Annalisa Marcja, Carlo Toffalori A Guide To Classical and Modern Model Theory Trends in Logic 2003

A GUIDE TO CLASSICAL AND MODERN MODEL THEORY
TRENDS IN LOGIC
Studia Logica Library
VOLUME 19
Managing Editor
Ryszard W6jcicki, Institute of Philosophy and Sociology,
Polish Academy of Sciences, Warsaw, Poland
Editors
Daniele Mundici, Department of Mathematics "Ulisse Dini ",
University of Florence, Italy
Ewa Orlowska, National Institute of Telecommunications,
Warsaw, Poland
Graham Priest, Department of Philosophy, Unive rsity of Queensland,
Brisbane, Australia
Krister Segerberg, Department of Philosophy, Uppsala University,
Sweden
Alasdair Urquhart, Department of Philosophy, University of Toronto, Canada
Heinrich Wansing, Institute of Philosophy, Dresden Uni versity of Technology,
Germany
SCOPE OF THE SERIES
Trends in Logi c is a bookseri es covering essentially the same area as the j ourn al
Studia Logica - that is, contemporary formal logic and its applications and
relations to other di sciplines. These include artifi ci al intelli gence, informatics,
cognitive science, philosophy of science, and the philosophy of language.
However, thi s list is not exhaustive, moreover, the range of applications, com-
parisons and sources of inspirati on is open and evolves over time.
Volume Editor
Ryszard Woj:kki
The titles published in this series are listed at the end of this volume.
A GUIDE TO CLASSICAL
AND
MODERN MODEL THEORY
by
ANNALISA MARCJA
University of Florence, Italy
and
CARLO TOFFALORI
University of Camerino, Italy
KLUWER ACADEMIC PUBLISHERS
DORDRECHTI BOSTON I LONDON
A C. LP. Catalogue record for this book is available from the Library of Congress.
ISBN 1-4020-1330-2 (HB)
ISBN 1-4020-1331-0 (PB)
Published by Kluwer Academic Publishers,
P.O. Box 17, 3300 AA Dordrecht , The Netherlands.
Sold and distributed in North, Central and South America
by Kluwer Academic Publishers,
101 Philip Drive, Norwell, MA 0206 1, U.S.A.
In all other countries, sold and distributed
by Kluwer Acade mic Publishers,
P.O. Box 322, 3300 AH Dordrecht , The Netherlands .
Printed on acid-free paper
All Rights Reserved
2003 Kluwer Academic Publishers
No part of this work may be reproduced, stored in a retrieval system, or transmitted
in any form or by any means, electronic, mechanical, photocopying, microfilming, recording
or otherwise, without written permission from the Publisher, with the exception
of any material supplied specifically for the purpose of being entered
and executed on a computer system, for excl usive use by the purchaser of the work.
Printed in the Netherlands.
Preface
This book deals with Mod el Theory. So the first question t hat a possible,
recalcitrant reader might ask is just: What is Model Theory? Which are it s
intents and applications? Why should one try to learn it? Another, mor e
parti cul ar question might be th e following one. Let us ass ume, if you like,
t hat Model Theor y deser ves some attent ion. Why should one use this book
as a guide t o it ?
The answer t o t he former questi on may sound problematic, but it is quite
simple, at least in our opinion. For , Model Theor y has been developing,
since it s bir th , a number of methods and concepts t hat do have t heir int rin-
sic relevance, but also provide fruitful and not abl e applications in various
fields of Mathematics. We could mention her e it s role in Algebra and Alge-
braic Geometry, for instance the analysis of differentially closed fields (and
th e resul t s on t he differential closure of a differential field) , or p-adic fields
(and th e asy mpt otic solut ion of Art in's Conjecture), as well as the recent
Hrushovski 's mod el t heoret ic approac h t o classical problems, like Mordell-
Lang's Conjecture or Manin-Mumford's Conjecture.
So Model Theory is today a lively, spright ly and fertil e research area, which
surely deser ves t he attent ion of t he mathemati cal world and, consequent ly,
it s own references. This recalls t he latter question above. Act ually there do
exist some excellent t extbooks explaining Model Th eory, such as [56] and
[57]. Also Poizat 's book [131] should be mentioned; it was writt en more th an
ten yea rs ago, but it is st ill up-to-dat e, and it has been recentl y t ranslated
in English. In addit ion more specialist ic references t reat adequat ely some
parti cul ar fields in Model Theory, such as stability t heory, simplicity t heory,
o-minimalit y, classification theory and so on.
Nevert heless, we believe t hat t his book has it s own role and its own origi-
nality in t his setting. Indeed we wish t o address t his work not onl y to t he
experts of th e area, but also, and mainly, to young people having a basic
knowledge of mod el theory and wishing to proceed towards a deeper anal-
ysis , as well as to mathematicians which are not directly involved in Mod el
v
VI PREFACE
Theor y but work in related and overlapping fields, such as Algebra and Ge-
ometry. Accordingly we will emphasize t he frequ ent and fruitful connections
between Model Theor y and t hese branches of Math ematics (different ially
closed fields, Artin's Conj ecture, Mordell-Lang's Conjecture and so on). In
each case, we aim at giving a detailed report or , at least , at sket ching th e
main ideas and techniques of the model t heoretic approach.
Our book wishes also to follow a historical perspective in introducing Model
Theory. Of course, t his does not mean to provide a full history of Model
Theory (although such a project could be interesting and wor thy of some
attention), but just to inser t any basic concept in th e historical framework
where it was born, and so to better clarify the reasons why it was introduced.
Hence, after shortly recalling in Chapter 1 basic Model Theory (structures
and theories, compactness and definability), we deal in Chapter 2 with
quantifier elimination, in particular with the work of Alfred Tarski on al-
gebraically closed fields and real closed fields. We will discuss the role of
quantifier eliminat ion in Mod el Theory, but we will t reat briefly also it s in-
t riguing role in the P = N P problem within the new models of computation
(such as t he Blum-Shub-Smale approach, and so on).
Cha pter 3 will be concerned with Abraham Robinson 's ideas: mod el com-
plet eness, model companions, existent ially closed struct ures . We will con-
sider again algebraically closed fields and real closed fields, but we will il-
lustrate also other crucial classes, like differentially closed fields, separabl y
closed fi elds, p-adically closed fields and, finall y, existent ially closed differ-
ence fi elds (a rather recent mat ter , with some remarkabl e applications t o
Algebraic Geometry).
Cha pt er 4 deal s with imaginary elements. They are essent ially classes of
definable equivalence relations in a structure A, so element s in some quoti ent
structure. We describe Shelah 's const ruct ion of Ae
q
, englobing these classes
as new element s in th e whole st ruct ure, and we show that th ese imaginary
elements can be somet imes eliminat ed, because the corresponding quotients
can be simulated by some suit able definable subsets of A.
Chapters 5 and 6 are devoted to Morley's Theorem on uncountable cate-
goricity. Actually it s proof will be given only in Chapter 7, but here we
describe Morley's ideas -algebraic closure, totally transcendental theories,
prime models, an so on- and we illustrate t heir richness and th eir applica-
t ions.
We will be led in this way t o one of t he main topics in Mod el Theory, namely
t he Classificat ion Problem. We will explain in Chapter 7 t he more relevant
ideas in t he formidabl e work of Shelah on t his mat t er (simplicity, stability,
superstability, modularity) , and we will discuss t heir significance in some
PREFACE
vu
important algebraic classes, like differential fields, difference fields, and so
on. We wish also to deal with the Zilber program of classifying structures
up to biinterpretability, in particular with Zilber's Conjecture on strongly
minimal sets, and its brilliant solution due to Hrushovski.
Also Chapter 8 largely owes to Hrushovski. In fact, after illustrating in
more detail the natural connection between Model Theory and Algebraic
Geometry, we will describe the Hrushovski proof of Mordell-Lang conjecture;
we will refer very quickly also to the Hrushovski solution of the related
Manin-Mumford conjecture. In particular we will realize how deeply Model
Theory, actually both pure Model Theory and Model Theory applied to
algebra are involved in these proofs.
The final Chapter is devoted to a (comparatively) recent and fertile area
in Model Theory: o-minimality. We will expound the basic results on 0-
minimal theories, and we will discuss some intriguing developments, includ-
ing Wilkie's solution of a classical problem of Tarski on the exponentiation
in the real field.
We assume some familiarity with the basic notions of Algebra, Set Theory
and Recursion Theory. [65], [66] or [78], and [121] respectively are good
references for these areas. Incidentally, let us point out that we are working
within the usual Zermelo - Fraenkel axiomatic system, including the Axiom
of Choice. We also assume some acquaintance with basic Model Theory,
such as it is usually proposed in any introductory course. However, Chapter
1 is devoted, as already said, to a short and somewhat informal sketch of
these matters.
As its title states, this book aims at being only a guide. We do not claim
to provide an exhaustive treatment of Model Theory; indeed our omissions
are likely to be much more numerous and larger than the topics we deal
with. But we have aimed at giving an almost complete report of at least
two crucial subjects (w-stability and o-minimality), and at providing the
basic hints towards some conspicuous generalizations (such as superstability,
stability, and so on).
In a similar way, we have treated in detail some key algebraic examples
(algebraically closed fields, real closed fields, differentially closed fields in
characteristic 0), but we have provided at least some basic information on
other relevant structures (like p-adic fields, existentially closed fields with
an automorphism, differentially and separably closed fields in a prime char-
acteristic). In conclusion, we do hope that the outcome of our work is a
sufficiently clear and terse picture of what Model Theory is, and provides
a report as homogeneous and general as possible. Incidentally, let 11S say
viii PREFACE
t hat t his book is not a lit er al t ranslation of t he form er it ali an version [108];
all t he material was revised and rewrit ten; our t reat ment of some t opics,
like quantifier elimination and model completeness , are entirely new; and
we have added some relevant matters , such as prime models and Morley's
T heorem on uncount able categorical t heories.
Contents
1 Structures
1.1 St ructures
1.2 Sentences
1.3 Embeddings
1.4 The Compactness Theorem
1.5 Elementary classes and t heories
1.6 Complete theori es
1.7 Definable sets
1.8 References.....
2 Quantifier Elimination
2.1 Elimination sets . . .
2.2 Discrete linear orders .
2.3 Dense linear orders . .
2.4 Algebraically closed fields (and Tarski)
2.5 Tarski again: Real closed fields . . . .
2.6 pp-elimination of quantifiers and modules
2.7 Strongly mini mal th eories .
2.8 o-minimal theories .
2.9 Computational aspects of q. e.
2.10 References .
3 Model Completeness
3.1 An introduction .
3.2 Abraha m Robinson 's t est .
3.3 Model complet eness and Algebra
3.4 p-adic fields and Artin's Conj ecture .
3.5 Existentially closed st ruct ures .
3.6 DCF
a
.
IX
1
1
5
9
18
20
30
35
42
43
43
47
52
54
61
68
76
78
79
82
85
85
88
91
96
103
109
x
3.7 SCF
p
and DCF
p
3.8 ACFA . . .
3.9 References.. ..
4 Elimination of imaginaries
4.1 Interpretabilit y .
4.2 Imaginary elements . . . .
4.3 Algebraically closed fields
4.4 Real closed fields . . . . .
4.5 The elimin ation of imaginaries sometimes fails .
4.6 References... . . . . . . . . . . . . . . . .. .
5 Morley rank
5. 1 A tale of two cha pt ers
5.2 Definable sets ..
5.3 Types .
5.4 Saturated models . . .
5.5 A parenthesis: pure injecti ve modules
5.6 Omi t ting ty pes . . . . . .
5.7 The Morley rank , at last .
5.8 Strongly minimal sets .
5.9 Algebraic closure and definable closure
5.10 References .
6 w -stabili t y
6.1 Tot ally transcendent al t heories
6.2 w-stable groups
6.3 w-stable fi elds .
6.4 Prime models .
6.5 DCF
o
revisited
6.6 Ryll-Nardzewski 's Theorem, and other things
6.7 References .
7 Classify ing
7.1 Shelah 's Classification Theor y.
7.2 Simple t heories .
7.3 St able theori es .
7.4 Superstabl e t heories
7.5 w-stable th eories ..
7.6 Classifiabl e th eories .
CONTENTS
112
115
119
121
121
123
126
129
131
132
133
133
133
136
143
150
156
158
168
172
180
181
181
184
192
196
209
217
220
221
221
227
235
239
242
261
CONTENTS
7.7 Shelah's Uniqueness Th eor em .
7.8 MorIey's Theorem .
7.9 Biint erpretability and Zilber Conjecture
7.10 Two algebraic examples
7.11 References .
8 Model Theory and Algebraic Geometry
8.1 Int roduction .
8.2 Algebraic varieties, ideals, types . .
8.3 Di mension and MorIey rank . . . .
8.4 Mor phisms and definable functions
8.5 Manifolds .
8.6 Algebraic gro ups .
8.7 The MordelI-Lang Conjecture
8.8 References .
9 O-minimality
9.1 Int roduct ion .
9.2 The Monotonicit y Theor em
9.3 Cells .
9.4 Cell decomposition and other theorems .
9.5 Their proofs .
9.6 Definable groups in o-rninimal struct ures.
9.7 O-minimalit y and Real Analysis .
9.8 Variants on th e o-minimal t heme
9.9 No rose wit hout thorns.
9.10 References .
Bibliography
Index
xi
270
273
279
286
289
291
291
292
294
297
299
301
304
310
313
313
318
320
324
329
339
341
346
347
348
351
363
Chapter 1
Structures
1.1 Structures
The aim of t his chapter is to sket ch out basic model t heory. We wish to
summarize some key facts for people already acquainted with them, but
also, at the same t ime, t o introduce t hem to people unfamili ar to logic, and
perh aps disliking too many logical details. Accordingly we will use a rather
colloquial tone. The fundament al question t o be answered is: what is Model
Theor y? As we will see in mor e detail in Section 1.2, Model Theory is -or,
mor e precisely, was at it s beginning- the study of t he relationship between
math ematical formul as and st ruct ures sat isfying or rejecting th em. But , in
ord er t o fully appreciate t his matter , it is advisable for us pr eliminarily to
recall what a st ruct ure is, and which kind of formul as we ar e dealing with.
This section is devoted to th e former concept .
Structures are an algebraic notion. Actually, since Galois, Algebra is not
only th e solving of equations, or literal calcu lus, but becomes the science of
st ruct ures (gro ups, rings, fields, and so on). This new direction gets clearer
at the beginning of t he last cent ury, with Steinitz's work on fields and, lat er ,
t he publication of th e Van der Waerden book . What is a st ruct ur e? Basi-
cally, it is a non-empty set A, with a collect ion of distinguished elements,
oper ations, and relations. For instance, t he set Z of int egers with th e usual
operations of addition + and multiplication . is a st ruct ure, as well as the
same set Z with th e ord er relation Note that , in th ese examples, th e un-
derl ying set is t he same (th e int egers) , but , of course, t he st ructure changes:
in t he former case we have t he ring of integers, in t he lat t er t he int egers as
an ordered set . To make t his kind of difference among struct ures clearer, we
have t o choose a language, in oth er words t o specify how many distinguished
1
2 CHAPTER 1. STRUCTURES
elements, how many n-ary operations and relations (for every natural n i= 0)
we want to involve in building our structure. So, when we discuss the in-
tegral domain of integers, our language needs two binary operations (for
addition and multiplication), while, in the latter case, a binary relation (for
the order) is enough. Notice that the language of the ring case works as well
for all the structures admitting two binary operations, and hence possibly
for structures which are not rings; for instance, the reals with the functions
f(x, y) = sin(x - y), g(x , y) = e
X
'
Y
for all x and y in R provide a new structure for our language, but, of
course, the algebraic features of this structure are very far from the basic
properties of integral domains. Accordingly it is advisable, from a general
point of view, to distinguish the constant, operation and relation symbols of
a language L and the elements, operations and relations embodying these
symbols in a given structure for L. Symbols are something like the characters
in a tragedy (like Hamlet) , while their interpretations in a structure are the
actors playing on the stage (Laurence Olivier, or Kenneth Branagh, or your
favourite "Hamlet").
In this framework, we can at last provide a sharp definition of structure.
We fix a language L. For simplicity, we assume that L is countable, hence
either finit e or denumerable (but most of what we shall say can be extended
without problems to uncountable languages).
Definition 1.1.1 A structure A for L is a pair consisting of a non empty
set A, called the universe of A, and a function mapping
(i) every constant c of L into an element cA of A,
and, for any positive integer n,
(ii) every n-ary operation symbol f of L into an n-ary operation fA of
A (hence a function from An into A),
(iii) every n-ary relation symbol R of L into an n-ary relation RA of A
(hence a subset of An).
The structure A is usually denoted as follows
Let us propose some examples, which will be useful later in this book.
1.1. STRUCT URES 3
Ex amples 1.1.2 1. A gr aph is a non empt y set A with a binary relation
P both irreflexive and symmet ric. Hence a gra ph can be viewed as a
struct ure A in the language L consisting of a uniqu e binary relation
symbol R, with RA = P. Also a non empty set A par ti ally ordered by
some relation can be regarded as a st ruct ure in t he same language
L ; this time, RA
2. A (mul tiplicative) groupy is a structureof th elanguageL = {I , ., -I},
where 1 is a constant , . and -I are oper ation symbols of arit y 2 and
1 respectively. 1
9
represent s the identi ty element in y, while .9 and
-1
9
denote the product and th e inverse operation in y. Actually one
might enrich L with some addit ional symbols; for instance, one might
introduce a new binary operation symbol [ ] corresponding to th e
commutat or operation in y. But, for a and bin G, [a , b] is just a . b
a-I . :' , so [ ] is not really new, and is impli citly defined by L .
Actually we will prefer L later ; but it is not ewor th y t hat L can capture
and express some fur th er operation s (and relations and constants) of
y besides t hose lit er ally int erpreting it s symbols.
3. A field K is a struct ure of t he language L = {O, 1, +, - , .} where 0
and 1 are const ant , and +, - and are oper ation symbols (each having
an obvious int erpret ati on in K). Alte rnatively, K ca n be viewed as a
st ruct ure in th e language L' = L U {- I } with a new operation symbol
-I ; obviously, -I has to be int erpreted wit hin K in t he inver se oper-
ation for nonzero elements of K . However , accor ding to t he gener al
definition of struct ure given before, - I JC should denote a 1-ary oper a-
t ion wit h dom ain K . So we run into t he problem of defining 0-
1
; thi s
can be over come by agreeing, for instance, 0-
1
= 0, but t his solut ion
may sound slight ly artificial. So we will pr efer t o adopt below t he
language L when dealing wit h fields. Indeed, when a and b are two
elements in a field K, then a = b-
I
can be equivalent ly expressed by
saying a . b = 1.
4. An ordered field is a st ruct ure in the language L = {+, - , " 0, 1,
obtained by adding a new binary relation symbol It s int erpret ation
in a given ordered field is clear: the order relation in t he fi eld.
5. Let N denote the set of natural numbers. 0 is an element of N ; t he
successor s (mapping each natural n into n + 1) is a 1-ary function
from N t o N. Giuseppe Peano point ed out t hat t he Induction Principle
(t ogether with t he auxiliary condit ions that s is 1 - 1 but 0 is not in
its image) fully characterizes (N , 0, s). A suitable language to discuss
this st ructure should include a constant symbol and a I-ary operations
symbol.
6. Let IC be a (countable) field. A vectorspace V over IC can be regarded
as a st ructure in the language LK = {O, +, - , r (r E K)}, where is
a constant , + and - are operation symbols with arity 2 and 1 respec-
t ively, and, for every r E K, r denotes in LK a l-ary operation symbol ,
to be interpret ed inside V in th e scalar multiplication by r . The other
symbols in LK are interpreted in the obvious way. The assumpt ion on
the cardinality of K has the only role of ensuring LK count able. More-
over, what we have said so far easily ext ends to right or left modules
over a (countable) ring R with identi t y; the corresponding language is
obviously denoted by L
R
.
As already said, we should distinguish symbols and interpretations, for in-
st ance, a binary relation symbol R and t he relation RA embodying it in a
st ructure A (sometimes an order relation in a partially ordered set, but else-
wher e possibly the adjacency relation in a graph). But , to avoid too many
complications, we will often confuse (and actually we already confused) t he
language symbols and their" most natural" interpretations. For instance, in
Example 1.1.2,6, we denoted in the same way the addition symbol + of LR
and its obvious interpretation in a given R-module M , namely the addition
in M .
We will be interested in several algebraic notions concerning st ruct ures. In
part icul ar embeddings play a crucial role in Model Theory. So let 's recall
their definition.
Definition 1.1.3 Let A and B two struct ures in a language L . A homo-
morphis m of A into B is a function f from A into B such that
(i) for every constant c of L, f(c
A)
= cB;
(ii) for every positive integer n , for eve ry n-ary operation symbol P in L
and for eve ry sequence ii = (aI , ... , an) in An, f(pA(ii)) = pB(J(ii))
(hereaf ter f( ii) abridges (J( al) , .. " f( an)) );
(iii) for eve ry positive integer n, for eve ry n-ary relation sy mbol R of L
and for every sequence ii in An, if ii E RA , then f (ii) E R
B
.
.r is called an e mbed ding of A into B if f is injective and, in (iii), f(ii) E R
B
implies ii E RA for every ii in An. When there is some embedding of A
1.2. SENTENCES 5
into E, we write A E. An isomorphism of A onto E is a surjective
embedding. When there is some isomorphism of A onto E, we say that
A and E are isomorphic and we write A E. Finally, an endomorphism
(automorphism) of A is a homomorphism (isomorphism) of A onto A.
Definition 1.1.4 Let A and E be two structures of L such that A B. If
the inclusion of A into B defines an embedding of A into E, A is called a
substructure of E, and E an extension of A .
Now let E be a st ruct ure of L , and A be a non- empty subset of B . We
wonder if A is the domain of a subst ruct ure of E. One promptly reali zes
t hat t his may be false. Indeed
(i) if c is a constant of L, it may happen that cB is not in A;
(ii) if F is an n-ary operation symbol in L , it may happen t hat t he restric-
ti on of F
B
to An is not an n- ary operation in A, in other words that A
is not closed under F
B
;
(iii) on the cont rary, if R is an n- ary relation symbol in L, then R
B
n An is
an n-ary relat ion in A.
So A is not necessarily t he domain of a substructure of E. However the
closure of A U {cB : c constant in L} with res pect t o t he ope rations pB,
when F ranges over the operation symbols in L , does form the domain
of a substructure of L , usually denoted (A) , and called the substructure
generated by A: in this case A is said to be a set of gener ators of (A) . Notice
that these noti ons can be introduced even in the case A = 0, provided that
L contains at least a const ant symbol. E is called finitely generat ed if there
exist s a finite subset A of B such that E = (A).
Finally, let L L
'
be t wo languages, A be an L-structure, A' be an L
'
-
st ruct ure such that A = A' and the int erpretations of the symbols in L are
the same in A and in A'. In this case, we say that A' expands A , or also
that A' is an expansion of A to L
'
; A is called a restriction of A'to L.
1.2 Sentences
Given a language L, after forming the st ruct ures of L, one builds, in a
complementary way, the formul as of L, in particular th e sentences of L, and
one defines when a formul a (a sente nce) is true in a given structure. This is
t he realm of Logic rather th an of Algebra.
Act ually t here ar e several possible ways of introducing formul as and truth,
according to our tastes or our math ematical purposes. We will limit our-
selves in t his book to the first order framework. Let us sket ch briefly how
formul as and truth are usually introduced in t he first order logic. For sim-
plicit y let us work in t he par t icul ar sett ing of na tural numbers (full general
det ails and sharp definitions can be found in any handbook of basic Math -
ema tical Logic, such as [153]).
Conside r the natural numbers and t he corresponding st r uct ure (N, 0, s),
where s denotes the successor function . The corres ponding language L in-
clud es a constant (for 0) and a 1-ary operation symbol (to be int erpret ed
in s). As an nounced at t he end of the pr evious section, we denote these
symbols by still using 0 and s: this is not completely correct , but simplifi es
our life. In the first ord er set t ing, formulas ca n be built by using additional
symbols
count ably many element vari abl es VD, V i, ... , V
n
, ... (just to respect
our count able framework ; oth er wise we can use as many variables as
we need) ,
t he basic connectives /\ (a nd) , V (or) , --, (not) (and even ---+ (if ...,
t hen) , H (if and onl y if) if you like) ,
t he quantifiers V (for all) a nd :J (t here exists) ,
parenth eses (, )
a nd a symbol = to be int erpreted everywhere by t he equality relation. At
this point one forms the terms of L . Essent ially t hey are pol ynomials; in
our case t hey are built st arting from t he constant 0 and t he variables V
n
(n
natural) and using t he oper ation symbols (so s in our sett ing) . The second
step is to construct the atomic formulas of L: basicall y they are equations
between t er ms, but , when th e language includ es a k-ary relation symbol R,
we have to include every state ment saying that a k-uple of t erms satisfies R.
At this point t he formul as of L are built from t he atomic ones inductively
in th e following way:
1. one can negate, or conj unct, or disjunct some given formulas 0:, (3, . . .
and get new formulas --'0: , 0: /\ (3, 0: V (3;
2. one ca n t ake a formul a 0: and a varia ble V
n
and form new formul as
Vvno: , :Jvno:;
3. nothing else is a formul a.
1.2. SENTENCES 7
For a and /3 formul as , a --+ /3, a H /3 just abr idge -,aV/3, (a --+ /3)1\ (/3 --+ a )
respectively. Let us propose some examples in our framework of natural
numbers. The inj ectivity of s can be expressed by t he following formul a in
our language L
while the formul a
Vvo-,(O = s(vo))
says that 0 is not in t he image of s. Actually these formulas are sentences
(each occurring variabl e is under the influence of a corresponding quantifier).
In general , an occurrence of a vari abl e v in a form ula a is bounded if it is
und er th e influence of a quantifier Vv, ::l v, and free otherwi se; a is called a
sentence if, as already said, each occurrence of a variable in a is bounded.
When writing a(v), we want to emphas ize th at th e variabl es freely occurring
in th e formula a are in the tuple v.
2. and 3. are very restrictive conditions, and are the distinctive peculiarity
of first ord er logic. Actually, in Mathematics, one sometimes uses Vand ::l on
subsets (rather t ha n on elements ) of a st ruct ure. This is just what happens
in our set t ing concerning (N, 0, s) with respect t o t he Indu cti on Principle.
In fact, Induction says
for every subset X of N , if X cont ains 0 and is dosed under s, then
X=N.
This statement uses Von su bset s, and t his is not allowed in fi rst order logic.
Accordingl y, the Induction Principle cannot be writ ten (at least lit er ally in
the form proposed some lines ago) in t he first ord er framework. This might
look very disappointing: consequent ly, one may search more powerful and
expressive ways of constructing formulas , for instance by allowing quant ifi-
cat ion on set variables (t his it the so-called second order logic). But act ually
first order logic enjoys several important and reasonable technical theorems,
th at get lost and do not hold any more in these alternative worlds. We
will discuss these results lat er , but it may be useful to quote already now a
theorem of Lindstrom saying (very roughly speaking) that "first ord er logic
is the best possible one" (see [11] for a det ailed expositi on of Lindstrom
theor em).
However , formul as and sentences ar e not sufficient to form a logic. Wh at
we need now to accomplish a complete description of our setti ng is a not ion
a truth. We want to define when a sentence of a language L is true in
a struct ure A of L, and, more generallly, when a sequence ii in A makes
8 CHAPTERl . STRUCTURES
a formul a a(v) true in A. This ca n be done in a very nat ur al way, saying
exactly what one expects to hear. For instance t he sentence 3v(v
2+1
= 0) is
t rue in t he complex field just because C contains some elements i satisfying
t he equation v
2
+ 1 = 0; and in t he ord ered field of reals yI2 makes t he
formul a v
2
= 2A v 2: 0 t rue because satisfies both it s condit ions, while -yI2,
or 1, or other element s cannot satisfy t he same formul a. See again [153], or
any handbook of Mathematical Logic for mor e det ails on t he definition of
first ord er t rut h. We omit t hem here.
Incident ally we note t hat, according t his not ion of t rut h, a V {3 just means
-{.a A ....,{3) , and Vvna says t he same thing as ....,3v
n(
...., a) . So we could avoid
t he connect ive V and t he quantifier V in our alphabet and, consequent ly,
in our inductive definition of formula, and to introduce a V {3 and Vvna as
abbreviat ions, just as we did for a ---+ {3 and a H {3. In this perspective,
formul as are obtained from t he atomic ones by using A, ...." 3 and nothing
else.
Moreover one can see t hat, according t o t his definiti on of t rut h, up t o suit-
able manipulations, each formula ep(w) can be wri t t en as
where QI , . . . , Qn are quantifiers , v = (VI, .. . , v
n
) and a(v, w) is a quanti -
fier free formula, and even a disj unction of conj unctions of atomic formul as
and negations. (*) is called t he normal form of a formul a. When ep(w) is in
its normal form and every quantifier Qi (1 i n) is uni versal V (existential
3), we say t hat ep(w) is universal (exist e nt ia l, respecti vely).
Before concluding t his section, we would like t o emphasize t hat t he st udy of
t his t rut h relation bet ween struct ures and sentences is just Model Theor y,
at least according t o the feeling in t he fifti es. In fact , one says t hat a
struct ure A is a model of a sentence a , or of a set T of sentences in the
language L of A, and one writ es A 1= a, A 1= T respectively, whenever a ,
or every sente nce in T , is true in A. Model Theor y is just t he st udy of this
relati onship between st ruct ures and (set s of) sentences. Tarski provides an
authoritative corroboration of t his claim, when he writes in 1954 [158J
Whit hin t he last years, a new branch of metamathematics
has been developing. It is called theory of models and can be
regarded as a part of t he semantics of formali zed t heories. The
problems st udied in the t heory of mo dels concern mut ual rela-
tions bet ween sentences offormalized theories and mathematical
systems in which these sentences hold.
1.3. EMBEDDINGS 9
It is notable that this Tarski quotation is likely to propose officially for the
first time the expression theory of models. Accordingly, one might fix 1954
as the birthyear -or perhaps the baptism year- of Model Theory (if one
likes this kind of matters). Actually, several themes related to the theory
of models predate the fifties; but one can reasonably agree that just in
that period Model Theory took its first steps as an autonomous subject in
Mathematical Logic and in general mathematics.
1.3 Embeddings
We already defined in 1.1 embeddings and isomorphisms among structures
of the same language L. We followed the usual algebraic approach. How-
ever there are alternative and equivalent ways, of more logical flavour , to
introduce these notions. Let us recall them. First we consider embeddings.
Theorem 1.3.1 Let A and B be structures of L , f be a function from A
into B. Then the following propositions are equival ent :
(i) f is an embedding of A into B;
(ii) for every quantifier free formula If'(if) in L and for every sequence ii
in A, A F If'(ii) if and only if B F 1f'(J(ii));
(iii) for every atomic formula If'(if) in L and for every sequence ii in A,
A F If'(ii) if and only if B F 1f'(J(ii)) .
The proof is just a straightfoward check using the definitions of embedding,
term and (atomic or quantifier free) formula. Referring to definitions is a
winning and straightforward strategy also in showing the following charac-
terizations of th e notion of isomorphism.
Theorem 1.3.2 Let A and B be structures of L, f be a surjective function
from A onto B. Then the following propositions are equivalent:
(i) f is an isomorphism of A onto B;
(ii) for every quantifier free formula If'(if) in L and for every sequence ii
in A, A F If'(ii) if and only if B F 1f'(J(ii));
(iii) for every atomic formula If'(if) in L and for every sequence ii in A,
A F If'(ii) if and only ifB F 1f'(J(ii));
(iv) for every formula If'(if) in L and for every sequence ii in A, A F If'(ii)
if and only if B F 1f'(J(ii)).
It can be obser ved t hat , when f is any embedding of A int o B, for every
qu anti fier free formul a a( 5, w) in L and every sequence ii in A,
if A F= 3wa(ii, w), t hen B F= 3wa(J(ii), w)
or also, equivalently,
if B F= Vwa(J (ii ), w) , t hen A F= Vwa(ii , w) .
Definition 1.3.3 T wo structures A and B of L are e le me nt a r ily equiv-
alent (A == B) if they satisfy the same senten ces of L.
As an eas y corollary of Theor em 1.3.2, we have:
Theorem 1.3.4 Isomorphic str uctures are elementarily equivalent.
Co nversely, it may happen t hat eleme ntarily equival ent st ructures A and B
are not isomorphic. We will see count erexamples below. However it is an
easy exercise t o show t hat , for fini t e struct ures, elementary equivalence and
isomorphism ar e j ust t he same t hing.
Now let us int roduce a related notion: parti al isomorphi sm.
Definition 1.3.5 Let A and B be structures of L. A partial isomorphism
between A and B is an isomorphis m bet ween a substruct ure of A and a
substructure of B. A and B are said to be partially ieomorphic A B
if there is a non empt y set J of partial isomorphisms bet ween A and B
satisf ying the back-and-forth propert y: for all f E I ,
(i) fo r every a (: A, there is some gE l such that f 9 and a is in the
domain of g,.
(ii) f or eve ry b E B , there is some g El such that f 9 and b is in the
image of g.
Example 1.3.6 Two dense linear ord erings wit hout endpoints A = (A, s;)
a nd B = (B, S;) are parti ally isomorphic.
In fact, let I include all t he possible isomor phisms bet ween a finit e substruc-
t ure of A and a finit e subst ruct ure of B. I is not empty, because, for every
a E A and b E B , a t-+ b defines a par ti al isomorphism in I. Now take any
f E I ; let ao < a l < . . . < an list t he element s in t he domain of f and
b
o
< b
l
< .. . < b
n
t hose in t he image of f ; so f (ai ) = b, for every i S; n.
1.3. EMBEDDINGS 11
Pick a E A, and notice that th ere exists some b E B such that, for every
i n,
a; a b, b.
This is trivial when a is in th e domain of f . Otherwise, one uses the facts
that B has no minimum when a < aa, that B has no maximum when a > an,
and, finall y, th at t he order of B is dense in the remaining cases. Define gEl
by putting
Domg = Domf U {a}, Img = Imf U {b} ,
g I, g(a) = b.
Clearly g satisfies (i). (ii) is proved in the same way.
Remark 1.3.7 If A B, then A B.
In fact , let f be an isomorphism of A onto B. I = {f} does satisfy (i) and
(ii) .
Conversely, partially isomorphic structures may not be isomorphic.
Indeed one can find two struct ures that admit a different cardinalit y, and
yet are partially isomorphic. For instance, t his is t he case of t wo dense
linear orderings without end points. We have j ust seen that th ey are al-
ways partially isomorphic, indipendently of their cardinalities; in particular
(Q,
But one can also find partially isomor phic non isomorphic struct ures with
the same cardinality. For instance, st ill consider dense linear orderings with-
out endpoint s, and notice t hat (R, (R +Q, ((R +Q, denotes
here the disjoint union of a copy of (R, and a copy of (Q, where
(R, precedes (Q, Both (R, and (R + Q, have t he contin-
uum power. But they cannot be isomorphic, because (R + Q, unlike
(R, cont ains some countable intervals, and any order isomor phism maps
countabl e int ervals onto countabl e int ervals.
However, with in countable models, parti ally isomorphic st ruct ures are also
isomorphic.
Theorem 1.3.8 Let A and B be countable partially isomorphic structures.
Then A B.
The pr oof is obtained as follows. First one list A and B in some way
A = {an: n E N}, B = {b
n
: n E N}.
Let J be a set of parti al isomorphisms bet ween A and B ensur ing A c::=. p B.
Due to (i) and (ii) one enlarges a given 10 E J by defining, for every nat ur al
n, a function In E I such t hat , for any n,
2. an is in the domain of [z,
3. b
n
is in the image of fz n+l.
Pu t 1= UnENln. Owing t o 1. , I is a function; 2. impli es t hat its domain
is A, and 3. ensures that its image is B. In ord er to conclude that I is an
isomorphism, we have t o check t hat , for every atomic formula cp (v) of L and
every sequence ii in A, A 1= cp (ii) if and onl y if B 1= cp (J (ii)). But t his is
easily don e, as t here is some n such t hat ii is in t he domain of In' and In
rest rict s I and is an isomorphi sm bet ween it s domain and it s image.
A notewort hy consequence of t he t heorem is
Corollary 1.3.9 (Cantor) Two countable dense linear orderings without
endpoints are isomorphic.
Hence lineari t y, densit y and lack of endpoint s cha racterize t he ord er of ra-
t ional s up t o isomorphism. It should be underlined t hat Cant or 's ori ginal
proof used a different argume nt; but a subsequent approach of I-I ausdorff
and Huntington inaugurat ed t he back-and-forth method. In fact, what th ey
did was just firstly to obser ve that two dense linear ord ers without end-
points are partially isomorphic (according to our modern t erminology) , and
consequently to deduce that , if one adds the countability ass umpt ion, th en
isomorphism follows; t he lat t er point can be easily generalized t o arbitrary
struct ures (and act ually t his is what Theorem 1.3. 8 says) . Now let us com-
pare c::=. p and =.
Theorem 1.3.10 Partially isomorphic structures are eleme ntarily equiva-
lent.
In fact let A and B be parti ally isomorphic st ruct ures in a language L, and
let I be a set of parti al isomorphisms between A and B witnessing A c::=. p B.
1.3. EMBEDDINGS 13
Then one can show that, for every choice of a formula B l= <p(J (ii)).
Note that , when <p ranges over th e sent ences of L, this impli es A == B. The
proof proceeds by a st raight forward induction on <p (v): t he condit ions (i)
and (ii) are useful in handling t he quantifier step. In fact suppose th at
<p(v) is of th e form 3wa(v, w) . If A l= <p (ii) , t hen there is some b E A
satisfying A l= a(ii , b) . According to (i), th ere is some 9 E J enlarging
f such t hat b is in the domain of g. By indu ction, B l= a (g (ii), g(b)) .
Therefore B F 3wa (g (ii) , w), B l= 3wa (J (ii) , w) and, at last, B l= <p(J(ii)) .
The converse is proved in a simi lar way, using (ii) instead of (i) .
As a consequence of t his theorem, one can deduce that t here exist elemen-
tarily equivalent struct ures which are not isomorphic. For instance, among
dense linear orders wit h no endpoints, (Q, :s;), (R, :S; ) and (R +Q, :S; ) are
partially isomorphic, hence elementarily equivalent. But t hey cannot be
isomorphic, as already obs erved.
Notice t hat Theorem 1.3.10 provides also anoth er proof that isomorphism
implies element ary equivalence (via partial isomorphism) . However, we will
see lat er t hat elementarily equivalent structures may not be part ially iso-
morphic (we will produce a counterexample). Let us also quote here th e
following result , characterizing element ary equivalence in t erms of partial
isomorphism.
Theor e m 1.3.11 (Fraisse] Let A and B be structures in a fini te language
L. Then A == B if and only if there is a decreasing sequences {In : n E N}
of non empt y set s In of partial is omorphisms between A and B suc h that,
for eve ry natural n and every f E Jn+1 ,
(a) for eve ry a E A , there is 9 E In su ch that 9 extends f and the
domain of 9 incl udes a,
(b) for every bEE, there is 9 E In su ch that 9 ex tends f and the image
of 9 incl udes b.
Now let us deal again with arbitrary embeddings. The charact erization of
isomorphism given in Theorem 1.3.2 (iv) suggests the following notion.
Definition 1.3.12 Let A and B be structures of L. An embedding f of
A into B is called elementary if, f or every f ormul a <p(v) in L and eve ry
sequence nin A , A l= <p(ii) if and only if B F <p(J(ii)) .
We say that A is elementarily embeddable in E, and we write A <
E, when th ere is some elementary embedding of A in E. A is called an
elementa r y s ubstruct ure of E if A B and the inclusion of A in B
defines an elementary embedding of A in E (hence, for every formula <p(v)
in L and every sequence a in A, A F <p(a) if and only if E F <p (a)). In this
case, we say also that E is an elementar y extension of A.
Remark 1.3.13
bedding.
1. Every isomorphism is, of course, an elementary em-
2. If A is element arily embeddable in E, then A and E are elementar-
ily equivalent (just restrict the definition of element ary embedding to
sentences <p in L) .
Examples 1.3.14
structures
(a) In the language L = { <}, consider the following
A = (N - {O} , <) , E = (N, <).
The inclusion of N - {O} in N defines an embedding of A in E which is
not elementary because 1 is a mini mal element in A, but not in E (in
other words, satisfies 3w(w < v) in E, but not in A) . On th e cont rar y,
t he successor function from N in N - {O} is an isomorphism between
E and A.
(b) The real field R is a subfield of the complex field C , and hence is
a substructure in th e language L = {O, 1, +, " - }. However R is not
an eleme ntary substructure of C . In fact, for every positive real a, a
satisfies t he formula 3w(w
2
+ V = 0) in C, but not in R . On t he other
hand, we will see later t hat every embedding of algebraically closed
fields (or real closed fields) is elementary.
Another remarkable class of embeddings concerns existe nt ial formulas.
Definition 1.3.15 Let A and E be structures in a language L. An embed-
ding f of A in to E is called ex istential if and only if, for every existential
formula <p(v) in L and for every sequence a in A,
A F <p(a) {::} E F <p(J(a)).
In more detail , we require that, for every qua ntifier free formul a a( a, w) in
L and for every sequence ain A,
(*) A F 3wa(a, w) if and only if E F 3wa(J(a), w).
1.3. EMBEDDINGS 15
When A Band f is the inclusion of A into B, we say that A is an
exist e ntial subst ruct ure of B. Wh en t here is some existe nt ial embedding
of A in B, we say t hat A is existentially e mbedde d in B and we write
A <1 B.
Let us discuss bri efly existent ial embeddings f . First of all, it is clear t hat
element ary embeddings ar e existent ial, t oo. Furthermore, notice that (*)
can be weakened to require
if B F 3wo:(J(a), w), then A F 3wo:(a, w)
because t he inverse implic ation
if A F 3wo:(a, w), then B F 3wo:(J(a) , w)
is satisfied by every embedding. Now, a quantifier free formula 0:( V, w) can
be equivalent ly writt en (by standard propositional tec hniques) as a disjunc-
t ion
vO:j (v, w)
j5,s
where s is a natural number and, for every j ::; s, O:j (v, w) is a conj unc-
t ion of atomic formul as (equations) and negations. One easily deduce t hat
3wo:(v, w) can be equivalent ly written as
V3wO: j( v, w).
j 5,s
It follows t hat f is exist enti al if and only if
B F 3wo:(J(a) , w) implies A F 3wo:(a, w)
for every a in A and for every finite conj unct ion 0: (V, w) of equations and
negations.
Now let us deal again with ar bit ra ry elementary embeddings. Let L be
a language, A be a structure of L. In ord er to examine the st ruct ures
element arily equivalent to A, L is just the language we need. But, within
t hese models, one meet s t hose where A is elementarily embeddable; and
t hese st ruct ur es actu ally require a richer language, emphas izing th e fact
t hat A embeds (element ari ly) in each of them. This lar ger language is buil t
by adding to L a constant symbol for every a E A, and will be called L (A) ;
t he new constant corresponding t o a could be written Ca, in ord er t o remind
a but t o avoid any confusion with a. But we will oft en denote it ambiguously
by a (for simplicit y's sake) .
A it self becomes a struct ure of L (A) provided we int erpret quite naturally,
for every a E A , t he constant cor responding t o a in a. The result ing struct ure
will be denot ed AA.
Wh ich are t he st ruct ures elementarily equivalent t o AA in L(A)? They are
obt ained as follows. Take an L-struct ure B where A embeds elementarily,
say by f ; for every a E A , let f (a ) int erpret t he constant of a. One get s in
t his way a struct ure BJ(A) of L (A), and it is easy to check t hat BJ(A) == AA
Conversely, every st ructure elementarily equivalent t o AA ca n be obt ained
in t his way.
More generally, for every struct ure A of L and for every subset X of A, one
can introduce a new language L(X) by adding to L a new const ant for every
element x in X. Ax is the L(X)-structure expanding A and interpreting,
for every x EX, the constant symbol corresponding t o x in x it self. Of
course, t here do exist ot her struct ures elementaril y equivalent t o Ax. Let
us see how to construct t hem. Let B be a st ruct ure of L.
Definition 1.3.16 A f unct ion f fro m X in to B is called elementary if ,
f or eve ry formula cp( if) in L and for every sequence i in X ,
A F cp (i ) {:} B F cp (J(i )).
Not ice t hat , when X = A, an element ary functi on from X in B is j ust
an eleme ntary embedding of A in B. Moreover , for any X , an elementary
functi on f from X in B enjoys t he following propert ies.
(i) f is 1 - 1 (use CP (VI' V2) : VI = V2) .
(ii) f-
I
itself is an elementary function (from f (X) in A).
(iii) A and B are elementarily equivalent (apply t he definition of elementary
function to the sente nces cp of L).
Now take an L-structure B. Let f be an element ary function from X into
B , and, for every x E X , let f( x) int erpret the constant of x in L(X). One
gets in thi s way a st ruct ure BJ(X ) of L(X) , and even a mod el elementarily
equivalent to Ax. Conversely, every st ruct ure == Ax ca n be obtained in t his
way.
Remark 1.3.17 Let cp(V, w) be a formula of L , d be a sequence in A, i be
a sequence in X. Fussy people will like t o distin guish
(a) A F cp( ii, i) (in t he sense t hat A satisfies t he L-formula cp (v, w) if
V, ware embodied by ii, i respect ively) ;
1.3. EMBEDDINGS 17
(b) Ax F i.p( ii , x) (in the sense that Ax sat isfies the L(X)-formula
i.p( if, x) if if is embodied by il);
(c) AA F i.p(il, x) (in th e sense that AA satisfies th e L(A)-sentence
i.p(il, x)).
However one easily shows that (a), (b) and (c) are equivalent. So we will
use t he simplest notation (that of (a)) to mean any of these conditions.
We conclude t his section by mentioning without proof two theorems on
elementary emb eddings. We will state them for simplicity in t he case when
the involved embeddings are just inclusions but they extend to arbitrary
(elementary) embeddings.
The for mer theorem provides a crit erion to check whether a given subset of
a struct ure B is the domain of an elementary substructure of B (remember
the discussion at th e end of 1.1) .
Theorem 1.3.18 (Tar ski-Vaught) Let B be a structure of L, A be a subset
of B. Th en the following propositions are equivalent:
(i) A is the domain of an eleme ntary substructure A of B;
(ii) for every fo rmula a( if , w) of L and for every sequence ii in A , if B F
:3wa(ii, w), then there exists an element b E A such that B F a(ii, b) .
Now let us introduce the latter result. Take a set I tot ally ordered by a
relation For every i E I , let A i be a struct ure of L. Suppose t hat , for
every choice of i j in I , A i is a subst ruct ur e of A j: this means th at
A- C A and
- J
for every constant c of L , cA; = c
Aj
,
for every n-ary operation symbol F of L, FA; is t he restriction of FAj
to Ai ,
for every n- ary relation symbol R of L, RA; = RAJ n Ai .
Therefore we can build a new structure A in L, having domain A = UiEI Ai
(and hence including Ai for all i E 1), and int er pret ing t he symbols of L as
follows:
for every constant c of L, cA = cA; where i is any element in J;
18 CHAPTER l . STRUCTURES
for every n-ary oper ation symbol F of L and for every choice of
aI , ... , an in A,
where i E I sat isfies aI, . . . , an E Ai;
for every n-ary relation symbol R of L and for every choice of aI , ... , an
in A,
where i E I sat isfies aI, ... , an E Ai.
It is clear that A is well defined and ext ends Ai for every i E I. Straight-
foward t echniques show:
T heorem 1.3.19 (Element ary Chain Th eor em) Suppose that , for eve ry
choice of i ::; j in I , Ai is an elementary substructure of Aj. Th en, for
every i E I , Ai is an elementary substruct ure of A.
1.4 The Compactness Theorem
The Compact ness Theorem is t he most powerful tool -and indeed a key
feature- in classical Model Theory. It states
Theorem 1.4.1 Let S be an infinite set of sentences in a language L . Sup-
pose that every fin it e subset of S has a model. Then S has a model.
Notice t hat the converse is obvious, because a model of S is a model of
every (finite or infinit e) subset of S . But t he t heorem ensures t hat , if every
finit e subset of S has it s own model (hence different subset s may admit
different models), then th ere is a global model sat isfying all the sentences
in S. In fact , there are several sit uat ions where the Compactness Theorem
applies and guarantees satisfiability for sets S of sente nces for which it is
very difficult to imagine a general model directl y, but it is quite simple t o
equip every finit e subset wit h a suit able privat e model: we will see some of
them lat er in 1.5. In fact t his section and (impli citl y) t he next one will be
devot ed t o discussing t his fundamental t heorem and, in det ail:
it s proof;
it s name;
1.4. THE COMPACTNESS THEOREM 19
it s role in producing " nonstandard" and, in some sense, unexpect ed
models, and, as th e reverse side of t he same medal , in bounding t he
expressiveness of first order logic and in excl uding t hat some familiar
principles, like induction on naturals, may be writ t en in any way in
t he first order framework ;
in spite of t his, it s plausibili t y, supported by metamathematical con-
siderations on t he nature of math ematical proofs;
finally, some words about t he already mentioned t heorem of Lindstrom
saying t hat the Compact ness Theorem, together with a related result
(th e downward Lowenheirn Skolem Th eor em) fully cha racterizes first
order logic.
As said, let us begin by discussing the proof. There are several possible
ways t o show t he Compactn ess Theorem. For instance, t here is an approach
based on t he algebraic noti on of ultraproduct and due t o Keisler (see [39]) .
Anot her classical pr oof was proposed by Henkin. Let us outline very qui ckly
its idea.
Wh at we have at t he beginning is a( n infinit e) set of sent ences S such t hat
every fi nit e subset has a model. Some technical -and non t rivial- pr eliminary
wor k shows t hat t here is no loss of gene rality in ass uming t hat S satisfies
t wo fur th er cond itions:
1. S is complete, in oth er words , for every sente nce rp in L , eit her rp or
-' rp is in S;
2. S is rich: if S contains a sentence of t he form :Jva:(v), t hen t here is a
constant symbol c in L such t hat a:(c) is in S .
At t his point a qui t e art ificial const ruct ion produces t he model we are look-
ing for. Basically, th e domain is just the set of t erms without variables in L
(the so-called Herbrand universe of L) ; this is non-empty owing to 2. The
L-structure arises in a rather reasonable way. 1 and 2 play a key role in
showing t hat what we build is a model of S .
It is worth emphasizing t hat t he model we get in our proof is count able (for
a countable L ; when t he language has a larger cardinality -X , it is easy t o
check t hat our argument st ill works and produces a model of power s: -X).
So, as a byproduct of t he Henkin proof, we have t hat , when S has a model,
t hen S has a countable model: t his is t he so called Downward Lowenli eim-
Skolem Theorem, and is a notable resul t . Vve shall discuss it s relevance in
1.5 .
Now let us t reat t he reason s of t he t heorem name. Act ually compactness
recalls t opology. In fact , we will see later in Chapter 5 that th e th eor em has
a topological content and implicitly says that a cert ain topological space is
compact .
We shall see within a few lines in 1.5 t hat t he Com pactness Theorem pro-
duces some strong and severe expressiveness rest rictions in first order logic.
For inst ance, we will show t hat , just owing to Compactness, condit ions like
finit eness, or popular statements such as t he Minimum Principle, cannot be
expressed in a first order way. On t he other hand, one should agr ee t hat
what t he Compact ness Theorem says is a qui t e reasona ble statement, espe-
cially if one considers t he following corollary. Let 5 be a set of sente nces of
L and a be a sentence of L ; we say that a is a logical consequ ence of 5, and
we wri t e 5 F a , when a is true in all the mod els of 5 .
Corollary 1.4.2 If 5 F a , then there is a fi nit e subset So of 5 such that
So Fa.
Proof. Clearly 5 F a if and onl y if 5 U {-,a} has no models. But , owing
to Compact ness, t his is equivalent t o say t hat t here is a finit e subset of
5 U {-,a} without any model s. Wi th no loss of generality we can assume
that this finite set is of the form So U {-,a} where So 5 . But , again,
stating t hat So U {-,a} has no models is equivalent t o say that So F !.p ...
Now, anot her fundament al result in first order logic, deeply related t o com-
pactness -the Completeness Theorem- says that one can explicit ly provide
a noti on of provabili ty accompanying and support ing this concept of con-
sequence in such a way t hat t he logical consequences of a given 5 are just
what is proved by 5 at the end of a sequence of rigorous deductions. So
what compact ness in conclusion emphas izes is t he finit ary nature of math-
emat ical pr oofs; t his feature can be regard ed as an aut horitat ive witness in
it s favour and, t rough it , as a support to first order logic.
1.5 Elementary classes and theories
When considering, for a given language L , struct ures, formul as and t ruth,
two problems arise quite naturally:
(a) given a set T of L-sentences, "classi fy" th e models of T (their class
will be denoted Mod(T));
1.5. ELEMENTARY CLASSES AND THEORIES 21
(b) given a class K of Le-structure, "characterize" t he set of t he L-sentences
true in all the structures of K (this set will be called t he t heory of K
and denoted T h( K )).
According to it s declar ed int ent s, Model Theory should be mainly concerned
with Problem (a) . However , also (b) arises quite naturally in the model
t heoretic framework. For inst ance, consider a class K formed by a single
structure, like t he complex field, or t he real field. We will see lat er that K
cannot be represented as Mod(T ) for any T. But it may be qui t e interesting
to realize in an explicit way which sentences are true in t he onl y structure
in K.
However we have to admit t hat the previous statements of (a) and (b) are
somewhat vague and unprecise. First of all, what do classifying or charac-
terizing mean? This is not a minor question; on the contrary, it is a very
delicat e and central matter. For inst ance, t he classification problem for a
class of structures touches and overlaps several basic ope n questions in Al-
gebra . So we should be more detailed about t his crucial point . Of course,
one can reasonably agree that a classification should ident ify isomorphic
structures. But t his is still a partial and indefinite answer; we should fix
more precisely which criteria, tools and invariants we want to use in our
classification problem. We shall try to clarify these fundamental questions
in the next chapters. Her e we limit ourse lves to discuss other points, mainly
concerning (a). T is a set of sentences in a language L .
1. Let a be a sentence of L, and suppose that every model of T is also a
model of a (so a a logical consequence of TT 1= a). Hence Mod(T) =
Mod (T U{a}). Consequent ly we can assume wit h no loss of generality
for our purposes t hat
for every sentence a of L, if T 1= a, t hen a E T.
A set of sentences in L with this property is called a t heory of L. It
is a simple exercise to show that, given a set T of sentences of L, T is
a t heory if and only if t here is a class K of st ruct ures of L such that
T =Th(K ) (hint : ({::: ) is clear; t o show (=?) use K = M od(T) ).
2. A theory T is called consistent if and only if T satisfies one of the
following (equivalent ) conditions:
(i) for every sentence a of L, either a rf. T or -,a rf. T;
(ii) there is some sentence of L which is not in T;
22
(iii) M od(T) -# 0.
CHAPTERl. STRUCTURES
To show t he equivalence among (i) , (ii) and (iii) is an easy exercise. (i) says
t hat t he consist ent theories are just t hose excluding any contradiction. (iii)
ensures t hat t hese th eori es are exactly t hose admitting at least one model. It
is clear t hat, within Problem (a), we are exclusively interested in consist ent
th eori es. So we can assume in (a) th at T is a consi stent th eor y; accordingly
her eafter theory will always abbreviat e consistent theory.
A rigid model theoretic perspective might limit th e classification analysis to
t he classes of models of (consistent) theories. But open minds could prefer
a more gener al study, providing an abstract t reat ment of th e class ification
problem for arbitrary classes of structures. Hence it is worth underlining
that there do exist classes K of L-struct ures which are not of th e form
K = M od(T) for any theory T of L. We propose here some examples; t he
Compactness Th eorem is a fundamental tool in t his set t ing.
Definition 1.5.1 A (non-empty) class K of struct ures of L is said to be
elementary (or also axiomat izable) if there is a set T of sentences of L
(without loss of generality, a theory T of L) such that K = M od(T).
Now let us propose a ser ies of examples, as promised. Par t of t hem aim
at pointing out t hat several classes of st ruct ures are explicit ly element ary
because t heir definitions can be naturally wri t t en in a first order way. But
oth er cases are not element ar y: it is here that the Compact ness Theorem
plays it s role and first ord er logic shows it s expressiveness bounds.
Examples 1.5.2 1. Let L = 0 (so the str uct ures of L are th e non-empty
sets), K be the class of infinit e sets . "Infinite" means " admitting at
lea st n + 1 elements for every natural n" . Given n, the property
"there are at least n + 1 elements" can be expressed in a first ord er
way by t he following sentence of L
o; :Jvo .. . :Jv
n
1\ -'(Vi = Vj ) .

Hence K = M od(T) where T = {(Yn : n E N}, and so K is elementary.
2. Let again L = 0, but now let K be th e class of finit e (non-empty)
sets. " Finite" means " having a t most n + 1 elements for some natural
n" . Given n, the proposit ion" there are at most n +1 elements" can
be expressed in a first order way by the sentence -,a
n
+l . But now
M od( {-,a
n
+! : n E N}) is not K, indeed it equals the class of the
sets having only one element. So the approach in 1 does not work any
longer. However, assume that K is elementary, hence K = M od(T)
for a suitable set T of sentences of L. Put
T' = Tu {an: nE N}.
Let T6 be a finite subset of T'. For some natural N,
T U {an: n EN, n < N}.
Notice that TU {an: nE N, n::; N} (and hence T6) has a model: it
suffices to take a finite set with at least N + 1 elements. At this point ,
owing to the Compactness Theorem, we deduce that T' itself has a
model. This is a set both finite (as a model of T') and infinite (as a
model of an for every natural n). We get in this way a contradiction.
Hence K is not elementary.
Notice that this argument works as well for every class K of finite
arbitrarily large structures (in the sense that, for every positive integer
n, there is a structure in K whose size is larger than n). A class K
of this kind cannot be elementary; in other words, the theory of K
does admit infinite models, too; notice that this applies, for instance,
to the class of finite groups, as well as to the class of finite fields. So
one can wonder which are the infinite models of the theory of these
finite structures. We will consider the particular case of fields later in
Example 6.
Now let us deal with orders.
3. Let L = {::;} where j, is a binary relation symbol (which we confuse,
for simplicity, with its interpretation below -an order relation-). In
L we consider the class K of linear orders. It is easily seen that K
is elementary, because the properties defining linear orders are first
order sentences (and indeed universal first order sentences) of L. For
instance, linearity can be expressed by
The set of the logical consequences of these sentences is the theory of
linear orders ; it is formed by the sentences true in every linear order.
In the sa me way the class of dense linear orders without endpoints is
element a ry; in fact, density is st at ed by
VvoVVI :J V2 (vo < VI -+ Vo < V21\ V2 < vd ,
and lackness of endpoints by
VVO:J VI ( VI < Vo),
VVO:JVI (vo < VI )
(vo < VI abbrevi ates here Vo:::; VII\ --, (VO= vd ).
The set of the logical conseque nces of the sent ences quoted so far is the
t heory of dense linear orders without endpoints; we will denote it by
DLO-. It is formed by t he sent ences true in every dense linear order
with no endpoint s. Recall t hat (Q, :::;) , (R, :::; ) are dense linear orders
without endoints, and consequently their theories include DLO- (and
one may wonder if they actually equal DLO-).
The reader can check directly t hat t he following clas ses of L-structures
are eleme nt a ry:
dense linear orders with a least but no last element, or a last but
no least element, or both a least and a last element,
infinite discrete linear orders with or without endpoints (an order
is discrete when ever y eleme nt, but t he leas t on e -if any- , has a
pr edecessor a nd eve ry eleme nt, but t he last on e -if any- , has a
success or) .
4. We still work in L = {:::;} (wh er e j, is a binary relation sy mbol), bu t
this t ime we deal with t he clas s K of well ordered sets (so ordered set s
where every non-empty subset has a least element ) . Hence (N , :::;) E K
owing to t he Minimum Principle, while any dense linear order (A, :::;),
even wit h a minimum, does not lie in K (in fact, given b > a in A,
which is t he least element > a in A?) . So the situation is , in some
sense, opposite t o the last (elementary) exa mple 3.
Suppose that K is element ary, so K = M od(T) for a suitable set T of
L-sent ences. Put
L' = L U {c
n
: n E N}
where, for every natural n , C
n
is a const ant symbol, and in L' look at
the following set of sentences
T' = TU {Cn+1 < C
n
: nE N }.
Let Tbbe a finite subset of 1", t hen t here is some natural N such that
Tb T U { Cn+l < C
n
: n EN, n N}.
Then Tb has a model because T U {Cn+l < C
n
: n EN, n N}
has : it suffices to t ake the well ordered set (N, to int erpret Co,
Cl , . . . , CN+! in N +1, N, . .. , respectively, and any furth er constant
C
n
(with n > N) ar bit ra rily. By the Compactness Theorem, 1" does
admit a model
A' = (A, (C;;') nEN) .
Let A = (A, th en A is a model of T, and hence is a well ordered
set; however it contains t he non- empt y subset
X={C;;': nEN}
admit t ing no minimum, because, for every natural n, < c;;'. So
we get a contradiction. Consequent ly K is not elementary. In oth er
words th ere are linearl y ordered sets which are not well ordered but
sat isfy the same first order sentence as well ordered sets.
5. Let now L = {O, 1, +, - , .} be our language for fi elds. We consider in
L t he class K of fields. K is elementary. In fact t he definition it self of
field can be wri t ten as a series of first order sentences (in most cases,
of universal first order sentences) in L . For instance
says that any non zero element has an inver se.
Also t he class of algebraically closed fields is elementary, although the
corresponding check is a lit tle subt ler. In fact what we have to say now
is that , for every natural n , any (monic) polynomial of degree n + 1
has at least one root. So the point is how to quantify over polynomials
of degree n + 1. However recall that such a polyn omi al is just an
ordered sequence of length n +2 of element s in th e field: the first is
t he coefficient of degree 0, t he last is t he coefficient of degree n + 1
(and equal s 1 if we deal with monic polynomials) ; so what we have to
write is j ust , for every n,
(where vi has the obvious meaning, for ever y i n +1).
The logical consequences of t he sent ences listed so far form the theory
of algebraically closed fields, usually denoted ACF. Now let p be a
pri me, or p = o. Also the class of (algebraically closed) fields of charac-
teristic p is elementary; for, it suffices to add to the previous sentences
the one saying that the sum of p t imes 1 is 0 when p is a prime, or,
when p = 0, the negations of all t hese sentences. In conclusion, for
every p prime or equal to 0, we can introduce the theory ACF
p
of
algebraically closed fi elds of charact eristic p. Among the algebraically
closed fields in characteristic 0 recall the complex field C , as well as
the (countable) field Co of complex algebraic numbers; th eir th eori es
contain ACF
o
, and one may wonder if they equal ACF
o
. Recall also
that ever y field K has a (minimal) algebraically closed extension K;
in particular, for p prime, Z/ pZ is an example of algebraically closed
field in characteristic p.
Since we are t reating fields , let us consider agai n finit e fields, and,
mor e exactly, the infinite models of their t heory we met in example
2: th e so called pseudofinite fields. As observed before, one can ask
which is the st ructure of these fields. J. Ax equipped them with a very
elegant axiomatization, explaining the essential nature of finite fields
in t he first order setting: in fact , pseudofinite fields are just the fields
K such that:
* K is perfect,
* K has exactly one algebraic ext ension of every degree,
* every absolutely irreducible vari et y over K has a point in K.
All these conditions ca n be written in a fi rst order way, although this
is not immediat e to check.
6. A first order language for the class K of ordered fields is L = {O, 1, +, - ,
., ::;}. K is elementary in L because it equals M od(T) where T is the
set of the following sentences in L:
(i) the field axioms (see Example 5) ;
(ii) those characterizing the linear orders (see Example 3);
(iii) the sentence saying that sums and products of nonnegative ele-
ments are non negative
Also t he class of real closed ordered fields (t hose satisfying t he Int er-
mediat e Value Propert y for polynomials of degr ee 1) is elementary,
it suffices to add t he new sentences:
(iv) for every natural n,
A 0 < Vo + vI . W+ ... +V
n
. ui" + w
n
+
1
A u < w -+
-+ u < v A v < w A Vo +vI . V+ ... +V
n
. v
n
+v
n
+
1
= 0).
T he logical consequences of (i), (ii), (iii) , (iv) form t he theory of real
closed ordered fields, usually denoted RC F. Examples of real closed
ord ered fields are t he ord ered field of t he real numbers R , as well as t he
(count abl e) ord ered field R
o
of real algebraic numbers. Their t heories
include RC F , and one may wonder if act ually t hey equal RC F.
7. Let R be a (countable) ring wit h identi t y. Consider t he language
L R = {O, +, - , r (r E Rn of (left ) R-modules. The class of left R-
modules is elementary because it equals t he class of models of t he
following sentences in L R:
(i) t hose axiomatizing t he abelian groups in t he language wit h 0, +
and - ;
(ii) for every r , s E R, if r +s and r s denot e t he sum and t he product
(respecti vely) of rand s in R,
\lvo( (r + s) vo = rvo +sVo) ,
\lvo((r s)vo = r( svo)) ,
\lvO\lvl (r( Vo +VI) = rvo + rvt} ,
(iii) finally, if 1 denot es th e identity element in R ,
\IVo(1Vo = vo).
The logical consequences of t he previous sentences form t he t heory
n T of left R-modules. Of course, t here is no reason to pr efer the
left to t he right , at least in t his case ; indeed, one can check t hat even
t he class of right R-modul es is element ary, and consequent ly one can
introduce t he theory T
n
of right R-modules.
28
CHAPTER 1. STRUCTURES
Let us come back to our classification probl em for element ary, or also non-
element ary classes. The following fundamental t heorem can suggest t hat ,
even in the elementary case, t his problem is not simple, as the class of models
of a theory T can include many pairwise non-i somorphic structures.
Theorem 1.5.3 (Lowenheim-Skolem] Let T be a theory in a (countable)
language L . Suppose that T has some infi nite model. Then, for every infinite
cardinal A, T admits some model of power A.
The proof just uses Compactness in the extended framework of languages
of arbitrary cardinali ti es. In fact one enlarges L by A many new constant
symbols c; (i El, III = A) and one gets in t his way an extended language L' .
In L' one considers the following set of sentences
T' = TU {--'(Ci = Cj) : i, j E I, i # j }.
Any finit e portion of T' has a mod el; in fact it turns out that , for some
finite subset 1
0
of I ,
so, in order t o obtain a model of it is sufficient to refer to an infinite
model A di T , as ensured by t he hypothesis, and t o interpret the finit ely
many cons t ants c; (i E 1
0
) in pairwise different elements of A. At this point ,
Compact ness applies and gives a model of T' (hence of T) of power :S: A.
But t his mod el has t o include the A many distint int erpret at ions of the c/s,
and so it s power is exactl y A.
T herefore, if a theory T of L has at least an infinite model then T has a model
in each infinite power (and two models with different cardinalit ies cannot
be isomorphic). Of course, one may wonder how strong is the assu mption
th at T has some infinit e model. Not so much, if one recalls that a th eory
T admit t ing finit e models of arbitrarily large size must admit also some
infinite models. Another reasonabl e question may concern how many models
T admit s in any fixed infinite cardinal A. One can check t hat t heir number
cannot exceed 2\ but this upper bound can be reached, for every A, by some
suit able T's. The opposite case, when T has just one mod el in power A (up
to isomorphism) , will be of some interest in the next chapters; we fix it in
t he following definition.
Definition 1.5.4 Let T be a theory with some infi nite model, A be an infi-
nite cardinal. T is said to be A-categorical if and only it any two models
of T of power A are isomorphic.
We wish to devote some more lines t o t he Lowenhcim-Skol em th eorem.
Among other things, it confirms th at elementary equivalence is a weaker
relation than isomorphism. In fact, t ake an infinite struct ure A, and use
the Lowenheim-Skolem to build a model A' satisfying the same first order
sentences as A but having a different cardinality, It is easily checked that
A , A' are elementary equivalent; but , of course, t hey cannot be isomorphic.
Now recall what we pointed out in 1.2 : the Induction Principle (in it s usual
form) cannot be written in the fi rst order style in t he language for (N, 0, s)
because first order logic forbids quantification on set vari abl es. However,
as far as we know, one might find an equivalent statement that can be ex-
pressed in t he first order set t ing; in this sense, Induction might become a
first orde r statement. Well, the Lowenheim-Skolem theorem excludes this
extreme possibility. For, t he Induction Principle cha racterizes (N, 0, s) up
to isomorphism, while the Lowenheim-Skol em theorem ensures us that any
tentat ive first order equivalent translation (even involving infinitely many
sentences) has some uncountable models. So this translation cannot exist .
The Lowenheim-Skol em t heorem emphasizes other similar expressiveness
restrictions in first ord er logic. For instance, it is well known t hat t he ord ered
field of teals is, up to isomorphism, t he onl y complete ordered field (here
completeness means t hat ever y non-empty upperly, lowerl y bounded set of
reals has a least upper bound, a greatest lower bound respectively). So
completeness cannot be expressed in a first order way, because any tentative
first order translation should be true in some real closed field wit h a non-
cont inuum power.
On t he ot her side, we will see that the Lowenheim-Skolcm t heore m is a
very useful and powerful technical tool in first order model theory (just as
the Compactness Theorem). And actually the expressiveness restrictions re-
marked before are only the other side of the picture of these technical advan-
tages. This is j ust th e content of the Lindst rom theorem quoted before in 1.2.
Indeed, what Lindstrorn shows is t hat, if you have a logic (namely a reason-
able syst em offormulas and truth) and you demand t hat your logic satisfies
the Compact ness Theorem and the weaker form of the Lowenheim-Skolem
Theorem, called Downward Lowenheim-Skolem Theorem, introduced in 1.4
and requiring -for countable languages- th at any set of sentences admitting
a mod el does have a count able model , then your logic is t he first order logic.
In t his sense the first order framework is (Leibni zianl y) the best possible
one .
1.6 Complete theories
Let us deal again now with one of the main themes in Model Theory, Le.
class ifying struct ures in a given class K . Due to our first order setting, we
limit our analysis to elementary classes K = Mod(T), where T is a first
order t heory. This choice is not so partial and na rrow as it may ap pear. In
fact, it certainly includes th e cases when T is explicit ly given and equips
K with an effect ive list of first order axioms, as in the positive examples of
the last section; but it is also conce rned wit h oth er , and worse sit uations.
For instance, think of the t heory T of finite sets, or groups, or fields, or,
in general, of a class of finite arbitrarily large structures, so t hat T has
also infinite models. Alternatively, think of the th eory T of a single infinite
structure A: due to the Lowenheim-Skol em Th eorem, T has some models
non-isomor phic to A. In t hese cases, T is introduced by specifying some
crucial models, but this does not det ermine in an explicit way a priori which
first order sentences belong to T , and which are excl uded; indeed we could
just be interested in finding an effective axiomatization as in the previous
examples, and we could aim both at descri bing T and also -as a relat ed
matter- at classifying its models.
These are th e settings we wish to consider. Actually we should also admit
that we have not clearly explained up to now which kind of classification
we pursue; however we have agreed t hat this classification should identify
isomor phic models but distinguish non isomorphic structures. Also, we have
seen that isomorphic models sat isfy t he same order orde r sentences. So a
pr eliminary classification is j ust up to elementary equivalence, and aims at
distin guishing non elementarily equivalent structures. Once this is done,
we could restrict our analysis to structures satisfying t he same first order
conditions; Le. fix a structure A and classify up to isomorphism t he models
of its theory T = Th({ A}) (by t he way, let us abbreviat e for simplicity
1'h( { A} ) by Th(A)).
Which is an intrinsic syntactical characterization of such a theory 1'? Basi-
cally it is "complet e" according to the following definit ion.
Definition 1.6.1 A (consist ent) theory T of L is sai d to be complete if,
for every sentence <p of L, either <p E T or -' <p ET.
In fact, it is easi ly observed on the ground of the definition of truth in first
order logic that , given a structure A in a language L, for every L-sentence ip,
either <p is t rue in A or -' <p is; equivalent ly, eit her <p E 1'h(A) or -' <p E Th( A).
1.6. COMPLETE THEORIES 31
On t he ot her hand, every complete t heory T ca n be represented in t his way.
In fact , fix any mod el A of T. Clearly T Th(A) . Conversely, let ep be
a sentence of Th (A), t hen -' ep (j. Th(A) and so -' ep (j. T; as T is complete,
ep E T .
Notice t hat t he same arg ume nt shows t hat, if T T' are consistent t heories
and T is cornplet e, t hen T = 'I",
Now not ice what follows.
Rem a r k 1.6.2 Every (consistent) t heory T of L can be enlarged in at least
one way to a complete t heory in L. In fact , it suffices t o consider Th(A)
where A is any mod el of T . A complet e t heory exte nding T is called a
complet ion of T .
So our classification project can be organized as follows.
First , determin e st ruct ures up to elementary equivalence, III other
words find all t he complet ions of a given t heory T;
t hen, class ify up t o isomorphism t he models of a complete T .
We deal in t his section wit h t he former problem, hence wit h complet ions
and, definitively, wit h complete t heor ies. Incomplet e t heor ies are easy to
meet .
Example 1.6.3 For inst ance, t he t heory of groups it is not complete (as
t here are both abelian and nonabelian groups, and commutativity can be
written in a universal first ord er sent ence) . In th e same way, the t heory
of fields is not complete (why?) . Also the th eor y of linear orders is not
complete (as t here are both dense and non dense t otal orders, as well as
ord ers with or without a minimum or a maximum).
On the other hand, the previous remark pointi ng out t hat a compl et e T
is the t heory of any model of T seems to provide a great deal of complete
t heories; but these exa mples are not satisfactory. In fact , as already said,
what we reasonabl y expect is t o have complet e t heories T equipped with
an explicit list of basic axioms, ensuring that t he sente nces in T are just
t he consequences of t hese axioms. Now, when we look at Th(A) for some
struct ure A (t he field of complex numbers, or t he ordered field of reals,
and so on) , t his list of axioms is lacking; indeed we could wish t o obt ain
such a basic axiomatization in t he mentioned cases . A possible strategy t o
solve t hese problems might be t he following. Given A, prepare a t entative
explicit axiornatization and t he corres ponding t heory T. Of course , A should
be a mod el of T . At t his point , check if T is complete, by some suitable
procedures. If yes, T = T h(A ).
Unfortunately, checkin g completeness for a theor y T as before is not sim-
ple. We mention here a celebrated sufficient (but non-necessary) condit ion,
found ed on t he notion of A-cat egor icit y.
Theorem 1.6.4 (Vaught) Let T be a theory of L. Suppose that every model
of T is infinite and T is A-categorical for some infinite cardinal A. Then T
is complete.
Proof. Suppose towards a contradi ction t hat T is not complet e; let <p be a
sent ence such that <p rt T and ' <p rt T. As <p rt T, there exists a (n infinite)
model A
o
of T such th at A
o
1= ' <p. In a similar way, there exists a( n infinite)
model AI of T such t hat Al 1= <p o For every i :::; 1, put T; = T h( A); t hen
T; 2 T , and Ti has an infinite model. By t he Li::i wenheim-Skolem t heorem,
T, has a model B, of power A. Both B
o
and B
1
are models of T , hence
t hey are isomorphic because T is A-cat egori cal. However B
o
= B
1
because
B
o
1= , <p , while B
1
1= <p o
Here are some consequences of Vaught 's Theor em.
Corollary 1.6.5 Let L = 0. Then the theory l oo of infinite sets is compl ete.
Proof. Clearly T has no finit e models. Moreover t wo (infini t e) sets in
t he same power are isomorphic; in oth er words l oo is A-cat egor ical in any
A Consequently l oo is complete .
Corollary 1.6.6 The theory DLO- of dense linear orders without end-
point s is complete.
Proof. DLO- has no finite models; this is easy to check, by using density
or th e absence of endpoints. Moreover th e famous theorem of Cantor on
dense linear orders recalled in 1.3 ensures t hat DLO- is
every countable linear ord er wit hout endpoint s is isomorphic to t he ord er of
rati onals. Hence DLO- is complete .
Recall t hat both (Q, :::;) and (R, :::;) are models of DLO- ; it follows t hat
DLO- = T h( Q,:::;) = T h(R, :::;) .
Corollary 1.6.7 For every p = 0 or prime, the theory ACFp of alge-
braically closed fi elds of characteristic p is compl ete.
1.6. COMPLETE THEORIES 33
Proof. An algebraically closed field lC is always infinite; in fact , if ao, .. . , an
are distinct elements of lC, the polynomial (x - ao) .... . (x - an) + 1 in K [x]
has a root cv in lC; cv cannot equal ao, ... , an, and so is a new element . At t his
point , in order to apply Vaught 's Theorem, we have t o prove '\-categoricity
for some infinite '\. But t his is just a consequence of St einitz's analysis of
algebraically closed fields . For, this analysis essent ially impli es (in our t ermi-
nology) that , fixed p = 0 or prime, t he t heory ACFp of algebraically closed
fields of characteristic p is '\-categorical for every uncount abl e cardina l ,\ (so
Vaught 's Theorem applies and yields cornplet eness) . Let us recall briefly
why (we will provide an alte rnat ive, detailed proof of t he complete ness of
AC Fp in Chapter 2). Any algebraically closed field in characteristic p can
be obtained as
lC = lCo(S)
where lCo is the prime subfield of lC (hence lCo is isomorphic to the rational
field if p = 0, or to the field with p elements if p is prime) , S is a t ra nscen-
dence basis of lC (namely a maximal algebraically inde pende nt subset ), and
- denotes the algebraic closure in lC. Fur t hermore t he isomorphism ty pe
of lC is fully determined by t he cardinality of S (the transcendence degree
of lC). Accordingly, one ca n realize that ACFp has
pairwise non isomorphic count able models (correspondingly to the
t ranscendence degrees 0, 1, ...,
for every uncountabl e cardinal '\ , exactly one isomorphism class of
models of power '\ , because all t hese models share the same t ranscen-
dence degr ee '\.
Hence ACFp is '\-categorical in ever y cardinal ,\ > and consequent ly
complete. ..
In particular, two algebraically closed fields lC
l
and lC
z
having the same
characteristic p but different transcendence degrees d
l
-I d
z
are elementarily
equivalent, but cannot be isomorphic. Hence, when lC
l
and lC
z
are countable,
t hey are not even partially isomorphic.
Notice also that t he field of complex numbers is a model of ACF
o
, and so
AC1"0 equals it s t heory: we find in this way an explicit list of axioms (that
of AC1"0) for the theory of the complex field.
Let us propose a further application of Vaught 's Theorem t o deduce com-
pleteness. Perhaps at t his point someone may expect t o meet RC1" and
the theory of the real field in ou r list of examples. But we have to delay
this appoint ment . Indeed, RCF is complete (and hence equals the theory
of t he ordered field of reals) , but RCF is not A-cat egori cal for any infini t e
ca rdinal A. So a different approach is necessary: we shall follow this new
strategy in t he next chapter. On t he contrary, Vaught 's Theorem applies t o
vectorspaces over a countable field. Let us see why.
Corollary 1.6.8 Let JC be a countable fi eld. Th en the theory KT' of infinite
vect orspaces over JC is complete.
Proof. Clearly KT' has no finite models. Moreover we know that two (infi-
nit e) vectorspaces with the same dimension over JC are isomorphic. Conse-
quently, for every cardinal A bigger than there is a unique isomorphism
class for all the JC-vectorspace of power A (in fact , each of them has dimen-
sion A). In other words, KT ' is A-cat egorical for every cardinal A > By
Vaught 's Theor em, KT ' is complete .
On t he cont rary, KT' may not be categorical in In fact , when JC is infinite,
JC, JC2, .. . , JC{ N
o
) are count abl e JC- vectorspaces wit h distin ct dimensions, and
so cannot be isomorphic, hence t hey are not even par ti ally isomorphic. So el-
eme ntary equivalence cannot impl y parti al isomorphism (and isomorphism).
The reader may check directly what happens when JC is finite .
We conclude t his section by introducing anot her noti on relat ed to complete-
ness. It will be used in Chapter 3 t o show th at RCF is complete. Recall
t hat a complete th eor y T equals Th(A) for every model A, and hence a
t heory T is complet e if and onl y if any t wo models of T a re element arily
equivalent .
Definition 1.6.9 A theory T is model complet e if every embedding of mod-
els of T is elementary.
It is eas y to exhibit t heories which ar e not mod el complete . For instance,
t he pr evious examples 1.3.14 ensure t hat t he t heory of (N , <) , as well as th e
t heory of fields, are not mod el complete . On th e cont rary, it is not simpie t o
give explicit examples of mod el complete t heories. Cha pter 3 will be devoted
t o t his point , and to discussing t he relevance of t his notion within Mod el
Theory.
1.7. DEFINABLE SETS
1.7 Definable sets
35
Formulas include equat ions, and Algebra aims at finding solut ions of equa-
tions. More gener ally, given a language L, a formula <p (v) of L and a struc-
t ure A of L, one could t ry to det ermine all t he sequences ii in A for which
A F <p(ii) . As we shall see in t he next cha pters, t his is not j ust a minor,
collateral exercise; on t he cont rary, in treating th is fr amework , we are mov-
ing t o t he core of modern model t heory. So let us give t he cor responding
definiti on.
Definition 1.7.1 Let A be a structure of L, n be a positive integer. A
subset D of An is called definable in A if there is a formula <p (v) of L(A)
such that D equals the set of the elements ii in An for which A F <p (ii) (n
is the length of ii, of course) .
In t his case one says t hat t he L(A)-formul a <p(v) defines D, and one writes
The elements of A occurring as constants in t he formul a <p (v) are called a
sequence of parameters defining D.
Let us propose a simple example.
Consider a polynomi al P(Xl ' X2 ) E R[Xl, X2 ], and look at t he algebraic cur ve
of t he solut ions of P(Xl ' X2 ) in R
2
. This is a defi na ble set , because it is
formed by t he elements in R 2 satisfying t he formula P(Vl' V2) = 0 (together
wit h t he coeffi cients of P(Xl' X2) -th e paramet ers of t he formul a in R-).
A functi on f having domain An and image As for some positi ve int eger
s is called definabl e when it s gra ph (hence t he set of sequences (ii, b) in An +s
such t hat f(ii) = b) is.
If X A and t he par ameters x in a formula defining D are in X , t hen D
is said t o be X -definable. In part icular D is 0-definable if and onl y if t here
exists a formula <p (v) in L such th at D = <p (An) is the set of t he sequences
of An satisfying <p (v) in A.
For D An ,
(i) D is definable if and only if D is A-definable;
(ii) if X Y A and D is X-definable, t hen D is V-definable, too;
(iii) D is defina ble if and onl y if t here exists a finit e subset X of A such t hat
D is X -definabl e (( <=) follows from (ii) ; in order to show (=}) , just let
X be t he set of t he paramet ers in a defini ng L(A)-formula <p (v)).
An element a E A is X -definabl e if it s singleton is. Of course, every a is
A-definabl e (by v = a). But , when a rJ. X , things are not so t rivial.
Remark 1.7.2 Fi x a struct ure A of L and a positive integer n.
1. Let X be a subset of A. The X-definable subset of An form a subal-
gebra of the Boolean algebra of all t he subset of An.
Inother words, both An and 0are X-definable (by t he formulas VI = VI
and '(VI = vd respectively) , and, if Do and D
I
are two X-definable
subsets of An, t hen even th eir union DoUD
I
, t heir intersection DonD
I
and the complement An - Do are X-definabl e (if <Po(v) and <pt(v) are
two L(X)-formulas defining Do and D
I
respectively, just look at the
formulas <Po(v) V <P I (v), <Po(v) 1\ <P I (v), ' <Po(v)).
Bn(X, A) will denot e below the Boolean algebra of the subsets of An
X -definabl e in A.
2. Definable sets are closed also under projections, in t he following sense.
Let D be a subset of An definable in A. Let Jr be t he projecti on of
A ont o some fixed i ::; n coordinates. Then Jr(D) (a subset of Ai) is
st ill definable. For instance, if t he L(A)-formula <p (v) defines D, then
:3v2" .:3v
n
<p (v) defines th e image of D by its projection onto the first
coordinate (of course, vabridges here (VI, V2 , , v
n
) ) .
3. Every finit e subset of An is definable.
In fact let D be a finit e subset of An, and let . . ., be it s elements .
For every j ::; t, put dj = (djl ' . . . , djn)' Then
V 1\ (Vi = dji)
j9 l:::; i:::;n
defines D in A. In particular, in a finit e struct ure A every subset of
An is definable; moreover , owing to 1, in any st ruct ur e A even the
cofinite subsets of An are definable.
4. A is infinit e, then there exist some subset s of An which are not defin-
able in A.
This follows from a simple cardi nal counting argument . In fact , we
know that there are 2
1A1
dist inct subsets of An, while the subset s of
An definabl e in A cannot exceed the L( A)-formul as defining t hem.
Consequentl y t he definable subsets of An ar e at most IAI (for a count-
able L, of course) .
1.7. DEFINABLE SETS 37
An explicit example of an infinit e struct ure wit h a non-definabl e subset
is t he following. Let L = 0, so t he struct ures of L are just t he non-
empty sets A. Take an infinit e set A. We have seen t hat every finite
or cofinite subset of A is definable. We claim t hat no other subset of
A is definable. In fact, let D be a subset of A such t hat both D and
it s complement A - D are infinite. Suppose t owards a cont radict ion
t hat D is definabl e, and so D = <p (A, b) for a suit able L-formula
<p(v, w) and a sequence bof parameters from A. Take d E D, d' E
A - D, d, d' out of b. We can find a bijecti on f of A onto A (and
hence an automorphism of A) fixing b point wise and mapping d in
d'. As d E D, A F <p( d, b); as isomorphisms preserve sat isfiability,
A F <p (d', b); consequently d' E D -a cont radict ion-. In conclusion, D
is not definable.
Let us concentrat e our attent ion on infinite structures from now on. In fact ,
owing t o Remark 1.7.2 , 3, t here is no point in exploring definabili t y in t he
finit e case . First let us give some more examples of definabl e sets, suggest ing
some int riguing connections between t his part of Model Theory and oth er
br anches of Mat hematics.
Examples 1.7.3 1. (Defi na ble sets and Complex Algebraic Ge-
ometry) Let L = {O, 1, +, " - } be t he language of fields. We have
seen at t he beginnin g of t his section t hat, over t he real field, polyno-
mials det ermine defi na ble sets. Let us investi gat e t his example more
generally and closely. Accordingly consider an arbitrary fi eld J( and a
positive int eger n . Algebr aic Geometry deals wit h algebraic varieti es
in K" : These are t he zero sets in K" of finite systems of polynomi als
Hence algebraic curves are exa mples of algebraic vari eties. Moreover
every algebraic vari et y as before is definabl e in J( , for instance by the
(quantifier free) formul a
1\ qj(v) = O.
j9
So one may wonder how close and deep t his connection bet ween de-
finabl e sets in J( and algebraic varieti es in J( is. Of course, we cannot
expect t hat any definabl e set is a variety (alt hough t his is cert ainl y
t rue when J( is finit e). In fact , owing t o Remark 1.7.2,1 before, ev-
ery finit e Bool ean combination of definabl e sets is also definable. But ,
wit h respect t o t his point , algebraic varieties behave in a different way.
The union of t wo (and consequently of finit ely many) algebraic
vari eties in an algebraic variety. This is a simple exercise of Al-
gebra, essenti ally using the fact that , in a field K , t he product of
t wo nonzero elements is different from O.
The intersection of two, or finitel y many, or even infinitel y many
algebraic vari eti es is st ill an algebraic vari et y. This is a trivi al
exercise in the finit e case, and a deep t heorem in Algebra -known
as Hilbert 's Basis Theor em- otherwise.
Notice that these properti es (together with the easy observation t hat
K" and 0 ar e algebraic variet ies -for, they are th e zero sets of t he
zero polynomial, and of any nonzero constant polynomial in K[x'J
respectively-) show t hat th e algebraic var ieti es of K" ar e the closed
sets in a suitable topology of K" (the Zariski topology) . However
t he complement of an algebraic vari et y of K" is not necessarily
an algebraic vari et y of K" :
So t here are definabl e sets of K" which are not algebraic vari eti es.
Indeed Algebraic Geometry introduces the noti on of construct ible set
t o define a finit e Boolean combination of algebraic variet ies of K":
Remark 1.7.2 ,1 before ensures t hat every constructible set is definable.
In certain fields K the converse is also true, and hence definable j ust
means constructible. For instance, this is what happens when K is an
algebraically closed field (and so, in parti cul ar , when K is t he complex
field). This is not a t rivial resul t , but a deep t heorem of Tarski and
Chevalley, and will be discussed in t he next Chapter .
2. (Defi na ble set s a nd Real Algebraic Geomet r y ) Let L = {O, 1, +,
" - ::;} be our language for ord ered fields. Fix an ordered field K, and
a positive integer n . Algebraic Geometry studies the sets of the ele-
ments of K" satisfying disequations like q(x) 0 where q(x) E I<[x'J,
and calls semialgebraic set any finite Boolean combinat ion of t hem.
It is clear t hat every semialgebraic set is definabl e in K , and even by
a quantifier free formula (a Boolean combinat ion of atomic formulas
q(v) 0) . A t heorem of Tarski and Seidenberg ensures t hat, when K
is a real closed ord ered field (in particular when K is t he ordered field
of reals) , t hen t he definable sets of K" are exactly t hose semialgebraic .
So a close connecti on arises bet ween Model Theory and Algebraic Ge-
ometry also in t his framework. The Tar ski and Seidenb erg t heorem
will be t reated in detail in t he next chapt er.
Notice also t hat the order relation is definable in the real field R
even within t he language of fields {O, 1, +, " - }: in fact, it suffices to
recall that the non negative reals are exactly the squares, and hence to
define
by t he for mul a
:3W(VI - V2 = w
2
).
Consequently every semialgebraic set D in R is definabl e in the real
field even wit hin the language of fields, just by replacing any formula
q(iJ) 0
(with q(x) E R[X]) by t he equivalent for mula
:3w(q(iJ) = w
2
) .
However , notice that the lat t er formula requires a quantifier.
3. (Definable sets and recursive sets) Now we consi der t he language
L = {+, .} and, in L , t he structure (N , +, .). First noti ce t hat every
natural n is 0-definabl e in (N, +, .). This can be eas ily shown by using
an induction argument on n; if n = 0 or n = 1, just take t he formulas
"VI = 0" : VI +VI = VI, "VI = 1" : VI . VI = VI A--, (" VI = 0" )
respectively, while, for n 1, n + 1 is 0-definable by
"VI = n + I" : :3z
0:3
z
1
(" Zo = I " A" Zl = n" AVI = Zo +zt).
Consequently in (N, +, .) defina ble sets just equal 0-definable sets.
Definability in (N , +, .) is deeply related to recursion theory. Let us
see why. A basic aim in recursion theory is to provide a sharp definition
of the notion of elgorithm. According to the Church and Turing thesis,
algorithm means Turing machine,
in t he sense t hat t he problems with a solving algorit hm are just t hose
handled by a Turing machine. Actually t he Church-Turing model of
computation dat es back to t he thirties, and, from the practical point
of view, is undoubt edl y sur passed by the new advances in computer
science. However, as an abstract and theoretic proposal , it is still valid
40 CHAPTERl . STRUCTURES
and commonly agreed, at least in any discrete setting. This is clearly
our framework when we are dealing with natural numbers. Inciden-
tally, recall that every discrete context can be eas ily translat ed int o
the world of natural numbers, owing to the G6del coding procedures,
equipping effectively each element with it s own natural label. So let 's
work wit h the struct ure (N, +, .) . On the ground of the Church and
Tur ing thesis, one can define in a sharp way which are the subsets
D <;;; N " (with n a positive integer) admitting a decision algorithm,
namely a procedure running in finitely many steps and establishing,
for every if E N " ; if if is in D or not. These sets D ar e called recur-
sive. A crucial result in Recursion Theory -indeed a key step within
t he proof of G6del First Incompleteness Theorem- ensures t hat every
recursive set is definable in (N, +, .) . More generally, any recursively
enumerable D <;;; N" is definabl e in (N, +, .). Recall t hat recurs ive
enumerability is a weaker notion t han rec urs iveness, and ju st requires
that there is some algorit hm effectively listi ng t he element s of the in-
volved set D. On the cont rary, there do exist some subset of N " which
are not recursively enumerable (and hence are not even recursive) , but
are definable. These remarks witness that now definable sets ar e a very
complicated class, because t hey inherit t he complexity of t he class of
recursively enumerabl e sets wit h t he cor res ponding int rinsic hierar-
chies, and possibl y more. So, when dealing with (N, +, -} , definable
sets are not so clean as in the complex, or in the real field .
4. (Definable sets and decidable theories) We again work in the
language L = {+, .}, but t his time we examine the structure (Z, +, .).
Ident ify nat ural numbers and non negative integers. Then N becomes
an 0-definabl e subset of Z; in fact , a celebrat ed theorem of Lagrange
ensures t hat, among the integers, t he non negative element s are just
the sums of 4 squares. Hence the formula
defines N inside (Z, +, .). Also the sum and product operations in N
are 0-definable in (Z, +, .), because they just restrict to N the addit ion
and multiplication in Z. So we can conclude that the whole structure
(N, +, .) is " 0-definable" in (Z, +, .).
Now it is well known that the t heory of (N, +, .) is not decidable, in
t he sense t hat the set of the natural codes of t he sentences t rue in
(N, +, .) is not recursive. In other words , no gener al algorit hm can
decide, for every sentence <p di L, if <p is t r ue in (N , +, .) or not. On
t he other hand, t he pr evious obse rvations let us effectively t ranslate
every sentence <p of L in a sent ence <p' of L such t hat
(N , +, .) F <p {:} (Z, +, .) F <p' .
Consequently even t he t heory of (Z, +, .) is undec idable.
T he method sketched here is oft en used t o show und ecidabili t y for
t heories int erpreting as described other t heo ries whose undecidability
is known, or also, spec ularly, t o deduce decidabili t y for t he t heories
that can be interpret ed in the previous way in some other decidable
theory. Hence definability plays a crucial rol e also within the deci sion
problem for theories.
5. (M o d ule s and pp-definable subgroups) Let R a (countable) ring
wit h identit y. Conside r the language LR = {a, +, - , r (r E Rn of
R-modules, and in LR t he class R - Mod of (left ) R -modules. A
formula <p (v) of LR is called a positi ve primiti ve formula (or also,
more synt het ically, a pp-formula) if <p(v) is of t he form
:J -(A t - B t - )
:JW V = . W
where A and B are matrices wit h coefficients in R an d suitable sizes,
. denot es t he usual row-by-column mul tiplicati on for matrices, and t
is t he t ranspose oper ation. Equivalentl y, if one put s v= (VI , ... , v
n
) ,
W= (WI , ... , wm), A = ( r i j)i,j and B = (Sih) i ,h , where j ra nges fro m
1 t o n , h from 1 t o m, and i fro m 1 t o some suitable positive int eger
t , then <p (v) can be wri tten
:JWI ... :Jw
m
1\ ( L r ij Vj = L Si hWh) .

Hence, in every R-module M , <p (Mn) is t he set of t he sequences ii =
(aI, ... , an) in M" for which t he linear system
has some solution in M m.
Let us examine some part icul ar cases.
42 CHAPTER 1. STRUCT URES
Let r E R, <p (v) : rv = O. Then <p (v) is a pp-formula (t ake an
empty W, or wri te <p (v) in the equival ent for m :3w(r v = Ow)) . For
every R-module M ,
<p (M ) = {a EM: ra = O}
is just the annihilat or of r in M , hence a (n addit ive) subgroup
of M , and even a submodule of A1 when R is commutative, or
when, simplerly, r is in the centre of R.
Take again r E R, and now consider <p (v) : :3w(v = rw). Then
<p (v) is a pp-for mul a (for which A e B?). If M is an R-module,
then
<p (M ) = rM
is a subgroup of M , and even a submodule at least when R is
commutative, or, simplerly, when r is in t he centre of R .
In general it is easy to check that , for ever y pp-formula <p (v), <p (Mn) is
an addit ive subgroup of M n; on the other hand, <p (Mn) is not always
a submodule, alt hough this is certainly true when R is commut at ive,
as we saw in the previous examples. <p (Mn) is called a pp-definable
subgroup of M" : Every coset D of <p (M
n
) in M" is definable in A1:
in fact , let if be any element in D, t hen t he formula <p (v- if) defines
t he whole coset D.
A t heorem of Baur and Monk ensures that , in every R-module M , any
defi nable set s is a finite Boolean combination of eosets of pp-defina ble
subgroups (see Chapt er 2).
1.8 References
There exist several excellent handbooks providing t he backgrounds of Math-
emat ical Logic necessary in this chapter. Among them, let us mention once
again Shoenfield [153], but also Malitz [103] or Ebbingha us-Flum-Thomas
[37]. The key reference for basic Model Theory is [18]. We also refer t o
Devlin [31] for Set Theory, t o Odifreddi [121] for Recursion Theory and to
Jacobson [65] for Algebra. [173] is the historical sour ce quot ed at t he end of
1.2. Ax' s analysis of pseudofini t e fields is given in [3]
Chapter 2
Quantifier Elimination
2.1 Elimination sets
Let L be a language. It may happen that two different L-formulas ip(v)
and ip'(v) admit t he same meaning in a structure A of L, or in a class
of L-structures, for inst ance among t he models of a given L-theory T .
For example, in t he ordered field of reals (and even in every real closed
field), t he formula ip(v) : v 0 (being nonnegati ve) is t he same t hing as
ip'(v) : :Jw(v = w
2
) (being a square) . Simi larly, in the ordered domain
of int egers, ip(v) : v 0 (being posi tive) has t he same int erpret at ion as
ip' (v) : :Jwt3w
2:J
w
3:J
w4 (v = L 1<i<4wr) (being the sum of four squares):
this is a celebrated t heorem of Lagrange, already mentioned in the last chap-
te r.
So, fix a consistent, possibly incomplete t heory T in a count able L . We shall
say t hat two L-for mulas ip( v) and ip' (v) are equivalent wit h respect to T,
and we shall write ip(v) rv T ip' (v), when
\fv(ip( v) +7 ip'(v)) E T ,
equivalently when
for all models A of T.
The notion of elimination set arises qui t e nat urally at t his point. An elimi-
nation set for T is a set F of L-formulas such that every L-for mula ip( v) is
T-equivalent to a suitable Boolean combination of formulas of F.
Clearly the set of all the L-formul as is an elimination set for T . But, of
course, t his is not an interesting case, and we reasonably expect simpler sets
43
44 CHAPTER 2. QUANTIFIER ELIMINATION
F. In particular, when the set of atomic formulas in L is an eliminat ion set
for T, we say t hat T has t he quantifier elimination in L. In detail
Definition 2.1.1 Let T be a theory in a language L. T has the elimination
of quantifiers (q.e.} in L if and only if every formula <p( iJ) of L is equivalent
in T to a quantifier free L-formula <p' (iJ) (so to a finit e Boolean combination
of atomic formulas) .
One easily realizes t hat every T gets t he elimi nation of qu antifiers in a
suit able language extending L. In fact , put L = La, T = To, and enlarge La
to a lan guage L
1
cont aining an n-ary relation symbol R<p for every for mul a
<p (iJ) of La (n is the lengt h of iJ, of course) ; t hen add the following sente nces
t o To
for every <p( iJ), and get a new t heory T
1
; it is clear t hat t he atomic for mulas
of L
1
form an elimination set in T
1
for t he formulas in La. By repeating t his
procedure countably many t imes, one event ually defines a language L' 2 L
an d a t heory T' of L' " nat ur ally" extending T and having t he elimination
of qu antifiers in L' .
Unfort unately t his procedure has a qui t e artifici al and abst ract flavour. In-
deed, what we would like to obt ain , given a theory T in a language L, is
showing t hat T has t he qua nt ifier eliminat ion directly in L or , otherwise,
determining a smallest ext ension L' 2 L, possibl y suggeste d by t he algebraic
analysis of t he models of T , wher e T (or, more exactly, it s natural extension
to L' ) has t he elimination of qu ant ifiers, or also a reasonably simple elimi-
nation set of form ulas , in L'. In fact, t here are good reasons t o believe t hat
such a language L' is, in some some, "the" proper language of T .
Which are the main advant ages of an elimination set, in part icular of quan-
tifier eliminat ion? They concern several applications.
1. The main one (at least from a historical poin t of view) is decidability .
Act ually t he first and most celebr ated quantifier elimination results
are relat ed to t he decision t heme. Let us explain why. Recall t hat a
t heory T is decidabl e if t here is an algorit hm checking in finit ely many
steps, for ever y sente nce 0:' in t he language L of T, whet her 0:' is in T
or not . Now suppose t hat F is an elimination set for T and that the
following are available:
an effecti ve procedure translating any L-sentence into a T-equiv-
alent Boolean combinat ion of sente nces in F (or even an effective
2.1. ELIMINATION SETS 45
reduction of any L-formula into a T-equivalent Boolean combi-
nation of formulas in F);
an algorithm to decide, for ever y Boolean combination 0' ofsen-
tences of F, whet her 0' is or not in T.
Then, clearly, T is decidable, and act ually we have got a decision
algorit hm (by successively applying t he pr evious t wo procedures).
2. Another noteworthy application of quanti fier elimi nation conce rns de-
finabilit y. In fact , if F is an elimination set for T , t hen the definabl e
sets of a mod el A of T reduce to
where <p (iJ, ill) is a finite Boolean combination of formulas of F and
xE A; in particular , if T has the quantifier elimination in L , then the
definabl e sets of A are just the ones of t he form
where <p (iJ, ill) is a quantifier free for mul a and xin A.
3. A third application regards the classification of complet ions of T . Re-
call th at T is possibly incomplete; but we know t hat T has some
(non-unique!) complete ext ension in L. So we are led to conside r th e
pro blem offi nding all t he complete extensions ofT in L , in other words
classifying t he isomorphism classes of model s of T up to elementary
equivalence. Now, if A and B are t wo models and A is not element arily
equivalent to B, t hen there is some sente nce <p in L such t hat A F= <p
and B F= -- <p. As F is an elimination set for T in L, we can ass ume
that <p is a Boolean combination of sente nces in F. Indeed, one easily
realizes that one can choose <p directly in F.
For inst ance, we will see in this chapter t hat t he theory ACF of
algebraically closed fi elds has th e quantifier elimination in L = {+, "
- ,0 ,1}. Consequent ly t he classification of algebraically closed fields
up to elementary equivalence depends on t he quant ifier free sente nces
in L , which are of the form m = n , where m and n are integers
(m abbreviates in the previous formula the addit ion of m summands
equal to 1 if m > 1, and - (- m) if m < -1; similarly for n). This
implies that the complet e extensions of ACF are fully determined by
t he characterist ic of t heir models , and hence coincide wit h t he theories
ACFp where p is 0, or a pri me.
4. Finally, let us deal with model completeness. Assume that T has quan-
tifier elimination in L. We claim that, in this case, every embedding
between models of T is elementary, in other words T is model com-
plete. In fact, let A and B be models of T, f be an embedding of A
into B. Given a formula <p(v) in L, let <p'(v) a quantifier free formula
equivalent to <p( v) in L. Take d in A. As f is an embedding,
A 1= <p'(ii) {:} B 1= <p'(J(ii)).
As VV(<P(V) +-+ <p'(v)) ET,
A 1= <p(ii) {:} B 1= <p(J(ii)).
Hence f is elementary.
This chapter is devoted to illustrating several key examples of quantifier
elimination, starting from the earliest (Langford's results on discrete or
dense linear orders) to include those perhaps most classical and celebrated
(Tarski's elimination procedures for the real and the complex fields). We
shall treat other eliminations sets as well (most notably, Baur-Monk's pp-
elimination theorem for modules over a given ring).
These examples will lead us to introduce two basic notions in Model Theory,
strong minimality and o-rninirnality respectively. We shall discuss them at
the end of the chapter. The final section will be devoted to some computa-
tional aspects of the quantifier elimination procedures.
It should be underlined that the interest in quantifier elimination arose sev-
eral years before the official birth of Model Theory. In fact it was at the
beginning of the twentieth century that Lowenheim and, later, Skolem pro-
vided some procedures translating formulas into a simpler form avoiding
quantifiers (they are, more or less, the artificial method we sketched at the
beginning of this section). Moreover, the earliest explicit examples of quanti-
fier elimination in some specific algebraic structures treat discrete and dense
linear orders and date back to the twenties (they were obtained by Langford
in 1927). In these results, as well as in Tarski's theorems, the major em-
phasis seems to be on decidability: the elimination of quantifiers is a step
towards decidability, just as described before. But over the years this em-
phasis on decidability reduced and was replaced by an increasing interest in
definability. Actually, definability is the main theme where Model Theory
and quantifier elimination meet.
2.2. DISCRETE LINEAR ORDERS
2.2 Discrete linear orders
47
We begin here our analysis of quantifier eliminable t heories. First we t reat
infini t e linear ord ers. Accordingly our basic language is L = {:s;}. More
pr ecisely we deal wit h:
t heories of discret e linear ord ers (in t his section) ,
t heories of dense linear ord ers (in t he next one).
As already said, the quantifier eliminat ion results in t hese cases were firstl y
shown by Langford in 1927; Tarski pursued t he analysis t o get decidabilit y
and to classify th e complete theories of infinite discret e and dense total
orders.
Recall t hat a (n infinit e) linear ord er A = (A, :S; ) is discrete if and onl y if
(i) 'Va E A, if there is some a' E A such th at a < a' , t hen t here exists a
least b E A for which a < b (b is called t he s uccessor of a and is denoted
s(a));
(ii) 'Va E A, if t here is some a' E A such t hat a' < a, t hen t here exist s
a maximal b E A for which b < a (b is called t he predecessor of a;
obviously a = s(b)).
Accordingly we ca n dist inguish 4 classes of (infinite) discret e linear orders:
1. t he class of discret e linear orders wit h a least , but no last element (like
(N, :S;));
2. t he class of discret e linear ord ers with a last , but no least element (for
instance, N with respect to t he relation reversing it s usual order);
3. t he class of (infinit e) discrete linear orders with both a least and a last
element (like the disjoint union of two discret e linear orders (A, :S; ),
(E , :S; ), the former in 1, th e latter in 2, with a < b for all a E A and
s B );
4. t he class of discrete linear orders without endpoints (like (Z, :s;)) .
Each of t hese classes is elementary. Moreover one can show t hat its t heory
has t he eliminat ion of quantifiers in a suitable language extending L, and is
complete even in L. Here we limit ourselves to prove, for simplicity, t hese
resu lts in t he case 1, in other words for discret e linear orders wit h a least
but no last element .
Accordingly consider the language L' = {::; , 0, s} where 0 is a constant (to
be interpreted in t he least element) and s is a 1-ary operation symbol (to
be interpreted in the function mapping any element into it s successor).
It is easy to write down a first order set of axioms for our class in L'. Let
dLO+ denote th e corresponding t heory. By t he way, notice that suitable
formul as in the restrict ed language L define th e minimal element and t he
successor function in every model of dLO+. This implies that the axioms of
dLO+ ca n be rewrit t en also in L , at t he cost of some more complicat ions (and
quantifiers). For inst ance, expressing the existence of a minimal element
requires the L-sentence
:3wVv(w::; v)
instead of
Vv(O ::; v).
But we momentarily prefer t o treat dLO+ in L'. Observe th at:
(N,::;, 0, s) is a model of dLO+;
if A is another mod el of dLO+ , then A contains a substructure ({s" (OA) :
n EN},::; , OA, sA) isomorph ic to a (N,::; , 0, s), and moreover some
furt her copies of (Z, ::; , s) (as OA is t he only element wit hout any pre-
decessor) .
Theorem 2.2.1 dLO+ has the elimination of quantifiers in L' .
Proof. Take a formula 'P( v) in our language L' ; we look for an equivalent
formul a 'P' (v) without quant ifiers.
Our first ste p is to show t hat we can ass ume that 'PCv) is of th e for m
:3w 1\ ai( v, w)
i<r
where each ai(v, w) is an at omic formula, or it s negation, and w actually
occurs in ai ( v,w) for ever y i ::; r .
We wish to underline that this step is qui t e general, and does not depend
on our part icular language L' . Let us see why. First of all, we can tacitly
ass ume that 'P(v) is of the for m
where the Qj's (1 ::; j ::; m) denote quanti fi ers Vor B, a(v, w) is a quantifier
free formul a, and even a disjunct ion of conj uncti ons of atomic formul as and
2.2. DISCRETE LINEAR ORDERS 49
negations (and wabridges (Wl, . . . , W
m
), of course) . Th e st rat egy at this
point is first to eliminate Qm, and then to repeat the procedure and remove
t he quantifier st ring completely. We recall t hat V is equivalent to -,:3-, and
consequently agree that it is enough to deal with the case when Qmis :3,
namely with
:3wa(v, w)
where a is a disjunction of conjunctions of atomic formulas or negations, W =
W
m
and vis possibly enlar ged to include Wl , ... , W m -l ' As :3 is distributive
with respect t o V, namely :3w(a' V a" ) is equivalent to (:3wa' ) V (:3wa" ),
there is no loss of generality for our purposes in assu ming that a is just a
conj unction of atomic formulas or negations. In conclusion we ar e dealing
with
:3w 1\ ai(v, w)
i<r
where each ai(v, w) is an atomic formul a, or its negation. vVe can also
assume t hat W actually occurs in ai(v, w) for every i r: oth erwi se let
j r deny t his condit ion and notice t hat our formula
:3w 1\ ai(v, w)
i<r
is equivalent to
aj(v) 1\ :3 w 1\ ai(v, w);
i =h
at t his point it suffices t o eliminate t he quantifi er :3 in the lat t er part of t he
formul a
:3 w 1\ ai(v, w).
i =j:j
This completes our preliminary st ep. As alr eady said, this does not depend
on our particular framewor k.
Now let us wor k with our formula
:3w 1\ ai(v, w)
i <r
and our language L' . We wonder which is the form of any ai. A look at L'
shows that ai is t = t' , or t t' , or the negat ion of one of t hese formul as,
where t and t' are terms in 0, w, v (and w actually occurs in tor t'). Recall
th at -, (t t') means t > t', and so on. Deduce that ai is, with no loss of
generality, eit her t = t' or t > t', with t and t' as before. Notice t hat t and
t' are of the form sP(11,) where p is a nonnegative integer and 11, ranges over
0, w, if. Now recall that s is injective and deduce that the formula under
exam ensures that there is a solution w for a finite set of conditions saying
that w or a successor sq(w) (q a non negative integer) is equal, or bigger, or
smaller than a term sP(11,) where p is, again, a nonnegative integer and 1J,
ranges over w, 0, if: as s is injective, we can assume that, in each of these
equations and inequations, s occurs only in one side (on the left, or on the
right). Our aim is to translate this formula into an equivalent one avoiding
wand simply stating quantifier free conditions on if (and 0).
To obtain this, proceed as follows. Trivialities like w = w or w < sP(w) for
a positive p can be ignored and deleted (they can be preliminarily listed and
hence are easily recognized); if nothing else occurs, then replace the whole
formula by 0 = O. On the contrary, when meeting a condition that cannot
be satisfied by any w, like w = sP(w), or 0 = sP(w) for a positive p (also
these conditions can be preliminarily listed), then replace our formula with
-.(0 = 0) (or with -'(Vi = vi) if you like and if is not empty).
Otherwise, as soon as you meet one equation like w = SP(Vi), delete wand
:3, and replace w with sP(Vi) throughout our formula. Proceed in the same
way if an equation w = sP(O) occurs. Similarly, when meeting a condition
sq(w) = Vi, consider any further occurrence of w in t he formula and, again
using the injectivity of s, represent it as sql (w) for a suitable nonnegative
integer q' 2:: q;finally delete wand :3 and replace each occurrence sql (w) by
sq-q'(Vi).
At last, assume that only disequations occur. Again using the injectivity
of s, we can suppose that all of them concern the same term sq(w) in w.
So our formula states that sq(w) is smaller that certain terms to, ... , th
in 0 and if , and larger that some other terms th+i, ... , tk. We obtain an
equivalent formula avoiding wand its quantifier in the following way. List
(in a suitable disjunction) all the possible orderings of to, ... , tk in ::; ac-
cording to which to, ... , th precede th+l , ... , tk; for every ordering, let t ,
t' denote respectively the greatest element among to, ... , it. and the least
among th+l, ... , tk; add s(t) < t' (in order to provide sq(w) with suitable
room).
This concludes the elimination procedure. ..
Corollary 2.2.2 dLO+ is model complete (in L') and complete (both in L'
and L).
Proof. Clearly dLO+ is model complete in L'. Moreover (N,::;, 0, s) 1=
dLO+, and (N, ::;,0, s) is embeddable in every model of dLO+. As dLO+ is
2.2. DISCRETE LINEAR ORDERS 51
model complet e, all t he corresponding embeddi ngs are elementary. Accord-
ingly all t he models of dLO+ are elementarily equivalent t o (N, ::;, 0, s) and
hence t o each oth er. This shows t hat dLO+ is complete in L' . Now recall
t hat both the minimal element and t he successor function (so t he int erpre-
tations of t he symbols in L' - L) are 0-definable by L-for mul as in t he models
of dLO+. Then it is an easy exercise t o deduce t hat dLO+ is complete in
L, t oo. ..
Corollary 2.2.3 dLO+ is decidable (in L' and in L).
Proof. Reduce any sente nce of L' into an equivalent quantifier free state-
ment. This is a Boolean combinat ion of formulas sm(o) sn(o) where m
and n are non-negative integers, and dLO+ ca n easily check it s membership.
This procedure works even for L-sent ences. ..
Corollary 2.2.4 Let A 1= dLO+. Th e subsets of A defi nable in A (in L or
in L') are j ust the fin it e unions of (open or closed) intervals in A (possibly
having +00 as a right endpoint) .
Proof. Let cp(v , w) be a I-formula. As dLO+ has t he elimination of quanti -
fiers in L' , we can ass ume t hant cp (v, w) is quanti fier free; owing to Theorem
2.2.1 (and its proof) , for every ii in A, cp (A, ii) is a union of int ersecti ons of
int er vals, and hence a union of int ervals. ..
This accomplishes our analysis of discrete linear ord ers wit h a last but no
least elements. How to deal wit h t he other t hree cases of infinit e discret e
linear ord ers list ed before? They can be handled in a similar way t o get
quantifier eliminat ion and consequent ly complet eness. In particular it turns
out that th e four cases exha ust all the possible complet ions of the th eor y
of infinite discrete linear orders; in other word s, these complet ions are fully
characterized by saying if t he corres ponding models admit or lack a least
and a greatest element.
Act ually t he case without endpoints deserves some more comments . In
fact , in t his fra mework, t he enlarged language L' needs no natural "con-
stant " symbol (just because endpoint s are lacking), and takes t he only addi-
t ional operation symbol s . Accordingly, properl y spea king, t he elimination
of quantifiers fails in t his extended language, because we have no constant t o
buil d atomic sentences . For inst ance t he (t rue) sentence 3w(w = w) ca nnot
be t ranslated into an equi valent quantifier free sentence; the same happens
for t he (false) sentence 3w(s(w) = w). So t he right statement here is as
follows: an elimination set for the theory of discrete linear orders without
end points is the set of atomic formulas plus a unique sentence (such as "there
is no least element", or " t here is no last element"). We do not discuss the
proof here. In fact, we shall treat this case in detail when considering dense
linear orders in the next section.
Finally, notice that decidability can be shown (in L) in the 4 possible cases.
Consequently the (incomplete) theory dLO of infinite discrete linear orders
is decidable, too; in fact , a sentence <p of L belongs to dLO if and only if it
is in each of it s 4 completions.
2.3 Dense linear orders
Now we deal with dense linear orders. The plan here is exactly the same
as in the discrete case. We use the language L = and we distinguish 4
possi ble cases:
1. there is a least element , but no last element (just as among non-
negative rationals with respect to the usual order);
2. there is a last element, but no least element (now non-positive rationale
provide an example);
3. there are both a least element and a last element (look at the rationale,
or even at the reals, in the closed interval [0, 1]);
4. there are no end points (this is the case of (Q,
In 1, 2, 3 one shows elimination of quantifiers in a language with one or
two additional constants to be intepreted into the endpoints; 4 deserves a
more specific treatment, because quantifier free formulas need an auxiliary
single L-sentence to form an elimination set (even in L): we provide full
details below. In all these cases it is easy to deduce completeness in L.
This implies that these 4 classes exhaust all the possible completions of the
theory of dense linear orders.
As already said, here we limit our analysis to dense linear orders without
endpoint s. We just met their theory in Chapter 1; we called it DLO- and we
observed that it is hence complete. 'Ve treat now quantifier
elimination (in L), and in this way we provide an alternative and detailed
proof of its completeness.
Theorem 2.3.1 The quantifier free formulas of L together with a single
sentence of DLO- (such as 3v(v = v)) are an elimination set of DLO-.
2.3. DENSE LINEAR ORDERS 53
Proof. We follow th e same approach as in th e discrete case. But now t he
successor symbol does not make sense, our language is smaller and hence
our setting is simpler: L-terms are just variables (no constant arises because
there are no endpoints) . Accordingly what we have to do is to eliminate the
quantifier in a formula
3w 0'(ii, w)
where 0'(il, w) is a conj unct ion of condit ions saying t hat w is equal, or
smaller, or lar ger t ha n some v in il. To obtain t his, proceed as follows. Again
ignore trivialities like w = w (th ey can be preliminarily listed and easily rec-
ognized) ; if nothing else occurs, just repl ace our formula with 3v( v = v).
On the cont rary, when meeting a condit ion that cannot be satisfied by any
w, like w < w (also th ese negative st ate ments ca n be prelimin arily listed) ,
repl ace our formul a with ,3v(v = v) . Otherwise, as soon as you meet one
equation w = Vi, delet e wand 3, and replace w wit h Vi t hroughout our
formula. At last , if only disequation s occur and hence our formula states
t hat w is smaller t han certain variables (VI , .. , Vh wit h no loss of general-
it y) and larger t hat oth ers (Vh+l , ... , Vk), t hen get t he required quantifier
free formula in th e following way. List (in a suitable disjunction) all t he
possible ord erings of VI, .. , Vk in :::; according to which VI, , Vh pr ecede
Vh +l , ... , Vk ; for every ordering, let v, v' denote respectively t he maximal
element among VI , . .. , Vh and the least among Vh+ l , ... , Vk ; V < v' and th e
densit y ass umpt ion are sufficient t o ensure t hat an intermedi ate w exist s.
Wh en h = 0 or h = k, one uses t he lack of end points.
This conclu des t he elimination procedure. ..
Corollary 2.3.2 DLO- is model complete and complete.
Proof. Model complete ness is a st raight forward consequence. Complete ness
can be deduced as follows, using model complet eness. First notice that any
two dense linear ord ers with no endpoints (A, :::; ) and (B, :::; ) embed in a
common exte nsion (for inst ance, t heir sum (A +B, :::;) , where A +B is t he
disjoint union of A and B, :::; enlarges t he orderings in A and B and , in
addition, satisfies a < b for every choice of a E A and b E B ). As DLO-
is model complete, each of t hese embeddings is elementary, in parti cul ar
(A, :::;) and (B, :::;) are elementarily equivalent t o t heir sum, and hence to
each other. ..
As already recalled, complet eness was also observed in the pr evious chapter
via and the Vaught crite rion. By th e way, notice that th e
54 CHAPTER 2. QUANTIFIER ELIMINATI ON
Vaught Theorem provides a complete ness pro of even when endpoin ts arise.
In fact, also t he remaining classes of dense linear orders (wit h least and/or
last element) have an t heory.
The decidability of DLO- can be easi ly shown. Indeed, by proceeding as
in t he discrete case , one sees t hat even t he t heory of arbit rary dense ord ers
(with or wit hout endpoint s) is decidable.
Now let us deal wit h definabili t y.
Corollary 2.3.3 Let A F DLO-. The subsets of A definable in A are
j ust the finite unions of (open or closed) int ervals, possibly with infin ite
endpoints.
Proof. Proceed as for dLO+. ..
2.4 Algebraically closed fields (and Tarski)
Tarski obtained his celebrated quantifier elimin ati on procedures for t he com-
plex field and t he ord ered field of reals in t he t hirties. Owing to t he stop due
to t he World War, he published his results only in 1948. We cons ider here
t he complex case, and we delay t he real one to the next sect ion. We should
underline t hat Tarski dealt wit h theories of single st ruct ures (t he complex
field , t he ord ered field of reals) rat her t ha n on axiomatizable classes (ACF ,
RCF) . But a careful analysis of t he proofs singles out which kind of alge-
braic condit ions are necessary to ensure t he quantifier elimination result: so
one realizes t hat what makes t he machinery wor k is just being algebraically
closed in t he complex case, and t he int ermedi at e value pro perty for polyno-
mials in t he real case. This is a crucial result , specially towards t he aim of
findin g a nice axiomat ization for t he t heory of t he complex field, or of the
ordered field of reals.
As promised, her e we consider t he complex case, but we prefer an approach
dealing wit h t he whole class of algebraically closed fields.
Theorem 2.4.1 (Tarski) The theory ACF of algebraically closed fi elds has
the elimination of quantifiers in the language L = {+, " - , 0, I}.
Proof. Take a formul a <p (if) of L, we are looking for an equivalent quant ifi er
free formula <p'(if) . As before, we can limit our analysis to t he case when
<p (if) is of t he for m
3w Q'(if, w)
2.4. ALGEBRAICALLY CLOSED FIELDS (AND TARSKI) 55
where a(V, w) is a finite conj unct ion of at omic for mulas and negations, all
containing ui. In our language, at omic form ulas are j ust equalit ies of t erms,
hence equat ions. Using - , one can express each of t hem as
p(v, w) = 0
where p(if, x) is a polynomial with integer coefficients. Accordin gly <p (v) is
where the Pi'S and th e qj' s are polynomials with integer coefficient s, all
having a positi ve degree, ti; and mj respectively, in x, hence with respect t o
w.
Basic field t heory tells us that a sequence of elements in a fi eld excludes 0
if and only if it s product is not O. Accordingly, we can ass ume t hat at most
one inequation occurs in <p(ii, w), say
-,(q(v, w) = 0).
where q(if, x) is the pr oduct of t he polynomials qj (if, x) when j ::; h; let m
denote the degree in x of q(if, x).
At this point one might wonder whether we can reduce t he number of equa-
tions in our formul a <p( v, w) to get at most a single equation. This is t rue,
and can be shown by using again pure field theory (so without appealing to
algebraic closure ), but requires some more subtlety. The idea is that, for a
given field K and a sequence b in K, t he common roots of t he polynomials
Pi (b, x) ar e just t he roots of their greates t common di visor, and t hat t here is
a quantifier free formula in v, defining the coefficient s (in x) of this greatest
common divisor, and independent of J( and b. The former claim is clear.
Let us explain the det ails of t he latt er.
Consider Pi(if, x) for i ::; k. For every i , write Pi (if, x) as a polynomi al in x
with coefficients in Z[YJ
Pi(if, x) = L Pi,r(f!) X
r.
r :::;ni
Take two of t hese polynomials, for instance Po and PI , and suppose for
simplicity no nl . We clai m t hat t here is quantifier free formul a in v
yielding, whenever PI (v, x) is not t he null polynomial (in x), the coefficient s
in x of the quotient and t he remainder of the di vision of Po (v, x ) by PI (v, x ).
To get this formula, just follow the usual divi sion procedure for polynomials.
This is a tedious but st raight forward exercise. For inst ance, t he first ste p is
t o wri t e that eit her
PI , n l (v) = 0
or the coefficients of Po(v, x) and P I (v, x) satisfy
where P(v, x) is a polynomial of degree < no in x, and, in t he lat t er case,
t he requi red quoti ent admits

nO -nl
PO, no V PI,nl V X
as a coefficient of maximal degree.
At t his point , recall t he Euclidean algorit hm of repeated di visions, yielding
the greatest common divisor (in x) of our polynomi als Pi (ij, x) (i ::; k) as t he
last nonzero remainder in a finit e sequence of successive divisions. Again, a
suitable quantifier free formula in vdetermines the coefficients in x of our
greatest common divisor , whenever t he Pi(V, x) 's (i ::; k) are not all zero .
In concl usion, we can ass ume that our formul a <p(v, w) has one of t he fol-
lowing t hree forms:
1. :3w(p(v, w) = 0),
2. :3w--, (q(v, w) = 0),
3. :3w(p(v, w) = 0 1\ --,(q(v, w) = 0))
where p and q are as before .
First cons ider 1. In any field, 1 is equivalent to say t hat, if v annihilates
all t he coefficients of p(fl, x) in x of posit ive degree in x, t hen v assigns t he
value 0 also t o th e term of degree 0 in x ; this can be writ t en as a suitable
quantifier free for mul a in v.
Now consider 2. In any infinite field, 2 is equivalent t o say t hat vdoes not
an nihilate t he coefficients of t he polynomial p(fl, x) in x. Again, the lat t er
statement can be expressed by a quantifier free formula in ii.
Finally let us deal with 3. We claim that , in any algebraically closed field,
3 is equivalent to the statement
(*) p(v, x) does not divide q(v, z}";
recall t hat n is the degree of p(fl, x ) wit h respect to x, and not ice t hat (*) ca n
be expressed as a quantifier free formula in v (just use t he previous remark s
about divisibility, and write that t he remai nder of the division in (*) is not
0). The dir ection from left to right is true in every field K : if, for a given
sequence bin K , the annihilator of p(b, x) is not included in t he annihilator
of q(b, x), then p(b, x) cannot divide q(b, x) and (q(b, x) )n. Conversely, take
K and bas before; assume t hat every root of p(b, x) annihilates q(b, x), t oo;
for K algebraically closed, this implies t hat every linear fact or of p(b, x)
divides q(b, x) and hence that p(b, x) divides q(b, x)n.
This accomplishes our proof.
Now let us comment t his quant ifier elimination result , and propose some
noteworth y consequences. First of all , we want to emphasize t hat the quan-
t ifier elimin ation property charact erizes t he algebraically closed fields among
infinite fields. In fact , it is a profou nd result of Macintyre, McKenna e Van
den Dries that an infinite field whose t heory eliminates the quantifiers in t he
language L = {+, - , " 0, I } must be algebraically closed.
An obvious consequence of quantifier elimin ation is the following.
Corollary 2.4.2 ACF is model complete.
Clearly ACF is not complet e. In fact , for every prime p, t he sentence p = 0
is true in every algebraically closed field of characterist ic p and false in
ever y algebraicall y closed field of characteristic -# p. However , as we alr eady
showed in Chapter 1,
Corollary 2.4.3 For every p = 0 or prime, the theory ACFp is complete .
Proof. In Chapter 1 we provided a proof founded on Vaught 's Theorem.
An alt ernative approac h, using quantifier elimination (indeed model com-
plet eness) , is t he following. Fix p. There is a minimal algebraically closed
field IC
p
of characterist ic p: t his is the algebraic closure of t he prime subfield.
IC
p
is embeddable in every algebraically closed field of the same character-
istic. Owing to the model completeness of ACF
p
, all t hese embedding are
elementary. In particular, all t he algebraically closed fields of characteristic
p are eleme ntarily equi valent t o IC
p
, and consequentl y to each other. ..
As we already observed in sect ion 2.1, the theories ACFp exhaust all t he
possible completi ons of AC F in L when p ranges over t he primes and 0
(furthermo re, each of t hem has the quantifier eliminat ion in L , j ust because
it ext ends ACF).
An applicat ion of the Compactness Theor em let s us say even mor e. In fact ,
we have seen t hat t he theory of the complex field is just ACF
o
, and so is ax-
iomatized by ACF and, in addition, by t he infinit ely many sentences stating
-,(p = 0) for every prime p. Let (1 be any sentence in ACFa. Compactess
tells us t hat (1 is a consequ ence of ACF and finit ely many sentences con-
cerning the characterist ic. Hence (1 is true in every algebraically d osed field
of prime characterist ic p for all but finit ely many p' s. So we have shown t he
following res ult.
Theorem 2.4.4 Let (1 be a sentence of the language L . (1 is true in some
(equivalently every) algebraically closed field of characteristic 0 if and only
if (1 is tr ue in some (equivalently every) fiel d of characteris tic p for all but
finitely many primes p.
Hence what is true in the complex field (and in any algeb raically closed
field of characteristic 0) is satisfied by the algebraically dosed fields of char-
acteristic p for almost all primes p. We'll see later in this section a nice
application of this model theor etic transfer principle to Algebra.
Now let us deal bri efly with decision problems. As already said, decidability
follows in a very simple way from quantifier eliminat ion.
Corollary 2.4.5 Th e theory ACF of algebraically closed fi elds is decidabl e.
Proof. It suffices to decide if a given quantifier free sentence (1 of L is in
ACF or not . With no loss of generality, (1 is a conjunction of disjunctions
of atomic sentences and negat ions. As a conjunction is in a theory if and
onl y if each conj unct is, we can write (1 (up t o equivalence, using -) as
(Vm; = 0) V (V-,(nj = 0)).
z j
where the m;'s and th e nj' s are positive int egers. So our sent ence just says
t hat the characteristic divides f1 m; or is coprime with some nj (or suitable
variants when no equation, or inequation arises). This ca n be easily checked
in the fixed framework. ..
We shall add some more comments about t he decidability of ACF in the
last section of this chapter.
Now let us deal with definability. We have seen in Ch ap ter 1 that , in any
fi eld /C, const ructible set s (in particular algebraic vari eties) are definabl e.
Theorem 2.4.1 implies that , within algebraically d osed fields , the converse
is also true.
Corollary 2.4.6 In an algebrai cally closed field /C, for every positiv e integer
n , a subset of K " is definable if and only if is consiructible.
Proof. It suffices to show t hat, if X K " is definable, then X is con-
structible. Let <p (V, ill) be a formula of L and ii be a sequence in I<satisfying
wher e n is the length of V. Using qu antifier elimination, we can repl ace
<p (v, ill) by an equivalent formula which excludes quantifiers and conse-
quently is a finit e Bool ean combination of equat ions
q(v, ill) = 0
wher e q(x, ff) is a polynomial with coefficients in the subring generated by
1. Hence X = <p(Kn, il) is the Bool ean combination of t he algebraic varieties
defined by the formulas
q(v, il) = 0,
and so is a const ruct ible set .
Not ice that in every field K t he subsets of K" defi nable by qua nt ifier free
formulas are constructible. Qu antifier elimination ensures that , when K is
algebraically closed, no further definable set ari ses.
The following proposition underlines the geometrical content of Tarski's The-
orem.
Theorem 2.4.7 (Chevalley) Let K be an algebrai cally clos ed field, n be a
positive integer, X I<n+l , X' be the projection of X onto the first n
coordinates. If X is constructible, then X' is also constructible.
Proof. If <p(V, w) defines X , then :3w<p(V, w) defines X' .
Now let us consider l-ary definabl e sets in an algebraically closed field K .
In t his restrict ed framework, t he following proposition holds.
Corollary 2.4.8 Let K be an algebraically closed fie ld, X I< be definable
in K. Th en X is ei ther fin it e or cofini le.
Proof. For every q( x , il) E I<[x], q( v , il) = 0 defines I< if q( x , il) is zero, and
a finit e set otherwise. A finit e Boolean combinat ion of finit e or cofinite sets
is still finit e or cofinite.
Actually we can say even more. Indeed , in any (possibly non- algebraically
closed) field K , a subset X of I<definable by a quantifier free formula is either
60 CHAPTER 2. QUANTIFIER ELI MI NATI ON
finit e or cofinit e. Qu antifi er elimination extends t his proper t y t o arbi trar y
l -ary subsets when J( is algebraically closed.
To conclude t his section , we wa nt t o propose a nice application of Model
Theory t o Algebra within algeb raically closed fields. This is t he so called
inj ecti vity-implies-surjecti vity Theorem, du e t o J. Ax [3]. Compact ness , and
t he consequent remark t hat t he sentences t rue in t he complex field are just
t hose satisfied by t he algebraically closed fields of characterist ic p for almost
all primes p , ar e used to ded uce
Theorem 2.4.9 Any inj ective morphism f from an algebraic variety V over
the complex field into V itself is surjecti ve.
Proof. We already noticed that any algebraic variety is a definable set , and
is even defined by a finit e conj unction of equations (possibly with parame-
te rs) . In part icular let t he for mul a
1\ Pj (il, ii) = 0
j 9
give V in t his way (t he pj's are polynomials with int eger coefficients, and
ii denotes a sequence of com plex par amet ers). Analogously, a morphism
bet ween varieties is a map defined by a finit e conj unction of equations.
Acco rdingly let
1\ qi (V, e, ii) = 0
i:s
yield f (t he qi' s are again polynomi als with int eger coefficients; we can freely
use her e t he sa me paramet ers ii as before; if necessary, we extend ii t o include
new complex numbers). At t his point it is an eas y exercise to wri te a first
order sentence in t he language L (without paramet ers) saying:
for all z, if Ai<s qi (v, w, Z) = 0 defines a morphism from the
variety given by Aj <t Pj(v, Z) = 0 into itself, and t his morphism
is injective, th en it I S also surjective.
Let ti denot e, as usu al , t he length of V. What we have t o show is t hat
t he complex field is a model of all t hese sentences when t he pj's and t he qi' s
range over t he polynomials wit h int eger coefficients, equivalent ly t hat ACF
o
includes t hese statements. Using compactness , we can alternatively check
what happens in ACF
p
when p is a prime, and so if t he pr evious sentences
are true in every algebraically closed field of characterist ic p; as ACFp is
com plete, it suffices to look at t he behavi our of a single model of ACF
p
,
2.5. TARSKI AGAIN: REAL CLOSED FIELDS 61
for instance of the algebraic closure Fp of t he field Fp with p elements: a
positive answer in Fp for sufficientl y many p implies a positive answer for
the complex field. So take an algebraic variety V over Fp and an injective
morphism f from V to V over Fp. Use the algebraic fact t hat F p is locally
finit e and represent V as t he union of it s int ersections with F" where F
ranges over t he finit e subfields of F p (containing t he parameters defining V
and f ). Recall t he t rivial principle t hat any injecti ve function from a finit e
set t o it self is also surjecti ve. Deduce t hat t he rest rict ion of f t o V n F" is
surjective for every F . Ext end t his resul t to f: f is surjective, as requi red.
..
2.5 Tarski again: Real closed fields
In t his section we deal wit h t he qu antifier elimination t heorem for real closed
fields . This is t he main result of Tarski in t his framework , not only because,
as we shall see, t he proof is deeper and more cornplicat ed t han in t he cornplex
case , but also because t he ordered field of reals is intrinsecally related to ge-
ometry. It is certainly needless to recall t hat, for instance, in t he Euclidean
plane equipped with some fixed Cart esian axes, every point is essent ially an
ordered pair of real s, every straight line is the vari ety given by a polyno-
mial with degree 1 and 2 unknowns over th e reals, and so on. Accordingly,
statement s about point s, lines, ... can be easily t ra nslated into st at ements
about reals, addit ion, multipli cati on (oft en in a first order way). In part ic-
ular, a decision algorit hm about th e t heory of t he ordered field of reals (t he
elementary algebra accor ding to Tarski's te rminology) should work for (first
order ) Euclidean geomet ry as well.
Act ually Tarski's quantifier elimin ation procedure dealt with t he reals rather
than with t he t heory RCF. But , just as in th e complex case, one can realize
that t he basic ingredients of the proof concern ar bit rary real closed fi elds. So
we state (and show) t he resul t in t his (seemingly enlarged) setting; but we
shall deduce quickl y t hat RCF is complete and hence equals t he th eory of
(R, +, - , ,,0, 1, We follow t he elegant approac h of Cohen [27] rather
t han Tarski 's or iginal proof.
Theorem 2.5.1 Th e theory RC F of reals closed fiel ds has the elimination
of quant ifiers in the language L = {+, - , " 0, 1,
Proof. By proceeding as in the case of algebraically closed fields, one pr e-
liminarily reali zes that the heart of th e matter is to eliminate th e quantifi er
62
:3 in a formul a
CHAPTER 2. QUANTIFIER ELIMINATION
:3Wct(w, v)
where ct(w, v) is t he conjuncti on of at most one equation p(w, v) = 0 and a
finite (possibly empty) set of disequations qj( w, v) > 0 (with j :::; m), where
p( x, Y) and qj (x, if) (j :::; m) are polynomials with integer coefficient s.
So let us open a (long) parenthesis and examine an arbitrary polynomial
f( x) = L i<t f i
Xi
with coefficients in a real closed field K . It is known that ,
if f i is not -0 for all i :::; t , t hen f( x ) has at most t roots in the field. Fix t.
Then it is easily seen th at
1) the funct ion calculating, for every polynomial f( x) as before, equiva-
lently for every non-zero sequence (fa, . . . , f t) in Kt+
I
, how many
roots f (x) admit s
as well as , for every r and s with 1 :::; r :::; s :::; t ,
2) the set of nonzero sequences (fa, .. . , f t) in K
t
+
I
such t hat f( x) has
exactly s roots,
3) t he funct ion mapping any nonzero (fa, .. . , ft) into the r-th root of f( x)
are definabl e in any ordered field K in a unifor m way (independent of K).
We claim that , within real closed fi elds, for every t, th ese object s are de-
finabl e by quantifier free formul as, st ill in a uniform way (independent of
the underlying field). To see this, one uses t he Sturm t heory of real root
counting. We proceed by induction on t.
The case t = 0 is clear: t he number of roots is 0 if fa i= 0, and undefi ned
ot herwise; 2 and 3 are empty objects.
So assume t > 0 and suppose our claim true for every natural value < t , in
order t o extend it to t . The idea here is t o relate the zeroes of f( x) t o th e
roots of it s derivative and th e sign of f( x) in these root s. Hence build t he
formal derivative f'( x) of f(x) with respect to x
f
' ( ) ,"",, 'f i - I
X = LJ Z i
X

a<i9
Preliminaril y, notice t hat f'( x) = 0 if and only if (fl " ' " fd = (0, . . . , 0).
Except this case , induction equips us with quantifier free formul as defining
(with respect to (fa, ... , ft) via (fI , 212, ... , tft))
1) t he funct ion counting, for every sequence (fa, ... , ft) with (fI , ... , fd i=
(0, .. . , 0), how many root s f'( x) ad mits,
2.5. TARSKI AGAIN: REAL CLOSED FIELDS
and , for 1 ::::; r ::::; s < t ,
63
2) the set of the sequences (fo, .. . , ft ) in K
t
+l such t hat (fI, . . . , ft) is not
zero and f' (x) has exact ly s roots,
3) the function mapping any (suitable) non-zero (fo , . . . , fd int o the r-th
root of f' (x) .
Now ord er the root s of f' (x)
PI < .. . < Ps
The intermediate value property, holding in every real closed field, ensures
that f' (x) cannot change its sign between two successive root s. Can we
deduce th at J( x) is monotone (increasing or decreasing according to the sign
of f'( x in t he same int erval ? Cert ainly yes in th e case of th e real field: thi s
is a well known resul t in elementary real analysis. But a complete algebraic
(although non trivial) proof can be done for polynomials by using onl y t he
axioms of RCF. Consequent ly, in every real closed field K , f( x) is monotone
(increasing or decr easing according t o the sign -positi ve or negative- of it s
der ivative) in each int erval (pi, Pi+I), 1 ::::; i < s. Now look at f(Pi) and
f (Pi+d
(i) If t hey are not 0 and t heir sign is t he same, th en (pi, pi+d does not
contain any root of f( x) because f( x) is monotone in th e int erval (notice
th at t he cases when exact ly one between Pi and Pi+l annihilat es f (x )
ca n be handled in a similar way).
(ii) If f(P i) and J(Pi+d admit opposite signs, then (pi, pi+d does cont ain
a root of f( x) by the intermediate value property. The uniqueness of
this root might follow from Rolle 's Theor em (two distinct roots of f( x)
Pi < a < b < Pi+! determine a new int ermediat e root of f' (x), and
t his is impossibl e) . Elementary analysis ensures t hat Rolle's Theor em
is certainly true for t he reals; but , again, one can give an alt ernative
algebraic and non t rivial proof (for polynomials) holding in every real
closed field.
(iii) Assume at last f (Pi) = f(Pi+d = O. The argument in 2 again excludes
any additional int ermediate root of f( x) .
This machinery lets us count th e roots in t he int erval [PI , Ps]' But what can
we say in (-00, pI) and (Ps, + oo)? The same arguments as before ensure
that f(x) is monotone, and at least one root occurs in each of these half-
lines. But our setting changes when we examine the existence of this root.
For, every interval (pi, Pi+!) (1 i 5) is bounded, while our half-lines are
not. However the following algebraic fact helps us.
Let f(x) = I:i<t fiXi as before. Then f(x) has no roots out of the
interval [-a, aT where a = 3tmax{(lft-d
t-
1
1 : 0 < i t} + 1.
(The proof only uses the axioms of ordered fields). So a possible root less
than PI should lie in [-a, PI), and a possible root greater than Ps should
belong to (Ps, a]; moreover the absolute value function I I can be defined
in a quantifier free way, because, for every b E K, Ibl is b when b 0 and -b
otherwise. Hence we are led to a bounded framework, and we can proceed
as in the previous cases.
In conclusion , we have provided a uniform procedure counting, for every
nonzero (fa, .. . , ft) in K, how many roots f (x) admits. The function cal-
culating their number 5 is defined by a quantifier free formula (essentially
checking the sign of f (x) in the roots of its derivative and in a). Similarly
the set of non-zero sequences (fa, .. . , ft) in K for which f(x) has exactly 5
roots can be defined by checking these sign relations and forming a suitable
first order disjunction to list the cases when 5 occurs. Finally, the function
producing, for every non-zero (fa, ... , ft) and 1 r 5, the r-th root of
f (x) is easily defined on the same basis.
This accomplishes the proof of th e claim and ends our parenthesis. Now
we come back to quantifier elimination. Recall that we are considering a
formula
(a) :3w(p(w, 17) = 0 1\ 1\ qj(w, 17) > 0)

or
(b) :3w 1\ qj(w, 17) > 0

where p(x, jJ) and qj(x, jJ) (j m) are polynomials with integer coefficients.
Each of them can be written as a polynomial with coefficients in Z[yJ in the
following way
p(x, jJ) = LPi(y)X
i
,
i<t
(a) is quickly reduced to (b) because its formula is equivalent to
2.5. TARSKI AGAIN: REAL CLOSED FIELDS
(/\ Pi(V) = 0 1\ :3w 1\ qj(w, v) > 0) V
i5, t j5,m
Vl<r<s<t ("p(x , v) has s roots" 1\
1\ " t he r -t h root Pr(v) satisfies !\ j5, mqj(Pr(v) , v) > 0") ,
65
where t he latter disjunct ca n be expressed by a quantifier free first order
formul a. So look at (b) . For every j ::; m and for every Sj ::; tj , t here are
quantifier free formul as defining , for every real closed field 1<, the set of th e
sequences bsuch that q(x, b) has Sj roots in x, and listing t hese roots
One can cornpute the sign of qj (x, b) in t he intervals
(-00 , Pj,1(b)) ,
(pj,i(b), pj,i+db)) (1::; i < Sj),
(Pj,Sj (b) , +(0)
in a uniform way (independent of 1< and b) by looki ng at t he (sign) value of
qj(pj ,l(b) -1, b)
.(Pj,i(b) + (pj,i+l(b) b)
qJ 2 '
qj(pj,sj(b) + 1, b)
respectively. List all the possible orderings of t he roots (in x) of t he qj (x, b) 's
when j ranges over the natural numbers ::; m, and di vide in every case 1<
int o finit ely many int er vals such that the qj (x, b) 's have a constant sign
(with respect t o x) in each of t hem; check these signs (in the way suggested
before) and for m a suitable disjunct ion picking the intervals where all t hese
signs are positive. This procedu re is independent of 1< and band provides
the required quantifier free formul a.
Corollary 2.5.2 RCF is model complete.
Corollary 2.5.3 RCF is complet e; in particular, RCF is the theory of the
ordered field of reals (as well as of every real closed field).
66 CHAPTER 2. QUANTIFIER ELI MINATI ON
Proof. There is a minimal ordered real closed field , embedded in any model
of RCF. This is the ordered field Ra of real algebraic numbers. The model
completeness of RCF ensures that every real closed field is an elementary
extension of Ra. In particular all the real closed fields are element arily
equivalent to Ra and, consequently, t o each other. '"
This is t he first completeness proof we give about RCF; in fact Vaught 's
criterion does not appl y because RCF is not categorical in any infini t e
power .
We have seen th at real closed fields eliminate quantifiers in t heir language
L = {+, -, ' ,0, 1, Notably, they are fully charact erized by t his prop-
erty: for, Macintyre, McKenna and Van den Dries showed t hat an or-
dered field, whose theory has the quantifier eliminat ion in L , must be real
closed . We also noti ce that RCF does not pr eserve quantifier elimination
in t he restricted language L' = {+, -, " 0, I} wit hout ord er . Act ually one
can remember t hat , even in checking solvability of the popular equation
ax
2
+ bx + c = 0 wit h degree 2 and 1 unknown over t he reals (or over any
real closed field), one needs a disequ ation b
2
- 4ac 0 to ensure roots, and
hence to eliminate 3 in the formul a 3w(V2 W2 +VIw+va = 0) . More form ally,
recall that , with resp ect to the theory of the real field , the formulas
',o( v):
',o' (v) : 3w(v = w
2
)
are equivalent. As RCF is complete and hence equals the t heory of t he real
field, t he same holds in every real closed field. Consequently t he I-formula
(wit h t he quantifier 3)
defines th e set of non-negative element s in ever y real closed field. However
',o(v) cannot be equivalent in RCF to any quantifier free I-formula ',o" (v). In
fact ',O(R ) is the half-line [0, +00) of R , and so is both infinite and coinfini te ,
whil e ',O"(K) is either finite or cofinite for ever y field K: see the proof of
Corollary 2.4. 8.
Now we discuss decidabili t y: as already said, t his was t he main consequence
of elimination of qu antifiers, according t o t he general feeling in t he fourties.
Corolla r y 2.5.4 RCF is decidabl e.
Proof. Owing to quantifier elimination, every L-sentence (J is equivalent in
RCF to a Boolean combination of sente nces m = n or m < n where m and
2.5. TARSKI AGAIN: REAL CLOSED FIELDS 67
n are integers. This quantifier free statement can be easily checked in our
framework. ..
We shall comment this result later in 2.9. Now we examine another remark-
able consequence of quantifier elimination, namely definability. Recall that,
in an ordered field K, every semialgebraic set (in other words, every finite
Boolean combination of sets of solutions of disequations
q(x) 2:: 0
with q(x) E K[X]) is definable.
Corollary 2.5.5 In a real closed ordered field K, the definable sets are ex-
actly the semialgebraic ones.
Proof. Let n be a positive integer, X S;; K" be a set definable in K. So
there are a formula <p( il, w) of L and a sequence if E J( such that
Owing to Tarski's quantifier elimination theorem, we can assume that <p(il, w)
has no quantifier and hence is a finite Boolean combination of disequations
q(c, w) 2:: 0
with q(x, if) E Z[x, Y'J. Consequently X is a finite Boolean combination of
sets of solutions of disequations
q(il, if) 2:: 0,
and so is a semialgebraic set. ..
Here is a geometric restatement of Tarski's Theorem.
Theorem 2.5.6 (Tarski-Seidenberg) Let K be a real closed ordered field, n
be a positive integer, X S;; J(n+l, X' be the projection of X onto the first n
coordinates. If X is semialgebraic, then X' is semialgebraic, too.
This formulation is due to Thorn, who also coined the name semialgebraic
set. It has a more geometric flavour. Some mathematicians might appreciate
this alternative terminology, for instance because it allows to state several
results in a (seemingly) more agreable way (avoiding logic). Nevertheless,
Tarski's original approach (via quantifier elimination and formulas) often
provides quicker proofs, even in this geometric framework. Let us quote the
following example from [168].
Exa m p le 2.5.7 Consider the following statement: For a semialgebraic func-
tion f from R n+! to R, the set A of the sequences xin R n+! s uch that
limy--t oo f(x, y) is in R
is semialgebraic.
If one replaces everywhere semialgebraic by definable, this proposition may
lose part of it s (mathematical) glamour. Bu t , using logic, one obtains a
short proof: A is definable via t he formula
<p( V) :3w'liE(E > 0 --+ :3rVy(y > r --+
--+ :3 z(J(v, y) = z A [z - wl < E)))
(recall that both f and the absolu t e value are definabl e) . A direct approach
via semialgebraic sets and projections is longer.
In a real closed field K , the definable subsets X J( have a very simple
form.
Corollary 2.5.8 Let K be a real closed fi eld, X be a definable subset of K.
Then X is a finit e un ion of intervals (closed or open, possibly wit h infinite
endpoints) .
Proof. Let q(x) E J([x]. We know th at q(v) = 0 defines J( if q(x) = 0 and a
finit e set (that is, th e set of th e roots ao < .. . < as of q(x ) in K) oth erwi se.
On the other hand, q(v) > 0 defines 0if q(x) = 0; otherwise q(v) > 0 defines
t he union of some int er vals among ] - 00, ao[, ]ao, al[, ..., ]a
s,
-l-oo] (recall
that K sat isfies t he int ermediat e value proper ty for polynomials) . Hence
any definable (equivalent ly, semialgebraic) set X J( is a finite Boolean
combination of intervals, and so a finite union of intervals.
2.6 pp-elimination of quantifiers and modules
In this section we deal wit h (left) modules over a (countable) ring R with
identi t y. In Chapter 1, we introduced a suitable language LR = {O, +,- , r
(r E R)} for t hese struct ur es, and we saw how t o axiomatize their class
by first ord er sente nces in LR. Let nT denote the corresponding t heory.
A qui ck look at the axioms of nT shows t hat each of them is a universal
sentence Vva( v) where a(v) is an at omic formula of LR; this confirms that
the class of t he models of nT (namely of t he R-modules) is closed under
2.6. PP-ELIMINATION OF QUANTIFIERS AND MODULES 69
substruct ures. Now we wonder whether nT has quant ifier elimination in
LR. A t rivial example shows th at t his is false even in t he simple case when
R is t he ring Z of integer s. In fact , just consider Z as a module over it self.
In Z the formula
<p (v) : 3w(v = 2w)
defines t he set 2Z of even integers. On t he oth er side , every atomic formula
<p' (v) in Lz is equivalent within zT, and hence in t he t heory of t he Z-module
Z, t o
rv = 0
for some non-negative int eger r . This formula defines in Z {O} if r =1= 0 and
Z otherwise. No Boolean combinat ion of t hese sets can equal 2Z. Therefore
no quantifi er free formula <p' (v) of Lz is equivalent to <p (v) in Th( Z) , and
so in zT. It follows that zT does not eliminat e th e quantifiers in Lz.
However notice that <p (v) is a ty pical pp-formula in Lz. Indeed we will see
that , for any R, t he pp-formulas of LR are just t he onl y obs truction to th e
elimination of quantifiers of nT in LR. Let us see why.
Take any (countable) ring R with identit y. Recall th at a pp-formula of LR
is an existent ial formula of t he form
<p (v) : 3w(A v = B w)
where A and B are matrices with coeffi cients in R with suitable sizes, .
denot es t he usual row-by-column product between matrices, and V, wshould
be viewed as column vectors (wit h suit ably many rows) . So, when w= 0,
pp-formulas includ e th e atomic formulas of LR .
In Chapter 1 we point ed out th at , for every pp-formula <p (v) of LR and
every R-module M , <p (M n) is a subgroup of M
n
(called a pp-definable
subgroup), but is not in general a submodule. Let us add here some mor e
remarks about pp-formulas.
Rema rk 2.6.1 1. If <p( v), 7jJ (v) are pp-formulas of LR , t hen also <p (v) 1\
7jJ (v) is (equivalent in n T t o) a pp-formula.
2. Let <p (v, Z) be a pp-formula of LR, <p (v, Z) : 3w (A
t
(v, Z) = B w).
Then <p (v, O) is a pp-formula, and hence, for every R-module M ,
<p(Mn, O) is a pp-definabl e subgroup of M n. Furth ermore, for ev-
ery ii EM, <p(M
n
, ii) = 0 or <p (Jlvt
n,
ii) is a coset of <p (Mn, 0) in M (in
fact , given bE <p (Mn, ii), it is easy t o check <p (M n, ii) = <p (Mn, o)+b).
3. Let SO( v) , 'Ij; (if) be pp-formulas of LR with n free variables, and let k
be a positive int eger. It is simple t o writ e a sent ence in LR ensuring
that , in a given R-module M , t he index of SO (Mn)n 'lj;(Mn) in SO (M n)
is k; in det ail t his sente nce says
We will denote it by (SO : 'Ij;) k. Any sent ence of this form is called
an invariant statement (we will see later the reason why) . Notice th at
the finite Boolean combinations of invariant st at ements include t he
sent ences sayin g:
t he index of SO (Mn ) n'Ij; (Mn) in SO (Mn ) is k (" = k" means "
k" but " 1. k + 1") ; we will denote this formula by (SO : 'Ij; ) = k;
t he index of SO(M
n)
n 'Ij; (M
n
) in SO (M
n
) is :::; k (" :::; k" means
" 1. k + 1" ); we shall denot e this formul a by (SO: 'Ij;) :::; k.
At this point we can state and show the following fund amental theorem (of
pp- elimination of quantifi ers for modules).
Theorem 2.6.2 (Baur - Monk) Let R be a (countable) ring with identi ty.
Then the pp-formulas of LR together with the invariant statements form an
elimination set for RT in LR. More precisely: for every formula a(v) of LR,
there are a Boolean combination j3 of invariant stateme nts and a Boolean
combinati on ,,(if) of pp-formulas such that
'v'v(a(v) +-7 (3/\ ,,(if) ) E RT.
We shall use in our proof the following result of group th eor y.
Lemma 2.6.3 (B. H. Neumann) Let g be a group, a, ai E G, H, Hi be
subgroups of g (where i ranqes among the naiurals less than some fi xed
N ), aH U <N a.Hi. Let I be the set of the tuiturals i < N f or which
IH : H n Hil :::; N !. Then aH UiEI aiHi .
Now let us begin the proof of t he t heorem.
Proof. We proceed by induction on a(v). If a(v) is an atomic formula, then
a(v) is directly a pp-formula. The cases --, and /\ are easy to handle. So
suppose t hat a (v) is of t he form 'v' wa' (w, v), where th e induction hypothesis
2.6. PP-ELIMINATION OF QUANTIFIERS AND MOD ULES 71
ensures that there exist an invariant st ate ment (3' and a Bool ean combination
, '(w, V) of pp-formulas such that
VwV v(0:' (w, V) +-+ (3' 1\ " (w, V)) E RT.
pt reduction: without loss of generality, 0:' (W , V) is a disjunction of pp-
formulas or negations. In fact , Vv(o:(v) +-+ (3' 1\ Vw, '(w,V)) EnT. Accord-
ingly we can replace o:' (w, V) by , ' (w, V) , whi ch is a Boolean combination
of pp-formulas , and hence is equivalent to a conj unction of disjunctions
of pp-formulas or negations. Correspondingly put o:'(w, V) : I\ j<s o:j (w, V)
where, for every j ::; s, o:j (w, v) is a disjunction of pp-formulas or negations ,
Vwo:'(w , v) is equivalent in RT to I\ j<sVwo:j( w, v). Then we can handle
o:j(w, V) (with j ::; s) instead of o:'(w,V) .
2
n d
reduction: o:'(w,V) is of th e form B(w,v) -+ Vi<NBi (W, V) where N is a
positive int eger , B(w, V) and Bi(W, V) (wit h i < N) are pp-formulas. In fact
0:' (w, v) is a single disjunction of pp-formul as and negations. But we know
t hat any conj unction of pp-formulas is (equivalent in RT to) a pp-formula,
and hence any disjunction of negations of pp-formulas is t he negation of a
single pp-formula. This clearly impli es our claim .
Let us summarize the sit uation. We want to find (3 and , (v) such th at , for
every R-module A1 and every sequence ii in M,
(1) B(M ,il) UBi(M,il).
i<N
if and onl y if A1 F (3 1\, (il). We know t hat, given A1 a nd ii , eit her B(M , il) =
oor B(M , il) is a coset of th e pp-definable subgroup B(A1, 0). The same ca n
be said a bout Bi(M , il) for every i < N . By t he way, notice t hat 3wB(w, V) ,
3WOi (w, V) (with i < N) are pp-formulas. (1) is certainly true when ii
sat isfies -,3wB(w, v) (the negation of a pp-formula) in M , and certainly
false when il satisfies
3wBi(W, v) 1\ V-,3wB
i
(w, v)
i<N
III M . So there is no loss of generality for our purposes in assuming
B(M , il) =I- 0 and Bi(M , il) =I- 0 for every i < N . Let 5 be t he set of
t he indi ces i < N satisfying
IB(A1 , 0) : B(M , 0) n Bi(M, 0)1::; N ! .
Notice that 5 depends on M (and on ii, of course). However there are only
finitely many poss ible ways of choosing 5, and each of them is described by
72 CHAPTER 2. QUANTIFIER ELIMINATI ON
a suit able invariant state ment . Let us assume, with no loss of generalit y,
that 8 is just the set of t he positive integers m for some m N . We can
apply B. H. Neumann's Lemma and deduce t hat (1) is equivalent to
(2) B(M , ii) UBi(M, ii) .
i<m
Put K = B(M , 0) n n i < m Bi(M, 0) . As B(M , ii) and Bi(M , ii) for i < m are
union of cosets of K in 1'1,1 , (2) can be equi valently writ t en
(3) B(M ,ii) /K UBi(A1,ii) /K.
i <m
As B(M ,ii)/K is finit e, we can use some (hopefully) well known combina-
torial arguments and restate (3) in the equi valent form
(4) 2)-I)IXII(B(M,ii) n nOi(M,ii)) /KI =
x i EX
where X ranges over t he subsets of {a, 1, ... , m-I}. For ever y X , put
k(X) = IB(M, 0) n nOi(M, 0) : KI ;
i EX
notice that , when B(M , ii) n n iEX Bi (A1, ii) =f- 0,
k(X) = I(B(M,ii) n nOi(M,ii)) /KI
i EX
Moreover k(X) N !N. Hence we have shown that M sat isfies a(ii) if and
only if I:(-I)
IX1
k(X ) = 0, where the sum concerns all the subsets X of
{ a, 1, ... , m - I} such that M F :3w(B(w, ii) A A iEX Oi(W, ii)), and hence if
and only if M satisfies a convenient disjunction of conj unctions of invari ant
statements and pp-formulas. This is what hap pens for a given 8 . As t here
are onl y finit ely many possible S'e, one ca n find some suit able f3 and -y (V),
valid for every R- mod ule M.
Remark 2.6.4 1. Not ice that the pro cedure given in the proof of Theo-
rem 2.6.3 is effect ive, and provides explicit ly for every a(v) the required
formul as f3 and -y(0 . Furthermore f3 is act ually a finit e Bool ean com-
bination of invariant statements conce rni ng pp-for mulas cp( v), 'IjJ (v)
(with at most one free var iable) .
2.6. PP-ELIMINATION OF QUANTIFIERS AND MODULES 73
2. In particular, when a is a sentence of LR, what the previous proce-
dure produces is just a Boolean combination /3 of invariant statements
(concerning pp-formulas <p(v), 1f;(v) with at most one free variable)
such that a +-+ /3 EnT.
3. Now fix an R-module M. Then, for every formula a(v) of Ln, there
exists a Boolean combination I(v) of pp-formulas such that M F
Vv(a(v) +-+ I(V)) (in fact, we know that a(v) is equivalent to /3A
1
(v)
for some Boolean combination I(v) of pp-formulas and some sentence
/3; so, if M F /3, then a(v) is equivalent to I(V), while, if M F -,/3,
then a(Mn) is empty and consequently a(v) is equivalent to l5(v) A
-,15(v), where l5(v) is an arbitrary pp-formula).
With respect to definable sets in modules, this is what Theorem 2.6.3 implies.
Corollary 2.6.5 Let M be an R-module, n be a positive integer. Then
every set X M" definable in M is a finite Boolean combination of cosets
of pp-definable subgroups.
Proof. There exist an Ln-formula a( ii, w) and a sequence a In M such
that X = a()\1n , a). We can assume that a(v, w) is a Boolean combination
of pp-formulas, and we know that, for every pp-formula <p(v, w) , <p(Mn, a),
when it is not empty, is a coset of the pp-definable subgroup <p(Mn, 5). '"
We can also characterize the complete extensions of nT (and hence the
=:-classes of R-modules).
Corollary 2.6.6 Let M ;M' be two R-modules. Then M =: M' if and only
if, for every choice of two pp-formulas <p( v), 1f;( v), the indices of <p(M) n
1f;(M) in <p(M) and <p(M') n 1f;(M') in <p(M') are either finite and equal,
or both infinite.
Proof. It is clear that, if M and M' are elementarily equivalent, then, for
every <p(v),1f;(v) as before, and for every positive integer k,
M F (<p: 1f;) k {::} M' F (<p: 1f;) k.
The inverse implication follows from Remark 2.6.4,2. '"
The previous result explains why "invariant statements" are called in this
way: actually these sentences fully characterize any R-module )\1 up to
elementary equivalence.
74 CHA PTER 2. QUANTIFIER ELIMINATION
Now let us discuss th e content of t he previou s resul t s in some particular case .
We deal with a principal ideal domain n (this setting includes the ring Z of
integers, as well as any field). First let us examine a generic pp-formula of
LR
O' (v) : :3w (Av= Bw).
A and B can obt ain a simpler form when n is a principal ideal domain.
For, it is a fact of Algebra t hat , in t his framework, t here are two inver ti ble
matrices X, Y wit h coeffi cients in R such t hat t he prod uct B
'
= X BY is
diagonal. So 0'(v) is equivalent to
a nd, unl ess replacing 1V by y-1w, A by XA and B by B
'
, one can suppose
B diagonal in O' (v). Consequent ly O'(v) becomes of the form
n
:3wl ... :3 wm ( 1\ ( L aijVj = biiWd )
j = l
Now let us moment arily restrict our analysis to a smaller setting.
Case 1: n = K is a field (so we are dealing wit h vectorspaces over K).
Assume that , for some i with 1 i m, bii i- o. Then we ca n divide t he
i-th equation in 0'(v) by b
ii
and consequently ass ume bii = 1. At this point
it is easy t o show t hat O' (v) is equivalent to
n
1\ (LaijVj = 0),
j = l
which is a conj unct ion of atomic formulas .
Combine t his obser vation and Baur-Monk's Theorem, and dedu ce t hat every
LK-formul a is equivalent in KT t o a conj unction of a quantifier free formul a
and a Boolean combination of invariant state ment s.
Moreover t he pp-formul as with a unique free vari abl e v reduce to r v = 0
for some rE I<, and hence to eit her v = 0 when r i- 0 or to v = v when
r = o. Consequent ly t he only pp-definable subgroups of a vectorspace V are
{O} and V. Owing to Corollary 2.6.5, t he subsets of V defina ble in V are
just t he finite Boolean combinations of t he cosets of t hese subgroups, and
so redu ce t o t he finit e or cofinite subsets.
Now let us examine invariant statements. In par t icular we dir ect our atten-
tion on t he sentences of t he form
2.6. PP- ELIMINATION OF QUANTIFI ERS AND MODULES 75
wher e k is a positi ve int eger. In any given vectorspace V over IC, t hey witn ess
if the size IVI di V is finit e or not , and, in t he positive case, it s value. We
claim that they can even determine th e =-type of t he vectorspace. Let us
see why.
First ass ume J( infinit e. We know t hat , in t his case, all t he nonzero vec-
t orspaces over IC are elementarily equivalent (for, t heir t heory is complet e).
In oth er words, when J( is infinite , t here are only two =-classes of IC-
vectorspaces: t he former contains all t he nonzero vectorspaces, and t he
latter reduces to t he zero space. But a vect orspace V is {O} if and only
if V 1= (v = v : v = 0) = 1. In parti cul ar , t he statements (*) determine t he
= -t ype of any vect orspace .
Now assume I< finit e (say of size q). Now we meet infinitely many =-classes
of IC-vecto rspaces. In fact, there is a class for every natural n , consist ing
of the (pairwise isomorphic) vect orspaces of di mension n over IC (hence size
qn), while infinit e vectors paces form again a uniqu e class. The sente nces
(v = v : v = 0) k can obviously dist inguish t hese classes.
One can ded uce t hat every invaria nt statement in LK is a Boolean combi -
nation of sent ences (v = v : v = 0) k where k ranges over t he positi ve
integers.
In conclusion, given a fi eld IC , t he atomic formulas of LK t ogether wit h t he
invariant state ments (v = v : v = 0) k, with k a positive int eger , form an
eliminat ion set for K;T. In particul ar t his yields quantifier elimination in LK
for th e t heory of infinit e IC-vectorspaces.
Case 2: Now let us enlarge our setting t o arbitrary principal ideal domai ns
R: First let us examine a pp-for mul a O'(v) . We cannot expect any longer
t hat, whenever b i= 0 in t he i-t h equation of 0'(v), one can divide t he whole
equation by bii and so obt ain b
ii
= 1. However O'(v) is equivalent in n T to
a conj unction of formulas
(1) I:/l=iaijVj = 0 (for b
ii
= 0) ,
(2) :JW(:L5=i aij Vj = qw) (where q = b
ii
i= 0).
The lat t er ones are divisibility condit ions: we can abbreviate each of t hem
by
Of course, when q is a unit in R: t his is a trivial condit ion and can be
forgo t t en. In the remaining cases, recall that q decomposes (uniquely) as a
product of powers of pairwise distinct primes in n, and q divid es an element
r E R if and only if all th ese pri me powers divide r . So there is no loss of
generality in ass uming that , in (2) , q is a prime power .
The situation becomes clearer if we restrict our analysis to formu las hav-
ing only one free variables. In fact, in this case, our pp-formula a(v) gets
equivalent in nT to a conjunction of form ulas
(1) ' rv = 0 (a torsion condit ion),
(2) ' pi I sv (a divisibilit y condition) ;
here p, r, s E R, p is a prime and I is a non-negative integer. Again, simple
algebraic fact s about principal ideal domains let us ass ume t hat s itself
is a power of p, s = ph for some non-n egative integer h < l. Every pp-
formula in at most one free variable is a conj unct ion of t orsion and divisibility
conditions as before.
This result helps also the analysis of invari ant statements (<p (v) : 'Ij7 (v)) 2': k.
We avoid her e too many details. However we wish to menti on t he following
sentences (r, p are elements of R, p is prime, n, k are positive integers):
(3) (pv = 0 A pn-1lv : pv = 0 A pnlv) 2': k ,
(4) (pv = 0 ApnIv : v = 0) 2': k,
(5) (pn-1 1v : pnIv) 2': k,
(6) (v = v : rv = 0) 2': k .
The reader may check their truth (at least when R is the ring of integers)
in some familiar abelian groups, like Zj qhZ, t he Priifer groups Zj qOOZ , th e
localizations of Z at q (when q ranges over the primes, and h over the natural
numbers), and the addit ive group of rationals, and realize in this way their
meaning.
Indeed Wanda Szmielew (a student of Tarski's) showed t hat , for every R-
module M, the = -t ype of A1 is fully determined by the invariant statements
(3)- (6) satisfied by M .
In concl usion, owing to Baur-Monk's theorem, the formulas (1)-(6) are an
elimination set for the t heory nT when R is a principal ideal domain.
2.7 Strongly minimal theories
We saw in 2.4 that the onl y (l-ary) definable sets in an algebraically closed
field are the finite and cofinite ones. 2.6 told us that the same happens, for
2.7. STRONGLY MINIMAL THEORIES 77
instance, in every (infinite) vectorspace over a fixed countable field. One
can also check that even pure sets (in a language with no symbols besides
equality =) enjoy this feature; again, every definable set is either finite or
cofinite (this was implicitly shown when we treated definable sets in 1.7,
in particular when we provided an example of an infinite coinfinite non
definable set).
So we find some non-trivial algebraic structures A whose definable l-ary
subsets reduce to the ones definable in the pure set A (with the equality
relation =) by quantifier free formulas. Let us name these structures in the
following way.
Definition 2.7.1 An infinite structure A is said to be minimal if and only
if the only definable subsets of A are those finite or cofinite. A complete
theory T is said to be strongly minimal if and only if every model A of T
is minimal.
Hence any algebraically closed field is a minimal structure, and any theory
ACFp (with p = 0 or prime) is strongly minimal. The same is true for
infinite vectorspaces, or pure sets.
It should be underlined that the minimality of a structure is not preserved
by elementary equivalence. In other words there are minimal structures
A such that the theory of A is not strongly minimal , and so admits some
non-minimal models. Here is an example.
Example 2.7.2 Consider the theory dLO+ of discrete orders with a least
element 0 but no last element. We know that dLO+ is complete, and has
quantifier elimination in a language L with a constant (for 0) and a l-ary
operation symbol (for the successor function s) in addition to the relation
symbol Consequently every definable subset of a model of T is a finite
Boolean combination of intervals (possibly with an infinite right end point).
Therefore 0, s) is a minimal model of T, because every interval in
0, s) is either finite or cofinite. However no other model A = (A,
OA, sA) of T is minimal. In fact, let a E A satisfy a =J. (sA) n (OA) for any
natural n. Then both [OA , a[ and [a, -l-oc] are infinite intervals in A.
We shall examine again strongly minimal theories in the next chapters. In
particular, with respect to algebraically closed fields, we will prove a theorem
of Macintyre showing (among other things) that the only integral domains
with indentity having a strongly minimal complete theory are just the alge-
braically closed fields.
2.8 a-minimal theories
Turning now to linearly ordered st ruct ur es A = (A, ::; , . . .), we met in the
previous sect ions some examples where t he definable subsets of A reduce
to the finit e unions of int ervals (possibly with infinite end points) in (A, ::; ).
This is what happens in real closed fields (as observed in 2.5), but also in
dense or discrete (infinit e) linear orders (see t he sections 2.2 and 2.3) . This
suggests t he following definition.
Definition 2.8.1 An infi nite linearly ordered struct ure A = (A, ::;, . ..) is
called o-minimal if and only if every subset of A definable in A is a finit e
union of intervals (closed or open, possi bly with infinite endpoints) . A com-
plete theory T of infinite linearly ordered structures is called o-minimal if
and only if every model of T is o-minimal.
"0" abridges "order", of course. This o-rninimal setting clearl y reminds min-
imalit y. In fact t he minimal st ruct ures (and t he strongly minimal theories)
are t he ones where every definable (l -a ry) set is already defined by a quan-
t ifier free formula involving th e only (language) symbol =. Similarl y t he
o-minimal struct ures (and t heori es) are just those admit t ing a total ord er
relation j; such t hat every definable (l-ary) is already defined by a quantifi er
free formu la involving the only (language) symbol ::;.
In t his sense the o-minimal structures and th eor ies are th e simplest ones
in th e presence of a t ot al ord er relation. Nevert heless th ey includ e several
non- trivial algebraic exa mples. We will st udy in mor e detail t hese st ruct ures
and t heories in the last cha pter of t his book.
But it is worth emphasizing since now that , in spite of t he similarit ies un-
derlined above between minimality and o-minimality, a relevant difference
ari ses. In fact , we not iced t ha t th e theory of a minimal structure may admit
some non-minimal models, and so fail to be st rongly minimal. This does
not happen in the o-minimal setting. In fact the following t heorem hold.
Theore m 2.8.2 (Knight - Pillay - St einhorn) 1fT is the theory of a linearly
ordered o-minimal structure, then every model of T is o-minimal.
Accordingly, we ca n spa re the adverb "st rongly" in defining a t heory with
o-rninimal models.
Coming back to real closed fields, we would like t o mention her e a result quite
similar to t he one recalled at t he end of t he previous sect ion on algebraically
closed fields. In fact , it was shown by Pill ay and Steinhorn that t he only
2.9. COMPUTATIONAL ASPECTS OF Q. E. 79
ordered rings (with identity) having an o-minimal theory are the real closed
fields.
The proofs of these theorems will be provided in Chapter 9.
2.9 Computational aspects of q. e.
In this section we shortly discuss the quantifier elimination procedures with
respect to effectiveness and fastness. Actually these criteria did not corre-
spond to the spirit of the forties (and some decades later), when the main
quantifier elimination results were proved. For , those times lived the influ-
ence of Godel incompleteness and undecidability phenomena; so, according
to that feeling, any decision algorithm (such as Tarski's method for real
elementary algebra), or even a decidability theorem simply ensuring the ex-
istence of such a procedure without explicitly exhibiting it, were exactly
the best answer one might expect. But later, in the seventies, the birth
of modern computers and the beginning of their science changed this set-
ting and inspired a prevalent interest in quickly running algorithms. Hence
complexity theory introduced
* the class P of the problems having a fast procedure to find solutions,
* the class N P of the problems having a fast procedure to verify solutions
(namely to check that a solution works).
We agree that a problem has a solving procedure when there is a Turing
machine handling it and that an algorithm is fast when it runs in a polyno-
mial time with respect to the length of the input . To realize the difference
between finding or verifying solutions, look at the problem of factoring in-
tegers. To decompose a natural number 2:: 2 into its prime factors -more
precisely, to find these factors- can be significantly slower (at least with re-
spect to the currently available algorithms) than to check this decomposition
when done. Just to quote a famous historical example, F. Cole announced
during an AMS meeting in 1903 that the Mersenne number 2
67
- 1 is not
prime. Factoring 2
67
- 1 is not easy (and certainly it was not in 1903, when
computers were not available). But Cole's proof is quite short to write and
needs only one line
2
67
- 1 = 193797721 X 761838257287
and can be checked very quickly.
Coming back to P and N P, it is trivial to observe that P N P, because
a procedure yielding solutions implicitly confirms these solutions. A fun-
damental problem in complexity theory (and, more generally, in the area
linking computer science and mathematics) asks whether P = N P, hence
whether, whenever a problem has a fast procedure verifying solutions, then
it admits a (possibly slower but still) fast (=polynomial) procedure finding
solutions.
According to the new spirit, what is primary even in quantifier elimination,
mainly towards decidability, is to get fast methods. Let us discuss this
feature for the elimination of quantifiers of real closed fields: this is perhaps
the most interesting case, owing to its connection with elementary geometry.
However a devastating result of Fischer and Rabin shows
Theorem 2.9.1 Any algorithm deciding, or even verifying membership to
RCF for sentences in the language of ordered fields requires a running time
at least exponential with respect to the length of the input sentence.
Algorithm still means Turing machine. Notice that Fischer-Rabin's Theorem
only refers to the additive structure of the reals (R, +, -); recall that, in
this restricted language, the reals inherit in an obvious way a structure of Q-
vectorspace, because the scalar multiplication by any rational can be defined
using +, and so form a structure elementarily equivalent to (C, +, -). In
this sense, the theorem applies also to the complex field, yielding the same
negative lower bound for the decision procedures concerning algebraically
closed fields.
By the way, it is not known any decidable theory with infinite models and
a decision procedure running in polynomial time. Indeed, if such a theory
exists, then P = N P.
In particular, Tarski's original elimination procedure is very inefficient and
slow. However, recently faster and more powerful elimination methods for
real closed fields have been introduced. We wish to quote the Collins proce-
dure, called cylindrical algebraic decomposition CAD, working in the worst
cases in a doubly exponential time with respect to the number of variables
in the input formula. Implementations of CAD, and other real quantifier
elimination methods are discussed, for instance, in [33].
But now we want to treat briefly another intriguing relationship between
complexity theory and quantifier elimination (for arbitrary theories and
structures). A few lines ago we have said algorithm = Turing machine,
in accordance with the Church-Turing Thesis. But the Turing model of
computation has an intrinsic discrete character, so that its applications to
2.9. COMPUTATIONAL ASPECTS OF Q. E. 81
a continous fra mework (like R or C) seem laborious and unnat ural (how-
ever, see [121] for a discussion of this point). Consequently, new models of
computation, incl uding real and complex numbers as possible inputs, and
even working in arbit rary structures, have been introduced. We quote the
Blurn-Shub-Smale BSS model ([14], [13]), or also t he Poi zat approach [133].
These new perspectives extend t he Turing point of view: t he classical com-
put abili ty is just the comput abilit y over the field F
2
with 2 elements; but
now computability over ar bitrary st ructures is allowed. In part icular , for
ever y structure A , one can define in a suitable sense the classes P and N P
(over A), one can cornpare t hese classes and check if P = N P over A. To
introduce t hese matters in detail would require a long time, so we refer
t he interested reader t o the bibliography quot ed at t he end of the chapter.
Remarkably, quantifier elimination arises in t his setti ng. In fact , Poizat
observed
Theorem 2.9.2 If P = N P over A , then the theory of A eliminates the
quantifi ers.
It is comparat ively easy to realize why. Let us refer for simplicity t o t he Cole
example quoted before. To prove that 2
67
- 1 is composite req uires to find
a non trivial divi sor , and hence to obtain some witn esses of the (exist enti al)
sent ence 3u3v(2
67
- 1 = uv). But , after Cole , we have simply t o check t he
quantifier free sentence
2
67
- 1 = 193797721 X 761838257287.
In ot her words, what P = N P asks here is a procedure (indeed a quick
procedur e) of eliminat ion of quant ifiers . Theorem 2.9.2 provides several ex-
amples of structures for which P -I N P. For instance, recall t he Macintyre-
McKenna-Van den Dri es theorems characterizing the infinit e fields whose
theory eliminat es the quantifiers in the language for fields (they are t he al-
gebraically closed fields) , or the ord ered fields whose theory elimin ates th e
quantifiers in the language for ordered fields (the real closed fields) . One
easily deduces
Corollary 2.9.3 1. P -I N P over the field of rationale, or over the field
of reels (without the order relation).
2. P -I N P over the ordered field of rationale.
On the other side, quantifier elimination is only a necessary condit ion to-
wards P = N P over a given st ructure. There do exist quantifier elimina ble
struct ures A such t hat P =1= N P over A : t his is the conte nt of the following
nice resul t of Meer.
Theorem 2.9.4 P =1= N P over (R , +, - ).
In fact t he t heory of (R, +, -) is essent ially t he t heory of nonzero vec-
t orspaces over the rational field, and so admits t he elimination of quantifiers
in t he language LQ and even in {+ , -} because t he action of any rational
is easily defined by the addit ive st ruct ure without using quantifiers.
Wh at can we say about t he complex field, or t he ordered field of reals?
Their t heor ies eliminat e t he qua nt ifiers in t he corres ponding language. Nev-
ert heless P = N P is still an open question over these st ruct ur es. Notabl y,
in t he complex case, a key (NP-complet e) problem towards a definitive an-
swer is related to t he celebrated Hilbert Nullstellensatz (a classical algebraic
resul t closely relat ed t o Model Theory, as we will see in t he next chapter) :
in fact, it asks a quick procedure checking t he solvability of a given finit e
system of polyno mials over C (in arbit rarily many variables) . One shows
t hat P = N P over t he cornplex field if and onl y if t his fast procedure exists.
N P -com plete problems over the ord ered field of reals are discussed in t he
references quoted below.
2.10 References
Van den Dri es [168] and Doner-Hodges [34] are two excellent and enjoyable
exposito ry papers, explaining Tarski's work on t he quantifier elimination
and, mor e generally, t he hist or y of this matter. They also include a rich
list of references. Here let us mention [87] and [154] on t he pioneeristi c
contributions of Lowenheim and Skolem t o the elimination of quantifiers.
Langford 's elimination met hods for dense or discret e orders are in [81] and
in [82], while Tarski 's subsequent cont ribut ions in t his sett ing are in [160].
Tarski 's eliminat ion pro cedures for the real field and t he complex field are
given in [157], while Cohen's method in the real case is in [27]. [98] contains
t he Macintyre-McKenna-Van den Dries t heorems saying that algebraically
closed fi elds are t he only infini t e fi elds whose first order t heory eliminates t he
quantifiers , and real closed fields are t he onl y ord ered fields wit h t he same
feat ure. Let us menti on t hat R. Thorn coined t he word "semialgebraic" in
[162].
Some more details about pp-elimination in modules (and a proof of Neu-
mann's Lemma 2.6.3) can be found in M. Prest 's book [136]; t he Eklof-Fisher
2.10. REFERENCES 83
paper [40] deals with t he particular case of abelian groups (and modules over
Dedekind domains).
The computat ional aspects of the quantifier eliminat ion for the real field ar e
discussed in [33], while t he cylindrical algebraic decomposition algorit hm
CAD is in the Collins paper [29]. The Fisher-Rabin t heorem ensuring that
no decision algorithm for the real field runs in polyn omial t ime is in [47].
[13], [14] describe th e new Blum-Shub-Smale model of computat ion; [133]
provides Poizat 's approach to this th eme. K. Meer 's th eorem th at P =J N P
over (R, +,-) (although the corresponding t heory eliminates t he quanti-
fi ers) is in [112] .
Chapter 3
Model Completeness
3.1 An introduction
We already defined model completeness in Chapter 1: a t heory T is called
model complete if every embedding bet ween models of T is elementary. We
dealt with this notion also in Chapter 2, where we considered it s connection
with quantifier elimination and completeness. But now we wish to examine
model completeness in a closer and more dir ect way, t o discuss it s genesis
and motivations, as well as it s importance and applications.
Model completeness deals with embeddings between st ruct ur es. This per-
spect ive might look slight ly oblique with res pect t o t he fundamental purpose
in mod el t heory, namely to connect sentences and struct ures via t ruth; under
this point of view, the most genuine relation among st ruct ures is elemen-
tary equivalence (that is, to satisfy the same sente nces). Nevert heless some
basic theorems in model theory, such as t he Lowenheim-Skolem t heorems,
involve pret ty naturally extensions, subst ruct ures, embeddings, and so draw
attenti on t o thi s subject . Furthermore, as we will see in Section 3.2, th ere
are other possible ways of introducing model completeness. The first one
st ill deals with embeddings and says that a theory T is model complete
when each embedding between models of T preserves existential formulas.
But anot her cha rac terizat ion is quite syntactical and resembl es the way we
defined quantifier eliminat ion; it says that a t heory T is model complete
exactly when any formula !.p(v) in the language of T is equivalent in T to an
appropriate existential formula !.p' (v).
The main motivations leadin g t o model completeness come from algebra. For
instance, consider field th eory. Given a field K , one looks at t he irreducible
polynomials f( x) E I<[x]. Algebra builds richer and richer extensions of
85
86 CHAPTER 3. MODEL COMPLETENESS
J( , equ ipping t hese polynomials with a single root , or all the possible roots.
Eventually, one reaches the algebraic closure J( of J( : a minimal extension
where every nonconstant polynomial f(x) in I< [x], and even in I<[x ] it self,
splits into linear factors, and so get s its own roots. Notice that, from a
logical poi nt of view, adding a root of a pol ynomial f( x) means t o satisfy the
sent ence 3w(f (w) = 0) with paramet er s from I< (the coefficients of f( x)) . J(
is algebraically closed when it equals J( and hence when it is able to sati sfy
all t hese sentences when f( x) ran ges over t he nonconstant pol ynomials over
I< it self. Pursuing t his logical approach, one ca n generalize and look at
a rbit rary L-structures A instead of pure fields, towards t wo possible objects:
* to enlarge A to a richer A sat isfying every existent ial sent ence 3wa(w)
(with a qu antifier free a(w)), or even ever y sent ence in L(A), or (why
not?) in L(A) , t oo;
* to examine closely t he structures A.
This program recalls A. Weil's notion of universal domains in [176]. Weil 's
idea (for t he class of fields) was t o determine large and rich struct ures, em-
bedding every fi eld under consider ation. Of course , in the cas e of fields ,
universal domains are just algebraically closed fields of infinite transcen-
dence degree. This strategy has now fallen into disuse within Algebraic
Geomet ry, but it is still alive in Model Theory (and certainl y it was in t he
sixt ies) . Model completeness arises quite naturally in this fr amework: for , in
a model complet e theor y T , for eve ry model A and for every L(A)-sent ence
<p , whenever <p is t rue in some model extending A, t hen A it self satisfies <p;
so, it is wor th devoting some speci fic pages to t his matter. T his is what
we will do in t his cha pter. First we will give an abstract an alysis of model
complet eness. Then we will emphasize it s strong connection with Algebra.
In fac t , Algebra inspires the notion of model complet eness , and several re-
lat ed concepts; but, conversely, we will see that some developments in Model
Theory concern ing model completeness do produce a significant progress in
Algebra; indeed some alternat ive elegant proofs of the celebrat ed Hilber t
Nullstellensatz, or of th e I-lilber t Sevent eenth Problem, and, more not abl y,
t he solution of Artin's Conj ecture on p-adic fields wit ness t hese fr uit ful con-
t ributions . Act ually, t his was the dream of Abraham Robinson (t he father
of model completeness) : t o quot e his own words in his address t o the 1950
ICM,
"Symbolic Logic can produce useful tools for the developments of
actual mathematics, more par ticularly of Algebra and, it would
3.1. AN INTRODUCTION 87
appear, of Algebraic Geometry. This is the realization of an
ambition ... expressed by Leibniz in a let ter t o Huyghens as long
ago as 1679".
This point of view is developed one year later in [140]. The algebraic theo-
rems recalled before do corroborate t his program. Other deep confirmations
(also concerning Geometry) will be provided in the next cha pters.
Let us conclude t his sect ion by recalling some connect ions between model
complet eness, elimination of quant ifiers and complet eness.
First of all, remember t hat elimination of quantifiers impl ies model complete-
ness. The converse is not true. For instance, we saw in t he last chapt er that
the t heory of real closed fields RCF loses t he quant ifier eliminat ion proper t y
if one removes t he relation symbol for ~ from it s language L: act ually the
order ~ is definable in the rest rict ed language L
o
= {+, - , " 0, 1}, as v ~ 0
is equivalent in RCF t o
:3w( w
2
= v),
but any possible definit ion needs quantifiers. However t o forget ~ does not
affect model complet eness: in fact every embedding f : A -+ B of real
closed fields in the rest ricted language L
o
enlarges naturally and involves ~
(beca use t he nonnegative elements mus t equal t he squares) , so is elementary
in both L and L.
On t he other side, model completeness can yield complet eness under some
suit able additional hypotheses. We saw t hat this happens, for inst ance,
for real closed fields (or also for discret e linear orders) . The reason was
that RCF has a "minimal" model , emb eddabl e in every real closed field:
the ordered field of real algebraic numbers (hence t he real closure of t he
rat ionals) . To extend this example t owar ds a general set t ing, we need t he
following
Definition 3.1.1 Let T be a theory. A model of T is prime if it is embed-
dable in every model of T.
Examples 3.1.2
of ACF
o
.
1. The (complex) algebraic numbers are a prime model
2. The real algebraic numbers are a prime model among real closed or-
dered fields.
3. (N, ~ 0, s) is a prime model of dLO+.
Proposition 3.1.3 Let T a model complete theory. 1fT has a prime model
A , then T is complete.
Proof. Just adapt the argument of RCF. For every model B of T , there is
an embedding of A into B. Owing to model complete ness, t his embedding
is elementary. In part icular B is elementarily equivalent t o A. Hence all th e
models of T are elementarily equivalent to each other. Consequently T is
complet e. '"
Another useful criterion deducing complet eness from model completeness is
t he following one (we used it when dealing with dense linear ord ers in t he
last chapter).
Proposition 3.1.4 Let T a model complete theory. Assume that any two
models A and B of T admit a common extension C in M od(T). Then T is
complete.
Proof. Given A and B, form a common ext ension C. Owing to model
completeness, t he embeddings of both A and B in C ar e elementary. So A
and B are elementarily equivalent t o C, and consequent ly to each oth er . '"
However , be careful: model completeness does not imply completeness in
general. Just to avoid any temptation about t his point, recall algebraically
closed fields. ACF is model complete, but it is not complet e: one needs t o
specify the characteristic to get a prime model and hence t o ded uce, even
via model completeness, that ACF
p
is complet e for every p = 0 or prime.
3.2 Abraham Robinson's test
Let T be a t heory in a language L . We know t hat T eliminates t he quantifiers
if and only if every L-formula is equivalent in T to a suitable quantifier
free formula (with the same free variables). Model complet eness can be
characterized in a similar way. Indeed one can show that T is model complete
if and onl y every L-formul a is equivalent in T to an existential formula, or
also, if you like V rat her than 3, to a universal formula (in fact , assume
that every L-formula ip(v) admit s an exist ent ial L-formula equival ent in T ;
apply this property to 'ip(v) and yield the corresponding existential formula
ip' (v); conclude that ip(v) is equivalent in T to ' ip'(v) , which in it s t urn is
obviously equivalent to a universal formula; the converse can be shown in
t he same way) .
There is another remarkabl e related characterization of model completeness
via embeddings. Recall th at T model complete just means that every em-
bedding between two models of T is element ary. But , notably, this is also
3.2. ABRAHAM ROBINSON'S TEST 89
equivalent to require that every embedding between two models of T is
existential (at a first sight , a weaker condition). This is the content of the
so called Abraham Robinson Test for model complet eness. We will apply
t his crit erion t o some algebraic set t ings in the next section. Now we want to
show Robinson 's Test and , at th e same t ime, the previous charact erizat ions
of model completeness in t erms of existential , or universal formulas.
Theorem 3.2.1 (A. Robinson) Let T be a theory of L. The following propo-
sitions are equivalent:
(i) T is model compl ete;
(ii) every embedding from A into B, where A and B are models of T, is
existential;
(iii) for every L-formula (ii) is clear , and (iii){:} (iv) was alr ead y established.
Let us consider now (ii)=>(iii) ,(iv) .
We preliminarily show that , if (ii) holds, t hen every existent ial formula <p (if)
admits an equivalent universal formul a, and conversely. So assume <p(if)
existe nt ial. Let ti be th e lengt h of if. Look at t he set S of the universal
formul as 0"( if) in L satisfying
Vif(<p(if) -t O"(if)) E T.
Notice that S is closed under conjunctions 1\ (up to straightforward ma-
ni pulations) . What we need is a formilia 0" (if) in S sat isfying the further
condition
Vif(O"(V) -t <p (if)) ET;
in fact, in this case, er( if) is a universal formula equivalent to <p (if) in T , and
we are done. Suppose towards a cont radiction that no er( if) E S works. We
can express t his assumption in t he following way: in a language L' ext ending
L by a sequence cof n new constant symbols,
for every er(if) E S, T U {O" (C) , -' <p( C)} has a model
(otherwi se, for every mod el M of T and every min cr(Mn), mE rp (Mn),
and so V' v(cr(V) -+ rp(v)) is in T ). Recall t hat S is closed under conjunctions
and use compactness t o dedu ce t hat
T U{cr(C) : er(V)ES}U{. rp (C)}
has a mod el. In oth er words , t here are a model A of T and a sequence ii in
An (th e int erpretation of cin A) such t hat
ii E n a (iI)ES cr(A
n
)
but ii ~ rp (An). Put ii = (aI, . . . , an)' Now take any quantifier free for-
mul a O(V, w) in L and suppose that t here is a sequence bin A for which
A F O(ii, b). Let h denote the lengt h of w(and of b); in particul ar put
b= (b
l
, . .. , bh). Then V'W.O(V, w) is a universal L-formula, but ii cannot
sat isfy it . Hence V'W.O(V, w) ~ S . It follows t hat , in some model A* of T ,
a suitable sequence a* in rp (A*n) does not sat isfy V'w.OCu, w); hence, for
some b-;, A* F O(a* , b-;). We can express t his fact in t he following way.
Consider t he language L (A) ; t o avoid any danger of confusion, distin guish
t he elements of A and t heir names in L (A), and denot e by Ca th e constant
symbol corresponding to a, for every a E A. Wh at we have j ust shown is
t hat
TU {rp (c
aIl
. .. , can)} U {O(cap . . . , Can' Cb I' .. . , Cb
h
)}
has a model. Again using compact ness and t he fact t hat t he quantifier
free formul as of L are closed under conj unction, we see t hat t he union of
TU {rp (Ca! , . . . , Can)} and of th e whole collectio n of t he for m_ulas O(Ca!, ... ,
Can ' Cb!, .. . , Cb
h
) wher e h ranges over non negati ve int egers, b = (b
1
, ... , b
h
)
over Ah, O( v, w) over quantifier free L-formulas and A F O(ii, b) , has a
model. Let A' denote th e restriction of this st ruct ure to L -hence a model of
T-, and let a
'
= ct for every a E A. Therefore, for every quantifier free L-
formulaO(v, w) and every bin A, A' F O(;;} , hi) if and only if A F O(ii, b) . So
a f---7 a
'
for every a E A embeds A into A' . (ii) impli es t hat this embedding
is existe nt ial. Consequently, as A' F rp (a
l
), one dedu ces A F rp( ii) -a
cont radi ction- . So a universal er(v) equivalent to rp (v) exists .
The converse redu ction (from universal to exist enti al formul as) ca n be han-
dled by passing to negations.
At t his point , the proof is straight forward. Take an ar bit ra ry L-formula
rp (v). We are looking for an equivalent existenti al , or uni versal formul a. If
rp( V) has no quantifier , t hen t here is nothing t o prove, and we are don e.
Other wise writ e rp( v) as
QI ZI ... Qkzka (V, Zl , .. . , Zk)
3.3. MODEL COMPLETENESS AND ALGEBRA 91
where k is a positive int eger, the Qj's (1 ::; j ::; k) are quantifiers (\7' or
:3) and a(v, Zl, ... , Zk) is quantifi er free. Proceed by induct ion on k, using
what was preliminarily shown: t he det ails are an easy exer cise.
At last, let us show that (iv) *(i) . This is quite simple. We know t hat ,
if f : A -+ B is any embedding between model s A, B of T, then, for
ever y universal L-formul a <p (v) and every sequence a in A, B F <p (J(a))
impli es A F r.p(a). But (iv) ensures that any L-formul a is equivalent in T
to a suitable universal formula , and so enlarges the previous statement to
arbitrary for mulas. ..
As an immediate consequence of (i)*(iii) , let us point out the following
not eworthy fact.
Corollary 3.2 .2 In a model complete theory, the defi nable sets are just the
project ions of sets definable by quantifier f ree formulas.
3.3 Model completeness and Algebra
Model complet eness of AC F and RC F was shown in the last chapter as
a consequence of elimi nation of quantifi ers. Robi nson's Test provides a di-
rect proof (in these and in other relevant cases). Here we wish t o illust rate
this new approac h. However , the main object in this sect ion is to empha-
size t he role of model complete ness towards some noteworthy applications
to Algebra. We underlined in 3. 1 A. Robinson 's program, and his hope
t hat a progress in model th eor y could supply Algebra with new important
and fruitful tools and techniques. Mod el complete ness really exemplifies this
project. In fact , we will see that , just using the model complete ness of ACF
and RCF , A. Robinson found neat and elega nt proofs of classical res ults,
such as t he Hilbert Nullst ellensat z and the solution of Hilbert's Seventeenth
Problem (a theorem of Art in) . But we want to underline that the model th e-
oretic approach can provide not only alternative proofs of previously known
theorems, but also , and more not ably, new and original answers to some
formerly open famous algebraic probl ems, for instance Artin's Conjecture
on p-adi c fields (this will be treated in 3.4).
But now let us show t he model complet eness of ACF and RCF via the A.
Robinson Test , as promised.
Theorem 3.3.1 (A. Robinson) ACF is model complete.
Proof. Suppose not. Owing to the Robinson Test, there is some embedding
between algebraically closed fields which is not existential. Let f : K -7 1-l
be a counterexample. With no loss of generality, we can assume that K
is a su bfield of 1-l and f is just the inclusion (otherwise, replace K by its
isomorphic copy inside 1-l). There are a quantifier free formula cv(V, w) in
the language L = {+, - , " 0, I} and a sequence ilin K such that :3wcv(v, w)
is true in 1-l and false in K. Incidentally, recall that cv(V, w) is (equivalent
to) a disjunction of conjunctions of equations and inequations; however, as
:3 is distributive with respect to V, we can assume that cv(V, w) is directly
a single conjunction of equations and inequations. Let bsatisfy cv(il, b) in
1L Form the extension K(b) of K by i, its algebraic closure K(b) embeds
itself in H. and satisfies :3wcv(il, w) because it includes b. So there is no loss
of generality in replacing 'H. by K(b) and hence in assuming that 1-l = K(b)
has a finite transcendence degree (> 0) over K. Now take a transcendence
basis tl, .. . , t
s
of 1-l over K, and split the embedding K ~ K(tl , .. . , t
s
) by
where K, = K(t
1
, ... , ti) for every i = 1, ... , s. There is some i = 1, ... , s
such that :3wcv(il, w) is true in K, and false in Ki-l. We can replace K ~ H.
by Ki-l ~ x; and to assume
1-l = K(t)
for a single transcendental element t over K. So :3wcv(il, w) is true in K(t)
and false in K.
Now take any algebraically closed extension K' of K having transcendence
degree 2 1. Hence K' enlarges K(u) for some transcendental element u over
K. Steinitz's analysis of algebraically closed fields tells us that K(t) and
K(u) are isomorphic via a function enlarging the identity of K and mapping
t into u. Then K(u) and, consequently, K' satisfy :3wcv(ii, w): our sentence
is true in every algebraically closed extension K' of K with transcendence
degree 2 1. Equivalently, in the language extending L(K) by a new constant
c, :3wcv(il, tu) is a consequence of Th(KK) plus the infinitely many sentences
ensuring that c is transcendental over K, i. e. does not solve any nonzero
polynomial with 1 unknown and coefficients from K. Use compactness and
deduce that finitely many sentences suffice (in addition to Th(KK)) to imply
:3tucv(ii, w); in particular, c can be interpreted by a suitable element of K
(out of the roots of a finite system of polynomials in K[x]). In conclusion, K
satisfies :3wcv(il, w): a contradiction. Hence ACF is model complete. ..
Which are t he main algebraic ingredients of thi s Robinson proof? Basically,
Steinitz 's analysis of algebraically closed fields . More specifically, two key
points should be underlined:
1. every field has an algebraic closure, and this is unique up to isomor-
phism enlarging t he identity in the ground field;
2. if K is an algebraically closed field and t is transcendental over J( ,
then t he isomorphism class of K(t) over K is uniquely det ermined (in
t he sense that two ext ensions of this kind are isomorphic via a map
extending the ident it y of J().
One can realize t hat real closed fields satisfy similar propert ies:
1. every ord ered field has a real closure (a minimal real closed exten-
sion), and this is algeb raic over t he ground field, and unique up to
isomorphism enlarging th e identity in the ground field;
2. if K is a real closed ord ered field and t is t ranscendental over J( , th en
the isomorphism class of K(t) over K is fully charact erized by t he cut
t det ermines over J(.
So, when dealing with model completeness for RC F, one can reproduce the
proof of the algebraically closed case (with some complicat ions due to th e
order) and deduce
Theor em 3.3.2 (A . Robinson) RCF is model complete.
Robinson 's Test can also be used t o prove th e mod el complet eness of sever al
theories we met in t he previous chapter : discre t e linear orders, dense linear
orders and so on. The reader may check this, as an exer cise. But here we
prefer to discuss some very noteworthy applications of t he model complet e-
ness of ACF and RC F to Algebra. They provide new elegant proofs of
known algebraic facts .
First let us deal with algebraically closed fields and Hilbert 's Nullstellensatz.
Theorem 3.3.3 (Hilbert Nullst ellensatz) Let K be an algebrai cally closed
field, J be an ideal of the ring K[X] (where x abridges, as usual, the sequen ce
of unknowns (Xl , ... , x
n
) ) . Th en, fo r eve ry polynomial f( x) E J([X] ,
f(ii) = 0 f or every d E K" such that g(ii) = 0 fo r all g( x) E I
if and only if
94 CHAPTER 3. MODEL COMPLET ENESS
for some positive integer m f m(x) El .
Proof. The direction from right t o left is clear. Conversely, suppose t owards
a cont radict ion t hat t here exists some polynomial f( x) in K[x] such that
f (ii) =0 for every ii E K" such t hat g(ii) =0 for all g(x) E I
but
f m(x) rJ. I for every positi ve int eger m.
Let J be an ideal of K[X] such that J ;2 I , no power 1
m
(x) of 1(x) (with m a
positive int eger ) lies in J, and J is maximal with respect t o th ese conditions
(Zorn's Lemma ensures that such a J exists). We claim that J is prime. In
fact, take two polynomi als go (x) , gl (x) in K[X] - J; t hen t he ideals
J
o
generated by J and go(x) ,
J1 generat ed by J and gl (x)
st rict ly include J ; accordingly t here are two positi ve int egers mo , m 1 such
t hat f m
o
(x) E J
o
, I": (x) E J
1
. So t here exist two polynomi als qo (x),
q1(x) E K[X] such t hat
F" (x) - s.(x) qi(x) E Ji' Vi = 0,1.
Consequent ly
r-r: (x)
is in t he ideal gener at ed by J and go(x) . gl (x), and so go(x) . gl (x) rJ. J .
Hence J is prime, and R = K[X] / J is an int egral domain exte nding K
by th e function mapping any a E K t o a +J. Th en K embeds into the
algebraic closure F of the field of quotients of R . As AC F is model com-
plet e, this embedding is elementary. Now take any (finit e) set of generators
10(x), ... , 18(X) of I (I is finit ely gene rated because K[X] is Noetherian),
and noti ce that the L(K)-sentence
::I v (/\ f i(V) = 0 1\ -, (J (v) = 0))
i ~
is t rue in F (owing t o t he seque nce (X l +J, ... , X
n
+J )). Consequent ly t his
sentence is t rue also in K. So t here exists some ii in K" such t hat ii satisfies
10(x), ... , f 8(X) and consequent ly all the polyno mials in I , but ii does not
annihilate f (x) -a contradictio n-. ..
Now we deal with RCF, and with Hilbert 's Seventeenth Problem. This was
solved by Artin in 1927. Indeed Artin himself and Schr eier developed t he
algebraic notion of real closed field just to answer Hilbert 's question. Later
A. Robinson proposed a very nice and simple proof, founded on th e model
complete ness of RCF. Here we want t o report A. Robinson's approac h.
First let us introduce Hilbert's problem in detail.
Indeed the seventeenth question in the celebrated Hilbert 1900 list just con-
cerns ordered fields (more properly, the ordered field of real s ) . Recall t hat,
in any ordered field K , a rational function f (x) E K (x) is said t o be semidef-
inite positive if and only if, for every sequence ii in K (such that f(ii) is
defined) , f(ii) ~ O. Of course, t he sums of squares in K(x) are semidefinite
positi ve. I-Tilbert 's Seventee nt h Problem conject ures t hat the converse is
also true when K is t he ordered field of real numbers. As already said, Art in
solved positively t his question; indeed he extended th e result to arbit ra ry
real closed ordered fields K. Now we provide A. Robinson 's proof of t his
t heorem.
Theorem 3.3.4 (Ar tin) Let K be a real closed ordered fiel d, f (x) a semidef -
inite positive rat ional funct ion in K(x). Then f (x) can be expressed as a
sum of squares in K(x).
Proof. We need t he following algebraic fact .
Fact 3.3.5 Let K be a fiel d, and assume that, fo r every natural t and for
every choice of aa, .. . , at E K , if L i<t at = 0, then aa = . . . = at = 0 (su ch
a fie ld K is usually called formally real). Let a be an element of K such that
a cannot be represented as a sum of squares in K . Th en there exists a total
order:::; in K making K an ordered fiel d and a < O.
Now let us begin our proof. Take a semidefinit e positive f(x) E K(x) .
Suppose t owards a contradiction that f(x) cannot be expressed as a sum
of squares. Clearl y K(x) is a formally real field. So, owing t o Fact 3.3.5,
there is some total order relation :::; in K(x) with respect t o which K(x)
becomes an ordered field and f( x) < O. Notice that this order relation
:::; in K( x), when rest ricted to K , does equal the primitive order of K: in
fact, in both these relations, t he non-negative elements of K are j ust the
squares in K . In ot her words, K is a substruct ure of K (x) in our language
for ordered fields . Recall that every ord ered field admits a minimal real
closed extension (it s real closure) , and accordingly embed K(x) into the real
closur e R. Alt oget her we obtain an embedding of K into R. As RCF is
model complet e, t his embedding is elementary. But K( x) , and consequentl y
R , satisfy t he L(K)-sent ence
3iJ(f (iJ) < 0)
(owing t o t he sequence x). Accordingly also K sat isfies 3iJ(f(iJ) < 0). But
t his cont radicts our ass umption that f( x) is semidefinite positive.
3.4 p-adic fields and Artin's Conjecture
Real numbers complete the rational field with respect to the usual metric
t opology. For every prime p, p-adic numbers complete th e rationals with
res pect t o an alt ernat ive t opology, called p-adic. Let us short ly remind th is
t opology. First write (uniquely) any nonzero rational a as
a = ph:
s
where h, rand s are int egers, s is positive and h, r , s are pairwise coprime;
t hen define a function v
p
from t he mul tiplicati ve group Q* of nonzero ra tio-
nals into t he ordered additi ve group of int egers by put tin g, for a as befor e,
One gets in th is way a gro up homomorphism from Q* int o t he integers,
satisfying t he addi tional condit ion
vp(a +b) ~ min{ vp (a), vp (b)} 'tI a, se Q*
(this is st raightforward to check) . v
p
can be form ally extended to 0 in some
a rt ificial way; putting vp(O) = 00 is a reasona ble choice, as ph divides 0
among th e integers for every natural number h. Now put , for every positive
integer h,
Oh = {a E Q : v
p
( a) ~ h}
and take the Oh'Sas basic open neighbourhoods of 0, hence their translations
Oh + b wit h b E Q as basic open set s. One gets a new topology of th e
rationals: t he p-adic topology. As already said, t he set Qp of p-adi c numbers
is t he complet ion of Q with respect t o t his t opology.
In order to realize as well as possible what a p-adi c number is, and so to
introduce Qp in a more det ailed way, one can follow several equivalent ways:
see [36] for a general outli ne of t his point . Here we limi t ourselves to sketch
3.4. P-ADIC FIELDS AND ARTIN'S CONJECTURE 97
some basic ideas. A possible approach to p-adic numbers might use the well
known fact t hat every positive integer a has a unique p-adic decomposition
where n and t he ai's are natural numbers, an =1= 0 and a; < P for all i ::; n.
Under this perspective, vp(a) is just the least i ::; n for which a; =1= o. Hence,
for every nonzero natural a
n
+l < p, it is t rivial to realize
So, when considering Cau chy sequences of int egers with respect to the p-adic
t opology and equipping such a sequence with a limit , one naturally builds
infinite sums
where the ai'S are non negative integers < p.
Enlarging the analysis from positive int egers to arbitrary nonzero rationals
leads to consider general infinite sums
00
(*) L aipi
i=N
where the ai'S are, as above, natural numbers < p, aN =1= 0 but N is now
any fixed -even negat ive- integer: in this sense (*) exhibits a typical p-adic
number. 0 can be easily recovered in this framework as the infinite sum
whose coeffi cients are constantly O.
Hence, for every prime p, Qp is t he set of t hese infinite sums. One defines
in a suitable way addition + and multiplication in Qp, extending the usual
operations in Q. But here we have to be careful: sum and product are not
computed componentwise, but as suggested by t he algebraic framework. For
inst ance, the trivial identity 1 + (p- 1) = P must be read
(1 + 0 . p) + ((p - 1) + 0 . p) = 0 + 1 p.
However t hese operations equip Qp with a field structure, extending the
rational field; v
p
can be enlarged to Qp in very simple way, as, for a =
I : ~ N aipi as before (and aN =1= 0) , vp(a) is just N. One sees t hat
is a local subring (t he ring ofp-adic integers), and
Ip ={aE Qp : vp(a O}
is it s maximal ideal. One easily checks that the quo ti ent field Zp/ Ip is
isomor phic to the field Fp with p elements .
We want to underline t wo fur th er basic properties of Qp.
(a) The first one is quite trivial, and simply point s out that vp (p) = 1.
(b) The other is mor e subst ant ial and concerns the so called Hensel's
Lemma. This is a key result in locating root s of polynomials in Qp, in
fact it states t hat,
if f(x) is a monic polynomial in Zp[x], t hen any decomposi-
tion of f( x) modulo I
p
in Fp[x] as a product of two relat ively
coprime monic polynomials lifts to a decomposit ion of f( x)
as a product of two mon ic polynomials in Zp[x].
This concludes our short summary about th e algebraic struct ure of Qp for
every prime p. What we have sket ched suggests some similarit ies with the
reals. Actually both Qp and R have a common topological genesis from t he
rationals (and, under this point of view, a common topological st ruct ure of
locally compact field); mor eover th ere do admi t some reasonable crite ria to
locat e roots of polynomials (t he Sign Change prop er t y for t he reals, Hensel's
Lemma in t he p-adic case ).
Now we want to discuss another example, closely recalling Qp. We take any
field K (but below we will be pr imarily interested in the field Fp with p
element s for any prime p). We look at the formal Laur ent seri es
00
a(t) = L: ai
i = N
where a; E K for every i ~ N and N is a given int eger. The cor responding
set K((t )) inherits a fi eld st ruct ure extending K , provided that we define
the addition + componentwise and the mult iplication . in the obvious way
enlarging t he product in K . Again th e set K [[t]] of formal power series
00
L: ai
i = O
3.4. P-ADIC FIELDS AND ARTIN'S CONJECTURE 99
is a local subring, whose maximal ideal just contains the power series with
ao = O. One eas ily deduce that t he quotient ring is (isomorphic to) K.
The function VK from the mul tiplicative group K ((t ))* (where ", as befor e,
means t o exclude 0) into t he (ordered) additive group of integers sending
any nonzero Laurent series a(t) as before (with aN i= 0) to N again yields a
group homomorphism sharing with v
p
and Qp th e following property
VK(a(t ) + b(t)) 2 min{vK(a(t )), vK(b(t ))}
for all a(t) and b(t) in I<((t)) *.
Now assume K = Fp for a given prime p. In this restricted framework
(a) VFp (P) i= 1. Notice t hat th is distinguishes Fp((t)) and Qp.
(b) However Fp((t )), just as Qp, satisfies Hensel' s Lemma.
An abst ract not ion including p-adics as well as formal Laurent series and
other relat ed examples t owards a common gener al t reat ment is the concept
of val ued field: this is a str uct ure (K , 9, v) where K is a field, 9 is an ordered
a belian group, and v is a group homomorphism from th e mult iplicative group
K* in 9 satisfying the further ass umpt ion
v(a +b) 2 min{ v(a), v(b)} 'tf a, s 1(*.
The function v is called the valuation map. A general algebraic analysis
promptl y confirms some basic proper ti es observed in th e previous examples:
in par ticu lar , for every valued field (K, 9, v), {a E I< : v(a) 2 O} is a local
subring (the val uation ring) , and it s maximal ideal is just {a E I< : v(a) >
O} ; th e corr esponding quotient field is called t he residue field of (K , 9, v ),
and will be denoted by K below (hence all throughout t his section K denotes
residue field rather than algebraic closure).
A valued field is called Henselian when it satisfies Hensel's Lemma, hence
when the following holds:
Let f( x) be a monic polynomial in R [x] and let l(x) denote its
projection in K[ x]. Th en any decomposition of](x) in K[x] as a
product of two relatively coprime monic polynomials
f( x) = ,(x) . A:( x)
lifts to a decomposition of f( x) into the product of t wo monic
polynomials in R[ x]
f (x ) = g(x ) . k (x )
where g (x ) = ,(x) and k(x) = A: (x ) .
For every prime p, both (Qp , Z, v
p)
and (Fp((t)), Z, VF
p
) are Henselian
valued fields (although they do not admit the same characteristic, and hence
their valuation of p is not the same).
Now let us come back to our original framework and hence to p-adic numbers.
Here Artin proposed a famous conjecture.
Conjecture 3.4.1 (Artin's Conjecture) Let p be a prime. For all positive
integers nand d with n > d
2
, every homogeneous polynomial f(Xl ' .. . , x
n
) E
Qp[Xl' ... , x
n
] of degree d has a nonzero root in Qp.
The conjecture is inspired by the underlined resemblance between Qp and
Fp((t)) for every prime p. Actually in the valued field Fp((t)) the claim is
true for every choice of p, nand d, as proved by Lang. But the behaviour
of the p-adics in this setting is not the same, in fact Artin's Conjecture fails
for some p, nand d. This was observed by Terjanian, who in [161] did
exactly what one is expected to do in disproving a statement, and exhibited
a counterexample for some suitable p, nand d.
However an asymptotic form of the conjecture is true: for any choice of n
and d, Artin's statement is satisfied by all but finitely many values of p. This
was the content of a celebrated theorem ofAx, Kochen and Ershov in 1965,
which, combined with Terjanian's counterexample, provides a sufficiently
complete answer to the question. The Ax-Kochen-Ershov approach is es-
sentially model theoretic. Indeed, they developed a general analysis of the
model theory of valued fields, and deduced the asymptotic form of Artin's
Conjecture as a consequence, using compactness and transfer techniques.
Let us briefly survey their work.
First of all, we have to clarify how to handle valued fields from a model the-
oretic point of view: which language to use, and so on. The more convincing
approach views a valued field as a two-sorted structure, in other words as a
structure with two sorts of variables, where
* the former sort of variables concerns the elements of the field,
* the latter sort of variables is devoted to the elements of the valuation
group;
moreover there are the usual field symbols for the former sort , and (dis-
jointly) the symbols of ordered groups for the latter; finally a l-ary oper-
ation symbol v is reserved for the valuation map. Valued fields are easily
axiomatized in a first order way in this language.
This approach emphasizes, within valued fields (!C, 9, v), the role of three
underlying structures:
3.4. P-ADIC FIELDS AND ARTIN'S CONJECTURE
the ori ginal field K ,
the ordered abelian group Y,
and finally, to capture the valu ation map v,
t he residu e fi eld K.
101
Ax, Kochen and Ershov show t hat t hese st ruct ures rul e the behaviour of
t he whole valued field (K, y, v) with resp ect to elementary equivalence . In
fact , t heir main general result says
Theorem 3.4.2 Let (K , y, v) be an Henselian valued fi eld, whose residue
field has characteristic o. Th en the complet e theory of (K , y, v) is fully
det ermined by the complete theories of the ordered valued group Y and the
residue fi eld K.
Warning: The theorem does not apply directl y t o Qp or t o Fp((t)) for
any prime p because t hese valued fields do not respect t he hypothesis on
t he characterist ic of the residue field. Nevert heless t he Ax-Koc hen-Ershov
main t heorem is enough t o t hrow a bridge bet ween t he valued fields of
Laurent series Fp((t)) and t he p-adic valued fields Qp with res pect to Ar t in's
Conjecture, and to deduce its asymptotic solution.
Theorem 3.4.3 (Ax-Kochen-Ershov) For all posit ive in tegers n , d wi th
n > d
2
, for all but fi nit ely ma ny primes p, every homogeneous polynomial
f( XI' . . . , x
n
) E Qp[x }, . .. , xnJ of degree d has a nontrivial root in Qp.
Proof. (Sketch) We can limit ourse lves t o treat t he case n = d
2
+ 1.
Other wise put Xi = 0 for n > i> d
2+1
and work wit h X l , , xd2 +l : up to
rearranging t he indices of our unknowns, we can ass ume t hat what we get
in this way is a homogeneous polynomial of degree d in X l, . . . , Xd2+I , and
every nontrivial zero of this polynomial clearly produces a nontrivial root of
f(XI ' ... , x
n
) . Needl ess t o say, for any fixed d (and n), Ar tin's Conjecture
becomes a first order sente nce Cid in t he language of valued fields: t his is
a rou tine exe rcise, easy to check. We know from Lang t hat t his sentence
Cid is t rue in t he valued field (Fp((t)) , Z, VF p) for all primes p. This fact ,
in part icular it s uniform validi t y for every p, is not ewor thy; act ually, we
are in a sit uation qui t e similar t o t he one described in Chapter 2, 4, for
algebraically closed fields. In part icular , one ca n apply th e same transfer
machiner y, combine the Ax-Kochen-Ershov main theorem and Lang's resul t ,
and in conclusion deduce that Cid is true in every valued field (K, y, v) such
that
102
* K is I-I enselian ,
CHAPTER 3. MODEL COMPLETENESS
* 9 is elementarily equivalent (as an ordered group) t o t he integers (9 is
called a Z-group in t his case) ,
* t he residue field K is a pseudofinite field of characterist ic O.
Of course, this does not concern directly any p-ad ic field Qp, because it s
residue fi eld has characterist ic p. But , in proving ad, onl y finite ly many
sentences about the characterist ic 0 of th e residue fi eld are necessary. This
impli es that a d is true in Qp for almost all primes p, as claimed . '"
Owing to the Terjanian counterexample quoted befor e, this Ax-Kochen-
Ershov answer t o the Artin Conject ure is the best possible.
None of these result s refers directly to mod el complete ness. Nevertheless
the spirit and the techniques of t he model theor etic approac h ofAx, Kochen
and Ershov clearly owe to A. Robinson 's ideas, and are int imately related
to his dr eam of linking Algebra and Mat hematical Logic via Mod el Theory.
One should also rememb er t hat t he Ax-Kochen-Ershov main t heorem does
not apply t o t he p-adic fields Qp, because, as already observed, the char ac-
t eristi c of t heir residue fields is not O. So one may wonder, for every pr ime p,
how to cha racterize Qp, and even it s t win Fp((t )), up to elementary equiva-
lence, in other words how t o axiomatize in a first order way t heir complet e
theories (incidentally recall t hat Th( Qp) -I Th( Fp((t)), because t he involved
charact eristics are not equal) . Here model cornplet eness sounds useful. In-
deed, we already underlined t hat p-adics and reals share seve ral relevan t
similarit ies: model complet eness, and a precise first ord er axiomat izat ion,
are among them. In fact , t he following theorem holds.
Theorem 3.4.4 (Ax-Kochen-Ershov) For every prime p, let T
p
be the the-
ory of the valued fields (K, g, v) such that
* K is Henselian and has characte ristic 0,
* 9 is a Z-group,
* the residue fi eld K is (elementarily equivalent to) the fi eld with p eleme nts,
* v( p) is 1 -the least positive element in G- .
Then T
p
is model complete and complete. In part icular T
p
is the theory of
the valued fi eld of p-adic numbers.
3.5. EXISTENTIALLY CLOSED STRUCTURES 103
What can we say about definable set s in Qp, or also about quantifier elimina-
tion for p-adics? Maci ntyre showed quantifier elimination in a very natural
language with addit ional relation symbols for t he valuation ring and the set
of n-th powers for every n. This provides, of course, some mor e information
on definable set s; in particular, one can see that the p-adics sat isfy
Theorem 3.4.5 (Macintyre) Eve ry infi nit e definable subset of Qp has a
nonempty int erior.
Notice that t his is exact ly what happens for the reals. Finally, let us remind
that T
p
is not t he t heory of Fp((t)); actually, the question of det ermining
the first order theory of this valued field seems still open.
3.5 Existentially closed structures
Example 3.5.1 The class of fields and t he subclass of algebraically closed
fields satisfy t he following properti es:
(i) every field embeds into an algebraically closed fields (for instance, into
it s algebraic closure);
(ii) every embedding between algebraically closed fields is elementary (in
other words, the theory ACF is model complete ).
Consequentl y (in our language L for fields)
(iii) for every embedding of an algebraicall y closed fi eld /C int o a field 1l ,
and for every quantifier free for mula <p ew) of L(K) , if 1l F :Jw<p(w) ,
then /C F :Jw<p( w) (in other words the embedding is existential).
In fact, owing to (i), 1l embeds in some algebraically closed field 1l ; as
1l F :Jw<p( w), also 1l sat isfies :Jw<p(w). Clearly /C embeds into 1l through
1l , and th is embedding is elementary because of (ii) . Hence /C F :Jw<p(w).
Not ice that (iii) is not a secondary property, but does include the definition
it self of an algebraically closed field /C. In fact , t he latt er requires that every
non-constant polynomial f( x) E K[x] has a root /C, or, equivalently, that /C
satisfies all the L(K)-formulas
:Jw(J(w) = 0)
(true in some extension of /C) . (iii) says that this still holds when we replace
few) = 0 by any quantifier free L(I<)-formula <p ew) (with arbitrarily many
variables) .
From an algebraic point of view, (iii) implies that any finite sys te m of equa-
tions
10(x) = 0, .. . , 1t(x) = 0,
with 10(x) , . . . , ft( x) E K[X), has a solut ion in our algebraically closed field
/C whenever it finds a zer o in some extension of /C: to see t his, just apply
(iii) to
<p (v) : 1\ 1i(V) = 0.
i <t
Example 3.5.2 A similar analysis can be developed in t he language L of
ordered fields with respect to t he subclass of real closed fields.
In fact
(i) any orde red field has some real closed extension (for instance, it s real
closure) ;
(ii) any embedding between real closed fields is eleme ntary (as t he t heory
ReF is model complet e).
(i) and (ii) imply, j ust as in th e previous case, that , in the language L of
ordered fields,
(iii) for every embedding of a real closed field /C into an ordered field 'H.
and for every quantifier free formula <p (w) of L(K), if}{ 1= ::!w<p(w) ,
t hen /C 1= ::!w<p( w).
Algebraically speaking, (iii) implies that , in a real closed field /C, any finit e
syste m of equ ations and disequations
10(x) = 0, . .. , 1t(x) = 0, 90( X) > 0, ... , 98( X) > 0,
with 10(x) , . . ., 1t(x), 90( X) , . . ., 98( X) E K[X), admitting a solution in some
ordered field extending /C, does have some solut ion even in /C.
So (iii) is not a minor property of real closed fields but includes their defini-
ti on it self, or, mor e precisely, their characterization saying that real closed
fields are just the ordered fields /C with no proper ordered algebraic exten-
sion: if 1( x) E K[x] has a root in some ordered extension of /C , then it finds
a solut ion also in /C.
Now consider any class K of structures ill a language L . The previous
examples suggest the following notion.
Definition 3.5.3 A structure A in K is existentially closed (e. c.)
if and only if every embedding of A in some B E K is existential (i. e.,
for every quantifier f ree L(A)-formula a(w), if B F :3wa(w), then A F
:3wa(w)) .
Let E(K) denote t he class of e. c. struct ures in K. Clearl y, among fields,
e. c. struct ures are just t he algebraically closed fields, and, among ord ered
fields, e. c. st ruct ures are just t he real closed fields.
Given a class K =1= 0, first we can wonder whether t here exist some e. c.
st ruct ures in K , and hence whether E(K) =1= 0. T he answer is posit ive, at
least under some simple condit ions on K. For inst ance, one can see what
follows.
Example 3.5.4 Let K be a class of struct ures . Assume th at , whenever
(I , ::;) is a tot ally order ed set and each i E I indicates a st ruct ure Ai E K
in such a way t hat, for i ::; j in J, Ai is a subst ructure of Aj , t hen t he
union A = UiE! Ai -as defined in Cha pt er 1- is still in K . In t his case , every
A E K embeds in some structure of E(K ); in parti cul ar, E(K) is not empty.
Notice also t hat, in t he examples above, K is elementary, as well as E(K) .
Accordingly one can wonde r whether , in general, if K is elementary, t hen
E(K ) is. The answer is negative.
Examples 3.5.5 (a) If K is t he -element ary- class of groups, t hen
E(K) is not elementary any more (thi s is a result of Eklof and Sa b-
bagh).
(b) If K is t he -elementary- class of commutative rings, t hen E(K ) is
not elementary any mor e (as shown by Cherlin) .
The pr oofs mix compactness and some algebraic facts. We shall provide
t heir det ails at t he end of t his section. It should be emphasized that , while
e. c. fi elds are fully characterized in a first ord er way, e. c. rings are
not . The same happens for gro ups. Conversely, among ab elian groups, e.
c. struct ures are an elementary class: t hey are exactly t he divisible ab elian
groups admitting infinit ely many elements of period p for every pr ime p;
t his is shown in [41], where e. c. closed modules over suitable rings are also
discussed.
However , take an element ary class K ; let T denote it s t heory. If E(K) is
element ary, t hen T* = Th( E(K)) is said t o be a model companion of T .
Clearly T ~ T* . Moreover t he following result holds (generalizing what was
observed in our start ing examples).
Theorem 3.5.6 (Eklof - Sabbagh) Let K be an elementary class of L-
structures, T = Th(K), T* be an L-theory containing T . Then T* is a
model companion ofT (and (K) = Mod(T*)) if and only if
(i) for every A E K, there exists some B E M od(T*) where A embeds
zn;
(ii) T* is model complete.
Existentially closed structures, and model companions, were intensively
studied in the sixties and in the seventies. Several -elementary- algebraic
classes K were considered, to check whether existentially closed structures
in K were or not an element ary class, to provide a satisfactory characteriza-
tion in the positive case, and to analyse their complexity in the negative one.
This was not (and is not) a barren and unproductive exercise: in fact, it
brought to light some very interesting classes of (existentially closed) struc-
tures, such as differentially closed fields. We shall treat them (and other
key examples) in the next sections; but we want to emphasize since now
that th e notion itself of differentially closed field -a quite algebraic concept-
arises for the first time within the framework of e. c. structures: no algebraic
treatment preceded the model theoretic approach. Moreover, it should be
pointed out that new interesting elementary classes of existentially closed
structures have been considered quite recently, for instance among fields
with a distinguished a ut omor phism. This framework, too, will be explained
later in this chapter.
Also in the negative case, when existentially closed structures cannot form
an elementary class , their analysis has some intriguing features. The main
purpose here is to understand the reasons of nonelementarity and, hence, in
some sense, to measure how complicated the class of e. c. objects is. To
illustrate this point , we show, as promised, that existentially closed groups
are not an elementary class, and the same is true for existentially closed
commutative rings (with identity), and we discuss bri efly these negative
results. First let us deal with groups.
Theorem 3.5.7 (Eklof - Sabbagh) Let K be the -elementary- class of groups.
Then (K) is not elementary.
Proof. We work in the language L = {-, -1 , I}. We need two preliminar y
facts. 9 denotes an e. c. group.
Fact 3.5.8 For every positive integer n, 9 admits some elements of period
n.
Proof. Take t he direct product 9' of 9 and a cycli c group of order n.
Clearly 9' has some elements of period n , in other words it sat isfies the
existenti al sentence 3w(w
n
= 11\!\d/ n,d=/=n'(W
d
= 1)). As 9 is e. c. , 9
satisfies the same condit ion, hence it has some elements as required.
Fact 3.5.9 Two elements of infi ni te period in 9 are conjugate.
Proof. This is a more delicate point , and refers to a theorem of Higman ,
B. H. Ne umann and H. Neu mann , building, for any two eleme nts a and b
of infinite period in 9, a group 9' extendi ng 9 and a n element t E C' such
that taCl = b. So 9' F 3w(waw-
l
= b). 9 inherits this property because
it is existentially closed .
Now ass ume (K) elementary, (K) = M od(T) for a suitable L-theory T .
Fact 2 says that , in a lan guage L' extending L by t wo new constant symbols
c and d,
T U{,(en = 1), ,(d
n
= 1) : nE N, n > O} F 3w(wcw-
l
= d).
By compactness t here is some posit ive int eger N for which
T U{, (en = 1), ,(d
n
= 1) : 0 < n ~ N } F 3w( wcw-
l
= d) .
Hence take an e. c. group 9 and two elements a, b in G of period N + 1,
N +2 respectively: owing to Fact 1, this can be done. Then a and b are
conjugate in C: but this clearly contradicts the fact that their periods are
differe nt. '"
Actually existent ially closed gr oups are a ver y complicated class: [91] con-
tains several results illustrating t heir wildn ess. In particular, we like t o
mention a noteworthy connection wit h t he word problem for groups.
Theorem 3.5.10 (Macintyre, Neumann) A finit ely generated group can be
presented with a solvable word problem if and only if it is embeddable in all
e. c. groups.
Now let us deal with rings.
Theorem 3.5.11 (Cherlin) Let K be the -elementary- class of commutative
rings with identit y. Then (K) is not elementary.
Proof. Throughout t he proof, ring abbreviates commutative ring with
identity. Accordingly we work in the language L = {+, -, " 0, I}. Again,
we need two pr eliminary fact s. R denotes an e. c. ring.
Fact 3.5.12 Th e nilradical of R (i. e. the ideal of nilpotent elements) is
0-definable in R in a un iform way (valid in any e. c. ring) .
Proof. We claim that , for every a E R, a is nilpot ent if and onl y if a
does not divid e any nonzero idempotent element . This is clearl y enough
for our pu rposes, because the latter condition can be easily written as a (n
existent ial) first order formula in L
:3w(v w = v
2
w
2
1\ ,(v w = 0)).
The implication from the left to the right is quite simple. For, let a divid e
some idempot ent e E R; as an = 0 for some integer ti 2:: 2, it follows
o= en = e. Conversely, ass ume that a is not nilpot ent. Form the quotient
ring R' = R[x]/I where I is t he ideal generat ed by the polynomial ax-a
2
x
2
Then R' ext ends R in the obvious way, by the embedding of R in R[x] and
the project ion of R[x] onto R' . In R' t he image I + a of a divides t he
idempotent 1+ ax; moreover 1+ ax =f 0, otherwise in R[x]
(*) ax = (ax - a
2
x
2)f(
x)
for some f( x) = L i<n f i xi E R[x]. A compa rison of the coeffi cients of t he
same degree in (*) yields
afo = a,
a
2
f i = af i+l Vi = 1, . .. , n - 1,
and, event ually,
a
2
in= O.
In particular a
n
+
2
= 0 -a contradict ion-. So t he image of a in R' divides a
nonzero idempotent ; as R is e. c., a satisfies t he same cond ition in R .
Fact 3.5.13 For every positive integer n, R contains some nilpotent a sat-
isf yi ng an+! = 0 and an =f O.
Proof. Again enlarge R t o suitable quoti ents of the polynomial ring R[x].
This ti me, for every positive integer n, form R' = R[x]/ I where J is the
ideal generated by x
n
+
1
One easily checks t hat R embeds in R' by t he map
a 1---7 I +a for every a E R ; moreover (I + x )n+l = I , while (I +x )n =f I.
Hence R' satisfies :3w(w
n+1
= 0 1\ ,(w
n
= 0)). As R is e. c. , t he same is
true in R.
Now we ca n conclude our proof. Again we use compactness. Enlarge L by
a new const ant symbol c and, in t he new language L
'
, look at
T' = Th([(K)) U {,(c
n
= 0) : n E N , n > O}U
3.6. DCF
o
109
Uf 3w(c w = c
2
w
2
1\ --, (c w = O))}.
Every finite T ~ ~ T' has it s own mod el: for , th ere is a positive integer N
such t hat
and t he latter set of sentences does admit a mod el: it suffices to take an
e. c. ring and interpret c in a nilpotent element a satisfying aN =f:. 0 (Fact
2 ensures that a exists ) . By compact ness T' has a model: this is a ring
sat isfying the same sente nces as e. c. rings , but it is not e. c. , because
it contains a nonnilpotent element (the interpretation of c) dividing some
nonzero idempotent. In conclusion, E(K) is not element ary, because it s
theory has non e. c. models. ..
Notably, nilpotents ar e t he key obstacle t o the non element arit y of the class
of e. c. rings. In fact , for the restrict ed class K of reduced rings (i. e.
rings without nonzero nilpot ents) E(K) is elementary. On t he other side, if
one allows nilpot ent elements, even of bounded exponent (for inst ance one
considers rings whose nilpot ents a satisfy a
2
= 0), t hen t he element arity of
e. c. object s gets lost [164].
3.6 DCF
o
A differential ring is a commutative ring K with identi t y, having an addi-
tional 1-ary operation D (called derivation) such th at
D(a+b) = Dc v Db,
D(a b) = a Db +b Da
for every a and bin K. A differential field is a differential ring which is also
afield. So a suitable first ord er language L' for differential fields enlarges
our language L for fields by a new 1-ary operation symbol (to be represent ed
by D).
Differential fields include
(K, D) where K is any field and D = 0,
but also more significant and int erestin g st ruct ures, like
(K(x ), l;J where K is any field, K(x) is t he field of rational functions
wit h coefficients from K in t he unknown x , and txis t he deri vative
operation;
HO CIIA PTER 3. MODEL COMPLETENESS
for any nonempty connected open subset U of C, the field of mero-
morphic functions from U to C , with respect to the usual derivative.
Differential fields and rings were introduced by Ritt in the thirties, and
differential algebra developed greatly since then. Now classical and excellent
references, like [76] or the nimbl er [67], expound its foundations. For our
purposes in this section, we just need a few algebraic crumbs on differential
fields (/C, D) .
First of all , notice that the usual derivation rules for power s and quotients,
like
D(a
n
) = na
n
-
1
Da
for every element a in I< and every int eger n, still hold and can be easily
deduced from the definition.
Moreover, the elements a E I< such that Da = 0 form a subfield of /C, called
the constan t su bfield and denoted C (/C).
Instead of the usual (algebraic) polynomials f(x) E I<[x], now differential
polynomials are considered: they are algebraic polynomials in the unknowns
x , Dx , .. . , Dnx, ...
where n is a natural, and form a differential ring /C{x} with respect to the
obvious operations.
For every non-zero differential polynomial f (x) E I<{x} we can define
the degree of f( x) (with respect to x , or Dx, and so on)
but also
the order of f(x): thi s is the maximal natural n such that Dnx occurs
in f(x) , if there is some n with this property, and -1 otherwise, so
when f(x) = a E K (clearly one agrees DOx = x).
Differential polynomial rings in more variables /C{i} are introduced in the
same way.
What is the role of model theory in this setting? As we shall see in a few lines,
it is quite relevant , mainly with respect to existentially closed structures.
Actually differential algebra did not provide, before model theory, any notion
of differentially closed field -something resembling algebraically closed fields
among fields, or real closed fields among orderedfields-. But the interest in
existentially closed structures, and hence, particularly, in existentially closed
differential fields , led A. Robinson to consider this question from the model
theoretic point of view.
3.6. DCF
o
111
Hence take t he -elernent ary- class of differential fields in our language i:
We wonder if t he class of existentially closed objects is element ary, in oth er
words if t he t heory of differ ential fields has a model companion. At t his point
it is advisable for us t o distinguish in our t reatment t he characterist ic 0 case
from t he prime charact eristi c case . We shall devot e t he next sect ion to t he
lat t er ; hence, now we limit ourselves t o differ ential fields of characterist ic O.
In t his restrict ed framework, A. Robinson point ed out
Theorem 3.6. 1 (A. Robinson, 1959) Th e theory of differential fi elds of
characteris tic 0 has a model compani on, and thi s eliminates the quantifiers
in the language i:
This model companion was denoted DCF
o
; it s models were called differen-
ti ally closed fields (of charact eristic 0) . Unfortunately, A. Robinson's ap-
proach was not able t o det ermine any incisive first ord er axiomat izat ion for
DCF
o
. Actually, what was lacking at t hat time was a suit able background .
In fact , in t he paradi gmati cal cases of ACF and RCF, th e notions t hem-
selves of algebraically closed field and real closed field clearl y preceded A.
Robi nson 's t reat ment, and t he existe nce and uniqu eness of an algebraic clo-
sure and a real closure played a key role in t he model completeness proofs.
On t he cont rary, in t he different ial case, t he concept of differenti ally closed
fields was just rising wit hin t he model t heoretic approac h, and, consequent ly,
nothing resembling a differenti al clos ure of a differenti al field (as a minimal
different ially closed extension) was known , even in cha racterist ic O.
Accordingly one had to wait for new significant progr ess in model theory,
mainl y due to Michael Morley, befor e over coming t his algebraic gap. We
shall refer in det ail Morley's ideas, and their effects for differentially closed
fi elds, in Chapte r 6. But we wish to anticipate here a short report on th e
end of th e affair (so far) . In particular , we want to recall t hat in 1968 L.
Blum, in her PhD thesis, found an elegant and nice axiomatization of DCF
o
.
What she showed was
Theorem 3.6.2 (1. Blum) Among differential fi elds (K , D) in characteris-
t ic 0, the existentially closed object s are exact ly those satisf ying the follo wing
property :
(*) fo r every choice of f (x) and g(x ) in K{ x} - {O} such that the order of f
is larger than the order of g, there is some a E [( fo r whic h f (a) = 0
but g(a) f. O.
It is easy t o express (*) in a first order way by infinit ely many sentences
in L', L. Blum's work is described in Sacks's book [146]: she confirmed
quantifier elimination, model complet eness (and completeness) of DCF
o
;
but her analysis went fart her and, combined with a quit e abstract mod el
t heoretic resul t of Shelah, yielded
Theorem 3.6.3 Any differential fi eld (K, D ) of charact eristic 0 has a dif-
f erential closure (a minimal differentially closed extension), and this is
unique up to isomorphism fi xing I< pointwise.
We shall refer in mor e det ail on this point in Chapter 6. But we want
to emphas ize that , at last , differential closures do exist and are uniqu e (in
cha racteristic 0) , and t he model th eoretic approach of A. Robinson and L.
Blum was the very first not only in introducing differenti ally closed fields,
but also in showing th ese existence and uniqueness resul t s. Of course, some
myster y still per sist s. Most not abl y, it is surprising, and perh aps regrettable,
to learn t hat , presentl y, no explicit example of differenti ally closed field is
known.
However , we conclude t his section by discussing some minor but useful
point s.
First of all, noti ce that every different ially closed field (K, D) of character-
isti c 0 is algebraically closed: t o see t his, j ust apply Blum's t heorem t o t he
par t icul ar case when f (x ) E I<[x ], g(x ) = 1. Similarly t he constant field
C(K) is algebraically closed (we will see why in Chapter 6) .
Secondly, what can we say about definabl e sets in a differenti ally closed field
(K, D) of characterist ic O? Here
(*) t he basic definabl e sets are, of course, t he sets of roots of finit e systems
of different ial polynomials in I<{i} ; t hey are j ust t he closed set s in
a Noether ian topology (t he Kolchin topology) in K " (where n is t he
length of i);
(*) the Kolchin constructible sets -i. e. th e finit e Boolean combinat ions of
Kolchin closed sets- are st ill definable, of course;
(*) owing to the elimination of quantifiers, no furt her definabl es occur; in
oth er words, definabl e j ust means Kolchin constructible.
3.7 SCF
p
and DCF
p
As promi sed, we want to deal here wit h differenti ally closed fields in pri me
characteristic p. But our t reat ment needs a pr elimi nary remark. Indeed we
3.7. SCFp AND DCFp 113
underlined in the previous section t hat differentially closed fields of char-
acterist ic 0 are algebraically closed as well. This cannot hold any more in
the prime characterist ic case. In fact , t ake any differential field (K, D) in
cha ract erist ic p: notice that I<P< C (I<) becau se
D(a
P
) = p a
P
-
1
Da = 0 Va E I<.
Now any algebr aically closed field of characteristic p is perfect ; in other
words, every element can be (uniquely) expressed as a p-th power. It fol-
lows that no differential field of characterist ic p can be algebraically closed,
except the trivial case when th e der ivation D is identically O. In particular
we cannot expect that the underlying field of a differentially closed field is
st ill algebraically closed. This remark threatens some major complications
with respect to the charact erist ic 0 case; and anyhow suggests a prelim-
inary analysis on some possible weak closure notions for pure fields K in
cha racteristic p, compatible with I< =1= to:
Accordingly, t ake a field K (in any cha racte rist ic) . A polynomial f (x) E I< [x]
of degree 2: 1 is said to be separable if and onl y if f (x) has no multiple roots
(in the algebraic closure of K).
When K has characterist ic 0, every irr educible polynomial f (x) E I<[x] is
also separable. Otherwise a multiple root a of f( x) annihilat es also th e
formal derivative f'( x) of f( x) -still a polynomial in I<[x]-, and hence t he
greatest common divisor q(x) of f( x) and f'( x) (in I<[x]). So q(x) is not
constant; however the degr ee of q(x) is st rict ly smaller than the degr ee
of f(x ), because it is less or equal to the degree of f'(x). Hence f( x) is
reducible.
Now assume t hat K has a prime characteristic p. Recall t he Frobenius
morphism Fr in K (and in every extension of K) , the one sending any
element a into its p-th power a", Fr is injective. This t ime, I<[x] may include
some irreducible non- separable polynomials: for instance, given a E I< - I<P
and a positive integer h,
h
f( x) = x
P
- a
is irreducible, but has a unique root in the algebraic closure of K because
Fr is 1-1; notice t hat f'(x) = o.
We say that K is separably closed if and only if every sepa rable polyno mial
f( x) in I<[x] has a root in K .
Separably closed fields K form an elementary class in the language L of
fields; in fact , they are axiomat ized by t he infinit ely many sentences saying
that every polynomial f( x) with a non-zero derivative has a root in K .
However, in characteristic 0, separably closed just means algebraically closed.
On the contrary, in a pri me characteristic p, there exist some separably
closed, non-algebraically closed fields; indeed one shows that a field J( is
algebraically closed
if and only if
J( is separably closed and Fr is onto (that is, J( = J(P is perfect).
If this is not the case, we are led to consider the field extension J( :2 J(P and
hence J( as a vectorspace over J(P. A set B <;;;; K is called a p- basis of J( over
J(P if the monomials
(when n, eo, ... , en range over natural numbers, aa, .. . , an are pairwise
distinct element s in Band e; < P for every i ::; n) form a basis of J( over
J(P (as a vectorspace) . One shows that a p-basis B always exists and it s
cardinality depends only on J( : it is call ed the imperfection degree of J(.
The model th eory of separably closed (possibly non algebraically closed)
fields J( of prime characteristic p was investigated by Ershov in 1967 [46].
Let seFp denot e t heir t heory. Of course, seFp is not complete, because
t he imperfect ion degr ee of these fields may change, and suitable first order
formulas in the language offields can express it s valu e, when finit e, or wit ness
its infinity otherwise. But Ershov observed that this imperfection degree is
the only obstruction to completeness; in fact, he showed
Theorem 3 . 7.1 (Ershov) Let J( be a separably closed fi eld of prime char-
act eristic p. Then the elementary equivalence class of J( is fully det ermined
by its imperfection degree, if finite, or by the f act that this degree is infinite,
otherwise.
Ershov's proof shows (an d uses) model completeness in a suit able enriched
language capturing the notion of p-basis. Notice that separably closed fields
do have some noteworthy algebraic connect ions with differential fields. For
instance, given any separably closed field J( and a p-basis B of J( , one can
see t hat any function 0 from B to K enlarges uniquely to a derivation D of
J( , and so equips J( with a structure of differential field.
This relationship between separably closed fields and differential fields in
prime characteristic p gets stronger if we enter t he model theoretic frame-
work and we consider differentially closed fields. But, before providing more
details, we have to explain what a differentially closed field in characteristic
3. 8. ACFA 115
p is. Indeed , we still have t o clarify if, even in this set t ing, existe ntially
closed objects are an element ary class: so far , we have simply realized t hat
the an alysis should be more complicated t han in characteristic 0, because
no existe nt ially closed differ enti al field ca n be algebraically closed .
Well, the answer is again positive: t he theory of differential fields of charac-
teristic p does admit a model companion. This is the cont ent of a theorem
of Carol Wood , who also found a nice axiomatizat ion of existe nt ially closed
objects.
T heorem 3. 7. 2 (C. Wood) A differential fi eld (/C, D ) in characteristic p
is existentially closed if and only if
(i) C(K) equals KP,
(ii) fo r every choice of two differential polynomials f (x ) and g(x ) in
K{x } - {O} such that the order of f (x ) is great er than the order of g(x )
and the formal derivat ive of f (x ) with respect to D" x is not 0, there is
some a E K such that f(a) = 0 and g(a) =1= o.
The model companion of the t heory of differ enti al fields in characteristic p is
usually deno t ed DCFp; it s models -hence t he exist ent ially closed differ enti al
fields- are again called differenti ally closed fields. Notably, t hey are separably
closed: t his is implicitly said in (ii), provided t hat we restrict to pol ynomi als
f( x) of ord er 0 and we t ake g( x) = 1. Moreover their imperfection degree is
infinite, because t he dimension itself of /C over C(K) = K Pis infinite: t o see
this, just use (ii) again and take, for every positive int eger rn, an element
X
m
in K satisfying D m x m = 0 and D
m- 1
X
m
=1= 0; noti ce t hat t he xm's are
linearly ind ependent over C( K) (this is an easy exercise) .
DCFp is mod el complet e and complete, but does not eliminate t he quanti-
fiers in the language L' : quantifi er eliminat ion needs a larger setting, with
a further operation ext racting p-th roots when possibl e, and valuing 0 oth-
erwise.
Again, general resul t s of pure model t heory ensure t he exist ence and unique-
ness of a differenti al closure for differenti al fields (/C, D) satisfying C( K) =
K P.
3.8 ACFA
In t his section we deal with differ ence fields. These are struct ures (/C, 0" )
where /C is a field and 0" is a distingui shed (surjective) automorphism. More
generally, difference rings can be introduced in t he same way. Difference
fields include some t rivial examples like
(!C, id) for every field !C
but also
(!C, Fr)
where !C is a perfect field of pr ime cha racterist ic p and Fr is t he Frobenius
morphism in !C: for every a E I<, Fr( a) = al', Notice t hat t hese examples
include any finit e fi eld !C.
At a super ficial sight, this set t ing resembles different ial fields. In both cases
t he underlying field is enr iched by a l -ary operation, and the only dissonance
concerns the rules that this new function has t o sat isfy: the derivation
laws in th e differenti al case, and t he morphism laws pr esently. Of course,
t his connection is shallow, and sharp distin ctions arise in examining t hese
struct ures in a deeper way. Nevert heless some (minor) similarit ies persist .
For inst ance, given a difference field (!C, <T ), one can look at t he fixed subfield
Fi x(<T) = {a E I< : <T (a) = a} .
T his resembles in some way t he constant subfield of a differential field.
Similarly, inst ead of algebraic polynomi als f (x) E I<[x], one can form differ-
ence polynomials in x (or, possibly, in more unk nowns) : t hey are algebr aic
polynomials in
x, <T (x), <T
2(
x), . .. , <Tn(x), . ..
(wit h n natural). So formally t his is the same set as in t he differential case;
but , of course, the new rules relating <T and t he operati ons of addition and
multiplication dict at e a different ring struct ure. This ring is usually denot ed
!C(x) and gets a difference ring struct ure extending (!C, <T) in th e obvious
way.
Difference algebra was began by Rit t in the thirties; now it is largely devel-
oped and includes some fund ament al references, such as [28] (warning: the
t erminology used in [28] is not t he same as here; in fact in t hat book a dif-
ference fi eld is a field wit h a distinguished monomorphism <T ; when <T is also
surjective, the difference field is called inversive). Wh at ca n we say within
t he model t heoretic approach? First of all, notice t hat t he language for dif-
ference fields enlarges {+, " - , 0, I } by a new l -ary operation symbol. So
t he resul tin g setting again reminds differenti al fields, alt hough now we prefer
to denot e t he new symbol by <T. Difference fields are eas ily axiomatized in
a first order way in t his language.
3.8. ACFA 117
But our main question is how t o characterize existent ially closed difference
fields. Is their class axiomatiza ble? In other words, does the theory of
differ ence fields admit a model companion ?
We should emphasize t hat this inter est in existent ially closed difference fields
is comparatively recent. Indeed , at t he begi nning of the nineties, Macintyre,
Van den Dries and Wook showed t heir eleme ntarity an d found a nice (al-
though non trivial) axiornatization: t his is report ed in [95]. In detail
Theorem 3.8.1 A difference fiel d (K , a) is existenti ally closed if and only
if
(i) K is algebraically closed;
(ii) for every irreducible affine variet y U over I< and fo r every subvariety
V of U X a( U) such that both the projections of V into U and a(U) are
dense with respect to the Zariski topol ogy, there is a point a of U (over
I<) for which (a, a (a)) is in V .
Expressi ng (ii) in a first or der way is not immedi ate, because it requires,
after all, to qu ant ify with res pect t o irreducibl e varieti es. However this can
be done, owing t o general boundedness resul t s for (algebraic) polynomials
over fields [172].
The t heory of existent ially closed differ ence fields is denoted ACFA; their
models are called algebraically closed fields with an automorphism, although
this name is a lit tl e misleading, and we have to be ca reful about it. So recall
that they are not simply algebr aically closed fields enriched by any arbitrary
automorphism, but j ust existent ially closed structures among fields wit h an
automorphism: (ii) has t o be resp ected . Of course ACFA is model complete,
as a model companion. But ACFA is not complete. However one shows
t hat the key features fully det ermining t he elementary equ ivalence class of
a model (K, a) of ACFA are
(*) t he characteristic of the underlying field K ,
and
(*) the action of a on the algebraic closure of t he prime subfield (up to
isomorphism).
The interest in ACFA is also related t o its role t owards a model theoretic
proof of classical questions in Algebraic Geometry, like t he Manin-Mumford
Conjecture. We will discuss t his point in Ch apt er 8. However t hese intrigu-
ing connections led to a syste mat ic st udy of ACFA, mainly pursued by
Chatzidakis and Hrushovski in [20], and later by Chatzidakis, Hrushovski
and Peterzil in [21]. Here we limit ourselves to some very basic information.
First of all, it turns out that , for any algebraically closed field with an
automorphism (K, 0"), the fixed subfield Fix(O") is pseudofinite, so it satisfies
the axioms of finite fields, but (consequently) it is not algebraically closed.
What about definable sets?
First one meets the zero sets of finite systems of difference polynomials
in K(X';; they are, again, the closed sets in a Noetherian topology for
K" (where n is the length of x). Fix(O") is an example, because it is
defined by O"(v) = v.
Secondly, one has to include the constructible sets in this topology, i.
e. the finite Boolean combinations of closed sets.
However, this is not enough, because quantifier elimination fails for
ACFA in its original language. So definable does not imply con-
structible. Here is a possible counterexample. We work inside an
"algebraically closed field with an automorphism" (K,O") of charac-
teristic 0 and suitably large cardinality. Due to this assumption and
existential closedness, we can find inside K two elements a and b tran-
scendental over the prime subfield Q, a and b in Fix(O") , and two
square roots a' and b' of a, b, respectively, such that
O"(a') = a', O"(b') = -b' .
a and b satisfy the same quantifier free formulas, because they are
fixed by 0" and f (a), f (b) i= 0 for every nonzero polynomial f (x) over
the rationals. Nevertheless
O"(V) = v A :3w(w
2
= v A --'(O"(w) = w))
holds for b and not for a, and hence defines a non-constructible set. Of
course, owing to model completeness, the projections of constructible
sets exhaust the definable ones.
Of course, owing to model completeness, the projections of constructible
sets exhaust the definable ones.
A final important remark. We saw that no example of differentially closed
field (even in characteristic 0) is known so far. This is not the case for
ACFA. In fact, Hrushovski and, independently, Macintyre built some mod-
els (K, 0") of ACFA explicitly; they are obtained by considering suitable
pseudofinite fields with a rather natural extension of the Frobenius mor-
phism.
3.9. REFERENCES
3.9 References
119
An excellent introduction to model completeness is in Macinty re expository
paper [94]. But also t he A. Robinson books [141] and [143] are still an
enjoyable reading. Hilber t Nullstellensatz and Hilbert Seventeent h Problem
are described in [94]; p-adi c numbers ar e also t he mat t er of a (particul ar )
chapter in [36] ; the Ax-Kochen-Ersov theorems are in [4, 5, 6] and [44, 45],
while Macinty re's quantifier elimination results on p-adics are in [93].
A very good refer ence for existentially closed st ruct ures (and model com-
plet eness, too) is Cherlin 's book [23]. See also the Eklof-Sabbagh paper [41].
Macintyr e's analysis of existe nt ially closed groups is in [91], and Cherlin 's
treatment of e.c. commutative rings in [22]: in th e latter set t ing, the role of
nilpotent elements is also discussed by Carson [17], Lipshitz-Saracino [86],
and Macinty re again [92], who observed that the t heory of commutative
rings without nonzero nilpot ent elements do have a model companion. See
also [164]. Differenti ally closed fields are introduced in [142], but t heir mod el
th eor y (in characteristic 0) is principally developed in L. Blum's Ph.D. The-
sisi [12], also relat ed in t he Sacks book [146]. By t he way, both [76] and [67]
provide complete t reat ments of t he basic differenti al algebra.
T he Ers hov model t heoretic analysis of separably closed fields is in [46], while
C. Wood 's t heorem on differenti ally closed fields of prime cha racterist ic is in
[179] and, wit h respect to prime models, in [180]; t he Marker and Mess mer
contributions t o [110] provide a complete and excellent exposit ion of t hese
matters.
The Maci ntyre - Van den Dri es - Wood axiomatization of existent ially closed
fields with an automorphism is related in [95]; t heir lar ge model theoretic
analysis due to Chat zidakis, Hrushovski and (later) Pet erzil is in [20] and in
[21]. The Hrushovski and Macint yr e explicit example of an e.c. field with
an aut omorphism ca n be found in [88] and [58].
Chapter 4
Elimination of imaginaries
4.1 Interpretability
Let A be a structure for a language L. We already dealt with definabil-
ity in A when we introduced definable sets and, mor e generally, definable
struct ures in A (see example 1.7.3 , 4) . The latter are th e st ruct ur es A' for
a language L
'
(possibly different from L) such that both t he universe A' of
A' and t he interpretations of symbols of L
'
in A' are definable in A . As in
t he case of sets, we can introduce also the concept of X -de finable struct ure
for X A. In t he quoted example, we observed t hat (N, +,. ) is a st ruct ure
definabl e in (Z, +,.). Here we provide some fur th er examples.
Examples 4. 1. 1 1. Let L = {x , e,- l } be th e language for groups, 9 be
a group. The centre Z(9) of 9 is t he set of elements of G commuting
with any element of G, and so it is 0-definable as a set, by th e formula
Vw(v x w = w X v) .
But Z( 9) is also a subgroup of 9, and hence a gro up wit h respect to t he
restrictions to Z (9) of t he operations of 9. Clearly t hese operations
are 0-definable in 9. It follows t hat Z(9) is 0-definable in 9 also as a
group (and hence as an L-structure).
2. Let L = {a, 1, +," - } be t he language for fields, IC be a field. Let us
consider, in t he language L
'
= {x , e,- l } for groups, th e linear group
of degree n (n a positi ve int eger) over IC
GL(n, IC),
121
122 CHAPTER 4. ELIMINATION OF IMAGINARIES
in other words the group of n X n matrices with entries in J{ and deter-
minant i= 0, with the usual row-by-column product . Then GL(n, K)
is a structure 0-definable in K.
In fact, the n X n matrices with entries in J{ can be viewed as tuples
if = (aijkj in J{n
2
So the element s of GL(n, K) are exactly those
tuples in J{n
2
satisfying
-,(det if = 0)
(here if abridges (Vi j kj): recall that det is a homogeneus polynomial
of degree n with coefficients in the substructure of K generated by 0, 1.
The multiplication is defined by
n
ifxw=z 1\ (zij=LvihWhj) ,
Isi,jsn h=l
the identical matrix In by
1\ Vii = 1 /\ 1\ Vij = 0,
l<i<n Isi,jsn,i#j
and, finally, the inverse operation by
if = w-
1
: "if X w = In" .
But Algebra deals not only with substructures, but also with quotient struc-
tures. For instance, in the examples quoted before, one can observe what
follows.
1. Look at the quotient group 9/Z(9). As Z(9) is 0-definable in 9, the
equivalence relation in G
whose equivalence classes are just the elements of 9/Z(9) is also 0-
definable. It follows that 9/Z(9) , as a quotient set, but also as a
quotient group, "lives" in 9.
2. The special linear group of degree n over K
SL(n,K)
(a subgroup of GL( n, K)) is 0-definable in K: just consider the formula
det(if) = 1.
4.2. IMAGINARY ELEMENTS 123
Accordingly, just as in t he previous example, t he quotient group
GL(n, K ) j S L(n, K) can be recovered in K by a suitable 0-definabl e
equivalence relation. On th e ot her side GL (n, K) j SL(n , K) is isomor-
phic to t he mul t iplicative group K* of nonzero element s of I<, and K*
it self is dir ectl y 0-definable in K (as a gro up, wit hout involving any
quotient const ruction) .
These remarks int roduce the following
Definition 4.1.2 Let A be a structure f or L . A structure A' f or L' is said
to be interpretable in A if and only if there exist a positive integer n, a
su bset S An defi nable in A, an equivalence relation E over A n defi nable
in A such that A' = S ] E and
(i) f or every costant sy mbol c of L', { a E An : al E = cA'} is definable in
A;
(ii) for every k-ary function symbol f of L' , {(al , " " a) E A n(k+l) :
fA' (aIIE, ... , aklE) = alE} is definable in A;
(iii) for every k-ary relation symbol R of L' , {(al , . . . , ak ) E A nk :
(aIIE, . . . , aklE) E RA' } is defi nabl e in A.
The concept of X-interpretable struct ure (for X A) is defined in t he
usual way. Every structure definable in A is int erpret abl e in A (t hrough t he
equality relation) .
Example 4.1.3 GL(n, K ) j SL(n , K) is 0-interpretable in K.
4.2 Imaginary elements
The examples of t he previous section show that, in a given L-struct ure A,
one meets not only
real elements
(t hose of t he domain A) , but even
imaginary elements,
t. e , equivalence classes of 0-definable equivalence relations E .
T he following technique, essent ially due to Shelah, shows how to expand in
a natu ral way t he structure A (and it s language L) in order to make t he
imaginary elements real.
Definition 4. 2.1 U
q
is the language obtained by L taking, f or every equiv-
alence relation E over An (f or some positive integer n) 0-definable in A
a 1-ary relation symbol AE (and among these also A=);
an n-ary function symbol 7rE .
Definition 4.2 .2 Aeq is the structure for u
q
such that
the universe of Aeq is the disjoint union of the interpretations in Aeq
of the relation symbols AE;
f or every equivalence relat ion E, AE is interpreted as the quotient
set An j E (in particular A= as {{a} : a EA}, which can be canoni-
cally identifi ed with A) and 7rE as the natural proj ect ion from An onto
AnjE;
the symbols from L have the same interpretation as in A (aft er iden-
tifying A and the interpretation of A=).
One can check that , if A and B are st ruct ures for L and A == B, then
Aeq == B'", This allows t o define, T' " = Th (Ae
q
) when T = Th (A). Many
significant properties of T are preserved under passing from T to T'", On
t he other hand, in T '" = Th (Ae
q
) t he imaginary elements of A get real. For
example, t ake a positive int eger n and the relation = in An; = is 0-definabl e,
and hence, for every a = (a1 ' ... , an) E An, t he class {a} of a modulo =, is
a real element of Ae
q
; so we can view any n-tuple in An as a real element.
Indeed it is a common agreement in Model Theory to consider the tuples
from a st ruct ure A as element s of A: t hey are imaginary elements in A , and
hence real element s in Aeq.
Sometimes imaginary elements can be naturally ident ified with real elements
of A (or wit h finit e sequences of real element s of A), and hence referring to
Aeq is no more necessary. Let us pro pose a simple example.
Example 4.2.3 Let K be a field. We alr eady pointed out that
GL(n, K)j SL(n, K) is isomorphic to the multiplicative group J(* (0-definabl e
in K). It follows that t he imaginary elements SL(n, K)a with a E GL(n, K)
can be ident ified with the real element s det(a) of J(* , by th e determinant
function (more precisely by the isomorphism from GL(n , J()j SL(n , K) to
J(* induced by det).
In general:
4.2. IMAGINARY ELEMENTS 125
Definition 4.2.4 A struct ure A for L has the elimination of imagi-
naries if and only if, for every equivalence relation E over An (n a positive
integer) 0-definable in A and aE An, there are a formula r.p( v, w) of Land
a unique sequence bin A such that
A structure A for L has a uniform elimination of imaginaries if and
only if, for every E as above, there exist a formula r.p (ii, w) of L and, for
each aE An, a unique sequence bin A such that
Theorem 4.2.5 Let A be a structure for L . Th en the follo wing proposit ions
are equivalent.
(i) A has a uniform elimination of imaginaries;
(ii) for every positive integer ti and equivalence relat ion E over An 0-
definable in A, there is a map Fi; 0-definable in A, with domain An
and range Am (for some posit ive integer m), such that, Va, a' E An,
a, a' are equivalent in E FE(a) = FE(a' ).
Hence, if A has a uniform eliminat ion of imaginaries, th en, for every equiv-
alence relation E as above, t he equivalence classes of E ca n be thought as
tupies of real elements of A, provided we ident ify, Va in An,
al E
wit h FE(a).
It follows that referring t o A eq is not necessary, because what is 0-inter-
pretable in A is even 0-definable in A.
Proof. (i)=?(ii) For each a E A , define FE(a) as the unique element bsuch
that alE = r.p (An, b) .
(ii)=?(i) Take r.p (v, Z) : "FE(V) = i" e b= FE(a). .,
The t heory of eliminat ion of imaginaries was essentially developed by Poizat.
Actually Poi zat's treat ment is slight ly different from ours. They do coincide
for st ruct ures admitting at least two constants, or even two distinct definable
elements (see [131], Theorem 16.16). As we are mainly int erest ed in fields,
and fields clearly sat isfy thi s condit ion, we can proceed with no anguishes.
4.3 Algebraically closed fields
Notably, many famili ar struct ures admit a uniform eliminat ion of imaginar-
ies. Let us propose some exa mples. The first case we deal with concerns
algebraically closed fields: t he aim of this section is just t o prove
T heor e m 4. 3.1 Every algebraically closed fi eld IC has a uniform elimina-
tion of imaginaries.
The proof is a direct consequence of the following two lemmas.
Lemma 4. 3. 2 If IC is an algebraically closed field, then IC has the elimina-
tion of imaginaries.
Lemma 4. 3. 3 Let L be a language with (at least) two cons tant symbols,
IC be an L-structure interpreting these two constant symbols in different
elements. If IC has the eliminati on of imaginaries, then IC has a uniform
elimination of imaginaries.
Obviously an algebraically closed field IC satisfies also t he hypotheses of
lemma 4.3.3: th e language for fields has t wo constant symbols 0, 1 int er-
pr et ed in IC as two differ ent elements . Hence IC uniformly eliminates imagi-
naries (provided it eliminates t hem) .
Now we show Lemma 4.3.2. The proof we provide here uses t he minimal-
it y of any algebraically closed fi eld IC, hence t he fact t hat every definable
subset of 1< is finit e or cofinite . An alternative approach (working even for
differentially closed fields) will be produced in Chapt er 6.
Proof. Let n be a positive int eger , E be an equi valence relation 0-definabl e
in K" : Let E (if, ill) indicat e t he formula defining E . Fix aE K", Consider
the formula
El (VI , ill) : :JV2. . . :JvnE (if, ill) .
El(lC, a) is a definable set, and hence by Corollary 2.4.8 is eit her finit e or
cofinite . If El (IC, it) is finit e, th en it contains only elements algebraic over
(th e subfield generated by) a; in fact t wo elements t ra nscendent al over (t he
subfield generated by) aare linked by some au tomorphism of IC fixing every
element of a; hence if one of th em occurs in El (IC, a) , t hen t he latter is in
El (IC, a) as well, and hence all t he elements of IC t ra nscendental over aar e in
El (IC, a), and El (IC, it) is infinit e. On t he other side, if El (IC, a) is cofinit e,
t hen El (IC, a) contains at least an algebraic element, because IC has finit ely
many elements algebraic over t he subfield generat ed by a.
4.3. ALGEBRAICALLY CLOSED FIELDS 127
In any case we can choose Cl = Cl (it) E E l (1( , a) algebraic over a. Now let
E
2
( Cl , /C, it) is either finit e or cofinite, and as above we can pick out in
E
2
( Cl , /C, it) an element algebraic over (a,Cl ), and hence over a. Repeating
th e procedure we build c= ( Cl , . ,C
n
) E E(J{n, a) such that Cl , , C
n
are
algebraic over a. We can even form a finit e set X (a) K" a-definable such
th at cE X (a); it suffices t o consider t he set X (a) defined by
1\ "fi (V ) = 0"
l < i < n
where, for each i = 1, ... , n , f i(X) is t he minimum polynomial of Ci over
t he su bfield generated by a. Without loss of generality, X(a) E(J{n, a)
(oth erwise substit ute X(a) for X(a) n E(J{n, a)).
Suppose X(a) = {c(O) , . .. ,c(m) } where c= c(O) and c(j) = (Cl (j) , .. ',C
n
(j))
for every j :::; m. Then we have
E(J{n, a) = E(IC , c(j))
for every j :::; m. Recall th at we are looking for a formula t.p (V, Z) and a
unique sequence bsuch t hat
Then consider the following polynomial f (y, x) E /C[y , X] (with x = (Xl , ... ,
X
n
) ) :
n
f(y , x ) = IT (y - L c; (j ) Xi ) .
j Scm i = l
Let b be the sequence of coefficient s of f(y , x) (with respect t o some pre-
established ordering) . It is clear that bis uniquely defined by X(a) =
{c(O) , ... , c (m) } . Conversely, as /C[y ,X] is a unique factori zat ion domain,
given the sequence bof th e coefficients of f(y , x), in other words given f( y, x),
we know that f(y , x) deco mposes in at most one way as
n
IT (y - L Ci (j) Xi)
jScm i =l
(up t o permuting fact ors) . So blet s recover X( a) = { c(O) , . . . , c (m) } (al-
t hough it may not provide the single sequences c (O), .. . , c( m)) . Then we can
pick out bas the sequence of coeffi cients of !(y, i) , and define 'P(71, i) in such
a way that
'P(71, b)
is th e formula
n
Vw( "(y- L Wi Xi ) divides j' (y, i)" -7 E(71, w)).
i=l
Obviously 'P(71, i) depends on a. ...
Let us prove now Lemma 4.3.3.
Proof. Let 0, 1 be two differ ent constant symbols in L such that K F= .(0 =
1). We know t hat, if E is an equivalence relation over K" 0-definabl e in K ,
then for every aE K" there exist a formul a 'PCi (71, i) and a unique sequence
b= b(a) in K such that
By compactness there exist a natural number h and aa ,... ,ah E K" such
t hat, for each aE K" ,
K F= V ::J ! zi V71(E(71,a) H 'Pai (71,ii) ).
i<h
Wi thout loss of generality, we ca n ass ume that Zo , . .. , Zh have the same
length m (oth erwise add some O's t o the shorter sequences) . Consider t he
formul a
'P' (71, il, Z)
(where ii = (uo , . . . , Uh) and Z has length m) conj unct ing:
the formula saying that , Vi :::; h, Ui E {O, I} and there exist s a unique
i :::; h such that Ui = 1;
the formula saying t hat, Vi:::; h, u; = 1 if and only if 'Pai(71,i)) , z is
the unique sequence Esuch that 'Pai uc-,Z) = 'Pai itc, i) and, Vj :::; h
with j < i, it is not true t hat t here is a unique sequence t such that
'Pai (K", Z) = 'Paj (K", i) .
It follows t hat for ever y a E K'" , there exists a uniqu e bin I<m+h+
1
such
that
K F= V71 (E(71, a) H 'P' (71, b)).
The formula 'P' (71, ii, i) depends only on E. Hence K has a uniform elimina-
t ion of imaginaries. ...
4.4. REAL CLOSED FIELDS
4.4 Real closed fields
129
T he aim of t his section is t o show t hat real closed fields unifor mly eliminat e
imaginaries.
Theorem 4.4.1 In every real closedfieldK imaginary elements can be uni-
formly eliminated.
Proof. First of all, let us emphasize a relevant proper t y of definabl e set s
in real closed fields, esse nt ially related t o t heir o-minimal it y. Take any for -
mul a O( v, w) in t he language L = {+, " - ,0, 1, of ordered fields. We
know t hat 'l9 (v, w) is (equivalent in ReF to) a finite bool ean combinati on
of formulas
(*) f n(w)v
n+
.. . +f1(W) v+fo(w) 0
where n is a natural number and f oCi) , . .. , fnUi) are polynomials wit h int e-
gral coefficients. Let q(y , x) = I:i<n f i( X)yi , q(y , x) - as a pol ynomial in y -
has at most n root s de noted by ... , Sr -1(x) (where clearly r depends
on x, but r n). Then (*) is equivalent t o the formula saying that eit her
w is equal t o some so(v) , ... , Sr- 1(iJ) or it is in a suit able union of intervals
among
] - oo, so(v)[, ]sO(V) , S1(V)[ , . . . , ]Sr-1( V) ,+00[.
It follows that in a real closed field K, for every a in J( , 'l9 (K, a) is a fi-
nite union of intervals whose endpoint s anni hilate some polynomials q(y , a);
hence these intervals are a-d efinabl e by a formul a onl y depending on 'l9 (v, w);
even t he number of t hese intervals is uniformly bounded wit h respect t o
'l9 (v, w). Let bdenote the sequence of these endpoints: bdep ends on a, as
said before, indeed every element in bis a-definable. At this point we may
wonder how to determine, for a given a, the zeros so(a), . . . , sr-da) (wit h
r = r(a) n) belonging to O(K, a) and, above all, t he intervals among
] - 00, so(a)[, ]so(a), s1(a)[, .. . , ]Sr-1(a) , +oo[
contained in O(K, a) . Of course, this depends on t he sign (+, - or 0) of the
polynomials q(y , a) in each point of b, and inside t he intervals. This can
be checked directly for the roots, and choosing in each int erval a b-definable
(hence a-definabl e) witness c to test . T he latter operation ca n be eas ily done
for every possibl e choice of the endpoints -00 d < e +00. For, take
c = 4 if -00 < d < e < +00,
c = d + 1 if -00 < d < e = +00,
c = e - 1 if -00 = d < e < +00,
c = 0 if -00 = d < e = +00.
and notice t hat C is a-defi nable when d and e are a-definabl e (or infinit e) .
In parti cul ar, obser ve that we can pick a unique a-definabl e c in (} (/C, a):
just choose t he lowest root , or t he wit ness in t he left most int er val. Hence,
what we have seen so far is t hat , given a formul a () (v, w) , one can build
a for mul a 0- (v, Z) and, for every /C F RCF and every sequence ain I< , a
unique sequence b in I< for which O' (/C , b) = (} (/C, a): b is t he ord ered t uple of
t he roots of t he polynomials q(y, a) (and so is a-defina ble) , o'(v, b) specifies
which roots and which intervals with endpoint s among b and oo form
(} (/C, a). A patient reader may wri t e down 0'(v, Z) in detail as an exercise.
Now let us come back t o elimination of imaginari es. Let E be an equival ence
relation on K" ; and ass ume E 0-definabl e in /C. Let E (V, w) the formul a
defining it. We are looking for a formula <p (V, Z) and, for every a E K ";
a sequence bsuch t hat E(I<n, a) = 1. Pu t
3V1 .. . 3v
n
E(v, w).
Again using our preliminar y wor k, and applying it to (}1(VI , w), we deduce
t hat there are a formula 0'1(VI , zi ) and, for each aE K"; a unique sequence
in I< such that
0'\(/C, = (}1 (/C, ii) .
Moreover there is a formul a X1(V\ , zi ) such t hat X1(V\, picks a unique
element Cl in 0'1 (/C, = (}1 (/C, a). Now consider
As before we obt ain t he existe nce of a formula 0'2(V2, Z2) and, for every
aE K "; of a unique sequence in I< such that
as above, t here is a for mul a X2 (V2 , select ing a unique element C2 in
O'2(/C' = (}2(C1 , /C, a).
Cont inuing t he procedure one defines a formul a X(v, Z) and, for each a in
I< , a unique sequence b such t hat
x (/C
n
, b) consists of a unique element if = (Cl , . . . , Cn)
and
4.5. THE ELIMINATION OF IMAGINARIES SOMETIMES FAILS 131
K F E(c, a), hence E(K
n
, a) = E(Kn, C).
Hence
t.p (e. Z) : ::Iw (x( e , Z) 1\ E (V, w))
is the required formula. It
4.5 The elimination of imaginaries sometimes fails
Differenti ally closed fields elimi nate imaginari es. This will be shown in
Chapter 6, in the section devoted to developing in detail their model t heory.
Similarly, algebraically closed fields with an automorphism (i. e. models of
ACFA) eliminate imaginaries [20], as well as separably closed fields of finit e
imp erfection degr ee in an enr iched language capt uring p-bases (see [110)).
But there do exist struct ures wit hout t his property. This is the case, for
instance, of the separably closed fields themselves in the pure language of
fields (again, see [110)). Let us propose here a simpler example.
Proposition 4.5.1 Let K be a finit e field with at least 3 element s and V
be a vectorspace of dimension 2: 2 over K. Then V does not eliminate the
. . .
unaquiaries.
Proof. Suppose yes. As K is finit e, one ca n consider in LK the formula
E (v, w) : V (v=kw)
kEK- {O}
This formula defines in Van equivalence relation E , identifying t wo elements
a, at E V if and only if a, at generate the same subspace. Take a E V , ai-D.
Then t here exist a formula t.p (v, Z) of LK and a unique sequence b in V
such t hat E (V, a) = t.p(V, b). As I< has at least three elements, fi x k E I<,
k i- 0, 1; for every x E V,
Since th e mult iplication by k is an automor phism of V, we have, for every
x E V ,
x E E (V, ka) x E t.p (V, kb).
Then E'(V, a) = E(V, ka) = t.p (V, kb). It follows b = kb; but k i- 1, hence
necessarily b = O. In other words, we can assume t hat z = 0 and t.p (v, Z)
reduces to a LK-formula t.p( v) with at most one free variable. The th eory
KT has t he quant ifier eliminat ion in LK, and hence <p(v) is equivalent to a
boolean combination of for mulas
v = v, v = o.
But V has dimension at least 2, therefore no such boolean combination ca n
equal E (V, a).
4.6 References
An exhaustive int roduction to A eq can be found in Hodges' book [56]. The
elimination of imaginaries was introduced by Poizat [130]; a complete treat-
ment can be found in Poizat 's book [131], recently t ranslat ed in english
[134].
The results on the elimination of quantifiers for separably closed fields can
be found in [110], while the cor responding analysis for e.c. fields with an
automorphism is provided in [20].
Chapter 5
Morley rank
5.1 A tale of two chapters
In 1965, M. Morley's work [116] proposed new ideas, new tools and , alto-
gether , new fertil e perspect ives in Model Theory. Actually Morley's main
t heorem is very simple to st at e: for , it says t hat a t heory T categorical
in some uncount able power is, consequentl y, categorical in every uncount -
able power. This is noteworthy, but perhaps not so dramatic and relevant .
However th e germs of Morley's ideas went much fur ther, and their richness
permeat ed the development of Model Theory for several years.
Accordingly, t his chapter , and t he following one , will be devoted t o preparing
a proof of Morley 's th eorem (alt hough t his pro of will be done only in chapter
7); but our main int ent t hroughout t hese pages will be to introduce, to
discuss and t o apply several Morley t ools: types, sat urat ed models, algebraic
and definable closure, t otally transcendental t heories.
We will deal also wit h some relat ed questi ons, both model t heoret ical (like
pri me models) and algebraic (such as differenti ally closed fi elds, or w-stabl e
groups and fields) .
5.2 Definable sets
Definable sets were int rodu ced in Cha pter 1, and were a constant leit moti v
t hroughout t he subsequent pages. They arise qui t e nat urally within a given
st ruct ure A from t he basic operations and relat ions in A, and fully charac-
t erize A. Indeed t he st ruct ure A could be t hought as a non empty set A
plus t he collecti on of it s definable sets . This new outlook is less formal t han
t he t radit ional definiti on given in Cha pter 1, and may sound a lit t le puz-
133
134 CHAPTER 5. MORLEY RANK
zling. However it is quite free and easy. Just to refer to our basic examples,
it is instinctive to view an algebraically closed field as naturally endowed
with the collect ion of its const ruct ible sets (including varieti es), or , in the
same way, a real closed field with the collection of it s semialgebraic sets , or
a differenti ally closed field with the famil y of it s Kolchin constructible sets.
Besides, should you like to fix exact ly this new perspective in introducing
st ruct ures , you could take note that the definable sets in a given A can be
characterized in a form ally pr ecise way, as follows.
Theorem 5.2.1 Th e sets definable in a structure A are the smalles t family
D of subs ets of Un>o An such that
(i) An is in D for every positive integer n; more generally for i ::; j ::; n
posit ive integers, {5 E An : a; = aj} is in D;
(ii) eve ry relation of A, as well as the graph of ever y operation in A, i s
in D;
(iii) D is closed under union, intersect ion and compleme nt;
(iv) D is closed under project ions: if X An is in D, then the image
of X by the proj ection of An onto any i ::; ti coordinates is also in D;
(v) D is closed under fibr es: if X An is in D, i is a posi ti ve integer
smaller than nand bE An-i, then the set of the sequences 5 E Ai fo r
which (5, b) E X is in D .
Just a few words to comment . A careful and st raight forward check easily
confirms that t hese condit ions are satisfied by the definable set s in A and
even characterize them. Indeed, (i)-(v) allow sets defined by at omic formu-
las (i)-(ii), Boolean combinations (iii) , quantifiers (iv) , par amet ers (v) and
nothing else.
In order t o realize how complicated a struct ure or, more generally, a class
of structures -for instance, the models of a given theory T- is, one can try
to measure the complexit y of it s, or t heir, definable sets. A possible way to
accomplish this program is to assign , to every definable, a value (such as an
ordinal , or something similar) satisfying some reasonable assumptions, like
monotonicit y (for C D definabl e sets, t he measure of D should not be
smaller than the measure of C), and so on.
To prepare this assignment , let us consider again definables. Fi x a set X
of parameters. Correspondingly one ca n form the Boolean algebra of X-
definable sets , as we saw in Chapter 1. In introducing our measure, we
may refer to t his algebraic framewor k. But we have to be careful; for ,
t he Boolean algebra of X-definable sets, as sketched in Chapter 1, seems to
depe nd on the structure A where X lies, and clea rly, even among the models
of a given complete t heory T , the choice of A may vary in an essential way.
Also, if we enlarge X to a bigger set of paramet ers X' , the X-definable sets
become auto matically X '-defi nable, but in this extended setting t hey gain
more complexity because t hey have to be compared wit h more object s; so a
finer analysis must be expected.
To cla rify t hese dou bt s (at least t he former one) , consider our set X, a
st ruct ure A containing X and a positi ve int eger n. Take a model Bf(X) of
Th(Ax); accordingly B is a st ruct ur e for Land j is an elementary func-
tion from X into B, in particular j(X) is a subset of Band Ax == Bf( x ).
As already said, for every positive integer n we wish to compare th e alge-
bras Bn(X, A) and Bn(J(X ), B) int roduced before. So consider any L(X) -
formula <p (v, i ) (v denot s here t he sequence (VI , . . . , v
n
)). Then <p(v, i)
defines two sets
<p (A
n,
i ) in A, and <p (B
n,
j (i )) in B.
It is easy to realize that they sat isfy the following properties.
1. If A is an elementary subst ruct ure of B (and j is th e inclusion of X
in B) , th en <p (An,i) <p( Bn, i) .
In fact, for every ii E An, if A F <p (ii,i), then B F <p (ii ,i), j ust
beca use B elementari ly extends A.
2. If <p (An, i ) is finite (of power k) , then 1<p (Bn, j (i ))1= k as well.
In fact , as Ax == Bf(X) , Ax and Bf (X) satisfy th e sentence :3! k v<p( v, i) .
In particular, under the ass umpt ions in 1, <p (An, i ) = <p( Bn, i).
On t he contrary, for an infinite <p( An, i ), <p (B
n,
j (i)) is infinite as well,
owing to elementary equivalence, but is possibly larger when B is an ele-
mentary extension of A. What we want to emphasize here is t hat , anyway,
t he Boolean algebras B
n
(X, A) and Bi,(J(X), B) are isomorphic; in this
sense our analysis of X-definable sets and t he corresponding assignment of
a measure does not depend on t he choice of A or B.
For t his pur pose , we need a preliminary analysis of Bn(X, A) (and, in par-
allel, of Bn(J(X ), B)). We know t hat every D in Bn(X , A) , na mely every
subset D An X-definable in A, can be viewed as D = <p(An, i) for some
L(X)-formula <p (v, i) . Of course, it may happen that two different L(X)-
formulas <p (V) and 'l/J (V) define th e same subset of An. This means that <p (v)
and 'IjJ (v) are logically equi valent wit hin Th(Ax) in the usual sense
Vv( cp(V) H 'IjJ (V) ) E Th(Ax).
Let rv denote t his (equivalence) relation. Elementary mathematical logic
t ells us that the quot ient set of L(X)-formulas cp (v) with respect t o rv gets
a nat ur al Boolean alge bra st ruct ure, provided we put , for cp (v) and 'IjJ (v) in
L(X) :
* the meet of t he <-classes of cp (v) and 'IjJ (v) is the rv-class of t heir conjunc-
t ion,
* the join of the rv-classes of cp (v) and 'IjJ (V) is the -v-class of t heir disjunction,
* t he complement of t he <-class of cp (v) is the <-class of its negation,
so t he bottom element is th e rv-class of --.(VI = vd or any contradict ion,
and t he top element is the rv-class of VI = VI or any tautology; in part icular
the -v-class of cp (v) is smaller or equal to the rv-class of 'IjJ (V) if and only if
<p (An) <;;; 'IjJ (An), and hence if and only if
Vv(cp(V) --+ 'IjJ (V) ) E Th(Ax ).
T herefore it is easy t o check that this quotient algebra is isomorphic t o
Bn(X, A) via t he function mapping t he rv-class of cp (v) into cp (An). The
same applies to Bn(J(X) , B), of course, and hence, in conclusion , Bn(X, A)
and Bn(J(X) , B) ar e isomorphic, as claimed, by t he function mapping
cp (An) into cp (Bn) for every cp(v). Hence t he isomorphi sm type of Bn(X, A)
depends on t he t heory of Ax rather t ha n t he mere st ruct ure Ax. Moreover
we can view t he elements of Bn(X, A) not only as X-definable subsets of
An, but also as -v-classes of formulas in L(X) . We will oft en confuse t hese
different points of view below, and even we will directl y t hink of t he elements
in Bn(X, A) as L(X)-formul as (ident ifying <-equivalent formul as).
5.3 Types
People acquainted with Boolean algebras B know what an ultrafilt er is and
remember t hat t he ult rafilt ers in a given B ca n be naturally endowed with
a topology which makes t heir set a Boolean (i.e. Hausdorff, compact and
tot ally disconnect ed) space: the dual space of B. Of course this procedure
applies to t he Boolean algebras of definable sets Bn(X, A). Even in t his
particular set t ing one can look at the ultrafilters: they are called complete
5.3. TYPES 137
n-types over X and their space is usually denoted Sn(X, A). That's all
about types.
But who is not familiar with Boolean algebras may appreciate some more
details. It is worthy satisfying her (him), also in order to realize explicitly
what a type is. This is the aim of this section.
Let us begin with a simple example. Everybody knows how real numbers
are introduced starting from the rationals and their usual order. But let
us summarize very briefly their construction (in our style using formulas
and structures). Accordingly consider the language L = and the L-
structure (Q, Partition Q in two non-empty subsets A, B such that, for
every a E A and b E B, a < b, A has no maximum and B has no minimum.
Then consider t he following set of L(Q)-formulas
p = {a < v: a EA} U {v < b: b E B}.
No element r E Qcan satisfy all the formulas in p. However, for every finite
conjunction <p( v) of formulas of p, there does exist r E Q for which Q F= <p(r);
in fact, let ao, ... , ah E A , bo, ... , b
m
E B, a be the maximal element among
ao, ... , ah and b be the minimal element among bo, ... , b
m
; then a < b
and, as the order of Q is dense, there is some rational r larger than a, and
consequently than ao, ... , ah, and smaller than b, hence than bo, ... , b
m
.
Now a simple application of Compactess Theorem (to be explained in detail
in the next Theorem 5.3.4) proves that in some elementary extension of
(Q, there is some element realizing p. On the other hand, even forgetting
Compactness Theorem and Model Theory, we do know that p is realized in
(R, just by the real irrational number corresponding to the section (A, B)
in Q.
The notion of type provides an abstract framework where to study the situ-
ation just sketched in the case of (Q, and (R, Accordingly take any
structure A for a language L, a subset X of A and a positive integer n.
Definition 5.3.1 A consistent n-type over X in Th(Ax) is a set p of
L(X)-formulas <p(v) (where v abbreviates (Vi' ... ' v
n))
such that every finite
conjunction of formulas of p is satisfied in A - more precisely in Ax - by
some suitable tuple aE An.
Definition 5.3.2 A complete n-type over X in Th(Ax) is a consistent
n-type maximal with respect to inclusion. Sn(X, A) denotes the set of com-
plete n-types over X in Th(Ax) .
In the sequel "n-type" will abbreviate "complete n-type", unless otherwise
stated.
Example 5 .3.3 Consider a mod el BJ(x) of Th (Ax ). Accordingly B == A
and f is an element ary function of X into B. For simplicity, assume t hat
X B and f is t he inclusion of X into B . Let bE B" ; p be t he set of
th e L( X )-formulas Cf/( v) such t hat B F Cf/( b). It is easy to check t hat p is a
complete n- t ype over X . Let us see why.
For every finit e conjunction Cf/ (v) of formul as of p, B F Cf/ (b), whence
B F :JVCf/(V) and A F :JVCf/ (v).
Enlarge p by a new formul a 'IjJ (if); 'IjJ (v) rf- p impli es B 'IjJ( b) and
so B F -,'IjJ (b) and -,'IjJ (v) E p; hence no set of formulas extending
p U { 'IjJ( if)} is a consistent n-t ype any mor e, because it contains two
formulas - 'IjJ (if) and it s negation - whose conjuncti on cannot be satis-
fied by any tu pie in An.
p is called th e type of b over X and is denoted tp(bj X) .
The next t heorem shows t hat any n-type over X ca n be obtained in t his
way.
T heorem 5.3.4 Let p be a set of L(X)-formulas Cf/(v) (where v abbreviates
(VI , . , V
n
) as before). The follo wing propositions are equivalent:
(i) pE Sn(X , A );
(ii) there are a model B of T h(A) such that X B and the inclusi on
of X into B is an elementary functi on, and a tuple b E B" such that
tp (bj X ) = p.
Proof. (ii)=}(i) was shown before.
(i)=}(ii) Enlarge L( X ) by a t uple cof n new constant symbols. Let L' be
th e language obtained in this way. Consider the following set of sentences
in L'
T' = Th(Ax) U {<p(C) : <p(v) E p}.
Every finit e subset Tb of T' has a model. In fact
Tb T h(Ax) U { <p (C) : <p (v) E Po}
for some finit e subset Po of p. As p is consistent , t here is some ii E An
satisfying in A th e conj unction of t he formul as of Po - At t his point look at
th e L' -structure expa nding A and int erpreting t he new constants cin ii and
not ice t hat t his provides a model of Tb.
Now apply Compact ness Theorem and deduce t hat t here is a model B' of
T'. Restrict B' t o L(X) and get a model BJ(x) of T h(Ax) (here B denot es
5.3. TYPES 139
a mod el of T h(A ) and f is an elementary function of X into B, but th ere is
no loss of gener ality in suppos ing X B , f = inclusion of X int o B ). Let
bE B" int erpret cin B'. In Bx , p tp( bjX). Owing t o the maximal ity of
p, p = tp( bj X) .
When p = tp(bj X ) for b E B" ; one says t hat b is a realization of p (or
also t hat brealizes p), and one writes b1= p. Notice t hat t he argument in
(i)=;.(ii) actually wor ks also when p is a consistent n-type over X , wit h t he
only exception of the last point (deducing p = tp(bj X) from p tp(bj X));
accordingly, if p is a cons istent n-type over X , then there are B and bsuch
t hat p tp( bjX), and we can state:
Theorem 5.3.5 Let p be a consistent n-type over X . Then p enlarges to a
comp lete n-type over X (possibly in several ways) .
In the particular case when X = A, Theorem 5.3.4 says that every complete
n-type over A is realized in some elementary extension of A. Notice also
t hat, when X A, any complete n-type over X can be viewed as a consistent
n-type over A and can be enlarged to a compl et e n-t ype over A, whence it
is realized in some elementary extension of A.
Theorem 5.3.4 (t oget her wit h the definition of complete type) also implies:
Corollary 5.3.6 Let p E Sn(X, A) , <p (v), 'IjJ (if) be L(X) -formulas.
(i) If <p( v) E p and 'IjJ (if) E p, then <p (v) 1\ 'IjJ (v) E p;
(ii) if <p (v) E p and VV(<P(if) ----+ 'IjJ (if)) E Th(Ax) , then 'IjJ (v) E p.
(iii) either <p (v) E p or ' <p(v) E P (and each case excludes the other).
Just t o summarize, we might say t hat , as t he rati onal order (Q, ::;) implicitl y
contains t hro ugh it s sections all th e real numbers -even t he irrational ones-
as ideal elements, similarly, for any A , X and n, t he n-types over X in
Th( Ax) tell us which new n-tuples of eleme nts can arise in the structures
Bj(X) == A x , in particular in t he elementary extensions of A. Under t his
point of view, the notion of n-type seems to deserve a good deal of attent ion.
So let us explore it. First we consider t he set Sn (X, A). We already linked
types and to pology at the beginni ng of this section. Let us examine t his
connection in mor e detail.
For every L(X)-formula <p (v) di L(X) , put
Urp(v) = {p E Sn(X, A) : <p (v) E p} .
140 CHA PTER 5. MORLEY RANK
Notice t hat, for cp(v) , 7j; (v) L (X )-formulas,
U<p(v) n U'Ij;( v) = U<p( v)I\ 'Ij;( v)'
In oth er words, for every t ype p E Sn(X, A) ,
cp(v) E p, 7j; (v) E P <=> cp(v) 1\ 7j; (if) E p.
In fact (= is just (i) in Corollary 5.3.6, while is a simpIe consequence
of (ii) .
Consequently t he sets U<p(v) for m a basis of open neighbourhoods for a to pol-
ogy on Sn(X, A ) when cp(v) ranges over L(X )-for mulas. Notice also t hat
U-'( Vl=vI) =0, UV1=Vl =Sn(X , A).
Furth ermore, for every L(X)-formula cp (v) ,
S; (X , A ) - U<p(v) = U-, <p( v)
(as implicitl y stated in Corollary 5.3.6, (iii)).
So t he topological space Sn( X, A ) is:
Hausdorff (in fact, for p, q E Sn(X, A ) and p =1= q, choose cp( v) E q - p
and obser ve -,cp( v) E p, whence p EU-, <p( v) , while q E U<p (v) ) );
totally disconn ected (because every open set of th e given basis is also
closed) .
Sn(X, A ) is also compact . In fact , take an open covering of Sn(X, A). With
no loss of generality we can ass ume t hat every open set of t his cover ing is
basic. So for some suitable set I of indexes our covering is just { U<pi (V) :
i E I} where, for any i E I , cpi (if) is an L(X )-formul a. Notice t hat { U<p i (V) :
i E I} covers Sn(X, A ) if and only if, for every model Bx of Th(Ax) and
bE B" ; t here is some i E I such t hat tp(bj X ) E U<pi(V) , namely B po cpi(b),
and hence if and only if, in a language with an n-t uple cof new constant s,
has no model. So, by Compactness Theor em, t here exists a finite subset l a
of J for which
Th(Ax) U {-,cpi(C) : i E l a}
has no model, equivalently, {U<Pi : i E l a} is a (finite) subcoveri ng of
Sn(X , A ).
By th e way, t his to pological application of Compact ness T heorem is just
the reason of it s name. But the topological framework also suggests t he
following definition.
5.3. TYPES 141
Definition 5.3.7 A type p E Sn(X, A) is said to be isolated if and only if
p is isolated as a point of the topological space Sn(X, A) , and so if and only
if there is some L(X)-formula cp(iJ) such that p is the only element in U<p(v) ,
namely the only n-type over X containing cp (iJ).
The next notion has a prevalent model theoretic flavour.
Definition 5.3.8 An n-type p E Sn(X, A) is said to be algebraic if and
only if there is a formula cp (iJ) E P such that cp (An) is finit e (and hence
cp(Bn) is also finite , and even of the same cardinality as cp (An ), for every
model Bf (X) of T h(Ax)).
Let us cornpare these notions. p is any t ype in Sn(X, A).
(i) If p is algebraic, then p is isolated.
In fact t ake a formula cp (iJ) in p such that cp (An) is finit e of minimal size
k. We claim that p is the only t ype in Sn(X, A) containing cp (iJ). Let
q be another t ype over X including cp (iJ) . For every formula 'l9( iJ) E p,
cp (iJ) A'l9 (iJ) is also in p, and so, owing t o t he choice of cp (iJ), cp (An) n
'l9(An) = cp (An), namely cp (An) 'l9(An). Hence 'l9(iJ) E q. It follows
p q, and so p = q by the maximality of p.
(ii) If X = A and p E Sn(A, A) is isolated, t hen p is algebraic.
Let cp( iJ) be an L(A)-formula isolating p. p is realized in some elemen-
tary extension B of A ; consequently B F ::I iJcp( iJ). As an element ary
substructure, A sat isfies ::IiJcp( iJ) as well, and so includes a realizat ion a
of p, and p = tp(a/A). But tp(a/A) contains t he formula iJ= a and so
ais it s only realization.
However we will see wit hin a few lines that an isolated t ype over an arbi trary
subset X may be non- algebraic. But , more gener ally, let us pro pose now
some specific examples of types.
Examples 5.3.9 1. Let L = 0, A be an infinit e set (viewed as a struct ure
for L) , X A. For ever y a EX, tp( a/ X) is the only l-'type containing
v = a, and so is both isolated and algebraic. Now t ake two elements
a and a' in A - X ; t here does exist a bijection from A onto A, hence
an aut omorphism of A, fixing X pointwise and mapping a in a'. So,
for every L(X)-formula cp( v), A F cp (a) if and only if A F cp (a' ); in
other words, tp( a/ X ) = tp( a' / X). This shows t hat all t he elements
in A - X realize the same type over X. Notice th at , for a finit e
X = {xo, .. . , x
n
} , this t ype is isolated, for instance by Ai<k--,( v = Xi );
however , it has infinit ely many realizati ons (t he elements in A - X)'
whence it cannot be algebraic.
2. Now take t he language L = {+, ., - , 0, I} for fields and an alge-
braically closed field /C. Let X be a subset of K. For simplicity we
ass ume t hat X is t he domain of some subfield 1i of /C. This is not so
restrictive. Indeed, not ice that each element in t he field 1i generated
by X is X -definable, in ot her words it is t he only poi nt in /C satisfying
a suitable L(X)-formula. Consequent ly there is no great loss of gen-
erality, when discussing t he types over X , in replacing X by H and
hence in ass uming t hat X = H is t he domain of a subfield 1i.
Owing to quantifier elimi nation, every n-type p over H is fully de-
termined by it s for mulas f( if) = 0 where f( x) ranges over polyno-
mials in H[X] . Indeed every Lwformul a is equivalent in ACF to a
Boolean combination of equations over 1i. Notice t hat t he polynomi-
als f(x) E H[X] for which f(if) = 0 E P form a proper prime ideal
in 1i[X]. This is easy to check. J ust take a realization ii of p in a
suitable extension , and notice t hat , for f( x) , g(x) , h(x) in H[ X] , if
f(ii) = g(ii) = 0, t hen f(ii) +g(ii) = f(ii) . h(ii) = 0; moreover, if ii is
a root of f(x) . g(x), th en it ann ihilates f(x) or g(x). Conversely, let
I be a (proper) prime ideal in 1i [X] and look at the set of formulas
{f (if) = 0 : f (x) E I} U{--, (g(V) = 0) : 9(x) E H [X] - I} .
This defines a (complet e) n-type over H , t he one of I +x viewed as
an element of t he extension of 1i provided by t he field of quotient s of
t he integral domain 1i [x]/ f.
In t his way we obtain a connection (indeed a bijection) between n-
types over H and prime ideals in 1i[x]. We shall provide a more de-
tailed analysis of this point in Cha pter 8. Now let us examine closely
t he case n = 1. When I is t he ideal generated by some monic irr e-
ducible f( x) E H[x], t hen we obtain t he ty pe defined by t he single
equation f(v) = 0 (and it s consequences). So the roots of f( x) share
a common I-t ype over H . This type p is clea rly realized in /C because
/C is algebraically closed; actually /C contains all t he realizati ons of p.
p is isolat ed by f(v) = 0, and algebraic , because f( x) has only finit ely
many roots.
Ot herwise I = O. Now t he corresponding I-type p conce rns all th e
elements which are transcendental over H . When the transcendence
degree of /C over 'H is bigger than 0, t hen p has (infinitely many)
5.1. SATURATED MODELS 143
realizations in K , in particular it is not algebraic. Otherwise, if K is
just the algebraic closure of 1{, t hen there is no room to satisfy p in
K. One easily sees t hat p is not even isolated.
3. Now take the language L = {+, ., -, 0, 1, :S} for ordered fields and
a real closed field K. We discuss directly t he t ypes over K , for sim-
plicity. Owing to quantifier elimination, every n-type over K is fully
det ermined by its equat ions and disequations f(v) = 0, g(iJ) 2: 0,
where f(x) and g(x) range over polynomials in n unknowns with co-
efficients in K. In particular, when we look at a I-type p over K ,
we have to consider polynomials in K[x]. But then, as explained in
Chapter 2, 5, p is completely determined by it s for mulas v :S a, b :S v
(and negations) , wit h a and b in K. So a nonalgebraic p is given by
the cut it defines in K.
4. Let L = {:S}, consider the L-st ructure A = (R, :S) and the subset
X = Q. If a, a' E R and a f a' (say a < a'), there is r E Q such
that a < r and r < a'. So v < r E tp(ajQ), but v < r rt. tp(a' jQ). In
conclusion, there are at least over Q in (R, :S).
In general , for a countable language , there are at most n-types
over a set X for every posit ive integer n.
Now let us mention another relevant t echnical fact about types; it will be
useful several times later. Let A be a model ofT, X A, f be an element ary
function from X into a model B of T (for instance, let f be t he rest rict ion
to X of some isomorphism between A and B). For pE S(X, A), put
f(p) = { <p(v,f(x)) : <p (v, x) Ep (and xinX)}.
Then f(p) is a ty pe over f(X) in B. Indeed f det ermines in this way a
homeomorphism between S(X, A) and S(J(X) , B). In particular f(p) is
isolated, or algebraic, exactly when pis.
To conclude, it is worth underlining that , just as th e algebra B
n
(X, A) of
X-definabl e subsets of An , similarly the space Sn(X, A) of n-types over X
does not depend dire ctl y on th e model A where X is element arily embedded,
but only on the theory ofAx .
5.4 Saturated models
Let T be a complete theory with infinite models in a countable language
L. A model A of T may not realize all the l-typ es over a subset X: a
t rivial counterexample is provided, when X = A, by th e consiste nt I- t ype
{--,( v = a) : a E A} , which cannot sat isfied by any element in A. The
sat urat ed models of T are, roughly speaking, those realizing as many types
as possible over their subset s. In part icular, for an infinit e cardinal A, the A-
sat urated models of T are those able to satisfy any I- t ype over an arbitrary
subset of power less than A.
Definition 5.4.1 Let A be an infinit e cardinal. A model B of T is sai d t o
be A-saturated if and only if, for every subset X of B or power < A, B
real ize every i -type over X.
Of course this definition makes sense when A IBI (as the previous coun-
t erexample shows); it is also clear t hat, if A 2:: f-L 2:: are cardinal numbers
and B is A-saturat ed, t hen B is f-L- saturated, t oo. The first quest ion one may
raise at this point just concerns the exist ence of A-satu rated models. The
answer is positive, as the next theorem shows.
Theorem 5.4.2 Let A be an i nfi nite cardinal. The n every model A of T
has a A-saturated elementary exte nsion.
Proof. We provide a comparatively simple argument working when A =
For an uncountable A this ap proach does not work any more, and one needs
a more complicated proof of a qui t e different style (but t he result remains
true). So take a model A of T, we are looking for an elementary ext ension
A' > A realizing any l -type over a finit e subset of A'. We do know that
every I-type over any (finit e or infinit e) subset of A is satisfied in some
elementary extension of A. Accordingly, well ord er the set P of l-t ypes over
finite subsets of A, P = {Pv : v < o:} and associate wit h any v < 0: an
elementary extension A(v) of A realizing p. Do this in such a way that , if
v < fL < 0: t hen A(v) is an elementary substr uct ure of A(f-L) (the reader may
check t he det ails as an exercise). Now use t he Elementary Chain Theorem
(1.3.19) and deduce that A* = Uv<a A(v) is an elementary extension of A,
and even of A(v) for any v < 0:; it is also clear that A* reali ze any ty pe
pEP and so every I- t ype over a finit e subs et of A. But this does not imply
that A* is namely realizes every I-type over a finit e subset of
it self. However we can repeat the previous procedu re and define, for every
natural n, an elementary extension An of A in such a way that A
o
is just A
and, for every n, A
n
+
1
is an elementary extension of An realizing any l-t ype
over an arbitrary finit e subset of An. Using the Elementary Chain Theorem
once again, we deduce that A' = U nEN An is an elementary extension of A,
and even of An for every n, Moreover A' is In fact, let X be
a finit e subset of A'; there is some n for which X An; consequent ly every
I-type over X is realized in An+! and , t hrough it , in A'. eT-
Now let us show some basic properties of A-saturated models.
Theorem 5 .4.3 (Weak Homogeneit y Theorem) Let B be a A-saturated model
of T. Let A be a model of T, X be a su bset of A of power < A and f be
an elementary funct ion from X into B. Th en, for eve ry a E A there is an
elementar y funct ion of X U {a} into B enlarging f . (A model B with this
property is called weakly A- ho moge neo us) .
Proof. Let p = tp(aj X), so f(p) is a complet e n-type over f(X) . As f is
1-1, If (X)1 < A. As B is A-satu rat ed, B contains some realization b of f(p);
hence, for every L(X)-formulal'(v, i),
A F l' (a, i) ? l' (v, i ) E P <=> l' (v, f(1)) E f(p) <=> B F l' (b, f(1))
So enlarge f to a function g of X U {a} into B put ting g(a) = b. Clearl y g
is what we are looking for. ..
By definition, a A-saturated model B of T realizes any I- type over a subset
of power < A. But actually B satisfies every ty pe (even with more t han 1
variable) over such a subset . Let us see why.
Coro ll ary 5 .4.4 Let A be an infinite cardinal, B be a A-saturated model of
T, X be a subset of B of power < A, n be a posit ive in teger. Th en every
n -type p over X is realized in B.
Proof. We know that p is realized in some model Ax of Th( Bx), say
by ii = (al ,"" an) E An . As Ax == Bx , t he identity map of X is an
element ary function from X into B. Using the Weak Homogeneit y Theore m
over and over again, one exte nds this map to an elementary function g from
X U {aI , .. . , an} into B. Let b= g(ii). For every L(X)-formu la r.p(iJ, i),
A F r.p (ii, i) <=> B F r.p (b, i ).
Hence tp(bj X) = p. eT-
Theorem 5 .4.5 (Universali ty Theorem) Let B be a A-saturated model ofT,
A be a model ofT of power < A. Th en there is an elementar y embeddi ng of
A in to B (and so A is an elementar y substructure of B up to isomorphism) .
In this case B is call ed A- universal.
Proof. Well order A = {a
v
: v < a} and, for v < a , put Av = {a
jj
:
Il < a }. As A == B, th e empty map is an elementary function from A
o
= 0
(viewed as a subset of A) into B. Using t he A-saturation (actually t he weak
A-homogeneit y) of B, ext end t his map to elementary functions from Av
into B for every t/ < a, whose domain progressively includes every element
in A. One eventually gets an elementary function (and so an elementary
embedding) of A into B. ..
Now we deal with t he "most saturated" models of T; as obser ved before,
t hey are t he mod els sat urated in t heir own power.
Definition 5.4.6 A model B of T is sai d to be saturated if and only if B
is IBI-saturated (and consequentl y realizes eve ry i-type over subsets o] B of
power < IBI).
Such a model B is (also) weakly IBI-homogeneous and IBI-universal.
First let us obser ve t hat, for every infinit e cardinal A, t here is at most one
sat ur ated model of T of power A (uP to isomorphism).
Theorem 5.4.7 (Uniqueness Theor em) Let Band B' be saturated models
of T o] the same power. Th en B B'.
Proof. Recall the Weak Homogeneit y Theor em: as both B and B' are
sat urated, if X B, X' B' , IXI < IBI and I is an elementary function of
Xinto B' wit h image X' (whence IX'I = IXI < IE'I and I-I is an elementary
function from X' into B), t hen
(i) for every a E B, one can enlarge f t o an elementary function g from
X u {a} into B' ,
(ii) for every a' E B', one can enlarge I- I t o an elementary function g from
X' u {a'} into B.
Now observe t hat, as B == B' , the empt y map is an elementary function of
o B int o B' (and conversely) . St art from t his function , well ord er both B
and B' and use alte rnatively (i) and (ii) in a suitable (possibly t ransfinite)
indu ction procedure. One event ually gets an isomorphism bet ween B and
B' .
Let us provide the details of this construction. Let A denot e th e common
power of B and B' ; view A as an initi al ordinal; let {b
v
: t/ < A }, : v <
A }. List B, B' respectively. For every v A, one builds t wo elementary
functions l v, t he former from a subset of B including all t he bjj's wit h
Il < Aand having power < Ainto B' , t he lat t er from a subset of B' including
all t he with J.l < A and having power < A into 8 , in such a way that ,
for u ::; v < A, f v enlarges I; and and enlarges and t;' , and
event ually 1>. and are two isomorphisms, t he one inverse of t he other one,
between 8 and 8'. We proceed by induction on i/,
When t/ = 0, we know what t o do: fa and are just t he empty function.
For a limit t/ , put I; = Uj.l<v fj.l and = Finally suppose v = J.l+1
success or. First build t.: It s domain is I m U{a
v
} ; using (i) , enlarge -1 ,
hence fj.l ' t o include a
v
in t he domain; I, is what we form in this way. Now
const ruct Its domain is I m f v U {b
v
} ; using (ii), enlarge t:' to include
b., in t he domain; let be t he resul ting function. Clearly this machinery
produces 1>. and as required. ...
As a consequence we obtain:
Corollary 5.4.8 (Strong Homogeneity Theorem) Let 8 a saturated model
of T , X be a subs et of B of power < IBI , f be an eleme ntary function of
X into 8. Then f can be enlarged to an automorphism of 8 (and then 8 is
called homogeneous).
Proof. In t he language L(X) consider t he struct ures 8 X and 8 f (X) ' As f is
an elementary function, 8x == 8
f
(x ), and so Bx and 8
f
(x ) are models of t he
same complete theory. As IXI = If( X)1< IBI , 8x and 8
f
(x ) are saturated
also in L(X ). Fin ally 8x and 8
f
(x ) have t he same power. Accordingly t here
is some isomorphism (in L(X )) bet ween 8x and 8
f
(x) , whose restriction to
L det ermines an automorphism of 8 enlarging f . ...
As a particular case, conside r X B, IXI < IBI and two t uples a, a' in
B" having the same n- type over X . The function fixing X pointwise and
mapping ainto a' is elementary, hence can be enl arged t o an automorphism
of 8. As said, this automorphism acts identically on X and maps ainto a'.
Now we wonder for which ca rdinals A a complete theory T may have a
sat urated model of power A. This is a quite delicate and deep question, and
t he answer is not easy. Of course, st ronger ass umpt ions on T may somet imes
ensure t hese existence results. For instan ce, one can see:
if T is A-categorical in some infinite cardinal A, t hen t he unique model
of T of power A is saturated (we will see why for a countable A in the
final section of Chapter 7, where we will also discuss t he uncount abl e
case; t he examples at t he end of t he pr esent section partl y concern
t his point ).
Furthermore the (complete) theories having a countable saturated model
can be characterized in the following way.
T has a saturated model of power if and only if T has at most
countably many n-types over 0for every positive integer n (the reader
may check this as an exercise).
But what we can say from a general perspective, when dealing with arbitrary
complete theories T? In this abstract framework, proving the existence of
saturated models seeems related to some deep set theoretic assumptions,
like the existence of large cardinals. In fact one shows:
Theorem 5.4.9 Let A > be an inaccessible cardinal. Then T has a
saturated model of power A.
The existence of an uncountable inaccessible cardinal is a quite delicate
matter. But assume momentarily that such a saturated model n exists (for
a given inaccessible A > Recall that n is unique up to isomorphism.
Furthermore
n is A-universal: every model of T of power < A can be embedded as
an elementary substructure in n;
n is A-homogeneous: if X is a subset of n of power < A and a, a'
are two tuples in n having the same type over X, then there is an
automorphism of n fixing X pointwise and mapping ainto a'.
It is a general agreement (and habit) in Model Theory to assume that such
a model n exists. This makes things easier and, on the other hand, is
sufficiently plausible; in particular, one can check that everything is shown
inside n, so assuming that n exists, can be proved (at the cost of some
major complications) even avoiding any reference to n.
Which are the benefits of working in n? As we said, the cardinality of n
is very large, and so one can reasonably suppose that all the models we
expect to handle have a smaller power. But this implies that they actually
are elementary substructures of n of a smaller size (up to isomorphism).
Under this perspective, the subsets of models of T can be directly viewed as
subsets of n with power < 1nl: we will call these subsets small subsets of n,
just to tell them from the other subsets of n admitting its same inaccessible
cardinality.
As already said, referring to small subsets of n instead of subsets of arbitrary
models makes our life, and also our notation simpler. For instance, take a
small X and a posit ive int eger n. When defining Bn(X , A) and Sn(X , A), we
had t o explicitly refer to a model A of T (now an elementary substructure of
Q) where X is embedded. We also emphasized that Bn(X , A) and Sn(X, A)
do not de pen d directly on A , but only on the t heory of Ax: t he choice
of A is qui t e arbitrary wit hin these bounds. But t hen it is convenient to
refer to A = Q , as a uni versal model where X is embedded. This is what
we will do from now on. Ind eed we will write, when t here is no danger of
misunde rstandi ng, Sn(X) to mean Sn(X, Q) for every small Xj so S(X) will
denote t he union of t he spaces Sn(X) when n ra nges over posit ive int egers.
At last , let us propose some algebraic examples concerning more or less
sat urated st ruct ures.
Examples 5.4.10 1. Which algebraically closed fields K are sat urated
in a given cardinality A 2:: To answer , we may recall t hat, for any
p = 0 or prime, t he t heory ACF
p
is categorical in every uncount abl e
power; so, according to what we said before, every uncou ntable alge-
braically closed field (in any characteristic) is sat urated. To confir m
t his from t he algebraic point of view and to discuss t he existence of sat-
urated models in t he countable case, we ca n refer to Steinitz's analysis
of algebraically closed fields K and recall that such a K is the algeb raic
closure of Ko(S) , where Ko is th e prime subfield and S is a t ranscen-
dence basis of K , so a maximal algebraically independent subset. The
isomorphism type of K is fully determined by it s characteristic and
it s t ranscendence degr ee, i. e. t he power of S. Moreover, when S is
infini t e, t his transcendence deg ree equals 1/(1. Now let 1/(12:: A and
take a subset H of K of power < A. As we saw in t he last section , H
ca n be replaced by t he subfield generated by H j t his does not change
its size, except when H is fini t e; however, even in t his case, each point
in t his subfield is /i-definable. Every algebraic l -fype over H is clearly
realized in K because K is algebraically closed. So t he point is: ca n
we realize the remaining l-nype, the one of t he elements which are
transcende nt al over H , for any H?
If K has an infini t e t ransce ndence degree, t hen t his degree is just
1/(12:: A > IHI , so it is st rictly larger t han t he t ransce ndence degree
of t he field generated by If. This means t hat we can realize our type
inside K for every H.
Otherwise, let H be generated by a finite t ranscendence basis S of K ,
Clearly IHI < A, but now t here is no way to realize our type inside K.
In conclusion , for 1/(12:: A, K is A-sat urat ed if and only if it s transcen-
dence degree is infinit e. In parti cul ar every algebraically closed field
satisfying t his condit ion is sat ur ated (in it s own power ). So every un-
countable algebraically closed field is sat urated (t his applies also t o t he
complex field), while t he onl y countable sat urated algebraically closed
field (in a fixed characterist ic) is t he algebraic closure of /Co(S ) where
/Co is t he pri me subfield and S is a countable (infinite) algebraically
independent set.
2. The real ord ered field is not This is perhaps remarkabl e,
if one remembers t hat t he reals are a complete ord ered field (in t he
sense t hat every Cauchy sequence has a limit) , and even the onl y
complete ordered field up to isomorphism: in other words, given a
Cauchy sequence (rn)n>o in R , the l -type defined by
1
Iv - rnl < - ,
n
when n ranges over positi ve int egers, has a (unique) realization in t he
real fi eld.
However , look at t he ty pe of a positive infinit esimal nonzero element .
This is defined by t he cut
1
0 < v < - ,
n
for n as before. So it is a ty pe over t he empt y set. Any finite portion
is satisfied among t he reals, so this infinit esimal element lives in some
element ary extension of t he real field. Nevert heless t he ty pe cannot
be realized in R.
Furt her examples will be discussed in th e next section within mod ules.
5.5 A parenthesis: pure injective modules
Saturated models are very large, powerful and rich. Wi thin algebraically
closed fields, t hey remind, and actually coincide with t he uni versal domains
int roduced by Andre Weil in his "Foundations ofAlgebraic Geometry" [176] ;
in fact, as we saw in t he last section, t hey are just t he algebraically closed
fields wit h an infinit e t ranscende nce degr ee (over t he prime subfield) . Hence
it is not surprising t o realize t hat t here do exist in other par ts of Algebra
some notions resembling sat uration, perhaps in a weaker form. This is t he
case of pure injecti vity wit hin module t heory.
5.5. A PARENTHESIS: PURE INJECT/VE MODULES 151
So, fix a (countable) ring R (with identity), and consider (say left) R-
modules. We want to recall in this section what a pure injective R-module
is, and why these modules are so important. First we need introduce a
particular class of embeddings concerning modules.
Definition 5.5.1 Let M and N be R-modules. An R-module homomor-
phism f from M into N if called pure if, for every pp-formula <p(v) in LR
and every ii in M,
Notice that a pure homomorphism is 1-1 and hence is an embedding. To
see this, just apply the definition to the formula Vi = V2. Moreover the
implication =? in (*) is trivial, hence the qualifying point in the definition
of purity When M is a submodule of N, M is called pure in N if
its inclusion embedding is pure; in this case, one says also that N is a pure
extension of M.
The algebraic content of purity is the following: if a linear system Ail = Bx
with parameters ii from M admits a solution in the extension N, then it
admits a solution already in M; here A and B are, of course, matrices with
entries in R and suitable sizes.
Examples 5.5.2 1. If M is an elementary substructure of N, then M
is pure in N. Indeed, (*) holds for arbitrary formulas.
2. If M is a direct summand of N , then Ai is pure in N. In fact, any
solution in N of a linear system as before projects itself to a solution
inM.
3. Let R = Z, so let us deal with abelian groups. We saw in Chapter
2 that, in this particular framework, pp-formulas reduce to torsion or
divisibility conditions. Hence it is not difficult to realize that, over Z,
M is pure in N if and only if rM = M n rN for every integer r , as
usually required in the definition of purity in the handbooks of abelian
group theory.
Incidentally, let us quote a noteworthy result of Sabbagh.
Theorem 5.5.3 Let M and N be R-modules, M be a submodule of N .
Then M is an elementary substructure of N if and only if M is pure in N
and M and N are elementarily equivalent.
Now we ca n int roduce pure injectivity: t his notion restricts in some sense
the usual injectivity requirement to pure embeddings.
Definition 5.5.4 An R-module M is called pure injective p. i. (but some-
one prefers to say algebraically compact) if and only if one of the following
equivalent conditions holds :
(i) For every choice of R -modules N and N', f or which N is a pure
submodule of N' , every homomorphism f from N to M lift s to a ho-
momorphism l' from N' to M .
(ii) M is a direct summand of every pure extension.
The equivalence bet ween (i) and (ii) is proved in the references quoted at
the end of t he cha pter. (i) and (ii) chacterize pure injectivity from t he
algebraic point of view. But t here is a t hird equivalent definit ion, disclosing
a model t heoretic flavour and showing that pure injectivity is directly relat ed
to saturation.
Theorem 5.5.5 An R -module M is pure injecti ve if and only if, for every
countable subset X of M and every (incomplet e) i-type p of pp-formulas
<p (v, ii) with paramet ers ii from X , when every fin ite port ion of p has a
realization in M, then the whole p can be satisfied in M.
Hence pure injectivity is j ust a weak form of sat uration, restricted to incom-
plete l-types of pp-formulas (they are usually called pp-types) . Consequently
every R- module M is p. i.; t his fact , and Theorem 5.4.2 clearly
imply t hat any R-module enlarges to a p. i. elementary extension. But
somet hing much stronger holds.
Theorem 5.5.6 Let M be an R -module. Then M has a p.i. pure extension
M which is minimal in the f ollowing sense: for every pure and p.i. extension
N of M , N contains M as a pure submodule up to isomorphism. Moreover
M is unique up to isomorphism fixing M pointwise. Finally, M is an
eleme ntary extension of M.
For a proof, see t he refer ences at th e end of th e chapter.
These exist ence and uniqueness result s j ustify the specific symbol A1 to
denote this p.i. extension of A1 and, furthermore, a special name to dub it;
actually, M is called "the" pure injective hull of M .
5.5. A PARENTHESIS: PURE INJECTIVE MODULES 153
Up to eleme ntary equivalence, p. i. R- modules represent t he whol e class of
R-modules, hence their algebraic analysis can help a description of the pos-
sible complet ions of RT. But what can we say about a possibl e classification
of pure injective R-modules? A usual and general t echnique in studying a
given class of modules is
(a) to look for a possible (and hopefully unique) direct sum decompositi on
of any module of the class into indecomp osable object s,
and t hen
(b) t o classify the indecomposable modules of the class up to isomorphism.
By the way, recall t hat an R-module M is call ed indecomposable if M =J. 0
but there is no way to express M as a direct sum of t wo nonz ero submodules.
This procedure is not always successful within arbitrary classes of modules,
but is sufficient ly sat isfactory when applied to p. i. modules. Let us explain
why. First of all it is comparatively easy t o check t hat pure inj ectivity
is pr eserved by direct summands . Moreover (a) does wor k wit hin p. i.
modules, due to t he following result of Fisher and Ziegler.
Theorem 5.5.7 (Fisher-Ziegler) Let M be a p.i. R-module. Then M de-
composes uniquely (up to isomorphism) as
ffiiUi ffi E ,
where E has no indecomposable direct summand, denotes p. i. hull
and, for every i , Ui is an indecomposable p.i. R-module. Moreover M is
elementarily equivalent to ffi# i .
Incidentally, pure inj ectivity is not preserved under infinit e direct sums . To
summarize what we have seen so far:
every R-module is an elementary subst ruct ure of its p.i . hull ;
every p.i. R-module is elementarily equivalent to a direct sum of in-
decomposable p.i. 's, and even isomorphic to th e p.i. hull of this sum
up to a summand with no indecomposable direct factors.
So what we have to do now is t o study isomorphism classes, or elementary
equivalence classes of indecomposabl e p.i. R-modules.
Cl assification up to isomorphism is sometimes easy. For instance, when R is
a field, so when we are dealing with R-vectorspaces , t he onl y indecomposable
object is just R (as a vectorspace over itself) and is pu re injective.
But over oth er rings t he situation changes dramat ically and in some cases
one has t o conclude t hat no classification is possible. This is what hap-
pens, for inst ance, when R = K (x, y) is t he ring of polyno mials over a
field K in two non-commuting variables x and y , so when one considers
K-vect orspaces wit h two additional linear operators, corresponding to t he
action of x, y respecti vely; in fact , t his class of mod ules encodes in some way
t he word pro blem for groups, and t here are some good reasons to believe
t hat t his forbids any class ification of indecomposable p.i. objects. See [136]
for a det ailed discussion of this point. As a consequence, even modules over
K[x, y] (wit h two commut ing unknowns x and y) and, in another di recti on,
over K (Xl, ... , x
n
) wit h n > 2 inherit t he same "wild" sit uat ion, excludi ng
any possible classi ficat ion.
So t he key point is t o realize for which rings R indecomposable p.i. modules
ca n be classified up to isomorphism. In t his perspect ive we would like to
quote a beau tiful result of Ziegler, using model t heory to equip in a nat ural
way indecomposable p.i. mod ules over any fixed ring R wit h a to pological
space struct ure.
In fact , let RZg denote t he set of (isomorphism classes of) indecomposable
p. i. R-modules. For every choice of pp-formulas rp (v) and 'IjJ (v) (in one free
variable) in LR, let (rp / 'IjJ ) be the set of the elements U in RZg such that
rp (U) properly includes rp(U) n 'IjJ (U). Then the (rp / 'IjJ )' s are a basi s for a
topology of RZg, which is always compact and seldom Hausdorff. RZg with
t his topology is called the (left) Ziegler spectrum of R. Again, see [136] or
directly [181] for more details. What we can say in t he restricted framewor k
of these pages is t hat t he knowledge of t he Ziegler spectrum of R is a sort of
fixed course towards t he solution of several significant mode l t heoretic (and
also algebraic) pro blems about R- modules.
In particular, a successfu l analysis of the Ziegler spectrum can provide some
useful informat ion on t he elementary equivalence class of any R- module M .
Let us see very briefly and roughl y why. We saw in Cha pter 2 t hat t he
complete t heory of M is fully det ermined by t he values (modulo 00) of
[rp( M) : 'IjJ (M )] where rp (v) and 'IjJ (v) range over pp-formulas in one free
variable in LR. Now, up to elementary equivalence, M ca n be replaced by
a direct sum of indecomposabl e p. i . R-modules G7iUi , where
is th e canonical decomposit ion of the pure injective hull A1 of A1 according
to the Fisher-Ziegler Theorem. Consequently, for rp( v) and 'IjJ (v) as before,
[rp(M) : 'IjJ (M )] = [rp( G7iUi) : 'IjJ (G7iUi )]
5.5. A PARENTHESIS: PURE INJECTIVE MODULES 155
modulo 00 (in the sense t hat th ese values are eit her finit e and equal or both
infinit e) . Now it is easy t o realize that
[<p (EBiUd : 7Ji (EBiUi )] = : 7Ji (Ui )]
1
modulo 00. So, when in t he Ziegler spectrum t he points (Le. t he indecom-
posable p.i. R-modules) are classified as well as th eir basic open neighbour-
hood s, th e element ary equivalence t ype of M can be det ermined by saying
which U ER Zg are involved in th e decomposition EBi Ui and how many t imes
t hey occur. For instance, if U is known to be isolat ed -say by (<p / 7Ji )-, then
t he value (modulo (0) of [<p(M) : 7Ji (A1)] can witness if U occurs in th e de-
composit ion, and how many ti mes. Of course, for a non-isolated U, a finer
analysis is necessary. In conclusion , when the spect rum is known, looki ng
at th e points in RZg and at their open neighbourhoods, one can effect ively
list t he complet e extensions of RT . Let us illustrate this by some examples.
Examples 5.5.8 1. Let R be t he ring Z of integers. Recall that every
pp-formul a <p (v) of Lz is logically equivalent within zT to a conjunc-
t ion of t orsion or di visibilit y condit ions
where p is a prime, l < k are positive int egers and r is an integer. An
effect ive list of indecomposable p. i. Z-modules is known , and includes
(up to isomorphism) exactly th e following object s:
the finit e modules Z/ pnZ ,
th e Priifer groups Z/ pco ,
th e p-adic int egers Z (p) ,
the additive group of rationals Q
where p ranges over primes and n over positive int egers; t he p-adic
integers are j ust th e p. L hull of the localization Z (p) of Z at p. It is not
prohibitive to realize t hat t he pr evious examples are indecomposabl e
and p. i.; for instance, t he pure injectivity of th e Priifer groups and
t he rationals comes directl y from th eir divi sibili t y. But t he point is
to show t hat t heir list exha usts all t he possible cases: see [181] for a
discussion, and a complete proof of t his.
Now we have to study t he t opology of t he Ziegler spectrum z Z g. Here
is it s Cant or-Bendixson analysis.
The isolated point s (those having rank 0) are t he finit e modules
Z/pnz; in fact, each of t hem is t he only eleme nt in t he open set
((pn- 1Iv 1\ pv = 0) / (pnlv 1\ pv = 0)).
Now forget the Z/pnZ's; t hen a Priifer group Z/ p= gets isolat ed
by ((pnlv 1\ pv = 0) / v = 0) for any positive n, while t he p-adi c
int eger s are isolat ed by (pn- 1Iv / pnlv) for any n; hence all t hese
groups have ra nk l.
The rational group is t he only remaining point , and so gets rank
2.
This analysis was pursued by Eklof-Fisher in t he particular case when
R is a principal ideal domain, and supports and clarifies the previ-
ous work of Wanda Szmielew (sket ched in our Cha pter 2) for ab elian
groups . In particul ar, as a consequence, it impli es t he decidability of
t he t heory z T. Ziegler 's approach was inspired by t hese par ti cul ar
cases, but is fully general and covers any ring R.
2. Hence what has been said on Z act ually enlarges t o principal ideal
domains and, partl y, to Dedekind domains. For instance, it applies t o
t he ring K[x] of polynomi als over a field K wit h a single unknown: also
in t his case one can accomplish a quite satisfactory descripti on of t he
Ziegler spectrum, essent ially repeating t he analysis for the int egers
wit h t he necessary variant s. Compare t his and what was obser ved
befor e when t he number of unknowns increases.
5.6 Omitting types
Sat ura ted models reali ze many ty pes . But non isolated t ypes are not easy to
sat isfy in ar bit rary models. Let us see why. We conside r a complet e theory
T, and we denote it s uni verse by n. Recall that we met isolated types in
5.3: a ty pe p over a set X n is isolated when p is t he onl y type over X
containing some L(X)-formula cp(v). It is an easy exercise t o check what
follows.
Fact 5.6.1 For p E S(x), t he following propositions are equivalent .
1. P is isolat ed;
2. t here is some L(X )-formula cp (v) E p such t hat p is just t he set of
L(X )-formul as 7f; (v) for which cp( nn) 7f; (nn);
5.6. OMITTING TYPES 157
3. there is some L(X)-formula tp(v) such that tp(Dn) =f. 0 and p is just
t he set of the L(X)-formulas 'IjJ (if) for which tp(Dn) 'IjJ (Dn).
Actually what distinguishes 2 and 3 is th at in 2 we require t he additional
condition tp(v) E p. But it is easy to see t hat tp(v) (j. p forces ,tp(v) E p,
hence tp(Dn) ,tp(Dn) , and so tp(Dn) = 0, a cont radict ion.
Now an isolated t ype p over X is trivially realized in every mod el M of
Th(Dx ). For , D does cont ain some tuple satisfying p, and in particular the
formul a tp(v) isolating p. So :3vtp(v) E Th(Dx).
On the contrary a non-isolated type p is not always realized. This is the
content of t he following result.
Theorem 5.6.2 (Omitting Types) Let T be a complete theory in a language
L, p be a non-isolat ed type over 0. Then there exists a (countable) model
M of T omitting p (in the sense that p(M) = 0).
Proof. For simplicity, assume p E SI (0) (the reader can easily adapt the
following argument to more variables, if he (she) likes). We use some usual
techniques of classical Model Theor y. In detail , first we exte nd our language
L by countably many new constant symbols
co, Cl, . .. , C
n
, . .. (n natural).
List the sentences of the enlarged language L' = L U {c
n
: n E N}
Now we build an increasing sequence of consistent theories in I
To T
l
... T; .. . ,
all enlarging T and satisfying the following conditions, for every n:
1. T
n
is axiomat ized by To and a unique further I-sent ence ()n;
2. eit her tpn or ' tpn is in Tn+I;
3. whenever tpn or ' tpn is in T
n
+
l
and has the form :3v'IjJn(v) , then
'ljJn (c
m
) E T
n
+
l
for some new constant Cm not already occuring in
any tpi, ()i for i ::; n;
4. finally, t here is some formula O"n( v) E P for which 'O"n( c
n)
E T
n
+
l
.
158 CHAPTER 5. MOREY RANK
To is just T, or , more precisely, t he I- theory axiomatized by T. Now,
ass ume T
n
given, and let us build Tn+l'
Let Cia ' ... , Ci
h
_
1,
C be t he constants in ' - L occuring in ()n ; replace t hese
constants by new variables Wo, ... , Wh- l , v chosen out of ()n ; place before
:Jwo, ... , :JWh-l and get a for mul a (v) in L , with t he only free variable
v . Take a model of T
n
; as o; E T
n
, e; is not empty (it includes
As p is not isolated, t he condit ion 3 in Fact 5.6.1 ensures t hat, for
some formula un(v) E p,
Pu t ,Un(C
n
) in Tn+l ' This gua ra ntees 4. Now look at 'Pn. If T
n
U {,Un(c
n),
'Pn} has some mod el, let T
n
+
1
include also 'Pn . Oth erwi se ' 'Pn is a conse-
quence of T
n
U {,Un(c
n)}.
This gives 2. When 'Pn (or ''Pn) is of t he form
:J v'Pn(v), pick t he least m such t hat Cm does not occur in 'Pi, ()i for i n,
and put 'ljJn(c
m
) in T
n
+
1
This ensures 3.
Let T
n
+
1
be t he t heory obtained in t his way. Clearly Tn+l is consistent and
satisfies l.
Now form T' = Unr; T' is consistent. Take any model l v1' of T'. By 2,
T' is complete , T' = Th( M '): t he sentences :Jv'ljJ (v) M' satisfies are just
those occu ring in some T
n
. Owing to 3, we can apply the Tarski-Vaught
Theorem and deduce that : n natural} is the domain of a (countable)
elementary substruct ure of M' , and hence a countable model of T' . Its
restricti on t o L is a countable model of T , and omits p because, for every n,
can not satisfy un(v). ..
5.7 The Morley rank, at last
Let T be a complet e theor y in a count able language. We remind the pro-
gram outlined in 5.2: we aim at measuring t he complexity of definable sets
X in n (equivalently, of r-v-classes of formul as with par amet ers in n) and,
more generally, of sets of formulas with par amet ers in n, including ty pes
over small subsets of n. A reason abl e way to get t his is to define some
function assigning, if possible, to every definabl e, or formul a, or type, an
ordi nal value, according to some reasonabl e condit ions, like monotonicit y.
An axiomatic introducti on to t his complexity measures ca n be found in [8]
or [131, 134]; t hese functi ons are usually called ranks. There are several
possible ways t o define a rank, according to some part icular algebraic or
model t heoretic featu res of t he theory T . Morley's rank was t he very first ,
5.7. THE MORLEY RANK, AT LAST 159
and fund amental , example in t his direction. Historically, it was inspired
by t he Cantor-Bendixson analysis of t opological spaces S. Recall t hat t his
equips any point in S with a rank, whose value is eit her an ordinal number
or 00 (where one assumes t hat 00 is gr eater than any ordinal) ; in parti cul ar
the Cant or-Bendixson ra nk assigns the value 0 t o t he isolated points of S,
th e value 1 t o t he point s in S t hat get isolated wit hin t he relati ve t opology
once th e isolated point s are forgotten, and so on. Of course one can apply
this Cant or-Bendixson machiner y t o t he topological space S(C) , where C
is a small subset of n , and hence to the types over C . Using t he Stone
du ality between topological Bool ean spaces and Bool ean algebras , one can
define a Cantor-Bendixson rank also for th e elements of t he dual algebra of
S(C) , and hence for C-defina ble sets. However , if one translates literally
th e Cantor-Bendixson analysis into our particular set t ing, then the same
definabl e object may obtain several possible ranks, according to which basic
set of parameters C we refer t o; in particular , if we replace C by a larger C',
t hen we ca n expect t hat, over C', C-definable sets get a st ronger complexity
and consequent ly a bigger Cant or-Bendixson rank: some examples of t his
phenomenon will be provided lat er in t he present section.
Morley's rank adapts t he Cant or- Bendixson approach t o obviate this diffi-
culty. The recipe is just to refer to our uni verse n and so t o evaluate any
definabl e within ar bit rary n -definabl e sets. But it is t ime t o give, at last, t he
exact definition of t he Morley rank; we will introduce also a related notion
(Morley degree).
vVe consider a complete t heory T in a countable language L: n denotes,
as usual, t he universe of T . Let X be a definable subset of n n (for some
positive int eger n), we wan t t o define the Morley rank of X RM(X). First
let us say what
RM(X) ::::: 0:
means for every ordinal 0:. We proceed by induction on 0: .
Defi nition 5.7.1 When 0: = 0, put RM(X) ::::: 0: if and only if X i= 0.
When 0: is a limit ordinal, put RM(X) ::::: 0: if and only if RM(X) ::::: 1/
fo r every ordinal 1/ < 0: . Fi nally, when 0: = 1/ + 1 is a successor ordinal,
put RM(X) ::::: 0: if and only if there are infi ni tely many pairwise disjoint
defi nabl e subsets Xi (i E N) of X such that RM(Xi ) ::::: 1/ for every i EN.
It is easy t o observe t hat, if 0:, {3 are ordinals and 0: ::::: {3, t hen
RM (X) ::::: 0: impli es RM (X ) ::::: {3.
Accordin gly it makes sense to put:
160
Definition 5.7. 2
CHAPTER 5. MORLEY RANK
1. If X = 0, then RM(X) = - 1.
2. Let X#-0 and suppose that there is some ordinal a such that RM(X) t.
a (so RM(X) t. (3 for every ordinal (3 2: a); the least ordinal aD with
this property is a successor; if aD = v + 1, then put R.A1(X) = t/,
3. Finally suppose X #- 0, RM(X) 2: a for every ordinal a. Then put
RM(X) = 00.
(Of course we assume - 1 < a < 00 for every ordinal a) . When <.p( v) is a
formula (possibl y with parameters from n), define
(where n is the length of v).
Examples 5.7.3 1. RM(X) = - 1 if and only if X = 0.
2. RM(X) = 0 if and only if X is finite and non-empty. For instance, in
algebraically closed fields, t he (definable) zero set of a given polynomial
of degree 2: 1 in one unknown is finite and non-empty, and hence has
Morley rank O.
RM(X) = 1 if and only if X is infinite, but cannot partition as the
union of infinitely many pairwise disjoint infinite definable subsets.
For example, in algebraically closed fields, the only definable infinite
1-ary sets are cofinite. This implies t hat their Morley rank is 1. Notice
that the same is true for infinite vectorspaces over a countable field.
4. Let T = DLO- be the theory of dense linear orders without endpoint s.
Fix a < b in n and consider the int erval I =]a, b[= {x En: a <
x < b}. I is definable, and RM(J) = 00. In fact, for every a, b as
before and ordinal a , RM(]a, bD 2: a. This can be eas ily checked by
induction on a. The cases a = 0 and a limit are trivial. So suppose
a = v + 1 for some u, As the ord er S; is dense, we can choose c E
]a, b[, and observe that ]a, b[ includes the disjoint intervals ]a, C[, ]c, b[,
both having Morley rank 2: u, because of th e induction hypothesis.
Repeating this procedure, one finds infinitely many pairwise disjoint
open intervals in ]a, b[, all of Morley rank 2: u, Hence RM(]a, bD 2:
t/ +1. Notice that even [a, b[, ]a, b], [a, b] (for a < b) have Morley rank
00.
Let us point out now some simple properties of RM:
P roposi tion 5.7.4 Let X , Y be two definable subsets of nnfor some pos-
itive integer n,
(i) If X Y, then RM(X) :S RM(Y ).
(ii) For every automorphism f of n, RM(X) = RM(J(X)) .
(iii) RM(X U Y) = max{RM(X) , RM(Y )}.
Proof. (i) Every definable subset of X is clea rly a definable subset also
of Y. So a simple induction argument shows t hat, for every ordinal a, if
RM(X) 2:: a, then RM(Y) 2:: a .
(ii) It suffices to show RM(X) :S RM(J(X)) (as th e opposite relation
RM(X) 2:: RM(J(X)) can be obtained just by reversing t he roles of X
and f(X ) and repl acing f by f-l) . In oth er words, we have to check t hat,
for every ordinal a, RM (X) 2:: a impli es RM (J (X )) 2:: a. Proceed again
by induction on a (and observe th at , if Z is a defina ble subset of X , t hen
f(Z) is a definable subset of f(X)) .
(iii) 2:: follows from (i) . To check :S, it suffices to prove th at, for every ordinal
a, RM(X UY) 2:: a impli es eit her RM(X) 2:: a or RM(Y) 2:: a . As befor e,
we can proceed by induction on a . If a = 0 and RM (X U Y) 2:: a, t hen
X U Y i= 0 and consequent ly eit her X or Y is not empty. Now suppose
a limit and RM(X U Y) 2:: a, namely RM(X U Y) 2:: t/ for every ordinal
t/ < a. If th ere exists some v < a such th at RM(X) 2:: J1 for every ordinal
J1 sat isfying v :S J1 < a , t hen RM(X) 2:: a . Ot herwi se, for every ordinal
v < a, th ere is some J1 2:: t/ such t hat J1 < a and RM (X) t J1 . By indu ction,
RiV1 (Y ) 2:: J1 , whence RM (Y ) 2:: u, It follows RM (Y ) 2:: a . At last , take a
successor a = v+ 1 and suppose RM (X UY ) 2:: a . Then t here are infinit ely
many pairwi se disjoint infinite definabl e subsets Z, (i E N) of X UY, all
having RM 2:: u, Look at t he sets
(i E N) .
For every i, eit her RM (Xi) 2:: t/ or RM (Y;) 2:: u, Accordingly RM (Xi) 2:: u,
or RiV1(Y;) 2:: v for infinitely many i 's. In t he former case RM (X) 2:: a; in
t he latter RM(Y) 2:: a .
Now we are going t o associate with any definable subset X nn (whose
Morley rank is an ordinal) a positi ve int eger: t he Morley degree of X
GM (X).
Proposition 5 .7.5 Let X nnbe a definabl e set whose Mo rley rank is an
ordinal 0:' . Th en there is a maximal positive integer d such that X partit ions
as the union of d pairwise disjoint definable subsets of Morley rank 0:' .
Proof. Suppose not. T hen, for every positi ve int eger d, one can partition
X as t he union of d definabl e subsets of Morley rank 0:'
X(d,O), .. . , X (d, d - 1).
We ca n assume t hat, for every d and i ::; d, ther e is some j < d such
that X (d + 1,i) X(d,j). In fact, X(d+ 1,i) decomposes as th e disjoint
union of it s subsets X (d + 1, i) n X (d, j) for j < d. Owing to Proposition
5.7.4, (iii) , at least one of these subsets has the same Morley rank 0:' as
X(d + 1, i). By Proposition 5.7.4, (i), all th ese subsets have Morley rank
::; 0:' . Accordingly we can repl ace X(d+1 , i) by a subset X(d+l , i )nX (d, j)
(to whi ch one possibly adds other subsets X (d, j) of Morley rank s, 0:') . Now
a combinatorial argument (K6nig's Lemma) ens ures t hat t here are posi ti ve
int eger s
do < d
l
< .. . < d
m
< ...
and natural number s
. . .
ZO,ZI , ,Zm,
such t hat, 'IIm EN, i
m
< d
m
and
(m E N )
(m E N)
We obt ain in this way infinitely many pairwise disjoi nt definabl e subsets of
X
X(d
m
, i
m
) - X(d
m
+
l
, i
m
+
l
)
(wher e m ranges over N), all having Morley rank 0:' . Consequently RM(X) 2:
0:' +1, and t his is a contradiction. '"
Owing t o t his proposition, we can at last introduce t he Morley degree as
follows.
Defi nit io n 5. 7.6 Let X be a defi nable subset of nn(f or some posit ive inte-
ger n) whose Morley rank is an ordinal 0:'. Th e Morley degree of X (den oted
GM (X)) is the maximal posit ive integer d such that X can partit ion as the
union of d defi nable subsets of Morley rank 0:'.
Remarks 5.7.7 1. If X is finit e and non-empty (in other words RM(X) =
0) , GM(X) is just the size IXI of X. In particular, when X is the zero
set of a polynomial f(x) of positive degr ee in one unknown over an
algebraically closed field /C, then the Morl ey degr ee of X is just the
number of roots of f( x) in K.
2. A definable X having Morley rank 1 and Morley degree 1 (so an infinite
definabl e X admit t ing no infinit e coinfinite definable subset) is called
strongly minimal. A t heory T is st rongly minimal if it s universe (so
the formula v = v) is: we met this notion in Chapter 2, and we shall
deal again with it in the next sect ions. Recall t hat algebraically closed
fields, as well as infinite vectorspaces over a countable field, admit a
st rongly minimal t heory.
3. Let X , Y be two definable disjoint subsets of nn such t hat both
RM(X) and RM(Y) are ordinals, and RM(X) < RM(Y). Then
GM(X U Y) equals
GM (X ) + GM(Y) if RM(X) = RM(Y) ,
GM(Y) otherwise
(the reader may check this as an exercise).
When <p (v) is a formula (possibl y with parameters from n) and RM(<p (nn))
is an ordi nal , we put GM(<p(v)) = GM(<p(nn)).
Now we want to discuss the following problem. Let X be a definable subset
of nn (for some positive integer n), 0' denot e t he Morley rank of X and d it s
Morley degree (if any) : both 0' and d can be calculated by looking at the de-
finabl e subsets of X . But suppose t hat X is M-definable for some mod el M
of T. Then we wonder wheth er t he values of 0' and d are already witnessed
by t he M-definabl e subsets of X , in ot her words whether 0' and d remain
t he same after repl aci ng "defina ble" by "M-definabl e" in t he definition of
RM and GM.
Examples 5.7.8 1. We know that RM(X) = 0 if and only if X is finite ;
furthermore, in this case , GM(X) is just t he size IXI of X. Suppose
X = <p (nn, if) for a suit able tuple if of parameters of a model M of
T. As M is an elementary subst ruct ure of n, X = <p (M n, if) . So the
elements of X in M can witness RM(X) = 0, GM(X) = IXI.
2. RM(X) = 1 if and only if X is infinite, but X cannot contain infinit ely
many pairwise disjoint infinit e definabl e subsets. As before, suppose
x = <p(nn, a) for a in M F T. As M < n, <p( Mn, a) is infini te, so
M can witness RM(X) 2: 1. It is also clear t hat <p( Mn, a) cannot
contain infinitel y many pairwise disjoint infinite AI -defina ble set s.
3. Now consider a str uct ure Mo = (kI
o
, E) where E is an equivalence
relation having exactly one equivalence class of size k for every posi-
t ive integer k , but no infinite class. Let T denot e t he theory of Mo.
One easily sees t hat Mo is element arily embeddable in every mod el
of T . But a simple a pplicat ion of Compactness Theorem shows that
there exist some other models of T containing also infinite equivalence
classes, and indeed, for every positive integer t , T has a count a ble
model M
t
with exactly t infinite classes, while the universe n has in-
finitely many infinit e classes. Now consider th e formula v = v. The set
it defines is just n, and consequentl y has Morley rank 2: 2, because it
partitions into infini t ely many infinite pairwise disjoint definable sub-
sets (the infinite classes of E). Indeed one sees t hat n has Morley
rank 2 and Morley degr ee 1. On t he oth er side, t here is onl y one non-
algebraic I-t ype over Mo as, for every a and b in n - Mo, one can
find an A1
o
-aut omor phism of n mapping a in b. In parti cular, every
lVf
o
-definable set is finit e or cofinite according to whether it excludes
or includes the elements in n - M. Hence n cannot parti tion in t wo
infinite Mo-definabl e subsets, and M o cannot wit ness RM(n) 2: 2.
So we ca n wonder which models A1 of a given T can witn ess t he Morley
rank and degr ee of every M-defina ble set . As a par tial answer, let us show
that models have t his feature.
P roposition 5.7.9 Let T be a complete theory, M be an model
of T . Then A1 can witness the Morley rank and (if it exists) the Morley
degree of any M-definable set X.
Proof. Put X = <p(nn, a) for some suitable positive integer n and a E M .
First let us deal with RM(X) . It suffices to show t hat , for every ordinal
u, if RM(X) 2: v + 1, th en th ere ar e infinitel y many pairwise disjoint M-
definable subsets of X , all having Morley rank 2: u. On t he other side we
do know t hat t here are infini te ly many pairwise disjoin t subsets Xi S;; X
(i EN) , all having RM 2: u, For every i E N, let Xi = <Pi(nn , ai) with
ai E n. As M is for every natural k t here ar e ... , b: in
M such t hat tp(aa , .. . , ak la) = ... , b:la). Consequentl y, for every
i k , S;; <p(n n, a) because V'W( <Pi(W, vi) -+ <p(w, a)) lies in the
ty pe of (aa, .. . , ak ) over a. Moreover , for i < j k , n <p( nn, bj) =
obecause .....,:3 W(ipi(W, Vi) /\ ipj(w,vj) ) is in tp(aa , . .. , a"k l a) . Not ice also
t hat, Vi :s; k , ai , correspond t o each other by some automorphism of n
fixing a pointwise, and consequent ly ipi( nn, ai) and ipi( nn, b; ) have the same
RM u, Finally, for every i :s; k , ip(nn, is M-defina ble because
is in M . In conclusion, for every natural k , X contains at least k + 1 M -
definabl e pairwise disjoint subsets of Morley rank u, So, t he combi nat orial
argument in Proposition 5.7.5 ca n be repeat ed and shows our claim.
Now let RM(X) = 0' where 0' is an ordinal , d = GM(X) . We have to show
t hat X can partition as the union of d M-defin abl e subsets of Morley rank
0'. Again we know that X does admit such a decomposition in definabl e
subsets
ipo(nn, a"O) , .. . , ipd-I (nn, ad-I)
with aa, .. . , in n. On t he oth er hand tp(aa, . .. , ad-=-I la) can be realized
in M , say by ... , bd-I, because M is By proceeding as
before, one sees that
provide a partition as required.
Now we want t o define t he Morley rank and degr ee of a complete ty pe p over
a small subset A of n: t he Morley rank of p RM (p) is an ordinal or 00, while
t he Morley degree GM(p) is defined onl y when RM (p) is an ordinal , and is
a positi ve int eger. The reader ca n eas ily observe t hat t his ass ignment of a
Morley rank and degr ee can be eas ily extended to arbit ra ry set of formul as
over A (alt hough, in t his enla rged framework, the Morley rank may ass ume
also the value -1 -for inconsistent set of forrnul as-).
Defi nition 5.7.10 Let A be a small subset of n, p E S(A) . The Morl ey
rank of p RM(p) is the least Morles) rank of a formula in p. If RM(p) is
an ordinal 0' , GM(p) (the Morley degr ee of p) is the least Morley degree of
a formula in p of Morley rank 0' .
Accordingly, with any p E S(A) it is associated a for mul a ip(v) E p such that
RM(ip(v)) = RM(p) and, when RM(p) is an ordinal , GM(ip(if)) = GM(p) .
Of course, one may wond er whether t his formul a ip(v) is unique . To clarify
t his question, first noti ce:
P roposit io n 5. 7. 11 Let p E S(A) satisf y RM (p) < 00, ip(v) be a formula
in p having the same Morleu rank and degree as p. Let ip'(v) be any for-
mula with parameters f rom A. Then ip'(v) E p if and only if RM (ip(nn) -
ip'( n n)) < RM (p).
Proof. First suppose <p'(v) E p. Th en <p(if) /\ <p'(v) E p; as RM (<p( v) /\
ip' (v)) ::; RM(<p (if)), <p (v) /\ <p' (v) have t he same Morley rank and degree as
<p(v) , hence as p. Consequentl y RM (<p (if) /\ ' <p' (v)) < RM(p).
Conversely, suppose RM (<p(n n)_<p' (n n)) < RM (p), so <p (v)/\ ,<p' (v) ca nnot
belong to p. In parti cul ar ' <p' (if) rt p, and so <p' (if) E p. ..
At t his point one deduces:
Corollary 5.7.12 Let pE S(A) satisf y RM(p) < 00, <p(v) be a formula in
p having the same Morley rank and degree as p. Let r.p' (v) be a formula of
L(A) also having the same Morley rank and degree as p. Then <p' (v) E p if
and only if Rj\![(r.p(nn) 6. r.p' (nn )) < RM(p).
Here 6. denotes symmet ric difference: for X , X' sets , X 6. X' = (X - X') U
(X' - X ).
Proof. Let r.p'( v) E p. Owing to Proposition 5.7.11, RM( r.p(nn) _ <p'(nn)) <
RM (p). But now we can reverse t he roles of <p (v) and <p' (if) , and so deduce
RM (<p (n n) - <p'( nn)) < RM (p). By Proposition 5.7.4, (iii) , RM(<p(nn) 6.
<p'( n n)) equals max{RM (r.p (n n)-<p' (n n)), RM (r.p' (n n)_ r.p (n n))} , and hence
is < RM (p). Conversely suppose RM (<p (nn) 6. r.p'( n n)) < RM (p). Th en
RM(<p(n n) - r.p'( n n)) < RM (p) as well, and so <p'(if) E p. ..
It is easy to realize t hat , for every ordinal a , t he relation =aidentifying two
L(A)-formul as <p (if), <p' (if) if and only if
is an equivalence relation. Accordingly, when RNf (p) is an ordinal a, a
for mul a r.p(v) E p with t he same Morley rank and degr ee as p is uniquely
det ermined (in p) up to =a.
Remarks 5.7.13 1. Let A B be small subsets of n, p E S(A) , q E
S(B) , p q. Then RM(p) ;::: RM(q) and, when RM(p) = RM(q) ,
GM(p) ;::: GM (q). For , all t he formul as in p belong t o q as well.
2. Let p E S(A). Then p is algebraic if and only if RM(p) = 0 (recall
t hat p is algebraic if and onl y if t here is a formul a r.p( if) E P such t hat
<p( n n) is finit e and non-empty, in other words has Morley rank 0). We
saw t hat, for ty pes in a suitable setting, algebraic just means isolat ed:
in t his way we recover in our frame work a very basic and familiar
propert y of t he Cant or-Bendixson rank. For algebraically closed fields
(of a fixed characteristic) t he only nonalgebraic I- typ e over a small
5.7. THE MORL EY RANK, AT LAST 167
subset of n has Morley ra nk 1. The same is true for every strongly
minimal t heory.
So we have just introduced Morley ra nk and degree for t ypes p E S(A) by
referring t o Morl ey rank and degr ee of formul as . Conversely, for every L (A)-
formul a <p (v), we can recover the Morley rank and degree of <p(v) (and of
th e corresponding A-definable set <p (nn)) from the Morley rank and degree
of the types over A cont aining <p (v) (provided t hat, of course, t here is some
t ype over A cont aining this formula, in other words <p(nn) i= 0) .
Proposition 5.7.14 Let A be a small subset ofn, <p (v) be an L(A)-formula
(in the free variables v = (Vt, ... , v
n
)) such that <p (nn ) i= 0. Then
RM(<p(nn)) = max{RM(p) : pE Sn(A) , <p (v) E p}
and, when R1v1.(<p(nn)) is an ordinal 0' ,
GM(<p(nn)) = L GM (p)
p
where p ranges over the n- types over A containing <p (v) and having rank 0' .
Proof. Clearl y, if pE Sn(A) and <p(v) E p, then RM (p) RM(<p (nn )). So
we have preliminarily to show that there exists at least one ty pe p E Sn(A)
containing <p(v) and having it s rank.
First suppose RM(<p(nn)) = 00. By compactness the set of L( A)-formulas
is consistent. In fact , let 'l?o(v), . . . , 'l? s(v) be L( A)-formul as of rank <
00; by Proposition 5.7.4, (iii) , also Vi<s'l?i(V) have rank < 00. Moreover
<p (nn) Cl:. Ui<s 'l?i(nn), otherwise even i(v) has Morley rank < 00. Accord-
ingly <p (nnf n (ni<s -,'l?i(nn)) i= 0. So exte nd Po to a type p E Sn(A);
<p( v) E p, and p has rank 00.
Now suppose RM(<p(n n)) = 0' where 0' is an ordinal. Decompose <p (v) as
Vi<s <Pi(V) where <po(v)"",<Ps(v) are L(A)-for mulas of Morley rank 0' and
minimal Morley degr ee and, for i < j < s, RM(<pi(nn) n <pj (nn )) < 0' .
The formul as <Po(v) , . .. , <Ps (v) occurring in this decomposition are uniquely
determined up to = CI'l as they have mini mal degree. Furthermore the Morley
degree of <p (v) is the sum of the degrees of <Pi (V) for i s. Now fix i s
and consider
qi = {<Pi (v)} U {-,'l?(v) : 'l?(v) L(A)-formula, RM(<pi(V) A 'l?(v)) < O' }.
By proceeding as in t he previous case, one checks t hat qi is consist ent, and
so enlarges to a type Pi E Sn(A). Furthermore t his ty pe Pi is unique: in
fact , if fl( v) is an L(A)-formula and RM(<pi(V) 1\ fl(v)) = 0: , t hen, owing
to th e choice of <p (v) (and the minimalit y of it s degree over A), it follows
RM(<p (v) 1\ -,fl( v)) < 0:, whence fl(v) is in Pi. Clearly RM (Pi ) = 0:. Using
t he minimalit y of t he degree of <pi (V) once again, one deduces GM(Pi) =
GM(<pi(V)) . Finally <pi(V) E Pi , and so <p (v) E Pi.
So we have found ty pes Po, . . . , Ps E Sn(A) as claimed. To accomplish t he
proof (in particular , t o show t he equality about degrees) it suffices t o observe
that the only n-types P over A containing <p( v) and having rank 0: are j ust
Po, ... , Ps In fact , if <p (v) E P, t hen there is a unique i s such t hat
<Pi (v) E p. As P has rank 0: , it follows qi P, whence Pi = p. ..
5.8 Strongly minimal sets
As said before, a definable set is strongly minimal if it is infinit e, but admits
no par ti ti on int o 2 disjoint infinit e definabl e subsets in D. A t heory is called
strongly minimal if the domain of it s universe is.
Recall also from Chapter 2 that a struct ur e A is said t o be st rongly minimal
if it s th eory is, and minimal when it s domain A admits no partition into 2
disjoin t A-definable infinit e subsets. Of course , st rongly minimal struct ures
are minimal , but t he conve rse is not t rue (see, again, Example 2.7.2) .
Clearly strongly mini mal sets are t he simplest infinit e definabl e sets and
hence provide a matter wort hy of some int erest. Incident ally, due to t he
new approach t o struct ures provided in 5.2, any strongly minimal set , and,
more generally, any defina ble D in D can be nat ur ally regarded as a structure
in it s own right. It suffices to ass ume as definabl e sets in D t he t races in D
of the definable sets in D, hence D" n X for every definable X Dn. It is
easy t o check that t he res ult ing collection satisfies t he condition in Theorem
5.2.1. Wh en D is strongly minimal , t he resul ting new struct ure is strongly
minimal , too.
Accordingly, one ca n examine strongly mini mal structures and t heor ies in-
stead of st rongly minim al sets. We already met some examples in t his area.
In fact we know that infinit e vectorspaces over a countable fields, as well
as algebraically closed fields (in a fixed characteristic) , and even mere infi-
nite sets, admit a st rongly minimal complete t heory. Let us analyse closerl y
these examples.
5.8. STRONGLY MINIMAL SETS 169
Example s 5.8.1 1. Let K be a countable field and consider the theor y
KT' of infinite vectorspaces over K. Recall t hat each of t hem is fully
determined, up to isomorphism, by a cardinal number (it s dimension
over K). In view of a possible generalization, let us remind very quickly
why. So let V be any vectorspace over K. For a E V , X V define
a -< X (alinearly dependent on X )
if and only if a is in t he subspace (X) of V spanned by X (in oth er
words, a is a finit e linear combination of elements in X U {O}). One
introduces in this way a subset -< of V X P(V) (the linear depen dence
relation). Moreover one sees that -< satisfies th e following properties:
for a, b in V and X , Y V ,
(D1) if a E X , t hen a -< X ;
(D2) if a -< X , t hen t here is some finit e X
o
X such th at a < X
o
;
(D3) if a -< X and, for every x E X , x -< Y, t hen a -< Y;
(D4) if a -< X U {b} but a -f< X , then b -< X U {a} .
Checking (D1) and (D2) is t rivial, while (D3), (D4) just need some
simpIe calculations.
Now define a subset B of V linearly independent if and only if, for
every b E B , b -f< B - {b}; and say th at B V is a basis of V if and
only if B is linearly independ ent and, for every a E V, a -< B .
Using (D1) - (D4) and Zorn 's Lemma (wit h no specific reference t o t he
algebraic fr amework of vectorspaces), one shows t hat V admi t s some
basis B and t hat two bases of V have the same power. Accordingly
one defines the dimension of V over K as the power of any basis of V.
Finally one proves that two vectorspaces over K are isomorphic if and
only if they have the same dimension. This t ime, t he proof needs also
th e following fact : if V is a vectorspace over K , X V , a, b E V
and both a and b are linearly independent of X , t hen t here is an
automor phism of V fixing X pointwise and mapping a into b (in other
words, a and b have t he same ty pe over X ).
2. Let p = 0 or p be prime, consider now t he t heory ACFpof algebraicall y
closed fields of cha racteristic p. Also in this case it is known t hat every
model K of ACF
p
is fully det ermined up t o isomorphism by a cardinal
number: it s transcendence degr ee (over t he prime subfield) . Under
th is point of view, t he setting is qui t e similar to what we saw in t he
prev ious example. In fact, for every algebraically closed field K of
characteristic p, one int roduces a subset -< of J( x P (J( ) (now called
t he alge braic dep endence relation ) in t he following way: for a E J(
and X J( , one puts
a -< X (a algebraically dependent on X)
if and only if a is algebraic over t he subfield generated by X. One
checks that (D1) - (D4) still hold for a, b E J( and X, Y J( (although
t he proofs, especially t hat of (D3) and (D4) , require now some major
algebraic difficulties) .
Any how, as in th e previous example, we can define a subset B J(
algebraically independent if and only if, for every b E B , b -< B - {b};
and we can say t hat B is a transcendence basis of K if B is alge-
br aically independent and, for every a E J( , a -< B. The existe nce of a
t ranscendence basis of K and t he fact t hat t wo transcendence bases of
K have t he same power are shown exactly as in t he previous examples,
despi t e t he different algebraic fr amework , by using onl y (D1) - (D4)
(and Zorn 's Lemma) . The cardinality of a t ra nscendence basis of K
is called t he transcendence degree of K. As in t he previous examples,
one shows t hat an algebraically closed field K of characterist ic p is
fully det ermined up to isomorphism just by it s t ranscendence degr ee
(so by a cardinal number). The reason is, again, t he fact t hat , if a,
b E J( , X J( and a, b -I< X (so both a and b are t ranscendental
over t he subfield generated by X ), t hen t here is an aut omor phism of
K fixing X pointwise and mapping a into b (whence a and b have t he
same ty pe over X ).
3. Let us propose a mor e trivi al example. Let L = 0, M be an infinite
set . For a E M , X M , put
a -<X {:} aEX;
t his again defines a subset -< of M x P(M) clearly satisfying (D1) -
(D4) . In t his case t he only possible basis of M is just M, and act ually
t he power of M fully determines t he isomorphism class of M among
L-structures (i. e. nonempty sets) .
Before cont inuing our analysis of strongly mini mal t heories and sets, in order
to exclude any possible misun derst andings and t o prepar e the next Chapter
5.8. STRONGLY MINIMAL SETS 171
7, let us point out a key remark: we cannot expect t hat, for any complet e
t heory T , every model of T is determined up to isomorphism by a single
cardi nal invari ant. Here is a simple "count erexample" .
Example 5.8.2 Let T be the t heory of th e st ruct ures (A, B) where B A
and bot h B and A - B are infinite. Not ice t hat T is complete, for example
because it sat isfies the assumpt ions of Vaught 's Theorem: it has no finite
models and is as the onl y countable model (up to isomor-
phism) must satisfy IBI = lA- BI = In order to characterize a model
(A, B) of T up to isomorphism, one needs an ordered pair (IBI, lA - BI)
of ca rdinal numbers.
Of course this example can be easily gene ralized. Just think of a structure
A = (A, (Bn)n<N) where 2 ::; N ::; w and t he Bn's are pairwise disjoint
infinite subsets of A. In t his ext ended fra mework the isomorphism t ype of
A is given by the ord ered sequence of cardinals
notice that this sequence is infinite when N = w.
Now let us restrict again our attent ion to st rongly minimal t heories. We
wond er if t he similar behaviour observed in Examples 1, 2 and 3 in 5.8.1 for
infinite K-vectorspaces, algebraically closed fields and infinite sets is found ed
on some common basis. To clarify this point , let us give a furt her glance at
these examples.
Examples 5.8.3 1. Let V be a vectorspace over K , a E V , X V .
Recall that a -< X means that there are n EN, k
o
, ... , k
n
E K ,
Xo , . .. , X
n
E X such that a = Li<n kiXi ; in other words a is the only
element of V sat isfying the L(X)=-formula v = L i<n kiXi .
2. Let K F ACF
p
, a E K , X K. Now a -< X if and onl y if th ere is a
polynomial f(t) E K[t]- {O} with coefficients in th e subfield generat ed
by X such that f(a) = O. There is no loss of generality in ass uming
that t he coefficient s of f(t) are also in the subring generated by X (so
in the L-substructure generated by X). Accordingly every coefficient
of f(t) gets X-definabl e, and "f (v) = 0" can be written as an L(X )-
formul a. Of course there are onl y finitely many elements satisfying
t his formul a, and a is one of them.
3. Let A be an infinite set, a E A, X A. Now a -< X simply means
a EX, hence that a is the unique element in A sat isfying the L(X)-
formula v = a (a is in X!) .
5.9 Algebraic closure and definable closure
What we observed at the end of the previous section suggest s the following
definitions. Let M be a st ruct ure in a language L, X M.
Definition 5.9 .1 1. The algebra ic closur e of X (denoted by acl(X))
is the union of all the finit e X -definable subsets of M (and so is the
set of the elements a E M such that, for some L(X)-formula <p(v),
a E <p(M) and <p(M) is finit e) .
2. The definable cl osure of X (denoted by dcl(X)) is the se t of all the
X -definable elements of M (and hence of the elements of M such that,
for some L(X)-formula <p (v), <p (M ) = {a} ) .
Clearly dcl(X) acl(X) for every X and A1. Sometimes dcl(X) = acl(X).
For instance, this is the case when M expands a linear order :S and X is
any subset of M.
In fact , let a E acl(X) , <p(v) be an L(X)-formula such t hat <p(M) is finit e
and includes a. More pr ecisely, let ao < . . . < a; be t he elements of <p (M),
and let a = a; for a unique i :S r . Then a is t he onl y eleme nt in M satisfying
t he formula
<p (v) A :J! i w (<p(w) A w < v).
However there do exist some st ructures ;\,1 and subsets X M such that
dcl(X) f. acl(X), as we will see in the next lines.
Remarks 5.9.2 1. Let M, N be structures for L such that M is an
elementary substructure of N, X M (hence X N) . Of course, we
might expect to have to form two algebraic closures of X , the former
in M and the latter in N. However these t wo sets coincide . Let us see
why. First notice t hat, for every L(X) -formula <p (v), <p (M ) is finit e if
and only if <p (N ) is , and t hese sets have the same power. In fact , if
1<p(M)1= k, then A1 1= :J!k v (<p( v)) ; as M < N , N 1= :J!k v (<p(v)) ,
whence 1<p(N) I = k. Of course this argument holds even in the other
direction, from N to M . Moreover, using A1 < N once again, one
sees t hat every element of a finit e <p (M) satisfies <p( v) in N as well.
Consequently, for every L(X)-formula <p(v) such t hat <p(M ) is finit e,
5.9. ALGEBRAIC CLOSURE AND DEFINABLE CLOSURE 173
rp(M) = rp(N). In conclusion, the algebraic closure of X is the same
in M and in N.
Clearly the same can be said about the definable closure. Conse-
quently, for a complete theory T with universe n and for a small
X n, ael(X) and del(X) do not depend on t he model of T contain-
ing X one refers to .
2. Let T be a complete theory, X be a small subset of n, a E n. Then
a E ael(X) {:} there are at most finitely many elements f(a) when f
ranges over the automorphisms of n fixing X point wise.
In fact, take a E ael(X) . Let rp(v) be an L(X)-formula such that
a E rp(n) and rp(n) is finite. For every automorphism f of n fixing X
pointwise, f(a) E rp(n) as well. Accordingly the images of a under the
automorphisms of n fixing X pointwise form a finite set .
Conversely, let a ael(X), so, for every L(X)-formula rp(v), if a E
rp(n) , then rp(n) is infinite. Let ao, .. . , an be different elements, all
realizing tp( a/ X) . For every formula rp( v) E tp( a/ X), there is some bE
rp(n), bf. ao, ... , an' As tp( a/X) is closed under finite conjunction,
tp( a/X) U {-,( v = ai) : i ::; n} is consistent and can be enlarged to
a complete type over X U {ao, ... , an}: any element an+! real izing
this type satisfies a
n
+l 1= tp(a/X) , an+! f. ao, . .. , a
n
. So there ar e
infinitely many realizations of tp(a/X) in n, and each of them is the
image of a under some automorphism of n fixing X pointwise.
A similar argument shows
a E del(X) {:} f(a) = a for every automorphism f of n fixing X
pointwise.
Now let us come back to the Examples 5.8.1 of the last section. We want to
examine ael and del in their frameworks.
Examples 5.9.3 1. Let V be a vectorspace over a countable field K,
X V . Then del(X) = ael(X) coincide with t he subspace (X) of V
spanned by X.
In fact, we know that del(X) ael(X) in general , and we have already
seen that (X) del(X) . Hence it suffices to prove ael(x) (X).
Notice that, if a and b are two elements of n, and both of them are
linearly independent of X, then there is some automorphism of nfixing
X pointwise and mapping a into b. So, for a (X) , every element
in n - (X) is the image of a under some aut omor phism of n act ing
identically on X ; accordingly art ael(X).
2. Now consider an algebraically closed field IC (of some given characte r-
istic p). If X J(, then ael(X) coincide wit h the algebraic closure (in
t he usual field theoretic sense) of t he subfield of IC generated by X.
In fact , we already saw t hat :2 holds . Conversely, if a, bEn are
t ra nscendental over the subfield generated by X , then there is some
aut omorphism of n fixing X point wise and mapping a into b. Just as
in 1, this excludes that a, or b, or any eleme nt transcendent al over t he
subfield generated by X can belong to ael(X) .
This t ime, del(X) i- ael(X) . The reason is very simple: in fact , if
a, b E J( are two different roots of the same irr educible polynomial
with coefficient s in the subri ng generated by X , th en one can build an
automorphism of IC fixing X pointwi se and mapping a into b. Hence
a E ael(X) , but a rt del(X).
Indeed one can check t hat, if IC has characteristic 0, th en del(X) is just
t he subfield of IC generated by X , while, if IC has a prime characteristic
p, then del (X) coincides with th e closure of t he subfield of IC generated
by X under the inver se function of t he Frobenius morphism Fr (the
one taking any a E J( to Fr (a) = a
P
) .
3. Finally let M be an infinit e set, X M . Now ael(X) = del(X) = X.
In fact ael(X) :2 del(X) :2 X is clear. So it suffices t o show ael(X)
X. Owing to the Remark 5. 9.2.1, we can ass ume IXI < IMI . Any two
elements of M - X correspond t o each other by a permutation of M
fixing X pointwise. Hence any X -definabl e subset of M overl apping
M -X includes M -X, and so is infinit e. In conclusion , ael(X) X ,
as claimed.
Anyhow, in t hese examples, we recognize a common feature: in fact , for
every small subset X of n and a E n,
a -< X {:} a E ael (X) .
But , before examining closerl y this point, let us propose a couple of further
examples.
Examples 5.9.4 1. Let T = DLO- , X be a small subset of n. As T is
concerned with linear orders, ael(X) = del(X) . We claim ael(X) = X.
5.9. ALGEBRAIC CLOSURE AND DEFINABLE CLOS URE 175
In fact , take a E n - X. Owing to Theorem 2.3.1 , every element
bEn - X such that
b : {:::} a < x
for every x E X sati sfies the same L(X)-formulas as a. Hence, if
a E <p (n ) for some L(X)-formula <p (v), <p (n) contains all these elements
b. By Compactness, there are infinitely many b's; hence <p (n) must be
infini t e. In conclusion arf.acl (X) .
2. Let T = Re F , X be a small subset of n. As before acl (X) = dcl (X).
But now acl (X) equals t he real closure of t he subfield gener ated by X ,
that is t he least real closed field including X (inside n). This can be
equivalently introduced as the ord ered field of th e elements in n alge-
braic over (the subfield generated by) X . So it is obvious t hat acl(X)
includes it . Conversely let a E acl(X) = dcl (X), accordingly a is t he
only element in n satisfying a suit able L(X)-formula <p(v) . Owing
the quantifi er elimination, we ca n ass ume that 0 wher e f( x) and g(x) are nonzero polynomials
with coefficients in the subfield generat ed by X. As a must be the
onl y element sat isfying t hese condi tions, at least one equat ion occurs,
whence a is algebraic over the subfield generated by X , as claimed.
The read er may check th e details of t he last example, and also calculate
acl(X) , dcl(X) in other familiar cases, for inst ance when X is a small subset
of the universe n of dLO+.
Now let us come back to a general framework . Accordingly, let A be a
struct ure for a given language L. For a E A, X A, put
a -< X {:::} a E acl (X).
We wonder whether thi s relation < sat isfies t he conditions (D1) - (D4) in
5.8 (just as in th e t hree main examples before) . (D1) - (D3) st ill hold in
this general set t ing. Let us see why.
(D'l ) For every a E A and X A, if a E X , t hen a -< X.
In fact a is the only element satisfying v = a (and so is even in dcl (X)).
(D2) For every a E A and X A, if a E acl(X) , then t here is a finite
subset X
o
of X such that a E acl(X
o
).
In fact take a E acl (X) ; th ere is some L(X)-formula <p(v) such t hat
<p(A) is finit e and includes a. Let X
o
be t he set of t he paramet ers
from X occurring in <p(v) . Clearly X
o
is what we are looking for.
(D3) For every X A, acl(acl(X)) acl(X) .
Let a E acl(acl(X) ), cp(v, X) be a formula with parameters F .e- (xo, . . . ,
x
m
) in acl(X) such t hat a E cp (A, x) and cp (A, x) is finite. Put
Icp(A, x)1= k. For every i m, there is a for mula t9i(V, Y"i) of L(X)
such that t9
i
(A, Yi) is finite and includes Xi . Put fj = (Yo , . . . , Y;") and
consider the formula
o; v, fj) : :.Jwo ... :.Jw
m
( ./\ t9i(Wi, Yi) !\ :.J !kz cp (z , w) !\ cp (v, 10) )
t<m
where wabbreviates th e tuple (wo, .. . , w
m
) . Clearly a E t9(A, 17) .
Furthermore t9(A,17) is th e union of the sets cp (A , ; ') when X' =
. . . , satisfies E t9i(A, Yi) and Icp(A, Xl ) 1 = k. Accordingly
t9(A, 17) is a finite union of sets of size k, an hence is finite .
Now let us wonder if even (D4) holds in any st ruct ure A. Recall what (D4)
states.
(D4) For every a,b E A and X A, if a E acl(X U {b}) - acl(X), th en
b E acl(XU {a} ).
Examp le 5.9.5 Let A = (N
2
, J) wher e f is the 1-ary function such t hat,
for every X, YEN, f (x, y) = (x , 0). Pu t
a= (0,0) , b= (0,1 ).
Then a is t he only element in N
2
sat isfying v = f(b) , whence a E acl(b).
Moreover a rt acl(0) because any element (x,O) with x E N is the image of
a in some automorphism of N
2
. But b rt acl(a) because any element (0, y)
with yEN- {a} is the image of b = (0,1) under some automorphism fixing
a = (0,0).
Accordingly (D4) fails in general. However the following proposition holds.
Proposition 5. 9.6 Let A be a minimal structure. Then (D4) holds in A:
if a,b E A, X A and a E acl(XU {b}) - acl(X) , then b E acl(XU {a}).
Recall that A is minimal if and onl y if any subset of A definabl e in A is either
finit e or cofinite. Algebraically closed fields, vectorspaces (over coun t abl e
fi elds) and pure infini t e sets are minimal structures. On t he contrar y, t he
struct ure A of t he last example is not minimal: for , t he formul a f (v) = (0, 0)
defines an infinite coinfinite set .
5.9. ALGEBHAIC CLOSUHE AND DEFINABLE CLOSURE 177
Proof. As t he minimality of A is preserved under adding or forgetting pa-
rameters from A, there is no loss of generality in assuming X = 0 (otherwise
replace A by Ax). Suppose towards a contradiction that t here are a and b
in A such that
a E acl(b) - acl(0), b f- acl(a).
As a E acl(b) , t here is an L-formula w) for which b) is finit e and
includes a; let k be the size of b) . Accordingly bsat isfies the formul a
'l9( v, a) : v) !\ 3!k z v).
As b f- acl(a), 'l9(A, a) must be infinit e, hence cofinite. Let m denote the
(finite) cardinality of A - 'l9(A, a); so a sat isfies
As a f- acl(0), also ')' (A) must be infinit e. Pick k + 1 pairwise different
elements aa, . .. , ak in ')' (A). For every i ::; k, t here are exactly m elements
in -,'l9(A, ai) . Exclude all these element s for i ::; k. We can st ill find some
b' E A satisfying
A 1= 1\ 'l9(b' , ai),
i<k
hence
A 1= 1\ b') !\ 3!k Z b').
i< k
Consequently aa, ... , ak E contradicting = k. This ac-
complishes th e proof. '"
Hence, when T is a strongly minimal t heory, -< sat isfies (D1)-(D4) in every
mod el A of T. This allows to generalize the notions already introduced in
the main examples: for B A, one can say that
B is independent if and onl y if, for every b E B, b f- acl(B - {b}),
B is a basis of A if and onl y if B is independent and A = acl(B) .
Using Zorn 's Lemma, one can st ill deduce the existence of a basis of A, and
the fact t hat two bases have the same ca rdinality; in t his way A is naturally
associated with a ca rdinal numbe r, t hat is t he power of any basis. This is
called the dimension of A. Of course , all this directly depends on (D1)-(D4).
However the fact that this ca rdinal det ermines A up to isomorphi sm refers
to another basic property of st rongly minimal theories, ensuri ng that , for
every X A, in particular for X = 0, t here is a unique non- algebraic l -type
px over X : t he one containing every formula <p(v) such that <p( n) is cofinite .
In this perspective a basis B of A is act ually a subset of P0(A), and indeed
any element bin B realizes t he unique non-algebr aic I- t ype over B - {b}.
Consequently, if A , A' F T have the same dimension and B, B
'
are t wo
bases of A , A' respectively, t hen th ere does exist some bijection between B
and B
'
, but (which is more relevant) such a bijection is even an elementary
function . Furthemore, as A = ael(B) and A' = ael(B
')
, it can be extended
wit h a little patience to an isomor phism from A onto A' .
In conclusion, t he models of a st rongly minimal theory can be classified quite
satisfactorily up to isomorphism in t he way we have just described, and the
main ingredient s of this class ificat ion result are:
(a) -< satisfies (Dl)-(D4);
(b) t here is a unique non-algebraic l-type over any small subset X of n.
To underline once again the relevance of (b) , let us mention the behaviour of
a-minimal t heories. Recall that t hey are t he complete theories of t he infini te
st ructures A expanding linear orders in such a way t hat th e only definable
subsets of A are t he finite unions of singletons and open int ervals (possibly
wit h infinit e endpoints) . We will see in the Chapter 7 t hat t here are good
reasons to agree t hat t he models of t hese theories cannot be classifi ed up
to isomorphism. In spit e of t his, in an a-minimal structure A -< can sat isfy
(Dl) - (D4)' hence (a) holds (but (b) does not ) . As (Dl) - (D3) are always
true, we have to check (D4) . Here is th e proof.
Proposition 5.9.7 Let A be an o-minimal structure expanding the linear
order (A, <). a, b E A, X A. If a E ael(X U {b}) - ael(X) , then
b E ael (X U {a}).
Proof. As the o-minimalit y of A is preserved under adding or forgetti ng
paramet ers from A, t here is no loss of generality in ass uming X = 0. Let a,
b E A satisfy a E ael(b) - ael(0) , we have to show b E ael(a) . As ael and
del coincide in linearly ordered st ruct ures, a is in del( b) and so t here is an
L-formula <p (v, w) such that a is t he only element in <p (A, b). We can even
assume that <p(v, w) defines a partial function f such that f( b) = a. Let B
denote t he set of the element s x E A such that f( x) = a. B is {a }-definable
because f is 0-definable. By o-rninimality, B is the union of finitely many
int ervals 1
0
, ... , Ik (where each int erval may be open, closed or semi-closed,
come down to a singleton, or admit infinit e endpoints) . Take k minimal and
ass ume 1
0
< .. . < I i : Not ice t hat all t he end points of t hese int er vals in A
5.9. ALGEBRAIC CLOSURE AND DEFINABLE CLOSURE 179
are {a}-definabl e. Consequently, if b equals one of them, t hen b E acl(a).
So suppose t hat b is in the int erior of some interval among 1
o
, . .. , Ik. Let
I denote this int erval , d
1
< d
2
be its end points (where d
1
, d
2
may possibly
equal - 00, +00 respectively). If I is finite, then it is easy to deduce again
s acl(a). Accordingly assume I infinit e. Notice that d
1
cannot equal - 00,
otherwise a is t he only element in A satisfying :3wYz(z ::; w -+ f (z) = v),
whence a E acl(0). Similarly, d
2
< +00.
For a suitably large positive int eger n, let D
n
denote t he (0-defina ble) set
of t he left endpoint s d E A of some open int erval J having size nand
satisfying t he following furt her ass umptions: f is (defined and) constant in
J , and one cannot enlarge J to t he left keeping this condit ion. Associate
any d E D
n
with the right endpoint g(d) E AU {+oo} of a maximal int erval
J. By o-minimalit y, g(d) is well defined: in fact, t he elements> d lying
in the domai n of f and having the same value in f as immediat ely after d
form a defina ble set, and so a finite union of intervals (in t he broader sense
said before) . Moreover , for d < d' in D
n
, ]d, g(d)[n]d', g(d' )[= 0. Recall
t hat D
n
is definable, so, by o-minimality, D
n
must be finite . As d
1
E D
n
,
d
1
E acl(0) = dcl(0). As a is t he image of any element in ]d
1
, d
2
[ by J, a
'is in acl(0) as well, and this contradicts our hypot heses. So I cannot be
infinit e, and we are done. .,
Before concluding th is section and the whole chapter , we would like to un-
derline two more point s.
The former specifically concerns a minimal, or o-rninimal L-struct ure M.
Both t hese prop erti es (minimality and o-minimalit y) are preserved under
adding parameters from a subset X of M, so under passing from L to L(X)
and from M t o M x: M x remains mini mal , or o-minimal. Indeed , if the
t heory of M is st rong ly minimal, t hen Th(Mx) is: in fact it s models are t he
st ruct ures Nf (x ) where N is an L-st ruct ure elementarily equivalent to M
-so a mini mal st ruct ure-, and f is an elementary function from X into N .
The same happens for an o-minimal Th(M ). So, for M as before, we can
int roduce for every X M a new dependence relation <x in M X P (M )
by put t ing, for a E Me S M, a <x S if and only if a E acl(S) in M x , in
other words if and only if a E acl(X US) in M. Moreover <x st ill satisfies
(D1) -(D4) .
Secondly, when our relation -< on M X P(M) satisfies (D1) - (D4) , we can
equip not only M, but also every subsets S of M wit h a dimension dimS:
this can be formally int roduced as the minimal power of a subset B of S such
t hat S acl(B) (in fact , it is easy to check that such a B is independent ).
Of course, t his also concerns the framework sket ched a few lines ago: for
x M, the dimension of X over X dim(Sj X) is just the dimension of S
with respect to <x
A final remark: in order to pursue and possibly accomplish our analysis of
strong minimality, we can wonder how general the examples we proposed
(pure sets, vectorspaces and algebraically closed fields) are. Do there exist
any" new" significant instances, or do they exhaust all the possible cases?
In Chapter 7 we will discuss the relevance of this question (due to Zilber)
and we will provide its solution (due to I-Irushovski).
5.10 References
The papers where Morley developed his ideas and, in particular, introduced
his rank are [116, 117, 119, 120] and [118]. The duality between Boolean
algebras and Boolean spaces is in [52]. The proof of Theorem 5.4.2 in the
general setting, for any cardinal A can be found in Poizat's book [131],
which also includes full details on saturated models. Inaccessible cardinals
are described in [66] or [78] .
Now let us deal with modules: both [181] and [136] treat the matters of
Section 5.5 (pure injectivity, Ziegler topology, and so on). Tame and wild
classes of modules are also discussed in [136]. The classification of pure
injective objects among abelian groups, and more generally among modules
over a Dedekind domain, is given in [40].
An abstract treatment of dependence relations, including the main examples
in Section 5.8, was developed in [173]. The details on the real closure of an
ordered field can be found in [65].
An axiomatic introduction to ranks, and a comparison of Morley rank with
other rank notions, are in [8]: see also [131].
Chapter 6
w -stability
6.1 Totally transcendental theories
We cont inue her e our t reat ment of Morley's Theor em and related ideas. All
t hroughout t his section, T is a complete t heory wit h no finit e models in a
countable language L , and S1 denot es a big sat urated model of T . In t he last
cha pter we defined Morley rank and Morley degree as a complexity measu re
for definabl e sets. In parti cul ar we st udied t he simplest infinit e definable sets
wit h res pect t o t his measure , i. e. t he strongly minimal sets, t hose having
Morley rank 1 and Morley degree 1; we considered also st rongly minimal
theori es, i. e. t he complete t heories whose universe is strongly minimal.
More generally, one can examine t he t heories in which every non empt y
definable set gets a(n ordinal) Morley rank. These were called by Morley
tot ally t ranscendent al.
Defi ni tion 6.1.1 T is called totally transcendental if and only if, for
every non empty definable set X S1n (wit h n a positive integer) , RM (X )
is an ordinal number.
Hence totally transcendental theories exclude t he th eor y DLO- of dense
linear orderings without endpoint s, as observed in Example 5.7.3,4, but
include, for inst ance, t he strongly mini mal t heories, which sat isfy RM(S1) =
GM(S1) = 1; indeed, as we will see within a few lines, strong minimality
impli es RM (S1n) = n, GM (S1n) = 1 for ever y n, and so RM(X) :S n for
every definable set X S1n.
Tot ally transcendental theori es can be characterized in t he following alt er-
nati ve and equivalent way.
181
182 CHAPTER 6. w -STABILITY
Definition 6.1.2 T is cc-stable if and only if, for every countable A D,
the space 5(A) of types over A is also countable.
One can equivalently ask that , for every countable A D and positive
integer n, 5
n(A)
is countable. Actually it suffices to require 51(A) countable.
In fact let n be an integer bigger than 1, pE 5
n(A)
, if = (aI , . . . , , an) realize
p in Dn. Then p is fully determined by:
(i) the (n - l.j-type of (a1, "" an-I) over A (equivalently, the orbit of
(a1,"" an-d with respect t o t he A-automorphisms of D);
(ii) the l-type of an over A U {aI , . .. ,an-d (hence over a countable set )
up to A-automorphism.
So, when 5 1(A) is countabl e, a simple induct ion proves t hat 5
n(A)
is count-
abl e for every posit ive integer n.
Observe also t hat T is w-stable if and only if, for every countable model M
of T , 51(M) is countable. In fact , for every countable A D, t here exists
some countable model M of T including A (just apply the Lowenheim-
Skolem Theore m t o the t heory of DA). Every l-type over A extends t o
some (possibl y non unique) I-type over M, and conversely restricting to
L( A) a l-type over M determines a l-type over A. So 15
1(A)1
::; 15
1(M)I
.
Now let us point out:
Proposition 6.1.3 A strongly minimal theory T is eo-stable.
Proof. Let A be a count able subset of D. The language L( A) is countable,
as well as the set of the L(A)-formulas <p (v) such that <p(D) is finit e. Ac-
cordingly acl(A) is countable, and the algebraic I-types over A also form a
countable set. As there is only one non- algebraic I-type over A, 51(A) is
count able, too. ..
Example 6.1.4 DLO- is not w-st abl e. In fact (Q,::;) is count able, but
it is easy t o realize t hat t here are I-types over Q: just observe that ,
for every r, s E R with r < s , there is some a E Q such that r < a < s.
Accordingly tp(r/Q) f tp(s/Q) as the former t ype contains v < a, and the
lat t er does not.
As said before, w-stability is j ust equivalent to total transcendency. Indeed
the following theorem holds.
Theorem 6.1.5 T is totally tran scendental if and only if it is eo-stable.
6.1. TOTALLY TRANSCENDENTAL THEORIES 183
Proof. Let T be t otally t ranscendent al. Then every ty pe p over a small set
A n is fully determined by an L (A)-formula <p(il) in p having t he same
Morley rank and degree as p. Then t here are at most as many ty pes over
A as L (A)-formulas, and in conclusion at most IAI + ty pes over A. In
par ticul ar , when A is countable, S(A) is coun t abl e, t oo.
Now ass ume T w-st able. First we claim t hat there exists an ordi nal aT such
th at , for every positive int eger n and definable set X nn, if RM(X) aT,
th en R M (X) = 00.
In fact , let <p (ii. w) be a formul a of L, ii, bbe two sequences in n havin g t he
same length as wand t he same t ype over 0; so t here is an aut omorphism of n
mapping ii into b, hence <p (nn, ii) onto <p(nn, b) . It follows t hat <p (nn, ii) and
<p(nn, b) have the same Morley rank. On the other side, S (0) is countable,
as well as t he set of all possible formulas <p( il, w). So there are at most
count ably many Morley rank s =1= 00 of definable sets in n, and we can take
a T as t he least ordinal great er than all these values. By t he way, one ca n
even show that aT is coun t abl e.
Now suppose towards a cont radict ion th at T is not totally t ra nscendental.
T hen t here exists a definabl e set X = X
0
having Morley rank 00, namely
aT. Hence RM(X
0)
aT + 1, and so t here exist two disjoint definable
subsets X
o
and Xl of X 0 both having Morley rank a T and consequent ly
00. Repeating t his procedure one builds, for every finit e ordered sequence
a of 0 and 1, hence for every element in {O, 1} <w, a definabl e subset X
u
of
X
0
of Morley rank 00 such that , for every a E {O, 1} <w and i E {O, I} ,
The set A of the paramet ers needed t o define th e X a 's when a ranges over
{O, 1} <w is countable. For every a E {O, I}W, th e L (A)-formulas " v belongs
to X
u1n
" (with n a nat ural number) form a consiste nt t ype, which can be
extended to a (complet e) t ype Pu E St(A) . Of course different sequences
a, a
'
E {O, 1} <w yield different t ypes Pu =1= Accordingly ther e are at least
2
No
I-types over A, and this cont radict s t he fact that T is w-stable. ..
A consequence of th is theorem is that t ot al transcendence (in fact w-stability)
is preserved by int erpret ability.
Corollary 6.1.6 Let A be a struc ture for L, A' be a structure for a possibly
different language L
'
such that A , A' are infi nite and A' is interpretable in
A. If Th(A) is w-stable, then T h( A
')
is.
Proof. (Sketch) We know t hat , for a suitable choice of a positi ve int eger
n , a subset S of An definable in A and an equivalence relation E in An
definable in A, A' is just t he quotient set SI E and t he whole L' -structure
A' can be definabl y recover ed by A. Furthermore, if D is a big saturat ed
elementary extension of A , t hen t he machiner y defining A' inside A singles
in D a big sat urated elementary extension D' of A' (of t he same power as D)
and every set X' definabl e in D' is t he quot ient set of some X definable in
D wit h respect to E. At t his point a simple arg ument of ord inal induct ion
shows t hat t he Morley rank of X ' in Th(A' ) is smaller or equal t o the Morley
rank of X in Th(A) . Consequent ly, when RM(X ) is an ordinal , RM(X' ) is
an ord inal, too. In conclusion, if Th(A) is w-stable, th en Th(A' ) is. .,
Now let us deal again with Morley rank and degree for a st rongly minimal
theory T. We already seen t hat
RM(D) = 1, GM(D) = 1.
Recall t hat t here is a uni que l -f.ype of Morley rank 1 over any small subset
X of D, t he one of t he elements out of acl(X). Furth ermore, for a, S in D,
a <x S <=? a E acl(X U S)
determines a dependence relation satisfying (Dl) - (D4) .
More gene ra lly, one can see t hat , for every positi ve integer nand n-t uple
ii in D, RM(tp(ii]X)) equals t he dimension of ii wit h respect t o <x - In
par ti cular RM(tp(iil X)) cannot exceed n; indeed t here is a unique n-type
over X of Morley rank n , t he one of an n-tuple ii --<x-independent.
Furt hermore
as already said at t he beginning of t his sect ion. The proofs of t hese fact s re-
quire patience rat her t han ingenuousness, and can be found in t he references
quoted at the end of t he chapter.
Not ice t hat , in t he particular case of algebraically closed fields, the Morley
rank of a t uple d over a subfield F just equals t he t rascendence degree of
F(ii) over F.
6.2 w-st able groups
A paradi gmatical example of w-stable t heory is t he t heory ACF
p
of alge-
br aically closed fields of a fixed characteristic p = 0 or pri me. But conse-
quentl y every struct ure defina ble, or interpretable in an algebraically closed
field IC has an w-stable t heory. This is t he case of t he groups GL(n, IC) and,
6.2. w-STABLE GROUPS 185
mor e generally, of th e algebraic groups over K, as we will see in Chapter 8.
A common feature of t hese examples is t he presence of a binary operation
(th e addition in K, t he usual row-by-column multiplication in GL(n, K))
makin g t hem a group. Accordingly we propose t he following definitions.
Defi nition 6.2.1 An w-st a ble structur e is a structure having an w-stable
theory; an w-s table group is an w-stable structure with a binary operat ion
making it a group (and possibly other operations and relati ons).
Then any algebraically closed field K is an w-stable group (wit h respect to
the sum operation) , as well as any group GL(n, K) and, more generally, any
algebraic group over K (as we will see in Chapter 8). Other examples can
be found among abeli an groups; for instance, we know that the theory of
non-z ero vectorspaces over Q is strongly minimal and consequent ly w-stable.
Accordingly every non-zero Q-vect orspace is an w-stabl e group (with respect
to addit ion), and so every divi sible torsionfree abelian group A is an w-stable
group (for, A can be naturally equipped with a st ruct ure of Q- vect orspace).
The aim of t his section is to begin t he analysis of w-stabl e groups. In fact,
part of it s results will be accomplished, and hence fully understood and
ap preciated, onl y in t he next Chapters 7 and 8, where mor e powerful t ools
will help our st udy. However t he present pages will develop several basic
noti ons in t his area. Most of t hem clearly owe t o t he t heory of algebraic
groups, and par t of t hem refer to t he study of finit e groups. Of course, we
are also int erest ed in classifying w-stable groups and in understanding t heir
connections with algebraic groups, or finit e groups, or abelian groups; but
we will examine in a closer and deeper way th ese t hemes in Chapte r 8.
First let us observe what follows. Let 9 be an w-st able group, H be a
definabl e subgroup, a E G. Notice t hat, if X is a subset of H definable in
g, t hen
aX = {ag : g E X} is a definable subset of th e coset aH
and
X a = {ga : g E X} is a definable subset of Ha.
It follows that , for any a, both aH and Ha have the same Morley rank and
degree as H. Consequently, if H K are definable subgroups of 9 and
Il i- K , t hen
RM (Il ) R M (K) and, if = holds, GM( Il) < GM (K)
(indeed K is a union of cosets of H ). Now we show a chain condit ion for
definabl e subgroups of w-stable groups.
Theorem 6.2.2 Let 9 be an eo-stable group. For every natural n, let H
n
be
a definable subgroup of 9 such that, for every n, H
n
+
1
is a subgroup of H
n
.
Th en there is natural n such that, for every m n in N, H
m
= H
n
.
Proof . Otherwi se, for every natural n , there exists mEN such t hat
m > n and n.; #- n.. Hence RM(H
m
) <RM(H
n
) and, when R M (H
m
) =
RM(H
n
), GM(H
m
) < GM(H
n
) . In part icular , fixed n , there is some s E N
tale che s > n e RM(H
s
) < RM(H
n
) . But in this way one forms a st rictl y
decreasi ng infinite sequence of ordinals, and this is a cont radiction. '"
Corollary 6.2.3 Let 9 be an eo-stable group and, for every natural n, let
K; be a definabl e subgroup of g. Then n nEN K; is definable.
Proof. For every nat ural n, put H
n
= n i <n Ki . Then H'; is a definable
subgroup of 9 and H
n
+
1
H
n
for every n. Consequently t here exists mEN
such t hat, for every i :s: m, Hi = H
m
So n nEN Kn = n nEN Hn = Hm =
n n<mtc; and n n<m; is definable. '"
- -
Corollary 6.2.4 Let 9 be an ea -st able group, f be a defi nable group homo-
morphism from 9 int o 9 having a fin ite kernel. Th en the image f(G) of G
has a fin it e index in g.
Proof. Otherwise one can build a st rict ly decreasing infinit e sequence of
subgr oups
G f( G) f 2(G) .. . r(G) ...
contradict ing Theorem 6.2.2. In fact , as f is definable, r (G) is a definable
subgroup of 9 for ever y natural n. Fur th ermor e, if f(G) has an infinit e
index in g, t hen, for every n EN, r:' (G) has an infinite index in r (G),
and so rH(G) #- r(G). Let us see why. The case n = 0 is just our
assumption. So t ake n > 0, and suppose that the index of r (G) in r:' (G)
is infinite. Notice t hat, for a and b in r(G) , f (a) - f(b) E r(G) if and
only if a - b E r:' (G) + K er f . As K er f is finit e and r(G) has an infinit e
index in r:' (G), t here ar e infinit ely many elements in r+
1
(G) pairwise
inequivalent modulo r(G) , and so r +
1
(G) has infinite index in r (G).
'"
Now take any st ruct ure 9 expanding some given group (G, .) (so 9 is not
necessaril y an w-stable group). As said before, if X is a definabl e subset of
di G, t hen also
aX={ag :gEX}, Xa={ga :gEX}
are for every a E G, as well as X -
i
= {g- i : g E X}. Notice also t hat, for
every ty pe p over G and a E G, we can form the sets ap, pa of the formulas
cp(v) of L(G) such that cp(av) E p, cp(va) E p respectively. It is easy to see
that ap, pa are types over G. More precisely, if x denotes any realization of
p, then ap is t he ty pe of a-ix over G and pa is the type of xa-
i
over G.
The funct ions from G X S (G) to S (G) mapping any ordered pai r (a, p) of
G X S(G) into ap and pa define two actions (a left action and a right action
respecti vely) of G on S(G). Correspondingly we ca n consider, for every
pES(G), t he left stabilizer {a E G : ap = p} of p in G and t he right
stabilizer {a E G : pa = p} of p in G. It is well known t hat bot h are
subgroups of g.
In a similar way, for every type p over G, we ca n form t he set p-i of t he
formul as cp(v) in L(G) such t hat cp (v-
i
) E p. Even p- i is a t ype over G;
indeed, if x is any realization of p, t hen p-i is just t he type of x-lover G.
Now let 9 be an w-stable gro up. As w-stability is preser ved by =, t here is
no loss of generality for our purposes in replacing 9 by a suitably saturated
elementary extension, and so in ass uming t hat 9 itself is saturated in some
uncountable power (and so is the universe of it s complete theory).
Let X G be definable. As observed before X has the same Morley rank
and degree as aX and X a for every a E G, and also as X-i . Consequently,
for every type p E S(G) , RM(p) = RM(ap) = RM(pa) per ogni a E G and
RM(p) = RM(p-i) . Moreover
Lemma 6.2.5 Let 9 be an w-stable group, p be a type over G. Then both
the left and the right stabilizers of p in G are definable subgroups of g.
Proof. We treat t he left case (t he right one can be han dled in a similar
way). For every L-formul a cp( v, ill) let G(p, cp) denote the set of the elements
a E G such t hat, for every 9 and h in G,
cp (hv,9) E p {:} cp (hv, 9) E ap.
It is easy to check that Gip, cp ) is a subgroup of g. In fact , let a, bE G(p, cp);
t hen, for every h and 9 in G,
cp(hv, if) E P {:} cp( hv, if) E ap {:} cp(hav, if) E P {:} cp(hav, if) E bp {:}
{:} cp(habv, if) E P {:} cp( hv, if) E abp;
whence ab E G(p, cp). Furthermore
whence a-I E G(p, <p), too.
Now let us momentainly ass ume what we will actually show only in Chapter
7, Theorem 7.5.5: due to w-stabiIity, G(p, <p) is definable in 9 because th er e
are some formulas with parameters in G defining the sets of th e tuples (h, )
in G for which
<p (hv, 9J E p, or <p(hv, 9J E ap (t hat is <p(hav, 9J E p).
Clearly the left stabilizer of p is th e intersection of the subgroups G(p, <p)
when <p (v, 'Ill) ranges over L-formulas. Hence, by Corollary 6.2.3, the left
stabilizer is definable. '"
Let us exte nd again out interest to arbit rary expansions 9 of groups (G, .)
(so we ar e moment arily forgetting the w-stable ass umpt ion) . A definable
subgroup H of 9 is called connect ed (in 9) if and only if H has no proper
subgroup definable (in g) and of finit e index. Clearly 9 has at most one
definable connect ed subgroup of finit e index (if Ho and HI satisfy t hese con-
dit ions, t hen Ho n HI must equal both Ho and HI , so Ho and HI coincide) .
However 9 may lack such a subgroup. For instance, the additive group of
integers Z has a definable subgroup nZ of finit e index n for every natural
n =1= O. But a simple application of Theorem 6.2.2 ensures th at every w-stable
9 has a connected definable subgroup of finit e index.
Theorem 6 .2.6 Let 9 be an to-stable group. Then there exists a uniqu e
subgroup GO of 9 which is definabl e, connect ed and of fi nite index in g.
Every definable subgroup of fin ite index of 9 includes GO.
Proof. Suppose that 9 has no definable connected subgroup of finite index.
Accordingly every definable subgroup H of 9 of finite index has in its turn
a proper subgroup definable in 9 and of finite index (in H and consequently
in G) . Th en one can build a strictly decreasing infinite sequence of definable
subgroups of 9
each of finit e index in it s predecessor , and so in g. But t his cont radicts
Theorem 6.2.2. This shows the existe nce of a subgroup as required. It s
uniqueness was already observed. Of course t his unique subgroup does de-
serve a specific symbol (Go , for instance) . Fin ally notice that , if H is a
definable su bgroup of finite index and H 1. GO, then If nGO contradicts the
fact that GO is connected. ..
GO is called the connected component of 9 . GO is a normal subgroup of 9.
In fact both the right and the left multiplication by a given element a E G
are definable. Hence, for every a E G, a-1GOa is a definable subgroup of 9
of the same index as GO, and consequently equals GO. But we can say even
more.
Proposition 6.2.7 Let 9 be an w-stable group. The connect ed component
GO of 9 is an 0-definable subgroup; in particular it is in variant under any
automorphism of 9.
Proof. GO is definable, and so th ere are a formula <p(v, w) in the language
of 9 and a t uple if in G such that GO = <p(9, if) . Let k denote t he index of
GO in 9 . The formula
<p( v) : :3w("<p (9 , w) is a subgroup of9 of index k" 1\ <p (v, w))
defines a subgroup <p0(9) of index k of 9. Then <p(9 ) ;;2 GO. But GO and
<p0(9) have the same index in 9 , whence <p0(9) = GO.
The connected component greatly clarifies the role of Morley degree of an
w-stable group 9 . In fact t he following result holds.
Theorem 6.2.8 Let 9 be an eo-stable group, and let GO be its connect ed
component. Th en GO has the same Morley rank as G, and Morley degree 1.
The former claim is almost trivial. Let k still denote t he index of GO in G.
Of course G is the union of th e k cosets of GO in G. Each coset is of th e
form aGo for some a E G, and so has t he same Morley rank as GO. Hence
RM(G) = RM(GO).
Wit h respect to the latter claim, again it is easy to observe that , if 9 is an
w-stable gro up of Morley degree 1, then 9 has to be connected. But showing
t hat in general, for any w-stable G, GO has Morley degree 1 is not so simple.
Indeed we have to prove t hat GO does not decompose as the union of two
disjoint subsets of the same Morl ey rank, and so that there is a uniqu e 1-
type p over G containing the formula <po(v) defining GO and having the same
Morl ey rank as G. This needs some further mor e sophist icated ideas and
tools (to be introduced in Chapter 7) . So we postpone th e full details of
Theorem 6.2.8 to Theor em 7.5.10. Here we limit ourselves to a preliminary
result , saying t hat , if p is a type over G as required before (so cont aining
t he formula <po (v) defining GO and having the same Morley ran k as G), t hen
GO equals bot h t he left and the right st abilizers of p.
Le m ma 6.2.9 Let 9 be an cc-stable group, GO be it s connected component,
p be any t ype containing the formula vP( v) defining GOand having the same
Morley rank as G. Then GO equals the left and the right stabilize r of p.
Proof. It suffices t o handle t he left case . Let S denot e t he left st abilizer
of p. Then S is a definable subgroup of g. Moreover, for every a in G , ap
has t he same Morley rank as p ; as t here are only finit ely many ty pes over
G having t he same Morley rank as G, there are onl y finit ely many different
types ap when a ranges over G . On t he other side, for a and bin G , ap = bp
holds if and onl y if a and b are in th e same left coset of S in G. So S
has finit e index in G , whence S ;2 GO. Conversely take a E S. For every
realizat ion x of p in n, ax sat isfies p as well. In particul ar both x and ax lie
in (l (n); hence a = ax x-
l
sat isfies <p (v ) . As a E G , a E <p (G ) = GO. ..
To summarize, if 9 is an w-st able group (of Morl ey rank a), GO is it s con-
nect ed component and 1 = aD, al , . . . , ak-l are a set of representatives of th e
cosets of GOin G , t hen G = Ui<kaiG
o
where each aiG
o
has Morley rank a
and Morley degree 1. So G has Morley degree k = [G : GO], t here are exactly
k I -types over G of Morley rank a, and each of t hem is isolat ed (among t he
I-typ es of Morley rank a) by a formula v E aiG
o
(more precisely <pO(a; lv))
for some i < k. These ty pes are usually called t he generic types of g.
More gene rally, when 9 is an w-stable group and A is any small subset of
t he uni verse n of t he t heory of g,
a ty pe p E S(A) is said t o be generic if and onl y if it has t he same
Morley rank as G;
an element x of n is said t o be generic over A if and onl y if it s t ype
over A is.
Clearly there are only finit ely many generic t ypes over any small A <;;;; n.
Moreover, if x is generic over A, th en also x-
l
, ax and x a (for a in A) are.
It is worthy underlining th at every g E G can be written (in n) as t he
product of t wo elements generic over G: in fact , if x E n is generic over G,
t hen also gx-
l
is, and g = (gx -l) x .
Let us conclude th is section by proposing anot her fund amental tool in t he
analysis of w-st able groups, in particular of w-stable groups of finit e Morley
rank (i. e. of Morley rank < w): t his is t he so called Zilber's Indecompos-
a bility Th eorem. We need t he following definition.
Defi nition 6.2.10 Let 9 be a structure expanding a group (G, .), X be a
(non empt y) definable subset of G . X is sai d to be (left) indecomposable
if and only if, for every definable subgroup H di g, either X is included in
a unique left cosel of H in G, or X overlaps infinitely many such cosets.
Right indecomposability is defined in a specular way. It is easy to see that,
for every definable X , X is left indecomposable if and only if X-I is right
indecomposable. In particular, when X = X-I , X is left indecomposable if
and only if it is right indecomposable.
Now we can state and prove Zilber's Indecomposability Theorem.
Theorem 6.2.11 (Zilber) Let 9 be an w-stable qroup of finit e Morley rank
(where finite means < w). For every i in a set J of indexes, let Xi denote
a definable left (right) indecomposable subset of G containing the identity
element 1 of g. Then the subgroup of 9 generated by UiEI Xi is definable and
connected in g. Moreover only finitely many Xi's are sufficient to generate
it.
Proo]. Assume for simplicity Xi left indecomposable for every i: the right
case can be handled in a similar way. As G has a finite Morley rank, we
can choose a subset B of G such that B is a finite product of sets Xi or
X
i
-
1
with i E J (and hence is definable) and B has a maximal Morley rank.
Accordingly, for every i E J, RM(XiB) = R1W.(X
i-
1
B) = RM(B); in fact
1 E Xi n X
i-
1
and hence B XiB, B < X
i-
1
B. Let p be a type over G
containing the formula "v E B" defining B and having the same Morley
rank as B. Moreover let S denote the (left) stabilizer of p. Recall that S is
definable (Lemma 6.2.5). We claim that S is just the subgroup generated by
the Xi'S when i ranges over J, and that only finitely many Xi's are sufficient
to generate S.
First notice that, for every i E J, Xi S. It suffices to show that Xi
intersects at most finitely many left cosets of S in G; in fact, in this case,
as Xi is left indecomposable, Xi is wholly included in a single left coset of
S in G, which must coincide with S as 1 E Xi. Accordingly take a E Xi;
the formula "v E X
i-
1
B" defining X
i-
1
B is in ap. We know that RM(ap) =
RM(p) = RM(B) and that "v E X
i-
1
B" occurs in at most finitely many
types having its rank. Hence there are only finitely many pairwise different
types of the form ap with a E X, otherwise RM(X
i-
1B)
> RM(B), which
contradicts what we have observed before. Consequently there are only
finitely many left cosets of S in G of the form as with a E X (recall as =
bS {:} b-
1a
E S {:} bp = ap for every a and b), as claimed. So S is a subgroup
containing each Xi.
Now we show S BB-I. Let a E S, x FP. SO even ax realizes p, and both
x and ax satisfy the formula "v E B" . Consequently a = axx-
1
satisfies
"v E BB-I " in n and hence in g. In par t icul ar S is gener at ed by t he Xi's
(act ually, by a finit e subfa mily of them).
At t his point it remains t o show t hat S is connected in g. So take a subgroup
H of S definable in 9 and having a finit e ind ex in S . Accordingly, for every
i E I , X i overl aps onl y finit ely many left cosets of H , whence X i H. Bu t
t his for ces 1I = S. ..
6.3 w-st a ble fields
The aim of this section is t wofold. First we want to apply what we have seen
about w-stable th eori es and, more particularly, w-stable groups to prove a
beautiful t heorem du e to Macintyre and characteri zing the fields having an
w-stable theory. We already know t hat they include the algebraically closed
fields ; but now we will show t hat no further examples arise, so t he w-st abl e
complete theories of (pure) fields ar e just the t heories ACFp where p is 0
or a prime. Secondly, we will provide a new proof of t he fact t hat alge-
br ai cally closed fields eliminate the imaginaries; t his alternative approach
mainly refers t o t heir w-stability and, owing t o t his feature, applies t o other
w-stable settings, including differenti ally closed fields of cha racteristic 0 (as
we will see lat er in t his chapter) .
Now let us state Macintyre's T heorem.
Theorem 6.3.1 (Maci ntyre) Let I( be an infinit e in tegral domain with iden-
t ity 1, and let I( have an w-stable theory. Then I( is an algebraicall y closed
fi eld.
Proof. First let us see t hat I( is a field. Take a E J( , a i= O. Let C denote
prop er inclusion. If C J( , t hen a
n
+
1
J( C anJ( for every positive integer
n because I( is a domain and so, for b E J( - o.K, anb E anJ( - a
n
+
l
J(. So
I( is an w-stable group (with respect to the sum oper ation) with a st rict ly
decreasing infinite sequence of definabl e subgroups J( J aJ( J a
2
J( J ... ,
whi ch cont radicts Theorem 6.2.2. Accordingly o.K = K , whence there is
some c E J( such t hat ac = 1. So I( is a field. Now suppose towards
a contradiction that I( is not algebraically closed. Hence I( has a Galoi s
extension F of finit e degr ee > 1. Consequently t here is some intermedi at e
field E extending I( and included in F such t hat t he Galois group of F over
E is (cyclic) of pri me or der q. On t he other side E is an extension of finit e
degree of 1( , and hence is defi nable in 1(: in fact , let d denote t he dimension
of L as a vectorspace over 1( , t hen E ca n be viewed as I( d equipped with
6.3. w-STABLE FIELDS 193
suitably defined field operations. Hence E has an w-stable t heory, and an
extension F with a (cyclic) Galois group of pri me order q. Wi th no loss
of generality replace K by and hence ass ume that K it self has a Galoi s
extension F of prime degree q. Field t heory tells us t hat, in t his setting,
F = K(a) where t he minimum polynomial a over K is eit her
x
q
- a wit h a E K , q =1= ca r K
or
x
q
- x - a with a E K , q = car K.
To get a contradiction, we show that every polynomial of this form must
be reducible. For this purpose, first recall th at K is an w-stabl e group with
respect to addition +. Moreover J( coincid es with it s connected component
(with respect to +). In fact , take a E J( and look at t he mul tiplication by
a. This gives an aut omor phism of t he addit ive group of K , and indeed an
aut omorphism definabl e in K. Consequent ly t he multiplication by a fi xes
t he connected component . So a J( = J(o for every a E J( , in oth er words
aJ(o is an ideal of K. Hence J(o = J( as J( is infinite and so x cannot
equal {a} .
Now consider J(* = J( - { a}; J(* is an w-stable gro up wit h respect to
mul tipli cati on. As J( is infinite, J(* has t he same Morley rank and degr ee as
K , in parti cul ar J(* has degree 1, and consequent ly J(* equals it s connect ed
component . Notice t hat t he function of J(* int o J(* mapping any element
k E J(* into k
q
is a definabl e group homomorphism having a finite kernel
{a E J( : a
q
= I} . Owing to Corollary 6.2.4, t he image of th is function
has a finit e index in 1<* and consequent ly equals J(* as J(* is connected.
Moreover a
q
= a and so k f---t k
q
defines a surjective function from J( ont o
J( . In particul ar, for every a E K th e polynomial x
q
- a is reducible.
In t he same way, when q = ca r K, th e function h of J( int o J( mapping any
element k E J( int o k
q
- k is a definable endomorphism of t he additive group
K and has a finit e kernel {a E J( : a
q
= a} . Accordingly its image has finit e
index in J( , and so coincides with J( because J( is connected. Then h is
surjective and, for every a E J( , even t he polynomi al x
q
- x - a is reducible.
This yields t he required cont radict ion. '"
It is worth underlini ng once again t hat Macinty re's Theore m deals not onl y
wit h infinit e fields, but more gene rally wit h int egral domains wit h identi t y;
moreover, as algebraically closed fields are st rongly minimal, it classifies
even t he strongly mini mal examples in t his framewor k.
Now let us deal with the second matter of this section: elimination of imag-
inaries.
Theorem 6.3.2 Every algebraically closed field I( eliminates uniformly the
. . .
tmaquuiries.
Proof. Our new approach needs the following non trivial algebraic pre-
liminaries, regarding arbitrary fields I( and ideals I in I([X] (where x still
abbreviates (Xl, ... , x
n
) ) : their details can be found in the references quoted
at the end of the chapter.
1. The first concerns the fields of definition of I: these are subfields of I(
containing the coefficients of the polynomials of some generating subset
of I. One can see that there is a (countable) field of definition I( (1) of I
satisfying the following additional condition: for every automorphism
a of 1(,
a (more exactly, its natural extension to I([X]) fixes I setwise
if and only if
a acts identically on K (1).
2. Incidentally one can notice that, for every automorphism a of I , if
a(1) 2 I then a(1) = I.
3. Now consider prime ideals in I([X]. More particularly, take pairwise
different prime ideals 10 , ... , I
m
of I([X] in the same conjugacy orbit:
so, for every j < h S; m, there is some automorphism of I( taking
Ij to h. Correspondingly one can find a subfield 1(0 of I( such that,
for every automorphism a of 1(, a permutes 10, ... , I
m
if and only
if a fixes [(0 pointwise. It suffices to form I = nj<m Ij and choose
as 1(0 the field of definition 1((1) associated with I as in 1. The key
point here is that an automorphism a fixes I setwise if and only if it
permutes the Ij's. This is a consequence of two facts: the former is
that, owing to 2, I = nj<mIj is an irredundant decomposition of the
(radical) ideal 1 as an intersection of prime ideals, and the latter is
that this irredundant decomposition is unique up to the order of the
involved prime ideals.
Let us deal at last with algebraically closed fields I( and with Model Theory:
we want to show that I( uniformly eliminates the imaginaries. As observed
in Chapter 4, it is sufficient for our purposes to prove that any such I(
eliminates the imaginaries and so to find, for every 0-definable equivalence
relation E = E (v, w) in K" and i1 in K";
6.3. w-STABLE FIELDS
* a formul a <p (v, Z),
* a unique sequence bin K
195
such t hat E(Kn, a) = <p( K
n,
b) (here t he common length of z and bmay be
f n) . There is no loss of generality in ass uming K in fact,
given ain K"; if t here exists a unique bin some extension K'
of K such t hat K' F Vv(E(v, a) H <p (v, b) ), t hen t his t uple bmust belong
to dcl (a) and consequently t o K.
Now we use the w-stabilit y -rnore pr ecisely th e total t ra nscendence- of K ,
as promised. In fact, given a E K" and t he formula E(v, a) , there are
onl y finit ely many n-types Po, ... , Pqover K containing E (v, a) and having
th e same Morley rank as E (v, a). Let 10, . .. , Iq be th e prime ideals of K[X]
corres ponding to these t ypes. Partition these ideal s into conj ugacy orbits, so
into equivalence classes with respect to the relation linking two ideals exactly
when t hey corres pond to each oth er by some au tomorphism of K . Take any
orbit O. Up to rearranging t he indexes, one ca n ass ume 0 = {I
o
, .. . , I
m
}
for some m ::; q. Appl y 3 and find a (count able) subfield K(O) of K such
t hat , for every aut omorphism (J of K, (J permut es 10, .. . , I
m
if and onl y if
(J fixes K(O) pointwi se. Let Ko be t he count abl e subfield of K generated by
t he K(O) 's when 0 ranges over t he orbi t s of 10, ... , Iq. We claim t hat, for
every (J , (J fixes E( Kn, a) setwise if and onl y if (J acts identi cally on Ko.
In fact, if (J pr eserves E( Kn, a), t hen (J permut es the ty pes Po , ... , Pq con-
taining E(v, a) and having t he same Morley rank as E(v, a); hence (J per-
mutes t he corres ponding ideals 10, ... , Iq as well, and in par ticular any
single orbit O. Hence (J fixes pointwi se each K(O) and in conclusion Ko.
Conversely, let (J act identically on K, and so on any K (O) . It follows
t hat (J permutes th e ideals in any orbit 0 and so t he whole set 10, ... , Iq.
Equivalent ly (J permutes Po, . . . , Pq . But (J has to preserve E because E
is 0-definable, and so takes E-classes to E-classes. In particular (J fixes
E(Kn, a) because the formul a E(v, a) lies in all th e t ypes Po , ... , Pq.
Therefore any automorphism of K fixing a pointwise, and so preserving
E(v, a), fixes Ko pointwise as well; in other words, Ko dcl(a).
Now assemble all the L(1<o)-formulas 'Ij; (V) for which E(Kn, a) 'Ij; (Kn).
They form a possibly incomplet e type P over Ko. We claim
E( K
n,
a) = n 'Ij; (K
n
).
'ljJ (V)Ep
Oth erwise t here is some l E K " sat isfying P and a). This remains
t rue for every ({1 E K" having t he same type as lover Ko; in fact , as K is
196
CHAPTER 6. w -STABILITY
and 1(0 is countable, d! correspond to each other by some
automor phism (Y of K fixing 1(0 poi ntwi se, hence E( Kn, ii) setwise. Then
tp(d/I(o) impli es -,E(v, ii) in K. Use compact ness and det er mine B(if) E
tp(d/I(o) such t hat B(Kn) -,E(Kn, ii ), equivalent ly -,B(Kn) 2 E(Kn, ii ).
But t his forces -,B(v) E p tp(d/I(o), a cont radict ion.
Consequent ly E(v, iD is a consequen5e of p. Again use compactness and
find a formula <p(v, b) in p - so wit h bin 1(0 - impl ying E(v, ii), and hence
equivalent t o it. At t his point it remains t o show t he uniqueness of b. This
can be obtained by some simple manipulations, slight ly cha nging <p (v, b). In
fact , t ake any fi # bin I( having th e same type as bover 0. There is some
aut omor phism (Y of K sending bt o fi , so (Y does not act identically on 1(0
because it moves b
i
; accordingly (Y does not fix setwise E(Kn, ii) = <p(Kn, b),
in other words <p(Kn, b) # <p (Kn, b') . This is t rue for every b
i
# b realizing
tp(bj0). By compact ness, t here is some formul a 1J( Z) in tp( bj0) for which
K F= VZ(1J (Z) 1\ -, (z = b) -+ -,(Vv(<p(v, Z) <p (v, b)))).
So, unless replacing <p (v, Z) by r.p' (v, D : r.p(v, Z) 1\ 1J (Z) , we t hat bis
t he onl y t uple in I( such t hat r.p'( Kn, b) = E( Kn, ii): in fact , if b' does not
satisfy 1J (Z) , t hen r.p( Kn, b
i
) = 0. This accomplishes our proof. ..
6.4 Prime models
T his section t reats a fund ament al feature of w-stable t heories, namely the
existence and uniqueness of prime models over subsets. This property will
be useful in 7.8, in proving Morley's Theorem on uncoun t abl y categorical
t heories; but a remarkabl e application will be provided also in 6.5, when
dealing with differ ential fields; in fact we will observe t hat DCF
o
is w-stable,
and so the machinery of this sect ion will apply and yield both existence and
uniqueness of prime models over subsets among differentially closed fields of
characterist ic 0 - in algebraic t erms, exist ence and uniqueness of differenti al
closures among differenti al fields of characterist ic 0 -.
The first ste p of our t reat ment is clearly to define what a prime mod el is.
Let T be any (possibl y non w-stable) complet e t heory wit h infinite mod els
in a count able language L . n still denot es a big sat urated model of T .
Defi nition 6.4.1 Let X be a small su bset of n. A model A of T is called
prime over X if and only if:
(i) there is an elemen tary function f of X in to A,
6.4. PRIME MODELS 197
(ii) for every model B of T and elementary f unct ion g from X into B,
there is an elementary embedding h of A into B for which h f = g.
Two problems spring qui t e naturally in t his setting. Let X n be given.
(a) (Existence) Is t here any model A of T prime over X?
(b) ( Uniquen ess) Is t his model unique? More precisely, let A
o
and A
l
be t wo models of T prime over X - via t he elementary funct ions f o,
!l respect ively -. Are A
o
and A
l
isomorphic by a map h satisfying
hfo = i l ?
Intui tively speaking, a model of T pnme over X is a "minimal" model
exte nding X by some elementary function . Of course a sharp definition in
a genera l set t ing fatally invol ves all the details given before. In particular
uniqueness has to be required up to isomorphism, in th e way stated some
lines ago. On t he other hand, t his generalit y has it s positi ve sides, and even
allows some well accepted freedom. For inst ance, just owing to (b), t here is
no loss of generality in assuming t hat a model A prime over X , if any, does
extend X (and so f is an inclusion).
We will see t hat existence and uniqueness of prime models are not guar an-
teed in general, for ar bit rary t heor ies T and over arbit rary sets X . However
t hey do hold when T is w-stable. But now let us propose some examples.
Examples 6.4.2 1. Let T = ACF or , if you like to keep our complete-
ness ass umpt ion on T , let T = ACF
p
for some p = 0 or prime. When
X is an int egr al domain wit h identit y in cha racterist ic p, a mod el of
ACFp prime over X is t he algebraic closure of t he field of quoti ents of
X (in t he field th eor et ic sense), and is unique (as t he Algebra hand-
books relate).
2. Let T = RCF. A model of RCF prime over an ord ered field X is it s
real closure, and is uniqu e up to isomorphism. Again, any handbook
of Algebra (such as [65]) provid es t he details.
3. Now let T = DCF
o
. Take a differential field X of characterist ic O. Now
Algebra does not tell us anyt hing similar t o t he previous examples; in
fact it clarifies neither t he existence nor t he uniqueness of a differenti al
closure of X , indeed Algebra ca nnot eit her define what a differential
closure is.
So t he existe nce and uniqueness probl ems of prime models overlap some non
trivial (unsolved?) algebraic questions.
Now let us summarize briefly how we plan to organize this section. First we
introduce and discuss some notions closely related to prime models: atomic
sets and constructible sets. Indeed we will see that over countable sets a
model is prime if and only if it is atomic and countable, and if and only
if it is constructible (warning: here constructible has a specific meaning,
having nothing to do with what we introduced in Chapters 1, 2 and 3 in the
particular framework of fields). In this perspective, a theorem of Ressayre
stating the uniqueness of a model constructible over a given subset is very
not eworthy, as it implies, among other t hings , th e uniqueness of a prime
model over a countable set. At t his point we will discuss the existence
problem of prime models. We will give a condition of topological flavour
characterizing the theories T such that every small subset X of n admits
a prime model over itself: they are exactly t hose where the isolated types
over every X are den se in S(X). This will accomplish the general analysis
for arbit rary theories.
At this poi nt we will ass ume T w-stable, and we will show a theorem of
Morley proving that, in this setting, the pr evious topological condit ion is
satisfied and, consequently, t he existence of prime models is always guar-
anteed.Finally we will treat a nice and deep theorem of Shelah showing
th e uniqueness of prime models in the w-stable framework. Act ually what
Shelah proves is that , for an w-stable T, a model prime over a set X is also
constr uct ible over X; so Ressayre's Theorem applies and gives the required
uniqueness result. As already said, Shelah's Th eorem is quite complicated,
and needs some subtle notions and tools to be introduced lat er in this book.
So we will postpone its proof unti l Chapter 7. To conclude, we will see th at
both exist ence and uniq ueness of prime models fail when the w-st ability
assumption is dropped.
This is th e sketc h of this section. Now let us begin our report by introducing
atomic and constructible sets, as promised. We ass ume some basic acquain-
tance wit h ordinal numbers. X, Y , Z , ... denote small subsets of n, a, b,
c,.. . tuples in n. For simplicity we sometimes ident ify a tuple aand t he
(finit e) set of th e elements in a.
Definit ion 6.4.3 Y X is atomic over X if and only if every tuple in Y
has an isolated type over X .
Recall th at every model A of T extending X via an elementary function
realizes all the isolat ed types over X; on the contrary, non-isolat ed types
can be omitted by suit ably chosen models, as t he Omitting Types Theorem
shows. So a mod el prime over X, being elementarily embeddable in every
model A extending X as before, should not be far from being atomic over
X.
Here is a list of simple and useful facts concerning atomic sets and isolated
types.
Remarks 6.4.4 1. If tp(b, clX) is isolated, then tp(bl X) is.
In fact, let <p(iJ, w) be an L(X)-formula isolating tp(b, cIX). An easy
check shows that :Jw<p(V, 117) isolates tp(bl X). In fact b does satisfy
:Jw<p(V, 117). Furthermore, for every fji realizing :Jw<p(iJ, w), there is
some c' for which 1= <p(fji, c'} Then (b
',
c') has the type as (b, c)
over X, and consequently b' has the same type as b over X.
2. If tp(b, cl X) is isolated, then tp(bl X U C) is.
In fact, if <p(v, w) is an L(X)-formulaisolating tp(b, clX) , then <p(v, C)
isolates tp(bl X U C).
3. Let X Y. Suppose tp(C/Y) isolated by some L(Y)-formula <p(iJ, b)
with parameters bin Y, and tp(bIX) also isolated. Then tp(C/X) is
isolated.
In fact, take an L(X)-formula 7jJ(w) isolating tp(bIX). Look at the
L(X)-formula 8(v) : :JW(7jJ(1V) A <p(v, w)). Clearly n 1= 8(C). We
claim that 8( iJ) isolates tp(clX). For, let 2 satisfy 8(v) , so for some
fji , n 1= 7jJ (fji ) A <p(2, fji). In particular b
'
has the same type over
X as band hence there is an automorphism f of n fixing X point-
wise and taking b to fji. Then <p(v, b
')
isolates a type over f(Y); as
n 1= <p(2, fji), tp(2I f(Y)) equals this type and so contains every for-
mula O'(v, f(d)) with O'(v, d) E tp(C/Y). In particular, as f fixes X
pointwise, tp(2I X) = tp(Cl X).
Notice that 3 implies that atomicity is transitive: for X Y Z, if
Z is atomic over Y and Y is atomic over X, then Z is atomic over X.
4. Let Y :2 X be atomic over X, cbe in Y. Then Y is atomic over X Uc.
In fact , for every bEY, tp(b, clX) is isolated. So 2 implies tp(blxuc)
isolated, too.
Definition 6.4.5 Y :2 X is constructible over X if one can list the ele-
ments ofY by ordinal indexes b
v
(with v < 0', 0' a suitable ordinal) in such
a way that, for every v < 0' , b; has an isolated type over X U {bJ1 : fL < v}.
In this case the sequence (b
v
: u < 0') is called a construction of Y over
X.
We put for simplicity X; = X U {b/-l : fL < v}. So X
o
= X , for a limi t
v X; = U/-I <v XJ.L and, for a successor v = fL + 1, X; = X/-I U {b/-l}' Of
course Y = Uv<a X v. It is quite trivial that, if b; E X , then tp(b
v/
X), and
consequently tp( b; / Xv), is isolated (by v = b
v)'
Now we discuss t he relationship between atomicity and const ructibility.
First we propose a simple resul t.
Lemma 6.4.6 Let Y ;2 X be atomic over X. If Y - X is countable, then
Y is constructi ble over X .
Proof. Let ba, b
l
, ..., b
n
, ... (n natural) enumerate Y - X . It is enough
to show th at , for every n, th e type of b
n
over X
n
= X U {b
m
: m < n} is
isolated. As X
n
- X is finit e, t he previous Remark 6.4.4.4 ensur es that Y
is isolat ed over X
n
, and so that tp(b
n
/ X
n
) is isolated, as required. .,
So at omicity sometimes implies const ruct ibility. But a stronger and deeper
resul t holds in t he opposite dir ecti on: in fact , as we will see wit hin a few
lines, if Y ;2 X is construct ible over X, t hen Y is atomic over X.
Now we relat e const ruct ibility and prime models. Fi x a small X 12 and
t ake a model A of T exte nding X by some given elementary function. As
already said, there is no loss of gener ality for our purposes in ass uming th at
X is just a subset of A and that t he corresponding inclusion is elementary.
Theorem 6.4.7 If A is construc ti ble over X , then A is prime over X .
Proof. Let B be a model of T and 9 be an elementary function from X int o
B. We are looking for an elementary embedding h of A into B extending
9. Let (b
v
: v < a ) be a const ruct ion of A over X. The st rat egy we devise
to build h is t he following: we progressively exte nd 9 to larger and larger
elementary functions 9v from X; into B (for v < a ); at last , we check that
Uv<a 9v is just t he required embedding h.
We proceed by inducti on on u, For v = 0, we put 90 = 9. Wh en v is limit ,
we set 9v = U/-I <v9/-1 , a function from X; = U/-I <vX /-I into B. It is easy t o
check t hat 9v is elementary. In fact let ii in X; sat isfy F ep(ii) for some L-
formula ep (v);for a suffi cient ly large fL < u, ii lies in X I-'" So t he elementarity
of 9/-1 ensures B F ep (9/-1 (ii)). As 9v ;2 9/-1 ' B F ep (9v(ii)).
Finally let t/ = fL + 1 be a successor ordinal. Look at t he isolat ed ty pe
tp (b/-l/ X /-I)' Call it P/-l for simplicity ; 9/-1(P/-l ) is in it s turn an isolated ty pe
over 9/-1(X/-I) B , and so it is sat isfied by some element E B. Ext end 91
L
by including the new element b/-l in the domain and mapping b/-l into Let
9v be the resulting function. Then the domain of 9v is just Xv = Xjt U{bjt};
moreover, owing to the choice of 9v is elementary.
At last form f = Uv<a 9v. The same argument as before shows that h is an
elementary embedding of A into B (recall A = Uv<a Xv). ..
So constructible models are prime. At this point, we show that constructible
models satisfy the uniqueness assumption. This is a non trivial result of
Ressayre. But, before stating it exactly and beginning its proof, we need
some technical preliminaries. So let Y be constructible over X, fix a con-
struction (b; : v < Cl:) of Y over X. For every v < Cl: choose an L(Xv)-
formula <Pv(v, a-:) isolating the type of b; over Xv, in particular fix the
parameters a-: from Xv' When b
v
EX, choose v = b
v
as an isolating for-
mula. In this framework, a subset Z of Y is called closed (with respect to
the given construction of Y over X) if, for every v < Cl: with b
v
E Z, even
the defining tuple a-: is in Z. Clearly X and Y are closed; moreover closed
sets are preserved under (finite) union and intersection.
Lemma 6.4.8 Let X Z Y, Z be closed. Then the theory of Dz is
axiomatized by Th(Dx) U {<p(b
v
, a-:) : v < Cl:, b; E Z}.
Proof. For every J-L ::; Cl: put Zjt = {b
v
E Z : t/ < J-L}. It is an easy exercise
to see that Zjt is closed for every J-L ::; Cl:. Moreover Za coincides with Z. We
claim that, for every J-L ::; Cl: ,
Th(Dxuz,J is axiomatized by Th(Dx) U{<Pv(b
v,
a-:) : b
v
E Zjt}.
This is clearly enough, as the case J-L = Cl: is just our thesis. Now it is obvious
that
Th(Dx) U {<Pv(b
v,
a-:) : b; E Zjt} Th(Dxuzl')
because, for every v and J-L, when b; E Zjt, a-: is in Zjt, too. So we have
to prove that , conversely, every sentence in Th(Dxuzl') is a consequence of
Th(Dx) U {<Pv(b
v,
a-:) : b; E Zjt}. We proceed by induction on J-L. When
/J, = 0, Zo is empty and so the claim is trivial. When J-L is limit, we can
observe that a sentence in Th(Dxuzl') involves only finitely many elements
in Zjt and so belongs to Th(DxuzJ for some suitable u < J-L. Hence the
induction hypothesis applies. At last , take a successor J-L = v + 1. We
can assume b., E Z, otherwise there is nothing to prove. As <Pv(v, a-:)
isolates tp(bv!X
v),
the sentences concerning b; in Th(Dxl') are proved by
Th(Dxl')U{<Pv(b
v,
a-: n. When we restrict our attention to Z, we can deduce
by standard arguments of elementary logic that
because b; and consequent ly a: are in Z . Notice xJ.Lnz = XUZ
JLl
X vnz =
X U Z; and deduce that
Th(Ox uzl') is axiomat ized by Th(Ox uzJ U {<Pv(b
v
, a: n.
But owing to t he induction hypothesis Th(OxuzJ is axiomatized in it s turn
by Th(Ox) U{<PA(b
A
, a)J : A < 1/, b
A
E Z }, whence the last set of sent ences,
together wit h <Pv(b
v
, a:) , axiomatizes the t heory of OXUZI', as claimed. ...
Still keeping the same not ation, now we show:
Lemma 6.4.9 If Z Y and Z is clos ed, then Y is at omic over X U Z .
Notice t hat t his applies also to Z = X and impli es
Corollary 6.4.10 IfY 2 X is const ructi ble over X , then Y is atomic over
X.
Proof. (of the lemma) Let hbe in Y , we have to show that t he ty pe of h over
X U Z is isolated. Extend hto a larger and closed tuple cin t he following
way. For every b E h, pick 1/ < 0: for which b = b
v,
and consider the formul a
<p (v , a:) isolating t he type of b over X v. For every a E a: , take f1 1/
such t hat a = bJ.L ' and repeat t he previous procedure. As there is no infinit e
st rict ly decreasing sequence of ordinals, this machinery stops after finit ely
many steps. Let cabsorb all t he elements arising in this procedure. Clearly
cis closed , as well as Z U cand X U Z. Owing to Lemma 6.4.8, the t heor y
of Oxuzue is axiomatized by Th(Oxuz) U {<Pv(b
v
, a:) : 1/ < 0: , b; E 2}.
Hence /\ bvEC<Pv (vv, a:) isolates the ty pe of cover X U Z. As cenlarges h,
tp(hj X U Z ) is isolated, too. ...
Now we can st at e and prove Ressayre's Uniqueness Theore m.
Theorem 6.4.11 (Ressayre) Let X be a small subset of 0 , A
o
and Ai be
t wo models of T containing X via some elementary i nclusions, and con-
structi ble over X. Then there is an isomorphism h of A
o
onto Ai fi xing X
pointwis e.
Proof. Look at th e triples (Y
o
, Y
i
, f) where, for every i = 0, 1, X Yi
Ai , Yi is closed with respect to some given const ruct ion of Ai over X and f
is an element ary function of Y
o
int o Ai acting identically on X and having
Y
i
as image. The collection of th ese triples is not empty because it cont ains
(X, X, i dx) (in fact , X is closed) ; moreover it is parti ally ordered by t he
relation ::S such that
(Y
o
, Y
l
, J) ::S Y{, 1')
if and only if Yi Y/ for every i = 0, 1 and f 1'. A st raight forward
check ensures t hat Zorn 's Lemma applies to ::S and yields a maxi mal t riple
(Y
o
, Y
l
, ] ). We claim t hat Yi = Ai for every i = 0, 1, so ] is the required
isomorphism.
Suppose not . For inst ance, ass ume Y
o
f= A
o
. Use t he same technique as in
th e proof of Lemma 6.4.10 and enl ar ge Y
o
t o a closed Y
o
o 2 Y
o
U {b} wit h
Y
o
O - Y
o
finit e. Owing to Lemma 6.4.10 t he sequence Co of t he element s in
Y
oo
- Y
o
has an isolat ed type over Yo. Call it p. Not ice th at even ](p) is
an isolated type (over YI) because] is elementary. In par ticular ] (p) is
reali zed in AI, say by Cl . Enlarge Y
l
by Cl and get a new set Y
lO
2 Y
l
in
Al . Clearly ] can be extended t o an elementary function f Ofrom Y
o
o into
. Al wit h image Yd: j ust map Co in Cl ' If Yd is closed wit h respect to t he
fixed const ructi on of Al over X , t hen we are done; in fact , we have found
a counterexample (Ycf , YlO, f a) to t he maximality of (Y
o
, Y
l
, ] ); t his is a
cont radiction, and so shows Y
o
= A
o
as claimed. Otherwise we apply t he
previous pr ocedure t o Y
lO.
Accordingly we enlarge Y
l
O t o a closed Yl 2 Y
l
O
with Yl - Y? finit e, and correspondi ngly we build Ycf Yd A
o
and an
elementary func tion i' from Yd in A l with image Yl. If Y
o
l
is closed, th en
we are done; ot herwise we cont inue our procedure. In t he worst case, we get
for every natural n a t riple (Y
o
n
, Yr, f n) such t hat , for every nand i = 0, 1,
(Y
o
, Y
l
, ]) ::S (Yon, Yr , I" ) , X C Yi
n
+
l
, I" is an element ary map from
Y
o
n
into Al wit h image Yt, Yi
n
- Yi is finit e and both Y
0
2n
+
l
and y
1
2n
+
2
are
closed. Owing to t he last condit ion both UnYon and Unyr are closed. So
(UnYon, UnYt , Unr) again cont radicts t he choice of (Y
o
, Y
l
, ]).
In conclusion Y
o
= A
o
, and, similarly, Y
l
= A l .
But now it is time to come back t o t he existence and uniqueness problems
of prime mod els. As a warm-up, let us first discuss the case X = 0. Here
we show uniqueness and we characterize existence by some equivalent con-
di tions. Not ice t hat t he ass umption X = 0 is not so rest ricti ve as it may
look. Indeed it impli citl y includes t he seemingly mor e general case when X
is countable: it suffices to repl ace L by t he (countable) language L(X) and
T by t he t heory of D
x
, and t o observe th at a model of T h(Dx) prime over
ois just , as a st ruct ure of L , a mod el of T prime over X . Not ice also t hat
a model of T prime over 0 is element arily embeddable in every mod el B
of T ; in fact, as T is complete, the inclusion of 0 in B is element ary. Th e
uniqueness of a model prime over 0 is a direct consequence of Ressayr e's
Theorem and the following result.
Theorem 6.4.12 Let A be a model of T. Th en the following propositions
are equivalent:
(1 ) A is prime over 0,.
(2) A is countable and atomic over 0,.
(3) A is constructible over 0.
Proof. (1) =} (2) As already said, a model A of T pr ime over 0elementarily
embeds it self in any mod el ofT. This clearly implies that A is countable, as
T does admit some countable mod els. Furthermore ever y non-isolated ty pe
over 0 is omitted in some countabl e model of T (by the Omitting Types
Theorem) and consequentl y in A: so A is atomic over 0.
(2) =} (3) Just apply Lemma 6.4.6 t o Y = A and X = 0.
(3) =} (1) This is a particular case of Theorem 6.4.7. ..
At this point Ressayre's Uniqueness Theor em applies and shows:
Theorem 6.4.13 Any two models of T prime over 0 are isomorphic.
In fact t hese mod els are const ructible over 0. Now let us discuss th e exis-
te nce of a mod el of T prime over 0. The following t heorem provides a nice
equivalent condit ion.
Theor e m 6.4.14 There exists a model ofT prime over 0 if and only if, for
every positi ve integer n, the isolat ed n-types over 0 are dense in the topologi-
cal space Sn(0) (thi s means that , for every L-formula <p (v) in n free variables
v, if <p (nn) =F 0, then there is some isolated n-type over 0 containing <p (v)) .
Proof. (=}) Assume that A is a mod el ofT pri me over 0. As T is complet e,
cp (A)n is not empty, in oth er words t here is some tuple ii in A sat isfying
cp (v) . As A is atomic over 0, th e ty pe of ii over 0 is isolat ed (and includes
<p(v)) .
({::: ) The st ra t egy here is t o use some classical t echniques (adding constants,
referring t o the Tarski-Vaught Theorem and so on) and to construct a count-
able model A cons tructible over 0: as said before , th is is also prime over
0. In order t o form A we build pr eliminarily a sequence of isolated types
PI <;:;: P2 <;:;: ... <;:;: Pn <;:;: .. (with n a posit ive int eger) such that , for every n,
(a) Pn is an n-type over 0 (in t he free vari abl es (Vi, . .. , v
n)
= v(n));
(b) if 3w<p(w, v( n)) E Pn for some formula <pew, v( n)), t hen for some
integer m > n <p(v
m,
v (n)) E Pm.
These ty pes ar e buil t by induction on n. Pi is any isolat ed I-t ype over 0.
Now assume t hat Pn is defined, we show how t o form Pn+!' Let /"n (v(n))
be a formul a isolating Pn, list all t he formul as 3w<p( w, v(n )) occuring in Pn.
As Pn 2 Pk for k n, t his list includes all the formulas of t he same kind
occurring in Pk with k < n. Take t he first formula 3w<pn(w, v (n)) in this list
not already considered in t he previous steps < n. Observe 3w<pn(w, v(n)) 1\
/"(v(n )) E Pn, so <Pn(w, v(n)) I\/"(v(n)) occurs in some (n+l)-type over 0. Use
our ass umpt ion that isolated (n + 1)-types over 0 ar e dense in Sn+!(0) and
deduce that <Pn(w, v(n)) 1\ /" (v(n )) actually occurs in some isolated (n + 1)-
ty pe. Choose such a ty pe as Pn+l .
Actually t his cons tructi on should involve sooner or lat er any formula 3w<p( w,
if) occurring in some Pk when k ranges over positi ve int egers. But t his can
be obtained by using a suitable diagonal Cantor like procedure.
Now extend t he language L of T by count abl y many new constants C =
{ Ch : h positi ve int eger} . For every h replace Vh by Ch in t he formul as of
P = UnPn One get s a set T
p
of sentences in t he enlarged language L UC: T
p
is a cons istent t heory, and indeed a complete t heory, as t he reader ca n easily
check. Now take a model Bp of Tp. Tp is just t he t heory of Bp, hence t he
(LUC)-sentences of t he form 3w<p(w, C) (wit h c in C) t rue in Bp are exactly
t hose occurring in T
p
Look at t he set A = : h positi ve int eger} and
not ice t hat, owing to (b) and t he Tar ski-Vaught Theorem, A is t he domain
of an element ary substr ucture of Bp. Restrict this subst ruct ure to L and get
a mod el -actually a countable model- A of T. Moreover A is atomic over 0,
as required: in fact, given a sequence c = (Cl , ... , c
n
) from C , the type of
i!3pover 0 cont ains Pn, hence equals Pn and is isolated. ..
This concludes our discussion of th e case X = 0. As already said, when
we enlarge our attention to countable sets X , we immediately obtain what
follows.
Corollary 6.4.15 (1) There exists a model of T prime over X if and only
if, for every posit ive integer n, the isolated n-types over X are dense
in Sn( X) ,
(2) Two models ofT extending X by elementary inclusions are isomorphic
by a fu nct ion fi xing X pointwise.
But t he pr evious analysis discloses some useful informati on even in t he gen-
eral case, when X is any, possibl y uncountable , small subset of D. In partic-
ular, it clarifi es under which condit ions on T any X admits a prime model.
This is what the following t heorem explains.
Theorem 6.4.16 The following propositions are equivalent.
(1) For every X, there is a model of T prime over X.
(2) For every X and positive integer n, the isolated n-types over X are
dense in Sn(X) .
(3) For every X , the isolated I -types over X are dense in SI(X).
(1)' For every countable X , there is a model ofT prime over X.
(2)' For every countable X and posit ive integer n, the isolated n-types
over X are dens e in Sn(X) .
(3)' For every countable X , the isolated I- types over X are dens e m
SI( X ).
Proof. (I)::::} (1)' , (2) ::::} (2)' and (3) ::::} (3)' are trivial , as well as (2) ::::} (3)
and (2)' ::::} (3)'. Moreover (1)' {:::? (2)' is just Corollary 6.4.15. So, in order t o
accomplish the proof, it suffices to show (3)'::::} (3) , (3) ::::} (2) and (3) ::::} (1).
(3)' ::::} (3) Suppose towards a contradict ion t hat there are X , a tuple ain X
and a formula <p (v, w) such that <p (v, a) occurs in no isolat ed I-ty pe over X.
We want to extract a countable X ' X such that X' contains abut <p (v, a)
occurs in no isolated l-t ype over X': this is clearly enough because such a
set X' contradicts (3)/. Let X
o
denote the set of t he elements in a. Is X
o
a
reasonable ca ndidate as X '? No, it is not , because it may happen t hat, in
t he rest ricted framework of X
o
, some L(Xo)-formula 'Ij1 (v ) isolates a single
L-t yp e cont aining <p (v, a). However 'Ij1 (v ) cannot preserve t his property over
X, so there are at least two different l-types over X containing 'Ij1 (v)A<p( v , a);
consequent ly there exists some formula O( v, b) wit h parameters bfrom X
occurring in t he former ty pe but not in the lat t er . For every 'Ij1(v) as above,
fix a corresponding b (and O( v, b)) . Enlarge X
o
by adding t he elements
of these t upies bwhen 'Ij1 (v ) ranges over the L(Xo)-formulas isolat ing some
type over X
o
containing <p (v, a) . One gets a count able Xl X enlarging a
and such that no L(Xo)-formula 'Ij1 (v ) isolates a I-type over Xl containing
<p (v, a) . Indeed, for every 'Ij1 (v), t here is some for mula O( v , b) in L(X
I
) for
which both 1jJ(v) 1\ r.p(v, if) 1\ ()(v, b) and 1jJ(v) 1\ r.p(v , if) 1\ -,()(v, b) enlarge to
consistent l-types over Xl' For any L(Xo)-formula 1jJ(v) which cannot isolate
any l-type containing r.p( v, if) even in 51 (X
o),
()( v, b) can be also found in
L(X
o
) ; otherwise ()(v, b) is produced by the previous procedure. We repeat
this machinery and we build, for every natural n, a countable X
n
X
enlarging if in a way such that, for any n, no L(Xn)-formula can isolate a
l-type over X
n+!
containing r.p(v, if). At this point, consider X' = Un X
n.
X' in countable. Moreover, for every L(X')-formula 1jJ (v), there is a suitable
n for which 1jJ(v) belongs to the language L(X
n)
and hence cannot isolate
any l-type containing r.p(v, if) over X
n
+
l
, and consequently over X'.
(3) ::::} (2) We proceed by induction on n. As the case n = I is just our hy-
pothesis, we have simply to see how to pass from any n to n+1. Let r.p(v, w)
be an L(X)-formula with n+1 free variables (v, w) (w = (WI, ... , w
n
)) and
r.p(nn+!) =I- 0. So :3vr.p( v, w) lies in some n-type over X, and hence in some
isolated n-type p over X. Let b realize p, so n p= :3vr.p(v, b), and r.p(v, b) is in
some I-type over X Ub. By hypothesis, there is some isolated l-'type q over
X U b containing r.p(v, b). Let a realize q, so p= r.p(a, b). Look at the type of
(a, b) over x: it contains r.p(v, w) and is isolated because both tp(ajXUb)
and tp(bj X) are.
(3) ::::} (I) Fix X, we have to find a model of T prime over X . As shown
in Theorem 6.4.7, it suffices to build a model A of T constructible over X.
Let us work in a model M of T containing X via an elementary inclusion.
To obtain A inside M, we extend progressively X and form larger and
larger sets XJ.L 2 X in M (for It an ordinal), all constructible over X and
elementarily included in M, until we meet in this increasing sequence the
domain of an elementary substructure of M and, in this way, a model of
T constructible over X. Start putting X
o
= X. For a limit It, set X J.L =
UV<J.L XV, as it is right to expect: indeed this preserves constructibility and
the other assumptions on X
w
The crucial step concerns successor ordinals
It = v+1. Suppose that Xv is not the domain of an elementary substructure
of M (otherwise we are done). By the Tarski-Vaught Theorem, there is an
L(Xv)-formula r.p(v) such that :3vr.p(v) is true in M - in other words r.p(M) =I-
o- but X; contains no realization of r.p(v). Use (3) and get some isolated
l-type p over Xv including r.p(v). There is some element a
v
E M satisfying p.
Put Xv+! = XvU{a
v};
Xv+! is constructible over X because p = tp(avj Xv)
is isolated; moreover the inclusion of X
v
+
l
in M is elementary. Now notice
that the length of this procedure cannot exceed IMI. Accordingly XJ.L =
XJ.L+! for some It IMI; but this means that XJ.L is the domain of the
required model. ..
Let us underline one more time t hat what we get in (3) =} (1) is actually a
model of T constructible over X .
Now assume, at last , T w-stable. We show t hat, under this assumption, th e
existence of pr ime models is ensured over any X. This is a result of Morley
and is obtained as a direct consequ ence of t he previous theor em.
Theorem 6 .4.17 (Morley) Let T be ta-stable. Th en, for every X , there is
a model of T prime over X .
Proof. Owing to Theorem 6.4.16, it suffices to show t hat, for every count-
able X , t he isolated I-types over X are dense in 51(X) . Use w-stability and
deduce that, for a countable X, 51(X) is countable, to o. Now refer t o topol-
ogy and specifically recall a theorem of Cantor and Bendixson saying t hat
in every count able compact Hausdorff space - like 51(X) - isolated points
are den se. This is just what we need. ..
Indeed th e previou s t heorem says even more: in fact , for an w-st able T , over
any X we can find a constr uctible mod el.
Finally let us deal with uniqueness in th e w-stable set t ing.
Theorem 6.4.18 (Shela h) Let T be w-stable, X be a small subset of rl , Aa
and A
1
be two models of T prime over X . Th en there is an isomorphism h
between Aa and A
1
fixing X pointwise.
Actually what Shelah proves is t hat , for an w-stable T , a mod el A prime
over X must be constructible over X as well. So Ressayr e's Theor em applies
and ensures the uniqueness of t he prime model over X . However , Shelah 's
tools in th e proof ar e quite sophist icated and require new progress in t he
general theory. We will develop t hese preliminari es in Chapter 7, so we have
to delay th e full details of the proof to t hat chapter.
We concl ude t his section observing that neither exist ence nor uniqueness of
pr ime models are guarant eed when T is not w-stable. In particular we pro-
pose a simple counterexample to existe nce. Uniqueness is also cont radicted
by some suit a ble th eori es T , but t he corresponding examples are more t ech-
nical and complicated, and may make this section still heavier. We omi t
them.
Example 6.4.19 In the language L = {S; , P} , where P is a l -ary relation
symbol, look at the st ruct ure (R, S; , Q). This is a dense linear order without
endpoints (R, S;) with a subset Q both dense and codense in R. All th ese
6.5. DCF
o
REVISITED 209
proper ti es can be easily wri tten as first ord er sentences in L. Let T be t he L-
theory axiomatized in t his way. A simple arrangement of Cantor 's Theorem
on DLO- shows that T is and so, by Vaught's Test , complet e.
Hence T = Th(R , ::;, Q). Of course T is not w-stable (just as DLO- ).
Now suppose t hat T has a prime model over Q inside (R, ::;, Q) . Up to
isomorphi sm this model will be of t he form (A, ::;, Q) for some countable
subset A of R properl y including Q. Let a E A - Q. If we take a away
from A, t hen we get a new model (A - {a} , ::; , Q) of T ; t his contains th e
rationals, but cannot embed (A , ::;, Q). In fact t he cut a fills in A over Q
is not realized in (A - {a}, ::;, Q).
6 .5 DCF
o
revisited
We pursue here th e analysis of existenti ally closed differential fields of cha r-
acterist ic 0 begun in Chapte r 2. We said at t hat t ime t hat their class
is elementary and we provided (wit hout proof ) t heir nice first order char-
acterization DCF
o
due t o Lenore Blum, according t o which t hey are th e
differenti al fields K of characterist ic 0 such t hat , for every non zero j (x)
and 9(x) in [({x } such t hat t he ord er of 9(x) is smaller t ha n t he ord er of
j (x), t here is some a E [( such t hat j (a) = 0 and g(a) =I 0 (let us mo-
ment aril y call t hese fields differenti ally closed fields; act ually we are going
to show t hat t hey are just t he existent ially closed differenti al fields). Now it
is t he right moment to prove at last t hat result , as well as th e related fact ,
mentioned several t imes, th at every different ial field of charact erist ic 0 has
it s own differential closure. In fact, we will show t hat DCF
o
is w-st able,
and so t he machinery developed in t he last section will apply and produce
a prime model of DCF
o
(so a differential closure) over any differen ti al field
of cha racterist ic O. The w-stability of DCF
o
will be also used to prove, in a
way closely resembling that pursued for ACF in 6.3, th at DCF
o
uniformly
eliminat es the imaginari es.
This is the plan of t his section. In order to begin our t reat ment and, in
particular, to focus existent ially closed differential fields, let us preliminarily
examine how a differential fi eld K can be enlarged and which is t he st ruct ure
of a differ ential extension H. of K. In par t icular, for a E H - K; we want
to single out in 'H. t he smallest differential subfield K (a) containing [( and
a (i. e. t he int ersection of all t he differential subfields of}{ cont aining both
[( and a); t his requires t o clarify and und erst and how a is relat ed t o K , in
oth er words which differenti al polynomials over [( a sat isfies. Of course t his
is basic (Different ial) Algebra rath er t han Mod el Theory; and in fact t here is
210 CHAPTER 6. w -STABI LI TY
a straight forward and general algebraic technique to approac h t his question,
t hat is t o consider t he function Fa from the differential domain K{ x} int o
1 taking any differenti al polynomial f (x) E K{ x} to it s value f (a) in a .
This is a different ial ring homomorphism, and it s kernel [ (aIK) = {f (x) E
K {x } : f (a) = O} is a proper pri me ideal of K { x }, and even a differential
ideal (i. e. an ideal closed under deri vation D, as it is easy t o check) .
Incidentally define a differenti ally algebraic over K if [(al K ) =1= 0 (so if
f (a) = 0 for some nonzero differenti al polynomial f( x) E K{ x} ), and dif-
ferentially t ranscendent al otherwise.
When a is differenti ally transcendental over K , t here is nothing t o add. For,
Fa is a differential ring isomor phism between K{ x} and a suitable differential
subring of K containing K and a . It is straightforward t o deduce that the
differential fi eld K(a) we are looking for is isomorphic to t he field of quotients
of K{ x} (which inheri t s a natural st ruct ure of differenti al field regul ated by
t he usual derivation rul es for quo ti ents) .
But t he most cr itical point of our analysis concerns a differentially algebraic
element a over K. In fact , on t he one hand it is easy to realize t hat even
in t his case t he quoti ent ring K{x}I [ (a]K ) gets a natural struct ure of dif-
ferenti al domai n, j ust because [ (a]J() is a prime differenti al ring and so it
makes sense to put , for f (x) E J({x },
D(J (x) +[ (al K )) = Df (x) +[ (aIK);
t he field of quotients of K{x} I [ (o]K ) inheri t s in its tu rn an obvious struc-
t ure of differenti al field, and K(a) is isomorphic to t his field. This is qui t e
general, and formally satisfact ory. But we aim at underst anding t he internal
struct ure of K(a), and t his clearly requires t o st udy t he prime differenti al
ideal [( al J() and to provide an intrinsic description of its polynomials.
First let us observe t hat every prime differenti al ideal 1 =1= {O} , K {x} of
K{x} can be represent ed as I = [(aI K) for some suitable 1 and a E II
(differentially algebraic over K) . This is st raight forward and quite gener al.
In fact , look at the differenti al domain K{ x}I L, form it s differential field of
quoti ent s Q, noti ce t hat Q extends K provided one identifies each k E K
and it s class I +k in Q, and t hat a = I +x just annihilates t he differenti al
polynomials in I over K.
So take any nont rivial pri me differenti al ideal I of K{x}. Choose a nonzero
polynomial f (x) E I having
* a minimal ord er n,
then
6.5. DCPo REVISITED
* a minimal degr ee in D" (x),
and finally
* a minimal total degree.
211
As I is prime, f( x) is irreducible as an algebraic polynomial in ,'C[Dm(x) : m
natural]. We wonder whether I equals by chance the differential ideal (J(x))
gener at ed by f( x) , in other word s t he algebraic ideal of ,'C[Dm(x) : m
natu ral] generated by f(x) and it s derivatives. This is not true in general,
and I is somewhat more complicated . But cla rifying this point requires a
further notion.
Definition 6.5.1 Let f(x) E K {x} have order n. The formal partial deriva-
tive of f( x) with respect to Dn(x) is called the separant of f(x) and denoted
sf (x).
For example, the separant of f( x) = (D
3
x) 2+x .D
3x+D
x .D
2x
is 2D
3
x+x.
Now, given a different ial polynomial f( x) E K{x} , look at t he set I(J( x))
of the polynomi als t(x) E K {x} such that for some natural m
SJ(x) . t(x) E (J(x) );
one sees that I(J( x)) is a differential ideal and, when f( x) is irr edu cible,
it is also prime. Of course the membership to I(J(x)) is not eas y and
immediate to check ; but one can show t hat, for an ir reducible f( x) of order
n, a polynomial g(x ) E I(J( x)) must have order 2:: n and , if it s order is just
n, then f( x) directly divides g(x). This an alysis implies that , if a is a root
of f( x) , then I(J( x)) = I(ajK).
Mo re gener ally, let us come back to our nont rivi al prime differential ideal I
and to t he polynomial f (x) E I chosen before. One proves
Theorem 6.5.2 I = I(J( x)) .
This is a noteworthy and deep algebraic fact; it s proof can be found in
t he references quoted at t he end of this chapt er. This also concludes our
preliminary outline of basic Differential Algebra. Now let us deal at last
with Model Theory. In fact , owing to what we said about I(J(x)) we are in
a posit ion to show:
Theorem 6.5.3 Existenti ally closed differential fields J( of characteristic 0
are differentially closed.
Proof. We have to check t hat K satisfies Blum's axioms: for every nonzero
polynomials f (x) , g(x) E K {x} such t hat the or der of f (x) is larger t han
the order of g(x) , t here is et is K sat isfying f( et) = 0 and g(et) i: O. As t he
existence of et can be expressed by an existe nt ial sentence wit h par ameter s
in K and K is existentially closed, it suffices to find such an element et in
some differenti al field 'H. extending K . Pi ck an irredu cible factor f' (x) of
f (x) in K{ x} of t he same order as f(x), and form I = I (J' (x )). Clearly
f (x) E I , while g(x) ca nnot belong to I because its order is less than t he
order of f '(x) . Now enlarge K to t he differenti al field of quotients H of
K{ x }/ I and not ice t hat in Il I + x sat isfies f (v) = 0 and ca nnot realize
g(v) = 0, because f( x) E I and g(x) rJ. I. ..
In order to conclude t hat DCF
o
is just th e t heory of existent ially closed
differenti al fi elds of cha racterist ic 0, we have to show t he inverse impl icatio n
(saying t hat any differenti ally closed fi eld of characteristic 0 is existent ially
closed) and definit ively to check two more point s:
(i) every differential field of characteristic 0 has a differenti ally closed
extension;
(ii) DCF
o
is model complete .
Her e ar e t heir proofs.
Theorem 6.5.4 Every differential field K of characteristic 0 has a differ-
entially closed extension.
Proof. This ca n be shown by using some familiar chain arguments. In
detail , list in some (possible t ra nsfinite) way t he pai rs (J (x), g(x)) of nonzero
pol ynomials in K {x } such t hat t he order of f (x) is larger t han t he order
of g(x). Apply t he sa me tec hnique as in Theor em 6.5.3 and, for every pa ir
(J (x), g(x)), enlarge the ground field in ord er t o include a roo t et of f (x) such
t hat g(et) i: o. Repeat t his procedure and event ually obt ain a differenti al
extension K' of K with t he following propert y: for every choice of f( x), g(x )
in K{ x} - {O} suc h t hat t he orde r of f (x) is larger t ha n t he orde r of g(x),
there is some et E K' such t hat f(et) = 0 and g(et) i: o. However t his does
not mean t hat K' is different ially closed (in fact , what happens for f (x )
and g(x) in K' {x}?). But now we can form a new sequence of di fferenti al
exte nsions K
n
of K , by put tin g Ko = K and, for every n, K
n
+
1
= (a
differ enti al fi eld enlarging K
n
and cont aining, for every choice of nonzero
polynomials f( x) , g(x ) in Kn{x} such that the orde r of f( x) is larger t han
the order of g(x) , a root et of f( x) t hat does not annihilate g(x)). tc = UK
n
6.5. DCF
o
REVISITED 213
is the domain of a differential field ext ending /C. A straightforward check
shows that f( is differentially closed . ..
Theorem 6.5.5 DCF
o
is complet e and eliminate the quantifiers in the lan-
guage L = {+, " - , 0, 1, D}. In particular DCF
o
is model complet e.
Proof. Fi rst we show t hat DCF
o
is complete, and hence th at any two
models /C, /C' of DCF
o
are elementarily equivalent . As every structure has
some t:io-sat urated elementary ext ension, we can assume that both /C and /C'
are t:io-saturated. As parti al isomorphism impli es element ary equivalence , it
suffices to show /C /C'. Accordingly let J be the set of all the isomorphisms
between finitely generated differenti al subfields of /C and /C' respect ively. J
is not empt y because both /C and /C' include the rational field (with a zero
derivative) as the minimal differential subfield generated by 0. Now let /Co
and /Cb be t wo finitely generated differential subfields of /C, /C' respectively,
corr esponding to each other by some isomorphism h. Of course, if ii is a tuple
generating K, then h(ii) = a
'
generates /Cb ' So, algebraically speaking, any
element of Ko can be obtained from ii by the usual elementary operations
and D, and t he same can be said about Kb and O!. Incidentally notice that
Ko dcl(ii) and, parallely, Kb dcl(a
l
). We have to control that h sat isfies
the back- and-forth properties (i) and (ii) . Clearly it suffices to deal wit h (i),
as (ii) can be handled in a similar way. So t ake a E K - Ko. We look for
t wo finitely generated extensions /Cl, of /Co, /Cb respectively such t hat
a E K
1
and h can be enlarged to an isomorphism between /Cl and We
distinguish two cases, according t o whether a is differentially algebraic, or
differentially transcendental over /Co.
When a is differentially algebraic, consider the nonzero prime differential
ideal I(a/K
o)
and th e (irreducible) polynomial j( x) in l(a/K
o)
having
minimal order n , then minimal degree in D": and, finally, minimal total
degree. Then j'( X) = h(J(x)) is an irr educible polynomial of order n in
/C
o{
x }. Look at t he ty pe p given by th e formulas
j'(v) = 0, -,(g'(V) = 0)
where g' (x) ranges over t he nonzero polynomials of order < n in Kb{x}. As
Kb dcl(a
'),
p can be viewed as a type over O! . As /C' is differ entially closed,
any finite portion of p is sati sfied in /C'. As /C' is t:io-satu rat ed, p it self is
realized in K
'
- Kb by a suitable element a'. Moreover I (J' (x)) = I (a
'
/ Kb)
So one can easily enlarge h to an isomorphism between /Co(a) and /Cb(a
/
)
mapping a in a ' (for , I (a/K
o
) and I (a' /Kb ) correspond to each oth er by
h). Accordingly K
i
= Ko(a) and = Kb(a' ) yield (i) .
The case when a is differentially t ranscendental over Ko is simpler. In
fact , using t he of K' , one easily finds an element a ' E K'
algebraically t ranscendental over Kb del(at) . At t his point , one obser ves
t hat Ko(a) and Kb(a') are isomorphic by a function extending h and taking
a to a' . This ensures (i) even in t his case, and eventually accomplishes t he
proof t hat DCF
o
is complet e.
Now we prove t hat DCF
o
eliminates t he quantifiers in L. Owing to com-
pleteness, every L-sent ence is DCFo-equivalent to 0 = 0 when belonging
to DCF
o
, and to 0 = 1 otherwi se. So we have to eliminate t he qu anti-
fiers in t he L-formulas <p (v) when t he sequence v of free vari abl es is not
empty. For t his purpose we arra nge t he previous completeness proof in t he
following way. Let ii, at be two t uples in t he uni verse n of DCF
o
satisfy-
ing the same quantifier free L-for mulas ; we claim t hat t hey have t he same
ty pe over 0. There is no loss of generality in assuming ii in K and at in
K' for some suitable different ially closed fields K and K'. T he
hypoth esis t hat ii and at realize t he same quantifi er free formul as easily im-
plies that ii, at generate isomorphic differenti al subrings, and consequently
isomorphic differential subfields in K, K' respectively, by ii f---7 at. We look
now at t he set of th e isomorphisms between finit ely generat ed differenti al
subfields of K containing ii and finit ely generated differenti al subfields of K'
containing at which extend ii f---7 ;{t. By proceeding as before, one obt ains
t his t ime (K, il) == (K' , ;{t) in t he language enlarging L by suit ably many
new constants to be interpreted in ii, ;{t respect ively. In other words, ii, at
have t he same ty pe over 0, as claimed. Hence, two sequences satisfying t he
same quantifier free formul as do admit t he same ty pe; in order t o concl ude ,
we have simply to realize that t his is just th e eliminat ion of quant ifiers for
L-formulas <p (v) with a non-empty v. Let us see why.
If <p (V) does not occur in any t ype over 0, t hen it is clearl y equivalent to
-, ( Vi = Vi) ' Oth er wise we have just seen that , for every type p contain-
ing <p( V) , <p (v) is a consequence of t he quant ifier free for mulas in p. By
compactness, only finitely many formulas of t his kind , and even a single
quanti fier free formul as <pp (v) -their conj unction-, are enough to impl y <p (v)
wit hin DCF
o
. Topologically speaking, when p ranges over t he ty pes over
containing <p (V), t he neighbour hoods Ucpp(v) form an open cover ing of t he
closed , hence compact set Ucp(v)' Again using compact ness, one finds a finit e
set Po of types p containing <p (v) such th at Ucp(v) equals U PEP
o
Ucpp(v) ' In
6.5. DCF
o
REVISITED
oth er words
Vv(<p( if) f-----+ V <pp (v)) E DCFa .
pEP
o
As t he lat t er formul a is quantifier free, we are don e. ..
215
Hence DCF
o
is t he mod el compa nion of t he t heory of differential fields
of cha racterist ic O. But t he pr evious analysis says mor e and in particular
ensures:
T heorem 6.5.6 DCFa is eo -sta ble.
Proof. What we have to check is t hat , for every differenti ally closed field /C
in characterist ic 0, IS1(K) 1= IKI. Actually we will show somet hing more,
resembling what we did for algebraically closed fields. In fact, we will see
that , for every /C, there is a natural bijection between l-types over K in
DCFa and (proper) prime differential ideals in /C{x}: owing to what we saw
about t hese ideals and t heir st ruct ure, this impli es 15
1(1<)1
IK{ x}1 = IKI ,
whence 15
1
(K)1 = IKI.
First take a l-t ype p over K and form
I (p) = {j (x) E K{ x} : j (v) =0 E p}.
Checking t hat I (p) is a prime different ial ideal in /C {x} is a straightforwa rd
exe rcise. Indeed every (proper) prime differential ideal I of /C{x} can be
obtai ned in this way: in fact , build t he (different ial) field of quoti ents Q
of /C{x} / I , enlarge Q t o a differenti ally closed K and, at last , take t he
ty pe of I +x over (t he isomorphic copy of ) /C (in K): clearl y a polynomi al
j (x) E K{ x} sat isfies j (v) = 0 E Pif and only if j (x) E I . Furth ermore t wo
different l-types p =1= q over /C define different ideals. In fact t here is some
L(K)-formula <p (v) in p and not in q. Now use elimination of quantifiers:
<p (v) can be chosen as a disjunction of conj unct ions -and even as a unique
conj unction- of equations j(v) = 0 or negations ---,(g( v) = 0) with f(x) and
g(x) in K {x}. But t hen any single equation and disequation is in p, and at
least one of them is not in q. Consequent ly I(p) =1= I(q) . ..
As a corollary, let us point t hat t he constant subfield C(/C) = {a E K : DO' =
O} of a differenti ally closed /C, as a struct ure definabl e in /C, is w-stable, and
so algebraically closed (due t o Macinty re' s Theor em).
Incident ally, we should also say t hat DCF
o
is not strongly minimal ; indeed
it s Morley rank is wand it s degr ee is 1. However , owing to t he t heorems
of Morley and Shelah on t he existence and uniqu eness of prime models over
subsets in w-stable t heories, we can deduce:
216
CHAPTER 6. w -STABILITY
Corollary 6.5.7 Over any differential field IC of characteristic 0 there is a
prime model of DCP
o
(hence a differential closure), and this is unique up
to isomorphism fixing K pointwise.
We already emphasized several times that this approach -using Model The-
ory, in particular w-stability, Morley's Theorem and Shelah's Theorem- is
the very first proof, and virtually the only one known till now, of the exis-
tence and uniqueness of differential closures.
Again using w-stability we show now the last result of this section.
Theorem 6.5.8 DCF
o
uniformly eliminates the imaginaries.
Proof. We repeat almost wholly the approach followed in 6.3 for alge-
braically closed fields. Indeed the first two algebraic preliminary steps in
that proof concern arbitrary fields and ideals, so apply to our present set-
ting as well. The same can be said about the use of w-stability, owing to
Theorem 6.5.6, although now types correspond to (proper) prime differential
ideals, and not directly to prime ideals. In conclusion what we have to check
is that the third preliminary step in the proof of Theorem 6.3.2 still works in
the new framework. Accordingly t ake a differential field IC of characteristic
0, and 1
0
, , I
m
prime differential ideals of IC{ x} in the same conjugacy
class (with respect to differential field automorphisms of IC) . We look for
a(n even algebraic) subfield ICo of IC such that, for every automorphism (J
of IC as a differential fi eld,
(J permutes 1
0
, .. , I
m
if and only if (J fixes Ko pointwise.
This can be obtained by rearranging the approach in the field case. Form
I = nj<m Ij. This is a differential ideal, and even a radical ideal in the usual
sense: for every f(x) E K {x}, if ft(x) E I for some positive integer t, then
f(x) itself is in I. A differential version of the Basis Theorem, due to Ritt
and Raudenbush and working over differential fields of characteristic 0, says
that every differential ideal of IC{x} -like I, but also 1
0
, , I
m
- is finitely
generated (as a differential ideal). Fix a finite set of differential polynomials
including generators of I, 1
0
, ... , I
m
. Let]-f be their maximal order, so
all these polynomials can be viewed as algebraic polynomials over K in the
unknowns tr, for h ::; n, 1 ::; i ::; m. Put t: = IC[Dhxi : h <u, 1 ::; i ::;
m], and look at In R, 1
0
n R, ... , i; n R. Then In R = nj<m(Ij n R),
and 1
0
n R, ..., I
m
n R are prime ideals in n in the same conjugacy class
with respect to field automorphisms of IC: in fact , for i , j' ::; m, there is
some automorphism of IC as a differential field, hence as a field, taking Ij to
6.6. RYLL-NARDZEWSKI'S THEOREM, AND OTHER THINGS 217
Ij , and so t, n R t o Ij n R. Use St ep 3 in Theorem 6.3.2 and dedu ce that
t here is an algebraic subfield Ko of K such that, for every aut omor phism 0-
of K, viewed as a field and in particular as a differenti al field, 0- permut es
t he Ij nR's if and only if it fixes Ko pointwise. But each Ij nR generates Ij
as a differential ideal in K{ x} , so, for i. j' ::; m, o-(Ij n R) = Ij n R impli es
o-(Ij) = Ij . Consequentl y, for every aut omorphism 0- of the differenti al field
K, 0- permutes 1
0
, .. , I
m
if and onl y if 0- act s identically on Ko. ..
6.6 Ryll-Nardzewski's Theorem, and other things
We conclude our outline of Morley's ideas in t hese two cha pters with a
perhaps oblique and superficially related argument. In fact we want t o
treat t heories. As already underlined, they behave qui t e au-
tonomo usly and incl ude examples which are not categorical in any uncount-
able cardinal, and neither are w-stable: for instance, t hink of DLO-. On
t he other hand, there do exist uncountably categorical struct ures which are
not (such as ACF
p
for any fixed p). However
theories have their peculiar and specific propert ies. In part icular t hey enjoy
t he following nice characterization, due t o Ryll-Nardzewki and others.
Theorem 6.6.1 (Ryll-Nardzewski) Let T be a complet e theory in a count-
able language L. Then the following propositions are equivalent:
(i) T is
(ii) for every positive integer n, there are only fini tely many n-types over
oin T ;
(iii) for every positive integer n, there are only fin itely many formulas
in n f ree vari ables pairwise inequivalent in T.
Proof. The equi valence between (ii) and (iii) is clear: it can be deduced
directly, or using the classical du ality bet ween Boolean spaces and Boolean
algebras (recall t hat, for every n, t he n-t ypes over 0 form the du al space
Sn(0) of t he Boolean algebra of 0-definable subsets of o n).
So let us compare (i) and (ii). First suppose Sn(0) infinite for some n.
As Sn(0) is cornpact, it contains some non-isolat ed type p. The Omit t ing
Types Theorem ens ur es that p is omitted in some countable model of T;
on the other hand, there does exist some countable model of T realizing
p , The latter mod el cannot be isomorphic t o t he former; hence T is not
Now ass ume Sn(0) finit e for every n. Consequently Sn (0) is
a discrete space, and so every n-type over 0 is isolated. As t his is true for
every n , every model of T is atomic over 0, hence every count able model of T
is prime over 0. Then all the countabl e models of Tare pairwise isomorphic.
..
Let us underline a simple conseque nce showing in the cou ntable case t he
more gene ral fact t hat , for a A-cat egorical T, t he only model of power A is
sat urated (in t he uncountable case, this res ult is much mo re complicated
and refers to Morley's Categoricity Theorem) .
Corollary 6.6.2 Let T be an theory. Then the only count able
model M of T is
Proof. Let abe a tuple (of length m) in M . For every positive integer
n , there are onl y finit ely many n-types over a, ind eed onl y finit ely many
(n + m)-t ypes over 0. So every n-type over ais isolated, and consequently
is realized in M. ..
Now let us propose some examples illustrating Ryll- Na rdzewski's Theorem.
Examples 6.6.3 1. We know that dens e linear orders (with or ) without
endpoints have an t heory (see Chapt er 1) . Let us con-
firm t his result via t he Ryll-Nardzewski Theorem. First take a positive
int eger m and ao, ... , am, bo, . . . , b
m
in the universe of DLO-, with
ao < ... < am, b
o
< .. . < b
m
. Then t here is an automorphism of n
mapping aj in bj for every j m. In fact t he intervals in n
] - 00, ao[, ] - 00, b
o[,
]aj , aj+d, ]bj , bj+d Vj < m,
Jam, +00[' ]b
m,
+oo[
are models of DLO- and ar e saturated in the same power as n. Conse-
quent ly t hey are isomorphic to each other. In part icular (ao, . .. , am),
(b
o
, . . . , b
m
) have the same t ype over 0. This implies t hat, for ev-
er y positi ve int eger n and sequence (Cl, . .. , C
n
) E nn, the type of
(Cl , . . . , c
n
) over t he empty set is fully determined by t he isomorphism
ty pe of t he ordered set ({Cl, ... , c
n
} , So we get onl y finit ely many
n-types over 0for every n , and the Ryll- Nardzewski Theorem applies
to confir m t hat DLO- is
6.6. RYLL-NARDZEWSKI'S THEOREM, AND OTHER THINGS 219
2. On the contrary, ACFp is not for any p = 0 or prime.
Let us see why by using the Ryll-Nardzewski Theorem. Is suffices
to notice that there are infinitely many pairwise distinct 0-definable
subsets in the universe n of ACFp, and even that there are infinitely
many pairwise distinct finite 0-definable subsets of n, in other words
that ad(0) is infinite. But this is well known because ad(0) equals the
(field theoretic) algebraic closure of the prime subfield of n.
Of course, no (expansion of a) field of characteristic 0 can admit an
categorical theory, for the same reason as in Example 6.6.3, 2. This applies
in particular to real closed fields. Checking what happens for differentially
closed fields, or separably closed fields, or ACFA, could be a useful exercise.
We conclude this section and the whole chapter by discussing a related
matter, again concerning definable sets and types. In particular we wonder
which information we can obtain about a complete theory T by looking at
the isomorphism types of the Boolean algebras of definable sets of its models.
Can the knowledge of these algebras say anything essential about the model
theoretic complexity of T? Incidentally, recall that, at least in the countable
case, there does exist a satisfactory classification of Boolean algebras up to
isomorphism, provided by Ketonen.
In this framework, A-categoricity can be replaced, for every infinite cardinal
A, by a seemingly weaker notion, called Boolean A-categoricity. A complete
theory T is said to be Booleanly A-categorical if and only if all its models of
power A have isomorphic Boolean algebras of definable 1-ary sets. Clearly
categoricity implies Boolean categoricity in every A. Notably
Theorem 6.6.4 (Mangani-Marcja) For an uncountable A, a complete the-
ory T is Boolean A-categorical if and only if is A-categorical.
But this is false when A= Indeed
Theorem 6.6.5 (Marcja-Toffalori) Every -categorical theory is Boolean
-categorical.
So algebraically closed fields have Boolean theories. One can
see that the same is true also for real closed and differentially closed fields,
as well as for every module.
Other connections between isomorphism types of Boolean algebras of defin-
able sets and structural properties of theories are investigated in the papers
quoted at the end of the chapter.
6.7 References
w-stable t heories and struct ures are deal t with in [8]. The det ails about
Morley rank and degree in st rongly minimal t heories ca n be found in [108],
or in the Ziegler contribution to [16]. [132] (and it s recent english transla-
tion) provide a nice and clear treat ment of w-stable groups; also [15] is a rich
source on this subject . Macintyre's Th eor em on w-stable fields is in [90]; by
t he way, let us mention it s generalization t o superstable fields in [25] saying
t hat any superstable field is algebraically closed (superstability is a notion
enlargi ng w-stability, and will be int roduced in t he next chapter) .
Prime mod els over t he empty set were st udied by Vaught [174]; t he anal-
ysis over arbitrary sets was developed in [118]. Ressayre's contributions to
constructible sets, and in particular the Uniqueness Theorem, were never
publi shed, while Shelah 's Uniqueness Theor em is in [148].
As already said, t he t heory of differentially closed fields of characteristic 0
was developed in Blum 's Ph .D. Thesis [12] (see also [146]). Good references
are also [110] or [131].
Ryll-Nardzewski 's Theorem is in [145]; see also [43] and [156].
The classification of count abl e Boolean algebras is in [69] . Boolean A-
categoricity was introduced in [104] and int ensively st udied in [105] and
[165]; actually, Boolean categoricity was called pseudo categoricity in t hose
papers. The st udy of t he st ruct ural properti es of t heories by t he isomor-
phism t ypes of Boolean algebras of definabl e sets of their models is also
pursued in [106] and [107].
Chapter 7
Classifying
7.1 Shelah's Classification Theory
A cent ral subject in Mathematics is classifying, t hat is characterizing t he
objects of a given class up to equivalence relat ions, for instance t he struct ures
of a given language, or t he models of a given theor y up to isomorphism. Let
us mention some classical examples.
Examples 7.1.1 1. An infinite set (without additional struct ure) is fully
characterized up t o isomorphism by its power, so by a (n infinite) car-
dinal numb er.
2. Let J( be a fixed (cou nt abl e) field. A nonzero vectorspace over J( is
completely determined up to isomorphism by its di mension over J( , so
again by a cardinal number.
3. An algebraically closed field of a given char acteristic is fully det ermined
up to isomorphism by its transcendence degree over its pr ime subfield,
so, once agai n, by a ca rdinal number.
Notice t hat t he pre vious examples concern element ary classes and even
strongly mini mal th eories (provided we restrict our attention to infinite
vectors paces in 7.1. 1, 2). But, of course, the classification purposes
touch larger horizons.
4. It is well known that every finitely gener at ed abelian group A decom-
poses as a finit e direct sum of cyclic grou ps, and even of copi es of Z,
Z/pn where p ranges over pr imes and n over positi ve int eger s. More-
over t he lat t er decomposit ion is unique up to isomorphism (and up
221
222 CHAPTER 7. CLASSIFYING
to permuting the summands); so the isomorphism type of A is de-
termined by saying how many copies of Z, Z/pn (p prime, n positive
integer) are involved in this decomposition of A. Notice that finitely
generated abelian groups are not an elementary class (why?).
5. Consider torsionfree abelian groups of rank at most n, where n is a
given positive integer: they are the subgroups of the additive group
(Qn, +). Baer found in the thirties a nice classification of these groups
up to isomorphism when n = 1. However it has been a longstanding
open question to accomplish a general satisfactory classification for
every n 2: 2, and actually no classification is presently known. We will
go deep in this question later.
6. Let K be a given field, and look at the ring K(x, y) of polynomials
with coefficients in J( and two non-commutings unknowns x and y.
Consider K(x, y)-modules finite dimensional over K towards a possi-
ble characterization of their isomorphism classes. Well, this problem
is generally believed intractable. In fact, finite dimensional K(x, y)-
modules interpret the word problem of groups, and classifying them
as said means solving this problem.
Of course, classification problems do not restrain themselves to the iso-
morphism relation; for instance one can try to classify square matrices by
similarity, and so on. Also, it should be underlined that a classification is
sometimes (often?) very unlikely to be obtained (as Examples 7.1.1, 5 and
7.1.1,6 witness). However looking for a classification, and even realizing its
unfeasibility, can disclose new connections between different areas, generate
new techniques and, definitively, enlight new horizons within Mathematics.
From an abstract point of view, a natural problem arises in this framework,
that is developing a theoretical treatment to recognize which classes and
relations admit a classification, and to propose general tools to accomplish
this classification when possible, or to measure its difficulty in the hard cases.
A closely related and preliminary question is just: what does classifying
mean?
These issues are intensively studied within Mathematical Logic. For in-
stance, Descriptive Set Theory (and consequently, through it, both Set The-
ory and Recursion Theory) are concerned with the classification matter.
In fact Descriptive Set Theory deals with definable sets of R, viewed as a
separable complete metric space, and of similar topological spaces (Polish
spaces). This does overlap classification. Indeed the objects that are to
classified often belong to a Polish space X , or at least to a Borel subset of
7.1. SHELAH'S CLASSIFICATION THEORY 223
X , and t he classifying equivalence relation E is a Borel, or analyt ic, etc .,
subset of X
2
: in fact, t his is the case of the similarity relat ion for square ma-
trices, as well as the isomorphism relation for den umerable struct ures (such
as groups and fields) in a given language. In this setting it is sometimes
possible t o associat e in a Borelian way to any object of X a point (called
"invariant") in another Polish space such that two different objects have the
same associated point if and only if they are equivalent in E. The relation E
is called smooth when it satisfies these requirements. For example, Jordan 's
canonical form is an invariant of the similarity relation for mat rices, which is
therefore smooth. In Ergodic Theory, smooth relations are just t hose which
are considered act ually classifiable. But not all equivalence relati ons one
meets are smooth. In the latest years a general theory of classification of
equivalence relations which live in Polish spaces has been developed within
Descriptive Set Theory: it allows to est ablish, for E and F equivalence
relations, when classifying modulo E is more complex than classifying mod-
ulo F. In particular it includes some recent and beautiful results of Simon
Thomas (and others) on isomorphism for torsionfree abelian groups of ra nk
n ; these works show that t his relation is of increasing complexity when n
grows, and is not smoot h (and so is likely to be intractable) when n 2.
But now let us come back to Model Theory. Of course, also Model The-
ory is interested in classifying (structures, definable sets, and so on). But
its approach is peculiar, and differs from the perspectives and the aims of
Descriptive Set Theory: and indeed, very roughly speaking, one could say
that Model Theory and Descriptive Set Theory are complementary in this
framework. In fact, one could agree that the latter is mainly concerned in
measuring how ha rd and difficult classifying is in the worst "wild" cases ,
while Model Theory aims at determining t he "tame" settings in Mathe-
matics, and accordingly at clarifying where and why a classification can be
done.
Classification in Model Theory promptly recalls Shelah. In fact, it was She-
lah who approached, treated and essentially solved the classification problem
in Model Theory since the end of the sixties until the early eighties. So She-
lah's formidable work dates back to almost twenty years ago, but is far from
having exhausted its st imulus, as we will explain later in more det ail. The
Shelah strategy was to det ermine a series of successive key properties (sim-
plicity, stability, superstability and so on) concerning complete theories and
having a twofold role: in fact , each of them allows a new significant step
towards classifiability, while its negation excludes any hope of classification.
At the end of these successive dichotomies Shelah exactly det ermi nes which
abstract conditions ensure class ifiability.
The aim ot this chapter is to outline Shelah's classification program, its con-
cept s, tools and t echniques (forking, stabilit y, superstabilit y and so on) , as
well as it s recent development s regarding simplicity and spect rum problem.
We will lay emphasis on ideas and motivations rather than on proofs. So
we will sacrifice full details in general. However we will provide a complete
report in t he w-st able case .
The first point to be clarifi ed is: what do we want to classify? Here th e
answer is quick and ready: we will deal with classes of struct ures . Just to
avoid too many complicat ions, we will limit our analysis here to element ary
classes, and even to classes Mod(T) of models of count able complet e first
order theories T. We illustrated this choice in Chapt er 1. As usual, we will
work inside a big saturated model n of T.
The second preliminary question is: with respect to which relation will we
classify? Also in t his case the reply is fast and easy: we will classify up to
isomorphi sm.
But now we have to answer a more delicate and fundament al quest ion, th at
is: what does it mean to classify the models of a complete T up to iso-
morphism? By which "invariants" will we accomplish our classificati on pr o-
gram?
When T is st rongly minimal - just as in t he Examples 7.1.1, 1-3 quoted
before - classifying means assigning to every model a cardinal number - it s
dimension - in such a way t hat two models are isomorphic if and onl y if t hey
are given th e same ca rdinal. But we already saw that, for some theories,
a classification cannot be accomplished by assigning a single cardinal and
often requires somet hing mor e complicated, like pairs of cardinals, or even
(possibly infinite) sequences of cardinals, and so on. This is what happens in
Exampl e 7. 1.1, 4, where act ually the involved class is not elementary; ot her
elementary cases were discussed in Chapter 5. Here is anot her exa mple.
Example 7.1.2 Consider the theory of the structures (A, Eh E
2
) where
El, E
2
are equivalence relations, E
2
El and
every E
2-cla
ss has infinitely many element s,
every El -class contains infinitely many E
2-cla
sses;
t here are infinitely many E l-classes.
T is complet e (for instance because it is categorical in In a mod el
of T any El -class of power is determined up to isomorphism by the
function mappi ng any infinite cardinal ::; to the number of t he E
2
-
subclasses of power and hence by an ordered sequence of a cardinals
7.1. SHELAH'S CLASSIFICATI ON THEORY 225
Con sequently, if A = (A, E t , E
2
) is a model of T of power A, th en
the isomorphism type of A is completely given by the function associating
to any ordered sequence of (11 cardinals A) t he number of th e
Et -classes in A corresponding to this sequence.
T his construction can be it erat ed to produce more and more complicated ex-
amples, needing more and more sophist icat ed isomorphism invariants. This
suggests t he following definition.
Definition 7.1.3 Let C denote the class of cardinal numbers. For every
ordinal (11, we define a class C" by induction on (11 in the f ollowing way.
CO is C,
if (11 = {3 +1 is a successor ordinal, then we put C" = C
f3
U P (C
f3
) ;
if (11 is limit, then we put C" = Uf3 <O' C
f3
Definition 7.1.4 Let T be a complete theory, (11 be an ordinal. T is said

to have an invariant system of rank (11 if and only if there is a funct ion f
associat ing eve ry model of T with an eleme nt of CA such that two models A
and B of T are isomorphic if and only if f(A) = f( B) .
Examples 7.1.5
of rank O.
1. A st rongly minimal theory has an invariant syst em
2. An ordered sequence of cardinals is an element of C
2
So th e theories
in Example 5.8.2 have an invariant system of rank 2.
3. The t heory in Example 7.1.2 has an invariant syst em of rank 4.
Shelah's proposal is to assume t hat a complete theory T is classifiable if and
only if T has an invariant system of ra nk (11 for some ordinal (11.
We will discuss Shelah 's point of view lat er in this cha pter, after describing
Shelah' s classification th eory. In particular, we will see that th ere are some
good reasons to agree wit h it . Once t his is don e, we can deduce:
Theorem 7.1.6 Let T be a complet e theory such that, for every uncountable
power A, T has 2'\ pairwise non isomorphi c models of cardinality A. Th en
T is not classifiable.
In fact some cardinal and ordinal computations exclude that such a T has
an invariant system of any possible ra nk. This is quite remarkable, because
there are some very familiar t heories sat isfying the ass umptions of Theor em
7.1.6. For instance, this is the case of the theory DLO- of dense linear or-
ders without endpoints: in fact DLO- is and so has only one
countable model (up to isomorphism) but it gets 2'\ pairwise non isomorphic
models in any uncountable power >.. Accordingly DLO- is not classifiable.
Mo re generally Theorem 7.1.6 applies, as we will see in the next section, to
any complete theory of linearly ordered infinite structures. None of these
theories is classifiable. In particular, no o-minimal theory is classifiable, al-
though any such theory satisfies the conditions (Dl)-(D4) of the dependence
relation
a -< A a E acl(A)
for a E n, A a small subset of n. This is a little surprising and will be
discussed in more detail later in this chapter, and then in Chapter 9.
Anyhow, if we (momentarily) agree with Shelah's definition of classifiable
theory, then we have to take note that too many models (in other words 2'\
non-isomorphic models in every uncountable >.) exclude classifiability. On
the other side, we would like also to determine the key criteria ensuring the
classifiability of an arbitrary complete theory: this will be the matter of
the forthcoming sections. To conclude the present one, let us spend some
more words about the close relationship between classifying theories T (in
the Shelah sense) and counting the models of T. This is already explicit
in Theorem 7.1.6. More generally, for every countable complete T, one can
define a function I(T, . ..) associating to every infinite cardinal >.
I(T, >.) = number of the isomorphim types of models of T of power >..
xf---7 I(T, >') is called the spectrum function ofT. Recall that 1 ::; I(T, >.) ::;
2'\ for every X, owing to the Lowenheim-Skolem Theorem and cardinal com-
putations. The content of Theorem 7.1.6 is just that I(T, >.) = 2'\ for every
uncountable >. excludes the classifiability of T. Indeed there was a conjec-
ture of Morley, preceding Shelah's work and, in some sense, originating it,
saying:
Conjeet ure 7.1. 7 (Morley) Let T be a countable complete first order the-
ory. Then the spectrum function>. f---7 I(T, >.) is increasing among uncount-
able cardinals: for < x< f-t, I(T, >.) < I(T, f-t).
Shelah positively answered this conjecture, as a non-minor consequence of
his classification analysis. The problem of determining all the possible spec-
trum functions>' f---7 I(T, >.) when T ranges over countable complete the-
ories (and>. is uncountable) was solved only in 2000 by Hart , Hrushovki
7.2. SIMPLE THEORIES 227
and Laskowski, who explicitly listed all these functions. Notably their proof
involves some arguments from Descriptive Set Theory.
Finally, what can we say when A = As we observed in th e last chapter
when talki ng about this case is a little oblique with respect
to the general analysis and requi res peculi ar approaches and techniques. A
classical conject ure in this framework was raised by Vaught .
Conjecture 7.1. 8 (Vaught) For a countable complete first order theory T ,
either I(T, ::; or I(T, = 2
No
(apart from the Continuum Hypoth-
esis, of course) .
As in the case of Morl ey' s Conjecture, this question is not only a mer e cardi-
nal investigation; what is more relevant is to underst and the structure of the
countable models of T. Shelah (together with Harrington and Makkai) pos-
itively answered Vaught's Conjecture for w-stabl e theories T . Other partial
posit ive answers are known. Indeed in t he lat est months the news of a coun-
terexample (a theory with exactly count able models) due to R. Knight
[75] has been spreading, but this negative solut ion st ill seems (october 2002)
under exami nation.
7. 2 Simple theories
All throughout this section T is a complete first order th eory in a coun table
language L, T has no finite models and n denotes t he universe of T.
We aim at determining which key properties make T classifiable. A classifi-
cation is very easy in the strongly minimal case . In fact, when T is strongly
mini mal , every mod el of T is labelled by a cardinal number - its di mens ion
- classifying it up to isomorphism; what rules this dimension and its as-
signme nt is a notion of dependence, based on the model theoretic algebraic
closure ad. Unfortunately the ad dependence does not work any more when
we enlarge our setting and we leave the strongly minimal framework.
So we need a more general notion of (in)dependence, still including the
classical cases of linear independence in vectorspaces, algebraic independence
in algebraically closed fields and, definitively, ad independence in strongly
minimal theories, but applying to a wider context. In other words we ai m
at defining for any T
ii is independent from B over A
where ii is a tuple in n and A B are small subsets of n. As not ions require
abst ract symbols to be presented let us denote by I (I for independence)
the set of all the triples (ii, B, A) in n such t hat
228
ais independ ent from B over A
CHAPTER 7. CLASSI FYING
in a sense t o be made more preci se. It is reasonable t o expect t hat 1 satisfies
the following proper ties:
(11) (invariance) for ever y (a,B, A) E 1 and aut omorphism f of n,
(1(a), f(B) , f(A)) is still in 1;
(12) (local character) for every aand B , th ere is a count able A B such
t hat (a, E , A) E 1;
(13) (fini t e character) for every a, A and B , (a,B , A) E 1 if and onl y if, for
all finit e tuples bin B, (a,AUb,A) E I;
(14) (ex tension) for every a, A and B, t here is a t uple a' having t he same
length and t he same ty pe over A as asuc h t hat (a', B , A) E 1;
(15) (symmetry) for every a, b and A, (a, A U b,A) E 1 if and onl y if
(b, A U a, A) E 1;
(16) (transitivity) for ever y a and A B C, (a,C, A) E 1 if and onl y if
(a, B, A) E 1 and (a,C, B) E 1.
Definition 7.2.1 A set 1 of triples (a, B, A) with a in n and A B small
subset ofn is called an independence system ofT if and only if I satisfies
(ll) - (16).
An eas y applicat ion of (13) and (15) shows t hat, if B , B' :::> A are small
subsets of n, t hen
(b,B' , A) E 1 VbE B
if and only if
(b
'
, B, A) E 1 V;} E B'.
We will say t hat B and B' are independ ent over A when t his happens.
Notice also t hat, if a' is a subsequence of a and (a, B, A) E 1, t hen
(a' , B, A) E 1 as well. By (13), it suffices t o check (a' , A Ub, A) E 1 for all
bin B. We know (a, A U i, A) E 1. By symmet ry (15), (b, A U a, A) E 1,
whence (b, AUa', A) E I by (13), and (a' , AUb, A) E 1 by (15) once again.
Let us propose some exa mples of independ ence systems, both t o illustrate
t he meaning of (11) - (16) and to confirm t hat t hey provide a reason abl e
axiomatic ground t o introduce an abst ract notion of independ ence.
Examples 7.2.2 1. First let us check that the old independence notion
in strongly minimal theories T corresponds naturally to an indendence
system in the new sense. In fact, let T be any theory. For a E nand
A <;;:; B small subsets of n put
(a, B, A) E I if and only if a E acl(B) implies a E acl(A)
or, equivalently but more transparently,
(a, B, A) rf. I if and only if a E acl(B) - acl(A).
We claim that, if the acl dependence relation in T satisfies (Dl), (D2)
and (D4) (in particular, if T is strongly minimal, or also o-rninimal),
then the triples (a, B , A) in I do satisfy (11) - (16). We underline that
we are momentarily dealing with elements a rather than with tuples a
in n.
(11) is clear.
(12) follows from (D2) , which says even more; in fact, it ensures that,
for every a and B, if a E acl(B), then there is a finite A <;;:; B
such that a E acl(A).
(13) says in our particular setting that, for every a and A <;;:; B with
a rf. acl(A), a E acl(B) if and only if there exists a tuple bin
B such that a E acl(A U b). So the direction from left to right
just follows from the definition of acl, and the converse is a direct
consequence of (D2).
(14) Let (a, B, A) rf. L, so a E acl(B) - acl(A). tp(a/A) is not alge-
braic, and so is realized even out of acl(B). Take any a' rf. acl(B)
such that tp(a'/A) = tp(a/A) and notice that (a', B, A) E I,
(15) when restricted to elements a and b in n is just the Exchange
Principle (D4).
(16) is clear.
It is straightforward to see that the previous analysis extends to arbi-
trary tuples ain n, provided we put , for a= (al' ... , an),
(a, B, A) E [
if and only if, for every i n,
(ai, B U {aa, ... , ai-d, A U {aa, ... , ai-d) E I,
230 CHAPTER 7. CLASSI FYING
The resulting I is an independence syste m of T. As al ready said , this
exa mple includ es both strongly minimal and o-rninimal theories. In
part icular it applies to t he following classes of struct ures .
(a) (Pure) infinite sets: here , for a, A, E as above, (a, E , A) et I (so
a dependent from E over A) means a E E - A.
(b) Vectorspaces over a given (countabl e) field K: in this case,
(a, E , A) et I means that a is in the subspace spanned by E
but is not in t he subspace spanned by A ; hence I recovers linear
independence.
(c) Algebraicall y closed fields: this time, for A and E subfields,
(a, E , A) et I means that a is algebraic over E but is transcen-
dental over A ; accordingly I recovers algebraic independence in
t his case.
(d) Real closed fields (such as t he ordered field of reals): recall t hat
these fi elds are o-mini mal ; moreover, for ordered subfields A and
E, t he model theoretic algebraic closure acl equals t he real closure
in t he usual algebraic sense; accordingly, for a, A and E as before,
(a, E, A ) et I means that a is in the real closure of E but is not
in the real closure of A.
But independence systems do go beyond the st rongly minimal and 0-
minimal set ti ngs, and apply to larger sceneries, for instance to w-stable
t heor ies, and other theories as well.
2. Let T be w-stable. Define I by putting, for ii, A and E as before ,
(ii, E, A) E I RM(tp(ii, E)) = RM(tp(ii, A)).
Then I is an independence system of T . The details will be provided
in Section 7.5 below. We give here just a few comments. Firstly recall
that RM(tp(ii, E)) ::; RM(tp(ii, A)) for every s and A E. Secondly
not ice t hat, for a strongly minimal T and for a E n, RM(tp(a, E)) <
RM(tp(a, A)) j ust means a E acl(E) - acl(A) . In other words, when
we rest rict our attention to st rongly mini mal theori es, we recover t he
independence syst em in 1. But , of course, independence concerns now
a much larger framework, including, for instance, differentially closed
fields of charact erist ic 0 (the differential case will be discussed in more
detail in Section 7.10) .
3. Now consider random graphs. They are (infinite ) graphs (C, R) (with
R a symmetric irreflexive binary relation on C) satisfying t he following
addit ional condit ion:
(*) for every m, n E N and aa, . .. , an, ba, ... , b
m
E C , t here is some
g E C satisfying R( g, ad for all i ::; n and -,R( g, bj ) for all j ::; m.
Clearly (*) ca n be exp ressed by infinitely many first order sente nces
in t he language L = {R} , so random gra phs are an elementary class.
Their t heory T is w-categorical and consequentl y complete ; fur th er-
mor e it elimina tes t he quantifiers in L. Notice t hat T is not w-stabl e.
In fact, given a random graph (C, R) , for any subset 5 of C, th e
formulas
R( v, s) Vs E 5, -,R(v,g) vs E C - 5
defines a I- t ype over C, and different subset s produce different ty pes;
so 5t{C) contains at least 2
1G1
elements. This implies t hat T is not
w-stable.
Now define, as in t he case of infinit e sets, for a in n and A B,
(a, B , A) E fi f and onl y if a E B - A.
It is easy to check t hat one determines in t his way an independence
syst em f.
4. Any cornpleti on of t he t heory ACFA of existent ially closed fi elds wit h
an aut omorphism has an independence system: thi s will be described
in mor e detail in t he final section of this chapter , among other algebraic
exa mples.
However t heories having an independence syst em (such as t hose considered
in th e previous examples) do not behave in the same way. For instance, one
realizes that th e following amalgamat ion property is a crucial dividing line:
(17) (amalgamation) let A be a model ofT, B , B' ;2 A be independent over
A, b, fi be tuples in n having the same ty pe p over A and satisfying
(b, B,A) E I , (fi , B' ,A) E frespecti vely; t hen t here is some c in n
realizing t he same ty pe as bover B and as fi over B' , and satisfying
(c, B U B' , A) E J.
Definition 7.2.3 An independence system I of T is good if and onl y if it
satisfi es (17).
232 CHA PTER 7. CLASSIFYING
Let us run t hrough t he previous examples t o illustrat e (17) and it s relevance.
Examples 7.2 .4 1. Let T be st rongly minimal , I be defined as before.
Then (17) holds. Let us check t his when dealing with t uples of length
1, so for element s of n. If b or b' is in A , t hen b = b' and c = b = b'
works. So ass ume b, b' t/. A; independence impli es b t/. acl(B) and
b' t/. acl(B') , in other words b, b' realize the onl y non-algebraic I-t ype
over B , B' respectively; so any realization of t he unique non-algebraic
I-type over B UB' works as c. Not ice that , in this case, (17) is a direct
consequence of th e fact th at, for every small subset X of n, there is
a uniq ue non-algebraic l -type px over X and consequent ly over every
small Y ;;2 X there is a unique non-algebraic l -fype ext ending ux .
However (17) fails within o-rninimal theories: / is not good in this
set t ing. Let us see why in t he particul ar case when T = RCF is th e
theor y of real closed fields. Consider th e ordered fi eld R of reals and ,
in a suitably sat urated exte nsion n of R , 4 positive element s
b < a < a' < b'
each infinit ely larger t ha n t he previous ones, and b infinitely lar ger
t ha n R . Then a and a' are independent over R ((a, R U{a' }, R) E /) ,
b, b' realize t he same ty pe over R and
(b , R U{a} , R ) E I , (b' , R U{a'} , R ) E /.
However no c E n can satisfy t he same ty pe as b over R U {a} and as
b' over R U {a'} because b < a < a' < b' and so c < a excludes a' < c.
2. The independence system associated with an w-stable T as before is
good. This will be checked in detail in 7.5.
3. Also t he independence syst em of random graphs enjoys amalgamation:
the reader may easily check this.
4. The same is true for ACF A and it s complet ions.
Definition 7.2.5 T is called simple when T has a good independence sys-
tem. A struct ure A is simple when its theory is.
Therefore w-stable t heories (and in par t icular st rongly minimal theories) are
simple; random gr aphs are simple, as well as any existent ially closed field
with an automor phism: in all these cases , an independence notion sat isfying
not only (11) - (16) but also the additional condit ion (17) can be introduced.
But we can say even more. Indeed, for a simple T, t his independence not ion
is unique; in other words, there is only one good independence system I of
T making T simple. I can be cha racterized as follows.
Definition 7.2.6 (Shelah) Let A E be small subsets of n, p be a type
over E . p is sai d to fork over A if and only if there are a formula <p eV, w)
and iuples i. (wh ere v < A and A is a sui t able infi nite ordinal) , all having
the same length as w, such that
(i) the bll 's have the same type over A;
(ii) <p cV, b
o
) E p;
(iii) n ll<'\ <p (nn, b
v
) = 0 (where n is, obviously, the length of v) .
In any simple t heory T , a good independence system necessaril y arises from
forking: this is t he content of t he following theor em of Kim and Pill ay.
Theorem 7.2.7 (Kim - Pill ay) Let T be simple, I be a good independence
system of T. Then, fo r eve ry tuple a in n and A E small subsets of n,
(a , E , A) E I if and only if t p( aj E ) does not fork over A.
So t he onl y possible good independence system of a simple t heory T is t he
set I of the triples (a, E, A) as before such t hat t p(ajE) does not fork over
A: this is what is necessarily meant when we say t hat aindependent from E
over A in a simple T; we will write a .J.. A E when t his happens. When A and
B are any small sets (so possibly A 'l:. E) , a .J.. A E abbreviates a .J..A A U B ;
one usually omit t he subscript A when A is empt y; so .J.. just means .J..0 ; for
B and E' sets, E .J..A E' abridges b .J.. A E' for every bin B, or equivalent ly
f} .J.. A E for every f} in E'.
For tec hnical convenience, let us rest at e (11) - (17) in t his new notation for
a simple T.
(ll) (Invariance) for ever y a, E and A as before and ever y aut omorphism
f of n, if a.J..A B , then f(a) .J.. f( A) feE);
(12) (local character) for every aand E , t here is a countable A E such
th at a.J.. A E ;
(13) (fini t e character) for every a, A and E , a .J..A E if and onl y if, for all
finite tuples bin E , a .J.. A b.
(14) (ex tension) for every ii, A and B , t here is a t uple 5' having the same
length and t he same ty pe over A as 5 and satisfying ii
'
4- AB ;
(15) (symmetry) for every s, band A, n4-Abif and only if b4- A5;
(16) (transiti vit y ) for every nand A B C, ii 4-A C if and onl y if s4- AB
and ii 4-B C.
(17) (amalgamation) let A be a model ofT, B , B
'
"2 A satisfying B 4- AB
'
,
b, iI be tuples in n havin g t he same t ype p over A and satisfying
b4- AB , b' 4- AB
'
respectively; then there is some s in n reali zing t he
same type as bover B and as iI over B
'
, and sat isfying s4-A B UB' .
The read er may directly reali ze what t hese properties imply when t uples ii
are repl aced by arbitrary small subsets, so when one deal s with B 4- AB
'
,
and rewrit e them in this enlarged setting.
Remarkably symmetry - namely (I5) - is a key proper ty of the forking in-
dependence within simple t heories : indeed , simplicity is equivalent to t he
symmet ry of forking. Also t ransit ivity (16) and local character (12) have t he
same crucial rol e. This was observed by Kim and Pillay as t o local character,
and has been recentl y shown by Kim in the other two cases.
Theorem 7.2.8 A theory T is simple if and only if, one of the follo wing
proposit ion s holds:
(i) the fo rking independenc e satisfies symmetry (15);
(ii) the fo rking independence satisfies transit ivity (16) ;
(iii) the forking independence satisfie s local characte r (12).
In conclusion, simple t heories are those wher e independ ence can be reason-
abl y introduced and developed in the axiomatic way we said before. Of
course, this is a pos itive property towards classifiability, But one can won-
der which is the reverse of t he medal , in other words what happens in non-
simple t heories. Well , Shelah proved a qui te negative resul t about them: in
fact , t hey are not classifiabl e, and so must be rejected in our classification
progr am.
Theorem 7. 2.9 (Shelah) If T is not simple, then I (T , A) = 2'\ for every
uncountable cardinal A.
Simple t heories exclude o-rninimal t heories, as obser ved before. More gen-
erally one can show t hat no com plete t heory of linear order s can be simple.
7.3. STABLE THEORIES 235
Theorem 7.2.10 No linearly ordered infinite structure has a simple theory.
In particular, no complete theory of linearly ordered infinit e structures is
classifiable.
It is perhaps worth repeating t hat, in spite of th is result , infinit e linearly
ord ered st ructures are not so bad . On t he contrary, some of them, like dense
or discr ete linear orders, real closed fields and, more generally, o-rninimal
mod els satisfy some nice and powerful model theor eti c properties; in all th ese
frameworks a suit able independence notion, sat isfying the axioms (11)- (16)
can be introduced. So things are not so sick as t hey look. We will come
back to t his point at th e end of Chapt er 9.
To conclude t his sect ion, let us sket ch a br ief history of simple t heori es.
Simplicity was first int roduced by Shelah in 1980, but its relevance was not
complet ely realized at that t ime; indeed Shelah laid a major emphasis on a
st ronger notion - st abilit y, the matter of the next section - and regarded sim-
plicity as a generalization of st ability, and not directly as a key dichotomy.
The role of simplicity in introducing independence and so towards classi-
fiability was neglect ed until t he nineti es, when one obser ved t hat several
int erest ing algebraic st ruct ures - most notably, exist ent ially closed fields
with an aut omorphism - have a simple unstable theory. Then Kim in 1996
showed that the forking independence sat isfies symmet ry (15) wit hin simple
theories and, together wit h Pillay, proved the close connection bet ween in-
dependence and simplicit y introduced before (in Theorem 7.2.7). Kim again
realized in 2001 t he key role of symmet ry for the for king independence in
the simple framework.
Now simplicit y t heory has deser ved it s own room in Model Theory, as an
autonomous and relevant part of t he classification program.
7.3 Stable theories
As said in t he last sect ion, Shelah's original analysis put its emphasis on
st abilit y rat her than simplicity. St abi lity looked t he key property towards
independence. But, aft er Kim and Pillay, one realized that it is simplicity
th at plays t he cent ral role here , while st abilit y strengths simplicity in t he
sense we are going t o explain.
There are several ways to int roduce stability, and their equivalence is not
so immediat e as one would like. We adopt th e following definition, which
underlines t he connection wit h simplicit y: T st ill denotes a complete first
ord er theory with no finite models in a countable language L, and n is a big
saturated model of T .
Definition 7.3.1 T is stable if and only if T has an independence system
I satisf ying the follo wing additional assumption:
(18) (stationarity over models) for ever y model A of 1', tupl es a, ;; in
f2 and small subset B ;;2 A in f2, if a, a
'
have the same type over A
and are independent from B over A (with respect to 1), then they have
the same type also over B.
T is called unstable when it is not stable. A struct ure is said to be stable
or unstable when its theory is.
The content of (18) is clear: t here is a unique way t o ext end a ty pe p over a
model A of T t o a ty pe over B ;;2 A of tuples independent fro m B over A.
Let us propose an easy example to illust rat e this condit ion (18), and so
stability.
Example 7.3.2 Just for a change(?!) , we deal with a strongly minimal T
and I-types over a model A ofT. The algebraic l-types are realized in A and
so extend uniquely over any B ;;2 A. But th e only non- algebraic p E 51 (A)
may have several extensions over B; however onl y one of th em does not fork
over A, and this is t he unique non- algebraic I-type over B , in other words
the ty pe of any element out of acl(B). What happens if we enl arge our
analysis from models of T to arbitrary small sets A? Now even an algebraic
l -type p over A may have several extensions over a B ;;2 A, and all of t hem
do not fork over A; however th eir number is finit e and cannot exceed 1<p(f2) I
where <p (v) is an L( A)-for mula isolating p; on the cont rary t he non-algeb raic
I-type over A has again a unique nonforking ext ension over p, t hat is t he
non- algebraic l -type over B.
Hence st rongly minimal t heor ies are stable. But this property may fail
elsewhere; indeed we are going to see that it does not hold any longer in non-
simple t heories, and even in certain simple struct ures, like random graphs
(and others).
Theorem 7.3.3 1f T is stable, then T is simple.
Proof. Let I be an independ ence system of T sat isfying (18). We have
to check amalgamation (17). So t ake a model A of 1', t wo sets Band B
'
ind ependent over A (with res pect to I) , and t wo tuples b, iJ having t he same
type over A and satisfying (b, B , A) E I , (iJ ,B
'
, A ) E I respectively. Using
extension (13) , one finds two tuples J! having the same t ype as bover B
and as b
'
over B
'
res pect ively, and both independent over B U B
'
over A.
7.3. STABLE THEORIES 237
In particular, J, Jt have the same type over A. By (18), they have the same
type also over B UB'. Hence lhas the same type as Jt, and consequently
as ii, over B
';
furthermore l has the same type as bover B. So l satisfies
(17) as c.
Then any independence system I satisfying (18) in a stable T is good, and
must equal the forking independence: for a, A and B as usual,
(a,B,A) E I atA B.
Accordingly (18) can be restated as follows for a simple T.
(18) Let A be a model of 1' , B be a small subset of n including A. Then
every type over A has a unique non-forking extension over B.
Once again recall that the existence of a non-forking extension is just stated
in (14). Uniqueness may fail for a stable T over arbitrary subsets A of n; but
also in this extended framework stability bounds the number of the possible
non-forking extensions over B of a type over A, in the following sense.
Theorem 7.3.4 T is stable if and only if T is simple and, for every small
A nand pES(A) , there is a cardinal K, (less than 1nI) such that, for any
small B :2 A, p has at most K, non-forking extensions in S (B).
Another equivalent way of defining stability relies upon a counting type
characterization (resembling w-stability).
Theorem 7.3.5 T is stable if and only if there is some infinite cardinal A
such that, for every small A n of power A, IS(A)I = A.
We wish to underline another equivalent characterization, saying that sta-
bility just excludes definable infinite linear orders, in the following sense.
Theorem 7.3.6 T is unstable if and only if it satisfies the order property:
there is a formula <p( v, 'Ill) - where v and 'Ill have the same length n - such
that <p(n
2n)
linearly orders an infinite (possibly non definable) subset of n n.
So the order property is the key obstruction to stability. On the other side
we saw in the last section that no infinite linearly ordered structure has a
classifiable theory; this applies more generally to theories with the order
property, hence to unstable theories.
Theorem 7.3.7 (Shelah) If T is not stable, then for every uncountable
power A I (1', A) = 2.\. In particular, T is not classifiable.
Hence instability is a negative condit ion in the classification perspective. On
the other hand, which are t he benefits of the stability assumption? Some of
them are already described by t he definition and t he equivalent (positive)
characterizations listed before. In particular we emphasi ze once again t hat ,
for a stable T and for a model A of T, ever y t ype p over A enlarges uniquely
to a non-forking extension over any set B 2 A. Let plB denote this extension
in S(B) ; plB can be characterized as follows
* for every L-formula <p(V, w), if there is some bin B for which <p (V, b) is
in pi B , then there is ii in A such that ip(v, ii) E Pi
** for every formula <p (v, b) E pi B, there is some ii in A for which F <p( ii, b).
For an arbitrary T, a type q over B sat isfying * (in the place of plB) is
called an heir of p, and a type q over B satisfying ** is called a coheir of p.
So, within st able theories and for t ypes over mod els,
heir = coheir = non-forking ex tension.
We will check t hese equ aliti es in the restrict ed w-stable framework in 7.5.
Anot her benefit of t he stability ass umption concerns definability. To de-
scribe t his, we need t he following definition .
Defi nition 7.3.8 Let A be any small subset of n. A type p over A is said
to be d efinable over A if and only if, for eve ry L -formula <p( ii, w), the set
of the tuples ii for which <p( ii, ii) occurs in p is A -defi nable; in oth er words,
there is an L(A)-formula d<p(w) (possibly with paramet ers) such that, for
every ii in A , <p (v, ii) E P if and only if F d<p( ii) .
Theorem 7.3.9 T is stable if and only if eve ry type over a model A of T
is definable; moreover, for every small B 2 A , eve n plB is definable over
B , and indeed, for every formula <p( v, w) the same L (A) -f ormula d<p( w)
working for p and A goes right also over B.
Th e last theorem impli es in it s turn
Cor olla ry 7.3.10 Let A -< B be models of a stable theory T . If <p (v) is a
formula wit h param et ers from B (and n is the length of il}, then <p (A n)
is A -defi nable.
We will check t hese results in det ail in 7.5 in t he restrict ed framework of w-
stable t heories. Now, to conclude t his sect ion, let us propose some examples
of stable or unstable t heories.
7.4. SUPERSTABLE THEORIES 239
Examples 7.3.11 1. Every w-stable theory T is st able. This will be
deduced by the definition of st ability in 7.5 below; it follows even
more directly from other charact eri zations (for instance, from Theorem
7.3.5). Indeed, the part icular case of strongly minimal t heories was
alread y dealt with a few lines ago. So any infinit e vectorspace, as well
as any algebraically closed field, or any differentially closed field of
characteristic 0, has a st able t heory.
2. Any module (over a countable ring R) has a stable theory: this ext ends
what we have just observed about vectorspaces. The proof requires
very basic preli minaries about the model theory of modules, and can
be found in the references quoted at the end of the cha pt er.
3. Real closed fields do not have a stable theor y. Indeed RCF is not even
simple, and anyhow it houses infinite linear orders.
4. The theory of random gr ap hs is unstable, although simple. In fact, let
A = (A, R) be a random gr aph, B A be a small subset of n. Use
compact ness and find a and at in n such that n 1= R( a, x) for every
x E Band n 1= R(a, b) for a unique b E B , with b 1- A. Then a and
at have t he same t ype over A and each of them is independent from B
over A. But they do not have th e same t ype over B.
5. Similarly, existentially closed fields with an aut omorphism are simple
and unstable.
6. On t he cont ra ry, sepa rably closed fields and differenti ally closed fields
in pr ime characteristics are stable.
7.4 Superstable theories
Also in t his section T is a complet e first order t heory with infinite models
in a countable language L, and n denotes a big saturated model of T .
Let us underline once again a particular feature of the st rongly minimal
case: this assumption implies, among other thi ngs, that T is stabl e and,
over any small subset A of n, th ere is a unique non-algebraic I-type p,
which is realized by all t he elements outside acl(A); moreover, for a and B
in p(n), so out of acl(A) ,
a.J!A B means a E acl(B) .
In particul ar YAsatisfies wit hin p(Q) t he properties (0 1) - (D4) st ated in
Chapter 5.
Now consider an ar bit rary stable T. In t his enlarged setting t here may be
several non- algebraic I-types over a small A; fix one of th em, call it p and
look at t he set p(Q) of it s realizations in Q. We wond er whet her YAstill
sat isfies (Dl) - (D4). Assume momentarily t he following condition.
(*) For every small B 2 A, p has a unique non-forking ext ension over B.
Recall t hat p has at least one non-for king extension over any B: t his is just
what is stated in (14). (*) requires t hat all t he elements a E p(Q) satisfying
a t AB realize t he same ty pe also over B. Not ice t hat, as T is stable, every
p over a model of T sat isfies (*). Call p stationary when (*) holds.
So take a non- algebraic stationa ry t ype p over a small A. It is easy to check
t hat YA satisfies (D'l), (D2) and (D4) in p(Q). In fact , let a, bE p(Q) and
B p(Q).
(D1) If a E B , t hen a YA B.
As p in not algebraic we can find infinitely many elements b
o
= a, bl, ... ,
b
n
, ... realizing p in Q . The formul a v = b
o
is in tp(a/ B) but no other b
n
can satisfy it. Hence tp( a/B) forks over A according to Shelah's definition ,
and so a YA B.
(D2) If a YA B , t hen t here is some finite B
o
B for which a YA B
o.
This follows directly from (13).
(D4) If a YA B U {b} but a tA B , t hen b YA B U {a}.
Oth er wise transit ivity (16) implies b tB a, whence a tB b by symmet ry (15);
as a tA B , (16) applies again and gives a tA B U {b}.
On t he oth er side, t here do exist non algebraic stationary p's for which YA
does not obey th e last condi tion:
(D3) if a YA B U {b} and bYA B , t hen a YAB.
Here is a counterexample.
Example 7.4.1 Consider t he t heory of t he usual Cartesian plane R
2
wit h
the two proj ections
(x, y) 1--7 (x, 0), (x, y) 1--7 (0, y), Vx, yE R
onto t he z-axis and t he y-axis respectively. There are t hree non-algebraic
I-types over t he empty set; t hey cor respond in R 2 to:
7.4. SUPERSTABLE THEORIES
(i) any element (x, 0) with x =1= 0,
(ii) any element (0, y) wit h y =1= 0,
(iii) any element (x, y) wit h x, y =1= 0
241
respectively (the only algebraic l-type concerns (0,0) , that is the unique
point lying in both the projection set s). Let p denot e t he third ty pe in t he
previous list. Now let us consider t he algebraic l-types over R
2
; t hey can
be parti tioned int o t hree classes:
(a) a new element (00, y) with an old ord inate y E Rand 00 1. R ;
(b) a new element (x, (0) with an old absciss x E R and 00 1. R;
(c) a ent irely new (00, (0) wit h 00 1. R .
Act ually the same analysis works over every model of T . The ty pes in
(a) and (b) have Morley rank 1, while th e t hird ty pe has Morley rank 2.
Accordingly T is w-stable with Morl ey rank 2 and Morley degr ee 1. The
thi rd type is t he only non-forking extension of p; actually, even t he ty pes
in (a), (b) wit h x , y =1= 0 enlarge p, but t hey fork over 0 (when y = 0 and
x = 0 t he ty pes in (a) and (b) extend t hose in (i), (ii) respecti vely) . It is
also straight forward to check t hat p is stationa ry. Now take
A = 0, B = R
2
, a = (00, (0), b = (x, (0)
with x =1= 0 in R . Clearly a and b realize p, in parti cul ar a -l- R
2
. But
a R
2
U {b} and b R
2
This cont radicts (D3) .
Definition 7.4.2 A (non-algebraic stati onary) i-type p over a small A is
called regular when (D3) holds in p(f2).
Our classification purposes suggest t o handle stable t heories T admitting
suitably many regul ar ty pes. In fact , roughl y speaking, t he more t hey are,
t he bet t er one can expect to approach T . Now put:
Definition 7.4.3 A stable theory T is called superstable if and onl y if,
for every small Band pE S (B ), there is a fi nite subset A of B over which
p does not fork. A structure A is said to be superstable when its theory is.
Notice t hat superstability generalizes what happens in t he strongly minimal
case . But what is more remarkabl e for our purposes is t he following resul t.
Theorem 70404 Let T be a superstable theory, A and B be two models of
T such that A is an elementary su bstructure of B and A =f:. B. Then there
is some b E B - A such that the type of b over A is regular.
So regular types are not rare within superstable theories. On the other hand,
if superst abilit y fails , then no classification can be expected.
Theorem 704.5 Let T be sta ble unsuperstable. Then, f or every uncountable
power A, I (T , A) = 2,\. In particular T is not classifiable.
Accordingly superstability provides a new dichot omy in Shelah's approach
to classificat ion. For, it ens ures sufficient ly many regular types, while its
negation excl udes any classification. Supe rstability can be introduced in
other equivalent ways. Let us quote a cha racterization based on a count ing
types cr iterion (just as we did in t he stable setting) .
Theorem 704.6 T is superstable if and only if, for every A 2
No
and every
set A of power A, there are at most A i -t ypes over A .
Let us concl ude t his section by list ing some examples and counterexamples
of superstable theories.
Examples 704.7 1. Separably closed fields , or even differentially closed
fields in a prime characteristic are not superstable (although t hey ar e
stable).
2. On t he other hand , all w-stable t heories are superst able. This can be
eas ily deduced from t he previous definiti on, but we will postpone t he
details to 7.5 below.
3. There do exist superstable non w-stable theory. For instance t he t heory
of the additi ve group of int egers is so.
7.5 w-st able theories
Just to have a short break in our outline of Shelah's classification, let us
come back in t his section to a more familiar setting, t he one of w-stable
theories. What we plan to do is to examine closerly within th is part icular
framework t he main notions int roduced so far in our report . Accordingly
we will check that w-stable theories are simple, stable and superstable; we
will characterize forking, independence and so on in the w-stable setting; we
7.5. w-STABLE THEORIES 243
will provide proofs and details of t he main t heorems stated t ill now. So,
all t hroughout t his sect ion, T will denot e an w-stable t heory in a coun t abl e
language L.
The first fact we want t o check is th at w-stable t heories are st able as well.
Indeed wit hin Shelah's dichoto mies, and in our outline, simplicity pr eceded
stability, and consequentl y one might reason abl y expect t o meet simplicity
before stability at this point ; we will explain lat er why we are reversi ng
t his ord er and postponing simplicity in t his sect ion. We provided several
equivalent characterizations of stability in Section 7.3; in parti cular we said
that stable theories are just t hose failing to have t he ord er property. And
act ually it is easy enough to show t hat w-stabilit y impli es stabilit y if we
refer to t his characterization.
Theorem 7.5. 1 No w-stable theory T satisfi es the order property. Equiva-
lently, any w-stable T is stable.
Proof. Let T be w-stable. Suppose t owards a cont radict ion t hat t here are
an L-for mul a cp (v, 'Ill) and t uples ai (i E N) in n (say of length n) such t hat,
for every choice of i and j in N ,
n F cp (ai , aj) {::} i::; j.
A straightforward applicati on of th e Compact ness Theorem shows that we
can repl ace N by Q above; in other words, t here are tupies ai (i E Q) in n
such t hat , for every choice of i and j in Q,
n F cp(ai, aj) {::} i::; j .
Notice t hat t here do exist definabl e subsets X in n such t hat t he indexes
i E Q such t hat ai E X form an int er val of positive length I(X) in Q; for
instance, when X = n n, I (nn) = Q is so. Choose X as before with minimal
Morley rank and degree. Fix a point j in the int erior of th e corres ponding
interval I(X) and look at
X
o
= X n cp (nn, aj) , Xl = X - X
o
.
Both X
o
and Xl are definable subsets of X . Furt hemore I (X
o
) = {i E
I (X) : i ::; j} and I (X
I
) = {i E I (X) : i > j} are int ervals of posit ive
length . So X
o
and Xl have t he same Morley rank and degree as X ; but t his
cont radict s t he fact th at t hey parti tion X.
Another equivalent way of introducing st able th eori es concerns definability
of types: let us check that w-stability implies stabilit y by referr ing to t his
characterization. Recall t hat a t ype p over a small subset A of the uni verse
n of T is said to be definable over A if and only if, for every L-formula
<.p (V, w), the set of t he t uples a in n for which <.p (v, a) E p is A-definable. T
is st able exactl y when any ty pe over any mod el M of T is definable over M:
t his is what we want to check for an w-stabl e T. We begin by t wo t echnical
lemmas.
Lemma 7.5.2 Let X nnbe a definable set, A be a small subset of n such
that every automorphism f of n fi xing A pointwise fixes X setwise. Then
X is A-definable.
Proof. Let <.p (v) be a formula defining X, and let P denote the set of the
ty pes p over A of elements ain X . Notice t hat, in this case, every realization
bof p is in X ; in fact there is an A-automorphism f of n sending ato b;
f fixes X setwise, and so, as ais in X, bis in X , too. In conclusion, for
every pEP and a 1= p, n 1= <.p (a). Now repeat the same argument as in
Theorem 6.5.5: by compact ness, for every pEP t here is a finite conjunction
of formulas in p, and hence a single for mul a <.pp (v) of p such that
Furt hermore, for every a EX, P = tp(a/A) is in P, and a E <.pp (nn). In
conclusion X = UPEP <.pp (nn), in other words {Urpp : pEP} is an open
covering of Urp. By compact ness agai n, there is a finite subset Po of P such
that {Urp p : p E Po} cover s Urp , hence X = U PEP
o
<.pp (nn). Accordingly X is
defined by t he L(A)-formula V PEP
o
<.pp(v).
Lemma 7.5.3 Let T be an w-stable theory, M be an model of
T, X nnbe a non-empty M -definable set, Y be any definable (poss ibly
non M -definable) subset of n such that Y X and Y has the same Morley
rank as X . Then Y n u i= 0.
Proof. Put 0' = RM(X) , d = GM (X). Recall S; 0' < 00. Then proceed
by induction on t he pair (0' , d) (with respect to t he lexicogr aphic order) . If
0' = 0, then X is finit e, and so X u. This implies Y u. Further more
y i= 0 because has t he same Morl ey rank as X .
Now suppose cv > 0, d = 1. Then RM(X - Y) < cv. Put (3 = RM(X - V) .
Owing to Proposition 5.7.9, there are infinitely many pairwise disj oint M-
definable subsets X i (i E N) of Morley rank (3 inside X . In part icular
there exist s some natural i for which RM((X - Y) n X i) < (3 . Accordingly
RM(Y n Xi ) = (3, and by induction Y n X i n u i= 0. Hence Y n M n i= 0.
At last assume a > 0 and d 1. By Proposition 5.7.9 once again, X can be
decomposed as the disjoin t union of d i\1-definable sets X
o
, ... , Xd-l having
Morl ey rank a and degree 1. For some i < d, Y n Xi has Morley rank a and
hence Morley degree 1. By induction Y nXinMn i= 0, whence YnMn i= 0.
..
The previous lemmas have a prevalent technical flavour. The next lemma
is more substant ial and will be used several t imes lat er in this section; it
ensures t hat, in an w-stable t heory T , for every definabl e X c n n and
formula <p(v, w) , the set of t he t uples afor which <p (nn, a) nX has t he same
Morl ey rank as X is definable.
Lemma 7.5.4 Let T be an w-stable theory, A be a small subset of n, n be
a positive integer, X be an A-definable subs et of n n, a be the Morley rank
of X. Let <p (V, w) be a formula of the language L of T (and n be the length
of v, m denote the length of w) . Th en both the sets of the tuples aE n
m
satisfying
respect ively are A-defi nable.
Proof. For every a E n
m
, RM (<p (n n, a) n X ) < a; accordingly it is
sufficient t o show that t he former set - that of t he tuples a E n
m
such
th at RM( <p (n n, a) n X) < a is A-definable; for , t he lat t er set is just it s
complement . Moreover notice that any A-auto morphism di n fixes th e set
of t he tuples a E n
m
for which RM(<p(nn, a) n X ) < a; consequent ly, if
thi s set is definabl e, t hen it is even A-defi nable, owing to Lemma 7.5.2.
In conclusion, it is enough to prove that t he former set is definable. Now
look at the Morley degree d of X: X decomposes as th e disjoint union of
d definable subsets, all having Morley rank a (and degree 1) X = Ui<dX i.
Clearly, Va E n
m
,
Hence it suffices t o show t hat, Vi < d, t he set of th e tuples aE n
m
sat isfying
RM(<p (nn, a) n Xi) < a is definabl e. In other words, we ca n also ass ume
d = 1.
Now let us sket ch the plan of t he proof.
The first step will produce a finit e subset D of X such t hat, for every aE n
m
,
if D n <p(n n, a) = 0, t hen RM(<p (n n, a) n X ) < a.
The second step will even prove th at , for every ii E n
m
sat isfying
RM (<p (nn, ii ) n X ) < Q', t here exists a finite subset D(ii ) such that D(ii ) n
<p (nn, ii ) = 0and, for a' E n
m
satisfying D(ii )n<p (nn, = 0, RM(<p(n, a')n
X ) < Q' .
Ass ume moment arily t hese pr elimin ar y steps shown, and consider all t he
finit e sets D X such t hat, Vii E n
m
,
Let denote the class of t hese sets may have t he same power as n) .
Not ice that, for every D E one can writ e a suitable formula Q'D(w)
(possibly with parameter s) just st ating "D n <p(nn, w) = 0" . Observe that
U DE6. Q'D( n
m
) is the set of th e tuples ii E n
m
such that
In fact , if ii E n
m
satisfies Q'D(w) for some D E t hen Dn <p (nn, ii ) = 0and
hence RM (<p (nn, ii) n X) < Q'. Conversely, suppose RM (<p (nn, ii ) n X) < Q'
for ii E n
m
and look at t he corres ponding set D(ii) (determined in t he second
step); D(ii ) is in and satisfies D(ii) n <p( nn, ii ) = 0, whence nE Q'D( n
m
) .
So t he t hird (and final) step of t he proof will aim at showing t hat U DE6. Q'D( n
m
)
is definable. Act ually, t here is a subtle objection one might make wit h re-
spect t o t his claim. In fact , one might notic e t hat U DE6. Q'D(n
m
) joins
many sets, and may admit t he same power as n, and so have an uncount-
able inaccessible ca rdinality; accor dingly our proof might require her e some
hypoth eses on inaccessibl e ca rdinals. However we ca n avoid any refer ence
t o t hese deep set theoretic condit ions by proceedin g in t he way we are going
t o describe. Let D(v) be a formula defining X , fix an w-saturated model M
of T containing the paramet ers in D(v), and rearrange t he first t wo ste ps of
th e proof as follows.
First ste p: there exists a finit e subset D D(Mn) such that , for every
s E nn,
D n<p(nn, ii ) = 0 :::} RM(<p(nn, ii) n X) < Q' .
Oth erwi se, for every natural i , one can find two t uples ai E n
m
, E nn
such t hat , for any i , j E N ,
RM (<p (nn, ai ) n X ) = Q',
n F .<p(bj , ai) j < i .
Let us see how t o int roduce ai, for a given natural i . Suppose aj , bj
given for all j < i . Then D = {bj : j < i } is a finit e subset of 'l9 (Mn), and
consequent ly there is ai E n
m
for whi ch
but
RM (<p (nn, ai) n X) = a.
As Rlvf(X ) = a and GM(X) = 1, RM (X n n j <i <p(nn, aj)) = a . By
Lemma 7. 5.3 , as X is M-definable, t here is some E Mn in X n
n j::;i <p (nn, aj ) such t hat Vj i n F aj). T his yield s t he t uples ai ,
as required. But at t his point it is easy to cont radict the w-stability (indeed
t he stability) of T . In fact t he formul a <p' (if, w; if, w') : ---' n 1= aj) <=> j < i
for eve ry i and j in N , and so obeys t he order pr oper t y.
In t he same way one shows the next st ep.
Second step: for ever y ii E n
m
for which RM(<p(nn, ii) n X) < a, ther e is a
finit e set D(ii) contain ed in 'l9(Mn) such that D(ii) n <p(nn, ii) = 0 and, for
a; E n
m
,
It suffices to apply t he same procedu re as in the first step to 'l9( M n) -
<p (nn, ii) (and use t he fact t hat X - <p(nn, ii) has t he same Mo rley rank and
degree as X ).
At t his point, just as said before, we can consider t he set .6. of all t he finit e
subsets D 'l9(Mn) such that , vs e n
m
,
.6. has power < 1nl , and aD( n
m
) (where, for every D E .6. , aD is
introduced as befor e, and hence is M-definable because D M
n
) equals
the set of t he t uples ii E n
m
for which RM (<p (nn, ii) n X ) < a . It remains
to show
Third step: aD( n
m
) is definabl e.
To get t his, first noti ce t hat what we have proved with resp ect t o <p (if, w)
applies to ---' <p( if, w) as well. So t he set of the t upIes d E n
m
sat isfying
RM (X - <p( nn, a)) < 0', t hat is RM (<p (nn, a) n X) = 0', decomposes as
a union of M-definable sets f3a( n
m
) (where G ranges over a suit able set
I"). Accordingly t he formul as O'D(W) wit h D E and f3a(w) with G E r
for m an open covering of Sm(M) . By compact ness we ca n extract a finit e
subcovering, and in parti cul ar we can find a finit e subset such t hat ,
Va En
m
,
RM(<p(nn, a) n X) < 0' aE U O'D(n
m
) .
DEt::..
o
As UDEt::..
o
O'D(n
m
) is definabl e, we are done.
At last we are in a position to show:
Theorem 7.5.5 Let T be an w- stable theory. Th en every type p over a
small A n is A-definable.
Proof. Let '19 (v) be a for mul a of p having t he same Morley rank and degr ee
as p. In particular , let 0' denote t he rank of p. For every formula <p (v,w) in
the language of T and aE Am (where m is the length of w),
<p(v, a) E p 'l9(V) A <p(V, a) E p.
Moreover, if t his is t he case, then '19 (v) A<p (v,a) has t he same Morley rank and
degree as p , and so RM ('19 (nn) - <p (nn, a)) < 0' . Conversely, RM ('19 (nn) -
<p( nn, a)) < 0' impli es t hat 'l9(v) A -' <p (v, a) rf. p, t hat is 'l9(v) A <p(v, a) E p. In
conclusion, for every formul a <p(v, w) and aE Am,
By Lemma 7.5.4, th e t uples aE n
m
satisfying
(that is RM('19(nn) n <p(nn, a)) = 0') form an A-definabl e class. The L(A)-
formula d<p(w) defining t his class is just what we are looking for. -"
Another remarkable consequence of Lemma 7.5.4 is t he following
Theorem 7.5.6 Let p be a type over any model M of an w-stable theory
T. Th en p has Morley degree 1.
Proof. Let '19 (v) be a formul a of p havin g t he same Morley rank 0' and
t he same Morley degree d as p. Accordingly t here are d formulas 'l9
i
(v, Xi )
(i < d) wit h parameter s xi in n such t hat
7.5. w-STABLE THEORIES
(ii) 'l9i(nn ,xj) n 'l9 j(nn, xi) = 0 for j < i < d,
(iii) RM ('l9 (n n) n 'l9
i(
n n, xi )) = Q' for ever y i < d.
249
The class of the sequences xi in n sat isfying (iii) for a given i < d is M-
definable becaus e 'l9 (iJ) has got it s own paramet ers in M. Accordingly t he
existence of some sequences i'o, . . . , Xd- l satisfying (i) , (ii) and (iii) can be
expressed by a suitable first ord er sente nce wit h parameters in M . Conse-
quently t hese t uples xo ,... , Xd- l can be chosen in M . Correspondingly we
ca n ass ume t hat, for every i < d, 'l9( iJ, xi ) is a formul a with paramet ers in
M; so there is i < d such t hat 'l9(iJ, xi) E p, as Vi<d'l9( iJ, xi) E p. If d > 1,
t hen (iii) implies RM('l9(iJ) A'l9 j(iJ, xj )) = Q' for every j < d, j =1= i, and hence
GM('l9(nn) n 'l9
i
(n n, xi )) < d. But t his cont radict s 'l9(iJ) A 'l9i(iJ, xi) E p. So
d = 1, as claimed. ...
Now we t urn our attent ion t o simplicit y. Our aim is t o prove t hat w-stable
th eories are simple. More specifically, we will show that , for an w-st able
theor y T , th e t riples (ii, E, A) such that ii is a t uple in n , B 2 A are small
subsets of nand RM(tp(ii/A)) = RM (tp(ii/B)) form a good independence
system of T . This confirms t hat T is simple, and proves also t hat, for ii, A
and B as before, ii is independent of B over A ii -!-AB if and only if tp(ii]A)
has the same Morley rank as its extension tp(ii/ B). Incident ally recall t hat,
for any ii, A and B, tp(ii/B) always includes tp(ii/A), and consequent ly
RM(tp(ii/B)) <RM(tp(ii/A)) .
Theorem 7.5.7 Let T be an cc-stable, and let I be the set of the triples
(ii, B, A) where ii is a tuple in n, A B are small subsets of n and
RM (tp(ii/B)) = RM (tp(ii/ A)). Then I is a good independence system of
T .
Proof. We have to check t hat I sat isfies th e condit ions (11)-(17) list ed in
Section 7.2.
(11) is trivially t rue, as automorphisms preserve both Morley rank and Mor-
ley degree.
(12) Fix ii, B, and pick a formula <p(iJ) of tp(ii/ B) having th e same Morley
rank (and degree) as tp(ii/B). Let A be the set of th e paramet ers in <p (iJ).
T hen A is finite (hence countable) and <p( iJ) belongs t o tp(ii/ A), when ce t he
Morley rank oftp(ii/ A) equals t hat of <p(iJ) , and so t hat oftp(ii/B).
Before (13) , let us t reat (16) .
(16) Fix A B C and a. Recall RM (tp(ajA) ) RM (tp(ajB))
RM(tp(ajC)) . Consequent ly RM (tp(aj A)) and RM(tp(ajC)) coincide if
and onl y if both t hese ranks equal RM (tp(aj B)).
(13) Take A B e a. If R1vf(tp(ajA)) = RM (tp(ajB)), t hen we can
apply what we have just checked in (16) and deduce RM(tp(aj A U b) ) =
RM(tp(ajA)) for every bEB.
Conversely suppose RM(tp(ifjA)) > RM (tp(aj B)). Choose a formula <p(v, b)
in tp(ii]B) having the same Morley rank as tp(ii]B). Look at tp(ajA U b):
t his ty pe includes <p(v, b) and consequent ly has the same Morley rank as
tp(aj B), and anyhow a Morley rank smaller t han tp(ajA) .
(14) Now we have t o show th at, for every choice of A B and if, th ere is a
tuple at having th e same t ype over A as if and sat isfying RM(tp(at j A)) =
RM (tp(at j B)). In oth er words, we claim t hat every ty pe p over A has
some extension q in 5(B) having t he same Morley rank (t o realize t his, put
p = tp(atj A) , q = tp(;' j B)). Let a be t he Morley rank of p. Choose a
formula <p(v) of p having t he same Morley rank a and the same Mo rley
degree as p. We know that t here do exist (fi nitely many) ty pes q in 5 (B)
containing <p (v) and havin g rank a ; indeed the sum of t heir Morl ey degrees
equals t he degree of <p(V) , and so t he degree of p. It suffi ces to show t hat
the extensions of p in 5( B) with rank a coincide wit h t hese types. Now it
is obvious t hat any extension of p contains <p( v) . On t he ot her side , take
any ty pe q over B containing <p(v) and having rank a. Pick f)(v) E p, t hen
<p (v) A f)(v) is also in p, and consequent ly has th e same Morl ey rank and
degr ee as p. Accord ingly the Morley rank of <p (v) A -.f)(v) is less t han a and
<p( v) A -.f)(v) ca nnot belong to p. So neith er -.f)(v) is in p. As q is cornplete,
q includes f)(v). In concl usion p q.
In part icular obser ve t hat, if M is a model of T and B is a small subset of
n including M, t hen every ty pe p over M has a unique extension q over B
with the same Morley rank. In fact, as we saw befor e in t his section, t he
Morley degree of p is 1. This remark is useful t o show (17).
(17) Let M be a model of T, B and B
'
be two small sets including M .
Owing t o t he last observation, any ty pe p over 1\1 has a unique extension
wit h t he same Morley rank over any set extending 1\1, in part icular over B,
over B
'
and over B UB' . A t uple realizing th e last ty pe (over B UB') must
satisfy th e only ext ension of p of the same rank both over B and over B
'
.
It remains to show (15).
(15) Let A be a small subset of n, if, bbe two sequences in n. We have to
prove t hat RM (tp(aj A U b)) = RM (tp(ifj A)) implies RM (tp(bj A U a)) =
RM(tp(bjA)). Let n, m denote the length of ii, b respectively.
First assume that A = M is the domain of an model of T. Let
<p(v) be a formula oftp(iijM) with the same Morley rank a as tp(iijM) (and
degree 1), and let 'l/;(v) be a formula oftp(bjM) with the same Morley rank (3
as tp(bjM) (and degree 1). Suppose towards a contradiction RM(tp(bjMU
ii)) < (3. Then there is a formula X(v, w) E tp(ii, bjM) such that X(ii, n
m
)
has Morley rank < (3. Without loss of generality we can assume
n FViNw(X(v, w) -7 <p(v) 1\ 'l/; (w));
otherwise we replace X(v, w) by its conjunction with <p(v) 1\ 'l/; (w). The set
of the tuples cin n for which RM(X(c, n
m
) ) < (3 (equivalently RM('l/;(nm)n
X(c, n
m
) ) < (3) is M-definable because 'l/;(n
m
) is. Furthermore RM(X(ii, n
m
) )
< (3. Accordingly we can assume RM(X(c, n
m
) ) < (3 for every tuple c in n
such that X(c, n
m
) i= 0 (just replace <p(v) by <p(v) 1\ "RM(X(v, n
m
) ) <
(3"). It follows x(nn, b) n M
n
= 0, otherwise, if a' is any element in
this intersection, then RM(x(a', n
m
) (3 as x(a', w) E tp(bjM). By
Lemma 7.5.3, RM(x(nn, b)) < RM(<p(nn)) = a because x(nn, b) <p(nn),
but x(nn, b) n u = 0. On the other side X(v,b) E tp(iijM U b), so
RM(tp(iijM U b)) < a = RM(tp(iijM)) (a contradiction).
Now let A be any small subset of n. Let M be an model
of T extending A. We can assume RM(tp(bjM)) = RM(tp(bjA)) and
RM(tp(iijM U b)) = RM(tp(iijA U b)). In fact , if this is not the case , then
use (14) and get a tuple P realizing tp(bjA) and satisfying RM(tp(b'jA)) =
RM(tp(bjA)). As i, b
'
have the same type over A, there is an A-au to-
morphism f of n mapping b into b
'.
Of course f(ii) has the same type
over A as ii. Now use again (14) and replace f(ii) by a tuple ;;; having
the same type over A U P as f (ii) and satisfying RM(tp(a'j M U p)) =
RM(tp(iijAUP)). As before, there is an automorphism 9 ofn fixing AUP
pointwise and mapping f(ii) into;;;. So gf is an A-automorphism sending
ainto a' and binto P. Owing to (11), if a' and Psatisfy our claim, then
the same holds for ii and b. In conclusion, we can assume RM(tp(bjM)) =
RM(tp(bjA)) and RM(tp(iijMUb)) = RM(tp(iijAUb)) , as said before. Now
recall what our hypothesis says , that is RM(tp(iijA U b)) = RM(tp(iijA)) .
Furthermore RM(tp(iijM U b)) = RM(tp(iijA U b)). Hence RM(tp(iijM U
b)) = RM(tp(iijA)) and consequently RM(tp(iijM U b)) = RM(tp(iijM)).
Owing to what we showed before when dealing with an model ,
RM(tp(bjM U ii)) = RM(tp(bjM)). As RM(tp(bjM)) = RM(tp(bjA)) ,
it follows RM(tp(bjM U ii)) = RM(tp(bjA)) , whence RM(tp(bjA U ii)) =
RM(tp(bjA)). This accomplishes the proof of (15) and the whole theorem.

At t his point we can also check (18) and confirm once again in t his way
t hat an w-st able theory T is stable. Indeed, given any t ype P over a small
A n and B 2 A, t he number of non-forking extensions of p in S( B)
ca nnot exceed the Morley degree of p , and so is anyhow finit e. So, when
A is t he domain of some mod el of T, (18) is just a rephras ing of Theor em
7. 5.6: in fact , the Morley degree of p is 1 in this case .
At t his point , it is immediat e t o deduce also t hat w-st ability impli es super-
stability.
Corollary 7.5.8 An w-stable theory T is superstable.
Proof. T is stable, and the set A in (12) is finite . '"
Now we want to discuss in our w-stable setting the concepts of heir and
coheir introduced in Section 7.3, and to show their equivalence with the
notion of non-forking exte nsion. Let us recall briefly how heir and coheir
are defined. First we fix our framework: T is a( n w-stable) theory, M is a
mod el of T, B is a small subset of n including M , n is a positive int eger , p
is an n-type over M and q is an extension of p in Sn(B) . In t his setting
q is an heir of p if and only if, for every formul a iJ(v, b) E q, t here exists
a t uple a in M such t hat iJ( ii, a) E p,
q is a coheir of p if and only if, for every formula iJ(V, b) E q, iJ(M
n
, b) =f:.
0.
'rVe are going t o show t he equalit ies heir = coheir = non-forking extension
among t he ty pes over a small set B extending a given ty pe p over a mod el
M of T. First let us recall what we showed before in t his section, namely
t hat a ty pe p over any small set A is A-definable and so, for every L-formula
iJ( v, w), finds an L (A )-formula diJ(w) such th at a t uple ain A sat isfies diJ(w)
if and only if iJ( v, a) E Pi indeed, if <p (v) is a formul a of p with t he same
Morley rank and degree as p, then diJ( w) defines th e set of all tuples a in
A for which RM(<p(v) 1\ -,iJ (v, a)) < RM(p) . But now, after seeing how
independence in T is charact erized (by the Morley rank) , we can say even
more. In fact , suppose t hat A = M is the domain of some model M of T
and let B be a small set including M. Then p (and <p(v)) have Morley degr ee
1, and so t here is a unique non-forking extension q of p over B , and <p (v)
is a formul a of q havin g its Morley rank and degree. Accordingly, for ever y
L-formula iJ(v, w), t he L( M)-formula diJ (w) still satisfies n F= diJ (b) {:}
iJ(v, b) E q for ever y bin B.
Now we can stat e our t heorem (we are still keeping here t he notation int ro-
duced a few lines ago) .
7.5. w-STABLE THEORIES
Theorem 7.5.9 The following propositions are equivalent:
253
(i) q does not fork over A;
(ii) q is an heir of p;
(iii) q is a coheir of p.
In particular the heir (and coheir) of p over B is unique.
Proof. Let q E S( B ) be a non-forking extension of p. As already underlined,
q is unique as p has Morley degr ee 1, and, for every L-formul a 'O(v, ill), there
is an L(M)-formula d'O( ill) such that, for every b in B ,
'O(v, b) E q {::} n F d'O(b)
(and, for every a in M , 'O(v, a) E p if and onl y if M F d'O(a)) . So suppose
th at , for some bin n, 'O(v,b) E q. Then n F :JilldO(ill) , and consequent ly
M F :Jill dO(w) because dO(w) is an L(M )-formula. Let a E M sat isfy
M F dO(a) , then '0(v, a) is in p and in q. So q is an heir of p. Now suppose
that q' E S(B) is anot her heir of p. Th ere are a formula '0(v,ill) in L and a
tuple bin B such that oio, b) E q - q'; in particular n F d'O(b) , and hence
-,'0(v, 01) 1\ d'O(a!) E p.
So M F d'O( a
l
), while 'O(V, a!) 1. p: t his is a cont radict ion. It follows th at q
is the onl y heir of pin S (B ). This implies t hat (i) and (ii) are equivalent .
Now we check (ii) {::} (iii) . First suppose B - M finite; let blist the elements
of B - M , and let arealize q. Observe that
tp(a/M Ub) is an heir of tp(a/M)
if and onl y if
for every L(M)-formula t.p (v, ill) satisfying n F t.p (a, b) , t here exist s mE M
such that n F t.p(a, m)
and hence if and onl y if
tp(b/M U a) is a coheir of tp(b/M) .
So t he equivalence (i) {::} (ii) and (15) imply that tp(a/ M U b) is an heir of
tp(a/M) if and only if tp( b/ MUa) is a coheir of tp( b/ M ). This est ablishes t he
equivalence of (ii) and (iii) when B - M is finite. To ext end t his conclusion
to the gener al setting, it suffices to obser ve that q is an heir (a coheir) of p
if and only if, for every finit e subset Bi, of B, t he rest rict ion of p to MU Bi,
is an heir (a coheir respectively) of p. ...
At this point we can complete what we observed in 6.2 about the connected
component of an w-stable group.
Theorem 7.5.10 Let 9 be an eo-stable group. Then the connected compo-
nent GO of 9 has Morley degree 1.
Proof. Take two ty pes p, q over G having th e same Morley rank as G and
containing th e formula <p(v) defining GO. Let x FP, Y F q, x -J- c y (namely
RM(tp(x/G)) = RM(tp(x/Gu {y}))) . For every formul a <p (v, w) and a in
G, <p( v, a) E tp(yx/GU {y}) if and onl y if <p (yv, a) E tp(x/GU {y}) (or also
<p0(y) 1\ <p (yv, a) E tp( x/G U {y})). As tp(x/G U {y}) is the heir of tp( x/G) ,
there is some b E GO= <p(G) such that <p (bv, a) E tp( x/ G) = p. As GOis
the left st abilizer of p, <p (v , a) E p. Hence t he t ype of yx over G coincides
with p. But just inverting th e roles of x and y one deduces in th e same way
that yx realizes q. So p = q. ...
As said before , w-stability impli es superstability. Recall th at , roughly speak-
ing, t he superstable theories T are t hose ensuring that every proper elemen-
tary ext ension M < N , M =f N between mod els of T realizes some regular
type over M ; and t he regul ar t ypes are just the non- algebraic types p over
M such t hat A sat isfies (D3), and consequently (Dl) - (D4), in p(D).
Ou r aim is to examine closerly these matters in t he w-st abl e framewor k,
where an elegant approach proposed by Lascar allows a simpler t reatment .
Indeed stronger tools are available in the restrict ed w-st able setting; for
instance, we will use over and over again (exist ence and uniqueness of)
prime models over subset s below: this is a feature of w-stable th eories which
may fail in arbitrary superst abl e theories. Accordingly fix an w-stable T
and a model M of T ; we wish t o examine types over M. By the way, let
us introduce t he following notation: for X a small subset of D, let M (X)
denote the model of T prime over M U X; when X is t he domain of some
finit e tuple a, we write M (a) inst ead of A1(X) ; notice th at , if a and a' have
t he same ty pe over M , then M (a) and M (a') ar e isomorphic by a function
fixing M pointwi se and mapping ainto a'; in ot her words, the isomorphism
type of M(a) over A1 does not depend di rectly on a, but on t he t ype of a
over M .
Wh at we fi rst need in our analysis is a crit erion compa ring types p and q
over M , and measuring whether realizing p is easier or harder than realizing
q. This is provided by t he so called Rudin-Keisler relation ?RK (although it
was Lascar who introduced t his noti on, faintl y inspi red by a Rudin-Keisler
order relation concerning ultrafilt ers). Here is it s definition.
Definition 7.5.11 Let p and q in S(M) . We put p '2RK q if and only if,
f or every tupl e ii satisfying p, the model M (ii) prime over M U ii con tains
some realizati on bof q, an d we put p r- tuc q if and on ly if p '2RK q and
q '2RK p.
Notice t hat , owing t o what we said ab out M (ii) before, we could equivalent ly
write "for some if' inst ead of "for every if' in t his definition.
'2RK is a preorder relati on: it is reflexive and t ra nsitive, but it is not an-
ti symmetric. So r-ruc is an equivalence relation and '2RK determines an
orde r relation in the quotient set S (M )/ r-ruc among rvRK-classes of t ypes.
Roughl y speaking, p '2RK q means t hat, wherever p is realized, q is reali zed
as well; and consequent ly p <nx q says th at p and q are realized by t he
same mod els exte nding M.
It is also clear that t he types over M minimal wit h respect t o '2RK are t he
algebraic ones, as t hey are already satisfied in M . But t hey do not interest
us, so let us exclude t hem and call a type p over M RK -minimal when it is
mini mal wit h respect to '2RK among non-algebraic ty pes.
Definition 7.5.12 A non-algebraic type p E S(M) is RK-minimal if it is
rvRK- equivalen t to every non-algebraic type q over M such that p '2RK p.
Notice t hat every RK-mini mal type p is rvRK-equivalent to some I- t ype. In
fact t ake ii 1= p; as p is not algebraic, t here exists some a in ii out of M;
t hen t p(a/M) is not algebraic and t p( a/M) S,RK p, whence tp(a/M) r-tuc P
by RK- mi nimalit y.
There is anot her relevant relation between t ypes arising in t his set t ing, and
closely connected with rvRK ; it is called orth ogonality and is defined as
follows.
Definition 7.5.13 Let p and q in S(M) . We say that p is orthogonal to
q and we writ e p 1. q if and only if, for every ii 1= p and b1= q, s {.M b.
Notice t hat 1. is symmetric. Roughl y speaking, p 1. q means t hat realizing
p and realizing q are "independent " facts. Here are some simple examples
concerning '2RK , rvRK and 1. .
Examples 7.5.14 1. Let K be an algebraically closed field. Recall t hat
t here is a unique non-algebraic I- t ype over K , that of t he t ra nscen-
dental elements over K. More generally look at arbitrary ty pes p and
q over K. Let nF= p and bF= q (where s, bmay have different length s).
Then p q if and only if t he t ranscendence degr ee of K(il) over
K is not smaller t ha n t he t ranscendence degree of K(b) over K, and
p rvRK q if and only if t hese t ranscendence degrees coincide. In par tic-
ular t he Rudin-Keisler relation det ermines a t otal ord er in t he quoti ent
set S (K) I <n , and t he unique non- algebraic I- t ype over K is t he
onl y RK-minimal ty pe (up to r-tuc ).
Notice t hat the same analysis can be repeated in any st rongly minimal
st ruct ure M (wit h respect t o t he acl dependence relation): for p,
q E S(M), il F= p and bF= q, P q if and only if th e dimension of il
over M is not smaller t ha n th e dimension of bover M , and p rvRK q if
and only if these dimensions coincid e. So the Rudin-Keisler order in
S (M )I r-tuc is st ill linear , and the only RK-minimal class if th at of
t he unique non-algebraic I-t ype over M .
2. Take the theory T of an equivalence relati on E with infinit ely many
infinit e classes (and no finite class) . T is complete (why?) . Let us de-
scribe t he non- algebraic l -types over a model M = (M , E) ofT. First ,
for every a E M; we have t o consider t he I-t ype Pa of a new element
in t he E-class of a; moreover t here is another I- t ype q, concerning
t he elements in a new E-class; no fur th er I- typ e ar ises. In par t icular
ISl (M) I = IM\, and T is w-stable. Notice also t hat , for every ty pe p
listed before and z F= p, t he domain of M (x) is just MU{x} (MUxl E
wit h a countable x]E in t he case of q). Accordingly t hese t ypes are
pairwi se rvRK-inequivalent and RK-minimal. Not ice also that t hey
are pairwise orthogonal.
Now let us introduce strongly regular t ypes . As we will see below, t hey can
reasonably replace regular t ypes in the w-stable fr amework. A1 st ill denotes
a model of a countabl e w-stable t heory T.
Definition 7.5.15 A i-type p over M is called strongly regular if p is
not algebraic and there is some formula <p( v) E p such that, for every a F= p
and se M (a) - M , b F= p if and only if F= <p (b) .
Notice that , owing t o what we observed before about A1(a) , we could say
"for some a F= p" inst ead of "for every a F= p" in t his definition. Now let us
propose some examples.
Examples 7.5.16 1. The only non- algebraic I- t ype over an algebraically
closed field, or also over an arbit rary strongly minimal struct ure, is
strongly regular: v = v works as <p( v) .
2. Take a struct ure (M, E) where E is an equivalence relation with in-
finitel y many infinit e classes and no finit e class, as in Example 7.5.13,
2. Each ty pe list ed in 7.5.13,2 is st rongly regular , Pa because of E(v, a)
and q because of v = v.
The basic fact about st rongly regul ar t ypes is:
Theorem 7.5.17 Let T be w-stable, M and N be models of T such that
M is an elementary substructure of N and M i= N . Then there is some
a E N - M for which tp( a/ M ) is strongly regular.
Proof. Let a E N - M be such t hat p = t p(a/M) has a minim al Morl ey
rank, and let th e formula <p (v) isolate p among the types of th e same Morley
rank in SI(M) ; in particular <p (v ) has t he same Morley rank as p. We claim
t hat <p( v) makes p strongly regul ar. In fact <p(v) E p. Furt hermore, let
bE <p (M (a)) - M ; owing t o t he choice of a, R M (tp (b/ M )) R M (tp (a/ M ));
as <p (v) E t p( b/M) , R M (tp (b/ M )) = R M (t p(a/ M )) and even tp( b/M) =
tp(a/M) . ..
Corollary 7.5.18 For every non-algebraic q E S( M ) there is some strongly
regular type p E S I (M) for which p q. In part icular every RI( -minimal
q is '"RK-equivalent to some strongly regular p.
Proof. Let b F q. As q is not algebraic, M (b) properl y ext ends M. So
apply th e pr evious t heorem and gets some strongly regular p reali zed in
M (b) : p q. ..
On the oth er side
Theorem 7.5.19 Eve ry strongly regular p E S I(M) is RI( -minimal.
Proof. This needs a more laborious approac h, and some preliminary ste ps
which perh aps have t heir own intrinsic int erest. First let us fix our setting.
There is some L(M)-formula <p (v ) wit nessing t hat p is strongly regular. We
take a non-algebraic q E S(M) satisfying q p and we have t o show
t hat q "' RK p. There is no loss of generality in ass uming q E S I (M) . Gi ven
a F p, th ere is some b E M (a) - M realizing q. It is an easy exercise t o
realize th at b M(a) (just compare the Morley rank of th e t ypes of b over
M (a) and over M ). The first pr eliminary step st ates t hat act ually b Ma.
In fact we have:
Step 1. Let a M, b a. Then b M(a) .
Notice t hat we do not use here th e hypothesis that p is st rongly regul ar.
What we want to show is t hat tp(b/ M( a)) does not fork over M , mor e
precisely is t he heir of tp(b/M). Accordingly take a formul a X(v, C) in
tp(b/M(a)) with paramet ers cin M( a). The ty pe of cover M U {a} is iso-
lat ed, say by "7 (w, a). We claim t hat "7 (w, a) isolat es tp(CjMU{a, b}) as well.
Otherwise t here is some t uple 2 such t hat 1= "7(2, a) but tp(c/M U{a, b}) =1=
tp(2/M U {a, b}), and so t here exist s some formul a 'I/; (w, a, b) (possibl y
with further paramet ers from M) for which 1= 'I/; (c, a, b) A -,'1/; (2, a, b) .
Consequent ly 3w3W" ("7 (w, v) A "7 (W" , v) A 'I/; (w, v, b) A -,'1/; (W" , v, b)) is in
tp(a/M U {b}). As a tM b and so tp(a/M U {b}) is t he heir of tp(a/M) ,
th ere is some m E M sat isfying
but this gives a contradiction because "7 (w, a) isolat es a single type over
M U {a} . Accordingly "7( w, a) isolates t he t ype of calso over M U {a, b}.
This t ype contains X(b, w) , so
'Vw("7 (w, a) -7 X(v, w)) A 3w"7(w, a) E tp(b/M U {a} ),
whence, as before, x(v, m) E tp(b/M) for some min M. In conclusion ,
tp(b/ M(a)) is the heir of tp(b/M), as claimed.
The second preliminary st ep proves:
Step 2. <p (M (b)) =1= <p( A1).
Oth erwi se a t M b, which cont radicts th e previous conclusion. In fact, let
O( v, a) be a formula in tp(b/ MU {a}) . As b E M(a) , M(a) 1= O(b , a), and
consequent ly M(a) 1= 3w(O(b, w) A <p (w)) (a sente nce with par ameters in
Mu {b}) . As M(b) is an elementary subst ruct ure of M(a) , even M(b)
sat isfies this sentence. So 1= O(b, c) for some c E <p (M (b)) M. Hence
O( v, c) E tp(b/M). In conclusion, tp(b/M U {a}) is the heir of tp(b/M) .
Now we can conclude our proof. Let dE <p (M (b)) - M; so d E <p (M (a)) as
well, and consequently tp(d/M) equals p. As d E M(b) , p -::;' RK q, whence
p r"VRK q. '"
Using mor e or less t he same t echni cal premises one shows also t he following
not abl e characterization of r"VR!{ within RI<-minimal ty pes.
T heorem 7.5.20 Two RI< -minimal types over M are r"VRK-equivalent if
and only if they are orthogonal.
We omi t the details. There are other relevant t hings t o underline about
st rongly regular t ypes. First of all , strongly regul ar t ypes ar e regul ar as
well: checking t his requires t he same ingredient s as in t he last t heorem and
some equivalent charact erizations of both t he involved notions. Again , we
omi t a full t reatment of t his point (t his section is going t o be quite long, and
actually we already proved about strongly regul ar ty pes what we will need
in t he rest of the cha pter) . However , it is worthy pointin g out that t here do
exist even in t he w-stable setting some regular ty pes which are not st rongly
regul ar (and indeed are not "'RK-equivalent t o any strongly regular t ype).
There is anot her point in t his framework deserving a few words. In fact ,
(and consequently '"RK) as well 1.. are preserved und er passing from
types P, q over a mod el M of an w-stable theory T to their non-forking
exte nsions over some elementary extension M': P q, P "'RK q, P 1.. q if
and onl y if plM' qlM' , plM' "'RK qlM' , plM' 1.. qlM' respectively. The
same ca n be said about RK-minimalit y. With respect t o strongly regul ar
ty pes , one sees that , if M < M' , P is a l-t yp e over M and cp(v ) is an
L( M)-formula making P st rongly regular , th en th e same formul a makes plM'
st rongly regul ar as well.
The last matter we want t o deal wit h in t his section st ill concerns non-
forki ng extensions and w-stable t heories. It is a t heorem (act ually valid
even for stable t heories) which will t urn out t o be a useful t ool in t he proof
of Shelah's Uniqueness Theor em in Sect ion 7.7. It is generally called t he
Open Mapping Theorem, a name clearly evoking t opology; of course, t he
topological spaces we are concerned with are t he Boolean spaces of t ypes
over small subsets of n. In parti cul ar , we consider two small subsets A B
in an w-stable t heory T, and we look at t he spaces S(A) and S (B) . It is
easy to realize that the restriction map r from S(B) in S( A) , sending every
t ype q over B into it s restriction to A (so into th e set of t he L(A)-formulas
of q) is cont inuous and surject ive: cont inuit y follows from th e trivial remark
that L(A) L(B) , hence L(A)-formulas are L(B)-formulas as well, and any
open set in S(A) can be regarded as the image und er r of some open set in
S(B); surject ivit y is obvious.
Now consider the set of t he types q over B t hat do not fork over A. Let
N (B , A) denote t his set. First notice
Proposition 7.5.21 N( B , A) is a closed subset of S( B ).
Proof. The clai m is obvious when A is t he domain of some model A of
T . In fact we know t hat , for every q E S (B ), q E N( B, A) if and only if
q is t he coheir of it s restricti on t o A, and it is easy t o realize t hat t he last
260 CHAPTER 7. CLASSI FYI NG
condit ion just ensures that q is in t he closure of t he set of t he types over B
of t he t uples in A . Accordingly N( B, A) equ als t his closed set .
Now let A be any subset of B. Use extension (13) and compactness, and
build a model M of T independent of B over A. Notice t hat every type
in N( B, A) has a non-forking extension over B U M , and so by t ransit ivity
enlarges to some ty pe in N( B U M, A) , which, again by t ransit ivity, lies
also in N (B U M , M). On t he other hand , every type q E N (B U M, M)
restricts t o a type in N( B, A ). In fact, let c reali ze q, henc e let c-!-M B .
Use symmet ry and deduce B -!-M M U c. As B -!-A M, t ransit ivit y yields
B -!-A MU s and consequentl y B -!- A A U c. Use symmet ry and deduce
c-!- AB, in ot her word s t hat the type of cdoes not fork over A, as claimed .
Then the restriction map from S (B UM) onto 8(B) - a cont inuos function
between compact spaces - sends N (B U M, M) just onto N(B, A). As the
for mer set is closed , it s image is, t oo. ..
Now we are in a position to prove t he Op en Mapping Theorem.
T heorem 7. 5.22 Let A B be small sets, r denote the restriction map
fro m N (B, A) onto 8(A) . Then r is open. In other words, for every L(B)-
fo rmula O(V) there is an L(A)-formula O' (V) suc h that a type q over B non-
forking over A contains O(v) if and only its restrict ion to A in cludes 0' (V) .
Proof. There is no loss of generality in repl acing B by a mod el A1 of
T elementarily incl uding B and saturated in some power > IAI. In fact ,
let r', r'' denote t he restricti on maps from N (NI, A) onto 8(A) and from
N(M, A) in 8 (B) respecti vely; it is an easy exercise t o check that the lat t er
functi on maps N(M, A) onto N( B, A) and t hat r' is just t he composit ion
of r" and r . Consequent ly, if r' is open, then r is, as r" is cont inuous. Hence
we ca n replace B by M , as clai med; accordingly, let r denote from now
on t he restriction map from N (M, A) onto 8(A). Let U be an open set of
N(M, A) . We can assume t hat U is t he set of the types over M which do
not fork over A and include a given formula <p( v, b) with parameters bfrom
M. We have to show t hat r (U) is open. Let U* denote r -1r( U) (so t he set of
t he t ypes q E N (M, A) having t he same restricti on to A as some type in U) ;
elementary t opology ensu res t hat it suffices t o prove t hat U* is open: in fact ,
in t his case, t he complement U; of U is closed, as well as it s image r( U; ) =
(r (U))' because r is closed ; whence r (U) is open , as claimed. To show t hat
U* is open , first obser ve t hat U* equals Ug{q E N(M, A) : <p(v, g(b)) E q}
where g ranges over t he automor phisms of M fixing A pointwi se. :2 is
clear. In fact , for <p(v, g(b)) E q, <p(v, b) E g-l (q) , and t he las t type g-l (q)
do es not fork over A as non-forking is preserved und er aut omorphism; hence
7.6. CLASSIFIABLE THEORIES 261
: (q) E U, and so, as q and g-1 (q) have t he same rest riction t o A, q E U* .
Now let us deal wit h t he other implication. Let q E U*, hence t here is some
t ype p E U such that p and q have t he same rest rict ion t o A. For every
for mula X(1o) of tp(bj A), ep(v, b) A X(b) is in p; as p does not fork over A,
t here is some t uple if in A for which ep (iJ, if) A X(if) occurs in t he restricti on
of p t o A and consequent ly in q. Recall t hat q E S(M) is definabl e, whence
t here is some L(M)-formula "ep(iJ, 10) E q" defining t he t upies csuch t hat
ep (iJ, C) E q. Accordingly t he set of L(M)-formulas
{ "ep( e. 10) E q"} u tp(bj A)
det ermines a (possibly incompl et e) t ype over M. As M is sat ur at ed in some
power > IAI, we can even find a t uple cin M realizing t hese for mul as; in
par t icular ep (v, C) E q. Enlarge t he identity map of A by bt----7 c, and get an
automorphism 9 of M such t hat ep(iJ, g(b)) E q, as requi red. '"
7.6 Classifiable theories
We cont inue and conclude in t his section our outline of Shelah's classification
progr am. T still denot es a complete first order theory with infinit e models
in a countable language L , n a big sat ur ated model of T . As unsuperst abl e
t heories have t oo many models and are not classifiable, we should ass ume T
superstable. However , to simplify our exposit ion, we will even assume T w-
stable, and we will t reat in det ail t his restrict ed framework. In fact w-stable
t heories are superstable as well (but not conversely) and enjoy some st ronger
properti es (such as existence and uniqu eness of prime models) making t hem
more t ract able.
To introduce our next ste ps, let us refer once again t o st rongly minimal
t heories, and even to the more particular case when T = AC Fp for a given
p = 0 or prime. Recall St einitz's analysis of an algebraically closed field IC
of cha racterist ic p: given
t he prime subfi eld IC<> of IC,
a t ranscendence basis B of IC (over IC< ,
IC is fully det ermined up to isomorphism as t he algebraic closure of t he
extension of IC<> by B , which in it s t urn depends only on t he cardinality of
B (t he t ranscendence degree of IC). This provides a qui t e sat isfact ory answer
to our classification purposes. But , in view of a possible generalization, we
have to be very car eful in weighing t he specific advantages of th e particul ar
algebraically closed (or also st rongly minimal) setting and in checking which
of t hem ca n be extended to a broader framework. Above all, we have to
recall t hat over K<> (as well as over any subfield of K) t here is a unique
non- algebraic l-type, t hat of t ranscendental elements; by t he way, t his ty pe
is st rongly regular.
Now t ake any w-stable T and a mod el M of T. Keeping t he previous exampie
in mind, we examine t he struct ure of M. M (element arily) includes t he
model M <> of T prime over 0. To construct A1 upon )\,-1 <>, look at the non-
algebraic I- t ypes over M <> realized in M . Unlike t he example of fields, in
this general sett ing we should expect to meet quit e a lot of ty pes; however we
can restrict our attention to t he RK-minimal ones, partition th em according
to t he equivalence relation r-tuc and choose a st rongly regular represent ati ve
in each rvRK- class. Incident ally recall t hat t he result ing ty pes are pairwise
ort hogo nal. At t his point take a maximal independent set of realizations
for each of t hese types, glue t hese sets, form t heir union X and t he model
M <>(X) prime over M<> U X in A1. If t his model equals M (just as in
algebraically closed fi elds) , t hen we are don e: t he isomorphism type of M
should be given by t he sizes of t he independent sets of realizations of t he
involved strongly regul ar ty pes. But it may happen t hat, given a strong ly
regul ar type over M<> realized in M , say by a, after for ming t he model
A1<> (a) ofT prime over lVf<>U{a} , one meets some new non-algebraic ty pes
over M<>(a) which are orthogonal to any type over M<> and are realized
in M (as t he next Example 7.6.1 will show) . One sees t hat t his excludes
M = M <>(X). Accordingly we have to repeat t he previous machinery
over and over again until M is reached (provided t hat A1 can be event ually
reac hed) .
Example 7.6.1 Let T be t he t heory of a l -ary function 1 such that each
element has infinitely many pr eimages via 1, t here is an element 0 for which
1 (0) = 0 and no a =I- 0 satisfies r (a) = a for any positive int eger n. The
reader may check t hat T is complete : we do not wish t o linger over t his point
and we pr efer t o descri be in detail t he non-algebraic I-types over a model M
of T . In t his setti ng, first we have to consider, for any posit ive integer nand
a E M, t he type PM(n, a) of an element satisfying r(v) = a, -,(fn-l( v) =
b) for every b E M. In addition t here is a t ype qM determined by t he
for mul as -,(r(v) = a) for any a E M and positive integer n. No fur th er
I- typ e arises. In pa rt icular SI(M) is countable when M is, and so T is
w-stable. Incident ally noti ce t hat t he set t heoretic union of two models of
T is again a model of T.
Now observe that all t he I-types mentioned above are strongly regular.
Let us check t his. First consider PM(n, a) for a E M , n a positive int e-
ger. Let x F PM(n,a). To form A1(x) , one t akes M , and one adds first
z , f (x), ... , r:' (x), and th en pr eimages for each of th em, and for ever y
pr eimage, and so on. Consequently th e only elements of M( x) - M real izing
PM(n, a) are t hose sat isfying t he formula r(v) = a. This confirms t hat
PM(n, a) is st rongly regu lar. Now look at x F qM. This time, to form
A1( x) we t ake M and we add f n(x) for every posi ti ve int eger n , and t hen
preimages for each of them, and for every preimage, and so on. All t hese
elements sat isfy as , and so th e point s of M (x) - M realizing qM ar e exactly
those satisfying, for instance, v = v . So also qM is st rongly regular.
Now take an model M of T. M <> (the model pri me over
0) is an element ary subst ruct ur e of M up to isomorphism. As M <> is
count able, M contains a realization x of PM<> (1, a) for every a E M, Also
A1' = M <>(x) is an elementary substructure of M . Now consider PM' (I , x),
which is a st rongly regular type, and again can be realized in M because
M is and M' is countable . Furth ermore PM'(I , x ) is orthogo-
nal (equivalently is not f"VRK-equivalent) to any non-algebraic I-type P over
M<>; in other words PM,(I , x ) 1.. P/M" In fact , if P = PM<>(n, a') for some
a' E M<> and positive integer n, t hen one eas ily observes PIM' = PM,(n, a') ,
while, if P = qA1<>, t hen pIM
'
= qM' , In any case, piM' is st rongly regu-
lar. To conclude PM,(I , x ) 1.. P!M' , we can equivalent ly check pIM' rfRK
PM,(I ,x), which is easily proved, as, for ever y y F PM,(I , x) , M'(y) con-
tains no realization of piM' (and conversely). In conclusion t here exist some
st rongly regular t ypes over M' realized in A1 but ort hogonal t o t he non-
algebraic types over M <>. Moreover t his pro cedure can be repeated as
many t imes as you like (afte r realizin g PM' (I , x ) by y in lVI , one can form
PM'( y) (1, y), and so on) .
Let us come back to our general analysis. We wonder und er which conditions
any model M of our arbitrary w-stable t heory T can be reached at th e top
of th e construction descr ibed before. To clarify this point we have to make
our framework mor e preci se. So let us first recall what follows.
Definition 7.6.2 For any non-zero ordinal A, let A<w denote the set of the
fini te ordered sequences of elements of A. A<w is partially ordered by the
relation ::::; according to which, for s, tE A<W, s ::::; t if and only if s is an
initi al segment of t . A<w has a least element in ::::; (the empt y sequence < ,o
every S EA<w different from <> has a (uni que) predecessor (a greatest
element < s}, which will be denoted by s: ; s is called a succe ssor of S- .
264 CHA PTER 7. CLASSIFYING
A tree of A<w is a downward closed subs et C : if SEC, tE A<w and t s,
then t E C as well.
The following definition describes in detail the models M which can be built
in the way sketched before: th ey are exactly those having a presentation of
the form we are going to explain .
Definition 7.6.3 Let M be a model of T of power A> A presentation
of M is a pai r (C , ((Ms, as) : s EC)) such that C is a tree of A<W, fo r every
sE C M s is a model ofT and as E Ms, and:
(i) M <> is prime over 0;
(ii) if s, t E C and s = t'>, then )\I{t = M s(at} and Pt = tp(at/M
s)
is
strongly regular;
(iii) if s, t , t
f
E C and s = i " = i':' , then ei ther Pt = Pt' or Pt .L Pt' ;
(iv) for every SE C, (at : t E C, t " = s) is independent over Ms,
namely, for every t E C such that C = s , at tM
s
{at ' : t
f
E C, t
f
=1=
t , t': = s};
(v) if s , t E C and s = t" >, then Pt is orthogonal to all the non-algebraic
types over Ms;
(vi) )\1{ is prime over UsEc u. .
Of course we are int erest ed in those t heories T such that every model gets
such a presentation.
Definition 7.6.4 An to-stable theory T is called presentable when, fo r all
models )\I{ o, M
1
, M
2
and )\I{ oj T such that
M o is an elementary substruct ure of both M
1
and M
2
,
M
1
tM
o
M
2
(in other words ai t M
o
M
2
f or every ai E M
1
, or, equzv-
alently, a2 tM
o
M
1
f or every a2 in M
2
),
M is prime over M1 U M2 ,
for every non-algebraic type pES(M) there exists some type over M 1 or
over M
2
that is not orthogonal to p.
Presentability is a new dichotomy within t he classification problem. In fact ,
on the one side, one shows:
Theorem 7.6.5 (Shelah) If T is an eo-stable theory and T is not pre-
sentable, then I(T, A) = 2-\ for every uncountable cardinal A. In particular,
T is not classificable.
On the other hand
Theorem 7.6.6 (Shelah) If T is ea-stable and presentable, then every un-
countable model M of T has a presentation.
A model M may admit several presentations. Actually there is a "quasi
uniqueness" theorem stating under which conditions two "different" pre-
sentations yield isomorphic models; but it is impossible to discuss it here
shortly, so we omit its treatment, and we conclude our report about pre-
sentability and presentations by proposing some examples and, in particular,
an w-stable theory which is not presentable.
Examples 7.6.7 1. Let T be the theory of two equivalence relations
El, E
2
such that any El-class and any E
2-class
share infinitely many
common elements. One checks that T is complete. Here are the non-
algebraic I-types over a model M of T:
for a E M, the formulas Et{v, a), E
2(v
, a) and ,(v = b) for all
b M determine a type PM(a);
for a E M and i = 1,2, the formulas Ei(V, a) and ,E
3-i(V,
b) for
all sM give a new type PM(a, i);
finally there is a type qM determined by the formulas ,El (v, b)
and ,E
2
(v, b) for all s M.
In particular, if A1 is countable, then Sl (M) is. Hence T is w-stable.
Moreover it is straightforward to check that all the types listed above
are strongly regular and, for all a E M, PM(a, 1) ..1 PM(a, 2). Now fix
a model Mo of T and a E 111
0
For i = 1, 2, let Xi realize PMo (a, i)
and Mi denote MO(Xi): so Mi is built by taking Mo and adding
a new E3_i-class (the class of Xi) having countably many common
elements with any Ei-class in M o- One can see that both M 1 and M 2
are elementary extensions of Mo ; furthermore PM(a , 1) ..1 PM(a, 2)
implies Xl -!-M
o
X2, whence M
1
-!-M
o
M
2
(see the proof of Theorem
7.5.19), Now form the model M of T prime over M
1
U M
2
; M is
obtained just by adding countably many elements to M
1
U M
2
in the
intersection between the E
2-class
of Xl and the El-class of X2. Let
X be an element in this intersection, and consider p = PM(X) : P is
a non-algebraic t ype over M , and one can check t hat P is or thogonal
(equivalent ly, is not "'RK-equivalent) to any ty pe over M
1
or M
2
So
T is not present abl e.
Let us also check t hat, for every uncount abl e cardinal >. , J(T, >' ) j ust
equals 2'\. In fact take t he disjoint uni on I of two sets h and J
2
of
power >.; let R be an irreflexive symmet ric binary relation in I such
t hat, if 81, 82 E I and ( 81 , 8 2) E R, th en 8 1 E h and 8 2 E h or
conve rsely (so (1, R ) is a bipartit e graph). Now build a model MR
of T where the Ei-classes correspond to th e element s of I; for ever y
i = 1, 2 and, for every El -class Xl and E
2
-class X
2
,
and
IX
1
n X
2
! = otherwi se.
It is clear th at MR has power>. for ever y R. Furt hermore, for R =1= R' ,
A1R i:- A1R' . So I (T , >' ) 2.\ and consequent ly I (T, >. ) = 2.\: in fact ,
t here exist 2.\ many relations R as befor e.
2. On t he contrary t he t heory T in Example 7.6.1 is present able. In fact
take t hree models M o, A11> M
2
of T such t hat JV1
0
is an element ary
substructu re of both M 1 and JV1
2
and M; -!-Mo M
2
. We observed t hat
M
1
U M
2
is t he domain of a mod el M of T extending M
1
and M
2
;
of course, M is prime over M
1
U M
2
Let p be a non-algebraic ty pe
over j1;[ (and keep th e same notat ion as in 7.6.1). If P = PM( n, a) for
some a E M and some positive integer n , t hen P"' RK PMi (n , a) where
i =1, 2 sat isfies a E Ms, If P =qM , t hen P is '"RK-equivalent to both
qM
l
and qM2' So T is pr esentable, as claimed.
Nevertheless T is not classifiable, as it has 2.\ many pairwise non iso-
morphic models in every uncountable power >.. Let us see why. Indeed,
for every X, we ca n build 2.\ non isomorphic models sat isfying the fur-
t her assumpt ion
for every a E M and for some natural n f n(a) = O.
Notice t hat such a mod el M ca n be viewed as a t ree of >. <w with
respect t o t he relati on defined as follows: \la , b E M, a b if and
only if r(b) = a for some natural n ; in parti cular a = b: if and
only if f( b) = a, and < >= O. The isomorphism class of M clearl y
det ermines t he isomorphism ty pe of (M, as a t ree. Moreover every
point in has infinit ely many successors, and IMI = A. So it
suffices to show t hat t here exist 2-\ pairwise non isomorphic trees of
A<w sat isfying t he last addit ional proper ties.
Associate a t ree C(v) of A<w with any ordinal v < A as follows.
(a) First let v = 0: in C(O) th e root <> has A successors, while any
fur t her vertex has successors.
(b) Now let v = J-l + 1: in C(v) <> has successors, and each of
them is t he root of a t ree isomorphic t o C(J-l) .
(c) Finally let v be a limit ordinal: in C(v), < > has a successor s/'
for every J-l < v , and s/' is the root of a tree isomorphic to C(J-l) .
At this point , let us bui ld for ever y S <;;; A a tree C(5) of A<was
follows: < > has a success or s., for every v < Aand, for every t/ < A,
SV is in it s turn th e root of a t ree isomorphic t o C( v) if v E 5 and t o
w<w otherwise. It is clear t hat jC(5) I = A for every 5 A and that
different subsets 5 I: 5' yield non isomorphic trees C(5) 'f!. C( 5' ).
At t his point one may wonder what is so wrong in t he last example to exclude
any classification of models. More generally one may ask why a present abl e
t heory may be non-classifiabl e. Recall th at , if T is presentabl e, t hen any
uncount abl e model A1 of T has a present ation, whose "skelet on" is a t ree C
of IMI<w. Of course t his is a good feature t owa rds a general class ificat ion of
models. But th e point is t hat some involved t ree might be non well found ed.
Let us recall what th is means.
Let C be any t ree (say in A<W). One associates with any point s in C a rank
r (s) (an ordinal , or 00) in t he following way. First we define r (s) a for
any ordinal a . We pro ceed by induction on a :
1. r (s) a if a = 0;
2. if a is limit , then r(s) a means r(s) {3 for every ordinal {3 < a;
3. if a = (3 + 1, th en r(s) a means t hat , for some t E S with s = t r ;
r(t) (3.
If t here is an ordinal a for which r(s) 'l a, t hen t he least ordinal with t his
proper t y is necessarily a successor ao + 1, and we put r( s) = ao. Otherwi se
(when r(s) a for every ordinal a) we put r(s) = 00. We say t hat C is
well founded if r( <"> is an ordinal.
Now let us arrange t hese definitions in our setting, when T is any w-st abl e
present abl e t heory. So every uncoun t abl e mod el of T has a present at ion,
where th e points of th e involved t ree corr espond to strongly regular types.
Take any model M of T and any st rongly regu lar type p over M. We want
to associate wit h p a depth Dp(p) (an ordinal or 00) in a way inspired by th e
above assignment of a rank to the points of a tree. First we define Dp(p) a
for every ordinal a . As usual, we proceed by induction on a .
Definit ion 7.6.8 1. If a = 0, then Dp(p) a .
2. If a is limit, then we put Dp(p) a when Dp(p) f3 fo r eve ry ordinal
f3 < a .
3. If a = f3 + 1 is a successor f3, then we put Dp(p) a exactl y when, for
every x realizing p, there is some strongly regular type q over M (x)
such that Dp(q) f3 and q is orthogonal to all the types over M .
Now we ca n introduce Dp(p) .
Defi nition 7.6. 9 If there is an ordinal a such that Dp(p) 'i. a , then the
least ordinal with this property is a succes sor a o+1, and we put Dp(p) = ao .
Otherwise (when Dp(p) a for every ordinal a) we put Dp(p) = 00 .
Now we define the depth of the whole theory T Dp(T) . As usual, we agree
that 00 is great er than any ordin al.
Definition 7.6.10 The depth of T Dp(T ) is the least upper bound of the
depths of the strongly regular types over models of T. T is called deep if
Dp(T) = 00, and s hallow if Dp(T ) is an ordinal.
Actu ally t he original definition of depth requires some rearrangements due
to a technical convenience in th e proofs. But thi s does not concern our
sket ched treatment here.
At t his point it is easy to realize that the (pr esentable) th eor y T in 7.6.1
(and lat er in 7.6.7.2) is deep. In fact , let M be a model of T , a E Mo, p
be the strongly regul ar ty pe PM(l , a) (accor ding t o t he not ation introduced
in 7.6.1) . We have seen t hat , for every x 1= p, S(M(x)) cont ains a st rongly
regul ar ty pe q = PM( x)(l , x) having depth 0 and or thogonal t o all t he
non- algebraic ty pes over M . Consequently Dp(p) 1. But th en even q
has depth 1, and so Dp(p) 2. It erating this procedure, one eventu ally
gets Dp(p) = 00. Recall t hat T is not classifiabl e, because it gets too many
models in any uncountable power. But this is a general fact, as th e following
theorem clarifies.
Theorem 7.6.11 (Shel ah) 1fT is an cc -stable, presentable and deep theory,
then I (T , A) = 2>' fo r eve ry uncountable cardin al A. I n part icular T is not
classifiable.
This is not the case for a shallow T. In fact , first one shows t hat Dp(T ) < Wl
(and indeed, for every ordinal et < Wl , ther e is some w-stable present abl e
shallow th eor y TOt having depth et). Moreover , for every uncount abl e A, one
ca n upp erl y and underl y bound I (T , A) with respect to Dp(T) in an effecti ve
way, impl ying, among oth er t hings, I (T , A) < 2>' for some A. In t his sense,
we ca n conclude
Theorem 7.6.12 Let T be an w- st able theory. Th en T is classifiabl e if and
only if T is presentable and shallow.
Of course, t his statement makes sense provided t hat we agree wit h Shelah's
proposal t hat T is not classifiabl e if and only if I (T , A) = 2>' for every un-
countable Aj should we think t hat a classifiabili t y pr oof requires t o show t he
existence of an invari an t system of some ordinal rank, we might reasonabl y
d oubt t hat Theor em 7.6.12 is t he last word about classification, and wonder
if Shelah' s classification progr am is fully reached in t his t heorem. This is a
subtle and delicate matter, and may be discussed as long as one likes. So
here we limit ourselves t o a few comments which just aim at introducing a
possible debate and do not claim t o suggest any conclusion.
First one should und oubt edly acknowledge how for midable Shelah's work
was; in some sense t he "dept h" it self of t he new notions and tec hniques pro-
posed by Shelah wit ness it s validity and aut horitat iveness. Not surp risingly,
Shelah himself celebrated it s conclusion even in t he t it le of a preliminary
version of t he final pa per, which just stated "Why am I so hap py". Act ually
Shelah's happiness is quite easy to underst and and t o sha re.
It should be also mentioned t hat Melles proposed some years ago an alter-
native approach to classifiability of a mor e recurs ion t heoretic flavour, so
looking for effective class ificat ion invariants; however Melles himself proved
t hat his perspecti ve event ually agrees wit h Shelah's point of view. It should
be also said t hat Appenzeller (and others) observed some intriguing con-
nections bet ween t he main notions ar ising in Shelah' s classification, and
some concepts coming from Stat iona ry Logic and Descriptive Set Theory.
This is quite interesting, and provides a fur th er corroboration of Shelah 's
perspective.
Fi nally let us discuss what happens wit hin superstable (an d possibl y non
w-stable) t heories T. The main t rouble in t his enlarged framework is t hat
prime models may fail. However prime models (and existence and uniqu e-
ness properties) st ill make sense provided we restrict our attent ion to a
suitable subclass of M od(T), formed by sufficientl y sat ur ated models (called
a-models) . This allows t o define what T presentable, or T deep, or T shallow
means, and t o deduce t hat , when T is not pr esent abl e, or is presentable and
deep, T is not classifiable: in fact , in t hese cases T has too many a- models,
and consequent ly t oo many models, t o get a classification.
But t he point is whether a present abl e shallow superstable T is classifiabl e.
According to Shelah 's perspecti ve, t his requires to upperly bound t he num-
ber of models of a given uncountable power (warning: we have said models,
and not a-models, bounding th e latt er ones does not impl y a priori restrict-
ing the number of t he former ones). Again the main difficulty in handling
this sett ing is the possible lack of prime models, especially in t he framework
proposed in t he definiti on of pr esent abl e theor y. And actually Shelah singled
out t he following key propert y, called t he existence property.
Definition 7.6.13 A superstable T is sai d to satisfy the existence prop-
erty if and onl y if, for every choice of models A1
0
, A1
1
, ) \.1 2 of T such
that
Mo is an elementary substruct ure of both M
1
and M
2
,
M
1
tMo M
2
,
there is a model M of T prime and atomic over M
1
U M
2

This is t he last dichot omy in t he superstable case , and provides t he final
dividing line between class ifiable and non- classificabl e t heories. In fact t he
following resul t s hold .
Theorem 7.6.14 (Shelah) 1f T is superstable but fails to have the existence
property, then l(T , A) = 2'\ for every uncountable cardinal A.
Theorem 7.6.15 (Shelah) 1fT is a superstable, presentable, shallow theory
and has the existence property, then T is classifiable.
Of course, t he comments following Theorem 7.6.12 are still valid and do
concern also Theorem 7.6.15.
7.7 Shelah's Uniqueness Theorem
In the forth coming sections we discharge t wo debts we contract ed in Chapter
6: the proofs of Shelah 's Uniqueness Theorem for prime models in w-stabl e
7.7. SHELAH'S UNI QUENESS THEOREM 271
t heories and Morley's Theorem on uncount ably categorical t heories. In fact
bot h require, in addit ion to t he techniques developed in Chapters 5 and 6,
t he mor e sophisticated tools introduced in t his Chapter , in par ticular t he
material of Section 7.5 on w-stability. First we deal wit h Shelah's Uniqueness
Theorem. vVe recall its framework: we are concern ed wit h an w-stable t heory
T and a small subset X of it s uni verse n. We have t o show t hat , if A
o
and
A 1 are t wo models of T both elementarily including X and prime over X ,
t hen t here is an isomorphism between A
o
and Al fixing X pointwise. As
we saw in 6.4 , Ressayr e's Uniqueness Theorem for const ructible sets reduces
th e whole problem to prove:
Theor e m 7.7.1 (Shelah) Let T be an w- stable theory, X be a small subs et
of n, A be a model of T eleme ntarily including X . If A is prime over X ,
then A is constructible over X.
Proof. According to Morl ey's Existence Theorem (6.4.17) t here is some
model B of T constructible over X . Fix such a mod el B. We can freely
assume t hat X is a subset of B and t hat t he corres ponding inclusion is
elementary. As A is prime over X , t here is an eleme ntary embedding of A
into B fixing X point wise. Again, wit hout loss of generality we can ass ume
t hat A is an eleme ntary subst ruct ure of B. In par t icular X A B, and
all t hese inclusions are elementary. We claim that t hese ass umptions
X A B, B const ructible over X
(so for get ting t he additional hypothesis t hat both A and B are models of
T ) are sufficient t o impl y, for T w-stable, t hat A is const ructible over X , as
expected.
We proceed by induction on the power IB - XI of B - X. If IB - XI is
countable, then lA - XI is count able, too; furt hermore B is at omic over X
because B is constructible over X ; hence we can apply Lemma 6.4.6 and
conclude that A is const ructible over X, as claimed.
So ass ume IB - XI = A uncount able. Fix any const ruction (b
v
: t/ < A) of
B over X : her e A is viewed as an initi al ordinal. Correspondingly we build
a n increasing sequence of subsets C; (v < A) of B satisfying
1. B=XUUv<'\ C
v
and, for every v < A,
2. ICvl = Ivl +
3 . Cv is closed (in t he sense of Section 6.4) ,
4. for ever y t uple c in Cv, the type of c over A has t he same Morley rank
as it s extension tp(cjA UCv).
To for m this sequence, start by putting Co = 0. Moreover , for a limit u, let
C; be t he union of the preceding sets in the sequence. It is straightfo rward
t o check t hat 2, 3, and 4 are satisfied in these cases . Now take a successor
ordinal v + 1. If b; E Cv, then put C
v
+
1
= Cv. Otherwise form Cv U {by}
and call it Cv(O). For every tuple bin Cv(O) th ere is a finite subset A(b) of
A for which
RM(tp(bjA)) = RM(tp(bjA(b))) .
Take Cv(O) U Ub'A(b) and enlarge it to a minimal closed ext ension C
v
(1 ) in
B, in t he way we saw in Section 6.4. Apply t he same procedure used to
enlarge Cv(O) to C
v(l),
and then to Cv(n) for every natural n, so building
for every n a closed set Cv(n + 1) 2 Cv(n) in B such t hat, for each bin
Cv(n), there is a finite subset A(b) of A n Cv(n + 1) for which
RM (tp(bjA)) = RM(tp(bjA(b))).
Finally put C
v
+
1
= UnCv(n) . Again , a st raightforward check proves 2, 3
and 4 for Cv+!,
It is also clear that B = X UUv<"' C
v
and hence 1 holds. Now notice t hat,
for a given t/ < A,
5. Cv+! is constructi ble over X UCv,
The latter claim follows directly from 2, as IB-XI = A > 5 requires
some more work . First form a (possibl y transfinite) list of the elements in
C
v
+
1
- (X U Cv)
(dJ1- : J-L < a)
extracting t hem from the given const ructi on of B over X . Notice th at, for
ever y /3 < a , C; U {dJ1- : J-L < /3} is closed because both C; and Cv+! are
closed. So use Lemma 6.4.9 and deduce that the type of d(3 over XUCvU{dJ1- :
J-L < /3}is isolated. Hence {dJ1- : J-L < a} is a construction of C
v
+
1
over XUC
v
,
and 5 is proved.
Owing t o 5 and 6, t he induct ion hypothesis applies to
X Uc, X Uo, U (C
v
+
1
n A) X UC
v
+
1
7.8. MORLEY'S THEOREM 273
and shows that xUC
v
U (Cv+! n A) is constructible over x UC
v.
For every
v < x, fix a const r uction of X Uc, U (Cv+! n A) over X Uc. , and glue all
t hese constructions t o buil d a (t ra nsfinit e) list of t he elements of A
where, for every v , (all : fl v :S fl < fl v+d constructs C
v
+
1
n A over X U Cv.
We claim that (*) is act ually a construct ion of A over X: in oth er words, for
every fl , tp(a
ll
/ X U{aT] : 1] < fl}) is isolat ed. In fact , suppose fl v :S fl < fl v+l ,
it suffices t o show
In fact , the former type is isolated and consequent ly th e Open Mapping
Theor em ensur es t hat t he latter is isolated, too. Now, for every tuple c in
c.,
RM(tp(CjA )) = RM(tp(CjA n Cv))
by 4. Now we use what we saw in Section 7.5 about independ ence in w-stable
th eori es. First noti ce t hat t ra nsit ivity (16) implies
Then symmetry yields
As t his holds for every t uple cin Cv, (13) implies th at t he ty pe of all over
X UC; U{aT] : 1] < fl} does not fork over X U{a T] : 1] < fl} , and this is just
what we have claimed. So A is const ructible over X , and we are done. ..
7.8 Morley's Theorem
At last here are to show Morley's Theorem on uncountabl y categorical t he-
ori es. Let us recall it s framework: for a complete t heory T wit hout finit e
mod els in a count able language L , t he theorem says t hat T is categorical
eit her in all uncount able cardinals, or in none of t hem. In oth er words
Theor e m 7. 8.1 (Morley) 1f T is categorical in some uncountable cardinal,
the T is categorical in every uncoun table cardinal.
Just to prepare the proof of Morl ey's Theorem, let us premit a particular
case as a warm-up.
Proposition 7.8.2 If T is stmngly minimal, then T is cat egorical in every
uncountable power.
Proof. For every mod el B of T , t he isomor phism class of B is given by its
dimension with respect to the ad dependence relation. Wh en B is uncount-
able, t his dimension just equals th e power of B becau se B = ad(X) forces
IXI = IBI for every X B . ..
So the point is: how far is a t heory T categorical in some uncount able power
f-l from being st rongly minimal? Can we recover inside T a possibly "weaker"
for m of "st rong" minimality making t he previous argument work?
Not ice that there do exist some t heor ies T which are categorical in any
infinit e cardinal but are not strongly minimal: for inst ance, consider t he
theory of t wo infinite disjoint sets A and A' with a bijection f between
them.
Anyhow take a f-l-cat egorical T where f-l is a fixed uncountable cardinal. We
keep thi s ass umpt ion all t hroughout t his section, unless explicit ly st at ed.
First we observe:
Lemma 7.8.3 T is w-sta ble.
Proo]. The strate gy is simple. We deny w-st ability and consequently we
succeed in building t wo models of T of power f-l which cannot be isomorphic,
so contradicting the hypothesis that T is f-l-categorical. To do th is, first we
recall that , if T is not w-stable, th en there is a countable X n over which
there are unco untably many I-types. So th ere is some model of T extending
X and realizing uncount ably many types, and using t he Lowen heim-Skolem
T heorem we can get a model B of T sat isfying these propert ies and having
power f-l. The second part of t he proof uses a method due to Ehrenfeucht and
Mostowski and valid for every th eory T and every uncount a ble cardinal f-l : it
yields in t his general framework a model Cof T of power f-l realizing at most
I-types over any count able subset. Ehrenfeucht-Mostowski 's method is
quite sophisticated, and its int erest for our purposes in this book is confined
to this lemma; so we omit its detailed t reat ment, and we limit ourselves to
apply it to our par ti cular set t ing. In fact, for a f-l-categorica l T, t he model
C just built in t his way cannot be isomorphic to the structure B obtained
before. ..
Of course , Lemma 7.8.3 ca nnot imply t hat T is st rongly minim al. But , as
a consequence of t he w-stability of T, indeed as a consequence of this only
hypothesis (we do not need fl-categoricit y here) , we show t hat t here is some
strongly minimal formula -possi bly involving paramet ers- in T .
Lemma 7.8.4 Th ere is some strongly minimal formula cp (v ) in T.
Just t o avoid any misunderst anding lat er , let us und erline once again t hat
here formul a means for mul a with paramet ers.
Proof. Oth erwise, for every formula cp(v) for which cp( n) is infinit e, one can
find a formul a cp' (v) such t hat cp' (n) cp (n ) and both cp' (n) and cp( n )_ cp' (n)
are infinite. Using t his fact and start ing from CPO (v) : v = v, one defines ,
for every finit e ord ered sequence s of 0 and 1, a formul a CPs (v) in such a way
t hat , for any s , cps(n) is infinit e and cpso(n ) and cps1(n ) parti tion cps (n ).
T he paramet ers involved in these formulas are as many as t he sequences s ,
and so form a countable set X . On t he other side, every bra nch t of t he t ree
{a, 1}wdet ermines a( n incomplet e) l-Lype Pt over X given by t he for mul as
CPt ln (v) wher e n ranges over naturals and t in denot es t he restriction of t
to t he first n naturals; furthermore different br anches yield different ty pes.
This gives at least 2
No
I- t ypes over X and contradicts w-st ability.
But fl-cat egorici t y implies even more. In fact , for a fl-cat egor ical T , t he
formul a cp (v) ca n be chosen wit h paramet ers in t he model A of T pri me over
o(owing to w-stability this models exists and is unique up to isomorphism).
This result requires the following premise, which is now just a t echnical
lemma, but will play a key role lat er in t he proof of Morley's Theorem.
Lemma 7.8.5 Let a( v, il) be a fo rmula (wi th paramet ers ii in n) such that
a( n , ii) is infini te, and let B
o
and B
l
be two models of T such that B
o
is an
elementary substruct ure of B
l
, B
o
=1= B, and both B
o
and B
1
elementarily
include ii. Th en a(B
o
, ii) is properly included in a(B
1
, ii).
Proof. Suppose not. Accordingly t here are t wo model s B
o
and B
1
of
T elementarily including ii such t hat B
o
is an elementary substr ucture of
B
1
, B
o
=1= B
1
but a(B
1
, il) equals a(B
o
, il) (an infinit e set ) . Apply t he
Lowenh eim-Skolem Theor em t o t he t heory of (B
1
, B
o
) in a language ex-
t ending L by a l-ary relat ion symbol for B
o
and new const ants for t he
elements in ii, and get a countable model (B'1 , Ba) of t his t heory: so Br
and Ba are still models of T elementarily including ii, B
o
is an elementary
substruct ure of Bi , B
o
=1= Bi , a(Bi , il ) = a(B
o
, ii), B
o
,e; and consequent ly
a(Bi, il ) = a(B
o
, ii) are countable. Hence, unless replacing B
o
and B
1
by B
o
and Br resp ecti vely, we can ass ume t hat a(B
l
, if) = a(B
o,
if) is count able.
Now we define for every or dinal t/ > 0 a model B
v
of T elementarily extend-
ing Ba (and if) and even t he Bp's for p < u, properly including all of them
and stilI satisfying a(B
v
, if) = a( B
o
, if) . Notice that this is sufficient for our
purposes because, pr oceedi ng in t his way, we event ually build a mod el B of
T of power f-l elementarily extending Ba but satisfying a( B, if) = a( B
o
, if),
hence admitting a countable a( B, if); on t he ot her side , a simple use of
Compactness Theorem yields anot her model C of T of power f-l such t hat ,
for every t uple if in C admitting t he same type as if over t he empty set ,
la (C, 01= f-l . Conseq uently B and C cannot be isomorphic, and t his con-
tradicts t he f-l-cat egoricit y of T .
So let us build the Bv's. vVe alr eady introduced B
l
. For a limit v, put
B; = Up<vBp: t his is a model of T and satisfies our condit ions above. So it
remains to treat the case of a successor or dinal v = p + 1 with p > O. For
simplicity we limi t ourselves t o p = 1, v = 2; indeed what we are going to say
in t his case applies to any p > 0 as well, and so is generally valid. First use
ind ependence t heory, more precisely (13) and (14), and find an isomorphic
copy of B
l
inside n, corresponding t o B
l
by an isomorphism fixing Ba
poin t wise, and satisfying ../.-Bo Bi . To buil d consider a language L*
enlarging L by a constant b* for every b E B
l
, and in L * t he t heory T * saying
that , for every bin B
l
,
b* satisfies t he non-forking extension of tp(bj Ba) over Bi .
Any fini t e por tion To of T* has a model; in fact , let bglue all t he t upIes from
B
l
arising in t he sent ences of To, use (14) and obtain a t upIe b* realizing t he
non-forking extension of tp(bj Ba) over Bi ; recall that every subsequence of
b* has t he same propert y. By compactness , T* has a model. The elements
b' embodying t he constants b* in t his model form a st r uct ure isomorphic
t o B
l
over Ba (as, for every L (Bo)-formul a <p (v) and bin B
l
, F <p (b) if and
only if F <p (b' )) and independent of Bi over Ba. In particul ar if) =
a( B
o,
if). Let B
2
be t he model of T prime over e, U We claim th at
B
2
is just t he model we are looking for. The key point to check is t hat
a( B
2
, if) = a( B
o,
if). 2 is clear. On t he other hand, take d E a( B
2
, if). Then
t he type of d over B, is isolat ed by some suitable for mula b1 , z) (in
t he free variable z) wit h par amet ers from e, and b1 fro m Recall t hat
Ba) is definabl e: for every L-formula X(v, w) t here is an LBo-formul a
dX(w) such t hat, for any bin Ba,
n F b) <=? n F dx(b).
This remains true if we enlarge Ba to U {d} : for every bin U {d} , so
also for b= (iJi, d) with iJi in
n F= iJi , d) {:? n F= dX(iJi , d).
Let us see why. Oth erwise n F= iJi , d) H dX(iJi , d)) and consequentl y
n F= 3z(a( z, ii) 1\ b', z) H dx(b' , z)));
in other words,
3z(a(z, ii) 1\ -' (x(v, iJi , z) H dx(iJi , z))) E BD.
As B, and are independent over Ba , there is some b
"
in Ba for which
n F= :3 z(a(z, ii) 1\ b
"
, z) H dx(b
"
, z))).
Then t here is some d" E B
l
(and indeed in a(B
l
, ii) ) such t hat
8
1
F= b
"
, d") H dX(b
"
, d")).
Recall a( B
l
, ii) = a(B
o,
ii), so d" E Ba. But t his cont radicts t he choice of dX
over Ba. Now apply what we have j ust observed to t he formul a 1/; (v, d).
As n F= d), it follows n F= d), where d1/; is t he defining for-
mula. So F= 3zd1/; (b1 , z), and we find d' E such t hat F= d1/; (b1, d'),
and hence n F= b1 , d'). This means that d' has t he same ty pe as d over
B
l
U and so d = d' E but ii) = a(B
o
, ii), hence d E a(B
o
, ii) .
This accomplishes our proof. ..
Now we are in a position to show, as promised
Lemma 7.8.6 Let A be a model ofT prime over 0. Then there is a strongly
min imal f ormula ep (v) in T with parameters from A .
Proof. Proceed as in Lemma 7.8.5, but work in A instead of n. Using
w-stability, find again a formul a ep (v) with paramet ers in A such t hat ep(A)
is infinit e but has no par ti t ion into 2 A-definable infinit e subsets. This does
not mean a priori t hat ep (v) is strongly mini mal , as in general we cannot
imagine what happens if we allow paramet ers out of A. However we claim
t hat in our parti cul ar setting, for a JL-categorical T, ep( v) is just st rongly
minimal. Let us see why. Suppose not , so, enlarging our perspective to
n, we can find an L-for mul a 1/; (v, Z) and par amet ers b in n such that both
<p(v) 1\ 'IjJ (v, b) and <p(v) 1\ -,'IjJ (v, b) define infinit e sets. Now let us restrict
our horizon t o A: for every natural n, we can pick a t uple ii(n) in A for
which
1<p(A) n 'IjJ (A, ii(n)l , 1<p( A) - 'IjJ (A, ii(n) I n.
However , just owing t he choice of <p( v), at least one of these sets is fi nit e,
and consequent ly equals <p (n) n 'IjJ (n, ii(n)), or <p (n) - 'IjJ (n , ii(n)) . So eit her
<p(A) n 'IjJ(A, ii(n)) is finite and equals <p (n ) - 'IjJ (n , ii(n)) for infinit ely many
n, or <p (A) - 'IjJ (A, ii(n)) satisfies t he same cond it ion. Assume for simplicity
that the former option holds. Unless forget ting ii(n) for every except ional n,
we can even ass uming t hat <p (A) n'IjJ (A, ii(n)) is finit e of size n and equals
<p (n ) - 'IjJ (n , ii(n)) for each n. Now a simple application of Compactness
Theor em, using these features of A, n and t he ii(n)'s, yields two models B
and C of T and a tuple ii in B such th at B is an elementary substruct ure of
C, B =/: C and
<p (B) n 'IjJ (B, ii) = <p (C) n 'IjJ (C, ii)
is infinit e. T his clearl y cont radicts Lemma 7.8.5. ..
But we can even ass ume t hat our st rongly mini mal <p(v) is actually an
L-formula, and needs no addit ional paramet er s. Let us explain why. Esse n-
t ially what we have to do is to inser t in t he language new constants for t he
paramet ers ii E A involved in <p (v) and t hen, in th is extended fra mewor k,
to consider T' =Th(A, ii) inst ead of T = Th (A). Notice t hat t he ty pe of ii
over 0 is isolat ed - say by t he L-formula O(w) - because A is atomic over 0,
so tp(ii/0) is realized in every model of T and indeed t he models of T' are
just t he struct ures (B, b) where B is a mod el of T and bE B has t he same
type as ii over 0; in oth er words, bsat isfies O(W) in B. In this set t ing it is
not prohibitive t o show:
Lemma 7. 8.7 For every infi nite cardinal A, T is A-categorical if and only
if T ' is.
The reader may t ry to check t his as an exercise. In parti cul ar , if T is J.L-
categorical, t hen T' is; and, if we succeed in proving t hat T' is categorical
in every uncountable power A, th en we can say t he same of T . So, wit h
no loss of generality, we can repl ace T by T', in other words to assume
t hat T is a J.L-cat egori cal t heory wit h a strongly mini mal L-formul a <p( v)
(wit hout paramet ers). This resembles t he plainer framework outlined at t he
beginning of this section, when T it self is strongly minimal. So t he point
is: for a model B of our t heory T, can the dimension of <p(B) with respect
7.9. BIINTERPRETABILITY AND ZILBER CONJECTURE 279
to ad, and so (at least in uncount able powers) t he cardi nality of <p (B), play
the same role as the dimension of B in t he strongly minimal case?
At t his point Lemma 7.8. 5 applies once again, and ensures t hat <p (B
o
) is
properly included in <p (B
1
) when B
o
and B
1
are models of T , B
o
is an ele-
ment ary substructure of B
1
and B
o
=1= B
1
But now we can say much more.
In fact we have what follows.
Lemma 7.8.8 For every model B of T , B is prime over <p(B). Moreover
1<p(B)1 = IBI
Proof. As T is w-stabl e, t here exists a model B' of T pr ime over <p (B).
B' element arily embeds into B via a function fixing <p (B) point wise; again,
we can ass ume t hat B' is an element ary subst ruct ure of B. Consequent ly
<p(B' ) = <p (B) n B' ; on the other side <p (B' ) <p (B), and so <p (B' ) is equal
to <p (B). Apply Lemma 7.8.5 and deduce B = B'. Hence B is prime over
<p (B). It remains t o check t hat 1<p(B)1 = IBI. ::; is trivial. On the other
hand an easy application of t he Lowenheim-Skolem T heor em t o t he t heory
of Bc.p(B) yields a model C of T such t hat ICI = 1<p(B)1 and <p (B) G by an
element ary inclusion . As B is prime over <p (B), B elementarily embeds in C.
In particular 1<p(B)1 = IGI ;::: IBI, as required.
At last we are in a position to conclude ou r way.
Proof. (Morley's Theor em) . Let B
o
and B
1
be t wo models of T having
the same uncount abl e power ).. . Owing to Lemma 7.8.8, each B, (i = 0, 1)
is prime over Then X
o
and Xl correspond to each other by some
elementary biject ion , which enlarges t o an eleme ntary bijecti on h between
<p (B
o
) and <p (B
1
) ; h defines in it s t urn an isomorphism between t he models
B
o
and B
1
prime over <p (B
o
) and <p (B
1
) respectively.
7.9 Biinterpretability and Zilber Conjecture
We have devoted several sections t o t he problem of classifying struct ures
up to isomorphism and to Shelah's analysis of this question. But , needl ess
to say, isomorphism is not the onl y possible classifying equivalence rela-
tion, even within structures. Anot her possible criterion, deeply related to
Model Theory, is interpret abili ty. The following t wo examples illustrate t his
alternative perspective and it s underl ying idea.
Examples 7.9.1 1. Nat ur al numbers (viewed as non-negati ve int egers)
form a definable set in t he ring (Z, +, .) of int egers: as recall ed in
Chapter 1, a celebrated Theorem of Lagr ange says t hat they are ex-
actly t he sums of four squa res in (Z, +, .). So t he whole struct ure
(N, +, .) is definabl e in (Z, +, '), because t he addit ion and mul tipli -
cation in N ar e just t he rest rict ions of t he corres ponding operations of
Z. This is a fund ament al resul t: in fact , as (N, +, .) lives in (Z, +, .)
as a definable st ruct ure , (Z, +, .) inheri t s it s undecidabili t y phenom-
ena related t o God el Incompleteness Theor ems, and in this sense is a
"wild" st ruct ure.
2. In t he same way (N, +, .) lives in the rational field (Q, +, .) as a
definable structure. This is a deep theor em of Julia Robinson. So
even the rational field inherits th e complexity of (N, +, .) and it s
undecidabili ty.
So, genera lly speaking, when we meet a st ruct ure A in a language L and we
realize t hat A defines, or also int erpret s anoth er struct ure A' (possibly of a
different language L' ), t hen we can reasonabl y agree t hat A inherit s th e full
complexity of A' , and consequent ly is at least as difficult to dominate as A'
is. Of course t his ca n be extended t o classes of st ruct ures. In t his enlarged
framework , we compare two classes of struct ures, K in a language L and K'
in a language L' respecti vely. For simplicity, we can agree t hat both K and
K' are elementary. We ass ume t hat t here ar e suitable L-formulas defining,
or also int erpreting, in any struct ure A E K a st ruct ure A' E K' and t hat,
conversely, every A' E K' ca n be recovered by some A E K in t his way;
we ass ume also that t hese formul as do not depend on t he choice of A in
K . Then we say t hat K int erpret s K' and in t his case we can agree that K
inheri t s t he complexity of K' . Here are some further examples illustrating
this point.
Examples 7.9.2 1. Recall that a graph is a st ruct ure (G, R) where R
is a symmet ric irr eflexive binary relation , usually called adjacency.
The (elemen tary) class of graphs int erpret s any class of st ruct ures,
and so inherit s in t his way t he full complexity of math emat ics. The
proof of thi s fact requires patience rath er t ha n ingenuousness. To
avoid t oo many tedious det ails, let us illustrat e its idea in a par t icular
case, and see how graphs int erpret arbitrar y binar y relation s. So take
any struct ure (A, R) where R is a binary relation on A, and form
a graph (A', R' ) as follows. Let A' include A. Moreover, for every
a E A, add t wo new vertices ao and al in A', both adjacent t o a (so
(a, ao ), (a, ad E R' ). Finally, for every pair e = (a, b) E R, add
in A' t wo new ver tices eo, el sat isfying (a, eo), (eo, el ), (el , b) E R' ,
t hree more ver tices adjacent to eo, and four mor e vertices adj acent t o
el . For inst ance, here is a pictur e of (A', R' ) when A = {a , b, c} and
R = { (a, b)} .
b
o
A
Co
It is an easy exercise to realize that any (A, R) can be definably re-
covered inside t he cor res ponding (A' , R' ) in a way independent of t he
particul ar choice of (A, R). In fact , A is just t he set of t he ver tices
in A' having eit her 2 adjacent nod es, or 3 adjacent nodes such t hat 2
of t hem have no fur th er adjacent nod e, while R is t he set of t he pairs
(a, b) E R such t hat t here are eo, el E A' for which (a, eo), (eo, el)
(el, b) E R' , eo has t hree adjacent nodes in addit ion to a and el has
four adj acent nodes besides eo and b.
2. The same can be said of t he class of groups, and even of the class of
nilpoten t groups of class 2 (a compa rat ively slight generalizat ion of
abelian groups) . This is a beautiful result of A. Mekler , showing that
nilpotent groups of class 2 int epret graphs and so, t hrough them, any
class of structures. The proof uses brilliantly some non-trivial notions
and tools from group th eory. The conclusion is clear: groups, and even
nilpotent groups of class 2, are a class as bad as possible, and inherits
t he full complexity of mathematics.
3. We said in 7.1 that , for a given (count able) field K, K(x, y)-modules (i.
e. K-vectorspaces with t wo distinguished endomorphisms x and y) ar e
an intractable class: no classification can be expected, even for finit e
dimensional objects, otherwi se the word problem for groups would be
solvable. Of course, every class interpretin g K(x, y)-modules is at least
as complicat ed as t hey are, and hence definiti vely a bad class. Several
not abl e classes of modules share t his negative feature. For instance,
t his is t he case of K[x, y]-modules (wit h t wo commuting unknowns x
and y), or Z[x]-modules, and so on. The book of Prest quoted in t he
references at t he end of t he cha pter includes a discussion of t his point
and a great deal of not ewort hy examples.
So a possible way of classifying st ruct ures, or even classes of st ruct ur es, is up
t o mutual interpretabilit y (biinterpretability) . Accordingly, one could t ry t o
characterize struct ures by looking at what is definabl e, or also int erpretable,
in t hem: gro ups, fields, and so on.
Incident ally notice t hat a relevant emphas is on t he role of definabilit y al-
ready arises within Shelah 's classification analysis (for instance, t hink of th e
ord er property, lookin g at t he orders definable in a given struct ure). How-
ever the st udy of mutual int erpretability in mathematics did pr ecede Model
Theory, or , at least , mod ern Model Th eory. For example, let us mention
t he celebrated Malcev correspondence between groups and rings, essent ially
showing t hat t he class of uni t ary rings is biint erpret able with a suitable class
of nilpot ent groups of class 2, and confirming in this way how complicated
t hese groups are.
But who mainly developed t he biint erpretabilit y program in Model Theory
was Boris Zilber. Zilber 's original project concerned t he class ification of
uncount abl y categorical t heories (t hose where Morley's Theorem applies)
by looking at which gro ups, or fields , and so on, are definabl e in t hem.
As already observed, uncoun t abl y categorical t heories include the strongly
minimal ones, and t he lat t er t heories are t he simplest possible (if we excludes
finit e struct ures) . Accordingly t heir exam is a reasonabl e first step t owards a
general approach. So t he question is: what strongly minimal struct ures look
like? Can we reason abl y classify t hem by lookin g at what t hey int erpret ?
Let us outline Zilber analysis in t he strongly minim al setting. Recall t hat
a st ruct ure A is said to be st rongly minimal if it s complete theory is: so,
for every B element arily equivalent to A, the only definable subsets of B
are t hose finit e or cofinite . We introduced in Cha pte r 5 several examples of
st rongly minimal struct ures . To summarize them, let us work inside a fixed
algebraically closed fi eld A.
Examples 7.9.3 1. Firstl y view A as a struct ure in t he language 0, so
as a mere infinite set . In t his case , t he t heory of A equals t hat of
infinit e sets, and is strongly minimal. T he st ruct ure of A is very poor ,
and no (infinite) gro up is definable here. Moreover
for every subset X of A, acl (X ) = X ,
7.9. BIINTERPRETABJLITY AND ZJLBER CONJECTURE 283
for every positive int eger n, the definable subset s of An are the
finit e Boolean combinations of
with i , j = 1, .. . , n, bE M .
2. Let Aa be t he pr ime subfield of A, and look at A as a vectorspace over
Aa in the appropriate language. Now t he complete theory of A is t hat
of infinite vectorspaces over Aa, and is again strongly minimal. But
t his time
for every subset X of A , acl(X) is the subspace of A spanned by
X,
for every positive integer n, the definable subsets of An are t he
finite Boolean combinat ions of cosets of pp-definable subgroups of
A, and for n > 1 their class is larger t han in Example 1. Notice
that an infinite group is trivially definable in A, but no field can
be int erpret ed inside A.
3. At last , view A just as an algebraically closed field. If p is its character-
istic, then the complet e theory of A is ACF
p
and is strongly minimal.
Moreover
for every subset X of A, acl(X) is t he algebraic closure of A in
the field theoretic sense,
for every positive integer n, t he definabl e subsets of An are the
construct ible ones , in other words t he finite Boolean combinations
of algebraic varieties of A n.
Now take any strongly minimal st ruct ure A.
Definition 7.9 .4 A is called trivial if, for every X A,
acl (X) = UxEXacl(X) .
Every (pure) infinit e set is trivial. But, of course, vectorspaces and alge-
braically closed fields are not. Moreover no trivial st rongly minimal struc-
ture can interpret an infinite group.
Definition 7.9.5 A is called locally modular if, for eve ry choice of X ,
Y A suc h that X n Y i= acl (0),
(*) dim (X UY) +dim (X n Y) = dimX +dim Y.
Every t rivial struct ure A is locally modular (in fact , for every X and Y,
acl (X UY ) = acl (X) U acl (Y), so a basis of X U Y can be formed by taking
a basis a X n Y , ext ending it t o a basis of X and a basis of Y, and gluing
t hese bases together). But also vect orspaces are modular: in fact , in t his
case , (*) is just t he Grassman formula (and does not need t he ass umpt ion
X n Y :f: acl (0)) .
On t he cont rary, no algebraically closed field A (of t ra nscendence degree
2': 4) is locally modular. In fact, choose ao, aI , a2, a3 E A algebraically
independent over the prime subfield Aa , and form th e extensions
Then dim X = dim Y = 3 but dim(X n Y) = dim Ao(aa) = 1, whence
X n Y :f: acl(0) , and dim(X U Y) = dim(Aa(ao , aI , a2, a3)) = 4. So (*)
fails.
What is t he significance of local modularit y? Basi cally a locally modular A
eit her is trivial or can define an infinite group. Furth ermore one obser ves
th at any group 9 definabl e in a locally modular A is abelian-by-finite (in
other wor ds , it has a normal abelian subgroup of finit e index); and every
subset of any cartesian power endefinabl e in A is a finit e Boolean combi-
nation of cosets of definable subgroups of en. Accordingly, no infinit e field
is definabl e in A.
In t his setting Zilber raised in 1984 t he following problem, generally called
Zilber Trichotomy Conject ure.
Conjecture 7.9.6 (Zilber ) Let A be a st rongly minimal non locally modular
struct ure. Then A interprets an infi ni te fi eld K . Furthermore, for every
posit ive integer n, the subsets of K " definable in A are ju st thos e definable
in K (and hence coincide with the construciible ones).
Recall that , owing to Macintyr e's Theor em, any infinite field interpretable
in an w-stabl e st ruct ure must be algebraically closed. Hence the importance
of this conj ecture is clear: according to it any st rongly minimal st ruct ure A
eit her is t rivial, and so looks like an infinite set (as in Exa mple 7.9.3.1) , or is
locally modular and not t rivial, and t hen resembl es a module (as in Exam-
ple 7.9.3.2), or looks like an algebraically closed field, because it interpret s
such a field (Example 7.9. 3.3). Hence t he conj ecture would provide a qui t e
sat isfact ory classification of strongly minimal st ruct ures (and t heories) up to
biint erpret abili t y. But in 1993 Hrushovski showed t ha t Zilber 's Conjecture
is false.
Theorem 7.9.7 (Hrushovski) There do exist strongly minimal structures
A which are not trivial but cannot int erpret any infinite group.
Clearly such a struct ure A is not locally modular and does not interpret any
infinite field.
However Zilber Conjecture (more precisely, a suitable restatement) does hold
in certain topological structures deeply related to st rongly minimal models:
the so called Zariski geometries. To introduce them, let us come back to
Example 7.9.3.3, so dealing with algebraically closed fields A.
We know that , for every positive integer n, the algebraic varieties of An are
preserved under finite union and arbit rary intersection, and form the closed
sets in the Zariski topology on An. These topologies are Noeth eri an: none of
them admits any infinite strictly decreasing sequence of closed sets. More-
over they sat isfy th e following properties (m and n denote below positive
int eger s) .
(Zl ) Let f = Ut, ... , f m) be a function from An in Am. Assume that each
component f i (1 i m), as a function from An in A , either projects
An onto A or is constant . Then f is continuous.
( Z2 ) Every set {iI EA : a; = aj} with 1 i , j n is closed.
(Z3) The projection of a closed set of An+! onto An is a constructible set
in An.
(Z4) A, as a closed set, is irr educible.
(Z5) Let X be a closed irreducible subset of An. For every ii E An-I , let
X (iI) denot e th e set of the elements b E A such that (iI, b) EX. Then
there is a natural N such t hat, for every ii E An -I, either IX (iI)I N
or XCiI) = A. In particular, when n = 1, every closed proper subset
of A must be finite.
(Z6) Let X be a closed irreducible subset of An, d denote the (topological)
dimension of X. Then, for every i , j among 1, . . . , n, X n {iI E An :
ai = aj} has dimension 2 d - 1.
(Zl) - (Z6) restate in a topological style some properties which are well
known, or easy to check. For inst ance (Z5) follows directly from some simple
algebraic facts and the st rong minimality of A by a compactness argument:
the reader may check this in detail as an exercise.
Now it is easy to realize that even infinite set s and vectorspaces over a
countable field satisfy (Zl) - (Z6) provided one takes as closed subsets the
finite Boolean combinations of t he following sets:
the set s of the tuples admit t ing a fixed coordinate in a given place,
or equal coordinates in two different places, when dealing with pur e
infinite sets;
the coset s of pp-definabl e subgroups when dealing with vectorspaces.
It is easy to control t hat in both cases t hese sets ar e act ually t he closed sets
in a suit able topology.
Definition 7.9.8 A Za riski st ructure (or geometry) is a collection (A,
{T
n
: n positive integer}) where A is a non-empty set, for every n T
n
is a
Noetherian topology on An and (Zl) - (Z6) hold.
Hence the examples 7.9.3 produce Zariski structures. Conversely, let (A, {T
n
:
n positive integer} ) be any Zari ski structure. Assume t hat A is t he domain
of some st ruct ure A (in a language L) such that, for every positive int eger n,
the subsets of An definabl e in A are j ust the finit e Boolean combinat ions of
closed sets in T'; (and so coincide with th e const ruct ible sets in T
n
) . Then it
is easy t o check that A is st rongly minimal; mor eover t he possible triviality,
or local modulari t y of A does not depend on L , or on t he L-structure of
A, but only relies upon th e charac terizat ion of th e definabl e sets of A and
so, afte r all, upon (Zl ) - (Z6) . More notably, in t he restricted framework
of Zari ski st ruct ures, t he Zilber Trichotomy Conjecture is t rue, as shown by
Hrushoski and Zilber himself.
T heor em 7.9.9 (Hr ushovski-Zilber) Let (A, {T
n
: n posit ive integer}) be
a Zariski struct ure, A be a strongly minimal st ructure with domain A such
that, for every posit ive integer n , the subset s of An definable in A are ju st
the cons iructible sets in T
n
. If A is not locally modular, then A interpret s an
algebraically clos ed fi eld J( , and J( is unique up to definable isomorphism.
Moreover, for every positive integer n , the subsets of K " definable in A
coincide with the ones definable in J(.
7.10 Two algebraic examples
Let us summarize briefly some of the main notions introduced in t his chapter
by examining t wo relevant classes of algebraic struct ures, and their first
order th eori es: differentially closed fields of charact eristic 0, and existent ially
closed fields with an automorphism (again in charact eristic 0). Both these
examples play an important role in the mod el t heoretic solution of some
not abl e questions of Algebraic Geomet ry: we will describe t hese probl ems
and t heir solut ion in t he next cha pter.
7.10. TWO ALGEBRAIC EXAMPLES 287
1. DC Fa . First let us deal wit h differenti ally closed fields of characteris-
t ic O. Let us recall once again t hat t heir t heory DC F
o
is complete and
quantifier elimina ble in it s natur al language L, containing the sym-
bols +, " - , 0, 1, D and nothing mor e. So definabl e sets are easy t o
classify: as we saw in Chapter 3, t hey include t he zero sets of (finite)
systems of differenti al polynomials - in oth er words, t he closed sets in
t he Kolchin t opology - as well as t heir finit e Boolean combinat ions -
t he const ruct ible sets in t his topology -, but nothing else. As a ty pical
Kolchin closed set in a different ially closed field (K, D) let us mention
t he field of const ant s C(K) = {a E K : Da = O}. This is an alge-
braically closed field - just as K -, and is strongly minimal even in L;
in fact , D is identically 0 on C(K) and so adds no fur ther definabl e
objects to th e field st ruct ure on C(K).
DCFa eliminates the imaginaries, t oo. Moreove r DCFa is w-st abl e
with Morley rank w, so independence makes sense in DC Fa , and indeed
it is ruled by Morley rank: for ii , A and B as in Section 7.2,
s+A B {:::=;> R M (tp (il/ B )) = R M (tp (il/ A )).
Of course t his raises t he question of characterizing algebraically Mor-
ley ra nk wit hin differenti ally closed fields of cha racterist ic O. But t here
are also oth er ways of describing forking and inde pendence in DCFa,
having a pret t y algeb raic fl avour. For inst ance, one can pr eliminar-
ily obser ve t hat , for every small A , acl (A) - in t he model t heoret ic
sense - is just t he (field t heoretic) algebraic closure of t he differen-
t ial subfield generated by A Q(Dia : a E A , i EN); at t his point ,
one can realize t hat, for ii, A and B as usual , ii +AB just means
t hat acl(AUil) and acl(B) are (algebraically) independ ent over acl( A).
Among other things, this cha racteri zation sugges ts an alternative rank
notion, specifically concerning th e differential fr amework: this is called
differential degree or D-degree and denoted D-dg: for H a differential
field and ii a tuple of elements in n, D-dg(il/H) is th e t ranscende nce
degre of the differenti al fi eld generated by H U ii over H . So, for ii, A
and B as before, and A and B differential subfi elds for simplicity,
il +AB {:::=;> D-dg(il/B) = D-dg(il/ A ).
However we have t o be ca reful here: t he last equivalence does not mean
t hat R M and D-dg coincide. Their relationshi p, an algebraic cha r-
acterization of R M in DC Fa and t he connection among differenti al
degr ee, Morl ey rank and other possible ranks in DCF
o
are described
in the refer ences ment ioned at the end of th e chapte r.
Now let us deal with biinterpret ability, in particular let us consider
st rongly minimal sets in differ entially closed fields IC of charact eristic
O. They incl ude the cons tant subfield C(IC) (which also has different ial
degree 1, as it is easy to check). C(IC) is not locally modular, in
fact the argume nt proposed in t he last sect ion for algebraically closed
fields applies to C(IC) as well. But what is most remarkabl e in this
sett ing is a theorem of Hru shovski and Sokolovic sayi ng that Zilber
Trichot omy Conjecture holds withi n strongly minimal sets in DCF
o
.
In fact all t hese sets are Zariski struct ures, and so obey the Hrushovski-
Zilber Th eor em. We can say even more: any strongly minimal set S
which is not locally modular, and hence interpret s an infinite field, does
interpret t he field of constants C(IC) up to a definable isomorphism.
We will provide more details about t hese mat te rs in Section 8.7.
2. ACF A. Now we deal with existe nt ially closed fields with an automor-
phi sm. For simplicit y, we still work in characterist ic o. Let ADFA
o
de-
note the corresponding theory in the natural language
L = {+, ., - , 0, 1, O" } where 0" is the symbol representing the au-
tomorphism. Recall that, t his ti me, fixing t he characteristic is not
sufficient to ensure completeness: in order to characterize a model
of ACFAo up to element ary equivalence, one has also to describe
the act ion of t he aut omorphism on the prime subfield Q. Moreover
ACF A
o
does not eliminate t he quantifiers in L, alt hough it is obvi-
ously model complete (as a model companion) . Accordingly definabl e
sets exhibit some more complications than in the differential case. In
fact , t hey include the zero sets of (finite) syst ems of difference polyno-
mials, as well as their finite Boolean combinat ions; the former are the
closed set s, and t he lat t er t he constructible ones in a suitable topology.
But now, as quantifier eliminat ion fails, we have to consider also the
projections of const ruct ible sets - and nothing else, owing to model
completeness - to capt ure the whole class of definabl e sets. An exam-
ple of a closed set in a model (IC, 0") of ACF A
o
is it s fixed subfield
Fix (O" ) = {a E I< : er(a) = a}. This is not algebraically closed (in
particular it is not st rongly minimal); but one can see t hat it is a
pseudofinite field, so an infinite model of the theory of finite fields.
ACF A
o
elimi nat es the imaginaries. This time no existentially closed
field (IC, 0") with an aut omor phism is w-st able , or even st able. How-
7.11. REFERENCES 289
ever (K, a) is simple (as well as it s fixed subfield, and any pseudofinit e
field) . So independence makes sense in ACFAa, and comes directl y
from forkin g, but cannot be ruled by Morley ran k. Any how an explicit
algebraic characterization ca n be don e as follows. We work for sim-
plicity in a big sat urated model (K, a) of ACF Aa. First one obser ves
t hat , for every small A, acl(A) coincides wit h t he algebraic closure -
in t he field t heoret ic sense - of t he differ ence subfield generated by A
Q (a i (a) : a E A, i E Z)
(here we use t he cha racterist ic 0 assumpt ion; prime cha racterist ics
cause some major trouble). At thi s point one shows th at, for ii, A
and B as usual , ii t A B just means that acl(A U il) and acl(B) are
(algebraically) independent over acl (A). This yields an appropriat e
notion of rank, of a pret t y algebraic flavour , called difference degree
or a-degree and denot ed o-dq: for H a difference fi eld and ii a t uple of
elements in n, a -dg(il/ H) is t he t ra nscendence degree of t he difference
fi eld generated by H U ii over H . When finite, t he difference degree
can reason abl y repl ace Morley rank and provides a good noti on of
dimension in t his unstable setting; on t he oth er side , when a-dg(il/ H)
is infinit e, t hen clearly t he a
i
(a)'s (when i ra nges over int egers) are
algeb raically independent over H. So, for d, A and B as before, and
A and B difference subfields for simplicity,
ii tA B {:::::::? a - dg (il/B) = a - dg (il/A)
at least when t he lat t er degrees are finit e. Wh en X is any definabl e set
in f{ n , the difference degree of X over H a - dg (X/H) is t he maximal
difference degree of a t uple ii in X over H . In particul ar the fixed
subfield of K get s difference degr ee 1. In this sense, Fi x(a) is a "mini-
mal" definabl e infinite set of K. Notably, an adapted version of Zilber
Tri chotomy Conj ecture holds for these "minimal" sets in ACFAa, and
even ensur es in part icular t hat, very roughl y speaking, Fix (a) is t he
only non "locally modul ar" exam ple among t hese structures.
7.11 References
The classification issue from t he point of view of Descripti ve Set Theory
is treated in [68]; t he par ticul ar and int riguing case of torsionfree abelian
groups of finit e rank is dealt wit h in [163]. Finit e dimensional vect orspaces
wit h two distinguished endomorphisms over a fixed field, and the wildness
of t heir classification problem are described in [136] or, more specifically, in
[137].
The mai n references on Shelah's classificat ion theory are just t he Shelah
book [149] and its revised and updated version [151]. Most of t he topics of
this chapter are treated there in detail. Another good and per haps mor e ac-
cessible source on classification and stability t heory is [8]. See also Makkai's
paper [101] . Vaught's Conjecture is proposed in [174], and its solution in
the w-stable framework can be found in [152]. Lascar's paper [84] provides
an enjoyable discussion of t his matter.
The (uncountable) spectrum problem for complet e countable first order the-
or ies is fully solved in [53].
Simple theor ies were introduced in [149], but it was Kim who showed, to-
gether with Pillay, t heir relevance within the classification program: see [71]
and [73]. Kim again observed the key role of symmetry, transitivity and
local character [72]. Wagner's book [175] provi des a general and exhaustive
report on this theme.
As already said, stability, superst ability and the further dichotomies ari sing
in t he classification program are treated in Shelah books [149, 151], in [8] or
also in [101]. [85] provides a nice and terse introduction to stability, using
and emphasizing t he notion of heir and coheir. [83] pursues t his approach
and deals in particular wit h Rudin-Keisler order and strong regularity.
The effectiveness aspects of Shelah 's classification progr am is discussed in
[114], while [2] examines it s connect ions with Stationary Logic.
Turning our attention to t he algebraic examples, let us mention [136] or also
[56] for modules, and [19] for pseudofinite fields . [110] treats differentially
closed and separably closed fields, and includes a wide bibliographical list on
them. [62] and [155] provide ot her key references on DCF
o
; see also [134].
Existentially closed fields wit h an automorphism are j ust t he subject of [20].
An explicit example of such a field can be found in [58] and [88]; see also
[95].
Shelah's uniqueness theorem is in [148], Morley's theorem in [117]. Anot her
proof of Morley's Theorem is given in [9]; see also Sack's book [146], or [57].
The Ehrenfeucht - Mostowski models quoted in Sect ion 7.8 are int roduced
in [38].
Finally, let us deal with biinterpretability. Malcev's correspondence is in
[102] while Mekler's th eorem on nilpotent groups of class 2 is in [113]. Zil-
ber 's program is developed in [182], where Zilber's Conjecture is also pro-
posed. The negative solution of this conjecture is in [59], and the Hrushovski-
Zilber theorem on Zar iski structures in [64].
Chapter 8
Model Theory and Algebraic
Geometry
8.1 Int roduction
We have often emphasized in t he past chapters t he deep relati onship bet ween
Model Theor y and Algebraic Geometry: we have seen, and we are going t o
see also in t his cha pter t hat several relevant notions arising in Algebraic Ge-
ometry (like variet ies, morphisms, manifolds, algebraic groups over a field
K) are definabl e object s and are consequent ly concerned wit h t he model t he-
or eti c machinery developed in t he previous pages. For inst ance, when K is
algebraicall y closed, t hey are w-stable st ruct ures.This connection ca n yield,
a nd is act ually yielding, significant frui t s in both Model T heory and Alge-
br aic Geometry. On th e one hand, several t echniques and ideas originated
and employed within t he specific sett ing of Algebraic Geomet ry can inspire
a more abstract model th eor etic treatment , applying to arbit rary classes of
st ruct ures. In this sense Algebraic Geometry over algebraically closed fields
can suggest new directions in th e study of w-stability: we will describe this
connection in many sect ions of t his chapter. However a parallel analysis
can be developed inside other relevant areas, like differentially closed fi elds
(and Differential Algebraic Geometry) , or existe nt ially closed fields with an
a ut omor phism, and so on.
On t he oth er hand, it is right t o observe that t he benefits of t his relationship
regard not only Model Theor y, but also, and relevant ly, Algebraic Geome-
t ry. In particular, we will propose some prominent problems in Algebraic
Geometry, whos e solution does profi t by Model Theor y and it s t echniques.
This will be the aim of t he final section of this cha pte r.
291
292 CHAPTER 8. MODEL THEORY AND ALGEBRAIC GEOMETRY
8.2 Algebraic varieties, ideals, types
Let J( be a field, n be a positive integer. We already introduced in th e
past chapters the (algebraic) varieties of K" : They are t he zero sets of finite
systems of pol ynomials of J([X] (where xabbreviates, as usual , (Xl , .. . , x
n
)),
and so are definabl e in J(. Moreover they form the closed sets in the Zari ski
t opology of K"; accordingly even the Zariski open or const ructible sets ar e
definable in J( .
But the varieti es of K" are also closely related to the ideal s of J([X]. In fact
one can define a function I from vari eties to ideals mapping any variety V
of K" into the ideal I(V) of t he polynomials f(x) in J([X] such that f(a) = 0
for every ain V . Checking that I(V) is indeed an ideal is straight for war d;
I(V) is even a radical ideal , in other words it coincides with its radical
radI (V ): if f(x) E [([X] and, for some positive int eger k, fk( x) E I(V) ,
t hen f (x) already occurs in I (V). In particular I is not onto.
But there is also another function V in the other direction, from ideals to
varieties, mapping any ideal 1 of J([X] (in part icular any radical ideal) into
t he set V(I) of t hose element s aE K" annihilating all the pol ynomials of I
V(1) = {a E J(n : f (a) = 0 vf (x) El}.
Due to t he Hilbert Basis Theorem, I is finit ely generated, and so V(I) is a
variety. Indeed, any vari ety V can be obt ained in this way by definition; in
oth er words V is onto. Not ice also t hat V(I) = V(rad I) for every ideal I .
The definition of I and V trivially impli es t hat, for every ideal I of J([X] ,
I(V(1)) "2 I. As I(V(I)) is a radical ideal , I(V(I)) "2 rad T, Hilbert 's
Nullstellensatz (see Ch apt er 3) ensur es that , when J( is an algebraically
closed field, equalit y holds: for every ideal I of J([X] ,
I(V(1)) = rad I ,
It is easy t o deduce that , if J( is algebraically closed, t hen I and V deter mine
t wo bijecti ons, the one inverse of t he other, between varieties of K" and
radical ideals of J([X] . We will st ill denote these restricted bijections by I ,
V respectively. Notice t hat both reverse inclusion: for instance if V , W are
two varieties of K" , then
I(V) "2 I(W) ~ V ~ W.
We assume from now on that J( is an algebraically closed field. By t he way
recall that, under t his condit ion, what is definable in J( is w-stable of finite
Morley rank, because J( is st rongly minimal.
8.2. ALGEBRAIC VARIETIES, IDEALS, TYPES 293
Let us restrict our at tention from radical ideal s of K[X] t o prime ideals (t hose
ideals I in K[X] such that , if a product of two polynomials of K[XJ lies in I ,
th en at least one factor polynomial is in I as well) . Parallely we consider,
among vari eties of K" , the irreducible ones, so t he non-empty vari eties V
t hat cannot decompose as a union of t wo proper subvariet ies . It is easy
to check that the previous bijections I and V between vari eties of K" and
radical ideals of K[X] link irreducible varieties of K" and prime ideals of
K[XJ: for every vari ety V of K"; V is irreducible if and only if I(V) is prime.
Notice also that I(0) = K[XJ .
A closer relationship links irreducible varieties and prime ideals. For in-
st ance, it is known that every non-empty variety V of K" can be expressed
as a finite irredundant union of irreducible varieties, and that this decompo-
sit ion is unique up to permuting the involved irreducible varieties (which ar e
consequently called the irreducible components of V) ; th e irredundancy of
the decomposit ion just means that no irreducible component of V is included
in t he union of the other components.
Specularly, every proper radi cal ideal I of K[XJ can be expressed as a finit e
int ersection of prime ideals minimal with respect t o inclusion ; even t his
representation is unique up to permuting the involved minimal prime ideals.
One can also realize th at , und er t his poi nt of view, for every non- empt y
vari et y V in K" , t he irredu cible components of V correspond to t he minimal
prime ideals occurring in th e decomposit ion of I (V ).
So far we have summarized-some very famili ar t opics of basic Algebraic
Geometry. Now let Model Theor y intervene. As we saw in Section 5.3,
prime ideal s of K[XJ - and hence, through t hem, irreducible variet ies of
K" - directly and naturally corres pond to n-types over I<. In fact , for an
algebraically closed field K , th er e are t wo bijections i and p , one inverse of
the other, between n-types over I< and prime ideal s of K[XJ. Basically, for
every n-type p over I<,
i(p) is the ideal of the polynomials f(x) E I<[XJ such that the formula
f(11) = 0 is in p,
and, conversely, for every prime ideal I of K[XJ ,
{J(11) = 0 : f( x) E I} U{.(g(iJ) = 0) : g( x) E I<[XJ - I}
enl arges t o a unique n-t ype p(I) over I<. Accordingly, for every irreducible
variety V of K" and polynomial f (x) E I<[ xl,
f (ii) = 0, Vii E V {::} f (x) E I(V) {::} " f(11) = 0" E p(I (V) );
294 CHAPTER 8. MODEL THEORY AND ALGEBRAIC GEOMETny
p(I(V)) is called a generic type of the variety V . More generally a generic
type of an arbit rary non- empty variety V of K" is a generic type of an
irr educible component of V . T his model t heoretic notion of generic type
just corresponds to the idea of a generic point of an irreducible variety V
int roduced in Algebraic Geometry: the latter is a point just annihilat ing t he
pol ynomials in I(V) and not hing more, and so can be equivalent ly defined as
a realizati on of the generic t ype p(I(V)). It can be obtained as follows. As
V is irreducible, I( V) is prime an d consequent ly t he quotient ring /C[X] /I(V)
is an integral domain; /C[X]/I( V) contains a subring {k +I(V) : k E I<}
isomorphic to /C and a t uple x+ I(V) annihilat ing just the polynomial s
of I(V) ; so t he field of fractions /C(V) of /C[X]/I(V) extends /C - up to
isomorphism - and includes a generic point a(V) = x+I(V) of V .
To conclude this section let us observe what follows.
Proposition 8.2.1 Let n be a posit ive integer, /C be an algebraically closed
fie ld, V be an irreducible variet y of K" , 0 be a Zariski open set of K"
satisf ying V nO ::j:. 0. Then the generic type p of V contains the formula
"v EO".
Proof. In fact V - 0 is a variety properly included in V ; consequently
I(V - 0) =:> I(V). If p does not cont ain "v E 0 " , then p has t o incl ude
"v E V - 0 " , and hen ce ever y formula "f (v) = 0" when f (x) ranges over
I(V-0) ; accor dingly, for some pol ynomial f( x) (j. J(V) , "J(v) = 0" belongs
to p: a contradiction. .,.
8.3 Dimension and Morley rank
We maint ain throughout t his section our assumption t hat /C is an alge-
br aically closed field. Let V be an irreducible variet y of K" : Algebraic
Geometry equ ips V with a dimension in the followi g way. As we saw in
the last section, there is a minimal field /C(V) extending /C by a realiza-
tion a(V) of the generic type of V , in ot her words by a generic point of V .
The dimension of V (dim( V)) is just t he transcendence degree of /C CV) over
/C. T his makes sense for an irreducible V , but can be eas ily extended t o
any non-em pty variet y V of K", In this enlarged setting the dimension of
V (di m(V)) is t he maximal di mension of an irr educible component of V.
Finally, t he dimension of a constructible non-empty set X dim(X) is the
dime nsion of t he closure of X in t he Zariski t opology (a non-empt y variety) .
8.3. DIMENSION AND MORLEY RANK 295
On the other side every vari ety V of K" (irreducible or not) and, more
generally, every constructible set X ~ K" is definable in K; K is algebraically
closed, hence w-stable; accordingly V and X have their Morley rank. We
want t o compare the dimension and the Morley rank of V , or X , and to
show that they coincide. We start by examining an irreducible var iety V.
Lemma 8.3.1 Let V be an irreducible variety of K" , p be its generic type.
Then dim(V) = RM(p) .
Proof. p is realized by a(V) in K(V) . So the MorIey rank of p equals
the transcendence degree of K(V) over K (see Section 6.1), and hence the
dimension of V.
Lemma 8.3.2 Let V , W be two irreducible varieties of x, p, q denote
their generic types. If V C W , then RM(p) > RM(q).
Proof. V C W impl ies I(W) "2 I (V) and hence
(*) for every f(x) E K[X] , if f (a(V)) = 0, t hen f(a(W)) = 0 as well.
In parti cul ar t he t ranscendence degree of K(V) over K is not smaller t ha n
that of K(W ) over K, and so RM(p) 2: RM(q) . Now ass ume RM(p) =
RM(q), t hen K(V) , K(W) have the same transcendence degr ee over K.
Recall t hat K(V) = K(a(V )), and similarly for W . (*) impli es that , if
i }, .. . , i
m
~ n and ail (W ), , aim(W) form a t ranscendence basis of
K(W) over K, t hen ail (V), , aim(V) are algebraically independent over
K and so form a transcendence basis of K(V) over K. Accordingly one
can define an isomor phism of Kw = K(ail (W ), .. . , aim(W)) onto Kv =
K(ai
l
(V) , . .. , aim (V )) fixing K pointwise and mapping ai(W) into ai (V)
for every i = i}, .. . , i
m
. This isomorphism can be enlarged in t he usual way
to an isomorphism between Kw [x] and Kv [X]. By using (*) once again, one
sees t hat , for every j = 1, ... , n with j =I- i} , ... , i
m
, the mini mum poly-
nomial of aj(W) over Kw must correspond to the minimum polynomial of
aj(V) over Kv in this isomorphism. Accordingly we obtain an isomorphism
of K(W) onto K(V) fixing tc pointwise and mapping a j (W) into a j (V) for
every j = 1, . . . , n. But a(V) realizes p and a(W) realizes q, whence p = q
and, in conclusion , V = W.
Now we can show t hat dimension and Morley rank coincide for an irreducible
variet y V.
Theor e m 8.3.3 Let V be an irreduci ble variety of K", Then RM(V ) =
dim(V ), GM(V) = 1.
Proof. Let p be t he generic ty pe of V . p is t he onl y n-type over /C
containing t he formul a "5 E V " th at defines V, and satisfying V( i(p)) =
V. If q is anot her n-type over /C cont aining "5 E V" , t hen i(q) :J i (p),
whence V (i (q)) C V (i (p)) = V. By Lemma 8.3.2, RM(q) < R M (p). Then
R M (V ) = RM (p) and GM(V) = 1. By Lemma 8.3.1, RM (V ) = R M (p) =
dim(V ).
Cor ollary 8. 3.4 Let V be an irreducible variety of K" (with the relat ive
topology of the Zariski topology) . If 0 is a non-empty open set of V , then
RM(O) = RM(V) . IfW C V is a closed set of V, then RM(W) < R M (V).
Proof. The forme r claim follows from Propositi on 8.2. 1 and t he fact t hat,
if p is t he generic point of V , t hen RM (p) = R M (V ). At t his poi nt t he
lat t er clai m is a consequence of t he fact t hat V has Morley degree 1. ...
Corollary 8. 3.5 Let V be a non- empty variet y of K", Th en R M (V) =
dim(V), furthermore GM (V ) equals the number of the irreducible compo-
nents of V having the same dimension as V.
Proof. V is t he (finit e) union of it s irreducible components. Then t he
Morley rank of V coinci des wit h t he maximal ra nk of it s components. So by
Theorem 8.3.3 and the definiti on of di m (V) R M (V ) = di m (V ). Moreover,
if V
o
and VI are two different irr educible components of maximal rank of V,
t hen V
o
n VI is a closed subset of V
o
properl y included in Vo. By Corollary
8.3.4, R M(VonVI) < R M(V
o).
This implies t hat GM (V) equals the number
of the irr educible components of maximal rank in V. ...
Dimension and Morl ey rank coincide even for constr uct ible sets X ~ K" :
Recall t hat, owing to Tarski's quantifier elimination Theor em, constructible
just means defina ble in /C. Furth ermore a const ruct ible X ~ K" ca n be
represented as a union of finit ely many sets, which are in t heir t urn t he
intersection of a variety W - so a Zari ski closed set, t he zero set of a finite
system of equations - and a Zari ski open set 0 - t he set of t he elements of
K" satisfying finit ely many inequ ations
8.4. MORPHISMS A ND DEFINABLE FUNCTIONS
with 90(), " " 9s() E K[Xj, or also, equivalent ly, a single inequation
II9j() # 0 - .
j ss
An open set defined by a unique inequation is called principal.
297
Corollary 8.3.6 Let X ~ tc- be construct ible (hence definable). Th en
RM(X) = dim(X) . Moreover, if X denotes the Zaris ki closure of X , then
R M(X) = R M(X) and RM(X - X) < RM(X).
Proof. We know dim(X) = dim(X) , and dim(X) = RM(X) (by Corollary
8.3. 5) . So the former claim is don e if RM(X) = RM(X). As observed
before, X a finite union of sets of t he form W nOwher e W is a non-
empty variet y of JCn and 0 is open in JCn. As any non- empt y variety of
JCn decomposes in its t urn as the union of it s irreducible components, we
ca n assume that every variety W occurring in t he above representation of
X is irreducible. By Corollary 8.3.4, RM(W n 0) = RM(W) = dim(W).
Consequently the Morl ey rank of X (that equals the maximal rank of the sets
W n 0) coincides with the Morley rank of X . This proves t he former claim.
But , by Corollary 8.3.4 once again, RM(WnO) < RM(W) for Wand 0 as
before and W irreducible. So, in conclusion, RM(X - X) < RM(X) . ...
8.4 Morphisms and definable functions
JC still denotes an algebraically closed field. Let n , m be positive integers, V ,
W be two algebr aic vari et ies in JCn, JCm respectively. Algebraic Geomet ry
defines what a morphism from V to W is: it is a function f from V int o
W such t hat, for every i = 1, . .. , m, t he composit ion f i of f and t he i-
th projection of K m onto K is a polynomial map. One easil y sees that a
morphism is a cont inuous functi on with respect to the to pology induced on
V and W by the Zariski topology.
But what is remarkabl e for our purposes is t hat a variety morphism is always
definabl e in JC. For inst ance, if f is, as above, a morphism from V to W ,
t hen "f (v) = W" is defined by the formula
"vE V" 1\ " $ E W" 1\ 1\ "f i (V) = uu",
l <i <m
Conversely what can we say about an arbitrary definabl e function f in JC?
Certainly both t he domain and the image of f are definable sets . Further-
more, if the image of f is a subset of K m, t hen, for every i = 1, . .. , m,
the composit ion f i of f and t he i-th projection of K": onto J( is definable.
So we are restricted to characterize definable functions from a subset of K"
into J( , for some positive integer n. In this framework one can observe what
follows .
Theorem 8.4.1 Let f be a definable function from K" into K.. . Then there
are a non-empty open subse t 0 of K.. n, a rational function r and (when K..
has a prime characteristic) a positive integer h such that
if car K.. = 0, then f equals r on 0;
if car K.. is a prime p and Fr denotes the Frobenius map from K.. to
K.. (that mapping any a E J( into Fr(a) = a/'], then f coincides with
r -:>. r on O.
Just to underline th e power of this result, let us recall that , owing to Corol-
lary 8.3.4, every non- empty open subset 0 of K" has th e same Morley rank
n as K" ; while th e Morley rank of K.. n-o is smaller (for , K" is an irreducible
vari ety of rank n). Secondly, it is worth recalling th e general fact that , if f
is a function from a vari ety V of K" into K.. and, for every aE V , there is
an open neighbourhood 0 of asuch t hat f equals some rational funct ion in
O n V , t hen f can be globally expressed as a polynomial function.
Now let us show Theor em 8.4. l.
Proof. Let Q be the uni verse of the theory of K.. , t} , .. . , t E Q be alge-
braically independent over K.. . As f is J(-definabl e, any automorphism of
n fixing K.. and t
l
, , t
n
pointwise acts identically also on f(i) (f denotes
here th e tuple (t}, ... , tn)). Whence j(i) is in dclt K n i) . So, if K.. has
characterist ic 0, then
f(i) = r(i)
for some suitable rational function r with coefficients in J(, while, if K.. has
prime characteristic p, then
where r is as before and hp is a positive integer. Put s = r when car K.. = 0,
s = r-: .r otherwi se. The elements in K" where f coincides with s form a
set X definable in K.. (just as f and s) , and th e form ula "s(v) = f (v)" defining
it is in th e type of f over J(. As t} , ... , t
n
are algebraically independent
over K.. , t his t ype has Morley rank n. Hence RM(X) 2: n. As X ~ K" ; t he
Morl ey rank of X must equal n , and RM(J( n - X ) < n. It follows th at t he
8.5. MANIFOLDS 299
Zari ski closure of K " - X has Morley ra nk < n, and so it s complement is
an open set 0 in X of Morley rank n where f = s.
8.5 Manifolds
Throughout this section IC still denotes an algebraically closed field and n a
positive int eger. We deal here with (abstract) manifolds in ICn and we show
that th ey are definabl e object s. First let us recall their definition.
Definition 8 .5.1 A manifold of ICn is a structure V = (V, (Vi , f i) i5:
m
)
where m is a natural and
V is a subset of K" (called atlas);
V
o,
... , V
m
are subsets of V , and V is the union Ui<mVi ;
fo r every i ~ m, f i is a bij ection of Vi onto a Zariski closed set Vi (a
coordinate chart of the atlas V) ;
for i ,j ~ m and i i- i , f i(Vi n Vj) = o., is an open subset of Vi;
for i , j ~ m and i i- i . f ij = fi . f T 1 (a biject ion between Vji and Vij )
can be locally expressed as a tuple of rat ional fu nct ions.
Manifolds include several familiar exa mples.
Examples 8.5.2 1. Every algebraic vari ety (so every Zari ski closed set )
V of ICn is a manifold, provided we set m = 0 and choose Ui, = V
o
= V
as the only coordinat e cha rt of t he atlas; the resulting manifold is called
affine.
2. Let 0 be an open principal set of ICn; 0 is defined by a single inequa-
t ion -' ''g(v) = 0" ; notice that the formul a "g(v) . V
n
+l = 1" defines
a closed V in ICn+l , and it is easy to control t hat the projection of
K
n
+
l
onto K" by t he first n coordi nates determines a bijection f of
V onto O. Accordingly (V, (V, f -l )) is a manifold with t he only chart
O. Such a manifold is called semiaffine.
3. Also the projective space p n(IC) can be equipped with a manifold
structure. In fact, view p n(IC) as t he quoti ent set of K n+l - {a}
with respect to the equivalence relation rv linking two non zero tu ples
x = (xo , Xl, . . . , x
n)
and fl = (Yo, YI , . .. , Yn) in K n+l if and onl y
if t here is some k E J( such that Xi = kYi for every i ~ n. For
x E J( n+l - {O}, let (xo : Xl : . : x
n
) be t he class of x with respect
t o t his relation. Moreover , for i ~ n , let
Ai denote th e set of th e element s (x o : Xl : : x
n
) in p n(!C)
such that Xi i= 0,
f i be the function of Ai int o K " mapping any (xo : Xl : ... : x
n
)
. (3Zll. Xi _l 5..!. ~
In Xi ' , ~ , x i , . , xi
It is st raight forward to check that (pn(!C), (Ai, Ji)i5:.
n
) is a manifold.
Notice t hat affine and semiaffine var iet ies are definable -even as st ructures-
in /C. Moreover p n(!C ) is interpretable in !C both as a set (since t he relation
rv is 0-definable) and as a manifo ld. But algebraically closed fields uniformly
eliminate th e imaginaries, so we can view p n(!C) even as a definable st ructur e
in !C.
Mor e gener ally one can show
T heorem 8.5.3 Let V = (V, (Vi , Ji) i5:.
m
) be a manifold of . Then V is
a structure definable in !C.
Proof. As algebraically closed fields have t he uniform eliminat ion of imag-
inari es, it is sufficient t o show t hat V is a struct ure int erpretable in !C. In
fact, every map U, of t he atlas V (i ~ m) is definable in !C. V can be re-
garded as the quotient set of t he disjoint uni on of t he cha rts U, (wit h i ~ m)
with respect to the equivalence relat ion ident ifying Uij and Uji via f ij for
every i < j ~ m; moreover , for every i ~ m, Vi can be definably recovered
as t he image of U, by th e projection into the quotient set V , and f i is given
by the inverse funct ion of this projection (restrict ed to Ui) . So our claim
is proved if we show t hat, for every i, j ~ m with i i= i , f ij is definable.
But f ij can be locally expressed as a rat ional function, and its domain Uij
is an open subset of U, and accordingly can be written as a finit e union
of principal open sets. So th e th eor em is a direct consequence of the next
result.
Lemma 8.5.4 Let 0 be a prin cipal open of !C
n
, and let q(x ) E [([X] be a
polynomial satisfying 0 = {ii E K " : q(ii) i= O}. Let f be a funct ion of 0
into K" which can be locally expressed as a rat ional funct ion. Then there
are a polynomial r( x) E [([X] and a positive integer m such that f(ii ) =
r(ii)/qm(ii) for eve ry ii EO. In particular f is defi nabl e.
8.6. ALGEBRAIC GROUPS 301
Proof. We know that 0 is canonically homeomorphic to the closed subset
V of Kn+l defined by q(v) . Vn+l = 1. Under this per spective f can be
replaced by the functi on 1* of V into 1< mapping any tuple (ii, an+d E V
in f(ii) . Even 1* can be locally expressed as a rational function , and hence
as a polynomial function in (x, xn+d (by th e general fact we recalled before
the proof of Theorem 8.4.1). So th ere is some polynomial s(x, xn+I} E
1<[x, xn+l ] such t hat
J*(ii, (q(ii))-l) = s(ii, (q(ii))- l) vs En.
Let m denot e t he degree of s(x, Xn+l ) with respect t o Xn+l' Then there is
some polynomial r (x) E 1<[X] such that , for every ii EO,
f(ii) = J*(ii, (q(ii))-l) = r a - ~ . ..
qm a
Notice th at a manifold, when regarded as a definable struct ur e, may lose
part of it s geometric features. For instance the Zariski closed subset of K
2
defined by
Xl . (X2 - 1) = X2 . (X2 - 1) = 0
is an affine manifold, formed by the line X2 = 1 and t he point (0,0) , and
hence is th e disjoint union of two closed sets. However , consider the manifold
given by t he proj ective line pl(K) (as seen in Example 8.5.2, 3). Fro m the
definabl e point of view, it s atl as has two charts, each of them is a line and
th ese lines coincide except a single point . So the resulting manifold is again
the union of a line {(1 : xI) : Xl E 1<} and a point (0 : 1), and as a definabl e
object is qui t e similar to the previous one. But p l (K) is not the union of
two distin ct closed sets.
On t he oth er side, it is noteworthy t hat every definable subset X ~ K"
can be easily equipped with a manifold st ruct ure. In fact X decomposes
as a union Ui<m(Vi n Oi) where m is a natural and, for every i ::; m, Vi
is a Zariski closed of K" and O, is a pr inci pal open set , so that Vi n O, is
canonically homeomorphic to a closed U, of Kn+l , as observed before. On
this basis, it is easy t o build a manifold structure on X (with U
o
, ... , U
m
as atl as maps) .
Furt her more a manifold V , viewed as a definable structure, is w-stable.
8.6 Algebraic groups
A basic example of algebr aic group over a field K is the linear group of
degree n over K GL(n, K) , where n is a positive integer. Observe:
GL(n , /C) is a prin cipal open of /Cn
2
, because it is defined by the dis-
equat ion -, (det (if) = 0) ; hence GL(n , /C) is canonically homeomorphic
t o t he Zariski closed of /Cn
2
+1 given by t he equat ion det(if) . v
n
2+ 1 = 1;
t he product oper ation in GL(n , /C) is a morphism of variet ies and so
is definable.
Accordingly the group GL (n, /C) is a st ructure definable in /C. Moreover , if
we ass ume /C algebraically closed, GL(n, /C) is an w-st able group of finit e
Morley rank.
Also linear algebraic groups are examples of algebraic groups. Recall t hat a
linear algebraic group over a field /C is a closed subgroup 9 of some linear
group GL (n , /C ). In particular a linear algebraic group 9 is a variety over
/C, hence is defina ble (as a group) in /C and is w-stable of finite Morley rank
when /C is algebraically closed. Under t he last ass umption, we can say even
more: indeed , for an algebraic ally closed /C , t he linear algebraic gro ups are
just t hose subgroups of t he linear groups GL(n, /C) which are definable in
/C. Let us see why.
Theorem 8 .6.1 Let /C be an algebraically closed fi eld, n be a posit ive in-
teger, 9 be a subgroup of GL(n, /C). Th en 9 is closed if and only if 9 is
defi nabl e (in /C) .
Proof. Clearly, if 9 is closed -i n other words 9 is a variety- , then 9 is
definabl e. Conversely suppose t hat 9 is a definabl e group; let C be t he
closure of G with respect to t he Zari ski topology, and let a be an element of
C. Every open set of /Cn
2
containing a overlaps G. Consequent ly, for bE G ,
every open 0 including ba overlaps G; in fact , if ba E 0 , then a E b-10;
b-10 is open because th e left multiplication by b is a morphism of varieties,
and so is cont inuous; whence (b- 10) n G i= 0 and so 0 n bG i= 0; as bEG,
bG is just G, hence 0 n G i= 0. In conclu sion , for every a E C, Ga < C. If
th ere is any a E C - G, then Ga ~ C is disjoint from G and has the same
Morl ey rank as G. So RM(G - G) ~ RM(G) , whi ch contradicts Corollary
8.3.6. Hence C ~ G and G is closed . ..
As already said, linear groups and linear algebraic groups exemplify algebraic
groups. In fact an algebraic group over a field /C is defined as a manifold
over /C ca rrying a group str uct ure whose product and inverse operations are
(manifold) morphisms.
Of course, understanding t his definition preliminarily requires to stat e what
a product of t wo manifolds is, and t o realize that t his product is a manifold
8.6. ALGEBRAIC GROUPS 303
as well; moreover we should specify what a manifold morph ism is. We omit
here the det ails. The former point is comparat ively simple and natural to
cla rify, whil e the concept of morphism is more complex t o introduce; but ,
once this is done, one ca n easily show th at manifold morphisms are definable
(as it is reasonabl e to expect) .
Hence, if K is algebraically closed, t hen every algebraic group g over K is a
st r ucture definable in K and so w-stable of finite Morley rank.
As already said, linear algebraic groups are algebraic gr oups; in fact they
just correspond to affine (or also semiaffine) manifolds. And indeed Theorem
8. 6.1 can be regarded as a parti cular case of a general result ensuring that
every group definabl e in an algebraically closed field K is an algebraic group
over K up t o definabl e isomorphism. This fact was shown by Hrushovski
(and Van den Dries) , who observed t hat it is impli citl y contained in some
results of A. Weil. So it is commonly quoted as the Hrushovski-Weil Theo-
rem.
Theorem 8.6 .2 (Hrushovski-Weil) Let K be an algebraically closed fi eld
and g be a group definable in K. Th en g is isomorphic to an algebrai c
group over K by a function definable in K
Let us spend some more words about the connection between algebraic
groups and w-st able groups. We have seen that every algebraic group over
an algeb raically closed field K is w-st abl e of finit e Morley ra nk. Of course
w-st able groups include furt her relevant examples: for ist ance, any divisibl e
torsionfree abelian group - so basically any vectorspace over Q - is w-stable,
and even strongly minimal. More generally, it was shown by Angus Mac-
intyre that the w-stable (pure) abelian groups are just the direct sums of a
divisible abelian group and an abelian group of finite exponent .
However the techniques used in t he investigat ion of abelian groups do not
seem appropriat e to handle w-stable groups. On the cont rary, algebraic
groups and t heir t heory fit very well for w-stable groups. This is not sur-
prising. In fact , even neglecting the similarit ies we emphasized in t he last
sect ions between varieties and definabl e set s, or di mension and Morl ey rank,
and so on , we can recall t hat several notions introduced in Chapter 6 for
st udying w-st abl e groups clearly come from Algebraic Geometry. In this set-
t ing, it is worth ment ioning t he following conjecture, proposed by Cherlin
in 1979.
Conjecture 8 .6.3 (Cherlin) Let g be an w- slable group of finit e Morley
rank. If g is si mple, then g is an algebraic group over an algebraically
clos ed fi eld.
Cherlin 's Conjecture is still an open questi on. We should also point out t hat
some remarkable progress in st udying w-stable groups of finit e Morl ey rank
has been obtained by using ideas and tec hniques coming from finite groups,
in particular from t he classification program of finit e simpl e groups.
8.7 The Mordell-Lang Conjecture
Every promise must be honoured. And consequently, after emphasizing
so many t imes t hat Model Theor y does significant ly apply t o Algebraic
Geomet ry, here we are t o propose one application (indeed a beau tiful and
deep application, in our opinion): Hrushovski 's proof of a question of Lang
usually called Mordell-Lang Conjecture. Why is this solut ion notewor thy?
Basically because it is t he very first proof of t his conject ure, at least in t he
general for m we will state in 8.7.6 below; but also, and mainly, because it
largely involves Model Theor y (strongly minimal sets, Zari ski str uct ures,
differentially closed and separably closed fields, as well as the mat eri al of
t his chapter).
So let us introduce t he Mordell-Lang Conject ure, and briefly sket ch it s his-
t ory. We assume some acquaintance with Algebraic Geometry. The questi on
originally rose within Diophantine Geomet ry, which deals wit h the root s of
sys te ms of polynomials over th e ration al field Q or also over a number field
F (that is an extension F of Q of finit e degree). This was t he setting where
Mordell raised in 1922 t he following probl em.
Conjecture 8.7.1 (Mordell) Let F be a number fi eld, X be a curve of genus
> l over F. Then X has onl y finit ely many F -rationals points.
Incident ally recall t hat a curve X of genus 1 is an ellipt ic curve, and so is
naturally equipped with a gr oup struct ure.
In a mor e abst ract perspective, on can obser ve what follows.
Remarks 8.7.2 1. A curve X of genus 2': lover F is a Zariski closed
subset of it s J acobian J (X ).
2. A t heorem of Riemann says t hat t he J acobian J (X ) is an abelian
variety (t hat is a connected complete algebraic group); by t he way,
every abelian variety is act ually an abelian group.
3. A t heorem of Mordell and Weil ensures t hat, if A is an abelian vari et y
in t he complex fi eld defined over our number field F , t hen t he set 9
of t he F-rational points in A is a finit ely generat ed subgroup.
8.7. THE MORDELL-LANG CONJECT URE
So Mordell's Conj ecture ca n be restat ed mor e generally as follows.
305
Conjecture 8.7.3 Let A be an abelian varie ty of C, X be a curve of C
embedded in A, 9 be a fi nit ely generated subgroup of A. Then either X is
an elliptic curve, or X nG is fi nite.
A similar question was raised by Manin and Mumford in 1963.
Conject ure 8. 7.4 (Manin-Mumford) Let A be an abelian variety ofC, X
be a curve of C embedded in A, 9 = Tor A be the torsion subgroup of A.
Then ei ther X is an elliptic curve, or X n G is fi nite.
Actually the original form of Manin-Mumford Conjecture said that, if X is
a cur ve of genus > 1 and A = J( X) is it s J acobian, th en X nTor A is finit e.
But t he more general statement given in 8.7.4 is easily obt ained as in th e
Mordell case .
A possibl e unifying approac h covering both the Mordell and t he Manin-
Mumford problem uses t he notion of group of fini te type: in our charac-
te ristic 0 fr amework, t his ca n be introduced as an abelian group 9 wit h a
finit ely generated subgroup S such t hat , for every g E G, t here is some posi-
t ive int eger m for which mg is in S . In fact , every finite ly generated abelian
gro up 9 is of finit e type (j ust take S = G) , and every to rsion group is of
finit e ty pe as well (via S = {O}) . Accordingly the conjectures of Mordell
and Manin-Mumford can be regarded as two parti cul ar cases of t he following
more general questi on.
Conject ure 8. 7.5 Let A be an abelian varie ty over C, X be a Zari ski closed
subset of A , 9 be a subgroup of fi nit e type of A. Then X n G is a (possibly
empt y) fin it e uni on of cosets of subgroups of g.
This statement was formul at ed by Lang in t he sixt ies, and is usually called
th e Mordell-Lang Conjecture. As underlined before, it impli es a positi ve
answer t o Mordell 's Conject ure: to see this, just take a curve X
o
of genus
> lover a number field F , embed X =Xo(C) int o it s J acobian A = J(X)
and apply t he Mordell-Weil Theore m ensuring th at t he group 9 of F-rational
point s in A is finit ely generated. Accordingly decompose X
o
= X n G as a
finit e union of cosets a + H where a E G and H is a subgroup of g. Take
a coset a +H . It s closure a +H is included in X
o
- an irr educible set of
di mension 1 -. Consequent ly, if a +H is infinit e, then X
o
just equals a +H
and so inherit s a group st ruct ure , and genus S; 1. This means t hat, if t he
genus of X
o
is > 1, t hen every coset a +H must be finite, whence X
o
it self
is finite .
These are the questions we wish to deal with. Now let us report about their
solution. Mordell's Conjecture was proved by Faltings in 1983. The echoes
of this result spread far and wide, also because it implied an asymptotic so-
lution of Fermat's Last Theorem: in fact, by applying Mordell's Conjecture
(or , more precisely, Falting's Theorem) to the projective curve over Q
for n ;::: 3, one gets only finitely many zeros for every n.
Also the Manin-Mumford Conjecture was positively answered by Raynaud
in 1983. The Mordell-Lang Conjecture (as stated before) was solved just a
few years ago: first Faltings handled the particular case when the group 9 is
finitely generated, and then McQuillan provided a general positive solution,
using Falting's work and other contributions of Hindry.
So far we have limited our analysis essentially to the characteristic 0, and
to number fields. What can we say when passing to function fields, or
prime characteristics? First let us deal with function fields. Still working
in characteristic 0, Manin had proved in the sixties the following analogue
of Mordell 's Conjecture in this setting: if K is a function field over an alge-
braically closed field Ko (of characteristic 0) and X is a curve of genus> 1
over K, then either X does not descend to Ko (in which case X(K) is finite),
or X is isomorphic to a curve X
o
defined over Ko (and all but finitely many
points of X(K) come from elements of X(K
o)).
When considering prime characteristics p, even the notion of group 9 of
finite rank must be rearranged. In fact, what we have to require now is that
9 has some finitely generated subgroup 5 such that, for every g E G, there
exists a positive integer m prime to p satisfying mg E S .
However 8.7.5 - as it was stated before - does not hold any more. In fact A
itself is a torsion group without element s of period p; but there may exist
some curves of A which are not finite unions of cosets of subgroups of A.
A reasonable restatement of 8.7.5 in the general setting, for an arbitrary
characteristic (0 or prime), is the following.
Conjecture 8.7.6 Let Ko -< K be algebraically closed fields, A be an abelian
variety over K having trace 0 over Ko (this means that A has no non-zero
abelian subvarieties isomorphic to abelian varieties over Ko). If X is a
Zariski closed subset of A and 9 is a subgroup of A of finite rank, then
X nG is a (possibly empty) finite union of cosets of subgroups of g.
In 1994 A. Buium proved this form of the Mordell-Lang Conjecture in char-
acteristic O.
8.7. THE MORDELL-LANG CONJECTURE 307
Theorem 8.7.7 (Buium) 8. 7.6 is a true statement when /Co -< /C are alge-
braically closed fiel ds of characteristic o.
What is not eworth y for our purposes in Buium's line of proof is his use
of Differenti al Algebraic Geometry; indeed Differenti al Algebra promptl y
recalls Model Theory and its treatment of differenti ally closed fields. So
it is right t o spend a few words to descri be Buium's strategy: one equips
/C wit h a deri vation D whose constant fi eld is just /Co, one embeds 9 in a
differenti al algebraic group 91 and, finally, one shows by analytic argument s
that X nCl is a finite union of cosets of 91 and one t ra nsfers this proper t y
to 9.
But it was Ehud Hrushovski who first proved t he Mordell-Lang Conjecture
in it s more general form , in any cha racterist ic, following the initial Buium
approac h and then using mod el th eor etic methods and, above all, Zari ski ge-
ometries, differentially closed fields in charac te rist ic 0 and separably closed
fields in prime characteristic, in addit ion t o Morley ra nk, elimin ation of
imaginari es and t he definabilit y resul t s of t his chapte r. It should be empha-
sized t hat no alte rnative general proof of the conjecture is known; and indeed
Hrushovski proposed, some time lat er , a new model t heoretic proof of t he
Manin-Mumford Conjecture, based on a crucial use of existentially closed
fields with an aut omorphism (in pa rticular Zilber 's Trichotomy in ACFAa) ,
a nd getting in t his way nice effective bounds of the number of involved cosets
in a decompositi on of X n C. Coming back t o t he Mordell-Lang conjecture,
we ca n say
Theorem 8. 7.8 (Hr ushovs ki) 8. 7.6 is a true statement in any charact eris-
tic.
This concludes our short and lacunose history of Mordell-Lang, Mordell and
Manin-Mumford Conjectures. Which is our purpose now? Certainly we do
not aim at providing a complete report of Hrushovski 's proof: t his would
requ ire many pages and serious effort s; moreover th ere do exist sever al nice
expository pap ers and books wholly devoted to a det ailed exposit ion (some
of t hem are mentioned among the references at t he end of th is chapter) .
On t he oth er hand, we would like t o spend a few words about Hrushovski 's
approach, just to explain where Model Theory int ervenes an d why it plays
a decisive role. With t his in mind, we will sketch Hrushovski 's proof in t he
characterist ic 0 case, where some old friends of ours - differenti ally closed
fi elds - are involved. Then we will shortly comment the prime character-
istic case, where differenti ally closed fields are profitabl y replaced by the
separably closed ones.
So take t wo algebraically closed fields 1(0 -< I( of charact eristic o. Let A be
an abelian variety over I( wit h trace 0 over 1(0, X be a Zari ski closed set
in A, y be a subgroup of A of finit e rank. Our claim is that X n G is a
(possibly empt y) finit e union of cosets of subgroups of y.
(a) Without loss of generality one ass umes that I( has infinite transcen-
dence degree over 1(0. Then one equips I( with a derivation D making
I( a differential field, and even a differenti ally closed field, whose con-
stant subfield C (I( ) coincides with 1(0. J ust to fix our symbols, let L
denot e from now on th e usual language for fields, and L' = L U{D}
that of differential fields. So 1(0 is strongly minimal both as a struc-
ture of Land L': in fact D is identically 0 on 1(0 and so adds no
definable objects to the pure field 1(0 . On t he cont rary, I( is a st rongly
minimal struct ure in L, as an algebraically closed field, but it is not
any more as a differenti ally closed field; indeed 1( , although w-stable,
has Morley rank w in L'. Not ice also t hat, owing to what we saw in
t he past sect ions, the abelian variety A is definable (even in L) in 1(.
(b) At this point one recalls a general result of Manin on differenti al
fields: the derivation D yields a group homomorphism p (definabl e in
L') from A onto (I(+) d, where 1(+ is the addit ive group of I( and d is
t he dimension of A. The kernel of p is definabl e in L' and has a finite
Morley rank. Now we deal with y. As y has finite rank and 1(+ has
no nonzero torsion elements, t he grou p p(G) is finitely generated and
t here are go, . . . , gm E K such that
p(G) ~ L Q . s. ~ L ](0 gi
i<m i<m
- -
Let II denot e L i<m ](0 . gi. H is definabl e (in L' ) and has finite
Morley ra nk. Hence p-l (H) is a subgroup of A exte nding y; moreover
p-l(H) is definabl e (in L') and has finit e Morley rank because both H
and t he kernel of p are definabl e of finit e Morley rank. Without loss
of gener alit y for our purposes, we can replace Y by p-l(H). In fact ,
if X n p-l (H) is a finit e uni on of cosets of subgroups of p-l (H) , t hen
t he same can be said about X n G and G. So we can ass ume t hat y
it self is definabl e and has a finit e Morley rank.
Now, just to explain Hr ushovski's ap proac h in a mor e accessible way,
let us restrict a little more our framework to the particular case when X
is an irreducible curve (the setting of the original Mordell Conjecture).
If X nG is finite , then we are don e. Otherwise X nG - as a definabl e
8.7. THE MORDELL-LANG CONJECTURE 309
set of finit e Morley rank - contains a definable strongly minimal subset
5 . As usual, 5 can be viewed as a strongly minimal str uct ure.
(c) Now we use t he result of Hr ushovski and Sokolovic saying that Zil-
ber's Tri chot omy Conjecture holds for strongly minimal sets definabl e
in differenti ally closed fields of characteristic O. In fact , t hese strongly
minimal sets are Zari ski struct ures, and so obey t he Hrushovski-Zilber
Theor em. This applies to 5, of course. But what Hrushovski also
point s out is t hat 5, as a Zari ski st ruct ure , is locally modular. In fact,
as 5 is st rongly minimal, it suffices to excl ude finit ely many points
from 5 to get an indecomposable set . So t here is no loss of generality
in ass uming th at 5 it self is indecomposable, and conseque nt ly each
t ranslate bS with bE G is also indecomposable. Up to replacing 5
by b-
1
5 for a suitable bE G we ca n even assume t hat t he identi t y
element 1G of G is in 5 . Hence we are just in a positi on to apply Zil-
ber 's Indecomposability T heorem; accordingly, one ded uces t hat the
subgroup generated by 5 in g is definable, and indeed every element
in t his subgrou p can be expressed as a c-
1
wit h a and c in 5. Hence 5
int erpret s an infinit e gro up and so, as a Zariski struct ure , it cannot be
t rivial. This means t hat eit her 5 is locally mod ular or 5 int erpret s an
infinite (algebraically closed) field. We have to exclude t he lat t er op-
t ion. To obtain t his, one uses a result of Sokolovic already ment ioned
in 7.10 and saying what follows.
(d) An infinit e field definable in a differenti ally closed field of charac-
te rist ic 0 and having finit e Morley rank is isomorphi c to t he constant
subfield by a defi nable function.
Recall t hat, owing to t he elimination of imaginaries, t here is no dif-
ference between definabl e or interpretable wit hin different ially closed
fields. So, if 5 int erpret s any infinite field, t hen it defines even C(K ) =
Ko up to an I-definabl e isomorphism. Consequent ly the subgroup
t hat 5 generates is isomorphi c t o some group go I-definable in Ko by
a function I also definabl e in L'. As D = 0 in [(0, go is definabl e even
in L (just as I and 1-1). Then we can apply t he Hrushovski-Weil
Theorem and deduce t hat go is an algebraic group over Ko up t o an
L-definabl e isomorphi sm. At t his point one checks t hat go defines an
abelian subvariet y of A in K, which cont radict s the hypoth esis t hat A
has t race 0 over Ko. In concl usion, 5 must be locally modular.
Now 5 is of t he form (a + L) - {b
o
, .. . , btl for some strongly min-
imal subgroup L of g and suitable a, b
o
, . . . , b, E G. X is Zariski
closed, and so contains t he closure a +L of S as well. Hence X equals
a + L and consequently inherits a group structure, as Mordell 's Con-
jecture requires. This concludes our outline of Hru shovski 's proof in
cha ract erist ic 0 (when X is a curve).
But , as already said , the real novelty of Hru shovski 's Theorem concerns
prime characterist ics p. So it is worth spending some words also on this case.
The pla n of the proof is similar, but requires some necessary rearrangements.
In particular, one can avoid to refer to differentially closed fields and directly
handle separably closed fields (with no derivation) .
(a) First one replaces with no loss of generality K by a separably closed non
algebraically closed extension having finit e degree over KP. Observe
that now the theory of K is not w-stable, although it is stable.
(b) The role of the kernel of J-L is now played by n npnA, which is not a
definable set, but is the intersection of infinitely many definable sets.
The other crucial point s in th e proof are:
(c) any strongly minimal struct ure definabl e in K is st ill a Zari ski st ruct ure
(so Zilber's Trichotomy holds);
(cl) a field definable in K and having Morl ey rank 1 is isomorphic to Ko by
a definabl e function (a result of M. Messmer).
For more details, look at the references quoted below.
8.8 References
The connection between Model Theory and Algebraic Geometry is clearl y
explained by Poizat's book [131] (recentl y translated in English [135]): this
is a very rich reference on t his matter. In particular it includes a proof
of Hru shovski-Weil Theorem 8.6.2 . Cherlin's Conjecture was raised in [24].
The classification of w-stable groups was given by Macint yr e in [89]. Groups
of finit e Morl ey rank are examined in [15].
Now let us deal with Mordell-Lang Conjecture and Manin-Mumford Conjec-
ture. A geometrical int roduction ca n be found in Lang's book [80]; a short
but resonabl y exha ust ive history of these two questions is also in the recent
Pillay paper [128]. Pillay's book [126] explains the main model theoretic
techniques involved in Hru shovski's approach. Hrushovski 's original proof
8.8. REFERENCES 311
of t he Mordell-Lang Conjecture is in [60]. A det ailed exposit ion is in th e
book [16]. [50] and [127] are shorter surveys , both very read abl e.
As already said, Hrushovski 's proof in cha racteristic 0 case is based on some
pr elimin ari es concerning differenti ally closed fields: t hey can be found in
[155] and [62]. For a prime characterist ic, separa bly closed fields are enough:
th e basic preliminari es are described in [115].
Finally, let us deal with Manin-Mumford Conjecture: Hrushovski 's proof is
now in [61] . It is based on t he model t heory of ACFA, as developed in [20]
and [21] .
Chapter 9
O-minimality
9.1 Introduction
The last part of this book is devoted to o-minimal st ruct ures . As we saw
in t he past chapt ers, t hey are the infini t e expansions M = (M, :S; , . . .) of
linear orderings such that th e subsets of M definabl e in M are as t rivial as
possibl e, and restrict to th e finit e unions of singlet ons and open int ervals,
possibl y with infinite end points oo (equivalent ly to t he finit e unions of
open, closed , ... int ervals in th e broad sense including half-lines and the
whol e M).
a-minimal str uct ures are not simple according to t he definition provided in
Ch ap ter 7, just because t hey define and even expand infinite linear orders;
consequent ly no good independence system can be developed inside th em,
and t hey are not classifiable in Shelah 's sense. Despi t e thi s, and just owing
th e relative triviality of th eir 1-ary definable sets , one ca n see that t hey enjoy
several relevant model th eor etic properties and, among th em, a sat isfact or y
notion of independence with a related di mension partly resembling Morley
rank .
Furthermore, they include a lot of noteworthy algebraic exampies. Indeed we
have already seen that th e ordered field of real s (R , +, " - , 0, 1, :S;), as well
as any real closed field, is o-rninimal; discrete or dense infinit e linear orders,
like (N , :S; ), or (Z, :S; ), or (Q, :S;), or (R , :S; ), ar e o-minimal as well; and one
can show t hat even divi sible ordered abelian groups, such as (Q, +, 0, :S; )
and (R , +, 0, :S; ), are o-rninimal.
In particular , considering t he order of th e reals, or expa nding it by addition,
or addit ion and multiplication together , yield o-rninimal st ruct ures. On the
oth er side, t here do exist some expansion of (R, :S;) which are not o-rninimal.
313
314 CHA PTER 9. O-MINIMALITY
For inst ance, ext end the order of reals by t he sinus function si n; in t his
enlarged st ruct ure, Z -a denumer abl e set of isolated points- get s definabl e
(by sin 1r V = 0) , and so o-mi nimality get s lost. Of course the same argument
applies to cos .
This cha pter has a twofold aim. On t he one hand, we will provide an ab-
st ract structure t heory for o-minimal models M. The st art ing poi nt will
be just the definit ion of o-mi nimalit y, and t he consequent class ification of
t he definable subsets of M . But we will see t hat, on t hese seemingly poor
grounds, one can develop a significant t heory, including a nice character-
ization of definabl e subset s of M" for every positi ve int eger n, as well as
of definabl e functions from M" into M. This will also lead t o the alr eady
recalled notion of dimension, satisfying several remarkable properties resem-
bling those of Morley rank in w-st abl e t heories. Act ually th ere is a good deal
of similarity bet ween t he o-mini mal framewor k and t he w-stable set t ing, just
in t hese di mensions, but also in definabl e groups and in other matters. We
will emphasize t hese connections in Sections 9.2-9.6. In part icular Section
9.4 and Section 9.5 will prove, among other th ings, t he already ment ioned
and not ewor th y fact t hat o-mini malit y, unlike minimality, is preserved by
elementary equivalence: if M is an o-minimal st ruct ure, t hen every model
of the t heory of M is o-minimal as well. Accordingly a complet e th eory
T is said to be o-minimal when some (equivalently every) model of T is
o-minimal. The subseq uent section 9.6 will treat definabl e groups, definabl e
manifolds, and so on, in o-minimal st ruct ures.
On t he other side , we will propose ot her relevant examples of o-minimal
struct ures (t o which t he previous gene ral t heorems apply) . T his will be t he
t heme of Section 9.7, where we will see t hat certain expa nsions of t he real
field by familiar functions, such as exponent iation, or suitably restrict ed an-
alytic functions, are o-minimal. Here o-minimality largely overlaps real alge-
braic geomet ry and real analytic geomet ry, both in acquiring techniques and
constructi ons from t he geomet ric framework t owards a general and lar ger
spect re of applications, and in providing a new light and in opening new
per specti ves wit hin t he geometrical setting it self.
The subsequent section 9.8 will deal with some variations on t he o-minimal
t heme, most not ably wit h a noti on of weak o-minimality, enlarging t he 0 -
minimal setting and intensively st udied in t he lat est years.
At last , t he final section 9.9 will int roduce very shortly t he qui t e recent and
attracti ve work of A. Onshuus about a noti on of independence enlarging
both forking independence in simple t heor ies and algebraic independence in
o-minimal th eories t owards a common general framework.
Now a few historical not es. O-minimality began it s life in th e eight ies; it s
9.1. INTRODUCTION 315
origin refers to a classical problem of Tarski, asking whether the real field,
expanded by the exponential function x t--+ eX, still has a decidable theory.
As we know, decidability closely overlaps definability, so a deeply related
question is what is definable in (R, +, " -, 0, 1, :S, exp): is the theory
of this structure quantifier eliminable, or model complete? This was the
scenery where 1. Van Den Dries introduced o-minimality at the beginning
of the eighties. But who gave a considerable impulse to this notion were
A. Pillay and C. Steinhorn, who proved the basic structure theorems on
o-rninimal structures and greatly developed their abstract theory. In 1991
A. Wilkie partly solved Tarski 's Conjecture, showing that the theory of
(R, +, " -, 0, 1, :S, exp) is model complete, and even o-minimal (as we
will see in 9.7, decidability is still an open question, involving a deep number
theoretic problem, usually known as Schanuel's Conjecture, while quantifier
elimination fails). This emphasized t he connection with analytic geometry,
mentioned some lines ago. And indeed o-minimality became, and still is, a
matter of interest not only to model theorists, but to geometers and analysts
as well.
To conclude this section, we give the proof that any o-rninimal ordered field
is real closed. This is the converse of a result we already know, ensuring
that every real closed field is o-minimal, and can be viewed as the o-rninimal
analogue of Macintyre's theorem saying that any w-stable field must be
algebraically closed. The proof requires a very basic machinery from 0-
minimality -just the definition itself- in addition to the necessary algebraic
grounds. Let us preliminarily examine o-minimal groups. Here (and later
in this chapter) intervals possibly admit infinite endpoints, and so include
half-lines, and the whole line, in case.
Lemma 9.1.1 Let A = (A, 0, +, -, :S , . .. ) be an o-minimal structure ex-
panding an ordered group (A, 0, +, -,:S) and let H be a subgroup of
(A, 0, +, -, :S) definable in A. Then H = {U} oppure H = A.
Proof. Suppose towards a contradiction that there exists some subgroup
H i- {O}, A of (A, 0, +, -, :S) definable in A. Owing to o-minimality, H
decomposes as a finite union of pairwise disjoint intervals (possibly closed,
or with infinite endpoints). Accordingly write
(0) H = Ut,
j :::; s
where 10, ... , Is are intervals and s is minimal. Notice that H is infinite,
because it must contain all the multiples nh of any nonzero element h E H
316 CHAPTER 9. O-MINIMALITY
when n ranges over int egers. Hence t here is j ::; s for which I j is infinite.
Without loss of generalit y put j = O. Moreover we can even ass ume that
l a contai ns 0 and is sy mmetric with respect to 0 (in t he sense t hat -c E l a
for every c E la). This is because H is a subgroup, and so includes 0 and is
preser ved under + and -. As H =f:. A, la is [-a, a] or] - a, a[ for some a > 0
in A. In t he former case, every b in A satisfying a < b ::; 2a decomposes as
b = a +(b - a) where a E la Hand 0 < b - a ::; a, so even b - a is in
10 and consequently in H ; hence b E H. Therefor e [-2a, 2a] is an interval
in H properly including 10' This impli es t hat [-2a , 2a] shares at least one
element with some interval Ij where 0 < j ::; s , say with I s. Put
= [-2a, 2a]Ul
s
.
So Jb is an interval including 10 U I s and contained in H . Consequent ly
H = U U i . ,
O<j<s
and t his cont radicts t he choice of s in (0).
In t he lat t er case, fix b E A such t hat 0 < b < a, t hen 0 < a - b < a, and
consequent ly even a - b is in l a. It follows a = b + (a - b) E H. We get in
t his way an interval [- a, a] properl y including l a and contained in H . But
t his contradicts as before t he minimalit y of s in (0). .,
As a consequence, one can give a full characteri zation of o-mi nimal ordered
groups. They ar e exactl y t hose listed before, namely t he ord er ed divisibl e
a belian groups.
Theor em 9.1.2 An o-minimal ordered group A = (A , 0, +, - , ::;) is abelian
and divi sibl e.
Proof. For every a EA, th e cent ralizer C(a) of a is a definable subgrou p of
A, and consequently equ als eit her {O} or A. If a = 0, then clearly C(a) = A.
On t he other side, when a =f:. 0, C (a) = A as well, because a E C (a) and t his
excludes C (a) = {a}. Hence C (a) = A for every a E A, and A is abelian .
Now take a positive int eger n : nA is a definabl e subgroup of A , clearl y
equalling A when A = {a}; on t he ot her side, if A =f:. {a}, t hen nA =f:. {a}
and consequent ly nA = A. Then nA = A for every positi ve int eger n , III
other words A is divisibl e. .,
Coming back t o ordered fields, we can event ually prove
9.1. INTRODUCTION 317
Theorem 9.1.3 Let /C = (K, 0, 1, +, " -,::;) be an a-minimal ordered
ring with identity 1 (in particular let /C be an a-minimal ordered field). Then
/C is a real closed field.
Proof. First we claim that /C is an ordered field, in other words the set /C*
of the nonzero elements in K is an abelian group with respect to -, It suffices
to show that the set > of the elements > 0 of K is so. In fact, for every
a E >, aK is a definable subgroup of (K, 0, +, -, ::;) , and aK -I {O}
because a > O. So, owing to Lemma 9.1.1, aK = K, and in particular there
exists some b E K satisfying ab = 1. As a > 0, b is positive as well. This
proves that x> is a(n ordered) group with respect to -, It remains to check
that > is also abelian. To show this, it suffi ces to observe that > is
o-minimal , and then to use Theorem 9.1.2. In fact > is definable (as a
group) in /C, and consequently every subset X of > definable in > is
also definable in /C, and hence is a finite union of non-empty intervals of
K. All the end points of these intervals lie in > U {+oo}, with the only
possible exception of the left most end point in the first interval, that might
equal 0, but can be replaced in this case by -00 in >. In conclusion, X
is actually a finite union of intervals in >. Hence > is o-minimal and
consequent ly abelian, as claimed.
Now let us prove that /C is real closed. Accordingly take a polynomial
f(x) E K[x] and two elements a < b in K satisfying f(a) . f(b) < 0, for
example f(a) < 0 < f(b). We have to show t hat there is some c E K such
th at a < c < band f(c) = O. Recall that /C is an ordered field, and hence
::; is dense in K and in la, b[. The polynomial function that f defines is
continuous (with respect to the order topology) , so both
la, b[+= {d E K : a < d < b, f(d) > O}
and
la, b[-= {d E K : a < d < b, f(d) < O}
are open sets. If la, b[-= 0, then the continuity of f is contradicted in a.
Hence la, b[- and similarly la, b[+ are not empty. Moreover both la, b[+ and
la, b[- are definable, and accordingly decompose as finite unions of intervals,
indeed of open intervals. As la, b[+ and la, b[- are disjoint , t here is some
c E]a, b[ out of both la, b[+ and la, b[-: so c is a root of [, f( c) = O. '"
9.2 The Monotonicity Theorem
Let M = (M, . ..) be an a-minimal struct ure. First obser ve t hat M
is a t opological space with respect to t he ord er t opology and so, for every
positi ve int eger n , M " is a t opological space as well, with respect t o th e
product t opology.
As already recall ed, t he o-minimalit y of M just requires t hat t he onl y de-
finabl e subsets are th e finit e union of singlet ons and open int er vals (possibl y
including half-lines and t he whole M) . But what can we say about th e
definable subsets of M" when n > I ? We give here a very partial and pr e-
liminary answer , dealing with l-ary definable functions f. We show t hat, if
the domain of f is an open interval la , b[ with a < b in M U {oo}, then
one can partition la , b[ into finitely many intervals such that, in each of
t hem, f is either const ant , or st rict ly incr easing, or strict ly decreasing, and
anyhow cont inuous according to t he order topology. This is t he so-called
Monotonicit y Theorem, saying in detail what follows.
Theorem 9.2.1 Let A1 be an o- minimal structure, X M, a, b E M U
{oo}, a < b, a and b be X -defi nabl e when belonging to M . Let f be an
X -defi nable f unction of la , b[ in to M . Th en there are a posit ive in teger n
and aD, al , .. . , an E MU {oo} suc h that
1. a = aD< al < ... < an = b, and aI , ... , an- l are X -defi nable;
2. fo r every i < n, f is either constant or strictly monotonic in ]ai , ai+d;
3. fo r every i < n , if f is strictlu monotonic in uu, ai+l[, then f( ]ai, ai+l [)
is also an in terval, and f contai ns a bij ect ion preserving or reversing
bet ween ]ai , ai+d and f (]ai ' ai+d )
In particular f is conti nuous in every interval ]ai , ai+d for i < n .
Notice t hat, when mor e generally f is an arbitrary definabl e function with
both domain and image in M , t hen the domain of f is definable as well,
and consequent ly is a finit e union of singletons and open int erval s. Each
of t hese int ervals satisfies t he ass umpt ions of Theorem 9.1.2 , and hence
inheri t s it s conclusions. Notice also t hat Theor em 9.1. 2 implies, as a simple
consequence, t hat, if M is an a-minimal st ruct ure, a, b E M U{oo}, a .
A full proof of Theorem 9.1.2, as stated before, would require several tech-
nical det ails and would be quite long. We prefer to propose here a simpler
9.2. THE MONOTONICITY THEOREM 319
argument showing only the continuity result when M expands (R, :S;) (ac-
tually this is more or less what we will need later).
Proposition 9.2.2 Let M be an o-minimal structure expanding (R, :S;) ,
X M, f be an X -definable function from ]a, b[ into R . Then there are
aI, ... , an E]a, b[ such that al < ... < an, aI, ... , an are X -definable and
f is continuous in ]a, al [, ]a
n,
b[ and each interval ]ai , ai+l [ for 1 :s; i < n.
Proof. Let S denote the set of the points in ]a, b[ where f is not continuous.
It is an easy exercise to prove that S is X -definable. If S is finite, then we are
done: for, aI, .. . , an are just the elements of S. Hence suppose S infinite.
So S contains an infinite open interval I. For every natural n, build two
infinite open intervals In and I
n
such that, for every n,
(i) In I ,
(ii) the (topological) closure I
n+l
of I
n+l
is included in In'
(iv) the length of I
n
is smaller than nil'
Let us see how to define these intervals. First put ID = I. Then take any
natural n , suppose In given and form I
n
and I
n+l
as follows. If f(In) is
finite , then for some d E f(1n) the preimage {c E In : f(c) = d} is infinite.
But {c E In : f( c) = d} is definable, and hence includes some infinite
interval; f is const ant , hence continuous, on this interval , which contradicts
In I S. Accordingly f(In) must be infinite. But f(In) is definable, too,
and hence includes in its turn an infinite interval; let I
n
be such an interval,
notice that we can assume with no loss of generality that the length of I
n
is
< nil' The preimage of I
n
in In is also definable, and contains some infinite
interval. Let I
n+l
denote such an interval; we can assume In+! In. This
determines the In's and the In's for every n. Now put I' = nnEN In. Clearly
I' = nnEN In' whence I' =1= 0 because R is compact. Pick d E 1' , we claim
that f is continuous in d (this contradicts dES and so accomplishes our
proof). Take any interval U containing f(d). Owing to (iv), there is some
natural n for which U ;2 I
n.
But this implies U;2 f(In+l) where I
n+l
is an
open neighbourhood of d. ...
A final remark. When M expands the field of reals, we can say even more,
and state a smooth version of the theorem: in fact, one can partition ]a, b[
into finitely many intervals where f is of class cm for every positive integer
m.
9.3 Cells
Now we move to characterize in an o-minimal st ruct ure M = (M, :S , . . .)
th e definable subsets of M": in particular t he definable n-ary functions, for
every positive int eger n. To make our life easier , we assume all through-
out t his sect ion (and also in th e following ones) t hat :S is dense without
endpoints: in particular every interval ]a, b[ with - 00 < a < b < + 00 in
M must be infinite. As alr eady said, the reason of t his restriction is just
to make our treatment and our proofs simpler; in fact , all th e results we
will show below can be extended -by the appropriate arr angement s- to any
o-rninimal st ruct ure M. On the other side , t he ord er of reals is j ust dense
without end point s, so our framework incl ude all the expansions of (R, :S ),
in particular all the st ruct ures enlarging t he real field; as we said before, the
notable o-minimal examples we will propose in Section 9.7 lie in t his set t ing.
We know that the basic definable subsets of M are t he int ervals and t he
singlet ons. More gener ally, th e basic definable subsets of M " are t he cells.
So let us define what a cell is, mor e preci sely what a k-cell in M " is.
Definition 9 .3.1 First suppose n = 1. A subset C of M is a O-cell if and
only if C is a si ngleton, and a 1-cell if and only if C is a non- empty open
in ter val, possibly with infi nite endpoints .
Now let n > 1, and let k a natural number :S n. A subset C of M " is called
a k -cell if and only one of the following conditions hold:
1. there are a k- cell D of M
n
-
1
and a continuous and defi nable fun ct ion
f of D into M such that C is the graph of I, namely
C = {(11, b) E M " : 11 E D , b = f(I1)} ;
2. k 1 and there are a (k - I )-cell D of M n-l and two functions f and
g with domain D such that
(i) either the image of f is a subset of M and f is both continuous
and definable, or f(l1) = -00 VI1 E D ,
(ii) either the image of g is a subset of M and g is both continuous
and definable, or g(l1) = + 00 VI1 E D ,
(iii) f(l1) < g (l1) VI1 E D,
(iv) C = {(11, b) E M n : 11 E D, f (l1) < b < g(I1)}.
One easily sees t hat every k-cell of Mn is definabl e, and t hat a k-cell of jU
n
is open in M" if and only if k = n . One can also observe
9.3. CELLS 321
Proposition 9.3.2 Let M be an o-minimal str ucture, k S; n be posit ive
integers. For every k-cell C of M" , there is a definable homeomorphism of
C onto a k- cell of M
k
Proof. It suffices to see, for n > k and n > 1, how to det ermine, for
every k-cell C of M
n
, a definable homeomorphism 1re onto some k-cell C' of
Mr:" , and then t o iterate th is procedure as long as one needs. First ass ume
that C is t he gr aph of some cont inuos definable func tion from a k-cell D of
M n-l into M. In this case it suffices to put C' = D, and choose as 1re th e
proj ection of C onto C' . Now ass ume
C = {(a, b) E M
n
: a E D, f (a) < b < 9 (a)}
where D is a (k - I)-cell of M n-l , and I, 9 sat isfy the cond it ions in 9.3.1,
2. We proceed by ind uction on n,
If n = 2, then k - 1 = 0, and D reduces t o a single point a of M. Put
C' =JJ(a) , g(a)[, 1rc(a, b) = b for every b E C'. Now let n > 2. We know
that there is some definabl e homeomorphism 1rD of D onto a (k- 1)-cell D' of
M n-2. Consider the t wo functions one gets by composing l , 9 respectively
and the inver se of 1rD. They have domain D' and sat isfy t he ass umpt ions in
9.3.1,2. Accordingly they define a k-cell C' of M
n
-
l
; furt her more
1rc(a, b) = (1rD(a) , b) V(a, b) E C
det ermines a definable homeomorphism of C onto C'. ...
When M expands a real closed field (for instance, when M is the real field,
or even an expansion of it) , the cells in M can be charact erized as follows.
Proposition 9.3.3 Let M be an o-minimal expansion of a real closed fiel d,
k , n be two natural nu mbers satisf ying n 2:: k , 1. If C is a k -cell of M n,
then there exists a definable homeomorphism of C onto ]0, I [k.
Proof. When k = 0, our claim is trivial, because a O-cell reduces t o a
singlet on. So t ake k > 0. Owing to the previous proposition, we can assume
n = k. We proceed by induction on n. If n = 1, t hen C = ]a, b[ where
a, b EMU {oo} and a < b. So the required homeomorphism between C
and ]0, I[ is easily obtained: for instance, if both a and b are in M , th en it
suffices to map every x in C = ]a, b[ into
x - a
b - a '
in t he remaining cases one proceeds in a similar way. Now suppose n > 1
(and our claim true for n - 1). Let C be an n-cell of M n. There do exist an
(n -I)-cell D of M n-l and two functions f and 9 as in 9.3.1 ,2 such t hat C
is t he set of t hose t uples (ii, b) in M" for which ii E D and f (ii) < b < g(ii).
By induction, t here is a definable homeomorphism h of D ont o ]0, l[n-l. If
both f and 9 take th eir values in 1\1, t hen

b) b- f (ii) )
a, a , g(ii) _ f(ii)
V(ii, b) E C
is the required homeomorphism. All t he other cases can be handled III a
similar way, just as when for n = 1.
In particul ar, when M expands th e real field Rand k > 0, every k-cell of
M is connected in th e ord er t opology; for, ]0, l[k is. This does not hold any
longer when R is replaced by any real closed field. For instance, if Ra is t he
ordered field of real algebraic numbers, th en Ra is real closed, and hence
o-minimal; Ra is a Lcell of it self, but is not connected because, for every
real t rascendental t , Ra parti tion s as
Ra = {r E Ra : r < t } U{r E Ra : r > t}.
Hence connect ion gets lost. However ever y cell in an o-minimal st ruct ure
M sat isfies a weak form of connection, wit h respect to open definable sets .
In fact , consider t he following notion.
Definition 9.3.4 Let M be an o-minimal structure, n be a posit ive integer.
A defi nable set X M" is said to be definably connected if and only if
X cannot part it ion as the disjo int union of two non-empty open definable
subsets.
Wh at we will see now is t hat every cell in an o-minimal st ruct ure M is de-
finably connected. First we give an equivalent charact erization of definabl y
connected set s.
An open box of M
n
is t he cartes ian product of n open int erval s of M . Hence
open boxes form a basis of open neighbourhoods of t he product topology
of M n. Furthermore, for Y X u, an element s of X is called a
boundary point of Y in X if and only if every open box of M n containing
ii overlaps both Y and X - Y.
Lemma 9.3.5 Let A1 be an o-minimal structure, n be a posit ive integer,
X be a definable subset of M": Then X is defi nably connect ed if and onl y
if, for every proper non-empty defi nable subset Y di X , X contains at least
one boundary point of Y in X.
9.3. CELLS 323
Proof. X is definably connected if and only if, for every definable Y X
such t hat Y =1= 0, X , eit her Y or X - Y is not op en , in other words ther e is
eit her a point aE Y such that every open box cont aining aoverl aps X - Y ,
or a point aE X - Y such that every open box cont aining aoverl aps Y .
Accordingly X is definably connected if and only if, for every Y as above,
t here is aE X which is a boundary point of Y in X . ..
Now we can show, as promised,
T heorem 9.3.6 Let M be an a-minimal structure, n be a posit ive integer.
Then every cell C of M n is defi nably connected.
Proof. We proceed by induction on n .
If n = 1, t hen t he claim is trivial becau se the cells of M red uce to singletons
and open int ervals.
Hence assume n > 1. Let C be a cell of J\;f n . If C is a O- cell, so a singleton,
t hen C is definably connected. If C is a k-cell for some positi ve int eger
k < n, t hen, owing t o Corollary 9.3.3, C is definably homeomor phi c t o a cell
C' of M
k
; by t he induction hypothesis, C' is definably connected, whence
C is definably connected , t oo. Finall y suppose t hat C is an n-cell of M n.
Then ther e exist a cell D of M n-l and two fun ctions 1 and 9 as in 9.3.1, 2
such that
C = {(a, b) E M
n
: a E D, 1(a) < b < 9(a) } .
D is definably connected by t he induction hypoth esis. To deduce that even
C is definably connected we use Lemma 9.3.5. Accordingly take a definabl e
subset Y of C such t hat Y =1= 0, C. First suppose t hat, for some aE D , t here
are t wo eleme nts b an d b
'
in M such t hat (a, b) E Y and (a, b
')
E C - Y.
Then t he int erval ]1(a), g(a)[ contains at leas t one boundary point b
o
of t he
definable set {b E M : (a, b) E Y}; t his implies t hat (a, b
o)
is a bo undary
point of Y in C. Now suppose t hat , for every aE D , either
{ (a, b) E C : bE M} Y
or
{ (a, b) E C : s M } C - y.
Let Z denot e t he set of t he eleme nts aof D satisfying t he for mer condit ion.
Y =1= 0, C implies Z =1= 0, D . As D is definabl y connected, t here is some
boundary point ao of Z in D. Let b E M sat isfy (aa , b) E C, we claim that
(aa, b) is a boundary point of Y in C. Let B an op en box of M n containing
(aa, b). As C is open , we can suppose B C. Let B
'
denote the projection
of B in M
n
-
1
. Then ao E B
'
and B
'
D. Since ao is a boundary point
of Z in D, there are a1 and a2 in B
'
such t hat a1 E Z and a2 tj Z . Conse-
quently, if b
1
, b
2
E M satisfy (a1 ' b
1
), (a2 ' b
2
) E C respectively (in particular
(a1 ' b1), (a2' b2) E B) , then (a1' b1) E Y, (a2' b2) tj Y. Hence (ao, b) is a
bou ndary point of Y in C. In conclusi on, C is defina bly connect ed. ..
9.4 Cell decomposition and other theorems
The aim of this secti on is to introduce and to state t he basic general theorems
for o-rninimal struct ures M = (M, . . .): we want to characterize all the
definable sets and funct ions in M , and we want also to emphasize some
relevant consequences and, among t hem, the already ment ioned fact t hat 0-
minimalit y is preserved under elementary equivalence. As th e proofs of t hese
fundamental results are qui t e long and intricate, we will defer part of t hem,
and t he cor responding details, to t he next section; here we provide just a
basic outline, illustrating these cent ral cores of the theory of o-minimalit y.
So a reader simply int erested in a general view may limit her , or his attent ion
t o t his section, and to skip the next one. Let us remind once again that , for
simplicit y's sake, we are assuming that (M, is dense without endpoints:
this is tacitly acce pted all throughout these sections, unless otherwise stated.
The first result we propose j ust describes definabl e sets (and functi ons) in
o- rnini mal struct ur es. It is a beau ti ful and powerful characterization, called
Cell Decomposition Theorem. In fact , it says t hat every definabl e set de-
composes as a fi nite union of cells.
Theorem 9.4.1 Let A1 be an o-minimal structure, n be a positive integer.
1. Eve ry definable set X M " can be expressed as a fi nit e (disjoint)
uni on of cells in M n.
2. Furthermore, if X is the domain of a definable funct ion f with values
in M, then one can decompose X as a fin it e disjoint union of cells,
such that f is contin uous on each of them.
Not ice that t his generalizes what we know when n = 1; in fact, in that case
every definable X M is a finit e union of point s and open int ervals (in
other words, of O-cells and I-cells respectively), and, when X is th e domain
of some definable function I, one can also suppose that f is continuous on
each of these pieces, owing to the Monotonicity Theorem. But now we can
ext end t hese results t o any n .
9.4. CELL DECOMPOSITION AND OTHER THEOREMS 325
As already said , t he proof of the Cell Decomposition Theorem will be given
in detail in t he next sect ion. But Cell Decomposition, and it s analysis
of definable sets, is also t he key t ool in showing another basic fact in 0-
minimality, that is it s preserving under elementary equivalence.
Theorem 9.4.2 If M is an a-minimal structure, then the theory of M i s
a-minimal.
In fact , which is t he trouble in this claim? Actually we do know that, for
an o-minimal M, for every formula O( v, w) in the language L of M and
t upl e d in M , O(M , il) is a finite union of (possibly closed) intervals . But
suppose that , when ii ranges over M, the minimal number of t he intervals
involved in these decompositions of O(M , il) cannot be upperly bounded ,
and so, for every natural N, one finds some tuple il( N) in M such that any
decomposition of O(M , il( N)) requires at least N intervals. If this is t he case ,
then it suffices a st raightforward application of Compactness Theorem to
provide some M' == M and some t uple a' in M' for which O(M' , a') cannot
be expressed as a finite union of int ervals , and consequently to conclude t hat
the t heory of M is not o-rninimal.
Hence the cr ucial point in showing that the o-rninimality of A1 is preserved
by element ary equivalence is to uniformly bound, for every form ula O( v, w)
as befor e, the minimal number of intervals necessary to decompose O(M , il)
when il ranges over M. This is a definability question concerning formulas in
arbitrarily many free vari ables, and so directly refers to Cell Decomposition.
On the other hand, boundi ng t he number of the invol ved int ervals in O(M, il)
is the same as bounding the total number of their endpoints (forming a
definable, and finite set). So the key ste p towards the proof of Theor em
9.4.2 is
Theorem 9.4.3 Let M be an a-minimal structure in L , Ifl(V, w) be an L-
formula such that, for all il in M , Ifl(A1, il) is finit e. Th en there is a positive
integer N such that, for every il in M , IIfl(M, il) I ::; N.
The proof of Theorem 9.4.3 will be deferr ed until the next section. But , as
we have just pointed out , Theorem 9.4.2 is an almost immedi ate consequence
of Theorem 9.4.3. Let us see in detail why.
Proof. (Theorem 9.4.2) Let L be the language of M , and let 7)( v , w) be an
L-formula; in particular, let n deno te the lengt h of w. For every d E M";
O(M, il) is a finit e union of intervals. Let Ifl(v, w) be t he L-formula saying
v is an endpoint of 0(... , w).
Then, for every ii E M."; cp( M, ii) is finite; hence, by Theor em 9.4.3, there
exists a posi tive int eger N such that , for every ii E M" , cp (M, ii) cont ains
at most N elements, whence B(M , ii) has at most N endpoint s. But t he
sentence
v cp( v, w)
remains t rue in every st ruct ure M' == M. Accordingly, for every A1' == M
and bE u, B(M' , b) has at most N endpoint s, and hence is a finit e union
of int ervals. In conclusion t he t heory of M is o-minimal. ..
The bounds given by Theor em 9.4.3 on formul as B(v, w) ca n be extended
to formulas B(V, w) with an ar bit rarily long v. In fact the following result
holds.
Theore m 9 .4.4 Let M be an o-minimal structure in L, B(V, w) an L-
formula (n be the length of vand m be that of w) . Then there exists a posit ive
integer N such that , for every M' == M and tupl e ii in M' , B(M, n, ii) is the
union of at most N cells in M' .
Before beginning t he proof, and just for pr eparing it , let us premi t a simple
exa mple. Assume n = 2. Let B(VI, V2, w) be an L-formul a, ii be a t uple in
M. Suppose t hat t he definabl e set B(M
2
, ii) decomposes as a disjoint union
of 2 cells in NI : the former is a singlet on, so a O-cell, while t he lat ter is a
I-cell, and more precisely is t he graph of a cont inuous definable func tion f
whose domain is t he open int erval ]a, b[ wit h a < bin M . Notice t hat t here
are an L-formula 1](V I, V2 , Z) and a sequence ein J\1 such t hat , for every Cl
and C2 in M with a < Cl < b,
Now consider the L-formula
1\ "1] (', " Z) defines a cont inuous function of domain ]u, w["
1\ 1]( Vl, V2, Z))) 1\ -, (u < UI < W 1\ 1]( UI ' U2, Z) ))).
It is clear t hat t he t uples bin M , or even in a model M' of t he t heory
of M , satisfying N F B(i (b) are just t hose for which B(M,2, b) has a cell
decomposit ion as 'l9( M
2,
ii) (t he disjoint union of a singlet on and a graph of
9.4. CELL DECOMPOSITION AND OTHER THEOREMS 327
a definabl e continuous function) . It is also obvious t hat t his is quite gener al:
given an L-formula B(v, w) as in t he statement of 9.4.4, a t uple 5 of th e same
length as win t he universe n of t he t heory of M and a cell decomposition of
'l9(nn , 5), one can build an L-formula B
ii
( w) such t hat a t uple bin n satisfies
B
ii
if and onl y 'l9(nn , b) has a cell decomposition just as 'l9(n n, 5).
Proof. (Theorem 9.4.4). Owing t o Theorem 9.4.2, any st ructure M' ele-
ment aril y equivalent to M is o-minimal, and hence, by Theorem 9.4.1, any
definabl e subset of u (in part icular B(M
m
, 5) for 5 E M' m) is a finit e
union of cell. We have seen t hat t here is an L-formula B
ii
( w) (without pa-
ramet ers) describing t he form of a given cell decomposition of B(M, n, 5).
Let denote the (countabl e) set of all these formul as B
ii
(1V) when 5 ranges
over M' n and M' is a model of th e th eory of M. Use Compactness Theorem
and get finitely many formul as
such t hat
Vw V'l9i(W) E Th (M ).
i :s
Consequent ly t here is a positi ve int eger N such th at , for every M' == M
and 5 E M'?" , B(M,n, 5) decomposes as the union of at most N cells: N is
just t he maximal number of involved cells in t he decomposit ions described
by 'l9
o
(w), ... , 'l9
s
(w).
Recall t hat , when A1 expands t he real field R , every cell of M is also
connected. Hence in t his case, for every formul a B(V, w), t here is a positive
integer N such that , for a t uple 5 in Mm, B(M
n
, 5) is t he union of N
connected components . vVe will see in Section 9.7 several relevant examples
of o-rninimal expansions of t he real field R. In this fr amework it is worth
stat ing the following result of Wilki e' s.
Theorem 9.4.5 (Wilkie) Let M expand the ordered fi eld of reals by Coo
fu nctions from some cartesi an powers R
t
of R into R. Assume that, for
every quantifier free fo rmula B(v, w) in the languag e L of M , there is a
posit ive integer N such that, for any tupl e 5 E Mm, B(Mn, 5) decomposes
as the union of N connect ed components . Then the same is true for every
fo rmula of L. In part icular, M is o-minimal.
Cell Decomposition is an import ant t ool also in developing a dimension
t heory inside o-mini mal struct ures M , and in equipping every definable X
in M with a natural number (it s dimension). Recall t hat no o-minimal M is
simple, and hence w-stable; accordingly t he Morley rank of X might be 00.
More generally, as simplicity fails, other possibl e rank notions ari sing wit hin
simple, or stable, or superstable settings and replacing RM in t hese enlarged
fr ameworks lose t heir interest in o-minimal mod els. But Cell Decomposition
does assign a dimension t o X in a qui te reasonabl e way. Let us see how.
Definition 9.4.6 The dimension of a k- cell of M
n
is k. The dimension
of a definabl e X in M is the maximal dimension of a cell arising in a cell
decomposition of X.
Of course, a cell decomposition of X is not unique; but the maximal di-
mension of an involved cell is, and does depend only on X. So the previous
definition makes sense for every X. There is another alte rnat ive way t o
introduce a dimension notion in M. In fact , as we saw in Chapt er 5, t he
alge braic closure acl det ermines a dependence relation in M (and in every
model M' of t he t heory of M ). T his relation generates in it s t urn an inde-
pendence system as axiomatized in sect ion 7.2, so satisfying t he condit ions
(11)-(16) (but not t he fur th er requirement (17) )). Wi th respect to this in-
dependence notion, we ca n define the dimension of a t upie ii = (al' . . . , am)
in M'": as t he size of a minimal subsequence b such t hat ii lies in acl(b) .
Then we can introduce, just as in t he st rongly mini mal case , t he dimen-
sion of a definable set <p (M ,m, b) as t he maximal dimension of a tuple ii in
<p (/vt"m, b) where M" is an elementary extension of A1'.
Notably t hese dimension notions (t he former ari sing from topology, the latter
more directl y related t o mod el t heory) coincide and share good properties
and, afte r all , a satisfactory behaviour.
To conclude this section, let us mention some other nice properties of 0-
minimal struct ures and theories , closely resembling what happens in th e
w-stable set t ing.
Theorem 9.4.7 (Pillay-St einhorn) Let T be a (compl ete) o-minimal theory,
A be a small subset of the universe n of T. Then there is a model of T prime
over A, and this is uni que up to isomorphism fi xing A point wise.
Act ua lly t his resul t does not need t he Cell Decompositio n Theor em, but can
be proved by referring directl y t o t he definition of o- rninimalit y and t o t he
Mo notonicity Theor em. The same ingredients yield a nice classification of
o-minimal t heories, again due to Pill ay and Steinhorn.
9.5. THEIR PROOFS 329
The uncountable spectrum of an o-minimal T th eory takes everywhere t he
maximal value
I(T, >. ) = 2" V>. >
With respect to th e count able framework, it is wort h emphasizing t ha t
Vaught 's Conject ure holds, in t he following st rong form.
Theorem 9.4.8 (Mayer) Let T be a (complete) o-minimal theory. Th en,
up to isomorphism, either T has conti nuum many countable models, or there
are two naturals n and m such that T has 3
n
. 6
m
countable models.
Of course, one might get curious in reading t he st ate ment of this th eorem:
why, and where, do n and m arise? Basically, they depend on a careful
analysis of types in o-minimal structures. The interested read er may directly
consult Laura Mayer 's work, quoted below.
9.5 Their proofs
This section provides t he proofs of Theorems 9.4.1 and 9.4.3 , stat ed in Sec-
tion 9.4. As said, th ey are long and intricat e. In spit e of this, we think
it right to propose th em for at least two reasons. Th e form er (and th e
principal) is t hat we believe t hat Theorem 9.4.1 (t he Cell Decomposition
Theorem) and Theor em 9.4.2 (the one saying t hat o-rninimalit y is pres erved
under element ary equivalence) are two beau tiful and fundamental results
and deserve a full report , including a t echni cal preliminary like Theorem
9.4.3. The latter reason just concerns t he intricacy of t he proof; actually
this is due t o it s length and ingenui t y, but does not depend on a relevant
and deep t heory, indeed t he premi ses it needs are rather element ary and
accessible (they just include the definition of o-minimality, t he Monotonic-
ity Theor em, some properties of cells and an induction argument). So our
exposition should require no pa rticular efforts but a lit tl e attention and pa-
t ience . And anyhow the reader who is not interested in th ese details may
neglect this section and proceed directly to th e next ones, t hat will not use
these proofs.
So conside r an o-minimal st ructure A1 in a language L ; for simplicity, we
keep our ass umption t hat t he order of A1 is dense without endpoint s.
What we said in the last section in introducing Theorem 9.4.3 suggests t he
following definition.
Definition 9.5.1 Let <p( v, ill) be a formula of L(M) ! n be the length of ill,
X be a subset of M": We say that:
tp(v, w) is finite in X if and only if, for every ii E X, tp(M, ii) is
fin it e;
tp(v, w) is uniformly finite in X if and only if there is a posit ive
integer h such that , fo r every ii EX, the size of tp(M, ii) is at most h.
When tp(v, w) is finite in X , we can introduce two parti al functions tp_
and tp+ mapping any ii E X into t he minimal and the maximal element in
tp(M, ii) respectively -provided t hat tp(M, ii) is not empty, of course-. If X
is definable, then tp_ and tp+ are definable as well.
Definition 9.5.2 Let n be a positive integer, X be a definable subset of
M": A decomposition of X is a part it ion of X into fin it ely many cells.
If y X is defi nabl e, we say that a decomposition P of X partitions Y
when no cell in P overlaps both Y and X - Y.
Here is anot her t echnical preliminary notion.
Definition 9.5 .3 Let C be an open cell of M n (so an n-cell} , tp(v, w) be a
fo rmula fin it e in C . Call a point ii E C good for' tp(v, w) if and only if the
following conditions hold:
1. for every b E tp(M , ii), there are an open box B C containing ii and
an open interval I containing b such that tp(A1
n
+l ) n (I X B) is the
graph of some continuous funct ion of B in I ;
2. fo r every b E M - tp(M, ii) , there is an open neighbourhood of (ii, b)
in lU
n
+
1
disjo int f rom tp(Mn+l ).
A point ii E C which is not good for tp(v, w) is called (with no particular
imaginat ion) nasty for tp(v, w). Not ice t hat both good and nast y point s for
tp(v, w) form definabl e subsets of t he cell C.
At this point we ca n begin our proof. The following lemma is it s crucial
step.
Lemma 9.5.4 Let M be an o-minimal structure, n be a positive integer, C
be a cell of M
n.
(l) n For every eleme nt i in a fin it e set I of indexes, let Xi denote a definable
subset of C. Then there exists a decomposition of C partit ioning each
Xi.
(2)n Let f be a definable function of C into M . Then there is a decomposi-
tion P of C such that f is continuous on any cell of P.
(3)n A formula <p(v , w) of L(M) finite in C is uniformly finite in C.
(4) n Let <p(v , w) be a formula finite in C. If the functions <p_ e <p+ are
defined and continuous in C and every point of C is good for <p(v, w),
then the size of <p(M, if) is constant when if ranges over C.
Lemma 9.5.4 immediately implies both Theorem 9.4.1 and Theorem 9.4.3.
Proof. (9.4.1.1) M" is an n-cell. If X is a definable subset of M", then (l)n
in Lemma 9.5.4 provides a decomposition P of M" partitioning X. Hence
X is the (finite) union of the cells of P it contains. ..
Proof. (9.4.1.2) Just apply (2)n to the cells of the decomposition of X given
by 9.4.1 , 1. ..
Proof. (9.4.3) This is just (3)n when C = M": ..
Now let us show Lemma 9.5.4. We proceed by induction on n.
(1h This just rephrases o-rninimality.
(2h This is the Monotonicity Theorem (in the weak form we saw in 9.2).
(4h If C reduces to a singleton, then the claim is trivial. Hence assume
that C is an open interval ]a, b[ where -00 ::; a < b ::; +00. Suppose
that, for some positive integer h, the set Y of the points c in ]a, b[ such
that 1<p(M, c)1 = h is not empty and does not equal ]a, b[. As <p_ and <P+
are defined throughout the interval C, we can assume h > 1. Let c be an
end point of Y in C, and put <p(M, c) = {do, ... , dd where L is a suitable
natural and do < ... < dt. . As c is good for <p, there are an interval I b[
containing c and L +1 pairwise disjoint intervals .la, ... , JL such that , for
every i ::; L, J; includes d, and <p(M
2
) n(Ji X 1) is the graph of a continuous
function gi of I into Ji, Each function gi is clearly definable. Furthermore,
for every c' El,
go(c
/)
< ... < gL(c
/);
consequently 1<p(M, c') I 2 L + 1. But c is an end point of Y in C, and
1<p(M, c) I = L + 1. Therefore, for some open interval I' having c as a left
or right endpoint,
j<p(M , c/)1 > L +1, 'Vc' El
'.
Assume for simplicity that c is a left endpoint of 1'. Define the following
function 9 in 1': for every c' E 1
',
g(c') is the least element d' E tp(M, c') such that d' =F gi(C') for every i L.
9 is definable and its domain coincides with the whole interval I'. So, as
observed in 2, 9 has a limit in MU {oo} when x -+ c+ .
d = lim
x
--+
c
+ g(x).
Moreover
tp_(c') < g(c') < tp+(c') Vc' El' ,
and so d cannot equal +00 or -00, in other words d E M. If 1= tp(d, c),
then d = d; for some i L , and consequently
on the other side , we know that, for every c' E I' ,
g(c') =F gi(C') , g(c'), gi(C') E tp(M, c').
But we contradict in this way the fact that c is good for tp (recall Definition
9.5.3, 1). Accordingly 1= -.tp(d, c), which again excludes that c is good for
tp (this time by 9.5.3 ,2) . This yields the required contradiction. Hence (4h
holds.
(3h The claim is trivial when C is a singlet on. Accordingly suppose C =
la, b[ where -00 a < b +00. The set of the element s c of C for which
tp(M, c) =F f/J is definable, and consequently is a finite union of singletons
and open intervals in C. Of course, it is enough to show that tp(v, w) is
uniformly finite on every interval in this decomposition. Accordingly we can
even assume
tp(M , c) =F f/J Vc E]a, b[;
in particular tp_ and tp+ are defined throughout ]a, b[. Owing to (2h , we can
even suppose (up to replacing again la, b[ with a suitable subinterval) that
both ip: and tp+ are continuous in ]a, b[. Now let Y denote the set of those
points in C =la, b[ that are nasty for tp. If Y is finite , say Y = {co, ... , Ct }
with a < Co < .. . < Ct < b, then ip: and tp+ are cont inuous in each interval
la, co[, ]Ci, ci+d for i < t, ]Ct, b[, and every point in these int ervals is good
for tp. So we are just in a position to apply (4h, and accordingly the size of
tp(M, c) is constant throughout every interval, and, in conclusion, tp(v , w)
is uniformly finite in la, b[.
Hence it suffices to show that Y is finite. Suppose not. Anyhow Y is
definable, and hence, by o-rninimality, it contains some infinite int erval.
Without loss of generality, we can assume that this interval just equals
C =]a, b[, and so that every point e of ]a, b[ is nasty for <.p. Then, for every
e E]a, b[, there exists d E M satisfying one of the following conditions:
(i) d E <.p(M, e) but, for no pair of open intervals I and J containing e and
d respectively, <.p(M
2)
n (J X 1) is the graph of a function of I in J;
(ii) d rJ. <.p(M, e) but every open box B containing (d, e) overlaps <.p(M)2
as well.
In both cases we say that (d, e) is a black sheep of <.p (of type (i) or (ii)
according to whether (i), or (ii), holds).
Observe that, for every e E]a, b[, there is a minimal dE M for which (d, e)
is a black sheep of <.p. In fact , as <.p( v , w) is finite in ]a, b[, there are at
most finitely many d E M such that (d, e) is a black sheep of type (i).
Consequently it suffices to prove that , if d
l,
d
2
E M, d
l
< d
2
and, for every
d E]d
l
, d
2
[, (d, e) is a black sheep of type (ii) of ip, then also (d
l
, e) is a black
sheep of <.p. This is certainly true when d
l
does not belong to <.p(M, e) (in this
case (d
l
, e) is a black sheep of type (ii)). Accordingly assume d
l
E <.p(A1, e)
and fix two open intervals I and J containing e and d
l
respectively. If
d E In]d
l,
d
2[,
then d rJ. <.p(M, e) and any open box including (d, e) -in
particular, any open box in J X I including (d, e)- intersects <.p(M
2).
Then
<.p(M
2)
n(J X 1) cannot be the graph of a continuous function of J in J , and
(d
l,
e) is a black sheep of type (i) of <.p.
Therefore we can consider the function g mapping any e E]a, b[ into the
minimal d E M for which (d, e) is a black sheep of <.p. g is definable, and
so, owing to (2h, we can find an open interval I in ]a, b[ such that g is
continuous in I and one of the following conditions holds:
(iii) for every e E 1, (g(e), e) is a black sheep of type (i) of sp;
(iv) for every eEl , (g(e), e) is a black sheep of type (ii) of <.p.
As before, we can suppose that I is just ]a, b[. Assume (iii). Introduce two
functions gl and g2 as follows. For every e E]a, b[,
glee) is the maximal element dE <.p(M, e) satisfying d < gee), if such
an element exists, and is -00 otherwise;
g2(e) is the minimal element dE <.p(M, e) satisfying d > gee), if such
an elements exists, and is +00 otherwise.
Both gl and g2 are definable (in the obvious sense) ; moreover we can sup-
pose that gl is cont inuous or const ant ly -00 in ]a, b[ and, similarly, g2 is
continuous or constant ly +00 in ]a, b[. Choose c E]a, b[, d
1
, d
2
E M such
that
By the definit ion of gl , g2 and the continuity of gl, g2, g, t here exists an
int erval I ]a, b[ containing c such t hat, for every c' E J,
Hence again t he definition of gl and g2 implies that <p(M)2 n(]d
1
, d
2
[x I) is
the graph of the cont inuous function g, whence (g(c), c) is not a black sheep
of <p oSo (iii) cannot hold .
Accordingly assume (iv). Let c E]a, b[. Clearly g(c) =I <p_(c). If g(c) <
<p_(c) , then one can use the cont inuity of .p: and 9 and determine an open
box containing (g(c), c) and disjoint from <p (M ). But this cont radicts the
fact that (g(c), c) is a black sheep of <po So <p_(c) < g(c). In t he same way
one proves g(c) < <p+(c). Now let us introduce two fun ctions gl and g2 as
follows. For every c E] a, b[,
gl (c) is the maximal element d E g(c).
Both gl and g2 have domai n ]a, b[ because, for every c E]a, b[, .p: (c) <
g(c) < <p+(c) an d <p(M, c) is finit e. We can eve n ass ume t hat gl and g2 are
cont inuous in ]a, b[. The Monotonicity Theorem provides, for any c E] a, b[,
two intervals I , .I containing c, g(c) respectively and sat isfying, for every
c' E I ,
9 (c') E .I, gl (c'} , g2 (c') rJ. J.
Hence <p (M
2
) n (.I X I) = 0, which contradicts the fact that (g(c), c) is a
black sheep of type (ii) . This excludes also (iv) .
In conclusion Y cannot be infinite. As already poi nt ed out , this implies our
claim: 1. Ass ume (l) j , (2) j , (3) j and (4)j for 1 ::; j < n.
(l) n Let C be a cell of M n. For every i E I let Xi be a definable subset of
C. We are looking for a decomposition of C part it ioning each Xi. If C is a
k-cell for k < n , then there is some definab le homeomorphism ITa of C onto
a cell C' of M n-1 . By (1)n- 1t here is a decomposition p I of C' part itioning
each 7!'C (Xi ). Through 71' 0
1
, p' can be lift ed to a decomposition P ofC
partitioning every Xi.
So ass ume t hat C is just an n-cell . Consequently t here are an (n - I )-cell
C' of A1
n
-
1
and two funct ions f and 9 satisfying 9.3. 1, 2; in part icul ar
C = Ha, b) E M" : a E C', f(a) < b < g(a)}.
Let 7!'C be the projection of C onto C'. By (1)n-1 there exists a decompo-
sition P' of C' partitioning each 7!' c(Xi )' For every Y' E P' , let
Y = {(a, b) E C : a E Y'} .
Clearl y these sets Y partition C when Y' ra nges over P'. So it suffices to
show that , for every Y,
(v) t here is a decomposition of Y partitioning each set Xi for which 7!' c(Xi)n
Y'-# 0.
To simplify t he notation, assume without loss of generality that 7!'c(Xi) n
Y' -# 0for every i E I. Fi x i E I and, for ever y aE Y' , consider
Xi(a) = {bE M: (a, b) E Xi} '
Th ere is a formula <Pi (v, w) (possibly wit h parameters from M) such t hat,
for every a E Y' , <pi (M, a) = Xi(a) . Moreover Xi(a) is a non-empty subset
of ]j(a), g(a)[. For every a E Y' , let Bi(a) denote the set of the end points
of Xi(a) in ]f(a) , g(a)[. There is a formula (}i(V, w) (with parameters in
M) such t hat, for every a E Y' , (}i(M, a) defines Bi(a) . (}i( V, w) is finite in
Y' and Y' is a cell. (3)n- l implies that (}i( V, w) is uniformly finit e in Y'.
Accordingly t here exists a positive int eger hi such t hat, for every aE Y' ,
IBi(a)1 hi . Xi is th e finit e union of t he tupIes (a, b) where a E Y',
s Xi(a) and IBi(a)1 = h for h = 1, . .. , hi. Wi thout loss of generalit y, we
ca n even suppose that IBi(a)1 = hi for every a E Y'. Consequently we can
define hi functions from Y' into M
ft , ... , n,
mapping any a E Y' int o t he first , . .., t he hi-th element of Bi (a). All t hese
functi ons are definable; by (2) n-1, we can assume that t hey are cont inuous
on Y'. Unless partitioning agai n each Xi , we ca n also suppose t hat, for every
i and j in I , h = 1, .. . , hi and h' = 1, ... , hj , exactly one of t he following
cases holds:
Va E Y' , f i(a) =
vsE Y' , <
vsE Y' , >
Accordingly we can rearrange t he functions (i E I , 1 h hi ) and form
a new sequence
go , . .. , gt
such t hat, for r 8 t, gr(a) < gs(a) for every a E Y'. But t his implies
that t he sets of t he tuples (a, b) E Y such that aE Y' and
f(a) < b < go(a),
gs(a) < b < gs+t{a) (8 < t) ,
gt(a) < b < g(a) ,
b=gs(a)
respecti vely, form a decomposition of Y parti tioning each X i , as claimed.
This concludes t he proof of (l)n.
Acco rdingly ass ume fro m now on also (l)n .
(2)n Let C be a cell of A1
n
and let f be a definabl e function from C into
M . Wh at we have t o find is a decomposition P of C in cells wher e f is
continuous. If C is a k-cell for some k < n , t hen t here exists a defin able
homeomorphism KC of C onto a cell C' of J\I1 n- l . By (2)n-I ' ther e is a
decomposition P' of C' in cells such that f K
C
I
is continuous on each of
them, and K
C
I
lift s P' to a decomposition P as required .
So assume that C is an n- cell, in other words an open cell in M n . Let
Cl be the set of tuples (a, b) E C such that
defines a continuous function f b on some open box B of M n-l con-
t aining aand satisfying B X {b} C,
C
2
be t he set of tupies (a, b) such t hat X
n
J----7 f (a, x
n
) defines a function fa
eit her constant or strictly monotonic on some open int er val I cont ain-
ing b and satisfying {a} x I X , and, in t he lat t er case, also f(l ) is
an op en interval and f is a bijection bet ween I and f(l ).
Use (l)n and get a decomposition P of X partitioning both Cl and C
2
So
it is enough to show that f is continuous on every cell of P, and even on
every open cell D of P, owing to what we observed at the beginning of this
point. Then there are a cell D' of Mn-l and two functions ft, h satisfying
9.3.1,2 and
D = {(a, b) E l'vr a E D' , ft(a) < b < h(a)}.
Notice:
(vi) D < Cl .
In fact let b E M satisfy (a, b) E D for some suitable a E u:', As the
domain of 9 includes D, fb is defined in some open subset of D'. By (l)n-l
and (2)n-l, there is some open cell in D' where fb is continuous. For ain
this cell, (a, b) E Cl. Hence D nCl =I- 0. As P partitions Cl, D Cl.
Now we claim
(vii) D C
2
and, for every a E D' , fit is either constant or strictly mono-
tonic in]ft (a), h(a)[ (and, in the latter case, the image of]ft (a), h(a)[
is an open interval and fit is a bijection between these intervals).
D C
2
can be shown by proceeding as for Cl. Now take aE D'. Owing to
the Monotonicity Theorem, there are a natural m and b
l
, ... , b
m
E M such
that
fl(a) < bl < .. . < b-: < ft(a),
fit is either constant or strictly monotonic in each interval J among jjj (a), b
l
[,
]bj, bj+l[ (for 1 ::; j < m) and ]b
m
, h(a)[, and, in the latter case, even fit(J)
is an interval and fit induces a bijection between J and fit (J). Choose m
minimal. If m = 0, then the only involved interval J is just ]fl(a), h(a)[,
and (vii) is trivial. On the other side, if m > 0, then fit is neither constant
nor strictly monotonic in any open interval containing b
l
; as (a, bd E D
and D C
2
, (a, bt) E C
2
, which contradicts the definition of C
2
So m = 0,
and we are done.
At this point, we are in a position to conclude the proof of (2)n' In fact, let
(a, b) E D, J be an open interval containing f(a, b); we are looking for an
open box B of Mn including (a, b) and satisfying f(B) J. Owing to (vii),
there is a closed interval I = [b
l
, b
2
] of M such that b is in the interior of I,
I h(a)[ and fit(I) J. By (vi) there exist two open boxes n, and
B
2
of M
n
-
l
both containing a, and satisfying the additional conditions
n. X {bi} D, f(Bi ' bi) J Vi = 1, 2.
Let B' be an open box of M
n- 1
such that B' B
1
n B
2
, ii E B'. Then
f (B'x 1) J . In fact, let a
'
E B' , b' E I , so f( a
l
) < b
1
because B' x{bd D
and b
2
< 12 (;' ) because B' X {b
2
} D. Consequent ly
it (a
')
< b
1
< b' < b
2
< 12 (a
l
) .
Furt hermore f (;' , bt) E J , f(;' , b
2
) E J. As fa! is eit her constant or st rictly
monotonic in ]f1(;') , 12(;')[, f (;' , b') E J. Accordingly f(B' X 1) J. This
accomplishes our proof of (2)n.
So we can ass ume even (2) n tr ue from now on. At t his point let us deal with
(3)n.
(3)n Let i.p (v, W1 , ... , w
n
) be a formula, C be a cell of M n such t hat
i.p (v, W1, . . . , w
n
) is finit e in C. The case when the dimension of C is strictly
smaller t ha n n can be handled as befor e. So we can limit our analysis to
t he case when C is an n-cell, in other words is open. Let
Cl be t he set of those point s (ii, b) E C such t hat ii is good for i.p (v, W1 , ... ,
W
n-1
, b) ,
C
2
be t he set ot the point s (ii, b) E C such t hat b is good for i.p (v, ii, w
n
)
respectively (ii abbreviat es here (a1 ' . . . , a
n-1)).
By (l) n' there is a decom-
position P of C parti tioning both Cl and C
2
We claim
(viii) for every open Y in P, Y Cl and Y C
2
In fact , let (ii, b) E Y , and let B be an open box in Y including (ii, b). Th e
projecti on B' of B onto t he first n - 1 coordinates is an open box of M n- 1.
By (3)n-1 , i.p (v, W1, . . . , W
n
-1 , b) is uniforml y finite in B' . By (1)n-1 and
(2) n-1, there are an open cell C' B' and a posi t ive integer h such t hat, for
everya
'
E C', i.p (M, ai, b) has size h and t he functions mapping any;' E C'
into
the first , ..., the h-th element of i.p (M, ai, b) respectively
are continuous. Let;' E C'. It is easy to see t hat ;' is good for i.p (v, W1 , . . . ,
W
n-1
, b).
Then (;' , b) E Cl, and Y n Cl =1= 0. Hence Y Cl . Y < C
2
is shown in a
similar way.
In conclusion, if Y is an open set of P, then for ever y (ii, b) E Y, d is good
for i.p (v, W l , . . . , Wn-l , b), b is good for i.p (v, ii , w
n
). So it suffices to show
what follows.
9.6. DEFINABLE GROUPS IN O-MINIMAL STRUCTURES 339
(ix) IfY is an open cell of M
n
, <p(v, WI, . , w
n
) is an L(M)-formulafinite
in Y and , for every (a, b) E Y, a isgood for <p(v, WI, , Wn- l , b) an d
b is good for <p(v, a, W
n),
t hen t he size of <p(M , C) is constant when C
ranges over Y.
Otherwise t here is a nat ural h such t hat t he set Y' of t he elements C E Y
satisfying 1<p(M, C) I = h is differ ent from Y and 0. But Y' is definable and
so, as Y is definably connected, Y cont ains some boundary point cof Y': any
open box B Y includi ng coverlaps both Y' and it s complement Y - Y' in
Y. Hence it is sufficient t o prove (ix) when Y just coincides with some open
box B. Let (aI, bd, (a2, b
2
) E B = Y. Not ice that every a in t he projection
B' of B onto the first n - 1 coordinates is good for <p (v, WI, .. , Wn-l, b
l
) :
in fact (a, bd E Y. Hence (4) n-l ensures
Similarly (4)1 implies
In conclusion
1<p( A1, aI , bdl = 1<p(M , a2, b
2
)1,
and t his accomplishes t he proof of (3)n'
The last claim t o be examined is (4)n' But now t he proof is a direct conse-
qu ence of what we have just observed. In fact , recall t hat , if Y is an open
box of M n and t he points (a, b) of Y are good for <p (v, Wl, . , w
n
), then,
for every (a, b) E Y, a is good for <p (v, WI , , Wn- l , b) and b is good for
<p (v, a, w
n
). Use this and (ix) and deduce (4)n'
Hence t he proof of Lemma 9. 5.4 is concluded, and at last we can also end
this section. ..
9.6 Definable groups in a-minimal structures
Which st ruct ures are definable in o-minimal models? The aim of this sec-
t ion is just t o measure how complicated o-mini mal struct ures are up t o
biint erpret a bilit y, and so t o answer t he previous question, and to reali ze
which groups, or rings, or manifolds are definabl e in t hem. In par t icul ar t he
int er est in definable manifolds arises qui t e naturally from t he connection
bet ween o-minimality and (analytic) geometry und erli ned at t he beginning
340 CHAPTER 9. O-MINIMALlTY
of this chapt er. We saw what a manifold is (inside an algebraically closed
field K) in Section 8.5 , where we observed that every manifold in K is de-
finable in K. In an ar bitrary o-rnin imal st ruct ure M a manifold may not
be definable. Accordingly, first we fix what definable manifold means. It is
just a finite family (V, (Vi, where m is a natural number and
* V = Vi is t he at las,
* each f i is a bijection from Vi onto a definable open subset U, = f i(Vi) of
M" for some natural n independent of i ,
* for i, j m and i =I- i , Ui,j = f i( Vi n\!j) is, again, definable and open
in Ui;
* for i, j m and i =I- i , f i, j = f i i;' is a definable homeomorphism
between Ui, j and Uj,i'
After fixing this definition, let us look for groups and manifolds definable
in o-rninimal st ructur es. Even at a first superficial sight one can meet some
non-trivial examples: for instance, it is quite obvious t hat , for a real closed
field K, the linear groups GL(n, K) are definable in K. Indeed, a sharp
analysis displays some notable similarities with the w-stable framework. In
particul ar, by ad apting the Hru shovski-Weil Theorem 8.6.2, Pillay showed
Theorem 9.6.1 Let M be an o-minimal st ructure, and g be a group de-
finable in M. Then g can be equipped with a (uniqu e) definable manifold
structure making it into a topological group.
When M expands th e real field , t he manifold t opology makes g into a locally
Euclidean t opological group and in conclusion, owing to the Montgomery,
Zippen and Gleason solut ion of Hilber t 's Fifth Problem, into a Lie group.
Definable groups have been intensively studied in o-rninimal st ructures. In
particular we would like to mention an o-rninimal analogue of Cherlin's
Conjecture, proved by Peterzil , Pillay and Starchenko.
Theorem 9.6.2 Let M be an o-minimal structure, g be a connected group
definabl e in M and having no definable non-trivial normal abelian subgroup.
Then there is a definable isomorphism of g onto the connected component
of an algebraic group over a real closed field.
Not ably, a local version of Zilber 's Conjecture is t rue in t he o-rninimal set-
t ing, as shown by Peter zil and Starchenko. Let us discuss briefly t his matter.
9.7. O-MINIMALITY AND REAL ANALYSIS 341
Given an o-rninimal M, call an element a E M t rivial when there are no
open interval/including a and no definable f from / 2 into I which is st rictly
monotone in each variable.
T heore m 9.6.3 (Tri chotomy Theorem) Let M be an 'No-saturated o-minimal
structure, a E M . Then exactly one of the following conditi ons holds:
(i) a is tri vial,
(ii) there is some convex neighbourhood of a where M induces a structure
of an ordered vect orspace over an ordered divi sion ring,
(iii) there is some open interval including a where M induces a structure
of a real closed field .
9.7 0-minimality and Real Analysis
In t his section, we introduce some new examples of o-minimal st r uctures.
They concern some expansions of t he real field closely related t o Real
An alysis and Geometry. Indeed Model Theory meets t hese areas within
o-rninimali t y, and provides new ideas , new t ools and, definit ively, new per-
spectives in studying t he involved st ructures .
1. R
ex p
The first example we wish to deal with is the most famous as well. It
concerns the exponent iat ion in t he real field . We have seen in Chap-
t er 2 Tarski 's Theorem showing that the t heory of th e real field R
has the elimination of quantifiers in the language L of ordered fields:
accordingly
definable = semialgebraic
in this set t ing. Tarski also gave an effective procedure reducing any
formula <p (v) di L into an equivalent quan tifier free L-formula <p'( V).
When applying this reduction method to a sentence <p of L , we get
explicit ly a quantifier free sentence <p' of L equivalent t o <p in RCF; <p'
is a finite Boolean combination of sentences n m where n and m are
int egers; accordingly it is easy to check whether <p' (a nd so <p) is t r ue
in R or not. In conclusion, t he t heory of R -in other words RCF- is
decidable.
Tarski also proposed t he following question. Expan d R to a st ruct ure
R
ex p
= (R, 0, 1, +, " - , exp)
where exp is the l-ary function mapping any real x into eX. Accord-
ingly add a l-ary operation symbol (for exp) t o L and denote by L
ex p
t he enlarged language: hence
L
ex p
= L U {exp} = {O, 1, +, " - , ex p}.
Conjecture 9 .7.1 (Tarski) The theory of R
ex p
is decidabl e.
One can observe t hat the th eor y of R
ex p
is not quant ifier elimina ble
in L
ex p
: this was shown by Van den Dries in 1982. However in 1991
Wilkie proved it s model completeness.
Theorem 9.7.2 (Wilkie) The theory of R
ex p
is model complet e.
In particul ar , the definable sets in R
ex p
ca n be obt ained as follows.
For n any positive int eger , call a subset E of R" exponential when it
has t he form
for a sui t able real polynomial f wit h 2n unknowns. Notice t hat ex-
ponenti al sets are closed und er finite union and intersect ion (as the
points annihilating at least one of finit ely many polynomials are just
the zeros of their product , and the point s annihilating a finit e system
of real polynomials are just the zeros of the sum of their squa res). But ,
according to the Van den Dries remark on (the failure of) quantifier
eliminat ion, the definable sets in R
ex p
are something larger than the
finit e Boolean combinat ions of exponent ial sets. So let us introduce
subexponential sets. A subexponent ial set in R" is just t he image
of an exponential set of R
n
+
m
(for some m) under t he proj ection
map of Rn+m onto the first n coordinat es in R" . Clearly exponential
and subexponent ial sets are definable. What Wilkie showed is that
subexponent ial set s are closed under complement. This implies model
complete ness, and proves that in R
ex p
definable = subexponential.
9.7. O-MINIMALJTY AND REAL ANALYSIS 343
At this point, one can use a theorem of Khovanskii saying that ev-
ery exponential set, and consequently every subexponent ial set , has
onl y finit ely many connected components. Just apply this result to
definable (equivalently subexponential) subsets of R and get
Corollary 9.7 .3 R
exp
(and its theory) are o-minimal.
Later Ressayre gave a nice axiomatization of T h( R
exp),
showing t hat
it s mode l theory requires very simple global infor mat ion about expo-
nentiation. But now let us come back to Tarski's Conject ure. What
can we say about it ? Well, there is a famous conjecture in transcen-
dental number theory, due to Schanu el and saying:
Conjecture 9.7.4 (Schanuel) Let n be a positive integer', al, .. . , an
be complex numbers linearl y independent over the rational fi eld Q.
Then the transcendence degree of Q(al ' .. . , an, e
a j
, , e
a n
) over Q
is at least n.
Remarks 9.7 .5 (a) Schanuel's Conjecture has been proved in some
particular cases, for example when n = 1, or al , . .. , an are alge-
braic (Lindemann) .
(b) 1, e are linearly independent over Q. Hence Schanuel's Conjec-
turewould imply t hat Q(l , e, el, ee), in other words Q (e, ee),has
transcendence degree 2 over Q, and hence e, e" are algebraically
independent. Nevertheless, as far as one presently knows, it is
still an open question whet her e" is irrational.
(c) 1, in are linearly independent over Q. Hence Schanuel's Con-
ject ure would imply that Q(l , in, el , e
i
7l" ) , hence Q (e, i1r) (as
e
i
7l" + 1 = 0), has transcendence degree 2 over Q, and conse-
que nt ly that e and n are algebraically independent : but this is
still an open question, as well known.
It is generally felt that a solution of Scha nuel's Conjecture is vary far ,
and should go beyond the present knowledge in Mathematics. However
a positive answer to the question of Schanuel would imply a solution
of Tarski 's Conjecture as well.
Theorem 9.7.6 (Macintyre-Wilkie) If Schanuel's Conjecture holds,
then Th( R
exp)
is decidable.
2. R an
Now we deal with real analytic functions f. Her e we have t o be very
careful in fixing our setting. In fact , we have t o recall what happens
when we expand the reals by sin (or cos ): o-minimality gets lost. How-
ever one sees t hat o-rninimalit y is preserved if we restrict t he domain
of t he sinus function t o a suitable interval] - Accordingly one
takes a language L
an
enlarging the language L of ord ered fields by a
l-ar y operation symbol j for every function f analytic on some open
subset U of R" containing th e cube [0, l]n (n ranges, as usual , over
positive integers, and th e only reason to choose [0, 1] instead of an-
other interval is just to fi x and normalize our set t ing); th en one t akes
the Lan-structure R an expanding the real fi eld R and int erpreting any
symbol j in t he funct ion equalling f in [0, I]" and ass uming th e con-
stant value 0 elsewhere. Notice that L
an
is uncoun t abl e. R an is called
t he real field with restri cted analytic functions.
Theorem 9.7.7 (Van den Dri es) Th(R
an
) is model complete and
o-minimal.
Van den Dri es' analysis also det ermines what is definabl e in R an. In
fact , t he definabl e sets are exactly t he so called globally subanalytic
sets. They are obt ained as follows. Call a subset A of an analyt ic
manifold X semianaly tic in X if t here is an open covering U of X such
t hat , for ever y U E U, A n U is a finite union of sets
{a E U : f (a) = 0, 9o(a), ... , 9k(a) > O}
where f and t he 9 'S ar e analyt ic functions on U. At t his point call a
subset B of X subanalytic in X if there is an open covering U of X
such that, for every U E U, B nU is th e image of some A semianalyt ic
in U X R'" by t he projection map from U X R'" onto U (here m may
depend on Band U). Finally call S R" globally subanaly tic if it
is subanalyt ic in th e analyt ic manifold (P
1
(R) )n (where P
1
(R) is t he
real projective line).
Gabrielov showed a "theorem of t he complement " for subanalyt ic sets
in an analyt ic manifold X, ensuring th at t hey are just closed und er
complement . This is t he key resul t in showing t he model completeness
and t he o-minimalit y of the t heory of R an' and also in proving that
definabl e = globally subanalytic
9.7. O-MINIMALITY AND REAL ANALYSIS 345
in R an.
Does Th( R
an)
admit quantifier elimination? Denef and Van den Dri es
showed t hat the answer is positive, provided one extends the language
L
an
by a symbol -I for the inverse funct ion (with the usual convention
0-
1
= 0).
Theorem 9.7.8 (Denef-Van den Dries) Th( R
an)
eliminates the quan-
tifiers in t.. ; U {-I}.
3. R an, exp
Finally let us examine what happens when we expand t he reals bot h
by t he exponent ial funct ion and t he rest rict ed analytic funct ions . Let
L
an
, exp = L
an
U { exp} the corresponding language, R an, exp t he re-
sult ing struct ure in Lan, exp. First Van den Dri es and Miller, adapting
Wilkie's wor k on exponentiation, pr oved
Theorem 9.7.9 (Van den Dri es-Miller ) Th e theory oj R an, exp is model
complete and o-minimal.
Subsequently, Van den Dries, Macintyre and Marker found a differ-
ent proof providing a nice axiomatization of the theory of R an, exp in
Ressayre's style. They got also quantifier elimination in a language
extending L
an
, exp by the logari thm function log .
Theorem 9.7.10 (Van den Dries-Macint yre-Marker ) Th e theory o]
R an, exp eliminates the quantifiers in the language L
an
, exp U {log} .
Notably, t he logari t hm funct ion cannot be ignored to obtain quantifier
eliminat ion. The Van den Dri es-Macint yr e-Marker approac h also pro-
vides an explicit descript ion of t he definabl e sets in R an,ex p , following
t he same lines as in t he cases before.
Of course t hese examples are very far from exhausting a general displ ay of
t he a-minimal expansions of the real field (a wider information can be found
in the references quoted at the end of t he chapter) . But t hey can illust rat e
how rich and int eresti ng t his research field is.
Let us conclude t his section with some final remarks partly exceeding the
a-minimal limits. In fact, it is notewort hy that , although Model Theory
and Real Analysis closely interact via o-minimality, Complex Analysis has
raised a lot of difficulties to a model t heoret ic t reat ment . For inst ance,
whi le expanding the reals by exponentiation gives an o-mi nimal structure
(by Wi lkie's Th eorem), (C, +, " - , 0, 1, ex p) defines the integers by the
formula
exp(27riv) = 1,
so t hat there is a very lit tl e hope to dominat e its first order t heory, it s
definable sets, and so on. Indeed, the zero sets of complex analytic functions
can be quite pat hological.
Some years ago , Boris Zilber proposed a satisfactory st rategy to develop th e
mod el theory of the complex exponentiation, but his program needs some
very strong conjectures on t ranscendental numbers (even beyond Schanuel's
P roblem). Zilber also followed a more successful approach, looking at ana-
lytic compact manifolds X rather than at analytic functions: in fact , t hese
manifolds can be viewed as first order structures in a language with a re-
lat ion for any subanalytic subset of every power of X. In this setting one
shows
Theorem 9.7.11 (Zilber) The theory of a compact complex manifold elim-
inates the quantifiers and has a finite Morley rank.
9.8 Variants on the a-minimal theme
Strongly minimal theories have a natural enlargement to totally transcen-
dent al (i. e. w-stable) theories via Morley ra nk. In the ordered setting
nothing is known ext ending sistematically o-minimality in a parallel way.
However just the ordered framework suggests several notions widening 0 -
minimality : they have been intensively studied in th e latest years.
In pa rticular we want to discuss here briefly weak o-minimality. As said, we
still work wit hin linearly ordered structures M = (M, ::; , ...). Recall that
M is o-minimal when every definable subset D of M is a finite union of
interval s (possibly with infinite endoints); not ice t hat intervals ar e convex.
Definition 9.8 .1 M is called weakly a-minimal when every definable
subs et D of M is a finit e union of convex (definable) sets.
Remark 9.8.2 Of course, o-rninimality implies weak o-minimality. More-
over , among expansions of the real line (R, ::;), the converse is also true, and
weakly o-minimal just means o-minimal. This is because (R, ::;) is Dedekind
complete, and every bounded set has its own least upper bound and its own
9.9. NO ROSE WI THOUT THORNS 347
greatest under bound; in particular every convex set is an int erval (in t he
broader sense recalled before).
However t here do exist weakly a-minimal structures which are not a-mini mal.
Example 9.8.3 Take t he ordered field of real algebraic numbers Ra. This
is a real closed field, and so an a-minimal st ruct ure. Add a 1-ar y relati on
select ing t he elements of Ra lying between - IT and IT (or, if you like, bet ween
any two reals a < b wit h a or b t ranscendental) . The resulting st ruct ure is
not a-minimal any more, because D = {r E Ra : -IT < r < IT} is convex
and definable, but cannot be expressed as a finit e union of int ervals with
real algebraic endpoint s. But act ually D is convex, and indeed one can see
that t he new structure is weakly a-minimal.
Notably, every expa nsion of an a- minimal struct ure by convex subsets is
weakly o-minimal. This is a beau tiful result of Baisalov-Poizat , generalizing
t he last example and answering in t his way a question of Cherlin. Oth er
relevant examples, ar ising from several frameworks in Algebra, can be pro-
posed. By t he way, weak o-rninimalit y was first introduced by M. Dickmann
in 1985, dealing wit h certain ord ered rings extending real closed fields.
Not surprisingl y, weakl y a-minimal st ruct ures do not behave so well as 0 -
minimal do. In particular
weak o-minimalit y is not preserved by elementary equivalence
(so t here are weakly a-minimal struct ures whose t heory has some non weakly
a-minimal models) . Furt hermore
Monotonicity and Cell Decomposition fail
as well as existence and uniqueness of prime models. However some " weaker"
versions of these results ca n be recovered, and a relevant, alt hough not so
fluent , th eory has been developed.
9.9 No rose without thorns
We have seen t hat a- minimal t heories admit an independence noti on re-
lated to algebraic closure and satisfying t he same basic proper ti es (11)-(16)
forkin g independence has in simple t heor ies. However t hese independence
notions -forking indipendence in simple t heories and algebraic independence
in a-minimal th eories- were introduced in a different way and were devel-
oped independently. So a natural question arises in Model Theory, Le. to
find a new conce pt of independence so convincing t o satisfy (11)-(16) in
most t heories and so general t o enlarge both t he pr evious cases. This is the
cont ent of a recent work of Alf Onshuus who, following suggest ions from
Thomas Sca nlon, int rodu ced
a new noti on of independence (called thorn-independence)
and
a related class of t heor ies (named rosy theories)
whe re t horn-i nde pende nce enjoys all t he basic ass umptions (11 )-(16) , so lo-
cal character, symmet ry and so on. Rosy t heories include both si mple and 0 -
minimal t heories, as well as fur ther relevant examples. Thorn-independence
agrees wit h algebraic independence in t he o-minimal case and with forking
ind ependence in stable t heories: in fact , when t hese pages are written (at
t he end of 2002) , it is not clear whether t horn- independence equals forking-
independence even in the simple setting, although t his has been checked to
be t rue in all t he known key examples of simple t heor ies.
Notably symmetry, or also local character, is a key property towards rosi-
ness. In fact , a t heory T is rosy if and onl y if t horn- inde pende nce satisfies
symmet ry or local character .
9.10 References
O- minimal t heories where introduced by Van den Dries [166] and extensively
st udied by Pill ay and Steinhorn in [129] and (together with J. Knight ) in
[74]. Van den Dries' book [169] provides a nice and stimulating treatment
of o-rnini mali t y, also describing it s genesis and moti vat ions, and emphasiz-
ing it s con nect ions wit h real analysis and real algebraic geometry. T hese
int er acti ons ar e illust rat ed in t he mor e recent survey [170], where t he 0 -
minimal expa nsions of t he real field are examined. Also [109] gives a short,
but captivating introd uction to o-rninimalit y,
A general proof of Monotonicity Theorem 9.2.1 can be found in [129]. Wilki e's
Complement Theorem 9.4 .5 is shown in [178], and t he Pill ay-St einhorn The-
orem 9.4.7 on prime models in o-rnin imal t heories is in [129] again . Laura
Mayer 's sol ution of Vaught Conjecture in t he o-rninimal setting is in [111].
Pill ay's analysis of t he groups definable in o-minimal struct ures (Theo-
rem 9.6.1) is in [125], while t he o-rninimal analogue of Cherlin Conjecture
shown by Pet erzil , Pill ay and Starchenko (9.6.2) is in [123] and the Pet erzil-
Starchenko Theorem 9.6.3 is in [124].
9.10. REFERENCES 349
As already said [170] provides a general sur vey of the main o-minimal expan-
sions of the real field, and a rich and detailed bibliography on this matter. In
particular Wilkie's Theorem 9.7.2 on R
ex p
is in [177], and the Khovanskii's
results on exponential sets in [70]; Ressayre's approach to the theory of R
ex p
can be found in [139]; the Macintyre-Wilki e Theorem on the decidability of
the theory of R
ex p
and its relationship to Schanuel's Conjecture is in [99].
The o-minimality of the theory of Ran is proved in [167], using [49], while
the Denef- Van den Dries treatment -including the quantifier elimination re-
sult in a language with the inverse operation- is in [30]. The o-minimality of
the theory of Ran, exp is already shown in [171], but the subsequent analysis
of Macintyre, Marker and Van den Dries is in [97].
Zilber 's Theorem 9.7.11 on compact complex manifolds can be found in
[183]; see also [128] .
Now let us deal with weak o-minimality. This was introduced in [32], and
extensively examined by Macpherson, Marker and Steinhorn in [100]. The
nice theorem of Baisalov-Poizat (mentioned at the end of Section 9.8) is in
[7].
Rosy theories and thorn-independence are just the matter of [122].
Bibliography
[1] J . W. Addison, L. Henkin, and A. Tarski, editors. The theory of models.
Nort h Holland, Amsterdam, 1965.
[2] F. Appenzeller . Classification Theory through St at ionar y Logic. Ann. Pure
Appl. Logic, 102:27-68, (2000) .
[3] J . Ax. The elementary theory of fi nite fields. Ann. Math . , 88:239-271 , (1968) .
[4] J . Ax and S. Kochen. Diophanti ne probl ems over local fields I. Amer. J.
Math. , 87:605-630, (1965) .
[5] J . Ax and S. Kochen. Diophantine problems over local fields 11: A complete
set of axioms for p-adic number theory. Amer. J. Math . , 87:631-648, (1965) .
[6] T. Ax and S. Kochen. Diophantine pr oblems over local fields Ill: Decidabl e
fields . Ann. Math . , 83:437-456, (1966).
[7] Y. Baisalov and B. Poizat . Paires de structures o-minirnales. J. Symbolic
Logic, 63:570-578, (1998) .
[8] .1. Baldwin. Fundamentals of Stability Theory. Spr inger Verl ag, Berlin , 1988.
[9] J . Baldwin and A. Lachlan. On strongly minimal sets. 1. Symbolic Logic,
36:79-96, (1971) .
[10] J. Barwise, edit or . Handbook of Mathematical Logic. Nort h Holland , Ams-
terd am, 1977.
[11] .1. Barwise and S. Feferman, editors. Model-theoretic logics. Springer, New
York, 1985.
[12] L. Blum. Generalized Algebraic Theories. PhD thesis, M.I.T., Cambri dge,
1968.
[13] L. Blum, F. Cucker , M. Shub, and S. Smale. Complexit y and Real computa-
tion. Springer, New York, 1998.
[14] L. Blum, M. Shub, and S. Smale. On a t heory of comput at ion and complexity
over t he real numbers. Bull. Amer. Moth . Soc. , 21:1-46, (1989) .
351
352 BIBLIOGRAPHY
[15] A. Borovik and A. Nesin . Groups of finit e Morl ey rank. Oxford University
Press, Oxford, 1994.
[16] E. Bous caren, edit or . Model Theory and Algebraic Geomet ry, volume 1696 of
Lecture Notes in Math ematics. Springer , Berlin , 1998.
[17] A. B. Ca rson. The model complet ion of the th eory of commut ative regul ar
rings. J. Algebra, 27:136-146, (1973).
[18] C. Chang and H. Keisler. Model Theory. North Holland, Amst erdam, third
edit ion, 1990.
[19] Z. Chat zidakis. Theori e des rnodeles des corps finis et pseudo-finis.
Prepublications 59, Equipe de Logique Mathematique, Universit e Paris VII ,
(1996) .
[20] Z. Chat zidakis and E. Hrushovski . The model th eory of difference fields .
Trans . Amer. Math . Soc. , 351:2997-3071, (1999) .
[21] Z. Chat zidakis, E. Hrushovski , and Y. Pet erzil. Model t heory of difference
fields , 11: Peri odi c ideals and th e trichotomy in all characterist ics. Preprint .
[22] G. Cherlin. Algebraically closed commut ative rings. J. Symbolic Logic,
38:493- 499, (1973) .
[23] G. Cherlin. Model Theoreti c Algebra, volume 521 of Lecture Not es in Math e-
mat ics. Springer Verlag, Berlin , 1976.
[24] G. Cherlin. Groups of small Morl ey rank. Ann. Math. Logic, 17:1-28, (1979).
[25] G. Cherlin and S. Shelah. Superstable fields and rings. Ann. Math . Logic,
18:227-270, (1980).
[26] P. Clote and J . Kr aji cek, edit ors. Arithmeti c, Proof Theory and Computa-
tional Complexity. Oxford Universit y Press, Oxford , 1993.
[27] P. Cohen. Decision procedures for real and p-adic fields . Comm. Pure Appl.
Math ., 22:131-151, (1969).
[28] R. M. Cohn. Difference algebra, volume 17 of Tracts in Mathematics. Inter-
science , New York , 1965.
[29] G. E. Collins. Quantifier eliminat ion for real closed fields by cylindrical alge-
br aic decomposition. In Automata Theory and Formal Language, volume 33
of Lecture Notes in Computer Science, pages 134-1 83, Berlin, 1975. Second
GI Conference, Kai serslautern, Springer .
[30] J. Denef and 1. Van den Dri es. p-adi c and real subanalyti c sets . Ann. Math .,
128:79-138, (1988).
[31] K. Devlin . The j oy of sets. Undergraduate Text s in Math . Springer Verlag,
Berlin, 1993.
BIBLIOGRAPHY 353
[32] M. Dickmann. Elimination of quantifiers for ord ered val uat ion rings . In Pro-
ceedings 3rd East er Model Theory Confer ence, Humb oldt Unive rsitiit , pages
64-88, 1985.
[33] A. Dolzmann , T . Sturm, and V. Weispfenning. Real quant ifier eliminat ion in
practi ce. In Algorithmic Algebra and number theory, pages 221-247. Springer ,
Berli n, 1999.
[34] J. Doner and W. Hodges. Alfred Tarski and decidable th eori es. J. Symbolic
Logic, 53:20-35, (1988) .
[35] J. 1. Duret . Les corps pseudo-finis ont la propriet e d'i ndependence. C. R.
Acad. Sci . , 290:981-983, (1980) .
[36] H. D. Ebbinghaus. Nu mbers, volume 123 of Graduate Texts in Math. Springer
Verl ag, New York , 1990.
[37] H. D. Ebbingha us, J. Flum , and W. Thomas. Mathema ti cal Logic. Under-
grad uate Text s in Mat h. Spri nger Verlag, Berlin, 1979.
[38] A. Ehrenfeucht and A. Mostowski. Models of axiomatic th eori es admitting
automorphisms . Fund. Math. , 43:50-68, (1956) .
[39] P. Eklof. Ult raproducts for Algebraist s, pages 105- 137. In Barwise [10], 1977.
[40] P. Eklof and E. Fisher. The element ary t heory of abelian groups . Ann. Math.
Logic, 4:115- 171, (1972).
[41] P. Eklof and G. Sabbagh. Model complet ions and modules. Ann. Math . Logic,
2:251-295, (1970-71).
[42] H. Endert on. Elements of recursion theory, pages 527-566. In Barwise [10],
1977.
[43] E. Engeler. A charact erization of th eories with isomorphic denumerable mod-
els. Notices Amer. Math. Soc., 6:161, (1959).
[44] Vu. L. Ershov. On th e element ary t heory of maximal normed fields . Algebra
i Logika, 4:31- 70, (1965).
[45] Vu. L. Ershov. On t he elementary t heory ofmaximalnormed fields 11. Algebra
i Logika, 5:5-40, (1966).
[46] Vu. L. Ershov. Fields wit h a solvable t heory. Soviet Math. Dokl. , 8:575-576,
(1967) .
[47] M. Fisher and M. Rabin. Super exponent ial complexit y of Presburger 's arit h-
meti c. SIAM-AMS Proc. , 7:27- 41, (1974).
[48] J. Flum. First order Logic and its extensions. Number 499 in Lecture Notes
in Math ematics. Springer Verl ag, Berlin, 1975.
[49] A. Gabrielov. Proj ections of semianalytic sets . Functional Analysis and its
Applicat ions, 2:282-291, (1968).
354 BIBLIOGRAPHY
[50] J . B. Goode. H.L.M. (Hrushovski-Lang-Mordell) . Technical Report 48me
annee, Serninaire Bourbaki, 1995-96.
[51] R. C. Gunning and H. Rossi . Analytic functions of several complex variables.
Princeton Hall Inc., 1965.
[52] P. Halmos. Lectures on Boolean Algebras. Van Nostrand, Princeton, 1963.
[53] B. Hart , E. Hrushovski , and M. Laskowski . The uncountable spectrum of
countable complete theories. Ann. Math . , 152:207-257, (2000) .
[54] D. Haskell , A. Pillay, and C. St einhorn, editors. Model Theory, Algebra and
Geometry. MSRI Publ. , Cambridge UP, 2000.
[55] L. Henkin. The completeness of t he first order functional calculus. J. Symbolic
Logic, 14:159-166, (1949) .
[56] W. Hodges. Model Theory. Cambridge University Press, Cambridge, 1993.
[57] W. Hodges. A shorter model theory. Cambridge Universit y Press, Cambridge,
1997.
[58] E. Hrushovski . The elementary theory of the Frobenius. Preprint.
[59] E. Hrushovski . A new strongly minimal set . Ann. Pure Appl. Logic, 62:147-
166, (1993) .
[60] E. Hrushovski . The Mordell-Lang Conjecture for funct ion fields. J. Amer.
Math . Soc., 9:667-690, (1996) .
[61] E. Hrushovski. The Manin - Mumford conjecture and the model theory of
difference fields. Ann. Pure Appl. Logic, 112:43-115, (2001) .
[62] E. Hrushovski and Z. Sokolovic. Minimal subsets of differentially closed fields .
(to appear) .
[63] E. Hrushovski and B. Zilber . Zaris ki 's geometries. Bull. Amer. Math. Soc.,
28:315-324, (1993) .
[64] E. Hrushovski and B. Zilber. Zariski 's geometries. J . Amer. Moth. Soc. ,
9:1-56, (1996).
[65] N. .Jacobson . Basic Algebra. Freeman, San Francisco, 1974-1980.
[66] T . J ech. Set Theory. Academic Press, New York - London, 1978.
[67] I. Kaplansky. Differential Algebra. Hermann, 1957.
[68] A. Kechris . New dir ections in Descriptive Set Theory. Bull. Symbolic Logic,
5:161-174, (1999) .
[69] J. Ketonen. The structure of countable Bool ean algebras. Ann. Math ., 108:41-
89, (1978) .
[70] A. Khovanskii . On a class of systems of transcendental equations. Sovie t
Math . Doklady, 22:762- 765, (1980) .
BIBLIOGRAPHY 355
[71] B. Ki m. Forking in simple unst able t heories. 1. London Mat h. Soc. , 57:257-
267, (1998) .
[72] B. Kim. Simplicity and stabilit y in t here. J. Symboli c Logic, 66:822-836,
(2001).
[73] B. Kim and A. Pillay. Simple t heories. Ann. Pure Appl. Logic, 88:149- 164,
(1997) .
[74] J . Knight , A. Pi llay, and C. St einhorn. Definable sets in ordered struct ures
11. Trans. Amer. Math. Soc., 295:593- 605, (1986).
[75] R. Knight . The Va ught Conj ect ure: a count erexample. Techni cal repor t ,
Oxford University, 2002.
[76] E. Kolchin. Differential Algebra and Algebraic Groups. Academi c Press, 1973.
[77] T. Kucera and M. Prest o Imaginary modules. J. Symbolic Logic, 57:698- 723,
(1992) .
[78] K. Kun en. Set Theory. An introduction to independence proof s. North Hol-
land, Amsterdam, 1983.
[79] S. Lang. Algebra. Add ison-Wesley, Reading, 1971.
[80] S. Lang. Number Theory Ill: Diophant ine Geometry. Encycl opedia of Mat h-
emat ical Sciences. Springer , Berl in, 1991.
[81] C. H. Langford. Some t heorems on deducibility. Ann. Mat h. , 28:16-40, (1926) .
[82] C. H. Langford. Theorems on deducibility. Ann. Math. , 28:459-471 , (1926).
[83] D. Lascar. Ordre de Rudin-Keisler et 'poids dans les t heories w-st ables.
Zeits chr. Math. Logik, 28:413- 430, (1982) .
[84] D. Lascar . Why some people are excit ed by Va ught 's Conje ct ure. J. Symbolic
Logic, 35:973- 982, (1985) .
[85] D. Lascar and B. Poizat . An int oduct ion to forkin g. J. Symbolic Logic,
46:781- 788, (1981).
[86] L. Lipschit z and D. Sar acino. The model companion of t he t heory of commu-
tative rings without nilpot ent element s. Proc. Amer. Math . Soc. , 37:381-387,
(1973).
[87] L. Lowenheim. Uber Moglichkeit en in Relat iv Kalkiil. Moth. Ann. , 76:447-
470, (1915).
[88] A. Macintyre. Non-standard Frob enius and generic automorphisms . Preprint .
[89] A. Macintyre. On wt -categor ical t heories of abelian groups. Fund. Mat h. ,
70:253-270, (1971) .
[90] A. Macintyre. On wt -categorical t heories of fields. Fund. Math. , 71:1- 25,
(1971).
356 BIBLIOGRAPHY
[91] A. Macintyre. On algebraicall y closed groups. Ann. Math. , 95:53-97, (1972) .
[92] A. Macintyr e. Model complete ness for sheaves of structures. Fund. Math. ,
81:73- 89, (1973).
[93] A. Macintyre. On definable subsets of p-adic fields . J. Symbolic Logic, 41:605-
610, (1976).
[94] A. Macintyre. Model- compl et eness, pages 139- 180. In Barwise [10], 1977.
[95] A. Macintyre. Gen eri c automorphisms of fields . Ann. Pure Appl. Logic,
88:165- 180, (1997) .
[96] A. Macint yre, edit or. Connecti ons bet ween Model Theory and Algebraic and
Analytic Geometr y, volume 6 of Quaderni di Mat ematica. Dipartimento di
Mat emat ica II Universita di Napoli, 2000.
[97] A. Macintyre, D. Marker , and L. Van den Dries. The element ar y theory
of restrict ed analyt ic fields with exponentiation. Ann. Math . , 140:183- 205,
(1994) .
[98] A. Macintyre, K. McKenna, and L. Van den Dries. Elimination of quantifi ers
in algebraic st ruct ures . Adv. Math. , 47:74-87, (1983) .
[99] A. Macintyre and A. Wilkie. On t he decidability of the real exponential field .
In Kreiseliana (P. Odifreddi ed.), pages 441-467. A. K. Peters, Wellesley,
1996.
[100] D. Macph erson, D. Marker , and C. Steinhorn. Weakly o-rninimal st ructures
and real closed fields . Trans. Amer. Math. Soc. , 352:5435-5483, (2000).
[101] M. Makkai . A survey of basic stability t heory, with particular emphasis on
or thogonalityand regular typ es. Israel J. Math., 49:181- 238, (1984) .
[102] A. Malcev. A correspondence between groups and rings. In The metamath-
ematics of algebraic syst ems , pages 124-137. North Holl and, Amsterdam,
1971.
[103] J . Mali t z. Int roduction to Math ematical Logic. Undergraduat e Texts in Math.
Springer Verlag, Berlin, 1979.
[104] A. Marcja and P. Mangani . Shelah rank for Boolean algebras and some
applications to element ary t heories. Alg. Univ., 10:247- 257, (1980) .
[105] A. Marcja and C. Toffalori. On pseud o-No-cat egorical t heories. Zeilschr.
Math. Logik, 30:533- 540, (1984) .
[106] A. Mar cja and C. Toffalori. On t he Cantor-Bendixson Spect ra containing
(1,1)-1. In Logic Colloquium '83, pages 331-350, Berlin, 1984. Spri nger .
[107] A. Marcja and C. Toffalori . On t he Cant or-Bendixson Spect ra containing
(l ,l)-II. J. Symbolic Logic, 50:611-618, (1985) .
BIBLIOGRAPHY 357
[108] A. Marcja and C. Toffalori . Introduzion e alta Teoria dei Modelti, volume 43
of Quaderni dell'Unione Matematica Italiana. Pitagora Editrice, Bologna,
1998.
[109] D. Marker . Model theory and exponent iat ion. Notices Amer. Math . Soc.,
43:753-759, (1996) .
[110] D. Marker , M. Messmer , and A. Pillay. Model Th eory of fields . Lecture Notes
in Logic. Springer, Berlin, 1996.
[111] L. Mayer. Vaught 's conjecture for o-minimal theories. J. Symbolic Logic,
53:146-159, (1988) .
[112] K. Meer. A note on a P =F N P result for a restricted class of real machines.
J. Complexity, 8:451-453, (1992).
[113] A. Mekler. Stability of nilpotent groups of class 2 and prime exponent . J.
Symbolic Logic, 46:781-788, (1981) .
[114] G. Melles. An exposition of Shelah's classification theory and classification
by set recursive functions (More for Thomas the doubter) . Preprint.
[115] M. Messmer. Groups and fields interpretable in separably closed fields. Trans.
Amer. Math . Soc., 344:361-377, (1994).
[116] M. Morl ey. On theories categorical in uncountable powers. Proc. Nat. Acad.
Se. U.S.A., 49:213-216, (1963) .
[117] M. Morley. Categoricity in power. Trans . Amer. Math . Soc., 114:514-538,
(1965).
[118] M. Morley. Omitting classes of elements, pages 265-273. In Addison et al.
[1], 1965.
[119] M. Morley. Countable models of theories. Israel J. Math.,
5:65-72, (1967) .
[120] M. Morley. The number of countable models. J. Symbolic Logic, 35:14-18,
(1970).
[121] P. Odifreddi . Classical recursion theory. North Holland, Amsterdam, 1989.
[122] A. Onshuus. Thorn-independence in rosy theories. PhD thesis, Berkeley,
2002.
[123] Y. Peterzil , A. Pillay, and S. Starchenko. Definably simple groups in 0-
minimal structures. Trans. Amer. Math . Soc., 352:4397-4419, (2000).
[124] Y. Pet erzil and S. Starchenko. A Trichotomy Theorem for o-minirnal struc-
tures. Proc. London Moth. Soc. , 77:481-253, (1998).
[125] A. Pillay. On groups and fields definable in o-rninimal structures. J. Pure
Appl. Algebra , 53:239-255, (1988) .
358 BIBLIOGRAPHY
[126] A. Pillay. Geometric Stability Theory. Oxford University Press, Oxford, 1996.
[127] A. Pillay. Model Theory and Diophantine Geometry. Bull. Amer. Math. Soc.,
34:405-422, (1997) .
[128] A. Pillay. Some model th eory of compact complex spaces. Contemp. Math .,
270:323-328, (2000).
[129] A. Pillay and C. St einhorn . Definabl e sets in ordered st ruct ures I. Trans.
Amer. Math. Soc. , 295:565- 592, (1986) .
[130] B. Poizat. Une theori e de Galois imaginaire. J. Symbolic Logic, 48:1151- 1170,
(1983) .
[131] B. Poizat . Cours de Theori e des Modeles. Nur Al - Mantiq Wal - Ma ' rifah ,
Villeurbanne, 1985.
[132] B. Poizat . Groupes Stables. Nur Al - Mantiq Wal - Ma 'rifah , Villeurbanne,
1987.
[133] B. Poizat . Les peti ts cailloux, volume 3. Nur Al - Mantiq Wal - Ma'rifah ,
Aleas , Lyon , 1995.
[134] B. Poizat. Course in Model Theory: An introduction in Contemporary Math-
ematical Logic. Springer Verl ag, 1999.
[135] B. Poizat. Stabl e groups. Ameri can Mathemati cal Society, 2001.
[136] M. Prest o Model Theory and Modul es. London Math. Soc. Colloquium Pub-
licat ions. Cambridge Universit y Press, Cambridge, 1988.
[137] M. Pr est o Model th eory and represent ati on type of algebras . In Logic Collo-
quium '86, pages 219- 260, Amst erdam , 1988. Nort h Holland.
[138] M. Rabin. Decidable Theori es, pages 595-629. In Barwise [10], 1977.
[139] J. P. Ressayre. Int eger parts of real closed exponential fields, pages 278-288.
In Clote and Kr aji cek [26], 1993.
[140] A. Robinson. On the metamathematics of Algebra. North Holland, Amst er-
dam, 1951.
[141] A. Robinson. Complete theori es. Nort h Holland, Amsterdam, 1956.
[142] A. Robinson . On th e concept of differenti all y closed field. Bull. Res. Council
Israel, 8F:1l3- 128, (1959) .
[143] A. Robinson . Introduction to Model Theory and to the Metamathemati cs of
Algebra. North Holl and, Amst erd am , 1963.
[144] H. Rogers. Theory of recursive fun ct ions and effective computabilit y. MIT
Press, Cambridge, 1988.
[145] C. Ryll-Nardzewski . On the categori city in power x, No. Bull. Acad. Pol on.
Sciences, 7:545-548, (1959) .
BIBLIOGRAPHY 359
[146] G. E. Sacks. Saturated model theory. W. A. Benjarnin, Reading, 1972.
[147] T . Scanlon. Diophantine Geometry from Model Theory. Bull. Symbolic Logic,
7:37- 57, (2001) .
[148] S. Shelah. Uniqueness and characterizat ion of prime models over sets for to-
tally transcendental first ord er th eory. J. Symbolic Logic, 37:107-113, (1972) .
[149] S. Shelah. Classification theory and the number of non-isomorphic models.
North Holl and, Amsterdam, 1978.
[150] S. Shelah. Simple unstable theori es. Ann. Math . Logic, 19:177-203, (1980) .
[151] S. Shelah. Classifi cation theory. North Holland, Amsterdam, 1990.
[152] S. Shelah, L. Harrington, and M. Makkai . A proof of Vaught 's Conjecture
for w-st abl e theories. Israel J. Math. , 49:259-280, (1984) .
[153] J .R. Shoenfield . Math ematical logic. Addison-Wesley, Reading, 1967.
[154] T . Skolem. Untersuchungen iiber die Axiome des Klassen Kalkiils and iiber
Produktations-und Summationsprobleme, welche gewisse Klassen von Aus-
sagen betreffen. Skrifter Vindenskapsakademi et i Kristania, 3:37-71, (1919).
[155] Z. Sokolovic. Model theory of different ial fields . PhD th esis, Notre Dame,
1992.
[156] L. Svenonius. No-categoricity in first ord er predicate calculus. Theoria, 25:82-
94, (1959).
[157] A. Tarsk i. A decision method for elementary algebra and geometry. University
of California Press, Berkeley and Los Angeles, second edit ion, 1951.
[158] A. Tarski . Cont ributions to the Theory of Models 1. Indag. Math ., 16:572-581,
(1954) .
[159] A. Tarski . Contributions to t he Theory of Models 11. Indag. Math., 16:582-
588, (1951) .
[160] A. Tarski . Foundations of th e calculus of systems . In Logic, semantics, meta-
mathematics. Papers from 1923 to 1938, pages 342- 383. Clar endon Press,
Oxford, 1956.
[161] G. Terjanian . Un contre-example a une conjecture d'Artin. C. R. Acad. Se.
Paris Ser A, 262:612, (1966) .
[162] R. Thorn. La stabilite topologique des applications polynomiales.
L'Einseignement Mathemaiiques, 8:21-33, (1962) .
[163] S. Thomas. On the complexit y of the classification problem for torsion-free
abelian groups of finit e rank. Bull. Symbolic Logic, 7:329- 344, (2001) .
[164] C. Toffalori . Strutture esistenzialmente complet e per certe classi di anelli.
Rend. Sem. Mat. Unio. Padova , 66:51-71, (1981) .
360 BIBLIOGRAPHY
[165] C. Toffalori . On pseud o structures. In Logic Colloquium '84,
pages 303- 327, Amsterdam, 1986. North Holland.
[166] L. van den Dri es. Remarks on Tarski 's problem concern ing (R , +, ',exp) . In
Logic Colloquium '82, pages 97-121. Nor th Holland, 1984.
[167] L. van den Dries . A genera lization of the Tarki-Seidenb erg Theorem and
some nondefinability results . Bull. Amer. Math. Soc. , 15:189-193, (1986) .
[168] L. van den Dries. Alfred Tarski 's eliminat ion theory for real closed fields. J .
Symbolic Logi c, 53:7-19, (1988) .
[169] L. van den Dries. Tame Topology and o-minimal Structures, volume 248 of
London Math. Soc . Lecture Not e Series. Cambridge University Press, Cam-
bridge, 1998.
[170] L. van den Dri es. a-minimal St ructures and Real Analytic Geometry. In
Current Developments in Mathematics (B . Ma zur ed.) , pages 105-152. Int .
Press, Sommerville MA, 1999.
[171] L. van den Dries and C. Miller . On the real exponent ial field with restricted
analytic functions . Israel J . Math., 85:19-56, (1994) .
[172] L. van den Dries and K. Schmidt . Bounds in the polynomial rings over fields .
A non standard approach. Inu. Ma th. , 76:77-91 , (1984).
[173] B. L. van den Waerden. Moderne Algebra. Springer, Berlin, 1930-31.
[174] R. Vaught . Denumerable models of complet e theories . In Infi nitistic Methods,
pages 303-321. Pergamon Press, London, 1961.
[175] F. Wagner . Simple theories. Kluwer , Dordrecht , 2000.
[176] A. Weil. Foundations of algebrai c geometry, volume 29 of Arn er. Ma th. Soc.
Colloquium Pubblics. Amer . Math. Soc., New York , 1946.
[177] A. Wilkie. Mod el completeness results for expansions of the ordered field of
real numbers by rest rict ed Pfaffian functions and the exponenti al function.
J. Amer. Math. Soc. , 9:1051-1094, (1996) .
[178] A. Wilkie. A general theorem of th e complement and some new o-minimal
structures. Selecta Math.(N.S.) , 5:397-421, (1999) .
[179] C. Wood. The model t heory of differential fields of charact erist ic p i O. Proc.
Amer. Math. Soc. , 40:577-584, (1973).
[180] C. Wood. Prime model extensions for differential fields of characteristic p i o.
J. Symbolic Logi c, 39:469-477, (1974) .
[181] M. Ziegler. Model theory of modules. Ann. Pure Appl. Logi c, 26:149-213,
(1984) .
[182] B. Zilber . The structure of models of uncountably cat egorical theories. In
ICM- Varsavia 1983, pages 359-368. North Holland, 1984.
BIBLIOGRAPHY 361
[183] B. Zilber. Model Theory and Algebraic Geometry. In Proceedings 10th Easter
Conference, Berlin, 1993.
Index
adiacency relation, 280
algebraic geometry, 291
complex, 37
real , 38, 341
algebra ic numbers
complex, 87
real, 87
algebraically independent set , 170
algorithm, 39, 80
amalgamation, 231, 234
analysis
complex, 345
annihilator , 42
Artin's conj ect ure , 86, 96-103
Art in's theorem, 95
atl as , 299
at omi c set , 198
automorphism, 5
Ax 's theorem, 60
Ax-Kochen-Ershov t heorem, 100
back-and-forth property, 10
basis, 169
Baur-Monk theorem, 70, 74
biinterpretability, 279-286, 288, 290,
339
Boolean space, 136, 180
boundary point, 322
Cantor 's t heorem, 12
Cant or-Bendixson rank, 155, 159
cell , 320
cell decompositi on th eorem, 324, 329
charact er
fini t e, 228, 233
local , 228, 233
363
Cherlin's conject ur e, 303, 310
Chevall ey's theorem, 59
Church-Turing thesis, 39, 80
class
element ary, 22, 105, 221
of fields, 25, 103
of finite sets, 22
of graphs, 280
of infini t e sets, 22
of modules, 27
of nilpotent groups of class 2, 281
of ordered fields , 26
classifi cation problem, 20, 221-227, 289
closure
algebraic, 133, 172, 287, 328
definable, 133, 172
differential , 209
real, 175
238, 252, 290
compact ness, 18
theorem of, 18, 243, 325, 327
connected component , 189
constructible set , 38, 58, 118, 134, 198,
199, 220, 271, 292, 295
const ruc t ion, 199
coordinate chart , 299
cylindrical algebra ic decomposition, 80
definable manifold, 340
definable sets, 35-42, 45, 59, 78, 112,
121, 126, 129, 133-136, 222,
280, 320, 344, 345
Bool ean algebra of, 36, 134, 143,
219
convex, 346
definably connect ed, 322
364
ind ecomposable, 190
dependence, 227
dependence relation, 180, 184, 328
algebraic, 170
linear , 169
derivation, 109, 307
descriptive set th eory, 222, 289
difference degree, 289
differential algebra, 209
differential degree , 287
differentially algebraic, 210
differentially transcendental , 210
dimension, 294-297
of a definable set, 328
of a tuple, 328
of a vectorspace, 169
dimensi on theory, 327
effect ive procedure, 40, 44, 79
Ehrenfeucht-Mostowski model , 274, 290
element ary cha in t heorem, 18, 144
element ary equivalence, 324, 347
eliminat ion of imaginari es, 125, 194,
307
uniform, 125, 300
eliminat ion sets, 43
embedding, 4,9-18, 85, 88, 103
element ary, 13, 85, 103
existent ial , 14, 89, 105
endomorphism, 5
exist ence property, 270
exponent ial set , 342
ext ension, 5, 228, 234
eleme nt ary, 14
non forking, 237, 238, 240
field
w-stable, 220
algebraic closure of, 33, 86, 93,
103, 117
algebraically closed, 26, 32, 45,
80- 82, 86, 103,105, 117, 134,
160, 184, 230, 239, 255, 284,
287, 291, 294
INDEX
eliminat ion of imaginaries for ,
126
model completeness of, 91
quantifier eliminat ion for , 54-
61
with an automorphism, 117
complex
eliminat ion of quantifiers for ,
82
const ant subfi eld of, 110, 116
difference, 115
existent ially closed, 117
inversive, 116
differential , 109, 209
existent ially closed, 110, 115,
209
differential closure of, 111, 115
differenti ally closed, 106, 109-115,
119, 134,211,220,239,242,
290, 291, 304, 307
axioms for, 111
existe nt ially closed, 239,290, 307
with an automorphism, 291
fixed subfield of, 116
formall y real , 95
Henselian , 99
imperfecti on degree of, 114
locall y compact, 98
of complex numbers, 33
of meromorphic functi ons, 110
of ra ti onal funct ions, 109
ord ered, 129, 316, 327
o-minimal ,317
real closure of, 180
p-adic, 86
perfect , 113
pseudofinite, 26, 102, 118, 288,
290
real , 321
82
real closed, 43, 78- 82, 87, 95,104,
105, 129, 134, 239, 313, 317
eliminat ion of imaginaries of,
129-131
INDEX
model completeness of, 93
quantifier eliminat ion for , 61-
68
theory of, 27
residue, 101, 102
separably closed , 112-115, 119,
239, 242, 290, 307
structure of, 3
superstable, 220
transcendence basis of, 33
transcendence degree of, 33, 86,
184
valued, 99, 100, 102
of p-adi c numbers, 102
field of definition, 194
finit e in
uniformly, 330
Fi scher-Rabin theorem, 80
forking, 233, 237, 287
formal Laurent ser ies, 98
formul a, 5
positive primitive, 41
atomic, 6
existential , 8, 88
finit e, 330
normal form of, 8
quantifier free , 8
T-equivalent , 43
true in a structure, 5, 7
universal , 8, 88
Fr aisse's th eorem, ] 3
Frobenius morphism, 113, 116, 118,
174
function
definable, 297-299
elementary, 16, ]35 , 196
function field , 306
generic element , 190
geometry, 286
graph, 280
random, 231
structure of, 3
group
pp-definable, 283
365
w-st abl e, 184-192, 220, 302, 303,
310
abelian- by-finite , 284
algebraic, 301-304
cent re of, 121
definable, 302, 317
existent ially closed, 106, 119
linear, 121,301
linear algebraic, 302
o-minimal , 315
of finit e Morl ey rank, 303, 310
of finit e type, 305
ordered abelian, 101
divisible, 313
quotient , 122
special, 122
structure of, 3
torsionfree abelian, 222, 289
groups
elementary class of, 105
Godel Incomplet eness theorem, 40, 280
heir , 238, 252, 290
Hensel's lemma, 98
Herbrand universe, 19
Hilbert 's Basis t heorem, 38, 292
Hilbert 's Nullstellens atz, 82, 86, 93,
119, 292
Hilbert's Seventeent h problem, 86, 95,
119
homomorphism, 4
pure, 151
Hrushovski- Weil theorem, 303, 310,
340
ideal
differ ential , 210
prime, 210
prime, 293
radical, 292
ideal element , 139
independence, 287
independence system, 236
good, 231, 249
independ ent set , 177
366
Induct ion Principle, 7, 29
infini t e sets
theory of , 22
inj ectivity-implies-surj ectivity th eorem,
60
int erpret ability, 279
invariance, 228, 233
invariant statement , 70, 73
invariant system, 225
irr educible components , 293
isomorphism, 5
partial , 10
Knight - Pillay - St einhorn th eorem,
78
Kolchin constructible, 112, 134
Kolchin topology, 112
Lagrange's th eorem, 43, 280
language, 1
Lindstrorri's theorem, 7, 29
linear order
expansion of, 313
linear orders, 78
class of, 23
dense, 52-54, 93
52
dense wit hout endpoints, 24, 32
24, 47-52, 93
eliminat ion of qu antifiers for ,
48
th eory of , 23, 218
linearl y independe nt set, 169
locall y modular , 288
Lowenheim-Skolem theorem, 28, 85,
182, 226, 274, 275
downward, 19, 29
Macintyre's theorem, 192, 284
manifold, 299-302
affine, 299
semiaffine, 299
Manin-Mumford conj ecture, 305,310
model, 8
INDEX
A-satu rated, 144
A-universal, 145
homogeneous, 147
minimal , 87
pri me, 87, 133, 196-209,216,220,
254, 270, 328
saturated, 133, 143-150, 180
weakly A-homogeneous, 145
model companion, Ill , 115, 117, 215
module, 239
algebraically compact, 152
indecomposable, 153
pure inj ective, 152
pure inj ect ive hull of, 152
modules, 27, 41, 290
t heory of, 68-76
monotonicity th eorem, 348
Mordell's conjecture , 304
Mordell-Lang conj ect ure , 304-310
Morl ey degree, 161, 189, 220, 241
of a type, 165
Morl ey rank, 158-168, 180, 220, 230,
241, 288, 294, 298, 307, 313,
346
of a type, 165
Morl ey's existence t heorem, 271
Morl ey's th eorem , 133, 181,271 ,273-
279, 290
mor phism, 297, 302
n- t ype, 293
complet e, 137
Neumann's lemma, 70,72
nilradical, 108
non-for king ext ension, 252, 259
number field, 304
omitting types th eorem, 157, 198, 217
open box , 322
open mapping th eorem, 259, 273
order property, 237, 243
ordered field, 104, 313
of reals, 65, 95
real closur e of, 93
st ructur e of, 3
INDEX
orthogonality, 255
p-adic topology, 96
p-b asis , 114
P=NP problem, 80
parameters , 35
partition, 330
polish space, 222
polynomial , 110
difference, 116
differential, 110, 115
separable, 113
pp-elimination of quantifiers, 76
pp-formula, 41, 69, 151
pp-type, 152
pr edecessor , 47, 263
presentation, 264
proj ective space, 299
Priifer group, 155
pure injectivity, 150
quantifier elimination, 43-82, 87, 88,
345
random graph
t heory of, 239
rank, 158
real analysis, 341
recursive sets, 39
recursively enumerable set , 40
residue field , 99
Ressayre's Uniqueness theorem, 202,
271
ring
commutative, 107
different ial , 109
existent ially closed , 106
ordered , 317
reduced, 109
rmgs
elementary class of, 105
Robinson's t est , 88-91
Rudin-Keisler relation, 254, 290
Ryll-Nardzewski 's theorem, 217, 220
Schanuel's Conjecture, 343
367
semialgebraic set , 38, 67, 134
semidefinite positive, 95
sentence, 5, 7
separant , 211
Shelah's uniqueness theorem, 220, 270-
273,290
sign change property, 98
small subset of n, 148
smooth equivalence relation, 223
spectrum function, 226
stationarity over models, 236
stationary logi c, 290
strong homogeneity th eorem, 147
strongly minimal set , 163, 168-172,
288, 304
structure
X-definable, 121
X-interpr etable, 123
w-stable, 185, 220, 291, 308
R an, etcp, 345
Ran , 344
R
exp
, 341
basis of, 177
definable, 40, 121, 302
dimension of, 177
exist ent ially closed , 105, 110, 119
expansion of, 5
extension of, 85
interpretable, 123
locally modular, 283
minimal , 77, 168, 176, 179
o-minimal, 78, 178, 179, 313, 318
restriction of, 5
simple, 232
stable, 236
strongly minimal , 168, 256, 282,
287, 292, 308
superstable, 241
trivial , 283
two-sorted , 100
universe of, 2
unstable, 236
structures, 2
elementarily equivalent , 10
subanalytic set, 344
368
globally, 344
subexpon ential set , 342
subgroup
definable
connected, 188
pp-definabl e, 41, 69
submodule
pure, 151
subst ruct ure , 5, 85
element ary, 14
existent ial, 15
finitely generated, 5
generated, 5
successor , 47, 263
symmet ry, 228, 234
Tarski 's t heorem, 54, 296, 341
Tarski-Seidenberg theorem, 38, 67
Tarski-Vaught theorem, 17, 158
Terj ani an 's counterexample, 100, 102
terms, 6
t heory, 21,22
A-categorical , 28
of vectorspaces, 34
ACF, 26,45, 54, 58, 88, 91, 111
ACFA, 117, 118, 131,231,288
ACF
o
, 33, 57,87
ACF
p
, 26, 32, 33, 45, 57,77, 88,
149, 169, 184, 197,219, 261
DCF
o
, Ill , 112, 197, 209-217,
287
DCF
p
,115
dLO ,52
dLO+, 48, 50, 51, 77, 87
DLO- , 24, 32, 52, 53, 160, 174,
181, 182, 217, 218, 226
RCF , 27, 33, 61, 65, 66, 80, 87,
88,91, 93, 95, 104, 111, 129,
175, 197, 341
SCF
p,
114
T
p
, 102
Tn, 27
w-stable, 181-184,220,230,242-
261, 270, 346
K. T' , 169
INDEX
n T , 68, 70, 73
Booleanl y A-categori cal , 219
categorical, 133, 274
classifiabl e, 225, 227, 261-270
complete , 19, 30, 102
complet ions of, 31, 45
consistent, 21
decidable, 40, 44, 342
deep, 268
depth of, 268
independence syst em of, 228
model companion of, 105, 117
mod el complete, 34, 46, 85-96,
102, 103, 117, 119, 212, 288,
342, 344, 345
not classifiable, 237
o-minimal, 78-79,178,226,234,
313, 344, 345, 348
of a 1-ar y functi on, 262
of a class of models, 21
of a model, 31
of an equivalence relati on, 256
of infini te sets , 32, 282
of t wo equivalence relati ons, 265
present abl e, 264
rich, 19
rosy, 348
shallow, 268
simple, 227- 235, 249, 289, 290,
328
st abl e, 235-239, 243
strongl y minimal , 76-77,163,168,
182, 184,220,221,225,227,
236, 239, 261, 274, 346
superstable, 239-242, 252
tot ally t ranscendent al, 133, 181-
184, 346
unstable, 236, 237
weakly o-minimal, 346
t horn-i ndependence, 348
topological space
compact , 140
Hausdorff, 140
totally disconnected, 140
t ranscendence basis, 170, 295
INDEX
t ra nscendence degree, 33, 92, 149, 170,
294
transiti vity, 228, 234
tree, 264
rank of, 267
well founded , 267
trichotomy t heorem, 341
Turing machi ne, 39, 79, 80
ty pe, 133, 136-1 43
RK-minimal , 255
algebraic, 141, 166
complet e, 137
consistent , 137
definabl e, 238, 244
depth of, 268
generic, 190, 294
isolat ed , 141, 156, 198
reali zati on of, 139
regul ar , 241
stabilizer of, 188
stationary, 240
st rongly regul ar , 256
type of, 138
ultrafilter , 136
uniqueness t heorem, 146
universal domain, 86
universal ity t heorem, 145
valuation map , 99, 101
valuation ring, 99
variety
abeli an , 304
algebra ic, 37, 285, 292, 299
irredu cibl e, 293, 294, 298
Vaught 's conject ure , 290, 329
Vaught 's t heorem, 32
vectorspace, 230, 239
st ruct ure of, 4
t heory of
quant ifier eliminat ion for , 75
weak homogeneity th eorem, 145
well ordered sets
class of, 24
369
word probl em, 107
Zar iski geomet ries, 285
Zariski st ruct ure, 286
Zari ski to pology, 38, 117, 292, 296,
297, 302
Ziegler spect rum, 154
Zilber 's conject ur e, 279-286,289,309,
340
Zilber 's Indecomposabilit y t heorem, 190,
309
Zorn's lemma, 94
TRENDS IN LOGIC
1. G. Schurz: The Is-Ought Problem. An Investigation in Philosophical Logic. 1997
ISBN 0-7923-4410-3
2. E. Ejerhed and S. Lindstrom (eds.): Logic, Action and Cognition. Essays in Philo-
sophical Logic. 1997 ISBN 0-7923-4560-6
3. H. Wansing: Displaying Modal Logic. 1998 ISBN 0-7923-5205-X
4. P. Hajek: Metamathematics of Fuzzy Logic. 1998 ISBN 0-7923-5238-6
5. HJ. Ohlbach and U. Reyle (eds.): Logic, Language and Reasoning. Essays in Honour
of Dov Gabbay. 1999 ISBN 0-7923-5687-X
6. K. Dosen: Cut Elimination in Categories. 2000 ISBN 0-7923-5720-5
7. R.L.O. Cignoli, LM.L. D'Ottaviano and D. Mundici : Algebraic Foundations ofmany-
valued Reasoning. 2000 ISBN 0-7923-6009-5
8. E.P. Klement, R. Mesiar and E. Pap: Triangular Norms. 2000
ISBN 0-7923-6416-3
9. Y.E Hendricks: The Convergence of Scientific Knowledge. A View From the Limit.
2001 ISBN 0-7923-6929-7
10. J. Czelakowski: Protoalgebraic Logics. 2001 ISBN 0-7923-6940-8
11. G. Gerla: Fuzzy Logic. Mathematical Tools for Approximate Reasoning. 2001
ISBN 0-7923-6941-6
12. M. Fitting: Types, Tableaus, and Godel's God. 2002 ISBN 1-4020-0604-7
13. E Paoli: Substructural Logi cs: A Primer. 2002 ISBN 1-4020-0605-5
14. S. Ghilardi and M. Zawadowki: Sheaves , Games, and Model Completions. A Cat-
egorical Approach to Nonclassical Propositional Logics. 2002
ISBN 1-4020-0660-8
15. G. Coletti and R. Scozzafava: Probabili stic Logic in a Coherent Setting. 2002
ISBN 1-4020-0917-8; Pb: 1-4020-0970-4
16. P. Kawalec: Structural Reliabilism. Inductive Logic as a Theory of Justification. 2002
ISBN 1-4020-1013-3
17. B. Lowe, W. Malzkorn and T. Rasch (eds.): Foundations of the Formal Sciences
II. Applications of Mathematical Logic in Philosophy and Linguistics, Papers of a
conference held in Bonn, November 10-13,2000.2003 ISBN 1-4020-1154-7
18. R.J.G.B. de Queiroz (cd.): Logic for Concurrency and Synchronisation. 2003
ISBN 1-4020-1270-5
19. A. Marcja and C. Toffalori: A Guide to Classical and Modern Model Theory. 2003
ISBN 1-4020-1330-2; Pb 1-4020-1331-0
KLUWER ACADEMIC PUBLISHERS - DORDRECHT / BOSTON / LONDON

Annalisa Marcja, Carlo Toffalori A Guide To Classical and Modern Model Theory Trends in Logic 2003

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Annalisa Marcja, Carlo Toffalori A Guide To Classical and Modern Model Theory Trends in Logic 2003

Uploaded by

Copyright:

Available Formats

A GUIDE TO CLASSICAL AND MODERN MODEL THEORY

Definition 7.1.4 Let T be a complete theory, (11 be an ordinal. T is said

252 CHAPTER 7. CLASSIFYING

You might also like