INTRODUCTION TO MODEL-THEORETIC STABILITY

L. VAN DEN DRIES

1. Introduction
Stability theory is a set of ideas and techniques in model theory that
originated in Morleys proof in the 60s of the Los conjecture: if a theory
in a countable language is categorical in one uncountable power, then it is
categorical in all uncountable powers. Morley discovered that such a theory
must be omega-stable, which means that its models and their denable sets
have some highly restrictive and therefore useful properties. Shelah then
took up the ambitious problem of describing explicitly all functions I(T, )
for complete theories T in countable languages, where I(T, ) is the number
of models of T of size (up to isomorphism) with ranging over innite
cardinals. On general grounds (which?) there can only be a limited number
of possibilities for such functions. A full solution was achieved by Shelah
in a remarkable and extensive body of work, but more important than the
precise answer are the intrinsic dividing lines between (complete) theories
exposed in this way. Here intrinsic means, roughly speaking, invariant
under bi-interpretability. One such dividing line is
stable versus unstable,
and Shelah was able to associate dimension-like quantities to types and
denable sets in models of stable theories.
Later applications in algebra and number theory are closely connected to
the new insights concerning denable sets gained in this way.
It may be surprising that a problem on how many models a theory has of a
given size can be relevant for the structure of denable sets in a given model.
The way this happens is via types: elementary extensions of a given model
/are constructed from /by realizing types over /; in particular, the size
of various type spaces over /determines more or less how many elementary
extensions of / of various kinds there can be. But a type space over /
is just the Stone dual of a boolean algebra of denable sets in /, and the
size of a type space more or less reects the complexity of the corresponding
boolean algebra of denable sets. (I am intentionally vague about size
and complexity; there are various ways to make this precise.) Perhaps
the remarkable properties of denable sets in stable structures could have
been discovered in another way. As a fact of history, however, they came
Date: November 2005.
1
2
to light via the study of type spaces forced by questions on the number of
models of a theory.
Throughout, m, n (sometimes decorated with subscripts or accents) range
over the set N = 0, 1, 2, . . . of natural numbers. We let ,

, etc., range
over innite cardinals. For a set X we let [X[ be the size (cardinal) of X.
Given sets P, Q and R P Q, we dene for a P,
R(a) := b Q : (a, b) R Q (the section of R above a),
and we view R as describing the family
_
R(a)
_
aP
of subsets of Q. (Of
course, this notation is only justied when P and Q are clear from the
context.) A partial map f from the set P to the set Q (notation: f : P Q)
is a map f : P

Q with P

P.
For an equivalence relation E on a set P we let P/E be the quotient set (its
elements are the E-classes E(p) with p P). Abusing language, we call E
nite if P/E is nite.
2. Boolean algebras with rank
The Cantor rank
1
of a denable set X in a model is, roughly speaking, an
ordinal that measures to what extent X can be split up in smaller denable
sets. It behaves as a kind of dimension of X, and
Cantor rank = Morley rank
if the ambient model is
0
-saturated, as we shall see later.
A key point is that the Cantor rank of X can be dened purely in terms
of the boolean algebra of denable subsets of X. This suggests introducing
such a notion of rank for elements of any boolean algebra.
We also intend this section as a review of the Stone representation of a
boolean algebra, which underlies our later use of types.
Let Or be the class of ordinals. For convenience we add two extra elements
and + to Or, and extend the usual linear ordering on Or to a linear
ordering on Or

:= Or , + by letting < < + for each

ordinal . Below we let , , range over ordinals.
Terminology and notations concerning boolean algebras. Recall
that a boolean algebra B is a set with distinguished elements 0 and 1, binary
operations and (join and meet), and a unary operation (complement),
such that certain equational laws are satised.
Let B be a boolean algebra. For a, b B we dene
a b : a = a b,
which makes B into a poset (partially ordered set), and we say that a and
b are disjoint if a b = 0. An atom of B is an a > 0 in B such that there is
no b B with a > b > 0. We let at(B) be the set of atoms of B.
1
The term Cantor-Bendixson rank is more common.
3
Given any set X, the (boolean) algebra of subsets of X has the subsets
of X as its elements, with and X as its 0 and 1, the binary operations
of taking union and intersection as its join and meet operations, and the
unary operation of taking the complement relative to X as its complement
operation. The atoms of this boolean algebra are the singletons x with
x X. These atoms generate the (boolean) subalgebra whose elements
are the nite and conite subsets of X. (A conite subset of X is the
complement in X of a nite subset of X.) Part of the Stone representation
theorem says that any boolean algebra embeds into the boolean algebra of
subsets of X, for some set X.
We now x a boolean algebra B, and let a, b, c (with or without subscripts)
denote elements of B. By b = b
1
+ +b
k
we mean that b = b
1
b
k
with pairwise disjoint b
1
, . . . , b
k
. We also put a b := a (b).
An ideal of B is a set I B such that 0 I, and for all a, b,
a b I a I, a, b I a b I.
Let I be an ideal of B. This yields an equivalence relation =
I
on B by
setting
a =
I
b a i = b i for some i I.
Let a/I be the equivalence class of a, and make the set B/I := a/I : a B
into a boolean algebra by requiring that a a/I : B B/I is a boolean
algebra homomorphism.
Given any set H B, the ideal (H) of B generated by H (i.e. the smallest
ideal of B, under set inclusion, that contains H) is given by
(H) := x B : x h
1
h
n
for some h
1
, . . . , h
n
H.
Thus
_
at(B)
_
is the ideal generated by the atoms of B, and its elements are
the a
1
+ + a
m
with atoms a
1
, . . . , a
m
, with m = 0 yielding the element
0 B. Each a
_
at(B)
_
has a unique representation as a sum of atoms: if
a = a
1
+ +a
m
= b
1
+ +b
n
with atoms a
i
and b
j
, then m = n and a
1
, . . . , a
m
= b
1
, . . . , b
n
.
Dening Cantor rank. By transnite recursion we assign to each ordinal
an ideal I

of B such that I

for :
(1) I
0
:=
_
at(B)
_
, the ideal generated by the atoms of B;
(2) for > 0, assume inductively that I

is an ideal of B for all < ,

and that I

whenever < ; put I

<
:=

<
I

; then I

is dened to be the ideal of B that contains I

<
and whose image in
B/I
<
is
_
at(B/I
<
)
_
, the ideal generated by the atoms of B/I
<
.
For convenience we put I
<0
:= 0, the trivial ideal of B.
Lemma 2.1. If b / I

, then there is an innite subset of B whose elements

are pairwise disjoint, < b, and outside I
<
.
4
Proof. We rst do the case = 0. Let b / I
0
. If there are innitely many
atoms < b, then the set of such atoms has the desired property. If there
are only nitely many atoms < b, then by subtracting these atoms from b
we reduce to the case that there are no atoms < b. In this case we take
some nonzero b
0
< b, then some nonzero b
1
< b b
0
, next some nonzero
b
2
< b b
0
b
1
, and so on. Then the set of all b
n
has the desired property.
The general case follows likewise by considering B/I
<
instead of B.
Note that if I

= I
+1
, then I

= I

for all ordinals .

We now assign to each b its Cantor rank, an element of Or

and denoted
by CR(b), and by CR
B
(b) if we wish to indicate the dependence on B. We
set CR(0) := ; if b ,= 0 and b I

for some , then we let CR(b) be the

least such ; if b ,= 0 and b / I

for all , then CR(b) := +. Thus

I

= b : CR(b) .
For example, if X is innite and B is the algebra of nite and conite subsets
of X, then the nonempty nite subsets of X have Cantor rank 0, and the
conite subsets of X have Cantor rank 1. Note that the lemma above has
the following reformulation:
CR(b) > there is an innite sequence a
0
, a
1
, a
2
, . . . with a
n
< b and
CR(a
n
) for all n, and a
m
a
n
= 0 for all m ,= n.
This equivalence is constantly and tacitly used in inductive proofs below.
Our rank function clearly satises
(i) CR(b) = b = 0,
(ii) a b = CR(a) CR(b),
(iii) CR(a b) = max(CR(a), CR(b)),
(iv) if < CR(b) < +, then = CR(a) for some a.
Property (iv) says that the ordinals that occur as Cantor ranks of elements
of B form an initial segment of Or. In particular, these ordinals are < [B[
+
,
the least cardinal > [B[.
We say that b is ranked if < CR(b) < +, that is, CR(b) is an
ordinal. We dene the Cantor degree CD(b) (or CD
B
(b)) of a ranked element
b to be the largest d 1 such that there are b
1
, . . . , b
d
of the same Cantor
rank as b with b = b
1
+ + b
d
; in other words, if CR(b) = , then CD(b)
is the number of atoms of B/I
<
that are b/I
<
.
Lemma 2.2. Cantor rank and degree are related as follows:
(v) If < CR(a) = CR(b) < + and a b, then CD(a) CD(b).
(vi) If < CR(a) = CR(b) < +, then CD(a b) CD(a) + CD(b)
with equality if a b = 0,
(vii) If CR(a) < CR(b) < +, then CD(a b) = CD(b).
Call b Cantor irreducible if b is ranked and CD(b) = 1; in that case there
are no b
1
and b
2
of the same Cantor rank as b such that b = b
1
+b
2
. Clearly:
5
(viii) (CR(a) = 0, CD(a) = 1) a is an atom;
(ix) CR(b) = 0 b = a
1
+ +a
d
for some d 1 and atoms a
1
, . . . , a
d
;
(x) if b is ranked of Cantor degree d and b = b
1
+ + b
d
with each b
i
of the same Cantor rank as b, then each b
i
is Cantor irreducible.
Cantor rank and degree behave as expected under morphisms:
Proposition 2.3. Let : B C be a boolean algebra morphism and b B.
(1) If is injective, then CR(b) CR((b)), and in case of equality with
b ranked we have CD(b) CD((b)).
(2) If is surjective, then CR(b) CR((b)), and in case of equality
with b ranked we have CD(b) CD((b)).
Proof. Suppose is injective. Then an easy induction shows that CR(b) >
implies CR((b)) > , from which the rst part of (1) follows. The second
part of (1) is then immediate from the denition of degree. We prove (2)
by induction on CR(b). The case CR(b) = is trivial, so let CR(b) = ,
and assume that the desired result holds for smaller values of CR(b). Let
c := (b), and suppose c = c
1
+ + c
k
with k 1 and CR(c
i
) for
i = 1, . . . , k. Take b
1
, . . . , b
k
B with (b
i
) = c
i
. Replacing b
i
by b
i
b if
necessary, we may assume b
i
b for all i. Note that

_
b
i

j=i
b
j
_
= (b
i
)

j=i
(b
j
) = c
i

j=i
c
j
= c
i
,
so replacing each b
i
by b
i

_
j=i
b
j
we may assume b
1
, . . . , b
k
are pairwise
disjoint. If CR(b
i
) < , then by the inductive hypothesis CR(c
i
) < , a
contradiction. Hence CR(b
i
) = for all i, and thus k CD(b). This
bound on k shows that CR(c) cannot be larger than , and also implies the
inequality on the degrees in case CR(b) = CR(c).
Let B[b := a[a b, and consider B[b as a boolean algebra in its own right,
by restricting and to B[b. Note that B[b has b as largest element, so
B[b is not a boolean subalgebra of B when b ,= 1, but the map a a b :
B B[b is a (surjective) morphism of boolean algebras. The following is
almost obvious from the denitions.
Lemma 2.4. Let a b. Then CR
B
(a) = CR
B|b
(a). If in addition a is
ranked, then CD
B
(a) = CD
B|b
(a).
In combination with a previous remark this lemma implies:
If < CR(b) < +, then there is a b with CR(a) = .
When is every nonzero element ranked? To answer this question, we
rst review lters and the Stone representation. A lter of B is just the
dual of an ideal of B: it is a set F B such that 1 F, and for all a, b we
have a b F a F, and a, b F a b F. Equivalently, it is a set
F B such that F := b : b F is an ideal of B. Each a determines a
lter F
a
:= b : b a of B.
6
A lter F of B is said to be proper if 0 / F, that is, F ,= B. An ultralter
of B is a proper lter F of B such that for each a, either a F or a F.
Note that F
a
is an ultralter of B i a is an atom. Using Zorn it is easy to
see that every proper lter of B is contained in a maximal proper lter of
B. One also checks easily that the maximal proper lters of B are exactly
the ultralters of B. Finally, each proper lter F of B is the intersection
of the ultralters of B that contain F. These facts easily yield the Stone
representation, to which we now turn. The Stone space St(B) of B is the
set of ultralters of B. To a we assign
[a] := F St(B) : a F, a subset of St(B),
and this assignment satises [0] = , [1] = St(B), and
[a] [b] = [a b], [a] [b] = [a b], [a] = St(B) [a].
Thus a [a] is a morphism of B into the boolean algebra of subsets of St(B).
It is in fact an embedding, called the Stone representation of B. We use
the word Stone space because St(B) is given the so-called Stone topology,
which has the sets [a] as basis. It makes St(B) a compact hausdor space
whose clopen sets are exactly the sets [a]. Thus the Stone representation of
B maps B isomorphically onto the algebra of clopen subsets of St(B).
Next a combinatorial notion: an innite binary tree in B is a family (a
j
) of
nonzero elements of B, where j = j
1
, . . . , j
n
ranges over the nite sequences
of zeros and ones, with a

= 1 (with the empty sequence), and such that

for each j as above we have a
j
= a
j,0
+a
j,1
.
This combinatorial notion is related to an algebraic notion: B is said to
be atomless if 1 ,= 0 but B has no atom. Note that if (a
j
) is an innite
binary tree in B, then the elements a
j
generate an atomless subalgebra of
B. Conversely, if B has an atomless subalgebra, then B has an innite
binary tree.
Example of an atomless boolean algebra: take the subsets of [0, 1) R
that are nite unions of intervals [a, b) with 0 a < b 1; these subsets are
the elements of an atomless subalgebra of the algebra of subsets of [0, 1).
We can now state an answer to the question above:
Theorem 2.5. The following four conditions are equivalent:
(i) B has no innite binary tree,
(ii) B has no atomless subalgebra,
(iii) each non-zero element of B is ranked,
(iv) St(A) is countable for each countable subalgebra A of B.
Moreover, (i) implies (v), and is equivalent to (v) if B is countable:
(v) [ St(B)[ [B[.
Towards the proof of this result we rst observe:
Lemma 2.6. Suppose CR(a) = +. Then there are a
1
, a
2
such that
a = a
1
+a
2
, CR(a
1
) = CR(a
2
) = +.
7
Proof. Take an ordinal > [B[
+
. From CR(a) > it follows that a = a
1
+a
2
with CR(a
1
), CR(a
2
) . Then CR(a
1
) = CR(a
2
) = +.
For Cantor irreducible a we put F(a) := b : CR(a b) = CR(a), hence
also F(a) = b : CR(a b) < CR(a). Clearly F(a) is then an ultralter on
B, and if a is an atom, then F(a) = F
a
. For Cantor irreducible a, b we have
F(a) = F(b) CR(a) = CR(b) = CR(a b).
We say that an ultralter on B is ranked if it contains a ranked element.
Given a ranked ultralter F on B, take an a F of minimal Cantor rank,
and of minimal Cantor degree of that rank. Then a is Cantor irreducible,
and one checks easily that then F = F(a). So a F(a) maps the set of
Cantor irreducible elements of B onto the set of ranked ultralters of B.
Proof of the Theorem. We already observed the equivalence of (i) and
(ii). To prove that (i)(iii), assume (i) and suppose B has a nonranked
non-zero element. Then CR(1) = +, so by Lemma 2.6 we get an innite
binary tree (a
j
) in B such that all a
j
have Cantor rank +, contradicting
(i). To prove (iii)(ii), assume (iii), and let A be a subalgebra of B. We
can assume 1 ,= 0. Since 1 is ranked in B, it is ranked in A by (1) of
Proposition 2.3, so CR
A
(a) = 0 for some a A, so A has an atom, and
is therefore not atomless. We have now shown that (i), (ii) and (iii) are
equivalent.
Next we prove (iii)(v). Assume (iii). Then each ultralter of B is
ranked, hence of the form F(a) for some irreducible a B, so (v) holds.
Assume B is countable. To show (v)(i), we prove the contrapositive:
Suppose we have an innite binary tree (a
j
) in B. Each innite sequence
j
1
, j
2
, j
3
, . . . of zeros and ones leads to a subset 1, a
j
1
, a
j
1
,j
2
, a
j
1
,j
2
,j
3
, . . .
of B that is contained in some ultralter on B, and dierent such innite
sequences give rise in this way to necessarily dierent ultralters. Thus B
has 2

0
many ultralters.
The last two arguments also give the equivalence of (i) with (iv) since
condition (i) is inherited by subalgebras. This concludes the proof of the
theorem.
In the presence of a set of generators. Let our boolean algebra B be
equipped with a distinguished subset G that generates B. It turns out that
we can recover the notion of (un)stability in this setting in a very natural
way. The combinatorial fact behind this is the following:
Lemma 2.7. Suppose H is an innite set and ( is a collection of subsets
of H such that [([ > [H[. Then for each n there are distinct elements
h
1
, . . . , h
n
H and sets C
0
, . . . , C
n
( such that
h
1
, . . . , h
n
C
j
= h
1
, . . . , h
j
for j = 0, . . . , n,
in other words: h
i
C
j
i j, for i = 1, . . . , n, j = 0, . . . , n.
8
Proof. For given n the property holds for ( i it holds for the collection
H ( := H C : C (: if h
1
, . . . , h
n
H and C
0
, . . . , C
n
witness this
property, then h
n
, . . . , h
1
and HC
n
, . . . , HC
0
witness it as well. We now
proceed by induction on n. The case n = 0 is obvious. Suppose the desired
result holds for a certain n, for all H and ( satisfying the hypothesis of the
lemma. Take a set C ( and put := [H[. We make the following case
distinction:
Case 1. ( C := X C : X ( has size > . Then there is a c C
such that more than sets Y ( C do not contain c: otherwise, each
c C would be outside at most many Y ( C, hence C ,= Y for at
most many Y ( C, contradicting the assumption of case 1. Take such
a c, and put T := Y ( C : c / Y , a collection of subsets of C with
[T[ > . Hence by the inductive hypothesis applied to C and T there are
distinct h
1
, . . . , h
n
C and sets Y
0
, . . . , Y
n
T such that
h
1
, . . . , h
n
Y
j
= h
1
, . . . , h
j
for j = 0, . . . , n.
We have Y
j
= C
j
C with C
j
(, j = 0, . . . , n. Put h
n+1
:= c and
C
n+1
:= C. Then
h
1
, . . . , h
n+1
C
j
= h
1
, . . . , h
j
for j = 0, . . . , n + 1.
Case 2. ( C := X C : X ( has size . Then ( (H C) has size
> , so (H () (H C) has size > . Then case 1 applies to H ( and
HC in place of ( and C, and we are done by the remark at the beginning
of the proof.
This result of Erdos-Makkai is a perfect example how the uncountable can
reect combinatorial complexity of a nite nature. (Is there a nite version
of this result?)
To apply Lemma 2.7, we note that dierent ultralters of B have dierent
intersections with G, because each ultralter of B equals b : (b) = 1 for
some boolean algebra morphism : B 0, 1, and each such morphism is
uniquely determined by its restriction to G.
Corollary 2.8. Suppose CR(1) = +. Then for each n there are distinct
elements g
1
. . . , g
n
and ultralters F
0
, . . . , F
n
of B such that
g
1
, . . . , g
n
F
j
= g
1
, . . . , g
j
for j = 0, . . . , n.
Proof. By the assumption B has a countable atomless subalgebra. Hence
we have a countable subset H of G such that the subalgebra A generated by
H has an atomless subalgebra. Thus A has uncountably many ultralters.
Dierent ultralters of A intersect H in dierent subsets, and each ultralter
of A extends to an ultralter of B. Now apply the lemma above to H and
the collection ( of intersections of ultralters of B with H.
9
3. Many-sorted Structures
Model theory is traditionally developed for one-sorted structures, but it is
now generally recognized that the many-sorted setting is better suited for
expressing various basic constructions like adding imaginaries. Also, natural
mathematical structures are often many-sorted to begin with. A many-
sorted (or multi-sorted) structure is a family of sets (M
s
)
sS
equipped with
relations
R M
s
1
M
sm
, (s
1
, . . . , s
m
S)
and functions
f : M
s
1
M
sn
M
s
n+1
, (s
1
, . . . , s
n+1
S).
The elements of the index set S are called sorts, and M
s
is the underlying
set of sort s. For example, an incidence geometry is a two-sorted structure,
consisting of a set P whose elements are called points, a set Q whose ele-
ments are called lines, and a relation R P Q between points and lines.
(Depending on the situation this incidence relation R is also assumed to
satisfy certain axioms such as for any two distinct points p
1
, p
2
there is
exactly one line q such that R(p
1
, q) and R(p
2
, q).)
We shall need a bit of syntax in connection with many-sorted structures
even though we wish to deal with issues that transcend our choice of syntax.
A many-sorted language L is a triple (S, L
r
, L
f
) consisting of
(1) a set S whose elements are the sorts of L,
(2) a set L
r
whose elements are the relation symbols of L,
(3) a set L
f
whose elements are the function symbols of L,
where L
r
and L
f
are disjoint, each R L
r
is equipped with an arity
(s
1
, . . . , s
m
) S
m
, and each f L
f
is equipped with an arity (s
1
, . . . , s
n+1
)
S
n+1
. A function symbol of L of arity (s) with s S is also called a con-
stant symbol of L of sort s. The elements of L
r
L
f
are also called nonlogical
symbols of L.
Let from now on L be a language, with S as its set of sorts, unless specied
otherwise. The size of L is the cardinal
[L[ := max
0
, [S[, [L
r
L
f
[,
and we say that L is countable if [L[ =
0
.
An L-structure is a many-sorted structure
/=
_
M; (R
M
)
RL
r , (f
M
)
fL
f
_
, where M = (M
s
)
sS
,
such that for R L
r
of arity (s
1
, . . . , s
m
) its interpretation R
M
in / is
a subset of M
s
1
M
sm
, and for f L
f
of arity (s
1
, . . . , s
n+1
) its
interpretation f
M
in / is a function M
s
1
M
sn
M
s
n+1
. For a
constant symbol c of L of sort s the corresponding M
s
-valued function c
M
is identied with its unique value in M
s
, so c
M
M
s
.
From now on / denotes an L-structure (M; ) with M = (M
s
)
sS
,
unless specied otherwise. If / is understood from the context we often
10
omit the superscript / in denoting the interpretation in / of a nonlogical
symbol of L. Elements of M
s
are also called elements of / of sort s.
Given any tuple (family) s = (s
i
)
iI
of sorts in S, we let M
s
denote the
corresponding product set:
M
s
:=

iI
M
s
i
.
Variables and Terms. We assume that we have available innitely many
symbols, called unsorted variables; these are given once and for all, indepen-
dent of the language L we are dealing with. Given any sort s, the variables
of sort s are the pairs v
s
:= (v, s) where v is an unsorted variable. This con-
vention is to guarantee that if s and s

are dierent sorts, then no variable

of sort s is a variable of sort s

. We also assume that no variable of any sort

is a nonlogical symbol of any language.
A variable of L is a variable of sort s for some s S, and a multivariable
of L is a tuple (x
i
)
iI
of distinct variables of L, where distinct means
that x
i
,= x
j
for i ,= j. The size of the index set I is called the size of x,
and the x
i
are called the variables in x. Often the index set I is nite, say
I = 1, . . . , n, so that x = (x
1
, . . . , x
n
). Given a multivariable x = (x
i
)
iI
of L, with x
i
of sort s
i
for i I, we dene the x-set of / to be the product
set
M
x
:= M
s
=

i
M
s
i
, with s = (s
i
)
iI
,
and we think of x as a variable running over M
x
.
Multivariables x = (x
i
)
iI
and y = (y
j
)
jJ
of L are said to be disjoint if
x
i
,= y
j
for all i I and j J, and in that case we put M
x,y
:= M
x
M
y
.
If in addition I = J and x
i
and y
i
have the same sort for all i I (so that
M
x
= M
y
), then we call x and y disjoint and similar. From now on x and
y denote multivariables of L, unless specied otherwise.
Instead of x has nite size we also say x is nite. A nite tuple a in
/ is a tuple a M
x
for a nite x.
We dene L-terms to be words on the alphabet consisting of the function
symbols and variables of L, obtained recursively as follows: each variable of
sort s S is an L-term (of sort s) when viewed as a word of length 1, and
if f is a function symbol of L of arity (s
1
, . . . , s
n
, s
n+1
) and t
1
, . . . , t
n
are L-
terms of sort s
1
, . . . , s
n
respectively, then ft
1
. . . t
n
is an L-term of sort s
n+1
.
(In practice we often write f(t
1
, . . . , t
n
) to denote ft
1
. . . t
n
, and use similar
devices to increase readability.) Each variable-free L-term t determines in
the usual way an element t
M
of / of the same sort as t, and often we drop
the superscript / in t
M
if it is clear from the context that we mean to refer
to an element of /.
An (L, x)-term is a pair (t, x) where t is an L-term such that each variable
in t is a variable in x. Such an (L, x)-term is also written more suggestively
as t(x), and referred to as the L-term t(x).
11
We extend L to the language L
M
by adding for each s S and a M
s
a
new constant symbol c(a, s) of arity s, and we expand /to an L
M
-structure,
also denoted by / for convenience, by setting c(a, s)
M
:= a. When s and
/ are clear from context, then we call c(a, s) the name of a and denote it
also just by a. Given an L
M
-term t(x) and a tuple a M
x
we obtain in this
way by substitution a variable-free L
M
-term t(a) of the same sort as t, so
t(a) denes an element t(a)
M
of / of the same sort as t, and denoted also
by t(a) when / is clear from context. In other words, we obtain a function
a t(a) : M
x
M
s
, (t of sort s).
Formulas. There is a slightly annoying problem with variables: to dene,
say, the theory Th(/) of / as a set, our usual set theory ZFC seems to
require the variables to be the elements of a set. But for some purposes we
do not want to limit how many variables there are, or even assume that all
variables are elements of a single set. The way out is to assume that bound
(quantied) variables are always taken from a xed countable set (for each
sort), while free (unquantied) variables are not limited in this way. More
precisely, we assume from now on as given a countably innite set of unsorted
variables whose elements are called unsorted quantiable variables. For each
sort s the quantiable variables of sort s are the variables v
s
where v is an
unsorted quantiable variable.
We now x once and for all the usual eight logical symbols:
=
These are assumed to be distinct from all relation symbols, function symbols
and variables of every language.
The L-formulas are words on the alphabet consisting of the nonlogical
symbols of L, the variables of L, and the eight logical symbols, and are
dened recursively in the usual way from the atomic L-formulas with the
proviso that every occurrence of and in a formula is followed immediately
by a quantiable variable. The atomic L-formulas are the words = t
1
t
2
(usually written as t
1
= t
2
for readability) where t
1
and t
2
are L-terms of
the same sort, together with the words Rt
1
. . . t
m
where R L
r
has arity
(s
1
, . . . , s
m
), and t
1
, . . . , t
m
are L-terms of sorts s
1
, . . . , s
m
, respectively. An
(L, x)-formula is a pair , x where is an L-formula all whose free variables
are in x. Such an (L, x)-formula is also written more suggestively as (x),
and referred to as the L-formula (x). Sometimes we separate the free
variables in a formula into two disjoint multivariables: for example, when
referring to an L-formula (x, y) we really mean a triple , x, y consisting
of an L-formula and disjoint multivariables x and y such that each free
variable of is in x or in y.
A formula without free variables is called a sentence, so all variables in a
sentence are quantiable variables. Thus the set of L-sentences has size [L[.
Truth and Denability. We dene in the usual recursive way what it
means for an L
M
-sentence to be true in / (notation: / [= ). The
12
theory of /, denoted by Th(/), is the set of all L-sentences true in /. If
we wish to express the dependence on L we call it the L-theory of / and
denote it by Th
L
(/). Given an L
M
-formula (x) with x = (x
i
) and a tuple
a M
x
we obtain, by substituting for each i the name of a
i
for the free
occurrences of x
i
in , an L
M
-sentence (a). Thus each L
M
-formula (x)
denes the subset
(M
x
) := a M
x
: / [= (a)
of M
x
in /. We also use this notation for a set (x) of L
M
-formula (x):
(M
x
) := a M
x
: /[= (a)for all =

(M
x
).
The denable subsets of M
x
in / are the subsets of M
x
dened in /
by an L
M
-formula (x), and are exactly the elements of a boolean algebra
Def
x
(/) of subsets of M
x
. The 0-denable subsets of M
x
in / are exactly
the sets dened in / by an L-formula (x), and are exactly the elements
of a boolean algebra Def
x
(/[0) of subsets of M
x
.
A partial map h : M
x
M
y
is said to be denable in / if its graph is a
denable subset of M
x
M
y
in /. Also, 0-denability of such a map refers
likewise to 0-denability of its graph.
A denable set in /is a pair X, s with s S
n
for some n and X a denable
subset of M
s
in /. The role of s is just to specify the intended ambient set
M
s
of X. The reason for this convention should be clear: it may happen
that, for example, M
s
= M
s
where s and s

are dierent sorts. In that case,

a set X M
s
could be a denable subset of M
s
in /, according to the
denition above, while not being a denable subset of M
s
in /. Usually
we refer to a denable set X, s in / just by its rst component X, and s is
clear from the context or left implicit.
For an L
M
-formula (x), / [= (x) means that / [= (a) for all a M
x
.
Thus if (x) is a second L
M
-formula, then / [= (x) (x) means
(M
x
) (M
x
), and / [= (x) (x) means (M
x
) = (M
x
).
The primitives of / (that is, the interpretations in / of the nonlogical
symbols of L) are secondary in the kind of model theory we are going to do;
their role is just to generate the 0-denable sets in /. (The denable sets
in / can be dened purely in terms of the 0-denable sets in /; how?) To
elaborate on this, dene a structure on M = (M
s
)
sS
to be a family
_
D(s)
_
s
indexed by the nite sequences s = s
1
, . . . , s
n
in S, such that for each such
s = s
1
, . . . , s
n
:
(1) D(s) is a boolean algebra of subsets of M
s
:= M
s
1
M
sn
;
(2) whenever X D(s) and s

S, then
M
s
X D(s

, s), X M
s
D(s, s

);
(3) whenever s

S, then (a, b, a) : a M
s
, b M
s
D(s

, s, s

);
13
(4) whenever s

S and X D(s, s

), then (X) D(s) where :

M
s
M
s
M
s
is the obvious projection map.
It is easy to check that the family
_
Def
s
(/[0)
_
is a structure on M that
contains the primitives of /, and that it is the smallest structure on M
containing these primitives.
Parameter sets. A parameter set in / is a family A = (A
s
) where
A
s
M
s
for each s. (We omit in / if /is clear from the context.) Such
A gives rise to a sublanguage L
A
of L
M
in the obvious way. If X M
x
is
dened by an L
A
-formula (x), then we also say that X is A-denable in
/, or denable over A in /. The subsets of M
x
that are A-denable in /
are the elements of a boolean subalgebra Def
x
(/[A) of Def
x
(/). A partial
map h : M
x
M
y
is said to be A-denable in / or denable over A in /
if its graph is an A-denable subset of M
x
M
y
in /. Given parameter
sets A = (A
s
) and B = (B
s
) in / we dene
A B : A
s
B
s
for all s,
AB := (A
s
B
s
) (a parameter set in /),
A B := (A
s
B
s
) (a parameter set in /).
Given a parameter set A as above, an A-tuple (in /) is a pair a, s with
s = (s
i
) S
I
and a M
s
, such that a
i
A
s
i
for all i I; this A-tuple is
also referred to as the A-tuple a M
s
, or even the A-tuple a.
Let A = (A
s
) be a parameter set in /. One checks easily that for nite
y, a set Y M
y
is A-denable i for some nite x disjoint from y, some
A-tuple a M
x
and some 0-denable Z M
x,y
we have Y = Z(a).
A nite tuple a M
x
is said to be A-denable in / (or denable over A
in /) if a is an A-denable subset of M
x
, and is said to be A-algebraic
in / (or algebraic over A in /) if a belongs to a nite A-denable subset
of M
x
. (We can omit in / if / is clear from the context.) We let dcl(A)
be the parameter set such that dcl(A)
s
= a M
s
: a is denable over A
for each s, and we dene the parameter set acl(A) likewise, with algebraic
instead of denable. We call dcl(A) (respectively, acl(A)) the denable
closure of A in / (respectively, the algebraic closure of A in /). We say
that A is denably closed in / (respectively, algebraically closed in /) if
dcl(A) = A (respectively, acl(A) = A). It is easy to check that dcl(A) is
denably closed in /, and that acl(A) is algebraically closed in /.
The size of A is the cardinal sum
[A[ :=

s
[A
s
[,
so [A[ < if [S[ < and [A
s
[ < for all s. If L, A and x all have size < ,
then the set of (L
A
, x)-formulas has size < . We say that A is nite if [A[
is nite, innite if [A[ is innite, and countable if [A[
0
.
14
For x = (x
i
)
iI
where each x
i
is of sort s
i
, we put
A
x
:=

iI
A
s
i
.
The parameter set A with A
s
= for all s is also denoted by 0, in accordance
with how we have been using terminology like 0-denable. Sometimes we
shall use a tuple a M
s
with s = (s
i
) S
I
as if it were a parameter set,
for example, when notations like dcl(a) or acl(a) occur. In such a context a
stands for the parameter set (A
s
) such that A
s
is the set of all a
i
with s
i
= s,
for each s. Thus if B is a parameter set and a M
s
as above, then aB and
Ba denote again a parameter set: aB = Ba = AB with A the parameter
set for which a stands. The terminology a-denable and a-algebraic is
used in the same vein.
Exercises and Denitions. Let a M
x
and b M
y
be nite tuples. Show
that b is a-denable i there is a 0-denable partial function h : M
x
M
y
with a in its domain such that h(a) = b. Show that b is a-algebraic i there
is a 0-denable Z M
x
M
y
such that b Z(a) and Z(a) is nite.
Below, notions such as a and b are interdenable are understood to be
with respect to the ambient structure /, and we add in / if we wish to
specify / as the ambient structure. We say that a and b are interdenable
if a is b-denable and b is a-denable. Show the equivalence of (i), (ii), (iii):
(i) dcl(a) = dcl(b);
(ii) a and b are interdenable;
(iii) there is an injective 0-denable partial function h : M
x
M
y
with
a in its domain such that h(a) = b.
Let in addition A be a parameter set. Then b is said to be a-denable
over A if b is a-denable in the L
A
-structure /. Show that b is a-denable
over A i b is Aa-denable. Also, a and b are said to be interdenable over
A if a and b are interdenable in the L
A
-structure /. Show that a and b
are interdenable over A i dcl(Aa) = dcl(Ab).
Likewise, b is said to be a-algebraic over A if b is a-algebraic in the L
A
-
structure /. Show that b is a-algebraic over A i b is Aa-algebraic. Finally,
a and b are said to be interalgebraic over A if a and b are interalgebraic
in the L
A
-structure /. Show that a and b are interalgebraic over A i
acl(Aa) = acl(Ab).
Types. Let A be a parameter set in /. A partial x-type over A in /
is a set (x) of L
A
-formulas (x), each nite subset of which is realized in
/ by some element of M
x
. An x-type over A in / is a partial x-type
p(x) over A in / such that for each L
A
-formula (x), either (x) p(x)
or (x) p(x). For b M
x
we let tp
M
x
(b[A) be the x-type over A in /
realized by b (and we leave out the superscript / or subscript x if / or
x are clear from context). The set of x-types over A in / is denoted by
15
St
M
x
(A), or St
x
(A) if / is clear from context, and we have a bijection
p(x) (M
x
) : (x) p(x) : St
x
(A) St
_
Def
x
(/[A)
_
of this set of x-types onto the Stone space of the boolean algebra Def
x
(/[A).
Given parameter sets A and B in /such that A B, we have the surjective
restriction map
p pA : St
x
(B) St
x
(A)
given by pA := (x) p : (x) is an L
A
-formula.
We declare / to be -saturated if for each parameter set A in / of size
< and each variable v of L, each v-type over A in / can be realized in
/. The following is proved as in the one-sorted case:
Lemma 3.1. If / is -saturated, A a parameter set of size < and x of
size , then each partial x-type over A in / can be realized in /, that
is, (M
x
) ,= for each partial x-type over A in /.
Note also that with the assumptions of this lemma, if X is a denable subset
of M
x
contained in a union

jJ
Y
j
of A-denable subsets Y
j
of M
x
, then
X

jJ
0
Y
j
for some nite J
0
J.
Exercise. With the assumptions of the last lemma, suppose

iI
X
i

_
jJ
Y
j
where the X
i
and Y
j
are A-denable subsets of M
x
. Show that there are
nite I
0
I and J
0
J such that

iI
0
X
i


jJ
0
Y
j
.
Comment. The result of the last exercise can sometimes be used to prove
that a set X M
x
with nite x is A-denable, where / is -saturated and
[A[ < : if one manages to represent X both as an intersection X =

iI
X
i
and as a union X =

jJ
Y
j
of A-denable subsets X
i
and Y
j
of M
x
, then
by this exercise there are nite I
0
I and J
0
J such that

iI
0
X
i
= X =
_
jJ
0
Y
j
,
so X is A-denable.
Elementary maps. Let ^ =
_
(N
s
);
_
be a second L-structure, and f
a partial map from / to ^. The latter means that f is a family (f
s
) with
f
s
: A
s
N
s
and A
s
M
s
for each s S; the parameter set A = (A
s
) is
called the domain of f. Given an A-tuple a M
s
, s = (s
i
) S
I
, we dene
fa = f(a) :=
_
f
s
i
(a
i
)
_
N
s
. We call f a partial elementary map from /
to ^ if
/ [= (a) ^ [= (fa)
for every nite A-tuple a M
x
and L-formula (x). Note that then each
f
s
: A
s
N
s
is injective.
Let f be a partial elementary map from /to ^ as above. If A
s
= M
s
for
all s, then we call f an elementary embedding from / into ^. If A
s
= M
s
16
and f
s
(M
s
) = N
s
for all s, then we call f an isomorphism from / onto ^,
and if in addition / = ^, we call f an automorphism of /.
We dene / to be strongly -homogeneous if any partial elementary map
from / to itself with domain of size < can be extended to an automor-
phism of /.
The automorphism group. The automorphisms of / form a group
Aut(/) under the obvious composition operation. A parameter set A in /
yields the subgroup Aut(/[A) of Aut(/) consisting of the automorphisms
f of / such that f
s
(a) = a for each s S and a A
s
. Note that if /
is strongly -homogeneous, [A[ < and a, b M
x
have the same type over
A, with x of size < , then there is an f Aut(/[A) such that fa = b.
For X M
x
and f Aut(/) we dene f(X) := fa : a X M
x
,
and we note that if X Def
x
(/[A), then f(X) Def
x
(/[fA), where
fA =
_
f
s
(A
s
)
_
. Thus we have a group action
(f, X) f(X) : Aut(/[A) Def
x
(/[A) Def
x
(/[A).
For an automorphism f of / and a map : X M
y
with X M
x
we
dene the map f() : f(X) M
y
by f()(fa) = f((a)) for a X. In
other words, f(graph()) = graph(f()).
Automorphisms also act on types: for f Aut(/) and p St
x
(A) we
dene f(p) St
x
(fA) by
f(p) := (x, fa) : (x, y) an L-formula, a A
y
, (x, a) p.
If we identify these types p St
x
(A) and f(p) St
x
(fA) with the corre-
sponding ultralters of the boolean algebras Def(/[A) and Def(/[fA),
this means f(p) = f(X) : X p. In particular, we have a group action
(f, p) f(p) : Aut(/[A) St
x
(A) St
x
(A).
Expansions and Reducts. An extension of L = (S, L
r
, L
f
) is a language
L

= (S

, L
r
, L
f
) such that S S

, L
r
L
r
, and L
f
L
f
. In that case, an
L

-structure /

=
_
(M

s
)
sS
;
_
is said to be an L

-expansion of / (and
/ is said to be the L-reduct of /

) if M
s
= M
s
for all s S, and each
nonlogical symbol of L has the same interpretation in / as in /

. Given
such an L

-expansion /

of / we sometimes abuse notation by letting a

parameter set A = (A
s
) in /denote also the parameter set (A
s
)
sS
in /

,
where A
s
= for s S

S.
Exercise. Let L
1
and L
2
be languages with disjoint sets of sorts S
1
and
S
2
, and let L be the disjoint union of L
1
and L
2
, dened in the obvious way,
in particular, with S = S
1
S
2
as set of sorts. Let /
1
= (M
1
; . . . ) be an
L
1
-structure and /
2
= (M
2
; . . . ) an L
2
-structure. Let / = (M; . . . ) be
the L-structure such that
M
s
= M
1
s
for s S
1
, and M
s
= M
2
s
for s S
2
;
each nonlogical symbol of L
1
(respectively, L
2
) has the same inter-
pretation in / as in /
1
(respectively, as in /
2
).
17
(So / is like a disjoint sum of /
1
and /
2
.) Let x be a multivariable of
L
1
and y a multivariable of L
2
(so x and y are disjoint multivariables of L),
and let Z M
x,y
. Show: Z is 0-denable in / i Z is a nite union of
cartesian products X Y where X M
x
is 0-denable in /
1
and Y M
y
is 0-denable in /
2
. (When y is the empty multivariable of L
2
, this says
that a subset of M
x
is 0-denable in / i it is 0-denable in /
1
.)
4. Imaginaries
Let E be a 0-denable equivalence relation on M
s
, where s = (s
1
, . . . , s
n
)
S
n
, that is, E is an equivalence relation on M
s
, and is 0-denable as a subset
of M
s
M
s
. It is often useful to treat the equivalence classes of E on an
equal footing with the elements of /. This can be done as follows. Let
f : M
s
M
E
be a surjective map onto a set M
E
such that for all a, b M
s
,
f(a) = f(b) aEb.
(For example, take M
E
= M
s
/E and f(a) = E(a), the E-equivalence class
of a M
s
.) Extend L by a new sort s

/ S and a new function symbol

f

of arity (s
1
, . . . , s
n
, s

) to give the language L

, and expand / to an L

-
structure /

=
_
(M

s
)
sS
; . . .
_
by setting M

s
= M
E
and interpreting f

in
/

as the function f.
For our purposes this is an innocuous way of expanding /, and to explain
this, let x = (x
1
, . . . , x
n
) and y = (y
1
, . . . , y
n
) be disjoint with x
i
and y
i
quantiable of sort s
i
for i = 1, . . . , n, and let (x, y) be an L-formula that
denes E M
s
M
s
in /.
Let T := Th(/) and let T

be the L

-theory axiomatized by T together

with the L

-sentence
vx
_
v = f

(x)
_
xy
_
(x, y) f

(x) = f

(y)
_
,
where v is a quantiable variable of sort s

.
Lemma 4.1. Let ^ be a model of T. Then
(1) ^ has an expansion to a model ^

of T

;
(2) if ^

is an expansion of ^ to a model of T

, then any isomorphism

/ ^ expands uniquely to an isomorphism /

;
(3) for any L

-formula (u, v
1
, . . . , v
k
) with u a multivariable of L and
each variable v
j
of sort s

, there is an L-formula

(u, v
1
, . . . , v
k
)
such that T

[=
_
u, f

(v
1
), . . . , f

(v
k
)
_

(u, v
1
, . . . , v
k
) where
v
j
= (v
j
1
, . . . , v
j
n
) and v
j
i
is of sort s
i
for i = 1, . . . , n and j = 1, . . . k.
(4) if ^

is an expansion of ^ to a model of T

, then any elementary

embedding / ^ expands uniquely to an elementary embedding
/

.
Items (1) and (2) are almost obvious, (3) follows by an easy induction on
formulas, and (4) is a consequence of (3). As a special case of (3), each
L

-sentence is T

-equivalent to an L-sentence. Thus T

is complete. Another
consequence of (3) is that /

induces no new structure on /: for any

18
parameter set A in /, a set X M
x
is A-denable in /i it is A-denable
in /

.
Expanding / to /
eq
. We now carry out the above for all 0-denable
equivalence relations simultaneously. For each s = (s
1
, . . . , s
n
) S
n
and
0-denable equivalence relation E on M
s
, we pick an L-formula = (x, y)
with quantiable x, y that denes E in /, and introduce a new sort s,
put M
s
:= M
s
/E, and let f
E
: M
s
M
s
be the function that assigns to
each a M
s
its E-equivalence class f
E
(a) = E(a). Let L
eq
be the language
obtained from L by adding, for each E and corresponding as above, s as
a new sort and a function symbol f

of arity (s
1
, . . . , s
n
, s). Let S
eq
S
be the set of sorts of L
eq
. We then expand / to the L
eq
-structure
/
eq
:=
_
(M
s
)
sS
eq;
_
by interpreting each new function symbol f

in /
eq
as the function f
E
dened above. For each E and corresponding as above and v a quantiable
variable of sort s we call the sentence
vx
_
v = f

(x)
_
xy
_
(x, y) f

(x) = f

(y)
_
the dening axiom of f

. We let T
eq
be the L
eq
-theory axiomatized by T
and the dening axioms for the new function symbols. Thus /
eq
[= T
eq
,
but L
eq
and T
eq
depend really only on the theory T of /, rather than on
the model / of T. In particular, every model ^ of T expands likewise to
an L
eq
-structure ^
eq
[= T
eq
.
Lemma 4.2. Any expansion of / to a model of T
eq
is isomorphic to /
eq
via a unique isomorphism that is the identity on /. If ^ is a model of
T, then any isomorphism / ^ expands uniquely to an isomorphism
/
eq
^
eq
.
In particular, each automorphism f of / expands uniquely to an automor-
phism of /
eq
, which we also denote by f. In this way the group Aut(/)
gets identied with the group Aut(/
eq
).
Lemma 4.3. Let (u, v
1
, . . . , v
k
) be an L
eq
-formula with u a multivariable
of L and each v
j
a variable of sort s
j
S
eq
S corresponding to a 0-
denable equivalence relation E(j) on M
s(j)
, s(j) = (s
1
(j), . . . , s
n(j)
(j))
S
n(j)
. Then there is an L-formula
eq
(u, v
1
, . . . , v
k
) such that
T
eq
[=
_
u, f

1
(v
1
), . . . , f

k
(v
k
)
_

eq
(u, v
1
, . . . , v
k
)
where v
j
= (v
j
1
, . . . , v
j
n(j)
) and v
j
i
is of sort s
i
(j) for j = 1, . . . k and i =
1, . . . , n(j).
In particular, T
eq
is complete. The lemma also yields that if ^ is a model
of T, then any elementary embedding / ^ expands uniquely to an
elementary embedding /
eq
^
eq
. One more consequence of the lemma is
that /
eq
induces no new structure on /: for any parameter set A in /,
a set X M
x
is A-denable in / i it is A-denable in /
eq
.
19
Elimination of Imaginaries and Coding Denable Sets. First some
terminology in connection with an arbitrary map g : P Q. The kernel of
g is the equivalence relation E
g
on its domain P dened by
aE
g
b g(a) = g(b).
More generally, given an equivalence relation E on its codomain Q, the pull
back of E by g is the equivalence relation g
1
E on P dened by
a(g
1
E)b g(a)Eg(b).
We say that T has elimination of imaginaries (EI, for short) if each 0-
denable equivalence relation E on a set M
x
with nite x is the kernel of
some 0-denable map f : M
x
M
s
, s = s
1
, . . . , s
n
.
While this denition refers to the particular model / of T, any model of
T instead of / would give the same notion of EI. We say that / has EI
if its theory T has EI. If / has EI, then the eq-construction is superuous
(for most purposes), by the following lemma.
Lemma 4.4. Suppose / has EI. Let A be a parameter set in / and X an
A-denable set in /
eq
. Then there is an A-denable bijection : X X
between X and an A-denable set X in /.
Proof. Since X is given as a subset of a cartesian product M
E(1)

M
E(k)
where E(1), . . . , E(k) are 0-denable equivalence relations on the sets
M
s(1)
, . . . , M
s(k)
, respectively, with s(i) S
n(i)
for i = 1, . . . , k, it is enough
to establish the following. Let E be a 0-denable equivalence relations on
M
s
with s S
n
; then there is a 0-denable bijection
E
: M
E
Y onto a
0-denable set Y in /. Because / has EI, we have a 0-denable f : M
s

M
x
with nite x such that the kernel of f is E; setting Y := f(M
s
) M
x
this yields a 0-denable bijection f
E
(a) f(a) : M
E
Y as desired.
Exercise. Show that if / has EI and A is a parameter set, then / as an
L
A
-structure also has EI.
Lemma 4.5. T
eq
has EI.
Proof. Let E(1), . . . , E(k) be 0-denable equivalence relations on the sets
M
s(1)
, . . . , M
s(k)
, with s(i) S
n(i)
for i = 1, . . . , k. Let E be an equivalence
relation on M
s
M
E(1)
M
E(k)
that is 0-denable in /
eq
. It suces
to nd an s S
N
with N N, an equivalence relation E

on M
s
that is
0-denable in /, and a map
h : M
s
M
E(1)
M
E(k)
M
E

that is 0-denable in /
eq
and has E as kernel. Put
s := (s, s(1), . . . , s(k)) S
n+n(1)++n(k)
, so
M
s
= M
s
M
s(1)
M
s(k)
.
20
Then the map g : M
s
M
s
M
E(1)
M
E(k)
given by
g(a, b
1
, . . . , b
k
) := (a, E(1)(b
1
), . . . , E(k)(b
k
))
is 0-denable in /
eq
, and the pull back of the equivalence relation E by g
is an equivalence relation E

on M
s
that is 0-denable in /. The map h
with the domain and codomain indicated above and given by
h(a, E(1)(b
1
), . . . , E(k)(b
k
)) = E

(a, b
1
, . . . , b
k
), (a, b
1
, . . . , b
k
) M
s
,
is easily seen to have the desired property.
Given nite x, y, a tuple a M
x
is said to code the set Y M
y
(in /) if
there is a 0-denable Z M
x
M
y
such that Y = Z(a) and Y ,= Z(b) for
all b ,= a in M
x
. The reader should check that then any nite tuple c M
z
that is interdenable with a (in /) also codes Y .
Lemma 4.6. If T has EI, then every denable set in / has a code.
Proof. Let x and y be nite and Y = Z(a) M
y
where Z M
x
M
y
is
0-denable. The equivalence relation E on M
x
dened by
aEb Z(a) = Z(b),
is 0-denable. Hence, if T has EI, then E is the kernel of a 0-denable
g : M
x
M
s
, s = (s
1
, . . . , s
n
), and then g(a) codes Y .
The converse of this lemma is also true under some natural assumptions.
One of these assumptions is that T has a home sort. By a home sort of
T we mean a sort s
0
S such that for every s S there is an n and a
surjective 0-denable map M
n
s
0
M
s
. While this denition refers to the
particular model / of T, any model of T instead of /would give the same
notion of homesort. In many cases, a many-sorted structure / is either
one-sorted, or arises from a one-sorted structure by adding a few carefully
chosen quotient sets, and then a homesort comes for free.
Lemma 4.7. Let T have home sort s
0
with two distinct 0-denable elements
in M
s
0
. Then, given any 0-denable sets P and Q in /, there is an s S
n
and there are 0-denable injective maps i : P M
s
and j : Q M
s
with
disjoint i(P) and j(Q).
Proof. Let P M
x
and Q M
y
be 0-denable with nite x and y. The
assumption on T yields 0-denable elements p
0
M
x
, q
0
M
y
, and 0, 1
M
s
0
with 0 ,= 1. Then dene i : P M
x
M
y
M
s
0
and j : Q
M
x
M
y
M
s
0
by i(p) = (p, q
0
, 0) and j(q) = (p
0
, q, 1).
Lemma 4.8. Suppose T has a home sort s
0
with two distinct 0-denable
elements in M
s
0
. Assume also that / is
0
-saturated and every denable
set in / has a code. Then T has EI.
21
Proof. Let x be nite and E a 0-denable equivalence relation on M
x
. For
each equivalence class E(a) there is a 0-denable Z M
s(a)
M
x
with
s(a) S
n(a)
such that E(a) = Z(c) for exactly one c M
s(a)
(so this c
codes E(a)). A standard saturation argument shows that as a varies over
M
x
we can choose such s(a) and Z from xed nite collections, with the code
c a piecewise 0-denable function of a. More precisely, there are 0-denable
X
1
, . . . , X
k
M
x
that cover M
x
, tuples s(1) S
n(1)
, . . . , s(k) S
n(k)
, and
0-denable sets
Z
1
M
s(1)
M
x
, . . . , Z
k
M
s(k)
M
x
with for each i 1, . . . , k a 0-denable function c
i
: X
i
M
s(i)
, such that
if a X
i
, then E(a) = Z
i
(c
i
(a)) and E(a) ,= Z
i
(c) for all c ,= c
i
(a) in M
s(i)
.
To keep notations simple we assume k = 2 in what follows. First we increase
X
1
to the 0-denable set X

1
M
x
consisting of the elements of M
x
that are
E-equivalent to some element of X
1
, and extend c
1
to a 0-denable function
c

1
: X

1
M
s(1)
by c

1
(a) := c
1
(b) if b X
1
and aEb. We then decrease X
2
to X

2
= M
x
X

1
and let c

2
: X

2
M
s(2)
be the restriction of c
2
to X

2
.
Up to this point we have not used the home sort assumption. This comes
into play in producing from c

1
and c

2
a single 0-denable function
g : M
x
M
s
, s S
n
,
such that E is the kernel of g. The previous lemma gives an s S
n
and
0-denable injective maps i : M
s(1)
M
s
and j : M
s(2)
M
s
with disjoint
images. Now dene g : M
x
M
s
by g(a) := i
_
c

1
(a)
_
if a X

1
and
g(a) := j
_
c

2
(a)
_
if a X

2
. Then c has the desired property.
Exercise. Let x, y be nite, and suppose a M
x
codes an A-denable set
Y M
y
. Show that the tuple a is A-denable.
5. The Monster Model
From now on we x a complete L-theory T. By the completeness of T,
any two models of T can be elementarily embedded into a third model of T.
More generally, any commutative diagram of models of T and elementary
embeddings between them can be realized inside a single model of T, where
the embeddings become inclusions among elementary submodels. (Try to
formulate this precisely, and prove it!) For our purpose there is indeed no
serious loss of generality to have all action take place in a single big model
of T, and there are advantages in doing so. For example, we can in this way
import Galois-theoretic ideas.
We refer to basic texts in model theory for the fact that for any there
is a model of T that is -saturated and strongly -homogeneous.
In what follows we x a model M=
_
(M
s
);
_
[= T and a cardinal (M) >
[L[ such that M is (M)-saturated and strongly (M)-homogeneous. We
also call M the monster model of T. In particular, every model of T of size
(M) has an elementary embedding into M.
22
The following conventions are in force, unless specied otherwise. Notions
of denable and algebraic for sets and tuples will be relative to M, in
the same way they were relative to / in earlier sections. Small will mean
of size < (M). Multivariables are assumed to be small and in L, and
A, A

, B denote small parameter sets in M. An elementary submodel of M

is completely determined by its underlying family of sets, and this family is
also a parameter set in M. We let M, M

and N denote small parameter

sets underlying an elementary submodel of M, and we denote this elementary
submodel also by M, M

and N; usually we signal this by phrasing like the

model M. On the other hand, if we use the more elaborate notation /,
then / can be any structure, not necessarily a small elementary submodel
of M, and in that context M will denote the underlying family of sets of
/, as before. Likewise with ^ and N. (We hope this does not confuse the
reader.)
Thus a partial x-type over A is a small set of formulas according to these
conventions. We shall write [= (x) to indicate that (x) is an L
M
-formula
and M [= (x). Likewise, (x) [= (x) will mean that (x) is a small set
of L
M
-formulas (x), and (x) is an L
M
-formula such that every a M
x
realizing (x) also realizes (x). Note that in that case there is a nite
subset
0
(x) of (x) such that
0
(x) [= (x). (This is how compactness is
built into the monster model via saturation.)
Note that M
eq
is also (M)-saturated and strongly (M)-homogeneous,
and (M) > [L
eq
[ = [L[, so M
eq
can serve as the corresponding monster
model of T
eq
. Likewise, given any A, our M remains (M)-saturated and
strongly (M)-homogeneous as an L
A
-structure, so M can serve as monster
model of T
A
, the theory of the L
A
-structure M.
Relative denability and automorphisms. Given a partial type (x)
over A, its set (M
x
) of realizations is a nonempty subset of M
x
. Note that
if p(x) is a type over A realized by a M
x
, then
p(M
x
) = fa : f Aut(M[A)
that is, p(M
x
) is the orbit of a under the action of Aut(M[A) on M
x
.
Lemma 5.1. Given A and a nite tuple a, we have:
(1) a is A-denable if and only if fa = a for each f Aut(M[A);
(2) a is A-algebraic if and only if fa : f Aut(M[A) is nite;
(3) a is A-algebraic if and only if fa : f Aut(M[A) is small.
Proof. Let a M
x
, and p(x) := tp(a[A). Suppose fa = a for all f
Aut(M[A). Then p(x) has a as its only realization, so with y disjoint from
and similar to x, we have p(x) p(y) [= x = y, so there is a (x) p(x)
such that (x) (y) [= x = y, so (M
x
) = a. The other direction of (1)
is obvious.
Next, suppose that fa : f Aut(M[A) is nite, say of size m. Then
p(x) p(y
1
) p(y
m
) [= x = y
1
x = y
2
x = y
m
,
23
where x, y
1
, . . . , y
m
are pairwise disjoint and similar. Hence there is a (x)
p(x) such that
(x) (y
1
) (y
m
) [= x = y
1
x = y
m
,
so (M
x
) is nite, A-denable, and a (M
x
). The other direction of (2)
is obvious. We leave (3) as an exercise.
By coding, this lemma yields an analogue for denable sets (instead of nite
tuples). To see this, suppose a M
x
codes the set Y M
y
, so we have a
0-denable Z M
x
M
y
such that Y = Z(a) and Y ,= Z(b) for all b ,= a
in M
x
. For f Aut(M) we have f(Y ) = Z(fa) with f(Y ) ,= Z(b) for
all b ,= fa in M
x
, so fa codes f(Y ), in particular, fa = a if and only if
f(Y ) = Y . A denable set Y M
y
with nite y is said to be A-algebraic if
there is a nite A-denable equivalence relation E on M
y
such that Y is a
union of E-classes.
Corollary 5.2. Given A and denable Y M
y
with nite y, we have:
(1) Y is A-denable if and only if f(Y ) = Y for each f Aut(M[A);
(2) Y is A-algebraic if and only if f(Y ) : f Aut(M[A) is nite;
(3) Y is A-algebraic if and only if f(Y ) : f Aut(M[A) is small.
Proof. Assume without loss of generality (why?) that T has EI. Then (1)
follows from part (1) of Lemma 5.1 by using a code of Y . In (2), the if
direction requires a little argument: Assume f(Y ) : f Aut(M[A) is
nite. Then this nite set generates a nite boolean algebra F of subsets of
M
y
. The sets in F are denable subsets of M
y
, so the equivalence relation
E on M
y
whose equivalence classes are the atoms of F is denable. Since
each f Aut(M[A) permutes these atoms, we have f(E) = E for such f,
so E is A-denable by (1). Now use that Y is a union of atoms of F.
We leave (3) to the reader.
Corollary 5.3. Suppose T has EI and let Y M
y
, with nite y. Then Y
is A-algebraic if and only if Y is acl(A)-denable.
Proof. Take a code a for Y . Then for all f Aut(M[A) we have the equiv-
alence f(Y ) = g(Y ) f(a) = g(a). By our earlier criteria,
Y is A-algebraic a is A-algebraic a is acl(A)-denable
Y is acl(A)-denable,
which contains the desired equivalence.
The next characterization of acl(A) plays no role in this course, but its proof
nicely illustrates how to use automorphisms in the monster model:
acl(A) =

MA
M.
To prove this, let a M
s
and a / acl(A)
s
. It suces to show that then
there is a model M A such that a / M
s
. We take any model N A,
24
and observe that by Lemma 5.1, (3), there is an f Aut(M[A) such that
fa / N. Then M := f
1
(N) has the desired property.
Strong types. For a parameter set C in M
eq
, let dcl
eq
(C) and acl
eq
(C)
denote its denable and algebraic closure in M
eq
. The strong type stp(a[A)
of a nite tuple a M
x
over A, is the x-type of a over acl
eq
(A) in M
eq
:
stp(a[A) := tp
_
a[ acl
eq
(A)
_
.
Corollary 5.4. Given A, and nite tuples a, b M
x
, we have: stp(a[A) =
stp(b[A) if and only if E(a) = E(b) for each nite A-denable equivalence
relation E on M
x
.
Proof. By the above it suces to prove the following claim: Let E be a nite
A-denable equivalence relation E on M
x
, and let f Aut
_
M
eq
[ acl
eq
(A)
_
.
Then E(fa) = E(a).
To prove this claim, take a 0-denable Z M
eq
s
M
x
(s (S
eq
)
n
) such
that E(a) = Z(c) for a unique c M
eq
s
. It follows that for this c its orbit
gc : g Aut(M
eq
[A) is nite, so c is algebraic over A in M
eq
, hence fc = c,
and thus E(fa) = E(a).
Exercises. Let Y M
y
with nite y be denable and let c be a nite tuple.
Show: c codes Y if and only if for all f Aut(M), f(Y ) = Y f(c) = c.
(This equivalence is used frequently.)
Let h : P Q be 0-denable and surjective, with P M
x
and Q M
y
and
nite x, y, let Y Q, and let c be a nite tuple. Show that c codes Y M
y
if and only if c codes h
1
(Y ) M
x
.
Show that if c and d are nite tuples coding the nonempty sets X M
x
and Y M
y
, then (c, d) codes X Y M
x,y
.
Let g : X M
y
and h : X M
z
be denable, with X M
x
and nite
x, y, z. Show: g and h have codes if and only if (g, h) : X M
y,z
has a
code.
Application to EI. The next lemma basically shows that coding denable
relations can be reduced to coding unary denable functions. (This may
be reminiscent of the fact that proving QE reduces to eliminating a single
existential quantier in front of a conjunction of atoms and negated atoms.)
In the proof we use the results stated in the exercises above.
Lemma 5.5. Suppose T has a home sort s
0
, every denable subset of M
s
0
has a code, and every denable partial function M
s
0
M
s
, s S, has a
code. Then every denable set in M has a code.
Proof. Assume as an inductive hypothesis that for a certain n > 0 all de-
nable subsets of M
n
s
0
have codes. Let Y M
1+n
s
0
be denable. We shall
prove that Y has a code. For each p M
s
0
the section Y (p) M
n
s
0
has a
25
code. Hence, by saturation, there are tuples s(1) S
m(1)
, . . . , s(k) S
m(k)
and 0-denable relations
Z
1
M
s(1)
M
n
s
0
, . . . , Z
k
M
s(k)
M
n
s
0
such that for each p M
s
0
there is an i 1, . . . , k for which there is
exactly one c M
s(i)
with Y (p) = Z
i
(c). Dene X
1
, . . . , X
k
M
s
0
by
X
i
:= p M
s
0
: Y (p) = Z
i
(c) for exactly one c M
s(i)
,
so we have a denable map
i
: X
i
M
s(i)
such that if p X
i
, then
Y (p) = Z
i
_

i
(p)
_
and Y (p) ,= Z
i
(c) for all c M
s(i)
with c ,=
i
(p). Note
that X
1
, . . . , X
k
cover M
s
0
. Let now f Aut(M), and identify
i
with its
graph, a subset of M
s
0
s(i)
.
Claim. f(Y ) = Y f(X
i
) = X
i
and f(
i
) =
i
for i = 1, . . . , k.
To prove this claim, let q M
s
0
and put p := f
1
(q), so q = f(p). Then
q f(X
i
) p X
i

!
c
_
Y (p) = Z
i
(c)
_

!
c
_
f(Y (p)) = Z
i
(fc)
_

!
d
_
f(Y )(q) = Z
i
(d)
_
for i = 1, . . . , k, with c and d ranging over M
s(i)
. Note also that if q f(X
i
),
then f(Y )(q) = Z
i
(f(
i
)(q)). The claim now follows easily.
By the assumption of the lemma, the sets X
1
, . . . , X
k
M
s
0
have codes
a
1
, . . . , a
k
. This assumption and one of the exercises above also yield that
the maps
1
, . . . ,
k
have codes b
1
, . . . , b
k
. By the claim and one of the
exercises above it follows that then Y has code (a
1
, . . . , a
k
, b
1
, . . . , b
k
).
We have now shown that for every n, every denable subset of M
n
s
0
has a
code. An earlier exercise then yields the conclusion of the lemma.
Corollary 5.6. Suppose T has a home sort s
0
with two distinct 0-denable
elements in M
s
0
, and that every denable partial function M
s
0
M
s
, s S,
has a code. Then T has EI.
Decomposing a type space. We nish this section with a fact on type
spaces. Let u and v be disjoint (small) multivariables. We wish to describe
u, v-types in terms of u-types and v-types. In detail: Each u, v-type p(u, v)
over A yields a v-type p
u
(v) over A, such that if (a, b) M
u
M
v
realizes
p(u, v), then b realizes p
u
(v). (This is clear if one thinks of types as orbits.)
Choose for each type q = q(v) St
v
(A) a realization b
q
M
v
. Then we can
dene a map
St
u,v
(A)
_
qStv(A)
St
u
(Ab
q
) (disjoint union)
as follows: given p = p(u, v) St
u,v
(A), pick a realization (a, b) such that
b = b
q
for q = p
u
(v), and assign to p the element tp(a[Ab) St
u
(Ab).
26
Lemma 5.7. The above map
St
u,v
(A)
_
qStv(A)
St
u
(Ab
q
)
is injective.
This injectivity is easily veried, and is useful later in bounding the size of
the type space St
u,v
(A) in terms of the size of simpler type spaces.
6. Definable Types
Denability of types is a consequence of stability, as we shall see in the
next section. Denable types, however, also occur signicantly in unstable
settings. For example, all types over the eld of real numbers are denable,
and this fact has interesting consequences for limit sets of semialgebraic
families. Therefore it makes good sense to introduce denable types before
discussing stability.
Throughout this section, x and y are small multivariables, although there
would be no loss of generality in assuming y to be nite.
Let (x, y) be an L-formula. Below we abuse notation by having stand for
(x, y). By a -instance over B we mean a formula (x, b) with b B
y
, and
by a -formula over B we mean an L
B
-formula (x) that is equivalent to a
boolean combination of -instances over B. The subsets of M
x
dened by
-formulas over B form a boolean subalgebra Def

(M[B) of Def
x
(M[B). By
a partial -type over B we mean a partial x-type (x) over B consisting of
-formulas over B. By a -type over B we mean a partial -type p(x) over
B such that for each -formula (x) over B, either (x) p or (x) p.
Note that a -type p(x) over B is uniquely determined by its subset
(x, b) p : b B
y

of -instances. Let St

(B) be the set of -types over B. Then we have a

bijection
p(x) (M
x
) : (x) p(x) : St

(B) St
_
Def

(M[B)
_
of St

(B) with an actual Stone space. Given a M

x
we let tp

(a[B) be the
-type over B realized by a.
Let p(x) be a -type over B and (y) an L
B
-formula. We say that (y)
denes p(x) if for all b B
y
,
(x, b) p(x) [= (b).
Note that (y) can dene at most one -type over B. We say that p(x) is
denable if some L
B
-formula (y) denes p(x). The following is immediate
from the denitions:
Lemma 6.1. Let a M
x
and let M be a model. Then the set (a, M
y
) M
y
is denable in M if and only if tp

(a[M) is denable.
27
Note that (a, M
y
) is dened with a parameter a possibly outside M!
Let p(x) be again a -type over B. When A B we say that p(x) is denable
over A if some L
A
-formula (y) denes p(x), that is, for all b B
y
,
(x, b) p(x) [= (b).
When B is a model M, then an L
M
-formula (y) that denes p(x) is clearly
determined up to equivalence in the L
M
-structure M by p and (x, y), and
we let d
p
x(x, y) denote such a formula (y); one may think of d
p
x as for
x realizing p and should view the x in d
p
x(x, y) as a bound multivariable
and y as free.
Consider now a type p(x) St
x
(B). Given an L-formula (x, y) we dene
p(x) := (x) p(x) : (x) is a -formula over B,
which is a -type over B. Note that p(x) is the union of the p(x) with
(x, y) ranging over the L-formulas of the indicated form. A dening scheme
for p(x) is a map assigning to each L-formula (x, y) an L
B
-formula (y)
that denes p(x), that is,
(B
y
) = b B
y
: (x, b) p(x).
We say that p(x) is denable if it has a dening scheme, in other words,
p(x) is denable if and only if p(x) is denable for each L-formula (x, y).
When B is a model M, then an L
M
-formula (y) that denes p is
determined up to equivalence in the L
M
-structure M by p and (x, y), and
we let d
p
x(x, y) denote such a formula (y).
Lemma 6.2. Let x be nite. Suppose that each x-type over B is denable
and [B[ [L[. Then
[ St
x
(B)[ [B[
|L|
.
Proof. Use that each x-types over B has a dening scheme.
This upperbound [B[
|L|
is generally much smaller than the trivial upper
bound 2
|B|
for [ St
x
(B)[ when [B[ [L[.
Let p(x) St
x
(B), and let A B. Then we say that p is denable over A
if for each L-formula (x, y) there is an L
A
-formula (y) such that
(x, b) p(x) [= (b), for all b B
y
,
in other words, for each L-formula (x, y) the -type p(x) is denable
over A. The following is immediate from the denitions
Lemma 6.3. If p(x) St
x
(B) is denable over A B, and f Aut(M[A)
satises f(B) = B, then f(p) = p.
Extending denable types. In this subsection we x M and B M.
28
Let (x, y) be an L-formula and p(x) a denable -type over M. We extend
p to a denable -type q = p B in St

(B) as follows. Pick an L

M
-formula
(y) dening p, and declare
(x, b) q [= (b), for all b B
y
.
Let us check that this declaration determines indeed a -type q over B. Let
b
1
, . . . , b
n
B
y
be such that
[= (b
1
) (b
m
) (b
m+1
) (b
n
).
It is enough to show that then
[= x
_
(x, b
1
) (x, b
m
) (x, b
m+1
) (x, b
n
)
_
.
Take disjoint y
1
, . . . , y
n
similar to y and disjoint from x, y. Then
M [=

(y) x

(x, y), where y = (y

1
, . . . , y
n
),

(y) := (y
1
) (y
m
) (y
m+1
) (y
n
),

(x, y) := (x, y
1
) (x, y
m
) (x, y
m+1
) (x, y
n
).
Hence [=

(y) x

(x, y), as desired. Thus the equivalence above does

determine a -type q over B. Moreover, q M = p, and any L
M
-formula
dening p also denes q.
The considerations above can easily be adapted to denable types p
St
x
(M): each such p extends to a denable type q = p B in St
x
(B) by
requiring that any dening scheme for p is also a dening scheme for q, in
other words,
q = (p)B for each L-formula (x, y).
Lemma 6.4. Let p St
x
(M) be denable. Then p has a unique extension
to a type q St
x
(B) that is denable over M.
Proof. The type q := p B extends p and is denable over M. Suppose
q

St
x
(B) also extends p and is denable over M. Given an L-formula
(x, y), let (y) and

(y) be L
M
-formulas that dene q and q

. Then
(M
y
) =

(M
y
), since q and q

have the same restriction p to M. Hence

(M
y
) =

(M
y
), so q and q

coincide. As is arbitrary, we conclude

q = q

.
Global types and canonical bases. Let St
x
(M) be the set of x-types
over the parameter set (M
s
) in M. (This parameter set is in general not
small!) A type in St
x
(M) is also called a global x-type, and, depending on
the context, is either viewed as a set of formulas, or as the corresponding
ultralter of the boolean algebra Def
x
(M). A global x-type p is said to be
denable if for each L-formula (x, y) there is an L
M
-formula (y) such that
for all b M
y
,
(x, b) p [= (b).
29
If in addition we can always choose such (y) to be an L
A
-formula, then p
is said to be denable over A. For denable p St
x
(M) and any L-formula
(x, y) we let d
p
x(x, y) denote an L
M
-formula (y) as above.
Exercise. Suppose the (global) type p St
x
(M) is denable. Show that p
is denable over some A. Show that for any A,
p is denable over A f(p) = p for all f Aut(M[A).
Denition. Let p St
x
(M). A canonical base for p is a parameter set A
such that for all f Aut(M) we have
f(p) = p f Aut(M[A).
If A is a canonical base for p, then B is a canonical base for p if and only if
dcl(A) = dcl(B). If p has a canonical base, then it has exactly one denably
closed canonical base, which shall be referred to as the canonical base of p
and denoted by cb(p).
Lemma 6.5. Suppose T has EI and the global type p St
x
(M) is denable.
Then p has a canonical base, and cb(p) is the smallest denably closed
parameter set over which p is denable.
Proof. Here we restrict to nite y. For each L-formula (x, y) , let c

code
the denable set d
p
x(x, M
y
). Then we have for f Aut(M):
f(p) = p f(c

) = c

for all L-formulas (x, y).

Therefore, if A consists of the components of all the c

, then A is a canonical
base of p. The claim about cb(p) now follows from the exercise at the end
of Section 4 and the exercise in this subsection.
Heirs. Let p = p(x) St
x
(M). Given M B, a heir of p over B is a
son q St
x
(B) of p such that for each L
M
-formula (x, y), if (x, b) q for
some b B
y
, then (x, a) p for some a M
y
.
Let M B C, and suppose r St
x
(C) is a son of q St
x
(B) and q is
a son of p. Then it follows easily that if r is a heir of p, then q is a heir of
p, and also that if q is a heir of p and B = N is an elementary submodel of
M and r is a heir of q, then r is a heir of p.
Lemma 6.6. Let p St
x
(M), and M B C. Then
(1) p has a heir over B;
(2) if q is a heir of p over B, then q extends to a heir of p over C;
(3) if p is denable, then p has exactly one heir over B, namely pB.
Proof. Since (1) is a special case of (2) we prove (2). Let q be a heir of p
over B. Claim 1: let (x, y) be an L
M
-formula and let c C
y
be such that
q(x) [= (x, c); then (x, a) p for some a M
y
. To see this, note rst
that q(x) tp
y
(c[B) [= (x, y), since for any realization (d, e) M
x
M
y
of this partial type over B in M there is an f Aut(M[B) with f(e) = c.
30
Compactness yields an L
M
-formula (y, z) and a tuple b B
z
such that
(y, b) tp
y
(c[B) and q(x) (y, b) [= (x, y). Hence
q(x) [=
_
y(y, b)
_
y
_
(y, b) (x, y)
_
.
Since q is a heir of p this gives b

M
z
such that
p(x) [=
_
y(y, b

)
_
y
_
(y, b

) (x, y)
_
.
Hence M [= y(y, b

), so we have a M
y
with [= (a, b

), and thus (x, a)

p(x). This proves Claim 1.
Let (x) be the set of L
C
-formulas (x, c) where (x, y) is an L
M
-formula
such that (x, a) p(x) for every a /
y
, and c C
y
. Claim 2: q(x)(x)
is a partial x-type over C. If Claim 2 holds, then we can take r(x) St
x
(C)
containing this partial type, and then r is heir of p that extends q. Suppose
Claim 2 is false. This gives L
M
-formulas
1
(x, y), . . . ,
n
(x, y) and c C
y
such that
(1)
i
(x, a) p(x) for i = 1, . . . , n and all a /
y
;
(2) q(x)
1
(x, c), . . . ,
n
(x, c) cannot be realized over C.
Then q(x) [=
1
(x, c)
n
(x, c), and thus by Claim 1 there is an a M
y
such that
1
(x, a)
n
(x, a) p(x), a contradiction.
Coheirs. Let p St
x
(M). Given B M, a coheir of p over B is a son
q St
x
(B) of p such that every formula in q(x) is realized by some a M
x
.
Note that if a M
x
and b M
y
and tp(a[Mb) is a coheir of tp(a[M) over
Mb, then tp(b[Ma) is a heir of tp(b[M).
Let M B C, and suppose r St
x
(C) is a son of q St
x
(B) and q is
a son of p, and r is a coheir of p. Then it follows easily that q is a coheir of
p, and also that r is a coheir of q when B = N is an elementary submodel
of M.
Lemma 6.7. Let p St
x
(M), and M B C. Then
(1) p has a coheir over B;
(2) if q is a coheir of p over B, then q extends to a coheir of p over C;
Proof. It is enough to prove (2), since (1) is a special case. We shall use
a topological argument. Let q be a coheir of p over B. Let R be the
closure in St
x
(C) of tp(a[C) : a M
x
. For every (x) q the closed
set [(x)] St
x
(C) intersects tp(a[C) : a M
x
, and thus R. Hence by
compactness R has a point r(x) that lies in every [(x)] with (x) q(x).
Then r extends q and is a coheir of p.
Lemma 6.8. Let M B and suppose every y-type over M is denable, for
every nite y. Then every p St
x
(M) has a unique coheir over B.
Proof. Let p St
x
(M), and suppose that a, a

M
x
realize dierent co-
heirs q and q

of p over B. Take a nite y disjoint from x, b B

y
, and
an L
M
-formula (x, y) such that [= (a, b) (a

, b), and thus (a, y)

31
tp(b[Ma) and (a

, y) / tp(b[Ma

). But tp(b[Ma) and tp(b[Ma

) are heirs
of tp(b[M), which gives a contradiction using (3) of Lemma 11.16 and
tp(a[M) = tp(a

[M).
7. Indiscernibles and NIP
Let I be a chain, that is, a linearly ordered set. An I-sequence is just a
family (a
i
)
iI
, and is said to be an I-sequence in the set X if all a
i
X.
For I = N, ordered in the usual way, we just say sequence instead of
N-sequence. An I-sequence (a
i
) in M
x
, is said to be indiscernible over A
if for every L
A
-formula (x
1
, . . . , x
n
), with x
1
, . . . , x
n
similar to x, and all
i
1
< < i
n
and j
1
< . . . , j
n
in I we have
[= (a
i
1
, . . . , a
in
) (a
j
1
, . . . , a
jn
).
For A = we just say indiscernible.
Denable types and coheirs generate indiscernible sequences:
Lemma 7.1. Let q St
y
(M) be denable and a, b /
x
with tp(a[M) =
tp(b[M). Suppose c /
y
realizes q Ma and d /
y
realizes q Mb. Then
tp
_
a, c)[M
_
= tp
_
(b, d)[M
_
.
Proof. Let (u, x, y) be an L-formula, e M
u
. Then
[= (e, a, c, ) (e, a, y) q [= d
q
(e, a) [= d
q
(e, b)
(e, b, y) q [= (e, b, d).

Corollary 7.2. Let q St

y
(M) be denable, and take a sequence (a
n
) in
M
y
such that a
0
realizes q and a
n+1
realizes q Ma
0
. . . a
n
for each n. Then
(a
n
) is indiscernible over M, and for each m the type tp
_
(a
0
, . . . , a
m
)[M
_
is independent of the choice of the sequence (a
n
).
Proof. Let i
0
< < i
n
< i
n+1
in ^, and suppose inductively that
tp
_
(a
0
, . . . , a
n
)[M
_
= tp
_
(a
i
0
, . . . , a
in
)[M
_
.
Since a
n+1
realizes q Ma
0
. . . a
n
and a
i
n+1
realizes q Ma
i
0
. . . a
in
, it follows
from the previous lemma that
tp
_
(a
0
, . . . , a
n
, a
n+1
)[M) = tp
_
(a
i
0
, . . . , a
in
, a
i
n+1
)[M
_
.
The second claim of the lemma is proved in the same way.
A sequence (a
n
) as in this lemma is called a Morley sequence of q over M.
For M B, a type q St
y
(B) is said to be nitely satisable in M if
every formula (y) q is realized by some element of M
y
, in other words, q
is a coheir of its restriction to M.
Let M N, and assume in the next lemma and its corollary that N is
-saturated with > [M[.
32
Lemma 7.3. Let q St
y
(B) be nitely satisable in M and let a, b N
x
be such that tp(a[M) = tp(b[M). Suppose c N
y
realizes q[Ma and d N
y
realizes q[Mb. Then tp
_
(a, c)[M
_
= tp
_
(b, d)[M
_
.
Proof. Let (x, y) be an L
M
-formula such that [= (a, c), so (a, y) q;
it suces to show that then (b, y) q. Towards a contradiction, assume
that (b, y) q. Then (a, y) (b, y) q, so we have e M
y
such that
[= (a, e) (b, e), contradicting tp(a[M) = tp(b[M).
Corollary 7.4. Suppose q St
y
(B) is nitely satisable in M. Take a
sequence a
0
, a
1
, a
2
, . . . in N
y
such that a
0
realizes q[M and a
n+1
realizes
q[Ma
0
. . . a
n
for all n. Then the sequence (a
n
) is indiscernible over M, and
for each n the type tp
_
(a
0
, . . . , a
n
)[M
_
depends only on M and q, not on
the choice of the sequence (a
n
).
Proof. Let i
0
< i
1
< < i
n
< i
n+1
(in ^), and assume inductively that
tp
_
(a
0
, a
1
, . . . , a
n
)[M
_
= tp
_
(a
i
0
, a
i
1
, . . . , a
in
)[M
_
.
Now observe that a
n+1
realizes q[Ma
0
. . . a
n
and a
i
n+1
realizes q[Ma
i
0
. . . a
in
.
Then by the lemma above we have
tp
_
(a
0
, a
1
, . . . , a
n
, a
n+1
)[M
_
= tp
_
(a
i
0
, a
i
1
, . . . , a
in
, a
i
n+1
)[M
_
.
The second part of the lemma follows in the same way.
A sequence (a
n
) as in this lemma is called a coheir sequence of q over M.
Dependence and independence. In this section, x and y are disjoint,
and R M
x,y
= M
x
M
y
is a denable relation. We say that R, and any
L
M
-formula (x, y) that denes R, has the independence property (or IP) if
for every I N there are a
i
M
x
for i N and a b
I
M
y
such that
R(a
i
, b
I
) i I, for all i N.
Equivalently, for each n and I 1, . . . , n there are a
i
M
x
for i = 1, . . . , n
and a b
I
M
y
, such that
R(a
i
, b
I
) i I, for all i 1, . . . , n.
The ordered index set N in this denition is just for convenience: using the
nite version it is easy to see that in the denition of IP one can replace N
by any small innite chain.
Note also that in this denition we single out not only a particular product
set M
x
M
y
but also M
x
as its rst factor and M
y
as its second factor. In
this connection, recall that an L
M
-formula (x, y) is formally a triple , x, y;
let (y, x) be the triple , y, x, so if (x, y) denes R, then (y, x) denes
the reversed relation

R M
y
M
x
given by
R(a, b)

R(b, a), for a M
x
, b M
y
.
Our main interest here is when R is dependent (that is, R does not have
IP), in which case also any L
M
-formula (x, y) that denes it is said to be
dependent.
33
Lemma 7.5. Let (x, y) be a dependent L
M
-formula. Then
(1) the formula (y, x) is dependent;
(2) the formula (x, y) is dependent;
(3) if the L
M
-formula

(x, y

) is dependent with y

disjoint from x and

y, then (

)(x, (y, y

)) and (

)(x, (y, y

)) are dependent;
(4) if y = (u, v) and c M
v
, then
1
(x, u) := (x, u, c) is dependent.
8. Stability
In this section, x and y are disjoint and nite, and R M
x,y
= M
x
M
y
is a
denable relation. We say that R, and any L
M
-formula (x, y) that denes
R, is unstable if there are a
i
M
x
and b
i
M
y
for i N such that
R(a
i
, b
j
) i < j, for all i, j N.
Equivalently, for each n there are a
i
M
x
and b
i
M
y
for i = 1, . . . , n, such
that
R(a
i
, b
j
) i < j, for all i, j 1, . . . , n.
The ordered index set N in this denition is just for convenience: using
the nite version it is easy to see that in the denition of unstable one can
replace N by any small innite linearly ordered set.
Note also that in this denition we single out not only a particular product
set M
x
M
y
but also M
x
as its rst factor and M
y
as its second factor. In
this connection, recall that an L
M
-formula (x, y) is formally a triple , x, y;
let

(y, x) be the triple , y, x, so if (x, y) denes R, then

(y, x) denes the
reversed relation

R M
y
M
x
given by
R(a, b)

R(b, a), for a M
x
, b M
y
.
As the terminology suggests, our main interest here is when R is stable
(that is, not unstable), in which case also any L
M
-formula (x, y) that denes
it is said to be stable.
Lemma 8.1. Let (x, y) be a stable L
M
-formula. Then
(1) the formula

(y, x) is stable;
(2) the formula (x, y) is stable;
(3) if the L
M
-formula

(x, y

) is stable with y

nite and disjoint from

x and y, then (

)(x, (y, y

)) and (

)(x, (y, y

)) are stable;
(4) if y = (u, v) and c M
v
, then
1
(x, u) := (x, u, c) is stable.
To prove these items, consider the contrapositives. For example, in (3),
suppose (

)(x, (y, y

)) is unstable. Take a
i
M
x
, b
i
M
y
, b

i
M
y
for
i N, such that for all i, j N: [= (a
i
, b
j
)

(a
i
, b

j
) i < j. Let
P := (i, j) N
2
: i < j and [= (a
i
, b
j
),
P

:= (i, j) N
2
: i < j and [=

(a
i
, b

j
).
Then P P

= (i, j) N
2
: i < j, hence by Ramseys theorem there is an
innite subset I of N such that either all pairs (i, j) I
2
with i < j belong
34
to P, in which case (x, y) is unstable, or all pairs (i, j) I
2
with i < j
belong to P

, in which case

(x, y

) is unstable.
Exercise. Let f : M
x
M
x
and g : M
y
M
y
be denable, with disjoint
nite x

, y

, and dene the relation R

M
x
M
y
by
R

(a, b) R(fa, gb).

Show that if R is stable, then R

is stable. Show that if f and g are surjective

and R

is stable, then R is stable.

Let (x, y) be an L-formula, not just an L
M
-formula, and let /= (M; )
be any model of T. Then (x, y) is unstable if and only if for each n there
are a
i
M
x
and b
i
M
y
for i = 1, . . . , n, such that
/ [= (a
i
, b
j
) i < j, for all i, j 1, . . . , n.
It follows that the stability of (x, y) depends only on the theory T, not on
the particular monster model M of T used in the denition.
The next results show how unstability relates to linear orderability.
Lemma 8.2. Suppose R is unstable. Dene R

M
x,y
M
x,y
by
R

(a, b; a

, b

) R(a, b

), (a, a

M
x
, b, b

M
y
).
Then there are c
i
M
x,y
for i N such that for all i, j N,
R

(c
i
, c
j
) i < j.
Lemma 8.3. Let (x, y) be an unstable L-formula, and [L[. Then there
is a model / =
_
M;
_
of size such that St
M
x
(M) has size > .
Proof. First construct a linearly ordered set I of size with more than
upward closed subsets (see exercise below). With i and j ranging over I, we
use compactness to obtain a model / =
_
M;
_
with elements a
i
M
x
and b
i
in M
y
such that
/[= (a
i
, b
j
) i < j, for all i, j.
Using Lowenheim-Skolem, arrange that / has size . To each upward
closed set U I with U ,= I we associate the set of L
M
-formulas

U
(x) := (x, b
j
) : j U (x, b
j
) : j / U.
One checks easily that
U
(x) is a partial x-type over M in /. Let p
U
(x)
be an x-type over M in / that extends
U
(x). Then p
U
(x) ,= p
U
(x) if
U ,= I and U

,= I are distinct upward closed subsets of I.

Exercise. Let be given, and let be the smallest cardinal such that
2

> , so . We order the set G of all functions g : 2 = 0, 1

lexicographically, that is, g < h i for some ordinal < we have
g() < h(), g() = h() for all ordinals < .
Observe that this makes G a linearly ordered set. Let I be the (ordered)
subset of G consisting of those i G that are eventually constant. For
35
g G, put U(g) := i I : i g. Show that I has size , and that if
g, h G, g ,= h, then U(g) ,= U(h).
If (x, y) is stable, then for some positive integer N there are no a
i
M
x
and b
i
M
y
for i = 1, . . . , N such that
[= (a
i
, b
j
) i j, for i, j = 1, . . . , N,
and below N() denotes a positive integer N with this property. (Here we
abuse language by letting stand for (x, y).) The next result is crucial.
Lemma 8.4. Let (x, y) be a stable L-formula, a M
x
, and M a model.
Then there are positive integers I, J and elements a
i
j
M
x
for 1 i I
and 1 j J, such that for each b M
y
,
[= (a, b)
I

i=1
_
J

j=1
(a
i
j
, b)
_
.
Proof. Put J := N(), and let (x) tp(a[M).
Claim. There are a
1
, . . . , a
n
M
x
realizing (x), with 1 n J, such
that for all b M
y
we have [=
_ _
n
i=1
(a
i
, b)
_
(a, b).
To prove this claim, dene a -sequence of length n to be a sequence
(a
1
, . . . , a
n
, b
1
, . . . , b
n
), (a
1
, . . . , a
n
M
x
, b
1
, . . . , b
n
M
y
)
such that for all i, j 1, . . . , n,
[= (a
i
, b
j
) i j, [= (a
i
) (a, b
i
).
Note that we have a -sequence of length 0. Let (a
1
, . . . , a
n
, b
1
, . . . , b
n
) be a
-sequence, so n < J. The L
M
-formula (x)
_
i
(x, b
i
) is realized by a,
and thus realized by an element a
n+1
M
x
. If there is a b M
y
such that
[=
_
n+1
i=1
(a
i
, b)(a, b), then we let b
n+1
be such a b, and we have extended
our sequence to a longer -sequence (a
1
, . . . , a
n+1
, b
1
, . . . , b
n+1
). If there is
no such b, then the claim holds for a
1
, . . . , a
n+1
M
x
. Since -sequences
have length < J, the extension process must come to a halt. Thus the claim
is established. Note that in this claim we can arrange n = J. For use in the
second part of the proof, we say that a tuple a = (a
1
, . . . , a
J
) M
J
x
satises
the claim if for all b M
y
we have
[=
_
J

j=1
(a
j
, b)
_
(a, b).
Now the second part of the proof. Introduce a multivariable x = (x
1
, . . . , x
J
)
where the x
j
are multivariables similar to x, pairwise disjoint, and disjoint
from y. Consider the L-formula

J
(x, y) := (x
1
, y) (x
J
, y).
36
Then
J
(x, y) is stable, and so is
J
(x, y). Put I := N(
J
). Suppose the
tuples a
1
, a
2
, . . . , a
n
M
J
x
= M
x
satisfy the claim, and b
1
, . . . , b
n
M
y
are
such that for all i, j 1, . . . , n
[=
J
(a
i
, b
j
) i > j, [= (a, b
j
).
Let us express this by saying that (a
1
, a
2
, . . . , a
n
, b
1
, . . . , b
n
) is a -sequence.
Note that then n < I, and that for n = 0 there exists a -sequence. We now
try to extend our -sequence. Applying the claim with (x) :=
_
n
j=1
(x, b
j
)
yields a tuple a
n+1
M
x
such that [=
J
(a
n+1
, b
j
) for j = 1, . . . , n, and for
each b M
y
, [=
J
(a
n+1
, b) (a, b). If there is a b M
y
such that
[= (a, b) and [=
J
(a
i
, b) for i = 1, . . . , n+1, then we let b
n+1
be such a b,
and we have constructed a longer -sequence
(a
1
, a
2
, . . . , a
n+1
, b
1
, . . . , b
n+1
).
If there is no such b, then for all b M
y
we have [= (a, b)
_
n+1
i=1

J
(a
i
, b),
and thus
[= (a, b)
n+1

i=1

J
(a
i
, b).
Then the lemma holds for the above values of I and J by taking the a
i
j
such
that a
i
= (a
i
1
, . . . , a
i
J
) for i = 1, . . . , n + 1, and a
i
j
= a
1
j
for n + 1 < i I.
The process of extending a -sequence must come to a halt.
Remarks. The proof shows that we can take I and J to depend only on
(x, y), not on a or M. In addition, if (x) tp(a[M), we can take the a
i
j
to realize (x). Also, if the model M is -saturated and A M has size
< , then we can choose the a
i
j
to realize tp(a[A).
A striking consequence of the lemma is that the subset (a, M
y
) of M
y
is
denable in the model M, although a might not be an M-tuple!
Corollary 8.5. Given the L-formula (x, y), the following are equivalent:
(1) (x, y) is stable;
(2) every -type over any M is denable;
(3) [ St

(M)[ , for each [L[ and each M of size .

(4) for some small [L[ we have [ St

(M)[ for each M of size .

Proof. For (1) (2), assume (1) and let p(x) be a -type over M. Take
a M
x
realizing p(x). Then Lemma 8.4 yields an L
M
-formula (y) such
that (a, M
y
) = (M
y
), that is, (y) denes p(x).
The implications (2) (3) and (3) (4) are clear. The implication
(4) (1) is clear from the proof of Lemma 8.3.
We say that T is stable if each L-formula (x, y) is stable. To prove that
stability of T is equivalent to other conditions on T we use that a type is
determined by the -types contained in it for relevant . It follows that if T
is stable, then each x-type over M has a dening scheme, hence [ St
x
(M)[
37

|L|
for M of size [L[. Here is a useful consequence of stability for global
types.
Corollary 8.6. If T is stable and has EI, then every global x-type has a
canonical base.
Proof. Stability of T implies that global types are denable, since we may
view M as a small elementary submodel of an even bigger monster. Now
apply Lemma 6.5.
From now on we shall also assume (as we may) that
(M) > 2
|L|
.
This will be used in proving some of the implications below.
Corollary 8.7. The following are equivalent:
(1) T is stable;
(2) no 0-denable binary relation R M
x
M
x
with nite x linearly
orders any innite subset of M
x
when restricted to that subset;
(3) whenever Y M
y
with nite y is denable, then the subset Y M
y
of M
y
is denable in the model M;
(4) [ St
x
(M)[
|L|
, for each nite x, [L[ and M of size ;
(5) for some small [L[ we have [ St
x
(M)[ for each nite x and
M of size ;
(6) for some small [L[ we have [ St
v
(M)[ for each variable v
and M of size ;
(7) each L-formula (v, y) where v is a variable and y is nite, is stable.
Proof. The equivalence of (1) and (2) is clear from the denition of stable
and Lemma 8.2. The equivalence of (1) and (3) follows from the equivalence
of (1) and (2) in Corollary 8.5. The implication (1) (4) was already
noted. From (4) we obtain (5) with = 2
|L|
, since this satises
|L|
= .
The implication (5) (6) is obvious. The implication (6) (5) follows by
induction on the number of variables in x from Lemma 5.7, and (5) (1)
follows from (3) (1) of Corollary 8.5. It remains to show that (7) is
equivalent to (6). The implication (6) (7) follows from (4) (1) of
Corollary 8.5. Assume (7), and let v be a variable. Then each v-type over
any M has a dening scheme, hence [ St
v
(M)[
|L|
for M of size [L[,
and thus (6) holds with = 2
|L|
.
Remark. The reader may nd it annoying that some of the above conditions
refer implicitly or explicitly to the monster model. Since M and (M) can
be chosen as large as we want, it is easy to see that if T is stable, then
(3) actually holds with M and M replaced by any / and ^ such that
/ _ ^ [= T, and that (4) holds with M replaced by arbitrary models /
of T. Likewise, (6) (1) holds without restricted to being small and with
M replaced by arbitrary models / of T.
38
Exercise. Suppose T is stable. Show that for each A the L
A
-theory T
A
of
M is stable. Show that T
eq
is stable.
Suppose now that (x, y) is a stable L-formula and p(x) is a -type over
M dened by (y). Lemma 8.4 shows that (y) is a

-formula over M, in
particular, either (y) q or (y) q whenever q is a

-type over M.
This remark is relevant in connection with the following symmetry result.
Lemma 8.8. Let (x, y) be a stable L-formula and A M. Let p(x) be a
-type over M dened by the L
A
-formula (y), and let q(y) be a

-type over
M dened by the L
A
-formula (x). Then
(y) q(y) (x) p(x).
Proof. Suppose towards a contradiction that (y) q(y) but (x) / p(x).
Let := max[A[, [L[
+
. By working in a big enough monster model we
can pass to a -saturated N M and replace p and q by p N and q N
we reduce to the situation that M is -saturated. Take a
0
M
x
realizing
p(x) A, so [= (a
0
), hence

(y, a
0
) q(y), that is, (a
0
, y) q(y).
Next, take b
0
M
y
realizing q(y)Aa
0
, so [= (a
0
, b
0
) and [= (b
0
), hence
(x, b
0
) p(x). Now, take a
1
M
x
realizing p(x) Ab
0
, so [= (a
1
, b
0
) and
[= (a
1
), and thus (a
1
, y) q(y). Continuing this way, we construct
innite sequences a
0
, a
1
, a
2
, . . . in M
x
and b
0
, b
1
, b
2
, . . . in M
y
, with a
n+1
realizing p(x) Ab
1
. . . b
n
and b
n
realizing q(y) Aa
1
. . . a
n
, for all n. One
shows by induction that [= (a
n
) (b
n
), and thus, for all m, n:
[= (a
m
, b
n
) m > n,
contradicting the stability of (x, y).
Indiscernibility. Let (a
i
)
iN
be a sequence in M
x
. We say that (a
i
) is
indiscernible over A if
tp((a
i
1
, . . . , a
in
)[A) = tp((a
j
1
, . . . , a
jn
)[A)
for all n and all n-element sets i
1
, . . . , i
n
, j
1
, . . . , j
n
N. We say that
(a
i
) is order-indiscernible over A if the displayed identity holds for all n and
i
1
< < i
n
and j
1
< < j
n
in N. In the presence of stability these two
notions coincide:
Lemma 8.9. Suppose T is stable and the sequence (a
i
)
iN
in M
x
is order-
indiscernible over A. Then (a
i
) is indiscernible over A.
Proof. To keep notation simple, assume A = 0. We rst prove the special
case tp(a
0
, a
1
, a
2
) = tp(a
1
, a
0
, a
2
). Suppose tp(a
0
, a
1
, a
2
) ,= tp(a
1
, a
0
, a
2
).
Then we have an L-formula (x, y, z) with y and z similar to x such that
[= (a
0
, a
1
, a
2
) and [= (a
1
, a
0
, a
2
), hence
[= (a
i
, a
j
, a
k
) and [= (a
j
, a
i
, a
k
) for all i < j < k.
39
By saturation we can take a M
x
such that [= (a
i
, a
j
, a) and [= (a
j
, a
i
, a)
for all i < j, contradicting stability. Having now established the special case,
it follows that tp(a
i
, a
j
, a
k
) = tp(a
j
, a
i
, a
k
) for all i < j < k in N.
For the general case, proceed likewise, and use that the permutations of
1, . . . , n are generated by the transpositions (i, i + 1) with 1 i < n.
9. Rank and degree in model theory
From now on x, y, z are nite multivariables, unless specied otherwise.
Let / be an L-structure and Y a denable set in /, say Y M
y
.
Then we dene the Cantor rank of Y (in /) to be the Cantor rank of Y
as an element of the boolean algebra Def
y
(/) of denable subsets of M
y
,
equivalently, it is the Cantor rank of Y as an element of the boolean algebra
of denable subsets of Y . We denote it by CR(Y ), and by CR
M
(Y ) if we
wish to indicate the ambient structure /. If CR(Y ) is an ordinal, then we
call Y ranked, and dene CD(Y ) (the Cantor degree of Y ) to be the Cantor
degree of Y as an element of the boolean algebra of denable subsets of M
y
.
(If we wish to indicate the ambient structure / we write this degree as
CD
M
(Y ).) Suppose now that X is also a denable set in /, say X M
x
,
and that the map f : X Y is denable. Then
(i) if f(X) = Y , then CR(X) CR(Y ), and in case of equality with X
ranked we have CD(X) CD(Y ).
(ii) if f is injective, then CR(X) CR(Y ), and in case of equality with
X ranked we have CD(X) CD(Y ).
To see this, note that U f
1
(U) is a morphism from the boolean algebra
of denable subsets of Y into the boolean algebra of denable subsets of X,
and that this morphism sends Y to X. If f(X) = Y , then this morphism is
injective, so we can apply part (1) of Proposition 2.3. If f is injective, then
this morphism is surjective, so we can apply part (2) of that proposition.
Unfortunately, the Cantor rank of a denable set may change in passing
to an elementary extension of /. Take for example / = (N, <) and let
^ = (N, <) be an
0
-saturated elementary extension of /. The denable
subsets of N in / are exactly the nite and conite subsets of N (exercise),
so N has Cantor rank 1 as a denable set in /. But N as a denable set
in ^ has Cantor rank > 1, as is easily veried.
More generally, let /_ ^, so we have a boolean algebra embedding
: Def
y
(/) Def
y
(^), (a, M
y
) (a, N
y
),
where (x, y) ranges over L-formulas of the indicated form with a M
x
.
By Proposition 2.3 we have CR(Y ) CR(Y ) for Y Def
y
(/). With a
mild saturation assumption we have equality:
Lemma 9.1. Let / be
0
-saturated. Then for all Y Def
y
(/),
CR(Y ) = CR(Y ), CD(Y ) = CD(Y ) in case Y is ranked.
40
Proof. By increasing ^ we can assume ^ is strongly
1
-homogeneous. Let
Y Def
y
(/). The rank equality follows from
CR(Y ) > = CR(Y ) > , for all ,
which we prove by induction on . Suppose CR(Y ) > . Then we have
disjoint denable sets Y
n
Y of Cantor rank in ^. Take a nite tuple
a in /such that Y is a-denable in /. Take a tuple b in ^ of countable size
such that all Y
n
are b-denable in ^. By the saturation and homogeneity
assumptions on / and ^ we have an automorphism f Aut(^[a) such
that fb is an M-tuple. Then f(Y ) = Y and f(Y
n
) is denable over M, so
f(Y
n
) = X
n
with X
n
Y denable in /. Since
CR(X
n
) = CRf(Y
n
) = CR(Y
n
) ,
we can assume inductively that CR(X
n
) . The X
n
being disjoint subsets
of Y , we obtain CR(Y ) > . The degree equality follows in a similar way.
Let Y Def
y
(/) be given. Choose an
0
-saturated elementary extension
^ of / and let be as above. Then by the lemma CR(Y ) is independent
of the choice of ^, so we can now dene the Morley rank of Y by
MR(Y ) = MR
M
(Y ) := CR(Y )
In case this rank is an ordinal, the Morley degree of Y is dened likewise by
MD(Y ) = MD
M
(Y ) := CD(Y ).
Often / is already
0
-saturated, and then MR(Y ) = CR(Y ), and in case
MR(Y ) is an ordinal, MD(Y ) = CD(Y ); this applies in particular to the
denable sets in our monster model M.
Exercise. Suppose X M
x
is M-denable (so the set X(M) := X M
x
is denable in the model M). Prove that
CR
M
_
X(M)
_
MR(X).
Totally transcendental theories. We dene T to be totally transcenden-
tal if every nonempty denable set in M is ranked. By Lemma 9.1 this is
indeed a property of T, that is, it does not depend on the particular monster
model M of T.
Lemma 9.2. Suppose T is totally transcendental. Then
(1) the theory T
A
of the L
A
-structure M is totally transcendental;
(2) T
eq
is totally transcendental;
(3) [ St
x
(M)[ [M[ for [M[ [L[;
(4) T is stable.
41
Proof. Item (1) is obvious, and (2) follows from the fact that every denable
set in M
eq
is the image of a denable set in M under a denable map. Item
(3) follows from the implication (iii)(iv) of Theorem 2.5 applied to the
subalgebras B = Def
x
(M[M) of Def
x
(M) for [M[ [L[, since [B[ [M[ for
such M. Item (4) follows from (3) and Corollary 8.7.
When the language is countable we can say a bit more:
Corollary 9.3. Let L be countable. The following are equivalent:
(1) T is totally transcendental;
(2) [ St
x
(M)[ [M[ for each x and innite M;
(3) [ St
x
(M)[
0
for each x and countable M;
(4) [ St
v
(M)[
0
for each variable v and countable M;
(5) each nonempty denable subset of M
v
is ranked, for each variable v.
Proof. The implication (1) (2) is part of the previous lemma. The im-
plications (2) (3) and (3) (4) are obvious. The equivalence (4) (5)
follows from the equivalence (3) (4) of Theorem 2.5 by noting that each
countable subalgebra of Def
v
(M) is contained in a subalgebra of the form
Def
v
(M[M) with countable M. The implication (4) (3) follows by in-
duction on the size of x from Lemma 5.7, and (3) (1) is again an easy
consequence of Theorem 2.5.
By Theorem 2.5, all nonzero elements in a boolean algebra are ranked if
each countable subalgebra has this property. Thus:
Corollary 9.4. The following are equivalent:
(1) T is totally transcendental;
(2) for each countable sublanguage L
0
of L the theory of the L
0
-reduct
of M is totally transcendental;
(3) each nonempty denable subset of M
v
is ranked, for each variable v.
Proof. The equivalence of (1) and (2) follows easily from Theorem 2.5. We
already know that (1) implies (3), and the implication (3) (2) follows
from the previous corollary.
For countable L, a totally transcendental T is also said to be omega-stable
(or -stable).
Stability and rank. Let (x, y) be an L-formula, and consider the boolean
subalgebra Def

(M) of Def
x
(M) that is generated by the sets (M
x
, b) with
b M
y
. For X Def

(M) we dene the -rank R

(X) of X to be its Cantor

rank in this boolean algebra, and if this rank is an ordinal, we let D

(X) be
the Cantor degree of X in this boolean algebra.
Proposition 9.5. The formula (x, y) is stable if and only if each nonempty
X Def

(M) is ranked.
42
Proof. We prove the contrapositives. Suppose (x, y) is unstable. Then we
have a model M of size [L[ such that St

(M) > . Then by Theorem 2.5,

the rank of M
x
as an element of the boolean subalgebra Def

(M) of Def

(M)
is +, and thus R

(M
x
) = +.
For the converse, suppose R

(M
x
) = +. Then we take an M of size
[L[ such that M
x
as an element of the boolean algebra Def

(M) has
rank +. For each n, Corollary 2.8 yields elements b
1
, . . . , b
n
M
y
and
types p
0
(x), . . . , p
n
(x) St
x
(M) such that
(x, b
i
) p
j
(x) i j, for i = 1, . . . , n, j = 0, . . . , n.
With such b
i
and p
j
we take a
j
M
x
realizing p
j
and obtain
[= (a
j
, b
i
) i j, for i = 1, . . . , n, j = 0, . . . , n.
Hence (x, y) is unstable.
10. Basic facts on Morley rank
From now on we restrict our attention mainly to Morley rank and totally
transcendental theories, although much will go through, by similar argu-
ments, for -rank and stable theories. We begin with improving the result
that MR(X) MR(Y ) if there is an injective denable map from X to Y .
Lemma 10.1. Let the sets X, Y and the map f : X Y be denable in M
such that f
1
(b) is nite for all b Y . Then MR(X) MR(Y ).
Proof. By induction on MR(Y ). The case that MR(Y ) , 0, + is
obvious, so we can assume MR(Y ) = 1, and make a further reduction
to MD(Y ) = 1. Saturation gives an n such that [f
1
(b)[ n for all b Y .
Suppose that MR(X) > . Take disjoint denable subsets X
1
, . . . , X
n+1
of
X with MR(X
i
) for all i. If MRf(X
i
) < for some i, then by the
induction hypothesis,
MR(X
i
) MRf(X
i
) <
for such i, a contradiction. Hence MRf(X
i
) = for all i. As MD(Y ) = 1,
we have a b f(X
1
) f(X
n+1
), contradicting [f
1
(b)[ n.
Corollary 10.2. Let the sets X, Y , and R XY be denable in M such
that for each a X the section R(a) is nite and for each b Y there is an
a X with (a, b) R. Then MR(X) MR(R) MR(Y ).
Proof. All bers of the map (a, b) a : R X are nite, so MR(X)
MR(R) by the previous lemma. The map (a, b) b : R Y is surjective,
so MR(R) MR(Y ).
Exercise. Show that if X M
x
, Y M
y
and f : X Y are denable,
and MR
_
f
1
(y)
_
1 for all y Y , then MR(X) MR(Y ) + 1.
43
For denable X, Y M
x
we set
X =

Y : MR(XY ) < ,
X

Y : MR(X Y ) < ,
X := M
x
X,
or in terms of the boolean algebra Def
x
(M) and its ideal I := I
<
:
X =

Y X =
I
Y, X

Y X/I Y/I.
Let X be a denable set in M. Clearly, MR(X) = 0 i X is nite and
nonempty. If MR(X) = 0, then MD(X) = [X[. The next possible size
is when MR(X) = 1, MD(X) = 1, and such X is called strongly minimal
(in M), and this is equivalent to X being innite such that each denable
subset of X is nite or conite in X. We say that X is A-irreducible if X is
A-denable and ranked, and X is not the union of two disjoint A-denable
subsets of the same Morley rank. For example, if X is A-denable and
ranked with MD(X) = 1, then X is A-irreducible.
Exercise. Show that if X, Y and R X Y are denable sets in M and
Y is strongly minimal, then there is an m such that for all a X, either
[R(a)[ m or [Y R(a)[ m.
Types over A and A-irreducible sets. Let p(x) be an x-type over A. We
shall identify p(x) with the corresponding ultralter (M
x
) : (x) p(x)
of the boolean algebra Def
x
(M[A), so for each A-denable X M
x
, either
X p or X p.
If p does not contain any ranked set we put MR(p) = +. If p contains
a ranked set, we choose X p of minimal Morley rank and of minimal
Morley degree d among the sets in p of Morley rank , and put MR(p) = ,
MD(p) = d. This set X is then A-irreducible, and
p = Y Def
x
(M[A) : X

Y .
Lemma 10.3. Suppose X M
x
is ranked with MR(X) = . Then
(1) If X is A-irreducible, then p := Y Def
x
(M[A) : X

Y is an
x-type over A with MR(p) = MR(X), MD(p) = MD(X).
(2) MR(X) = maxMR(p) : p St
x
(A), X p.
(3) The set p : p St
x
(A), X p, MR(p) = is nite and
MD(X) =

p
MD(p) where the sum is over the p in this nite set.
Proof. For (1), use that if X is A-irreducible, then for each A-denable
Y M
y
, either MR(XY ) < (in which case X

Y ), or MR(XY ) <
(in which case X

Y ). To obtain (2) and (3), decompose X as X

1
X
n
with n 1 and disjoint A-irreducible X
1
, . . . , X
n
of Morley rank . Now
apply (1) to each X
i
.
44
We say that an A-irreducible set X M
x
determines the type p dened in
(1) of Lemma 10.3. Let p(x) St
x
(A); then
MR(p) = 0 p(M
x
) is nite;
we say that p is algebraic if MR(p) = 0; in that case, MD(p) = [p(M
x
)[.
For a M
x
we dene
MR(a[A) := MR
_
tp(a[A)
_
.
Thus MR(a[A) = 0 a is A-algebraic. Note that if MR(a[A) = , then
there is an A-irreducible X M
x
such that a X and MR(X) = and
there is no A-denable X M
x
with a X and MR(X) < . Conversely,
if X M
x
is A-irreducible, then there is an a X such that MR(a[A) =
MR(X), and such an a is called a generic element of X over A (or just a
generic of X over A).
From now on we let a, a

, a
1
, a
2
, b, b

, c denote nite tuples in M.

Lemma 10.4. Suppose b is a-algebraic over A. Then MR(b[A) MR(a[A).
Proof. Let a M
x
and b M
y
, and take an A-denable X M
x
with
a X and MR(X) = MR(a[A). Take an A-denable R M
x
M
y
such
that b R(a) and R(a) is nite. By shrinking R we can arrange that
R X M
y
and [R(a

)[ [R(a)[ for all a

X. Let
Y := b

M
y
: R(a

, b

) for some a

X.
Then Y is A-denable and R X Y satises the conditions of Corol-
lary 10.2, so MR(b[A) MR(Y ) MR(X) = MR(a[A).
In particular, if a and b are interalgebraic over A, then MR(a[A) = MR(b[A).
Denable types again. Following Ziegler we shall give another proof of
denability of types in the case of totally transcendental theories.
Lemma 10.5. Suppose the model M is
0
-saturated, and X M
x
is den-
able and ranked, and contained in an M-denable subset of M
x
of the same
Morley rank. Then X M
x
,= .
Proof. Let Y M
x
be M-denable with MR(X) = MR(Y ) = . We
proceed by induction on . If = 0, then Y is nite, so X Y M
x
. Let
> 0 and decompose Y as a disjoint union Y
1
Y
d
of M-denable sets
of Morley rank and Morley degree 1. (This is possible by Lemma 9.1.)
Then MR(X Y
i
) = for some i, so by replacing X and Y by X Y
i
and
Y
i
for suitable i we have made a reduction to the case that MD(Y ) = 1.
We can also assume that X ,= Y . So MR(Y X) = < . Take disjoint
M-denable sets Y
0
, Y
1
, Y
2
, Y of Morley rank and < . Not all Y
n
can be contained in Y X, so we can take n such that X Y
n
,= . Since
MR(Y
n
) < we can assume inductively that X Y
n
M
x
,= .
45
Given M and an M-denable X M
x
we put
X(M) := X M
x
(the set of M-points of X).
Theorem 10.6. Let (x, y) be a stable L-formula and let X M
x
be de-
nable with MR(X) = . Then the subset
b M
y
: X

(M
x
, b)
of M
y
is denable.
Proof. By decomposing X into irreducible sets of Morley rank we can
assume that MD(X) = 1. Take an
0
-saturated model M such that X is
M-denable.
Claim 1. X(M) has a nite subset such that for all b M
y
,
(M
x
, b) = X

(M
x
, b).
To prove this claim, suppose there is no such . Then we obtain a contra-
diction by constructing innite sequences a
1
, a
2
, X(M) and b
1
, b
2
,
M
y
such that for all i, j
[= (a
i
, b
j
) i j.
We obtain such sequences as follows: Assume a
1
, . . . , a
n
X(M) and
b
1
, . . . , b
n
M
y
are such that this equivalence holds for i, j = 1, . . . , n and
such that X ,

(M
x
, b
i
) for i = 1, . . . , n. Then
X ,

(M
x
, b
1
) (M
x
, b
n
),
so by the previous lemma we have an a
n+1
X(M) outside each (M
x
, b
i
).
Since a
1
, . . . , a
n+1
cannot serve as , we obtain b
n+1
M
y
such that
a
1
, . . . , a
n+1
(M
x
, b
n+1
) and X ,

(M
x
, b
n+1
), so the displayed
equivalence holds for i, j = 1, . . . , n + 1.
Claim 2. Suppose that X

(M
x
, c), c M
y
. Then there exists a as
in Claim 1 such that (M
x
, c).
The proof is a variant of that of Claim 1: take the a
i
(M
x
, c).
For each as in Claim 1, let Y

:= b M
y
: (M
x
, b), so Y

M
y
is denable. Claims 1 and 2 yield that
b M
y
: X

(M
x
, b) =
_

where the union is over the small set of s as in Claim 1. Since X is

irreducible, the complement in M
y
of the set on the left is
b M
y
: X

(M
x
, b),
which is likewise a union of a small number of denable sets. Now use an
exercise from Section 2.
In Section 5 we proved that if T is stable, then types over a model M are
denable over M. Here we extend this to types over parameter sets, for
totally transcendental T.
46
Corollary 10.7. Assume T is totally transcendental, and let p(x) St
x
(A).
Then p is denable over A: for each L-formula (x, y) there is an L
A
-
formula (y) such that for all b A
y
:
(x, b) p(x) [= (b).
Proof. Take an A-irreducible set X M
x
that determines p, so MR(X) =
MR(p) = , say. Then for all b A
y
,
(x, b) p(x) (M
x
, b) p X

(M
x
, b).
Now apply Theorem 10.6.
Morley rank of global types. Let p St
x
(M) be a global type, viewed
as an ultralter of Def
x
(M). If p does not contain any ranked set, we put
MR(p) = +. Suppose that p contains a ranked set. Then we choose
X p of minimal Morley rank , and put MR(p) = . Such an X can be
chosen to have Morley degree 1, and then determines p in the sense that
p = Y Def
x
(M) : X

Y .
Conversely, if X M
x
is denable, MR(X) = , MD(X) = 1, then X
determines the global type p = Y Def
x
(M) : X

Y of Morley rank .
11. Forking and Independence
In this section we assume that T is totally transcendental.
Let p St
x
(A), A B and suppose q St
x
(B) extends p, that is, p q,
equivalently, q A = p. Then MR(p) MR(q), and in case of equality we
say that q is a nonforking extension of p (to B). We also say that q does
not fork over A if it is a nonforking extension of p; this terminology makes
sense, since q determines p by p = q A.
For example, if p is algebraic, then q is necessarily a nonforking extension
of p, and if MR(p) = 1, then q is a nonforking extension of p i q is not alge-
braic. We also say that a global type p St
x
(M) is a nonforking extension
of p if p p and MR(p) = MR(p).
Lemma 11.1. Let p St
x
(A), A B. Then p has a nonforking extension
to B, and has at most MD(p) such extensions:
MD(p) =

q
MD(q),
where the sum is over the nonforking extensions q of p to B.
Proof. Take an A-irreducible set X that determines p, so MR(X) = MR(p)
and MD(X) = MD(p). Take disjoint B-irreducible sets X
1
, . . . , X
n
of the
same Morley rank as X such that X = X
1
X
n
, n 1. Then each X
i
determines a nonforking extension p
i
St
x
(B) of p, with MD(X
i
) = MD(p
i
).
Also each nonforking extension of p to B contains an X
i
, and thus equals a
p
i
. It remains to note that MD(X) =

i
MD(X
i
).
47
Almost the same proof, with irreducible in place of B-irreducible, yields
a similar result for global nonforking extensions:
Lemma 11.2. Each p St
x
(A) has exactly MD(p) global nonforking exten-
sions in St
x
(M).
Corollary 11.3. Let p St
x
(A). Then MD(p) = 1 if and only if p has
exactly one nonforking extension to every B A.
Proof. The forward direction is immediate from Lemma 11.1. For the con-
verse, suppose n := MD(p) > 1. Take an A-irreducible set X that deter-
mines p, so MD(X) = n. Then X is a disjoint union of denable subsets
X
1
, . . . , X
n
of the same Morley rank. Take a B A such that X
1
, . . . , X
n
are B-denable, hence B-irreducible. Then X
1
, . . . , X
n
determine distinct
types q
1
, . . . , q
n
St
x
(B), and all are nonforking extensions of p.
A type p St
x
(A) with MD(p) = 1 is also said to be stationary, because
it has the property expressed in this corollary. Note that if p St
x
(A) is
stationary, then p has also a unique global nonforking extension p St
x
(M).
It is obvious that nonforking is monotone and transitive:
Lemma 11.4. Let A B C and p St
x
(A), q St
x
(B), r St
x
(C) be
such that p q r. Then r is a nonforking extension of p if and only if r
is a nonforking extension of q and q is a nonforking extension of p.
Nonforking is also continuous:
Lemma 11.5. Let q St
x
(B). Then
(1) there is a nite B
0
B such that q does not fork over B
0
and q is
the only nonforking extension of q B
0
to B;
(2) if A B and q forks over A, then there is a nite B
0
B such that
q (A B
0
) forks over A.
Proof. Take a set X q with MR(X) = MR(q), and MD(X) = MD(q).
Take a nite B
0
B such that X is B
0
-denable. Then B
0
witnesses (1) by
Lemma 11.1. For (2), assume A B and q forks over A, and put p := q A.
Then we can take X q such that MR(X) = MR(q) < MR(p). Take a
nite B
0
B such that X is (A B
0
)-denable. Then B
0
witnesses the
conclusion of (2).
We say that a is independent from B over A, notation:
a
[

A
B,
if MR(a[AB) = MR(a[A), in other words, tp(a[AB) does not fork over A.
In this notation, Lemma 10.4 and the lemmas above in this section yield:
a is A-algebraic = a
[

A
B,
if a and b are interalgebraic over A, then: a
[

A
B b
[

A
B,
48
p St
x
(A) = p has a realization a such that a
[

A
B,
a
[

A
B and B

B = a
[

A
B

,
_
a
[

A
B and a
[

AB
C
_
a
[

A
BC,
a
[

A
B a
[

A
b for each nite B-tuple b.
We shall use these rules freely from now on. Note the special case
a
[

A
BC = a
[

AB
C.
Symmetry. In the next symmetry lemma, a and b also stand for certain
parameter sets, according to the convention in Section 2.
Lemma 11.6. a
[

A
b b
[

A
a.
Proof. We rst assume that A = M is an
0
-saturated model. Set :=
MR(a[M) and := MR(b[M). Let a M
x
and b M
y
with disjoint x, y,
and take M-denable X M
x
and Y M
y
that determine tp(a[M) and
tp(b[M). Suppose that b is not independent from a over M. Then
MR(b[Ma) < MR(b[M) = ,
so we have an M-denable Z M
x
M
y
such that
(a, b) Z, MR
_
Z(a)
_
< .
We can assume that Z X Y . We have
a

X : MR
_
Z(a

)
_
< = a

X : Y ,

Z(a

),
so by Theorem 10.6, the set X

:= a

X : MR
_
Z(a

)
_
< is M-
denable, and contains a. After replacing Z by Z (X

Y ) we can assume
that MR
_
Z(a

)
_
< for all a

X. Hence

Z(b)M
x
= , so by Lemma 10.1
we have MR
_

Z(b)
_
< MR(X) = . So a is not independent from b over A.
We now consider any A, and take an
0
-saturated model M A. Take a
b

M
y
such that
b

realizes tp(b[A), b

A
M.
Take f Aut(M[A) with fb = b

, and take an a

M
x
such that
a

realizes tp(fa[Ab

), a

Ab

Mb

,
and take g Aut(M[Ab

) with g(fa) = a

. Then h := gf Aut(M[A)
satises h(a) = a

and h(b) = b

, so
a
[

A
b a

A
b

, b
[

A
a b

A
a

,
49
Suppose now that a is independent from b over A. Then a

A
b

, so a

A
Mb

(since a

Ab

Mb

), so a

M
b

. By the rst part of the proof, this gives b

M
a

,
which in view of b

A
M yields b

A
a

, so b is independent froma over A.

Corollary 11.7. a
[

A
acl(A).
Proof. For A-algebraic b we have b
[

A
a, hence a
[

A
b.
The next result is needed in dealing with one-basedness in Section 13.
Corollary 11.8. Suppose a
[

A
b. Then acl(Aa) acl(Ab) = acl(A).
Proof. Let c be algebraic over Aa and over Ab. Then
MR(a[Abc) = MR(a[Ab) = MR(a[A),
and thus MR(a[Ac) = MR(a[A). Hence a
[

A
c, so c
[

A
a, that is, MR(c[A) =
MR(c[Aa) = 0, so c is algebraic over A.
Lemma 11.9. Let p St
x
(A). Then
(1) each q p in St
x
(acl(A)) is a nonforking extension of p;
(2) any two extensions of p in St
x
(acl(A)) are conjugate over A.
Proof. Item (1) is immediate from Corollary 11.7 by taking a realization
a M
x
of q. For (2) we take an A-irreducible X that determines p, and take
an acl(A)-irreducible Y X with MR(X) = MR(Y ). Take an A-algebraic b
such that Y is b-denable. Then clearly f(b) : f Aut(M[A) is nite, and
whenever f(b) = g(b), then f(Y ) = g(Y ), for f, g Aut(M[A), so Y has only
nitely many conjugates over A. Let Y
1
, . . . , Y
n
be the distinct conjugates
of Y over A. Then Y
1
Y
n
X is A-denable by Corollary 5.2, so
X =

Y
1
Y
n
where = MR(X). The sets Y
1
, . . . , Y
n
determine
extensions q
1
, . . . , q
n
St
x
(acl(A)) of p, and q
1
, . . . , q
n
are conjugate over A
the same way Y
1
, . . . , Y
n
are.
Suppose that q St
x
(acl(A)) extends p. Since Y
1
Y
n
p we have
an i 1, . . . , n with Y
i
q. Since MR(q) = by (1), this Y
i
determines
q, so q = q
i
.
Theorem 11.10. Suppose T has EI and A is algebraically closed. Let A
M, and let p, p
1
, p
2
St
x
(M). Then
(1) if p does not fork over A, then p is denable over A;
(2) if p
1
, p
2
are denable over A, and p
1
A = p
2
A, then p
1
= p
2
;
(3) each x-type over A is stationary.
Proof. For (1), assume p does not fork over A. Take a global nonforking
extension p St
x
(M) of p. The conjugates f(p) with f Aut(M[A) are
all nonforking extensions of p A, so there are only nitely many such
50
conjugates. Let (x, y) be an L-formula, and Y := d
p
x(x, M
y
). Then by
the above, Y has only nitely many conjugates over A, so Y is A-algebraic
by Corollary 5.2, hence Y is A-denable by Corollary 5.3.
For (2), let p
1
, p
2
be denable over A, and p
1
A = p
2
A. Let (x, y)
be an L-formula, and let
1
(y) and
2
(y) be L
A
-formulas that dene p
1

and p
2
. It is enough to show that then
1
(M
y
) =
2
(M
y
). Let b

1
(M
y
), and let q St
y
(M) be a nonforking extension of tp(b[A). Then q is
denable over A by (1), so we have an L
A
-formula (x) dening q

. Since

1
(y) q(y) we can apply Lemma 8.8 to p
1
and q

to get (x) p
1
(x),
hence (x) p
2
(x). Now apply the same lemma to p
2
and q

to get

2
(y) q, that is, b
2
(M
y
).
Item (3) follows from (1) and (2).
In the next three corollaries we do not assume that T has EI.
Corollary 11.11. Every x-type over a model M is stationary.
Proof. Each p St
x
(M) extends uniquely to a type p
eq
St
x
(M
eq
) in
M
eq
. Since T
eq
has EI, and M
eq
is algebraically closed in M
eq
, the type p
eq
is stationary, that is, MD(p
eq
) = 1. But Def
x
(M[M) and Def
x
(M
eq
[M
eq
)
are the same boolean algebra, and, viewed as ultralters on these boolean
algebras, p and p
eq
are also the same, so MD(p) = 1, so p is stationary.
Exercises. Show that if X M
x
is M-denable, MR(X) = , MD(X) = d,
then there are M-denable disjoint X
1
, . . . , X
d
M
x
such that
X = X
1
X
k
, and MR(X
i
) = , MD(X
i
) = 1 for i = 1, . . . , d.
Show that for a, b M
x
, stp(a[A) = stp(b[A) if and only if there is a model
M A such that tp(a[M) = tp(b[M).
Elaborating on the proof of Corollary 11.11 we note that each x-type over
A in M, as an ultralter of the boolean algebra Def
x
(M[A) is also an x-
type over A in M
eq
since Def
x
(M[A) and Def
x
(M
eq
[A) are the same boolean
algebra. In view of Aut(M[A) = Aut(M
eq
[A) (see Section 3), this yields
Corollary 11.12. Any two global nonforking extensions of p St
x
(A) are
conjugate over A.
Proof. By the observation preceding the statement of this corollary it suces
to prove this for types in M
eq
instead of M, that is, we can assume that T
has EI. Let p
1
and p
2
be nonforking extensions of p St
x
(A), and let
p
i
:= p
i
acl(A) for i = 1, 2. By Lemma 11.9, (2), we have an f Aut(M[A)
such that f(p
1
) = p
2
. Then by Theorem 11.10, (3), we have f(p
1
) = p
2
.
Corollary 11.13. Suppose M B and q St
x
(B) does not fork over M.
Then q is denable over M.
Proof. As subalgebras of the boolean algebra Def
x
(M) = Def
x
(M
eq
) we have
Def
x
(M[B) = Def
x
(M
eq
[B), Def
x
(M[M) = Def
x
(M
eq
[M
eq
).
51
Viewing in this way q as an x-type over B in M
eq
, Theorem 11.10, (1), yields
that q is denable over M
eq
. It follows that q is denable over M.
Corollary 11.14. Suppose T has EI, and p St
x
(M). Then
(1) p does not fork over A if and only if cb(p) acl(A);
(2) p does not fork over A and pA is stationary if and only if
cb(p) dcl(A).
Proof. Assume cb(p) acl(A). Then p is dened over acl(A). Take a
nonforking extension q St
x
(M) of p acl(A). Then q is also dened over
acl(A) by Theorem 11.10, (1), so p = q by part (2) of that theorem, so p
does not fork over acl(A), and therefore does not fork over A by Lemma 11.9.
Conversely, if p does not fork over A, then p does not fork over acl(A), hence
is denable over acl(A) by Theorem 11.10, and thus cb(p) acl(A). This
proves (1).
For (2), if cb(p) dcl(A), then p does not fork over A by (1), and has no
conjugates over A dierent from itself, hence pA is stationary. Conversely,
if p does not fork over A and p A is stationary, then f(p) = p for all
f Aut(M[A), and thus cb(p) dcl(A).
Corollary 11.15. Suppose T has EI, and p St
x
(M). Then cb(p) = dcl(a)
for some a.
Proof. Let A := cb(p). Then by (2) of the previous corollary, p is the unique
nonforking extension of p := p A, and MD(p) = 1. Take an irreducible
A-denable X M
x
that determines p, and let a be a code of X. Then
X is a-denable, so X p a, so p a is stationary, and p is a nonforking
extension of pa. Hence by (2) of the previous corollary we have A dcl(a).
But a is A-denable, so dcl(a) = A.
Suppose T has EI, and p St
x
(A) is stationary. Then we put
cb(p) := cb(p), where p St
x
(M) is the nonforking extension of p.
Also, given a M
x
such that tp(a[A) is stationary, we put
cb(a[A) := cb(tp(a[A)).
Theorem 11.16. Let p St
x
(M), M B and p q St
x
(B). Then
q is a nonforking extension of p q is a heir of p.
Proof. Let q be the nonforking extension of p to B. Then q is denable
over M by Corollary 11.13. Let (x, y) be an L
M
-formula, b B
y
, and
(x, b) q. Take an L
M
-formula (y) that denes q . Then [= (b),
which in view of M _ M gives a M
y
such that [= (a), so (x, a) q,
and thus (x, a) p. This proves =. For =, note that p is denable, so
p has exactly one heir over B.
Morley sequences. This material will be needed in Section 13. Given a
stationary p St
x
(A), a Morley sequence in p is a sequence (a
i
)
iN
in M
x
52
such that a
i
realizes the unique nonforking extension of p to Aa
0
. . . a
i1
for each i. (In particular, all a
i
realize p.) Note that there exists a Morley
sequence in p.
Lemma 11.17. Let p St
x
(A) be stationary, and let (a
i
) be a Morley
sequence in p. Then
(1) a
i
[

A
(a
0
, . . . , a
i1
) for all i;
(2) the sequence (a
i
) is indiscernible over A;
(3) if (b
i
) is also a Morley sequence in p, then there is an f Aut(M[A)
such that f(a
i
) = b
i
for all i;
(4) suppose T has EI; then cb(p) dcl(a
i
: i N);
Proof. Item (1) is evident from the denition of Morley sequence. For (2) it
is enough, by Lemma 8.9, to show that for all i
0
< < i
n
,
tp((a
0
, . . . , a
n
)[A) = tp((a
i
0
, . . . , a
in
)[A).
We prove this by induction on n. The case n = 0 is clear. Let i
0
< < i
n
<
i
n+1
. Now a
n+1
realizes pAa
0
. . . a
n
and a
i
n+1
realizes pAa
0
. . . a
N
where
N = i
n+1
1, hence a
i
n+1
realizes pAa
i
0
. . . a
in
. Assuming inductively the
displayed identity above, it follows that
tp((a
0
, . . . , a
n
, a
n+1
)[A) = tp((a
i
0
, . . . , a
in
, a
i
n+1
)[A).
For (3), let (b
i
) be a Morley sequence in p. Induction on n as in the proof of
(2) yields tp((a
0
, . . . , a
n
)[A) = tp((b
0
, . . . , b
n
)[A), and (3) follows. To prove
(4), let p be the global nonforking extension of p. Consider an L-formula
(x, y). The proof of Lemma 8.4 shows that p is dened over (b
i
) for
some Morley sequence (b
i
) in p. But p and thus p is also dened over A.
Then by (3) p is dened over (a
i
). Since is arbitrary, it follows that p,
and thus p is dened over a
i
: i N.
There is a problem with this proof since we only dened p B for types p
over models. For stationary p it should be dened in general.
Independence of parameter sets. For use in Section 13 we extend our
notion of independence from nite tuples to parameter sets: A is indepen-
dent from B over C, notation: A
[

C
B, if for each A-tuple a we have a
[

C
B.
It is easily checked that if A and B are the parameter sets corresponding
to tuples a and b, then A
[

C
B is equivalent to a
[

C
b as dened previously.
Thus our notation for independence of parameter sets agrees with the con-
vention of letting a tuple in M stand for the corresponding parameter set
when it suits us.
12. Combinatorial Geometries and Strongly Minimal Sets
Strongly minimal sets are basic building blocks of totally transcendental
structures. As we shall see, some very robust features of a strongly minimal
set are in turn controlled by something more primitive: a pregeometry. We
53
begin with discussing pregeometries. As the name suggests, a pregeometry
can be turned into a geometry.
Pregeometries. A pregeometry is a set with an operation
cl : T() T()
such that for all E and a, b :
(1) E cl E;
(2) cl E =

cl F : F a nite subset of E;
(3) cl(cl E) = cl E;
(4) a cl(E b), a / cl E = b cl(E a).
In most cases (1) and (2) are trivially satised, and (3) and (4) may require
a little work. Condition (4) is called the Steinitz Exchange Axiom and is
the most signicant of the four conditions.
Examples.
(1) Let be any set. For E , let cl E := E. This makes into a
(rather trivial) pregeometry.
(2) Let V be a (left) vector space over a division ring k. For E V ,
put cl E := k-linear span of E. This makes V into a pregeometry.
(3) Let K be a eld. For E K, let cl E be the set of all a K that
are algebraic over the subeld of K generated by E. This makes K
into an pregeometry.
Let be a pregeometry as above, and E . Condition (2) yields
E E

= cl E cl E

.
We call E closed if cl E = E. The intersection (inside ) of any collection
of closed subsets of is also closed, and cl E is the smallest closed subset of
that contains E.
We say that E is independent if a / cl(E a) for all a E, that E
generates or spans if cl E = , and that E is a basis of if E is both
independent and generates . It is easy to show that the following three
conditions are equivalent:
(1) E is a maximal independent subset of ;
(2) E is a minimal generating set of ;
(3) E is a basis of .
For example, to show that (2) = (3), prove rst that if E spans and
F E is independent, then has a basis B such that F B E. (Use
Zorn.) Special cases: if F is independent, then it is a subset of a basis of ,
and if E spans , then E has a subset that is a basis of . In particular,
has a basis. We leave the proof of the following exchange lemma as an
exercise.
Lemma 12.1. Let E and F be bases of and e E F. Then there is an
f F such that (E e) f is also a basis of .
54
Now a key result:
Proposition 12.2. All bases of have the same size.
Proof. Let E and F be bases of . Consider rst the case that F is nite.
We claim that then E is nite. To see why, note that each f F lies
in cl(E
f
) for some nite E
f
E, so F cl(E

) with nite E

E, so
E = cl(F) cl(E

). Since E is independent, this gives E = E

, so E
is nite. With E and F both nite, one uses the exchange lemma above to
obtain [E[ = [F[. It remains to consider the case that E and F are both
innite. Then we argue as before: for each f F we take a nite E
f
E
such that f cl(E
f
). Then
F
_
fF
cl E
f
cl
_
_
fF
E
f
_
= cl E

, E

:=
_
fF
E
f
,
so E = cl F cl E

with E

E, so E

= E. Hence [E[ = [E

[
[F[
0
= [F[. Likewise, [F[ [E[.
The rank of the pregeometry, denoted by rk , is the size of any basis of .
Some people call it the dimension of the pregeometry, but in some cases, like
projective geometry, this conicts with more natural notions of dimension,
so the neutral term rank is to be preferred.
Note that in example (1) the set is itself a basis, so rk = [[, and
in example (2), the notions of spanning set, independent set, and basis are
the familiar ones in vector spaces, and thus rk = dim
k
. In example (3),
independent means algebraically independent, so rk is the transcendence
degree of the eld over its prime eld.
Let X . Then we consider X as a pregeometry with respect to the
closure operation E cl(E) X : T(X) T(X). Note that a set E X
is independent in the pregeometry X if and only if E is independent in ,
so, with harmless ambiguity:
rk X = size of any maximal independent subset of X.
We also have the pregeometry [X (Omega over X) which has as its
underlying set, and closure operation
E cl(E X) : T() T().
A set F that is independent in the pregeometry [X is also said to be
independent over X or X-independent. If E is a basis of X and F a basis
of [X, then E F = , and E F is a basis of . Thus
rk = rkX + rk [X (additivity of rank)
We also dene rk(E[X) to be the rank of the pregeometry (E X)[X.
Note that any maximal X-independent subset of E is a basis of (E X)[X,
so rk(E[X) is the size of any maximal X-independent subset of E. The
pregeometry [X is also referred to as the localization of at X.
55
From a pregeometry to a geometry. A (combinatorial) geometry is a
pregeometry such that cl = and cla = a for each a . The
pregeometry on a set with cl E = E for each E is a geometry. But
the pregeometry of a vector space over a division ring is not a geometry,
since it has cl = 0.
Let be a pregeometry. By a line of we mean a set cla with a
cl . Note that any two distinct lines of have the trivial intersection
cl . The sets E, F are said to intersect nontrivially if their intersection
contains an element outside cl .
To the pregeometry we associate a geometry

with closure operation

cl

as follows: The points (elements) of

are the lines of ; for X

, the
union X is the set of all elements of that lie on some line p X, and we
dene cl

to be the set of all lines of that are contained in cl X.

Let E , and put E

:= cla : a E cl , that is, E

is the
set of lines of that intersect E nontrivially. It is easy to see that cl

is the set of lines of that intersect cl E nontrivially. Note that we have a

bijection E E

from the set of closed subsets E of onto the set of closed

subsets of

. Note also that if E is independent, then E

is independent.
In particular, rk E = rk E

for all E.
Let be the pregeometry of a vector space over a division ring k. Then the
points of

are the lines ka (a , a ,= 0), that is

= P(), the projective

space associated to the vector space . Note also that if E is closed,
that is, E is a k-linear subspace of , then E

= P(E) := ka : 0 ,= a E.
Thus our construction of a geometry from a pregeometry generalizes the
construction of the projective space associated to a vector space. We call

the projective geometry associated to the vector space . Note that if

as a vector space over k has dimension n, then rk = n = rk

, but
the projective space P() is regarded as a space of dimension n 1 for
n > 0. So our combinatorially dened rank does not always agree with
more geometrically inspired notions of dimension.
We also associate to a vector space V over a division ring k another com-
binatorial geometry, namely its ane geometry: dene a at to be either
a translate a + E of a linear subspace E or the empty subset of V . Then
the ats are the closed sets of the ane geometry of V , which has V as its
underlying set. Localizing this geometry at 0 we get back the pregeometry
of the vector space V .
The pregeometry on a strongly minimal set. We now return to the
setting of our monster model Mof T, but we do not assume that T is totally
transcendental unless we say so.
Let M
x
be innite and type-denable over A, such that for each denable
X M
x
, either X is nite, or X is nite. (For A-denable M
x
this just says that is strongly minimal, but there are situations of interest
where the weaker assumption is relevant.)
56
Then we have the following exchange lemma:
Lemma 12.3. Let a , b M
y
, and suppose b is Aa-algebraic, but not
A-algebraic. Then a is Ab-algebraic.
Proof. Take an A-denable relation R M
x
M
y
such that R(a) is nite,
say [R(a)[ n, and b R(a). By shrinking R we can arrange that [R(a

)[
n for all a

M
x
. If the set

R(b) is nite, then it is Ab-denable, which
in view of a

R(b) yields that a is Ab-algebraic, and we are done.
Assume

R(b) is innite, so [

R(b)[ = m, say. Representing as an
intersection of A-denable subsets of M
x
, one of these A-denable sets, call
it X, satises
X M
x
, X , [X

R(b)[ = m.
By shrinking R further we can arrange that for all b

M
y
,
either [X

R(b

)[ m, or

R(b

) = .
The set Y := b

M
y
: [X

R(b

)[ m is A-denable and contains b, so

Y is innite. Take distinct b
1
, . . . , b
n+1
Y . Then the sets

R(b
1
) , . . . ,

R(b
n+1
)
are conite in , so have a common element a

, hence [R(a

)[ n + 1 > n,
a contradiction.
This allows us to introduce a pregeometry on , but to do this we need to
view a set E M
x
as a parameter set. More precisely, we assign to such
E the parameter set [E] consisting of the components of the elements of
E; that is, if x = (x
i
)
iI
with each variable x
i
of sort s
i
, then [E] is the
parameter set given by
[E]
s
:=
_
{iI:s
i
=s}
e
i
: e E.
(Note that if E is not small, then [E] is not small.) When using terminology
like E-denable and E-algebraic for E M
x
we regard E as standing
for the parameter set [E]. Likewise, AE is the parameter set A[E].
For the rest of this section E ranges just over subsets of , and we dene
the closure operation cl
A
: T() T() by
cl
A
(E) := a : a is algebraic over AE.
Here we have xed the (small) parameter set A over which is type-
denable. Of course, is then also type-denable over any B A, and
this yields likewise a closure operation
cl
B
: T() T().
Theorem 12.4. The set with the closure operation cl
A
is a pregeometry,
denoted by
A
. For any a, b outside cl
A
(E) there is an f Aut(M[AE)
such that f(a) = b.
57
Proof. The rst three axioms dening pregeometries are obviously satised,
and the Exchange Axiom is satised because of Lemma 12.3 applied to AE
instead of A. Let a, b be outside cl
A
(E). It clearly suces to show:
Claim. tp(a[AE) = tp(b[AE). Suppose this claim fails. Then we have an
AE-denable X M
x
such that a X and b / X. We can assume that
X is nite. (Otherwise, interchange the roles of a and b, replacing X
by its complement in M
x
.) Then X is AE-denable, so a cl
A
(E), a
contradiction.
We let rk
A
E denote the rank of E in the pregeometry
A
, that is, rk
A
E
is the size of any maximal independent subset of E (independent in the
sense of the pregeometry
A
).
We have a global type p St
x
(M) associated to :
p := X Def
x
(M) : X is nite .
If is strongly minimal, then p is the global type determined by , and
MR(p) = 1. By the result above and its proof, an element a realizes
p AE i a / cl
A
(E). An n-tuple (a
1
, . . . , a
n
)
n
is said to be E-
independent if a
1
, . . . , a
n
are distinct and a
1
, . . . , a
n
is E-independent
in the pregeometry
A
that is, independent in the pregeometry
A
[E. For
E = we write independent instead of -independent.
Lemma 12.5. Let n 1, and suppose (a
1
, . . . , a
n
)
n
is E-independent
and (b
1
, . . . , b
n
)
n
is E-independent. Then there is an f Aut(M[AE)
such that f(a
i
) = b
i
for i = 1, . . . , n.
Proof. By the theorem above we can take a g Aut(M[AE) such that
g(a
1
) = b
1
. By replacing (a
1
, . . . , a
n
) by its image under g (and renaming)
we reduce to the case that a
1
= b
1
. If n > 1, use that (a
2
, . . . , a
n
) and
(b
2
, . . . , b
n
) are E a
1
-independent, and proceed inductively.
Note that for each n there are a
1
, . . . , a
n
such that (a
1
, . . . , a
n
) is inde-
pendent: just take a
i
to be outside cl
A
(a
1
, . . . , a
i1
) for i = 1, . . . , n.
Thus the pregeometry
A
has innite rank.
We now continue with the more restrictive strongly minimal case, that is,
in the rest of this section we assume:
is A-denable , MR() = 1, MD() = 1.
(But we do not assume that T is totally transcendental.) It follows by
induction on n and an exercise in Section 2.3 that MR(
n
) n.
Lemma 12.6. Let E be small. Then for all (a
1
, . . . , a
n
)
n
,
MR((a
1
, . . . , a
n
)[AE) = n (a
1
, . . . , a
n
) is E-independent.
58
Proof. By induction on n. The case n = 0 is trivial. Assume the lemma
holds for a certain n, and let a = (a
1
, . . . , a
n
, a
n+1
)
n+1
.
To obtain the direction , let a be E-independent, and let X
n+1
be AE-denable with a X; it remains to show that MR(X) = n + 1. By
identifying
n+1
with
n
we make X into a binary relation:
X
n
.
The set X(a
1
)
n
is denable over AEa
1
and contains the n-tuple
(a
2
, . . . , a
n+1
), which is (E a
1
)-independent, so MR
_
X(a
1
)
_
= n by the
inductive assumption. Hence MR
_
X(b)
_
= n for b outside cl
A
(E), and
there are innitely many such b. These b produce innitely many disjoint
denable subsets bX(b) of X of Morley rank n. Hence MR(X) = n+1.
For the direction , suppose a is not E-independent. By permutating
the coordinates of the (n + 1)-tuple a we arrange that a
n+1
is algebraic
over AEa

where a

:= (a
1
, . . . , a
n
), so we have an AE-denable relation
Y
n
=
n+1
such that a Y and Y (a

) is nite, say [Y (a

)[ = m.
We can shrink Y to arrange that [Y (b

)[ m for all b

n
. It follows that
MR(Y ) MR(
n
) n, and thus MR(a[AE) n.
Corollary 12.7. Given any a
1
, . . . , a
n
, we have
MR
_
(a
1
, . . . , a
n
)[A
_
= rk
A
a
1
, . . . , a
n
.
Proof. Suppose rk
A
a
1
, . . . , a
n
= k. By permutating the coordinates of the
tuple (a
1
, . . . , a
n
) we can arrange that (a
1
, . . . , a
k
) is independent, and that
a
k+1
, . . . , a
n
are algebraic over A(a
1
, . . . , a
k
). Then MR
_
(a
1
, . . . , a
n
)[A
_
=
MR
_
(a
1
, . . . , a
k
)[A
_
= k by Lemma 10.4 and Lemma 12.6.
Corollary 12.8. Let a = (a
1
, . . . , a
m
)
m
and b = (b
1
, . . . , b
n
)
n
, so
(a, b) := (a
1
, . . . , a
m
, b
1
, . . . , b
n
)
m+n
. Then
MR((a, b)[A) = MR(a[bA) + MR(b[A).
Proof. Let E = a
1
, . . . , a
m
and F = b
1
, . . . , b
n
. Then MR((a, b)[A) =
rk
A
(E F), MR(a[bA) = rk
bA
(E), and MR(b[A) = rk
A
(F). Now use that
rk
A
(E F) = rk
bA
(E) + rk
A
(F).

For denable Y
n
we can determine MR(Y ) inductively. To see how,
let n > 0 and use the identication
n
=
n1
to view Y as a binary
relation: Y
n1
. By an exercise in Section 10 we have an m such
that for all b
n1
, either [Y (b)[ m or [ Y (b)[ m. Put
Y

:= b
n1
: 1 [Y (b)[ m,
Y

:= b
n1
: [ Y (b)[ m.
Then Y

and Y

are denable, and Y is the disjoint union of (Y

)Y and
(Y

)Y , whose Morley ranks are MR(Y

) and MR(Y

)+1, respectively.
59
(Why?) Thus
MR(Y ) = maxMR(Y

), MR(Y

) + 1.
This argument can be done with parameters to obtain the denability of
Morley rank within denable families of subsets of
n
:
Corollary 12.9. Let X be a denable set in M and R X
n
a denable
relation. Then a X : MR
_
R(a)
_
d is denable for d = 0, . . . , n.
Proof. We proceed by induction on n. The case n = 0 being obvious, assume
n > 0. Identify
n
with
n1
and take m such that for all a X and
b
n1
, either [X(a)(b)[ m or [ X(a)(b)[ m. Put
R

:= (a, b) X
n1
: 1 [X(a)(b)[ m X
n1
,
R

:= (a, b) X
n1
: [ X(a)(b)[ m X
n1
.
By the argument preceding this corollary we have for each a X:
MR
_
R(a)
_
= maxMR
_
R

(a)
_
, MR
_
R

(a)
_
+ 1.
The desired result now follows by applying the inductive assumption to the
denable relations R

and R

.
Extension to algebraic elements. In this subsection we assume
T is totally transcendental.
An element h M
y
is said to be A-algebraic over if h is algebraic over
AE for some (nite) E. We extend some of our results on tuples in to
elements that are A-algebraic over .
Lemma 12.10. Suppose that h M
y
is A-algebraic over . Then there are
a = (a
1
, . . . , a
m
)
m
and b = (b
1
, . . . , b
n
)
n
such that the tuple
(a, b) := (a
1
, . . . , a
m
, b
1
, . . . , b
n
)
m+n
is independent in the pregeometry
A
, h is independent from a over A, and
h and b are interalgebraic over Aa (so MR(h[A) = MR(h[Aa) = n).
Proof. Take a nite E of minimal size such that h is algebraic over AE.
Note that then E is independent in
A
. Let a
1
, . . . , a
m
E be distinct
such that a
1
, . . . , a
m
is a basis of E in the pregeometry
Ah
, and let E =
a
1
, . . . , a
m
, b
1
, . . . , b
n
with [E[ = m+n. Then each b
i
cl
Ah
a
1
, . . . , a
m
,
so (b
1
, . . . , b
n
) is algebraic over Aa
1
. . . a
m
h. Hence h and (b
1
, . . . , b
n
) are
interalgebraic over Aa
1
. . . a
m
. Also,
MR
_
(a
1
, . . . , a
m
)[Ah
_
= m = MR
_
(a
1
, . . . , a
m
)[A
_
,
hence (a
1
, . . . , a
m
)
[

A
h, and thus h
[

A
(a
1
, . . . , a
m
), as desired.
In this lemma, with n = MR(h[A), we can take for (b
1
, . . . , b
n
) any element
of
n
that is independent in the pregeometry
Ah
, since all such n-tuples
are conjugate under Aut(M[Ah).
60
Lemma 12.11. Let g M
y
and h M
z
be A-algebraic over . Then
MR
_
(g, h)[A) = MR(g[hA) + MR(h[A).
Proof. By the previous lemma we can take a
p
and b
m
such that
g
[

A
a and g is interalgebraic with b over Aa,
and we can take c
q
and d
n
such that
h
[

A
c and h is interalgebraic with d over Ac.
We arrange that for (g, h) M
y
M
z
and (a, c)
p+q
we have
(g, h)
[

A
(a, c).
This is done as follows: Take an automorphism of M over Ag that sends a
to a realization of a nonforking extension of tp(a[Ag) to Agh, and replace a
and b by their images under this automorphism; this achieves
MR(a[Agh) = MR(a[Ag) = MR(a[A).
Next, take an automorphism of M over Ah that sends c to a realization of
a nonforking extension of tp(c[Ah) to Agha and replace c and d by their
images under this automorphism; this guarantees
MR(c[Agha) = MR(c[Ah) = MR(c[A).
Hence
MR((a, c)[A) = MR(c[Aa) + MR(a[A) = MR(c[Agha) + MR(a[Agh)
= MR((a, c)[Agh),
so (a, c)
[

A
(g, h), and thus (g, h)
[

A
(a, c) by symmetry. Then
MR
_
(g, h)[A) = MR
_
(g, h)[Aac
_
= MR((b, d)[Aac)
= MR(b[Aacd) + MR(d[Aac)
= MR(b[Aacd) + MR(h[Aac).
From (g, h)
[

A
(a, c) we obtain g
[

Ah
(a, c), so MR(g[Ah) = MR(g[Aach) =
MR(b[Aach) = MR(b[Aacd). Likewise, using h
[

A
(a, c) we get
MR(h[A) = MR(h[Aac).

A set Z M
z
is said to be almost strongly minimal with respect to , A if
Z is denable and each element of Z is A-algebraic over .
An easy saturation argument yields:
61
Lemma 12.12. A denable set Z M
z
is almost strongly minimal with
respect to , A if and only if there are k, N N and A-denable relations
R
1

n
1
M
z
, . . . , R
k

n
k
M
z
whose sections are all nite of size at
most N, such that each element of Z lies in a section of some R
i
.
With Lemmas 12.11 and 12.12 some properties of almost strongly minimal
sets follow easily from similar properties of denable subsets of
n
:
Exercises. Let Z M
z
be almost strongly minimal with respect to , A.
Show that MR(Z) < . Let also Y M
y
and R Y Z be denable.
Prove:
(1) g Y : MR
_
R(g)
_
d is denable;
(2) if MR
_
R(g)
_
= d for all g Y , then MR(R) = MR(Y ) +d.
13. Modularity
In this section we view the collection of closed sets in a pregeometry as a
lattice with respect to inclusion. The meet and join operations of this lattice
are given by
E F := E F, E F := cl(E F).
We begin with a short excursion on (modular) lattices in general.
Modular Lattices. A poset (partially ordered set) is said to be a lattice
if any elements a, b in it have a least upperbound a b (their join) and
a greatest lowerbound a b (their meet). For example, a boolean algebra
viewed as a poset is a lattice.
In this subsection P is a lattice, and a, b, c range over P. The partial
ordering of P can be recovered from its join operation as well as from its
meet operation:
a b a b = b a b = b.
The join and meet operations of P are idempotent (a a = a a = a),
commutative, associative, and satisfy the absorption identities:
a (a b) = a (a b) = a.
A sublattice of P is just a subset of P closed under and , and is thus a
lattice with respect to the induced partial ordering. For example, given a, c
we have the sublattice [a, c] := b : a b c of P. Note that for all a, b, c:
a c = a (b c) (a b) c.
The lattice P is said to be modular if for all a, b, c
a c = a (b c) = (a b) c.
This identity implies its own dual: if P is modular, then for all a, b, c
a c = a (b c) = (a b) c.
Of course, a sublattice of a modular lattice is modular as well.
62
Example. Let G be an abelian (additively written) group and consider the
lattice of all subgroups of G with respect to inclusion. Its join and meet
operations are given by A B = A + B, A B = A B. It is easy to
check that this lattice is modular. Thus the lattice of linear subspaces of a
vector space over a division ring is modular.
There are many results about modular lattices, but we develop here only
what is needed for our purpose. Let ^
5
be the lattice with exactly ve
distinct elements o, p, q, r, i with o as least element, i as the greatest element,
and p < q, p r = i, q r = o. This pentagon shaped lattice ^
5
is not
modular: q p, but q (r p) = q and (q r) p = p. The signicance of
^
5
is that modularity of P is equivalent to P not containing a copy of ^
5
:
Proposition 13.1. The following conditions on P are equivalent:
(1) P is modular;
(2) a (b c) = a
_
(b (a c)) c
_
for all a, b, c;
(3) P has no sublattice isomorphic to ^
5
.
Proof. For (1) (2), assume P is modular. Since a c c, this yields
(b (a c)) c = (b c) (a c),
hence a
_
(b (a c)) c
_
= a
_
(b c) (a c)
_
= a (b c).
For (2) (3), note that the identity of (2) fails in ^
5
:
q (r p) = q i = q and q
_
(r (q p)) p = q ((r q) p) = q p = p.
For (3) (1), suppose P is not modular. Take a, b, c such that a c but
a (b c) ,= (a b) c. We leave it as an exercise to show that
a b, (a b) c, a (b c), b c, b
are distinct, and are the elements of a sublattice of P isomorphic to ^
5
.
Given a second lattice Q, we make the cartesian product set P Q into
a lattice by dening (for d, e Q):
(a, d) (b, e) a b and d e,
so that (a, d) (b, e) = (a b, d e), (a, d) (b, e) = (a b, d e). The
following isomorphisms are very useful, and easily veried:
Proposition 13.2. Let P be modular. Then we have an isomorphism
s s b : [a b, a] [b, a b],
of sublattices, with inverse isomorphism
t t a : [b, a b] [a b, a].
Moreover, the map
(s, s

) s s

: [a b, a] [a b, b] [a b, a b]
is an isomorphism of the product lattice [ab, a] [ab, b] onto a sublattice
of [a b, a b], with inverse given by c (c a, c b).
63
We now return to the setting of a pregeometry with its lattice of closed sets.
Modular pairs. In this subsection E and F are closed sets of nite rank
in a pregeometry . Consider the inclusion diagram
.........................
Take bases B, C, D of E F, E[E F and F[E F, respectively. Then
B, C, D are nite, pairwise disjoint, B C is a basis of E, B D is a basis
of F, and B C D generates E F. Hence
rk E F rk(E F) + rk(E[E F) + rk(F[E F).
In view of
rk(E[E F) = rk(E) rk(E F) and rk(F[E F) = rk(F) rk(E F),
this inequality becomes
rk E + rkF rk(E F) + rk(E F),
in other words, rk(E[E F) rk(E F[F). This argument also shows:
rk E + rk F = rk(E F) + rk(E F) C is independent in [F
rk(E[E F) = rk(E F[F).
Let us call the pair E, F modular if
rk E + rkF = rk(E F) + rk(E F),
equivalently, rk(E[E F) = rk(E F[F). Note that if rk(E[E F) 1,
then E, F is modular. (This is because C F = .)
Next, let E F D E with D closed. Then the inclusion diagram
above consists of two subdiagrams:
.............
Take a basis C
1
of E[D, and a basis C
2
of D[E F, and let C = C
1
C
2
(a
disjoint union) be the corresponding basis of E[E F. Note that then C
1
generates E F[D F and C
2
generates D F[F. Hence E, F is modular
if and only if C
1
is a basis of E F[D F and C
2
is a basis of D F[F,
that is, if and only if rk(E[D) = rk(E F[D F) and D, F is modular.
Claim. rk(E[D) = rk(E F[D F) if and only if E (D F) = D and
E, D F is modular.
The if direction is clear, and the only if direction follows from
rk(E[D) rk(E[E (D F)) rk(E F[D F).
64
Corollary 13.3. Suppose E, F is modular. Consider the maps
up : [E F, E] [F, E F], up(D) := D F
down : [F, E F] [E F, E], down(G) := E G.
Then down up = the identity on [E F, E], in particular, up is injective.
If in addition E, G is modular for each G [F, E F], then both maps are
isomorphisms between sublattices of the lattice of closed sets in .
Proof. The arguments above proves the rst assertion. Let G [F, E F]
and put D = E G. Then D F G, and D, F is modular, so if E, G is
also modular, then necessarily G = D F, as is easily seen in an inclusion
diagram.
Modular pregeometries. Let be a pregeometry; we let E, F range over
subsets of . Modularity is a very strong condition, and appears in model
theory in the company of other properties that a pregeometry may or may
not have:
(1) Triviality: cl E =

aE
cla, for all nite E;
(2) Modularity: E, F is modular for all closed E and F of nite rank;
(3) Local modularity: E, F is modular for all closed E and F of nite
rank with E F ,= cl ;
(4) Local niteness: cl E is nite for all nite E;
(5) Homogeneity: for any nite E and a, b cl E the pregeometry
has an automorphism xing E pointwise and sending a to b.
In the rst example of the previous section, where cl E = E for all E, the
pregeometry is trivial, modular, locally nite, and homogeneous. The second
example, where is the pregeometry of a vector space over a division ring k,
is modular and homogeneous; it is also locally nite if k is a nite eld. If in
the third example the eld K is algebraically closed, then the pregeometry
is homogeneous. Note also that the pregeometry
A
of Theorem 12.4 is
homogeneous. If is the ane geometry of a vector space over a division
ring k, then is homogeneous, and locally nite if in addition k is nite.
Note that if is a geometry, then triviality of means that cl E = E for
all E.
Combinatorial geometries became important in model theory when Zilber
conjectured the following theorem, subsequently proved by him and inde-
pendently by Cherlin and by Mills:
Theorem 13.4. Suppose is a homogeneous locally nite geometry of in-
nite rank. Then one of the following happens:
(1) is trivial;
(2) is isomorphic to the projective geometry associated to a vector
space of innite dimension over a nite eld;
(3) is isomorphic to the ane geometry of a vector space of innite
dimension over a nite eld.
65
We do not prove this here, and just mention it to give some perspective on
what we are engaged in. The next result has a routine proof.
Lemma 13.5. Each of the ve properties listed above is inherited from by
the geometry

associated to , and by each localization [X with X .

Lemma 13.6. is modular if and only if its lattice of closed sets is modular.
Proof. The lattice of closed sets is modular i its sublattice of closed sets of
nite rank is modular. This follows easily from the denitions and the fact
that a closed set is the directed union of its closed subsets of nite rank.
Suppose now that is modular. To show that the lattice of closed sets of
nite rank is modular, assume towards a contradiction that this lattice has
a sublattice isomorphic to ^
5
. Then we have closed sets E, F of nite rank,
and a closed set D strictly between E F and E such that DF = E F,
but this contradicts Corollary 13.3.
For the converse, assume the lattice of closed sets is modular, and let E
and F be closed of nite rank. Let n := rk(E[F), so we have a strictly
increasing sequence
E F = D
0
D
1
D
n
= E
of closed sets. By Proposition 13.2 this yields a strictly increasing sequence
F = D
0
F D
1
F D
n
F = E F
of closed sets, hence rk(E F[F) n = rk(E[E F). Also
rk(E F[F) rk(E[E F),
and thus E, F is modular.
The next lemma has a routine proof.
Lemma 13.7. is locally modular if and only if each localization [a
with a cl is modular. If is homogeneous and [a is modular for
some a cl , then is locally modular.
Lemma 13.8. If E, F is modular for all closed E, F of nite rank with
rk(E[E F) = 2, then is modular.
Proof. We prove the contrapositive. Assume is not modular, so E, F is
not modular for certain closed E and F of nite rank; take such E and F
for which n := rk(E[E F) is minimal, so n 2 and rk(E F[F) < n. Let
C = c
1
, . . . , c
n
be a basis of E[E F, and put
E
1
:= cl((E F) c
1
, . . . , c
n1
,
so rk(E
1
[E F) = n 1. Also E
1
, F is modular by the minimality of n,
hence rk(E
1
F[F) = n 1, and thus E
1
F = E F. Next, let
E
2
:= cl((E F) c
1
, . . . , c
n2
,
so rk(E
1
[E
2
) = 1, and thus for G := E
2
F we have rk(E F[G) = 1 and
E
1
G = E
2
, by Corollary 13.3.
66
Claim. E G = E
2
.
To prove the claim, note that E
2
EG. Suppose E
2
is properly contained
in EG. Then rk(EG[EF) = n1, so EG, F is modular by minimality
of n. Hence rk(G[F) rk((E G) F[F) = rk(E G[E F) = n 1,
contradicting rk(G[F) = n 2. This proves the claim.
It follows that E, G is not modular, since rk(E[E
2
) = 2 and rk(EG[G) =
rk(E F[G) = 1. Thus E, G has the desired property.
We can go further along these lines:
Lemma 13.9. If E, F is modular for all closed E, F of nite rank with
rk(E[E F) = 2 = rk(F[E F), then is modular.
Proof. Again, we prove the contrapositive, and assume is not modular.
By the previous lemma we can take closed E, F of nite rank such that
E, F is not modular and rk(E[E F) = 2. Take such E, F with minimal
m := rk(F[E F), so m 2. Let b
1
, . . . , b
m
be a basis of F[E F, let
H := cl(E b
1
). Then
rk(H[E F) = rk(H[E) + rk(E[E F) = 3
= rk(H[H F) + rk(H F[E F),
so rk(H F[E F) equals 1 or 2. It it equals 1, then rk(H[H F) = 2 and
rk(F[HF) = m1, and H, F is nonmodular, contradicting the minimality
of m. So rk(H F[E F) = 2, and thus E, H F is nonmodular of the
desired form.
Lemma 13.10. The following are equivalent:
(1) is modular;
(2) for all closed E and F such that rk E = 2, F has nite rank, and
rk(E F[F) = 1 we have E F ,= cl ;
(3) for each nonempty closed E and b we have
cl(E b) =
_
aE
cla, b;
(4) for all nonempty closed E, F we have
E F =
_
aE,bF
cla, b.
Proof. The direction (1) =(2) is clear. For the converse, suppose is not
modular. Then Lemma 13.8 gives closed E, F of nite rank with
rk(E[E F) = 2, rk(E F[F) = 1.
Let a, b be a basis of E[E F and put E

:= cla, b. Then rk E

= 2
and rk(E

[E

F) = 2, so E

F = cl . Also E

F = E F, so
rk(E

F[F) = 1. This contradicts (2), and proves (the contrapositive of)

(2) = (1). For (1) = (3), assume (1), let E be closed and nonempty,
67
b and c cl(E b). To nd a E such that c cla, b we can
assume E has nite rank and c / clb. By modularity,
rk(E b, c) = rk E + rkb, c rk(E clb, c), and
rk(E b, c) = rk(E b) = rk E + rkb rk(E clb).
Also, rkb, c = rkb +1, so E clb is properly contained in E clb, c.
Take a E clb, c such that a / clb. Then c cla, b by exchange.
To prove (3) =(4), assume (3), and let E, F be closed, nonempty, and,
without loss of generality, of nite rank. Let c E F. We have to nd
a E and b F such that c cla, b. We proceed by induction on rk F.
If rk F = 0, we take any b F, and apply (2). Next, let F = cl(F
0
b
1
)
with closed F
0
and rkF = rk F
0
+ 1. Then by (3) we have an a
0
E F
0
such that c cla
0
, b
1
. The inductive assumption gives a E and b
0
F
0
such that a
0
cla, b
0
. Then c cla, b
0
, b
1
cl(F a), so by (3) there
is b F such that c cla, b.
We leave (4) =(1) to the reader.
14. Modularity, One-basedness, and Linearity
We now return to the setting of a totally transcendental theory T with
monster model M and a strongly minimal A-denable set M
z
. This
gives the pregeometry
A
with closure operation cl
A
. We shall prove in this
section that local modularity of
A
is equivalent to various other conditions.
Given a tuple a = (a
1
, . . . , a
n
)
n
and E , we put
cl
A
(a) := cl
A
a
1
, . . . , a
n
, a subset of ,
rk
A
(a[E) := rk
A
(a
1
, . . . , a
n
[E).
Lemma 14.1. The following are equivalent:
(1)
A
is modular;
(2) for all a
m
and b
n
: a
[

B
b, where B := A
_
cl
A
(a) cl
A
(b)
_
.
Proof. Recall rst that by our notational conventions
rk
A
(a[E) = rk
A
_
cl
A
(a)[E
_
, (a
n
, E ).
Suppose
A
is modular, and let a
m
and b
n
. Then
rk
A
_
a[ cl
A
(a) cl
A
(b)
_
= rk
A
_
(a, b)[ cl
A
(b)
_
= rk
A
_
a[ cl
A
(b)
_
,
hence MR
_
a[A(cl
A
(a) cl
A
(b))
_
= MR(a[Acl
A
(b)) by Corollary 12.7, that
is, a
[

B
b, with B = A
_
cl
A
(a) cl
A
(b)
_
.
Suppose next that
A
is not modular. Then by Lemma 13.8 there are a
2
and b
n
such that rk
A
_
a[ cl
A
(a) cl
A
(b)
_
= 2 but rk
A
_
a[ cl
A
(b)
_
< 2.
Hence
MR
_
a[A(cl
A
(a) cl
A
(b))
_
= 2, MR
_
a[Acl
A
(b)
_
< 2,
68
so a
[

B
b fails, with B as above.
One-basedness. In the rest of this section T has EI, in addition to being
totally transcendental. This assumption is just for the convenience of having
simpler statements of some denitions and results. (To avoid assuming that
T has EI one would have to work in M
eq
instead of M.)
Consider a parameter set P in M that is A-invariant, that is, f(P) = P
for all f Aut(M[A). (We do not assume P is small.) We say that P is
one-based over A if for each P-tuple a and each B A such that tp(a[B)
is stationary we have cb(a[B) acl(Aa). Clearly, if P is one-based over A
and A A

, then P is one-based over A

.
Lemma 14.2. The following are equivalent:
(1) P is one-based over A;
(2) for each P-tuple a and each B A we have
a
[

C
B, where C := acl(Aa) acl(B).
Proof. Assume (1), let a be a P-tuple, and let B A. Since tp(a[ acl(B)) is
stationary, we have cb(a[ acl(B)) C := acl(Aa)acl(B). Hence tp(a[ acl(B))
does not fork over C, that is, a
[

C
acl(B), and thus a
[

C
B.
The converse follows by similar reasoning.
Lemma 14.3. If P is one-based over A, then acl(P) is also one-based over
A. If A

A and P is one-based over A

, then P is one-based over A.

Proof. Assume P is one-based over A. Let a be an acl(P)-tuple such that
tp(a[B) is stationary, where B A. To obtain cb(a[B) acl(Aa) we can
assume B is algebraically closed, replacing B by acl(B) if necessary. Take
a P-tuple b such that a is b-algebraic. We can arrange that b
[

Aa
B, by
replacing b by a realization of a nonforking extension of tp(b[Aa) to AaB.
Then C := acl(Ab) B acl(Aa) by Lemma 11.8, so C = acl(Aa) B. By
the lemma above, b
[

C
B, hence a
[

C
B, so cb(a[B) C acl(Aa).
Next, assume P is one-based over A

, where A

A. Let a be a P-tuple
and let B A be such that tp(a[B) is stationary. We can arrange that B
is algebraically closed, and that A

A
Ba, so A

AB
a, that is, A

B
a, so
a
[

B
A

, that is, tp(a[A

B) does not fork over B. So cb(a[B) = cb(a[A

B).
By assumption, cb(a[A

B) acl(A

a), hence cb(a[B) acl(A

a) acl(B).
Since B
[

Aa
A

a, it follows that cb(a[B) acl(Aa).

Lemma 14.4. Suppose
A
is modular. Then is one-based over A.
69
Proof. Let a
n
and let B A be such that tp(a[B) is stationary. Take b
such that cb(a[B) = dcl(b). By Lemma 11.17 the tuple b is denable over a
tuple c
m
, and we arrange in the usual way that c
[

Ab
a, so a
[

Ab
c. Note
that then
cb(a[B) = cb(a[Ab) = cb(a[Abc) = cb(a[Ac).
By the modularity assumption and Lemma 14.1 we have
a
[

C
c, where C := A(cl
A
(a) cl
A
(c)).
Note that Ac Cc acl(Ac), so
cb(a[Ac) = cb(a[Cc) = cb(a[C) dcl(C) acl(Aa),
hence cb(a[B) acl(Aa).
Linearity. We say that is A-linear if for all , and each algebraically
closed B A such that MR((, )[B) = 1 there is a c such that MR(c[A) 1
and tp((, )[B) is dened over c.
Note that if is A-linear, , , B A is algebraically closed,
MR((, )[B) = 1, and cb((, )[B) = dcl(c), then necessarily MR(c[A) 1.
Note also that if is A-linear and A A

, then is A

-linear.
Lemma 14.5. The following are equivalent:
(1) is A-linear;
(2) for all , and each algebraically closed B A there is a c such
that tp((, )[B) is denable over c and c is algebraic over A.
Proof. Assume is A-linear, let , , and let B A be algebraically
closed. The type tp((, )[B) is stationary, so it has a unique global nonfork-
ing extension p. Take c such that dcl(c) = cb(p). Thus p and its restriction
tp((, )[B) are denable over c. We shall prove that c is algebraic over A.
If MR((, )[B) = 2, then MR((, )[A) = 2, so p does not fork over A,
hence c is algebraic over A by Corollary 10.13(1). If MR((, )[B) = 0, then
(, ) is the only realization of tp((, )[B), so c is interdenable with (, ).
Suppose MR((, )[B) = 1, the remaining case. Then by the A-linearity of
we have MR(c[A) 1.
If MR((, )[A) = 1, then again c is algebraic over A. So we can as-
sume MR((, )[A) = 2. Then (, ) [

/
A
B, hence (, ) [

/
A
c : otherwise,
MR((, )[c) 2, but p is a nonforking extension of tp((, )[c), so MR(p) =
2, hence MR((, )[B) = 2, a contradiction. Therefore, c is not alge-
braic over A, and thus MR(c[A) = 1. Also c [

/
A
(, ) by symmetry, so
MR(c[A) > MR(c[A, so MR(c[A) = 0, hence c is algebraic over A.
For the converse, assume (2), let , , and let B A be a parameter
set in M such that MR((, )[B) = 1. The type tp((, )[B) is stationary,
so it has a unique global nonforking extension p. Take c such that dcl(c) =
cb(p). Then p and its restriction tp((, )[B) are denable over c, so c is
70
algebraic over B. By (2) the tuple c is also algebraic over A, so MR(c[A)
MR((, )[A) 2. We shall prove that MR(c[A) 1. Assume towards a
contradiction that MR(c[A) = MR((, )[A) = 2. By Lemma 11.11,
MR((, , c)[A) = MR((, )[Ac) + MR(c[A) = MR(c[A) + MR((, )[A),
so MR((, )[Ac) = MR(c[A) = 0. But c is algebraic over B, so
MR((, )[Ac) MR((, )[B) 1,
a contradiction.
Lemma 14.6. Suppose is A-linear. Then
A
is locally modular.
Proof. Take a
0
cl
A
. We claim that
A
[a
0
is modular. By 12.10 is
is enough to show:
Let a
1
, a
2
and rk
A
(a
1
, a
2
[a
0
) = 2, and let F be closed in
A
of nite rank with a
0
F and rk
A
(a
1
, a
2
[F) = 1. Then
cl
A
a
0
, a
1
, a
2
F ,= cl
A
a
0
.
By Lemma 11.17 we can take c such that dcl(c) = cb((a
1
, a
2
)[ acl(AF)),
so c is algebraic over AF. By A-linearity and the previous lemma, c is also
algebraic over Aa
1
a
2
. Note that MR((a
1
, a
2
)[Ac) = 1 and a
i
is not algebraic
over Ac for i = 1, 2. So a
2
is algebraic over Aca
1
. Also, a
0
is not algebraic
over Aa
1
a
2
, so a
0
is not algebraic over Ac. Hence a
0
and a
1
are conjugate
over Ac, so (a
0
, c) and (a
1
, c) are conjugate over A. Take a

2
such that
(a
0
, a

2
, c) and (a
1
, a
2
, c) are conjugate over A. Then a

2
is algebraic over
Aca
0
, so a

2
F and a

2
/ cl
A
a
0
. Since a

2
cl
A
a
0
, a
1
, a
2
, this completes
the proof.
We can summarize most of the above as follows.
Theorem 14.7. The following conditions on , A are equivalent:
(1) there is B A such that
B
is modular;
(2) is one-based over A;
(3) is A-linear;
(4)
A
is locally modular.
Proof. The direction (1) = (2) follows from Lemmas 14.3 and 14.4. The
direction (2) =(3) follows from Lemma 14.5. The direction (3) =(4) is
Lemma 14.6. For (4) = (1), assume
A
is locally modular. Take a
such that
A
[a is modular. Let B := Aa, and note that the closure
operations of
B
and
A
[a coincide. Thus
B
=
A
[a is modular.
Condition (1) in this theorem raises the question whether
B
is always of the
form
A
[E for a suitable E . An obvious candidate to try is E = cl
B
.
Linearity and plane curves. A family of curves in
2
is a pair (X, C)
where X M
x
and C X
2
are denable such that C(a) is strongly
minimal for all a X.
71
Let (X, C) be a family of curves in
2
. We think of
2
as the plane, and
of each section C(a)
2
as a plane curve. Note that for all a, b X, either
C(a) C(b) is nite, or C(a)C(b) is nite. Thus by saturation, there is a
natural number N such that for all a, b X, either [C(a) C(b)[ N, or
[C(a)C(b)[ N. In the former case we wish to consider C(a) and C(b)
as essentially dierent curves, and in the latter case as essentially the same
curve. How many essentially dierent curves are there in this family? To
make sense of this question we introduce the (denable) equivalence relation
on X by:
a b C(a)C(b) is nite.
Since T has EI we have a denable surjective map f : X X

with X

a
denable set in M, such that is the kernel of f. Thus MR(X

) does not
depend on the choice of f, X

. We call MR(X

) the essential Morley rank of

(X, C) and think of it as a rough measure of how many essentially dierent
curves there are in the family.
Examples. In these examples we just consider one-sorted M.
(1) Let M = be an innite set (no further structure), and let C

2
be given by C = (b, , )
3
: = b, so each section C(b)
is a horizontal line. Then (, C) is a family of curves in
2
, the
equivalence relation on the parameter space is just equality, so
the essential Morley rank of the family is 1.
(2) Let M be an innite vector space over a division ring k with under-
lying set , x a scalar k

, and let C
2
be given by
C = (b, , )
3
: = + b. Then (, C) is a family of curves
in
2
, the equivalence relation on the parameter space is just
equality, so the essential Morley rank of the family is 1.
(3) Let M be an algebraically closed eld with underlying set (so M,
being large, is actually of innite transcendence degree). Then
C := (a, b, , )
2

2
: = a +b.
Then (
2
, C) is a family of curves in
2
, the equivalence relation
on the parameter space
2
is just equality, so the essential Morley
rank of the family is 2.
The signicance of A-linearity of is that there are few curves in
2
. The
following result makes one direction of this precise. (There is also a sort of
converse.)
Proposition 14.8. Suppose is A-linear. Then every family of curves in

2
has essential Morley rank 1.
Proof. Let (X, C) be a family of curves in
2
, X M
x
. To keep notations
simple we assume that , X and C are all 0-denable. (Why is this no
loss of generality?) Let f : X X

be a 0-denable map with kernel

as above. Let a X, and let (, ) be a generic point of C(a) over a, so
MR((, )[a) = MR((, )[ acl(a)) = 1. Let cb((, )[ acl(a)) = dcl(c). Then
72
MR(c[0) 1 by linearity. Since tp((, )[ acl(a)) is denable over c, there is
an L
c
-formula (x) such that for all b X we have:
C(b) tp((, )[ acl(a)) [= (b).
The set C(a) determines the type tp((, )[ acl(a)), in the sense of Section
9, so for all b X we have:
C(b) tp((, )[ acl(a)) b a f(b) = f(a).
So f(a) is the unique element of X

equal to f(b) for some b such that

[= (b). Hence f(a) is c-denable, so MR(f(a)[0) MR(c[0) 1. Since
a X was arbitrary, this yields MR(X

) 1.