An Infinite Ramsey Theorem and Some Banach Space Dichotomies - Gowers

Annals of Mathematics
An Infinite Ramsey Theorem and Some Banach-Space Dichotomies

Author(s): W. T. Gowers
Source: Annals of Mathematics, Second Series, Vol. 156, No. 3 (Nov., 2002), pp. 797-833
Published by: Annals of Mathematics
Stable URL: http://www.jstor.org/stable/3597282 .
Accessed: 12/08/2013 22:34
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Annals of Mathematics is collaborating with JSTOR to digitize, preserve and extend access to Annals of
Mathematics.
http://www.jstor.org
This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

All use subject to JSTOR Terms and Conditions
Annals of Mathematics, 156 (2002), 797-833
An infinite Ramsey theorem and

some Banach-space dichotomies
BY W. T. GOWERS
Abstract
A problem of Banach asks whether every infinite-dimensionalBanach

space which is isomorphicto all its infinite-dimensionalsubspaces must be
isomorphicto a separableHilbert space. In this paper we prove a result of
a Ramsey-theoreticnature which implies an interesting dichotomyfor sub-
spacesof Banachspaces. Combinedwith a resultof Komorowskiand Tomczak-
Jaegermann,this gives a positive answerto BanachXs problem.We then gener-
alize the Ramsey-theoreticresult and deduce a furtherdichotomyfor Banach
spaces with an unconditionalbasis.
1. Introduction
This paper containsa completeproofof a result announcedin [G1]which

has been circulatingin preprintformfor severalyears [G2,3]. Our main theo-
rem, when combinedwith a very differentresult of Komorowskiand Tomczak-
Jaegermann,solves a problemfrom the famous 1932 book of Banach [B]. He
askedwhethera separableHilbertspaceis the only infinite-dimensional Banach
space, up to isomorphism,which is isomorphicto every infinite-dimensional
closed subspace of itself. The answerturns out to be yes) but most of the
proof appears to have little to do with the problem. Therefore,in order to
motivate the rest of the paper, we shall begin by explaining (in this section
and the next) how Banach'sproblemcan be reducedto a questionwith a much
more combinatorialflavour.
To save writing, let us assume from now on that all Banach spaces and
subspacesare infinite-dimensionalunless they are specifiedotherwise. Recall
that a (Schauder) basisof a Banachspace X is a sequence(xn)°n°=l such that
everya E X can be writtenuniquelyas a norm-convergent sum a = £°n°-lanXn
for some sequence(an)°c=lof scalars. If everyxn has normone then the basisis
normalized.It can be shownthat if (xn)°n°l is a basis, then there is a constant
C such that all the projectionsPN: @°n°=l anXn 1 5£,nN=l antn have norm at

798 W. T. GOWERS
most C. The smallest constant with this propertyis called the baszsconstant
of the basis. If the basis constant is 1, then the basis is called monotone. A
basis (Xn)n=l is unconditionalif the sum En-l anXn convergesunconditionally
wheneverit converges. This, it turns out, implies that there is aconstant
C such that, for every subset A c 1S (it is sufficientto consideronly finite
subsets), the projectionPA: En=l anXn 1 ngA antn has norm at most C.
The smallest constant with this propertyis called the ?l,nconditzonal constant
of the basis. A sequence(xn)°°=1which is a basis/unconditionalbasis of the
closed subspace that it spans is called a basic sequence/unconditionalbasic
sequence.
The result of Komorowskiand Tomczak-Jaegermann we shall use is the
following[K T-J].
THEORE1W 1.1. Let X be a Banach space with cotypeq for some q < oo.
Then eitherX has a subspacewzthoutan unconditionalbasis or X has a sub-
space isomorphicto T2
Let us followusual practiceand definea Banachspace to be homogeneous

if it is isomorphicto all its subspaces.It can be shownthat everyhomogeneous
Banach space satisfies the cotype condition above. Indeed, an old result of
Mazurstates that every Banach space has a subspacewith a basis. On the
other hand, Szankowskihas shown [Szl (generalizingEnflo's solution of the
basis problem)that if p < 2 and q > 2, then any Banach space which either
fails to have type p or fails to have cotype q has a subspacewithout a basis.
Hence,such a space cannot be homogeneous.So an immediateconsequenceof
Theorem1 and the definitionof homogeneityis the followingadditionalresult
of Komorowskiand Tomczak-Jaegermann.
COROLLARY 1.2. Let X be a homogeneousBanach space. Then either

X is isornorphicto T2or X fails to have an unconditionalbaszs.
Notice that in the second case no subspace of X has an unconditional

basis, since X is homogeneous. Thus Corollary1.2 appearsto imply a very
strong propertyof homogeneousspaces not isomorphicto 12. Indeed,it is not
at all obviousthat there exists a Banach space such that no subspacehas an
unconditionalbasis.
As it happens,the existenceof such spaceswas itself a long-standingopen
problemuntil 1991, when counterexampleswere found [GM1](see also [AD],
[F1], [G6],[GM2],[H]).However,these counterexamplesdid not come close to
being homogeneous.In fact, many of them had a propertywhichis almostthe
reverseof homogeneity.A Banachspace X is called decomposable if it can be
written as a direct sum Y + Z with the projectionsto Y and Z continuous.
Equivalently,X is decomposableif it admits a nontrivialprojection(that is,

ANINFINITE RAMSEY THEOREM 799
oneof infinite rank and infinite corank). Otherwise, X is tndecomposable.

Itis hereditarilyindecomposable if it has no decomposablesubspace. This is
theremarkablepropertywhich was enjoyedby many of the counterexamples.
Laterwe shall have moreto say about hereditarilyindecomposablespaces,but
fornow we shall simply quote the followingresult from [GM1].
THEOREM1.3. Banach space is isomor-
A hereditarilyindecornposable
phicto no propersubspaceof itself.
In particular,a hereditarilyindecomposablespace is certainlynot homo-
geneous.
We can now state our first Banach-spacedichotomy,which is one of the
mainresults of this paper.
THEOREM1.4. Every Banach space X has a subspaceW which either
has an unconditionalbasis or is hereditarilyindecomposable.
Note that this is a genuinedichotomy,since an unconditionalbasis allows
W to be decomposedin uncountablymany differentways. Note also that the
last two resultsimplythat a homogeneousspaceX must havean unconditional
basis, since otherwiseno subspaceof X has an unconditionalbasis, which im-
plies, by Theorem1.4, that some subspaceof X is hereditarilyindecomposable,
which implies, by homogeneity,that X is hereditarilyindecomposable,which
implies, by Theorem 1.3, that X is not homogeneous. However,we have al-
ready seen that a homogeneousspace with an unconditionalbasis must be
isomorphicto /2, SOBanach's problemis solved once Theorem 1.4 is estab-
lished.
The definitionswe have given of an unconditionalbasis and a hereditarily

indecomposableBanachspace are not the most conrrenientfor provingTheo-
rem 1.4. Instead,we use simplecharacterizationsof these conceptsin terms of
finite sequencesof vectors in the space. First we need some more definitions
and elementaryfacts concernedwith bases.
Because of MazurXs result that every Banach space has a subspacewith
a basis, and indeed a basis with basis constant arbitrarilyclose to 1, we shall
restrict attention to spaces with a monotone basis. Two bases (xn)°n°l and
(yn)ono=l are definedto be C-equivalentif there exist constants A and B such
that B/A < C and
oo oo oo
A|n=l n=lanUn||
% ||
E anXn|| % B) E anZnt|
n=l
for every sequenceof scalars (an)n=l such that either of the two sums con-
verges. If X is a Banachspace with a given basis (en)n=l: then the supportof
a vector x-E°n°=l xnen, written supp(x), is the set of n for which xn 78 0.

800 W. T. GOWERS
If max supp(x) < min supp(y) we write x < y. A block basis of X is defined to
be a sequence x1 < x2 < ... of nonzero vectors. A subspace of X generated
by a block basis is called a block subspace. The following very useful lemma of
Bessaga and Pelczynski (see [LT]) allows us to restrict our attention to block
bases and block subspaces.
LEMMA 1.5. Let X be a Banach space with a basis (en)n=l, let Y be a

subspace of X and let 6 > 0. Then Y has a subspace Z generated by a basis
which is (1 + E)-equivalent to a block basis of (en)n=l1
In fact, more is true. For every sequence (Jn)n = of positive real numbers,
we can find a subspace Z with a basis (zn)n=l such that there is a normalized
block basis (xn)n=l1 of (en)n I_with lIxn - Znl < n for every n. In other words,
Z can be chosen to be an arbitrarily small perturbation of a block subspace
of X. Moreover, given any basic sequence in Y, we can choose Z to be spanned
by a subsequence.
We now state two simple lemmas. Proofs can be found in [LT]and [GM1].
From now on, when we use block-basis terminology, we shall assume implicitly
that the Banach space under discussion comes with some particular chosen
normalized monotone basis (en)n l. This will save a good deal of writing.
LEMMA1.6. Let X be a Banach space. The following two statements are

equivalent:
(i) X has no subspace with an unconditional basis.
(ii) For every block subspace Y of X and every real number C there is a
sequence yi < ... < Yn of vectors in Y such that
n n
Slyi > c S(-l)iy
i=1 i=1
LEMMA 1.7. Let X be a Banach space. The following two statements are
equivalent:
(i) X is hereditarily indecomposable.
(ii) For every pair of block subspaces Y, Z of X and every real number C
there is a sequence yi < zi < Y2 < Z2 < ... < Yn < Zn of vectors such
that yi E Y and zi C Z for every i, and such that
(yi + i)| > c|(yi- i)|

i=1 i=l

<
AN INFINITE RAMSEY THEOREM 801
Notice that if we insist that Y = Z in condition (ii) of Lemma 1.7 then

we recover condition (ii) of Lemma 1.6 (with n replaced by 2n). This makes it
clearer what we need to do to prove Theorem 1.4. Let us say that a sequence
Yi < ... < Yn is C-conditional if
n n
Yi > Cl0 (-l)iyi .
i=1 i--l
Then we begin with a space X such that every block subspace contains a
C-conditional sequence for every C, and must find a (block) subspace with the
stronger property that for any two block subspaces and any C, a C-conditional
sequence can be found with its odd terms in one of the subspaces and its even
terms in the other. As we shall see later in the paper, a far more general result
is true, one which does not use in a strong way the definition of C-conditional.
2. Further definitions and some preliminary results
In this section we shall state the first of our main Ramsey-theoretic results,
which implies Theorem 1.4 very easily. It will be proved in the next section.
Given a Banach space X (with a specified basis) let us define Zf = Ef (X) to
be the set of all finite sequences xl < ... < Xn of nonzero vectors in the unit
ball of X. Given an arbitrary subset a C Ef (in other words, given some set
of finite block bases) the following two-player game can be defined, between
players S and P. On the nth move of the game, player S chooses a block subspace
Xn C X (infinite-dimensional), and player P chooses some point xn E Xn. The
aim of P is to construct a sequence (xi,..., Xn) E Cr,whereas the aim of S is
that at no stage should the sequence constructed by P belong to a.
Formally, a strategy for P is a function qfwhich, for any finite block basis
x1,..., Xn and any subspace Y c X, gives a vector x = $(xi,... .,xn; Y) E Y.
The strategy 0qis a winning strategy for P if, given any sequence XI, X2, ... of
subspaces of X, the sequence (x1, x2,...) defined inductively by xl =- q(0; X1)
and Xn+l = (xi, . .., xn; Xn+l) is in u.
If A = (51,2,...) is a sequence of positive scalars and a C Zf, then
the A-expansion of r, denoted oA, is the set of block bases (xi,... , Xn) E Ef
such that there exists (Yl,... ,Yn) E aowith Iyi- xi\[I 6i for every i. We
shall also define o-_A to be (((cr))A)c, or in other words the set of block bases
(xl,... , Xn) e Ef such that every (Yi,..., Yn) z a with \yi - xi 5i for all i
is in a. It is easy to check that (a-A)A C U C (crA)-A.
We now introduce some notation. If A = (i, ... ,Ym) and Y C X is a
subspace, then [A;Y] will stand for the set of sequences (Zn)N l E Zf such
that zi - yi for i < m and zi E Y for m < i < N. If A is the null sequence then
we shall write [Y]. Given a C E f, we write c[A; Y] for the set of sequences

802 W. T. GOWERS
(Xn)N=i such that (i,..., ym, xi,..., XN) E [A; Y] n o. Finally, when we say
that P has a winning strategy for the game cr[A;Y], S's moves are understood
to have to be subspaces of Y.
Our first Ramsey-type theorem is the following.
THEOREM2.1. Let X be a Banach space, let a be any subset of Ef(X)

and let A be a sequence of positive real numbers. Then X has a subspace Y
such that either a[Y] - 0 or P has a winning strategy for the game CA4[Y].
We refer to this as a Ramsey theorem for two reasons. The first is that its
proof closely resembles existing arguments in Ramsey theory. However, even
the statement can be regarded as Ramsey-theoretical. If we call sequences in
acblue and those not in a red, then the theorem gives us a subspace such that
either every finite block sequence is red or there is such an abundance of small
perturbations of blue sequences that P has a winning strategy for obtaining
them. One might hope to find a subspace where all sequences were close to
blue sequences, but a strengthened statement along these lines is false (for
nontrivial reasons, see the appendix).
Many arguments in infinite Ramsey theory depend on one particular di-
agonalization procedure. It will be used often enough in this paper for it to be
well worth stating as an abstract principle. We shall need some more notation.
A *-pair is a pair (A, Z) where A E Ef, Z is an infinite-dimensional block
subspace and x < z for every x E A and z E Z. We shall sometimes refer to
*-pairs simply as pairs, and we shall use the notation A < Z as shorthand for
the support condition above. If A is a sequence (i, 62,... ) of positive reals
and A = (xi, ..., xn) and B = (yi,..., Yn) are block bases of the same size, we
use the shorthand notation d(A, B) <cA to mean d(xi, yi) < &ifor each i. The
definition of a A-net of a set of block bases is obvious. If TIis a set of *-pairs,
we write HA for the set of pairs (A, Z) such that there is a block basis B with
d(A, B) < A and (B, Z) E H. If xi < ... < xn, then we shall write (xi, . X..,xn)
for the subspace generated by xi,..., Xn. Given a sequence (xi,... ,Xn) C E,
we shall write Ef(xi,... , Xn) for the set of all sequences (Yli,..., Yk) E Ef such
that every yi belongs to the subspace (xi,..., xn).
LEMMA2.2. Let A1, A2,... be a sequence of positive real sequences and

let IIi, II2,... be a sequence of sets of *-pairs satisfying the following conditions:
(i) for every pair (A, Z) and every n there is Z' C Z such that (A, Z') E Hn;
(ii) if (A, Z) E IIn and Z' C Z then (A, Z') Ec n.
Then there exists a subspace Y C X such that (A, Z) E (1Hn)A for every pair
(A, Z) such that A is of length at least n and both A and Z are subsets of Y.

Proof. Choose a block basis yi, Y2, ... and a sequence of block subspaces
X = Yo D Yi D Y2 D ... inductively as follows. Once we have chosen
Y1i,...,n-l and Yi, ...,Yn-i, let yn E Yn-i be arbitrary (except that
(yi,..., yn) E Ef) and let A1,..., AN be a An-net of the set Ef(yi,..., yn).
By property (i) we can pick a sequence of subspaces
Yn-1 D Vll D ... D Vn D V21 D ... D V2n D ... D VNI D... D VNn
such that (At, Vij) e IIj for every 1 < i < N and 1 < j < n. Let Yn = VNn.
By property (ii) we have (Ai, Z) E IIj for every i, j and every Z C Yn. Since
(An)N= is a An-net, we find that (A, Z) E (HII)Al for< every j n and every
pair (A, Z) such that A E E(yi,..., Yn) and Z C Yn.
Let Y = (Y, Y2,...) and suppose that (A, Z) is a *-pair inside Y, with
A of length m > n. Then there exists k > n such that A C (Y1,... ,Yk) and
Z C (Yk+i,Yk+2, *) C Yk. By our construction therefore, (A,Z) E (ln)An.
Hence, Y will do. D2
We shall often use specializations of Lemma 2.2. It may be convenient

for the reader if we state them separately. Let us define a singleton *-pair to
be a pair (x, Z) where x is a nonzero vector of norm at most 1, Z is a block
subspace and x < z for every z E Z (or in other words a *-pair (A, Z) for
which A is a singleton). Given 6 > 0 and a set II of singleton *-pairs, write
Ha for the set of *-pairs (x, Z) for which there exists x' such that d(x, x') < 5,
x' < z for every z E Z and (x', Z) e I.
COROLLARY 2.3. Let 6 > 0 and let H be a set of singleton *-pairs satis-
fying the following conditions:
(i) for every pair (y, Z) there is Z' C Z such that (y, Z') HI;
(ii) if (y, Z) E H and Z' C Z then (y, Z') E II.
Then there exists a subspace Y C X such that every pair (y, Z) with y E Y
and Z C Y belongs to Ha.
Proof. Apply Lemma 2.2 with the following choices for the Hi and Ai. Let
Hi be the set of all *-pairs (A, Z) such that if A is a singleton, then (A, Z) E II.
(Thus, every pair (A, Z) for which A is not a singleton belongs to HII.) If i > 1,
then let Hi be the set of all *-pairs. Let all Ai be the sequence (6, 1, 1, 1,...)
(but all that matters is that the first term of Ai should be 6). l
COROLLARY2.4. Let Al, A2,... and II1, II2,... be as in Lemma 2.2 but
satisfying the following additional condition:
(iii) if (A, Z) E Hn then A has length n.

804 T. GOWERS
W.
everypair
Then there exists a subspaceY c X such that (A, Z) E (rln)/\r, for
(A,Z) such that A and Z are subsets of Y and A has lengthn.
iS the set of >R-
Proof. ApplyLemma2.2 replacingeachHn by Hn) whereTI/n
Corollary2.3,
pairs(A, Z) suchthat if A has length n then (A, Z) E un. (As in
n
A has any other length then (A, Z) belongsto Hn.)
if
3. A Banach-space dichotomy
detail how it
In this section we shall prove Theorem 2.1 and show in
Banach's problem
impliesTheorem 1.4. This will complete the solution of
us make two more
onhomogeneousspaces. Beforestating the next result, let
Ef (X),
definitions. If X is a Banachspace, Y is a blocksubspaceof X and ff c
of Y contains a
weshall say that ff is large for Y if every block subspace
strategy for
sequencein , and strategicallylarpefor Y if P has a winning
say that ff is large
thegame cr[Y]. More generally,if (A,Y) is a *-pair, we
B E Ef such that
for[A;Y] if every block subspaceof Y containsa sequence
strategy for
(A,B) E , and strategicallylarpefor [A;Y] if P has a winning
Y] and only if
if
thegame v[A;Y]. Note that ff is (strategically)large for [A;
cr[A;Y] is (strategically)large for Y.
monotone basis.
THEOREM3.1. Let X be a Banach space with a given
Lete}= (b3n)°n°=l and A = ((5n)°n°=l be sequencesof positive real numbersszbch
then X has a block
that2 EtN ai % SN for every N. If v_e is larpefor X,
subspaceY such that v2a is strategicallylargefor Y.
is false. Then
Proof. Supposethat cr c Ef is a set for which the result
other hand, cr2/\
a_<3is large for X, so in particularCJis large for X; on the
set of sequences
is not strategicallylarge for any subspaceof X. Let p be the
is a proper subspace
(x1,. . ., xn) E ff suchthat if Y1< . . . < y, and (Y1,. . ., y,)
p is still large for
of (x1, ,xn)7 then (Y1,...,y) f v. It is easy to see that
of X.
X and that P2a is not strategicallylarge for any subspace
- An =
For each n ) O Let An = (61) . . * >(5n: 0: 0: * * *) and let rn-2A
(¢6l7***7(6n:2/6n+l72(6n+2v*. .) * We now constructsequencesx1, x2, . . . and X =
n:
Xo 2 X1 2 X2 2 ... with the followingpropertiesfor every
(i) Zn E Xn-1;
(ii)PAnis large for [x1,. . ., Xn;Xn];
Z c Xn.
(iii) Prn is not strategicallylarge for any [x1,. . ., Xn;Z] with
Xo but PrO
The inductionstarts with the space Xo) since p/\Ois large for
Xrland X1, , Xn7
is not large for any subspaceof Xo. Havingfound xl, . . .,
Then for every
supposewe cannotfind suitablecandidatesfor Xn+l and Xnal.

x E Xn and every subspaceY of Xn we can find a subspaceZ c Y such that

either
(a) Pi\n+1n [Xl) . ., Xn,x; z] is empty
or
(b) Prn+] is strategicallylarge for [x1, , xn)x; Z].
Let TIbe the set of singleton *-pairs (x, Z) with x E Xn and Z c Xn
such that either (a) or (b) holds. We shall apply Corollary2.3. Condi-
tion (ii) of this corollary is obvious, and we have just shown that condi-
tion (i) holds as well. Applying the corollary with a = dr+1) we obtain
a subspace Y c Xn+1 such that, for every *-pair (y,Z) with y E Y and
Z c Y, there exists x with IIY-XlI S Dn+1 such that either (a) or (b) holds
for the pair (x, Z) . Since IIY-X|l% Dn+1 ) the first alternativeimplies that
P/\nn Lxl:...,zn,y;z] is empty and the second implies that Prn is strategi-
cally large for [x1,. . ., xn, y; Z]. In particular,one of these two conclusionsis
true when Z = Y. (Thereis a small technicalpoint which is that (y, Y) is not
a *-pair. However,the conclusioncan be seen by consideringfor each y E Y
the *-pair (y, Z), where Z = {z E Y: y < z}.)
The set of y such that Pi\nn [xl...,zn,y;r] is empty cannot contain a
subspaceZ of Y since then P\nn [xl: . . ., xn;z] is empty,whichcontradicts(ii)
of ourinductivehypothesis. Therefore,the set of y suchthat prn is strategically
largefor [x1,. . ., xn, y; Y] is largefor Y. But this gives P a winningstrategyfor
the game prn EX1) . . ., Xn; Y] and contradicts(iii) of our inductivehypothesis.
We now claim that the subspacegeneratedby (X,7,)n=l has empty inter-
section with v_z3. Indeed, let (Z1,. . . ,z,) E Ef be a sequencecontainedin
the subspace (x1,...,xn). By property (ii), the sequence(x1,...,xn) can be
extendedto a sequence(x1, . . ., xm) in pa . Choose (x1, . . ., ' ) E p such that
llxi-x' 11 < di for every i S rn. Let (z1, . . . , zk) be the corresponding pertur-
bation of (z1, . . ., z), and notice that the minimalityconditionsatisfiedby p
ensuresthat (Z1,. . . ) Zzc)¢ v. Our choiceof A guaranteesthat ||zj - zji| < (3j
for every j < #. But this proves that (Z1,...,Z) , _43 and our claim is
proved. This contradictsthe assumptionthat (r_zg3 was large. n
It is not hard to see that the above result is equivalentto Theorem2.1.
Proof of Theorem2.1. Let T = C/\/2.The assumptionof the theoremand

tell us that 7-/2 is large for X. Therefore,applying
the fact that ff c Or-/\/2
Theorem 3.1 to T (with (3 replacedby A/2) we can find some positive real
sequencer S Al2 such that Tr is strategicallylarge for some subspaceof X.
But Tr C cra, so the result is proved. C:

806 W. T. GOWERS
Beforewe apply Theorem2.1, let us introducetwo furtherdefinitions.We

shall say that a blockbasis (xn)°n°=l
is C-unconditionalif it generatesa subspace
which contains no C-conditionalfinite sequences of blocks. (C-conditional
sequencesweredefinedjust afterLemma1.7.) We shallsay that a Banachspace
X is C-hereditarilyindecomposableif for every pair Y, Z of block subspaces
of X we can find y C Y and z E Z such that IIY+ZII, CIIY-Zll This
is equivalentto condition (ii) of Lemma 1.7 for that particularvalue of C.
Lemma 1.7 asserts that X is hereditarilyindecomposableif and only if it is
C-hereditarilyindecomposablefor every C.
COROLLARY 3.2. Let X be a Banach space. Then either X contains a

C-anconditionalblockbasis or for every E > O it has a (C-£)-hereditarily
indecomposableblocksubspace.
Proof. Let ff be the set of all sequences (x1,. . . ,zn) E Ef that are
C-conditionaland contain at least one vector xi of norm 1. If X contains
no C-unconditionalblock basis, then cris large, since every subspacecontains
a C-conditionalsequence(Y1: . .: Yn):and if we dividethis by the largestvalue
of [lYillwe obtain a sequencein v. Therefore,by Theorem2.1, for any A > 0
we can find a blocksubspaceW of X such that , is strategicallylargefor W.
Let us choose A such that 1°°-1 di = r1for some r7> Osatisfyingthe inequality
(1 + 271)-l(C - 277)) C - s.
Now let Y and Z be arbitraryblock subspacesof W and considerthe
strategy for S, the subspaceplayer,of alternatingY and Z. Since az\ is strate-
gically largefor W, P can defeatthis strategy,whichmeansthat P can choose
a sequence (g1vZ1:g2:Z2: vYn:Zn)e va such that Yi belongs to Y and Zi
belongs to Z for every i. (An unimportanttechnicalpoint is that P's strat-
egy may succeed after an odd number of moves. One can either alter the
statement of Lemma1.7 or let P choose a sufficientlysmall Znto finish. The
second approachis possiblebecauseof the strict inequalityin the definitionof
C-conditionalsequences.) We can then find (y1,Z1, , Yn:Zn)E ff such that
IIYi-Yill S 02i-l and ||zi-Zi'|| S a2i for every i % n. Since the basis of X is
monotone, the norms ||n=l(y' + z') \ and t|n=l(y' _ z') are both at ]east
1/2. We thereforeknow that
n n
lE(Y
t=1
+Zi/) t> C lE(Yi
i=l
- Z/)lI ) C/2
By the triangleinequalityand our choice of A, we know that

n n
iLf ) Iti i,ll

and
n n
i=1
(Yi- zi)i I(Y
T - Zi,)+n
=1
It follows from our choice of r7that

n n
(yi +zi) >(C- ) (yi - zi)

i=1 i=1
Since Y and Z were arbitrary subspaces of W, we have shown that W is
(C - e)-hereditarily indecomposable. D
A simple diagonalization now completes the proof of Theorem 1.4.
Proof of Theorem 1.4. If X has no subspace with an unconditional basis,

then Lemma 1.6 tells us that, for every C, every block subspace of X contains a
C-conditional block sequence. By Corollary 3.2 we can therefore find a nested
sequence W1 D W2 D ... of block subspaces of X such that, for every n, the
subspace Wn is n-hereditarily indecomposable. Let (wn),=l be a block basis
of X such that WnE Wn for every n and let W be the subspace generated by
(W7n)=1. We claim that W is hereditarily indecomposable. To see this, let Y
and Z be arbitrary block subspaces of W and let C be any real number. Choose
a positive integer n > C. Then Y n Wn and Z n Wn are infinite-dimensional.
Since Wn is n-hereditarily indecomposable, we can find a sequence satisfying
condition (ii) of Lemma 1.7 for this particular C. But C was arbitrary, so the
condition holds in general. Lemma 1.7 therefore implies that W is hereditarily
indecomposable, as claimed. C]
4. Definitions and preliminary results for infinite sequences
We begin this section with a brief discussion of the connections between

our results so far and known results of infinite Ramsey theory. The statement
of Theorem 2.1 is strongly reminiscent of a result in infinite Ramsey theory
due to Nash-Williams [N-W], which says the following. Let Nw be the set of
all infinite subsets of N. If A is an open subset of N' (in the product topology)
then either A or its complement contains all infinite subsets of some X E N'.
A set A with this property is called a Ramsey set.
It is easy to check that Nash-Williams's result is equivalent to the following
statement. Let N<Wbe the set of all finite subsets of N and let A C N<'. Then
there is an infinite subset X of N such that either no finite subset of X is in A or
for every infinite subset Y of X there exists n such that Y n {1, 2,..., n} E A.
Notice that if the first alternative does not hold, then A is large in an obvious
sense. So Nash-Williams's theorem asserts that if A is a large subset of N<W

808 W. T. GOWERS
then there is an infinite subset X of N for which A has a much stronger largeness
property. This formulation makes the resemblance with Theorem 2.1 very
clear.
Nash-Williams's theorem was extended to all Borel sets by Galvin and
Prikry [GP]. A combinatorial lemma of theirs inspired our proof of Theo-
rem 2.1. Silver [S] proved that all analytic sets are Ramsey, and Mathias
proved that in a model constructed by Solovay all sets are Ramsey. From
these proofs there emerged a natural strengthening of the Ramsey property,
which was shown by Ellentuck [E] to be equivalent to the property of Baire
in a certain topology. For proofs of the Galvin-Prikry lemma and Ellentuck's
theorem, see [Bo] and for further results in this direction, see [M].
The main aim of the next two sections is to extend Theorem 2.1 in a
similar way. However, as we will explain later (see the appendix), the obvious
analogue of Ellentuck's characterization is false, so we must be satisfied with
a result more like Silver's. In order to state it, we shall need some definitions
concerning infinite sequences and sets of infinite sequences.
The obvious way to extend our earlier results would be to redefine Y(X)
as the set of all infinite sequences xi < x2 < ... of nonzero vectors in X
such that llxnll < 1 for every n. Instead, for technical reasons, our defini-
tion will be slightly different: let S(X) be the set of all sequences of pairs
(xl, A1), (x2, A2),... where x < x2 < ... are vectors of norm 1, and Ai, A2,...
are real numbers in the interval [0,1]. Of course, we can usually identify the
pair (x, A) with the vector Ax. The main difference between our definition and
the obvious definition is that if x -/ y, then we distinguish between the pairs
(x, 0) and (y, 0). In order to save writing, we shall usually use a single letter to
denote one of these pairs, unless it is important to be careful. We shall refer to
elements of E as a block bases. Sometimes we shall discuss finite block bases.
Let us now redefine Ef to be the set of finite sequences ((xi, A1),... , (Xn,An))
with the (xi, Ai) as above and xl < ... < xn. The support of a pair (Xn, An)
is defined to be the support of the vector Xn. The subspace generated by (the
pairs) (xi, A (x2,A2), ... is defined to be the (block) subspace generated by
the vectors x, 2,.... -
Several definitions to do with finite sequences can be easily adapted for
infinite sequences. For example, if X is a Banach space, Y is a block subspace of
X and A = (yl,..., ym) is a finite sequence of blocks (that is, pairs of the above
kind) then [A;Y] is now defined as the set of all infinite sequences (Zn)1 E S
such that zi = Yi for i < m and zi E Y for i > m. Again we denote this by [Y]
if A is the null sequence. If a is a subset of E, then we write ar[A;Y] for the
set of all sequences (wn) Z(Y) such that (yl,...,
=l E ym, w,w2,...) E u. We
say that a is large for [A;Y] if every block subspace of Y contains a sequence
in c[A; Y]. Once again, a is large for [A;Y] if and only if r[A;Y] is large for Y.

Given a subset crof , we can define an infinitegame just as we did for

finite sequences. The only differenceis that P's aim is to producean infinite
sequencethat lies in a, so the game alwayslasts for ever (althoughone player
may have a guaranteedwin after finite time regardlessof future moves). We
say that ff is strategicallylarpefor [A;Y] if P has a winning strategy for the
set cn[A;Y] when all of S's moves are requiredto be subspacesof Y.
Finally, we consider two topologies orl S. Most of the time it will be
convenientto take as basic open sets all sets of the form {(Xn)°c=l xrs =
Ynfor 1 S n S N}. In other words, we put the discrete topology on X and
then take the producttopologyon S(X) . However,for our mainresultwe take
a differenttopology. First, definea metricon S(X) x [O,1] by d((x, A),(y, ,)) =
llx-Yll + 1S-,aI (S(X) is the unit sphereof X. Roughlyspeaking,we have
made a hole wherezero used to be.) Fromthis we derivea topology on X and
hence a differentproduct topology on S(X). A basic open neighbourhoodof
the sequence((cn)An))°°=l is now a set of the form
{((Ynvlln))°°=l:
||Yi-xi|| + |,tli-Ail < si for every i < N}
for some positive integer N and positive real sequence (61,...,EN). The ad-
vantage of the first topology is that it is less messy to talk about open and
closed sets, and the advantageof the second is that it makes S(X) a Polish
space (see Lemma4.2 below), which is more convenientfor talking about an-
alytic sets. (It is in orderto make E completemetrizablethat we allow "zero
vectors"in a block basis.) However,a perturbationis involvedin our result,
so the distinctionbetweenthe two topologiesis not at all importantfor appli-
cations. We shall refer to D-open and N-open sets and so on (for "discrete"
and "norm")when it is not otherwiseclear which topology is meant.
We have not yet defined perturbationsof infinite sequences. If ff is a
subset of S(X) and A > O is an infinite sequenceof positive real numbers,
then let v\ denote the set of all sequences (xn)n_l such that there exists a
sequence(yn)n=l E ff with d(28n, Yn)S dnfor every n. (Note that the Scnand
Ynare elementsof S(X) x [0,1] and d is the metric definedabove.) We now
have enoughnotation to state the main theoremof this and the next section.
THEOREM 4.1. Let X be a Banach space, let ff c S(X) be N-analytic
and larpefor X and let A > O. Then thereis a subspaceY of X such that (J2
is strategicallylarpefor Y.
Let us give a definitionwhich will be useful for the rest of the paper.
Definition. A set ff c S(X) is weaklyRamseyif for every A > Othere is
a subspaceY C X such that either ff n [V] is empty or v\ is strategicallylarge
for Y.

810 W. T. GOWERS
Thus,Theorem4.1 statesthat N-analyticsets areweaklyRamsey,anda

set ff c S(X) is weaklyRamseyif the conclusionof Theorem4.1 holdsforv.
Forthe rest of this sectionwe shallprovethat D-opensets andN-closed
sets are weaklyRamsey,and then we shall provea few lemmaswhichare
neededfor Theorem4.1. First, we checkthat Theorem2.1 still holdsnow
that we have redefinedEf. One way of doingthis is simplyto checkthat
the proofwe gaveis still validunderthe newinterpretation. However,it may
reassurethe readerto see that one can deducethe resultforthe newEf from
Theorem2.1.
LEMMA4.2. LetX bea Banachspace,let sr beanysubsetof Ef (X) and
let A be a sequenceof positiverealnumbers.ThenX has a subspaceY such
thateither[Y] = 0 or P has a winningstrategyfor thegameva [Y].
Proof.Definep to be the set of all sequences
((1 + >1)t1/2, * * ., (1 + An)Xn/2)
suchthat the sequence

((z1 7 A1 ) 7 * * * v (Xn v An ))
is in v. It followsfromthe triangleinequalitythat
11(1 + A)X-(1 + H)Y|| ) maX{|S-81, ||x-Yll-1>-81} .
Therefore,applyingTheorem2.1 to p with A replacedby A/4 we obtainthe

desiredconclusion. g
THEOREM4.3. 4.3. All D-opensets areweaAcly

Ramsey.
Proof.This is just a reformulation of Lemma4.2. Let ff be a D-open
set. Let ¢f be the set of finitesequences(X1: . . . ) Xn) E Ef suchthat all their
extensionsto sequencesin S areelementsof v. Sincea is a D-openset, every
sequencein ff has an initialsegmentbelongingto ¢ff. If cris largefor X, it
followsthat ¢f is largefor X as well. Therefore,by Lemma4.2, for every
A > Othereis a subspaceY of X suchthat (ff)X\is strategically largeforY.
If P usesthis strategyandplaysforever,then the resultinginfinitesequence
belongsto vz,. Thus,Crz\ is alsostrategically
largeforY. Thisshowsthat ff is
weaklyRamsey,as claimed. C:
Althoughit is weakerthanourmainresultandnot neededforthe proof,
we shallnowshowthat N-closedsets are weaklyRamsey,sincethe proofis
quiteshort,andwe shallapplythis weakerresultdirectlyin §7. As a matter
of fact,we proveslightlymore,by defininga finertopology,whichweEshall
call
the D-*-topology,verysimilarto the so-called*-topology,or
Mathiastopology,

INFINITERAMSEY THEOREM
AN
811
show that D-*-closed sets are

the infinite subsets of N. We shall then
on
sets are genuinelyweakly
to being weakly Ramsey and that N-closed
close
to obtain a second Banach
Ramsey. In Section 7, we shall apply this result
spacedichotomy.
by the sets [A;Y] defined
The D-*-topology is the topology generated
basic open neighbourhoodof a se-
the beginningof this section. Thus, a
at
where A is an initial segment
quence (xn)°°=lis some set of sequences[A;Y],
xi for every i > m. (Simi-
(x1,...,xm)and Y is a block subspacecontaining
basic open sets of the following
one can definethe N-*-topologyto have
larly,
that llxi-Yi 11< d for i ( n and
form:the set of all sequences(Xn)n_l E S such
E Y for i > n, where d > O,Y1
< < Ynand Y is an iniinite-dimensional
Xi
Ramseyif, whenevercris large
block subspace.) A set ff is completelyweakly
Z of Y such that P has a winning
in[A;Y] and A > O, there is a subspace
analyticset is weaklyRam-
strategy for the game aa [A;Z]. Note that if every
Ramsey. We shall show in
sey,then every analytic set is completelyweakly
of Ellentuck'stheorem, that a set is
theappendixthat the natural analogue
set in the N-*-topology,is
completely weaklyRamseyif and only if it is a Baire
of two completelyweaklyRamsey
false.In fact, we show that the intersection
setsneed not be completelyweaklyRamsey.
Galvin and Prikry by saying
Given A and Y as above, we shall follow
and that Y rejectsA (fromCJ)if no
thatY acceptsA (into a) if [A;Y] c ,
Y rejects the null sequencefrom the
subspaceZ c Y accepts A. Sayingthat block subspace Z c Y
complement of ff is equivalentto saying that every
that ff is large in Y.
containsa sequencein (x,or in other words
times. However,strictly
We are about to apply Corollary2.4 several
modifiedblocks (x, A). Such
speakingone needs a differentresult concerning
way, so we shall simply apply
a result can easily be proved in an identical
as pairs (x, A) E S(X) x [O,1J.
Corollary2.4, interpretingthe vectorsthere
let So > A1 >
LEMMA4.4. Let ff c E be a set of infinite sequencesand
a subspace
2 > ... be a seqtbence of positive real sequences. Then there is
(Y1, , Yk) E Sf(Y)v eitherY
Y c X such that, for any k and any seqllence
Y1,. . ., Yk from C/\k .
acceptsY1,. . , Yk into CJ<k_l or Y rejects
+ /\;c) and let (3k = 2 (S,-1-/\k).
Let
Proof. Foreach k let rk-2 (k-l
is a sequenceof length k and Z either
fI, be the set of pairs (A, Z) such that A
It is easy to see that the conditions
accepts A into crrkor rejectsA from vrk.
a subspacesuch that, for every k and
of Corollary2.4 hold, so we can pass to
(H#)@k.Given any pair (A, Z) with
every A of size k, every pair (A, Z) is in
d(A, A') < #) and (A', Z) E uk. If Z
A of length k, choose A' E Ef such that
and if Z rejectsA' fromvrk
accepts A' into vrk, then Z accepts A into
(7Z\k-l
the result. O
then Z rejects A from ,\k. This proves

812 W. T. GOWERS
THEOREM 4.5. Let a be a D-*-closed subset of , let T be the complement

of a and let A > 0. Then there exists a subspace Y C X such that either P
has a winning strategy for the game r[Y] or [Y] c TA.
Proof. Suppose there is no subspace Y such that [Y] C rT. Then X

rejects 0 from TA. For k = 0,1, 2,... let Ak = 2-kA. By Lemma 4.4, we may
assume that, for every sequence (xi,..., xn), X either accepts (xi,..., Xn) into
TAn-I or rejects it from Tr/A.
Now let P play the following strategy: for the nth move, choose Xn such
that X rejects (xi,... ,Xn) from TAn. We shall show both that this is possible
and that it is a winning strategy.
Suppose then that P has played the strategy successfully for n moves.
We know that X rejects (xi, ... ., n) from Tr/N. Now let S play the subspace
Xn+i. If P cannot continue the strategy, then, for every x E Xn+i, X does
not reject (xi,..., Xn,X) from xTAn+l.But because of our assumption about
X, this implies that X accepts (xi,... ,Xn7,x) into -TA for every x c Xn+i. It
follows that Xn+i accepts (xi,... , -xn) into TA/\. But this contradicts the fact
that X rejects (x1,... , Xn) from TAn.
We have shown that the strategy is possible. Now let (xi,x2,...) be a
sequence produced by B when using it. Then we certainly know that, for every
n, X rejects (x1i,... , xn) from T. That means that, given any subspace Y C X,
[xi,..., Xn; Y] n cr# 0. Since cr is *-closed, it follows that (xi, x2,...) E a. O
Because of the way we defined weakly Ramsey sets, Theorem 4.5 does not
imply that D-*-closed sets are weakly Ramsey, although this is true up to an
arbitrarily small perturbation. Moreover, if aois D-*-closed, it does not follow
that crAis D-*-closed. Although it is not important for applications, we shall
now show that N-closed sets are weakly Ramsey, using the following simple
result.
LEMMA4.6. Let ar be N-closed and let A > 0. Then c/x is also N-closed.
Proof. Since orA- (/A/2)A/2 we can assume that every 6n is at most 1/2.
We shall use bold face letters to denote sequences in E. Let x belong to the
closure of c(A. Then there exist sequences ym E ruAconverging pointwise to
x. For every m we can find a sequence z'm E a such that d(ym, zM) < A.
For fixed n, the vectors Zn (or strictly speaking elements of S(X) x [0,1])
have bounded support, as otherwise d(z+i, x4+1) would be at least 2/3 for
sufficiently large m. Therefore we can find a sequence ml, m2,... such that zTm
converges pointwise, to z, say. Since aois N-closed, z must be in U. Because
Zm -> z and ym -_ x pointwise we must have d(zn, Xn) < 6n for every n, which
proves that x C oA, as required. C
COROLLARY4.7. Every N-closed set is weakly Ramsey.

RAMSEY THEOREM
INFINITE
AN
by 813
> 0. Then (Tz\/2 iS N-closed,
let A
Proof.Let ff be N-closed and ApplyingTheorem4.5
to 5</2 with
and in particular D-*-closed. O
Lemma-4.6, result.
A we obtain the
replacing
Z\/2
used to prove
is devoted to lemrnaswhich will be largefor
Therest of this section :Ehas the propertythat it is not
If sr C
result)Theorem4.1.
Inain
our nowhere and-
larpe, otherwisethat it is
that it is
Y], then we shall say
[A;
any Ramsey theorems,it is shown
that being
In the usual infinite here (see
somewhere.
larpe being *-nowhere dense. That is false
same as states
large is the
nowhere have the following usefullemma,which
We do: however, under countable
appendix).
the
sets are, up to perturbation,closed
the nowherelarge
that
unions.
let An > Ofor every

n andsupposethat
LEMMA 4 8 Letff = U°nl Sn)
nowherelarpe.
is larpe
nowhere for everyn. Thena is
(n)an for is not large
We must show that
a
W) be any pair. A is a
Proof.Let (B, the set of pairs (A, Z) in W such that
W].For each n let IIn
be (We mean by
[B; n [(B7A);Z] is empty.
of length exactly n and B.) Our hypothesisimplies that the sets
(5n)An
sequence
of A and find a subspace
A) the concatenation
(B, 2.4. We can therefore
for Corollary A of length
nsatisfy the conditions for every pair (A, Z) in Y with
YW C such that (A)Z)
E (IIn)Z\n
A);Z] n -0 for every such
this implies that [(B,
It is easy to see that the end of
n. C of length n in Y can be addedto
pairThis says that no sequence in Cn It followsthat
[B;Y]n ff-0
in Y to a sequence E an then the
Bandthen extended
D S(Y) and n such that (B:D)
because if we could find
E
C contradicting what we havejust
us a sequence
n terms of D wouldgive
first arbitrary, we have shownthat cr
is nowhere
Since B and W were O
established.
large.
of Lemma4.8 (Corollary
reasons we need a slight extension
Fortechnical
4.10below)*
[B;Y] and A > thenfor everym E N 07
LEMMA 4.9. If ff is larpefor in Y and a subspaceZ c Y suchthat

A of lengthm
thereexzstsa seq?bence
va is larpetn [(B)A);Z]
the set
to the assertion. Let IIm be
Proof.Supposeff is a counterexample A);Z] =
A is a sequence of lengthm and , n [(B, the
that assumptionimpliesthat
of pairs(A,Z) in Y such of all pairs. Our
be the set find a subspace
0. For n + m let [In Corollary 2.4. Thereforewe can m,
of
Hn satisfy the
conditions
every pair (A, Z) in W with A of length
Z) E um for
W C Y such that (A,

WGOWERS
T.
814
says that
that a n l(B, A); Z] = 0 for every such pair. But this
implies
which
be put on to the end of B and extended
sequencefrom W of length m can
no O
a sequence in . Therefore cr is not large for [B;Y].
intoW
ff = U°C-lan7 A1,A2,
COROLLARY 4.10. If a is larpe somewhere,
and k1,ks,... is a seq?>ence of integers, then there
positive
are real sequences of size at
[A;Z] withA a seq?bence
n szuchthat (n)An is larpefor some
escists
kn
least
n and a pair (EX, Y) such that C/\n/2 iS
Proof. By Lemma4.8 we can find A = An/2)
for [B; Y]. Then by Lemma 4.9 with m = kn7ff = ¢An/2and
large knand a subspaceZ C Y such that
can find a sequenceC E Ef of length
we at least
large in [B7 C; Z]. The length of the sequence(B, C) is certainly
is
cr\n 12
so we are done.
kn
spirit to Lemma4.4.
The next lemma is very similarin
let be a sequenceof
4.11. Let X be a Banach space,
¢lncr2:-**
LEMMA
of S(X) and let be a sequenceof positive real seqtuences.
subsets 1:A27***
n and everyseqllence
exists a subspace Y c X such that, for every
Then there
eitherCJn n [>4;r] is emptyor (n)an is larpe
AE Ef (Y) of lengthat least n,
Jor[A;Y]
that either
Proof. For each n let 77, be the set of all *-pairs(A, Z) such
to
n [A; Z] is empty or (°n)/\n/2 is large for [A;Z]. It is very easy
(5n)/\n/2 of Lemma2.2. Therefore,there
check that the sets Hnsatisfy the conditions at least
*-pair (A, Z) in Y with A of length
isa subspaceY such that every An/2
to This says that there is some A' with d(A, A') <
nbelongs
[A',Z] = 0 then ¢nn [A;Z] = 0
(tIn)An/2
If (C7n)An/2
with(A', Z) E fIn. If (5n)/\n/2 n subset of Z. We can choose an
[A'; Z], then let W be an arbitrary
islarge for . Then
that A < B and (A', B) E (n)Z\n/2
infinitesequenceB E S(W) such is large for [A;Z]. The proof is
(n)/\n. This shows that (n)an
(A,B) E O
complete.
of a more technicalnature.
We end this section with a lemma
= U°O=1 let A > O be a real
4.12. Let cr c S be larpe,let a
¢n)
LEMMA Y C X let p(Y)

be a function from N to N. For any subspace
sequence and let ?t f (n)}
such that there exists n with max{n:
be the set of sequencesB:E Ef (Y) a subspace
largefor [B;Y]. Then there exists
at most the size of B and (<7n)
largefor Y.
Y c X such that p(Y) is strategically
Proof. By Lemma 4.11 with each
an replacedby (CT,)A/3and each An
every
A/3, we can find a subspace Z such that, for every n E N and
equal to is empty or
least n: either (5n)A/3 n [A;Z]
sequenceA E Ef(Z) of length at
(crn)2/\/3is large in [A;Z].

Now a is large in Z and therefore in any subspace of Z. Therefore, by

Corollary 4.10, for any subspace W c Z there exist n, a subspace V c W and a
sequence B E Ef (W) of length at least max{n, b(n)} such that (rn)A/3 is large
in [B; V]. In particular, (crn)A/3 f [B; Z] is nonempty, so by our construction
of Z we know that (Crn)2A/3 is large for [B; Z].
Define r to be the set of sequences B E Ef (Z) such that there exists
n with max{n, /(n)} at most the length of B and (rn)2A/3 large for [B;Z].
We have just demonstrated that r is large for Z. By Theorem 2.1 there is a
subspace Y C Z such that 7TA/3is strategically large for Y. The result follows,
since rA/3 n Sf (Y) c p(Y). E
5. Analytic sets are weakly Ramsey
In preparation for the proof of Theorem 4.1, we recall the definition of

and a basic fact about analytic sets. For a proof, and for general background
in descriptive set theory, we recommend Chapter 7 of Jech's excellent book [J].
(This is the last chapter but the relevant part is very clear and more or less
independent of the rest of the book.)
The Baire space Af is the set NN with the product topology. Using con-
tinued fractions, one finds that KAis homeomorphic to the set of irrational
numbers. If X is a topological space, an analytic subset of X is the continuous
image of a Borel subset. If X is a Polish space, this can be shown to be equiv-
alent to saying that the subset is a continuous image of M. This equivalent
definition is usually the easiest to use. It is clear that every Borel set is ana-
lytic. Souslin's theorem (which we do not need here) states that two disjoint
analytic sets can be contained in two disjoint Borel sets, which implies that a
set is Borel if and only if it and its complement are analytic.
Before giving the proof of Theorem 4.1, let us quickly verify that the set
we shall be considering is indeed a Polish space.
LEMMA5.1. Let X be a Banach space. Then E(X) is a Polish space in
the N-topology.
Proof. We must find a complete metric which gives rise to the N-topology
on E(X). Such a metric is given by
d ((Xn, An) i, (yn ln), = ) =maxn-ld ((xn, An), (Yn,i,n))
where on the right-hand side we have taken the metric defined earlier on S(X) x
[0,1]. Let Sm be the sequence ((Xmn, Amn)) and suppose that the sequence
(Sm)m=li is Cauchy in the above metric. The only small point to note is that
the supports of the vectors Xmn are bounded for fixed n, as otherwise the

816 W. T. GOWERS
vectors Xm,n+l would not form a Cauchy sequence in X. It is therefore easy

to see that the sequences Sm converge pointwise and hence in the metric, and
that the limit is in E(X). It is also simple to check that this metric gives rise
to the N-topology on E(X). O
Proof of Theorem 4.1. Since E is a Polish space (in the N-topology) we

can use the equivalent definition of analytic sets. Suppose therefore that a C S
and that there is a continuous map f : JA -+ E such that f(Af) = c. Given
a finite sequence ni,... , nk of integers, let u(ni,...,nk denote the image under f
of the set of sequences in Af starting i, .. , nk. Note that, for any A > 0,
(o7nl,...,nk)A = (nl,...,nk+l)A-
Unk+l-
Let 0 = Ao, A1,A2,... be a (pointwise) strictly increasing sequence of
real sequences less than A. For n > 0 let Fn be the sequence (An+l - An)
and for i = 1, 2, 3 let An,i = An + iFn.
Let b be a bijection between the set of finite sequences of integers and N.
By Lemma 4.11 we can find a subspace W of X such that, for every (ni,..., nk)
and every sequence A E Ef(W) of length at least b(ni,... ,nk), either
(Uni,...,nk)Ak, n0[A; W] is empty or (onl,...,nk)Ak,2 is large for [A; W].
For every (ni,... , nk) and every pair (A, Z) in W, define Pnl,...,nk[A; Z]
to be the set of all sequences B E Ef(Z) such that there exists nk+i with
0(ni, .. ., nk+l) at most the length of B and (UC ,...,nk+l)Ak,3 large for [A, B; Z].
Let Inl...,nk be the set of pairs (A, Z) such that if A has length at least
Oq(nl,..., nk) then either
(a) (ln1,...,nk)Ak,, n [A; Z] is empty
or
(b) pn,...,nk [A; Z] is strategically large for Z.

We shall apply Lemma 2.2 to the sets IInl,...,n. However, unlike with
our previous diagonalization arguments, it will not be trivial to verify that the
relevant conditions are satisfied. Instead, we use Lemma 4.12. Condition (ii)
of Lemma 2.2 is of course not a problem. As for condition (i), let (A, Z) be
any pair and suppose that the length of A is at least O(n, ... , nk). If (a) is
not true for the pair (A, Z), then by our construction of the subspace W we
know that (0nl,...,nk)k,2 is large for [A;W] and hence for [A;Z]. Now
00
(Unl,...,nk)Ak2 = U (0nl,..-,nk+l) k,2

nk+l=l
Lemma 4.12 applied to this union with A = 'k implies that there exists a
subspace Z' C Z such that Pnh,...,nk [A; Z'] is strategically large for Z', which
tells us that (b) is true of the pair (A, Z'). This shows that condition (ii) of
Lemma 2.2 is indeed satisfied.

RAMSEY THEOREM
INFINITE
AN
817
Y C W such that everypair
a subspace
Using that lemma,we can find +(nl..*nk) belongsto (llnl>..>nk)rk
A of length at least Z]
in Y with
(A,Z) sets Z] by defining tnl,...,nk [A;
of the Pnl,...,nk [A;

with
us perturbations B E Ef (Z) suchthat thereexists nk+l
define
Let
thebeset of all sequences of B and
to largefor[A:B; Z]
. at mostthe
. ., nk+l) length (5nl,...,nk+l)i\k+l
implies that either

+(nl, Z) belongs to (IInl,...,nk)rk
that (A, Z.
statement
The
or tnl,...,nk[A;Z] is strategicallylargefor
n [A;Z] is empty = Y.
Z
istruein particularwhen Y, and to
(5nl,...,nk)<k
This is strategically largefor the subspace

Weclaimnowthat va
strategy for P. The game starts with
suitable
it we shall describea
prove of Y, the set r is strategically
large
our construction moves
craO
ff =largeforY- By ensure that aftersomefinitenumber
of
that P can with 0(nl) at
Y.That means
for Ef(Y) such that thereexistsnl
B1
In general,supposethat
E
willreacha sequence
(s)he large for [B1; Y].
the lengthof B1 and (nl)/\l
most Ef(Y) such
.,nk) with
that thereexists(nl,. . Y] The
reacheda sequence Bk E
largefor [Bk;
has
P of Bk and (vrll "...:nk )/\k
most the length Y, which
. . ., nk) at
s6(nl, that is strategicallylargefor
of Y guarantees Bk to a sequenceBk+lsuch
Tnl,...,nk
construction a way as to extend

means that P canplayin such at most the lengthof
Bk+l and
thereexists n,+l with +(nl,...,nk+l)
that Y]
largefor [B1c+l; finite
,...,nk+l )Z\k+l
let B be the infinite sequencewhichhasthe P
((7nl
LetP playas above and by

That is, B is the sequenceproduced by
sequences Bk as initialsegments. to provethat B E 5t. We do this
It remains
the end of the game.
at Write B = (xn)ono-l and f (nl, n2, . . .) =
bt.
t suchthat d(xt,Yt)>
% .
showing that d(B)f (nl, n2 . . .))
there exists
then
(yn)°n°=l. If our assertionis false large k and for everysequence
every sufficiently
By the continuityof f, for ni for all i % k) if z = f(ml,m2,...), then
mi =
(ml,m2,...)E J\Fwith r ) max{t,Q(nl, . . .,ns)},wehavejustproved
Yt). For any ) /x k iS
d(Zt,Yt) < o$t-d(xt,
xr ; Y] is empty, which impliesthat we find
(CJnl , . . . ,7?,k
n [X1,. . ., that 0(nl, . . ., nk) > t,

that ((7nl , . . . ,nk ) /\
Y]. Choosing k such for
notlargefor [xl, . . ., xr; ) +(nl, . . . ,nk) suchthat (Jnl,..,nk)/\kis large g
r
thattheredoesnot exist
[B,;Y], whichis a contradiction.
result can be strengthened

6. Spaces where the
Theorem 4.1 to a genuineRamseyresult
Weshowfirsthowto strengthen are
at the beginningis co and the blockbases
given the following
whenthe Banachspace between co and otherspacesis
normalized.The big difference
theorem[GS].

818 W. T. GOWERS
THEOREM 6.1. Let A1 U ... U Ar be a partition of the unit sphere of co

and let E > 0. Then there exists an infinite-dimensional block subspace Y of
co and some j < r such that every point in the unit sphere of Y is within E of
some point of Aj.
Define a subset A of the unit sphere of co to be large if it has a nonempty

intersection with every block subspace of co. (The word asymptotic is more
standard in Banach space theory, but less in keeping with the terminology of
this paper and other papers in Ramsey theory.) An equivalent formulation of
Theorem 1 is that if A is a large subset of co and E > 0 then there exists an
infinite-dimensional subspace Y of co such that every point of Y lies within E
of some point of A.
We shall now prove something similar for analytic sets of normalized block
bases of co (or equivalently, of block subspaces of co). Rather than giving full
details, which are very similar to the proof of Theorem 4.1, we merely indicate
the points at which that argument must be changed. Let ES(X) denote the
set of all normalized block bases of X and let us now interpret notation such
as [X] to refer to normalized block bases. (Alternatively, one could consider
sets cr C S(X) such that ((xn, An))nl E cr implies that ((xn, pn))n=l for all
sequences (p/i)~l of numbers in the interval [0,1].) Let us define a subset ac
of El(co) to be Ramsey if for every A > 0 there is a block subspace Y of co
such that either can [Y] is empty or [Y] C crzA.
LEMMA 6.2. N-closed subsets of El(co) are Ramsey.
Proof. Corollary 4.7 implies that N-closed subsets of E(co) are weakly
Ramsey. Let cr be N-closed and large for co, let A > 0 and let X be a block
subspace of co such that or-/2 is strategically large for X. Let A/2 - A0 <
A+ < A1 < A1 < A+ < A2 < A2 < ... be a strictly increasing sequence of
sequences bounded above by A. For each k let IIk be the set of pairs (A, Z)
such that A has length k and either
(a) for every z E Z, cr+ is not strategically large for [(A, z); X]
or
(b) for every z C Z, crais strategically large for [(A, z); X].
Clearly, if (A, Z) E Ilk and Z' is a subspace of Z then (A, Z') E 11k. We now
verify condition (i) of Lemma 2.2. Let (A, Z) be a pair such that (a) is not
true for any subspace of Z. This says that the set of z E Z for which vA,+ is
strategically large for [(A, z); X] is large for Z. Hence, by Theorem 6.1 there
is a subspace Z' C Z such that for every z E Z' the set /\- is strategically
large for [(A,z); X]. (Note that we did not use the full strength of the fact
that A+ < Ak+l. The strict inequality was important only for the kth terms
of these sequences.)

We may therefore apply Corollary 2.4, which, for a suitable choice of

perturbations, gives us a subspace Y of X such that, for every finite normalized
block basis A in Y, either
(c) for every z C Y, eAk is not strategically large for [(A, z); X]
or
(d) for every z E Y, 7A\k+l is strategically large for [(A, z); X].
We claim now that [Y] C cA. The proof is by induction. We start with
the statement that cAo is strategically large for [X]. It follows that there
exists z E S(Y) such that ao0 is strategically large for X (or else S would
have a winning strategy for the game c7Ao[X]by starting with the subspace
Y). By construction of [Y], for every zi C S(Y) the set oAl is strategically
large for [zli;X]. By a similar argument, for any fixed zi we then have that for
every Z2 e S(Y) the set cr2 is strategically large for [zi, z2; X]. Continuing in
this way we find that for any finite normalized block basis (zi,... , zk) in Y,
(CAk is strategically large for [zi,..., zk; X]. It follows that (zi,..., z) can be
extended to a sequence in cTA. By Lemma 4.6, (A is N-closed, and therefore
D-closed. It follows that every sequence (zn)n= E Si(Y) belongs to rA. C
6.3.
COROLLARY N-open subsets of Ei(co) are Ramsey.
Proof. Let a be an N-open subset of Ei (co) and let T be the complement

of aOA.It is not hard to show that cA is N-open, so r is N-closed. If no subpace
Y can be found such that [Y] c o/A, then r is large for co. Then, by Lemma
6.2, there is a subspace Y such that [Y] c TA. However, TA Cnu is empty, so
a n [Y] is empty. DI
Given a subspace Y of co and a collection p of finite sequences, let us say
that p is very large for Y if every sequence in ES(Y) has an initial segment
in p.
LEMMA 6.4. Let ao C Ei(co) be large, let a- = Un=i a, let A > 0 be a

real sequence and let 4 be a function from N to N. For any subspace Y C X let
p(Y) be the set of finite normalized block bases B in Y such that there exists n
with max{n, (n)} at most the length of B and (on)A large for [B; Y]. Then
there exists a subspace Y C X such that p(Y) is very large for Y.
Proof. For each Y, let r(Y) be defined in the same way as p(Y) but with
A replaced by A/2. Lemma 4.12 implies that there is a subspace W such that
r(W) is strategically large for W. In particular, r(W) is large for W. Hence, by
Corollary 6.3, there is a subspace Y c W such that [Y] C r(W)A/2. The result
now follows from the fact that r(W)A/2 C p(W) and that p(Y) C [Y] n p(W).
n

820 W. T. GOWERS
THEOREM6.5. Let a be an N-analytic subset of El(co) and let A > 0.

Then there is a subspace X C co such that either [X] n a = 0 or [X] C uA.
Proof (Sketch). The proof is very similar to that of Theorem 4.1. The
main changes are that the phrase "strategically large" should be replaced by
"very large" and that instead of Lemma 4.12 one should use Lemma 6.4. In
the penultimate paragraph of the proof, instead of considering a game between
P and S, one considers an arbitrary sequence B in ES(Y). The construction
of Y now guarantees that this sequence has an increasing collection of initial
segments B1, B2,... such that for some sequence ni,n2,... of positive inte-
gers, we have for every k that 0(nl,... ,nk) is at most the length of Bk and
(cnl,...,nk)Ak is large for [Bk; Y]. The proof that the distance between B and
f(nl, n2, ...) is at most A is the same as in the proof of Theorem 4.1. El
The above result also applies, for trivial reasons, to spaces containing c0.
It follows from results of Milman [Mi] and Odell and Schlumprecht [OS] that
these are the only spaces for which this stronger Ramsey result is true. The
result can in fact be strengthened further; it is basically symmetrical and so
an Ellentuck-type theorem holds.
Going to the opposite extreme, let us consider the "worst" spaces and
return to nonnormalized sequences. Recall that a space X (which we assume
for convenience has a basis here) is hereditarily indecomposableif it has one of
the following equivalent properties:
(i) X has no subspace that can be decomposed as a topological direct sum

Y e Z with Y and Z infinite-dimensional;
(ii) for any pair of infinite-dimensional subspaces Y and Z of X and any

e > 0 there exist y E Y and z E Z such that eIIY+ zll > IIY- zll;
(iii) for any pair of infinite-dimensional block subspaces Y and Z of X and any
positive real sequence A there are block subspaces Y' C Y and Z' C Z
such that d(Y', Z') < A.
The third property above, which we shall use, states that up to small
perturbations the subspaces of X form a filter. The existence of such spaces
is proved in [GM1]. The observation we make here is that if X is hereditarily
indecomposable and a C E(X) is a set such that, for some subspace Y C X,
P has a winning strategy for the game r[Y], then for any A > 0 P has a
winning strategy for the game ora[X]. Indeed, let Xi, X2,... be any sequence
of (infinite-dimensional block) subspaces of X. Using property (iii) with the
perturbations sufficiently small, we can find a sequence of subspaces Y1,Y2, ...
such that for every n c N and every y E Yn of norm 1 there exists x E Xn of

INFINITE RAMSEY THEOREM
AN 821
norm 1 such that IIY-S11 < Bn. If P plays a (fixed) winningstrategy against
in va. Thus we
themoves Y1,Y2,. . . from S, then P has produceda sequence
have the followingresult.
THEOREM6.6. Let X be a hereditarilyindecornposable
space, let A > O
Y s?lchthat
andlet ff c S (X) be N-analytic. Then either there is a subspace
ffn [r] = 0 or there is a winningstrategyfor
P for the game A[X].
7* Applications
Our aim in this section is to obtain the beginning;sof a classification

mean by this a
ofseparableinfinite-dimensionalBanach spaces. We do not
since such a task
completedescriptionof all such spaces up to isomorphism,
spaces with the
isclearlyhopeless. Rather, we wish to Snd classes of Banach
followingproperties:
X.
(a) If X is in one of the classes, then so is every subspaceof
(b) Every space has a subspacein one of the classes.
(c) The classes are very obviouslydisjoint.
a lot of informa-
(d) Knowingthat a space belongsto a particularclass gives
be defined
tion about the structureof the space and what operatorscan
on it.
with the
Properties (c) and (d) above are of course related. We begin
Theorem 4.1
commentthat it is slightly easier to deduce Theorem 1.4 from
the end of §3)
than from Theorem2.1, becausethe extra diagonalization(see
(xn, An)n=1
can be avoided. One definesff to be the set of all infinitesequences
the sequence
suchthat for everypositiveintegerm thereexist r and s suchthat
that X con-
(ArXr,Ar+1Xr+l,..,Asxs) is m-conditional. Lemma 1.6 implies
It is easy to
tains no unconditionalbasic sequenceif and only if ff is large.
basic sequence
check that a is a Ga-set. So if X contains no unconditional
is strategically
and A > 0, then we can find a subspaceY of X such that cr/\
proof of Corollary
large for Y. For sufficientlysmall /\, this shows, as in the
hereditarily
3.2) that Y satisfiescondition (ii) of Lemma 1.7, and is therefore
indecomposable .
hereditarily
In [GM1]it was shown that every operatoron a (complex)
of a multiple of the
indecomposablespace is a strictly singularperturbation
provedthat every
identity. This result was strengthenedby Ferenczi[F2],who
space into the space
operatorfroma subspaceof a hereditarilyindecomposable
We havetherefore
itself is a strictly singularperturbationof the inclusionmap.

822 T. GOWERS
W.
(d) is
foundtwo classes satisfying properties(a) to (d). However,property
to classify the
amatter of degree. It is not clear whether one would want
of view
indecomposablespacesany further(at least fromthe point
hereditarily
is a wide variety
ofunderstandingthe operatorsdefinedon them) but there
ways. At one
ofspaces with an unconditionalbasis differingin important
at the
extreme there are the Sp spaces, the ones with most symmetry,and
to any proper
otheris a space constructedin [G4] which is not isomorphic
that every operator on the
subspace of itself. In fact, it is shown in [GM2]
(It is not known
spaceis a strictly singularperturbationof a diagonalmap.
singular
whether everyoperatorfroma subspaceto the wholespaceis a strictly
maps
perturbation of the restrictionof a diagonalmap.) Since the diagonal
such a space
withboundedentriesdown the diagonalare trivially continuous,
basis, just
canbe regardedas the "worst"type of space with an unconditional
of generalspace.
asa hereditarilyindecomposablespace is the "worst"type
space with an
It is natural to ask, therefore,what can be said about a
unconditional basis if it does not containthe "worst:'kindof subspace. Is there
We shall
someresult stating that such a space must have a "nice"subspace?
givea result of this type, but we need some definitionsfirst.
Recallfirst that an infinite-dimensional BanachspaceX is said to be min-
imalif everyinfinite-dimensional subspaceof X has a furthersubspaceisomor-
that every
phicto X. This term was coined by Rosenthal,who conjectured
infinite-dimensional space has a minimalsubspace. However,it was shownby
Of
Casazzaand Odell [CO]that Tsirelson'sspace gives a counterexample.
counterex-
course,a hereditarilyindecomposablespace gives a much stronger
interestingone. It
ample,but for this very reasonTsirelson'sspace is a more
discuss in some
has in particulara propertywhich we shall now define and
incomparableif
detail. Recall that two spaces X and Y are said to be totally
of Y. Let
no infinite-dimensionalsubspaceof X is isomorphicto a subspace
a pair of totally
us say that a space X is quasi-minimalif it does not contain
given by the next
incomparablesubspaces. The reasonfor this terminologyis
to mean Y embeds
simplelemma. If Y and Z are two spaces, we write Y < Z
into Z.
LEMMA 7.1. A spaceX is quasi-rninirnal if and only if thereis a collec-
tion S of subspacesof X with the followingproperties:
(i) The set S is partiallyorderedby (.

and Z > W.
(ii) If Y, Z C S then there exists W E S such that Y > W
in S.
(iii) Every s?lbspaceof X has a subspaceisomorphicto a subspace

Proof. Write Y - Z if Y < Z and Z < Y. Then ~ is an equivalence

relation. If S consists of a representative from each equivalence class then it
clearly has the properties stated. The converse is obvious from properties (ii)
and (iii). Oi
Remarks. The important property is of course (ii). If a quasi-minimal

space X has a minimal subspace Y then Y embeds into every subspace of X.
(Under these circumstances X is said to be Y-saturated.) For if Z were a
subspace into which Y did not embed then Y would embed into every subspace
of Y and no subspace of Z, so Y and Z would be totally incomparable. It
follows that X has a minimal subspace if and only if S has cardinality one
for every choice of S. If S has cardinality greater than one, in which case we
shall say that X is strictly quasi-minimal, then S is uncountable. To see this,
suppose S is countably infinite (the finite case is easier). We can pick a chain
Y1 > Y2 > ... of subspaces in S such that every Y E S satisfies Y > Yn for
some n. Now let zl < z2 < ... with Zn E Yn and let Z be the subspace spanned
by Zl3,Z35 .... Then some Yn embeds into Z. However, Z is isomorphic to
the subspacespanned by Zn+, Zn+2, ... Z2n, Z2n+1 Z2n+3, Z2n+5*... which is a
subspace of Yn+1, contradicting the fact that Yn > Yn+1. Finally, observe
that a hereditarily indecomposable space is quasi-minimal for trivial reasons.
The definition will therefore only interest us in the context of spaces with an
unconditional basis.
We say that two block subspaces Y and Z of a space X are disjointly
supported if the support of every y E Y is disjoint from the support of every
z c Z. We can now state and prove a dichotomy for spaces with an uncondi-
tional basis. Later we shall strengthen it a little.
THEOREM7.2. Let X be a Banach space with an unconditional basis.

Then either X has a quasi-minimal subspace or X has a subspace Y such that
no two disjointly supported subspaces of Y are isomorphic (and hence any two
disjointly supported subspaces are totally incomparable).
Although it is possible to prove Theorem 7.2 in one go using Theorem

4.1, it is perhaps of interest to show that the full strength of Theorem 4.1
is by no means necessary. Instead we shall use Corollary 4.7 (that N-closed
sets are weakly Ramsey) and a diagonalization. In order to do this, we must,
as with Theorem 1.4, introduce some more quantitative definitions. Given
two block subspaces Y, Z of a Banach space X, we shall say that they are
C-block isomorphic if the linear map from Y to Z taking the normalized block
basis of Y to the normalized block basis of Z is a C-isomorphism. We shall
say that a space X is C-quasi-minimal if any two block subspaces Y, Z have

824 W. T. GOWERS
further subspaces V, W which are C-block isomorphic. In the next lemma,

when A and B are finite subsets of N, we write A < B for the statement that
max A < min B.
LEMMA7.3. Let X be a Banach space with an unconditional basis and

let C > 1 and e > O. Then either X has a (C + E)-quasi-minimal subspace or
X has-a subspace V such that no two disjointly supported block subspaces of V
are C-block isomorphic.
Proof. For convenience we shall assume that X has the property that
1x + yI > Ix Ilwhenever x < y are nonzero vectors in X. Standard arguments
tell us that every Banach space (with a given monotone basis) is almost iso-
metric to a space with this property. Suppose that every block subspace W of
X contains a pair of disjointly supported C-block-isomorphic subspaces Y, Z.
By passing to subsequences we may assume that Y and Z are generated by
normalized block bases (Yn) =l and (Zn) 1= such that yi < zi+l and zi < Yi+l
for every i > 1. We know that for every i > 1 the vectors Yi and zi are dis-
jointly supported linear combinations of the vectors in the block basis of W.
Therefore, we can find a sequence wi < w2 < ... of nonzero vectors in the unit
ball of W and a partition A1 < A2 < ... of N such that, if we let yi be the
sum of all wj such that j E Ai is odd, and zi be the sum of all wj such that
j E Ai is even, then the sequences Yi, Y2,... and zl, z2,... are C-equivalent
normalized block bases. Let a be the set of sequences (w)__ 1 for which such
a partition of N exists. (Of course, strictly speaking we should replace each wi
by the pair (wi/ ||wi||, |lwil).)
Notice that, given our sequence Wi,w2,..., the sets Ai, and hence the
vectors yi and zi, are uniquely determined. For example, A1 - {1,..., n},
where n is the unique positive integer such that the sum of all wj with j < n odd
has norm 1 and so does the sum of all wj with j < n even. Once A1 has been
determined, the uniqueness of A2 follows similarly, and so on. The sequences
yi, Y2, ... and zl, Z2, ... are C-equivalent if and only if all their initial segments
are C-equivalent. By the uniqueness of the sets Ai, this is a property of initial
segments of the sequence wl, W2,.... Moreover, if wl, W2,... is a sequence not
in r, either because we cannot find the sets Ai making the vectors yi and zi
normalized, or because for some n the sequences yi, . .., Yn and zl,..., Zn are
not C-equivalent, then a sufficiently small perturbation of wl, w2,... will also
not be in a. This shows that cris N-closed.
Since ao is closed and large for X, Corollary 4.7 allows us, for any
A > 0, to pass to a subspace V C X such that crAis strategically large for
Y. It is not hard to show that if A is sufficiently small, then for any sequence
(Wn)' 1l EE A there are normalized sequences (Yn)n = and (zn)n 1 which are
(C + c)-equivalent such that the yi are generated by Wl, w3,... and the zi are

AN INFINITE RAMSEY THEOREM
generatedby w2:w4, ... Since v\ is strategicallylarge for V, if S chooses 825

any two subspacesY,Z of V and plays the strategy Y, Z, Y,Z, . . ., P can find a
sequence(Wn)n=l E CA with wi E Y for i odd and wi E Z for i even. Choosing
sequences(Yn)°°=land (zn)n_l as above, we find block subspacesof Y and Z
which are (C + s)-block isomorphic. O
Proof of Theorem7.2. Suppose that X has no quasi-minimalsubspace.

Then obviouslyX has no rn-quasi-minimalsubspacefor any positive integer
m. Thereforeby Lemma7.3 we can find a nested sequenceof subspacesWl 2
W2 2 ... such that no two disjointly supported block subspacesof Ym are
(m-l)-block-isomorphic. Let w] < w2 < . . . be a normalizedblockbasis with
WmE Wm for every m and let W be the subspacegeneratedby W1:W2:....
We claim that no two disjointlysupportedsubspacesof W are isomorphic.
To see this, suppose that Y and Z are two disjointly supported block
subspacesof W. If they are isomorphic,then by standard argumentsthere
exist equivalentblock bases (yn)n=l and (zn)n=l with every Yiin Y and every
Ziin Z (but not necessarilygeneratingthe wholeof Y and Z). They must be C-
equivalentforsomeC, so let us choosea positiveintegerm ) C+2, and suppose
(by removingthe beginnings)that everyYiand everyZilies in Wm. We are not
quite finishedbecause the yz and Zi are not necessarilynormalized.However,
the ratio [lYill/ llzillis boundedand boundedawayfromzero. Therefore,we can
drop to subsequencessuch that this ratio convergesvery quicklyto a nonzero
limit, and then a small perturbationgives us normalizedblockbases whichare
(C + l)-equivalent. This then contradictsour choiceof Wm. O
It is clear that the dichotomyabove is genuinelya dichotomy.If a space

even containstwo disjointlysupportedtotally incomparablesubspacesthen it
is not quasi-minimal.As we mentionedabove, it is possibleto proveTheorem
7.2 more directly. If we let Cm be the set ff definedin the proofof Lemma7.3
when C = rn and let ff = U°°=l) then ff is an F-set. If everysubspacecontains
two isomorphicdisjointlysupportedfurthersubspaces,then the argumentin
the deductionof Theorem7.2 fromLemma7.3 showsthat ff is large. Applying
Theorem4.1, we can find a subspaceVVwhere < is strategicallylarge which
proves,as above, that W is quasi-minimal.
Notice that the proofof Theorem7.2 actuallygaveus somethingmore:ev-
ery Banachspace X has a subspaceY such that either Y is C-quasi-minimal
for some C or Y does not contain two disjointlysupportedisomorphicsub-
spaces. In particular, every quasi-minimalspace has a subspace which is
C-quasi-minimalfor some C.
This last result is not unlike a consequenceof our solution to Banach's
problem on homogeneousspaces. It follows from the fact that a homoge-
neous space X must be a Hilbert space that there exists a constant C such

826 W. T. GOWERS
that all subspaces of X are C-isomorphic. (One might call such a space
C-homogeneous.)However,no direct proof of this fact is known.
We now give some more informationabout the "bad"spaces in Theo-
rem 7.2. First, we make a simple observationusing a result of Casazza [C].
(The proof can also be found in [G4].)
LEMMA7.4. Let X be a space with a basis such that, for any bloc1z basis
Y1,Z1,Y2,Z2, . . the blocSc
basesY1,Y2,. . . andzl, Z2,. . . are not equivalent.Then
X ts isomorphicto no propersubspaceof itself (and hence the sarneis true in
any subspaceof X).
COROLLARY 7.5. Let X be a spacewith an unconditionalbasissuch that
no two disdointlysupportedsubspacesare isomorphic. Then no subspaceof X
is isomorphicto a furtherpropersubspace.
Proof. Clearlysuch a space satisfiesthe above criterionof Casazza. C1
We shall now strengthenCorollary7.5 considerably.In the proof below, we
identify a set A c N with the obvious projectionon X associated with A.
The result basically states that two block subspacesof a space X satisfying
the conditionsof Corollary7.5 are only isomorphicif they are isomorphicfor
trivial reasons.
LEMMA7.6. Let X be a space with an uncondttionalbasis such that no
two disjointlysupportedsubspacesare isomorphic,and let Y andZ besubspaces
of X generatedby equivalentblocScbases. Let T: Y ) Z be the corresponding
tsomorphistn.Then there is an invertiblediagonaloperatorD: X f X such
that T-Dly and T-1-D-l lz are strictly singtblar.
Proof. Let (yn)n=l be a normalizedblock basis generatingY and let ZrR =
TYnfor every n. Without loss of generalitythe basis of X is l-unconditional.
Let c, C be constants such that c IIYII % IITYII
S C IIYII for every y E Y. For
each n let Jn)Kn and Ln be respectivelythe sets of integersm such that O<
2CIYnmlS lZnmlr2C|Ynm|> lZnml> (cl2)lUnml> O and (C/2)18nml ) lZnml
Let J, K and L be the unions of the Jn, Kn and LnbNotice that these sets
are disjoint.
Now let D' be the uniquediagonaloperatoron KX sendingKyn to Kzn
for every n. Extend D' to an invertiblediagonaloperatorD on the whole of
X by mappingx C (1-K)X to Ax, where A is a constant between c and C.
We claim that D has the propertyrequired.
We show first that the map U definedon KY by KyrlW> (J + L)ynis con-
tinuous. If this were not so, then since (J8n)°n°=1
and (Lyn)n=lare not equiv-
alent we would be able to find y C Y such that either ilJYtI> 2 Il(K+ L)yll or
{lLyll> (4C/c) 11(J + K)yll. Now since the basis of X is l-unconditionaland

(Zn)=1 is a block basis, we have in the first case
IlTryll IITJy\I- jT(K + L)yjI > 2C \\Jyll - C I(K + L)yjj

> C(ljJy + jI(K+L)yjI) C IIyI
contradicting the fact that I[TI[I C. Similarly, in the second case, we have
IlTyll IJITLyll+ I\T(J+K)y\\

< (J + K)yjI < (3c/4) IlLyll< c IlYll
(c/2) IILyll+ C 11
contradicting the fact that IrT-11 cC-1. Since J, K and L are disjoint, we
also know that U cannot be an isomorphism on any subspace of KY, so it is
strictly singular.
It follows that the sequences (yn)n=l and (Kyn)=lI are equivalent. Thus
the map V defined on KY by Vyn = (1 - K)zn is continuous and therefore
also strictly singular. But T = (D + V)K = D - D(J + L) + VK so T - D is
strictly singular as claimed. The argument for T-1 is similar. D
Putting all these results together, we get the following theorem.

THEOREM 7.7. Let X be an infinite-dimensional Banach space. Then
X has a subspace Y with one of the following properties, which are mutually
exclusive and all possible.
(1) Y is hereditarily indecomposable, and therefore (by [F2]) every operator

from a subspace Z of Y into Y is a strictly singular perturbation of a
multiple of the inclusion map.
(2) Y has an unconditional basis and every isomorphism between block sub-
spaces W and Z of Y is a strictly singular perturbation of the restriction
of some invertible diagonal operator on Y.
(3) Y has an unconditional basis and is strictly quasi-minimal.
(4) Y has an unconditional basis and is minimal.
How might Theorem 7.7 be extended? The obvious class to look at is (3).
Notice that it is not at all clear that a strictly quasi-minimal space need be
isomorphic to a proper subspace. In fact, it is almost certainly not even true.
Indeed, a suggested counterexample, with a very sketchy argument about why
it was a counterexample, was given in [G3], but so far nobody, the author
included, has checked whether the details can be filled in. If such a space
exists, it is likely to be the "worst" quasi-minimal space in the sense that the
only operators on the space and its subspaces are essentially those guaranteed

828 W. T. GOWERS
to exist by quasi-minimality.One can then ask what there is to say about a

space not containingone of these '4worst"subspacesand hope to divide (3)
further.
The ultimate extension of Theorem7.7 would be a sequenceof possible
and mutuallyexclusiveproperties(1) to (k) such that, for any Banach space
X, there would exist j and a subspaceY C X such that Y was the "worst"
sort of space satisfyingproperty(j) in the sense that the only operatorson Y
werethose guaranteedby property(j), and such that property(k) was that of
being isomorphicto some Tp. Tsirelson'sspace ought to be an example of a
typical space having a propertya little strongerthan (3) but not as strong as
(4). Such a result would in a sense be a complete theory of the structureof
operatorson subspacesof Banach spaces, as it would show that every space
had a subspacethat was somehow "generic"of a certain kind, and therefore
could be describedin some detail.
Unfortunately,it is not altogetherobviouswhat the next step is. Suppose
X is strictly quasi-minimalwith an unconditionalbasis and suppose that for
every subspace Y of X there are isomorphicsubspaces W, Z of Y with W
properly contained in Z. Then no subspace of X satisfies the criterionof
Casazza(Lemma7.4) so in some furthersubspacethere is a winningstrategy
for P for producingsequences(xn)n_1suchthat the sequenceof odd-numbered
vectorsis equivalentto the sequenceof even-numberedones. This, althougha
strongerpropertythan quasi-minimality,is ratherartificial,and it is not clear
what of interest can be deducedfrom it. We can also find a winningstrategy
in some subspacefor sequences(Xn)n_l such that there is a partitionof N into
sets A1,A2, . . with max(Ai) < min(Ai+1)with Yi-EjGAi zj of norm 1 for
every i and with (x/ sIx2ll)Z°°land (yn)n=l equivalentblock bases. Again, it is
not clear how to use this fact.
It may well be that there is no tidy dichotomyconnectedwith isomor-
phisms to propersubspaces. If the followingconjectureis true, then it would
certainlyplace a limit on any dichotomythat one might wish to prove.
CONJECTURE 7.8. There exists a Banach space X such that every sub-
spaceY of X has furthers?>bspaces Z and 1X such that Z is isomorphicto no
propersubspaceof itself and W is isomorphicto its hyperplanes.
The difficultyin provingthis conjectureis that the knowntechniquesfor
findingspaces that are not isomorphicto any propersubspaceall produceex-
amples such that all subspaceshave the same property.A new idea is needed
for ruling out operatorson the whole space without doing the same for sub-
spaces.
It might be a good idea to focus on the followingrathergeneralquestion.
Recallthe definitionsconnectedwith Lemma7.1, that Y < X if Y is isomorphic
toasubspaceofXandYXifY4XandX<Y.

AN INFINITE RAMSEY THEOREM
829
PROBLEM7.9. Given a Banach space X, let P(X) be the set of all
equivalenceclasses of subspacesof X, partially orderedas above. For which
posets P does there eacista Banach spaceX such that every subspaceY of X
contains a f?lrthersubspaceZ with P(Z) = P?
Gharacterizingall such posets might be ratherdifficult,if it involvedcon-
structing exotic Banach spaces. However,even a strong necessarycondition,
perhapsprovedusing the methods of this paper, would be interesting.
Let us concludethis section with a vaguelystated problem.
PROBLEM 7.10. Are therefurthernonartificialapplicationsof Theorern4.1?
In particular,is there any appltcationwhichneeds the f?lllstrengthof the the-
orem?
Although I have searchedunsuccessfullyfor such an application,I still
believe that Theorem4.1 can be exploited further. Many natural classes of
sequences,such as the set of all sequencesthat generatea subspaceisomorphic
to an fp space, concernhereditarypropertiesof subspaces.However,Theorem
4.1 has nothing interestingto say about such properties,since it is trivial that
if ff is a hereditarypropertyof subspaces,then there is a subspacesuch that
either all its subspacesare in ff or none of them are. This indicatesthat any
applicationof Theorem4.1 is likely to involvea "clever"choice of sequences,
which gives me hope that there are interestingapplicationsthat have not yet
been found. As for using the full strength of the theorem, this could mean
applyingit to a set of sequenceswhich is analytic but not Borel (see [Bos]for
some good examplesof natural Banach-spacepropertieswhich are genuinely
analytic), but it would also be good to see an applicationwhich involved a
strategy for S that was-morecomplicatedthan simply alternatingbetweentwo
subspaces.
8. Appendix
In this sectionwe indicatebrieflydirectionsin whichthe resultsof Sections
2 to 5 cannotbe strengthened.The most obviousthing to showis that not every
set is weakly Ramsey. That followsfrom the followingstatement, that there
need not be a subspacein which the game is even approximatelydetermined.
The proof, of course, uses the axiom of choice, and in fact the continuum
hypothesis as well, although this is not so obviously necessary. (In fact, in
[BLA1]it is shown that Martin'saxiom suffices.) It would be interestingto
know whether the axiom of determinacyimplied that every set was weakly
Ramsey.
In orderto applythe continuumhypothesis,it is essentialthat the number
of strategiesfor both playersof the game is at most 28°. However,this is not
true with the definitionof the game given earlier,since both playershave 28°

830 W. T. GOWERS
possible moves at each stage. Fortunately,this problemis not too serious.

First of all, one can restrict each of P's moves to a countabledense set. (For
example, insist that they are rational linear combinationsof the block basis
presentedby S.) To deal with S, notice that it is not importantfor the game
that S should declare an entire block subspace all at once-it is enough
if S presentsthe block basis one vector at a time, letting P choose a linear
combinationof blocks at some finite stage. If in addition S is restrictedto a
countabledense set, then the numberof strategies for both playersis 28° as
required. I am gratefulto Joan Bagaria and Jordi Lopez Abad for pointing
out the necessityof this remark.
THEOREM8.1. Let A be the constantsequence(1/2,1/2,...). For every
X there exists a set ff c S(X) such that in no subspaceY c X doesP have a
wznningstrategyfor the game ,[Y] and in no subspaceY c X does S have
a winningstrategyfor the game vC[Y].
Proof. Well orderthe set of all possible pairs consistingof a pair (Y,f),
whereY is a block subspaceof X and Q is either a strategyfor P or a strategy
for S inside Y. Do this so that each pair has countablymany predecessors.
We inductivelydefine for each ordinal oe a sequence sa as follows. Suppose
the 01th pair in the well orderingis the pair (YOU,
,, ) and that Q(,is a strategy
for P in YO. Since there are only countablymany sequencess,g with p < og,
S can easily find a sequenceof moves such that the sequencechosen by P is
not within A of any of these s. In that case, let sa be the sequenceresulting
from such a play from S (with P playing the strategy fO). Call such an sa
an S-sequence. Similarlyvif , is a strategy for S, then P can easily choose a
sequences,, againstthis strategy not within /\ of any sg with S < oe. Call this
sort of sa a P-sequence.Note that, in particular,we have guaranteedthat no
S-sequenceis within A of any P-sequence.
Now let crbe the set of all P-sequences.Given any strategy for S and any
subspaceY, P can play againstthis strategy and producea P-sequence.Thus,
S has no strategy for any crC[Y].Similarly,given any strategy for P and any
subspaceY, S can forceP to producean S-sequence,whichis not in ,. So P
does not have a winningstrategy for a,, [Y]. g
Recallthe definitionof an asymptoticset givenearlier.A fundamentalfact

in Banachspace theory is that there exist spaces X such that the unit sphere
S(X) containstwo asymptotic sets A and B and d > 0 with llx-Yll > b for
everyx E A and y E B. It followsfromworkof Milman[Mi]that suchsets exist
in Tsirelson'sspace. A major breakthroughdue to Odell and Schlumprecht
[OS]was the discoveryof such sets in Tpfor 1 < p < oo. The results of [Mi],
[OS]and [G5]in fact show that a space has this propertyif and only if it does
nat contain co [OS, Corollary1].

INFINITE RAMSEY THEOREM
AN 831
COROLLARY8.2. If X doesnot containco:thenthereeacistsubsetsff and

rof S(X), andr > o, suchthatP has a winning
strategyfor everyv[Y] and
has no
everyr[Y] butP has no winningstrategyfor any (r n tr)[Y] andS
winningstrategyfor any ( n )c[V].
for every
Proof.Let A and B be asymptotic sets such that llx-Yll ) a
ff be the set of
xE 4 and y E 1S. Let p be a set given by Theorem8.1. Let
T be the same
allsequences(xn)°°=leither in p or such that xl E 4 and let
for cr[YX and for
withB. ClearlyP has winningstrategiesin every subspaceY
n Tr = Pr C pa
r[Y].Let r be a sequencewith tY1= min(a/2, 1/39. Then vr
anda n T = p, so the result followsfromthe propertyof p.
g
are obviously
If ff and r are given by this corollary,then (Ja/2 and 7A/2
weaklyRamsey,but their intersectionis not even in an approximate
completely
sensecompletelyweaklyRamsey.
8.3.
LEMMA If X doesnot containco then thereexzstsa larpesubset
ff c
dense.
S(X) andr, o s?lchthatvr zs D-*-nowhere
Let ff
Proof.Let A and r be as in the proof of the previouscorollary.
Clearlyff is large.
bethe set of sequences(SC,7,)n=l with SCn E XAfor every n.
[(A,z); Z] n vr
However,given any basic open set [A;Z], pick z E Z n 8. Then
O
=0.
9. Recent developments
version of a
As mentioned in the introduction, this paper is the final
no excuse for
preprint[G3]that has been aroundfor severalyears. I can offer
but am glad to remedy it now. I am gratefulto certain
this state of aflTairs,
ValentinFerenczi
peoplefor drawingmy attentionto mistakesin the preprint.
that D-
pointedout that the proofof Theorem8 in that paper (the statement
the proofin
open sets are weaklyRamsey)was incorrect. (On the other hand,
Joan Bagaria and
[G2]was correct.) This was later pointedout to me againby
maintheorem
JordiLopezAbad, who also said that they foundthe proofof the
In response,I
(correspondingto Theorem4.1 in this paper) hard to follow.
also noticed and
have made my presentationclearerin many respects. I have
to Lemma
correcteda mistake in the proof of Lemma 9, which corresponds
preprint,and was
4.12 here. (Lemma4.11 does not have a counterpartin the
the missingstep.)
of the
Recently Bagaria and Lopez Abad have given alternativeproofs
[BLA1,2]. Their
main results of this paper, and have extended some of them
expressed in a
argumentsare along similar lines to the ones here, but are

832 W. T. GOWERS
language more familiar to logicians. For example, they speak of dense generic
filters when carrying out their diagonalizations. Under additional set-theoretic
hypotheses they extend Theorem 4.1 to El-sets, that is, continuous images of
coanalytic sets. Their work was carried out before I revised my preprint.
DPMMS, UNIVERSITY OF CAMBRIDGE, WILBERFORCE ROAD, CAMBRIDGE CB3 OWB, UK

E-mail address: wtgl0@dpmms.cam.ac.uk
REFERENCES
[AD] S. A. ARGYROS and I. DELIYANNI, Examples of asymptotic iL Banach spaces, Trans.

Amer. Math. Soc. 349 (1997), 973-995.
[BLA1] J. BAGARIA and J. LOPEZ-ABAD,Weakly Ramsey sets in Banach spaces, Adv. Math.
160 (2001), 133-174.
[BLA2] , Determinacy and weakly Ramsey sets in Banach spaces, Trans. Amer.
Math. Soc. 354 (2002), 1327-1349.
[B] S. BANACH, Theorie des Operations Lineaires, Editions Jacques Gabay, Sceaux, 1993
(Reprint of 1932 original).
[Bo] B. BOLLOBAS,Combinatorics. Set Systems, Hypergraphs, Families of Vectors and
Combinatorial Probability, Cambridge University Press, Cambridge, 1986.
[Bos] B. BOSSARD, Theorie descriptive des ensembles en geometrie des espaces de Banach,
Ph.D. Thesis, Universite de Paris VI (1994).
[CS] T. J. CARLSON and S. G. SIMPSON, A dual form of Ramsey's theorem, Adv. in Math.
53 (1984), 265-290.
[C] P. G. CASAZZA, handwritten notes.
[CO] P. G. CASAZZA and E. ODELL, Tsirelson's space and minimal subspaces, Longhorn
Notes, University of Texas, Texas Functional Analysis Seminar (1982-3), 61-72.
[E] E. ELLENTUCK,A new proof that analytic sets are Ramsey, J. Symbolic Logic 39
(1974), 163-165.
[Fl] V. FERENCZI, A uniformly convex hereditarily indecomposable Banach space, Israel
J. Math. 102 (1997), 199-225.
[F2] , Operators on subspaces of hereditarily indecomposable Banach spaces,
Bull. London Math. Soc. 29 (1997), 338-344.
[GP] F. GALVIN and K. PRIKRY, Borel sets and Ramsey's theorem, J. Symbolic Logic 38
(1973), 193-198.
[G1] W. T. GOWERS,A new dichotomy for Banach spaces, Geom. Funct. Anal. 6 (1996),
1083-1093.
[G2] , A new dichotomy for Banach spaces, preprint.
[G3] , Analytic sets and games in Banach spaces, preprint IHES M/94/42.
[G4] , A solution to Banach's hyperplane problem, Bull. London Math. Soc. 26
(1994), 523-530.
[G5] , Lipschitz functions on classical spaces, Europ. J. Combin. 13 (1992), 141-
151.
[G6] , A Banach space not containing co, l1or a reflexive subspace, Trans. Amer.
Math. Soc. 344 (1994), 407-420.
[GM1] W. T. GOWERSand B. MAUREY, The unconditional basic sequence problem, J. Amer.
Math. Soc. 6 (1993), 851-874.
[GM2] , Banach spaces with small spaces of operators, Math. Ann. 307 (1997),
543-568.
[H] P. HABALA, A Banach space all of whose subspaces fail the Gordon-Lewis property,
Math. Ann. 310 (1998), 197-219.

[J] T. JECH, Set Theory, Academic Press, New York (1978).

[K T-J] R. A. KOMOROWSKI and N. TOMCZAK-JAEGERMANN, Banach spaces without local un-
conditional structure, Israel J. Math. 89 (1995), 205-226.
[LT] J. LINDENSTRAUSS and L. TZAFRIRI, Classical Banach Spaces. I. Sequence Spaces,
Springer-Verlag, New York, 1977.
[M] A. R. D. MATHIAS,Happy families, Ann. Math. Logic 12 (1977), 59-111.
[Mi] V. D. MILMAN, Spectrum of bounded continuous functions which are given on the
unit sphere of a Banach space, Funct. Anal. Appl. 3 (1969), 67-79 (translated from
Russian).
[N-W] C. ST. J. A. NASH-WILLIAMS, On well-quasi-ordering transfinite sequences, Proc.
Cambridge Philos. Soc. 61 (1965), 33-39.
[OS] E. ODELLand T. SCHLUMPRECHT, The distortion problem, Acta Math. 173 (1994),
259-281.
[S] J. SILVER,Every analytic set is Ramsey, J. Symbolic Logic 35 (1970), 60-64.
[Sz] A. SZANKOWSKI, Subspaces without the approximation property, Israel J. Math. 30
(1978), 123-129.
(Received March 30, 2000)


An Infinite Ramsey Theorem and Some Banach Space Dichotomies - Gowers

Uploaded by

Document Information

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

An Infinite Ramsey Theorem and Some Banach Space Dichotomies - Gowers

Uploaded by

Copyright:

Available Formats

Annals of Mathematics

An Infinite Ramsey Theorem and Some Banach-Space Dichotomies

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

An infinite Ramsey theorem and

A problem of Banach asks whether every infinite-dimensionalBanach

This paper containsa completeproofof a result announcedin [G1]which

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

Let us followusual practiceand definea Banachspace to be homogeneous

COROLLARY 1.2. Let X be a homogeneousBanach space. Then either

Notice that in the second case no subspace of X has an unconditional

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

oneof infinite rank and infinite corank). Otherwise, X is tndecomposable.

The definitionswe have given of an unconditionalbasis and a hereditarily

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

LEMMA 1.5. Let X be a Banach space with a basis (en)n=l, let Y be a

LEMMA1.6. Let X be a Banach space. The following two statements are

(i) X has no subspace with an unconditional basis.

(i) X is hereditarily indecomposable.

(yi + i)| > c|(yi- i)|

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

AN INFINITE RAMSEY THEOREM 801

Notice that if we insist that Y = Z in condition (ii) of Lemma 1.7 then

2. Further definitions and some preliminary results

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

THEOREM2.1. Let X be a Banach space, let a be any subset of Ef(X)

LEMMA2.2. Let A1, A2,... be a sequence of positive real sequences and

(ii) if (A, Z) E IIn and Z' C Z then (A, Z') Ec n.

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

We shall often use specializations of Lemma 2.2. It may be convenient

(ii) if (y, Z) E H and Z' C Z then (y, Z') E II.

(iii) if (A, Z) E Hn then A has length n.

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

x E Xn and every subspaceY of Xn we can find a subspaceZ c Y such that

It is not hard to see that the above result is equivalentto Theorem2.1.

Proof of Theorem2.1. Let T = C/\/2.The assumptionof the theoremand

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

Beforewe apply Theorem2.1, let us introducetwo furtherdefinitions.We

COROLLARY 3.2. Let X be a Banach space. Then either X contains a

By the triangleinequalityand our choice of A, we know that

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

It follows from our choice of r7that

(yi +zi) >(C- ) (yi - zi)

Proof of Theorem 1.4. If X has no subspace with an unconditional basis,

4. Definitions and preliminary results for infinite sequences

We begin this section with a brief discussion of the connections between

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

Given a subset crof , we can define an infinitegame just as we did for

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

Thus,Theorem4.1 statesthat N-analyticsets areweaklyRamsey,anda

((1 + >1)t1/2, * * ., (1 + An)Xn/2)

suchthat the sequence

11(1 + A)X-(1 + H)Y|| ) maX{|S-81, ||x-Yll-1>-81} .

Therefore,applyingTheorem2.1 to p with A replacedby A/4 we obtainthe

THEOREM4.3. 4.3. All D-opensets areweaAcly

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

show that D-*-closed sets are

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

THEOREM 4.5. Let a be a D-*-closed subset of , let T be the complement

Proof. Suppose there is no subspace Y such that [Y] C rT. Then X

COROLLARY4.7. Every N-closed set is weakly Ramsey.

This content downloaded from 129.173.72.87 on Mon, 12 Aug 2013 22:34:03 PM

let An > Ofor every