A Crèche Course in Model Theory

Domenico Zambella
Università di Torino

April 2018

Preface 4

1 Preliminaries and notation 5

1 Structures 5
2 Tuples 6
3 Terms 7
4 Substructures 8
5 Formulas 9
6 Yet more notation 11

2 Theories and elementarity 13

1 Logical consequences 13
2 Elementary equivalence 15
3 Embeddings and isomorphisms 17
4 Quotient structures 19
5 Completeness 20
6 The Tarski-Vaught test 21
7 Downward Löwenheim-Skolem 22
8 Elementary chains 23

3 Ultraproducts 25
1 Filters and ultrafilters 25
2 Direct products 26
3 Łoś’s Theorem 27

4 Compactness 30
1 Compactness via syntax 30
2 Compactness via ultraproducts 31
3 Upward Löwenheim-Skolem 32
4 Finite axiomatizability 33

5 Types and morphisms 35

1 Semilattices and filters 35
2 Distributive lattices and prime filters 37
3 Types as filters 39
4 Morphisms 41

1
6 Some relational structures 44
1 Dense linear orders 44
2 Random graphs 46
3 Notes and references 49

7 Fraïssé limits 50
1 Rich models. 50
2 Weaker notions of universality and homogeneity 55
3 The amalgamation property 56

8 Some algebraic structures 59

1 Abelian groups 59
2 Torsion-free abelian groups 60
3 Divisible abelian groups 61
4 Commutative rings 63
5 Integral domains 64
6 Algebraically closed fields 65
7 Hilbert’s Nullstellensatz 66

9 Saturation and homogeneity 69

1 Saturated structures 69
2 Homogeneous structures 71
3 The monster model 72

10 Preservation theorems 77
1 Lyndon-Robinson Lemma 77
2 Quantifier elimination by back-and-forth 79
3 Model-completeness 81

11 Geometry and dimension 83

1 Algebraic and definable elements 83
2 Strongly minimal theories 85
3 Independence and dimension 86

12 Countable models 89
1 The omitting types theorem 89
2 Prime and atomic models 91
3 Countable categoricity 92
4 Small theories 94
5 A toy version of a theorem of Zil’ber 96
6 Notes and references 97

13 Definability and automorphisms 98

1 Many-sorted structures 98
2 The eq-expansion 99
3 The definable closure in the eq-expansion 101
4 The algebraic closure in the eq-expansion 102
5 Elimination of imaginaries 103

2
 6 Imaginaries: the true story 106
7 Uniform elimination of imaginaries 106
8 Notes and references 108

14 Invariant sets 109

1 Invariant sets and types 109
2 Invariance from the dual perspective 110
3 Morley sequences and indiscernibles 111
4 From coheirs to Ramsey to indiscernibles 113

15 Lascar invariant sets 115

1 Expansions 115
2 Lascar strong types 116
3 The Lascar graph and Newelski’s theorem 118
4 Kim-Pillay types 121
5 Notes and references 123

16 Externally definable sets 124

1 Approximable sets 124
2 Ladders and definability 126
3 Vapknik-Chevronenkis dimension 128
4 Honest definitions 129
5 Stable theories 130
6 Stability and the number of types 132
7 Notes and references 133

17 Topological dynamics 134

1 T.b.c. 134

18 Keisler measures 135

1 T.b.c. 135

3
 Preface

This book was written to answer one question

“Does a recursion theorist dare to write a book
on model theory?”
Gerald E. Sacks
Saturated Model Theory (1972)

These are the notes of a course that I have given for a few years in Amsterdam and
many more in Turin. A few chapters are missing and will be added soon. I keep
the most recent version in
https://github.com/domenicozambella/creche
B A warning sign in the margin indicates that the notation is nonstandard. Occasion-
ally, the whole exposition is substantially nonstandard. Below are a few examples.
. Fraïssé limits are presented in a general setting that accommodates a large vari-
ety of examples (and I’ll add a few more). This setting is used, for example, to
discuss saturation.
. Quantifier elimination for ACF and Hilbert’s Nullstellensatz are presented in
more detail than is usual. (This may annoy some readers, but I hope it will help
others.)
. The proof of the Omitting Types Theorem uses a model theoretic construction.
. Imaginaries and the eq-expansion are introduced from the (equivalent) dual per-
spective that the canonical name of a definable set is the set itself.
. Ramsey’s Theorem is derived from the existence of coheir sequences. This is not
the shortest proof, but it should be interesting.
. Lascar and Kim-Pillay types are introduced in a slightly unconventional way.
. Newelski’s Theorem on the diameter of Lascar types is proved in an elementary
self-contained way.
. Stability and NIP are introduced very briefly. We only discuss the properties of
externally definable sets, which we identify with approximable sets.

4
 Chapter 1
Preliminaries and notation
This chapter introduces the syntax and semantic of first order logic. We assume
that the reader has at least some familiarity with first order logic.
The definitions of terms and formulas we give in Section 3 and 5 are more formal
than is required in subsequent chapters. Our main objective is to convince the
reader that a rigorous definition of language and truth is possible. However, the
actual details of such a definition are not relevant for our purposes.

1 Structures
A (first order) language L (also called signature ) is a triple that consists of
1. a set Lfun whose elements are called function symbols ;

2. a set Lrel whose elements are called relation symbols ;

3. a function that assigns to every f ∈ Lfun , respectively r ∈ Lrel , non-negative inte-

gers n f and nr that we call arity of the function, respectively relation, symbol.
We say that f is an n f -ary function symbol, and similarly for r. A 0-ary function
symbol is also called a constant .
Warning: it is customary to use the symbol L to denote both the language and the
set of formulas (to be defined below) associated to it. The cardinality of Lfun ∪ Lrel ∪
ω is denoted by | L| .
A (first order) structure M of signature L (for short L-structure ) consists of
1. a set that we call the domain or support and denote by the same symbol M
used for the structure as a whole;
2. a function that assigns to every f ∈ Lfun a total map f M : M n f → M ;

3. a function that assigns to every r ∈ Lrel a relation r M ⊆ M nr .

We call f M , respectively r M , the interpretation of f , respectively r, in M.

Recall that, by definition, M0 = {∅}. Therefore the interpretation of a constant c is
a function that maps the unique element of M0 to an element of M. We identify c M
with c M (∅).
We may use the word model as a synonym for structure. But beware that, in some
contexts, the word is used to denote a particular kind of structure.
If M is an L-structure and A ⊆ M is any subset, we write L( A) for the language
obtained by adding to Lfun the elements of A as constants. In this context, the
elements of A are called parameters . There is a canonical expansion of M to an
L( A)-structure that is obtained by setting a M = a for every a ∈ A.

1.1 Example The language of additive groups consists of the following function sym-

5
 bols:
1. a constant (that is, a function symbol of arity 0) 0

2. a unary function symbol (that is, of arity 1) −

3. a binary function symbol (that is, of arity 2) + .

In the language of multiplicative groups the three symbols above are replaced by
1, −1 , and · respectively. Any group is a structure in either of these two signa-
tures with the obvious interpretation. Needless to say, not all structures with these
signatures are groups.
The language of (unitary) rings contains all the function symbols above except −1 .

The language of ordered rings also contains the binary relation symbol <. 

The following example is less straightforward. The reason for the choice of the
language of vector spaces will become clear in Example 1.9 below.

1.2 Example Let F be a field. The language of vector spaces over F , which we denote
by L F , extends that of additive groups by a unary function symbol k for every k ∈ F.
Recall that a vector space over F is an abelian group M together with a function
µ : F × M → M satisfying some properties (that we assume well-known, see Exam-
ple 2.4). To view a vector space over F as an L F -structure, we interpret the group
symbols in the obvious way and each k ∈ F as the function µ(k, -). 

The languages in Examples 1.1 and 1.2, with the exception of that of ordered rings,
are functional languages , that is, Lrel = ∅. In what follows, we consider two
important examples of relational languages , that is, languages where Lfun = ∅.

1.3 Example The language of strict orders only contains the binary relation symbol <.
The language of graphs , too, only contains a binary relation symbol (for which
there is no standard notation). In combinatorics, a graph is a set of unordered
pairs. To view a graph as a first order structure, we identify sets of unordered pairs
with symmetric irreflexive relations. 

2 Tuples
A tuple of elements of A is a function a : α → A from an ordinal α to a set A. We
call α the length of a and denote it by | a| . If a is surjective, it is said to be an
enumeration of A. Occasionally, we may say sequence for tuple – a word that is
also used when the domain is not necessarily an ordinal.
We write ai for the i-th element of a, that is, the element a(i ), where i < α. We may
use the notation a = h ai : i < αi for a tuple; when α is finite, we may also write
a = a0 , . . . , aα−1 . We write an for the restriction of a to n.
B In general, for any I ⊆ α, we write a I for the tuple h aik : k < ni where i0 , . . . , in−1
are the elements if I in increasing order.
B We sometimes overline the symbol that denotes a tuple. For ease of notation, the
elements of the tuple ā are denoted by ai (without the bar). Similarly for an and
a I .

6
 The set of tuples of elements of length α is denoted by Aα . The set of tuples of
length < α is denoted by A<α . For instance, A<ω is the set of all finite tuples
of elements of A. When α is finite we do not distinguish between Aα and the α-th
Cartesian power of A. In particular, we do not distinguish between A1 and A.
If a, b ∈ A|α| and h is a function defined on A, we write h( a) = b for h( ai ) = bi . We
often do not distinguish between the pair h a, bi and the tuple of pairs h ai , bi i. The
context will resolve the ambiguity.
Note that there is a unique tuple of length 0, the empty set ∅, which in this context is
called empty tuple . Recall that by definition A0 = {∅} for every set A. Therefore,
even when A is empty, A0 contains the empty string.
We often concatenate tuples. If a and b are tuples, we write a b or, equivalently,
a, b .

3 Terms
Let V be an infinite set whose elements we call variables . We use the letters x, y,
z, etc. to denote variables or tuples of variables. We rarely refer to V explicitly, and
we always assume that V is large enough for our needs.
We fix a signature L for the whole section.

1.4 Definition A term is a finite sequence of elements of Lfun ∪ V that are obtained induc-
tively as follows:
o. every variable, intended as a tuple of length 1, is a term;

i. if f ∈ Lfun and t is a tuple obtained by concatenating n f terms, then f t is a term. We

write f t for the tuple obtained by prefixing t by f .
We say L-term when we need to specify the language L. 

Note that any constant f , intended as a tuple of length 1, is a term (by i, the term
f is obtained concatenating n f = 0 terms and prefixing by f ). Terms that do not
contain variables are called closed terms .
The intended meaning of, for instance, the term + + x y z is ( x + y) + z. The first
expression uses prefix notation ; the second uses infix notation . When convenient,
we informally use infix notation and add parentheses to improve legibility and
avoid ambiguity.
The following lemma shows that prefix notation allows to write terms unambigu-
ously without using parentheses.

1.5 Lemma (unique legibility of terms) Let a be a sequence of terms. Suppose a can be ob-
tained both by concatenating the terms t1 , . . . , tn and by concatenating the terms s1 , . . . , sm .
Then n = m and si = ti .
Proof By induction on | a|. If | a| = 0 than n = m = 0 and there is nothing to prove.
Suppose the claim holds for tuples of length k and let a = a1 , . . . , ak+1 . Then a1 is
the first element of both t1 and s1 . If a1 is a variable, say x, then t1 and s1 are the
term x and n = m = 1. Otherwise a1 is a function symbol, say f . Then t1 = f t̄
and s1 = f s̄, where t̄ and s̄ are obtained by concatenating the terms t10 , . . . , t0p and

7
 s10 , . . . , s0p . Now apply the induction hypothesis to a2 , . . . , ak+1 and to the terms
t10 , . . . , t0p , t2 , . . . , tn and s10 , . . . , s0p , s2 , . . . , sm . 

If x = x1 , . . . , xn is a tuple of distinct variables and s = s1 , . . . , sn is a tuple of terms,

we write t[ x/s] for the sequence obtained by replacing x by s coordinatewise.
Proving that t[ x/s] is indeed a term is a tedious task that can be safely skipped.
If t is a term and x1 , . . . , xn are (tuples of) variables, we write t( x1 , . . . , xn ) to declare
that the variables occurring in t are among those that occur in x1 , . . . , xn . When a
term has been presented as t( x, y), we write t(s, y) for t[ x/s].
Finally, we define the interpretation of a term in a structure M. We begin with closed
terms. These are interpreted as 0-ary functions, i.e. as elements of the structure.

1.6 Definition Let t be a closed L( M ) term. The interpretation of t , denoted by t M , is

defined by induction of the syntax of t as follows.
i. if t = f t̄, where f ∈ Lfun and t̄ is a tuple obtained by concatenating the terms
t1 , . . . , tn f , then t M = f M (t1M , . . . , tnMf ).
Note that in i we have used Lemma 1.5 in an essential way. In fact this ensures that the
sequence t̄ uniquely determines the terms t1 , . . . , tn f . 

The inductive definition above is based on the case n f = 0, that is, the case where
f a constant, or a parameter. When t = c, a constant, t̄ is the empty tuple, and so
t M = c M (∅), which we abbreviate as c M . In particular, if t = a, a parameter, then
t M = a M = a.
Now we generalize the interpretation to all (not necessarily closed) terms. If t( x ) is
a term, we define t M( x ) : M| x| → M to be the function that maps a to t( a) M .

4 Substructures
1.7 Definition Fix a signature L and let M and N be two L-structures. We say that M is a
substructure of N, and write M ⊆ N , if
1. the domain of M is a subset of the domain of N

2. f M = f N  M n f for every f ∈ Lfun

3. r M = r N ∩ Mnr for every f ∈ Lrel .

Note that when f is a constant 2 becomes f M = f N , in particular the substructures

of N contains at least all constants of N.
If a set A ⊆ N is such that
1. f N [ An f ] ⊆ A for every f ∈ Lfun

then there is a unique substructure M ⊆ N with domain A, namely, the structure

with the following interpretation
2. f M = f N  M n f (which is a good definition by the assumption on A);

3. r M = r N ∩ M nr .

It is usual to confuse subsets of N that satisfy 1 with the unuque substructure they

8
 support.
It is immediate to verify that the intersection of an arbitrary family of substructures
of N is a substructure of N. Therefore, for any given A ⊆ N we may define the
substructure of N generated A as the intersection of all substructures of N that
contain A. We write h Ai N . The following easy proposition gives more concrete
representation of h Ai N

1.8 Lemma The following hold for every A ⊆ N

n o
1. h Ai N = t N : t a closed L( A)-term
n o
2. h Ai N = t N ( a) : t( x ) an L-term and a ∈ A| x|
[
3. h Ai N = An , where A0 = A
n∈ω n nf
o
A n +1 = An ∪ f N ( a) : f ∈ Lfun , a ∈ An . 

1.9 Example Let L be the language of groups. Let N be a group, which we consider
as an L-structure in the natural way. Then the substructures of N are exactly the
subgroups of N and h Ai N is the group generated by A ⊆ N. A similar claim is true
when L F is the signature of vector spaces over some fixed field F. The choice of the
language is more or less fixed if we want that the algebraic and the model theoretic
notion of substructure coincide. 

5 Formulas
Fix a language L and a set of variables V as in Section 3. A formula is a finite
.
sequence of symbols in Lfun ∪ Lrel ∪ V ∪ {=, ⊥, ¬, ∨, ∃}. The last set contains the
logical symbols that are called respectively
.
= equality ⊥ contradiction ¬ negation
∨ disjunction ∃ existential quantifier.
.
Syntactically, = behaves like a binary relation symbol. So, for convenience set n=. =
.
2. However = is considered as a logic symbol because its semantic is fixed (it is
always interpreted in the diagonal).
The definition below uses the prefix notation which simplifies the proof of the
unique legibility lemma. However, in practice we always we use the infix nota-
.
tion: t = s, ϕ ∨ ψ, etc.

1.10 Definition A formula is any finite sequence is obtained with the following inductive
procedure
.
o. if r ∈ Lrel ∪ {=} and t is a tuple obtained concatenating nr terms then r t is a formula.
Formulas of this form are called atomic ;
i. if ϕ e ψ are formulas then the following are formulas: ⊥, ¬ ϕ, ∨ ϕ ψ, and ∃ x ϕ, for
any x ∈ V. 

We use L to denote both the language and the set of formulas. We write Lat
for the set of atomic formulas and Lqf for the set of quantifier-free formulas i.e.

9
 formulas where ∃ does not occur.
The proof of the following is similar to the analogous lemma for terms.

1.11 Lemma (unique legibility of formulas) Let a be a sequence of formulas. Suppose a

can be obtained both by the concatenation of the formulas ϕ1 , . . . , ϕn or by the concatenation
of the formulas ψ1 , . . . , ψm . Then n = m and ϕi = ψi . 

A formula is closed if all its variables occur under the scope of a quantifier. Closed
formulas are also called sentences . We will do without a formal definition of occurs
under the scope of a quantifier which is too lengthy. An example suffices: all occur-
rences of x are under the scope a quantifiers in the formula ∃ x ϕ. These occurrences
.
are called bonded . The formula x =y ∧ ∃ x ϕ has free (i.e., not bond) occurrences
of x and y.
Let x is a tuple of variables and t is a tuple of terms such that | x | = |t|. We
write ϕ[ x/t] for the formula obtained substituting t for all free occurrences of x,
coordinatewise.
We write ϕ( x ) to declare that the free variables in the formula ϕ are all among those
of the tuple x. In this case we write ϕ(t) for ϕ[ x/t].
We will often use without explicit mention the following useful syntactic decompo-
sition of formulas with parameters.

1.12 Lemma For every formula ϕ( x ) ∈ L( A) there is a formula ψ( x ; z) ∈ L and a tuple of

parameters a ∈ A|z| such that ϕ( x ) = ψ( x ; a). 

Just as a term t( x ) is a name for a function t( x ) M : M| x| → M, a formula ϕ( x ) is a

name for a subset ϕ( x ) M ⊆ M| x| which we call the subset of M defined by ϕ( x ) .
It is also very common to write ϕ( M) for the set defined by ϕ( x ). In general sets
of the form ϕ( M ) for some ϕ( x ) ∈ M are called definable .

1.13 Definition of truth For every formula ϕ with variables among those of the tuple x we
define ϕ( x ) M by induction as follows
.
n o
o1. (= t s)( x ) M = a ∈ M| x| : t M( a) = s M( a)
n o
o2. (r t1 . . . tn )( x ) M = a ∈ M| x| : ht1M (a), . . . , tnM( a)i ∈ r M

i0. ⊥( x ) M = ∅

¬ξ ( x ) M = M|x| r ξ ( x ) M


i1.

∨ ξ ψ ( x ) M = ξ ( x ) M ∪ ψ( x ) M


i2.

ϕ[y/a] ( x ) M
[
∃y ϕ ( x ) M =



i3.
a∈ M

Condition i2 assumes that ξ and ψ are uniquely determined by ∨ ξ ψ. This is a guaranteed

by the unique legibility o formulas, Lemma 1.11. Analogously, o1 e o2 assume Lemma 1.5.

10
 The case when x is the empty tuple is far from trivial. Note that ϕ(∅) M is a subset
of M0 = {∅}. Then there are two possibilities either {∅} or ∅. We wil read
them as two truth values : True and False , respectively. If ϕ M = {∅} we say that
ϕ is true in M , if ϕ M = ∅, we say that ϕ is false M . We write M  ϕ , respectively
M 2 ϕ . Or we may say that M models ϕ , respectively M does not model ϕ . It is
immediate to verify at
ϕ( M) = a ∈ M|x| : M  ϕ( a) .


Note that usually, we say formula when, strictly speaking, we mean pair that consists
of a formula and a tuple of variables. Such pairs are interpreted in definable sets
(cfr. Definition 1.13). In fact, if the tuple of variables were not given, the arity of the
corresponding set is not determined.
In some contexts we also want to distinguish between two sorts of variables that
play different roles. Some are placeholder for parameters, some are used to define
a set. In the the first chapters this distinction is only a clue for the reader, in the last
chapters it is an essential part of the definitions.

1.14 Definition A partitioned formula (strictly speaking, we should say a 2-partitioned for-
mula) is a triple ϕ( x ; z) consisting of a formula and two tuples of variables such that the
variables occurring in ϕ are all among x, z. 

We use a semicolon to separate the two tuples of variables. Typically, z is the

placeholder for parameters and x runs over the elements of the set defined by the
formula.

6 Yet more notation

Now we abandon the prefix notation in favor of the infix notation. We also use the
following logical connectives as abbreviations
> stands for ¬⊥ tautology
ϕ∧ψ
 
stands for ¬ ¬ ϕ ∨ ¬ψ conjunction
ϕ→ψ stands for ¬ϕ ∨ ψ implication
ϕ↔ψ
   
stands for ϕ→ψ ∧ ψ→ϕ bi-implication
ϕ↔
 
/ ψ stands for ¬ ϕ↔ψ exclusive disjunction
∀x ϕ stands for ¬∃ x ¬ ϕ universal quantifier

We agree that → e ↔ bind less than ∧ e ∨. Unary connnectives (quantifiers and

negation) bind stronger then binarary connectives. For example
h  i h  i
∃ x ϕ ∧ ψ → ¬ξ ∨ ϑ reads as ∃ x ϕ ∧ ψ → ¬ξ ∨ ϑ
We say that ∀ x ϕ( x ) and ∃ x ϕ( x ) are the universal , respectively, existential closure
of ϕ( x ). We say that ϕ( x ) holds in M when its universal closure is true in M. We
say that ϕ( x ) is consistent in M when its existential closure is true in M.
The semantic of conjunction and disjunction is associative. Then for any finite set
of formulas { ϕi : i ∈ I } we can write without ambiguities

11
 ^ _
ϕi ϕi
i∈ I i∈ I

When x = x1 , . . . , xn is a tuple of variables we write ∃ x ϕ or ∃ x1 , . . . , xn ϕ for

∃ x1 . . . ∃ xn ϕ. With first order sentences we are able to say that ϕ( M) has at least
n elements (also, no more than, or exactly n). It is convenient to use the following
abbreviations.
∃≥n x ϕ( x ) stands for ∃ x1 , . . . , xn
h ^ i
˙ xj .
^
ϕ ( xi ) ∧ xi 6 =
1≤ i ≤ n 1≤ i < j ≤ n

∃≤n x ϕ ( x ) stands for ¬∃ ≥ n +1

x ϕ( x )

∃=n x ϕ ( x ) stands for ∃≥n x ϕ ( x ) ∧ ∃≤n x ϕ ( x )

1.15 Exercise Let M be an L-structure and let ψ( x ), ϕ( x, y) ∈ L. For each of the follow-
ing conditions, write a sentence true in M exactly when

a. ψ( M ) ∈ ϕ( a, M ) : a ∈ M ;

b. ϕ( a, M ) : a ∈ M contains at least two sets;

c. ϕ( a, M ) : a ∈ M contains only sets that are pairwise disjoint. 

1.16 Exercise Let M be a structure in the signature of graphs as defined in Example 1.3
(but not necessarily a graph). Write a sentence ϕ such that,
a. M  ϕ if and only if there is an A ⊆ M such that r M ⊆ A × ¬ A.

Remark: ϕ assert an asymmetric version of the property below

b. M  ψ if and only if there is an A ⊆ M such that r M ⊆ ( A × ¬ A) ∪ (¬ A × A).

Assume M is a graph, what required in b is equivalent to saying that M is a bipartite

graph, or equivalently that it has chromatic number 2 i.e., we can color the vertices
with 2 colors so that no two adjacent vertices share the same color. 

12
 Chapter 2
Theories and elementarity

1 Logical consequences
A theory is a set T ⊆ L of sentences. We write M  T if M  ϕ for every ϕ ∈ T. If
ϕ ∈ L is a sentence we write T ` ϕ when
MT ⇒ Mϕ for every M.
In words, we say that ϕ is a logical consequence of T or that ϕ follows from T. If
S is a theory T ` S has a similar meaning. If T ` S and S ` T we say that T and S
are logically equivalent . We may say that T axiomatizes S (or vice versa).
We say that a theory is consistent if it has a model. With the notation above, T is
consistent if and only if T 0 ⊥ .
The closure of T under logical consequence is the set ccl( T ) which is defined as
follows:
n o
ccl( T ) = ϕ ∈ L : sentence such that T ` ϕ

If T is a finite set, say T = ϕ1 , . . . , ϕn we write ccl( ϕ1 , . . . , ϕn ) for ccl( T ). If
T = ccl( T ) we say that T is closed under logical consequences .
The theory of M is the set of sentences that hold in M and is denoted by Th( M) .
More generally, if K is a class of structures, Th(K) is the set of sentences that hold
in every model in K. That is
Th(K)
\
= Th( M )
M ∈K
The class of all models of T is denoted by Mod( T ) . We say that K is axiomatizable
if Mod( T ) = K for some theory T. If T is finite we say that K is finitely axiomatizable .
To sum up
n o
Th( M) = ϕ : Mϕ
n o
Th(K) = ϕ : M  ϕ for all M ∈ K
n o
Mod( T ) = M : MT

2.1 Example Let L be the language of multiplicative groups. Let Tg be the set contain-
ing the universal closure of following three formulas
1. ( x · y ) · z = x · ( y · z );
2. x · x −1 = x −1 · x = 1;
3. x · 1 = 1 · x = x.
Then Tg axiomatizes the theory of groups, i.e. Th(K) for K the class of all groups.
Let ϕ be the universal closure of the following formula

13
 z · x = z · y → x = y.
As ϕ formalizes the cancellation property then Tg ` ϕ, that is, ϕ is a logical conse-
quence of Tg . Now consider the sentence ψ which is the universal closure of
4. x · y = y · x.
So, commutative groups model ψ and non commutative groups model ¬ψ. Hence
neither Tg ` ψ nor Tg ` ¬ψ. We say that Tg does not decide ψ. 

Note that even when T is a very concrete set, ccl( T ) may be more difficult to grasp.
In the example above Tg contains three sentences but ccl( Tg ) is an infinite set con-
taining sentences that code theorems of group theory yet to be proved.

2.2 Remark The following properties say that ccl is a finitary closure operator.
1. T ⊆ ccl( T ) (extensive)


2. ccl( T ) = ccl ccl( T ) (increasing)
3. T ⊆ S ⇒ ccl( T ) ⊆ ccl(S) (idempotent)
S
4. ccl( T ) = ccl(S) : S finite subset of T . (finitary)
Properties 1-3 are easy to verify while 4 requires the compactness theorem. 

In the next example we list a few algebraic theories with straightforward axiomati-
zation.

2.3 Example We write Tag for the theory of abelian groups which contains the univer-
sal closure of following
a1. ( x + y) + z = y + ( x + z);

a2. x + (− x ) = 0;

a3. x + 0 = x;

a4. x + y = y + x.

The theory Tr of (unitary) rings extends Tag with

a5. ( x · y) · z = x · (y · z);

a6. 1 · x = x · 1 = x;

a7. ( x + y) · z = x · z + y · z;

a8. z · ( x + y) = z · x + z · y.

The theory of commutative rings Tcg contains also com of examples 2.1. The theory
of ordered rings Tor extends Tcr with
o1. x < z → x + y < z + y;

o2. 0 < x ∧ 0 < z → 0 < x · z. 

The axiomatization of the theory of vector spaces is less straightforward.

2.4 Example Fix a field F. The language L F extends the language of additive groups
with a unary function for every element of F. The theory of vector fields over F
extends Tag with the following axioms (for all h, k, l ∈ F)

14
 m1. h ( x + y) = h x + h y

m2. l x = h x + k x, where l = h + F k
m3. l x = h (k x ), where l = h · F k
m4. 0 F x = 0

m5. 1 F x = x

The symbols 0F and 1F denote the zero and the unit of F. The symbols + F and · F
denote the sum and the product in F. These are not part of L F , they are symbols we
use in the metalanguage. 

2.5 Example Recall from Example 1.3 that we represent a graph with a symmetric
irreflexive relation. Therefore theory of graphs contains the following two axioms
1. ¬r ( x, x );

2. r ( x, y) → r (y, x ). 

Our last example is a trivial one.

2.6 Example Let L be the empty language The theory of infinite sets is axiomatized
by the sentences ∃≥n x ( x = x ) for all positive integer n. 

2.7 Exercise Prove that ccl( ϕ ∨ ψ) = ccl( ϕ) ∩ ccl(ψ). 

2.8 Exercise Prove that T ∪ { ϕ} ` ψ then T ` ϕ → ψ. 

2.9 Exercise Prove that Th(Mod( T )) = ccl( T ). 

2 Elementary equivalence
The following is a fundamental notion in model theory.

2.10 Definition We say that M and N are elementarily equivalent if

ee. N  ϕ ⇔ M  ϕ, for every ϕ ∈ L.
In this case we write M ≡ N . More generally, we write M ≡ A N and say that M and
N are elementarily equivalent over A if the following hold
a. A ⊆ M ∩ N

ee’. equivalence ee above holds for every sentence ϕ ∈ L( A). 

The case when A is the whole domain of M is particularly important.

2.11 Definition When M ≡ M N we write M  N and say that M is an elementary substructure

of N. 

In the definition above the use of the term substructure is appropriate by the follow-
ing lemma.

15
2.12 Lemma If M and N are such that M ≡ A N and A is the domain of a substructure of M
then A is also the domain a substructure of N and the two substructures coincide.
Proof Let f be a function symbol and ler r be a relation symbol. It suffices to prove
that f M ( a) = f N ( a) for every a ∈ An f and that r M ∩ Anr = r N ∩ Anr .
If b ∈ A is such that b = f M a then M  f a = b. So, from M ≡ A N, we obtain
N  f a = b, hence f N a = b. This proves f M ( a) = f N ( a).
Now let a ∈ Anr and suppose a ∈ r M . Then M  ra and, by elementarity, N  ra,
hence a ∈ r N . By symmetry r M ∩ Anr = r N ∩ Anr follows. 

It is not easy to prove that two structures are elementary equivalent. A direct
verification is unfeasible even for the most simple structures. It will take a few
chapters before we are able to discuss concrete examples.
We generalize the definition of Th( M ) to include parameters
n o
Th( M/A) = ϕ : sentence in L( A) such that M  ϕ .
The following proposition is immediate

2.13 Proposition For every pair of structures M and N and every A ⊆ N ∩ N the following
are equivalent
a. M ≡ A N;

b. Th( M/A) = Th( N/A);

c. N  ϕ( a) ⇔ M  ϕ( a) for every ϕ( x ) ∈ L and every a ∈ A| x| .

d. ϕ( M ) ∩ A| x| = ϕ( N ) ∩ A| x| for every ϕ( x ) ∈ L. 

If we restate a and c of the proposition above when A = M we obtain that the

following are equivalent
a’. M  N;

c’. ϕ( M ) = ϕ( N ) ∩ M| x| for every ϕ( x ) ∈ L.

Note that c’ extends to all definable sets what Definition 1.7 requires for a few basic
definable sets.

2.14 Example Let G be a group which we consider as a structure in the multiplicative

language of groups. We show that if G is simple and H  G then also H is simple.
Recall that G is simple if all its normal subgroups are trivial, equivalently, if for
every a ∈ G r {1} the set gag−1 : g ∈ G generates the whole group G.


Assume H is not simple then there a, b ∈ H be such that b is not product of elements
of hah−1 : h ∈ H . hence for every n


H  ¬∃ x1 , . . . , xn b = x1 ax1−1 · · · xn axn−1

By elementarity the same hold in G. Hence G is not simple. 

2.15 Exercise Let A ⊆ M ∩ N. Prove that M ≡ A N if and only if M ≡ B N for every

finite B ⊆ A. 

2.16 Exercise Let M  N and let ϕ( x ) ∈ L( M). Prove that ϕ( M) is finite if and only if

16
 ϕ( N ) is finite and in this case ϕ( N ) = ϕ( M). 

2.17 Exercise Let M  N and let ϕ( x, z) ∈ L. Suppose there are finitely many sets of
the form ϕ( a, N ) for some a ∈ N | x| . Prove that all these sets are definable over M. 

2.18 Exercise Consider Zn as a structure in the additive language of groups with the
natural interpretation. Prove that Zn 6≡ Zm for every positive integers n 6= m. Hint:
in Zn there are at most 2n − 1 elements that are not congruent modulo 2. 

3 Embeddings and isomorphisms

Here we prove that isomorphic structures are elementary equivalent and a few
related results.

2.19 Definition An embedding of M into N is an injective map h : M → N such that

1. a ∈ r M ⇔ ha ∈ r N for every r ∈ Lrel and a ∈ Mnr ;
2. f M ( a) = f N (h a) for every f ∈ Lfun and a ∈ Mn f .
Note that when c ∈ Lfun is a constant 2 reads h c M = c N . Therefore that M ⊆ N if and
only if id M : M → N is an embedding.
An surjective embedding is an isomorphism or, when domain and codomain coincide, an
automorphism . 

Condition 1 above and the assumption that h is injective can be summarized in the
following
1’. M  r ( a) ⇔ N  r (ha) for every r ∈ Lrel ∪ {=} and every a ∈ Mnr .
Note also that, by straightforward induction on syntax, from 2 we obtain
2’ h t M ( a) = t N (h a) for every term t( x ) and every a ∈ M| x| .
Combining these two properties, and by straightforward induction on the syntax,
we obtain
3. M  ϕ( a) ⇔ N  ϕ(ha) for every ϕ( x ) ∈ Lqf and every a ∈ M| x| .
Recall that we write Lqf for the set of quantifier-free formulas. It is worth noting
that when M ⊆ N and h = id M then 3 becomes
3’ M  ϕ( a) ⇔ N  ϕ( a) for every ϕ( x ) ∈ Lqf and for every a ∈ M| x| .
In words this is summarized by saying that the truth of quantifier-free formulas is
preserved under sub- and superstructure.

Finally we prove that first order truth is preserved under isomorphism. We say that
a map h : M → N fixes A ⊆ M (pointwise) if id A ⊆ h. An isomorphism that fixes
A is also called an A-isomorphism .

2.20 Theorem If h : M → N is an isomorphism then for every ϕ( x ) ∈ L

# M  ϕ( a) ⇔ N  ϕ(ha) for every a ∈ M| x|
In particular, if h is an A-isomorphism then M ≡ A N.

17
 Proof We proceed by induction of the syntax of ϕ( x ). When ϕ( x ) is atomic # holds
by 3 above. Induction for the Boolean connectives is straightforward so we only
need to consider the existential quantifier. Assume as induction hypothesis that
M  ϕ( a, b) ⇔ N  ϕ(ha, hb) for every tupla a ∈ M| x| and b ∈ M.
We prove that # holds for the formula ∃y ϕ( x, y).
M  ∃y ϕ( a, y) ⇔ M  ϕ( a, b) for some b ∈ M
⇔ N  ϕ(ha, hb) for some b ∈ M (by induction hypothesis)
⇔ N  ϕ(ha, c) for some c ∈ N (⇐ by surjectivity)
⇔ N  ∃y ϕ(ha, y). 
 
2.21 Corollary If h : M → N is an isomorphism then h ϕ( M) = ϕ( N ) for every ϕ( x ) ∈ L. 

We can now give a few very simple examples of elementarily equivalent structures.

2.22 Example Let L be the language of strict orders. Consider intervals of R (or in Q)
as structures in the natural way. The intervals [0, 1] and [0, 2] are isomorphic, hence
[0, 1] ≡ [0, 2] follows from Theorem 2.20. Clearly, [0, 1] is a substructure of [0, 2].
However [0, 1]  [0, 2], in fact the formula ∀ x ( x ≤1) holds in [0, 1] but is false in
[0, 2].
The example above shows that M ⊆ N and M ≡ N does not imply M  M.
Now we prove that (0, 1)  (0, 2). By Exercise 2.15 above, it suffices to verify that
(0, 1) ≡ B (0, 2) for every finite B ⊆ (0, 1). This follows again by Theorem 2.20 as
(0, 1) and (0, 2) are B-isomorphic for every finite B ⊆ (0, 1). 

For the sake of completeness we also give the definition of homomorphism.

2.23 Definition A homomorphism is a total map h : M → N such that

1. a ∈ r M ⇒ ha ∈ r N for every r ∈ Lrel and a ∈ Mnr ;

2. h f M ( a) = f N (h a) for every f ∈ Lfun and a ∈ Mn f .

Note that only one implication is required in 1. 

2.24 Exercise Prove that if h : N → N is an automorphism and M  N then h[ M ]  N. 

2.25 Exercise Let L be the empty language. Let A, D ⊆ M. Prove that the following are
equivalent
1. D is definable over M;

2. either D is finite and A ⊆ A or ¬ D is finite and ¬ D ⊆ A.

Hint: as L-structures are plain sets, every bijection f : M → M is an automorphism. 

2.26 Exercise Prove that if ϕ( x ) is an existential formula and h : M → M is an embed-

ding then
M  ϕ( a) ⇒ N  ϕ(ha) for every a ∈ M| x| .

18
 Recall that existential formulas as those of the form ∃y ψ( x, y) for ψ( x, y) ∈ Lqf .
Note that Theorem 10.6 proves that the property above characterizes existential
formulas. 

2.27 Exercise Assume that the language L contains only the symbol +. Consider Z as
a structure of signature L with the usual interpretation of +. Prove that there is no
existential formula that defines the set of even integers. Hint: use Exercise 2.26. 

2.28 Exercise Let N be the multiplicative group of Q. Let M be the subgroup of those
rational numbers that are of the form n/m for some odd integers m and n. Prove
that M  N. Hint: use the fundamental theorem of arithmetic and reason as in
Example 2.22. 

4 Quotient structures
The content of this section is mainly technical and only required later in the course.
Its reading may be postponed.
If E is an equivalence relation on N we write [c] E for the equivalence class of c ∈ N.
We use the same symbol for the equivalence relation on N n defined as follow: if
a = a1 , . . . , an and b = b1 , . . . , bn are n-tuples of elements of N then a E b means
that ai E bi holds for all i. It is easy to see that b1 , . . . , bn ∈ [ a, ..., an ] E if and only if
bi ∈ [ ai ] E for all i. Therefore we use the notation [ a] E for both the equivalence class
of a ∈ N n and the tuple of equivalence classes [ a1 ] E , . . . , [ an ] E .

2.29 Definition We say that the equivalence relation E on a structure N is a congruence if for
every f ∈ Lfun
c1. aEb ⇒ f N a E f N b;
When E is a congruence on N we write N/E for the a structure that has as domain the set
of E-equivalence classes in N and the following interpretation of f ∈ Lfun and r ∈ Lrel :
f N/E [ a] E =
 N 
c2. f a E;
c3. [ a] E ∈ r N/E ⇔ [ a] E ∩ r N 6= ∅.
We call N/E the quotient structure . 

By c1 the quotiont structure is well defined. The reader will recognize it as a familiar
notion by the following proposition (which is not required in the following and
requires the notion of homomorphism, see Definition 2.23. Recall that the kernel
of a total map h : N → M is the equivalence relation E such that
a E b ⇔ ha = hb
for every a, b ∈ N.

2.30 Proposition Let h : N → M be a surjective homomorphism and let E be the kernel of h.

Then there is an isomorphism k that makes the following diagram commute

19
 h
N M
k
π

N/E

where π : a 7→ [ a] E is the projection map. 

B Quotients clutter the notation with brackets. To avoid the mess, we prefer to reason
in N and tweak the satisfaction relation. Warning: this is not standard.
Essentially, we work in N while thinking of N/E. So, we replace equality with E and
adapt the interpretation of relation symbols according to c3 above. The proposition
below shows that the definition does what is intended to do.

2.31 Definition For t2 , t2 closed terms of L( N ) define

1∗ N/E * t1 = t2 ⇔ t1N E t2N
For t a tuple of closed terms of L( N ) and r ∈ Lrel a relation symbol
2∗ N/E * rt ⇔ t N E a ∈ r N for some a ∈ M|t|
Finally the definition is extended to all sentences ϕ ∈ L( N ) by induction in the usual way
3∗ N/E * ¬ ϕ ⇔ not N/E * ϕ
4∗ N/E * ϕ ∧ ψ ⇔ N/E * ϕ and N/E * ψ
5∗ N/E * ∃ x ϕ( x ) ⇔ N/E * ϕ( â) for some a ∈ N. 

Now, by induction on the syntax of formulas one can prove * does what required.
In particular, N/E * ϕ( a) ↔ ϕ(b) for every a E b.

2.32 Proposition Let E be a congruence relation of N. Then the following are equivalent for
every ϕ( x ) ∈ L
1. N/E * ϕ ( a );


2. N/E  ϕ [ a] E . 

5 Completeness
A theory T is maximal consistent if it is consistent and there is no consistent theory
S such that S ⊂ T. Equivalently, T contains every sentence ϕ consistent with T,
that is, such that T ∪ { ϕ} is consistent. Clearly a maximal consistent theory is closed
under logical consequences.
A theory T is complete if cclT is maximal consistent. Concrete examples will be
given in the next chapters as it is not easy to prove that a theory is complete.

2.33 Proposition The following are equivalent

a. T is maximal consistent;

b. T = Th( M ) for some structure M;

20
 c. T is consistent and ϕ ∈ T or ¬ ϕ ∈ T for every sentence ϕ.

Proof To prove a⇒b, assume that T is consistent. Then there is M  T. Therefore

T ⊆ Th( M ). As T is maximal consistent T = Th( M). Implication b⇒c is immediate.
As for c⇒a note that if T ∪ { ϕ} is consistent then ¬ ϕ 6∈ T therefore ϕ ∈ T follows
from c. 

The proof of the proposition below is is left as an exercise for the reader.

2.34 Proposition The following are equivalent

a. T is complete;

b. there is a unique maximal consistent theory S such that T ⊆ S;

c. T is consistent and T ` Th( M ) for every M  T;

d. T is consistent and T ` ϕ o T ` ¬ ϕ for every sentence ϕ;

e. T is consistent and M ≡ N for every pair of models of T. 

2.35 Exercise Prove that the following are equivalent

a. T is complete;

b. for every sentence ϕ, o T ` ϕ o T ` ¬ ϕ but not both.

By contrast prove that the following are not equivalent

a. T is maximal consistent;

b. for every sentence ϕ, o ϕ ∈ T o ¬ ϕ ∈ T but not both.

Hint: consider the theory containing all sentences where the symbol ¬ occurs an
even number of times. This theory is not consistent as it contains ⊥. 

2.36 Exercise Prove that if T has exactly 2 maximal consistent extension T1 and T2 then
there is a sentence ϕ such that T, ϕ ` T1 and T, ¬ ϕ ` T2 . State and prove the
generalization to finitely many maximally consistent extensions. 

6 The Tarski-Vaught test

There is no natural notion of smallest elementary substructure containing a set of
parameters A. The downward Löwenheim-Skolem, which we prove in the next
section, is the best result that holds in full generality. Given an arbitrary A ⊆ N we
shall construct a model M  N containing A that is small in the sense of cardinality.
The construction selects one by one the elements of M that are required to realise
the condition M  N. Unfortunately, Definition 2.11 supposes full knowledge of
the truth in M and it may not be applied during the construction. The following
lemma comes to our rescue with a property equivalent to M  N that only mention
the truth in N.

2.37 Lemma (Tarski-Vaught test) For every A ⊆ N the following are equivalent
1. A is the domain of a structure M  N;

2. for every formula ϕ( x ) ∈ L( A), with | x | = 1,

21
 N  ∃ x ϕ( x ) ⇒ N  ϕ(b) for some b ∈ A.
Proof 1⇒2
N  ∃ x ϕ( x ) ⇒ M  ∃ x ϕ( x )
⇒ M  ϕ(b) for some b ∈ M
⇒ N  ϕ(b) for some b ∈ M.
2⇒1 Firstly, note that A is the domain of a substructure of N, that is, f N a ∈ A for
every f ∈ Lfun and every a ∈ An f . In fact, this follows from 2 with f a = x for ϕ( x ).
Write M for the substructure of N with domain A. By induction on the syntax we
prove
M  ϕ( a) ⇔ N  ϕ( a) for every a ∈ M| x| .
If ϕ( x ) is atomic the claim follows from M ⊆ N and the remarks underneath Defini-
tion 2.19. The case of Boolean connectives is straightforward, so only the existential
quantifier requires a proof
M  ∃ x ϕ( a, x ) ⇔ M  ϕ( a, b) for some b ∈ M
⇔ N  ϕ( a, b) for some b ∈ M
⇔ N  ∃ x ϕ( a, x ).
The second equivalence holds by induction hypothesis, in the last equivalence we
use 2 for the implication ⇐. 

2.38 Exercise Prove that, in the language of strict orders, R r {0}  R and R r {0} 6' R. 

7 Downward Löwenheim-Skolem
The main theorem of this section was proved by Löwenheim at the beginning of the
last century. Skolem gave a simpler proof immediately afterwards. At the time, the
result was perceived as paradoxical.
A few years earlier, the works of Zermelo and Fraenkel had convinced the mathe-
matical community that the whole of set theory could be formalised in a first order
language. The downward Löwenheim-Skolem theorem implies the existence of an
infinite countable model M of set theory: this is the so-called Skolem paradox . The
existence of M seems paradoxical because, in particular, a sentence that formalises
the axiom of power set holds in M. Therefore M contains an element b which, in
M, is the set of subsets of the natural numbers. But the set of elements of b is a
subset of M, and therefore it is countable.
In fact, this is not a contradiction, because the expression all subsets of the natural
numbers does not have the same meaning in M as it has in the real world. The notion
of cardinality, too, acquires a different meaning. In the language of set theory, there
is a first order sentence that formalises the fact that b is uncountable: the sentence
says that there is no bijection between b and the natural numbers. Therefore the
bijection between the elements of b and the natural numbers (which exists in the
real world) does not belong to M. The notion of equinumerosity has a different
meaning in M and in the real world, but those who live in M cannot realise this.

22
2.39 Downward Löwenheim-Skolem Theorem Let N be an infinite L-structure and let
A ⊆ N. Then there is a structure M of cardinality ≤ | L( A)| such that A ⊆ M  N.
Proof Set λ = | L( A)|. Below we construct a chain h Ai : i < λi of subsets of N.
The chain begins at A0 = A and is continuous at limit ordinals. Finally we set
M = i<λ Ai . All Ai will have cardinality ≤ λ so | M | ≤ λ follows.
S

Now we construct Ai+1 given Ai . Assume as induction hypothesis that | Ai | ≤ λ.

Then | L( Ai )| ≤ λ. For some fixed variable x let h ϕk ( x ) : k < λi be an enumeration
of the formulas in L( Ai ) that are consistent in N. For every k pick ak ∈ N such that
N  ϕk ( ak ). Define Ai+1 = Ai ∪ { ak : k < λ}. Then | Ai+1 | ≤ λ is clear.
Now use Tarski-Vaught test to prove M  N. Suppose ϕ( x ) ∈ L( M ) is consistent in
N. As finitely many parameters occur in formulas, ϕ( x ) ∈ L( Ai ) for some i. Then
ϕ( x ) is among the formulas enumerated above and Ai+1 ⊆ M contains a solution
of ϕ( x ). 

We will need to adapt the construction above to meet more requirements on the
model M. To better control the elements that enters M it is convenient to add one
element at the time (above we add λ elements at each stage). We need to enumerate
formulas with care if we want to complete the construction by stage λ.

2.40 Second proof of the downward Löwenheim-Skolem Theorem From set theory
we know there is a bijection π : λ → λ2 such that π1 (i ) and π2 (i ), the two compo-
nents of π (i ), are both ≤ i. Suppose we have defined the set A j for very j ≤ i. Let
h ϕk ( x ) : k < λi be an enumeration of the formulas in L( Aπ1 (i) ). Let b be a solution
of the ϕk ( x ) for k = π2 (i ) and define Ai+1 = Ai ∪ {b}.
We use Tarski-Vaught test to prove M  N. Let ϕ( x ) ∈ L( M) be consistent in N.
As π is a bijection, there is an i such that ϕ( x ) is the π2 (i )-th formula enumerated
in L( Aπ1 (i) ). Hence a witness of ϕ( x ) is enumerated in M at stage i + 1. 

2.41 Exercise Assume L is countable and let M  N have arbitrary (large) cardinality.
Let A ⊆ N be countable. Prove there is a countable model K such that A ⊆ K  N
and K ∩ M  N (in particular, K ∩ M is a model). Hint: adapt the construction used
to prove the downward Löwenheim-Skolem Theorem. 

8 Elementary chains
An elementary chain is a chain h Mi : i < λi of structures such that Mi  M j
for every i < j < λ. The union (or limit ) of the chain is the structure with as
S
domain the set i<λ Mi and as relations and functions the union of the relations
and functions of Mi . It is plain that all structures in the chain are substructures of
the limit.

2.42 Lemma Let h Mi : i ∈ λi be an elementary chain of structures. Let N be the union of the
chain. Then Mi  N for every i.
Proof By induction on the syntax of ϕ( x ) ∈ L we prove
|x|
Mi  ϕ ( a ) ⇔ N  ϕ ( a ) for every i < λ and every a ∈ Mi
As remarked in 3’ of Section 3, the claim holds for quantifier-free formulas. In-

23
 duction for Boolean connectives is straightforward so we only need to consider the
existential quantifier
Mi  ∃y ϕ( a, y) ⇒ Mi  ϕ( a, b) for some b ∈ Mi .
⇒ N  ϕ( a, b) for some b ∈ Mi ⊆ N
where the second implication follows from the induction hypothesis. Vice versa
N  ∃y ϕ( a, y) ⇒ N  ϕ( a, b) for some b ∈ N
Without loss of generality we can assume that b ∈ M j for some j ≥ i and obtain
⇒ M j  ϕ( a, b) for some b ∈ M j
Now apply the induction hypothesis to ϕ( x, y) and M j
⇒ M j  ∃y ϕ( a, y)
⇒ Mi  ∃y ϕ( a, y)
where the last implication holds because Mi  M j . 

2.43 Exercise Let h Mi : i ∈ λi be an elementary chain of substructures of N. Let M be

the union of the chain. Prove that M  N. 

2.44 Exercise Give an alternative proof of Exercise 2.41 using the downward Löwenheim-
Skolem Theorem (instead of its proof). Hint: construct two countable chains of
countable models such that Ki ∩ M ⊆ Mi  N and A ∪ Mi ⊆ Ki+1  N. The
required model is K = i∈ω Ki . In fact it is easy to check that K ∩ M = i∈ω Mi .
S S


24
 Chapter 3
Ultraproducts
In these notes we only use ultraproducts to prove the compactness theorem. Since
a syntactic proof of the compactness theorem is also given, this chapter is, strictly
speaking, not required. However, the importance of ultraproducts transcends its
application to model theory.

1 Filters and ultrafilters

The material in this section will appear again in a more general setting in Chapter 5.
Let I be any set. A filter on I (or a filter of P( I ) ), is a non empty set F ⊆ P( I )
such that for every a, b ∈ P( I ):
f1 a ∈ F and a ⊆ b ⇒ b ∈ F
f2 a ∈ F and b ∈ F ⇒ a ∩ b ∈ F
A filter F is proper if F 6= P( I ), equivalently if ∅ ∈ / F. Otherwise it is improper .
By f1 above, F is proper if and only if ∅ ∈ / F. A filter F is principal if F = { a ⊆
I : b ⊆ a} for some set b ⊆ I. When this happens, we say that {b} generates F.
When I is finite every filter F is principal and generated by {∩ F }. Non-principal
filters on I exist as soon as I is infinite. In fact, if I is infinite it is easy to check that
the following is a filter:
n o
F = a ⊆ I : a is cofinite in I .
where cofinite (in I) means that I r a is finite. This is called the Fréchet filter on I .
The Fréchet filter is the minimal non-principal filter.
A proper filter F is maximal if there is no proper filter H such that F ⊂ H. A
proper filter F is an ultrafilter if for every a ⊆ I
a∈
/ F ⇒ ¬a ∈ F where ¬ a = I r a .
Below we prove that the ultrafilters are exactly the maximal filters.

3.1 Exercise Let I be infinite. Prove that every non-principal ultrafilter on I contains
Fréchet’s filter. Show that this does not hold for plain filters. 

Let B ⊆ P( I ). Then the filter generated by B is the intersection of all the filters that
contain B. It is easy to check that the intersection of a family of filters is a filter,
so the notion is well defined. The following easy proposition gives a workable
characterization of the filter generated by a set.

3.2 Proposition The filter generated by B is a ⊆ I : C ⊆ a for some finite C ⊆ B .
T


C 6= ∅ for every finite C ⊆ B.

T
We say that B has the finite intersection property if
The following proposition is immediate.

25
3.3 Proposition The following are equivalent:
1. the filter generated by B is non principal;

2. B has the finite intersection property. 

A proper filter F is prime if for every a, b ⊆ I.

a ∪ b ∈ F ⇒ a ∈ F or b ∈ F
Prime filters coincide with maximal filters. However, in Chapter 5, we introduce a
more general context where primality is distinct from maximality.

3.4 Proposition For every filter F on I, the following are equivalent:

1. F is a maximal filter;

2. F is a prime filter;

3. F is an ultrafilter.

Proof 1⇒2. Suppose a, b ∈ / F, where F is maximal. We claim that a ∪ b ∈ / F. By

maximality, there is a c ∈ F such that a ∩ c = b ∩ c = ∅. Therefore ( a ∪ b) ∩ c = ∅.
Hence a ∪ b ∈
/ F.
2⇒3. It suffices to note that I = a ∪ ¬ a ∈ F.

3⇒1. If a ∈
/ F, where F is an ultrafilter, then ¬ a ∈ F and no proper filter contains
F ∪ { a }. 

3.5 Proposition Let B ⊆ P( I ) have the finite intersection property. Then B is contained in a
maximal filter.
Proof First we prove that the union of a chain of subsets of P( I ) with the finite
intersection property has the finite intersection property. Let B be such a chain and
suppose for a contradiction that B does not have the finite intersection property.
S

Fix a finite C ⊆ B such that C = ∅. As C is finite, we have C ⊆ B for some B ∈

T T

B. Hence B does not have the finite intersection property, which is a contradiction.
Now apply Zorn’s lemma to obtain a B ⊆ P( I ) which is maximal among the sets
with the finite intersection property. It is immediate that B is a filter. 

3.6 Exercise Prove that all principal ultrafilters are generated by a singleton. 

2 Direct products
In this and in the next section h Mi : i ∈ I i is a sequence of L-structures. (We are
abusing of the word sequence, since I is only a set.) The direct product of this
sequence is a structure denoted by

∏ Mi (below this product is denoted by N ).

i∈ I

and defined by conditions 1-3 below. If Mi = M for all i ∈ I, we say that N is a

direct power of M and denote it by M I .
The domain of N is the set containing all functions

26
 [
1. â : I → Mi
i∈ I

â : i 7→ âi ∈ Mi
We do not distinguish between tuples of elements of N and tuple-valued functions.
For instance, the tuple â = h â1 . . . ân i is identified with the function â : i 7→ âi =
h â1 i, . . . , ân i i. On a first reading of what follows, it may help to pretend that all
functions and relations are unary.
The interpretation of f ∈ Lfun is defined as follows:
f N â i = f Mi ( âi )


2. for all i ∈ I.
The interpretation of r ∈ Lfun is the product of the relations r Mi , that is, we define
3. â ∈ r N ⇔ âi ∈ r Mi for all i ∈ I.
The following proposition is immediate.

3.7 Proposition If ∧ and ∀ are the only connectives and ∃ the only quantifier that occur in
ϕ( x ) ∈ L, then for every â ∈ N | x|
] N  ϕ( â) ⇔ Mi  ϕ( âi ) for all i ∈ I.
Proof By induction on syntax. First note that we can extend 2 to all terms t( x ) as
follows:
20 . t N â i = t Mi ( âi )


for all i ∈ I.
Combining 3 and 20 gives that for every r ∈ Lrel ∪ {=} and every L-term t( x )
30 . N  rt â ⇔ Mi  rt âi for all i ∈ I.
This shows that ] holds for ϕ( x ) atomic. Induction for the connectives ∧, ∀ ed ∃ is
immediate. 

A consequence of Proposition 3.7 is that a direct product of groups, rings or vector

spaces is a structure of the same sort. However, a product of fields is not a field.

3 Łoś’s Theorem
Assume the notation of the previous sections. In particular, h Mi : i ∈ I i is a se-
quence of structures and N is the direct product of this sequence.
Let F be a filter on I. We define the following congruence on N (see Definition 2.29):

â ∼ F ĉ ⇔ i ∈ I : âi = ĉi ∈ F .
To check that ∼ F is indeed a congruence, first we need to check that it is an equiv-
alence relation. Reflexivity and symmetry are immediate, and transitivity follows
from f2 in Section 1. Then we check that ∼ F is compatible with the functions of L,
that is, that c1 of Definition 2.29 is satisfied. This follows from f1 in Section 1.
For brevity, we write N/F for N/∼ F and [ â] F for [ â]∼ F . Then c3 of Section 2.4 gives
c30 . [ â] F ∈ r N/F ⇔ i : âi = b̂i ∈ r Mi ∈ F for some b̂ ∈ N.


The structure N/F is called the reduced product of the structures h Mi : i ∈ I i or,
when Mi = M for all i ∈ I, the reduced power of M. When F is an ultrafilter we
say ultraproduct , respectively ultrapower .

27
 3.8 Łoś’s Theorem Let ϕ( x ) ∈ L and let F be an ultrafilter on I. Then for every â ∈ N | x|
the following are equivalent:
1. N/F * ϕ( â) (see Definition 2.31);

2. i : Mi  ϕ( âi ) ∈ F.
Proof We proceed by induction on the syntax of ϕ( x ). Suppose ϕ( x ) is of the form
rt( x ) for some tuple of terms t( x ) and r ∈ Lrel ∪ {=}. Then 1⇔2 is clear.
We prove the inductive step for the connectives ¬, ∧, and the quantifier ∃. We
begin with ¬. This is the only place in the proof where the assumption that F is an
ultrafilter is required. By the inductive hypothesis,
N/F * ¬ ϕ â ⇔ i : Mi  ϕ( âi ) ∈


/ F
So, as F is an ultrafilter

⇔ i : Mi  ¬ ϕ( âi ) ∈ F.
Now consider ∧. Assume inductively that the equivalence 1⇔2 holds for ϕ( x ) and
ψ( x ). Then
N/F * ϕ â ∧ ψ â ⇔ i : Mi  ϕ( âi ) ∈ F and i : Mi  ψ( âi ) ∈ F .


 

As filters are closed under intersection, we obtain


⇔ i : Mi  ϕ( âi ) ∧ ψ( âi ) ∈ F.
Finally, consider ∃y. Assume inductively that the equivalence 1⇔2 holds for ϕ( x, y).
Then
N/F * ∃y ϕ â, y ⇔ N/F * ϕ â, b̂



for some b̂ ∈ N

⇔ i : Mi  ϕ( âi, b̂i ) ∈ F for some b̂ ∈ N.
We claim this is equivalent to

⇔ i : Mi  ∃y ϕ( âi, y) ∈ F.
The ⇒ direction is trivial. For ⇐, we choose as b̂ a sequence that picks a witness of
Mi  ∃y ϕ( âi, y) if it exists, and some arbitrary element of Mi otherwise. 

Let a I denote the element of M I that has constant value a. The following is an
immediate consequence of Łoś’s theorem.

3.9 Corollary For every a ∈ M

M I /F * ϕ( a I ) ⇔ M  ϕ( a) . 

We often identify M with its image under the embedding h : a 7→ [ a I ] F , and say
that M I /F is an elementary extension of M.
The following corollary is an immediate consequence of the compactness theorem
that we prove in the next chapter. The construction in the proof uses ultrapowers.

3.10 Corollary Every infinite structure has a proper elementary extension.

Proof Let M be an infinite structure and let F be a non-principal ultrafilter on ω. It
suffices to show that h[ M], the image of the embedding defined above, is a proper
substructure of Mω /F. As M is infinite, there is an injective function dˆ ∈ Mω . Then
ˆ = a is either empty or a singleton and, as F is non

for every a ∈ M the set i : di

28
 principal, it does not belong to F. So, by Łoš Theorem, we have Mω /F * dˆ 6= a I
for every a ∈ M, that is, [dˆ] F ∈
/ h [ M ]. 

3.11 Exercise Consider N as a structure in the language of strict orders. Let F be a

non-principal ultrafilter on ω. Prove that in Nω /F there is a sequence h âi : i ∈ ω i
such that âi+1 < âi . 

3.12 Exercise Let I be the set of integers i > 1. For i ∈ I, let Zi denote the additive
group of integers modulo i, and let N denote the product ∏ Zi . Prove that, if F is
a non-principal ultrafilter on I, i∈ I

1. for some F, N/F does not contain any element of finite order;

2. for some F, N/F has some elements of order 2;

3. for all F, N/F contains an element of infinite order;

4. for some F, N/F  ∀ x ∃y my = x for every integer m > 0;

5. for some F, N/F contains an element â such that N/F  ∀ x mx 6= â for every
positive integer m. 

29
 Chapter 4
Compactness
We present two proofs of the compactness theorem. The first is syntactic, the second
uses ultrapowers.
Somewhat surprisignly, the compactness theorem is not strctly required for the
next few chapters (only from Chapter 9). So the reading of this chapter may be
postponed.

1 Compactness via syntax

Here we prove the compactness theorem using the so-called Henkin method. We
divide the proof in two steps. Firstly, we observe that when the language is rich
enough to name witnesses for all existential statements of the theory, these wit-
nesses (Henkin constants) form a canonical model. Secondly, we show that we can
add the required Henkin constants to any finitely consistent theory.

4.1 Definition Fix a language L. Assume for simplicity that formulas use only the connectives
∧, ¬ and ∃. We say that T is a Henkin theory if for every formulas ϕ and ψ
0. ϕ ∈ T ⇒ ¬ϕ ∈
/T
1. ¬¬ ϕ ∈ T ⇒ ϕ∈T
2. ϕ∧ψ ∈ T ⇒ ϕ ∈ T and ψ ∈ T
3. ¬( ϕ ∧ ψ) ∈ T ⇒ ¬ ϕ ∈ T or ¬ψ ∈ T
4. ∃x ϕ ∈ T ⇒ ϕ[ x/a] ∈ T for some closed term a
5. ¬∃ x ϕ ∈ T ⇒ ¬ ϕ[ x/a] ∈ T for all closed terms a.
Moreover, the following holds for all closed terms t and all closed terms a, b, c
.
a. a= a ∈ T
. .
b. a=b ∈ T ⇒ b= a ∈ T
. . .
c. a=b, b=c ∈ T ⇒ a=c ∈ T
.
d. a=b, ϕ[ x/a] ∈ T ⇒ ϕ[ x/b] ∈ T. 

Fix a theory T and let M be the structure that has as domain the set of closed
terms. Define for every relation symbol

rM = h a1 , . . . , a n i : r ( a1 , . . . , a n ) ∈ T .
Define for every function symbol f
fM =

h a1 , . . . , an , ti : a1 , . . . , an ∈ M, t = f a1 . . . an .
An easy proof by induction shows that t M = t for all closed terms t.
.
Finally, let E be the relation on M that holds when a=b ∈ T.

30
4.2 Lemma The relation E is a congruence on M (as defined in Section 2.4).
Proof Axioms a-c ensure that E is an equivalence. Axiom e that it is a congruence. 

4.3 Theorem If T is Henkin theory then M/E  T.

Proof By induction on the complexity of the formula ϕ in T we prove that
1. ϕ ∈ T ⇒ M/E * ϕ for the notation cf. Definition 2.31
2. ¬ ϕ ∈ T ⇒ M/E * ¬ ϕ
Induction is immediate by 1-5 of Definition 4.1. We verify the claim for atomic
.
formulas. Consider first the formula ϕ = (t=s) where t and s are closed terms.
.
Recall that t M = t and s M = s, so claim 1 is clear. As for 2, suppose M/E * t=s,
. .
that is tEs. Then t=s ∈ T and ¬t=s ∈ / T follows from axiom 0.
Now assume ϕ = rt for a relation r and a tuple of closed terms t. The argument is
similar: 1 is immediate; to prove 2 suppose that M/E * rt. Then tEs ∈ r M for some
tuple of closed terms s. Then rs ∈ T, and by d rt ∈ T. Finally from 0 we obtain
¬rt ∈
/ T. 

4.4 Proposition If every finite subset of T has a model then there is a Henkin theory T 0
containing T. (The theory T 0 may be in an expanded language.)
Proof Set λ = | L|. Let hci : i < λi be some constants not in L. Write Li be the
language L expanded with the constants ci and fix an enumeration of all sentences
in Lλ .
If α is 0 or a limit ordinal define
Tα0 Ti0
[
= Tα = T ∪
i <α
As for successor ordinals, let Ti+1 be a maximal finitely consistent set of Li -formulas
containing Ti0 . It is immediate that Ti0 satisfy Definition 4.1 but possibly for 3.

Let ∃ x ϕ be the least sentence in Lλ contradicting 3 and set Ti+1 = Ti ∪ ϕ[ x/ci ] .
It is immediate that Ti+1 is finitely consistent.
At stage λ all possible counterexamples to 3 have been ruled out, then T 0 = Tλ is
the required Henkin theory. 

A theory is finitely consistent if all its finite subsets are consistent. The following
theorem is an immediate corollary of the proposition above.

4.5 Compactness Theorem If every finite subset of T is consistent then T is consistent. 

2 Compactness via ultraproducts

We repeat, a theory is finitely consistent if all its finite subsets are consistent. The
following theorem is the fiat lux of model theory.

4.6 Compactness Theorem Every finitely consistent theory is consistent.

Proof Let T be a finitely consistent theory. We claim that the structure N/F which
we define below is a model of T. Let I be the set of consistent sentences I in the

31
 language L. For every ξ ∈ I pick some Mξ  ξ. For any sentence ϕ ∈ L we write
X ϕ for the following subset of I
n o
Xϕ = ξ∈I : ξ`ϕ
Clearly ϕ is consistent if and only if X ϕ 6= ∅. Moreover X ϕ∧ψ = X ϕ ∩ Xψ . hence,

as T is finitely consistent, the set B = X ϕ : ϕ ∈ T has the finite intersection
property. Therefore B extends to an ultrafilter F on I. Define
N = ∏ Mξ .
ξ∈I

We claim that N/F  T. By Łoš Theorem, for every sentence ϕ ∈ L

n o
N/F  ϕ ⇔ ξ : Mξ  ϕ ∈ F .

By the definition of F, for every ϕ ∈ T, the set X ϕ ⊆ ξ : Mξ  ϕ belongs to F.
Therefore N/F  T, et lux fuit. 

The compactness theorem can be formulated in the following apparently stronger

way.

4.7 Corollary If T ` ϕ then there is some finite S ⊆ T such that S ` ϕ.

Proof Suppose S 0 ϕ for every finite S ⊆ T. Then for every finite S ⊆ T there
is a model M  S ∪ {¬ ϕ}. In other words, T ∪ {¬ ϕ} is finitely consistent. By
compactness T ∪ {¬ ϕ} hence T 0 ϕ. 

4.8 Exercise Let Φ ⊆ L be a set of sentences and suppose that ` ψ ↔ Φ for some
W

sentence ψ. Prove that there is a finite Φ0 ⊆ Φ such that ` ψ ↔ Φ0 .

W


3 Upward Löwenheim-Skolem
Recall that a type is a set of formulas. When we present types we usually declare
the variables that may occur in it – we write p( x ) , q( x ) , etc. where x is a tuple of
variables. Clearly, when x is the empty tuple, p( x ) is just a theory. We identify a
finite types with the conjunction of the formulas contained in it.
We write M  p( a) if M  ϕ( a) for every ϕ( x ) ∈ p. We say that a is a solution or
a realization of p( x ). An equivalent notation is M, a  p( x ) or, when M is clear
from the context, a  p( x ) . We say that p( x ) is consistent in M it has a solution
in M. In this case we may write M  ∃ x p( x ) . We say that p( x ) is consistent if it
is consistent in some model.
We say that a type p( x ) is finitely consistent if all its finite subsets are consistent.
If they are all consistent in the same model M, we say that p( x ) is finitely consistent
in M . The following theorem shows that the latter notion, which is trivial for
theories, is very interesting for types.

4.9 Compactness Theorem for types Every finitely consistent type p( x ) ⊆ L is consistent.
Moreover, if p( x ) ⊆ L( M ) is finitely consistent in M then it is realised in some elementary
extension of M.
Proof Let L0 be the expansion of L obtained by adding the fresh symbols c, a tuple
of constants of the same length as x. Then p(c) is a finitely consistent theory in the

32
 language L0 . By the compactness theorem there is an L0 -structure N 0  p(c). Let N
be the reduced of N 0 to L. That is the L-structure with the same domain and the
same interpretation of N 0 on the symbols of L. Note that, though the constants c
0 0
are not in L, the elements c N remain in N. Then N, c N  p( x ).
As for the second claim, let a be an enumeration of M. We can assume that p( x )
has the form p0 ( x ; a) for some p0 ( x ; z) ∈ L. Define

q(z) = ϕ(z) : M  ϕ( a)
Clearly, p0 ( x ; z) ∪ q(z) is finitely consistent in M. By the first part of the proof there
is a model N such that N  p0 (c0 ; a0 ) for some c0 , a0 ∈ N | x,z| . Let h = {h a, a0 i}. We
claim that h : M → N is an embedding and that h[ M ]  N. So the theorem follows
by identifying M with h[ M ].
For any ϕ(z) ∈ L we have
h[ M]  ϕ(ha) ⇔ M  ϕ( a) ⇔ ϕ(z) ∈ q(z) ⇒ N  ϕ( a0 ).
Then h[ M]  N follows because all sentences in L( M ) have the form ϕ( a) for some
ϕ(z) ∈ L. (Using a term that will be introduced only in Section 5.4, we have proved
that h : M → N is an elementary embedding.) 

The following corollary is historically important.

4.10 Upward Löwenheim-Skolem Theorem Every infinite structure has arbitrarily large
elementary extensions.
Proof Let x = h xi : i < λi be a tuple distinct variables, where λ is an arbitrary

cardinal. The type p( x ) = xi 6= x j : i < j < λ is finitely consistent in every
infinite structure and every structure that realises p has cardinality ≥ λ. Hence the
claim follows from Theorem 4.9. 

4 Finite axiomatizability
A theory T is finitely axiomatizable if ccl(S) = ccl( T ) for some finite S. The fol-
lowing theorem shows that we can restrict the search of S to the subsets of T.

4.11 Proposition For every theory T the following are equivalent

1. T is finitely axiomatizable;

2. there a finite S ⊆ T such that S ` T.

Proof Only 1⇒2 requires a proof. If T is finitely axiomatizable, there is a sentence

ϕ such that ccl( ϕ) = ccl( T ). Then T ` ϕ ` T. By Proposition 4.7 there is a finite
S ⊆ T such that S ` ϕ. Then also S ` T. 

If L is empty, then every structure is a model. The theory of infinite sets is the set
of sentences that hold in every infinite structure.

4.12 Example The theory of infinite sets is not finitely axiomatizable. Define
T∞ = ∃≥n x ( x = x ) : n ∈ ω


Every infinite set is a model of T∞ and, vice versa, every model of T∞ is is an infinite
set. Then ccl( T∞ ) is the theory of infinite sets. Suppose for a contradiction that T∞

33
 is finitely axiomatizable. By Proposition 4.11, ∃≥n x ( x = x ) ` T∞ for some n. Any
set of cardinality n + 1 proves that this is not the case. 

The following is a less trivial example. It proves a claim we made in Exercise 1.16.

4.13 Example Write Tgph for the theory of graphs, see Example 2.5. Let K be the fol-
lowing class of structures
n o
M
K = M  Tgph : r ⊆ ( A × ¬ A) ∪ (¬ A × A) for some A ⊆ M
We prove that K is axiomatizable but not finitely axiomatizable, i.e. K = Mod( T )
for some theory T, but T cannot be choosen finite.
A path of length n in M is a sequence c0 , . . . , cn ∈ M such that M  r (ci , ci+1 ) for
every 0 ≤ i < n. A path is closed if c0 = cn . We claim that the following theory
axiomatizes K.
( " # )
2n
^
T = ¬∃ x0 , . . . x2n+1 r ( xi , xi+1 ) ∧ x0 = x2n+1 : n∈ω
i =0
In words, T says that all closed paths have even length. Inclusion K ⊆ Mod( T ) is
clear, we prove Mod( T ) ⊆ K. Let M  T and let Ao ⊆ M contain exactly one point
for every connected component of M. Define
−1
( " # )
2n
^
A = b : M  ∃ x0 , . . . , x2n r ( xi , xi+1 ) ∧ a= x0 ∧ x2n =b , a ∈ Ao , n ∈ ω
i =1

We claim that r M ⊆ ( A × ¬ A) ∪ (¬ A × A), hence M ∈ K. We need to verify that

if r (b, c) then neither b, c ∈ A nor b, c ∈ ¬ A. Suppose for a contradiction that
r (b, c) and b, c ∈ A (the case b, c ∈ ¬ A is similar). As b and c belong to the same
connected component, there are two paths b0 , . . . , b2n and c0 , . . . , c2m that connect
a = b0 = c0 ∈ Ao to b = b2n and c = c2m . Then a, b1 , . . . , b2n , c2m , . . . , c1 , a is a closed
path of odd length. A contradiction.
We now prove that K is not finitely axiomatizable. By Proposition 4.11 it suffices to
note that no finite S ⊆ T axiomatizes K. 

34
 Chapter 5
Types and morphisms
In Section 1 and 2 we introduce distributive lattices and prime filters and prove
Stone’s representation theorem for distributive lattices. In Section 3 we discuss lat-
tices that arise from sets of formulas and their prime filters (prime types). These
sections are mainly required for the discussion of Hilbert’s Nullstellensatz in Sec-
tion 8.7 below.
There is a lot of notation in this chapter. The reader may skim through and return
to it when we refer to it.

1 Semilattices and filters

A preorder is a set P with a transitive and symmetric relation which we usually
denote by ≤. If A, B ⊆ P we write A ≤ B if a ≤ b for every a ∈ A and b ∈ B. We
write a ≤ B and A ≤ b for { a} ≤ B and A ≤ {b} respectively.
Quotienting a preorder by the equivalence relation
a∼b ⇔ a≤b and b≤a
gives a (partial) order . We often do not distinguish between a preorder and the
partial order associated to it. Preorders are very common; here we are interested in
the one induced by the relation of logical consequence
ϕ ≤ ψ ⇔ ϕ ` ψ,
or, more generally, by the relation of logical consequence over a theory T , that is
ϕ ≤ ψ ⇔ T ∪ { ϕ} ` ψ.
A partial order P is a lower semilattice if for each pair a, b ∈ P there is a maximal
element c such that c ≤ { a, b}. We call c the meet of a and b. The meet is unique
and is denoted by a ∧ b . Dually, a partial order is an upper semilattice if for each
pair of elements a and b there is a minimal element c such that { a, b} ≤ c. This c is
called the joint of a and b. The joint is unique and is denoted by a ∨ b . A lattice
is simultaneously a lower and an upper semilattice.
An element c such that c ≤ P is called a lower bound or a bottom . An element
such that P ≤ c is called an upper bound or a top . Lower and upper bounds are
unique and will be denoted by 0 , respectively 1 . Other symbols common in the
literature are ⊥ , respectively > . A semilattice is bounded if it has both an upper
and a lower bound.
For the rest of this section we assume that P is a bounded lower semilattice.
The meet is associative and commutative
( a ∧ b) ∧ c = a ∧ (b ∧ c)
a∧b = a ∧ b.

35
 Hence we may unambiguously write a1 ∧ · · · ∧ an . When C ⊆ P is finite, we write
∧C for the meet of all the elements of C. We agree that ∧∅ = 1.

In an upper semilattice, the dual properties hold for the joint. We write ∨C for the
joint of all elements in C and we agree that ∨∅ = 0.
A filter of P is a non-empty set F ⊆ P that satisfies the following for all a, b ∈
P
f1. a ∈ F and a ≤ b ⇒ b ∈ F

f2. a, b ∈ F ⇒ a ∧ b ∈ F.
We say that F is a proper filter if F 6= P, equivalently if 0 ∈/ F. We say that F is

principal if F = b : a ≤ b for some a ∈ P. More precisely, we say that F is the
principal filter generated by a . A proper filter F is maximal if there is no filter H
such that F ⊂ H ⊂ P.
We say that a filter F is maximal relative to c if c ∈
/ F and c ∈ H for every
H ⊃ F. So, a filter F is maximal if it is maximal relative to 0. We say that F is
relatively maximal when it is maximal relative to some c.
For B ⊆ P we define the filter generated by B to be the intersection of all the filters
containing B. It is easy to verify that this is indeed a filter. When B is a finite set,
the filter generated by B is the principal filter generated by ∧ B. In general we have
the following.

5.1 Proposition For every B ⊆ P, the filter generated by B is the set


a : ∧C ≤ a for some finite non-empty C ⊆ B . 

This has the following important consequence.

5.2 Proposition Let B ⊆ P and let c ∈ P. If ∧C 6≤ c for every finite non-empty C ⊆ B, then
B is contained in a maximal filter relative to c.
Proof Let F be the set of filters F such that B ⊆ F and c ∈
/ F. By Proposition 5.1,
F is non-empty. It is immediate that F is closed under unions of arbitrary chains.
Then, by Zorn’s lemma, F has a maximal element. 

5.3 Exercise Prove that the following are equivalent

1. F is maximal relative to c;

2. for every a ∈
/ F there is d ∈ F such that d ∧ a ≤ c. 

5.4 Exercise (A generalization of the exercise above.) Let B ⊆ P and let c ∈ P be

such that ∧C 6≤ c for every finite non-empty C ⊆ B. Prove that the following are
equivalent
1. B is a maximal filter relative to c;

2. a ∈
/ B ⇒ b ∧ a ≤ c for some b ∈ B. 

5.5 Exercise Let F ⊆ P be a non-principal filter. Is F always contained in a maximal

non-principal filter ? 

36
2 Distributive lattices and prime filters
Let P be a lattice. We say that P is distributive if for every a, b, c ∈ P
a ∧ (b ∨ c) = ( a ∧ b) ∨ ( a ∧ c)
a ∨ (b ∧ c) = ( a ∨ b) ∧ ( a ∨ c)
Throughout this section we assume that P is a bounded distributive lattice.
A proper filter F is prime if for every a, b ∈ P
a∨b ∈ F ⇒ a ∈ F or b ∈ F.

5.6 Proposition Every relatively maximal filter of P is prime.

Proof Let F be maximal relative to c and assume that a ∈ / F and b ∈/ F. Then, by
Exercise 5.3, there are d1 , d2 ∈ F such that d1 a ≤ c and d2 b ≤ c. Let d = d1 ∧ d2 .
∧ ∧

Then d ∧ a ≤ c and d ∧ b ≤ c and therefore (d ∧ a) ∨ (d ∧ b) ≤ c. Hence, by distributivity,

d ∧ ( a ∨ b) ≤ c. Then a ∨ b ∈
/ F. 

The Stone space of P is a topological space that we denote by S(P) . The points of
S(P) are the prime filters of P. The closed sets of the Stone topology are arbitrary
intersections of sets of the form
n o
[ a ]P = F : prime filter such that a ∈ F .
for a ∈ P. In other words, the sets above form a base of closed sets of the Stone
topology. Using 1 and 3 in the following proposition the reader can easily check
that this is indeed a base for a topology.

5.7 Proposition For every a, b ∈ P we have

 
1. 0 P = ∅;
 
2. 1 P = S (P);
     
3. a P ∪ b P = a ∨ b P;
    
4. a P ∩ b P = a ∧ b ]P .
Proof The verification is immediate. Only 3 requires that the filters in S(P) are
prime. 

The closed subsets of S(P) ordered by inclusion form a distributive lattice. The
following is a representation theorem for distributive lattices.

5.8 Theorem The map a 7→ [ a]P is an embedding of P in the lattice of the closed subsets of
S(P). In particular
1. 0 7→ ∅;
2. 1 7 → S (P);
   
3. a ∨ b 7→ a P ∪ b P ;
   
4. a ∧ b 7→ a P ∩ b P .

Proof It is immediate that the map above preserves the order and Proposition 5.7
shows that it preserves the lattice operations. We prove that the map is injective. Let
   
a 6= b, say a  b. We claim that a P * b P . There is a filter F that contains a and is

37
 maximal relative to b. By Proposition 5.6 such an F is prime. Then F ∈ [ a]P r [b]P . 

5.9 Theorem With the Stone topology, S(P) is a compact space.

Proof Let h[ ai ]P : i ∈ I i be basic closed sets such that for every finite J ⊆ I
\
a. [ ai ]P 6= ∅.
i∈ J

We claim that
\
b. [ ai ]P 6= ∅.
i∈ I

By 4 of Proposition 5.7 and a we obtain that ∧C  0 for every finite C ⊆ { ai : i ∈ I }.

By Proposition 5.2, there is a maximal (relative to 0) filter containing { ai : i ∈ I }. By
Proposition 5.6, such filter is prime and it belongs to the intersection in b. 

Let a, b ∈ P. If a ∧ b = 0 and a ∨ b = 1, we say that b is the complement of a (and

vice versa). The complement of an element need not exist. If the complement exists
it is unique, the complement of a is denoted by ¬ a .

5.10 Lemma Let U ⊆ S(P) be a clopen set. Then U = [ a]P for some a ∈ P, and ¬ a exists.
Proof As both U and S(P) r U are closed, for some sets A, B ⊆ P
\
[ x ]P = U
x∈ A
\
[ y ]P = S(P) r U.
y∈ B

By compactness, that is Theorem 5.9, there are some finite A0 ⊆ A and B0 ⊆ B such
that
\ \
[ x ]P ∩ [ y ]P = ∅
x ∈ A0 y∈ B0

Let a = ∧ A0 and b = ∧ B0 . From claim 4 of Proposition 5.7 we obtain [ a]P ∩ [b]P = ∅.

Therefore U ⊆ [ a]P ⊆ S(P) r [b]P ⊆ U. Hence U = [ a]P and b = ¬ a. 

A Boolean algebra is a bounded distributive lattice where every element has a

complement. In a Boolean algebra, the sets [ a]P are clopen and they form also a
base of open set of the topology of S(P). A topology that has a base of clopen sets
is called zero-dimensional . By the following proposition the Stone topology of a
Boolean algebra is Hausdorff.
A proper filter of a Boolean algebra is an ultrafilter if either a ∈ F or ¬ a ∈ F for
every a.

5.11 Proposition Let P be a Boolean algebra. Then the following are equivalent
1. F is maximal;

2. F is prime;

3. F is an ultrafilter.

Proof Implication 2 ⇒ 3 is obtained observing that a ∨ ¬ a ∈ F. The rest is immedi-

ate. 

38
5.12 Exercise Prove that the stone topology on S(P) has a base of open compact sets. 

5.13 Exercise Suppose we had defined S(P) as the set of relatively maximal filters.
What could possibly go wrong ? 

3 Types as filters
This section we work with a fixed set of formulas ∆ all with free variables among
those of some fixed tuple x . In this section we do not display x in the notation.
Subsets of ∆ are called ∆-types .
We associate to ∆ a bounded lattice P(∆) . This is the closure under conjunction
and disjunction of the formulas in ∆ ∪ {⊥, >}. The (pre)order relation in P(∆) is
given by
ψ ≤ ϕ ⇔ T ∪ {ψ} ` ϕ,
for some fixed theory T. In this section, to lighten notation, we absorb T in the
symbol `. If p is a ∆-type we denote by h piP the filter in P(∆) generated by p.

5.14 Lemma For every ∆-type p

n o
h p iP = ϕ ∈ P( ∆ ) : p ` ϕ .

In particular p is consistent if and only if h piP is a proper filter.

Proof Inclusion ⊆ is clear. Inclusion ⊇ is a consequence of the Compactness The-
orem. In fact p ` ϕ implies that ψ ` ϕ for some formula ψ that is conjunction of
formulas in p. Then ϕ ∈ h piP follows from ψ ∈ h piP and ψ ≤ ϕ. 

We say that p ⊆ ∆ is a principal ∆-type if h piP is a principal. The following

lemma is an immediate consequence of the Compactness Theorem. Note that in 3
the formula ϕ is arbitrary, possibly not even in P(∆).

5.15 Lemma For every ∆-type p the following are equivalent

1. p is principal;

2. ψ ` p ` ψ for some formula ψ (here ψ is any formula, it need not be in ∆);

3. ϕ ` p where ϕ is conjunction of formulas in p. 

Proof Implications 1⇒2 is immediate by Lemma 5.14. To prove 2⇒3 suppose ψ `

p ` ψ. Apply compactness to obtain a formula ϕ, conjunction of formulas in p,
such that ϕ ` ψ. Implications 3⇒1is trivial. 

5.16 Definition We say that p ⊆ ∆ is a prime ∆-type if h piP is a prime filter. We say that p
is a complete ∆-type if h piP is a maximal filter. 

Though in general neither ∆ nor P(∆) are closed under negation, Lemma 5.14 has
the following consequence.

5.17 Proposition For every consistent ∆-type p the following are equivalent
1. p is complete;

39
 2. p is consistent and either p ` ϕ or p ` ¬ ϕ for every formula ϕ ∈ ∆.

Proof 2⇒1. Assume 2. As p is consistent, h piP is a proper filter. To prove that it is

maximal suppose ϕ ∈ / h piP . Then p 0 ϕ and from 2 it follows that p ` ¬ ϕ. Hence
no proper filter contains p ∪ { ϕ}.
1⇒2. As h piP is proper, p is consistent. Suppose p 0 ϕ. Then ϕ ∈
/ h piP and, as
h piP is maximal, p ∪ { ϕ} generates the improper filter. Then p ∪ { ϕ} is inconsistent.
Hence p ` ¬ ϕ. 

Given a model M and a tuple c ∈ M| x| the ∆-type of c in M is the sets

n o
∆-tp M (c) = ϕ( x ) ∈ ∆ : M  ϕ(c)
When the model M is clear from the context we omit the subscript. When x and c
are the empty tuple, we write Th∆ ( M) for ∆-tp M (c).

5.18 Lemma For every ∆-type p the following are equivalent

1. p is prime;

¬ ϕ : ϕ ∈ ∆ such that p 0 ϕ

2. p ∪ is consistent;
ϕ∈∆:p`ϕ = ∆-tp M (c) for some model M and some tuple c ∈ M| x| .

3.

Proof Implications 2⇒3⇒1 are clear, we prove 1⇒2. By compactness if the type
in 2 is inconsistent then there are finitely many formulas ϕ1 , . . . , ϕn ∈ ∆ such that
p 0 ϕi and
n
_
p ` ϕi .
i =1
hence p is not prime. 

The following corollary is immediate. When it comes to verifying that a given

∆-type is prime, it simplifies the proof.

5.19 Corollary For every ∆-type p the following are equivalent

1. p is prime;
n
for every n and every ϕ1 , . . . , ϕn ∈ ∆.
_
2. p ` ϕi ⇒ p ` ϕi for some i ≤ n, 
i =1

The set ∆ above contains only formulas with variables among those of the tuple x.
In the following it is convenient to consider sets ∆ that are closed under substitution
of variables with any other variable. The set of prime ∆-types is denoted by S(∆) ,
and we write Sx (∆) when we restrict to types in with variables among those of the
tuple x.
The most common ∆ used in the sequel is the set of all formulas in L( A) and the
underlying theory is Th( M/A) for some given model M containing A. In this case
we write tp M (c/A) for ∆-tp M (c) or, when A is empty, tp M (c) . The set Sx (∆) is
denoted by Sx ( A). The topology on Sx ( A) is generated by the clopen
  n o
ϕ( x ) = p ∈ Sx ( A) : ϕ( x ) ∈ p .
Sometimes ∆ is the set of all formulas of a given syntactic form. Then we use some

40
 more suggestive notation that we summarize below.

5.20 Notation The following are some of the most common ∆-types and ∆-theories
1. at-tp(c) , That ( M ) when ∆ = Lat
2. at± -tp(c) , That± ( M ) when ∆ = Lat±
3. qf-tp(c) , Thqf ( M ) when ∆ = Lqf .
Clearly, the types/theories in 2 and 3 are equivalent. Most used is the theory
That± ( M/M ) which is called the diagram of M and has a dedicated symbol: Diag( M ) . 

5.21 Remark Let A ⊆ M ∩ N. The following are equivalent

1. N  Diagh Ai M ;

2. h Ai M is a substructure of N. 

4 Morphisms
First we set the meaning that the word map has in these notes.

5.22 Definition A map consists a triple f : M → N where

1. M is a set (usually a structure) called the domain of the map ;

2. N is a set (usually a structure) called the codomain the map ;

3. f is a function with domain of definition dom f ⊆ M and image img f ⊆ N.

By cardinality of f : M → N we understand the cardinality of the function f . 

If dom f = M we say that the map is total ; if img f = N we say that it is surjective .
The composition of two maps and the inverse of a map are defined in the obvious
way.

5.23 Definition Let ∆ be a set of formulas with free variables in the tuple x = h xi : i < λi.
The map h : M → N is a ∆-morphism if it preserves the truth of all formulas in ∆. By
this we mean that
p. M  ϕ( a) ⇒ N  ϕ( ha) for every ϕ( x ) ∈ ∆ and every a ∈ (dom h)| x| .
Notation: if a is the tuple h ai : i < λi then ha is the tuple hhai : i < λi. 

When ∆ = L we say elementary map for ∆-morphism. When ∆ = Lat we

say partial homomorphism and when ∆ = Lat± we say partial embedding or
partial isomorphism . The reason for the latter name is explained in Remark 5.24.
It is immediate to verify that a partial embedding which is total is an embedding
(so, there is not conflict with Definition 2.19 above). Similarly, a partial embedding
which is a total and surjective is an isomorphism. The precise connection between
partial and total homo/iso-morphisms is discussed in following remark.

5.24 Remark For every map h : M → N the following are equivalent

1. h : M → N is a partial isomorphism;

41
 2. there is an isomorphism k : hdom hi M → himg hi N that extends h.

Moreover, k unique. The equivalence holds replacing isomorphism by homomor-

phism (in which case we in 2 we obtain an surjection). The extension k is obtained


defining k t( a) = t(ha).
A similar fact holds for partial homomorphims with epimorphism (surjective ho-
momorphism) for isomorphism.
Unfortunately, there is no hope to generalize this fact to ∆-morphisms as, in general,
there is no notion of generated ∆-elementary substructure. 

We use ∆-morphisms to compare, locally, two structures (or two loci of the same
structure). There are different ways to do this, in the proposition below we list a
few synonymous expressions. But first some more notation. When x is a fixed tuple
of variables, a ∈ M| x| and b ∈ N | x| we write
M, a V∆ N, b if M  ϕ( a) ⇒ N  ϕ(b) for every ϕ( x ) ∈ ∆.
M, a ≡∆ N, b if M  ϕ( a) ⇔ N  ϕ(b) for every ϕ( x ) ∈ ∆.
The following equivalences are immediate and will be used without explicit refer-
ence.

5.25 Proposition For every given set of formulas ∆ and every map h : M → N the following
are equivalent
1. h : M → N is a ∆-morphism;

2. M, a V∆ N, ha for every a ∈ (dom h)| x| ;

3. ∆-tp M ( a) ⊆ ∆-tp N ( ha) for every a ∈ (dom h)| x| ;

4. N, ha  p( x ) for every a ∈ (dom h)| x| and p( x ) = ∆-tp M ( a). 

5.26 Remark Condition p in Definition 5.23 apply to tuples x of any length, in particular
to the empty tuple. In this case ϕ( x ) is a sentence, a ∈ (dom h)0 = {∅} is the empty
tuple, and p asserts that Th∆ ( M ) ⊆ Th∆ ( N ). When h = ∅ this is actually all what p
says. In fact ∅| x| = ∅ unless | x | = 0. Still, Th∆ ( M ) ⊆ Th∆ ( N ) may be a non trivial
requirement. 

5.27 Definition We call Th∆ ( M ) the ∆-characteristic of M. We say that T decides the
∆-characteristic if Th∆ ( M) = Th∆ ( N ) for all M, N  T. In other words, for any pair of
models of T, the empty map ∅ : M → N is a ∆-morphism. When ∆ = Lat± we say simply
characteristic , then T decides the characteristic h∅i M ' h∅i N for all M, N  T. 

We conclude this section with a couple of propositions that break Theorem 2.20
into parts. More interestingly, in Chapter 10 we shall prove a sort of converse of
Propositions 5.29 and 5.30.
This is not immediately required in the following. There is a relation between
some properties of ∆-morphisms closure of ∆ logical connectives. When C ⊆
{∀, ∃, ¬, ∨, ∧} is a set of connectives, we write C∆ for the closure of ∆ with re-
spect to all connectives in C. We write ¬∆ for the set containing the negation of
the formulas in ∆. Warning: do not confuse ¬∆ with {¬}∆.

42
 It is clear that ∆-morphisms are {∧, ∨}∆-morphisms.

5.28 Proposition For every given set of formulas ∆ and every injective map h : M → N the
following are equivalent
a. h : M → N is a ¬∆-morphism;

b. h−1 : M → N is a ∆-morphism. 

5.29 Proposition For every set of formulas ∆, every ∆-embedding h : M → N is a {∃}∆-mor-

phism.
Proof Formulas in {∃}∆ ave the form ∃y ϕ( x, y) where y is a finite tuples of vari-
ables and ϕ( x, y) ∈ ∆. For every tuple a ∈ (dom h)| x| we have:
M  ∃y ϕ( a, y) ⇒ M  ϕ( a, b) per una tupla b ∈ M|y|
⇒ N  ϕ(ha, hb)
⇒ N  ϕ(ha, c) per una tupla c ∈ N |y|
⇒ N  ∃y ϕ(ha, y).
Note that the second implication requires the totality of h : M → N which guaran-
tees that ϕ( a, y) has a solution in dom h. 

When h : M → N is injective the following is a corollary of Propositions 5.28

and 5.29. The general proof is the ‘dual’ of that of Propositions 5.29.

5.30 Proposition For every set of formulas ∆, every surjective ∆-morphism h : M → N is a

{∀}∆-morphism. 

43
 Chapter 6
Some relational structures
In the first section we prove that theory of dense linear orders without endpoints is
ω-categorical. That is, any two such countable orders are isomorphic. This is an
easy classical result of Cantor. In this chapter we examine Cantor’s construction (a
so-called back-and-forth construction) in great detail. In the second section we apply
the same technique to prove that the theory of the random graph is ω-categorical.

1 Dense linear orders

The language of strict orders , which in this section we denote by L, contains only a
binary relation symbol <. A structure M of signature L is a strict order if it models
(the universal closure of) the following formulas
1. x 6< x irreflexive;
2. x < z < y → x < y transitive.
Note that the following is an immediate consequence of 1 and 2.
x < y → y 6< x antisymmetric.
We say that the order is total or linear if
li. x < y ∨ y < x ∨ x = y linear or total.
An order is dense if
nt. ∃ x, y ( x < y) non trivial;
d. x < y → ∃z ( x < z < y) dense.
We need to require the existence of two comparable elements, then d implies that
dense orders are in fact infinite. We say that the ordering has no endpoints if
e. ∃y ( x < y) ∧ ∃y (y < x ) without endpoints.
We denote by Tlo the theory strict linear orders and by Tdlo the theory of dense
linear orders without endpoints. Clearly, these are consistent theories: Q with the
usual ordering is a model of Tdlo .
We introduce some notation to improve readability of the proof of the following
theorem. Let A and B be subsets of an ordered set. We write A < B if a < b for
every a ∈ A and b ∈ B. We write a < B e A < b for { a} < B, respectively A < {b}.
Let M  Tlo . Then M  Tdlo if and only if for every finite A, B ⊆ M such that A < B
there is a c such that A < c < B. In fact axiom d is evident and axioms nt and e are
obtained taking replacing A and/or B by the empty set.
Now we prove the first of a series of lemmas that we call extension lemmas . Re-
call that in the language of strict orders an injective map k : M → N is a partial
isomorphism if

44
 M  a < b ⇔ N  ka < kb for every a, b ∈ dom k.
(When M, N  Tlo direction ⇒ suffices.)

6.1 Lemma Fix M  Tlo and N  Tdlo . Let k : M → N be a finite partial isomorphism and
let b ∈ M. Then there is a partial isomorphism h : M → N that extends k and is defined in
b.
Proof Given a finite partial isomorphism k : M → N define
A− = a ∈ dom(k) : a < b ;


A+ = a ∈ dom(k) : b < a .


The sets A− and A+ are finite and partition dom k, and A− < A+ . As k : M → N
is a partial isomorphism, k[ A− ] < k [ A+ ]. Then in N there is an element c such that
k[ A− ] < c < k [ A+ ]. It is easy to check that setting h = k ∪ hb, ci we obtain the


required extension. 

The following is an equivalent version of Lemma 6.1.

6.2 Corollary Let M  Tlo be countable and let N  Tdlo . Let k : M → N be a finite partial
isomorphism. Then there is a (total) embedding h : M ,→ N that extends k.
Proof Let h ai : i < ω i be an enumeration of M. Define by induction a chain of finite
partial isomorphisms hi : M → N such that ai ∈ dom hi+1 . The construction starts
with h0 = k. At stage i + 1 we chose any finite partial isomorphism hi+1 : M → N
that extends hi and is defined in ai . This is possible by Lemma 6.1. In the end we
set
[
h = hi .
i ∈ω
It is immediate to verify that h : M ,→ N is the required embedding. 

6.3 Exercise Prove that the extension Lemma 6.1 characterizes models of Tdlo among
models of Tlo . That is, if N is a model of Tlo such that the conclusion of Lemma 6.1
holds, then M  Tdlo . 

We are now ready to prove that any two countable models of Tdlo are isomorphic
which is a classical result of Cantor’s. Actually what we prove is slightly more
general than that. In fact Cantor’s theorem is obtained from the theorem below by
setting k = ∅, which we are allowed to, because all models have the same empty
characteristic (cfr. Remark 5.26).

6.4 Theorem Every finite partial isomorphism k : M → N between countable models of Tdlo
∼
extends to an isomorphism g : M → N.

The following is the archetypal back-and-forth construction. It is important to note

that it does not mention linear orders at all. It only uses the extension Lemma 6.1.
In many other contexts, as soon as we have an extension lemma, we can apply
verbatim the same construction (cfr. Theorem 7.6).
Proof Fix some enumerations of M and N, say h ai : i < ω i, respectively hbi : i < ω i.
We define by induction a chain of finite partial isomorphisms gi : M → N such that
ai ∈ dom gi+1 and bi ∈ img gi+1 . In the end we set

45
 [
g = gi
i ∈ω
We begin by setting g0 = k. The inductive step consists of two half-steps that we
may call the forth-step and back-step. In the forth-step we define gi+1/2 such that
ai ∈ dom gi+1/2 . in the back-step to define gi+1 such that bi ∈ img gi+1
By the extension lemma 6.1 there is a finite partial isomorphism gi+1/2 : M → N that

−1
extends gi and is defined in ai . Now apply the same lemma to extend gi+1/2 :
N → M to a finite partial isomorphism ( gi+1 )−1 : N → M defined in bi . 

Let λ be an infinite cardinal. We say that a theory is λ-categorical if any two

models of T of cardinality λ are isomorphic. From Theorem 6.4, taking k = ∅, we
obtain the following.

6.5 Corollary The theory Tdlo is ω-categorical. 

We also obtain that Tdlo is a complete theory. This is consequence of the following
general fact.

6.6 Proposition If T is λ-categorical, for some λ ≥ | L|, then T is complete.

Proof Let M and N be any two models of Tdlo . Applying the upward and/or
downward Löwenheim-Skolem theorem, we may assume they both have cardinality
λ. (Here we use that λ ≥ | L|.) Hence M ' N and in particular M ≡ N. 

6.7 Exercise Prove that Tdlo is not λ-categorical for any uncountable λ. 

6.8 Exercise Prove that, in the language of strict orders, Q  R. 

6.9 Exercise Let L be the language of strict orders augmented with countably many

constants ci : i ∈ ω . Let T be the theory that extends Tdlo with the axioms
ci < ci+1 for all i. Prove that T is complete. Find three non isomorphic models of
this theory. For N a rightly chosen model of T, prove the statement in Lemma 6.1,
with M any model of T. 

6.10 Exercise Show that in Theorem 6.6 the assumption λ ≥ | L| is necessary. (Hint.
Let ν be an uncountable cardinal. The language contains only the ordinals i < ν
as constants. The theory T says that there are infinitely many elements and either
i = 0 for every i < ν, or i 6= j for every i < j < ν. Prove that T is ω-categorical but
incomplete.) 

2 Random graphs
Recall that the language of graphs , which in this section we denote by L, contains
only a binary relation r. A graph structure of signature L such that
1. ¬r ( x, x ) irreflexive;
2. r ( x, y) → r (y, x ) symmetric.

46
 An element of a graph M is called a vertex or a node . An edge is an unordered
pair of vertices { a, b} ⊆ M such that M  r ( a, b). In words we may say that a is
adjacent to b.
A random graph is a graph that also satisfies the following axioms for every n
nt. ∃ x, y, z ( x 6= y 6= z 6= x ) non trivial;
n
^ n 
^
for every n ∈ Z+ .

rn . xi 6 = y j → ∃ z r ( xi , z) ∧ ¬r (z, yi )
i,j=1 i =1

The theory of graphs is denoted by Tgph and the theory of random graphs is
denoted by Trg . The scheme of axioms rn plays the same role of density in the
previous section. It says that given two non empty disjoint sets A+ and A− of
cardinality ≤ n there is a vertex z that is adjacent to all vertices in A+ and to no
vertex in A− . Note that it might well be that z ∈ A− , however z ∈ A+ would
contradict 1.
Some readers may doubt if Trg is consistent.

6.11 Proposition There exists a random graph.

Proof The domain of N is the set of natural numbers. Let N  r (n, m) hold if the
n-th prime number divides m or vice versa m-th prime number divides n. 

6.12 Proposition Random graphs are infinite.

Proof Suppose for a contradiction that M is a finite random graph. As M contains
at least 3 elements, we can fix two distinct elements a, b ∈ M and require a z
adjacent to all elements of M r { a, b} and no element of { a, b}. A remarked above,
z ∈ { a, b}, then a and b are not adjacent. Now let w be adjacent to all vertices in
M r { a} but not to a. The only option is w = b, but this implies that a is adjacent
to b. Contradiction. 

The following is the analogous of Lemma 6.1 for random graphs. Recall that in the
language of graphs a map k : M → N is a partial isomorphism if it is injective and
M  r ( a, b) ⇔ N  r (ka, kb) for every a, b ∈ dom k.

6.13 Lemma Fix M  Tgph and N  Trg . Let k : M → N be a finite partial isomorphism and
let b ∈ M. Then there is a partial isomorphism h : M → N that extends k and is defined in
b.
Proof The schema of the proof is the same as in Lemma 6.1, so we use the same
notation. Assume b ∈
/ dom k and define

A+ = x ∈ dom k : M  r ( x, b) e A− = y ∈ dom k : M  ¬r (y, b) .

 

These two sets are finite and disjoint, then so are k[ A+ ] and k [ A− ]. Then there is a
c such that
^ ^
r (ka, c) ∧ ¬r (ka, c).
a∈ A+ a∈ A−

As k [ A+ ] ∪ k [ A− ] = img k, it is immediate to verify that h = k ∪ hb, ci


is the
required extension. 

47
 With the same proof of Corollary 6.2 obtain the following.

6.14 Corollary Let M  Tgph be countable and let N  Trg . Let k : M → N be a finite partial
isomorphism. Then there is a (total) embedding h : M ,→ N that extends k. 

The proof of Theorem 6.4 gives the following theorem and its corollary.

6.15 Theorem Every finite partial isomorphism k : M → N between models of Trg extends to
∼
an isomorphism g : M → N. 

6.16 Corollary The theory Trg is ω-categorical (and therefore complete). 

6.17 Exercise Let a, b, c ∈ N  Trg . Prove that r ( a, N ) = r (b, N ) ∩ r (c, N ) occurs only in
the trivial case a = b = c. 

6.18 Exercise Prove that for every b ∈ N  Trg the set r (b, N ) is a random graph.
Assume N is countable. Is every random graph M ⊆ N of the form r (b, N ) for
some b ∈ N ? 

6.19 Exercise Let N be free union of two random graphs. That is, N = N1 t N2 and
r N = r N1 t r N2 , with N1 and N2 some countable random graphs. By t we denote the
disjoint union. Prove that N is not a random graph. Show that N1 is not definable
without parameters. Write a first order sentence ψ( x, y) true if x and y belong to
the same connected component of N. Axiomatize the class K of graphs that are free
union of two random graphs. 

6.20 Exercise Prove that Trg is not λ-categorical for any uncountable λ. Hint. Prove that
there is a random graph N of cardinality λ where every vertex is adjacent to < λ
vertices. Compare it with its complement graph (the graph that has edges between
pairs that are non adjacent in N). 

6.21 Exercise Prove that Trg is not finitely axiomatizable. Hint. Given a random graph
N and a set P of cardinality n + 1 show that you can add edges to M t P and make
it satisfy axiom rn but not rn+1 . 

6.22 Exercise Let A ⊆ N  Trg and let ϕ( x ) ∈ L( A), where | x | = 1. Prove that if ϕ( N )
is finite then ϕ( N ) ⊆ A. 

6.23 Exercise (Peter J. Cameron) Prove that for every countable graph M the following
are equivalent
1. M is either the infinite complete graph, the infinite empty graph, or the random
graph;
2. if M1 , M2 ⊆ M are such that M1 t M2 = M, then M1 ' M or M2 ' M.

Hint. For 2⇒1 first show that if r ( a, M) = ∅ for some a ∈ M then the graph is null
and if r (b, M ) = M r {b} for some b, then the graph is complete. Clearly, 2 implies
that any finite partition of M contains an element isomorphic to M. It follows thats

48
 T
if 2 and there is a finite A such that a∈ A r ( a, M ) = ∅ then the graph is null and if
there is a finite B such that B ∪ b∈ B r (b, M) = M then the graph is complete.
S

Suppose M is not a random graph. Fix some finite, disjoint A and B such that no
c satisfies both r ( A, c) = A and r ( B, c) = ∅. Let M1 = {c : r ( A, c) 6= A} r B and
M2 = B ∪ {c : r ( B, c) 6= ∅} r M1 . This is a partition of M. Now note that
\ [
r ( a, M1 ) = ∅ and r (b, M2 ) = M2 . 
a∈ A b∈ B

3 Notes and references

The following is a well-written accessible survey on the amazing model theoretic
properties of the random graph.

[1] Peter J. Cameron, The random graph (2013), arXiv:1301.7544.

49
 Chapter 7
Fraïssé limits
We introduce Fraïssé limits, aleas homogeneous-universal or generic structures, which
here we call rich models, after Poizat. Rich models generalize the examples in
Chapters 6 and the many more to come.
Elimination of quantifiers is briefly discussed at the end of Section 1. For the time
being we identify quantifier elimination with the property that says that all partial
isomorphisms are elementary maps. Proofs are easier with this notion in mind. The
equivalence with its syntactic counterpart is only proved in Chapter 10, when the
reader is more familiar with arguments of compactness.

1 Rich models.
We now define categories of models and partial morphisms. These are example of con-
crete categories as intended in category theory. However, apart from the name, in
what follows we dispense with all notions of category theory as they would make
the exposition less basic than intended (without providing more technical instru-
ments).
B Warning: the terminology introduced in this section is not standard. In the litera-
ture many details are left implicit. This is generally safe when the category used
is fixed. Here we prefer to be more explicit because, when discussing saturation,
quantifier elimination, and model completeness, it helps to compare different cate-
gories.
A category (of models and partial morphisms) is a class M which is disjoint union
of two classes: Mob and Mhom . The first is the class of objects and contains
structures with a common signature L which we call models . The second is the
class of morphisms and contains (partial) maps between models. We require that
the identity maps are morphisms and that composition of two morphism is again
morphism. This makes M a well-defined category.
For example, M could consist of all models of some theory T0 and of all partial iso-
morphisms between these. Alternatively, as morphisms we could take elementary
maps between models. At a first reading the reader may assume M is as in one of
these two examples. In the general case we need to make some assumptions on M.

7.1 Definition For ease of reference we list together all properties required below
c1. the identity map id A : M → M is a morphism, for any A ⊆ M;

c2. if k 0 : M → N is a morphism for every finite k 0 ⊆ k, then k : M → N is a morphism.

c3. morphisms are invertible maps and the inverse of a morphism is a morphism;

c4. morphisms preserve the truth of Lat -formulas;

c5. if M is a model, every elementary map k : M → N is morphism (and, N is a model). 

50
 The connected component of a model M is the subclass of models N such that
there is any morphism with domain M and codomain N (or vice versa, by c3). By
axiom c1 the restriction of a morphism is a morphism, therefore M and N are in
the same connected component if and only if the empty map ∅ : M → N is a
morphism. If the whole category M consists of one connected component we say
that M is connected .
We call c2 the finite character of morphisms . Note that it implies the following
[
c6. if k i : M → N is a chain of morphisms, then k i : M → N is a morphism.
i <λ

A consequence of c5 is that Mob is closed under elementary equivalence. In fact, if

N ≡ M then ∅ : M → N is elementary, hence it is a morphism and, in particular,
and N is a model.
Notably, the following two definitions require c3. The generalization to non injec-
tive morphisms is not straightforward (in fact, there are two generalizations: pro-
jective and inductive). These generalizations are not very common and will not be
considered here.

7.2 Definition Assume that M satisfies c1-c3 of Definition 7.1. We say that a model N is
λ-rich if for every model M, every b ∈ M and every morphism k : M → N of cardinality
< λ there is a c ∈ N such that k ∪ {hb, ci} : M → N is a morphism. We say that N is
rich if it is λ-rich for λ = | N |. When Mob = Mod( T0 ) for some theory T0 and Mhom is
clear from the context, we say rich model of T0 . 

Rich models are also called Fraïssé limits or homogeneous-universal for a reason that
will soon be clear; they are also called generic. Unfortunately these names are either
too long or too generic, so we opt for the less common term rich that was proposed
by Poizat.
The following two notions are closely connected with richness.

7.3 Definition Assume that M satisfies c1-c3 of Definition 7.1. We say that a model N is
λ-universal if for every model M of cardinality ≤ λ in the same connected component
of N there is an embedding k : M ,→ N. We say that a model N is λ-homogeneous if
every k : N → N of cardinality < λ extends to a bijective morphism h : N ∼ → N (an
automorphism when c4 below holds).
Note that the larger Mhom , the stronger notion of homogeneity. When Mhom contains all
partial isomorphisms between models (the largest class of morphisms considered here), it is
common to say λ-ultrahomogeneous for λ-homogeneous.
As above, when λ = | N | we say universal , homogeneous and ultrahomogeneous . 

In Section 6.1 we implicitly used Mob = Mod( Tlo ) and partial isomorphisms as
Mhom . In Section 6.2 we used Mob = Mod( Tgph ) and again partial isomorphisms as
Mhom . Corollary 6.2 proves that every model of Tdlo is ω-rich. Corollary 6.14 claims
the analogous fact for Trg .
In the following we frequently work under the following assumption (even when
not all properties are strictly necessary).

51
7.4 Assumption Assume | L| ≤ λ and suppose that M satisfies c1-c5 of Definition 7.1

The assumption on the cardinality L is only necessary to apply the downward

Löwenheim-Skolem Theorem when required.

7.5 Proposition (Assume 7.4) The following are equivalent

1. N is a λ-rich model;

2. for every model M of cardinality ≤ λ and every morphism k : M → N of cardinality

< λ there is a morphism h : M → N that extends k.
Proof Closure under union of chains of morphisms, which is ensured by c6, im-
mediately yields 1⇒2. As for implication 2⇒1 we only need to consider the case
λ < | M|. By the downward Löwenheim-Skolem theorem there is an M0  M of
cardinality λ containing b. Let h : M0 ,→ N be the embedding obtained from 1. By
c4, the map h : M → N is a composition of morphisms, hence a morphism. 

The following theorem subsumes both Theorem 6.4 and Theorem 6.15.

7.6 Theorem (Assume 7.4) Let M and N be two rich models of the same cardinality λ. Then
every morphism k : M → N of cardinality < λ extends to an isomorphism.
Proof When λ = ω, we can take the proof of Theorem 6.4 and replace partial isomor-
phism by morphism and the references to Lemma 6.1 by references to Proposition 7.5.
As for uncountable λ, we only need to extend the construction through limit stages.
By c6 we can simply take the union. 

7.7 Corollary (Assume 7.4) Then all rich models of cardinality λ in the same connected
component are isomorphic. 

It is obvious that λ-rich models are λ-universals and, by Theorem 7.6, they are λ-ho-
mogeneous. These two notions are weaker than richness. For instance, when M is
as in Section 6.2, the countable graph that has no arc is trivially ultrahomogeneous
but it is not universal and a fortiori not rich. On the other hand if we add to a
countable random graph an isolated point we obtain an universal graph which is
not ultrahomogeneous. However, when taken together, these two properties are
equivalent to richness.

7.8 Theorem (Assume 7.4) The following are equivalent:

1. N is rich;

2. N is homogeneous and universal.

Proof Implication 1⇒2 is clear as noted above, so we prove 2⇒1. We use the
characterization of richness given in Proposition 7.5. Let k : M → N be a morphism
such that |k| < | N | and | M | ≤ | N |. As N is universal, there is a total morphism
f : M ,→ N. By c3 the map k ◦ f −1 : N → N is a morphism of cardinality < | N |. By
∼
homogeneity it has an extension to an automorphism h : N → N. It is immediate
that h ◦ f : M ,→ N is the required extension of k. 

7.9 Exercise Let N be the structure obtained by adding to a countable random graph
an isolated point. Show that N is homogeneous if morphisms are elementary maps

52
 but it is not if morphisms are simply partial isomorphisms. 

A consequence of Theorem 7.6 is that morphisms between rich models of the same
cardinality are elementary maps. However, we do not know what happens when
the models have different cardinality or when they are merely λ-rich. Next theorem
extend this property to any pair of λ-rich models.
To test the theorem below in a simple case we propose the following exercise.

7.10 Exercise Every partial isomorphism k : M → N between models of Tdlo (or mod-
els of Trg ) is an elementary map. Hint: use downward Löwenheim-Skolem and
Theorem 6.4. 

7.11 Theorem (Assume 7.4) Every morphism between λ-rich models is elementary. In partic-
ular, λ-rich models in the same connected component are elementary equivalent.
Proof Let k : M → N be a morphism between rich models. It suffices to prove that
every finite restriction of k is elementary. By c1, we may as well assume that k itself
is finite. It suffices to construct M0  M and N 0  N together with a morphism
h : M → N that extends k and maps M0 bijectively to N 0 . Then by c5 the map
∼
h : M0 → N 0 is the composition of morphisms, hence it is a morphism. Finally, by
c4, it is an isomorphism, in particular an elementary map.
We construct simultaneously a chain h Ai : i <λi of subsets of M, a chain h Bi : i <λi
of subsets of N, and a chain functions hhi : i <λi such that hi : M → N are mor-
phisms, dom hi ⊆ Ai and img hi ⊆ Bi . In the end we set
M0 = N0 =
[ [ [
Ai , Bi , h = hi .
i <λ i <λ i <λ
We interweave the Löwenheim-Skolem constraction 2.40 with the usual back-and-
forth-argument.
The chains start with A0 = dom k, B0 = img k and h0 = k. At limit stages we
take the union. Now we consider successor stages. Fix some enumerations of M
and N. At stage i + 1 we also fix some enumerations of Ai , Bi , L x ( Ai ), and L x ( Bi ),
where | x | = 1. Let hi1 , i2 i be the i-th pair of ordinals < λ. If the i2 -th formula in
L x ( Ai1 ) is consistent in M, pick any solution a. Also, if the i2 -th formula in L x ( Bi1 )
is consistent in N, pick any solution b. Let a0 be the i2 -th element of Ai1 and let b0
be the i2 -th element of Bi1 .
Let hi+1 : M → N be an extension of hi such that dom hi+1 = dom hi ∪ a0 and

 0
img hi+1 = img hi ∪ b . Note that the latter requirement is met applying Proposi-
tion 7.5 to hi−1 : N → M. Define Ai+1 = Ai ∪ a, hi−+11 b0 and Bi+1 = Bi ∪ b, hi+1 a0 .
 

The elements a and b, added to M0 and N 0 , ensure that M0  M and N 0  N, by

the Tarski-Vaught test. In fact every formula in L( M0 ) belongs to L( Ai1 ) for some
i1 and it occurs as i2 - formula in the enumeration that has been fixed along with
the construction. Therefore at stage i = hi1 , i2 i a witness is added to M0 . The same
happens for formulas in L( N 0 ).
A similar argument shows that the elements a0 and b0 , added to the domain and
∼
range of h, ensure that the h : M0 → N 0 is both total and surjective. 

The theory of rich models of M is the set T1 of sentences that hold in all rich mod-
els (or equivalently | L|-rich models), assuming they exist. By the theorem above, T1

53
 is complete as soon as M is connected.
Assume for simplicity that L is countable and that M is connected. Then all ω-rich
models belong to Mod( T1 ) for some complete theory T1 . It is interesting to ask
if the converse is true: if T1 is the theory of the ω-rich models of M, do all mod-
els in M ∩ Mod( T1 ) are rich? The answer is affirmative when T1 is either Tdlo or
Trg , but this is not always the case, see Example 7.15 for an easy counterexample.
An affirmative answer implies that the theory of rich models is ω-categorical. An
interesting variant of this question is considered in Theorem 9.8.
For convenience of the reader, we instantiate Theorem 7.11 with the two categories
used in Chapter 6.

7.12 Corollary Partial isomorphisms between models of Tdlo and between models of Trg are
elementary maps. 

When partial isomorphism between models of a given theory T coincide with ele-
mentary maps, it is always by a fundamental reason. Let us introduce some termi-
nology. Let T be a consistent theory. We say that T has (or admits) elimination of
quantifiers if for every ϕ( x ) ∈ L there is a quantifier-free formula ψ( x ) ∈ Lqf such
that
T ` ψ ( x ) ↔ ϕ ( x ).
We will discuss general criteria for elimination of quantifiers in Chapter 10, see for
instance Corollary 10.11 and 10.12. Here we report without proof the following
theorem.

7.13 Theorem The following are equivalent

1. T has elimination of quantifiers;

2. every partial isomorphism between models T is an elementary map. 

For the time being when saying that T has quantifier elimination we intend 2 of the
theorem above. For instance we rephrase Corollary 7.12 above as follows.

7.14 Corollary The theories Tdlo and Trg have elimination of quantifiers. 

In the next chapter we introduce important examples of ω-rich models that do not
have an ω-categorical theory. These are algebraic structures (groups, fields etc.)
hence more complex than pure relational structures. So, we conclude this section
with an example of this phenomenon in almost trivial context.

7.15 Example Let L contain a unary predicate rn for every positive integer n. The theory
T0 contains the axioms ¬∃ x rn ( x ) ∧ rm ( x ) for n 6= m and ∃≤n x rn ( x ) for every n.
 

Work in the the category of models of T0 and partial isomorphisms. Let T1 be the
theory that extends T0 with the axioms ∃=n x rn ( x ) for every n. Let q( x ) be the type

¬rn ( x ) : n ∈ ω . There are models of T1 that do not realize q( x ), hence T1 is not
ω-categorical. It is easy to verify that the following are equivalent.
1. N is an ω-rich model;

2. N  T1 and q( N ) is infinite.

54
 The reader may use Theorem 7.11 and Compactess Theorem for Types 4.9 to prove
that T1 is complete and has elimination of quantifiers. It is also easy to verify that
every uncountable model of T1 is rich and consequently that T1 is uncountably
categorical. 

7.16 Exercise Let T0 and M be as in Example 7.15 except that we restrict the language
to the relations r0 , . . . , rn for a fixed n. Do ω-rich models of T0 exist? Is their theory
ω-categorical? What if we add to T0 the axiom r0 ( x ) ∨ · · · ∨ rn ( x ) ? 

7.17 Exercise The language contains only the binary relations < and e. The theory T0
says that < is a strict linear order and that e is an equivalence relation. Let M
consists of models of T0 and partial isomorphisms. Do rich models exist? Can we
axiomatize their theory? Is it ω-categorical? 

7.18 Exercise The language contains only two binary relations. The theory T0 says that
they are equivalence relations. Let M consists of models of T0 and partial isomor-
phisms. Do rich models exist? Can we axiomatize their theory? Is it ω-categorical? 

7.19 Exercise In the language of graphs let T0 say that there are no cycles (equivalently,
there is at most one path between any two nodes). In combinatorics these graphs
are called forests, and their connected components are called trees. Let M consists of
models of T0 and partial isomorphisms. Do rich models exist? Can we axiomatize
their theory? Is it ω-categorical? 

2 Weaker notions of universality and homogeneity

We want to extend the equivalence in Theorem 7.8 to λ-rich models. For that we
need to weaken the notions of λ-universality and λ-homogeneity. This section is
more technical and could be skipped at a first reading.

7.20 Definition We say that a structure N is weakly λ-homogeneous if for every b ∈ N

every morphism k : N → N of cardinality < λ extends to one defined in b. The term
back-and-forth λ-homogeneous is also used. 

The following easy exercise on back-and-forth is required in the sequel.

7.21 Exercise (Assume 7.4) Prove that any weakly λ-homogeneous structure of cardi-
nality λ is homogeneous. 

7.22 Definition We say that a structure N is weakly λ-universal if for every model M in
the connected component of N and every A ⊆ M of cardinality < λ there is a morphism
k : M → N such that A ⊆ dom k. 

7.23 Lemma (Assume 7.4) Let N be a weakly λ-homogeneous model. Let A ⊆ N have cardi-
nality ≤ λ and let k : N → N be a morphism of cardinality < λ. Then there is a model
∼
M  N containing A and an automorphism h : M → M that extends k.

55
 Proof Similar to the proof of Theorem 7.11. We shall construct simultaneously a
chain h Ai : i < λi of subsets of N and a chain functions hhi : i < λi, such that
hi : N → N are morphisms. In the end we will set
[ [
M = Ai and h = hi
i <λ i <λ
The chains start with A0 = A ∪ dom k ∪ img k and h0 = k. As usual, at limit
stages we take the union. Now we consider successor stages. At stage i we fix
some enumerations of Ai and of L x ( Ai ), where | x | = 1. Let hi1 , i2 i be the i-th pair
of ordinals < λ. If the i2 -th formula in L x ( Ai1 ) is consistent in N, let a be any
of its solutions. Also let b be the i2 -th element of Ai1 . Let hi+1 : N → N be a
minimal morphism that extends hi and is such that b ∈ dom hi+1 ∩ img hi+1 . Define
Ai+1 = Ai ∪ a, hi+1 b, hi−+11 b .



7.24 Theorem (Assume 7.4) For every model N the following are equivalent
1. N is λ-rich;

2. N is weakly λ-universal and weakly λ-homogeneous.

Proof Implication 1⇒2 is clear. To prove 2⇒1 we generalize the proof of Theo-
rem 7.8. We assume 2 fix some morphism k : M → N of cardinality < λ and let
b ∈ M. By weak λ-universality there is a morphism f : M → N with domain of
definition dom k ∪ b . The map f ◦ k−1 : N → N has cardinality < λ and, by

∼
Lemma 7.23, it has an extension to an automorphism h : N 0 → N 0 for some N 0  N
containing img k ∪ img f . Then h ◦ f : M → N extends k and is defined on b. 

3 The amalgamation property

In this section we discuss conditions that ensure the existence of rich models.
We say that M has the amalgamation property if for every pair of morphisms
f 1 : M → M1 and f 2 : M → M2 there are two of embeddings g1 : M1 ,→ N and
g2 : M2 ,→ N such that g1 ◦ f 1 ( a) = g2 ◦ f 2 ( a) for every a in the common domain
of definition, dom( f 1 ) ∩ dom( f 2 ).

f1
M1 g1

M N
f1 g2
M2

As we assume that morphisms are invertible, we may express the amalgamation

property in a more concise form. Namely, for every morphism k : M → N there is
morphism g : M0 → N 0 such that the following diagram commutes

k
N id N

M g N0

It is convenient to use the following terminology. We write M ≤ N for M ⊆ N and

id M : M ,→ N is a morphism. We say that k : M → N extends to g : M0 → N 0 if
k ⊆ h, M ≤ M0 , and N ≤ N 0 .

56
7.25 Proposition Assume c3, then the following are equivalent
1. M has the amalgamation property;

2. every morphism k : M → N extends to an embedding g : M ,→ N 0 .

Proof 1⇒2 Given k : M → N, the amalgamation property yields the following

commutative diagram which can be simplified to the diagram at the right
N g1
k N g1
M N0 k

id M g2
M g2 N0
M

Up to isomorphism we can assume g2 = id M , i.e. that N ≤ N 0 . Hence g2 : M ,→ N 0

is the required extension of k : M → N.
2⇒1 Let f 1 : M → M1 and f 2 : M → M2 be given. Let k = f 2 ◦ f 1−1 : M1 → M2 and
let g : M1 ,→ N 0 be the extension ensured by 2. Then we obtain
M1
f1 g

M f 2 ◦ f 1−1 N0

f2 id M2
M2

as required. 

We say that h Mi : i < λi is a ≤-chain if Mi ≤ M j for all i < j < λ. For the next
theorem to hold we need the following property:

7.26 Definition We say that M is closed under union of ≤-chains if

[
c7. if h Mi : i < λi is a ≤-chain, then Mi ≤ M j for all i < λ. 
j<λ

The following is a general existence theorem for rich models. This general form
requires large cardinalities. We leave to the reader to verify that if the number if
finite morphisms is countable (up to isomorphism) then countable rich models exit.

7.27 Theorem Assume 7.4. Assume further c7 and that M has the amalgamation property. Let
λ be such that | L| < λ = λ<λ . Then there is a rich model N of cardinality λ.
Proof We construct N as union of a ≤-chain of models h Ni : i < λi such that
| Ni | = λ. Let N0 be any model of cardinality λ. At stage i + 1, let f : M → Ni be the
least morphism (in a well-ordering that we specify below) such that | f | ≤ | M| < λ
and f has no extension to an embedding f 0 : M ,→ Ni . Apply the amalgamation
property to obtain an total morphism f 0 : M ,→ N 0 that extends f : M → Ni . By the
downword Löwenheim-Skolem Theorem we may assume | N 0 | = λ. Let Ni+1 = N 0 .
At limit stages take the union.
The well-ordering mentioned needs to be chosen so that in the end we forget no-
body. So, first at each stage we well-order the isomorphism-type of the morphisms
f : M → Ni such that f ≤ | M | < λ. Then the required well-ordering is obtained by

57
 dovetailing all these well-orderings. The length of this enumeration is at most λ<λ ,
which is λ by hypothesis.
We check that N is rich. Let f : M → N be a morphism and | f | < | M| ≤ λ. As
| L| < λ we can approximate M with an elementary chain of structures of cardinality
< λ. Hence we may as well assume that | f | ≤ | M| < λ. The cofinality of λ is larger
than | f |, hence img f ⊆ Ni for some i < λ. So f : M → Ni is a morphism and at
some stage j we have ensured the existence of an embedding of f 0 : M ,→ Nj+1 that
extends f . 

7.28 Proposition Let M consist of all structures of some fixed signature and the elementary
maps between these. Then M has the amalgamation property.
Proof Let k : M → N be an elementary map. Let a enumerate dom k and let b
enumerate M. Let p( x ; z) = tp M (b ; a). The type p( x ; a) is consisent in M, in
particular, it is finitely consistent and, by elementarity, p( x ; ka) is finitely consistent
in N. By the compactness theorem, there is N 0  N such that N 0  p(c ; ka) for
some c ∈ N 0| x| . Hence g = {hb, ci} : M → N 0 is the required elementary map that
extends k : M → N. 

58
 Chapter 8
Some algebraic structures
The main result in this chapter is Corollary 8.24, the elimination of quantifiers in
algebraically closed fields, which in algebra is called Chevalley’s Theorem on con-
structible sets. From it we derive Hilbert’s Nullstellensatz 8.24. Finally, we isolate
the model theoretic properties of those types that correspond to the algebraic no-
tions of prime and radical ideal of polynomials.
The first sections of this chapter are not a pre-requisite for Sections 4 – 7, at the cost
of a few repetitions in the latter.

8.1 Notation Recall that when A ⊆ M we denote by h Ai M the substructure of M

generated by A. Then h Ai M ⊆ N is equivalent to N  Diag h Ai M . The diagram of
a structure has been defined at the end of Section 5.3.
In this chapter, whenever some A ⊆ M  T are fixed, by model we mean super-
structures of h Ai M that models T. The notions of logical consequence, consistency,
completeness, etc. are modified accordingly, and we write ` for T ∪ Diagh Ai M ` .
We say that a type p( x ) is trivial if ` p( x ). 

1 Abelian groups
In this section the language L is that of additive groups. The theory Tag of
abelian groups is axiomatized by the universal closure of the following axioms
a1 ( x + y) + z = y + ( x + z);

a2 x + (− x ) = (− x ) + x = 0;

a3 x + 0 = 0 + x = x;

a4 x + y = y + x.

Let x be a tuple of variables of length α, an ordinal. We write Lter,x for the set
of terms t( x ) with free variables among x. On this set we define the equivalence
relation
t( x ) ∼ s( x ) = Tag ` t( x ) = s( x ).
We define the group operations on Lter,x /∼ in the obvious way. We denote by Z⊕α
the set of tuples of integers of length α that are almost always 0. The group oper-
ations on Z⊕α are defined coordinate-wise. The following immediate proposition
implies in particular that Lter,x /∼ is isomorphic to Z⊕α .

8.2 Proposition Let A ⊆ M  Tag . Then for every formula ϕ( x ) ∈ Lat( A) there are n ∈ Z⊕α
and c ∈ h Ai M such that
` ϕ( x ) ↔ ∑ ni xi = c.
i <α

59
 where n = hni : i < αi and x = h xi : i < αi. 

Proof Up to equivalence over Tag the formula ϕ( x ) has the form s( x ) = t( a) for
some parameter-free terms s( x ) and t(z). Over Diag h Ai M , we can replace t( a) with
a single c ∈ h Ai M and write s( x ) as the linear combination shown above. 

8.3 Definition Let M  Tag . For A ⊆ M and c ∈ M. We say that c is independent from A

if h Ai M ∩ hci M = 0 . Otherwise we say that c is dependent from A . The rank of M is
the least cardinality of a subset A ⊆ M such that all elements in M are dependent from A.
We denote it by rank( M) . 

Note that when M is a vector space the condition h Ai M ∩ hci M = 0 is equivalent
to saying that c is not a linear combination of vectors in A. Then rank M coincides
with the dimension of M. In fact, what we do here for abelian groups could be
easily generalized to D-modules, where D is any integral domain, and in particular
to vector spaces.
The following proposition gives a convenient syntactic characterization of indepen-
dence. An element c of a group is called nilpotent , or in abelian groups also a
torsion element , if cn = 0 for some n.

8.4 Proposition Let A ⊆ M  Tag . Let c ∈ M be not a torsion element. Then the following
are equivalent.

1. h Ai M ∩ hci M = 0 ;

2. p( x ) = at-tp M (c/A) is trivial (see Notation 5.20 and 8.1).

Proof 1⇒2 By Proposition 8.2, a non trivial formula in p( x ) may be assumed to

have the form nx = a for some n and some a. If such a formula is satisfied by c then
a ∈ h Ai M ∩ hci M . As c is not a torsion element, n = 0 and the equation is trivial.

2⇒1 If h Ai M ∩ hci M 6= 0 then nc = a for some a ∈ h Ai M r {0}. Then c satisfies
some non trivial equation nx = a. 

8.5 Remark Let k : M → N be a partial embedding and let a be an enumeration of


dom k. We claim that k ∪ hb, ci : M → N is a partial embedding for every b ∈ M
and c ∈ N that are independent from a, respectively ka. In fact, it suffices to check
that M, b, a ≡at N, c, ka. Suppose ϕ( x ; z) ∈ Lat is such that M  ϕ(b ; a). Then by
independence ϕ( x ; a) is trivial, i.e.
Tag ∪ Diagh ai M ` ϕ ( x ; a ).
As h ai M and hkai N are isomorphic structures
Tag ∪ Diaghkai N ` ϕ( x ; ka).
Therefore N  ϕ(c ; ka). The proof that N  ϕ(c ; ka) implies M  ϕ(b ; a) is similar. 

2 Torsion-free abelian groups

The theory of torsion-free abelian groups extends Tag with the following axioms
for all positive integers n

60
 st nx = 0 → x = 0.

We denote this theory by Ttfag . It is not difficult to see that in a torsion-free abelian
group every equation of the form nx = a has at most one solution.

8.6 Proposition Let M  Ttfag be uncountable. Then rank M = | M|.

Proof Let A ⊆ M have cardinality < | M |. We claim that M contains some element
that is independent from A. It suffices to show that the number of elements that are
dependent from A is < | M |. If c ∈ M is dependent from A then, by Proposition 8.7,
it is a solution of some formula Lat( A). As there is no torsion, such a formula has
at most one solution. Therefore the number of elements that are dependent from A

is at most | Lat( A)|, that is max | A|, ω . If M is uncountable the claim follows. 

8.7 Proposition Let A ⊆ M  Ttfag . Let p( x ) = at± -tp(b/A), where b ∈ M. Then one of
the following holds:
1. b is independent from A;

2. M  ϕ(b) for some ϕ( x ) ∈ Lat( A) such that ` ϕ( x ) → p( x ).

Note the similarity with Example 7.15, where the independent type is q( x ) and the
isolating formulas are the ri ( x ).
It is important to observe that the set A above may be infinite. This is essential to
obtain Corollary 8.11, and it is one of the main differences between this example
and the examples encountered in Chapter 6.
Proof If b is dependent from A, then b satisfies a non trivial atomic formula ϕ( x )
which we claim is the formula required in 2. It suffices to show that ϕ( x ) implies
a complete Lat±( A)-type. Clearly this type must be p( x ). Let ϕ( x ) have the form
nx = a for some n ∈ Z r {0} and a ∈ h Ai M r {0}. We show that for every m ∈ Z
and every c ∈ h Ai M one of the following holds
a. ` nx = a → mx = c;

b. ` nx = a → mx 6= c.

Suppose not for a contradiction that neither a nor b holds and fix models N1 , N2
and some bi ∈ Ni such that
a’. N1  nb1 = a and N1  mb1 6= c;

b’. N2  nb2 = a and N2  mb2 = c.

From b’ we obtain N2  ma = nc. As N1 is torsion-free, from a’ we obtain N1 

ma 6= nc. But ma = nc is a formula with parameters in h Ai M , so it should have the
same truth value in all superstructures of h Ai M , a contradiction. 

3 Divisible abelian groups

The theory of divisible abelian groups extends Ttfag with the following axioms for
all integers n 6= 0
div y 6= 0 → ∃ x nx = y.

We denote this theory by Tdag .

61
 8.8 Proposition Let A ⊆ M  Ttfag and let ϕ( x ) ∈ Lat( A), where | x | = 1, be consistent.
Then N  ∃ x ϕ( x ) for every model N  Tdag .

Note that in the proposition above consistent means satisfied in some M0 such that
h Ai M ⊆ M0  Ttfag .
The claim in the proposition holds more generally for all ϕ( x ) ∈ Lqf and also when
x is a tuple of variables. This follows from Lemma 8.10, whose proof uses the
proposition.
Proof We can assume that ϕ( x ) has the form nx = a for some n ∈ Z and some
a ∈ h Ai M . If n = 0, then a = 0 since ϕ( x ) is consistent, and the claim is trivial. If
n 6= 0 then by consistency a 6= 0, hence a solution exist in N by axiom div. 

8.9 Exercise Prove a converse of Proposition 8.8. Let A ⊆ N  Tag and let x be a
single variable. Prove that if N  ∃ x ϕ( x ) for every consistent ϕ( x ) ∈ Lat( A), then
N  Tdag . 

We are ready to prove that divisible abelian groups of infinite rank are ω-rich.

8.10 Lemma Let k : M → N be a partial isomorphism of cardinality < λ, where M  Ttfag

and N  Tdag is a model of rank ≥ λ. Then for every b ∈ M there is c ∈ N such that

k ∪ hb ; ci : M → N is a partial isomorphism.
Proof Let a be an enumeration of dom k and let p( x ; z) = at± -tp(b ; a). The required
c has to realize p( x ; ka). We consider two cases. If b is dependent from a, then
Proposition 8.7 yields a formula ϕ( x ; z) ∈ Lat such that
i. ` ϕ( x ; a) → p( x ; a)
ii. ϕ( x ; a) is consistent.
By isomorphism, i and ii hold with a replaced by ka. Then by Proposition 8.8 the
formula ϕ( x ; ka) has a solution c ∈ N.
The second case, which has no analogue in Lemma 6.1, is when b is independent
from a. Then by Remark 8.5 we may choose c to be any element of N independent
from ka. Such an element exists because N has rank at least λ. 

8.11 Corollary Every uncountable model of Tdag is rich in the category of models of Ttfag and
partial isomorphisms. In particular Tdag is uncountably categorical, complete, and has
quantifier elimination.
Proof Any uncountable N  Tdag has rank | N |, therefore it is rich by Lemma 8.10;
categoricity and completeness follow. As for quantifier elimination, let k : M → N
is a partial isomorphism between models of Tdag . If M  M0 and M  N 0 are ele-
mentary superstructures of uncountable cardinality then k : M0 → N 0 is elementary
by Theorem 7.11 and this suffices to conclude that k : M → N is elementary. 

8.12 Exercise Prove that every model of Tdag is ω-ultraomogeneous (indipendenlty of

cardinality and rank). 

62
 4 Commutative rings
In this section L is the language of (unital) rings . It contains two constants 0 and
1 the unary operation − and two binary operations + and ·. The theory of rings
contains the following axioms
a1-a4 as for abelian groups

r1 ( x ·y)·z = y·( x ·z),

r2 1· x = x ·1 = x,

r3 ( x + y)·z = x ·z + y·z,

r4 z·( x + y) = z· x + z·y.

All the rings we consider are commutative

c x ·y = y· x.
We denote the theory of commutative rings by Tcr .
In what follows the theory Tcr ∪ Diagh Ai M is implicit in the sense of Notation 8.1,
so it is important to remember that Diagh Ai M is not trivial even when A = ∅. In
fact, Diagh∅i M determines the characteristic of the models.
Let A ⊆ M  Tcr and let x be a tuple of variables of length α, an ordinal. We write
Lter,x ( A) for the set of terms t( x ) with free variables among x and parameters in A.
On this set we define the equivalence relation
t( x ) ∼ s( x ) = ` t ( x ) = s ( x ).
On Lter,x ( A)/∼ we define the ring operations in the obvious way so that Lter,x ( A)/∼
is a commutative ring. We denote by A[ x ] the set of polynomials with variables
among x and parameters in h Ai M . The ring operations on A[ x ] are defined as
usual. The following proposition (which is clear, but tedious to prove) implies in
particular that Lter,x ( A)/∼ is isomorphic to A[ x ]. For simplicity we state it only for
| x | = 1.

8.13 Proposition Let A ⊆ M  Tcr and let x be a single variable. Then for every formula
ϕ( x ) ∈ Lat( A) there is a unique n < ω and a unique tuple h ai : i ≤ ni of elements of
h Ai M such that an 6= 0 and
` ϕ( x ) ↔ ∑ ai x i = 0. 
i ≤n

The integer n in the proposition above is called the degree of ϕ( x ) .

8.14 Definition Let A ⊆ M  Tcr . We say that an element b ∈ M is transcendent over A if

the type p( x ) = at-tp M (b/A) is trivial (see Notation 5.20 and 8.1). Otherwise we say that
b is algebraic over A . The transcendence degree of M is the least cardinality of a subset
A ⊆ M such that all elements of M are algebraic over A. 

8.15 Remark Remark 8.5 holds here with ‘independent’ replaced by ‘transcendent’ and
Tag replaced by Tcr . 

63
5 Integral domains
Let a ∈ M  T. We say that a is a zero divisor if a b = 0 for some b ∈ M r {0}.
An integral domain is a commutative ring without zero divisors. The theorey of
integral domains contains the axioms of commutative rings and the following
nt. 0 6= 1

id. x ·y = 0 → x = 0 ∨ y = 0.

We denote the theory of integral domains by Tid .

8.16 Proposition Let M  Tid be uncountable. Then M has transcendence degree | M|.
Proof In an integral domain every polynomial has finitely many solutions and there
are | L( A)| polynomials over A. 

8.17 Proposition Let A ⊆ M  Tid . For b ∈ M let p( x ) = at± -tp(b/A). Then one of the
following holds
1. b is transcendental over A;

2. M  ϕ(b) for some ϕ( x ) ∈ Lat( A) such that ` ϕ( x ) → p( x ).

Note the similarity with Example 7.15, where the transcendental type is q( x ) and
the isolating formulas are the ri ( x ).
As in Proposition 8.7, the set A may be infinite. This is essential to obtain Corol-
lary 8.20.
Proof Suppose b is not transcendental, i.e. it satisfies a non trivial atomic formula.
Let ϕ( x ) ∈ Lat( A) be a non trivial formula with minimal degree such that ϕ(b). We
prove that ϕ( x ) implies a complete Lat±( A)-type. Clearly this type must be p( x ).
We prove that for any ξ ( x ) ∈ Lat( A) one of the following holds
1. ` ϕ( x ) → ξ ( x )
2. ` ϕ ( x ) → ¬ ξ ( x ).
Let us write a( x ) = 0 and a0 ( x ) = 0 for the formulas ϕ( x ) and ξ ( x ), respectively.
Let n be the degree of the polynomial a( x ). If h Ai M is a field, choose a polynomial
d( x ) of maximal degree such that
a. d( x )t( x ) = a( x )

a’. d( x )t0 ( x ) = a0 ( x ),

for some polynomials t( x ) and t0 ( x ).

If h Ai M is not a field, polynomials d( x ), t( x ) and t0 ( x ) as above exist with coeffi-
cients in the field of fractions of h Ai M . Then a and a’ hold up to a factor in h Ai M
which we absorb in a( x ) and a0 ( x ).
From a we get d(b) = 0 or t(b) = 0. In the first case, as a( x ) has minimal degree,
we conclude that t( x ) is constant. This implies that any zero of a( x ) is also a zero
of a0 ( x ), that is, it implies 1.
Now suppose t(b) = 0. Then the minimality of the degree of a( x ) implies that
d( x ) = d, where d is a nonzero constant. If h Ai M is a field, apply Bézout’s identity

64
 to obtain two polynomials c( x ) and c0 ( x ) such that d = a( x )c( x ) + a0 ( x )c0 ( x ). Then
a( x ) and a0 ( x ) have no common zeros, and 2 follows. If h Ai M is not a field, we use
Bézout’s identity in the field of fractions of h Ai M and, for some d0 ∈ h Ai M r 0 ,


obtain d0 d = a( x )c( x ) + a0 ( x )c0 ( x ). Then we reach the same conclusion. 

6 Algebraically closed fields

Let a, b ∈ M  Tcr . We say that b is the inverse of a if a·b = 1. A field is a com-
mutative ring where every non-zero element has an inverse. The theory of fields
contains Tid and the axiom
 
f. ∃y x 6= 0 → x ·y = 1 .

Fields are structures in the signature of rings: the language contains no symbol for
the multiplicative inverse. So, substructures of fields are merely integral domains.
The theory of algebraically closed field , which we denote by Tacf , also contains
the following axioms for every positive integer n
ac. ∃ x x n + zn−1 x n−1 + · · · + z1 x + z0 = 0

This not a complete theory because it does not decide the characteristic of the field.
p
For a prime p, we define the theory Tacf , which contains Tacf and the axiom
ch p . 1 + . . . (p times) · · · + 1 = 0.
0 contains the negation of ch for all p. Note that all models of T p
The theory Tacf p acf
have the same characteristic in the model theoretic sense defined in 5.27. In fact,
any two fields M and N with the same (algebraic) characteristic have isomorphic
prime subfield, that is, h∅i M ' h∅i N .

8.18 Proposition Let A ⊆ M  Tid and let ϕ( x ) ∈ Lat( A), where | x | = 1, be consistent.
Then N  ∃ x ϕ( x ) for every model N  Tacf .

Note that in the proposition above consistent means satisfied in some M0 such that
h Ai M ⊆ M0  Tid .
The claim in the proposition holds more generally for all ϕ( x ) ∈ Lqf when x is a
tuple of variables. This follows from Lemma 8.19 whose proof uses the proposition.
Proof Up to equivalence ϕ( x ) has the form an x n + · · · + a1 x + a0 = 0 for some
ai ∈ h Ai N . Choose n minimal. If n = 0 then a0 = 0 by the consistency of ϕ( x ) and
the claim is trivial. Otherwise an 6= 0 and the claim follows from f and ac. 

8.19 Lemma Let k : M → N be a partial isomorphism of cardinality < λ, where M  Tid and
N  Tacf has transcendence degree ≥ λ. Then for every b ∈ M there is c ∈ N such that

k ∪ hb, ci : M → N is a partial isomorphism.

The following is the proof of Lemma 8.10 which we have pasted here for the sake
of the lazy reader.
Proof Let a be an enumeration of dom k and let p( x ; z) = at± -tp M (b ; a). The
required c has to realize p( x ; ka). We consider two cases. If b is algebraic over a,
then Proposition 8.17 yields a formula ϕ( x ; z) ∈ Lat such that

65
 i. ϕ( x ; a) → p( x ; a)
ii. ϕ( x ; a) is consistent.
By isomorphism i and ii hold with a replaced by ka. Then by Proposition 8.18 the
formula ϕ( x ; ka) has a solution in c ∈ N.
The second case, which has no analogue in Lemma 6.1, is when b is transcendent
over a. Then by Remark 8.15 we may choose c to be any element of N transcendent
over ka. This exists because N has transcendence degree ≥ λ. 

p
8.20 Corollary Every uncountable model of Tacf is rich in the category of integral domains of
p
characteristic p and partial isomorphisms. In particular Tacf is uncountably categorical, and
it is complete. Moreover, Tacf has quantifier elimination.

The proof requires the same argument as Corollary 8.11.

p
Proof Every uncountable N  Tacf has transcendence degree | N |, therefore it is rich
by Lemma 8.19. Categoricity and completeness follow.
For quantifier elimination, let k : M → N be a partial embedding between models of
p
Tacf . Then M and N have the same charactiristic, i.e. they are model of Tacf . Let M0
and N 0 be elementary superstructures of M and N respectively of sufficiently large
cardinality. As M0 and N 0 are rich, k : M0 → N 0 is elementary by Theorem 7.11.
Hence k : M → N is also elementary. 

8.21 Exercise Prove that every model of Tacf is ω-ultrahomogeneous (indipendently of

cardinality and transcendence degree). 

7 Hilbert’s Nullstellensatz
In this section we fix a tuple of variables x and a subset A of an integral domain.
We denote by ∆( A) the set of formulas of the form t( x ) = 0 where t( x ) is a term
with parameters in A. So ∆( A)-types are (possibly infinite) systems of polynomial
equations with coefficients in h Ai M .
For convenience we define the closure of p( x ) under logical consequences as follows
n o
ccl p( x ) = t( x ) = 0 : p( x ) ` t( x )=0
Remember that we always work under the assumptions made in Notation 8.1. In
particular, in this section we work over the theory Tid ∪ Diagh Ai M . In general,
closure under logical consequences is an elusive notion. Hence Propositions 8.22
and 8.23 are useful because they give a model theoretical, respectively algebraic,
characterisation of ccl p( x ).

8.22 Proposition Let A ⊆ M  Tid and let p( x ) be a ∆( A)-type. Fix some N of sufficiently
large cardinality such that h Ai M ⊆ N  Tacf . Then
n  o
ccl p( x ) = t( x ) = 0 : N  ∀ x p( x ) → t( x )=0 .

It suffices to choose N such that | A| < | N | and | x | ≤ | N |; note that x has pos-
sibly infinite length. In Corollary 8.24 below we will considerably strengthen this
proposition for finite x.

66
 Proof Only the inclusion ⊇ requires a proof. Fix some polynomial t( x )=0 ∈ /
ccl p( x ). Then p( x ) ∧ t( x ) 6= 0 is consistent, so there is a model M0 such that
|x|
M0  p( a) ∧ t( a) 6= 0 for some a ∈ M0 . Then there is a partial isomorphism
h : M0 → N that extends id A and is defined on a, provided N is large enough to
accommodate a. This implies that p( x ) ∧ t( x ) 6= 0 has a solution in N. Hence t( x )
does not belong to the set on the r.h.s. 

Recall that A[ x ] denotes the ring of polynomials with variables in x and coefficients
in h Ai M . We identify ∆( A) nd A[ x ] in the obvious way. Consequently, a ∆( A)-type
p( x ) is identified with a set of polynomials p ⊆ A[ x ]. For p ⊆ A[ x ] we write rad p
for the radical ideal generated by p , that is, the intersection of all prime ideals
containing p. When p = rad p, we say that p is a radical ideal . Recall from algebra
that if p ⊆ A[ x ] is an ideal then
n o
# rad p = t( x ) : tn ( x ) ∈ p for some positive integer n .

An identity that justifies the name.

8.23 Proposition Let A ⊆ M  Tid and let p( x ) be a ∆( A)-type. Then ccl p( x ) ' rad p.

The proposition holds with a similar proof for the broader class of rings without
nilpotent elements. (Which does not come as a surprise.)
Proof (⊇) We claim that ccl p( x ) is an ideal. In fact, for every pair of L( A)-terms
t( x ) and s( x )
t( x ) = 0 ` s( x )t( x ) = 0
s( x ) = t( x ) = 0 ` s( x ) + t( x ) = 0
Moreover, as integral domains do not have nilpotent elements
tn ( x ) = 0 ` t( x ) = 0
By # above, ccl p( x ) is a radical ideal which proves ccl p( x ) ⊇ rad p.
(⊆) We fix some t( x ) ∈ / rad p and prove that p( x ) ∧ t( x ) 6= 0 is consistent. Let q
be some prime ideal containing p such that t( x ) ∈
/ q. As q is prime, the ring A[ x ]/q
is an integral domain. The polynomials that vanish in A[ x ]/q at x + q are exactly
those in q. Hence A[ x ]/q witnesses the consistency of q( x ) ∧ t( x ) 6= 0. 

When x is a finite tuple of variables, we can extend the validity of Proposition 8.22
to the case A = M = N.

8.24 Corollary (Hilbert’s Nullstellensatz) Let N  Tacf and let p( x ), where | x | < ω, be a
∆( N )-type. Then
n o
rad p ' ccl p( x ) = t ( x ) : N  ∀ x [ p ( x ) → t ( x ) = 0] .

Proof Let N 0 be a large elementary extension of N. By Proposition 8.22, the claim

holds for N 0 . By Hilbert’s Basis Theorem, the ideal generated by p is finitely gen-
erated hence p( x ) is equivalent to a formula (cfr. Exercise 8.27). Therefore, by ele-
mentarity, the claim holds for N. 

Hilbert’s Nullstellensatz comes in two variants. The one in Corollary 8.24 is some-
times referred to as the strong Nullstellensatz. The weaker variant is stated in Exer-
cise 8.26.

67
 We conclude this section by showing that the notions of primeness for types and
ideals coincide (if we restrict to types closed under logical consequences).

8.25 Proposition Let A ⊆ M  Tid and let p( x ) be a ∆( A)-type closed under logical conse-
quences. Then the following are equivalent
1. p( x ) is a prime ∆( A)-type;

2. p is a prime ideal.

Proof 1⇒2 Assume 1 and suppose that the polynomial t( x ) · s( x ) belongs to p.

Clearly, over Tid ∪ Diagh Ai M we have ` t( x ) · s( x ) = 0 → t( x ) = 0 ∨ s( x ) = 0. As
p( x ) is a prime ∆( A)-type, p( x ) ` t( x ) = 0 or p( x ) ` s( x ) = 0. By Proposition 8.23,
the type p( x ) is closed under logical consequences, therefore t( x ) ∈ p or s( x ) ∈ p.
2⇒1 Assume p is a prime ideal and for some ti ( x ) = 0 ∈ ∆( A)
n
_
p( x ) ` ti ( x ) = 0.
i =1
Then
n
p( x ) ` ∏ ti ( x ) = 0
i =1
Since p( x ) is closed under logical consequences, and p is a prime ideal, ti ( x ) ∈ p for
some i. Hence p( x ) contains the equation ti ( x ) = 0. By Corollary 5.19 this suffices
to prove that p( x ) is a prime ∆( A)-type. 

8.26 Exercise Let N  Tacf and let p( x ) be a ∆( N )-type where | x | < ω. Prove that the
following are equivalent
1. p is a proper ideal;

2. p( x ) has a solution in N. 

8.27 Exercise Let A ⊆ M  Tid and let p( x ) be a ∆( A)-type. Prove that the following
are equivalent
1. p( x ) is a principal ∆( A)-type;

2. the ideal generated by p is finitely generated. 

68
 Chapter 9
Saturation and homogeneity
The first two section introduce saturation and homogeneity and Section 3 presents
the notation we shall use in the following chapters when working inside a monster
model.

1 Saturated structures
Recall that a type p( x ) ⊆ L( M) is finitely consistent in M if every ϕ( x ), conjunc-
tion of formulas in p( x ), has a solution in M. When A ⊆ M we write Sx ( A) for the
set of types that are complete and finitely consistent in M. We never display M in
the notation as it will always be clear from the context. When A is empty it usual
to write Sx ( T ) for Sx ( A) where T = Th( M ). We write S( A) for the union of Sx ( A)
as x ranges over all tuples of variables. Similarly for S( T ).
The following remark will be used in the sequel without explicit reference.

9.1 Remark Let k : M → N be an elementary map and let a be an enumeration of

dom k. If p( x ; a) is finitely consistent in M, then p( x ; k a) is finitely consistent in N.
(We can drop finitely in the antecedent but not in the consequent.) 

9.2 Definition Let x be a single variable and let λ be an infinite cardinal. We say that a
structure N is λ-saturated if it realizes every type p( x ) such that
1. p( x ) ⊆ L( A) for some A ⊆ N of cardinality < λ;

2. p( x ) is finitely consistent in N.

We say that N is saturated if it is λ-saturated and | N | = λ.

9.3 Exercise Suppose | L| ≤ ω and let M be an infinite structure. Then for every
non-principal ultrafilter F on ω the structure Mω /F is a ω1 -saturated elementary
superstructure of M.
Hint: the notation is as in Chapter 3. Let | x | = 1 and |z| = ω. It suffices to consider
the types form p( x ; ĉ) where p( x ; z) = ϕi ( x ; z) : i < ω ⊆ L and ĉ ∈ ( Mω )|z| .


Without loss of generality we can also assume that ϕi+1 ( x ; z) → ϕi ( x ; z), and that
all formulas ϕi ( x ; ĉ) are consistent in Mω .
Let h Xi : i < ω i be a strictly decreasing chain of elements of the ultrafilter such that

Xi+1 ⊆ j : M  ∃ x ϕi ( x, ĉj) . Let â ∈ Mω be such that ϕi ( âj, ĉj) holds for every
j ∈ Xi r Xi+1 . Then â realizes p( x ). 

9.4 Theorem Assume | L| < λ where λ is be such that λ<λ = λ. Then every structure M of
cardinality λ has a saturated elementary extension of cardinality λ.
Proof We construct an elementary chain h Mi : i < λi of models of cardinality λ.
The chain starts with M and is the union at limit stages. Given Mi we choose as

69
 Mi+1 any model of cardinality λ that realizes all types in Sx ( A) for all A ⊆ Mi of
cardinality < λ. The required Mi+1 exists because there are 2| L( A)| ≤ λ types in
Sx ( A) and there are λ<λ = λ sets A.
Let N be the union of the chain. We check that N is the required extension. Let
p( x ) ∈ S( A) for some A ⊆ N of cardinality < λ. As λ is a regular cardinal, A ⊆ Mi
for some i < λ. Then Mi+1 realizes p( x ), and so does N, by elementarity. 

9.5 Theorem Assume | L| ≤ λ and let N be an infinite structure. Let M be the category (see
Section 7.1) that consists of models of a compete theory T and elementary maps between
these. Then the following are equivalent
1 N is a λ-saturated structure;
2 N is a λ-rich model;
3 N realizes all types p(z) ⊆ L( A), with |z| ≤ λ and | A| < λ, finitely consistent in N.

Note that it is the completeness of T which makes the category M connected.

Proof 1⇒2. Let k : M → N be an elementary map of cardinality < λ. It suffices

to show that for every b ∈ M there is a c ∈ N such that k ∪ hb, ci : M → N is an
elementary map. Let a be an enumeration of dom k and define p( x ; z) = tp M (b ; a).
The required c is any element of N such that N  p(c ; k a). As p( x ; a) is finitely
consistent in M then p( x ; k a) is finitely consistent in N. As | a| < λ, saturation
yields the required c.
2⇒3. Let p(z) be as in 3. By the compactness theorem N  K  p( a) for some
model K and a ∈ K |z| . By the downward Löwenheim-Skolem theorem there is a
model A, a ⊆ M  K of cardinality ≤ λ. (Here we use | L|, | A|, |z| ≤ λ.) By 2, there
is an elementary embedding h : M ,→ N that extends id A . Finally, as M  p( a),
elementarity yields N  p(h a).
3⇒1. Trivial. 

Two saturated structures of the same cardinality are isomorphic as soon as they are
elementarily equivalent (i.e. as soon as ∅ : M → N is an elementary map.) In fact,
from Theorem 9.5 and 7.6 we obtain the following.

9.6 Corollary Every elementary map k : M → N of cardinality < λ between saturated

models of the same cardinality λ extends to an isomorphism. 

As it turns out, we already have many examples of saturated structures.

9.7 Corollary The following models are ω-saturated

1 models of Tdlo ;
2 models of Trg ;
3 models of Tdag with infinite rank;
4 models of Tacf with infinite degree of transcendence.
Countable models of Tdlo and Trg and uncountable models of Tdag or Tacf are saturated.
Proof By quantifier elimination embeddings coincide with elementary embeddings.
Then saturation is proved applying Theorem 9.5 and the extension lemmas proved

70
 in Chapter 6 and 8. 

The following is a useful test for quantifier elimination.

9.8 Theorem Assume | L| ≤ λ. Consider the category that consists of models of some theory
T0 and partial isomorphism. Suppose λ-rich models exist and denote by T1 their theory.
Then the following are equivalent
1. every λ-saturated model of T1 is λ-rich;

2. T1 has elimination of quantifiers.

Proof 2⇒1. Let N  T1 be λ-saturated. Fix a morphism k : M → N of cardinality <

λ, some b ∈ M and let p( x ; z) = qf-tp M (b ; a), where a enumerates dom k. The type
p( x ; k a) is realized in any λ-rich model N 0 that contains hk ai N . By 2, N ≡ka N 0 ,
so p( x ; k a) is finitely consistent in N. By saturation, it is realised by some c ∈ N.
Then k ∪ {hb, ci} : M → N is the required extension.
1⇒2. Let k : M → N be a finite partial isomorphism between models of T. We
claim that it is an elementary map. Let M0  M and N 0  N are saturated models
of equal cardinality. As these are rich, k : M0 → N 0 extends to an isomorphism
∼
h : M0 → N 0 and the claim follows. 

2 Homogeneous structures
Definition 7.3 introduces the notion of universal and homogeneous structures a
the general context. When the morphisms of the underlying category are the ele-
mentary maps we refer to these notion as elementary homogeneity and elementary
universality. Though one often omits to specify elementary. We repeat the definition
in this particular case.

9.9 Definition A structure N is (elementarily) λ-universal if every M ≡ N of cardinality

≤ λ there is a elementary embedding h : M ,→ N. We say universal if it is λ-universal
and of cardinality λ.
We say that N is (elementarily) λ-homogeneous if every elementary map k : N → N
of cardinality < λ extends to an automorphism. We say that N is homogeneous if it is
λ-homogeneous and of cardinality λ. 

As saturated structure are rich, the following theorem is an instance of Theorem 7.8.

9.10 Theorem For every structure N of cardinality ≥ | L| the following are equivalent
1. N is saturated;

2. N is elementarily universal and homogeneous. 

Given A ⊆ N we denote by Aut ( N/A) the group of A-automorphisms of N. That

is the group of automorphisms that fix A point-wise. Let a be a tuple of elements
of N. The orbit of a over A in N is the set
n o
O N ( a/A) = f a : f ∈ Aut ( N/A)
When the model N is clear from the context we omit the subscript.

71
 Orbits are particularly interesting if taken in a homogeneous structure. The follow-
ing proposition is immediate but its importance cannot be overestimated.

9.11 Proposition Let N be a λ-homogeneous structure. Let A ⊆ N have cardinality < λ and
let a ∈ N <λ . Then O N ( a/A) = p( N ), where p( x ) = tp N ( a/A). 

Finally, we want to extend the equivalence in Theorem 9.10 to λ-saturated struc-

tures. For this we only need to apply Theorem 7.24.
When the morphisms of the underlying category are the elementary maps, we say
weakly λ-saturated for weakly λ-universal (cfr. Definition 7.22). This is not the
standard definition (i.e. 2 of the proposition below) but the reader can easily verify
that it is equivalent by reasoning as in the proof of theorem 9.5.

9.12 Proposition The following are equivalent

1. N weakly λ-saturated;

2. N realizes every type p( x ) ⊆ L, where | x | < λ, that is finitely consistent in N. 

The following is an instance of Theorem 7.24 that the reader may prove directly as
an exercise.

9.13 Corollary Let | L| ≤ λ. The following are equivalent

1. N is λ-saturated;

2. N is weakly λ-saturated and weakly λ-homogeneous. 

9.14 Exercise Let M be an arbitrary structure. Prove that M has an ω-homogeneous

elementary extension of the same cardinality. (There is no assumption on the cardi-
nality of the language.) 

9.15 Exercise Let M and N be elementarily homogeneous structures of the same cardi-
nality λ. Suppose that M  ∃ x p( x ) ⇔ N  ∃ x p( x ) for every p( x ) ⊆ L such that
| x | < λ. Prove that the two structures are isomorphic. 

9.16 Exercise Let L be a language that extends that of strict linear orders with the con-
stants {ci : i ∈ ω }. Let T be the theory that extends Tdlo with the axioms ci < ci+1
for every i ∈ ω. Prove that T has elimination of quantifiers and is complete (it can
be deduced from what is known of Tdlo ). Exhibit a countable saturated model and
a countable model that is not homogeneous. 

3 The monster model

In this section we present some notation and terminology frequently adopted when
dealing with a complete theory T. We fix a saturated structure U of cardinality
larger than | L|. We assume U to be large enough that among its elementary sub-
structures we can find any model of T we might be interested in. For this reason
U is nicknamed the monster model . We denote by κ the cardinality of U when
necessary we assume κ to be inaccessible.

72
Some terms acquire a slightly different meaning when working inside a monster
model.

truth we say that ϕ( x ) holds if U  ∀ x ϕ( x );

small/large cardinalities smaller than κ are called small;

models are elementary substructure of U of small cardinality, they are

denoted by the letters M and N, and derived symbols;

parameters are always in U; the symbols A, B, C, etc. denote sets of pa-

rameters of small cardinality; calligraphic letters as A, B, C,
etc. are used for sets of arbitrary cardinality;

tuples have length < κ unless otherwise specified;

global types are complete types over U; the set of global types is denoted
by S(U);

formulas have parameters in U unless otherwise specified;

definable sets are sets of the form ϕ(U) for some formula ϕ( x ) ∈ L(U); we
may say A-definable if ϕ( x ) ∈ L( A);

type-definable sets are sets of the form p(U) for some p( x ) ⊆ L( A) where, as the
symbol suggests, A has small cardinality;

types of tuples we write tp( a/A) for tpU ( a/A) and a ≡ A b for U, a ≡ A U, b;

orbits of tuples under the action of Aut (U/A) are denoted by O( a/A).

Let x be a tuple of variables. For any fixed A ⊆ U we introduce a topology on

U| x| that we call the topology induced by A or, for short, A-topology . (This is non
standard terminology, not to be confused with the logic A-topology in Section 14.4.)
The closed sets of the A-topology are those of the form p(U) where p( x ) ⊆ L( A) is
any type over A.
For ϕ( x ) ∈ L( A) the sets of the form ϕ(U) are clopen in this topology (and vice
versa by Proposition 9.17). They form both a base of closed sets and base of open
sets, which makes these topologies zero-dimensional. By saturation, the topology
induced by A is compact. Actually, saturation is equivalent to the compactness of
all these topologies as A ranges over the sets of small cardinality.
These topologies are never T0 as any pair of tuples a ≡ A b have exactly the same
neighborhoods. Such pairs always exist by cardinality reasons. However it is im-
mediate that the topology induced on the quotient U| x| / ≡ A is Hausdorff (this is
the so-called Kolmogorov quotient). Indeed, this quotient corresponds to Sx ( A) with
the topology introduced in Section 5.3.
The following proposition is an immediate consequence of compactness. When
A = B it says that the topology induced by A is normal: any two closed sets are
separated by open sets. It could be called mutual normality (not a standard name)
because the two closed sets belong to different topologies and the separating sets
are each found in the corresponding topology.

73
9.17 Proposition (mutual normality) Let p( x ) ⊆ L( A) and q( x ) ⊆ L( B) be such that
p( x ) → ¬q( x ). Then there are ϕ( x ), conjuction of formulas in p( x ), and ψ( x ), conjuction
of formulas in q( x ), such that ϕ( x ) → ¬ψ( x ).
Proof The assumptions say that p( x ) ∧ q( x ) is inconsistent. Then the formulas ϕ( x )
and ψ( x ) exist by compactness (i.e. saturation). 

9.18 Remark There are many forms in which the proposition above can be applied.
For instance, assuming for brevity that p( x ) and q( x ) are closed under conjunc-
tions,
a. if p( x ) ↔ ¬q( x ) then p( x ) ↔ ϕ( x ) for some ϕ( x ) ∈ p( x );

b. if p( x ) ↔ ψ( x ) for some ψ( x ) ∈ L(U) then p( x ) ↔ ϕ( x ) for some ϕ( x ) ∈ p;

d. if p( x ) → ψ( x ) for some ψ( x ) ∈ L(U) then ϕ( x ) → ψ( x ) for some ϕ( x ) ∈ p;

n
ψ( x ), where |Ψ| < κ, then p( x ) → ψi ( x ) for some ψi ∈ Ψ.
_ _
c. if p( x ) → 
ψ∈Ψ i =1

A definable set as the form ϕ(U ; b) for some formula ϕ( x ; z) ∈ L and some b ∈ U|z| .
If f ∈ Aut (U/A) then
f ϕ (U ; b ) = a ∈ U| x |
  
f a : ϕ ( a ; b ),
f a : ϕ ( f a ; f b ), a ∈ U| x |

=
= ϕ (U ; f b ).
Hence automorphisms act on definable sets in a very natural way. Their action of
type-definable sets is similar.
We say that a set D ⊆ U| x| is invariant over A if f [D] = D for every f ∈ Aut (U/A).
Equivalently, if O( a/A) ⊆ D for every a ∈ D. By homogeneity this is equivalent to
requiring that q( x ) → x ∈ D for every q( x ) = tp( a/A) where a ∈ D.
Proposition 9.21 below an important fact about invariant type-definable sets. It may
clarify the proof to consider first the particular case of definable sets.

9.19 Proposition Let ϕ( x ) ∈ L(U) then the following are equivalent

1. ϕ( x ) is equivalent to some formula ψ( x ) ∈ L( A);

2. ϕ(U) is invariant over A.

We give two proofs of this theorem as they are both instructive.

Proof 1⇒2 Obvious.
2⇒1 From 2 and homogeneity we obtain
_
ϕ( x ) ↔ q( x )
q( x )→ ϕ( x )

where q( x ) above range over all types in Sx ( A). By compactness, we can rewrite
this equivalence
_
ϕ( x ) ↔ ϑ( x)
ϑ ( x )→ ϕ( x )

where ϑ ( x ) ranges over all formulas in L( A). This latter equivalence says that ¬ ϕ( x )

74
 is equivalent to a type over A. Again by compactness we obtain
n
_
ϕ( x ) ↔ ϑi ( x )
i =1
for some formula ϑi ( x ) ∈ L( A). 

Second proof of Proposition 9.19 2⇒1 Let ϕ(U ; b), where ϕ( x ; z) ∈ L, be a for-
mula invariant over A. Let p(z) = tp(b/A). As f [ ϕ(U ; b)] = ϕ(U ; f b) for every
f ∈ Aut (U/A), homogeneity and invariance yield
 
p(z) → ∀ x ϕ( x ; z) ↔ ϕ( x ; b) .
By compactness there is a formula ϑ (z) ∈ p such that
 
ϑ (z) → ∀ x ϕ( x ; z) ↔ ϕ( x ; b) .
Hence ϕ(U ; b) is defined by the formula ∃z ϑ (z) ∧ ϕ( x ; z) , which is a formula in
 

L( A) as required. 

9.20 Exercise Let ϕ( x ) ∈ L. Prove that the following are equivalent

1. ϕ( x ) is equivalent to some ψ( x ) ∈ Lqf ;

2. ϕ( a) ↔ ϕ( f a) for every partial isomorphism f : U → U defined in a.

Use the result to prove Theorem 7.13 for T complete. 

9.21 Proposition Let p( x ) ⊆ L(U) then the following are equivalent

1. p( x ) is equivalent to some type q( x ) ⊆ L( A);

2. p(U) is invariant over A.

We give two proofs of this theorem. The second one requires Proposition 9.22 below.
Proof 1⇒2 Obvious.
2⇒1 It suffices to show that for every formula ψ( x ) ∈ p( x ) there is a formula
ϕ( x ) ∈ L( A) such that p( x ) → ϕ( x ) → ψ( x ). Fix ψ( x ) ∈ p( x ). By invariance, any
q( x ) ∈ S( A) consistent with p( x ) implies p( x ), hence
_
p( x ) → q( x ) → ψ( x )
q( x )→ψ( x )

where q( x ) above range over all types in Sx ( A). By compactness we can rewrite
this equivalence as follows
_
p( x ) → ϑ( x ) → ψ( x )
ϑ ( x )→ψ( x )

where ϑ ( x ) ranges over all formulas in L( A). Applying mutual normality (Propo-
sition 9.17) to the first implication we obtain a finite number of formulas ϑi ( x ) such
that
n
_
p( x ) → ϑi ( x ) → ψ ( x ) .
i =1
This completes the proof. 

The following easy proposition is very useful. Its proof is left to the reader. Note

75
 that it would not hold without saturation saturation: for instance, consider R as
a structure in the language of strict orders and let q( x, y) = tp(−1, 0 / A), where
A = {1/n : 0 < n}. By quantifier elimination, −1 ≡ A 0. But 0 2 ∃y q( x, y).

9.22 Proposition Let p( x ; z) ⊆ L( A). Then ∃z p( x ; z) is equivalent to a type over A, namely


to the type ∃z ϕ( x ; z) : ϕ( x ; z) conjunction of formulas in p( x ; z) . The theorem holds
also when x and z have length κ. 

As an application we give a second proof of the proposition above.

Second proof of Proposition 9.21 2⇒1 Write p( x ) as the type q( x ; b) for some
q( x ; z) ⊆ L and some b ∈ U|z| . Let s(z) = tp(b/A). By invariance and homogeneity
the types p( x ; f b) for f ∈ Aut (U/A) are all equivalent. Therefore
_
p( x ) ↔ q( x ; f b)
f ∈Aut (U/A)
_
p( x ) ↔ q( x ; c)
c ≡A b
 
↔ ∃z s(z) ∧ q( x ; z) .

Hence, by Proposition 9.22, p( x ) is equivalent to a type over A. 

9.23 Exercise Let p( x ) ⊆ L( A), with | x | < ω. Prove that if p(U) is infinite then it has
cardinality κ. Show that this may not be true if x is an infinite tuple. 

9.24 Exercise Let ϕ( x, y) ∈ L(U). Prove that if the set ϕ( a, U) : a ∈ U| x| is infinite



then it has cardinality κ. Does the claim remains true with a type p( x, y) ⊆ L( A)
for ϕ( x, y)? 

9.25 Exercise Let ϕ( x ; y) ∈ L(U). Prove that the following are equivalent
1. there is a sequence h ai : i ∈ ω i such that ϕ(U ; ai ) ⊂ ϕ(U ; ai+1 ) for every i < ω;

2. there is a sequence h ai : i ∈ ω i such that ϕ(U ; ai+1 ) ⊂ ϕ(U ; ai ) for every i < ω. 

9.26 Exercise Let ϕ( x ; y) be such that for every b ∈ U|y| there is a ψ( x ) ∈ L such that
ψ( x ) → ϕ( x ; b). Prove that we can choose ψ( x ) in a finite subset of L. 

76
 Chapter 10
Preservation theorems
In this chapter we present a few results dating from the 1950s that describe the re-
lationship between syntactic and semantic properties of first-order formulas. These
results characterize the classes of formulas that preserved under various sorts of
morphisms. Criteria for quantifier-elimination follow from these theorems, see for
instance the frequently used back-and-forth method of Corollary 10.12.

1 Lyndon-Robinson Lemma
In this section T is a consistent theory without finite models and ∆ is a set of formu-
las closed under renaming of variables. At a first reading the reader is encouraged
to assume that ∆ is Lat .

10.1 Definition If C ⊆ {∀, ∃, ¬, ∨, ∧} is a set of connectives, we write C∆ for the closure of

∆ with respect to all connectives in C. We may write ∆± for {¬}∆. 

Recall that ∆-morphism is a map k : M → N that preserves the truth of formulas

in ∆. It is immediate that ∆-morphism are automatically {∧∨}∆-morphisms. As
∆ is closed under renaming of variables, {∃}∆-morphism are {∃∧∨}∆-morphisms
and similarly {∀}∆-morphism are {∀∧∨}∆-morphisms.
Below we use the following proposition without further reference.

10.2 Proposition Fix M  T and b ∈ M| x| . Let q( x ) = ∆-tp M (b). Then for every ϕ( x ) ∈ L
the following are equivalent
1. N  ϕ(kb) for every k : M → N  T that is a ∆-morphism defined in b;

2. T ` q( x ) → ϕ( x ).

Proof 2⇒1 Immediate.

1⇒2 Negate 2, then there are N  T and c ∈ N | x| such that q(c) ∧ ¬ ϕ(c). Therefore
the map k : M → N, where k = {hb, ci}, contradicts 1. 

The following is sometimes referred to as the Lyndon-Robinson Lemma.

10.3 Lemma For every ϕ( x ) ∈ L the following are equivalent

1. ϕ( x ) is equivalent over T to a formula in {∧∨}∆;

2. ϕ( x ) is preserved by ∆-morphisms between models of T.

Proof 1⇒2 Immediate.

2⇒1 We claim that from 2 it follows that
p( x ) ⊆ ∆ : T ` p( x ) → ϕ( x )
_
# T ` ϕ( x ) ↔

Direction ← is clear. To verify →, fix M  T and let b ∈ M| x| be such that M  ϕ(b).

77
 From 2 it follows that ϕ( x ) satisfy 1 of Proposition 10.2. Therefore T ` q( x ) → ϕ( x )
for q( x ) = ∆-tp(b). Hence q( x ) is one of the types that occur in the disjuction in #
which therefore is satisfied by b.
From # and compactness we obtain
ψ( x ) ∈ {∧}∆ : T ` ψ( x ) → ϕ( x )
_
T ` ϕ( x ) ↔
Applying compactness again, we replace the infinite disjunction above with a finite
one and prove 2. 

10.4 Proposition Let N be λ-saturated and let k : M → N be a ∆-morphism of cardinality

< λ. Then the following are equivalent
1. k : M → N is a {∃}∆-morphism;

2. for every b ∈ M some {∃}∆-morphism h : M → N defined in b extends k;

3. for every b ∈ Mω some ∆-morphism h : M → N defined in b extends k.

Proof 1⇒2 Let a enumerate dom k. Define p( x ; z) = {∃}∆-tp M (b ; a). By 1,

p( x ; ka) is finitely consistent in N. By saturation there is a c ∈ N that realizes
p( x ; ka). Therefore, h : M → N where h = k ∪ {hb, ci}, witnesses 2.
2⇒3 Iterate ω-times the extension in 2.
3⇒1 Let a enumerate dom k and let |z| = | a|. Formulas in {∃}∆ with free variables
among z are of the form ∃ x ϕ( x ; z) where ϕ( x ; z) is in ∆ and x is some fixed tuple
of length ω. Assume M  ∃ x ϕ( x ; a)) and let b be such that M  ϕ(b ; a). By 3, we
can extend k to some ∆-morphism h : M → N defined in b. Then N  ϕ(hb ; ha)
and therefore N  ∃ x ϕ( x ; ka). 

Iterating the lemma above we obtain the following.

10.5 Corollary Let N be λ-saturated and let | M| ≤ λ. Let k : M → N be a ∆-morphism of

cardinality < λ. Then the following are equivalent
1. k : M → N is a {∃}∆-morphism;

2. k : M → N extends to an {∃}∆-embedding;

3. k : M → N extends to an ∆-embedding. 

The following theorem is often paraphrased as follows: a formula is existential if

and only if (its truth) is preserved under extensions of structures.

10.6 Theorem For every ϕ( x ) ∈ L the following are equivalent

1. ϕ( x ) is equivalent over T to a formula in {∃∧∨}∆;

2. ϕ( x ) is preserved by ∆-embedding between models of T.

Proof 1⇒2 Immediate.

2⇒1 Negate 1. By Lemma 10.3 there is a {∃}∆-morphism k : M → N between
models of T that does not preserve ϕ( x ). We can assume that N is λ-saturated for
some sufficiently large λ. By Corollary 10.5 there is a ∆-embedding h : M ,→ N that
extends k and contradicts 2. 

78
 A dual version of the results above is obtained replacing embeddings by epimor-
phisms (surjective homomorphisms) and {∃} by {∀}. If ∆ contains the formula
x = y and is closed under negation, then k : M → N is a ∆-morphism if and only
is k−1 : N → M is a ∆-morphism. In this case the dual version follows from what
proved above. Without these assumptions the results need a similar but indepen-
dent proof.

10.7 Proposition Let M be λ-saturated and let k : M → N be a ∆-morphism of cardinality

< λ. Then the following are equivalent
1. k : M → N is a {∀}∆-morphism;

2. for every c ∈ N some {∀}∆-morphism h : M → N extends k and c ∈ img h;

3. for every c ∈ N ω some ∆-morphism h : M → N extends k and c ∈ (img h)ω .

We write ¬∆ for the set containing the negation of the formulas in ∆. Warning: do
not confuse ¬∆ with {¬}∆.
Proof Left as an exercise for the reader. Hint: to prove implication 1⇒2 define
p( x, y) = ¬{∀}∆-tp N (ka, c), where a is a tuple that enumerates dom k. From 1
obtain that p( a, y) is finitely consistent in M. Then proceed as in the proof of
Proposition 10.4. 

10.8 Corollary Let M be λ-saturated and let | N | ≤ λ. Let k : M → N be a ∆-morphism of

cardinality < λ. Then the following are equivalent
1. k : M → N is a {∀}∆-morphism;

2. k : M → N extends to an {∀}∆-epimorphism;

3. k : M → N extends to an ∆-epimorphism. 

Finally we obtain the following.

10.9 Theorem The following are equivalent

1. ϕ( x ) is equivalent to a formula in {∀}∆;

2. every ∆-epimorphism between models of T preserves ϕ( x ). 

2 Quantifier elimination by back-and-forth

We say that T admits (or has ) positive ∆-elimination of quantifiers if for every
formula ϕ( x ) in {∃ ∀∧∨}∆ there is a formula ψ( x ) in {∧∨}∆ such that
T ` ϕ ( x ) ↔ ψ ( x ).
When ∆ is closed under negation the attribute positive becomes irrelevant and will
be omitted. When ∆ is Lat± or Lqf , we simply say that T admits elimination of
quantifiers. This is by far the most common case.
Quantifier elimination is often used to prove that a theory is complete because it
reduces it to something much simpler to prove. The following is an immediate
consequence of the definition above with x replaced by the empty tuple.

79
10.10 Remark If T has elimination of quantifiers then the following are equivalent
1. T decides all quantifier free sentences;

2. T is complete.

Hence a theory with quantifier elimination is complete if it decides the characteris-

tic, see Definition 5.26). 

The following is a consequence of Lemma 10.3.

10.11 Corollary The following are equivalent

1. T has ∆-elimination of quantifiers;

2. every ∆-morphism between models of T is both a {∃}∆ and a {∀}∆-morphism.

Proof 1⇒2 Immediate.

2⇒1 We prove by induction of syntax that ∆-morphism preserve the truth of all
formulas in {∃ ∀∧∨}∆, this suffices by Lemma 10.3. Induction for the connectives
∨ and ∧ is trivial. So assume as induction hypothesis that the truth of ϕ( x, y) is
preserved. By Lemma 10.3 ϕ( x, y) is equivalent to a formula in {∧∨}∆, hence by 2
the truth of ∃y ϕ( x, y) and ∀y ϕ( x, y) is preserved. 

Condition 2 of the corollary above may be difficult to verify directly. The following
corollary of Proposition 10.4 and 10.7 gives a back-and-forth condition with is easier
to verify.

10.12 Corollary Let | L| ≤ λ. The following are equivalent

1. T has ∆-elimination of quantifiers;

2. for every finite ∆-morphism k : M → N between λ-saturated models of T

a. for every b ∈ M some ∆-morphism h : M → N extends k and b ∈ dom h;

b. for every c ∈ N some ∆-morphism h : M → N extends k and c ∈ img h. 

Note that when ∆ contains the formula x = y and is closed under negation, then
k : M → N is a ∆-morphism if and only if k−1 : N → M is a ∆-morphism. In this
case a and b are equivalent.

10.13 Exercise Let T be a complete theory without finite models in a language that con-
sists only of unary predicates. Prove that T has elimination of quantifiers. 

10.14 Exercise Let L be the language of strict orders. The theory of discrete linear orders
extends the theory of linear orders without endpoints (see Section 6.1) with the
following two of axioms
 
dis↑. ∃z x < z ∧ ¬∃y x < y < z ;
 
dis↓. ∃z z < x ∧ ¬∃y z < y < x .

Let ∆ be the set of formulas that contains (all alphabetic variants of) the formulas
x <n y := ∃≥n z ( x <z<y) and their negations (read <0 as <). Prove that the theory
of discrete linear orders has ∆-elimination of quantifiers. Prove that the structure
Q × Z ordered with the lexicographic order

80
 ( a1 , a2 ) < (b1 , b2 ) ⇔ a1 < b1 or (a1 = b1 e a2 < b2 )
is a saturated model of T. 

10.15 Exercise Let T be a consistent theory. Suppose that all completions of T are of the
form T ∪ S for some set S of quantifier-free sentences. Prove that, if all completion
of T have elimination of quantifiers, so does T. Show that this fails when the
completions of T have arbitrary complexity.
Note. Thought the claim follows immediately from Corollary 10.11, a direct proof
is also instructive. Prove that for every formula ϕ( x ) there are some quantifier-free
sentences σi and quantifier-free formulas ψi ( x ) such that
n
_
σi ` ϕ( x ) ↔ ψi ( x ), T` σi , and σi ` ¬σj for i 6= j.
i =1
For a counter example consider the empty theory in the language with a single
unary predicate. 

3 Model-completeness
We say that T is model-complete if every embedding h : M ,→ N between models
of T is an elementary embedding.
The notion was introduced by Abraham Robinson and has become standard. It
is inspired by the fact that T is model-complete if and only if T ∪ Diag( M) is a
complete theory, in the language L( M ), for every M  T.
To stress positivity in the next proposition, we generalize the definition as follows.
We say that T is ∆-model-complete if every ∆-embedding h : M ,→ N between
models of T is a {∀∃}∆-embedding.
Model-completeness is equivalent a sort of elimination of quantifiers.

10.16 Proposition The following are equivalent

1. T is ∆-model-complete;

2. T has {∃}∆-elimination of quantifiers.

Proof 1⇒2 By 1, every formula {∀∃∧∨}∆ is preserved by ∆-embeddings there-

fore, by Theorem 10.6, it is equivalent to a formula in {∃∧∨}∆.
2⇒1 Clear, because ∆-embeddings preserve formulas in {∃∧∨}∆. 

The theory of discrete linear orders defined in Exercise 10.14 is an example of a

model-complete theory without elimination of quantifiers.

10.17 Exercise Prove that the theory of discrete linear orders is model-complete. 

The difference between quantifier elimination and model-completeness subtle. It

boils down to models of T having or not the amalgamation property.

10.18 Proposition Assume T is model-complete. Let M be the category that consists of models
of T and partial isomorphisms. Then the following are equivalent

81
 1. M has the amalgamation property;

2. T has elimination of quantifiers.

Proof 1⇒2 By Proposition 7.25 every partial morphism k : M → N extends to

an embedding g : M ,→ N 0 which, by model-completeness, is an elementary em-
bedding. Model-completeness also implies that N  N 0 . Hence k : M → N is an
elementary map. This proves 2.
2⇒1 If all morphisms are elementary maps, amalgamations follows from Propo-
sition 7.28. 

Note however that the models of a model-complete theory T do have amalgamation

when the proper notion of morphism is chosen.
Let M0 be the category that consists of models of T and the maps k : M → N such
that there is a partial isomorphism h : M0 → N 0 with
1. k ⊆ h; M  M0 ; N  N0;
2. dom h contains a substructure of M 0 that models T (equivalently img h and N 0 ).

It is clear that if T is model-complete then the morphisms of M0 are exactly the

elementary maps. In this case M0 has amalgamation. Vice versa if M0 has amalga-
mation, the theory of rich models is model-complete.

10.19 Exercise Prove that M0 satisfies finite character of morphisms, c6 of Definition 7.1. 

82
 Chapter 11
Geometry and dimension
In this chapter we fix a signature L, a complete theory T without finite models, and
a saturated model U of inaccessible cardinality κ larger than | L|. The notation and
implicit assumptions are as in Section 9.3.

1 Algebraic and definable elements

Let a ∈ U and let A ⊆ U be some set of parameters (of arbitrary cardinality).
We say that a is algebraic over A if ϕ( a) ∧ ∃=k x ϕ( x ) holds for some formula
ϕ( x ) ∈ L(A) and some positive integer k. In particular when k = 1 we say that a
is definable over A . We write acl(A) for the algebraic closure of A , that is, set of
all elements algebraic over A. If A = acl(A), we say that A is algebraically closed .
The definable closure of A is defined similarly and is denoted by dcl(A) .
A formula ϕ(u) ∈ L( A), or a type p(u) ⊆ L( A), where u is a finite tuple, with
finitely many solutions are called an algebraic .

11.1 Proposition For every A ⊆ U and every type p(u) ⊆ L( A), where |u| < ω, the following
are equivalent
1 ∃ ≤ n u p ( u );
2 ∃≤n u ϕ(u) for some ϕ(u) which is conjunction of formulas p(u).
Proof The non trivial implication is 1⇒2. Let { ai , . . . , an } be all the solutions of
p(u). Then
n
_
p(u) ↔ ai = u
i =1
Then 2 follows by compactness (cfr. Remark 9.18.b). 

11.2 Exercise Prove that Proposition 11.1 does not hold in general for infinite tuple u. 

11.3 Exercise For every a ∈ U|u| and A ⊆ U, the following are equivalent
1. a is solution of some algebraic formula ϕ(u) ∈ L( A);

2. a = a1 , . . . , an for some a1 , . . . , an ∈ acl( A).

Prove also that if a is definable then so are a1 , . . . , an , and vice versa. 

11.4 Theorem For every A ⊆ U and every a ∈ U the following are equivalent
1 a ∈ dcl( A);
2 O( a/A) = a .


83
 Proof Implication 1⇒2 is obvious. As for 2⇒1, recall that O( a/A) is the set of
realizations of tp( a/A), then the theorem follows from Proposition 11.1. 

11.5 Theorem For every A ⊆ U and every a ∈ U the following are equivalent
1 a ∈ acl( A);
2 O( a/A) is finite;
3 a belongs to every model containing A.
Proof 1⇔2. This is proved as in Theorem 11.4.
1⇒3. Assume 1, then there is a formula ϕ( x ) ⊆ L( A) such that ϕ( a) ∧ ∃=k x ϕ( x )
for some k. By elementarity ∃=k x ϕ( x ) holds in every model M containing A. Again
by elementarity, the k solutions of ϕ( x ) in M are solutions in U, therefore a is one
of these.
3⇒2. Assume O( a/A) is infinite and fix any model M containing A. By Exer-
cise 9.23, O( a/A) has cardinality κ, hence O( a/A) * M. Pick any f ∈ Aut (U/A)
such that f a ∈ / f −1 [ M], so f −1 [ M] is a models that contradicts 3.
/ M. Then a ∈ 

11.6 Corollary For every A ⊆ U and every a ∈ U the following hold

1 Se a ∈ acl A then a ∈ acl B for some finite B ⊆ A; finite character
2 A ⊆ acl A; extensivity
3 se A ⊆ B allora acl A ⊆ acl B; monotonicity
4 acl A = acl(acl A); idempotency
\
5 acl A = M.
A⊆ M

Properties 1-4 say that the algebraic closure is a closure operator with finite charac-
ter.
Proof Properties 1-3 are obvious, 4 follow from 5 which in turn follows from The-
orem 11.5. 

11.7 Proposition If f ∈ Aut (U) then f acl( A) = acl f [ A] for every A ⊆ U.

 

 

Proof We prove f acl( A) ⊆ acl f [ A] . Fix a ∈ acl( A) and let ϕ( x ; z) ∈ L and
b ∈ A|z| be such that ϕ( x ; b) is algebraic formula satisfied by a. By elementarity,
ϕ( x ; f b) is algebraic and satisfied by f a. Therefore f a is algebraic over f [ A], which
proves the inclusion.
The converse inclusion is obtained by substituting f −1 for f and f [ A] for A. 

11.8 Exercise Let ϕ(z) ∈ L( A) be a consistent formula. Prove that, if a ∈ acl( A, b) for
every b  ϕ(z), then a ∈ acl( A). Prove the same claim with a type p(z) ⊆ L( A) for
ϕ ( z ). 

11.9 Exercise Let a ∈ U r acl ∅. Prove that U is isomorphic to some V  U such

/ V. Hint: let c be an enumeration of U and let p(u) = tp(c) prove that
that a ∈
p(u) ∪ ui 6= a : i < κ is realized in U and that any realization yields the required


84
 substructure of U. 

11.10 Exercise Let C be a finite set. Prove that if C ∩ M 6= ∅ for every model M contain-
ing A, then C ∩ acl( A) 6= ∅. Hint: by induction on the cardinality of C. Suppose
there is a c ∈ C r acl( A), then there is V ' U such that A ⊆ V  U and c ∈/ V, see
0
Exercise 11.9. Apply the induction hypothesis to C = C ∩ V with V for U. 

11.11 Exercise Prove that for every A ⊆ N there is an M such that acl A = M ∩ N. Hint:
add the requirement acl( Ai ) ∩ N ⊆ acl( A) to the construction used to prove the
downward Löwenheim-Skolem theorem. You need to prove that every consistent
ϕ( x ) ∈ L( Ai ) has a solution a such that acl( Ai , a) ∩ N ⊆ acl( A). The required a has
to realize the type
n  o
ϕ( x ) ∪ ¬ ψ(b, x ) ∧ ∃≤n y ψ(y, x ) : b ∈ N r acl( A), ψ(y, x ) ∈ L( Ai ), n < ω
 

whose consistence need to be verified. (See Exercise 11.12 for a different hint.) 

11.12 Exercise Prove that for every A ⊆ N there is an automorphism f ∈ Aut (U/A) such
that acl A = f [ N ] ∩ N. (This is a stronger version of the claim in Exercise 11.11.)
Hint: let c be an enumeration of N. Let p( x ) = tp(c/A). Consider the type
n  o
p( x ) ∪ ¬ ψ(b, x ) ∧ ∃≤n y ψ(y, x ) : b ∈ N r acl( A), ψ(y, x ) ∈ L x ( A), n < ω


Any a  p( x ) enumerates a model A-isomorphic to N. 

11.13 Exercise Let ϕ( x ) ∈ L(U) and fix an arbitrary set A. Prove that the following are
equivalent
1. there is some model M containing A and such that M ∩ ϕ(U) = ∅;

2. there is no consistent formula ψ(z1 , . . . , zn ) ∈ L( A) such that

^n
ψ ( z1 , . . . , z n ) → ϕ ( z i ).
i =1

Hint: let c = hci : i < λi be an enumeration of N | x| , where N is any model

containing A. Let p(z) = tp(c/A). Prove that 2 implies the consistency of the type

p(z) ∪ ¬ ϕ(zi ) : i < λ and deduce the existence of the required M. 

2 Strongly minimal theories

Finite and cofinite sets are always (trivially) definable in every structure. We say that
M is a minimal structure all its definable subsets of arity one are finite or cofinite.
Unfortunately, this notion is not elementary, i.e. it is not a property of Th( M ). For
instance N with only the order relation in the language is a minimal structure but
none of its elementary extensions is. Hence the following definition: we say that
M is a strongly minimal structure if it is minimal and all its elementary extensions
are minimal.
We say that T, a consistent theory without finite models, is strongly minimal if for
every formula ϕ( x ; z) ∈ L, where x has arity one, there is an n ∈ ω tale che
h i
T ` ∀ z ∃≤n x ϕ ( x ; z ) ∨ ∃≤n x ¬ ϕ ( x ; z ) .

85
 We show that the semantic notions match with and the syntactic one.

11.14 Proposition The following are equivalent

1. Th( M ) is a strongly minimal theory;

2. M is a strongly minimal structure;

3. M has an elementary extension which is minimal and ω-saturated.

Proof Implications 1⇒2⇒3 are immediate, we prove 3⇒1. Let ϕ( x ; z) ∈ L and fix
an elementary extension N as required by 2. Let p(z) ⊆ L be the following type
n o
p(z) = ∃>n x ϕ ( x ; z ) ∧ ∃>n x ¬ ϕ ( x ; z ) : n ∈ ω .
As N is minimal, N 2 ∃z p(z). By saturation p(z) is not finitely consistent in M.
Hence, for some n
h i
M  ∀z ∃≤n x ϕ ( x ; z ) ∨ ∃≤n x ¬ ϕ ( x ; z ) .

which proves that Th( M ) is strongly minimal. 

By quantifier elimination, Tacf and Tdag are strongly minimal theories.

11.15 Exercise Let T be a complete theory without finite models. Prove that the following
are equivalent
1. M is minimal;

2. a ≡ M b for every a, b ∈ U r M. 

3 Independence and dimension

Thoughout this section we assume that T is a complete strongly minimal theory.
When a ∈ / acl B we say that a is algebraically independent from B. We say that B

is an algebraically independent set if every a ∈ B is independent from B r a .
 
Below we shall abbreviate B ∪ a by B, a and B r a by B r a .
The following is a pivotal property of independence that holds in strongly minimal
structures. It is called symmetry or exchange principle . For every B and every
pair of elements a, b ∈ U r acl B
b ∈ acl( B, a) ⇔ a ∈ acl( B, b)
Note that when T = Tacf this principle is the so called Steinitz exchange lemma .

11.16 Theorem (T strongly minimal.) Independence is symmetric.

Proof Suppose b ∈ / acl( B, a) and a ∈ acl( B, b). We prove that a ∈ acl B. Fix a
formula ϕ( x, y) ∈ L( B) such that ϕ( x, b) witness a ∈ acl( B, b), i.e. for some n
ϕ( a, b) ∧ ∃≤n x ϕ( x, b).
As b ∈
/ acl( B, a), the formula
ψ( a, y) = ϕ( a, y) ∧ ∃≤n x ϕ( x, y).
is not algebraic. Therefore, by strong minimality, ψ( a, y) has cofinitely many so-
lutions. Hence every model containing B contains a solution of ψ( a, y). As a is

86
 algebraic in any of these solutions, a belongs to every model containing B. There-
fore, a ∈ acl B by Theorem 11.5. 

We say that B ⊆ C is a base of C if B is an independent set and C ⊆ acl B. The

following theorem proves that all bases have the same cardinality, which we call the
dimension of C and denote by dim C . First we need the following lemma.

11.17 Lemma (T strongly minimal.) If B is an independent set and a ∈

/ acl B then B, a is also
an independent set.
Proof Suppose B, a is not independent and that a ∈/ acl B. Then b ∈ acl( B r b, a) for
some b ∈ B. As a, b ∈
/ acl( B r b), from symmetry we obtain a ∈ acl( B r b, b) = acl B.
Hence B is not an independent set. 

11.18 Corollary (T strongly minimal.) For every B ⊆ C the following are equivalent
1. B is a base of C.

2. B is a maximally independent subset of C. 

We may write acl( B/A) per acl( B ∪ A). The notation is used to suggest to the
reader that the role of A in the argument is irrelevant and that it could be absorbed
in the language. Independent sets over A and bases over A are defined in the
obvious way.

11.19 Theorem (T strongly minimal.) Fix some arbitrary set C, then

1 every independent set B ⊆ C can be extended to a base of C;
2 all bases of C have the same cardinality;
3 claims 1 and 2 hold over any set of parameters A.
Proof By the finite character of algebraic closure, the independent set form an in-
ductive class. Apply Zorn lemma to obtain a maximally independent subset of C
containing B. By Corollary 11.18 this set is a base of C. This prove 1 and it is
immediate that the argument can be generalized as required in 3.
As for 2, assume for a contradiction that B1 , B2 ⊆ C are two bases of C and that
| B1 | < | B2 |. First consider the case when B2 is infinite. For each b ∈ B1 fix a set
Db ⊆ B2 such that b ∈ acl( Db ). Let
[
D = Db .
b∈ B1

Then B1 ⊆ acl D and | D | < | B2 |. By transitivity, C ⊆ acl D which contradicts the

independence of B2 . Again the argument generalizes immediately as required in 3.
Now we suppose that B2 is finite and prove directly the generalized version. Let n
be the least integer such that, for some set of parameters A, and some B1 and B2
bases of C over A, we have | B1 | < | B2 | = n + 1. Fix any c ∈ B2 . Then B2 r c is a
base over A, c. Note that B1 is not independent over A, c. In fact, as c ∈ acl( B/A)
by symmetry we obtain that b ∈ acl( B r b/A, c) for any b ∈ B1 . Now, let D1 ⊂ B1
be maximal independent over A, c. Then D1 and B2 r c are two bases over A, c such
that | D1 | < | B2 r c| = n, which contradicts the minimality of n. 

87
 
11.20 Proposition (T strongly minimal.) Let k be an elementary map. Then k ∪ hb, ci is also
an elementary map for every b ∈
/ acl(dom k) and c ∈
/ acl(img k).
Proof Let a be an enumeration of dom k. We need to show that ϕ(b ; a) ↔ ϕ(c ; ka)
holds for every ϕ( x ; z) ∈ L. As k is elementary, the formulas ϕ( x ; a) and ϕ( x ; ka)
are either both algebraic or both co-algebraic. As b ∈ / acl( a) and c ∈/ acl(ka), they
are both false or both true respectively. So the proposition follows. 

11.21 Corollary (T strongly minimal.) Every bijection between independent sets is an elemen-
tary map. 

Finally we show that dimension classifies models of T.

11.22 Theorem (T strongly minimal.) Models of T with the same dimension are isomorphic.
Proof Let A e B be some bases of M and N respectively. By Corollary 11.21, any
bijectiion between A and B is an elementary map. By Proposition 11.7, it extends to
the required isomorphism between acl A = M and acl B = N. 

11.23 Corollary (T strongly minimal.) Strongly minimal theories are λ-categorical for every
λ > | L |.
Proof As | acl A| ≤ | L( A)|, all models of cardinality λ have the same dimension λ. 

11.24 Proposition (T strongly minimal.) For every model N of cardinality ≥ | L| the following
are equivalent
1. N is saturated;

2. dim N = | N |.

Proof 2⇒1. Assume 2 and let k : M → N be an elementary map of cardinality

< | N | and let b ∈ M. We want an extension of k defined in b. If b ∈ acl(dom k)
the required extension exists by Proposition 11.7. Otherwise, pick any c ∈ N r

acl(img k), which exist as |k| < dim N = | N |. By Proposition 11.20, k ∪ hb, ci is
the required extension.
1⇒2. If B ⊆ N is a base of N the following type is not realized in N
n o
p( x ) = ¬ ϕ( x ) : ϕ( x ) ∈ L( B) is algebraic
Therefore, if N is saturated, | B| = | N |. 

11.25 Exercise (T strongly minimal.) Prove that every infinite algebraically closed set is
a model. 

11.26 Exercise (T strongly minimal.) Prove that every model is homogeneous. 

11.27 Exercise (T strongly minimal.) Prove that if dim N = dim M + 1 then there is no
model K such that M ≺ K ≺ N. 

88
 Chapter 12
Countable models
In this chapter we fix a signature L, a complete theory T without finite models,
and a saturated model U of inaccessible cardinality κ larger than | L|. (Expect that
all relevant theorems are proved under the assumption that L is countable.) The
notation and implicit assumptions are as in Section 9.3.

1 The omitting types theorem

We say that the formula ϕ( x ) isolates the type p( x ), if ϕ( x ) is consistent and
ϕ( x ) → p( x ). When ∆ is a set of formulas, we say that ∆ isolates p( x ) if some
formula in ∆ does. In this chapter ∆ is always a set of the form L x ( A) and we say
A isolates p( x ) or every simply p( x ) is isolated when A is clear. We say that a
model M omits p( x ) if it does not realized it.
Observe that if p( x ) ⊆ L( M) then M realizes p( x ) if and only if M isolates p( x ).
Therefore if A isolates p( x ), then every model containing A realizes p( x ). Below
we prove that the converse also holds when L and A are countable. This a famous
classical theorem which is called the omitting types theorem.
The construction of a model M omitting p( x ) proceeds by stages. Along the con-
struction we preserve the property of being isolated since that of being realized is not
well behaved in sets which are not models. In the limit, when a model is obtained,
these two properties coincide.
The omitting types theorem require the following lemma which is an analogue of
the Kuratowski-Ulam theorem (i.e., the so-called Fubini theorem for Baire category).

12.1 Lemma Assume L( A) is countable. Let p( x ) ⊆ L( A) and suppose that A does not isolate
p( x ). Then every consistent formula ψ(z) ∈ L( A) has a solution a such that A, a does not
isolate p( x ).
Proof We construct sequence of formulas hψi (z) : i < ω i such that any realization

a of the type q(z) = ψi (z) : i < ω is as required by the lemma.
To begin with, we fix an enumeration hξ i ( x ; z) : i < ω i of all formulas in L x ;z ( A)
and set ψ0 (z) = ψ(z). At stage i + 1, if ξ i ( x ; z) ∧ ψi (z) in inconsistent, let ψi+1 (z) =
ψi (z), otherwise let
 
ψi+1 (z) = ψi (z) ∧ ∃ x ξ i ( x ; z) ∧ ¬ ϕ( x )
for some formula ϕ( x ) ∈ p such that the resulting ψi+1 (z) is consistent. If possible,
this guarantees that ξ i ( x ; a) for any a  q(z) does not isolate p( x ). The proof
is complete if we can show that it is always possible to find a formula ϕ( x ) as
required.
Suppose by way of contradiction that no formula makes ψi+1 (z) consistent, that is,
ξ i ( x ; z) ∧ ψi (z) → ϕ( x )

89
 for every ϕ( x ) ∈ p. This immediately imply that
 
∃z ξ i ( x ; z) ∧ ψi (z) → p( x ),
that is, p( x ) is isolated by a formula in L x ( A), which contradicts our assumptions. 

12.2 Theorem (Omitting types theorem) Assume L( A) is countable. Then for every consis-
tent p( x ) ⊆ L( A) the following are equivalent
1. all models containing A realize p( x );

2. p( x ) is isolated.

The theorem owes its name to the contrapositive of implication 1⇒2.

Proof Implication 2⇒1 is clear, we prove 1⇒2. Assume A does not isolate p( x ).
The model M is the union of a chain h Ai : i < ω i of countable subsets of U that
starts with A0 = A. Along the construction we require as inductive hypothesis that
Ai does not isolate p( x ). At the end, M will not isolate p( x ) which for a model is
equivalent to omitting.
We proceed as in the proof of the downward Löwenheim-Skolem theorem (pre-
cisely, we use the second proof 2.40). Lemma 12.1 ensure that we can satisfy any
consistent formula in Ai and preserve the inductive hypothesis. 

Gerald Sacks once famously remarked: Any fool can realize a type but it take a model
theorist to omit one. (Though, as remarked above, the proof reminds more of set
theory than model theory.)

12.3 Example The following example shows that in the omitting types theorem we can-
not drop the assumption that L( A) is countable. Let F be the set of all bijection
between two uncountable sets, X and Y. Let M be the model with domain the dis-
joint union of F, X and Y. The language has a ternary relation symbol for f ( x ) = y
and unary relations for X and Y. Let U be a saturated elementary extension of M.
Let Y1 ⊆ Y be countable. Let c ∈ Y r Y1 and set p(y) = tp(c/X, Y1 ). We claim that
p(y) is not isolated but that it is realized in every model containing X, Y1 .
No formula could be equivalent to p(y) (not even a type with finitely many param-
eters). In fact, for every finite sets Y0 ⊆ Y1 and X0 ⊆ X there is a automorphism of
M that fixes Y0 ∪ X0 and maps c to some (any) element of Y1 r Y0 . This automor-
phism extends to an automorphism of U and maps c to an element that does not
realize p(y).
The type p(y) cannot be omitted because every model N containing X has neces-
sarily uncountably many elements of sort Y and therefore realizes p(y). 

12.4 Exercise Let p( x ) ⊆ L( B) and pn ( x ) ⊆ L( A), for n < ω, be consistent types such
that
_
p( x ) → pn ( x )
n<ω
Prove that there is an n < ω and a formula ϕ( x ) ∈ L( A) consistent with p( x ) such
that

p ( x ) ∧ ϕ ( x ) → p n ( x ). 

90
12.5 Exercise Let M be a second countable topological space (i.e. the topology has a
countable base). We say that A ⊆ M is meager if it is the countable union of
nowhere dense sets.
Use Lemma 12.1 to prove the Kuratowski-Ulam Theorem, i.e. that for A ⊆ M2 the
following are equivalent (w.r.t. the product topology)
1. A is meager in M2 ;
n o
2. x ∈ M : A ∩ { x }× M is not meager is meager in M.

Hint: Use the base of the topology as predicates of a first-order language. 

2 Prime and atomic models

We say that M is prime over A if A ⊆ M and for every N containing A there is an
elementary embedding h : M → N that fixes A. When A is empty we simply say
that M is prime .
There is no syntactic analogue of primeness; the notion that comes most close (and
works pretty well when everything is countable) is atomicity. For a ∈ U| x| we
say that a is isolated over A if the type p( x ) = tp( a/A) is isolated. Note that this
equivalent to claiming that a is an isolated point in U| x| w.r.t. the A-topology defined
in Section 9.3. We say that M is atomic over A if A ⊆ M and every a ∈ M<ω is
isolated over A. When A is empty we say that M is atomic .

12.6 Proposition Let b and a be finite tuples. Then the following are equivalent
1. A isolates b, a;

2. A, a isolates b and A isolates a.

Proof Let p( x, z) = tp(b, a/A) and note that p( x, a) = tp( x/A, a) and ∃ x p( x, z) =
tp( a/A).
1⇒2 Let ϕ( x, z) ∈ p be such that ϕ( x, z) → p( x, z). It follows that ϕ( x, a) →
p( x, a) and ∃ x ϕ( x, z) → ∃ x p( x, z). Therefore 2 holds.
2⇒1 Fix ϕ( x ; z), ψ(z) ∈ L( A) such that ϕ( x ; a) isolates p( x ; a) and ψ(z) iso-
lates ∃ x p( x ; z). Let ξ ( x ; z) ∈ p be arbitrary. As ϕ( x ; a) → ξ ( x ; a), the formula
   
∀ x ϕ( x ; z) → ξ ( x ; z) belongs to ∃ x p( x, z). Hence ψ(z) → ∀ x ϕ( x ; z) → ξ ( x ; z) .
As this holds for all ξ ( x ; z) ∈ p, we conclude that ψ(z) ∧ ϕ( x ; z) isolates p( x, z). 

The straightforward direction of the proposition above yields the following useful
proposition.

12.7 Proposition If M is atomic over A then M is atomic over A, a for every finite a ∈ M<ω .
Proof Let b ∈ M| x| be a finite tuple. Then A isolates b, a hence A, a isolates b. 

12.8 Proposition Let k : M → N be an elementary map and suppose that M is atomic over

dom k. Then for every b ∈ M there is a c ∈ N such that k ∪ hb, ci : M → N is
elementary.
Proof Let p( x ; z) = tp(b ; a) where a is an enumeration of dom k. Let ϕ( x ; z) ∈ L
be such that ϕ( x ; a) → p( x ; a). Note that, by elementarity, ϕ( x ; ka) → p( x ; ka).

91
 Hence the required c is any solution of ϕ( x ; ka) in N. 

A limiting assumption in Proposition 12.7 is that a need to be finite. Therefore the

following proposition is restricted to countable models.

12.9 Proposition Any two is countable models atomic over A are isomorphic.
Proof Easy, using Propositions 12.7 and 12.8 and back-and-forth. 

12.10 Proposition Assume L( A) is countable. Then for every model M the following are equiv-
alent
1. M is countable and atomic over A;

2. M is prime over A.

Proof 1⇒2 By Propositions 12.7 and 12.8.

2⇒1 Some countable model containing A exists, as M embeds in it, M has also to
be countable. Now we prove that M is atomic over A. Suppose for a contradiction
that there is some b ∈ M<ω such that p( x ) = tp(b/A) is not isolated. By the
omitting types theorem there is a model N containing A that omits p( x ). Then
there cannot be any A-elementary embedding of M into N. 

12.11 Proposition Assume L( A) is countable. Then the following are equivalent

1. there are models atomic over A;

2. for every |z| < ω, every consistent ϕ(z) ∈ L( A) has a solution that is isolated over A.

Note that 2 says that in U|z| isolated points are dense w.r.t. the topology defined in
Section 9.3.
Proof Implication 1⇒2 hold by elementarity. To prove 2⇒1 we construct by induc-
tion a sequence h ai : i < ω i. Reasoning as in (the second proof of) the downword

Löverheim-Skolem theorem ensures that A ∪ ai : i < ω is a model. To obtain an
atomic model we require that ai is isolated over A.
Suppose ai has been defined and let ϕ(z) ∈ L( A) be the formula that isolates
tp( ai /A). Let ψ( x ; z) ∈ L( A) be such that ψ( x ; ai ) is consistent (we leave to the
reader the details of the enumeration of such formulas). Then ψ( x ; z) ∧ ϕ(z) is
also consistent and by assumption it has a solution b ; c that is isolated over A. As
ai ≡ A c, there is an A-automorphism such that f c = ai . Therefore f b ; ai is a
solution ψ( x ; z) that is also isolated over A. Then we can set ai = f b. 

3 Countable categoricity
Here we present some important characterizations of ω-categoricity. The second
property below can be stated in different equivalent ways; for convenience, these
equivalents are considered in a separate proposition. For the time being we intro-
duce the following generalization (which we will prove is completely unnecessary):
we say hat T is ω-categorical over A if any two countable models containing A are
isomorphic. We say ω-categorical for ω-categorical over ∅.

92
12.12 Theorem (Engeler, Ryll-Nardzewsky, and Svenonius Theorem) Assume L( A) is
countable. The following are equivalent:
1. T is ω-categorical over A;

2. every type p( x ) ⊆ L( A) with | x | < ω is isolated.

The set A is introduced for convenience. By 3 of Proposition 12.13 below, no theory

is ω-categorical over an infinite set, and categoricity over some finite A is equivalent
to categoricity over ∅ (see Exercise 12.14).
Proof The implication 1⇒2 is an immediate consequence of the omitting types
theorem. In fact, if p( x ) is a non-isolated A-type, then there are two countable
models M and N containing A such that M realizes p( x ) ⊆ L( A) while N omits
it. Then M and N cannot be isomorphic over A. As for implication 2⇒1, observe
that 2 implies that every countable model containing A is atomic over A. But, by
Proposition 12.9, countable atomic models are unique up to isomorphism. 

12.13 Proposition Fix a set A and a finite tuple of variables x. The following are equivalent
1. every A-type p( x ) is isolated;

2. Sx ( A) is finite;

3. L x ( A) is finite up to equivalence;

4. in U| x| there is a finite number of orbits under Aut (U/A).

Proof To prove the implication 1⇒2 observe that U| x| is the union of sets of the
form p(U) where p ∈ Sx ( A). If these types are isolated then U| x| is the union of
A-definable sets. By compactness this union has to be finite. To prove 2⇒1 let
p ∈ Sx ( A). If Sx ( A) is finite, ¬ p(U) is the union of finitely many type definable
sets. A finite union of type definable sets is type definable. So ¬ p(U) is type
definable. Hence p(U) is isolated. We prove implication 2⇒3 observe that each
formula in L x ( A) is equivalent to the disjunction of the types in Sx ( A) that contain
this formula. If Sx ( A) is finite, L x ( A) is finite up to equivalence. Implication 3⇒2
is clear and equivalence 2⇔4 follows from the characterization of orbits as type-
definable sets. 

12.14 Exercise Prove that the following are equivalent for every finite set A
1. T is ω-categorical;

2. T is ω-categorical over A. 

12.15 Exercise Prove that the following are equivalent

1. T is ω-categorical;

2. there is countable model that is both saturated and atomic. 

12.16 Exercise Assume L is countable and that T is complete. Suppose that for every
finite tuple x there is a model M that realizes only finitely many types in Sx ( T ).
Prove that T is ω-categorical. 

93
 4 Small theories
Let T be, as always in this chapter, a complete theory without finite models. We say
that T is small over A if Sx ( A) is countable for every x of finite length. When A
is empty, we simply say that T is small . The set A is introduced for convenience,
in the literature it is used only for A = ∅. A different term is used in another very
interesting case, e.g. a theory small over every countable set A is said to be ω-stable.

12.17 Proposition No T is small over an uncountable set and if T is small (over ∅) then it is
small over any finite A.

Proof Let a be an enumeration of a finite set A. As Sx ( A) = p( x ; a) : p ∈
Sx ;z ( T ) the proposition is immediate. 

Below we identify Sx ( A) with U| x| /≡ A

12.18 Definition Let ∆ be a set of formulas (we mainly use ∆ = L x ( A) in this section). A
binary tree of formulas in ∆ is a sequence h ϕs : s ∈ 2<λ i of formulas in ∆ ∪ {>} such
that

1. for each s ∈ 2λ the type ps = ϕsn : n < |s| is consistent;

2. ps ∪ pr is inconsistent for any two distinct s, r ∈ 2λ .

(Condition 2 is usually obtained by taking ϕs0 ↔ ¬ ϕs1 for every s.) We call λ the height
of the tree. If the height is not specified, we assume it is ω. We may depict a binary tree of
formulas as follows
...
ϕ11
...
ϕ1
...
ϕ10
...
ϕ∅
...
ϕ01
...
ϕ0
...
ϕ00
...

Where branches are consistent types and distinct branches are inconsistent. 

Let S(∆) be denote the set of maximal consistent ∆-types.

12.19 Lemma Suppose ∆ is countable and closed under negation. Then the following are equiv-
alent
1. there is a binary tree of formulas in ∆;

2. S ( ∆ ) = 2ω ;
3. S(∆) > ω.
Proof As implications 1⇒2⇒3 are clear, it suffices to prove 3⇒1. We assume that
S(∆) is uncountable and define a tree of formulas in ∆ by induction. Begin with
ϕ∅ = >. For s ∈ 2<ω let ps be as in the definition above. Assume as induction
hypothesis that ps has uncountably many extensions in S(∆). This will guarantee
the consistency of the branches.

94
 It suffices to show that there is a formula ψ ∈ ∆ such that both ps ∩ {ψ} and
ps ∩ {¬ψ} have uncountably many extensions in S(∆). In fact, we if we define
ϕs0 = ψ and ϕs1 = ¬ψ this preserves the induction hypothesis.
Suppose for a contradiction that such a formula ψ does not exist. Consider the
following type
n o
q = ξ ∈ ∆ : ps ∩ {¬ξ } has ≤ ω extensions in S(∆) .
This type is consistent otherwise ¬ξ 1 ∨ · · · ∨ ¬ξ n would hold for some ξ i ∈ q. This
cannot happen because ps has uncountably many extensions in S(∆) while the
ps ∩ {¬ξ i } have not.
If the formula ψ required above does not exist, q is complete hence belongs to S(∆).
But there are countably many types distinct from q, so this contradicts 3. 

12.20 Proposition Suppose L( A) is countable. The following are equivalent

1. T is small over A;

2. there exists a countable saturated model containing A;

3. there is no binary tree of formulas in L x ( A) for any finite x.

Proof 1⇒2 There is a countable model M containing A that is weakly saturated

(see Proposition 9.12). There is a countable homogeneous model N containing M
(see Exercise 9.13). Clearly N is also weakly saturated. Then it is saturated by
Corollary 9.13 .
2⇒3 Clear.
3⇒1 By Lemma 12.19. 

12.21 Proposition A small theory has countable atomic models over every countable set A.
Proof We prove that every formula in L x ( A), where | x | < ω, has a solution isolated
over A. Then it suffices to apply Proposition 12.11.
Suppose for a contradiction that ϕ( x ) ∈ L( A) is consistent but has no solution
isolated over A. Then there is a formula ψ( x ) ∈ L( A) such that both ϕ( x ) ∧ ψ( x ) and
ϕ( x ) ∧ ¬ψ( x ) are consistent, otherwise ϕ( x ) would imply a complete type and every
solution of ϕ( x ) would be isolated. Fix such a ψ( x ). Clearly neither ϕ( x ) ∧ ψ( x )
nor ϕ( x ) ∧ ¬ψ( x ) have a solution isolated over A. This allows to construct a tree of
formulas in L x ( A) and prove that T is not small over A. 

12.22 Exercise Suppose ∆ is countable and closed under negation. Prove that if there
is a binary tree of formulas then there is a binary tree such that ϕs0 = ¬ ϕs1 and
ϕ∅ = >. 

12.23 Exercise Let | x | = 1. Prove that if Sx ( A) is countable for every finite set A, then T
is small. 

12.24 Exercise (Vaught) Prove that no complete theory has exactly 2 countable models
(assume L is countable – though it is not really necessary).
Hint: suppose T has exactly two countable models. Then T is small and there

95
 are a countable saturated model N and an atomic model M ⊆ N. As T is not
ω-categorical, M 6' N and there is finite tuple a that is not isolated over ∅. Let K
be an atomic model over a. Clearly K 6' M and, by Exercises 12.14 and 12.15, also
K 6' N. 

5 A toy version of a theorem of Zil’ber

As an application we prove that if T is ω-categorical and strongly minimal then it
is not finitely axiomatizable.
We say that T has the finite model property if for every sentence ϕ ∈ L there is a
finite substructure A ⊆ U such that
fmp Uϕ ⇔ Aϕ
The property is interesting because of the following proposition.

12.25 Proposition If T has the finite model property then it is not finitely axiomatizable.
Proof Assume fmp and suppose for a contadiction that there is a sentence ϕ ∈ L
such that T ` ϕ ` T. Then A  T for some finite structure A. But T ` ∃>kx ( x = x )
for every k. A contradiction. 

We need the following definition. We say that C ⊆ U is an homogeneous set if for

every pair of tuples a, c ∈ C <ω such that a ≡ c and for every b ∈ C there is a d ∈ C
such that a, b ≡ c, d.

12.26 Lemma Suppose L is countable. If T is ω-categorical and every finite set is contained in a
finite homogeneous set, then T has the finite model property.
Proof We prove fmp also for formulas with parameters. We prove that for all n there
is a finite structure A ⊆ U where fmp holds for all sentences ϕ ∈ L( A) such that
# number of parameters in ϕ + number of quantifiers in ϕ ≤ n.
Fix n and pick some finite set A that is homogeneous and such that all types p(z) ⊆
L with |z| ≤ n have a realization in A|z| . Now we prove fmp by induction on the
syntax of ϕ.
The claim for atomic formulas is witnessed by any finite structure that contains
the parameters of the formula. Such finite substructure exists in fact it suffices
to take the algebraic closure which, in an ω-categorical theory, is finite (by 3 of
Lemma 12.13). Induction for Boolean connectives is straightforward. As for in-
duction step for the existential quantifier, consider the formula ∃ x ϕ( x ; c), where
c ∈ A<n and | x | = 1. Implication ⇐ of fmp follows immediately from the induction
hypothesis and from the fact that, if ∃ x ϕ( x ; c) satisfy #, also ϕ(d ; c) satisfies it. As
for ⇒, assume U  ∃ x ϕ( x ; c). Let a, b ∈ A<ω be a solution of ϕ( x ; z) such that
a ≡ c. Such a solution exists because all types with ≤ n variables are realized in A.
By homogeneity there is a d ∈ A such that a, b ≡ c, d and therefore A  ϕ(d ; c). 

12.27 Proposition If T is strongly minimal, then every algebraically closed set is homogeneous.
Proof Let A be argebraically closed and fix some a, c ∈ A<ω such that a ≡ c. Let b
be an element of A. Suppose first the case b ∈ acl a. Fix an f ∈ Aut (U) such that

96
 f ( a) = c. Then d = f b is the required element, in fact a, b ≡ c, d and d ∈ acl c ⊆ A.
Now, suppose instead that b ∈ / acl a. Then any d ∈
/ acl c satisfies a, b ≡ c, d. Such a d
exists in A otherwise A = acl c 6= acl a which contradicts a ≡ c. 

From the propositions above we finally obtain the following.

12.28 Theorem A theory which is ω-categorical and strongly minimal is not finitely axiomatiz-
able.
Proof If T is ω-categorical the algebraic closure of a finite set is finite. Therefore
from Proposition 12.27 we infer that T satisfies the assumptions of Lemma 12.26.
Hence T has the finite model property, so, by Proposition 12.25 it is not finitely
axiomatizable. 

12.29 Exercise Assume L is countable and let T be strongly minimal. Prove that the
following are equivalent
1. T is ω-categorical;

2. the algebraic closure of a finite set is finite.

Implication 1⇒2 does not require the strong minimality of T. 

6 Notes and references

An uncountable, non-isolated, complete type that cannot be omitted was produced
by Gebhard Fuhrken in 1962. Example 12.3 is inspired by a post of Alex Kruckman
to StackExhange [1]. I am not aware of other expositions of this (or other) example.
A famous theorem of Boris Zil’ber claims that Theorem 12.28 holds for any totally
categorical theory. The same theorem has been proved independently by Cherlin,
Harrington and Lachlan with a proof that uses the classification of finite simple
groups. This theorem signs the birth of a subject known as geometric stability theory
which studies in depth the geometric properties of model theories which we briefly
hinted to in Chapter 11. The interested reader may consult Pillay’s monograph [2].

[1] Alex Kruckman, Counterexample to the omitting types in uncountable language

(2017), available at https://math.stackexchange.com/q/2434851. URL
accessed 2017-11-19.
[2] Anand Pillay, Geometric stability theory, Oxford Logic Guides, vol. 32, 1996.

97
 Chapter 13
Definability and automorphisms
The description of first-order definability is simplified if we allow definable sets
to be used as second-order parameters in formulas. This leads to the theory of
(elimination of) imaginaries. The technical reason that induced Shalah to introduce
imaginaries will only be clear later, see Sections 16.5, but the theory is of indepen-
dent interest.
In this chapter we fix a signature L, a complete theory T without finite models,
and a saturated model U of inaccessible cardinality κ strictly larger than | L|. The
notation and implicit assumptions are as in Section 9.3.

1 Many-sorted structures
A many-sorted language consists of three disjoint sets. Besides the usual Lfun and
Lrel , we have a set Lsrt whose elements are called sorts . The language also includes
a (many-sorted) arity function that assigns to function and relation symbols r, f a
tuple of sorts of finite positive length which we call arity .
A many-sorted structure M consists of
1. a set Ms , for each s ∈ Lsrt ;

2. a function f M : Ms1 × · · · × Msn → Ms0 , for each f ∈ Lfun of arity hs0 , . . . , sn i;

3. a relation r M ⊆ Ms0 × · · · × Msn , for each r ∈ Lrel of arity hs0 , . . . , sn i.

For every sort s we fix a sufficiently large set of variables Vs . Now we define terms
and their respective sorts by induction.
All variables are terms of their respective sort. If t1 , . . . , tn are terms of sorts
s1 , . . . , sn and f ∈ Lfun is of arity hs0 , . . . , sn i then f t1 , . . . , tn is a term of sort s0 .
Formulas are defined as follows. If r ∈ Lrel has arity hs0 , . . . , sn i then r t0 , . . . , tn is
a formula. Also, t1 = t2 is formula for every pair of terms of equal sort. All other
formulas are constructed by induction using the propositional connectives ¬ and ∨
and the quantifier ∃ x (or any other reasonable choice of logical connectives).
Truth of formulas is defined as for one-sorted languages, except that here we require
that the witness of the quantifier ∃ x belongs to Ms , where s is the sort of the variable
x.
Models of second-order logic are arguably the most widely used examples of many-
sorted structures. They may be described using a language with a sort n for every
n ∈ ω. The sort 0 is used for the first-order elements; the sort n > 0 is used for
relations of arity n. For every n > 0 the language has a relation symbol ∈n of arity
h0n , ni, where
0n = 0, . . . , 0 .
| {z }
n-times

98
 There are also arbitrarily many function and relation symbols of sort h0n i for any
n > 0.

2 The eq-expansion
B Warning: the structure Ueq and the theory T eq defined below do not coincide with
the standard ones introduced by Shelah. As the difference is merely cosmetic, in-
troducing new notation would be overkill and we prefer to abuse the existing ter-
minology. In Section 6 below we compare our definition with the standard one.
Given a language L we define a many-sorted language Leq which has a sort for
each partitioned formula σ ( x ; z) ∈ L and a sort 0 which we call the home sort .
(Partitioned formulas have been introduced in Definition 1.14.)
For better legibility, we will pretend that all formulas σ depend on the same pair of
tuples x ; z, so we assume that x ; z are infinite tuples.
The home sort is also called the first-order sort . All other sorts are second-order .
First of all, Leq contains all relations and functions of L. The many-sorted arity of
a relation r is h0nr i, where nr is the arity of r in L. Similarly, the many-sorted arity
of a function f is h01+n f i, where n f is the arity of f in L. Moreover, Leq contains
a relation symbol ∈σ( x ;z) for each sort σ ( x ; z). These relation symbols have arity
0| x| , σ ( x ; z) . As there is no risk of ambiguity, in what follows we write ∈ without
subscript.
The model Ueq has U as domain for the home sort. The domain for the sort
σ ( x ; z) contains the definable sets A = σ(U ; b) for some b ∈ U|z| . The symbols in
L have the same interpretation as in the one-sorted case, and ∈ is interpreted as set
membership. We write T eq for Th(Ueq ).
As usual Leq also denotes the set of formulas constructed in this language and,
if A ⊆ Ueq , we write Leq ( A) for the language and the set of formulas that use
elements of A as parameters. We write L( A) for the set of formulas in Leq ( A) that
contain no second-order variables (neither free nor quantified).
We use the symbol X to denote a generic second-order variable.
It is important to note right away that this expansion of U is a mild one: the de-
finable subsets of the home sort of Ueq are the same as those of U. In particular,
iterating the expansion would not yield anything new.

13.1 Proposition Let X = X1 , . . . , Xn be a tuple of second order variables of sort σi ( x ; z) for

i = 1, . . . , n. Let ϕ(u ; X) ∈ Leq where u is a tuple of first-order variables. Then there is a
formula ϕ0 (u ; z) ∈ L such that
 n 
0
∀X, z Xi = x : σi ( x ; z) → ∀u ϕ(u ; X) ↔ ϕ (u ; z)
^   
i =1

eq
When n = 0 the proposition asserts that Lu and Lu have the same expressive power.
Proof By induction on syntax. When ϕ is atomic, we set ϕ0 = ϕ unless it is of the
form t ∈ X for some tuple of terms t or it has the form X1 = X2 . In the first case ϕ0
 
is the formula σ (t ; z). In the second case it is the formula ∀ x σ1 ( x ; z) ↔ σ2 ( x ; z) .
The connectives stay unchanged except for the quantifiers ∃X, where X is a second-

99
 order variable, say of sort σ ( x ; z). These quantifiers are replaced by ∃z. 

Proposition 13.1 implies in particular that we can always eliminate second-order

quantifiers. This is an obvious fact that may be obscured by the notation: we can
replace ∃X by ∃z if we substitute σ (t ; z) for t ∈ X in the quantified formula.
Proposition 13.1 should convince the reader that the move from U to Ueq is almost
trivial.

13.2 Remark From Proposition 13.1 it follows that for every A ⊆ Ueq , there is a B ⊆ U
such that L( B) is at least as expressive as L( A). The point of Ueq is that there might
not be any B ⊆ U such that L( B) is exactly as expressive as Leq ( A). For instance,
suppose L contains only an equivalence relation with infinitely many classes, all
with infinite elements. Let A be an equivalence class. Then A is is definable in L( B)
if and only if B ∩ A 6= ∅. But no element of A is definable in Leq ({A}). 

If V  U we write Veq for the substructure of Ueq that has V as domain of the home
sort and the set of definable sets of the form σ (U ; b) for some b ∈ V|z| as domain of
the sort σ ( x ; z). The following proposition claims that the elementary substructures
of Ueq are exactly those of the form Veq for some V  U.

13.3 Proposition The following are equivalent for every structure V† of signature Leq
1. V†  Ueq ;

2. V† = Veq for some V  U.

Proof Implication 2⇒1 is a direct consequence of Proposition 13.1. We prove 1⇒2.

Let V be the domain of the home sort of V† . It is clear that V  U. Let A ∈ Veq have
sort σ ( x ; z), say A = σ (U ; b) for some b ∈ V|z| . As ∃=1 X ∀ x x ∈ X ↔ σ( x ; b) holds
 

in Ueq , by elementarity it holds in V† and therefore A ∈ V† . This proves Veq ⊆ V† .

A similar argument proves the converse inclusion. 

13.4 Proposition Let A ⊆ Ueq . Then every type p(u ; X) ⊆ Leq ( A) finitely consistent in Ueq
is realized in Ueq . That is, Ueq is saturated.
Proof By Remark 13.2, there are some B ⊆ U and some q(u ; X) ⊆ Leq ( B) equiv-
alent to p(u ; X). This already proves the proposition when X is the empty tuple.
Otherwise, let q0 (u ; z) be obtained replacing every formula ϕ(u ; X) in q(u ; X) with
the formula ϕ0 (u ; z) given in Proposition 13.1. Then q0 (u ; z) is finitely consistent
in U. Assume for clarity of notation that X is a single variable of sort σ ( x ; z). If
c ; b  q0 (u ; z), then c ; σ (U ; b)  q(u ; X). 

Automorphisms of a many-sorted structure are defined in the obvious way: sorts

are preserved as well as functions and relations. Every automorphisms f : U → U
extends to an automorphism f : Ueq → Ueq as follows. If A = σ (U ; b) we define
f A = σ (U ; f b) = f A , which clearly preserves the sort and the relation ∈. Clearly,
 

this extension is unique.

The homogeneity of Ueq follows by back-and-forth as in the one-sorted case.

13.5 Proposition Every elementary map k : Ueq → Ueq of cardinality < κ extends to an
automorphism of Ueq . 

100
 3 The definable closure in the eq-expansion
We may safely identify automorphism of U with automorphisms of Ueq . Let A ⊆
Ueq and let a be a tuple of elements of Ueq . We denote by Aut (U/A) the set of
automorphisms (of Ueq ) that fix all elements of A. The symbol O( a/A) denotes
the orbit of a over A . This has been defined in Section 9.2 and now we apply it to
Ueq
O( a/A) = f a : f ∈ Aut (U/A) .


By homogeneity, O( a/A) = p(Ueq ) where p(v) = tp( a/A). When O( a/A) = { a}

we say that a is invariant over A or A-invariant , for short.

13.6 Definition Let A ⊆ Ueq and a ∈ Ueq . When ϕ( a) ∧ ∃=1 v ϕ(v) holds for some formula
ϕ(v) ∈ Leq ( A), we say that a is definable over A . We write dcleq ( A) for the set of those
a ∈ Ueq that are definable over A. We write dcl( A) for dcleq ( A) ∩ U. This is the natural
generalization of the notion of definability introduced in Section 11.1. 

The saturation and homogeneity of Ueq allow us to prove the following proposition
with virtually the same proof as for Theorem 11.4

13.7 Theorem For any A ⊆ Ueq and a ∈ Ueq the following are equivalent
1. a is invariant over A;

2. a ∈ dcleq ( A). 

Note that a tuple is invariant if and only it all its component are definable. Then in
Theorem 13.7 we may replace a and/or A by a tuple.
Definition 13.6 treats first- and second-order elements of Ueq uniformly. However,
for sets it can be rephrased in a more natural way.

13.8 Proposition Let A ⊆ Ueq and let A ∈ Ueq have sort σ( x ; z). Then the following are
equivalent
1. A is definable over A;

2. A = ψ(U) for some ψ( x ) ∈ L( A).

Proof Implication 2⇒1 is clear because extensionality is implicit in the definition

of Ueq . We prove 1⇒2. Let ϕ(X) ∈ Leq ( A) be a formula A is the unique solution of.
Then the required formula ψ( x ) is ∃X x ∈ X ∧ ϕ(X) .
 


By the characterization above, Theorem 13.7 when applied to a definable set A gives
an alternative proof of Proposition 9.19.
The formula ψ( x ) in Proposition 13.8 need not be related to the sort σ ( x ; z). For ex-
ample, consider the theory of a binary equivalence relation e( x ; z) with two infinite
classes, let A be one of these classes and let A 6= ∅ be such that A ∩ A = ∅. Then A
is definable over A though not by some formula of the form e( x ; b) for some b ∈ A.
Things change radically if we replace A with a model.

13.9 Proposition Let M be a model and let A be an element of sort σ ( x ; z). Then the following
are equivalent

101
 1. A ∈ dcleq ( M );

2. A = σ (U ; b) for some b ∈ M|z| .

In particular Meq = dcleq ( M).

Proof Assume 1 and let ψ( x ) ∈ L( M ) be such that A = ψ(U). Such a formula exists
by Proposition 13.8. Then ∃z ∀ x ψ( x ) ↔ σ ( x ; z) holds in Ueq . By elementarity it
 

holds in Meq , therefore ∃z has a witness in M. This proves 1⇒2, the converse
implication is obvious. 

4 The algebraic closure in the eq-expansion

Let A ⊆ Ueq and a ∈ Ueq . We say that a is algebraic over A if ϕ( a) ∧ ∃=k v ϕ(v)
holds for some formula ϕ(v) ∈ Leq ( A) and some positive integer k. We write
acleq ( A) for the set of those a ∈ Ueq that are algebraic over A. We write acl( A)
for acleq ( A) ∩ U. This is the natural generalization of the notion introduced in
Section 11.1.
The following proposition is proved with virtually the same proof of Theorem 11.5

13.10 Theorem For every A ⊆ Ueq and every a ∈ Ueq the following are equivalent
1. O( a/A) is finite;

2. a ∈ acleq ( A);

3. a ∈ Meq for every model such that A ⊆ Meq . 

We say finite equivalence relation for an equivalence relation with finitely many
classes. A finite equivalence formula or type is a formula, respectively a type,
that defines a finite equivalence relation. Theorem 13.11 belows proves that sets
algebraic over A are union of classes of a finite equivalence relations definable over
A.

13.11 Theorem Let A ⊆ Ueq and let A ∈ Ueq be an element of sort σ ( x ; z). Then the following
are equivalent
1. A ∈ acleq ( A)

2. for some finite equivalence formula ε( x ; y) ∈ L( A) and some c1 , . . . , cn ∈ U|y|

n
x∈A ↔
_
ε ( x ; c i ).
i =1
 
m
Proof 2⇒1 If ε( x ; y) has m classes, then O(A/A) contains at most sets.
n
1⇒2 Let ϕ(X) ∈ Leq ( A) be an algebraic formula that has A among its solutions
and define
 
ε ( x ; y ) = ∀X ϕ (X) → x ∈ X ↔ y ∈ X
 

If ϕ(X) has n, solutions, then ε( x ; y) has at most 2n equivalence classes. Clearly, A

is union of some these classes. 

102
 Sh
13.12 Definition We write a ≡ A b when ε( a ; b) holds for every finite equivalence formula
ε( x ; y) ∈ L( A). In words we say that a and b have the same Shelah strong-type over A. 

By the following proposition, the Shelah strong type of a over A is tp( a/ acleq A).

13.13 Proposition Let A ⊆ Ueq and let a, b ∈ U| x| . Then the following are equivalent
Sh
1. a ≡ A b;

2. a ≡acleq A b.

Proof 2⇒1 Assume ¬1 and let ε( x ; y) ∈ L( A) be a finite equivalence formula

such that ¬ε( a ; b). Let D = ε(U ; b), then b ∈ D and a ∈ / D. As ε( x ; y) is an
A-invariant finite equivalence formula, D ∈ acleq ( A), and ¬2 follows.
1⇒2 Assume ¬2 and let ϕ( x ) ∈ L acleq ( A) be such that ϕ( a) ↔


/ ϕ(b). Let
eq
D = ϕ(U), then D ∈ acl ( A). Therefore, by Proposition 13.11, the set D is union
of equivalence classes of some finite equivalence formula ε( x ; y) ∈ L( A). Then
¬ε( a ; b) and ¬1 follows. 

We may write S( a/A) be the intersection of all definable sets that contain a and
Sh
are algebraic over A. Then by the proposition above S( a/A) = b : b ≡ A a .


13.14 Exercise Let p( x ) ⊆ L( A) and let ϕ( x ; y) ∈ L( A) be a formula that defines, when

restricted to p(U), an equivalence relation with finitely many classes. Prove that
there is a finite equivalence relation definable over A that coincides with ϕ( x ; y) on
p (U). 

13.15 Exercise Let A ⊆ U, let A be a definable set with finite orbit over A. Without using
the eq-expansion, prove that A is union of classes of a finite equivalence relation
definable over A. 

13.16 Exercise Let T be strongly minimal and let ϕ( x ; z) ∈ L( A) with | x | = 1. For

arbitrary b ∈ U|z| , prove that if the orbit of ϕ(U ; b) over A is finite, then ϕ(U ; b) is
definable over acl A.
Hint: you can use Theorem 13.11. 

5 Elimination of imaginaries
For the time being, we agree that imaginary is just another word for definable set.
Though this is not formally correct (cfr. Section 6), it is morally true and helps to
understand the terminology. The concept of elimination of imaginaries has been
introduced by Poizat who also proved Theorem 13.23 below. A theory has elimina-
tion of imaginaries if for every A ⊆ Ueq , there is a B ⊆ U such that L( B) and L( A)
have the same expressive power (i.e. they are the same up to to equivalence).

13.17 Definition We say that T has elimination of imaginaries if for every definable set A there
is a formula ϕ( x ; z) ∈ L such that
 
=1
ei ∃ z ∀ x x ∈ A ↔ ϕ( x ; z)

103
 We say that the witness of ∃=1 z in the formula above is a canonical parameter of A. It is
also called a a canonical name for A. A set may have different canonical parameters for
different formulas ϕ( x ; z).
We say that T has weak elimination of imaginaries if
 
wei ∃=k z ∀ x x ∈ A ↔ ϕ ( x ; z )

for some positive integer k. 

In the formulas above we allow z to be the empty string. In this case we read ei and
wei omitting the quantifiers ∃=1 z, respectively ∃=k z. Therefore ∅-definabile sets
have all (at least) the empty string as a canonical parameter.
To show that the notions above are well-defined properties of a theory one needs to
check that they are independent of our choice of monster model. We leave this to
the reader as an exercise.

13.18 Exercise Let M be an arbitrary model. Prove that ei and wei hold (in Ueq ) if and
only if they hold in Meq . 

We say that two tuples a and b of elements of Ueq are interdefinable if dcleq ( a) =
dcleq (b). By Theorem 13.7 this is equivalent to saying that Aut (U/a) = Aut (U/b),
that is, the automorphisms that fix a fix also b, and vice versa.

13.19 Theorem The following are equivalent

1. T has weak elimination of imaginaries;

2. every definable set is interdefinable with a finite set;

3. every definable set A is definable over acl {A} .

Proof 1⇒2 Assume 1 and let B be the set of solutions of the formula
 
∀ x x ∈ A ↔ σ( x ; z) .

Hence B is finite and B ∈ dcleq {A}. We also have A ∈ dcleq {B} because A is
definibile by the formula ∃z z ∈ B ∧ σ ( x ; z) . Therefore dcleq {A} = dcleq {B}.
 

2⇒3 Assume dcleq {A} = dcleq {B} for some finite set B. The elements of B,
say b1 , . . . , bn , are the (finitely many) solutions of the formula z ∈ B. Therefore
b1 , . . . , bn ∈ acl(B) = acl(A). Let ϕ(X ; B) be a formula that has A as unique solu-
tion. Then ∃Y Y = b1 , . . . , bn ∧ ϕ(X ; Y) is the formula that proves 3.
  

3⇒1 Assume 3. As wei holds trivially for all ∅-definable sets, we may assume
A 6= ∅. Let σ ( x ; z) be such that
 
∀ x x ∈ A ↔ σ( x ; b) .

for some tuple b of elements of acl {A} . Fix some algebraic formula δ(z ; A)

satisfied by b and write ψ(z ; X) for the formula

 
∀ x x ∈ X ↔ σ ( x ; z ) ∧ δ ( z ; X).

The following formula is clearly satisfied by b therefore, if we can prove that it has

104
 finitely many solutions, wei follows from Proposition 13.1
 
] ∀ x x ∈ A ↔ σ ( x ; z ) ∧ ∃X ψ ( z ; X) .

We check that any c that satisfies ] also satisfies δ(z ; A). As A is non empty,
ψ(c ; A0 ) holds for some A0 . By ] and the definition of ψ(z ; X) we obtain A0 = A
and δ(c ; A). 

13.20 Theorem The following are equivalent

1. T has elimination of imaginaries;

2. every definable set is interdefinable with a tuple of real elements;

3. every definable set A is definable over dcl {A} .

Proof Implications 1⇒2⇒3 are immediate. Implication 3⇒1is identical to the ho-
mologous implication in Theorem 13.19, just substitute algebraic with definable. 

We now consider elimination of imaginaries in two concrete structures: algebraically

closed fields and real closed fields. The following lemma is required in the proof of
Theorem 13.22 below.

13.21 Lemma The following is a sufficient condition for weak elimination of imaginaries
] for every A ⊆ Ueq , every consistent ϕ(z) ∈ L( A) has a solution in acl A.
Proof Let A ∈ Ueq be a definable set of sort σ ( x ; z). Then ∀ x x ∈ A ↔ σ ( x ; z)
 

is consistent and, by ] it it has a solution in acl {A} . Hence weak elimination

follows from Theorem 13.19. 

13.22 Theorem Let T be a complete, strongly minimal theory. Then, if acl ∅ is infinite, T has
weak elimination of imaginaries.
Proof If acl ∅ is infinite, acl A is a model for every A (cfr. Exercise 11.25) so condi-
tion ] of lemma 13.21 holds by elementarity and the theorem follows. 

p
13.23 Theorem The theories Tacf have elimination of imaginaries.
p
Proof By Theorem 13.22 we know that Tacf has weak elimination of imaginaries.
Therefore, by Theorem 13.19 it suffices to prove that every finite set A is interde-
finable with a tuple. Let A = a1 , . . . , an where each ai is a tuple ai,1 , . . . , ai,m of


elements of U. Given A we define the term

n  m 
tA ( x ; y) = ∏ x − ∑ ai,k yk . where y = y1 , . . . , ym .
i =1 k =1
Note that (the interpretation of) the term tA ( x ; y) is independent on particular in-
dexing of the set A. So, any automorphism that fixes A, fixes the tA ( x ; y). Now
rewrite tA ( x ; y) as a sum of monomials and let c be the tuple of coefficients of
these monomials. The tuple c uniquely determines tA ( x ; y) and vice versa. There-
fore every automorphism that fixes A fixes c and vice versa. Hence A and c are
interdefinable. 

13.24 Exercise Let T have elimination of imaginaries and ϕ( x ; z) ∈ L( A). For arbitrary
c ∈ U|z| , prove that if the orbit of ϕ(U ; c) over A is finite, then ϕ(U ; c) is definable

105
 over acl A. 

6 Imaginaries: the true story

The point of the expansion to Ueq is to add a canonical parameter for each definable
set. In fact, in Ueq every definable subset of U|z| is the canonical parameter if itself.
This allows to deal with theories without elimination of imaginaries in the most
straightforward way.
The expansion to Ueq that was originally introduced by Shelah (and still used ev-
erywhere else) is slightly different from the one introduced here. For a given set
A = σ (U ; b) Shelah considers the equivalence relation defined by the formula
 
ε(z ; z0 ) = ∀ x σ ( x ; z) ↔ σ ( x ; z0 ) .
def

The equivalence class of b in the relation ε(z ; z0 ) is what Shelah uses as canonical
parameter of the set A.
Shelah’s Ueq has a sort for each ∅-definable equivalence relation ε(z ; z0 ). The do-
main of the sort ε(z ; z0 ) contains the classes of the equivalence relation defined by
ε(z ; z0 ). These equivalence classes are called imaginaries . Shelah’s Leq contains
functions that map tuples in the home sort to their equivalence class.
Tortuous routes are somewhat typical of Shelah. This did not hinder him from
becoming one of the most productive mathematician of his century. Lesser minds
may find it useful to take a simplified route.

7 Uniform elimination of imaginaries

Sometimes in the literature elimination of imaginaries is confused with uniform
elimination of imaginaries, see e.g. [1, Definition 8.4.2]. Theorem 13.26 below
shows that the difference is immaterial. This section is more technical and could be
skipped at a first reading.
Let us rephrase the definition of elimination of imaginaries: for every formula
ϕ( x ; u) ∈ L and every tuple c ∈ U|u| there is a formula σ ( x ; z) ∈ L such that
 
∃ =1 z ∀ x ϕ ( x ; c ) ↔ σ ( x ; z ) .

A priori, the formula σ ( x ; z) may depend on c in a very wild manner. We say that
T has uniform elimination of imaginaries if for every ϕ( x ; u) there are a formula
σ ( x ; z) and a formula ρ(z) such that
 
=1
 
uei ∀u ∃ z ρ(z) ∧ ∀ x ϕ( x ; u) ↔ σ( x ; z) .

The role of the formula ρ(z) above is mysterious. It is clarified by the following
propositions. In fact, uniform elimination of imaginaries is equivalent to a very
natural property which in words says: every definable equivalence relation is the
kernel of a definable function. Recall that the kernel of the function f is the relation
f a = f b.
Uniform elimination of imaginaries is convenient when dealing with interpretations

106
 of structure inside other structures. Let us consider a simple concrete example.
Suppose U is a group and let H be a definable normal subgroup of U. The elements
of the quotient structure U/H are equivalent classes of a definable equivalence
relation. If there is uniform elimination of imaginaries we can identify U/H with
an actual definable subset of G (the range of the function f above). Moreover,
the group operation of G are definable functions. As working in U/H may be
notationally cumbersome, G may offer a convenient alternative.

13.25 Proposition The following are equivalent

1. T has uniform elimination of imaginaries;

2. for every ϕ( x ; u) such that ∀u ∃ x ϕ( x ; u) there is a formula σ ( x ; z) such that

 
∀ u ∃ =1 z ∀ x ϕ ( x ; u ) ↔ σ ( x ; z ) ;

3. for every equivalence formula ε( x ; u) ∈ L there is given by σ ( x ; z) ∈ L such that

 
∀ u ∃ =1 z ∀ x ε ( x ; u ) ↔ σ ( x ; z ) .

Note that the definable function f u = z mentioned above is defined by the formula
 
ϑ (u ; z) = ∀ x ε( x ; u) ↔ σ( x ; z) .

Proof 1⇒2. If ∀u ∃ x ϕ( x ; u) we can rewrite uei as

 
=1
∀u ∃ z ∀ x ϕ( x ; u) ↔ ρ(z) ∧ σ( x ; z) .

2⇒3. Clear.
3⇒1. Apply 3 to the equivalence formula
 
ε(u ; v) = ∀ x ϕ( x ; u) ↔ ϕ( x ; v)

Let ϑ (u ; z) be defined as above. The reader may check that uei holds substituting for
 
δ(z) the formula ∃u ϑ (u ; z) and and for σ ( x ; z) the formula ∃u ϕ( x ; u) ∧ ϑ (u ; z) . 

By the following theorem, uniformity comes almost for free.

13.26 Theorem The following are equivalent

1. T has uniform elimination of imaginaries;

2. dcl ∅ contains at least two elements and T has elimination of imaginaries.

Proof 1⇒2. Let ϕ( x, u) be the formula u1 = u2 . From 1 we obtain a formula

σ ( x ; z) ∈ L such that
 
=1
 
∀ u1 , u2 ∃ z ρ ( z ) ∧ ∀ x u1 = u2 ↔ σ ( x ; z ) .

Therefore the formulas ∃=1 z ρ(z) ∧ ∀ x σ ( x ; z) and ∃=1 z ρ(z) ∧ ∀ x ¬σ ( x ; z) are

   

both true. The witnesses of ∃=1 z in these two formulas are two distinct elements of
dcl ∅.
2⇒1. Assume 2 and fix a formula formula ϕ( x ; u) such that ϕ( x, a) is consistent
for every a ∈ U|u| . We prove 2 of Proposition 13.25.

107
 Let p(u) be the type that contains the formulas
 
=1
¬ ∃ z ∀ x ϕ( x ; u) ↔ σ( x; z) ,

where σ ( x; z) ranges over all formulas in L. By elimination of imaginaries p(u) is

not consistent. Therefore, by compactness, there are some formulas σi ( x, z) such
that
n  
=1
_
] ∀u ∃ z ∀ x ϕ( x ; u) ↔ σi ( x ; z) .
i =0
To prove the theorem we need to move the disjunction in front of the σi ( x ; z).
We can assume that if σi ( x ; b) ↔ σi0 ( x ; b0 ) for some hb, i i 6= hb0 , i0 i then σi ( x ; b) is
inconsistent. Otherwise we can substitute the formula σi ( x ; z) with
 
^
σi ( x ; z) ∧ ¬∃y 6= b ∀ x σj ( x ; y) ↔ σi ( x ; z) .
j ≤i

As ϕ( x, a) is consistent for every a, the substitution does not break the validity of ].
Fix some distinct ∅-definable tuples d0 , . . . , dn of the same length (these are easy to
obtain from two ∅-definable elements). We claim that from ] it follows that
" #
n

∀u ∃=1 z, y ∀ x ϕ( x ; u) ↔
_
σi ( x ; z) ∧ y = di .
i =1
(The tuple z, y plays the role of z.) We fix some a and check that the formula below
has a unique solution
" #
 n
_
[ ∀ x ϕ( x ; a) ↔ σi ( x ; z) ∧ y = di .
i =1
Existence follows immediately from ]. As for uniqueness, note that if b, di and b0 , di0
are two distinct solution of [ then σi ( x ; b) ↔ σi0 ( x ; b0 ) for some hb, i i 6= hb0 , i0 i. By
what assumed on σi ( x ; z), we obtain that ϕ( x ; a) is inconsistent. A contradiction
which proves the theorem. 

8 Notes and references

[1] Katrin Tent and Martin Ziegler, A course in model theory, Lecture Notes in Logic,
vol. 40, Association for Symbolic Logic, Cambridge University Press, 2012.

108
 Chapter 14
Invariant sets
In this chapter we fix a signature L, a complete theory T without finite models, and
a saturated model U of inaccessible cardinality κ strictly larger than | L|. Notation
and implicit assumptions are as presented in Section 9.3.

1 Invariant sets and types

Let D ⊆ U|z| , where z is a tuple of length < κ. We say that D is an A-invariant set
if it is fixed (set-wise) by A-automorphisms. That is, f [D] = D for every automor-
phism f ∈ Aut (U/A) or, yet in other words,
is1. a∈D ↔ fa ∈ D for every a ∈ U|z| and every f ∈ Aut (U/A),
which, by homogeneity is equivalent to,
is2. a∈D ↔ b∈D for all a, b ∈ U|z| such that a ≡ A b.
This yields the following bound on the number of invariant sets

14.1 Proposition Let λ = | Lz ( A)|. There are at most 22 sets D ⊆ U|z| that are invariant
λ

over A.
Proof By is2, sets that are invariant over A are union of equivalence classes of the
relation ≡ A , that is, union of sets of the form p(U) where p(z) ∈ S( A). Then the
number of A-invariant sets is 2|Sz ( A)| . Clearly |Sz ( A)| ≤ 2λ . 

We say that D is an invariant set if it is invariant over some (small) set A. As we

require κ to be inaccessible, there are exactly κ invariant sets.
In this chapter we work with ∆-types where ∆ may be either L(A) or the set of
Boolean combinations of ϕ( x ; b) for b ∈ A|z| and some fixed ϕ( x ; z) ∈ L. In the
latter case ∆-types are called a ϕ-types . We denote by S ϕ (A) the set of complete
ϕ-types with parameters in A. Typically, A is either the whole of U or some small
set A ⊆ U. Types in S ϕ (U) are called global ϕ-types .
Let p( x ) ⊆ L(U) be a consistent type. For every formula ϕ( x ; z) ∈ L we define
n o
D p,ϕ = a ∈ U| z | : ϕ ( x ; a ) ∈ p
We can read the notation in two ways. Either the tuple z has infinite length and is
the same for all formulas, or it is finite and depends on ϕ. This is possible because
adding or erasing dummy variables to the second tuple of ϕ( x ; z) does not change
D p,ϕ in any relevant way, in particular invariance is preserved.
Let p( x ) ⊆ L(U) be a consistent type. We say that p( x ) is an A-invariant type , if
for every formula ϕ( x ; z) ∈ L,
it1. ϕ( x ; a) ∈ p ⇔ ϕ( x ; f a) ∈ p for every a ∈ U|z| and every f ∈ Aut (U/A).
By invariant type we mean invariant over some (small) set A.

109
 A global ϕ-type can be identified with D p,ϕ and a global type can be identified with
the collection of all these sets. The notions of invariance for types and sets coincide.
We say that the type p( x ) ⊆ L(U) does not split over A if
 
it2. a ≡A b ⇒ ϕ( x ; a) ∈ p ⇔ ϕ( x ; b) ∈ p for all a, b ∈ U|z|

for every formula ϕ( x ; z) ∈ L. For global type (or a global ϕ-types when ϕ is fixed)
it2 is equivalent to

it2’. a ≡ A b ⇒ ϕ( x ; a) ↔ ϕ( x ; b) ∈ p
By homogeneity, non splitting is equivalent to invariance.
The following is yet another important equivalent of invariance over A that easily
follows from it2’
it3. a ≡ A b ⇒ a ≡ A,c b for all a, b ∈ U|z| and for all c  p A,a,b .
Note that it3 applies to global types p( x ) ∈ S(U) but not to ϕ-types.

2 Invariance from the dual perspective

B The following terminology in not standard. We say that the set B ⊆ U| x| , typically
a definable set, is quasi-invariant over A if for any finitely many automorphisms
f 1 , . . . , f n ∈ Aut (U/A) the sets f i [B] have non-empty intersection.
We say that the type p( x ) ⊆ L(U), typically a global type, is quasi-invariant over A
if ϕ(U) is quasi-invariant over A for every ϕ( x ) conjunction of formulas in p( x ). For
types of small cardinality, this is equivalent to requiring that p(U) is quasi-invariant.
For global types quasi-invariance coincides with invariance.

14.2 Proposition Let p( x ) ∈ S ϕ (U) be a global ϕ-type, then the following are equivalent
1. p( x ) is invariant over A;

2. p( x ) is quasi-invariant over A.

Proof 1⇒2. Assume p( x ) is invariant and let ψ( x ; b) ∈ p and some b ∈ U|z| . Then
ψ( x ; f b) ∈ p for every f ∈ Aut (U/A), so 2 follows from the finite consistency of
p ( x ).
2⇒1. Assume p( x ) is not invariant. Then there is b ∈ U|z| such that ϕ( x ; b) ∈ p and
/ p for some f ∈ Aut (U/A). By completeness, p( x ) contains the formula
ϕ( x ; f b) ∈
ϕ( x ; b) ∧ ¬ ϕ( x ; f b) which clearly is not consistent with its f -translate. 

14.3 Exercise We say that B strongly quasi-invariant if for every definable set D at least
one of B ∩ D and B ∩ ¬D is quasi-invariant. Strongly quasi invariant types are
defined in analogy to the definition of quasi-invariant type. Note parenthetically
that for global types the two notions coincide. Prove that every strongly quasi-
invariant type has an extension to a global invariant type.
Hint: it may help to prove that if B is strongly quasi-invariant then for every de-
finable D either B ∩ D or B ∩ ¬D is strongly quasi-invariant. Then the maximal
strongly quasi-invariant set is the required global extension. 

110
14.4 Exercise Let p( x ) ∈ L(U). Let V be a saturated elementary extension of U. Prove
that the following are equivalent:
1. the type p( x ) is quasi-invariant over A;

2. the set p(V) is quasi-invariant over A (defined as above with V for U). 

The easiest way to obtain quasi-invariant types is via types that are finitely satisfi-
able. We say that a type p( x ) is finitely satisfiable in A if ϕ(U) ∩ A| x| 6= ∅ for every
ϕ( x ) that is a conjunction of formulas in p( x ).
Note that by elementarity every type over a model is finitely satisfiable in the model,
that is, if p( x ) ⊆ L( M) for some model M, then p( x ) is finitely satisfiable in M. For
this reason the notion is mainly applied to models.

14.5 Proposition Let p( x ) ∈ S ϕ (U) be a global ϕ-type finitely satisfiable in A. Then p( x ) is

A-invariant.
Proof By Proposition 14.2, because finitely satisfiable types are quasi-invariant. 

14.6 Proposition Every type q( x ) ⊆ L(U) finitely satisfiable in A has an extension to a global
type finitely satisfiable in A.
Proof Let p( x ) ⊆ L(U) be maximal among the types containing q( x ) and finitely
satisfiable in A. We prove that p( x ) is complete. If for a contradiction that p( x )
 
contains neither ψ( x ) nor ¬ψ( x ). Then neither p( x ) ∪ ψ( x ) nor p( x ) ∪ ¬ψ( x )
is finitely satisfiable in A. This contradicts the finite satisfiability of p( x ). 

A global type p( x ) ∈ S(U) may be called a global coheir of p M ( x ) if it is finitely

satisfiable in M.

3 Morley sequences and indiscernibles

In the following α is some ordinal ≤ κ, typically ω, and x is a tuple of variables
of length < κ. Let p( x ) ∈ S(U) be a global type invariant over A. We say that
c̄ = hci : i < αi is a Morley sequence of p( x ) over A if for every i < α
Ms. ci  p A,ci ( x )
In particular, when p( x ) is finitely satisfiable in some model M, we say that c̄ is
a coheirs sequence . We may say that c̄ is a Morley (or coheir) sequence without
specifying a global type. Then we mean that c̄ is a Morley (respectively coheir)
sequence of some suitable global type.
The following is a convenient characterization of coheir sequences.

14.7 Lemma The following are equivalent

1. c̄ = hcn : n < αi is a coheir sequence over A;

2. cn+1 ≡ A,cn cn and ϕ(U) ∩ A| x| 6= ∅ whenever ϕ( x ) ∈ L( A, cn ) and ϕ(cn ).

Proof 1⇒2. Assume 1 and let p( x ) ∈ S(U) be a global type finitely satisfiable in A
and such that ci  p A,ci ( x ). The requirement cn+1 ≡ A,cn cn is clear. Now, suppose
ϕ(cn ) for some ϕ( x ) ∈ L( A, cn ). Then ϕ( x ) belongs to p( x ) hence ϕ(U) ∩ A| x| 6= ∅
because p( x ) is finitely satisfiable in A.

111
 2⇒1. Assume 2 and define

q( x ) = ϕ( x ) : cn  ϕ( x ) ∈ L( A, cn ) for some n < α .
Clearly c̄ is a Morley sequence of any global type finitely satisfiable in A that ex-
tends q( x ). Hence it suffices to show that such type exists. So, we prove that q( x )
is finitely satisfiable in A and apply Proposition 14.6.
Note that 2 requires that every single formula in q( x ) is finitely satisfiable in A.
Then it suffices to show that q( x ) is closed under conjunctions. Let ϕi ( x ) ∈ L( A, cni ),
for i = 1, 2, be two formulas in q( x ) satisfied by cni . By the first requirement in 2,
both formulas are satisfied by cm where m = max{n1 , n2 }. Hence their conjuction
is in q( x ). 

Let I, < I be a linear order. We call a function ā : I → U| x| an I-sequence , or

simply a sequence when I is clear. Often we introduce an I-sequence by writing
ā = h ai : i ∈ I i.
If I0 ⊆ I we call a I0 a subsequence of ā. Most relevant are the subsets I0 ⊆ I that
are well-ordered by < I and, in particular, the finite ones. When I0 has order-type α,
an ordinal, we identify a I0 with a tuple of length α.

14.8 Definition Let I, < I be an infinite linear order and let ā be an I-sequence. We say that a
is a sequence of indiscernibles over A or, a sequence of A-indiscernibles , if a I0 ≡ A a I1
for every I0 , I1 ∈ (nI ) and n < ω. As usual, (nI ) is the set of subsets of I of cardinality n. 

The condition above can be formulated in a number of clearly equivalent ways. We

can require that that ϕ( ai0 , . . . , ain ) ↔ ϕ( a j0 , . . . , a jn ) for every formula ϕ( x1 , . . . , xn ) ∈
L( A) and every pairs of tuples in I n such that i0 < · · · < in and j0 < · · · < jn . Or we
can simply say that for every i0 , . . . , in ∈ I the type tp( ai0 , . . . , ain /A) only depends
on the order type of i0 , . . . , in .

14.9 Proposition Let p( x ) ∈ S(U) be a global A-invariant type and let c̄ = hci : i < αi be a
Morley sequence of p( x ) over A. Then c̄ is a sequence of indiscernibles over A.
Proof We prove by induction on n < ω that
] cn ≡ A c I0 for every I0 ⊆ α of cardinality n.
For n = 0 the claim is trivial, then we assume ] above is true and prove that
cn , cn ≡ A c I0 , ci for every I0 < i < α.
As c is Morley sequence, cn ≡ A,cn ci whenever n < i. Hence we can equivalently
prove that
cn , ci ≡ A c I0 , ci ,
which is equivalent to
cn ≡ A, ci c I0 .
The latter holds by induction hypothesis ] and the invariance of p( x ) as formulated
in it3 of Section 14. 

112
 4 From coheirs to Ramsey to indiscernibles
Let X be an arbitrary infinite set an n, k < ω. A total function
 
X
f : →k
n
is also called a coloring of ( Xn ), the set of the subsets of X of cardinality n. This is
identified with a partition of ( Xn ) into finitely k classes. A set H ⊆ X such that f
is constant on ( H
n ) is called monochromatic set . In the literature it may be called
a homogeneous set , but we already use this word with a different meaning. The
following is a most famous theorem which here we prove it in unusual way.

14.10 Ramsey’s Theorem For every infinite set M, every n, k < ω and every coloring f :
(M
n ) → k there is an infinite monochromatic set.
Proof Let L be a language that contains k relation symbols r0 , . . . , rk−1 of arity n.
Given f : ( Mn ) → k as above we define now a structure with domain M. The
interpretation the relation symbols is
n o
riM =


a 0 , . . . , a n −1 ∈ M : f a 0 , . . . , a n −1 = i .
Then
 
^ _
# M  ∀ x 0 , . . . , x n −1  xi 6 = x j → r i ( x 0 , . . . , x n −1 )  .
0≤ i < j < n i <k

We may assume that M is an elementary substructure of some large saturated model

U. Pick any type p( x ) ∈ S(U) finitely satisfied in M but not realized in M and let
c̄ = hci : i < ω i be a coheir sequence of p( x ). By # and indiscernibility, all tuples of
n distinct elements of c have the same color. The theorem is proved if we can find
in M a sequence ā = h ai : i < ω i with the same property. (Note parenthetically that
no element of c̄ is in M.) We construct ai by induction on i as follows.
Assume as induction hypothesis that the subsequences of length n of ai , cn are
monochromatic. Our goal is to find ai such as the same property holds for ai , ai , cn .
By the indiscernibility of c̄, the property holds for ai , cn , cn . And this can be written
by a formula ϕ( ai , cn , cn ). As c̄ is a coheir sequence, by Lemma 14.7 we can find
ai ∈ M such that ϕ( ai , cn , ai ). So, as the order is irrelevant, ai , ai , cn satisfies the
induction hypothesis. 

Let I, < I be an infinite linear order and let ā be an I-sequence. Fix a sequence of
distinct variables x̄ = h xi : i < ω i. We write p( x̄ ) = EM-tp( ā/A) and say that
p( x̄ ) is the Ehrenfeucht-Mostowski type of ā over A if p( x̄ ) is the set of formulas
ϕ( xn ) ∈ L( A) such that ϕ( a I0 ) holds for every I0 ∈ (nI ).
Note that x̄ is always of order-type ω, while ā is an arbitrary infinite sequence.
Clearly, xi and ai are tuple of the same sort.
Note also that if ā is A-indiscernible then EM-tp( ā/A) is a complete type, and vice
versa. So, if ā and c̄ are two A-indiscernible I-sequences with the same Ehrenfeucht-
Mostowski type over A, then ā ≡ A c̄.

14.11 Ehrenfeucht-Mostowski’s Theorem Let I, < I and J, < J be two infinite linear orders
such that | J | ≤ κ. Then for every sequence ā = h ai : i ∈ I i there is an J-sequence of

113
 A-indiscernibles c̄ such that EM-tp( ā/A) ⊆ EM-tp(c̄/A).
Proof We prove the theorem for I = J = ω and leave the general case to the reader.
Let c̄ be any realization of the following type yields a J-sequence of A-indiscernibles.
 
n ω o
q( x̄ ) ∪ ϕ( x I0 ) ↔ ϕ( x J0 ) : ϕ( xn ) ∈ L( A), I0 , J0 ∈ , n<ω .
n
We will prove that any finite subset of the type above is realized by a finite subse-
quence of ā. First note that any finite subset of q( x̄ ) is realized by any subsequence
of ā of the proper length by the definition of EM-type.
Then we only need to pay attention to the set on the right. We prove that for m and
n arbitrary large and every ϕ0 , . . . , ϕm−1 ∈ L xn ( A) there is an infinite H ⊆ I such
that a H realizes
 
n ω o
# ϕi ( x J0 ) ↔ ϕi ( x I0 ) : I0 , J0 ∈ and i < m .
n
Let f : ( n ) → P(m) map I0 ∈ ( n ) to the set i < m : ϕi ( a I0 ) . By the Ramsey
ω ω 

theorem with 2m colors, there is some infinite monochromatic set H ⊆ I. As I has

order type ω, so does H, and it is immediate that a H realizes #, as required to prove
the theorem with I = J = ω. 

14.12 Proposition Let ā = h ai : i ∈ I i be a sequence of A-indiscernibles. Then ā is indiscernible

over some model M containing A.
Proof Fix an arbitrary model M containing A. By Theorem 14.11 there is an
I-sequence of M-indiscernibles c̄ such that EM-tp( ā/M) ⊆ EM-tp(c̄/M ). As ā is an
A-indiscernible sequence ā ≡ A c̄. Therefore h c̄ = ā for some some h ∈ Aut (U/A).
Hence ā is indiscernible over h[ M ]. 

14.13 Exercise Let ā = h ai : i ∈ I i be an A-indiscernible sequence and let I ⊆ J with

| J | ≤ κ. Then there is an A-indiscernible sequence c̄ = hci : i ∈ J i such that c I = ā. 

14.14 Exercise Let p( x ) ∈ S(U) be a global type invariant over A. Let a and b two
realizations of p A ( x ). Prove that there is a sequence c̄ = hci : i < ω i such that a, c̄
and b, c̄ are both sequences of A-indiscernibles. 

114
 Chapter 15
Lascar invariant sets
In this chapter we fix a signature L, a complete theory T without finite models,
and a saturated model U of inaccessible cardinality κ strictly larger than | L|. The
notation and implicit assumptions are as in Section 9.3.

1 Expansions
This section is only marginally required in the present chapter so it can be post-
poned with minor consequences.
We will find it convenient to expand the language L with a predicate for a given
D ⊆ U|z| . We denote by hU, Di the corresponding expansion of U. Generally, we
write L(X) for the expanded language but, if the intended interpretation of X is
only going to be D, we may write L(D) and abbreviate hU ; Di  ϕ(X) as ϕ(D).

15.1 Remark The definitions above are straightforward when z finite tuple. When z
is an infinite tuple the intuition stays the same but a more involved definition is
required. In fact, first-order logic does not allow infinitary predicates. We think of
L(X) for a two sorted language. The home sort, denoted by 0, and the z-sort, denoted
by z. The expansion hU, Di has domain U for the home sort, and U|z| for the z-sort.
Besides the symbols of L, there is a function symbol πi for every i < |z| which is
interpreted as the projection to the i-coordinate. These functions have arity hz, 0i
(see Section 13.1 for the notation). There is also a predicate of sort hzi interpreted
as D. 

What said above is adapted to define the expansion L(Xi : i < λ) , where Xi are
predicates of arity |zi |. Again, when the sets Di ⊆ U|zi | are the only intended
interpretation of Xi , we may write L(Di : i < λ) .
If C, D ⊆ U|z| we abbreviate hU, Ci ≡ A hU, Di as C ≡ A D. We also say that D is
saturated if so is the model hU ; Di.

15.2 Remark For every D ⊆ U|z| there is a saturated C ≡ D. In fact, it suffices to find
a saturated model hU0 , D0 i ≡ hU, Di of cardinality κ. By saturation, there is an
isomorphism f : U0 → U. Therefore f [D0 ] is the required set C. 

15.3 Exercise Prove that if D ⊆ U|z| is saturated and invariant over A than it is definable
over A. 

15.4 Proposition If D ⊆ U|z| is invariant over A then every A-indiscernible sequence is in-
discernible in the language L( A ; D).
Proof Let c̄ = hci : i ∈ I i be an A-indiscernible sequence. For every I0 , I1 ∈ I [n]
there is an f ∈ Aut (U/A) such that f c I0 = c I1 , Hence

115
 ϕ c I0 ; D ↔ ϕ f c I0 ; f [D] ↔ ϕ c I1 ; D .






15.5 Exercise Prove that if C ⊆ U|z| is type-definable over B then ≡ A,B implies ≡ A ;C . 

2 Lascar strong types

Let D ⊆ U|z| , where z is a tuple of length < κ. The orbit of D over A is the set
n o
o (D/A) = f [D] : f ∈ Aut (U/A)
def

So, D is invariant over A when o (D/A) = D . We say that D is Lascar invariant



over A if it is invariant over every model M ⊇ A. Recall that this means that if
a ≡ M c for some model M containing A then a ∈ D ↔ c ∈ D.

15.6 Proposition Let λ = | Lz ( A)|. There are at most 22 sets D ⊆ U|z| that are Lascar
λ

invariant over A.
Proof Let N be a model containing A of cardinality ≤ λ. Every set that is Lascar
invariant over A is invariant over N. As | Lz ( N )| = λ the bound follows from
Proposition 14.1. 

15.7 Theorem For every D and every A ⊆ M the following are equivalent
1. D is Lascar invariant over A;

2. every set in o (D/A) is M-invariant;

| L( A)|
3. o (D/A) has cardinality ≤ 22 ;
4. o (D/A) has cardinality < κ;
5. c0 ∈ D ↔ c1 ∈ D for every A-indiscernible sequence hci : i < ω i.

Proof 1⇒2. This implication is clear because all sets in o (D/A) are Lascar invari-
ant over A.
2 ⇒ 3. When | M | ≤ | L( A)| the implication follows from the bounds discussed in
Section 14.1. We temporary add this assumption on M. Once the proof of the
proposition is completed, is easily seen to be redundant.
3 ⇒ 4. This implication holds because κ is a strong limit cardinal.
4 ⇒ 5. Assume ¬5. Then we can find an A-indiscernible sequence hci : i < κ i such
that c0 ∈ D ↔
/ c1 ∈ D. Define
E(u ; v) ⇔ u ∈ C ↔ v ∈ C for every C ∈ o (D/A).
Then E(u ; v) is an A-invariant equivalence relation. As ¬ E(c0 ; c1 ), by Proposi-
tion 15.4, indiscernibility over A implies that ¬ E(ci , c j ) for every i < j < κ. Then
E(u ; v) has κ equivalence classes. As κ is inaccessible, this implies ¬4.
5⇒1. Fix any a ≡ N b where A ⊆ N. It suffices to prove that a ∈ D ↔ b ∈ D.
Let p(z) ∈ S(U) be a global coheir of tp( a/N ) = tp(b/N ). Let c̄ = hci : i < ω i be
a Morley sequence of p(z) over N, a, b. Then both a, c̄ and b, c̄ are A-indiscernible
sequences. Therefore, from 5 we obtain a ∈ D ↔ c0 ∈ D ↔ b ∈ D. 

116
 For definable sets Lascar invariance reduces to definablity over the algebraic closure.

15.8 Corollary For every definable set D the following are equivalent
1. D is Lascar invariant over A;

2. D is definable over acleq ( A). 

The following corollary easily follows from Theorem 15.7 and Proposition 15.4. As
an exercise the reader may wish to prove it using Proposition 14.12.

15.9 Corollary The following are equivalent

1. D is Lascar invariant over A;

2. every sequence of A-indiscernibles is L( A ; D)-indiscernible. 

B Given a tuple a ∈ U|z| , we write L( a/A) for the intersection of all sets containing
a that are Lascar invariant over A. Clearly L( a/A) is Lascar invariant over A. The
symbol L( a/A) is not standard.
L
15.10 Definition We write a ≡ A b and say that a and b have the same Lascar strong type over
A if the equivalence a ∈ D ↔ b ∈ D holds for every set D that is Lascar invariant over A
or, in other words, if L( a/A) = L(b/A). The notation L-stp( a/A) = L-stp(b/A) and
aEL/A b are also common in the the literature. 

L
15.11 Proposition The relation ≡ A is the finest equivalence relation with < κ classes that is
invariant over A. 
L
Proof Clearly ≡ A is an equivalence relation invariant over A. Each equivalence
class is Lascar invariant over A, hence the number of equivalence classes is bounded
L
by the number of Lascar invariant sets over A. To see that ≡ A is the finest of such
equivalences. Suppose D is an equivalence class of an A-invariant equivalence
relation with < κ classes. Then o (D/A) has also cardinality < κ. Then D is Lascar
L
invariant and as such it is union of classes of the relation ≡ A . 

Let p( x ) ∈ S(U) be global type; we say that p is Lascar invariant over A if the
sets D p,ϕ for ϕ( x ; z) ∈ L all Lascar invariant over A. (The sets D p,ϕ are defined in
Section 14.1.)

15.12 Proposition Let p( x ) ∈ S(U) be a global type. Then the following are equivalent
1. p( x ) is Lascar invariant over A;

2. every A-indiscernible sequence c̄ = hci : i < ω i is A, a-indiscernible for every a  p A,c̄ ;

3. every A-indiscernible sequence c̄ = hci : i < ω i is indiscernible over any a  pc̄ .

For convenience all tuples ci have length |z| = ω.

Proof We prove 1⇔3. Equivalence 1⇔2 can be proved similarly because the Las-
car invariance of p( x ) easily implies the Lascar invariance of the sets D p,ϕ for all
ϕ ( x ; z ) ∈ L ( A ).
3⇒1 If p( x ) is not Lascar invariant over A then c0 ∈ D p,ϕ ↔/ c1 ∈ D p,ϕ for some
A-indiscernible sequence c = hci : i < ω i and some ϕ( x ; z) ∈ L. Then p( x ) contains

117
 the formula ϕ( x ; c0 ) ↔
/ ϕ( x ; c1 ). Hence, c̄ is not indiscernible over any realization
of p( x )c0 ,c1
1⇒3 Assume 1 and fix an A-indiscernible sequence c̄ = hci : i < ω i and some
a  pc̄ . We need to prove that for every formula ϕ( x ; z0 ) ∈ L, where z0 = z1 , . . . , zn ,
ϕ ( a ; c 0 , . . . , c n −1 ) ↔ ϕ ( a ; c i 1 , . . . , c i n −1 ) .
holds for every i0 < · · · < in−1 < ω. Suppose not and let m be any integer larger
than in−1 . Then the following equivalences cannot both be true
ϕ ( a ; c m , . . . , c m + n −1 ) ↔ ϕ ( a ; c 0 , . . . , c n −1 );
ϕ ( a ; c m , . . . , c m + n −1 ) ↔ ϕ ( a ; c i 0 , . . . , c i n −1 ) .
If the first is false, define c0k = ckm , . . . , ckm+n−1 for all k < ω. Otherwise, do this only
for positive k and set c00 = ci0 , . . . , cin−1 . In either cases hc0k : k < ω i is a sequence of
A-indiscernibles and c00 ∈ D p,ϕ ↔ / c10 ∈ D p,ϕ . This contradicts 1. 

L
15.13 Exercise Prove that the equivalence relation a ≡ A b is the transitive closure of the
relation: there is a sequence hci : i < ω i indiscernible over A such that c0 = a and
c1 = b. Hint: apply Theorem 15.7. 

3 The Lascar graph and Newelski’s theorem

Here we study Lascar strong types from a different viewpoint. The Lascar graph
over A has an arc between all pairs a, b ∈ U|z| such that a ≡ M b for some model
M containing A. We write d A ( a, b) for the distance between a and b in the Lascar
graph over A. Let us spell this out: d A ( a, b) ≤ n if there is a sequence a0 , . . . , an
such that a = a0 , b = an , and ai ≡ Mi ai+1 for some models Mi containing A. We
write d A ( a, b) < ∞ if a and b are in the same connected component of the Lascar
graph over A.

15.14 Proposition For every a ∈ U|z|

L( a/A) = c : d A ( a, c) < ∞ .


Proof To prove inclusion ⊇ it suffices to show that every Lascar A-invariant set
containing a contains the set on the r.h.s. Let D be Lascar A-invariant, and let
b ∈ D. Then D contains also every c such that b ≡ M c for some model M containing
A. That is, D contains every c such that d A (b, c) ≤ 1. It follows that D contains
every c such that d A ( a, c) < ∞.
To prove inclusion ⊆ we prove the set on the r.h.s. is Lascar A-invariant. Suppose
the sequence a0 , . . . , an , where a0 = a and an = c, witnesses d A ( a, c) ≤ n and
suppose that c ≡ M b for some M containing A, then the sequence a0 , . . . , an , b
witnesses d A ( a, b) ≤ n + 1. 

We write Autf (U/A) for the subgroup of Aut (U/A) that is generated by the auto-
morphisms that fix point-wise some model M containing A. (The “f” in the symbol
stands for fort, the French word for strong.) It is easy to verify that Autf (U/A) is a
normal subgroup of Aut (U/A). The following is a corollary of Proposition 15.14.

15.15 Corollary The following are equivalent

118
 L
1. a ≡ A b

2. a = f b for some f ∈ Autf (U/A). 

It may not be immediately obvious that the relation d A (z, y) ≤ n is type-definable.

15.16 Proposition For every n < ω there is a type pn (z, y) ⊆ L( A) equivalent to d A (z, y) ≤ n.
Proof It suffices to prove the proposition with n = 1. Let λ = | L( A)| and fix a tuple
of distinct variables w = hwi : i < λi, then p1 (z, y) = ∃w p(w, z, y) where
n o
p(w, z, y) = q(w) ∪ ϕ(z, w) ↔ ϕ(y, w) : ϕ(z, w) ∈ L( A)

and q(w) ⊆ L( A) is a consistent type with the property that all its realizations
enumerate a model containing A.
It remains to verify that such a type exist. Let hψi ( x, wi ) : i < λi be an enumeration
of the formulas in L x,w ( A), where x is a single variable. Let
n o
q(w) = ∃ x ψi ( x, wi ) → ψi (wi , wi ) : i < λ .
Any realization of q(w) satisfy the Tarski-Vaught test therefore it enumerates a
model containing A. Vice versa it is clear that we can realize q(w) in any model
containing A. 

We conclude this section with a theorem of Ludomir Newelski.

The following notions apply generally to any group G acting on some set X and
to any set D ⊆ X. Below we always have G = Autf (U/A) and X = U|z| . We say
that D is drifting if for every finitely many f 1 , . . . , f n ∈ G there is a g ∈ G such
that g[D] is disjoint from all the f i [D]. We say that D is quasi-invariant if for every
finitely many f 1 , . . . , f n ∈ G the sets f i [D] have non-empty intersection. We say that
a formula or a type is drifting or quasi-invariant if the set it defines is.
The union of drifting sets need not be drifting. However, by the following lemma it
cannot be quasi-invariant.

15.17 Lemma The union of finitely many drifting sets in not quasi-invariant.
Proof It is convenient to prove an apparently more general claim. If D1 , . . . , Dn are
all drifting and L is such that for some finite F ⊆ G

L f [ D1 ∪ · · · ∪ D n ] ,
[
] ⊆
f ∈F

then L is not quasi-invariant. The claim is vacuously true for n = 0. Now, assume
n is positive and that the claim holds for n − 1. Define C = D1 ∪ · · · ∪ Dn−1 and
rewrite ] as follows
L f [C] ∪ h [Dn ]
[ [
⊆
f ∈F h∈ F

Since Dn is drifting there is a g ∈ G such that g[Dn ] is disjoint from h[Dn ] for every
h ∈ F, which implies that

L ∩ g [Dn ] f [C].
[
⊆
f ∈F

Hence for every h there holds

119
 hg−1 [L] ∩ h[Dn ] hg−1 f [C]
[
⊆
f ∈F

So, from ] we obtain

L∩ hg−1 [L] f [C] ∪ hg−1 f [C].

\ [ [ [
⊆
h∈ F f ∈F h∈ F f ∈ F

By the induction hypothesis, the set on the r.h.s. is not quasi-invariant. Hence
neither is L, proving the claim and with it the lemma. 

The following is a consequence of Baire’s category theorem. We sketch a proof for

the convenience of the reader.

15.18 Lemma Let p( x ) ⊆ L( B) and pn ( x ) ⊆ L( A), for n < ω, be consistent types such that
_
1. p( x ) → pn ( x )
n<ω
Then there is an n < ω and a formula ϕ( x ) ∈ L( A) consistent with p( x ) such that

2. p( x ) ∧ ϕ( x ) → pn ( x )
Proof Negate 2 and choose inductively for every n < ω a formula ψn ( x ) ∈ pn ( x )
such that p( x ) ∧ ¬ψ0 ( x ) ∧ · · · ∧ ¬ψn ( x ) is consistent. By compactness, this contra-
dicts 1. 

Finally we can prove Newelski’s theorem on the diameter of Lascar types.

15.19 Theorem (Newelski) For every a ∈ U|z| the following are equivalent
1. L( a/A) is type-definable;

2. L( a/A) =

c : d A ( a, c) < n for some n < ω.
Proof Implications 2⇒1 holds by Proposition 15.16. We prove 1⇒2. Suppose
L( a/A) is type-definable, say by the type l (z). Let p(z, y) be some consistent type
(to be defined below) such that and p(z, y) → l (z) ∧ l (y). Then, in particular
_
p(z, y) → d A (z, y) < n.
n<ω
By Proposition 15.16 and Lemma 15.18, there is some n < ω and some ϕ(z, y) ∈
L( A) consistent with p(z, y) such that

]1 p(z, y) ∧ ϕ(z, y) → d A (z, y) < n.

Below we define p(z, y) so that for every ψ(z, y) ∈ L( A)

]2 p(z, a) ∧ ψ(z, a) is non-drifting whenever it is consistent.

Drifting and quasi-invariance are relative to the action of Autf (U/A) on U|z| . Then,
in particular, p(z, a) ∧ ϕ(z, a) is non-drifting and the theorem follows. In fact, by
non-drifting, there are some a0 , . . . , ak ∈ L( a/A) such that every set p(U, c) ∩ ϕ(U, c)
for c ∈ L( a/A) intersects some p(U, ai ) ∩ ϕ(U, ai ). Let m be such that d A ( ai , a j ) ≤ m
for every i, j ≤ k. From ]1 we obtain that d A ( a, c) ≤ m + 2n. As c ∈ L( a/A) is
arbitrary, the theorem follows.
The required type p(z, y) is union of a chain of types pα (z, y) defined as follows

120
 p0 (z, y) = l (z) ∪ l (y);
n o
]3 pα+1 (z, y) = pα (z, y) ∪ ¬ψ(z, y) : pα (z, a) ∧ ψ(z, a) is drifting ;
[
pα (z, y) = pn (z, y) for limit α.
n<α
Clearly, the chain stabilizes at some stage ≤ | L( A)| yielding a type which satisfies
]2 . So we only need to prove consistency. We prove that pα (z, a) is quasi-invariant
(so, in particular, consistent). Suppose that pn (z, a) is quasi-invariant for every
n < α but, for a contradiction, pα (z, a) is not. Then for some f 1 , . . . , f k ∈ Autf (U/A)
k
[
pα (z, a) ∪ pα (z, f i a)
i =1
is inconsistent. By compactness there is some n < α and some ψi (z, y) as in ]3 such
that
m ^
^ k
pn (z, a) → ¬ ¬ψj (z, f i a)
j =1 i =1

As pn (z, a) is quasi-invariant, from Lemma 15.17 we obtain that pn (z, f i a) ∧ ψj (z, f i a)

is non-drifting for some i, j. Clearly we can replace f i a with a, then this contradicts
the construction of pα (z, y) and proves the theorem. 

15.20 Exercise Let L be quasi-invariant and let D be drifting, prove that L r D is quasi-
invariant. 

4 Kim-Pillay types
Given a tuple a ∈ U|z| , we write K( a/A) for the intersection of all type-definable
sets containing a that are Lascar invariant over A. Or, more concisely, the intersec-
tion of all sets that are type-definable over a model containing A. We call K( a/A)
the Kim-Pillay strong type over A. Clearly K( a/A) is Lascar invariant over A. It
also easy to see that K( a/A) is type-definable. In fact, by invariance, we can assume
that all types in the intersection above are over M, for any fixed model containing
A. Hence K( a/A) is the minimal type-definable set containing a and closed under
L
the relation ≡ A . It follows that if b ∈ K( a/A) then K(b/A) ⊆ K( a/A).
To summarize, we recall that we have defined a whole hierarchy of strong types
obtained from the intersection of different sets with various sots of invariance
L( a/A) ⊆ K( a/A) ⊆ S( a/A) ⊆ O( a/A).
Recall that S( a/A) was defined after Proposition 13.13.
If K( a/A) = K(b/A), we say that a and b have the same Kim-Pillay strong type
KP KP
over A. We abbreviate this by a ≡ A b . In other words, we write a ≡ A b when
a ∈ D ↔ b ∈ D for every type-definable set D that is Lascar invariant over A.
KP
B Warning: the symbol K( a/A) is not standard. The symbol a ≡ A b is not unusual,
but some author write KP-stp( a/A) = KP-stp(b/A) or a EKP/A b.

15.21 Proposition Fix some a ∈ U|z| and some A ⊆ U. Then there is a type e(z ; w) ⊆ L( A)

121
 such that K(b/A) = e(U ; b) for all b ∈ O( a/A) and e(z ; w) defines an equivalence
relation on O( a/A).
Proof Notice that K( a/A) is type-definable over A, a. In fact, if f ∈ Aut (U/A, a)
and D is a set containing a that is type-definable and Lascar invariant over A, then
so is f [D]. Therefore K( a/A) is invariant over A, a. As K( a/A) is type-definable,
invariance implies that it is type-definable over A, a. Let e(z ; w) ⊆ L( A) be such
that K( a/A) = e(U ; a).
We prove that K(b/A) = e(U ; b) for all b ∈ O( a/A). Let f ∈ Aut (U/A) be such
that f a = b. If D is a type-definable Lascar invariant over A, then so is f [D].
Therefore, f is a bijection between type-definable sets that are Lascar invariant over
A and contain a and analogous sets containing b. Then f [K( a/A)] = K(b/A) and
K(b/A) = e(U ; b) follows.
We prove that e(b ; U) ∩ O( a/A) is Lascar invariant over A. Let f ∈ Autf (U/A) and
c ∈ O( a/A). Then e(b ; f c) is equivalent to e( f −1 b ; c) which in turn is equivalent to
e(b ; c), by the invariance of e(U ; c).
Finally we are ready to prove that e(z ; w) defines a symmetric relation on O( a/A).
From what proved above, e(b ; U) ∩ O( a/A) is a type-definable Lascar invariant set
containing b and therefore it contains K(b/A). We conclude that for all b, c ∈
O( a/A)
e(c ; b) ↔ c ∈ e(U ; b) ↔ c ∈ K(b/A) → e(b ; c)
Reflexivity is clear; we prove transitivity. As remarked above, K(b/A) ⊆ K(c/A),
for all b ∈ K(c/A) or equivalently e(b ; c). Hence if e(b ; c) then
e(d ; b) → d ∈ K(b/A) ↔ d ∈ K(c/A) → e(d ; c).
Which completes the proof. 

15.22 Corollary For every a, b ∈ U|z| and A ⊆ U the following are equivalent
1. a ∈ K(b/A);

2. b ∈ K( a/A);

3. K( a/A) = K(b/A). 

The following useful lemma is the key ingredient in the proof of Theorem 15.24.

15.23 Lemma Let p(z) ⊆ L( A) and let e(z ; w) ⊆ L( A) define a bounded equivalence relation
on p(U). Then there is a type e0 (z ; w) ⊆ L( A) which defines a bounded equivalence relation
(on U) and refines e(z ; w) on p(U).
Proof Let C = hCi : i < λi enumerate the partition of p(U) induced by e(z ; w).
Note that each Ci is type-definable over A, a for any a ∈ Ci . If x = h xi : i < λi, we
write x ∈ C for the type that is the conjunction of xi ∈ Ci for i < λ.
We claim that the required type is
e0 ( a ; b) = ∃ x 0 ∈ C ∃ x 00 ∈ C a, x 0 ≡ A b, x 00
def

As the type above is invariant over A, we can assume that e0 ( a ; b) is type-definable

over A. Moreover, it is clearly a reflexive and symmetric relation, so we only check
it is transitive. Suppose a, x 0 ≡ A b, x 00 and b, y0 ≡ A c, y00 for some x 0 , x 00 , y0 , y00 ∈ C.

122
 Let z be such that a, x 0 , z ≡ A b, x 00 , y0 then z ∈ C and a, z ≡ A c, y00 .
The relation e0 ( a ; b) clearly refines the equivalence defined by e(z ; w) when re-
stricted to to p(U). To prove that it is bounded fix some c ∈ C and note that e0 ( a ; b)
is refined by a ≡ A,c b, which is bounded. 

Finally, we have the following.

15.24 Theorem Denote by e A (z ; w) be the finest bounded equivalence relation on U|z| that is
type-definable over A. Then for every a, b ∈ U|z| the following are equivalent
KP
1. a ≡ b

2. e A ( a ; b). 
KP
Proof As a ≡ b is equivalent to a ∈ K(b/A), it suffices to prove that e A (U ; b) =
K(b/A) for every b ∈ U|z| .
⊆ The orbit of e A (U ; b) under Aut (U/A) has cardinality < κ. Hence, by The-
orem 15.7, it is Lascar invariant over A. As K(b/A) is the least of such sets,
K(b/A) ⊆ e A (U ; b).
⊇ Let e(z ; w) be the type-definable equivalence relation given by Proposition 15.21.
The orbit of K(b/A) under Aut (U/A) has cardinality < κ. Hence the equivalence
relation that e(z ; w) defines on O(b/A) is bounded. By Lemma 15.23 there is be
a type-definable bounded equivalence relation e0 (z ; w) ⊆ L( A) that refines e(z ; w)
on O(b/A). As e0 (z ; w) is refined by e A (z ; w), we obtain e A (U ; b) ⊆ e(U ; b) =
K(b/A). 

The logic A-topology , or simply the logic topology when A is empty, is the topol-
ogy on U|z| whose closed sets are the type-definable Lascar A-invariant sets. It is
KP
clearly a compact topology. Its Kolmogorov quotient, that is U|z| / ≡, is Hausdorff.
KP
In fact, by Corollary 15.22, the open sets ¬K( a/A) and ¬K(b/A) separate b 6≡ a.

15.25 Exercise Prove that the clopen sets in the logic A-topology are exactly the sets that
are definable over acleq A. 

5 Notes and references

The original proof of Newelski theorem is rather long and complex. A simplified
proof appears in Rodrigo Peláez’s thesis [1, Section 3.3]. The proof here is a stream-
lined version the latter.
We are grateful to Volodya Shavrukov for fixing a wrong proof of Lemma 15.17.

[1] Rodrigo Peláez, About the Lascar group, PhD Thesis, Universitat de Barcelona,
Departament de Lógica, História i Filosofia de la Ciéncia, 2008.
[2] Pierre Simon, A guide to NIP theories, Lecture Notes in Logic, vol. 44, Association
for Symbolic Logic, Cambridge University Press, 2015.

123
 Chapter 16
Externally definable sets
In this chapter we fix a signature L, a complete theory T without finite models,
and a saturated model U of inaccessible cardinality κ strictly larger than | L|. The
notation and implicit assumptions are as in Section 9.3.

1 Approximable sets
Let C, D ⊆ U|z| . The set D ∩ A|z| is called the trace of D over A. We write C = A D
if C and D have the same trace on A.
We say that D is externally definable if it is of the form D p,ϕ for some global type
p ∈ Sx (U) and some formula ϕ( x ; z) ∈ L, see Section 14.1. We may say that D is
externally definable by p . Equivalently, a set D is externally definable if it is the
trace over U of a set which is definable in some elementary extension of U. Precisely,
D is the trace on U of a set of the form ϕ(b0 ; U0 ) where U0 is elementary extension
of U and b0 ∈ U0| x| . The latter interpretation explains the terminology.
B We prefer to deal with external definability in a different, though equivalent, way.
This is not the most common approach.

16.1 Definition We say that D is approximated by the formula ϕ( x ; z) if for every finite B
there is a b ∈ U| x| such that ϕ(b ; U) = B D. If in addition we have that ϕ(b ; U) ⊆ D, we
say that D is approximated from below . Similarly, when D ⊆ ϕ(b ; U) we say that D is
approximated from above . We may call ϕ( x ; z) the sort of D. 

The following proposition is clear by compactness.

16.2 Proposition For every D the following are equivalent

1. D is approximable;

2. D is externally definable. 

Approximability from below is an adaptation to our context of the notion of having

an honest definition in [1].

16.3 Definition We say that the global type p ∈ Sx (U) is honestly definable if for every
ϕ( x ; z) ∈ L the set D p,ϕ is approximated from below (by some formula). We say that p is a
definable if the sets D p,ϕ are all definable (over U). Note that the terminology is misleading:
honestly definable is a weaker than definable.

16.4 Example Every definable set is trivially approximable. Clearly, only by the right
formulas. Let T = Tdlo . Then D = {z ∈ U : a ≤ z ≤ b} is approximable both from
below and from above by the formula x1 < z < x2 though it is not definable by this
formula.

124
 Now, let T = Trg . Then every D ⊆ U is approximable and, when D has small infinite
cardinality, it is approximable from above but not from below, see Exercise 6.22. 

In Definition 16.1, the sort ϕ( x ; z) is fixed (otherwise any set would be approx-
imable) but this requirement of uniformity may be dropped if we allow B to have
larger cardinality.

16.5 Proposition For every D the following are equivalent

1. D is approximable;

2. for every C ⊆ U of cardinality ≤ | T | there is ψ(z) ∈ L(U) such that ψ(U) =C D.

Similarly, the following are equivalent:

3. D is approximable from below;

4. for every C ⊆ D of cardinality ≤ | T | there is ψ(z) ∈ L(U) such that C ⊆ ψ(U) ⊆ D.

Proof To prove 2⇒1 assume 2 and negate 1 for a contradiction. For each formula
ψ( x ; z) ∈ L choose a finite set B such that ψ(b ; U) 6= B D for every b ∈ U| x| . Let C be
the union of all these finite sets. Clearly |C | ≤ | T |. By 2 there are a formula ϕ( x ; z)
and a tuple c such that ϕ(c ; U) =C D, contradicting the definition of C.
Implication 1⇒2 is obtained by compactness and the equivalence 3⇔4 is proved
similarly. 

16.6 Remark Let C ⊆ U|z| . If D is approximated by ϕ( x ; z) then so is any C such that

C ≡ D, see Section 15.1 for the definition. In fact, if the set D is approximable by
ϕ( x ; z) then for every n
n 
ϕ ( x ; zi ) ↔ zi ∈ D .
^ 
∀ z1 , . . . , z n ∃ x
i =1
So the same holds for any C ≡ D. A similar remark apply to approximability
from below and from above. For approximability from below, add the conjunct
∀z ϕ( x ; z) → z ∈ D to the formula above. Similarly for approximability from
 

above. 

From the following easy observation of Chernikov and Simon [1] we obtain an
interesting (and misterious) quantifier elimination result originally due to Shelah
(see also Corollary 16.20 below).

16.7 Proposition Let C ⊆ U|y z| be approximated from below by the formula ϕ( x ; y z). Then
D = z : ∃y y z ∈ C is approximated from below by the formula ∃y ϕ( x ; y z).


Proof Let B ⊆ U be finite. We want a ∈ U| x| such that

∃y y b ∈ C ↔ ∃y ϕ( a ; y b) for every b ∈ B|z|


a.
h
i
b. ∀z ∃y ϕ( a ; y z) → ∃y y z ∈ C
Let D ⊆ U be a finite set such that
∃y ∈ D |y| y b ∈ C ↔ ∃ y y b ∈ C for every b ∈ B|z|



c.
As C is approximable from below, there is an a such that

|y z|
a’. d b ∈ C ↔ ϕ( a ; d b) for every d b ∈ D ∪ B

125
 h i
b’. ∀y z ϕ( a ; y z) → y z ∈ C
We obtain b from b’ simply by logic. Implication → in a follows from a’ and c.
Implication ← follows from b. 

16.8 Corollary If p ∈ Sx (U) is honestly definable then the family of sets externally definable
by p is closed under quantifiers and Boolean combinations.
Proof The sets externally definable by p are always closed under Boolean opera-
tions. By the proposition above, they are closed under quantifiers. 

2 Ladders and definability

Partitioned formulas have been introduced in Definition 1.14. Here we write ϕ( x ; z)∗
for the partitioned formula where the order of the two tuples is inverted (i.e. it is
the opposite of what displayed).
Let ϕ( x ; z) ∈ L(U) be a partitioned formula. We say that h ai ; bi : i < αi is a
ladder sequence for ϕ( x ; z) if
# i<j ⇔ ϕ ( ai ; b j ) for all 0 ≤ i, j ≤ α
We say that ϕ( x ; z) is stable if all ladders have length α < ω. Otherwise we say
it is unstable . Note that by compactness the ladder sequences of a stable formula
cannot have unbounded finite length.
It is evident that ϕ( x ; z) is stable if and only if ϕ( x ; z)∗ is stable.
Suppose ϕ( x ; z) be unstable as witnessed by h ai ; bi : i < ω i. By padding redundant
variables, we can read ϕ( x ; z) as a formula ϕ( x z ; y w). Then ϕ( x z ; y w) defines the
order of the sequence h ai bi : i < ω i. For this reason it is also common to say that
ϕ( x ; z) has the order property . It simply means that it is unstable.
The following is arguably one of the most important fact about stable formulas.

16.9 Theorem Any D ⊆ U|z| approximated by a stable formula is definable. Precisely, if

ϕ( x ; z) is a stable formula that approximates D then there are a1,1 , . . . , an,m ∈ U| x| such
that
n ^
_ m
z∈D ↔ ϕ( ai,j ; z)
i =1 j =1

Proof The theorem follows immediately from the the three lemmas below. 

16.10 Remark Theorem 16.9 is often stated in the following apparently more general
form. For every A ⊆ U there are a1,1 , . . . , an,m ∈ A| x| such that
n ^
m
for every z ∈ A|z|
_
z∈D ↔ ϕ( ai,j ; z)
i =1 j =1

Note that the equivalence above is restricted to A.

The proof is exactly the same, in fact elementarity and saturation are never used (to
a great extent, no model theory is used either, just finite combinatorics). 

126
16.11 Lemma If D is approximated from below by a stable formula ϕ( x ; z) then there are
a1 , . . . , an ∈ U| x| such that
n
_
z∈D ↔ ϕ ( ai ; z )
i =1

Proof Choose an ∈ U| x| such that b0 , . . . , bn−1 ∈ ϕ( an ; U) ⊆ D. Then, if possible,

choose bn such that
n
Dr ϕ ( a i ; U)
_
bn ∈
i =0
and iterate the procedure. By stability, the procedure has to stop at some n, yielding
the required a1 , . . . , an . 

16.12 Lemma If D is approximated by a stable formula ϕ( x ; z) with ladder order m. Then the
formula
m
^
ψ ( x1 , . . . , x m ; z ) = ϕ( x j ; z)
j =1

approximates D from below.

Proof Let B ⊆ D be finite. It suffices to prove that there are some a1 , . . . , am such
that B ⊆ ψ( a1 , . . . , am ; U) ⊆ D. Choose an ∈ U| x| such that
B ⊆ ϕ( an ; U) ⊆ U|z| r {b0 , . . . , bn−1 }.
Then, if possible, let bn such that
n
ϕ ( a i , U) r D
^
bn ∈
i =0
and iterate the procedure. The procedure has to stop at some n ≤ m. Hence the
required parameters are a1 , . . . , an = an+1 = · · · = am . 

16.13 Lemma If ϕ( x ; z) is a stable formula then for every m the formula

m
^
ψ ( x1 , . . . , x m ; z ) ↔ ϕ( x j ; z)
j =1

is stable.
Proof For legibility, we only prove that ϕ( x1 ; z) ∧ ϕ( x2 ; z) is stable. Suppose not
and let a1i , a2i ∈ U| x| and bi ∈ U|z| be such that
i<j ⇔ ϕ( a1i ; b j ) ∧ ϕ( a2i ; b j ) for all i, j < ω
For n = 1, 2 let Hn ⊆ (ω2 ) contain those pairs j < i such that ¬ ϕ( ain ; b j ). By the
equivalence above H1 ∪ H2 = (ω2 ). By the Ramsey Theorem there is an infinite set
H such that ( H2 ) ⊆ Hn for at least one of n = 1, 2. Suppose H1 for definiteness. So,
we obtain an infinite sequence a1i , bi such that
j < i ⇔ ¬ ϕ( a1i ; b j ) for all i, j < ω
which contradicts the stability of ϕ( x ; z). 

This last lemma concludes the proof of Theorem 16.9.

16.14 Exercise The following is a version of Harrinton’s mysterious Lemma (cfr. [1,

127
 Lemma 8.3.4]). Let ϕ( x ; z) ∈ L be a stable formula and suppose B ⊆ U|z| and
A ⊆ U| x| are approximated by ϕ( x ; z) and ϕ( x ; z)∗ , respectively. Then at least one
of the conditions 1 and 2 below occurs
1. a. ϕ( x ; z) ∧ x ∈ A approximates B and
b. ϕ( x ; z)∗ ∧ z ∈ B approximates A.
2. a. ϕ( x ; z) ∧ x ∈/ A approximates B and
b. ϕ( x ; z)∗ ∧ z ∈
/ B approximates A.
Hint: Note that at least one of 1a or 2a occurs. Hence it sufficies to prove that
1a⇒1b and 2a⇒2b. The two implications are essentially equivalent.

3 Vapknik-Chevronenkis dimension
We say that the formula ϕ( x ; z) ∈ L has Vapknik-Chevronenkis dimension n if this
is the largest finite cardinatity of a set A ⊆ U| x| such that |S ϕ ( A)| = 2n . If such n
does not exist, we say that we say that ϕ( x ; z) has infinite VC-dimension .
Note that the condition |S ϕ ( A)| = 2n is equivalent to saying that every subset of A
is the trace of some definable set of sort ϕ( x ; z).
For instance, the formula z1 < x < z2 in Tdlo has VC-dimension 1.
Arguing by compactness we obtain the following proposition whose proof is left as
an exercise for the reader.

16.15 Proposition The following are equivalent

1. ϕ( x ; z) ∈ L has finite VC-dimension;

2. there is no infinite set A ⊆ U| x| such that every subset of A is the trace of some definable
set of sort ϕ( x ; z). 

From the proposition above and Proposition 16.28 below it follows that all stable
formulas have finite VC-dimension.
We say that the a sequence of sentences h ϕi : i < ω i converges if the truth value of
ϕi is eventually constant.

16.16 Lemma The following are equivalent

1. ϕ( x ; z) ∈ L has finite VC-dimension;

2. h ϕ( ai ; b) : i < ω i converges for any b and any indiscernible sequence h ai : i < ω i.

Proof 1⇒2 Assume ¬2. It suffices to prove that for every I ⊆ ω the type that says
ϕ ( ai ; z ) ⇔ i ∈ I
is finitely consistent. Let ψ I n ( x0 , . . . , xn−1 ) the formula that says that there is a z
such that
ϕ ( xi ; z ) ⇔ i ∈ I for all i < n.
is finitely consistent we need to prove that ψ I n ( a0 , . . . , an−1 ) is true. If there is a b
such that the truth value of h ϕ( ai ; b) : i < ω i oscillates at least n times, then we can
find k0 < · · · < k n−1 such that

128
 ϕ( aki ; b) ⇔ i ∈ I.
This proves that ψ I n ( ak0 , . . . , akn−1 ) is true. By indiscernibility, also ψ I n ( a0 , . . . , an−1 )
holds true.
2⇒1 Let I ⊆ ω be the set of even integers. Assume ¬1 and let hci : i < ω i be an
infinite sequence that is shattered by ϕ( x ; z). Let h ai : i < ω i be an indiscernible
sequence with the same EM-type as hci : i < ω i. Then h ai : i < ω i satisfies
ψ I n ( x0 , . . . , xn−1 ) for all n. By compactness there is a b such that
ϕ ( ai ; z ) ⇔ i is even.
This proves ¬2. 

In the next section we need the following corollary.

16.17 Corollary If C ⊆ U| x| is a set approximable by a formula with finite VC-dimension, then

h ai ∈ C : i < ω i converges for any indiscernible sequence h ai : i < ω i. 

4 Honest definitions
In this section we present a beautiful theorem of Chernikov and Simon [1] and their
alternative proof of a famous quantifier elimination result of Shelah.
We write ¬n for ¬ . .n. .times /n .
. . . . . ¬. We abbreviate ¬n (· ∈ ·) as ∈

16.18 Lemma Let C be a set approximable by a formula with finite VC-dimension and let A be
any (small) set of parameters. Then every global A-invariant type p( x ) contains a formula
ψ( x ) such that either ψ( x ) → x ∈ C or ψ( x ) → x ∈
/ C. Moreover, for every sufficiently
saturated model M containing A, we can require that ψ( x ) ∈ L( M )
Proof The property we want to prove is elementary in L(C) therefore, by Re-
mark 15.2 we can assume that C is saturated. By Corollary 16.17, there is no infinite
sequence h ai : i < ω i
ai  / i C}
p A,ai ( x ) ∪ { x ∈
Let n be the largest integer such that h ai : i < ni satisfies the condition above. Then
/n C
p A,an ( x ) → x ∈
and the first claim of the lemma follows by compactness.
Now let M be a model that contains A and such that h M, Ci  hU, Ci. Assume
also that h M, Ci is saturated, we can further require that all ai ∈ M| x| . Hence
ψ ( x ) ∈ L ( M ). 

16.19 Theorem Let C be a set approximable by a formula with finite VC-dimension. Then C is
approximable from above and from below.
Proof Let A ⊆ C be any non empty (small) set. As above, by Remark 15.2 we can
assume that C is saturated. Let M be a saturated elementary substructure of U, in
the language L(C). Define
ϕ ( x ) ∈ L ( M ) : A ⊆ ϕ (U) .

q( x ) =
By compactness, it suffices to prove that q( x ) → x ∈ C. Suppose not and fix some
/ C. Let p( x ) = tp( a/M ). As p( x ) is finitely satisfied in A and
a  q( x ) ∧ x ∈

129
 M is saturated, from Lemma 16.18 yields a formula ψ( x ) ∈ p( x ) such that either
ψ( x ) → x ∈ C or ψ( x ) → x ∈/ C. But the first contradicts a  ψ( x ) and the second
cannot hold because ψ( x ) is satisfied in A ⊆ C. 

When all formulas have finite VC-dimension, we say that the theory T has the
non-independence property or, for short, that T is nip .
Let hDi : i < λi be the collection of all subsets of U, of arbitrary finite arity, that are
externally definable. The expansion of U to the language L(Xi : i < λ) is called the
Shelah expansion of U and is denoted by USh .
From Theorem 16.19 and Proposition 16.7 we obtain the following.

16.20 Corollary If T is nip then USh has L-elimination of quantifiers. (I.e. every formula is
Boolean combination of formulas in L and formulas of the form z ∈ Di .) 

5 Stable theories
We say that T is a stable theory if every formula is stable.
Let p( x ) ∈ S(U) be a global type. We say that p( x ) is a definable type if for all for-
mulas ϕ( x ; z) ∈ L the sets D p,ϕ are definable. We say that p( x ) is definable over A
if so are all the sets D p,ϕ .
If p( x ) ∈ S(U) is a global type, a canonical base of p( x ) is a definably closed set
Cb(p) ⊆ Ueq such that an automorphism f ∈ Aut (U) fixes p( x ) if and only if
it fixes Cb ( p) pointwise. When they exist, canonical bases are unique, see Exer-
cise 16.27.
Clearly, any definable type has a canonical base, namely
 
Cb ( p) = dcleq D p,ϕ : ϕ( x ; z) ∈ L .
Therefore if T is stable, all global types have a canonical base.
We now turn to Lascar invariance. Quite interestingly when T is stable this reduces
to a more manageable kind of invariance.

16.21 Proposition Let T be stable and let p( x ) ∈ S(U). Then the following are equivalent
1. p( x ) is Lascar invariant over A;

2. p( x ) is definable over acleq A;

3. D p,ϕ ∈ acleq A for all ϕ( x ; z) ∈ L.

Proof 3⇒2⇒1 are clear (stability is not required).

1⇒3 The sets D p,ϕ are externally definable therefore, by Theorem 16.9, definable
(over U). As p( x ) is Lascar invariant over A, so are the sets D p,ϕ . Hence they belong
to acleq A by Theorem 13.10. 

A type q( x ) ⊆ L( A) is stationary if it has a unique global extension that is Las-

car invariant over A. The following proposition says that in a stable theory with
elimination of imaginaries types over algebraically closed sets are stationary.

16.22 Proposition If T is stable then every type q( x ) ∈ S acleq A is stationary.

130
 Proof Let pi ( x ) ∈ S(Ueq ), for i = 1, 2, be two global types that extend q( x ) and are
invariant over acleq A. To prove that p1 ( x ) = p2 ( x ), it suffices to show that for every
formula ϕ( x ; z) ∈ L
# D p1 ,ϕ = D p2 ,ϕ
Note that, by Proposition 16.21 both sets belong to acleq A. Clearly, for i = 1, 2, the
formula ∀z ϕ( x ; z) ↔ z ∈ D pi ,ϕ belongs to pi ( x ). Then both formulas belong to
 

q( x ) and # follows. 

16.23 Corollary If T is stable then the following are equivalent

L
1. a ≡ A b, see Definition 15.10;
Sh
2. a ≡ A b, see Definition 13.12.

Proof 1⇒2. This is left as an exercise to the reader (stability is not required).
Sh
2⇒1. Assume a ≡ A b. By Proposition 13.13 this is equivalent to a ≡acleq A b. Let
q( x ) = tp( a/acleqA) = tp(b/acleqA). Let p( x ) ∈ S(Ueq ) be the unique global type
that is invariant over acleq A and extends q( x ) which we obtain from by Proposi-
tion 16.22. Let c̄ = hci : i < ω i be such that ci  p acleq ( A), a, b, ci . Then a, c̄ and
b, c̄ are A-indiscernible sequences, which proves 1, see Exercise 15.13. 

We end this section with a characterization of stability which is not directly related
with the properties discussed above.

16.24 Proposition The following are equivalent

1. T is stable;

2. every A-indiscernible sequence is totally A-indiscernible.

Proof 2⇒1. Assume ¬1 and let ϕ( x ; z) be an unstable formula witnessed by the

ladder sequence h ai ; bi : i < ω i. Let h ai0 ; bi0 : i < ω i be indiscernible sequence
with the same Ehrenfeucht-Mostowski type as h ai ; bi : i < ω i. This is not totally
indiscernible because ϕ( ai0 ; b0j ) if and if i < j.
1⇒2. Assume ¬2 and let h ai : i < ω i be an A-indiscernible sequence, which is
not totally A-indiscernible. Then there is a formula ϕ( x, y) ∈ L( A) and some i < j
such that ϕ( ai , a j ) ∧ ¬ ϕ( a j , ai ). By indiscernibility ϕ( ai , a j ) holds if and only if i ≤ j.
Hence ϕ( x ; y) is not stable. 

16.25 Exercise Prove that if every formula ϕ( x ; z) ∈ L with | x | = 1 is stable then T is

stable. Hint: by compactness, if all sets approximable by ϕ( x ; y, z) are definable, so
are the sets approximable by ϕ( x, y ; z). 

16.26 Exercise Prove that strongly minimal theories are stable. 

16.27 Exercise Let p( x ) ∈ S(U). Prove that there is at most one definably closed set
A ⊆ Ueq such that Aut (U/A) is the set of automorphisms that fix p( x ). 

131
 6 Stability and the number of types
The following proposition highlights the connection between stability and the car-
dinality of types.
Binary trees of formulas have been introduced in Definition 12.18. Here we restrict
to trees hψs : s ∈ 2<λ i of a particular form. Namely, ψ∅ = > and for every s ∈ 2<λ
we have ψs1 = ϕ( x ; bs ) and ψs0 = ¬ ϕ( x ; bs ). When such a binary tree of height ω
exists, we say that ϕ( x ; z) has the binary tree property .

16.28 Proposition The following are equivalent

1. ϕ( x ; z) is a stable formula;

2. S ϕ ( A) = | A| for all infinite sets A;

3. S ϕ ( A) < 2| A| for all infinite sets A;
4. ϕ( x ; z) does not have the binary tree property.

Proof 2⇔3⇔4. This is proved as in Lemma 12.19.

3⇒1. Assume ¬1. Let I, < I be a dense linear order of cardinality | A|. Let xi and zi ,
for i ∈ I, be a tuples of variables. The type that says that
i<j ⇔ ϕ ( xi ; z j ) for all i, j ∈ I
is finitely consistent by ¬1. If ai , bi is a realization of this type, then for every
< I -upward closed set C ⊆ I the type
 
ϕ( x ; b j ) : j ∈ C ∪ ¬ ϕ( x ; b j ) : j ∈
/C
is a finitely consistent. Any two such types are mutually inconsistent. As there are
2| A| Dedekind’s cuts in I, ¬3 follows.
1⇒4 Assume ϕ( x ; z) has the binary tree property witnessed, say, by hbs : s ∈ 2<ω i.
We show that there is a ladder sequence of length n. For i = 0, . . . , n abbreviate b1i
by bi . Let a j be such that
j −1
^
aj  ¬ ϕ( x ; b j ) ∧ ϕ ( x ; bi ) .
i =0
Then
i < j ⇔ ϕ ( a i ; b j ),
so h ai ; bi : i < ni is a ladder sequence of length n. 

16.29 Theorem The following are equivalent

1. T is stable;

2. |S( A)| ≤ | A| some infinite cardinal λ, and all sets A of cardinality ≤ λ.

3. |S( A)| ≤ | A| for every set A such that | L| < cf( A).

Proof 2⇒1. Suppose a formula ϕ( x ; z) is unstable. Let hzs : s ∈ 2<λ i be a sequence

variables of length |z|. Let p(zs : s ∈ 2<λ ) be the type that says that zs : i < 2<λ
witnesses a binary tree of height λ. As ϕ( x ; z) is unstable, p is finitely consistent.
As λ is an arbitrary infinite cardinal, this contradicts 2.
3⇒2. Trivial.

132
 1⇒3. Proposition 16.28 implies that |S( A)| ≤ | A|| L| . Therefore, when | L| < cf( A),
we obtain |S( A)| ≤ | A|. 

16.30 Exercise Prove that ϕ( x ; z) is stable if and only if |S ϕ (U)| ≤ 2κ . 

16.31 Exercise Prove that ϕ( x ; z) is stable if and only if every set approximated by ϕ( x ; z)
is definable. 

7 Notes and references

[1] Artem Chernikov and Pierre Simon, Externally definable sets and dependent pairs,
Israel J. Math. 194 (2013), no. 1, 409–425, arXiv:1007.4468.
[2] Katrin Tent and Martin Ziegler, A course in model theory, Lecture Notes in Logic,
vol. 40, Association for Symbolic Logic, Cambridge University Press, 2012.

133
 Chapter 17
Topological dynamics

1 T.b.c.

134
 Chapter 18
Keisler measures

1 T.b.c.

135