Lindenstrauss, Tzafriri - Classical Banach Spaces I - Sequence Spaces - (1977)

Ergebnisse der Mathematik
und ihrer Grenzgebiete 92

A Series of Modern Surveys in Mathematics
Editorial Board: P. R. Halmos P. J. Hilton (Chair-

man) R. Remmert B. Sz6kefalvi-Nagy
Advisors: L. V. Ahlfors R. Baer F. L. Bauer
A. Dold J. L. Doob S. Eilenberg
K. W. Gruenberg M. Kneser G. H. MUller
M. M. Postnikov B. Segre E. Sperner
Joram Lindenstrauss Lior Tzafriri
Classical
Banach Spaces I
Sequence Spaces
Springer-Verlag
Berlin Heidelberg New York 1977
Joram Lindenstrauss
Lior Tzafriri
Department of Mathematics, The Hebrew University of Jerusalem
Jerusalem, Israel
AMS Subject Classification (1970): 46-02, 46 A4S, 46Bxx, 46Jxx
ISBN-13: 978-3-642-66559-2 e-ISBN-13: 978-3-642-66557-8

DOl: 10.1 007/978-3-642-66557-8
Library of Congress Cataloging in Publication Data. Lindenstrauss, 10ram, 1936-.

Classical Banach spaces. (Ergebnisse der Mathematik und ihrer Grenzgebiete; 92).
Bibliography: v. HI, p. Includes index. CONTENTS: I. Sequence spaces. I. Banach spaces.
2. Sequence spaces. 3. Function spaces. I. Tzafriri, Lior, 1936- joint author. II. Title. III.
Series QA322.2.L56. 1977- 515'.73. 77-23131.
This work is subject to copyright. All rights are reserved, whether the whole or part of the
material is concerned, specifically those of translation, reprinting, re-use of illustrations,
broadcasting, reproduction by photocopying machine or similar means, and storage in
data banks. Under § 54 of the German Copyright Law where copies are made for other
than private use, a fee is payable to the publisher, the amount of the fcc to be determined
by agreement with the publisher.
© by Springer-Verlag Berlin Heidelberg 1977.

Softcover reprind of the hardcover 1st edition 1977
Typesetting: William Clowes & Sons Ltd., London, Beccles and Colchester.
2141/3140-543210
To Naomi and Marianne
Preface
The appearance of Banach's book [8] in 1932 signified the beginning of a syste-
matic study of normed linear spaces, which have been the subject of continuous
research ever since.
In the sixties, and especially in the last decade, the research activity in this area
grew considerably. As a result, Ban:ach space theory gained very much in depth as
well as in scope: Most of its well known classical problems were solved, many
interesting new directions were developed, and deep connections between Banach
space theory and other areas of mathematics were established.
The purpose of this book is to present the main results and current research
directions in the geometry of Banach spaces, with an emphasis on the study of the
structure of the classical Banach spaces, that is C(K) and Lip.) and related spaces.
We did not attempt to write a comprehensive survey of Banach space theory, or
even only of the theory of classical Banach spaces, since the amount of interesting
results on the subject makes such a survey practically impossible.
A part of the subject matter of this book appeared in outline in our lecture notes
[96]. In contrast to those notes, most of the results presented here are given with
complete proofs. We therefore hope that it will be possible to use the present book
both as a text book on Banach space theory and as a reference book for research
workers in the area. It contains much material which was not discussed in [96], a
large part of which being the result of very recent research work. An indication to
the rapid recent progress in Banach space theory is the fact that most of the many
problems stated in [96] have been solved by now.
In the present volume we also state some open problems. It is reasonable to
expect that many of these will be solved in the not too far future. We feel, however,
that most of the topics discussed here have reached a relatively final form, and
that their presentation will not be radically affected by the solution of the open
problems. Among the topics discussed in detail in this volume, the one which seems
to us to be the least well understood and which might change the most in the
future, is that of the approximation property.
We divided our book into four volumes. The present volume deals with
sequence spaces. The notion of a Schauder basis plays a central role here. The
classical spaces which are in the most natural way sequence spaces are Co and i p ,
J ~p ~ 00. Volumes II and III will deal with function spaces. In Volume II we shall
present the general theory of Banach lattices with an emphasis on those notions
concerning lattices which are related to Lip.) spaces. Volume III will be devoted
to a study of the structure of the spaces Lp(O, 1), C(K) and general preduals of
viii Preface
L 1 (/1,) spaces. The division of the common Banach spaces into sequence and function
spaces is made according to the usual practice. It should be remembered, however,
that several spaces have natural representations both as sequence and function spaces.
The best known example is the separable Hilbert space, which can be represented
both as the sequence space 12 and as the function spaceL2 (O, 1). A less trivial example
is the space Ip, l,,;,p";' 00, which is isomorphic to the function space HiD) of the
analytic functions on the disc D={z; Izl<l} with Ilfll=(SJlf(z)IP dx dy)1!P<oo
(cf. [88]). Also, the spaces C(O, 1) and Lp(O, 1), 1 ,,;,p < 00, have Schauder bases, and
thus it is convenient sometimes to use their representations as sequence spaces.
In Volume IV we intend to present the local theory of Banach spaces. This theory
deals with the structure of finite-dimensional Banach spaces and the relation
between an infinite-dimensional Banach space and its finite-dimensional subspaces.
A central part in this approach to Banach space theory is played by the evaluation
of various parameters of finite-dimensional Banach spaces. The role of the classical
finite-dimensional spaces, that is of the spaces I;, l,,;,p";' 00, n = 1, 2, ... in the local
theory of Banach spaces is even more central than the role of the classical spaces
in the gerieral theory of Banach sequence spaces and function spaces.
We sketch now briefly the contents of this volume. Chapter 1 contains a quite
complete account of the main results on Schauder bases in general Banach s')Jaces.
Several notions related to Schauder bases-the various approximation properties,
general biorthogonal systems and Schauder decompositions-as well as some
examples are discussed in detail.
Chapter 2 is devoted to a study of the spaces Co and Ip, l,,;,p < 00, and to some
extent also of I",. Section a is devoted to an examination of the basic properties of
these spaces, some of which are shown to characterize these spaces among general
Banach spaces. The other sections of Chapter 2 are basically independent of each
other and can thus be read in any order. In Sections band C we discuss certain ideals
of operators on general Banach spaces and show how they can be used in the study
of the structure of the classical sequence spaces. Section d contains a structure
theorem for "nice" subspaces of Co and Ip as well as examples of subspaces which
are not "nice" (i.e. subspaces which fail to have the approximation property).
This section contains also a discussion of general results related to the approxima-
tion property which complement the treatment of this property in Section e of
Chapter 1. Section f contains an example of an infinite-dimensional Banach space
which fails to have any of the classical sequence spaces as a subspace and also
criteria for general Banach spaces to have subs paces isomorphic to Co and especially
to 11 , The final section of Chapter 2 deals with the extension properties of Co and I""
the lifting property of 11 , and the closely related topic of the automorphisms of
these spaces.
In Chapter 3 we discuss the special properties of symmetric bases and the
relation between symmetric bases and general unconditional bases. A large part
of this chapter is devoted to results and examples related to the possible charac-
terizations of Co and Ip , 1 ,,;,p < 00, in the class of all spaces with a symmetric basis.
The final chapter of this volume is devoted to a detailed study of the structure of
some particular classes of spaces with symmetric bases, mainly Orlicz sequence
spaces. The main emphasis is again on the relation between these spaces and the
Preface ix
spaces Co and Ip. Several examples given there demonstrate how much more
complicated the structure of general Orlicz sequence spaces is, as compared to that
of Ip spaces. In section d it is shown that Orlicz sequence spaces enter naturally into
the study of spaces like Ip EB IT with p-/=r. In Vol. III it will be shown that Orlicz
sequence spaces arise naturally in the study of the structure of subspaces of
Ll(O, 1).
We assume that the reader is familiar with the basic results of real analysis and
functional analysis which are usually covered in first year graduate courses in these
subjects. An acquaintance with the main results in chapters I-VI of [33] will
certainly suffice (much less is actually needed for being able to read this book).
The bibliography contains only those papers which are actually quoted in the
text. We tried to indicate in the text the source of the main results which we
present. The reference list is, however, far from being complete. Reference to papers
where the basic results in Banach space theory were first proved can be found, for
example, in [28] and [33]. Further references on bases may be found in [135].
References to further literature on Orlicz spaces may be found in [75].
The overlap between this book and existing books on related topics is very small.
We would like to acknowledge the contribution of G. Schechtman, who read
the entire manuscript and made several very valuable suggestions, and of J. Arazy,
who was very helpful in proofreading this volume and eliminated several mistakes.
We also wish to thank Z. Altshuler and Y. Sternfeld for their help. We are grateful
to Danit Sharon who expertly carried out the task of typing the manuscript of this
book. We are also indebted to the U.S.-Israel Binational Science Foundation for
partial support.
Finally, we would like to thank Springer-Verlag and especially Roberto Minio
for their cooperation and help in all the stages of the preparation of this book.
Jerusalem Joram Lindenstrauss

January 1977 Lior Tzafriri
Table of Contents
1. Schauder Bases .
a. Existence of Bases and Examples
b. Schauder Bases and Duality 7
c. Unconditional Bases. 15
d. Examples of Spaces Without an Unconditional Basis 24
e. The Approximation Property 29
f. Biorthogonal Systems 42
g. Schauder Decompositions 47
2. The Spaces Co and Ip 53

a. Projections in Co and Ip and Characterizations of these Spaces 53
b. Absolutely Summing Operators and Uniqueness of Unconditional
Bases . 63
c. Fredholm Operators, Strictly Singular Operators and Complemented
Subspaces of Ip EB IT . 75
d. Subspaces of Co and Ip and the Approximation Property, Complement-
ably Universal Spaces 84
e. Banach Spaces Containing Ip or co. 95
f. Extension and Lifting Properties, Automorphisms of leo, Co and 11 104
3. Symmetric Bases. 113

a. Properties of Symmetric Bases, Examples and Special Block Bases 113
b. Subspaces of Spaces with a Symmetric Basis. 123
4. Orlicz Sequence Spaces . 137

a. Subspaces of Orlicz Sequence Spaces which have a Symmetric Basis 137
b. Duality and Complemented Subspaces 147
c. Examples of Orlicz Sequence Spaces . 156
d. Modular Sequence Spaces and Subspaces of lp EB IT • 166
e. Lorentz Sequence Spaces 175
References 180
Subject Index 185

Standard Definitions, Notations and
Conventions
For most of the results presented in this book it does not matter whether the field
of scalars is real or complex. In the isometric theory there are some differences
(usually minor) between real and complex spaces. As a rule we shall work with real
scalars and, in a few places, we shall indicate the changes needed in the complex
case. In a few instances e.g. where spaces of analytic functions are involved or
where spectral theory is used we shall use complex scalars.
By L 1'{JL)=L;(!2, E, JL), 1 ~p~oo we denote the Banach space of equivalence
classes of measurable functions on (D, E, JL) whose p'th power is integrable
(respectively, which are essentially bounded if p=oo). The norm in L1'{JL) is defined
by 11/11 = (f1/{w)jP dJL{w»l/1' (ess sup I/{w)1 ifp = 00). If{D,E, JL) is the usual Lebesgue
measure space on [0, I] we denote L1'{JL) by L1'(O, 1). If (r, E, JL) is the discrete
measure space on a set r, with JL{{Y}) = 1 for every Y e r, we denote Lp(JL) by
11'{T). If r is the set of positive integers we denote 11'{r) also by 11' while if r=
{I, 2, ... , n}, for some n<oo, we denote Ip(r) by I;. The subspace of I,.,{T), of those
functions which vanish at 00, is denoted by co(T) (if r is the set of positive integers
we denote this space by co). The subspace of I,., consisting of convergent sequences
is denoted by c. For a compact Hausdorff space K we denote by C{K) the Banach
space of all continuous scalar-valued functions on K with the supremum norm.
If K is the unit interval [0, 1] in its usual topology we denote C(K) by C(O, 1).
In a Banach space X we denote the ball with center x and radius r, i.e. {y;
lIy-xll ~r}, by Bx(x, r). If the space X is clear from the context, we simply write
B{x, r). The unit ball Bx{O, 1) of X is denoted also by Bx. For a sequence {Xn}:'=l
of elements of X we denote by span {X n}:'=l the algebraic linear span of {Xn}:'=l
i.e. the set of all finite linear combinations of {Xn}:'=l' The closure of span {Xn}:'=l
is denoted by [xn ]:'=l' A similar notation is used for the span of a set other than a
sequence. For a set A c X its norm closure is denoted by A, e.g. [Xn]:'=l=span
{Xn}:'=l' The convex hull of a sequence {xn}:'=l'is denoted by conv{xn}:'=l; the
closed convex hull by conv {Xn}:'=l'
The term "operator" means a bounded linear operator unless specified other-
wise. The space of all operators from X to Y with the usual operator norm is
denoted by L(X, Y). An operator T e L{X, Y) is called compact if T Bx is a norm
compact subset of Y. The identity operator of a Banach space X is denoted by Ix
(or simply by I if X is clear from the context). For an operator TeL{X, Y) the
notation T 1z denotes the restriction of T to the subspace Z of X.
Two Banach spaces X and Yare called isomorphic (denoted by X ~ Y) if there
exists an invertible operator from X onto Y. The Banach-Mazur distance coeffiCient
xu Standard Definitions, Notations and Conventions
d(X, Y) is defined by inf IITIIIIT-III, the infimum being taken over all invertible
operators from X onto Y (if X is not isomorphic to Y we put d( X, Y) = 00). Notice
that d(X, Y):;:::; I, for every X and Y, and that d(X, Y) d( Y, Z):;:::; d(X, Z), for every
X, Y and Z. If there exists an invertible operator T from X onto Y so that IITII =
liT-III = I (i.e. IITxl1 = IIxll, for every x E X) we say that X is isometric to Y. In
this case d(X, Y)= I (the converse is false in general; it is possible that d(X, Y)= I
but that the infimum in the definition of d(X, Y) is not attained i.e. X is not
isometric to Y). An operator T E L(X, Y) is said to be an isomorphism into Y if
°
there is some constant C> so that IITxll :;:::; C/ Ixll for every x E X. In this case T-1
is a well defined element in L(TX, X).
A closed linear subspace Y of a Banach space X is said to be a complemented
subspace of X if there is a bounded linear projection from X onto Y, or what is the
same, if there exists a closed linear subspace Z of X so that X is the direct sum of Y
and Z, i.e. X = Y ED Z. We shall also use some direct sums of infinite sequences of
Banach spaces. If {Xn}:; 1 is a sequence of Banach spaces we define the direct sum
of these spaces in the sense of lp, I <p < 00, namely C~lED Xn) p, as the space of all
sequences x=(x1 , X2, ... ), with Xn E Xn for all n, for which Ilxll = C~lllxnW)lIP<
00. Similarly, C~/B Xn)o denotes the direct su~ of {Xn}:;l in the sense of Co
i.e. the space of all sequences x=(x 1 , x 2 , ••• ), with Xn E Xn for all n, for which
lim IIxnll =0. The norm in this direct sum is taken as IIxll = max IIxnll. We shall
n n
occasionally use also other types of infinite direct sums. These will be defined in the
proper places in the text.
Besides the norm (or strong) topology of a Banach space X we often use some
other topologies. If Y is a subspace of the dual X* of X then the Y-topology of X
is the weakest topology making all the elements of Y continuous. A basis for the Y
°
topology is obtained by taking all the sets of the form Vex, e, A)={u; Ix*(u)-
x*(x)1 < e, x* E A}, where x E X, e > and A is a finite subset of ;Y. If Y = X* the Y
topology is called the weak topology (w topology). If X=Z* and we take as Ythe
canonical image of Z in Z** = X* we obtain the w* topology induced by Z (if Z
is clear from the context we simply talk of the w* topology). Convergence of
sequences in the w topology (resp. w* topology) is denoted by Xn ~ x or w lim Xn =
n
X (resp. Xn ~x or w* lim Xn = x). An operator T E L(X, Y) is said to be w compact
n
if T Bx is a compact set in Y, in its w topology (i.e. a w compact set in Y).
Whenever we consider a Banach space X as a subspace of its second dual X* *
we assume that it is embedded canonically. For a subset A c: X we denote by A l.
the subspace {x*; x*(x) = 0, x E A} of X*. For a subset A c: X* we denote by AT
the subspace {x; x*(x) =0, x* E A} of X. For every subset A c: Xwe have Al.T ::::> A
and equality holds if and only if A is a closed linear subspace.
Besides subspaces of Banach spaces we shall also study quotient spaces. An
operator T: X -+ Y is called a quotient map if T Bx=By. A Banach space Y is
isomorphic to a quotient space of a space X if and only if there exists an operator
T from X onto Y. If such a T exists then Y~X/ker T, where ker T={x; Tx=O},
Standard Definitions, Notations and Conventions xiii
and y* is isomorphic to the subspace (ker T).L of X*. Similarly, if Z is a subspace

of X then Z* is isometric to the quotient space X/(Z.L).
Among the general notations used in this book we want to single out the
following. For a positive number S we denote by [S] the largest integer:::;; S. For
a set A we denote by A the cardinality of A. If A and B are sets we put A,..., B=
{x, X E A, x ¢ B}.
1. Schauder Bases
a. Existence of Bases and Examples
The aim of this volume is to describe some results concerning sequence spaces, i.e.
those Banach spaces which can be presented in some natural manner as spaces of
sequences. In general, such a repre~entation is achieved by introducing in the space
a sort of "coordinate system". There are, obviously, many different ways of giving
a precise meaning to the terms "Banach sequence spaces" and "coordinate systems".
The best known and most useful approach is by using the notion of a Schauder
basis.
Definition l.a.l. A sequence {Xn};:'~l in a Banach space X is called a Schauder

basis of X if for every x E X there is a unique sequence of scalars {an};:'~ 1 so that
OJ
X= L anxn. A sequence {Xn};:'=l which is a Schauder basis of its closed linear span
n=l
is called a basic sequence.
In this book we shall not consider any type of bases in infinite-dimensional

Banach spaces besides Schauder bases. We shall therefore often omit the word
Schauder. In addition to Schauder bases we shall only encounter algebraic bases
in finite-dimensional spaces. This should not cause any confusion. As a matter of
fact, quantitative notions concerning Schauder bases (like the basis constant defined
below) have a meaning and will be used also in the context of algebraic bases in
finite dimensional spaces.
Evidently, a space X with a Schauder basis {Xn};:'=l can be considered as a
00
sequence space by identifying each x = L

anxn with the unique sequence of
n=l
coefficients (aI' a2, a3'''')' It is important to note that for describing a Schauder
basis one has to define the basis vectors not only as a set but as an ordered sequence.
Let (X, II II) be a Banach space with a basis {Xn};:'=l' For every X= ~ anxn in
n=l
X the expression 1/ Ixlll ~ s~p Ili~l a/x/II is finite. Evidently, 111·111 is a norm on X
and Ilxll ~ Illxlll for every x E X. A simple argument shows that X is complete also
with respect to 111·111 and thus, by the open mapping theorem, the norms 11·11 and
111·111 are equivalent. These remarks prove the following proposition [8].
Proposition l.a.2. Let X be a Banach space with a Schauder basis {Xn};:'=l' Then the
2 1. Schauder Bases
projections P n: X -+ X, defined by P n (~ a1xi ) =

1=1
i
j=1
tljXh are bounded linear operators
and sup
n
IIPnl1 <00.
The projections {Pn}~= 1 are called the natural projections associated to

{Xn}~=1; the number sup IIPnl1 is called the basis constant of {Xn}~=1' A basis whose
n
basis constant is 1 is called a monotone basis. In other words, a basis is monotone if,
for every choice of scalars {a"}~=h the sequence of numbers {IIJ1 alxIII}~=l is non-
decreasing. Every Schauder basis {Xn}~=1 is monotone with respect to the norm
IlIxlli = sup IIP"xll which was already used above. Indeed,
n
Thus, given any Schauder basis {X"}~=1 of X, we can pass to an equivalent norm
in X for which the given basis is monotone.
There is a simple and useful criterion for checking whether a given sequence is
a Schauder basis.
Proposition 1.a.3. Let {X,,}:'=1 be a sequence of vectors in X. Then {Xn}~=1 is a

Schauder basis of X if and only if the following three conditions hold.
(i) xn:f. 0 for all n.
(ii) There is a constant K so that, for every choice of scalars {al}~ 1 and integers
n<m, we have
(iii) The closed linear span of {X"}~=1 is all of X.
The proof is easy. The necessity of (i) and (iii) is clear from the definition, while that
c:o
of (ii) follows from l.a.2. Conversely, if (i) and (ii) hold then L: anxn=O implies
,,=1
that a,,=O for all n. This proves the uniqueness of the expansion in terms of
{X"}~=1' In order to prove that every x E X has such an expansion it is enough, in
c:o
view of (iii), to show that the space of all elements of the form L: anx" is a closed
n=1
linear space. This latter fact can be easily proved by using (ii). 0
Obviously conditions (i) and (ii) of 1.a.3, by themselves, form a necessary and
sufficient condition for a sequence {X"}~=1 to be a basic sequence. It is also worth-
while to observe that in case we can take K= 1 it is enough to verify (ii) for m=n+ 1.
A basis {Xn}~=1 is called normalized if Ilxnll = 1 for all n. Clearly, whenever
{Xn}~=1 is a Schauder basis of X, the sequence {xil/llxnJJ}~=1 is a normalized basis
in X.
a. Existence of Bases and Examples 3
Before proceeding with the general discussion we present some examples of

" ... ) form a monotone and normalized
bases. The unit vectors en = (0, 0, 0, ... ,1,0,
basis in each of the spaces Co and lp, l::;;,;p <00. An example of a basis in the space c,
of convergent sequences of scalars, is given by
Xl=(l, 1, 1, ... ) and, for n> 1, x n=en-l'
The expansion of x=(al> a2,"') E c with respect to this basis is
x=(lim an)xl + (aI-lim an)x2 + (a2- lim a,,)xs+'" .

" n "
An important example of a Schauder basis is the Haar system in Lp(O, 1),

l::;;,;p<oo.
Definition 1.a.4. The sequence of functions {Xn(t)};-'=l defined by Xl(t) = 1 and, for
k=O, 1, 2, ... , 1= 1, 2, ... , 2",
I if t E [(2/-2)2- k -\ (21-1)2-k-l]
{
°
X2"+Z(t)= -1 if t E «2/-1)2- k -\ 2/·2- k - 1 ]
otherwise
is called the Haar system.
The Haar system is (in its given order) a monotone (but obviously not normal-
ized) basis of Lp(O, 1) for every l::;;,;p<oo. Indeed, since the linear span of the
Haar system contains all the characteristic functions of dyadic intervals (i.e.
intervals of the form [/·2-", (/+ 1)·2- k ]), it is clear that (iii) of 1.a.3 holds. We have
only to verify that (ii) holds with K= 1. Let {at}:; 1 be any sequence of scalars, let
n ,,+1
n be an integer and letf(t)= L: a.Xt(t)
t=l
and g(t) = L:
t=l
OtXt(t). The only difference
betweenfand g is that on some dyadic interval Iwherefhas the constant value b,
say, g has the value b+a"+l on the first half of I and b-an+l on the second half.
Since, for every p~ 1, Ib+a"+lI P+ Ib-a"+1IP~2IbIP we get that IIfll::;;,; IIgli.
By integrating the Haar system or more precisely by putting
t
fPn(t)= fo Xn-l(U) du, n> 1
we obtain another famous and important basis. The sequence {fPn};-'=l is called the
Schauder system. The Schauder system is a monotone basis of C(O, 1). Indeed,
the linear span of the {fPn};-'=l consists, exactly of the continuous piecewise linear
functions on [0, 1] whose nodes are dyadic points. This shows that (iii) of l.a.3
is satisfied. Since, for every integer n, the interval on which the function fP,,+l(t)
°
is different from is such that on it all the functions {fPi(t)}f=l are linear it follows
immediately that (ii) of l.a.3 holds with K = 1.
4 1. Schauder Bases
Schauder bases have been constructed in many other important Banach spaces
appearing in analysis. Of particular interest in this direction are the results of Z.
Ciesielski and J. Domsta [18] and S. Schonefeld [132] who proved the existence of
a basis in Ck(Jn) (=the space of all real functions f(t 1, t2"'" tn), ti E [0, 1] which
are k times continuously differentiable, with the obvious norm) and the result of
S. V. Botschkariev [13] who proved the existence of a basis in the disc algebra
A (= the space consisting of all the functions f(z) which are analytic on Izl < 1 and
continuous on Izl:::;; 1, with the sup norm). In these papers an important role is
played by the Franklin system. The Franklin system consists of the sequence
{fn(t)};;"=1 of functions on [0, 1] which are obtained from the Schauder system
{Pn};;"=1 by applying the Gram-Schmidt orthogonalization procedure (with respect
to the Lesbegue measure on [0, 1J). The Franklin system is (by definition) an ortho-
nomal sequence which turns out to be also a Schauder basis of C(O, 1). For a
detailed study of the Franklin system we refer to the above mentioned papers as
well as to [17J.
The fact that in the common spaces there exists a Schauder basis led Banach to
pose the question whether every separable Banach space has a basis. This problem
(known as the basis problem) remained open for a long time and was solved in the
negative by P. Enflo [37]. We shall present later on in this book (in Section 2.d)
a variant of Enflo's solution.
The question whether every infinite-dimensional Banach space contains a basic
sequence has, however, a positive answer. This simple fact was known already to
Banach.
Theorem 1.a.5. Every infinite dimensional Banach space contains a basic sequence.
The proof, due to S. Mazur, is based on the following lemma.
Lemma 1.a.6. Let X be an infinite dimensional Banach space. Let Be X be a finite-

dimensional subspace and let s> O. Then there is an x E X with Ilxll = 1 so that
Ilyll:::;; (1 +s)IIY+AXII for every y E B and every scalar A.
Proof of l.a.6. We may clearly assume that s < 1. Let {Yi}7'= 1 be elements of norm
1 in B such that for every y E B with II yll = 1 there is an i for which II y - Yill < s/2.
Let {yt}7'= 1 be elements of norm 1 in X* so that yt(Yi) = 1 for all i, and let x E X
with Ilxll = 1 and yt(x)=O for all i. This x has the desired property. Indeed, let
Y E Y with Ilyll = 1, let i be such that IIY-Yill:::;; s/2 and let A be a scalar. Then
Ily+.\x11 ~ IIYi+AXII-s/2~ yt(Yi+.\x)-s/2= 1-s/2~ Ilyll/(l +s). D
Proof of l.a.5. Let s be any positive number and let {Sn};;"=1 be positive numbers
00
such that Il (1 + sn):::;; 1 + s. Let Xl be any element in

X with norm 1. By 1.a.6 we
n=l
can construct inductively a sequence of unit vectors {X n};;"=2 so that for every n~ 1
Ilyll:::;;(1 +sn)IIY+AXn+lll for all Y E span {Xl,'''' xn} and every scalar A.
a. Existence. of Bases and Examples 5
The sequence {Xn},~)=1 is a basic sequence in X whose basis constant is ,;; 1 +0

(observe that IIPnll,;;
i=n
f:r (1+0;), n=l, 2, ... ). D
Remark. It is useful to note that in the proof of 1.a.6 it is enough to take a vector x
with Ilxll=1 and Iyi"{X) I<0/4 for i=I, ... ,m. Indeed, if IIYII=l and 1,\1;;::2 then
Ily+'\xll;;::llyll while if 1'\1<2 the computation in the proof of l.a.6 gives that
Ily+,\x11 > (l-o)IIYII. It follows from this observation and the proof of l.a.5 that
if {Xn}:'= 1 is a sequence of vectors in X such that lim inf Ilxnll > 0 and Xn ~ 0 then
n
{Xn}:'=1 has a subsequence {X nk }:'=1 which is a basic sequence.
Once it is known that a Banach space has a Schauder basis it is natural to raise
the question of its uniqueness. In order to study this question properly we intro-
duce first the notion of equivalence of bases.
Definition l.a.7. Two bases, {X n}:'=1 of X and {Yn}:'=1 of Y, are called equivalent
00 00
provided a series 2: anxn converges if and only if 2: anYn converges.

n=1 n=l
Thus the bases are equivalent if the sequence space associated to X by {Xn}:'=l
is identical to the sequence space associated to Yby {Yn}:'=1' It follows immediately
from the closed graph theorem that {X n}:'= 1 is equivalent to {Yn}:'= 1 if and only if
there is an isomorphism T from X onto Y for which TXn = Yn for all n.
Using the notion of equivalence, the uniqueness question can be given a mean-
ingful formulation. It turns out however that even up to equivalence bases, if they
exist at all, are never unique.
Theorem l.a.S [120]. Let X be an infinite dimensional Banach space with a Schauder
basis. Then there are uncountably many mutually non-equivalent normalized bases in X.
We shall discuss in detail some aspects of uniqueness of bases in Chapters 2 and

3 below. We shall show there that if we restrict the discussion to bases which have
some nice properties then it is possible to have uniqueness in some interesting
special cases. In this context we shall also present a proof of a weak version of
1.a.8 (namely that there are at least two non-equivalent normalized bases in every
space having a basis).
Schauder bases have certain stability properties. If we perturb each element
of a basis by a sufficiently small vector we still get a basis. The perturbed basis is
equivalent to the original one. The simplest result in this direction is the following
useful proposition [76].
Proposition l.a.9. (i) Let {Xn}:'=l be a normalized basis of a Banach space X with
00
basis constant K. Let {Yn}:'= 1 be a sequence of vectors in X with 2: Ilx n - Ynll < 1/2K.
n=1
Then {Yn}:'=1 is a basis of X which is .equivalent to {Xn}:'=1 (if {Xn}:'=l is just a basic
sequence then {Yn}:'= 1 will also be a basic sequence which is equivalent to {Xn}:'= 1)'
6 1. Schauder Bases
(ii) Let {X,,}:'=1 be a normalized basic sequence in a Banach space X with a basis
constant K. Assume that there is a projection P from X onto [Xn]~=1' Let {y,,}:'=1 be
a sequence of vectors in X such that ~ Ilx"-y,,lI::::; 1/8KIIPII. Then Y=[y,,]:'=1 is
,,=1
complemented in X.
00 00
Proof. For X= L: a"xn E X define Tx= ,,=1

n=1
L: anYn' The series converges and
00 00
(*) Ilx-Txll::::; L la"lllx"-y,,lI::::;max

n=1
la,,1 L IIx,,-y,,1I
"n=1
L Ilx"-y,,ll·
00
::::;2Kllxll
n=1
To prove (i) we have just to observe that under its assumptions III-Til < I and
hence T is an automorphism of X
To prove (ii) we have to observe that if we puty=Tx, then Ily-xll < IIxi1/4 and
in particular IIxll <211yll and IITII <2. Thus
for some 8< 1. Thus S=TP1y-is an invertible operator on Yand S-1TP is a pro-
jection from X onto Y. 0
A very useful method to obtain new basic sequences, starting from a given basis
or basic sequence, is by considering block bases.
Definition l.a.lO. Let {xn}:'= 1 be a basic sequence in a Banach space X. A sequence

Pj+l
of non-zero vectors {Uj}j'=1 in X of the form Uj= L: anx", with {a,,}:'=1 scalars
"=Pj+1
and P1 <P2 < ... an increasing sequence of integers, is called a block basic sequence
or briefly a block basis of the {X,,}:'=1'
It is obvious that a block basis {Uj}J'~1 of {Xn}:'=1 is a basic sequence whose

basis constant does not exceed that of {Xn}:'=1' The usefulness of the notion of
block basis rests very much on the following simple observation [9].
Proposition l.a.H. Let X be a Banach space with a Schauder basis {x,,}:'= l' Let Y
be a closed infinite dimensional subspace of X. Then there is a subspace Z of Y which
has a basis which is equivalent to a block basis of {X,,}:'=1'
Proof. Observe first that since Y is infinite dimensional there is, for every integer p,
00
an element y E Ywith lIyll = I of the form y= L:

,,=p+1
a"x". We construct the block
00
basis of {X,,}:'=1 inductively. Pick any Y1 = L: a",1xn E Y with Ilhll = 1. Let P1 be

n=1
Pl
an integer so that IIYI-Ulll < 1/4Kwhere Ul = L: an,lXn and Kis the basis constant
n=l
of {Xn};'= l' Next we pick a Y2 = 1
an,2Xn E Y with II Y211 = 1 and an integer P2
n=Pl +1
P2
so that IIY2-U211 < 1/42K where u2= L: an,2Xn' We continue in an obvious
n=Pl+1
manner. The sequence {U;}i'=1 obtained in this way is a block basis of {Xn};'=l'
Since ~ IIYj-Uill < 1/3Kit follows by 1.a.9 that {y;}j..l is a basic sequence which
1=1
is equivalent to {Uj}j.. l' The space Z = [Yj]j.. 1 has the desired property. 0
In the proof of 1.a.ll we used the vectors Yi E Ywhose expansions with respect
to the basis {Xn };'= 1 started arbitrarily far. In some instances it is important to be
able to choose a basic sequence out of a subset Y of X which is not a subspace.
For this purpose it is of interest to observe that what we actually need in the proof
00
of 1.a.ll is the following. For every e >0 and every integer p there is a Y= L: anx"
t
. n=l
in Y with II yll ~ 1 and\\n anXnl1 ~ e. This remark proves the following.
Proposition l.a.12. Let {Xn};'= 1 be a Schauder basis of a Banach space X. Let

00
Yk= L: an,kXm k= 1,2, ... ,

,,=1
be a sequence of vectors such that lim sup IIYkl1 >0 and
k
lim an k = 0 for every n (this is the case in particular if Yk ~ 0 and II Ykll
k '
-+- 0). Then
there is a subsequence {Ykj}j..l of {Yk}k'=l which is equivalent to a block basis of
{Xn};'=l'
Proposition 1.a.11 enables us to give an alternative proof of 1.a.S. It is clearly

enough to prove 1.a.S for separable Banach spaces X. Every such Xis isometric to
a subspace of C(O, 1). Hence, by l.a.11, X has a subspace with a basis which is
equivalent to a block basis of the Schauder system in C(O, 1).
h. Schauder Bases and Duality
Let X be a Banach space with a Schauder basis {X n};'= l' For every integer n the
linear functional x~ on X defined by X~(Jl ajxi ) =a" is, by 1.a.2, a bounded linear
functional. In fact Ilx~11 ~2K/llxnll where K is the basis constant of {Xn };'=l' These
functionals {X~};'= 1, which are characterized by the relation x~(xm) = S:;', are called
the biorthogonal functionals associated to the basis {Xn};'=l' Let {Pn};'=l be the
natural projections associated to the basis, i.e. P n (Jl ajxj) = Jl aiXj. For every
1
choice of scalars {aj}f'= and for all integers n < m we have P,f (Jl ajxr) = jt ajXr.
8 1. Scbauder Bases
Hence, by l.a.3, the sequence {X:}:'~l is a basic sequence in X* whose basis con-
stant is identical to that of {Xn}:'=l. Since lim IIPnx-xll =0, for every x E X, we
n
<Xl
get that, in the sense of convergence in the w* topology, x* = 2: x*(xn)x: for every
n=l
x* E X*. In general, this expansion does not converge in norm. We have con-
vergence in norm for every x* E X* if and only if the sequence {X:}:'=l is a basis of
X* i.e., (by l.a.3) if and only if the closed linear span of {X:}:'=l is all of X*. For
this to happen X* must in particular be separable; thus, for X=/r or X=C(O, I)
this cannot happen for any basis. On the other hand this is always the case if X
is reflexive. The following proposition gives a very simple but useful criterion for
checking whether {X:}:'=l is a basis of X*.
Proposition l.b.l. Let {Xn}:'=l be a basis of a Banach space X. The biorthogonal

functionals {X:}:'=l form a basis of X* if and only if, for every x* E X*, the norm
of x~xtl~n (= the restriction of x* .to the span of {x;}f;n) tends to 0 as n --+co. A
basis {xn}:'=;t which has this property is called shrinking.
Proof. If {X~}:'=l is a basis of X* then, for every x* E X*, IIP:x*-x*II--+ O. Since

(P:- 1x*)'[Xill":.n=0 it follows that lim IIx*'[Xll~nll=O. Conversely, assume that
n
IIX*'[Xd~nll--+ 0 and let x E X be any element of norm 1. Then,
where K is the basis constant of {Xn}:'=l. Hence, IIP:x* -x*II--+ 0 0
If X has a shrinking basis it is possible to give a convenient representation of

X** by using the basis.
Proposition l.b.2. Let {Xn};'=l be a shrinking basis of a Banach space X. Then X**
can be identified with the space of all sequences of scalars {a n}:'=1 such that
S~P"J1 aixill <co. This correspondence is given by x**~ (x**(xf), x**(xf), ... ). The
norm of x** is equivalent (and in case the basis constant is 1 even equal). to
s~p IIJ1 x**(Xf)Xill·
Proof. We may clearly assume that the basis constant of {Xn}:'=l is 1. If x** E X**
and {Pn}:'=l are the projections on X associated to the basis then P:*x**=
n
2: x**(xf)xt and IIx**1I = lim
1=1 n
IIP:*x**11 = sup IIP:*x**II. Conversely, if {a n};:'=l is
n
such that s~p IIJ1 aixi// <co then any w* limit point x** of the bounded set
{t aix :=1 satisfies x**(xf)=al for all i. (In particular, the uniqueness of x**
i}
implies that {J1 a1xi}:=1 must tend w* to x**.) 0

Observe that the canonical image of X in X** corresponds to those sequences

{a n}:'=l for which {t aix}~=l is not only bounded but converges in norm.
Another important notion concerning bases, which is in a sense dual to "shrink-
ing", is that of "boundedly complete".
Definition 1.b.3. A basis {X n}:'=l of a Banach space is called boundedly complete if,
for every sequence of scalars {a n}:'=l such that s~p IIJ1 aixill <00, the series n~l anxn
converges.
A typical example of a non-boundedly complete basis is the unit vector basis of

co. The unit vector basis is boundedly complete in all the lp spaces, 1 ~p <00.
If {Xn}:'=l is a shrinking basis in X then {X;n:'=l is a boundedly complete basis
in X*. Indeed, if sup
n
I ;=1~ aixrll,<00 then the expansion of a w* limit point x* of
{J1 a,xr }~=1 with respect to the basis {X;n:'=l must be n~l anx:I'. The converse of
this remark is also valid:
Proposition 1.b.4. A Banach space X with a boundedly complete basis {X n}:'=l is

isomorphic to a conjugate space. More precisely, X is isomorphic to the dual of the
subspace [X,'i'J:'=l of X*.
Proof Let Z = [X:l'J:'=l and let J be the canonical map from X to Z* defined by
Jx(z)=z(x). We claim that J is an isomorphism onto. Indeed, let x E span {Xi}:'=b
for some integer n, and x* E X* be such that [[x*[[ = 1 and x*(x) = [[xli. Then,
P:I'x*(x)=x*(x), P;x* EZ and [[P:I'x*[[~K, where K is the basis constant of
{Xn};=l' Hence, [[x[[/K~ [[Jx[[~ [[xli and this shows that J is an isomorphism. To
show that J is onto observe that {JXn}:'=l are functionals biorthogonal to {Xn~=l
in Z*. Let z* EZ*. The sequence {t z*(XnJXi}~=l is bounded in norm (by
K[[z*[[) and thus, since {X n}:'=l is boundedly complete, the series I
n=l
z*(x:l')Xn con-
verges in X to an element x. Clearly z* =Jx. D
By combining the notions of shrinking and boundedly complete we get the

following elegant characterization of reflexivity in terms of bases.
Theorem l.b.5 [53]. Let X be a Banach space with a Schauder basis {Xn }:'=l' Then
X is reflexive if and only if {X n}:'=l is both shrinking and boundedly complete.
Proof Assume first that X is reflexive. We already observed above that {Xn}:'=l
n 1=1
I
must be shrinking. Also if sup ~ aixill <00 then any w limit point x of {~ aixi}'"
1=1 n=l
'"
must be of the form L anxn and in particular this series converges.
n=l
10 1. Schauder Bases
The converse assertion follows immediately from l.b.2. We shall give an addi-
tional proof of the converse assertion since it is somewhat more convenient to
use it in more general situations in which an analogue of l.b.5 is valid. Let {Yk}:'=l
be a sequence of vectors of norm 1 in X. By the diagonal procedure we can find
a subsequence {Ykj}f'=1 of {Yk}:'=l so that an = lim X;1:(Ykj) exists for every n. We
j
have that ~
1=1
ajxj=limPnYkj and hence, sup
j n
II i aixill~K. Since {Xn}:=1 is boun-
1= 1 .
00
dedly complete Y= 2: anxn exists and, by definition, lim x;1:(Ykj)=x;f(y), for every
n=l j
n. Since the basis {X n}:=l is also shrinking [X;1:]:=l = X* and therefore w lim Yk;= y.
j
This proves that X is reflexive. 0
In l.b.5 we considered a single basis in X. In [147] M. Zippin showed that if

we consider all the bases in a given space it is enough to use only one of the two
properties appearing in l.b.5. More precisely: a Banach space X with a basis is
reflexive if (and clearly only if) every basis in X is shrinking or, alternatively, if
every basis in X is boundedly complete.
We pass now to questions relating existence of bases and duality. These
questions are non-trivial (and thus of interest) only for non-reflexive Banach
spaces. If a Banach space X has a basis its dual X* need not have a basis even if
X* is separable (cf. Section e below). An interesting and rather deep result of
W. B. Johnson, H. P. Rosenthal and M. Zippin [61] shows that the existence of a
basis in X* does imply that also X has a basis.
Theorem l.h.6. Let X be a Banach space such that X* has a basis. Then X has a
shrinking basis and therefore X* has a boundedly complete basis.
The proof of this theorem is closely related to the local theory of Banach spaces
and therefore we shall give it only in Vol. IV.
In Section a we proved that every Banach space X contains a basic sequence.
For a reflexive X this result implies that X has a quotient space with a basis. W. B.
Johnson and H. ·P. Rosenthal proved in [60] that the same holds for general
separable Banach spaces.
Theorem l.h.7. Every separable infinite dimensional Banach space has an infinite-
dimensional quotient space with a basis.
For the proof of l.b.7 we introduce first the following.
Definition l.h.S. A basic sequence {X;f}:=l in the dual X* of a Banach space X is

called a w* basic sequence if there exists a sequence {X n}:= 1 in X for which x;f(xm) = 8;;'
and such that, for every x* in the w* closure of span {x:}:=l> we have
n
x* = w* lim
n
2: x*(x,)X[.
j=1
The next proposition clarifies the meaning of the notion of a w* basic sequence
and its relation to 1.b.7.
Proposition l.b.9. A sequence {X;n:'=l E X* is a w* basic sequence if and only if

there is a basis {Yn}:'=l 01 Y=X/([X;n:'=l)T so that x:=T*y;l', n=l, 2, ... , where
T: X -+ Y is the quotient map and {y;l'}:'=l are the lunctionals biorthogonal to
{Yn}:'=l'
This proposition is proved by a straightforward verification. Note that T* is
an isometry from y* onto the w* closure of [X:]:'=l since it is w* continuous.
In the proof of l.b.7 we shall use also the following fact which is a direct con-
sequence of the w* density of the unit ball of a Banach space X in the unit ball of
X**. Let B be a finite dimensional subspace of X* and let S > O. Then, there exists
a finite set F of elements of norm 1 in X so that, for every IE B* with IIIII = 1,
there is an x E Fwhich satisfies I/(x*)-x*(x)l::::;sllx*ll, for all x* E B.
Prool 01 l.b.7. For every infinite-dimensional Banach space X the set

{X*EX*; Ilx*ll=l} is w* dense in the unit ball of X*. Since Xis separable the
unit ball in X* is w* metrizable and hence there is a sequence {Xnk'=l in X* so
that IIxi.\' II = 1 for all k and w* lim xi.\' =0. We shall construct a subsequence of
{Xi.\'}k'=l which is w* basic (this will conclude the proof in view of l.b.9).
'" Sn <00. By the
Let {Sn}:'=l be a sequence of numbers so that O<sn < 1 and 2:
n=l
remark above and the separability of X we can choose inductively a sequence of
integers kl < k2 < ... and a sequence of finite sets Fl c F2 C ... of elements of
norm 1 in X so that
(i) X=span U Fn
n=l
(ii) For everyI E ([x~,]f= 1)* with IIIII = 1 there is an x E Fn so that I/(x*) - x*(x) I
<snllx*11/3 for every x* E [4,lf=1.
(iii) IxL\'n+l(x)l::::;sn/3 for all x E Fn.
We claim that {xL\'n},':"=1 is a w* basic sequence. Note first that, by the proof of
l.a.5 (and the remark following it), this is a basic sequence. Moreover, if {Pn }:'=l
denote the natural projections (on [4n ]:'=1) which are associated to this basic
sequence, then
(*)
Let {Yn}:'=l c([xi.\'n]:'=l)* be the functionals biorthogonal to {xL\'n}:'=l' In view

of l.b.9 it suffices to show that the operator T: X -+ ([xL\'n]:'=l)* defined by
Tx(x*)=x*(x), x* E [xL\'nl:'=l maps X onto [Ynl:'=l and thus, in particular, is a
quotient map (note that the kernel of Tis ([XL\'nJ:'=l)T).
We show first that TXc [Ynl:'=l' This follows from (i) and (iii) since if x E Fn
for some n then ~ Ix,t.(x)
i=l ~
I <00 and hence Tx= ~ X,t.(X)Yi E [Ynl:'=l'
i=l t
To prove that TX exhausts all of [Ynl:'=l it suffices to show that, for every
Y E span {Yn};''' 1 of norm 1 and every e> 0, there exists an x E X with Ilxll = 1 and
IITx-yll <4e (the desired result will then follow by successive approximation). We
may assume that Y E span {YI}f=l and that n is so large that e> ~ el and IIPml1 <
I=n
1+e for m~n (use (*)). Foru E span {yaf=l we denote IIUI[xihlf=lll by Ilu111' Observe
that for every such u
Choose anx E Fn which satisfies (ii) for z=y/llyII1' We get that Itt x:1(x)YI-ZII1 <
en /3 <e/3 and hence 'I ~ x:1(x)Yi-ZII <2e/3. Since IIYIII = IIPi-Pi-11l ~4 for i~n
I
11=1
we deduce from (iii) that ~ x~(X)Yjll<4e/3. Hence

l=n+1
~2e/3+4e/3+2e~4e .
This concludes the proof of 1.b.7. 0
It is not known whether 1.b.7 is true without the separability assumption on X.

This question is clearly equivalent to the following: Does every infinite-dimensional
Banach space have an infinite dimensional and separable quotient space?
Another open problem which arises naturally in view of 1.a.5 and 1.b.7 is the
following
Problem l.b.IO. Let X be an infinite-dimensional separable Banach space. Does there

exist a subspace Y of X so that both Yand X/ Y have a Schauder basis?
In Section g below we shall present a partial positive answer to 1.b.lO.

The construction used in the proof of 1.b.7 yields under an additional assump-
tion a boundedly complete basis. Before stating this precisely we prove a useful
renorming result due to Kadec [65] and Klee [71]. The proof we present is taken
from [25].
Proposition l.b.ll. Let X be a separable Banach space and let Y be a separable

subspace of X*. Then, there is an equivalent norm 111·111 on X so that, whenever
x;t'~x* with {X;t'};'=lcX*,X*E Y and IIlx:III-+lllx*lIl, we have also that
II Ix: - x* III -+ O. ClII·1I1 denotes here the new norm in X as well as the new norm
induced by it in X*.)
Proof. Let B1 c B2 C ••• be a sequence of finite dimensional subspaces of Y whose

union is dense in Y. Define a new norm on X* by putting
co
Illx*III=llx*ll+ 2: 2- n d(x*,Bn)
n=1
where d(x*, Bn) denotes the distance with respect to 11·11 of x* from Bn (i.e., the
norm of the canonical image of x* in the quotient space X*/Bn). Clearly 111·111 is
an equivalent norm on X* and since its unit ball is w* closed (observe that each
Bn is w* closed) it is induced by an equivalent norm in X which is also denoted by
111·111·
Assume now that x: ~ x* with x* E Y and Illx:11I ~ Illx*llI. Then, since
liminfd(x:,Bk)~d(x*,Bk) for all k and liminfllx:II~llx*ll, we deduce that
n n
d(x*, Bk)=limd(x:, Bk). From this and the fact that limd(x*,Bk)=O it follows
n k
that for every e > 0 there exists an integer k and elements u: E B k , n = 1, 2, ... , such
that Ilx: - u;t II < 8/4 for n sufficiently large. Since Bk is finite-dimensional, by pass-
ing to a subsequence if needed, we can assume that u: ~ u* for some u* E B k. Thus,
for n large enough, Ilx: -u*11 <e/2 which, by taking the w*-limit, gives Ilx*-u*ll~
e/2 i.e., Ilx*-x:ll~e. This proves that Ilx:-x*11 ~O or, equivalently, that
Illx:-x*111 ~O. 0
Observe that if X* is separable we can take Y=X*. In this case we obtain a

renorming of X so that in X*, w* convergence on the boundary of the new unit
ball is equivalent to norm convergence.
We can now prove the result ensuring the existence of boundedly complete
basic sequences [60] at which we hinted above.
Proposition l.b.12. Let X be a Banach space whose dual X* is separable. Then every
sequence {X!n~=1 in X* such that x~ ~ 0 and lim sup Ilx~1I >0 has a boundedly
k
complete basic subsequence {x~n}:= l'
Proof. By l.b.ll there is no loss of generality to assume that in X*, y: ~ y* and

lIy:11 ~ Ily*11 => Ily: -y*1I ~ O. In the proof of l.b.7we showed that{x~}~=1 has a
n
w* basic subsequence {x!:n}:= 1 for which II P nil ~ 1 (see (*». Thus, if y: = i~1 atx!:i'
n= 1,2, ... , is a bounded sequence it follows from the fact that {X~n}:=1 is w* basic
that y: converges w* to a limit y*. Since
lIy*11 ~lim inf Ily:1I ~ lim sup Ily:11 = lim sup IIP;ty* II~ Ily*11
n n n
itfollowsthatIIY:-y*II~Oi.e., ~ anx~nconverges. 0
n=1
The dual result to l.b.12 is also true. This was observed in [30].
Proposition l.h.13. Let X be an infinite-dimensional Banach space with a separable

dual. Then X contains a shrinking basic sequence.
Proof Let {yl!'}k'=l be a sequence which is norm dense in the unit ball of X*. The
construction of {Xn};:"=l in the proof of 1.a.5 can clearly be carried out so that in
addition to the properties required there we have also that yi!'(xn)=O for n>k.
The basic sequence {Xn};:"=l obtained in this manner is obviously shrinking. 0
The preceding propositions imply an interesting result whose statement has

nothing to do with bases. This result, stated by V. D. Milman [106] and first
proved by W. B. Johnson and H. P. Rosenthal [60], is a very good illustration
of the use of bases in the investigation of linear topological properties of Banach
spaces.
Theorem l.h.14. (i) Let X be a Banach space whose dual X* is separable. Let Y
be an infinite dimensional subspace of X* with a separable dual Y*. Then Y has an
infinite-dimensional refleXive subspace.
(ii) Let X be an infinite-dimensional Banach space whose second dual X** is
separable. Then every infinite dimensional subspace of X or of X* contains an infinite-
dimensional reflexive subspace.
Proof [60]. (i). By l.b.13 Y contains a shrinking basic sequence {Yk}k'=l with
iiYkii = 1 for all k. Clearly, Yk ~ 0 and thus also Yk ~ 0 (as an element of X*). By
l.b.l1 there is a subsequence {Yk n};:"=l of {Yk}k'=l which is a boundedly complete
basic sequence. Since {Yk n};:"=l is of course also shrinking we deduce from l.b.5
that [Ykn ];:"=l is reflexive.
(ii) Both assertions are immediate consequences of (i). For example, let Y
be an infinite-dimensional subspace of X. Then y* is separable and Y is a subspace
of the separable conjugate X** so (i) applies to Y. 0
We conclude this section by mentioning the following result [24].
Theorem l.h.15. Let X be a Banach space whose dual X* is separable. Then there
is a Banach space Y with a shrinking basis which has X as a quotient space. In
particular, X* is isomorphic to a subspace of a space with a boundedly complete
basis, namely Y*.
The proof of this theorem is quite long. Since the theorem will not be used in
the sequel we omit its proof and refer the interested reader to [24].
Note that every separable Banach space is a quotient space of h which has a
boundedly complete basis. Thus, in one sense, the dual to l.b.15 is trivially true.
In another sense the dualization of l.b.15 leads to the following open problem.
Problem l.b.16. Let X be a Banach space with a separable dual. Is X isomorphic to

a subspace of a space with a shrinking basis?
c. Unconditional Bases 15
c. Unconditional Bases
The existence of a Schauder basis in a Banach space does not give very much
information on the structure of the space. If one wants to study in more detail the
structure of a Banach space by using bases one is led to consider bases with
various special properties. In Section b we encountered already two such useful
types of bases, namely shrihking and boundedly complete bases. Undoubtedly,
the most useful and widely studied special class of bases is that of unconditional
bases.
Before studying unconditional bases we present some general facts concerning
unconditional convergence.
Proposition l.c.l. Let {Xn}:=l be a sequence of vectors in a Banach space X. Then

the following conditions are equivalent.
co
(i) The sertes 2: X,,(n) converges for every permutation
n=l
7T of the integers.
co
(ii) The series 2:
1=1
xnj converges for every choice ofnl <n2<na .. ·.
co
(iii) The series2: 8nxn converges for every choice of signs 8n (i.e. 8n= ± 1).
n=1
(iv) For every e> 0 there exists an integer n so that Illter XIII < efor every finite
set of integers a which satisfies min {i E a} > n.
co
A series 2:
Xn which satisfies one, and thus all of the above conditions, is said
n=1
to be unconditionally convergent.
Proof. The equivalence of (ii) and (iii) is obvious. If (iv) holds then the partial
sums of the series appearing in (i) and in (ii) satisfy the Cauchy condition and thus
(iv) => (i) and (iv) => (ii). Assume that (iv) is not satisfied. Then, there is an e>O
and finite sets {an}:= 1 of integers so that
and II ieern U an is a subsequence of the integers

2: XIII;:,: e, for all n. It is clear that a = n=l
for which 2: Xi does not converge (hence (ii) => (iv». Also if
ieer 7T is a permutation of
the integers which, for every n, maps the set {i; Pn ~ i ~ qn} onto itself in such a
manner thabr- 1 (an) = {Pn, Pn + 1, ... , Pn +kn}, where k n is the cardinality of am then
co
2: X"(i) does not converge (and hence (i) => (iv».
;=1
0
co
It is easily verified that if 2:
Xn converges unconditionally then the sum of
co
n=1
2:
n=l
X,,(n) does not depend on the permutation 7T. The set of vectors of the form
""
L 0nXn, 0" = ± 1 forms a norm compact set (by (iv) the map from {-I, l}N into
n=l
'" 8nxn is continuous). It is also easy to verify
X which assigns to {On}:=l the point L
,,=1
"" Xn converges unconditionally then, for every bounded sequence of
that if L
,,=1
00
scalars {a n}:=l, the series L anxn converges and the operator T: f"" -J>- X, defined
n=l
00
by T(ah a 2, ... )= L anxn , is a bounded linear operator.

n=l
00
In finite-dimensional spaces a series L x" converges unconditionally if and

n=l
L Ilxnll <00. In every infinite-dimensional space
00
only if it converges absolutely, i.e.

n=l
00
there exists a series L Xn which converges unconditionally but not absolutely.

n=l
More precisely we have the following result of Dvoretzky and Rogers [34].
Theorem l.c.2. Let X be an infinite-dimensional Banach space. Let {A n}:=l be a

"" A; <00. Then, there is an unconditionally
sequence of positive numbers such that L
n=l
"" Xn in X such that
convergent series L Ilxnll = A", for every n.
n=l
For the proof of 1.c.2 we need the following result, due to Auerbach, which
is useful in many contexts.
Proposition l.c.3. Let B be a Banach space of dimension n. Then, there exist n vectors
{xi}f=l of norm 1 in Band n vectors {Xnf=l of norm 1 in X* so that xj(x,)=Si.
Proof Introduce a coordinate system in B and, for Yl, ... , Yn in the unit ball of B,
let V(YbY2, ... ,Yn) be the determinant of (al,f)4f=1, where (al,ha,,2, ... ,ai,n)
denote the coordinates of Yi, 1:( i:( n. The function V attains its maximum at an
n-tuple {Xb X2,'''' xn} of vectors of norm 1. Put
The n-tuples {Xi}r=l and {XnY=l have the desired property. D
The n-tuples of vectors {XI}f=l, whose existence is ensured by 1.c.3, is called an

Auerbach system.
The main step in the proof of 1.c.2 is the next lemma (the proof we present is
taken from [43]).
Lemma l.c.4. Let B be a Banach space of dimension n 2 and norm 11·11. Then there
is an n-dimensional subspace C of B and an inner product norm III Ilion C so that
II yll :( III YIII, for all Y E C, and an orthonormal (with respect to 111·11)) basis {Yi}Y= 1
of C with II Yill ~ 1/8, for every i.
1/2
IIIxll11 =n
n2 )
Proof Let {XJY!l be an Auerbach system in B and put (
}~1 xj(x)2 .
Then, 111·1111 is an inner product norm on Band
n2
IllxII11In2~m~x Ixj(x)I~llxll~
1
I
}=1
Ixt(x)I~lllxII11'
Consider the following statement

(t) Every subspace C of B with dim C>dim BI2 contains a vector Y with
IIIyll11 = 1 and Ilyll > 1/8.
If (t) is true we can construct inductively at least n2 /2-1 elements Yi which
are orthonormal with respect to 111·1111 and satisfy II yd I~ 1/8, for every i. In this
case there is nothing more to prove.
If (t) does not hold there is a subspace B2 of B with dim B2 > dim BI2 and an
inner product 111,1112= 111·1111/8 on B2 so that 8111x11121n2~ Ilxll~ Illxllb for every
x E B2. Consider now the statement obtained from (t) by replacing B by B2 and
111·1111 by 111·1112. If the statement we get is true we have nothing more to prove. If
it fails there is a subspace Bs of B2 with dim Bs > dim B2/2 and an inner product
norm 111·llls on Bs so that 82111x11131n2~ Ilxll ~ Illxllls, for every x E Bs.
We continue in an obvious way. The process must end after /-1 steps for an
integer I such that 81 ~ n2 • The space BI will be of dimension ~ n2 /2 /- 1 • In this space
we can find at least dim Bzl2-1 vectors Yl which are orthonormal with respect to
111·1111 and satisfy IIYil1 > 1/8, for every i. Since n2~ 81 we get that n2·2- 1-1 >n and
this concludes the proof. 0
Observe that the unit vectors uj=ydIIY111, l~i~n, whose existence was proved
in l.c.4, satisfy
for every choice of scalars {aiH=l'
2: Ar <00.
00
Proof of I.c.2. Let t\ be positive numbers such that Choose an in-

1= 1
00
creasing sequence of integers {n/c}k'=l such that 2: Ar~2-2k, k= 1,2, .... By the
i=nk
preceding lemma we can find in any Banach space of dimension ~ (nlc+ 1 - nk)2
(and thus in every infinite-dimensional Banach space) unit vectors {Ui}~~~; -1 so
that if we put Xi = t\Ui then, for every choice of signs ()i,
00
For i < n1 we take as Xi any vector in X of norm t\. The series 2: Xi converges
!= 1
unconditionally and clearly Ilxd 1= t\, for all i. 0
Theorem I.c.2 is the strongest possible general result in this direction. It follows
readily from the parallelogram identity in Hilbert space that if {XI}f=1 are any n
vectors in 12 then the average of It~1 8i xf taken over all 2" choices of signs {8 }f=1 i
is equal to 2:
"
1=1
II XI I12. Consequently, if 2:
00
1=1
XI is an unconditionally convergent
00
series in 12 then 2:
1=1
IlxII12<oo. We shall discuss possible converse statements to
I.c.2 in other examples of Banach spaces and their relation to uniform convexity
in Vol. II (see also Remark 2 following 2.b.9 below).
We pass now to unconditional bases.
Definition l.c.5. A basis {x,,}:'= 1 of a Banach space X is said to be unconditional if for

00
every x E X, its expansion in terms of the basis 2: a"x" converges unconditionally.

,,=1
The following proposition is an immediate consequence of I.c.I.
Proposition l.c.6. A basic sequence {x,,}:'= 1 is unconditional if and only if any of the
following conditions holds.
(i) For every permutation 7T of the integers the sequence {x,,(,,)}:'=1 is a basic
sequence.
00
(ii) For every subset u of the integers the convergence of 2: a"x"

,,=1
implies the
convergence of 2: a"x".
"e" 00 00
(iii) The convergence of 2: a"x"

,,=1
implies the convergence of 2:
,,=1
b"x" whenever
Ib,,1 ~ lanl, for all n.
It follows from (ii) and the closed graph theorem that if {Xn}:'=1 is an uncon-
ditional basic sequence and u is a subset of the integers then there is a bounded
linear projection p .. defined on [X,,]:'=1 by P .. (
,,=1 "e ..
i
a"x,,) = 2: a"x". These projec-
tions are called the natural projections associated to the unconditional basic sequence.
For the sets u of the form u={I, 2, ... , n} we get that the projections p .. coincide
with the projections PIt which are the natural projections associated to the basic
sequence {x,,}:'= l ' Similarly, for every choice of signs 8 ={ 8n}:'= 1> we have a bounded
linear operator Me on [X,,]:'=1 defined by MeC~1 a"xn) = "~1 a"8"x,,. Observe that
if u={n; 8" = I} then P.. =(I+M..)/2, and if 1) = {7Jn}:'= 1 is another sequence of signs
then MeM1/=Me1/ where (81))"=8,,1),,. The uniform boundedness principle implies
.. ..
that sup IIP.. II and sup IIMel1 are finite. These numbers are related by the in-
equality
sup IIP.. II~sup IIMell~2 sup IIP.. II.

.. e ..
The number sup IIMel1 is called the unconditional constant of {X,,}:'=l' Observe
e
that the unconditional constant of a basis is always larger or equal to the basis
constant. If {Xn}:'=l is an unconditional basis of X we can always define on X an
equivalent norm so that the unconditional constant becomes 1. We have simply to
take as a new norm the expression Illxlll = sup IIMaxll. Every block basis of an
a
unconditional basis is again unconditional. The unconditional constant of a block
basis is smaller or equal to the unconditional constant of the original basis. If
{Xn}:'=l is an unconditional basis of X then the biorthogonal functionals {X,n:'=l
form an unconditional basic sequence in X* whose unconditional constant is the
same as that of {Xn}:'=l'
Another often used trivial observation concerning the unconditional constant
is the following.
Proposition 1.c.7. Let {Xn}:'=l be an unconditional basic sequence with an uncondi-

00
tional constant K. Then, for every choice of scalars {an}:'=l such that L anxn con-
n=l
verges and every choice of bounded scalars {An}:'=b we have
(in the real case we can take K instead of 2K).
Proof Assume the scalars are real and pick an X* E X*, with Ilx*II=I, so that
n~l Ananx*(xn)=IIJIAnanxnll· Let {On}:'=l be defined by On=1 ifanx*(xn);:.O and
On = - 1 if anx*(xn) < 0. Then
Iln~1 Ananxnll ~ n~l IAnllanx*(xn)1 ~ s~p IAnl n~ Onanx*(xn)
~ s~p IAnl x* (Ma C~l anxn)) ~ s~p IAnl' Klln~l anxnll'

If the scalars are complex we get the desired result by considering separately
00
the real and imaginary parts of L anx*(xn).

n=l
0
. The simplest examples of unconditional bases are the unit vector bases in Co
or in [p, 1 ~p <ct). A much more interesting example of an unconditional basis is
the Haar system in LiO, 1), 1 <p <00. We shall prove the unconditionality of this
basis in Vol. II. There is a general simple procedure of constructing an uncon-
ditional basis. Let {Xn}:'=l be a sequence of non-zero vectors in a Banach space
X. Let Xo be the completion of the space of all sequences of scalars y = (aI' a2 , ..• ),
which are eventually zero, with respect to the norm
The unit vectors form an unconditional basis of Xo with unconditional constant 1.

Obviously, this basis is equivalent to {X n};:'= 1 if and only if {X n };:'= 1 is itself an un-
conditional basic sequence.
A simple and important example of a basis which is not unconditional is the
summing basis in c. It consists of the vectors
n-l
~
Xn = (0, 0, ... , 0, 1, 1, ... ), n = 1, 2, 3, ....
The norm of n~l anxn is l~~~m IJI £7;1· The basis {Xn};:'=l is a monotone and normal-
ized basis of c which is not unconditional since IIJI Xiii =n while IIJI (-lYXill = 1,
for all n. It is of interest to note that if we apply the previously described general
procedure of constructing unconditional bases to the sequence {Xn};:'=l in c we get
as Xo the space 11' As we shall see in the next section, the Schauder system in
ceO, 1) and the Haar basis of Ll(O, 1) are not unconditional. In the next section
we shall present several other examples of bases which are not unconditional.
Bounded linear operators which map one sequence space into another sequence
space have a natural representation by an infinite matrix. If {Xi}~l is a basis of X
and {Yj}i'= 1 is a basis of Y, the matrix A = (ai. j) corresponding to a bounded linear
00
operator T: X ---'r Y is defined by the relation Tx; = L ai ' jYj.

j=l
This representation
is especially useful in the case when both bases are unconditional. We shall prove
here only one simple fact (cf. [139]), concerning matrix representation, which will
be applied in the sequel. A study of some related and deeper questions may be
found in [78].
Proposition :i.c.S. Let the matrix A = (ai, j) represent a bounded linear operator T
from a Banach space X into a Banach space Y with unconditional bases {Xi};";, 1 and
{Yj}i'=b respectively. Then the diagonal of A (i.e. the matrix (8jai,j)) also represents
a bounded linear operator D from X into Y. if the unconditional constants of {Xi}~ 1
and {Yj}i'=l are 1 then IIDII::::;IITII.
Proof It is clearly enough to prove only the second part of the statement. Assume
.)
therefore that the unconditional constants are 1. Notice that the matrices
C~"
-al,2 -al,3 al,2 al,3
C"'" 'J
A _ a2,1 a2,Z aZ,3 ... -a21 aZ,2 aZ,3 ..,
1- A2 = '
a3,1 a3,2 a3,3 -- a3, 1 a3,Z a3,3
represent operators with the same norm as that of T. Hence, the matrix (AI + A z)/2
i.e.,
C~'"
° °
aZ,2
a3,Z
aZ,3
a3,3
H)
...
represents an operator of norm ~ IITII. Applying a similar procedure to this matrix,

using the second column and row, we get that
- a1,l 0 0 0
0 -a2,2 0 0
0 0 a3,3 a3,4
0 0 a4,3 a4,4
also represents an operator of norm ~ IITII. By continuing inductively we obtain

the desired result. 0
Remarks 1. The same method of proof shows that I.c.8 is also valid for "block
diagonal" matrices. More precisely~ if {m/c}k'=l and {n/c}k'=l are increasing sequences
of integers and
for m/c~i<m/C+l> n/c~j<n/C+l> k= 1,2, ...

otherwise
then (dt ,1) represents a bounded linear operator (whose norm does not exceed that
of T if the unconditional constants are 1).
2. A variant of l.c.8 which is valid even when only one of the spaces has an
unconditional basis will also be used later on. Let Z be a Banach space with an un-
conditional basis {Zn}:'=l and let {Yn}:'=l be a sequence of non-zero vectors in Z. Let
V be the completion of the space of all finite sequences of scalars v=(al> a2,"')'
which are eventually zero, equipped with the norm
<Xl
Put Yt = L at ' 1Z1

1=1
and observe that the operator T: V ~ Z, defined by
<Xl
Tv= L anYn, has norm ~ I. Since the unit vectors form an unconditional basis of
n=l
Vit follows from I.c.8 that the diagonal of (a;, 1)1':1= 1 defines a bounded operator
from V into Z. More precisely, for each sequence of scalars {an}:'=l, we have
where K denotes the unconditional constant of {Zn}:'=l'
For Banach spaces having an unconditional basis there are two fundamental
structure theorems which are due to R. C. James [53].
Theorem 1.c.9. Let X be a Banach space with an unconditional basis {Xn}:'=l' Then
{Xn}:'=l is shrinking if and only if X does not have a subspace isomorphic to 11'
Proof. If h is isomorphic to a subspace of X then X* is non-separable and therefore

no basis of X can be shrinking. Conversely, if {Xn};:'=l is not shrinking there is an
x* E X* with Ilx*1/ = 1, an e> 0 and a normalized block basis {Uj}j.,l of {Xn};:'=l so
that x*(uj)~e for every j. Hence, for every choice of positive {aj}T=l,
It follows that, for every choice of scalars {aj}T= 1, Itt ajUjll ~ e j~l lafllK, where K
is the unconditional constant of {Xn};:'=l. The {Uj}j.,l are thus equivalent to the
unit vector basis of h. 0
Theorem l.c.lO. Let X be a Banach space with an unconditional basis {Xn};:'=l. Then,
the following three assertions are equivalent.
(i) The basis is boundedly complete.
(ii) X is weakly sequentially complete (i.e. if {YI}~ 1 C X are such that lim X*(Yi)
1
exists for every x* E X* then there is ayE X such that X*(y) = lim X*(YI)
j
for every X* E X*).

(iii) X does not have a subspace isomorphic to co.
Proof. We assume, as we clearly may, that the unconditional constant of {Xn};:'=l

is 1. Since Co is not w sequentially complete it is clear that (ii) =;> (iii). We shall
prove now that (iii) =;> (i). Assume that the basis is not boundedly complete. Then
there exist scalars {an};:'=l such that Itt alxill~ 1, for every n, but n~l anxn does not
converge. It follows that there is an e> 0 and a sequence of integers PI < ql <
qj
P2<q2". so that ifuj= Z

1=1'j
alXj then Ilujll ~e, for every j. It follows from I.c.7 that
for every choice of {,\j}T= 1>
On the other hand, 11ft ,\jUjll~B s~p 1\1 and hence, {Uj}i=l is equivalent to the
unit vector basis in co.
In order to prove the remaining implication in I.c.10 (i.e. that (i) =;> (ii» we
need the following lemma.
Lemma l.c.H. Let {Xn};:'=l be an unconditional basis of a Banach space X with

biorthogonalfunctionals {X:};:'=l. Let {Yi}~l be a bounded sequence in X such that
lim X*(Yi) exists for every x* E X*, and such that lim X:(YI) = 0 for every n. Then,
1 1
lim x*(Yj) = 0 for every X* 6 X*.
j
Proof. Assume that, for some x* E X* and some e > 0, X*(YI) ~ e for all i. Since
lim X~(YI)=O for every n, there is a subsequence {Yl k }:'=1 of {Y,}f=1 so that
I
IIYlk -ukll <2-"', for some block basis {Uk}:'=1 of {X n}:=I' Since X*(Uk) > ej2, for
sufficiently large k, the proof of l.c.9 shows that {Uk}:'=I, and therefore also
{Y 1k}:'=1> are equivalent to the unit vector basis of 11' Thus, there is an element
y* E X* such that Y*(YIJ=( -1)k, for every k. This contradicts the assumption
that lim Y*(YI) exists. 0
;
We now prove (i) ~ (ii) of l.c.lO. Assume that {Xn}:=1 is boundedly complete and
that lim x*(y;) exists for every x* E X*. Put an = lim X~(YI)' n= 1,2, .... For every
; I
integer m, IIn~ 1 anxnll =li:n IIPmYIl1 ~ s~p IIYIII· Hence, n~1 anxn converges to an ele-
ment Y EX. By applying l.c.ll to {YI-Y}f=1 we get thaty ~ y. 0
The following, theorem is an immediate consequence of l.c.9, l.c.1O and the

results l.bA, l.b.5 of the previous section.
Theorem I.c.n. (a) A Banach space X with an unconditional basis which does not
have subspaces isomorphic to Co or h must be reflexive. In particular, if X has an
unconditional basis and X** is separable then X is reflexive.
(b) A weakly sequentially complete Banach space with an unconditional basis is
isomorphic to a conjugate space.
(c) If X has an unconditional basis and X* is separable then X* has an uncon-
ditional basis.
In connection with assertion (c) above and Theorem l.b.6 it is worthwhile to

remark that the existence of an unconditional basis in X* does not imply that X
has an unconditional basis. For example, let K be the set of ordinals ~ w"', en-
dowed with their usual order topology. We shall prove in Vol. III that C(K) does
not have an unconditional basis. On the other hand C(K)*, which is isometric to
h, has an unconditional basis.
The preceding results were generalized by Bessaga and Pelczynski [10, 11], as
follows
Theorem l.c.B. Let Y be a closed subspace of a Banach space X with an uncon-

ditional basis. Then,
(a) Y is weakly sequentially complete if and only if Y contains no subspace
isomorphic to co.
(b) Each norm bounded set in Y is weakly conditionally compact (i.e. each
bounded sequence in Y contains a subsequence which is Cauchy in the weak
sense) if and only if Y contains no subspace isomorphic to 11'
(c) Y is reflexive if and only if Y contains no subspace isomorphic to Co or 11'
We shall prove l.c.13 in Vol. II in the more general setting of spaces Ywhich
are subspaces of suitable Banach lattices. H. P. Rosenthal has proved that part
(b) above holds without any assumption on Y (Le. for this we do not have to assume
the existence of x~ Y with an unconditional basis). This is a deeper result and we
shall discuss it in detail in Section 2.e below.
d. Examples of Spaces Without an Unconditional Basis
Since the notion of an unconditional basis is much stronger than that of a basis
it is naturally much easier to exhibit examples of separable spaces which fail to
have an unconditional basis than to exhibit spaces which fail to have a basis
altogether. The most common non-reflexive classical function spaces, i.e. C(O, 1)
(cf. [68]) and L 1 (0, 1), fail to have an unconditional basis. That L 1 (O, 1) fails to have
an unconditional basis can be deduced from 1.c.12(b) sinceL1 (0, 1) is w sequentially
complete and not isomorphic to a conjugate space (this latter fact will be proved
in Vol. III). We present here a short proof of the fact, due to Pelczynski [115], that
L 1 (0, 1) is not even isomorphic to a subspace of a space with an unconditional
basis. The proof we give is due to V. D. Milman [107].
Proposition I.d.l. The space L 1 (O, 1) is not isomorphic to a subspace of a space

with an unconditional basis.
Proof Let {r n(t)}:'=l be the Rademacher functions on [0,1] defined by rit)=

sign sin 2 nTTt. Then, for every x E L 1 (O, 1), we have
[[x(t)+x(t)rn(t)[[-+ [[x(t)[[
(to check, e.g., the second statement observe that if x is the characteristic function
of an interval [k2- n, (k+ 1)2- n] then [[x+rmx[[ = [[x[[ for m>n).
Assume now that L 1(O, l)c Yand that Y has an unconditional basis {yj}~l'
Pick any Xl EL1 (0, 1) with [[x 1 [[=1. By the observation above we can define
inductively vectors {X n }:'=2 in L 1(0, 1) of the form
so that -!~[[xn[[=[[Xl+X2+ "'+xn-1[[~2, for all n, and so that [[xn-un[[~2-n,

where {U n}:'=l is a suitable block basis of {yj}~ l ' It follows from these relations that
{un}:'= 1> and therefore {x n}:'= 1, is equivalent to the unit vector basis of co. This is
impossible since L 1 (0, 1) is w-sequentially complete. 0
Since every separable Banach space is isometric to a subspace of C(O, 1) it

follows from 1.d.1 that also C(O, 1) cannot be embedded in a space with an uncon-
ditional basis.
We shall consider now another example of a separable Banach space which,
among its other interesting properties, cannot be embedded in a space with an
unconditional basis. This example, which is due to R. C. James [53, 54], had an
important role in the development of Banach space theory and is still now a source
of inspiration to many constructions in the theory.
Example I.d.2. The space J: A Banach space with a Schauder basis whose canonical
image is of codimension 1 in its second dual J** and is also isometric to J**.
Observe that, in spite of the fact that J is isometric to J**, the space J is not
reflexive. Since J** is separable it cannot have a subspace isomorphic to Co or 11'
By 1.c.13, J is not isomorphic to a subspace of a space with an unconditional basis.
The space J consists of all sequences of scalars x=(al, az, ... , an,. .. ) for which
(i)
and
(ii) lim an =0.

n
The supremum in (i) is taken over all choices of m and PI <pz < ... <Pm. It is
useful to note that
is an equivalent norm on J (the special form of 11·11 is of importance only for

proving that J** is isometric and not merely isomorphic to J).
It is easy to verify that J is a Banach space and that the unit vectors {e n }:=1
form a monotone basis with respect to both norms. The vectors {el + ez + '" + en}:= 1
have all norm 1 but they have no w limit point in J; therefore, J is not reflexive.
The unit vector basis is a shrinking basis 'of J. Indeed, assume that for some
x* E J*, some e> 0 and a normalized block basis {Uk}k'=1 of {en}:=1 we have
CD
X*(Uk) ~ e, for all k. It is easily verified (by using 111·111) that L uk/k converges in J.
k=1
00
Since L x*(uk)/k does not converge we arrived at a contradiction. By l.b.2, the

k=1
space J** consists of all sequences (al> a 2 , .. ·, an, ... ) for which s~p Ilj~l aieill <00, i.e.
all sequences for which (i) in the definition of J holds. Since (i) implies the existence
of lim an we infer that J** is the linear span of J (or, more precisely, of the canon-
n
ical image of J in J**) and the functional xt* defined by xt*(e;D= 1 for all n (i.e.
the functional which corresponds to the sequence (1, 1, 1, ... )). The map U: J** - J
defined by
UX**=(-A, x**(et}-,\, x**(e:)-,\, ... ),

where ,\ = li;n x**(e;), is an isometry of J** onto J. (Recall that the norm in J**
is given by Ilx**11 = s~p IIJl x**(et)ei//.)
It is noteworthy to remark that the basis {e;}:=l of J* has the property that
//Jl cne;11 = n~l cn, whenever all the Cn are non-negative. Since J* does not contain
a subspace isomorphic to 11 no subsequence of {e;}:=l can be unconditional.
A useful variant of l.d.2 shows that not only the one-dimensional space but
every separable Banach space X can be realized as Z**/Z for a suitable Z. This
was proved in [55] under some restrictions on X and in the general form in [85].
Theorem l.d.3. Let X be a separable Banach space. Then there is a separable Banach
space Z such that Z** has a boundedly complete basis and Z**/Z is isomorphic to X.
For the proof of l.d.3 we refer to [85]; we just mention here the definition of
Z. Let {xn }:= 1 be a sequence which is dense on the boundary of the unit ball of X.
The space Z consists of all the sequences Z= (al, a2, ... ) of scalars for which
and
(ii) L'" atxi=O.

,=1
The supremum in (i) is taken over all choices of integers m and 0 = Po < Pl < ... < Pm.
The main point in the proof is to show that Z** can be identified with the space of
all sequences (aI' a 2, ... ) for which condition (i) above holds. Observe that con-
'" aiXi converges. The operator Tfrom Z** to X defined by
dition (i) implies that L
i=l
00
T(al' a2, ... )= L

i= 1
aixi is bounded and, by the density of {Xi}i'=l, it is actually a
quotient map. The kernel of T is, by definition, the canonical image of Z in Z**
and thus Z**/Z is isomorphic to X. From the description of Z** it is obvious that
the unit vectors {en}:=l form a boundedly complete basis of Z**. From this fact
and l.b.6 it follows that Z has a shrinking basis.
In view of l.b.2 it is possible to phrase l.d.3 as follows. Let X be a Banach
space. Then, there is a norm 11·11 on the space of the sequences of scalars which are
eventually 0 so that X is isometric to Zl/Z2, where Zl consists of all sequences
(aI' a2,· .. ) such that Sl~p IIJI aieill <00 while Z2 consists of all those sequences for
which {Jl aie}~=l is a Cauchy sequence.

Another construction of a space Z, having the properties required in l.d.3,
is given in [24]. This paper contains a construction of spaces Z satisfying Z**/Z;.::: X
also for a large natural class of non-separable spaces X. It seems to be unknown

whether an arbitrary non-separable X can be represented as Z** /Z, for a suitable Z.
Every space Z obtained in l.d.3 is non-reflexive but has a separable second dual.
Thus, by l.c.13, it is not isomorphic to a subspace of a space with an unconditional
basis. In all examples considered thus far the non-reflexivity of the space played
a crucial role in establishing the non-existence of an unconditional basis.
The first example of a separable reflexive space which cannot be embedded in a
space with an unconditional basis was given in [78]. Later it was shown in [88]
that no space with an unconditional basis can contain isometric copies of LiO, 1),
with p arbitrarily close to 1 or to 00 (we shall present the proof of this fact in Vol.
II). Thus, for every sequence {Pn};;'= 1 of numbers so that 1 <Pn <00 for every nand
either inf Pn = 1 or sup Pn =00, the reflexive space
n n
(I
n= 1
EB Lpn(O, 1)) is not iso-
2
morphic to a subspace of a space with an unconditional basis.
The previous examples cannot be embedded isomorphically into a Banach
space having an unconditional basis. There are also examples of spaces which fail
to have an unconditional basis but do embed into a space having such a basis. The
first and perhaps simplest example of this kind (cf. [82]) is the subspace D of 11>
spanned by xn=en-(e2n+e2n+l)/2, n=l, 2, ... , where {e n};;'=l are the unit vectors
in II. The sequence {X n };;'=1 forms a monotone basis of D. The fact that D does not
have an unconditional basis follows from l.c.12 since D is weakly sequentially
complete without being isomorphic to a conjugate space (this latter fact will be
proved in Vol. IV).
More complicated but also more interesting examples are obtained by using
Enflo's solution to the basis problem (and its modification in [20] and [40]); there
exist subspaces of Co and lp, 2<p<oo, which fail even to have a basis. We shall
discuss this in detail in the next chapter.
We mention now two questions related to the existence of unconditional bases
which are still open.
Problem l.d.4. Let X be a Banach space with an unconditional basis and let Y be a
complemented subspace of X. Does Y have an unconditional basis?
The second question asks whether Banach's theorem on the existence of basic
sequences can be strengthened in the following sense.
Problem l.d.S. Does every infinite dimensional Banach space X contain an uncon-
ditional basic sequence?
The answer to l.d.5 is positive whenever X is isomorphic to a subspace of a

space with an unconditional basis (use l.a.l1), or more generally, to a subspace of
a nice Banach lattice (this will be made precise and proved in Vol. II). In a Banach
space which is a subspace of a space with an unconditional basis an unconditional
basic sequence can be obtained from every sequence of vectors {U n };;'=1 with
Ilunll = 1 and Un ~ 0, by passing to a subsequence. A natural approach to 1.d.5 is
therefore to investigate whether in an arbitrary Banach space a sequence {U n };;'=l,
for which Ilunll = 1 for all n and Un ~ 0, must have an unconditional basic sub-
sequence. B. Maurey and H. P. Rosenthal [104] proved recently that this is not the
case.
Example 1.d.6. There is a Banach space having a Schauder basis {en}~=1 which
converges weakly to 0 but which has no unconditional subsequence.
Let 1 >8>0 and let M={mn}~=1 be an increasing sequence of integers with

m1 =1 so that
(*)
We shall construct a collection Ll of sequences 8={Un}~=1 of finite subsets Un

of the integers N whose main properties are:
1. If 8={Un}~=1 E Ll then, for every n, the largest integer in Un is smaller than
the smallest integer in Un + 1 (i.e. max Un < min Un + 1)'
2. For8={un}~=1 ELl the cardinalities an of the sets Un satisfy 1 =<11 <0'2 <0'3 < ...
and an E M, n= 1,2, ....
3. If, for a pair of sequences 81={u*}~=1 and 82={u~}~=1 in Ll, we have a}+l =
o'~ + 10 for some integers j and k, then necessarily j = k and ut = up for 1::::; i::::; k.
4. For every infinite subset N1 of the integers there is a 8={Un}~=1 in Ll so that
Un C N1 for all n.
The family Ll is defined as follows: Pick a one to one mapping if; from the
set of all finite subsets of the integers into M which satisfies if;(u»O', for every
00, and take as Ll the family of all 8={Un}~=1 which satisfy 1 above and for which
0'1 = 1 and an+1 =if;(u1 U 002 U ... U Un), n= 1,2, .... With this definition of Ll it is
obvious that 1, 2 and 4 hold. Let us verify that also 3 holds. We observe first that
the growth condition (*) on {mn}~=1 is such that the following is true. If
for some choice of n1 < n2 < ... < nk and S1 < S2 < ... < s" then k = hand 1l; = Sf for
1::::; i::::;k. Assume now that, as in 3, iij+1 =a~+l' By the definition of Ll and the fact
j k j k
that if; is one to one it follows that U ut = U uf and hence, L at = L ap. By 2
1=1 1=1 1=1 1=1
ur
and the remark made above on M it follows that j = k and at = for 1::::; i::::; k.
From this and the fact that if; is one to one we get that ut = uf for 1 ::::; i::::; k, as desired.
We pass now to the construction of the space E of Maurey and Rosenthal. It
is the completion of the space of sequences x=(a1o a2,"') of scalars, which are
eventually 0, with respect to the norm
where the supremum is taken over all sequences 8 = {un}~= 1 in Ll. It is clear that the
unit vectors {en }:=l form a monotone basis of E. Let a= {aj}i'= 1 E LI and put
uj{j)=(2: ei)/a}/2,J= 1,2, .... Let {Cj}i'=l be scalars and let n be an integer. By
IE "1
using in the definition of 11·11 elements 1)k={~}i'=lELI, l~k~n, with TJ=aj if
I ~j~ k and min T~ + 1 > max am it follows that
It follows from (*) and condition 3 on LI, by an easy computation, that
Hence, {Uj{j)}i'=l is equivalent to the summing basis in c. Consequently, by 4, every

subsequence of {e n}:= 1 has a block basis which is equivalent to the summing basis
and thus no subsequence of {e n }:=1 can be unconditional.
It remains to show that {e n}:=1 tends weakly to O. This follows from the fact
that, for every increasing sequence {nk}k."=1 of integers, we have
lim
k-co
IIe +en2 + ···+enkll/k=O.
n1
k
Indeed, by 2 and the definition of 11·11, we obtain that II en1 + en2 + ... + enkll ~ 2: at>
1= 1
where aj = mi 1/2 whenever

j-l
2: mh < i ~ h=1 j
2: mho Obviously, lim (k )
2: aj /k = O.
h=1 k j=1
The space E described here is not reflexive. In Vol. II we shaH show how to
modify the construction of E in order to get even a uniformly convex space with a
basis which tends weakly to 0 but has no unconditional sUbsequence. Thus, there
is also a uniformly convex space which is not isomorphic to a subspace of a space
with an unconditional basis.
e. The Approximation Property
A result which goes back to the beginnings of functional analysis asserts that the
compact operators on a Hilbert space are exactly those operators which are limits
in norm of operators of finite rank. One part of this assertion, namely that every
TEL(X, Y) for which IIT-Tnll-""O for suitable {Tn}:=1 EL(X, Y) with
dim TnX <w is compact, is trivially true for every pair of Banach spaces X and Y.
It was realized long ago that the converse assertion is also true for many examples
of spaces X and Y besides Hilbert spaces. For example, if Y has a Schauder basis
{Yn}:=1 then, for every compact TEL(X, Y), IIT-PnTII-""O, where the {Pn}:=l
are the projections associated to the basis {Yn}:= 1. The question whether the con-
verse assertion is true for arbitrary.Banach spaces X and Y (which was called for
obvious reasons the approximation problem) was open for a long time. This
problem was solved (in the negative) by P. Enflo [37]. The observation above shows
that this solution provides also a negative solution to the basis problem. We
shall present Enflo's solution or, more precisely, a simplified version of it, due to
Davie [20], in the next chapter. Our purpose in this section is to investigate those
Banach spaces which have one of the many variants of the "approximation
property", i.e. those spaces X for which every compact TinL(X, Y), or L(Y, X),
(Yarbitrary) is a limit in norm of a suitable sequence of finite rank operators. The
investigation of the various variants of the approximation property and the relations
between them was initiated by Grothendieck [48]. Many of the results presented
here are taken from Grothendieck's memoir. The real impetus to the investigation
of the approximation properties was however given by Enflo's result which en-
sured that the class of spaces which fail to have any of the approximation properties
is not void. Much progress has been done in this study but, generally speaking,
the situation is still very far from being clear. We shall mention below several of
the many natural open problems concerning the approximation property. What is
clear by now is that there are many examples of spaces which fail to have the
approximation property even among spaces which are "nice" in other respects
(there is, for example, a Banach lattice which fails to have the approximation
property [138]. We shall present this example in Vol. II). It is also clear that the
study of the approximation property is important in many contexts in Banach
space theory and even in some areas of analysis outside the framework of this
theory.
Definition l.e.I. A Banach space X is said to have the approximation property

(A.P. in short) if, for every compact set K in X and every e > 0, there is an operator
n
T: X -»- X of finite rank (i.e. Tx= 2: Xt(X)Xb for some {X;}f=l c X and {Xt}f=l c X*)
1=1
so that IITx-xlI::::;e, for every x E K.
Every space with a Schauder basis has the A.P. Indeed, for every compact K
and every e> 0, we have II P nX - xii < e for every x E K provided n > n(e, K), where
{Pn};'=l are the projections associated to the basis.
In order to study the A.P. we need two general facts-one concerns the structure
of compact sets in a Banach space and the other gives a concrete representation
of the dual of some space of operators.
Proposition l.e.2. A closed subset K of a Banach space X is compact if and only if

there is a sequence {Xn};'=l in X such that Ilxnll-»- 0 and Kcconv {Xn};'=l'
Proof It is easily checked that if Ilxnll -»- 0 then t~ 1 AnXn; An ~ 0, J An::::; I} is

1
compact and coincides with conv {Xn };'=l' This proves the "if" part. We prove
now the "only if" part. Let K be compact. Let {Xi, 1}f,; 1 be a finite set of elements of
nl
X so that 2Kc U B(x! ' h
1=1
1/4), Put
U {(B(Xi,l, 1/4) n
nl
K2 = 2K)-XI,l}'
1=1
Then K2 is a compact subset of B(O, 1/4). Pick next {X;.2m1 in B(O, 1/2) so that
n2
2K2c U B(Xi ' 2, 1/42) and put
1=1
U
n2
K3= {(B(Xi,2' 1/42) n 2K2)-Xi ,2}'
i= 1
We continue the inductive construction of {Xj,j}r~l,j= 1,2,3, ... in an obvious way.

For every x E K there is an 1 ~ i1 ~n1 so that 2X-XiI ,l E K 2; hence an 1 ~ i2~n2
so that 4x - 2XiI , 1 - X I2 , 2 E K3 and, in general,
It follows that x E conv {xi,j; 1 ~ i~nj,j= 1,2, ... }. Since IlxI,jl1 ~ 2·4 -H\ for j> 1
and every i~nj, our assertion is proved. 0
Proposition 1.e.3. Let X and Y be Banach spaces and put on L(X, Y) the topology r
of uniform convergence on compact sets in X (this is the locally convex topology
generated by the seminorms of the form IITIIK = sup {IITxll, x E K}, where K ranges
over the compact subsets of X). Then, the continuous linear functionals on (L(X, Y), r)
consist of all functionals rp of the form
2: YT(Tx;),
00
2:
00
rp(T) =
1=1
{Xi}I";"l cX, {yn";"l c Y*,
i= 1
IIx1111lytii <00.
°
Proof Assume that rp has such a representation. We may clearly assume that
Xi # for every i. Let bi}t= 1 be a sequence of positive scalars tending to 00 so that
00
2: 1)iII XiIIIIYTII =
j=l
C<OO. Put K={x;/llxilhj}~l U {O}. Then K is compact and
2: IIYTllllxill1)iIIT(x;/llx;lhj)ll~ ClITIIK'
00
Irp(T)1 ~
j= 1
Conversely, assume that rp is a linear functional on L(X, Y) so that Irp(T) I~ C IITIIK'

for some constant C and some compact set Kc X. By l.e.2 we may assume without
loss of generality that K=conv{xn}~=l' where Ilxnll--+O. Let S:L(X, Y)->
(YEB YEB "-)0 be defined by S(T) = (TX1' TX2'''')' Since Irp(T)I~ClIS(T)11 itfol-
lows that there is a linear functional if; defined on the closure of SL(X, Y) so that
rp(T)=if;(S(T)). By the Hahn-Banach theorem we may extend if; to a continuous
linear functional on (Y EB Y EB .. ')0, i.e. to an element of (y* EB y* EB .. ')1' In
other words there exist {y;D:=l in y* so that ~ Ily;11 <00 and rp(T) = ~ y;T(xn)'
n=l n=l
o
The next theorem, due to Grothendieck [48], clarifies the relation between the
approximation property and the question of approximating compact operators,
with which we started this section.
Theorem 1.e.4. Let X be a Banach space. The following five assertions are equivalent.
(i) X has the approximation property.
(ii) For every Banach space Y the finite rank operators are dense in L( Y, X),
in the topology T of uniform convergence on compact sets.
(iii) For every Banach space Y the finite rank operators are dense in L(X, Y),
in the topology T of uniform convergence on compact sets.
(iv) For every choice of {Xn};;"=lcX, {X~};;"=lCX* such that I Ilx~llllxnll <00
n=l
and i:
n=l
x~(x)xn=O,for all x E X, we have i:
n=l x~(xn)=O.
(v) For every Banach space Y, every compact TEL(Y, X) and every e>O there
is a finite rank operator Tl E L( Y, X) with liT - Tlil < e.
Proof The equivalence of (i) and (iv) is a consequence of l.e.3. Indeed, by defini-
tion, (i) means that the identity operator is in the T closure of the space of finite
rank operators in L(X, X). This happens if and only if every T continuous linear
functionl;ll rp on L(X, X), which vanishes on operators of rank 1, vanishes also on
the identity operator. By l.e.3 this is exactly what (iv) means.
It is clear that (ii) or (iii) (with Y = X) imply (i). We shall show that (i) implies
(ii) and (iii). Let T E L( Y, X). For every compact set Kc Y the set TK is compact
in X. Hence, given e> 0, we have by (i) a finite rank operator Tl on X so that
IITlTy - Tyll : s; e, for y E K. Since TlT is of finite rank we proved (ii). Let now
Ole- T E L( X, Y), let K be a compact set in X and let e > O. By (i) there is a finite
rank operator Tl on X so that IITlx-xll ::S;e/IITII, for x E K. Then, IITTlx-Txll::S;e
for x E K and this proves (iii).
It remains to prove the equivalence of (i) and (v). Assume that (i) holds and let
TEL(Y, X) be a compact operator. The set K=TBy(O, 1) is compact and hence,
for every e > 0, there is a finite rank operator Tl on X so that IITlx - xii : s; efor
XEK. Then, IITlT-TII::S;e and thus (v) holds.
Assume that (v) holds and let K be a compact subset of X and e>O. By l.e.2
we may assume without loss of generality that K=conv {Xn};;"=l, with Ilxnll t 0 and
Ilxlll::s; 1. Put U =conv {± xn/llxnlll'2};;"=l' Clearly, U is a compact convex set in X
which is symmetric with respect to the origin. Let Y be the linear span of U in X,
00
i.e. Y = U n U,
n=l
and introduce in Ya norm 111·111 which makes U its unit ball (i.e.
Illylll =inf {A>O; ylll E U}). A routine argument shows that (Y, 111·111) is a Banach
space (in particular, it is complete). The formal identity map from Y to X is com-
pact and hence, by (v), there are {y{ }rl= 1 C y* and {udr'= 1 C X so that
Ilit YT(X)U; - xll::s; e12, for every x E U, and hence x E K. The {YnV~ 1 are con-
tinuous with respect to 111·111 but need not be continuous with respect to 11·11 (and
thus are not in general restrictions of elements of X* to Y). In order to conclude
the proof it is enough to verify the following statement. Given any y* E y* and
I» 0 (in our case we take D= el2m· max Iluill) there is an x* E X* such that
i
ly*(x)-x*(x)1 < Dfor x E K i.e. ly*(xn)-x*(xn)1 < D, for every n.
Observe that, since xn/llxnlll'2 E U, we have Illxnlll::S; Ilxn I1 1'2, for every n,
and thus IIIXnlll--7- O. For n~no we have therefore Iy*(Xn) I< 8/2. Put Ko=
28- 1 conv {±Xn};;'=no+l (notice that the closures in 11·11 and 111·111 are the same) and
F={x; x E span {Xn}~'?",I' y*(X) = I}. Then, F is 11·11 closed, Ko is 11·11 compact and
Ko (") F= 0. By the geometric version of the Hahn-Banach theorem there is a
11·11 closed hyperplane F in X so that Fe F and F (") Ko = 0. Let x* E X* be such
that F={x; x*(x) = I}. Then, X*(Xn) = y*(xn ) for n~no and Ix*(xn)1 < 8/2 for n > no·
Consequently, Ix*(x n) - y*(Xn) I< 8 for every n, as desired. 0
In view of (ii), (iii) and (v) of 1.eA it is natural to ask what is the situation if
we reverse the roles of X and Y in (v). The answer is given by the following result
which is also due to Grothendieck [48].
Theorem t.e.S. Let X be a Banach space. Then, X* has the approximation property
if and only if, for every Banach space Y, every e > 0 and every compact T E L( X, Y),
there is a finite rank operator Tl E L( X, Y) such that II T - TIll ~ e.
Proof Assume that, for every Y, every compact T E L(X, Y) can be approximated
as above. Let Z be any Banach space, let T E L(Z, X*) be compact and let e > O.
By considering the compact operator ni: X --7- Z* it follows from our assumption
that there are {X!}f=1 e X*, {Zt}r=1 eZ* so that, for every x with Ilxll ~ 1,
Hence, for every z E Z with Ilzll ~ 1,
i.e.II Tz - Jl ztcz)xtll ~ e. By 1.e.4, X* has the A.P.

Assume, conversely, that X* has the A.P. Let T E L(X, Y) be compact and
1/2> e > O. The operator T*: y* --7- X* is also compact and hence there are
{yt*}f=1 in y** and {X!}f=1 in X* so that IIT*Y*-i~1 yt*(Y*)x!lI~e, whenever
Ily*II~1. It follows that IITx-JI xt(x)Y!*II~e, whenever Ilxll~1. This does not
conclude the proof since the {Y!*}f=1 are not necessarily contained in Y. We have
to "push" the {yt*}f=1 into Y. This is done by using the following lemma.
Lemma t.e.6 [90]. Let X be a Banach space, let D be a finite-dimensional subspace

of X** and let e> O. Then, there is an operator S: D --7- X such that liS II ~ 1 + e and
SID()X is the identity.
Let us first see how l.e.6 is used to conclude the proof of 1.e.5. By the com-
pactness of T there are {Xj}T= 1 in the unit ball Bx(O, 1) of X so that TBx(O,l)e
m
U By(Txj' e).
;=1
Apply l.e.6 to D=span {TXj}T=1 u {yt*}f=I' We claim that
\\TX- Jl xNX)SYt*\\~4e, for every x E Bx(O, 1). Indeed, fix x E Bx(O, 1) and pick a
j such that IITxj-Txll~e. Then, \\TXj-Jl xt(X)Yt*\\~2e. By applying S we get
that \\TXj-Jl XNX)SYt*I\~2e(l+e)~3e and this proves our assertion and con-
cludes the proof of 1.e.5. 0
Lemma 1.e.6 is a special instance of a result which plays a central role in the
local theory of Banach spaces and which will be discussed in detail in Vol. IV. For
the sake of completeness we present its proof (cf. [29]) also here.
Proof of I.e.6. First we notice that L(l~, X**) =L(lr, X)**. This follows from the
fact that the correspondence T_ {Yt=Tei}f=l (where {eaf=l are the unit vectors of
n times
m is an isometry from L(n, X) onto (X EEl ••• EEl X) co.
Let I: D - X** be the identity mapping and let e> 0. There exist an n and vec-
tors {Uj}7 =1 in D of norm ~ I + e so that conv {uj}f =1 ::::> B D(O, I). Hence, there is an
operator' V:lr -D, for which 11V11~I+e and VBI~(O, 1)::::>BD(O, 1). Since
IV EL(lr, X)** there is a net {S,,}cL(I~, X) with IIS"II~ IIIVII ~ 1+e for all a and
{Sa} converges to IV in the w* topology of L(l~, X)**. Any pair e E n, x* E X*
defines a functional (e, x*) EL(I~, X)*, by setting (e, X*)(S) = x*Se. This implies
that, for every e E n, S"e ~ IVe. Put B= {e En; Ve E D n X} and observe that
Sae ~ IVe for all e E B. Thus, by taking a suitable convex combination of S,,'s
and by using a standard perturbation argument, we can construct an operator
T: l~ - X such that T1B=IV1B and IITII < 1+2e. If D 3 d= VVl = VV2, for some
Vh V2 E lr, then Vl-V2 E B and therefore TVl = TV2' Hence, by setting Sd=Tv,
lr
where v E is any vector satisfying Vv = d, we define an operator S E L(D, X) for
which IISII < 1+2e and SID()x=IID()x, This completes the proof. 0
The relation between the properties appearing in 1.e.4 and 1.e.5 is clarified in
the following result.
Theorem l.e.7. (a) Let X be a Banach space. If X* has the AP. then X has the
AP. In particular, if X is reflexive then X has the AP. if and only if X* has the
AP.
(b) There is a separable Banach space having a Schauder basis whose dual is
separable but fails to have the AP.
Proof Assertion (a) follows immediately from the equivalence (i) <? (iv) of 1.e.4.
co co
l~{xn};=lcXand{x;n;=lcX*aresuchthat 2: Ilxnll Ilx;11 <00 and n=l
n=l
2: x;(x)xn=O,
co
for every x E X, then also 2:
x*(xn)x;=O, for every x* E X*. In other words
n=l
~ JXn(x*)x;=O, where J: X - X** is the canonical embedding. Since X* has
n=l
the AP. ~ Jxn(x;)= ~ X;(X n ) =0.
n=l n=l
The proof of assertion (b) uses of course the fact that there is a Banach space
which fails to have the A.P. (this will be proved in Section 2.d below). Let X be a
separable Banach space which does not have the A.P. By l.d.3 there is a space Z
so that Z** has a basis and Z**/Z is isomorphic to X. By passing to the duals we
get that Z***~Z* EEl X* (observe that for every Banach space Z there is a pro-
jection of norm 1 from Z*** onto Z*. The projection is the map which assigns to
every functional on Z** its restriction to Z). Since X fails to have the A.P. the
same is true for X*, by assertion (a). It is trivial to verify that a complemented
subspace of a space having the A.P. has also the A.P. Hence Z*** (which is a dual
of a space Z** with a basis) fails to have the A.P. If X is such that X* is separable
(e.g. if X is isomorphic to a subspace of co; see the remark following l.e.S) then
Z*** is separable. 0
While investigating the approximation problem Grothendieck found many

nice equivalent formulations of this problem. We give two of those formulations
here. As stated, this result is of course only of historical interest. However, from a
different point of view it is still useful since it shows the connection between the
A.P. and some problems arising in classical analysis. Moreover, the simple and
explicit proof of the result enables us to transfer each counterexample to one of the
versions of the approximation problem to a counterexample to the others.
Proposition l.e.S. The following three assertions are equivalent.

(i) Every Banach space has the A.P.
(ii) Every matrix A=(ai,N:i=l of scalars, for which lim ai,i=O, i= 1,2, ... ,
i
.L max lai,,1 <00 and A2=0, satisfies trace A= n=l

.L ann=O.
00 00
1=1 ,
(iii) Every continuous function K(s, t) on [0, 1) x [0, 1], for which
1 1
JK(s, t)K(t, u) dt=O for every sand u, satisfies
o
J K(t, t) dt=O.
0
Prooj. (ii) => (i). Let X be a Banach space which fails to have the A.P. Then, by
l.eA, there are {Xn}:'=l eX, {X;}:'=l c X* with i
n=l
Ilxnllllx;11 <00 and
n=l
i x;(x)xn=O
for every x E X but

n=l
i x;(xn) =f 0. There is clearly no loss of generality to assume
that Ilxnll-» and

00
° n=l
i Ilx;11 <00. The matrix A = (xt(X,))l;'i=l satisfies lim Xt(Xj) =0,
1
L max Ixi(Xj)1 <00 and also A2 =0 (the entries of A2 are expressions of the form
j= 1 1
00 00
L X;(Xi)X~(Xn)). However, trace A= n=l

n=l
L x;(xn)=fO.
(iii) => (ii). Assume that (ii) fails and that A = (ai, j) is a counterexample. Put
ai = max lai, 11 and choose a sequence of positive numbers {7J;}f;,1 such that 7Ji -»00
i
00
and L ai7Ji < 1.

i= 1
Put hi j=ai
"
l/a j7Ji;then lim hi ,=0 (in the sense that, for every
i, i '
e>O, there are only finitely many pairs (i,j) for which Ibd>e). Let 1=t1>s1>
t2 >S2'" be numbers such that lim tn =0 and t;-Si > CX(1Ji> for every i (this is possible
n
co
since L: CXiT]. < 1). Let {'Pi}t=1 be continuous functions on [0, 1] such that 0:::;; 'Pi:::;; 1,
i=1
'Pi vanishes outside [s" til and jf 'PI(t) dt=cxIT]I'
Sf
It is easy to verify that K(s, t)=
~ ~ V'Pi(S)'P;(t) bl ,1 is continuous on [0, 1] x [0, 1] (at each point the sum con-
1=11=1
1
sists of at most one summand). The fact that A2=0 implies thatf K(s, t)K(t, u)dt=O
for every sand u while °
1 co 1 co co
L f 'Pn(t) dt·bn,n= n=1
f K(t, t) dt= n=10
o
L cxnT]nbnn= n=1
L an,n#O.
(i) =:> (iii). Assume that there is a counterexample K(s, t) to (iii). For every
s E [0, 1]. letls(t) E ceO, 1) be defined by ls(t)=K(s, t). The continuity of K(s, t)
implies that Ko={Is}o';S';1 is a compact subset of C(O, 1). We claim that
Xo= span Ko is a subspace of C(O, 1) which does not have the A.P. Indeed, assume
that e> °is such that there is an operator T of finite rank Tg=
n
L: xf(g)fi on
1=1
Xo
such that IITg-gll < e for all g E Ko. Since span Ko is dense in Xo there is no loss
of generality to assume (by increasing n) that It E Ko for all i, i.e. It = Isf' for suitable
S; E [0, 1]. The functionals xt extend to functionals on C(O, 1) and can thus be
considered as measures on [0, 1]. Since the measures with finite support are w*
dense in the set of all measures it follows easily that we may assume that each xT
is a finitely supported measure and thus (by increasing n again) that each xf is of
the form xt(g)=A;g(uI)' for suitable U; E [0,1] and scalars {A;}f=1' In conclusion
we get that, for every s E [0, 1], t E [0, 1],
and, in particular,
1
This however is impossible for small enough e since f K(t, t) dt#O while
o
1
f K(s;, t)K(t, Ut) dt=O, for every i. 0
o
It is also instructive to note that it is very simple to verify directly that (i) =:> (ii)
in 1.e.8. Indeed, let A = (at, 1) be a matrix which fails (ii) and put x; =(at,l' a;,2, ... )ECo,
for i= 1,2, .... Then, the closed linear span X of {Xt}!':1 in Co is a space which fails
the AP. (apply 1.eA to the vectors {Xi}r;,1 E X and {e i }r;,1 E X* where ei is the
restriction to X of the i'th unit vector in II)' Thus, the analysis of Grothendieck of
the approximation property shows that if there is a Banach space which does not
have the AP. then there is also a subspace of Co which does not have the AP. In
the next chapter we shall show that not only Co but also the sequence spaces lp,
for p > 2, have subspaces which fail to have the AP.
We mention now some of the open problems concerning the AP.
Problem l.e.9. Let X be a Banach space such that every compact T:X-,;-X is a limit
in norm offinite rank operators from X into itself. Does X have the AP. ?
Problem l.e.IO. Does the space L(l2, 12 ), of all bounded operators on 12 with the
usual operator norm, have the A.P. ? Does the space H ro, of all the bounded analytic
functions on {z; Izl < I} with the supremum norm, have the AP.?
Observe that the two concrete spaces appearing in 1.e.lO, for which the AP.
has not yet been verified, are non-separable. The common separable spaces which
appear in analysis have the AP. and, as a matter of fact, as mentioned already in
Section a they even have a Schauder basis. We would like to point out that it is
usually much easier to verify that a given space has the AP. than to construct a
basis in this space. Let us illustrate this by considering the disc algebra A. As
mentioned in Section a it is known by now that A has a basis. However, it is not
easy to construct such a basis (and its existence was open for a long time). On the
other hand it is very easy to verify that A has the AP. Indeed, for f (x) = ao +
alz+a2z2+ ... EA put Snf=aO+alz+", +anzn and U n f=(Slf+S2f+'" +Snf)/n,
n = 1,2, .... The classical result of Fejer states that II Un II ~ 1 for all n and that
lIunf-fll-,;- 0, for every fE A. This shows that the disc algebra has the AP. (even
the M.AP. defined below).
In the definition I.e.I of the AP. we imposed no requirement on the norm of
the operator T. For a Banach space with a basis the operator T required in I.e.I
can be chosen to be bounded by a constant independent of the compact set K
(namely, by the basis constant). We shall study now this stronger version of the AP.
Definition l.e.H. Let X be a Banach space and let 1 ~ A<00. We say that X has the
A-approximation property (A-A.P. in short) if, for every 8> 0 and every compact
set K in X, there is a finite rank operator T in X so that IITx-xll ~8, for every
x E K, and IITII ~ A. A Banach space is said to have the bounded approximation
property (B.AP. in short) if it has the A-A.P., for some A. A Banach space is said
to have the metric approximation property (M.AP. in short) if it has the I-AP.
Observe that in I.e.II it is enough to take instead of a general compact set K

n
finite sets only. Indeed, given K and 8 we find {Xi}f=1 so that Kc U
j= 1
B(x;, 8/3,1.). If
IITII~A and IITxj-Xill~8/3, for every i, then IITx-xlI~8, for every XEK (in
I.e.l it is of course essential that K is infinite; for finite sets KaT satisfying the
requirements in I.e.1 always exists trivially).
As we already observed, a space with a basis has the B.A.P. (and a space with a
monotone basis has the M.A.P.). It is not known whether the converse is true.
Problem 1.e.12. Does there exist a separable Banach space which has the B.A.P.
but fails to have a basis?
It is likely that the answer to 1.e.12 is negative and that the "right" relation
between bases and the B.A.P. is the one given by the following result of [118] and
[61].
Theorem 1.e.13. A separable Banach space X has the B.A.P. if and only if X is
isomorphic to a complemented subspace of a space with a basis.
Proof [118]. The "if" part is trivial and so we have merely to prove the "only if"
part. We start by making two observations.
1. Assume that X is separable and has the A-A.P. Then there exists a sequence
00
of finite rank operators {Sn};'=l on X so that X= L Snx, for every x E X, and

n=l
IIJl Sill::;; A, for every n. Indeed, let {y!}~ 1 be a dense sequence in X. There exist,
for n=I, 2, ... , Tn EL(X, X) of finite rank such that !!Tn!!::;;A and !!TnYi-yd!::;;n- 1
for I::;;i::;;n. The operators {Sn};'=l defined by Sl=T1 and Sn=Tn-Tn-l' for n> 1,
have the desired property.
2. Let B be a Banach space with dim B=n. Then, there are operators {Ui}f:l in
L(B, B) such that dim Uk B=I, IIJl Uill::;;2, for every l::;;k::;;n 2, and !~l Uix=x for
every x E B. Indeed, let {Xj}J=l and {Xj}7=1 be an Auerbach system for X (see
1.c.3). For i=rn+j, O::;;r<n, 1 ::;;j::;;n put Uix=xj(x)xj/n. Then, for every k=rn+j,
we get
(/ denotes the identity operator on B).

Let now X be a separable space having the B.A.P. and choose {Sn};'=l as in
observation 1. Since every space SnX is finite dimensional we can construct, for each
n, operators {Ui.n}f'~l on SnX, as in observation 2, where mn = (dim SnX)2. Put
Vj=Ui,nSn if j=ml+m2+"'+mn-l+i, I::;;i::;;m n,n=I,2, .... Then, for every
x E X, X= jt Itt
Vjx, vjll::;; 5A for every k and dim VjX = 1 for every j. Let Vj be
a vector of norm 1 in VjX,j= 1,2, .... Let Ybe the space consisting of all sequences
co
of scalars y=(al> a2,"') such that L:
j=l
ajVj converges and put
The unit vectors form clearly a monotone basis of Y. Define V: X ~ Y by

Vx=(al> a2, ... ), where aj is the scalar determined by Vjx=ajvj,j= 1,2, .... Clearly,
IIVII ~ SA and 11V- 1 11 ~ 1, i.e. V is an isomorphism (into). Let U: Y ~ X be the
'" ajVf. Clearly, UV is the identity operator of
operator defined by U(al' a2,' .. )= 2
j=1
X and hence VX is complemented in Y. D
A part of Theorem LeA can be generalized to the setting of B.A.P. or M.A.P.

without any change in the proof. We state this for the M.A.P.
Proposition l.e.14. Let X be a Banach space. The following four assertions are
equivalent.
(i) X has the M.A.P.
(ii) For every Banach space Y the finite rank operators of norm ~ 1 are dense
in the unit ball of L( Y, X) in the topology T.
(iii) For every Banach space Y the finite rank operators of norm ~ 1 are dense
in the unit ball of L(X, Y) in the topology T.
co
(iv) For every choice of {X n}::'=1 c X, {X;n::'=1 c X* such that 2 Ilxnllllx:11 <00
n=1
and IJI X:(TXn)l~ IITII, for every operator T of finite rank in L(X, X),
we have In~1 X:(Xn)l~ 1.

Condition (v) of LeA does not generalize to this setting since the Tl given there
satisfies automatically II TIll ~ II Til + e and it can actually be chosen always so that
IITlll=IITII·
We conclude this section with a discussion of the relation between the A.P.
and the M.A.P. Grothendieck [48] proved the surprising result that in many cases
the A.P. implies the M.A.P.
Theorem I.e.15. Let X be a separable space which is isometric to a dual space and
which has the A.P. Then X has the M.A.P.
For the proof of l.e.I5 we need two simple lemmas.
Lemma l.e.16. Let X be separable and let e > O. Then there exists a sequence of
functions {j;}~1on the unit ball Bx of X so that X= 2 '" j;(x), for every x in B x ,
1= 1
co
each j;(x) is of the form j~l XEijX)Xl,h where {E',f}i'=l are disjoint Borel sets of
ro
Bx, {Xi,f}i'=l CBX and.2 11j;11",<I+e where 11j;11", = sup IIj;(x)ll=s~p Ilxdl·
.=1 x 3
Proof We construct the {j;}j;, 1 inductively. Choose first anf1 of the suitable form
so that Ilf111",~1 and Ilx-f1(x)llro~e/2, then an h so that Ilf21Ico~e/2 and
Ilx-f1(x)-f2(x)llco~e/4 and continue in an obvious manner. D
Lemma 1.e.17. Let X = Y*. The space of all operators T of the form
n
(*) Tx= L X(Yi)Xi,
;=1
with {xi }f=l c X and {Yi}f=l c Y,
is 7-dense in the space of all finite rank operators from X into itself.
Proof It is enough to note that every x* E X* = y** is a limit (in the sense of
uniform convergence on compact sets of X) of elements from JYc Y**. D
Proof of 1.e.15. Let X= y* be a space having the A.P. By I.e.14 we have to show
that if cp is a 7-continuous linear functional on L( X, X) such that Icp(T) I~ II Til,
for finite rank operators, then Icp(T)1 ~ IITII, for every T EL(X, X). We shall prove
this in the following manner. For every e > 0 we shall construct a 7 continuous
linear functional !fe on L(X, X) such that !f.(T)=cp(T), for T ofthe form (*), and
for which it will be evident that l!fe(T)I~(I+e)IITII for every TEL(X, X). By
the assumption that X has the A.P. and l.e.17 it follows from this that !fiT) = cp(T),
for all T E L(X, X), and thus Icp(T)1 ~ (1 +e)11 Til for every T. Since e > 0 is arbitrary
this gives the desired result.
For the construction of !fe we use I.e.16. First we let K=Bx x Bx •. This is a
compact metric space if we endow Bx with the w* topology induced by Y and
Bx> with the w* topology induced by X. To every T of the form (*) we assign
a function gT E C(K) by gT(X, x*)=x*(Tx). The special form of Tensures thatgTis
continuous. The map T -'.>- gT is an isometry. By the Hahn-Banach and the Riesz
representation theorems it follows that there is a measure fl- of norm 1 on K so that
cp(T)= f x*(Tx) dfl-,

K
T of the form (*).
Apply now I.e.16. Then for T of the form (*), we get
L f x*T(fj(x»dfl-= L j=l
L x~,iTxi. i
00 00 00
(!) cp(T)=
;=1 K ;=1
where xt i is the functional on X defined by xt;(x) = f x*(x)dfl-. Clearly

Et,l x Bx*
Ilxt ill ~ Ifl-I(E;, i x B x .) and hence ~ Ilxt ill ~ 11fl-11 ~ 1, for every i. Also
j=l
00
2: sup IIx;, jll ~ 1 + e. It follows that the right-hand side of (!), which we denote
i= 1 i
by !feeT), is a 7 continuous functional on L(X, X) which satisfies
L IlxtJllllxdl :;::;(1 +e)IITII·

00
l!fe(T)1 ~ IITII' 0
;,i=l
It follows from I.e.15 that, for separable reflexive spaces, the A.P. implies the
M.A.P. The same is true for nonseparable reflexive spaces. This follows e.g. from
the fact (cf. [83]) that if X is reflexive and Xo is a separable subspace of X then there
is a separable space Z, XocZc X so that there is a projection of norm 1 from X

onto Z.
In general, the A.P. does not imply the B.A.P. and the B.A.P. does not imply
the M.A.P. This was shown by T. Figiel and W. B. Johnson [41]. In order to
present their example it is convenient to use the following variant of the A-A.P.
A Banach space X is said to satisfy the (e, A)-A.P. if, for every finite dimensional
subspace B of X and every 13 > 0, there is a finite rank operator T on X so that
IITx-xll:(e+13)llxll for XEB and IITII:(A+13. It is easily checked that the
A-A.P. is the same property as the (0, A)-A.P. The (e, A)-A.P. implies the N-A.P,
for some N = A/CA, e). More precisely,
Lemma l.e.1S. A Banach space which has the (e, A)-A.P. with O<e< I has also the
(l-e)-lA-A.P.
°
Proof Let 13 > be such that e + 13 < 1 and let Bc X with dim B <00. By our assump-
tion we can find .inductively a sequence {Tn}:~ 1 of finite rank operators of norm
:( A+ 13 on X so that IITI x-xii :(e+ 13)llxll, for x E B, and
IITn+lx-xll:(e+13)llxll, forxEspan{BU Q TiX}'
For n"? 1 let Sn EL(X, X) be defined by (I-Sn)=(I-Tn)(I-Tn_l) ... (I-Tl)' Then.

for x EB, 11(I-Sn)xll:(e+13)nllxll. Also,
Sn =(1 -Tn)(I-Tn_l) ... (I-T2)Tl +(I-Tn).··(I-T3)T2+·· ·+Tn .
Hence, IISnl1 =(A+ 13)«e+13)n-l+(e+ 8)n-2+ ... +(e+ 8) + 1):( (A+ 8)/(1-e- 8).
Since 8 can be taken arbitrarily small this proves our assertion. 0
The main step in the construction of Figiel and Johnson is contained in the next
lemma.
Lemma 1.e.19 Let X be a Banach space and A"? 1 be a number such that X has the
A-A.P. in every equivalent norm. Then, for every e> 0, the dual X* of X has the
(e, A(1 +2e- 1 A))-A.P.
°
Proof. Let B be a finite dimensional subspace of X*, let ?l> and (3 = A+ 8. Let C
be a finite-dimensional subspace of X such that, for every x* EB, Ilx*ll:(
(1 + 8) sup {lx*(x)l; x E c, Ilxll =l}. We fix e>O and introduce a new norm 111·111
on X* by Illx*111 = Ilx*11 +2e- 1(3d(x*, B). Since the unit ball of 111·111 is w* com-
pact this norm is induced by an equivalent norm in X which is also denoted by
111·111·
Since, by our assumption, (X, 111·11 J) has the A-A.P. there is a finite rank operator
T on X so that II Tx-xll:( 811xll, for x E C, and IIITIII :((3. Passing to the dual we
get, for x* E X*
(t) II T*x*11 +2e- 1(3d(T*x*, B):((3(llx*11 +2e- 1(3d(x*, B)) .

It follows that II T*x*11 ~,8(1 +2e- 1,8)llx*1I and hence, II Til ~,8(1 +2e- 1,8). For
x* E B we get from (t) that d(T*x*, B)~ ellx*II/2, i.e. there is a y* E B such that
IIT*x*-y*ll~ellx*II/2. Let y E C be any element of norm 1. Then, IT*x*(y)-
x*(y)1 = Ix*(Ty- y)1 ~ Sllx*lI. Hence,
Ilx*-y*11 ~(1 + S) sup {lx*(y)-y*(y)l, y E C, Ilyll = I}

~(I+S)(S+e/2)lIx*11 ,
and consequently, IIT*x* -x*11 ~ «1 + S)(S+e/2)+e/2)llx*ll. Since S >0 is arbitrary

the lemma is proved. 0
Example l.e.20. There is a separable Banach space X (which has even a separable
dual) such that X has the A.P. but not the B.A.P.
Proof Let Z be a space su~h that Z has the A.P. but Z* fails to have the A.P. and
is separable (see I.e.7(b». By I.e.I8 and I.e.I9 we can find, for every integer n,
an equivalent norm 1I1·lIln on Z so that (Z, III· II In) fails to have the n-A.P. The
space Ct EB (Z, III· II In») 2 = X clearly has the A.P., fails to have the B.A.P. and
has a separable dual. 0
Note that if Z has the B.A.P. and Z* fails to have the A.P. then (Z, III· II In), for
n> 1, is an example of a space which has the B.A.P. but not the M.A.P.
In connection with I.e.15 and I.e.20 we mention the following open problem.
Problem l.e.21. Let X be a Banach space having the B.A.P. Does there exist an
equivalent norm 111·111 on X so that (X, III· liD has the M.A.P.?
The investigation of the approximation property will be continued in Section
2.d. This section (starting from Theorem 2.d.3) can be read directly after the
present section.
f. Biorthogonal Systems
The existence of separable Banach spaces which fail to have a basis motivates the
attempts to try to use some weaker forms of coordinate systems. One approach,
which has been studied for a long time and for which strong existence theorems
are now available, is that of using biorthogonal systems.
Definition I.U. Let X be a Banach space. A pair of sequences {Xn};:"=1 in X and

{X~};:'=1 in X* is called a biorthogonal system if x;(xn)=S:i'. A sequence {Xn};:"=l
in X is called a minimal system if there exists a sequence {X~};:"=1 in X such that
({xn};:'=l> {X~};:"=l) is a biorthogonal system.
It is clear that a sequence {Xn};:'=l is minimal if and only if, for every integer n,
Xn ¢= [Xd~1.i*n' Observe that if ({xn};:"=l> {X~};:"=1) forms a biorthogonal system then
both {Xn};'=l and {X~};'=l are minimal systems. Every basic sequence {Xn};'=l is a
minimal system. The functionals {X~};'= 1 are in this case the biorthogonal functionals
discussed in Section b (more precisely, extensions of those functionals from [X n ];'= 1
to all of X). Observe that, unlike the notion of a basis, the notions of a minimal
system and of a biorthogonal system involve no natural ordering and they should
therefore be considered as countable sets of elements rather than sequences. There
are important examples of minimal systems which do not form a basic sequence in
any ordering. For example, take xn(t) =eint , n=O, ± 1, ± 2, ... in C(O, 27T) (=the sub-
space of C (0, 27T) consisting of those functions f for which f(O) = f(27T )). The corres-
ponding functionals x~ in C(O, 27T)* are in this case the measures given by
x; = rint dt, n = 0, ± 1, ± 2, .... The fact that {x n};'= _ does not form a basis under
<Xl
any ordering follows e.g. from a result of P. Cohen [19].

In Section c above we proved the existence of a nice biorthogonal system in any
finite dimensional space (called there an Auerbach system). We shall present now
existence theorems in the infinite dimensional case. Since we want to construct a
biorthogonal system which has to playa role similar to that of a basis rather than
a basic sequence we introduce first a definition of a suitable notion of completeness
of a biorthogonal system.
Definition 1.f.2. A minimal system {X n};'=l is called fundamental if [X n ];'=l is all

°
of X (i.e. x*(x,,) = for all n ~ x* = 0). A minimal system {x;};'= 1 in X* is called
total if X;(X) =0 for all n ~ x=o (i.e. X* is the w* closed linear span of {X;};'=l)'
If {x n};'= 1 is a basis in X then it is clearly fundamental and its biorthogonal
functlonals are total. The trigonometric system mentioned above is total and funda-
mental in C(O, 27T). The following general existence theorem is known for a long
time (it was proved first by Markushevich [101]).
Proposition 1.f.3. Let X be a separable Banach space. Then X contains a funda-

mental minimal system whose biorthogonal functionals are total (or more briefly,
with a slight abuse of language, X contains a total and fundamental biorthogonal
system).
Proof. Let {Yn};'= 1 be a sequence of non-zero elements in X such that [Yn];'= 1 = X

and let {Y;};'=l be a sequence in X* such that y;(x)=O for all n implies x=O.
We shall construct inductively elements {X n};'=l eX, {X;};;'=l c X* such that
x;(xn ) = 8;;,', span {X n};'= 1 => span {Yn};'= 1 and span {X;};'= 1 => span {Y;};'= 1.
We start by taking Xl = Y1 and put xI = yZ)yZ/Y1), where k1 is any integer
such that yZ1 (Y1) of- 0. Next, we take the smallest integer h2 such that yit" ¢= span XI.
Put x~=Y:2-xI'Y:2(X1) and let X2=(Xk2-X1·xi(x/c2))lx~(X"2)' where k2 is any
index such x;l'(Xk2) of- 0. It is easily checked that with this choice x;(x m ) = 8;;,' for
1 ::;; n, m:;;; 2. In the next step we let h3 be the smallest integer such that
YIl3 ¢= span {Xl, X2}' We put
and
yt
where k3 is such that 3 (X3)#O. We continue in an obvious way. In the step 2n
we start in X* and construct first the element xtnwhile in the step 2n + I we start
by constructing X 2n +l. It is clear that span {X;}r~l ~span {Yi}f=l and span {X;t'}r~l ~
span {y;t'}f= 1, for every n, and that xj(x;) = Sf. 0
Remark. The proof given above shows that we can assure that the {x;};;'= 1
are not only total but also that [X;];;'=l is norming, i.e. that for every x EX,
Ilxll =sup {lx*(x)l; Ilx*11 ~ 1, x* E [X;];;'=l}. Indeed, we have simply to start the con-
struction with a sequence {Y;};;'=l so that [Y;];;'=l is norming. Similarly, we can
assure that [x;];;'= 1 = X* if X* is separable.
The simple proposition 1.f.3 can be used for replacing bases (which may not
exist) in several situations. There is, however, one obvious drawback to the con-
struction given in l.f.3. We have no control on Ilxnll and Ilx;ll; in general,
sup Ilxnllllx;ll=oo (of course we can always normalize the system so that, e.g.
n
Ilxnll = 1, for every n, but then sup Ilx;11 may be 00). If {Xn };;'=l is a basis then clearly
n
sup Ilx!llllxnll <00. Solving a problem which was open for a long time, R. Ovsepian
n
and A. Pelczynski [112] proved that 1.f.3 can be strengthened to ensure that we
get also sup Ilx;llllxnll <00.
n
Theorem 1.f.4. In every separable and infinite-dimensional Banach space X there is

afundamentaland total biorthogonal system ({x n};;'= 1, {x;};;'= 1) SO that Ilxnll·llx;II~20,
for every n. If X* is separable the system may be chosen so that, in addition,
[X;];;'=l = X*.
Proof The proof is divided into two steps. The first step is a slight refinement of
the proof of 1.f.3 which shows that it is possible to choose a fundamental and total
biorthogonal system ({Un};;'=l, {U;};;'=l) so that, for some subsequence {nkh~=l of
the integers, IlunJI·llu;kll ~ 3, k= 1,2, .... The second step shows how, by using this
well behaved subsequence, it is possible to replace the biorthogonal system by
another one in which the entire sequence behaves well.
We begin with the proof of the first step. As in the proof of 1.f.3 we construct
{u n};;'= 1 and {u;};;'= 1 inductively. The inductive construction will depend onn (mod 3).
If n=3j+l, resp. n=3j+2, we do exactly the same which we did in the proof of
1.f.3 for n odd, resp. n even. This, by itself, will ensure that span {un };;'= 1 ~
span {Yn};;'= 1 and span {u;};;'= 1 ~ span {y;};;'= 1. The inductive construction for
n = 3j will be such that Ilutillllu3ill ~ 3. Assume that the Un and u; have been chosen
for n < 3j. By 1.a.6 there is a vector U3j in X such that IIu3jll = 1, U;(U3j) = 0 if n < 3j
and Ilxll~21Ix+'\U3jll, x E span {un}~j=-/, ,\ scalar. The functional on span {Un}~~l'
which assigns the value 0 to Un with n < 3j and the value 1 to U3j' has norm ~ 3. We
take as utj any Hahn-Banach extension of it to all of X*. This concludes the proof
of the first step.
For the second step we need the following lemma
Lemma 1.f.5. Let X be a Banach space and let {u;}r: 1 C X and {ut}r: 1 c X* be such
that Uj(Ui) = 8{. Then, there exists a real unitary matrix A = (ale, i) of order 2n x 2n
so that ~f
2n
Xle = 2: ale, iU;,
;=1 -
k= 1,2, ... , 2n
then
1. max Ilx"II~(l+V2) max
1 ~ k ~ 2n 1 ~ i < 2n
Iludl+2-nI21IU2nll
2. max
1~k~2n
Ilx;11 ~ (l + V2) max
1~i<2n
Ilutll +2-n/21Iu:nll
Proof. The relations 3 and 4 hold for any choice of a unitary matrix A. We get that
also 1 and 2 hold if we choose A = (a",;) so that, for k=l, ... , 2n ,
2"
2:
i= 1
la",il~1+V2 and la",2nl~2-nI2.
Such an orthogonal matrix A exists; put for 0 ~ s ~ n - 1, 0 ~ r~ 2s - 1
2n- s- 1 2r< k~ 2n- S - 1 (2r+ 1)

2n - S - 1 (2r+ 1) < k~ 2n - s - 1 (2r+2)
otherwise
and a",2n =2- nI2 for every k. 0
We give now the proof of step 2 of 1.f.4. We take the biorthogonal sequence which
was constructed in step 1 and reorder it in such a manner that in the new order
"most" elements are nicely bounded. More precisely, we give this sequence new
indices such that with the new indices the following holds. There is a sequence of
integers {nj}i= i so that if k is not of the form 2n1+ 2n2+ ... + 2n, for some i then
Ilu"ll~l and Ilu:II~3 while, for k=2n1+2n2+"'+2n" Ilu"I12-n,/2<1/20,
Ilu;112 -n,/2 < 1/20. It is clear that such a choice of indices can be made. Indeed, pick
the first element in the original sequence, sayan element u E X with a correspond-
ing u* E X*. We find an 111 so that IIuI12 -n1 /2 < 1/20 and Ilu*112 -n1 /2 < 1/20 and give
this element the index 2n1. For 1 ~ k < 2n 1 we take out of the original sequence the
first 2n 1 - 1 elements whose norms are 1 and whose biorthogonal functionals have
norms ~ 3. We consider now the first element in the original sequence which was
not chosen so far and use it to determine 112 , We continue in an obvious manner.
For every integer i we apply 1.f.5 to construct Xlc E X and x; E X* for 2n1+
2 2+'" +2n'-1 < k~ 2n1 +2n2+'" +2n, so that the span of the Xic with these indices
n
(resp. of the xt with these indices) coincides with that of the Uk (resp. the ut) with
the same indices, xt(Xj) = 15k, and
Ilxkll~(1 +'\1'2).1 + 1/20<2.5

/lxt/l~(I+v'2).3+1/20<7.5. 0
The proof gives clearly a constant which is better than the constant 20 appear-
ing in the statement of l.fA. By modifying the proof, Pelczynski [119] was able to
show that 1.fA remains true if 20 is replaced by any constant > 1. It seems to be
open whether it is actually possible to get /lxn/l /lx:/I = 1, for every n (in the finite
dimensional case this is possible; this is exactly the assertion of Auerbach's Lemma
1.c.3).
Many of the notions we encountered in Sections a and b can be defined in a
meaningful way also in the context of general biorthogonal systems. A minimal
system {Xn};'=l is said to be equivalent to a minimal system {Yn};'=l if there is an
isomorpllism T of [Xn]:'=l onto [Yn];'=l such that TXn=Yn for all n. In analogy
to l.a.9 it is easy to prove the following stability theorem. If {Xn};'= 1 eX
and {X:}:'= 1 c X* form a biorthogonal system and if {Yn};'= 1 C X satisfy
~ IIxn-Ynll /lx:/I < 1 then {Yn};'=l is a minimal system which is equivalent to
n=l
{xn};'= l' A minimal system {Xn}:'= 1 is said to be shrinking if the following relation
holds [X:]:'=l = ([X n]:'=l)* (here the x: are considered as functionals on [Xn]:'=l only).
A minimal system {X n}:'=l is said to be boundedly complete if, for every bounded
sequence {Yi}t~, 1 in [x n]:'= b the existence of lim x:(Yt) = an for every n implies the
i
existence of a vector yin [xnJ:'= 1 such that x:CY) = an for every n. Clearly, if {Xn}:'= 1
is a basic sequence these notions agree with those defined in Section b. The natural
generalization of l.b.5 is true in the present setting (with the same proof); let
({Xn};'=b {X:}:'=l) be a fundamental and total biorthogonal system for a Banach
space X. Then X is reflexive if and only if {xn};'= 1 is both shrinking and boundedly
complete.
A detailed discussion of minimal systems and their applications is given in
[106]. In this paper V. D. Milman uses, e.g. biorthogonal systems for proving
l.b.14. (In the original proof of l.b.14 given by Milman there is however a wrong
statement. He stated that a block minimal system (in the obvious definition of this
notion) of a boundedly complete minimal system is again boundedly complete.
As observed in [26] this is not true even for bases. What is true however is that
every block minimal system of a boundedly complete system has a boundedly
complete subsequence. This is a direct consequence of l.b.12.)
We conclude this section by showing that there is no non-trivial generalization
of the notion of an unconditional basis to the setting of biorthogonal systems. This
was proved by various authors in several degrees of generality (see, e.g. [7] and
[23]). We present here, following [98], a very simple variant.
Proposition l.f.6. Let {X n}:'=l be afundamental minimal system in a Banach space X

such that {X:}:'=l is total. Assume that, for every x E X and every subset a of the
integers, there is an element x" in X such that X:(X,,) = X: (x) for n E a and x:(x,,)=O
for n 1= a. Then, {Xn};:'=1 is already an unconditional basis of X.
Proof It is clearly enough to show that {xn};:'= 1 is a basis of X (the assumptions are
independent of the order). By 1.a.3 we have to show that the operators {Pn};:'=b
n
defined by Pnx= L Xt(X)Xh are uniformly bounded. Observe first that it follows
1=1
from our assumption and the closed graph theorem that, for every subset a of the
integers, there is a bounded linear operator PeT on X defined by P"x=x". If the
{Pn};:'=1 are not uniformly bounded we can construct inductively a sequence
of integers I =P1 <q1 <P2<Q2'" and vectors {uf}i'=1 so that lIuill =2- f,
<t>
UfESpan{Xi}r;,~~-1 and IIPq1uill~1 forj=I,2, .... Put a= U {i;Pf~i~Qi}' Then

j=1
<t> <t>
L Uj converges but L P"Uj fails to converge and this contradicts the continuity
1=1 j=1
of P(J' 0
g. Schauder Decompositions
A Schauder basis decomposes, in a sense, a Banach space into a sum of one-

dimensional spaces. It is sometimes useful to consider cruder decompositions where
the components into which we decompose a given Banach space are subspaces of
dimension larger than 1.
Definition l.g.l. Let X be a Banach space. A sequence {Xn};:'=l of closed sub-

spaces of X is called a Schauder decomposition of X if every x E X has a unique
<t>
representation of the form x = L

,,=1
x"' with x" E X" for every n.
Observe that if dim X,,=1 for every n, i.e. X,,=span{x,,} then {X,,};:'=l is a
Schauder decomposition of X if and only if {x,,};:'= 1 is a Schauder basis of X.
Many results concerning bases generalize trivially to the setting of Schauder decom-
positions. Every Schauder decomposition {Xn };:'= 1 of a Banach space X determines
'" Xj= L" Xl' These pro-
a sequence of projections {Pn};:'=l on X by putting P n L
t=l 1=1
jections are bounded linear operators and sup IIPnl1 <00. The number sup IIP"II
n n
is called the decomposition constant of {Xn};:'=1' Conversely, every sequence of
bounded projections {P,,};:'=l on X such that PnPm=Pm!n(n.m) and limPnx=x for
"
every x E X determines a unique Schauder decomposition of Xby putting Xl =P1X
and Xn = (Pn - Pn -l)X for n> 1. As in 1.a.3 it is easily seen that it is possible to
replace the condition lim P nX = X by the apparently weaker conditions sup liPnil <00
n n
<t>
and U PnX=X.
n=l
A decomposition {Xn}:=l is called boundedly complete if, for every sequence

{Xn}:=l with Xn E Xno n= 1,2,3, ... for which sup II
n
i.
1=1
XIII <00, the series ~
1=1
Xl
converges. The decomposition is called shrinking if, for every X* E X*, we have
IIP;x*-x*ll-+ O. If this is the case the sequence {P;}:=l determines a boundedly
complete Schauder decomposition of X*. A decomposition {Xn}:=l is unconditional
co
if, for every X E X, the series L: Xno which represents x, converges unconditionally.
n=l
In this case, for every sequence 8=(81, 82, ... ) of signs, the operator Mo defined by
co co
Mo L: Xn = L: 8nxn is a bounded linear operator. The constant sup IIMol1 is called
n=l n=l 0
the unconditional constant of the decomposition.
The decompositions, which are most useful in applications, are those in which
dim Xn <00 for all n (sup dim Xn need not be finite). Such decompositions are
n
called finite dimensional Sc.hauder decompositions or F.D.D. in short. The same
proof as that of l.b.5 shows that if {Xn}:= 1 is an F.D.D. of a Banach space X then
X is reflexive and only if {Xn}:=l is boundedly complete and shrinking.
Of course, the interest in Schauder decompositions does not stem from results
which generalize trivially theorems on bases. Their importance stems from the
fact that there are results on F.D.D.'s whose analogues for bases are not known
(and perhaps false) or do not even have a meaningful analogue in terms of bases.
We shall illustrate this by proving here two results on F.D.D.'s which will be
applied in the next chapter. The first theorem shows that the answer to Problem
l.b.10 concerning bases has a positive answer in the setting ofF.D.D.'s. The second
result is a theorem which makes sense only in the setting of F.D.D.
Theorem 1.g.2 [60]. Let X be a separable infinite-dimensional Banach space. Then

there exists a subspace Y of X such that both Yand XI Y have a F.D.D. Moreover,
if X* is separable Y may be chosen so that Yand XI Y have a shrinking F.D.D.
Proof By l.f.3 (and the remark following it) there is a biorthogona1 system
{xn}:='lcXand {X:}:=lCX* so that [x n]:=l=Xand [X:]:=l is norming over X.
We can therefore choose inductively finite sets 0"1 c 0"2 C • •• and 1]1 C 1]2 C • •• so
co co
that 0" = U O"n and 1] = U 1]n are complementary infinite subsets of the positive
n=l n=l
integers and, for n= 1,2, ... ,
IIxlI::;;; (1 +n- 1) sup {lx*(x)l; IIx*1I ~ 1, x* E [XneO"n Ul1n}' for every X E [XI]leO"no
IIx*lI::;;; (1 +n- 1) sup {lx*(x)l, IIxll = 1, X E [XI]lel1nUO"n+l}' for every x* E [X?]lel1n.
For every n let Sn and Tn be the projections on X defined by Snx= L: 4(x)x"

teO'n
respectively Tnx= L: x?{x)x;. We claim that
lel1n
(i) IIT~[x!li'"eun+1I1::;;;1+n-1
(ii) IISnl[x1ll:'~nll::;;; 1 +n~l .

Indeed, let x* E [x;]te<1 n+l and pick an x E [X;]ien n U<1n+l so that IIxll = 1 and
IIT:x*1I ~(1 +n-1)IT:x*(x)l· Since Tnx - x E [Xj]ie<1n+l and IT:x*(x) I= Ix*(x) +
x*(Tnx-x)1 = Ix*(x)1 we get that IIT:x*II~(I+n-l)llx*ll. This proves (i); the
proof of (ii) is similar.
We show next that for x* E [xi]te<1' T:x* ~ x*. By (i) the sequence {T:X*}:'=l
is bounded. Let y* be any w* limit point of {T:X*}:'=l' Then, clearly
Y*(Xi) = x*(Xj) =0 for i E a and Y*(Xi)=X*(X;) for i E 7]. Since a U 7] is the set of all
positive integers we deduce that y* = x* and thus, indeed, T:x* ~ x*. It follows
also that [Xj]te<1 is the w* closure of [xt];en. Put Y= [xr]Ten= [Xi];e<1. By the
analogue of l.b.9 for F.D.D.'s it follows that XI Y has an F.D.D. From (ii) it
follows that {Snly}:'=l determines an F.D.D. in Y. This concludes the proof of the
first assertion of the theorem.
Assume now that X* is separable. Then the {Xn}:'=l can be chosen so that, in
addition, [X:]:=l=X*, Also, we may assume that the norm in X is such that
{Y:}:'=l C X*, Y: ~ y* and Ily:11 ~ Ily*11 => IIY:-Ynll ~ 0 (use l.b.ll). The proof
given above shows that in this case liT: x* - x* II ~ 0 for every x* E [xdte u (it is
here that we use the factor 1 +n- 1 in (i); for the proof of the first assertion of
l.g.2 it would have been enough to replace 1 +n- 1 by 2, say). This shows that
[XTJi e n is w* closed and that the decomposition of XI Y constructed above is
shrinking. Similarly, it follows from [X:]:=l =X* that {Snly}:'=l determines a
shrinking decomposition of Y. D
For stating the next result we need first a definition.
Definition 1.g.3. Let {Xn}:=l be a Schauder decomposition of X. Let 1 =kl <

k2 < k3 < ... be an increasing sequence of integers and put Y; = X ki EiJ X ki + 1 EiJ •.. EiJ
Xki+l- h i=l, 2, .... Then, the decomposition {lI};;'l of Xis said to be a blocking
of the decomposition {Xn}:'=l'
If {Pn}:'=l is the sequence of projection associated to an F.D.D. {Xn}:'=l then
the subsequences of {Pn}:=l are exactly the sequences of projections associated to
the blockings of {Xn}:'=l' The decomposition (resp. the unconditional decomposi-
tion) constant of {lI};;'l is smaller or equal to the decomposition (resp. the uncon-
ditional decomposition) constant of {Xn}:'=l' One word of caution should be said
concerning this definition. The notion of the blocking of a decomposition is not
the obvious generalization of the notion of a block basic sequence. The qirect
generalization of the notion of a block basic sequence to the setting of Schauder
decomposition would be to consider decompositions {lI}t'= 1 of subspaces of X
such that Yjc X ki EiJ X ki + l EiJ ••. EiJ Xki+l- l for every i.
The following result concerning blockings of F.D.D.'s was proved in [63] and
[59]. It turns out to be very useful in the study of the structure of subspaces of Lp
spaces (and other spaces as well). In view of its general nature we state and prove
it here; however, its significance will become clear only in view of its applications
(for example, in Section 2.d below).
Proposition 1.g.4. (a) Let T: X ~ Y be a bounded linear operator. Let {Bn}:'=l be

a shrinking F.D.D. of X and let {en}:'=l be an F.D.D. of Y. Let {Bi};;'l be a sequence
of positive numbers tending to O. Then there are blockings {B;}~I of {B n};:"=1 and
{Cal'= I of{ Cn};:"= I so thatJor every x E B[, there is ayE C;-l EEl C; so that II Tx- yll:::;
eillxll·
(b) Let T: X -->- Y be a quotient map. Let {Bn};:"=1 be an F.D.D. of X and {Cn};:"=1
a shrinking F.D.D. of a subspace of Y. Let {8i}~ I be a sequence of positive numbers
tending to O. Then there is a constant K and blockings {B;}~ I of {B n};:"= I and {C;h""= I
of {Cn};:"= I so thatJor every y E C;, there is an x E B; EEl B; +1 such that IITx- yll :::;
etllYl1 and Ilxll:::;Kllyll·
Both parts of l.gA (which are in a sense dual to each other) assert that after a
suitable blocking the given operator is close to being diagonal with respect to the
blockings; e.g. in (a) TBUs "essentially" contained in C[-I EEl C;.
Proof of (a). Let {Pn};:"=1> resp. {Qn};:"=1> be the projections associated to the given
decomposition of X, resp. Y. We note first that, for every 8> 0 and integer n, there
is an integer m such that if x E X with PmX = 0 then II QnTx11 :::; elixil. Indeed, other-
wise there would be an n, an 8> 0 and a sequence of vectors {XIe}k'=I in X so that
Ilxlell=l, IIQnTxlell;;::8 for all k and limPmxle=O for every m. Since QnYis finite-
Ie
dimensional we may assume (by passing to a subsequence if necessary) that, for
some y* E y* with II y*11 = 1, T*y*(xlc);;:: 812 for every k. This however contradicts
the assumption that {Bn};:"=1 is shrinking.
Using this observation we construct two sequences of integers 1 =ml <m2 <
ms<'" and l=kl <k2<ks'" as follows. We pick m2 so that if P m2 x=0 then
IIQ Ie1Txll:::;elllxII/2. Next, we let k2 be such that, for every X EPm2 X,
IITx- QIe2Txll:::;811IxII/2. Then, we pick ms so that if Pmax=O then IIQ Ie2 Txll:::;
8211xII/2 and ks so that IITx-QleaTxll:::;821IxII/2 for every X EPmaX. We continue
in an obvious manner. The sequences {I, m2+ 1, m3+ I, ... } and {I, k2+ 1, k3+ I, ... }
determine blockings with the desired properties.
Proof of (b). Let again {Pn};:"=1 and {Qn};:"=1 be the projections which correspond
to the given decompositions (note that the {Qn};:"=l are defined only on the sub-
space Yo= [Cn];:"=1 of Y). Let K=4+4 sup IIPnll. As above, we shall define the
n
suitable sequences of integers 1 = ml < m2 < ... and 1 = kl < k2 < . " inductively. It
will be clear how to choose the {mi}~ I and {k!}~1 once we show the following. For
every e > 0 and every integer n there is an integer k so that if y E Yo with QIeY = 0
there is an x E X with Pnx=O, Ilxll:::;Kllyll and IITx-yll:::;ellyll. Suppose this
were false for some 8> 0 and integer n. Then, there is an 8> 0 and a sequence
{Yj}i'=1 of vectors of norm 1 in Yo so that dey}, TU);;::e, j=l, 2, ... (where
U={x EX; PnX=O, Ilxll :::;K}) and lim QIeYj=O for every k(and thus, since {Qn};:"=l
j
is shrinking, Yj ~ 0). Since T is a quotient map there are {Xj}i'= I in X such that
Ilxjll:::;2 and TXj=yjoj=l, 2, .... Put Vj=PnXj and Uj=Xj-Vj. Clearly, Ilujll:::;Kl2
for every j. Since PnX is finite dimensional we may assume (by passing to a sub-
sequence if necessary) that Ilvj-VIII:::; el2 for every j. Hence,
Since Yj ~ 0 the point Y1 belongs to the closed convex hull of {Y1 - Yj}f= 2 and
consequently, d(Yl, TU)~ e/2. This contradicts the choice of Y1 and concludes the
proof. 0
Remark. The proof of l.g.4 ensures that not only the specified blockings have the
desired property but that the same is true for suitable blockings of these blockings.
More precisely, assume for simplicity that e1 > e2 > e3'" . Then, the blockings chosen
in (a) have the following property. For every blocking {B'j}f=1 of {B:};'~1 there is
a blocking {C'j}j=1 of {Cm~1 so that, for every x E Bi, there is ayE Ci-1 EEl Ci
with IITx-YII~ejllx/l. Similarly, the blockings chosen in (b) have the following
property. For every blocking {Ci}f=l of {C[}t'=1 there is a blocking {B'j}f=1 of
{Bn~1 so that, for every Y E C;, there is an x E Bi EEl Bj'+l with II Tx-yll ~ej"y"
and "x"~K/lyll.
We turn now to a discussion of the relation between the existence of F.D.D.'s
and that of existence of bases.
Since obviously the existence of anF.D.D. of a Banach space X implies that
Xhas the B.A.P; we note first that, by Enflo's example, there are separable Banach
spaces which fail to have an F.D.D. It is not known whether the existence of an
F.D.D. implies the existence of a basis (this is a special instance of problem I.e.12).
What is trivially true is the following fact: Let {Bn }:'=l be an F.D.D. of a Banach
space X. Assume that every Bn has a basis {Xi, n}~~ 1 with basis constant Kn and
sup Kn <00. Then, the sequence
n
forms a basis of X whose basis constant is ~ K· sup Km where Kis the decomposition
n
constant of {B n }:'=l'
For unconditional bases the situation is more involved. Let, for example, X be
the space of all compact operators' T on 12 which have a triangular representing
n
matrix with respect to the unit vector basis (i.e. Te n = L an,mem, n=I, 2, ... ).
m=l
Let Bn be the subspace of X consisting of those T E X such that Tej = 0 for j # n
(i.e. for which aj, m = 0 unless j = n). It is clear that Bn is isometric to I!}:, n = I, 2, ....
Moreover, it is trivial to check that {Bn}:'=l forms an unconditional decomposition
of X. Nevertheless, it follows from the results of [47] that X does not have an
unconditional basis and it is not even complemented in a space with an unconditional
basis. This space does however embed in a space with an unconditional basis.
This is a special case of the following general result.
Theorem l.g.5. Let X be a Banach space admitting an unconditional F.D.D. {B n}:'= l'
Then X is isomorphic to a subspace of a space with an unconditional basis.
Proof Without loss of generality we may assume that the unconditional constant
of {B n}:'= 1 is 1. For each n we choose a set of non-zero elements {xt n}~;; 1 in B;
such that //xt n/ I~ 1 for all i and so that, for every x* in the unit ball of B;, there
is an i such that Ilx*-xtn/I~4-n.
Define a map T from Xo=span {Bn}:'=l into the space Yo of sequences of

scalars which are eventually 0 by
m
T L xn=(XL(Xl)' XL(Xl),""
n=1
XZ1 ,1(Xl), XI,2(X2),"" xZ2 ,ix2),"')'
Let U be the unit ball of Xo and let V be the convex hull of U

e
MeTU, where
the union is taken over all sequences of signs ()=«()h ()2,''') and Me is the operator
on Yo defined by Me(al, a2'''')=«()lal, ()2a2,"')' We introduce a norm in Yo, by
putting Ilyll =inf {,\>o; y/'\ E V}, and we let Y be the completion of (Yo, II·ID.
It is clear that the unit vectors form an unconditional basis of Yand that T extends
to an operator of norm 1 from X into Y. We have to show that Tis an isomorphism
into.
m
Let U= 2:
Un be an element of norm 1 in Xo. By the Hahn-Banach theorem
n=1
and the fact that the unconditional constant of the decomposition is 1 there is an
m m
x*= 2: x; in X* with Ilx*II=I,x;EB,'t and l=x*(u)= 2:
x;(un) (we identify
n=l n=l
in an obvious manner B,'t with a suitable subspace of X*). For each 1 :{"n:{"m let in
be such that Ilx;-xt.nll:{,,4- n. Let rp be the linear functional on Yo defined by
m
rp(ah a2, as, ... )= La
n=l
jn
wherejn is the index assigned to xt,n by T (i.e. jn=k 1 +k2 +,,·+kn- 1 + in). By the
m m m
definition of rp we get that rp(Tu)= 2: xt.n(un)~ 2: x;(un ) - 2: 4-n~ 1/2. Also,
n=1 n=l n=1
co
for every x = 2:
n=1
Xn E Xo with Ilxll:{" 1 and for every (),
m m co
rp(MoTx)= L ()nxt,n(xn):{" n=1
n=l
L ()nx;(x n)+ n=l
L 4- n=
Hence, by the definition of the norm in Y, rp E y* and Ilrpll:{" 3/2. Consequently,

II Tujj ~ 2rp(Tu)J3 ~ 1/3 and this concludes the proof. 0
2. The Spaces Co and lp
a. Projections in Co and Ip and Characterizations of these

Spaces
The simplest examples of infinite-dimensional Banach spaces are lp, 1 ~p~oo,

and Co. These spaces appeared in many problems in analysis much before a system-
atic theory of normed linear spaces was developed, and, as a result of continuous
efforts, their geometry is quite well known today.
The unit vectors {e n}:=1 form an unconditional basis (with unconditional
constant 1) of Co and lp, 1 ~p <00. Some very simple but important properties of
this basis are exhibited in the following proposition.
Proposition 2.a.l. Let X be either Co or lp, 1 ~p <00, and let {uj}j= 1 be a normalized
block basis of the unit vector basis {e n}:=1' Then,
(i) {Uj}i'=1 is eqUivalent to {en }:=1 and [Uj]i'=1 is isometric to X.
(ii) There is a projection of norm 1 from X onto [Uj]i'=1'
Proof We present the proof in the case of lp. The proof for Co is the same but the
mj+l mj+l
notation is somewhat different. Let Uj= L:
i=mj+1
.\e, with L: I.\IP= 1, j= 1,2, ....
i=mj+1
co
For every sequence of scalars {aj}i'= 1 with L:
}=1
lajlP <00 we have
and this proves (i).

To prove (ii) choose, for every j, an element uj E span {e'}~;;'~+1 c so that I:
Ilujll=uj(Uf)=1. Then uj(ul<)=O for k#j and the operator P defined by Px=
~ uj(x)Uj is a projection of norm 1 from X onto [Uj]i'= l' Indeed, if x = ~ a,e; E Ip
j=1 mj+1 co
i=1
then luj(x)lp~ L: la;jP for every j and thus IIPxW= L: luj(x)IP~ IlxW· 0
i=mj+ 1 j= 1
From 2.a.l we get immediately (by using l.a.9) the following result.
Proposition 2.a.2. Let X be either Co or Ip, 1 ~p <00. Then every infinite dimensional
subspace Y of X contains a subspace Z which is isomorphic to X and complemented
in X (and therefore also in Y).
54 2. The Spaces Co and /p
Another interesting fact which follows immediately from 2.a.1 (and l.a.l2) is
that no space of the family Co and lp, I ~p <00, is isomorphic to a subspace of another
member of this family. By using 2.a.2 A. Pelczynski gave in [114] a complete
characterization of the complemented subspaces of Co and lp, 1 ~p <00.
Theorem 2.a.3. Let X be either Co or lp, 1 ~p <00. Then every infinite dimensional
complemented subspace of X is isomorphic to X.
Proof Let Y be an infinite-dimensional complemented subspace of X; then

X=YEBXl for some Banach space Xl' By 2.a.2, Y=ZEB Y1 withZ;::::Xand a
suitable Banach space Yl • Then,
since XEB X;:::: X. Furthermore, observe that if X=lp, l~p<oo (resp. X=co) then
X is isometric to (X EB X EB .. ')p (resp. (X EB X EB .. ')0); in other words, we can
write for every such X, X = (X EB X EB .. ·h. Consequently,
XEB Y=(XEBXEB'''hEB Y;::::«YEBX1)EB(YEBX1)EB'''hEB Y

;::::(Xl EB Xl EB "'h EB (YEB YEB "'h EB Y
;::::(Xl EB Xl EB '''h EB (YEB YEB '''h
;::::«YEB Xl) EB (YEB Xl) EB ·"h=X.
Hence X;:::: X EB Y;:::: Y and this concludes the proof (the verification that all the
computations done here with infinite direct sums are valid, is straightforward). 0
The elegant method of proof of2.a.3 is called Pelczynski's decomposition method.

This method (as well as several variants of it) is very useful in many contexts.
Observe that all that we used in the computations above was that X and Yare
each isomorphic to a complemented subspace of the other space and that X is
isomorphic to an infinite direct sum of itself with respect to a suitable norm (the
fact that X EB X is also isomorphic to X is a consequence of this assumption).
Let us point out that it is unknown whether the assumption that X and Yare each
isomorphic to a complemented subspace of the other space suffices to ensure that
X is isomorphic to Y.
The decomposition method has however one drawback: it is hard to give an
explicit form of an isomorphism whose existence is proved by the decomposition
method. From the practical point of view (though not from the formal theoretical
point of view) the decomposition method is basically an existence proof. In simple
cases where other methods are available they usually give more information than
the decomposition method. We shall illustrate this by considering projections of
norm 1 in Co or lp, 1 ~p <00. In this case it is possible to give explicit representations
of the projections and their ranges. In particular, if X is Co or lp, 1 ~p <00 and P
is a projection of norm 1 on X with infinite dimensional range than P X is isometric
to X (this is a special case of a result due essentially to Ando [5] cf. also [79]). We
shall prove this here only in the case 1 <p <00, pof.2. For p=2 the result is of course
a. Projections in Co and lp and Characterizations of these Spaces 55
trivial. The cases 11 and Co are somewhat different; the proofs in these cases are
simpler.
Theorem 2.a.4. Let X=lp,for some 1 <p<00,p0/2 and let P be a projection of norm
1 in X. Then, there exist vectors {UJ1'=1 of norm 1 in X (where m=dim PX is either
an integer or 00) of the form
(*) Uj= L
ie:O"f
,\e;, 1 ~j~m, with aj n a/c= 0 for k=h
m
so that PX= 2: uj(x)Uj,
j=1
where {Uj}1'=1 E X* satisfy Ilujll =uj(Uj) = 1, j= 1, ... , m.
In particular, PX=[Uj]1'=1 is isometric to 1';;.
Observe that the proof of 2.a.l(ii) shows that, conversely, every P defined as
above is a projection of norm 1 on X. Note also that the {Uj}7= 1 do not necessarily
form a block basis of {ei }t"'=1 according to definition 1.a.lO; first, since the sets
{aj}1'=1 may be infinite and even if all the sets {aj}1'=1 are finite the {Uj}1'=l are then
only a block basis of a suitable permutation of {ei}~1' What 2.a.4 asserts is that
the obvious generalization of 2.a.l(ii) to the setting in which a block basis of
{e n};:'=1 is replaced by vectors satisfying (*) gives the most general projection of
norm 1 on Ip, 1 <p<00,p0/2.
Proof We introduce first some notations. For Oo/x E Ip the support supp x of x
is the set of integers i such that xCi) (the i'th coordinate of x) is 0/ O. The function
sign x on the integers is defined by sign xCi) = 1 if xCi) > 0, = -1 if xCi) < 0 and = 0
if xCi) = 0 (we give the proof in the case where the scalars are real; the same argu-
ment with some slight changes works also in the complex case). For any Oo/x E Ip
let tfop(x) be the unique element in Iq=l; for which tfoix)(x) = Iltfop(x)!lqiixiip= iixil~.
The element tfop(x) is given explicitly by tfop(x) = ixi P -1 sign x (i.e. tfop(x)Ci) =
ix(i)i p- 1 sign xCi) for 1 ~i<oo); for x=o put tfop(x)=O. Observe that tfoitfoix))=x
for every x E [p.
Let now P be a projection of norm 1 in X = Ip with 2 <p <00 (by duality it is
enough to prove the theorem in this case) and let Y=PX. Note first that if x=Px
then P*tfoix)(x) = tfop(x)(Px)=tfop(x)(x) and hence, by the uniqueness oftfop(x), we get
that P*tfop(x) = tfop(x). By duality we get conversely that P*tfop(x)=tfop(x) ~ PX=x.
Thus,
ker P={x*; P*x* =x*F = {tfop(y) ; y E YF =tfoP( YF .

Hence, since Y determines both P X and ker P, it follows that Y determines P
uniquely, i.e. on a subspace of lp there is at most one projection of norm 1. The
preceding remarks show also that tfop( Y) =p* X* is a linear subspace of lq. We shall
exploit this fact in proving the following lemma.
Lemma 2.a.S. Let Y be a subspace of lp, 2 <p <00, on which there is a projection of
norm 1. Let y, z E Y be two non-zero vectors. Then rzCy) E Y, where rzCy) is the
restriction of y to the support of z, i.e. rz(y)(i) = y(i) if i E supp z and rly)(i) =0 if

i 1= supp z.
Proof Since if;P( Y) is a linear subspace of Iq we get that, for every real t,
if;p(z+ ty)-if;p(z) E if;p(Y). A simple computation shows that lim (if;p(z+ ty)-if;p(z»jt
t-+ 00
exists (in norm) and is equal to (p-l)lzIP-2y and thus IzIP-2yE if;P(Y). Hence,
U1 = IzI1-<q-1)lylq-1 sign y=if;ilzIP-2y) E if;iif;P(Y» = Y.
By repeating the same argument we get that
U2 = IzI1-<q-1)lu1Iq-1 sign U1 = IzI1-<q-l)2Iyl<Q-1)2 sign y E Y,
and, in general, Un = IzI1-<Q-l)nl y l<Q-1)n signy E Y, n=l, 2, .... Since p>2 it follows
that q -1 < 1 and hence, U = lim Un = Izl sign y E Y. Consequently,
n
rz(y) = Iyl sign (izlsigny)= Iyl sign U E Y. 0
We return to the proof of 2.aA. We show first the following. Let io be an integer
which belongs to the support of some element in Y. Then, among all 0 =f y E Y such
that io E supp y, there is an element Yo whose support is minimal. In order to show
this it is enough to verify that if supp Y1 => supp J2=>'" 3 i o, with {yj}j'= 1 e Y, then
n
00
there is ayE Y such that supp y= supp yj. Such an element is lim ry/Yl),
j=l j
which belongs to Y by 2.a.5.
The subspace of Y consisting of those elements y for which supp ye supp Yo is
one-dimensional. Otherwise, there would exist y, z E Y such that supp ye supp Yo,
supp zesupp Yo, z(io)=y(io) = 1 and z(i)=fy(i) for some i E supp Yo "'{io}. Then,
;0 E supp (z(i)y- y(i)z)esupp Yo "'{i} and this contradicts the minimality ofsupp Yo.
lt follows from the preceding observation that there is a set {aj}T= 1 (m finite or
infinite) of disjoint subsets of the integers such that every element of Y vanishes
m
outside U
j=l
aj and elements {Uj}T=l of norm I in Y such that aj=SUpp Uj and
{y; Y E Y, supp yeaj} is the one dimensional space spanned by Uj; j= 1,2, ... , m.
From this and 2.a.5 we deduce that Y= [Uj]T=l' The fact thatP has the desired form
follows from the fact that P is determined uniquely by Y. 0
In terms of the Banach-Mazur distance, 2.aA (and its analogue for Co and 11 )
states that if X is either Co or Ip , for 1 ~p <00, and P is a projection of norm 1 on
X with dim P X =00 then d(X, P X) = 1. From 2.a.3 it is easy to deduce that there
is a functionf('\) defined on'\ ~ 1 (and independent ofp) such that if Pis a projection
on X with dim PX=oo then d(X, PX)~f(IIPIi). Sincef(1)=f 1 this does not reduce
to 2.aA if liP II = 1 and leaves open the question whether the following "perturbed
form" of 2.aA is valid. Does there exist a function g('\) defined on ,\ ~ 1 such that
lim g('\) = 1 and such that, for every projection P on X with dim P X =00,
1<-1
d(X, PX)~g(IIPII)?
The property of Co and Ip , 1 ~p <00 exhibited in 2.a.3 is of enough interest to
justify a special terminology.
a. Projections in Co and 11' and Characterizations of these Spaces 57
Definition 2.a.6. An infinite-dimensional Banach space X is said to be prime if

every infinite-dimensional complemented subspace of X is isomorphic to X.
Besides Co and Ip, 1 ~p <00 the only known example of a prime space is I",. It
is very likely that there are many other examples of prime spaces. In Section 4.c
we shall give some examples of OrIicz sequences spaces which are conjectured to
be prime. We present now the proof that 100 is prime. It is clear that this case requires
a proof different from that of 2.a.3 since 100 is not separable and thus does not have
a Schauder basis.
Theorem 2.a.7 [84]. The space 100 is prime.
Proof Let Y be an infinite dimensional complemented subspace of rX). The main

point in the proof is to show that Y has a subspace Z isomorphic to 1 Such a 00 ,
subspace Z is necessarily complemented in Y. Indeed, if T: Z --+ 100 is an isomorph-

ism define an extension T of T, from Yinto I"" by Ty=(yt(y), y~(y), ... ), where y{
is a Hahn-Banach extension of the functional z --+ Tz(i), i= 1,2, ... , (for u E 100 , u(i)
denotes the i'th coordinate of u). The operator Q=T-lT is a projection from Y
onto Z. We are now in a position to apply Pelczynski's decomposition method to
conclude that Y';:::,/oo.
The proof that Y contains a subspace isomorphic to 100 is based on the follow-
ing two general facts concerning C(K) spaces which will be proved in Vol. III.
(i) Let T be an operator from a C(K) space into a Banach space X which does
not have a subspace isomorphic to Co. Then, T is weakly compact .
. (ii) IfT: C(K) --+ C(K) is weakly compact then T2 is compact.
By applying (i) and (ii) to a projection operator P on a C(K) space we see that
if the range of P is infinite dimensional then P is not compact, hence not even
weakly compact (note that p=p2). Thus, the range of P contains a subspace iso-
morphic to co. Since 100 is a C(K) space (K=the Stone-Cech compactification of
the integers) it follows that the complemented subspace Y of 100 contains a subspace
isomorphic to co. Thus, there exist vectors {Yn};:'=l in Yand a constant M such that
00
It follows from (t) that, for every integer i, 2: lynCnl ~ M.
n=l
Hence, the series
n~l anYn is w* convergent for every bounded sequence {a n};:'=1 of scalars (i.e.
nt anYn(i) converges for every i). The right-hand inequality of (t) remains valid
for every choice of a bounded sequence of scalars {an};:'=l' It is not clear however
that the same is true for the left-hand inequality of (t). Moreover, since Y is not
00
necessarily w* closed the expression 2: anYn need not be in

n=l
Y if {an};:'=l does not
converge to O. We shall see that by passing to a subsequence we can overcome
these two difficulties.
Lemma 2.a.S. Let {Yn}:'=l be a sequence of vectors in leo which satisfies (t). Then
there is a subsequence {Ynk}Z'= 1 of {Yn}:'= 1 so that
Proof For xE/oo and 8>0 we put N(x, e)={i; Ix(i)l<e}. We observe first that,
for every e, there is an integer no so that, for infinitely many indices n, the restriction
of Yn to the subset N(Yno' e) of the integers N has norm ~ 1. This follows from the
00 T
fact that 2: IYn(OI ~M and thus, if r> Mle, n=l

n=l
U N(Yn, e)=N. Hence, we can take
as no one of the integers from 1 to f.
Using this observation we can find an integer n1 and an infinite sequence of
integers N1 so that if n E N1 the restriction of Yn to N(Ynl' 1/8) has norm ~ 1. Pick
i1 such that IYn/i1) I~ 3/4. By passing to a subsequence of N 1, if necessary, we may
assume also that 2: IynCi1 ) I< 1/8.
. neNl
Next we pick an n 2 E N1 and an infinite subsequence N 2cN1 so that, for n E N 2,
the norm of the restriction of Yn to N(Ynl' 1/8) () N(Y n2' 1/8 2 ) is ~ 1. By our assump-
tion on N1 there is an i2 E N(Ynl' 1/8) so that IYn2(i2) I> 3/4. By passing to an
infinite subsequence of N 2 , if necessary, we may assume also that 2: IynCi2 ) I< 1/8 2 •
neN2
We continue this inductive construction in an obvious manner and obtain
subsequences {n",}Z'=l and {ik}Z'=l of the integers so that IYnkCik) I> 3/4 for every k
and
00
2:
j= 1
IYniik)I~1/8+1/82+"·=1/7.
Nk
The subsequence {Y nk}Z'=l satisfies m. Indeed, if sup lakl = 1 pick a ko such that
k
lakol > 9/10. Then,
~ (3/4)· (9/10) -1/7 > 1/2 .

This proves the left-hand side of m. We observed already that the right-hand side
of mis automatically true. D
We return to the proof of 2.a.7. For simplicity of notation we assume, as we

may, that {Yn}:'= 1 itself satisfies (:j:) For every infinite subset No of the integers we
00
let XNo be the subspace of 100 consisting of all vectors of the form 2:
anyno where
n=l
{a n}:'=l is bounded and a,,=O if n 1= No. By (t) each such X No is isomorphic to 100'
We shall show now that fhere is an No such that X No c Yand this will conclude the
proof. Let {NyhEr be an u.ncountable collection of infinite subsets of the integers
such that NH () Ny, is finite for every Y1 #Y2 (to prove the existence of such a
a. Projections in Co and Ip and Characterizations of these Spaces 59
collection assign to each real number t a sequence of rational numbers converging

to t. In this way we get an uncountable collection of subsets of a countable set (the
rationals) such that the intersection of every two different sets in the collection is
finite). If, for each y E r, the space X Ny is not contained in Y we can find, for each
y, an Xy E X Ny with Ilxyll = 1 and Txy=I=O, where T: leo -+ I,,) Yis the quotient map.
By our assumption on the collection {NY}YEr and the fact that all the {Yn};'=l
belong to Y it follows that, for every choice of scalars {b k}r=l and every choice of
distinct indices {YIJr= 1 in r, Ilkt bkTXYkll~2M s~p Iblel. Hence, for every cp E (leol Y)*
and every e > 0, there are only finitely many Y E r such that Icp(Txy)1 > e and thus,
there are only countably many Y E r such that cp(Txy) =1= O.
Since Y is complemented in leo the space 1001 Y is isomorphic to a subspace of
leo. Hence, there is a sequence offunctionals {CPj}J'=l in (leoIY)* which is total (i.e,
if U E leol Yis such that cpiu)=O, for every j, then u=O). Since r is uncountable we
deduce that there is ayE r such that cplTxy)=O for j= 1,2, .... This however
contradicts the assumption that the {CPj}i'=l are total and TXy=l=O. 0
We already noted above that it seems likely that the fact that Co and lp, 1 ~p <00
are prime does not characterize these spaces among the class of all separable spaces
and probably not even among all spaces having an unconditional basis. The main
ingredients used in the proof of 2.a.3, namely the two parts of the simple Proposition
2.a.l, do however characterize Co and lp, 1 ~p <00 up to isomorphism. This (and
its analogues in the function space and "local" settings) explains why the structure
of LifL) and C(K) spaces is far simpler than that of general function spaces. We
present first a result of M. Zippin [146] which shows that 2.a.l(i) characterizes in
a very strong sense the unit vectors of Co or lp.
Theorem 2.a.9. Let X be a Banach space with a normalized basis {X n };'=l' Assume
that {X n};'=l is equivalent to all its normalized block bases. Then, {X n };'=l is equivalent
to the unit vector basis in Co or in some lp, 1~p <00.
Proof First notice that since {X n};'=l is equivalent to {8 nXn};'=1, for every choice of
signs {8 n};'=1, we get that {X n };'=l is unconditional. Hence, we can assume without
loss of generality that the unconditional constant of {X n };'= 1 is 1. Next, using a
uniform boundedness argument, the following is proved: there exists a constant
M so that, for every normalized block basis {Uj}J'= 1 of {xn};'=l> the operator Twhich
exhibits the equivalence of these basic sequences satisfies II Til, II T -111 ~ M or,
equivalently,
eo
for all choices of scalars {aj}J'= 1 such that 2: ajxj converges. Taking in (*) Uj =
j=l
xmr
where m1 < m2 <"', aj= 1 for j= 1, ... , nand aj=O for j>n, we get
Let nand k be integers and construct blocks {Uj}i'= 1 as follows:
Ul = ni~l Xi Illni~ Xi II ,
k
-
1 k
-
1
By applying (*) to these {UAi=l and suitably chosen {aJi'=l we get that
where '\(n) = It~l XII · It follows easily by induction that
Let m, nand k be integers and denote by [x] the integer part of a positive real
number x. By using the preceding inequality we get that
M -2[lclogml'\(nylclogml~ '\(n[lcl0gml)~ '\(m[ldOgnl+l)

~ M2[lclognl+2'\(mYIc log nl+1 •
It follows from this, after some easy computations and letting k -+ 00, that
IIOg,\(n)
log n
log'\(m)I~210
log m g
M(_l_+_l_).
log n log m
Hence, c=lim log '\(n)/log n exists. Passing to the limit, as m -+00, in the pre-
n
ceding inequality we get
Since 1 ~ '\(n)~n it follows that o~ c~ l. If c=o we get that '\(n)~ M2 and thus
for every sequence of scalars {a n}:'= 1 which are eventually 0, i.e. {X n}:'= 1 is equivalent
to the unit vector basis of Co.
If c> 0 we put p= l/c. To prove the equivalence of {Xn}7~=1 with the unit vector
basis of Ip we let {rj}J"=l be positive rational numbers with rj=kjlk, where {k j }7'=l
and k are positive integers. It follows from (**) that
a. Projections in Co and /p and Characterizations of these Spaces 61
where
IIj~1 A(kj)Xjll ~M-t~1 ajXjll~M-2I1j~1 ajujll=M-2A(~1 k j)

~M-4(i k j)1lP =M- 4 k 1IP
j=1
(i1=1 rf)1lP ,
and this completes the proof. D
Remarks. 1. The following terminology is used in several places. A normalized

basis {X n};:'=1 is said to be perfectly homogeneous if it is equivalent to any of its
normalized block bases. Since 2.a.9 states that the perfectly homogeneous bases
are exactly those which are equivalent to the unit vector basis of Co or some
1m 1 ~p <00 this terminology is no longer very useful; it is just an abbreviation of
"being equivalent to the unit vector basis of Co or some lp, 1 ~p <00".
2. The proof of 2.a.9 uses only the fact that {X n};:'=1 is equivalent to each of its
normalized block bases with constant coefficients, i.e. blocks having the form
Uj = mJ+l
n=mj+ 1
III
2: Xn mj+l 2: XnII , for some increasing sequence of integers {mj};' 1. The
n= mj+1
significance of this remark will become clear in the next chapter.
We show now that a modified version of 2.a.l(ii) also characterizes the unit
vector basis of Co or lpo
Theorem 2.a.tO [92]. Let {X n};:'=1 be a normalized unconditional basis of a Banach

space X. Assume that, for every permutation 7r of the integers and for every block
basis {Uj};'1 of {x"Cn)};:'=b the subspace [Uj];'1 is complemented. Then {X n};:'=1 is
equivalent to the unit vector basis of Co or lp for some 1 ~p <00.
The proof of 2.a.1O is based on a lemma which will be useful also in other con-
texts.
Lemma 2.a.n. Let X be a Banach space with an unconditional basis {X n};:'=1' Let
{An};:'=1 be a sequence of scalars tending to 0 and let Vj= 2: anx n, Wj= 2: anx n,
nE6j nEGj
j= 1,2, ... be normalized block bases of a permutation of {X n};:'=1 such that 8t n Uj= 0
for every i and j. Assume that [Vj + AjWj];' 1 is complemented in X. Then, for every
00 00
choice of scalars {7)j}i= 1 such that 2: 7)jVj converges, the series 2: Aj7)jWj converges
j=1 j=1
too.
62 2. The Spaces Co and lp
Proof Let Uj=Vj+,\wj,j=l, 2, ... and let Q be a projection from X onto [Uj]f=l'
Put
2: bi. jUj, 2: Ci.jUj,

o:J 00
QVi = QWi= i=l, 2, ....

j=1 j=1
Then, for every i, b;,;+A;c;,;= 1 and, since sup
i
Ic;,;1 <00, we get that lim b;,;= 1.
i
Let P a be the projection associated to the unconditional basis {X n};:'=1 and the set
ro 00
a= U aj
j=1
(i.e. Paxn=xn if n E a and P"xn=O if n ¢: a). Then, P"Qv;=.2 bi,jAjw"
J=1
i=l, 2, .... We regard P"Q as an operator from V=[V;];"~'1 into W=[Wj]i'=1' The
matrix corresponding to PaQ with respect to the given bases is (b;,jAj). By 1.c.S
the operator D: V - ? W, which corresponds to the diagonal of this matrix (i.e.
defined by DVj=Ajbj,jwj,j=l, 2, ... ), is also bounded. Thus the convergence of
00 00 00
j~1 7] jVj implies that of j~1 7] jDvj and therefore also that of j~1 Aj7]jwj (recall that
li~bj,j=l). 0
J .
Proof of 2,a,lO. Let {Vj}i'= 1 and {Wj}f'= 1 be normalized block bases of {X2n};:'=1,
respectively {X 2n +1};:'=1' We shall show that {Vj}i'=1 is equivalent to {Wj}f'=1' This
will imply that {X2n};:'=1 and {X2n+1};:'=1 are both perfectly homogeneous and
equivalent to each other. An application of 2.a.9 will thus conclude the proof.
00
Let {7]j}f'= 1 be a sequence of scalars such that 2 7] jVj converges. By the as sump-
j=1
tion in 2.a.1O and Lemma 2.a.11 we get that, for every sequence {Aj}i'=1 tending to
00 00
0, the series
j=1
2: 7]jWj fails
2 Aj7]jwj converges. If j=1 to converge we can find disjoint
finite sets of integers {a;};~1 and an e>O such that Iljfa,7]jwjll;:;,e for every i. Put
00
U; = 2:
jE~
7]jWj, i = 1, 2, ... , Since 2: A;Ui
;=1
converges whenever A; - ? 0 we get that
{U;}~1 is equivalent to the unit vector basis of Co. By using again 2.a.lI it follows
that in this case {vAf=1 is also equivalent to the unit vector basis of Co and (apply
00
2.a.ll once more) so is {Wj}f'=1' Thus, 2: 7]jWj must converge. 0

j=1
The use of permutations in the statement of 2.a.1O is necessary. This follows

from the next proposition.
Proposition 2.a.12 [15]. Let X be a Banach space with an unconditional basis

{X n};:'=1' Let {mj}f= 1 be an increasing sequence of integers with m1 =0 and
put Bj=span {xn}:i'~+Jj+ bj= 1,2, .... If every block basis of {Xn};:'=1 spans a comple-
mented subspace of X then the same is true for the natural basis of Y = (J1 EB Bj)p,
wherep=O or I~p<oo.
Proof We may assume without loss of generality that the unconditional constant of
{Xn};:'=1 is 1. By the "natural basis" of Y we mean the vectors {Yn};:'=1 defined as
follows: if mj <n~mi+1 we let Yn be the element in the direct sum whose only non-
b. Absolutely Summing Operators and Uniqueness of Unconditional Bases 63
zero component is x" in thej'th place. It is obvious that {Y,,};'=l is an unconditional

basis of Y whose unconditional constant is 1. For the proof we need the following
trivial observation. If {Pj}i=l are projections in {Bj}i=l such that s~p IIP,II <00 then
1
Q(Zlo Z2,"')=(Pl Zlo P 2Z2, ... ) is a projection on Y with IIQII =s~p IIP,II (here
1
Zj E Bj for every j). Q is called the projection determined by {Pj}i=l'
qk+l
Let now Uk= L a"Y", k= 1,2, ... be a normalized block basis of {Yn};'=l'
,,=qk+ 1
Let Nl be the set of those integers k for which there exists a j such that Uk E B,
(Le. mj~qk<qk+l~mi+l)' By our assumption on {X,,};'=l there is a bounded linear
projection P from X onto [qki 1 a"X"]kENl' By I.c.8 (and the remark thereafter)
j=qk+ 1
qk+l
there is no loss of generality to assume that if qk < n ~ qk +1 then PX" = A" L ajXi>
j=qk+ 1
where An =0 whenever k 1= N l . Hence, PBjcBj for every j. Let Ql be the projection
on Ydetermined by {P1B1}i=l' Then QIY=[Uk]kENl and Qluk=Oifk1=N1 •
We divide now N", Nl into two sets N2 u N3 by taking every second element in
N", Nl (i.e. the first, third, ... ) into N2 and the rest (i.e. the second, fourth, ... ) into
N 3. By this choice of N2 we ensure that if k < h are two integers in N2 then there is
ajsuch thatqk+l ~mj<q" and a similar statement holds for N 3. Thus, the sequence
{Uk} kEN 2 is equivalent to the unit vector basis of lp, resp. Co (if p=O). (If N",Nl
is finite then N2 is finite and we get only the unit vector basis of l~ for some finite y.)
Also, there exist {Uth EN2 such that Ilu;11 =U;(Uk) = 1 for every k E N 2, u;(u,,)=O
for every h #- k, and {U;}kEN2 is equivalent to the unit vector basis in I:, resp. ct.
The operator Q2 on Y defined by Q2U= L U;(Y)UI< is a projection of norm 1
kEN2
onto [Uk] kEN 2 so that Q2Uk=0 if k 1= N 2. In a similar manner we define Q3' Then
Q = Ql + Q2 + Q3 is a bounded linear projection from Yonto [Udk = l' 0
It follows from 2.a.12 that, e.g. the natural basis of X= (J 1 EEl l:)p has, for
every choice of 8 and p, the property that all its block bases span a complemented
subspace of X. For p #-8 the natural basis of this space is not perfectly homogeneous.
b. Absolutely Summing Operators and Uniqueness of

Unconditional Bases
In this section we shall present some of the basic facts concerning the class of
p-absolutely summing operators. These operators are, by definition, closely con-
nected to the spaces Lp(p.). They are of importance in the geometric theory of general
Banach spaces and, in particular, in the study of the structure of the classical spaces.
In this section we apply them in proving that Co and It have up to equivalence,
only one normalized unconditional basis (namely the unit vector basis). Further
applications of p-absolutely summing operators to the study of classical spaces
will be presented in Vol. II.
Definition 2.b.t. Let X and Y be Banach spaces and let p ~ 1. An operator

64 2. The Spaces Co and i p
T E L(X, Y) is called p-absolutely summing if there is a constant K so that, for

every choice of an integer n and vectors {xi}f= 1 in X, we have
The smallest possible constant K is denoted by rriT). The class of all p-absolutely
summing operators in L(X, Y) is denoted by niX, Y).
For T¢ niX, Y) we shall put rriT) =00. The I-absolutely summing operators
will be simply called absolutely summing operators. A straightforward verification
shows that, for every p, niX, Y) is a linear subspace of L(X, Y) and rriT)
defines a norm on np(X, Y) in which this space is even complete (i.e. a Banach
space). It is also trivial to verify that if S and Tare bounded linear operators whose
composition is defined then rrp(ST)<rriS)'11 Til and rriST) < IISlh(T). The
name "absolutely summing" will become clear if we consider the definition for p = 1.
Observe· that sup i
[[x*[[=lt=l
Ix*(xt)1 =sup {II i
1=1
8iXtll, 8i = ± 1,1 <i<n}. This implies
that an operator T E L( X, Y) is absolutely summing if and only if, for every sequence
00 00
{X n};:'=l in X such that 2: Xn converges unconditionally, the series n=l

n=l
2: TX n converges
absolutely (i.e. n~lIITXnll<oo).
The following factorization theorem, due to A. Pietsch [121], clarifies the notion
of p-absolutely summing operators for a general p and is a basic tool in several
applications.
Theorem 2.b.2. An operator T E L(X, Y) is p-absolutely summing for some 1 <p <00
if and only if there is a regular probability measure fL on the unit ball B x * of X* (in its
w* topology) and a constant K so that IITxll<K( J IX*(XWdfL(X*))l/P. Moreover,
Bx*
the smallest possible constant Kfor which such a measure fL exists is equal to rriT).
This theorem can be interpreted as follows. Let I: X -i>- C(Bx.) be the canonical
isometry defined by Ix(x*)=x*(x). For a probability measure fL on B x * let Jil be
the formal identity map from C(Bx .) into Lp(Bx., fL). Then, T: X -i>- Y is p-abso-
lutely summing if and only if there is a measure fL and a bounded linear operator
S from J,JX =Z into Y such that the following diagram commutes
X --~T'----+) Y
In general, S cannot be extended to an operator from Lp(Bx *, 1-') into Y (of course,
if p = 2 this is always possible since Z is then complemented in the Hilbert space
L 2(B x" 1-')).
Proof of 2.b.2. Assume first that such I-' and K exist. Let {xj}f= 1 be elements of X;
then
n n n
2:
j=l
IITxiW~KP f 2:
Bx*1=1
IX*(XiW dl-'(x*)~KP sup 2:
11.:<*11 .. 1 1=1
Ix*(xjW
and thus TTvCT)~K<oo. Assume conversely, that TEIIvCX, Y) and TTp(T)=1.

Consider the following subsets of C(Bxo)
F1={fEC(Bx o); sup f(x*)<I},

x*eBx*
F 2 =conv {f;f(x*) = Ix*(x)iP, IITxll=I}.
The sets F1 and F2 are convex, F1 is open and the assumption that TTp( T) = 1 im-
plies that F1 rJ F2 = 0. By the Hahn-Banach and the Riesz representation theorems
there exists a positive constant'\ and a regular measure I-' on Bxo such thatfE F1 =>
f f(x*) dl-'(x*) ~,\ and f E F2 => f f(x*) dl-'(x*) ~'\. Since F1 contains all the
B~ B~
negative functions the measure I-' must be a positive measure and thus we may
assume without loss of generality that it is a probability measure. Since F1 con-
tains the open unit ball of C(Bxo) we get that ,\ ~ 1. Hence if x E X with II Txll = 1
then f Ix*(x)lp dl-'(x*)~ I, i.e. for every x E X, IITxW~ f Ix*(x)lp dl-'(x*). 0
Bx* Bx*
Several interesting facts can be read off directly from 2.b.2.

1. Since, for a probability measure 1-', the norm of a function in Lp(l-') is always
smaller than its norm in Lr(I-'), if p < r, we get that for all Banach spaces X and Y,
IIp(X, Y)cIIr(X, Y) whenever I ~p<r<oo and, moreover, TTvCT)~TTr(T)for every
TEL(X, Y).
2. Every p-absolutely summing operator is. weakly compact. For p> I this is
evident from the factorization diagram (since Z is reflexive). For p= I use the pre-
ceding observation.
'3. Theorem 2.b.2 provides a proof of a weak version of the Dvoretzky-Rogers
theorem I.c.2. In every infinite dimensional Banach space X there is a sequence
ro
{x.}f;l such that L: Xj cOlwerges unconditionally but not absolutely. Indeed,
i=l
assume that X is an infinite dimensional Banach space in which unconditional
convergence implies absolute converges. Then the identity operator of X is
I-absolutely summing and hence also 2-absolutely summing. By 2.b.2 there is a
Hilbert space H and operators Sl: X -+ H, S2: H -+ X such that S2S1 is the
identity of X. thus, X is isomorphic to a subspace of H and is therefore isomorphic
to a Hilbert space. It is however, obvious that in an infinite-dimensional Hilbert
space there are unconditionally converging series which are not absolutely con-
verging.
Before proceeding with the study of absolutely summing operators we prove a

classical inequality of Khintchine which has very many applications in the study of
the spaces Ip and Lp(f-l,,), in general.
Theorem 2.b.3. Let r,.(t) = sign sin 2n7Tt, n=O, 1,2, ... be the Rademacher functions
on [0, 1]. For every 1 ~p <00 there exist positive constants Ap and Bp so that
for every choice of scalars {a n}:;'=l'
Proof It is trivial that we can take A2 = B2 = 1 and also Ap = 1 if p ~ 2 and Bp = 1

if 1 ~p < 2. It is clearly enough to show that we can find a suitable Bp if p is an even
integer and a suitable Ap if p = 1. By considering the real and purely imaginary
parts of the {a n }:;'=l separately we see immediately that it is enough to consider
only real coefficients.
Observe that J o
1
r~i(t)r~~(t) ... r~s(t) dt, where nl < n 2 <". < ns, is
s
unless all the °
{ki}f = 1 are even in which case the integral is equal to 1. A direct calculation shows
therefore that let anrn(t)) 2k dt is equal to})y(2kl> 2k 2, ... , 2ks)a;'~1 a;'~2 ... a~~s,
where the sum is taken over all choices of 'h, n2,"" ns between 1 and m and all
s
choices of positive integers {k i}f=l such that k= L ki' The explicit form of y
i= 1
is given by y(2k1 , 2k 2, ... , 2ks) = (2kl +·"+2ks)!/(2k1 )!(2k2)! ... (2k s)!. By ex-
panding C~l a;)k we get a very similar expression the only change being
that y(2k 1 , 2k2'"'' 2ks) is replaced by y(kl' k 2 , ... , k s). Thus, the desired inequality
for p=2k holds if we take m~=sup y(2k1 , 2k 2, ... , 2ks)/y(kl' k 2, ... , k s). A short
computation shows that this gives B2k~ k 1 / 2 • It remains to verify the existence of
m
AI, Putf(t)= L anrn(t). Then, by Holder's inequality,
n=l
f If(t)12 dt=[ If(t)12/3If(t)14/3 dt~

1 1 ( 1
[If(t)1 dt
) 2/3 (1
[ If(t)1 4 dt
) 1/3
1
~ ( [If(t)1 dt
) 2/3
. Bt/3 [If(t)12 dt ) 2/3
( 1
This gives the desired result (with Al =Bi 2 ). 0
We shall apply now Khintchine's inequality for identifying the class of p-abso-
lutely summing operators in L(X, X) for the simplest possible case, i.e. when
X=12' It turns out that in this special case the space IIp(X, X) does not depend
onp.
Theorem 2.b.4 [116]. For every 1 ~p <<Xl the space IIp(l2' 12) consists exactly of the
Hilbert-Schmidt operators.
We recall that an operator Tin L(12' 12 ) is called a Hilbert-Schmidt operator if

00
L IITen I1 2 <oo, where {e n},';'=l is the unit vector basis in 12 , Also, if {Yn},';'=l is any
n=l
orthonormal basis in 12 then
L IITYnI1 L L IIT* e 11
00 00 00
2 = I(TYn, em)12= m 2 ,
n=l n,m=l m=l
<Xl <Xl
hence, L IITYnI1 2 = n=l

n=l
L IIT*YnI1 2 and this expression is independent of the par-
ticular choice of {Yn},';'=l and is denoted by IITII~s.
Every Hilbert-Schmidt operator T is compact. Hence, as for any compact T
in L(l2' 12 ), there is an orthonormal basis {X n},';'=l of 12 , an orthonormal system
{Yn},';'=l in 12 and scalars {A·n},';'=l such that TX n = AnYn for all n. An operator T given
00
in this form is Hilbert-Schmidt if and only if L IAnl2 <<Xl.

n=l
Proof of 2.b.4. We show first that every Hilbert-Schmidt operator Tis I-absolutely
summing. By observation 1 following 2.b.2 this will prove that Tis p-absolutely
summing for every p > 1.
Let T be given by TXn=AnYm n= 1,2, ... and let A= ( n~l IAnl2
00 )112 <<Xl. Let
00
{uj}J'= 1 be any m-tuple of vectors in 12 and, for O~ t~ I, let vet) = A-I L r nCt)Anxn E 12 ,
n=l
where {rn},';'=l are the Rademacher functions. Then by 2.b.3,
m m
sup L
Ilxll';l j=l
ICui> x)l?: sup
O.;t';l
L ICui> vCt))1
j=l
and thus 1TICT)~Al1.A.

Assume, conversely, that Tis p-absolutely summing for some p?: 2. It follows
that, for every orthonormal basis {U n},';'=l of 12,
and hence lim II TUn II= O. This implies easily that T is compact and thus there are
n
orthonormal bases {Xn},';'=l and {Yn},';''';l in 12 and scalars {An},';'=l such that TX n = AnYn.
68 2. The Spaces Co and Ip
m
Let m be an integer and put w(t)= 2: r,,(t)xm 0::::; t::::; 1, where {rn}:i'=1 are the
n=1
Rademacher functions. Then, by 2.b.3,
::::;'TTiT) sup
IIxll=1
(f1I(w(t), x)IP dt )UP
0
=7Tp(T) sup
Ilxll=1
( L: rnCt)(xn, x) IP dt )1/P
f10 I,,=1
m
<t:J
Hence, .2: 1",,1 2 <00 and T is a Hilbert-Schmidt operator. 0

,,=1
Many interesting results concerning p-absolutely summing operators as well as

several applications of these operators are based on the following inequality due
to Grothendieck [49].
Theorem 2.b.S. Let (ai,iW,i=1 be a matrix ofscalars such that I1,1=1

~ ai, AS1::::; 11 for
every choice of scalars {tl }r=1 and {si}f=1 satisfying Itil::::; 1, ISil::::; 1. Then, for any
choice of vectors {XI}r=1 and {Y1}f=1 in a Hilbert space,
where KG is Grothendieck's universal constant (in case the scalars are real KG::::;
(e n/2 _ e- n/2)/2).
Proof. It is clearly enough to prove the theorem for real scalars. Also, it is easily
seen that it is enough to consider vectors {Xi }r=1 and {Yi}f=1 so that IlxI11 = IIYIII = 1
for every i. Since every finite dimensional subspace of a Hilbert space is isometric
to l~ for some k we may assume that {xi}f=b {Yi}f=1cl~ for some k. Let
S={u; u E l~, Ilull = I} and let p. be the unique probability measure on S which is
rotation invariant. A simple two-dimensional computation shows that, for every
choice of x, YES,
f sign (x, u) sign (y, u) dp.(u) = 1-28[x, Y]/'TT

s
where 8[x, y] is the angle between x andy (i.e. 0::::; 8[x, Y]::::;'TT and cos 8[x, y] = (x, y».
In view of our assumption on (ai,j) we have, for every u E S and for every {ti}f=l,
{SfW=l with Itil~l, IsJ~;l, that
-1 ~
i,
2:j= 1
ai,jSitjsign (x;, u) sign (Yh u)~ 1 .
By integrating with respect to fL we get that
n
-1 ~ 2:
i, j;::::.l
ai,jsitll-2B[xi,yj]/7T)~ 1 .
Hence, the matrix (aij1- 2B[x;, Yj]/7T) also satisfies the assumptions made on (ai, j).
Thus, by iterating the same argument, we get for every integer m that
Since, for every l~i,j~n,
2:
00
= (-1)m(7T/2-B[xi'Yi])2m+l/(2m+l)!
m=O
we conclude that
Several other proofs of 2.b.5 can be found in the literature (cf. [12, lO3, 123]).
Some give also a better estimate for KG' The best possible value of KG seems to be
unknown.
The next result, due to Grothendieck [49], is actually a restatement of 2.b.5
in terms of absolutely summing operators.
Theorem 2.h.6. Every bounded linear operator T from 11 into 12 is absolutely summing
and 7T1(T)~KGIITII.
m
Proof Let {ej}j=l be the unit vector basis of 11 and Ui= L ai jej,
j= 1 '
i=l, 2, ... , n be
n
vectors in If', for some m, such that L IX*(Ui)1 <s; IIx*lI, for every x* E It. Let
i=l
{Sj}7'=l be scalars of absolute value ~ 1 and let x; E rt
be defined by x;(ej)=sj if
1 <S;j<S;m and x*(ej)=O ifj>m. For every choice of {ti}f=l such that It;\<s; 1, 1 ~i<s;m,
For every 1:::;;i:::;;n let Y; E 12 be such that IIYil1 = 1 and (TUj, Yi)= IITuill. Then, by
2.b.S
Let us point out that 2.b.6 is in a sense a joint characterization of hand 12 , It

is proved in [87] that if X has an unconditional basis and, for some Banach space
Y, n 1(x, Y)=L(X, Y) then X is isomorphic to hand Y is isomorphic to a Hilbert
space. If we do not assume that X has an unconditional basis then there are more
n
spaces which satisfy 1(X, 12)=L(X, 12), For example, X=L1(0,1) has this
property.
There are many examples of pairs of spaces X and Y such that L(X, Y)=
np(X, Y), for some p ~ 2. For instance, we have the following:
Theorem 2.b.7 [87]. Every bounded linear operator T from Co into Ip, with l:::;;p:::;; 2,
is 2-absolutely summing and 772(T):::;; KGII Til.
Proof. Let {et}j;l and {jj}j;,l be the unit vector bases of co, respectively Ip, and let
<Xl
TEL(c o, Ip). Let (a; 1);'':1=1 be the matrix defined by Te;=

,
2: ai 1JJ, i=l, 2, .... For
;=1 '
every choice of y*=(al>a2, ... )E/; with Ily*ll=l and every choice of scalars
{ti}j;l> {S1}j'=l of absolute value:::;; 1 such that lim t;=O we have
i
where i; = (SIal> S2a2"") E I;. Let {Xk}~= 1 be n vectors in Co of the form

m n
Xk= 2: bk.ieb for some integer m, which satisfy k=l
i=1
2: IX*(Xk)12:::;; 1 whenever x* E ct
with Ilx*ll:::;; 1. In particular, by taking as x* the unit vectors in h we get that, for
n
1:::;;i:::;;m, 2: b~ I:::;; 1. By considering the m vectors u;=(b 1 b b2 b"" bn i), 1:::;; i:::;;m
k=l ' " ,
in I~ it follows from (*) and 2.b.S that
JJII~ a;a l , 1ul lt =1~l ttl (~ a1 bk, jal, 1) 2) 1/2 :::;; KGII Til.
Since this holds whenever II(al' a2,. .. )II=1 (in I;) we get that
(;) (Jl ttl (~ bk, ;aj, J) 2) P/2) lIP:::;; KGII Til.
Put C;,k = IJl bk, ;al, jr By using the triangle inequality in 12Jp (recall that p:::;; 2), i.e
( 2:n (2: Cj,k

00 ) 2JP) pJ2
. :::;; 2:
00 (n2: cJ!~
)pJ2
k=l j=l j=l k=l
b. Absolutely Summing Operators and Uniqueness of Unconditional Bases ?1
we deduce from mthat

m C~l C~Ji~ bk.ifti·,n2/pr2~KGIITII.
Since Txk = 1 bk,; Te;= j=ll=l
;=1
I 1 bk,ift;,d" k=l, 2, ... , n, (:)* states that
and thus 7T2(T)~KGIITII. 0
We state now without proof a result of L. Schwartz [133] and S. Kwapien [77]
which shows what happens if we take p > 2 in 2.b.?
Theorem 2.b.S. Let 2<p<r<00; then L(co, Ip) = IIr(c o, Ip). There are however
operators in L(co, Ip) which are not p-absolutely summing.
There are of course many other situations which can be investigated, e.g. the
structure of IIr(lp, Is) for arbitrary r, p, and s. The only operators T which are
relatively easy to investigate in the general case are the diagonal operators (i.e.
operators of the form Te; = A;/;, i = 1, 2, ... , where {e;};"~ 1, resp. {/;}~ 1, are the unit
vector bases in Ip , resp. Is. For a discussion of these results and several of their
variants we refer to [122].
We turn now to the question of uniqueness of unconditional bases in the spaces
\r
Co and lp, 1 ~p <00. We consider first the simplest case, i.e. p = 2. It follows from the
parallelogram identity in Hilbert space that the average of IIJ1 8 iX i over all
n
choices of signs {8i W=b is equal to L Ilxil12. Thus, if {U n}:=l is a normalized un-
i= 1
00 00
conditional basis in 12 then L anun converges if

n=l
and only if L
n=l
a~ <00. In other
words, every normalized unconditional basis of 12 is equivalent to the unit vector
basis (this observation goes back to G. Kothe [72] and E. R. Lorch [98]). The
situation in Co and 11 is similar to that in 12 but the proof is more difficult (since it
is based on 2.b.5).
Proposition 2.b.9 [87]. Every normalized unconditional basis in 11 or in Co is equivalent

to the unit vector basis of the space.
Proof Let {Xn}:= 1 be a normalized unconditional basis of 11' Let {a n}:= 1 be a

00
sequence of scalars such that L anxn converges. Let S be the operator from Co into 11
n=l
defined by S(A1' A2, .. ·)= J1 Ananxn. By 2.b.? 7T2(S)~KGIISII ~KGMllnt anxnll, where
M is the unconditional constant of {Xn}:=l' Hence, tt lanl2tZ ~KGMllnt anXnll·

It follows that the operator T: 11 -+ 12, defined by TC~1 anXn) = (al' a2,"')' is
a bounded linear operator with IITII~KGM. By 2.b.6, 7T1(T)~KGIITII~KeM.
co
Hence, for every X= 2: anXn in h,
n=1
Since obviously IL~1 anXnll~n~1 lanl we deduce that {Xn}:'=l is equivalent to the
unit vector basis in h.
Let now {Yn}:'=l be normalized unconditional basis in Co. By l.c.9 the basis
{Yn}:'=l is shrinking and thus the biorthogonal functionals {Y~}:'=l associated to
{Yn}:'=l form an unconditional basis of 11 such that I~ IIY~II~M, for some M and
n= 1,2, .... By the first part of the proposition the sequence {Y~}:'=l is equivalent
to the unit :vector basis in h. This implies immediately that {Yn}:'=1 is equivalent to
the unit vector basis of Co. 0
Remarks. 1. The same argument as that used in the proof of 2.b.9 shows that if
{Xn}:'=l is an unconditional Schauder decomposition of 110 respectively Co, then
co
there is a constant M so that, for every choice of Xn E X n, n= 1,2, ... with 2:
n=1
Xn
converging, we have n~lIIXnll~MILt Xnll, respectively IL~1 Xnll~MS~P Ilxnll·

2. The arguments used in the beginning of the proofs of 2.b.9 and 2.b.7 show
also the following. Let {Un };'=1 be a sequence of elements in [p, l~p<2 such that
co co
2: Un converges unconditionally; then 2: Ilun Il 2 <00. This fact was first proved
n=l n=l
by Orlicz [110] and his proof is simpler than the one given here (which relies on
co
2.b.S). Similarly, we deduce from 2.b.S that if 2: Un converges unconditionally in
co
n=l
Ip, for p > 2, then 2: Ilunll <00 for every r > p. This however is not the best possible
n=l
T
result. We shall see in Vol. II that we can take also r=p.

In contrast to the cases of Co, hand 12 there are normalized unconditional bases
in Ip, 1 <p <00, p=ft2 which are not equivalent to the unit vector basis [114]. In
order to see this we observe that the Khintchine inequality 2.b.3 shows that the
co
mapping T: 12 -+ Lp(O, I), defined by T(al, a2,''')= 2: anrn(t), is an isomorphism
n=1
for every I~p<oo.
Let now p~2 and let P be the orthogonal projection from Lp(O, 1) (which is,
under our assumption on p, a linear subspace of L 2(0, 1) onto [rn];'=l' The pro-
jection P is defined by
L ff(s)rn(s) ds·rn(t).
co 1
Pf(t)=
n=lO
It is bounded in Lp since
For every n let Fn be the subspace of LiO, 1) spanned by the characteristic functions
of the intervals [kI2n, (k+ l)/2n], k=O, 1, ... , 2n-1. Clearly, Fn=> En = [rj]f=l> Fn is
isometric to l:n and PFn = En for every n. In other words, there is a constant Kp so
that, for every n, there exist a subspace en of l:n with d( en, l~) ~ Kp and a projection
from l:n onto en of norm ~Kp. Consequently, the space (I
n=1
EB l~) is, for 2~p <00,
z:nt
p
isomorphic to a complemented subspace of (J1 EB =lp. By 2.a.3, C~1 EB l~t

is isomorphic to lp for 2~p <00 and, by duality, also for 1 <p < 2. The obvious unit
vector basis in Ct EB l~t is therefore, for 1 <p<00,p=l2, an example of a
normalized unconditional basis in If, (more precisely, in a space isomorphic to lp)
which is not equivalent to the unit vector basis of lp. Observe that the basis we have
just discussed has, by 2.a.12, the property that all its block bases span a comple-
mented subspace of lp. Observe also that by 2.b.9 the space It is not isomorphic to
C~1 EB l~t and, similarly, Co is not isomorphic to (Jl EB l~)o'
The preceding discussion concerning uniqueness of unconditional bases proves
a part of the following theorem.
Theorem 2.b.tO [97]. A Banach space has, up to equivalence, a unique unconditional

basis if and only if it is isomorphic to one of the following three spaces: co, It or 12,
We shall conclude the proof of 2.b.l0 in Section 3.a below.

We would like to show now how the weak form of 1.a.8 (namely, that every
infinite-dimensional Banach space with a basis has at least two non-equivalent
normalized bases) can be deduced from 2.b.l0. Let {Xn}~=1 be a normalized basis
of a Banach space X. If {Xn}~=1 is not unconditional then, for some sequence of
signs {8n}~=1' the normalized basis {8nXn}~=l of X is not equivalent to {Xn}~=1' If
{Xn}~=l is unconditional then, by 2.b.1O, X has a normalized unconditional basis
which is not equivalent to {Xn}~=1 unless X is co, It or 12, Thus, we have only to
show that these three spaces have conditional (i.e. not unconditional) bases. For
11 and Co this assertion is trivial. Th6 vectors el> e2 - el, ea -:- e2,'" form a conditional
basis of 11 ({en}~=l denotes as usual the unit vector basis). The summing basis is
a conditional basis of c (which is of course isomorphic to Co via the map
T(al> a2,"') = (lim an> al-lim an> a2-lim an, .. ,), from c onto co). For 12 the asser-
n n n
tion is by no means obvious.
Proposition 2.b.n. The space 12 has a conditional basis.
This proposition was proved first by Babenko [6] who used harmonic analysis
to construct concrete examples of COl1ditional bases in Hilbert function spaces. For
example, he showed that, for O<a< 1/2, the sequence Itl a eint, n=O, ± 1, ±2, ...
is a conditional basis of L 2( -TT, TT). The proof we present here is completely different
and is due to C. A. McCarthy and J. Schwartz [105].
Proof It is clearly enough to construct, for every even integer n, a set {Q,}f=l of
projections in l~ so that dim Q;l~=l for every i, QiQj=O if i'/=j, Itt Qill~2 for
1 ~ k~ n and lim II Ql + Q3 + ... + Qn-lll =00. Indeed, these assumptions on {Qi}f= 1
n
n
imply that L Qi=! and thus if O'/=Xi E Qila, l~i~n, then {X;}f=l is a basis of l~
i=l
whose basis constant is ~ 2 and whose unconditional constant tends to 00 with n.
From this it is obvious how to construct a conditional basis in 1= (~
2
n=l
EB l~) .
2
In order to construct suitable Ql we pick a sequence {lXk}:'= 1 of positive numbers
such that
(take, e.g. IXk = '8jk log k, for k> 1 and a suitable '8 > 0).
Let now n = 2m and let Qi be the operator on la whose matrix with respect to
the unit vector basis is defined as follows:
°
If i=2j-l,j= 1, ... , m the matrix which represents Qi has all its columns equal
to except the i'th column which is equal to (0,0, ... ,0,1, 1X1, 0, 1X2, 0, ... , IXm-l+l)
°
(for typographical reasons we write this column here as a row). If i=2j,j= 1, ... , m
the matrix which represents Qi has all rows equal to except the i'th row which
. i
is equal to (-lXl' 0, -IX) -1, 0, ... , -1X 1, 1, 0, 0, ... , 0). It is clear that each Qi is a
rank one operator and also that Qf = Qi (observe that the matrices representing the
°
°
Qi are triangular with a diagonal which is equal to except that i'th element which
is equal to 1). It is also easy to verify that Qil' Qi2 = for i 1 '/= i2.
In order to estimate the norm of Ql + Q2 + ... + Qt, 1 ~ i ~ n, we write this
operator in the form of D, + Ri where Di is the operator which is represented by
the diagonal of the matrix representing Ql + Q2 + ... + Qi' (Thus the matrix repre-
i
senting Di is a diagonal matrix whose diagonal is (1, 1, ... , 1,0, ... ,0).) Observe
that the non-zero entries in the matrix representing R, (i.e. the off diagonal entries
in the matricial representation of Ql + Q2 + ... + Qi) consist of at most k times the
term ak, k= 1,2, ... , m. Thus the Hilbert Schmidt norm of Ri is ~ (Jl ka~t2.
Hence,
In order to estimate the norm of Ql + Q3 + ... + Q2m -1 let x be the vector

(1,0,1, ... , 1,0) in lao Then Ilxll =ml/2 and
c. Fredholm Operators, Strictly Singular Operators and Complemented Subspaces of lp EEl lr 75
Hence
c. Fredholm Operators, Strictly Singular Operators and

Complemented Subspaces of Ip EB Ir
We have proved in Section a that each infinite-dimensional complemented sub-

space of Co or Ip, I <p <00 is isomorphic to the entire space. In this section we are
going to present a result of!. S. Edelstein and P. Wojtaszczyk [35] which shows that
the only isomorphism types of infinite dimensional complemented subs paces of
lp EEl Ir are the obvious ones, i.e. Ip, Ir and lp EEl IT" The proof of this result makes use
of some notions (like strictly singular and Fredholm operators) which enter into
many other contexts in functional analysis. A large part of this section will be
devoted to proving the basic results concerning these important notions. Our
presentation will go a little beyond the material which is strictly needed for the
proof of the result of Edelstein and W ojtaszczyk. Concerning the proof itself let
us point out already here one interesting feature. It uses spectral theory in an
essential way. This topic had so far only few applications in the structure theory of
Banach spaces. The section concludes with a study of the unconditional bases of
the spaces lp EEl Ir'
The fact that makes the approach of Edelstein and W ojtaszczyk work is that,
for p #- r, the space lp does not have an infinite-dimensional subspace isomorphic
to a subspace of lr (see the remark following 2.a.2). Let us give this property a
formal name.
Definition Z.c.I. Two infinite-dimensional Banach spaces X and Yare called totally
incomparable if there exists no infinite dimensional Banach space Z which is
isomorphic to a subspace of X and to a subspace of Y.
As mentioned above, any two different spaces of the set {co} u {/p; l::;;,.p <oo} are
totally incomparable. The notion of "totally incomparable spaces" was introduced
by H. P. Rosenthal [126]. In this paper it is proved that X and Yare totally incom-
parable if and only if, whenever U is a Banach space lvith subspaces Xl and Y1
which are isomorphic to X, respectively Y, the algebraic sum Xl + YI is closed in U.
Definition 2.c.2. An operator T: X --+ Y is called strictly singular if the restriction

of T to any infinite-dimensional subspace of X is not an isomorphism.
If X and Yare totally incomparable any operator from X to Y is strictly

singular. For general spaces X and Y, every compact operator is strictly singular.
The formal identity map from 11' to In if r > p, is an example of a strictly singular
operator which is not compact. In this connection it is worthwhile to mention the
following result of H. R. Pitt.
Proposition 2.c.3. Let 1<.p < r <00. Then, every bounded linear operator from IT
into 11' is compact. The same is true for every linear operator from Co into lpo
Proof Assume that T is a non-compact operator from IT into 11" Then there is a
sequence {X n},';'= 1 in IT so that Xn ~ 0 and IITxnll:;?: e for some e> 0 and all integers
n. By passing to a subsequence we may assume by l.a.9 and 2.a.l that {X n },';'=1 is
equivalent to the unit vector basis in IT and {TXn}::'=1 is equivalent to the unit vector
basis in 11" Since the formal identity map from IT into /p is not bounded (recall that
r > p) we arrived at a contradiction. The proof in the case where Co replaces IT is
the same. 0
Observe that the proof of 2.c.3 shows also that aTE L(lp, lp) is strictly singular
if and only if it is compact.
We present now two simple propositions concerning strictly singular operators
due to T. Kato [69J.
Proposition 2.c.4. Let X and Y be infinite-dimensional Banach spaces. Assume that

T: X -> Y is an operator such that the restriction of T to any subspace of X offinite
codimension is not an isomorphism (this is the case in particular if TX is not closed
in Y). Then, for every e> 0 there is an infinite-dimensional subspace Z of X so that
T 1z is compact and II Tizil <. e.
Proof For every I:l > 0 and every finite set {xn\,,= 1 of elements in X* there is, by our
assumption, an element XEX with Ilxll=l, IITxll<1:l and xi(x)=O,I<'i<.m.
Hence, for every e> 0 (see the proof of 1.a.5), there exists a normalized basic
sequence {Xn }::'=1 in X, with basis constant <.2, so that IITxn ll<,e.8- n , n=l, 2, ....
The space Z= [Xn J::'=1 has the desired property. 0
Observe that if, for some X oC X of finite codimension, TIXo is an isomorphism

then TXo, and thus also TX, are closed in Y.
Proposition 2.c.S. (i) The sum of two strictly singular operators is strictly singular.
(ii) The composition of a strictly singular and a bounded operator is strictly
singular.
Proof (i) Let T and S be strictly singular operators from X to Y and let Z be an
infinite-dimensional subspace of X. By 2.cA there is an infinite dimensional sub-
space Z1 of Z such that T 1Z1 is compact. By applying 2.cA again we get an infinite-
dimensional subspace Z2 ofZ1 such that SI Z 2 is compact. Clearly, (S + T)I Z 2 is com-
pact and thus S + T is not an isomorphism on Z. The verification of (ii) is straight-
forward. 0
c. Fredholm Operators, Strictly Singular Operators and Complemented Subspaces of Ip EB Ir 77
Another way of formulating 2.c.5 is to say that the strictly singular operators
form an operator ideal.
It should be pointed out that, in contrast to the ideal of compact (or weakly
compact) operators, the dual operator T* of a strictly singular operator T need not
be strictly singular. For example, let X be a separable Banach space which does
not contain a subspace isomorphic to 11 (e.g. X = lp, for p > 1, or X = co). Let T be a
quotient map from 11 onto X (cf. 2.f). By 2.a.2 T is strictly singular. The operator
T*: X -> 100 is an isometry (into) and thus is not strictly singular.
We introduce now another important notion, namely that of the index of an
operator.
Definition 2.c.6. Let T: X -» Y be a bounded linear operator for which TX is

closed. Put
0:( T) = dim ker T, ,B(T)=dim Y/TX.
If either 0:( T) < 00 or ,B( T) < 00 we define the index i( T) of T by i( T) = 0:( T) - ,B( T).
If o:(T) and ,B(T) are both finite (i.e. if i(T) is defined and is finite) then T is called
a Fredholm operator.
Observe that if the algebraic dimension of Y/TX is finite then, by the open
mapping theorem, it follows that TX is closed in Y.
°
A typical example of a Fredholm operator of index k > is the operator, in
Co or lp, which sends (al> a2,"') into (ak+l> ak+2,"')'
We shall prove now several general results concerning Fredholm and strictly
singular operators \fhich are due to 1. E. Gohberg and M. G. Krein [46J and to
T. Kato [69]. Observe first that T: X -» Y is a Fredholm operator if and only if
there exist subspaces Xl and B of X; Y 1 and C of Y, so that X = Xl EB B, Y = Y 1 EB C,
T1B=O, T 1X1 is an isomorphism onto Y 1 and dim B<oo, dim C<oo. With this
notation we have o:(T)=dim Band ,B(T)=dim C. This observation and simple
finite-dimensional linear algebra prove the following proposition.
Proposition 2.c.7. (i) Let T: X -» Y be a Fredholm operator and let S: X -» Y be

afinite rank operator. Then, T+S is a Fredholm operator and i(T+S)=i(T).
(ii) Let T 1 : X -» Y, T 2: Y -» Z be Fredholm operators. Then, T2Tl is a Fredholm
operator and i( T 2T 1 ) = i( T 2) + i( T 1 ).
(iii) Let T: X -» Y be a Fredholm operator. Then T*: y* -» X* is also a
Fredholm operator for which 0:( T*) =,B( T), ,B( T*) = 0:( T) and therefore, i( T*) = - i(T).
In the sequel we shall need the following finite dimensional lemma.
Lemma 2.c.S. Let X be a Banach space and let Band C be subspaces of X with
dim B<oo and dim C>dim B. Then, there is an x E C such that Ilxll =d(x, B)= 1.
Proof Assume first that X is a strictly convex Banach space. For every x E C
with Ilxll = 1 let f(x) be the unique point in B for which IIx-f(x)1 I=d(x, B).
Clearly, f is continuous and f( -x)= -f(x). By the antipodal map theorem of
78 2. The Spaces Co and /"
Borsuk [32, p. 347] and the fact that dim C> dim B we infer that there is an x with
Ilxll = 1 such thatj(x) =0, i.e. d(x, B)= 1.
In the general case we observe that we may assume without loss of generality
that dim C=dim B+ 1 <00 and that X=span {B, C}. Hence, Xis finite-dimensional
and thus has an equivalent strictly convex norm 111·111. For every integer n let 11·lln
be the strictly convex norm on X defined by Ilxlln=llxll+lllxlll/n, where 11·11 is
the given norm in X. By the first part of the proof we can find, for every integer n,
an Xn E C with Ilxnll n= 1 and inf {lIxn- Ylln; Y E B} = 1. If x is any limiting point of
the sequence {Xn }'':'=l then Ilxll=d(x, B)=l, as desired. 0
Proposition 2.c.9. Let T: X ~ Y be an operator jor which i(T) is defined. There is

a number A( T) > 0 so that if S: X ~ Y satisfies II SII < A( T) then
(i) a(T+S)~a(T).
(ii) T+S has a closed range and f1(T+S)~f1(T).
(iii) i(T+S)=i(T).
Proof. Every operator T with a closed range admits a unique factorization of the
form
1fT t iT
X~ X/kerT~ TX~ Y
where TTT is the natural quotient map, T is an isomorphism and iT is the inclusion
map. Put A( T) = II T -111-1. Since A( T) = A( T*) we may assume without loss of
generality that aCT) <00 (otherwise, we work with T*).
Let IISII<A(T) and suppose that a(T+S»a(T). Then, by 2.c.8, there
is an x E ker (T + S) such that 1 = Ilxll = d(x, ker T) = II7TTxl I· Hence,
II Txll = IIT7TTxll ~A(T).
On the other hand, Sx=-Tx and thus IITxll=IISxll~IISII. This contradicts the
assumption that IISII < A(T) and (i) is thus proved.
Assume that T+S does not have a closed range. Let O<e<A(T)-IISII. By
2.cA there is an infinite dimensional subspace Z of X so that II( T + S)lzll < e. By
2.c.8 there is an z E Z such that 1 = Ilzll = I17TTzll. Hence,
A(T) > IISII+e~ IISzll+II(S+T)zll ~ IITzl1 = IIT7TTzl I ~A(T),
and we arrived at a contradiction. If f1( T) = 00 this concludes the proof of (ii). If

f1( T) <00 we can apply (i) to T* and thereby conclude the proof of (ii).
In order to prove (iii) it is enough to verify it with some e> 0 instead of the A( T)
defined above. Indeed, once we show that 11U11~eT implies i(T+ U)=i(T), we
can apply this fact (by (i) and (ii» to T + tS for every t E [0, 1] provided of course
IISII < A( T). The function i( T + tS) is thus a continuous function of t which takes
only discrete values; hence, it must be constant, i.e. i( T) = i( T + S).
We prove the existence of such an e> 0 first in the case a( T) = 0, i.e. when T
is an isomorphism. Let U: X ~ Y satisfy 11U11 < A(T)/2 and let x E X with Ilxll = 1.
c. Fredholm Operators, Strictly Singular Operators and Complemented Sunspaces of lp E8 lr 79
Then II(T+U)xll;?:A(T)-llUxll>A(T)j2 and hence, A(T+U»A(T)j2. By (i)

and (ii) we have a(T+ U)=O and f3(T+ U),,;;f3(T). Since 11U11 <A(T+ U) we may
apply (ii) also to T + U and deduce that f3( T)";; f3( T + U). Consequently, f3( T) =
f3(T+ U) and thus also i(T)=i(T+ U).
The general case (i.e. a( T) > 0 but finite) reduces easily to the previous case.
Indeed, let X = Xl EB ker T and let W = Y EB ker T. Let T l : X -+ W be such that
T l1x1 =T1X1 and Tl is the identity on ker T. Then, a(Tl)=O and f3(Tl )=f3(T), i.e.
i(Tl) = i( T) - a( T). Let P be the natural projection from Wonto Yand let U: X -+ Y.
Since a(Tl)=O we have that if 11U11 < A(Tl)j2 then i(Tl + U)=i(Tl ) =i(T) - aCT).
By 2.c.7(i) the index of T+ U=P(Tl + U), as an operator from X to W, is the
same as that of Tl + U. Hence, the index of T + U, as an operator from X to Y,
is i( Tl + U) + a( T), i.e. is equal to i( T). 0
Proposition 2.c.l0. Let T: X -> Y be an operator with closed range for which
a(T) <00. Let S: X -+ Y be strictly singular. Then, a(T+S)<oo, T+S has a closed
range and i(T+S)=i(T).
Proof Let Xl e X be such that X = Xl EB ker T. Then, T 1X1 is an isomorphism.

If dim ker (T+S)=oo there would exist an infinite dimensional subspace Z of Xl
on which T + S = 0, i.e. S = - T is invertible. This contradicts the assumption that
S is strictly singular. If (T + S)X were not closed we could, by 2.cA, find an infinite
dimensional subspace Z of Xl so that II(T+S)lzll<IITixill-l. But then Sizis an
isomorphism and we reached again a contradiction.
By what we have already proved the number i(T + tS) is defined for all t E [0, 1].
By 2.c.9, i(T+tS) is a continuous function of t and thus a constant, i.e.
i(T)=i(T+S). 0
Proposition 2.c.ll. Let T: X -+ X be a Fredholm operator. There is an e > 0 so that

a(T+M) is constant on the set {A; 0< IAI";;e}.
Proof All the iterates of T have a closed range and thus Y =

n=l
n
Tn X is a closed
subspace of X (which may consist of the vector 0 only). Clearly, TYe Y; we shall
prove that TY= Y. Observe first that since a(T) <00 there is an integer k so that
TkXnkerT=TnXnkerT, for every n>k. Let yE Y and let uETkX and
v E TnX(for some n>k) be such that Tu=y=Tv. Then, u-v E TkX n ker TeTnx,
i.e. u E Tn X. Since this is true for every n > k we get that u E Y =
n=1
n
Tn X.
Let Tl =T1y. We have shown that f3(Tl )=O and thus, by 2.c.9, f3(Tl +My)=O
for A sufficiently small. By using 2.c.9 again, we deduce that i(Tl + My) = i( T l ) and
thus also a(Tl + My)=a(Tl) if A is sufficiently small. To conclude the proof we have
only to observe that, for A#O, ker (T+A/)=ker (Tl + My). Indeed, if Tx= -AX,
n rnx= Y.
00
with A#O, then X E 0

n=l
The next lemma involves spectral properties of operators and it therefore requires
that we work with a complex Banach space.
Lemma 2.c.12. Let X be a complex Banach space and let P be a projection in X. Let
S: X --J>- X be a strictly singular operator and let Q=P+S. Then, the spectrum a(Q)
of Q is countable and its only possible limit points are 0 and 1.
Proof Let C denote the complex plane. By 2.c.1 0 the operator Q - AI is a Fredholm
operator of index 0 for every AE C",{O, I}. By 2.c.ll, any two points in C",{O, I}
can be connected by a finite chain of discs such that a( Q - AI) is constant on each
disc (with the possible exception of the center of the disc). Consequently, a(Q-AI)
is constant on C",{O, I} except for a set of isolated points. This constant value is
o since Q-AI is invertible for IAI > IIQI!. Since i(Q-AI)=O it follows that also
{3(Q-AI) =0 except for a set of isolated points in e",{O, I}. D
We are now ready to prove the theorem of!. S. Edelstein and P. Wojtaszczyk
[35] and P. Wojtaszczyk [143].
Theorem 2.c.13. Let X and Y be two Banach spaces so that every operator from Y
into X in'strictly singular. Let P be a projection of X EB Yonto an infinite-dimensional
subspace Z. Then there exist an automorphism TO of X EB Y and complemented
subspaces Xo of X and Yo of Y such that TOZ = Xo EB Yo.
Proof We shall first assume that X and Yare complex Banach spaces. As every
operator from X EB Y into itself, the projection P has a natural representation as
a matrix P= (~~ ~:), where SI: X --J>- X, S2: X --J>- Y, S3: Y --J>- X and S4: Y --J>- Y.
Since S3 is strictly singular the same is true for P- Q, where Q= (~~ s~). We want
to replace Q by a projection P having a similar matricial representation (i.e. zero
in the upper right corner). We achieve this by putting
- I . f R (A, Q) dA, ,
P=-2
TTl r
r
where R(A, Q) is the resolvent of Q and is a closed simple curve which does not
intersect a(Q), has the number I in its interior and 0 and 2 in its exterior. Such a
r exists by 2.c.12.
The operator P is a projection in X EB Y s6 that a(Q'P(xffiy»)cinterior of r
and a(Q,(I-P>(Xffi Y»)c exterior of r (cf. [33, VII. 3. 11]). Moreover, itis easily checked
that P= (;~ p~), where PI is a projection in X, P 4 a projection in Yand P 2 an
operator from X into Y.
The operator 1+P- Q on X EB Y leaves P(X EB Y) invariant and is invertible
there (since (/+P- Q)'P(XffiY) = (2/- Q),P(XffiY) and 2 is in the exterior of This n.
operator also leaves (1- P)( X EB Y) invariant and is invertible on this subspace
(since I is in the interior of n.
Hence, 1+ P- Q is an invertible operator on
XEB Yand thus, by 2.c.1O, Tl =I+P-P=(I+P- Q)+(Q-P) is a Fredholm
operator of index 0 on X EB Y.
c. Fredholm Operators, Strictly Singular Operators and Complemented Subspaces of I" Ei1/r 81
We shall show next that T 1 Z is a subspace of finite codimension in P(X EEl Y)

Since T 1P=PP it follows that T l Z=T1P(X EEl Y) e P(XEEl Y). To show that TlZis
of finite co dimension in P(X EEl Y) is the same as to show that Z is of finite codimen-
sion in T1l(P(X EEl Y)) (this follows from the fact that T1 is a Fredholm operator;
in particular, it is obvious that T 1Z is closed). If dim T1l(P(X EEl Y))/Z =00 there
would exist an infinite dimensional subspace W of X EEl Y so that P1W=O, T l1W is
an isomorphism and T1 wep(X EEl Y). For WE W we have T 1w=w+Pw and thus
w=Tlw-PWEp(XEEl Y), i.e. We P(XEEl Y). On P(XEEl Y) the operator Q is
invertible (since 0 is in the exterior of r). But QIW=(Q-P)IW and, since Q-P is
strictly singular, we arrived at a contradiction.
Consider now the projection P= (~1 ~ J on X EEl Y. The next step in the proof
is to construct an automorphism T2 of X EEl Yso that T 2PTi 1=P. We put S=P-P
and notice that S2=SPS=PSP=O. Hence, S=SP-PS satisfies
It follows that T2 = 1- S is an automorphlsm of X EEl Y with Til = 1+ S and we

have
The operator T= T2T1Ti \ which is also a Fredholm operator of index 0 in

X EEl Y, maps Z into a subspace of finite codimension of P(X EEl Y) = Xl EEl Y 1,
where Xl =PlX and Yl =P4 Y. Since T is a Fredholm operator of index 0 there is
a finite rank operator R on X EEl Y so that 71 = T + R is an automorphism of
X EEl Y. There are finite dimensional subspaces Band C of X EEl Y so that
71(Z) EEl B=Xl EEl Y 1 EEl C. In particular, 71(Z)eX2 EEl Y 2 with X 2=X1 EEl Cx
and Y 2 = Y 1 EEl Cy where Cx and Cy are suitable finite dimensional subspaces of
X, respectively Y.
We recall now the fact that if V l and V2 are two closed subspaces of a Banach
space U with dim U/V1 = dim U/V2 <00 then there is an automorphism of U
which maps V1 onto V2 (the automorphism can be chosen to be the identity on
V1 n V2 since dim Vl /(V1 n V2)=dim V 2/(V1 n V2)<00).
Hence, if Xo and Yo are subspaces of finite codimension in X 2, respectively
Y 2, so that dim (X2 EEl Y 2)/71(Z) = dim (X2 EEl Y 2)/(Xo EEl Yo) then there is an
automorphism 72 of X 2 EEl Y 2 onto itself which maps 7l(Z) onto Xo EEl Yo. Since
X 2 EEl Y2 is complemented in X EEl Y, 72 can be extended to an automorphism ';2
of X EEl Y. Consequently, 70 = ';271 is an automorphism of X EEl Y so that 70(Z) =
Xo EEl Yo. Since Xo is complemented in X and Yo is complemented in Y this con-
cludes the proof in the complex case.
We point out now the modification needed in the case where the scalars are
real. If X and Yare real Banach spaces we pass to their natural "complexifications"
X, respectively f. For example, X={(u, v); u, v E X}, where we put (a+ib)(u, v)=
(au-bv, av+bu) and lI(u, v)ll=max {llau-bvI1 2 + Ilav+buI1 2 )lf2; a2 +b2 =1}. The
given projection P on X EEl Y induces in an obvious way a projection 13 on X EEl f.
The proof proceeds as in the complex case. We have onl~ to ensure that the operator
(27Ti)-1 J R(A, Q) dA, used in the proof, is of the form 15, for a suitable operator j3
r
on X EB Y. This is ensured if, e.g. we take as r a rectangle symmetric with respect
to the real axis. 0
Remarks. 1. Observe that the assumptions on X and Yin 2.c.13 are satisfied in
particular if X and Yare totally incomparable. In this case the proof 2.c.13, can
be simplified somewhat. Indeed, we could take here Q= (gl ~J and get that
- (P1
P= 0 0 4 ) • So the step of producing P~ becomes unnecessary in this case.
P
2. It follows from 2.c.13 that if Y is an arbitrary Banach space and X is lp,
1 <"p <00 or Co then any complemented subspace of X EB Y is of the form Yo or
X EB Yo for some complemented subspace Yo of Y. Indeed, if there is a non-
strictly singular operator from Y into X then, by 2.a.2, Y has a complemented
subspac;e isomorphic to X and thus Y EB X~ Y.
From 2.c.13 and 2.a.3 we get, by an obvious induction argument, the following
result
Theorem 2.c.14. Let {Xi }f'=l be distinct spaces out of the set {co} u {lp; 1 <"p<oo}.
Then any infinite-dimensional complemented subspace of Xl EB X 2 EB"'EB Xm is iso-
morphic to Xii EB X;2 EB ... EB X;"" for some 1<" i1 < i2 < ... < ik<" m.
There is no analogue to 2.c.14 if we consider general subspaces instead of

complemented subspaces. We shall discuss this situation in Section 4.d.
We shall apply now 2.c.14 to describe the structure of unconditional bases in
spaces of the form lpi EB Ip2 EB ... EB lpm'
Theorem 2.c.15 [35]. Let {X;}f'=1 be distinct spaces out of the set {co} u {lp; 1 <"p <oo}.
Let {Zn};'=l be an unconditional basis of Xl EB X 2 EB ... EB X m • Then we can partition
the integers N into m disjoint sets {Ni}f'=l such that [Zn]neNI is isomorphic to X;,
i=I, ... , m.
This theorem reduces the question of describing the unconditional bases in

m
the direct sum 2: EB IPI to that of describing the unconditional bases in each com-
i= 1
ponent. In particular, we get by 2.c.15 and 2.b.9 (cf. also the remark preceding it)
the following result.
Theorem 2.c.16. Let Z be one of the spaces, 11 EB 12 , Co EB 110 Co EB 12 or Co EB 11 EB 12 ,

Let {Z~};'=l and {Z~};'=l be two normalized unconditional bases ofZ. Then there is a
permutation 7T of the positive integers such that {Z~};'=l is equivalent to {z~(n)};'=l'
The proof of 2.c.15 is by induction on m. The argument needed in the general

c. Fredholm Operators, Strictly Singular Operators and Complemented Sunspaces of Ip (B lr 83
induction step is very similar to that used in the case m=2. We present here only
the proof for m = 2. Also, there is some difference between the reflexive and non-
reflexive case; we shall treat here only the reflexive case. The main step of the proof
consists of the following lemma.
Lemma 2.c.17. Let I<p<r<oo and let zn=(xmYn),n=I,2, ... be a normalized

unconditional basis of Ip EB IT with unconditional constant K (lp EB IT is normed by
II(x,y)II=llxll+llyllfor xElp,YEIT)· Let {n"Jk'=l be a subsequence of the integers
so that cx=sup Ilxnkll <K- 1 • Then, [znk]k'=l is isomorphic to lr.
k
Proof The proof we give here is very similar to that of 2.a.II. We claim first that,
00
for every sequence of scalars {Ak}k'=1 such that L: AkYnk converges unconditionally,
k=1
the series '"
L AkZnk converges too. To prove this let {a~},':'=l and {b~},';'=1 be scalars
k=1
so that
(xnk' 0) = 'L'" a~zm

n=l
(0, Ynk) = 'L'" b~zn'
n=1
k=I,2, ....
Then, a~k +b~k= 1 and, since la~J ::::;Kllxnkll, we get that b~k;?;: l-Kcx>O for every k.
By Remark 2 following I.c.8 we deduce that
00
Thus, if {Ak}k'=1 is such that L AkYn/c

k=l
converges unconditionally the series
~ Akb~kznk' and therefore also ~ AkZnk , converge too.

k=1 k=l
If [znk]k'=1 is not isomorphic to IT it would follow from 2.c.14 that [znk]k'=1 con-
tains a subspace isomorphic to lp. Hence, by I.a.9, there is a normalized block
basis Wj= L cxkznk,j=I, 2, ... of {znk}k'=1 which is equivalent to the unit vector
kE(Jj
basis of Ip (here {aj}~ 1 denotes a sequence of disjoint finite subsets of the integers).
Let Q be the natural projection from lp EB IT onto IT and let {Pm};:'=1 be the projec-
tions on lr which are associated to the unit vector basis. Since Zn ~ 0 we get that,
for every m, lim IIPmQ![Zk1r=nll =0. Hence, by passing to a subsequence if necessary,
n-> 00
we may assume that there exists a sequence {mj}~1 of integers so that

II(Pmj+1-Pmj)Qz- Qzll::::;2- j llzll, for every Z E B j = [znkhE(Jj' In particular, if
Uj E B j for j= 1,2, ... , s then
Let{,8j}~1 be a sequence of scalars SQ that ~ l,8jl' <00 and let {8k}k'=1 be a sequence
j=1
84 2. The Spaces Co and /"
of signs. PutUj= L 8/ca/czn/<'j=I,2, .... Then,

/ceCIl
lIujll~Kllwjll~Kandhence,
<Xl
In other words, L L {3ja/cYn/< converges unconditionally.

j=l/ceClj
By the first part of the
<Xl co
proof this implies that L {3iWj= L L {3ja/czn/< converges. This however contra-
j=l i=l/ceClI
dicts the assumption that {Wj}t,l is equivalent to the unit vector basis of 11' and
thatp<r. D
.Lemma 2.c.18. Let l<p<r<oo and let Zn=(Xn,Yn), n=l, 2, ... , be a normalized
unconditional basis of 11' EEl/r. Let {n/c}k'=l be a subsequence of the integers so that
IIxn/<1I ~a,for some a>O and every integer k. Then [Zn/<]k'=l is isomorphic to 11"
Proof. Every subsequence {nnk'=l of {n/c}k'=l has, by l.a.I2 and 2.a.I, a sub-
sequence {nk}k'=l such that {Xn~}k'=l and {YnkiIlYn;;;II}k'=l are equivalent to the unit
vector bases of 11" respectively Ir (if Yn~ = 0 the situation is even simpler). For such
a sequence {nk}k'=l the basic sequence {Zn;;}k'=l is also equivalent to the unit vector
basis of 11" It follows from this remark and 2.c.14 that [Zn/<]k=l is isomorphic to
either 11' or 11' EEllr.
Assume that [zn/<]k'= 1 is isomorphic to 11' EEl/r and let z:/< = (Y:/<, x:/<), k = I, 2, ...
be the functionals biorthogonal to zn/< in (/1' EEl/r)*=I: EEl r:.Since [Z:/<]k=l~
([zn/<]k'=l)*~/r* EEl I: it follows by 2.c.17, that lim sup lIy:/<1I >0 and hence, by the
/c
first part of this proof, there is a subsequence {n~}k'=l of {n/c}k'=l such that [Z:;.]k'=l
is isomorphic to Ir*. Hence, [Zn~]k=l is isomorphic to Ir. Using again the first part
of this proof we arrive at a contradiction ({Zn~}k'=l cannot have a subsequence
equivalent to the unit vector basis of 11')' D
It is now obvious how to conclude the proof of 2.c.15 (for 11' EEllr with
I <p<r<oo). We simply take N1={n; II Xn II > 1/2K} and N 2 ={n; IIxnll~I/2K}.
Then, [zn]neNl is isomorphic to 11' and [zn] neN2 is isomorphic to Ir. D
A stronger version of 2.c.15 was proved recently by P. Wojtaszczyk [143].
We conclude this section by mentioning (an admittedly vaguely stated) prob-
lem arising from 2.c.16,
Problem 2.c.19. Describe all the separable Banach spaces which have, up to equiva-
lence and to a permutation, a unique normalized unconditional basis.
d. Subspaces of Co and lp and the Approximation Property,

Complementably Universal Spaces
This section is devoted mainly to the study of subspaces of Co and 11" I <p <00. We
present first a result of W. B. Johnson and M. Zippin on the structure of subspaces
d. Subspaces of Co, /p and Approximation Property, Complementably Universal Spaces 85
of Co or lp, 1 <p<oo having a shrinking basis. The main part of this section is
devoted to Davie's proof of the fact that Co and lp, 2 <p <00 have subspaces which
fail to have the approximation property. The section concludes with some general
results concerning universal spaces and the approximation property.
We start with the result of Johnson and Zippin [63]. This result gives actually
more information than indicated in the preceding paragraph. It describes even the
structure of subspaces of quotients of Co or lp, 1 <p <00 and the assumptions con-
cerning the existence of a basis can be somewhat relaxed. Before stating this result
let us observe that the notions of a quotient of a subspace and that of a subspace
of a quotient coincide. Let X be a subspace of a quotient space Y of Z and let
T: Z -+ Y be the quotient map. Then X is a quotient of the subspace T -1 X of Z.
Conversely, assume that X is a quotient space of a subspace Y of Z and let T: Y -+ X
be the quotient map. Then X is isometric to Y/ker T which is a subspace of the
quotient space Z/ker T of Z.
Theorem 2.d.1. Lt:t {Bn}:'=1 be a sequence offinite-dimensional Banach spaces. Let

X be an infinite dimensional subspace of a quotient space of Ct EB Bn)p' 1 <p<oo
( resp. (~
n=1
EB Bn) ) having a shrinking F.D.D. Then X is isomorphic to (~ EB Dk)
0 k=1 p
(resp. t~1 EB Dk)J,for a suitable sequence offinite dimensional spaces {Dk}k'=l'
Proof. We prove the theorem first in the case 1 <p<oo. Let {Ck}k'=l be an F.D.D.
of X and let T: C~l EB Bnt -+ Y be a quotient map, where Y::::l X. By 1.g.4(b)
there exists a constant K, a blocking {B~}:'=l of {Bn}:'=l and a blocking {Ck}k'=l
of {Ck}k'=l so that, for every x E Ck, there is a u E Bk EB Bk + l with Ilull~Kllxll
k=l
00
so that IITu-xlI~2-kllxll. Let X= ~ Xk be an element in X with Xk E Ck, for
k=l
00
every k, and assume that ~ IlxkW<oo. For k= 1,2, ... let Uk E Bk EB Bk + l satisfy
IIUkll~Kllxkll and IITuk-Xkll~2-kllxkll. Put U=Ul+U3+U5+ .. ·. Then,
lIull=(Jlllu2i-lWY'P ~Kt~11IxkWY'P. Since IITu-Jl X2i-lll~k~l 2- kllxkll we
get that IIJl X2i-lll~ lIull+ k~l 2- kllxkll and thus IIJl X2i-lll~Mtt IlxkWY'P,
for a suitable constant M.The same calculation shows that this estimate holds also
for IIJl X2il and thus
In order to obtain an estimate on Ilxll from below we pass to the dual. The space
X* is a subspace of a quotient space of (I
k=l
EB (B~)*) , where q-l+p-l= 1. By
q
applying 1.g.4(b) once more and repeating the argument above it follows that
there is a blocking {D k};:'=l of {C~};:'=l so that, whenever x*= ~ x~ with x~ EDt,
k=l
k= 1,2, ... , we have
(**) Ilx*11 ~M'(~l IlxtW) l/q
for a suitable constant M'. Clearly, (*) remains true if we consider decompositions
00
with respect to the blocking {D k};:'=l, i.e. if X= I Xk with Xk E DIe> k= 1,2, ...
k=l
(see the remark following 1.g.4). From (*) and (**) it follows that X is isomorphic
to CJl EB Dkt
The proof for the case (Jl EB Bnt is simpler. In this case it is enough to do
the first step. The inequality t
(*) reads Ilk Xkll ~ 2M s~p Ilxkll while the inequality
~ xkl/, ~ sup
II k-l Ie
Ilxkll/M ', for some M', follows from the fact that {Ck};:'=l, and thus
{C~};:'=l' are F.D.D.'s. We remark that only in this case is the assumption that
{Ck};:'=l is shrinking a real restriction. In the case 1 <p<oo this assumption holds
automatically since X is reflexive. 0
Remarks. 1. Theorem 2.d.1 is no longer true if p = 1. This is completely obvious

if we consider subspaces of quotients or even only quotients since every separable
Banach space is a quotient space of 11' If we consider only subspaces of 11, Theorem
2.d.1 is valid for a trivial reason-no infinite dimensional subspace of 11 has a
shrinking F.D.D. Without the requirement that the F.D.D. be shrinking the
theorem fails even for subspaces. We mentioned already in Section 1.d that
{en - (e2n + e2n + 1)/2},';'= 1 is a basic sequence in 11 whose span is not isomorphic to
a conjugate space and thus is not of the form C~l EB D n )l'
2. It is unknown whether the assumption of the existence of a F.D.D. can be
replaced by the weaker assumption that Xhas the approximation property. This is
a special instance of Problem 1.e.12.
Theorem 2.d.1 does not answer of course all the natural questions on subspaces
of Ip, 1 <p <00 or Co, even on those subspaces which have a basis. For example,
the answer to the following problem is unknown even under the assumption of the
existence of a basis.
Problem 2.d.2. Let X be an infinite-dimensional Banach space and let 1 < p < 00 be
such that X is a subspace, as well as a quotient space, of Ip. Is X isomorphic to Ip?
We pass now to the construction of subspaces of Co and Ip, 2 <p <00 which do
not have the approximation property. We reproduce here the proof due to A. M.
d. Subspaces of Co, lp and Approximation Property, Complementably Universal Spaces 87
Davie [20, 21]. Our approach will be to construct an infinite matrix of the type
appearing in l.e.8 (this is the main step) and then deduce from the existence of
such a matrix the existence of the desired subspaces.
Theorem 2.d.3. There exists an infinite matrix A =(ai.j)i~j=1 of scalars such that for
00
every i, ai, j # 0 only for finitely many indices j, L (max

i=1 j
lai, jlY <00 for every r > 2/3,
00
A2=0 andtraceA= 2:
i= 1
ai i#O.
'
Proof We shall work with complex scalars since this is somewhat more convenient.
For every k=O, 1,2, ... we let Uk be a unitary matrix of order 3·2 k (the specific
2(k+l)/2P )
choice of Uk will be made later). We partition Uk as follows ( 2k/2Qk k = Uk,
where 2(k+1)/2Pk is the 2k+1 X 3 ·2k matrix consisting of the first 2k+1 rows of Uk
and 2k/2 QTe is the 2k x 3 ·2k matrix consisting of the last 2k rows of UTe' Since
UkUZ=I3.2k (where 1m denotes the identity matrix of order m) we get that
k=O, 1,2, ....
Consider now the matrix
A=(ai. j )
ptpo PtQl 0 0 0
-QUo PUI-QrQl PtQ2 0 0
0 -Q~Pl P:P2-Q~Q2 P:Q3 0
0 0 -QtP2 ptp3- QtQ3 PiQ4
It is easily verified that A2=0. Clearly, trace (PZP k - Q;Qk)=O for k= 1,2, ...
and thus trace A = trace Pcipo = 1. We shall construct the Uk'S so that each element
in the k'th block of rows in A (i.e. each element of the matrices - Q;Pk - 1 ,
PZPk-Q;Qk andPZQk+l) is ~C(k+1)1/22-3k/2, where Cis a constant. Since the
k'th block contains 3· 2k rows this will imply that, for r> 2/3,
00 00
i~ (m~x ladr~k~O 3·2TcC"(k+1Y/22-3rk/2<00.
Observe that, since PZQk+l =(QZ+IPk)*, it is enough to examine only the

matrices Q;Pk - 1 and pZPk - Q;Qk' For the construction of the {Uk}k'=o we need
two lemmas.
Lemma 2.d.4. (a) Let {Ctj}7=l be complex numbers and let {ej}7=l be independent
88 2. The Spaces Co and /1'
random variables each taking the values +1 and -1 with probability 1/2. Then
there is an absolute constant K so that
(b) The same assertion as (a) with the only difference that each OJ takes the value
2 with probability 1/3 and -1 with probability 2/3.
Proof We shall prove assertion (a) only; the proof of (b) is very similar. By con-
sidering separately the real and the imaginary parts it follows that it is enough to
n
prove (a) for real {aj}7=l. Also, there is no loss of generality to assume that L a~= 1.
j=l
We write 8=(81,82 , ••• , 8n) and f(8)=IJ1 Ojajl. We denote by E the expectation
with respect to the probability distribution of 8 (i.e. the average over all 2n possible
choices of 8). Then, for every A> 0,
E(ei\f(O»:(, E( it Of a, + e -;..p,a,) =2 fI (eAaf + e-;..a )/2 .

j=l
f
The desired result follows by taking, e.g. A=(3log n)1/2. 0
Lemma 2.d.S. Let G be an Abelian group of order 3· 2k for some integer k. Then it
is possible to divide the characters ofG into two sets: one consisting of2 k + 1 elements,
denoted by {'Tj}1~;\ and another consisting of 2k elements, denoted by {aj}~~h so
that for every g E G
where L is an absolute constant.
Recall that a character y of an Abelian group G is a homomorphism from G

into the multiplicative group {z; JzJ = I} in the plane. An Abelian group of order m
has exactly m characters and any two different characters are orthogonal, i.e.
L y(g)y'(g) =0 if y#y'. For our purpose it is enough to consider a cyclic group G
gEG
of order m. In this case the characters of G are given by Yk(gb) = e - 2nik!/m, 0:(' k < m,
0:(' 1< m, where go is a generator of G.
Proof Let {yj}Y3~ be an enumeration of all the characters of G. By 2.d.4(b) there

is a choice of {OJ}7;:~ such that OJ is either 2 or -1, for every j, and so that, for
some constant L,
I~t: Bffig)I::::;;L(k+ l)l/22k12, for every g E G.
The number of these inequalities is n= 3 ·2k and thus such {OJ}7=l exist, by 2.d.4(b),
whenever n<n 3/K, i.e. for n>no=VK. (For n<no there is nothing to prove.)
By taking, in particular, g to be the identity element of G we get that
I;t:Ojl::::;;L(k+ 1)1/22kI2. Thus, by changing if necessary at most 2L(k+ 1)1/22kI2 of
3·2 1c
the OJ and replacing L by 2L, we may assume that L Bj=O, i.e. exactly 2k+1 of
j=l
the OJ are equal to -1 and 2k of the OJ are equal to 2. Those Yj for which the corres-
ponding OJ are equal to -1 are dep.oted by h}7~;1 and the rest by {ajH~l' From
the choice of {OJ}7;:~ it is clear that this partition of the characters into two sets
has the required 'properties. 0
We return to the proof of 2.d.3. We make the set {I, 2, ... , 3· 2k} into an Abelian
group G ,c and define the T~ and aJ as in 2.d.5. We let the rows of P'C be
3- 1/22-<2k+1l/2TJ, 1::::;;j::::;;2k +1 (i.e. the entries of P k are 3- 1/22-<2k+ll/2TJ(i), where
i E Gk , i.e. 1::::;; i::::;; 3 ·2k) and let the rows of Qk be 3- 1I22- k BJaJ, 1::::;;j::::;; 2k , where the
OJ are either + 1 or -1 and will be determined below. Whatever is the choice of
{BJ};~l the matrix Ule defined in this manner is unitary (this follows from the
orthogonality property of the characters).
We now write down explicitly the elements of QZPIe -1 and pZPk - QZ QIe' They
are
2k
(1) 3-1·21/2-21e L BJaJ(h)TJ-l(g), hE Gk , g E GIe - 1
j=l
In the derivation of (2) we used the fact that the TJ and aJ are characters, i.e. that
TJ(h 1 0 h 2 ) = TJ(h 1 )TJ(h 2), where hI 0 h2 denotes the multiplication in G k • We have
to show that, for suitable choices of {BJH~l' k=O, 1,2, ... all these terms are in
absolute value ::::;; C (k + 1)1/22 - 3k/2. The expressions appearing in (2) are inde-
pendent of BJ and they are of the right order of magnitude by Lemma 2.d.5.
Observe that the number of terms written in (1) is 3·21e - 1·3·21e ::::;;5·n2 , where
n=21e. Hence, by Lemma 2.d.4(a), we infer that, whenever 5.221e < 231e/K, the
{BJH~l can be chosen so that all terms in (1) are in absolute value of the required
order of magnitude (for the finitely many k for which 5· 22k > 231e/ K there is of
course nothing to prove). This concludes the construction of {Uk }r=l and thus also
the proof of 2.d.3. 0
The matrix A, constructed in 2.d.3, satisfies in particular L max lad <00. Thus,
i j
by 1.e.8, the rows of A span a subspace of Co which does not have the A.P. The
fact that the matrix of 2.d.3 has a stronger property enables us to show, by the
same argument which was used in the proof of 1.e.8, that also the spaces Ip, for
2 <p <00 have subspaces which fail to have the A.P.
Theorem 2.d.6. The spaces Co and Ip, 2<p<00, have subspaces which do not have
the approximation property.
Proof. We have to prove the theorem only in the case of Ip. Let A be a matrix satis-
fying the requirements of 2.d.3 (the assumption that for every i only finitely many
j satisfy ai,j#O will not be used here). Put Ai=max lad, I::S;i<oo and
J
bi,j=(Aj/Ai)1/(P+1)ai,j,I::S;i,j<00. The matrix B=(b i,j)t:j=1 satisfies B2=0 and
trace B=trace A #0. Let Yi =(bi ,1, bi , 2, ... ). Then, since p/(p+ 1) > 2/3, Yi E Ip and
for some constant L. Consequently, IllYil1 <00. Denote by {ei}t~1 the unit
i= 1
00
unit vector basis of lq. For every Y in [ydt-;"1 c Ip we have L Yiei(Y)=O,

i=1
00
but L ei(Yi) = trace B # 0. Hence, by 1.e.4, the space [YdY'= 1 does not have the
i=1
A.P. 0
Remarks. 1. Since every subspace of 12 has the A.P. the argument used in the
preceding proof shows that if A=(ai. j)i:j=1 is a matrix such that A2=0 and
co
L (sup lad)2/3<00 then trace A=O. Thus, the exponent 2/3 appearing in 2.d.3,
i=1 j
is the smallest possible one.

2. In vol II we shall present a proof of the fact that the spaces lp for I::S;p < 2
also have subspaces which fail to have the A.P.
3. In 1.e.8 we presented a result of Grothendieck which shows that the approxi-
mation problem has also an equivalent formulation in terms of continuous functions
K(s, t) on [0, 1] x [0, 1]. We can now give a precise answer also to this problem.
By using the proof of (iii) => (ii) of 1.e.8 and the matrix given in 2.d.3 we obtain
a function K(s, t) on [0, 1] x [0, 1] which satisfies a Lipschitz condition of every
1 1
order < 1/2 and so that J K(s, t)K(t, u) dt=O, for every sand u, while J K(t, t) dt#O.
o 0
On the other hand it can be shown that there is no function which satisfies a
Lipschitz condition of order 1/2 and has these properties.
It is not easy to construct an infinite dimensional subspace X of Ip, for 2 <p <00,
which is not isomorphic to /p. In fact, prior to the work of Davie [20] and Figiel
[40] who exhibited such X which fail to have the A.P., no such examples were
known. Once we know of the existence of such X it is possible to show that there
are "many" such examples. We show first that, for 2 <p <00 (and, in fact, for every
d. Subspaces of Co, tp and Approximation Property, Complementably Universal Spaces 91
1 ~p <00, p of 2), there is an infinite dimensional subspace X of lp so that X has the

A.P. but it is not isomorphic.to lp.
Proposition 2.d.7. (a) Let 1 <p<oo and let Y be a Banach space which is not iso-
00
morphic to a complemented subspace of Lp(O, 1). Assume that Y = U Bn with

n=l
B ncBn+ 1 and dim Bn <00 for every n. Then, C~l E8 BnL is not isomorphic to lp.
(b) For every 1~p<00,pof2 there is a subspace of lp which is isomorphic to
C~l E8 BnL, for suitable finite dimensional {B n};:'=l, and which is not isomorphic
to lp itself.
Outline afproof The proof uses arguments whose natural setting is the local theory
of Banach spaces. These arguments will be discussed in detail in Vol. IV; here we
only outline them. If C~l E8 BnL is isomorphic to lp then there is a'\ and operators
Un: Bn -i>-lp, Vn: lp -i>- Bn, n = 1, 2, ... so that VnUn = IBn and l\UnllllVnll ~,\, for every
n. Using this fact and a compactness argument we get that Y must be isomorphic
to a complemented subspace of LiO, 1), contradicting the assumption in (a).
To prove part (b), let first 2<p<00 and let Ybe a subspace of Ip which does
not have the A.P. Let {Yn};:'=l be a sequence which spans Yand put Bn= [Ydf=b
n = 1,2, .... It is easily seen that (I
n= 1
E8 Bn) is isomorphic to a subspace of
p
lp
which, by part (a), is not isomorphic to Ip itself.
For 1 ~p < 2 we take p < r < 2 and apply part (a) to Y = Ir• It will be shown in
Vol. II that C~1 E8 l~) p is isomorphic to a subspace of lp but that IT is not isomorphic
to a complemented subspace of Lp(O, 1). Thus, (Jl E8l~L is not isomorphic to
lp. 0
The examples constructed for 1 ~p < 2 have an unconditional basis. Also for
2 <p <00, lp has a subspace with an unconditional basis which is not isomorphic to
lp itself. This is a consequence of Szankowski's result [138] that a certain lattice
does not have the A.P. We shall discuss this fact in detail in Vol. IV.
We conclude this section with some results involving universal spaces. These
results show that there are "many" spaces which fail to have the A.P. even if
we consider only subspaces of lp, for some fixed 2 <p <00. It is well known that
C(O, 1) is a universal Banach space in the sense that every separable Banach space
Y is isometric to a subspace of C(O, 1). In general, Y is not isomorphic to a comple-
mented subspace of C(O, 1). The question whether there is a separable Banach
space X which contains isomorphic copies of all separable spaces as complemented
subspaces turns out to be closely related to the approximation property. The
following two theorems answer this question.
Theorem 2.d.8 [118, 66]. There is a Banach space X having a basis such that every
92 2. The Spaces Co and Ip
separable Banach space having the B.A.P. is isomorphic to a complemented subspace

ofX.
Theorem 2.d.9 [62]. There is no separable Banach space X so that every separable
Banach space Y is isomorphic to a complemented subspace of X. Moreover, such an
X fails to exist even if we consider only those separable spaces Y which have the A.P.
or, alternatively, all the subspaces Y of lp which fail to. have the A.P. for any fixed
p,2<p<00.
Since for every sequence {Yn}:'=l of separable Banach spaces there is trivially a
separable Banach space X which contains all the Yn as complemented subspaces
(e.g. X= Ct EB Yn) J, the last statement in 2.d.9 asserts in particular that, for
every 2 <p <00, there are uncountably many mutually non-isomorphic subspaces
of lp all of which fail to have the A.P.
We prove first Theorem 2.d.8. This theorem is an immediate consequence of
Theorem l.e.13 and the second assertion of the following theorem due essentially
to Pe1czynski [117].
Theorem 2.d.l0. (a) There exists a separable Banach space U1 having an uncon-
ditional basis {Xi}i~l such that every unconditional basic sequence (in an arbitrary
separable Banach space) is equivalent to a subsequence of {X;}l~ l'
(b) There exists a separable Banach space U2 having a Schauder basis {Xn}:'=l
such that,for any basic sequence {Yk}f= b there is a subsequence {nk}f= 1 of the integers
°
such that {Yk}f=l is equivalent to {x"k}f=l and the natural projection P on [X"Ic]f=i
(defined by PXnlc = X"Ic' k = 1, 2, ... and PXn = if n ¢ {nk}f= 1) is bounded.
The spaces U1 and U 2 are determined uniquely, up to isomorphism, by the pro-
perties appearing in (a), resp. (b).
Proof. We shall present a proof due to Schechtman [131] which is considerably

shorter than the original proof. We start with the construction of U1. Let {U n}:'=l
be a sequence which is dense in C(O, 1). We introduce a norm 111·1111 in the space Uo
of all sequences x=(a1' a2, a3,"') of scalars which are eventually 0, by putting,
" 0, ... ), n = 1, .... The sequence {xn}:'=1 is an unconditional basis

Let x" = (0, 0, ... , 1,
of the completion U1 of Ua with respect to 111·llk Let {Yk}f=l be any unconditional
basis of a separable Banach space. By the universality of C(O, 1) we may assume
that YTc E C(O, 1), for every k. Let K be the unconditional constant of {Yl,}f=l'
Choose integers {n k}f=l so that IIYTc-unlcl I:;;; IIYkll/K2k+z for every k. By l.a.9
the sequence {Unlc}f=l is an unconditional basic sequence equivalent to {Yk}f=l'
From the definition of 111·1111 it follows that {xnlc}f=l is equivalent to {u n1Jf=1 and
thus to {YTc}f=l'
We pass now to the proof of the existence of U z . Without the requirement of the
existence of P we could have proceeded in a way very similar to that used in the
preceding case. In the case of an unconditional basis the natural projection is of
course defined and bounded for every subsequence {nk}k'=l. The main point in our
construction now will be that, in spite of the fact that we deal with a conditional
basis, the natural projection on a large collection of subsequences of the integers
(called "branches") will be bounded.
Let {U n}:'=1 be a sequence which is dense in C(O, 1). Let t:p be a one to one
function from the set of all finite sequences of positive integers onto the positive
integers so that t:p(i1, i2 , ••• , ik) < t:p(i1, i2 , ••• , ik, i k +1) for every choice of k and
{ij}J,;; f. A subsequence of the integers is called a branch if it has the form
for a suitable choice of integers {ij}f'= 1. The set of all branches will be denoted by!JI.
For n=t:p(i1, i2 , .•• , i k ) we put Vn=Uik. We define now a norm 111·1112 on the space
Uo of sequences of scalars x = (al> a2, as, ... ) which are eventually 0, by putting,
° n
where XB(j) is 1 if j E B and is if j 1= B. Let Xn = (0, 0, ... , 0,1,0, ... ), n= 1,2, ....
Then {Xn}:'=l forms a basis of the completion U2 of Uo with respect to 111·1112.
It is clear that if B={nk}k'=l E!JI then the natural projection from U 2 onto [Xnk]k'=l
is bounded (in fact, it has norm 1).
Let {Yk}k'=l be a basic sequence in C(O, 1) and let Kbe its basis constant. Choose
integers {ik}k'=l so that Ilu ik - Ykll ~ IIYkll/2k+2K, and let nk = t:p(i1 , i 2, ... , i k), k= 1,2, ....
Then, {nk}k'=l E!JI and, by l.a.9, {Yk}k'=l is equivalent to {U1k}k'=l> i.e. to {Vnk}k'=l.
It follows immediately from the definition of 111·1112 and the fact that the inter-
section of two different branches is an initial segment of these branches that
{Xnk}k'=l is equivalent to the basic sequence {Vnk}k'=l and thus to {Yk}k'=1. This
proves that U 2 has the desired properties.
We prove now the uniqueness of U1 and U 2 • We actually prove the uniqueness
in the following stronger sense: every separable Banach space which has an un-
conditional basis (respectively, a basis) and which contains isomorphic copies of
all separable spaces with an unconditional basis (resp. with a basis) as comple-
mented subspaces must be isomorphic to U1 (resp. U 2 ). We present the proof in
the case of U1 (the proof for U2 is identical). Let V be another space which has the
above mentioned property of U1 • Then, there exist Banach spaces X and Y such
that V~ U1 EB X and Ul~ V EB Y. Also, there is a Banach space Z so that
Ul~(Ul EB U1 EB .. ·)2 EB Z and thus,
Similarly, V EB V~ V. Thus, we get that

Remarks. 1. Some further interesting properties of the space U1 will be presented

in Section 3.b.
2. In connection with 2.d.1O (a) it is worthwhile to mention the following result
of Schechtman [131] concerning unconditional basic sequences in lp. Let 1 < P < 00 ;
then, there exists an unconditional basic sequence {Xn}~=1 in lp so that every uncon-
ditional basic sequence in Ip is equivalent to a subsequence {Xn,Jk'=1 of {Xn}~=1' The
proof of this fact is based on 2.d.1.
We shall now give a partialproofof2.d.9. We have seen that for every 2<p<00
there is a subspace Yp of Ip which does not have the A.P. Actually, the spaces
obtained in Davie's construction fail to have an apparently weaker property
namely, the compact approximation property (C.A.P. in short). A Banach space Y
is said to have the C.A.P. if the identity operator on Y is in the closure of the set
of compact operators from Y into itself with respect to the topology T of uniform
convergence on compact subsets of Y. That the spaces Yp fail to have the C.A.P.
is evident from the argument in [20] but it is not as apparent from the approach
presented in this section via infinite matrices. In [20] Davie defines a sequence of T
continuous linear functionals f3n( T), n = 1,... so that f3( T) = lim f3n{ T) exists for
n
every T and is a T continuous linear functional. The {f3n}~=1 satisfy f3n(I) = 1 for
every n while lim f3n{ T) = 0 for every compact operator T: Yp -.,.. Yp. Thus, f3 is
n
a T continuous functional which vanishes on the compact operators and is equal
to 1 on the identity operator.
Assume now that X is a separable space such that, for every 2 <p <00, there is
a subspace Zp of X which is isomorphic to Yp and so that there is a bounded linear
projection Qp from X onto Zp. There is an uncountable subset A of (2, (0) and
a .:\<00 so that IIQpll~ Afor every pEA.
Since each Zp does not have the C.A.P. it follows that there are finite sets
{z" p}f~i of unit vectors in Zp and an ep> 0 so that, whenever T is a compact
operator on Zp for which IIZi.P-Tz1,pll <epfor l~i~n(p), then IITII >.:\2. Let B be
an uncountable subset of A so thatn(p) is constant (say =n) on Band inf ep=e>O.
pEE
Since B is uncountable and X is separable there exist p < r in B so that
IIZI,p-Zi,rll < {.:\+.:\2)-18 for 1 ~i~n.
The proof of 2.c.3 shows also that every operator from a subspace of Ir into Ip
is compact. Thus, every operator from Zr to Zp is compact. In particular, T= QrQplZr
is a compact operator from Zr into itself with IITII ~ .:\2. Note that, for 1 ~ i~n,
IIZi,r-Tzl,rll = IIZI,r- QrQpzi,rll ~ IIZi,r- Qrzi,pll + II QrZt,p- QrQpzi,rll

= II Qr(Zi, r - Zi, p)11 + II QrQp{z;, p- z;, r)11
~{A+.:\2)IIz;,r-Zi.pII <8,
but this contradicts the choice of {Zi,r}f=1' This proves the first assertion in the
statement of 2.d.9. The proofs of the other two statements of 2.d.9 are more com-
plicated and we do not reproduce them here (we refer the reader to [62]). We just
remark that the proof of the fact that there is no separable space which contains
e. Banach Spaces Containing /p or Co 95
isomorphic copies of all separable spaces having the A.P. as complemented sub-
spaces is based on the construction presented in l.e.20 of a space which has the
A.P. but fails to have the B.A.P. The proof of the fact that there is no separable
space which contains isomorphic copies of all subspaces of lp (for a fixed 2 <p <(0)
which fail to have the A.P., as complemented subspaces is obtained by modifying
Davie's construction. (Instead of constructing one space the same method is used
to construct a suitable uncountable family of spaces.)
e. Banach Spaces Containing lp or Co
A long standing open problem going back to Banach's book was the following:
Does every infinite-dimensional Ban.ach space have a subspace isomorphic to either
Co or lp, for some 1 ~p <<Xl? For the common examples of Banach spaces it was
easy to show that the answer is positive while, in general, the problem seemed to be
quite difficult. Therefore, it came as a surprise when a rather simple counter-
example was constructed by B. S. Tsirelson [140]. Tsirelson constructed an example
of a reflexive space with an unconditional basis which contains no isomorphic
copy of any lp space, 1 <p<oo (actually, his example contains no uniformly con-
vexifiable subspace). We shall present here the dual of Tsirelson's original example
which also solves the question stated above. In our presentation we follow T.
Figiel and W. B. Johnson [42].
Example 2.e.1. There is a reflexive Banach space T with an unconditional basis which
contains no isomorphic copy of any lp space, 1 ~p <<Xl.
Proof. We start by defining a sequence of norms on To, the space of all sequences
of scalars which are eventually zero. We denote by {tn}~=l the unit vector basis of
co
To and set, for x= 2: antn ETa,
n=l
Ilxllo=max lanl, and for m;;:,O
n
where the inner max is taken over all choices of k~Pl <P2 < ... <Pk+l, k= 1,2, ....
Obviously, Ilxll = lim Ilxll m exists for all x E To and defines a norm on To. The
__ n~ 00
unit vectors {tn}~=l form a normalized unconditional basis of the completion T

of To. It is also clear that
Ilxll = max {m:x lanl, 2- 1 sup C~l:~:l antnll'

k~Pl <P2 < ... <Pk+l> k= 1,2, ... )},
96 2. The Spaces Co and II'
CD
for every x= ~antn E T. Consequently, for any k and any sequence of k normal-
n=l
1'J+1
ized blocks Uj= ~ antm I~j~k with k~Pl <P2<···<Pk+1, we have
n=pJ+1
for every choice of scalars {Cj}7=1. This fact and l.a.12 show that T contains no
subspace isomorphic· to Co or to some II" I <p<oo.
To show that T contains no subspace isomorphic to h is more difficult. To this
end we have to apply a result of R. C. James (cf. Proposition 2.e.3 below) which
ensures that if T has a subspace isomorphic to h then there is a normalized block
basis {Vj}j"=o of {tn}'':'=l so that, for every choice of scalars {btl;;,o,
and, in particular,
Consider now integers k~Pl <P2<···<Pk+1 and let {P1}7=1 be the projections
associated to the basis {tn}:=l so thatPjtn=tn ifp1<n~p1+1 and PJtn=O, otherwise.
Let no be the largest integer for which tn belongs to the support of Vo. If k ~ no then
k k
L IIPivo+r- 1(v1 +···+vT»11 = J=l
1=1
L IIPJr-1(V1 + .. ·+VT)II~2.
If k<no we set
8={i; IIPJVt II :;60 for at least two values ofj} ,

u={i; IIPJvtll:;60 for at most one value ofj}.
Then, since 8 has at most k -I elements, we get that

k
L IIP;(vo+r- 1(v1 +···+vT»11
J=l
~ j~ IIP1voll+r-1(~ J~ IIPjvtll+t~ J~lIIPjvtll)

~211voll +r-1(2 j~ Ilvdl + &:" Ilvdl)
~2+r-1(2(k-I)+r-k+ 1)
~3+(k-l)r-l~3+(no-l)r-1 .
Therefore, by taking r ~ 2no, we get that
2:
k
IIPlvo +r- 1(v1 + ... + v II~ 7/2.
r ))
;=1
It follows from the definition of the norm in T that IIvo+,-1(v1 +"'+vr)II~7/4

and this contradicts (!).
Since T has an unconditional basis but does not contain subspaces isomorphic
to Co or h we deduce from l.c.12 that T is reflexive. 0
Remark. The relation (*) actually shows that T has no uniformly convexifiable
infinite-dimensional subspace. Figiel and Johnson show in [42] how to modify T
so as to obtain an example of a uniformly convex space with an unconditional basis
containing no isomorphic copy of any lp; 1 ~p <00 (we shall discuss this matter in
Vol. II).
Closely related to the problem with which we started this section and which
was solved by Tsirelson's example is the problem whether a space which contains
an isomorphic copy of some lp must actually contain almost isometric copies of
this space. More precisely,
Problem 2.e.2. Let X be the space lpfor some 1 <p<oo, with the usual norm 11·11.
Let 111·111 be an equivalent norm on X. Given e > 0, does there exist a subspace Y of
X so that d(( Y, 111·111), (X, 11·11)) < 1 + e?
This problem, which is called for obvious reasons the "distortion problem",
is still open. Some partial positive answers to it will be described in Vol. III and
Vol. IV. Some negative results concerning the analogous distortion problem in
Lp(O, I), pi'2, are given in [88]. The most interesting case in 2.e.2 is the case p=2.
Notice that in the statement of 2.e.2 we excluded the cases X=/1 and X=c o.
In these cases the answer to the problem is known to be positive. This is a result of
R. C. James [56] which has already been used in the proof of 2.e.1.
Proposition 2.e.3. Let (X, 11·11) be the space 11 or Co with its usualn,orm. Let 111·111
be an equivalent norm on X. Then, for every e> 0, there is a subspace Y of X with
d((Y, 111·111), (X, 1I·1I)<I+e.
Proof. Assume that X=ll and that alllxlll~lIxll~lIlxlll, for some a>O and all
XE X. Let e>O and let {Pn}:'=l be the natural projections induced by the unit
vector basis. For every n put An = sup {lIxll; IlIxlll = I, Pnx=O}. Clearly, An t A, for
some I~A~a. Let no be such that Ano<A(I+e). By the definition of the {An}:'=l
there is a block basis {Yk}k'=l of the unit vector basis so that, for all k, IIIYklll=l,
PnoYk=O and IIYkll > ,\j(l +e). For every choice of scalars {ak}k'=l we have
PnoC~l akYk) =0 and hence
co
2: lakl~(1+e)-2 k=l
2: lakl·
00
~A~/(1+e)-l,\
k=l
On the other hand, by the triangle inequality, IlIlet akYklll~ k~l lakllllYkll1 ~ k~l lakl
and thus d(([Yklr=l, 111·111), 11)~(1+B)2.
The proof for Co is similar. By replacing the "sup" in the definition of An, by
"inf", we get that, for some constant A, there is a block basis {Yk}r= 1 of the unit
vector basis of Co with IIIYklll=l, IIYkll<A(l+B) for every k, and Illk~l akYklll~
(1 +B)2· max lakl for every sequence of scalars {ak}r=l tending to O. An estimate
k
from below can again be deduced from the triangle inequality. Indeed, assume
that lakol = max lakl; then
k
The preceding proof does not work for 1 <p <00 since the triangle inequality
cannot be used to get one of the desired inequalities automatically. The only thing
which can obviously be done is to choose a block basis {Yk}r=l of the unit vector
basis so that IllYkll1 = 1 for every k and IIlk~1 akYkll1 ;:, (J11ak1P t P/ (1 + B) for every
choice of scalars {ak}r=1, and a block basis {Zk}l~=1 of the unit vector basis so that
Illzklll = 1 for every k and II/kt akzl,III~(l +B)(J1 laklPtP for every choice of
scalars {ak}r= l'
We return to the discussion centered around the question with which we started
this section. While there are no known good criteria for a space to contain sub-
spaces isomorphic to lp, for some 1 <p <00, the situation is different for Co and
especially for 11 , We shall first present a simple characterization for spaces con-
taining Co and then present a deep result characterizing spaces containing h.
Proposition 2.e.4 [9]. A Banach space X has a subspace isomorphic to Co if and only
00
if there is a sequence {X n};:'=l in X so that 2: Ix*(xn)1 <00 for every x* E X* but

n=l
00
2: xnfails to converge.
n=l
Proof The "only if" part is trivial: we simply take as {X n};:'=l a basic sequence
which is equivalent to the unit vector basis of co. To prove the "if" part let {X n};:'=l
be such that ~ Ix*(xn)1 <00 for every x*
n=l
E X* and i
n=l
Xn diverges. It follows from
the uniform boundedness principle that there is a constant M so that i

n=l
Ix*(xn)1 ~
Mllx* II for every x* E X*. Since ~ Xn diverges there is an 10 > 0 and integers
n=l
P1 <q1 <P2 <q2 < ... so that IIn~~ Xnll;:'B for every k. Put Yk=n=~k Xm k= 1,2, ....
Since i
k=l
IX*(Yk)1 <00 for every x* E X* it follows that Yk:'+ O. By l.a.12 (and the
remark following it) we may assume without loss of generality that {ydk'=l forms
a basic sequence with basis constant K (otherwise, pass to a subsequence). For every
finite sequence of scalars {ak}~=l we have 112
k=1
aky"ll?: 8 max la"I/2K and also
Ie
Thus, {Y"}'~=l is equivalent to the unit vector basis of Co. 0

co co
Remark. A series 2: Xn for which 2: Ix*(xn)1 <00 for every x* E X* is said to be
n=l n=l
co
weakly unconditionally convergent (w.u.c.). It is easily seen that 2: Xn is a w.u.c.
n=l
co
series in an arbitrary Banach space X if and only if 2: anxn converges uncon-
n=l
ditionally whenever an ---;.. O.
We pass now to a fundamental result due to H. P. Rosenthal [129].
Theorem 2.e.S. Let {Xn }:'=l be a bounded sequence in a Banach space X. Then,
{Xn}:'=l has a subsequence {X n ,}i'=l satisfying one of the two mutually exclusive
alternatives:
(i) {Xn,}i'= 1 is equivalent to the unit vector basis of 11 ,
(ii) {X n ,}i'=l is a weak Cauchy sequence.
Consequently, the unit ball of X is weakly conditionally compact if and only if no

closed subspace of X is isomorphic to 11 ,
The proof given in [129] is valid only for real spaces. L. Dor [31] adapted the
proof to the complex case. We shall reproduce here a simplified proof due to J.
Farahat [38].
Proof By using the canonical embedding of X into X** we can consider each Xn
as an affine continuous functionfn on the unit ball S of X*. In this setting we have
to prove that if {fn}:'=l does not have a subsequence which converges pointwise
then it has a subsequence which is equivalent, in the sup norm on S, to the unit
vector basis of 11 ,
Suppose that no subsequence of {fn}:'=l converges pointwise on S. Let
E&={D~, Dnk'=l be the countable family of all pairs of open discs in the complex
plane for which both centers and radii are rational and such that
diam m=diam D~<d(m, D~)/2, k=1,2, ....
Then there exists an index ko and an infinite subsequence {fn}nEM, MeN, such
that for every subsequence {fn}nEL with Le M there is an SL E S for which the
100 2. The Spaces Co and I"
sequence of scalars {fn(SL)}neL has points of accumulation in both D~o and D~o'
Indeed, if this were false we could construct a sequence of infinite subsequences of
the integers N'=> Ml :::> M2 ... :::> M k :::> ••• such that, for every k and every s E S, the
sequence {fn(s)}neMk does not have points of accumulation in both discs D~ and
D~. Let L={mk}k=l be a subsequence of the integers so that mk E Mk for every k.
In view of our assumption, {fn}neL does not converge pointwise on S and thus
there is an So E S for which {fn(So)}neL has at least two distinct points of accumula-
tion dl and d2 • This, however, contradicts our assumption concerning M k , for k
chosen so that dl E D~ d2 E D~.
Let a be the center of D~o and fJ the center of D~o' There is no loss of generality
to assume that fJ-a is real and positive (otherwise, replace {fn}neM, D~o and D~o
by {yJ,.}neM' yD~o and yD~o' respectively, where y=lfJ-al/(fJ-a».
To continue the proof we need the following lemma.
Lemma 2.e.6. Let {Ai' Bi};=l be a sequence of pairs of subsets of a set S so that
Aj n Bj = 0 ,j= 1,2, .... Assume that there is no subsequence {Ai'" Bin}:=l so that,
for every s E S, either lim XAj (s)=O or lim XBj (s)=O (here XA denotes the character-
h" h"
istic function of A). Then there exists an infinite subsequence {Aj, Bihel so that
for any pair of disjoint finite sets 8, CTCJ (such a sequence {Ai' BAiel is called Boolean
independent).
Once 2.e.6 is proved, the proof of 2.e.5 can be completed as follows.

Let ko and M={ni}f"=l be as above. Set
Bj={s;fnis) E D~o}' j= 1,2, ....
The properties of ko and M show that the assumptions in 2.e.6 are satisfied. Thus,
there exists a Boolean independent infinite subsequence {A" Bjhel' We shall show
that if d=d(D~o' D~o) and cj=ai+ib"j E J are arbitrary complex scalars then
i.e. that {fnjhel is equivalent to the unit vector basis of h. Let CT be any finite subset
of J and assume, for simplicity, that 2: lail ~ 2: Ibil. Set CT + ={j;j E CT, ai~O},
jeu jeu
CT _ = CT '" CT + and choose
S2 E ((
ieu+
n Aj) n (n
feu.
B f ).
Since, for Zl E D~o and Z2 E D~o' we have

e. Banach Spaces Containing lp or Co 101
it follows that
1 1
~2 .L:
feu
af Re (fnls1)-f",(s2»-2 .L:
feu
Ib f 1m (fnis1)-fnls2»1
d d d d
~2 .L: lafl-4 feu.L:
feu
Ibfl ~4 .L: lafl ~8 feu
feu
.L: ICfl· 0
Proof of 2.e.6. The proof is based on the following combinatorial result, due to
C. st. J. A. Nash-Williams [108] (for a simpler proof, as well as a more general
result, see [36]): Let .9'oo(L) denote the set of all infinite subsets of a countable set L
and let .9'c.9'oo(N) be a closed subset (.9'oo(N) is identified with {D, I}N, endowed with
the product topology). Then .9' is a. Ramsey set, i.e. for every ME .9'oo(N), there
exists an L E .9'oo(M) such that either .9'oo(L)c.9' or .9'oo(L)c.9'oo(N) -.9'.
Let the sequence {Af, BAf=l satisfy the assumptions of2.e.6 and, for notational
convenience, denote Bf by -Ajoj= 1,2, .... For each integer k consider the subset
~c.9'oo(N) consisting of all M={nh}h"=l for which n (-I)hA nh # 0. Each of the
k
h=l
n ~. Thus.9' is a Ramsey set; this means that there
00
sets ~ is closed and so is.9'=

k=l
exists an L={mph~'=l E.9'oo(N) so that either .9'oo(L)c.9' or .9'oo(L)c.9'oo(N)-.9'. In
our case we must have .9'oo(L)c.9'. Indeed, in view of our assumption, there is an
So E S so that both sets {n E L; SO E An} and {n E L; SO E Bn} are infinite which
implies the existence of an infinite subsequence L O ={nh}h"=l of L so that
So E (-I)hA nh , h= 1,2, .... Thus, Lo E.9' and therefore .9'oo(L)c.9'.
Let J={m2P};'=1; we claim that {Ai> Bj}le! is Boolean independent. Indeed, let
{8p}~=1 be a finite sequence of signs. Construct a subset L1 = {nh}h"= 1 of L which
contains the integers m2, m4,"" m2k, scattered among nh n2,"" n2k, so that if
nh=m2pthen8p=(-I)l". We have
k
p=l
n
8pA m2p::>
2k
h=l
n
(-I)hA nh #0. 0
We pass now to another result which characterizes Banach spaces containing

11 , This result is due to Odell and Rosenthal [109] and Rosenthal [130].
Theorem 2.e.7. Let X be a separable Banach space. Then, the following assertions
are equivalent
(i) X does not contain a subspace isomorphic to h.
(ii) Every element in X** is the w*-limit of a sequence of elements in X (i.e. in
the canonical image of X in X**).
(iii) The cardinality of X** is equal to that of X (i.e. the cardinality of the
continuum).
(iv) Every bounded sequence in X** has a w* convergent subsequence.
We shall outline the proof of a part .of this theorem; namely of the equivalence of
(i), (ii) and (iii). Assume that every element of X** is a w* limit of a sequence of
102 2. The Spaces Co and I"
elements in X. This sequence can obviously be taken out of any fixed countable
dense set in X. Hence, the cardinality of X** is at most the cardinality of the set
of all subsets of the integers, i.e. that of the continuum. Thus (ii) =? (iii). It is also
trivial that (iii) =? (i) since /t* has a cardinality larger than that of the continuum
(observe, e.g. that if r is a set of the cardinality of the continuum then
11 (r)cC(o, l)*c/",=/t and thus Ir has I",(r) as a quotient space).
Outline of the proof of (i) =? (ii). Assume that there is an xt* E X** with Ilxt*11 = 1
which is not a w* limit of a sequence of elements in X. Let S be the unit ball of X*
with the w* topology. Since X is separable S is a compact metric space. The function
xt*(x*) on S is not the pointwise limit of a sequence of continuous functions on S,
i.e. does not belong to the first Baire class on S. This fact is however not obvious;
our assumption on xt* implies immediately only that it is not the pointwise limit
of a sequence of affine continuous functions on S. A direct proof of this fact is
given in [109]; it can also be deduced from a general result of Choquet concerning
functions of Baire class 1 (cf. [1, p. 16]).
Once we know that xt*(x*) is not in Baire class 1 on S it follows from the
classical characterization of Baire class 1 functions (cf. [51, p. 288]) that there is
a closed non-empty set Kin S so that the restriction of xt* to K has no points of
continuity. Let ~={D~, D~}k=l be the family of discs used in the proof of 2.e.5
and let, for k= 1,2, ... ,
Fk={x* E K; in any w* neighbourhood G of x* there are y* E K (') G

and z* E K (') G so that xt*(y*) E D~, xt*(z*) E D~} .
'" Fk. By the Baire category

Then, clearly, each Fk is a closed subset of K and K= U
k=l
theorem there is an integer ko so that Fko has a non-empty interior, say Go, relative
to K. Thus, for every non-empty open subset G of Go, there are y*, z* E G so that
xt*(y*) E D~o and xt*(z*) E D~o'
We shall prove that there is a sequence {Xn}:'=l in the unit ball of X so that the
sequence {An> Bn}:'=l> where
is Boolean independent. Once this is proved it follows, as in the proof of 2.e.5,

that {Xn}:'=l i.e. equivalent to the unit vector basis in 11 ,
We choose the {Xn}:'=l inductively. By the definition of Go there are y*, z* E Go
so that xt*(y*) E D~o and xt*(z*) E D~o' By the w* density of the unit ball of X
in the unit balI of X** there is an Xl E X with Ilxlll ~ 1 so that Y*(Xl) E D~o and
Z*(Xl) E D~o' With this choice of Xl we have that Alof 0, Blof 0 and obviously
Al (') Bl = 0. Assume that {Xn};;'=l have already been chosen in the unit ball of X
n 8nAnof 0
m
so that, for every choice of signs 8=(81 ,82 , ... , 8m), (where for nota-
n=l
n 8nAn is an open subset of Go there
m
tional convenience we put - An = Bn). Since
n=l
are, for every such 8, elements y; and z; in n 8nAn so that xt*(yt) D~o and
m
E
n=l
e. Banach Spaces Containing lp or Co 103
xt*(zt) E D~o. By the w* density of the unit ball of X in that of X** we get that
there is an Xm+1 with Ilxm+111 ~ 1 so that Y;(X m+1) E D~o and Z;(Xm+1) E D~o for all
the 2m possible choices of ({II, (l2, ... , 8m). With this choice of Xm+l it is clear that
n
m+1
n=l
{lnAn# 0 for all (m+ 1)-tuples of signs (8 1 , (l2,.." (lm+l)' 0
There are also other interesting theorems characterizing Banach spaces con-
taining II' These theorems involve the function spaces C(O, 1) and L 1 (0, 1) and
we shall discuss them in Vol. III.
A question which goes back to Banach is whether every separable Banach
space X whose dual is non-separable must contain a subspace isomorphic to II'
This question was answered negatively by two, independently constructed, counter-
examples. One (cf. [57] and also [91]), denoted by JT and called the James tree, is
obtained from the space J (cf. Example 1.d.2) by replacing its index set (i.e. the
integers) by an infinite tree. The second example (cf. [91]), denoted by JF and called
the James function space, is the continuous analogue of J. This space is easier to
define (but more difficult to analyse) than JT. The space JF is the completion of the
linear span of characteristic functions of subintervals of [0, 1] with respect to the
cr
norm
Ilfll = sup C~ f(t) dtrr2
where the supremum is taken over all partitions 0= to < t1 < ... < tn = 1 of [0, 1].
For both spaces the verification that their duals are non-separable is trivial while
the proof that they do not contain II is more difficult. Since the discovery of these
two examples several other counterexamples have been found. We mention, in
particular, an example due to Hagler [50] of a space X which, among its other
interesting properties, satisfies the following: X is separable, X* is non-separable,
every infinite-dimensional subspace Y of X has a subspace isomorphic to Co and
every infinite-dimensional subspace Z of X* has a subspace isomorphic to It. We
shall present in Vol. IV still another counterexample to Banach's question which is
obtained by using a counterexample of James [58] to an important question in the
local theory of Banach spaces.
We pass now to some questions involving duality which concern Banach spaces
containing Co or II'
Proposition 2.e.8 [9]. Let X be a Banach space such that X* contains a subspace
isomorphic to co. Then, X has a complemented subspace isomorphic to II' Conse-
quently, X* has a subspace isomorphic to 100 ,
Proof Let T: CO --?- X* be an isomorphism into and let {e n};:'=l denote the unit
vector basis of co. The map x --?- Sx=(Te 1 (x), Te 2(x), ... ) is the restriction of T*
to Xc X** and thus it maps X into 11' Since T* maps X** onto 11 and the unit
ball of X is w* dense in the unit ball of X** there exists a constant K such that, for
n-1
every n, there is an Xn E L: ITei(x n)I< lin. The
X with Ilxnll ~ K, Ten(xn) = 1 and
i=l
sequence {SX n};:'=l has, by 1.a.9, 1.a.12 and 2.a.I, a subsequence {SXnk }t'=l which
is equivalent to the unit vector basis of II and whose span is complemented in 11 by

a projection P. Hence, for some constant M and every choice of scalars {ak}f~l>
we have
Thus, S is invertible on Y = [xnk]f~1 which implies that Y is isomorphic to 11 and

S-IPS is a projection from X onto Y. 0
Proposition 2.e.8 shows in particular that Co is not isomorphic to a subspace of a

separable conjugate space.
We cannot interchange the roles of Co and 11 in 2.e.8. If, e.g. X=11, then X*
contains 11 as a subspace without X having Co as a subspace. It is also possible for
X* to contain 11 as a subspace without X having Co as a quotient space. This is
e.g. the case for X=I", (it will be proved in Vol. II that every separable quotient
space of 1", is reflexive (cf. Proposition 2.f.4 below)). However, for separable spaces,
we have the following result [60].
Proposition 2.e.9. Let X be a separable space such that X* contains a subspace

isomorphic to 11' Then Co is isomorphic to a quotient space of X.
Proof. Let {X;D:~1 be a sequence in X* which is equivalent to the unit vector basis
of 11' Since X is separable there is a subsequence {x;k}f~1 which converges w*.
The sequence Y: =X;2k+l -X;2k' k= 1,2, ... is also equivalent to the unit vector
basis of 11 and converges w* to O. By the proof of l.b.7 there is a subsequence
{Y:j}i'=1 of {Ynf~1 which is a w* basic sequence. Hence, by l.b.9, the space
X/([Y:j]i'= l)T is isomorphic to co. 0
We conclude this section with an open problem which is closely related to the
results discussed above.
Problem 2.e.10. Does every infinite-dimensional Banach space X contain an infinite

dimensional subspace which is either reflexive or isomorphic to Co or to h?
By l.c.12 the answer to 2.e.1O is positive if X has a subspace with an uncon-

ditional basis (thus, 2.e.1O is a weak version of Problem l.d.5). Another partial
answer to 2.e.1O is given in l.b.l4.
f. Extension and Lifting Properties, Automorphisms of rm

Co and 11
The spaces I", and Co have important "extension properties" which characterize
them while 11 is characterized by a "lifting property". Some of these properties
f. Extension and Lifting Properties, Automorphisms of I"" CO and l, 105
were mentioned briefly in the previous sections. Here we shall treat them in detail.
The main part of this section will be devoted to the study of automorphisms (i.e.
invertible linear operators) of I"" Co and 11 , It turns out that these spaces are
surprisingly rich in automorphisms.
We start by considering the extension property of I",.
Definition 2.f.I. A Banach space X is said to be injective if, for every Banach space
Y containing X as a subspace, there is a bounded linear projection from Yonto X.
Injective spaces can be characterized by extension properties for operators, as

shown in the following simple proposition.
Proposition 2.f.2. The following three assertions concerning a Banach space X are
equivalent.
(i) X is injective.
(ii) For every Banach space Y:;) X, every Banach space Z and every T E L(X, Z)
there is aTE L( Y, Z) which extends T.
(iii) For every pair of Banach spaces Z:;) Y and every T E L( Y, X) there is a
T E L(Z, X) which extends T.
Proof Assertion (i) is a special case of both (ii) and (iii) (take e.g. in (ii) Z = X
and T the identity operator) and thus (ii) => (i) and (iii) => (i). Assume now that
(i) holds and let Y, Z and T be given as in (ii). Let P be a projection from Y onto
X. The operator T= TP has the desired property. It remains to prove that (i) => (iii).
Let {xnYer be a set of functionals of norm 1 on X such that Ilxll = sup Ixi(x) I for
r
every x E X (if X is separable r can of course be taken to be countable). Define
an isometry S, from X into I",(r), by Sx(y)=xi(x), y E r. For every y E r let
zi E Z* be a norm preserving extension of T*xi from Y to Z and let To E L(Z, I",(F))
be defined by Toz(y)=zi(z), y E r. Let P be a projection from I",(r) onto SX (here
we use (i)). Then, T=S-lPTo has the desired property. 0
The proof of (i) => (iii) above shows in particular that, for every set r, the
space I",(r) is injective (there is even a projection of norm 1 from any Z:;) l",(r)
onto lro(F)). The spaces lro(r) are not the only injective spaces. We shall discuss
in detail the question of the structure of injective spaces in Vol. III. We shall show
here only that I", is characterized by the fact that it is the "smallest" injective space.
Theorem 2.a.7 states that every injective infinite-dimensional subspace of t", is
isomorphic to tro and, in particular, that there are no separable infinite-dimensional
injective spaces. Essentially the same argument as that used in 2.a.7 proves however
the following stronger version of 2.a.7.
Theorem 2.f.3 [125]. Every infinite-dimensional injective Banach space has a sub-
space isomorphic to lro.
This theorem is a consequence of

Proposition 2.f.4. Let r be a set and T a non-weakly compact operator from I",(r)
into some Banach space Z. Then, there exists a subspace U of I",(r) which is iso-
morphic to I", so that T 1u is an isomorphism.
We show first that 2.f.4 implies 2.f.3. Let X be an infinite-dimensional injective

space and let r be such that X is isometric to a subspace Xo of I",(r). Let P be a
projection from [",(r) onto Xo. By fact (ii) stated in the proof of 2.a.7 P is not w
compact and thus, by 2.f.4, Xo has a subspace isomorphic to I",.
Proof of 2.f.4. The starting point of the proof is a stronger variant of fact (i)
which was used in the proof of 2.a.7: every bounded non-w compact operator T,
defined on a C(K) space, is an isomorphism on a suitable subspace which is iso-
morphic to co. This result is due to Pelczynski [114] and will be proved in Vol. III.
Let T: I",(r) -+ Z be non-w compact. Then, by the preceding remark, there are
vectors {Yi}j';,l in I",(r) wlpch are equivalent to the unit vector basis of Co and so
that T1 = TI(ytlt~ 1 is an isomorphism. Since [Ydj';, 1 is separable we may consider it
as a subspace of I",. Since I", and I",(r) are injective there are bounded linear
operators Sl: 100 -+ I",(r) and S2: Z -+ I", so that Sll[lItll".,l = identity and
S21[Tlltl;"=1 =Tl\ i.e. the following diagram commutes
s.
100 ~ loo(r) ~ Z ~ I",
U U U u
Let I be the identity operator of 100 • The operator 1- S2TS1: I", -+ I", vanishes
on [Yt]j';,l. By the proof of2.a.7, I-S2TS1 must vanish on a subspace Vof 100 which
is isomorphic to I",. Hence, S2TS1 is the identity on V and therefore U=Sl V is
isomorphic to I", and TIU is an isomorphism. 0
We pass now to the extension property of co. The space Co is, by 2.a.7 (as
every other separable space), not injective. If we consider however only separable
spaces containing Co then Co behaves as an injective space. More precisely, we have
the following theorem due to Sobczyk [136].
Theorem 2.f.S. Let Y be a separable Banach space containing co. Then there is a
projection of norm ~ 2 from Yonto co.
Proof(due to Veech [142]). Let {e~},~)=l denote the unit vector basis in 11 =ct. For
every n let y~ E y* be a norm preserving extension of e~ from Co to Y. Let
F=By. (\ (c&) and let d be a translation invariant metric on y* which induces on
By. the w* topology (here, the separability of Y is used). Since every w* limit point
of the set{Y~};'=l belongs to Fit follows thatd(y~, F) -+ O. Let {Z:};'=l be a sequence
of elements in F such that lim dey:, z:)=O, i.e. w* lim (y:-z:)=O. The operator
"
P: Y -+ co, defined by Py=(y!(y)-z!(y), y:(y)-z:(y), ... ), is a projection of
norm ~2. 0
f. Extension and Lifting Properties, Automorphisms of /"" Co and h 107
The property of Co, which was exhibited in 2.f.5, actually characterizes Co

among the separable Banach spaces. More precisely, a separable infinite-dimen-
sional Banach space which is complemented in every separable space containing
it must be isomorphic to Co. This is a difficult result, due to M. Zippin [149], and
will be presented in Vol. III.
We shall present now a result on quotient spaces of Co whose proof uses 2.f.5.
Theorem 2.f.6 [64]. Every quotient space of Co is isomorphic to a subspace of co.
Proof. Let X be a quotient space of Co. Since X* is clearly separable we deduce

from l.g.2 that there is a closed subspace Y of X so that both Y and XI Y have
shrinking F.D.D.'s. By 2.d.1 there exist sequences {B n}:'=l and {Cn }:'=l of finite
dimensional Banach spaces so that Y is isomorphic to C~l EEl Bnt and XIY is
isomorphic to (~ EEl Cn) . Since (~ EEl B n): is clearly isomorphic to a subspace

n=l 0 n=1 0
of (co EEl Co EEl •• ')0 = Co we get that Y, and similarly XI Y, are isomorphic to sub-
spaces of Co. Let T: Y - Co and S: XI Y - Co be isomorphisms into and let
Q: X _ XI Y be the quotient map. The operator T can be extended to a bounded
linear operator T: X - Co. Indeed, by 2.f.2, T can be extended to an operator
To: X -100' Since ToX is separable there is, by 2.f.5, a projection P from
span {ToX, co} onto Co. The operator T=PTo maps X into Co and extends T. It is
easily verified that the operator R: X-co EEl Co, defined by Rx=(Tx, SQx), is
an isomorphism into. 0
We turn now to the study of the space h which has a property dual to the exten-
sion property, namely the "lifting property".
Proposition 2.f.7. Let X and Y be Banach spaces so that there is an operator Sfrom
Y onto X. Then, for every T E L(h, X), there is aTE L(11, Y) for which ST = T.
Moreover, if S is a quotient map then, for every e>O, T may be chosen so that
ITII~(l +e)IITII·
Proof. Let {en}:'= 1 be the unit vector basis of /1 and put Xn = Ten> n = 1, 2, .... Then,
sup Ilxnll <00. It follows from the open mapping theorem that there is a bounded
n
sequence {Yn}:'=l in Y such that SYn=xno n=l, 2, .... The operator T, defined by
00
T(a1, a2,"')= 2: anYn, has the desired property. If S is a quotient map then, for
n=l
every e>O, we can choose the {yJ:'=l so that IIYnll~llxnll(1+e), for all n. This
proves the second assertion. 0
The property of h exhibited in 2.f.7 characterizes 11 , We say that a Banach

space Z has the lifting property if, for every operator S from a Banach space Y
onto a space X and for every T E L(Z, X), there is aTE L(Z, Y) so that T= ST.
Every separable infinite dimensional Banach space with the lifting property is
isomorphic to 11, Indeed, if we take Y=h, X=Z, S a quotient map from 11 onto
Z and T=identity of Z, in the definition of the lifting property we deduce that there
is aTE L(Z, 11) so that ST= I z . Consequently, T is an isomorphism into and TS
is a projection from h onto TZ. Thus, by 2.a.3, Z~ 11 ,
The spaces having the lifting property have been characterized also in the non-
separable case. It is clear from the proof of 2.f.7 that, for every set F, the space
11(F) has the lifting property. Kothe [73] generalized 2.a.3 to the non-separable case
and showed that, conversely, every ·space with the lifting property is isomorphic to
11(F),for some set F.
We used in the preceding paragraph (and in several places in the preceding

sections) the fact that every separable Banach space X is a quotient space of h. Let
us recall the proof of this simple fact. We let {Xn }:'=l be a dense sequence in the
00
unit ball of X and define T: h - X by T(a1o a2"')= 2: anxn.

n=l
It is clear that
IITII ~ 1 and, since the image under T of the unit ball of 11 is dense in the unit ball
of X, we get that Tis a quotient map. This proof shows that T is highly non-unique;
it depends on the choice of the sequence {X n }:'=l (actually the {X n }:'=l need not even
be dense in the unit ball of X. It is easily seen that a necessary and sufficient con-
dition for T to be a quotient map is that Bx=conv {X n }:'=l)' It is therefore some-
what surprising that, up to an automorphism of 110 T is actually unique.
Theorem 2.f.8 [89]. Let T1 and T2 be two linear operators mapping 11 onto the same
Banach space X which is not isomorphic to h. Then there exists an automorphism T
of 11 (i.e. an invertible linear operator from 11 onto 11) so that T1 = T 2T. In particular,
ker T1 is isomorphic to ker T 2.
Proof. By 2.a.2 and 2.a.3 there are subspaces U and V of 11 so that 11 = U EEl V, Uc
ker T1 and U ~ V ~ 11, Let P be the projection of 11 onto V which maps U to O. Then,
T1 =T1P and thus T 11V maps V onto X. Using twice the lifting property for T 11V
(acting on V~/1) and T2 we conclude that there are operators T1: V -/1 and
T 2: /1 - V so that T2T1 =T1IV and T 1T 2=T2. Let R be an isomorphism from U
onto /1 normalized so that IIR- 1 11 > 1 and put S=(I-T1T 2)R(I-P)+T1P, where
I is the identity on II' The operator S maps II into /1 and satisfies
T2S=(Tz-TzT1T2)R(I-P) + T zT 1P=T1P= T1 .
Passing to the duals we get that S*=(I-P*)R*(I-T:Tn+p*Tt and therefore,
IIS*x*11 ~-!IIPll-1 max (11(I-P*)R*(I-T:Tnx*ll, IIP*T'tx*l[)
for every x* E It'. Since IIP*T'tx*11 ~ IIT'tx* II and, similarly,

f. Extension and Lifting Properties, Automorphisms of l"" Co and 1, 109
we get that
IIS*x*1I ~tIlPII-11IR-111-1 max (1Ix*II-llt2 111Ittx*ll, Ilttx*ID

~(41IPIIIIR-111(1 + Ilt2 ID)-1Ilx*ll·
Hence, S* is an isomorphism and thus S is an operator from h onto 11 , By using

again the lifting property of h we deduce that there is an i E L(/1' h) such that
si=I. The operator I-is is a projection from h onto ker S.
Let Q be a projection in h such that QI1 c ker T2 and QI1::::: (I - Q)/1::::: h. Then
h can be decomposed into the direct sum h =ker S EEl W 1 EEl W 2 so that SW1 = QI1
and SW2 = (1- Q)/1 • Let T E L(h, 11) be an operator which maps ker S EEl Wl
isomorphically onto Qh and whose restriction to W 2 is equal to SI W 2' Then, T is
an automorphism and T1 = T2T since ker S EEl W 1c ker T1 and T1 = T 2S. 0
Remarks. 1. The requirement that X is not isomorphic to h was used in the proof
only to ensure that ker T1 and ker T2 are both infinite-dimensional. If X = 11 and
if dim ker T1 =1= dim ker T2 (i.e. if one of them is finite and the second infinite or if
both are finite but different) then, clearly, there does not exist an automorphism
T so that Tl =T2T.
2. Theorem 2.f.8 shows, in particular, that to every separable Banach space X
(with X';f, 11) there corresponds an, up to isomorphism unique, subspace of h,
namely the kernel of any quotient map from h onto X. This correspondence is
not however, one to one. For example, if T is a quotient map from h onto some
Banach space X then T: h EEl 11 -+ X EEl 110 defined by T(u, v) = (Tu, v), is also a
quotient map. Clearly, ker T= ker T though, in general, X is not isomorphic to
XEElh.
It is possible that 2.f.8 or even some weaker versions of it characterize the space
h. For example, the following problem is open.
Problem 2.f.9. Let X be an infinite-dimensional separable Banach space. Assume

that,for every pair ofquotient maps T1 and T 2 from X onto the same space Y (with
Y';f, X), there is an automorphism T of X so that Tl = T 2T. Is X isomorphic to either
h or 12?
Notice that, for X = 12 , the assumptions in 2.f.9 are trivially satisfied.
Theorem 2.f.8 can be dualized to theorems concerning extension of isomorph-
isms in Co and 100 , We state first the result for co.
Theorem 2.f.l0 [89]. Let Y be a closed subspace of Co and let T be an isomorphism

from Y into Co so that dim col Y = dim co/TY =00. Then, there is an automorphism T
of Co so that Tly=T. In particular, co/Y:::::co/TY.
The proof of 2.f.l0 is very similar to the proof of the first part of the corres-
ponding result for 100 (Theorem 2.f.12 below) and therefore we omit it. Theorem
2.f.l0 enables us to associate to every Banach space X isomorphic to a subspace
of Co (with X';f,c o) a, unique up to isomorphism, quotient space of Co (denoted by
col X). In analogy to 2.f.9 we have the following problem.
Problem 2.f.H. Let X be a separable infinite-dimensional Banach space. Assume

that, for every pair Y, Z of isomorphic subspaces of X of infinite-codimension, there
is an automorphism T of X so that T Y = Z. Is X isomorphic to Co or 12 ?
We pass now to the space 100 ,
Theorem 2.f.12 [89]. Let Y be a subspace of 100 and let T be an isomorphism from Y
into 100 so that dim 100/ Y = dim Ioo/TY =00. Then,
(i) If I",,/Y and I",,/TY are both non-reflexive there is an automorphism T of 100
which extends T (i.e. Tly=T).
(ii) If only one of the spaces 100/ Y and Ioo/TY is reflexive T cannot be extended
to an automorphism of I"".
(iii) If Ioo/Y and I",,/TY are both reflexive then every extension t of T to an
operator from 100 into itselfis a Fredholm operator. The index i(t) oft does
not depend on the particular choice of t and thus defines an integer valued
invariant of T denoted by f(T). The operator T can be extended to an auto-
morphism of 100 if and only iff(T) =0.
Since assertion (ii) is trivial only (i) and (iii) require a proof.
Proof of 2.f12(i). We show first that if Y is a subspace of 100 with 1",,/ Y non-
reflexive then there is a projection P on I"" with P Y ={O} and dim PI"" =00 (i.e.
PI"" ~ I",,). Indeed, since the quotient map ({l: I"" ~ 1",,/ Y is not w compact there is,
by 2.f.4, a subspace U of I"" with U~/"" so that ({l1U is an isomorphism. Since I""
is injective there is a projection Po from 1",,/ Yonto ({lU. The projection P= C({l I U)-lPo({l
has the desired properties.
Let T: Y ~ I"" be an isomorphism such that 1",,/ Y and I",,/TY are both non-
reflexive. By the preceding remark there are projections P and Q on I"" so that
PY= QTY={O} and PI",,~ QI""~/",,. Let I denote the identity operator of I"". The
subspace (1- Q)/oo contains TY and hence, by the injectivity of 100 , there is an
Sl: 100 ~ (1- Q)/oo such that Slly=T. Similarly, there is an S2: I"" ~ (I-P)/oo so
that S2ITy';'T-1. Let R be an isomorphism from 100 onto Q/oo normalized so that
IIR-III > 1 and define t: 100 ~ I"" by t=Sl +R(I-S2S 1). Since (I-S2 S 1)IY=0
it follows that t1y=T. We show next that tis an isomorphism into 100 , Indeed, let
x E 100 ; then
Iltxll ~max GIQtxll/IIQII, IICI- Q)txIIlIlI- QID

~!IIQII-l max (IiRCI-S2S I )xll, IISIXID
~!IIQII-IIiR-lll-l max (1Ixll-IIS2111ISIXII, IISIXID
~ Ilxll/21IQIIIIR- 1 1IGI S 211+1).
The operator t is not necessarily an automorphism of 100 since t need not map
100 onto 100 , In order to replace t by an automorphism consider the subspace
tCI-P)loo of 100 , This is a complemented subspace of infinite-codimension in 100
(since (I -P)/oo is of infinite codimension in 100)' Hence, by 2.a.7, there is a subspace
W of 1<1) so that 1<1) = WE8 TCI-P)loo and W~loo. Let Ro be an isomorphism from
f. Extension and Lifting Properties, Automorphisms of 1 Co and 11
00 , 111
Ploo onto W. The operator T=RoP+T(I-P) is an automorphism of 100 and

Tly=RoPly+T(I-P)ly=Tly=T. D
Proof of 2.f12(iii). We remark first that if K is a w compact operator on 100 then

1+ K is a Fredholm operator of index O. This follows from the fact that K2 is a
compact operator (cf. fact (ii) in the proof of 2.a.7) and that the classical Riesz
theory for compact operators applies also to operators which have a compact
power (cf. [33, VII.4.6]). Alternatively, we could apply 2.c.1O and use the fact that
K is strictly singular. Indeed, let VC 100 be such that KIV is an isomorphism. By the
injectivity of 100 there is a Ko: 100 --+ 100 such that Ko IKV = K -1. Then KoKI v = identity
and since (KoK)2 is compact it follows that dim V <00.
Let now Rand S be operators from 100 into itself such that R1y=T, SITY=T-l
(R and S exist since I", is injective). The operator 1- SR vanishes on Yand thus it
factors through the reflexive space 1",/ Y. Consequently, 1- SR is w compact and
hence SR is a Fredholm operator of index O. Similarly, since loo/TY is reflexive,
RS is a Fredholll1 operator of index O. Hence, both Rand S are Fredholm operators
and i(R) + i(S) = O. This proves that i(R) does not depend on the choice of Rand,
thus, f(T) is well defined. If there is an automorphism T which extends T then
f(T)=i(T)=O. Conversely, if f(T)=O then T can be extended to an R: I", --+ I",
with index 0. By adding to R a finite rank operator (which maps ker R isomorphic-
ally on a subspace U of I", such that U EEl RI", =/ 00 ) we get an automorphism T
of I", which extends T. D
Remarks. There are subspaces Y of 100 such that 1001 Y is reflexive and infinite-
dimensional. For example, 12 is isomorphic to a quotient space of Ioc;. This follows
from the following facts. By 2.b.3 L 1 (O, 1) has a subspace isomorphic to 12 and
thus 12 is a quotient space of Loc;(O, 1). The space Loo(O, 1) is isomorphic to 100 [113].
Indeed, Loo(O, 1) as a dual of a separable space is isometric to a subspace of 1 00 ,
The usual proof of the Hahn-Banach theorem can be used to show that, for every
Y:::J X and every T E L(X, Loo(O, 1)), there is a l' EL(Y, Loo(O, 1)) with 111'11 = IITII
and T,x=T. Thus, Loo(O, 1) is injective and its isomorphism with 100 follows by
using the decomposition method or, alternatively, from 2.a.7.
Let us verify now that cases (ii) and (iii) of2.f.12 can actually occur. Let loolY
be isomorphic to 12 , Since 100 EEl 100 = 100 there is a subspace Yo of 100 so that Yo is
isometric to Yand 1001 Yo~ 1001 Y EEl 1m i.e. lcol Yo is non-reflexive. By letting T be
an isometry from Y to Yo we get an instance where case (ii) occurs. Let now
S: Zoo ->-/00 be the shift operator defined by Seal, a2,"')=(O, aI, a2"")' For a
positive integer k let Tk=SFy. Then, loolY and loolTkY are both reflexive and
f(Tk)=k, and i(T;;l) = -k.
In connection with 2.f.12 it is of interest to study further the subspaces Yof
100 so that 1001 Y is reflexive. In this direction we have the following.
Proposition 2.f.13 [89]. Let Y be a subspace of 100 such that lcol Y is reflexive. Then
Y has a subspace isomorphic to 100 ,
Proof We show first that Y has a subspace isomorphic to co. Otherwise, Yand Co
112 2. The Spaces Co and i p
would be totally incomparable and thus (cf. the result quoted following 2.c.1)
Y + Co would be a closed subspace of 100 with dim (Y nco) <00. Consequently, the
quotient map 100 -3>-/ 00 1Y would be an isomorphism on an infinite-dimensional
subspace of Co which contradicts the reflexivity to 1001 Y.
By using 2.a.8 we get that there is a sequence {Yn};:'=l in Y which is equivalent
to the unit vector basis of Co so that, for every infinite subset M of the integers, the
w* closure X M of [Yn]nEM is isomorphic to 100 , Let {NY}YEr be an uncountable set
of subsets of the integers so that N Yl n N Y2 is finite for every Y1 =1= Y2' We claim that,
for some Y E r, X Ny C Y. Assume that, for every y, there is an Xy E X Ny with
Ilxyll = 1 and Xy rj Y. Let T: 100 -3>- 1001 Y be the quotient map. Since r is uncountable
there is a sequence {Yi}i'= 1 of elements in r and a 0> 0 such that IITxy!11 ~ 0 for
every i. Since [Yn];:'=l C Y it follows that, for every choice of scalars {ai}{=b
II ±alTxytll<::; max lall and thus S: Co -3>-/oo IY, defined by S(alo a2,···)= ~
1=1 l""j""j ;=1
alTxy!,
is a bounded operator, and Tx y ! -3>- 0 weakly in 1001 Y. Since 1001 Y is reflexive S is
weakly compact and therefore also compact (recall that in 11 a sequence converges
w if and only if it converges in norm and that an operator is compact, resp. w
compact, if and only if its conjugate is such an operator). Thus, {TXyJi'=l has a
subsequence which converges in norm. The limit of this subsequence must be 0
but this contradicts the fact that IITxy! II ~ 0 for all i. 0
It follows from 2.f.12 and 2.f.13 that if Y is separable then, for any pair of iso-
morphisms T1 and T2 of Yinto 100' there is an automorphism T of 100 so that TTl =T2
and, in particular, loolTl Y;:;;;; loolT2 Y. It makes sense therefore to talk of the space
1001 Y.
Theorems 2.f.8, 2.f. 10 and 2.f.12 show that there are many automorphisms of
II. Co and 1 The automorphisms constructed in these theorems are in general not
00 ,
isometries. The isometries form a rather restricted class and it is easy to determine
their most general form.
Proposition 2.f.14. Let T be an isometry of Co or Ip, 1 <::;p<::;oo, p=l=2 onto itself. Then
there is a sequence {8i }i'=1 of signs and a permutation 7r of the integers so
that T(ab a2, ... )=(81a,,(1), 82a"(2), 83a"(3»"')'
Proof Let us first consider the spaces Ip , 1 <::;p <00, p =1= 2. It is not hard to verify
(and for p= 1 it is trivial) that if x=(a1, a2,' .. ) and Y= (blo b2,···) then Ilx+ yW=
Ilx - yIIP = Ilxil P + II yIIP if and only if x and Y have disjoint supports (i.e. aibi = 0 for
every i). Hence, if T is an isometry of Ip into itself and {e;}i'=l denotes the unit
vector basis of Ip then {Tei};""=l have mutually disjoint supports. Consequently, if
T is, in addition, onto then there must exist a permutation 7r of the integers and
signs {8i }i'= 1 so that Tel = 8,e,,(!) for every i. The case of Co can be proved similarly
or by using the fact that if T is an isometry of Co onto Co then T* is an isometry of
11 onto h. In the case of 100 the simplest way to prove 2.f.14 is to deduce it as a
special case of the general result on the structure of isometries of C(K) spaces (the
classical Banach-Stone theorem which will be discussed in Vol. III). 0
3. Symmetric Bases
a. Properties of Symmetric Bases, Examples and Special

Block Bases
It is a trivial observation that the unit vector basis {en};;"=1 of Co or [p, p ?::- 1, besides
being unconditional, is equivalent to any of its permutations. Moreover, the unit
ball of each of these spaces is a symmetric body in the sense that it contains a
sequence (al, a2, ... ) if and only if it contains the sequence (a"(l)' a,,(2)," .), for every
permutation 1T of the integers. The purpose of this chapter is to study those Banach
spaces which share this property with Co and lp.
Definition 3.a.l [134]. A basis {x n};;"= 1 of a Banach space X is said to be symmetric

if, for any permutation 1T of the integers, {x,,(1t)};;"=l is equivalent to {Xn};;"=l'
It is interesting to compare this definition with l.c.6 where it is shown that a

basis {x n};;"= 1 is unconditional if and only if, for each permutation 1T, {x,,(n)};;"= 1 is
a basic sequence. For {Xn};;"=l to be a symmetric basis one has to require that, in
addition, the basic sequence {x,,(n)};;"=l be equivalent to {X n};;"=l for each such 1T.
This remark shows that every symmetric basis is also unconditional.
Let {x n};;"= 1 be a symmetric basis of a Banach space X. Then, for each permuta-
tion 1T of the integers, the operator V,,: X -0>- X defined by V,,(~ anxn) = ~ anx,,(n)
is evidently an automorphism of X. Moreover, the family {V,,} of all these auto-
morphisms induced by permutations forms a uniformly bounded group. Indeed,
if this were not true then, by the uniform boundedness principle, we could produce
00
a vector X= L anXn E X, a sequence {1Tj}j=l of permutations of the integers and a

n=l
sequence {uJi= 1 of finite subsets of the integers with
Uj n Uk= 0 and 1TlUj) n 1Tk(Uk) = 0 ,
for j=/= k, such that Iln~"f anXnf(nJII?::- 1 for all j. Then, for any permutation 1To such
00
that 1To(n)=1T2in) for n E U2j,j= 1,2, ... , it would follow that L

anx"o(n) diverges,
n=l
contrary to our assumption.
Since, as remarked above, {X n};;"=l is also unconditional, we get that
K=sup IIMeV,,1 I<00 (recall that, for every choice of signs 8={8n };;"=b the operator
e."
Me is defined by MeC~l anXn) = n~l an8nxn). The number K, thus defined, is called
114 3. Symmetric Bases
the symmetric constant of {Xn};'=1. Notice that the symmetric constant of any
symmetric basis is always larger than or equal to the unconditional constant of
co
that basis. If, for X= L: anXn E X, we put
n=1
IIxllo = 6:;t~1s~p Iln~1 anOnX,,<n)11

then 11·110 is a new norm on X such that IIxll ~ IIxllo~KllxlI for all x E X. This new
norm satisfies 111 anOnX,,<n)// =111n=1 anXnl1
In=1 0 0
for every permutation 7T of the integers
and every choice of signs On = ± I. If X is equipped with such a norm (called some-
times a symmetric norm) then its unit ball is a symmetric body with respect to all
permutations of the coordinates and all changes of signs.
We introduce now a notion which is slightly weaker than that of a symmetric
basis.
Definition 3.a.2. A basis {X n};'=1 of a Banach space X is called subsymmetric if it

is unconditional and, for every increasing sequence of integers {nt}j';., 1, {xnJj';., 1 is
equivalent to {Xn}~=l'
The requirement that {X n };'=1 be unconditional is not redundant: a basis which

is equivalent to each of its subsequences need not be unconditional. For instance,
the summing basis of the space c, introduced in Section I.c, is subsequence equiva-
lent but not unconditional.
Let {X n};'=1 be a sub symmetric basis of a Banach space X. It is easily checked
that, for any increasing sequence of integers {nt}j';., h the operator S{nj}; X ~ X,
defined by S{nj}C~l anXn) = t~l atXnj' is an isomorphism from X onto [XnJj';.,l.
Moreover, the family of all these isomorphisms is uniformly bounded; hence, we
can define the subsymmetric constant of {Xn};'=l as the number
If, for X= nt anXn EX, we put IIxllo= 6~;t~1 ~~r IIJl atOtxn,1I then 11·110 is a new
norm on X which satisfies IIxll ~ IIxllo~KlllxlI and IIJl a;OtXnjll o= IL~l anXnllo for
all x E X. Such a norm is called a subsymmetric norm.
The relationship between symmetric and sub symmetric bases is described in
the following simple result.
Proposition 3.a.3 [134]. Every symmetric basis is subsymmetric.
Proof Let X be a Banach space with a symmetric basis {Xn};'=l whose symmetric
constant is 1. For every increasing sequence of integers {niH"= 1 and every integer k
there is a permutation 7T such that 7T(i)=nt for I~i~k. Then, for any choice of
scalars (ah'''' ak), we have 111~1 a,x;ll= It~l aiX"(i)II= IIJ1 aiX"tll, which completes
the proof since k is arbitrary. 0
It follows easily from 3.a.3 and its proof that, for any symmetric basis, the
symmetric constant is larger than or equal to its subsymmetric constant.
That the notions of a symmetric and of a subsymmetric basis do not coincide
is shown by the following example, due to D. J. H. Garling [44]. Let Ybe the space
of all sequences of scalars y=(ah a2,"') for which
L la",li- 1/2 <00,

00
IIYII=sup
1=1
where the supremum is taken over all increasing sequences of integers {ni}f;.l' It
is easily checked that Y, endowed with the norm defined above, is a Banach space
whose unit vectors {e,,}:'=l form a ~ubsymmetric basis. However, the unit vectors
do not form a symmetric basis in Y. Indeed, for each fixed k, the vector y(k) =
(1, 2- 112, ... , k- 1/2 , 0, 0, ... ) is obtained from z(k)=(k- 1/2 ,(k-1)-1I2, ... , 1,0,0, ... )
k
by a suitable permutation of the integers but, in spite of this, Illk)ll= L n- 1 and
,,=1
k
Iiz(k)11 = ~ (k-n+ 1)-1/2n -1/2, i.e. sup Ii y(k) II=00 while sup IIZ(k) II<00.
n t 1 k k
In the past, spaces with a subsymmetric basis were studied only very little.
However, quite recently, there has been a growing interest in the notion of a sub-
symmetric basis especially since it has turned out that such bases arise naturally
in the local theory of Banach spaces. We shall discuss this in more detail in Vol. m.
The most interesting class of spaces with a symmetric basis is perhaps that of
Orlicz sequence spaces which were originally introduced by W. Orlicz [111]. An
Orlicz function M is a convex, non-decreasing continuous function on [0,(0) such
that M(O) =0 and lim M(t)=oo. To any Orlicz function M one can associate the
t-+ 00
00
space 1M of all the sequences x=(ah a2,''') of scalars such that L M(lanl/p)<oo
,,=1
for some p > 0. The conditions imposed on M make 1M into a Banach space when
it is equipped with the norm
The space 1M is called an Orlicz sequence space and it can be easily seen that, for
M(t)=t P with p";3I, 1M coincides with the classical space [p, The unit vectors
{e,,}:'=l form clearly a symmetric basis of hM= [e n ]:'=l and, in general, hM does not
coincide with 1M , We shall discuss this matter, as well as many other questions
regarding the structure of Orlicz spaces, in the next chapter.
Another class of spaces with a symmetric basis related to the lp spaces is that
of Lorentz sequence spaces. For every p ~ 1 and every non-increasing sequence
of positive numbers w={Wn}:'=l we consider the space d(w,p) of all sequences of
scalars x={al> a2,"') for which IIxll = sup C~l la"cnlIPwnf'P <00, the supremum
being taken over all permutations 7T ofthe integers. It is easily checked that dew, p),
endowed with the norm 11·11 defined above, is a Banach space. If inf Wn > 0 then
n
<Xl
dew, p) is isomorphic to Ip. Another trivial case occurs when 2: Wn <00: then,
n=l
d{w,p)~/<Xl' We shall exclude these two trivial cases, i.e. we shall assume that
<Xl
lim wn=O,
n-toO'J
2:
n=l
Wn=OO and wl=1 (to normalize the unit vectors). When these
conditions are satisfied dew, p) is called a Lorentz sequence space. We shall discuss
these spaces in Section 4.e.
There is also a general procedure to construct symmetric bases starting with an
arbitrary sequence of vectors. Let X be a Banach space and let {xn}~=1 be an arbi-
trary sequence of non-zero vectors in X. Denote by Z the completion of the space
of all sequences of scalars z={alo a2,"') which are eventually equal to zero, with
respect to the norm IIzllz= 9~~£1 s~p IIJI an8nx"Cnl/l. The space Z, equipped with
the nOrIn 1I·lIz, is a Banach space in which the unit vector basis is a symmetric
basis.
New spaces with a symmetric basis can be constructed from old ones by some
averaging processes. Sometimes, we obtain in this way new interesting spaces. For
instance, let fL be a probability measure on the interval [1, 2] and let X/L be the space of
all sequences ofscalarsx={al, a2,"') with the norm IIxll/L=[ Ct lanl P) IIp dfL(P) <00.
We shall see in Vol. II that X/L is a subspace of Ll{O, 1). These spaces, as well as
spaces obtained by averaging suitable OrIicz norms, playa role in the investigation
of the structure of Ll{O, 1).
Let X be a Banach space with a normalized symmetric basis {Xn}~=l whose
symmetric constant is 1. In the study of subspaces of X a special role is played by
block bases with constant coefficients {with respect to {Xn}~=l)' Let a={aj}i'=l be
a sequence of consecutive disjoint finite subsets of the integers (i.e. the greatest
integer in aj precedes the smallest integer in ai+l, for allj), let Uj denote the number
of elements in aj and define the operator P q on X, as follows.
The operator P q, thus defined, is called the averaging projection or the conditional
expectation with respect to a and has the following property.
Proposition 3.a.4. For every X, {Xn}~=l and a as above, the averaging projection P q
is a norm-one linear projection in X whose range is the closed linear span of the vectors
Uj= 2: xn,j= 1,2, .... In other words, the closed linear span of any block basis with
neq
constant coefficients is complemented.
.if {Xn }:'=l is only a normalized subsymmetric basis whose subsymmetric constant
is 1, the same holds with the exception that IIP"II~2.
Proof Fix an integer k and let {1Tj}r~ki, with m(k) = al . a2" .ak, be the set of all
distinct permutations 1T such that 7r leaves invariant all the integers outside
Ul U U2 U ... U Uk and, for l~j~k, 7T(UJ=Uj and 1T restricted to Uj acts as a cyclic
permutation on Uj (by cyclic permutation T on a set {nj}f=b where nl <n2 < ... <nk>
we mean a permutation of the form Tlnj)=ni+j(mOdk), l~i~k, l~j~k). For every
vector y= n~1 anXn E X and every 1 ~ i~m(k) we have lint a",(n)Xnll= IIYII, in the
symmetric case, and lint a"t(n)Xn"~21IYII, in the subsymmetric case.

By averaging over i=l, 2, ... , m(k) we get that
and therefore also "j~JCfuj an)/aj)ujll, are ~llyll, in the symmetric case, and
~21Iyll, in the subsymmetric case. Since k is arbitrary it follows that P(J is a well-
defined operator on X whose norm is ~ 1, respectively ~ 2. Obviously, P (J is a
projection whose range is [Uj]i'= l ' 0
In the case {X n};:'=1 is a symmetric basis 3.aA clearly remains true even if the
u/s are taken to be a block basis with constant coefficients of some permutation of
{Xn };:'=I' As pointed out by J. T. Woo [145], there exist also non-symmetric un-
conditional bases which have this property, i.e. in which every block basis with
constant coefficients of any permutation of the basis spans a complemented sub-
space. Such a basis is, for example, the natural basis of the space X oo ." r~ 1 (see
the remark following 4.d.7 below).
As a first application of 3.aA, we bring the following.
Proposition 3.a.5. Let X be a Banach space with a symmetric basis {X n};:'=I' For every
u={Uj}i'=1 denote by U the closed linear span of the vectors Uj = 2: xnoj=1,2, ....
nECTj
Then X is isomorphic to X EB u.
Proof We shall use a simplified argument due to W. B. Johnson. Let {V n };:'=1 be
a normalized block basis with constant coefficients of {X n };:'=1 so that every possible
normalized block with constant coefficients of {X n };:'=1 appears infinitely many
times in the sequence {V n};:'=I' This property of {V n };:'=1 ensures that V= [V n ];:'=1
satisfies V EB V~ V and V EB U~ V for every U as above. It follows that we can
apply the decomposition method for the spaces X EB V and X since X EB X~ X,
XEB V~(XEB V) EB (XEB V), and, by 3.aA, XEB Vis a complemented subspace
of X. We conclude that
X~XEB V~XEB VEB U~XEB u. 0
We are now prepared to complete the proof of the fact that co, II and l2 are the
only Banach spaces which have, up to equivalence, a unique unconditional basis.
Proof of 2.b.lO. Assume that a Banach space X has, up to equivalence, a unique

normalized unconditional basis {Xn}:=l' Then, for each permutation Tr of the
integers, {x,,(n)}:=l is equivalent to {Xn}:=l and therefore {Xn}:=l is a symmetric
basis of X. Let {Uj}i'=l be any normalized block basis with constant coefficients of
{Xn}:=l and let U=[Uj]i'=l' By 3.a.5, X~XEB U, which implies that
{Xl' U1, X2, U2"'" X", Un,"'} is a normalized unconditional basis of X. Thus, {Uj}i'=l
is equivalent to {Xn}:=l and, by the second remark following 2.a.9, it follows that
{Xn}:=l is a perfectly homogeneous basis, i.e. {X n}:=l is equivalent to the unit vector
basis of Co or Ip for some p'?J; 1. Since each of the spaces Ip with p # 1, 2 has an
unconditional basis which is not equivalent to the unit vector basis (see the dis-
cussion preceding 2.b.1O), the only remaining possibilities are Co, hand 12 , 0
It is interesting to point out that a Banach space with an unconditional basis

can have either a unique normalized unconditional basis, as in the case of 11, 12 and
co, or uncountably many non-equivalent unconditional bases. The construction of
infinitely many unconditional bases is quite simple (cf. [52]). Let X be a space
having a least two non-equivalent normalized unconditional bases. Then, we can
easily check that X admits a normalized unconditional basis, say {Xn}:=l, which
is not symmetric. For such a basis there is a partition of the integers into disjoint
infinite subsets {Nj}i'=l so that, for eachj, {xn}neNi is not a symmetric basis sequence.
Let Trj,l be a permutation of N j such that {xn}neNi and {X"i,l(n)}neNi are not equiva-
lent. For each sequence {7JU)}i'=l of zeros and ones, we consider the basis
a>
jl}l {x"i .•Ul(n)}neN" where Trj, 0 denotes simply the identity on N j • It is easily seen that
distinct sequences of zeros and ones give rise to non-equivalent unconditional
bases of X.
Proposition 3.a.4 is also useful if we desire to study the behavior of the sums of
vectors of a normalized symmetric or subsymmetric basis.
Proposition 3.a.6. Let {X n}:=l be a normalized symmetric (subsymmetric) basis

of a Banach space X with symmetric (subsymmetric) constant equal to 1. Let {X:},~)=l
be the sequence of the functionals biorthogonal to {Xn}:=l and, for each n, put
A(n)=iix1 +x2+"'+xnii and p,(n)=iixt+x:+···+x:lI. Then,
A(n)p,(n)=n (n~A(n)p,(n)~2n) ,
for all integers n.
Proof. Since C~lX:)(J1 xk)=n the inequality A(n)p,(n)'?J;n is true for every
space with a basis. By 3.a.4, applied to a single block a, we obtain the reverse
inequality (i.e. A(n)p,(n)~n in the symmetric case or A(n)p,(n)~2n in the sub-
symmetric case). Indeed,
p,(n)=llik= x://=sup (ik ak)//Iik= akXk//~n/A(n),

1 {a,,} =1 1
(respectively,2n/A(n». 0
For every space with a normalized symmetric basis {Xn}:'=1 with symmetric
constant equal to 1, {>t(n)}:'=1 is a non-decreasing sequence which tends to 00
unless {Xn }:'=1 is equivalent to the unit vector basis of Co (see e.g. the proof of
2.a.9). In most of the known examples >.(.) is even a "concave" function on the
integers in the sense that {>t(n + 1) - >.(n)}:,= 1 is a non-increasing sequence. This
fact is not true in general but we have the following result (cf. [148]).
Proposition 3.a.7. Let (X, II·ID be a Banach space with a symmetric basis {Xn }:'=l
whose symmetric constant is equal to 1. Then there exists a new norm 11·110 on X
such that:
(i) Ilxll:r;; Ilxllo:r;;21lxllfor all x EX,
(ii) The symmetric constant of {Xn }:'=1 with respect to 11·110 is equal to 1,
(iii) if we put >'o(n)=/tt xillo' n= 1,2, ... then {>'o(n+ I)->'0(n)}:'=1 is a non-
increasing sequence, i.e. >'0(·) is a concave function on the integers.
Proof Let {X~}:'=1 denote again the sequence of the functionals biorthogonal to
{X n }:'=1 and let >.(n)=lllt XIII, fL(n)=ltt xiii for n~ 1 and >'(0)=0. For each in-
n
teger n, we put further >'o(n) = sup L [>'(k;) - >'(ki -1)], where the supremum is
1=1
taken over all n-tuples of integers {k1 < k2 < ... < k n}. By 3.a.6 we get, for every
n~I,
>.(n):r;; >'oCn)= sup{ L (>'(ki)- >'(ki -1))+ L CkdfLCki)- (ki -I)/fL(k; -I)}
~~n ~>n
:r;;>.(n)+n/fL(n)=2>.(n) .
Now, fix 8> 0 and n and choose suitable integers so that
n-1
>'0(n-I)<8/2+ L
;=1
(>'(hi)->'(h;-I)),
n+l
>'0(n+I)<8/2+ L (>'U;)->'Ui-I)).
;=1
It follows immediately that >'o(n-I) + >'o(n+ I):r;; 8+ 2>'0(n) which shows that
>'oCn+I)->'o(n):r;;>'o(n)->'o(n-l) since ~>O is arbitrary. The sequence >'o(n) is
therefore concave. Observe that a similar argument shows also that k>'o(n):r;; n>'o(k)
for all I:r;; k:r;; n. In order to complete the proof it suffices to construct a new norm
11·110 satisfying (i), (ii) and such that 11ft xillo = >'o(n) for all n. This is achieved by
setting,
L anXn EX,
00
for X=
n=1
where the inner supremum ranges over all the permutations of the integers. The
norm 11·110 has indeed all the desired properties since, by 3.aA,
for all x, nand l' as above. 0
The converse to 3.a.7 is also true in the sense that, for every concave non-
decreasing sequence of positive numbers P'n}:'=l, there exists at least one Banach
space X having a symmetric basis {Xn}:'=l with symmetric constant equal to 1
such that IIJI Xtll=An for every n. For instance, we can take the space dew, 1) with
the sequence w={Wn}:'=l defined as follows: Wl=A 1 and, for n>l, wn=An-An-l'
The sequence {A n}:'=l does not determine the equivalence class of {X n}:'=l except
for the following two extreme cases: if sup An <00 then {Xn}:'=l must be equivalent
n
to the unit vector basis of Co and, if lim sup A(n)jn>O then {X n}:'=l is necessarily
n
equivalent to the unit vector basis of II (this last fact is an immediate consequence
of 3.aA or of 3.a.6). If we exclude these two cases there are always many mutually
non-isomorphic spaces with a symmetric basis {Xn}:'=l such that IIJI XII = Am for
all n. For example, in addition to the space dew, 1) considered above, we can take
the closed linear span of the functionals biorthogonal to the unit vector basis of
d(w', 1), where w'={W~}:'=l is defined as follows: W~=ljAl and, for n>l,
w~=njAn-(n-l)jAn_l (see the discussion following 4.eA below).
Besides block bases with constant coefficients there is another type of special
block bases which is of interest in the study of symmetric bases.
Definition 3.a.8 [4]. Let X be a Banach space with a symmetric basis {Xn }:'=l and
let Nt = {ni, 1 <ni, 2 < ...}, i= 1,2, ... be any sequence of disjoint infinite subsets of
00
the integers. For every O#a= 2: anxn E X, the sequence {u\a»);";"l> defined by
n=l
00
2: ajxnI,'- -, i= 1,2,00', is called a block basis generated by the vector a.

ui a)= j=l
Clearly, {Ui a)};";"l is a symmetric basic sequence and its equivalence type does
not depend on the particular choice of the sets {NI);";"l' Observe also that we use
the term "block basis" though, strictly speaking, the u\a),s do not form a block
basis of {Xn}:'=l'
The following elegant result concerning block bases generated by one vector
was proved in [15].
Theorem 3.a.9. Let X be a Banach space with a symmetric basis {X n}:'=l' Let
00
a= 2: anxn#O and let {U\a)}i";"l be a block basis generated by the vector a. Then
n=l
Ma)}i";"l is equivalent to {X n}:'=l if and only if [ul a)]i'=l is a complemented subspace
of X.
Proof. Fix an integer h so that a,,#O. Then, for every x E X, put
2: (bl,,,/a,,)ujIX J,
00
Px=
1=1
where bl ,,, is the coefficient of x n !.,. in the expansion of x with respect to the basis
{Xn};:"=1' If {uja J}j;,1 is equivalent to {Xn};:"=1 then P is a bounded linear projection
from X onto [ujIX)]l'= l ' Indeed, if Ka is chosen so that lIi~1 ClujlXJl1 ~ Kalll~1 CiXt11 for
every choice of {Ct}t~1 then IIPII~KIX/lahl.
<Xl
To prove the converse we first notice that whenever 2: ciuja) converges so does
1=1
~ Cia"Xn! • h'• and thus also 1=1

1=1
~ CiXt. Assume now that [uja J]I";,,1 is a complemented
<Xl
subspace of X and consider a series 2: CiXi which converges in

1=1
X. Choose an in-
creasing sequence of integers {km}~=1 so that III~m CIXIII~2-m for all m and, for
m
km~i<km+h set via)=;~1 ajxnt .1, ,\=lluja)-vla)11 and wi IX )= (ujaJ_ViIX»)/A i. Then
uja) = via) + AiWia) for all i (notice that this is true even if Ai=O in which case wi a ) is
not well-defined and can be taken to be an arbitrary unit vector). We have,
<Xl
which shows that 2: civja) converges. In order to conclude the proof we have to
1=1
prove that ~ CiAiWia) converges too. This however is an immediate consequence

1=1
of 2.a.lI. 0
In the spaces Co and lp, p> 1 all the block bases generated by one vector are
equivalent to the unit vector basis. This can also happen in other spaces with a
symmetric basis, for example, in the Lorentz sequence spaces dew, p) with
w={n- 1}:'=1 and p> 1 arbitrary (see 4.e.5). However, if this situation occurs both
in the space and in its dual then the underlying space has to be isomorphic to Co
or to lp for some p> 1. This was proved by Z. Altshuler [2].
Theorem 3.a.10 . Let {Xn};:"=1 be a symmetric basis of a Banach space X and let
{X;}:'=1 be the functionals biorthogonal to {X n};:"=1' Then, X is isomorphic to either
Co or to lpfor some p > 1 if(and only if) all the block bases of {xn};:"= 1 and of {x;};:"= 1,
which are generated by one vector, are equivalent to {Xn};:"=1, respectively {X;};:"=1'
Proof. We assume, as we may without loss of generality, that the symmetric con-
co <Xl
stant of {X n}:'=1 is equal to 1. Next, we observe that if a= 2: anxn and /3= 2: bnxn
n=1 n=1
are two non-zero vectors in X then IIJl aiU}!l)11 IIJl biUl"'lll.

= Using this fact and the
uniform boundedness principle it follows immediately that if all block bases of
{X n};:'=l, which are generated by one-vector, are equivalent to {X n};:'=l then there
exists a constant K so that IIJl blul"')II~K Ilexll·II.BI1 for all ex, f3 EX. Using the
similar inequality for {Xi};:'=l and a standard duality argument we get that
K- 1 1Iexll·IIf311 ~ IIJl biu;"')II·
Put A(n) = lilt Xii and use the preceding inequalities for ex = J1 Xi and f3 = Jl Xi'
with nand k being arbitrary integers. We get that
K-l~ A(nk)/A(n)'A(k)~K, n, k= 1,2, ....
Hence (see the proof of 2:a.9), either sup A(n)<oo or K-l~A(n)/nl/P~Kfor some
n
p ~ 1. If sup A(n) <00 then X is isomorphic to Co. If p = 1 then X is isomorphic to
n
i 1 , by the remark following 3.a.7. We can therefore assume that p > 1. By duality
and 3.a.6 it suffices to prove that, for every choice of scalars {b i }Y=l, we have
We shall prove this fact only for 11=2; it will become clear from the proof how to
generalize the argument for a general n.
Put f31 = b1X1 + b2X2' By the inequalities established above we have
where f33=bh1 +bib2(X2+X3+X4)+b1b~(X5+X6+X7)+bgX8' Continuing induc-

tively we get that
m
where f3m= L
j= 0
bT-jb~ L
ieClj
Xi with {Uj}f=o being disjoint subsets of integers for
which O'j= (}). Hence,
m m
11f311Im~Km-lllf3mll ~Km-l L Ibllm-jlb2IjA(O'j)~Km j=O
j=O
L Ibllm-jlb2Ij(O'j)1/P
b. Subspaces of Spaces with a Symmetric Basis 123
By Holder's inequality we deduce that
and thus, by taking the m'th root and letting m ~OO, we obtain the desired in-
equality, i.e. 11,8111 ~ K(lb 1 P + Ib 2 Ip )1/p • 0
1
b. Subspaces of Spaces with a Symmetric Basis
The properties of the spaces with a symmetric basis might create the impression
that their subspaces, or perhaps their complemented subspaces, should inherit
some of the symmetric structure. The next result (cf. [86]) shows that this is not
the case.
Theorem 3.b.t. Every Banach space with an unconditional basis is isomorphic to a

complemented subspace of a space with a symmetric basis.
Proof We may assume without loss of generality that X is a Banach space with a
normalized unconditional basis {Xn }:=l whose unconditional constant is equal to 1.
We denote by Xo the algebraic span of {Xn}:=l and by Yo the vector space of all
sequences of scalars which are eventually equal to zero.
Let {IXtH"=lo {,8t}i'=1 be decreasing sequences of positive numbers and {nlh~l
an increasing sequence of integers such that IX1 =,81 =n1 = 1 and, for i> 1, IXt/IXt-1 <
1
2-t,,8jNt_l<2-t-1 and ex;,8;n;=I, where N i = 2:. nj. We consider now the linear
j=l
Ie
operator Tfrom Xo into Yo which maps a vector X= 2: AtXt E Xo into the sequence
;=1
y=Tx=(al, a2,"') E Yo so that a1 =A1IXl> af=AjIXj for N t - 1<j~Ni> i=2, 3, ... , k
and af=O for j>NIe •
LetKbe the convex hull of all sequences z=(cl> C2,"') E Yo such that Icnl ~ la,,(n)j,
n= 1,2, ... , for some permutation.". of the integers and some sequence y=(al> a2,"')
which belongs to the image, under T, of the unit ball of Xo. With the aid of K we
introduce a norm on Yo by putting
Ilzll=inf{p>O; p-1zE K}, ZE Yo.
Then the closure of K in the normed space Yo coincides with its unit ball and
therefore Tbecomes a norm-one operator. It follows that Thas a unique continuous
extension to a norm-one operator from X into the norm completion Yof Yo. We
Ie
shall show that T is an isomorphism. Fix an x = 2: AtX; E Xo with Ilxll = 1. By the
1=1
Ie Ie
Hahn-Banach theorem there exist scalars {tIW= 1 such that 2:
t=l
t;-'t = 1 and 2:
1=1
Itift! I~ 1
whenever IIJ1fL;xill~1 (in particular, It;l~l for i=I,2, ... ,k). We define now a
linear functional y* on Yo by putting
Ie N,
y*(z)= 2:
1=1 j=N,-l +1
2: f;{3icj, Z=(C1, C2,"') E Yo.
Since aifJ;n; = 1 for all i we get that

Ie N, Ie
y*(Tx) = 2:
1=1 j=N,_l+l
2: ti/3;A;a, = 2: t,A; = 1 ,
1=1
i.e. IITxl1 ~ l/lly*ll. Hence, T would be an isomorphism provided we could show

that Ily*ll, which a priori depends on x, is bounded by a number independent of
the choice of x.
co
Let 7T be a permutation of the integers, let u= L: fLlX;
1=1
be a vector in Xo with
norm ~l and put Tu=(b h b 2, ... )E Yo. For every l~i~k we split the set ai=
U; N t - 1 <j~N.} into three disjoint (possibly empty) subsets: ai, 1 =U E ai; 7T(j) Eat},
ai, 2 = U E a;; 7T(j) ~ N;-l} and ai, 3 = U E ai; 7T(j) > Nt}. Then, by the choice of
{a;}l=l, {fJ;h";.l and {nih";. 1, we have for every i~k
2: ItifJib,,(j) I~ N i - 1fJi ~ 2 -;-1 ,

jerTt,2
2: ItifJ;b,,(j) I~n;fJiai+1 ~2-i-1 .

f ea t,3
Hence, if fiTu denotes the vector (b,,(ll' b,,(2l"") E Yo then

Ie Nt Ie
ly*(fiTu)l~ 2:
t=l f=Nt-l+1
2: ItifJib,,(j)I~ 2:
;=1
(ItifLd+2.2-1-1)~2,
i.e. Ily*11 ~2. This implies that IITxl1 ~ 1/2, for all x E X with Ilxll = 1, i.e. T is an
isomorphism from X into Y. This completes the proof since the unit vectors of
Yo form a symmetric basis in Yand TX is complemented in Y in view of 3.aA (TX
is clearly the closed linear span of a block basis with constant coefficients). 0
Two variants of 3.b.l are presented in the following theorem.
Theorem 3.h.2. Every reflexive (uniformly convex) Banach space with an unconditional
basis is isomorphic to a complemented subspace of some reflexive (uniformly convex)
space with a symmetric basis.
The reflexive case has been originally proved by A. Szankowski [137] and the
uniformly convex one by W. J. Davis [22]. We present here the approach of Davis
which, besides proving the uniformly convex case, provides an alternative proof
of 3.b.l, as well as of the reflexive case. The approach of Davis, which in essence
is an interpolation method (based on some ideas from [24]) proceeds as follows.
Let E and F be two Banach sequence spaces so that the unit vectors {ek}:~l
form a symmetric basis with symmetric constant equal to 1 in both E and F. We
assume that
(i) For each sequence of scalars a=(ah a2, ... ) E F, IIaIIE::::; lIaliF and
(ii) lim Ain)/Ap(n)=O, where,
n .... '"
for n~ 1, we set AE(n) = II i
k~l
ekll
E
and
For every number m ~ 1 we define a new norm on E by putting
lIallm=inf {(II,BII~+ IIYII~)1/2; a=m-1,B+my with,B E E and y E F}.
Let a=m-1,B+my be as above; then lIaIlE::::;(m-2+m2)1/2(1I,B1I~+ IIYII~)1/2::::;

2m(II,BII~+ lIyIlW /2 , which implies that IIaIIE::::;2mllali m for every a E E. On the
other hand, we clearly have Iiallm::::; mllallE' a E E. A better estimate can be obtained
for y E F, namely lIyllm::::;m- 1 I1yIlF'
It follows from this discussion that, for every m ~ I, the space Em = (E, II· 11m)
is isomorphic to E and the identity mapping of Finto Em has norm ::::;m- 1 •
We compute now the value of AEm(n)=/ikt eklL, n~ 1.
Lemma 3.h.3. With AE, AF and AEm as above we have
for every integer n and every m ~ 1. For a fixed n the maximal value of AEm(n) is
equal to
and is attained at m=(AF(n)/AECn»1/2.
Proof Since the norm of any sequence in E or in F does not increase when the
sequence is replaced by its restriction to the first n components or by its average
over all possible permutations of the integers {I, 2, ... , n} it follows that
However, for any constants Band C so that Bm -1 + Cm = I, we have
This proves the first assertion of the. lemma. The second assertion is an immediate
consequence of the first. 0
Let now X be a Banach space with a normalized unconditional basis {Xn}:'=l

whose unconditional constant is equal to 1. For every increasing sequence of
<Xl
numbers {1~m1<m2<···} so that L m;;l<oo we let Y=Y(E, F, X, {m n}:'=l)

n=l
be the space of all sequences of scalars ex E E for which
The condition imposed on the sequence {m n}:'=l ensures that the identity mapping
of F into Y is a bounded operator. Indeed, for y E F, we have
In particular, it follows that the unit vectors belong to Y and, as easily checked,
they form there a symmetric basis with symmetric constant equal to 1. We can
prove now the embedding result (cf. [22]).
Proposition 3.h.4. For every E, F and X as above there exists an increasing sequence
co
of numbers {m n}:'=l with L
m;;l <00 such that Y = Y(E, F, X, {m n}:'=l) contains
n=l
a complemented subspace isomorphic to X. Moreover, every infinite dimensional
subspace of Y contains an infinite dimensional subspace which is isomorphic to a
subspace of E or to a subspace of x.
Proof. We recall that by 3.b.3, for each fixed n, the maximal value of AEm(n) is
attained at m(n) = (AF(n)/Ain»1/2. The condition (ii) imposed on E and F therefore
implies that men) -00 as n _00. Thus, we can construct an increasing sequence of
integers {nj}i'=l so that the corresponding numbers mj=m(n,),j= 1,2, ... satisfy
i-1
L m;+mr ;=j+1
L m;-1<2-
co
m,-l j - 1 for allj.
;=1
Next, we choose a sequence {O"j}J'=l of disjoint subsets of integers so that u;=nj

and put
Then, by 3.b.3 and the definition of the m/s, we have JJu;JJmj= 1 for allj and, for
i=lj,
JJUjJJm, = AEm ,(nj)/(2 -1 AE(n,)· AF(n,»1/2
= 2 1/2/(m;- 2. Ai l(n,)· Ainj) + m~ . Ainj)· Ai 1(nj»1/2
=21/2/(mi-2m~+m~mi2)1/2~21/2 min (m;/mj, m,/m;) .
The lacunarity condition imposed on the sequence {mAi'=l implies that

2: Ilujllm,~21/2.2-i-l for all i. Hence, for any choice of {Cj}i'=1, we have
jM
which further implies that in Y= Y(E, F, X, {mi}I"=1) we have
I J=1i CjUj11 = I 1=11=1

Y
~ II.~ CjUj11 m,Xiii ~ I 1=1
X
~ cix,11 X
+ 2 -1/2 max
'
led ,
This shows that {X n };;'=1 is equivalent to the block basis {Uj}j=l in Yand, since the
u/s have constant coefficients, [Uj]j=1 is complemented there, by 3.aA.
Let now Vbe an infinite dimensional subspace of Yand suppose that the formal
identity mapping of V into E is a strictly singular operator. Then, there exists a
basic sequence {WAi'=1 in V with Ilwjlly= 1 for allj and Ilwjllr-?o- O. Since each of
the norms 11·llml' i = 1, 2, ... used in the construction of Y is equivalent to the norm
of E we can find an increasing sequence of integers {qk}k'=1 and a subsequence
{Wik}k'=1 of {Wj}i'=1 so that I IIWiJcllm,Xll1 x <2- k -\ k?:-l and
II I=qk
The basic sequence {wik}k'= 1 is easily seen to be equivalent to the block basis
qk-1
Yk = 2: I[wik[ [mix;, k= 1, 2, ... of {Xn };;'=1. 0
j=qk-l
We note that 3.bA already provides an alternative proof for 3.b.l. We can
also give now the
Proof of 3.h.2. Let X be a space with an unconditional basis {X n};;'=1. We may

assume without loss of generality that {X n };;'=1 is normalized and has unconditional
constant equal to l. Take E=lq and F=lp with 1 <p<q<oo and observe that
conditions (i) and (ii) are satisfied with this choice of E and F. Using 3.bA we
construct the space Y = Y(E, F, X, {m n };;'= 1) so that it contains a complemented
subspace isomorphic to X. We recall that the unit vectors {e n };;'=1 form a symmetric
basis in Y.
Assume now that X is reflexive. Then, by the last part of 3.bA, Y contains no
subspace isomorphic to 11 or to co. Consequently, by l.c.12, Y is a reflexive space
too.
We consider now the case when X is a uniformly convex space. Uniform con-
vexity will be studied quite in detail in Vol. ll. For the present proof we just recall
that the modulus of convexity Ou of a Banach space U is defined by
ou(e)=inf {1-llu+vll/2; Ilull = Ilvll = I, Ilu-vll ~e}, 0<e<2.

A Banach space Uis uniformly convex provided ou(e»O for every 0<e<2.
We note that in our case the spaces Em, m ~ 1 are uniformly uniformly convex in
the sense that their moduli of convexity OEm satisfy inf OEm(e»O for every 0<e<2.
m
Indeed, the mapping Qm: (E EB F)2 -+ Em, defined by Qm(f3, y)=m- 1{3+my for
{3 E E and y E F, maps the unit ball of (E EB F)2 onto the unit ball of Em. Thus, for
any m~ 1 and 0< e<2, we have OEm(e)~oCE(j) Fl2(e). By using an argument of M. M.
Day [27] we shall show that, under the present assumptions, the direct sum
Z= C~l EB Em n) x is also a uniformly convex space. This direct sum is defined as
the space of all the sequences {aCnl}:'=l with aCnl E EmJor alln for which II{aCnl}:'=lllz=
IIJl lIacnlllmnXnllx <00. Since Y is clearly a subspace of Z (more precisely, the
"diagonal" of Z) the uniform convexity of Z would imply that Y is also a uniformly
convex space, thus completing the proof.
Let {aCnl}:'=l and {{3C nl}:=l be two elements of Z so that
for some 0 < e < 2. We first consider the case where Ilacnlllmn = II{3cnlllmn for all n. Put
an = Ilacnll Imn' Cn= Ilacnl - {3c nlll mn , n= 1,2, ... , a={n; cn/an > e/2} and o(e)=inf 0EmCe).
m
II{acnl+{3cnl}:=lllz~ 211n~1 anCI- o(cn/an))XnIL

~ 211C1- 0(e/2)) 2:
nEd
anXn + 2: anXnl1 X .
n~d
We also have 1~llntaanXnllx~C2/e)lInt(,cnXnllx' which implies that IIn~aanXnllx~

!lIn~a CnXnllx ~ e/4. Thus, by the uniform convexity of X, we get that
II{aCnl+{3Cnl}:=11Iz~2(1-1]Ce)) ,
where 1](e) = 0xC eo(e/2)/8).

In the general case put yCnl = Ilacnll Imn . {3c nl/1 I{3CnlIImn' n= 1,2, .... Since IlyCnlllmn=
Ilacnlilmn for all n we have II{yCnl}:=lllz= 1. We observe that
II{acnl _yCnl}:=lllz ~ II{acnl - {3Cnl}:=lllz -11{{3Cnl _yCnl}:= lliz
~e-IIJ111{3Cnl_yCnlllmnXnllx
=e-Iln~llilacnlilmn -11{3cnlllmnIXnIL·
If Iln~111Ia(n)II7nn-II,8(n)II7nnIXnllx~7](e/2) then, by the uniform convexity of X, we

get that
Otherwise, lI{a(n)-y(n)}:=11Iz~e-7](e/2)~e/2 which, by the first part of this proof,

implies that
Ij{a(n) +.a<n)}:= lllz:::; II{a(n) +y(n)}:=lliz + II{,8(n) _y(n)}:= lliz
:::; 2(1-7](e/2))+7](e/2):::; 2(I-7](e/2)/2) .
Hence, Z and therefore also Y, are uniformly convex spaces. D
As an immediate application of 3.b.2 we get that each of the spaces Lp(O, 1),
1 <p<oo, p=l=2 is isomorphic to a complemented subspace of a uniformly convex
space with a symmetric basis. It will be shown in Vol. III that none of these spaces
themselves has a symmetric basis.
An interesting application of 3.b.I is connected with the universal space U1 of
PeIczynski, which was introduced in 2.d.IO(a). We recall that U1 is a space with
an unconditional basis {Un}:=l, which is universal in the sense that every other
space with an unconditional basis is isomorphic to a complemented subspace of
U1 • The space U1 is determined uniquely, up to isomorphism, by this universality
property.
It turns out that U1 has a symmetric basis. Indeed, by 3.b.I, this universal space
U1 isisomorphic to a complemented subspace of a space Y with a symmetric basis.
It follows that Yitselfis a universal space for all spaces with an unconditional basis
and therefore, by the uniqueness of Ul> we get that Y is isomorphic to U1 • Hence,
U1 has a symmetric basis. This fact is somewhat surprising if we recall that, by its
construction, the natural basis {Un}:=l of U 1 has the property that every other un-
conditional basis is equivalent to a subsequence {Unj}i'=l of {Un}:=l.
We pass now to the question of uniqueness, up to equivalence, of a symmetric
basis. This property, which is shared by a relatively large class of spaces with a
symmetric basis, is quite useful in applications and it is applied mostly in the
following typical situation. Let X and Y be two Banach spaces with symmetric
bases {Xn}:=l, respectively {Yn}:=l, and assume that {Xn}:=l is, up to equivalence,
the unique symmetric basis of X. In this case, if we wish to check whether X is
isomorphic to Yor not it suffices to check whether {Xn}:=l and {Yn}:=l are equiva-
lent which, in general, is much easier.
Proposition 3.b.5. The spaces Co and lp with l:::;p <00 have, up to equivalence, a
unique symmetric basis.
Proof. Let X be either Co or lp for some I:::;p <00 and let {Xn}:=l be a symmetric
basis of X. Let {e~}:=1 denote the sequence of the functionals biorthogonal to the
unit vector basis {en}:=l of X. Choose a subsequence {XnJj;,l of {Xn}:=l such that
lim e;xnl exists for each In. If all these limits are equal to 0 it follows from l.a.12
i~ 00
and 2.a.1(ii) that {Xn.}i'=b and thus also {Xn},';'=b are equivalent to the unit vector
basis of X. If lim e;xnl"",O for some m then a subsequence of {Xn,H';'l' and therefore
!~OO
also {X n },';'=l, are equivalent to the unit vector basis of II' 0
Remark. The proof actually shows that every symmetric basic sequence in X is
equivalent to the unit vector basis.
In contrast to the case of Co and Ip , the universal space U1 (for all the spaces
with an unconditional basis) has uncountably many mutually non-equivalent sym-
metric bases. Moreover, for every p?;; 1, the space U1 has a symmetric basis
{U~fl},';'=l such that, for every e > 0, there is a constant KE > 0 with the property that
KeCI[an[P+B)l/(P+B)~
n
IIIn anuii'JII ~ (In [an[P)lIP ,
for every choice of {an}. Indeed, take F=lp and E=IM p, where Mp(t) = tPI(l + [log ff),
and construct the space Yp= Y(E, F, U1 , {m n },';'=l) so that Yp contains a comple-
00
mented subspace isomorphic to U1 and 2: m;; 1 < 1. This is possible in view of

n=l
3.bA and it implies that U1 ;:;; Yp , by the uniqueness of the universal space U1 •
To complete the proof we observe that
for some constant K> 0 and every sequence ex E lp. This means that the unit vector
basis of Yp is mapped, by the isomorphism between Yp and U1 , into a symmetric
basis {uii'J},';'= 1 of U1 having all the desired properties.
Unlike the case of uniqueness of unconditional bases (cf. 2.b.1O) it does not
seem likely that there is a way to describe those spaces having a unique symmetric
basis. It is of interest to point out that from the isometric point of view a symmetric
basis is unique. More precisely, we have the following result due to A. Pelczynski
and S. Rolewicz (cf. [124, Th. IX.8.3]). Let X and Y be Banach spaces having
normalized bases {Xn},';'=b resp. {Yn},';'=l, whose symmetric constants are equal to l.
Then, X is isometric to Y if and only if the basis {x n},';'=1 is isometrically equivalent
to {Yn},';'=b i.e. the map T: X --;.- Y defined by TC~l anxn) = n~l anYn is an isometry.
Another problem related to the uniqueness of symmetric bases involves the
possible number of mutually non-equivalent symmetric bases of one space. In all
the examples studied so far (see also the Section 4.b, 4.c and 4.e for the cases of
Orlicz and Lorentz sequences spaces) either there is a unique symmetric basis or
there are uncountably many non-equivalent symmetric bases. Hence, we have the
following question.
Problem 3.b.6. Is there any Banach space with exactly 2 (or any other finite number,
or ~o) non-equivalent symmetric bases?
In 2.a.6 we have introduced the notion of a prime space, i.e. a Banach space
in which all infinite-dimensional complemented subspaces are isomorphic to each
other. There are relatively few prime spaces (the only known examples are Co and
lp, l~p~CX)). We introduce now a related class of spaces which is however much
larger.
Definition 3.b.7. A Banach space X is said to be primary if, for every bounded
projection Q on X, either QX or (J - Q)X is isomorphic to X.
In Vol. II we shall prove that the common classical function spaces C(K) (with
Kbeing a compact metric topological space) and LiO, 1), 1 ~p~CX) are all primary.
We shall show now that the universal space U1 (of 2.d.1O(a)) is primary. This fact
is proved by using the decomposition method of Pe1czynski described in the proof of
2.a.3 (observe that this method can be applied since
Ul~(Ul ® U1 ® ... EB U1 EB "')2)

and the following result of P. G. Casazza and Bor-Luh Lin [16].
Proposition 3.b.S. Let X be a Banach space with a subsymmetric basis {Xn}:=I'

Then, for every bounded projection Q on X, either QX or (J - Q)X contains a sub-
space isomorphic to X which is complemented in X.
Proof We may assume that {Xn}:=1 is a normalized basis with subsymmetric

constant equal to 1. We can also assume that Xn ~ 0 since, otherwise, X is iso-
00
morphic to 11 which is prime. By putting QXn= 2:An, kXk, n = 1, 2, ... we can find
k=1
an infinite subset Nl of the integers so that either IAn, nl ;;::: 1/2 or 11- ~n, nl ;;::: 1/2 for
all n E N 1 • If, for instance, the first alternative holds then, by l.a.12, there are
an infinite subset N 2 ={nl <n2<"'<nj<"'} of Nl and a block basis {Uj}f'=1 of
{X n}:=1 such that {QXnj}i'=1 satisfies IIQxnj-ujll < 1/5.2i+ 3 ·IIQII for allj.
Let {X;n:=1 be the sequence of the functionals biorthogonal to {Xn}:=I' Then,
IX;jujl;;:::IX;jQxnjl-I/2i+3;;:::IAnj,njl-I/24;;:::1/4 for allj. It follows that the con-
vergence of a series ~ ajQxn-> which is equivalent to that of ~ ajuj, implies the
j=1 ' j=1
convergence of I
j=1
ajXnj' On the other hand, it is obvious that ~ ajQxnJ converges
j=1
00
whenever 2: ajxnj is a convergent series. Hence, since {Xn}:= 1 is sub symmetric,

j=1
we get that [QXnj]i'=1 is a subspace of QX which is isomorphic to X. Using the
equivalence between {Uj}i'=1 and {X n}:=1 it is easy to show that the operator
P: X -> X, defined by
is a bounded projection onto [Uj]i'=1 with IIPII~51IQII. Thus, by l.a.9(ii), [QXnj]i'=l

is also complemented in X. .
If the second alternative holds, i.e. if 11- An, nl ;?; 1/2 for all n E Nl we use
1- Q instead of Q. 0
Proposition 3.b.S suggests the following
Problem 3.b.9. Is every Banach space X, with a symmetric or even subsymmetric

basis, primary?
Observe that, by the decomposition method, this problem would have a

positive answer provided we could show that the factor containing a complemented
subspace isomorphic to X, say QX, satisfies QX:::::: QX EB QX. In general, however,
a space with a symmetric basis might have complemented subspaces which are not
isomorphic to their square (this follows from T. Figiel [39] and 3.b.l).
Let us mention another space which is primary; th~space J of James, intro-
duced in l.d.2 (cf. P. G. Casazza [14]).
We conclude this section by discussing some examples of spaces with a sym-
metric basis which have some special properties. T. Figiel and W. B. Johnson [42]
have constructed a Banach space with a symmetric basis which contains no sub-
space isomorphic to Co or to lp, p;?; 1. Their example disproved the feeling that, by
some fixed point argument, it might be possible to come up with a positive answer
to the question whether every space with a symmetric basis contains Co or lp (such
a method works for Orlicz sequence spaces, as shown in 4.a.9 below). The construc-
tion of Figiel and Johnson is based on the procedure described in 3.bA and the
proof of 3.b.2. More precisely, in the notation used there, their example is just
the space Y(c o, h, T, {mn}:'=l), where {mn},':'=l is any increasing sequence satisfying
00
m n ;?; 1 and 2: m;; 1 < 00, and T is the dual of the space of Tsirelson introduced in
n=l
2.e.1. We recall that Tis a reflexive space with an unconditional basis {tn}:'=l' The
proof of 2.e.1. actually shows that T contains no subsymmetric basic sequence.
Instead of discussing here in detail the example of Figiel and Johnson we present
a modified example, due to Z. Altshuler [3], which has some additional interesting
properties.
Example 3.b.tO. A Banach space Y with a symmetric basis {en}:'=l in which all
symmetric basic sequences are equivalent to {en}:'=l but which contains no subspace
isomorphic to Co or lp, p;?; 1.
The additional interest in this example stems from the fact, proved in 3.a.1O,
that a space X with a symmetric basis {Xn}:'=l has to be isomorphic to Co or to lp
for some p;?; 1 if all symmetric basic sequences in X are equivalent to each other
and the same holds in X*. The example 3. b.l 0 shows that it is not enough to assume
that this property holds only in X.
We first define a sequence of symmetric norms on Co as follows. For any integer
n and any a=(a1> a2,''') E Co we put
k
where Sk= L j-l, k= 1,2, ... and the inner supremum ranges over all permutations
j:1
of the integers. It is easily checked that
for all n and all a E Co, where 11·110 denotes the norm in Co.
Let Y be the space of all sequences a E Co for which Ilally=llnt Ilal ntnllT <<X),
where {tn}:':1 is the unconditional basis of the space Tintroduced in 2.e.1. The unit
vectors {ej}i';' 1 belong to Y since, for each i,
Moreover, they form a symmetric basis of Y with symmetric constant equal to 1.

We also note that lim
k-+oo
I ±ell =00. Indeed, let nand k be such that Sk/2 < 22n::;; 2sk.
j:1 Y
Then, IIJ1 et;;"lt~1 el;;"slJ(2n+2-nSk);;"S~/2/3, which proves our assertion since

lim Sk=oo. It follows from 3.b.5 that Y is not isomorphic to Co.
k-+ 00
In order to prove that Y has the desired properties we first remark that every
symmetric basic sequence in Y is equivalent to a (symmetric) block basis of
{e i }f'=,1 (apply the argument used in the proof of 3.b.5). We begin by treating some
particular cases of such block bases.
qm+l
Lemma 3.b.11. Let U m= L 1 Cjei, m = 1, 2, ... be a normalized block basis of the
j:qm+
unit vector basis {ej}t~1 of Y.lflim Ct=O then there exists a subsequence {Umj }j':1 of
t-+ 00
{Um}::1 which is equivalent to a block basis {Xj}f',.1 of{tn}:':1' the natural basis ofT.
Prooj. The relation between the norms 11·lln and 11·110 of Co shows that, for each
fixed m and N, we have
2: Ilumlln:; n:1
N-1 N-1
n:1
2: 2nllumllo::;;2 N max{lct!;qm<i::;;qm+1}.
Therefore, we can construct inductively two increasing sequences of integers,

{mj}f',.1 and {Nj}f'=,1' such that
As easily checked, this implies that the basic sequence {Um)f'=,1 is equivalent to the
Nj-1
block basis Xj= L Ilumjllntmj= 1,2, ... of {tn}:':l' D
n:Nj_l
It follows from 3.b.ll and the fact that T contains no subsymmetric basic
qm+l
sequences that there is no symmetric block basis U m= 2: Ctej, m= 1,2, ... of
t;qm+ 1
{et}f'..1 so that the sequence {Cj}f'..1 tends to zero. This implies that {ej}t';1 is not
equivalent to the unit vector basis of lp,p';:; 1 and thus, by 3.b.S, Y is not iso-
morphic to any lp with p ';:; 1.
We consider next block bases of the unit vector basis of Ywhich are generated
by one vector (defined in 3.a.8).
Proposition 3.b.12. Every block basis {Ui a)}f'..1 of {et}t";'1, which is generated by a
vector a E Y, i~ equivalent to {el}t''';1'
00
Proof We first recall that it suffices to show that, for any fixed O~a= 2:
j;1
aiel E Y,
a series 1 bju)a) converges whenever fJ= ~

j;1 j;1
bjej E Y. We further observe that it is
00
actually enough to prove that

j;1
2: atUla) is a convergent series for every O~a=
00
2:
j;1
atet E Y. Indeed, if this were the case then, for a and fJ as above with at';:; 0 and
00 00
b;';:;O for all i, we would get that 2: (al+bj)uia+/J), and therefore also 2: btu/a), is
j;1 ;;1
a convergent series.
00
Fix a = 2:
j;1
atej E Y with 1 ';:; a1 ';:; a2';:; ... ';:; at ';:; ... ';:; 0 and notice that in order
2: atuia) converges in Y we have to compute the II· lin-norms of

00
to check whether
t;1
the double sequence {a j aj}j':'j;1 (the numbers ajaj, i,j=l, 2, ... are the coefficients
00
in the expansion of 2: atUla) with respect to {ei }t';1)'

j;1
Let aCt) be a non-increasing
function on [1,00) such that a(i) = at for all integers i. If, for some integer m,
i·j=m then at least one of the integers i or j is ,;:;m1/2 and therefore a;aj::;;a(m1/2).
It follows that the non-increasing rearrangement of {a;ajh':'j;l (as a single sequence)
is majorated by the sequence fJ=(b 1, b2, ... ) whose explicit form is
7(1)-times 7(2)-times 7(m)-times
~~, .A. ..
fJ=(a(P/2), a(21/2), a(21/2), ... , a(m1/2), ... , a(m1/2), ... ) ,
where T(m) is the number of distinct divisors of m. Thus, for every n, we have
IIJI alufat ::;;
IlfJlln.
For each integer m let cp(m) be the first place where a(m) appears in the sequence
fJ. Then, for cp(m)::;;k<cp(m+l), we have
(~1 bJ- 1 )/(2n+2- nS k)::;; (J1 rp~%j~1 bji-1)f(2n+2-nsk)

: ; (~1 brp(j)cp(j)-1(cp(j+ 1) -cp(j)) /(2n+ 2 -nsrp(m»
: ; (~1 ajcp(j)-1(cp(j+ 1) -cp(j)) /(2n+ 2 -ns<p(m») .
Since 81P(m)~log cp(m) we get that
11.Blln~s~p (~1 ajcp(j)-1(cp(j+ 1)-cp(j)))/(2n+2- n log cp(m)),

n=I,2, ....
k
2: 7{i)=k log k+(2y-I)k+
To estimate further the norm of.Bwe use the fact that
1=1
O(k1/2), where y=0.57721... is the constant of Biller (this formula is proved in
many books on number theory; e.g. see [80]). It follows that there are constants
C1 and C2 so that, for j~ 1, we have
j2-1
cp(j)= 1 + L T(i)~ 1+C1P logj,
1=1
and cp(j+1)-cp(j)~C2(1+jlogj). Since 8m behaves asymptotically as logm we

get, by substituting the estimates of cp(j) and cp(j+ I)-cp(j) in that of 11.Blln, that
for all n and for some constant C3 <00. This implies that II.BIIY~ C3 110:11y, i.e.
lilt tltu~a)IIY ~ C3110:1Iy for all 0: EY. 0
We pass now to the study of general symmetric block bases of {ei}t"=1' Let
qm+l
U m= 2: Cjeb m= 1,2, ... be a normalized symmetric block basis of {ej}~1 in Y.
j=qm+ 1
We can assume without loss of generality that in each block Um the coefficients
Cj are arranged in non-increasing order and that they are non-negative.
Suppose first that, for every e>O, there exists an integer r=r(e) such that
III=qm+
q;tl Cte;// <e for all m for which qm+1-qm~r. In this case {Um}:=1 is equivalent
r Y
to a block basis generated by one vector and thus, by .3.b.I2, it is equivalent to
qm+l-qm
{ei}~1' Indeed, by putting V m= 2:
j=1
Cj+qmeb m=I, 2, ... and using a standard
diagonal argument we can find a subsequence {V mj }j'=1 of {V m}:=1 such that
at = lim Cj+qmj exists for every i. For e > 0 and r=r(e) we have Ilqm +±-qmj Cj+qmjeill < e
j-+oo
j
j=r Y
for allj which implies that II:± atetll ~ e,'ror every 8 > r. Hence, 0: = ~ tltej E Y and
!=r Y j=1
It follows that a suitable subsequence of {Um j }i=1 is equivalent to a block basis

generated by 0: (recall that each vmj is a "translation" of um).
Finally, suppose that there exist an 8> 0 and an increasing sequence of integers
{mh}h"=l such that qmh+l-qmh>h and II
qmh+l
2: I
Ctet ~8 for all h. Put
l=qmh+h+ l y
qmh+h qmh+l
Vh = 2: Cjej and Wh = 2: Cjej. Then, u",h = Vh + Wh and iiwhiiy ~ 8 for every
l=qmh+1 i= qmh+ h+1
integer h. Notice also that Cqmh+h~C, for some constant c>O and every h, would
imply 1 ~ IIVhiiY~CI11 ell
1=1 y
,h= 1, 2, ... , i.e. c=o. Thus, lim Cqmh+h=O which means
h .... '"
that {Wh}h"=1 is a block of {ej}l"=1 with coefficients tending to zero.
By using 3.b.ll and passing to a subsequence if necessary we can assume that
{Wh}h"=1 is equivalent to a block basis {Xh}h"=1 of {tn}:'=l, the unit vector basis of T.
The definition of the norm in T implies the existence of a constant Al > 0 such that,
for every k, we have IIh=k+1
~ Xhll T ~ Alk. It follows that, for every integer k and
some constant A 2 ,
Since {U mh}h"=1 is a symmetric basic sequence we get, by the discussion following

3.a.7, that {umh}h"=l> and thus also {Um}~=l' are equivalent to the unit vector basis
of h. This would imply that there exists in Ya block basis of {ei}i'=1 with coefficients
tending to zero which is also equivalent to the unit vector basis of h. This however
is impossible in view of 3.b.ll and the discussion thereafter. 0
Remark. It is interesting to compare the construction of Yin 3.b.1O with the

general method of constructing spaces with a symmetric basis which has been
described in the proof of 3.b.2. It can be shown that the space Yof 3.b.l0 coincides
with the space denoted, in the notation of 3.b.4, by Y(co, dew, 1), T, {2n}:'=l),
where w={n-l}:'=l.
4. Orlicz Sequence Spaces
a. Subspaces of Orlicz Sequence Spaces which have a

Symmetric Basis
Most of this chapter is devoted to a quite detailed study of Orlicz sequence spaces,
of particular interest being the relation between these spaces and the Ip spaces.
Later on in this chapter we study the subspaces and the quotients of subspaces
of the direct sum Ip EB IT" It turns out that this topic is closely connected to the study
of OrIicz sequence spaces. In the last section of this chapter we present some results
on another class of spaces with a symmetric basis, namely Lorentz sequence spaces.
The introduction of OrIicz functions has been inspired by the obvious role
played by the functions t P in the definition of the spaces Ip or, more generally
Lp(l'.,). It is quite natural to try to replace t P by a more general function M and then
to consider the set of all sequences of scalars {an};:'=l for which the series ~ M(la n[)
n=l
converges. W. OrIicz [111] has checked the restrictions which have to be imposed
on the function M in order to make this set of sequences into a suitable Banach
space. His study led to the following definition of the so-called OrIicz functions
and Orlicz sequence spaces (for basic material on OrIicz spaces the reader is
referred also to [75]).
Definition 4.a.l. An Orlicz function M is a continuous non-decreasing and convex

function defined for t~O such that M(O)=O and lim M(t)=oo. If M(t)=O for
t-+ 00
some t> 0, M is said to be a degenerate Orlicz function.
To any Orlicz function M we associate the space 1M of all sequences of scalars

00
x=(ab a2,···) such that 2: M([anl!p)<oo

n=l
for some p>O. The space 1M equipped
with the norm
is a Banach space usually called an Orlicz sequence space.

Of particular interest is the subspace hM of 1M consisting of those sequences
00
x= (aI' a2, ... ) E 1M for which 2: M(la n[/ p) <00 for

n=l
every p > O. Some basic proper-
ties of hM are collected in the following proposition.
138 4. Orlicz Sequence Spaces
Proposition 4.a.2. Let M be an Orlicz function. Then hM is a closed subspace of 1M

and the unit vectors {en}:'= 1 form a symmetric basis of hM•
Proof. It is clear that the unit vectors form a symmetric basic sequence in 1M,
Therefo!e, both assertions of the proposition will be proved if we show that hM
coincides with [en l:'=l' An element x=(a1, a2,''') belongs to [en l:'=l if and only if,
for every p>O, there exists an integer N=N(p) such that Iln~N anenll :::;;p, i.e. if and
co
only if L M(la nl/p):::;;1. D
n=N
It is easily verified that if M is a degenerate Orlicz function then IM-;:::,I co and

hM-;:::,co. Since this case is not interesting in the present context we shall assumefrom
now on that all the Orlicz functions considered in the sequel are non-degenerate unless
specified otherwise.
In general, the spaces 1M and hM are distinct. In order to give conditions for 1M
to coincide with hM we need the following definition.
Definition 4.a.3. An Orlicz function M is said to satisfy the L1 2-condition at zero

if lim sup M(2t)/M(t) <00.
t->o
It is easily checked that the L1 2 -condition at 0 implies that, for every positive
number Q, lim sup M(Qt)/M(t) <00 (this condition is sometimes called the
t->o
L1 Q -condition). The importance of the L1 2 -condition is illustrated by the following
result.
Proposition 4.a.4. For an Orlicz function M the following conditions are equivalent.
(i) M satisfies the L1 2-condition at O.
(ii) IM=h M.
(iii) The unit vectors form a boundedly complete symmetric basis of 1M,
(iv) 1M is separable.
(v) 1M contains no subspace isomorphic to lco.
'" M(lanl/p) implies that of

Proof. The fact that the convergence of a series L
n=l
L'" M(la nlll7) follows easily from the L1 Q -condition at zero with Q = pll7. This
n=l
proves the implication (i) (ii). In order to prove that (ii) (iii) we use 4.a.2
lilt aie;":::;;
=? =?
and the fact that s~p 1, for some sequence {at}t'=l> implies that
co '"
L M(lail):::;; 1,
1=1
i.e. (al> a2,"') E IM=hM and thus L aiel
1=1
converges. It is obvious
that (iii) =? (iv) =? (v). Assume now that an Orlicz function M does not satisfy
the L1 2 -condition at zero: Then, we can find a sequence {In}:'=l such that
M(2t n)/M(tn»2n+1 and M(tn):::;;2- n. Let k n be integers chosen so that 2-(n+1)<
00
k n M(l n)::;;2- n for all n. Then L knM{tn)::;; 1 while k nM(2tn) > 1. Thus, for any
n=l
choice of scalars {an};:'=lo we have that
kl times k2 times k n times
,---"-----..,---"-----.. ,---"-----..
2 -1 SUp lanl::;; II(a1tlo ••. , altl , a2t2,"" a2t2,"" antm ... , antn>-' .)11::;; SUp Janl
n n
and, therefore, (v) => (i). 0

The definitions of 1M and of hM show that, up to an isomorphism, what really
matters is the behavior of M in the neighborhood of 1= 0: if two Orlicz functions
MI and M2 coincide on an interval 0::;; t::;; to then IMl and 1M2 consist of the same
sequences and the norms induced by MI and M2 are equivalent. The same is true
for hMl and hM2 . More generally, we have the following result.
Proposition 4.a.S. Let Ml and M2 be two Orlicz functions. Then, the following
assertions are equivalent.
(i) IMl = 1M2 (i.e. both spaces consist of the same sequences) and the identity
mapping is an isomorphism between IMl and 1M2 ,
(ii) The unit vector bases of hMl and hM2 are equivalent.
(iii) MI and M2 are eqUivalent at zero, i.e. there exist constants k > 0, K> 0
and 10>0 such that,for all 0::;; t::;; to, we have
The proof of this proposition is very simple. The implication (ii) => (iii), for
instance, is proved by comparing the norms of {el +"'+en};:'=l in tMl and 1M2 ,
If at least one of the functions satisfies the Ll 2 -condition at zero then the equiva-
lence at zero of Ml and M2 can be expressed in a simpler form: there exist constants
K and to >0 such that K- 1 ::;;MI(t)/M2(t)::;;K for all 0< t::;; to.
There are many instances where an Orlicz function M is defined only in a
neighborhood of zero. In this situation the function M can be extended for t> to
so that it becomes an Orlicz function on the entire positive, line. By 4.a.5 the
corresponding spaces 1M and hM will be the same regardless of the way we have
extended M. The norms associated to two distinct extensions might be different
but always equivalent.
Every Orlicz function M, being non-decreasing and convex, has a right-
t
derivative pet) for every t>O and M(t)=f pes) ds. Since the function p is non-
o
negative and non-decreasing it follows that 1::;; tp(t)/ M(t) for all t> O. In particular,
tllis implies by differentiation that M(t)/t is a non-decreasing function. This fact
can be also obtained directly from the convexity of M since, for 0 < s < t, we have
M(s)::;; (s/t)M(t)+(1-s/t)M(O)=(s/t)M(t).
For every Orlicz function M there exists an Orlicz function Mo which is
i
equivalent to M and has a continuous derivative. We simply put MoCt) = f (M(s)/s)ds
o
t
and get M(t)~Mo(t)~ f (M(s)/s)ds~M(t/2) for all t>O.
t/2
140 4. OrIicz Sequence Spaces
The ratio tp(t)/M(t) is also related to the Ll 2 -condition. More precisely, an

Orlicz function M satisfies the Ll 2-condition at zero if and only if
lim sup tp(t)/M(t) <00.

t->o
Indeed, if M(2t)~ KM(t), for some constant K and for o~ t~ to, then
2t
tp(t)~ ft pes) ds=M(2t)-M(t)~KM(t).
Conversely, if tp(t)/M(t)~Kl' °
< t~ to then
2t
log (M(2t)/M(t)) = ft (p(s)IM(s)) ds~Kl log 2, 0< t~ to/2,
i.e. M(2t)~2KIM(t).
The fact that OrIicz sequence spaces are a natural generalization of lp spaces
might suggest that they have a structure almost as simple as that of lp spaces.
Already the study of subspaces of OrIicz sequence spaces shows that this is not the
case. For instance, there is no simple description of a general subspace with a basis
of an OrIicz sequence space similar to that for lp spaces (cf. 2.d.l). As we shall see
in the next section, the structure of complemented subspaces of 1M is much more
involved. It is however possible to describe quite satisfactorily those subspaces
of an hM space which themselves possess a symmetric basis. To present this ap-
proach we need the following lemma which was proved in [81] and [93] under tlfe
assumption that M satisfies the Ll 2 -condition; the fact that this lemma, as well as
4.a.7 and 4.a.8 below, are valid without the Ll 2 -condition was noticed in [70].
Lemma 4.a.6. Let M be an Orlicz function and consider the following subsets of
C(O, t)
where the closure is taken in the norm topology ojC(O, t). Then EM, I!> EM, CM,A
and CM are non-void norm compact subsets of C(O, t) consisting entirely of Orlicz
functions which might be degenerate.
Proof Let p be the right-derivative of M. Then, for A> 0, we have

iI iI
M(A) = fo pes) ds~ f pes) ds~t,\p(,\f2).
iI/2
Hence, for O~tl' t2~1/2 and any A>O, we get that

which shows that the functions in EM,ro, considered as elements of C(O, t), are
equi-continuous. In addition, these functions are uniformly bounded by 1. Thus,
EM,ro is a norm compact subset of C(O, t) and so are all the other sets defined
above. It is quite clear that every N E CM,co is an Orlicz function on [0, tl (as a
uniform limit of Orlicz functions) though it is even possible that N(t) =0 for every
t E [0, tl. The functions of CM,co will be extended, for convenience, to Orlicz
functions defined on [0,(0). 0
Remark. It is easily checked that the statement of 4.a.6 remains valid if, instead
of [0, H we use any other interval [0, tal with 0< to < 1. If the Ll 2 -condition at
zero does hold for M then, for any 0< A <00, sup (Ap(A)/ M(A)) <00 which
O<i\~A
implies that the set CM,A is compact also when considered as a subset of C(O, 1),
the closure being subsequently taken also in C(O, 1). We shall use this observation
later on in this section.
nj
Let M be an Orlicz function, {e n}:= 1 the unit vectors in 1M and Uj= 2:
i=nj_l + 1
Ciel,
j= 1,2, ... (no=O) any normalized block basis of {en}:=l' To every vector Uj we
nj nj
associate the function Mit) = 2: M(lcilt). Since ;=nj_l

i=nj_l + 1
2: + 1 M(lcd)= 1 it follows
immediately that the functions {Mj}i'=b as elements of C(O, t), belong to the set
CM,l' By 4.a.6 there exists a subsequence {Mjk}k'=l of {Mj}i'=l and an Orlicz
function NE CM,l, which might be degenerate, so that IMjk(t)-N(t)I~2-k,
O~ t~t, k= 1,2, .... Assume that N is not degenerate; then, we get that
00 00
2: Mik(iakl) <00 if and only if 2: N(iakl) <00, i.e. the subsequence {UhJk'=l is
k=l 1<:=1
equivalent to the unit vector basis of hN and [UjJk'=l;:::;h N. Moreover, the map
(ab a2, ... ) -+ (alcl,"" alcn1 , a2cnl+l,"" a2cn2'''') is an isomorphism from IN into
1M, If N(t)=O for some t>O then {Uik}k'=l is equivalent to the unit vector basis of
Co which, in this case, is isomorphic to hN •
We also note that every sub symmetric basic sequence in hM is equivalent to

some normalized block basis of {e n }:=1 (see the proof of 3.b.5 and the remark
thereafter). These facts and 1.a.II prove the following result (cf. [81]).
Proposition 4.a.7. For every Orlicz function M the following assertions are true.
(i) Every infinite-dimensional subspace Y of hM contains a closed subspace Z
which is isomorphic to some Orlicz sequence space hN.
(ii) Let X be a subspace of hM which has a subsymmetric basis {Xn}:=l' Then X
is isomorphic to some Orlicz sequence space hN and {Xn}:=1 is equivalent to
the unit vector basis of hN.
The functions N appearing in (i) and (ii) might be degenerate. The discussion
above shows that the function N appearing in 4.a.7 belongs to the compact convex
set CM,l introduced in 4.a.6. The next result proved in [93, 94] shows that CM,l
actually "coincides" with the collection of all subspaces of hM which have a sub-
symmetric (or a symmetric) basis.
Theorem 4.a.8. Let M be any Orlicz function. An Orlicz sequence space hN' where
N might be a degenerate Orlicz function, is isomorphic to a subspace of hM if and only
if N is equivalent to some function in e M,1'
Proof. We have to prove only the "if" part. Let N E eM, 1 and observe that the
extreme points of eM, 1 are contained in the compact set EM,l' The correspondence
>"-+ M(M)jM(>..) is a continuous map from the interval 10=(0, 1] into EM, 1 and,
therefore, it may be extended uniquely to a map W -+ M ro from fJlo, the Stone-Cech
compactification of 10 , onto E M ,1' By the Krein-Milman theorem there exists a
probability measure I-' on Wo so that
N(t)= f
610
Mro(t)dl-'(w), O~t~t·
For every integer n put An= 1 if 1-'(/0»0 and An=lj2n+1 if 1-'(/0)=0. Then, choose
a sequence of probability measures {l-'n}:=1 on fJlo so that I-'n is supported by the
interval (0, An) and
An
IN(t)- fo (M(M)jM(>..» dl-'n(>")1 < Ij2n+1
for all n and for all t E [0, t]. Fix 0 < T < 1 and, for every nand j, set
,i-1An
CX1,n= f dl-'nC>..)jM(>..). Then,
, 1An
L L [cx1,n]M(~-1Ant)
00 00
[cx;,n]M(~Ant)-lj2n+1~N(t)~
1=1 1=1
for all n and all t E [0, t]. Choosing integers k n so that
L
00
[cxj,n]M(T1-1Anj2)~lj2n+1, n =1,2, ...

1=kn+l
we get that
kn
for all n and all t E [0, t], where Fn(t)= 2:
;=1
[cx;
'
n]M(~-lAnt). We also observe
that, in the case 1-'(/0»0, N(t)~yM('Aot) for some constants 1 >y, >"0>0 and for
every t E [0, tJ. Thus, in this case
Let {?7n}:=l be disjoint subsets of integers and let {?71, n}1=0 be a disjoint splitting
of "In so that "10, n consists only of one element mn and 7Jj, n has [aj, n] elements
forj=l, 2, ... , k n and is void forj>k n. Then, the vectors
form a basic sequence which, by the inequalities (*) and (:), is equivalent to the
unit vector basis of hN ({ei}~l denotes here the unit vector basis of hM). 0
We have presented in 2.e.l the negative solution of Tsirelson to the question

whether every Banach space contains a subspace isomorphic to Co or to Ip for some
1 ~p <00. Nevertheless, there are some important classes of Banach spaces which
are not connected a priori to some lp or to Co but for which this problem has a
positive answer. The next theorem gives a positive answer in the case of Orlicz
sequences spaces (and in view of 4.a.7(i) also for their subspaces) in very precise
terms (cf. [93] and [95]).
Theorem 4.a.9. The space lp, or Co ifp=oo, is isomorphic to a subspace of an Orlicz

sequence space hM if and only if p E [aM' ,sM] where
aM=SUp {q; sup M(M)jM(A)tq<OO} and

O<'\,t.,l
,sM=inf{q; inf M(M)jM(A)t q> Otf .

O<A,t~l
Proof It is easily checked that we always have l~aM~,sM~oo and ,sM<OO if and
only if M satisfies the Ll 2 -condition at zero. It follows from 4.aA that the Ll 2 -
condition holds for M if and only if the unit vectors of 1M form a boundedly com-
plete basis. Hence, by 1.e.lO, ,sM=OO if and only if Co is isomorphic to a subspace
of h M. From now on we consider only finite values of p. For p 1= [aM' ,sM] it is
easily seen that the function t P is not equivalent to any function in CM ' l and there-
fore, by 4.a.8, lp is not isomorphic to a subspace of hM • If aM=,sM<oo the Ll 2 -con-
dition holds for M and using the remark following 4.a.6 we can consider .CM , l as a
convex compact subset of C(O, 1). Let 0 < 'T < 1 and consider the map T, on CM , l
defined by T,N(t)=N( 'Tt)jN(T). Since CM,l c C(O, 1) T, is well-defined, continuous
and it maps CM,A into CM,A<for every 0< A~ 1. Hence, by the Schauder-Tychonoff
fixed point theorem [33, V.lO.5] T. has a fixed point N. E CM which, by definition,
satisfies N.('Tt)=N.('T)N.(t),O~t~1. Putting q,=logN,('T)jlog'T, we get that
N.('Tn) = 'Tnq, for all n. Let t E (0, 1) and choose an integer n such that 'Tn < t~ 'Tn-I.
Then,
IN,(t) - tq'l ~ IN.(t) - N.('Tn) I+ l'Tnq,- tq'l ~ 2('T(n-l)q, - 'Tnq,)
~2q.(1-'T), O~t~l.
By letting 'T ---+ 1 and using the compactness of CM we get that t q E CM for some
q;;:, 1 (observe that sup q. <00 because of the Ll 2 -condition). As remarked above,
,
this q must be equal to aM=,sM'
In the case aM <13M we choose p E (aM' 13M)' By the definition of aM and 13M there
are numbers 0 < Un < Vn < Wn ~ 1 with Wn -?> 0, un/vn -?> 0 such that ncp(un) < cp(vn/2),
ncp(wn) < cp(vn/2), where cp(t) =M(t)jtP. Put
1
Mn(t)=A;l f M(tsWn)s-P-l ds,
un/wn
1
where An= f M(swn)S-P-l ds. Clearly, Mn E CM,w n for all n. Let an = un/wn and
unl w n
t
bn=vn/wn; by substituting y=ts we get that Mn(t)=A;lt P f M(ywn)y-P-l dy.
ant
t 1 an 1
Since f = f + f - f it follows that Mn(t) = t P+fn(t)-,gn(t), where
ant an ant t
an
fn(t)=A;lt P f M(ywn)y-P-l dy~A;la;PM(un)
ant
and
1
gnCt)=A;ltP f M(ywn)y-P-l dy~A;lM(wn)'
t
On the other hand, since bn/an= Vn/Un -?>CX), we get, for n sufficiently large, that
f
bn
An?;< M(swn)S-P-l ds?;<2- 1 M(v n/2)b;P .
b n /2
It follows that
i.e. Mn(t) -?> t P uniformly on [0, t]. Hence, t PE CM and this concludes the proof
in view of 4.a.8 and the fact that CMis closed. D
Let us make a few comments on 4.a.9. The proof above shows that, for any
finite value of p in [aM' 13M], we have t P E CM; hence, as easily checked in the proof
of 4.a.8, we get that hM actually contains almost isometric copies of lp. In the case
p=oo the same is true for Co (use 2.e.3).
For the actual computation of the interval [aM' 13M] it is of some interest to
point out (cf. [95 p. 374]) that aM = SUp aN and f3M=inf bN if 13M <00, where the
supremum, respectively the infimum, is taken over all Orlicz functions N which are
equivalent to M at zero, and
aN = lim inf tN'(t)/N(t), bN = lim sup tN'(t)/N(t) .

t--+o t--+ 0
Theorems 4.a.8 and 4.a.9 give a complete description of those Orlicz sequence
spaces and, in particular, of those lp spaces which embed isomorphically into a
given Orlicz sequence space. There are, however, some interesting questions of a
more specific nature on mappings between Orlicz sequence spaces. We conclude

this section by presenting some results, due to N. J. Kalton [67], which give
necessary and sufficient conditions for the identity mapping between two Orlicz
sequence spaces to be an isomorphism on some infinite dimensional subspace.
Theorem 4.a.l0. Let M and £1 be two Orlicz functions satisfying the iJ 2 -condition
°
at zero so that the formal identity mapping T: 1M -7 1M is bounded (i.e. M(t)?:AM(t)
for some constant A> and for every t E [0, 1D. Then T is strictly singular if and
only if, for every constant B<oo, there exists afinite sequence {'Tj};"=l c(O, 1] so that
m m
2: Mht)?: B 2: Mht)
j=l 1=1
for every t E [0, 1].
Proof We shall prove here only the necessity since this is the part which will be
used in Section 2.c below. Fix B<oo and suppose that the formal identity mapping
T from 1M into hi is a strictly singular operator. By using the iJ 2 -condition and the
fact that M(t)?:AM(t) for all t E [0, 1], it is easily checked that the set
D=D(M, £1, B) = conv {(M(M)-BM(M))/M(}o..); O<}o..~ I}
is a norm compact subset of C(O, 1).

We shall show now that the set D contains no non-positive function. Assume
the contrary, i.e. that there isfE D withf(t)~O for all t E [0, 1]. Let W -7 Mro and
W -7 £1ro denote the continuous mappings from [310' the Stone-Cech compactification
of the intervallo = (0, 1], onto EM, 1, respectively EM, 1> such that Mi',ct) = M(M)/M(}o..)
and MA(t)=M(M)/M(}o..) for }o.. E 10, Let Rro be the unique extension to [310 of the
(bounded) function M(}o..)/M(A), A E 10 , Then, by the definition of D, there is a
probability measure fL on f3Io so thatf(t) = f (Mro(t)-BRroMro(t)) dfL(W), O~ t~ 1.
/310
It follows that
A S Mro(t)Rro dfL(W)~ SMro(t) dfL(W)~B S Mro(t)Rro dfL(w), O~ t~ 1 .

/310 /310 /310
PutN(t)= f Mro(t)dfL(W),N(t)= f Mro(t)RrodfL(w) and notice that NE CM , l and

/3~ B~
N(t)/N(l) E C,W, l' Hence, by 4.a.8, the unit vectors in IN (respectively in hi) form
a basis equivalent to a block basis {U n}:'=l (respectively {Un}:'=l) of the unit vectors
in 1M (respectively in 1M)' Reviewing the proof of 4.a.8 it is easy to check that the
blocks Un, n = 1, 2, ... can be constructed as to actually coincide with Un> n = 1, 2, ....
Indeed, if we keep the notation of 4.a.8 and approximate fL by the same sequence
of measures {fLn}:'=l on (0, An), for both Nand N, the forms of Un and Un depend
only on the numbers
tJ-1A n '(;1 -IAn
ctj,n= S dfLn(}o..)/M(}o..), respectivelyaj,n= S (RA/M(A)) dILn(}o..) .

,fAn ,fAn
But CXj,n=a1,n for allj and n which proves that u~iin' Hence, TI[Unl~=l is an iso-
morphism and this contradicts our assumption.
Since the compact set D is disjoint from {f;f(t):s;;O}, by the geometric form
of the Hahn-Banach theorem and by the Riesz representation theorem for func-
tionals in C(O, 1), we get that there exists a probability measure v on [0, 1] such
that inf {tf(t) dv(t);fE D }>o. This measure v can be approximated by a convex
combination, with rational coefficients, of point-mass measures. Thus, there
k
L: niM(Atj)-
exist integers {n1H=1 and reals (t1}~=l' with 0< tj:S;; 1 for allj, so that
1=1
BS1(Atj»>O, O<A:s;;1. To complete the proof of the necessity we just take as
h}r'=1 the points {tjH=h each t1 being repeated nj times. 0
The condition appearing in 4.a.1O takes a particularly simple form if M(t)= tV.
Corollary 4.a.1l [67]. Let M be an Orlicz function satisfying the .tJ 2 -condition at
zero and such that M(t)~AtV for some A >0, 1:S;;p<oo and for every O:s;;t:s;; 1. Then
the identity mapping from 1M into Iv is a strictly singular operator if and only if
· 'nf . f
I1ml 1 f1 M(st) d
8-+0
In 1-1/e 8 S vv+1
0<8';;1 og t t=oo.
Proof. Again, as in 4.a.l0, we shall prove only the necessity so we assume that the
identity mapping from 1M into Iv is strictly singular. By 4.a.IO, for every B>O,
m m
there are {TI}r'=1C::(0, 1] so that L: M(TISU)~BsVuV L: Tf for every s, UE [0,1].
1=1 1=1
Put T= min Ti' Then, for 0<e<T2 , we have
1E;IE;m
M(st) M(st)
L Tf L Tf f
m tf m 1
= f .mtV+1 dt:S;; Vt v +1 dt .
I=l tflt 3" 1=1 8 S
Thus, for 0< e < T2 and for every s E (0, 1], we get that
_1_ f1 M(st) dt~Blog T/e~~

log l/e 8 sVt v+1 log I/e r 2
which proves our assertion since B is arbitrary. 0
Using 4.a.ll it is easy to check that the identity operator from IMp' with
Mit) = tV(I + Ilog tl), into Iv is strictly singular in spite of the fact that IMp contains
complemented subspaces isomorphic to Iv (cf. 4.c.I below).
h. Duality and Complemented Subspaces
Let M be a non-degenerate Orlicz function whose right-derivative p satisfies

p(O)=O and limp(t)=<Xl. These restrictions exclude only the case when M(t) is
t-+ 00
equivalent to t, i.e. 1M ;:::; 11' Consider the right-inverse q of p which is defined by

q(u)=sup {t;p(t)~u}, u:;?:O. It is easily verified that q is a right-continuous non-
decreasing function such that q(O)=O and q(u»O whenever u>O. Put M*(u)=
u
f q(v) dv for u:;?:O. Then, M* is also a non-degenerate Orlicz function and q is its
o
right-derivative. The function M*, defined in this way, is called the function com-
plementary to M. It is clear that M is the function complementary to M*, i.e.
M**=M.
A quick glance at the graph of p shows that, for any t and u:;?: 0, we have the
so-called Young inequality, namely
tu~M(t)+M*(u) ,
with equality holding if u=p(t) (or t=q(u)). In other words, for any u:;?: 0, we have
uq(u) = M(q(u)) + M*(u) ,
i.e. M* satisfies
M*(u) = max {tu-M(t); 0< t<<Xl}.
With the aid of the complementary function M* we can introduce a new norm
on 1M by putting,
fot x=(ar, a2,''') E hr. The norm III·IIIM satisfies
for every x E 1M , The right-hand side inequality follows directly from Young's
inequality. In order to prove the left-hand side inequality let x=(a1 , a 2, ... ) E 1M be
such that IllxlllM = 1 and take bn = p(lanD. Then lanlb n = M(lanl) + M*(b n) for all n.
If it M*(bi) > 1 it would follow by the convexity of M* that M*(bn/i~l M*(b i)) ~
M*(bn)/Jl M*(bi) and thus, n~l M*(bn/j~lM*(bj))~1. Using the fact that
IllxlllM = 1 we would get that
co
.L M*(bi):;?:.L lanlbn=.L M(lanD+ .L M*(bn)
00 00 00
i=1 n=l n=l n=l

00 00
and this is a contradiction. Thus, .L

j~l
M*(bi )::;:; 1 and therefore 1;;.: .L
n~l
lanlb n;;':
~ M(ianj), i.e. IlxllM::;:; 1.

n~l
The complementary function can be used to describe the dual space of an

Orlicz sequence space.
Proposition 4.b.l. Let M and M* be complementary Orlicz functions. Then

ht.';::;:,IM*, ltr';::;:,ht.t and if, in addition, M* satisfies the Ll 2 -condition at zero then
ht.*';::;:,IM'
Prooj. Let x* E ht. and put Cn =x*(en), n= 1,2, .... Then,
II(cI, c2,· .. )IIM'=sup t~l ancn; n~l M{ian/)::;:; I}

=sup {x* C~l anen); II(al, a2,· .. )IIM::;:; 1}= IIx*1I .
As easily checked, this implies that the map x* ----'7- (x*(el)' x*(e2)"") defines an
isometry from ht. onto 1M>, endowed with III· 111M>, and therefore an isomorphism
from ht. onto 1M>, The other two assertions folIow immediately from the existence
of this isomorphism and 4.aA. 0
Combining 4.b.1 with 4.aA we get the folIowing criterion for reflexivity of
OrIicz sequence spaces.
Proposition 4.b.2. Let M and M* be complementary Orlicz function. Then hM (or

1M) is reflexive if and only if both M and M* satisfy the Ll 2 -condition at zero.
We have already remarked that M satisfies the condition Ll2 at zero if and only
if bM=lim sup tp(t)/M(t) <00. If we assume, in addition, that both p and q
t-+o
are continuous functions then M* satisfies the Ll 2 -condition if and only if
1 < aM = lim inf tp(t)/M(t). More precisely, we have aM=b M./(bM*-l). Indeed,
t-+o
ifuq(u)/M*(u)::;:;bM*+e for some e>O and O<u::;:;uo then
Because of the continuity of q we further get tp(t)/M(t);;.:(b M*+e)/(b M *+e-1) in

some neighborhood of t = O. This shows that aM;;': bM./Cb M• -1). The opposite
inequality is proved in a similar manner.
Sometimes the following terminology is used: an OrIicz function M is said to
satisfy the Llt-condition at zero if lim inf tp(t)/M(t) > 1. We can thus restate 4.b.2
t-+o
in the following form.
Proposition 4.b.2', Let M be an Orlicz function with a continuous strictly-increasing

derivative. Then 1M the (or h M) is reflexive if and only if 1 < aM~ b M<00 (i.e. both the
Ll z- and the LI~-conditions hold for M).
Observe that M' is strictly increasing if and only if q, the derivative of M*,
is continuous.
For reflexive Orlicz sequence spaces the duality between subspaces and quotient
spaces is reflected by the following result.
Theorem 4.h.3. Let 1M be a reflexive Orlicz sequence space and let M* be the function
complementary to M. Then the following assertions hold.
(i) An Orlicz sequence space IN is isomorphic to a quotient space of 1M if and
only if N* is equivalent to a function in eM',l'
(ii) An Orlicz function N is equivalent at zero to a function in EM,l if and only
if N* is equivalent to a function in EM-, I'
(iii) a"i+,sM;=l, aM;+,sM I =1.
(iv) 1M contains a subspace isomorphic to lp for some p:;:; 1 if and only if 1M has
a quotient space isomorphic to lp.
Before proving the theorem let us mention that Examples 4.c.l and 4.c.2, to be
presented in the sequel, show that, in general, property (ii) is not shared by the sets
eM, I and eM-,1 or by eM and eM>' In other words, condition (iv) above is not
necessarily valid when the space Ip is replaced by a general Orlicz sequence space.
Proof of 4.b.3. The assertion (i) is just a restatement of 4.a.8. To prove (ii) we
notice that for any 0 < '\, u
(M(,\t)fM(A»*(u)=max {tu-M(,\t)fM(A); 0< t<oo}

=M*(A-IM(,\)u)fM(A)
= (M*(fLU)fM*(fL»'(M*(p,)fM('\» ,
where fL=A-IM(A). It is easily checked that the reflexivity of 1M and 4.b.2' imply
that fL -+ 0, while the ratio M*(fL)fM(A) remains bounded and bounded away from
zero, as 1.-+0. This completes the proof of (ii), and therefore also that of (iii),
which is an immediate consequence of (ii). Finally, (iv) follows from (iii) and
4.a.9. 0
A simple application of 4.a.9 and 4.b.3 is the following generalization of 2.c.3.

Let M and N be two Orliczfunctions satisfying the LIz-condition at zero. Then, every
bounded linear operator from 1M into IN is compact if and only if aM > ,sN' The "if"
part of this assertion is proved by using arguments similar to those used in 2.c.3.
In order to prove the "only if" part assume that aM~,sN' Let TI be a quotient map
from 1M onto laM' let I be the formal identity map from laM into IPN and let T z be
an isomorphism from IPN into IN' Then, the operator T= Tz1TI is a non-compact
operator from 1M into IN'
We turn now to the study of complemented subspaces of an Orlicz sequence
space. The results proved so far yield immediately a necessary condition as well
as a sufficient condition on an Orlicz function N, for IN to be isomorphic to a
complemented subspace of 1M, Assume that 1M is reflexive; then, by 4.b.3, a necessary
condition is
(*) N is equivalent to a function in CM , l and N* is equivalent to a function in
CM -,l'
In order to derive a sufficient condition observe that in an Orlicz space 1M
normalized block bases of the unit vector basis with constant coefficients corres-
pond (via the general correspondence between blocks and functions in CM, h
which was described in 4.a.8) to functions in EM, l' Hence, by 3.aA, a sufficient
condition is
(~) N is equivalent to a function in E M , l '
We shall present in the next section examples which show that (*) is not a
sufficient condition and that (~) is not a necessary condition (in both cases 1M will
be reflexive and N(t) equivalent to t P for some p).
There is, however, a weaker version of (~) which is already a necessary condition
for IN to be isomorphic to a complemented subspace of 1M , To explain the definition
below we first write down the negation of (~) in an explicit manner: N is not equiva-
lent to any function in EM, 1 if and only if
(t) For every K?:-1 there exist mK points tl E (0, 1/2) such that, for every
,\ E (0, 1), there is at least one index i, 1 ~ i ~ mK for which
Definition 4.b.4. Let M be an Orlicz function. A function N is said to be strongly

non-equivalent to EM, 1 if (t) holds with the additional requirement that mK can be
chosen so that mK=o(Ka) as K--+oo, for every a>O.
Theorem 4.h.5 [94]. Let 1M be a separable Orlicz sequence space and Nan Orlicz
function which is strongly non-equivalent to EM. l ' Then IN is not isomorphic to a
complemented subspace of 1M ,
Proof. Suppose that IN is isomorphic to a complemented subspace of 1M and let

{e n};'=l denote the unit vector basis of 1M , By l.a.12 and l.a.9(ii) there exists a
normalized block basis Wj= L aleioj= 1,2, ... of {en};'=l such that {WAi=l is
ie aj
equivalent to the unit vector basis of IN and there exists a projection P from 1M
°
onto [Wj]i'= l ' By passing to a subsequence and changing the signs of the coefficients,
if necessary, we may assume that at> for all i E aj and that the functions
NtCt)= L M(alt) satisfy lNit)-N(t)I~2-j,j=I,2, ... for some Orlicz function
iE "1
N E CM , l equivalent to N and for every t E [0,1]. We also observe that, since M
satisfies the L1 2 -condition at zero, there exists a p <00 so that M(st)~sP M(t) for
s> 1 and every t>O.
Assume now that N is strongly non-equivalent to EM, l ' Then, there are a num-
ber K and mK points th E (0, 1), h= 1,2, ... , mK so that
and, for every AE (0, I), there exists at least one h, 1 ~h~mK for which
M(Ath)/M(A)N(th) 1= [K-l, K]. By passing to a subsequence of {Nj}i=l, if necessary,
we may assume without loss of generality that INit) - N(t) I~ min {N(th); 1 ~ h~ mK}
for all t E [0, 1].
We split now each of the sets aj into 2mK disjoint subsets of integers 8J ami
7]J so that, for every l~h~mK' we have M(ai t h)/M(ai )N(th)<K-1 if iE 8J and
M(aith)/M(ai)N(th»Kfor iE7]J. Then, forj~1 and l~h~mK' we have
KN(th) 2 M(ai)< 2 M(aith)~Nith)~2N(th)

iE1/~ iEn7
mK
which implies that L M(ai)~2/K. Thus, L L M(ai)~2mK/Kfor j~ 1.
iEn~ h=1 jEn~
Every function FE CM ,l satisfies F(st)/F(t)~sP for all s> 1 and t>O. In par-
mK
ticular, if we put Fit) = L L M(ait) then Fit)/F;Cl) E C M ,l and therefore
h=1 iEn~
The condition imposed on the ratio mK/K implies that Fi21IPI[)~ I, i.e. the vectors
inK
Vj= L L aie!,j~ 1, have norms ~ 1/21IPII.

h=l iEn~
Put, for 1 ~h~mK andj= 1,2, ... , uJ= L aiei and let Qj be the norm one pro-
iEt5~
inK
jection from 1M onto [et1!E".' Then, Wj= QjPWj= L QjPuJ+ QjPVj, j= 1,2, ...
1 h=1
which implies that
mK
.2 II QjPuJl1 ~ IIWjll-IIQjPVjll ~ I-IIPII·IIVjll~!·
h=1
Hence, for every j~l, there exists at least one index l~hj~mK such that
00
II QjPuJ111 ~ 1/2mK' It follows that if we put PUJ1=.L

.=1
di,jwj then
for all j~ 1. By 1.c.8 the linear operator D: [UJ1]f'=1 ~ [Wj]f'=b defined by

DuJ1 = dj.jwi>j~ I (i.e. the "diagonal" of P), is bounded and II DII ~ IIPII. Con-
sequently, for any set of coefficients {h j } ; = b we have
J _ J
Choosing an integer J so that 2~ L N(thj) ~ 3 we get that I ~ L Nlth1 ) ~ 4. Thus,
j=1 j=1
in view of the correspondence between the functions {Nj}!=l and the blocks
{Wj}f=l, we have
On the other hand,
Therefore, by the fact that for all FE CM,l and t>O we conclude
lIit thPJjll~3l/PK-l/P.
F(st)/F(t)~sP,
that This implies that 1/2mKIIPII~3l/PK-l/P, which contra-

dicts the choice of K and mK' 0
The results on subspaces of Orlicz sequence spaces (and especially the method
used to prove 4.a.9) put in evidence the mappings T,,: CM,l -?>- CM ,1 defined by
(T"N)(t)=N(At)/N(A). The pair (C M, 1, {T"n forms what is called a flow in topo-
logical dynamics. Some standard notions and reasonings from topological dynamics
yield interesting facts on Orlicz sequence spaces if we consider this particular
flow. For instance, the notion of a minimal set from topological dynamics has
some applications in our context. The notion of a minimal OrIicz function M will
be defined only for those M which satisfy the Ll 2 -condition. Therefore, using the
remark following 4.a.6, we can and shall consider the sets EM, 1, CM ,1> etc. as
compact subsets of C(O, 1) (rather than subsets of C(O, 1-)). In this way the
mapping T" is well-defined for every 0< A< 1.
Definition 4.h.6. An Orlicz function M satisfying the Ll 2 -condition at zero is called

minimal if the set EM, 1 has no proper closed subsets which are invariant under the
flow (T,,; O<A< 1). In other words, if EN,l=EM,l for every NEEM,l'
Let M be any OrIicz function satisfying the Ll 2 -condition at zero. A standard

application of Zorn's lemma to the set E M ,1> endowed with the order F-<G-¢>
FE E a, 1, shows that EM, 1 contains at least one minimal Orlicz function.
Minimal Orlicz sequence spaces have the following property.
Proposition 4.h.7 [94]. Let M be a minimal OrUcz function. Then every block basis
with constant coefficients of the unit vector basis of 1M spans a subspace which is
isomorphic to 1M itself.
Proof It follows from 3.a.5 that if U is a subspace of 1M which is spanned by a

block basis with constant coefficients then lM~/M EB U. On the other hand, by
4.a.7 and its proof, U~/N EB V for some Orlicz function N E EM, 1 and some
Banach space V. Since Mis minimal we have M E EN,l and thus, by (~), U~/MEB W
for some space W. Consequently,
It is worthwhile to compare 4.b.7 with the second reinark following 2.a.9: for
a minimal Orlicz function M which is not equivalent to any t'P the isomorphism
between U and 1M cannot be always induced by mapping the n'th block to the n'th
unit vector.
We shall present later on in this chapter some examples of minimal Orlicz
functions which are not equivalent to any t'P. Actually, we will see that for any
interval [a,,8] there is a minimal Orlicz function M with aM=u and ,8M=,8; the
only restriction being that a> 1 (it is easily verified that if M is minimal and
aM=l then M(t)=ct).
The proposition above states that for minimal Orlicz sequence spaces the
"obvious" complemented subspaces of 1M (i.e. those spanned by block bases with
constant coefficients) are necessarily isomorphic to 1M , It is possible that this is
also true for any other complemented subspaces. We thus formulate.
Problem 4.b.S. Assume that M is a minimal Orliczjunction. Is then 1M a prime Banach

space?
It would be actually of interest to decide whether 1M is prime even for a single

example of a minimal function M other than t'P.
In Section 3.b we have seen that the universal space U1 of Pelczynski has un-
countably many mutually non-equivalent symmetric bases. Among the Orlicz
sequence spaces there are also many examples of spaces having at least two non-
equivalent symmetric bases. The construction of such spaces is based on the
following remark: if 1M is isomorphic to IN but M is not equivalent to N then 1M has
at least two non-equivalent symmetric bases, namely the unit vector bases of 1M and
of IN'
This observation can be used in the case of minimal Orlicz functions in order
to prove the next result.
Theorem 4.b.9 [95]. Let M be a minimal Orlicz function which is not equivalent to
any t'P, 1~p <00. Then 1M has uncountably many mutually non-equivalent symmetric
bases.
Proof It follows from the definition of minimality, condition <=) and the decom-
position method of Pelczynski that, for every N E EM,!> the space IN is isomorphic
to 1M, Therefore, in view of the preceding remark, it suffices to show that EM, 1
contains uncountably many mutually non-equivalent functions.
Assume that there are only countably many equivalence classes in EM,l and
denote by {Mn};'=l their representatives. Then, by Baire's category theorem, one
of the sets
whose union covers entirely the set EM,!> contains a relatively open set G. By
minimality there exists, for every N E EM, 1, a A E (0, 1) such that N(>.t)/N(A) E G
154 4. Orlicz Sequence Spaces.
and thus, E M ,1 consists of exactly one equivalence class. In order to complete

the proof it suffices to show that all the functions in EM, 1 are uniformly equivalent
to M. For 0< A< 1 we put GlI. ={N E EM • 1 ; N(M)/N(A) E G}. The sets GlI. are open
and, as remarked above, E M ,1=U {GlI.; 0< A:;::; I}. Hence, by the compactness
of E M,1, there is a Ao > 0 such that EM, 1 = U {GlI.; Ao:;::; A:;::; I}. It follows that, for
every O<fL:;::;l, there is a A with Ao:;::;A:;::;1 such that M(AfLt)/M(AfL)EG, i.e.
k- 1 :;::;M(AfLt)/M(AfL)M(t):;::;k (if we assume, as we may, that GCFn,k and Mn = M).
Since A~ Ao the Ll 2 -condition implies the existence of a constant A >0 so that, for
all O<fL, t:;::;l,
A- 1 :;::;M(fLt )/M(fL)M(t):;::;A.
This implies that M is equivalent to t P for some l:;::;p <00, contrary to the assump-
tion. 0
The concept of a minimal Orlicz sequence space can be extended in a natural

manner to the more general setting of spaces with a symmetric basis. A symmetric
basis {Xn}:'=1 of a Banach space X is said to be minimal symmetric provided every
block basis of {X n }:'=1 with constant coefficients spans a subspace which is iso-
morphic to the whole space X.
In view of 4.b.9 it is natural to ask whether the only spaces with a minimal
symmetric basis which is unique, up to equivalence, are Co and lp, l:;::;p <00.
There are many interesting examples of Orlicz sequence spaces which do have,
up to equivalence, a unique symmetric basis. A sufficient condition for this to
happen is given in the next proposition (cf. [93]).
Proposition 4.h.l0. Let M be an Orlicz Junction Jor which the set eM contains no
OrliczJunction equivalent to M itself. Then the unit vector basis is, up to equivalence,
the unique symmetric basis oj h M.
Proof. If hM has in addition to the unit vector basis {en}:'=1 another symmetric
basis {fn}:'=1 then each ofthese two bases is equivalent to a block oasis of the other.
It is easily seen that if, in both block basis representations, the coefficients do not
tend to zero then {fn}:'=1 is equivalent to {en}:'=1' If this is not the case then {e n}:'=1
is equivalent to a block basis of itself with coefficients tending to zero. In view of
4.a.8 and its proof this implies that eM contains a function which is equivalent to
M itself. 0
A simple consequences of 4.b.1O is that Orlicz sequence spaces hM Jor which

lim tM'(t)/M(t) exists (i.e. aM=b M) have, up to equivalence, a unique symmetric
t .... o
basis. This follows from the fact that, in this case, eM consists only of one Orlicz
function, namely t P , provided p=lim tM'(t)/M(t) is finite or J(t) =0 in [0,1/2]
t .... o
if p=oo.
Besides minimal Orlicz fJlnctions it is of interest to consider some maximal ones.
More precisely, we shall construct Orlicz functions Uc , cl which are universal for
the class of all Orlicz functions M with [aM. f3M]C(C, d). We need first the following
lemma.
Lemma 4.h.ll. Let F and G be two continuous non-decreasing convex functions,

defined on an interval [T, 1] with 0 < T < 1. Assume that
(i) F(I)=G(1)=I, O<F(T)<I, O<G(T)<l and,

(ii) for some numbers c and d such that 1 <c<d, F'(I)=G'(1)=c and c~
tF'(t)jF(t)~d, c~tG'(t)jG(t)~d for every tE [T, 1] (here F' and G' stand
for the right-derivatives except for t= 1 where they mean the left-derivatives).
Then the function
{ F(t),
H(t)= F(T)G(tjT),
is continuous, convex, non-decreasing and H'(1)=c~ tH'(t)/H(t)~d for every

tE [T2, 1].
The proof is straightforward.
Theorem 4.h.12. For every l~c<d<oo there exists an Orlicz function

U = Uc. d such that
(i) c~tU'(t)/U(t)~dforalltE[O,1],
(ii) for every Orliczfunction M with c~tM'(t)IM(t)~dfor all tE(O, 1] there
exists in Eu a function equivalent to M,
(iii) there exists a constant Kc. d such that, for every M with c~ tM'(t)/M(t)~d,
there is a norm-one projection PM in lu such that d(PMl u, IM)~Kc. rj.
Proof. Assume first that c> 1 and choose a sequence {Nn(t)};:'=1 of Orlicz functions
which is dense in the set ff of all Orlicz functions N satisfying N(1)= 1, N'(1)=c
and c~ tN'(t)/N(t)~d for all t E (0, 1]. Put Tn =2- 2n -\ n= 1,2, ... and define
In view of 4.b.ll, U is an Orlicz function defined on the entire interval [0,1] and
such that c~ tU'(t)jU(t)~d for all 0< t~ 1, Moreover, for every n and every
Tn~ t~ 1, we get that U(Tnt)jU(Tn)=Nn(t) which implies that Eu contains all the
functions of ff.
Let now M be any Orlicz function such that M(1)= 1 and c~ tM'(t)jM(t)~d
for all O<t~1. Choose t 1=tC • d so that t1c(d-l)j(c-l)d=lj2 and t2 so that
c(M'(t1)(t2- t1) + M(t1» = M'(t 1 )t2. It is easily verified that
Define
0~t~t1
t1 <t~t2
t2<t~1 .
Then the function M 1(t) = Mo(t)/Mo(1) E~ and K-1M1(t)~M(t)~KM1(t) for
all IE [0, 1] and for some constant K=Kc. d, independent of M. This concludes
the proof in the case c> 1.
The case c= 1 can be reduced to c> 1 in the following manner. Fix d> 1
and let U = U2 • d + 1 be a universal function corresponding to the numbers 2
and d+ 1 (whose existence follows from the previous case). Put Uo(t) = U(t)/t.
In general, Uo need not be convex so, instead, we consider the function
t
U1 (/) = J(Uo(s)/s) tis. It is easily checked that U1 is an Orlicz function satisfying
o
I ~ tUf(t)/Ul(t)~d and Ul(1)~ Uo(t)~dUl(t) for all 0< t~ 1. We shall show that
U1 = U1 • d has all the desired properties. Let N be any Orlicz function such that
N(1)=1 "and l~tN'(t)/N(t)~d for all O<t~1. Then, M(t)=tN(t) satisfies
M(l)= 1 and 2~tM'(t)/M(t)~d+ 1 for all O<t~ 1; hence, by the first part of the
proof, Eu contains a function M1 such that K-1Ml(t)~M(t)~KMl(t), O~t~ 1,
with K being a constant independent of M. This means that, for some sequence
An -+ 0, U(Ant)/U(An) -+ M 1 (t) for all 0 < t~ 1. By passing to a subsequence of
{A n}:'=l, if needed, we can assume with no loss of generality that, for some Nl E E U1 ,
U1 (A nt)/U1 (An) -+ N 1 (t) uniformly for O~ t~ 1. It is easily verified that
(dK)-lN(t)~N1(t)~dKN(t) for all O~ t~ 1. D
Remarks. It follows immediately from (~) and Pelczynski's decomposition method

that luc.a is determined uniquely, up to an isomorphism, by c and d. If we choose c
and d so that c -1 + d -1 = 1 then U:, d is also universal for the same c and d. Con-
sequently, IUc.a is isomorphic to ltc• a' This constitutes a non-trivial example, i.e.
different from 12 , of a Banach space with a symmetric basis which is isomorphic to
its own conjugate. The fact that all the universal spaces corresponding to the same
c and d are isomorphic to each other shows that IUc.a is another example of a space
with infinitely many mutually non-equivalent symmetric bases. Indeed, it suffices
to observe that, by proper rearrangements of the sequence {Nn}:'=l in the con-
struction of Uc•d , we obtain uncountably many universal functions {U~~~}a such
that, for ex ¥: [3, U~~~ is not equivalent at zero to U~%.
c. Examples of Orlicz Sequence Spaces

The difference between the structure of Orlicz sequence spaces and that of Ip
spaces is best illustrated by considering suitable examples. In the beginning of
this section we present some examples of Orlicz functions which are given by concrete
formulas. In addition to th()se examples we shall describe a general method of
constructing Orlicz functions by using sequences of zeros and ones.
c. Examples of Orlicz Sequence Spaces 157
Example 4.c.l. A reflexive Orlicz sequence space 1M with a unique symmetric basis
such that the sets C M <,1 and CZ, 1 ={N*; N E C M ,1} have a totally different structure
(recall that, up to equivalence at zero, the set E M ',1 coincides with EZ,1 in view of
4.b.3.(ii).
Let M(t)=t P Ilog tl" with 1 <p<oo and a>O. It is easily checked that Mis
an Orlicz function on some interval [0, to] with to> 0. Thus, M can be extended to
°
an Orlicz function on [0,00) but for the present discussion the values of M out-
side a neighborhood of t = are of no importance. A trivial computation shows that
lim tM'(t)/M(t)=p. Hence, by 4.b.2', 1M is reflexive and, by 4.b.l0, the unit
t->o
vectors are, up to equivalence, the unique symmetric basis of 1M , We also have
lim M(..\t)/M(A) = lim t P(1 +log tflog A)" =t P,

A->O A->O
° < t~ to.
Therefore, the sets EM and CM consist both of only one function, namely t P • It
follows that the set E M ,1 has exactly two equivalence classes: t P and functions
equivalent to M itself. Also CM ,1 consists of these two equivalent classes. Indeed,
to
let N(t) = J (M(..\t)/M(A» df.L(A), for some probability measure f.L on [0, to] (where
o
M(Ot)/M(O) stands for t P). Observe that, for a fixed
increasing function of A. Hence, for every 0< t1 < to,
° < t~ to, M(..\t)/M(A) is an
and, unless f.L is concentrated in the origin, N is equivalent to M. We conclude

that, up to an isomorphism, the only Orlicz sequence subspaces of 1M are Ip and
1M itself, and both are also complemented subspaces of 1M ,
In order to study the quotient spaces of 1M we have to compute the comple-
mentary function M*. The exact computation of M* is quite complicated but, for
our purposes, it suffices to find a function which is equivalent to M*. Observe first
t
that M is equivalent at zero to M 1(t) = J f(s) ds, where f(S)=SP-1 Ilog sl". Take q
o
so thatp-1+ q -1= 1 and put g(t)=t q- 1 Ilog tl,,(1- q). Since
f(g(t»=gP-1(t) Ilog g(t)I" = t Ilog tl- al(q-l) log t+a(1-q) log Ilog til"
we get that lim f(g(t»/t=(q-l)". This implies that g is equivalent at zero to the
t->o
t
inverse function of f and thus, M* is equivalent to J g(s) ds which, in turn, is
o
equivalent to the function t q Ilog tla(l-q). By abuse of notation we shall put
M*(t) = t q Ilog tl,,(1-q).
It is easily verified that EM> and CM > consist only of one function, namely t q •
However, the set CM>,l turns out to contain infinitely many equivalence classes:
for example, for every O<e<a(q-l), C M *,1 contains a function equivalent to
e- 1
t q Ilog tl-e. Indeed, the function Ne(t)= J (M*(M)JM*(A))A- 1 [log A[-I-e dA
o
clearly satisfies Ne(t)JNe(I) E CM., 1 and
log A ),,(q-1l
N(t)=t q
e
S
e- 1
0
(
log A+log t
A-1[logA[-1-edA
u )a(q-l) 1 e
=t q S
co (
1 u+ [log tl
u- - du
=tq[logt[-e --),,(q-l l v-
S (V
co
1- e dv
Ilog II - 1 1+v
for some constant C>O. It follows that Ne is equivalent to t q [log t[-e.

It is clear from the discussion in Section b above that, for any N E CM ',1 such
that N is not equivalent to a function in E M >,1 (i.e. to M* itself or to t q), IN is
isomorphic to a subspace of 1M , (or, equivalently, IN> is isomorphic to a quotient
space of 1M ) but not to a complemented subspace of 1M " 0
We consider next an example introduced in [81] and investigated in [81] and

[94].
Example 4.c.2. A reflexive Orlicz sequence space 1M with a unique symmetric basis
for which ctM<f3M and C M> is different from C~={N*; N E CM} (recall that, in
4.c.l, CM.=C~={tq}).
Let M(t)=tP+s!n(logIIOgt[); it is easily checked that, for p-V2>1, M is an

Orlicz function in some neighborhood of t=O. Put
U(A)=AM'(A)JM(A)=p+sin (log [log A[)+COS (log [log A[).
Then, for any 1> 0>0, lim (U(AS)- U(A))=O uniformly for s E [0,1]. Suppose
iI-->O
now that lim U(An)=r for some r E [aM' bM] and some sequence An -+ O. By using
n--> co
the uniformity of the limit above and by passing to a subsequence, if needed, we
can assume that
Integrating this inequality between t and 1 we get that
i.e. M(Ant)JM(An) tends to t T • It follows that E M={t';p-V2::(n,:;p+V2} while

C M consists of all the functions N which can be represented as
p+';"2
N(t)= S t T df1-(r)
p-';"2
c. Examples of OrIicz Sequence Spaces 159
for some probability measure p. on [p-V2,p+V2]. By taking, for instance, p.

to be uniformly distributed on the interval [r, p+ V2] we get a function equivalent
to tT/llog tl. A simple computation shows that if N E CM is represented by a measure
p. and r is the smallest number in the support of p. then lim N(>..t)/N(A)=tT. Con-
" .... 0
sequently, CM contains no function equivalent to M itself and, by 4.b.lO, 1M has,
up to equivalence, a unique symmetric basis.
In order to prove that the sets C;; and CM' are different we shall show that
every Orlicz function, which is simultaneously equivalent to a function in CM
and to a function in C;;" is already equivalent to a function in EM, i.e. to 1'0 for
somep-V2~ro~p+V·2. Indeed, let NE CM and let ro be the smallest number in
the support of the measure p. representing N. Then, as remarked above, EN={tTa}
and
p+../2
N(t)/t Ta = f tT-Ta dp.(r)~p.([ro, r])+t T- Ta for ro<r.,;;p+V2.
TO
If N is not equivalent to fTa then p.({ro}) =0 and hence, lim N(t)/tTa=O. It follows
t .... o
that lim N*(t)Wa=oo, where qol+rol= 1.
t ... o
On the other hand, if N* is equivalent at zero to a function in
S2
N*(t) = ft S dll(s)
Sl
for some probability measure II on [Sl' S2]' Since EN. = {tqa} the smallest number in
the support of II must be qo. This however is a contradiction since it implies that
lim sup N*(t)/t qa .,;; 1. 0
t ... o
The following example, constructed by N. J. Kalton [67], shows that Theorem

4.b.5 is no longer valid if we replace "strong non-equivalence" by "non-equiva-
lence".
Example 4.c.3. A separable Orlicz sequence space 1M which contains complemented

subspaces isomorphic to lp for some 1.,;;p <00 but t P is not equivalent to any function
in EM,l'
We first define a sequence of functions {fn}:=l on [0,(0) in the following way:

for each integer n, fn is the function of period P n = 2 22n such that
It is easily seen that f" is a continuous function on [0,(0) whose maximal value is
equal to 2". Let f(t) = max {f,,(t); n= 1, 2, ... } and observe that If(t1)-f(t2)t~
tlt1-t21 for any tlo t 2>0. Fixp>3/2 and put M(t)=tPef(-IOgt), for O<t~l, and
M(O) =0. The function M is continuous but not necessarily convex on [0,1].
However, it is easily checked that, with M' standing for the right-derivative of M,
we have
p-t~tM'(t)/M(t)=p-f'( -log t)~p+t
for every t E (0, 1]. By the condition imposed on p we get that M(t)/t is an in-
creasing function and thus, M is equivalent at zero to an Orli¢z function Mlo
t
defined on [0,(0) (take, e.g. M1(t)=f(M(u)/u)du). Moreover, M1 satisfies the
o
Ll 2 -condition at zero since tM'(t)/M(t)~p+t.
It is evident from the definition of M that the formal identity mapping T from
1M! into Ip is a bounded operator. We shall show that T is not strictly singular. This
would ill}.ply that IMl contains a complemented subspace isomorphic to Ip. Indeed,
if W is an infinite-dimensional subspace of IMl for which T1 = T,w is an isomorphism
then, by 2.a.2, TW contains a subspace V~ Ip so that there exists a bounded pro-
jection P from Ip onto V. It is easily checked that Q = TIlPT is a bounded projection
from 1M! onto its subspace TIl V, which is clearly isomorphic to Ip.
To prove that Tis not a strictly singular operator we use 4.a.l1. Put e" = e- Pn ,
n= 1,2, .... Then,
1 1 M(t) 1 1 e'(-logt) 1 Pn
I 1/ f p+1 dt=p f dt=p- f ef(u) du .
og e" Bn t " Bn t "0
Notice that, for t E [0, P,,].!(t) = max (h(t); 1 ~ i~n}. Hence, by using the definition
of};, we get that
~4 L 2-22!.2t·e2i<00,
<Xl
i=l
for every integer n, and this proves that T is strictly singular.

It remains to show that t P is not equivalent to any function in EM!. l' Assume
to the contrary that there exists a constant K> 1 such that, for any T>O, there is
a U=U(T) for which e-KtP~M(e-Ut)/M(e-U)~eKtP, e-$~t~ lor, equivalently
If(u+v)-f(u)I~K, O~V~T. Thus, for U~tlo t2~U+T, we have
For each t>O let n(t) be the least integer for whichf,,(t)=f(t). Choose T>3·2 64K2
and observe that if 0 <f(t) <2n (t)-2Kfor some u+T/3~ t~U+2T/3 then there exists
a tl so thatJ(t)+ 3K> fn(t/t l ) > J(t)+2K and It1 - tl =2(fn(t)(tl )-J(t» < 6K < T/3. It
follows that u<tl <U+T and this contradicts the fact thatJ(tl )-f(t»2K. Thus,
for every u+T/3~t~u+2T/3, either J(t)=O or J(t)~2n(t)-2K. Since T/3~P2
there exists atE [u+ T/3, U+ 2T/3] so that n(t) > 1. We also notice that net) is not a
constant on [u+ T/3, U+2T/3] since this would imply that the variation ofJ on this
interval exceeds T/12>2K. Hence, if t l , t E [u+T/3, U+2T/3] are chosen so that
n(t)-n(tl)~1 then J(t)#O (otherwise n(t)=I) and therefore J(t)~2n(t)-2K~
2· 2n(t 1 ) - 2K~ 2J(t1) - 2K, i.e. J(t 1) ~ 2K + J(t) - J(t 1) ~ 4K. Consequently,
It follows that, for every t E [u+ T/3, U+2T/3], we haveJ(t) = max {ftCt); 1 ~ i~no},
where no is the largest integer satisfying 2no ~ SK. Hence, on this interval, J has
period Pno = 222no ~ 264K2 < T/3 which means that J takes both values 0 and 2n o there.
This however implies that the variation ofJon [u+T/3, U+2T/3] is equal to 2no ~4K,
which leads to a contradiction. D
We present now a general procedure of constructing (or representing) Orlicz

functions M in a form in which the set EM, 1 can be easily described (cf. [95]).
Fix 0 < T < 1 and let F and G be two strictly increasing continuous convex
functions on the interval [T, 1] such that
(i) F(l)=G(1)= 1, O<F(T)< 1, O<G(T)< 1.

(ii) F'(l)=G'(I), F'(1)~tF'(t)/F(t) and G'(l)~tG'(t)/G(t) for all tE hi].
(iii) F(T)=TPI and G(T)=T P2 for some 1 <PI <P2'
For every sequence of digits 7] = {l'](n)}:= 1 with 7](n) equal to 0 or to 1 for each n
we define a function Mn on [0, 1] in the following way. We put Mn(l) = 1, MnCO) =0
and, for Tn~t<Tn-\ n=l, 2, ...
Using 4. b.11 it is easily verified that M n is indeed an Orlicz function on [0, 1]

which satisfies the Lf 2 -condition at zero.
The function Mn has the following quite obvious properties.
k
(a) Mn{T k ) = TPlk+(P2-Pl)n~1 n(nl, k= 1,2, ... which implies that, up to equivalence,
Mn is solely determined by Pl, P2 and 7] and does not depend on the particular
choice of F and G.
(b) For the same T, P1 and P2 and for two different sequences of digits
7]={7](n)}:=l and p={p(n)}:=l, the functions Mn and Mp are equivalent if and
only if
162 4. OrIicz Sequence Spaces
(c) For fixed T, PI and P2, the set of all the functions of the form Mm with 7J
being a sequence of zeros and ones, is a norm compact subset of C(O, 1) and the
map 7J --? M7I is a homeomorphism from {O,I}lto, equipped with the product
topology, into C(O, 1).
(d) Consider the map T defined by (TN)(t)=N(Tt)/N(T) and let ([J be the shift
by one to the left, i.e. «([J7J)(n)=7J(n+l). Then TM7I=M~71'
(e) Up to equivalence, EM~, 1 consists offunctions ofthe form Mp (for the same
T, P1 and P2) with p being a pointwise limit of sequences having the form {([Jki7J}i'= l'
The following proposition describes the interval [aM~' .BM~] corresponding to
an Orlicz function of the type M7/ (see 4.a.9).
Proposition 4.c.4. Let T, Ph P2, 7J and M 7/ be as above. Then,
where 8(n+ 1, n+k) denotes the density of ones between the numbers n+ 1 and n+k,
n+k
i.e. 8(n+ 1, n+k)= L
i=n+1
7J(i)/k.
Proof. Since M7I satisfies the Ll 2 -condition at zero it suffices to consider in the
definition of aM~ only expressions of the form
Hence, aM. is the supremum of all numbers P for which
A simple argument shows that aM. is therefore equal to the expression given in the
statement. The proof for .BMn is similar. 0
We present now a characterization of minimal Orlicz function of the form M7I'
Proposition 4.c.S. Let M7I be as above. The function M7I is equivalent to a minimal
Orlicz function if and only if there exists a constant K such that, for every integer k,
there is an integer n=n(k) with the following property: for every integer h there is
an m~n such that
Proof. Assume first that such K and n(k) do exist. Let {hl}j'~l be an increasing
sequence of integers for which the pointwise limit of {rph l7)}Z"=l exists. Denote this
limit sequence by p and fix k. Then there exists an index hI so that
p(i)=(rph l 7))(i)=7)(h l +i) for l::;;;i::;;;k+n(k).
By our assumption, there is an l::;;;m::;;;n(k) so that
Hence, by using the observations (e) and (b) above, it follows that EM., 1 contains a
function equivalent to Mn. By the remark preceding 4.b.7 it follows that Mn is
equivalent to a minimal Orlicz function.
In order to prove the converse we suppose that there exists a minimal function
N such that A-1::;;;N(t)/MnCt)::;;;A for some constant A>O and every O<t::;;;1.
Assume now that the condition described in the statement is not satisfied for any
K and n(k). This )1leans that, for every integer K, there exist an integer k(K) and
sequences {h1(n, K)}:'=l> {h 2(n, K)}:'=l for which hin, K)-hICn, K)=n, n= 1,2, ...
and such that, for every h1(n, K)::;;;h<hin, K), there is a j::;;;k(K) with
For a fixed Klet7)K be any limit point of the sequence {rph 1 (n,K)7)}:'= l' Then MnKEEMn , 1
and, for every integer m, there is aj::;;;k(K) so that
It follows from the observations (a), (b) and (e) that, for any No E EMr,K' there
exists O<t::;;;1 for which the ratio No(t)/Mn(t) is not in the interval [B-I~(P2-Pl),
BT- K(P2- Pl)], where B is a constant depending only on F and G. This however
contradicts the minimality of N when K is chosen so that BT- K (P2 -P1 l > A4. D
The previous propositions will now be used to study some concrete minimal
Orlicz sequence spaces.
Example 4.c.6 [94, 95]. A minimal Orlicz sequence space 1M whose interval [aM' fiM]
reduces to a single point p and which does not have any complemented subspace
isomorphic to lp.
°
Let < T < 1 and 1 <PI <P2 be arbitrary. We construct simultaneously two
sequences of zeros and ones, 7) ={7)(i)}~1 and p= {p(i)}t'= 1, as follows. Put 7)(1)=0,
p(1)=l and, for n=O, 1,2, ... ,
7)(2 3n + i) = p(i) and p(23n + i) = p(i)

7)(23n+I+i)=7)(i) and p(23n+I+i)=p(i) for l::;;;i::;;;2 3n +I ,
7)(23n+2+i)=7)(i) and p(2 3n +2+i)=7)(i) for 1 ::;;;i::;;;23n +2 •
Thus, these two sequences begin as follows
'"",-.'-, ~ , ,
1)=(0,1,0,1,0,1,0,1,1,1, 1, 1,0,1,0,1, ... ),
,-"-.~~, .A \.
p=(1, 1, 1, 1,0,1,0, 1, 1, 1, 1, 1,0,1,0,1, ... ).
We shall prove first that Mn (and also Mp) is equivalent to a minimal Orlicz
function. For every n let An (respectively Bn) be the block consisting of the first
23n digits in 1) (respectively p). By the inductive definition of 1) and p both An+1
and Bn + 1 contain a block equal to An and a block equal to Bn. Since p can be written
as a succession of blocks C1 C 2 C 3 ••• , where each block Cj is equal either to An+1
or to B n+1, it follows that every block of p of length ~3·23(n+1) contains in it
either An + 1 or Bn + 1 and thus, both An and Bn. This implies that the condition
4.c.5 holds for Mn and also for M p, with K=O and n(k)=3·23(n+1) whenever
k~23n. This proves that both Mn and Mp are equivalent to minimal functions.
The functions Mn and Mp are not equivalent at zero because
3n
.L
2
(p(i)-1)(i))=2n for all n.
1=1
On the other hand, 1) is the pointwise limit of the sequence {W 23n + 2 p}:=1 and p that
of {W 23n7]}:=1' Hence, Mn E EMp , 1 and Mp E EM", 1 which, by condition (:) of the
preceding section, implies that IMp';::;:, 1M", We therefore conclude that neither Mn
nor Mp is equivalent to t P for some p ~ 1. We note in passing that, by 4.b.9, 1M "
has uncountably many mutually non-equivalent symmetric bases.
In order to determine the interval associated to Mn let us denote by an (respec-
tively bn) the number of times the digit 1 appears in the block An (respectively Bn).
Then, ao = 0, bo = 1 and
The density of ones in the block An (respectively Bn) is equal to an/23n (respectively
n-1
bn/23n). Easy computations show that (b n-an)/23n = 1/4n and an/2 3n = 2: 1/2.41 for
i=O
all n. It follows immediately that lim an/2 3n = lim bn/2 3n =2/3. Let 8>0 be given
n--..oo n-..oo
and choose an integer n so that Ja n/2 3n -2/3J<e and Jb n/2 3n -2/3J<8. It is easily
checked that the density of ones in any block of p of length 23n ·1 is between
(/- 2)(2/3 -e)/I and «(1-2)(2/3+8)+2)/1. Therefore, for I large enough, the density
is between 2/3 - 2e and 2/3 + 2e. By 4.c.3 it follows that aMp = f3Mp= PI + 2(P2 - PI)/3.
Since IMp is isomorphic to 1M" we get that Mn has the same interval.
It remains to show that, for p = P1 + 2(P2 - PI)/3, IMp contains no complemented
subspace isomorphic to Ip. In order to prove this we use 4.b.5 and show that the
function t P is strongly non-equivalent to E Mp , l ' For an integer n put
m(n)=3·23(n+1)+2 3n and assume the existence of an integer k and of a constant
K>O so that
Let 1 ~j~ 3· 23(n + 1); by using the above inequality with i = j and i = j + 23n we get
Recall now that every block of p of length .~ 3 . 23 (n + 1) contains a block equal to

An and a block equal to Bn. This implies that there are integers j1,j2, 1 ~j1>j2 <
3·23(n+1) for which
This means that for T-(V2-Vl)2n-l~K<T-(V2-Vl)2n-2 it suffices to take

mK =3·23(n+1)+2 3n points in order to prove that condition (t) of 4.b.4 holds.
Since, for any a >0, m K =o(KO:) we get that tV is strongly non-equivalent to E Mp . l . 0
The next example is similar to 4.c.6 with the exception that its interval can be
chosen arbitrarily.
Example 4.c.7 [95]. A minimal Orlicz sequence space 1M with aM<f3M which con-
tains no complemented subspace isomorphic to Ip for p ~ 1.
Let °
< T < 1 and 1 < P1 < P2 be arbitrarily chosen. For every 0 < a < 13 < 1 con-
(X)
struct a sequence of positive integers {nj}i=1 such that 2

j=1
Ifnj~a and
n
co
(I-Ifn j) ~ 13. We define two sequences p and 'Y) of zeros and ones, as follows.
j=1
Let mj=n1·n2 ... nj_l(ml = 1) and let Aj (respectively B j) denote the block of the
first mj digits of p (respectively 'Y)). Al consists of the digit 1 and B1 of the digit 0
while, for j> 1, Aj and Bj are defined inductively by
The same argument as in 4.c.6 shows that, for this p, the Orlicz function Mp is
minimal (the condition in the statement of 4.c.5 is satisfied with K=O and n(mj_I)=
3mj). To estimate the size of the interval of Mp we remark that the density of ones
j-1
in Aj is larger than n
i=1
(I-Ifni) ~ 13 while the density of ones in B j is less than
j-l
.2 Ifni ~ a. It follows from 4.c.4 that aMp ~Pl + a(p2 - PI) and f3Mp~ PI + f3(P2 - PI),
,=1
i.e. f3Mp-aMp =(f3-a)(p2 - P1) > 0.
In spite of the fact that IMp has subspaces isomorphic to Ip for an entire interval
of p's it does not admit any Iv as a complemented subspace. Since the proof is
similar to that presented in 4.c.6 we do not reproduce it here. 0
We conclude the study offunctions of the form Mn by proving that they actually
represent all reflexive Orlicz sequence spaces.
Proposition 4.c.S. For every Orlicz Junction M such that 1M is reflexive there exist
I<Pl<P2' O<T<1 and a sequence 1)={7](n)}:=l oJ zeros and ones such that the
corresponding Junction M n is equivalent to M.
Proof. We may assume that M(1)= 1. Since 1M is reflexive it follows from 4.b.2'
that, for some I <Pl <P2 and all 0< t~ I,Pl ~ tM'(t)/M(t):(P2' Choose 0< T< 1 so
thatplT-Pl + 1 = TP2. Then, the functions F(t) = tPl and G(t)=P1t-Pl + 1 satisfy in
the interval h I] all the assumptions appearing in the definition of functions of
the type Mn.
Now, we construct inductively a sequence 1)={1)(n)}:=l as follows. We set
1)(1) = 1 and if M n(Tn)TPl ~ M( Tn + 1) we put 1)(n + 1) = 0; otherwise, 1)(n + 1) = 1. It
is easily verified that
Mn(Tn):(M(Tn)~ TPl -P2 MiT n) for all n;::: 1,
i.e. Mn is equivalent to M. 0
Remark. The construction above does not work for a non-reflexive space 1M
satisfying the L1 2 -condition because, in this case, Pl is necessarily equal to 1 and
thus, 4.b.ll cannot be used. However, the same construction can be still performed
with F(t)=t, G(t)=t P2 and O<T<1 arbitrarily chosen. The function Mit) ob-
tained in this way is equivalent to M(t) and Mn(t)/t in an increasing function but,
in general, Mn(t) need not be convex.
d. Modular Sequence Spaces and Subspaces of ip EB ir
In the study of Orlicz sequence spaces we have already encountered spaces which
are generated by a sequence of Orlicz functions. More precisely, we have seen that,
for any block basis {Un}:=l of the unit vector basis of an Orlicz sequence space 1M,
00
there exists a sequence of Orlicz functions {Mn}:=l such that a series 2:

n=l
anun
00
converges if and only if 2:

n=l
Milanl/p)<oo for some p>O. It is therefore natural
to consider the following class of sequence spaces.
Definition 4.d.l. Let {Mn}:=l be a sequence of Orlicz functions. The space liMn}'
00
is the Banach space of all sequences x=(aba2,.,,) with 2: Mn(\anl/p)<oo for

n=l
some p> 0, equipped with the norm
The space liMn} is called a modular sequence space.

An important subspace of liMn} is h{Mn} which consists of those sequences
00
x=(ab a2,"') E liMn} such that 2:

n=l
Mn(\anl/p) <00 Jor every p>O.
d. Modular Sequence Spaces and Subspaces of Ip EEl Ir 167
The notion of equivalence between sequences of Orlicz functions is defined in

the following way: two sequences oj Orlicz Junctions {Mn}:'=l and {Nn}:'=l are said
to be equivalent ifl{Mn} and lINn} are equal as sets, i.e. they consist oj the same sequences.
It is easily checked, using the closed graph theorem, that if {Mn}:'=l and {Nn}:'=l
are equivalent then the identity map from lIMn} onto l{Nn}· is an isomorphism.
Analytic conditions for the equivalence of two sequences of Orlicz functions are
in general quite unnatural. A sufficient eondition, used often in the sequel, is the
following: {Mn}:'=l and {Nn}:'=l are equivalent provided there exist numbers K>O,
til ~ 0, n = I, 2, ... and an integer no so that
The proof is straightforward.

In order to avoid technical difficulties which arise in some non-interesting cases
we assume, unless stated otherwise, that the functions {Mn}:'=l, {Nn}:'=l etc., are
nondegenerate, strictly increasing and have a derivative for every t ~ 0. A sequence
ofOrlicz functions {Mn}:'=l is said to be normalized if Mil)= 1 for all n. If {Mn}:'=l
is not normalized we put NnCt)=MiTnt), where Mn(Tn)=1. Then, {Nn}:'=l is
normalized and lINn} is isomorphic to l{Mn}' Since we consider here only linear
topological properties of modular sequence spaces we can and shall assume that
{Mn}:'=l is a normalized sequence.
In the study of Orlicz sequence spaces a crucial role is played by the Ll 2 -
condition at zero. What we need in the present case is a "uniform" Ll 2 -condition.
Definition 4.d.2. A sequence of Orlicz functions {Mn}:'=l is said to satisfy the
uniform Ll 2 -condition (LI:-condition) at zero if there exist a number p>1 (r>l)
and an integer no such that, for all tE(O, 1) and n~no, we have tM~(t)/Mit)~p
CtM~(t)/MnCt) > r).
An equivalent definition of the uniform Ll 2 -condition at zero is obtained by

requiring the existence of a constant K <00 and of an integer no such that
Mn(2t)/Mn(t)~K for all n~no and t E (0, 1/2J. The uniform Ll 2 -condition is not
preserved by equivalence. For example, let {Mn}:'=l be any sequence of Orlicz
functions for which the uniform Ll 2 -condition at zero holds. Choose real numbers
00
tn>O, n= 1,2, ... so that L

n=l
Mn(tn) <00; then, we can easily construct 'Orlicz
functions N n , n= 1,2, ... which do not satisfy the Ll 2 -condition (even individually)
but such that NnCt)=MnCt) for all n and t~tn' As observed above, {Mn}:'=l and
{Nn}:'=l are equivalent.
The importance of the uniform Ll 2 -condition is illustrated by the following
result.
Proposition 4.d.3 [144]. For any sequence oJOrliczJunctions {Mn}:'=l theJollowing
conditions are equivalent.
(i) The sequence {Mn}:'=l is equivalent to a sequence {Nn}:'=l which satisfies the
uniform Ll 2 -condition at zero.
(ii) I{Mn}=h{Mn}'
(iii) The unit vectors form a boundedly complete normalized unconditional basis
ofl{Mn}'
(iv) l{Mn} is separable.
(v) I{Mn} contains no subspace isomorphic to I",.
Notice that the unit vectors are normalized for it is assumed that MnCl)= 1 for
all n. The proof of 4.d.3 is quite similar to that of 4.aA and we do not reproduce
it here.
We pass now to another routine matter, namely that of duality of modular
spaces. Let {Mn}:'=l be a sequence of Orlicz functions such that none of which is
equivalent to t. Let {M,n:'=l be the corresponding sequence of complementary
functions. In general, the functions M; do not satisfy M;(1) = 1 for all n but we
still can consider the space I{M~}' For every y=(al> a2,"') E I{M~} we put
IIIYIII=~upt~l anbn; n~l M(lbnD~I}'

It follows from Young's inequality that II yll ~ III ylll ~ 211 yll for every y E l{~}.
Let x* E htMn} and put cn=x*(en), n= 1,2, ... , where {en}:'=l denotes, as usual,
the sequence of t1ie unit vectors. Then, we have
i.e. the mapping x* -+ (x*(el), x*(e2)"") defines an isomorphism from htMn} into
I{M~}' It is easily verified that this isomorphism is onto, i.e. htMn}~ l{~}.
We present now without proof a result which generalizes 4.b.2' to the case of
modular sequence spaces.
Proposition 4.d.4 [144]. A modular sequence space I{Mn} is reflexive if and only
if {Mn}:'=l is equivalent to a sequence of Orlicz functions {Nn}:'=l for which the
uniform .d 2 - and .d~-conditions hold.
In the sequel we shall be interested only in modular sequence spaces satisfying

the uniform .d 2 -condition at zero. These spaces are related to Orlicz sequence
spaces in a very simple way.
Theorem 4.d.S. Let 1 ~ q < s < 00. A Banach space is a modular sequence space I{Mnl>
with Mn satisfying q~ tM~(t)/Mn(t)~s for all nand t E (0, 1), if and only if it is
isomorphic to the closed linear span of a block basis of the unit vectors in some Orlicz
sequence space 1M, with M satisfying q~tM'(t)/M(t)~sfor every t>O.
Proof, The "if" part is immediate (use, e.g. the argument preceding 4.a.7). Con-
versely, let {Mn}:'=l be a sequence of Orlicz functions satisfying our hypotheses for
some 1 ~q<s<oo. Let U= Uq ,. be the universal Orlicz function whose existence is
ensured by 4. b.13. The function Uhas the property that there are a constant K = K q , s
and functions N n E Eu, n= 1,2, ... such that K-l~Nn(t)/Mn(t)~K for all nand
d. Modular Sequence Spaces and Subspaces of lp EB lr 169
t E (0, 1). The arguments already used in the proof of 4.a.8 show that there is a
block basis {Un};;"=l (with constant coefficients) of the unit vector basis of lu such
that [un];;"=l~I{Nn}~I{Mn~' We remark that [Un];;"=l is actually a complemented sub-
space of lu. 0
We turn now our attention to a special class of modular sequence spaces. Fix
l::S;r<p<co and let {fn};;"=l and {gn};;"=l be the unit vector bases of lp, respectively
Ir. For a sequence w={Wn};;"=l of positive reals put en=fn+wngmn=I,2, ... and
let Xp,r,w be the closed linear span of {en};;"=l in (lp EB IT)"" In many cases, the space
Xp, r, w is isomorphic to either Ip, Ir or to Ip EB Ir. A non-trivial case which is of im-
portance occurs when the sequence w={Wn};;"=l satisfies the condition
(*) ~
00
"wPr/(p-r)=oo
n=l
n W -*
,n °and Wn < 1 for all n •
The spaces Xp , 2, w were introduced by H. P. Rosenthal [127]. These spaces play

an important role in the study of complemented subspaces of LiO, 1) and they will
be further investigated in Vol.· II. Some of the results, to be presented in this section,
have been originally proved by probabilistic methods in the case when r=2 (cf.
[128]). The probabilistic methods do not work for other values of r and therefore
we present here an approach based on modular and Orlicz sequence spaces. This
method is due to J. T. Woo [145].
Before describing these results we prove that X p , T, W does not really depend on
the sequence W provided condition (*) holds.
Proposition 4.d.6 [127, 145]. Let w={Wn};;"=l and w' ={W~};;"=l be two sequences both
satisfying the condition (*). Then the spaces X p, r, wand Xp, r, w' are isomorphic.
Proof Since W satisfies (*) there are disjoint finite subsets h}i= 1 of integers so that
the numbers vj =( 2: w~r/(p-r)(p-r)/pr satisfy w;::s;vj::S;2wj for allj.
nEaj
Let {en};;"=l be the natural basis of Xp,r,w and put
Then, for every choice of scalars {aj}i= 1, we have

This shows that {hj}f.= 1 is equivalent to the natural basis of Xp, r. w'. We shall prove
now that [hj]f.=l is a complemented subspace of Xp,r,w' For any sequence of
scalars {a n}:=l which are eventually equal to zero we set
We have
IIp( i: anen) II=max {[.i: 12: anw;t'T-r)/(p-T)IP/ (2: w~r/(p-r))P-l] liP,

n:::=l 1=1 neGj nEaj
[.i: 12: anW;t'r - r)/(v - r)IT/ (2: w~rl(p - r)) -1] 1/r}
1=1 nEGj neC1'j
r
and, using twice Holder's inequality, we obtain
Hence, the operator P extends to a norm-one projection from X p, r, w onto [hj]f.= l '
We remark that the proof that [hj]i=l is complemented in Xp,r,w is valid for any
choice of {u}}i= 1 (i.e. it does not depend on the particular relation between Vj and
wj which was taken into account when the sets {Uj}f.=l were constructed). The
existence of the projection P shows that Xv,r,w' is isomorphic to a complemented
subspace of Xp,r,w and, by symmetry, also that Xp,r,w is isomorphic to a comple-
mented subspace of Xp,r,w" Thus, by Pelczynski's decomposition method (see the
proof of 2.d.l 0), it would follow that Xv. r, w~ X p, r, w' provided we prove that each
of the spaces X p , r, wand X p , r, w' is isomorphic to its own square. This means that
in order to complete the proof it suffices to prove the following lemma.
Lemma 4.d.7. For each sequence w={Wn}:=l satisfying the condition (*) we have
Xp,r,w EJj Xp,r,w~Xp,r,W'
Proof We first split the integers N into disjoint infinite subsets N 1 , N 2 , .. ·, N k, ...
so that, for each k, 2: w~r/(p-T)=oo. Put w(k)={Wn}nENk' By the first part of the
neNk
proof of 4.d.6 there exists a constant A <00 such that, for each k, there is a block
basis {ujkl}}ENk of the natural basis of Xp,r,w(kJ, considered as a subspace of Xp,r,w,
which is equivalent to the natural basis of Xp,r,w, d([uJk)]jENk,Xp,r,w)~ A, and whose
span [ujk)LENk is the range of a norm-one projection in Xv, r, wCk ).
Using the remark made in the proof of 4.d.6 we get that the space
Y=[ujkl]}ENk,kEN is also complemented in Xp,r,w, i.e. Xp,r,w= YEJjZ for some
00
Banach space Z. Every y E Y can be represented uniquely as y= 2: y<k) with

k=l
d. Modular Sequence Spaces and Subspaces of Ip EB Ir 171
y(k)= L C?)ujk) E [ujk)]jENk for all k. Moreover, for some constant C depending
jeNk
only on A, we have for every y E Y that
This implies that Y~{y= ~ y<k); y<k)=O for k> I} EB {y= ~ y<k); y(l)=O}, i.e.
k=l k=l
Y~Xp,r,w EB Y. It follows that Xp,r,w EB Xp,r,w~Xp,r,w EB YEBZ~ YEBZ=
X p , r, wand this completes the proof. 0
Until now we have considered only the spaces Xp,r,w with 1 ~r<p<oo. If, in
the definition of Xp;r,w, we replace the space /p by Co we obtain a subspace of
(co EB Ir) 'M denoted by X"" r, w' The condition (*) is replaced in this case by
(*)'" 2'"
n=l
w~ =00, Wn -7- 0 and Wn < 1 for all n .
By the same method as in 4.d.6 it can be shown that, for any two sequences
w={Wn}:'=l and w~={W~}:'=l which both satisfy the condition (*)"" the spaces
X "',r,w and X "',r,w' are isomorphic. The blocks {hj}i'=l used in the proof are in this
case blocks with constant coefficients, The proof that P is a norm-one projection
shows in this case that every block basis with constant coefficients of any permuta-
tion of the natural basis of X""r,w spans a complemented subspace.
In view of 4.d.6 and the preceding remark it is justified to use the notation
Xp,n l~r<p~oo instead of Xp,r,w without reference to the particular sequence W
(satisfying the condition (*), respectively (*)",).
We shall show now that, for any fixed l~r<p~oo and w={Wn}:'=l satisfying
(*), respectively (*)"" the natural basis of X p , r, w is equivalent to the unit vector
basis of some modular space.
We first consider the case 1 ~ r <p <00. For every n we define the function
_ p r r _{w~tr, O~t~wi!(p-r)
Mwn(t)-max{t,wnt}- p r/(p-r)
t, Wn <t<oo.
It is clear that each Mwn is an Orlicz function satisfying r~ tM:.vn(t)/Mwn(t)~p for

every t> 0 for which the derivative exists. Moreover, it is easily checked that a series
n~l Mwn (lanD converges if and only if max
00 {(
n~l lanl P
'" (
)l/P 'n~l
00
lanlrw~
)l/r} <00.
This implies that the natural basis of X p , r, w is equivalent to the unit vector basis
of I{M wn )' Consequently, X p , r~ l{Mwn}'
In the case whenp=oo we put
00
Then, for any p > 0, a series L Mwn(lanll p) converges if and only if

n=1
It follows from this fact that the natural basis of X,Xl,r,w is equivalent to the unit
vector basis of h{M"'n}' Obviously, in this case the sequence {MW ):=1 is not equivalent
to any sequence satisfying the uniform Ll 2 -condition at zero.
In order to determine the dual of Xp,r we restrict ourselves to the case
1 <r<p~oo (for r= 1, the functions complementary to Mwn are degenerate). We
take q and s so thatp-1+ q -1=1 and r- 1+s- 1=1 (q=1 ifp=oo). For p<oo the
function M:t", complementary to. M wn , can be computed directly from the definition
and we get
{
tSlsrS-lW~, o~ t~ rw~(S-q)
M:tn(t) = tw~,q-s)/(s-q)- w~ql(s-q), rw'j/(S-q) < t~pw'j/(s-q)
. tqlqpq-\ pw'j/(S-q)<t<oo.
For computations it is however more convenient to use the non-convex function

N wn , defined by
which has the property that A-l~Nwn(t)IM:tn(t)~A for every t>O and for some
°
constant A> which depends only on p and r but not on t and n. This formula is
valid also for p=oo, i.e. q= 1.
It follows from this discussion that the dual Y q,., 1 ~q<s<oo, of Xp,n con-
sists of all sequences y=(al, a2,"') for which ~ Nwn(lani) <00.
n=l
Before we state the main result of this section we prove an elementary tech-
nical lemma which will be used in the sequel.
Lemma 4.d.S. [102] Let Qo be a continuous junction on an interval [0, to] so that
° °
(i) Qo(O) = and Qo(t) > for 0< t ~ to,
(ii) Qo has a right derivative which satisfies ° ~ tQ~(t)1 Qo(t) ~ 1 jar 0< t < to.
°
Then there exists a concave increasing twice continuously differentiable junction
Q so that Qo(t)/8 ~ Q(t) ~ Qo(t) for ~ t ~ to.
Proof Let fP(t) be a continuously differentiable function on [0, toJ so that

1/4 ~ rp(t) ~ 1/2, fP/(t) ~ 0, 4trp'(t) + tQ~(t)/ Qo(t) ~ 1 for 0< t < to and so that fP/(t) > 0
whenever t is an interior point of {t; Q~(t)=O}. These properties of fP ensure that
Ql = QofP is a continuous strictly increasing function on [0, to] satisfying
tQ~(t)1 Q1(t) ~ 1 on that interval. Therefore, the inverse function Nl of Ql is also
strictly increasing and satisfies tN~(t)1 N(t) ~ 1, in a certain neighbourhood of 0. It
t
follows that N 2 (t) = J(N1(u)/u)du is a continuously differentiable convex strictly
o
d. Modular Sequence Spaces and Subspaces of Ip EB Ir 173
increasing function such that N 1(t/2) < N 2 (t) <N 1(t) in some neighborhood ofO. In
t
order to get a twice continuously differentiable function we take N(t) = f(N2 (u)/u)du
o
which satisfies N 1(t/4) <N(t) <N 1(t). The inverse function Q of N is a twice
continuously differentiable concave function on [0, to] which satisfies
Q(t)/4 < Q1(t) < Q(t) and thus also Qo(t)/8 < Q(t) < Qo(t) for t E [0, to]. D
We are now ready to prove the theorem of J. T. Woo [145].
Theorem 4.<1.9. Let 1 <q<s< 00. A Banach space X with a symmetric basis is iso-
morphic to a subspace of Yq,s if and only if X is isomorphic to an Orlicz sequence
space 1M with M satisfying q< tM'(t)/M(t) <s for every t>O.
Proof Let Xbe a Banach space with a symmetric basis {X n};:'=l which is isomorphic
to a subspace of Yq,s' Since Yq,s~I{M;:'n} we can assume with no loss of generality
that {X n};:'=l is equivalent to a normalized block basis {Uj}i'=l of the unit vector basis
of I{M;:'n}' As in the case of Orlicz sequence spaces (see the discussion preceding
4.a.7), there exists a sequence of Orlicz functions {Nj}i'=l> with each N j being a
normalized convex combination of the M;n's, such that a series
j=l
lajuj converges
00
if and only if 2: Nj(Jaji) <00. The functions {M;n};:'=l, and thus also {Nj}i'=b satisfy
j=l
the uniform L1 2 -condition at zero; more precisely, we have q<tN;(t)/Nit)<s for
all j and t > O. Hence, for any N which is a uniform limit of a subsequence of
{Nj}i'=l, {Uj}i'=l is equivalent to the unit vector basis of IN' Thus, X~IN and
q< tN'(t)/N(t)<s for every t> O. (Instead of repeating the argument used in 4.a.7
we could have also applied 4.d.5.)
In order to prove the converse we consider an Orlicz function M(t) which
satisfies the assumptions of the theorem, is not equivalent to t q and for which
M(1) = 1. We can exclude the case when M(t) is equivalent at zero to t q since Yq, s
contains even it complemented subspace isomorphic to lq. Indeed, by taking a
00 00
subsequence of integers {n i }f;"l for which 2: w~~/(p-r)<oo ifp<oo, or 2: w~!<oo if
i= 1 i=l
p=oo, we get that I{M;:'nj} is a complemented subspace of 1{M';,;n}~ Yq,s which is iso-
morphic to lq.
Put Qo(t)=M(t 1/(S-q))/t q /(S-q). In view of the conditions imposed on M we
get that Qo satisfies all the requirements of 4.d.8. Hence, there is a concave
increasing function Q with two continuous derivatives and Q(O) = 0 which is
equivalent to Qo. Let W={W n};:'= 1 be a sequence satisfying (*) if p <00 or (*)00 if
p=oo, let an=w~(s-q) and bn=a~(s-q)/(2S-q). Put
f Nwn(bntu) [ -(S-q)uq-2S-1Q"(a~-qb~-Suq-S)] du,

1
Gn(t)=C;;l
an
where
1
Cn= f Nwn(bnu) [ -(S-q)uq-2S-1Q"(a~-%~-Suq-S)] du.
an
This formula is inspired by the method used in the proof of 4.a.9. By substituting
the explicit value of N wn and putting v=a:.-qb~-·uq-· we get that
+tq J -(s-q)b~u2q-2s-1Q"(a:.-qb~-Suq-S) dU]

anlbnt
tS -" (anlbn)S-" ]
=C;l [ tS)s Q"(V) dv+t q tsi. VQ"(V) dv
n
= C;1{t8[Q'(t S- q)- Q'(b~-S)] + tq[(an/bn)S-qQ'«an/bn)S-q)
-tS-qQ'(tS-q)]+tq 7"
(anlbn)S-q
Q'(v) dV}
= C;l{ - tSQ'(b~-·) + tq(an/bn)S-qQ'«an/bn)S-q)

+ tqQ(tS- q)- tqQ«an/bn)S-q)} .
Since the behavior of M at 00 is not important we may suppose that lim Q'(t)=O.
t-+ co
This implies that Q'(b~-8) ~ 0 as n ~oo. Moreover, from the fact that lim tQ'(t)=
t-+o
lim Q(t) =0 it follows that also the second term and the fourth term in the paren-
t-+o
thesis tend to zero since an/bn =a~/(28-q) ~ 0 as n ~oo. Thus, for t= 1, we get that
CnGn(1)=Cn ~ Q(1). Consequently, the following limit exists uniformly for
tE [0,1]
G(t) = lim Gn(t) = tqQ(t S- q)/ Q(I),

n-+co
and, since Q is equivalent to Qo, we get that G is equivalent to M. From the

definition of Gn it follows that, for every integer k, there is an integer n(k) and real
1(1e)
numbers {Ui.k}f~i and {AI.k}f~i in [0, 1] with L: At.k=1 such that
1=1
Since ANwn(bt) = Nvn(ct) for suitable Vn and c we can write the previous inequality
also in the form
for some Vt. k and Ct. k' This clearly shows that IG' or equivalently 1M , is isomorphic
to a subspace of the modular sequence space I{N. i • le ) (recall that the non-convex
functions N Vi • k can be always replaced by proper Orlicz functions). In order to
conclude the proof we remark that the sequence {vt.k}{~i k=l tends to zero and if
co 1(1e) co 1(k)
L L vf,'lc(p-r)=oo in the case when p<oo, or L L VLk=oo in the case when
k=lt=l k=lt=l
co 1(k)
p=oo, then I{N v .}~ Yq,s'
41<
If, on the other hand, L L
k=lt=l
vf,'Ic(P-T)<oo, respectively
co 1(k)
L L vL k <00, then by applying Holder's inequality we get that the unit vectors
k=lt=l
of I{N v" ,,} form a basis equivalent to the unit vector basis of Iq. Thus, IM~/q and
we have already proved that Yq, s contains a subspace isomorphic to Iq. 0
Using 4.d.9 together with 4.d.5 we obtain the following corollary.
Theorem 4.d.l0. Let l~q<s<oo and let {Mn}:'=l be a sequence ofOrliczfunctions

satisfying q~tM~(t)/Mn(t)~sfor all nand t E (0,1). Then Yq,s contains a subspace
isomorphic to I{Mn}'
By its construction the space Yq, s is a quotient of Iq EEl Is. Thus, every Orlicz
sequence space "between" q and s is isomorphic to a subspace of a quotient of
Iq EEl Is. Comparing 4.d.9 and 4.d.l0 with 2.c.14 and 2.d.l we note the marked
difference between the behavior of subspaces of quotients of Iq EEl Is from that of
complemented subspaces of Iq EEl Is and also from that of subspaces of quotients
of Ip.
e. Lorentz Sequence Spaces
The Lorentz sequence spaces d(w,p) which have already been mentioned in
Section 3.a, as well as the Lorentz function spaces (cf. [99]), were introduced in
connection with some problems of harmonic analysis and interpolation theory.
We do not study here this aspect; instead, we present briefly some results regarding
their geometric structure. Let us first recall the definition of a Lorentz sequence
space.
Definition 4.e.l. Let l~p<oo and let w={Wn}:'=l be a non-increasing sequence of

co
positive numbers such that Wl = 1, lim wn=O and
n-+aJ n=l
L Wn=OO. The Banach space
of all sequences of scalars x=(al> a2'''') for which
where 7T ranges over all the permutations of the integers, is denoted by dew, p)
and it is called a Lorentz sequence space.
If {a:}:'=l is a non-increasing rearrangement of the sequence {an}:'=l> i.e. {a:}:'=l

is a non-increasing sequence obtained from {lan l}:'=1 by a suitable permutation of

the integers then Ilxll = (
00
n~1 a;Pwn
)1iP for x=(a1, a2, ... ) E dew, p).
Lorentz sequence spaces are, in general, different from the previously defined
classes of spaces with a symmetric basis. For instance, the uniqueness of the
symmetric basis in lp (see 3.b.5) easily implies that no Lorentz sequence space is
isomorphic to an lp space. In some cases a Lorentz sequence space is isomorphic to
an Orlicz sequence space. An easy computation shows that this is the case, for
example, if Wn = 1/(1 + log n) for n ~ 1 and p ~ 1 arbitrary: the corresponding space
dew, p) is isomorphic to 1M, where M(t) = t P/(l + Ilog tl). By the uniqueness of the
symmetric basis in any Lorentz sequence space (see 4.eA below), a space dew, p)
is isomorphic to an Orlicz sequence space 1M if and only if they are identical, i.e.
they consist of the same sequences. In other words, dew, p);::;:,IM if and only if
2: A;:Wn <00 2: M(An) <00,

00 00
(t) ¢>
n=1 n=l
whenever {An}:'=1 is a non-increasing sequence of reals tending to zero.

G. G. Lorentz [100] has found necessary and sufficient conditions on w for the
existence of an Orlicz function M for which (t) holds and also necessary and suffi-
cient conditions on M so that there exists a sequence w={W n}:'=1 for which (t)
is satisfied. The conditions given by G. G. Lorentz were actually stated in the more
general context of function spaces but they can be easily translated into the
sequence spaces language, as follows.
Let w={Wn}:'=1 be a strictly decreasing sequence of reals such that w1=1,
00
lim wn=O and L Wn=OO (the requirement that W be strictly decreasing sequence
n-+co n=l
is not really a restriction since every dew, p) space is isomorphic to a space dew', p)
for which w' is a strictly decreasing sequence). Construct two continuous functions
Wet) and Set), defined both on [1,(0), such that W is strictly decreasing, S is
n
concave and strictly increasing, W(n)=w n and S(n)=sn= L w/
/=1
for n=l, 2, ....
Since S-l is a strictly increasing convex function on [1,(0) it follows easily that
F(t)=1/S-1(I/t) satisfies l~tF'(t)/F(t) for all tE(O, 1) for which the derivative
exists. This implies that F(t)/t is non-decreasing and therefore, F is equivalent at
t
zero to the Orlicz function N(t) = f(F(u)/u) duo With these notations we are pre-
o
pared to state the results of G. G. Lorentz [100].
Theorem 4.e.2. A Lorentz sequence space dew, 1) is isomorphic to an Orlicz sequence

space 1M (i.e. dew, l)=IM as sets) if and only if
00
(i) there exists a constant y>O such that L

I/W- 1(yw n) <00.
n=l
In this case M is equivalent to the Orlicz function N, defined above.
We also have
Theorem 4.e.2'. An Orlicz sequence space 1M is isomorphic to a Lorentz sequence

space dew, 1) if and only if
00
(i') there exists a constant 8>0 so that f M*(8M*-1(I/t)) dt<oo.

1
lfthis is the case the sequence w may be taken to be wn=M*-I(1/n), n=l, 2, ....
Since these two theorems are not used in the sequel we do not reproduce their
proofs here. The fact that both 4.e.2 and 4.e.2' apply only in the case p = 1 is not
really a restriction. It is relatively easy to check that, for any p> 1, we have
d(w,p)~IMp if and only if dew, l)~IMl' where the Orlicz functions Mp and Ml are
connected by the relation Mp(t) = M1(t P ).
Proposition 4.e.3. Let {e n},';"=1 be the unit vector basis of a Lorentz sequence space
qn+l
dew, p) with p?:-1. Then every normalized block basis un = 2: ajeio n= 1,2, ... such
i =qn +1
that lim ai =0 contains,for every e > 0, a subsequence {U n;}i'=1 which is 1 +e-equivalent
,_00
to the unit vector basis of lp and so that [Unj]i'=1 is complemented in dew, p).
Consequently, every infinite dimensional subspace of dew, p) contains comple-
mented subspaces which are nearly isometric to Ip.
Proof Since every change of signs and every permutation of the integers induces
an isometry in dew, p) we may assume, by switching to a subsequence if necessary,
that {a i }i"=1 is a non-increasing sequence of positive numbers.
Fix e> 0 and construct by induction two increasing sequences of integers
j-l
{n j}i'=1 and {rj}i=1 such that qn;<rj<qn;+l, QJ-l= 2: (qnlc+l-qnlc)~rj-qn; and
k=1
(. .:t
t=qnj+l
afWi_qnj)lIP <e/21+1 for all j. Then, for any set of coefficients {,\}i=l, we
have
On the other hand, it is easily checked that, for every normalized block basis and
therefore in particular for {Unj}i=lo we have IIJI '\Unjll~ C~1 1/\IPf'P.
In order to prove that [Unj Ji'=l is complemented in d(w,p) we set
Then, using Holder's inequality we get
i.e. P is a bounded linear projection from dew, p) onto [unjJi'= l ' The last assertion
follows easily from 1.a.I1. D
The following result from [4] is an immediate corollary of 4.e.3.
Theorem 4.e.4. Let X be a subspace of a Lorentz sequence space dew, p), p ~ 1 which
has a symmetric basis {Xn}~=l' Then, up to equivalence, {Xn}~=l is the unique sym-
metric basis of this subspace.
Proof Assume that {yn}~=l is another symmetric basis of X. If X,;:;;lp then, by

3.b.5, {Xn}~=l and {yn}~=l are equivalent. Otherwise, by 1.a.I2, {Xn}~=l is equivalent
qn+l
to a block basis Un = .z;
i=qn+ 1
aiel> n = 1, 2,... of {en}~= 1, the unit vector basis of
Tm+l
dew, p), and {Ym}:=l is equivalent to a block basis Vm = .z;
n=rm+l
bnxn, m= 1,2, ...
of {xn}~=l> with limsupbn#O (for, by 4.e.3, lim bn=O implies that {Yn}~=l is
n-+ co n-+ 0:>
equivalent to the unit vector basis of lp, contrary to our assumption). Hence, by
the symmetry of {Yn}~=l> we may assume without loss of generality that there is
an e>O so that, for every m, Ibnl~e for some rm<n<rm+1' Consequently, if a
.z; AnYn converges so does n=l
.z; AnXn.
00 00
series By interchanging the roles of {Xn}~=l

n=l
and {Yn}~=l we deduce the equivalence of these two bases. D
It follows from 4.e.3 and l.c.l2(a) that a Lorentz sequence space dew, p) is
reflexive if and only if p > 1. The dual d*(w, p) of dew, p) is never isomorphic to a
Lorentz sequence space. Indeed, assume that d*(w, p) = dew', q) for some sequence
w' and for some q~ 1. By 4.e.3 we must have p-1+ q -1=I (if p=I the assertion
is entirely obvious so we assume that p> 1). Let {en}~=l and {fn}~=l be the unit
.z;
00
vector bases of dew', q), respectively d*(w, p). Then, a series Anen converges
n=l
co
whenever A=(A1, A2, ... ) E lq while there is a A ¢ lq for which L: .>..,.en converges.
n=l
00
On the other hand, the convergence of L: Anfn implies that AE lq while there is a
n=l
co
AE lq for which L: Anfn fails to converge. This contradicts 4.eA.
n=l
It is easy to give an explicit representation of d*(w,p). In the case p> 1 this
space consists of all sequences x=(ab a2,.,,) for which
where the infimum is taken over all y=(Y1, Y2,''') E lq with Ilyll = 1, where
p-1+ q -1=1 and {at}~l denotes a non-increasing rearrangement of {lajl}~l (cf.
[45]).
In the remainder of this section· we present, without proofs, some results con-
cerning symmetr~c basis sequences in dew, p) spaces. These sequences have been
studied in [4]. It has been proved there that every symmetric basic sequence in a
Lorentz sequence space dew, p) is equivalent either to the unit vector basis of lp or to
00
a block basis {u~")}:=l generated by a vector O#a= L: anenEd(w,p) (recall

n=l
definition 3.a.8).
In many cases, block bases generated by one vector are themselves equivalent
to the unit vector basis of some Lorentz sequence space but this is not true in general.
For example, it has been shown in [4] that in dew, p), with Wn = l/n 1/2 (1 + log n)2,
00
n= 1,2, ... andp;?: 1 arbitrary, the block basis generated by the vector a= L: en/nl/2P
n=l
spans a subspace (with a symmetric basis) which is not isomorphic to any Lorentz
sequence space or to any lp space. On the other hand, there are cases where all
block bases generated by one vector are equivalent to each other. These Lorentz
sequence spaces have been characterized in [4].
Theorem 4.e.5. In a Lorentz sequence space dew, p) there are exactly two non-
equivalent symmetric basic sequences (namely, the unit vector basis of lp and that of
n
dew, p) itself) if and only if sup Snk/SnSk <00, where Sn = L: Wj'
n, k j = 1
References
1. Alfsen, E. M.: Compact convex sets and boundary integrals. Berlin-Heidelberg-New York:
Springer, 1971.
2. Altshuler, Z.: Characterization of Co and lp among Banach spaces with a symmetric basis.
Israel J. Math. 24, 39-44 (1976).
3. Altshuler, Z.: A Banach space with a symmetric basis which contains no lp or Co, and all its
symmetric basic sequences are equivalent. Compositio Math. (1977).
4. Altshuler, Z., Casazza, P. G., Lin, B. L.: On symmetric basic sequences in Lorentz sequence
spaces. Israel J. Math. 15, 140-155 (1973).
5. Ando, T.: Contractive projections in Lp-spaces. Pacific J. Math. 17, 391-405 (1966).
6. Babenko, K. 1.: On conjugate functions. Dokl. Akad. Nauk. SSSR 62, 157-160 (1948)
[Russian].
7. Bachelis, G. F., Rosenthal, H. P.: On unconditionally converging series and biorthogonaI
systems in a Banach space. Pacific J. Math. 37, 1-5 (1971).
8. Banach. S.: Theorie des operations lineaires. Warszawa, 1932.
9. Bessaga, C., Pelczynski, A.: On bases and unconditional convergence of series in Banach
spaces. Studia Math. 17, 151-164 (1958).
10. Bessaga, c., Pelczynski, A.: A generalization of results of R. C. James concerning absolute
bases in Banach spaces. Studia Math. 17, 165-174 (1958).
11. Bessaga, c., Pelczynski, A.: On subspaces of a space with an absolute basis. Bull. Acad. Sci.
Pol. 6, 313-314 (1958).
12. Blei, R. c.: A uniformity property for A(2) sets and Grothendieck inequality. to appear.
13. Botschkariev, S. V.: Existence of a basis in the space of analytic functions, and some proper-
ties of the Franklin system. Matern. Sbornik 24, 1-16 (1974), [translated from Russian].
14. Casazza, P. G.: James' quasi-reflexive space is primary. Israel J. Math. 26, 294-305 (1977).
15. Casazza, P. G., Lin, B. L.: On symmetric basic sequences in Lorentz sequence spaces II. Israel
J. Math. 17, 191-218 (1974).
16. Casazza, P. G., Lin, B. L.: Projections on Banach spaces with symmetric bases. Studia Math.
52, 189-193 (1974).
17. Ciesielski, Z.: Properties of the orthonormal Franklin system. Studia Math. 23, 141-157
(1963).
18. Ciesielski, Z., Domsta, J.: Construction of an orthonormal basis in ern(/d) and W;'(/d).
Studia Math. 41, 211-224 (1972).
19. Cohen, P. J.: On a conjecture of Littlewood and idempotent measures. Amer. J. Math. 82,
191-212 (1960).
20. Davie, A. M.: The approximation problem for Banach spaces. Bull. London Math. Soc. 5,
261-266 (1973).
21. Davie, A. M.: The Banach approximation problem. J. Approx. Theory 13, 392-394 (1975).
22. Davis, W. J. Embedding spaces with unconditional bases. Israel J. Math. 20, 189-191
(1975).
23. Davis, W. J., Dean, D. W., Singer, I.: Multipliers and unconditional convergence ofbiortho-
gonal expansions. Pacific J. Math. 37, 35-39 (1971).
24. Davis, W. J., Figiel, T., Johnson, W. B., Pelczynski, A.: Factoring weakly compact operators.
J. Funct. Anal. 17, 311-327 (1974).
25. Davis, W. J., Johnson, W. B.: A renorming of nonreflexive Banach spaces. Proc. Amer.
Math. Soc. 37, 486--488 (1973).
References 181
26. Davis, W. J., Singer, I.: Boundedly complete M bases and complemented subspaces in
Banach spaces. Trans. Amer. Math. Soc. 175, 299-326 (1973).
27. Day, M. M.: Some more uniformly convex spaces. Bull. Amer. Math. Soc. 47, 504-507
(1941).
28. Day, M. M.: Normed linear spaces, Third Edition. Berlin-Heidelberg-New York: Springer
1973.
29. Dean, D. W.: The equation L(E, X**)=L(E, X)** and the principle of local reflexivity.
Proc. Amer. Math. Soc. 40, 146-148 (1973).
30. Dean, D. W., Singer, I, Sternbach, L.: On shrinking basic sequences in Banach spaces.
Studia Math. 40, 23-33 (1971).
31. Dor, L. E.: On sequences spanning a complex h space. Proc. Amer. Math. Soc. 47,515-516
(1975).
32. Dugundji, J.: Topology. Boston: Allyn and Bacon Inc. 1966.·
33. Dunford, N., Schwartz, J.: Linear operators, Vol. I. New York: Interscience 1958
34. Dvoretzky, A., Rogers, C. A.: Absolute and unconditional convergence in normed linear
spaces. Proc. Nat. Acad. Sci. (U.S.A.) 36,192-197 (1950).
35. Edelstein, I. S., Wojtaszczyk, P.: On projections and unconditional bases in direct sums of
Banach spaces. Studia Math. 56, 263-276 (1976).
36. Ellentuck, E.: A.new proof that analytic sets are Ramsey. J. Symbolic Logic 39,163-165
(1974).
37. Enflo, P.: A counterexample to the approximation property in Banach spaces. Acta Math.
130, 309-317 (1973).
38. Farahat, J.: Espaces de Banach contenant h d'apres H. P. Rosenthal. Seminairc Maurey-
Schwartz, Ecole Polytechnique, 1973-74.
39. Figiel, T.: An example of an infinite dimensional Banach space non-isomorphic to its
Cartesian square. Studia Math. 42, 295-306 (1972).
40. Figiel, T.: Further counterexamples to the approximation problem, dittoed notes.
41. Figiel, T., Johnson, W. B.: The approximation property does not imply the bounded ap-
proximation property. Proc. Amer. Math. Soc. 41,197-200 (1973).
42. Figiel, T., Johnson, W. B.: A uniformly convex Banach space which contains no [p. Com-
positio Math. 29, 179-190 (1974).
43. Figiel, T., Lindenstrauss, J., Milman, V.: The dimension of almost spherical sections of convex
sets. Acta Math. (1977).
44. Garling, D. J. H.: Symmetric bases of locally convex spaces. Studia Math. 30, 163-181
(1968).
45. Garling, D. J. H.: A class of reflexive symmetric BK spaces. Canad. J. Math. 21, 602-608
(1969).
46. Gohberg, I. C., Krein, M. G.: Fundamental theorems on deficiency numbers, root numbers
and indices of linear operators [Russian]. Usp. Math. Nauk 12, 43-118 (1957) [Translated
in Amer. Math. Soc. Translations, vol. 13].
47. Gordon, Y., Lewis, D. R.: Absolutely summing operators and local unconditional structures.
Acta Math. 133, 27-48 (1974).
48. Grothendieck, A.: Produits tensoriels topologiques et especas nucleaires. Memo. Amer.
Math. Soc. 16 (1955).
49. Grothendieck, A.: Resume de la theorie metrique des produits tensoriels topologiques. Bol.
Soc. Mat. Sao Paulo 8, 1-79 (1956).
50. Hagler, J.: A counterexample to several questions about Banach spaces. Studia Math. to
appear.
51. Hausdorff, F.: Set theory. New York: Chelsea, 1962.
52. Hennefeld, J.: On nonequivalent normalized unconditional bases for Banach spaces. Proc.
Amer. Math. Soc. 41,156-158 (1973).
53. James, R. C.: Bases and reflexivity of Banach spaces. Ann. of Math. 52, 518-527 (1950).
54. James, R. C.: A non-reflexive Banach space isometric with its second conjugate. Proc. Nat.
Acad. Sci. (U.S.A.) 37,174-177 (1951).'
55. James, R. C.: Separable conjugate spaces. Pacific J. Math. 10, 563-571 (1960).
182 References
56. James, R. C.: Uniformly non-square Banach spaces. Ann. of Math. 80, 542-550 (1964).
57. James, R. C.: A separable somewhat reflexive Banach space with non-separable dual. Bull.
Amer. Math. Soc. 80, 738-743 (1974).
58. James, R. C.: A nonreftexive Banach space that is uniformly nonoctahedral. Israel J. Math.
18, 145-155 (1974).
59. Johnson, W. B.: On quotients of Lp which are quotients of [p. Compositio Math. 33, (1976).
60. Johnson, W. B., Rosenthal, H. P.: On w*-basic sequences and their applications to the study
of Banach spaces. Studia Math. 43, 77-92 (1972).
61. Johnson, W. B., Rosenthal, H. P., Zippin, M.: On bases, finite-dimensional decompositions
and weaker structures in Banach spaces. Israel J. Math. 9, 488-506 (1971).
62. Johnson, W. B. Szankowski, A.: Complementably universal Banach spaces. Studia Math.
58,91-97 (1976).
63. Johnson, W. B., Zippin, M.: On subspaces of quotients of (2:Gn)zp and (2:Gn)co. Israel J. Math.
13, 311-316 (1972).
64. Johnson, W. B., Zippin, M.: Subspaces and quotient spaces of (2:Gn)zp and (2:Gn)o. Israel J.
Math. 17, 50-55 (1974).
65. Kadec, M. 1.: On the connection between weak and strong convergence. Dopovidi Akad.
Nank Ukrain 9, 949-952 (1959), [Ukrainian].
66. Kadec, M. 1.: On complementably universal Banach spaces. Studia Math. 40, 85-89 (1971).
67. Kalto'n, N. J.: Orlicz sequence spaces without local convexity. to appear.
68. Karlin, S.: Bases in Banach spaces. Duke Math. J. 15, 971-985 (1948).
69. Kato, T.: Perturbation theory for nullity deficiency and other quantities of linear operators.
J. Analyse Math. 6273-322 (1958).
70. Kircev, K. P., Troianskii, S. L.: On Orlicz spaces associated to Orlicz functions not satisfying
the ..:1 2 -condition [Russian]. Serdica 1, 88-95 (1975).
71. Klee, V. L.: Mappings into normed linear spaces. Fund. Math. 49, 25-34 (1960/61).
72. Kothe, G.: Das Triigheitsgesetz der quadratischen Formen im Hilbertschen Raum. Math.
Z.41, 137-152 (1936).
73. Kothe, G.: Hebbare lokalkonvexe Riiume. Math. Ann. 165, 181-195 (1966).
74. Kothe, G.: Topological vector spaces I. Berlin-Heidelberg-New York: Springer 1969.
75. Krasnoselskii, M. A., Rutickii, Ya. B.: Convex functions and Orlicz spaces. Groningen,
Netherlands 1961 [Translated from Russian].
76. Krein, M. G., Milman, D. P., Rutman, M. A.: On a property of the basis in Banach space.
Zapiski Mat. T. (Kharkov) 16, 106-108 (1940) [Russian].
77. Kwapien, S.: On a theorem of L. Schwartz and its applications to absolutely summing opera-
tors. Studia Math. 38, 193-201 (1970).
78. Kwapien, S., Pelczynski, A.: The main triangle projection in matrix spaces and its application.
Studia Math. 34, 43-68 (1970).
79. Lacey, E.: The isometric theory of classical Banach spaces. Berlin-Heidelberg-New York:
Springer 1974.
80. LeVeque, W. J.: Elementary theory of numbers. London: Addison-Wesley 1962.
81. Lindberg, K. J.: On subspaces of Orlicz sequence spaces. Studia Math. 45,119-146 (1793).
82. Lindenstrauss, J.: On a certain subspace of h. Bull. Acad. Polon. Sci. 12, 539-542 (1964).
83. Lindenstrauss, J.: On non-separable reflexive Banach spaces. Bull. Amer. Math. Soc. 72,
967-970 (1966).
84. Lindenstrauss, J.: On complemented subspaces of m. Israel J. Math. 5, 153-156 (1967).
85. Lindenstrauss, J.: On James' paper "separable conjugate spaces". Israel J. Math. 9, 279-
284 (1971).
86. Lindenstrauss, J.: A remark on symmetric bases. Israel J. Math. 13, 317-320 (1972).
87. Lindenstrauss, J., Pelczynski, A.: Absolutely summing operators in !&'p spaces and their
applications. Studia Math. 29, 275-326 (1968).
88. Lindenstrauss, J., Pelczynski, A.: Contributions to the theory of the classical Banach spaces.
J. Funct. Anal. 8, 225-249 (1971).
89. Lindenstrauss, J., Rosenthal~ H. P.: Automorphisms in Co, h and m. Israel J. Math. 7, 227-
239 (1969).
References 183
90. Lindenstrauss, J., Rosenthal, H. P.: The!l'p spaces. Israel J. Math. 7, 325-349 (1969).
91. Lindenstrauss, J., Stegall, C.: Examples of separable spaces which do not contain 11 and
whose duals are non-separable. Studia Math. 54, 81-105 (1975).
92. Lindenstrauss, J., Tzafriri, L.: On the complemented subspaces problem. Israel J. Math. 9,
263-269 (1971).
93. Lindenstrauss, J., Tzafriri, L.: On Orlicz sequence spaces. Israel J. Math. 10, 379-390 (1971).
94. Lindenstrauss, J., Tzafriri, L.: On Orlicz sequence spaces II. Israel J. Math. 11, 355-379
(1972).
95. Lindenstrauss, J., Tzafriri, L.: On Orlicz sequence spaces III. Israel J. Math. 14, 368-389
(1973).
96. Lindenstrauss, J., Tzafriri, L.: Classical Banach spaces. Lecture Notes in Math. Berlin-
Heidelberg-New York: Springer 1973.
97. Lindenstrauss, J., Zippin, M.: Banach spaces with a unique unconditional basis. J. Funct.
Anal. 3, 115-125 (1969).
98. Lorch, E. R.: Bicontinuous linear transformations in certain vector spaces. Bull. Amer.
Math. Soc..45, 564-569 (1939).
99. Lorentz, C. G.: Some new functional spaces. Ann. of Math. 51, 37-55 (1950).
100. Lorentz, G. G.: Relations between function spaces. Proc. Amer. Math. Soc. 12, 127-132
(1961).
101. Markushevich, A I.: On a basis in the wide sense for linear spaces. Dokl. Akad. Nauk 41,
241-244 (1943).
102. Matuszewska, W.: Regularly increasing functions in connection with the theory of L*'P-
spaces. Studia Math. 21, 317-344 (1962).
103. Maurey, B.: Une nouvelle demonstration d'un theoreme de Grothendieck. Seminaire
Maurey-Schwartz, 1972-1973.
104. Maurey, B., Rosenthal, H. P.: Normalized weakly null sequences with no unconditional
subsequence. Studia Math. to appear.
105. McCarthy, C. A, Schwartz, J.: On the norm of a finite Boolean algebra of projections and
applications to theorems of Kreiss and Morton. Comm. Pure Appl. Math. 18, 191-201 (1965).
106. Milman, V. D.: The geometric theory of Banach spaces, Part I. Usp. Math. Nauk. 25, 113-
173 (1970) [Russian]. English translation in Russian Math. Surveys 25,111-170 (1970).
107. Milman, V. D.: The geometric theory of Banach spaces, Part II. Usp. Math. Nauk 26,73-
149 (1971) [Russian]. English translation in Russian Math. Surveys 26, 79-163 (1971).
108. Nash-Williams, C. St. J. A.: On well-quasi-ordering transfinite sequences. Proc. Cambridge
Phil. Soc. 61, 33-39 (1965).
109. Odell, E., Rosenthal, H. P.: A double dual characterization of separable Banach spaces
containing 11 • Israel J. Math. 20, 375-384 (1975).
110. Orlicz, W.: Dber unbedingte Konvergenz in Funktionraumen. Studia Math. 1, 83-85 (1930).
111. Orlicz, W.: Dber eine gewisse Klasse von Raumen vom Typus B. Bull. Intern. Acad. Pol.
8, 207-220 (1932).
112. Ovsepian, R. I., Pelczynski, A: The existence in every separable Banach space of a funda-
mental total and bounded biorthogonal sequence and related constructions of uniformly
bounded orthonormal systems in L2. Stuejia Math. 54, 149-159 (1975).
113. Pelczynski, A.: On the isomorphism of the spaces m and M. Bull. Acad. Pol. Sci. 6, 695-696
(1958).
114. Pelczynski, A: Projections in certain Banach spaces. Studia Math. 19, 209-228 (1960).
115. Pelczynski, A: On the impossibility of embedding of the space L in certain Banach spaces.
ColI. Math. 8, 199-203 (1961).
116. Pelczynski, A: A characterization of Hilbert Schmidt operators. Studia Math. 28, 355-360
(1967).
117. Pelczynski, A.: Universal bases. Studia Math. 32, 247-268 (1969).
118. Pelczynski, A.: Any separable Banach space with the bounded approximation property is
a complemented subspace of a Banach space with a basis. Studia Math. 40, 239-242 (1971).
119. Pelczynski, A.: All separable Banach'spaces admit for every e>O fundamental and total
biorthogonal sequences bounded by 1 + e. Studia Math. 55, 295-304 (1976).
184 References
120. Pe1czynski, A., Singer, I.: On non-equivalent bases and conditional bases in Banach spaces.
Studia Math. 25, 5-25 (1964).
121. Pietsch, A.: Absolute p-sumrnierende Abbildugen in normierten Riiumen. Studia Math. 28,
333-353 (1967).
122. Pietsch, A.: Theorie der Operatorenidea1e. A forthcoming book.
123. Reitz, R. E.: A proof of the Grothendieck inequality. Israel J. Math. 19, 271-276 (1974).
124. Rolewicz, S.: Metric Linear Spaces. Monografie Matematyczne Warsaw 1972.
125. Rosenthal, H. P.: On complemented and quasi-complemented subspaces of quotients of
C(S) for Stonian S. Proc. Nat. Acad. Sci. (U.S.A.) 60, 1165-1169 (1968).
126. Rosenthal, H. P.: On totally incomparable Banach spaces. J. Funct. Anal. 4,167-175 (1969).
127. Rosenthal, H. P.: On the subspaces of U (p>2) spanned by sequences of independent
random variables. Israel. J. Math. 8, 273-303 (1970).
128. Rosenthal, H. P.: On the span in U of sequences of independent random variables (II).
Proc. of the 6th Berkeley Symp. on Prob. and Stat., Berkeley, Calif. 1971.
129. Rosenthal, H. P.: A characterization of Banach spaces containing /r. Proc. Nat. Acad. Sci.
(U.S.A.) 71, 2411-2413 (1974).
130. Rosenthal, H. P.: Point-wise compact subsets of the first Baire class. Amer. J. Math. to
appear.
131. Schechtman, G.: On Pe1czynski's paper "Universal bases". Israel J. Math. 22, 181-184 (1975).
132. Schonefeld, S.: Schauder bases in the Banach spaces Ck(P). Trans Amer. Math. Soc. 165,
309-318 (1972).
133. Schwartz, L.: Probabilites cylindriques et applications radonifiantes. C. R. Acad. Paris 268,
646-648 (1969).
134. Singer, I.: On Banach spaces with a symmetric basis. Rev. Math. Pure Appl. 6,159-166 (1961).
135. Singer, I.: Bases in Banach spaces I. Berlin-Heidelberg-New York: Springer 1970.
136. Sobczyk, A.: Projection of the space m on its subspace Co. Bull. Amer. Math. Soc. 47, 938-
947 (1941).
137. Szankowski, A.: Embedding Banach spaces with unconditional bases into spaces with sym-
metric bases. Israel J. Math. 15, 53-59 (1973).
138. Szankowski, A.: A Banach lattice without the approximation property. Israel J. Math. 24,
329-337 (1976).
139. Tong, A. E.: Diagonal submatrices of matrix maps. Pacific J. Math. 32, 551-559 (1970).
140. Tsirelson, B. S.: Not every Banach space contains lp or Co. Functional Anal. Appl. 8, 138-
141 (1974) [translated from Russian].
141. Urbanik, K.: Some prediction problems for strictly stationary processes. Proc. of the 5th
Berkeley Symp. II (part I), pp. 235-258.
142. Veech, W. A.: Short proof of Sobczyk's theorem. Proc. Atner. Math. Soc. 28, 627-628 (1971).
143. Wojtaszczyk, P.: On projections and unconditional bases in direct sums of Banach spaces II.
Studia Math. to appear.
144. Woo, J. Y. T.: On modular sequence spaces. Studia Math. 48, 271-289 (1973).
145. Woo, J. Y. T.: On a class of universal modular sequence spaces. Israel J. Math. 20, 193-215
(1975).
146. Zippin, M.: On perfectly homogeneous bases in Banach spaces. Israel J. Math. 4, 265-
272 (1966).
147. Zippin, M.: A remark on bases and reflexivity in Banach spaces. Israel J. Math. 6, 74-79
(1968).
148. Zippin, M.: Interpolation of operators of weak type between rearrangement invariant
function spaces. J. Funct. Anal. 7, 267-284 (1971).
149. Zippin, M.: The separable extension problem. Israel J. Math. 26, 372-387 (1977).
Subject Index
absolutely summing operators (see operators) basis, shrinking 8, 10, 14,21,22, 85

absolute convergence 16, 64, 65 -, subsymmetric (see subsymmetric basis)
approximation problem 29 -, symmetric (see symmetric basis)
- property (A.P.) 29, 30, 32, 35, 40-42, 84, -, unconditional (see unconditional basis)
86,90,94,95 -, uniqueness of 5
- - , bounded (B.A.P.) 37,38,41,42,51,92, -, weak null 28
95 biorthogonal functionals (see basis)
- -, compact (C.A.P.) 94 - system 42, 44, 46
- - , (e, A) 41 Boolean independence 100, 101
- - in subspaces of Co and I. 90, 91 Borsuk's antipodal map theorem 77
- - i n X· 33,34,41,42 bounded approximation property (see approxi-
- - , A 37 mation property)
- -, metric (M.A.P.) 37-42 boundedly complete basic sequence (see basic
Auerbach system 16,17,38,43,46 sequence)
averaging projections (see projections) - - basis (see basis)
- - minimal system (see minimal system)
Baire class 102

basic sequence 1,2,4, 5-7, 92 compact sets in Banach spaces 30
- -, boundedly complete 13 - - , uniform convergence on 31,32
- - , block 6 complementably universal spaces (see universal
- - i n X* 8,10 spaces)
- - , shrinking 14 complexification 81
- - , weak* 10,11 condition .::12 at zero 138-140, 143, 145, 146,
basis 1,7, 30, 34, 37, 38, 43, 44, 47, 48, 91 148, 149, 152, 166
- , biorthogonal functionals associated to 7, - - - - , uniform 167,168,172,173
11,22, 118 - .::I~ - -, 148, 149
- , block 6, 7, 19,49, 53, 59, 61, 62, 73, 133, - - - - , uniform 167,168
135,141 -.::I Q 138
- . boundedly complete 9, 10, 12, 14,22, 26, conditional basis 74
138, 143, 168 - expectation 116
- , conditional (see conditional bases)
- , complemented subspace of a space with a
38 Davie's construction 87, 90, 94, 95
- , constant 2,19 decomposition method (see Pelczynski's de-
- . duality of 7,10 composition method)
- , monotone 2, 38 disc algebra, basis in 4, 37
-. normalized 2 distortion problem 97
-of X* 8-10 Dvoretzky-Rogers theorem 16,65
- , perfectly homogeneous 61,62
- , perturbation of a 5
- , problem 4,27,30,34,51 equiValence of bases 5,59, 71, 73, 129
-, projections associated to 2, 7, 30 - - Orlicz functions 139, 153, 159, 166
186 Subject Index
equivalence of sequences of Orlicz func- norming functionals 44

tions 167
- - subsequences of a basis 114
extension of operators (see operators)
operator ideal 77
- property 106, 107
operators, absolutely summing 63, 64, 67
-, between Orlicz sequence spaces 149
-, compact 29, 32, 33, 51, 57, 67, 76, 94, 149
Fredholm operators (see operators) -, diagonal 20,21,71, 151
fundamental system 43, 44, 46 -, diagonal with respect to blocking 50
Franklin system 4 - , extension of 105,110
-, Fredholm 77-79,110,111
-, from Co into Ip 70, 71, 76
Grothendieck's inequality 68 - from It into 12 69
- universal constant 68, 70 - from I, into Ip 76
-, Hilbert-Schmidt 67, 74
-, index of 77, 79, 80, 110
Haar system 3, 19, 20 -, matrix representation of (see matrix)
-, p-absolutely summing 63-65, 67, 69-71
- of finite rank 29, 32, 33, 39
independent random variables 87 -, space of 37
index of an operator (see operators) -, spectrum of 80
injective spaces 105, 106, 111 -, strictly singular 75, 77, 79, 80, 145, 146,
interpolation 125 160
-, weakly compact 57,65, 106, 111
Orlicz functions 115, 137, 138, 140, 141, 145,
James' function space 103 146, 150, 173, 176
- tree space 103 - - , complementary functions of 147-149
- space J (see space J) - -, degenerate 137, 138, 140, 141, 147, 167
- - , equivalence of (see equivalence)
- -, interval associated to 143, 144, 149,
Khintchine's inequality 66, 72 153-155, 158, 162, 163, 165, 168, 175
- - , minimal 152-154,162,163,165
- -, right derivative of 139, 140, 143, 147,
149, 155, 172, 173
lifting property 104, 107-109
- sequence spaces 115, 116, 130, 137, 141,
Lorentz sequence spaces 115, 116, 130, 132,
143,145, 150, 166, 169, 173, 176, 177
137,175-179
- - -, dual of 147
- - -, unit vector basis of J 77
- - -, quotient spaces of 149
- - - , unit vector basis of 141, 150, 152,
166, 168
matrix 35, 68, 87, 89
-, block diagonal 21
-, diagonal of 20,21, 151
-, representation of operators 20, 83 PeIczynski's decomposition method 54, 57,
-, trace of 35, 90 93,111, 117, 131, 132, 156, 170
-, unitary 45, 87, 89 perturbation of a basis (see basis)
Maurey-Rosenthal example 28 Pietsch's factorization theorem 64
metric approximation property (see approxi- primary spaces 131
mation property) prime spaces 57, 131, 153
minimal system 42, 43, 46 projections, averaging 116
- - , block 46 -, associated to a basis (see basis)
- -, boundedly complete 46
- -, shrinking 46
modular sequence spaces 166-169
modulus of convexity 128 quotient map 11, 50, 78, 86, 108, 109
Subject Index 187
quotient map of an Orlicz sequence space (see space d (w,p) (see Lorentz sequence spaces)
Orlicz sequence spaces) - hM (see Orlicz sequences spaces)
- space 10, 12, 14, 26, 35, 85, 86, 104 -H", 37
- J 25, 103, 132
- JF (see James's function space)
Rademacher functions 24, 66, 67 - JT (see James's tree space)
Ramsey sets 101 - L1 (0,1) 24
rearrangement of a sequence 175 - Lp(O, 1) 27, 63, 65, 72, 91, 97, 129, 131
reflexivity 9, 14, 23, 27, 34, 40, 46, 95, 104, - - , basis of 3, 19, 20
110, 124, 127 - L",(O, 1) 111
- , local 33 - h 21, 23, 27, 69, 71-73, 97, 99, 101, 103,
- of Lorentz sequence spaces 178 104,118
- of modular sequence spaces 168 - - , automorphisms of 108
- of Orlicz sequence spaces 148-150, 157, - - , quotient spaces of 108
158, 165, 166 - h(r) 108
- of quotient spaces 11 0, 111 -/2 18,71,73,110,111,118,153
representation of second dual 8 - 1M (see Orlicz sequence spaces)
-I" 27,53,54,70,73,84,86,90,91,95, 131,
132, 143, 144, 146, 149, 159, 163
Schauder basis (see basis) - - , basis of 3,19,53,59,61,73,94,99,121,
- decomposition 47,49 129, 177 .
- - , blocking of 49 - - , isometries of 112
- - , boundedly complete 48 - - , projections in 55, 56
- - , constant 47, 49 - - , quotient spaces of 86, 92, 175
- - , finite dimensional (F.D.D.) 48, 49, 85, - Ip EEl I, 75, 82, 137, 166, 169
86 - - , unconditional bases of 82, 83
- - , shrinking 48, 49, 50 - - , quotient spaces of 175
- - , unconditional 48, 72 -I", 57,58, 103, 105, 111, 138, 168
- - , unconditional constant 49 - - , automorphisms of 110
- system 3, 20 - - , quotient spaces of 111,112
- Tychonoff fixed point theorem 143 - /",(r) 105, 106
sequences of zeros and ones 156,161,164,166 - X"., 171, 172
set CM 140, 143, 144, 158 - X".,.,. 169
- CMA 140, 143, 149, 150, 157 - X".2.,. 169
-EM 140 - X"", 171
- EMA 140, 149, 157, 161 - X.,.,.,. 171
shrinking basic sequence (see basic sequence) - Yq•• 172,173, 175
- basis (see basis) strong non-equivalence of OrIicz functions
- minimal system (see minimal system) 150, 159, 164
space C(O, 1) 24 subsymmetric basis 114, 116, 131, 132, 141
- - , basis of 3, 43 - - , block basis with constant coefficients of a
- - , compact sets in 141,152,162 116
- C(K) 57, 131 - - , sums of vectors of a 118
- Ck(ln), basis of 4 - constant 114, 115
- c, basis of 3, 20, 29 (see also summing -norm 114
basis) summing basis 20, 29, 73
- Co 22,23,27,37,53,54,70-73,84,86,90, symmetric basis 113, 114, 116, 117, 123, 124,
97,98, 103, 104, 106, 110, 118, 130, 132, 143, 129,132,136-138,140,153,178,179
144 - - , block bases generated by one vector of a
- - , automorphisms of 109 120, 121, 134, 135, 179
- - , basis of 3,19,53,59,61,120,121,129 - - , block bases with constant coefficients of a
- - , isometries of 112 116,117,127,150,152,171
- - , projections in 56 - - , functionals biorthogonal to 118, 120,
- - , quotient spaces of 107, 109 121
- Co EEl I, 171 - - , minimal 154
188 Subject Index
symmetric basis, sums of vectors of a 118, unconditional basis, uniqueness of 63, 71,
125, 139 73, 84, 117, 118, 130
- - , uniqueness of 129, 130, 153, 154, 157, - constant 18, 19
158, 164, 176, 178 - convergence 15,64,65,72,99
- constant 114, 115, 119 - F.D.D. (see Schauder decomposition)
-norm 114 uniform convexifiability 97
- convexity 97, 124, 127-129
universal Orlicz sequence spaces 154, 155, 156
- spaces 84, 91, 92, 129-131
Total system 43, 44, 46
- - in /p 94
totally incomparable spaces 75, 82, 112
trigonometric system 43
Tsirelson's example 95, 97, 132, 133, 136, 143 weak Cauchy sequences 99
- conditional compactness 23
- sequential completeness 23, 27
unconditional basic sequence 18,28, 92 - unconditional convergence (w.u.c.) 99
- basis 15, 18, 19,21-23,27,46,47,51,61, weak* convergence 101
62,70,91,92,95, 113, 123, 124,,126, 168 -limit 101
- - i n X* 23
- -, projections associated to 18
- -, spaces without an 24 Young's inequality 147
Ergebnisse der Mathematik und ihrer Grenzgebiete
1. Bachmann: Transfinite Zahlen
2. Miranda: Partial Differential Equations of Elliptic Type
4. Samuel: Methodes d'algebre abstraite en geometrie algebrique
5. Dieudonne: La geometrie des groupes cIassiques
7. Ostmann: Additive Zahlentheorie. 1. Teil: Allgemeine Untersuchungen
8. Wittich: Neuere Untersuchungen tiber eindeutige analytische Funktionen
10. Suzuki: Structure of a Group and the Structure of its Lattice of Subgroups. Second edition in
preparation
11. Ostmann: Additive Zahlentheorie. 2. Tei1:Spezielle Zahlenrnengen
13. Segre: Some Properties of Differentiable Varieties and Transformations
14. Coxeter/Moser: Generators and Relations for Discrete Groups
15. Zeller/Beckmann: Theorie der Limitierungsverfahren
16. Cesari: Asymptotic Behavior and Stability Problems in Ordinary Differential Equations
17. Severi: II teorema di Riemann-Roch per curve, superficie e varieta questioni collegate
18. Jenkins: Univalent Functions and Conformal Mapping
19. Boas-Buck: Polynomial Expansions of Analytic Functions
20. Bruck: A Survey of Binary Systems
21. Day: Normed Linear Spaces
23. Bergmann: Integral Operators in the Theory of Linear Partial Differential Equations
25. Sikorski: Boolean Algebras
26. Ktinzi: Quasikonforme Abbildungen
27. Schatten: Norm Ideals of Completely Continuous Operators
30. Beckenbach/Bellman: Inequalities
31. Wolfowitz: Coding of Theorems of Information Theory
32. Constantinescu/Cornea: Ideale Rander Riemannscher Flachen
33. Conner/Floyd: Differentiable Periodic Maps
34. Mumford: Geometric Invariant Theory
35. Gabriel/Zisman: Calculus of Fractions and Homotopy Theory
36. Putnam: Commutation y Properties of Hilbert Space Operators and Related Topics
37. Neumann: Varieties of Groups
38. Boas: Integrability Theorems for Trigonometric Transforms
39. Sz.-Nagy: Spektraldarstellung linearer Transformationen des Hilbertschen Raumes
40. Seligman: Modular Lie Algebras
41. Deuring: Algebren
42. Schtitte: Vollstandige Systeme modaler und intuitionistischer Logik
43. Smullyan: First-Order Logic
44. Dembowski: Finite Geometries
45. Linnik: Ergodic Properties of Algebraic Fields
46. Krull: Idealtheorie
47. Nachbin: Topology on Spaces of Holomorphic Mappings
48. A. Ionescu Tulcea/C. Ionescu Tulcea: Topics in the Theory of Lifting
49. Hayes/Pauc: Derivation and Martingales
50. Kahane: Series de Fourier absolument convergentes
51. Behnke/Thullen: Theorie der Funktionen mehrerer komplexer Veranderlichen
52. Wilf: Finite Sections of Some Classical Inequalities
53. Ramis: Sous-ensembles analytiques d'une variete banachique complexe
54. Busemann: Recent Synthetic Differential Geometry
55. Walter: Differential and Integral Inequalities
56. Monna: Analyse non-archimedienne
57. Alfsen: Compact Convex Sets and Boundary Integrals
58. Greco/Salmon: Topics in m-Adic Topologies
59. Lopez de Medrano: Involutions on Manifolds
190
60. Sakai: C*-Algebras and W*-A1gebras

61. Zariski: A1gebtaic Surfaces
62. Robinson: Finiteness Conditions and Generalized Soluble Groups, Part 1
63. Robinson: Finiteness Conditions and Generalized Soluble Groups, Part 2
64. Hakim: Topos anneles et schemas relatifs
65. Browder: Surgery on Simply-Connected Manifolds
66. Pietsch: Nuclear Locally Convex Spaces
67. Dellacherie: Capacites et processus stochastiques
68. Raghunathan: Discrete Subgroups of Lie Groups
69. Rourke/Sanderson: Introduction of Piecewise-Linear Topology
70. Kobayashi: Transformation Groups in Differential Geometry
71. Tougeron: Ideaux de fonctions diiferentiables
72. Gihman/Skorohod: Stochastic Differential Equations
73. Milnor/Husemoller: Symmetric Bilinear Forms
74. Fossum: The Divisor Class Group of a Krull Domain
75. Springer: Jordan Algebras and Algebraic Groups
76. Wehrfritz: Infinite Linear Groups
77. Radjavi/Rosenthal: Invariant-Subspaces
78. Bognar: Indefinite Inner Product Spaces
79. Skorohod: Integration in Hilbert Space
80. Bonsall/Duncan: Complete Normed Algebras
81. Crossley/Nerode: Combinatorial Functors
82. Petrov: Sums of Independent Random Variables
83. Walker: The Stone-Cech Compactification
84. Wells/Williams: Embeddings and Extensions in Analysis
85. Hsiang: Cohomology Theory of Topological Transformation Groups
86. Olevskii: Fourier Series with Respect to General Orthogonal Systems
87. Berg/Forst: Potential Theory on Locally Compact Abelian Groups
88. Weil: Elliptic Functions according to Eisenstein and Kronecker
89. Lyndon/Schupp: Combinatorial Group Theory
90. Edwards/Gaudry: Littlewood-Paley and Multiplier Theory
91. Gunning: Riemann Surfaces and Generalized Theta Functions
92. Lindenstrauss/Tzafriri: Classical Banch Spaces I

Lindenstrauss, Tzafriri - Classical Banach Spaces I - Sequence Spaces - (1977)

Uploaded by

Copyright:

Available Formats

Lindenstrauss, Tzafriri - Classical Banach Spaces I - Sequence Spaces - (1977)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lindenstrauss, Tzafriri - Classical Banach Spaces I - Sequence Spaces - (1977)

Uploaded by

Copyright:

Available Formats

Ergebnisse der Mathematik

und ihrer Grenzgebiete 92

Editorial Board: P. R. Halmos P. J. Hilton (Chair-

AMS Subject Classification (1970): 46-02, 46 A4S, 46Bxx, 46Jxx

ISBN-13: 978-3-642-66559-2 e-ISBN-13: 978-3-642-66557-8

Library of Congress Cataloging in Publication Data. Lindenstrauss, 10ram, 1936-.

© by Springer-Verlag Berlin Heidelberg 1977.

Jerusalem Joram Lindenstrauss

2. The Spaces Co and Ip 53

3. Symmetric Bases. 113

4. Orlicz Sequence Spaces . 137

Subject Index 185

and y* is isomorphic to the subspace (ker T).L of X*. Similarly, if Z is a subspace

a. Existence of Bases and Examples

Definition l.a.l. A sequence {Xn};:'~l in a Banach space X is called a Schauder

In this book we shall not consider any type of bases in infinite-dimensional

sequence space by identifying each x = L

projections P n: X -+ X, defined by P n (~ a1xi ) =

The projections {Pn}~= 1 are called the natural projections associated to

Proposition 1.a.3. Let {X,,}:'=1 be a sequence of vectors in X. Then {Xn}~=1 is a

(iii) The closed linear span of {X"}~=1 is all of X.

Before proceeding with the general discussion we present some examples of

Xl=(l, 1, 1, ... ) and, for n> 1, x n=en-l'

The expansion of x=(al> a2,"') E c with respect to this basis is

x=(lim an)xl + (aI-lim an)x2 + (a2- lim a,,)xs+'" .

An important example of a Schauder basis is the Haar system in Lp(O, 1),

is called the Haar system.

The proof, due to S. Mazur, is based on the following lemma.

Lemma 1.a.6. Let X be an infinite dimensional Banach space. Let Be X be a finite-

Ily+.\x11 ~ IIYi+AXII-s/2~ yt(Yi+.\x)-s/2= 1-s/2~ Ilyll/(l +s). D

such that Il (1 + sn):::;; 1 + s. Let Xl be any element in

The sequence {Xn},~)=1 is a basic sequence in X whose basis constant is ,;; 1 +0

provided a series 2: anxn converges if and only if 2: anYn converges.

We shall discuss in detail some aspects of uniqueness of bases in Chapters 2 and

Proof. For X= L: a"xn E X define Tx= ,,=1

(*) Ilx-Txll::::; L la"lllx"-y,,lI::::;max

Definition l.a.lO. Let {xn}:'= 1 be a basic sequence in a Banach space X. A sequence

It is obvious that a block basis {Uj}J'~1 of {Xn}:'=1 is a basic sequence whose

an element y E Ywith lIyll = I of the form y= L:

basis of {X,,}:'=1 inductively. Pick any Y1 = L: a",1xn E Y with Ilhll = 1. Let P1 be

Proposition l.a.12. Let {Xn};'= 1 be a Schauder basis of a Banach space X. Let

Yk= L: an,kXm k= 1,2, ... ,

Proposition 1.a.11 enables us to give an alternative proof of 1.a.S. It is clearly

h. Schauder Bases and Duality

Proposition l.b.l. Let {Xn}:'=l be a basis of a Banach space X. The biorthogonal

Proof. If {X~}:'=l is a basis of X* then, for every x* E X*, IIP:x*-x*II--+ O. Since

where K is the basis constant of {Xn}:'=l. Hence, IIP:x* -x*II--+ 0 0

If X has a shrinking basis it is possible to give a convenient representation of

implies that {J1 a1xi}:=1 must tend w* to x**.) 0

Observe that the canonical image of X in X** corresponds to those sequences

A typical example of a non-boundedly complete basis is the unit vector basis of

Proposition 1.b.4. A Banach space X with a boundedly complete basis {X n}:'=l is

By combining the notions of shrinking and boundedly complete we get the

This proves that X is reflexive. 0

In l.b.5 we considered a single basis in X. In [147] M. Zippin showed that if

For the proof of l.b.7 we introduce first the following.

Definition l.h.S. A basic sequence {X;f}:=l in the dual X* of a Banach space X is

Proposition l.b.9. A sequence {X;n:'=l E X* is a w* basic sequence if and only if

Prool 01 l.b.7. For every infinite-dimensional Banach space X the set

Proof. If {X~}:'=l is a basis of X* then, for every x* E X, IIP:x-x*II--+ O. Since

Proof. By l.b.ll there is no loss of generality to assume that in X, y: ~ y and

Iln~1 Ananxnll ~ n~l IAnllanx(xn)1 ~ s~p IAnl n~ Onanx(xn)

UX=(-A, x(et}-,\, x**(e:)-,\, ... ),