {'P E C(X)* : II'PII = 1 = 'P(l)}, the state space of C(X), where the norm in
A(K) is the usual supremum norm. Here K is a Bauer simplex and consequently
there is a natural reformulation of the Banach-Stone theorem in the context of
affine geometry as follows. Two Bauer simplexes K and S are affinely homeomor-
phic if and only if their corresponding affine function spaces A(K) and A(S) are
linearly isometric.
Clearly an affine homeomorphism between any two compact convex sets K
and S induces a linear isometry between A(K) and A(S). However, an example
of J.T. Chan given in [9] shows that the converse of this cannot hold for arbitrary
compact convex sets in locally convex (Hausdorff) spaces, even in finite dimensions.
Thus it is natural to ask what conditions on K and S imply that this converse does
hold.
Lazar [11] showed that the converse holds if both K and S are Choquet sim-
plexes. Ellis and So [9] extended this to the case when both K and S have the
property that every pair of complementary closed faces is split, which applies to
the state spaces of function algebras.
In this paper we use another geometric condition on K and S, namely the
unique decomposition property of Ellis [6, 7, 8], under which K and S are affinely
homeomorphic whenever there is a linear isometry T between A(K) and A(S).

2000 Mathematics Subject Classification. Primary 46A55; Secondary 46B04.


Key words and phrases. Affine function space, isometry, unique decomposition property.
*Partially supported by Taiwan National Science Council Grants: 89-2115-Mll0-009,37128F.
151
152 AUDREY CURNOCK, JOHN HOWROYD, AND NGAI-CHING WONG

Moreover, we show that in this case T can be written as a weighted composition


operator. This includes Lazar's result as a special case and, more generally, the
state spaces of C*-algebras.
Further results by the same authors relating to skew-symmetries are to be found
in [4].
We would like to express our deepest gratitude to Professor Cho-Ho Chu for
many useful discussions.

2. Preliminaries
We note that every compact convex set K (in a locally convex space) can be
embedded into A(K)* as the state space {cp E A(K)* : Ilcpll = 1 = cp(l)} of A(K)
with the weak* topology, where x E K is identified with the linear functional
f 1-+ (j, x) = f(x)
for all f in A(K). If Y is a subset of K then we denote the convex hull of Y by
co (Y). A point x in K is called an extreme point of K if K \ {x} is a convex set.
The set of extreme points of K is denoted by 8K. In the above setting the closed
unit ball BA(K)' of A(K)* is the convex hull of K and -K; namely co (K U -K)
where -K denotes {-k: k E K}. Thus 8B A(K)' is contained in 8KU8(-K). The
reverse inclusion follows easily from the fact that for cp E BA(K)" \ (K U -K) we
have -1 < cp(l) < 1. Therefore, 8B A(K)" is equal to 8KU8(-K).
A convex subset F of K is called a face of K if whenever x E F with x =
>..y + (1 - >..)z for some y, z E K and >.. E (0,1), then both y and z are in F.
A pair of faces (FI, F2) is said to be complementary whenever FI n F2 = 0 and
K = co (FI U F 2 ). Thus, in this case, each x in K has a decomposition relative
to (FI' F2); namely x = >..y + (1 - >..)z for some y E F I , Z E F2 and>" E [0,1]. If
>.. is unique, for each x E K \ (FI U F 2 ) but independent of y and z, then FI and
F2 are said to be parallel faces; parallel faces are automatically norm closed. If in
addition y and z are unique then FI and F2 are called split faces. See, for example,
[1, 2] for the general theory of compact convex sets and related topics. Let K be
a compact convex set.
Recall that the facial topology on 8K is given by defining
{F n 8K : F is a closed split face of K}
to be the family of all closed sets. The facial topology is weaker than the relative
topology on 8K. The centre Z(A(K)) of A(K) is the set of all those functions
in A(K) whose restriction to 8K is facially continuous. The central functions
h E Z(A(K)) are characterised by the following property (see, for example, [1,
Corollary 11.7.4] or [2, Theorem 3.1.4]): for all f E A(K), there exists 9 E A(K)
such that g(x) = h(x)f(x) for all x in 8K. The uniqueness of the continuous affine
function 9 is clear, since a continuous affine function on a compact convex set is
completely determined by its values on the extreme boundary, and consequently
we may write 9 = h . f. In this way it is useful to think of the central functions as
the multipliers of A(K).
A compact convex set K is a Choquet simplex whenever for all bounded (real)
linear functionals cp in A(K)* and 0 > 0 the set K n (cp + oK) is either empty or
of the form 'IjJ + (3K for some ' in A(K)* and {3 ~ 0 (see, for example, [12]); note
that {3 = 0 allows K n (cp + oK) to be a singleton. In a Choquet simplex every
closed face is split (see [2, Theorem 2.7.2]).
THE UNIQUE DECOMPOSITION PROPERTY AND BANACH-STONE THEOREM 153

In [6, 7, 8], Ellis defined the unique decomposition property of K by the condi-
tion that for every <P in A(K)*, there is a unique decomposition <P = <PI - <P2 with
<PI, <P2 ~ 0 and Ii<pll = Ii <PI Ii + 1i<P21i. In [6] he showed that K has the unique decom-
position property if and only iffor every <P in A(K)* and 0> 0 the set Kn(<p+oK)
is either empty or a singleton, or contains some 1/1+{3K where 1/1 E A(K)* and (3 > O.
In particular, every Choquet simplex has the unique decomposition property.
A result of Grothendieck [10], see also [5, p. 272], shows that the state space of
a (unital) C*-algebra has the unique decomposition property. Thus the results of
this paper apply, giving an affine homeomorphism between the state spaces K and
S of two C*-algebras whenever the associated (real) affine function spaces A(K)
and A(S) are linearly isometric.
We give some examples below, the second and third of which show that the
unique decomposition property and the condition of Ellis and So are independent
geometric properties of a compact convex set.
EXAMPLE 2.1. Let K be the state space of the C*-algebra M 2 , of all 2 x 2
matrices over C. Then, by Grothendieck's result, K has the unique decomposition
property. Also, K is affinely homeomorphic to a closed ball in 1R3 (see [2, p. 241])
and satisfies the condition of Ellis and So since it has no proper complementary
faces.
EXAMPLE 2.2. Let K be a triangular bi-simplex in 3-dimensional space as in the
figure below, Figure 1. Then no proper face is (geometrically) parallel to any other
and hence K n (<p + oK) is either empty, a singleton or has non-empty interior, for
<P E A(K)* and 0 > o. It follows, by the geometric characterisation of Ellis, that K
has the unique decomposition property. However (F, F') is a pair of complementary
faces of K, but not split, and hence K does not satisfy the condition of Ellis and
So.

EXAMPLE 2.3. Let K be an icosahedron in 3-dimensional space. Then K


satisfies the condition of Ellis and So because it has no proper complementary
faces. However it does not have t.he unique decomposition property because it has

FIGURE 1. The bi-simplex of Example 2.2


154 AUDREY CURNOCK, JOHN HOWROYD, AND NGAI-CHING WONG

faces (geometrically) parallel to each other and hence K n (cp + aK) can have empty
interior for some cp E A(K)* and a > O.

3. Results
Throughout this section K and 8 will denote compact convex sets of locally
convex (Hausdorff) spaces. We will also let T: A(K) ~ A(8) denote a surjective
linear isometry. Notice that if T1 = 1 then for cp E A(8)* with Ilcpli = 1 = cp(1), we
have
IIT*cpll = licp 0 Til = licpli = 1 = cp(1) = cp(T1) = T*cp(1);
here T*: A(8)* ~ A(K)* denotes the dual map of T. Consequently T*(8) = K
and hence T* induces an affine homeomorphism a: 8 ~ K such that Tf(s) =
T*s(f) = f(a(s)) for all s E 8; that is, T is a composition operator f f-+ f 0 a. In
Proposition 3.3 we see that T is a weighted composition operator if and only if T1
is central. To do this we 'decompose' 8 by defining
(3.1) 81 = {s E 8: (Tl)(s) = I} and 82 = {s E 8 : (T1)(s) = -I}.
It is clear that 8 1 and 8 2 are closed faces of 8.
LEMMA 3.1. Let 8 1 and 8 2 be as in (3.1). Then 88 ~ 8 1 U 8 2, and 8 1 and 8 2
are closed parallel faces of 8.
PROOF. Observe that the dual map T* is a linear isometry from A(8)* onto
A(K)*. Hence T* maps the extreme points of the closed unit ball of A(8)* onto
the extreme points of the closed unit ball of A(K)*. Thus
T*(88 U 8( -8)) = 8K U 8( -K).
Consequently, for each s E 88 we have T* s is in 8K or 8( - K) and hence
Tl(s) = T*s(l) = 1.
Therefore 88 ~ 8 1 U 8 2 and thus by the Krein-Milman Theorem 8 = co (81 U 8 2),
since 8, 8 1 and 82 are all (weak*) compact. Thus 8 1 and 8 2 are complementary
faces since they are clearly disjoint.
For each s in 8 with s = >.x + (1 - >')y, where x E 8 1 , Y E 8 2, and>' E (0,1),
we have T* s = >'T*x + (1 - >')T*y. Thus,
Tl(s) = T*s(1) = >'T*x(1) + (1- >')T*y(1) = >. - (1- >.) = 2>' - 1.
This establishes the uniqueness of>. = ((Tl)(s) + 1)/2 and the result follows. 0
We now specialise to the case when K and 8 have the unique decomposition
property.
LEMMA 3.2. Let 8 1 and 8 2 be as in (3.1). 8uppose that K has the unique
decomposition property. Then 8 1 and 8 2 are complementary split faces of 8.
PROOF. By Lemma 3.1, for each s E 8\(81 U 8 2) we may write s = >.x +
(1 - >.)y where x E 8 1 , Y E 8 2 and 0 < >. < 1, and>' is unique. We consider the
decomposition T* s = >'T*x - (>. -1)T*y. Since x E 8 1 we have T*x E K and hence
>'T*x is positive. Similarly, T*y E -K and hence (>. - I)T*y is positive. Also
IIT*sll = 1 = >. + (1- >.) = Ii>'T*xli + 11(>' -
I)T*yli
Thus, by the unique decomposition property, >'T*x and (>. -1)T*y are unique. By
the uniqueness of >., we have T*x and T*y are unique. Since T is surjective, T* is
injective and the result follows. 0
THE UNIQUE DECOMPOSITION PROPERTY AND BANACH-STONE THEOREM 155

By replacing T by T- 1 in (3.1) we may 'decompose' K by defining


(3.2) K1 = {k E K : T- 11(k) = I} and K2 = {k E K : T- 11(k) = -I}.
Applying Lemmas 3.1 and 3.2 to T- 1 we see that K1 and K2 are complementary
split faces of K whenever S has the unique decomposition property.
We say that T is a weighted composition operator whenever there exists a
central function h in A(S) and a continuous affine mapping (1: S -+ K such that
Tf = h f 0 (1 for all f E A(K); that is, Tf(s) = h(s)f((1(s)) for all sEaS. The
following proposition asserts that the linear isometry T is a weighted composition
operator, with Tl(s) = 1 on as, if and only if Tl is central.
PROPOSITION 3.3. Let T: A(K) -+ A(S) be a linear mapping. Then the fol-
lowing are equivalent:
a) T is an isometry and Tl is central;
b) T is a weighted composition operator of the form T f = h . f 0 (1 for all
f E A(K) where (1 is an affine homeomorphism and h(s) = 1 for all
sEas.
PROOF. See [4, Theorem 3.3] or [13]. o
We now apply the above decompositions of K and S to prove our main theorem.
THEOREM 3.4. Suppose that K and S have the unique decomposition property.
Then the real affine function spaces A(K) and A(S) are linearly isometric if and
only if K and S are affinely homeomorphic. Moreover, every linear isometry from
A(K) onto A(S) may be written as a weighted composition operator.
PROOF. It suffices to show necessity. Suppose that T is a linear isometry from
A(K) onto A(S). We 'decompose'S into the complementary split faces Sl and S2
of (3.1) and, similarly, K into the complementary split faces K1 and K2 of (3.2).
Since (T-1)* = (T*)-l, we have T*(Sd = K1 and T*(S2) = -K2' and hence we
may define (1: S -+ K by
(1(,xx + (1 - ,x)y) = ,xT*(x) - (1 - ,x)T*(y)
whenever x E Sl, Y E S2 and 0 :5 ,x :5 1. We see that (1 is an affine homeomorphism
from S = co (Sl U S2) onto K = co (K1 U K 2).
Moreover, to show that T is a weighted composition operator it suffices, by
Proposition 3.3, to show that h = Tl is central. Let f E A(S) then, since (1: S -+ K
is an affine homeomorphism, we may write f = gO(1 for some g E A(K). Note that
for x E Sl we have Tg(x) = T*x(g) = g((1(x)) = h(x)f(x). Similarly for x E S2 we
have Tg(x) = T*x(g) = -g((1(x)) = h(x)f(x). Therefore, for all x E as ;; Sl U S2
we have Tg(x) = h(x)f(x), and the result follows. 0
References
[1] E.M. Alfsen, Compact convex sets and boundary Integmls, Ergebnisse der Mathematik, 57,
(Springer-Verlag, Berlin-Heidelberg-New York, 1971).
[21 L. Asimow and A.J. Ellis, Convexity theory and its applications in functional analysis, Lon-
don Math. Soc. Monograph, 16 (Academic Press, London, 1980).
[3] E. Behrends, M -Structure and the Banach-Stone Theorem, Lecture Notes in Mathematics
736, (Springer-Verlag, Berlin-Heidelberg-New York, 1979).
[4] A. Curnock, J. Howroyd, and N.-C. Wong, Isometries of affine function spaces, preprint.
[5] J. Dixmier, C*-Algebms (North-Holland Publishing Co., Amsterdem-New York-Oxford,
1982).
156 AUDREY CURNOCK, JOHN HOWROYD, AND NGAI-CHING WONG

[6] A.J. Ellis, An intersection property for state spaces, J. London Math. Soc., 43 (1968), 173-
176.
[7] A.J. Ellis; Minimal decompositions in partially ordered normed spaces, Proc. Camb. Phil.
Soc., 64 (1968),989-1000.
[8] A.J. Ellis, On partial orderings of normed spaces, Math. Scand., 23 (1968), 123-132.
[9] A.J. Ellis and W.S. So, Isometries and the complex state spaces of uniform algebras, Math.
Z., 195 (1987), 119-125.
[10] A. Grothendieck, Un result at sur Ie dual d'une C*-algebre, J. Math. Pures Appl., 36 (1957),
97-108.
[11] A.J. Lazar, Affine products of simplexes, Math. Scand., 22 (1968), 165-175.
[12] R.R. Phelps, Lectures on Choquet's Theorem, Second Edition, Lecture notes in Mathematics
1757 (Springer-Verlag, Berlin, 2001).
[13] T.S.R.K. Rao, Isometries of Ac(K), Proc. Amer. Math. Soc., 85 (1982), 544-546.

SCHOOL OF COMPUTING, INFORMATION SYSTEMS AND MATHEMATICS, SOUTH BANK UNIVER-


SITY, LONDON SE1 OAA, ENGLAND.
E-mail address:curnocaOsbu.ac.uk
DEPARTMENT OF MATHEMATICAL SCIENCES, GOLDSMITHS COLLEGE, UNIVERSITY OF LONDON,
LONDON SE14 6NW, ENGLAND.
E-mail address:masOljdhOgold.ac.uk
DEPARTMENT OF ApPLIED MATHEMATICS, NATIONAL SUN YAT-SEN UNIVERSITY, KAOHSJUNG
80424, TAIWAN, R.O.C.
E-mail address:wong~ath.nsysu.edu.tw
Contemporary Mathematics
Volume 328, 2003

A Survey of Algebraic Extensions of


Commutative, Unital Normed Algebras

Thomas Dawson

ABSTRACT. We describe the role of algebraic extensions in the theory of com-


mutative, unital normed algebras, with special attention to uniform algebras.
We shall also compare these constructions and show how they are related to
each other.

Introduction
Algebraic extensions have had striking applications in the theory of uniform
algebras ever since Cole used them (in [5]) to construct a counterexample to the
peak-point conjecture. Apart from this, their main use has been in (a) the con-
struction of examples of general, normed algebras with special properties and (b)
the Galois theory of Banach algebras. We shall not discuss (b) here; a summary of
some of this work is included in [29].
In the first section of this article we shall introduce the types of extensions and
relate their applications. The section ends by giving the exact relationship between
the types of extensions. Section 2 contains a table summarising what is known
about the extensions' properties.
A theme lying behind all the work to be discussed is the following question:
(Q) Suppose the normed algebra B is related to a subalgebra A by some specific
property or construction. (For example, B might be integral over A: every
element b E B satisfies ao + ... +an_1bn-1 +bn = 0 for some ao, ... ,an -1 EA.)
What properties of A (for example, completeness or semisimplicity) must be
shared by B?
This is a natural question, and interesting in its own right. Many special cases of
it have been studied in the literature. We shall review the related body of work in
which B is constructed from A by adjoining roots of monic polynomial equations.
Throughout this article, A denotes a commutative, unital normed algebra, and
A its completion. The fundamental construction of [1] applies to this class of
1991 Mathematics Subject Classification. Primary 46J05, 46J10.

This research was supported by the EPSRC.


2003 American Mathematical Society

157
158 THOMAS DAWSON

algebras. Algebraic extensions of more general types of topological algebras have


received limited attention in the literature (see [19], [21]).
If E is a subset of a ring then (E) will stand for the the ideal generated by E.

1. Types of Algebraic Extensions and their Applications


1.1. Arens-Hoffman Extensions. Let a(x) = ao + ... + an_lx n- l + xn be
a monic polynomial over the algebra A. The basic construction arising from A and
a(x) is the Arens-Hoffman extension, Ao. This was introduced in [1]. Most of the
obvious questions of the type (Q) for Arens-Hoffman extensions were dealt with in
this paper and in the subsequent work of Lindberg ([18]' [20], [13]). See columns
two and three of Table 2.2.
All the constructions we shall meet are built out of Arens-Hoffman extensions.
DEFINITION 1.1.1. A mapping 0: A - B between algebras A and B is called
unital if it sends the identity of A to the identity of B. An extension of A is a commu-
tative, unital normed algebra, B, together with a unital, isometric monomorphism
O:A-B.
The Arens-Hoffman extension of A with respect to a(x) is the algebra Ao :=
A[x]/(a(x)) under a certain norm; the embedding is given by the map v: a t-+
(a(x)) + a.
To simplify notation, we shall let x denote the coset of x and often omit the
indeterminate when using a polynomial as an index.
It is a purely algebraic fact that each element of Ao has a unique representative
of degree less than n, the degree of a(x). Arens and Hoffman proved that, provided
the positive number t satisfies the inequality t n ~ ~~:~ lIakll tk, then

I~ bkX
k
= ~ IIbkll t k (bo, ... , bn - l E A)

defines an algebra norm on Ao.


The first proposition shows that Arens-Hoffman extensions satisfy a certain
universal property which is very useful when investigating algebraic extensions. It
is not specially stated anywhere in the literature; it seems to be taken as obvious.
PROPOSITION 1.1.2. Let A(l) be a normed algebra and let 0: A(l) _ B(2) be a
unital homomorphism of normed algebras. Let al (x) = ao + ... + an_lXn - l + xn E
A(l) [x] and B(1) = A~l}. Let y E B(2) be a root of the polynomial a2(x) :=
O(ad(x) := O(ao) + ... +O(an_l)X n - l +xn. Then there is a unique homomorphism
1>: B(l) _ B(2) such that
B(1) ~ B(2)
r II

A(l)
/9 is commutative and 1>(x) = y.

The map 1> is continuous if and only if 0 is continuous.


PROOF. This is elementary; see [7] o
ALGEBRAIC EXTENSIONS OF COMMUTATIVE NORMED ALGEBRAS 159

1.2. Incomplete Normed Algebras. A minor source of applications of


Arens-Hoffman extensions fits in nicely with our thematic question (Q): these ex-
tensions are useful in constructing examples to show that taking the completion of
A need not preserve certain properties of A.
The method uses the fact that the actions of forming completions and Arens-
Hoffman extensions commute in a natural sense. A special case of this is stated
in [17]; the general case is proved in [7), Theorem 3.13, and follows easily from
Proposition 1.1.2.
It is convenient to introduce some more notation and terminology here.
Let O(A) denote the space of continuous epimorphisms A --+ Cj when n appears
on its own it will refer to A. As discussed in [1], this space, with the weak *-topology
relative to the topological dual of A, generalises the notion of the maximal ideal
space of a Banach algebra. In fact, it is easy to check that 0 is homeomorphic to
0(..4), the maximal ideal space of the completion of A.
The Gelfand transform of an element a E A is defined by
a: n --+ C; W 1--+ w(a)
and the map sending a to a is a homomorphism, r, of A into the algebra, C(O), of
all continuous, complex-valued functions on the compact, Hausdorff space O. We
denote the image of r by A. A good reference for Gelfand theory is Chapter three
of [24].
DEFINITION 1. 2.1 ([1)). The algebra A is called topologically semisimple if r
is injective.
If A is a Banach algebra then this condition is equivalent to the usual notion of
semisimplicity. The precise conditions under which Aa is topologically semisimple
if A is are determined in [1]. In [17] Lindberg shows that the completion of a
topologically semisimple algebra need not be semisimple.
In order to illustrate Lindberg's strategy we recall two standard properties of
normed algebras.
DEFINITION 1.2.2. The normed algebra A is called regular if for each closed
subset E ~ 0 and wE O-E there exists a E A such that a(E) ~ {O} and a(w) = 1.
The algebra is called local if A contains every complex function, f, on 0 such that
every wE 0 has a neighbourhood, V, and an element a E A such that flv = alv.
It is a standard fact that regularity is stronger than localness; see Lemma 7.2.8
of [24).
EXAMPLE 1.2.3. Let A be the algebra of all continuous, piecewise polyno-
mial functions on the unit interval, I, and a(x) = x 2 - id/ E A[x]. Let A have the
supremum norm. By the Stone-Weierstrass theorem, A = C(I) and hence n is iden-
tifiable with I. Clearly A is regular. We leave it as an exercise for the reader to find
examples to show that Aa is not local. This is not hardj it may be helpful to know
that in this example the space O(Aa) is homeomorphic to {(s, oX) E I xC: oX 2 = s}.
This follows from facts in [1].
In the present example, neither localness nor regularity is preserved by (incom-
plete) Arens-Hoffman extensions.
160 THOMAS DAWSON

Finally we can explain the method for showing that some properties of normed
algebras are not shared by their completions because, in the above, 'non-regularity'
is not preserved by completion of Ao (nor is 'non-localness'). To see this, note that
Ais clearly regular if A is and so by a theorem of Lindberg (see Table 2.2) the Arens-
Hoffman extension (A)o is regular. But, by a result of [17] referred to above, this
algebra is isometrically isomorphic to the completion of Aa. Of course Lindberg's
original application was much more significant; there are simpler examples of the
present result: for example the algebra of polynomials on I.
1.3. Uniform Algebras. It is curious that the application of Arens-Hoffman
extensions to the construction of integrally closed extensions of normed algebras
did not appear in the literature for some time after [1]. It was seventeen years later
until a construction was given in [22]. Even then the author acknowledges that
the constuction was prompted by the work of Cole, [5], in the theory of uniform
algebras.
Cole invented a method of adjoining square roots of elements to uniform al-
gebras. He used it to extend uniform algebras to ones which contain square roots
for all of their elements. Apart from feeding back into the general theory of com-
mutative Banach algebras (mainly accomplished in [22] and [23]) his construction
provided important examples in the theory of uniform algebras. We shall describe
these after recalling some basic definitions.
DEFINITION 1.3.1. A uniform algebra, A, is a subalgebra of C(X) for some
compact, Hausdorff space X such that A is closed with respect to the supremum
norm, separates the points of X, and contains the constant functions. We speak
of 'the uniform algebra (A, X)'. The uniform algebra is natural if all of its homo-
morphisms wEn are given by evaluation at points of X, and it is called trivial if
A = C(X).
Introductions to uniform algebras can be found in [4], [11], [26], and [16]. An
important question in this area is which properties of (A, X) force A to be trivial.
For example it is sufficient that A be self-adjoint, by the Stone-Weierstrass theorem.
In [5] an example is given of a non-trivial uniform algebra, (B,X), which is natural
and such that every point of X is a 'peak-point'. It had previously been conjectured
that no such algebra existed.
We shall describe the use of Cole's construction in the next section, but now
we reveal some of the detail.
PROPOSITION 1.3.2 ([5],[7]). Let U be a set of monic polynomials over the
uniform algebra (A, X). There exists a uniform algebra (AU, XU) and a continuous,
open surjection 7r: XU ---> X such that
(i) the adjoint map 7r*: C(X) ---> C(XU) induces an isometric, unital monomor-
phism A ---> AU, and
(ii) for every a E U the polynomial 7r*(a)(x) E AU[x] has a root Po E AU.
PROOF. We let XU be the subset of X x cU consisting of the elements (K, >.)
such that for all a E U
f(a)(K)
JO
+ ... + /(0)
n(a)-1
(K)>.n(0)-1 + >.n(o)
a 0
= 0
ALGEBRAIC EXTENSIONS OF COMMUTATIVE NORM ED ALGEBRAS 161

where o:(x) = f~Ol) + ... + f~(l)_lXn(Ol)-l + xn(Ol) E U. The reader can easily check
that XU is a compact, Hausdorff space in the relative product topology and so the
following functions are continuous:
11": XU -+ X; (~, >.) ...... ~
POI: XU -+ C; (~, >.) ...... >'01 (0: E U).

The extension AU is defined to be the closed subalgebra of C(XU) generated by


1I"*(A) U {POI: 0: E U} where 11"* is the adjoint map C(X) -+ C(XU) ; g ...... go 11".
It is not hard to check that AU is a uniform algebra on XU with the required
properties. 0

We shall call AU the Cole extension of A by U. Cole gave the construction for
the case in which every element of U is of the form x 2 - f for some f E A. It is
remarked in [22] that similar methods can be used for the general case; these were
independently, explicitly given in [7].
By repeating this construction, using transfinite induction, one can generate
uniform algebras which are integrally closed extensions of A. Full details of this,
including references and the required facts on ordinal numbers and direct limits of
normed algebras, can be found in [7]. Again this closely follows [5]. Informally the
construction is as follows.
Let v be a non-zero ordinal number. Set (Ao,Xo) = (A,X). For ordinal
numbers T with 0 < T ~ V we define
(A~" , X!:" ) if T = a +1 and
(A.,-, X T ) = { l~
. (A u,Xu)u<.,-, (*
1I"p,u, 1I"p,u ) p~U<T ) if T is a limit ordinal.

The construction requires sets of monic polynomials, Uu ~ Au [x], to be chosen in-


ductively. The notation (Au, Xu )u<.,-, (1I";,u, 1I"p,u )P~U<T) is not standard; it means
that Xu)u<.,-, (1I"p,U)p~U<T) is an inverse system of (non-empty) compact, Haus-
dorff spaces, while at the same time the induced direct system of uniform alge-
bras, (C(Xuu<.,-,(1I";,U)P~U<T)' 'restricts' in a natural way to the direct system
(Au)u<.,-, (1I";'U)p~U<T)'
The connection between Cole extensions and Arens-Hoffman extensions will be
elucidated in Section 1.6.
DEFINITION 1.3.3. Let (A, X) = (Ao, Xo) be a uniform algebra and v > 0 be
an ordinal number. Then (A.,-, X T)T<V' as above, is a system of Cole extensions of
(A, X). -
Thus (A1' Xl) is just a Cole extension of (Ao, Xo). When U1 is a singleton
we call Al a simple extension of Ao; the same adjective can be applied to Arens-
Hoffman extensions.
An integrally closed extension, (Av, Xv), is obtained by taking v to be the first
uncountable ordinal. At the successor ordinals the whole set of monic polynomials
is frequently used to extend the algebra, but this set is larger than necessary. The
same procedure is used to obtain the integrally closed extensions in other categories
(to be discussed in Section 1.6).
162 THOMAS DAWSON

1.4. Some Applications of Cole's Construction. Cole's method has been


developed by others, including Karahanjan and Feinstein, to produce examples of
non-trivial uniform algebras with interesting combinations of properties. We cite
the following example of Karahanjan.
THEOREM 1.4.1 (from [15], Theorem 4). There is a non-trivial, antisymmetric
uniform algebra, A, such that (1) A is integrally closed, (2) A is regular, (3) n is
hereditarily unicoherent, (4) G(A) is dense in A, and (5) the set of peak-points of
A is equal to n.
In the above, G(A) is our notation for the invertible group of A. We refer the
reader to [15] and the literature on uniform algebras for the definitions of other
terms we have not defined here.
A further example in [15] also strengthens Cole's original counter-example.
Both examples (of non-trivial, natural uniform algebras on compact, metriseable
spaces, every point of which is a 'peak-point') are regular. Feinstein has varied the
construction to obtain such an example which is not regular in [10].
The same author also used Cole extensions in [9] to answer a question of Wilken
by constructing a non-trivial, 'strongly regular', uniform algebra on a compact,
metriseable space.
Returning to the sample theorem quoted above, note that some of these prop-
erties (for example the topological property of 'hereditary unicoherence') are con-
sequences of the combination of other properties of the final algebra. By contrast,
(2) and (4) hold because they are true for the base algebra on which the example
is constructed. It is therefore very useful to know exactly when specific properties
of a uniform algebra are transferred to those in a system of Cole extensions of it.
The known results on this problem are summarised in the first column of Table
2.2.
Determining if an algebra's property is shared by its algebraic extensions has
led to some interesting devices. We shall elaborate on this topic in the next section.
We remark in passing that the methods used in [15] to show that the final algebra
has a dense invertible group have been simplified in [8]; in particular there is no
need to develop the theory of 'dense thin systems' in [15].
1.5. A Further Remark on Cole Extensions. The reader will notice
from Table 2.2 that virtually all properties of uniform algebras are preserved by
Cole extensions. The key to obtaining most of these results is the following result,
originating with Cole.
PROPOSITION 1.5.1 ([5]'[23]). Let (Ar,Xrk:;v be a system of Cole exten-
sions of (A, X). There exists a family of unital contractions (T.,.,r: C(Xr) -+
C(X.,.)).,.::;r::;v such that for all a ~ T ~ v
(i) T.,.,r(A r ) ~ A.,., and
(ii) T.,.,r 07r;,r = idc(X u )'
PROOF. See [23]. o
For example, it is easy to see from the existence of T: C(Xu) -+ C(X) that the
Cole extension AU is non-trivial if A i= C(X).
ALGEBRAIC EXTENSIONS OF COMMUTATIVE NORMED ALGEBRAS 163

The operator T was constructed in [5] for extensions by square-roots. In the


case of a simple Cole extension, (A{o}, X{o}), there are at most two points y(x:)
in the fibre 7r- l (x:) for each x: E X and they correspond to the roots of the equation
x 2 - f(x:) = 0 where o(x) = x 2 - f. The operator is then defined by

(g E C(X{o}), x: EX).

For other sorts of monic polynomials it was not so obvious how to construct
T. The basic techniques appeared in [22] (see the proof of Theorem 3.5) for simple
extensions, and were further developed in the proof of Theorem 4 of [15], but it was
not until [23] that a comprehensive construction was given. We must also mention
the role of E. A. Gorin: he appears to have paved the way for [15] and [23].
1.6. Algebraic Extensions of Normed and Banach Algebras. As we
have seen, algebraic extensions have had striking applications in the theory of uni-
form algebras. They have long been used as auxiliary constructions in the general
theory of Banach algebras. Notable examples of this are in [14] and [25]; the latter
explicitly uses Arens-Hoffman extensions.
However algebraic extensions for Hormed algebras were apparently only studied
in their own right in order to generalise the work of Cole and Karahanjan. We now
turn to these generalisations.
The basic extension generalising Arens-Hoffman extensions is called a standard
normed extension. It is defined in the following theorem of Lindberg.
THEOREM 1.6.1 ([22]). Let A be a normed algebra and U a set of monic
polynomials over A. Let ~ be a well-ordering on U with least element 00' Then
there exists a normed algebra, B u , with a family of subalgebras, (Bo)oEU, such
that:
(i) for all 0, (3 E U, Bo ~ Bf3 if 0 ~ (3, and,
(ii) for all (3 E U, Bf3 is isometrically isomorphic to an Arens-HoffmBll extension of
B<f3 by (3(x) where

PROOF. See [22]. o


Lindberg shows how this leads to the construction of Banach algebras with
interesting combinations of properties, one of which is integral closedness.
Let the isometric isomorphism B<f3[x]/(3(x)) -+ Bf3 in (ii) above be denoted
by 1jJf3 (3 E U). We shall refer to 1jJf3(x) as the standard root of (3(x) in Bu. This
helps us when we show that standard extensions share a similar universal property
to the one described in Proposition 1.1.2. Again the following lemma has not
been explicitly given in the literature but is probably regarded as obvious by those
working in the field. We state the result in full as we shall make use of it later.
LEMMA 1.6.2. Let A (1) be a normed algebra and U a non-empty set of monic
polynomials over A(l). Let B(l) = AiJ) be a standard extension of A(l) Witll
respect to U and (): A(l) -+ B(2) be a unital homomorphism of normed algebras.
164 THOMAS DAWSON

Let ~Ot be the standard root of a E U, witll associated norm parameter tOt, and
suppose (T/Ot)OtEU ~ B(2) is such that I(a)(T/Ot) = 0 for all a E U. Tllen there is a
unique, unital homomorphism : B(l) --+ B(2) such that the following diagram is
commutative
B(1) ~ B(2)

r~
A(1)
/9

(Note added in proof: The map is continuous if and only if I is continuous and

L (n(a) -l)log+ (11~OtII) < +00


OtEU Ot
where log+ denotes the positive part of the logarithm, max(log,O).)
PROOF. A simple application of transfinite methods and Proposition 1.1.2. 0

Purely algebraic standard extensions are defined in [22] and the main content
of Lemma 1.6.2 is a statement about these.
Narmania gives ([23]) an alternative construction for integrally closed exten-
sions of a commutative, unital Banach algebra, A. His method is rather more
conventional than the one used to define standard extensions. If U is a set of
monic polynomials over A then the Narrnania extension of A by U is equal to the
Banach-algebra direct limit of (As: 8 is a finite subset of U) where each As is
isometrically isomorphic to A extended finitely many times by the Arens-Hoffman
construction. As this paper is not readily available in English and we shall refer
to the explicit construction of Narmania's extensions in the next result, we stop to
report the precise details of this.
If E is a set, the set of all finite subsets of E will be written E<wo. Let
8 = {ai, ... , am} ~ U and let tOt (a E U) be a valid choice of Arens-Hoffman norm-
parameters (see Section 1.1). It is important to insist that distinct elements a, /3 E U
are associated with distinct indeterminates XOt , x{3. Thus 8 is an abbreviation for
{a1(xoJ, ... ,am(xo,J}.
It is proved carefully in [23] that for q = Ls qsx~ll ... x~;;, E A[x U1 ' ,xo: m ],
the algebra of polynomials in m commuting indeterminates over A (s is a mult.iindex
in No where No = {O}UN), then (8) +q has a unique representative whose degree in
3:O: j is less than than n(aj), the degree of aj(:ro: j ) (j = 1, ... ,m). For convenience
we shall call such representatives minimal. Then if q is the minimal representative of
(8) +q, 11(8) + qll := Ls IIqsll t~ll t~: defines an algebra norm on As. The index
set, U<wo is a directed set, directed by ~. The connecting homomorphisms VS.T (for
8 ~ T E U<wo) are the natural maps; they are isometries. Thus (sec [24] Section
1.3) Au is the completion of the normed direct limit, D := UsEU<wo As / "', where
'" is an equivalence relation given by (8) + q '" (T) + r if and only if q - r E (8 UT)
for 8, T E U<wo. Furthermore, the canonical map, Vs, which sends an element of
As to its equivalence class in D, is an isometry. Note that A0 is defined to be A.
We can now show how the types of extensions we have been considering are
related. Many of the idea.'3 behind Proposition 1.6.3 are due to Narmania but we
take the step of linking them to Cole and standard extensions.
ALGEBRAIC EXTENSIONS OF COMMUTATIVE NORMED ALGEBRAS 165

PROPOSITION 1.6.3. Let A be a commutative, unital Banach algebra and U a


set of monic polynomials over A. Then, up to isometric isomorphism, Au = Bu. If
A is a uniform algebra then we have

u -- --
A = (Au)" = (Bur,

where the closures are taken with respect to the supremum norm.

PROOF. It is easily checked that if B is a normed algebra then the homeomor-


phism 12(B) ---t O(B) induces an isometric isomorphism B" ---t (B)". It is therefore
sufficient to prove that Au = Bu and that AU = (Bu)". The last equality follows
very quickly from the universal property of standard extensions mentioned above
and the simplicity of the definition of AU. We shall only prove the first identifi-
cation; the second can be proved by a similar approach. Although what follows is
routine, we hope that it will help to clarify the details of standard and Narmania
extensions.
As before let to (0 E U) be a valid choice of Arens-Hoffman norm-parameters
for the respective extensions Ao. We shall show that there is then an isometric
isomorphism between Bu and D (when defined by these parameters); the result
then follows from the uniqueness of completions.
For each a E U let Yo be the equivalence cla..<;s [({o(:c o )}) + xo] E D. Since
Yo is a root of V0(0)(X) in D there exists, by the universal property of standard
extensions, a (unique) homomorphism : Bu ---t D such that I A = V0 and for all
a E U, (~o) = Yo' Here, ~o is the the element of Bu associated with x by the
isometric isomorphism 1/)0 : B<o [x]j(o(x)) ---t Bo in the notation of Theorem 1.6.1.
Thus 11~01l = to. Note that, by its definition in [22], 'l/J" satisfies 'l/Jo(a) = a for all
a E B<o'
It is clear that is surjective; we now use the transfinite induction theorem,
as is customary for proving results about standard extensions, to show that is
isometric.
Let :J = {,6 E U: IBIl is isometric}. It should be clear to the reader that
ao E:J. Let,6 E U and suppose that [00,,6) ~:J. Let b E B(3. Then, writ-
ing n(,6) for the degree of ,6(x), there exist unique b1 ,. , bn ((3)-l E B<!3 such
that b = L,j~)-1 bj~~. We have, by hypothesis, that Ilbll = L,;~)-l Ilbj II tb =
L,j!!j-111(bj)11 tb. Since the algebras vs(As) are directed there exists 8 E U<W(I
such that (bj ) E vs(As) (j = 0, ... ,n(,6) -1). We can assume that 8 =
{al,'" ,am} and ol(X) = ,6(x). Let qo,. ,qn((3)-IE A[XQ1"" ,XOm] be the
minimal representatives such that (bj ) = [(8) + qj] (j = 0, ... ,n(,6) -1). So
Ilbjll = Ilqjll (j = 0, ... ,n(,6) - 1). A routine exercise in the transfinite induction
theorem shows that for all 'Y E U, (B"() ~ UTE[O,"(]<w" vT(AT). It follows that the
166 THOMAS DAWSON

degree of qj in XO<l is zero. Hence

n({3)-l

11(b)11 = L reS) + qjX~l] =


j=O
n(tJ)-l

(S) + L qj.T~l
j=O
n({3)-l

= L
j=O
IIqj II t~l = Ilbll ,

from above. The penultimate equality above follows from noting that the represen-
tative of the coset is minimal and then expanding and collecting terms.
By the transfinite induction theorem, .:J = U as required. 0

2. A Survey of Properties Preserved by Algebraic Extensions


2.1. Introduction. We summarise in Table 2.2 what is currently known
about the behaviour of certain properties of normed algebras with respect to the
types of extensions we have been considering. Some preliminary explanation of the
entries is in order first.
Extra information about the polynomial(s) generating an algebraic extension
can help to determine whether certain properties are preserved or not. For ex-
ample if a(x) has degree nand factorises completely over A with distinct roots
AI, ... ,An E A such that for all W E fl, .x:(w) =I- :X;(w) if i =I- j then n(Ao<) decom-
poses into n disjoint homeomorphs of fl in which case very many properties of A,
for example localness, are shared by Ao<' This property, referred to as 'complete
solvability', is investigated in [12].
The condition on a(x) most frequently encountered in the literature is that
it should be 'separable'. This means that its 'discriminant', which is a certain
polynomial in the coefficients of a(x), is invertible in A. It is interesting to compare
columns two and three.
Of course one can make additional assumptions on the algebra (for example
that A be regular and semisimple) but the resulting table would become too large
and we have restricted it to three popular categories.
References to the results follow the table. We should mention that some of the
entries have trivial explanations. For example Sheinberg's theorem, that a uniform
algebra is amenable if and only if it is trivial, explains the entries for amenability in
column one. Also, applying the Arens-Hoffman construction to a uniform algebra
need not result in a uniform algebra so not all the entries make sense.
We have already met most of the properties listed in the table. We end this
section by discussing the ones which have not yet been specially mentioned.
1. Denseness of the invertible group. Although this property is self-explanatory
it might not be obvious why it is listed. However, the condition G(A) = A
appears in the literature in various contexts; see for example [8].
ALGEBRAIC EXTENSIONS OF COMMUTATIVE NORMED ALGEBRAS 167

2. The Banach algebra, A, is called sup-norm closed if A is uniformly closed in


C(f2) (and therefore a uniform algebra). It is called symmetric if A is self-
adjoint.
3. For the definitions of 'amenability' and 'weak amenability' we refer the reader
to section 2.8 of [6].
All the properties in the table are preserved by forming the standard unitisation
of a normed algebra. Most of these results are standard facts or ea'iy exercises; some
are true by definition. However this question does not fit into our scheme because
the embedding is not unital in this case.

2.2. Table. Cole extensions have only been defined for uniform algebra'l;
the algebra is therefore assumed to be a uniform algebra throughout column one.
Colulllns two and three, a'l mentioned above, refer to Arens-Hoffman extensions of
a normed algebra, A, by a monic polynomial a(x); in column three it is given that
a(:1:) is separable.

Type of Extension: Cole A.-H. A.-H. standard Narmania


Property: Ac. a sep.

for normed algebras

1. complete 0

2. topologically semisimple 0
0 0

3. non-local
4. local ? 0 ? '? ?
5. regular 0 '? 0

for Banach algebras

6. local ? '? '! ? ?


7. regular ?
S. dense invertible group
9. sup-norm closed 0
0 0

10. symmetric 0
0 0

11. amenable 0 ? 0 0

12. weakly amenable ? 0 ? 0 0

for uniform algebms

13. non-trivial
14. trivial
15. natural
Key
property is always preserved
o property is sometimes, but not always preserved
168 THOMAS DAWSON

? not yet determined


- it doesn't always make sense to consider this property here
References for the Entries. If we do not mention an entry here, it can be
taken that the result is an immediate consequence of the definition or was proved
in the same paper in which the relevant extension was introduced (that is in [5],
[1], [22], or [23]).
The results of row three are not hard to obtain, using appropriate versions of
Proposition 1.5.1.
Localness and regularity were discussed in Section 1.2. The main result about
this is due to Lindberg in [18]; the same section of his paper also deals with the
results on the symmetry of Arens-Hoffman extensions. That regularity passes to
direct limits of such extensions has been widely noted by many authors, for example
in [15].
Results of row eight follow from [8]; the case of Cole extensions was partially
covered in [15], but the reasoning is not clear.
The property of being sup-norm closed was investigated in [13]; this work was
generalised in [28].
Finally, examples of amenable Banach algebras which do not have even weakly
amenable Arens-Hoffman extensions have been known for a long time. For example,
the algebra C E9 C under the multiplication (a, b) (e, d) = (ae, be + ad) is realisable
as an Arens-Hoffman extension of C. Examples with both A and Ao semisimple
have been found by the author. However the entries marked "?' in rows eleven and
twelve represent intriguing open problems.

3. Conclusion
The table in Section 2.2 still has gaps, and there are many more rows which
could be added. For example it would be interesting to be able to estimate various
types of 'stable ranks' (see [2]) of the extensions in terms of the stable ranks of the
original algebras. (The condition G(A) = A is equivalent to the 'topological stable
rank' of A not exceeding 1.) Remember too that there are many more questions
which can be asked, of the form: 'if n has the topological property P, does n(Ao)
have property P?'
By way of a conclusion we repeat that algebraic extensions have proved im-
mensely useful in the construction of examples of uniform algebras. There is there-
fore great scope for and potential usefulness in augmenting Table 2.2. It might also
be valuable to reexamine the techniques used to obtain the entries to produce more
general results (of the kind in [28] for example) in the context of question (Q).

References
1. Arens, R. and Hoffman, K., Algebraic Extension of Normed Algebras., Proc. Am. Math. Soc.
7 (1956), 203-210.
2. Badea, C., The Stable Rank of Topological Algebras and a Problem of R. G. Swan., J. Funet.
Anal. 160 (1998), 42-78.
3. Batikyan, B. T., Point Derivations on Algebraic Extension of Banach Algebra., Lobachevskii
J. Math. 6 (2000), 3-37.
4. Browder, A., Introduction to Function Algebras., W. A. Benjamin, Inc., New York, 1969.
ALGEBRAIC EXTENSIONS OF COMMUTATIVE NORMED ALGEBRAS 169

5. Cole, B. J., One-Point Parts and the Peak-Point Conjecture., Ph.D. Thesis, Yale University,
1968.
6. Dales, H. G., Banach Algebms and Automatic Continuity., Oxford University Press Inc., New
York,2000.
7. Dawson, T. W .. Algebmic Extensions of Normed Algebms., M.Math. Dissertation, accessible
from the web at: http://xxx.lanl.gov/abs/math.FA/0102131, University of Nottingham,
2000.
8. Dawson, T. W., and Feinstein, J. F., On the Denseness of the Invertible Group in Banach
Algebms., Proc. Am. Math. Soc. (to appear).
9. Feinstein, J. F., A Non-Trivial, Strongly Regular Unif01m Algebm., J. Lond. Math. Soc. 45
(1992), 288-300.
10. Feinstein, J. F., Trivial Jensen Measures Without Regularity., Studia Math. 148 (2001).67-
74.
11. Gamelin, T. W., Uniform Algebms., Prentice-Hall Inc., Engelwood Cliffs, N. J., 1969.
12. Gorin, E. A., and Lin, V ..J., Algebmic Equations with Cont'inuous Coefficients and Some
Problems of the Algebmic Theory of Bmids., Math. USSR Sb. 7 (1969), 569-596.
13. Heuer, G. A., and Lindberg, J. A., Algebmic Extensions of Continuous Function Algebms.,
Proc. Am. Math. Soc. 14 (1963),337-342.
14. Johnson, B. E., Norming C(O) and Related Algebms., Trans. Am. Math. Soc. 220 (1976),
37-58.
15. Karahanjan, M. I., Some Algebmic Chamcterizations of the Algebm of All Continuous Func-
tions on a Locally Connected Compactum., Math. USSR Sb. 35 (1979),681-696.
16. Leibowitz, G. M., Lectures on Complex Function Algebm,~., Scott, Foresman and Company,
Glenview, Illinois, 1970.
17. Lindberg, J. A . On the Completion of Tractable Normed Algebms., Proc. Am. Math. Soc. 14
(1963),319-321.
18. Lindberg, J. A., Algebmic Extensions of Commutative Banach Algebms., Pacif. J. Math. 14
(1964), 559-583.
19. Lindberg, J. A., On Singly Genemted Topological Algebras, Function Algebras (ed. Birtel,
F. T.), Scott-Foresman, Chicago, 1966, pp. 334-340.
20. Lindberg, J. A., A Class of Commutative Banach Algebms with Unique Complete Norm
Topology and Continuous Derivations., Proc. Am. Math. Soc. 29 (1971), 516-520.
21. Lindberg, J. A .. Polynomials over Complete l.m.-c. Algebms and Simple Integml Extensions.,
Rev. Roumaine Math. Pures Appl. 17 (1972), 47-63.
22. Lindberg, J. A., lntegml Extensions of Commutative Banach Algebms., Can. ,/. Math. 25
(1973), 673-686.
23. Narmaniya, V. G., The Construction of Algebmically Closed Exten,~ions of Commutative
Banach Algebms., Trudy Tbiliss. Mat. Inst. Razmadze Akad. 69 (1982), 154-162.
24. Palmer, T. W., Banach Algebras and the Geneml Theory of *-Algebms. Vol. 1, Cambridge
University Press, Cambridge, 1994.
25. Read, C. J., Commutative, Radical Amenable Banach Algebms., Studia Math. 140 (2000),
199-212.
26. Stout, E. L., The Theory of Uniform Algebms., Bogden and Quigley Inc., Tarrytown-on-
Hudson, New York, 1973.
27. Taylor, J. L., Banach Algebms and Topology, Algebras in Analysis. (ed. Williamson, J. H.),
Academic Press Inc. (London) Ltd., Norwich:, 1975, pp. 118-186.
28. Verdera, J., On Finitely Genemted and Projective Extensions of Banach Algebms., Proc. Am.
Math. Soc. 80 (1980), 614-620.
29. Zame, W. R., Covering Spaces and the Galois Theory of Commutative Banach Algebms., J.
Funet. Anal. 27 (1984), 151-171.

Acknowledgements
The author would like to thank the Division of Pure Mathematics and the
Graduate School at the University of Nottingham for paying for his expenses in
170 THOMAS DAWSON

order to attend the 4th Conference on FUnction Spaces (2002) at the Southern
Illinois University at Edwardsville.
The author is grateful to Mr. Brian Lockett who provided him with a transla-
tion of the paper [23].
Special thanks are due to Dr. J. F. Feinstein who offered much valuable advice
and encouragement and also proofread the article.

DIVISION OF PURE MATHEMATICS, SCHOOL OF MATHEMATICAL SCIENCES, UNIVERSITY OF


NOTTINGHAM, UNIVERSITY PARK, NOTTINGHAM, NC7 2RD, UK.
E-mail address:pmxtwdlDnottingham.ac.uk
Contemporary Mat.hematics
Volume 328, 200a

Some more examples of subsets of Co and Ll[O, 1] failing the


fixed point property

P.N. Dowling, C.J. Lennard, and B. Thrett

ABSTRACT. We give examples of closed, bounded. convex, non-weakly com-


pact subsets of Co on which the right shift is expansive, and we construct two
nonexpansive self-mappings (one affine and one non-affine) on these sets which
fail to have a fixed point. We also prove that every closed, bounded, convex
subset of L1 [0,1] with a non-empty interior fails the fixed point property for
nonexpansive mappings. Finally, we extend this result by showing that every
closed, bounded, convex subset of L1 [0,1] that contains a non-trivial order
interval must fail the fixed point property.

1. Introduction
In [5], Llorens-Fuster and Sims construct examples of closed, bounded, convex
subsets of Co that are not weakly compact but are compact in a topology that is
slightly coarser than the weak topology, and they nonetheless fail the fixed point
property for nonexpansive mappings. These examples led Llorens-Fuster and Sims
to conjecture that closed, bounded, convex non-empty subsets of Co have the fixed
point property if and only if they are weakly compact. This conjecture has been
recently settled in the affirmative [2, 3]. All the examples constructed in [5] had a
common feature - they all support a nonexpansive right shift.
In the first part of this short note we produce a collection of sets of the type
considered by Llorens-Fuster and Sims, but which do not support a nonexpansive
right shift and yet they fail the fixed point property for nonexpansive mappings. In
fact, we will produce two nonexpansive fixed point free mappings: one affine and
the other non-affine. Variations on the themes of these examples are important in
the papers [2, 3].
Llorens-Fuster and Sims [5] also proved that a closed, bounded, convex subset of
Co with non-empty interior fails the fixed point property for nonexpansive mappings.
In the second part of tllis note we prove an analogous statement in the setting of
L1 [0, 1]. We also generalize this result to show that every closed, bounded, convex
2000 Mathematics Subject Classification. Primary 47HI0, 47H09, 46E30.
The authors wish to thank Professor Kaz Goebel for his helpful suggestions concerning the
proof of Theorem 3.2. The second author thanks the Department of Mathematics and Statistics
at Miami University for their hospitality during part of the preparation of this paper. He also
acknowledges the financial support of Miami University.

2003 American l\1athematical Society


171
172 P.N. DOWLING, C.J. LENNARD, AND B. TURETT

subset of L1 [0, 1J that contains a non-trivial order interval must fail the fixed point
property.
We refer the reader to the text of Goebel and Kirk [4J for any unexplained
terminology.

2. The fixed point property in Co


We begin this section with the Llorens-F\lster and Sims examples. We have
slightly modified their examples to simplify the computations.
Let r denote the set of sequences 'Y = ("(n)n in the interval (0,1). For all 'Y E r,
define K., by
K..., := {x = ("(ntn)n E Co 112: h 2: t2 2: t3 2: ... 2: O} .
We define the right shift Ton K..., by
T( ("(1 t 1 , 'Y2t2, 'Y3 t3, 'Y4t4, ... )) := ("(1, 'Y2t 1, 'Y3 t2, 'Y4t3, ... ).
Note that if x = ("(ntn)n and y = ("(nsn)n are elements of K..." then Ilx - yll
sUPn 'Ynlt" - 8 n l and IIT(x) - T(y)11 = SUPn 'Yn+1Itn - 8 n l Clearly, if the sequence
("(n)n is decreasing, then T is a nonexpansive mapping on K..., - this is the case
considered by Llorens-Fuster and Sims [5J. However, it is equally obvious that if
the sequence ("(n)n is strictly increasing, then T is an expansive mapping on K...,;
that is, IIT(x) - T(y)11 > IIx - yll whenever x i= y. We will show that even though
some of these sets do not support a nonexpansive right shift, they do support
nonexpansive fixed point free mappings.
To simplify our computations we will only consider the sets K..., where the
sequence 'Y = ("(n)n is in (0,1), is strictly increasing and satisfies 1 - 'Yn < 4- n for
all n E N.
EXAMPLE 2.1. Let I be the identity mapping on K..." let T be the right shift
defined above, T2 = ToT, T3 = ToT 0 T, and so on. Define R: K..., --+ K..., by
R := ~I + -b T + -1a T2 + 21 4 T3 +... .
A simple calculation shows that if x = ("(1 t1, 'Y2t2, 'Y3t3, ... ) E K..." then
R(x) = ('Y1(~t1 + ~),'Y2(~t2 + it1 + i),'Y3(~t3 + it2 + kt1 + ~), ... ).
It is easily seen that if R(x) = x, then tn = 1 for all n E N, and thus x is
not an element of K...,; that is, R is fixed point free on K...,. To see that R is
nonexpansive on K..." let x = ("(ntn)n and y = ("(nsn)n be elements of K...,. Then,
12 + - 4
since 1 - 4- n < 'Yn < 1, we have 'Yn( -...,,, 1 + ... + 2,,1...,1 ) < 1 for each n E N.
"",,-1
Consequently,
IIR(x) - R(y)11 sup hnl~(tn - sn) + i(tn-1 - sn-d + ... + 2~ (t1 - st)l}
n

< sup hn(~ltn - snl + iltn-1 - 8 n -11 + ... + 2~' It 1 - Sl J)}


n

< sup
n
{'Yn(_l-
2...,,,
+ _1_
4"",,-1
+ ... + +) max 'Yilti - Sil}
2""1 l::;'i::;n

< sUP'Ynltn - 8n l
n
Ilx-yll
Thus R is a nonexpansive mapping on K...,.
EXAMPLES OF FIXED POINT FREE MAPS 173
The mapping R, given in example 2.1, is an affine mapping on K"I' Our next
example is non-affine on K"I'

EXAMPLE 2.2. For an element x = (xn)n in K"I' we denote by bl,X) the


sequence bl,Xl,X2,X3, ... ). We define a mapping S: K"I -+ Co by S(x):= X, for
each x E K"I' where x = (Xl, X2, X2, ... ) is the decreasing rearrangement of the
sequence bl' x); that is, X = bl, x)*. Note that

X= (Xl,X2,X3, ... ) = (11 (~~) ,12 (~:) ,13 (~:) , ... ).


Also Xl 2: X2 2: X3 2: ... 2: 0 and 0 < 11 < 12 < 13 < ... < 1. Therefore
~ > ~ > ~ > ... > O. Since ~ = max("("x) > 1 S(x) does not necessarily belong
~-n-n- - ~ ~ -,
to K"I' However, the mapping S is nonexpansive on K"I because the operation of
decreasing rearrangement is nonexpansive on Co, so for all x and y in K"I' we have

IIS(X) - S(Y)II = IIx - yll = IIbl,X)* - bl,Y)*11 :::; Ilb1.x) - b1.y)11 = Ilx - YII
We now introduce a modification U of S that will be nonexpansive and fixed
point free on K"I' Define U : K"I -+ Co by

Since 11j AXj - 1j AYj I :::; IXj - Yj I for all j E N, it follows that
IIU(x) - U(Y)II :::; Ilx - yll = IIS(x) - S(Y)II :::; IIx - YII
Thus U is a nonexpansive mapping on K"I'
Furthermore, since 1j AXj = 1j (1 A ~) for all j E Nand 1 2: 1 A ~ 2: 1 A ~ 2:
1 A ~ 2: ... 2: 0, U maps K"I into K"I' To finish, we will show is fixed point free on
K"I' Suppose, to get a contradiction, that there exists x E K"I such that x = U(x).
Thus, for all j EN, Xj = 1j AXj.
A well-known fact about decreasing rearrangements that we will use is that for
all W E ct,for each mEN,

WI + ... + Wm :::; wi + ... + w;;'.


Since 1 = (1n)n is strictly increasing with limit 1, while X is decreasing with
1 > Xl 2: 11, there exists a unique kEN such that Xk+1 < 1k+l and Xk 2: 1k.
Thus, for all mEN with m > k, we have,

k + Xl + ... + Xm+l > (Xl + ... + Xk) + (Xl + ... + xm+d


(Xl + ... + Xk) +
bl A Xl + ... + 1k A Xk + 1k+l A Xk+1 + ... + 1m+1 A Xm+1
= (Xl + ... + Xk) + bl + ... + 1k + Xk+l + ... + xm+d
bl + ... + 1k) + (Xl + ... + Xk + Xk+1 + ... + Xm+1)
bl + ... + 1k) + (Xl + ... + Xm+l)
> bl+"'+1k)+bl+Xl+"'+Xm),
174 P.N. DOWLING, C.J. LENNARD, AND B. TURETT

It follows that for all Tn E N with Tn > k,


k
Xm +l > bl+"'+'Yk)-k+'Yl='YI-L(1-'Yj)
j=1
= =. 1 1
> 'Y1 - L (1 - 'Yj) ~ 'Yl - L 4- J = 'Yl - :3 > :3'
j=1 j=1

This contradicts the fact that x E Co and so completes the proof that U is fixed
point free on K-y.

3. The fixed point property in 1 [0, 1]


One of the most notable works in metric fixed point theory is the construction
of Alspach [1] of a non-empty weakly compact convex subset of Ll[O, 1] which fails
the fixed point property. We begin this section by recalling some of the details of
Alspach's construction.
Let C:= {f E Ll[O, 1] : 0::; f(t) ::; 1, for all t E [0, I]}. Now define T: C -+ C
by
Tf(t) := {min{2f (2t), I} for 0 ::; t ::; ~
max{2f(2t - 1) - 1, O} for ~ < t ::; 1.
for all f E C.
Alspach showed that the mapping T is an isometry on C which has two fixed
points; namely 0 and X[O,lj' Alspach also showed that T is an isometric self map of
the closed convex subset Co := {f E C : J~ f dm = 1/2} of C, such that T is fixed
point free on Co. Here, m denotes Lebesgue measure.
We now follow a modification of Alspach's example due to Sine [7]. Define
S : C -+ C by S(f) := X[O,lj - f, for all f E C. The mapping S is clearly an
isometry of C onto C. Thus the mapping ST is a nonexpansive mapping on C.
Sine proved that ST is fixed point free on C.
In [5], Llorens-FUster and Sims prove that a closed bounded convex subset of
Co with non-empty interior fails the fixed point property. We will use the above
construction of Alspach, and modification by Sine, to prove a result analogous to
the Llorens-FUster and Sims result in the setting of Ll [0, 1]. Specifically we prove
the following result.
THEOREM 3.1. Let K be a closed, bounded, convex subset of Ll [0, 1] with non-
empty interior. Then K fails the fixed point property for non expansive mappings.
PROOF. By translating and scaling, we ca.n assume that K contains the unit
ball of 1 [0, 1]. Consequently, the set C, constructed above, is a subset of K. Define
the mapping R : K -+ K by
Rf(t) := min{lf(t)l, I}, for 0::; t::; 1, for all f E K.
It is easily seen that R is a. nonexpansive mapping on K and R(f) E C for all
f E K. Now define U: K -+ K by
U(f) := ST(R(f)), for all f E K.
The mapping U is nonexpansive since all of the mappings, R, S, and Tare
nonexpansive.
EXAMPLES OF FIXED POINT FREE MAPS 175

We now show that U is fixed point free. Suppose that f E K is a fixed point
of U, that is, U(f) = f. Since f E K, R(f) E C, and since ST maps C into
C, f = U(f) = ST(R(f)) E C. Note that the mapping R restricted to C is the
identity on C. Therefore, f = ST(R(f)) = ST(f) and so f is a fixed point of ST
in C. This contradicts Sine's result that ST has no fixed point in C [7], and thus
the proof is complete. D

THEOREM 3.2. Let K be a closed, bOllnded, convex sllbset of Ll[O, 1] that con-
tains an order interval [h,g] := {f E L1[0, 1] : h ::; f ::; 9 a.e.}, for some
h,g E Ll[O, 1] with h ::; 9 a.nd h ::f g. Then K fa'ils the fixed point property for
nonexpansive mappings,

a
PROOF. By translating by -h, we may assume that h = and 9 ~ a.e. with a
a
9 non-trivial. Next, note that there exists a real number c > and a measurable
set E with Lebesgue measure rn(E) > 0, such that 9 ~ CXE. By rescaling K by
lie, we may assume without loss that c = 1.
Now, define the mapping R : K ~ [0, xel ~ K by R(f) := ifi /\ XE. Note that
R is nonexpansive and R equals the identity on [0, xel.
At this stage, consider E. There exists to in the interval [0,1] such that rn(En
[0, to]) = ~ rn(E), Let E1.1 := En [0, to] and E1.2 := En (to, 1]. Clearly E is
the disjoint union of E1.1 and E 1,2 and rn(E 1,1) = rn(El,2) = ~ rn(E). Proceed
iteratively from here. Similarly to above, there exist pairwise disjoint measurable
subsets E2,1, E 2,2, E2,3 and E2,4 of [0, 1] such that El,l = E 2,1 U E2,2, E 1 ,2 = E 2,3 U
E 2.4 , and each rn(E2,k) = ~ rn(E). Repeating this construction inductively, we
produce a family of measurable subsets (EO,1 := E, En,k : n E N, k E {I, ... , 2n}) of
[0,1] such that (XE",k)n.k is a dyadic tree in Ll [0,1]. Moreover, letting the measure
v be defined on the measurable subsets of E by v = (1/rn(E))rn, it follows that the
Banach space L 1 ( E, v) is isometrically isomorphic to L 1 ( [0, 1], m) = L 1 [0, 1] via the
mapping Z defined as follows:

Z(XEn, k):= X[k-l k),


2l'r'2"'t"

for each XE",k' Then Z is extended to L := the linear span of the functions XE",k
in the usual way. Of course, Z is an isometry on L. Finally, since L is dense in
Ll(E, v), with dense range in Ll [0,1], Z extends to a linear isometry from L1(E, v)
onto Ll[O, 1].
Let W := ST be Sine's variation on Alspach's example, as described above,
and note that W maps the order interval C := [0, X[O,I]] into C. Let's use W to
define E : [0, xel ~ [0, XE], by E := Z-1 W Z. We have that E is a fixed point free
Ll [0, 1]-isometry on [0, xel.
Finally, we define U : K ~ [0, XE] ~ K via U := E R. In a manner analogous
to the argument in the proof of Theorem 3.1 above, we see that U is a fixed point
free Ll[O, 1]-nonexpansive mapping on K. D

REMARK 3.3. In [6], MatlI'ey proved that closed, bounded, convex, non-empty
subsets of reflexive subspaces of Ll [0, 1] have the fixed point property for nonexpan-
sive mappings. Consequently, Maurey's result, in tandem with Theorem 3.2, shows
that reflexive subspaces of U [0,1] cannot contain a non-trivial order interval. In
fact, as pointed out by the referee, the argument in the proof of Theorem 3.2 shows
176 P.N. DOWLING, C.J. LENNARD, AND B. TURETT

that infinite-dimensional subspaces of 1[0,1] which contain non-trivial order in-


tervals actually contain isometric copies of 1 [0, 1] and thus are nonreflexive. The
authors thank the referee for his/her comments.

References
1. D. Alspach, A fixed point free nonexpansive mapping, Proc. Amer. Math. Soc., 82 (1981),
423-424.
2. P.N. Dowling, C.J. Lennard and B. Thrett, Characterizations of weakly compact sets and new
fixed point free maps in co, to appear in Studia Math.
3. P.N. Dowling, C.J. Lennard and B. Thrett, Weak compactness is equivalent to the fixed point
property in co, preprint
4. Kazimierz Goebel and W.A. Kirk, Topics in metric fixed point theory, Cambridge University
Press, Cambridge, 1990
5. Enrique Llorens-Fuster and Brailey Sims, The fixed point property in co, Canad. Math. Bull.
41 (1998), no. 2, 413-422.
6. B. Maurey, Points fixes des contractions de certains faiblement compacts de Ll, Seminaire
d'Analyse Fonctionelle, 1980-1981, Centre de Mathematiques, Ecole Polytech., Palaiseau,
1981, pp. Exp. No. VIII, 19.
7. R. Sine, Remarks on an example of Alspach, Nonlinear Anal. and Appl., Marcel Dekker,
(1981), 237-241.

DEPARTMENT OF MATHEMATICS AND STATISTICS, MIAMI UNIVERSITY, OXFORD, OH 45056


E-mail address: dowlinpnGmuohio. edu

DEPARTMENT OF MATHEMATICS, UNIVERSITY OF PITTSBURGH, PITTSBURGH, PA 15260


E-mail address: lennard+lDpi tt. edu

DEPARTMENT OF MATHEMATICS AND STATISTICS, OAKLAND UNIVERSITY, ROCHESTER, MI


48309
E-mail address: turettlDoakland. edu
Contemporary Mathelnatics
Volume 328. 2003

Homotopic composition operators on Hoo (Bn)

Pamela Gorkin, Raymond Mortini, and Daniel Suarez

ABSTRACT. We characterize the path components of composition operators on


Hoo(B n ), where Bn is the unit ball of en. We give a geometrical equivalence
for the compactness of the difference of two of such operators. For n = 1, we
give a characterization of the path components of the algebra endomorphisms.

1. Introduction
Consider the Hardy space H2 on the unit disk D. Littlewood's subordination
principle tells us that for an analytic self-map of D and a function f in H2, the
function f 0 is once again in H2. Thus one defines the composition operator C'"
on H2 by C",(f) = f 0 . The interplay of operator theory and function theory
leads to several interesting results. One of these results is Berkson's theorem on
isolation of composition operators (see [1] and [14]):
THEOREM 1 (Berkson). Let be an analytic self-map of D. If has mdial
limits of modulus one on a set E of positive measure, then for every other analytic
self-map 'l/J of D, the following estimate holds:

IIC", - C",II2: Jmea;(E),


where C'" and C'" are the corresponding composition opemtors on H2.
Thus, Berkson's theorem tells us that every such operator is isolated in the set
of composition operators in the operator norm topology. For example, the identity
operator, C z , is at least a distance of ..[f72 from every other composition operator
on H2 (as is C"', where is any inner function). However, not every composition
operator is isolated. If is analytic and : D ~ sD for some s with 0 < s < 1,
then it is easy to check that

Thus, C'" is not isolated.


Inner functions induce highly noncompact operators, as well as isolated opera-
tors. The operators C'" for which (D) is contained in sD for some s with 0 < s < 1

2000 Mathematics Subject Classification. Primary 47B33; Secondary 47B38.


Key words and phrases. composition operator, path components, compact differences.

2003 American lvlathematical Society


177
178 PAMELA GORKIN. RAYMOND MORTINI. AND DANIEL SUAREZ

are compact. As Shapiro and Sundberg [14J indicate in their paper, "compact com-
position operators are dramatically nonisolated." They show that the set of com-
pact composition operators is path connected, and therefore these operators are
never isolated. It is interesting to ask which composition operators are, in fact,
isolated. Shapiro and Sundberg studied this problem, and showed (among other
things) that if is au analytic self-map of D that is not an extreme point of the
algebra HOC(D), then C</> is not isolated; in their words, "isolated composition op-
erators can only be induced by extreme points." This allowed them to exhibit an
example of a non-compact non-isolated operator. They also raised several questions
at the end of their paper:
(1) Characterize the components in Comp(H2), the space of all composition
operators on H2.
(2) Which composition operators are isolated?
(3) Characterize composition operators whose difference is compact.

Before stating the final question, we remind the reader that the essential norm
of an operator T defined on a Banach space H is the distance to the compact
operators; that is,

liT lie = inf{IIT - KII : K compact on H}.

It is clear that IITII ~ IITlle, and therefore every essentially isolated operator is
isolated. In fact, because of the abundance of weakly null sequences in H2, all
the results on isolation appearing in Shapiro and Sundberg's paper hold true if we
replace the norm with the essential norm (see [14], p. 148). Thus they raised the
following question.

(4) Is every isolated operator essentially isolated?


Other papers of interest on this subject include [10J. Of course, one is not
limited to the space H2, and the study of composition operators on various spaces
has lead to a large body of literature.
In this paper, we are interested in the same problems for composition operators
on HOC (Bn), where Bn denotes the open unit ball in en.
While the problem 011 H2 seems to be difficult, MacCluer, Ohno and Zhao [l1J
were able to obtain partial results about operators on the algebra HOO(D). They
showed that for two analytic self-maps of the disk, C</> and C.p are in the same path
component in the space of composition operators, Comp(HOC(D)), if and only if
IIC", - c.pll < 2. In particular, an operator C</> is isolated in Comp(HOC(D)) if and
only if IIC", - C",II = 2 for any other analytic self-map 'IjJ of the disk. The authors
show that this result can be rephra..<;ed in terms of the pseudohyperbolic metric,
and they posed the question of whether or not every isolated composition operator
on HOC is, in fact, essentially isolated.
The answer to this la..<;t question was given by Hosokawa, Izuchi and Zheng
[8J. Their technique was to develop something called asymptotic interpolating se-
quences, or a.i.s. for short. Essentially, this definition allowed them to interpolate
sequences with a good bound on the norm (see [6J for more information about
these sequences). They then used these sequences and Blaschke products to ob-
tain HOC(D) functions that provide good estimates for the essential norm of the
difference of two composition operators.
HOMOTOPIC COMPOSITION OPERATORS ON Hoo(Bn) 179

In this paper, we give simpler proofs of the results obtained by Hosokawa,


Izuchi and Zheng, and combine them with the proofs of MacCluer, Ohno and
Zhao. Because our proofs are significantly simpler and do not refer to asymptotic
interpolating sequences or Blaschke products, we are able to obtain the results on
the ball in en. While our results do not rely on interpolating sequence results, they
do rely on a construction of Gamelin and Garnett relating interpolating sequences
to peak sets [5]. The proof we will provide is simple enough to be applicable to other
algebras. We conclude the paper with an example of such an application: Using
these same techniques, we are able to "lift" these results to obtain a characterization
of the path components of endomorphisms of HOO(D).
After we completed this paper, we learned that some of the results on com-
position operators on Hoo(Bn) were also obtained by Carl Toews [15]. Two other
papers directly related to the results described here are [3] and [9]. Finally, we
mention that recent results on Shapiro and Sundberg's first question in the space
H2 can be found in [12].

2. Preliminary results
Our goal is to prove results about composition operators on Hoo(Bn) and
endomorphisms of HOO(D). In this section, we present proofs of several lemmas
that will be important in obtaining estimates on norms and essential norms of
operators.
Our discussion begins with functions of n variables and Z = (Zl' Z2, .. , zn),
where each Zj is a complex number. As usual, the associated norm of Z is given by

Izi = (z, z)1/2,

and the unit ball Bn is the set of all Z E en for which Izi < 1. We let Hoo(Bn)
denote the space of bounded holomorphic functions on Bn. If n = 1 our situation
reduces to the familiar space of functions on the unit disk: HOO(D). We need some
background on general uniform algebras, some information specific to Hoo(Bn) and
some deeper results for HOO(D). Everything we need is presented in this paper.
Let A be a uniform algebra. The maximal ideal space of A, denoted M(A), is
the set of complex-valued, linear, multiplicative maps of A that map the identity
of A to the value 1 E C. Since evaluations at points of Bn are linear multiplicative
functionals on Hoo(Bn), we may think of Bn as a subset of M(Hoo(Bn)).
It is well known that M(A) is a compact Hausdorff space when endowed with
the weak-* topology induced by A * . We will always consider this topology for
M(A). For an element a E A, the Gelfand transform of a is a complex-valued map
defined on M(A) by a(x) = x(a). This map establishes an isometric isomorphism
between A and a closed subalgebra of C(M(A)). It is usual to identify the function
with its Gelfand transform, since the meaning is generally clear from the context.
For x, y E M(A), the pseudohyperbolic and hyperbolic metrics are defined, respec-
tively, by

p(x, y) = sup{l/(y)1 : I E A, 11/11 ::; 1, and I(x) = O} and

h( x, y ) -- Iog 11 + p(x,
(
y)
).
-p x,y
180 PAMELA GORKIN, RAYMOND MORTINI, AND DANIEL SUAREZ

It is well-known that P is a [0, 1]-valued metric and that h is a [0, +oo]-valued


metric on M(A). The triangle inequality for h immediately implies that the condi-
tion p(x, y) < 1 (i.e. h(x, y) < 00) is an equivalence relation on M(A).
If A is a uniform algebra and : M(A)-+M(A) is a continuous map such that
a 0 E A for all a E A, then the map C'" defined by
C",a = a 0
is an endomorphism of A. We will write C'" E End (A). Conversely, if E E End (A),
we may define : M(A)-+M(A) by (x)(a) = x(E(a)), obtaining E = C"'.
The Shilov boundary of A is the smallest closed subset of M(A) on which every
function in A attains its maximum. We denote the Shilov boundary by 8A.
LEMMA 2. Let A be a uniform algebra and let C"', C1/J E End (A). Let V c
M(A) be a set whose closure contains 8A. Then,
2p((x),1/I(x))
(2.1) sup J :::; IIC", - C1/J1I :::; 2 sup p((x), 1/I(X)).
xEV 1 + 1 - p((x), 1/I(x))2 xEV

PROOF. First we show that if x, y E M(A), then

(2.2 ) 2p(x, y) II II 2 ( )
1
+ Jl - p(x,y )2:::; X - Y A*:::; p x, y .

The proof uses the techniques of [4, p. 144]. Let I E A with Ilfll < 1. It is clear

that 1 = U - f(y))/(1 - I(y)f) E A, l(y) = and 11111 :::; 1. By definition of p
then lj(x)1 :::; p(x, y), and consequently
If(x) - f(y)1 :::; 11 - f(y)f(x)1 p(x, y) :::; 2p(x, y).
Taking the supremum over Ilfll < 1 we get the upper inequality.
Now we turn to the lower inequality. For simplicity, we write p = p(x, y).

Choose fn E A with I!fnl! < 1, In(x) = and fn(y) > p-l/n. Let Pn = fn(Y) and
Ln(z) = (tn - z)/(I- tnz), where tn = (1- Jl- P~)/Pn' Therefore Ln 0 fn E A,
IILn 0 Inll :::; 1 and
IIx - yl!A* ~ I(Ln 0 fn)(x) - (Ln 0 fn)(Y) I = Itn - L .. (Pn)l
A simple computation shows that

Itn - Ln(Pn)1 = 1Pn(1 - t;) 1-+ 2p .


I-tnPn 1+~
The lemma will follow immediately from (2.2) and the following chain of identities:
IIC",-C1/J1I sup sup IU 0 )(x) - U 0 1/I)(x) I
11/11=1 xE8A
sup sup IU 0 )(x) - U 0 1/I)(x) I
1I/1I=lxEV
sup sup 11((x)) - f(1/I(x))1
xEV 11/11=1
sup 11(x) -1/I(x)IIA*.
xEV
o
LEMMA 3. Let A be a uniform algebra and let C"', C,,} E End (A). The condition
IIC", - C1/J1I < 2 is an equivalence relation.
HOMOTOPIC COMPOSITION OPERATORS ON Hoo(Bn) 181

PROOF. It is obvious that the relation is reflexive and symmetric. So, suppose
that ,'l/J and t.p define endomorphisms of A such that IICeI> - Cvlli < 2 and IICob -
C<pll < 2. By (2.1) we see that
0'1 = sup p((x),'l/J(x)) < 1 and 0'2 = sup p('l/J(x),t.p(x)) < 1.
xEM(A) xEM(A)

Using the hyperbolic metric h on M(A) we obtain


1 + 0'1 1 + 0'2
sup h((x), t.p(x)) :::; log - - + log - - = /3,
XEM(A) 1 - 0'1 1- 0'2

and consequently sUPXEM(A) p((x), t.p(x)) :::; (ei3 -1)/(ei3 + 1) < 1. A new applica-
tion of (2.1) yields the desired result. 0

LEMMA 4. Let A be a uniform algebra. If {fn} is a sequence of functions in


the unit ball of A tending pointwise to zero on 8A, then {fn} tends to zero weakly
in A.
PROOF. As indicated in the introduction, using the Gelfand transform, we
may think of A ~ C(8A). Let x be any element of the dual space of A. By the
Hahn-Banach theorem, x has a continuous norm preserving extension to the space
of continuous functions on the Shilov boundary. Therefore, there exists a finite
measure ILx on 8A such that

x(f) = r
loA
f dJtx

But Ilfnll :::; 1 for all n, and fn ...... 0 pointwise on 8A, so we may apply the
Lebesgue dominated convergence theorem to conclude that 3;(fn) ...... O. Therefore,
the sequence {fn} converges to zero weakly. 0

We will need another estimate, but this will depend on the pseudohyperbolic
metric particular to the ball, Bn. For a, z E Bn, let Sa = lal 2 , and the Jl -
projections Pa and Qa be given by Pa(z) =
((z' a)) a and Qa = 1- Pa. Relevant
a,a
computations can be found in [13, p. 25]. On Bn, the pseudohyperbolic metric
induced by Hoo(Bn) is given by

(
p a, z
)= la - Pa(z) - saQa(z) I
1 _ (z, a) ,
In what follows, for points a and z in the ball and numbers sand t in the closed
interval [0,1]' we let as = a + s(z - a) and asH = a + (s + t)(z - a).
LEMMA 5. Let a, z E Bn. For s, t E [0,1] .satisfying t :::; 1 - s we have
tp(a,z)
(2.3) p(a + s(z - a), a + (s + t)(z - a)) :::; 1 _ (1 _ t)p(a, z)

PROOF. We can assume that a =I z, because otherwise there is nothing to


prove. We consider first the case s = O. Since Pa and Qa are linear operators
satisfying Pa(a) = a and Qa(a) = 0, the nmnerator of p(a, a + t(z - a)) satisfies
a - Pa(a + t(z - a)) - saQa(a + t(z - a)) = ta - tPa(z) - tsaQa(z).
182 PAMELA GORKIN, RAYMOND MORTINI, AND DANIEL SUAREZ

We have assumed that a =I- z, and therefore a - Pa(z) - saQa(z) =I- O. A simple
computation gives

p(a,a+t(z-a)) Ia - Pa{a + t(z - a)) - satQa(z) I


1 - (a + t(z - a), a)
= I t(a - Pa(z) - saQa(z)) I
1 - (z, a) + (z, a) - (a, a) - t(z - a, a)

< 11/p(Z,a) -1(1- t)(z - a,ta)I/la - Pa(z) - saQa(z)ll


But I(z - a,a)1 = I(a - Pa(z) - saQa(z),a)1 ::; la - Pa(z) - saQa(z)llal. Therefore
t tp(a,z)
(2.4) p(a. at) < = .
. - (1/ p(a, z)) - (1 - t) 1 - (1 - t)p(a, z)
This proves (2.3) for s = O. For the general case we can assume s =I- 1 since
otherwise t = 0 and there is nothing to prove. From (2.4) we obtain
p(a s, as + t(z - a))
p(as,a s + (t/(I- s))(z - as))
t/(1 - s)
<
(l/p(a s,z)) - (1- (t/(I- s)))'
But
p(a + s(z - a), z)
p(z + (1 - s)(a - z), z)
p(z, z + (1 - s)(a - z)).
So we may apply (2.4) again to conclude that
I-s
p(as,z)::; (l/p(a,z)) -s
Combining this with our estimate on p(a s , as+t) above, we see that
t
p(a s , as+t) ::; (1/ p(a, z)) - (1 - t)
Simplifying, we obtain the desired conclusion. o
3. Composition operators on Hoc (Bn)
In this section we study composition operators on Hoc(Bn); that is, given an
analytic self-map > of the unit ball, we look at maps C'" : Hoc(Bn) ~ Hoc(Bn)
defined by C",(f) = f 0 >. These maps are all endomorphisms of the algebra
HOC (Bn). For the special case of n = 1 we will say more in the final section of the
paper.
We are interested here in estimates on the essential norm of the difference of
two composition operators. If T is a bounded operator, we denote its essential norm
by IITlle.
THEOREM 6. Let > and '!/J be holomorphic self-maps of Bn such that
max{II>II, II'!/JII} = 1.
HOMOTOPIC COMPOSITION OPERATORS ON H"""(Bn) 183

Let
e= max {lim sup p((Z), 'lj!(Z)) , lim sup p((Z), '(Z))} .
1",(z)l-l 1'I/J(z)l-l
Then
(3.1)
PROOF. By hypothesis there is a sequence of points {Zj} in Bn, such that
p((Zj),-(Zj)) ~ e, and one of them, say {(Zj)}, converges to a point ( on the
boundary of Bn. Without loss of generality we may assume that p( ( Zj ), 'lj!( Zj)) >
e-1fj. Let kj E Hoo(Bn) be a function of norm one, whose existence is guaranteed
by the pseudohyperbolic distance definition, satisfying kj((zj)) > e - 1fj and
kj('lj!(zj)) = O. Consider the functions
f(z) = (1 + (z, () )/2 and g(z) = (1 - (z, () )/2.
Then f,g E HOO(B n ), f(() = 1, If(7])1 < 1 for all 7] E aBn satisfying 7] -I- (, and
g(() = O. We will now produce a sequence of functions, {hj}, tending to zero
weakly for which Ihj((zj)) - hj (1/I(Zj)) I ~ e. We proceed as follows.
Let j E N. Since f((zj)) ~ 1, we may choose Zmj so that If((zmj))lj >
J 1 - 1 fj. Now since 9 -I- 0 on Bn, there exists an integer Ij such that
Ig((zmj))1 1/ 1j 2: J1- 1fj.
Consider the functions hj = (gl/1 j )(J3)kmr Then Ilhjll ~ 1, hj (1/I(zmj)) = 0, and
Ihj((zmJ)1 > (1 - 1fj)(e - 1/mj). We note that hj ~ 0 on the Shilov boundary
and, by Lemma 4, hj ~ 0 weakly. Thus for any compact operator K we have
IIC", - C,p + KII > IIC",hj - C,phj + Khjll
2: Ihj((zmj)) - hj('lj!(zmj)) + (Khj)(zmJI
= Ihj((zmJ) + (Khj)(zmj)1 ~ e
This proves the lower inequality in (3.1).
For the upper inequality, let I: > 0 and choose 8 with 0 < 8 < 1 close enough
to 1 so that
p((Z), 'lj!(z)) ~ e + I: on the set {1(z)1 > 8} U {1'lj!(z)1 > 8}.
Now choose a = a(I:,8) E (0,1) close enough to one so that p((z),a(z)) < I: and
p('lj!(Z) , ml'{z)) < I: on the set {1(z)1 ~ 8}n{I1/I(z)1 ~ 8}. Since max{llalI, Ila'lj!ll} ~
a < 1, the operator K ~f Co", - Co,p is compact. If Z and ware any two points of
B'\ we may view Z and w as elements of the dual space of Hoo(Bn) in the obvi-
ous way. Therefore, for any function f in the unit ball of Hoo(Bn) we may apply
(2.2) to conclude that If(z) - f(w)1 ~ 2p(z,w). Henceforth, applying (2.2) to the
functions f and fo(z) = f(az), we have
I(C",f)(z) - (C,pf)(z) - (Kf)(z)1 ~ If((z)) - f(a(z))1 + If('lj!(z)) - f(a'(z))1
< 2p((z), a(z)) + 2p('lj!(z), a'lj!(z)) < 4c
when z E {II ~ 8} n Hl/)I ~ 8}, while
I(C",f)(z) - (C,pf)(z) - (Kf)(z)1 < If((z)) - f('lj!(z)) I + If(a(z)) - f(a1/l(z))1
< 2p((z), 'lj!(z)) + 2p((z), 'lj!(z))
< 4e + 41:
184 PAMELA GORKIN. RAYMOND MORTINI, AND DANIEL SUAREZ

when Z E {1cf>1 > 8} U {I'!/JI > 8}. Since the function f is arbitrary, IIG</> - G", - KII ~
4e + 4c, and since c is arbitrary we obtain (3.1). D
COROLLARY 7. Let cf> and '!/J be two holomorphic self-maps of the unit ball.
Then G</> - G", is compact if and only if either max {11<;b11, II'!/JII} < 1, or
lim sup p(cf>(z), '!/J(z)) = lim sup p(cf>(z), '!/J(z)) = O.
leI>(z)I~1 1"'(z)I~1

PROOF. It is clear that G</> and G", are compact, if max {11cf>11, 11'!/J11l < 1. On
the other hand, if max {11cf>11, II'!/JII} = 1 and e is the parameter of Theorem 6, then
(3.1) says that Gel> - G", is compact if and only if e = o. D
Our next goal is to characterize the path components of composition operators
on Hoo(Bn). We write G</> '" G", to indicate that there is a norm-continuous
homotopy of composition operators joining Gel> with G",. Also, if K denotes the
ideal of compact operators, we write G</> "'e G", to indicate that there is an essential
norm-continuous homotopy of classes {G", + K: cp: Bn -4 B n holomorphic} joining
Gel> + K with G", + K.
Let cf> be a holomorphic self-map of Bn. For x E M (HOO (Bn)) we can define
cf>(x) E M(Hoo(Bn)) by the rule <;b(x)(f) ~f x(f 0 cf. Thus we can extend cf> :
B n -4Bn to a self-map of M(Hoo(Bn)), which we also denote by cf>. The continuity
of this extension is immediate.
We now have everything we need to prove the main theorem of this paper. As
indicated in the introduction, this theorem unifies and extends many of the results
appearing in [11], as well as [8].
THEOREM 8. Let cf> and '!/J be holomorphic self-maps of the unit ball in en.
Then the following are equivalent.
(a) G</> '" G",.
(b) G</> "'e G",.
(c) IIGeI> - G",II < 2.
(d) SUPzEBn p(cf>(z), '!/J(z)) < 1.
PROOF. (a) => (b) is obvious.
(c) :} (d). A boundary for HOCJ (Bn) is a closed set F c M (HOCJ (Bn)) such that
Ilfll = SUPxEF If(x)1 for all f E HOCJ(Bn). It is clear that the closure B n of B n in
M(HOCJ(Bn)) is a boundary for HOCJ(B n ), and since oHOCJ(Bn) is the intersection of
all the boundaries [4, p. 10], then oHOCJ(Bn) c F. The equivalence then follows
from (2.1).
(b) => (c). By hypothesis there is a family {cf>t}, with t E [0, 1], of holomorphic
self-maps of Bn such that cf>o = cf>, cf>1 = '!/J and for every c > 0 there is some 8 > 0
satisfying
< c: if It - sl < 8.
IIG</>t - Gel>. lie
Then we can take finitely many points ty = 0 < ... < tm = 1 in [0,1] such that
IIG</>t. - GeI>tHl lie < 1/2 for every i = 1, . .. , m - 1. We claim that
(3.2) sup p(cf>t. (z), cf>t'+l (z)) < 1
zEBn

for every i. In fact, if r = max{llcf>d,lIcf>tHlll} < 1, then both functions map


B n into the closure of r Bn, and since the pseudohyperbolic diameter of this ball
HOMOTOPIC COMPOSITION OPERATORS ON Hoo(Bn) 185

is smaller than 1, we are done. If some of the maps have norm 1, then the first
inequality of (3.1) tells us that there is some 0 < 8 < 1 close enough to 1 such that
sup p(t;{Z), tHl (Z)) < 3/4,
{I</>., 1~t5}U{I</>"+11~6}
while the set {It, I < 8} n {It'+ll < 8} is mapped by both functions into the ball
8Bn, whose pseudohyperbolic diameter is smaller than 1. Our claim follows.
Since the closure of Bn in M(Hoo(Bn)) contains the Shilov boundary, (3.2)
and (2.1) imply that IIC</>" - C</>'<+1 11 < 2 for i = 1, ... , Tn - 1. Lemma 3 now says
that (c) holds.
(d) =} (a). By (d) there exists 0: < 1 such that
sup p((z), 'l/J(z)) -:; 0:.
zEBn

We define a map t = + t('l/J - ) for t E [0,1]. Now, if t < s < 1


IIC</>, - C</>. II < 2 sup p(t(z), s(z))
zEBn
2(s - t)p((z), 'l/J(z))
<
1 - (1 - (s - t)) p((z), 'l/J(z))
20:
< (s - t) 1 _ 0:'
where the first inequality holds by (2.1), since 8Hoo(Bn) C 13", and the second
inequality from (2.3). From this, we see that t 1--+ C</>t is a continuous mapping.
Therefore C'" lies in the same path component as C</>. 0
As a corollary, we obtain the following generalization of the work on isolated
points in [11].
COROLLARY 9. Let be a holomorphic self-map of the ball. Then C</> is isolated
in the set of composition operators if and only if C</> is essentially isolated.
PROOF. Since IIC</> - C",II ~ IIC</> - C",lIe, it is clear that if C</> is essentially
isolated, then it is isolated. If C</> is isolated and 'l/J =I- , Theorem 8 implies that
SUPzEBn p((z),'(z)) = 1. This can only happen if there are points z E Bn such
that 1(z)I--+1 or I'l/J(z) 1--+1, and p((z), 'l/J(z))--+1. Hence, Theorem 6 says that
IIC</> - C'" lie ~ 1, and C</> is essentially isolated. 0

4. Examples
So what are some examples of isolated operators? If : Bn --+ Bn has radial
limits of Euclidean norm 1 on a set of positive measure, we claim that C</> is isolated.
If 'l/J =I- there must exist a set of positive measure in 8Bn on which has
radial limits of norm 1 and 'l/J does not equal (see [13, Ch. 5]). Thus, there
exists a sequence {zd c Bn for which (Zk) --+ ( E 8Bn and 'l/J(Zk) --+ 'Tf, with
'Tf =I- (. Therefore p((Zk), 'l/J(Zk)) --+ 1 and Lemma 2 tells us that IIC</> - C",II = 2.
In particular, the automorphisms of B n induce isolated composition operators.
It is clear from Theorem 8 that if , 'l/J are holomorphic self-maps of Bn and
O</> - C'" is compact, then C</> '" C"'. It is not completely clear, though, that the
converse fails. In [8] Hosokawa, Izuchi and Zheng constructed an example that
shows this for n = 1. By eliminating variables, every example that works for
n = 1 can be made to work for general n. Here we construct a simpler example of
186 PAMELA GORKIN, RAYMOND MORTINI, AND DANIEL SUAREZ

holomorphic self-maps of the ball, 4> and '1/), such that C'" '" C'" but C<I> - C,,) is not
compact.
Let n = {UJ ED: ~1-=-!:1 > ~} be a nontangential region in D at the point
Z = 1. We want to estimate p(w, (w + 1)/2) for wEn. We recall that p(z, w) =
Iz - wi/II - zwl for z, wED. By straightforward calculation,
( ( + 1)/2) -1 = 11 -lwl 2+ 1- wi (1 -lwl)(1 + Iwl) 1 3
p w, w Il-w I ~ Il-w I + ~
and

p(w, (w + 1)/2)-1 11+ 1l-w


- Iwl21

> ~ (1 + 1- Iw12)
1- w
1-lwl 2
1+ 11_wI2(1-~w)

> 1+ 1+ Iwl > ~


2 - 2
when wEn. That is,
1 2
(4.1) 3 ~ p(w, (w + 1)/2) ~ 3
for all wEn. Let c.p : D---+n be a one-to-one and onto holomorphic function and
define 4>, 'I/J : Bn---+Bn by
4>(Z1,"" zn) = (c.p(Z1), 0, ... ,0) and 'I/J(Z1, ... , zn) = ((c.p(zt) + 1)/2,0, ... ,0).
It is clear that 114>11 = 11'1/)11 = 1. For z E Bn, a straightforward calculation shows
that
p((z), 'I/J(z)) = p(c.p(zd, (c.p(zd + 1)/2).
Since c.p(zd E n, the inequalities in (4.1) show that p(4)(z), 'I/J(z)) E [1/3,2/3].
Therefore Theorem 8 says that C'" '" C"', while Corollary 7 says that C<I> - C'" is
not compact..

5. Endomorphisms of HOO(D)
In this section we investigate the path components of elldomorphisms of H OO (D).
For x E M(HOO(D)), the Gleason part of x is P(x) = {y E M(HOO(D)) : p(x,y) <
I}. Since the condition p(:r, y) < 1 is an equivalence relation, the Gleason parts form
a partition of M(HOO(D)). In [7] Hoffman produced a continuous and onto map
Lx: D---+P(x) such that Lx(O) = x and foLx E HOO(D) for every x E M(HOO(D))
and f E HOO(D). There are two possibilities: either Lx(z) = x for all zED (so
P(x) = {x}) or Lx is one-to-one. We write
G = {x E M(HOO) : Lx is one-to-one} and r = {x E M(HOO) : Lx = {x}}.
It is well-known that every endomorphism T of HOO(D) can be factored as T =
C",CL", , where 4> is a holomorphic self-map of D and x E M(HOO(D)). Although it
is clear that this factorization is not unique, two different factorizations of the same
endomorphism are related in the following way (see [2]): if p(x, y) < 1, then there
is a biholomorphic map r of D (depending on x and y) such that Ly(z) = Lx(r(z))
HOMOTOPIC COMPOSITION OPERATORS ON Hoo(B n ) 187

for every zED. This means that every endomorphism of the form T = Cq,CLy can
also be factored as
T = Cq,CLy = Cq,CTCL", = CToq,CLx '
Of course, if x E rand p(x, y) < 1, then x = y and T = C Lr .

LEMMA 10. Let x E G and A = Ej=1 )..,jCq,j' where )..,j E C and Cq,j are
composition operators on H'XJ(D). Then IIACLJI = IIAII.
PROOF. Since IIACLxll ::; IICLxllliAIl ::; IIAII, one direction is easy. For the
other direction, if 0 < f < 1, there exists a function f in the ball of HOO(D) such
that IIA(f)11 > (1 - f)IIAIi. By the definition of the norm, there exists r with
o < r < 1 such that
n
sup IA(f)(z)1 = sup
ZETD
L )..,jf(cPj(z))
ZETD j=l
> (1 - f)2I1AII

By a result of Hoffman [7, p. 91]' there exist Blaschke products bk such that (bk 0
Lx)(z) ~ z uniformly on compact subsets of D. But rD is a precompact subset
of D, and therefore cPj (r D) is precompact for each j. That is, there is 0 < 0: < 1
such that Uj=l cP j(rD) C o:D. Fix {J with 0: < {J < 1. Since f is analytic, there is
8> 0 such that for z, wE (JD with Iz - wi < 8 we have If(z) - f(w)1 < f. Clearly
we can also require 8 < (J - 0:. Therefore we may choose k sufficiently large so that
I(b k 0 Lx)(cPj(z)) - cPj(z)1 < 8 for all z E rD. Thus, for k that large, z E rD and f
as above, (b k o Lx)(cPj(z)) E (3D and consequently If(bk(Lx(cPj(z))) - f(cPj(z))1 < f.
Therefore there exists a constant At depending only on nand )..,l"",)..,n such that
II ACLx II ~ sup I(ACLx(f 0 bk))(z)1
zErD
n
sup IL )..,j(f 0 bk 0 Lx)(cPj(z))1
zErD j=l
n
> sup IL )..,jf(cPj(z))I- Mf
zErD j=l
> (1 - f)2I1AII- Mf.
Letting f ~ 0 yields the desired result. o
THEOREM 11. Let T 1, T2 E End(HOO(D)). Then the following ar'e equivalent.
(a) T1 rv T2 in End(HOO(D)).
(b) IIT1 - T211 < 2.
(c) There exist x E M(HOO(D)) and holomorphic self-maps cP. 'ljJ of D such
that T1 = Cq,C Lx T2 = CtfJCLx and IICq, - CtfJlI < 2.
PROOF. Suppose that (a) holds. Then there is a homotopy
G: [0, 1] ~End(HOO(D))
with G(O) = T1 and G(I) = T 2. We can find finitely many points 0 = it < ... <
tn = 1 such that IIG(tj) - G(tj+d II < 2 for j = 1, ... , n - 1. Lemma 3 then says
that IIG(O) - G(I)1I < 2.
Suppose that (b) holds and write T1 = CLxoq, and T2 = CLyo"" where x,y E
M(HOO(D)) and <p, r.p : D ~ Dare holomorphic. If p(x, y) = 1, we then have that
188 PAMELA GORKIN, RAYMOND MORTINI, AND DANIEL SUAREZ

p(Lx((O)), Ly(<p(O))) = 1, and (2.1) tells us that IICL"ot/> - CLyo'!'lI = 2. Thus


(b) implies that p(x,y) < 1. If x E r, then Tl = T2 = C Lx ' If x E G and we
write T2 = CLxo.p, where 1/J is a holomorphic self-map of D. Lemma 10 says that
2> IITI - T211 = IICt/> - C,pll, so (c) holds.
If (c) holds Theorem 8 says that there is a homotopy of composition operators
F(t), with t E [0,1] such that F(O) = Ct/> and F(l) = C,p. By Lemma 10, G(t) ~f
F(t)CLx is a homotopy of endomorphisms connecting Tl with T 2, which proves
(a). 0
Acknowledgement. The last author thanks Bucknell University for its hos-
pitality and peaceful environment during the preparation of this paper.

References
[1] E. Berkson, Composition opemtors isolated in the uniform opemtor topology, Proc. Amer.
Math. Soc. 81 (1981),230-232.
[2] P. Budde, Support sets and Gleason parts, Michigan Math. J. 37 (1990), 367-383.
[3] P. Galindo and M. Lindstrom, Factorization of homomorphisms through HOO(D), preprint.
[4] T. Gamelin, Uniform algebras, Prentice-Hall, Inc., Englewood Cliffs, N. J., 1969.
[5] T. Gamelin and J. Garnett, Distinguished homomorphisms and fiber algebras Amer. J. Math.
92 (1970), 455-474.
[6] P. Gorkin and R. Mortini, Asymptotically interpolating sequences in uniform algebms, J.
London Math. Soc., to appear.
[7] K. Hoffman, Bounded analytic functions and Gleason parts, Ann. of Math. (2) 86 (1967),
74-111 .
[8] T. Hosokawa, K. Izuchi, and D. Zheng, Isolated points and essential components of compo-
sition opemtors on HOC, Proc. Amer. Math. Soc. 130 (2002), no. 6, 1765-1773.
[9] U. Klein, Kompakte multiplikative Opemtoren auf uniformen Algebren, Mitt. Math. Sem.
Giessen 232 (1997), 1-120.
[10] B. MacCluer, Components in the space of composition opemtors, Integr. Equ. Oper. Theory
12 (1989), 725-738.
[11] B. MacCluer, S. Ohno, and R. Zhao, Topological structure of the space of composition oper'-
ators on HOC, Integr. Equ. Oper. Theory 40 (2001), 481-494.
[12] J. Moorhouse and C. Toews, Differences of composition opemtors, preprint.
[13] W. Rudin, Function theory in the unit ball of en, Grundlehren der Mathematischen
Wissenschaften [Fundamental Principles of Mathematical Science], 241. Springer-Verlag, New
York-Berlin, 1980.
[14] J. Shapiro and C. Sundberg, Isolation amongst the composition opemtors, Pacific J. Math.,
145 (1990), 117-151.
[15] C. Toews, Topological components of the set of composition opemtors on HOC (B N ), preprint.

DEPARTMENT OF MATHEMATICS, BUCKNELL UNIVERSITY, LEWISBURG, PENNSYLVANIA 17837


E-mail address: pgorkinlDbucknell. edu

DEPARTEMENT DE MATHEl\fATIQUES, UNIVERSITE DE METZ, ILE DU SAULCY, F-57045 METZ,


FRANCE
E-mail address:mortinilDponcelet.univ-metz.fr

DEPARTA~IENTO DE MATEMATICA, UNIVERSIDAD DE BUENOS AIRES, PAB. I, CIUDAD UNIVER-


SITARIA, (1428) NUNEZ, CAPITAL FEDERAL, ARGENTINA
E-mail address:dsuarezlDdm.uba.ar
Contemporary Mathematics
Volume 328, 2003

Characterization of conditional expectation in terms of


positive projections

J.J. Grobler and M. De Kock

ABSTRACT. A description of positive, order continuous projections in ideals of


measurable functions is given in terms of conditional expectation-type opera-
tors. The dual of such an operator can also be represented as a conditional
expectation operator. We use this result to characterize conditional expecta-
tion in terms of a positive, order continuous projection that preserves one, and
such that an extension of the dual also preserves one.

1. Introduction
Dodds, Huijsmans and De Pagter (see [2]) give a complete description of positive
projections in ideals of measurable functions in terms of conditional expectation-
type operators. Let E be an ideal of measurable functions such that Loo(O, E, Jl) C
E C L1 (0, E, Jl). They characterize a positive, order continuous projection T : E -+
E with the property that T and the operator dual T' preserves one, in terms of
conditional expectation.
We extend this result to an ideal of measurable functions L with the prop-
erty that Loo(O, E, Jl) C L, but we omit the assumption L C Ll (0, E, Jl). In this
case we prove that a positive, order continuous projection S : L -+ L, which is one-
preserving and with the property that an extension of its dual is also one-preserving,
can be characterized in terms of the conditional expectation. In order to prove this
extension, we need the fact that if an operator can be characterized in terms of con-
ditional expectation, then its dual can also be characterized in terms of conditional
expectation.

2. Preliminaries
Let (0, E, Jl) be a a-finite measure space. The vector lattice of all Jl-a.e. finite
E-measurable functions on 0, with the usual identification of Jl-a.e. equal functions,
is denoted by Lo(O, E, Jl). M+(O, E, Jl) denotes the set of all (equivalence classes
of) real, positive, Jl-measurable functions into [0,00]. We define x Vy := sup{x, y}.
A linear subspace G of a vector lattice E is called a sublattice if x Vy belongs to G
for all x, y E G. A linear subspace A in which Ixl :::; IYI, with yEA implies x E A,
is called an ideal.
A subset AcE is order bounded if A is contained in an order interval, i.e. A is
bounded from above and below. We denote the set of order bounded linear operators
2003 American Mathematical Society
lRO
190 J.J. GROBLER AND M. DE KOCK

from the vector lattice E into itself by Cb(E). Cb(E) is a Dedekind complete vector
lattice (see [3), p3). All vector lattices considered will be Dedekind complete.
A net (xQ)QEf in E is called order convergent if there exists a net (YQ)QEr
satisfying (YQ) 1 0 and Ix - x", I ~ y", for all Q E r, where r is an index set. We
write X ---> x in order.
Q

If Eo is a sub-a-algebra of E, then we denote the restriction of p, to Eo by


IL again. Et denotes the collection of subsets of Eo of positive measure. The
characte1"istic function of a set A E E is denoted by lA. We write 1 rather than
In. For any measurable function Jon 0, the support of f is denoted by supp(J) =
{w EO: f(w) =1= o}.
Our attention will, to a great extent, be focused on ideals of measurable func-
tions on (0, E, I.L), i.e., on ideals L in the vector lattice Lo(O, E, p,). The set Z E E
is called an L-zero set if every .f E L vanishes p,-a.e. on Z. There exists (modulo
IL-llUll sets) a maximal L-zero set ZI in E and the set 0 1 = 0 \ ZI is called the
carrier of the ideal L. There exists a sequence An i 0 1 in E such that 1.L(An) < 00
and IAn E L for all n E N, (see [5], p143). Clearly, the carrier of L is equal to n, if
and only if L is order dense in Lo(n, E, p,).
Let L c Lo(O, E, p,) be an order dense ideal with order continuous dual L~. We
identi(y L~ with an ideal L' of functions in Lo(n,E,p,), and we will assume that
L' is again an order dense ideal (which is always the case if L is a Banach function
space; (see [5), Theorem 112.1). Equivalent to this assumption is that L ~ separates
the points of L. The duality relation between Land L' is given by (J, g) = Inf gd,.L
for f ELand gEL' (see [5), Section 86). Let 8 E Cb(L) with L an ideal of
functions in Lo(O, E, p,). We define its order continuous adjoint 8' : L' ---> L' by
(g, 8' J) = (8g, J) for all f E L' and gEL (see [5], Section 97). Then 8' E Cb(L').
If there is no reason for confusion, we will denote (n, E, p,) by (E) only.
DEFINITION 1. Let E and F be vector lattices and let T : E ---> F be a linear
operator. Then
(i) T is positive (denoted by T ~ 0) whenever Tx ~ 0 for all x ~ 0; T is called
strictly positive (denoted by T 0) if Tx > 0 for all x > O.
(ii) T is order continuous whenever Txc< -+ 0 in order for every net (x Q )

satisfying X -+ 0 in order.
Q

(iii) T is order bounded if it maps order bounded subsets into order bounded
subsets.
For a Banach function space (E, II . liE) defined on some finite measure space
(n, E, p,) for which Loo(E, p,)
~ E ~ Ll (E, I.L), we define the following (see [2), p69).

DEFINITION 2. (i) The linear map T : E -+ E is called averaging if for


all f E Loo(E) and all gEE we have that T(JTg) = Tf Tg.
(ii) T: E -+ E is called contractive if II T II ~ 1.
DEFINITION 3. Let (n, E, p,) be a probability space (i.e. 1.L(n) = 1) and let Eo
be a sub-a-algebra of E. For fELl (E), we denote by lFP(J I Eo) the IL-a.e. unique
Eo-measurable function with the property that

i lFf'(J I Eo)dl.L = i fdp,


for all A E Eo. The function lFP(J I Eo) is called the conditional expectation of f
with respect to Eo.
If there is no reason for confusion, we will denote the p,-a.e. Eo-measurable
function lElL('IE o) by lE('IE o) only. The existence of lE(J I Eo) is a consequence of
CHARACTERIZATION OF CONDITIONAL EXPECTATION. 191

the Radon-Nikodym theorem. The conditional expectationlE('IEo) can be extended


from a mapping from Ll (E) into itself, to a mapping from M+(E) into itself. If
f E M+(E,), then 1E(f I Eo) E .l\J+(E) is defined by 1E(f I Eo) = suplE(fn I Eo),
where 0 :::; fn E Ll (E) (71. = 1,2, ... ) satisfy 0 :::; fn i f J,L-a.e.

The conditional expectation operator has the following properties. For a proof
of properties (i) to (vi) we refer to [4], p7; for property (vii) we refer to [3], p7.

PROPOSITION 1. (i) lE(o:f + /1g I Eo) = 0:1E(f I Eo) + /11E(g I Eo) for all
f,g E M+(E) and for all 0:::; 0:,/1 E R
(ii) 0 :::; f :::; 9 in M+(E) implies that 0 :::; 1E(f I Eo) :::; lE(g I Eo) and if
1E(lfll Eo) = 0, then it follows that f = O. By virtue of positivity we
have 11E(f I Eo)1 :::; 1E(lfll Eo).
(iii) 0:::; fn i f IJ,-a.e. implies that 0:::; lE(fn I Eo) i 1E(f I Eo) J,L-a.e.
(iv) lE(gf I Eo) = glE(f I Eo) for all f E M+(E) and all 9 E M+(Eo).
(v) If 9 E M+(Eo) and f E M+(E), then fA gdJ,L = J~ fdJ,L for all A E Eo if
and only if 9 = 1E(f I Eo) IJ,-a.e.
(vi) If Eo c Ao are sub-a-algebras of E, then 1E(f I Eo) = 1E(1E(f I Ao) I Eo) for
all 0:::; f E M+(E).
(vii) If f E M+(E) is such that 1E(f I Eo) E Lo(E), then we also have that
f E Lo(E).
DEFINITION 4. The domain domlE('IE o) of 1E('IEo) is defined by
domlE('IE o) := {f E Lo(E) : 1E(lfll Eo) E Lo(Eo)}.
Clearly, domlE(IE o) is an ideal in Lo(E) which contains L 1 (E).
For f E dom 1E(1 Eo), we define:
1E(f I Eo) := 1E(f+ I Eo) -1E(r I Eo).
This defines a positive linear operator
1E('IEo) : domlE('IE o) --+ Lo(Eo) C Lo(E).
Let (n, E, l.l) be a probability space and let L c Lo(n, E, J,L) be an ideal with
carrier n. Set
M(L) = {m E Lo(E) : 1E(lmfll Eo) E L V f E L}.
Since L C Lo(E), we have that mf E domlE('IE o) for all m E M(L) and f E L.
For m E M(L) we define Smf : L --+ L by
Smf := lE(mf I Eo) V f E L.
Sm is order bounded and ISml :::; Simi' Sm is also order continuous.
The following proposition will be applied in the sequel. A proof can be found
in [3], (p8).
PROPOSITION 2. Let (n, E, J,L) be a probability space and Eo C E a sub-a-
algebra.
(i) If f E domlE('IE o) and 9 E Lo(Eo), then it follows that gf E domlE('IE o)
and lE(gf I Eo) = glE(f I Eo).
(ii) If f E Lo(E), then f E domlE('IE o) if and only if there exists a sequence
{A n }:'=1 in Eo such that An in and .

{ IfldJ,L < 00 V 71. = 1,2, ....


JAn
192 J.J. GROBLER AND M. DE KOCK

Moreover, if f E domlE(IE o), then, for all A E Eo with fA Ifldft < 00,

i lEU I Eo)dft = i fdft

The following lemma will be applied in t.he sequel.


LEMMA 3. For a linear subspace N of Loo(E) the following statements are
equivalent.
(i) There exists a sub-a-algebra Eo such that N = Loo(Eo).
(ii) N is a subalgebra of Loo(E) containing the constants such that fn E
N, Ifni::; u E Loo(E) (11. = 1,2, ... ) and fn - t f a.e. imply that fEN.
The proofs of the following propositions and corollaries rely mainly on t.he proofs
by Dodds, Huijsmans and De Pagter (see [2]).
Let L c Lo(E) be an ideal of measurable functions such that Loo C L. We then
have the following.
PROPOSITION 4. Let S : L -t L be an or'der continuous, positive linear opemtor
for which
(i) Sf E Loo(E) whenever f E Loo(E),
(ii) SUSg) = Sf Sg for' all f E Loo(E) and all gEL.
Then there exists a sub-a-algebra Eo of E and there exists a 0 ::; m E M(L) such
that Sf = lE(mf I Eo) for all f E L.
We use the following proposition in the proof of the main result.
PROPOSITION 5. For a linear operator S : L -t L, the following statements are
equivalent.
(i) S is positive and or'der continuous, S2 = S, SI = 1 and the range R(S)
of S is a sublattice.
(ii) There exist a sub-a-algebra Eo of E and a function 0 ::; m E M(L) with
lE(mIEo) = 1 such that Sf = lE(mf I Eo) for all f E L.
Because the range of a strictly positive projection is a sublattice, we obtain the
following result.
COROLLARY 6. For a linear operator S : L -t L the following statements are
equivalent.
(i) S is a strictly positive, order continuous project'ion with SI = 1.
(ii) There exists a sub-a-algebra Eo of E and a strictly positive function m E
M(L) with lE(ml Eo) = 1 such that Sf = lE(mfl Eo) for all f E L.
In the following proposition we consider the case where the operator no more
preserves onc, but where the image of the indicator function is strictly positive. We
derive a similar result as in Corollary 6 for S strictly positive.
PROPOSITION 7. Let S : L -t L be a linear operator, then the following state-
ments are equivalent.
(i) S is a positive or'der continuous projection onto a sublattice such that SI
is strictly posit'ive.
(ii) There exist a s'nb-a-algebra Eo ofE, 0 ::; m E Lo(E) and a strictly positive
function k E Lo(E) with lE(mk I Eo) = 1, such that Sf = klE(mf I Eo) for
all f E L.
CHARACTERIZATION OF CONDITIONAL EXPECTATION. 193

We have the following basic characterization of conditional expectation on


L1 (E). A proof can be found in [2], p71.
PROPOSITION 8. (Douglas R.G. and Seever) If T is a continuous linear map
on L 1(E), then the following statements are equivalent.
(i) There exists a sub-a-algebra Eo of E such that for all f E L1 (E) we have
that Tf = lEU I Eo).
(ii) T is a contractive projection which preserves 1.

3. Main characterization of conditional expectation.


We prove that if an operator can be characterized in terms of the conditional
expectation, then its dual can also be characterized in terms of the conditional
expectation. As before, we let L c Lo(E) be an ideal of measurable functions
which contains Loo(E).
LEMMA 9. If S : L --+ L is a linear operator such that Sf = lEU IEo) for all
f E L, then S' : L' --+ L' satisfies S' 09 = lE(g I Eo) for all E-measurable 09 E L'.
ProoF. Let f ELand gEL'. Then
(j, S'g) (Sf,g)

l SfgdJ.l

l !l
lE iL U I Eo)gdJ.l

l n
lEiL(glEiLU I Eo) I Eo)dJ.l

llEiLU I Eo) 1E1L(g I Eo)dJ.l

l f S'lE iL (g I Eo )dJ.l

Thus, we have proved that


(3.1) (j,S'g) = (j,S'IE(gIE o )).
For any Eo-measurable g, we have that

(3.2)

It follows from (3.2) that for Eo-measurable 9 we have that S'g = 9 and from (3.1)
and (3.2), for arbitrary gEL' that S'g = S'lE(g I Eo) = lE(g I Eo). 0

Now we are able to prove the main result, where we characterize conditional
expectation in terms of a positive, order continuous projection and an extension of
its dual.
PROPOSITION 10. If S is a linear map, then the following statements are equi-
valent.
(i) Ther'e exists a sub-a-algebra Eo of E such that Sf = lEU I Eo) for' all
f E L.
(ii) S: L --+ L is a positive order continuous projection such that SI = 1 and
S' has an extention S' : L' + Loo --+ L' + Loo satisfying S'I = 1.
194 J.J. GROBLER AND M. DE KOCK

Proof. (i) =} (ii) It follows from Proposition 5 that S is a positive, order


continuous projection such that SI = 1. Since the conditional expectation operator
is defined on L1 (E) into L1 (E), it follows that
S : L n L 1(E) -4 L n L 1(E).
Denote the restriction of S to L n L1 (E) by 8. Then
8' : L' + Loo(E) -4 L' + Loo(E)
and 8' is an extension of S' : L' -4 L'. Because for f E L' and gEL n L1 (E), we
have
(3.3) (g, 8' J) = (8g, J) = (Sg, J) = (g, s' J).
Since L n L1 (E) is dense in L, (3.3) holds for all gEL, and so 8' f = S' f for all
f E L'. For all gEL' + Loo(E), it follows from Lemma 9 that 8'g = lE(g 1 Eo), so
8'1 = 1, by the properties of conditional expectation.
(ii) =} (i) We first note that since Loo (E) c L we have L' C L1 (E) and also
that L' is dense in L 1 (E). An argument of Ando (see [1], (p401)) shows that S' is
contractive for the L1 (E)-norm. In fact, if gEL', it follows from the assumption
SI = 1 that,

In IS' gld{t In S' 9 sgn S' gd{t

< In IgIIS(1 sgnS'g)ld{t

< InlgIS(lsgnS'gl)dJL

< In IglSld{t

In Igld{t.
Since L' is dense in L1 (E) for the L1 (E)-norm, we can extend S' to a contraction
on LdE). Since S is a projection, the same holds for S'. Thus, by Proposition 8,
there exists a sub-a-algebra Eo of E such that S' 9 = lE(g 1 Eo) for every 9 E Ll (E)
and therefore also for all gEL'. It follows from Lemma 9 that S" f = lEU 1Eo) for
all f E L". By restricting S" to L, we therefore have that Sf = lEU 1 Eo) for all
f E L. 0

References
[1] ANDO, T., 1966, Contmctive projections in Lp-spaces, Pacific J. Math., 17,391-405.
[2] DODDS, P.G., HUIJSMANS, C.B., and DE PAGTER, B., 1990, Chamcterizations of Con-
ditional Expectation type-opemtors, Pacific J. Math., 141,55-76.
[3] GROBLER, J.J. and DE PAGTER, B., 1999, Opemtors representable as Multiplication Condi-
tional Expectation opemtors, To appear in J. of Operator Theory.
[4] NEVEU, J., 1975, Discrete-pammeter martingales, North Holland/American Elsevier, Am-
sterdam Oxford New York.
[5] ZAANEN, A.C., 1982, Riesz spaces II, North Holland, Amsterdam, New York.

SCHOOL FOR BUSINESS MATHEMATICS, POTCHEFSTROOM UNIVERSITY FOR CHE, POTCHEF-


STROOM 2520, SOUTH AFRICA" MATHEMATICS DEPARTMENT, KENT STATE UNIVERSITY, KENT,
OH 44240
E-mail address:srsjjgClpuknet.puk.ac.za mdekockClmath. kent. edu
Contemporary Mathematics
Volume 328, 2003

The Krull nature of locally C* -algebras

Marina Haralampidou

ABSTRACT. Any complete locally m-convex algebra, whose normed factors in


its Arens-Michael decomposition are Krull algebras is also Krull. In particular,
any locally CO-algebra is a Krull algebra. Considering perfect projective sys-
tems, we give another proof of the fact that any Frechet locally CO-algebra is
a Krull algebra. Furthermore, a proper complete locally m-convex H* -algebra
with continuous involution and a normal unit is a locally C* -algebra, hence
Krull. The class of Krull (topological) algebras is closed with respect to carte-
sian products, topological algebra isomorphic images, and perfect projective
limits.

1. Introduction and preliminaries


Every closed (left) ideal of a CO-algebra E is the intersection of the (closed)
maximal regular (left) ideals containing it (see, for instance, [2: p. 56, Theorem
2.9.5]. Thus, E is a Krull algebra in the sense of Definition 1.1. A natural question
arises here whether, in general, any locally CO-algebra is Krull. In that direction,
using the Arens-Michael decomposition, we get that a complete locally m-convex
algebra (E,(Po)oEA) is a Krull algebra, if each factor Eo. = E/ker(po.)' Q E A
is a Krull algebra (Proposition 2.1). As a consequence, we get that any locally
CO-algebra is Krull (Corollary 2.2). Besides, a proper complete locally m-convex
H*-algebra (E, (Po.)o.EA) with continuous involution and a unit element e, so that
po.(e) = 1 for every Q E A, is a Krull algebra (Corollary 2.6). Based on the notion
of a perfect projective system (Definition 2.7), we provide another proof of the fact
that any Fn3chet locally C* -algebra is Krull (Theorem 2.10).
By the term topological algebm we mean an algebra, which is a topological
linear space such that the ring multiplication is separately continuous (see [11: p.
4, Definition 1.1] and/or [12: p. 6]). A topological algebra E is called a Q'-algebm,
if every maximal regular left or right ideal in E is closed (see [5: p. 148, Definition
1.1]).
A locally m-convex algebm is a topological algebra E whose topology is defined
by a family (Po.)o.EA of submultiplicative seminorms, i.e. Po.(xy) :::; Po.(x)Po.(y) for

1991 Mathematics Subject Classification. Primary 46H05, 46H10, 46H20.


Key words and phrases. Krull algebra, Q'-algebra, Arens-Michael decomposition, locally C*-
algebra, perfect projective system of topological algebras, perfect projective limit algebra, Frechet
locally CO-algebra, proper algebra, locally m-convex H*-algebra.

2003 American Mathematical Society


195
196 MARINA HARALAMPIDOU

all x, y E E, 0: E A (see for instance [11] and/or [12]). Such a topological algebra
is denoted by (E, (PoJ"'EA)' A complete metrizable locally m-convex algebra E is
called a Prechet locally m-convex algebra. In this case, the topology of E is defined
by a countable family (Pn)nEN of submultiplicative seminorms.
A C*-seminorm is a seminorm P on an involutive algebra E, satisfying the
C*-condition, namely, p(x*x) = p(x)2 for every x E E [13: p. 1, Definition 1].
Such a seminorm is submultiplicative and *-preserving [ibid. p. 2, Theorem 2].
A locally pre-C* -algebra is an involutive locally (m-) convex algebra (E, (P",)",EA),
such that each p"" 0: E A is a C* -seminorm, while a complete algebra, as before, is
called a locally C* -algebra [8: p. 198, Definition 2.2]. A Frechet locally C* -algebra
is an involutive Frechet locally (m-) convex algebra (E, (Pn)nEN) where each Pn is
a C* -seminorm.
A locally m-convex H* -algebra is an algebra E equipped with a family (P"')"'EA
of Ambrose seminorms in the sense that P"" 0: E A arises from a positive pseudo-
inner product <, >"" such that the induced topology makes E into a locally m-
convex topological algebra. Moreover, the following conditions are satisfied:
For any x E E, there is an x* E E, such that
< xy,z >",=< y,x*z >",
< yx, z >",=< y, zx* >",
for any y, z E E and 0: E A. x* is not necessarily unique. In case, E is proper
(viz. Ex = (0), implies x = 0), then x* is unique and * : E -+ E : x f-+ x* is an
involution (see [4: p. 451, Definition 1.1 and p. 452, Theorem 1.3]).
Throughout of this work the considered algebras are over the field of complexes.
To fix notation we recall the following. Let (E, (P"')"'EA) be a complete locally
m-convex algebra and
P'" : E -+ E/ker(p",) == E", : x f-+ p"'(x) :=x + ker(p",)
(1.1)

the respective quotient maps. Then Ilx",ll", := p"'(x), x E E, 0: E A defines on E",


an algebra norm, so that E", is a normed algebra and the morphisms P"" 0: E A are
continuous. E"" 0: E A denotes the completion of E", (with respect to II . II",). A
is endowed with a partial order by putting 0: :::; /3 if and only if p"'(x) :::; P(3(x) for
every x E E. Thus, ker(p(3) <;;;; ker(p",) and hence the continuous (onto) morphism
(1.2) j",(3 : E(3 -+ E", : x(3 f-+ j",(3(x(3) := x"" 0::::; /3
is defined. Moreover, j ",(3 is extended to a continuous morphism
!",(3 : E(3 -+ E"" 0::::; /3.
Thus, (E"" j",(3), (E",'/",(3), 0:, /3 E A with 0: :::; /3 are projective systems of normed
(resp. Banach) algebras, so that
(1.3) E 2:!
- -
lim E", 2:! lim E", (Arens-Michael decomposition)
within topological algebra isomorphisms (cf., for instance, [11: p. 88, Theorem 3.1
and p. 90, Definition 3.1] and/or [12: p. 20, Theorem 5.1]).
Concerning the following notion see [7].
DEFINITION 1.1. A topological algebra is called a Krull algebra, if every proper
closed left (resp. right) ideal is contained in a closed maximal regular left (resp.
right) ideal.
THE KRULL NATURE OF LOCALLY C'-ALGEBRAS 197

For the statements (i) and (iii) in the next proposition see [7: Lemma 3.8).
PROPOSITION 1.2. Let E, F be topological algebms and : E ---7 F a continuous
epimorphism. Then the following hold true:
(i) If E is a Krull algebm and closed, then F is a Krull algebm.
(ii) If E is a Krull algebm and F a Q'-algebm, then F is a Krull algebm.
(iii) If F is a Krull algebm and closed with ker() ~ I for every proper closed
left or right ideal in E, then E is a Krull algebm.
(iv) If F is a Krull algebm and a closed injection, then E is a Krull algebm.
PROOF. (ii) For a proper closed left ideal J in F, -l(J) is a proper closed
left ideal in E with ker() ~ -l(J) and J = (-l(J)) (see also [3: p. 316,
Proposition B.5.4)). Thus, -l(J) ~ M for some closed maximal regular left ideal
Min E. Hence J ~ (M), so that (M) is a maximal regular left ideal in F (ibid.),
closed by Q'. Similarly, for proper closed right ideals.
(iv) Immediate from (iii). 0

COROLLARY 1.3. A topological algebm is a Krull algebm if and only if a topo-


logical algebm isomorphic image of it is so.

PROPOSITION 104. Let (E"')"'EA be a family of topological algebms and


F = II"'EA E", the respective cartesian product topological algebm. Then F is a
Krull algebm, if each E"" Q E A is a Krull algebm. The converse is true in case
the factors are Q' -algebms.

PROOF. Consider the canonical continuous epimorphisms (projections)


(1.4) 11"", : F ---7 E", : x = (X"')"'EA 1--+ 1I"",(x) := x"" Q E A.
Let I be a proper closed left ideal in F. Since the multiplication is separately con-
tinuous, it follows that the closure 11"",(1) of the left ideal 11"", (I), is a closed left ideal
in E",. Moreover, 11"",(1) i- E", for some Q E A. Otherwise, II"'EA 11"",(1) = F. It is
easily seen that n"'EA 11";;1 (E",) = ILEA E",. Besides, I = I = n"'EA 11";;1(11"",(1)).
Hence I = II"'EA E", = F, a contradiction. Now, since E", is a Krull algebra,
11"",(1) ~ M, for some closed maximal regular left ideal M. Hence I ~ 1I";;1(M) with
1I";;1(M) a closed maximal regular left ideal in F. Similarly, for proper closed right
ideals. The above argument shows that F is a Krull algebra.
For the rest of the assertion apply (ii) of Proposition 1.2. 0

2. The Krull property for locally C*-algebras


We provide first the following result akin to that of Proposition 104.
PROPOSITION 2.1. Let (E, (P"')"'EA) be a complete locally Tn-convex algebm,
such that the norrned algebms E"" Q E A in its Arens-Michael decomposition, are
Krull algebms. Then E is a Krull algebm, as well.
On the other hand, if E is a Krull algebm, then a factor E", is a Krull algebm
if it is also a Q' -algebm.

PROOF. E ~ lim f--


E", within a topological algebra isomorphism, say (see
(1.3)). Consider the continuous epimorphic image p",(I) (see (1.1)) of a proper
closed left ideal I in E. Claim that the closed left ideal p",(I) is proper in E", for
198 MARINA HARALAMPIDOU

some Q E A. Suppose the contrary. Then based on M. Exarchakos, concerning the


first equality of the next rels, we get
(E) = ll!!!Ea = ll!!!Pa(I) = ll!!!(fa((I))) = (I),
here fa denotes the restriction to ll!!! Eo of the projection map 7ra : TIaEA Eo
-+ Eo, Q E A (cf. also [11: p. 87, Lemma 3.2 and p. 89, (3.24)]; we note that (I)
is a closed left ideal in lim
t-
Eo). Thus, E = I, which is a contradiction. So, since
Ea, Q E A is a Krull algebra, it follows that Pa(I) ~ M for some closed maximal
regular left ideal M, and hence I ~ p~l(Pa(I)) ~ p~l(M). Besides, p~l(M) is a
maximal regular left ideal in E, closed by the continuity of Po. An analogous result
holds for proper closed right ideals.
The last part of the assertion follows from (ii) of Proposition 1.2. 0
By [1: p. 32, Theorem 2.4], the factor normed algebras, in the Arens-Michael
decomposition of a locally C*-algebra, are C*-algebras and hence Krull (see, for
instance [2: p. 56, Theorem 2.9.5]). Thus, Proposition 2.1 implies the next.
COROLLARY 2.2. Every locally C*-algebm is a Krull algebm.
By Proposition 1.4 and Corollary 2.2, we get the next.
COROLLARY 2.3. The cartesian product of locally C* -algebras is a Krull (locally
C* -) algebm.
In view of Corollary 2.2, Theorem 4.7 in [6: p. 3732] is improved as follows:
THEOREM 2.4. A locally C* -algebm is dual if and only if it is complemented.
Let (E, (Pa)aEA) be a proper complete locally m-convex H* -algebra with con-
tinuous involution. Then E can be made into a locally pre-C* -algebra, via a family
(qa)aEA of C*-seminorms given by
(2.1)
so that,
(2.2) qa(X) ~ Po(x) for every x E E, Q E A.
(Namely, the respective topology on E is weaker than the given one).
Moreover,
(2.3) Po(xy) ~ qa(x)Pa(Y) for every x, y E E, Q E A.
(See [9: p. 265, Proposition 2.3]).
In that framework we get the next two results.
PROPOSITION 2.5. Let (E, (Po)oEA) be a proper complete locally m-convex H*-
algebm with continuous involution and a unit element e. Then the following are
equivalent:
1) po(e) = 1 for every Q E A (: normal unit).
2) (E, (Pa)aEA) is a locally C*-algebm.
PROOF. 1) ==> 2): Let qa, Q E A be the seminorms given by (2.1). By (2.3),
p",(x) ~ qo(x) for every x E E, Q E A. Hence (see also (2.2)) Po = qa for every
Q E A. Namely, (E, (Pa)aEA) is a locally C*-algebra.

2) ==> 1): C*-property implies Pa(e)(l-po(e)) = 0 for every Q E A. If po(e) = 0


for some Q E A, then Ileoli o = 0, where eo = e + ker(Pa) is the respective unit
THE KRULL NATURE OF LOCALLY C'-ALGEBRAS 199

element in the factor algebra EOl == EOl (see aslo [1: p. 32, Theorem 2.4] and [11:
p. 91, Theorem 4.1]). Thus eOl = 0, which is a contradiction. Thus, pOl(e) =I- 0 for
every Q E A, hence POl (e) = 1 for every Q E A. 0
As a consequence of Corollary 2.2 and Proposition 2.5 we have the next.
COROLLARY 2.6. Every proper complete locally m-convex H* -algebm with con-
tinuous involution and a normal unit is a Krull algebm.
Our next aim is to provide another proof to the fact that a F'rechet locally
C*-algebra is Krull (see Corollary 2.2). To do this, we use the notion involved in
the next.
DEFINITION 2.7. A projective system {(EOl , fOl,8)}OlEA of topological algebras
is called perfect, if the restrictions to the projective limit algebra
(2.4) E = ~EOl = {(x Ol ) E II EOl: fOl,8(X,8) = XOl , if Q:S; (3 in A}
OlEA
of the canonical projections 7r0l : I10lEA EOl -+ E Ol , Q E A, namely, the (continuous
algebra) morphisms
(2.5) fOl = 7rOl IE =limE", : E -+ EOl , Q E A,
t--
are onto maps. The resulted projective limit algebra E = lim EOl is called perfect
~

(topological) algebm.
LEMMA 2.8. Every Frechet locally m-convex algebm (E, (Pn)nEN) gives a perfect
projective system of normed algebms.
PROOF. For any n :s; m in N, the connecting maps
(2.6)
with
fnm(X + ker(Pm)) = x + ker(Pn)
are onto algebra morphisms (see, for instance, [11: p. 86, (3.6) and (3.7)]). So,
since {(En, fnm)}nEN is a denumerable projective system of normed algebras, it
follows that fn, n E N (see (2.5)) are onto, as well (see [10: p. 229, Theorem
8]). 0
The proof in the next result is an adaptation of that in Proposition 2.1.
PROPOSITION 2.9. Any perfect projective limit of Krull algebms is a Krull
algebm.
PROOF. Let {(EOl,fOl,B)}OlEA be a perfect projective system of Krull algebras.
Consider the projective limit algebra E = limEOl (see (2.4)), which is a closed
~

subalgebra of the cartesian product topological algebra I10l EA EOl (see, for instance,
[11: p. 84, Lemma 2.1]). For a proper closed left ideal I in E, fOl(I) is a (closed)
left ideal in E Ol , Q E A. If f Ol (I) = EOl for every Q E A, then
1= limfOl(I)
~
= limfOl(I) = limEOl = E,
~ ~

(see also [ibid. p. 87, Lemma 3.2]), which is a contradiction. Thus, fOl(I) =I- EOl
for some Q E A. Since E Ol , Q E A is a Krull algebra, there exists a closed maximal
regular left ideal, say M, with fOl(I) ~ M and hence I ~ f;;l(1Ol(1)) ~ f;;l(M),
200 MARINA HARALAMPIDOU

where J;;l(M) is a closed maximal regular left ideal in E and this terminates the
proof for closed left ideals. Similarly, for closed right ideals. D
THEOREM 2.10. Any Prichet locally C*-algebm is a Krull algebm.
PROOF. Let (E, (Pn)nEN) be an algebra as in the statement. By [1: p. 32,
Theorem 2.4], the respective normed algebras En, n E N in the Arens-Michael de-
composition of E, are C* -algebras and hence Krull (see, for instance, [2: p. 56,
Theorem 2.4.5]. In particular, {(En' Jnm)}nEN is a perfect system of normed alge-
bras (see Lemma 2.8 and relation (2.6)). Proposition 2.9 assures that the projective
limit algebra lim En is a Krull algebra and hence E is a Krull algebra, as it fol-
+--
lows from Corollary 1.3 and the fact that E ~ lim En within a topological algebra
+--
isomorphism (see (1.3)). D

References
[1] C. Apostol, b*-algebms and their representation, J. London Math. Soc. 3(1971), 30-38. MR
44:2040.
[2] J. Dixmier, C*-Algebms, North-Holland, Amsterdam, 1977. MR 56:16388.
[3] R.S. Doran and V.A. Belfi, Chamcterizations of CO-Algebras. The Gel'fand-Na'tmark Theo-
rems, Marcel-Dekker, 1986. MR 87k:46115.
[4] M. Haralampidou, On locally convex H*-algebms, Math. Japon. 38(1993), 451-460. MR
94h:46088.
[5] M. Haralampidou, Annihilator topological algebras, Portug. Math. 51(1994), 147-162. MR
95f:46076.
[6] M. Haralampidou, On complementing topological algebms, J. Math. Sci. 96(1999), 3722-3734.
MR 2000j:46085.
[7] M. Haralampidou, On the Krull property in topological algebms (to appear).
[8] A. Inoue, Locally C* -algebras, Mem. Faculty Sci. Kyushu Univ. (SerA) 25(1971), 197-235.
MR 46:4219.
[9] A. EI Kinani, On locally pre-C*-algebm structures in locally m-convex H*-algebms, Thrk. J.
Math. 26(2002), 263-271.
[10] G. Kothe, Topological Vector Spaces, I, Springer-Verlag, Berlin, 1969. MR 40:1750.
[11] A. Mallios, Topological Algebms. Selected Topics, North-Holland, Amsterdam, 1986. MR
87m:46099.
[12] E.A. Michael, Locally multiplicatively-convex topological algebms, Mem. Amer. Math. Soc.
11(1952). (Reprinted 1968). MR 14,482a.
[13] Z. Sebestyen, Every C*-seminorm is automatically 8ubmultiplicative, Period. Math. Hung.
10(1979), 1-8. MR 80c:46065.

DEPARTMENT OF MATHEMATICS, UNIVERSITY OF ATHENS, PANEPISTIMIOPOLlS, ATHENS 15784,


GREECE
E-mail address:mharalamOcc.uoa.gr
Contemporary Mathematics
Volume 328, 2003

Characterizations and automatic linearity for ring


homomorphisms on algebras of functions

Osamu Hatori, Takashi Ishii, Takeshi Miura, and Sin-Ei Takahasi

ABSTRACT. Automatic linearity results for certain ring homomorphisms be-


tween two algebras, in particular, semi-simple commutative Banach algebras
with units are proved. For this purpose a representation by using the induced
continuous mapping between the maximal ideal spaces and ring homomor-
phisms on the field of complex numbers is given. Ring homomorphisms on
certain non-complete metrizable algebras into the algebras of analytic func-
tions are also considered. A characterization of the kernel of complex-valued
ring homomorphism on a commutative algebra is given. As a corollary of the
results a complete description of ring homomorphisms on the disk algebra into
itself is given in terms of prime ideals.

Introduction
A ring homomorphism between two algebras is a mapping which preserves ad-
dition and multiplication. If we assume that the mapping is linear, then it is an
ordinary homomorphism. In the case where the two algebras are just the field C
of complex numbers, the assumption cannot be avoided; there are ring homomor-
phisms of C into C which are not linear nor conjugate linear (cf. [9]). The history
of ring homomorphisms on C probably dates back to the investigation of Segre [19]
in the nineteenth century and that of Lebesgue [12]. A similar remark applies to
finite-dimensional Banach algebras. But this is not the case for several infinite-
dimensional ones; for instance, Arnold [1] proved that a ring isomorphism between
the two Banach algebras of all bounded operators on two infinite-dimensional Ba-
nach spaces is linear or conjugate linear (cf. [5]). Kaplansky [8] proved that if p
is a ring isomorphism from one semi-simple Banach algebra A onto another, then
A is a direct sum Al EB A2 EB A3 with A3 finite-dimensional, p linear on All and
p conjugate linear on A 2 . It follows that a ring isomorphism from a semi-simple
commutative Banach algebra onto another with infinite and connected maximal
ideal space is linear or conjugate linear.

2000 Mathematics Subject Classification. Primary 46JlO, 46E25; Secondary 46J40.


The first, the second, and the fourth author were partialy supported by the Grants-in-Aid
for Scientific Research, The Ministry of Education, Science, Sports and Culture, Japan.

2003 AJnerican Mathematical Society


201
202 O. HATORI, T. ISHII, T. MIURA, AND S.-E. TAKAHASI

It is interesting to study ring homomorphisms on Banach algebras which are


not necessarily injective or surjective. We may expect that a number of ring ho-
momorphisms on infinite-dimensional Banach algebras are automatically linear or
conjugate linear. By a routine work we see that a a ring homomorphism is real-linear
if it is continuous. On the other hand, we can also arrive at automatic linearity for
several ring homomorphisms by results in [14, 15, 20, 21, 13]; they studied and
characterized *-ring homomorphisms between commutative Banach algebras with
involutions and ring homomorphisms on regular commutative Banach algebras with
additional assumptions. The heart of this paper is automatic linearity results for
certain ring homomorphisms of a much more general nature.
Throughout the paper A and B denote semi-simple commutative Banach alge-
bras with units eA and eB respectively. The maximal ideal space for A is denoted
by MA. In this paper, we denote the Gelfand transform of a E A also by a. For a
ring homomorphism of C into C, we simply say a ring homomorphism on C.
Let T be a ring homomorphism on C and x E MA. Then the complex-valued
mapping p on A defined by
p(a) = T(a(x)), aEA
is a typical example of a ring homomorphism. Semrl [20, Example 5.4] showed that
there exists a complex-valued ring homomorphism other than this type.
In section 2 we show that if a ring homomorphism of A into B satisfies a
certain condition, say (m), then it is represented by a modified version of the
above. Many ring homomorphisms satisfy this weak and rather natural condition
(m): *-ring homomorphisms on involutive algebras; p{A)(y) = C for every y E M B ;
p(A) contains a subalgebra of B. Thus our result generalizes the previous ones in
[14, 20, 21, 13].
In section 3, by using results in section 2, we deduce some automatic linearity
results for ring homomorphisms: p with (m) is real-linear on a closed ideal of finite
co dimension in A; if p(CeA) = CeB and p(A) contains an element with an infinite
spectrum, then p is linear or conjugate linear. It is a natural question: under the
two hypotheses (1) p(CeA) c CeB and (2) p(A) contains an element with an infinite
spectrum, does it follow that p is linear or conjugate linear? We give an affirmative
answer under stronger hypotheses: (1) and (2)' p(A) contains an element whose
spectrum contains a non-empty open subset.
Problems in the same vein are also considered not only for Banach algebras
but also for algebras of analytic functions. Bers [3] proved that if U and V are
plane domains and H(U), H(V) are the rings of analytic functions on U, V respec-
tively, then any ring isomorphism of H(U) onto H(V) is induced by a conformal
(or anti-conformal) equivalence of V with U, thus the ring isomorphism is linear (or
conjugate linear). Nakai [17] and Rudin [18] have shown this also holds for open
Riemann surfaces. Ring homomorphisms which are not necessarily injective or sur-
jective are also considered by many mathematicians (cf. [7, 10]). Among them,
Becker and Zame [2] have proved automatic continuity and linearity for ring homo-
morphisms from certain complete metrizable topological algebras into the algebra
of analytic functions on connected, reduced analytic spaces.
In section 4 we also consider ring homomorphisms into the algebras of analytic
functions. In particular, we consider the case of a ring homomorphism p from the
algebra Rs of rational functions on C with poles off a subset SeC into an algebra
of analytic funtions. Here Rs is a metrizable topological algebra, but it cannot be
AUTOMATIC LINEARITY FOR RING HOMOMORPHISMS 203

a Banach algebra by the Baire category theorem. We show that, for certain subsets
S, p is automatically linear or conjugate linear, if the range of p contains a 11011-
constant function. We also give an example of S such that a ring hommomorphism
that is neither linear nor conjugate linear, and whose range contains non-constant
functions, is possible.
In the final section we study ring homomorphisms into C: ring homomorphisms
whose ranges contain only constant functions. We characterize the kernels of ring
homomorphisms from a unital commutative algebra into C, which is compared with
the one to one correspondence between maximal ideals and complex (linear) homo-
morphisms on commutative Banach algebras. As a corollary we show that there
exists an injective ring homomorphism from an algebra which consists of analytic
functions into C. We also give a complete description of the ring homomorphisms
on the disk algebra in terms of prime ideals.
We say that a ring homomorphism 7 on C is trivial if 7 = 0 or 7(Z) = Z (resp.
z) for every Z E C. Other ring homomorphisms on C are said to be non-trivial.
We note some properties of non-trivial ring homomorphisms on C, which are used
later in this paper. For a proof of the existence of non-trivial ring homomorphisms,
historical comments, and further properties, see [9]. It is easy to see that every
non-zero ring homomorphism 7 on C fixes rational real numbers and 7(i) = i or
-i. If 7 is non-trivial, then 7 does not preserve complex conjugation. (This is
a standard fact. Here is a proof. Suppose 7 does preserve complex conjugation:
7(Z) = 7(Z) for every Z E C. Then 7(JR) C JR, that is, 7 is a ring homomorphism on
the set of all real numbers R If x> 0, then 7(X) = (7( JX))2 > O. It follows that 7
is order preserving on R Since 7(r) = r for every rational real number r, we have
7(X) = x for every real number x. Thus 7(Z) = Z (resp. 7(Z) = z) for every Z E C
if 7(i) = i (resp. 7(i) = -i), which is a contradiction.) It is easy to see that T is
non-trivial if and only if 7 is discontinuous at every (resp. one) point in C. Thus, if
T is non-trivial, then it is unbounded on every neighborhood of zero. It follows that
there exists a sequence {w n } of complex numbers which converges to 0 such that
IT(Wn)1 tends to infinity as n -> 00 if 7 is non-trivial. If the ring homomorphism
on C is onto, then it is said to be a ring automorphism on C. Note that there is
a non-zero ring homomorphism on C which is not a ring automorphism. We also
note that there is a non-trivial ring automorphism on C (cf. [9, 11]).

1. Partial representation
If is a non-zero complex homomorphism on A, then there exists a unique
x E MA such that (a) = a(x) for every a E A. By this fact a well-known
representation of a (linear) homomorphism VJ from A into B follows: There exists
a continuous mapping <I> defined on {y E MB : VJ(a)(y) :I 0 for some a E A} into
MA such that
VJ(a)(y) = a(<I>(y)), a E A, y E {y E MB : VJ(a)(y) :I 0 for some a E A}.
On the other hand, if <I> is a continuous mapping of MB into AfA and Ty is a ring
homomorphism on C for every y E AfB , then
p(a)(y) = Ty(a(<I>(y)), a E A, y E MB
defines a ring homomorphism from A into the algebra of all complex-valued func-
tions on M B . Thus it defines a ring homomorphism from A into B under the
condition that T. (a (<I> (. )) is in B for every a E A, and this is the case when M B
204 O. HATORI, T. ISHII, T. MIURA, AND S.-E. TAKAHASI

is finite. A problem is the converse: Is every ring homomorphism represented as


above? A negative answer is known even in the case where B = C by the example
due to Semd [20, Example 5.4]. Nevertheless, we show that a partial representation
is still possible in this section.

DEFINITION 1.1. Let p be a ring homomorphism of A into Band y a point in


M B . The induced ring homomorphism py of A into C is defined by

py(a) = p(a)(y), aE A.

Let Ie : C -+ A be defined by IdA) = AeA for every A E C. We denote Ty = pyole.


For every y E M B , the induced mapping Ty is a ring homomorphism on C.

DEFINITION 1.2. Let p be a ring homomorphism of A into B. We denote:


Mo = {y E MB : Ty = a};
!vlt = {y E MB : Ty(Z) = Z for every Z E C};
ALl = {y E MB : Ty(Z) = Z for every Z E C};
Md,l = {y E MB : Ty is non-trivial and Ty(i) = i};
Md,-l = {y E !vIB : Ty is non-trivial and Ty(i) = -i}.

LEMMA 1.3. Let p be a ring homomorphism of A into B. Then M o, Ml U Md,l


and M-1 UMd,-l are clopen (closed and open) subsets of MB. The subsets M1 and
M-l are closed in M B .

PROOF. By the definitions it is easy to see that Mo = {y E MB : p(ieA)(y) =


a}, M1 U Md,l = {y E MB : p(ieA)(y) = i}, and ALl U Md,-l = {y E MB :
p(ieA)(y) = -i}, so they are clopen since p(ieA) is continuous on M B .
Next we show that Jl,ft is a closed subset of MB. Let y E Md,l' Since Ty is
non-trivial, there exists a complex number A such that Ty(A) =I- A. Put

Then G is an open neighborhood of y. We also see that G n M1 = 0. It follows


that ]\,{1 is a closed subset of MB since M1 U M d,l is clopen. In the same way, we
see that M-1 is a closed subset of M B . 0

Suppose that p is a ring homomorphism of A into B. If y E M 1 , then it is


easy to see that Py is a non-zero complex homomorphism on A. Thus there exists
a unique cp(y) in MA with

p(a)(y) = a(CP(y)), a E A.

In a way similar to the above we arrive at a partial representation as follows:

a, yEMo,
p(a)(y) = { a(CP(y)), yE M 1 ,
a(CP(y)), y E M_ 1 .

If y E Md,l U Md.-1, then the situation is complicated, in particular, ring homo-


morphisms with large Md,l U Md,-l are possible (cf. [20, Examples 5.3 and 5.4]).
AUTOMATIC LINEARITY FOR RING HOMOMORPHISMS 205

2. Ring homomorphisms which satisfy the condition (m)


In general the kernel of a non-zero ring homomorphism of A into C is a prime
ideal and need not be a maximal ideal. (See section 5 in this paper.) In this section
we consider ring homomorphisms P of A into B which satisfy the condition that
the kernel of the induced ring homomorphism Py for each y E MB defined by
Py(f) = p(f)(y), f EA
is a maximal ideal.
DEFINITION 2.1. Let P be a ring homomorphism of A into B. We say that P
satisfies the condition (m) if Py is zero or ker Py is a maximal ideal of A for every
yEMB.
By the following Lemma 2.2, if py(A) = C for every y E M B , in particular,
if p(A) :J CeB, then (m) is satisfied. A *-ring homomorphism also satisfies the
condition (m). (See the proof of Corollary 2.5.)
LEMMA 2.2. Let Po be a non-zero ring homomorphism of A into C. Then the
following are equivalent.
(1) The kernel ker Po of Po is a maximal ideal of A.
(2) The equation po(A) = PO(CeA) holds.
(3) There exist a non-zero ring homomorphism 7 on C and an x E MA such
that the equation po(a) = 7(a(x)) holds for every a E A. In this case 7 = Po 0 Ie.
Such a 7 and x are unique.
(4) The mnge Po(A) is a subfield ofC which contains a non-zero complex num-
ber.
PROOF. First we show that (1) implies (2). Suppose that ker Po is a maximal
ideal. Then there exists a non-zero complex homomorphism <p on A such that
ker Po = ker<p. Let a E A. Then a - <p(a)eA E ker<p, so po(a) = po(<p(a)eA) E
PO(CeA). Thus (2) holds.
Next suppose that (2) holds. We show (3). Let 7 be a function on C defined
by 7('x) = PO('xeA) for every complex number ,x, that is 7 = Po 0 Ie. Then 7- 1
is well-defined on po(A) onto C. By a simple calculation we see that 7- 1 0 Po is
a non-zero complex homomorphism on A. So there corresponds an x E MA such
that 7- 1 0 po(a) = a(x) holds for every a E A. Uniqueness for 7 and x is trivial.
Suppose that (3) holds. Then Po(A) = 7(C) and it is a subfield of C which
contains a non-zero complex number since 7 is a non-zero ring homomorphism.
Thus (4) holds.
Finally we show that (4) implies (1). Since ker Po is a proper algebra ideal of
A, there is an x E MA such that ker Po C {a E A : a(x) = a}. Assume that (4)
holds. We show ker Po = {a E A : a(x) = a}. Let a E A with a(x) = O. Suppose
that po(a) =I- O. Then there is abE A with Po(b) = 1/ po(a) for Po(A) is a field. It
follows that ab - e A E ker Po, which is a contradiction since (ab - eA)( x) = -1. We
conclude that ker Po = {a E A : a(x) = O} and ker Po is a maximal ideal of A. D
THEOREM 2.3. Let p be a ring homomorphism of A into B. Suppose that p
satisfies the condition (m). Then there exists a continuous function <P of MB \ Mo
into MA such that for every a E A

p(a)(y) = {7y(a(<p(y))),
0, yEMo.
206 O. HATORI, T. ISHII, T. MIURA, AND S.-E. TAKAHASI

In particular, p(a)(y) = a(q>(y)) for every y E Ml and p(a)(y) = a(q>(y)) for every
y EM_I' The set q>(Md,l U Md,-d is a (possibly empty) finite subset of MA and
q>-l(X) n (Md.l U Md,-d is an open subset of MB for each x E q>(Md,l U Md,-d.

PROOF. Let y E MB \ Mo. Then by the condition (m) ker PY is a maximal


ideal of A, so by Lemma 2.2 there exists a unique q>(y) E MA such that
py(a) = PY o Ic(a(q>(y)))
holds for every a E A. By the definition Ty = py 0 Ie and since py(a) = p(a)(y) we
see that p(a)(y) = Ty(a(q>(y))) holds for every a E A. If y E M l , then Ty(A) = A
for every A E C, so p(a)(y) = a(q>(y)). If y EM_I, then Ty(A) =). for every A E C,
so p(a)(y) = a(q>(y)). If y E M o, then p(a)(y) = O.
Put Md = Md,l UMd,-l. We show that q>(Md) is a finite subset of MA. Suppose
not. Then there is a countable subset {Xn}~=l of q>(Md). For each n choose a point
Yn E Md with q>(Yn) = Xn Since TYI is unbounded near zero, there exists an al E A
such that lIalli < 2- 1 and hI (al(xd)1 > 2. By induction we can find, for every n,
an E A such that an(xd = ... = an(xn-d = 0, lIanll < 2- n , and
ITYn(an(xn))1 > 2n + ITYn(al(Xn) + ... + an-l(xn))l
(Choose b2 E A with b2 (xd = 0 and b2(X2) = 1. Since TY2 is unbounded near
zero, there is a non-zero complex number 02 such that 1021 < Ilb211- l 2- 2 and
ITY2(0)1 > 22+ITy2 (al(x2))I Then put a2 = 02b2. We have a2(xl) = 0, IIa211 < 2- 2,
and ITy2 (a n (x2))1 > 22+ITy2 (al(x2))1. Suppose that al, .. ' ,an-l E A are choosenso
that the conditions are satisfied. Choose bn E A with bn(xd = ... = bn(xn-d = 0
and bn(xn) = 1. Since TYn is unbounded near zero, there is a non-zero complex
number On such that 10nl < Ilbn ll- 1 2- n and
ITYn (on)1 > 2n + ITYn (al(xn) + ... + an-l(xn))l
Then a2 = onbn is a desired function for n.) Then E::'=l an converges in A, say to
a. Then a(xn) = al(x n ) + ... + an(xn) since the Banach norm on A dominates the
uniform norm on AlA. On the other hand

so that p(a) is unbounded, which is a contradiction proving that q>(Md) is a finite


set.
Let q>(Md) = {Xl, ... , xn} and Yj = q>-l(Xj) n Md for each j = 1,2, ... , n.
Choose an a E A such that a(xl) = 1, a(x2) = ... = a(xn) = O. Then p(a)(y) = 1
if y E Y l while p(a)(y) = 0 if y E Md \ Y l . Because p(a) is continuous, Y1 is clopen
in Md; but Md is open in M B , so Y l is open in MB. In the same way we see that
Yj is an open subset of MB for each j = 2,3, ... n.
Finally we prove that q> is continuous. Since Yj is open and q>(Yj) = Xj, we
only need to prove that q> is continuous at each point in Ml U M_ l . Let y E Ml
and {y>.hEA be a net which converges to y. Without loss of generality we may
assume that {y>.} c Ml U Md,l since Ml U Md,l is clopen. Suppose that {q>(y>.)}
does not converge to q>(y), that is, there is an open neighborhood G of q>(y) such
that for every A E A there exists a A' 2:: A with q>(YN) rf. G. There exist a finite
number of points aI, ... , am in A and a positive real number e such that
{x E MA : laj(x) - aj(q>(y))1 < e,j = 1,2, ... , m} C G.
AUTOMATIC LINEARITY FOR RING HOMOMORPHISMS 207

Since {p(aj)(y,x)} converges to p(aj)(Y) for each j, there exists a Ao E A such that
JTy>. (aj(iP(y,x))) - aj(iP(y))J < e
holds for every A ~ AO and j = 1,2, ... , m. Suppose that A ~ Ao, then there exists
a A' ~ A such that iP(y,x/) f/. G, so that Jajl(iP(y,x/)) - ajl(iP(y))J ~ e for some j'.
lt follows that YN E Md,l. We also see that iP(y,x/) E {Xl, ... ,Xn } \ {iP(y)}. There
exists an a E A such that a(iP(y)) = 1 and a = 0 on {Xl, ... ,Xn } \ {iP(y)}. We
conclude that for every A with A ~ Ao there exists a A' ~ A such that p(a)(y,x/) = 0
and p(a)(y) = 1, which is a contradiction since p(a) is continuous on M B . Thus we
have that {iP(y,x)} converges to iP(y), so iP is continuous at y. In the same way we
see that iP is continuous at each point in M_ I . We have proved that iP is continuous
on MB \Mo. 0
Note that the set Md,l U Md,-l need not be a finite set or even a closed subset
of AlB (cf. [20, Example 5.3]).
In [21] the authors proved the following corollary in the case where A is regular
and satisfies a certain additional condition. Now we can remove these conditions.
COROLLARY 2.4. Let p be a ring homomorphism from A into B. Suppose that
py(A) = C for every y E M B . Then there exists a continuous mapping iP of MB
into MA and a non-trivial ring automorphism Ty on C for every y E Md,l U Md,-l
s1Lch that
a(iP(Y))' Y E MI ,
{
p(a)(y) = a(iP(y)), Y EM_I,
Ty(a(iP(y))), Y E Md,l U Md,-l.
Moreover iP(Md,1 U Md,-d is a finite subset of MA.
PROOF. By Lemma 2.2 we see that ker py is a maximal ideal, so the condition
(m) is satisfied. The conclusion follows by Theorem 2.3. In particular, Ty = Py 0 Ie
is onto, thus it is a non-trivial automorphism on C for y E Md,l U Md,-l. 0
Theorem 2.1 in [13] for the case of unital and semi-simple commutative Banach
algebras is also deduced from Theorem 2.3
COROLLARY 2.5. Suppose that A is involutive and B is symmetrically involu-
tive. Let p be a *-ring homomorphism. Then MB = Mo U MI U M-I and there
exists a continuous function iP from MB \ Mo into AlA such that
a(iP(Y))' yE MI ,
p(a)(y) = { 0, yE Mo,
--:a('-=-iP..,...(
y77")) , y EM_I

PROOF. Since p is a *-ring homomorphism, it it easy to see that Ty(Z) = Ty(Z)


for every Z E C and for every y E MB \ Mo. It follows that Ty is 0 or linear or
conjugate linear. Thus the conclusion follows. 0

3. Automatic linearity
One of the reasons for ring homomorphisms between infinite-dimensional Ba-
nach algebras to be linear or conjugate linear is that the range contains an element
with large spectrum. In this section we show evidence of this.
208 O. HATORI, T. ISHII, T. MIURA, AND S.-E. TAKA HAS I

COROLLARY 3.1. Let p be a ring homomorphism of A onto B. If MB contains


no isolated point, then p is real-linear. If MB is infinite and connected, then p is
linear or conjugate linear.
PROOF. Since p is a surjection, py(A) = C for every y E M B , so the condition
(m) is satisfied by Lemma 2.2. We also have that the induced mapping <I> is injective.
Thus Md,l U Md,-l is a (possibly empty) finite set. Because Md,l U Md,-l is open
(by Lemma 1.3), each point of Md,l U Md,-l is isolated in M B .
If MB contains no isolated point, then Md,l U Md,-l = 0. Thus p is real-linear.
If lvIB is infinite and connected, then MB contains no isolated point, so Md,l U
Md,-l = 0. It follows by Lemma 1.3 that MB = lvh or MB = M_ l . Thus P is
linear or conjugate linear. 0
COROLLARY 3.2. Let p be a ring homomorphism of A into B. Suppose that
p satisfies the condition (m). Then there exists a (possibly empty) finite subset
{Xl, ... , xn} of M A such that p is real-linear on the finite-codimensional closed
ideal {a E A: a(xj) = O,j = 1,2, ... ,n} of A.
PROOF. Put {Xl,""X n } = <I>(Md,l UMd,-d. (The set is finite by Theorem
2.3.) Then for every a E {a E A: a(xj) = O,j = 1,2, ... ,n}

p(a)(y) = {Ty(a(<I>(Y))), y E Ml U M_l'


0, Y E Mo U Md,l U Md,-l.
Since Ty is real-linear for every y E Ml U M_l' the conclusion follows. o
COROLLARY 3.3. Let p be a ring homomorphism from A into B such that
p(CeA) = CeB. Then we have that Mo = 0, and there exists a continuous mapping
<I> from MB into MA such that one of the following three occurs.
(1) P is linear:
p(a)(y) = a(<I>(y)), a E A, y E MB .
(2) p is conjugate linear:
p(a)(y) = a(<I>(y)), a E A, y E MB .
(3) There exists a non-trivial ring automorphism T on C such that
p(a)(y) = T(a(<I>(y))), a E A, y E M B.
In particular, if there exists an a E A such that the spectrum of p( a) is an infinite
set, then p is linear or conjugate linear.
PROOF. For every y E MB, we have py(CeA) = C, so py(A) = py(CeA) = c.
Thus ker Py is a maximal ideal of A by Lemma 2.2, so that the condition (m) is
satisfied. Since p(ieA) E CeB, MB = Ml U Md,l or MB = M_l U Md,-l. Suppose
that Md,l U Md,-l = 0. Then (1) or (2) occurs. Suppose that there exists some
Yd E Md,l and some Yl E M l Then there is a complex number>' with Tyd (>') =I=- >.,
so that p(>.eA) is not a constant function, which contradicts our hypothesis. Thus
Md,l =I=- 0 implies that MB = Md,l. It is also easy to see that Ty is identical for
every y E Md,l since p(CeA) consists of constant functions. Thus (3) follows. In
the same way we see that (3) follows if Md,-l =I=- 0.
Suppose that there exists an a E A such that p(a)(MB) is infinite, then (3)
does not occur since <I> ( M B) is a finite set in this case. It follows that p is linear or
conjugate linear. 0
AUTOMATIC LINEARITY FOR RING HOMOMORPHISMS 209

Corresponding results for ring homomorphisms on rings of analytic functions


are proved by Kra [10, Theorem I].
Suppose that p is a ring homomorphism from A into B which satisfies two
conditions: p(CeA) C CeB; there exists an a E A such that p(a)(MB) is infinite.
Does it follow that p is linear or conjugate linear? Although the authors do not
know the answer, we can provide a positive answer under a stronger condition.
THEOREM 3.4. Let p be a ring homomorphism from A into B. Suppose that
the following two conditions are satisfied: (i) p(CeA) C CeB; (ii) there exists an
a E A such that p(a)(MB) contains a non-empty open subset ofC. Then p is linear
or conjugate linear.
PROOF. Since p(CeA) C CeB we may suppose that pole is a non-zero ring
homomorphism on C. We have two possibilities: po Ie(i) = i; po Ie(i) = -i. We
show that, in the first case, po Ie(z) = z for every complex number z, so it will
follow that p is linear on A. (In the same way we see that p is conjugate linear if
pole( i) = -i.) Suppose that pole( i) = i. We show that pole is continuous on
C. For this it is enough to show that pole is continuous at O. Suppose not. Then
there is a sequence {w n } of non-zero complex numbers which converges to 0 such
that {poIe(w n )} does not converge to O. Without loss of generality we may assume
that Ip 0 Ie(w n ) I ~ 00 as n ~ 00. Let a be in A such that p(a)(MB) contains a
non-empty open subset G of the complex plane. Let s be a complex number in G
such that the real part and the imaginary part of s are both rational numbers. Put
Zn = S + 1/ po Ie(w n ). Then there is a positive integer mo such that Zm E G for
every m ~ mo since Ip 0 Ie(w n )I ~ 00 as n ~ 00, so ZmeB - p(a) ~ B- 1 . Thus we
have (8 + l/wm)eA - a ~ A-I. Then 8 + l/wm is in the spectrum of a for every
m ~ mo, which is a contradiction since Is + l/wnl ~ 00 as n ~ 00. It follows that
pole is continuous at 0, thus on C, so pole( w) = w for every complex number w
since p(ieA) = i. Then we see that p is linear on A. D
Note that either of the two conditions (i) and (ii) in the above theorem itself
does not suffice for p to be linear or conjugate linear. Let T be a non-trivial ring
automorphism on C. Suppose that x E MA and tp from A into C is defined by
tp(a) = a(x) for every a E A. Put P = TO tp. Then p is a ring fomomorphism with
(i) since p(CeA) = C, but P is neither linear nor conjugate linear; p is not even
real-linear. Let D be the closed unit disk in the complex plane. Let D + 3 = {z E
C: Iz - 31::; I} and X = D U (D + 3). Define
zED
p(f)(z) = {f(Z)'
f(z - 3), zED+3
for every f E C(D). Then p is a ring homomorphism from C(D) into C(X) with
the condition (ii). But p is neither linear nor conjugate linear. Even more is true.
There is a ring homomorphism with the condition (ii) which is not real-linear.
Recall that the disk algebra A(D) is the algebra of all complex-valued continuous
functions on D which are analytic on the interior D of D. Suppose that
K = {O} U {l/n : n is a positive integer},
X = {2} U D
and
Y = K U {z E C : Iz - 31::; I}.
210 O. HATORI, T. ISHII, T. MIURA, AND S.-E. TAKAHASI

Let
A = {f E C(X) : flD E A(D)}
and B = C(Y), where CO denotes the algebra of all complex-valued continuous
functions on '. Let be the ring homomorphism from C into C(K) defined in [20,
Example 5.3]. For f E A, put
f(2), y=O,
p(f)(y) = { (f(2))(1/n), y = lin,
f(y - 3), Iy - 31 :::; 1.
Then p is a ring homomorphism and satisfies the condition (ii). But (i) is not
satisfied and p is not real-linear on A.

4. Ring homomorphisms into algebras of analytic functions


Suppose that A is a completely metrizable topological algebra with an identity
and r(X, Ox) is the algebra of global sections of a connected reduced complex-
analytic space (X, Ox). Becker and Zame [2] proved among other things that if p
is a ring homomorphism from A into r(X, Ox) such that the range of p contains
a non-constant section, then p is linear or conjugate linear. This is not the case
for ring homomorphisms on non-complete algebras. (Suppose that P is the algebra
of polynomials on C and H(CC) is the algebra of entire functions. Let T be a non-
trivial ring homomorphism on C and define p on P by p(Eanz n ) = ET(an)Zn
for every polynomial E anz n . Then p is a ring homomorphism, but it is neither
linear nor conjugate linear. By Theorem 5.1 there also exists an injective ring
homomorphism from Pinto C since {O} is a prime ideal in P.) Nevertheless we
show automatic linearity results for ring homomorphisms on certain non-complete
metrizable algebras.
THEOREM 4.1. Suppose that A is a complex (commutative or non-commutative)
algebra with unit e. Suppose that Y is a non-empty set and B is a complex algebra
of complex-valued functions on Y which contains the constant functions. Suppose
that for every non-constant function b E B the range of b contains a non-empty
open subset of c. Let p be a ring homomorphism from A into B. If there exists an
element a in A such that the resolvent set of a contains a non-empty open subset
G of C and p( a) is non-constant, then p is linear or conjugate linear.
PROOF. It is easy to see that p( e) = 0 or 1 since the range of a non-constant
function in B contains a non-empty open set. If p( e) = 0, then p is 0 on A and so
p is linear. Suppose that p(e) = 1. In the same way as above, we see that p(ie) = i
or -i. We show that p is linear if p(ie) = i. (If p(ie) = -i, then 15 defined by
15(a) = p(a) will be linear by what we will show, so p will be conjugate linear.) We
will show that p(,Xe) = ,X for every complex number 'x.
First we show that p(Ce) C cc. Suppose not. Then there is a complex number
,x such that p(,Xe) is a non-constant function. Note that Re'x or Im'x is irrational
since p(xe) = x for every rational real number x by a simple calculation. Since
p('xe)(Y) contains a non-empty open set, there exists r E p('xe)(Y) with Rer and
Imr both rational. Then p(,Xe) - r is not an invertible element in B. Therefore
(,x - r)e is not invertible in A, that is, ,x = r, which is a contradiction. Thus we
have proved that p(Ce) C C, or p induces a ring homomorphism on cc. We denote
the induced ring homomorphism also by p.
AUTOMATIC LINEARITY FOR RING HOMOMORPHISMS 211

Suppose that p is non-trivial on C. Since p(a) is not a constant, the interior


of p( a) (Y) contains a complex number s whose real and imaginary parts are both
rational, by our assumption for B. Since p is non-trivial, there exists a sequence
{w n } of non-zero complex number such that Ip(wn)1 tends to infinity as n --+ 00
and s+ l/w n E G for every n. Since p(i) = i we see that p(s+ l/wn ) = s+ 1/ p(w n )
and we may assume that
s + l/p(w n ) E p(a)(Y)
for every n. Thus p(a) - (s + 1/ p(wn )) is not invertible in B. It follows that
a - (s + l/w n )e is not invertible in A, which is a contradiction, proving that p is
trivial. Since p(i) = i, we have that p(A) = A for every complex number A. We
conclude that p is linear on A. 0
The spectrum of each element in a Banach algebra is compact, so the conditions
for A in Theorem 4.1 are satisfied by every Banach algebra with unit. Since the
range of non-constant analytic function is a non-empty open subset of C, algebras of
global sections on connected, reduced complex-analytic spaces satisfy the condition
for B in Theorem 4.1. Thus we have the following, which is a version of a more
general result of Becker and Zame [2, Theorem 3.1]. But our proof is considerably
simpler.
COROLLARY 4.2. Let Ao be a Banach algebra with unit. Suppose that p is
a ring homomorphism from Ao into r(X, Ox), the algebra of global sections on
a connected, reduced complex-analytic space (X , Ox). If p( Ao) contains a non-
constant section, then p is linear or conjugate linear.
Let S be a subset of C. We denote by Rs the algebra of all rational functions
on C with poles off S. Although Rs is a wlital algebra, it cannot be a Banach
algebra by the Baire category theorem. If S = C, then Rs = P, and so there is
a ring homomorphism p on Rc into r( X , Ox) for (X, Ox) = C such that p(Rc)
contains a non-constant function, while p is neither linear nor conjugate linear. In
the case where C \ S contains an interior point, the situation is different; in this
case we prove an automatic linearity result.
COROLLARY 4.3. Let S be a subset of C whose complement contains interior
points. Suppose that (X, Ox) is a connected, reduced complex-analytic space and
r(X, Ox) is the algebra of global sections. Suppose that p is a ring homomorphism
from Rs into r (X , Ox). If the range of p contains a non-constant section, then p
is linear or conjugate linear.
PROOF. In the same way as in the proof of Theorem 4.1 we see that p(C) C C.
Suppose that z denotes the identity function: z(w) = w for every complex number
w. Then we have that p(z) is non-constant. (Suppose not. Then p(f) is a constant
section for every f E Rs.) On the other hand z - A is invertible for every A E C \ S,
which contains a non-empty open set. Thus the conditions in Theorem 4.1 are
satisfied. It follows by Theorem 4.1 that p is linear or conjugate linear. 0
Note that every ring homomorphism p of R0 into r(X, Ox) is constant-valued
for the empty set 0. (We see that p(C) C C as before. Suppose that p(f) is not a
constant section for some non-constant rational function f. Then there is a complex
number r in p(f)(X) with rational real and imaginary parts. It follows that f - r
is not invertible in R0, which is a contradiction.)
212 O. HATORI, T. ISHII, T. MIURA. AND S.-E. TAKAHASI

Let A be one of the algebras P, njj or the disk algebra A(D), where jj denotes
the closed unit disk in C, and H(D) the algebra of analytic functions on the open
unit disk. Although both P and no are dense in the disk algebra, automatic
linearity results for ring homomorphisms on these algebras are different from each
other. Suppose that p is a ring homomorphism from A into H(D) such that the
range of p contains a non-constant function. If A = njj (resp. A(D), then p is
linear or conjugate linear by Corollary 4.3 (resp. Corollary 4.2). But that is not
the case for A = P. The ring homomorphisms defined by p(2:a n z n ) = 2:1'(a n )zn
for polynomials 2: anz n are neither linear nor conjugate linear for non-trivial ring
homomorphisms l' on C.

5. Complex-valued ring homomorphisms


In this section we consider ring homomorphisms into the complex number field
C. Suppose that A is a complex algebra and p is a non-zero ring homomorphism
from A into C. Then the kernel ker p of p is a prime ideal. Recall that a proper
ideal I of A is said to be a prime ideal of A if fg E I implies that f E I or gEl.
By using well-known results of algebra, we see the converse is also valid; for every
prime ideal such that the cardinal number of the quotient algebra of the algebra by
the ideal is equal to that of the continuum, there exists a ring homomorphism into
C whose kernel coincides with the ideal.
Let K be an extension field of a field k. (Here and after a field means a
commutative field.) We recall a subset S of K is said to be algebraically independent
over k if the set of all finite products of elements in S is linearly independent over
k. A subset T of K which is algebraically independent over k and is maximal with
respect to the inclusion ordering is said to be a transcendence base of Kover k.
By definition, for every transcendence base T of Kover k, K is algebraic over the
quotient field k(T) of the polynomial ring of T over k. There exists a transcendence
base of Kover k (cf. [11, Theorem X.l.I]). Using the same argument as in [9] we
can prove the following (cf. [11, 20]). (This might be a standard fact. But we
present here with a proof for the convenient of the readers.)

THEOREM 5.l. Let A be a commutative complex algebra with unit e. Suppose


that I is a prime ideal of A such that the cardinal number of AI I is that of the
continuum c. Then there exists a ring homomorphism p from A into C such that
kerp=I.

PROOF. The quotient algebra AI I has no non-zero divisor of zero, for I is a


prime ideal. We denote by K the field of fractions over AI I. Let Q be the field
of complex numbers whose real and imaginary parts are both rational. Let TK
be a transcendence base for Kover Q and T a transcendence base for Cover Q.
Then the cardinal number of TK (resp. T) is c since that of AI I (resp. q is
c. There exists au injection a defined from TK onto T. Since TK is algebraically
independent, there is a unique extension from Q(TK ) onto Q(T), which is also
denoted by a, and a is a ring homomorphism. Since C is algebraically closed and
K is an algebraic extension of Q(TK), there exists an extension of a which defines
a ring homomorphism of K into C by Theorem VII.2.8 in [11]. We also denote it
bya. Let h be the natural homomorphism of A onto AI I. Put p = a 0 h. Then p
is the desired ring homomorphism. D
AUTOMATIC LINEARITY FOR RING HOMOMORPHISMS 213

Note that the corresponding ring homomorphism p is not unique. Let T be any
non-zero ring homomorphism on C. Then TOp is a ring homomorphism on A with
ker p = ker TOp.
As a corollary of Theorem 5.1 we display a pathological feature of ring homo-
morphisms on algebras of analytic functions into C; even injection can be possible.
COROLLARY 5.2. Let A be a unital algebra which consists of holomorphic func-
tions on a domain in cn.
Then there exists an injective ring homomorphism of A
into Co
PROOF. Since the ideal containing only zero is a prime ideal and the cardinality
of A is the same as that of the continuum, there exists a ring homomorphism p of
A into C whose kernel consists only of zero, by Theorem 5.1. Then p is an injective
ring homomorphism. 0

Note that the injective ring homomorphism in Corollary 5.2 can never be sur-
jective if A contains non-constant functions since A is not a field. Note also that
every ring homomorphism from a unital commutative C* -algebra into C cannot be
injective if the dimension of the algebra is greater than one since {O} is not a prime
ideal in this case.
Together with the results in the previous sections we give a complete description
of ring homomorphisms on the disk algebra A(D).
COROLLARY 5.3. Let p be a non-zero ring homomorphism on the disk algebra
into itself. Then ker p is a prime ideal. If the range of p contains a non-constant
function, then p is linear or conjugate linear; there exists 'P E A(D) with 'P(D) c fJ
such that
p(f)(z) = f 0 'P(z), zED, f E A(D)
or
p(f)(z) = f 0 'P(z), z E fJ, f E A(fJ).
On the other hand, suppose that'P E A(D) with 'P(D) c D. Then
a(f)(z) = f 0 'P(z), ZED, f E A(fJ)

defines a linear ring homomorphism and


a(f)(z) = f 0 'P(z), zED, f E A(D)

defines a conjugate linear ring homomorphism.

PROOF. A(D) has no non-zero divisors of zero, so the kernel of any ring ho-
momorphism from complex algebra with unit element into A(D) must be a prime
(algebra) ideal. If p(A(D) contains a non-constant function, then by Theorem 4.1
we see that p is linear or conjugate linear. Suppose that p is linear. Then it is well
known and easy to prove, since the maximal ideal space of A(D) is the closed unit
disk D, that there exists 'P E A(fJ) with 'P(D) c D such that
p(f)(z) = f 0 'P(z)
holds for every f E A(D) and zED. Suppose that p is conjugate linear. Let
h : A(fJ) -+ A(fJ) be defined as

h(f)(z) = f(2), f E A(D), zED.


214 O. HATORI, T. ISHII, T. MIURA, AND S.-E. TAKAHASI

Then hop is a linear ring homomorphism on the disk algebra. It follows that there
exists r.p E A(D) with r.p(D) c D such that
h 0 p(f)(z) = 1 0 r.p(z) , zED, I E A(D).
Thus we see that
p(f)(z) = 1 0 r.p(z),
holds for every I E A(D) and zED.
Conversely, suppose that r.p E A(D) with r.p(D) c D. Then it is easy to see that
u(f)(z) = 1 0 r.p(z), zED, IE A(D)
defines a linear ring homomorphism and
u(f)(z) = 1 0 r.p(z), zED, IE A(D)
defines a conjugate linear ring homomorphism. o
Let n be a positive integer and An(D) the subalgebra of those I in A(D)
whose n-th derivative I(n) on D is continuously extended up to D. An(D) is a
unital commutative Banach algebra with the norm IIIlIn = L~=o III(k)lIoo/k! for
I E An(D), where II . 1100 is the supremum norm on D. Then Corollary 5.3 is
also valid for An(D). Prime ideals in A(D) and An(D) are studied in [16]. (See
also [4] for the case of A(D).) Mortini proved that every non-zero prime ideal
is contained in a unique maximal ideal. He in fact showed that a non-zero and
nonmaximal prime ideal in An(D) (resp. A(D)) is dense in exactly one of the
ideals {J E An(D) : 1(>") = 1'(>..) = ... = I(j)(>..) = O} for some 0 ~ j ~ n (resp.
{I E A(D) : 1(>") = O}), >.. E aD. We also see by a theorem of Dietrich [4] that
the cardinal number of the set of all prime ideals of A(D) which is contained in
a maximal ideal {J E A(D) : 1(>") = O}, >.. E aD is 2', the cardinal number of
the set of all the subsets of the continuum. Thus we see that there are 2' ring
homomorphisms on the disk algebra.
Acknowlegement. The authers would like to thank Professor Ken-Ichiroh
Kawasaki for his valuable comments. They also would like to thank the referees for
their careful reading of the paper and their valuable comments.

References
[1] B. H. Arnold, Rings of opemtors on vector spaces, Ann. of Math. 45(1944), 24-49
[2] J. A. Becker and W. R. Zame, Homomorphisms into analytic rings, Amer. Jour. Math.
101(1979), 1103-1122
[3] L. Bers, On rings of analytic junctions, Bull. Amer. Math. Soc. 54(1948), 311-315
[4] W. E. Dietrich, Jr., Prime ideals in uniform algebms, Proc. Amer. Math. Soc. 42(1974),
171-174
[5] M. Eidelheit, On isomorphisms of rings of linear opemtors, Studia Math. 9(1940),97-105
[6] J. B. Garnett, Bounded analytic functions, Academic Press, New York, 1981
[7] H. Iss'sa, On the meromorphic junction field of a Stein variety, Ann. Math. 83(1966), 34-46
[8] 1. Kaplansky, Ring isomorphisms of Banach algebms, Canadian J. Math. 6(1954),374-381
[9] H. Kestelman, Automorphisms of the field of complex numbers, Proc. London Math. Soc.
53(1951), 1-12
[10] 1. Kra, On the ring of holomorphic functions on an open Riemann surface, Trans. Amer.
Math. Soc. 132(1968),231-244
[11] S. Lang, Algebm (second edition), Addison-Wesley, California, 1984.
[12] M. H. Lebesgue, Sur les tmnsformations ponctuelles, tmnsformaant les plans en plans, qu'on
peut definir par des procedes analytiques, Atti della R. Acc. delle Scienze di Torino 42(1907),
532-539
AUTOMATIC LINEARITY FOR RING HOMOMORPHISMS 215

[13J T. Miura, Star ring homomorphisms between commutative Banach algebrns, Proc. Amer.
Math. Soc. 129(2001), 2005-2010
[14J L. Molnar, The rnnge of a ring homomorphism from a commutative C' -algebrn, Proc. Amer.
Math. Soc. 124(1996), 1789-1794
[15J L. Molnar, Automatic surjectivity of ring homomorphisms on H* -algebrns and algebrnic
differences among some group algebrns of compact groups, Proc. Amer. Math. Soc. 128(2000),
125-134
[16J R. Mortini, Prime ideals in the algebrn An(D), Complex Variables Theory AppJ. 6(1986),
337-345
[17J M. Nakai, On rings of analytic and meromorphic functions, Proc. Japan Acad. 39(1963),
79-84
[18J W. Rudin, An algebrnic charncterization of conformal equivalence, Bull. Amer. Math. Soc.
61(1955), 543
[19J S. de Corrado Segre, Un nuovo campo di ricerche geometriche, Atti della R. Acc. delle Scienze
di Torino 25(1889), 276-301
[20J P. Semrl, Non-linear perturbations of homomorphisms on C(X), Quart. J. Math. Oxford (2)
50(1999),87-109
[21J S.-E. Takahasi and O. Hatori, A structure of ring homomorphisms on commutative Banach
algebrns, Proc. Amer. Math. Soc. 127(1999),2283-2288

DEPARTMENT OF MATHEMATICAL SCIENCE, GRADUATE SCHOOL OF SCIENCE AND TECHNOL-


950-2181 JAPAN
OGY, NIIGATA UNIVERSITY, NIIGATA
E-mail address:hatorilDmath.se.niigata-u.ae.jp

NIIGATA CHUO HIGH SCHOOL, NIIGATA 951-8126 JAPAN

DEPARTMENT OF BASIC TECHNOLOGY, ApPLIED MATHEMATICS AND PHISICS, YAMAGATA UNI-


VERSITY, YONEZAWA 992-8510 JAPAN
E-mail address:miuralyz.yamagata-u.ae.jp

DEPARTMENT OF BASIC TECHNOLOGY, ApPLIED MATHEMATICS AND PHISICS, YAMAGATA UNI-


VERSITY, YONEZAWA 992-8510 JAPAN
E-mail address:sin-eiOemperor.yz.yamagata-u.ae.jp
Contemporary :r...1athematics
Volume 328. 2003

Carleson Embeddings for Weighted Bergman Spaces

Hans Jarchow and Urs Kollbrunner

ABSTRACT. We are going to discuss Carleson measures for the standard weigh-
ted Bergman spaces A~ (-1 < a < 00, 0 < p < (0). These are finite,
positive Borel measures J.L on the unit disk in IC such that, given 0 < q < 00,
A~ embeds, as a set, continuously into Lq(J.L). Such measures have been
closely investigated by V.L. Oleinikov and B.S. Pavlov [15], W.W. Hastings
[6] and D.H. Luecking [11], [12]. We complement their results, in particular
by characterizing compactness, order boundedness and related (absolutely)
summing properties of the canonical embedding A~ <-t Lq(J.L).

1. Introduction
The Carleson measures under investigation are finite, positive Borel measures
Jl. on the open unit disk l[J in the complex plane such that, given -1 < 0 < 00
and 0 < p, q < 00, the (classical) weighted Bergman space A~ is a subset of
Lq(J.l) and the embedding is bounded. Such measures have been characterized,
even in a more general setting, by V.L. Oleinikov and B.S. Pavlov [15J, W.W.
Hastings [6J and D.H. Luecking [11], [12J. Our first main topic is to complement
their results by characterizing when the canonical embedding I : A~ ~ Lq(J.l) is
compact. We will see that this is always the case if p > q. Our second main
topic is to characterize when I is order bounded, that is, the unit ball of A~ is
a subset of some order interval in Lq(J.l). As a consequence, we obtain necessary
and sufficient conditions for I to have specific (absolutely) summing properties.
Our results extend corresponding ones for composition operators which have been
obtained e.g. in [3]' [4], [20]' [21J. In fact, they can also be viewed as results on
composition operators which may have rather unusual range spaces. They apply,
for example, to pointwise multipliers.
Many of the results to be presented remain valid for measures J.l on D for which
f f-> f induces just a bounded linear map A~ ~ Lq(J.l) (not necessarily injective).
This is rather straightforward; precise formulations, however, require somewhat

1991 Mathematics Subject Classification. Primary 46 E 15,47 B 38, 47 B 10; Secondary 46


B 25, 30 H 05, 32 H 10.
Key words and phrases. Weighted Bergman spaces, Carleson measures, composition opera-
tors, compactness, order boundedness, absolutely summing operators.
The results of this paper are part of the dissertation of the second named author written at
the University of Ziirich under the supervision of the first.

2003 Alnerican Mathematical Society


217
218 HANS JARCHOW AND URS KOLLBRUNNER

clumsy notation. Also, there is an immediate extension to complex Borel measures


on 1lJ whose variation is (a, p, q) - Carleson.
We are indebted to the referee for providing Example 8 and for bringing to our
attention the paper [15] by V.L. Oleinikov and B.S. Pavlov.

2. Weighted Bergman spaces


Throughout the paper, we will use standard results and notation from (quasi-)
Banach space theory.
We will work on the open unit disk 1lJ = {z E C: Izl < I} in the complex plane.
The space 1i(1lJ) of all analytic functions 1lJ -> C is a F'rechet space with respect to
the topology of uniform convergence on compact subsets of 1lJ.
Let da be normalized area measure on 1lJ. For each a > -1,
dao(z) := (a + 1) (1 -lzI 2 )O da(z)
is a probability measure on 1lJ. For each 0 < p < 00, the corresponding weighted
Bergman space is defined to be
A~ := 1i(1lJ) n P(ao ).
A~ is closed in LP (a 0); it is a Banach space if p 2:: 1 and a p - Banach space if
o < p < 1. Its (p- ) norm will be denoted by II . Ilo,p.
A~ is a Hilbert space and has a reproducing kernel:
Ko(z, w) = K(z, w)o+2;
here K(z,w) = (1 - ZW)-l is the reproducing kernel for the Hardy space H2.
For reasons like this, the scale of Hardy spaces is often considered as the scale of
weighted Bergman spaces which corresponds to a = -1. Some of the results below
actually remain true for this case, and some can even be extended to analytic Besov
spaces B~ (f E B~ {::} f' E A~+p). Nevertheless, in this paper we will only deal
with the case -1 < a < 00.

3. Carleson measures
All measures on 1lJ will be finite, positive Borel measures. Let -1 < a < 00 and
o < p, q < 00 be given. We say that a measure /L on 1lJ is an (a, p, q) - Carleson
measure if A~ c Lq(/L) and the embedding A~ '---+ Lq(/L) is continuous: there is a
constant C> 0 such that IIfIILq(~) ::; C lIfIIA~ "If E A~.
Given an (a, p, q) - Carleson measure, the canonical embedding I : A~ -> Lq(/L)
will be referred to as a Carleson embedding.
As mentioned in the introduction, a number of the results to follow remain
true if we just require that f 1-+ f induces a bounded linear map A~ -> Lq(/L).
Also, complex measures whose variation is (a, p, q) - Carleson can be incorporated.
Moreover, there are extensions to analytic functions of several variables. However,
we are not going to discuss such generalizations in this paper.
We say that /L is a compact (a,p, q) - Carleson measure if the embedding
A~ '---+ Lq(/L) exists and is compact.
For example, an a.e. positive function h E Lq(a/3) defines the bounded mul-
tiplier Mh : A~ -> Lq(a/3) : f 1-+ fh iff the measure h q da/3 is (a,p, q) - Carleson.
Moreover, discrete (a, p, q) - Carleson measures on 1lJ can be defined using appro-
priate versions of 'sampling sequences', etc.
CARLESON EMBEDDINGS FOR WEIGHTED BERGMAN SPACES 219

An important example is obtained by looking at the composition operator C", :


J f-+ J 0 'P induced by a non-constant analytic function 'P : 10 ---> 10. Clearly,
C",: A~ ---> A~ exists iff a o o'P- 1 is (o:,p,q)-Carleson.
More generally, an arbitrary measure J.l on 10 is (0:, p, q) - Carleson if and only if,
for every analytic map 'P : 10 ---> 10, CI{J maps A~ boundedly into Aq(J.l) := 1i(1O) n
Lq(J.l). In fact, the condition applied to the identity of 10 shows that J.l is (o:,p,q)-
Carleson. On the other hand, if J.l is (0:, p, q) - Carleson and 'P : 10 ---> 10 is analytic,
then CI{J : A~ ---> Aq(J.l) is well-defined and bounded. For non-constant functions 'P,
the condition is further equivalent to J.l 0 'P -1 being (0:, p, q) - Carleson. This allows
an interpretation of Carleson embeddings, and in particular of multipliers as above,
as composition operators. However, in such a general setting the range space of
a composition operator might be unpleasent, and desirable properties may not be
available. For example, Aq(J.l) embeds continuously into 1i(1O) if and only if Aq(J.l)
is a closed subspace of Lq(J.l) and all point evaluations Aq(J.l) ---> C : J f-+ J(z),
z E 10, are continuous.
(0:, p, q) - Carleson measures have been characterized, even in a more general set-
ting, by V.L. Oleinikov and B.S. Pavlov [15], W.W. Hastings [6] and D.H. Luecking
[11],[12].
The hyperbolic metric on 10 is given by

e(z, w) := l~f
. lJd(J
"I 1 _ J(J

where the infimum extends over all smooth curves 'Y in 10 joining z and w. For
w E 10 and r > 0, let
Br (w ) = {z E 10 : e( z, w) < r}
be the corresponding hyperbolic disk. Actually, the particular choice of r > 0
doesn't really matter in our context.
Let us agree to write A~ '---+ Lq(J.l) if A~ is a subset of U(J.l) and the embedding
is continuous. Similar for other function spaces.
The Carleson measures under consideration can be characterized in terms of
the function

THEOREM 3.1. Let -1 < 0: < 00 and 0 < p, q < 00 be given.


(a) IJp::; q, then A~ '---+ Lq(J.l) iJ and only iJ Ho,p,q(w) is bounded on 10.
(b) IJp> q, then A~ '---+ Lq(J.l) if and only if Ho,p,q E L-/!!q(A).

Here dA(z) = da(z)(I-JzJ2)-2 is the Mobius invariant measure on 10. In [12]


Luecking presents an interesting proof of (b) which is based on the inequalities of
Khinchin and Kahane for Rademacher functions (see e.g. [2]).
It is well-known (compare K. Zhu [22]) that there is a constant C = C(o:, r) > 0
such that
1
C . ao(Br(w)) ::; (I-JwJ2)o+2 ::; C ao(Br(w)) 'Vw E 10 .
Therefore we may also say that Theorem 3.1 refers to properties of the function
w f-+ J.l(Br (W))I/ q(1 _JWJ2)-(o+2)/p .
220 HANS JARCHOW AND URS KOLLBRUNNER

It also follows that Ha,p,q E L~(A) if and only if J-t(B r ())/C7 a (B r ()) is in
LP/(p-q) (C7 a ). This will be used in the proof of Theorem 4.3 below.
If p ~ q; then the relevant parameter in Theorem 3.1 is q (0: + 2)/p, whereas
for p > q and fixed 0:, dependence is on p / q.
As a first immediate consequence we may state:
COROLLARY 3.2. For any -1 < 0: < 00,0 < p,q < 00 and t > 0, A~ "---> Lq(J-t)
if and only if A~ "---> Ltq(J-t).

In turn, this leads to:


COROLLARY 3.3. Let -1 < 0:,0:' < 00 and 0 < p, pi, q, q' < 00 be given.
(a) Ifp ~ q, pi ~ q' and q. (0: + 2)/p = q'. (0: ' + 2)/p', then A~ "---> Lq(J-t) iff
A~, "---> Lq' (J-t).
(b) Ifp> q and p/q = pi /q', then A~ "---> Lq(J-t) iff A~ "---> Lq' (J-t).
For a large range of parameters, this allows a reduction to Hilbert spaces as
follows:
COROLLARY 3.4. Suppose that -1 < 0:,0:' < 00 and 0 < p, q < 00.
(a) Ifp ~ q and 0:' + 2 = q. (0: + 2)/p, then A~ "---> Lq(J-t) iff A~, "---> L 2 (J-t).
(b) If p > q and pi /2 = 2/q' = p/q, then A~ "---> Lq(J-t) iff A~ "---> Lq' (J-t) iff
A~ "---> L 2 (J-t).
A special known case occurs when we take J-t = C7fJ for some (3 > -1 (see C.
Horowitz [7]).
COROLLARY 3.5. Suppose that -1 < 0:, (3 < 00 andO < p, q < 00.

(a) If p ~ q, then A~ "---> A~ iff (0: + 2)/p ~ ((3 + 2)/q.


(b) If p > q, then A~ "---> A~ iff (0: + l)/p < ((3 + l)/q.
There are several ways to modify the domain space of a composition operator.
In a systematic fashion, we may proceed as follows; cf. [4]. Each of the kernel
functions K a ( Z, .) is bounded (z E lV), and
._ ( 1 - IzI2 ) (a+2)/p
ka,p,z(w),- (1 _ ZW)2

has (p-) norm one in A~ (0 < p < 00). The functions f E A~ which admit a
representation
00

f(w) = L an ka,p,zn(w) Vw E lV ,
n=l
where the scalars an satisfy En lanl < 00 and the Zn's are taken from lV, form a
linear space, say

This is a Banach space with norm


00

IlfIIA~) := inf {L lanl: (*) holds} .


n=l

In fact, the map e1(lV) --+ A~) : (az)zEllJ 1--4 EZEllJ azka,p,z is a metric surjection.
CARLESON EMBEDDINGS FOR WEIGHTED BERGMAN SPACES 221

It is immediate that
if 0 <p,q < 00 and -1 < a,/3 < 00 are such that (a+2)/p = (/3+2)/q,
then A~) and A~q) coincide.
Atomic decomposition is available for weighted Bergman spaces (e.g. [10], [1],
[22]), hence
A~ S:! A~l) (with equivalent norms).
In particular, if a> p-2, then A~), alias Ab with /3+2 = (a+2)/p, is isomorphic
to fl.
Moreover:
If p 2: 1 then A~) '---+ A~ (boundedly and densely).
If 0 < p < 1, then A~ A~) = Ab (boundedly and densely).
'---+

In fact, it was shown by J.H. Shapiro [19] that in the latter case Ab is the Banach
envelope of A~, that is, the Banach space which is obtained by completing A~ with
respect to the biggest norm which is smaller than the given p-norm.
PROPOSITION 3.6. Let -1 < a < 00 and 0 < p ::; q < 00, q 2: 1. The following
are equivalent:
(i) J.L is an (a,p, q) - Carleson measure.
(ii) A~) '---+ Lq(J.L) (boundedly).
(iii) SUPzEV IIko,p,zlbu.t) < 00.
PROOF. (i) =? (ii) is obvious if p 2: 1 and follows from the above result of J.H.
Shapiro [19] if p < 1 ::; q: A~) is the Banach envelope of A~, and so the convex
hull of BAP is dense in B A(p), (ii) =? (iii) is trivial, and (iii) =? (i) is immediate from
the followi~g estimate in ;hich C is a constant depending only on 0 < r < 1:
J.L(B r (W))l/ q 1
(1 - IwI 2){o+2)/p = (1 -lwI 2){o+2)/p . Br{w) dJ.L(z)
(1
)l/q

- (1
- -
I
W
12){o+2)/p
.
(1 Br{w)
1 d ( )) l/q
(1 -lwI2)2q{o+2)/p J.L Z

<
-
C ( r1(1(1 -- wz)2{o+2)/p I
lv
IwI 2){o+2)/p qd Z) l/q
J.L()
= C lIko,p,wIILq{~).
Now apply Theorem 3.1(a). o
Actually, in the last step, no restriction on p and q is needed: Ho,p,q(w) is
bounded whenever A~) '---+ Lq(J.L).
There is another interesting consequence of Theorem 3.1, Corollary 3.2, and
Proposition 3.6:
COROLLARY 3.7. If 1 ::; q < p < 00 and (p/q) - 2 < a < 00, then Ho,p,q(w) E
L~(A) implies Ho,p,q(w) E LOO(A).
PROOF. Define a' > -1 by a' + 2 = q(a + 2)/p. Then, with T = T(A),
Ho,p,q E Lpq/{p-q) <=> A~ '---+ Lq(J.L) =? A~/q) = A~, '---+ L 1 (J.L)
<=> H O ',l,l E L oo <=> Ho,p,q E L oo . o
222 HANS JARCHOW AND URS KOLLBRUNNER

EXAMPLE 3.8.
(a) Let a, p, q be as in Corollary 3.7, let (an) be a sequence in foo \ fP/(p-q) ,
and let (1]n) be a sequence in llJ such that (!(1]n,1]k) :::: r 8nk for all n, k,
and such that llJ = Un Br(1]n). Consider the measure JL = En bn 811n on llJ
where bn = lanl' (1-I1]nI 2 )q("'+2)/p and note that (b n ) E fl. A calculation
reveals that H""p,q is in LOO(A) but not in uq/(p-q) (A). We conclude that
the converse in Corollary 3.7 doesn't hold.
(b) In Proposition 3.6, (ii) {::} (iii) is true for arbitrary 0 < p, q < 00, and
(i) ~ (ii) holds trivially whenever q :::: 1. But (iii) ~ (i) fails for 1 ~ q < p
and a + 2 > p. In fact, if JL is as in (a), then A~ 'f+ Lq(JL) since H""p,q tf-
Lpq/(p-q)(A), but H-y,l,q = H""p,q E LOO(A) if we put "I = (a + 2)/p - 2.
From A~) = A~l) = A~ we conclude that A~ '----> Lq(JL).

4. Compactness
We shall frequently make use of the following classical result:
THEOREM 4.1 (Pitt's Theorem). If 0 < p < q < 00 then every operator
f q --+ fP is compact.
This was obtained in 1936 by H.R. Pitt [16] for p :::: 1. For an extension see
H.P. Rosenthal [17]. The result as stated was proved recently by E. Oja [14].
By atomic decomposition, A~ and fP are isomorphic (see [10], [1]). Combining
this with Pitt's theorem we see that in particular the embedding in Corollary 3.5.(b)
is compact. It will follow from the next theorem that the embedding in 3.5.(a) is
compact iff (a + 2)/p < (f3 + 2)/q.
Our characterization of compactness of Carleson measures splits into two parts.
We consider the case p ~ q first. Here the characterization is as expected:
THEOREM 4.2. Let -1 < a < 00 and 0 < p ~ q < 00 with 1 ~ q. The following
statements are equivalent:
(i) JL is a compact (a,p, q) - Carles on measure.
(ii) A~) '----> Lq(JL) compactly.
limlllk""p,zIILq(ll)
(iii) z-+ = O.
(iv) lim Ho.,p,q(w) = O.
Iwl-+l
PROOF. (i) ~ (ii): If p :::: 1 then nothing is to prove since A~) '----> A~. If
p < 1 ~ q then, by [19], A~) is the Banach envelope of A~, and the convex hull
of B A~ is dense in B A!:)' Hence relative compactness of B A~ in the Banach space
Lq(JL) entails relative compactness of B A<.!) in Lq(JL). - Here we have used Bx to
denote the unit ball of a (quasi-) Banach space X.
(ii) ~ (iii): Suppose that (iii) doesn't hold. Then there exist an c > 0 and
a sequence (zn) in llJ such that limn--+ oo IZnl = 1 and Ilko.,p,zn IIL-(Il) > c for all
n. By (ii), we may assume that (k""p,zn)n converges to some f E Lq(JL). But
lim n -+ oo IZn I = 1 implies f = 0 since clearly (k""p,zn) tends to zero pointwise:
contradiction.
(iii) ~ (iv): The estimate proved in (iii) ~ (i) of Proposition 3.6 provides us
with a constant C = C(r) > 0 such that JL(Br(W))l/q ~ C, (1 - IwI 2 )(",+2)/p .
IIk""p,w IIL-(Il) for all W E llJ.
CARLESON EMBEDDINGS FOR WEIGHTED BERGMAN SPACES 223

(iv) =} (i): We apply Lemma 4.3.6 of K. Zhu [22]: there exists an integer N
such that for sufficiently small r there is a sequence (TJn) in llJ having the following
properties:
(1) llJ = U~=l Br(TJn),
(2) B r / 4 (TJm) n B r / 4 (TJn) = 0 whenever m -# n,
(3) Every Z E llJ is contained in at most N of the sets B 2r (TJn)'
Note that limn-+oo ITJnl = 1 follows from (2).
Let now (In) be a sequence in BA~' By Montel's theorem, some subsequence
(Ink) converges uniformly on compact sets to some I E 1i(llJ). By Fatou's lemma,
even I E BA~'
Put gk := I - Ink' Vk E N. By our hypothesis there exists, for any given c > 0,
an integer ne such that J.L(Br(TJn)) :::; c' (1 - ITJnI 2)q(a+2)/p if n ~ ne' Therefore,
with constants depending only on the indicated parameters,

n~, Lr(T/n) Igk(ZW dJ.L(z) = n~, L,.(T/,.) (lgk(ZW)q/PdJ.L(z)


< C(r) n~E Lr(T/n) COa(B:r(TJn)) L2,0(T/n) IgdwWdaa(w)) q/p dJ.L(z)

n~nE
r
< C(r) C(r, 0)' "~ (1 _IJ.L(B 2(17n))
)q(a+2)/p
TJn
1
(1 B 2r (T/n)
p
Igk(W)1 daa(w)
)q/P

< C(r)C(r,o)c L ({
n~nE J B2r(T/n)
Igk (wWda a (w)r/ p

< C(r) C(r, 0) . c ( L1


n~n. B2r(T/n)
Igk(wWdaa(w) riP

< C(1')' C(r, 0) . Nq/p . c (L Igk (wWda a (w) ) q/p :::; c c ;


here C = C(r,o,N,p,q).
If we choose now ke EN such that Ln<nE IBr(T/,,) IgklPdJ.L :::; c for k ~ ke, then
Iv Igk Iq dJ.L :::; (C + 1) . c for such k, and we may conclude that (gk) is a null sequence
in Lq(J.L). 0

A slight modification of the argument used to prove (i) =} (ii) shows that com-
pactness of A~ <--+ Lq(J.L) implies compactness of A~, <--+ Lq(J.L) if -1 < 0,0' < 00,
0< p < p' :::; q and (0 + 2)/p = (0' + 2)/p'.
It can be shown that (i) {::} (iii) {::} (iv) is true even for arbitrary 0 < q < 00.
In the case p > q, we can prove:
THEOREM 4.3. Suppose that -1 < 0 < 00 and 0 < q < p < 00. Regardless 01
the (0, p, q) - Carleson measure J.L, the embedding A~ <--+ Lq (J.L) is always compact.
For composition operators C<p : A~ --+ Ah this is due to W. Smith and L. Yang
[21]; see also [4].

PROOF. Put s = p/(p - q) and recall that J.L(Br(-))/aa(Br ()) E B(aa) when-
ever J.L is (0, p, q) - Carleson.
224 HANS JARCHOW AND URS KOLLBRUNNER

Let Un) be a bounded sequence in A~. For some subsequence, f = limk-+oo fnk
exists in 'H(lU). As before, f E A~, hence we may assume that Un) is a null sequence
in 'H(lU) and that IIfnllo,p :::; 1 for all n.
There are constants Cr , Cr > 0 such that, for each n,

Inf1lJ Ifnlq dp. < Cr' f


l1lJ a o
(;r (W)) lBr(w)
f Ifn(zWdao(z) dp.(w)

= Cr' flao(B~(w)) I Br (w)(z) Ifn(zW dp.(w) dao(z)


(4.1)
= Cr' f Ifn(zW lBr(z)a
f (B1 ( )) dp.(w) dao(z)
l1lJ rW O

- f q p.(Br(z))
< Cr' Cr" l1lJ1fn(z) I ao(Br(z)) dao(z) .
. f p.(Br(zW
By hypothesIs, l1lJ ao(Br(Z))8dao(z) < 00.

Therefore, given c > 0, there is an r E E (0,1) such that

(4.2) 11lJ\r 1lJ


E
_
rZ
p.(Br(zW
ao
( ) ( / )8
(B ( ))8 da o z < c 2 .

Un) tends to zero uniformly on rEV, so that


limn-+oo Ir,v Ifn (z)IPda o (z) = o. Ac-
cordingly, we may choose n E EN such that, for n 2: n E ,

(4.3)

(4.2) c
< -
2

Combine (4.1), (4.4) and (4.5) to find a constant Cr > 0 such that I1lJ Ifnlq dp. :::; Crc
for n 2: ne:' 0
Essential parts of Theorem 4.3, if not the entire theorem, can be proved by
using other methods. We sketch three possibilities:
Suppose that -1 < Q < 00, 1 :::; q :::; 2 < P < 00 and v is any measure.
Every operator u : A~ -+ Lq(v) is compact.
In fact, by Kwapien's theorem (see [8], or [2], 12.19), the operator u
e
admits a factorization u : A~ ~ 2 .2:. Lq(v). By atomic decomposition
and Pitt's theorem, w is compact, and so is u.
CARLESON EMBED DINGS FOR WEIGHTED BERGMAN SPACES 225

Rosenthal's extension [17J of Pitt's theorem admits the same conclusion


for q ~ 1 and p > max{2, q}.
Suppose that -1 < a < 00, 1 < p < 00, and pis (a,p, 1) - Carleson. Then
I : A~ '--+ 1 (p) is compact.
Since A~ is reflexive it suffices to verify that I is completely continuous.
Accordingly, let (fn)n be a weakly null sequence in A~. Then fn(z) ----t 0
for each z E llJ. Being weakly null in L1(p), (fn) is uniformly integrable.
Now limn ..... oo IIfnl11 = 0, by a theorem of Vitali (see W. Rudin [18], p.133).
Standard results from interpolation theory on the preservation of compactness
by interpolated operators lead from either of these special results to (at least parts
of) Theorem 4.3.

5. Order bounded and absolutely summing operators


Our Banach lattices will be complex Banach lattices; see e.g. P. Meyer-Nieberg
[13J for the construction of such an object from a real Banach lattice. Let X
be a Banach space and Y a closed subspace of a Banach lattice L. An operator
u : X ----t Y is called order bounded if there is a non - negative h E L such that
lufl :::; h for f in Bx, the unit ball of X. Thus we require u to map Bx into the
order interval {g E L : Igl :::; h} of L. Note that L is part of the definition! Every
1.L E C(X, Y) is order bounded when Y is considered as a subspace of C(K) for
some compact Hausdorff space K.
Let I be an order interval in the Banach lattice L. Its span, Z, is a Banach
lattice with respect to L's order and (a multiple of) I's gauge functional as its
norm. Z is an abstract M - space with unit and so, by a well-known theorem of S.
Kakutani, isometrically isomorphic (as a Banach lattice) to C(K) for some compact
Hausdorff space K; see again [13J.
It follows that every order bounded operator u : X ----t Y c L factorizes X ~
Z ~ C(K) L L where K is as before and j is the canonical embedding.
In this paper, L will be a space LP(p) which results in close ties with absolutely
summing operators. Recall that a Banach space operator u : X ----t Y is (q,p)-
summing (p:::; q), written u E IIq.p(X, Y) , if there is a constant C such that, for
every choice of n E N and x!, . . . ,X n EX,

In other words, u is (q,p)-summing iff every weak eP-sequence, i.e. every se-
quence (xn) in X which satisfies 2:::=11(x*,x n )iP < 00 for all x* E X*, is taken to
a strong eq - sequence, i.e. 2:::=1 Iluxnllq < 00 holds.
(p, p) - summing operators are called p - summing; the corresponding notation
is IIp(X, Y) = IIp,p(X, Y) .
We refer to [2J for details on these concepts and in particular for the following
facts:
If Hand K are Hilbert spaces and q ~ 2, then IIq,2(H, K) is the corre-
sponding Schatten q - class. Moreover, for any 1 :::; p < 00, IIp(H, K) is
the class of Hilbert - Schmidt operators.
If 1 :::; P :::; 2, then every operator from C(K) to LP(lI) is 2 - summing.
226 HANS JARCHOW AND URS KOLLBRUNNER

If p > 2, then every operator C(K) ----7 U(v) is (p, 2) -summing, and
r - summing for every r > p.
Moreover:
If u : X Lp(v) is order bounded then u is p - summing.
----7

Here v is any measure. In the last statement, the converse fails. But:
If u* p - summing then u is order bounded.
More precisely, we have the following result due to D.J.H. Garling [5]:
Let 1 ::; p < 00. A Banach space operator u : X ----7 Y has a p - summing
adjoint if and only if, for every measure v and operator v : Y ----7 LP(v),
the composition v 0 u : X ----7 LP(v) is order bounded.
In particular:
An operator u : L 2 (VI) ----7 L 2 (V2) is order bounded iff it is Hilbert-
Schmidt.
We are going to characterize order boundedness of Carleson embeddings A~ ~
Lq(J.L). To this end we introduce, for s > 0, the Banach space
Xs := {f: lU ----7 C: f measurable, sup(1-lzI 2 )Slf(z)1 < oo} .
zE1U

and its closed subspace


Xs := Xs n H(lU) .
It is easy to see that A~ ~ X(o.+2)/p and that the index (o.+2)/p is best possible.
THEOREM 5.1. Let -1 < 0. < 00, 0 < p < 00 and 1 ::; q < 00. Then, with
s := (0. + 2)/p, the following statements are equivalent:
(i) A~ ~ U(J.L) order boundedly.
(ii) A~) ~ U(J.L) order boundedly.
(iii) (1-lzI 2 )-S E U(J.L).
(iv) XS ~ Lq(J.L) boundedly.
(v) XS ~ Lq(J.L) order boundedly.
(vi) XS ~ Lq(J.L) boundedly.
(vii) XS ~ Lq(J.L) order boundedly.
PROOF. (i)::::} (ii) is obtained as before, by considering separately the cases
p ;:::: 1 and p < 1. In order to prove (ii)::::} (iii) it suffices to look at the functions
ka,p,q' (iii)::::} (iv) and (iv)::::} (vi) as well as (iv) <=?{v) and (vi) :} (vii) are easily
verified. Finally, for (vii)::::} (i), just observe that A~ ~ Xs. 0

Various statements related to boundedness of Carleson embeddings do have


'order bounded counterparts'. The first example is:
COROLLARY 5.2. Suppose that -1 < 0.,0.' < 00,0 < p,p' < 00 and 1::; q,q' <
00 are such that q. (0. + 2)/p = q'. (0.' + 2)/p'. Then

A~ ~ U (J.L) order boundedly :} A~, ~ U' (J.L) order boundedly.


Again, in many cases, reduction to Hilbert spaces is possible.
CARLESON EMBEDDINGS FOR WEIGHTED BERGMAN SPACES 227

COROLLARY 5.3. If -1 < a, ci < 00, 0 < p < 00 and 1 :::; q < 00 are such that
a' +2 = q. (o+2)/p, then A~ '---+ Lq(f.-L) order boundedly if and only if A~, '---+ L2(f.-L)
exists as a Hilbert - Schmidt operator.

A special case occurs when f.-L is a measure U'(3. From (iii) of Theorem 5.1 we
conclude:

COROLLARY 5.4. If -1 < a,(3 < 00, 0 < p < 00 and 1 < q < 00, then
A~ '---+ A~ order boundedly if and only if (a + 2)/p < ((3 + l)/q .

In such case, we can even factorize: A~ '---+ Xs '---+ A~ where s = (a + 2)/p.


- There is of course no problem in verifying this corollary directly.
If (a + 2)/p < ((3 + l)/q, then A~ '---+ A~ is order bounded and so q-summing;
here -1 < a, (3 < 00 and 1 :::; p, q < 00. In many cases, the converse (which doesn't
hold for general Banach space operators) is true for Carleson embeddings. To see
why, we need to look at special sets. With each z E U, we associate the 'interval'
in au:

I(z) := {I:I ei9 : -11"(1 -Izl) :::; () :::; 11"(1 - IZI)} ,

and the 'squares' in U:

R(z) := {w E U: Izl < Iwl < 1, (w/lwl) E I(z)}

and

Q(z) := {w E R(z) : Iwl < (1 + Izl)/2} .


It can be shown that the sets Q(Tln) form a partition of U whenever (TIn) is a
[sufficiently fine] sequence of points in U having the properties (1) - (3) listed in
part (iii) =} (iv) of the proof of Theorem 4.2.

PROPOSITION 5.5. Let f.-L be a (a, p, q) - Carles on measure with -1 < a < 00
and 1 < p, q < 00.
If p* :::; q < 00, then A~ '---+ L q(f.-L) is order bounded iff it is q - summing iff it is
(q, p*) - summing.

This is due to T. Domenig [3] for composition operators acting between weigh-
ted Bergman spaces. For the sake of completeness, we sketch a proof of the propo-
sition which follows closely Domenig's arguments.

PROOF. If I : A~ '---+ Lq(f.-L) is order bounded, then it is q-summing and so


(q,p*)-summing.
Suppose now that I is (q, p*) - summing. By standard - but lengthy - calcula-
tions it can be shown that the functions
228 HANS JARCHOW AND URS KOLLBRUNNER

form a weak ff -sequence in A~. Hence we get from our hypothesis that, with a
suitable constant C,

00 > L llvn(zW d/-L(z) 2 L 1


n IlJ n Q(T/n)
Ivn(zW d/-L(z)

L [ ((1 -11JnI 2 )l/p )q(a+2)

n 1Q (T/n) 11 -1Jn z l

> C L1 (1 -lzI 2)-q(<>+2)/Pd/-L(z) .


n Q(T/,,)

Thus (1 -lzI 2 )-(<>+2)/p E Lq(/-L), and so ] is order bounded by Theorem 5.1. 0


More is available. Consider the Rademacher functions
Tn : [0, 1] ~ 1R : t f-+ sign sin (2nrrt) , n E N
(or any sequence of independent symmetric Bernoulli variables). Given 0 < p < 00,
Khinchin's inequality assures the existence of positive constants Ap and Bp such
that, for any finite collection of scalars al,"" an:

L 10 IL akTk(t) I dt
1/2 ::; ([1 L 1/2 .
n ) n P ) l/p ( n )
Ap' ( lakl 2 ::; Bp' lakl 2
k=l k=l k=l
Pursuing the fate of this inequality within the framework of Banach spaces leads to
the theory of type and cotype of Banach spaces, and to the following related class
of operators (compare [2], Chs. 11- 12).
A Banach space operator u : X -> Y is almost summing,
u E IIas(X, Y) ,
if there is a constant C such that, for any choice of finitely many vectors Xl, ... , Xn
from X,

(10[III {; rk(t)UXk 112 dt )1/2 ::; C x.~~x. (


n
{;
n
I(x*, xk)1 2
)1/2
It is known that each of the operator ideals IIp is properly contained in IIas.
Moreover, if 1 ::; p < 00 and r = max{p,2} then IIash X) c II r ,2(', X) whenever
X is an P space, or the Schatten p-class Sp(H) for some Hilbert space H. In
addition, it was shown by S. Kwapien [9] that
if H is a Hilbert space and u is in IIas(H, Y) then the adjoint 11,* : Y* -> H
is 1 - summing.
See [2], p.255 for details.
We have the following application to Carleson embeddings. The argument is
the same as for composition operators between weighted Bergman spaces [3].
PROPOSITION 5.6. Let /-L be an (a,p, q) - Carles on measure where 1 ::; q < 00.,
and 2 ::; p < 00. The embedding] : A~ '--+ Lq(/-L) is almost summing if and only if
it is order bounded.
PROOF. Define "I > -1 by ("I + 2)/2 = (a + 2)/p. Since p 2 2, A~ '--+ A~.
:::::} ]: A~ '--+ Lq(/-L) is almost summing
:::::} ]* is I-summing (Kwapien)
:::::} ] is order bounded (Garling)
CARLESON EMBEDDINGS FOR WEIGHTED BERGMAN SPACES 229

Combining the preceding two propositions yields:

COROLLARY 5.7. Let -1 < ct < 00 and 1 ::; p, q < 00 be such that p 2:
min {q* ,2} and let 11, be an (ct, p, q) - Carleson measure. The embedding I : A~ "--->
Lq(/.l) is q - summing iff it is order bounded.

PROOF. Only sufficiency requires proof. If p 2: q*, then Proposition 5.5 settles
the case. And if p 2: 2, then I, being q - summing, is almost summing, and so order
bounded by Proposition 5.6. 0

References
[IJ R.R. Coifman, R. Rochberg, G. Weiss, Facto'rization theorems for Hardy spaces in seveml
variables. Ann. of Math. (2) 103. (1976),611-635.
[2J J. Diestel, H. Jarchow, A. Tonge: Absolutely Summing Opemtors. Cambridge University
Press 1995.
[3J T. Domenig, CompoS'ition opemtors on weighted Beryman sp(,ces and Hardy spaces. Disser-
tation University of Zurich 1997.
[4J T. Domenig, H. Jarchow, R. Riedl, The domain space of an analytic composition opemtor.
Journ. Austral. Math. Soc. 66 (1999), 56-65.
[5J D.J.H. Garling, Lattice bounding, Radonifying and summing mappings. Math. Proc. Camb.
Phil. Soc. 77 (1975), 327-333.
[6J W.W. Hastings, A Carleson measure theQrem for Beryman spaces. Proc. Amer. Math. Soc.
52 (1975), 237-241.
[7J C. Horowitz Zeros of functions in the Bergman spaces. Duke Math. Journ. 41 (1974), 693-710.
[8J S. Kwapien, On a theorem of L. Schwartz and its applications to absolutely summing opem-
tors. Studia Math. 38 (1970), 193-201.
[9J S. Kwapien, A remark on p - summing opemtors in fr - spaces. Studia Math. 34 (1970), 277-
278.
[lOJ J. Lindenstrauss, A. Pelczynski, Contributions to the theory of classical Banach spaces.
Journ. Funct. Anal. 8 (1971), 225-249.
[l1J D.H. Luecking, Multipliers of Bergman spaces into Lebesgue spaces. Proc. Edinb. Math. Soc.
29 (1986), 125-131.
[12J D.H. Luecking, Embedding theorems for spaces of analytic functions via Khinchine's inequal-
ity. Mich. Math. Journ. 40 (1993), 333-358.
[13J P. Meyer-Nieberg, Banach Lattices. Springer-Verlag 1991.
[14J E. Oja, Pitt Theorem for non-locally convex spaces f p Preprint.
[15J V.L. Oleinik, B.S. Pavlov, Embedding theorems for weighted classes of harmonic functions.
Zap. Nauchn. Sem. Leningrad Otdel. Mat. Inst. Steklov 22 (1971),94-102. Transl. in Journ.
Soviet Math. 2 (1974), 135-142.
[16J H.R.Pitt, A note on bilinear forms. Journ. London Math. Soc. 11, 171-174 (1936).
[17J H.P. Rosenthal, On quasi-complemented subspaces of Banach spaces with an appendix on
compactness of opemtors from LP(p,) to Lr(/I). Journ. Funct. Anal. 4 (1969), 176-214.
[18J W. Rudin, Real and Complex Analysis. 3 rd ed., McGraw-Hill 1987.
[19J J.H. Shapiro, Mackey topologies, reproducing kernels, and diagonal maps on the Hardy and
Bergman spaces. Duke Math. Journ. 43 (1976), 187-202.
[20J W. Smith, Composition opemtors between Bergman and Hardy spaces. Trans. Amer. Math.
Soc. 348 (1996) 2331-2348.
[21J W. Smith, L. Yang, Composition opemtors that impro1Je integmbility on weighted Beryman
spaces. Proc. Amer. Math. Soc. 126 (1998) 411-420.
[22J K. Zhu, Opemtor Theory in Function Space.~. Marcel Dekker, New York 1990.
230 HANS JARCHOW AND URS KOLLBRUNNER

INSTITUT FUR MATHEMATIK, UNIVERSITAT ZURICH, WINTERTHURERSTRASSE 190, CH 8057


ZURICH, SWITZERLAND
E-mail address:jarchowlDmath.unizh.ch

INSTITUT FUR MATHEMATIK, UNIVERSITAT ZURICH, WINTERTHURERSTRASSE 190, CH 8057


ZURICH, SWITZERLAND
E-mail address:kollbrunlDmath.unizh.ch
Contemporary Mathematics
Volume 328, 2003

Weak* -extreme points of injective tensor product spaces

Krzysztof Jarosz and T. S. S. R. K. Rao

ABSTRACT. We investigate weak* -extreme points of the injective tensor prod-


uct spaces of the form A . E, where A is a closed subspace of C (X) and E
is a Banach space. We show that if x E X is a weak peak point of A then
f (x) is a weak*-extreme point for any weak*-extreme point f in the unit ball
of A . E C C (X, E). Consequently, when A is a function algebra, f (x) is a
weak*-extreme point for all x in the Choquet boundary of A; the conclusion
does not hold on the Silov boundary.

1. Introduction
For a Banach space E we denote by E1 the closed unit ball in E and by BeE1
the set of extreme points of E 1 . In 1961 Phelps [16] observed that for the space
C(X) of all continuous functions on a compact Hausdorff space X every point f in
Be (C (X))1 remains extreme when C (X) is canonically embedded into its second
dual C (X)**. The question whether the same is true for any Banach space was
answered in the negative by Y. Katznelson who showed that the disc algebra fails
that property. A point x E OeE1 is called weak* -extreme if it remains extreme in
BeEi*; we denote by B;E1 the set of all such points in E 1. The importance of this
class for geometry of Banach spaces was enunciated by Rosenthal when he proved
that E has the Radon-Nikodym property if and only if under any renorming the
unit ball of E has a weak* -extreme point [19].
While not all extreme points are weak* -extreme the later category is among
the largest considered in the literature. For example we have:
strongly exposed S;; denting S;; strongly extreme S;; weak* -extreme.
We recall that x E E1 is not a strongly extreme point if there is a sequence Xn
in E such that Ilx xnll ---t 1 while IIxnll ~ 0 (see [3] for all the definitions). We
denote by O;E1 the set of strongly extreme points of E 1. It was proved in [14]
that e E O;E1 if and only if e E o;Ei* (see [9], [13], or [17] for related results).
Examples of weak* -extreme points that are no longer weak* -extreme in the unit
ball of the bidual were given only recently in [6].
In this paper we study the weak* -extreme points of the unit ball of the injective
tensor product space A,E, where A is a closed subspace of C(X). Since C(X),E

Both authors were supported in part by a grant #0096616 from DST/INT/US(NSF-


RP041)/2000.

2003 American Mathematical Society


231
232 KRZYSZTOF JAROSZ AND T. S. S. R. K. RAO

can be identified with the space C(X, E) of E-valued continuous functions on X,


equipped with the supremum norm, elements of A e E can be seen as functions on
X. We are interested in the relations between 1 E a; (A E) and 1 (x) E a;E1 ,
for all (some?) x E X. Since any Banach space can be embedded as a subspace A of
a C (X) space no complete characterization should be expected in such very general
setting. For example if A is a finite dimensional Hilbert space naturally embedded
in C (X), with X = Ai, and dim E = 1, then any norm one element 1 of A E is
obviously weak* -extreme however the set of points x where 1 (x) is extreme is very
small consisting of scalar multiples of a single vector in Ai. Hence in this note we
will be primarily interested in the case when A is a sufficiently regular subspace of
C (X) and/or x is a sufficiently regular point of X.
It wa..<; proved in [5] that
(1.1) 1 E a:C(X, Eh {::::::::} [I (x) E a:EI , for all x EX].
It follows from the arguments given during the proof of Proposition 2 in [6] that
for a function 1 E (A E)l we have
[I (x) E a;El for all x E X with c5 x E aeAi] => 1 E a; (A e E)I'
where we denote by c5x the functional on A of evaluation at the point x. In this
paper we obtain a partial converse of the above result (Theorem 1). Our proof also
shows that if 1 E a; (A E)l then 1 (x) E a;El for any weak peak point x (see
Def. 1), extending one of the implications of (1.1). It follows that when A is a
function algebra then any weak* -extreme point of Al is of absolute value one on
the Choquet boundary ChA (and hence on its closure, the Shilov boundary) and
consequently is a strongly extreme point [17]. Since we have concrete descriptions
of the set of extreme points of several standard function algebras (see e.g. [12], page
139 for the Disc algebra) one can give easy examples of extreme points that are not
weak* -extreme. Recently several authors have studied the extremal structure of the
unit ball of function algebras ([1], [15], [18]). It follows from their results that the
unit ball has no strongly exposed or denting points. Our description that strongly
extreme and weak* -extreme points coincide for function algebras and are precisely
the functions that are of absolute value one on the Shilov boundary completes that
circle of ideas. We also give an example to show that the weak* -extreme points of
(A , E)l in general need not map the Shilov boundary into aeE1 Considering the
more general case of the space of compact operators K (E, F) (we recall that under
assumptions of approximation property on E or F*, K (E, F) can be identified with
E* F) we exhibit weak*-extreme points T E K (fP)1 for 1 < p =I=- 2 < 00 for
which T* does not map unit vectors to unit vectors.
Our notation and terminology is standard and can be found in [3], [4], or [11].
We always consider a Banach space as canonically embedded in its bidual. By E(n)
we denote the n-th dual of E. By a function algebra we mean a closed subalgebra
of a C (X) space separating the points of X and containing the constant functions;
we denote the Choquet boundary of A by ChAo

2. The result
As noticed earlier, for A c C (X) and a point 1 E a;
(A Eh we may not
have 1 (x) E a;EI for all x E X even in a finite dimensional case. Hence we need
to define a sufficiently regular subset of X in relation to A.
WEAK' -EXTREME POINTS 233

DEFINITION 1. A point Xo E X is called a weak peak point of A C C (X) if for


each neighborhood U of Xo and > 0 there is a E A with 1 = a (xo) = Iiall and
la (xo)1 < for x E X\U; we denote by opA the set of all such points in X.

There are a number of alternative ways to describe the set opA. If Xo E X is


a weak peak point of A C C (X), /-l is a regular Borel measure on X annihilating
A, and al' is a net in A convergent almost uniformly to 0 on X\ {xo} and such
that a.., (xo) = 1 then /-l({xo}) = liml'Jxal' = O. Hence if a* E A* and VI,V2
are measures on X representing a* we have VI ({Xo}) = V2 ({xo}), consequently
V ~ v ( {xo}) is a well defined functional on A * .
On the other hand if X {xo} E A ** then /-l ( {xo}) = 0, for any annihilating mea-
sure /-l, and Xo is a weak peak point. To justify the last claim notice that Al is weak*-
dense in Ai* so X{xo} is in the weak*-closure of the set K = {f E Al : f (xo) = I}.
Let U be an open neighborhood of Xo and Ax\U be the space of all restrictions of
the functions from A to X\U. We define the norm on Ax\U as sup on X\U. Lct
K X\U be the set of restrictions ofthe functions from K and cl (Kx\U) be the norm
closure of Kx\U C Ax\U. If 0 ~ clKx\U then there is G E (Ax\U)* , represented
by a mcasure 1] on X\U and separating Kx\U from 0:

Re G (h) > a > 0 = X{xo} (/-l) I for all hE clKx\U


The measure 1] extends G to a functional on A so K is functionally separated from 0
in A contrary to our previous observation. Hence 0 E clKx\u so there is a function
in K that is smaller then outside U which means that Xo is a weak peak point.
The concept of weak peak points is well known in the context of function alge-
bras where opA coincides with the Choquet boundary ([8]' p. 58). For more general
spaces of the form Ao 'f1 {foa E C (X) : a E A} I where A C C (X) is a function al-
gebra and fo a nonvanishing continuous function on X we have ChA ~ opAo.
Spaces of these type appear naturally in the study of singly generated modules and
Morita equivalence bimodules in the operator theory [2J.

THEOREM 1. Let E be a Banach space, X a compact Hausdorff space, and A


a closed subspace of C (X). If f E A , E is a weak* (strongly) extreme point of
the unit ball then f (x) is a weak* (strongly) extreme point of the unit ball of E for
any x E opA.
In particular if f E (C(X,E))I is a weak*(strongly) extreme point then f(x)
is a weak* (strongly) extreme point of EI for all x EX.

We first need to show that for a weak peak point Xo E X there exists a function
in A not only peaking at Xo but that is also almost real and almost positive.

LEMMA 1. Assume X is (L compact Hausdorff space, A is a closed subspace of


C (X), and Xo is a weak peak point of A. Then for each neighborhood U of Xo and
> 0 there is g E A such that

Ilgll = 1 = g (xo) ,
(2.1) Ig (x)1 < , for all x E X\U, and

IIRe+ g - gil < ,


where Re+ z = max{O,Rez}.
234 KRZYSZTOF JAROSZ AND T. S. S. R. K. RAO

PROOF. Put U1 = U and let gl E A be such that


IIg111 = 1 = gl (xo) and Ig1 (x)1 < e for x i U1.
Put U2 = {x E U1 : Ig1 (x) - 11 < e} and let g2 E A be such that
IIg211 = 1 = g2 (xo) and Ig2 (x)1 < e for x i U2
Put U3 = {x E U2 : Ig2 (x) - 11 < e}. Proceeding this way we choose a sequence
{gn}n>l
_ in A. Fix a natural number k such that k> 1e: and put

1 k
g= k Lgj
j=l

We clearly have Ilgll = 1 = 9 (xo) and Ig (x)1 < e for x i U. Let x E U, then either
x belongs to all of the sets Uj , j :s
k, in which case Ig (x) - 11 < e, or there is a
natural number p < k such that x E Up \ Up+!. In the later case we have

Ig(X)_P~II=~ tgj-(P-l)
;=1

< .!. ( (lg1 (x) - 11 + ... + Igp-1 (x) - 11) )


- k + Igp (x)1 + (lgp+dx)1 + ... + Igk (x)!)
p-l 1 k-p
< -k- e + k + -k- e < e.
Hence IIRe+ 9 - gil < e. o
We are now ready to finish the proof of the Theorem.
PROOF. Suppose! (xo) is not a weak*-extreme point. Then by [9] there is a
sequence en in E1 and e* E Ei such that II! (xo) enll 1 + ~ and e* (en) ~ O. :s
By the Lemma there is a sequence gn in A such that
Ilgnll = 1 = gn (xo) ,

(2.2) ign (x)1 <.!.,


n
if II! (x) - ! (xo) II ~ .!.,
n
and

IIRe+ gn - gnll < .!.,


n
Hence
II! (x) 9 (x) e II < max { sUPllf(x)-f(xo)II~,* {II! (x)11 + Ign (x)llIenll}, }
n n - '*
sUPllf(x)- f(xo)lI< {II! (x) gn (x) en II}

:s max {I +.!.,.!.
n n
+ II! (xo) Re+ gn (x) en II + .!.}
n
3
<
- 1+-.
n
Therefore II! gnenll ---> 1 but (8 (xo) e*)(gnen) = gn (xo) e* (en) ~ O. This
contradiction shows that! (xo) is a weak*-extreme point.
The same line of arguments shows that! (x) is strongly extreme for any strongly
extreme! E (A , E)l' 0
Since for a function algebra A the Choquet boundary C hA coincides with 8p A
([8], p. 58) and the Shilov boundary 8A is equal to the closure of ChA we have:
WEAK' -EXTREME POINTS 235

COROLLARY 1. Let A be a function algebra, E a Banach space and f E (A e


Eh a weak* -extreme point. Then
f(x) E 8;E1 , for x E ChA, and Ilf (x)11 = 1, for x E 8A.
REMARK 1. Theorem 1 is not valid for the spaces WC(X, E) of E-valued con-
tinuous functions with E quipped with the weak topology. Even a strongly extreme
point of WC(X, Eh need not assume extremal values at all points of X [13].
We next give an example of a function algebra A and a 3-dimensional space
E showing that a weak* -extreme point f E (A e Eh need not take extremal
values on the entire Shilov boundary. Since E is finite dimensional this function f
maps the Choquet boundary into the set of strongly extreme points but f is not a
strongly extreme point.
EXAMPLE 1. Put
Q = {(z,w,O) E (:3: Izl2 + Iwl 2 ::; I} U {(O,w,u) E (:3: max{lwl, lui}::; I},

and B = convQ. Let 1111 be the norm on (:3 such that B is its unit ball. Note that
(z,w,O) is an extreme point ofB ifflwl =I 1 and Iz12+lw12 = 1. PutE = ((:3,1111),

X'!!:. {O} x {I} x lD>u {(sint,cost,O): 0::; t::; 11"},


df
fo : X -+ El, fo (x) = x, and
A = {h E C (X) : h (0, 1,) E A (lD>)},
where A (lD is the disc algebra. We have
ChA = {O} x {I} x 8lD> U {(sin t, cos t, 0) : 0 < t < 11"} .
The function fo is in A e E and takes extremal values on the Choquet boundary of
A so it is a weak* -extreme points of (A e E)l. However fo (0, 1, 0) = (0, 1, 0)
are not extreme points of Ei while (0, 1, 0) are in the Shilov boundary of A.
Since E is finite dimensional clearly the function fo maps the Choquet boundary
of A into the set of strongly extreme points of E 1 We next show that f is not a
strongly extreme point.
Let gn E A be such that

Ilgnll = 1 = gn (sin ~,cos ~,o) ,


gn (sint, cost, 0) = 0, for ~ < t::; 1, and
n
gn (0, 1, z) = 0, for z E lD>.
Put fn = (O,O,gn) E A e E. We have
(0,1, c) for (a,b,c) E {O} x {I} x 8lD>
(fo fn)(a, b, c) = { (a, b, gn) for (a, b, c) E {(sin t, cos t, 0) : 0 ::; t ::; 11"} .
Hence IIf fnll -+ 1 but Ilfnll ~ 0 so f is not a strongly extreme point.
In the next Proposition we consider a more general setting of compact oper-
ators. For a Banach space E we denote by C(E) the space of all linear bounded
maps on E, by K(E) the set of all compact linear maps, and by S(E) the set of
236 KRZYSZTOF JAROSZ AND T. S. S. R. K. RAO

unit vectors in E. Since K (E, C (X)) can be identified with C (X, E*) our result on
weak* -extreme points taking weak* -extremal values can be interpreted as follows
T E a;K(E,C(X))l ==} T* (a;C(X);) c a;Er.
Thus more generally one can ask whether T* (a; Ft) c a; Ei for any TEa; K(E, Fh.
We give a class of counter examples with the help of the following proposition.
PROPOSITION 1. Let E be an infinite dimensional Banach space such that K(E)
is an M-ideal in C(E). 1fT E K(Eh then T*(aeEi) ct. S(E*).
We recall that a closed subspace M of a Banach space E is an AI-ideal if there is
a projection P E C (E*) such that ker P = Ml. and liP (e*)II+lle* - P (e*)11 = Ile*ll,
for all e* E E* (see [11] for an excellent introduction to 1\/-ideals).

PROOF. Since qE) is an AI-ideal it follows from Corollary V1.4.5 in [11] that
E* has the Radon-Nikodym property and hence the IP (see [10]). Also since qE)
is a proper M-ideal it fails the IP. It therefore follows from Theorem 4.1 in [10]
a
that there exists a net {x~} c e Ei such that x~ ---> Xu in the weak* -topology with
Ilxoll < 1. Suppose T*(aeEi) c S(E*). Since T* is a compact operator by going
through a subnet if necessary we may assume that T*(x~) ---> T*(xo) in the norm.
Thus 1 = IIT*(xo)11 < 1 and the contradiction gives the desired conclusion. 0

EXAMPLE 2. Banach spaces E for which K(E) is an AI ideal in C(E) have been
well extensively studied. Chapter VI of [11] provides seveml examples including
e
E = p , 1 < p < 00, as well as properties of these spaces. It was observed in [6]
that for p # 2 there are weak*-extreme points in the space K(ePh. It follows from
the last proposition that the adjoint of these weak" -extreme points do not even map
extreme points to unit vectors.
A strongly extreme point remains extreme in all the dual spaces of arbitrary
even order. A weak* -extreme point remains extreme in the second dual but may
not be extreme in the fourth dual. Hence the property of remaining extreme in all
the duals of even order is placed between the strong and the weak* type of extreme
points. It would be interesting to describe that property in terms of the original
Banach space alone. A procedure for generating extreme points which have this
property but are not strongly extreme was described in [6].
PROPOSITION 2. Let X be a compact Hausdorff space, A a closed subspace of
C(X), and E a Banach space. Suppose Xo E X is a weak peak point and f E A. E
is an extreme point in the unit ball of all the duals of even order. Then f (xo) is
an extreme point of the unit ball of all the duals of E of even order.
PROOF. Since the space A . E** can be canonically embedded in (A . E)**
[7] we have, for any natural number n
A . E(2n) C (A ( E(2n-2))** C (A . E)(2n).
If f E A . E is an extreme point of (A . E)(2n+2) then it is a weak*-extreme
point of (A . E)(2n), as it also belongs to A . E(2n) it is a weak*-extreme point
of A . E(2n). Hence by our theorem f (xo) is an extreme point of E~2n). 0

The next proposition characterizes strongly extreme points in terms of ultra-


powers.
WEAK' -EXTR.EME POINTS 237

PROPOSITION 3. An element e of a Banach space E is a strongly extreme point


of the unit ball El if and only if (e).:F is an extreme point of (E.:Fh-
PROOF. If e . O;El then there is a sequence {en}n~l eEl with lie enll --+ 1
and infnEN Ilenll > o. Thus II (e).:F (en).:F11 = 1 and II (en).:F11 =I- 0 so (e).:F is not an
extreme point.
If (e).:F . oe(E.:Fh then there is 0 =I- (en).:F E (E.:Fh with 1 = II (e).:F (en).:F11 =
lim.:F lie enll Thus for every > 0 the set {n E N: lie enll ~ 1 + } is none empty
as an element of F. Hence there exists a sequence {k n } such that lie ek n II --+ 1
but Ilek" II ---A- 0 so e is not a strongly extreme point. 0
References
[1] P. Beneker and J. Wiegerinck, Strongly exposed points in 'Uniform algebras, Proc. Amer.
Math. Soc. 127 (1999) 1567-1570.
[2] D. Blecher and K. Jarosz, Isomorphisms of function modules, and generalized approximation
in modulus, Trans. Amer. Math. Soc. 354 (2002), 3663-3701
[3] R. D. Bourgin, Geometric aspects of convex sets with the Radon-Nikodym property, LNM
993, Springer, Berlin 1983.
[4] A. Browder, [ntroduction to Function Algebras, W. A. Benjamin, New York 1969.
[5] P. N. Dowling, Z. Hu and M. A. Smith, Extremal structure of the unit ball of C(K, X),
Contemp. Math., 144 (1993) 81-85.
[6] S. Dutta and T. S. S. R. K. Rao, On weak*-extreme points in Banach spaces, preprint 2001.
[7] G. Emmanuele, Remarks on weak compactness of operators defined on injective tensor prod-
ucts, Proc. Amer. Math. Soc., 116 (1992) 473-476.
[8] T. Gamelin, Un'iform Algebras, Chelsea Pub. Comp., 1984.
[9] B. V. Godun, Bor-Luh Lin and S. L. Troyanski, On the strongly extreme points of convex
bodies in separable Banach spaces, Proc. Amer. Math. Soc., 114 (1992) 673-675.
[10] P. Harmand and T. S. S. R. K. Rao, An intersection property of balls and relations with
M-ideals, Math. Z. 197 (1988) 277-290.
[11] P. Harmand, D. Werner and W. Werner, M-ideals in Banach spaces and Banach algebras,
Springer LNM No 1547, Berlin 1993.
[12] K. Hoffman, Banach spaces of analytic functions, Dover 1988.
[13] Z. Hu and M. A. Smith, On the extremal structure of the unit balls of Banach spaces of
weakly continuous functions and their duals, Trans. Amer. Math. Soc. 349 (1997) 1901-1918.
[14] K. Kunen and H. P. Rosenthal, Martingale proofs of some geometric results in Banach space
theory, Pacific J. Math. 100 (1982) 153-175.
[15] O. Nygaard and D. Werner, Slices in the unit ball of a uniform algebra, Arch. Math. (Basel)
76 (2001) 441-444.
[16] R. R. Phelps, Extreme points of polar convex sets, Proc. Amer. Math. Soc. 12 (1961) 291-296.
[17] T. S. S. R. K. Rao, Denting and strongly extreme points in the unit ball of spaces of operators,
Proc. Indian Acad. Sci. (Math. Sci.) 109 (1999) 75-85.
[18] T. S. S. R. K. Rao, Points of weak-norm continuity in the unit ball of Banach spaces, J.
Math. Anal. Appl., 265 (2002) 128-134.
[19] H. Rosenthal, On the non-norm attaining functionals and the equivalence of the weak' -KMP
with the RNP, Longhorn Notes, 1985-86.

DEPARTMENT OF MATHEMATICS AND STATISTICS. SOUTHERN ILLINOIS UNIVERSITY, EDWARDSVILLE.


IL 62026-1653, USA
E-mail address: kjaroszlDsiue. edu
URL: http://www.siue.edu/-kjarosz/

INDIAN STATISTICAL INSTITUTE, R. V. COLLEGE POST, BANGALORE 560059. INDIA


E-mail address:tsslDisibang.ac . in
Contemporary Mathematics
Volume 328. 2003

Determining Sets and Fixed Points for Holomorphic


Endomorphisms

Kang-Tae Kim and Steven G. Krantz

ABSTRACT. The authors study the fixed point sets of a holomorphic endomor-
phism of a domain in complex space. Sufficient (and necessary) conditions are
given-on the number and configuration of the fixed points-for the endomor-
phism to be forced to be the identity. The proofs depend on certain key ideas
from differential geometry, particularly the notions of cut locus and Hadamard
length space.

1. Introduction
This article concerns the study of the concept of determining set for a collection
of holomorphic mappings. We first give the definition.
DEFINITION 1.1. Let M be a complex manifold, and let Aut (M) be the col-
lection of biholomorphic mappings of M into itself. We call a subset Z c M a
determining set for Aut (M) (or, equivalently, an Aut (M)-determining set), if any
map f E Aut (M) satisfying f(p) = p for every p E Z is in fact the identity map of
M.
We observe first that this article is related to the authors' collaboration with
Burna Fridman and Daowei Ma (see [FKKM]), which was originally inspired by the
following remarkable theorem in complex dimension one.
THEOREM 1.2. Let n be a domain in the complex plane C and let f : n --+ n be
a biholomorphic (conformal) mapping. If there are three distinct points Pl,P2,P3
in n such that f(pj) = Pj, for j = 1,2,3, then f is the identity map.
The higher-dimensional analog of this theorem given in [FKKM] is as follows:
THEOREM 1.3. (Fridman-Kim-Krantz-Ma [FKKM]) Let M be a connected,
complex manifold of dimension n admitting a complete invariant Hermitian metric.

2000 Mathematics Subject Classification. 32H02, 32H50, 32H99.


Key words and phrases. fixed point set, holomorphic mapping, cut locus, Hadamard length
space.
K.- T. Kim supported in part by grant ROl-1999-00005 from The Korean Science and Engi-
neering Foundation.
Steven G. Krantz supported in part by grant DMS-9988854 from the National Science
Foundation.

2003 American Mathematical Society


239
240 KANG-TAE KIl\1 AND STEVEN G. KRANTZ

Then a determining set consisting of n + 1 points exists for the automorphisms of


.M. Furthermore, the choice of such a determining set is generic.

Throughout this paper we shall discuss both endomorphisms and automor-


phisms. If M is a complex manifold then an endomorphism of .I'IJ is any holomor-
phic mapping <P : /1.1 -+ Iv! (not necessarily one-to-one or onto). An automorphism
is an endomorphism that is both one-to-one and onto. Observe that the collection
of automorphisms of !vI forms a group under composition of mappings. The col-
lection of endomorphisms forms a semi-group. In many contexts, the algebraic and
topological properties of the automorphism group reveal considerable information
about M. See [ISK] for a survey of these ideas.
The main theme of this article is to provide a fundamental principle for the
theorem of Fridman-Kim-Krantz-Ma ([FKKM]). We first investigate the ca.<;e of
strongly convex domains and the determining sets for general holomorphic endo-
morphisms, which is a broader cla.<;s than the collection of automorphisms. Then
we re-formulate the problem in the general context of Hadamard length spaces (a
concept to be defined below, and which has its own section). This enables us to
understand certain underlying principles for the study of the determining set prob-
lem. In particular, we give a new proof of the aforementioned theorem from this
Hadamard length space viewpoint. Then we end this article with a discussion of
the necessity of several geometric concepts such as web-spanning properties and cut
locus considerations for the determining set; examples are provided.
It is appropriate to note here our gratitude to John MCCarthy, who initially
steered our thoughts to mappings that are not biholomorphic. The three of us had
many stimulating conversations, and a number of the ideas here grew out of those
communications.
The underlying principle of the present article is that, in the appropriate
circumstances, a holomorphic endomorphism is distance-decreasing in a suitable
metric. This property obtains, for instance, with respect to the Kobayashi or
Caratheodory metric of a bounded domain in (Cn. Even in the over-simplified con-
text of linear mappings of the plane (which is often a good testing ground for our
ideas), it is clear that a linear mapping that fixes enough points (three points in
general position will do) must be the identity. The distance-decreasing property
gives us a similar handle for obtaining rigidity of holomorphic mappings. We couple
this metric geometry with ideas about complex geodesics originating with Lempert
to obtain a thoroughly geometric approach to the question of determining sets.
As we discovered in [FKKM], the cut locus is a decisive characteristic for any
mapping question. One's intuition dictates that if enough distinct points (in general
position) are fixed, then barycentric coordinates modeled on that fixed set will show
that all points in an open set are fixed, and thus the map must be the identity.
However, in a geometric situation in which two distinct points may be connected
by several different geodesics, the concept of "barycentric coordinates" is not well-
defined. [The cut locus is the set of points where this ambiguity occurs.] And the
argument just suggested is in fact incorrect, as simple examples show (see [FKKM]).
The geometric machinery developed in [FKKM] and in the present paper is designed
to address the breakdown of barycentric coordinates in such contexts. The ideas
have intrinsic interest, and should be of use in other contexts of geometric analysis,
in particular in the study of holoIllorphic mappings.
DETERMINING SETS FOR HOLOMORPHIC ENDOMORPHISMS 241

2. The Case of Convex Domains


Let us consider a bounded strongly convex domain n in en with a smooth (Ck,
k 2 6) boundary. By the well-known work of Lempert ([LEM]), for each pair of
distinct points p, q E n, there exists a unique holomorphic map 'P : ~ ---- n such
that
(1) 'P(O) = p and 'P(~) = q for some ~ E ~, and
(2) 'P*dn = dLl, where d denotes the Kobayashi distance.
We call such a map 'P a complex geodesic joining p and q.
We now consider the holomorphic endomorphisms of n fixing two given points.
LEMMA 2.1. Let n be a bounded, strongly convex domain in en with Ck
smooth boundary for some k 2 6. Let p, q E n be two distinct points and let 'P
denote a complex geodesic joining p and q. If a holomorphic mapping f : n ---- n
satisfies the condition that f(p) = p and f(q) = q, then it holds that f('P(()) = 'P(()
for every (E ~.

PROOF. Let f and p, q be as in the hypothesis. Let 'P : ~ ---- n be a complex


geodesic joining p and q, with 'P(O) = P and 'P(~) = q. Then let "( : [0, f] ---- ~ be
the unit speed geodesic in ~ with "((0) = 0 and "((e) =~, where = dLl(O,~). Let
o :S t :S f and let r = "((t).
Then we see that

dn(p, q) dn(f(p), f(q))


< dn (f(p), f('P(r))) + dn(f('P(r)), f(q))
< dn('P("((O)), 'P("((t))) + dLl('P("((t)), cp("(()))
< dLl(,,((O), "((t)) + dLl("((t), "((f))
dLl(O,~)
dn(p, q).
Because of the distance-decreasing property of the Kobayashi metric and the
fixed point conditions, we see from the above that

where dn(p, q) = . Notice that every Kobayashi distance ball is strictly convex,
as our domain n is a strongly convex domain with smooth boundary (see [LEM]).
Hence the above observations together with the fact that
dn(p,cp(r)) = t, dn('P(r)) = f - t
imply that f('P(r)) = 'P(r). Consequently, the map f fixes every point in the set
'P 0 "(( [0, f]). Hence the two maps f 0 'P and 'P of ~ into n coincide along a curve in
the unit disc ~. Therefore f 0 'P(() = 'P(() for every ( E ~, as claimed. 0

In other words, we have shown that any holomorphic endomorphism of a


bounded strongly convex domain in en fixing two distinct points must fix every
point that belongs to the complex geodesic passing through the two fixed points.
We immediately ohtain the following general result on the determining sets for
holomorphic endomorphisms of a bounded convex domain.
242 KANG-TAE KIM AND STEVEN G. KRANTZ

LEMMA 2.2. Let 0 be a bounded, strongly convex domain in en with C k


(k ~ 6) smooth boundary. Let PO,Pl,'" ,Pn be points in 0 chosen in such a way
that the complex geodesics passing through Po and Pj (j = 1, ... n) have tangent
vectors at Po that are linearly independent over IC. Then any holomorphic mapping
f: 0 ---> 0 fixing po, ... ,Pn must fix every point ofO.

PROOF. Notice that the current hypothesis together with the preceding lemma
implies that dfpo is the identity map. Therefore a theorem of H. Cartan implies
that f is in fact the identity mapping. 0

We remark that the choice for Po, ... ,Pn is generic. To formulate this notion
more precisely, we consider the cartesian product rrnHo of (n + 1) copies of O. In
fact it is shown in [FKKM] that there exists an open dense subset U of rrn+10 such
that any element of U gives (n + 1) points that satisfy the sufficiency condition
of the preceding lemma. We summarize the result more elegantly in the following
statement.
THEOREM 2.3. For a bounded, strongly convex domain in en, there exists a
collection of n + 1 points such that any holomorphic endomorphism of the domain
fixing them must fix every point in the domain. Moreover, the choice of such n + 1
points is generic.
REMARK 2.4. We point out that the result of this section concerns the class of
bounded, strongly convex domains, which is a rather special collection of objects.
However, in compensation, we emphasize that we have treated general holomorphic
endomorphisms, rather than just biholomorphic self-maps.

3. Hadamard Spaces
In light of the preceding section, we would like to present in this section a
description of the underlying metric space principles that we use in the study of
determining sets.
Let (X, d) be a metric space, equipped with the distance function d : X x X --->
lR. By an isometry we mean a self-mapping f : X ---> X satisfying the condition:
d(J(p), f(q)) = d(p, q), Vp, q E X.
We denote by Isom(X) the collection of isometries of (X, d).
Imitating the concept of "length spaces" that is commonly encountered in ge-
ometry (cf. [BUS]), we give the following definition.
DEFINITION 3.1. Let"(: (a,b) ---> X be a continuous curve in X. We call it
minimal if
d("((x), ,,((y)) = t - x + d("((t), ,,((y)),
for every t,x,y with a < x:::; t:::; y < b.
DEFINITION 3.2. A metric space (X, d) is called a length space if, for every pair
of points p, q EX, there exists a minimal curve "( : [a, b] ---> X such that "((a) = P
and "((b) = q. Furthermore, we call (X, d) a Hadamard space if the minimal curve
joining each pair of points is unique up to a reversal of parametrization.
Notice that any convex subset of Euclidean space is a Hadamard space with
respect to the standard Euclidean distance. A strongly convex domain in en,
equipped with the Kobayashi distance, is also a Hadamard space. Every complete,
DETERMINING SETS FOR HOLOMORPHIC ENDOMORPHISMS 243

simply connected Riemannian manifold with non-positive curvature is also an ex-


ample of a Hadamard space. These are often called Hadamard manifolds; from this
derives our terminology of Hadamard (length) space.
Now. for the study of determining sets, we present this lemma.
LEMMA 3.3. Let (X, d) be a Hadamard space and let p, q E X be two distinct
points. If an isometry f : X --> X fixes P and q, tllen f fixes every point on the
minimal curve passing through P and q.

PROOF. Let'Y : [O,f] --> X be a minimal curve with 'Y(O) = p,'Y(f) = q. It is


immediate to see that the isometry f of (X, d) has the property that f 0 'Y is also
a minimal curve. Since P and q are fixed by f, and since (X, d) is Hadamard, it
follows that f 0 'Y(t) = 'Y(t) for every t E [0, fl. Now consider the minimal curve
passing through P and q. So far, we have shown that the portion of this minimal
curve between P and q is pointwise fixed by f. We still must show that the other
portion of the minimal curve is fixed pointwise by f. But this is a simple matter
using the uniqueness of the minimal curve passing through any two given points and
the minimal-curve-preserving property of isometries. This completes the proof. D

We remark that it is possible to derive the same conclusions for locally Hadamard
spaces but, in order to keep our exposition concise, we do not introduce the concept
here.
LEMMA 3.4. Let (X, d) be a Hadamard space and let U be an open subset of
X. Then any isometry f fixing every point in U must be the identity map.

PROOF. Let P E U. Then, for every minimal curve 'Y emanating from p, f fixes
a point in 'Y n (U \ p). The preceding lemma implies now that f fixes every point
of 'Y. Since every point in X can be joined to P by a minimal curve, this completes
the proof. D

Now we consider the concept of convex hull in a Hadamard space. We say that
a set Q in a Hadamard space X is convex if every minimal curve joining P and q in
X is contained in Q. For a subset A of a Hadamard space X, its convex hull W (A)
is the smallest convex subset of X containing A.
DEFINITION 3.5. Let Po, . .. ,Pm be points in a Hadamard space (X, d) with
minimal geodesics 'Yl .... ,'Ym such that 'Yj passes through Po and Pi for every j =
1, .... m. We call the points Po, ... ,Pm spanning if the convex hull Whl U ... U'Ym)
has non-empty interior.
Now we have the following general result.
PROPOSITION 3.6. If a Hadamard space (X, d) admits m + 1 points Po, PI, ... ,
Pm ill X that are spanning, then these m + 1 points constitute a determining set
for the isometries of( X, d).

PROOF. Notice that the convex hull we obtain from the minimal curves through
Po and Pi is fixed pointwise by any isometry that fixes the points Po, ,Pm. Then
the preceding lemma finishes the proof. D

Observe that the full isometry condition is not really needed to prove the con-
clusion of the proposition. In fact, any distance-decreasing map will satisfy the
244 KANG-TAE KIM AND STEVEN G. KRANTZ

same conclusion (just use the unique continuation principle). Notice that this of-
fers an underlying principle for the determining set theorem for the holomorphic
endomorphisms of strongly convex domains in the preceding section.

4. CH-Subsets and Automorphisms


In this section we demonstrate how the principles of the preceding section reflect
upon the main theorem of [FKKM] and its proof.
Let M be a connected, complex manifold that admits a smooth invariant Her-
mitian metric. Here the invariance refers to the property that every holomorphic
automorphism of M is an isometry with respect to the Hermitian metric.
For a moment, we take the real part of the Hermitian metric, and consider
everything in terms of lliemannian geometry. Let P EM. Then we call q E M a
cut point of P if there are at least two distinct geodesics joining P and q with the
same minimal length. The collection of cut points for P is called the cut locus of p,
which we denote by Cpo
In [FKKMJ, a subset X of M was called Carlan-Hadamard ('CH' for short) if
there exists Xo E X so that X does not intersect the cut locus C(xo) of Xo in M.
Furthermore, we call such a CH-set X generating if the set
Ip(X) := {'Y~(O) I 'Y is the unique normal geodesic from Xo to p, Vp E Z}
spans TxoM over C.
Suppose now that X is a set of finitely many points, and that a certain holo-
morphic automorphism f fixes every point of X. Then, by complex differentiability,
one picks up more geodesics than just the geodesics joining Xo and the other points
of X. [That is to say, each geodesic tangent may be multiplied by i.] If x E X \ Xo
and if 'Yx is the unit speed geodesic from Xo to x, then 9x == expxo(i"(~(O)) is also
fixed, point by point, by f. Now it is not hard to see, using the exponential map
and the tangent space, that the convex hull W = Wbl U 91 U ... U 'Ym U 9m) has
non-empty interior. Notice that every point of the hull W is fixed by f pointwise.
We obtain the following result as a consequence of Proposition 3.6.
PROPOSITION 4.1. (Fridman/Kim/Krantz/Ma [FKKM]) Let M be a connected,
complex manifold with an invariant Hermitian metric. Let X be a generating CH-
subset of M. Then, whenever an automorphism fixes every point of X, it is in fact
the identity map. In other words, every generating CH-subset is a determining set
for automorphisms.
The method of choosing a smallest (in the sense of inclusion of sets) generating
CH-subset in a complex manifold with a complete invariant Hermitian metric has
been explained in detail in [FKKM]. We briefly describe the paradigm. Choose
an arbitrary p E M. Then the cut locus C(p) is a nowhere dense subset of M.
Thus choose PI E M \ (C (p) U {p}). Then choose P2 away from C (P) and the
complete geodesic through p and Pl. Then P3 will be chosen away from C(p) and
the geodesic cone generated by P,Pl and P2. An inductive construction lets us
choose p, PI, ... ,Pn which compose a spanning CH-subset of M. Thus we arrive at
THEOREM 4.2. (Fridman-Kim-Krantz-Ma [FKKM]) Let M be a connected n-
dimensional complex manifold admitting a complete invariant Hermitian metric.
Then a determining set, consisting of n + 1 points, exists for the automorphisms of
M. Furthermore, the choice of such a determining set is generic.
DETERMINING SETS FOR HOLOMORPHIC ENDOMORPHISMS 245

Here, by "generic", we mean that the the collection of (n + I)-tuples of points


in Al x ... x M that satisfy our conclusion form a dense, open set.
From the discussion above, if the metric happens to be distance-decreasing, in
the sense that all holomorphic endomorphisms are distance-decreasing with respect
to the given metric, then this theorem will hold for holomorphic endomorphisms.
This result of course uses the idea developed in the preceding section about the
distance-decreasing property together with the unique continuation property.
We remark at this point that the collection of complex manifolds admitting a
complete invariant Hermitian metric is rather large. For instance, every bounded
pseudoconvex domain in en is equipped with the complete Kiihler-Einstein metric.
See [MOY] (also [CHY], [OHS], [YAU]) for instance.

5. Examples, Counterexamples, and the Cut Locus


One might have the impression that some transversality condition for m + 1
points might be sufficient for the determining set problem for holomorphic automor-
phisms. However, it is shown in [FKKM] that a simplistic topological transversality
assumption will not be sufficient; consider the following statement.
THEOREM 5.1. ([FKKM]) Fix a finite set K = {PI, ... , Pk} in en, n > 1. There
exists a bounded domain containing K, and a subgroup H C Aut(D) isomorphic
to the unitary group U(n - 1) of en-I, such that each element of H fixes every
point of K.
Moreover, unlike the one-dimensional planar domain case, the consideration
of the cut locus seems essential even for one-dimensional Riemann surfaces. If
one considers the torus coming from the lattice generated by {I, i}, then the map
z --+ - z of e generates an automorphism on the torus. It is easy to see that it has 4
fixed points, and yet is different from the identity map. If one considers a two-holed
torus with a well-balanced fundamental domain centered at 0 in the Poincare disc,
then the same map z --+ -z of the disc will generate a non-trivial automorphism
with 6 fixed points. In this way, one can generate arbitrarily many fixed points for
a non-trivial automorphisms of compact Riemann surfaces of high enough genus.
Since our discussion has not depended upon the completeness of manifolds, simple
puncturing will create an arbitrary number of fixed points. Notice that all these
examples have fixed points in the cut loci.
246 KANG-TAE KIM AND STEVEN G. KRANTZ

References
[ALK] G. Aladro and S. G. Krantz, A criterion for normality in Cn, Jour. Math. Anal. and Appl.
161(1991), 1-8.
[BED] E. Bedford and J. Dadok, Bounded domains with prescribed group of automorphisms.,
Comment. !'v'lath. Helv. 62 (1987), 561-572.
[BUS] H. Busemann, The geometry of geodesics, Academic Press, New York, NY, 1955.
[CHY] S.Y. Cheng and S.T. Yau, On the existence of a compact Kahler metric, Comm. Pure
App!. Math., 33 (1980), 507-544.
[FIF] S. D. Fisher and John Franks, The fixed points of an analytic self-mapping, Proc. AMS,
99(1987), 76-78.
[FKKM] B. Fridman, K-T. Kim, S. G. Krantz, and D. Ma, On Fixed Points and Determining
Sets for Holomorphic Automorphisms, Michigan Math. Jour., to appear.
[FP] B. L. Fridman and E. A. Poletsky, Upper semicontinuity of automorphism groups, Math.
Ann., 299(1994), 615-628.
[GRK] R. E. Greene and S. G. Krantz, Stability properties of the Bergman kernel and curvature
properties of bounded domains, Recent Developments in Several Complex Variables (J. E.
Fornress, ed.), Princeton University Press (1979),179-198.
[GRW] R. E. Greene and H. Wu, Function theory on manifolds which possess a pole, Lecture
Notes in Mathematics, 699, Springer, Berlin, 1979.
[GKM] D. Gromoll, W. Klingenberg, and W. Meyer, Riemannsche Geometrie im Grossen, 2nd
ed., Lecture Notes in Mathematics, v. 55, Springer-Verlag, New York, 1975.
[ISK] A. Isaev and S. G. Krantz, Domains with non-compact automorphism group: A survey,
Advances in Math. 146(1999), 1-38.
[KLI] W. Klingenberg, Riemannian Geometry, 2nd ed., de Gruyter Studies in Mathematics,
Berlin, 1995.
[KOB] S. Kobayashi, Hyperbolic complex spaces, Springer, 1999.
[LEM] L. Lempert, La metrique Kobayashi et las representation des domains sur la boule, Bull.
Soc. Math. Prance 109(1981), 427-474.
[LES] K Leschinger, Uber fixpunkte holomorpher Automorphismen, Manuscripta Math., 25
(1978), 391-396.
[MA] D. Ma, Upper semicontinuity of isotropy and automorphism groups, Math. Ann., 292(1992),
533-545.
[MAS] B. Maskit, The conformal group of a plane domain, Amer. J. Math., 90 (1968), 718-722.
[MOY] N. Mok and S. T. Yau, Completeness of the Kahler-Einstein metric on bounded domains
and the characterization of domains of holomorphy by curvature conditions, Symposia in Pure
Math. The mathematical heritage ofH. Poincare, Amer. Math. Soc., 39, Part I. (1983),41-60.
[OHS] T. Ohsawa, On complete Kahler domains with C 1 boundary, Pub!. Res. Inst. Math. Sci.,
RIMS (Kyoto), 16 (1980), 929-940.
[PEL] E. Peschl and M Lehtinen, A conformal self-map which fixes 3 points is the identity, Ann.
Acad. Sci. Fenn., Ser. A I Math., 4 (1979), no. 1, 85-86.
[SUI] N. Suita, On fixed points of conformal self-mappings, Hokkaido Math. J., 10(1981), 667-671.
[Vll] J.-P. Vigue, Fixed points of holomorphic mappings in a bounded convex domain in Cn,
Proceedings of Symposia in Pure Mathematics, 52(1991), Amer. Math. Soc., 579-582.
[VI2] J.-P. Vigue, Fixed points of holomorphic mappings, Complex Geometry and Analysis (Pisa,
1988), Lecture Notes in Mathematics, v. 1422, Springer, Berlin, 1990, pp. 101-106.
[YAU] S. T. Yau, A survey on Kahler-Einstein metrics. Complex analysis of several variables
(Madison, Wis., 1982), 285-289, Proc. Sympos. Pure Math., 41, Amer. Math. Soc., Provi-
dence, RI, 1984.

KANG-TAE KIM, DEPARTMENT OF MATHEMATICS, POHANG UNIVERSITY OF SCIENCE AND


TECHNOLOGY, POHANG 790-784, KOREA

STEVEN G. KRANTZ, DEPARTMENT OF MATHEMATICS, CAMPUS Box 1146, WASHINGTON UNI-


VERSITY, ST. LOUIS, MISSOURI 63130 U.S.A.
E-mail address:kimktlpostech.ac.kr
E-mail address: sklDmath. wustl. edu
Contemporary Mathenlatics
Volume 328, 2003

Localization in the Spectral Theory


of Operators on Banach Spaces

T. L. Miller, V. G. Miller, and M. M. Neumann

ABSTRACT. In the first two sections of this article, we survey some of the re-
cent progress in the local spectral theory of operators on Banach spaces with
emphasis on the local spectrum and on restrictions and quotients of decom-
posable operators. In particular, the problem of characterizing restrictions
and quotients of generalized scalar operators with spectrum in the unit circle
in terms of suitable growth conditions is addressed in detail, with emphasis
on [11], [22], and [23]. The last two sections center around certain localized
versions of the single-valued extension property, Bishop's property (13), and
the decomposition property (8), mainly in the spirit of [2], [5], [6], and [13].
For each of these properties, we find a smallest closed set modulo which it
holds. For these residual sets, we establish a spectral mapping theorem with
respect to the Riesz functional calculus. We also obtain precise information
about the extent to which Bishop's property (3) holds on the essential or the
Kato resolvent set. Our results are exemplified in the case of weighted shifts.
Moreover, several of the outstanding open questions of the field are mentioned
in their natural context.

1. Decomposable operators and the local spectrum

Among the various aspects and levels of localization in spectral theory, we


choose decomposability as our starting point. Let X be a complex Banach space,
and let L(X) denote the Banach algebra of all bounded linear operators on X. For
T E L(X), let, as usual, a(T), ap(T), aap(T), r(T), and p(T) denote the spectrum,
point spectrum, approximate point spectrum, spectral radius, and resolvent set of
T, and let Lat(T) stand for the collection of all T-invariant closed linear subspaces
of X.
From [18J and [29J we recall that an operator T E L(X) is said to be decom-
posable provided that, for each open cover {U, V} of C, there exist Y, Z E Lat(T)
for which
X = Y + Z, a(T Iy) ~ U, and a(T I Z) ~ V.
By [18, 1.2.23J or [29, 4.4.28J, this simple definition is equivalent to the original
notion of decomposability, as introduced by Foi~ in 1963 and discussed in the
classical book by Colojoara and Foi~ [IOJ.

2000 Mathematics Subject Classification. Primary 47All, 47B40; Secondary 47B37.

2003 American Mathematical Society


247
248 T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN

As witnessed, for instance, by the monographs [10], [16], [18], and [29], the
theory of decomposable operators is now richly developed with many interesting ap-
plications and connections. Evidently, the class of decomposable operators contains
all normal operators on Hilbert spaces and, more generally, all spectral operators
in the sense of Dunford on Banach spaces. Moreover, a simple application of the
Riesz functional calculus shows that all operators with totally disconnected spec-
trum are decomposable. In particular, all compact and all algebraic operators are
decomposable.
An important subclass of the decomposable operators is formed by the gener-
alized scalar operators, defined as those operators T E L(X) for which there exists
a continuous unital algebra homomorphism <P : (C) ---t L(X) with <p(Z) = T.
Here (C) denotes the Frechet algebra of all infinitely differentiable complex-valued
functions on C, and Z stands for the identity function on C. More generally, de-
composability is shared by all operators with a non-analytic functional calculus on
an algebra of functions that admits partitions of unity. Finally, all multiplication
operators on a regular semi-simple Banach algebra are decomposable, and so are
many, but not all, convolution operators on the group and measure algebra of a
locally compact abelian group. In fact, as discussed in [18, 4.11], the problem
of decomposability of convolution operators is closely related to the Wiener-Pitt
phenomenon from commutative harmonic analysis.
To see more clearly how decomposability localizes the spectrum, we need to
review a few basic notions from local spectral theory that date back to Dunford;
see [12], [18], and [29] for details. For an open set U ; C, let H(U, X) denote the
Frechet space of all X-valued analytic functions on U, and, for arbitrary T E L(X),
let Tu denote the operator on H(U, X) given by
(Tu f)(J-t) := (T - J-t)f(J-t) for all f E H(U, X) and J-t E U.
It turns out that this operator dominates large parts of spectral theory.
The local resolvent set PT(X) of T at a vector x E X is defined to consist of
all >. E C for which there exists some f E H(U, X) on an open neighborhood U of
>. for which Tuf = x. Clearly, f(J-t) = (T - J-t)-lX for all J-t E Un p(T), so that
PT(X) is open and contains p(T). Hence the local spectrum aT(x) := C \ pT(X) is a
closed subset of a(T). In general, the various analytic functions that occur in the
definition of pT(X) need not be consistent. This issue is addressed by the following
definition.
The operator T E L(X) is said to possess the single-valued extension property
(SVEP), if Tu is injective for all open sets U ; IC. By [18, 3.3.2], T has SVEP
precisely when, for each x E X, there exists a unique function f E H(PT(X), X) for
which
(T - J-t)f(J-t) = x for all J-t E PT(X).
This function is then called the local resolvent function for T at x. In remarkable
contrast to the usual resolvent function, such functions may well be bounded; this
recent discovery of Bermudez and Gonzalez will be exemplified below.
One might expect a(T) to be the union of the local spectra aT(x) over all
x E X, but this is not true in general. In fact, this union coincides with the
surjectivity spectrum asu(T) of T, the set of all >. E C for which T - >. fails to be
surjective. However, if T has SVEP, then asu(T) = a(T), and aT(x) is non-empty
for all non-zero x E X, [18, 1.2.16 and 1.3.2]. As a powerful application, we obtain
THE SPECTRAL THEORY OF OPERATORS ON BANACH SPACES 249

that every surjective operator with SVEP is actually bijective, [18, 1.2.10]. A more
precise version of this result will be discussed in Section 3. For arbitrary T E L(X)
and F ~ C, let
XT(F) := {x EX: aT(x) ~ F}
denote the corresponding local spectml subspace. Evidently, XT(F) is a linear sub-
space of X, but need not be closed.
The following classical result illustrates that these subspaces playa basic role
for spectral decompositions, see [18, 1.2] and [29,4.4].
THEOREM 1. Suppose that T E L(X) is decomposable. Then T has SVEP, and,
for each closed set F ~ C, the space XT(F) is closed and satisfies a(T I XT(F)) ~ F.
In fact, XT(F) is the largest among all spaces Y E Lat(T) for which aT(T IY) ~ F.
Moreover, XT(F) ~ XT(Ut}+ +XT(U n ) for every finite open cover {U1 , ... , Un}
~F 0
The following examples may illuminate how spectral decompositions work in
some important cases; for details see [18] and, for the last assertion of Example
4, also [26, Th.16]. The extent to which the compactness of the group is essential
here remains a challenging open problem.
EXAMPLE 2. Let T E L(X) be a normal opemtor with spectml measure ~ on
a complex Hilbert space X. Then T is decomposable, and XT(F) = ran ~(F) for
every closed set F ~ Co Moreover, there exists a non-zero bounded local resolvent
function for T precisely when int a(T) =f:. 0. 0
EXAMPLE 3. Let X := C(O) be the space of continuous functions on a compact
Hausdorff space 0, and let T E L(X) denote the opemtor of multiplication by a given
function g E X. Then T is decomposable, and XT(F) = {f E C(O) : g(supp f) ~ F}
for every closed set F ~ Co Also in this case, there exists a non-zero bounded local
resolvent function for T precisely when int a(T) =f:. 0. 0
EXAMPLE 4. Let X := Ll(G) be the group algebm of a locally compact abelian
group G, and let T E L(X) denote the opemtor of convolution by a given function
g E X. Then T is decomposable, and XT(F) = {f E Ll(G) : g(suppj) ~ F} for
every closed set F ~ C, where j denotes the Fourier tmnsform. Moreover, at least
when G is compact, there exists a non-zero bounded local resolvent function for T
precisely when int a(T) =f:. 0. 0
On the other hand, there are important classes of operators which are not
covered by decomposability. For instance, by [18, 1.6.14] and [22], a unilateral
weighted right shift on the sequence space fP(N o) for arbitrary 1 ::::; p < 00 is
decomposable, or, equivalently, the quotient of a decomposable operator, only in
the trivial case when it is quasi-nilpotent, while unilateral weighted right shifts are
never generalized scalar. Moreover, as we shall see, there are many examples of
unilateral and bilateral weighted left shifts without SVEP.
Another illuminating case is that of isometries. By [18, 1.6.7], an arbitrary
Banach space isometry is decomposable, or, equivalently, the quotient of a decom-
posable operator or generalized scalar, precisely when it is invertible. On the other
hand, every isometry may be extended, by a classical result due to Douglas, recorded
in [18, 1.6.6], to an invertible isometry, and hence has a decomposable extension.
In the next section, we shall discuss a more general version of this result.
250 T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN

2. Moving beyond decomposability

Several years before decomposability was formally introduced by Foi~, Bishop


[9] investigated a number of spectral decomposition properties in an attempt to
extend some of the features of the theory of normal operators to the general set-
ting of Banach spaces. Among these properties, one turned out to be particularly
important.
We now say that an operator T E L(X) on a complex Banach space X has
Bishop's property ((J) provided that, for each open set U ~ C, the operator Tu on
H(U, X) is injective with closed range, equivalently, if, for each sequence (fn)nEN
in H(U, X) with
(T - )..)fn(>\) -> 0 asn->oo,
uniformly on each compact subset of U, it follows that fn()..) -> 0 as n -> 00, again
uniformly on the compact subsets of U, [18, 1.2.6]. Actually, by [18, 3.3.5], the
injectivity condition in this definition is redundant. Obviously, property ({J) implies
SVEP. It was shown a long time ago by Foi~ that all decomposable operators share
property ((J), but the precise relationship was discovered only recently by Albrecht
and Eschmeier [6].
THEOREM 5. An operator T E L(X) has Bishop's property ((J) precisely when
T is similar to the restriction of a decomposable operator to a closed invariant
subspace. Moreover, in this case, there exists a decomposable extension 8 for which
aCT) ~ a(8). 0
The result was, in part, inspired by the work of Putinar [27] who proved that
every hyponormal operator is subscalar, in the sense that it has a generalized scalar
extension. Thus all hyponormal operators have property ((J). In particular, all
unilateral weighted right shifts on f2(N o) with an increasing weight sequence w
have property ({J), but a characterization of ({J) in terms of w seems to be an
intriguing open problem. For partial results, see [11], [18], [22], and [23].
To discuss the dual notion of Bishop's property ((J), we need a slight variant of
the local spectral subspaces. For arbitrary T E L(X) and a closed subset F of <C,
the corresponding glocal spectral subspace is defined as
XT(F) := {x EX: x E ran TC\F }.
In this definition, the point is that the local resolvent function is defined globally
on the entire complement of F. Clearly, XT(F) is a linear subspace contained in
XT(F). Moreover, by [18, 3.3.2], the identity XT(F) = XT(F) holds for all closed
sets F ~ <C precisely when T has SVEP. As formalized by Albrecht, Eschmeier, and
the third author in 1986, T is said to have the decomposition property (8) provided
that
X = XT(U) + XT(V) for every open cover {U, V} of <C,
[18, 1.2.28]. Of course, the basic idea behind this definition is that of decomposition
of local spectra, but some care with the domains of the local resolvent functions is
needed.
It is not difficult to see that property (8) is inherited by quotients, and that T
is decomposable precisely when T has both properties ({J) and (8). More important
and much deeper is the following result due to Albrecht and Eschmeier [6] that,
in a sense, completed, after many years, the aspirations of the inquiry initiated by
Bishop [9].
THE SPECTRAL THEORY OF OPERATORS ON BANACH SPACES 251

THEOREM 6. Property (8) chamcterizes, up to similarity, the quotients of de-


composable opemtors by closed invariant subspaces. Moreover. the properties (13)
and (8) are dual to each other. in the sense that an opemtor T E L(X) has one of
these properties precisely when the adjoint T* E L(X*) on the dual space X*has
the other one. D

Here the hardest assertion to prove is that property (13) for T* implies prop-
erty (8) for T. The construction of decomposable extensions and liftings uses two
powerful functional models for operators on Banach spaces of independent interest.
These models are in the spirit of function-theoretic operator theory, and involve
the operator of multiplication by the independent variable on certain Sobolev-type
spaces together with the theory of topological tensor products. The complete dual-
ity between the properties (/3) and (8) employs the Grothendieck-K6the duality for
spaces of vector-valued analytic functions. All of this is described, in considerable
detail, in [18, Ch.2].
There are interesting applications to the invariant subspace problem. Indeed, if
the operator T E L(X) has either property (/3) or property (8), then Eschmeier and
Prunaru [14] established that Lat(T) is non-trivial provided that a(T) is thick, and
that Lat(T) is rich in the sense that it contains the lattice of all closed subspaces of
some infinite-dimensional Banach space provided that the essential spectrum ae(T)
is thick. Here we skip the formal definition of thick subsets of the complex plane,
but note that all compact sets with non-empty interior are thick. A streamlined
approach to this result and further references may be found in [18, 2.6].
Since all hyponormal operators have, by Putinar's result [27], property (/3),
the preceding result subsumes, in particular, Brown's celebrated invariant subspace
theorem for hyponormal operators with thick spectrum. In light of Read's recent
construction of a quasi-nilpotent, and hence decomposable, operator on a Banach
space without non-trivial invariant subspaces, it is clear that the condition of thick
spectrum cannot. be dropped in general. However, the invariant subspace problem
for operators on Hilbert spaces remains open, even for the class of hyponormal
operators.
As discussed in the monograph by Eschmeier and Putinar [16], there are also
interesting connections between property (/3) and the theory of analytic sheaves.
These connections are not only important for the spectral theory of several com-
muting operators, but they are also at the heart of some of the recent developments
in the case of single operators. Although, as witnessed by the exposition of local
spectral theory in [18], the explicit use of sheaf theory can be avoided in the case of
single operator theory, the reader should be aware of these connections. The basic
idea is sketched in [18, 2.2].
A classical issue of local spectral theory is to derive spectral decomposition
properties from growth conditions on the powers or the resolvent function of a given
operator. For instance, Levinson's log-log theorem from complex analysis may be
used to show that, for operators with spectrum in the real line or the unit circle 'll', a
very weak logarithmic growth condition on the resolvent function suffices to ensure
decomposability. The short approach from [18, 1.7] to this classical result due to
Radjabalipour is based on the fact that, by Theorem 6, an operator T E L(X) is
decomposable precisely when both T and T* have property (13). A very attractive
account of the local spectral t.heory for operators with thin spectrum may be found
in a recent survey article by Albrecht and Ricker [7].
252 T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN

Here we focus only on one aspect that leads to interesting open problems. For
arbitrary T E L(X), let K(T) := inf{IITxll : Ilxll = I} denote the lower bound of
T. Evidently, K(T)-l = II T-lil when T is invertible. Similar to the case of the
spectral radius, it is known that the sequence of numbers K(Tn) lin converges to
its supremum, denoted by i(T), and that aap(T) ~ {>' E C : i(T) ~ 1>'1 ~ r(T)},
[18, 1.6.1 and 1.6.2].
By a classical result due to Colojoara and Foi~, [10, 5.1] or [18, 1.5.12], a
generalized scalar operator T satisfies a(T) ~ T precisely when T is (T) -scalar, in
the sense that T admits a continuous functional calculus on the Frechet algebra (T)
of all COO-functions on T. Moreover, T is (T)-scalar if and only if T is invertible
and satisfies the condition of polynomial growth (P), in the sense that there exist
constants c, s > 0 such that
1
-s ~ K(Tn) ~ IITnl1 ~ cn s for all n E Nj
cn
indeed, in this case, a functional calculus cP for T is given by the formula
00

cpU):= 2: !(n) Tn for all f E (T) ,


n=-oo

where !(n) denotes the nth Fourier coefficient of f.


Evidently, all invertible isometries have property (P), and hence are (T)-
scalar. Also, it follows from the preceding characterization that an operator T E
L(X) is (T)-scalar precisely when its adjoint T* is (T)-scalar. Moreover, since
property (P) implies that i(T) = r(T) = 1 and consequently aap(T) ~ T, and since
aap(T) = a(T) when T has property (6), we are led to the following result.
PROPOSITION 7. For every T E L(X) with property (P), the following equiva-
lences hold:
T is invertible :} a(T) ~ T :} T has (6) :} T is decomposable :} T is (T)-scalar.
Moreover, ifT is not invertible, then aap(T) = T and a(T) is the closed unit disc.o
Evidently, every restriction of an (T)-scalar operator has property (P), but
the converse is open in general. The preceding proposition shows that this problem
is equivalent to the problem of extending an arbitrary operator with property (P) to
an invertible operator with property (P) for possibly larger constants c, s > O. Since
the extension provided by the Albrecht-Eschmeier functional model in Theorem 5
increases the spectrum, a different approach is needed here.
As noted above, for isometries, the desired extension is possible by a result of
Douglas. Also, for a certain class of operators that includes all unilateral weighted
right shifts, a positive solution was recently provided by Didas [11] and the authors
[23]. While Didas exploits the theory of topological tensor products in the spirit
of Eschmeier and Putinar [16], the more elementary approach from [23] uses a
modification of a construction provided by Bercovici and Petrovic [8] to charac-
terize compressions of (T)-scalar operators. For unilateral weighted right shifts
on fP(N o), the method developed in [23] leads to extensions as bilateral weighted
shifts on fP(Z) with sharp growth estimates.
To reduce the case of quotients of (T)-scalar operators to that of restrictions,
we recall that the minimum modulus 'Y(T) of a non-zero operator T E L(X) is
defined as 'Y(T):= inf{IITxll/dist(x,kerT): x (j kerT}. Clearly, 'Y(T) = K(T)
THE SPECTRAL THEORY OF OPERATORS ON BANACH SPACES 253

when T is injective. It is also well known that 'Y(T) = 'Y(T*), and that 'Y(T) > 0
precisely when T has closed range. Standard duality theory now leads to the
following result.
PROPOSITION 8. An operator T E L(X) is the restriction of an (1l')-scalar
operator if and only if its adjoint T* is the quotient of an (1l')-scalar operator.
Moreover, ifT is the quotient of an (1l')-scalar operator, then T* is the restriction
of an (1l')-scalar operator, and hence there exist constants c, s > 0 for which

~::;'Y(Tn)::;IITnll::;cns
cn
forallnEN. 0

In general, it is not known if the last growth condition characterizes the quo-
tients of (1l')-scalar operators, but, by Proposition 8 and [23, Prop.5], this is
the case for the class of all unilateral weighted left shifts on fP(N o ) for arbitrary
1 < p < 00. More precisely, a unilateral weighted left shift on fP(N o) satisfies the
growth condition of Proposition 8 if and only if it admits a bilateral weighted shift
lifting on fP(Z) that is (1l')-scalar.
Similar results hold for more general growth conditions; see [22] and [23].
For instance, by another classical result due to Colojoara and Foi~, an invertible
operator T E L(X) is decomposable provided that T satisfies Beurling's condition
(B), in the sense that
1
L
00

n 2 (llogK(Tn)1 + IlogliTnll1) < 00,


n=l
[10, 5.3.2] and [18,4.4.7]. Clearly, property (B) is inherited by restrictions, but it
remains open, if every operator with property (B) has an invertible extension with
property (B). In fact, it is not known, if property (B) implies property ((3). For
certain unilateral weighted right shifts, a positive answer was recently given in [22]
and [23].

3. Localization of the single-valued extension property

For an arbitrary operator T E L(X) on a complex Banach space X, here the


spaces K(T) := XT(C\ {O}) and Ho(T) := X T ( {O}) will be of particular importance.
Both spaces were, in some disguise, studied by Mbekhta and also by Vrbova; see
[19], [20], and [30]. By [18, 3.3.7], K(T) coincides with the analytic core of T,
defined to consist of all x E X for which there exist a constant c > 0 and elements
Xn E X such that

for all n E N.
By this characterization and the open mapping theorem, K(T) =X if and only if
T is surjective. In terms of local spectral theory, this follows also from the fact that
asu(T) is the union of all local spectra of T. On the other hand, by [18, 3.3.13],
Ho(T) is the quasi-nilpotent part of T, defined as the set of all x E X for which
IITnx11 1/ n _ 0 as n - 00.

In general, neither K(T) nor Ho(T) need to be closed, but, if 0 is isolated in a(T),
then, by [19, 1.6], both spaces are closed and X = K(T) EB Ho(T). For more on
operators with closed K(T) and Ho(T), see [1], [2], and [24]. For instance, by [24,
Cor.6], for any non-invertible decomposable operator T, the point 0 is isolated in
254 T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN

a(T) precisely when K(T) is closed. In particular, the analytic core of a compact
or, more generally, a Riesz operator T is closed exactly when T has finite spectrum,
[24, Cor.9].
The spaces Ho(T) and K(T) are related to the kernel N(T) and the range R(T)
of T as follows. With the notation

N(T) :=
oc

U N(Tn)
n=l
and R(T):= n00

n=l
R(Tn)

for the generalized kernel and range of T, there is an increasing chain of kernel-type
spaces
N(T) ~ N(Tn) ~ N(T) ~ Ho(T) ~ X T ({O})
and a decreasing chain of range-type spaces
R(T) 2 R(Tn) 2 R(T) 2 K(T) 2 X T (0)
for arbitrary n E N. [18, 1.2.16 and 3.3.1]. The geometric position of the kernel-
type spaces versus the range-type spaces turns out to be intimately related to a
certain localized version of SVEP for the operator T and its adjoint T*.
An operator T E L(X) is said to have SVEP at a point A E C, if, for every
open disc U centered at A, the operator Tu is injective on H(U, X). This notion
dates back to Finch [17], and was pursued further, for instance, in [1], [2], [3], [4],
[5], and [20]. Evidently, T has SVEP at A precisely when T - A has SVEP at 0,
while SVEP for T is equivalent to SVEP for T at A for each A E C.
Local spectral theory leads to a variety of characterizations of this localized
version of SVEP that involve the kernel-type and range-type spaces introduced
above. Our starting point is the following characterization from [3, 1.9]. The result
shows, in particular, that every injective operator T E L(X) has SVEP at 0, and
may be viewed as a local version of the classical fact that T has SVEP if and only
if X T (0) = {O}, [18, 1.2.16]. For completeness, we include a short new proof that
uses nothing but local spectral theory.
THEOREM 9. For every operator T E L(X), the following equivalences hold:
T has SVEP at 0 <=} N(T) n X T (0) = {O} <=} aT(x) = {O} for all 0 =f. x E N(T).
Proof. First suppose that T has SVEP at 0, and consider an arbitrary x E
N(T) for which aT(x) is empty. Then 0 E pT(X) so that there exists an f E H(U, X)
on some open disc U with center 0 for which (T-A)f(A) = x for all A E U. It follows
that (T - A)Tf(A) = Tx = 0 for all A E U, and therefore Tf(A) = 0 for all A E U,
since T has SVEP at O. Thus x = Tf(O) = 0, and hence N(T) n X T (0) = {O}.
Next observe that, for each x E N(T), the definition f(A) := -xl A yields
an analytic function for which (T - A)f(A) = x for all non-zero A E C. Thus
aT(x) ~ {O} for all x E N(T). Consequently, the second and third assertions are
equivalent.
Finally suppose that N(T) n XT(0) = {O}, let U be an open disc with cen-
ter 0, and consider a function f E H(U, X) for which Tu f = O. By [18, 1.2.14],
aT(f(A)) = aT(O) = 0 for all A E U. Now, for the power series representation f(A) =
L:~=o an An for all A E U, our task is to show that each of the coefficients an E X is
zero. For the case n = 0, this is immediate, since ao = f(O) E N(T) nXT (0) = {O}.
But then it follows that
for all A E U,
THE SPECTRAL THEORY OF OPERATORS ON BANACH SPACES 255

and therefore (T - >.) (at + a2 >. + a3 >.2 + ... ) = 0 first for all non-zero>. E U, and
then, by continuity, also for>. = O. Exactly as before, we conclude that at = 0 and
hence, by induction, an = 0 for all n ~ O. Thus f == 0 on U, as desired. 0
Since N(T) n K(T) ~ X T ( {O}) n XT(C \ {O}) = X T (0), it clearly follows that
N(T) n K(T) = N(T) n X T (0) for every T E L(X). Thus, by Theorem 9, T has
SVEP at 0 if and only if N(T) n K(T) = {O}. In particular, if T is surjective, then,
as noted above, K(T) = X, so that T has SVEP at 0 precisely when T is injective.
This characterization from [3, 1.11] extends a classical result due to Finch [17]. As
another immediate consequence of Theorem 9, we obtain the following result.
COROLLARY 10. An operator T E L(X) has SVEP at 0 provided that either
Ho(T) n K(T) = {O} or N(T) n R(T) = (0). 0
Recent counter-examples in [2] show that, in general, none of the latter condi-
tions is equivalent to SVEP of T at 0, thus disproving a claim made in [20, 1.4].
However, by [1, 2.7], [5, 1.3], and Theorem 12 below, equivalences do hold for
certain classes of operators.
We now describe how the localized SVEP behaves under duality. For a linear
subspace M of X, let Ml. := {cp E X* : cp(x) = 0 for all x E M}, and for a linear
subspace N of X*, let l.N := {x EX: cp(x) = 0 for all cp E N}. By the bipolar
theorem, l.(Ml.) is the norm-closure of M, and (l.N)l. is the weak-*-closure of
N. Moreover, for every T E L(X), it is well known that N(T*) = R(T)l. and
N(T) = l.R(T*), while R(T) is a norm-dense subspace of l.N(T*), and R(T*) is a
weak-*-dense subspace of N(T)l.. An elementary short proof of the following result
may be found in [2,4.1].
PROPOSITION 11. For every operator T E L(X), the following assertions hold:
(a) K(T) ~ l.Ho(T*) and K(T*) ~ Ho(T)l.;
(b) if Ho(T) + R(T) is norm-dense in X, then T* has SVEP at 0;
(c) if Ho(T*) + R(T*) is weak-*-dense in X*, then T has SVEP at o. 0
Even in the Hilbert space setting, the inclusions in part (a) of Proposition 11
need not be identities, and the implications of parts (b) and (c) cannot be reversed
in general; see [2] for counter-examples in the class of weighted shifts. However, for
suitable classes of operators, the results can be improved.
As usual, an operator T E L(X) is said to be a semi-Fredholm operator, if either
N(T) is finite-dimensional and R(T) is closed, or R(T) is of finite codimension in
X. Also, an operator T E L(X) is said to be semi-regular, if R(T) is closed and
N(T) ~ R(T); see [18], [19], and [21] for a discussion of these operators.
THEOREM 12. Suppose that the operator T E L(X) is either semi-Fredholm or
semi-regular. Then the following assertions hold:
(a) R(T) = K(T) = l.Ho(T*) = l.N(T*);
(b) R(T*) = K(T*) = Ho(T)l. = N(T)l.;
(c) N(T) n R(T) = {O} <=> T has SVEP at 0;
(d) N(T*) n R(T*) = {O} <=> T* has SVEP at 0;
(e) N(T) + R(T) = X<=> Ho(T) + R(T) = X<=> T* has SVEP at 0;
(f) N-;-;:-;:(T=*""7"")-+-:R~(T=*"-;-) w' = X* <=> Ho(T*) + R(T*) w' = X* <=> T has SVEP at 0,
where w* indicates the closure with respect to the weak-*-topology. o
256 T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN

Theorem 12 was recently obtained in [2], see also [5]. An important ingredient
of the proof is the fact that T has SVEP at 0 if and only Tn has SVEP at 0 for
arbitrary n E N. This equivalence is a special case of a spectral mapping formula
for the set 6(T) of all A E C at which T fails to have SVEP, namely 6(f(T)) =
f(6(T)) for every analytic function f on some open neighborhood of a(T); see [2,
3.1] and also Theorem 18 below. Further developments may be found in [1], [3],
[4], [5], [17], and [20]. Here we mention only one simple consequence of Theorem
12 for semi-regular operators from [3, 2.13].
For T E L(X), let PK(T) consist of all A E C for which T - A is semi-regular.
The Kato spectrum aK(T) := C \ PK(T) is a closed subset of a(T) and contains
oa(T); see [18, 3.1] and [21] for details. We include a short proof of the following
result, since the dichotomy for the connected components of the Kato resolvent set
PK (T) with respect to the localized SVEP will play an essential role in Section 4.

THEOREM 13. Let T E L(X) be semi-regular. Then T has SVEP at 0 precisely


when T is injective, or, equivalently, when T is bounded below, while T* has SVEP
at 0 precisely when T is surjective.
Moreover, for arbitrary T E L(X), each connected component n of PK(T) sat-
isfies either n ~ 6(T) or n n 6(T) = 0. The inclusion n ~ 6(T) OCC1J.rs pre-
cisely when n ~ ap(T), or, equivalently, when n n aap(T) i=- 0, while the iden-
tity n n 6(T) = 0 occurs precisely when n n ap(T) = 0, or, equivalently, when
n \ aap(T) i=- 0.
Proof. If T is semi-regular, then N(T) n n(T) = N(T) and N(T) + R(T) =
R(T) = R(T). Hence the first assertions follow from parts (c) and (e) of Theorem
12. For the last claim, it suffices to see that injectivity of T - A for some A E n entails
that T - f.J. is injective for every f.J. E n. But this is clear, since, by part (b) of Theorem
12, N(T - f.J.) = J..n(T* -f.J.) and, by [18, 3.1.6 and 3.1.11], n(T* - f.J.) = n(T* -A)
for all f.J. E n. 0
It is well known that the approximate point spectrum and the surjectivity
spectrum of an arbitrary operator T E L(X) are related by the duality formulas
aap(T) = asu(T*) and asu(T) = aap(T*), [18, 1.3.1]. Moreover, by [18, 1.3.2 and
3.1.7], asu(T) = a(T) and aap(T) = aK(T) if T has SVEP, and aap(T) = a(T) and
asu(T) = aK(T) if T* has SVEP. The following local version of these results is
immediate from Theorem 13.
PROPOSITION 14. For every operator T E L(X), the following assertions hold:
(a) If A E a(T) \ aap(T), then T has SVEP at A, but T* fails to have SVEP at A;
(b) if A E a(T) \ asu(T), then T* has SVEP at A, but T fails to have SVEP at A.
o
The next result from [2, 5.2] is a straightforward consequence of Proposition
14. For instance, it follows that 6(T*) is the open unit disc for every non-invertible
operator T with property (P) or (B). Further examples including analytic Toeplitz
operators, composition operators on Hardy spaces, and weighted shifts may be
found in [2].
COROLLARY 15. If aap(T) ~ oa(T), then T has SVEP and 6(T*) = int a(T).
Similarly, if asu(T) ~ oa(T), then T* has SVEP and 6(T) = int a(T). 0
THE SPECTRAL THEORY OF OPERATORS ON BANACH SPACES 257

4. Localization of the properties ((:J) and (8)

There is a natural extension of the class of decomposable operators for which


spectral decompositions are only required with respect to a given open subset U
of the complex plane. These operators were introduced by Vasilescu as residually
decomposable operators in 1969, shortly after the publication of the seminal mono-
graph [10]. They became also known as S-decomposable operators with S = c \ U,
and were studied by Bacalu, Nagy, Vasilescu, and others; see [29, eh.4].
As in [6] and [13], we now say that an operator T E L(X) on a complex Banach
space X is decomposable on an open subset U of C provided that, for every finite
open cover {Vl , ... , Vn } of C with C \ U ~ Vl , there exist Xl,"" Xn E Lat(T)
with the property that
X = Xl + ... + Xn and a(T I X k ) ~ Vk for k = 1, ... , n.
It is known, although certainly not obvious, that, in this definition, it suffices to
consider the case n = 2; see [6] and [29]. Evidently, classical decomposability occurs
when U = C. On the other hand, every operator T E L(X) is at least decomposable
on its resolvent set p(T).
Among the remarkable early accomplishments of the theory is the following
result due to Nagy [25] from 1979: For every T E L(X), there exists a largest
open set U ~ C on which T is decomposable. The complement of this set is Nagy's
spectral residuum Sr(T), a closed, possibly empty, subset of a(T).
In the present section, we shall employ the recent results of Albrecht and Es-
chmeier [6] to obtain a short proof for the existence and a useful description of
Nagy's spectral residuum. In particular, we shall see how Sr(T) is related to the
Kato spectrum aK(T) and the essential spectrum ae(T).
For this, we shall work with certain localized versions of property ((:J) and
property (8) from [6]. An operator T E L(X) is said to possess Bishop's property
((:J) on an open set U ~ C, if, for every open subset V of U, the operator Tv is
injective with closed range, equivalently, if, for every sequence of analytic functions
In: V ---+ X for which (T->')In(>\) ---+ 0 as n ---+ 00 locally uniformly on V, it follows
that In(>') ---+ 0 as n ---+ 00, again locally uniformly on V. It is straightforward to
check that this condition is preserved under arbitrary unions of open sets. This
shows that there exists a largest open set on which T has property ((:J), denoted by
U{3(T). Its complement S{3(T) := C \ U{3(T) is a closed, possibly empty, subset of
a(T). In fact, T satisfies Bishop's classical property ((:J) precisely when S{3(T) = 0.
Moreover, the operator T is said to have property (8) on U, if
X = XT(C \ V) + XT(W)
for all open sets V, W ~ C for which C \ U ~ V ~ V ~ W; see [6] and [13]. Quite
remarkably, as shown in [6, Th.3], this condition holds precisely when, for each
closed set F ~ C and every finite open cover {VI"'" Vn } of F with F \ U ~ VI,
it follows that XT(F) ~ XT(V d + ... + XT(V n); see also [18, 2.2.2] for the case
U=C,
These localized versions of ((:J) and (8) already proved to be useful in the theory
of invariant subspaces for operators on Banach spaces, [14]. The following result
summarizes the main accomplishments from [6].
258 T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN

THEOREM 16. For every operator T E L(X) and every open set U ~ C, the
following equivalences hold:
(a) T has (13) on U {::} T* has (8) on U;
(b) T has (8) on U {::} T* has (/3) on U;
(c) T is decomposable on U {::} T has both (13) and (8) on U;
(d) T has (13) on U {::} T is the restriction of a decomposable operator on U;
(e) T has (15) on U {::} T i,~ the quotient of a decomposable operator on U. 0
It is not at all obvious from the definition of (8) that there exists a largest open
set, say Uc5(T), on which the operator T has property (8), but this now follows from
the corresponding result for (13) by duality. In fact, Uc5(T) = U{3(T*) by part (b) of
the preceding result. More precisely, Theorem 16 leads to the following result.
COROLLARY 17. For every operator T E L( X), there exists a smallest closed set
Sc5(T) so that T has property (8) on its complement. Moreover, Sc5(T) = S{3(T*),
S{3(T) = Sc5(T*), and Sr(T) = S{3(T) U Sc5(T) = S{3(T) U S{3(T*) = Sr(T*). 0

Perhaps somewhat surprisingly, it will be possible to obtain general information


about the location of S{3(T), and hence of Sc5(T) and Sr(T). For this, another
localized version of SVEP will play a crucial role.
For consistency, we say that the operator T E L(X) has SVEP on an open set
U ~ C, if, for every open subset V of U, the operator Tv is injective, [13]. It is
straightforward to check that T has SVEP on U precisely when T has SVEP at each
point >. E U, as defined in the previous section. Obviously, there exists a largest
open set on which T has SVEP, and the analytic spectral residuum S(T) is defined
to be the complement of this set; see [29, 4.3.2] and [30]. Clearly, 6(T) ~ S(T),
but, since 6(T) is open and S(T) is closed, equality occurs only in the trivial case
when T has SVEP. Nevertheless, as noted in [2], a simple verification shows that
6(T) = S(T) ~ S{3(T).
It is interesting to observe that all these residual sets behave canonically with
respect to the Riesz functional calculus. As usual, for T E L(X) and any analytic
complex-valued function f on an open neighborhood 0 of a(T), the operator f(T) E
L(X) is defined by
f(T) := ~
2m
rf(>.)(>. - T)-l d>',
lr
where r denotes an arbitrary contour in 0 that surrounds a(T), [12] or [18, A.2].
The standard spectral mapping theorem asserts that a(f(T)) = f(a(T)). The
next result has a similar flavor, and may be viewed as an extension of the fact that
the classical versions of SVEP, property (/3), property (8), and decomposability
are all preserved under the Riesz functional calculus, [18, 3.3.6 and 3.3.9]. The
fact that the Riesz functional calculus respects Bishop's classical property (13) was
established by Eschmeier and Putinar [15]. The following proof involves a different
approach to this result.
THEOREM 18. Let T E L(X) be an arbitrary operator, let f : 0 -+ C be an
analytic function on an open neighborhood 0 of a(T), and suppose that f is non-
constant on each connected component ofO. If E denotes any of the symbols 6, S,
S{3, Sc5, or S., then
E(f(T)) = f(E(T)).
THE SPECTRAL THEORY OF OPERATORS ON BANACH SPACES 259

Proof. In the case of 6(T), the spectral mapping formula was recently ob-
tained in [2, 3.1]. The result for S(T) is a standard fact that may be found in [29,
4.3.14] and [30, 1.6]. Note, however, that the formula for S(T) is also an imme-
diate consequence of that for 6(T), because 6(T) = S(T). While the existence of
the residual set So(T) was most conveniently established by using S{3(T) and the
Albrecht-Eschmeier duality between the localized versions of (f3) and (8), for the is-
sue at hand it seems appropriate to switch the order. Indeed, since f(T*) = f(T) *,
Corollary 17 ensures that it suffices to prove the claim for So(T). For this, fortu-
nately, we may proceed as in the proof of [18,3.3.6 and 3.3.9], where property (8)
is shown to be stable under the Riesz functional calculus.
First, consider arbitrary open sets V, W ~ C for which f(So(T)) ~ V ~ V ~ W.
Then {f-l(C \ V), f- 1(W)} is an open cover of a(T) for which So(T) ~ f-l(W).
Thus, by the characterization of the localized property (8) mentioned above, we
obtain that
X = XT(a(T)) = XT (1-1(C \ V) n a(T)) + XT (1-1(W) n a(T)) .
Clearly, f- 1(C \ V) n a(T) ~ f- 1(C \ V) n a(T) and, similarly, f-l(W) n a(T) ~
f-l(W) n a(T). Since, by [18,3.3.6], the formula XT (f-l(F) n a(T)) = Xf(T) (F)
holds for every closed set F ~ C, we conclude that
X = Xf(T) (C \ V) + Xf(T) (W).
This shows that f(T) has (8) on C \ f(So(T)), thus C \ f(So(T)) ~ Uo(f(T)),
and hence So(f(T)) ~ f(So(T)). Note that this inclusion even holds without the
requirement that f be non-constant on each of the connected components of its
domain.
The reverse inclusion is less obvious, but may be obtained by a suitable mod-
ification of the proof of [18, 3.3.9]. Let S := f- 1 (So(f(T))) n a(T). Then the
desired inclusion f(So(T)) ~ So(f(T)) means precisely that the decomposition
X = XT(C \ V) + XT(W) holds for all open sets V, W ~ C for which S ~ V ~
V ~ W. Evidently, it suffices to show that X = XT(G) + XT(H) for every open
cover {G,H} of a(T) for which S ~ G, SnH = 0, and both G and H are compact
subsets of n. Ignoring momentarily the exceptional set S, we may proceed word
by word along the lines of the proof of [18, 3.3.9] to obtain a finite open cover
{WI, ... , W n } of a(T) in n for which

fork=I, ... ,n.

To handle the residual set, we note that the identity SnH = 0 may be reformulated
in the form So(f(T)) n f(a(T) n H) = 0. Hence, by continuity and compactness,
there exists an open neighborhood V of So (f(T)) for which V n f(a(T) n H) = 0.
This implies that f- 1 (V) n a(T) n H = 0, hence f- 1 (V) n a(T) ~ G, and therefore,
by [18, 3.3.6],
Xf(T)(V) = XT (J-l(V) n a(T)) ~ XT(G).
Now, since f(T) has (8) on C \ So(f(T)), and since {V, f(W1), ... , f(Wn )} is an
open cover of a(f(T)) = f(a(T)) for which So(f(T)) ~ V, we conclude that

X = Xf(T)(a(T)) = Xf(T)(V) + Xf(T) (f(Wd) + ... + Xf(T) (f(Wn)) '


260 T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN

again by the basic characterization of the localized version of (8) provided in [6,
Th.3]. Thus X = XT(G) + XT(H), as desired. 0
In the following results, we show how to verify property ({3) on the Kato resol-
vent set and its Fredholm counterpart.
THEOREM 19. For an arbitrary operator T E L(X) and every connected com-
ponent n of PK(T), the following equivalences hold:
T has ({3) on n {:} T has SVEP on n {:} n n O'p(T) = 0 {:} n \ O'ap(T) =F 0j
in particular, T has property ({3) on PK(T) precisely when T has SVEP on PK(T).
Proof. Clearly, the first of the displayed conditions implies the second one, and
the equivalence of the last three conditions follows from Theorem 13. Conversely, if
these three conditions hold, then T - A is injective with closed range for all A E n.
Thus f'i,(T - A) = "((T - A) > 0 for all A E n. Moreover, by [18, 3.1.10], for every
compact subset K of n, there exists a constant c > 0 such that f'i,(T - A) > 0 for
all A E Kj in fact, as shown, for instance, in [21,4.1], the function A t-+ "((T - A)
is continuous and strictly positive on PK(T). From this it is immediate that T has
({3) on n. 0
The next result is clear from Corollary 17, Theorem 19, and the well-known
identity PK(T) = PK(T*), [18, 3.1.6]. Part (c) of Corollary 20 may be viewed as
an extension of the fact that, by [18, 3.1.7], p(T) = PK(T) whenever both T and
T* have SVEP.
COROLLARY 20. For every operatorT E L(X), the following assertions hold:(a)
T has SVEP on PK(T) {:} S{3(T) ~ O'K(T)j
(b) T* has SVEP on PK(T) {:} SIi(T) ~ O'K(T)j
(c) T and T* have SVEP on PK(T) {:} Sr(T) ~ O'K(T). 0
To derive the companion result for the essential spectrum, we employ the fact
that, for every operator T E L(X) and every open subset V of the essential resolvent
set Pe(T) := C\O'e(T), the operator Tv has closed range (but need not be injective).
This interesting result was recently obtained by Eschmeier [13, 3.1], based on sheaf-
theoretic methods developed by Putinar [28] to show that quasi-similar operators
with property ({3) have the same essential spectrum. In tandem with Corollary 17,
we obtain the following extension of [13, 3.9].
COROLLARY 21. For every operator T E L(X), the following assertions hold:
(a) T has SVEP on Pe(T) {:} S{3(T) ~ O'e(T)j
(b) T* has SVEP on Pe(T) {:} SIi(T) ~ O'e(T)j
(c) T and T* have SVEP on Pe(T) {:} Sr(T) ~ O'e(T). o
We close with an application to the spectral theory of weighted shifts.
EXAMPLE 22. Let W := (Wn)nEN"o be a bounded sequence of strictly positive
real numbers, and let T E L(X) denote the corresponding unilateral weighted right
shift on the sequence space X := fP(N o) for some 1 ::; p < 00. Clearly,
i(T) = lim inf (Wk .. Wk+n_d 1 / n and r(T) = lim sup (Wk .. Wk+n_d 1 / n .
n-+oo k~O n-+oo k~O
THE SPECTRAL THEORY OF OPERATORS ON BANACH SPACES 261

Since T has no eigenvalues, T has SVEP and asu(T) = a(T) = {A E C : IAI s r(T)},
by [18, 1.3.2 and 1.6.15J. Moreover, as noted in [18,3.7.7],
ae(T) = aK(T) = aap(T) = {A E C : i(T) siAl s r(T)} ,
and therefore, by Corollaries 20 or 21, the annulus {A E C : i(T) siAl s r(T)} con-
tains S(3(T). On the other hand, by [2, 6.1J, 6(T*) = {A E C : IAI < c(T)} , where
c(T):= liminf(wl" 'Wn)l/n.
n->oo

Thus, by Corollary 17, it follows that S5(T) = S(3(T*) 2 {A E C : IAI S c(T)}. We


finally note that, by [22, 2.7J, condition (f3) on T implies that i(T) = r(T) and
aT(x) = a(T) for all non-zero x E X, while, by [22, 3.3J or [23, Prop.5], a certain
growth condition of exponential type for the weight sequence w suffices to ensure
that T has (f3). 0

References
[IJ P. Aiena, M. L. Colasante, and M. Gonzalez, Opemtors which have a closed quasi-nilpotent
part, Proc. Amer. Math. Soc. 130 (2002), 2701-2710.
[2J P. Aiena, T. L. Miller, and M. M. Neumann, On a localized single-valued extension property,
to appear in Proc. Royal Irish Acad.
[3J P. Aiena and O. Monsalve, Opemtors which do not have the single valued extension property,
J. Math. Anal. Appl. 250 (2000),435-448.
[4J P. Aiena and O. Monsalve, The single valued extension property and the genemlized Kato
decomposition property, Acta Sci. Math. (Szeged) 67 (2001), 791-807.
[5J P. Aiena and F. Villafane, Components of resolvent sets and local spectml theory, submitted
to this volume.
[6J E. Albrecht and J. Eschmeier, Analytic functional models and local spectml theory, Proc.
London Math. Soc. (3) 75 (1997), 323-348.
[7J E. Albrecht and W. J. Ricker, Local spectml theory f01 opemtors with thin spectrum, preprint,
University of Saarbriicken, 2002.
[8J H. Bercovici and S. Petrovic, Genemlized scalar opemtors as dilations, Proc. Amer. Math.
Soc. 123 (1995), 2173-2180.
[9J E. Bishop, A duality theory for an arbitmry opemtor, Pacific J. Math. 9 (1959), 379-397.
[lOJ I. Colojoara and C. Foi~, Theory of Genemlized Spectml Opemtors, Gordon and Breach,
New York, 1968.
[11J M. Didas, E(]"n )-subscalar n-tuples and the Cesaro opemtor on HP, Annales Universitatis
Saraviensis, Series Mathematicae 10 (2000), 285-335.
[12J N. Dunford and J. T. Schwartz, Linear Opemtors III, Wiley-Interscience, New York, 1971.
[13J J. Eschmeier, On the essential spectrum of Banach-space opemtors, Proc. Edinburgh Math.
Soc. (2) 43 (2000), 511-528.
[14J J. Eschmeier and B. Prunaru, Invariant subspaces for opemtors with Bishop's property ({3)
and thick spectrum, J. Funct. Anal. 94 (1990), 196-222.
[15J J. Eschmeier and M. Putinar, Bishop's condition ({3) and rich extensions of linear opemtors,
Indiana Univ. Math. J. 37 (1988), 325-348.
[16J J. Eschmeier and M. Putinar, Spectml Decompositions and Analytic Sheaves, Clarendon
Press, Oxford, 1996.
[17J J. K. Finch, The single valued exten.9ion property on a Banach space, Pacific J. Math. 58
(1975),61-69.
[18J K. B. Laursen and M. M. Neumann, An Introduction to Local Spectml Theory, Clarendon
Press, Oxford, 2000.
[19J M. Mbekbta, Genemlisation de la decomposition de Kato aux opemteurs pamnormaux et
spectmux, Glasgow Math. J. 29 (1987), 159-175.
[20J M. Mbekhta, Sur la theorie spectmle locale et limite des nilpotents, Proc. Amer. Math. Soc.
110 (1990), 621-631.
262 T. L. MILLER, V. G. MILLER, AND M. M. NEUMANN

[21] M. Mbekhta and A. Ouahab, Operateurs s-regulier dans un espace de Banach et theorie
spectrale, Acta Sci. Math. (Szeged) 59 (1994), 525-543.
[22] T. L. Miller, V. G. Miller, and M. M. Neumann, Local spectral properties of weighted shifts,
to appear in J. Operator Theory.
[23] T. L. Miller, V. G. Miller, and M. M. Neumann, Growth conditions and decomposable exten-
sions, to appear in Contemp. Math.
[24] T. L. Miller, V. G. Miller, and M. M. Neumann, On operators with closed analytic core, to
appear in Rend. Cire. Mat. Palermo (2) 51 (2002).
[25] B. Nagy, On S-decomposable operators, J. Operator Theory 2 (1979),277-286.
[26] M. M. Neumann, Recent developments in local spectral theory, Rend. Circ. Mat. Palermo (2)
Suppl. 68 (2002), 111-131.
[27] M. Putinar, Hyponormal operators are subsealar, J. Operator Theory 12 (1984), 385-395.
[28] M. Putinar, Quasi-similarity of tuples with Bishop's property (,8), Integral Equations Oper-
ator Theory 15 (1992), 1047-1052.
[29] F.-H. Vasilescu, Analytic FUnctional Calculus and Spectral Decompositions, Editura
Aeademiei and D. Reidel Publishing Company, Bucharest and Dordreeht, 1982.
[30] P. Vrbova., On local spectral properties of operators in Banach spaces, Czechoslovak Math.
J. 23 (98) (1973),483-492.

DEPARTMENT OF MATHEMATICS AND STATISTICS, MISSISSIPPI STATE UNIVERSITY, MISSISSIPPI


STATE, MS 39762, USA
E-mail address: neumannOmath.msstate.edu
Contemporary Mathematics
Volume 328, 2003

Abstract harmonic analysis, homological algebra, and


operator spaces

Volker Runde

ABSTRACT. In 1972, B. E. Johnson proved that a locally compact group G


is amenable if and only if certain Hochschild cohomology groups of its convo-
lution algebra Ll(G) vanish. Similarly, G is compact if and only if Ll(G) is
biprojective: In each case, a classical property of G corresponds to a cohomo-
logical propety of Ll(G). Starting with the work of Z.-J. Ruan in 1995, it has
become apparent that in the non-commutative setting, i.e. when dealing with
the Fourier algebra A(G) or the Fourier-Stieltjes algebra B(G), the canon-
ical operator space structure of the algebras under consideration has to be
taken into account: In analogy with Johnson's result, Ruan characterized the
amenable locally compact groups G through the vanishing of certain cohomol-
ogy groups of A(G). In this paper, we give a survey of historical developments,
known results, and current open problems.

1. Abstract harmonic analysis, ...


The central objects of interest in abstract harmonic analysis are locally compact
groups, i.e. groups equipped with a locally compact Hausdorff topology such that
multiplication and inversion are continuous. This includes all discrete groups, but
also all Lie groups. There are various function spaces associated with a locally
compact group G, e.g. the space Co(G) of all continuous functions on G that vanish
at infinity. The dual space of Co(G) can be identified with M(G), the space of all
regular (complex) Borel measures on G. The convolution product * oftwo measures
is defined via
(1,11-* v):= LLf(XY)dJ.L(X)V(Y) (J.L,V E M(G), f E Co(G))

and turns M(G) into a Banach algebra. Moreover, M(G) has an isometric involu-
tion given by

(I,J.L*):= Lf(x-1)dJ.L(X) (J.L E M(G), f E Co(G)).

1991 Mathematics Subject Classification. 22D15, 22D25, 43A20, 43A30, 46H20 (primary),
46H25, 46L07, 46M18, 46M20, 47B47, 47L25, 47L50.
Key words and phrases. locally compact groups, group algebra, Fourier algebra, Fourier-
Stieltjes algebra, Hochschild cohomology, homological algebra, operator spaces.
Financial support by NSERC under grant no. 227043-00 is gratefully acknowledged.

2003 American Mathematical Society


263
264 VOLKER RUNDE

The most surprising feature of an object as general as a locally compact group


is the existence of (left) Haar measure: a regular Borel measure which is invariant
under left translation and unique up to a multiplicative constant. For example,
the Haar measure of a discrete group is simply counting measure, and the Haar
measure of ]RN, is N-dimensional Lebesgue measure. The space Ll(G) of all in-
tegrable functions with respect to Haar measure can be identified with a closed
*-ideal of M(G) via the Radon-Nikodym theorem. Both M(G) and Ll(G) are
complete invariants for G: Whenver Ll(G 1 ) and 1(G 2 ) (or M(Gt} and M(G 2 ))
are isometrically isomorphic, then G 1 and G 2 are topologically isomorphic. This
means that all information on a locally compact group is already encoded in Ll (G)
and M(G). For example, Ll(G) and M(G) are abelian if and only if G is abelian,
and Ll (G) has an identity if and only if G is discrete.
References for abstract harmonic analysis are [Fol], [H-R], and [R-St].
The property of locally compact groups we will mostly be concerned in this
survey is amenability. A a mean on a locally compact group G is a bounded
linear functional m: LOC(G) ---+ C such that (1, m) = Ilmil = 1. For any function
I on G and for any x E G, we write LxI for the left translate of I by x, i.e.
(Lxf)(y) := I(xy) for y E G.
DEFINITION 1.1. A locally compact group G is called amenable if there is a
(left) translation invariant mean on G, i.e. a mean m such that
(, m) = (L x , m) ( E LOC(G), x E G).
EXAMPLE 1.2. (1) Since the Haar measure of a compact group G is finite,
LOC(G) C 1(G) holds. Consequently, Haar measure is an invariant mean
on G.
(2) For abelian G, the Markov-Kakutani fixed point theorem yields an invari-
ant mean on G.
(3) The free group in two generators is not amenable ([Pat, (0.6) Exanlple]).
Moreover, amenability is stable under standard constructions on locally com-
pact groups such as taking subgroups, quotients, extensions, and inductive limits.
Amenable, locally compact groups were first considered by J. v. Neumann
([Neu]) in the discrete case; he used the term "Gruppen von endlichem MaB".
The adjective amenable for groups satisfying Definition 1.1 is due to M. M. Day
([Day]), apparently with a pun in mind: They are amenable because they have an
invariant mean, but also since they are particularly pleasant to deal with and thus
are truly amenable - just in the sense of that adjective in everyday speech.
For more on the theory of amenable, locally compact groups, we refer to the
monographs [Gre], [Pat], and [Pie].

2. homological algebra, ...


We will not attempt here to give a survey on a area as vast as homological
algebra, but outline only a few, basic cohomological concepts that are relevant in
connection with abstract harmonic analysis. For the general theory of homological
algebra, we refer to [C-E], [MacL], and [Wei]. The first to adapt notions from
homological algebra to the functional analytic context was H. Kamowitz in [Kam].
Let 2l be a Banach algebra. A Banach 2l-bimodule is a Banach space E which is
also an 2l-bimodule such that the module actions of 2l on E are jointly continuous.
ABSTRACT HARMONIC ANALYSIS 265

A derivation from 2l to E is a (bounded) linear map D: 2l --+ E satisfying


D(ab) = a . Db + (Da) . b (a, bE !2l);
the space of all derivation from 2l to E is commonly denoted by ZI(!2l, E). A
derivation D is called inner if there is x E E such that
Da = ax-xa (a E !2l).
The symbol for the subspace of ZI (!2l, E) consisting of the inner derivations is
B 1 (2l,E); note that B 1 (!2l,E) need not be closed in ZI(!2l,E).
DEFINITION 2.1. Let !2l be a Banach algebra, and let E be a Banach 2l-bi-
module. Then then the first Hochschild cohomology group 'HI (2l, E) of 2l with
coefficients in E is defined as
'H 1 (!2l, E) := ZI(!2l, E)/B 1 (2l, E).
The name Hochschild cohomology group is in the honor of G. Hochschild
who first considered these groups in a purely algebraic context ([Hoch 1] and
[Hoch 2]).
Given a Banach !2l-bimodule E, its dual space E* carries a natural Banach
2l-bimodule structure via
(x,a ) := (x a,) and (x, a) := (a x,) (a E !2l, E E*, x E E).
We call such Banach !2l-bimodules dual.
In his seminal memoir [Joh 1], B. E. Johnson characterized the amenable lo-
cally compact groups G through Hochschild cohomology groups of Ll(G) with
coefficients in dual Banach l(G)-bimodules ([Joh 1, Theorem 2.5]):
THEOREM 2.2 (B. E. Johnson). Let G be a locally compact group. Then G is
amenable if and only if'Hl(Ll(G),E*) = {O} for each Banach Ll(G)-bimodule E.
The relevance of Theorem 2.2 is twofold: First of all, homological algebra is a
large and powerful toolkit - the fact that a certain property is cohomological in
nature allows to apply its tools, which then yield further insights. Secondly, the
cohomological triviality condition in Theorem 2.2 makes sense for every Banach
algebra. This motivates the following definition from [Joh 1]:
DEFINITION 2.3. A Banach algebra 2l is called amenable if 'Hl(!2l, E*) = {O}
for each Banach 2l-bimodule E.
Given a new definition, the question of how significant it is arises naturally.
Without going into the details and even without defining what a nuclear C* -algebra
is, we would like to only mention the following very deep result which is very much
a collective accomplishment of many mathematicians, among them A. Connes, M.
D. Choi, E. G. Effros, U. Haagerup, E. C. Lance, and S. Wassermann:
THEOREM 2.4. A C* -algebra is amenable if and only if it is nuclear.
For a relatively self-contained exposition of the proof, see [Run, Chapter 6].
Of course, Definition 2.3 allows for modifications by replacing the class of all
dual Banach 2l-bimodules by any other class. In [B-C-D], W. G. Bade, P. C. Cur-
tis, Jr., and H. G. Dales called a commutative Banach algebra !2l weakly amenable
if and only if 'HI (2l, E) = {O} for every symmetric Banach !2l-bimodule E, i.e.
satisfying
ax=xa (a E !2l, x E E).
266 VOLKER RUNDE

This definition is of little use for non-commutative 21. For commutative 21, weak
amenability, however, is equivalent to 'Jtl(21, 21*) = {O} ([B-C-D, Theorem 1.5]),
and in [Joh 2], Johnson suggested that this should be used to define weak amenabil-
ity for arbitrary 21:
DEFINITION 2.5. A Banach algebra 21 is called weakly amenable if 'Jtl(21, 21*) =
{O}.
REMARK 2.6. There is also the notion of a weakly amenable, locally compact
group ([C-H]). This coincidence of terminology, however, is purely accidental.
In contrast to Theorem 2.2, we have:
THEOREM 2.7 ([Joh 3]). Let G be a locally compact group. Then l(G) is
weakly amenable.
For a particularly simple proof of this result, see [D-Gh]. For M(G), things
are strikingly different:
THEOREM 2.8 ([D-Gh-H]). Let G be a locally compact group. Then M(G) is
weakly amenable if and only if G is discrete. In particular, M (G) is amenable if
and only if G is discrete and amenable.
Sometime after Kamowitz's pioneering paper, several mathematicians started
to systematically develop a homological algebra with functional analytic overtones.
Besides Johnson, who followed Hochschild's original approach, there were
A. Guichardet ([Gui]), whose point of view was homological rather than coho-
mological, and J. A. Taylor ([Tay]) and - most persistently - A. Ya. Helemskil
and his Moscow school, whose approaches used projective or injective resolutions;
Helemskil's development of homological algebra for Banach and more general topo-
logical algebras is expounded in the monograph [He} 2].
In homological algebra, the notions of projective, injective, and flat modules
play a pivotal role. Each of these concepts tranlates into the functional analytic
context. Helemskil calls a Banach algebra 21 biprojective (respectively biflat) if it
is a projective (respetively flat) Banach 21-bimodule over itself. We do not attempt
to give the fairly technical definitions of a projective or a flat Banach 21-bimodule.
Fortunately, there are equivalent, but more elementary characterizations of bipro-
jectivity and biflatness, respectively.
We use -y to denote the completed projective tensor product of Banach spaces.
If 21 is a Banach algebra, then 21 -y 21 has a natural Banach 21-bimodule structure
via
a(xy):=axy and (xy)a=:xya (a, x, y E 21).
This turns the multiplication operator
A: 21 -y 21 -+ 21, a b f-+ ab
into a homomorphism of Banach 21-bimodules.
DEFINITION 2.9. Let 21 be a Banach algebra. Then:
(a) 21 is called biprojective if and only if A has bounded right inverse which
is an 21-bimodule homomorphism.
(b) 21 is called biftat if and only if A * has bounded left inverse which is an
21-bimodule homomorphism.
ABSTRACT HARMONIC ANALYSIS 267

Clearly, biflatness is a property weaker than biprojectivity.


The following theorem holds ([Hell, Theorem 51]):
THEOREM 2.10 (A. Ya. Helemskil). Let G be a locally compact group. Then
Ll (G) is biprojective if and only if G is compact.
Again, a classical property of G is equivalent to a cohomological property of
Ll(G). The question for which locally compact groups G the Banach algebra Ll(G)
is biflat seems natural at the first glance. However, any Banach algebra is amenable
if and only if it is biflat and has a bounded approximate identity ([Hel 2, Theorem
Vii.2.20]). Since Ll(G) has a bounded approximate identity for any G, this means
that Ll (G) is biflat precisely when G is amenable.
Let G be a locally compact group. A unitary representation of G on a Hilbert
space jj is a group homomorphism 7r from G into the unitary operators on jj which
is continuous with respect to the given topology on G and the strong operator
topology on B(jj). A function
G-+C, Xf-+(7r(x)~,1J)

with ~, TJ E jj is called a coefficient function of 7r.


EXAMPLE 2.11. The left regular representation A of G on L2(G) is given by
A(X)~ := LX-l~ (x E G, ~ E L2(G)).
DEFINITION 2.12 ([Eym]). Let G be a locally compact group.
(a) The Fourier algebra A(G) of G is defined as
A(G) := {f: G -+ C : f is a coefficient function of A}.
(b) The Fourier-Stieltjes algebra B( G) of G is defined as
B( G) := {f: G -+ C : f is a coefficient function of a unitary representation of G}.
It is immediate that A(G) c B(G), that B(G) consists of bounded continuous
functions, and that A(G) C Co(G). However, it is not obvious that A(G) and B(G)
are linear spaces, let alone algebras. Nevertheless, the following are true ([Eym]):
Let C*(G) be the enveloping C*-algebra of the Banach *-algebra Ll(G).
Then B(G) can be canonically identified with C*(G)*. This turns B(G)
into a commutative Banach algebra.
Let VN(G) := A(G)" denote the group von Neumann algebra of G. Then
A(G) can be canonically identified with the unique predual of VN(G).
This turns A( G) into a commutative Banach algebra whose character
space is G.
A(G) is a closed ideal in B(G).
If G is an abelian group with dual group r, then the Fourier and Fourier-
Stieltjes transform, respectively, yield isometric isomorphisms A( G) ~ Ll (r) and
B(G) ~ M(r). Consequently, A(G) is amenable for any abelian locally compact
group G. It doesn't require much extra effort to see that A(G) is also amenable
if G has an abelian subgroup of finite index ([L-L-W, Theorem 4.1] and [For 2,
Theorem 2.2]). On the other hand, every amenable Banach algebra has a bounded
approximate identity, and hence Leptin's theorem ([Lep]) implies that the amena-
bility of A(G) forces G to be amenable. Nevertheless, the tempting conjecture that
A( G) is amenable if and only if G is amenable is false:
268 VOLKER RUNDE

THEOREM 2.13 ([Joh 4]). The Fourier algebra of SO(3) is not amenable.
This leaves the following intriguing open question:
QUESTION 2.14. Which are the locally compact groups G for which A(G) is
amenable?
The only groups G for which A( G) is known to be amenable are those with an
abelian subgroup of finite index. It is a plausible conjecture that these are indeed
the only ones. The corresponding question for weak amenability is open as well.
B. E. Forrest has shown that A( G) is weakly amenable whenever the principal
component of G is abelian ([For 2, Theorem 2.4]).
One can, of course, ask the same question(s) for the Fourier-Stieltjes algebra:
QUESTION 2.15. Which are the locally compact groups G for which B(G) is
amenable?
Here, the natural conjecture is that those groups are precisely those with a
compact, abelian subgroup of finite index. Since A( G) is a complemented ideal in
B( G), the hereditary properties of amenability for Banach algebras ([Run, Theo-
rem 2.3.7]) yield that A( G) has to be amenable whenever B( G) is. It is easy to
see that, if the conjectured answer to Question 2.14 is true, then so is the one to
Question 2.15.
Partial answers to both Question 2.14 and Question 2.15 can be found in
[L-L-W] and [For 2].
3. and operator spaces
Given any linear space E and n E N, we denote the n x n-matrices with entries
from E by Mn(E); if E = C, we simply write Mn. Clearly, formal matrix mul-
tiplication turns Mn(E) into an Mn-bimodule. Identifying Mn with the bounded
linear operators on n-dimensional Hilbert space, we equip Mn with a norm, which
we denote by I I
DEFINITION 3.1. An operator space is a linear space E with a complete norm
II lin on Mn(E) for each n E N such that
(R 1)

II ~ I~ Iln+m = max{llxll n, IIYllm} (n, mEN, x E Mn(E), Y E Mm(E))

and
(R 2)
EXAMPLE 3.2. Let fJ be a Hilbert space. The unique C*-norms on Mn(13(SJ)) ~
13(fJn) turn 13(SJ) and any of its subspaces into operator spaces.
Given two linear spaces E and F, a linear map T: E --+ F, and n E N, we
define the the n-th amplification T(n) : Mn(E) --+ Mn(F) by applying T to each
matrix entry.
DEFINITION 3.3. Let E and F be operator spaces, and let T E 13(E, F). Then:
(a) T is completely bounded if
IITllcb := sup IIT(n) IIB(Mn(E),Mn(F)) < 00.
nEN
ABSTRACT HARMONIC ANALYSIS 269

(b) T is a complete contraction if IITlicb ~ 1.


(c ) T is a complete isometry if T( n) is an isometry for each n EN.
The following theorem due to Z.-J. Ruan marks the beginning of abstract op-
erator space theory:
THEOREM 3.4 ([Rna 1]). Let E be an operator space. Then there is a Hilbert
space Sj and a complete isometry from E into B(Sj).
To appreciate Theorem 3.4, one should think of it as the operator space ana-
logue of the elementary fact that every Banach space is isometrically isomorphic
to a closed subspace of C(O) for some compact Hausdorff space O. One could thus
define a Banach space as a closed subspace of C(O) some compact Hausdorff space
O. With this definition, however, even checking, e.g., that 1 is a Banach space or
that quotients and dual spaces of Banach spaces are again Banach spaces is difficult
if not imposssible.
Since any C* -algebra can be represented on a Hilbert space, each Banach space
E can be isometrically embedded into B(Sj) for some Hilbert space Sj. For an
operator space, it is not important that, but how it sits inside B(Sj).
There is one monograph devoted to the theory of operator spaces ([E-R]) as
well as an online survey article ([Wit et al.]).
The notions of complete boundedness as well as of complete contractivity can
be defined for multilinear maps as well ([E-R, p. 126]). Since this is somewhat more
technical than Definition 3.3, we won't give the details here. As in the category of
Banach spaces, there is a universallinearizer for the right, i.e. completely bounded,
bilinear maps: the projective operator space tensor product ([E-R, Section 7.1]),
which we denote by .
DEFINITION 3.5. An operator space 2l which is also an algebra is called a
completely contractive Banach algebra if multiplication on 2l is a complete (bilinear)
contraction.
The universal property of ([E-R, Proposition 7.1.2]) yields that, for a com-
pletely contractive Banach algebra 2l, the multiplication induces a complete (linear)
contraction ~: 2l2l --+ 2l.
EXAMPLE 3.6. (1) For any Banach space E, there is an operator space
maxE such that, for any other operator space F, every T E B(E,F)
is completely bounded with IITlicb = IITII ([E-R, pp. 47-54]). Given a
Banach algebra 2l, the operator space max2l is a completely contractive
Banach algebra ([E-R, p. 316]).
(2) Any closed subalgebra of B(Sj) for some Hilbert space Sj is a completely
contractive Banach algebra.
To obtain more, more interesting, and - in the context of abstract harmonic
analysis - more relevant examples, we require some more operator space theory.
Given two operator spaces E and F, let
CB(E, F) := {T: E --+ F : T is completely bounded}.
It is easy to check that CB(E, F) equipped with 11llcb is a Banach space. To define
an operator space structure on CB(E, F), first note that Mn(F) is, for each n E N,
an operator space in a canonical manner. The (purely algebraic) identification
Mn(CB(E, F)) := CB(E, Mn(F)) (n E N)
270 VOLKER RUNDE

then yields norms 1Illn on the spaces Mn(CB(E, F)) that satisfy (R 1) and (R 2),
which is not hard to verify.
Since, for any operator space E, the Banach spaces E* and CB(E, C) are isome-
trically isomorphic ([E-R, Corollary 2.2.3]), this yields a canonical operator space
structure on the dual Banach space of an operator space. In partiuclar, the unique
predual of a von Neumann algebra is an operator space in a canonical way.
We shall see how this yields further examples of completely contractive Banach
algebras.
We denote the W* -tensor product by .
DEFINITION 3.7. A Hop/-von Neumann algebra is a pair (rot, V), where rot
is a von Neumann algebra, and V is a co-multiplication: a unital, injective, w*-
continuous *-homomorphism V: rot --+ rotrot which is co-associative, i.e. the dia-
gram
rot
v
---+ rotrot

vl 1Vid!lJl

rotrot I rotrotrot
id!lJlV
commutes.
EXAMPLE 3.8. Let G be a locally compact group.
(1) Define V: oo(G) --+ oo(G x G) by letting
(V</(xy) := </>(xy) (</> E oo(G), x, y E G).
Since oo(G)oo(G) ~ oo(G x G), this turns oo(G) into a Hopf-von
Neumann algebra.
(2) Let W*(G) be the enveloping von Neumann algebra of C*(G). There is
a canonical w* -continuous homomorphism w from G into the unitaries
of W* (G) with the following universal property: For any unitary rep-
resentation 7r of G on a Hilbert space, there is unique w* -continuous *-
homomorphism (J: W*(G) --+ 7r(G)" such that 7r = (J 0 w. Applying this
universal property to the representation
G --+ W*(G)W*(G), x 1--+ w(x) w(x)
yields a co-multiplication V: W*(G) --+ W*(G)W*(G).
Given two von Neumann algebras rot and 1)1 with preduals rot* and 1)1*, their
W*-tensor product rot1)1 also has a unique predual (rotI)1)*. Operator space
theory allows to identify this predual in terms of rot* and 1)1* ([E-R, Theorem
7.2.4]):
(rotI)1)* ~ rot*ci~m*.
Since VN(G) VN(H) ~ VN(G x H) for any locally compact groups G and H, this
implies in particular that
A(G x H) ~ A(G)A(H).
Suppose now that rot is a Hopf-von Neumann algebra with predual rot*. The
comultiplication V : rot --+ rotrot is w -continuous and thus the adjoint map of
a complete contraction V. : rot*rot. --+ rot.. This turns rot. into a completely
contractive Banach algebra. In view of Example 3.8, we have:
ABSTRACT HARMONIC ANALYSIS 271

EXAMPLE 3.9. Let G be a locally compact group.


(1) The multiplication on L1(G) induced by \7 as in Example 3.8.1 is just
the usual convolution product. Hence, L1 (G) is a completely contractive
Banach algebra.
(2) The multiplication on B( G) induced by \7 as in Example 3.8.2 is pointwise
multiplication, so that B( G) is a completely contractive Banach algebra.
Since A (G) is an ideal in B (G) and since the operator space strucures A ( G)
has as the predual of VN(G) and as a subspce of B(G) coincide, A(G)
with its canonical operator space structure is also a completely contractive
Banach algebra.
REMARK 3.10. Since A(G) fails to be Arens regular for any non-discrete or
infinite, amenable, locally compact group G ([For 1]), it cannot be a subalgebra
of the Arens regular Banach algebra B(f)). Hence, for those groups, A(G) is not of
the form described in Example 3.6.2.
We now return to homological algebra and its applications to abstract harmonic
analysis.
An operator bimodule over a completely contractive Banach algebra Il is an
operator space E which is also an !!-bimodule such that the module actions of Il on
E are completely bounded bilinear maps. One can then define operator Hochschild
cohomology groups 01t 1 (!!, E) by considering only completely bounded derivations
(all inner derivations are automatically completely bounded). It is routine to check
that the dual space of an operator !!-bimodule is again an operator Il-bimodule, so
that the following definition makes sense:
DEFINITION 3.11 ([Rua 2]). A completely contractive Banach algebra Il is
called operator amenable if 01t 1 (Il,E*) = {O} for each operator Il-bimodule E.
The following result ([Rua 2, Theorem 3.6]) shows that Definition 3.11 is in-
deed a good one:
THEOREM 3.12 (Z.-J. Ruan). Let G be a locally compact group. Then G is
amenable if and only if A( G) is operator amenable.
REMARK 3.13. A Banach algebra!! is amenable if and only if max!! is operator
amenable ([E-R, Proposition 16.1.5]). Since L1(G) is the predual of the abelian
von Neumann algebra LOO(G), the canonical operator space structure on L1(G) is
maxL1(G). Hence, Definition 3.11 yields no information on L1(G) beyond Theorem
2.2.
The following is an open problem:
QUESTION 3.14. Which are the locally compact groups G for which B(G) is
operator amenable?
With Theorem 2.8 and the abelian case in mind, it is reasonable to conjecture
that B(G) is operator amenable if and only if G is compact. One direction is
obvious in the light of Theorem 3.12; a partial result towards the converse is given
in [R-Sp].
Adding operator space overtones to Definition 2.5, we define:
DEFINITION 3.15. A completely contractive Banach algebra Il is called operator
weakly amenable if 01t 1 (!!, Il *) = {O}.
272 VOLKER RUNDE

In analogy with Theorem 2.7, we have:


THEOREM 3.16 ([Spr]). Let G be a locally compact group. Then A(G) is op-
erator weakly amenable.
One can translate Helemskir's homological algebra for Banach algebras rela-
tively painlessly to the operator space setting: This is done to some extent in [Ari]
and [Woo 1]. Of course, appropriate notions of projectivity and flatness play an
important role in this operator space homological algebra. Operator biprojectivity
and biflatness can be defined as in the classical setting, and an analogue - with
instead of Q$)-y - of the characterization used for Definition 2.9 holds.
The operator counterpart of Theorem 2.10 was discovered, independently, by
O. Yu. Aristov and P. J. Wood:
THEOREM 3.17 ([Ari], [Woo 2]). Let G be a locally compact group. Then G
is discrete if and only if A(G) is operator biprojective.
As in the classical setting, both operator amenability and operator biprojectiv-
ity imply operator biflatness. Hence, Theorem 3.17 immediately supplies examples
of locally compact groups G for which A( G) is operator biflat, but not operator
amenable. A locally compact group is called a [SIN]-group if Ll(G) has a bounded
approximate identity belonging to its center. By [R-X, Corollary 4.5], A(G) is also
operator biflat whenever G is a [SIN]-group. It may be that A( G) is operator biflat
for every locally compact group G: this question is investigated in [A-R-Sp].
All these results suggest that in order to get a proper understanding of the
Fourier algebra and of how its cohomological properties relate to the underlying
group, one has to take its canonical operator space structure into account.

References
[Ari] O. Yu. Aristov, Biprojective algebras and operator spaces. J. Math. Sci. (to appear).
[A-R-Sp] O. Yu. Aristov, V. Runde, and N. Spronk. Operator biflatness of the Fourier algebra.
In preparation.
[B-C-D] W. G. Bade, P. C. Curtis, Jr., and H. G. Dales, Amenability and weak amenability
for Beurling and Lipschitz algebras. Proc. London Math. Soc. (3) 55 (1987), 359-377.
[C-E] H. Cartan and S. Eilenberg, Homological algebra. Princeton University Press, Prince-
ton, 1956.
[C-H] M. Cowling and U. Haagerup, Completely bounded multipliers of the Fourier algebra
of a simple Lie group of real rank one. Invent. Math. 96 (1989), 507-549.
[D-Gh-H] H. G. Dales, F. Ghahramani, and A. Va. HelemskiY, The amenability of measure
algebras. J. London Math. Soc. 66 (2002), 213-226.
[Day] M. M. Day, Means on semigroups and groups. Bull. Amer. Math. Soc. 55 (1949),
1054-1055.
[D-Gh] M. Despic and F. Ghahramani, Weak amenability of group algebras of locally compact
groups. Ganad. Math. Bull. 37 (1994), 165-167.
[E-R] E. G. Effros and Z.-J. Ruan, Operator spaces. Clarendon Press, Oxford, 2000.
[Eym] P. Eymard, L'algebre de Fourier d'un groupe localement compact. Bull. Soc. Math.
France 92 (1964), 181-236.
[Fol] G. B. Folland, A course in abstract harmonic analysis. CRC Press, Boca Raton,
Florida, 1995.
[For 1] B. E. Forrest, Arens regularity and discrete groups. Pacific J. Math. 151 (1991),
217-227.
[For 2] B. E. Forrest, Amenability and weak amenability of the Fourier algebra. Preprint
(2000).
[Gre] F. P. Greenleaf, Invariant means on locally compact groups. Van Nostrand, New
York-Toronto-London, 1969.
ABSTRACT HARMONIC ANALYSIS 273

[Gui] A. Guichardet, Sur I'homologie et la cohomologie des algebres de Banach. C. R. Acad.


Sci. Paris, Ser. A 262 (1966), 38-42.
[Her] C. S. Herz, Harmonic synthesis for subgruops. Ann. Inst. Fourier (Grenoble) 23
(1973),91-123.
[H-R] E. Hewitt and K. A. Ross, Abstract harmonic analysis, I and II. Springer Verlag,
Berlin-Heideberg-New York, 1963 and 1970.
[Hell] A. Ya. Helemskil', Flat Banach modules and amenable algebras. Trans. Moscow Math.
Soc. 47 (1985), 199-224.
[HeI2] A. Ya. Helemskil, The homology of banach and topological algebras (translated from
the Russian). Kluwer Academic Publishers, Dordrecht, 1989.
[Hoch 1] G. Hochschild, On the cohomology groups of an associative algebra. Ann. of Math.
(2) 46 (1945), 58-67.
[Hoch 2] G. Hochschild, On the cohomology theory for associative algebras. Ann. of Math. (2)
47 (1946), 568-579.
[Joh 1] B. E. Johnson, Cohomology in Banach algebras. Mem. Amer. Math. Soc. 127 (1972).
[Joh 2] B. E. Johnson, Derivations from Ll(G) into Ll(G) and LOO(G). In: J. P. Pier (ed.),
Harmonic analysis (Luxembourg, 1987), pp. 191-198. Lectures Notes in Mathematics
1359. Springer Verlag, Berlin-Heidelberg-New York, 1988.
[Joh 3] B. E. Johnson, Weak amenability of group algebras. Bull. London Math. Soc. 23
(1991),281-284.
[Joh 4] B. E. JOHNSON, Non-amenability of the Fourier algebra of a compact group. J. London
Math. Soc. (2) 50 (1994),361-374.
[Kam] H. Kamowitz, Cohomology groups of commutative Banach algebras. Trans. Amer.
Math. Soc. 102 (1962), 352-372.
[L-L-W] A. T.-M. Lau, R. J. Loy, and G. A. Willis, Amenability of Banach and C"-algebras
on locally compact groups. Studia Math. 119 (1996), 161-178.
[Lep] H. Leptin, Sur l'algebre de Fourier d'un groupe localement compact. C. R. Acad. Sci.
Paris, Ser. A 266 (1968), 1180-1182.
[MacL] S. MacLane, Homology. Springer Verlag, Berlin-Heidelberg-New York, 1995.
[Neu] J. von Neumann, Zur allgemeinen Theorie des MaBes. Fund. Math. 13 (1929), 73-116.
[Pat] A. L. T. Paterson, Amenability. American Mathematical Society, Providence, 1988.
[Pie] J. P. Pier, Amenable locally compact groups. Wiley-Interscience, New York, 1984.
[R-St] H. Reiter and J. D. Stegeman, Classical harmonic analysis and locally compact
groups. Clarendon Press, Oxford, 2000.
[Rua 1] Z.-J. Ruan, Subspaces of C"-algebras. J. Funct. Anal. 76 (1988), 217-230.
[Rua2] Z.-J. Ruan, The operator amenability of A(G). Amer. J. Math. 117 (1995), 1449-
1474.
[R-X] Z.-J. Ruan and G. Xu, Splitting properties of operator bimodules and operator ame-
nability of Kac algebras. In: A. Gheondea, R. N. Gologan and D. Timotin, Operator
theory, operator algebras, and related topics, pp. 193-216. The Theta Foundation,
Bucharest, 1997.
[Run] v. Runde, Lectures on amenability. Lecture Notes in Mathematics 1774. Springer
Verlag, Berlin-Heidelberg-New York, 2002.
[R-Sp] V. Runde and N. Spronk, Operator amenability of Fourier-Stieltjes algebras. Preprint
(2001).
[Spr] N. Spronk, Operator weak amenability of the Fourier algebra. Proc. Amer. Math.
Soc. 130 (2002), 3609-3617.
[Tay] J. A. Taylor, Homology and cohomology for topological algebras. Adv. in Math. 9
(1970), 137-182.
[Wei] C. A. Weibel, An introduction to homological algebra. Cambridge University Press,
Cambridge, 1994.
[Wit et al.] G. Wittstock et al., What are operator spaces? - An online dictionary. URL:
http://wwv.math.uni-sb.de/~ag-wittstock/projekt2001.html (2001).
[Woo 1] P. J. Wood, Homological algebra in operator spaces with applications to harmonic
analysis. Ph.D. thesis, University of Waterloo, 1999.
[Woo 2] P. J. Wood, The operator biprojectivity of the Fourier algebra. Canadian .1. Math.
(to appear).
274 VOLKER RUNDE

DEPARTMENT OF MATHEMATICAL AND STATISTICAL SCIENC


MONTON, AB, CANADA T6G 2Gl
E-mail address:vrundeClualberta.ca
Contemporary Mathematics
Volume 328, 2003

Relative Tensor Products for Modules over von Neumann


Algebras

David Sherman

ABSTRACT. We give an overview of relative tensor products (RTPs) for von


Neumann algebra modules. For background, we start with the categorical
definition and go on to examine its algebraic formulation, which is applied to
Morita equivalence and index. Then we consider the analytic construction,
with particular emphasis on explaining why the RTP is not generally defined
for every pair of vectors. We also look at recent work justifying a representation
of RTPs as composition of unbounded operators, noting that these ideas work
equally well for LP modules. Finally, we prove some new results characterizing
preclosedness of the map (~, 7) f-> ~ 181",7).

1. Introduction
The purpose of this article is to summarize and explore some of the various
constructions of the relative tensor product (RTP) of von Neumann algebra mod-
ules. Alternately known as composition or fusion, RTPs are a key tool in subfactor
theory and the study of Morita equivalence. The idea is this: given a von Neumann
algebra M, we want a map which associates a vector space to certain pairs of a
right M-module and a left M-module. If we write module actions with subscripts,
we have
(XM,M!i)) f-> X 0M!i).
This should be functorial, covariant in both variables, and appropriately normal-
ized. Other than this, we only need to specify which modules and spaces we are
considering.
In spirit, RTPs are algebraic; a ring-theoretic definition can be found in most
algebra textbooks. But in the context of operator algebras, the requirement that
the output be a certain type of space - typically a Hilbert space - causes an ana-
lytic obstruction. As a consequence, there are domain issues in any vector-based
construction. Fortunately, von Neumann algebras have a sufficiently simple repre-
sentation theory to allow a recasting of RTPs in algebraic terms.
The analytic study of RTPs can be related nicely to noncommutative P spaces.
Indeed, examination of the usual (2) case reveals that the technical difficulties

2000 Mathematics Subject Classification. Primary: 46LIO; Secondary: 46M05.


Key words and phrases. relative tensor product, von Neumann algebra, bimodule.

2003 American Mathematical Society


275
276 DAVID SHERMAN

come from a "change of density". (We say that the density of an LP-type space
is lip.) Once this is understood, it is easy to handle LP modules [JS] as well.
Modular algebras ([Y], [S]) provide an elegant framework, so we briefly explain
their meaning.
The final section of the paper investigates the question, "When is the map
(~, 1]) f--+ ~ <p 1] preclosed?" This may be considered as an extension of Falcone's
theorem [F], in which he found conditions for the map to be everywhere-defined.
We consider a variety of formulations.
We have tried to make the paper as accessible as possible to non-operator
algebraists, especially in the first half. Of course, even at this level many results rely
on familiarity with the projection theory of von Neumann algebras; basic sources
are [TI], [T2], [KR]. Primary references for RTPs are [Sa], [P]' [F], [C2].

2. Notations and background


The basic objects of this paper are von Neumann algebras, always denoted here
by M, N, or P. These can be defined in many equivalent ways:
C*-algebras which are dual spaces.
strongly-closed unital *-subalgebras of B(i)). B(i)) is the set of bounded
linear operators on a Hilbert space i); the strong topology is generated by
the seminorms x f--+ Ilx~ll, ~ E i); the * operation is given by the operator
adjoint.
*-closed subsets of B(i)) which equal their double (iterated) commutant.
The commutant of a set S c B(i)) is {x E B(i)) I xy = yx, 'l:/y E S}.
As one might guess from the definitions, the study of von Neumann algebras turns
on the interplay between algebraic and analytic techniques.
Finite-dimensional von Neumann algebras are direct sums of full matrix alge-
bras. At the other extreme, commutative von Neumann algebras are all of the form
Loo(X, J-l) for some measure space (X, J-l)' so the study of general von Neumann al-
gebras is considered "noncommutative measure theory." Based on this analogy, the
(unique) predual M* of M is called Ll(M); it is the set of normal (= continuous
in yet another topology, the a-weak) linear functionals on M c B(i)), and can be
thought of as "noncommutative countably additive measures". A functional r.p is
positive when x> 0 =} r.p(x) 2: 0; the set of positive normal functionals is denoted
M;t. The support s( r.p) of a positive normal linear functional r.p is the smallest pro-
jection q E M with r.p(l - q) = O. So if M is abelian, r.p corresponds to a measure
and q is the (indicator function of the) usual support.
For simplicity, all modules in this paper are separable Hilbert spaces (except in
Section 6), all algebras have separable predual, all linear functionals are normal, and
all representations are normal and nondegenerate (MSJ or SJM is all of SJ). Two
projections p, q in a von Neumann algebra are said to be (Murray-von Neumann)
equivalent if there exists v E M with v*v = p, vv* = q. Such an element v is called
a partial isometry, and we think of p and q as being "the same size". Subscripts
are used to represent actions, so XM indicates that X is a right M-module, i.e. a
representation of the opposite algebra MOP. It is implicit in the term "bimodule",
or in the notation Mi)N, that the two actions commute. The phrase "left (resp.
right) action of' is frequently abbreviated to L (resp. R) for operators or entire
algebras, so that we speak of L(x) or R(M). Finally, we often write Moo for the
von Neumann algebra of all bounded operators on a separable infinite-dimensional
RELATIVE TENSOR PRODUCTS 277

Hilbert space, and Moc(M) for the von Neumann tensor product MocM. One
can think of this as the set of infinite matrices with entries in M; we will denote
by eij the matrix unit with 1 in the ij position and 0 elsewhere.
The (left) representation theory of von Neumann algebras on Hilbert spaces is
simple, so we recall it briefly. (Most of this development can be found in Chapters
1 and 2 of [JoS].) First, there is a standard construction, due to Gelfand-Neumark
and Segal (abbreviated GNS), for building a representation from 'P E Mt. To
each x E M we formally associate the vector x'P 1/ 2 (various notations are in use,
e.g. 7]",(x) or A",(.'1:), but this one is especially appropriate ([C2] V.App.B, [S])). We
endow this set with the inner product
< x'P1/2, Y'P 1/ 2 >= 'P(Y* x),
and set fJ", to be the closure in the inherited topology, modulo the null space.
The left action of M on fJ", = M'P1/2 is bounded and densely defined by left
composition.
When 'P is faithful (meaning x > 0 => 'P(x) > 0), the vector 'P 1/ 2 = I'P1/2 is
cyclic (M'P1/2 = fJ",) and separating (x =f. 0 => X'P1/2 =f. 0). Now all representa-
tions with a cyclic and separating vector are isomorphic - a sort of "left regular
representation"; we will denote this by ML 2 (M). It is a fundamental fact that the
commutant of this action is antiisomorphic to M, and when we make this iden-
tification we call ML2(M)M the standard form of M. If 'P is not faithful, the
GNS construction produces a vector 'Pl/2 which is cyclic but not separating, and a
representation which is isomorphic to ML2(M)s('P) ([T2], Ch. VIII, IX).
Now let us examine an arbitrary (separable, so countably generated) module
MfJ. Following standard arguments (e.g. [TI] I.9), fJ decomposes into a direct
sum of cyclic representations M(M~n), each of which is isomorphic to the GNS
representation for the associated vector state w~n (=< ~n, ~n . With qn = s(w~J,
we have
MfJ ~ EBMM~n ~ EBMfJw~n ~ EB M L 2(M)qn.
(Here and elsewhere, "~" means a unitary equivalence of (bi)modules.) Since this is
a left module, it is natural to write vectors as rows with the nth entry in L 2(M)qn:

We will call such a decomposition a row representation of MfJ. Here e nn are


diagonal matrix units in Moo, so (Eqn e nn ) is a diagonal projection in Moc(M).
The left action of M is, of course, matrix multiplication (by 1 x 1 matrices) on the
left.
The module (L2(M)L2(M) ... ) will be denoted R2(M) (for "row"). Since the
standard form behaves naturally with respect to restriction - L2(q,Nq) ~ qL 2(,N)q
as bimodules - it follows that L2(Moo(M)) is built as infinite matrices over L2(M)
(see (3.3)). Thus R2(M) can be realized as ellL 2(Moo (M)).

PROPOSITION 2.1. Any countably generated left representation of M on a


Hilbert space is isomorphic to R 2(M)q for some diagonal projection q E Moc(M).
Any projection ,in Moc(M), diagonal or not, defines a module in this way, and two
such modules are isomorphic exactly when the projections are equivalent. In fact
(2.2)
278 DAVID SHERMAN

So isomorphism classes correspond to equivalence classes of projections in lVIoo(M),


which is the monoid V(Moo(M)) in K -theoretic language [W-O}. The direct sum of
isomorphism ciasses of modules corresponds to the sum of orthogonal representatives
in V(Moo(M)), giving a monoidal equivalence.
We denote the category of separable left M-modules by Left L 2(M). For us,
the most important consequence of (2.2) is that
(2.3) C.(MR2(M)q) = R(qMoo(M)q),
where "c." stands for the commutant of the M-action. (In particular, the case
q = ell is just the standard form.) The algebra qMoo(M)q is called an amplification
of M, being a generalization of a matrix algebra with entries in M. Of course
everything above can be done for right modules - the relevant abbreviations are
C 2(M), for "column," and Right L 2(M).
Example. Suppose M = M3(C). In this case the standard form may be taken
as
M3L2(M3hf3; L2(M3) ~ (M3, < .,. , where < x, y >= Tr(y*x).
Note that this norm, called the Hilbert-Schmidt norm, is just the e2 norm of the ma-
trix entries, and that the left and right multiplicative actions are commutants. (If
we had chosen a nontracial positive linear functional, we would have obtained an iso-
morphic bimodule with a "twisted" right action ... this is inchoate Tomita-Takesaki
theory.) The module R2(M3) is M 3xoo , again with the Hilbert-Schmidt norm,
and the commutant is Moo(M3) ~ Moo. According to Proposition 2.1, isomor-
phism classes of left M 3-modules should be parameterized by equivalence classes of
projections in Moo. These are indexed by their rank n E (1:+Uoo); the correspond-
ing isomorphism class of modules has representative M 3xn . In summary, we have
learned that any left representation of M3 on a Hilbert space is isomorphic to some
number of copies of C 3. The same argument shows that V(Moo(Mk)) ~ (1:+ U 00)
for any k.
Properties of the monoid V ( Moo (M)) determine the so-called type of the alge-
bra. For a factor (a von Neumann algebra whose center is just the scalars), there
are only three possibilities: (1:+ U 00), (lR.+ U 00), and {O,+oo}. These are called
types I, II, III, respectively; a fuller discussion is given in Section 7.

3. Algebraic approaches to RTPs


When R is a ring, the algebraic R-relative tensor product is the functor, co-
variant in both variables, which maps a right R-module A and left R-module B to
the vector space (A alg B)/N, where N is the subspace generated algebraically
by tensors of the form ar b - a rb. In functional analysis, where spaces are
usually normed and infinite-dimensional, one obvious amendment is to replace vec-
tor spaces with their closures. But in the context of Hilbert modules over a von
Neumann algebra M, this is still not enough. Surprisingly, a result of Falcone
([FJ, Theorem 3.8) shows that if the RTP L2(M) M L2(M) is the closure of a
continuous (meaning III(~ 7])11 < ClI~IIII7]11) nondegenerate image ofthe algebraic
M-relative tensor product, M must be atomic, Le. M ~ EBnB(f.>n). We will discuss
the analytic obstruction further in Section 5. For now, we take Falcone's theorem
as a directive: do not look for a map which is defined for every pair of vectors.
If we give up completely on a vector-level construction, we can at least make the
functorial
RELATIVE TENSOR PRODUCTS 279

DEFINITION 3.1 (Sa). Given a von Neumann algebra M, a relative tensor


product is a junctor, covariant in both variables,
(3.1) RightL 2(M) x LeJtL2(M) -> Hilbert: (n,.ft) f-+ n M.ft,
which satisfies
(3.2)
as bimodules.
Although at first glance this definition seems broad, in fact we see in the next
proposition that there is exactly one RTP functor (up to equivalence) for each al-
gebra. The reader is reminded that functoriality implies a mapping of intertwiner
spaces as well, so it is enough to specify the map on representatives of each isomor-
phism class. In particular we have the bimodule structure
':(j) .ll(n M .ft)C(M.ll)
PROPOSITION 3.2. Let n ~ P C 2(M) E Right L2 M od(M) and.ft ~ R 2(M)q E
Left L2Mod(M) Jor some projections p,q E Moo(M). Then
n M.ft ~ P L2(Moo(M))q
with natural action oj the commutants.
PROOF. By implementing an isomorphism, we may assume that the projections
are diagonal: p = LPi eii, q = Lqj ejj. Using (3.2) and functoriality, we have
the bimodule isomorphisms
n M.ft ~ (EBPi L2(M)) M (EBL2(M)qj)

~ E9Pi L2(M) M L2(M)qj ~ E9Pi L2(M)qj ~ p L2(Moo(M))q.


i,j i,j
D

Visually,

(3.3)

where of course the 2 sums of the norms of the entries in these matrices are finite.

After making the categorical definition above, Sauvageot immediately noted


that it gives us no way to perform computations. We will turn to his analytic
construction in Section 5; here we discuss an approach to bimodules and RTPs
due to Connes. In his terminology a bimodule is called a correspondence. (The
best references known to the author are [C2l and [Pl, but there was an earlier
unpublished manuscript which is truly the source of Connes fusion.)
Consider a correspondence MnN. Choosing a row representation R2(M)q for
n, we know that the full commutant of L(M) is isomorphic to R(qMoo(M)q).
This gives us a unital injective *-homomorphism p : N '---+ qMoo(M)q, and from
the map p one can reconstruct the original bimodule (up to isomorphism) as
M(R 2(M)q)p(N)'
280 DAVID SHERMAN

What if we had chosen a different row representation R 2 (M)q' and obtained


P' : N ---+ q'Moo(M)q'? By Proposition 2.1, the module isomorphism
(3.4) MR 2(M)q ~ MR 2(M)q'
is necessarily given by the right action of a partial isometry v between q and q' in
Moo(M). Then P and P' differ by an inner perturbation: p(x) = v*p'(x)v. We con-
clude that the class of M - N correspondences, modulo isomorphism, is equivalent
to the class of unital injective *-homomorphisms from N into an amplification of
M, modulo inner perturbation. (These last are called sectors in subfactor theory.)
The distinctions between bimodules, morphisms, and their appropriate equivalence
classes are frequently blurred in the literature; our convention here is to use the
term "correspondence" to mean a representative *-homomorphism for a bimodule.
Notice that a unital inclusion N eM gives the bimodule ML 2(M)N.
The RTP of correspondences is extremely simple.
PROPOSITION 3.3. Consider bimodules MYJN and N.ftP coming from correspon-
dences PI : N '----> qMoo(M)q and P2 : P '----> q'Moo(N)q'. The bimodule M(YJN.ft)P
is the correspondence PI 0 P2, where we amplify PI appropriately.
We pause to mention that it is also fruitful to realize correspondences in terms
of completely positive maps. We shall have nothing to say about this approach;
the reader is referred to [P] for basics or [A2] for a recent investigation.

4. Applications to Morita equivalence and index


An invertible *-functor from Left L2 M od(N) to Left L2 M od(M) is called a
Morita equivalence [R]. Here a *-functor is a functor which commutes with the
adjoint operation at the level of morphisms. One way to create *-functors is the
following: to the bimodule MYJN, we associate
(4.1) FSj: Left L2 Mod(N) ---+ Left L2 Mod(M); N.ft 1---+ (MYJN) N (N.ft).
The next theorem is fundamental.
THEOREM 4.1. When L(M) and R(N) are commutants on YJ, the RTP functor
FSj is a Morita equivalence. Moreover, every Morita equivalence is equivalent to an
RTP functor.
This type of result - the second statement is an operator algebraic analogue of
the Eilenberg-Watts theorem - goes back to several sources, primarily the funda-
mental paper of Rieffel [R]. His investigation was more general and algebraic, and
his bimodules were not Hilbert spaces but rigged self-dual Hilbert C*-modules, fol-
lowing Paschke [Pal. From a correspondence point of view, rigged self-dual Hilbert
C*-modules and Hilbert space bimodules give the same theory; the equivalence is
discussed nicely in [A1]. (And the former is nothing but an L oo version of the
latter, as explained in [JS].) Our Hilbert space approach here is parallel to that of
Sauvageot [Sa], though modeled more on [R], and is streamlined by our standing
assumption of separable preduals.
We will need
DEFINITION 4.2. The contragredient of the bimodule MYJN is the bimodule
NfJM' where fJ is conjugate linearly isomorphic to fj (the image of ~ is written (J,
and the actions are defined by n~m = m*~n*.
RELATIVE TENSOR PRODUCTS 281

LEMMA 4.3. Suppose L(M) and R(N) are commutants on fl. Then
NfJM M MflN ~ NL2(N)N.
PROOF. If Mfl ~ MR 2(M)q, then N ~ qMoo(M)q by (2.3), and fJM
qC 2 (M)M. By Proposition 3.2 and the comment preceding Proposition 2.1,
NfJM M MflN ~ N(qL2(Moo(M))q)N ~ NL2(qMoo(M)q)N ~ NL2(N)N.
D
Lemma 4.3 was first proven by Sauvageot (in another way). In our situation it
means

FfJ 0 FSj(Nfi) ~ L2(N) N Nfi ~ Nfi.


(Here we have used the associativity of the RTP, which is most easily seen from
the explicit construction in Section 5.) We conclude that Fi) 0 FSj is equivalent to
the identity functor on Left L2Mod(N), and by a symmetric argument FSj 0 Fi)
is equivalent to the identity functor on Left L2Mod(M). Thus FSj is a Morita
equivalence, and the first implication of Theorem 4.1 is proved.
Now let F be a Morita equivalence from Left L2 M od(N) to Left L2 M od(M).
Then F must take NR2(N) to a module isomorphic to MR2(M), because each is
in the unique isomorphism class which absorbs all other modules. (This is meant
in the sense that NR2(N) ffiNfl ~ NR2(N); R2(N) is the "infinite element" in the
monoid V (Moo (N)).) Being an invertible *-functor, F implements a *-isomorphism
of commutants - call it a, not F, to ease the notation:

(4.2)
Apparently we have

(4.3)
Before continuing the argument, we need an observation: isomorphic algebras
have isomorphic standard forms. Specifically, we may write L2(Moo(N)) as the
GNS construction for tp E Moo(N)t and obtain the isomorphism
(a-1)t : L2(Moo(N)) ..::. L2(Moo(M)), (a-1)t : xtp1/2 ........ a(x)(tp 0 a- 1)1/2.
Note that (a- 1 )t(x~y) = a(x)[(a- 1 )t(~)la(y).
Now consider the RTP functor for the bimodule

MflN = u- 1 (M)a- 1(et'{)C 2(N)N.


By Proposition 3.2 and the comment preceding Proposition 2.1, its action is

R2(N)q ........ 17-1 (M)a- 1 (et'{)L2(Moo (N))q (17;:)' Met'{ L 2(Moo (M))a(q)
~ MR 2(M)a(q) ~ F(R 2(N)q).
We conclude that F is equivalent to FSj, which finishes the proof of Theorem 4.1.
Notice that (4.2) and (4.3) can also be used to define a functor; this gives us
COROLLARY 4.4. For two von Neumann algebras M and N, the following are
equivalent:
(1) M and N are Morita equivalent;
282 DAVID SHERMAN

(2) Moo(N) ~ Moo(M);


(3) there is a bimodule Mf)N where the actions are commutants of each other;
(4) there is a projection q E Moo(M) with central support 1 such that
qMoo(M)q ~ N.
(The central support of x E M is the least projection z in the center of M
satisfying x = zx.)
Example continued. M3 and M5 are Morita equivalent. This can be deduced
easily from condition (2) or (4) of the corollary above. It also follows from the
(Hilbert) equivalence bimodule M3HS(M3X5)Ms, where "HS" indicates the Hilbert-
Schmidt norm; this bimodule gives us an RTP functor which is a Morita equivalence.
Regardless of the construction, the equivalence will send (functorially) n copies of
C 5 to n copies of C 3 . Apparently Morita equivalence is a coarse relation on type I
algebras - it only classifies the center of the algebra (up to isomorphism). At the
other extreme, Morita equivalence for type III algebras is the same as algebraic
isomorphism; Morita equivalence for type II algebras is somewhere in the middle
([RJ, Sec. 8).

For a bimodule Mf)N where the algebras are not necessarily commutants, the
functor (4.1) still makes sense. To get a more tractable object, we may consider
the domain and range to be isomorphism classes of modules:

(4.4) 71'1) : V(Moo(N)) ---. V(Moo(M));


F1)(R 2(N)q) = Mf)N N R2(N)q ~ R2(M)7I'1) ([q]).
We call this the bimodule morphism associated to f), a sort of "skeleton" for
the correspondence. It follows from Proposition 3.3 that if the bimodule is p : N '---+
qMoo(M)q, then 71'1) is nothing but poo, the amplification of p to Moo(N), restricted
to equivalence classes of projections.
This has an important application to inclusions of algebras. We have seen in
Section 3 that a unital inclusion N t.
M is equivalent to a bimodule ML2(M)N.
When the correspondence p is surjective, the module generates a Morita equivalence
via its RTP functor, and the induced bimodule morphism is an isomorphism of
monoids. When N i= M, it is natural to expect that the bimodule morphism gives
us a way to measure the relative size, or index, of N in M. (For readers unfamiliar
with this concept, the index of an inclusion N c M is denoted [M : N] and is
analogous to the index of a subgroup. The kernel of this idea goes back to Murray
and von Neumann, but the startling results of Jones [J] in the early 1980's touched
off a new wave of investigation. We recommend the exposition [K] as a nice starting
point.)
For algebras of type I or II, the index can be calculated in terms of bimodule
morphisms. (There are also broader definitions of index which require a conditional
expectation (=norm-decreasing projection) from M ontoN.) This amounts largely
to rephrasing and extension of the paper [Jol] , and we do not give details here. Very
briefly, let 71' : V(Moo(M)) ---. V(Moo(M)) be the bimodule morphism for

(4.5)
RELATIVE TENSOR PRODUCTS 283

When M is a factor, 11" is a monoidal morphism on the extended nonnegative


integers (type I) or extended nonnegative reals (type II). It must be multiplication
by a scalar, and this scalar is the index. If M is not a factor, the index is the spectral
radius of 11", provided that V(Moo(M)) is endowed with some extra structure (at
present it is not even a vector space).
Example. Consider the correspondence

The image of M6L2(M6) under the RTP functor for (4.5) is

~ M6 L2 (M6)M3 M3 M3 L2 (M6)
(now counting the dimensions of the Hilbert spaces)

~ M6HS(M12X3)M3 M3 M3 HS (M3xd ~ M6HS(M12X12) ~ M6HS(M6X24).


We have gone from 6 copies of (:6 to 24 copies; that is,

61-+24.
Apparently the index is 4, which is also the ratio of the dimensions of the algebras.

5. Analytic approaches to RTPs


As mentioned in Section 3, we cannot expect the expression ~ M T/ to make
sense for every pair of vectors ~,T/. In essence, the problem is that the product
of two L2 vectors is Ll, and an Ll space does not lie inside its corresponding
L2 space unless the underlying measure is atomic. Densities add, even in the
noncommutative setting, and so the product in (3.3) "should" be an Ll matrix. To
make this work at the vector level, we need to decrease the density by 1/2 without
affecting the "outside" action of the commutants ... and the solution by Connes and
Sauvageot ([Sa], [C2]) is almost obvious: choose a faithful cp E Mt and put cp-l/2
in the middle of the product. That is,

(5.1)
This requires some explanation.
Fix faithful cp E Mt and row and column representations of 5) and .It as in
(2.1). We define

D(S;, 1") ~ { (::~:;:) E S; , ~>~Xn ex;"" in M}


V(5), cp) is dense in 5), and its elements are called cp-left bounded vectors [C1].
Now by (5.1) we mean the following: for ~ E V(5), cp), we simply erase the
symbol cpl/2 from the right of each entry, then carry out the multiplication. The
natural domain is V(5), cp) x .It. Visually, .
284 DAVID SHERMAN

(5.2) ( ( ::::;:) ,(" "' '" )) ~ (::::;:) (,-"') (" "' '" ) ~ (X;",
For cp =I- 1jJ E Mt, we cannot expect cp TJ = ,p TJ even if both are defined,
although the reader familiar with modular theory will see that

(5.3) cp TJ = (cp-l/2)TJ = (cp-l/21jJl/21jJ-l/2)TJ = (Dcp : D1jJ)i/2 ,p TJ.


(An interpretation of the symbols cpl/2, cp-l/2 as unbounded operators will be dis-
cussed in the next section.)
Now we define S) cp .!'t to be the closed linear span of the vectors cp TJ inside
L2(Moo(M)). Up to isomorphism, this is independent of cp. (We know this because
of functoriality; the "change of weight" isomorphism is densely defined by (5.3).)
The given definition for V(S), cp) c S) makes it seem dependent on the choice
of column representation. That this is not so can be seen by noting (as in (3.4))
that the intertwining isomorphism is given by L( v) for some partial isometry v E
Moo(M). But let us also mention a method of defining the same RTP construction
without representing S) and.!'t. First notice that V(S), cp) can also be defined as the
set of vectors for which
7rf() : L 2(M)M -+ S)M. cpl/2x 1-+ x,
is bounded. (A more suggestive (and rigorous) notation would be L(cp-l/2).) Now
we consider an inner product on the algebraic tensor product V(S), cp) .!'t, defined
on simple tensors by

(5.4)
The important point here is that 7rt(2)*7rf(6) E .c(L2(M)M) = M. The closure
of V(S), cp) .!'t in this inner product, modulo the null space, is once again S) cp .!'t.
(If we do choose a row representation as in (5.2), we have

The paper [F] contains more exposition of this approach, including some alter-
nate constructions.

6. Realization of the relative tensor product as composition of


unbounded operators
In this section we briefly indicate how (5.1) can be rigorously justified and
extended. Readers are referred to the sources for all details.
In his pioneering theory of noncommutative LP spaces, Haagerup [H] estab-.
lished a linear isomorphism between Mt and a class of positive unbounded oper-
ators affiliated with the core of M. (The core, well-defined up to isomorphism, is
the crossed product of M with one of its modular automorphism groups.) We will
denote the operator corresponding to the positive functional cp by cp also. These
operators are r-measurable (see the next section), where r is the canonical trace
RELATIVE TENSOR PRODUCTS 285

on the core, and so they generate a certain graded *-algebra: positive elements of
LP(M) are defined to be operators of the form '{)1/ p . The basic development of this
theory can be found in [Tej; our choice of notation is influenced by [Yj, where it is
called a modular algebra.
The composition of two L2 operators is an Ll operator, and it turns out that
(5.1) can be rigorously justified [Sj as an operator equation. (This is not automatic,
as operators like ,{)-1/2 are not 7-measurable and require more delicate arguments.)
In fact, there is nothing sacred about half-densities. With the recent development
of noncommutative LP modules [JS], one can allow relative tensor products to be
bifunctors on Right U(M) x Left Lq(M), with range in a certain L'" space. The
mapping is densely-defined by

~ cp 1/ ~ (~,{)~-*-~)1/.
In the case of an RTP of L oo modules (or more generally, Hilbert C*-modules),
the middle term is trivial and there is no change of density. This explains why there
are no domain issues in defining a vector-valued RTP of Hilbert C*-modules [Rj.
Let us mention that the recent theory of operator bimodules, in which vectors
can be realized as bounded operators, allows a variety of relative tensor products
over C*-algebras [APj. This can be naturally viewed as a generalization of the theory
of Banach space tensor products, which corresponds to a C*-algebra of scalars.

7. Preclosedness of the map (~, 1/) 1-+ ~ cp 1/


Our purpose in this final section is to study when the relative tensor map is
preclosed. This is a weaker condition than that of Falcone, who studied (effectively)
when the map was bounded. We begin with a base case: a fa.ctor, two standard
modules, and a simple product. With the usual notation Bcp for V(L 2 (M), '{)), the
relevant map is

Bcp x L2(M) :3 (~, 1/) 1-+ ~ cp 1/ E L 2(M).


This is bilinear: we take "preclosed" to mean that if ~n -+ ~ E Bcp, 1/n -+ 1/,
~n cp 1/n -+ (, then necessarily ( = ~ cp 1/. We will also consider several variations:
changing the domain to an algebraic tensor product, allowing non-factors, and
allowing arbitrary modules.
Readers unfamiliar with von Neumann algebras will find this section more
technical, and any background we can offer here is sure to be insufficient. Still, we
introduce the necessary concepts in hopes that the non-expert will at least find the
statements of the theorems accessible.

A weight is an "unbounded positive linear functional": a linear map from M+


to [0, +ooj. We will always assume that weights are normal, so Xc< / x strongly
=> '{)(xc<) / '{)(x); and semifinite, so {x E M+ I '{)(x) < oo} is a-weakly dense
in M+. We can still define RTPs for faithful weights, but now Bcp = {x'{)I/2 I
'{)(x*x) < oo} C L 2(M). For details of the representations associated to weights,
see [T2j.
A weight 7 which satisfies 7(XY) = 7(YX) on its domain of definition will be
called a trace (more properly called a "tracial weight"). An algebra which admits
a faithful trace 7 is semifinite; if in addition we can have 7(1) < 00, it is finite.
This facilitates the following classification of factors: a factor with n orthogonal
286 DAVID SHERMAN

minimal projections is type In (possibly n = 00); a semifinite factor without minimal


projections is type III if finite and type 1100 if not; a factor which is not semifinite
is type III. The reader will note that this refines our previous definitions of type,
as a trace is exactly the object which orders the equivalence classes of projections.
Obviously, there is much more to be said, and most of it can be found in [Tl].
For a faithful trace T on semifinite M, it is useful to consider the T-measure