Sei sulla pagina 1di 331

Progress in Mathematics

Volume 263

Series Editors
H.Bass
J. Oesterl
A. Weinstein
Stevo Todorcevic

Walks on Ordinals
and Their
Characteristics

Birkhuser
Basel Boston Berlin
Stevo Todorcevic
Universit Paris VII C.N.R.S. Department of Mathematics
UMR 7056 University of Toronto
2, Place Jussieu Case 7012 Toronto M5S 2E4
75251 Paris Cedex 05 Canada
France e-mail: stevo@math.toronto.edu
e-mail: stevo@math.jussieu.fr

and

Mathematical Institute, SANU


Kneza Mihaila 35
11000 Belgrad
Serbia
e-mail: stevo@mi.sanu.ac.yu

2000 Mathematics Subject Classication 03E10, 03E75, 05D10, 06A07, 46B03, 54D65,
54A25

Library of Congress Control Number: 2007933914

Bibliographic information published by Die Deutsche Bibliothek. Die Deutsche Biblio-


thek lists this publication in the Deutsche Nationalbibliograe; detailed bibliographic
data is available in the Internet at http://dnb.ddb.de

ISBN 978-3-7643-8528-6 Birkhuser Verlag AG, Basel Boston Berlin

This work is subject to copyright. All rights are reserved, whether the whole or part of
the material is concerned, specically the rights of translation, reprinting, re-use of
illustrations, broadcasting, reproduction on microlms or in other ways, and storage
in data banks. For any kind of use whatsoever, permission from the copyright owner
must be obtained.

2007 Birkhuser Verlag AG


Basel Boston Berlin
P.O. Box 133, CH-4010 Basel, Switzerland
Part of Springer Science+Business Media
Printed on acid-free paper produced from chlorine-free pulp. TCF
Printed in Germany

ISBN 978-3-7643-8528-6 e-ISBN 978-3-7643-8529-3

987654321 www.birkhauser.ch
Contents

1 Introduction
1.1 Walks and the metric theory of ordinals . . . . . . . . . . . . . . . 1
1.2 Summary of results . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Prerequisites and notation . . . . . . . . . . . . . . . . . . . . . . . 17
1.4 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 Walks on Countable Ordinals
2.1 Walks on countable ordinals and their basic characteristics . . . . . 19
2.2 The coherence of maximal weights . . . . . . . . . . . . . . . . . . 29
2.3 Oscillations of traces . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4 The number of steps and the last step functions . . . . . . . . . . . 47
3 Metric Theory of Countable Ordinals
3.1 Triangle inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2 Constructing a Souslin tree using . . . . . . . . . . . . . . . . . . 58
3.3 A Hausdor gap from . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4 A general theory of subadditive functions on 1 . . . . . . . . . . 66
3.5 Conditional weakly null sequences based on
subadditive functions . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4 Coherent Mappings and Trees
4.1 Coherent mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.2 Lipschitz property of coherent trees . . . . . . . . . . . . . . . . . . 95
4.3 The global structure of the class of coherent trees . . . . . . . . . . 108
4.4 Lexicographically ordered coherent trees . . . . . . . . . . . . . . . 124
4.5 Stationary C-lines . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5 The Square-bracket Operation on Countable Ordinals
5.1 The upper trace and the square-bracket operation . . . . . . . . . 133
5.2 Projecting the square-bracket operation . . . . . . . . . . . . . . . 139
5.3 Some geometrical applications of the
square-bracket operation . . . . . . . . . . . . . . . . . . . . . . . . 144
vi Contents

5.4 A square-bracket operation from a special Aronszajn tree . . . . . 152


5.5 A square-bracket operation from the complete binary tree . . . . . 157
6 General Walks and Their Characteristics
6.1 The full code and its application in
characterizing Mahlo cardinals . . . . . . . . . . . . . . . . . . . . 161
6.2 The weight function and its local versions . . . . . . . . . . . . . . 174
6.3 Unboundedness of the number of steps . . . . . . . . . . . . . . . . 178
7 Square Sequences
7.1 Square sequences and their full lower traces . . . . . . . . . . . . . 187
7.2 Square sequences and local versions of . . . . . . . . . . . . . . . 195
7.3 Special square sequence and the corresponding function . . . . . 202
7.4 The function on successors of regular cardinals . . . . . . . . . . 205
7.5 Forcing constructions based on . . . . . . . . . . . . . . . . . . . 213
7.6 The function on successors of singular cardinals . . . . . . . . . . 220
8 The Oscillation Mapping and the Square-bracket Operation
8.1 The oscillation mapping . . . . . . . . . . . . . . . . . . . . . . . . 233
8.2 The trace lter and the square-bracket operation . . . . . . . . . . 243
8.3 Projections of the square-bracket operation
on accessible cardinals . . . . . . . . . . . . . . . . . . . . . . . . . 251
8.4 Two more variations on the square-bracket operation . . . . . . . . 257
9 Unbounded Functions
9.1 Partial square-sequences . . . . . . . . . . . . . . . . . . . . . . . . 271
9.2 Unbounded subadditive functions . . . . . . . . . . . . . . . . . . . 273
9.3 Changs conjecture and 2 . . . . . . . . . . . . . . . . . . . . . . 277
9.4 Higher dimensions and the continuum hypothesis . . . . . . . . . . 283
10 Higher Dimensions
10.1 Stepping-up to higher dimensions . . . . . . . . . . . . . . . . . . . 289
10.2 Changs conjecture as a 3-dimensional
Ramsey-theoretic statement . . . . . . . . . . . . . . . . . . . . . . 294
10.3 Three-dimensional oscillation mapping . . . . . . . . . . . . . . . . 298
10.4 Two-cardinal walks . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Chapter 1

Introduction

1.1 Walks and the metric theory of ordinals


This book is devoted to a particular recursive method of constructing mathemati-
cal structures that live on a given ordinal , using a single transformation  C
which assigns to every ordinal < a set C of smaller ordinals that is closed and
unbounded in the set of ordinals < . The transnite sequence
C ( < )
which we call a C-sequence and on which we base our recursive constructions may
have a number of coherence properties and we shall give a detailed study of them
and the way they inuence these constructions. Here, coherence usually means
that the C s are chosen in some canonical way, beyond the already mentioned and
natural requirement that C is closed and unbounded in for all . For example,
choosing a canonical fundamental sequence of sets C for < 0 , relying
on the specic properties of the Cantor normal form for ordinals below the rst
ordinal satisfying the equation x = x , is a basis for a number of important
results in proof theory. In set theory, one is interested in longer sequences as well
and usually has a dierent perspective in applications, so one is naturally led to
use some other tools besides the Cantor normal form. It turns out that the sets
C can not only be used as ladders for climbing up in recursive constructions but
also as tools for walking from an ordinal to a smaller one ,
= 0 > 1 > > n1 > n = ,
where the step i i+1 is dened by letting i+1 be the minimal point of Ci
that is bigger than or equal to . This notion of a walk and the corresponding
characteristics and distance functions constitute the main body of study in
this book. We show that the resulting metric theory of ordinals is a theory of
considerable intrinsic interest which provides not only a unied approach to a
2 Chapter 1. Introduction

number of classical problems in set theory but is also easily applicable to other
areas of mathematics. For example, highly applicable characteristics of the walk
are dened on the basis of the corresponding traces. The most natural trace of
the walk is its upper trace dened simply to be the set

Tr(, ) = {0 > 1 > > n1 > n }

of places visited along the way, which is of course most naturally enumerated in
decreasing order. Another important trace of the walk is its lower trace, the set

(, ) = {0 1 n2 n1 },
i
where i = max( j=0 Cj ) for i < n. The traces are usually used in dening
various binary operations on ordinals < , the most prominent of which is the
square-bracket operation that gives us a way to transfer the quantier for every
unbounded set to the quantier for every closed and unbounded set. It is perhaps
not surprising that this reduction of quantiers has proven to be quite useful in
constructions of mathematical structures on where one needs to have some grip
on substructures of cardinality .
From the metric theory of ordinals based on analysis of walks, one also learns
that the triangle inequality of an ultrametric

(, ) max{ (, ), (, )}

has three versions, depending on the natural ordering between the ordinals ,
and . The three versions of the inequality are in fact of a quite dierent character
and occur in quite dierent places and constructions in set theory. For example,
the most frequent occurrence is the case < < , when the triangle inequality
becomes something that one can call transitivity of . Considerably more subtle is
the case < < of this inequality1 . It is this case of the inequality that captures
most of the coherence properties found in this article. It is also an inequality that
has proven to be quite useful in applications.
A large portion of the book is organized as a discussion of four basic char-
acteristics of the walk , 0 , 1 , 2 and 3 . The reader may choose to follow the
analysis of any of these functions in various contexts. The characteristic 0 (, )
codes the entire walk = 0 > 1 > > n1 > n = by simply listing the
positions of i+1 in the set Ci for i < n. While this looks simple-minded, the
resulting mapping 0 is a rather remarkable object. For example, in the realm of
the space 1 of countable ordinals, it gives us a canonical example of a special
Aronszajn tree of increasing sequences of rationals which has the additional re-
markable property that, when ordered lexicographically, its cartesian square can
be covered by countably many chains. In other words, the single characteristic
0 of walks on countable ordinals gives two critical structures, one in the class of
1 It
appears that the third case < < of this inequality is rarely a reasonable assumption to
be made in this context.
1.1. Walks and the metric theory of ordinals 3

so-called Lipschitz trees and the other in the class of linear orderings. For higher
cardinals , analysis of 0 leads us to some interesting nitary characterizations
of hyper inaccessible cardinals. This is given in some detail in Chapter 6 of this
book.
The characteristic 1 (, ) loses a considerable amount of information about
the walk as it records only the maximal order type among the sets

{C0 , C1 , . . . , Cn1 }.

Nevertheless it gives us the rst example of what we call a coherent mapping. The
class of coherent mappings and trees in the case = 1 exhibits an unexpected
structure that we study in great detail in Chapter 4 of the book. The ne structure
in the class of coherent trees is based on the metric notion of a Lipschitz mapping
between trees. The profusion of such mappings between coherent trees eventually
leads us to the so-called Lipschitz Map Conjecture that has proven crucial for
the nal resolution of the basis problem for uncountable linear orderings and that
is presented in the same chapter. For higher cardinals the characteristic 1 and
its local versions oer a rich source of so-called unbounded functions that have
some applications.
The characteristic 2 (, ) simply counts the number of steps of the walk
from to . While this also looks rather simple minded, the remarkable proper-
ties of the corresponding function 2 become especially apparent on higher car-
dinals . Important properties of this characteristic are its coherence and its un-
boundedness. The coherence property of 2 requires the corresponding C-sequence
C ( < ) to be coherent in the sense that C = C whenever is a limit
point of C . On the other hand, the unboundedness of 2 translates into a require-
ment that the corresponding C-sequence C ( < ) be nontrivial2 , a condition
that eventually leads us to a simple and natural characterization of weakly com-
pact cardinals that we choose to reproduce in some detail in Chapter 6.
Finally, the characteristic 3 (, ) attaches one of the digits 0 or 1 to the walk
according to the behavior of the last step n1 n = . The full analysis of
this characteristic is currently available only in the reals of the space 1 , where 3
becomes a rather canonical example of a sequence-coherent mapping with values
in {0, 1} and with properties reminiscent of those appearing in the well-known
notion of a Hausdor gap in the quotient algebra P()/n (another critical object
that shows up in many problems about this quotient structure).
The true metric theory of ordinals comes only with development of the
characteristic (, ) of the walk that takes advantage of the so-called full lower
trace of the walk. The depth of this characteristic is apparent even in the space 1
of countable ordinals, but its full power comes at higher cardinals and especially
at that are successors of singular cardinals. The full analysis of the characteristic
requires C ( < ) to be a so-called square sequence or in other words requires
2 We say that C ( < ) is nontrivial if there is no closed and unbounded set C such that.
for all limit points of C, there is such that C C .
4 Chapter 1. Introduction

the most widely known coherence condition on this sequence, which says that if
is a limit point of C , then C = C . It is not surprising that this characteristic
has the largest number of applications, many of which are reproduced in this book.
We have already mentioned that its development in the case = 1 was the initial
impulse for development of the so-called metric theory of countable ordinals that
has already a rich spectrum of applications. At higher cardinals the characteristic
can be used in facilitating set-theoretic forcing constructions of rather special
objects and we shall reproduce some of these constructions in Chapter 7 of this
book. While, as said above, the full development of requires C ( < ) to be
a square sequence, the function itself holds considerable information about
the notion of square sequences. For example in Chapter 7, we use to turn an
arbitrary square sequence C ( < ) into a non-special one by expressing the
usual order relation among ordinals < as an increasing union of tree-orderings
that come from square sequences on themselves. This again leads us to some
applications that we choose to reproduce in detail at the end of Chapter 7.
We have already mentioned that one of the important outcomes of our study
of walks on ordinals is the square-bracket operation, a transformation which to
every pair < of ordinals < assigns an ordinal [] belonging to the upper
trace Tr(, ) of the walk from to . We have also mentioned that the choice of
[] has to be rather careful in order to reduce an arbitrary unbounded subset A
of to the corresponding set of values

{[] : , A and < }

that contains a closed and unbounded subset of relative to some xed stationary
set , which the C-sequence C ( < ) avoids3. We present several variations
on the way [] is chosen, each of which works best in some particular context.
The common feature of these denitions of [] is that they are all based on the
oscillation mapping
osc : P()2 Card
dened by
osc(x, y) = |x \ (sup(x y) + 1)/ |,
where is the equivalence relation on x \ (sup(x y) + 1) dened by letting
if and only if the closed interval determined by the ordinals and contains no
point from y. In other words, osc(x, y) is simply the number of convex pieces that
the set x \ (sup(x y) + 1) is split into by the set y. The original theory of the
oscillation mapping osc has been developed in the realm of partial functions from
into . In other words, there is a well-developed theory of the oscillation mapping

osc(s, t) = |{ : s() t() but s( + ) > t( + )}|, 4


3 C ( < ) avoids if C = for all limit ordinals < .

4 Here, + is the immediate successor of in the common domain of s and t.
1.1. Walks and the metric theory of ordinals 5

(see, for example, [111]), but the general theory works equally well and it will be
in part reproduced here in Chapters 8 and 9. The common feature of all results
of such a theory is identication of the notion of unbounded, in either of the
two contexts, in such a way that the typical oscillation result would say that
the set of values osc(x, y) the oscillation mapping takes when x and y run inside
two unbounded sets is in some sense rich. In our context of dening the square-
bracket operation, the sets x and y in osc(x, y) are members of our C-sequence
C ( < ) on which we base the notion of walk, and the notion of unbounded
becomes the familiar notion of nontriviality of C ( < ). This makes the square-
bracket operation [] well dened in a wide variety of contexts and therefore
quite applicable.
Judging from the applications found so far, it appears that in order to make
a particular variation of the oscillation mapping or the square-bracket operation
useful, one needs to be be able to give a quite precise estimate of its behavior,
not only on unbounded subsets of but also on families of pairwise-disjoint
nite subsets of . It is for this reason that denite results about any particular
variation of osc and [] presented in this book will typically be about families
A of pairwise-disjoint nite subsets of . Given such a family A of cardinality
, by going to a subfamily, we may assume that elements of A have some xed
nite cardinality n. So, given a in A one can view it enumerated increasingly as
a(0), . . . , a(n 1). All variations of the square-bracket operation that we present
will have the property that the set of ordinals < that can be represented as

= [a(0)b(0)] = [a(1)b(1)] = = [a(n 1)b(n 1)],

for some a < b in A, contains a closed and unbounded set relative to some xed
stationary set which the C-sequence C ( < ) avoids. Many applications
however require that we know the values [a(i)b(j)] when i = j < n and when a
and b run through A. It turns out that modulo taking a projection [[]] of [], (or
in other words, modulo composing [] with a map from into ) for many of the
square-bracket operations that we dene in this book, the particular set

{[[a(i)b(j)]] : i, j < n and i = j}

of values will be independent of the choice of a = b in A. This turns out to be


crucial in several applications of [] presented in this book. Naturally, one would
also like to know whether one can dene a variation on [] where we would have
freedom of getting arbitrary values of the form [a(i)b(j)] independently of whether
i = j or not. It turns out that this is indeed possible for some choices of , though
the corresponding denitions are necessarily less general as they do not apply in
the case = 1 since otherwise one would be able to prove that the countable
chain condition is not productive without appealing to additional axioms of set
theory.5 In case = 1 , we do have symmetric binary operations with some degree
5 Recall that the countable chain condition is a productive property under MA1 (see [36], 41E).
6 Chapter 1. Introduction

of freedom in that direction (see Chapter 2), but the exact breaking point between
this and what requires additional axioms of set theory has yet to be determined.
This book will also present some higher-dimensional characteristics of the
walk, though in that context the full theory is yet to be developed. For example,
in Chapter 10, we consider the characteristic (, , ) which to any given three
ordinals < < < assigns the place where the walk from to branches from
the walk from to . It turns out that when C ( < + ) is a square sequence,
the characteristic can be used to step-up objects living on to objects on its
successor + . For example, one application of this characteristic is found in the
proof that Changs Conjecture6 is equivalent to a 3-dimensional Ramsey-theoretic
statement saying that, for every coloring of [2 ]3 with 1 colors, there is an un-
countable set B 2 which misses at least one of the colors. The 3-dimensional
characteristic (, , ) that simply measures the length of the common parts of
the walks and can be used for detecting when a subset of admits
a rich restriction of the 3-dimensional version of the oscillation mapping,
osc : [+ ]3 ,
dened on the basis of its 2-dimensional version as follows:
osc(, , ) = osc(Cs \ , Ct \ ),
where s = 0 (, )  (, , ) and t = 0 (, )  (, , ). Here s is the member
i of the trace of the walk = 0 > > l = whose code is the sequence s, i.e.,
0 (i , ) = s, and similarly t is the term j of the walk = 0 > > k =
whose code is the sequence t. In Chapter 10 we show that our analysis of the
3-dimensional version of the oscillation mapping leads naturally to a square-
bracket operation in that dimension, though the full analogy is yet to be com-
pleted as one still needs to determine the behavior of this operation on families of
pairwise-disjoint nite subsets of + (which at the moment seems elusive).
The nal section of Chapter 10 is concerned with generalizing the basic notion
of walks to the context of sets of ordinals rather than ordinals themselves, with
the goal of obtaining two-cardinal versions of the square-bracket operation. For
example, we show that for every pair of innite cardinals < with regular,
there is a mapping c : [[] ]2 such that, for every conal subset U of [] and
every , there exist x y in U such that c(x, y) = . Fuller analogues of the
square-bracket operation are however obtained only in certain cases. For example,
assuming that there is a stationary subset S of [] which is equinumerous with
a locally countable7 subset of [] , one can dene a square-bracket operation
[]S : [[] ]2 []
6 Recall that Changs conjecture is the model-theoretic statement claiming that every model of a

countable signature that has the form (2 , 1 , <, . . . ) has an uncountable elementary submodel
M such that M 1 is countable.
7 A subset K [] is locally countable, if the set {x K : x a} is countable for every

a [] .
1.1. Walks and the metric theory of ordinals 7

with the property that, for every conal subset U of [] , the set of all s S that
are not of the form [xy]S for some x y in U is not stationary in [] . Since under
the same assumption about the stationary set S, this set can be partitioned into |S|
pairwise-disjoint stationary subsets, one gets the following: For every stationary
subset S of [] that is equinumerous with a locally countable subset of [] , there
is a projection
[[..]]S : [[] ]2 S

of the square-bracket operation [..]S with the property that, for every conal set
U [] and every z S, there exist x y in U such that [[xy]]S = z.
An interesting phenomenon that one realizes while analysing walks on or-
dinals is the special role of the rst uncountable ordinal 1 in this theory. Any
natural coherency requirement on the sets C ( < ) that one nds in this theory
is satisable in the case = 1 . How natural the notion of walk in this context is
can be seen from the fact that basically all of its characteristics lead us in one way
or the other to some critical structure that shows up in various rough classica-
tions of mathematical structures. For example, any of the characteristics , 0 , 1 ,
2 and 3 of the walk that we study here lead us to the canonical linear ordering
appearing on the list of ve linear orderings that forms a basis for the class of all
uncountable linear orderings. The rst uncountable cardinal is the only cardinal
on which the theory can be carried out without relying on additional axioms of
set theory. The rst uncountable cardinal is also the place where the theory has
its deepest applications as well as its most important open problems. This special
role can perhaps be best explained by the fact that many set-theoretical problems,
especially those coming from other elds of mathematics, are usually concerned
only about the duality between the countable and the uncountable rather than
some intricate relationship between two or more uncountable cardinalities. 8 For
example, consider the classical problem coming from topology asking if the hered-
itary separability of a given regular space X is equivalent to its dual requirement
that every collection of open subsets of X has a countable subcollection with the
same union.9 It turns out that this problem has a reformulation in terms of the
behavior of mappings of the form c : [1 ]2 2 on uncountable families A of
pairwise-disjoint nite subsets of 1 . In other words, a mapping that has a cer-
tain complex behavior on uncountable families of pairwise disjoint nite sheafs
would lead to an example of a regular space in which one of the implications fails.
For example, one can produce a hereditary separable non-Lindelof space assuming
there is c : [1 ]2 2 with the property that, for every uncountable family A
of pairwise-disjoint nite subsets of 1 , all of some xed size n for every position

8 This is of course not to say that an intricate relationship between two or more uncountable

cardinalities may not be a protable detour in the course of solving such a problem. In fact, this
is one of the reasons for our attempt to develop the metric theory of ordinals without restricting
ourselves to the realm of countable ordinals.
9 In other words, every subspace of X is Lindelof.
8 Chapter 1. Introduction

i0 < n and every requirement h : n 2, there exist a < b in A such that

c(a(i0 ), b(j)) = h(j) for all j < n.

On the other hand, there is a regular hereditary Lindel of non-separable space if


there is c : [1 ]2 2 with the dual property that, for every uncountable family
A of pairwise-disjoint nite subsets of 1 , all of some xed size n for every position
j0 < n and every requirement h : n 2, there exist a < b in A such that

c(a(i), b(j0 )) = h(i) for all i < n.

It turns out that one cannot produce c having the rst property without appealing
to additional axioms of set theory, but that a variation of the oscillation mapping
that we reproduce in Chapter 2 of this book will give us a c with the second
property and therefore a regular hereditary Lindel of non-separable space.
One can nd this kind of application of the oscillation mapping or the square-
bracket operation on 1 in other areas of mathematics as well. For example con-
sider the problem of nding a large subspace on which a given homogeneous poly-
nomial P : X C is zero, where X is some Banach space over the eld C of
complex numbers.10 It is known that if X is innite-dimensional, then there will
always be an innite-dimensional subspace Y of X on which P is zero, but can one
nd larger Y assuming X is not separable? More precisely, if X is not separable,
can one nd a non-separable subspace Y of X on which P vanishes? Note that a
counterexample would be a polynomial P : X C which takes all the complex
values on any non-separable subspace of X. So it is natural to try appealing to
the square-bracket operation for constructing such a polynomial P . It turns out
that this is indeed possible and we shall reproduce it in Chapter 5 of this book
for the space X = 1 (1 ). In order to apply the square-bracket operation we need
to take one of its convenient projections [[]]G0 with only countably many values,
which one may assume to form a countable dense subset D of the unit disc of the
complex plane. Then the polynomial P : 1 (1 ) C is dened as

P (x) = [[]]G0 x x .

In order to establish that P takes all the values from C on any closed non-separable
subspace of 1 (1 ), one uses the following property of the projection [[]]G0 : For
every uncountable family A of pairwise-disjoint nite subsets of 1 , all of the
same size n, there is uncountable set B A, an equivalence relation on n =
{0, 1, . . . , n 1}, and a single mapping h : n n D such that

[[a(i)b(j)]]G0 = h(i, j) for all a < b in B and i  j.


10 Recall
that a homogeneous polynomial of degree n on a Banach space X is a mapping P :
X C of the form P (x) = (x, x, . . . , x), where : X n C is a bounded symmetric n-linear
form on X.
1.1. Walks and the metric theory of ordinals 9

Moreover, for every uncountable C B and for every g : n n D there exist


a < b in C such that
[[a(i)b(j)]]G0 = g(i, j) for all i, j < n with i j.
This particular application makes it clear how useful it is to know behavior of
the square-bracket operation not only on uncountable subsets of 1 but also on
arbitrary uncountable families A of pairwise-disjoint subsets of 1 .
Minimal walks in general and the metric theory of ordinals in particular
are applicable to problems that are not necessarily problems about uncountable
structures but rather about the behavior of countable substructures of a xed
large structure. We demonstrate this in Chapter 3 where we present some ap-
plications of the metric theory of countable ordinals to classical problems from
innite-dimensional geometry, such as, for example, the distortion problem or the
unconditional basic sequence problem. Given a function
: []2
that is locally nite11 and that satises the two ultrametric inequalities,12 one
can use it to dene special functionals on the vector space c00 () of all nitely
supported maps from the ordinal into the reals. Let e < be the standard
Hamel basis of c00 (). Let e < be the corresponding sequence of biorthogonal
functionals. We say that a sequence (Ei )i<n of nite subsets of 1 is special if
Ei < Ej 13 for i < j < n, and if
|E0 | = 1 and |Ej | =  (E0 , E1 , . . . , Ej1 ) for 0 < j < n.
Here  (E0 , E1 , . . . , Ej1 ) is the integer that codes in some natural way the nite
metric structure induced by on the union of these sets. To a nite set E , we
associate the vector and functional on c00 (),
1  1 
xE = e and E = e .
|E|1/2 E |E|1/2 E

Given a special sequence (Ei )i<n of nite subsets of 1 , the corresponding special
functional is dened by i<n Ei . Special functionals induce a norm on c00 (),
and if we let X denote the completion of c00 () under this norm, the sequence
e ( < ) will be a weakly null sequence in X with no innite unconditional
subsequence. More precisely, we show in Chapter 3 that, if (Ei )i< is any innite
special sequence of nite subsets of and, if for i < we let vi = xEi as dened
above, then for every n < and every sequence (ai )in [1, +1] of scalars,

k 
n 
k
max | ai |
ai vi
(3 + ) max | ai | .
0kn 0kn
i=0 i=0 i=0
11 I.e.,
the set { < : (, ) n} is nite for every < and n < .
12 I.e.,
(, ) max{(, ), (, )} and (, ) max{(, ), (, )} for all < < .
13 We let E < F if every ordinal from E is smaller than every ordinal from F .
10 Chapter 1. Introduction

We have already noted that the largest ordinal supporting such a -function is the
rst uncountable ordinal 1 . It follows therefore that one can have an uncountable
weakly-null sequence with no innite unconditional subsequence. It follows that
the function can be used to control all countably innite subsets of the long
weakly-null sequence e ( < 1 ). With a considerable amount of additional work
one can even build a reexive Banach space X1 with a Schauder basis of length
1 with no innite unconditional basic sequence. Extending these ideas further
one can use to build another Banach space X01 with a Schauder basis of length
1 , which on one hand keeps the distortion and all the conditional structure of the
space X1 at the level of its nonseparable subspaces, but on the other hand, every
innite-dimensional closed subspace of X01 contains an isomorphic copy of the
space c0 , or in other words separable subspaces of X01 do not hold any conditional
structure nor do they allow their norms to be distorted.

1.2 Summary of results


Many chapters of the book can be read independently from each other. Once the
basic denition of the walk is understood, the reader may choose to follow the
development of a particular characteristic in various contexts. The reader inclined
towards applications of the methods of this book may start by reading about a
particular application and go back towards background material that is needed.
In this section, we include a short summary that might help the reader in nding
specic results presented in this book.
The rst section of Chapter 2 is the one to be rst read as it presents the
notion of minimal walk along a given C-sequence and the corresponding charac-
teristic 0 that codes it. The resulting tree T (0 ) is a tree of height 1 that has
all of its levels countable and which admits a natural strictly increasing map into
the rationals. This could be the shortest construction of such a tree found in the
literature and surely is the most canonical one. This could be seen for example
on the basis of the fact that if we order T (0 ) lexicographically we obtain an un-
countable linearly ordered set whose cartesian square can be covered by countably
many chains. The proof of this fact depends on the development of the full lower
trace of the minimal walk and the reader is advised to skip this on the rst reading
and go instead to the second section of Chapter 2 which presents the characteristic
1 and the corresponding tree T (1 ) for which this fact is easier to establish. The
characteristic 1 has its own interesting application presented in the same section.
For example, we show that the tree T (1 ) naturally leads us to an example of
a homogeneous non metrizable compactum that can be represented as a weakly
compact subset of some Banach space. As another application of 1 presented in
the same section is a functor that transfers a given graph G on the vertex set 1
to a graph G on the same vertex set such that if G has an uncountable clique
then the vertex set 1 can be covered by countably many sets that are cliques of
G. Moreover, we show that G satises the countable chain condition provided G
1.2. Summary of results 11

satises a slightly stronger version of this condition saying that for every uncount-
able family F of pairwise disjoint nite subsets of 1 , we can nd a = b in F such
that {, } is an edge of G for every a and b. The point of this is that es-
sentially all known ccc graphs on 1 do satisfy this stronger form of the countable
chain condition and therefore this in particular shows that many of the standard
consequences of MA1 are Ramsey-theoretic in nature. Recall that it is still not
known if MA1 is in fact equivalent to the statement that every ccc graph on the
vertex set 1 contains an uncountable clique. We nish this section with a proof
that unlike T (0 ) the tree T (1 ) does not always admit a strictly increasing map
into the rationals. This amounts to measuring how much information about the
minimal walk is lost when one passes from the full code 0 (, ) to the maximal
weight 1 (, ) of the walk.

The third section of Chapter 2 presents an important theme that is going


to be developed in later parts of the book, the theme of the oscillation mapping.
More precisely, in Section 2.3, we present two oscillation mappings osc0 and osc1
corresponding to the upper and lower trace of the minimal walk, respectively. Of
specic applications of these two oscillation mappings, we present a rather abso-
lute decomposition of 1 into innitely many pairwise disjoint stationary subsets,
and an example of a regular hereditarily Lindel of topological space that is not
separable. In Section 2.3 we introduce two new characteristics 2 and 3 . The
characteristic 2 while quite simple minded in that it counts only the number of
steps of the minimal walk its full power will become apparent only in later parts
of the book when dealing with walks on larger cardinals. In fact, as far as we
know, 2 could be the rst nontrivial two-place mapping from a large cardinal
into in the sense that it takes arbitrarily high value from on any product
of two unbounded subsets of . Recall that large cardinals typically have the
Ramsey-theoretic properties saying that maps f : []2 are constant on []2
for large subsets . The mapping 3 while considerably more subtle makes
sense only in the realm of countable ordinals. To a given walk from a countable
ordinal down to a smaller ordinal , the characteristic 3 assigns one of the
digits 0 or 1 according to what happens on the last step of the walk. Since it is
a coherent mapping it shares many properties with the characteristic 1 though
we shall show that, unlike T (1 ), if we order lexicographically T (3 ), we are not
always guaranteed that the cartesian square of the corresponding uncountable lin-
ear ordering can be covered by countably many chains. On the other hand, it is
true that every uncountable subset X of T (3 ) contains an uncountable subset Y
which when ordered lexicographically has the property that its cartesian square
can be covered by countably many chains. This again amounts to measuring how
much of the information is lost about the minimal walks by passing from the char-
acteristic 1 (, ) to the characteristic 3 (, ). However, while 3 looses much of
the information about the minimal walk, it still gives us a rather interesting and
canonical object on 1 that is very much reminiscent of the classical notion of a
Hausdor gap in the quotient algebra P()/n.
12 Chapter 1. Introduction

In Chapter 3 we develop the metric theory of countable ordinals concentrated


around the two ultrametric triangle inequalities

d(, ) max{d(, ), d(, )} and d(, ) max{d(, ), d(, )},

whenever < < . We have already mentioned two applications of this theory
found in this chapter, a reexive Banach space X1 with a Schauder basis of length
1 with no innite unconditional basic sequence, and another Banach space X01
with a Schauder basis of length 1 that is, on one hand saturated with copies of
the space c0 but on the other hand its norm admits an arbitrarily high distortion
relative to the class of nonseparable subspaces of X01 . Chapter 3 presents also
the characteristic of the minimal walk which besides the two mentioned triangle
inequalities has many other interesting properties that show their full power on
larger cardinals than 1 . In Chapter 3 we do present two important objects that
are naturally derived from , the Cohen name for a Souslin tree and a particularly
canonical example of a Hausdor gap in P()/n. The Souslin tree has domain 1
and the property that no innite ground model set can be its chain or antichain.
In Section 3.4 we study the general theory of functions : [1 ]2 that satisfy
the two ultrametric triangle inequalities and that are locally nite in the sets that
the set { < : (, ) n} is nite for every < 1 and every n < . We show
that there is a vast variety of such mappings including the universal one whose
construction we reproduce in the same section.
In Chapter 4 we develop a metric theory of trees that lies behind the prop-
erties of the tree T (1 ). The basic notion here is that of a Lipschitz map between
trees, a level preserving map with the property that (g(x), g(y)) (x, y) for all
x and y in its domain. This leads us to the notion of a Lipschitz tree, an uncount-
able tree T with the property that every level preserving map from an uncountable
subset of T into T has a Lipschitz restriction on some uncountable subset of its
domain. The main purpose of Chapter 4 is to study the class C of Lipschitz trees
as a structure equipped with the quasi ordering S T dened to hold whenever
there is a Lipschitz map from the tree S into the tree T , or equivalently whenever
there is a strictly increasing map from S into T. We also examine the relationship
between the class C and the larger class A of Aronszajn trees. Most of the theory
is developed without appeal to additional axioms of set theory though in the same
places we have used either MA1 or the Proper Forcing Axiom. It turns out that
Lipschitz trees share many of the properties of the tree T (1 ) such as for example
the property that when ordered lexicographically the cartesian square of the corre-
sponding uncountable linear ordering can be covered by countably many chains. It
turns out that all trees T (0 ), T (1 ), T (2 ), T (3 ), and T () associated to various
characteristics of walks in 1 are Lipschitz. It turns out also that the class C is
totally ordered under . In fact the chain (C, ) is discrete in the sense that every
T in C admits a naturally dened shift T (1) that forms an immediate successor
of T in C. In fact, the shift T (1) is also an immediate successor of T even in the
bigger class A and this is essentially equivalent to Shelahs Conjecture saying that
1.2. Summary of results 13

an uncountable ordering either contains an uncountable well ordered or conversely


well ordered subset, an uncountable separable ordered subset, or an uncountable
subset whose cartesian square can be covered by countably many chains. It turns
out that the chain C is not well ordered and this in particular solves an old problem
of R. Laver asking if the class A is well quasi-ordered under a stronger ordering
than . Chapter 4 contains many other structural results about the classes C and
A. For example, we show that while A is not totally ordered under , the chain
C is both conal and coinitial in (A, ).
In Chapter 5 we introduce and study the square-bracket operation [] for
pairs and of countable ordinals. The ordinal [] is taken from the upper trace
of the minimal walk from to . In later parts of the book we shall see several
variation of this choice of [] but the basic idea is always the same and it is based
on the oscillation mapping osc0 of upper traces exposed above in Section 2.3. In
Chapter 5 itself we present two other variations on the square-bracket operation,
one based on special Aronszajn tree and the other on the tree of all nite binary
sequences. What the square-bracket operation adds to the space 1 of countable
ordinals is a rigidity which can formally be expressed by the fact that there is a
sentence of L(Q2 ) which has only rigid models. This is presented in Example 5.1.10
which solves a problem of Ebbinghaus and Flum who showed that every model
of a sentence of L(Q) has a nontrivial automorphism. In Section 5.3 we present
tree geometrical application of the square-bracket operation. The rst one is an
example of a projective geometry of points and hyperplanes in Rn . The second
one is the example of a complex polynomial on (1 ) with no nonseparable null
subspace already mentioned above. The third application of [] is an example of
a reexive Banach space X with a Schauder basis of length 1 in which every
operator T can be written as I + S, where S is an operator with separable range.
While in some sense this example is subsumed by the example of the Banach space
X1 from Chapter 3 we have included it as its construction uses a quite dierent
set of ideas. In Section 5.4 we give a formal explanation of how the square-bracket
operation reduces the quantication over uncountable subsets of 1 to that over
closed and unbounded subsets of 1 . Answering a question of W.H. Woodin, we
show that there is a natural functor based on [] which to every subset on 1
associates a graph K [1 ]2 such that, modulo PFA or Woodins axiom (), a
subset of 1 contains a closed and unbounded set if and only if K contains
[X]2 for some uncountable X 1 .
In Chapter 6 we develop the characteristic 0 of the minimal walk in the
general context. Recall that in the context of walks on 1 the tree T (0 ) that cor-
responds to the characteristics 0 is special in the sense that it can be decomposed
into countably many antichains. In Chapter 6 we show that a strongly inaccessible
cardinal is not Mahlo precisely when one is able to nd a C-sequence in for
which the corresponding tree T (0 ) special in the sense that it admits a regressive
mapping that is not constant on any subset of T (0 ) that cannot be covered by less
than antichains. The idea is then used in providing a unied approach towards
14 Chapter 1. Introduction

Ramsey-theoretic characterizations of n-Mahlo cardinals in terms of the existence


of min-homogeneous sets relative to regressive maps of the form f : []n+2
or f : []n+3 . In the same chapter we develop the general theory of the
characteristics 1 and 2 . In particular, in Section 6.2 we develop the local version
1 : []2 of 1 for an arbitrary regular cardinal < . The main interest in
this variation of 1 is that under some very mild assumptions about one obtains
mappings from []2 into that have strong unboundedness properties. Similar un-
boundedness property of the characteristic 2 : []2 is equivalent to the non
triviality property of the C-sequence C ( < ) saying that there is no closed
and unbounded subset C of such that for every < there is such that
C C . This gives a vast variety of cardinals for which one can nd a
C-sequence C ( < ) such that the corresponding function 2 : []2 is
strongly unbounded. This can be seen from the characterization of weakly com-
pact cardinals given in Section 6.3 which says that a strongly inaccessible cardinal
is weakly compact if and only if for every C-sequence C ( < ) there is a
closed and unbounded set C such that for all < there is such that
C = C . We nish Chapter 6 with a particular topological application of
the unboundedness property of 2

In Chapter 7 we study walks based on C-sequences C ( < ) that satisfy


the coherence property saying that C = C whenever is a limit point of C .
We call such C-sequences square sequences. It is this condition on the C-sequence
which permits that the full theory of walks on 1 be lifted to the level that
supports it. For example, we show that the full lower trace F(, ) of walks along
square sequence keeps all its properties from the context of countable ordinals. As
an application we show that the characteristic 2 in this context has an interesting
coherence property saying that sup< |2 (, ) 2 (, )| < for all < < .
This has an interesting topological interpretation saying that the square of the
sequential fan with edges has tightness equal to . Another interpretation of
the same fact is the statement that the P-ideal dichotomy is incompatible with
the existence of a nontrivial square sequence on any cardinal that is larger than
1 . The main reason of imposing the coherence property on a given C-sequence is
however motivated by attempts to develop the theory of the characteristic in the
general context. This is done in Section 7.2 where we also deal with local versions
of this characteristic. A typical application of this new theory is Theorem 7.2.14
saying that if a regular uncountable cardinal = 1 carries a nontrivial square
sequence then it also carries one for this the corresponding tree (, <2 ) is not
special. Complementing a well-known result of Laver and Shelah, this shows that
the existence of a weakly compact cardinal is the exact consistence strength of
the statement that all Aronszajn trees on 2 are special. Special square sequence
is something that is more closely related to Jensens notion of square sequences
on successor cardinals and we develop the corresponding theory of walks in the
following sections of Chapter 7. Particularly interesting is the case of successors
of regular cardinals which we consider in Section 7.4 where we develop the theory
1.2. Summary of results 15

of the corresponding characteristic and the derived set-mapping D : [+ ]2


[+ ]< dened by
D{, } = { min{, } : {, } {, }}.
In Section 7.5 we use the set-mapping D to construct interesting forcing notions.
For example, we show that if a regular cardinal is -inaccessible then there is a -
closed -cc forcing notion that introduces a Souslin tree on + as well as a forcing
notion that introduces a Kurepa family on . So, in particular, we show that if
21 holds then there is a property K forcing notion that introduces a Souslin
tree on 2 . In the same section we reproduce a proof that under 21 there is a
property K forcing notion that introduces a locally compact scattered topology
on 2 all of whose Cantor-Bendixson ranks are countable. We deal with successors
of singular cardinals in Section 7.5. The main applications of the characteristic
in this context are in producing Jensen matrices Jn ( < + , n < ) of subsets
of + . As one application of Jensen matrices we present a construction of a conal
Kurepa family of countable subsets of + . Conal Kurepa families are quite useful
objects. We show this by using them in constructing Bernstein decompositions
of arbitrary Hausdor spaces, or in constructing coherent families of nite-to-one
maps indexed by countable subsets of + .
In Chapter 8 we develop the general theory of the oscillation mapping of
traces and the corresponding square-bracket operation. In fact we give three vari-
ations on the basic idea behind the denition of the square-bracket operation with
a particular emphasis on the properties of the corresponding projections. Projec-
tions are something that is typically needed in applications and they do tend to
impose accessibility conditions on the cardinal . For example, we show that if, for
example, is the rst inaccessible cardinal then there is a projection o : []2
of the oscillation mapping which takes all the values from on any set of the
form []2 for unbounded subset of . However if one wants such a projection of
the oscillation mapping that could be useful in constructing interesting complex
polynomials on 1 () or interesting bilinear mappings one needs to impose the
condition that is not bigger than the continuum. It turns out that the basic de-
nition of the square-bracket operation from 1 lifts without diculties when 1 is
replaced by a cardinal that is a successor of a regular cardinal or more generally
that admits a non reecting stationary subset. A typical application of this result
is the statement that for every regular cardinal there is a 2-nilpotent group G
of cardinality + such that every abelian subgroup of G has cardinality at most
. To obtain a projection [[]] of [] that would have the generic behaviour14 one
needs to vary the original idea using the assumption that the cardinal is in some
sense large. We present two such variations with some topological applications of
the corresponding projections. For example, we show that for every regular un-
countable cardinal there is a topological group G of cellularity whose cartesian
14 Inother words, that would have the property that for every family A of pairwise disjoint
nite subsets of all of some xed size n and for every mapping h : n n there exist a < b
in A such that [ a(i)b(j)]] = h(i, j) of all i, j < n.
16 Chapter 1. Introduction

square has cellularity > . In fact, this result is true for arbitrary uncountable car-
dinal that is not necessarily regular. This is proved in a similar fashion using
the colorings dened on successors of singular cardinals exposed in Chapter 9 of
this book.
In the rst section of Chapter 9, we use the characteristic in producing par-
tial square sequence that can sometimes be used in places of full square sequences.
In the second section of Chapter 9, we show that for a regular uncountable cardinal
, there is a structure of the form (+ , , Rn )n< with no elementary substructure
B of size with B bounded in just in case there is a strongly unbounded func-
tion f : [+ ]2 satisfying the two ultrametric triangle inequalities mentioned
above. This gives us a useful reformulation of the well-known model-theoretic
statement known as Changs Conjecture which is simply the statement that every
structure of the form (2 , 1 , Rn )n< has an uncountable elementary substructure
B such that B 1 is countable. In Section 9.3 we use this reformulation to show
that Changs Conjecture is also equivalent to the statement that every ccc poset
forces that every mapping of the form g : 2 2 must be constant on some
product of two innite subsets of 2 . This is then used in Section 9.4 in analyzing
possible Ramsey-theoretic reformulations of the Continuum Hypothesis.
In Section 10.1 we develop a general stepping-up procedure that lifts struc-
tures living on a given cardinal to similarly behaved structure living on its
successor + . We manage this by just assuming the existence of a square sequence
on + though we are able to step-up some complex combinatorial statements like
these of Hajnal and Komjath where originally one would have expected that the
higher-gap morasses are needed. As an application, we show that using 21 one
can construct a reexive Banach space X with a Schauder basis of length 2 with
the property that every bounded linear operator T : X X can be written as
a sum of an operator with a separable range and a diagonal operator with only
countably many changes of eigenvalues. In Section 10.2 we show that Changs
Conjecture is equivalent to the purely Ramsey-theoretic statement saying that for
every f : [2 ]2 1 there is an uncountable set 2 such that f  []2 = 1 .
In Section 10.3 we introduce the 3-dimensional oscillation mapping and use it in
producing a coloring c : [2 ]3 that takes all of its values from on the
symmetric cube of any uncountable subset of 2 .
The last section of Chapter 10 is an attempt to start the corresponding
theory of two cardinal walks, an analogous theory that would work in the context
of structures of the form P () = {x : |x| < } in place of ordinals and
cardinals. In fact we were able to dene a notion of walk in this context that
is suciently rich for giving us the analogue of the square-bracket operation in
this context. For example, we show that for every pair < of cardinals with
regular there is a projection [[]] : [[] ]2 of the square-bracket operation
in the context of [] (or more formally, in the context of P+ ()) such that for
every conal subset U of [] and every there exist x y in U such that
[[xy]] = . A particularly interesting case of this result is when = which allows
1.3. Prerequisites and notation 17

a further elaboration. For example, we show that if there is a stationary subset


of [] that is equinumerous with a locally countable subset of [] then there is
a variation of the square-bracket operation []S : [[] ]2 [] with properties
quite analogous to those of the square-bracket operation of 1 (which is really the
case = and = 1 ) in the sense that for every conal subset U of [] the
set of all s S that are not of the form [xy]S for x y in U is not stationary
in [] . One of the points of this variation of the square bracket operation is that
the set S can be split into |S| pairwise disjoint stationary sets, and since quite
frequently S has cardinality that is bigger than , we can dene projections
[[]]S : [[] ]2 that take more than colors (more precisely, = |S| colors)
on any conal subset U of [] .

1.3 Prerequisites and notation


We have tried to keep the prerequisites needed for mastering the material of this
book to a minimum. Though no specic training in set theory is necessary, the
reader should be familiar with the notion of ordinal and recursive denitions and
inductive proofs on them. This all can be found in most of the introductory texts
on the subject such as, for example, the newer ones [51], [64] and [55], or the older
text [62] which has particularly detailed expositions of the recursive denitions over
ordinals. By comparing, the reader will notice that we are using standard notation
in essentially complete agreement with these textbooks. In these sources the reader
will nd all the operations and properties of ordinals that we will use, but if more
complete treatment is needed the reader may also consult one of the specialized
texts like [7]. We shall also look at ordinals with their natural topology induced by
order. It is in this context that we refer to closed subsets of a particular ordinal
. The notions of unbounded and conality in this context refer of course to the
natural ordering of . The combinatorially inclined readers will notice that we are
adopting the Erd osRado notation for the symmetric powers [S] , the collections
of all subsets of the set S of cardinality . Of special interest are of course the
nite symmetric powers [S]k and the mappings dened on them because of the
clear connection that this book has with Ramsey theory (see [40], [33] and [133]).
In fact most of our characteristics of walks are mappings with domains equal to
symmetric squares []2 of some ordinals . A pair {, } []2 is usually assumed
to be written such that < . In other words, sometimes it is convenient to
identify the symmetric square []2 with the set

{(, ) 2 : < }

of ordered pairs. In this way a given characteristic f : []2 of the walk can
be identied with the sequence f : ( < ) of ber mappings dened by

f () = f (, ).
18 Chapter 1. Introduction

Also we can more clearly express when the given characteristic f is coherent in
some way by referring to the coherence between the corresponding ber mappings.
Moreover, this identication allows us to use the notation f (, ) instead of the
more cumbersome notation f ({, }). Basically all our characteristics

f : []2

of walks are dened recursively in the sense that f (, ) is dened on the basis
of the values f ( ,  ) on lexicographically smaller pairs (, ). All the recursive
denitions require us to specify the boundary values f (, ) which are typically
taken to be constant values such us 0 or depending on the context. It is for this
reason that sometimes we implicitly assume that the diagonal {(, ) : < } is
a part of the domain of f .

1.4 Acknowledgements
We are grateful to a large number of people who have read versions of the manu-
script and suggested corrections and improvements or helped us with preparation
of the nal manuscript. Of these we would like in particular to mention Maxim
Burke, Christine H
artl, Akihiro Kanamori, Bernhard Konig, Piotr Koszmider, Car-
los Martinez, Justin Moore and Luis Pereira. Special thanks are due to Jordi
Lopez-Abad for helping us in designing the diagrams and gures of the book.
Chapter 2

Walks on Countable Ordinals

2.1 Walks on countable ordinals and


their basic characteristics
The space 1 of countable ordinals is by far the most interesting space considered
in this book. There are many mathematical problems whose combinatorial essence
can be reformulated as problems about 1 , which is in some sense the smallest
uncountable structure. What we mean by structure is 1 together with a system
C ( < 1 ) of fundamental sequences, i.e., a system with the following two
properties:

(a) C+1 = {},


(b) C is an unbounded subset of of order-type , whenever is a countable
limit ordinal > 0.

Despite its simplicity, this structure can be used to derive virtually all known
other structures that have been dened so far on 1 . There is a natural recursive
way of picking up the fundamental sequences C , a recursion that refers to the Can-
tor normal form which works well for, say, ordinals < 0 1 . For longer fundamental
sequences one typically relies on some other principles of recursive denitions and
one typically works with fundamental sequences with as few extra properties as
possible. We shall see that the following assumption is what is frequently needed
and will therefore be implicitly assumed whenever necessary:

(c) if is a limit ordinal, then C does not contain limit ordinals.


1 One is tempted to believe that the recursion can be stretched all the way up to and this is
1
probably the way P.S. Alexandro discovered the phenomenon that regressive mappings on 1
must be constant on uncountable sets (see [2] and [3, Appendix]).
20 Chapter 2. Walks on Countable Ordinals

Denition 2.1.1. A step from a countable ordinal towards a smaller ordinal is


the minimal point of C that is . The cardinality of the set C , or better
to say the order-type of this set, is the weight of the step.
Denition 2.1.2. A walk (or a minimal walk) from a countable ordinal to a
smaller ordinal is the sequence = 0 > 1 > > n = such that for each
i < n, the ordinal i+1 is the step from i towards .
Analysis of this notion leads to several two-place functions on 1 that have
a rich structure and many applications. Let us expose some of these functions.
Denition 2.1.3. The full code of the walk is the function 0 : [1 ]2 < ,
dened recursively by
0 (, ) = |C | 0 (, min(C \ )),
with the boundary value2 0 (, ) = where the symbol  refers to the sequence
obtained by concatenating the one-term sequence |C | with the already
known nite sequence 0 (, min(C \ ) of integers.
Denition 2.1.4. The upper trace of the walk from to is the nite set Tr(, )
of ordinals between and visited during the minimal walk from to . Note
the following recursive denition of this trace,
Tr(, ) = {} Tr(, min(C \ ))
where the boundary value is Tr(, ) = {}. Note also the following expression of
the upper trace in terms of the full code,
Tr(, ) = { : 0 (, )  0 (, )},
where  denotes the ordering of being an initial segment among nite sequences
of ordinals. One also has the recursive denition of this trace.
Denition 2.1.5. The lower trace of the walk from to is the set
L(, ) = {(, ) : Tr(, ) and = },
where
(, ) = max{max(C ) : Tr(, ) and = }.
Note the following recursive denition of this trace,
L(, ) = (L(, min(C \ )) \ max(C )) {max(C )}3
with the boundary value L(, ) = .
2 Note that while 0 operates on the set [1 ]2 of unordered pairs of countable ordinals, the
formal recursive denition of 0 (, ) does involve the term 0 (, ) and so we have to give it
as a boundary value 0 (, ) = . The boundary value is always specied when such a function
is recursively dened and will typically be either 0 or the empty sequence . This convention
will be used even when a function e with domain such as [1 ]2 is given to us beforehand and we
consider its diagonal value e(, ).
3 If C = , we let L(, ) = L(, min(C \ )).

2.1. Walks on countable ordinals and their basic characteristics 21

Figure 2.1: The walk and its traces.

Thus, if = 0 > 1 > > n = is the minimal walk from to (i.e.,


0 = , n = , and i+1 = min(Ci \ ) for i < n), then
Tr(, ) = {i : 0 i n} and L(, ) = {i : 0 i < n},
i
where i = max( j=0 Cj ). The following immediate facts about these notions
will be frequently and often implicitly used.
Lemma 2.1.6. For ,
(a) > L(, ) implies that 0 (, ) = 0 (, ) 0 (, ) and therefore that
Tr(, ) = Tr(, ) Tr(, ),
(b) L(, ) > L(, ) implies that L(, ) = L(, ) L(, ).4 
4 Fortwo sets of ordinals F and G, by F < G we denote the fact that every ordinal from F is
smaller than every ordinal from G. When G = {}, we write F < rather than F < {}.
22 Chapter 2. Walks on Countable Ordinals

Denition 2.1.7. The full lower trace of the minimal walk is the function
F : [1 ]2 [1 ]< , dened recursively by

F(, ) = F(, min(C \ )) F(, ),
C

with the boundary value F(, ) = {} for all .


Lemma 2.1.8. For all < 1 ,
(a) F(, ) F(, ) F(, ),
(b) F(, ) F(, ) F(, ).
Proof. The proof is by induction on . Let

1 = min(C \ ).

Towards proving (a), we rst show that the part F(, 1 ) of F(, ) is included
in F(, ) F(, ). Consider rst the case 1 < . By the inductive hypothesis,

F(, 1 ) F(, ) F(1 , ) F(, ) F(, ),

since 1 C , and therefore, F(1 , ) F(, ). If 1 , then 1 = min(C \


) and again F(, 1 ) F(, ). By the inductive hypothesis,

F(, 1 ) F(, ) F(, 1 ) F(, ) F(, ),

so we are done in this case as well. Let us now consider the factor F(, ) of F(, ),
where C . Using the inductive hypothesis we get

F(, ) F(, ) F(, ) F(, ) F(, ),

since C and F(, ) is by denition included in F(, ).


To prove (b) consider rst the case 1 < . By the inductive hypothesis,

F(, ) F(, 1 ) F(1 , ).

Since 1 C , the set F(1 , ) is included in F (, ). Similarly by the denition


of F(, ), the set F(, 1 ) is included in F(, ).
Suppose now that 1 . Note that in this case 1 = min(C ), so F(, 1 )
is included in F(, ). Using the inductive hypothesis for 1 we have

F(, ) F(, 1 ) F(, 1 ) F(, ) F(, ). 

Lemma 2.1.9. For all ,


(a) 0 (, ) = 0 (min(F(, ) \ ), ) 0 (, min(F(, ) \ )),
(b) 0 (, ) = 0 (min(F(, ) \ ), ) 0 (, min(F(, ) \ )).
2.1. Walks on countable ordinals and their basic characteristics 23

Proof. The proof is again by induction on . Let 1 = min(F(, ) \ ) and


1 = min(C \ ). Let 1 be the minimal (C ) {min(C \ )} such that
1 F(, ) (or F(, )).
Applying the inductive hypothesis to 1 (or 1 ) and noting
that 1 = min(F(1 , ) \ ) (or 1 = min(F(, 1 ) \ )), we get

0 (, ) = 0 (1 , ) 0 (, 1 ), (2.1.1)

the part (a) of Lemma 2.1.9. Note that depending on whether 1 or


1 , we have that F(1 , ) F(, ) or F(, 1 ) F(, ) because in the
latter case 1 is also equal to min(C \ ). Dene

2 := min(F(1 , ) \ ) or 2 := min(F(, 1 ) \ ),

depending on whether 1 or 1 , respectively. Note that in any


of the cases of the denition of 2 , we have that 2 1 . Applying the inductive
hypothesis to 1 (or 1 ) we get:

0 (, ) = 0 (2 , ) 0 (, 2 ), (2.1.2)

0 (, 1 ) = 0 (2 , 1 ) 0 (, 2 ). (2.1.3)

Comparing (2.1.1) and (2.1.2) we see that 0 (, 1 ) is a tail of 0 (, 2 ), so the


equation (2.1.3) can be rewritten as

0 (, 1 ) = 0 (2 , 1 ) 0 (1 , 2 ) 0 (, 1 ) = 0 (1 , 1 ) 0 (, 1 ). (2.1.4)

Finally since 0 (, ) = |C | 0 (, 1 ) = 0 (1 , ) 0 (, 1 ) holds, the


equation (2.1.4) gives us the desired conclusion (b) of Lemma 2.1.9. 

Denition 2.1.10. Recall that right lexicographical ordering on < denoted by <r
refers to the total ordering dened by letting s <r t if s is an end-extension of t,
or if s(i) < t(i) for i = min{j : s(j) = t(j)}. Using the identication between a
countable ordinal and the -sequence (0 ) of elements of < , the ordering <r
induces the following ordering <0 on 1 :

<0 i 0 (, ) <r 0 (, ), (2.1.5)

where = (, ) = min{ min{, } : 0 (, ) = 0 (, )}.5


Lemma 2.1.11. The cartesian square of the total ordering <0 of 1 is the union
of countably many chains.
5 Notethat when the ordinal is equal to min{, } then max{, } <0 min{, } which is a
consequence of the fact that the empty sequence is the <r -maximal element of < .
24 Chapter 2. Walks on Countable Ordinals

Proof. It suces to decompose the set of all pairs (, ) where < . To each such
pair we associate a hereditarily nite set p(, ) which codes the nite structure
obtained from F(, ) {} by adding relations that describe the way 0 acts on
it. To show that this parametrization works, suppose we are given two pairs (, )
and (, ) such that
p(, ) = p(, ) and <0 .
We must show that 0 . Let

= min(F(, ) \ (, )), and


= min(F(, ) \ (, )).

Note that F(, ) (, ) = F(, ) (, ) so and correspond to each


other in the isomorphism of the (, ) and (, ) structures. It follows that:

0 ( , ) = 0 ( , )(= t ), (2.1.6)

0 ( , ) = 0 ( , )(= t ). (2.1.7)
Applying Lemma 2.1.9, we get:

0 ((, ), ) = t  0 ((, ), ), (2.1.8)

0 ((, ), ) = t  0 ((, ), ). (2.1.9)


It follows that 0 ((, ), ) = 0 ((, ), ). Applying Lemma 2.1.9 for
and and the ordinal (, ) we get:

0 ((, ), ) = t  0 ((, ), ), (2.1.10)

0 ((, ), ) = t  0 ((, ), ). (2.1.11)


It follows that 0 ((, ), ) = 0 ((, ), ). This shows (, ) (, ). A
symmetrical argument shows the other inequality (, ) (, ). It follows
that

(, ) = (, )(= ).
) <r 0 (,
Our assumption is that 0 (, ) and since these two sequences have
t as common initial part, this reduces to
) <r 0 (,
0 (, ). (2.1.12)

On the other hand t is a common initial part of 0 (, ) and 0 (,


), so their
lexicographical relationship depends on their tails which by (2.1.10) and (2.1.11)
) and 0 (,
are equal to 0 (, ) respectively. Referring to (2.1.12) we conclude

that indeed 0 (, ) <r 0 (, ), i.e., <0 . 
The following results show that the ordering C(0 ) = (1 , <0 ) is a canonical
object in the class of uncountable linear orderings.
2.1. Walks on countable ordinals and their basic characteristics 25

Theorem 2.1.12. Under MA1 , the ordering C(0 ) is a minimal uncountable linear
ordering in the sense that every linear ordering that embeds into C(0 ) but is not
equivalent6 to it must be countable.
Proof. Let be a given uncountable subset of 1 . We need to construct a mapping
f : 1 that is strictly increasing relative to the ordering <0 . Choose a
suciently fast closed and unbounded subset D of 1 . For < 1 , let () denote
the maximal element of D that is smaller than or equal to and let + () denote
the minimal member of D strictly above . Let P be the collection of all nite
partial mappings p from 1 into such that:
(1) <0 in dom(p) implies p() <0 p(),
(2) (p()) = + () for all dom(p),
(3) ((, )) = ((p(), p())) for all = dom(p),
(4) p preserves splitting patterns in the sense that (, ) > (, ) if and only
if (p(), p()) > (p(), p()) for every triple , , dom(p).
We consider P as a partially ordered set under the inclusion ordering and we
claim that P satises the countable chain condition. So let p ( < 1 ) be a given
sequence of elements of P. Going to a subsequence, we assume that p ( < 1 )
form an increasing -system7 of functions with some root function r. The best
way to view p s is as nite partial functions from the tree8
T (0 ) = {(0 )  : < 1 },
identifying a countable ordinal with the function 0 : . So each p
generates two nite -closed subtrees of T (0 ), one with its domain and the other
with its range. Going to a subsequence of p ( < 1 ), we may further assume that
for every < < 1 , the pair of subtrees associated with p is isomorphic to the
pair of subtrees associated with p as witnessed by the natural order isomorphism
between the two nite sets of ordinals dom(p ) and dom(p ) that xes dom(r).
Note that by Lemma 2.1.11 not only the square of C(0 ) but also all of its nite
cartesian powers can be covered by countably many chains. So identifying p s
with the naturally chosen elements of C(0 )2k , where k is the common cardinality
of dom(p ) \ dom(r), and going to an uncountable subsequence of p ( < 1 ), we
may assume that it forms a chain in C(0 )2k which means that p p satises
(1) and (2) for all < < 1 . Referring to the proof of Lemma 2.1.11, we can
nd < < 1 (and in fact an uncountable subsequence of such ordinals) such
that for a single ordinal ,
( (i), (i)) = (p ( (i)), p ( (i))) =
6 Recall that for two linear orderings K and L, one writes K L if there is a strictly increasing

mapping from K into L. Recall also that two linear orderings K and L are equivalent, in notation
K L, if K L and L K.
7 Here, we use the well-known -System Lemma (see, for example, [64, II.1.5]).
8 We consider T ( ), as well as any other tree whose members are transnite sequences of objects,
0
ordered by the natural initial-segment relation . By <, we denote the strict version of .
26 Chapter 2. Walks on Countable Ordinals

for all i = 1, . . . , k, where (1), . . . , (k) and (1), . . . , (k) are increasing
enumerations of dom(p ) \ dom(r) and dom(p ) \ dom(r), respectively. It should
be clear that this means that p p satises (3) and (4) as well.
If we choose the closed unbounded set D to be the trace on 1 of some
continuous -chain of countable elementary submodels of (H(1 ), ), one easily
checks that for all < 1 , the set

D = {p P : dom(p)}

is a dense subset of P. So an application of MA1 gives us a mapping f : 1


that is strictly increasing relative to the ordering <0 . 
Corollary 2.1.13. Under MA1 , every uncountable linear ordering L whose carte-
sian square is the union of countably many chains contains an isomorphic copy of
the ordering C(0 ) or its reverse C(0 ) .

Proof. Choose an arbitrary one-to-one mapping f : C(0 ) L. Let P be the set


of all nite subsets of C(0 ) on which f is strictly increasing. We consider P as
a partially ordered set ordered by the inclusion. If P satises the countable chain
condition, then an application of MA1 would give us an uncountable subset K of
C(0 ) on which f is strictly increasing. It follows that K L. Since by Theorem
2.1.12, C(0 ) K, this gives us the conclusion of the corollary. So we are left with
analysing the case when P fails to satisfy the countable chain condition. So let X
be a given uncountable subset of P consisting of pairwise incompatible members
of P. Applying the -system lemma and noting that the root cant contribute
to the incompatibilities of members of P, subtracting it from each member of
P, we may assume that X consists of pairwise-disjoint nite subsets of C(0 )
all of some xed size k 1. We show that in this case there is an uncountable
subset K on which f is decreasing. Recall that C(0 ) can naturally be viewed
as a subset of the lexicographically ordered tree T (0 ). Similarly, the ordering
L can be identied with a subset of a lexicographically ordered Aronszajn tree
T . So considering the -closures of elements of X as well as their f -images, we
perform a -system argument on them, obtaining an uncountable subfamily X0
of X where the corresponding two families of trees form -systems with roots
R and S, respectively. Moreover, as in the previous proof, we assume that the
corresponding structures are isomorphic via natural isomorphisms. Note now that
since for two dierent elements p and q of X0 the mapping f is not increasing on
the union p q, there must be p and q occupying the same position in
the increasing enumerations of p and q, respectively such that

<0 i f () >0 f (). (2.1.13)

As above, we may identify a p X0 with the k-tuple p(1), . . . , p(k) of


elements of C(0 ) that enumerates p increasingly according to the ordering of
C(0 ). Similarly, we identify the image f  p with the k-tuple f (p(1)), . . . , f (p(k))
2.1. Walks on countable ordinals and their basic characteristics 27

of elements of L. So going to an uncountable subfamily of X0 , we may assume that


{p(1), . . . , p(k) : p X0 } is a chain in C(0 )k and that {f (p(1)), . . . , f (p(k)) :
p X0 } is a chain in Lk . Using 2.1.13 it follows now that f is decreasing on
K = {p(1) : p X0 }. Combining this with Theorem 2.1.12, we get the desired
conclusion C(0 ) K L of the corollary. 
Remark 2.1.14. It follows that, up to the equivalence, the ordering C(0 ) and its
reverse C(0 ) will appear in any basis9 of the class of uncountable linear orderings
as long as the axiom used in nding the basis includes the weak forms of MA1
used in the proofs of Theorem 2.1.12 and Corollary 2.1.13. In fact, the combination
of the results of [10] and [77] shows that under the Proper Forcing Axiom,10 the
list
1 , 1 , B, C(0 ), C(0 ) (2.1.14)
form a basis for the class of all uncountable linear orderings, where B is any set of
reals of cardinality 1 . We shall return to this subject matter in Section 4.4 below.
Notation 2.1.15. Well-ordered sets of rationals. The set < ordered by the right
lexicographical ordering <r is a particular copy of the rationals of the interval
(0, 1] which we are going to denote by Qr or simply by Q. The next lemma shows
that for a xed , the map  0 (, ) is a strictly increasing function from
into Qr . We let (0 ) , or simply 0 , denote this function and we will identify it
with its range, i.e., view it as a member of the tree Qr of all well-ordered subsets
of Qr , ordered by end-extension.
Lemma 2.1.16. 0 (, ) <r 0 (, ) whenever < < . 
It follows that the sequence 0 : Qr ( < 1 ) of mappings that one
associates to 0 denes the downwards closed subtree

T (0 ) = {(0 )  : < 1 }

of Qr . Note that by Lemma 2.1.9, for a xed and arbitrary , the restriction
(0 )  is determined by the way (0 ) acts on the full lower trace F(, ), a
nite subset of . Hence all levels of T (0 ) are countable, and therefore T (0 ) is
a particular example of an Aronszajn tree, a tree of height 1 with all levels and
chains countable.
Theorem 2.1.17. The subtree T (0 ) of Qr admits a strictly increasing map into
the rationals.11 
This is easily deducible from the following fact about 0 that is of independent
interest.
9 Recall that a class B of linear orderings is a basis for a class X of linear orderings if for every
L X there is K B such that K L. For example, Corollary 2.1.13 shows that under MA1 ,
the class of all uncountable linear orderings whose cartesian squares can be covered by countably
many chains has a two-element basis C(0 ), C(0 ) .
10 The statement of this axiom and some of its applications can be found for example in [13].
11 Trees of this kind are also known under the name special Aronszajn trees
28 Chapter 2. Walks on Countable Ordinals

Lemma 2.1.18. { < : 0 (, ) = 0 (, )} is a closed subset of whenever


< .

Proof. Let < be a given accumulation point of this set and let = 0 > >
n = and = 0 > > m = be the traces of the walks and ,
respectively. Being an accumulation point of the set { < : 0 (, ) = 0 (, )},
the ordinal is in particular a limit ordinal, so we can nd an ordinal < such
that 0 (, ) = 0 (, ) and

> max(Ci ) and > max(Cj ) for all i < n and j < m. (2.1.15)

Note that this in particular means that the walk is a common tail of
the walks and . Subtracting 0 (, ) from 0 (, ) we get 0 (, )
and subtracting 0 (, ) from 0 (, ) we get 0 (, ). It follows that 0 (, ) =
0 (, ). 

It follows that T (0 ) does not branch at limit levels. Now note the general
fact that any downward closed subtree of Q which is nitely branching at limit
nodes admits a strictly increasing mapping into Q, so the tree T (0 ) is a special
subtree of Q.
Denition 2.1.19. Identifying the power set of Q with the particular copy 2Q of
the Cantor set, dene for every countable ordinal ,

G = {x 2Q : x end-extends no 0  for }.

Lemma 2.1.20. G ( < 1 ) is an increasing sequence of proper G -subsets of the


Cantor set whose union is equal to the Cantor set. 
Lemma 2.1.21. The set X = {0 : < 1 } considered as a subset of the Cantor
set 2Q has universal measure zero.

Proof. Let be a given nonatomic Borel measure on 2Q . For t T (0 ), set

Pt = {x 2Q : x end-extends t}.

Note that each Pt is a perfect subset of 2Q and therefore is -measurable. Let

S = {t T (0 ) : (Pt ) > 0}.

Then S is a downward closed subtree of Q with no uncountable antichains. By


an old result of Kurepa (see [105]), no Souslin tree admits a strictly increasing
map into the reals (as for example Q does). It follows that S must be countable
and so we are done. 
2.2. The coherence of maximal weights 29

2.2 The coherence of maximal weights


Denition 2.2.1. The maximal weight of the walk is the characteristic
1 : [1 ]2
dened recursively by
1 (, ) = max{|C |, 1 (, min(C \ ))},
with the boundary value 1 (, ) = 0 for all . Thus 1 (, ) is simply the maximal
integer appearing in the sequence 0 (, ) dened above.
Lemma 2.2.2. For all < < 1 and n < ,
(a) { : 1 (, ) n} is nite,
(b) { : 1 (, ) = 1 (, )} is nite.
Proof. The proof is by induction. To prove (a) it suces to show that for every
n < and every A of order-type there is A such that 1 (, ) > n.
Let = sup(A). If = one chooses arbitrary A with the property that
|C | > n, so let us consider the case < . Let 1 = min(C \ ). By the
inductive hypothesis there is A such that:
> max(C ), (2.2.1)
1 (, 1 ) > n. (2.2.2)

Note that 0 (, ) = |C | 0 (, 1 ), and therefore
1 (, ) 1 (, 1 ) > n.
To prove (b) we show by induction that for every A of order-type there
exists A such that 1 (, ) = 1 (, ). Let = sup(A) and let 1 = min(C \).
Let n = |C | and let
B = { A : > max(C ) and 1 (, 1 ) > n}.
Then B is innite, so by the induction hypothesis we can nd B such that
1 (, ) = 1 (, 1 ). Then
1 (, ) = max{n, 1 (, 1 )} = 1 (, 1 ),
so we are done. 
Denition 2.2.3. Consider the following linear ordering <1 on 1 that is now
based on the mapping 1 rather than 0 :
<1 i 1 (, ) <r 1 (, ), (2.2.3)
where = (, ) = min{ < min{, } : 1 (, ) = 1 (, )}.12
12 When 1 (, ) = 1 (, ) for all < min{, } then it is understood that min{, } <1
max{, }.
30 Chapter 2. Walks on Countable Ordinals

Lemma 2.2.4. The cartesian square of the total ordering <1 of 1 is the union of
countably many chains.
Proof. It suces to decompose {,  : < < 1 } into countably many chains.
For < < 1 , let

D = { : = or < and 1 (, ) = (, )},

and let n be the maximal value 1 (, ) or 1 (, ) for in D . Let

F = { : = or < and 1 (, ) n, or 1 (, ) n, }.

Note that D F . The lemma is proved once we show that, for some <
and < , we have <1 and

n = n = n, 1  F
= 1  F 13 , 1  F
= 1  F 14 , (2.2.4)

and then prove that <1 . For this we rst observe that (, ) = (, ) is
an easy consequence of 2.2.4. Let denote the common value of the function .
If F \ F , then

1 (, ) = 1 (, ) > n 1 (, )

which means that <1 . If on the other hand F \ F , then

1 (, ) > n 1 (, )

contradicting our assumption <1 . If


/ F F , then

1 (, ) = 1 (, ) < 1 (, ) = 1 (, )

which gives us the desired inequality <1 . 


Let C(1 ) denote the linearly ordered set (1 , <1 ). Then working as in the
case of the ordering <0 , we have the following result showing that C(1 ) also
serves as a basis member for the class of uncountable linear orderings.
Theorem 2.2.5. Assuming MA1 , the ordering C(1 ) and its reverse C(1 ) serve
as a two-element basis for the class of all uncountable linear orderings whose carte-
sian squares can be covered by countably many chains. 
We now note the following unboundedness property of 1 which is an imme-
diate consequence of Lemma 2.2.2 and which will be elaborated on in later sections
of this book.
13 Here, 1 : are dened from 1 in the usual way, 1 () = 1 (, ).
14 Here the isomorphism relation = between two nite partial functions p and q from 1 into
refers to the fact that there is an order-preserving bijection between their domains transfering
one function to the other.
2.2. The coherence of maximal weights 31

Lemma 2.2.6. For every pair A and B of uncountable families of pairwise-disjoint


nite subsets of 1 and every positive integer n, there exist uncountable subfamilies
A0 A and B0 B such that, for every pair a A0 and b B0 such that a < b,15
we have 1 (, ) > n for all a and b. 
Corollary 2.2.7. The tree
T (1 ) = {t : : < 1 , t = 1 }16
is a coherent, homogeneous, Aronszajn tree that admits a strictly increasing map
into the reals. 
Denition 2.2.8. Consider the following extension of T (1 ):

T(1 ) = {t : : < 1 and t  T (1 ) for all < }.

If we order T(1 ) by the right lexicographical ordering <r we get a complete


linearly ordered set. It is not continuous as it contains jumps of the form
[t m, t m + 10],
where t T (1 ) and m < . Removing the right-hand points from all the jumps
 1 ).
we get a linearly ordered continuum which we denote by A(
 1 ) is a homogeneous nonreversible ordered continuum that can
Lemma 2.2.9. A(
be represented as the union of an increasing 1 -sequence of Cantor sets.
Proof. For a countable limit ordinal > 0, set
 (1 ) = {t A(
A  1 ) : l(t) }.

Then each A  1 ) with the level T (1 ) being a count-


 (1 ) is a closed subset of A(

able order-dense subset. One easily concludes from this that A  (1 ) ( limit < 1 )
is the required decomposition of A(  1 ).
To show that A(  1 ) is homogeneous, consider two pairs x0 < x1 and y0 < y1
of non-endpoints of A(  1 ). Choose a countable limit ordinal strictly above the
lengths of x0 , x1 , y0 and y1 . So we have a Cantor set A (1 ), an order-dense subset
 (1 ), and two pairs of points x0 < x1 and y0 < y1 in
T (1 ) of A
 (1 ) \ T (1 ).
A
 (1 ) A
By Cantors theorem there is an order isomorphism : A  (1 ) such
that:
 T (1 ) = T (1 ), (2.2.5)
(xi ) = yi for i < 2. (2.2.6)
15 For sets of ordinals a and b we write a < b whenever < for all a and b.
16 Here, = denotes the fact that the functions agree on all but nitely many arguments.
32 Chapter 2. Walks on Countable Ordinals

 1 ) by the formula
Extend to the rest of A(

(t) = (t  ) t  [, l(t)).

It is easily checked that the function is the required order-isomorphism.


Let us now prove that there is no order-reversing bijection
 1 ) A(
: A(  1 ).

Since T (1 ) with the lexicographical ordering does not have a countable order-
dense subset, the image  T (1 ) must have sequences of length bigger than any
given countable ordinal. So for every limit ordinal we can x t in T (1 ) of
length such that (t ) also has length . Let

f () = max{ < : t () = (t )()} + 1.

By the Pressing Down Lemma17 , nd an uncountable set of countable limit


ordinals, a countable ordinal and s, t T (1 ) such that for all :


f () = , (2.2.7)

t  = s and (t )  = t. (2.2.8)
Moreover we may assume that

{t  : } and {(t )  : }

are antichains of the tree T (1 ). Pick = in and suppose for deniteness that
t <l t . Note that the ordinal (t , t ) where this relation is decided is above
and is smaller than both and . Similarly the ordinal ((t ), (t )) where the
lexicographical ordering between (t ) and (t ) is decided also belongs to the
interval
min{, }).
(,
It follows that ((t ), (t )) = (t , t )(= ) and that

(t )() = t () < t () = (t ()),

contradicting the fact that is order-reversing. 


Denition 2.2.10. The set T(1 ) has another natural structure, a topology gener-
ated by the family of sets of the form

Vt = {u T(1 ) : t  u},

for t a node of T (1 ) of successor length as a clopen subbase. Let T 0 (1 ) denote


the set of all nodes of T (1 ) of successor length. Then T(1 ) can be regarded as
17 Thisis a well-known fact saying that regressive mappings dened on large subsets of 1 (or any
other uncountable cardinal) are constant on large subsets of 1 (see, for example, [64, II.6.15]).
2.2. The coherence of maximal weights 33

the set of all downward closed chains of the tree T 0 (1 ) and the topology on T(1 )
is simply the topology one obtains from identifying the power set of T 0 (1 ) with
the cube
0
{0, 1}T (1 )
with its Tychono topology18 . T(1 ) being a closed subset of the cube is compact.
In fact T(1 ) has some very strong topological properties such as the property
that closed subsets of T(1 ) are its retracts.
Lemma 2.2.11. T(1 ) is a homogeneous Eberlein compactum19 .

Proof. The proof that T(1 ) is homogeneous is quite similar to the corresponding
part of the proof of Lemma 2.2.9. To see that T(1 ) is an Eberlein compactum,
i.e., that the function space C(T(1 )) is weakly compactly generated, let {Xn } be
a countable antichain decomposition of T 0 (1 ) and consider the set

K = {2n Vt : n < , t Xn } { }.

Note that K is a weakly compact subset of C(T(1 )) which separates the points
of T(1 ). 
In the next application the coherent sequence 1 : ( < 1 ) of
nite-to-one maps needs to be turned into a coherent sequence of maps that are
actually one-to-one. One way to achieve this is via the following formula:

1 (, ) = 21 (,) (2 |{ : 1 (, ) = 1 (, )}| + 1).

Lemma 2.2.12.
(a) 1 (, ) = 1 (, ) for all < < < 1 .
(b) { : 1 (, ) = 1 (, )} is nite for all < < 1 .
Proof. Suppose < and that 1 (, ) = 1 (, ). Then 1 (, ) =
1 (, ) = n and

|{ : 1 (, ) = n}| = |{ : 1 (, ) = n}|.

Since the set on the left-hand side is an initial segment of the set of the right-hand
side, the sets must in fact be equal. It follows that = . This proves (a).
To prove (b), by Lemma 2.2.2 the set

D = { : 1 (, ) = 1 (, )}
18 This is done by identifying a subset V of T 0 (1 ) with its characteristic function V : T 0 (1 )
2.
19 Recallthat a compactum X is Eberlein if its function space C(X) can be generated by a subset
which is compact in the weak topology.
34 Chapter 2. Walks on Countable Ordinals

is nite. Let
m = max{(, ) : D , {, }} + 1.
By Lemma 2.2.2, the proof of (b) is nished if we show that for < , if 1 (, ) >
m and 1 (, ) > m, then 1 (, ) = 1 (, ). From the denition of m, we must
have that j = 1 (, ) = 1 (, ) > m, and that

{ : (, ) = i} = { : (, ) = i}.

It follows that if j denotes the cardinality of this set, then

1 (, ) = 2i (2j + 1) = 1 (, ).

This completes the proof. 


Dene 1 from 1 just as 1 was dened from 1 ; then the 1 s are one-
to-one and coherent. From 1 one also has a natural sequence r ( < 1 ) of
elements of dened as

r (n) = |{ : 1 (, ) n}|.

Note that r eventually strictly dominates r whenever < .


Denition 2.2.13. The sequences e = 1 ( < 1 ) and r ( < 1 ) can be used
in describing a functor
G  G ,
which to every graph G on 1 associates another graph G on 1 as follows:
1
{, } G i {e1
(l), e (l)} G (2.2.9)

for all l < (r , r ) for which these preimages are both dened and dierent20 .
Lemma 2.2.14. Suppose that every uncountable family F of pairwise-disjoint nite
subsets of 1 contains two sets A and B such that A B G21 . Then the same is
true about G provided the uncountable family F consists of nite cliques22 of G .
Proof. Let F be a given uncountable family of pairwise-disjoint nite cliques of
G . We may assume that all members of F are of some xed size k.
Consider a countable limit ordinal > 0 and an A in F with all its elements
above . Let n = n(A, ) be the minimal integer such that for all < in A {}:

(r , r ) < n, 23 (2.2.10)
20 As it will be clear from the proof of the following lemma, the functor G G can equally
be based on any other coherent sequence e : ( < 1 ) of one-to-one mappings and any
other sequence r ( < 1 ) of pairwise distinct reals.
21 Here, A B = {{, } : A, B, = }.
22 A clique of G is a subset C of with the property that [C]2 G .
1
23 Recall that for s = t , we let (s, t) = min{k : s(k) = t(k)}.
2.2. The coherence of maximal weights 35

e(, ) = e(, ) for some implies e(, ), e(, ) n. (2.2.11)


Let

H(, A) = { 1 : e(, ) n(, A) for some from A {}}.


A) of H(, A), the sequence
Taking the transitive collapse H(,

e  H(, A) ( A)

collapses to a k-sequence s(, A) of mappings with integer domains. Let r(, A)


denote the k-sequence
r  (n(, A) + 1) ( A)
enumerated in increasing order. Hence every A F above generates a quadruple

(H(, A) , r(, A), s(, A), n(, A))

of parameters. Since the set of quadruples is countable, by the assumption on the


graph G we can nd two distinct members A and B of F above that generate
the same quadruple of parameters denoted by

(H , r , s , n ).

Moreover, we choose A and B to satisfy the following isomorphism condition.


For every l n , i < k, if is the ith member of A , the
1
ith member of B and if e1 (l) and e (l) are both dened, (2.2.12)
then they are G-connected to the same elements of H .
Let m be the minimal integer > n such that

(r , r ) < m for all A and B . (2.2.13)

Let
I = { : e(, ) m for some in A B }.
Let p be the k-sequence r  m ( A ) enumerated in increasing order and
similarly let q be the k-sequence r  m ( B ) enumerated increasingly. Let
t and u be the transitive collapse of e  I ( A ) and e  I ( B ),
respectively. By the Pressing Down Lemma there is an unbounded 1 and
tuples
(H, r, s, n) and (I, p, q, t, u
, m)
such that for all :

(H , r , s , n ) = (H, r, s, n), (2.2.14)

(I , p , q , t , u
, m ) = (I, p, q, t, u
, m). (2.2.15)
36 Chapter 2. Walks on Countable Ordinals

Moreover we assume the following analogue of (2.2.12) where k1 = |I \| for some


(equivalently all) :
For every , and i < k1 , if is the ith member
of I and if is the ith member of I in their increasing
enumerations, then and are G-connected to the same
(2.2.16)
elements of the root I, and if happens to be the i th
member of A (or B ) for some i < k, then must be the
i th member of A (B resp.) and vice versa.
By our assumption about the graph G there exist < in such that:

max(I ) < , (2.2.17)

(I \ ) (I \ ) G. (2.2.18)
The proof of Lemma 2.2.14 is nished once we show that

A B G . (2.2.19)

Let i, j < k be given and let be the ith member of A and let be the jth
member of B .
Case 1: i = j. Pick an l < (r , r ). By (2.2.10), l < n. Assume that e1
(l) and
e1
(l) are both dened.
1
Subcase 1.1: e1 1
(l) < and e (l) < . By the rst choice of parameters, e (l)
1
and e (l) are members of the set H which is an initial part of H(, A ) and
H(, B ). Therefore, we have that the behavior of e and e on H is encoded by
the ith and jth term of the sequence s, respectively. In particular, we have

e1 1
(l) = e  (l), (2.2.20)

where  =the jth member of A . By our assumption that [A ]2 G we infer


that {,  } G . Referring to the denition (2.2.9) we conclude that e1
(l) and
e1
(l) must be G-connected if they are dierent.
1
Subcase 1.2: e1
(l) < and e (l) I \ . Let

 = the jth member of B .

By the choice of parametrization, the position of e1  (l) in I is the same as the


1
position of e (l) in I so by (2.2.16) their relationship to the point e1 (l) of the
root I is the same. Similarly, by the choice of the rst set of parameters, letting
 be the jth member of A , the position of e1 1
 (l) in H(, B ) and e  (l) in
H(, A ) is the same, so from (2.2.12) we conclude that their relationship with
e1
(l) is the same. However, we have checked in the previous subcase that e (l)
1
1
and e  (l) are G-connected in case they are dierent.
2.2. The coherence of maximal weights 37

1
Subcase 1.3: e1
(l) I \ and e (l) < . This is essentially symmetric to the
previous subcase.
1 1
Subcase 1.4: e1 1
(l) I \ and e (l) I \ . The fact that {e (l), e (l)} G
in this case follows from (2.2.18).
Case 2: i = j. Consider an l < (r , r ). Note that now we have l < m (see
(2.2.13)).
1 1
Subcase 2.1: e1 1
(l) < and e (l) < . Then e (l) and e (l) are elements of
I which is an initial part of both I and I . Therefore the jth (= ith) term of u
encodes both e  I and e   I (see the denition of  above). It follows that

e1 1
(l) = e  (l).

If l n then e1
 (l) belongs to H and since the jth(= ith) term of s
encodes both
e  H and e  H, it follows that
1
e1
(l) = e  (l).

1
If l > n then e1
(l) and e  (l) are not members of H so from (2.2.11) and the
denition of
H(, A ) = H = H(, B )
we conclude that
1 1
e (e1 1
(l)) = e (e (l)) = l = e  (e  (l)) = e (e  (l)).

Since e is one-to-one we conclude that


1 1
e1
(l) = e  (l) = e (l).

1
Subcase 2.2: e1 
(l) < and e (l) I \ . Recall that is the ith (= jth)
member of B , so by the rst choice of parameters the relative position of e1  (l)
in H(, B ) must be the same as the relative position of e1 (l) in H(, A ), i.e.,
it must belong to the root H. Note that if j is the position of in I , then j
must also be the position of  in I . It follows that the j th term of u
encodes
both
e  I and e   I = e   I.
So it must be that e1
(l) belongs to the root I, a contradiction. So this subcase
never occurs.
1
Subcase 2.3: e1
(l) I \ and e (l) < . This is symmetric to the previous
subcase, so it also never occurs.
1 1
Subcase 2.4: e1 1
(l) I \ and e (l) I \ . Then {e (l), e (l)} G follows
from (2.2.18).
This completes the proof of Lemma 2.2.14. 
38 Chapter 2. Walks on Countable Ordinals

Lemma 2.2.15. If there is uncountable 1 such that []2 G , then 1 can


be decomposed into countably many sets such that []2 G.
Proof. Fix an uncountable 1 such that []2 G . For a nite binary sequence
s of length equal to some l + 1, set

s = { < 1 : e(, ) = l for some in with s r }.

Then the sets s cover 1 and [s ]2 G for all s. 


Remark 2.2.16. Let G be the comparability graph of some Souslin tree T . Then
for every uncountable family F of pairwise-disjoint cliques of G (nite chains of T )
there exist A = B in F such that AB is a clique of G (a chain of T ). However, it is
not hard to see that G fails to have this property (i.e., the conclusion of Lemma
2.2.14). This shows that some assumption on the graph G in Lemma 2.2.14 is
necessary. There are indeed many graphs that satisfy the hypothesis of Lemma
2.2.14. Many examples appear when one is trying to apply Martins axiom to some
Ramsey-theoretic problems. Note that the conclusion of Lemma 2.2.14 is simply
saying that the poset of all nite cliques of G is ccc, while its hypothesis is a bit
stronger than the fact that the poset of all nite cliques of G is ccc in all of its
nite powers. Applying Lemma 2.2.15 to the case when G is the incomparability
graph of some Aronszajn tree, we see that the statement saying that all Aronszajn
trees are special is a purely Ramsey-theoretic statement in the same way Souslins
hypothesis is.
We nish this section by showing that the assumption (c) on a given C-
sequence C ( < 1 ) on which we base our walks, and the function 1 , is not
sucient to give us the stronger conclusion that the tree T (1 ) can be decomposed
into countably many antichains (see Corollary 2.2.7). To describe an example, we
need the following notation, given that we have xed one C-sequence C ( < 1 ).
Let C (0) = 0 and for 0 < n < , let C (n) denote the nth element of C according
to its increasing enumeration with the convention that C+1 (n) = for all n > 0.
We assume that the C-sequence is chosen so that for a limit ordinal > 0 and a
positive integer n, the ordinal C+1 (n) is at least n + 1 steps away from the closest
limit ordinal below it. For a set D of countable ordinals, let D0 denote the set
of successor ordinals in D. We also x, for each countable ordinal a one-to-one
function e whose domain is the set 0 of all successor ordinals < and range
included in , which cohere in the sense that e () = e () for all but nitely
many successor ordinals 0 0 . For each r ([]<) , we associate another
C-sequence Cr ( < 1 ) by letting Cr = C [r , ) n Dr (n), where

Dr (n) = { [C (n), C (n + 1))0 : e () r(n)},



and where r = sup( n Dr (n)). (This denition really applies only when
is a limit ordinal; for successor ordinals we put C+1 r
= {}.) This gives us a
C-sequence C ( < 1 ) and we can consider the corresponding r1 : [1 ]2
r
2.2. The coherence of maximal weights 39

as above. This function will have the properties stated in Lemma 2.2.2, so the
corresponding tree
T (r1 ) = {(r1 )  : < 1 }
is a coherent tree of nite-to-one mappings that admits a strictly increasing func-
tion into the real line. The following fact shows that T (r1 ) will typically not be
decomposable into countably many antichains.
Lemma 2.2.17. If r is a Cohen real, then T (r1 ) has no stationary antichain.
Proof. The family []< of nite subsets of equipped with the discrete topology
and its power ([]< ) with the corresponding product topology and the given
Cohen real r is to meet all dense G subsets of this space that  one can explicitly
dene. Thus in particular, r = , and therefore, Cr = n Dr (n) for every
limit ordinal . Since the tree T (r1 ) is coherent, to establish the conclusion of the
lemma, it suces to show that for every stationary 1 there exists ,
such that (r1 ) (r1 ) . This in turn amounts to showing that for every stationary
1 , the set
U = {x ([]< ) : (, ) (x1 ) < (x1 ) }
is a dense-open subset of ([]< ) . So, given a nite partial function p from
into []< , it is sucient to nd a nite extension q of p such that the basic open
subset of ([]< ) determined by q is included in U . Let n be the minimal integer
that is bigger than all integers appearing in the domain of p or any set of the form
p(j) for j dom(p). For , set
Fn () = { 0 : e () n}.
Applying the Pressing Down Lemma, we obtain a nite set F 1 and a station-
ary set such that Fn () = F for all and such that, if = max(F )+1,
then e  = e  for all , . A similar application of the Pressing Down
Lemma will give us an integer m > n and two ordinals < in such that
(1) C (j) = C (j) for all j m,
(2) C (m + 1) > + 1, and
(3) e  (C (m) + 1)0 = e  (C (m) + 1)0 .
Extend the partial function p to a partial function q with domain {0, 1, . . . , m}
such that:
(4) q(m) = {e ( + 1)}, and
(5) q(j) = for any j < m not belonging to dom(p).
Choose any x ([]< ) extending the partial mapping q. Then from the choices
of the objects n, , F , m, , and q, we have that
(6) + 1 Cx , and
(7) Cx is an initial segment of Cx .
40 Chapter 2. Walks on Countable Ordinals

It follows that, given a < , the walk from to along the C-sequence Cx ( <
1 ) either leads to the same nite string of the corresponding weights as the walk
from to , or else it starts with two steps to + 1 and , and then follows the
walk from to . Since x1 (, ) and x1 (, ) are by denition maximums of these
two strings of weights, we conclude that x1 (, ) x1 (, ). On the other hand,
note that by (7), in the second case, the weight |Cx | of the rst step from to
is less than or equal to the weight |Cx | of the rst step from to . It follows
that we have also the other inequality x1 (, ) x1 (, ). Hence we have shown
that x1 (, ) = x1 (, ) for all < , or in other words, that (x1 ) < (x1 ) . This
nishes the proof. 

Question 2.2.18. What is the condition one needs to put on a given C-sequence
C ( < 1 ) in order to guarantee that the corresponding tree T (1 ) can be
decomposed into countably many antichains?

2.3 Oscillations of traces


In this section we shall consider the version of the oscillation mapping

osc : ([1 ]< )2 Card

dened by
osc(x, y) = |x/E(x, y)|,
where E(x, y) is the equivalence relation on x dened by letting E(x, y) if and
only if the closed interval determined by and contains no point from y \ x.
So, osc(x, y) is simply the number of convex pieces into which the set x is split by
the set y \ x. The resulting oscillation theory reproduced here in part has shown
to be a quite useful and robust coding technique. In this section, we give some
applications to this eect, and in Sections 8.1 and 9.4 below we shall extend the
oscillation theory to a more general context.

Figure 2.2: The oscillation mapping osc.


2.3. Oscillations of traces 41

Denition 2.3.1. For < < 1 , set

osc0 (, ) = osc(Tr(0 (, ), ), Tr(0 (, ), )),

where 0 (, ) = min{ : 0 (, ) = 0 (, )} 124 .

Figure 2.3: The oscillation mapping osc0 .

The following result shows that the upper traces do realize all possible oscil-
lations in any uncountable subset of 1 .
Lemma 2.3.2. For every uncountable subset of 1 and every positive integer
n, there is an integer l such that for all k < n there exist , such that
osc0 (, ) = l + k.
Proof. Fix a sequence Mi (i n) of continuous -chains of length 1 of countable
elementary submodels of (H2 , ) containing all the relevant objects such that Mi
24 Note that by Lemma 2.1.18 this is well dened.
42 Chapter 2. Walks on Countable Ordinals

is an element of the minimal model of Mi+1 for all i < n. Let Mn be the minimal
model of Mn and pick a above n = Mn 1 . Since Mn1 Mn we can nd
Mn1 Mn1 Mn such that n1 = Mn1 1 > L(n , ). Similarly, we can nd
Mn2 Mn1 Mn1 such that n2 = Mn2 1 > L(n1 , n ), and so on. This
will give us an -chain Mi (i n) of countable elementary submodels of (H2 , )
containing all relevant objects such that if we let n+1 = and i = Mi 1 for
i n, then i > L(i+1 , i+2 ) for all i < n. Pick an ordinal C0 above L(0 , ).
Then (see Lemma 2.1.6 above), i Tr(, ) for all i n and in fact 0 is the
last step of the walk from to . Let t = 0 (, ), ti = 0 (i , ) for i n. Then
tn < tn1 < < t0 < t and t = t0  |C0 |. Let
0 = { : t = 0 (, ) and 0  = 0  }.
Then 0 , and since 0  M , we have that 0 M0 . For 0 , let

T = {x [1 ]|t| : x  Tr(, ) for some }


considered as a tree ordered by the relation . Call a subset of 0 large if
T contains a subtree T whose root is {} and in which a node x of cardinality
< |t| must either have exactly one immediate successor or there is i n and
uncountably many < 1 such that x {} T and 0 (, ) = ti for all
such that x {}  Tr(, ). A simple elementarity argument using models M0
M1 Mn shows that 0 itself is large. Note also that if one decomposes
a large subset of 0 into countably many subsets, at least one of them must be
large. So we can nd an ordinal 0 in the interval [, 0 ) and a large M0 such
that 0 = 0 (, ) for all . Let s = 0 (0 , ). Since the walk from to 0
must pass through all the i (i n) we know that s  t0 = t1 = = tn . Choose
a subtree T M0 of T witnessing that is large. Pick an immediate successor
{, } of {} such that > 0 . Then (see Lemma 2.1.6) for every ,
t0 = 0 (, ) and Tr(0 , ) = Tr(0 , ) Tr(, ).
Choose in M0 a branching node x1 = {, } whose immediate successors x1 {}
T have the property that 0 (, ) = t1 for all such that x {}  Tr(, ).
Let
l = osc(Tr(0 , ) x1 , Tr(0 , )).
Choose an arbitrary non-negative integer k < n. Working in M1 , we choose an
immediate successor x1 {1 } T of x1 such that 1 > 1 Tr(0 , ) and a
branching node x2 = x1 {1 } of T corresponding to t2 . Then working in M2 , we
choose an immediate successor x2 {2 } T of x2 such that 2 > 2 Tr(0 , )
and a branching node x3 = x2 {2 } of T corresponding to t3 , and so on. After
k steps we arrive at an immediate successor xk {k } T Mk of xk such that
k > k Tr(0 , ). Working in Mk we choose such that xk {k }  Tr(, ).
Let x = Tr(0 , ), let y = Tr(0 , ), and let z = Tr(0 , ) x1 . Then
x/E(x, y) = z/E(z, y) {x2 \ x1 , x3 \ x2 , . . . , x \ xk }.
2.3. Oscillations of traces 43

It follows that osc(, ) = l + k. Since k was an arbitrary non-negative integer


n, this nishes the proof. 

It should be clear that the idea of the previous proof with very minor mod-
ications also works in proving the following multidimensional version of Lemma
2.3.2.
Lemma 2.3.3. For every uncountable family G of pairwise-disjoint nite subsets
of 1 all of some xed size m, and every positive integer n, there exist integers
li (i < m) and a, b G such that osc0 (a(i), b(i)) = li + k for all i < m.25 

Consider the following projection of the oscillation mapping osc0 ,

osc0 (, ) = max{m : 2m |osc0 (, )}.

One would need to choose a more involved projection in order to take full advan-
tage of the multi-dimensional version of Lemma 2.3.2 but already this projection
is giving us interesting results.
Corollary 2.3.4. For every uncountable 1 and every positive integer k there
exist , such that osc0 (, ) = k. 
Corollary 2.3.5. Every inner model M of set theory which correctly computes 1
contains a partition of 1 into innitely many pairwise-disjoint subsets that are
stationary in the universe V of all sets.

Proof. Our assumption about M means in particular that the C-sequence C ( <
1 ) can be chosen to be an element of M . It follows that the oscillation mapping
osc0 is also an element of M . It follows that for each < 1 , the sequence of sets

Sn = { > : osc0 (, ) = n} (n < )

belongs to M . So it suces to show that there must be < 1 such that for
every n < the set Sn is stationary in V . This is an immediate consequence of
Corollary 2.3.4. 

Remark 2.3.6. In [67], Larson showed that Corollary 2.3.5 cannot be extended to
partitions of 1 into uncountably many pairwise-disjoint stationary sets. He also
showed that under the Proper Forcing Axiom, for every mapping c : [1 ]2 1
there is a stationary set S 1 such that for all < 1 ,

{c(, ) : S \ ( + 1)} = 1 .

Note that by Corollary 2.3.4 this cannot be extended to mappings c with countable
ranges.
25 Here a(i) denotes the ith member of a according to the increasing enumeration of a.
44 Chapter 2. Walks on Countable Ordinals

The original oscillation theory has been developed in the realm of functions
with domain and ranges included on , i.e., as a theory of the oscillation mapping

osc : ( )2 + 1

that counts the number of times two such functions change in dominating each
other. Some of this will be reproduced in Section 9.4 below. In a quite similar
manner, one can also count oscillations between two nite partial functions x and
y from 1 into as follows,

osc(x, y) = |{ dom(x) : x() y() but x( + ) > y( + )}|,

where + denotes the minimal ordinal in dom(x) above if there is one.

Figure 2.4: The oscillation mapping osc1 .

Denition 2.3.7. For < < 1 , set (see Figure 2.4),

osc1 (, ) = osc(1  L(, ), 1  L(, )).


2.3. Oscillations of traces 45

We have the following analogue of Lemma 2.3.2 which is now true even in its
rectangular form.
Lemma 2.3.8. For every pair and of uncountable subsets of 1 and every
positive integer n, there is an integer l such that for all k < n there exist
and such that osc1 (, ) = l + k.
Proof. Fix a continuous -chain N of length 1 of countable elementary submodels
of (H2 , ) containing all the relevant objects and x also a countable elementary
submodel M of (H2 , ) that contains N as an element. Let = M 1 . Fix
0 \ and 0 \ . Let L0 = L(, 0 ). By exchanging the steps in the
following construction, we may assume that 1 (, 0 ) > 1 (, 0 ) for = max(L0 ).
Choose 0 C such that 0 > L0 and such that the three mappings 10 , 10 ,
and 1 agree on the interval [0 , ). By elementarity of M there will be 0+ > ,
+ + + +
0 \ 0 , and 0 \ 0 realizing the same type as , 0 , and over the
parameters L0 and 0 . Let L+ + +
1 = L(, 0 ). Note that 0 min(L1 ). It follows
+
that, in particular, the mappings 1+ and 1 + agree on L1 . Choose an N N
0 0
belonging to M such that N 1 > L+ 1 and choose 1 C \ N . Then we can nd
1 > and + +
1 \ 1 and 1 \ 1 realizing the same type as , 0 , and 0
+
+ ++
over the objects accumulated so far, L0 , 0 , L1 , N , and 1 . Let L1 = L(, 1 )
and let t+1 be the restriction of 1+ on max(L+ 1 ) + 1. Then the set of whose
1
+
1 end-extends t1 belongs to the submodel N and is uncountable, since clearly
it contains the ordinal + 1 which does not belong to N . So by the elementarity of
N and Lemma 2.2.6 there is 1 \ such that 11 end-extends t+ 1 and

1 (, 1 ) > 1 (, 1 ) for all L++


1 . (2.3.1)

Let t1 be the restriction of 11 on max(L++1 ). Then t1 M and every


whose 1 end-extends t1 satises (2.3.1) and

1 (, ) = 1 (, 1 ) for all L+
1. (2.3.2)

Note also that L(, 1 ) = L0 L+ ++


1 L1 . Using similar reasoning we can nd
t2 M T (1 ) end-extending t1 , an 2 \ whose 12 end-extends t2 ,
nite sets > max(L++ + ++
2 ) > max(L2 ) > max(L1 ), and 2 \ such that
+ ++ + ++
L(, 2 ) = L0 L1 L1 L2 L2 and such that the analogues of (2.3.1) and
(2.3.2) hold for all whose 1 end-extends t2 , and so on. This procedure
gives us a sequence L0 < L+ 1 < L1
++
< < L+
n < Ln
++
of nite subsets of ,
an increasing sequence t1 < t2 < < tn of nodes of T (0 ) M , and a sequence
k (1 k n) such that for all 1 k n, and all whose 1 end-extends tk ,

L(, k ) = L0 L+ ++
1 L1 L+ ++
k Lk , (2.3.3)

1 (, ) > 1 (, k ) for all L++


k , (2.3.4)
1 (, ) = 1 (, k ) for all L+
k. (2.3.5)
46 Chapter 2. Walks on Countable Ordinals

Choose M whose 1 end-extends tn such that the set Ln = L(, ) lies


above L++
n and such that all the 1k s agree on Ln . Note that by Lemma 2.1.6
for each 1 k n,
L(, k ) = L0 L+ ++
1 L1 L+ ++
k Lk Ln . (2.3.6)
Let l = osc(1  (L0 Ln ), 1  (L0 Ln )) where is equal to one of the
k s. Note that l does not depend on which k we choose and that according to
(2.3.4), (2.3.5) and (2.3.6), for each 1 k n, we have that osc1 (, k ) = l + k,
as required. 
Note that in the above proof the set can be replaced by any uncount-
able family of pairwise-disjoint sets giving us the following slightly more general
conclusion.
Lemma 2.3.9. For every uncountable family G of pairwise-disjoint nite subsets of
1 all of some xed size m, every uncountable subset of 1 , and every positive
integer n, there exist integers li (i < m) and a G such that for all k < n there
exist such that osc1 (a(i), ) = li + k for all i < m. 
Having in mind an application we dene a projection of osc1 as follows. For
a real x, let [x] denote the greatest integer not bigger than x. Choose a mapping
1 1
: R such that (x) = n if x [x] In , where In = ( n+3 , n+2 ]. For x in
R we use the notation x for (x). Choose a one-to-one sequence x ( < 1 ) of
elements of the rst half of the unit interval I = [0, 1] which together with the
point 1 form a set that is linearly independent over the eld of rational numbers.
Finally, set
osc1 (, ) = (osc1 (, ) x ) .
Then Lemma 2.3.9 turns into the following statement about osc1 .
Lemma 2.3.10. For every uncountable family G of pairwise-disjoint nite subsets of
1 all of some xed size m, every uncountable subset of 1 , and every mapping
h : m , there exist a G and such that osc1 (a(i), ) = h(i) for all
i < m.
Proof. Let be half of the length of the shortest interval of the form Ih(i) (i < m).
Note that for a G, the points 1, xa(0) , . . . , xa(m1) are rationally independent,
so a standard fact from number theory (see for example [46]; XXIII) gives us that
there is an integer na such that for every y Im there exist k < na such that
|k xa(i) yi | < (mod 1) for all i < m. Going to an uncountable subfamily of G,
we may assume that there is n such that na = n for all a G. By Lemma 2.3.9
there exists a G and li (i < m) such that for every k < n there is k > a such
that osc1 (a(i), k ) = li + k for all i < m. For i < m, let yi = li xa(i) and let zi be
the middle-point of the interval Ih(i) . Find k < n such that for all i < m,
|k xa(i) (zi yi )| < (mod 1).
Then osc1 (a(i), k ) = h(i) for all i < m, as required. 
2.4. The number of steps and the last step functions 47

Corollary 2.3.11. There is a regular hereditarily Lindel


of space which is not sepa-
rable.
Proof. For < 1 , let f {0, 1}1 be dened by letting f () = 1, f () = 0 for
> , and f () = min{1, osc1 (, )} for < . Consider F = {f : < 1 } as
a subspace of {0, 1}1 . Clearly F is not separable. That F is hereditarily Lindelof
follows easily from Lemma 2.3.10. 
Remark 2.3.12. The projection osc0 of the oscillation mapping osc0 appears in
[109] as the historically rst such map with more than four colors that takes all of
its values on every symmetric square of an uncountable subset of 1 . The variations
osc1 and osc1 are on the other hand very recent and are due to J.T. Moore [78]
who made them in order to obtain the conclusion of Corollary 2.3.11. Concerning
Corollary 2.3.11 we note that the dual implication behaves quite dierently since,
assuming the Proper Forcing Axiom all, hereditarily separable regular spaces are
hereditarily Lindelof (see [111]).

2.4 The number of steps and the last step functions


In this section we show that a very natural characteristic associated to the mini-
mal walks between countable ordinals lead to functions that have coherence and
nontriviality properties very much reminiscent of the Hausdor gap phenomenon
that will be a subject of our study in Section 3.1 below.
Denition 2.4.1. The number of steps of the minimal walk is the two-place function
2 : [1 ]2 dened recursively by

2 (, ) = 2 (, min(C \ )) + 1,

with the boundary condition 2 (, ) = 0 for all .


This is an interesting mapping which is particularly useful on higher cardinal-
ities and especially in situations where the more informative mappings 0 , 1 and
lack their usual coherence properties. Later on we shall devote a whole section
to 2 but here we list only few of its basic properties. We start with the coherence
property that this function enjoys.
Lemma 2.4.2. sup{|2 (, ) 2 (, )| : < } < for all < < 1 .
Proof. Suppose the conclusion of the lemma fails for some < < 1 . Then for
every k < , we can nd k < such that |2 (k , ) 2 (k , )| > k. We may
assume that the sequence of k s is strictly increasing and let = supk k . Then
is a limit ordinal , so the lower traces of walks from to and to have a
common upper bound < . Then by Lemma 2.1.6, for every ordinal [, ),
we have that

0 (, ) = 0 (, ) 0 (, ) and 0 (, ) = 0 (, ) 0 (, ). (2.4.1)
48 Chapter 2. Walks on Countable Ordinals

It follows that for every [, ),

|2 (, ) 2 (, )| |2 (, ) 2 (, )|, (2.4.2)

and so in particular, k
/ [, ) for all k such that k > |2 (, ) 2 (, )|, a
contradiction. 
We mention also the following unboundedness property of this function which
introduces another theme to be explored fully in later sections of this book.
Lemma 2.4.3. For every uncountable family A of pairwise-disjoint nite subsets
of 1 , all of some xed size n, and for every integer k, there exist an uncountable
subfamily B of A such that for all a < b in B, we have 2 (a(i), b(j)) k for all
i, j < n.26
Proof. The proof is by induction on k. So suppose that our given family A already
satises that 2 (a(i), b(j)) k for all i, j < n and all a < b in A. We shall nd
uncountable B A such that 2 (a(i), b(j)) k + 1 for all i, j < n and all a < b
in B. To this end, for each limit ordinal < 1 , we x b > in A. Then there is
< such that

0 (, ) = 0 (, ) 0 (, ) for all [ , ) and b . (2.4.3)

Find a stationary set 1 and < 1 such that = for all . Choose a
stationary subset of such that

< b < < b for all < in . (2.4.4)

Form the family A = {{} b : } of pairwise-disjoint sets of size n + 1. By


the inductive hypothesis there is uncountable B A such that 2 (a(i), b(j)) k
for all i, j < n + 1 and all a < b in B . Then the uncountable subfamily

B = {b \ {min(b)} : b B }

of A has the property that 2 (a(i), b(j)) k + 1 for all i, j < n and all a < b
in B. 
The unboundedness property of 2 given by Lemma 2.4.3 shows, in particular,
that while the ber maps 2 (, ) ( < 1 ) are at a nite distance from each other,
there is no global map with domain 1 that has a nite distance from all of these
bers.
Corollary 2.4.4. For every g : 1 there is < 1 such that

sup{|2 (, ) g())| : < } = .


26 Here,a < b signies the fact < whenever a and b, while a(i) denotes the ith
element of a relative to its increasing enumeration {a(0), a(1), . . . , a(n 1)}.
2.4. The number of steps and the last step functions 49

Proof. Otherwise we can nd an integer k and an unbounded set 1 such that


sup{|2 (, ) g())| : < } k for all . Find unbounded 1 and an
integer l such that g takes the constant value l on . Forming a family of pairs
by taking an element from and the other from and applying Lemma 2.4.3, we
can nd uncountable sets 0 and 0 such that

2 (, ) > k + l + 1 for all 0 , 0 such that < . (2.4.5)

Consider an 0 that is above 0 = min 0 . Then

k |2 (0 , ) g(0 ))| > k + l + 1 l = k + 1, (2.4.6)

a contradiction. 

The main object of study in this section, however, is another natural charac-
teristic of the minimal walk. This new mapping 3 not only has strong coherence
properties but its full analysis will lead us naturally to some ner requirements
that one can put on the given C-sequences on which we base our walks.
Denition 2.4.5. The last step function is the characteristic 3 : [1 ]2 2 of the
walk dened by letting

3 (, ) = 1 i 0 (, )(2 (, ) 1) = 1 (, ).

In other words, we let 3 (, ) = 1 just in case the last step of the walk
comes with the maximal weight.
Lemma 2.4.6. { < : 3 (, ) = 3 (, )} is nite for all < < 1 .

Proof. It suces to show that for every innite there exists such that
3 (, ) = 3 (, ). Shrinking we may assume that for some xed F (, )
and all :
= min(F (, ) \ ), (2.4.7)
1 (, ) = 1 (, ), (2.4.8)
1 (, ) > 1 (
, ), (2.4.9)
1 (, ) > 1 (
, ). (2.4.10)
It follows (see 2.1.9) that for every :

, ) 0 (, ),
0 (, ) = 0 ( (2.4.11)

, ) 0 (, ).
0 (, ) = 0 ( (2.4.12)
So for any , 3 (, ) = 1 i the last term of 0 (, )
is its maximal term i
3 (, ) = 1. 
50 Chapter 2. Walks on Countable Ordinals

The sequence (3 ) : 2 ( < 1 )27 is therefore coherent in the sense


that (3 ) = (3 )  whenever < . We need to show that the sequence is
not trivial, i.e., that it cannot be uniformized by a single total map from 1 into 2.
In other words, we need to show that 3 still contains enough information about
the C-sequence C ( < 1 ) from which it is dened. For this it will be convenient
to assume that C ( < 1 ) satises the following natural condition:
(d) If = + for some limit ordinal , then C = { + n : m < n < } for
some non-negative integer m; if is a limit of limit ordinals and if occupies
the nth place in the increasing enumeration of C , then = + m for some
limit ordinal and integer m > n.
Denition 2.4.7. Let denote the set of all countable limit ordinals and for an
integer n , let + n = { + n : }.
Lemma 2.4.8. 3 ( + n, ) = 1 for all but nitely many n with + n < .
Proof. Clearly we may assume that = + . Then there is n0 < such
that for every n n0 the walk + n passes through . By (d) we know that
C = { + n : m < n < } for some non-negative integer m, so in any such walk
+ n there is only one step from to + n. So choosing n > n0 , m, 1 (, )
we will ensure that the last step of + n comes with the maximal weight,
i.e., 3 ( + n, ) = 1. 
Lemma 2.4.9. For all < 1 , n < , the set { : + n < and
3 ( + n, ) = 1} is nite.
Proof. Given an innite subset of ( + n) we need to nd a + n such
that 3 ( + n, ) = 0. Shrinking if necessary assume that
1 ( + n, ) > n + 2
for all + n . So if 3 ( + n, ) = 1 for some + n , then the last step
of + n would have to be of weight > n + 2 which is impossible by our
assumption (d) about C ( < 1 ). 
The meaning of these properties of 3 is perhaps easier to comprehend if we
reformulate them in a way that resembles the original formulation of the existence
of Hausdor gaps. It is not surprising that this sort of variation on the classical
Hausdor gap phenomenon has appeared rst in a topological study that analyses
how the space of subuniform 28 ultralters on 1 is embedded into the space of all
ultralters on 1 .
27 Recall the way one always denes the ber functions from a two-variable function applied to

the context of 3 : (3 ) () = 3 (, ).
28 An ultralter on is subuniform if it is nonprincipal and if it concentrates on a countable
1
subset of 1 . The space SU(1 ) of subuniform ultralters on 1 is an open subspace of the

CechStone remainder of the discrete space on 1 . Lemma 2.4.10 says that there is a continuous
{0, 1}-valued function on SU(1 ) which does not extend to a continuous function dened on the

whole CechStone remainder (see [131] and [23]).
2.4. The number of steps and the last step functions 51

Lemma 2.4.10. Let B = { < : 3 (, ) = 1} for < 1 . Then:


1. B = B for < ,
2. ( + n) B is nite for all n < and < 1 ,
3. { + n : n < } B whenever + . 
In particular, there is is no uncountable 1 such that B for
all < 1 , and so the tree

T (3 ) = {(3 )  : < 1 }

contains no uncountable chains. On the other hand, the P-ideal29 I generated


by B ( < 1 ) is large as it contains all intervals of the form [, + ). The
following general dichotomy about P-ideals shows that here indeed we have quite
a canonical example of a P-ideal on 1 .
Denition 2.4.11. The P-ideal dichotomy. For every P-ideal I of countable subsets
of some set S either:
1. there is uncountable X S such that [X] I, or
2. S can be decomposed into countably many sets orthogonal to I.
Remark 2.4.12. It is known that the P-ideal dichotomy is a consequence of the
Proper Forcing Axiom and moreover that it does not contradict the Continuum
Hypothesis (see [122]). This is an interesting dichotomy which will be used in this
article for testing various notions of coherence as we encounter them. In fact, this
was the original and still most important reason for isolating this dichotomy from
the rest of the consequences of PFA (see[111], Chapter 8).
Let us now turn our attention to the lexicographical ordering given by 3 , or
more precisely, the lexicographical ordering among the bers (3 ) of 3 induced
by the global lexicographical ordering of the coherent tree T (3 ).
Denition 2.4.13. Consider the linear ordering <3 on the set of countable limit
dened as follows:
<3 i 3 (, ) <r 3 (, ), (2.4.13)
where = (, ) = min{ < min{, } : 3 (, ) = 3 (, )}.30
It turns out that chain decompositions of the cartesian square of this lexi-
cographical ordering is more closely tied to the assumption that the whole tree
T (3 ) is coverable by countably many more antichains than in the case of the
ordering <1 and the corresponding tree T (1 ). Recall that in Lemma 2.2.17 we
29 Recall that an ideal I of subsets of some set S is a P-ideal if for every sequence A (n < ) of
n
elements of I there is B in I such that An \ B is nite for all n < . A set X is orthogonal to I
if X A is nite for all A in I.
30 If (, ) = (, ) for all < min{, }, then it is understood that min{, } <
3 3 3
max{, }.
52 Chapter 2. Walks on Countable Ordinals

have shown that there could be choices of the C-sequence C ( < 1 ) for which
the corresponding tree T (1 ) is not special, and in fact contains no stationary
antichain, though the cartesian square of the ordering C(1 ) = (1 , <1 ) is al-
ways decomposable into countably many chains. A similar generic choice of the
C-sequence (satisfying the condition (d) above) will give us that the corresponding
tree T (3 ) is also not special, or more precisely, that it contains no stationary an-
tichain. Moreover, a close examination of the proof shows that if 3 is based on the
generic C-sequence, then the cartesian square of (1 , <3 ) cannot be decomposed
into countably many chains. This leads us to the following natural restriction.
Denition 2.4.14. A set of countable ordinals is special if the corresponding
subset {(3 ) : } of the tree T (3 ) can be decomposed into countably many
antichains.
Note that since the restriction of the tree T (3 ) to the set of countable
limit ordinals admits a strictly increasing real-valued function,31 every uncountable
subset of 1 contains an uncountable special subset.
Lemma 2.4.15. If is a special subset of 1 , then the cartesian square of the total
ordering <3 on is the union of countably many chains.

Proof. Note that similarly to the tree T (1 ), the restriction T (3 )  of T (3 )


to the set of countable limit ordinals admits a strictly increasing real-valued
function. It follows that if 0 denotes a set of countable limit ordinals of the form
+ , then there is a : T (3 )  0 such that a(s) = a(t) for every pair of
distinct comparable nodes of T (3 )  0 . We also assume that a is one-to-one on
levels of T (3 )  0 .
Consider a pair and of ordinals from such that < . Let be the
maximal limit ordinal , and let

D = { : = or 3 (, ) = 3 (, ) for some [, + )}.

Then D is a nite set of countable ordinals. Let



+
D = {[, + ) : D }.

+
Let p be the isomorphism type of the structure on D whose relations code the
restrictions of the bers (3 ) and (3 ) on this set, as well as the two mappings
+
 a((3 )  + ) and  a((3 )  + ) dened on the subset D of D .
Since there exist only countably many isomorphism types of these structures
and since the set is special, the conclusion of the lemma would follow once
we establish the following: Suppose that we have two pairs < and < of
ordinals from such that <3 , that p = p , and that the pair of nodes
31 Note that if the whole tree T (3 ) admits a strictly increasing real-valued function, then it is
in fact decomposable into countably many antichains.
2.4. The number of steps and the last step functions 53

(3 ) and (3 ) and the pair of nodes (3 ) and (3 ) are incomparable in T (3 ).


We shall show that <3 .
Let = (, ). Note that is not a limit ordinal and that < min{, }.
Since the mappings  a((3 )  + ) and  a((3 )  + ) and the
mappings  a((3 )  + ) and  a((3 )  + ) are isomorphic, we
conclude that the set D = D is an initial segment of both sets D and D .
Let be the maximal limit ordinal . Using the fact that the restriction of the
+ +
bers (3 ) and (3 ) on D and D , respectively, are isomorphic, we conclude
that + < for = max D. Then in particular, / D D , so the
bers (3 ) and (3 ) and the bers ( )
3 and ( )
3 have identical restrictions
on the interval [ , + ). Let D+ = {[, + ) : D}. Then we get the
equality
(3 )  D+ = (3 )  D+
+
as an immediate consequence of the fact that (3 )  D is isomorphic to (3 ) 
+
D and the fact that D is an initial segment of the sets D and D . Since
+ +
D = D+ = D , we conclude that

(3 )  \ D+ = (3 )  \ D+ = (3 )  \ D+ = (3 )  \ D+ .

It follows that (3 )  = (3 )  . Combining this with the facts that the


bers (3 ) and (3 ) and the bers (3 ) and (3 ) have identical restrictions on
the interval [ , + ), we conclude that = (, ) and that

3 (, ) = (, ) < 3 (, ) = 3 (, ).

This gives us the desired conclusion <3 that nishes the proof. 
Let C(3 ) denote the linearly ordered set (1 , <3 ). Then as in the case of
C(0 ) and C(1 ), we have the following result proved along similar lines of rea-
soning and showing that C(3 ) is yet another realization of the canonical minimal
ordering on 1 whose cartesian square is the union of countably many chains.
Theorem 2.4.16. Assuming MA1 , the ordering C(3 ) and its reverse C(3 ) serve
as a two-element basis for the class of all uncountable linear orderings whose carte-
sian squares can be covered by countably many chains. 
Chapter 3

Metric Theory of
Countable Ordinals

3.1 Triangle inequalities


In this section we study a characteristic of the minimal walk that satises certain
triangle inequalities reminiscent of those found in an ultra-metric space. Some
applications of the corresponding metric-like theory of 1 will appear already in
this section and some of them will later on get separate treatments.
Denition 3.1.1. Dene : [1 ]2 recursively by,

(, ) = max{|C |, (, min(C \ )), (, ) : C }, (3.1.1)

with the boundary condition (, ) = 0 for all .

Figure 3.1: The two triangle inequalities of .


56 Chapter 3. Metric Theory of Countable Ordinals

The following lemma lists three basic properties of the -function. As we go


on, we will discover several other properties of this function that will be closely
tied with properties of the C-sequence on which we base our walks.
Lemma 3.1.2. For all < < < 1 and n < ,
(a) { : (, ) n} is nite,
(b) (, ) max{(, ), (, )},
(c) (, ) max{(, ), (, )}.
Proof. Note that (, ) 1 (, ), so (a) follows from the corresponding prop-
erty of 1 . The proof of (b) and (c) is by a simultaneous induction on the ordinals
, and .
To prove (b), consider n = max{(, ), (, )}. We have to show that
(, ) n. Let

= min(C \ ) and = min(C \ ).

Case 1b : = . Then by the inductive hypothesis,

(, ) max{(, ), (, )}.

From the denition of (, ) we get that (, ) (, ) n, so replacing


(, ) by (, ) in the above inequality we get (, ) n. Consider an ordinal
belonging to the set C = C . By the inductive hypothesis

(, ) max{(, ), (, )}.

From the denition of (, ) we see that (, ) (, ), so replacing (, )


with (, ) in the last inequality we get that (, ) n. Since

|C | = |C | (, ) n,

referring to the denition of (, ) we conclude that (, ) n.


Case 2b : < . Then C , so

( , ) (, ) n.

By the inductive hypothesis

(, ) max{(, ), ( , )} n.

Similarly, for every C C ,

(, ) max{(, ), (, )} n.

Finally |C | |C | (, ) n. Combining these inequalities we get the


desired conclusion (, ) n.
3.1. Triangle inequalities 57

To prove (c), consider now n = max{(, ), (, )}. We have to show that


(, ) n. Let and be as above and let us consider the same two cases as
above.
Then by the inductive hypothesis
Case 1c : = = .
(, )}.
(, ) max{(, ),

This gives the desired bound (, ) n, since (, ) (, ) n and


(, ) (, ) n.
Case 2c : < . Applying the inductive hypothesis again we get

(, ) max{(, ), ( , )} n.

This completes the proof. 


The following fact shows that the function has a considerably ner coher-
ence property than 1 .
Lemma 3.1.3. If < < and if (, ) > (, ), then (, ) = (, ).
Proof. Applying Lemma 3.1.2,

(, ) max{(, ), (, )} = (, ),

(, ) max{(, ), (, )} = (, ).
Note that max{(, ), (, )} must be equal to (, ) rather than (, ) since
by the assumption (, ), which is bounded by the maximum, is bigger than
(, ). 
Remark 3.1.4. Referring to the standard denitions of metrics and ultra-metrics
the properties in Lemma 3.1.2(b) and (c) could more properly be called ultra-
subadditivities though we shall keep calling them subadditivity properties of the
function . Recall the notion of the full lower trace F (, ) given recursively by

F (, ) = F (, min(C \ )) F (, ),
C

with the boundary value F (, ) = {} for all (see Denition 2.1.7 above). The
function d : [1 ]2 dened by

d(, ) = |F (, )|

is an example of a truly subadditive function, since by Lemma 2.1.8, we have the


following inequalities whenever < < :
(a) d(, ) d(, ) + d(, ),
(b) d(, ) d(, ) + d(, ).
58 Chapter 3. Metric Theory of Countable Ordinals

Denition 3.1.5. For two distinct countable ordinals and , set

< i (, ) < (, ), (3.1.2)

where = min{ < min{, } : (, ) = (, )}; if there is no < min{, }


such that (, ) = (, )}, put min{, } < max{, }. Clearly, < is a total
ordering on 1 , so we let C() denote the uncountable linearly ordered set (1 , < ).
It should be clear that the proof of Lemma 2.2.4 also shows the following
corresponding fact about < in place of <1 .
Lemma 3.1.6. The cartesian square of the linearly ordered set C() = (1 , < ) is
the union of countably many chains. 
The following fact is also established in the same way as the corresponding
results about the orderings C(0 ) and C(1 ) showing that C() is yet another
realization of the canonical linear ordering satisfying the conclusion of the previous
lemma.
Theorem 3.1.7. Assuming MA1 , the ordering C() and its reverse C() serve as
a two-element basis for the class of all uncountable linear orderings whose cartesian
squares can be covered by countably many chains. 
There is a common property of the trees associated with the mappings 0 ,
1 , and introduced on the basis of the metric notion of a Lipschitz map that lies
behind the common properties of the corresponding linear orderings C(0 ), C(1 ),
and C(). This will be explained in Chapter 4 below.

3.2 Constructing a Souslin tree using


The purpose of this section is to give a particular application of the function to
the theory of Cohen forcing.
Denition 3.2.1. Dene : [1 ]2 as follows:

(, ) = 2(,) (2 |{ : (, ) (, )}| + 1).

Lemma 3.2.2. For all < < < 1 ,


(a) (, ) = (, ),
(b) (, ) max{(, ), (, )},
(c) (, ) max{(, ), (, )}.
Proof. Only (b) and (c) require some argument. Consider rst (b). If (, )
(, ), then (, ) max{(, ), (, )} = (, ), so

{ : (, ) (, )} { : (, ) (, )}.
3.2. Constructing a Souslin tree using 59

Therefore in this case (, ) (, ). On the other hand, if (, ) (, ),


then (, ) max{(, ), (, )} = (, ), so

{ : (, ) (, )} { : (, ) (, )}.

So in this case (, ) (, ). This checks (b). Checking (c) is similar. 


Lemma 3.2.3. (, ) = (, ) for all < < < 1 .
Proof. Suppose the conclusion fails for some triple < < < 1 . Let n =
(, ) = (, ) and write it as n = 2i (2j + 1) for some integers i and j. Then
i = (, ) = (, ) and

|{ : (, ) i}| = j = |{ : (, ) i}|. (3.2.1)

Since i = (, ), the ordinal belongs to the set { : (, ) i} and so by


the subadditivity of the set { : (, ) i} is an initial segment of the set
{ : (, ) i}. Since the two sets have the same cardinality j they must be
equal. This is a contradiction, since clearly belongs to { : (, ) i} but
not to { : (, ) i}. 
We shall also need the following property of .

Lemma 3.2.4. Suppose = < min{, } and ( , ) = ( , ) = n. Then
( , ), ( , ) > n.
Proof. By symmetry, we may assume that < . Our assumption ( , ) =
( , ) yields that for some integers i and j, we have that ( , ) = ( , ) = i
and that

|{ : (, ) i)}| = |{ : (, ) i}| = j. (3.2.2)

Note that ( , )  i, since otherwise, the set { : (, ) i} would be


a proper subset of the set { : (, ) i} contradicting (3.2.2). It follows
that k = ( , ) > i, and therefore

{ : (, ) k} { : (, ) i}, (3.2.3)

and so in particular, we get that l = |{ : (, ) k}| j. It follows that

( , ) = 2k (2l + 1) > 2i (2j + 1) = ( , ) = ( , ) = n. (3.2.4)

This together with the subadditivity properties of yields that

n < ( , ) max{( , ), ( , )} = ( , ),

n < ( , ) max{( , ), ( , )} = ( , ).
This completes the proof. 
60 Chapter 3. Metric Theory of Countable Ordinals

Denition 3.2.5. For p < dene a binary relation <p on 1 by letting <p
i < , (, ) |p| and

p( (, )) for any < such that (, ) < |p|.


(, )) = p( (3.2.5)

Lemma 3.2.6.
(a) <p is a tree-ordering on 1 of height |p| + 1,
(b) p q implies <p <q .
Proof. This follows immediately from Lemma 3.2.2. 
Denition 3.2.7. For x , set


<x = {<xn: n < }.

Lemma 3.2.8. For every p < there is a partition of 1 into nitely many
pieces such that if < belong to the same piece of the partition, then there is
q p in < such that <q .
Proof. For < 1 , the set

F|p| () = { : (, ) |p|}

has size |p| + 1, so if we enrich F|p| () to a structure that would code the
behavior of on F|p| (), one can realize only nitely many dierent isomorphism
types. Thus it suces to show that if < < 1 with F|p| (), F|p| () isomorphic,
then there is an extension q of p such that <q .
Consider the graph G of all unordered pairs of integers < (, ) of the form

{(, ), (, )}

for some < . It suces to show that every connected component of G has at
most one point < |p|. For if this is true, extend p to a mapping q : (, )
which is constant on each component of G. Clearly, <q for any such q.
Consider an integer n < |p|. We claim that the connected component of n is a
star with center n which has no other points < |p|. This will of course be sucient
for our needs here. Otherwise, pick a simple G-path n = n0 , n1 , n2 which starts
with n. Pick ordinals 0 , 1 < and i {, } for i = 0, 1, 2, 3 which witness
these connections, or in other words,

(0 , 0 ) = n0 , (0 , 1 ) = n1 , (1 , 2 ) = n1 and (1 , 3 ) = n2 . (3.2.6)

Since n0 = n1 , by Lemma 3.2.2(a), we conclude that 0 = 1 . Similarly, we


conclude that 2 = 3 .
We claim that 0 = 1 . To see this, we suppose 0 = 1 and work for a
contradiction. Since 0 and 1 are distinct members of the doubleton {, }, either
3.2. Constructing a Souslin tree using 61

3 = 0 , or else 3 = 1 . If 3 = 0 , then n2 = (1 , 3 ) = (0 , 0 ) = n0 , a
contradiction. On the other hand, if 3 = 1 , then n2 = (1 , 3 ) = (0 , 1 ) = n1
which is also a contradiction.
The non-equality 0 = 1 , the two equalities (0 , 1 ) = n1 and (1 , 2 ) = n1 ,
and the property 3.2.2(a) of , give us the missing non-equality 1 = 2 between
two consecutive i s. So, in particular, we have that 2 = 0 .
By our assumption n0 < |p|, the ordinal 0 must belong to F|p| (0 ). However,
by our assumption about the isomorphism between the structures F|p| (0 ) and
F|p| (1 ), the ordinal 0 F|p| (0 ) cannot belong to the root

F|p| (0 ) F|p| (1 ).

It follows that 0
/ F|p| (1 ), and therefore,

n1 = (0 , 1 ) > |p| > n0 .

Note that this in particular shows that no immediate G-neighbor of an integer


n = n0 < |p| is also < |p|. Applying Lemma 3.2.4 to the two equalities

(0 , 1 ) = n1 = (1 , 2 ),

we conclude that (0 , 2 ) > n1 . Since 2 = 0 , this gives us

n0 = (0 , 0 ) = (0 , 2 ) > n1 ,

a contradiction. This nishes the proof. 

Theorem 3.2.9. For every innite subset 1 , the set

G = {x : <x for some , }

is a dense open subset of the Baire space.

Proof. This is an immediate consequence of Lemma 3.2.8. 

Denition 3.2.10. For < < 1 , let < denote the fact that (, ) = (, )
for all < . Then < is a tree-ordering on 1 obtained by identifying with the
member (, ) of the tree T (). Note that <<x for all x and that there
exists x such that <x =< (e.g., one such x is the identity map id : ).

Theorem 3.2.11. If is an innite <-antichain, the set

H = {x : x for some < in }

is a dense-open subset of the Baire space.


62 Chapter 3. Metric Theory of Countable Ordinals

Proof. Let p < be given. A simple application of Ramseys theorem gives us


a strictly increasing sequence i (i < ) of members of such that (i , j )
(j , k )1 for all i < j < k. Going to a subsequence, we may assume that either
(i , i+1 ) < (i+1 , i+2 ) for all i or else for some ordinal and all i < j,
(i , j ) = . In the rst case, by Lemma 3.2.2(a), we can nd i such that one
of the integers ((i , i+1 ), i ) or ((i , i+1 ), i+1 ) does not belong to the
domain of p, so there will be q p such that (i , i+1 ) < |q| and i q i+1 . This
guarantees that i x i+1 for every x q. In the second case, (, i ) (i < ) is
a one-to-one sequence of integers so there will be i < j such that neither of the
integers (, i ) or (, j ) belongs to the domain of p, so there is q p such that
(i , j ) < |q| and i q j . Hence, i x j for every x q. 
Denition 3.2.12. For a family F of innite <-antichains, we say that a real
x is F -Cohen if x G H for all F. We say that x is F -Souslin if no
member of F is a <x -chain or a <x -antichain. We say that a real x is Souslin
if the tree ordering <x on 1 has no uncountable chains nor antichains, i.e., when
x is F -Souslin for F equal to the family of all uncountable subsets of 1 .
Note that since every uncountable subset of 1 contains an uncountable
<-antichain, if a family F renes the family of all uncountable <-antichains,
then every F -Souslin real is Souslin. The following fact summarizes Theorems
3.2.9 and 3.2.11 and connects the two kinds of reals.
Theorem 3.2.13. If F is a family of innite <-antichains, then every F -Cohen
real is F -Souslin. 
Corollary 3.2.14. If the density of the family of all uncountable subsets of 1 is
smaller than the number of nowhere dense sets needed to cover the real line, then
there is a Souslin tree. 
Remark 3.2.15. Recall that the density of a family F of innite subsets of some
set S is the minimal size of a family F0 of innite subsets of S with the property
that every member of F is rened by a member of F0 . A special case of Corollary
3.2.14, when the density of the family of all uncountable subsets of 1 is equal to
1 , was rst observed by T. Miyamoto (unpublished).
Corollary 3.2.16. Every Cohen real is Souslin.
Proof. Every uncountable subset of 1 in the Cohen extension contains an un-
countable subset from the ground model. So it suces to consider the family F of
all innite <-antichains from the ground model. 
If 1 is a successor cardinal in the constructible subuniverse, then can be
chosen to be coanalytic and so the transformation x <x will transfer combina-
torial notions of Souslin, Aronszajn or special Aronszajn trees into the correspond-
ing classes of reals that lie in the third level of the projective hierarchy. This trans-
formation has been explored in several places in the literature (see, e.g., [9], [47]).
1 For < < 1 , (, ) = min{ : (, ) = (,
)}.
3.3. A Hausdor gap from 63

Remark 3.2.17. We have just seen how the combination of the subadditivity prop-
erties (3.2.2(b),(c)) of the coherent sequence : ( < 1 ) of one-to-one
mappings can be used in controlling the nite disagreement between them. It turns
out that in many contexts the coherence and the subadditivities are essentially
equivalent restrictions on a given sequence e : ( < 1 ). For example,
the following construction shows that, in the context of nite-to-one mappings
e : ( < 1 ), coherence does lead us naturally to the two subadditivity
properties.
Denition 3.2.18. Given a coherent sequence e : ( < 1 ) of nite-to-one
mappings, dene e : [1 ]2 as
2
e (, ) = max{max{e(, ), e(, )} : , e(, ) = e(, )}.

Lemma 3.2.19. For every < < < 1 ,

(a) e (, ) e(, ),
(b) e (, ) max{e (, ), e (, )},
(c) e (, ) max{e (, ), e (, )}.

Proof. To see (a) note that it trivially holds if e(, ) = 0 so we can concentrate
on the case e(, ) > 0. Applying the convention e(, ) = 0 we see that e(, ) =
e(, ) for = . It follows that e (, ) {max{e(, ), e(, )} = e(, ).
To see (b) let n = max{e (, ), e (, )} and let be such that e(, ) =
e(, ). We need to show that e(, ) n and e(, ) n. If e(, ) > n then
since e (, ) n we must have e(, ) = e(, ) and so e(, ) = e(, ). It
follows that e (, ) e(, ) > n, a contradiction. Similarly, e(, ) > n yields
e(, ) = e(, ) = e(, ) and therefore e (, ) e(, )) > n, a contradiction.
The proof of (c) is similar. 

3.3 A Hausdor gap from


The purpose of this section is to use the -function and give a natural denition
of an (1 , 1 )-gap in the quotient algebra P()/Fin.
Denition 3.3.1. We say that a : [1 ]2 is transitive if for < < < 1 ,

a(, ) max{a(, ), a(, )}.

Transitive maps occur quite frequently in set-theoretic constructions. For


example, given a sequence A ( < 1 ) of subsets of that increases relative
2 Note the occurence of the boundary value e(, ) = 0 in this formula as well.
64 Chapter 3. Metric Theory of Countable Ordinals

to the ordering of inclusion modulo a nite set, the mapping a : [1 ]2


dened by
a(, ) = min{n : A \ n A }
is a transitive map. The transitivity condition by itself is not nearly as useful as its
combination with the other subadditivity property (3.2.2(c)). Fortunately, there
is a general procedure that produces a subadditive dominant to every transitive
map.
Denition 3.3.2. For a transitive a : [1 ]2 dene a : [1 ]2 recursively
as follows:

a (, ) = max{|C |, a(min(C \ ), ), a (, min(C \ )),


a (, ) : C }.

Lemma 3.3.3. For all < < < 1 and n < ,


(a) { : a (, ) n} is nite,
(b) a (, ) max{a (, ), a (, )},
(c) a (, ) max{a (, ), a (, )},
(d) a (, ) a(, ).
Proof. The proof of (a), (b) and (c) is quite similar to the corresponding part of
the proof of Lemma 3.1.2. This comes of course from the fact that the denition
of and a are closely related. The occurrence of the factor a(min(C \ ), )
complicates a bit the proof that a is subadditive, and the fact that a is transitive
is quite helpful in getting rid of the additional diculty. The details are left to
the interested reader. Given < , for every step n n+1 of the minimal walk
= 0 > 1 > > k = , we have a (, ) a (n , n+1 ) a(n , n+1 ) by
the very denition of a . Applying the transitivity of a to this path of inequalities
we get the conclusion (d). 
Lemma 3.3.4. a (, ) a (+1, ) whenever 0 < < and is a limit ordinal.
Proof. Recall the assumption (c) about the xed C-sequence C ( < 1 ) on which
all our denitions are based: if is a limit ordinal > 0, then no point of C is a
limit ordinal. It follows that if 0 < < and is a limit ordinal, then the minimal
walk must pass through + 1 and therefore a (, ) a ( + 1, ). 
Let us now give an application of a to a classical phenomenon of occurrence
of gaps in the quotient algebra P()/n.
Denition 3.3.5. A Hausdor gap in P()/n is a pair of sequences A ( < 1 )
and B ( < 1 ) such that
(a) A A B B whenever < , but
(b) there is no C such that A C B for all .
3.3. A Hausdor gap from 65

The following straightforward reformulation shows that a Hausdor gap is


just another instance of a nontrivial coherent sequence

f : A 2 ( < 1 )

where the domain A of f is not the ordinal itself but a subset of and that
the corresponding sequence of domains A ( < 1 ) is a realization of 1 inside
the quotient P()/n in the sense that A A whenever < .
Lemma 3.3.6. A pair of 1 -sequences A ( < 1 ) and B ( < 1 ) form a
Hausdor gap i the pair

A = A ( \ B ) ( < 1 ) and B
= \ B ( < 1 )

has the following properties:


(a) A A whenever < ,
(b) B = B
A whenever < ,
= B A for all .
(c) there is no B such that B 
From now on we x a strictly -increasing chain A ( < 1 ) of innite
subsets of and let a : [1 ]2 be dened by

a(, ) = min{n : A \ n A }.

Let a : [1 ]2 be the corresponding subadditive dominant of a dened above.


For < 1 , set
D = A+1 \ A .
Lemma 3.3.7. The sets D \ a (, ) and D \ a (, ) are disjoint whenever
0 < < < and and are limit ordinals.
Proof. This follows immediately from Lemmas 3.3.3 and 3.3.4. 
We are in a position to dene a partial mapping m : [1 ]2 by

m(, ) = min(D \ a (, )),

whenever < and is a limit ordinal.


Lemma 3.3.8. The mapping m is coherent, i.e., m(, ) = m(, ) for all but
nitely many limit ordinals < min{, }.
Proof. This is by the coherence of a and the fact that a (, ) = a (, ) already
implies m(, ) = m(, ). 
Lemma 3.3.9. m(, ) = m(, ) whenever = < and , are limits.
Proof. This follows from Lemma 3.3.7. 
66 Chapter 3. Metric Theory of Countable Ordinals

For < 1 , set

B = {m(, ) : < and limit}.

Lemma 3.3.10. B = B A whenever < .


Proof. By the coherence of m. 
Note the following immediate consequence of Lemma 3.3.7 and the denition
of m.
Lemma 3.3.11. m(, ) = max(B D ) whenever < and is a limit. 
Lemma 3.3.12. There is no B such that B A = B for all .
Proof. Suppose that such a B exists and for a limit ordinal let us dene g() =
max(B D ). Then by Lemma 3.3.11, g() = m(, ) for all < 1 and all
but nitely many limit ordinals < . By Lemma 3.3.9, it follows that g is a
nite-to-one map, a contradiction. 
Theorem 3.3.13. For every strictly -increasing chain A ( < 1 ) of subsets of
, there is a sequence B ( < 1 ) of subsets of such that:
(a) B = B A whenever < ,
(b) there is no B such that B = B A for all . 
Remark 3.3.14. For a given countable ordinal let h : A 2 be such that
h1
(1) = B . Then by Lemma 3.3.10, the corresponding sequence h ( < 1 ) of
partial functions is coherent in the sense that for every pair < < 1 , the set

Dh (, ) = {n A A : h (n) = h (n)}

is nite. By Lemma 3.3.12, the sequence is nontrivial in the sense that there is no
g : 2 such that h = g  A for all < 1 . Hausdors original sequence,
when reformulated in this way, has the property that the maps dh (, ) :
dened by
dh (, ) = |Dh (, )|
are nite-to-one mappings. The P-ideal dichotomy, (see 2.4.11 above) asserts that
every coherent and nontrivial sequence must contain a subsequence with Haus-
dors property.

3.4 A general theory of subadditive functions on 1


The purpose of this section is a general theory of distance functions on 1 which
have the following property.
Denition 3.4.1. A -function on an ordinal is any function : []2 with
the following three properties:
3.4. A general theory of subadditive functions on 1 67

(1) (, ) max{ (, ), (, )} for all < < < .


(2) (, ) max{ (, ), (, )} for all < < < .
(3) { < : (, ) n} is nite for all < and n < .
Note that = 1 is the largest ordinal that allows a -function. We have seen
above that the function : [1 ]2 of Denition 3.1.1 satises this denition,
but the purpose of this section is to examine -functions that may not necessarily
come from an analysis of the minimal walk. We are of course motivated again
by the great potential for application of such a general metric theory of 1 . For
example, having in mind that : [1 ]2 is some sort of metric on 1 , the
following useful denitions are quite natural, where a -function : [1 ]2 is
xed and where, as usual, we are adopting the convention that

(, ) = 0 for all < 1 . (3.4.1)

Denition 3.4.2. The -number of a given nite set F 1 is dened to be

pF = p (F ) = max{ (, ) : , F }.

Denition 3.4.3. For a nite set F 1 and an integer k, let

(F )k = { max(F ) : (, ) k for some F \ }.

Note that by (3.4.1), F (F )k and that by 3.4.1(3), (F )k is also nite. We will


say that F is k-closed if (F )k = F . We will say that F is -closed if it is k-closed
for k = pF .
Note that the operator F  (F )k holds all the information about the -
function on which it is based, so one can use it in describing conditions that one
can put on such . One example of this is the following useful property that one
can require from a given -function.
|({})k |
Denition 3.4.4. A -function : [1 ]2 is smooth if limk k = 0 for
every countable ordinal .
Lemma 3.4.5. There is a smooth -function on 1 .
Proof. We shall dene : [1 ]2 recursively by the formula

(, ) = max{g (|C |), (, min(C \ )), (, ) : C } (3.4.2)

with the boundary condition (, ) = 0, where C ( < 1 ) is a C-sequence on


1 and where g ( < 1 ) is a sequence of nondecreasing maps from into
both to be dened recursively as we go on. However, note that independently of
how we choose C ( < 1 ) and g ( < 1 ), the equation (3.4.2) indeed gives
us a function satisfying the conditions in Denition 3.4.1(1), (2), (3) above. Note
that smoothness is a condition that needs to be checked only for countable limit
68 Chapter 3. Metric Theory of Countable Ordinals

ordinals , or in other words, we can already make the commitment that g+1 is
the identity map for all . Similarly, we can already make the commitment that
C+1 = {} for all countable ordinals .
Case 1. = + for some countable limit ordinal . Let g (n) = 2n and let
C = (, ). For and n < , let

Fn = ({})n = { < : (, ) n}.

By our inductive assumption |Fn |/n n 0 for all (limit) ordinals and we
need to check that this holds also for = . Note that by the formula (3.4.2) if
an ordinal of the form + k belongs to Fn then k log2 n. It follows that

|Fn | |Fn | + 1 + log2 n

and this, by the inductive hypothesis, implies that |Fn |/n n 0.


Case 2. = supk k for some strictly increasing sequence (k ) of countable limit
ordinals. Let C = {k + 1 : k < }. By the inductive hypothesis, we can choose
a strictly increasing sequence (nk ) of integers such that

|Fn0 |/n + + |Fnk |/n < 2k for all k and n nk . (3.4.3)

Fix > 0 and choose k such that 2k < . We show that |Fn |/n < for n nk .
Let j be the maximal integer such that nk nj n. Note that then,

Fn { < : i() = |C | j and Fn i() } = Fn0 Fnj . (3.4.4)

It follows that |Fn |/n |Fn0 |/n+ +|Fn j |/n < 2j 2k < , as required.


Recall that in the theory of metric spaces universal objects such as, for exam-
ple, the Urysohn space are interesting and useful. The same is true in the context
of metric theory of 1 . Before we introduce the corresponding notion, we need
some preliminary denitions.
Denition 3.4.6. A nite -model is a model of the form (M, <, M , pM ) where M
is a set, < is a total ordering on M , pM is an integer, and

M : [M ]2 {0, 1, . . . , pM }

is a function with properties 3.4.1(1) and (2) listed above. We also assume that
there exists x < y in M such that M (x, y) = pM .
Note that for every -function : []2 and for every -closed subset
M of , the structure (M, <,  [M ]2 , pM ) is an example of a -model. Note also
that an initial part M0 of a -closed set M is a -closed set and that its integer
pM0 might be smaller than pM . From this one concludes that an initial part of a
-model is also a -model with a possibly smaller integer pM .
3.4. A general theory of subadditive functions on 1 69

Denition 3.4.7. Two -models (M1 , <1 , 1 , p1 ) and (M2 , <2 , 2 , p2 ) are isomorphic
if there is a bijection : M1 M2 such that for all a, b M1 ,
(i) a <1 b implies (a) <2 (b),
(ii) 1 (a, b) = 2 ((a), (b)),
(iii) p1 = p2 .
We are now ready to introduce the notion of a universal -function that is
quite analogous to the corresponding theory of metric spaces.
Denition 3.4.8. A -function : []2 dened on some limit ordinal 1
is universal if for every nite -model (M, <, M , pM ), every -closed subset M0
of such that the -model

(M0 , <,  [M0 ]2 , pM0 )

is isomorphic to an initial segment of (M, <, M , pM ), and every ordinal such


that + , there is a -closed subset M1 of + such that:
(a) (M1 , <,  [M1 ]2 , pM1 )
= (M, <, M , pM ),
(b) M0 is an initial segment of M1 ,
(c) M1 \ M0 [, + ).
Theorem 3.4.9. There is a universal -function on 1 .

Proof. We will obtain universal : [1 ]2 by recursively constructing an


increasing sequence : []2 ( ) of universal -function. Let 0 = , and
suppose : []2 has been determined for some countable limit ordinal .
Let C be a subset of of order-type such that = sup C. Dene

+ (, ) = max{|C |, (, min(C \ )), (, ) : C }.

It can be checked that this denes a function + : [ + 1]2 having the


properties (1), (2) and (3) of Denition 3.4.1 (which we will refer to throughout
this proof). Starting with this extension of and the assumption that is
universal we build extensions

+ : []2 ( + 1 < + )

in such a way that at a given stage we take care about a particular instance of
universality of + . Thus, modulo some bookkeeping device, it suces to show
how does one deal with the single task of Denition 3.4.8.
Suppose we have already dened an extension + : []2 and that we
are given a nite -model (M, <, M , pM ) and a -closed subset M0 of such that
the -model,
(M0 , <, +  [M0 ]2 , pM0 )
70 Chapter 3. Metric Theory of Countable Ordinals

is isomorphic to a proper initial segment of the -model (M, <, M , pM ). Let

l = |M | |M0 |.

We need to extend + from []2 to [ + l]2 . First of all dene + on

[M0 [, + l)]2 \ [M0 ]2

in such a way that we have the isomorphism

(M0 [, + l), <, + , pM )


= (M, <, M , pM ).

Thus, it remains to dene (, ) for \ M0 and (, + l). If < and


> max M0 , then set

(, ) = max{pM + 1, (, 1), (, ) : M0 }. (3.4.5)

If max M0 , then set

(, ) = max{ (, min(M0 \ )), (, ) : M0 }. (3.4.6)

It remains to show that +  [ + l]2 satises properties 3.4.1(1) and (2). In


order to avoid lengthy formulas, let us simplify the notation as follows,

= (, ) and = max{ (, ), (, )}. (3.4.7)

We also adopt the notation in place of 1.


So let < < < + l be a given triple of countable ordinals.
Case 1. < < < + l. If M0 , then properties 3.4.1(1) and (2) for
< < follow from the fact that in the denitions of , and we have
copied the -model (M, <, M , pM ) which satises 3.4.1(1) and (2). If
/ M0 ,
then in both case > max M0 and max M0 we conclude that = , so
3.4.1(1) and (2) for < < follow immediately.

Case 2. < < < < + l.

Subcase 2.1. max M0 < < . Consider rst the inequality . The
quantities pM + 1 and ( M ) from the denition of are all present in the
denition of , so it remains only to show that the quantity is bounded by
. Applying the inequality 3.4.1(1) for +  []2 , we get

and so we are done as shows up in the denition of .


3.4. A general theory of subadditive functions on 1 71

Consider now the inequality . Applying inequality 3.4.1 for


+  []2 to the triple < < , we get

so we are done also in this case since the quantity on the right-hand side is bounded
by .
Subcase 2.2. max M0 < < . Consider rst the subcase when M0 .
To see that observe that pM < pM + 1 . To see that
observe that appears as a quantity in the denition of . Let
us consider the case
/ M0 and let

 = min(M0 \ ).

The quantity  from the denition of is bounded by , since by the


inequality 3.4.1(2) for +  []2 , we have that

  ,

and  appears in the denition of . Since the sequence of quantities

( M0 )

appear also in the denition of , this shows that .


Let us check now the inequality . Apply the inequality 3.4.1(1)
that is valid for +  []2 to the triple <  < , and get

  .

This nishes the checking, since  and  .


Subcase 2.3. < max M0 . If , M0 , then the inequalities 3.4.1(1), (2) for
< < follow from the fact that in the denitions of , and we copied
the -model (M, <, M , pM ).
Subcase 2.3.1. M0 and
/ M0 . Consider the inequality . This
follows from the fact that

pM <  , where  = min(M0 \ )

and the fact that in the denition of the quantity  appears. The inequality
in this subcase follows from the fact that the quantity appears
in the denition of .
Subcase 2.3.2.
/ M0 and M0 . Consider the inequality . Let

 = min(M0 \ ).
72 Chapter 3. Metric Theory of Countable Ordinals

We need to bound the quantities  and ( M0 ) by . Apply the


inequality 3.4.1(2) that is valid for +  []2 to <  , and get

  .

Since  pM and  > pM we conclude that  as required. Similarly


note that
= ,
since pM while > pM . It remains to check the inequality
in this subcase. As before, note that

  ,

and that >  since  pM while > pM . It follows that  ,


as required.
Subcase 2.3.3.
/ M0 and
/ M0 (and , max M0 ). So in this subcase both
quantities and are dened according to the second denition. Let

 = min(M0 \ ) and  = min(M0 \ ).

Note that   . We rst check the inequality . Consider rst the


quantity  that appears in the denition of . If  =  , then

  ,

as  appears in the denition of . Suppose that  <  i.e., that  < .


Then
 
as  appears as a quantity in the denition of . Consider now the quantity
for M0 . Note that

as appears in the denition of .


It remains to check the inequality in this subcase. If  =  ,
then we get that
  ,
as the quantity  =  appears in while  appears in . If  <  , i.e.,
 <  , then we get that

  ,

since  appears in and  appears in .


3.4. A general theory of subadditive functions on 1 73

This nishes our checking that the extension +  [ +l]2 remains to satisfy
the conditions 3.4.1(1),(2) and (3). Note that

(, ) > pM for all \ M0 and [, + l),

we conclude that the set M0 [, + l) is + -closed. It follows that the extension


+  [, + l]2 has the subset

M1 = M0 [, + 1) + l,

of its domain as a + -closed set, while the corresponding -model

(M1 , <, +  [M1 ]2 , pM )

is isomorphic to the given -model (M, <, M , pM ). This nishes the recursive
construction of a universal -function : [1 ]2 . 
Many -functions on 1 including the universal one lack the following prop-
erty frequently useful in applications.
Denition 3.4.10. A -function : [1 ]2 is unbounded if for every k < and
every innite set A 1 there exist < in A such that (, ) > k.
It should be clear that there is an unbounded -function on 1 . For example,
the function of Denition 3.2.1 is clearly unbounded. In fact there is a gen-
eral procedure of changing an arbitrary -function to an unbounded one via the
formula,
(, ) = max{ (, ), |{ : (, ) (, )}|}. (3.4.8)
Lemma 3.4.11. Suppose : [1 ]2 is an unbounded -function. Then for every
innite family A of pairwise-disjoint nite subsets of 1 , all of some xed size n
and for every k, there exist an innite subfamily B A such that for all a = b in
B and all a and b, we have that (, ) > k.
Proof. Fix an integer k and a nonprincipal ultralter U on A. A simple diagonal-
ization procedure reduces the conclusion of the lemma to showing that

(i, j < n)(Ua)(Ub) (a(i), b(j)) > k.3 (3.4.9)

Suppose otherwise that there exist i, j < n such that

(Ua)(Ub) (a(i), b(j)) k. (3.4.10)

Note that by the property (3) of , we must have that

(i) = limaU a(i) limaU a(j) = (j), (3.4.11)


3 Here a(i) and b(j) denote the ith element of a and jth element of b, respectively, relative to
their increasing enumerations.
74 Chapter 3. Metric Theory of Countable Ordinals

or else, (3.4.10) would fail. Let

X = {a(i) : (Ub) (a(i), b(j)) k}

and let Y = X (i). Then Y is, in particular, innite. We claim that (, ) k


for all and in Y , contradicting the assumption that is unbounded. To see
this consider < in Y . Then by (3.4.10) and the fact that

limaU a(j) = (j) > > , (3.4.12)

we conclude that there must be b in A such that b(j) > > and (, b(j)) k
and (, b(j)) k. Using the subadditivity of , this gives us (, ) k, as
required. 

Further analysis of -functions on 1 requires the following natural notion.


Denition 3.4.12. Given a -function on 1 and n < , we let Gn denote the
corresponding graph whose vertex set is 1 and where an unordered pair {, }
forms an edge if and only if (, ) = n.
Note the following immediate consequence of Lemma 3.2.2(a) that bears on
this notion.
Lemma 3.4.13. If is equal to of Denition 3.2.1, then all graphs Gn are acyclic.

Using deeper properties of one has the following information about the
graphs Gn .
Lemma 3.4.14. If is equal to of Denition 3.2.1, then for a given positive
integer n, the connected component of the graph Gn is a star whose center is its
minimum.

Proof. Consider a path 0 , 1 , . . . , k of some xed graph Gn . Note that by Lemma


3.2.3 for no i < k 1 do we have the inequalities

i < i+1 < i+2 or i > i+1 > i+2 .

On the other hand, by Lemma 3.2.2(a), the path does not contain three consecutive
points i , i+1 , i+2 with property

i < i+1 > i+2 .

So, if j is the minimal point of the path, then on one hand k j 1, and on the
other hand j 1. It follows that the path has length at most 2, as required. 
Corollary 3.4.15. The complete graph on 1 vertices can be split into countably
many graphs, none of which has paths of length 3. 
3.4. A general theory of subadditive functions on 1 75

Remark 3.4.16. It follows, in particular, that the complete graph on 1 vertices


can be split into countably many acyclic graphs. In fact, 1 is the largest vertex-set
that allows such a decomposition. This is an old result of Erdos and Kakutani [30].
Similar argument shows the following result that complements Corollary 3.4.15.
Lemma 3.4.17. For every partition of the complete graph on 2 into countably
many subgraphs Gn , one of the graphs Gn contains an innite path.
Proof. Let c : [2 ]2 be a given countable edge-partition of the complete
graph (2 , [2 ]2 ). The conclusion of the lemma will follow if we can nd an integer
n < , an increasing sequence i (i < ) of countable ordinals, and an increasing
sequence j (j < ) of uncountable ordinals < 2 such that
c(i , j ) = n for all i j < . (3.4.13)
Since then 0 , 0 , 1 , 1 , . . . , i , i , . . . would form an innite path of the graph
(2 , c1 (n)), as required. Find an n < such that the set B of all [1 , 2 ) for
which the corresponding set
A = { < 1 : c(, ) = n}
is unbounded in 1 is unbounded in 2 . We shall show that this n works. Find
0 B with the property that { B : c(, ) = n} is unbounded in 2 for every
A0 . Pick an arbitrary 0 A0 and set
B0 = { B : > 0 and c(0 , ) = n}.
Then B0 is an unbounded subset of 2 so we can nd 1 B0 with the property
that { B0 : c(, ) = n} is unbounded in 2 for every A1 . Pick an
arbitrary 1 A1 such that 1 > 0 and form the set
B1 = { B0 : > 1 and c(1 , ) = n}.
Then B1 is unbounded in 2 , so we can proceed further and nd 1 , 1 , B2 ,
2 , 2 , B3 , etc. Note that the construction is giving us the desired equalities of
(3.4.13). 
Remark 3.4.18. One may wonder if the previous proof can be augmented so that
we also get the equalities
c(i , j ) = n for all j < i < . (3.4.14)
besides those of (3.4.13). Theorem 9.3.1 below shows that, in fact, this is not
possible to achieve.
Clearly no graph Gn associated to a universal function can be acyclic, so
one can view and universal as extreme points of a rich spectrum of -functions
relative to the structure theory of the corresponding graphs Gn . So let us introduce
some interesting classes between these two extremes. First of all note the following
general fact about -functions.
76 Chapter 3. Metric Theory of Countable Ordinals

Lemma 3.4.19. Suppose that is a -function and that n is a positive integer. Let
F be a -closed set whose -number is equal to n. Then there is a decomposition

F = F0 F1 F2 , (3.4.15)

where F0 is an initial segment of F consisting of points that are isolated in Gn


while every point of F1 is connected to every point of F2 .
Proof. Note that by our assumption n > 0, since in case n = 0 the restriction of Gn
on the set F is a complete graph, and so while we can put F0 = and F1 = F2 = F ,
the decomposition is not as required since the pieces are not pairwise-disjoint. Let
E be the connected component of the point = max F relative to the graph which
Gn induces on F . Note that by the subadditivity properties of the component
E has at least one more point besides . By our assumption that Gn is bipartite
so we can decompose E as the union of two nonempty subsets F1 and F2 such
that all edges of Gn inside E are between a point of F1 and a point of F2 . We
assume that belongs to F2 . Let = max F1 . Recall the tree-ordering <n1 on
1 associated to and the integer n 1 as follows,

<n1 i < and (, ) n 1. (3.4.16)

Let Pn1 () = { : n1 }, and similarly, let Pn1 () = { : n1 }. Then

F = Pn1 () Pn1 (),

and moreover, the set F0 = Pn1 ()Pn1 () is an initial segment of both Pn1 ()
and Pn1 (). It follows that F0 is also an initial segment of F . Note that

(, ) n 1 for every {, } [F ]2 with {, } F0 = .

From this, we conclude that every point of the set F0 is an isolated point of
the graph Gn induces on F . Since every point of the connected component E is
connected to at least one more point of E, we conclude that E is disjoint from F0 .
Since
F1 Pn1 () and F2 Pn1 (),
we conclude that every point of F1 is <n1 -incomparable to every point of F2 .
This precisely means that every point of F1 is Gn -connected to every point of F2 .
This completes the proof. 
Corollary 3.4.20. If is a -function and if n is a positive integer, then the graph
Gn has no odd cycles. 
Denition 3.4.21. For a -closed set F with -number equal to n, we call the
decomposition (3.4.15) of Lemma 3.4.19, the Gn -decomposition of F , or more
simply the -decomposition of F . The -decomposition (3.4.15) is increasing if
F0 < F1 < F2 .
3.5. Conditional weakly null sequences based on subadditive functions 77

Denition 3.4.22. We call a -function : [1 ]2 a uniform increasing bipartite


-function if:
(1) Every -closed set H is nite and is included in a maximal -closed F which
is also nite and which has the same -number as H, and moreover, every
such set F admits an increasing -decomposition.
(2) Maximal -closed sets F and the pieces F0 , F1 and F2 of their increasing
-decompositions have cardinalities kn , ln , and ln , respectively, that depend
only on the -number n of F .
Theorem 3.4.23. There is a uniform increasing bipartite -function on 1 . 
Remark 3.4.24. We do not give a proof of Theorem 3.4.23 here, though it bears
a considerable similarity with the proof of Theorem 3.4.9 above. It would be of
interest, however, to determine the possible sequences (kn )n and (ln )n that wit-
ness uniformity of a given uniform increasing bipartite -function on 1 . Some
results to that eect are given by Velleman [126] though using a quite dierent
terminology. The relationship between the theory of -functions and Vellemans
theory of morasses was fully explained by Morgan [79].

3.5 Conditional weakly null sequences based on


subadditive functions
In this section we mention some applications of the concept of -function on 1
satisfying the inequalities of Lemma 3.1.2. As we shall see, any such a semi-distance
function can be used as a coding procedure in constructions of long weakly null
sequences in Banach spaces with no unconditional basic subsequences.
Example 3.5.1. A weakly null sequence with no unconditional basic subsequence.
We describe a weakly null normalized sequence (x )<1 of members of some
Banach space X which contains no innite unconditional basic subsequence4 . Fix
an 0 < < 1 and an innite set M of positive integers such that 1 M and

 m 
: m, n M, m < n . (3.5.1)
n
Fix also a natural bijection . : HF M , where HF denotes the family of all
hereditarily nite sets, so that in particular  = 1. Recall that the -number of
4 Recall that a basic sequence in a Banach space X is a seminormalized sequence (x )
n n with the

 that every x in its closed linear span span{xn : n } can be uniquely represented as
property
a sum n an xn . Note that this is equivalent to saying that the norms of projection operators
Pk : span{xn : n } span{xn : n < k} are uniformly bounded. When the norms of
projections PM : span{xn : n } span{xn : n M } over all subsets M of are uniformly
bounded, the sequence (xn ) is said to be unconditional. The natural basis (en ) of any of the
sequence-spaces p for p [1, ] is unconditional. A typical example of a basic sequence that is
not unconditional is the summing basis e0 , e0 + e1 , e0 + e1 + e2 , . . . of c0 or the summing basis
e0 + e1 + e2 + + en + , e1 + e2 + + en + , e2 + + en + , . . . in c.
78 Chapter 3. Metric Theory of Countable Ordinals

a nite set F 1 is the integer

(F ) = max{ (, ) : , F, }.

For F 1 and k , let

(F )k = { max(F ) : (, ) k for some F with }.

We say that F is k-closed if (F )k = F . The -closure of F is the set (F ) = (F )k ,


where k = (F ). This leads us to the function

: [[1 ]< ]< M

dened by letting () =  and


n
  
(F0 , F1 , . . . , Fn ) = ((E F0 , E F1 , . . . , E Fn ), ( Fi )),
i=0
n
where E = ( i=0 Fi ) and where E : E |E| is the unique order-preserving
map between the sets of ordinals E and |E| = {0, 1, . . . , |E| 1}. We say that a
sequence (Ei )i<n of nite subsets of 1 is special if
1. Ei < Ej 5 for i < j < n,
2. |E0 | = 1 and |Ej | = (E0 , E1 , . . . , Ej1 ) for 0 < j < n.
Let c00 (1 ) be the normed space of all nitely supported maps from 1 into the
reals. Let (e )<1 be the basis of c00 (1 ) and let (e )<1 be the corresponding
sequence of biorthogonal functionals. To a nite set E 1 , we associate the
following vector and functional on c00 (1 ),

1  1 
xE = e and E = e . (3.5.2)
|E|1/2 E |E|1/2 E

Given a special sequence (Ei )i<n of nite subsets of 1 , the corresponding special
functional is dened by i<n Ei . Let F be the family of all special functionals
of c00 (1 ). This family induces the following norm on c00 (1 ),


x
= sup{f, x : f F }. (3.5.3)

The Banach space X is the completion of the normed space (c00 (1 ),


.
). Note
that
e
= 1 and that (e )<1 is a weakly null sequence in X. To see this
note that given a nite set G with |G| M and a special sequence (Ei )i<n , if
5 Recall, that for two sets of ordinals E and F , by E < F we denote the fact that every ordinal

from E is smaller than every ordinal from F . Sequences of sets of ordinals of this sort are called
block sequences.
3.5. Conditional weakly null sequences based on subadditive functions 79


= i<n Ei is the corresponding special functional, then the inner product

between and the average |G|1 G e = |G|1/2 xG is not bigger than

|G|1/2 (1 + min((m/|G|)1/2 , (|G|/m)1/2 )
|G|=mM

1/2
 |G|
and therefore not bigger than (1 + ) by our assumption (3.5.1). It follows
that the averages
|G|1 G e
tend to 0 as |G| M goes to innity, so
(e )<1 must be weakly null in X. To show that no innite subsequence of
(e )<1 is unconditional it suces to show that for every set G 1 of order
type there is a seminormalized block-subsequence6 of (e )G which represents
the summing basis in c. More precisely, we shall show that if (Ei )i< is any innite
special sequence of nite subsets of G and if for i < we let vi = xEi as dened
in (3.5.2), then for every n < and every sequence (ai )in [1, +1] of scalars.


k 
n 
k
max | ai |
ai vi
(3 + ) max | ai | . (3.5.4)
0kn 0kn
i=0 i=0 i=0

To see the rst inequality, note that according to (3.5.2) n the absolute value of
the inner
n product between the
n special functional = i=0 Ei and the vector
x = i=0 ai vi is equal to | i=0 ai | giving us the 
rst inequality. To see the
m
second inequality, consider a special functional = i=0 Fi where (Fi )im is
some other special sequence of nite sets. In order to estimate the inner product
, x we need to know how the two special sequences interact and this requires
properties of the function that controls them. To see this, let l min(m, n)
be maximal
 with the property that |El | = |Fl |, and let E = ( i<l Ei ) and
F = ( i<l Fi ) . By the denition of a special sequence and |El | = |Fl |, we know
that:
 
3. ( i<l Ei ) = ( i<l Fi ),
 
4. E Ei = E Fi for i < l.
It follows that E and F have the same cardinality and the same -numbers, and
therefore by the properties of listed in Lemma 3.1.2 their intersection H = E F
is their initial segment. Hence, E and F agree on H. Let k be the maximal integer
such that Ek H. Then Ei = Fi for all i k, and using again the properties of
listed in Lemma 3.1.2, Ei Fj = for k + 1 < i, j < l (see Figure 3.2). Combining
all this we see that |, x| is bounded by

 k  m

ai + ak+1 Fk+1 , xEk+1  + al Fl , xEl  + {( )1/2 : m, n M, m < n}.
n
i=0

6 i.e.,
a seminormalized sequence (xi )i< of nitely supported vectors of the linear span of
(e )G such that supp(xi ) < supp(xj )for i < j.
80 Chapter 3. Metric Theory of Countable Ordinals

Figure 3.2: The interference of two special functionals.

Referring to (3.5.1) we get that |, x| is bounded by the right-hand side of the
second inequality of (3.5.4).
Remark 3.5.2. The rst example of a weakly null sequence (of length ) with
no unconditional basic subsequence was constructed by Maurey and Rosenthal
[74]. Sixteen years later Gowers and Maurey [39] solved the basic unconditional
sequence problem by constructing an innite-dimensional separable reexive Ba-
nach space with no innite unconditional basic sequence. The basic new ingredient
in the construction of [39] is the idea behind the Schlumprecht space [89] which
can in general be described as follows in terms of two increasing sequences (mk )
and (nk ) of positive integers dened by the recursive formulas:
k 3
mk = 24 and n0 = 4 and nk+1 = (4nk )log2 mk+1 . (3.5.5)
3.5. Conditional weakly null sequences based on subadditive functions 81

Denition 3.5.3. For an innite ordinal let G be the minimal subset of c00 ()
with the following properties:
(1) e G for all < ,

(2) (1/mj ) i<nj i G for every non-negative integer j and every block se-
quence 0 < 1 < < nj 1 of members of G ,
(3) G is closed under restrictions to intervals of ,
(4) G is closed under rational convex combinations.
Let T [m1
j , nj ] be the completion of c00 () equipped with the norm


x
= supG , x.

Remark 3.5.4. Schlumprechts original space [89] is really the space T [1/mj , nj ]
where mj = log2 (j + 1) and nj = j, though we will nd it more convenient to
work with T [1/mj , nj ] for the particularly chosen sequences (mj ) and (nj ) as in
(3.5.5) or their subsequences. Schlumprechts space was the rst known example
of an arbitrary distortable reexive Banach space, and his arguments form an
important part of basically all latter constructions of implicit norms including the
ones that we choose to present below.
The following result of [4] relates to Example 3.5.1 given at the beginning
of this section in the similar way the result of [39] relates to the example of [74].
While the basic idea of adding conditionality to the norm is supplemented by the
use of the -function, substantial new work is still needed to implement the ideas
of [89] and [39] to this setting.
Theorem 3.5.5. There is a reexive Banach space X1 with a transnite Schauder
basis of length 1 containing no innite unconditional basic sequence. 
The space X1 is the completion of c00 (1 ) under the norm given by the formula
(3.5.3), where now F = F1 is chosen as the minimal set of nitely supported
functionals on c00 (1 ) such that:
1. F {e : < 1 } and F is symmetric and closed under restriction on
intervals of 1 .
2. For every (i )i<n2k F such that supp(i ) < supp(j ) for i < j < n2k , the
functional m12k i<n2k i belongs to F .
3. For every special sequence (i )i<n2k+1 F the corresponding special func-

tional m1
2k+1 i<n2k+1 i belongs to F .
4. F is closed under rational convex combinations.
The notion of a special sequence of functionals is dened analogously to that
Example 3.5.1 as follows. First of all we say that F is of type 0 if =e for
some and of type I if it is obtained from a simpler one as the sum m1 k i<nk i
restricted to some interval; further, we call the integer mk , the weight of , and
82 Chapter 3. Metric Theory of Countable Ordinals

denote it w(). Note that a given functional of type I can have more than one
representation and therefore more than one weight. We say that is of type II
if it is a rational convex combination of functionals of type 0 and type I. The
denition of a special sequence of functionals will be again based on . For this
we need an injection
. : HF {2k : k odd },
such that if (i , wi , pi )i<j is a nite sequence of triples where wi , pi and
where i s are nite partial maps from into Q, then (i , wi , pi )i<j  is bigger
than the maximum of support of any of the i s, the square of any integer of
the form wi or pi , and the ratio of the square of any nonzero value of a i .
A sequence (i )i<n2k+1 of members of F is a special sequence of functionals, if

(a) supp(i ) < supp(j ) for i < j < n2k+1 ,


(b) each i is of type I and of some xed weight w(i ) = m2ki with k0 even and
satisfying m2k0 > n22k+1 ,
(c) for 0 < j <n2k+1 there is an increasing sequence (pi )i<j of integers such
7
 pi ( i <i supp(i )) and w(j ) = m(E i ,w(i ),pi )i<j  , where E =
that  

( i<j supp(i )) .

The condition 1 from the denition of F provides that (e )<1 is a bimonotone


Schauder basis of X1 , the condition 3 is responsible for the conditional structure
of the norm of X1 in a similar manner as Example 3.5.1, but in order to come
to the point where the idea can be applied, one needs the unconditional structure
given by the closure property 2 of F . This requires a rather involved tree-analysis
of the functionals of F that gives the required norm estimates, an analysis that is
beyond the scope of this presentation. We refer the reader to [4] for the details.
Here we just list some of the most interesting properties of X1 . For example,
the space X1 is reexive, has no innite unconditional basic sequence, and is not
isomorphic to any proper subspace or a nontrivial quotient. The basis (e )<1
allows us to talk about natural subspaces of X1 . For example, given an innite
interval I 1 , one can consider the corresponding restriction

XI = span({e : I}).

Then XI are all reexive Banach spaces with transnite Schauder basis (e )I .
Particularly interesting are the separable spaces of the form

X = span({e : < })

for a countable limit ordinal. It turns out that, choosing an appropriate function
, all these separable restrictions of X1 not only have transnite Schauder bases
but also Schauder bases of length .
7 The  is dened by the formula ( ()) = ().
collapse i = E i E
3.5. Conditional weakly null sequences based on subadditive functions 83

Lemma 3.5.6. Suppose the construction of X1 is based on a -function : [1 ]2


. Fix a countable limit ordinal . For n < , let

Fn = ({})n = { < : (, ) n},

and let
PFn : X XFn
be the projection from X onto XFn = span({e : Fn }). Then the sequence of
projections (PFn )n is uniformly bounded by 1 + supk |({})k |/k. 

Clearly, there exist -functions on 1 for which the sequence ( |({})


k
k|
) is bounded
for every . For example, the function of Denition 3.2.1 is one such -function.
Note also that every smooth -function constructed above in Lemma 3.4.5 also
has the sequence ( |({})
k
k|
) bounded for every countable ordinal and smoothness
is exactly what is needed to turn the sequence of projections into a Schauder basis
for X of length .
Corollary 3.5.7. If the construction of X1 is based on a smooth -function on 1 ,
then every subspace of X1 of the form XI , for I 1 a countable interval, has a
Schauder basis of length . In fact, a Schauder basis of XI of length is obtained
by re-enumerating e ( I). 
If we base X1 on a universal we get a rather interesting property that
every nite-dimensional subspace Y of X1 can be isomorphically embedded into
any subspace XI of X1 spanned by an innite interval of the transnite Schauder
basis (e )<1 by an operator T such that
T

T 1
4 + .
Denition 3.5.8. A transnite sequence (e )< is nearly subsymmetric if there is
a positive constant C such that for every > 0 and every nite block-sequence
Fi (i k) of nite subsets of and for every sequence I0 I1 Ik of
innite intervals of there is a sequence Gi Ii , |G
i | = |Fi | (i k) such
 that the
natural isomorphism T : XF XG , where F = ik Fi and G = ik Gi has
the property that
T

T 1
C + .
Theorem 3.5.9. If X1 is based on a universal -function, then its Schauder basis
(e )<1 is nearly subsymmetric with constant 4. 
Taking the restrictions XI = span({e : I}) for an appropriate family of
countable innite intervals I 1 , one obtains an uncountable family of separa-
ble reexive spaces which are asymptotic versions of each other. This is a new
phenomenon in the separable Banach-space theory not revealed by the previous
methods.
The basis (e )<1 allows us also to talk about diagonal operators of X1 . It
turns out that all bounded diagonal operators D of X1 are given by a sequence
( )<1 of eigenvalues that have the property that

= whenever < + .
84 Chapter 3. Metric Theory of Countable Ordinals

Moreover, there are only countably many dierent eigenvalues among ( )<1 .
Recall, that an operator between two Banach spaces is strictly singular if it is not
an isomorphism when restricted to any innite-dimensional subspace of its domain.
It turns out that every bounded operator T on X1 has the form DT + S, where
DT is a diagonal-step operator with countably many eigenvalues and where X is a
strictly singular operator. There is a natural candidate for DT . The problem is to
show that it is a bounded operator. The original way of handling this problem is
again based on an universal -function. To see this, suppose we are given a nitely
supported normalized vector x such that DT (x) is very large in comparison with
T (x). The vector x has a natural decomposition x = x1 + + xn such that

DT (x) = 1 x1 + + n xn ,

where 1 , . . . , n are eigenvalues of T . The universality of via Theorem 3.5.9


guaranties that xi s can be simultaneously moved by a (4 + )-isomorphism some-
where very close to eigenvectors with eigenvalues i s. This will give us a con-
tradiction. It follows that, in particular, every bounded linear operator on X1 is
a multiple of the identity plus the operator with separable range8 . It turns out
that we can actually identify the space D(X1 ) of bounded diagonal operators as
the dual of the James space JT0 (1 ) over the particular reexive Banach space
T0 = T [1/m2j , n2j ] and 1 as the corresponding ordered set. So let us recall the
functor (X, A)  JX (A) as introduced in [4].
Denition 3.5.10. Let X be a reexive Banach space with a 1-subsymmetric basis
(ek )k< , and let A be a set of ordinals naturally ordered. Then JX (A) is the
completion of c00 (A) under the norm given by

 


x
JX (A) = sup

x() e

k
: n < , I0 < < In1 intervals of A .
k<n Ik X

The natural Hamel basis (v )A of c00 (A) is a 1-subsymmetric Schauder basis


of JX (A) which is not unconditional since for example l1 does not embed  into
JX (A). Note also that for an interval I of A the functional I (x) = I x()

belongs to the dual JX (A) and in fact
I
= 1. Moreover, the dual space JX

(A)
is generated in norm by the family of functionals of the form I , where I is an

initial interval of A. It follows that the norm dimension of the dual space JX (A)
is equal to the cardinality of the set A. It turns out that the following concept is
quite useful in describing the phenomena encountered inside the space X1 .
Denition 3.5.11. Let X be a Banach space with a Schauder basis (xn )n< and
let Y be a Banach space with a transnite Schauder basis (y )< . We say that
X is nitely interval representable in Y if there is a constant C > 0 such that for
8A considerably easier example of a space with this property will be given in some detail in
Theorem 5.3.10 below.
3.5. Conditional weakly null sequences based on subadditive functions 85

every nite sequence I0 I1 In1 9 of innite intervals of there exist


zk span{y : Ik } (k < n) such that the natural isomorphism

H : span(xk )k<n  span(zk )k<n

has the property that


H

H 1
C.
Remark 3.5.12. The possible relevance of this concept becomes more clear if one
recalls that in Example 3.5.1 above we have really proved that the James-like
space Jc0 () is nitely representable in an arbitrary innite subsequence of the
weakly null sequence (e )<1 .
The following is the main technical result about the concept of nite interval
representability inside our space X1 .
Theorem 3.5.13. The James space JT0 10 is nitely interval representable inside an
arbitrary closed linear span of a transnite normalized block sequence (y )< of
X 1 . 
Corollary 3.5.14. The Banach space X1 contains no innite basic unconditional
sequence. 
Denition 3.5.15. A Banach space X is indecomposable if for every topological11
decomposition X = Y Z one of the spaces Y of Z must be nite-dimensional.
We say that X is hereditarily indecomposable if every closed subspace of X is
indecomposable.
Corollary 3.5.16. Every innite-dimensional closed subspace of X1 contains an
innite-dimensional hereditarily indecomposable subspace. 
Denition 3.5.17. A bounded operator S : X  Y is strictly singular if S  Z is
not an isomorphic embedding for every innite-dimensional subspace Z of X. Let
S(X, Y ) denote the collection of all bounded strictly singular operators from X
into Y .
The following result characterizes the operator space of an arbitrary closed sub-
space X of X1 spanned by a transnite normalized block sequence (x )< of
vectors.
Theorem 3.5.18. L(X)/S(X) = L(X, X1 )/S(X, X1 ) = JT0 (0 ()) for every
closed subspace X of X1 generated by a transnite basic block sequence (x )< ,
where 0 () denote the set of all limit ordinals that have the form +
for some ordinal < . 
Corollary 3.5.19. L(X1 )/S(X1 )
= JT0 (1 ). 
9 For two intervals of ordinals I and J, we let I J if either I = J or else every ordinal from I

is smaller than every ordinal from J


10 Here, J
T0 = JT0 () for T0 = T [1/m2j , n2j ], where (mj ) and (nj ) are the sequences of positive
integers as in (3.5.5).
11 I.e., Y Z spans X and d(S , S ) > 0.
Y Z
86 Chapter 3. Metric Theory of Countable Ordinals

Corollary 3.5.20. No closed subspace X of X1 generated by a transnite basic


block sequence is isomorphic to its proper subspace and no such a subspace X of
X1 can have a nontrivial quotient. 
Corollary 3.5.21. X1 is a reexive Banach space that is not isomorphic to any of
its proper subspaces and X1 has no nontrivial quotients. 
In fact one has a more informative result that lies behind Theorem 3.5.18. To see
this x a bounded operator T : X  X1 , where X is a closed linear span of some
transnite normalized block sequence (x )<1 of vectors of X1 . Recall that ()
is the set of limit ordinals and that 0 () is the set of isolated members of
(). Then to the given operator T we associate the function
T : 0 () R
by letting T () = if limn
T (yn ) yn
= 0 for every (3, )-rapidly increas-
ing sequence block subsequence (yn )n< of (x )<1 whose supports relative to
(x )<1 increasingly converge to . Here a (3, )-rapidly increasing sequence is
what lies behind Schlumprechts idea of recovering a biorthogonal block sequence
of vectors and functionals inside an arbitrary innite block sequence of vectors in
spaces that incorporated his norming set of functionals, which in our case is cap-
tured by the property 2 of F1 . Now, given an arbitrary mapping : 0 () R,
dene the corresponding diagonal operator D : X X by letting
D (x ) = ( + )x .
For a bounded operator X  X1 let DT : X X1 be dened as DT = DT . The
following, whose proof is based on Theorem 3.5.13, gives more precise information
about operator spaces than Theorem 3.5.18.
Theorem 3.5.22. Suppose X is a closed linear span of a transnite normalized
block sequence of vectors of X1 . Then for every bounded operator T : X X1 ,
the corresponding diagonal operator DT : X X1 is bounded as well and the
dierence T DT is strictly singular. 
While Theorem 3.5.18 gives us a fairly complete description of operator spaces of
closed linear spans of transnite basic block sequences of vectors of X1 , one can
construct subspaces of X1 that reveal some quite dierent and interesting new
phenomena about operator spaces such as the following.
Theorem 3.5.23. There is a closed subspace Y of X1 such that every bounded
operator T : Y Y is a strictly singular perturbation of a scalar multiple of
identity, while operator space L(Y, X1 ) is quite rich in the sense that
L(Y, X1 ) = JT0 S(Y, X1 ). 

The method of adding the conditional structure to norms using -functions


is actually quite exible. To this eect, we give yet another construction of a
conditional norm using . It will utilize the following notion.
3.5. Conditional weakly null sequences based on subadditive functions 87

Denition 3.5.24. Fix a -function : [1 ]2 . For two subsets s and t of 1 ,


set
(s, t) = min{ (, ) : s, t}.
Given an integer k we say that s and t are k-separated if (s, t) k. A sequence
(si )i is k-separated if it is pairwise k-separated, i.e.,
 si and sj are k-separated for
every i = j. The sequence (si )i is separated if it is | i si |-separated. These notions
are naturally transferred to vectors of X01 via their supports.
The following results show that separated sequences appear in profusion.
Lemma 3.5.25. Let (Ai )i<n be a block sequence of subsets of 1 , each of them of
order-type . Then for every block sequence (F )i<n Ai of nite sets of countable
ordinals and every integer k, there are i Ai (i < n) such that (Fi )i<n is k-
separated.
Proof. This is done by induction on n. Fix all data as in the statement for n > 1.
Then let n1 = min An1 . Since the set
(Fn1 )k = { < 1 : (, ) k for some (Fn1 \ )}
is (by property (iii) of ) nite, one easily obtains innite sets
Bi Ai (i < n 1)
such that F (Fn1 )k = for every Bi and all i < n 1. By inductive
hypothesis, there are i Bi (i < n 1) such that (Fi )i<n1 is k-separated.
Then (Fi )i<n is the desired k-separated sequence. 
Corollary 3.5.26. Let n be an integer and let (Fi )<1 be a block sequence of nite
sets of countable ordinals for every i < n. Then there are 0 < < n1 such
that (Fi i )i<n is a separated block sequence.
Proof. Let A be an uncountable set such that |Fi | = ki for every i A and every
i < n. Now for each i < n, let Ai A be of order type and such that Ai < Aj
and Fi < Fj if i < j < n and Ai , Aj are such that < . Then apply

the previous proposition to (E )A0 An1 and i<n ki , where E = Fi for
the unique i < n such that Ai . 
We are now ready to dene the new conditional norm on c00 (1 ) for which
the corresponding space behaves quite dierently than the space X1 considered
above.
Denition 3.5.27. Let K be the minimal subset of c00 (1 ) satisfying the following
conditions:
(i) e K for all < 1 , K is symmetric (i.e., K implies K), and K
is closed under the restriction on intervals of 1 .
(ii) For every separated block sequence (i )di=1 K, with d n2j , one has that
n2j
the combination (1/m2j ) i=1 i K.
88 Chapter 3. Metric Theory of Countable Ordinals

special sequence (i )di=1 K with d n2j+1 one has


(iii) For every separated 
n2j+1
that = (1/m2j+1 ) i=1 i is in K. The functional is called a special
functional.
(iv) K is rationally convex.

Whenever K is of the form = (1/mj ) i<d i given in (ii) or (iii) above we
say that has a weight w() = mj . Finally, the norm on c00 (1 ) is dened as


x
= sup{(x) = <1 () x() : K},

and we let X01 be the completion of the normed space (c00 ,



).

Let us list some properties of X01 that follow directly from this denition.
Remark 3.5.28.
(a) It is clear that the norming set K presented here is a subset of the norming
set of the Banach space X1 considered above. So, for x c00 (1 ), the norm of x
in the space X01 is not bigger than the norm of x in X1 .
(b) There is a natural notion of complexityof elements of K: Either = e ,
or is a rational
 convex combination = i<k ri fi of elements (fi )i<k of K, or
= (1/mj ) i<d fi for a separated block sequence (fi )i<d in K with d nj . And
in this latter case we say that w() = mj is a weight of .
(c) The property (i) makes the natural Hamel basis (e )<1 of c00 (1 ) a trans-
nite bimonotone Schauder basis of X01 , or in other words, X01 is a closed linear
span of (e )<1 , and for every interval I 1 the corresponding projection

PI : X01 XI0 = span({e : I})

has norm 1. Let us set P = P[0,] and X0 = X[0,]


0
for < 1 .
(d) The basis (e )<1 is shrinking in the sense that the subbasis (en )n< is
shrinking in the usual sense12 for every increasing sequence (n )n< of countable
ordinals. It follows that (e )<1 is an uncountable weakly-null sequence, i.e., for
every x (X01 ) the numerical sequence (x (e ))<1 belongs to c0 (1 ). This
last property readily implies that if T : X01 X01 is a bounded operator, then
for every uncountable subset A of 1 and every countable ordinal one has that
P (T (e )) = 0 for all but countably many A.
The property (ii) of the norming set is responsible for the existence of so-called
semi-normalized averages in the span of every uncountable block sequence of X01 .
This in combination with the third (iii) and fourth (iv) properties of the norming
set K is behind a rather strong control on operators that is expressible using the
following new concept.
12 I.e.,
the sequence {en : n < } of biorthogonal functionals is the Schauder basis of the dual
space (span({en : n < }) .
3.5. Conditional weakly null sequences based on subadditive functions 89

Denition 3.5.29. An operator T : Y X is an 1 -singular operator if it has


the property that T  Z is not an isomorphic embedding of Z into X for every
nonseparable closed subspace Z of Y .
Remark 3.5.30. Observe that strictly singular and separable range operators are
1 -singular. While the strictly singular operators and operators with separable
ranges form closed ideals of the Banach algebra L(X) of all bounded operators
from X into X, it is not clear if for general Banach spaces X this is also the case
for the family S1 (X) of 1 -singular operators of X. Indeed it is not even clear if
S1 (X) is closed under sums. We shall show however that for subspaces X of our
space X01 one does have that the S1 (X) form a closed ideal in the algebra L(X).
Theorem 3.5.31. Every operator T : Y X01 from a closed subspace Y of X01
into X01 has the form
T = IY,X0 + S,
1

where is a scalar, IY,X0 is the inclusion operator from Y into X01 and where
1
S is an 1 -singular operator. 

Here are some of the other main properties of the space X01 .
Theorem 3.5.32. The space X01 is c0 -saturated in the sense that every innite-
dimensional closed subspace of X01 contains an isomorphic copy of c0 . 
Corollary 3.5.33. The space X01 is not -distortable, or more precisely, for every
> 0 and every equivalent norm

 on X01 there is an innite-dimensional
subspace Y of X01 such that
x
 (1 + )
y
 for every x, y Y with

x
=
y
= 1. 
Theorem 3.5.34. The space X01 is arbitrarily 1 -distortable, or more precisely,
for every positive real number M there is an equivalent norm

 on X01 such
that for every non-separable closed subspace Y of X01 there exist x, y Y such
that
x
=
y
= 1 but
x
 M
y
 . 

Note that by Theorem 3.5.32, every innite-dimensional closed subspace of X01


contains an innite basic sequence. However, on the non-separable level, we have
a quite dierent behavior as the following result shows.
Theorem 3.5.35. The space X01 contains no uncountable unconditional basic se-
quence. 

This can be made more precise using the following natural variation of the standard
notion of decomposability.
Denition 3.5.36. A Banach space X is 1 -indecomposable if for every topological
decomposition X = Y Z one of the spaces Y of Z must be separable. We
say that X is hereditarily 1 -indecomposable if every closed subspace of X is
1 -indecomposable.
90 Chapter 3. Metric Theory of Countable Ordinals

This kind of indecomposability corresponds to the notion of 1 -singularity for


operators introduced above. Using this notion one can have the following sucient
condition for being 1 -hereditarily indecomposable.
Lemma 3.5.37. Suppose that X has the property that for every subspace Y of X
every bounded operator T : Y X is of the form T = IY,X + S for some scalar
and a 1 -singular operator S. Then X is 1 -hereditarily indecomposable.
Proof. Otherwise, x two non-separable subspaces Y and Z of X whose union
spans X and which have the property that

d(SY , SZ ) > 0.

It follows that the two natural projections PY : Y Z Y and PZ : Y Z Z


are both bounded. Fix a real number such that

T = IY,X PY = IY Z,X + S

with S an 1 -singular operator. Since T 2 = T , we have that

(2 )IY Z,X = ((1 2)IY Z,X S) S. (3.5.6)

Since it is clear that U S is 1 -singular if S is 1 -singular, it follows from (3.5.6)


that 2 = . Without loss of generality, we may assume that = 1 (if = 0 we
replace Y by Z in the preceding argument). Since PY  Z = 0, we obtain that
S = IZ,X , a contradiction. 
This leads us to the following more precise version of Theorem 3.5.35.
Theorem 3.5.38. The space X01 is 1 -hereditarily indecomposable. 
Remark 3.5.39. Recall that V. Ferenczi [34] has shown that if a complex Banach
spaces X is hereditarily indecomposable, then every operator from a subspace Y
of X into X is a scalar multiple of the inclusion operator plus a strictly singular
operator. We do not know if the analogous result is true for 1 -singular operators
or, in other words, if the converse implication of Lemma 3.5.37 is true in the case
of complex Banach spaces.
The reader is referred to [4] and [71] for more information about this new method
of using -functions in order to impose conditional structure on norms of Banach
spaces under construction. The following question now naturally suggests itself.
Question 3.5.40. What is the minimal cardinal u with property that every weakly
null sequence of length u contains an innite unconditional subsequence?
We have seen above that u 2 . From a result of Ketonen [58] we infer that
u is not bigger than the rst Ramsey cardinal. Not much more about u seems
to be known.
Chapter 4

Coherent Mappings and Trees

4.1 Coherent mappings


In this and the following section we present a general study of the following notion
already encountered at several places above.
Denition 4.1.1. A mapping a : [1 ]2 is coherent if for every < < 1
there exist only nitely many < such that a(, ) = a(, ), or in other words,
a = a  1 . We say that a is nontrivial if there is no h : 1 such that
h  = a for all < 1 .
Examples of coherent nontrivial mappings encountered so far are the map-
pings 1 , 3 , and . The general theory of coherent mappings is sometimes conve-
niently identied with the theory of corresponding trees dened as follows.
Denition 4.1.2. To every coherent mapping a : [1 ]2 one associates the
tree
T (a) = {a  : < 1 }
and its uniform closure
T (a) = {t : : < 1 and t = a }.
If the coherent mapping is given as a : [1 ]2 k where k is some specied
positive integer, then the corresponding uniform version of the associated tree is
denes as
T (a) = {t : k : < 1 and t = a }.
An arbitrary tree of height 1 is said to be a coherent tree if it can be represented
as a downward closed subtree of some tree of the form T (a), where a : [1 ]2
is a coherent mapping.
1A mapping a : [1 ]2 is naturally identied with a sequence a ( < 1 ), where a :
is dened by a () = a(, ).
92 Chapter 4. Coherent Mappings and Trees

Clearly, the tree T (a) corresponding to a coherent and nontrivial mapping


a : [1 ]2 is an Aronszajn tree (in short, A-tree). Note also that the existence
of a coherent and nontrivial a : [1 ]2 2 (such as, for example, the function
3 dened above) is something that corresponds also to the notion of a Hausdor
gap in this context. The theory of coherent mappings and coherent trees can
therefore be considered as a unied approach towards these two important classical
concepts.2
Denition 4.1.3. The support of a map a : [1 ]2 is the sequence

supp(a ) = { < : a(, ) = 0} ( < 1 )

of subsets of 1 . A set is orthogonal to a if supp(a ) is nite for all < 1 .


We say that a : [1 ]2 is nowhere dense if there is no uncountable 1
such that supp(a ) for all < 1 .
Note that 3 is an example of a nowhere dense coherent map for the simple
reason that 1 can be covered by countably many sets + n (n < ) that are
orthogonal to 3 . Recall, that here denotes the set of all countable limit ordinals
including 0 and that + n denote its translate { + n : }. The following
immediate fact shows that 3 is indeed a prototype of a nowhere dense and coherent
map a : [1 ]2 .
Proposition 4.1.4. Assuming the P-ideal dichotomy, for every nowhere dense and
coherent map a : [1 ]2 the domain 1 can be decomposed into countably
many sets orthogonal to a. 
The nowhere dense mappings like 3 occupy a special place in the theory of
coherent mappings and one of their special properties is revealed by considering
the following notion.
Denition 4.1.5. The shift of a : [1 ]2 is dened to be the mapping

a(1) : [1 ]2

determined by
a(1) (, ) = a( + 1, ),

where = min{ : }. The n-fold iteration of the shift operation is


dened recursively by the formula a(n+1) = (a(n) )(1) .
Theorem 4.1.6. If a is nontrivial, coherent and orthogonal to , then it holds that
T (a) > T (a(1) )3 .
2 Similarities between the notion of a Hausdor gap and the notion of an A-tree have been

further explained recently in the two papers of Talayco ([102], [103]), where it is shown that they
naturally correspond to rst cohomology groups over a pair of very similar spaces.
3 For two trees S and T , by S T we denote the fact that there is a strictly increasing map

f : S T . Let S < T whenever S T and T  S and let S T whenever S T and T S.


4.1. Coherent mappings 93

Proof. By the assumption a , the subtree S of T (a) consisting of all ts that


take the value 0 at every limit ordinal in their domains is an uncountable initial
part of T (a). Note that every node of T (a) is equal to the restriction of some nite
change of a node from S. For a node t at some limit level of S, let t(1) :
be determined by the formula t(1) () = t( + 1). Let S (1) = {t(1) : t S  }.
Every node of T (a(1) ) is equal to the restriction of some nite change of a node of
S (1) . Note also that t  t(1) is a one-to-one map on S  and that its inverse
naturally extends to a strictly increasing map from T (a(1) ) into T . This shows
that T (a(1) ) T (a).
It remains to establish T (a)  T (a(1) ). Note that if there is a strictly increas-
ing g : T (a) T (a(1) ), then there is one that is moreover level-preserving. For
example f (t) = g(t)  |t| is one such map. So let f : T (a) T (a(1) ) be a given
level-preserving map. For a countable limit ordinal , set
(1)
F = (supp(a ) ) { < : a () = f (a )()}. (4.1.1)

Then F is a nite set of ordinals < , so by the Pressing Down Lemma there
is a stationary set and F 1 such that F = F for all . Let
0 = max(F ) + 1. Shrinking if necessary, assume that for some s, t : 0 2,

a  0 = s and f (a )  0 = t for all . (4.1.2)

Choose < in such that a = a  and let = (a , a )(= min{ < :


a () = a ()}). Then by (4.1.1) and (4.1.2), is not a limit ordinal and
(1)
1 = (a , a(1)
) = (f (a ), f (a )). (4.1.3)

It follows that a and a split in T (a) after their images f (a ) and f (a ) split in
T (a(1) ), which cannot happen if f is strictly increasing. This nishes the proof. 

Corollary 4.1.7. If a is nontrivial, coherent and orthogonal to + n for all n < ,


then T (a(n) ) > T (a(m) ) whenever n < m < .

Proof. Note that if a is orthogonal to + n for all n < , then so is every one of
its nite shifts a(m) . 
(n) (m)
Corollary 4.1.8. T (3 ) > T (3 ) whenever n < m < . 
Remark 4.1.9. It follows that coherent trees are not well-quasi-ordered under any
quasi-ordering that is stronger than , and so in particular under the ordering
1 4 considered by Laver [68] when he proved that the class of -scattered trees is
w.q.o. under 1 . This in particular shows that the result of [68] does not extend
to any larger class of trees that would include the class of coherent trees.
4 Recall that S 1 T if there is a strictly increasing map f : S T with the property that
f (s t) = f (s) f (t) for all s, t S.
94 Chapter 4. Coherent Mappings and Trees

We now turn our attention to a metric theory of coherent mappings that will
give us an interesting 1 -denition of an ultralter on 1 . For this we need a piece
of notation.
Notation 4.1.10. To every a : [1 ]2 we associate the corresponding -
function a : [1 ]2 as follows:

a (, ) = min{ < : a(, ) = a(, )}

with the convention that a (, ) = whenever a(, ) = a(, ) for all < .
Given this notation, it is natural to let

a () = {a (, ) : , , < }

for an arbitrary set 1 .


Lemma 4.1.11. Suppose that a is nontrivial and coherent and that every uncount-
able subset of T (a) contains an uncountable antichain. Then for every pair and
of uncountable subsets of 1 there exists an uncountable subset of 1 such
that a () a () a ().
Proof. For each limit ordinal < 1 we x () and () above and
let
F = { < : a(, ()) = a(, ) or a(, ()) = a(, )}.
By the Pressing Down Lemma there is a stationary set 1 and a nite set F
such that F = F for all . Let 0 = max(F ) + 1. By the assumption about
T (a), nd an uncountable 0 such that a ( 0 ) is an antichain of T (a).
Moreover, we may assume that for some s, t and u in T (a):

a()  0 = s, a()  0 = t and a  0 = u for all 0 . (4.1.4)

It follows that for all < in 0 :

a ((), ()) = a (, ) = a ((), ()). (4.1.5)

So in particular, a (0 ) a () a (). 
Notation 4.1.12. For a : [1 ]2 , set

U(a) = {A 1 : A a () for some uncountable 1 }.

Note that Lemma 4.1.11 is saying that U(a) is a uniform lter on 1 for every
nontrivial coherent a : [1 ]2 for which T (a) contains no Souslin subtrees.
It turns out that under some very mild assumption, U(a) is in fact a uniform
ultralter on 1 .
Theorem 4.1.13. Under MA1 , the lter U(a) is an ultralter for every nontrivial
and coherent a : [1 ]2 .
4.2. Lipschitz property of coherent trees 95

Proof. Let A be a given subset of 1 and let P be the collection of all nite subsets
p of 1 such that a (p) A. If P is a ccc poset, then an application of MA1
would give us an uncountable 1 such that a () A, showing thus that A
belongs to U(a). So let us consider the alternative when P is not a ccc poset. Let
X be an uncountable subset of P consisting of pairwise incompatible conditions
of P. Going to an uncountable -subsystem of X and noticing that the root does
not contribute to the incompatibility, one sees that without loss of generality, we
may assume that the members of X are in fact pairwise-disjoint. Thus for each
limit ordinal < 1 we can x p in X lying entirely above . For < 1 , let
F = { < : a(, ) = a(, ) for some p }.
By the Pressing Down Lemma nd a stationary set 1 , a nite set F 1 ,
an integer n and t0 , . . . , tn1 in T (a) of height 0 = max(F ) + 1 such that for all
:
F = F, (4.1.6)
a  0 = ti with the ith member of p for some i < n. (4.1.7)
By MA1 the tree T (a) is special, so going to an uncountable subset of we may
assume that a ( ) is an antichain of T (a), and moreover that for every <
in :
a   a  for all p and p . (4.1.8)
It follows that for every < in :
a (, ) = a (, ) for all p and p . (4.1.9)
By our assumption that conditions from X are incompatible in P, we conclude
that for every < in , there exist p and p such that a (, )
/ A. So
from (4.1.9) we conclude that a () A = , showing thus that the complement
of A belongs to U(a). 
Remark 4.1.14. One may nd Theorem 4.1.13 a bit surprising in view of the fact
that it gives us an ultralter U(a) on 1 that is 1 -denable over the structure
(H2 , ). It is well known that there is no ultralter on that is 1 -denable over
the structure (H1 , ).

4.2 Lipschitz property of coherent trees


Given a tree5 T , let : T 2 Ord be dened by
(s, t) = otp{x T : x <T s and x <T t}.
5 The notion of a tree is to be interpreted in its order-theoretic sense, i.e., a partially ordered set
(T, T ) with the property that the predecessors of every point form a well-ordered chain. The
order-type of {x T : x <T t} is called the height of t in T and denoted by ht(t). The th level
of T is the set T = {t T : ht(t) = }. If ht(t), then t  denotes the s T t such that
ht(s) = . We make the implicit assumption that dierent nodes of the same level of T have
dierent sets of predecessors. This allows us to dene s t = max{x T : x s and x t}.
96 Chapter 4. Coherent Mappings and Trees

One should view as some sort of distance function on T by interpreting inequal-


ities like (x, y) > (x, z) as saying that x is closer to y than to z.
Denition 4.2.1. A partial map g from a tree S into a tree T is Lipschitz, if g is
level-preserving and
(g(x), g(y)) (x, y)
for all x, y dom(g).

Figure 4.1: Lipschitz map on a tree.

Note that this notion is equivalent to the notion of a strictly increasing level-
preserving map, when the domain of g is a downward closed subset of S, but
is otherwise more general. Any partial Lipschitz map g from S into T , however,
naturally extends to a strictly increasing level-preserving map g on the downward
closure of dom(g) in S, dened by letting g(x) = g(s)  ht(x) for some (equiva-
lently, for all) s dom(g) extending x.
Denition 4.2.2. A Lipschitz tree is any Aronszajn tree T with the property that
every level-preserving map from an uncountable subset of T into T is Lipschitz on
an uncountable subset of its domain.
Remark 4.2.3. It should be noted that if we restrict ourselves to countably branch-
ing trees of height 1 , there are essentially no Lipschitz trees outside the class of
Aronszajn trees. A tree of height 1 in which every node has extensions to all
higher levels cannot have an uncountable chain if it is to satisfy the requirement
of Denition 4.2.2. Using a bit of PFA one can also show that a Lipschitz tree
must have all levels countable.
The following fact shows that coherent trees that we have been encountering
so far all satisfy the strong metric-type condition about partial level preserving
maps and that therefore Lipschitz trees appear in profusion.
Lemma 4.2.4. A coherent tree T is Lipschitz if and only if every uncountable subset
of T contains an uncountable antichain.
4.2. Lipschitz property of coherent trees 97

Proof. The converse implication follows from Lemma 4.2.6 below, so let us con-
centrate on proving the direct implication. Let T be a given coherent downwards
closed subtree of the tree of all maps from countable ordinals into with the
property that every uncountable subset of T contains an uncountable antichain.
Thus, the height of a given node t of T is equal to its domain. Similarly, for two
dierent nodes s, t T of the same height ,

(s, t) = otp{x T : x <T s and x <T t} = min{ < : s() = t()}.

Consider a partial level-preserving map f from an uncountable subset X of T into


T . By our assumption, we may assume that both domain and the range of f are
antichains of T . Moreover, we may assume that X contains no two dierent nodes
of the same height. For t X, set

Dt = { < ht(t) : t() = f (t)()}.

Applying the -system lemma and shrinking X, if necessary, we may assume that
Dt (t X) forms a -system with root D. Shrinking X even further, we may
assume that:

(1) Ds \ D < Dt \ D whenever s, t X are such that ht(s) < ht(t),


(2) s  (max(D) + 1) = t  (max(D) + 1) for all s, t X.

Note that if Dt = D for all t X, then (1) and (2) ensure that

(s, t) = (f (s), f (t)) for all s = t in X, (4.2.1)

and so, in particular, the map f is Lipschitz. Thus we may assume that Dt \ D =
for all t X. For t X, let t = min(Dt \ D). By our assumption about T , we
can nd an uncountable subset Y of X such that both sets

{t  t : t Y } and {f (t)  t : t Y }

are antichains of T . It follows again that (s, t) = (f (s), f (t)) for all s = t in
X, as required. 

The following property of Lipschitz trees will be quite frequently and implic-
itly used in this and the following section.
Lemma 4.2.5. Suppose T is a Lipschitz tree, n is a positive integer, and that A
is an uncountable family of pairwise-disjoint n-element subsets of T . Then there
exists an uncountable B A such that (ai , bi ) = (aj , bj ) for all a = b in B
and i, j < n.

Proof. Replacing the family A by a family of the form a  a (a A) and applying


Lemma 4.2.4, we may assume that a given member of A is in fact a subset of some
98 Chapter 4. Coherent Mappings and Trees

level of T . Fix i, j < n and apply Denition 4.2.2 to the partial level-preserving
map
ai  aj (a A)

obtaining an uncountable A0 A such that

(ai , bi ) (aj , bj ) for all a, b A0 .

Applying Denition 4.2.2 to the inverse map aj  ai (a A0 ) will give us an


uncountable A1 A such that

(ai , bi ) = (aj , bj ) for all a, b A1 .

Repeating this procedure successively for every pair i, j < n, we reach the conclu-
sion of Lemma 4.2.5. 

Lemma 4.2.6. Every uncountable subset of a Lipschitz tree T contains an uncount-


able antichain. More generally, every family A of pairwise-disjoint nite subsets
of T contains an uncountable subfamily B such that B is an antichain of T .

Proof. Let X be a given uncountable subset of some Lipschitz tree T . It suf-


ces to nd an uncountable antichain in the downwards closure of X in T , so
we may in fact assume that X is already downward closed in T . We may also
assume that no x X is an end-node of T and in fact that every splitting
node x X has two successors x0 and x1 in X that split at x and have the
same height in T . Applying the fact that T is a Lipschitz tree to the partial map
f (x0 ) = x1 (x a splitting node of X) gives us an uncountable set Y of splitting
nodes of X such that f is Lipschitz on Y0 = {y0 : y Y }. It follows that Y0 is an
uncountable antichain of T . The second part of the lemma follows from the rst
and the fact that nite powers of T are Lipschitz as well (see Lemma 4.2.5). This
nishes the proof. 

The following result shows that under some mild axiomatic assumption the
class of Lipschitz trees in fact coincides with the class of coherent trees, the main
object of our study here.

Lemma 4.2.7. Under MA1 , every Lipschitz tree T is isomorphic to a coherent


downwards closed subtree of the tree of all maps from countable ordinals into .

Proof. Dene P to be the poset of all nite partial functions p from T 1 into
such that the following holds for all x = y in dom0 (p)6 :
6 Here and below, dom0 (p) = {t T : (t, ) dom(p) for some } and dom1 (p) = { 1 :
(t, ) dom(p) for some t T }.
4.2. Lipschitz property of coherent trees 99

(1) p(x, ) = p(y, ) for < (x, y),


(2) p(x, ) = p(y, ) for = (x, y)7 .
We let p extend q, if p extends q as a function, and
(3) p(x, ) = p(y, ) for all x, y dom0 (q) and < ht(x), ht(y) with the property
that / dom1 (q).
It is clear that a suciently generic lter will give us the desired embedding, and
so we concentrate on showing that P satises the countable chain condition.
So let p ( < 1 ) be a given sequence of elements of P. Let a be the
projection of dom0 (p ) onto the th level of T and let

h() = max{(s, t) + 1 : s, t a , s = t}.

Then there is a stationary set of countable limit ordinals on which the mapping
h, as well as the mapping

 dom(p ) (T  )

is constant. Let and F be these constant values respectively. We may assume that
all a s project to the same set a on the th level and all p s generate isomorphic
structures over , a and F . Thus, in particular we want the isomorphism between
the p s ( ) to respect a xed enumeration a (i) (i < n) of a , where n is the
common cardinality of these sets. As in the previous proof, we nd an uncountable
subset of such that for all = in :
(4) a (i) and a (j) are incomparable for all i, j < n,
(5) (a (i), a (i)) = (a (j), a (j)) for all i, j < n.
We claim that if = are in , then p and p are compatible in P. By (5), we
have an ordinal smaller than both and such that

(a (i), a (i)) = for all i < n.

Dene r P by letting its domain be

dom(p ) dom(p ) {(t, ) : t dom0 (p ) dom0 (p ), ht(t) > }

and letting r(t, ) = 0 if t dom0 (p ), ht(t) > and r(t, ) = 1 if t dom0 (p ),


ht(t) > . Note that r is indeed a member of P, as it clearly satises the conditions
(1) and (2) above. It is also easily checked that r extends both p and q, i.e.,
r(x, ) = r(y, ) for all x, y dom0 (p) or x, y dom0 (q) and is equal to , the
only new member of dom1 (r). This nishes the proof. 
7 In (1) and (2) we are making the implicit requirement that for every dom (p) and x =
1
y dom0 (p), if (x, y), then (x, ) dom(p) i (y, ) dom(p), and moreover, that always
(x, (x, y) and (y, (x, y)) belong to dom(p).
100 Chapter 4. Coherent Mappings and Trees

Denition 4.2.8. For a given tree T of height 1 ,

U(T ) = {A 1 : A (X) for some uncountable X T }.

Then one has the following analogue of Lemma 4.1.11 which in fact can be
deduced from it, though for a future motivation we include the more direct proof.
Lemma 4.2.9. The family U(T ) is a uniform lter on 1 for every Lipschitz tree
T . In particular, U(T ) is a uniform lter on 1 for every coherent tree T with the
property that every uncountable subset of T has an uncountable antichain.

Proof. Given two uncountable subsets X and Y of T , we need to nd an uncount-


able subset Z of T such that

(X) (Y ) (Z).

By the assumption about T , it is clear that we may replace X and Y by two


level-sequences x ( ) and y ( ) indexed by the same uncountable set
of limit ordinals. Moreover, we may assume that the x s and y s are all pairwise
incomparable (see Lemma 4.2.4). Apply Lemma 4.2.5 to the subset (x , y ) ( )
of T T and obtain an uncountable set such that

(x , x ) = (y , y ) for all , , = .

So we can take Z to be any of the sets {x : } or {y : }. This nishes


the proof. 

One also has the following analogue of Theorem 4.1.13 which can be in fact
deduced from it, but again, we give the direct proof.
Theorem 4.2.10. Under MA1 , the lter U(T ) is an ultralter for every coherent
tree T with no uncountable branches.

Proof. Let be a given subset of 1 . We need to nd an uncountable subset X of


T such that (X) is included in either or its complement. Let P be the poset
of all nite subsets p of T that take at most one point from a given level of T such
that
(p) = {(x, y) : x, y p, x = y} .
If P satises the countable chain condition, then a straightforward application
of (*) gives an uncountable X T such that (X) . So let us consider the
alternative that P fails to satisfy this condition. Let p ( < 1 ) be a sequence of
pairwise incomparable members of P . Going to a subsequence, we may assume
that every node of a given p has height at least , so we can dene a = the
projection of p on the th level of T . For , let

h() = {(x, y) + 1 : x, y a , x = y}.


4.2. Lipschitz property of coherent trees 101

Find stationary such that h is constant on . Let be the constant value of


h. Going to an uncountable subset of , we may assume that all a ( ) are of
equal size n and that they are given with an enumeration a (i) (i < n). Moreover,
we may assume that a (i)  = a (i)  for all , . Applying Lemmas 4.2.5
and 4.2.6, we obtain an uncountable such that for all = in :
(1) a (i) and a (j) are incomparable for all i, j < n,
(2) (a (i), a (i)) = (a (j), a (j)) for all i, j < n.
It follows that for all = in :
(3) (p p ) = (p ) (p ) {(a (0), a (0))}.
Since (p ) and (p ) are subsets of , (a (0), a (0))
/ must hold. This
gives rise to an uncountable set X = {a (0) : } with the property that
(X) = . This nishes the proof. 
The denition of U(T ) can be relativized to any countable elementary sub-
model M of (H2 , ) as follows.
Denition 4.2.11. For a coherent tree T of height 1 and a countable elementary
submodel M of H2 with T M , let UM (T ) be the collection of all subsets A of 1
for which one can nd uncountable X T with X M and t TM1 belonging
to the downward closure of X such that A (t, X) = {(t, x) : x X}.
It should be clear that the basic idea of the proof of Lemma 4.2.9 also gives
the following.
Lemma 4.2.12. For every coherent tree T and every countable elementary submodel
M of H2 such that T M , the family UM (T ) is a nonprincipal lter concentrating
on the set M 1 . 
From now on, we x a special coherent subtree T of the complete binary tree
2<1 that is closed under restrictions and nite changes of its elements. We also
x a subset K of T and use K c to denote its complement in T . Our goal is to nd
an uncountable subset X of T such that either (X) K 8 or (X) K c . Note
that in Theorem 4.2.10, assuming MA1 , we have achieved this in the case when
K is the union of levels of T . The general case requires a further analysis and a
stronger axiom. For J {K, K c } and t T , the corresponding ber is denoted by

J(t) = { < 1 : t  J}.



For a subset X of T and J {K, K c }, let J(X) = tX J(t). The lters of the
form UM (T ) are used to dene the following important notion.
Denition 4.2.13. A countable elementary submodel M of H2 rejects a nite
subset F of T if K c (F ) UM (T ).
8 Recallthat for x, y T by x y one denotes the maximal common predecessor of x and y, and
for a subset X of T one puts (X) = {x y : x, y X, x = y}.
102 Chapter 4. Coherent Mappings and Trees

Remark 4.2.14. Note that if an elementary submodel M of H2 such that {T, K}


M rejects a singleton {t}, then there is an uncountable antichain X of T such that
(X) K c . Note also that if M and N are two countable elementary submodels
containing T and K such that N M and M 1 = N 1 , and if N rejects F ,
then so does M . This follows from the inclusion UN (T ) UM (T ).
The following lemma suggests the direction of further analysis as it is def-
initely saying something about the poset of all nite approximations to an un-
countable subset X of T such that (X) K.
Lemma 4.2.15. Suppose we have an uncountable family A of pairwise-disjoint nite
subsets of T , all of some xed size n such that the collection M of all countable
elementary submodels of H2 which do not reject any member of the family A is a
stationary subset of [H2 ] . Then there exist a = b in A such that a(i) b(i) K
for all i < n.
Proof. Using Lemma 4.2.4 and going to an uncountable subfamily of A, we may
assume that
(ai , bi ) = (aj , bj ) for all a = b in A and i, j < n. (4.2.2)
Since M is stationary, we can nd M M such that {A, T, K} M . Pick a
b A such that the height of every node from b is bigger than = M 1 . Let
X = {a(0) : a A} and let t = b(0)  . Then (t, X) is a typical generator of
UM (T ) and by our assumption that M does not reject b, we conclude that
(t, X)  K c (b). (4.2.3)
Applying (4.2.2) and (4.2.3), we conclude that there must be a A M such that
a(i) b(i) K for all i < n. 
Corollary 4.2.16. Assuming MA1 , if there is an uncountable subset Y of T such
that the collection M of all countable elementary submodels of H2 which do not
reject any nite subset of Y is stationary in [H2 ] , then there is uncountable
X Y such that (X) K.
Proof. By Lemma 4.2.15, the poset of all nite F Y such that (F ) K
satises the countable chain condition. 
Let 0 = 2 (= (1 )+ ) and let C0 be the collection of all closed unbounded
subsets of [H0 ] consisting of countable elementary submodels M of (H0 , ) such
that {T, K} M . Similarly, let 1 = (21 )+ and let C1 be the collection of all
closed unbounded subsets of [H1 ] consisting of countable elementary submodels
1
M of (H1 , ) such that {T, K} M . Finally, let 2 = (22 )+ .
Lemma 4.2.17. Assume the Proper Forcing Axiom, and suppose a ( ) is a
given sequence of nite subsets of T indexed by some set 1 such that a T
for all . Then there is C C1 such that for every M C, if = M 1
then there is D M C0 such that every element of D M rejects a or no
element of D M rejects a .
4.2. Lipschitz property of coherent trees 103

Proof. Let P be the poset of all continuous -chains p = Np : p  of


countable elementary submodels Np of (H1 , ) containing all the relevant objects
such that, if Np 1 = belongs to and there D C0 Np such that no
N D Np rejects a , then there is < such that Np does not reject a for
all (, ); otherwise, there is < such that Np rejects a for all (, ).
It is easily checked that P ordered by end-extension is a proper poset, so an
application of PFA gives us a continuous -chain N : < 1  of countable
elementary submodels N of (H1 , ) with the property that if N = belongs to
, then either every term in a tail of the sequence N : <  does not reject a ,
or else, every member of a tail of the sequence N : <  rejects a depending
on whether D C0 N with no member of D N rejecting a , or not.
The conclusion of the lemma would follow if we show that for every countable
elementary submodel M of (H2 , ) containing all the objects accumulated so far
and having the property that = M 1 belongs to , there is D M C0 such
that every element of D M rejects a or no element of D M rejects a . We may
assume that the second alternative fails, i.e., that for every D C0 M there is
N D M rejecting a . Since N M , our assumption about M and a simple
elementarity argument shows that for every D C0 N there is N D N
rejecting a . It follows that there is 0 < such that N rejects a for all (0 , ).
Let D0 be the collection of all countable elementary submodels N of (H0 , )
such that 0 and N H0 : < 1  belong to N . Clearly, D0 is a closed and
unbounded subset of [H0 ] . We show that every N D0 M rejects a . So,
consider an N D0 M and let = N 1 . Then (0 , ) and so N rejects
a . Since N N , we conclude that N also rejects a . 

We are now in a situation to prove the following result of J.T. Moore [77].
Theorem 4.2.18. Assuming the Proper Forcing Axiom, for every coherent tree T
and every subset K of T there is an uncountable subset X of T such that either
(X) K or (X) K = . 

Proof. It suces to concentrate only on our xed special binary coherent tree
T . We shall assume that there is no uncountable X T such that (X) K c
and we shall show how to force something similar to the hypothesis of Corollary
4.2.16 so that its proof would give us uncountable X T such that (X) K.
There is a natural partial ordering PK for achieving this, the collection of all pairs
p = (Fp , Np ) such that:

(1) Fp is a nite subset of T ,


(2) Np is a nite -chain of countable elementary submodels of (H2 , ) contain-
ing all the relevant objects,
(3) for every N Np there is D C0 N such that Fp is not rejected by any
member of D N .
104 Chapter 4. Coherent Mappings and Trees

We consider PK as a poset ordered by coordinatewise inclusions. We shall prove


that PK is a proper partial order. This will nish the proof, since we are assuming
that no countable elementary submodel N of H0 rejects a singleton {t}, and any
condition of the form ({t}, {M }) with t TM1 forces that the union of the rst
coordinates of the generic lter is uncountable.
Suppose that for some large enough regular cardinal > 2 there is some
countable elementary submodel M of (H , ) containing all relevant objects and
there is some p PK M which cannot be extended to an M -generic condition
of PK . In particular, the condition

q = (Fp , Np {M H2 })

is not M -generic. So there is a dense-open subset D of PK such that D M and


there is extension r of q such that no r D M is compatible with r. Extending r,
we may assume that it belongs to D, so our assumption that the standard Fubini-
type argument (see, e.g., [111],8) that tries to build a copy r of r that belongs
to D M and that is compatible with r must fail. This results in two stationary
sets S [H1 ] and T [H2 ] and two mappings N  aN and N  bN with
domains S and T , respectively, and ranges included in [T ]< such that:
(4) for every N S there is a countable elementary submodel M of (H2 , )
such that N = M H1 ,
(5) for every N S there is D C0 N such that no model from D N rejects
the nite set aN ,
(6) for every N T there is D C0 N such that no model from D N rejects
the nite set bN ,
(7) for every N1 T and every N0 S N1 there is no D C0 N0 with
property that no member of D N0 rejects aN0 bN1 .
We may assume that S N for all N T . Furthermore, we may assume that for
every N T , the set bN has some xed size n and that bN TM1 . Moreover,
applying the Pressing Down Lemma and shrinking T , we may assume that for
some xed D0 C0 and some xed countable ordinal 0 and some xed n-element
subset of T0 , we have the following conditions on every N T :
(8) D0 N and no member of D0 N rejects bN ,
(9) s  [0 , N 1 ) = t  [0 , N 1 ) for all s, t bN ,
(10) bN  0 = F0 .
Let U be the collection of all n-element subsets a of T such that for some ordinal
> 0 , the set
T [a] = {N T : bN  = a}
is stationary. Note that the natural coordinatewise ordering makes U a tree which
is also coherent, being isomorphic to a subtree of {t T : t  0 = F0 (0)}. We shall
4.2. Lipschitz property of coherent trees 105

now see that there is a natural ccc poset which forces an uncountable antichain
A of U with the property that for all a = b in A there exist i < n such that
a(i) b(i)
/ K. Since no member of D0 rejects any member of A this would
contradict Lemma 4.2.15, and so the proof of Theorem 4.2.18 would be nished.
Let Q be the collection of all nite antichains q of U that meet a given level
of U in at most one point such that

{a(i) b(i) : i < n}  K for all a = b in q. (4.2.4)

We order Q by inclusion and we check that it satises the countable chain condi-
tion. So, let A be a given uncountable subset of Q. By a -system lemma, we may
assume that A forms a -system. Since the root cant contribute to incompati-
bility between two conditions from A, we may actually assume that it is empty
and that in fact A is an uncountable family of pairwise-disjoint antichains a of U
satisfying the condition 4.2.4 and all having some xed size m. We may view an
element a of A as a sequence a(i) (i < n) which enumerates a according to the
height in U . Since each a(i) is in turn an n-element subset of some level of T , we
can order it lexicographically as a(i)(j) = a(i, j) (j < n).
Choose N S such that 0 and A are elements of N and let M be a countable
elementary submodel of (H2 , ) such that N = M H1 . Let = N 1 and x
an element b of A, all of whose members are of height > . Fix also C0 C0 such
that no member of C0 N rejects aN . For each i < m, we choose Mi T such
that N Mi and bi is a restriction of bMi . Since every nite subset of T is equal
to the th member of a 1 -sequence belonging to M , applying Lemma 4.2.17 to
the restrictions of bMi s to the th level of T , and intersecting nitely many closed
and unbounded subsets of [H0 ] , and then relying on the condition (7), we can
nd N0 C0 N such that

K c (aN bi ) UN0 (T ). (4.2.5)
i<m

Let 0 = N0 1 . Pick 1 (0 , 0 ) such that elements of b agree on the interval


[1 , 0 ), and moreover, if two dierent nodes of b have dierent restrictions on ,
then they also have dierent restrictions on 1 . Using the elementarity of N0 , we
can choose a sequence b : < 1  N0 of elements of A such that b0  0 = b 
0 and such that for all < 1 , the nodes of b are of heights , they all agree
on the interval [1 , ), and if two of nodes of b have dierent restrictions on ,
then they also have dierent restrictions on 1 .
Using coherence of T , we nd uncountable subset of 1 in N0 such that the
node t = b(0, 0)  0 is in the downwards closure of the set X = {b (0, 0) : },
and such that

(t, X) K c (aN bi ). (4.2.6)
i<m
106 Chapter 4. Coherent Mappings and Trees

Since N0 belongs to C0 N , it does not reject aN , and so in particular, (t, X) is


not a subset of K c (aN ). So there must be N0 such that

aN (k)  (t, b (0, 0)) K for all k < |aN |. (4.2.7)

Using this and (4.2.6), we conclude that for all i < m there is j < n such that,

b(i, j)  (t, b (0, 0)) = b(i, j) b (i, j)


/ K. (4.2.8)

We shall show that b b satises the requirement (4.2.4) and this will nish the
proof. Consider i0 = i1 < m. If b(i0 )  0 = b(i1 )  0 , then there is j < n such that
b (i0 , j) b (i1 , j)
/ K. Then (b(i1 , j), b (i1 , j)) 1 > (b (i0 , j), b (i1 , j)),
and so we must have that

(b (i0 , j), b(i1 , j)) = (b (i0 , j), b (i1 , j)), (4.2.9)

and therefore, b (i0 , j) b(i1 , j) = b (i0 , j) b (i1 , j)


/ K. If, on the other hand,
b(i0 )  0 = b(i1 )  0 , then for all j < m,

(b (i0 , j), b(i1 , j)) = (b (i0 , j), b(i0 , j)) = (b (0, 0), b(0, 0)). (4.2.10)

By (4.2.8), there is j < m such that

b (i0 , j) b(i0 , j) = b(i0 , j)  (b (0, 0), b(0, 0))


/ K. (4.2.11)

Hence, b (i0 , j) b(i1 , j)


/ K. This nishes the proof. 
Remark 4.2.19. It turns out that the conclusion of Theorem 4.2.18 is a principle
of independent interest, so for future reference, we give it a name Coherent Tree
Axiom and denote it, in short, as CTA9 . Note that if T contains an uncountable
branch, the conclusion of CTA holds trivially, so the axiom is really about coherent
trees with no uncountable branches. It is also possible to show that, under MA1 ,
the axiom does not weaken if we restrict our attention to a xed coherent tree
with no uncountable branches such as for example T (3 ).
We nish this section with few concepts and facts about the general class of
A-trees10 .
Denition 4.2.20. A level-preserving partial map f : S T between two trees S
and T is pattern preserving if for every triple x, y, z dom(f ),

(x, y) > (x, z) i (f (x), f (y)) > (f (x), f (z)). (4.2.12)


9 The statement of a stronger axiom with T an arbitrary A-tree appears rst in a remark on page

79 of [1]. The relevance of the coherent version of the axiom became evident only after the paper
[123]. Needless to say, the deep arguments of [77] are built around the assumption that the tree
T is coherent.
10 Here, A-tree stands for an Aronszajn tree, an uncountable tree that has all levels countable as

well as all branches.


4.2. Lipschitz property of coherent trees 107

Remark 4.2.21. Note that if X = dom(f ) and Y = rang(f ) are antichains, this is
in particular saying that the corresponding subtrees (X) S and (Y ) T are
isomorphic. So the following fact will be quite useful in transferring statements
like CTA from coherent trees to other A-trees.
Lemma 4.2.22. Assuming the Proper Forcing Axiom, every level-preserving partial
map f : S T between two A-trees S and T of uncountable domain is pattern
preserving on an uncountable subset of its domain.
Proof. Let P be the set of all pairs p = (Xp , Cp ) of nite subsets of dom(f ) and
1 , respectively, such that 4.2.12 holds for all x, y, z Xp and such that for all
x, y Xp and Cp ,

(x, y) < i (f (x), f (y)) < . (4.2.13)

We let P be ordered by coordinatewise inclusions. To see that P is a proper


partial ordering, note that if M is a countable elementary submodel of some large-
enough structure of the form (H , ), then for every p P M , the extension
q = (Xp , Cp {M 1 }) is an M -generic condition of P. Note also that a condition
of the form ({x}, {}), where = M 1 and x dom(f ) \ M for such a model
M , forces the unions of the rst projection of the generic lter to uncountable.
Hence an application of the Proper Forcing Axiom gives us the conclusion of the
lemma. 
Lemma 4.2.23. Assuming the Proper Forcing Axiom, every level-preserving map
f from an uncountable subset of a coherent A-tree T into an A-tree U has an
uncountable X dom(f ) such that either
(1) (x, y) (f (x), f (y)) for all x = y X, or
(2) (x, y) < (f (x), f (y)) for all x = y X.
Proof. By Lemma 4.2.22, we may assume that f preserves the splitting patterns.
Choose an unbounded subset D of 1 and an uncountable subset X of dom(f )
such that
(x, y) < i (f (x), f (y)) <
for all x = y X and D. This can be done using the poset of nite approxi-
mations to X and D which is easily seen to satisfy the countable chain condition
using the fact that f preserves patterns. Taking the closure, we may assume that
D is actually closed and unbounded. Moreover, we may assume that both X and
its image f  X are antichains in the respective trees.
Rening the set X, we may assume that the downward closure T (X) of X in
T , the downward closure U (f  X) of f  X in U , and the downward closure S(X)
of the graph of f  X in the product tree T U 11 are all binary. That this can
be assumed follows from the easily checked fact that the poset of nite subsets
11 Recall that T U = {(x, y) T U : ht(x) = ht(y)} is a tree when ordered coordinatewise.
108 Chapter 4. Coherent Mappings and Trees

of an A-tree whose downward closures are binary satises the countable chain
condition. Rene X still further so that we can assume that the heights of any
pair of comparable splitting nodes of one of these three downwards closures are
separated by a member of the closed and unbounded set D. This again follows
from the fact that the poset of nite subsets p of X, with the property that the
comparable splitting nodes generated by p inside these three trees are separated by
D, satises the countable chain condition. Then to every splitting node t = x y
of T (X) there corresponds a unique splitting node ut = f (x) f (y) and this
association is independent of the choice of x = y X used to represent t.
Let K be the collection of all splitting nodes t of T (X) such that ht(t)
ht(ut ). Applying CTA, we nd an uncountable subset Y of X such that

(Y ) K or (Y ) K = . (4.2.14)

Then in the rst case, we have the alternative (1) while in the second case we have
the alternative (2) of the lemma. 

Remark 4.2.24. The conclusion of Lemma 4.2.23 in particular says that for every
level-preserving partial map f from an uncountable subset of a coherent A-tree T
into an A-tree U , there is an uncountable subset X of the domain of f such that
either the restriction f  X or its inverse is a Lipschitz map. We take the conclusion
of Lemma 4.2.23 as an axiom of independent interest since we are going to use it
at several places below. So we introduce a special name for it, the Lipschitz Map
Axiom and denote it as LMA .

4.3 The global structure of the class of coherent trees


In this section we study the class of coherent trees relative to the following quasi-
ordering relation.
Denition 4.3.1. For two trees S and T , by S T we denote the fact that there
is a strictly increasing map f : S T . Let S < T whenever S T and T  S
and let S T whenever S T and T S. When S T , we will say that the
trees S and T are equivalent trees.
Remark 4.3.2. Note that if there is a strictly increasing map f : S T from a
tree S into a tree T , then there is such a map g : S T that is moreover level
preserving, i.e., a strictly increasing map with the property that ht(f (x)) = ht(x)
for all x S. It suces to take g(x) = f (x)  ht(x). Note that strictly increasing
and level preserving maps from a downwards closed subtree of S into T are nothing
else than Lipschitz maps for this sort of domains. It is therefore natural to expect
some structure theory of the quasi-ordered class (C, ), where C denotes the class of
coherent trees of height 1 , or equivalently, the class of Lipschitz trees introduced
above in the previous section.
4.3. The global structure of the class of coherent trees 109

The basic fact which shows that coherent trees might have a special structure
under is the following.
Lemma 4.3.3. Under MA1 , every coherent tree T is irreducible in the sense that
T U for every uncountable downwards closed subtree U of T .
Proof. Let U be an uncountable downwards closed subset of T . We assume that
T has no uncountable branches and that every t U has extensions in U in all
levels above ht(t) and will use MA1 to produce a map witnessing T U . Let
P be the poset of all partial nite level-preserving Lipschitz maps from T into U ,
i.e., maps p with the property that
(p(x), p(y)) (x, y) (4.3.1)
for all x, y dom(p). Let us show that P satises the countable chain condition.
So let p ( < 1 ) be a given sequence of elements of P. Using the -system
lemma and arguments about coherent trees already encountered in the previous
section, we can nd < < 1 , an integer k, two nite subsets D and E of 1 ,
and an ordinal > max(D E) such that p p is a function:
(1) (dom(p ) dom(p ))12 = D {},
(2) (rang(p ) rang(p )) = E {},
(3) |dom(p ) \ D| = |dom(p ) \ D|,
(4) (x, y) = (p (x), p (y)), when x dom(p ) \ D and y dom(p ) \ D
occupy the same position in the increasing enumerations of these sets.
It should be clear now that p p satises the Lipschitz condition (4.3.1) and
therefore is a member of P.
For x T , let
Dx = {p P : x dom(p)}.
By our assumption on U , each Dx is a dense subset of P, so an application of
MA1 gives us the conclusion of the lemma. 
It follows that a coherent A-tree T is equivalent to all of its uncountable
downward closed subtrees. The following concept already appearing in the previous
section (though in a dierent guise) is quite useful in further analysis of the class
of Lipschitz tree, or equivalently, the class of coherent A-trees.
Denition 4.3.4. Given a pair of trees S and T of height 1 and having xed level
sequences s S ( < 1 ) and t T ( < 1 ), in order to make it clear where
the -function is being applied, we use the notation
s (, ) = (s , s ) and t (, ) = (t , t ).
12 Recallthat coherent trees are subtrees of trees of the form T (a) for a a coherent map, so the
notation (x, y) and (X) can be taken from the previous section. However, this could be as
easily dened for general trees T : For x, y T , we let (x, y) be the order type of the set of all
common (not necessarily strict) predecessors of x and y, and (X) = {(x, y) : x, y X} for
X T.
110 Chapter 4. Coherent Mappings and Trees

As a result of this notation, for 1 , we let

s () = {s (, ) : , , = } and t () = {t (, ) : , , = }.

For example, using this notation (xing arbitrary level sequences s S ( < 1 )
and t T ( < 1 )), and using the arguments from the previous sections, we
arrive at the following useful description of U(T ).
Lemma 4.3.5. For every U(T ) and uncountable 1 there is an uncountable
such that t () . 
It follows that sets of the form t () for uncountable 1 generate the lter
U(T ). Using these concepts one also gets a useful reformulation of the quasi-
ordering relation in the class of coherent A-trees.
Lemma 4.3.6. Assuming MA1 , the following are equivalent for every pair S and
T of coherent A-trees:
(a) S T ,
(b) there is an uncountable 1 such that s (, ) t (, ) for all , ,
= ,
(c) for every uncountable 1 there is an uncountable such that
s (, ) t (, ) for all , , = .
Proof. To deduce (b) from (a), suppose we are given a strictly increasing level-
preserving map f : S T . Apply Lemma 4.2.5 to (t , f (s )) ( 1 ) and obtain
an uncountable 1 such that for all = :
(1) t and t are incomparable,
(2) f (s ) and f (s ) are incomparable,
(3) (t , t ) = (f (s ), f (s )).
Clearly, this satises (b). Similarly one shows that (a) implies (c). Note that
clause (b) simply says that the map s  t ( ) is Lipschitz and it therefore
extends to a strictly increasing map from the downward closure S0 of the set
{s : } in S. By Lemma 4.3.3, S S0 T . This shows that (b), and
therefore the stronger (c), implies (a) and nishes the proof. 
The next lemma gives us a convenient reformulation of the inequality T  S.
Lemma 4.3.7. Assuming MA1 , the following are equivalent for every pair S and
T of coherent A-trees:
(a) T  S,
(b) there is an uncountable 1 such that s (, ) < t (, ) for all , ,
= ,
(c) for every uncountable 1 there is an uncountable such that
s (, ) < t (, ) for all , .
4.3. The global structure of the class of coherent trees 111

Proof. To see that (a) implies (b), let P be the poset of all nite p 1 such that
(1) t (, ) s (, ) for all , p, = .
If P would satisfy the countable chain condition, an application of MA1 would
give us an uncountable set 1 such that t (, ) s (, ) for all , ,
= which by Lemma 4.3.6 would give us T S, contradicting (a). So let
p ( 1 ) be a given sequence of pairwise incompatible members of P. We may
assume that min(p ) for all 1 . For 1 , let

a = {s  : p } and b = {t  : p }.

For , let h() be the maximum of all ordinals that have the form (x, y) + 1,
x, y a , x = y and (x, y) + 1, x, y b , x = y. Find a stationary such
that h is constant on and let be the constant value. Shrinking , we may
assume that all a ( ) are of some xed size m, and that all b ( ) are
of some xed size n. Let a (i) (i < m) and b (j) (j < n) be xed enumerations.
Applying Lemma 4.2.5, we can nd an uncountable such that for all =
in :
(2) a (i)  = a (i)  for all i < m,
(3) b (i)  = b (i)  for all i < n,
(4) a (i) and a (j) are incomparable for all i, j < m,
(5) b (i) and b (j) are incomparable for all i, j < m,
(6) (a (i), a (i)) = (a (j), a (j)) for all i, j < m,
(7) (b (i), b (i)) = (b (j), b (j)) for all i, j < n.
Consider = in . Then p p fails to satisfy condition (1), i.e., there exist
p and p such that (s , s ) < (t , t ). Let s  = a (i), s  =
a (j), t  = b (k) and t  = b (l). Using (2),(3),(4) and the fact that p and
p satisfy (1), we conclude that i = j and k = l and therefore, by (6) and (7), we
have the following:
(8) (a (0), a (0)) = (a (i), a (i)) < (b (k), b (k))
= (b (0), b (0)).
Of course, we may assume that the enumerations of a and b are given in a way
such that if () = min(p ), then s()  = a (0) and t()  = b (0) for all in
. Let = {() : }. Then satises clause (b), nishing thus the proof that
(a) implies (b). Similarly, one proves that (a) in fact implies (c). The implication
from (b) to (a) follows from Lemma 4.3.6. This nishes the proof. 
Denition 4.3.8. The uniform closure of a given downward closed subtree T of
<1 is dened to be

T = {x : : < 1 and x = t for some t T }.


112 Chapter 4. Coherent Mappings and Trees

If the tree T is given to us as a downward closed subtree of some tree of the form
k <1 where k < then it is natural to take its uniform closure T inside the set
k <1 as well in which case every node of T will branch into exactly k immediate
successors. In all these cases, we say that T is a uniform tree if T = T .
Uniform trees are natural whenever one considers the isomorphism relation
on some class of trees. The following lemma explains why in the context of coherent
trees it suces to study only the quasi-ordering and the associated equivalence
relation .
Lemma 4.3.9. Under MA1 , two uniform coherent A-trees of the same branching
degree k are isomorphic if and only if they are equivalent.
Proof. If S T , then by Lemma 4.3.6, there is an uncountable set 1 such
that s and t agree on pairs from . Let P be the poset of all partial nite level-
preserving isomorphisms p from S into T which extend to isomorphisms between
downward closures of dom(p) in S and rang(p) in T . Using the agreement of s
and t on and the argument from the proof of Lemma 4.3.7, one shows that P
satises the countable chain condition. The uniformity of the trees S and T are
used to show that all sets of the form

Dx = {p P : x dom(p)} (x S),

Ey = {p P : y rang(p)} (y T )
are dense-open in P. A lter of P which intersects all these sets gives us the
required isomorphism between S and T . 
Finally, we arrive at the following result that reveals another unexpected
structure in the class of coherent trees.
Theorem 4.3.10. Assuming MA1 , every two coherent trees are comparable.
Proof. Suppose we are given a pair S and T of coherent A-trees such that T  S.
By Lemma 4.3.7, there is an uncountable 1 such that s (, ) < t (, )
for all , , = . By Lemma 4.3.6, we conclude that T S. This completes
the proof. 
Remark 4.3.11. While under MA1 , the class of coherent trees is totally ordered by
, Corollary 4.1.8 gives us that this chain of trees is not well ordered. This should
be compared with an old result of Ohkuma [81] that the class of all scattered trees
is well ordered by and a result of Laver [68] that the class of -scattered trees
is w.q.o. under 1 . It turns out that the class of all A-trees is not totally ordered
under as the following result from [123] shows.
Theorem 4.3.12. There exist A-trees S and T such that S  T and T  S. In
fact, there is an uncountable family Ti (i I) of A-trees such that Ti  Tj for
i = j. 
4.3. The global structure of the class of coherent trees 113

The following concept (already appearing in the previous section though in a


dierent guise) will be quite useful in our further analysis of the class of coherent
A-trees (or, equivalently, the class of Lipschitz trees).
Denition 4.3.13. Two trees S and T are -equivalent on some set of ordinals ,
if there are level-sequences s S ( ) and t T ( ) such that for
every unordered triple , , of elements of ,
(s , s ) > (s , s ) i (t , t ) > (t , t ). (4.3.2)
Lemma 4.3.14. Assuming the Proper Forcing Axiom, every pair of coherent A-trees
S and T are -equivalent on some uncountable set of levels.
Proof. Fix level-sequences s S ( 1 ) and t T ( 1 ). Let P be the
set of all pairs p = (p , Cp ) of nite subsets of 1 such that S and T are equivalent
on p as witnessed by the xed level-sequences and such that for all , p and
Cp ,
(s , s ) < i (t , t ) < . (4.3.3)
We let P be ordered by coordinatewise inclusions. To see that P is a proper
partial ordering, note that if M is a countable elementary submodel of some large-
enough structure of the form (H , ), then for every p P M , the extension
q = (p , Cp {M 1 }) is an M -generic condition of P. Note also that a condition
of the form ({}, {}), where = M 1 for such a model M , forces the unions
of both projections of the generic lter to be uncountable. Hence an application
of the Proper Forcing Axiom gives us the conclusion of Lemma 4.3.14. 
Let us now turn our attention to the transformation T  U(T ) and show
that it provides a complete invariant of the quasi-ordered class (C, ) of coherent
A-trees.
Theorem 4.3.15. Under MA1 , two coherent A-trees are equivalent if and only if
the corresponding ultralters are equal.
Proof. Using Lemma 4.3.3, we may work with a pair of trees of the form T (a) and
T (b) where a and b are a pair of coherent and nontrivial mappings, and prove that
the condition T (a) T (b) is equivalent to the condition U(a) = U(b). Choose a
pair of strictly increasing mappings
f : T (a) T (b) and g : T (b) T (a).
Replacing f by the mapping t  f (t)  dom(t) and replacing g by the map-
ping t  g(t)  dom(t), we may assume that f and g are also level-preserving.
Recall our notation a () = a(, ) and b () = b(, ) that gives us particular
representatives a and b of the th level of T (a) and T (b) respectively. For ,
set
F = { < : a () = g(f (a ))() or b () = f (a )()}.
Then F is a nite subset of . Find a stationary subset of and F 1 such
that F = F for all . Find now an uncountable such that a ( )
114 Chapter 4. Coherent Mappings and Trees

is an antichain of T (a), b ( ) is an antichain of T (b), and the mapping


f (a )  ,
 (a  , g(f (a ))  ,
b  )

is constant on , where = max(F ) + 1. It follows that for = in ,


(a , a ) = (g(f (a )), g(f (a ))), (4.3.4)
(b , b ) = (f (a ), f (a )). (4.3.5)
Since f is strictly increasing, we must have (a ,a ) (f (a ),f (a )) = (b ,b ).
Since g is strictly increasing, we have that
(b , b ) = (f (a ), f (a )) (g(f (a )), g(f (a ))) = (a , a ).
It follows that (a , a ) = (b , b ) for all = in . From the proof of Lemma
4.1.11 we conclude that a () ( , uncountable) generates the lter U(a)
and similarly that b () ( , uncountable) generates the ultralter U(b). It
follows that U(a) = U(b).
Suppose now that U(a) = U(b). By symmetry, it suces to show that T (a)
T (b). Note again that the argument in the proof of Lemma 4.1.11 shows that there
is a stationary set such that a and b coincide on []2 . Let P be the poset
of all nite partial strictly increasing level-preserving mappings p from T (a) into
T (b) which (uniquely) extend to strictly increasing level-preserving mappings on
the downward-closures of their domains. By MA1 , it suces to show that P
satises the countable chain condition. So let p ( ) be a given sequence of
conditions of P. For , set
E = { < : a () = t() or b () = p(t)()
for some t dom(p ) of height }.
Then E is a nite subset of for each . Find a stationary set and
E 1 such that E = E for all . Shrinking , we may assume that the
mapping  p  is constant on . Let be the minimal ordinal strictly
above max(E) and the height of every node of dom(p )  for some (all) .
For , let p be the extension of p on the downward-closure of its domain
in T (a). Find a stationary such that each of the projections  p  ,
 a  and  b  is constant on . Find < in such that
a = a  , b = b  and p and p are isomorphic via the isomorphism that
xes their common restriction to the th levels of T (a) and T (b). It suces to
check that q = p p satises
(q(x), q(y)) (x, y) (4.3.6)
for every x dom(p ) and y dom(p ). If x and y correspond to each other in
the isomorphism between p and p , then either they belong to the common part
of p and p in which case (4.3.6) follows by the assumption that p and p are
strictly increasing level-preserving, or we have
(x, y) = (a , a ) = (b , b ) = (p (x), p (y)). (4.3.7)
4.3. The global structure of the class of coherent trees 115

If x is dierent from the copy y dom(p ) of y dom(p ), then (x, y) = (x, y)


and (p (x), p (y)) = (p (x), p (y)), so we conclude that (4.3.6) follows from
the fact that p is strictly increasing level-preserving. This nishes the proof. 
It turns out that the chain formed by the class of coherent trees in sequence
is discrete and in fact there is a natural shift operator giving the immediate suc-
cessor T (1) of a given coherent tree T .
Denition 4.3.16. For an integer m and a tree T , we let T (m) be its mth shift, the
downward-closure of {t(m) : t T  }, where for a limit node t of T , we let t(m)
be the function with the same domain as t dened by

t(m) () = t( m),

when m exists; otherwise (i.e., when m is positive and the largest limit ordinal
is less than m steps away), we let t(m) () = 0.
Remark 4.3.17. Note that a positive shift T (m) of any coherent tree T is coherent
and that the map t  t(m) is a strictly increasing map from T into T (m) . It
follows that T T (m) for all m 0. It is for this reason that we choose to switch
the direction of the shift operator on coherent mappings dened in the previous
section. Note also that for non-negative integers m and n,

T (m+n) = (T (m) )(n)

holds. Therefore we have that T (m) T (n) for every pair of non-negative integers
m and n such that m n.
Lemma 4.3.18. Suppose T is a coherent A-tree in which every uncountable sub-
set contains an uncountable antichain. Then T (m) < T (n) for every pair of non-
negative integers m and n such that m < n.
Proof. Suppose m < n are non-negative integers and consider a level-preserving
map f : T (n) T (m) . For each countable ordinal , pick a representative t from
the th level of T and let s T be such that
(n) (m)
f (t ) = s .

Working as before, we nd an uncountable set of countable limit ordinals such


that (t , t ) = (s , s ) for all , , = . Choose = in such that

= (t , t ) = (s , s )

is smaller than both and (i.e., t is incomparable to t and s is incomparable


to s ). Then
(n) (m) (m)
(t(n)
, t ) = + n > + m = (s , s ).

This shows that f is not a strictly increasing map, nishing the proof. 
116 Chapter 4. Coherent Mappings and Trees

The following lemma reveals that in the realm of coherent A-trees, the shift
T (1) is indeed the minimal tree above T .
Lemma 4.3.19. Under MA1 , for every pair S and T of coherent trees, the in-
equality S < T implies S (1) T .

Proof. Choose representatives s S ( < 1 ) and t T ( < 1 ). Working as


before, we show that there is an uncountable set of countable limit ordinals such
that (s , s ) < (t , t ) for all = in . We have already seen above during
the course of the proof of Lemma 4.3.18 that
(1)
(s(1)
, s ) = (s , s ) + 1 (t , t )

for all = in . This is sucient for helping us in showing that the poset
of all nite partial level-preserving Lipschitz maps from S (1) into T satises the
countable chain condition. An application MA1 will now give us the desired
conclusion S (1) T . 

It follows that (under MA1 ) for every coherent tree T , the chain

T (n) (n N)

of positive shifts is really an N-chain, i.e., its convex closure inside the class of
coherent A-trees is isomorphic to N as an ordered set. The case of negative shifts
is a bit more subtle though we shall see now that they also behave as expected.
Denition 4.3.20. A tree T is orthogonal to a set of ordinals , if there is an
uncountable subset X of T such that (x, y)
/ for all x, y X, x = y.

The proof of the following lemma is similar to the proof of Theorem 4.1.6
from the previous section.
Lemma 4.3.21. Suppose that n < m 0 and that T is a coherent A-tree which is
orthogonal to + k for all 0 k |n|. Then T (m) and T (n) are also coherent and
T (m)  T (n) . 

The following result gives us a sucient condition for the existence of a


Z-chain in the class of coherent A-trees under the total quasi-ordering .
Theorem 4.3.22. Under MA1 , for every coherent A-tree T which is orthogonal to
+ k for all k 0, the shifts T (n) (n Z) form a family of coherent A-trees with
the following properties:

(1) T (m+n) (T (m) )(n) ,


(2) T (n) < T (m) i n < m,
(3) no coherent tree S satises T (n) < S < T (n+1) for some n Z. 
4.3. The global structure of the class of coherent trees 117

Denition 4.3.23. Suppose that g is a partial map from 1 into 1 and that T is
a given tree represented as a downward-closed subtree of the tree of functions

{t : : < 1 }.

Then the g-shift of T , denoted by T (g) is the downwards-closure of

{t(g) : t T  },

where = { < 1 : g  } and t(g) is dened by

t(g) () = t(g())

if dom(g); otherwise t(g) () = 0.


Note that above, we considered the shifts T (g) for maps of the form g() =
m, but it should be clear that the arguments from that special case are sucient
to give us the following two facts.
Lemma 4.3.24. If g is a partial strictly increasing map on 1 and if rang(g) U(T )
for some Lipschitz tree T , then the g-shift T (g) is also a Lipschitz tree. 
Lemma 4.3.25. Suppose that T is a Lipschitz tree and g is a strictly increasing
partial map on 1 such that rang(g) U(T ). If g is regressive13 then T (g)  T .
On the other hand, if g is expanding14 then T  T (g) . 
Remark 4.3.26. Note that if g is a strictly increasing regressive partial map on
1 such that rang(g) = 1 , then T < T (g) holds for every Lipschitz tree T .
This observation can be used to construct both strictly increasing and strictly
decreasing 1 -sequences of Lipschitz trees.
Theorem 4.3.27. Assuming the Proper Forcing Axiom, for every pair S and T
of Lipschitz trees, there is a strictly increasing partial map g on 1 such that
S T (g) .
Proof. By Lemma 4.3.14, we have an uncountable 1 on which S and T are
-equivalent for some choice of level-sequences s S ( ) and t T (
). Dene

g : {(s , s ) : , , = } {(t , t ) : , , = }

by letting g((s , s )) = (t , t ). By (4.3.2), this is a well-dened strictly in-


creasing partial map on 1 such that dom(g) U(S) and rang(g) U(T ). Let

= { < 1 : g  }.

Assuming that {s : } and {t : } form antichains, then replacing each


s and t by their extensions on the ()th level of S and T respectively, where
13 I.e., g() < for all dom(g).
14 I.e., g() > for all dom(g).
118 Chapter 4. Coherent Mappings and Trees

() = min( \ + 1), and nally replacing with {() : }, we may assume


that is actually a subset of . By Lemma 4.3.24, the shift T (g) is a Lipschitz
(g)
tree and t ( ) is its level-sequence such that
(g)
(t(g)
, t ) = (s , s ) for all , , = . (4.3.8)

By Lemma 4.3.6 we conclude that S T (g) . This nishes the proof. 


Corollary 4.3.28. Assuming the Proper Forcing Axiom, for every pair S and T of
coherent A-trees, there is a strictly increasing partial map g on 1 which maps
U(S) into U(T ).

Proof. By Theorems 4.3.15 and 4.3.27 we may assume that in fact S = T (g) for
some strictly increasing partial map g on 1 with rang(g) U(T ). Now note that
dom(g) U(T (g) ) and that g  A U(T ) for every A U(T (g) ). 

The following result shows that the class of coherent trees has a very special
relationship with the class of all A-trees.
Theorem 4.3.29. Assuming the Proper Forcing Axiom, every coherent tree is com-
parable with every Aronszajn tree.

Proof. By Lemma 4.2.23, we can concentrate on showing that LMA together with
MA1 yield the conclusion of the theorem. Let T be a given coherent A-tree and
let S be any other A-tree. Let P be the poset of all nite partial level-preserving
Lipschitz maps from S into T . If P satises the countable chain condition, then
an application of MA1 would give us a total strictly increasing map f : S T
witnessing thus the relation S T . So we are left with the alternative that P
fails to satisfy the countable chain condition. Let X be an uncountable family of
pairwise incomparable members of P. We may assume that X forms a -system.
Removing the root, which obviously does not contribute to the incomparability
between members of X , we may in fact assume that dom(p) (p X ) is a family
of pairwise-disjoint nite subsets of S, all of some xed size n. For p X , let
si (p) (i < n) enumerate dom(p) and let ti (p) = p(si (p))(i < n) enumerate its
range. Applying the -System Lemma again, we obtain an uncountable subfamily
Y of X such that the families of nite subtrees

{(si (p), sj (p)) : i < j < n)} (p Y),

and
{(ti (p), tj (p)) : i < j < n)} (p Y)
both form -systems. Shrinking the family Y even further, we may assume that if
is the least upper bound of the roots of these two -systems, then for all p = q
in Y:
si (p)  = si (q)  and ti (p)  = ti (q)  for all i < n.
4.3. The global structure of the class of coherent trees 119

Since any pair p = q of conditions from Y are incompatible in P, it follows that


there exists i = i(p, q) < n such that:

(si (p), si (q)) > (ti (p), ti (q)).

Apply the LMA successively n times and obtain an uncountable subset Z of Y


such that for all i < n, the map

si (p)  ti (p) (p Z)

or its inverse is Lipschitz. If for every i < n, the map si (p)  ti (p) (p Z) is
Lipschitz, we would obtain two compatible members of X contradicting our initial
assumption about X . Hence, there must be i < n so that the inverse map

g : ti (p)  si (p) (p Z)

is Lipschitz. Let T0 be the downward-closure of {ti (p) : p Z} in T . Then T0 is


uncountable and g extends to a Lipschitz map g : T0 S which is in particular
strictly increasing. It follows that T0 S. Since T is irreducible, we have that
T T0 , and therefore T S. This nishes the proof. 
The following analogue of Lemma 4.3.19 shows that the shift transformation
is giving us a true successor operation valid in the class of all trees.
Theorem 4.3.30. Assuming the Proper Forcing Axiom, no tree lies strictly between
a coherent tree T and its shift T (1) .
Proof. Suppose there is a tree S such that T < S < T (1) . Clearly, we may assume
that S is Aronszajn. During the course of the proof of Theorem 4.3.29, we have
seen that, since the poset of all nite Lipschitz maps from S into T fails to satisfy
the countable chain condition (since we are assuming MA1 ), an application of
LMA to an uncountable family of pairwise incomparable members of this poset
will provide us with an uncountable set and level-sequences s S ( ),
t T ( ) such that

(s , s ) > (t , t ) for all = in .

Let S0 be the downward closure of {s : } in S. Consider now the poset of all


nite partial Lipschitz maps from T (1) into S0 . By MA1 , this poset cannot satisfy
the countable chain condition either. So applying LMA again to an uncountable
family of pairwise incompatible members of this poset would give us uncountable
(1) (1)
level-sequences r S ( ), q T ( ) such that
(1)
(q(1) , q ) > (r , r ) for all = in .

For each , choose () such that

() and r = s()  .
120 Chapter 4. Coherent Mappings and Trees

(1)
We may also assume that {q : }, {q : } and {r : } are all
antichains in their respective trees. For each , nd p() T() such that
p()  = q and therefore
(1)
p()  = q(1) .
Applying the already standard coherent tree argument, we get an uncountable
0 such that

(t() , t() ) = (p() , p() ) for all = in 0 .

Combining all this, we get the following for all = in 0 :


(1) (1) (1)
(p() , p() ) = (q(1) , q ) > (r , r ) = (s() , s() )

> (t() , t() ) = (p() , p() ).


It follows that for = in 0 :
(1) (1)
(p() , q() ) (p() , p() ) + 2,

a contradiction. This nishes the proof. 


Remark 4.3.31. The very special properties of the discrete chain of coherent trees
described above might give us an impression that the real reason behind the so
ne structure theory lies in the fact that the class of trees is rather small. The
results that follow show, however, that this is not the case.
Lemma 4.3.32. Assuming MA1 , for every A-tree S, there is a coherent A-tree T
such that S T .
Proof. Let P be the set of all nite partial functions p from S 1 into such
that:
(1) < ht(x) for all (x, ) dom(p),
(2) p(x, ) = p(y, ) for all (x, ), (y, ) dom(p) with < (x, y).
We let p extend q if p extends q as a function and
(3) p(x, ) = p(y, ) for all (x, ), (y, ) dom0 (q) and < ht(x), ht(x) such that
/ dom1 (q(x)) dom1 (q(y))15 ,
(4) p(x, ) = q(x, ) for all (x, ) dom(q) and (x, ) dom(p) \ dom(q).
A simple -system argument (contained in the proofs of Lemmas 4.3.3 and 4.2.7
above) shows that P satises the countable chain condition, so an application of
MA1 gives us a map g from S 1 into so that its bers gx () = g(x, ) are
total maps from ht(x) into for all x S and such that:
15 Recall
the notations dom0 (p) = {x : (x, ) dom(p) for some }, dom1 (p) = { : (x, )
dom(p) for some x}, and dom1 (p(x)) = { < ht(x) : (x, ) dom(p)}.
4.3. The global structure of the class of coherent trees 121

(5) gx : ht(x) is a nite-to-one map for all x S,


(6) (x, y) (gx , gy ) for all x, y S,
(7) { : gx () = gy ()} is nite for all x, y S.
It follows that the downwards-closure T of {gx : x S} is a coherent A-tree and
that x  gx is a Lipschitz map from S into T sucient for witnessing the relation
S T . This nishes the proof. 
Lemma 4.3.33. Assuming MA1 , for every A-tree S, there is a coherent tree T
such that T S.

Proof. Let P be the set of all nite partial functions p from S 1 into such
that:
(1) < ht(x) for all (x, ) dom(p),
(2) for every pair x and y of incomparable nodes from dom0 (p), there is
(x, y) with (x, ), (y, ) dom(p) and p(x, ) = p(y, ).
We let p extend q if p extends q as a function and
(3) p(x, ) = p(y, ) for all (x, ), (y, ) dom0 (q) and < ht(x), ht(x) such that
/ dom1 (q(x)) dom1 (q(y)),
(4) p(x, ) = q(x, ) for all (x, ) dom(q) and (x, ) dom(p) \ dom(q).
To prove that P satises the countable chain condition, we start with an uncount-
able subset X of P and perform the -system argument from the proof of Lemma
4.3.3, obtaining two conditions p and q in X such that for some < < ,
v0 , . . . , vn S, s0 , . . . , sn S and t0 , . . . , tn S we have:

(5) every node of dom0 (p) is either of height < or it extends some si (i n),
(6) every node of dom0 (q) is either of height < or it extends some ti (i n),
(7) dom1 (p) (, ) and dom1 (q) (, 1 ),
(8) vi = vj for i = j n,
(9) si and ti extend vi but are incomparable for all i n,
(10) p and q are isomorphic conditions via an isomorphism that is the identity on
vi (i n) and maps si to ti for all i n.
,
We claim that such p and q can be amalgamated into a condition r of P that
extends them both. Let

= min{(si , ti ) : i n}.

Then < . Let

k = max(rang(p)) = max(rang(q)).
122 Chapter 4. Coherent Mappings and Trees

Let dom(r) be equal to the union of dom(p), dom(q) and the following two sets:

D = {(x, ) : x dom0 (p), ht(x) },


E = {(y, ) : y dom0 (q), ht(y) }.

Dene r by giving it constant value k + 1 on D and constant value k + 2 on E.


Note that r satises (1) and (2) as well as conditions (3) and (4) for extending
both p and q.
Applying MA1 to P and an appropriate family of dense-open subsets of P
gives us a partial map g : S 1 so that if gx () = g(x, ) then:
(18) gx is a nite-to-one map from ht(x) into for all x S,
(19) (gx , gy ) (x, y) for all x, y S,
(20) { : gx () = gy ()} is nite for all x, y S.
It follows that the downwards-closure T of {gx : x S} is a coherent tree and
that gx  x is a partial Lipschitz map from T into S, sucient for witnessing
the relation T S. This nishes the proof. 
Theorem 4.3.34. Assuming MA1 , there is no maximal Aronszajn tree.
Proof. Given an Aronszajn tree S by Lemma 4.3.32 we nd a Lipschitz tree T
such that S T . By Lemma 4.3.18, T < T (1) , so in particular S < T (1) . 
Theorem 4.3.35. Assuming the Proper Forcing Axiom, for every coherent tree T
there is a coherent tree S such that S < T .
Proof. Fix a level-sequence t T ( 1 ) in a given Lipschitz tree T . Let
t : [1 ]2 1 be the corresponding distance function t (, ) = (t , t ).
For 1 , let
t () = {t (, ) : , , = }.
Let P be the set of all pairs p = (fp , p ) such that:
(21) p is a nite subset of 1 ,
(22) fp is a nite partial strictly increasing map from 1 into 1 which can be
extended to a total strictly increasing and continuous map f : 1 1 so
that rang(f ) is disjoint from t (p ) and separates16 the points of t (p ).
We order P by coordinatewise inclusion. To show that P is proper, consider a
countable elementary submodel M of some large-enough structure of the form
(H , ) such that M contains P, T and the level-sequence t ( 1 ). For a given
p P M , let
q = (fp {, }, p ),
where = M 1 . We claim that q is an M -generic condition of P. To show this,
consider a dense-open subset D of P such that D M and an extension r of q. We
16 i.e., between every two members of t (p ), there is a member of rang(f ).
4.3. The global structure of the class of coherent trees 123

need to show that r is compatible with a member of D M . Extending r, we may


assume r D. Let vi (i n) be a one-to-one enumeration of {t  : p \ }.
Let p = r  M . Then p P M and so we can nd an extension f : 1 1 of
fp satisfying (22) for p such that f M . Let (max(p), ) be a xed point of
f. Find a copy r of r in D M such that if we let and vi (i n) be its versions
of and vi (i n), then
(23) vi  = vi  for all i n,
(24) vi and vi are incomparable for all i n,
(25) (v0 , v0 ) = = (vn , vn ).
It is clear that we can combine the function f with the normal functions witnessing
(22) for r, r and obtain a strictly increasing continuous function f : 1 1
which xes and witnesses (22) simultaneously for r, r and moreover, the ordinal
(v0 , v0 ) is not in its range. Since

t (r r ) = t (r) t (r ) {(v0 , v0 )},

this shows that (fr fr , r r ) is a member of P witnessing the compatibility


of r and r.
Applying the Proper Forcing Axiom to P gives us an uncountable 1
and a closed unbounded set C 1 such that C t () = and C separates the
points of t (). For < 1 , let + be the minimal point of C above . Dene

C0 = { C : (, + ) t () = }.

Note that for C, there is only one point of t () in the interval (, + ). Call
this point g(). This denes a strictly increasing map g from C0 onto t (). So
in particular, rang(g) U(T ). Let S = T (g) (see Denition 4.3.23). From Lemma
4.3.24 we conclude that S is a Lipschitz tree. Since clearly g() > for all C0 ,
we conclude that T  S from Lemma 4.3.25. This nishes the proof. 
Corollary 4.3.36. Assuming the Proper Forcing Axiom, there is no minimal Aron-
szajn tree.
Proof. This follows from Lemma 4.3.33 and Theorem 4.3.35. 
Corollary 4.3.37. Assuming the Proper Forcing axiom, the class C of all coherent
Aronszajn trees is a discrete chain which is coinitial and conal in the quasi-
ordered class (A, ) of all Aronszajn trees. Moreover, neither the chain C nor the
class A has a maximum nor a minimum. 
124 Chapter 4. Coherent Mappings and Trees

4.4 Lexicographically ordered coherent trees


The results of the previous section suggest that coherent trees are likely to be
relevant to questions surrounding the class of Aronszajn trees and therefore in
many questions about uncountable structures. In this section we give one example
of such a phenomenon by considering the basis problem for uncountable linear
orderings. We start by analysing the lexicographical orderings of coherent trees.
Theorem 4.4.1. The cartesian square of a lexicographically ordered coherent tree
that admits a strictly increasing mapping into the rationals17 can be covered by
countably many chains.
Proof. Let T be a given special coherent tree. Identifying a node t : of T
with the characteristic function (t) of its graph
(t) = {(, t()) : < },
the natural bijection between and the ordinal will move (t) to a function
t : 2. Since the transformation t  t preserves the lexicographical ordering
and since the downward-closure of {t : t T } remains special, we may assume
that our tree T itself consists of maps whose ranges are included in {0, 1}. By our
assumption, the tree T can be decomposed into countably many antichains. So let
a : T be a xed map such that a(s) = a(t), whenever s <T t and such that
a is one-to-one on the levels of T . It suces to decompose the subset
{(s, t) T 2 : ht(s) ht(t)}
of the cartesian square of (T, <lex ) into countably many chains. Note that the
existence of a allows us to replace (modulo a countable decomposition) every such
ordered pair (s, t) by the pair (s, t  ht(s)). It follows that it suces to show that
if f : D T is a level-preserving map from a subset X of T which intersects a
given level of T in at most one point, then its graph f = {(x, f (x)) : x X} can
be decomposed into countably many cartesian chains.
For t X, let
Dt = { 1 : = ht(t) or < ht(t) and t() = f (t)()}.
Then by the coherence of the tree T , it follows that Dt is a nite set of countable
ordinals for all t X. Let pt : Dt and qt : Dt be the functions
dened by
pt () = a(t  ) and qt () = a(f (t)  ).
It suces to show that if for some s <lex t in X, we have that
t  Dt
= s  Ds , pt  Dt
= ps  Ds , and qt  Dt
= qs  Ds ,
then f (s) <lex f (t).
17 Recallthat trees admitting strictly increasing maps into the rationals are in the literature
frequently called special
4.4. Lexicographically ordered coherent trees 125

Note that from the isomorphisms between the parameters ps and pt it fol-
lows in particular that a(s) = a(t), and similarly from the isomorphism between
parameters qs and qt we can conclude that a(f (s)) = a(f (t)). It follows that s and
t as well as f (s) and f (t) are incomparable in T .
Claim. (s, t) = (f (s), f (t)).
Proof. Let us rst establish the inequality

(f (s), f (t)) (s, t).

To this end, let = (s, t). Then < ht(s), ht(t), since s and t are incomparable.
By the properties of our parameters s and t, we have that Ds = Dt and
that f (s) and f (t) agree on this set. From the denition of Ds and Dt , we conclude
that f (s) and f (t) must agree below , and so (f (s), f (t)) . A completely
symmetric argument will give us the other inequality

(s, t) (f (s), f (t)). 

Let = (s, t) = (f (s), f (t)). Note that cannot belong to the intersection of
Ds and Dt since then it would occupy the same position in these sets, and so using
the isomorphism t  Dt = s  Ds , we would have that s() = t(), contradicting
the assumption that = (s, t). So, we are left with considering the following
three cases.
Case 1. Suppose rst that Ds \ Dt . Then

s() < t() = f (t)(),

and since f (t)() can take only two values 0 and 1, we conclude that f (t)() = 1.
Since f (s)() = f (t)(), using the same reasoning, we conclude that f (s)() = 0,
giving as the desired inequality f (s) <lex f (t).
Case 2. Suppose now that Dt \ Ds . Then

f (s)() = s() < t(),

and so as before, we must have that f (s)() = 0. Since f (s)() =  f (t)(), we


must have that f (t)() = 1, giving us again the desired inequality f (s) <lex f (t).
Case 3. Finally, suppose that
/ Dt Ds . Then

f (s)() = s() < t() = f (t)()

giving us the desired inequality f (s) <lex f (t). 


Corollary 4.4.2. Assuming MA1 , the cartesian square of a lexicographically or-
dered coherent Aronszajn tree is the union of countably many chains. 
126 Chapter 4. Coherent Mappings and Trees

Let us now see how this is related to the basis problem for uncountable linear
orderings, or equivalently and more naturally, co-basis problem for the class of
countable linear orderings. Thus, one would like to have a nite list K of linear
orderings with the property that an arbitrary linear ordering L is countable if and
only if K  L for all K K. Note that if such a co-basis K exists and if it is
moreover a minimal such co-basis, then every ordering K in K will have to be
rather canonical or at least minimal as an uncountable linear ordering18 Clearly,
1 and 1 will be members of any co-basis for the countable linear orderings.
Another member of the co-basis is found by considering the following notion.
Denition 4.4.3. A linearly ordered set L is separable if it contains a countable
dense subset D with the property that for all x <L y in L there is d D such
that x L d L y.
Clearly, for every uncountable separable linear ordering L there is an un-
countable set of reals X such that X L. The following result of Baumgartner
[10] solves this part of the co-basis problem for the class of countable linear order-
ings.
Theorem 4.4.4. Assuming the Proper Forcing Axiom, every set of reals of size 1
embeds into any uncountable separable linear ordering. 
Recall, that a set of reals B of cardinality 1 can indeed be canonically cho-
sen on the basis of the given C-sequence C ( < 1 ). For example we could
take B = B(0 ) to be the set T (0 ) as a subset of the Cantor cube 2Q ordered
lexicographically. It follows that under the Proper Forcing Axiom any such set of
reals B will appear as a member of a co-basis for the class of countable linear or-
derings, if such a co-basis exists at all. It remains to concentrate on the orthogonal
to {1 , 1 , B}.
Denition 4.4.5. A linearly ordered set L is Aronszajn 19 if it is uncountable and
if it contains no uncountable subordering which is well ordered, conversely well
ordered, or separable.
It remains to nd a basis for the class of Aronszajn orderings. The following
result of J.T. Moore [77] accomplishes this.
Theorem 4.4.6. Assuming the Proper Forcing Axiom, every Aronszajn ordering
contains an uncountable subordering whose cartesian square can be covered by
countably many chains.
Proof. To attain the conclusion, we shall actually need only the Coherent Tree
Axiom (see Remark 4.2.19) and a bit of PFA restricted to orderings of size at
18 Recall that an uncountable linear ordering K is minimal, if it has the property that every

linear ordering L such that L < K must be countable.


19 The name comes from the fact that any partition tree of such an ordering is Aronszajn and

that a lexicographical ordering of any Aronszajn tree is an Aronszajn ordering (see, for example,
[105]).
4.4. Lexicographically ordered coherent trees 127

most 1 . So let L be a given Aronszajn ordering and let T be one of its binary
partition trees (see [105]). Thus L can naturally be identied with the subset
{{x} : x L} of T . Choose an uncountable subset X of T (3 ) which intersects a
given level of T in at most one point and a level-preserving map

f : X L T.

By Lemma 4.2.22, going to an uncountable subset of X, we may assume that f


preserves splitting patterns. Applying LMA and going to an uncountable subset
of X, we may assume that f is Lipschitz (the case that f 1 is Lipschitz is con-
sidered similarly). Let U (f ) be the downwards-closure of the graph of f in the
tree-product T (3 ) T . Shrinking X if necessary we may assume that U (f ) is also
a binary tree. This is so because the poset of nite subsets of the graph of f whose
downward-closures are binary satises the countable chain condition. Now, choose
an uncountable subset Y of X and an unbounded subset D of 1 , such that

[(x, y), (f (x), f (y))] D = for all x = y Y. (4.4.1)

That this can be done follows from the fact that the natural poset of nite ap-
proximations to D and T is easily seen to satisfy the countable chain condition
using the fact that f preserves splitting patterns. Taking the closure of D we may
assume that D is closed and unbounded. Going to an uncountable subset of Y ,
we may assume that the heights of every pair of comparable splitting nodes of
U (f  Y ) are separated by a member of D. That this can be done follows from the
fact that the poset of nite subsets of Y with this property satises the countable
chain condition. It follows that to every splitting node s = x y of the subtree
S(Y ) of T (3 ) generated by taking the downwards-closure of Y , there corresponds
a splitting node ts = f (x) f (y) and the association s  t does not depend on
which x and y we pick in Y to represent s. Let K be the collection of all split-
ting nodes s = x y of S(Y ) such that x <lex y i f (x) <L f (y). Note again
that the denition does not depend on the x and y in Y we choose to represent
s since the reason for the inequality f (x) <L f (y) relies on the fact that f (x)
must lie in the left and f (y) in the right immediate successor of the splitting node
ts = f (x) f (y). By CTA, there is an uncountable subset Z of Y such that either

(1) {x y : x, y Z, x = y} K, or
(2) {x y : x, y Z, x = y} K = .

In the rst case, the restriction f : (Z, <lex ) L is strictly increasing and in the
second case, it is strictly decreasing. In any case, by Corollary 4.4.2, the image
f  Z is an uncountable subordering of L whose cartesian square can be covered by
countably many chains. 

Combining Theorems 4.4.4 and 4.4.6 and Corollary 2.1.13, we obtain the
following.
128 Chapter 4. Coherent Mappings and Trees

Corollary 4.4.7. Assuming the Proper Forcing Axiom, a linearly ordered set L is
countable if and only if it contains no isomorphic copy of any ordering from the
list 1 , 1 , B, C(0 ), and C(0 ) . 
As indicated above the role of B in this result can be played by any set of
reals of size 1 and the role of C(0 ) in this result can be played by C(1 ), C(3 ),
or C(a), where a : [1 ]2 is any nontrivial coherent mapping. Here, as before,
we let C(a) be the linearly ordered set (1 , <a ), where

<a if and only if a(, ) < a(, )

for = min{ : a(, ) = a(, }. By Corollary 4.4.2, the cartesian square of any
such ordering can be covered by countably many chains.

4.5 Stationary C-lines


Recall that a C-line is any uncountable linear ordering A whose cartesian square
A2 can be covered by countably many chains. We have seen above many con-
structions of such objects based on dierent characteristics of the walk on . The
purpose of this section is to examine a particular class of C-lines from a topo-
logical viewpoint. So, let A be a given C-line and let D be a countable subset of
A. Then the complement of D in A is split into a family TD of convex sets, the
equivalence classes of the relation in which x and y are equivalent if no member
of D separates them. The members of TD will also be called the complementary
intervals of D. Note that TD must be countable or else A would contain a subset
order-isomorphic to a set of reals. For two countable subsets A0 and A1 , we say
that A1 properly extends A0 if A0 A1 and A1 has points in everyone of the
complementary intervals of A0 .
Denition 4.5.1. A proper decomposition of a C-line A is a sequence A ( < 1 )
of countable
 subsets of A such that A properly extends A if < , and such
that A = < A if is a limit ordinal.
Note that every 1 -sequence of properly extending countable subsets of A
must have union equal to A. On the other hand, trivially no properly extending
sequence can have length > 1 , so there is only one possibility for the length of
any proper decomposition {A } of A, it must be equal to 1 . Given such a proper
decomposition {A }, let

TA = T A .
<1

This is the partition tree of A associated with a proper decomposition {A } of A.


Needless to say this is an Aronszajn tree. Note also that T uniquely determines
the proper decomposition {A } of A, so their roles will be frequently interchanged
below. We shall also simplify the notation and write T for TA .
4.5. Stationary C-lines 129

Denition 4.5.2. If is a countable limit ordinal, then we say that a proper decom-
position {A } of a C-line A is continuous20 at if every complementary interval
of A has a minimal element.
This leads us to the following central notion of this section.
Denition 4.5.3. We shall say that A is a stationary 21 C-line if and only if ev-
ery proper decomposition {A } of A is (left) continuous at some countable limit
ordinal.
Note that if A is a stationary C-line, then the set of continuities of every
proper decomposition of A is a stationary subset of 1 , explaining thus our choice
of the name for this concept. It turns out that this stationarity of a C-line A
viewed as a topological space with its natural topology is the key to many of its
covering properties, but let us rst establish that these lines exist.
Theorem 4.5.4. The tree T (1 ) ordered lexicographically is a homogeneous sta-
tionary C-line.
Proof. First of all recall that

T (1 ) = {t : : < 1 , t = 1 }

ordered lexicographically is a dense subordering of the homogeneous continuum


 1 ) and that the proof of Lemma 2.2.9 actually shows that T (1 ) is homoge-
A(
neous as well. In order to show that T (1 ) is a C-line, it suces to decompose
the set
W = {(s, t) (T (1 ))2 : ht(s) ht(t)}
into countably many chains. To a given pair (s, t) of W with = ht(s) = ht(t),
we associate a hereditarily nite set that codes the behavior of the restrictions of
s and t on the nite set

Fst = { : s() = t()}.

Then working as in the proof of Lemma 2.2.4, we show that this parametrization
with hereditarily nite sets will give us a countable chain decomposition of W .
Thus, T (1 ) ordered lexicographically is a homogeneous C-line. To see that it is
stationary, note that sets of the form

T (1 )  = {t : : < , t = 1 },

for limit ordinals < 1 constitute a proper decomposition of T (1 ). This de-


composition is (left) continuous at every such , since a typical complementary
interval to T (1 )  has the form It = {x T (1 ) : t x} for some t T (1 )
of length and since clearly t = min It . 
20 This notion should be more accurately called -continuous from the left since one may also

dene its right analogue by requiring that every convex set from TA has a maximal element.
21 Or, more precisely, left-stationary
130 Chapter 4. Coherent Mappings and Trees

Recall that E. Borels notion of strong measure zero is a metric notion that
has several closely related topological notions due to K. Menger, F. Rothberger
and others. For example, Rothbergers property C  is a purely topological notion
which in the class of metric spaces is slightly stronger than the property C, i.e.,
being of strong measure zero. It says that for every sequence {Un } of 
open covers

of a space X one can choose Un Un for each n such that X = n=0 Un . A
considerably stronger covering property of a topological space is obtained via the
following notion.
Denition 4.5.5. An -cover U of a set X is a family of subsets of X with the
property that for every nite subset F of X there is U U such that F U .
Theorem 4.5.6. For every stationary C-line C and every open -cover U of C
there is a sequence {Un } U such that C = lim inf Un .22
Proof. Let {C } be a xed proper decomposition of C and let T be the correspond-
ing partition tree of complementary intervals to the sets C . Choose a countable
elementary submodel M of some large enough structure containing all the relevant
objects such that, if is its intersection with 1 , then {C } is continuous at .
Let {a2n }
n=1 be an enumeration of all elements of C and let {a2n1 }n=1 be an
enumeration of all minimums of the complementary intervals of C . The theorem
will be proved once we show that for every n there is an open set Un in U M
which includes

n 
{a2 , a4 , . . . , a2n } t2i1 ,
i=1

where t2i1 is the complementary interval of C whose minimum is equal to


a2i1 . Applying the fact that U is a 2n-cover of C, choose U in U containing
{a1 , a2 , . . . , a2n } and for each 1 i n choose xi < a2i1 in C (= C M ) such
that (xi , a2i1 ] U . Let Y be the set of all (y1 , . . . , yn ) in C n such that xi < yi
(1 i n) and such that

n 
{a2 , a4 , . . . , a2n } (xi , yi ]
i=1

is covered by a single element of U. Then Y M and (a1 , a3 , . . . , a2n1 ) Y .


Since C n is the union of countably many chains, there is a chain Y0 Y such that
Y0 M and (a1 , a3 , . . . , a2n1 ) Y0 .
Claim. There is (y1 , . . . , yn ) in Y0 M such that a2i1 < yi for all 1 i n.
Note that to prove the claim, it suces to nd such (y1 , . . . , yn ) with the property
that a1 < y1 because the other inequalities will follow from this one and the
fact that Y0 is a chain of C n . So, let c be the supremum (taken in the Dedekind
completion of C) of the set of all rst coordinates of elements of Y0 . Note that
22 Recall
 
that lim inf Un = m n>m Un .
4.5. Stationary C-lines 131

since C contains no uncountable well-ordered subsets, this supremum is equal to


the supremum of the projection of Y0 M on the rst coordinate. Note also that
c M and that a1 c. Since a1 is not a member of M , the equality is impossible,
so we have a1 < c. Since c = sup 1 (Y0 M ), there is (y1 , . . . , yn ) in Y0 M such
that a1 < y1 and this proves the claim.
Choose (y1 , . . . , yn ) as in the claim. The proof of the theorem is nished,
once we show that t2i1 is included in (xi , yi ] for all 1 i n or equivalently that
sup t2i1 yi for all 1 i n. But this is clearly so since ai = min t2i1 < yi and
since the point yi , being a member of C , cannot split the complementary interval
t2i1 . This completes the proof. 
Corollary 4.5.7. Suppose F is a family of continuous real-valued functions dened
on some stationary C-line C which pointwise accumulates to some continuous real-
valued function f on C. Then there is a sequence (fn ) F such that lim fn = f .
Proof. Clearly, we may assume that f (x) = 0 for all x C. Fix a sequence (xn )
of pairwise distinct elements of C. Let U be the set of all subsets of C of the form

U (f, n) = {x X : x = xn and |f (x)| < 2n }

where f F and n N. By our assumption about F and f , the family U is an


open -cover of C. By Theorem 4.5.6, there exist {U (fi , ni )} U such that

lim inf U (fi , ni ) = C.

It is easily checked that ni which in turn gives fi f . This completes the


proof. 
Denition 4.5.8. The left arrow topology on a given C-line C is the one deter-
mined by intervals of the form (x, y] for some x < y in C. Similarly, one denes
the right arrow topology generated by intervals of the form [x, y).
It should be clear that the proof of Theorem 4.5.6 also establishes the fol-
lowing slightly stronger conclusion.
Theorem 4.5.9. For every stationary C-line C and every -cover U of C consisting
of sets that are open relative to the left arrow topology of C there is a sequence
{Un } U such that C = lim inf Un . 
Corollary 4.5.10. Suppose F is a family of continuous real-valued functions dened
on the arrow space (C, ), where C is a stationary C-line. Suppose further that
some continuous real-valued function f on C is a pointwise accumulation point of
F . Then there is a sequence (fn ) F such that lim fn = f . 
At this point one may ask for the existence of a C-line that has proper
decompositions that are stationary both from the left and from the right so that
both of its arrow topologies and enjoy the strong covering property that
involves the notion of an -cover. For this, it suces to constrict a C-line A with
132 Chapter 4. Coherent Mappings and Trees

proper decomposition {A } such that for a stationary set of s the complementary


intervals to A have minimums as well as maximums. This could be done, for
example, by considering the following extension of T (1 ),

T (1 )+ = T (1 ) {t 1 : t T (1 )},

where for t T (1 ) of length , by t 1 , we denote the sequence s of length


+ extending t and having the property that s( + n) = 1 for all n < . It is
clear that the proof of Theorem 4.5.4 shows that T (1 )+ is indeed a C-line and
that its natural proper decomposition is continuous from both sides, or in other
words, that for the complementary intervals

It+ = {x T (1 )+ : t x},

we have that t = min It+ and t 1 = max It+ . This proves the following result.
Theorem 4.5.11. There is a C-line C with the property that, if U is an -cover
of C consisting of sets that are all open relative to the left arrow topology of C
or sets that are all open relative to the right arrow topology on C, then there is a
sequence (Un ) U such that C = lim inf Un . 
Chapter 5

The Square-bracket Operation


on Countable Ordinals

5.1 The upper trace and the square-bracket operation


Recall that a walk from a countable ordinal to a smaller ordinal along the
xed C-sequence C ( < 1 ) is a nite decreasing sequence

= 0 > 1 > > n = ,

where i+1 = min(Ci \ ) for all i < n. Recall also the notion of the upper trace
of the minimal walk,
Tr(, ) = {0 , 1 , . . . , n },
the nite set of places visited in the minimal walk from to . The following
simple fact about the upper trace lies at the heart of all known denitions of
square-bracket operations, not only on 1 but also at higher cardinalities.
Lemma 5.1.1. For every uncountable subset of 1 , the union of Tr(, ) for
< in contains a closed and unbounded subset of 1 .

Proof. It suces to show that the union of traces contains every countable limit
ordinal such that sup( ) = . Pick an arbitrary \ and let

= 0 > 1 > > k =

be the minimal walk from to . Let < be an upper bound of all sets of the
form Ci for i < k. By the choice of there is above . Then the
minimal walk from to starts as 0 > 1 > > k , so in particular belongs
to Tr(, ). 
134 Chapter 5. The Square-bracket Operation on Countable Ordinals

We shall now see that it is possible to pick a single place [] in Tr(, )


so that Lemma 5.1.1 remains valid with [] in place of Tr(, ). Recall that by
Lemma 2.1.18,
(, ) = min{ : 0 (, ) = 0 (, )}

is a successor ordinal. We shall be interested in its predecessor,


Denition 5.1.2. 0 (, ) = (, ) 1.

Thus, if = 0 (, ), then 0 (, ) = 0 (, ) and so there is a natural


isomorphism between Tr(, ) and Tr(, ). We shall dene [] by comparing the
three sets Tr(, ), Tr(, ) and Tr(, ).

Figure 5.1: The square-bracket operation.


5.1. The upper trace and the square-bracket operation 135

Denition 5.1.3. The square-bracket operation on 1 is dened as follows (see


Figure 5.1):
[] = min(Tr(, ) Tr(0 (, ), ))
= min(Tr(0 (, ), ) \ ).
Next, recall the function 0 : [1 ]2 < which codes the walk along the xed
C-sequence C ( < 1 ) and which is given by the recursive formula
0 (, ) = |C | 0 (, min(C \ )),
with the boundary value 0 (, ) = . Recall also the denition of the correspond-
ing tree
T (0 ) = {0 (, )  : < 1 }.
As before, for < 1 , we shorten the notation 0 (, ) to (0 ) of the corresponding
ber-mapping  < dened by (0 ) () = 0 (, ).
Lemma 5.1.4. For every uncountable subset of 1 , the set of all ordinals of the
form [] for some < in contains a closed and unbounded subset of 1 .
Proof. For t T (0 ) let
t = { : (0 ) end-extends t}.
Let S be the collection of all t T (0 ) for which t is uncountable. Clearly, S is
a downward-closed uncountable subtree of T (0 ). The lemma is proved once we
prove that every countable limit ordinal > 0 with the following two properties
can be represented as [] for some < in :
sup(t ) = for every t S of length < , (5.1.1)
every t S of length < has two incomparable successors
(5.1.2)
in S both of length < .
Fix such a and choose \ such that (0 )  S and consider the
minimal walk from to :
= 0 > 1 > > k = .
Let < be an upper bound of all sets of the form Ci for i < k. Since
the restriction t = (0 )  belongs to S, by (5.1.2) we can nd one of its end-
extensions s in S which is incomparable with (0 ) . It follows that for s , the
ordinal 0 (, ) has the xed value
= min{ < |s| : s() = 0 (, )} 1.
Note that , so the walk is a common initial part of walks and
for every s . Hence if we choose s above min(C \ ) (which
we can by (5.1.1)), we get that the walks and never meet after . In
other words for any such , the ordinal is the minimum of Tr(, ) Tr(, ),
or equivalently is the minimum of Tr(, ) \ . 
136 Chapter 5. The Square-bracket Operation on Countable Ordinals

It should be clear that the above argument can easily be adjusted to give us
the following slightly more general fact about the square-bracket operation.
Lemma 5.1.5. For every uncountable family A of pairwise-disjoint nite subsets of
1 , all of the same size n, the set of all ordinals of the form
1
= [a(1)b(1)] = [a(2)b(2)] = = [a(n)b(n)]

for some a = b in A contains a closed and unbounded subset of 1 . 


It turns out that the square-bracket operation can be used in constructions
of various mathematical objects of complex behavior where all known previous
constructions needed the Continuum Hypothesis or stronger enumeration prin-
ciples. The usefulness of [] in these constructions is based on the fact that []
reduces the quantication over uncountable subsets of 1 to the quantication
over closed unbounded subsets of 1 . For example composing [] with a unary
operation : 1 1 which takes each of the values stationary many times, one
gets the following fact about the mapping c(, ) = [] .
Theorem 5.1.6. There is a mapping c : [1 ]2 1 which takes all the values
from 1 on any square []2 of some uncountable subset of 1 . 
This result gives a denitive limitation on any form of a Ramsey Theorem
for the uncountable that tries to say something about the behaviors of restrictions
of colorings on an uncountable square. Recall that in Section 2.3.1 we have seen
that there are mappings of this sort with complex behavior on sets of the form

0 1 = {{0 , 1 } : 0 0 , 1 1 , 0 < 1 },

where 0 and 1 are two uncountable subsets of 1 . It turns out that the original
square-bracket operation has a simple behavior on 0 1 for every pair 0 and
1 of suciently thin uncountable subsets of 1 , a property that turns out to be
quite useful in applications. The property is made precise as follows.
Lemma 5.1.7. For every uncountable family A of pairwise-disjoint subsets of 1 ,
all of the same size n, there exist an equivalence relation on n = {0, 1, . . . , n1},
an uncountable B A, and a mapping hb : n n 1 for each b B such that
for all a < b 2 in B, we have [a(i)b(j)] = hb (i, j) whenever i  j. Moreover, for
every uncountable C B, the set of ordinals for which there exist a < b in C
such that [a(p)b(q)] = for all i, j < n with i j contains a closed and unbounded
subset of 1 .
Proof. Identifying an ordinal with (0 ) , we can think of A as a family of
n-tuples of nodes of the tree T (0 ). For a limit ordinal < 1 , choose a in A such
1 For a nite set x of ordinals of size n we use the notation x(1), x(2), . . ., x(n) or x(0), x(1), . . .,

x(n 1), depending on the context, for the enumeration of x according to the natural ordering
on the ordinals.
2 For a pair a and b of sets of ordinals, a < b denotes the fact that < for all a and b.
5.1. The upper trace and the square-bracket operation 137

that every node from a has height . Let c be the projection of a on the th
level of T (0 ). Let h() < be such that the projection of c on the h()th level
has size equal to the size of c , i.e.,

(s, t) < h() for all s = t in c . (5.1.3)

Find a stationary set 1 such that h takes a constant value 0 on and


that the 0 th projection takes the constant value on c ( ). Fix < in
and consider a and a that project to dierent elements of the 0 th
level of T (0 ). Then (, ) is independent of and or and as it is equal
to ((0 )  0 , (0 )  0 ). So the trace Tr(0 (, ), ) depends only on and
the projection (0 )  0 . Thus, going to a subset of , we may assume that its
size depends only on the position of in some a , and that the minimal point
of Tr(0 (, ), ) \ is the same for all in some a , for < , having a xed
projection on the 0 th level of T (0 ). This is the content of the rst part of the
conclusion of Lemma 5.1.7. To see the second part of the conclusion, we shrink
even further to assure that the union of c ( ) is an antichain of T (0 ). Now
note that for two typical < in , [] = [  ] for all ,  a and ,  a
with the property that (0 )  0 = (0 )  0 = (0 )  0 = (0 )   0 . So the
rest of the proof reduces to the proof of Lemma 5.1.5 above. 

Remark 5.1.8. To see the point of Lemma 5.1.7, consider the projection [] of
the square-bracket operation generated by a mapping from 1 onto rather than
1 . Then the sequence hb : n n (b B) of corresponding compositions will
have the same constant value h : n n modulo, of course, shrinking B to
some uncountable subset. Noting that the proof of Lemma 5.1.7 applied to a typical
A will actually result in the equality = as the equivalence relation , we conclude
that if and are coming from two dierent positions of some a and a in B,
then [] will depend on the positions themselves rather than and . It fact,
there is a natural variation of the square-bracket operation (see the mapping b of
[109, p. 287]), where, in Lemma 5.1.7, we always have the nest possible partition
Ii = {i} (i = 1, . . . , n) for every uncountable family A of pairwise-disjoint n-tuples
of countable ordinals. We shall reproduce this variation in Section 5.4 below, where
we shall see also one of its applications to the quadratic form theory.
Note that the basic C-sequence C ( < 1 ) which we have xed at the
beginning of this chapter can be used to actually dene a unary operation :
1 1 which takes each of the ordinals from 1 stationarily many times. So
the projection [] can actually be dened in our basic structure (1 , , C).
 We
are now at the point to see that our basic structure is actually rigid.
Lemma 5.1.9. The algebraic structure (1 , [], ) has no nontrivial automorphisms.

Proof. Let h be a given automorphism of (1 , [], ). If the set of xed points


of h is uncountable, h must be the identity map. To see this, consider a < 1 .
138 Chapter 5. The Square-bracket Operation on Countable Ordinals

By the property of the map c(, ) = [] stated in Theorem 5.1.6, there exist
< in such that [] = . Applying h to this equation we get
h() = h([] ) = (h([])) = [h()h()] = [] = .
It follows that = { < 1 : h() = } is in particular uncountable. Shrinking
and replacing h by h1 , if necessary, we may safely assume that h() > for all
. Consider a < 1 and let S be the set of all < 1 such that = .
By our choice of the set S is stationary. By Lemma 5.1.5 applied to the family
A = {{, h()} : }, we can nd < in such that [] = [h()h()]
belongs to S , or in other words,
[] = [h()h()] = .
Since [h()h()] = h([] ) we conclude that h() = . Since was an arbitrary
countable ordinal, this shows that h is the identity map. 
We give now an application of this rigidity result to a problem in Model
Theory about the quantier Qx = there exist uncountably many x and its higher-
dimensional analogues Qn x1 xn = there exist an uncountable n-cube many
x1 , , xn . By a result of Ebbinghaus and Flum [26] (see also [82]) every model
of every sentence of L(Q) has nontrivial automorphisms. However we shall now
see that this is no longer true about the quantier Q2 , for example.
Example 5.1.10. The sentence will talk about one unary relation N , one binary
relation < and two binary functional symbols C and E. It is the conjunction of
the following seven sentences:
(1 ) Qx x = x.
(2 ) Qx N (x).
(3 ) < is a total ordering.
(4 ) E is a symmetric binary operation.
(5 ) x < y N (E(x, y)).
(6 ) x < y < z E(x, z) = E(y, z).
(7 ) xn{N (n) Q2 uv[u < uv  < v(u = v  E(u , u) = E(v  , v) = n)
u < u v  < v(E(u , u) = E(v  , v) = n (C(u , v  ) = x C(u, v) = x))]}.
The model of that we have in mind is the model (1 , , <, c, e) where c(, ) =
[] and e : [1 ]2 is any mapping such that e(, ) = e(, ) whenever
< < (e.g., we can take e = 1 or e = ). The sentence 7 simply says that
for every < 1 and every uncountable family A of pairwise-disjoint unordered
pairs of countable ordinals, there exist a = b in A such that
c(min a, min b) = c(max a, max b) = .
This is a consequence of Lemma 5.1.5 and the fact that S = { : = } is a
stationary subset of 1 . These are the properties of [] and which we have used
5.2. Projecting the square-bracket operation 139

in the proof of Lemma 5.1.9 in order to prove that (1 , [], ) is a rigid structure.
So a quite analogous proof will show that any model (M, N, <, C, E) of must be
rigid. 
The crucial property of [] stated in Lemma 5.1.5 can also be used to provide
a negative answer to the basis problem for uncountable graphs by constructing a
large family of pairwise orthogonal uncountable graphs.
Denition 5.1.11. For a subset of 1 , let G be the graph whose vertex-set is 1
and whose edge-set is equal to

{{, } : [] }.

Lemma 5.1.12. If the symmetric dierence between and is a stationary subset


of 1 , then the corresponding graphs G and G are orthogonal to each other, i.e.,
they do not contain uncountable isomorphic subgraphs. 
We have seen above that comparing [] with a map : 1 I where I
is some set of mathematical objects/requirements in such a way that each ob-
ject/requirement is given a stationary preimage, gives us a way to meet each of
these objects/requirements in the square of any uncountable subset of 1 . This
observation is the basis of all known applications of the square-bracket operation.
A careful choice of I and : 1 I gives us a projection of the square-bracket
operation that can be quite useful. This is exposed in the next section.

5.2 Projecting the square-bracket operation


Denition 5.2.1. Let G be the collection of all maps g : 2m 2m 1 where m is
a positive integer denoted by m(g). Choose a mapping : 1 G which takes
each value from G stationarily many times. Choose also a one-to-one sequence
r ( < 1 ) of elements of the Cantor set 2 . Note that both these objects
can actually be dened in our basic structure (1 , , C).  Consider the projection
[[]] = [[]]G of the square-bracket operation [] dened as follows:

[[]] = ([])(r  m(([])), r  m(([]))).

Then the property of [] stated in Lemma 5.1.5 corresponds to the following


property of the projection [[]].
Lemma 5.2.2. For every uncountable family A of pairwise-disjoint nite subsets of
1 , all of the same size n, and for every n-sequence 0 , 1 , . . . , n1 of countable
ordinals there exist a and b in A such that [[a(i)b(i)]] = i for i = 1, . . . , n.

Proof. Find uncountable B A and m < such that for all b B,

r  m = r  m for all = in b.
140 Chapter 5. The Square-bracket Operation on Countable Ordinals

Moreover, we may assume that for some xed one-to-one sequence t0 , t1 , . . . , tn1
of elements of 2m ,

rb(i)  m = ti for all i < n and all b B.

Find g : 2m 2m 1 such that g(ti , ti ) = for all i < n. By our choice of , the
set = { < 1 : () = g} is stationary, so by Lemma 5.1.5, we can nd a < b
in B such that for some ,

[a(0)b(0)] = [a(1)b(1)] = = [a(n 1)b(n 1)] = .

It follows that for all i < n,

[[a(i)b(i)]] = g(ra(i)  m, rb(i)  m) = g(ti , ti ) = ,

as required. 

The following lemma describes a universality property of the projection [[]] =


[[]]G of the square-bracket operation [].
Lemma 5.2.3. For every positive integer n, every uncountable subset of 1 and
every symmetric n n-matrix M of countable ordinals there is a one-to-one map-
ping : n such that [[(i)(j)]] = M (i, j) for i, j < n.

Proof. Let be an uncountable subset of 1 and let M be a given n n matrix


on countable ordinals. Then for each we can nd t in the th level of the tree
T (0 ) and a nite set F \ of size n such that t (0 ) for all F .
Let n be the minimal integer such that r  n = r  n whenever = are
members of F . For a limit ordinal , let () < be an upper bound of all sets
of the form Ci where F , i < k < and = 0 > > k = is
the minimal walk from to along the C-sequence xed at the beginning of this
chapter. Then there is a stationary set 1 , m < , s0 , . . . , sn1 2m and
< 1 such that:
n = m and = for all , (5.2.1)
t ( ) form an antichain of T (0 ), (5.2.2)
if occupies the ith place in F for some , then
(5.2.3)
r  m = si .
Let h : 2m 2m 1 be an arbitrary map subject to the requirement that
h : (si , sj ) = M (i, j) for all i, j < n. Let h = { < 1 : () = h}. By the choice
of the set h is stationary in 1 . By the proof of the basic property Lemma
5.1.4 of the square-bracket operation we can nd h , and \
such that t  = t  , F and [] = . We claim that

[] = for all F and F . (5.2.4)


5.2. Projecting the square-bracket operation 141

To see this, rst note that 0 (, ) = 0 (, ) = for all F and F .


By the choice of and the fact that , the walk from some F to goes
through , so for any F and F we have

[] = min(Tr(, ) \ ) = min(Tr(, ) \ ) = [].

In fact the proof of Lemma 5.1.4 shows that we can nd 0 h , 0


and an uncountable set 1 \ 0 such that F0 0 , t0  = t  and
[0 ] = 0 for all 1 . It follows that

[] = 0 for all F0 , F and 1 . (5.2.5)

So we can repeat the procedure for 1 in place of and get 1 h , 1 1 0


and uncountable 2 1 \ 1 such that [1 ] = 1 for all 2 , and so on.
Repeating this n times we get a sequence 0 , . . . , n1 such that for all i < j < n,

([]) = h for all Fi and F with j . (5.2.6)

Dene : n 1 by letting (i) be the ith member of Fi . Going back to the


denition of the projection [[]] of [] one sees that indeed [[(i)(j)]] = M (i, j)
holds for all i < j < n. 
Projections of the square-bracket operation that have only countably many
values are also worth considering. In fact, in later sections of this book, we shall
give applications of this kind of projections in cases where the projections with
uncountably many values do not seem to t.
Denition 5.2.4. Let G0 be the collection of all maps g : 2n 2n where n is
a positive integer denoted by n(g). Choose a mapping : 1 G0 which takes
each value from G0 stationarily many times. Choose also a one-to-one sequence
r ( < 1 ) of elements of the Cantor set 2 . Consider the projection

[[]]G0 : [1 ]2

of the square-bracket operation [] dened as follows:

[[]]G0 = ([])(r  m(([])), r  m(([]))).

Theorem 5.2.5. For every uncountable family A of pairwise-disjoint nite subsets


of 1 , all of the same size n, there is uncountable set B A, an equivalence
relation on n = {0, 1, . . . , n 1}, and a single mapping h : n n such that

[[a(i)b(j)]]G0 = h(i, j) for all a < b in B and i  j.

Moreover, for every uncountable C B and for every g : n n there exist


a < b in C such that

[[a(i)b(j)]]G0 = g(i, j) for all i, j < n with i j.


142 Chapter 5. The Square-bracket Operation on Countable Ordinals

Proof. Take uncountable B A satisfying the conclusion of Lemma 5.1.7. Since


there exist only countably many mappings of the form

hb : n n G0 ,

for b B, shrinking B, we may assume that for some h : n n G0 and all b B,


we have hb = h. Choose an integer m such that for some uncountable B  B and
some one-to-one sequence ti (i < n) of elements of 2m , we have

(a) m > m(h(i, j)) for all i, j < n,


(b) rb(i)  m = ti for all b B  and i < n.
Dene h : 2m 2m , by

h (i, j)(s, t) = h(i, j)(t  m(i, j), s  m(i, j)).

To check that this subset B  of A and the mapping h satisfy the conclusion of the
lemma, consider a < b in B  and i, j < n such that i  j. Then by the conclusion
of Lemma 5.1.7, we know that [a(i)b(j)] = hb (i, j, ) = for some countable ordinal
. Let g = (). Then,

[[a(i)b(j)]]G0 = g(ti  m(i, j), tj  m(i, j)) = h (i, j).

Consider now an arbitrary g : n n and an arbitrary uncountable set


C B  . Choose a mapping f : 2m 2m such that f (ti , tj ) = g(i, j) for
i, j < n. Let f = { < 1 : () = f }. Then by our choice of , the set f
is stationary, so by the conclusion of Lemma 5.1.7, we can nd a < b in C and
such that [a(i)b(j)] = for all i, j < n such that i j. Going back to the
denition of [[]]G0 , we conclude that

[[a(i)b(j)]]G0 = ()(ra(i)  m(()), rb(j)  m(())) = f (ti , tj ) = g(i, j)

for all i, j < n such that i j. 

For a future use let us state and prove yet another application of the square-
bracket operation [].
Lemma 5.2.6. There is a mapping f : 1< 1 such that, for every sequence
( ) of elements of 1< indexed by some uncountable subset of 1 such that
rang( ) for all , and for every < 1 , there exist < in such
that f (  ) = whenever is a subsequence of with rang() and is a
subsequence of with rang().

Proof. Choose a one-to-one sequence r ( < 1 ) of elements of 2 . For 1< ,


set
( ) = max{(r (i) , r (j) ) : i < j < | |, (i) = (j)},
5.2. Projecting the square-bracket operation 143

where we stipulate that ( ) = whenever is a constant sequence. Set f ( ) = 0


for a constant sequence; otherwise, set

f ( ) = [[ (i) (j)]] for (i, j) = min{(p, q) : (r (p) , r (q) ) = ( )}. (5.2.7)

Note that the conclusion of the lemma in case the sequences for are all
constant reduces to the case n = 1 of Lemma 5.2.2, so we may as well assume that
no ( ) is a constant sequence. Going to an uncountable subset of , we
may assume that for some integers l and m,

| | = l and ( ) = m for all ,

and that rang( ) forms an increasing -system with root c. Moreover, going
to an uncountable subset of , we may also assume that the obviously dened nite
structures

M = (rang( ), <, c, {}, , {r (i)  (m + 1) : i < l}, . . . ) ( )

are all isomorphic. Applying Lemma 5.2.2, for a given < 1 , we get < in
such that
[[ (k) (k)]] = for all k < l with (k), (k)
/ c. (5.2.8)

Let us check that these and satisfy the conclusion of the lemma. So let be a
subsequence of with rang() and be a subsequence of with rang().
Let =  . Note that

() (r(i0 ) , r(j0 ) ) > m, (5.2.9)

where i0 and j0 are such that (i0 ) = and (j0 ) = . Let

(i, j) = max{(p, q) : (r(p) , r(q) ) = ()}.

Then by (5.2.9), the ordinals (i) and (j) must not belong to the same set rang( )
or rang( ), and moreover, they must occupy the same positions in the increasing
enumerations of the corresponding sets rang( ) and rang( ) to which they be-
long. Since they are dierent and since the isomorphism between the structures
M and M is identity on the root c, none of these two ordinals belongs to c.
Applying again the fact that M and M are isomorphic, we conclude that there
must be an index k < l such that (k) = (i) and (k) = (j), or vice versa,
(k) = (j) and (k) = (i). Applying (5.2.8), it follows that

f (  ) = [[(i)(j)]] = [[ (k) (k)]] = ,

as required. 
144 Chapter 5. The Square-bracket Operation on Countable Ordinals

5.3 Some geometrical applications of the


square-bracket operation
The purpose of this section is to show that the projections [[]]G and [[]]G0 of the
square-bracket operation [] could be used in constructing interesting examples
in the geometry of Banach spaces as well as in constructing some interesting
geometrical congurations in nite-dimensional Euclidian spaces.
Theorem 5.3.1. For every integer n 2 there are countably many collections Hk
(k < ) of hyperplanes3 of Rn such that:
(1) any n hyperplanes belonging to dierent Hk s can have at most one point in
common, and

(2) for every decomposition R = k< Pk there is k < and hyperplane H Hk
which is spanned by n points of the intersection H Pk .
Proof. Choose a set Rn of cardinality 1 such that any n distinct points
x1 , . . . , xn of are in general position, in the sense that they span a unique hy-
perplane H(x1 , . . . , xn ) of Rn . Moreover, we assume that the set is chosen in
such a way that if X1 , . . . , Xn is a sequence of n-element subsets of with the
property that
|Xi Xj | 1 whenever i = j,
then
|H(X1 ) H(X2 ) H(Xn )| 1.
We may assume that the projection [[]]G0 of the square-bracket operation [] has
domain []2 rather than [1 ]2 . We have seen in the previous section that this
projection of the square-bracket operation is quite rich in properties, but in this
proof, we shall only use the property of [[]]G0 saying that for every uncountable
subset of and every k < there exist x and y in such that [[xy]]G0 = k.
So for this particular proof even a simple projection of [] would work, such as
for example, the composition of [] with a partition of 1 into countably many
pairwise-disjoint stationary sets.
For an n-element subset X of , let H(X) denote the hyperplane of Rn
spanned by X if indeed there is a unique hyperplane of Rn that contains the
set X; otherwise H(X) is left undened. Let H(Rn ) denote the collection of all
hyperplanes of Rn , and for k < , set

Hk = {H(X) H(Rn ) : X []n and [[xy]]G0 = k for all x = y in X}.

Let us check that the sequence Hk (k < ) has the properties (1) and (2). The
fact that n hyperplanes coming from distinct Hk s can have at most one point
3A hyperplane in Rn is an arbitrary translate of an (n 1)-dimensional subspace of Rn . A set S
of points from R span a hyperplane H if H is the only hyperplane that contains all the points
of S.
5.3. Some geometrical applications of the square-bracket operation 145

in common is an immediate consequence of the property of the set , since if X


and Y are n-element subsets of such that the square-bracket operation takes
two dierent constant
 values on [X]2 and [Y ]2 , then |X Y | 1. Consider now
a partition R = k< Pk . Fix a k < such that = Pk is uncountable. By
the property of the square-bracket operation and the DushnikMiller theorem [25]
there must exist an n-element set X such that [[xy]]G0 = k for all x = y in
X. Then H(X) belongs to Hk and it is spanned by the set X H(X) Pk . This
nishes the proof. 
Remark 5.3.2. Theorem 5.3.1 was discovered by Erd os, Jackson and Mauldin [29]
where we refer the reader for further information.
Recall that a homogeneous polynomial of degree n on a (typically complex)
Banach space X is a mapping P : X C of the form

P (x) = (x, x, . . . , x),

where : X n C is a bounded symmetric multilinear form on X. Recall that


a multilinear form : X n C is bounded if there is a constant C such that

|(x0 , . . . , xn1 )| C
x0

xn1
for all (x0 , . . . , xn1 ) X n .

An arbitrary polynomial on X of degree n is the sum of the form

P (x) = P0 (x) + P1 (x)+, . . . , Pn (x),

where for k n the Pk (x) is a homogeneous polynomial of degree k and where


Pn (x) is not identically zero. Describing the set ker(P ) of zeros of a polynomial is
an old subject (see, e.g., [19]). For example, it is known that an arbitrary polyno-
mial P on a complex innite-dimensional Banach space X such that P (0) = 0 is
equal to zero on an innite-dimensional subspace of X (see, [83]). The following
example shows that this result does not extend to the other dimensions.
Theorem 5.3.3. There is a homogeneous polynomial P : 1 (1 ) C of degree
2 dened on the complex version of 1 (1 ) which maps every closed nonseparable
subspace of 1 (1 ) onto C.
Proof. We shall use the projection [[]]G0 and its properties listed in Theorem 5.2.5
in order to dene the symmetric bilinear form on which we base our polynomial
P . Recall that [[]]G0 has range and by indexing the set D of rational points of
the unit disc of C by , we shall actually assume that the range of [[]]G0 is D
itself rather than . It will also be convenient to extend [[]]G0 on the diagonal by
letting [[]]G0 = 0 for all < 1 . Let B = {1 : < 1 }, the Hamel basis of the
set c00 (1 ) of all nitely supported maps from 1 into C. The symmetric bilinear
form is the unique extension of its restriction on B B dened as

(1 , 1 ) = (1 , 1 ) = [[]]G0 (5.3.1)
146 Chapter 5. The Square-bracket Operation on Countable Ordinals

for < 1 . Note that |[[]]G0 | 1 for all and . It follows that,


  

|(x, y)| = x()y()[[]]G0 |x()| |y()| =
x
1
y
1 (5.3.2)
,

for all x and y in 1 (1 ). So is a bounded bilinear form on 1 (1 ) and so


P (x) = (x, x) is a homogeneous polynomial on 1 (1 ) of degree 2. We shall show
that P maps any non-separable closed subspace of 1 (1 ) onto C. Consider a closed
and non-separable subspace X of 1 (1 ). A simple consequence of the fundamental
theorem of algebra is that every nonzero complex polynomial Q dened on some
Banach space over the eld C must map the Banach space onto C. It therefore
suces to nd w in X such that P (w) = 0. Since X is not separable, we can
choose a rational 0 <  1, an uncountable subset Y from the unit ball of X, and
for each x in Y an ordinal x supp(x) such that:
(1) x = y for x = y in Y ,
(2) |x(x )|  for all x in Y .
For each x in Y choose a nite set bx supp(x) such that
2
/ x |x()| <  /32.
b (5.3.3)
Going to an uncountable subset of Y , we may assume that the family bx (x Y )
forms a -system with root b. Let us rst consider the case b = . Thus, we
may assume that the family bx (x Y ) is a family of pairwise-disjoint n-element
subsets of 1 for some xed integer n. Going to an uncountable subfamily of Y ,
we may assume that for all x and y in Y ,
(3) [[bx (i)bx (j)]]G0 = [[by (i)by (j)]]G0 for all i, j < n,
(4) |x(bx (i)) y(by (i))| < 2 /16n for all i < n.
Combining this and (5.3.3), we obtain
|P (x) P (y)| < 2 /4 for all x, y Y. (5.3.4)
Applying Theorem 5.2.5, we nd an uncountable set Z Y , an equivalence rela-
tion on n = {0, 1, . . . , n 1}, and h : n n D such that for all x and y in Z
with bx < by , we have
[[bx (i)by (j)]]G0 = h(i, j) for all i, j < n such that i  j. (5.3.5)
Rening Z, we may assume that for some xed integer i0 < n and all x in Z, we
have x = bx (i0 ). Let g : n n D be a mapping which agrees with h on pairs
(i, j) with property i  j such that g(i, j) = 0 for all pairs (i, j) such that i j.
By (5.3.5) and the second conclusion of Theorem 5.2.5, there exist x, y Z such
that bx < by , and
[[bx (i)by (j)]]G0 = g(i, j) for all i, j < n. (5.3.6)
5.3. Some geometrical applications of the square-bracket operation 147

Consider now the mapping g  : n n D which agrees with h on pairs (i, j) with
property i  j such that g  (i, j) = 0 for all pairs (i, j) with property i j with
the exception of the pair (i0 , i0 ) when we put g  (i0 , i0 ) = 1. By (5.3.5) and the
second conclusion of Theorem 5.2.5, there exist x , y  Z such that bx < by , and

[[bx (i)by (j)]]G0 = g  (i, j) for all i, j < n. (5.3.7)

We claim that P (x + y) = P (x + y  ). Otherwise, we would have the equation


P (x) + P (y) + 2(x, y) = P (x ) + P (y  ) + 2(x , y  ). Hence, by (5.3.4),

|(x, y) (x , y  )| < 1/2(2 /4 + 2 /4) = 2 /4. (5.3.8)

For w {x, y, x , y  }, let

w0 = w  (supp(w) \ bw ), w1 = w  (bw \ {bw (i0 )}), and w2 = w  {bw (i0 )}.

Thus, w = w0 + w1 + w2 for every w {x, y, x , y  }, and so, (x, y) and (x , y  ),


have the nine-term expansions

i,j<3 (xi , yj ) and i,j<3 (xi , yj ),

respectively. Note that by the setup above, |(xi , yj ) (xi , yj )| < /16 for all
i, j < 3 such that (i, j) = (2, 2). On the other hand, by (2), (5.3.6) and (5.3.7),

|(x2 , y2 ) (x2 , y2 )| 2 .

Combining all this with (5.3.8), we get that

2 /4 > |(x, y) (x , y  )| 2 8(2 /16),

a contradiction. It follows that one of P (x + y) or P (x + y  ) must be nonzero and


this nishes the proof.
To reduce the general case to the case when the root b of the -system is
empty, we start with an uncountable set Y of vectors of X of norm 1/2 satisfying
(1) and (2) for some  > 0. Find nite sets ax (x Y ) such that
2
/ x |x()| <  /128.
a

Assume that the family forms a -system with root a and such that for all x, y Y .

a |x() y()| < 2 /128. (5.3.9)

For x Z, set bx = ax \ a. Form an uncountable set Z of elements of the form


z = y x for x, y Y such that bx < by . We also assume that if z = y x
and z  = y  x are members of Z, then bx < by < bx < by , or vice versa
bx < by < bx < by . Note that Z still satises (1) and (2) if we put z = y
148 Chapter 5. The Square-bracket Operation on Countable Ordinals

for z = y x Z. Note also that bz = bx by (z = y x Z) is a family of


pairwise-disjoint nite subsets of 1 such that
2
/ z |z()| <  /32.
b

So we have reached the input of the rst part of the proof, i.e., we have reduced
the general case to the case when the root of the -system is empty. 

The reader is refereed to [6] for more information on this and other related
results. For example, in [6], we show that the above proof can be extended in the
following direction.
Theorem 5.3.4. For every separable Banach space V over the eld C, there is
a homogeneous polynomial P : (1 ) V of degree 2 dened on the complex
version of (1 ) which maps every closed non-separable subspace of (1 ) onto a
somewhere dense subset of V . 

As one may expect, Banach spaces over the eld C of complex numbers and
Banach spaces over the eld R of real numbers behave quite dierently when
considering zeros of polynomials. For example, on the version of the space 1 over
the eld R there exist homogeneous polynomials of any given even degree that
vanish only at 0. It is for this reason that, when studying zero spaces of polynomials
in the case of Banach spaces over R, one considers only polynomials of odd degrees.
Even under this restriction there is a considerable dierence between the real and
complex case as the following result from [5] shows.
Theorem 5.3.5. Let X be an innite-dimensional Banach space over the eld R of
real numbers whose dual X is weak -separable. Then for every odd integer n > 1
there is a homogeneous polynomial P : X R of degree n which is not zero on
any innite-dimensional subspace of X. 

On the other hand, we have the following result from [83] for polynomials
over the eld of complex numbers.
Theorem 5.3.6. Let X be an innite-dimensional Banach space over the eld C
of complex numbers and let P : X C be any polynomial such that P (0) = 0.
Then there is innite-dimensional subspace Y of X on which P is identically equal
to 0. 

Note the following immediate corollary of Theorem 5.3.5 which should be


compared both to Theorem 5.3.3 and Theorem 5.3.6.
Theorem 5.3.7. If is an index-set of cardinality at most the continuum (so in
particular, if = 1 ), then for every odd integer n > 1 there is a homogeneous
polynomial P : () R of degree n which is not zero on any innite-dimensional
subspace of (). 
5.3. Some geometrical applications of the square-bracket operation 149

On the other hand, in the case of the eld of complex numbers, we have the
following problem.
Question 5.3.8. Let be an index set of cardinality continuum and let us consider
the version of the Banach space () over the eld C of complex numbers. Is there
a homogeneous polynomial P : () C which is not zero on any non-separable
subspace of ()?
Remark 5.3.9. Recall that by the result of [83] there will be always an innite-
dimensional subspace of () on which P is identically equal to 0. Part of the
interest in this question comes from the fact that for sets of cardinality con-
tinuum there is essentially only one known example of a mapping f : []2 2
which is not constant on any square [A]2 of an uncountable subset A of . This
is the well-known partition of Sierpinski, obtained by comparing the usual order-
ing of the reals with a well-ordering. However, it can be shown, see [6], that if
f : []2 2 is the Sierpinski partition, then the corresponding homogeneous
polynomial Pf : () C dened by4

Pf (x) = f (, )x x
,

does admit a non-separable null subspace. The reader is refereed to papers [5]
and [6] where this problem is treated for bigger index-sets where the dierence
between the elds R and C is even more apparent.
We shall now see that the square-bracket operation can also be eectively
used in constructing Banach spaces X with some extreme properties of bounded
operators T : X X.
Theorem 5.3.10. There is a non-separable reexive Banach space E with the prop-
erty that every bounded linear operator T : E E can be expressed as T = I +S
where is a scalar, I the identity operator of E, and S a bounded operator with
separable range.
Proof. Let I = 3 [1 ]< and let us identify the index-set I with 1 , i.e., pretend
that the projection [[]] = [[]]G given above in Denition 5.2.1 takes its values in I
rather than in 1 . Let

[[]]0 : [1 ]2 3 and [[]]1 : [1 ]2 [1 ]<

be the rst and second coordinates of [[]] : [1 ]2 3 [1 ]< , respectively. Let

G = {G [1 ]< : [[]]0 = 0 for all {, } [G]2 },

H = {H [1 ]< : [[]]0 = 1 for all {, } [H]2 }.


4 Ascustomary, in this formula, we are identifying f with a symmetric function f : 2 2
which is identically equal to 0 on the diagonal.
150 Chapter 5. The Square-bracket Operation on Countable Ordinals

Let K be the collection of all nite sets {{i , i } : i < k} of pairs of countable
ordinals such that for all i < j < k:

max{i , i } < min{j , j }, (5.3.10)

[[i j ]]0 = [[i j ]]0 = 2, (5.3.11)


[[i j ]]1 = [[i j ]]1 = {l : l < i} {l : l < i}. (5.3.12)
The following properties of G, H and K should be clear:
G and H contain all the singletons, are closed under subsets
and they are 1-orthogonal to each other in the sense that (5.3.13)
G H contains no doubleton.
G and H are both 2-orthogonal to the family of the unions
(5.3.14)
of members of K.
If K and L are two distinct members of K, then there are
no more than ve ordinals such that {, } K and (5.3.15)
{, } L for some = .
For every sequence { , } ( < 1 ) of pairwise-disjoint
pairs of countable ordinals there exist arbitrarily large nite
(5.3.16)
sets , 1 such that { : } G, { : } H
and {{ , } : } K.
For a function x from 1 into R, set
1

x
H,2 = sup{(H x()2 ) 2 : H H},
1

x
K,2 = sup{({,}K (x() x())2 ) 2 : K K}.
2 = {x :
x
< }. Let 1
Let

= max{

,

H,2 ,

K,2 } and dene E
be the characteristic function of {}. Finally, let E2 be the closure of the linear
span of {1 : 1 } inside (E2 ,

). The following facts about the norm

are easy to establish using the properties of the families G, H and K listed above:

If x is supported by some G G, then


x
2
x
. (5.3.17)

If x is supported by K for some K in K, then
x
10
x
. (5.3.18)
The role of the seminorm

H,2 is to ensure that every bounded operator T :
E2 E2 can be expressed as D + S, where D is a bounded diagonal operator
relative to the basis5 1 ( < 1 ) and where S has separable range. Namely, if this
were not so, we could nd a sequence { , } ( < 1 ) of pairwise-disjoint pairs
from 1 such that T (1 )( ) = 0 for all < 1 . Thinning the sequence, we may
5 Indeed it can be shown that 1 ( < ) is a transnite basis of E in the sense of [100]. So
1 2
every vector x of E2 has a unique representation as <1 x()1 and the projection operators
P : E2 E2  ( < 1 ) are uniformly bounded.
5.3. Some geometrical applications of the square-bracket operation 151

assume that for some xed r > 0 and all < 1 , |T (1 )( )| r. By (5.3.16) for
every n < we can nd a subset of 1 of size n such that G = { : } G
and H = { : } H. Then, by (5.3.17),
12

2
T

T (G)
T (G )()2 r n.
H

Since n is arbitrary, this contradicts the fact that T is bounded. So from now on
we may concentrate on showing that every bounded diagonal operator has the
form I + S, where S is a bounded operator with separable range.
The role of the seminorm

K,2 is to represent every bounded diagonal
operator
D(x)() = x() ( < 1 )
as I + S for some scalar and operator S with separable range. This reduces to
showing that the sequence ( < 1 ) is eventually constant. If this were false,
we would be able to extract a sequence { , } ( < 1 ) of pairwise-disjoint
pairs of countable ordinals and an r > 0 such that | | r for all < 1 .
By (5.3.16), for every n < , we can nd a subset of 1 of size n such that
K = {{ , } : } belongs to K. Then, by (5.3.18),

10
D

D( K )

1
( (D( K )( ) D( K )( ))2 ) 2
1
= ( ( )2 ) 2

r n.

Since n was arbitrary this contradicts our initial assumption that D is a bounded
operator.
Note that
x
2
x
2 for all x 2 (1 ). It follows that 2 (1 ) E2
and the inclusion is a bounded linear operator. Note also that 2 (1 ) is a dense
subset of E2 . Therefore E2 is a weak compactly-generated space. For example,
W = {x 2 (1 ) :
x
2 1} is a weakly-compact subset of E2 and its linear
span is dense in E2 . To get a reexive example out of E2 one uses an interpolation
method of Davis, Figiel, Johnson and Pelczynski [17] as follows. Let pn be the
Minkowski functional of the set 2n W + 2n Ball(E2 ).6 Let
1
E = {x E2 :
x
E = ( 2 2
n=0 pn (x) ) < }.

By [17, Lemma 1], E is a reexive Banach space and 2 (1 ) E E2 are


continuous inclusions. Note that pn (x) < r i x = y + z for some y E2 and
z 2 (1 ) such that
y
< 2n r and
z
2 < 2n r. The following two facts about
6 In other words, pn (x) = inf{ > 0 : x B}, where B = 2n W + 2n Ball(E2 ).
152 Chapter 5. The Square-bracket Operation on Countable Ordinals

E correspond to the facts (5.3.17) and (5.3.18) and they are used in a similar way
in the proof that every bounded operator T : E E has the form I + S.

If x is the vector supported by some G G which can be


split into n blocks such that the kth one has size 24k and
(5.3.19)
x takes the value 22k on each term of the kth block, then

x
E 2.

If {{ , } : } is a member of K so that can be


4k
split into n blocks such that the k-th
 one k has size 2 (5.3.20)
and if x is the vector supported by K for k , x( ) =
22k and x( ) = 22k , then
x
E 12.
We leave the checking of (5.3.19) and (5.3.20) as well as the corresponding proof
that E has few operators to the interested reader. 
Remark 5.3.11. The rst example of a reexive Banach space X with a Schauder
basis of length 1 on which every bounded linear operator is the sum of a scalar
multiple of the identity and an operator with a separable range, is given by Wark
[130] who was motivated by a previous and similar construction due to Shelah and
Steprans [97], a construction in which the square-bracket operation also plays the
key role.

5.4 A square-bracket operation from a


special Aronszajn tree
In this section we try to show that the basic idea of the square-bracket operation
on 1 can perhaps be more easily grasped by working on an arbitrary special
Aronszajn tree rather than T (0 ). So let T = T, <T  be a xed special Aronszajn
tree and let
a : T
be a xed map witnessing this, i.e., a mapping with the property that for two
dierent nodes s and t of T , the equality a(s) = a(t) implies that s and t are
incomparable in T . We shall also assume that for every s, t T the greatest lower
bound s t exists in T . For t T and n < , set

Fn (t) = {s T t : s = t or a(s) n}.

Finally, for s, t T with ht(s) ht(t), let

[st]T = min{v Fa(st) (t) : ht(v) ht(s)}.7 (5.4.1)

The following fact corresponds to Lemma 5.1.4 in the case when T = T (0 ).


7 If ht(s) ht(t) we let [st]T = [ts]T .
5.4. A square-bracket operation from a special Aronszajn tree 153

Lemma 5.4.1. If X is an uncountable subset of T , the set of nodes of T of the form


[st]T for some s, t X intersects a closed and unbounded set of levels of T . 

We do not give a proof of this fact as it is almost identical to the proof of


Lemma 5.1.4 which deals with the special case T = T (0 ). But one can go further
and show that []T shares all the other properties of the square-bracket operation
[] described in the previous section. Some of these properties, however, are easier
to visualize and prove in the general context. For example, consider the following
fact, which in the case T = T (0 ) is the essence of Lemma 5.2.3.
Lemma 5.4.2. Suppose A T is an uncountable antichain and that for each t A
there is given a nite set Ft of its successors. Then for every stationary set 1
there exists an arbitrarily large nite set B A such that the height of [xy]T
belongs to whenever x Fs and y Ft for some s = t in B. 

Let us now examine in more detail the collection of graphs G ( 1 ) of


Denition 5.1.11 but in the present more general context.
Denition 5.4.3. For 1 , let

K = {{s, t} [T ]2 : ht([st]T ) }.

Working as in Lemma 5.1.12 one shows that (T, K ) and (T, K ) have no
isomorphic uncountable subgraphs whenever the symmetric dierence between
and is a stationary subset of 1 , i.e., whenever they represent dierent members
of the quotient algebra P(1 )/N S. In particular, K contains no square [X]2 of
an uncountable set X T whenever contains no closed and unbounded subset
of 1 . The following fact is a sort of converse to this.
Lemma 5.4.4. If contains a closed and unbounded subset of 1 , then there is a
proper forcing notion introducing an uncountable set X T such that [X]2 K .

Proof. The forcing notion P is dened to be the set of all pairs p = Xp , Np  where:

Np is a nite -chain of countable elementary submodels


(5.4.2)
of H2 containing all the relevant objects.

Xp is a nite subset of T such that ht([st]T ) for every


(5.4.3)
pair {s, t} of elements of Xp .
For all s = t in Xp there exist N Np such that s N i
(5.4.4)
t
/ N.
For every s = t in Xp , N Np , if t
/ N then the height of
(5.4.5)
min(Fa(st) (t) \ N ) belongs to .
The ordering of P is the coordinatewise inclusion. To show that P is a proper
forcing notion, let M be a given countable elementary submodel of some large
154 Chapter 5. The Square-bracket Operation on Countable Ordinals

enough H containing all the relevant objects and let p P M . We shall show
that
q = Xp , Np {M H2 }
is an M -generic condition extending p. So let D M be a dense-open subset of P
and let r be a given extension of q. Extending r we may assume that r D. Let
p0 = r M . Note that p0 P M . Let m0 be the maximum of the union of the
following two nite sets of integers

{a(s t) : s, t Xr },

{a(t  ) : t Xr , = N 1 for some N Np such that t


/ N }.
Let k be the size of the projection of the set Xr \ M on the level TM1 of
the tree T . Let F be the family of all k-element subsets F of T for which one can
nd r D realizing the same type as r over p0 such that, if N is the minimal
model of Nr \ Np0 , then

F = {x  (N 1 ) : x Xr \ Xp0 }.

Since the projection of Xr \ M onto TM1 belongs to F , the family is a rather


large subset of [T ]k . It follows that the family E of all k-element subsets E of T
for which we can nd < 1 such that

FE = {F F : {x  : x F } = E}

is uncountable, is also quite large. In particular, we can nd an ordinal 0 M 1


above the heights of all nodes from the set

{M Fm0 (t) : t Xr }

and E1 E M such that

{x  0 : x E1 } = {x  0 : x Xr \ M }

and such that every node of Xr \ M is incomparable with every node of E1 . Let

m+ = max{a(x y) : x Xr \ M, y E1 },

m = min{a(x y) : x Xr \ M, y E1 and a(x y) > m0 }.


Let 0 be an ordinal of M 1 above 0 which bounds the heights of all nodes
from the set 
{M Fm+ (t) : t Xr \ M }.
By the denitions of E and F there is r D M realizing the same type over p0
as r such that, if N is the minimal model of Nr \ Np0 , then N 1 0 and the
projection
F = {x  (N 1 ) : x Xr \ Xp0 }
5.4. A square-bracket operation from a special Aronszajn tree 155

dominates the set E1 . We claim that

q = Xr Xr , Nr Nr 

is a condition of P. This will nish the proof of properness of P since clearly q


extends both r and r. Clearly, only (5.4.3) and (5.4.5) for q require some argu-
mentation. We check rst (5.4.5), so let s and t be a pair of distinct members of
Xr Xr and let N Nr Nr be a model such that t / N . Let m = a(s t). If
t Xr and if there is s Xr such that s = t and s t = s t, the conclusion
that the height of min(Fm (t)\ N ) belongs to follows from the fact that r satises
(5.4.5). Similarly one considers the case when t Xr \ M and when s t = s t
for some s Xr , s = t. For if N belongs to Nr , the conclusion follows from the
fact that r satises (5.4.5) and if N Nr \ Nr , then by the choice of r,

Fm (t) \ N = Fm (t) \ M,

and so the conclusion that the height of min(Fm (t) \ N ) belongs to follows from
the fact that r satises (5.4.5) for s and t from Xr and M H2 from Nr . Let us
now consider the case t Xr \ M , s Xr \ Xr and s t = st for all s Xr \ {t}.
Note that in this case we have the following inequalities:

m0 < m m = a(s t) m+ .

So by the choice of m0 the set Fm (t) contains t  (N  1 ) for every N  Nr .


Since every ordinal of the form N  1 (N  Nr ) belongs to we are done in
case N belongs to Nr . So suppose N Nr \ Nr . Then by the choice of the bound
0 and the choice of r,
Fm (t) \ N = Fm (t) \ M
so the conclusion of (5.4.5) follows again from the fact that M H2 belongs to
Nr . Consider now the remaining case t Xr \ Xr and s Xr \ Xr. Note that in
this case the model N must belong to Nr , so if s t = s t for some s Xr,
s = t, the conclusion follows from the fact that r satises the condition (5.4.5). So
let us consider the subcase when s t = s t for all s Xr, s = t. Thus, s t
is the new splitting, so by the choice of 0 and E0 we have the inequalities:

m0 < m m = a(s t).

Recall that r was chosen to realize the same type over p0 as r so we in particular
have that Fm0 (t) and therefore its superset Fm (t) contains the projection t 
(N 1 ). It follows that min(Fm (t) \ N ) = t  (N 1 ), so its height is indeed a
member of (as N and contains a closed and unbounded subset of 1 ).
It remains to check (5.4.3) for q. So let s and t be a given pair of distinct
members of Xr Xr . We need to show that the height of [st]T belongs to . Clearly
the only nontrivial case is when, say, s Xr \ Xr and t Xr \ Xr. Suppose rst
that s t = s t for some s Xr , s = t. Since M H2 belongs to Nr and since
156 Chapter 5. The Square-bracket Operation on Countable Ordinals

r satises (5.4.5), the height of min(Fa(s t) (t) \ M ) belongs to . Since the height
of s is above the bound 0 for Fm+ (t) M ,

[st]T = min(Fa(st) (t) \ M ) = min(Fa(s t) (t) \ M ),

so we are done in this subcase. Suppose now that s t is a new splitting, i.e., that

m0 < m = a(s t) m+ .

So again we know that in this subcase, Fm (t) contains the projection t  (M 1 ).


By our choice of the bound 0 and the condition r we know that this projection
is the minimal point of Fm (t) whose height is bigger than or equal to the height
of s. By the denition of the square-bracket operation, we have

[st]T = t  (M 1 ).

Since M 1 belongs to this nishes our checking that q satises (5.4.3). More-
over, this also nishes the proof that the forcing notion P is proper.
It is clear that P forces the square of the union X of the Xp s for p belonging
to the generic lter to be included in K . What is not so clear is that P forces
the set X to be uncountable as promised. To ensure this we replace P with its
restriction P( r) to the condition

r = {t}, {M H2 },

where M is an arbitrary countable elementary submodel of some large enough H


containing all the relevant objects and where t is any node of T of height at least
M 1 . Note that the above proof of properness of P starting from p = , 
shows that
q = , {M H2 }
is an M -generic condition of P, so in particular its extension r is also M -generic.
It follows that if we let M [G] be the name for the set formed by interpreting all
P-names belonging to M itself. Then r forces X to be an element of the countable
elementary submodel M [G] of H [G]
which contains a point t of height above
M [G] 1 . It follows that r forces X to be uncountable. This nishes the proof
of Lemma 5.4.4. 
Corollary 5.4.5. The graph K contains the square of some uncountable subset of
T in some 1 -preserving forcing extension, if and only if is a stationary subset
of 1 .
Proof. If is disjoint from a closed and unbounded subset, then in any 1 -
preserving forcing extension its complement = 1 \ will be a stationary subset
of 1 . So by the basic property in Lemma 5.4.1 of the square-bracket operation,
no such a forcing extension will contain an uncountable set X T such that
[X]2 K . On the other hand, if is a stationary subset of 1 , going rst to
5.5. A square-bracket operation from the complete binary tree 157

some standard 1 -preserving forcing extension in which contains a closed and


unbounded subset of 1 and then applying Lemma 5.4.4, we get an 1 -preserving
forcing extension having an uncountable set X T such that [X]2 K . 
Remark 5.4.6. Corollary 5.4.5 gives us a further indication of the extreme com-
plexity of the class of graphs on the vertex-set 1 . It also bears some relevance
to the recent work of Woodin [135] who, working in his Pmax -forcing extension,
was able to associate a stationary subset of 1 to any partition of [1 ]2 into two
pieces. So one may view Corollary 5.4.5 as some sort of converse to this, since
in the Pmax -extension one is able to get a suciently generic lter to the forcing
notion P = P of Lemma 5.4.4 that would give us an uncountable X T such
that [X]2 K . In other words, under a bit of PFA or Woodins axiom (*), a set
1 contains a closed and unbounded subset of 1 if and only if K contains
[X]2 for some uncountable X T .

5.5 A square-bracket operation from the


complete binary tree
The variation of the square operation can also be based on the complete binary
tree of height rather than a special Aronszajn tree. To dene this variation we
start with a one-to-one sequence

{r : < 1 } 2

of branches through the complete binary tree R = 2< , and a mapping

e : [1 ]2

such that e(, ) = e(, ) whenever < < < 1 .


For n < and < 1 , set

Fn () = { : = or e(, ) n}.

For < < 1 , set

(, ) = min{n < : r (n) = r (n)}.

Finally,we are in a position to dene the variation of the square-bracket operation


on 1 by putting for < < 1 ,

[]R = min(F(,) () \ ). (5.5.1)

Then we have the following fact whose proof is very similar to the proof of Lemma
5.1.4, the corresponding fact for the original square-bracket operation.
158 Chapter 5. The Square-bracket Operation on Countable Ordinals

Lemma 5.5.1. If is an uncountable subset of 1 , the set of all ordinals of the form
[]R for some < in contains a closed and unbounded subset of 1 . 
A simple analysis of the proof of Lemma 5.1.7 for the original square-bracket
operation gives the following result for the variation []R .
Lemma 5.5.2. For every uncountable family A of pairwise-disjoint subsets of 1 , all
of the same size n, there exist an uncountable B A, and a mapping hb : n n
1 for each b B such that for all a < b in B, we have that [a(i)b(j)]R = hb (i, j)
whenever (i, j) n n \ 8 . Moreover, for every uncountable C B, the set of
ordinals for which there exist a < b in C such that [a(i)b(i)]R = for all i < n
contains a closed and unbounded subset of 1 . 
Let h ( < 1 ) be a list of all mappings of the form h : 2n 2n , where
n = n(h) < , such that every such mapping appears stationarily often on the
list. This leads us to a projection

[]0R : [1 ]2

of the square-bracket operation []R , dened as

[]0R = h[]R (r  n(h[]R ), r  n(h[]R )).

Then working as in the proof of Theorem 5.2.5, we conclude that Lemma 5.5.2
has the following immediate consequence.
Theorem 5.5.3. For every uncountable family A of pairwise-disjoint subsets of 1 ,
all of some xed size n, there exist an uncountable B A, and a single mapping
h : n n such that for all a < b in B, we have that

[a(i)b(j)]0R = h(i, j) whenever (i, j) n n \ .

Moreover, for every uncountable C B and for every mapping g : n 1 there


exist a < b in C such that

[a(i)b(i)]0R = g(i) for all i < n. 

We nish this section with an application of []0R . Suppose V and W are


vector spaces over the same eld K. Recall that a mapping : V V W is a
bilinear map if it is bilinear in each of the variables separately, or in other words,
if it satises the equalities of the form

(m n m n
1 i xj , 1 j yj ) = 1 1 i j (xi , yj ).

If (x, y) = (y, x) then we call a symmetric bilinear map. We say that (V, W, )
is a non-degenerate bilinear space, if for every x = 0 in V there is y in V such that
(x, y) = 0.
8 Here, = {(i, i) : i < n} is the diagonal of n n.
5.5. A square-bracket operation from the complete binary tree 159

Theorem 5.5.4. Suppose V and W are two vectors spaces over the same eld K
such that W is of countable (including nite) dimension while V has dimension
1 . Then there is a symmetric non-degenerate bilinear map : V V W such
that {(x, y) : x, y X, x = y} = W for every uncountable subset X of V . So in
particular, there is no uncountable subset X of V that is biorthogonal relative to
, i.e., that has the property that (x, y) = 0 for x = y in X.
Proof. Fix a linear basis B = {v : < 1 } of V . In order to simplify the
notation, we assume that []0R takes its values in W rather than . Since an
arbitrary symmetric mapping from BB into W extends to a uniquely determined
symmetric bilinear mapping from V V into W , it suces to dene  B B.
For < 1 , let (v , v ) be an arbitrary vector of W . For = , set
(v , v ) = (v , v ) = []0R .
Consider an uncountable subset X of V and x a vector t in W . We need to nd
x = y in X such that (x, y) = t. Since the eld K is countable, going to an
uncountable subset of X, we may assume that for some positive integer n and
some xed sequence i (i < n) of nonzero scalars, every vector x X, has a
representation of the form
x = i<n i vax (i) ,
for some subset ax of 1 of size n. We may assume that ax (x X) form a
-system with some root a.
Consider rst the case a = , or in other words, the case that ax (x X) is
a family of pairwise-disjoint sets. By the rst part of Theorem 5.5.3, we can nd
uncountable Y X and a mapping h : n n W such that for every x, y Y
with the property that ax < ay , we have
[ax (i)ay (j)]0R = h(i, j) for all (i, j) n n \ . (5.5.2)
Changing, h, we may assume that h(i, i) = 0 of all i < n. Let
u = (i,j)=(0,0) i j h(i, j).
Let v = t u and let w = (1/20 )v. Dene g : n W by letting g(i) = 0 for i = 0
and g(0) = w. By the second part of Theorem 5.5.3, we can nd x, y Y such
that ax < ay and
[ax (i)ay (i)]0R = g(i) for all i < n. (5.5.3)
Using (5.5.2) and (5.5.3) and referring to how u, v, and w were dened, we conclude
that
(x, y) = 20 [ax (0)ay (0)]0R + u = 20 (1/20 )v + u = t,
and so we are done in this case.
Let us now consider the case a = . Fix an element x0 of X. Find uncountable
Y X and w W such that (x0 , y) = w for all y Y . The family
{y x0 : y Y }
160 Chapter 5. The Square-bracket Operation on Countable Ordinals

is an uncountable family of disjointly supported vectors, so by the rst part of the


proof, we can nd y = z in Y such that

(y x0 , z x0 ) = t 2w + (x0 , x0 ).

Since (y x0 , z x0 ) = (y, z) 2w (x0 , x0 ), it follows that (y, z) = t, as


required.
Let V = {x V : (x, y) = 0 for all y V }. Then by the basic property of
that we have just established, it follows that V is a countable vector subspace
of V , so its linear complement V0 in V is of dimension 1 . Note that the restriction
 V0 V0 is non-degenerate while it still keeps the basic property of . So, taking
a linear isomorphism between V and V0 and transferring  V0 V0 to a bilinear
map on V V , we get the full conclusion of the theorem. 
Remark 5.5.5. As indicated above, the variation []R of the square-bracket opera-
tion appears rst as the mapping b of [109, p. 287]. Its application given above in
Theorem 5.5.4 is motivated by a similar application of []R appearing in [12] (see
also [96]).
Chapter 6

General Walks and


Their Characteristics

6.1 The full code and its application in


characterizing Mahlo cardinals
One of the most basic questions frequently asked about set-theoretical trees is the
question whether they contain any conal branch, a branch that intersects each
level of the tree. The fundamental importance of this question has already been
realized in the work of Kurepa [65] and then later in the works of Erdos and Tarski
[32] in their respective attempts to develop the theory of partition calculus and
large cardinals. A tree T of height equal to some regular cardinal may not have
a conal branch for a very special reason as the following denition indicates.
Denition 6.1.1. For a tree T = T, <T , a function f : T T is regressive if
f (t) <T t for every t T that is not a minimal node of T . A tree T of height
is special if there is a regressive map f : T T with the property that the
f -preimage of every point of T can be written as the union of < antichains of T .
This denition in case = 1 reduces indeed to the old denition of spe-
cial tree, a tree that can be decomposed into countably many antichains. More
generally we have the following:
Lemma 6.1.2. If is a successor cardinal, then a tree T of height is special if
and only if T is the union of < antichains. 
The new denition, however, seems to be the right notion of speciality as it
makes sense even if is a limit cardinal.
Denition 6.1.3. A tree T of height is Aronszajn if T has no conal branches
and if every level of T has size < .
162 Chapter 6. General Walks and Their Characteristics

Recall the well-known characterization of weakly compact cardinals due to


Tarski and his collaborators: a strongly inaccessible cardinal is weakly compact
if and only if there are no Aronszajn trees of height . We supplement this with
the following:
Theorem 6.1.4. The following are equivalent for a strongly inaccessible cardinal :
(1) is Mahlo.
(2) There are no special Aronszajn trees of height .
Proof. Suppose is a Mahlo cardinal and let T be a given tree of height all of
whose levels have size < . To show that T is not special let f : T T be a given
regressive mapping. By our assumption of there is an elementary submodel M
of some large enough structure H such that T, f M and = M is a regular
cardinal < . Note that T  is a subset of M and since this tree of height is
clearly not special, there is t T  such that the preimage f 1 (t) is not the
union of < antichains. Using the elementarity of M we conclude that f 1 (t) is
actually not the union of < antichains.
The proof that (2) implies (1) uses the method of minimal walks in a rather
crucial way. So suppose to the contrary that our cardinal contains a closed and
unbounded subset C consisting of singular strong limit cardinals. Using C, we
choose a C-sequence C ( < ) such that: C+1 = {}, C = ( , ) for limit
such that = sup(C ) < , but if = sup(C ) then take C such that:

tp(C ) = cf() < min(C ), (6.1.1)

= sup(C ) implies C, (6.1.2)


C and > sup(C ) imply = + 1 for some C. (6.1.3)
Given the C-sequence C ( < ) we have the notion of minimal walk along
the sequence and various distance functions dened above. In this proof we are
particularly interested in the function 0 from []2 into the set Q of all nite
sequences of ordinals from :

0 (, ) = tp(C ) 0 (, min(C \ ))

with the boundary value 0 (, ) = 0 for all . We would like to show that the
tree1
T (0 ) = {(0 )  : < }
is a special Aronszajn tree of height . Note that the size of the th level (T (0 ))
of T (0 ) is controlled in the following way:

|(T (0 )) | |{C : < }| + || + 0 . (6.1.4)


1 Recall,
that we consider T (0 ) ordered by the natural initial segment relation . Recall also
that < denotes the strict version of
6.1. The full code and its application in characterizing Mahlo cardinals 163

So under the present assumption that is a strongly inaccessible cardinal,


all levels of T (0 ) do indeed have size < . It remains to dene the regressive map

f : T (0 ) T (0 ) (6.1.5)

that will witness speciality of T (0 ). Note that it really suces dening f on all
levels whose index belong to our club C of singular cardinals. So let t = (0 ) 
be a given node of T such that C and < . Note that by our choice
of the C-sequence every term of the nite sequence of ordinals 0 (, ) is strictly
smaller than . So, if we let

f (t) = t  0 (, ), (6.1.6)

where  is a standard coding of nite sequences of ordinals by ordinals, we get


a regressive map. To show that f is one-to-one on chains of T (0 ), which would
be more than sucient, suppose ti = (0 )i  i (i < 2) are two nodes such that
t0 < t1 . Our choice of the C-sequence allows us to deduce the following general
fact about the corresponding 0 -function whose proof is similar to the proof of
Lemma 2.1.18 which treats the case = 1 .
If , is a limit ordinal, and if 0 (, ) = 0 (, )
(6.1.7)
for all < , then 0 (, ) = 0 (, ).
Applying this to the triple of ordinals 0 , 0 and 1 we conclude that

0 (0 , 0 ) = 0 (0 , 1 ). (6.1.8)

Now observe another fact about the 0 -function whose proof is almost identical to
that of case = 1 , seen above in Lemma 2.1.16.

If < , then 0 (, ) <r 0 (, ), (6.1.9)

where <r denotes the right lexicographical ordering between nite sequences of
ordinals < . Applying this to the triple 0 < 1 1 we in particular get that,

0 (0 , 1 ) = 0 (1 , 1 ). (6.1.10)

Combining (6.1.8) and (6.1.10), we get that 0 (0 , 0 ) = 0 (1 , 1 ) and therefore


that f (t0 ) = f (t1 ). 
It turns out that the regressive mapping dened during the course of the
previous proof leads us to the following purely Ramsey-theoretic characterization
of Mahlo cardinals.
Theorem 6.1.5. A cardinal is a Mahlo cardinal if and only if every regressive
map f dened on a cube [C]3 of a closed and unbounded subset of has an innite
min-homogeneous set X C.2
2 Recall, that X is min-homogeneous for f if f (, , ) = f ( ,  ,  ) for every pair < <
and  <  <  of triples of elements of X such that =  .
164 Chapter 6. General Walks and Their Characteristics

Proof. To see the direct implication, let f : []3 be a given regressive mapping.
Dene the tree-ordering f on by letting f if and
f (u {}) = f (v {}) for all u, v [{ : <f }]2 with min u = min v.
Since is a strongly inaccessible cardinal, one easily concludes that the levels
of the tree (, <f ) have cardinalities < . So, using the assumption that is a
Mahlo cardinal, we can nd an inaccessible cardinal < such that the set of
predecessors { : <f } has cardinality . Thus we have reduced the conclusion
to the fact that regressive mappings g : []2 for inaccessible cardinals
have innite min-homogeneous sets. This fact in turn is proved by considering the
corresponding tree-orderings g and noticing that, by the inaccessibility of , the
tree (, g ) must have an innite chain (and in fact a chain of arbitrary order type
< ). Now note that any chain of the tree (, g ) is min-homogeneous relative
to g.
For the converse implication, note that the statement about regressive maps
implies that must be a strongly inaccessible cardinal. So assume is a strongly
inaccessible non-Mahlo cardinal. Let C be a closed and unbounded subset of
consisting of strong limit singular cardinals. We shall nd a regressive map on
[C]3 without an innite min-homogeneous set. This will give us an opportunity to
further expose the tree T (0 ) and the regressive map
f : T (0 )  C T (0 )
dened above in (6.1.6) in the course of proving Theorem 6.1.4. We have already
observed that the length of every branching node t must be of the form + 1 for
some and in fact the following more precise description is true:
If t is a branching node of T (0 ), then its length is equal to
(6.1.11)
+ 1 for some C.
To see this, let t be a given branching node of T (0 ). By (6.1.7) we know that the
length of t is equal to + 1 for some ordinal . Let
= max(C ( + 1)) and + = min(C \ ( + 1)).
Let , > + 1 be two ordinals such that (0 ) and (0 ) split at t. Let =
0 > > k = and = 0 > > l = be the walks from to
and to respectively. From the fact that (0 ) and (0 ) agree below we
conclude that k = l and Ci = Ci for all i k. By the choice of the
C-sequence and the inequality > , we can deduce a sequence of stronger
agreements Ci + = Ci + for all i k. However, that would mean that
0 ( + 1, ) = 0 ( + 1, ), contrary to our assumption that (0 ) and (0 ) split
at t. It follows that = , and therefore C, which was to be shown.
Going back to the proof of Theorem 6.1.5, recall the denition of the regres-
sive map f : T (0 )  C T (0 ) given above in Theorem 6.1.4:
f ((0 )  ) = (0 )  0 (, ).
6.1. The full code and its application in characterizing Mahlo cardinals 165

We extend f to the whole tree T (0 ), by letting f (t) = t  , where is the


maximal element of C that is less than the length of t for any node t T (0 )
whose length does not belong to C. Then we know that f is one-to-one on any
chain whose nodes are separated by C. Using the mapping f we shall derive three
other regressive maps:

p : [C]3 , q : [C]3 and r : [C]2 .

First note that for = in C the functions (0 ) and (0 ) are incomparable, so it


is natural to denote by the node of T (0 ) where they split. For < < < ,
let
p(, , )=the length of f n ( ), where n is the minimal
(6.1.12)
integer such that f n ( )  (0 ) .
q(, , ) = min{n : f n ( ) = }. (6.1.13)
r(, ) = 0 (, ), 0 (, ), where = length( ). (6.1.14)
Now we claim that no innite subset of C can be simultaneously min-homo-
geneous with respect to all three regressive maps. For, given an innite increasing
sequence {i } of elements of C, using Ramseys theorem we may assume that
either
i j = k l for all i = j and k = l in , or (6.1.15)
i i+1 < i+1 i+2 for all i < . (6.1.16)
Note that (6.1.15) would violate min-homogeneity with respect to the mapping r.
If (6.1.16) holds and if the sequence is min-homogeneous with respect to p and q,
then another application of Ramseys theorem would give us an integer m and an
ordinal such that for all i < j < k < :

m = min{n : f n (j k ) (0 )i }, (6.1.17)

f m (j k ) = (0 )0  (= t). (6.1.18)
By (6.1.11) the chain {i i+1 : i 1} is separated by C, so using the basic
property of f we conclude that m > 1. Then the pre-image f 1 (t) includes the set

{f m1 (i i+1 ) : i 1}.

By (6.1.17) we have,

i i+1 < f m1 (i+1 i+2 )  i+1 i+2 , (6.1.19)

so {f m1 (i i+1 ) : i 1} is an innite chain of T (0 ) separated by C, a


contradiction. This nishes the proof of Theorem 6.1.5. 
Starting from the case n = 1, one can go further and obtain the following
characterization of n-Mahlo cardinals.
166 Chapter 6. General Walks and Their Characteristics

Figure 6.1: The regressive mapping f .

Theorem 6.1.6. The following are equivalent for an uncountable cardinal and a
positive integer n:

(1) is n-Mahlo.
(2) For every < and every regressive map dened on []n+2 , some unbounded
subset of has a min-homogeneous set of order type .
(3) Every regressive map dened on [C]n+2 for some closed and unbounded subset
C of has an innite min-homogeneous subset.

Proof. The proof of the direct implication of Theorem 6.1.5 gives us the necessary
inductive step for proving the implication from (1) to (2), so it remains only to
supply the argument for the implication from (3) to (1).
For a set X of ordinals and n , let W (X, n) denote the statement that
there exist f : [X]n+2 and regressive g : [X]n+3 sup(X) such that,
if H X is min-homogeneous for g and f  [H]n+2 = {k} for some k , then
6.1. The full code and its application in characterizing Mahlo cardinals 167

|H| n + k + 5. Note that in the course of proof of Theorem 6.1.5, we have


established W (C, 0) for every closed unbounded subset C of some limit ordinal
such that C contains no inaccessible cardinals, since the mapping q is really a
map from [C]2 into . The proof of the theorem is nished once we show that, in
general, W (C, n) holds for every set of ordinals C of limit order type that is closed
in its supremum and which contains no n-Mahlo cardinals. So let C be a given
closed and unbounded subset of some limit ordinal and suppose that C contains
no (n+1)-Mahlo cardinals. We need to show that W (C, n+1) holds. We rst show,
by induction on C, that W (C , n) holds for all C. Clearly, W (C , n)
holds for = min C. Suppose that W (C , n) holds and show that W (C + , n)
for + = min(C \( +1)). Let f and g be a pair of mappings witnessing W (C , n)
and assume, as we may, that the range of f does not include the value 0. Dene

f + : [C + ]n+2

by letting f + (0 ,...,n+1 ) = f (0 ,...,n+1 ) if n+1 < and f + (0 , . . . , n+1 ) = 0


if n+1 . Similarly, dene regressive

g + : [C + ]n+3

by letting g + (0 ,...,n+2 ) = g(0 ,...,n+2 ) if n+2 < and g + (0 ,...,n+2 ) = 0 if


n+2 . It is easily seen that the pair f + and g + of mappings witness W (C
+ , n). Assume now that is a limit point of C and that W (C , n) holds for all
C . For each C x a pair f and g witnessing W (C , n). Since
is not an (n + 1)-Mahlo cardinal, there is a closed and unbounded set C  C
containing no n-Mahlo cardinals. By the inductive hypothesis W (C  , n) holds, so
we can x a pair of mappings f  and g  witnessing this. For C , let

= sup(C  ( + 1)) and + = min(C  \ ( + 1)).

A C  -pattern of an element {0 , . . . , n+1 } of [C]n+2 written in its increasing


order is the equivalence relation E = E(0 , . . . , n+1 ) on the index-set {0, . . . , n +
1} where iEj if and only if
i = j . We can of course code (enumerate) all

C -patterns by natural numbers and assume that the ranges of the mappings f
and f  do not take any of these codes as values. Dene

f : [C ]n+2

by letting f (0 , . . . , n+1 ) = f  (
0 , . . . , n+1 ) if 0 < < n+1 , and by letting
f (0 , . . . , n+1 ) = f+ (
0 , . . . , n+1 ) if E(0 , . . . , n+1 ) is the trivial equivalence
0
relation where every pair of integers of its domain are equivalent; in all other
cases, f (0 , . . . , n+1 ) = k, where k is the integer which codes the C  -pattern of
{0 , . . . , n+1 }. Similarly, dene

g : [C ]n+3
168 Chapter 6. General Walks and Their Characteristics

by letting g(0 , . . . , n+2 ) = g  (


0 , . . . , n+2 ) if 0 < < n+2 , and by letting

g(0 , . . . , n+2 ) = g+ (0 , . . . , n+2 ) if E(0 , . . . , n+2 ) is the trivial equivalence
0
relation where every pair of integers of its domain are equivalent; in all other
cases, g(0 , . . . , n+2 ) = 0. It is easily checked that the mappings f and g witness
W (C , n). This completes our proof that W (C , n) holds for all C.
In order to prove that W (C, n + 1) holds we x for each C a pair of
mappings f and g witnessing W (C , n). Dene f : [C]n+3 and regressive
g : [C]n+4 as

f (0 , . . . , n+2 ) = fn+2 (0 , . . . , n+1 ) + 1,

g(0 , . . . , n+2 ) = gn+2 (0 , . . . , n+1 ).


Then f and g witness W (C, n + 1) nishing thus the proof. 
Remark 6.1.7. Theorems 6.1.5 and 6.1.6 are due to Hajnal, Kanamori and Shelah
[44]. An earlier characterization of this sort is due to Schmerl [91] who was the
rst to consider statements like W (C, n). In particular Schmerl [91] proves the
following characterization theorem using a bit dierent terminology (see also [44]),
a theorem which has some interesting metamathematical applications (see [38],
[56], and [54]).
Theorem 6.1.8. The following are equivalent for an uncountable cardinal and a
non-negative integer n:

(1) is n-Mahlo.
(2) For every positive integer k and every regressive mapping dened on []n+3 ,
some unbounded subset of has a min-homogeneous set of size k.
(3) Every regressive map dened on []n+3 for some unbounded subset of has
a min-homogeneous set of size n + 5.

The proof of Theorem 6.1.8 will be given in a sequence of lemmas. We start


with the following lemma which has essentially appeared above in the course of
the proof of Theorem 6.1.6.
Lemma 6.1.9. Suppose that n 3 is an integer, that > n is an ordinal, and
that is an innite limit ordinal. Suppose also that C and are subsets of \
with C closed and unbounded and min C min . Suppose further that there is a
regressive map f : [C]n with no min-homogeneous set of order type . If for
each < there is a regressive map g : [ ]n with no min-homogeneous
set of order type , then there is such a regressive map on []n as well.

Proof. By our assumption min C min and the fact that C is closed, for each
the ordinal = sup(C ( + 1)) belongs to C. For < , let

+ = sup(C \ ( + 1)).
6.1. The full code and its application in characterizing Mahlo cardinals 169

For a given n-tuple 0 < 1 < < n1 of elements of , one associates the
equivalence relation E(0 ,1 ,...,n1 ) on {0, 1, . . . , n 1} by letting

iE(0 ,1 ,...,n1 ) j if and only if


i = j .

We shall code such equivalence relations by integers E(0 ,1 ,...,n1 )  in some nat-
ural way. Choose a regressive map g : [C]n Ord which has no min-homogeneous
set of order type . Similarly, choose regressive mappings

f : [ ]n Ord, ( < )

that have no min-homogeneous set of order type . We may assume that the ranges
of these regressive mapping avoid the integers that code the equivalence relations
of the form E(0 ,1 ,...,n1 ) . Now, dene the regressive mapping f : []n Ord
according to the following three cases.

Case 1. If
0 < 1 < < n1 , let

f (0 , 1 , . . . , n1 ) = g(
0 , 1 , . . . , n1 ).

Case 2. If
1 = = n1 , let

f (0 , 1 , . . . , n1 ) = f+ (0 , 1 , . . . , n1 ).
1

Case 3. In all other cases, put,

f (0 , 1 , . . . , n1 ) = E(0 ,1 ,...,n1 ) .

Suppose there is a min-homogeneous subset X of of order-type . Let 0 < 1


be the least two elements of X. If
0 = 1 , then there will be no X with

> 1 , or else we would get two n-tuples with minimum 0 that realize two
dierent equivalence relations. It follows that in this case  is a constant
map on X. This means that Case 2 applies for all the n-tuples from [X]n . So, in this
case X would be a min-homogeneous set for f+ of order-type , a contradiction.
1

Assume now that


0 < 1 . Suppose rst that there is some X such

that > 1 . Then the mapping  must be strictly increasing on X,

or else we would nd two n-tuples of elements of X with minimum 0 realizing


two dierent equivalence relations, a contradiction. In follows that Case 1 applies
all the time for n-tuples from [X]n . So in this case { : X} would be a
subset of C of order-type that is min-homogeneous for the regressive mapping
g, a contradiction. It remains to consider the case
0 < 1 and = 1 for all
X with > 1 . This means that Case 2 applies again for the n-tuples from
[X]n . So, in this case X would be a min-homogeneous set for f+ of order-type ,
1
a contradiction. This nishes the proof. 
170 Chapter 6. General Walks and Their Characteristics

Lemma 6.1.10. Suppose that some limit ordinal > has the property that, for
every < and every regressive map f : [ \ ]3 , there exists a min-
homogeneous set of cardinality 4. Then is a strong limit cardinal.
Proof. Suppose that for some < , the cardinality of 2 is bigger than or equal
to the cardinality of . Fix a one-to-one sequence x ( < ) of elements of 2 ,
and dene a regressive map f : [ \ ]3 , by

f (0 , 1 , 2 ) = min{ < : x0 () = x1 ()}. (6.1.20)

It is clear that f has no min-homogeneous set of size 4. 


Lemma 6.1.11. Suppose that n > 0 is an integer, that is an inaccessible cardinal,
and that C is a closed and unbounded subset of consisting of strong limit cardinals
that are not (n 1)-Mahlo. Then there is a regressive mapping f : [C]n+3
which has no min-homogeneous set of size n + 5.
Proof. Let S1 (m, n, ) denote the statement that, for every inaccessible cardinal
and every closed and unbounded subset C of consisting of strong limit cardinals
that are not m-Mahlo, there is a regressive mapping f : [C]n which has
no min-homogeneous set of order-type . We need to prove that S1 (n 1, n +
3, n + 5) holds for every positive integer n. For this we will need another such a
statement S2 (m, n, ) saying that, for every inaccessible cardinal , every closed
and unbounded subset C of consisting of strong limit cardinals that are not
m-Mahlo, and every < , there is a regressive mapping f : [C ]n with
no min-homogeneous set of order-type . First of all note that

S2 (m, n, ) implies S1 (m, n + 1, + 1). (6.1.21)

To see this x an inaccessible cardinal and a closed unbounded set C


consisting of strong limit cardinals. By S2 (m, n, ), for each < , we can x also
a regressive mapping f : [C ]n with no min-homogeneous set of order-type
. Dene f : [C]n+1 , by

f (0 , 1 , . . . , n ) = fn (0 , 1 , . . . , n1 ).

It is clear that f is then a regressive mapping with no min-homogeneous set of


order-type + 1.
Claim. Suppose that n + 2, n 3. Suppose further that S1 (k, n, ) holds for
all k < m. Then S2 (m, n, ) holds.
Proof. Let C be a given closed and unbounded subset of some inaccessible cardinal
consisting only of strong limit cardinals that are not m-Mahlo. We need to
construct for each < a regressive mapping f : [C ]n with no min-
homogeneous set of order-type . This will be done by recursion on C. Suppose
rst that
= max(C ) < .
6.1. The full code and its application in characterizing Mahlo cardinals 171

We may assume that 0 is not in the range of f . Dene f : [C ]n by


f (0 , . . . , n1 ) = f (0 , . . . , n1 )
if n1 < ; otherwise put f (0 , . . . , n1 ) = 0. It is clear that f is a regressive
map with no min-homogeneous set of order-type . Suppose now that
sup(C ) = .
Then is either a singular strong limit cardinal or an inaccessible cardinal which
is not m-Mahlo. In any case, we can select a subset D C that is closed
and unbounded in and by the hypothesis of the claim (or singularity of ) a
regressive map g : [D]n which has no min-homogeneous set of order-type .
For < , set
= sup(D ( + 1) and + = min(D \ ( + 1).
As before, a D-pattern of an n-tuple {0 , 1 , . . . , n1 } of ordinals < written in
increasing order is the equivalence relation E = E(0 ,1 ,...,n1 ) on the index set
{0, 1, . . . , n 1}, where iEj if and only if
i = j . As before we code patterns
by positive integers as E(0 ,1 ,...,n1 )  and assume that neither of our regressive
mappings xed so far takes these integers as values. We dene a regressive mapping
f : [C ]n
according to the following three cases.
Case 1. If
0 < 1 < < n1 , let

f (0 , 1 , . . . , n1 ) = g(
0 , 1 , . . . , n1 ).

Case 2. If
0 = 1 = = n1 , let

f (0 , 1 , . . . , n1 ) = f+ (0 , 1 , . . . , n1 ).
0

Case 3. If E(0 ,1 ,...,n1 ) has at least two but less than n classes, we let
f (0 , 1 , . . . , n1 ) = E(0 ,1 ,...,n1 ) .
Suppose there is a min-homogeneous subset of C of order-type and let
= min . If the value of f at some n-tuple of elements of with minimum
gets assigned according to Case 1, then
= { : }
has order-type and is min-homogeneous for g, a contradiction. If the value of f
at some n-tuple of elements of with minimum gets assigned according to Case
2, then C + and is min-homogeneous for g+ , a contradiction. If the value
of f at some n-tuple of elements of with minimum gets assigned according
to Case 3, then this is true about all such n-tuples. However, our assumptions of
n + 2 and n 3 will give us two such n-tuples realizing dierent patterns, a
contradiction.This completes the proof of the claim. 
172 Chapter 6. General Walks and Their Characteristics

Note that the Claim in particular shows that S2 (0, 3, 5) holds. Applying (6.1.21)
and the claim repeatedly, we obtain that S1 (n 1, n + 3, n + 5) holds for every
positive integer n. This nishes the proof of Lemma 6.1.11. 

We are now ready to nish the proof of Theorem 6.1.8. The direct implication
from (1) to (2) (and, therefore (3)) uses the partition tree in a similar manner as
in the proof of Theorem 6.1.6, so we dont repeat it. So let us show that (3)
implies (1). First of all note that if (1) holds, the cardinal must be inaccessible
by Lemma 6.1.10. So assuming that is inaccessible, the implication from (3)
towards (1) follows directly from Lemma 6.1.11.
The following renement of Theorem 6.1.8 from [54] is also worth pointing out.
Theorem 6.1.12. The following are equivalent for every set of innite ordinals
and a non-negative integer n:
(1) is unbounded in for some n-Mahlo cardinal .
(2) Every regressive mapping f : []n+3 Ord has a min-homogeneous set of
size n + 5.

Proof. The direction from (1) to (2) follows from Theorem 6.1.8. To see the con-
verse implication, let be the minimal ordinal with the property that every
regressive mapping f : [ ]n+3 Ord has a min-homogeneous set of size n + 5.
Note that must be a limit ordinal. We show that is in fact an n-Mahlo cardinal.
By Lemma 6.1.9, we know that for every closed and unbounded set C with
properties
C = and min(C) min(),
every regressive map f : [C]n+3 Ord has a min-homogeneous set of size n + 5.
Applying this to closed and unbounded sets of the form

C = {min()} [, ) for < ,

by Lemma 6.1.10, we conclude that must be a strong limit cardinal. If were


a singular cardinal, then taking a closed unbounded subset D of of order-type
cf() with min(D) cf() and taking a regressive one-to-one mapping on [D]n+3
would contradict this property for the closed and unbounded set

C = {min()} D.

It follows that is an inaccessible cardinal, so if would not be an n-Mahlo


cardinal, we would have that n > 0 and that we have contradicted Lemma 6.1.11.
This nishes the proof. 

We nish this section by showing how the idea of proof of Theorem 6.1.4
leads us naturally to the following well-known fact, rst established by Silver (see
[75]) when is a successor of a regular cardinal.
6.1. The full code and its application in characterizing Mahlo cardinals 173

Theorem 6.1.13. If is a regular uncountable cardinal which is not Mahlo in


the constructible universe, then there is a constructible special Aronszajn tree of
height .
Proof. Working in L we choose a closed and unbounded subset C of consisting
of singular ordinals and a C-sequence C ( < ) such that
C+1 = {}, (6.1.22)
C = ( = sup(C ) < ,
, ) when (6.1.23)
while if is a limit point of C we take C to have the following two properties:
= sup(C ) implies C, (6.1.24)
> sup(C ) implies = + 1 for some C. (6.1.25)
It is clear that we can choose the C-sequence C ( < ) to have also the following
important property,
|{C : < }| || + 0 for all < . (6.1.26)
Then from what we have learned above these properties guarantee that the tree
T (0 ), where 0 is the full code of the walk along the C-sequence C ( < ), is
a constructible special Aronszajn tree of height . 
We are also in a position to deduce the following well-known fact.
Theorem 6.1.14. The following are equivalent for a successor cardinal :
(a) There is a special Aronszajn tree of height .
(b) There is a C-sequence C ( < ) such that tp(C ) for all and such
that {C : < } has size for all < .
Proof. If C ( < ) is a C-sequence satisfying (b) and if 0 is the associated
0 -function, then T (0 ) is a special Aronszajn tree of height . Suppose <T is a
special Aronszajn tree-ordering on such that [ , ( + 1)) is its th level.
Let C be the closed and unbounded set of ordinals < that are divisible by .
Let f : be such that the f -preimage of every ordinal < is an antichain
of the tree (, <T ). We choose a C-sequence C ( < ) such that C+1 = {},
C = ( , ) for limit with the property that = sup(C ) < , but if is a
limit point of C we take C more carefully as follows: C = { : < } where
= sup{ : < } for limit < , (6.1.27)
0 =the <T -predecessor of with minimal f -image, (6.1.28)
+1 =the <T -predecessor of with minimal f -image sub-
ject to the requirement that f (+1 ) > f (+1 ) for all (6.1.29)
< ,
is the limit ordinal where the process stops, i.e.,
(6.1.30)
sup{f (+1 ) : < } = .
174 Chapter 6. General Walks and Their Characteristics

Note that if and are two limit points of C and if <T , , then C =
C . From this one concludes that the C-sequence is locally small, i.e., that
{C : < } has size for all < . 

Corollary 6.1.15. If < = , then there exists a special Aronszajn tree of height + .

Corollary 6.1.16. In the constructible universe, special Aronszajn trees of any reg-
ular uncountable non-Mahlo height exist. 
Remark 6.1.17. In a large portion of the literature on this subject the notion of a
special Aronszajn tree of height equal to some successor cardinal + is somewhat
weaker, equivalent to the fact that the tree can be embedded inside the tree

S() = {f : : < & f is 1 1}.

One would get our notion of speciality by restricting the tree on successor ordinals,
losing thus the frequently useful property of a tree that dierent nodes of the same
limit height have dierent sets of predecessors. The result Corollary 6.1.15 in this
weaker form is due to Specker [101], while the result Corollary 6.1.16 in this weaker
form is essentially due to Jensen [52].

6.2 The weight function and its local versions


In this section we assume that = + and we x a C-sequence C ( < + ) such
that
tp(C ) for all < + . (6.2.1)
Let 1 : [+ ]2 be dened recursively by

1 (, ) = max{tp(C ), 1 (, min(C \ ))}

where we stipulate that 1 (, ) = 0 for all .


Lemma 6.2.1. |{ : 1 (, ) }| || + 0 for all < + and < .

Proof. Let + be the rst innite cardinal above the ordinal . The proof of the
conclusion is by induction on . So let be a given set of order-type + . We
need to nd such that 1 (, ) > . This will clearly be true if there is
such that tp(C ) > . So, we may assume that

tp(C ) for all . (6.2.2)

Then there must be an ordinal 1 C such that the set

1 = { : 1 = min(C \ )},
6.2. The weight function and its local versions 175

has size + . By the inductive hypothesis there is 1 such that (see (6.2.2))

1 (, 1 ) > tp(C ). (6.2.3)

It follows that

1 (, ) = max{tp(C ), 1 (, 1 )} = 1 (, 1 ) > .

This nishes the proof. 


Lemma 6.2.2. If is regular, then { : 1 (, ) = 1 (, )} has size < for
all < < + .
Proof. The proof is by induction on and . Let be a given set of order-type
. We need to nd such that 1 (, ) = 1 (, ). Let = sup() and

0 = max(C ), 0 = min(C \ ). (6.2.4)

Note that by our assumption on and the C-sequence, these two ordinals are well
dened and
0 < 0 < . (6.2.5)
By Lemma 6.2.1 and the inductive hypothesis, there is in (0 , ) such that

1 (, ) = 1 (, 0 ) > tp(C ). (6.2.6)

It follows that C = C and 0 = min(C \ ), and so

1 (, ) = max{tp(C ), 1 (, 0 )} = 1 (, 0 ) = 1 (, ).

This completes the proof. 


Remark 6.2.3. The assumption about the regularity of in Lemma 6.2.2 is essen-
tial. For example, it can be seen (see [13, p. 72]) that the conclusion of this lemma
fails if is a singular limit of supercompact cardinals.
Denition 6.2.4. Dene 1 : [+ ]2 by

1 (, ) = 21 (,) (2 tp{ : 1 (, ) = 1 (, )} + 1).

The following fact shows that this stretching of 1 keeps the basic coherence
property stated in Lemma 6.2.2 but it gives us also the injectivity property that
could be quite useful.
Lemma 6.2.5. If is a regular cardinal, then
(a) 1 (, ) = 1 (, ) whenever < < < + ,
(b) |{ : 1 (, ) = 1 (, )}| < whenever < < + .
176 Chapter 6. General Walks and Their Characteristics

Proof. Suppose that 1 (, ) = 1 (, ) for some < < + . Then


1 (, ) = 1 (, ) = , and
tp{ : 1 (, ) = } = tp{ : 1 (, ) = }.
This can only happen if = . This proves (a).
To prove (b), for a given < < + , set
D = { : 1 (, ) = 1 (, )}.
By Lemma 6.2.2, the set D has cardinality < , so by the regularity of ,
= sup{1 (, ) : D , {, } + 1 < .
It suces to show that, if some < is such that 1 (, ) > and 1 (, ) > ,
then 1 (, ) = 1 (, ). To see this, note that by the denition of , we must
have that 1 (, ) = 1 (, ) and that, if we let denote this equal value, then
{ : 1 (, ) = } = { : 1 (, ) = }.
So, if we let denote the order type of this set, we have that
1 (, ) = 2 (2 + 1) = 1 (, ),
as required. 
Remark 6.2.6. Note that Lemma 6.2.5 gives an alternative proof of Corollary
6.1.15 since under the assumption < = the tree
1 ) = {1 (, )  : < + }
T (
will have levels of size at most . It should be noted that the coherent sequence
) ( < + ) of one-to-one mappings is an object of independent interest which
(
can be particularly useful in stepping-up combinatorial properties of to + . It is

also an object that has interpretations in such areas as the theory of CechStone
compactications of discrete spaces (see, e.g., [131], [16], [86], [64], [23]). We have
already noted that if is singular, then we may no longer have the coherence
property of Lemma 6.2.2. To get this property, one needs to make some additional
assumption on the C-sequence C ( < + ), an assumption about the coherence
of the C-sequence. This will be a subject of some of the following chapters where
we will concentrate on a ner function instead of 1 .
We nish this section by exposing local versions of the maximal weight func-
tion. We start with a C-sequence C ( < )3 on a regular uncountable cardinal
and an innite cardinal < that is not necessarily its cardinal predecessor as
above. Let : []2 be dened by
(, ) = max{ C ( + 1) : divides tp(C )}, 4 (6.2.7)
3 As always, we are implicitly assuming that C+1 = {} for all < .
4 We say that divides an ordinal of = for some ordinal .
6.2. The weight function and its local versions 177

where we stipulate that max = 0 and that divides the ordinal 0. Using this
mapping we dene characteristic
1 : []2 (6.2.8)
by the recursive formula
1 (, ) = max{1 (, min(C \ )), |C [ (, ), )|} (6.2.9)
with the boundary value 1 (, ) = 0. Observe that when is a successor of
some cardinal , and if the C-sequence C ( < = + ) is chosen to have the
property that tp(C ) for all < = + , then 1 is nothing else than the
old characteristic 1 . The following unboundedness property is one of the main
reasons for introducing the local version of the maximal weight characteristic.
Theorem 6.2.7. Suppose that the C-sequence C ( < ) avoids5 a stationary set
and that tp(C ) for all . Then for every family A of size of
pairwise-disjoint subsets of such that cf() > |a| for all and a A, and
for every < , there is a subfamily B A of size such that for every pair
a = b in B, we have 1 (, ) > whenever a and b.
Proof. Choose a continuous -chain M of order-type consisting of elementary
submodels M of (H+ , ) such that M contains all the relevant objects and such
that M = M . Let
D = {M : M M}.
Then D is a closed and unbounded subset of , and so we can nd D.
Choose b A such that min(b) > . Fix b. Since our C-sequence avoids the
walk = 0 > 1 > > n1 >  n = from to passes through + 1, i.e.,
n1
n1 = + 1. In particular the set i=1 Ci is bounded in . Since cf() > |b|,
there is < such that for all b,
> max(C ) for all Tr(, ) \ {}. (6.2.10)
Moreover, we may assume that tp(C ) > . It follows that Tr(, ) for all
[, ) and b, and so in particular, since tp(C ) ,6
1 (, ) tp(C ) > for all [, ) and b. (6.2.11)
Let M M be such that M = . Then M satises the following sentence: For
every < there is b A with min(b) > such that 1 (, ) > for all [, )
and b. By elementarity of M this sentence is true, and so for each < ,
we can x b A satisfying it. Let E be the collection of all ordinals < such
that > and such that sup(b ) < for all < . Then E is a closed and
unbounded subset of and the family B = {b : E} satises the conclusion of
the theorem. 
5 We say that C ( < ) avoids if C = for all limit ordinals < .
6 and therefore, (, ) = min(C ) for all < .
178 Chapter 6. General Walks and Their Characteristics

6.3 Unboundedness of the number of steps


The purpose of this section is to isolate a condition on given C-sequences

C ( < )

dened on a regular uncountable cardinal that would correspond to the require-


ment that the corresponding function 2 : []2 , dened recursively by the
formulas
2 (, ) = 2 (, min(C \ )) + 1 and
2 (, ) = 0,
is in some sense nontrivial, and in particular, far from being constant. Without
doubt the C-sequence C = ( < ) is the most trivial choice and the corre-
sponding 2 -function, being constantly equal to 1, gives no information whatsoever
about the cardinal . The following notion of the triviality of a C-sequence on
appears to be only marginally dierent from this one, but we will soon see that it
actually captures a natural unboundedness property of the characteristic 2 that
counts the number of steps for the corresponding notion of walk.
Denition 6.3.1. A C-sequence C ( < ) on a regular uncountable cardinal
is trivial if there is a closed and unbounded set C such that for every <
there is with C C .
Theorem 6.3.2. The following are equivalent for any C-sequence C ( < ) on a
regular uncountable cardinal and the corresponding function 2 :
(i) C ( < ) is nontrivial.
(ii) For every family A of pairwise-disjoint nite subsets of and every integer
n, there is a subfamily B of A of size such that 2 (, ) > n for all a,
b and a = b in B.
Proof. Note that if a closed and unbounded set C witnesses the triviality of
the C-sequence, and if for < we let () be the minimal such that
C C() , then any disjoint subfamily of A = {{, ()} : C} of size
violates (ii) for n = 1.
Conversely, assume we are given a nontrivial C-sequence C ( < ). Con-
sider the following statement:
(iii) For every pair and of unbounded subsets of and every integer n, there
exist a tail subset 0 of , an unbounded subset 0 of and a regressive
strictly increasing map f : 0 such that 2 (, ) > n for all 0
and 0 with < f ().
To get (ii) from (iii) one assumes that the family A consists of nite sets of some
xed size m and performs m2 successive applications of (iii): this will give us an
unbounded A A and regressive strictly increasing maps fij : j [A]
(i,j < m),
6.3. Unboundedness of the number of steps 179

where j is the projection mapping that takes b to the jth member of b. We have
then that for all a, b A and i, j < m:
if max(a) < fij (j (b)), then 2 (i (a), j (b)) > n.
Going from there, we get a B as in (ii) by thinning out A once again.
Let us prove (iii) from (i) by induction on n. So suppose and are two
unbounded subsets of and that (iii) is true for n 1. Let M ( < ) be a
continuous -chain of submodels of H+ containing all relevant objects such that
= M . Let
C = { < : = }.
Then C is a closed and unbounded subset of , so by our assumption (i) on
C ( < ) there is C such that C  C for all . Pick an arbitrary
above . Then there is C such that
/ C . Then = sup(C ) < .
Using the elementarity of the submodel M we conclude that for every \
there is () \ ( + 1) such that sup(C() ) = . Applying the inductive
hypothesis to the two sets
\
and {min(C() \ ) : \
}
we get a tail subset 0 of \
, an unbounded subset 1 of \
and a strictly
increasing map g : 1 such that g() < min(C() \ ) and
2 (, min(C() \ )) > n 1 for all 0 , 1 with < g(). (6.3.1)
Let 0 = {() : 1 } and f : 0 be dened by
f (()) = min{, g()}.
Then for 0 and 1 with < f (()) < () we have
2 (, ()) = 2 (, min(C() \ )) + 1
= 2 (, min(C() \ )) + 1
> (n 1) + 1 = n.
This nishes the proof. 
The following consequence of Theorem 6.3.2 shows that no tree of the form T (0 )
can be Souslin.
Corollary 6.3.3. Suppose that C ( < ) is a nontrivial C-sequence, let
0 : []2 <
be the full code of the walk along this C-sequence, and let
T (0 ) = {0 (, )  : < + }
be the corresponding tree. Then every subset of T (0 ) of size contains an an-
tichain of size .
180 Chapter 6. General Walks and Their Characteristics

Proof. Consider a subset X of T (0 ) of size . Clearly, we may assume that X


takes at most one point from a given level of T (0 ). Replacing X by a set lying
inside its downward closure, we may assume that the set consists of successor
nodes of T (0 ). Let K []2 be such that

X = {0 (, )  ( + 1) : {, } K}.

Shrinking X, we may assume that K consists of pairwise-disjoint pairs. Shrinking


X further, we may also assume that 2 is constant on K. Let n be the constant
value of 2  K. Applying Theorem 6.3.2(ii) to K and n, we get K0 K of size
such that 2 (, ) > n for all {, } and {, } from K0 with properties < ,
< and < . Then

X0 = {0 (, )  ( + 1) : {, } K0 }

is an antichain in T (0 ). 
Remark 6.3.4. It should be clear that nontrivial C-sequences exist on any suc-
cessor cardinal. Indeed, with very little extra work one can show that nontrivial
C-sequences exist for some inaccessible cardinals quite high in the Mahlo-hierarchy.
To show how close this is to the notion of weak compactness, we will give the fol-
lowing characterization of it which is of independent interest.7
Theorem 6.3.5. The following are equivalent for an inaccessible cardinal :
(i) is weakly compact.
(ii) For every C-sequence C ( < ) there is a closed and unbounded set C
such that for all < there is such that C = C .

Proof. To see the implication from (i) to (ii), let us assume that C ( < ) is a
C-sequence and let 0 : []2 Q be the corresponding 0 -function. Set

T (0 ) = {(0 )  : < }.

Since is inaccessible, the levels of T (0 ) have size < , so by the tree-property


of (see [55]) the tree T (0 ) must contain a conal branch b. For each < x
() such that the restriction of (0 )() to belongs to b. For limit <
let () be the largest member appearing in the walk

() = 0 () > 1 () > > k() () =

along the C-sequence with the property that C is unbounded in . (Thus


= or = k()1 ().) By the Pressing Down Lemma there is a stationary set
7 It turns out that every C-sequence on being trivial is not quite as strong as the weak com-

pactness of . As pointed out to us by Donder and K onig, one can show this using a model of
Kunen [63, 3].
6.3. Unboundedness of the number of steps 181

and an integer k, a sequence t in Q and an ordinal


< such that for all
:
k() = k and sup(Ci () )
for all i < k for which (6.3.2)
this set is bounded in .
tp(Ci () ) = t(i) for all i < k for which this set is (6.3.3)
bounded in .
Shrinking we may assume that either for all , () = , in which case t
is chosen to be a k-sequence, or for all , () = k1 (), in which case t is
chosen to be a (k 1)-sequence. It follows that

0 (, ()) = t 0 (, ()) for all and [


, ). (6.3.4)

It follows that for <  in the mappings (0 )() and (0 )( ) have the same
restrictions on the interval [
, ). So in particular we have that

C() [ , ) whenever <  in .


, ) = C( ) [ (6.3.5)

Going to a stationary subset of we may assume that all C() ( ) have the
same intersection with . It is now clear that the union of C() ( ) is a
closed and unbounded subset of satisfying the conclusion of (ii).
To prove the implication from (ii) to (i), let T = (, <T ) be a tree on
whose levels are intervals of , or more precisely, there is a closed and unbounded
set of limit ordinals in such that if ( < ) is its increasing enumeration,
then [ , +1 ) is equal to the th level of T . Choose a C-sequence C ( < )
as follows. First of all let C+1 = {} and C = ( , ) if is the limit and
= sup( ) < . If is a limit point of , let

C = { < : <T }.

By (ii) there is a closed and unbounded set C such that for all < there is
() with C() = C . Let be the set of all successor ordinals <
such that C [ , +1 ) = . For let t = min(C [ , +1 )). Then it is seen
that {t : } is a chain of T of size . This nishes the proof. 

We have already remarked that every successor cardinal = + admits a


nontrivial C-sequence C ( < ). It suces to take the C s to be all of order-
type . It turns out that for such a C-sequence the corresponding 2 -function
has a property that is considerably stronger than 6.3.2(ii).
Theorem 6.3.6. For every innite cardinal there is a C-sequence on + such
that the corresponding 2 -function has the following unboundedness property: for
every family A of + pairwise-disjoint subsets of + , all of size < , and for every
n < , there exists B A of size + such that 2 (, ) > n whenever a and
b for some a = b in B.
182 Chapter 6. General Walks and Their Characteristics

Proof. We shall show that this is true for every C-sequence C ( < + ) with the
property that tp(C ) for all < + . Let us rst consider the case when
is a regular cardinal. The proof is by induction on n. Suppose the conclusion is
true for some integer n and let A be a given family of + pairwise-disjoint subsets
of + , all of size < + . Let M ( < + ) be a continuous -chain of elementary
submodels of H++ such that for every the M contains all the relevant objects
and has the property that = M + + . Let C = { : < + }. Then C
is a closed and unbounded subset of + . Choose = C of conality and
b A such that > for all b. Then
= sup{max(C ) : b} < .
By the elementarity of M we conclude that for every (, + ) there is b in A
above such that
= sup{max(C ) : b }.
Consider the following family of subsets of + ,
= b {min(C \ ) : b } ( (, + )).
a
By the inductive hypothesis there is an unbounded + such that for all <
in , b and 2 (, ) > n for all a
and a
. Let B = {b : } and
let b and b for < be two given members of B. Then
2 (, ) = 2 (, min(C \ )) + 1 n + 1 for all b and b .
This completes the inductive step and therefore the proof of the theorem when
is a regular cardinal.
If is a singular cardinal, going to a subfamily of A, we may assume that
there is a regular cardinal < and an unbounded set + with the property
that A can be enumerated as a ( ) such that for all :
|a | < , (6.3.6)
> for all a , (6.3.7)
|C | < for all a . (6.3.8)
Choose = C of conality and above and proceed as in the previous
case. This completes the inductive step and the proof of Theorem 6.3.6. 
Note that the combination of the ideas from the proofs of Theorem 6.3.2 and
6.3.6 gives us the following variation.
Theorem 6.3.7. Suppose that a regular uncountable cardinal supports a nontrivial
C-sequence and let 2 be the associated function. Then for every integer n and
every pair of -sized families A0 and A1 , where the members of A0 are pairwise-
disjoint bounded subsets of and the members of A1 are pairwise-disjoint nite
subsets of , there exist B0 A0 and B1 A1 of size such that for all a B0
and b B1 such that sup(a) < min(b), we have that 2 (, ) > n for every choice
of a and b. 
6.3. Unboundedness of the number of steps 183

Let us give an interesting application of this result found recently by Gruen-


hage [41]. Recall that for every perfect8 surjection f : X Y from a topological
space X onto a topological space Y there is a closed subset H of X on which f
is irreducible 9 . The set H is obtained as the intersection of a maximal chain C of
closed subsets of X such that f  K = Y for all K C. When the map f is closed
but not necessarily perfect, such a set may not exist, in which case we associate
to it the following natural cardinal characteristic,

(f ) = min{otp(C) : C ICf (X) and f  ( C) = Y }, (6.3.9)
where ICf (X) denotes the collection of all closed subspaces H of X such that
f  H = Y and |f 1 (y) H| = 1 for every isolated point y of Y . If the set on the
right-hand side of (6.3.9) is empty, which happens for example in the case when
the map f is perfect, set (f ) = . Clearly (f ) is a regular cardinal unless it is
equal to . The following result says that mild restrictions on spaces X and Y
force (f ) to be rather large, or in other words that when X and Y are relatively
small, then every closed surjection f : X Y will have an irreducible restriction
f : H Y on some closed subspace H of X.
Theorem 6.3.8. Let f : X Y be a given closed surjection from a topological
space X onto a topological space Y which has no irreducible restriction on a closed
subspace H of X. Assume that the space X has the property that every open cover
of X has a point-countable open renement and that the space Y has the property
that every open subspace of Y either contains an isolated point or a sequence of
open sets whose intersection is not open. Then the corresponding cardinal (f ) is
uncountable and it does not support a nontrivial C-sequence.
Proof. Clearly, (f ) < . Let us rst show that (f ) = . Otherwise, there is a
decreasing sequence Hn (n < ) of closedsubspaces of X such that f  Hn = Y
for all n, but f  H = Y , where H = n< Hn . Moreover, we are assuming
that |f 1 (y) H0 | = 1 for every isolated point y of Y , and therefore the set
U = Y \ f  H is open, nonempty, and has no isolated points. By our assumption
on Y, there is a decreasing sequence
 Un (n < ) of open subsets of U and a point
y n< Un such that y n< Kn , where Kn = Y \ Un for n < . Let

G= (Hn f 1 (Kn )).
n<

Note that f G = n< Kn , and therefore, y f  G = f  G. Pick x G f 1 (y).


Then for every n < ,



x (Hm f 1 Km ),
mn
8 Recall that a continuous surjection f : X Y is perfect if it is closed (i.e., it maps closed
subsets of X onto closed subsets of Y ) and it has the property that f 1 (y) is compact for all y
in Y .
9 Recall that a surjection f : H Y is irreducible if f  K = Y for every proper closed subset

K of H.
184 Chapter 6. General Walks and Their Characteristics

and therefore, x Hn . It follows that x H f 1 (y), and so, y = f (x) f  H ,


a contradiction.
It follows that = (f ) is a regular uncountable cardinal. We need to show
that does not support a nontrivial C-sequence. Suppose it does; x such a se-
quence C ( < ), and consider the corresponding characteristic

2 : []2

that counts the number of steps of the walk along C ( < ), i.e., a characteristic
dened by the recursive formula

2 (, ) = 2 (, min(C \ )) + 1

with the boundary value 2 (, ) = 0 for all . Choose a decreasing sequence H


 
 subsets of X such that f H = Y for1all < , but f H = Y ,
( < ) of closed
where H = < H . Moreover we assume that |f (y) H0 | = 1 for every
isolated point y of Y , and therefore the set Y \ f  H is open, nonempty, and
has no isolated points. So, as before, on the basis of our assumption on Y , we
can nd an increasing
 sequence Kn (n < ) of closed subsets of Y and a point
y n< Kn \ n< Kn . By our assumption on Y there is a point-countable open
cover V ( < ) of the closed set f 1 (y) such that V X \ H for all < .
For < and n < , set
Vn = V \ f 1 (Kn ). (6.3.10)

Then for every < , the open set < V2 {,} 10 covers the set f 1 (y), and
since the mapping f is closed, for each such ordinal < , we can select a neigh-
borhood W of y such that

f 1 (W ) V2 {,} . (6.3.11)
<

Using again the fact that f is a closed mapping, we can nd an open neighborhood
U of y such that 
f 1 (U ) (V f 1 (W )). (6.3.12)
<

Fix an n0 < such that Kn0 U = and pick a point z in this set. For x f 1 (z),
set
ax = { < : x V }.
Then ax (x f 1 (z)) is a family of countable subsets of . Note that ax > for
all < and x H . So one can nd a subset S f 1 (z) of cardinality such
that ax (x S) is a family of pairwise-disjoint countable subsets of . Let

B = { < : z W }.
10 Here, 2 {, } = 2 (, ) if < , 2 {, } = 2 (, ) if > , and 2 {, } = 0 if = .
6.3. Unboundedness of the number of steps 185

We claim that B is an unbounded subset of . Pick < . Then f 1 (z) H = ,


so we can x a point x0 in this set. Then

x0 f 1 (U ) (V f 1 (W )).
<

So we can nd < such that x0 V f 1 (W ). Then z = f (x0 ) W . Since


x0 V and since V H = for > , we conclude that is a member of
B \ , as required.
Applying Theorem 6.3.7 to the family ax (x S), the set B, and the integer
n0 , we can nd x S and B such that ax < and

2 (, ) > n0 for all ax .

On the other hand, f (x) = z Kn0 , and so x


/ Vn for all n n0 . It follows that

x/ V2 {,} ,
<

and so on the basis of (6.3.11), we conclude that z = f (x)


/ W , a contradiction.

Corollary 6.3.9. Let X and Y be two topological spaces that are not bigger than
the rst regular uncountable cardinal not supporting a nontrivial C-sequence.
Suppose that X every open cover of X has a point-countable open renement and
that every open subspace of Y either contains an isolated point or a sequence of
open sets whose intersection is not open. Then every closed surjection f : X Y
is irreducible when restricted on some closed subspace H of X. 
Thus, in particular, if X and Y are smaller than, say, the rst -Mahlo
cardinal, then every closed surjection f : X Y has an irreducible restriction
on a closed subspace of X.
Remark 6.3.10. As shown in [41] there is a closed surjection f : X Y for
a Lindel
of space X which admits no irreducible restriction. It turns out that for
that closed surjection f , we have (f ) = , so the restriction on the space Y in
Theorem 6.3.8, needed to eliminate this case, is in some sense necessary.
Chapter 7

Square Sequences

7.1 Square sequences and their full lower traces


The purpose of this section is to study walks along C-sequences that have the
following pleasant coherence property.
Denition 7.1.1. A C-sequence C ( < ) is a square sequence if and only if it is
coherent, i.e., if it has the property that C = C whenever is a limit point
of C .
Note that the nontriviality conditions appearing in Denition 6.3.1 and The-
orem 6.3.5 coincide in the realm of square sequences:
Lemma 7.1.2. A square sequence C ( < ) is trivial if and only if there is a
closed and unbounded subset C of such that C = C whenever is a limit
point of C. 
To a given square sequence C ( < ) one naturally associates a tree-
ordering <2 on as follows:

<2 if and only if is a limit point of C . (7.1.1)

The triviality of C ( < ) is then equivalent to the statement that the tree
(, <2 ) has a chain of size . In fact, one can characterize the tree-orderings <T on
for which there exists a square sequence C ( < ) such that for all < < :

<T if and only if is a limit point of C . (7.1.2)

Lemma 7.1.3. A tree-ordering <T on admits a square sequence C ( < )


satisfying (7.1.2) if and only if
(i) <T can hold only for limit ordinals and such that < ,
188 Chapter 7. Square Sequences

(ii) P = { : <T } is a closed subset of , which is unbounded in whenever


cf() > and
(iii) minimal as well as successor nodes of the tree <T on are ordinals of co-
nality .
Proof. For each ordinal < of countable conality we x a subset S of
order-type conal with . Given a tree-ordering <T on with properties (i)(iii),
for a limit ordinal < let P+ be the set of all successor nodes from P {}
including the minimal one. For P+ , let be its immediate predecessor in
P . Finally, set

C = P {S [ , ) : P+ }.
It is easily checked that this denes a square sequence C ( < ) with the
property that <T holds if and only if is a limit point of C . 
Remark 7.1.4. It should be clear that the proof of Lemma 7.1.3 shows that the
exact analogue of this result is true for any conality < rather than .
An important result about square sequences is the following fact that can be
deduced on the basis of the well-known construction of square sequences in the
constructible universe, a construction that is essentially contained in the classical
paper of Jensen [52] (see also 1.10 of [109]).
Theorem 7.1.5. If a regular uncountable cardinal is not weakly compact in the
constructible subuniverse, then there is a nontrivial square sequence on which is
moreover constructible. 
Corollary 7.1.6. If a regular uncountable cardinal is not weakly compact in the
constructible subuniverse, then there is a constructible Aronszajn tree on .
Proof. Let C ( < ) be a xed nontrivial square sequence which is constructible.
Changing the C s a bit, we may assume that if is a limit ordinal with =
min C , or if C but sup(C ) < , then must be a successor ordinal in
. Consider the corresponding characteristic

0 : []2 <

of the walk along C ( < ) dened by the recursive formula

0 (, ) = tp(C ) 0 (, min(C \ )),

where 0 (, ) = for all < . Consider also the corresponding tree

T (0 ) = {(0 )  : < }.

Clearly T (0 ) is constructible. By (6.1.4) the th level of T (0 ) is bounded by


the size of the set {C : }. Since the intersection of the form C is
7.1. Square sequences and their full lower traces 189

determined by its maximal limit point modulo a nite subset of , we conclude


that the th level of T (0 ) has size || + 0 . Since the sequence C ( < ) is
nontrivial, the proof of Theorem 6.3.5 shows that T (0 ) has no conal branches.

Lemma 7.1.7. Suppose C ( < ) is a square sequence on , <2 the associated
tree-ordering on . Let 0 : []2 < be the full code of the walk along C ( <
), and let T (0 ) = {(0 )  : < } be the associated tree. Then 
(0 ) is a strictly increasing map from the tree (, <2 ) into the tree T (0 ).
Proof. If is a limit point of C , then C = C , so the walks and
for < get the same code 0 (, ) = 0 (, ). 
Having explained the nontriviality condition on a given square sequence, for
the rest of this section, we x a regular uncountable cardinal and a nontrivial
square sequence C ( < ) on and study walks along C ( < ). In particular,
we will be interested in the full lower trace associated to these walks. For this it
is convenient to introduce the function
= : []2
given by the formula
(, ) = maximal limit point of C ( + 1),
where we stipulate that (, ) = 0 if C ( + 1) has no limit points.
Having this function, we can write a recursive trace formula describing the
full lower trace characteristic
F : []2 []< (7.1.3)
of the walk along C ( < ) as follows:

F(, ) = F(, min(C \ )) {F(, ) : C [(, ), )}, (7.1.4)

with the boundary value F(, ) = {} for all .


Lemma 7.1.8. For all < ,
(a) F(, ) F(, ) F(, ),
(b) F(, ) F(, ) F(, ).
Proof. The proof of the lemma is by simultaneous induction. We rst check (a).
Let
= min(C \ ), = min(C \ ) and = (, ). (7.1.5)
If < , then C = C and therefore F(, ) = F(, ). Applying the inductive
hypothesis (b) for < , we get the required conclusion
F(, ) = F(, ) F(, ) F(, ) F(, ) F(, ). (7.1.6)
190 Chapter 7. Square Sequences

So we may assume that . Then (, ) = . If < , applying the inductive


hypothesis (b) for < , we have

F(, ) F(, ) F( , ) F(, ) F(, ), (7.1.7)

since C [(, ), ). If , then = , and applying the inductive


hypothesis (a) for < = , we have

F(, ) F(, ) F(, ) F(, ) F(, ). (7.1.8)

Consider now C [, ). By the inductive hypothesis (b) for < we


have
F(, ) F(, ) F(, ) F(, ) F(, ), (7.1.9)
since C [(, ), ). It follows that all the factors of F(, ) are included in
the union F(, ) F(, ), so this completes the checking of (a).
To prove (b) dene , and as above (see 7.1.5) and consider rst the
case . Then C = C and therefore F(, ) = F(, ). Applying the
inductive hypothesis (a) for < we get

F(, ) F(, ) F(, ) F(, ) F(, ). (7.1.10)

So we may assume that < in which case we also know that (, ) = .


Consider rst the case < . Applying the inductive hypothesis (a) for
< we have

F(, ) F(, ) F( , ) F(, ) F(, ), (7.1.11)

since C [(, ), ). Suppose now that . Then = and applying


the inductive hypothesis (b) for the triple = we have

F(, ) F(, ) F(, ) F(, ) F(, ). (7.1.12)

This completes the proof. 


Lemma 7.1.9. For all < ,
(a) 0 (, ) = 0 (min(F(, ) \ ), ) 0 (, min(F(, ) \ )),
(b) 0 (, ) = 0 (min(F(, ) \ ), ) 0 (, min(F(, ) \ )).

Proof. The proof is by induction on , and . Let

= (, ), 1 = min(C \ ) and 1 = min(F(, ) \ ). (7.1.13)

Pick {min(C \ )} (C [, )) such that 1 F(, ). Then

1 = min(F(, ) \ ), (7.1.14)
7.1. Square sequences and their full lower traces 191

so applying the inductive hypothesis (b) for < we get

0 (, ) = 0 (1 , ) 0 (, 1 ) (7.1.15)

which checks (a) for . Suppose rst that > . Note that F(, ) is a
subset of F(, ), and therefore,

2 = min(F(, ) \ ) 1 . (7.1.16)

Applying the inductive hypothesis for , we get

0 (, ) = 0 (2 , ) 0 (, 2 ), (7.1.17)

0 (, ) = 0 (2 , ) 0 (, 2 ). (7.1.18)
Combining (7.1.15) and (7.1.17), we infer that 0 (, 1 ) is a tail of 0 (, 2 ), so
(7.1.18) can be written as

0 (, ) = 0 (2 , ) 0 (1 , 2 ) 0 (, 1 ) = 0 (1 , ) 0 (, 1 ). (7.1.19)

Since is a limit point of C , we have 0 (, ) = 0 (, ), so (7.1.19) gives us the


conclusion (b) of Lemma 7.1.9.
Consider now the case . Then 1 is either equal to min(C \ ) or it
belongs to [, ) C . In any case we have that F{1 , } F(, ), so

2 = min(F{1 , } \ ) 1 . (7.1.20)

Applying the inductive hypothesis for , 1 and we get

0 (, ) = 0 (2 , ) 0 (, 2 ), (7.1.21)

0 (, 1 ) = 0 (2 , 1 ) 0 (, 2 ). (7.1.22)
Combining (7.1.15) and (7.1.21) we see that 0 (, 1 ) is a tail of 0 (, 2 ), so
(7.1.22) can be rewritten as

0 (, 1 ) = 0 (2 , 1 ) 0 (1 , 2 ) 0 (, 1 ) = 0 (1 , 1 ) 0 (, 1 ). (7.1.23)

Since 0 (, ) = tp(C ) 0 (, 1 ), (7.1.23) gives us the conclusion (b) of


Lemma 7.1.9. This nishes the proof. 
Recall the function 2 : []2 which counts the number of steps in the
walk along the xed C-sequence C ( < ) which in this section is assumed to
be moreover a square sequence:

2 (, ) = 2 (, min(C \ )) + 1,

where we let 2 (, ) = 0 for all . Thus 2 (, ) + 1 is simply equal to the


cardinality of the trace Tr(, ) of the minimal walk from to .
192 Chapter 7. Square Sequences

Lemma 7.1.10. sup< |2 (, ) 2 (, )| < for all < < .

Proof. By Lemma 7.1.9,

sup< |2 (, ) 2 (, )| supF(,) |2 (, ) 2 (, )|. 

Denition 7.1.11. Set I to be the set of all countable such that


sup 2 (, ) = for all < and innite .
Lemma 7.1.12. I is a P-ideal of countable subsets of .

Proof. Let n (n < ) be a given sequence of members of I and x < such


that n for all n. For n < , set n = { n : 
2 (, ) n}. Since n
belongs to I, n is a conite subset of n . Let = n< n . Then is a
member of I such that n \ is nite for all n. 

Theorem 7.1.13. The P-ideal dichotomy implies that a nontrivial square sequence
can exist only on = 1 .

Proof. Applying the P-ideal dichotomy to the ideal I of Denition 7.1.11, we get
the following two alternatives (see Denition 2.4.11):

there is an uncountable such that [] I, or (7.1.24)



there is a partition = n< n such that n I for all n. (7.1.25)

By Lemma 7.1.10, if (7.1.24) holds, then must be countable for all <
and so the conality of must be equal to 1 . Since we are working only with
regular uncountable cardinals, we see that (7.1.24) gives us that = 1 must hold.
Suppose now (7.1.25) holds and pick k < such that k is unbounded in . Since
k I we have that (2 ) is bounded on k for all < . So there is an
unbounded set and an integer n such that for each , the restriction of
(2 ) on k is bounded by n. By Theorem 6.3.2 we conclude that the square
sequence C ( < ) we started with must be trivial. 

Denition 7.1.14. By S we denote the sequential fan with edges, i.e., the space
on ( ){} with as the only non-isolated point, while a typical neighborhood
of has the form Uf = {(, n) : n f ()} {} where f : .
Theorem 7.1.15. If there is a nontrivial square sequence on , then the square of
the sequential fan S has tightness 1 equal to .

The proof will be given after a sequence of denitions and lemmas.


1 Thetightness of a point x in a space X is equal to if is the minimal cardinal such that, if a
set W X \ {x} accumulates to x, then there is a subset of W of size that accumulates to x.
7.1. Square sequences and their full lower traces 193

Denition 7.1.16. Given a square sequence C ( < ) and the corresponding


function 2 : []2 that counts the number of steps of the minimal walks, we
dene d : []2 by letting

d(, ) = sup |2 (, ) 2 (, )|.


Lemma 7.1.17. For all < ,


(a) d(, ) 2 (, ),
(b) d(, ) d(, ) + d(, ),
(c) d(, ) d(, ) + d(, ).
Proof. The conclusion (a) follows from the fact that we allow = in the def-
inition of d(, ). The conclusions (b) and (c) are consequences of the triangle
inequalities of the  -norm and the fact that in both inequalities we have that the
domain of functions on the left-hand side is included in the domain of functions
on the right-hand side. 
Denition 7.1.18. For , let

W = {((, d(, )), (, d(, ))) : < < }.

The following lemma establishes that the tightness of the point (, ) of S2


is equal to , giving us the proof of Theorem 7.1.15.
Lemma 7.1.19. (, ) W but (, )
/W for all < .

Proof. To see that W accumulates to (, ), let Uf2 be a given neighborhood of


(, ). Fix an unbounded set on which f is constant. By Theorem 6.3.2 and
Lemma 7.1.17(a) there exist < in such that d(, ) f () = f (). Then
((, d(, )), (, d(, ))) belongs to the intersection W Uf2 . To see that for a
given < the set W does not accumulate to (, ), choose g : such that

g() = 2d(, ) + 1 for < .

Suppose W Ug2 is nonempty and choose ((, d(, )), (, d(, ))) from this set.
Then
d(, ) 2d(, ) + 1 and d(, ) 2d(, ) + 1. (7.1.26)
It follows that d(, ) d(, ) + d(, ) + 1 contradicting Lemma 7.1.17(c). 
Since = 1 admits a nontrivial square sequence, Theorem 7.1.15 leads to
the following result of Gruenhage and Tanaka [42].
Corollary 7.1.20. The square of the sequential fan with 1 edges is not countably
tight. 
Question 7.1.21. What is the tightness of the square of the sequential fan with 2
edges?
194 Chapter 7. Square Sequences

Corollary 7.1.22. If a regular uncountable cardinal is not weakly compact in the


constructible subuniverse, then the square of the sequential fan with edges has
tightness equal to . 
The way we have proved that S2 has tightness equal to is of independent
interest. We have found a point (, ) of S2 and a set W of isolated points which
accumulates to (, ) but has no other accumulation points and moreover, no
subset of W of size < has accumulation points at all. This is interesting in
view of the following result of Dow and Watson [24], where C denotes the category
of spaces that contain the converging sequence + 1 and is closed under nite
products, quotients and arbitrary topological sums.
Theorem 7.1.23. If for every regular cardinal there is a space X in C which has
a set of isolated points of size witnessing tightness of its unique accumulation
point in X, then every space belongs to C. 
Theorem 7.1.24. If every regular uncountable cardinal supports a nontrivial square
sequence, then every topological space X can be obtained from the converging se-
quence after nitely many steps of taking (nite) products, quotients, and arbitrary
topological sums.
Proof. This follows from combining (the proof of) Theorem 7.1.15 and Theorem
7.1.23. 
If is a strongly compact cardinal, then the tightness of the product of every pair of
spaces of tightness < is < and so C contains no space of tightness or larger.
It has been shown in [24] that if the Lebesgue measure extends to a countably
additive measure on all sets of reals, then there even is a countable space which
does not belong to C. On the other hand, by Theorem 7.1.5 and Theorem 7.1.24, if
no regular cardinal above 1 is weakly compact in the constructible subuniverse,
then every space can be generated by the converging sequence by taking nite
products, quotients and arbitrary topological sums.
Remark 7.1.25. We have seen that the case = 1 is quite special when one
considers the problem of existence of various nontrivial square sequences on . It
should be noted that a similar result about the problem of the tightness of S2 is
not available. In particular, it is not known whether the P-ideal dichotomy or a
similar consistent hypothesis of set theory implies that the tightness of the square
of S2 is smaller than 2 . It is interesting that considerably more is known about
the dual question, the question of initial compactness of the Tychono cube N .
For example, if one denes B = {f N : f (), f () d(, )} ( < < )
one gets an open cover of N without a subcover of size < . However, for small
such as = 2 one is able to nd such a cover of N without any additional
set-theoretic assumptions and in particular without the assumption that carries
a nontrivial square sequence.
7.2. Square sequences and local versions of 195

7.2 Square sequences and local versions of


In this section, we give the analysis of a family of -functions associated with a
(nontrivial) square sequence C ( < ) that lives on some regular uncountable
cardinal . Recall that an ordinal divides an ordinal if there is such that
= , i.e., can be written as the union of an increasing -sequence of intervals
of type . Let < be a xed innite cardinal. Let : []2 be dened by

(, ) = max{ C ( + 1) : divides tp(C )}, (7.2.1)

where we stipulate that max = 0 and that divides the ordinal 0. This translates
into the requirements that

(, ) = 0 if C = and (, ) = min(C ) if tp(C ) . (7.2.2)

Finally we are ready to dene the main object of study in this section:

: []2 (7.2.3)

dened recursively by

(, ) = sup{tp(C [ (, ), )), (, min(C \ )), (, ) :


C [ (, ), )},

with the boundary condition (, ) = 0 for all .


The following consequence of the coherence property of C ( < ) will be
quite useful.
Lemma 7.2.1. If is a limit point of C , then (,) = (,) for all < . 
Note that is something that corresponds to the function : [1 ]2
considered in Denition 3.1.1 (see also Section 7.3) and that the s are simply
local versions of the key denition. We shall show that they all have the crucial
subadditivity properties.
Lemma 7.2.2. If < < < , then
(a) (, ) max{ (, ), (, )},
(b) (, ) max{ (, ), (, )}.
Proof. The proof is by induction on , and . We rst prove the inductive step
for (a). Let = max{ (, ), (, )}. We need to show that (, ) . To
this end, let
= min(C \ ) and = min(C \ ). (7.2.4)
Case 1a : < (, ). Then by Lemma 7.2.1, (, ) = (, (, )). By the
denition of (, ) we have

( (, ), ) (, ) . (7.2.5)
196 Chapter 7. Square Sequences

Applying the inductive hypothesis (b) for the triple < (, ) we conclude
that (, (, )) .
Case 2a : (, ). Then (, ) = (, ) = .
Subcase 2a .1: = = . By the inductive hypothesis (a) for the triple <
we get
(, ) max{ (, ), (, )} , (7.2.6)
since (, ) (, ). Note that [, )C is an initial part of [, )C so its
order-type is bounded by (, ) and therefore by . Consider a [, ) C .
Applying the inductive hypothesis (b) for the triple < < we conclude that

(, ) max{ (, ), (, )} , (7.2.7)

since (, ) (, ) by the denition of (, ). This shows that all ordinals


appearing in the formula for (, ) are and so (, ) .
Subcase 2a .2: < . Then belongs to C [, ) and therefore ( , )
(, ) . Applying the inductive hypothesis (b) for < gives us

(, ) max{ (, ), ( , )} . (7.2.8)

Given C [, ) using the inductive hypothesis (b) for < < we get that

(, ) max{ (, ), (, )} , (7.2.9)

since (, ) (, ). Since tp(C [, )) tp(C [, )) (, )


we see again that all the ordinals appearing in the formula for (, ) are .
Let us now concentrate on the inductive step for (b). This time let =
max{ (, ), (, )} and let and be as in (7.2.4). We need to show that
(, ) .
Case 1b : < (, ). By Lemma 7.2.1,

(, (, )) = (, ) . (7.2.10)

From the formula for (, ) we see that ( (, ), ) (, ) . Applying


the inductive hypothesis (a) for the triple < (, ) we get

(, ) max{ (, (, )), ( (, ), )} . (7.2.11)

Case 2b : (, ).
Subcase 2b .1: = = . Then (, ) (, ) and (, )
(, ) . Applying the inductive hypothesis (b) for < < we get

(, ) max{ (, ), (, )} . (7.2.12)
7.2. Square sequences and local versions of 197

Subcase 2b .2: < . Then C [, ) and so ( , ) appears in the


formula for (, ) and so we conclude that ( , ) (, ) . Similarly,
(, ) (, ) . Applying the inductive hypothesis (a) for the triple
< we get

(, ) max{ (, ), ( , )} . (7.2.13)

This nishes the proof. 


The following is an immediate consequence of the fact that the denition of
is closely tied to the notion of a walk along the xed square sequence.
Lemma 7.2.3. (, ) (, ) whenever and belongs to the trace
of the walk from to . 
Lemma 7.2.4. Suppose < and that is a limit ordinal > 0. Then
(, ) (, ) for coboundedly many < .
Proof. Let = 0 > 1 > > n1 > n = be the trace of the walk from to
. Let = n1 if is a limit point of Cn1 , otherwise let = . Note that by
Lemma 7.2.1, in any case we have that

(, ) = (, ) for all < . (7.2.14)

Let < be an upper bound of all Ci (i < n) which are bounded in . Then
is a member of the trace of any walk from to some ordinal in the interval
). Applying Lemma 7.2.3 to this fact gives us
[,
).
(, ) (, ) for all [, (7.2.15)

Since (, ) = (, ) for all < (see (7.2.14)), this gives us the conclusion
of the lemma. 
Lemma 7.2.5. The set P () = { < : (, ) } is a closed subset of for
every < and < .
Proof. The proof is by induction on . So let < be a limit point of P (). Let
1 = min(C \ ) and let = (, ). If = , then (, ) = 0 and therefore
P (). So we may assume that < . Then

(, ) = for every [, ). (7.2.16)

Case 1: = sup(C ) < . Then 1 = min(C \ ) for every P () (


, ).
It follows that
P () (
, ) P (1 ), (7.2.17)
so is also a limit point of P (1 ). By the inductive hypothesis P (1 ), i.e.,
(, 1 ) . Since (, ) = for all [, ), we get

tp(C [, )) = sup{tp(C [, )) : P () [, )} . (7.2.18)


198 Chapter 7. Square Sequences

Consider C [, ). Then C [, ) for some P () [ , ) so by


the denition of (, ) we conclude that (, ) (, ) . By (7.2.17) we
also have (, 1 ) so by Lemma 7.2.2(a), (, 1 ) . Applying Lemma
7.2.2(b) to < < 1 , we conclude that

(, ) max{ (, 1 ), (, 1 )} . (7.2.19)

This shows that all the ordinals appearing in the formula for (, ) are and
therefore that (, ) .

Case 2: is a limit point of C . As in the previous case, to show that (, )


we need to have tp(C [, )), (, 1 ) and (, ) with C [, ) all
. Note that 1 = so (, 1 ) = 0. Similarly as in the rst case we conclude
tp(C [, )) . By Lemma 7.2.1,

(, ) = (, ) for all P () . (7.2.20)

Consider an C [, ). Since is a limit point of P (), there is from this set


above . By (7.2.16) we have that C [ (, ), ) so (, ) (, ) .
Applying 7.2.2(a) to < < we get

(, ) max{ (, ), (, )} . (7.2.21)

This nishes the proof. 

For < < and < , set

< if and only if (, ) . (7.2.22)

Lemma 7.2.6.

(1) < is a tree ordering on ,


(2) < < whenever < < ,

(3)  = < < .

Proof. This follows immediately from Lemma 7.2.2. 

Recall the notion of a special tree of height from Section 6.1, a tree T for
which one can nd a T -regressive map f : T T with the property that the
preimage of any point is the union of < antichains. By a tree on we mean a
tree of the form (, <T ) with the property that <T implies < .
Lemma 7.2.7. If a tree T on is special, then there is an ordinal-regressive map
f : and a closed and unbounded set C such that f is one-to-one on all
chains separated by C.
7.2. Square sequences and local versions of 199

Proof. Let g : be a T -regressive map such that for each < the preimage
g 1 () can be written as a union of a sequence A () ( < ) of antichains, where
< . Let C be the collection of all limits < with the property that <
for all < . Let ,  denote the Godel pairing function which in particular
has the property that ,  < for C and , , . Dene ordinal-regressive
map f : by letting f () = max(C ) if / C and for C, letting
f () = g(),  where < g() (< ) is minimal ordinal with the property that
A (g()). Then f is as required. 
By Lemma 7.2.6 we have a sequence < ( < ) of tree-orderings on . The
following lemma tells us that they are frequently quite large orderings.
Lemma 7.2.8. If > is not a successor of a cardinal of conality , then there
must be < such that (, < ) is a nonspecial tree on .
Proof. Suppose to the contrary that all trees are special. By Lemma 7.2.7 we may
choose ordinal-regressive maps f : for all < and a single closed and
unbounded set C such that each of the maps f is one-to-one on < -chains
separated by C. Using the Pressing Down Lemma, we nd a stationary set of
conality + ordinals < and < such that f () < for all and < .
If ||+ < , let = , = 0 and if ||+ = , represent as the increasing union
of a sequence ( < cf(||)) of sets of size < ||. Since = cf(||) there is a
< cf(||) and a stationary 0 such that for all 0 , f () for
many < . Let = . This gives us subsets and 0 of such that

||+ < and 0 is stationary in , (7.2.23)

= { < : f () } is unbounded in for all 0 . (7.2.24)


+ +
Let = || . Then < and so we can nd 0 such that 0 C
Then there will be 0 < and 1 0 C of size such that
has size .
(, ) 0 for all 1 . By (7.2.24) we can nd 2 1 of size and 1 0
such that f1 () for all 2 . Note that 2 is a <1 -chain separated by
C, so f1 is one-to-one on 2 . However, this gives us the desired contradiction
since the set , in which f1 embeds 2 has size smaller than the size of 2 . This
nishes the proof. 
It is now natural to ask the following question: under which assumption on
the square sequence C ( < ) can we conclude that neither of the trees (, < )
will have a branch of size ?
Lemma 7.2.9. If the set = { < : tp(C ) = } is stationary in , then none
of the trees (, < ) has a branch of size .
Proof. Assume that B is a < -branch of size . By Lemma 7.2.5, B is a closed
and unbounded subset of . Pick a limit point of B which belongs to . Pick
B such that tp(C ) > . By denition of (, ) we have that
200 Chapter 7. Square Sequences

(, ) tp(C ) > since clearly (, ) = 0. This contradicts the fact


that < and nishes the proof. 

Denition 7.2.10. A square sequence on is special if the corresponding tree (, <2 )


is special, i.e., there is a <2 -regressive map f : with the property that the
f -preimage of every < is the union of < antichains of (, <2 ).
Theorem 7.2.11. Suppose < are regular cardinals such that is not a successor
of a cardinal of conality . Then to every square sequence C ( < ) for which
there exist stationarily many such that tp C = , one can associate a sequence
C ( < , < ) such that:
(i) C C for all and ,

(ii) = < C for all limit ,
(iii) C ( < ) is a nonspecial (and nontrivial) square sequence on for all
< .
Proof. Fix < and dene C by induction on < . So suppose is a limit
ordinal < and that C is dened for all < . If P () is bounded in , let
be the maximal limit point of P () ( = 0 if the set has no limit points) and let

P () (C [max(P ()), )).



C = C (7.2.25)

If P () is unbounded in , let

C = P () {C : P () & = sup(P () )}. (7.2.26)

By Lemmas 7.2.2 and 7.2.5, C ( < ) is well dened and it forms a square
sequence on . The properties (i) and (ii) are also immediate. To see that for each
< the sequence C ( < ) is nontrivial, one uses Lemma 7.2.9 and the fact
that if is a limit point of C occupying a place in C that is divisible by ,
then < . By Lemma 7.2.8, or rather its proof, we conclude that there is <
such that C ( < ) is nonspecial for all . This nishes the proof. 
Lemma 7.2.12. For every pair of regular cardinals < , every special square
sequence C ( < ) can be rened to a special square sequence C ( < )
with the property that { < : tp C = } is stationary in and the sequence
C ( < ) avoids some stationary subset of this set.

Proof. Let <2 be the tree-ordering associated with C ( < ) and let f :
be a <2 -regressive map such that f 1 () is the union of < antichains of (, <2 )
for all < .
Case 1: = . By the Pressing Down Lemma there is a stationary set of
conality ordinals < on which f takes some constant value. Thus can be
decomposed into < antichains of (, <2 ) and so contains a stationary subset
7.2. Square sequences and local versions of 201

0 of pairwise <2 -incomparable ordinals. For 0 let C be any -sequence of


successor ordinals converging to . If is a limit ordinal not in 0 but there is (a
<2 in 0 , let C = C \
unique!) . In all other cases, let C = C . It is easily

seen that C ( < ) is as required.
Case 2: > . By Lemma 7.2.7 we may nd a closed and unbounded subset C
of and an ordinal-regressive map g : C which is one-to-one on <2 -chains
included in C (and which has the property that for C g() is the Godel pairing
of f () and the index of the antichain which contains in the xed antichain-
decomposition of the preimage f 1 (f ()). By the argument from Case 1 there is
a stationary set of limit points of C consisting of conality ordinals < such
that is an antichain of (, <2 ). Given , let h() be the minimal ordinal
< such that g() < for unboundedly many limit points of C that also
belong to C. Since g is regressive and since , the conality of , is uncountable,
the Pressing Down Lemma gives us that h() < . Choose a stationary set 0
such that h takes some constant value on 0 . Note that since g is one-to one on
<2 -chains included in C, the constant value cannot be a successor ordinal. It
follows that must be a limit ordinal of conality equal to . Let ( < ) be
a strictly increasing sequence which converges to .
We rst dene C for 0 as a strictly increasing continuous sequence
c () ( < ) of limit points of C that also belong to C and satisfy the following
requirements at each < :
c () = sup< c () with limit , (7.2.27)
c+1 () > c () is minimal subject to being a limit point
of C , belonging to C, and being > for every limit point (7.2.28)
of C such that C and g() < +1 .
If is a limit point of C for some 0 , set C = C . Note that by the
canonicity of the denition of c () ( < ) for 0 , we would get the same
result, as C = C for any other 0 for which is a limit point of C .
If is a limit ordinal which is not a limit point of any C for 0 and
there is (a unique!) <2 in 0 , let C = C \
.
If is a limit ordinal such that <2 for some 0 but is not a limit
point on any C for 0 , let C = C \ max(C ). Note again that there is
no ambiguity in this denition, as
C = C
holds for every and in 0 such that <2 and <2 . In all other cases
set C = C . It should be clear that C ( < ) is a nontrivial square sequence
satisfying the conclusion of Lemma 7.2.12. 
Before going further let us mention an unboundedness property of the func-
tion whenever it is based on a square sequence satisfying the conclusion of
Lemma 7.2.12
202 Chapter 7. Square Sequences

Theorem 7.2.13. Suppose that is a regular cardinal such that, for some innite
cardinal < , we base the mapping : []2 on a square sequence C
( < ) with the property that the set S = { < : tp(C ) = } is stationary in
and that the sequence C ( < ) avoids some stationary subset of S. Then for
every < and every family A of pairwise-disjoint subsets of that have sizes
< cf(), there exists B A of size such that for all a = b in B and all a,
b we have (, ) > .
Proof. The conclusion follows from Theorem 6.2.7 which claims the same thing
for the local version 1 of 1 , having in mind the obvious inequalities (, )
1 (, ) for < < . 
We are nally ready to state the main result of this section which follows
from Theorem 7.2.11 and Lemma 7.2.12.
Theorem 7.2.14. If regular uncountable cardinal = 1 carries a nontrivial square
sequence, then it also carries a nontrivial square sequence which is moreover non-
special. 
Corollary 7.2.15. If a regular uncountable cardinal = 1 is not weakly compact in
the constructible subuniverse, then there is a nonspecial Aronszajn tree of height .
Proof. By Theorem 7.1.5, carries a nontrivial square sequence C ( < ). By
Theorem 7.2.14, we may assume that the sequence is moreover nonspecial. Let 0
be the associated 0 -function and consider the tree T (0 ). As in Corollary 7.1.6 we
conclude that T (0 ) is an Aronszajn tree of height . By Lemma 7.1.7 there is a
strictly increasing map from (, <2 ) into T (0 ), so T (0 ) must be nonspecial. 
Remark 7.2.16. The assumption = 1 in Theorem 7.2.14 is essential as there is
always a nontrivial square sequence on 1 , but it is possible to have a situation
where all Aronszajn trees on 1 are special. For example MA1 implies this. In
[69], Laver and Shelah have shown that any model with a weakly compact cardinal
admits a forcing extension satisfying CH and the statement that all Aronszajn
trees on 2 are special. A well-known open problem in this area asks whether one
can have GCH rather than CH in a model where all Aronszajn trees on 2 are
special.

7.3 Special square sequence and the


corresponding function
The following well-known result of Jensen [52] supplements the corresponding
result for weakly compact cardinals listed above as Theorem 7.1.5.
Theorem 7.3.1. If a regular uncountable cardinal is not Mahlo in the constructible
subuniverse, then there is a special square sequence on which is moreover con-
structible. 
7.3. Special square sequence and the corresponding function 203

Today we know many more inner models with sucient amount of ne struc-
ture necessary for building special square sequences. So the existence of special
square sequences, especially at successors of strong-limit singular cardinals, is tied
to the existence of some other large cardinal axioms. The reader is referred to
standard textbooks in this area such as [55] for the specic information. In this
section we give the combinatorial analysis of walks along special square sequences
and the corresponding distance functions. Let us start by restating some results
of Section 7.2.
Theorem 7.3.2. Suppose < are regular cardinals and that carries a special
square sequence. Then there exist C ( < , < ) such that:
(1) C C for all and < ,

(2) = < C for all limit ,
(3) C ( < ) is a nontrivial square sequence on for all < .
Moreover, if is not a successor of a cardinal of conality , then each of the
square sequences can be chosen to be nonspecial. 
Theorem 7.3.3. Suppose < are regular cardinals and that carries a special
square sequence. Then there exist < ( < ) such that:
(i) < is a closed tree ordering of for each < ,

(ii)  = < < ,
(iii) no tree (, < ) has a chain of size . 
Let us now consider the case when is a successor cardinal, since in this case
some ner special square sequences exist.
Lemma 7.3.4. The following are equivalent when is a successor of some innite
cardinal :
(1) there is a special square sequence on ,
(2) there is a square sequence C ( < ) such that tp(C ) for all < .
Proof. Let D ( < + ) be a given special square sequence. By Lemma 6.1.2
the corresponding tree (+ , <2 ) can be decomposed into antichains. So, let
f : + be a xed map such that f 1 () is a <2 -antichain for all < .
Let < + be a given limit ordinal. If D has a maximal limit point < , let
C = D \ . Suppose now that { : <2 } is unbounded in and dene a
strictly increasing continuous sequence c () ( < ()) of its elements as follows.
Let c (0) = min{ : <2 }, c () = sup< c () for limit, and c ( + 1) is the
minimal <2 -predecessor of such that > c () and has the minimal f -image
among all <2 -predecessors that are > c (). The ordinal () is dened as the
place where the process stops, i.e., when = sup<() c (). Let

C = {c () : < ()}.
204 Chapter 7. Square Sequences

It is easily checked that this gives a square sequence C ( < + ) with the property
that tp(C ) for all < + . 
Denition 7.3.5. Square sequences C ( < + ) that have the property tp(C )
for all < + are usually called  -sequences.
So, from now on, we x a  -sequence C ( < + ) and concentrate on
dening and analysing the corresponding function , the main object of study in
this section. Let

(, ) = maximal limit point of C ( + 1) (7.3.1)

when such a limit point exists; otherwise (, ) = 0. The main purpose of this
section is to introduce a distance function

: [+ ]2 (7.3.2)

associated with the  -sequence C ( < + ) and dened recursively by the


formula

(, ) = max{tp(C ), (, min(C \ )), (, ) :


C [(, ), )},

with the boundary value (, ) = 0 for all . The following immediate conse-
quence of the coherence property of the  -sequence is quite useful.
Lemma 7.3.6. If is a limit point of C , then (, ) = (, ) for all < . 
For example, using Lemma 7.3.6, one can easily conclude that the mapping
is in fact equal to the function dened in Section 7.2. So, we have the following
crucial subadditivity properties of .
Lemma 7.3.7. For all < + ,
(a) (, ) max{(, ), (, )},
(b) (, ) max{(, ), (, )}. 
Clearly (, ) 1 (, ), so by Lemma 6.2.1, we have
Lemma 7.3.8. |{ : (, ) }| || + 0 for all < + and < . 
The following property is another immediate consequence of the fact that
the denition of is closely tied to the notion of a minimal walk along the square
sequence (compare with Lemma 7.2.3 above ).
Lemma 7.3.9. (, ) max{(, ), (, )} whenever and belongs
to the trace of the walk from to . 
Using Lemmas 7.3.6 and 7.3.9 one proves the following fact exactly as in the
case of of Section 7.2 (see the proof of Lemma 7.2.4).
7.4. The function on successors of regular cardinals 205

Lemma 7.3.10. If 0 < and if is a limit ordinal, then there is < such
).
that (, ) (, ) for all in the interval [, 
The proof of the following fact is also completely analogous to the proof of
the corresponding fact for the local version considered above in Section 7.2 (see
the proof of Lemma 7.2.5).
Lemma 7.3.11. P () = { < : (, ) } is a closed subset of for all
< + and < . 
Since in this case is equal to the local version , it is worth restating the
following unboundedness property that follows from Theorem 7.2.13, a property
that we will be continuously rening as we go on.
Theorem 7.3.12. It is possible to modify the  -sequence C ( < + ) in such a
way that the corresponding function : [+ ]2 has the property that for every
< and every family A of pairwise-disjoint subsets of + that have sizes < cf()
there is a subfamily B A of cardinality + such that for all a = b in B and all
a, b we have (, ) > . 
The discussion of : [+ ]2 now splits naturally into two cases depend-
ing on whether is a regular or a singular cardinal (with the case cf() = of
special importance). This is done in the following two sections.

7.4 The function on successors of regular cardinals


In this section is a xed regular cardinal, C ( < + ) a xed  -sequence and
: [+ ]2
is the corresponding -function dened by the recursive formula
(, ) = max{tp(C ), (, min(C \ )), (, ) :
C [(, ), )},
with the boundary condition (, ) = 0 for all .
For < and < < + , set
< if and only if (, ) . (7.4.1)
The following lemma summarizes some of the properties of these relations and they
are all immediate consequences of the corresponding facts about various -function
given in previous sections.
Lemma 7.4.1.
(a) < is a closed tree-ordering of + of height for all < ,
(b) < < whenever < ,

(c)  + = < < . 
206 Chapter 7. Square Sequences

Remark 7.4.2. Recall that by Theorem 7.3.3 for every regular cardinal
the -relation of + can also be written as an increasing union of a -sequence
< ( < ) of closed tree-orderings. When < , however, we can no longer
insist that the trees (+ , < ) have heights . Using an additional set-theoretic
assumption (compatible with the existence of a  -sequence) one can strengthen
(a) of Lemma 7.4.1 and have that the height of (+ , < ) is < for all < .
Without additional set-theoretic assumptions, these trees, however, do have some
properties of smallness not covered by Lemma 7.4.1.
Lemma 7.4.3. If > , then no tree (+ , < ) has a branch of size .

Proof. Suppose towards a contradiction that some tree (+ , < ) does have a
branch of size and let B be one such xed branch (maximal chain). By Lemmas
7.3.8 and 7.3.11, if = sup(B) then B is a closed and unbounded subset of of
order-type . Since is regular and uncountable, C B is unbounded in C , so
in particular we can nd C B such that tp(C ) > . Reading o the
denition of (, ) we conclude that (, ) = tp(C ) > . Similarly we can
nd a > belonging to the intersection of lim(C ) and B. Then C = C
so C and therefore (, ) = tp(C ) > contradicting the fact that
< . 
Lemma 7.4.4. If > , then no tree (+ , < ) has a Souslin subtree of height .

Proof. Forcing with a Souslin subtree of (+ , < ) of height would produce an


ordinal of conality and a closed and unbounded subset B of C forming a
chain of the tree (+ , < ). It is well known that in this case B would contain a
ground model subset of size contradicting Lemma 7.4.3. 

The unboundedness property of stated in Theorem 7.3.12 should be com-


pared to the following one that ts better the main theme of this section.
Theorem 7.4.5. If > , then for every < and every family A of pairwise-
disjoint nite subsets of + , there exists A0 A of size such that for all a = b
in A0 and all a, b we have (, ) > .

Proof. We may assume that for some n and all a A, we have |a| = n. Let
a(0), . . . , a(n 1) enumerate a given element a of A increasingly. Using Lemma
7.4.4 and shrinking A, we may assume that {a(i) : a A} is an antichain of
(+ , < ) for all i < n. For two elements a and b of A, we write a < b if every
ordinal from a is smaller than every ordinal from b. Going to a subfamily of A of
equal size we may assume that a < b or b < a for every pair a and b of distinct
elements of A. We now dene a coloring

f : [A]2 {0} {(i, j) : i, j < n}

as follows. Let f (a, b) = 0 if a b is an < -antichain. Otherwise, assuming a < b,


let f (a, b) = (i, j) where (i, j) is the minimal pair such that a(i) < b(j). By the
7.4. The function on successors of regular cardinals 207

DushnikMiller partition theorem ([25]), either there exists A0 A of size such


that f is constantly equal to 0 on [A0 ]2 or there exists a pair (i, j) of integers
< n and an innite A1 A such that f is constantly equal to (i, j) on [A1 ]2 . The
rst alternative is what we want, so let us see that the second one is impossible.
Otherwise, x a triple a < b < c of elements of A1 . Then a(i) and b(i) are both
< -dominated by c(j). Since < is a tree-ordering, we must have a(i) < b(i),
contradicting our initial assumption that the set {a(i) : a A} forms an antichain
of the tree (+ , < ). This completes the proof. 
We shall see later that the unboundedness property of Theorem 7.4.5 can be
quite useful in designing forcing notions satisfying good chain conditions. Having
such applications in mind, we shall now work on rening further this kind of
unboundedness property of the -function.
Lemma 7.4.6. Suppose > 0, let < + and let { , } ( < ) be a sequence of
pairwise-disjoint elements of [+ ]2 . Then there is an unbounded set such
that { , } min{{ , }, { , }} for all = in .2
Proof. Clearly we may assume that the sequences ( < ) and ( < ) are
strictly increasing. For deniteness we assume that for all . The case
> for all is considered similarly. Let

= sup and = sup . (7.4.2)

Then and are ordinals of conality , so the order-types of C and C are


both equal to . It will be convenient to assume that the two sequences { } and
{ } are actually indexed by C rather than . It is then clear we may assume
that
for all C . (7.4.3)
Case 1: = . By Lemma 7.3.10, for each limit point of C there is f () <
in C such that

(, ), (, ) (, ) = (, ) for all [f (), ). (7.4.4)

Fix a stationary lim(C ) such that f is constant on . Then

( , ) ( , ) for all < in , and (7.4.5)

( , ) ( , ) for all < in . (7.4.6)


Subcase 1a : . Using the two subadditive properties of in Lemma 7.3.7 one
easily concludes that (, ) = (, ) for any < such that (, ) > (, ).
So, going to a tail of we may assume that

( , ) = ( , ) and ( , ) = ( , ) for all . (7.4.7)


2 Here,and everywhere else later in this chapter, the convention is that, {, } is meant to be
equal to (, ) if < , equal to (, ) if < , and equal to 0 if = .
208 Chapter 7. Square Sequences

Consider < in . Combining (7.4.5) and the rst equality of (7.4.7) gives us

( , ) ( , ) = ( , ),

which is sucient for the conclusion of Lemma 7.4.6. Suppose > are chosen
from in this order. Combining (7.4.6) and the second equality of (7.4.7) gives us

( , ) ( , ) = ( , ),

which is also sucient for the conclusion of Lemma 7.4.6.


Subcase 1b : < . Using Lemma 7.3.8 and going to a tail of the set we may
assume that lies above and that

( , ), ( , ) > (, ) for all . (7.4.8)

Applying the subadditive property Lemma 7.3.7(b) of we get for all :

(, ) max{(, ), ( , )} = ( , ), and (7.4.9)

(, ) max{(, ), ( , )} = ( , ). (7.4.10)
If < are chosen from in this order, then (7.4.5) and (7.4.9) give us

( , ) ( , ) (, ),

which is sucient for the conclusion of Lemma 7.4.6. If > are chosen from
in this order, then combining (7.4.6) and (7.4.10) gives us

( , ) ( , ) (, ),

which is again sucient for the conclusion of Lemma 7.4.6.


Case 2: < . We may assume all > for all and working as in Case 1 we
nd a stationary set of limit points of C such that

( , ) ( , ) for all < in . (7.4.11)

By Lemma 7.3.8 for each there is g() C such that (, ) > (, ) for
all in the interval [g(), ). Using the two subadditive properties of from Lemma
7.3.7, one concludes that (, ) = (, ) for all [g(), ). Intersecting with
the closed and unbounded subset of C of all ordinals that are closed under the
mapping g we may assume

( , ) = ( , ) for all > in . (7.4.12)

Subcase 2a : < . Applying Lemma 7.3.8 again and going to a tail of we may
assume that lies above and

( , ) > (, ) for all . (7.4.13)


7.4. The function on successors of regular cardinals 209

Applying the subadditive properties of we get

(, ) max{(, ), ( , )} = ( , ) for all . (7.4.14)

If < are chosen from in this order, then combining the inequalities (7.4.11)
and (7.4.14) we get
( , ) ( , ) (, ),
which is sucient to give us the conclusion of Lemma 7.4.6. If > are chosen
from in this order, then combining (7.4.12) and (7.4.14) we get

( , ) = ( , ) (, ),

which is again sucient for the conclusion of Lemma 7.4.6.


Subcase 2b : . Going to a tail of we may assume that ( , ) > (, ), so
as before the subadditive properties give us that

( , ) = ( , ) for all . (7.4.15)

If < are chosen from in this order, then combining the inequality (7.4.11)
and equality (7.4.15) we have

( , ) ( , ) = ( , ),

which is sucient to give us the conclusion of Lemma 7.4.6. If > are chosen
from in this order, then combining the equalities (7.4.12) and (7.4.15) we have

( , ) = ( , ) = ( , ),

which is again sucient to give us the conclusion of Lemma 7.4.6. This nishes
the proof. 
Lemma 7.4.7. Suppose is -inaccessible3 for some < and that A is a family
of size of subsets of + , all of size < . Then for every ordinal < there is a
subfamily B of A of size such that for all a and b in B:
(a) {, } > for all a \ b and b \ a.
(b) {, } min{{, }, {, }} for all a \ b, b \ a and a b.
Proof. Consider rst the case = . Going to a subfamily, assume A forms a
-system with root r and that for some xed integer n 1 and all a A we have
|a \ r| = n. By Lemma 7.4.5, going to a subfamily we may assume that A satises
(a). For a A let a(0), . . . , a(n 1) be the increasing enumeration of a \ r. Going
to a subfamily, we may assume that A can be enumerated as a ( < ) in such
a way that a (i) ( < ) is strictly increasing for all i < n. By n2 |r| successive
3 I.e., < for all < and < .
210 Chapter 7. Square Sequences

applications of Lemma 7.4.6 we nd a single unbounded set such that for


all r, (i, j) n n and = in :

{a (i), a (j)} min{{a (i), }, {a (j), }}. (7.4.16)

This gives us the conclusion (b) of the lemma.


Let us now consider the general case. Using the -inaccessibility of we may
assume that A forms a -system with root r and that members of A have some
xed order-type, i.e., that for some xed ordinal < we can enumerate each
a \ r for a A in increasing order as a(i) (i < ). Using -inaccessibility again
and rening A we may assume that there is an enumeration a ( < ) of A such
that a (i) ( < ) is a strictly increasing sequence for all i < . For i < , set

(i) = sup a (i).

Then cf((i)) = and therefore tp C(i) = for all i < . Going to a smaller
cardinal we may assume that is a regular cardinal < . Applying Lemma 7.3.10
simultaneously times gives us a regressive map f : such that for every
< of conality , all [c(i) (f ()), c(i) ()) and i < :

(, a (i)) (, c(i) ()) = (, (i)). (7.4.17)

Here c(i) () ( < ) refers to the increasing enumeration of the set C(i) for each
i < . Pick a stationary set of conality ordinals < such that f is constant
on . It follows that for all i, j < , if (i) = (j), then

{a (i), a (j)} (a (i), (i)) for < in , (7.4.18)

{a (j), a (i)} (a (j), (i)) for > in . (7.4.19)


On the other hand, if i, j < such that (i) < (j) we have

{a (i), a (j)} (a (i), (i)) for all < in . (7.4.20)

Intersecting with a closed and unbounded subset of we also assume that

a (i) c(i) () for all i < and . (7.4.21)

a (i) a (j) for all and i, j < such that (i) (j). (7.4.22)
Since (, (i)) tp(C(i) ) for every i < and < (i), combining (7.4.18)
(7.4.22) we see that

{a (i), a (j)} min{, } for all i, j < and = in . (7.4.23)

So replacing with \ ( + 1) gives us the conclusion (a) of Lemma 7.4.7.


Suppose now we are given two coordinates i, j < and an ordinal in the
root r. Let
= a (i) and = a (j)
7.4. The function on successors of regular cardinals 211

for . Note that (7.4.18) and (7.4.19) give us the conditions (7.4.5) and (7.4.6)
of Case 1 (i.e., (i) = = = (j)) in the proof of Lemma 7.4.6 while (7.4.20)
gives us the condition (7.4.11) of Case 2 (i.e., (i) = < = (j)) in the proof
of Lemma 7.4.6. Observe that in this proof of Lemma 7.4.6, once the set was
obtained by a single application of the Pressing Down Lemma to nally give us
(7.4.5) and (7.4.6) of Case 1 or (7.4.11) of Case 2, the rest of the renements of
all lie in the restriction of the closed and unbounded lter to . In other words,
we can now proceed as in the case = and successively apply this rening
procedure for each r and (i, j) obtaining a set 0 with the
property that \ 0 is nonstationary and
{a (i), a (j)} min{{a (i), }, {a (j), }} for all r,
(7.4.24)
i, j < , and = in 0 .
Since this is clearly equivalent to the conclusion (b) of Lemma 7.4.7, the proof is
completed. 
Denition 7.4.8. The function D : [+ ]2 [+ ]< is dened by

D(, ) = { : (, ) (, )}.4

Note that D(, ) = { : (, ) (, )}, so we could take the formula

D{, } = { min{, } : (, ) {, }}

as our denition of D{, } which could serve when there is no implicit assumption
about the ordering between and , contrary to the case whenever we write
D(, ) where we implicitly assume that < .
Lemma 7.4.9. If is -inaccessible for some < , then for every family A of
size of subsets of + , all of size < , there exists B A of size such that for
all a and b in B and all a \ b, b \ a and a b:

(a) , > implies D{, } D{, } D{, },


(b) > implies D{, } D{, },
(c) > implies D{, } D{, },
(d) > , implies D{, } D{, } or D{, } D{, }.

Proof. Choose B A of size satisfying the conclusion (b) of Lemma 7.4.7. Pick
a = b in B and consider a \ b, b \ a and a b. By the conclusion of
Lemma 7.4.7(b), we have

{, } min{{, }, {, }}. (7.4.25)

a. Suppose , > . Note that in this case a single inequality (, ) {, }


or (, ) {, } given to us by (7.4.25) implies that we actually have both
4 Note that by our boundary requirements (, ) = 0, we have that in particular D(, ).
212 Chapter 7. Square Sequences

inequalities simultaneously holding. The subadditivity of gives us (, )


{, }, or equivalently (, ) {, } for any with (, ) (, )
or (, ) (, ). This is exactly the conclusion of 7.4.9(a).
b. Suppose that > > . Using the subadditivity of we see that in both
cases given to us by (7.4.25), we have that (, ) {, }. So the inclusion
D{, } D{, } follows immediately.
c. Suppose that > > . The conclusion D{, } D{, } follows from the
previous case by symmetry.
d. Suppose that > , . Then (, ) {, } gives D{, } D{, } while
(, ) {, } gives us D{, } D{, }.
This completes the proof. 
Remark 7.4.10. Note that min{x, y} D{x, y} for every {x, y} [+ ]2 , so the
conclusion (a) of Lemma 7.4.9 in particular means that < min{, } implies
D{, }. In applications, one usually needs this consequence of 7.4.9(a) rather
than 7.4.9(a) itself. We shall also need the convention that D{x, y} = {x} whenever
x = y.
We nish this section with a remark about the behavior of D on families of
sets of size + rather than .
Lemma 7.4.11. Suppose that < = and that A is a family of subsets of +
of size < . Then there is a subfamily B A of size + forming an increasing
-system with root r5 such that for all a, b B such that a \ r < b \ r, we have
that a ( + 1) D(, ) for all a \ r and b \ b.
Proof. By our assumption < = and the -system lemma at that level, and
by going to a subfamily of A of size + , we may assume that A already forms an
increasing -system with root r. Going to a further subfamily of A of equal size,
we may assume that for some ordinal < and all a A,
(, ) < for all in a. (7.4.26)
Since is assumed to be a regular cardinal, since is based on a  sequence
which obviously avoids conality ordinals, and since = , Theorem 7.2.13
applies giving us a subfamily B A of size + such that for all a, b B with
a \ r < b \ r,
(, ) > for all a \ r and b \ r. (7.4.27)
Combining these two last equations, we see that for a, b B with a \ r < b \ r, for
a \ r and b \ r, and for a ( + 1), we have
(, ) < < (, ). (7.4.28)
Since D(, ) = { : (, ) (, )}, this gives us the conclusion of the
lemma. 
5 We say that B forms an increasing -system with root r if for all a = b in B, we have that
a b = r and either a \ r < b \ r, or else b \ r < a \ r.
7.5. Forcing constructions based on 213

7.5 Forcing constructions based on


We start with the most typical application of Lemma 7.4.9 in forcing constructions.
Theorem 7.5.1. If  holds for some regular cardinal that is -inaccessible, then
there is a -closed -cc forcing notion P that introduces a Souslin tree of height + .
Proof. We let P be the collection of all pairs p = (Tp , p ), where Tp is a subset of
+ of size < and where p is a tree-ordering on Tp compatible with the usual
ordering between the ordinals and with the property that
p and p implies that there is D{, } Tp
(7.5.1)
such that p and p and p .
We let p be a stronger condition than q, and write p q, whenever Tp Tq and
p  Tq2 =q . Clearly, P is a -closed forcing notion, so let us check that it satises
the -chain condition. Consider a family A P of size . Applying Lemma 7.4.9,
we obtain a subfamily B A of size such that the family of sets Tp (p B)
satises the conditions (a), (b), (c), and (d) of the lemma. By a further thinning,
we may assume that Tp (p B) forms a -system with root T . We may also
assume that Tp s are isomorphic trees via the order-isomorphisms between their
domains as sets of ordinals and that xes T . So in particular p (p B) all agree
on T 2 . Moreover, we may assume that

D{, } (Tp \ T ) = for all , T and p B. (7.5.2)

Consider p = q in B. Let r = (Tr , r ), where Tr = Tp Tq , and where r is the


minimal tree-ordering on Tr containing p and q . We nish the proof of -cc for
P, by checking that r has the property (7.5.1).
Suppose that r and r . We need to nd D{, } Tr such
that r and r and r .
Case 1. T . Assuming that one of the of is in T would reduce our
problem to the property (7.5.1) of p of q. So we may assume that Tp \ T and
Tq \ T . Then Lemma 7.4.9 (a) gives us that D{, }, so we could take
= .
Case 2. / T . By symmetry, we may consider only the case Tp \ Tq .
Since p has the property (7.5.1), we may assume that one of the or does not
belong to Tp . By symmetry, let us assume that Tq \ T .
Subcase 2.1. Tq \ T . By the minimality of the ordering r , we can nd
p , p T such that p p q and p p q . By the property (7.5.1) for
p there is p D{p , p } Tp such that p p and p p p and p p p . By
(7.5.2), p must belong to T , and therefore to Tq . Applying (7.5.1) for q, we can
nd D{, } Tq such that p q and q and q . Then is as
required.
Subcase 2.2. T . By our assumption r and by the minimality of
r there must be p T such that p p q . By the property (7.5.1) of p
214 Chapter 7. Square Sequences

there is  D{, p } such that p  ,  p and  p p . By (7.5.2), 


must belong to T and therefore to Tq , so by the property (7.5.1) of q there must
be D{, } Tq such that  q and q and q . Then this is
a witness that r has the property (7.5.1) for the triple , and and so we are
done in this subcase as well.
Subcase 2.3. Tp \ T . Take again p T such that p p q . Applying
Lemma 7.4.9 (b) for the triple , , and p in place of , , and , respectively,
we conclude that D{, p } D{, }. Applying the property (7.5.1) for p in the
case of the inequalities p and p p , we can nd D{, p } Tp such
that p and p and p p . Since D{, p } D{, } the ordinal
belongs also to D{, } Tr , as required.
Since P obviously has cardinality + , one condition from P must force that
the generic tree T has cardinality + . So the proof is nished, once we show that
P forces that T has no chains nor antichains of cardinality + . This reduces to
showing that, given any family A P of size + and given an assignment p Tp
(p A) such that p = q whenever p = q, we can nd distinct p, q A and
r p, q such that p and q are r -incomparable, as well as distinct p, q A and
r p, q such that p and q are r -comparable. Let us concentrate on the rst of
these goals. Going to a subfamily of A of size + , we may assume that Tp (p A)
forms a -system with root T which forms an initial segment of all the Tp s which
we assume as isomorphic as trees via order-isomorphisms between their domains as
sets of ordinals, the isomorphisms that x Tp . Moreover, we may assume that the
order-isomorphisms serve also as isomorphisms between the structures (Tp ,  Tp2 )
(p A). Note that this in particular means that for some 0 < ,

sup{(, ) : , Tp } = 0 for all p A. (7.5.3)

By Theorem 7.3.12, we can nd B A of cardinality + such that

{, } > 0 for Tp \ T, Tq \ T, where p = q in B. (7.5.4)

Fix two dierent members p and q of B. Let Tr = Tp Tq and let r be the minimal
tree-ordering on Tr extending p and q . Since p Tp \ T and q Tq \ T , and
since T is a common initial segment of Tp and Tq , we conclude that p and q are
incomparable relative to r . So it remains only to show that r is a member of P,
i.e., that it satises (7.5.1). However, note that this has already been done during
the course of the proof that P satises the -chain condition because (7.5.3) and
(7.5.4) are giving us the necessary input for that argument.
Choose now p and q in B such that every ordinal of Tp \ T is smaller than
every ordinal of Tq \ T . Let Tr = Tp Tq and let r be the minimal tree-ordering
on Tr extending p and q such that p <r q . It remains to show that r satises
(7.5.1). Let r and r . Note that and belong to the set Tq \T and they
do not have common q -predecessors in that sets they do not have common r -
predecessors either. This takes care of the case , Tp \Tq . The case , Tp \T
7.5. Forcing constructions based on 215

is also immediate. So it remains to consider the case Tp \ T and Tq \ T .


Combining (7.5.3) and (7.5.4) we get that

D{, } Tq = Tp ( + 1). (7.5.5)


It follows that D{, } Tq , so we can take = to obtain the conclusion of
(7.5.1). This nishes the proof. 
Corollary 7.5.2. If 1 holds, there is a property K forcing notion that adds a
Souslin tree on 2 . 
Theorem 7.5.3. If  holds for some regular cardinal that is -inaccessible, then
there is a -closed -cc forcing notion P that introduces a mapping g : [+ ]2
with property that Eg (, ) = { < : g(, ) = g(, )} has cardinality < for
all < < + .
Proof. Let P be the collection of all mappings of the form p : [Dp ]2 , where
Dp is a subset of + of cardinality < and where the following condition is
satised for every triple < < of elements of Dp ,

/ D(, ) implies p(, ) = p(, ) (7.5.6)
For p and q in P, let p q if and only if Dp Dq , p  [Dq ]2 = q, and

p(, ) = p(, ) for every < in Dq and (Dp \ Dq ) . (7.5.7)

It is clear that P is a -closed forcing notion, so let us show that it satises the
-chain condition. Let A be a given subset of P of cardinality . By our assumption
about -inaccessibility of , we can nd B A of cardinality forming a -system
with root r. By Lemma 7.4.9 and by going to a subfamily of B of size , we may
assume that the family of sets Dp (p B) satises the conditions (a), (b), (c),
and (d) of the lemma. Choose two dierent members p and q from B and nd a
mapping s : [Dp Dq ]2 extending p and q such that s is one-to-one on the
set
[Dp Dq ]2 \ ([Dp ]2 [Dq ]2
and such that the s-image of this set is disjoint from the ranges of p and q. It
remains to check that s satises the condition (7.5.6) for every triple < <
of elements of Ds = Dp Dq . Fix such a triple Ds = Dp Dq and assume that
p(, ) = p(, ). We need to show that D(, ). From our denition of s it
follows that
{, }, {, } [Dp ]2 [Dq ]2 .
If {, }, {, } [Dp ]2 or if {, }, {, } [Dp ]2 the required conclusion
D(, ) follows from the fact that p and q satisfy the condition (7.5.6). So it
remains to consider the case when, say, {, } [Dp ]2 \ [Dq ]2 and {, } [Dq ]2 \
[Dp ]2 . Note that this means that

Dr = Dp Dq , Dp \ Dr and Dq \ Dr . (7.5.8)
216 Chapter 7. Square Sequences

This in turn means that the conclusion (a) of Lemma 7.4.9 applies, and so we have
that
(D(, ) D(, )) D(, ), (7.5.9)
as required. Having checked that P satises the -chain condition, note that the
generic lter of P gives us the mapping g : [+ ]2 with the property that
Eg (, ) has cardinality < for all < < + . 
Corollary 7.5.4. Suppose that  holds for some regular cardinal that is
-inaccessible and that = + for some cardinal . Then there is a -closed
-cc forcing notion P that introduces a mapping f : (+ )3 which is not
constant on any product A B C of three subsets of + of cardinality .
Proof. It suces to construct such a map f : (+ )3 on the basis of map
g : [+ ]2 satisfying the conclusion of Theorem 7.5.3. For this, we rst choose
a mapping e : []2 with the property that
e(, ) = e(, ) whenever < < . (7.5.10)
+ 3 4
Dene f : ( ) as follows:
f (, , ) = (p(, , ), e(g(, ), g(, )), e(g(, ), g(, )), e(g(, ), g(, ))),
where p(, , ) codes the order and equality relations between the ordinals ,
and . Consider three subsets A, B and C of + of cardinality . We need to show
that f is not constant on the product A B C. We may assume that A, B and
C have order-types equal to . Since we have incorporated the mapping p into f
the three sets must have dierent suprema. So we may, in fact, assume that
sup A < sup B < sup C.
Applying the basic property of g, we can nd subsets B0 B and C0 C of
cardinality and A such that
g(, ) = g(, ) for B0 and C0 . (7.5.11)
Going to subsets of B0 and C0 , we may also assume that
g(, ) < g(,  ) for <  in B0 , and (7.5.12)
 
g(, ) < g(, ) for < in C0 . (7.5.13)
Let = sup{g(, ) : B0 } and = sup{g(, ) : C0 }. We may assume
that . Let <  be the rst two elements of B0 . Choose C0 such that
g(, ) > g(,  ) > g(, ).
By the choice of the mapping e, we have that
e(g(, ), g(, )) = e(g(,  ), g(, )).
It follows that the second term of f (, , ) is not equal to the second term of
f (,  , ), and so in particular f (, , ) = f (,  , ), as required. 
7.5. Forcing constructions based on 217

Corollary 7.5.5. If c+ holds, then there is a -closed c+ -cc forcing notion that
introduces a mapping f : (c++ )3 which is not constant on any product
A B C of three uncountable subsets of c++ . 
Denition 7.5.6. The -function of some family F of subsets of some ordinal
(respectively, a family of functions with domain ) is the function : [F]2
dened by
(f, g) = min(f ! g),
(respectively, (f, g) = min{ : f () = g()}).
Note the following property of :
Lemma 7.5.7. (f, g) min{(f, h), (g, h)} for all {f, g, h} [F ]3 . 
Remark 7.5.8. This property can be very useful when transferring objects that
live on to objects on F. This is especially interesting when F is of size larger
than while all of its restrictions F  = {f : f F } ( < ) have size
< , i.e., when F is a Kurepa family (see for example [18]). We shall now see that
it is possible to have a Kurepa family F = {f : < + } whose -function is
dominated by , i.e., (f , f ) (, ) for all < < + .
Theorem 7.5.9. If  holds and is -inaccessible, then there is a -closed -cc
forcing notion P that introduces a Kurepa family on .
Proof. Going to a larger cardinal, we may assume that is regular. Put p in P,
if p is a one-to-one function from a subset of + of size < into the family of all
subsets of of size < such that for all and in dom(p):

p() p() is an initial part of p() and of p(), (7.5.14)

(p(), p()) (, ) provided that = . (7.5.15)


Let p q whenever dom(p) dom(q) and p() q() for all dom(q). Clearly
P is a -closed forcing notion. To see that it satises the -chain condition, let A
be a given subset of P of size . By the assumption that is -inaccessible, going
to a subfamily of A of size , we may assume that A forms a -system,
 or more
precisely, that dom(p) (p A) forms a -system with root d, that rng(p) (p A)
forms a -system with root c and that any two members of A generate isomorphic
structures via the natural isomorphism that xes the roots. By Lemma 7.4.7 there
exists B A of size such that for all p, q B, all dom(p) \ dom(q), all
dom(q) \ dom(p) and all dom(p) dom(q)(= d):

{, } > max(c), (7.5.16)

{, } min{{, }, {, }}. (7.5.17)


 
Suppose p, q B and that a = ( rng(p)) \ c lies entirely below b = ( rng(q)) \ c.
Dene r : dom(p) dom(q) []< as follows. For d, let r() = p() q().
218 Chapter 7. Square Sequences

If dom(q) \ dom(p) and if there is d such that (q(), q()) b, the


must be unique. Put r() = p() q() in this case. If such does not exist,
put r() = q() {max(c) + 1}. For dom(p) \ dom(q), put r() = p().
Note that this choice of r does not change the behavior of on pairs coming
from [dom(p)]2 or [dom(q)]2 , so in checking that r satises (7.5.14) it suces to
assume dom(p) \ dom(q) and dom(q) \ dom(p). If r() was dened as
p() q() for some d, we have that (r(), r()) = (p(), p()) and that
(q(), q()) b. It follows that
(r(), r()) = min{(p(), p()), (q(), q())}
(7.5.18)
min{{, }, {, }}.
Applying (7.5.17) we get the desired inequality (r(), r()) {, }. If r()
was chosen to be q() {max(c) + 1}, then (r(), r()) max(c) + 1 and this
is {, } by (7.5.16).
Let G be the generic lter of P and let F = {f : < + }, where f is the
A simple genericity argument shows that each f intersects
union of p() (p G).
every interval of the form [, + ) ( < ). It follows (see (7.5.14)) that F 
has size at most for all < . This nishes the proof. 
Corollary 7.5.10. If 1 holds, and so in particular if 2 is not a Mahlo cardinal
in the constructible universe, then there is a property K poset, forcing the Kurepa
hypothesis. 
Remark 7.5.11. This is a slight strengthening of a result of Jensen who proved
that under 1 there is a ccc poset forcing the Kurepa hypothesis. It should be
noted that Jensen also proved (see [53]) that in the Levy collapse of a Mahlo
cardinal to 2 there is no ccc poset forcing the Kurepa hypothesis. Velickovic
[125] was the rst to use to re-prove Jensens original result, though his proof
worked only in case = 1 and produced ccc rather than property K forcing. He
has also shown (see [125]) that easily yields an example of a function with the
-property in the sense of Baumgartner and Shelah [11]. Recall that a function
with the -property in the sense of [11] is a function that satises some of the
properties of the function D of Lemma 7.4.9 and a function that forms a ground
for the well-known forcing construction of a locally compact scattered topology on
2 , all of whose CantorBendixson ranks are countable. We now use the deeper
analysis of to improve the chain-condition of the BaumgartnerShelah forcing.
Theorem 7.5.12. If 1 holds, then there is a property K forcing notion that intro-
duces a locally compact scattered topology on 2 , all of whose CantorBendixson
ranks are countable.
Proof. Let P be the set of all p = Dp , p , Mp  where Dp is a nite subset of
2 , where p is a partial ordering of Dp compatible with the well-ordering and
Mp : [Dp ]2 [2 ]< has the following properties:
Mp {, } D{, } Dp , (7.5.19)
7.5. Forcing constructions based on 219

Mp {, } = {} if p and Mp {, } = {} if p , (7.5.20)
p , for all Mp {, }, (7.5.21)
for every p , there is Mp {, } such that p . (7.5.22)
We let p q if and only if Dp Dq , p  Dq =q and Mp  [Dq ]2 = Mq . To verify
that P satises Knasters chain condition, let A be a given uncountable subset of
P. Shrinking A we may assume that A forms a -system with root R and that
the conditions of A generate isomorphic structures over R. By Lemma 7.4.9 there
is an uncountable subset B of A such that for all p and q in B and all Dp \ Dq ,
Dq \ Dp and R = Dp Dq :

, > implies D{, } D{, } D{, }, (7.5.23)


> implies D{, } D{, }, (7.5.24)
> implies D{, } D{, }. (7.5.25)

Consider two conditions p and q from B. Let r = Dr , r , Mr  be dened as


follows: Dr = Dp Dq and r if and only if p , or q , or there is
R such that p q or q p . It is easily checked that r is a
partial ordering on Dr compatible with the well-ordering such that r  Dp =p
and r  Dq =q . Dene Mr : [Dr ]2 [2 ]< by letting Mr  [Dp ]2 = Mp ,
Mr  [Dq ]2 = Mq , Mr {, } = {} if r , Mr {, } = {} if r and if
Dp \ R and Dq \ R are r -incomparable, set

Mr {, } = { D{, } Dr : r , }. (7.5.26)

It is nontrivial to check that a so-dened r satises (7.5.22). So let r , be


given. If , Dp \ Dq and Dq \ Dp , then there exist , R such that
q p and q p . By (7.5.22) for q there exists Mq { , }
such that q . Since Mq { , } R we conclude that R. Thus p , ,
so by (7.5.22) there is Mp {, } such that p and therefore r . The
case R, Dp \ D and Dq \ Dp is quite similar.
Consider now the case when Dp \ Dq , Dq \ Dp and they are r -
incomparable, i.e., Mr {, } is given by (7.5.26). For symmetry we may consider
only the subcase r , with Dp . Then p and p q for some
R. By (7.5.22) for p there is Mp {, } such that p . It remains to show
that was put in Mr {, } in (7.5.26), i.e., that belongs to D{, }. Note that
since r agrees with the well-ordering, we have that > , so by (7.5.24) we have
that D{, } D{, }. By (7.5.19) for p we have that Mp {, } D{, }
so we can conclude that D{, }. The rest of the cases are symmetric to the
ones above.
Let G be a generic lter of P and let G be the union of p (p G).
For < 2 set B = { : G }. Let be the topology on 2 generated by
B ( < 2 ) as a clopen subbasis. It is easily checked that is a Hausdor locally
220 Chapter 7. Square Sequences

compact scattered topology on 2 . A simple density argument and induction on


shows that the interval [, + ) is the th Cantor-Bendixson rank of the
space (2 , ). 
Denition 7.5.13. A function f : [2 ]2 [2 ] has property if for every
uncountable set A of nite subsets of 2 there exist a and b in A such that for all
a \ b, b \ a and a b:

(a) , > implies f {, },


(b) > implies f {, } f {, },
(c) > implies f {, } f {, }.

Remark 7.5.14. This denition is due to Baumgartner and Shelah [11] who have
used it in their well-known forcing construction. It should be noted that they were
also able to force a function with the property using a -closed 2 -cc poset. As
shown above (a fact rst checked in [11, p. 129]), the function

D{, } = { min{, } : (, ) {, }}

has property . However, Lemma 7.4.9 shows that the function D has many other
properties that are of independent interest and that are likely to be needed in other
forcing constructions of this sort. The reader is referred to papers of Koszmider
[60] and Rabus [84] for further work in this area.

7.6 The function on successors of singular cardinals


In the previous section we have shown that the function : [+ ]2 based
on a  -sequence C ( < + ) can be a quite useful tool in stepping-up objects
from to + . In this section we analyse the stepping-up power of under the
assumption that is a singular cardinal of conality . So let n (n < ) be a
strictly increasing sequence of regular cardinals converging to xed from now on.
This immediately gives rise to a rather striking tree decomposition <n (n < )
of the -relation on + :

<n if and only if < and (, ) n . (7.6.1)


Lemma 7.6.1.

(1)  + = n< <n ,
(2) <n <n+1 ,
(3) (+ , <n ) is a tree of height +
n. 
The following information about the tree-orderings <n could sometimes be
useful.
7.6. The function on successors of singular cardinals 221

Lemma 7.6.2. For every n < the tree-ordering <n is a closed relation in the sense
that { < : <n } is a closed subset of for every < + . In, particular, the
tree (+ , <n ) has no chains of size +
n.

Proof. The rst conclusion follows from Lemma 7.4.1. To see the second conclusion
of the lemma, let b be a maximal chain of the tree (+ , <n ) and suppose that
= sup(b) is a limit ordinal of uncountable conality. Then by the rst conclusion,
the set b is a closed and unbounded subset of . It follows that if C denotes the set
of limit points of C , then the intersection c = b C is also closed and unbounded
in . Consider two ordinals < from c. Then <n , and therefore,

n (, ) tp(C ).

Since C = C and C = C , we conclude that tp(C ) n . Since


this is true for all c, it follows that C has order-type at most n . So the tree
(+ , <n ) cannot have chains of order-type + n. 
For example, one can use Lemma 7.6.2 to prove the following unboundedness
property of .
Theorem 7.6.3. Suppose that for some n < we have a family A of + n pairwise-
disjoint nite subsets of + . Then there is a subfamily B A of size +
n such that
for every a = b in B, we have that (, ) > n for all a and b.
Proof. By Lemma 7.6.2 the tree (+ , <n ) cannot contain a Souslin subtree of
height + +
n . So in particular, every subset of this tree of cardinality n contains
an antichain of equal size. Using this, and arguing as in the proof of Theorem
7.4.5, one sees that an application of the DushnikMiller theorem gives us the
conclusion. 
Denition 7.6.4. For < + and n < , set

f (n) = tp(Pn ()),

where Pn () = { : (, ) n }. Let

L() = {f : < + },

considered as a linearly ordered set with the lexicographical ordering.


Lemma 7.6.5. The linearly ordered set L() has an order-dense subset of size ,
and so in particular it contains no well-ordered subset of size + .
Proof. For each nite sequence t of ordinals < let (t) < + be chosen in the
following manner. If there is < + such that f is the lexicographically minimal
member of L() end-extending t we let (t) be equal to this . Otherwise, let (t)
be the minimal < + such that f end-extends t, and if such an cannot be
found, let (t) = 0. Then {f(t) : t < } is an order-dense subset of L() of
size . 
222 Chapter 7. Square Sequences

The following result shows that, while L() contains no well-ordered subset
of size + , every subset of L() of size smaller than + is the union of countably
many well-ordered subsets.
Lemma 7.6.6. For every < + , the set L () = {f : < } can be decomposed
into countably many well-ordered subsets.

Proof. Let Ln () = {f : Pn ()} for n < . Note that the projection


f  f  (n + 1) is one-to-one and lexicographically order preserving on Ln ().
It follows that each Ln () is a lexicographically well-ordered set. 

Example 7.6.7. Note that

K() = {{(n, f (n)) : n < } : < + }

is a family of countable subsets of which has the property that

K()  X = {K X : K K}

has size |X|+0 for every X of size < . In other words, when cf() = ,
the function : [+ ] is giving us a particular example of a Kurepa family on
. We now give an application of the existence of such a family of countable sets.
This requires a few standard notions from measure theory and real analysis.
Denition 7.6.8. A Radon measure space is a Hausdor topological space X with
a -nite measure dened on a eld of subsets of X containing open sets with
the property that the measure of every set is equal to the supremum of measures
of its compact subsets. When (X) = 1, then (X, ) is called Radon probability
space. Finally recall that the Maharam type of (X, ) is the least cardinality of a
subset of the measure algebra of which completely generates the algebra.
Denition 7.6.9. A directed set D is Tukey reducible to a directed set E, written
as D E, if there is f : D E which maps unbounded sets to unbounded sets.
In case D and E are ideals of subsets, this is equivalent to saying that there is
q : E D which is monotone and has conal range. We say that D and E are
Tukey equivalent, D E, whenever D E and E D.

We shall study Tukey reducibility in a class of directed sets that includes the
particular example of a directed set, the directed set [] of all countable subsets
of . The following turns out to be an important notion in this study.
Denition 7.6.10. A subset K of [] is locally countable if {x K : x a} is
countable for all a [] .

As a closely related notion, we have the following well-known classical concept of


local smallness introduced by Kurepa in [66].
7.6. The function on successors of singular cardinals 223

Denition 7.6.11. A subset K of [] is called a Kurepa family, or in short K-family


if the restriction
K  a = {x a : x K}
is countable for every countable subset a of .
Clearly, every Kurepa family is locally countable but not vice versa. But these two
notions are closely related as the following fact shows.
Lemma 7.6.12. If [] admits a locally countable family of some size , then it also
admits a Kurepa family of size .
Proof. Let K be a xed locally countable family of countable and innite subsets
of some innite cardinal and let = |K|. Let S be the set < of nite sequences
of ordinals < . Since S is equinumerous with , it suces to produce a Kurepa
family K [S] of cardinality . For x K, x an onto map fx : x, and
set x = {fx  n : n < }. Finally, set

K = {x : x K}.

Since x  x is a bijection between K and K , the two families are equinumerous.


To see that K satises Kurepas condition consider a countable subset a of S.
Replacing a by a bigger countable subset of S, we may assume that a is closed
under maps t  t  n (n < ). Let a be the countable subset of consisting of all
ordinals mentioned in a nite sequence t appearing in a . Let

Ka = {x K : x a}.

Then by our assumption about K, the family Ka is countable. On the other hand,
if an element x of K has an innite intersection with a , then it is actually a
subset of a . It follows that

{x a : x K } {x : x Ka } [a ]< .

It follows that {x a : x K } is a countable family, which was to be shown.



Here is a typical property that distinguishes between these two closely related
notions.
Lemma 7.6.13. If there is a locally countable subset of [] of size cf[] , then
there is a locally countable family K [] that is moreover conal6 in [] .
Proof. Fix a locally countable H [] of cardinality cf[] . It follows that there
is a mapping : H [] such that
(1) a (a) for every a H,
(2) K = {(a) : a H} is conal in [] .
6K [] is conal in [] if for every a [] there is x K such that a x.
224 Chapter 7. Square Sequences

Then for every countable subset b of ,

{x K : x b} {(a) : a H, a b},

so by our assumption that H is locally countable, we conclude that {x K : x b}


is also countable. Thus K is a conal locally countable subset of [] . 
Corollary 7.6.14. If for some innite cardinal we have that cf[] = , then
there is a conal locally countable subset of [] .
Proof. Any family of pairwise-disjoint countable subsets of satises the hy-
pothesis of Lemma 7.6.13. 
Corollary 7.6.15. For every positive integer n there is a conal locally countable
subset of [n ] . 
Here is a typical application of the concept of conal locally countable families
Theorem 7.6.16. If there is a locally countable subset of [] of size cf[] , or
equivalently, if there is a locally countable conal subset of [] , then the null ideal
of every atomless Radon probability space of Maharam type is Tukey equivalent
to the product N [] where N is the ideal of measure zero subsets of the unit
interval.
Proof. We shall give an argument for the case X = {0, 1} and the Haar measure
of X, referring the reader to the general result from [37] that the null ideal of any
atomless probability space of Maharam type is Tukey equivalent to the null
ideal N of {0, 1}. The only nontrivial direction is to produce a Tukey map from
N into N [] . Fix a locally countable conal family C [] . For c C
x a bijection ec : c . This gives us a way to dene a function c mapping
cylinders of {0, 1} with supports included in c into 2 in the natural way. Let Ncl
be the family of all null subsets of {0, 1} that are countably supported cylinders
of {0, 1}. Clearly, Ncl is conal in N , so it suces to produce a Tukey map

f : Ncl N [] .

Given N Ncl , nd c(N ) C such that supp(N ) c(N ) and put

f (N ) = (c(N )  N, c(N )).

It is straightforward to check that f maps unbounded subsets of Ncl to unbounded


subsets of the product N [] . 
Remark 7.6.17. Tukey theory of this kind originated in the two papers of Isbell [50]
and [49] though Theorem 7.6.16 is due to Fremlin [37]. Note that the hypothesis of
Fremlins theorem is true when [] has conality and so it is always satised for
< . By Example 7.6.7 above, the hypothesis of Theorem 7.6.16 is satised for
all if one assumes  and cf[] = + for every of countable conality. In fact,
7.6. The function on successors of singular cardinals 225

under these assumptions, we get a Kurepa family of cardinality cf[] = + . Note,


however, that even if K [] is assumed to be a Kurepa family of size cf[] ,
the transfer to a conal locally countable C = {a h(a) : a K} seen above does
not necessarily give us that C is a Kurepa family. To obtain a Kurepa family that
is conal in [] , one needs to work a bit harder and use stronger assumptions.
Denition 7.6.18. If a family K [S] is at the same time a Kurepa family and
conal in [S] , then we call it a conal Kurepa family (conal K-family in short).
Two conal K-families H and K are compatible if H K H K for all H H
and K K. We say that K extends H if they are compatible and if H K.
Remark 7.6.19. Note that the size of any conal K-family K on a set S is equal
to the conality of [S] . Note also that for every X S there is Y X of size
cf[X] such that K Y K for all K K.
Denition 7.6.20. Dene CK() to be the statement that every sequence Kn (n <
) of comparable conal K-families with domains included in which are closed
under , and \ can be extended to a single conal K-family on , which is also
closed under these three operations.
Lemma 7.6.21. CK(1 ) is true, and if CK() is true for some such that cf[] =
, then CK(+ ) is also true.
Proof. The easy proof of CK(1 ) is left to the reader, so we concentrate on the
rest of the conclusion of the lemma.
Suppose CK() and let Kn (n < ) be a given sequence of compatible conal
K-families as in the hypothesis of CK(+ ). By Remark 7.6.19 there is a strictly
increasing sequence ( < + ) of ordinals < such that Kn  Kn for all
< + and n < . Recursively on < + we construct a chain H ( < + ) of
conal K-families as follows. If = 0 or = + 1 for some , using CK( ) we
can nd a conal K-family H on extending H 1 and Kn  (n < ). If has
uncountable conality, then the union of H =
< H is a conal K-family with
domain included in , so using CK( ) we can nd a conal K-family H on
extending H and Kn  (n < ). If has countable conality, pick a sequence
{n } converging to and use CK( ) to nd a conal K-family H on extending
Hn (n < ) and Kn (n < ). When the recursion is done, set H = <+ H .
Then H is a conal K-family on + extending Kn (n < ). 
Corollary 7.6.22. For each n < there is a conal Kurepa family on n . 
Denition 7.6.23. Let be a cardinal of conality . A Jensen matrix on + is a
matrix Jn ( < + , n < ) of subsets of + with the following properties, where
n (n < ) is some xed increasing sequence of cardinals converging to :
(1) |Jn | n for all < + and n < ,
(2) for all < and n < there is m < such that Jn Jm ,
  
(3) n< [Jn ] = < n< [Jn ] whenever cf() > ,
 
(4) [+ ] = <+ n< [Jn ] .
226 Chapter 7. Square Sequences

Remark 7.6.24. The notion of a Jensen matrix is the combinatorial essence behind
Silvers proof of Jensens model-theoretic two-cardinal transfer theorem in the
constructible universe (see [52, appendix]), so the matrix could equally well be
called Silver matrix. It has been implicitly or explicitly used in several places in
the literature. The reader is referred to the paper of Foreman and Magidor [35]
which gives a quite complete discussion of this notion and its occurrence in the
literature.
Lemma 7.6.25. Suppose some cardinal of countable conality carries a Jensen
matrix Jn ( < + , n < ) relative to some sequence of cardinals n (n < )
that converge to . If CK(n ) holds for all n < , then CK(+ ) is also true.
Proof. Let Kn (n < ) be a given sequence of compatible conal K-families with
domains included in + . Given Jn , there is a natural continuous chain Jn ( <
1 ) of subsets of+ of size n such that J0n = Jn and +1
 Jn equal to the

union of all K n< Kn which intersect Jn . Let Jn = <1 Jn . It is easily

seen that Jn ( < + , n < ) is also a Jensen matrix. By recursion on and


n we dene a sequence Hn ( < + , n < ) of compatible conal K-families as
follows. If = + 1 or = 0 and n < , using CK(n ) we can nd a conal
K-family Hn with domain Jn compatible with Hm (m < n), H(1)m (m < )
and Km  Jn (m < ). If cf() = let n (n < ) be an increasing sequence of
ordinals converging to . Using CK(n ) we can nd a conal K-family Hn which
extends Hm (m < n), Km  Jn (m < ) and each of the families Hi k (i <
, k < and Ji k Jn ). Finally suppose that cf() > . For n < , set
 
Hn = [Jn ] ( Hm ).
< m<

Using the properties of the Jensen matrix (especially (3)) as well as the compati-
bility of Hm ( < ,m < ) one easily checks that Hn is a conal K-family with
domain Jn which extends each member of Hm (m < n) and Km  Jn (m < )
and which is compatible with all of the previously constructed families Hm ( <
, m < ). When the recursion is done we set
 
H= Hn .
<+ n<

Using the property (4) of Jn ( < + , n < ), it follows easily that H is a conal
K-family on + extending Kn (n < ). 
Theorem 7.6.26. If a Jensen matrix exists on any successor of a cardinal of co-
nality , then a conal Kurepa family exists on any domain. 
It should be noted that the -function : [+ ]2 associated with a
 -sequence C ( < + ) for some singular cardinal of conality leads to the
matrix
Fn () = { < : (, ) n}( < + , n < ) (7.6.2)
7.6. The function on successors of singular cardinals 227

which has all the properties (1)(3) of Denition 7.6.23 as well as some other
properties not captured by the denition of a Jensen matrix, so let us expose
some of them. The additional power of the -matrix is most easily seen if one
additionally has a sequence a ( < + ) of countable subsets of + that is conal
in [+ ] . Then one can extend the matrix (7.6.2) as follows:

Mn = (a {}) ( < + , n < ). (7.6.3)
<n

(Recall that <n is the tree-ordering on + dened by the formula <n i


(, ) n where n is a xed increasing sequence of cardinals converging to .)
The matrix
Mn ( < + , n < )
has properties not captured by Denition 7.6.23 that are of independent interest,
so let us state them in a separate lemma.
Lemma 7.6.27.

(1) <n implies Mn Mn .


(2) Mm Mn whenever m < n.

(3) If = sup{ : <n } then Mn = <n Mn .
+
(4) Every countable subset of is covered by some Mn .
+
(5) M = {Mn : < , n < } is a locally countable family if we have started
with a locally countable K = {a : < + }. 
Remark 7.6.28. One can think of the matrix M = {Mn : < + , n < } as a
version of a morass for the singular cardinal (see [127]). It would be interesting
to see how far one can go in this analogy. We give a few applications just to
illustrate the usefulness of the families we have constructed so far.
Denition 7.6.29. A Bernstein decomposition of a topological space X is a function
f : X 2N with the property that f takes all the values from 2N on any subset
of X homeomorphic to the Cantor set.
Remark 7.6.30. The classical construction of Bernstein [14] can be interpreted by
saying that every space of size at most continuum admits a Bernstein decomposi-
tion. For larger spaces one must assume Hausdors separation axiom, a result of
Nesetril and Rodl (see [80]). In this context Malykhin was able to extend Bern-
steins result to all spaces of size < c+ (see [73]). To extend this to all Hausdor
spaces, some use of square sequences seems natural. In fact, the rst Bernstein
decompositions of an arbitrary Hausdor space were constructed using  and
= + for every > c of conality by Weiss [132] and Wolfsdorf [134]. We
shall now see that conal K-families are quite natural tools in constructions of
Bernstein decompositions.
228 Chapter 7. Square Sequences

Theorem 7.6.31. Suppose every regular cardinal > c supports a conal Kurepa
family of size . Then every Hausdor space admits a Bernstein decomposition.

Proof. Let us say that a subset S of a topological space X is sequentially closed


in X if it contains all limit points of sequences {xn } S that converge in X. We
say S is -sequentially closed in X if it can be decomposed into countably many
sequentially closed sets. For a given cardinal , let B() denote the statement
that for every Hausdor space X of size , every -sequentially closed subset
S of X and every Bernstein decomposition f : S 2N , there is a Bernstein
decomposition g : X 2N extending f . Using the fact that a Hausdor space of
size c has at most c copies of the Cantor set and following the idea of Bernsteins
original proof, one sees that B(c) holds. Note also that by our assumption 0 =
for every cardinal c of uncountable conality. So if we have B() for such
, one proves B(+ ) by constructing a chain f ( < + ) of partial Bernstein
decompositions on any Hausdor topology on + , using the fact that the set C
of all < + which are -sequentially closed in (+ , ) is closed and unbounded
in + . Thus if ( < + ) enumerates C increasingly, we will have g : 2N
and g compatible with the given partial Bernstein decomposition f of a given
-sequentially closed set S + . Note that S will be -sequentially closed,
so the use of B( ) to get f will be possible. This inductive procedure encounters
a crucial diculty only when = + for some > c of conality . This is where
a conal K-family K on + is useful. Let be a given Hausdor topology on +
and for K K let K  be the collection of all limits of -convergent sequences
of elements of K. Note that if a set + is K-closed in the sense that K 
K (i.e., K K for all K K), then  will be equal to the union of
all K  (K K  ). So if || > c is a cardinal of uncountable conality, then
| | = ||. It follows that every < + with the property that = lim n for some
sequence n (n < ) of ordinals with the property that K  n K for all n
can be decomposed into countably many sets n (n < ) such that for each n,
the set n is -sequentially closed in and |n | is a cardinal > c of uncountable
conality. However, note that the set C of all such < + is closed and unbounded
in + so as above we use our inductive hypothesis B() ( < ) to construct a
chain g : 2N ( < + ) of Bernstein decompositions. This completes the
proof. 

It is interesting that various less pathological classes of spaces admit a local


version of Theorem 7.6.31.
Theorem 7.6.32. A sigma-product7 of the unit interval admits a Bernstein decom-
position if its index set carries a conal Kurepa family. So in particular, any metric
space that carries a conal Kurepa family admits a Bernstein decomposition.

7 Recall
that a sigma-product is the subspace of some Tychono cube [0, 1] consisting of all
mappings x such that supp(x) = { : x() = 0} is countable. The set is its index-set .
7.6. The function on successors of singular cardinals 229

Proof. So let K be a xed well-founded conal K-family on and let <w be a


well-ordering of K compatible with . For K K x a Bernstein decomposition
fK : XK 2N where XK = {x X : supp(x) K}. Dene f : X 2N by
letting f (x) = fK (x) where K is the <w -minimal K K such that supp(x) K.
It is easily checked that f is a Bernstein decomposition of the sigma-product. 

Denition 7.6.33. Recall the notion of a coherent family of partial functions in-
dexed by some ideal I, a family of the form fa : a (a I) with the property
that {x a b : fa (x) = fb (x)} is nite for all a, b I.

It can be seen (see [122]) that the P-ideal dichotomy (see Denition 2.4.11)
has a strong inuence on such families provided I is a P-ideal of countable subsets
of some set . The following result is one way to express this inuence.
Theorem 7.6.34. Assuming the P-ideal dichotomy, for every coherent family of
functions fa : a (a I) indexed by some P-ideal I of countable subsets of
some set , either

(1) there is an uncountable such that fa  is nite-to-one for all a I,


or
(2) there is g : such that g  a = fa for all a I.

Proof. Let L be the family of all countable subsets b of for which one can nd
an a in I such that b \ a is nite and fa is nite-to-one on b. To see that L is a
P-ideal, let {bn } be a given sequence of members of L and for each n x a member
an of I such that fan is nite-to-one on bn . Since I is a P-ideal, we can nd a I
such that an \ a is nite for all n. Note that for all n, bn \ a is nite and that fa
is nite-to-one on bn . For n < , let

bn = { bn a : fa () > n}.

Then bn is a conite subset of bn for each n, so if we set b to be equal to the union


of the bn s, we get a subset of a which almost includes each bn and on which fa
is nite-to-one. It follows that b belongs to L. This nishes the proof that L is
a P-ideal. Applying the P-ideal dichotomy to L, we get the two alternatives that
translate into the alternatives (1) and (2) of Theorem 7.6.34. 

This leads to the natural question whether for any set one can construct a
family {fa : a } of nite-to-one mappings indexed by [] . This question was
answered by Koszmider [59] using the notion of a Jensen matrix discussed above.
We shall present Koszmiders result using the notion of a conal Kurepa family
instead.
Theorem 7.6.35. If carries a conal Kurepa family, then there is a coherent
family fa : a (a [] ) of nite-to-one mappings.
230 Chapter 7. Square Sequences

Proof. Let K be a xed well-founded conal K-family on and let <w be a well-
ordering of K compatible with . It suces to produce a coherent family of nite-
to-one mappings indexed by K. This is done by induction on <w . Suppose K K
and fH : H is determined for all H K with H <w K. Let Hn (n < ) be
a sequence of elements of K that are <w K and have the property that for every
H K with H <w K there is n < such that H K = Hn K. So it suces
to construct a nite-to-one fK : K which coheres with each fHn (n < ), a
straightforward task. 
Corollary 7.6.36. For every non-negative integer n there is a coherent family

fa : a (a [n ] )

of nite-to-one mappings. 
Remark 7.6.37. It is interesting that nite-to-one cannot be replaced by one-
to-one in these results. For example, there is no coherent family of one-to-one
mappings fa : a (a [c+ ] ). We nish this section with a typical application
of coherent families of nite-to-one mappings discovered by Scheepers [87].
Theorem 7.6.38. If there is a coherent family fa : a (a [] ) of nite-to-
one mappings, then there is a mapping

S : [[] ]2 []<
 
with the property that n< an n< S(an , an+1 ) for every strictly -increasing
sequence an (n < ) of countable subsets of .
Proof. For a [] let xa : be dened by letting

xa (n) = |{ a : fa () n}|.

Note that xa is eventually dominated by xb whenever a is a proper subset of b.


Choose : with the property that x < y implies (y) < (x),
where < is the ordering of eventual dominance8 on . Dene another family of
functions ga : a (a [] ) by letting

ga () = (xa )(fa ()). (7.6.4)

Note the following interesting property of ga (a [] ):

S(a, b) = { a : gb () ga ()} is nite for all a  b in [] . (7.6.5)

Let an (n < ) be a given strictly -increasing sequence of countable subsets of


. Fix an ordinal belonging to some an . Then the sequence of integers
(
gan () n n < )
8 Dened by, x < y if and only if x(n) < y(n) for all but nitely many ns.
7.6. The function on successors of singular cardinals 231

cannot be strictly decreasing, so there must be some n n with the property that
ga ().
gan () In other words, for every n < and an there exists n n

n+1
such that S(an , an+1 ), as required. 
Remark 7.6.39. Note that if is a singular cardinal of conality with the prop-
erty that cf[] < for all < , then the existence of a conal Kurepa family
on + implies the existence of a Jensen matrix on + . So these two notions ap-
pear to be quite close to each other. The three basic properties of the function
: [+ ] (Lemmas 7.3.8 and 7.3.7(a),(b)) seem much stronger in view of the
fact that the linear ordering as in Lemma 7.6.6 cannot exist for above a super-
compact cardinal and the fact that Foreman and Magidor [35] have produced a
model with a supercompact cardinal that carries a Jensen matrix on any successor
of a singular cardinal of conality . Note that Changs conjecture of the form
(+ , )  (1 , ) for some singular of conality implies that every locally
countable family K [] must have size . So, one of the models of set theory
that has no conal K-family on, say +1 , is the model of Levinski, Magidor and
Shelah [70]. It seems still unknown whether the conclusion of Theorem 7.6.38 can
be proved without additional set-theoretic assumptions.
Chapter 8

The Oscillation Mapping and


the Square-bracket Operation

8.1 The oscillation mapping


In what follows, will be a xed regular innite cardinal.

osc : P()2 Card (8.1.1)

is dened by
osc(x, y) = |x \ (sup(x y) + 1)/ |, (8.1.2)

where is the equivalence relation on x \ (sup(x y) + 1) dened by letting


i the closed interval determined by and contains no point from y. Hence,
osc(x, y) is simply the number of convex pieces the set x \ (sup(x y) + 1) is split
by the set y (see Figure 8.1). Note that this is slightly dierent from the way we
have dened the oscillation between two subsets x and y of 1 in Section 2.3 above,
where osc(x, y) was the number of convex pieces the set x is split by into the set
y \ x. Since the variation is rather minor, we keep the same old notation as there is
no danger of confusion. The oscillation mapping has proven to be a useful device in
various schemes for coding information. Its usefulness in a given context depends
very much on the corresponding oscillation theory, a set of denitions and lemmas
that disclose when it is possible to achieve a given number as oscillation between
two sets x and y in a given family X . The following denition reveals the notion
of largeness relevant to the oscillation theory that we develop in this section.
Denition 8.1.1. A family X P() is unbounded if for every closed and un-
bounded subset C of there exist x X and an increasing sequence {n : n <
} C such that sup(x n ) < n and [n , n+1 ) x = for all n < .
234 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

This notion of unboundedness has proven to be the key behind a number of results
asserting the complex behavior of the oscillation mapping on X 2 . The case =
seems to contain the deeper part of the oscillation theory known so far (see [110],
[111, 1] and [121]), though in this section we shall only consider the case > .
We shall also restrict ourselves to the family K() of all closed bounded subsets
of rather than the whole power-set of . The following is the basic result about
the behavior of the oscillation mapping in this context.

Figure 8.1: The oscillation mapping on the power-set of .

Lemma 8.1.2. If X is an unbounded subfamily of K(), then for every positive


integer n there exist x and y in X such that osc(x, y) = n.
Proof. Choose an -chain M of length of elementary submodels M of some large
enough structure of the form H such that X M and M . Applying the
fact that X is unbounded with respect to the closed and unbounded set {M :
M M}, we can nd y X such that:
i = M i
/ y for all i < n, (8.1.3)
0 y = and y \ n1 = , (8.1.4)
[i1 , i ) y = for all 0 < i < n. (8.1.5)
Let
J0 = [0, sup(y 0 )], (8.1.6)
Ji = [i1 , max(y i )] for 0 < i < n, (8.1.7)
Jn = [n1 , max(y)]. (8.1.8)
Let Fn be the collection of all increasing sequences I0 < I1 < < In of closed
intervals of such that I0 = J0 and such that there is x = xI in X such that
x I0 I1 In , (8.1.9)

max(Ii ) x for all i n. (8.1.10)


8.1. The oscillation mapping 235

Clearly, Fn M0 and J0 , J1 , . . . , Jn  Fn . Let Fn1 be the collection of all


I0 , . . . , In1  such that for every < there exists an interval I of above
such that I0 , . . . , In1 , I Fn . Note that Fn1 M0 and that, using the
elementarity of Mn1 , one proves that J0 , . . . , Jn1  Fn1 . Let Fn2 be the
collection of all I0 , . . . , In2  such that for every < there exists interval I
above such that I0 , . . . , In2 , I Fn1 . Then again Fn2 M0 and, using
the elementarity of Mn2 , one shows that the restriction J0 , . . . , Jn2  belongs to
Fn2 . Continuing in this way we construct families Fn , Fn1 , . . . , F0 such that:

Fi M0 for all i n, (8.1.11)

F0 = {J0 }. (8.1.12)
I0 , . . . , Ini  Fni implies that for all < there exists
(8.1.13)
an interval I above such that I0 , . . . , Ini , I Fni+1 .
Recursively in i n we choose intervals I0 < I1 < < In such that:

I0 , . . . , Ii  Fi for all i n, (8.1.14)

Ji1 < Ii Mi1 for 0 < i n. (8.1.15)


Clearly, there is no problem in choosing the sequence using the elementarity of
the submodels Mi1 and the condition (8.1.13). By the denition I0 , . . . , In 
Fn means that there is x X satisfying (8.1.9) and (8.1.10). It follows that
max(x y) = max(I0 ) and that

x Ii = (0 < i n) (8.1.16)

are the convex pieces of x \ (max(x y) + 1) into which this set is split by y. It
follows that osc(x, y) = n. This nishes the proof. 
Lemma 8.1.2 also has a rectangular form.
Lemma 8.1.3. If X and Y are two unbounded subfamilies of K(), then for all
but nitely many positive integers n there exist x X and y Y such that
osc(x, y) = n. 
Recall the notion of a nontrivial C-sequence C ( < ) on from Section
6.3, a C-sequence with the property that, for every closed and unbounded subset
C of , there is a limit point of C such that C  C for all < .
Denition 8.1.4. For a subset D of let lim(D) denote the set of all < with
the property that = sup(D ). A subsequence C ( ) of some C-sequence
C ( < ) is stationary if the union of all lim(C ) ( ) is a stationary subset
of .
Lemma 8.1.5. A stationary subsequence of a nontrivial C-sequence on is an
unbounded family of subsets of .
236 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Proof. Let C ( ) be a given stationary subsequence of a nontrivial


C-sequence on . Let C be a given closed and unbounded subset of . Let
be the union of all lim(C ) ( ). Then is a stationary subset of . For
choose such that lim(C ). Applying the assumption that C ( )
is a nontrivial C-subsequence, we can nd lim(C) such that

C [, )  C for all < . (8.1.17)

If such cannot be found using the stationarity of the set lim(C), we would be
able to use the Pressing Down Lemma on the regressive mapping that would give
us an < violating (8.1.17) and get that a tail of C trivializes C ( ). Using
(8.1.17) and the fact that C is unbounded in , we can nd a strictly increasing
sequence n (n < ) of elements of (C ) \ C such that [n , n+1 ) C =
for all n. So the set C satises the conclusion of Denition 8.1.1 for the given
closed and unbounded set C. 
Notation 8.1.6. Recall that Q denotes the set of all nite sequences of ordinals
< and that we consider it totally ordered by the relation of right lexicographical
ordering.1 We need the notation for the following two further orderings on Q :

s  t if and only if s is an initial segment of t. (8.1.18)

s < t if and only if s is a proper initial segment of t. (8.1.19)


Denition 8.1.7. Given a C-sequence C ( < ) we can dene a partial action
(, t)  t of Q on recursively on the ordering  of Q as follows:

= , (8.1.20)

 = the th member of C if < tp(C ), (8.1.21)


 is undened if tp(C ), (8.1.22)
t  = (t ) . (8.1.23)
Whenever we write t , we implicitly assume that this ordinal is dened.
Remark 8.1.8. Note that if 0 (, ) = t for some < < , then t = . In fact,
if = 0 > > n = is the walk from to along the C-sequence, each
member of the trace Tr(, ) = {0 , 1 , . . . , n } has the form s where s is the
uniquely determined initial part of t. Note, however, that in general t = does
not imply that 0 (, ) = t.
Notation 8.1.9. Given a C-sequence C ( < ) on we shall use the shorter
notation
osc(, ) = osc(C , C )
for the oscillation between two members of the C-sequence.
1 Recall the notion of right lexicographical ordering: s <r t i either t is an initial segment of s
or else s() < t() for = min{ : s() = t()}.
8.1. The oscillation mapping 237

We are now ready to dene the higher cardinal analogue of the basic oscilla-
tion mapping on 1 exposed above in Section 2.3.
Denition 8.1.10. To every C-sequence C ( < ) we associate an oscillation
mapping
o : []2
as follows. Given a pair of ordinals < < , if there is t  0 (, ) such that
t is dened and
(i) osc(t , t ) > 1, but
(ii) osc(s , s ) = 1 for all s < t,
let o(, ) = osc(t , t ) 2; otherwise, let o(, ) = 0.

Figure 8.2: The oscillation mapping o.

Theorem 8.1.11. If C ( < ) is a nontrivial C-sequence on and if

o : []2

is the corresponding oscillation mapping, then for every unbounded set and
every non-negative integer n there exist < in such that o(, ) = n.
238 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Proof. Choose an elementary submodel M of some large enough structure of the


form H such that = M and M contains all the relevant objects. Choose
above and let
= 0 > 1 > > k =
be the minimal walk from to along the C-sequence. Let k = k 1 if Ck1 is
unbounded in ; otherwise k = k. Recall our implicit assumption about essentially
all C-sequences that we consider here:
(c) If is a limit ordinal, then an ordinal C occupying a successor place in
C must be a successor ordinal.
In this case, since is a limit ordinal and since Ck1 , either

(1) k1 is a limit ordinal and = sup(Ck1 ),


or
(2) k1 = + 1.
In any of the two cases, it follows that k is a limit ordinal. Set t = 0 (k , ).
Let

= t  0, t  1, . . . , t  k = t

be our notation for all the restrictions of the sequence t. Note that i = ti for
i k. Let 0 be the set of all such that:

= > t1 > > t , (8.1.24)

sup(Cti t ) = sup(Ci k ) for all i < k.


(8.1.25)
Note that 0 M and 0 . Let = {t : 0 }. Then M and k .
By the very choice of k, the ordinal belongs to lim(C ), and so, using the
k

elementarity of M , one concludes that C ( ) is a stationary subsequence of
our original sequence. Shrinking a bit if necessary, we may assume that for every
an ordinal 0 such that t = can be found below the next member of
above . By Lemma 8.1.5, the sequence C ( ) forms an unbounded family
of subsets of , and so we can apply the proof of our basic oscillation Lemma 8.1.2
and nd an ordinal < of the form t for some 0 such that

osc(t , t ) = n + 2.

Using the fact that < and the fact that both ordinals and satisfy (8.1.25),
we conclude that osc(ti , ti ) = 1 for all i < k.
It follows that o(, )
= n. This
nishes the proof. 
Remark 8.1.12. The class of all regular cardinals that carry a nontrivial
C-sequence is quite extensive. It includes not only all successor cardinals but also
some inaccessible as well as hyperinaccessible cardinals such as for example, rst
inaccessible or rst Mahlo cardinals. In other words, we would like to point out
that Theorem 8.1.11 has corollaries of the following kind.
8.1. The oscillation mapping 239

Corollary 8.1.13. If is the rst weakly inaccessible uncountable cardinal, then


there is a mapping
o : []2
which takes all the values from on any set of the form []2 for an unbounded
subset of . 
In view of the well-known Ramsey-theoretic characterization of weak compactness,
the statement of Theorem 8.1.11 leads us now to the following natural question.
Question 8.1.14. Can the weak compactness of a strong limit regular uncountable
cardinal be characterized by the fact that for every mapping f : []2 there
exists an unbounded set such that f  []2 = ? This is true when is
replaced by 2, but can any other number besides 2 be used in this characterization?
We shall now see how to adapt the basic idea of the proof of Theorem 8.1.11
in order to establish the following more informative and potentially more useful
property of the oscillation mapping o.
Theorem 8.1.15. If a regular uncountable cardinal carries a nontrivial C-se-
quence, then the corresponding oscillation mapping o : []2 has the property
that for every family A of pairwise-disjoint nite subsets of , all of some xed
size n, there exist a subfamily B A of size , a mapping h : n n , and an
equivalence relation on the index-set n = {0, 1, . . . , n 1} such that

o(a(i), b(j)) = h(i, j) for all a < b in B and i  j.

Moreover, for every subfamily C B of size there is g : n n which agrees


with h on non-equivalent pairs such that every integer m there exist a < b in C
such that

o(a(i), b(j)) = g(i, j) + m for all i, j < n such that i j.

Proof. As indicated above, we will build on the proof of Theorem 8.1.11 so a


familiarity with that proof will be assumed below. As before we choose an ele-
mentary submodel M of some large enough structure of the form H such that
= M and M contains all the relevant objects. Let b A be minimal such
that all of its ordinals are strictly above . For b, consider the minimal walk

= 0 () > > k() () =



along the xed nontrivial C-sequence C ( < ). Let k() = k() 1 if

sup(Ck()1 () ) = ;


otherwise, put k() = k(). Let t = 0 (k()
, ). Finally, for i, j < n, set i b j
if and only if when we set = b(i) and = b(j), then we have:
240 Chapter 8. The Oscillation Mapping and the Square-bracket Operation


(1) k() = k(), k()
= k() and t = t ,
(2) sup(Cl () ) = sup(Cl () ) for all l < k.
Using the elementarity of M we can nd a stationary set and for each
a member b of A above , a xed equivalence relation on {0, 1, . . . , n 1} and
for each equivalence class e of an integer ke , a sequence te of ordinals < , and
a sequence l (l < ke ) such that for each and for each , b occupying
ith and jth place, respectively, with i j, we have:

(3) k() = k(), k()
= k() and t = t = te ,
(4) sup(Cl () ) = sup(Cl () ) = l for all l < ke ,
where e denotes the equivalence class of i and j. Shrinking , we may assume that
for < in , we have that b . Let B = {b : }. By our denition of
the mapping o, it follows that for a < b in B,

o(a(i), b(j)) = 0 whenever i  j. (8.1.26)

For and i < n, let D (i) = Ck () , where b occupies the ith place.
Note that each D (i) is a closed and unbounded subset of . Clearly, we may
assume that all these objects belong to our submodel M , that = M belongs
to and that b = b . Using a sequence of continuous -chains Mp of elementary
submodels of H , such that Mp M and such that Mp min Mp+1 , working as
in the proof of Lemma 8.1.5 and applying the assumption about the nontriviality
of the C-sequence C ( < ), we can nd an -chain Mp (p < ) of elementary
submodels of H containing the relevant objects such that p = Mp , and

sup(D (i) p ) < p and D (i) [p , p+1 ) = for all i < n and p < . (8.1.27)

Dene h : n n by letting h(i, j) = 0 for i  j and let

h(i, j) = osc(D (i) 0 , D (j) 0 ) for i j. (8.1.28)

Let m be a given positive integer. Using (8.1.27) and the elementarity of the rst
m + 2 of our submodels Mp (p m + 1), we can nd an a = b B M that add
m + 2 more oscillations to the existing ones, or in other words,

o(a(i), b(j)) = osc(D (i) 0 , D (j) 0 ) + m whenever i j. (8.1.29)

It is clear that using the elementarity of M0 , we can also arrange that

osc(D (i) 0 , D (j) 0 ) = osc(D (i) 0 , D (j) 0 ) whenever i j. (8.1.30)

Referring to the denition of h, we infer that o(a(i), b(j)) = h(i, j) + m for all
i, j < n such that i j, as required. This proves the lemma for C = B. The
general case reduces to this one by getting rst B A of the particular form
B = {b : } as above for which the corresponding equivalence relation
cannot be made ner by going to such a further subset of B. 
8.1. The oscillation mapping 241

Denition 8.1.16. The mapping o of Theorem 8.1.15 has particularly useful pro-
jections when the cardinal is not bigger than the continuum. In order to dene
this projection, we need a one-to-one sequence r ( < ) of elements of 2 , and
a list hl (l < ) of all mappings of the form h : 2n 2n , where n = n(h) < ,
in such a way that for every such mapping h the set Ph = {l : hl = h} contains an
innite arithmetic progression. Then the projection of o is dened as

o (, ) = ho(,) (r  n(ho(,) ), r  n(ho(,) )).

Then Theorem 8.1.15 has the following immediate consequence deduced from it in
a similar manner as in the case of the oscillation mapping and the square-bracket
operation on 1 (see Chapter 5).
Theorem 8.1.17. Suppose that a regular uncountable cardinal that is not bigger
than the continuum carries a nontrivial C-sequence. Then the projection

o : []2

of the oscillation mapping has the property that, for every family A of pairwise-
disjoint nite subsets of , all of some xed size n, there is a subfamily B A
of size , a mapping h : n n , and an equivalence relation on the index set
n = {0, 1, . . . , n 1} such that

o (a(i), b(j)) = h(i, j) for all a < b in B and i  j.

Moreover, for every subfamily C B of size and g : n n there exist a < b


in C such that

o (a(i), b(j)) = g(i) for all i, j < n such that i j. 

We nish this section with some typical applications of this projection. For
example, the following result is proved on the basis of Theorem 8.1.17 in a similar
manner as Theorem 5.3.3 above was proven on the basis of the corresponding
result about the square-bracket operation.
Theorem 8.1.18. Let be an index-set of uncountable regular cardinality not
bigger than the continuum and supporting a nontrivial C-sequence and let 1 () be
the complex version of the standard Banach space. Then there is a homogeneous
polynomial
P : 1 () C
that takes all the complex values on any closed subspace of 1 () of density . 
Corollary 8.1.19. Suppose that the continuum c is regular and not bigger than
the rst weakly inaccessible cardinal. Then the complex version of the space 1 (c)
admits a complex-valued homogeneous polynomial of degree 2 which takes all the
values on any closed subspace of 1 (c) of density continuum. 
242 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Let us now mention another application of Theorem 8.1.17 deduced from it


in a similar manner as Theorem 5.5.4 above.

Theorem 8.1.20. Suppose that a regular uncountable cardinal that is not bigger
than the continuum carries a nontrivial C-sequence. Let V and W be two vector
spaces over the same eld K such that V has dimension while W is countable.
Then there is a non-degenerate symmetric bilinear mapping : V V W such
that {(x, y) : x, y X, x = y} = W for every subset X of V of cardinality . 

It should be clear (and we will come to this point in the next section) that
Theorem 8.1.20 allows also its asymmetric version with essentially the same proof.
To state this version, recall that a bilinear map : V V W is alternating if
(x, y) = (y, x) and (x, x) = 0 for all x, y V .

Theorem 8.1.21. Suppose that a regular uncountable cardinal that is not bigger
than the continuum carries a nontrivial C-sequence. Let V and W be two vector
spaces over the same eld K such that V has dimension while W is countable.
Then there is a non-degenerate alternating bilinear mapping : V V W such
that {(x, y) : x, y X, x = y} = W for every subset X of V of cardinality . 

Corollary 8.1.22. Suppose that the continuum is regular and not bigger than the
rst weakly inaccessible cardinal. Let V and W be two vector spaces over the same
eld K such that V has dimension continuum while W is countable. Then there is
a non-degenerate alternating bilinear mapping : V V W such that {(x, y) :
x, y X, x = y} = W for every subset X of V of cardinality . 

Recall that a vector space V over a eld K equipped with a bilinear form
: V V K is called symplectic space if the bilinear form is alternating on
V . The symplectic space (V, ) is non-degenerate if for all x V there is y V
such that (x, y) = 0.

Theorem 8.1.23. Suppose that a regular uncountable cardinal that is not bigger
than the continuum carries a nontrivial C-sequence. Then for every countable
(including nite) eld K and for every vector space V over K of dimension there
is a symplectic space of the form (V, ) without a null subspace of cardinality . 

Corollary 8.1.24. Suppose that the continuum is regular and not bigger than the
rst weakly inaccessible cardinal. Then for every countable eld K and for every
vector space V over K of dimension continuum, there is a symplectic space of the
form (V, ) without a null subspace of cardinality . 

Remark 8.1.25. We remark again that the results of this section cover a large
class of regular cardinals that are not necessarily successor cardinals. In the
case of successors of regular cardinals it is more eective to use the square-bracket
operation that forms the subject matter of the following section.
8.2. The trace lter and the square-bracket operation 243

8.2 The trace lter and the square-bracket operation


A square-bracket operation on a regular uncountable cardinal carrying a nontrivial
C-sequence is a transformation which to every pair < of ordinals below
assigns a point [] belonging to the trace
= 0 > 1 > > 2 (,) =
of the walk from to . We want the operation to be chosen in such a way that
the set of all < of the form [] for and running inside the same (and
sometimes two dierent) unbounded subsets of contains a closed and unbounded
set relative to some xed stationary subset of . The set has to be chosen
carefully relative to the C-sequence on which we base our walks. We have seen
above that 1 admits this kind of operation (see Denition 5.1.3) and the purpose
of this section is to show that the basic idea extends to a general setting on
an arbitrary uncountable regular cardinal that carries a nontrivial C-sequence
C ( < ). Not surprisingly, the idea is based again on the oscillation map dened
in the previous section and, in particular, on the property of this map exposed in
Theorem 8.1.11. Recall also the denition of the partial action (, t)  t , dened
recursively by = and
t = th member of Cs
for t = s , if is smaller than the order type of Cs ; otherwise, we leave t
undened.
Denition 8.2.1. For < < we set [] = t , where t  0 (, ) is minimal
for which t is dened and
th element of Ct = th element of Ct , (8.2.1)
where = tp(Ct ); if such t does not exist, we let [] = .
Remark 8.2.2. Thus, [] is the rst place m visited by on its walk down to
such that a nontrivial oscillation occurs between Cm and Cm , where m is the
corresponding place of a walk that starts with and has the same full code up to
that point. Note that Theorem 8.1.11 is telling us that, in particular, the nontrivial
oscillation indeed happens most of the time. Results that would say that the set
of values {[] : {, } []2 } is in some sense large no matter how small the
unbounded set is, would correspond to the results of Lemmas 5.1.45.1.5
about the square-bracket operation on 1 . It turns out that this is indeed possible
and to describe it we need the following denition.
Denition 8.2.3. A C-sequence C ( < ) on avoids a given subset of if
C = for all limit ordinals < .
Lemma 8.2.4. Suppose C ( < ) is a given C-sequence on that avoids a set
. Then for every unbounded set , the set of elements of not of the
form [] for some < in is nonstationary in .
244 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Figure 8.3: The square-bracket operation.

Proof. Let be a given stationary subset of . We need to nd < in such


that [] . For a limit x a = () above and let

= 0 > 1 > > k() =

be the walk from to along the sequence C ( < ). Since / C for any
limit and since C+1 = {} for all , we must have that k()1 = + 1. In
particular,
Ci is bounded in for all i < k(). (8.2.2)
Applying the Pressing Down Lemma gives us a stationary set 0 , an integer
k, ordinals i (i < k), and a sequence t Q such that for all 0 :

k() = k and 0 (, ()) = t, (8.2.3)

max(C()i ) = i for all i < k. (8.2.4)


Intersecting 0 with a closed and unbounded subset of , we may assume for every
0 that () is smaller than the next member of 0 above . Note that any
C-sequence on a regular uncountable cardinal that avoids a stationary subset of
the cardinal, in particular, must be nontrivial. Thus C ( 0 ) is a stationary
8.2. The trace lter and the square-bracket operation 245

subsequence of a nontrivial C-sequence, and so by Lemma 8.1.5, unbounded as a


family of subsets of . By the basic oscillation Lemma 8.1.2, we can nd < in
0 such that osc(C , C ) = 2. Then by the choice of , if = () and = (),
then = t , = t and < . Comparing this with (8.2.3) and (8.2.4) we
conclude that [] = t = . This nishes the proof. 

It is clear that the argument gives the following more general result.
Lemma 8.2.5. Suppose C ( < ) avoids and let A be a family of size
consisting of pairwise-disjoint nite sets, all of some xed size n. Then the set of
all elements of that are not of the form

= [a(1)b(1)] = [a(2)b(2)] = = [a(n)b(n)]


2
for some a < b in A is not stationary in . 

The following analogue of Lemma 5.1.7 gives us in particular an explanation


about the behavior of the square-bracket operation on rectangles over pairs of
unbounded subsets of .
Theorem 8.2.6. For every family A of pairwise-disjoint subsets of , all of same
xed size n, there exist an equivalence relation on n = {0, . . . , n1}, a subfamily
B A of size , and a mapping hb : n n for each b B such that,

[a(i)b(i)] = hb (i, j) for all a < b in B and i, j < n such that i  j.

Moreover, for every C B of size , for all but a nonstationary set of there
exist a < b in C such that

[a(i)b(j)] = for all i, j < n such that i j.

Proof. For each choose b A such that b > . Applying the Pressing
Down Lemma, we nd a stationary set 0 and a sequence ti (i < n) of
elements of < such that

0 (, b (i)) = ti for all i < n and 0 . (8.2.5)

For i < n, let li = |ti |. Since our C-sequence avoids , for each i < n, if

b (i) = (b (i))0 > (b (i))1 > > (b (i))li =

is the trace of the walk from b (i) down to , then

ik () = max(C(b (i))k ) < for all k < li . (8.2.6)


2 Recall that if a and b are two sets of ordinals, then the notation a < b means that max(a) <
min(b).
246 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

So, going to a further stationary subset of 0 , we may assume that for some xed
sequence ik (i < n, k < li ) of ordinals < , we have

ik () = ik for all i < n, k < li , and 0 . (8.2.7)

For 0 , set

T = {t < : ti < t  tj for some i, j < n such that (b (i))t is dened}.

Going to a stationary subset of 0 , we may assume that for some T and all 0 ,
we have T = T . For 0 and i < n, let

T (i) = {t T : t = ti and (b (i))t is dened}.

Going again to a stationary subset of 0 , we nd a sequence T (i) (i < n), such


that T (i) = T (i) for all 0 . Applying the Pressing Down Lemma, and going
to a stationary subset of 0 , we may assume to have a xed sequence t (t
T (i), i < n) such that

(b (i))t = t for all t T (i) and i < n. (8.2.8)

Moreover, we may assume that < in 0 implies b < and that for some
ordinal < ,

ik , t < C for all t T (i), i < n, k < li , and 0 . (8.2.9)

For i, j < n, set i j if and only if ti = tj and ik = jk for all k < li = lj .


Clearly, is an equivalence relation. Let us check that B = {b : 0 } satises
the second conclusion of the lemma. Consider < in 0 and i, j < n such that
i j. Then, applying 8.2.6 and 8.2.7, we conclude that for t = ti = tj ,
(1) 0 (b (i), b (j)) = t 0 (b (i), ),
(2) (b (i))t = , (b (j))t = ,
(3) (b (i))s = b (i)|s| and (b (j))s = b (j)|s| for s < t.
Note that by (8.2.6) and (8.2.7), we can also conclude that if for some k < li = lj ,
we let = b (i)k and = b (j)k , then osc(C , C ) = 1. Combining all this, we
conclude that

[b (i)b (j)] = [] for all i j and < in 0 , (8.2.10)

so the second conclusion of the lemma follows from Lemma 8.2.4.


Consider now i, j < n such that i  j. If ti = tj , then there must be
k < li = lj such that ik = jk . For 0 , put hb (i, j) = b (j)k for k < li = lj
minimal with this property. Then in this case, [b (i)b (j)] = hb (i, j) for all <
in 0 .
8.2. The trace lter and the square-bracket operation 247

Assume now that ti = tj and that there is k < min{li , lj } such that ik = jk .
For 0 , put hb (i, j) = b (j)k for k < min{li , lj } minimal with this property.
Going back to the denition of the square-bracket operation, we conclude that, in
this case as well, [b (i)b (j)] = hb (i, j) for all < in 0 .
Assume now that ti is a strict end-extension of tj but there is k < lj such
that ik = jk . For 0 , put hb (i, j) = b (j)k for k < li = lj minimal with this
property. Then again, [b (i)b (j)] = hb (i, j) for all < in 0 . Similar denition
and conclusion holds if tj strictly end-extends ti but there is k < li such that
ik = jk .
Assume now that ti is a strict end-extension of tj and ik = jk for all i < lj .
For 0 , put hb (i, j) = . Let us check that [b (i)b (j)] = hb (i, j) for all <
in 0 . First of all note that = (b (j))tj . Let = tp(C (b (i))tj ). Then either,

th element of C(b (i))tj ilj , or else th element of C(b (i))tj .

So, by (8.2.9), in both cases

th element of C(b (i))tj = th element of C .

On the other hand, for s < tj ,

th element of C(b (i))s = i|s| = j|s| = th element of C(b (j))s ,

for = tp(C(b (j))s (b (i))s ). It follows that [b (i)b (j)] = , as required.


Consider now the nal case that tj is a strict end-extension of ti and ik = jk
for all i < li . Fix a in 0 . If there is no < in 0 , set hb (i, j) = 0. Otherwise,
x  < in 0 . Note that by (8.2.8) and (8.2.9), it must be that

[b  (i)b (j)] = (b (j))t for some ti  t  tj . (8.2.11)

Moreover, the choice of ti  t  tj is independent of  . Namely, the fact that


there is a nontrivial oscillation between C(b (j))tj and C(b (i))tj , if dened, is clear.
Other than that, this equation depends on the relationships between the th point
of C(b  (i))s and the th point of C(b (j))s , for s < tj and

= tp(C(b (j))s (b  (i))s ) = tp(C(b (j))s ) = tj (|s|),

which according to (8.2.7), (8.2.8), and (8.2.9) are all independent on  . Thus, if we
put hb (i, j) = (b (j))t for t satisfying (8.2.11), we have that [b (i)b (j)] = hb (i, j)
for all < in 0 . 
For suciently large cardinals , and for some special choice of the set ,
we have the following variation on a theme rst encountered above in Theorem
6.3.2 that could be used in dening a square-bracket operation with an even more
complex behavior on families of pairwise-disjoint nite subsets of .
248 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Lemma 8.2.7. Suppose > c and that cf() > for all . Let A be a family
of pairwise-disjoint nite subsets of , all of some xed size n. Then for every
stationary 0 there exist s, t n and a positive integer k such that for every
l < there exist a < b in A and 0 > 1 > > l in 0 (max(a), min(b)) such
that:
(1) 2 (i+1 , i ) = k for all i < l,
(2) 0 (a(i), b(j)) = 0 (0 , b(j)) 0 (1 , 0 )  0 (l , l1 ) 0 (a(i), l ) for all
i, j < n,
(3) 2 (0 , b(j)) = tj and 2 (a(i), l ) = si for all i, j < n.

Proof. For and < set



S (A) = {2 (a(i), ) : i < n : a A  [, )}, (8.2.12)
<


T () = {2 (, ) : [, )}. (8.2.13)
<

Note that S (A) and T () range over subsets of n and , respectively, so if has
uncountable conality and if A and are unbounded in , then both of these sets,
S (A) and T (), are nonempty. Let 0 be a given stationary subset of . Since
has size bigger than the continuum, starting from 0one can nd an innite
sequence 0 1 of subsets of such that = i< i is stationary in
and
S (A) = S (A) = for all , 1 , (8.2.14)
T (i ) = T (i ) = for all , i+1 and i < . (8.2.15)
Pick such that is unbounded in and choose b A above . Let
tj = 2 (, bj ) for j < n. Now let s be an arbitrary member of S (A) and k an
arbitrary member of T ( ). To check that these objects satisfy the conclusion of
the theorem, let l < be given. Let 0 = and let 0 < 0 be an upper bound for
all sets of the form

C 0 ( Tr(0 , b(j)), = 0 , j < n).

Then the walk from any b(j) to any [0 , 0 ) must pass through 0 . Since
k T ( ) T (l ) we can nd 1 l (0 , 0 ) such that 2 (1 , 0 ) = k. Choose
1 (0 , 1 ) such that the walk from 0 to any [1 , 1 ) must pass through
1 . Since k T ( ) T (l1 ) = T1 (l1 ) there exists 2 l1 (1 , 1 )
such that 2 (2 , 1 ) = k. Choose 2 (1 , 2 ) such that the walk from 1 to any
[2 , 2 ) must pass through 2 , and so on. Proceeding in this way, we nd
= 0 > 1 > > l > l > l1 > > 0 such that:

i li+1 for all 0 < i l, (8.2.16)


8.2. The trace lter and the square-bracket operation 249

2 (i+1 , i ) = k for all i < l, (8.2.17)

for every i < l the walk from i to some [i+1 , i+1 )


(8.2.18)
passes through i+1 .
Since s S (A) = Sl (A), we can nd a A  [l , l ) such that 2 (a(i), l ) = s(i)
for all i < n. Now, given i, j < n, the walk from b(j) to a(i) rst passes through
all the i (i l), so 0 (a(i), b(j)) is simply equal to the concatenation

0 (0 , b(j)) 0 (1 , 0 )  0 (l , l1 ) 0 (a(i), l ).

This together with (8.2.17) gives us the three conclusions of Lemma 8.2.7. 

It turns out that Lemma 8.2.7 gives us a way to dene another square-bracket
operation which has complex behavior not only on squares of unbounded subsets
of but also on rectangles formed by two unbounded subsets of . To dene this
variation, we need a mapping n  n from into such that:

for every k, m, n, p < and s n there exist l < such


(8.2.19)
that (m + l k + s(i)) = m + p for all i < n.

Denition 8.2.8. Dene a square-bracket operation [] : []2 by letting,

[] = t where t = 0 (, )  (2 (, )) .

Thus, if = 0 > 1 > > 2 (,) = is the decreasing enumeration of


the upper trace of the walk from to , then

[] = k , where k = (2 (, )) .

It is clear that Lemma 8.2.7 and the choice of n  n in (8.2.19) give us the
following conclusion.
Lemma 8.2.9. Suppose > c and that cf() > for all . Let A be a family
of pairwise-disjoint nite subsets of , all of some xed size n, and let be
an unbounded subset of . Then almost every3 has the form [a(0)] =
[a(1)] = = [a(n 1)] for some a A, , a < . 

Let us now turn back to the case of an arbitrary C-sequence on a regular


uncountable cardinal avoiding an arbitrary stationary set . Since []
belongs to the trace Tr(, ) of the walk from to it is not surprising that []
strongly depends on the behavior of Tr. For example the following should be clear
from the proof of Lemma 8.2.4.
3 Here almost every is to be interpreted by all except a nonstationary set.
250 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Lemma 8.2.10. The following are equivalent for a given C-sequence and a set :

(a) \ {[] : {, } []2 } is nonstationary in ,



(b) \ {Tr(, ) : {, } []2 } is nonstationary in . 

This fact suggests the following denition.

 C-sequence C ( < ) is the normal


Denition 8.2.11. The trace lter of a given
lter on generated by sets of the form {Tr(, ) : {, } []2 } where is an
unbounded subset of .
Remark 8.2.12. Having a proper trace lter (i.e., a trace lter that does not contain
all subsets of ) is a strengthening of the nontriviality requirement on a given C-
sequence C ( < ). For example, if a C-sequence avoids a stationary set ,
then its trace lter is nontrivial and in fact no stationary subset of is a member
of it, and in fact, any member of the trace lter contains a closed and unbounded
subset relative to the set . Note the following analogue of Lemma 8.2.10: The
trace lter of a given C-sequence is the normal lter generated by sets of the form
{[] : {, } []2 } where is an unbounded subset of . So to obtain the
analogues of the results of Section 5.1 about the square-bracket operation on 1 ,
one needs a C-sequence C ( < ) on whose trace lter is not only nontrivial
but also not -saturated, i.e., it allows a family of pairwise-disjoint positive sets.
It turns out that the hypothesis of Lemma 8.2.4 is sucient for both of these
conclusions.
Lemma 8.2.13. If a C-sequence on avoids a stationary subset of , then there
exist pairwise-disjoint subsets of that are positive with respect to its trace lter.4

Proof. This follows from the well-known fact (see [55]) that if there is a normal,
nontrivial and -saturated lter on , then for every stationary there exists
< such that is stationary in (and the fact that the stationary set which
is avoided by the C-sequence does not reect in this way). 

Corollary 8.2.14. If a regular cardinal admits a nonreecting stationary subset,


then there is c : []2 which takes all the values from on any set of the form
[]2 for some unbounded set . 

To get such a c, one composes the square-bracket operation of some


C-sequence, that avoids a stationary subset of , with a mapping :
with the property that the -preimage of each point from is positive with re-
spect to the trace lter of the square sequence. In other words c is equal to the
composition of [] and , i.e., c(, ) = [] . Note that, as in Section 5.1, the
property of the square-bracket operation from Lemma 8.2.5 leads to the following
rigidity result which corresponds to Lemma 5.1.9.
4A subset A of the domain of some lter F is positive with respect to F if A F = for every
F F.
8.3. Projections of the square-bracket operation on accessible cardinals 251

Lemma 8.2.15. The algebraic structure (, [], ) has no nontrivial automorphisms.




Remark 8.2.16. Note that every which is a successor of a regular cardinal


admits a nonreecting stationary set. For example = { < : cf = } is such
a set. Thus any C-sequence on that avoids leads to a square-bracket operation
which allows analogues of all the results from Section 5.1 about the square-bracket
operation on 1 . We shall now explore this by examining properties of a particular
projection of the square-bracket operation that are analogous to those of Lemma
5.1.7 and Theorem 5.2.5.

8.3 Projections of the square-bracket operation


on accessible cardinals
Suppose now that is not a strong limit cardinal and that it carries a C-sequence
C ( < ) avoiding a stationary set . As indicated above, any successor
of a regular cardinal has these properties. Let [] be the corresponding square-
bracket operation. Let be the minimal cardinal such that 2 . Choose a
sequence r ( < ) of distinct elements of 2 . Let G be the collection of all nitely
supported5 maps g : 2 2 where = (g) < . Clearly, G has cardinality
, so we can nd a mapping : G such that 1 (g) is stationary for all
g G. Finally, dene a binary operation [[]] on as follows:

[[]] = ([])(r  (([])), r  (([]))). (8.3.1)

The following is a simple consequence of the property of the square-bracket op-


eration stated in Lemma 8.2.5 above and its proof is very similar to the proof of
Lemma 5.2.2 above (see also the proof of Theorem 8.3.2 below).

Lemma 8.3.1. Suppose that is an uncountable regular and not strong limit cardi-
nal carrying a C-sequence that avoids a stationary subset of . Then the projec-
tion [[]] of the square-bracket operation [] has the property that for every family
A of size consisting of pairwise-disjoint nite subsets of , all of some xed size
n, and every sequence 0 , . . . , n1 of ordinals < , there exist a = b in A such
that [[a(i)b(i)]] = i for all i < n. 

As noted before, a typical input to this result is a successor of a regular


cardinal = + and = { < + : cf() = .} The additional information
given in Theorem 8.2.6 allows us to obtain in this case a particularly interesting
projection of the square-bracket operation.

5 i.e., maps g with supp(g) = {(s, t) : g(s, t) = 0} nite.


252 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Theorem 8.3.2. Suppose that is a successor of a regular cardinal . Then the


square-bracket operation [] admits a projection

[[]] : []2

with the property that for every family A of size of pairwise-disjoint nite subsets
of , all of some xed size n, there exist B A of size , a mapping h : n n ,
and an equivalence relation on n = {0, 1, . . . , n 1} such that for all a < b in
B, we have that
[[a(i)b(j)]] = h(i, j) whenever i  j.
Moreover, for every subfamily C of B of size and for every mapping g : nn
there exist a < b in C such that

[[a(i)b(j)]] = g(i, j) for all i < n such that i j.

Proof. Let be minimal such that 2 . As before, choose a sequence


r ( < ) of distinct elements of 2 . Let G be the collection of all nitely
supported maps g : 2 2 where = (g) < . Clearly, G has cardinality
. Let : G be a map with the property that 1 (g) is stationary for
all g G . Finally, dene a binary operation [[]] : []2 as follows:

[[]] = ([])(r  (([])), r  (([]))). (8.3.2)

Let A be a given family of pairwise-disjoint nite subsets of , all of some


xed size n. Applying Lemma 8.2.6, we get a subfamily B of size , an equivalence
relation on n, and a mapping hb : n n such that, on one hand, for all
a < b in B, we have [a(i)b(j)] = hb (i, j) for i  j, while on the other hand, for
almost all < of conality , we can nd a < b in B such that [a(i)b(j)] =
whenever i j. Note that for each b B, the range of the composition mapping
hb : n n G has size , so there are at most such mappings. So going to
a subset of B of size , we may assume that for some xed map f : n n G ,
we have that
hb = f for all b B.
For i, j < n, let f (i, j) : 2(i,j) . Since |2 | for all < , shrinking the
family B, we can assume to have an ordinal < and a sequence ti (i < n) of
elements of 2 such that:
(1) > (i, j) for all i, j < n,
(2) rb(i)  = ti for all i < n and b B,
(3) ti = tj whenever i = j.
Dene h : n n , by

h(i, j) = f (i, j)(ti  (i, j), tj  (i, j)).


8.3. Projections of the square-bracket operation on accessible cardinals 253

Going back through the denitions, we see that for all a < b in B,

[[a(i)b(j)]] = f (i, j)(ti  (i, j), tj  (i, j)) = h(i, j) for all i  j,

as required.
To check the second conclusion of the lemma, suppose that g : n n is
a given map. Choose a nitely supported map g : 2 2 , such that

g(ti , tj ) = g(i, j) for all i, j < n.

By our choice of the set g = { < : cf() = and () = g} is stationary, so


by the conclusion of Theorem 8.2.6, we can nd g and a < b in B such that

[a(i)b(j)] = for all i, j < n, i j.

Going back through the denitions, we conclude that,

[[a(i)b(j)]] = g(ti , tj ) = g(i, j) for all i, j < n such that i j,

as required. 
Working as in the case of the corresponding projection of the square-bracket
operation on 1 , we get the following kind of application.
Theorem 8.3.3. Suppose that is a successor of a regular cardinal . Let V and
W be vector spaces over the same eld K such that W has cardinality at most
while the dimension, and therefore cardinality, of V is equal to . Then there is a
non-degenerate alternating bilinear mapping : V V W such that {(x, y) :
x, y X, x = y} = W for all X V of cardinality .
Proof. Let B = {v : < } be a xed linear basis of V and in order to simplify
the notation, we assume that the range of [[]] is equal to W rather than . Dene
 B B by letting (v , v ) = 0 for all < , while when , < are dierent,
we set

(v , v ) = [[]] for < and (v , v ) = [[]] for > .

Then  BB extend to a unique bilinear map : V V W . It is easily checked


that the extension is alternating, i.e., that (x, x) = 0 for all x, or equivalently
that (y, x) = (x, y) for all x and y.
To check the basic property of , consider an arbitrary subset X of V of
cardinality . Going to a subset of X of size , we may assume that for some xed
integer n and some xed sequence i (i < n) of elements of K, every vector x X
has a representation
x = i<n i vax (i) ,
where ax is some subset of of cardinality n. Going to a further subset of X of
size , we may assume that the ax s form a -system with root a. Let us consider
254 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

rst the case a = , i.e., the case when the sets ax are pairwise-disjoint. Applying
Theorem 8.3.2, we nd a Y X of size , a mapping h : n n W , and an
equivalence relation on n = {0, 1, . . . , n 1} such that [[ax (i)ay (j)]] = h(i, j)
for all x, y Y such that ax < ay and all i, j with i  j. Changing h, we may
assume that h(i, j) = 0 whenever i j. Let

w = (i,j)=(0,0) i j h(i, j).

Let u = t w and let v = (1/20 )u. Dene g : n n W by letting g(0, 0) = v


and g(i, j) = 0 for (i, j) = (0, 0). By the second conclusion of Theorem 8.3.2, we
can nd x, y Y such that ax < ay and [[ax (i)ay (j)]] = g(i, j) for all i < n such
that i j. Then

(x, y) = ij i j [[ax (i)ay (j)]] + ij i j [[ax (i)ay (j)]] = w + 20 (1/20 )u,

which gives us the desired conclusion (x, y) = t.


Assume now that the root a is not empty. Fix x0 in X and nd a subset Y
of X \ {x0 } of size such that for some w0 in W and all y Y , we have that
(x0 , y) = w0 . Since the supports of vectors from the family {y x0 : y Y }
are pairwise-disjoint, by the rst part of the proof, we can nd y, z Y such that
ax \ a < az \ a and
(y x0 , z x0 ) = t + (x0 , x0 ).
It follows that (y, z) = t, as required. 

Recall that vector space V over a eld K equipped with a bilinear form
: V V K is called symplectic space if the bilinear form is alternating, i.e.,
it has the properties (y, x) = (x, y) and (x, x) = 0 for all x, y V . The
symplectic space (V, ) is non-degenerate if for all x V there is y V such
that (x, y) = 0. A subspace W of a symplectic space (V, ) is called isotropic if
(x, y) = 0 for all x, y W .
Corollary 8.3.4. For every regular cardinal , every eld K of size at most ,
and every vector space V over K of dimension + , there is a non-degenerate
symplectic space over K of the form (V, ) which contains no isotropic subspace
of dimension + . 

Adjusting a construction from [8], Ehrenfeucht and Faber [27] have found
the following general application of bilinear mappings.
Lemma 8.3.5. Suppose : V V W is a bilinear form, where V and W are
vector spaces over the same eld K. Then the product V W can be given a
structure of a 2-nilpotent group, denoted by V W , in which the commutators6
have the form [(x, a), (y, b)] = (0, (x, y) (y, x)).
6 i.e., products of the form [x, y] = xyx1 y 1 .
8.3. Projections of the square-bracket operation on accessible cardinals 255

Proof. The multiplication operation of V W is dened by

(x, a)(y, b) = (x + y, a + b + (x, y)). (8.3.3)

This is easily seen to be a group operation whose identity is (0, 0) and where
(x, a)1 = (x, a + (x, x)). Checking the equation

(x, a)(y, b)(x, a)1 (y, b)1 = (0, (x, y) (y, x)) (8.3.4)

is also straightforward. 
Applying this construction to the bilinear mappings of Theorem 8.3.3, we get
the following consequence.
Corollary 8.3.6. For every regular cardinal there is a 2-nilpotent group G of
cardinality + such that every abelian subgroup of G has cardinality at most . 
In order to investigate the homomorphisms between groups of the form V W
we need to x some notation. Fixing a linear basis B = {v : < }, a vector x
from V has its support dened to be equal to the nite set

supp(x) = { < : = 0},

where x = < v is the unique representation of x in terms of B. For an


arbitrary element (x, a) of the group V W , we let

supp(x, a) = supp(x).

The following result shows that if the bilinear mapping is based on the projection
of the square-bracket operation as in Theorem 8.3.3, then the corresponding group
V W holds a considerable amount of rigidity.
Theorem 8.3.7. Let be the successor of some regular cardinal , let V and W be
vector spaces over the same eld K such that W has cardinality at most while
the dimension of V is equal to , and let : V V W be the bilinear mapping
of Theorem 8.3.3. Suppose is an isomorphism between two subgroups of V W .
Then there is < such that for all (x, a) V W ,

supp((x, a)) \ = supp(x, a) \ .

Proof. Suppose that the conclusion is false. Then for every < we can select
(x , a ) in V W such that

supp((x)) \ = supp(x) \ .

By the Pressing Down Lemma, we can nd a stationary set , vectors x , y,


and b from W such that for all ,
y , z , x , y ( ) from V , and vectors a
(1) a = a ) = (y , b),
and (x , a
256 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

(2) x  = x
and y  = y,
(3) x  [, ) = z + x and y  [, ) = z + y ,
(4) x , y , and z have pairwise-disjoint supports.

By symmetry, we may assume that supp((x)) \  supp(x) \ , or equivalently


that supp(y ) = for all . Our intention is to try to nd < in such
that (x , x ) = 0 but (y , y ) = 0 showing that (x , a ) and (x , a) commute
in V W though their images (x , a ) = (y , b) and (x , a ) = (y , b) do not
commute, giving us the desired contradiction. Since the mapping is alternating,
shrinking , we may assume that ( x, x ) ( ) and (
y , y ) ( ) are constant
sequences. It follows that for all < in ,

(x , x ) = (z + x , z + x ) and (y , y ) = (z + y , z + y ). (8.3.5)

For let s = supp(x ) supp(x ) supp(z ). Shrinking further, we may


assume that for some integer n and all , we have that |s | = n. Moreover,
we may assume that for some xed sequence i (i < n) of scalars and some xed
decomposition {0, 1, . . . , n 1} = H I J, we have that for all

x = iH i vs (i) , y = iI i vs (i) , and z = iJ i vs (i) . (8.3.6)

By Theorem 8.3.2, we can nd an unbounded subset of , an equivalence relation


on n, and h : n n W such that for all < in ,

[[s (i)s (j)]] = h(i, j) whenever i  j. (8.3.7)

Moreover, for every unbounded subset of and for every mapping g : nn W


there exist < in such that

[[s (i)s (j)]] = g(i, j) for all i < n such that i j. (8.3.8)

Let i0 = min(I) and let j0 = min(H J). We may assume that H J is indeed
nonempty, since in this case our job reduces to a straightforward application of
Theorem 8.3.3. Note that i0 = j0 . Let hx : (H J) (H J) W be a mapping
that extends the restriction of h to this domain Dx = (H J) (H J) and which
takes the value 0 otherwise. Let

ux = (i,j)Dx \{(j0 ,j0 )} i j hx (i, j)

and let hy : (I J) (I J) W be a mapping that extends the restriction


of h to this domain Dy = (I J) (I J) and takes the value 0 on all new
pairs except perhaps on the pair (j0 , j0 ) in case j0 J in which case we put
hy (j0 , j0 ) = (1/2j0 ux ). Let

uy = (i,j)Dy \{(i0 ,i0 )} i j hy (i, j).


8.4. Two more variations on the square-bracket operation 257

Choose a mapping g : n n W that extend hx and hy on all pairs from


their respective domains except (j0 , j0 ) and (i0 , i0 ) for which we put g(j0 , j0 ) =
(1/2j0 ux ) and g(i0 , i0 ) = uy if uy = 0 and put g(i0 , i0 ) to be an arbitrary nonzero
element of W , otherwise. Find < satisfying the equation (8.3.8) for this choice
of g. Then on one hand

(x , x ) = (z + x , z + x ) = ux + 2j0 (1/2j0 ux ) = 0,

and on the other hand

(y , y ) = (z + y , z + y ) = uy + 2i0 g(i0 , i0 ) = 0,

as required. 
Corollary 8.3.8. Suppose that is the successor of some regular cardinal , and
that V and W are vector spaces over the same eld K such that W has cardinality
at most while the dimension of V is equal to . Let : V V W be the bilinear
mapping of Theorem 8.3.3. Then the corresponding group V W has a family of
2 pairwise non-isomorphic subgroups.
Proof. For a subset of , let

G = {(x, a) V W : supp(x) }.

Clearly each G forms a subgroup of V W . By Theorem 8.3.7, G is not isomor-


phic to a subgroup of G whenever the dierence \ is unbounded in . Since
there is a family F of size 2 of subsets of such that \ whenever = F,
the conclusion of the corollary follows. 
Remark 8.3.9. The special case of Theorem 8.3.3, where W = K = Zp for some
prime p and where V is the vector space of nitely supported maps from into
Zp , produces a bilinear form with a particularly interesting group V Zp . Namely,
in this case the group G = V Zp is an example of what is called in [124] an
extraspecial p-group G whose derived group G is equal to its center Z(G) and is
in turn isomorphic to Zp , while the corresponding quotient G/G is isomorphic to
a direct product of copies of Zp . The rst use of the square-bracket operation to
construct an extraspecial p-group of cardinality without an abelian subgroup of
cardinality was done in [98].

8.4 Two more variations on the


square-bracket operation
In this section we present two variations of the square-bracket operation which
allow projections with quite dierent properties from those encountered in the
previous section. In fact these properties can be realized only at cardinals > 1
258 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

and as we will see this is reected in the fact that these square-bracket operations
can be dened only on the basis of C-sequences that avoid stationary subsets of
consisting of ordinals of high conalities. For example, while our rst variation
works on the basis that cf() > for all , we must assume that is bigger
than the continuum c. The second variation does not need the assumption > c
though it works only for C-sequences avoiding a stationary set with the property
cf() > 1 for all .
We start this section with a variation of the square-bracket operation that
works in the case > c and under the assumption that the C-sequence on avoids
some xed stationary set with the property that cf() > for all . In
order to dene it, we need to x a function   from to such that

n = { :  = n}

is stationary in for all n. The mapping  leads us naturally to the variation

 : []2 <

of the mapping 0 which we dene recursively by the formula

 (, ) =     (, min(C \ )), (8.4.1)

with the boundary value  (, ) =   for all . The variation  of 0 shares


many properties with 0 such as the following that will be frequently and implicitly
used below and where we are using the already standard notation

(, ) = max{max(C ) : Tr(, ) and = }

for the upper bound on the lower trace.


Lemma 8.4.1. If (, ) < < < , then  (, ) =  (, )  (, ). 
It should also be clear that the proof of Lemma 8.2.7 allows adding the
following additional conclusions.
Lemma 8.4.2. Under the hypothesis of Lemma 8.2.7, its conclusion can be extended
by adding the following two new statements:
(4)  (1 , 0 ) = =  (l , l1 ).
(5) The maximal term of the sequence  (1 , 0 ) = =  (l , l1 ) is bigger
than the maximal term of any of the sequences  (0 , b(j)) or  (a(i), l ) for
i, j < n. 
Having this in mind, the following variation on the square-bracket operation
suggests itself.
Denition 8.4.3. For < < , let [] = t for t the minimal initial part of
0 (, ) such that (t ) = max( (, )).
8.4. Two more variations on the square-bracket operation 259

In other words, if = 0 > 1 > > 2 (,) = is the decreasing


enumeration of the upper trace of the walk from to , then

[] = k , where k = min{j : (j ) = max{(i ) : i 2 (, )}}.

So [] is the rst place in the walk from to where the function 
reaches its maximum among all other places visited during the walk. Note that
combining the conclusions Lemmas 8.2.7 and 8.4.2, we arrive at the following
additional information.
Lemma 8.4.4. Under the hypothesis of Lemma 8.2.7, its conclusion can be extended
by adding the following:
(6) [a(i)b(j)] = [1 0 ] for all i, j < n. 
Having in mind the property of [] stated in Lemma 8.2.9, the following
variation is now quite natural.
Denition 8.4.5. [] = [[] ] for < < .
Using Lemmas 8.2.7, 8.4.2, and 8.4.4 one gets the following conclusion.
Lemma 8.4.6. Let A be a family of pairwise-disjoint nite subsets of , all of
some xed size n. Then for all but a nonstationary set of one can nd a < b
in A such that [a(i)b(j)] = for all i, j < n. 
Composing [] with a mapping : with the property that 1 ()
is stationary for all < , one gets a projection of [] for which the conclusion of
Lemma 8.4.6 is true for all < . Assuming that is moreover a non-inaccessible
regular cardinal bigger than the continuum (such as, for example, a successor of a
regular cardinal c, in which case can be taken to be the set { < + : cf =
}), and proceeding as in Lemma 8.3.1 above we get a projection [[]] with the
following property.
Theorem 8.4.7. Suppose that a regular non-strong limit cardinal > c carries a
C-sequence avoiding a stationary set of ordinals < of uncountable conality.
Then the square-bracket operation [] has a projection [[]] with the property that
for every family A of pairwise-disjoint nite subsets of , all of some xed size n,
and for every q : n n , there exist a < b in A such that [[a(i)b(j)]] = q(i, j)
for all i, j < n. 
Remark 8.4.8. The rst example of a cardinal with such a complex binary opera-
tion was given by the author [110] using the oscillation mapping described above
in Section 8.1. It was the cardinal b, the minimal cardinality of an unbounded
subset of under the ordering of eventual dominance. The oscillation mapping
restricted to some well-ordered unbounded subset W of is perhaps still the most
interesting example of this kind, due to the fact that its properties are preserved in
forcing extensions that do not change the unboundedness of W (although they can
260 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

collapse cardinals and therefore destroy the properties of the square-bracket oper-
ations on them). This absoluteness of osc is the key feature behind its applications
in various coding procedures (see, e.g., [118]).
We are now ready for another variation of the square-bracket operation. It
works for a regular cardinal that carries a C-sequence C ( < ) avoiding a
stationary subset of with the property that cf() > 1 for all . The
additional parameters in the denition are a mapping  from to 1 such
that
= { : = }
is stationary in for all < 1 , and a mapping  from 1 into with the
property that
n = { < 1 : = n}
is stationary in 1 for all n < . For < < , let
= 0 > 1 > > 2 (,) =
be the trace of the walk from to , and set,
u(, ) = {k > 2 (, )/2 : k > l for all l < 2 (, )/2}, (8.4.2)
v(, ) = {l < 2 (, )/2 : l > k for all k > 2 (, )/2, k
/ u(, )}. (8.4.3)
Let
k(, ) = min u(, ) and l(, ) = (min{l : l v(, )}) . (8.4.4)
If k(, ) l(, ) 2 (, , ) set
[] = k(,)l(,) . (8.4.5)
Otherwise, [] = .
We extend the mapping , to a mapping from < into 1<
in the natural way, by letting be a nite sequence of the same length as and
(i) = (i) for all i < | |. The mapping also leads us to a natural
variation : []2 1< of the mapping 0 dened by the recursive formula
(, ) =   (, min(C \ )),
with the boundary value (, ) =  . Thus, if
= 0 > > 2 (,) =
is the decreasing enumeration of the trace of the minimal walk from to along
our C-sequence, then
(, ) = , where = 0 , 1 , . . . , 2 (,) .
As before, the variation shares many of the common properties with the original
function 0 including the following.
8.4. Two more variations on the square-bracket operation 261

Lemma 8.4.9. If (, ) < < < , then (, ) = (, ) (, ). 


We shall need another piece of notation, where for a given (1 )1 and
< 1 , we let be the sequence in (1 + 1)< which has the length | | equal
to the length || of , and

(i) = (i) if (i) < ; (i) = 1 if (i) .

The analysis of the square-bracket operation [] is greatly facilitated by the


following notion.
Denition 8.4.10. Given an unbounded subset of and < , a sequence
(1 + 1)< is called a (, )-lower limit point if for every < 1 and <
there is (, ) such that (, ) = .
Lemma 8.4.11. For every unbounded and almost all ordinals < of
conality > there is a (, )-lower limit point in (1 + 1)< .
Proof.
 Choose an arbitrary < of uncountable conality belonging to the set
<1 lim( lim()). Consider rst the case cf() = 1 . Fix a strictly increasing
sequence ( < 1 ) converging to . For < 1 , choose lim() ( , ).
Then, since our C-sequence avoids the set , for each ,

( , ) = max{max(C ) : Tr( , ) and = } < ,

so we can nd ( , ) such that > ( , ). It follows that Tr( , )


for all < 1 , and therefore = is a term of the sequence ( , ) for all < 1 .
Find a stationary set 0 1 and n < such that | ( , )| = n, and such that
for all i < n, either ( , )() < for all 0 , or else ( , )() for all
0 . So, by a repeated application of the Pressing Down Lemma, we can nd
(1 + 1)< and a stationary set 1 0 such that

( , ) = for all 1 .

Then is one of the (, )-lower limit points.


Let us consider now the case = cf() > 1 . Choose a strictly increasing
sequence ( < ) converging to . Working as in the rst case, for each < ,
we nd a stationary set 1 , a sequence

( , ) ( ),

and (1 +1)< such that ( , ) = for all . Choose (1 +1)<


such that { < : = } has cardinality . Then is a (, )-lower limit
point. 
Denition 8.4.12. Given an unbounded subset of and < , a sequence
(1 + 1)< is called a (, )-upper limit point if for every < 1 there is
\ ( + 1) such that (, ) = . If in addition we are given a subset of 1 ,
262 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

we say that such a sequence is a (, , )-upper limit point if for every < 1 ,
we can nd \ ( + 1) such that
(, ) = and min{ (, )(i) : (i) = 1 } .
Lemma 8.4.13. For every unbounded and stationary 1 , for all but a
bounded set of < , there exists a (, , )-upper limit point.
Proof. For < 1 , let
= { < : is not a term of (, ) for all \ ( + 1)}.
Since for each < 1 , the set is stationary in and is avoided by our C-
sequence, one easily concludes that | | < for all < 1 . So it suces to show
that if < dominates all the s, then there is always a (, , )-upper limit
point. So x such and for each < 1 a \ ( + 1) such that appears
as a term of (, ). Then as before, we can choose a stationary 0 and
n < such that | (, )| = n for all 0 , and such that for all i < n, either
(, )(i) < for all 0 , or else (, )(i) for all 0 . So by a
repeated application of the Pressing Down Lemma, we can nd a stationary set
1 0 and (1 )< such that
(, ) = for all 1 .
Since for 1 ,
min{ (, )(i) : i < n, (i) = 1 } = 1 ,
we conclude that is a (, , )-upper limit point. 
Remark 8.4.14. Note that if cf() > 1 in the above proof, then the sequence
(, ) ( < ) is bounded in , so if < is one of its upper bounds, then we
say that the resulting sequence is a (, , , )-upper limit point.
We are now ready to state and prove the rst important property of the
square-bracket operation [] .
Lemma 8.4.15. For every pair and of unbounded subsets , almost all
are of the form [] for some and such that < .
Proof. Let  be a given stationary subset of . We need to nd and
with < such that []  . Consider  lim(). Since cf() > 1 , we
can nd (1 )< such that
sup{ : (, ) = } = . (8.4.6)
For the same  , we x also \ ( + 1) and set + = (, ). Since
(, ) < , we can nd a stationary set   , sequences , + (1 )< ,
and  < such that
(, ) =  , = , and + = + for all  . (8.4.7)
8.4. Two more variations on the square-bracket operation 263

By Lemma 8.4.11, there is a stationary set  and (1 + 1)< such that

is a ( , )-lower limit point for all . (8.4.8)

Let
i0 = min{i < || : (i) = 1 .} (8.4.9)
By Lemma 8.4.11, there is a stationary set  and (1 + 1)< such that

is a (, )-lower limit point for all  . (8.4.10)

By Lemma 8.4.13 and Remark 8.4.14, for every  there is (1 + 1)<


and  < < such that is a ( , , i0 +| | , )-upper limit point. So applying
the Pressing Down Lemma, we can nd ordinal  < , a sequence (1 + 1)< ,
and stationary   such that

is a ( , , i0 +| | ,  )-upper limit point for all  . (8.4.11)

By Theorem 6.3.2, we can nd  <  in  such that

2 (  ,  ) > | | + | + | + | | + ||. (8.4.12)

Choose 0 < 1 such that

0 > (rang( (  ,  )) rang() rang( + )) 1 . (8.4.13)

Since is a (,  )-limit point there is  ( ,  ) such that

(  ,  )0 = . (8.4.14)

Pick 1 < 1 such that


1 > rang( (  ,  )). (8.4.15)
   
By (8.4.11) and the fact that belongs to , we can nd \ ( + 1) such
that
(  , )1 = and (  , ) <  . (8.4.16)
Since  , by (8.4.7), we conclude that

(, ) = + and (, ) =  . (8.4.17)

Choose 2 < 1 such that


2 > rang( (  , )). (8.4.18)
    
Since , by (8.4.8), there is ( , ) such that

(  ,  )2 = . (8.4.19)

Since   , by (8.4.7), we can nd such that

> max{ , (  ,  )} and (,  ) = . (8.4.20)


264 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Let = . Then , , and

0 (, ) = 0 (, ) 0 (  , ) 0 (  ,  ) 0 (  ,  ) 0 (  ,  ) 0 (,  ).

Note that the middle point(s) of the trace = 0 > 1 > > 2 (,) = is
(are) positioned way inside the trace of the walk from  to  . The set u(, )
consists of the positions k belonging to the segment 0 (  ,  ) of the walk such
that (  ,  )(k) 2 . It follows that

p(, ) = m + |0 (  ,  )| + i0 = m + | | + i0 ,

where m is determined by the equation m =  . The set v(, ) consists of the


positions k belonging to the segment 0 (  , ) of the walk such that (  , )(k)
1 , since any such value is bigger than (, )(l) for every l > 2 (, )/2 with
l / u(, ). The portion corresponding to 0 (  ,  ) and the sequence with
values 0 was inserted in order to provide that no other k < 2 (, )/2 have
this property. Note also that by the choice of and  and the sequence , we
conclude that
min{k : k v(, )} i0 +| |.
It follows that q(, ) = i0 + | |, and therefore,

[] = p(,)q(,) = m =   ,

as required. 
In fact, we have here the following slightly stronger property of [..] .
Theorem 8.4.16. Suppose that a regular cardinal carries a C-sequence avoiding a
stationary set of conality > 1 ordinals. Then the corresponding square-bracket
operation [] has the property that for every family A of pairwise-disjoint nite
subsets of , all of some xed size n, for almost all there exist a < b in A
such that [a(i)b(j)] = for all i, j < n.
Proof. This has been really established during the course of the proof of Lemma
8.4.15, so let us only point out the adjustment needed for the formally stronger
result. As before we x a stationary set  . Consider  lim(). Since
cf() > 1 , we can nd (i) (1 )< for i < n such that for all < there is
a A that is included in the interval (, ) such that

(a(i), ) = (i) for all i < n. (8.4.21)

For the same  , we x also b A such that b > and set + (i) = (, b (i)).
Since (, b (i)) < for all i < n, we can nd a stationary set   , sequences
(i), + (i) (1 )< (i < n) and  (i) < (i < n) such that for all i < n,

(, b (i)) =  (i), (i) = (i), and + (i) = + (i) for all  .


8.4. Two more variations on the square-bracket operation 265

We now proceed as in the previous proof, choosing of course  <  such that

2 (  ,  ) > | | + || + (| (i)| + | + (i)|),
i<n

and choosing 0 such that



0 > (rang( (  ,  )) rang() rang( + (i))) 1 .
i<n

This will result in nding b = b > and a (max{ , (  ,  )},  ), both in A,


such that
[a(i)b(j)] =   for all i, j < n,
as required. 
Working as in the case of the original square-bracket operation [], one ob-
tains the following result.
Theorem 8.4.17. Suppose that a regular non-strong limit cardinal carries a C-
sequence avoiding a stationary set of conality > 1 ordinals. Then the corre-
sponding square-bracket operation [] has a projection [[]] with the property that
for every family A of pairwise-disjoint nite subsets of , all of some xed size n,
and for every q : n n , there exist a < b in A such that [[a(i)b(j)]] = q(i, j)
for all i, j < n. 
Remark 8.4.18. A typical cardinal satisfying the hypothesis of the previous result
is a successor of a regular cardinal > 1 . Thus, = 3 is the minimal instance
to which this result applies. The case = 2 requires a special treatment as we
will obtain it, not through projecting a square-bracket operation on 2 , but rather
through stepping up the square-bracket operation of 1 .
To describe a coloring that would cover the case = 2 , we choose a mapping
 from 2 onto 1 such that
= { < 2 : cf() = 1 and = }
is stationary in 2 for all < 1 . Our C-sequence C ( < 2 ) will have the usual
properties including the property that tp(C ) 1 for all . This in particular
gives that the C-sequence avoids all the sets ( < 1 ). As before, the mapping
 leads us to the variation : [2 ]2 1 of 0 dened by the recursive
formula
(, ) =   (, min(C \ )),
with the boundary value (, ) =   and which has the property that,
(, ) < < < implies (, ) = (, ) (, ). (8.4.22)
Finally, dene g : [2 ] 1 , by
g(, ) = f ( (, )), (8.4.23)
where f is the mapping of Lemma 5.2.6.
266 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Lemma 8.4.19. For every family A of 2 pairwise-disjoint nite subsets of 2 ,


all of some xed size n, and for every < 1 , there exist a < b in A such that
g(a(i), b(j)) = for all i, j < n.
Proof. Let

C = { < 2 : for all < there is a A, a (, )}.

Clearly, C is a closed and unbounded subset of 2 . For < 2 , we x b A such


that b > . Then, since our C-sequence avoids , for every ,

(, ) = max{max(C ) : Tr(, ) and = } < .

So, we can nd a single ordinal 0 < 2 , and for each < 1 a stationary set
C such that for all < 1 ,

(, ) < 0 for all and b . (8.4.24)

Choose an ordinal < of conality 1 that is a common limit point of all sets
( < 1 ). For < 1 choose \ ( + 1). Then as before, for all < 1
and b , we have (, ) < . So we can choose a strictly increasing sequence
( < 1 ) converging to such that for all < 1 ,

> (, ) for all b and < . (8.4.25)

Note that for the same reason for each < 1 , we have ( , ) < , so since this
ordinal has conality 1 and since it belongs to C, we can nd a < in A such
that
a > for all < , and (8.4.26)
a > ( , ) and a > (, ) for all b and < . (8.4.27)
Combining all this we conclude that for < < 1 ,

(, ) = ( , ) (, ) for a , and (8.4.28)

(, ) = (, ) ( , ) (, ) for a and b . (8.4.29)


For < 1 , let be concatenation of all sequences of the form (, ) for a
of the form (, ) for b in some order arbitrarily chosen. Clearly is a term
of for all < 1 . Given < 1 , by Lemma 5.2.6, we can nd < such that
f (  ) = whenever is a subsequence of and is a subsequence of and
appears as a term in both of them. Given a and b , letting = (, )
and = (, ), we see that they do have as one of their terms, and therefore,
by (8.4.28) and (8.4.29),

g(, ) = f ( (, )) = f ( (, ) (, )) = ,

as required. 
8.4. Two more variations on the square-bracket operation 267

Since 2 is not a strong limit cardinal, the already standard argument gives
us the following consequence of Lemma 8.4.19.
Lemma 8.4.20. The mapping g of Lemma 8.4.19 has a projection [g] with the
property that for every family A of 2 pairwise-disjoint nite subsets of 2 , all of
some xed size n, and every mapping h : n n 1 , there exist a < b in A such
that [g](s(i), b(j)) = h(i, j) for all i, j < n. 
We nish this section by giving some typical applications of these results.
Theorem 8.4.21. Suppose that = 2 , or is a regular cardinal carrying a C-
sequence avoiding a stationary set of conality > 1 ordinals, or > c and
carries a C-sequence avoiding a stationary set of conality > ordinals. Then
the -chain condition is not productive, i.e., there exist two partially-ordered sets
P0 and P1 satisfying the -chain condition but their cartesian product P0 P1
fails to have this property.
Proof. Since in all three cases the proofs are rather similar let us show how it
works in the second case. Fix two disjoint stationary subsets 0 and 1 of . For
i < 2, let
Pi = {p []< : [] i for all < in p}.
By Lemma 8.4.6, P0 and P1 are -cc posets. The cartesian product P0 P1 fails
the -chain condition since it has the sequence {}, {} ( < ) of pairwise
incomparable conditions. 
Corollary 8.4.22. For every regular uncountable cardinal , the + -chain condition
is not productive, i.e., there exist two partially-ordered sets P0 and P1 satisfying the
+ -chain condition but their cartesian product P0 P1 fails to have this property.

Corollary 8.4.23. The 2 -chain condition is not productive. 
Much ner examples of this sort can be obtained by relying on the projection
[[]] of the square-bracket operation [] rather than the square-bracket operation
itself. A typical application of this sort is given in the following result.
Theorem 8.4.24. For every regular uncountable cardinal there is a topological
group G of cellularity7 , though the cellularity of its square G2 is bigger than .
Proof. All that we will need in constructing the group G is a mapping f : [+ ]3 3
with the property that for every family A of + pairwise-disjoint nite subsets of
+ , all of some xed size n, and for every h : n n 3, there exist a < b in A
such that f (a(i), b(j)) = h(i, j) for all i, j < n. By going to an interval of the form
[, + ), we may assume that

{ < + : f (, ) = i} has size + for all < + and i < 3. (8.4.30)


7 Recall that cellularity of a topological space X is dened to be the supremum of cardinalities
of families of pairwise-disjoint open subsets of X.
268 Chapter 8. The Oscillation Mapping and the Square-bracket Operation

Let S = + 2. An element of S of the form (, i) will be denoted by i . Let G be


the collection of all nite subsets of S considered as a group with the symmetric
dierence +2 as the group operation.For a G, let

supp(a) = { < + : (, i) a for some i < 2}.

For < + , let G be the collection of all nite subsets a of S such that:
(1) i
/ a for all i < 2,
(2) < and f (, ) = i for some i < 2 imply that a1i
/ a,
(3) < and f (, ) = 2 imply that either i a for all i < 2, or else, i
/a
for all i < 2.

Clearly, each G is a subgroup of G and by (8.4.30), {} = <+ G . Thus, if

for subset b of + , we let Gb = b G , we get that Gb (b [+ ]< ) forms a
neighborhood base of in the topological group structure that we decide to put
on G. We shall rst show that the topological group G contains no family of more
than pairwise-disjoint open sets.
Suppose a +2 Gb ( < + ) is a given sequence of basic open sets of G. We
need to nd < such that

(a +2 Gb ) (a +2 Gb ) = .

We may assume that a s form a -system with root a and that b s form a
-system with root b. For < + , let a
= a \ a, let b = b \ b, and let

c = b supp(
a ).

We may assume that c s have some xed size n and that for some xed subset I
of n = {0, 1, . . . , n 1}, a xed mapping : I 3, and all < + , we have:
= {c (i) : i I},
(4) a
(5) (i) = 0 if c (i)0 a
and c (i)1
/a ,
(6) (i) = 1 if c (i)1 a
and c (i)0
/a ,
(7) (i) = 2 if c (i)0 a
and c (i)1 a
.
Choose a mapping h : n n 3 such that h(i, j) = (i) for i I and arbitrary
j < n. By the basic property of f , we can nd < such that

f (c (i), c (j)) = h(i, j) for all i, j < n.

Moreover, we may assume that ab < c < c . Going back through the denitions,
Gb , and this gives us that
one easily veries that a

a a = a + 2 a
a +2 Gb .
8.4. Two more variations on the square-bracket operation 269

Since the denition of G gives no restriction on sets whose supports lie above ,
we conclude that a Gb . It follows that
a a = a + 2 a
a +2 Gb ,
and therefore a a (a +2 Gb ) (a +2 Gb ), as required.
We prove the second conclusion by verifying that the members of the family
{ 0 } +2 G ) ({ 1 } +2 G ( < + )
of open subsets of G2 are pairwise-disjoint. Consider < < + , and suppose
that some pair (x, y) of nite subsets of S belongs to the intersection
({0 } +2 G ) ({1 } +2 G ) ({ 0 } +2 G ) ({ 1 } +2 G ).
Note that since no set from G can contain 0 or 1 , the set x must contain 0 .
Let x = { 0 } +2 x G . Then 0 x
for some x G , so by (2), we conclude
that f (, ) = 0. Working with the set y and the second coordinate, we similarly
conclude that f (, ) = 1. It follows that f (, ) = 2. Applying (3) and the fact
that 0 x G , we conclude that 1 x , and therefore 1 { 0 } +2 x
= x. It
follows that x { } +2 G , and so 1 must be a member of some nite set
1 0

belonging to G , contradicting (1). This nishes the proof. 


Corollary 8.4.25. There is a topological group G of cellularity 1 whose square has
cellularity > 1 . 
Remark 8.4.26. Theorem 8.4.21, Corollary 8.4.22 and Corollary 8.4.23 are due
to Shelah [93], [92], and [95] who also used the method of minimal walks for
proving these results. The example of topological group witnessing of the non-
productiveness of cellularity comes from p. 12 of [111] and from [115]. An exposi-
tion of the general problem of productiveness in the context of topological groups
can be found in [72]. Recall that the countable chain condition can be productive.
For example MA1 implies that the cartesian product of two posets satisfying
the countable chain condition also satises this chain condition( see, [36]). First
examples of cardinals with -cc non-productive without using additional set-
theoretical assumptions were given by the author in [107] (see also [111]). For
example, = b or = cf c are some of these cardinals. The rst examples of
non-productiveness of the + -chain condition that do not use additional axioms
of set theory were given by the author in [108] using what is today known under
the name pcf theory. For example, it has been proved in [108] that if = c+ ,
then the + -chain condition is not productive. After the full development of pcf
theory it became apparent that the basic construction from [108] applies to every
successor of a singular cardinal [94]. For an overview of recent advances in this
area, the reader is referred to [76]. The following problem, however, still seems
open.
Question 8.4.27. Suppose that is a regular strong limit cardinal and the -chain
condition is productive. Is necessarily a weakly compact cardinal?
Chapter 9

Unbounded Functions

9.1 Partial square-sequences


In this section is a regular cardinal and C ( < + ) is a xed C-sequence
with the property that tp(C ) for all < + . When the C-sequence is not
necessarily coherent, then it is natural to dene the corresponding mapping

: [+ ]2 (9.1.1)
as follows:

(, ) = sup{tp(C ), (, min(C \ )), (, ) : C }, (9.1.2)


with the boundary value (, ) = 0 for all < + , a denition that is slightly
dierent from the one given above in (7.3.2) above. Clearly,

(, ) 1 (, ) for all < < + , (9.1.3)


and so, using Lemma 6.2.1, we have the following fact.
Lemma 9.1.1. For < , < + the set P () = { : (, ) } has size
no more than || + 0 . 
The proof of the following subadditivity properties of is very similar to the
proof of the corresponding fact for the function from Section 7.3.
Lemma 9.1.2. For all < + ,
(a) (, ) max{(, ), (, )},
(b) (, ) max{(, ), (, )}. 
We mention a typical application of this function to the problem of existence
of partial square-sequences which, for example, have some applications in the pcf
theory (see [15]).
272 Chapter 9. Unbounded Functions

Theorem 9.1.3. For every regular uncountable cardinal < and stationary
{ < + : cf() = }, there is a stationary set and a sequence C ( )
such that:

(1) C is a closed and unbounded subset of ,


(2) C = C for every C C .

Proof. For each , choose = () < such that the set P< () = { < :
(, ) < } is unbounded in and closed under taking suprema of sequences of
size < . Then there is , < and stationary such that () = and
for all . Let C be a xed closed and unbounded subset of
tp(P< ()) =
of order-type . Finally, for set

C = { P< () : tp(P< ()) C}.

Using Lemma 9.1.2, one easily checks that C ( ) satises the conditions (1)
and (2). 

Another application concerns the fact exposed above in Section 7.6, that the
inequalities in Lemma 9.1.2(a), (b) are particularly useful, when has conality
. Also consider the well-known phenomenon rst discovered by K.Prikry (see
[55]), that in some cases, the conality of a regular cardinal can be changed to
, while preserving all cardinals.
Theorem 9.1.4. In any cardinal-preserving extension of the universe, which has no
new bounded subsets of , but in which has a conal -sequence diagonalizing the
lter of closed and unbounded subsets of restricted to the ordinals of conality
> , there is a sequence Cn ( lim (+ ), n < ) such that for all < in
lim(+ ):

(1) Cn is a closed subset of for all n,


(2) Cn Cm , whenever n m,

(3) = n< Cn ,
(4) lim(Cn ) implies Cn = Cn .

Proof. For < + , let D be the collection of all < for which P< () is
-closed, i.e., closed under supremums of all of its countable bounded subsets.
Clearly, D contains a closed unbounded subset of , restricted to conality >
ordinals. Note that

D and (, ) < imply that D . (9.1.4)

In the extended universe, pick a strictly increasing sequence n (n < ) which


converges to and has the property that for each < + , there is n < such
that m D for all m n. Let n() be the minimal such integer n.
9.2. Unbounded subadditive functions 273

Fix lim(+ ) and n < . If there is such that n n() and


lim(P<n ()), (9.1.5)
let () be the minimal such and let
Cn = P<n (()) . (9.1.6)
If such a does not exist, let Cn = for n < n() and let
Cn = P<n () (9.1.7)
for n n(). Clearly, this choice of Cn s satises (1),(2) and (3). To verify (4),
let be a limit point of some Cn .
Suppose rst that cf() = . If Cn was dened according to the case (9.1.6)
so is Cn . Since P<n (()) is -closed we conclude that P<n (()) which by
Lemma 9.1.2(a),(b) means that P< () = P< (()) for all n and there-
fore D() \ n D . It follows that n() n and therefore that () = . Hence,
Cn = Cn . If Cn was dened according to the case (9.1.7) then our assump-
tion lim(Cn ) and the fact that n n() means that Cn was dened accord-
ing to (9.1.6). Similarly as above we conclude that actually P<n () which
gives us n n() and () = . It follows that Cn = Cn in this case as well.
Suppose now that cf() > . If Cn was dened according to (9.1.6) then so
is Cn and so Pn (()) and Pn (()) are two -closed and unbounded
subsets of . So their intersection is unbounded in which by Lemma 9.1.2(a), (b)
gives us that Pn (()) = Pn (()) . It follows that Cn = Cn in this
case. Suppose now that Cn was dened according to (9.1.7). Then = satises
n n() and (9.1.5). So Cn was dened according to the case (9.1.6). It follows
that Pn (()) and Pn () are two -closed and unbounded subsets of
so their intersection is unbounded in . Applying Lemma 9.1.2(a),(b) we get that
Pn (()) = Pn (). It follows that Cn = Cn in this case as well. 
Remark 9.1.5. The combinatorial principle appearing in the statement of Theorem
9.1.4 is a member of a family of square principles that has been studied system-
atically by Schimmerling and others (see, e.g., [88]). It is denitely a principle
sucient for all of the applications of  appearing in Section 7.6 above.

9.2 Unbounded subadditive functions


We now turn our attention to various unboundedness properties of functions of
the form f : [+ ]2 .
Denition 9.2.1. A function f : [+ ]2 is unbounded if f  []2 is unbounded
in for every + of size . We shall say that such an f is strongly unbounded
if for every family A of size + , consisting of pairwise-disjoint nite subsets of + ,
and every < , there exists A0 A of size such that f (, ) > for all a,
b and a = b in A0 .
274 Chapter 9. Unbounded Functions

Lemma 9.2.2. If f : [+ ]2 is unbounded and subadditive (i.e., it satises the


two inequalities in Lemma 9.1.2(a), (b)), then f is strongly unbounded.
Proof. For < < + , set < if and only if f (, ) . Then our assumption
about f satisfying Lemma 9.1.2(a) and (b) reduces to the fact that each < is a
tree-ordering on + compatible with the usual ordering on + . Note that the
unboundedness property of f is preserved by any forcing notion satisfying the -
chain condition, so in particular no tree (+ , < ) can contain a Souslin subtree of
height . In the proof of Lemma 7.4.5 above we have seen that this property of
(+ , < ) alone is sucient to conclude that every family A of many pairwise-
disjoint subsets of + contains a subfamily A0 of size such that for every a = b in
A0 every a is < -incomparable to every b, which is exactly the conclusion
of f being strongly unbounded. 
Lemma 9.2.3. The following are equivalent for a regular uncountable cardinal :
(1) There is a structure (+ , , <, Rn )n< with no elementary substructure B of
size such that B is bounded in .
(2) There is an unbounded function f : [+ ]2 .
(3) There is a strongly unbounded subadditive function f : [+ ]2 .
Proof. Only the implication from (2) to (3) needs some argument. So let f :
[+ ]2 be a given unbounded function. Increasing f , we may assume that
each of its sections f : is one-to-one. For a set + , let the f -closure
of , CLf (), be the minimal with the property that f  []2 and
f1 () , whenever and < sup( ). Dene e : [+ ]2 by letting

e(, ) = sup( CLf ((, ) P(,) ())). (9.2.1)

We show the two subadditive properties for e:

e(, ) max{e(, ), e(, )} whenever . (9.2.2)

By Lemma 9.1.2(a) either (, ) (, ) in which case P(,) () P(,) ()


holds, or (, ) (, ) when we have P(,) () P(,) (). By the denition
of e, the rst case gives us the inequality e(, ) e(, ) and the second gives
us e(, ) e(, ). Then

e(, ) max{e(, ), e(, )} whenever . (9.2.3)

By Lemma 9.1.2(b) either (, ) (, ) in which case P(,) () P(,) ()


holds, or (, ) (, ) when we have P(,) () P(,) (). In the rst case
we get e(, ) e(, ) and in the second e(, ) e(, ).
By Lemma 9.2.2, to get the full conclusion of (3) it suces to show that e
is unbounded. So let be a given subset of + of order-type and let < be
a given ordinal. Since e(, ) (, ) for all < < + we may assume that
9.2. Unbounded subadditive functions 275

is bounded on , or more precisely, that for some and all < in ,


(, ) . For , let () be the minimal point of above . Then there
is and of size such that (, ()) = for all . Note that
the sequence of sets P () ( ) forms a chain under end-extension. Let
be the union of this sequence of sets. Since f is unbounded, there exist < in
such that f (, ) > . Pick above . Then , P () and therefore
e(, ()) f (, ) > . This nishes the proof. 
Remark 9.2.4. Recall that Changs conjecture is the model-theoretic transfer prin-
ciple asserting that every structure of the form (2 , 1 , <, . . .) with a countable
signature has an uncountable elementary submodel B with the property that B1
is countable. This principle shows up in many considerations including the rst
two uncountable cardinals 1 and 2 . For example, it is known that it is preserved
by ccc forcing extensions, that it holds in the Silver collapse of an 1 -Erdos cardi-
nal, and that it in turn implies that 2 is an 1 -Erdos cardinal in the core model
of Dodd and Jensen (see, e.g., [20],[55]).
Corollary 9.2.5. The negation of Changs conjecture is equivalent to the statement
that there exists e : [2 ]2 1 such that:
(a) e(, ) max{e(, ), e(, )} whenever .
(b) e(, ) max{e(, ), e(, )} whenever .
(c) For every uncountable family A of pairwise-disjoint nite subsets of 2 and
every < 1 , there exists an uncountable A0 A such that e(, ) >
whenever a and b for every a = b A0 . 
Remark 9.2.6. Note that if a mapping e : [2 ]2 1 has properties (a),(b) and
(c) of Corollary 9.2.5, then De : [2 ]2 [2 ]0 dened by

De {, } = { min{, } : e(, ) e{, }}

satises the weak form of Denition 7.5.13, where only the conclusion (a) is kept.
The full form of Denition 7.5.13, however, cannot be achieved assuming only
the negation of Changs conjecture. This shows that the function , based on a
1 -sequence, is a considerably deeper object than an e : [2 ]2 1 satisfying
Corollary 9.2.5(a),(b),(c).

Note that by Corollary 9.2.5, Changs conjecture is equivalent to the state-


ment that within every decomposition of the usual ordering on 2 as an increasing
chain of tree-orderings, one of the trees has an uncountable chain. Is it possi-
ble to have decompositions of  2 into an increasing 1 -chain of tree-orderings
of countable heights? It turns out that the answer to this question is equivalent
to a dierent well-known combinatorial statement about 2 rather than Changs
conjecture itself. Recall that f : [+ ]2 is transitive if

f (, ) max{f (, ), f (, )} whenever . (9.2.4)


276 Chapter 9. Unbounded Functions

Denition 9.2.7. Given a transitive map f : [+ ]2 , one denes

f : [+ ]2

recursively on < using the formula

f (, ) = sup{f (min(C \ ), ), tp(C ), f (, min(C \ )), f (, ) :


C },

with the boundary condition f (, ) = 0 for all .


Lemma 9.2.8. For every transitive map f : [+ ]2 the corresponding function
f : [+ ]2 has the following properties:
(a) f (, ) max{f (, ), f (, )} whenever ,
(b) f (, ) max{f (, ), f (, )} whenever ,
(c) |{ : f (, ) }| || + 0 for < and < + ,
(d) f (, ) f (, ) for all < < + .
Proof. Proofs of (a)(c) are almost identical to the corresponding proofs for the
function itself and so we avoid repetition. The inequality (d) is deduced from
the denition of f using the transitivity of f as follows:

f (, ) max f (i+1 , i ) f (, ),
i<n

where = 0 > 1 > > n1 > n = is the walk from to along the
given C-sequence. 
Remark 9.2.9. Transitive maps are frequently used combinatorial objects, espe-
cially when one works with quotient structures. Adding the extra subadditivity
condition Lemma 9.2.8(b), one obtains a considerably more subtle object which is
much less understood.
Denition 9.2.10. Let f : ( < + ) be a given sequence of functions
such that f < f whenever < (i.e., { < : f () f ()} is bounded
in whenever < ). Then the corresponding transitive map f : [+ ]2 is
dened by

f (, ) = min{ < : f () < f () for all }. (9.2.5)

Let f be the corresponding -function that dominates this particular f and


for < let <f be the corresponding tree-ordering of + , i.e.,

<f if and only if f (, ) . (9.2.6)

Lemma 9.2.11. Suppose f g for all < + where is the ordering of every-
where dominance. Then for every < the tree (+ , <f ) has height g().
9.3. Changs conjecture and 2 277

Proof. Let P be a maximal chain of (+ , <f ). f (, ) f (, ) for every


< in P . It follows that f () < f () g() for all < in P . So P has
order-type g(). 

Note that if we have a function g : which bounds the sequence


f ( < + ) in the ordering < of eventual dominance, then the new sequence
f = min{f , g} ( < + ) is still strictly < -increasing but now bounded by g
even in the ordering of everywhere dominance. So this proves the following result
of Galvin and Keisler (see [57], [48], [85]).
Corollary 9.2.12. The following two conditions are equivalent for every regular
cardinal .

(1) There is a sequence f : ( < + ) which is strictly increasing and


bounded in the ordering of eventual dominance.
(2) The usual order-relation of + can be decomposed into an increasing -
sequence of tree-orderings of heights < . 
Remark 9.2.13. The assertion that every strictly < -increasing + -sequence of
functions from to is < -unbounded is strictly weaker than Changs conjecture
and in the literature it is usually referred to as weak Changs conjecture. This
statement still has considerable large cardinal strength (see [21]). Also note the
following consequence of Corollary 9.2.12 which can be deduced from Lemmas
9.1.1 and 9.1.2 above as well.
Corollary 9.2.14. If is a regular limit cardinal (e.g., = ), then the usual order-
relation of + can be decomposed into an increasing -sequence of tree-orderings
of heights < . 

9.3 Changs conjecture and 2


Recall the ErdosRado arrow notation

()2

which succinctly expresses the statement that for every mapping f : []2 there
is some innite subset X of such that f is constant on its symmetric square [X]2 .
Recall also that the successor of the continuum is characterized as the minimal
cardinal having this property, or in other words, that

c+ = min{ : ()2 }.

We shall now see that considering an asymmetric version of this partition property,
i.e., replacing []2 by and replacing symmetric squares [X]2 by rectangles
278 Chapter 9. Unbounded Functions

X Y , one gets a characterization of a quite dierent cardinal. To this end, let


us use the arrow notation 1,1



to succinctly express the statement that for every map f : , there exist
innite sets A, B such that f is constant on their product. Let
1,1

2 = min{ : }.

Note that 1 < 2 c+ . The following result shows that 2 can have the minimal
possible value 2 and that, in general, the cardinal 2 can be considerably smaller
than the continuum.
Theorem 9.3.1. Changs conjecture is equivalent to the statement that every ccc
poset forces 2 > 2 .
Proof. We have already noted that Changs conjecture is preserved by ccc forcing
extensions, so in order to prove the direct implication, it suces to show that
Changs conjecture implies that every f : 2 2 is constant on the product
of two innite subsets of 2 . Given such an f , using Changs conjecture it is possible
to nd an elementary submodel M of H3 such that f M , = M 1 1
and = M 2 is uncountable. Choose i and uncountable such that
f (, ) = i for all . Choose 0 < such that

0 = { : f (0 , ) = i}

is uncountable. Such 0 exists by elementarity of M . Pick 0 0 above 1 . Note


that by the elementarity of M , for every 0 there exist unboundedly many
< such that f (, 0 ) = f (, ) = i. So we can choose 1 < above 0 such
that the set
1 = { 0 : f (1 , ) = i}
is uncountable. Then again for every 1 there exist unboundedly many <
such that f (, 0 ) = f (, 1 ) = f (, ) = i, and so on. Continuing in this way, we
build two increasing sequences {n } and {n } \1 such that f (n , m ) = i
for all n, m < . This shows that f is constant on a product of two innite sets,
nishing the proof of the direct implication.
Assume now that Changs conjecture fails and x e : [2 ]2 as in
Corollary 9.2.5. We shall describe a ccc forcing notion P which forces a mapping
f : 2 2 which is not constant on the product of any two innite subsets
of 2 . The conditions of P are simply maps of the form p : [Dp ]2 where Dp
is some nite subset of 2 such that
p(, ) = p(, ) whenever < < belong to Dp and
(9.3.1)
have the property that e(, ) > e(, ).
9.3. Changs conjecture and 2 279

We order P by letting p extend q if and only if p extends q as a function and


p(, ) = p(, ) whenever < belong to Dq and be-
(9.3.2)
longs to Dp \ Dq .
Let A be a given uncountable subset of P. Rening A, we may assume that
Dp (p A) forms a -system with root D and that Fp = e [Dp ]2 (p A)
form an increasing -system on 1 with root F . Moreover, we may assume that
for every p and q in A, the nite structures they generate together with e are
isomorphic via the isomorphism that xes D and F . By Corollary 9.2.5(c) there
is uncountable A0 A such that for all p, q A0 ,
e(, ) > max(F ) whenever Dp \ D and Dq \ D. (9.3.3)
We claim that every two conditions p and q from A0 are compatible. To see this,
extend p q to a mapping r : [Dp Dq ]2 which is one-to-one on new
pairs, avoiding the old values. To see that r is indeed a member of P we need
to check (9.3.1) for r. So let < < be three given ordinals from Dr such
that e(, ) > e(, ). We have to prove that r(, ) = r(, ). By the choice of
r we may assume that {, } and {, } are old pairs, i.e., that they belong to
[Dp ]2 [Dq ]2 . Since p and q both satisfy (9.3.1), we may consider only the case
when (say) Dp \ Dq , Dq \ Dp and D. Recall that e(, ) > e(, ) and
the subadditive properties of e (see Corollary 9.2.5(a),(b)) give us the equality
e(, ) = e(, ), i.e., e(, ) belongs to F contradicting (9.3.3). To see that r
extends p and q, by symmetry it suces to check that r extends, say, p. So suppose
< < are given such that < belong to Dp and belongs to Dq \ Dp . We
need to show that r(, ) = r(, ). By the choice of new pairs, this is automatic
if one of the two pairs is new. So we may consider the case when both and
are in D. Since r(, ) = q(, ) and r(, ) = q(, ), in this case the inequality
r(, ) = r(, ) would follow from (9.3.1) for q if we show that e(, ) > e(, ).
Otherwise e(, ) would belong to the root F and by the fact that all conditions
are isomorphic, e((s), ) F holds for all s A0 , where (s) Ds (s A0 )
are the copies of . This of course contradicts the property Corollary 9.2.5(c) of e,
since (s) = (t) when s = t.
Forcing with P gives us a mapping g : [2 ]2 such that
{ < : g(, ) = g(, )} is nite for all < < 2 . (9.3.4)
Dene f : 2 2 by letting f (, ) = 2g{, } + 2 when < , f (, ) =
2g{, } + 1 when > , and f (, ) = 0 when = . Then f is not constant
on any product of two innite sets. 
In view of considerable technical diculties one encounters while trying to
design a partial ordering that would force that 2 > 3 , the following question is
natural.
Question 9.3.2. Can one prove a bound like 2 3 without appealing to addi-
tional set-theoretic axioms?
280 Chapter 9. Unbounded Functions

Remark 9.3.3. One can of course consider also the higher-dimensional analogue
n of 2 dened as the minimal cardinal with property that every mapping
f : n is constant on a product of n innite subsets of . The problem of
comparison of the two sequences of cardinals 2 , 3 , 4 ,. . . and 2 , 3 , 4 , . . . is
of considerable interest, both in set theory and model theory (see, e.g., [90], [114],
[120]). In particular, the following natural questions are still unanswered.
Question 9.3.4. Can one prove any of the bounds of the form n n+1 , or in
general, can one determine the function : Ord satisfying n = (n) without
appealing to additional axioms of set theory? Is n n for all n ?
It turns out that the cardinal 2 is also related to the following standard
concept from set theory.
Denition 9.3.5. A cardinal is a J onsson cardinal if every algebra of the form
, (fi )i<  has a proper subalgebra of cardinality , or equivalently, if for every
mapping f : []< there is X of cardinality such that f  [X]< = .
Note that our previous results about the square-bracket operation show that
a Jonsson cardinal cannot be a successor of a regular cardinal and, in fact, the
following simple stepping-up procedure shows that a long initial segment of suc-
cessors of singular cardinals also fail to be J
onsson.
Lemma 9.3.6. Suppose that n (n < ) is a strictly increasing sequence of regu-
lar cardinalswith the property that for some non-principal ultralter U on the
ultrapower n n /U has conality + where = sup n . Suppose that for every
n < there cn : [n ]2 n such that cn [A]2 = n for all A n of size n . Then
there is c : [+ ]2 + such that c [A]2 = + for all A + of size + .
Proof.
 Let f ( < + ) be an increasing and conal sequence in the ultrapower
+ 2
n n /U. Dene c : [ ] by

c({, }) = cn ({f (n), f (n)}) where n = min{i : f (i) = f (i)}. (9.3.5)


 2 +
We claim that c [A] = for all unbounded A . To see this, xing unbounded
A + and < , let n be the minimal integer i such that
n > and sup{f (k) :
A} = i . Then we can nd B A of size n and t i<n i such that:
(1) f  n = t for all B, and
(2) f (n) < f (n) whenever < belong to B.
Since {f : B} is unbounded in n , by the property of cn , we can nd <
in B such that cn ({f (n), f (n)}) = . Since n = min{k : f (i) = f (i)}, we
conclude that c({, }) = cn ({f (n), f (n)}) = .
For each < + , x a one-to-one mapping e : . Dene c : [+ ]2 + ,
by
c(, ) = e1
(c(, )), (9.3.6)
if c(, ) rang(e ); put c(, ) = 0, otherwise. Then it is easily checked that
c [A]2 = + for all A + of size + . 
9.3. Changs conjecture and 2 281

Corollary 9.3.7. There is c : [+1 ]2 +1 such that c [A]2 = +1 for all
unbounded A +1 .

Proof. This follows from Lemma 9.3.6 and a standard fact from the PCF theory
(see, e.g., [94]). 

Singular cardinals of larger conalities are handled by the following stepping-


up procedure which uses a slightly dierent assumption.
Lemma 9.3.8. Let ( < ) be a sequence of regular cardinals such that for
every < , we have that > sup< and we can nd c : [ ]2 such
that c [A]2 = for all unbounded A . Suppose that there is a sequence f

( < + ) of elements of < that is increasing and conal in this product
relative to the ordering of eventual dominance. Let = sup . Then there is
c : [+ ]2 + such that c [A]2 = + for all unbounded A + .

Proof. Working as before, it suces to construct c : [+ ]2 such that c [A]2 =


for all unbounded A + . For < < + , let

c(, ) = c (f (), f ()), where = max{ < : f () f ()}, (9.3.7)

if such maximum < exists and if f () = f (); otherwise, let c(, ) = 0.


Consider an unbounded A + and < . For each < and < ,
pick,if possible, an ordinal = (, ) A with property that f () = . Dene
g , by

g() = sup{f () : = (, ) for some < < }. (9.3.8)



Since f ( < + ) is conal in relative to the ordering of eventual domi-
nance, we can nd conal B A and 0 < such that

g() < f () for all 0 and B. (9.3.9)

Then, in particular, we have that

(, ) < for all < , < and B. (9.3.10)



Since f ( B) is conal in , we can nd 1 0 such that the set
X = {f (1 ) : B} is unbounded in 1 . So by the property of c1 , we can nd
< in X such that c1 (, ) = . Then = (1 , ) is dened. Pick B such
that = f (1 ). Then by (9.3.9) and (9.3.10), we have that < , f (1 ) > f (1 ),
and f () g() < f () for all > 1 . It follows that

c(, ) = c1 (f (1 ), f (1 )) = c1 (, ) = ,

as required. 
282 Chapter 9. Unbounded Functions

Corollary 9.3.9. Suppose that is a singular cardinal and that for some < ,
for every regular cardinal [, ) there is c : []2 such that c [A]2 = for
all unbounded A . Then there is c : [+ ]2 + such that c [A]2 = + for all
unbounded A + .

Proof. This follows again from Lemma 9.3.8 and a standard fact from the PCF
theory (see, e.g., [94]). 

Corollary 9.3.10. There is c : [1 +1 ]2 1 +1 such that c [A]2 = 1 +1 for all


unbounded A 1 +1 . 

The following fact explains the relationship between the class of J


onsson
cardinals and the cardinal 2 .
Theorem 9.3.11. There are no J
onsson cardinals < 2 .

Proof. Suppose that is a Jonsson cardinal and let f : 2 be a given mapping.


Then we can nd an elementary submodel M of (H+ , ) such that f M , such
that M has cardinality , but  M . Let = min( \ M ) and let + be
the minimal cardinal > . Choose B M of size + and n < such that
f (, ) = n for all B. Then by the elementarity of M , for every nite subset
F B, the set
AF = { < : f (, ) = n}
is unbounded in . Since B has bigger cardinality than , there must be 0 <
such that
B0 = { B : f (0 , ) = n}
has cardinality + . Let 0 = min B0 . Then A{0 ,} is unbounded in for all
B0 . So there must be 1 < , such that 1 > 0 and such that

B1 = { B0 : > 0 and f (1 , ) = n}

has cardinality + , and so on. Proceeding in this way we constrict an increasing


sequence i (i < ) of ordinals < and an increasing sequence i (i < ) such
that
f (i , j ) = n for all i, j < .
Since f was an arbitrary mapping this shows that 2 , as required. 

It turns out also that cardinals < 2 share also the following strong un-
boundedness property.
Theorem 9.3.12. For every < 2 there is f : 2 such that, for every
uncountable family A of pairwise-disjoint nite subsets of , all of some xed size
n, and for every k < , there exist a = b in A such that f (a(i), b(j)) > k for all
i, j < n.
9.4. Higher dimensions and the continuum hypothesis 283

Proof. From the denition of the cardinal 2 , we infer that there exist a mapping
f : 2 which is not constant on any product of two innite subsets of .
Suppose that for some k < such pair a = b in A cannot be found. Let A0 and A1
be two disjoint subsets of A such that A0 is countable and innite and such that
A1 is uncountable. Choose a uniform ultralter U0 on A0 and a uniform ultralter
U1 on A1 . Then there exist k k and i0 , j0 < n such that

(U1 b)(U0 a) f (a(i0 ), b(j0 )) = k. (9.3.11)

Let X = {a(i0 ) : a A0 } and let



Y = {b(j0 ) : b A1 and (U0 a) f (a(i0 ), b(j0 )) = k}.

Then X is countable and innite while the set Y is uncountable as it belongs to


U1 . Moreover, by (9.3.11), for every nite F Y , the set

XF = { X : f (, ) = k for all F }

is innite as it belongs to U0 . So working as in the previous proof, we can select


a one-to-one sequence l (l < ) of elements of X and a one-to-one sequence l
(l < ) of elements of Y such that

f (l , m ) = k for all l, m < .

This contradicts our initial assumption that the mapping f is not constant on any
product of two innite subsets of . 

9.4 Higher dimensions and the continuum hypothesis


Recall that the minimal cardinal for which there is a ccc poset of size that is
not -centered1 is equal to the Baire category number m, the minimal cardinal
for which one can nd a compact ccc space without isolated points that can
be decomposed into nowhere-dense subsets (see [104]). What is the minimal
cardinality of a ccc poset that cannot be decomposed into 1 centered subsets?
In fact, it is natural to consider this question by restricting ourselves to a particular
nite dimension n.
Denition 9.4.1. For an integer n 2, an n-hypergraph is simply a pair of the form
(V, H), where H is a collection of n-element subsets of V , i.e., H [V ]n . Thus, a
graph is a 2-hypergraph and a triple-system is a 3-hypergraph. An n-hypergraph
(V, H) satises the countable chain condition if for every uncountable family A of
nite subsets of V , either there is a A such that [a]n H = , or else there exist
a = b in A such that [a b]n H = .
1A subset X of a partially ordered set P is centered if for every nite subset F of X there is
q P such that q p for all p F . A poset P is -centered if it can be decomposed into at
most centered subsets. Of course, -centered signies 0 -centered.
284 Chapter 9. Unbounded Functions

Let n be the minimal cardinal for which one can nd a ccc n-hypergraph
on a vertex set of size that is not 1 -chromatic2 . First of all, we should mention
the following immediate inequalities.
Lemma 9.4.2. n n+1 > 1 for every integer n 2. 
It turns out that these cardinals do have dramatic inuence on standard
cardinal characteristics of the continuum. For example, we have the following facts
proved in [114].
Theorem 9.4.3. 2 > b implies b = 1 , while 3 > cf(c) implies cf(c) = 1 and
3 > d implies d = 1 . 
It follows that cardinal inequalities like n > c can be considered as special
forms of the Continuum Hypothesis, one for each nite dimension n. In fact, it is
still unknown if the inequality > 1 is equivalent to CH. So, in some sense, it
is not surprising that these inequalities do bear some relationship to the cardinal
2 which was meant to be the Ramsey-theoretic analogue of the successor of the
continuum.
Theorem 9.4.4. If 3 > c, then 2 = 2 .
Proof. Suppose towards a contradiction that 2 > 2 and x p : 2 2
that is not constant on any product of two innite subsets of 2 . By Theorem
9.4.3, we can also x a sequence f ( < 1 ) of increasing functions from into
which is unbounded in relative to the ordering of eventual dominance. We
shall also consider as a partially ordered set with the ordering of everywhere
dominance:
f g i f (n) g(n)for all n <
and so we will say that f, g are incomparable if f  g and f  g. Thus,
an antichain of is any subset A of with the property that every pair of
distinct members of A are incomparable. Similar terminology will be adopted when
working with the nite powers m and their coordinatewise orderings rather
than the innite power and its coordinatewise ordering . Since the Continuum
Hypothesis must fail, we may also assume to have a one-to-one sequence r ( <
2 ) of elements of 2 . Fix also e : [2 ]2 1 such that e(, ) : 1 is one-to-
one for all < 2 . Now we are ready to dene our triple-system (2 , H). In fact,
it will be more convenient to dene the complement H c = [2 ]3 \ H. So, let H c
be the collection of all triples < < such that

r  p(, ) = r  p(, ) implies {fe(,) , fe(,) } is an antichain. (9.4.1)

The proof that the triple-system (2 , H) satises the countable chain condition
and the nal step towards a contradiction depends on the oscillation theory of ,
2 Recall
that an n-hypergraph (V, H) is -chromatic if the set V can be decomposed into subsets
X with the property that [X]n H = .
9.4. Higher dimensions and the continuum hypothesis 285

so let us reproduce some facts from this theory. More information can be found in
Chapter 1 of [111].
For f and g in , let

osc(f, g) = |{n < : f (n) g(n) and f (n + 1) > g(n + 1)}|. (9.4.2)

Note that if osc(f, g) = 0 and if f (0) g(0), then f (n) g(n) for all n < ,
i.e., f g. Conversely, if f g, then osc(f, g) = 0. The analysis of the oscillation
mapping requires its natural extension to ( )2 as follows. Given s and t in ,
let

osc(s, t) = |{n < min(|s|, |t|) 1 : s(n) t(n) and s(n + 1) > t(n + 1)}|. (9.4.3)

For f and g in , and m < , set

f <m g i f (n) < g(n) for all n m. (9.4.4)

We are now ready to state explicitly a consequence of Lemma 1.0 of [111] that will
be useful in the rest of the proof of Theorem 9.4.4.
Lemma 9.4.5. Suppose X and Y are two subsets of that are totally ordered
and unbounded relative to the ordering < of eventual dominance. Suppose that
there is a topologically dense subset D of X such that for some m < , h ,
and s, t < with |s| = |t| > m, we have:
(a) f  |s| = s for all f X,
(b) g  |t| = t for all g Y ,
(c) f < h <m g for all f D and g Y ,
(d) sup{g(|t|) : g Y } = .
Then there exist increasing sequences sl (1 l < ) and tl (1 l < ) of
elements of < and sequences fl (l < ) and gl (l < ) of elements of D and Y ,
respectively, such that for all l 0 :
(e) s1 = s, t1 = t, and |tl1 | < |sl | < |tl |,
(f) fl  |sl | = sl , and gl  |tl | = tl ,
(g) fl <|sl | h and g2l (|t2l1 |) > h(|s2l |),
(h) sup{f (|sl |) : f  |sl | = sl and f X} = ,
(i) sup{g(|tl |) : g  |tl | = tl and g Y } = .
So, in particular osc(fl , gl ) = osc(s, t) + l for all 0 l < . 
Corollary 9.4.6. For every subset X of that is totally ordered and unbounded
in the ordering of eventual dominance and for every l < , there exist f and g in
X such that osc(f, g) = l. 
286 Chapter 9. Unbounded Functions

We shall also need the following fact that is an immediate consequence of the
Claim on p. 62 of [111].
Lemma 9.4.7. Let X be a subset of that is totally ordered in the ordering of
eventual dominance. Suppose that A is a family of pairwise-disjoint nite subsets
of X, all of some xed size n, such that for every a = b in A either a < b or
b < a.3 Let P(A) be the poset of all nite subsets F of A such that a(i)  b(j)
for all a = b in F and i, j < n. Then P(A) satises the productive countable chain
condition.4 
We are now ready to continue the proof of Theorem 9.4.4. We rst prove
that the triple system (2 , H) satises the countable chain condition. So let A be
a given uncountable family of nite subsets of 2 such that [a]3 H = for all
a in A. We need to nd a = b in A such that [a b]3 H = . We may assume
that elements of A are all of some xed size n and that A forms a -system with
root w. Shrinking A further, we may assume that for some xed integer m and all
a A,
p(, ) < m and (r , r ) < m for all = in a. (9.4.5)
Moreover, we may assume that for all a A,

{fe(,)  m, fe(,)  m} is an antichain for all {, , } [a]3 H c . (9.4.6)

Applying the -system lemma again, we may assume that for each w and
l < m, the family

a (l) = {e(, ) : a and p(, ) = l} (a A) (9.4.7)

of nite subsets of 1 forms a -system with root v (l). Applying Lemma 9.4.7,
and going perhaps to some uncountable subset of A in some ccc forcing extension,
we may assume that for all a = b in A, and w and l < m,

f  f for all a (l) \ v (l) and b (l) \ v (l). (9.4.8)

Moreover, we may assume that the structures of the form

(a, <, w, e  [a]2 , p  [a]2 , r  m, v (l))a,l<m (9.4.9)

are all isomorphic. Split A into two uncountable disjoint subsets A0 and A1 such
that for some m m and all a A0 and b A1 ,

for all a and b.


(r , r ) < m (9.4.10)

Applying Theorem 9.3.12, we can nd a A0 and b A1 such that

for all a \ w and b \ w.


p(, ) > m (9.4.11)
3 Here a< <
b signies the fact that a(i) b(j) for all i, j < n.
4 I.e., the product of P(A) with every ccc poset Q satises the ccc.
9.4. Higher dimensions and the continuum hypothesis 287

We claim that [a b]3 H = . Consider a triple < < from a b. We need


to check that it belongs to H c , i.e., that it satises the condition (9.4.1). So let us
assume that
r  p(, ) = r  p(, ) (9.4.12)
and let l = p(, ) = p(, ). Then either l < m or l > m. Consider rst the case
l < m. Then the pairs {, } and the pairs {, } must be included in one of the
sets a or b. It follows that in particular w. Only the subcase a \ w and
b \ w requires some argument. Let  a be the copy of with respect to
the isomorphism of structures of the form 9.4.9 associated with a and b. If = 
then {fe(,) , fe(,  ) } forms an antichain since [a]2 H = . Then by (9.4.6), the
pair
{fe(,)  m, fe(,  )  m}
forms an antichain of the poset ( m , ), and since by the isomorphism

fe(,)  m = fe(,  )  m,

we get the desired conclusion that {fe(,) , fe(,) } forms an antichain.


Consider now the subcase =  , i.e., the case when and correspond to
each other in the isomorphism of the structures (9.4.9). Then the ordinal e(, )
belongs to the root v (l) if and only if the ordinal e(, ) belongs to v (l) in which
case they are equal, and therefore, {fe(,) , fe(,) } forms the trivial one-element
antichain. If these two ordinals do not belong to v (l), then by (9.4.8), the pair
{fe(,) , fe(,) } forms an antichain as well.
It remains to consider the case l > m. This can only happen if belongs to
one of the set a \ w or b \ w and and to the other. So, in particular and
belong to the same set a or b. But then by (9.4.5), r  m = r  m, and therefore,
r  l = r  l, contradicting (9.4.12) and the fact that l = p(, ) = p(, ). This
shows that this case is in fact not possible and nishes the proof that the triple
system (2 , H) satises the countable chain condition.
Since the triple system (2 , H) satises the countable chain condition, our
assumption 3 > c applies to it, and so we have a decomposition of 2 into 1
subsets X with the property that [X]3 H = . So, in particular, there is an
unbounded subset X of 2 such that [X]3 H = . Since the mappings e(, )s
are one-to-one there must be X, l < and t l such that

Y = {e(, ) : X, > , p(, ) = l and r  l = t}

is unbounded in 1 . By Corollary 9.4.6, there exist < in Y such that f and


f are comparable, and in fact f f . Choose and in X above such that

p(, ) = l = p(, ) and r  l = t = r  l,

and = e(, ) and = e(, ). Going back to the denition (9.4.1) of H c , we


conclude that the triple {, , } does not belong to H c , since on one hand, the
288 Chapter 9. Unbounded Functions

hypothesis of (9.4.1) holds but f = fe(,) and f = fe(,) are comparable. This
shows that [X]3 H = , a contradiction. The proof of Theorem 9.4.4 is now
completed. 
We nish this section with the following natural question asking if the di-
mension in Theorem 9.4.4 can be lowered.
Question 9.4.8. Does 2 > c imply that 2 = 2 ?
Of course, one is also interested in determining how much of the Continuum
Hypothesis is captured by inequalities between the continuum and the cardinal
numbers like or n .
Question 9.4.9. Is any of the inequalities like > c or n > c equivalent to the
Continuum Hypothesis?
Chapter 10

Higher Dimensions

10.1 Stepping-up to higher dimensions


The reader must have already noticed that in this book so far, we have only
considered functions of the form f : []2 I or equivalently sequences
f : I ( < )
of one-place functions. To obtain analogous results about functions dened on
higher-dimensional cubes []n , one usually develops some form of stepping-up pro-
cedure that lifts a function of the form f : []n I to a function of the form
g : [+ ]n+1 I. The basic idea seems quite simple. One starts with a coherent
sequence e : ( < + ) of one-to-one mappings and wishes to dene
g : [+ ]n+1 I as follows:
g(0 , 1 , . . . , n ) = f (e(0 , n ), . . . , e(n1 , n )). (10.1.1)
In other words, we use en to send {0 , . . . , n1 } to the domain of f and then
apply f to the resulting n-tuple. The problem with such a simple-minded denition
is that for a typical subset of + , the sequence of restrictions e  ( ) ( )
may not cohere, so we cannot produce a subset of that would correspond to
and on which we would like to apply some property of f . It turns out that
the denition (10.1.1) is basically correct except that we need to replace en by
e (n2 ,n1 ,n ) , where
: [+ ]3 + (10.1.2)
is dened as follows (see Denition 8.1.7):
(, , ) = t , where t = 0 (, ) 0 (, ). (10.1.3)
The function 0 to which (10.1.3) refers is of course based on some C-sequence
C ( < + ) on + . The following result shows that if the C-sequence is carefully
chosen, the function will serve as a stepping-up tool.
290 Chapter 10. Higher Dimensions

Figure 10.1: The characteristic (, , ).

Lemma 10.1.1. Suppose 0 and are based on some  -sequence C ( < + ) and
let be a regular uncountable cardinal . Then every set + of order-type
contains a conal subset such that, if = sup() = sup(), then 0 (, ) =
0 (, (, , )) for all < < < in .

Proof. For lim(C ), let be the minimal element of above . Let be the
maximal point of the trace Tr(, ) of the walk from to along the  -sequence
C ( < + ). Thus = or = min(Tr(, )\( +1)) and C is bounded in
for all Tr(, ) strictly above . Let f () be a member of C which bounds
all these sets. By the Pressing Down Lemma there is a stationary set lim(C )
and C such that f () for all . Shrinking we may assume that
< whenever < are members of . Note that

= min(Tr( , ) Tr( , )) for every < < in . (10.1.4)

In other words, ( , , ) = for all < < in . By the denition of


for lim(C ), belongs to C and so by the implicit assumption that nonlimit
points of C are successor ordinals, we conclude that is a limit point of C .
10.1. Stepping-up to higher dimensions 291

Thus the section (0 ) of the function 0 is an initial part of both the sections
(0 ) and (0 ) , which is exactly what is needed for the conclusion of Lemma
10.1.1. 
Recall that for a given C-sequence C ( < + ) such that tp(C ) for
all < + , the range of 0 is the collection of all nite sequences of ordinals
< . It will be convenient to identify Q with via, for example, G odels coding
of sequences of ordinals by ordinals. This identication gives us a well-ordering
<w of Q of length . We use this identication with lift-up of an arbitrary map
f : []n I (really, a map of the form f : []n Q ) to a map g : [+ ]n+1 I
by the following formula:

g(0 , . . . , n1 , n ) = f (0 (0 , ), . . . , 0 (n1 , )), (10.1.5)

where = (n2 , n1 , n ).
Let us examine how this stepping-up procedure works on a particular example
motivated by a problem from [43].
Theorem 10.1.2. Suppose is an arbitrary cardinal for which  holds. Suppose
further that for some regular > and integer n 2 there is a map f : []n
[[]< ]< such that:
(1) A min(a) for all a []n and A f (a).
(2) For all < and of size there exist a []n and A f (a) such that
tp(A) and A .
Then + and satisfy the same combinatorial property, but with n + 1 in place
of n.
Proof. We assume that the domain of f is actually equal to the set [Q ]n rather
than []n and dene g by

g(0 , . . . , n1 , n ) = (0 )1
(f (0 (0 , ), . . . , 0 (n1 , ))), (10.1.6)

where = (n2 , n1 , n ), and where 0 and are based on some xed  -


sequence. Note that the transformation does not necessarily preserve (1), so we
intersect each member of a given g(a) with min(a) in order to satisfy this condition.
To check (2), let be a given set of size . By Lemma 10.1.1, shrinking we
may assume that has order-type and that if = sup(), then

0 (, ) = 0 (, (, , )) for all < < in . (10.1.7)

It follows that g restricted to []n+1 satises the formula

g(0 , . . . , n1 , n ) = (0 )1
(f (0 (0 , ), . . . , 0 (n1 , ))). (10.1.8)

Shrinking further we assume that the mapping (0 ) , as a map from the ordered
set (+ , ) into the ordered set (Q , <w ) is strictly increasing when restricted to .
292 Chapter 10. Higher Dimensions

Given an ordinal < , we apply (2) for f to the set = {0 (, ) : } and


nd a []n and A f (a) such that tp(A) and A . Let {0 , . . . , n1 }
1
be the increasing enumeration of the preimage (0 ) (a) and pick n above
1
n1 . Let B be the preimage (0 ) (A). Then B g(0 , . . . , n1 , n ), tp(B)
and B . This nishes the proof. 
Remark 10.1.3. It should be noted that Velleman [128] was the rst to attempt
to step-up the combinatorial property of Hajnal and Komjath [43] appearing in
Theorem 10.1.2 using his version of the gap-2 morass. Unfortunately the stepping-
up procedure of [128] worked only up to the value n = 3 so it remains unclear if
the higher-gap morass could be used to reach all the other values of n.
If we apply this stepping-up procedure to the projection [[]] of the square-
bracket operation dened in Denition 5.2.1, one obtains analogues of families
G, H and K dened in the course of the proof of Theorem 5.3.10 for 2 instead of
1 . To make this precise, we rst x a 1 -sequence C ( < 2 ) and consider
the corresponding mappings,

0 : [2 ]2 1< and : [2 ]2 1 .

Let : [2 ]2 1 be dened by

(, ) = 2(,) (2 tp{ : (, ) (, )} + 1).

Then as in the proof of Lemma 3.2.2 one shows that has the following three
useful properties.
Lemma 10.1.4. For all < < < 2 ,
(a) (, ) = (, ),
(b) (, ) max{(, ), (, )},
(c) (, ) max{(, ), (, )}. 
Moreover, we have the following coherence property of that is also quite
useful.
Lemma 10.1.5. =  whenever lim(C ). 
Denition 10.1.6. The ternary operation [[[]]] : [2 ]3 2 is dened as follows:

(,,))1 [[
[[[]]] = ( (,,)() (,,)()]].

It should be clear that the stepping-up method lifts the property of Lemma
5.2.2 of [[]] to the following property of the ternary operation [[[]]]:
Lemma 10.1.7. Let A be a given uncountable family of pairwise-disjoint subsets of
2 , all of some xed size n. Let 1 , . . . , n be a sequence of ordinals < 2 such that
for all i = 1, . . . , n and all but countably many a A we have that i < a(i). Then
there exist a, b and c in A such that [[[a(i)b(i)c(i)]]] = i for i = 1, . . . , n. 
10.1. Stepping-up to higher dimensions 293

As in Theorem 5.3.10, let us identify 2 and 3 [2 ]< and view [[[ ]]] as a
function
[[[ ]]] : [2 ]3 3 [2 ]<
having the range 3 [2 ]< rather than 2 . Let
[[[ ]]]0 : [2 ]3 3 and [[[ ]]]1 : [2 ]3 [2 ]<
be the two projections of [[[]]]. Set
G = {G [2 ]< : [[[]]]0 = 0 for all {, , } [G]3 },
H = {H [2 ]< : [[[]]]0 = 1 for all {, , } [H]3 }.
Note the following immediate consequence of Lemma 10.1.7.
Lemma 10.1.8. For every uncountable family A of disjoint 2-element subsets of
2 there is an arbitrarily large nite subset B of A with {b(0) : b B} G and
{b(1) : b B} H. 
Recall that the parameter in all our denitions so far is a xed 1 -sequence
C ( < 2 ) and the corresponding characteristics of the walk 0 , , , , etc. We
use them to dene an analogue of the family K of Theorem 5.3.10 as follows. Let
K now be the collection of all nite sets {ai : i < n} of elements of [2 ]2 such that
for all i < j < k < n:
a i < aj < ak , (10.1.9)
(, ) (, ) holds if ai , aj and ak , (10.1.10)
[[[ai (0)aj (0)ak (0)]]]0 = [[[ai (1)aj (1)ak (1)]]]0 = 2, (10.1.11)

[[[ai (0)aj (0)ak (0)]]]1 = [[[ai (1)aj (1)ak (1)]]]1 = al . (10.1.12)
l<i
The following property of K is a consequence of Lemma 10.1.7 and the tri-
angle inequalities of .
Lemma 10.1.9. For every sequence a ( < 1 ) of 2-element subsets of 2 such
that a < a whenever < , and for every positive integer n, there exist 0 <
1 < < n1 < 1 such that {ai : i < n} K. 
The particular denition of K is made in order to have the following property
of K which is analogous to (5.3.15) of Section 5.1:
If K and L are two distinct members of K, then there are
no more than nine ordinals such that {, } K and (10.1.13)
{, } L for some = .
The denitions of G, H and K immediately lead to the following analogues of
(5.3.13) and (5.3.14):
G and H contain all singletons, are closed under subsets
(10.1.14)
and are 3-orthogonal to each other.
G and H are both 5-orthogonal to the family of all unions
(10.1.15)
of members of K.
294 Chapter 10. Higher Dimensions

So we can proceed as in Theorem 5.3.10 and for a function x from 2 into R,


dene
  12


x
H,2 = sup x()2 :HH ,

H

12





x
K,2 = sup (x() x())2 : K K .

{,}K

2 = {x :
x
< }, and let E2
Let

= max{

,

H,2 ,

K,2 }, let E
be the closure of the linear span of {1 : 2 } inside (E 2 ,

). Then as
in Theorem 5.3.10 one shows, using the properties of G, H and K listed above in
Lemmas 10.1.8 and 10.1.9 and in (10.1.1010.1.15), that E2 is a Banach space with
{1 : 2 } as its transitive basis and with the property that every bounded
linear operator T : E2 E2 can be written as S +D, where S is an operator with
separable range and D is a diagonal operator relative to the basis {1 : 2 }
with only countably many changes of the constants. The new feature here is the
use of (10.1.10) in checking that the projections on countable sets of the form
{ < : (, ) < } ( < 2 , < 1 ) are (uniformly) bounded. This is needed
among other things in showing, using Lemma 10.1.8, that T can be written as a
sum of one such projection and a diagonal operator relative to {1 : 2 }.
Using the interpolation method of [17], one can turn E2 into a reexive
example. Thus, we have the following:

Theorem 10.1.10. Assuming 1 , there is a reexive Banach space E with a tran-


sitive basis of type 2 with the property that every bounded operator T : E E
can be written as a sum of an operator with a separable range and a diagonal op-
erator (relative to the basis) with only countably many changes of constants. 

Remark 10.1.11. In [61], P. Koszmider has shown that such a space cannot be
constructed on the basis of the usual axioms of set theory. We refer the reader to
that paper for more details about these kinds of examples of Banach spaces.

10.2 Changs conjecture as a 3-dimensional


Ramsey-theoretic statement
In this section we shall examine the stepping-up method with less restrictions on
the given C-sequence C ( < + ) on which it is based. This will lead us to
an interesting formulation of Changs Conjecture as a purely Ramsey-theoretic
statement.
10.2. Changs conjecture as a 3-dimensional Ramsey-theoretic statement 295

Theorem 10.2.1. The following are equivalent for a regular cardinal such that
log + = . 1
(1) There is a substructure of the form (++ , + , <, . . .) with no substructure B
of size + with B + of size .
(2) There is f : [++ ]3 + which takes all the possible values on the cube of
any subset of ++ of size + .
Proof. To prove the nontrivial direction from (1) to (2), we use Lemma 9.2.3 and
choose a strongly unbounded and subadditive e : [++ ]2 + . We also choose
a C-sequence C ( < + ) such that tp(C ) for all < + and consider
the corresponding function : [+ ]2 dened above in (9.1.2). Finally we
+
choose a one-to-one sequence r ( < ++ ) of elements of {0, 1} and consider
the corresponding function : [++ ]2 + :

(, ) = (r , r ) = min{ : r () = r ()}. (10.2.1)

The denition of f : [++ ]3 is given according to the following two rules


applied to a given triple x = {, , } [++ ]3 ( < < ):
Rule 1: If (r , r ) < (r , r ) and r <lex r <lex r or r >lex r >lex r , let

f (, , ) = min(P ((, )) \ (, )),

where

= (min{ (, ) : (, (, )) = (, (, ))}, (, )).

Rule 2: If x is such that r is lexicographically between the other two r s for


x, if x \ {} is such that (r , r ) > (r , r ), where is the remaining
element of x and if x does not fall under Rule 1, let

f (, , ) = min(P (e(, )) \ e(, )),

where = {(, ), e(, )}.


The proof of the theorem is nished once we show the following: for every
stationary set of conality ordinals < + and every ++ of size + there
exist < < in such that f (, , ) . Clearly we may assume that has
order-type + . Let
R = {r : }.
Case 1: R contains a subset of size + which is lexicographically well ordered or
conversely well ordered. Going to a subset of , we may assume that

(r , r ) = (r , r ) for all < < in . (10.2.2)


1 log = min{ : 2 }.
296 Chapter 10. Higher Dimensions

It follows that Rule 1 applies in the denition of f (, , ) for any triple < <
of elements of . Let = {(r , r ) : {, } []2 }. Then is a subset of +
of size + and the range of f on []3 includes the range of [] on []2 , where []
is the square-bracket operation on + dened as follows:
[] = min(P () \ ), (10.2.3)
where = (min{ : (, ) = (, )}, ). Note that this is the exact
analogue of the square-bracket operation that has been analysed above in Section
5.1. So exactly as in Section 5.1, one establishes the following fact.
For + unbounded, the set {[] : {, } []2 }
(10.2.4)
contains almost all ordinals < + of conality .
Case 2: R contains no subsets of size + which are lexicographically well ordered
+
or conversely well ordered. For t {0, 1}< , let t = { : t r }. Let
+
T = {t {0, 1}< : |t | = + }.
Shrinking we may assume that T is either a downwards closed subtree of
+
{0, 1}< of size and height < + , or a downwards closed Aronszajn sub-
+
tree of {0, 1}< of height + . For , let
e  + = {e{, } : , r <lex r }. (10.2.5)

Case 2.1: There exist + many such that the set e  + is bounded for all
. Choose an elementary submodel M of some large enough structure of the
form H such that = M + and M contains all the relevant objects. By
the elementarity of M , strong unboundedness of e and the fact that T contains
no + -chain, there exist and in \ M such that
r <lex r , = e{, } and (r , r ) < . (10.2.6)
Let = (, ). Note that there exist many > (r , r ) such that (r  ) 0
belongs to M and r () = 1. This is clear in the case when T is an Aronszajn tree,
by the elementarity of M and the fact that cf() = . However, this is also true in
case T has height < + and has size , since from our assumption log + = the
ordinal must have conality exactly . So by the property 9.1.1 of there is one
such such that = (, ) . Let t = (r  ) 0. Then t T M and so t is
a subset of of size + which belongs to M . Using the strong unboundedness of
e and the elementarity of M , there must be an in t M such that
e(, ) > sup(P () ). (10.2.7)
Since e(, ) belongs to e  + , a set which, by the assumption of Case 2.1, is
bounded in + , we conclude that e(, ) < . It follows that f (, , ) is dened
according to Rule 2, and therefore
f (, , ) = min(P () \ e(, )) = . (10.2.8)
10.2. Changs conjecture as a 3-dimensional Ramsey-theoretic statement 297

Case 2.2: For every in some tail of the set e  + is unbounded in + . Going
to a tail, we assume that this is true for all . Let M and = M +
be chosen as before. If T is Aronszajn, using elementarity of M , there must be
M and < of conality such that there exist unboundedly many
below such that the set

= {e{, } : , r <lex r , (r , r ) = } (10.2.9)

is unbounded in + . If T is a tree of height < + , = will satisfy this for all


but many , so in particular, we can pick one such M . Choose
such that

r <lex r , = e(, ) and (r , r ) < . (10.2.10)

< such that


Let = (, ). By the property of in Lemma 9.1.1 of , there is
(, ) for all [ Choose [
, ). above (r , r ) such that the set
, )
is unbounded in + . Let = (, ). Then . The set P () being of
size < is bounded in . Since M , there is M above such that

(r , r ) = , r <lex r and e(, ) > sup(P () ). (10.2.11)

Note that r <lex r , so the denition of f {, , } obeys Rule 2. Applying this


rule, we get that
f {, , } = min(P () \ e(, )) = . (10.2.12)
This nishes the proof of Theorem 10.2.1. 
Theorem 10.2.2. If is a regular strong limit cardinal carrying a nonreecting
stationary set, then there is f : [+ ]3 which takes all the values from on
the cube of any subset of + of size .
Proof. This is really a corollary of the proof of Theorem 10.2.1, so let us only
indicate the adjustments. By Corollary 9.2.14 and Lemma 9.2.2, we can choose a
strongly unbounded subadditive map e : [+ ]2 . By the assumption about
we can choose a C-sequence C ( < ) avoiding a stationary set and
consider the corresponding notion of a walk, trace, 0 -function and the square-
bracket operation [] as dened in (8.2.1) above. As in the proof of Theorem
10.2.1, we choose a one-to-one sequence r ( < + ) of elements of {0, 1} and
consider the corresponding function : [+ ]2 :

(, ) = (r , r ) = min{ < : r () = r ()}. (10.2.13)

The denition of f : [+ ]3 is given according to the following rules, applied


to a given x [+ ]3 .
Rule 1: If x = { < < }, (r , r ) < (r , r ) and r <lex r <lex r , or
r >lex r >lex r , let

f {, , } = [(, )(, )].


298 Chapter 10. Higher Dimensions

Rule 2: If x is such that r is lexicographically between the other two r s for


x, if x \ {} is such that (r , r ) > (r , r ), where is the remaining
element of x, and they do not satisfy the conditions of Rule 1, set

f {, , } = min(Tr((, ), e{, }) \ e{, }),

i.e., f {, , } is the minimal point on the trace of the walk from e{, } to (, )
above the ordinal e{, }; if such a point does not exist, set f {, , } = 0.
The proof of the theorem is nished once we show that for every stationary
and every + of size , there exist x []3 such that f (x) . This
is done, as in the proof of Theorem 10.2.1, by considering the two cases. Case 1
is reduced to the property in Lemma 8.2.4 of the square-bracket operation. The
treatment of Case 2 is similar and a bit simpler than in the proof of Theorem
10.2.1, since the case that T is of height < is impossible due to our assumption
that is a strong limit cardinal. This nishes the proof. 
Since log 1 = , the following is an immediate consequence of Theorem
10.2.1.
Theorem 10.2.3. Changs conjecture is equivalent to the statement that for every
f : [2 ]3 1 there is uncountable 2 such that f  []3 = 1 . 
Remark 10.2.4. Since this same statement is stronger for functions from higher-
dimensional cubes [2 ]n into 1 Theorem 10.2.3 shows that they are all equivalent
to Changs conjecture. Note also that n = 3 is the minimal dimension for which this
equivalence holds, since the case n = 2 follows from the Continuum Hypothesis,
which has no relationship to Changs conjecture.

10.3 Three-dimensional oscillation mapping


In this section we examine the basic stepping-up procedure without the assumption
that some form of Changs conjecture is false. So let be a given regular uncount-
able cardinal and let C ( < + ) be a xed C-sequence such that tp(C ) for
all < + . Let : [+ ]2 be the -function dened above in (9.1.2). Recall
that, in case C ( < + ) is a  -sequence, the key to our stepping-up procedure
was the function : [+ ]3 + dened by the formula (10.1.3). Without us-
ing the assumption that C ( < + ) is a  -sequence, one can build a related
function which turns out to be a good substitute. This is the function,

: [+ ]3 (10.3.1)

dened by
(, , ) = |0 (, ) 0 (, )|. (10.3.2)
Thus, (, , ) is equal to the length of the common part of the walks
and .
10.3. Three-dimensional oscillation mapping 299

Denition 10.3.1. A subset of + is stable if is bounded on []3 .


Lemma 10.3.2. Suppose that is a stable subset of + of size . Then

sup{(, ) : {, } []2 } =

for every of size .


Proof. Suppose is bounded by on []2 for some + of size . We need to
show that is unbounded on []3 . Therefore we may assume that is of order-
type and let = sup(). Construct a sequence = 0 1 of subsets of
of order-type such that:

(, , ) n for all {, , } [n ]3 . (10.3.3)

Given n , choose an -chain M of length of elementary submodels M of some


large enough structure of the form H such that M and M contains all
the relevant objects. Now choose another elementary submodel N of H which
contains M as well as all the other relevant objects such that N . Let
N = sup(N ). Choose n above N . Let = 0 > > k = N be the
walk from to N . Let be the i with the smallest index such that Ci N is
unbounded in N . Then is either equal to N or k1 . Choose 0 N such
that:
0 > max(Ci N ) for all i < k for which Ci N is
(10.3.4)
bounded in N .
Then {i : i < k} \ Tr(, ) for all N n above 0 . So if we choose
< in N 0 such that C [, ) = , then

Tr(, ) Tr(, ) = {i : i k} \ . (10.3.5)

From this and our assumption (, , ) n we conclude that this set has size at
least n. For M M, let M = sup(M ) and C = {M : M M}. Then C is
a closed and unbounded subset of of order-type . By our assumption that is
bounded by on []2 , we get:

tp(C N ) sup{tp(C ) : [0 , N )}
sup{(, ) : [0 , N )} (10.3.6)
.

Since C[0 , N ) has order-type > there must be M MN such that M / C


and M > 0 . Thus, we have an elementary submodel M of H which contains
all the relevant objects and a n above M such that the walk M
contains all the points from the set {i : i < k} \ , but includes at least the point
min(C \ M ) which is strictly above M . Hence, if = 0 > 1 > > l = M
is the trace of M and if is the i (i l) with the smallest index, subject
300 Chapter 10. Higher Dimensions

to the requirement that C M is unbounded in M , then {i : i < l} \ has


0 M be such that:
size at least n + 1. Let
0 > max(C M ) for all i < l for which C M is
i i (10.3.7)
bounded in M .

Then {i : i < l}\ Tr(, ) for all M n above 0 . Using the elementarity
of M we conclude that for every < there exist n \ and a set a [, )
of size at least n + 1 such that a Tr(, ) for all n [ 0 , ). Hence we can

choose a closed and unbounded set D of [0 , ) of order-type , for each D an
() n \ and an a() [, ()) such that for every D:

0 , ),
a() Tr(, ()) for [ (10.3.8)

() < min(D \ ( + 1)). (10.3.9)


Finally let n+1 = {() : D}. Then

a() Tr((), ()) Tr((), ()) for all < < in . (10.3.10)

This gives us that (, , ) n + 1 for every triple < < in n+1 , nishing
the inductive step and therefore the proof of Lemma 10.3.2. 

Denition 10.3.3. The 3-dimensional version of the oscillation mapping,

osc : [+ ]3 , (10.3.11)

is dened on the basis of the 2-dimensional version of Section 8.1 as follows:

osc(, , ) = osc(Cs \ , Ct \ ), (10.3.12)

where s = 0 (, )  (, , ) and t = 0 (, )  (, , ).
In other words, we let n be the length of the common part of the two walks
and . Then we consider the walks

= 0 () > > k () = and = 0 () > > l () = (10.3.13)

from to and to , respectively. If both k and l are bigger than n (and


note that k n), or in other words, if n () and n () are both dened, we let
osc(, , ) be equal to the oscillation of the two sets Cn () \ and Cn () \ . If
not, we let osc(, , ) = 0.
Lemma 10.3.4. Suppose that is a subset of + of size , a regular uncountable
cardinal, and that every subset of of size is unstable. Then for every positive
integer n, there exist < < in such that osc(, , ) = n.
10.3. Three-dimensional oscillation mapping 301

Figure 10.2: osc(, , ) = 3.

Proof. We may assume that is actually of order-type and let = sup().


Let M be an -chain of length of elementary submodels M of H++ such that
M and M contains all the relevant objects. For M M, let M =
sup( M ). Now choose another elementary submodel N of H++ containing M
as well as all the other relevant objects with N . Let N = sup( N ) and
choose above N and let = 0 > > k = N be the walk from to
N . Let and 0 be chosen as above in (10.3.4). Then = k or = k1 and

Tr(, ) Tr(, ) = {i : i k} \ for all < in


(10.3.14)
[0 , N ) such that C [, ) = .

So if a tail of the set {M : M M N } is included in C , then using the


elementarity of N , we would be able to construct a conal subset 0 such
that is constantly equal to k on [0 ]3 , where k is the place of in the sequence
0 , . . . , k . This, of course, would contradict our assumption that contains no
large stable subsets. It follows that the set of all N for M M N which do not
belong to C is conal in N . So we can pick M1 M2 Mn+1 in M N
302 Chapter 10. Higher Dimensions

such that, if i = Mi (1 i n + 1), then:

i [0 , ) \ C for all 1 i n + 1, (10.3.15)

[i , i+1 ) C = for all 1 i n. (10.3.16)


Fix M1 above max(C 1 ). For 1 i n, pick i [i , i+1 )Mi+1 such

that i max(C i+1 ). For 1 i n, let Ii = [i , i ], and let In+1 = [n , ].

Now set F to be the collection of all increasing sequences I1 < I2 < < In+1
of closed intervals of ordinals < with the property that there is = ( I ) in
In+1 such that, if = 0 > > l = is the walk from to , then:

l > k and k = min(Tr(, ) In+1 ), (10.3.17)


n+1
Ck \ Ii , (10.3.18)
i=1

Ck Ii = for all 1 n n + 1. (10.3.19)


Clearly, F M1 and I1 , . . . , In+1  F, as is a witness of this. So working
as in the proof of Lemma 8.1.2, we can nd F0 F in M1 such that, if some
sequence J of length n (including the case J = ) end-extends to a member of
F0 , then for every < there is a closed interval K included in [, ) such that
the concatenation J K also end-extends to a sequence in F0 . Working inductively
on 1 i n + 1 we can select a sequence I1 , I2 , . . . , In such that:

I1 , . . . , Ii Mi+1 for 1 i n, (10.3.20)


i , i+1 ) for 1 i n.
Ii [ (10.3.21)
I1 , . . . , Ii  end-extends to a member of F0 . (10.3.22)
Clearly, there is no problem in choosing these objects, using the elementarity of
the models Mi (1 i n + 1) and the splitting property of F0 . Working in
Mn+1 , we nd an interval In+1 such that I0 , . . . , In , In+1  belongs to F0 . Let
1 ) it separates from
= (I0 , . . . , In+1 ). Since C intersects the interval [1 ,

, so by (10.3.14) we get that (, , ) = k, hence by denition

osc(, , ) = osc(Ck \ , Ck \ ). (10.3.23)

Recall that k = , so referring to (10.3.16), (10.3.18) and (10.3.19) we conclude


that I1 Ck , . . . , In1 Ck and (In In+1 ) Ck are the n convex pieces into
which the set C \ splits the set Ck \ . Hence osc(Ck \ , Ck \ ) = n. This
completes the proof of Lemma 10.3.4. 
Applying the last two lemmas to the subsets of + of size , we get an
interesting dichotomy:
10.3. Three-dimensional oscillation mapping 303

Lemma 10.3.5. Every + of size can be rened to a subset of size such


that either:
(1) is unbounded and therefore strongly unbounded on , or
(2) the oscillation mapping takes all its possible values on []3 . 
The proof of the following result contains a typical application of this di-
chotomy where a projection of the square-bracket operation of + is lifted up to
a mapping of the form f : [++ ]3 .
Theorem 10.3.6. Suppose is a regular cardinal such that log + = . Then a
natural projection of the square-bracket operation on + can be lifted to a mapping
f : [++ ]3 which in particular takes all the values from on the cube of any
subset of ++ of size + .
Proof. We choose two C-sequences C ( < + ) and C+ ( < ++ ) on + and
++ respectively, such that tp(C ) for all < + and tp(C+ ) + for all
< ++ . Let
: [+ ]2 and + : [++ ]2 +
be the corresponding -functions dened above in (9.1.2). Choose also a one-to-one
sequence
+
{r : < ++ } {0, 1}
and consider the corresponding function : [++ ]2 + dened by ( see (10.2.1
above),
(, ) = (r , r ) = min{ : r () = r ()}. (10.3.24)
The mapping
f : [++ ]3 +
that will satisfy the conclusion of the theorem is dened according to the following
three cases where a given triple < < of elements of ++ can fall.
Case 1: (Cs Ct ) \ = , where
s = 0 (, )  (, , ) and t = 0 (, )  (, , ),
assuming of course that 0 (, ) has length at least (, , ).
Rule 1: If (r , r ) < (r , r ) and r <lex r <lex r or r >lex r >lex r , set
f (, , ) = min(P ((, )) \ (, )),
where = (min{ (, ) : (, (, )) = (, (, ))}, (, )).
Rule 2: If {, , } is such that r is lexicographically between the other two
r s for {, , }, if {, , } \ { } is such that (r , r ) > (r , r ),
where is the remaining member of {, , }, and if {, , } does not fall under
Rule 1, let
f (, , ) = min(P (+ {, }) \ + (, )),
where = {(, ), + (, )}.
304 Chapter 10. Higher Dimensions

Case 2: (Cs Ct ) \ = , where


s = 0 (, )  (, , ) and t = 0 (, )  (, , ),
assuming again that 0 (, ) has length at least (, , ). Let
f (, , ) = osc(, , ).
Case 3: If a given triple < < does not fall into one of the Cases 1 and 2, let
f (, , ) = 0.
The proof of Theorem 10.3.6 is nished if we show that for every ++ of
size + , the image f  []3 either contains all positive integers or almost all ordinals
< + of conality .
So let be a given subset of ++ of order-type + and let = sup(). By
the proof of Lemma 10.3.2, if is stable, then it must contain a (stable) subset
such that for some n < and all < < in :
(, , ) = n, (10.3.25)
(Cs Ct ) \ = where s = 0 (, )  n and t = 0 (, )  n. (10.3.26)
In other words, Case 1 of the denition of f (, , ) applies to any triple < <
from . Note that in this case, the denition follows exactly the denition of the
analogous function in the proof of Theorem 10.2.1 with + in the role of the
strongly unbounded subadditive function e : [++ ]2 + . By Lemma 10.3.2,
+ is strongly unbounded on subsets of , so by the proof of Theorem 10.2.1 we
conclude that in this case the image f  []3 contains almost all ordinals < + of
conality .
Suppose now that no subset of of size + is stable. By Lemma 10.3.4 (in
fact, its proof) for every integer n 1 there exist < < in such that for
s = 0 (, )  (, , ) and t = 0 (, )  (, , ), where 0 (, ) has length
> (, , ):
osc(, , ) = n, (10.3.27)
(Cs Ct ) \ = . (10.3.28)
This nishes the proof. 
Corollary 10.3.7. There is f : [2 ]3 which takes all the values on the cube
of any uncountable subset of 2 . 
Remark 10.3.8. Note that the dimension 3 in this corollary cannot be lowered to 2
as long as one does not use some additional axioms to construct such f . Note also
that the range cannot be replaced by a set of bigger size, as this would contradict
Changs conjecture. We have seen above that Changs conjecture is equivalent
to the statement that for every f : [2 ]3 1 there is an uncountable set
2 such that f  []3 = 1 . Is there a similar reformulation of the Continuum
Hypothesis? More precisely, one can ask the following question.
Question 10.3.9. Is CH equivalent to the statement that for every f : [2 ]2
there exists an uncountable 2 with f  []2 = ?
10.4. Two-cardinal walks 305

10.4 Two-cardinal walks


The basic notion of minimal walk also makes sense in the context of structures
P () = {x : |x| < } and the purpose of this section is to expose some of this
theory. The two-cardinal walk is based on a two-cardinal version of a C-sequence,
a sequence of the form
Cx (x P ()),
where for each x in P (), the sequence Cx ( x {sup(x)})2 is an ordinary
C-sequence modulo taking the transitive collapse of x. In other words, if x is the
order-type of x and if x : x {sup(x)} x + 1 is the order-isomorphism, then

C = x Cx1 () ( x )

is an ordinary C-sequence on x + 1. Given x P () and x, one denes the


minimal walk from x to to be the sequence 0 = sup(x) > 1 > > n = ,
where
i+1 = min(Cxi \ )
for i < n. The integer n is the length of the walk and is denoted by 2 (, x), while
the upper trace of the walk from x to x is the nite set

Tr(, x) = {0 , 1 , . . . , n }.

The sequence
0 (, x) = |Cxi | : i < n
is the the full code of the walk from x to . Note that for every i n, the walk
from x to i is the segment 0 = sup(x) > 1 > > i of the walk from x to .
In other words, Tr(i , x) = {0 , . . . , i } holds for every i n. So if, in general, for
y P (), we put

(, y) = max( Cy ),
Tr(,y)\{}

we get the following expression for the lower trace of the walk from x to ,

(, x) = {(i , x) : i < n}.

The theory of minimal walks in the context of P () comparable to the theory of


minimal walks on ordinals is yet to be developed, but we do have some important
analogues such as, for example, the square-bracket operation.
Denition 10.4.1. Consider x, y P (). If x {sup(x)}  y put [xy] = 0. If
x {sup(x)} y, let

0 = sup(y) > 1 > > n = sup(x)


2 Here, sup(x) = min{ : < for all x}. Hence, Cx = Cx for = sup(x).
306 Chapter 10. Higher Dimensions

be the minimal walk from y to sup(x). Let m n be maximal with the property
that there is x such that 0 (, x) = 0 (m , y) and let 0 = sup(x) > 1 >
> m = be the minimal walk from x to . Finally, let [xy] = j where j < m
is determined by the following two properties:
(1) max(Cxi ) max(Cyi sup(x)) for all i < j.
(2) max(Cxj ) > max(Cyj sup(x)).
If such a j < m cannot be found, set [xy] = 0.
Recall that analysis of the square-bracket operation in the realm of walks on
ordinals requires that the C-sequence avoids certain stationary sets. In particular,
ordinary square-bracket operations work well on successors of regular cardinals. In
the two-cardinal context this corresponds to considering P+ (), which we denote
more simply as [] , assuming that is a regular cardinal, and choosing the
C-sequence Cx (x [] ) in the following way. One rst chooses a C-sequence
C ( < + ) which avoids the set { < + : cf() = }, or in other words,
C+1 = {} and for limit the set C is a closed and unbounded in , has order
type cf() and it contains no ordinals of conality . Given x [] one puts

Cx = x1 C ,

where = tp(x) and where x : x is the unique order isomorphism.


Lemma 10.4.2. If > are regular cardinals, then for every conal subset U of
[] , the set of all ordinals < of conality which are not of the form [xy] for
some x y in U is not stationary in .
Proof. Let S be a given stationary subset of { < : cf() = }. We need to
nd x y in U such that [xy] S. Choose large enough regular cardinal and
continuous -chain M of elementary submodels M of (H , ) of size < such that
M and such that M contains S and the C-sequence Cx (x [] ). Choose
now an elementary submodel N of (H , ) of size such that N + + , =
sup(N ) S, and such that N contains S, the C-sequence Cx (x [] ), and the
chain M. Since U is conal in [] , we can nd y U such that (N ) {} y.
Let
0 = sup(y) > 1 > > j =
be the minimal walk from y to . Note that by our choice of the C-sequence Cx
(x [] ), the set Cyi is bounded in for all i < j. Since = sup(N ),
for each i < j, the minimum

i = min((N ) \ max(Cyi ))

exists and is smaller than . Let

I = {i < j : i = max(Cyi )}.


10.4. Two-cardinal walks 307

For M M, let M = M . Let

D = {M : M M}.

Then D is a closed and unbounded subset of . Let D be the set of all limit points
of D of conality at most . Then D N and the intersection D M has an
indecomposable order-type bigger than . So in particular we can nd two models
M0 , M1 M N such that M0 and M1 belong to the set D but M0 / Cy ,
M1 / Cy , and moreover,

{i : i < j} M0 M1 , and Cy [M0 , M1 ) = . (10.4.1)

Let N0 = N M0 and N1 = N M1 . Then N0 and N1 are also elementary


submodels of (H , ) and they have the property that

M0 = sup(N0 ) and M1 = sup(N1 ).

Since Cy is a closed subset of y modulo taking the transitive collapse, it follows


that max(Cy M0 ) and max(Cy M1 ) exist and that they are strictly smaller
than M0 and M1 , respectively. By the conality requirement on M0 and M1 , the
sets N0 and N1 are conal in M0 and M1 , respectively. So we can dene 0
to be the minimal point of N0 that is greater than or equal to max(Cy M0 ).
Similarly, we let 1 be the minimal point of N1 that is greater than or equal
to max(Cy M1 ). Let

0 = tp(Cy M0 ) and 1 = tp(Cy M1 ).

Let t = 0 (, y). Note that t N0 . For < , let U be the set of all x U for
which there is a minimal walk 0 = sup(x) > 1 > > j inside x with the
following properties:
(i) 0 (, x) = t,
(ii) max(Cxi ) = i for i I,
(iii) max(Cxi ) < i for i < j, i
/ I,
(iv) tp(Cx ) = and tp(Cx 0 ) = 0 ,
(v) Cx [, ) = .
Note that for every N0 , the set U belongs to N0 . Note also that by our
choices of the objects appearing in the denition, the set y belongs to U for all
N0 . Since y N0 , by the elementarity of N0 , we conclude that U
must be conal in [] for all < . Applying this in the model M1 to the set
U for = 1 + 1, we nd x U1 +1 M1 such that {i : i < j} {0 , 1 } x.
Let 0 = sup(x) > 1 > > j be the minimal walk with properties (i), (ii),
(iii),(iv), and (v) above witnessing the membership of x in U1 +1 . Let Cxj
be such that tp(Cxj ) = 1 . Note that 1 is a successor ordinal so Cxj is
308 Chapter 10. Higher Dimensions

bounded in x, so the maximum max(Cxj ) exists, and since sup(x) < M1 ,


we have that
max(Cxj ) > 1 max(Cyj sup(x)).
By (ii) and the denition of the set I, we conclude that

max(Cxi ) = i = max(Cyi sup(x)) for all i I.

By (iii), for i < j with i


/ I, we must have

max(Cxi ) < max(Cyi sup(x)),

since otherwise, max(Cxi ) x M1 N would be a point of N that is


strictly smaller than i and which still dominates max(Cyi ) contradicting the
choice of i . Combining all this we see that the pair x {sup(x)} y satises the
requirements of Denition 10.4.1 which permits us to conclude that [xy] = S.
This nishes the proof. 
In order to treat the case of singular cardinals , we need the following
relativization of the square-bracket operation.
Denition 10.4.3. Let . Consider an arbitrary pair x, y P (). If
x {sup(x )}  y, put [xy] = 0. If x {sup(x )} y, let

[xy] = [(x )(y )]

as dened in P () according to the Denition 10.4.1. In other words, let

0 = sup(y ) > 1 > > n = sup(x )

be the minimal walk from y to sup(x ). Let m n be maximal with the


property that there is x such that 0 (, x ) = 0 (m , y ) and let
0 = sup(x ) > 1 > > m = be the minimal walk from x to .
Finally, let [xy] = j where j < m is determined by the following two properties:
(1) max(Cxi ) max(Cyi sup(x)) for all i < j.
(2) max(Cxj ) > max(Cyj sup(x)).
If such a j < m cannot be found, set [xy] = 0.
Then the proof of Lemma 10.4.2 adjusts easily to the proof of the following
slightly more general fact.
Lemma 10.4.4. Suppose is a regular uncountable cardinal and that is an
arbitrary cardinal. Let 0 and 1 be two distinct regular uncountable cardinals such
that 0 , 1 . Let S0 and S1 be stationary subsets of 0 and 1 , respectively,
consisting only of conality ordinals. Then for every conal subset U of []
there exist x y in U such that [xy]0 S0 and [xy]1 S1 . 
10.4. Two-cardinal walks 309

Taking an appropriate projection of the square-bracket operation one gets


the following result.
Theorem 10.4.5. For every pair of innite cardinals < with regular, there is
a mapping c : [[] ]2 such that for every conal subset U of [] and every
there exist x y in U such that c(x, y) = .
Proof. If is a regular cardinal, then according to Lemma 10.4.2, composing the
square-bracket operation of [] with a partition of { < : cf() = } into
disjoint stationary sets, gives us the conclusion. So let us assume that is a
singular cardinal. Let = max(cf(), + ) and choose a strictly increasing sequence
( < cf()) of regular cardinals bigger than whose supremum is equal to . Let
S ( < cf()) be a partition of { < : cf() = } into cf() disjoint stationary
sets. For each < cf(), let T ( < ) be a partition of { < : cf() = } into
disjoint stationary sets. Finally, for x, y [] , set

c(x, y) = ,  if [xy] S and [xy] T ,



and if these conditions are not satised for any pair from the set P = <cf() {}
, let c(x, y) be an arbitrary member of P . Then by Lemma 10.4.4, for every
conal U [] and every ,  P there exist x y in P such that [xy] S
and [xy] T . Since P has cardinality , this nishes the proof. 
Remark 10.4.6. The rst attempt to generalize the square-bracket operation into
the two-cardinal context was made by Velleman [129] who proved the conclusion
of Theorem 10.4.5 for = and for cardinals for which [] admits a stationary
subset of cardinality . The full conclusion was obtained by the author in [113].
In certain cases when the cardinal is singular, one can increase the number of
colors in Theorem 10.4.5. The following result of Shioya [99] gives one condition
when this is possible.
Theorem 10.4.7. Suppose is a regular cardinal and > is such that cf() = .
Then there is a mapping c : [[] ]2 + such that for every conal subset U of
[] and every + , there exist x y in U such that c(x, y) = . 
Further advances in this area are likely to involve walks that step from one
element y of P () to one of its subsets x that is not necessarily its initial segment.
This in turn is likely to involve two-cardinal analogues of square-sequences that
would facilitate these kind of walks. The case = 1 is of course the rst to be
understood. So let us present something in that direction, a result which states that
the quantier reduction analogous to the one proved before in the case = = 1
is possible for many other cardinals .
Theorem 10.4.8. Suppose S is a subset of [] that is equinumerous with a locally
countable subset of [] . Then there is a square-bracket operation

[]S : [[] ]2 []
310 Chapter 10. Higher Dimensions

such that for every conal subset U of [] the set of all s S that are not of the
form [xy]S for some x y in U is not stationary in [] . 
Recall that a locally countable subset of [] is an arbitrary subset K of []
with the property that {x K : x y} is countable for every y [] . Clearly
every family of pairwise-disjoint sets is locally countable, so there is always a
locally countable subset of [] of cardinality . The interest of Theorem 10.4.8
lies in the fact that in many cases there exist locally countable subsets of []
that are equinumerous with some stationary subsets of [] . In order to dene the
square-bracket operation of Theorem 10.4.8, we need the following simple fact.
Lemma 10.4.9. There is r : [] {0, 1} such that rx = ry for x y.
Proof. Choose a one-to-one sequence p ( < 1 ) of elements of {0, 1} and choose
for each ordinal < of conality a strictly increasing sequence (i ) of smaller
ordinals conal in . It will be convenient for us to extend the domain of r to
include all nite subsets of as well and to have range of r included in {0, 1}
{0, 1} rather than in {0, 1} . Let rx = ptp(x) , qx , where qx is dened recursively
on sup(x) as follows. If x has a maximal element , set qx (0)