Model Theory

Model theory
Course Notes
November 30, 2023

Manuel Bodirsky, Institut für Algebra, TU Dresden
Disclaimer: this is a draft and probably contains many typos and mistakes.
Please report them to Manuel.Bodirsky@tu-dresden.de.
The text contains 98 exercises; the ones with a star are harder.
Acknowledgements. The author is grateful to Leopold Schlicht, Stephanie Feil-
itzsch, Florian Starke, and to the students at TU Dresden from the summer semester
2021 for reporting typos and mistakes.
Contents
Chapter 1. First-order Logic 5

1.1. First-order Structures 5
1.2. Formulas, Sentences, Theories 8
1.3. Vaught’s Conjecture 11
Chapter 2. Compactness 13
2.1. Filter 14
2.2. Ultrafilter 14
2.3. Ultraproducts 16
2.4. The Theorem of Łoś 16
2.5. Proof of the Compactness Theorem 18
2.6. Proving the Ultrafilter Lemma with Compactness 18
Chapter 3. The Löwenheim-Skolem Theorem 21

3.1. Chains 21
3.2. Proving Löwenheim-Skolem 24
3.3. Vaught’s test 24
Chapter 4. Fraïssé Amalgamation 27

4.1. The Age of a Structure 27
4.2. Amalgamation Classes 27
4.3. Fraïssé’s theorem 28
4.4. Strong Amalgamation 30
4.5. Further Examples: C-relations 32
Chapter 5. Types 35
5.1. Stone Spaces 37
5.2. Saturated Structures 37
5.3. Omitting Types 40
Chapter 6. Countably Categorical Structures 43

6.1. Algebraicity 45
6.2. Quantifier Elimination 46
6.3. First-Order Interpretations 47
Chapter 7. Model Completeness and Quantifier-Elimination 53

7.1. Preservation Theorems 53
7.2. Model Completeness 56
7.3. Model Companions 57
7.4. Existentially Closed Structures 58
7.5. The Kaiser Hull 59
7.6. Quantifier Elimination 61
7.7. Algebraically Closed Fields 62
3
4 CONTENTS
Chapter 8. Stability 65
8.1. The Order Property 66
8.2. Indiscernible Sequences 68
8.3. ω-stable Theories 71
8.4. The Classification Picture 71
Chapter 9. NIP Theories 75

9.1. VC Dimension 75
9.2. The Lemma of Sauer-Shelah 76
Chapter 10. Geometric Model Theory 79
10.1. Pregeometries 79
10.2. Minimal and Strongly Minimal Sets 84
10.3. Uncountably Categorical Structures 86
Appendix A. Set Theory 87
A.1. Sets and Classes 87
A.2. Zorn’s Lemma 88
A.3. Ordinals 88
A.4. Cardinals 89
Appendix B. Rings and Fields 93
B.1. Integral Domains 93
B.2. Ideals and Quotient Rings 94
B.3. Polynomial Rings 98
B.4. Field Extensions 98
Appendix C. Ramsey’s theorem 105
Bibliography 109
CHAPTER 1
First-order Logic
There are many excellent text books about model theory. A classic is Hodges’
model theory [18]; a short version is also available [19]. Another introduction is given
by Marker [22]. A more recent book has been written by Tent and Ziegler [29].
1.1. First-order Structures

1.1.1. Signatures. A signature τ is a set of relation and function symbols, each
equipped with an arity k ∈ N. Examples of important signatures:
• τGraph = {E}
• τAGroup = {+, −, 0}
• τRing = τAG ∪ {1, ·}
• τGroup = {◦,−1 , e}
• τLO = {<}
• τORing = τRing ∪ τLO
• τArithmetic = {0, s, +, ·, <}
• τSet = {∈}
Here, +, ◦, and · are binary function symbols, −, −1 , and s are unary function
symbols, e, 0, and 1 are 0-ary function symbols, and E, ∈, and < are binary relation
symbols.
1.1.2. Structures. A τ -structure A is a set A (called the domain, or the base

set, or the universe of A) together with
• a relation RA ⊆ Ak for each k-ary relation symbol R ∈ τ . Here we allow the
case k = 0, in which case RA is either empty or of the form {()}, i.e, the set
consisting of the empty tuple;
• a function f A : Ak → A for each k-ary function symbol f ∈ τ ; here we also
allow the case k = 0 to model constants from A.
Unless stated otherwise, A, B, C, . . . denotes the domain of the structure A, B, C,
A A A A
. . . , respectively. We sometimes write (A; R1 , R2 , . . . , f1 , f2 , . . . ) for the structure
A A A A
A with relations R1 , R2 , . . . and functions f1 , f2 , . . . Well-known examples of struc-
tures are (Q; +, −, 0, 1, ·, <), (C; +, −, 0, 1, ·, <), (Z; 0, s, +, ·, <), etc. We say that a
structure is infinite if its domain is infinite. If all symbols in τ are function symbols,
then a τ -structure A is also called an algebra (in the sense of universal algebra); in the
other extreme, if all symbols in τ are relation symbols, then A is called a relational
structure.
Example 1. A (simple, undirected) graph is a pair (V, E) consisting of a set of
vertices V and a set of edges E ⊆ V2 , that is, E is a set of 2-element subsets of V .
Graphs can be modelled as relational structures G using a signature that contains a
single binary relation symbol R, putting G := V and adding (u, v) to RG if {u, v} ∈ E.
If we insist that a structure with this signature satisfies (x, y) ∈ RG ⇒ (y, x) ∈ RG
and (x, x) 6∈ RG , then we can associate to such a structure an undirected graph and
5
6 1. FIRST-ORDER LOGIC
obtain a bijective correspondence between undirected graphs with vertices V and

structures G with domain V as described above. 4
Example 2. A group is an algebra G with the signature τGroup = {◦,−1 , e} such
that for all x, y, z ∈ G we have that
• x ◦ (y ◦ z) = (x ◦ y) ◦ z,
• x ◦ x−1 = e,
• e ◦ x = x and x ◦ e = x. 4
1.1.3. Substructures. A τ -structure A is a substructure of a τ -structure B iff
• A ⊆ B,
• for each R ∈ τ , and for all tuples ā from A, ā ∈ RA iff ā ∈ RB , and
• for each f ∈ τ we have that f A (ā) = f B (ā).
In this case, we also say that B is an extension of A. Substructures A of B and
extensions B of A are called proper if the domains of A and B are distinct.
Note that for every subset S of the domain of B there is a unique smallest sub-
structure of B whose domain contains S, which is called the substructure of B gen-
erated by S, and which is denoted by B[S].
Example 3. When we view a graph as an {E}-structure G, then a subgraph is
not necessarily a substructure of G. In graph theory, the substructures of G are called
induced subgraphs: the difference is that in an induced subgraph (V 0 , E 0 ) of (V, E) the
0
edge set must be of the form E 0 := E ∩ V2 instead of an arbitrary subset of it. 4
Example 4. Due to the choice of our signature τGroup , the subgroups of G are
precisely the substructures of G as defined above. 4
1.1.4. Expansions and Reducts. Let σ, τ be signatures with σ ⊆ τ . If A is a
σ-structure and B is a τ -structure, both with the same domain, such that RA = RB
for all relations R ∈ σ and f A = f B for all functions and constants f ∈ σ, then A is
called a reduct of B, and B is called an expansion of A.
Let A be a τ -structure and B ⊆ A. We write AB for the expansion of A by the
constant symbol b for every element b ∈ B (we set bAB := b).
1.1.5. Homomorphisms. In the following, let A and B be τ -structures. A

homomorphism h from A to B is a mapping from A to B that preserves each function
and each relation for the symbols in τ ; that is,
• if (a1 , . . . , ak ) is in RA , then (h(a1 ), . . . , h(ak )) must be in RB ;
• f B (h(a1 ), . . . , h(ak )) = h(f A (a1 , . . . , ak )).
A homomorphism from A to B is called a strong homomorphism if it also preserves
the complements of the relations from A. Injective strong homomorphisms are called
embeddings. Surjective embeddings are called isomorphisms.
Example 5. Group homomorphisms, field endomorphisms, ring endomorphisms,
linear maps between vector spaces. 4
Example 6. The graph colorability problem is an important problem in discrete
mathematics, with many applications in theoretical computer science (it can be used
to model e.g. frequency assignment problems). As a computational problem, the
graph n-colorability problem has the following form.
Given: a finite graph G = (V, E).
Question: can we colour the vertices of G with n colours such that adjacent vertices
get different colours?
1.1. FIRST-ORDER STRUCTURES 7
The n-colorability problem can be formulated as a graph homomorphism problem: is

there a homomorphism from G to
Kn := ({1, . . . , n}; E Kn ) where E Kn := {(u, v) ∈ V 2 | u 6= v} .
We also refer to these homomorphisms as proper colourings of G, and say that G is n-
colourable if such a colouring exists. The chromatic number χ(G) of G is the minimal
natural number n ∈ N such that G is n-colourable. For example, the chromatic
number of Kn is n. 4
Example 7. We present a concrete instance of a colouring problem from pure
mathematics. Let (V, E) the unit distance graph on R2 , i.e., the graph has the vertex
set V := R2 (we imagine the nodes as the points of the Euclidean plane) and edge set
E := (x, y) ∈ V 2 |x − y| = 1 .

In other words, two points are linked by an edge if they have distance one. What is
the chromatic number of this graph? 4
The following statement follows from the compactness theorem of first-order logic,
as we will see later in this chapter.
Proposition 1.1.1. Let G be a graph such that all finite subgraphs of G are
k-colourable. Then G is k-colourable.
The problem to determine the chromatic number χ of this graph is known as the
Hadwiger-Nelson problem. It is known that χ ≤ 7. We have seen that 4 ≤ χ. In
April 11, 2018, Aubrey de Grey announced a proof that 5 ≤ χ. The precise value of
χ ∈ {5, 6, 7} is not known.
1.1.6. Products. Let A and B be τ -structures. Then the (direct, or categorical )
product C = A×B is the τ -structure with domain A×B, which has for each k-ary R ∈
τ the relation that contains a tuple ((a1 , b1 ), . . . , (ak , bk )) if and only if R(a1 , . . . , ak )
holds in A and R(b1 , . . . , bk ) holds in B. For each k-ary f ∈ τ the structure C has
the operation that maps ((a1 , b1 ), . . . , (ak , bk )) to (f (a1 , . . . , ak ), f (b1 , . . . , bk )). The
direct product A × A is also denoted by A2 , and the k-fold product A × · · · × A,
defined analogously, by Ak .
Note that the product of two groups, viewed as structures, is precisely the well-
known product of groups.
We generalise the definition of products to infinite products Q as follows. For a
sequence of τ -structures
Q (A )
i i∈I , the direct product B = i∈I i is the τ -structure
A
on the domain i∈I Ai such that for R ∈ τ of arity k
(a1i )i∈I , . . . , (aki )i∈I ∈ RB iff (a1i , . . . , aki ) ∈ RAi for each i ∈ I ,

and for f ∈ τ of arity k, we have

f B (a1i )i∈I , . . . , (aki )i∈I = f Ai (a1i , . . . , aki ) i∈I .

If Ai = A for all i ∈ I, we also write AI instead of i∈I Ai , and call it the I-th power
Q
of A.
Exercises.
(1) Show that there is a homomorphism from A2 to A. Find an example of a
structure A where all such homomorphisms are strong, and another example
where all such homomorphisms are not strong.
(2) Find a signature τ for vector spaces and describe how a vector space may
be viewed as a τ -structure. Your definition should have the property that
homomorphisms between the structures you consider correspond precisely
to linear maps.
(3) Let A and B be τ -structures and suppose that G ⊆ A generates A, i.e.,

A = A[G]. Then every homomorphism h : A → B is determined by its
values on G.
(4) Let A and B be τ -structures, and suppose that τ has no relation symbols.
Show that every bijective homomorphism from A to B is an isomorphism.
(5) Find an example of a τ -structure A and a bijective homomorphism from A
to A which is not an isomorphism.
1.2. Formulas, Sentences, Theories

To define the syntax of first-order logic, we first have to introduce terms, before
we can define (first-order) formulas and (first-order) sentences, and finally (first-order)
theories.
1.2.1. Terms. Let τ be a signature. In this section we will see how to use the
function symbols in τ to build terms.
Definition 1.2.1. A (τ -)term is defined inductively:
• constants from τ are τ -terms;
• variables x0 , x1 , . . . are τ -terms;
• if t1 , . . . , tk are τ -terms, and f ∈ τ has arity k, then f (t1 , . . . , tk ) is a τ -term.
We write t(x1 , . . . , xn ) if all variables that appear in t come from {x1 , . . . , xn };
we do not require that each variable xi appears in t.
Example 8. Well-known examples of terms are polynomials over a ring R: they
are terms over a signature that contains a binary symbol + for addition and a binary
symbol ∗ for multiplication, together with a constant symbol for each element of
R. 4
1.2.2. Semantics of Terms. In this section we describe how τ -terms over a
given τ -structure describe functions (in the same way as polynomials describe poly-
nomial functions over a given ring).
Let A be a τ -structure and let x1 , . . . , xn be distinct variables. Every τ -term
t(x1 , . . . , xn ) describes a function tA : An → A as follows:
(1) if t equals c ∈ τ then tA is the function (a1 , . . . , an ) 7→ cA ;
(2) if t equals xi then tA is the function (a1 , . . . , an ) 7→ ai ;
(3) if t equals f (t1 , . . . , tk ) for a k-ary f ∈ τ then tA is the function
A A
(a1 , . . . , an ) 7→ f A t1 (a1 , . . . , an ), . . . , tk (a1 , . . . , an )

Note that item 1 is a special case of item 3. The function described by t is also called
the term function of t (with respect to A).
Exercises.
(6) Let B be a τ -structure and G ⊆ B. Let A be the substructure of B generated
by G. Show that for every element a ∈ A there exists a τ -term t(x1 , . . . , xn )
and elements g1 , . . . , gn ∈ G such that tB (g1 , . . . , gn ) = a.
1.2.3. Formulas and sentences. Let τ be a signature. The relation symbols
in the signature τ did not play any role when defining τ -terms, but they become
important when defining τ -formulas. Moreover, the equality symbol = is ‘hard-wired’
into first-order logic; we can use it to create formulas by equating terms. Finally, we
can combine formulas using Boolean connectives, and quantify over variables.
Definition 1.2.2. An atomic (τ -)formula is an expression of the form
• t1 = t2 where t1 and t2 are τ -terms;
1.2. FORMULAS, SENTENCES, THEORIES 9
• R(t1 , . . . , tk ) where t1 , . . . , tk are τ -terms and R ∈ τ is a k-ary relation

symbol.
Formulas are defined inductively as follows:
• atomic formulas are formulas;
• if φ is a formula, then ¬φ is a formula (negation);
• if φ and ψ are formulas, then φ ∧ ψ is a formula (conjunction);
• if φ is a formula, and x is a variable, then ∃x. φ is a formula (existential
quantification)
Atomic formulas and negations of atomic formulas are sometimes called literals.
Similarly as for terms, we write φ(x1 , . . . , xn ) if all non-quantified variables in the
formula φ come from {x1 , . . . , xn }. A (first-order) sentence is a formula without free
variables, i.e., all variables that appear in the formula are bound, i.e., quantified by
some quantifier.
Example 9. Let τ = {R} where R is a binary relation symbol. Then the following
formula is an example of a first-order sentence

∀x1 , x2 , x3 R(x1 , x2 ) ∧ R(x2 , x3 ) ⇒ R(x1 , x3 ) . 4
1.2.4. Semantics of formulas. So far, we have just introduced the syntax of
first-order logic, i.e., the shape of formulas and sentences, without discussing what
these expressions actually mean. In this section we discuss their semantics; the idea
is that formulas can be used to define new relations and functions in a given structure,
and that sentences can be used to describe properties that a structure might have or
not.
Let A is a τ -structure. Then every τ -formula φ(x1 , . . . , xn ) describes a relation
φA ⊆ An as follows:
• if φ equals t1 = t2 then φA is the relation
A A
(a1 , . . . , an ) | t1 (a1 , . . . , an ) = t2 (a1 , . . . , an ) ;
• if φ equals R(t1 , . . . , tk ) then
A A
φA := (a1 , . . . , an ) | (t1 (ā), . . . , tk (ā)) ∈ RA ;

A A
• if φ equals φ1 ∧ φ2 then φA := φ1 ∩ φ2 ;
• if φ equals ¬ψ then φA := An \ ψ A ;
• if φ equals ∃x. ψ(x, x1 , . . . , xn ) then
φA := {(a1 , . . . , an ) | there exists a ∈ A s.t. (a, a1 , . . . , an ) ∈ ψ A }.
A relation R ⊆ Ak is called (first-order) definable in A if there exists a τ -formula
φ(x1 , . . . , xk ) such that R = φA ; we also say that φ defines R over A. An function
f : S → T , for S ⊆ Ak and T ⊆ Al , is called definable if the relation
{(a1 , . . . , ak , b1 , . . . , bl ) | f (a1 , . . . , ak ) = (b1 , . . . , bl )}
is definable in A.
We freely use brackets to avoid ambiguities when writing down terms. That is,
we write ¬(x ∧ y) to distinguish it from (¬x) ∧ y. If brackets are omitted, there is the
convention that negation ¬ binds stronger than conjunction ∧; that is, ¬x ∧ y would
stand for (¬x) ∧ y.
For φ(x1 , . . . , xn ) we write
A |= φ(a1 , . . . , an )
A
instead of (a1 , . . . , an ) ∈ φ . In particular, if φ is a sentence, i.e., if n = 0, we write
A |= φ, and say that A satisfies φ, if () ∈ φA (that is, if φA 6= ∅).
Shortcuts:
• Disjunction: φ ∨ ψ is an abbreviation for ¬(¬φ ∧ ¬ψ)
• Implication: φ ⇒ ψ is an abbreviation for ¬φ ∨ ψ
• Equivalence: φ ⇔ ψ is an abbreviation for φ ⇒ ψ ∧ ψ ⇒ φ
• Universal quantification: ∀x. φ(x) is an abbreviation for ¬∃x. ¬φ(x)
• Inequality: x 6= y is an abbreviation for ¬(x = y)
• False: ⊥ is an abbreviation for ∃x.x 6= x.
• True: > is an abbreviation for ¬⊥ (the same as ∀x.x = x).
Moreover, when A is a unary relation symbol, then we may write ∃x ∈ A. φ instead
of ∃x (A(x) ∧ φ) and ∀x ∈ A. φ instead of ∀x (A(x) ⇒ φ).
Example 10. The following statements about well-known structures follow straight-
forwardly from the definitions.
• (Z; <) |= 0 < 1
• (Q; <) |= ∀x, y x < y ⇒ ∃z (x < z ∧ z < y) (density)
• (Z; <) |6 = ∀x, y x < y ⇒ ∃z (x < z ∧ z < y) 4
1.2.5. First-order Theories. Let τ be a signature. A (τ -)theory is a set of

first-order τ -sentences. Let A be a τ -structure and T a τ -theory. Then A is a model
of T , in symbols A |= T , if A |= φ for all φ ∈ T . A τ -theory T is called
• satisfiable (or consistent) if T has a model.
• complete if T is satisfiable and for every τ -sentence either φ ∈ T or ¬φ ∈ T .
The (first-order) theory of A is defined as the set of all first-order τ -sentences
that are satisfied by A. Note that Th(A) is always a complete theory.
Example 11. A famous example of a first-order theory is the set ZF of axioms of

Zermelo-Fraenkl set theory. Recall that this is an infinite set of first-order sentences
over the signature τ = {∈}. An example of a sentence in ZF is
∃x∀y. ¬(y ∈ x)
stating the existence of the empty set. Note that ZF is not complete: for example the
Axiom of Choice (C) is independent from ZF in the sense that both ZFC := ZF ∪ {C}
and ZF ∪ {¬C} have a model. 4
Example 12. The theory TAGroup of abelian groups is over the signature τAGroup =
{+, −, 0} and contains the following axioms:
∀x, y, z. (x + y) + z = x + (y + z) (associativity)
∀x. 0 + x = x + 0 = y (neutral element)
∀x. x + (−x) = 0 (inverse elements)
∀x, y. x + y = y + x (Abelian) 4
Example 13. The theory TCRing of commutative rings is over the signature τRing ,
contains TAGroup and the following additional axioms:
∀x, y, z. (xy)z = x(yz) (associativity)

∀x. 1 · x = x (multiplicative unit)
∀x, y. xy = yx (commutativity)
∀x, y, z. x(y + z) = xy + xz (distributivity) 4
1.3. VAUGHT’S CONJECTURE 11
Example 14. The theory TField of fields is over the signature τRing , contains
TCRing and the following additional axioms:
¬(0 = 1)
∀x (¬(x = 0) ⇒ ∃y.xy = 1) 4
If S and T are τ -theories then we write
S |= T
if every model of S is also a model of T . We also write S |= φ instead of S |= {φ}.
Exercises.
(7) Let A and B be τ -structures. If a ∈ An and b ∈ B n satisfy the same
quantifier-free formulas, then the substructure of A generated by the entries
of a is isomorphic to the substructure of B generated by the entries of b.
(8) Show that for every first-order formula φ(x) and for every ` ∈ N there exists
a first-order sentence ψ such that a structure A satisfies ψ if and only if
there are precisely ` elements a ∈ A such that A |= φ(a). The sentence ψ is
often denoted by ∃x=` .φ.
(9) Write down the axioms of algebraically closed fields in first-order logic.
(10) A formula is in prenex normal form if it is of the form Q1 x1 . . . Qn xn . φ where
Qi is either ∃ or ∀ and φ is without quantifiers. Show that every formula
φ(y1 , . . . , yn ) is equivalent to a formula ψ(y1 , . . . , yn ) in prenex normal form,
that is, for every structure M we have M |= ∀ȳ(φ(ȳ) ⇔ ψ(ȳ)).
(11) Show that every quantifier-free formula is equivalent to a formula in conjunc-
tive normal form, i.e., to a conjunction of disjunctions of atomic formulas
and negated atomic formulas.
1.3. Vaught’s Conjecture

Let T be a complete theory over a countable signature and let κ be a cardinal.
How many models of cardinality κ can T have, up to isomorphism? Let IT (κ) be the
number of isomorphism types of models of T of cardinality κ, also called the spectrum
of T . Examples:
ITh(K5 ) (ℵ0 ) = 0 (See Exercise 12)
ITh(Q;<) (ℵ0 ) = 1 (see Example 15)
ITh(Z;{(x,y)|y=x+1}) (ℵ0 ) = ℵ0 (exercise)
ITh(R;+,<) (ℵ0 ) = 2ℵ0 (exercise)
Example 15. ITh(Q;<) (ℵ0 ) = 1. The standard approach to verify this is a so-
called back-and-forth argument. Let A be a countable model of the first-order theory
T of (Q; <). It is easy to verify that T contains (and, as this argument will show, is
uniquely determined by)
• ∃x. x = x (no empty model)
• ∀x, y, z ((x < y ∧ y < z) ⇒ x < z) (transitivity)
• ∀x. ¬(x < x) (irreflexivity)
• ∀x, y (x < y ∨ y < x ∨ x = y) (totality)
• ∀x ∃y. x < y (no largest element)
• ∀x ∃y. y < x (no smallest element)
• ∀x, z x < z ⇒ ∃y (x < y ∧ y < z) (density).
An isomorphism between A and (Q; <) can be defined inductively as follows. Suppose
that we have already defined f on a finite subset S of Q and that f is an embedding
of the structure induced by S in (Q; <) into A. Since <A is dense and unbounded, we
can extend f to any other element of Q such that the extension is still an embedding
from a substructure of Q into A (going forth). Symmetrically, for every element v of A
we can find an element u ∈ Q such that the extension of f that maps u to v is also an
embedding (going back ). We now alternate between going forth and going back; when
going forth, we extend the domain of f by the next element of Q, according to some
fixed enumeration of the elements in Q. When going back, we extend f such that
the image of A contains the next element of A, according to some fixed enumeration
of the elements of A. If we continue in this way, we have defined the value of f on
all elements of Q. Moreover, f will be surjective, and an embedding, and hence an
isomorphism between A and (Q; <). 4
Vaught’s theorem states that for any theory T the spectrum IT (ℵ0 ) cannot be
two. Morley showed that if IT (ℵ0 ) is infinite, then
IT (ℵ0 ) ∈ {ℵ0 , ℵ1 , 2ℵ0 }
Vaught’s Conjecture: if IT (ℵ0 ) is infinite, then IT (ℵ0 ) ∈ {ℵ0 , 2ℵ0 }.

Exercises.
(12) Show that two finite structures are isomorphic if and only if they have the
same theory.
(13) Show that if τ is a countable signature, then for any infinite cardinal κ there
are at most 2κ many non-isomorphic τ -structures of cardinality κ.
(14) Let A be a {E}-structure where E is an equivalence relation on A with
infinitely many infinite classes. Show that ITh(A) (ℵ0 ) = 1.
(15) (∗) Prove the claim that ITh(Z;{(x,y)|y=x+1}) (ℵ0 ) = ℵ0 .
(16) (∗) Determine the value of ITh(Z;<) (ℵ0 ).
CHAPTER 2
Compactness
A special case of the compactness theorem was already proved by Gödel; the
general case below was proved by A. Maltsev in 1936.
Theorem 2.0.1. A first-order theory T is satisfiable if and only if all finite subsets
of T are satisfiable.
Remark 2.0.2. The name compactness theorem comes from the fact that the
compactness theorem is equivalent to the statement that the following natural topo-
logical space is compact: the space is the set T (τ ) of all complete τ -theories, and the
basic open sets are the sets Tφ of the form {T ∈ T (τ ) | φ ∈ T }.
To see the equivalence, let C be Sa covering of T (τ ) by open subsets of T (τ ). This
means that T (τ ) can be written as φ∈S Tφ for some set S of τ -sentences. Consider
the set S 0 of all τ -sentences that are not contained in S. If S 0 has a model B, then
the theory Th(B) is not covered by C. The compactness theorem implies that there
is a finite subset F of S 0 which is unsatisfiable. But then {T¬φ | φ ∈ F } is a finite
subset of C covering T (τ ), showing that T (τ ) is compact.
Conversely, suppose that T (τ ) is compact, and that the τ -theory S is unsatisfiable.
Then {T¬φ | φ ∈ S} is an open covering of T (τ ). So by compactness
S it has a finite
subcovering, i.e., there is a finite subset F of S such that {T¬φ | φ ∈ F } = T (τ ).
Hence, F is unsatisfiable, which is the statement of the compactness theorem.
There are several different proofs of the compactness theorem. It is a consequence

of the completeness theorem of first-order logic, and one can show it using Henkin’s
construction. In this course we present a proof based on ultraproducts; they are an
important tool of model theory to build interesting structures and have found many
applications in mathematics, e.g., in topology, set theory, and algebra.
Exercises.
(17) Show that if a first-order theory has infinite models, then it also has arbi-
trarily large models (for every set X there is a model M with |M | ≥ |X|).
(18) Let K be a field. The characteristic of K is the smallest n ∈ N such that
1 + ··· + 1 = 0
| {z }
n times
if such an n exists, and 0 otherwise. Let T be a theory that contains the

theory of fields.
• Show that if there are models of T of arbitrarily large finite character-
istic, then there exists a model of T of characteristic 0.
• Show that there is no finite first-order theory whose models are precisely
the fields of characteristic 0.
13
14 2. COMPACTNESS
2.1. Filter
Let X be a set. A filter on X is a certain set of subsets of X; the idea is that the
elements of F are (in some sense) ‘large’; it helps thinking of the elements F ∈ F as
being ‘almost all’ of X.
Definition 2.1.1. A filter F on X is a set of subsets of X such that
(1) ∅ ∈/ F and X ∈ F;
(2) if F ∈ F and G ⊆ X contains F , then G ∈ F.
(3) if F1 , F2 ∈ F then F1 ∩ F2 ∈ F.
Note that filters have the finite intersection property:
A1 , . . . , An ∈ F ⇒ A1 ∩ · · · ∩ An 6= ∅ (FIP)
Lemma 2.1.2. Every subset S ⊆ P(X) with the FIP is contained in a smallest
filter that contains S; this filter is called the filter generated by S.
Proof. First add finite intersections, and then all supersets to S.
Example 16. For a non-empty subset Y ⊆ X, the family

F := {Z ⊆ X | Y ⊆ Z}
is a filter, the filter generated by a {Y }; such filters are called principal . 4
Example 17. The Fréchet filter : for an infinite set X this is the filter F that
consists of all cofinite subsets of X, i.e.,
F := {Y ⊆ X | X \ Y is finite}. 4
2.2. Ultrafilter
A filter F is called a ultrafilter if F is maximal, that is for every filter G ⊇ F we
have G = F.
Lemma 2.2.1. Let F be a filter. Then the following are equivalent.
(1) F is a ultrafilter.
(2) For all A ⊆ X either A ∈ F or X \ A ∈ F.
(3) For all A1 ∪ · · · ∪ An ∈ F there is an i ≤ n with Ai ∈ F.
Proof. (1) ⇐ (2): No A ⊆ X can be added to F. Hence, F is maximal.
(2) ⇐ (3): Note that A ∪ (X \ A) = X ∈ F.
(1) ⇒ (3): If there is an i ≤ n such that F ∪{Ai } has the FIP, then by Lemma 2.1.2
there is a filter that contains this set, and hence F was not maximal. Otherwise, there
are S1 , . . . , Sn ⊆ F with Ai ∩ Si = ∅. Then Si ⊆ X \ Ai and thus S1 ∩ · · · ∩ Sn ⊆
X \ (A1 ∪ · · · ∪ An ) ∈/ F, a contradiction.
A filter F is principal if it
T contains a inclusionwise minimal element. Note that
this is the case if and only if F ∈ F.
Lemma 2.2.2. Let F be a filter on a set X. Then the following are equivalent.
(1) F is a principal ultrafilter;
(2) F contains {a} for some a ∈ X.
(3) F is of the form {Y ⊆ X | a ∈ Y } for some a ∈ X.
(4) F is an ultrafilter and contains a finite set.
2.2. ULTRAFILTER 15
T
Proof. (1) ⇒ (2): let A := F ∈ F. If |A| > 1 then we can write A = B1 ∪ B2
for B1 , B2 ⊆ X non-empty. But then Lemma 2.2.1 (3) implies that B1 ∈ F or B2 ∈ F,
in contradiction to the definition of A. So A = {a} for some a ∈ X.
(2) ⇒ (3). Clearly, {Y ⊆ X | a ∈ Y } ⊆ F since F is closed under supersets, and
F ⊆ {Y ⊆ X | a ∈ Y } since F does not contain the empty set.
(3) ⇒ (4): Clearly F contains a finite set; use Lemma 2.2.1 (2) to check that F
is an ultrafilter. T
(4) ⇒ (1): If A ∈ F is finite, then B := F is finite, and hence B is the
intersection of finitely many elements in F, and hence in F since F is a filter. This
shows that F is principal.
Are there non-principal ultrafilters?
Lemma 2.2.3 (Ultrafilter Lemma). Every filter F is contained in a ultrafilter.
Proof. Let M be the set of all filters on X that contain F, partially ordered
by containment. Note that unions of chains of filters in this partial order are again
filters. By Zorn’s lemma (Theorem A.2.1), M contains a maximal filter.
Non-principal ultrafilters are also called free ultrafilters. In particular the Fréchet
filter is contained in an ultrafilter, which must be free:
Lemma 2.2.4. An ultrafilter is free if and only if it contains the Fréchet filter.
Proof. Let U be a free ultrafilter on X and let x ∈ X. Either {x} ∈ U or

X \ {x} ∈ U. As U is free, {x} ∈
/ U (Lemma 2.2.2). Hence, X \ {x} ∈ U for every
x ∈ X. Let F ⊆ X be finite. Then
\
X \F = (X \ {x}) ∈ U .
x∈F
Now let U be a principal ultrafilter, i.e., there is x ∈ X with {x} ∈ U (Lemma 2.2.2).
Then the element X \ {x} of the Fréchet filters is not in U.
Exercises.
(19) Show that a set of subsets of a set X can be extended to an ultrafilter if and
only if it has the FIP.
(20) Show that a set F of subsets of a set X can be extended to a free ultrafilter
if and only if the intersection of every finite subset of F is infinite.
(21) Show that every filter F on a set X is the intersection of all ultrafilters on
X that extend F.
(22) Show that if U is a free ultrafilter on X, S ∈ U, and T ⊆ X is such that the
symmetric difference S∆T is finite, then S ∈ U.
|X|
(23) Show that there are 22 many ultrafilters on an infinite set X. Hint:
first show that there is a family F of 2|X| subsets of X such that for any
A1 , . . . , An , B1 , . . . , Bn ∈ F
A1 ∩ · · · ∩ An ∩ (X \ B1 ) ∩ · · · ∩ (X \ Bn ) 6= ∅.
(24) True or false: if U and V are free ultrafilters on an infinite set X, is there is
a permutation π of X such that S ∈ U if and only if π(S) ∈ V?1
1Thanks to Lukas Juhrich for the idea for this exercise.

16 2. COMPACTNESS
2.3. Ultraproducts
Let τ be a signature, let U be an ultrafilter on X, and for each a ∈ X let M a be
a τ -structure. The idea of ultraproducts is to define an “average” structure of all the
structures M a .
Definition 2.3.1. We write
Y
M a /U
a∈X
for the τ -structure M whose domain are the equivalence classes of the equivalence
relation ∼ defined on the set
Y [
Ma := {g : X → Ma | for all x ∈ X : g(a) ∈ Ma }
a∈X a∈X
as follows:
g ∼ g 0 :⇔ {a ∈ X | g(a) = g 0 (a)} ∈ U.
The equivalence class of a function g with respect to the equivalence relation ∼ will
be denoted by [g]. The relations and functions of M are defined as follows:
• for constant symbols c ∈ τ :
cM := [a 7→ cM a ]
• for relation symbols R ∈ τ of arity k:
RM ([g1 ], . . . , [gk ]) :⇔ {a ∈ X | RM a (g1 (a), . . . , gk (a))} ∈ U
• for function symbols f ∈ τ of arity k:
f M ([g1 ], . . . , [gk ]) = [g0 ] :⇔ {a ∈ X | f M a (g1 (a), . . . , gk (a)) = g0 (a)} ∈ U
It is straightforward (but tedious) to verify that this is indeed well defined, i.e.,
the interpretation of function and relation symbols is independent from the choice of
the representatives2.
Exercises.
(25) Show that Definition 2.3.1 is well-defined.
(26) Let n ∈ N and let M be a ultraproduct of finite structures each of which
has at most n elements. Then M has at most n elements, too.
(27) (Exercise 1.31 in [17]). Let X be an index set and U an ultrafilter on X. Let
(Aa )a∈X and (B a )a∈X be families
Q of τ -structures. If Aa can be Q embedded in
B a for all a ∈ X, show that a∈X Aa /U can be embedded in a∈X B a /U.
2.4. The Theorem of Łoś

Let {M a | a ∈ X} be a family of τ -structures, and let U be an ultrafilter on X.
Theorem 2.4.1 (Łoś). Let φ(x)Q be a first-order τ -formula and [g] be a tuple of
elements of the ultraproduct M := a∈X M a /U. Then
M |= φ([g]) ⇔ {a ∈ X | M a |= φ(g(a))} ∈ U
Proof. By induction over the syntactic form of first-order formulas.
• if φ is atomic and of the form R(x1 , . . . , xn ), then the statement follows from
the definition of RM . If φ contains terms then we have to do an additional
induction over the recursive form of terms, which we omit.
2Note that the same definition works even if U is not an ultrafilter, but just a filter: in this case
the resulting structure is called a reduced product.
2.4. THE THEOREM OF ŁOŚ 17
• Suppose that the statement holds for φ and for ψ. Then

M |= (φ ∧ ψ)([g])
⇔ M |= φ([g]) and M |= ψ([g]) (semantics of conjunction)
⇔ {a ∈ X | M a |= φ(g(a))} ∈ U
and {a ∈ X | M a |= φ(g(a))} ∈ U (inductive assumption)
⇔ {a ∈ X | M a |= φ(g(a))} ∩ {a ∈ X | M a |= ψ(g(a))} ∈ U (U is filter)
⇔ {a ∈ X | M a |= (φ ∧ ψ)(g(a))} ∈ U (semantics of conjunction)
• Suppose that the statement holds for φ. Then
M |= ¬φ([g])
⇔ M 6|= φ([g]) (semantics of negation)
⇔ {a ∈ X | M a |= φ(g(a))} ∈
/U (inductive assumption)
⇔ {a ∈ X | M a 6|= φ(g(a))} ∈ U (U is ultrafilter)
⇔ {a ∈ X | M a |= ¬φ(g(a))} ∈ U (semantics of negation)
• Finally, suppose that the statement holds for the formula φ(x, ȳ).
M |= (∃x.φ)([g])
⇔ there is [f ] ∈ M with M |= φ([f ], [g])
⇔ there is [f ] ∈ M with {a ∈ X | M a |= φ(f (a), g(a))} ∈ U
⇔ {a ∈ X | M a |= (∃x.φ)(g(a))} ∈ U
To see the final equivalence, note that for every [f ] ∈ M
{a ∈ X | M a |= φ(f (a), g(a))} ⊆ {a ∈ X | M a |= (∃x.φ)(g(a))}.
S
Conversely, there is [f ] ∈ M with ⊇. Define f : X → a Ma as follows.
For every a ∈ X with M a |= (∃x.φ)(g(a)) choose a witness ha ∈ Ma for x.
Define (
ha if M a |= (∃x.φ)(g(a))
f (a) :=
arbitrary otherwise.
Then {a ∈ X | M a |= φ(f (a), g(a))} = {a ∈ X | M a |= (∃x.φ)(g(a))}.
The following can be obtained from the theorem of Łoś by applying it to sentences
instead of formulas, and to ultrapowers instead of ultraproducts. An ultrapower is an
ultraproduct where all the factors M a are equal to a single structure A; in this case
we write AX instead of a∈X M a .
Q
Corollary 2.4.2. Let M := AX /U. Then Th(M ) = Th(A).

The statement of the corollary is trivial if U is a principal filter (why?). The
statement is interesting if U is free, as we see in the following example.
Example 18. Consider A := (N; +, ∗, <, 0, 1) and let U be a free ultrafilter on ω.
There is an element u of M := Aω /U such that for every n ∈ N
M |= 1 + · · · + 1 < u
| {z }
n times
for example u = [(1, 2, 3, . . . )] ∈ M . This distinguishes M from A. (Why is there no
contradiction to Corollary 2.4.2?) 4
18 2. COMPACTNESS
Example 19. Let U be a free ultrafilter on ω, and consider A := (R; 0, 1, +, ∗, ≤).

Then
Th(A) = Th(Aω /U)
How big is Aω /U?
|Aω /U| ≤ |Rℵ0 | = (2ℵ0 )ℵ0 = 2ℵ0
ω
Note that A /U has ‘infinitesimal’ elements, i.e., elements x such that for all n ∈ N
Aω /U |= 0 < x ∧ (1 + · · · + 1) ∗x < 1
| {z }
n times
for example
x = [(1, 1/2, 1/3, 1/4, . . . )].
Abraham Robinson used this idea to develop ‘nonstandard analysis’ which gives a for-
mal interpretation to the reasoning with infinitely small positive entities à la Leibniz,
Euler, and Cauchy. 4
2.5. Proof of the Compactness Theorem

Let T be a theory.
Proof of the compactness theorem, Theorem 2.0.1. Clearly, if a theory
T is satisfiable then all finite subsets of the theory are satisfiable as well. For the
converse, assume that every finite subset S of T has a model M S . Let X be the set
of all finite subsets of T . For φ ∈ T , let
Xφ := {S ∈ X | M S |= φ} .
Then the set {Xφ | φ ∈ T } has the FIP: if φ1 , . . . , φn ∈ T , let S := {φ1 , . . . , φn } ∈ X
and note that M S |= φi for all i, and S ∈ Xφi . Hence, S ∈ Xφ1 ∩ · · · ∩ Xφn shows
that Xφ1 ∩ · · · ∩ Xφn is non-empty, proving the FIP.
By Lemma 2.1.2 and the ultrafilter lemma Q (Lemma 2.2.3) there is an ultrafilter
U that contains {Xφ | φ ∈ T }. Then M := S∈X M S /U is a model of T : for φ ∈ T
we have Xφ ∈ U and M |= φ because of the theorem of Łoś (Theorem 2.4.1).
2.6. Proving the Ultrafilter Lemma with Compactness

Let F be a filter on a set X. We want to show that F is contained in an ultrafilter
U. Let τ = {P } ∪ {cS | S ⊆ X} be a signatur with a unary relation symbol P and
constant symbols cS . Informally, our idea is to construct a theory T such that for
every model M of T we have
M |= P (cS ) ⇔ S ∈ U .
And here is such a theory.
T := {P (cA ) ⇒ P (cB ) | A ⊆ B ⊆ X}
∪ {(P (cA ) ∧ P (cB )) ⇒ P (cA∩B ) | A, B ⊆ X}
∪ {P (cA ) ⇔ ¬P (cX\A ) | A ⊆ X}
∪ {P (cA ) | A ∈ F}
Claim: Every finite T 0 ⊆ T is satisfiable. 0
T In T there are only finitely many constant0
symbols cS1 , . . . , cSn . Then there is x ∈ P (cS )∈T 0 Si because F has the FIP. Let M
i
be a τ -structure with domain P(X) and
M0
• cS := S
0
• P M (S) iff x ∈ S.
2.6. PROVING THE ULTRAFILTER LEMMA WITH COMPACTNESS 19
We have M 0 |= T 0 . The compactness theorem asserts the existence of a model M of

T . Then
U := {S ⊆ X | P M (S)}
is an ultrafilter that extends F.
CHAPTER 3
The Löwenheim-Skolem Theorem
The theorem of Löwenheim and Skolem implies that if T is a theory with infinite
models, then T has models of arbitrary infinite cardinality κ, unless the signature is
larger than κ. The proof of the theorem of Löwenheim and Skolem and has two parts,
upwards and downwards (Theorem 3.2.1). Going upwards is essentially an application
of the compactness theorem. For going downwards, we need to introduce elementary
chains.
3.1. Chains
We will need another important operation to build structures, namely the for-
mation of limits of chains. Chains of τ -structures are a fundamental concept from
model theory. Let I be a linearly ordered index set. Let (Ai )i∈I be a sequence of
τ -structures. Then (Ai )i∈I is called a chain if Ai is a substructure of Aj for all i < j.
Definition 3.1.1. The union of a chain (Ai )i∈I is a τ -structure B = limi∈I Ai
defined as follows.
S
• The domain of B is B := i∈I Ai .
• For each relation symbol R ∈ τ put a ∈ RB if a ∈ RAi for some i ∈ I.
• For each function symbol f ∈ τ put f B (a) := b if f Ai (a) = b for some i ∈ I.
Example 20. For each n ∈ N, let
An := {−n, −n + 1, . . . , 0, 1, 2, . . . } ⊆ Z
and let An := (An ; ≤) be the substructure of (Z; ≤) induced by An . Then Th(An ) =
Th(A0 ) for all n ∈ N, but Th(limi∈I Ai ) is different (why?). 4
Let A and B be τ -structures. A function f : A → B preserves a first-order τ -
formula φ(x1 , . . . , xn ) if and only if for all a1 , . . . , an ∈ A
A |= φ(a1 , . . . , an ) implies B |= φ(f (a1 ), . . . , f (an )).
Remark 3.1.2. This definition is compatible with the notion of preservation as
introduced in Section 1.1.5, in the following sense: if A0 is the τ ∪ {R}-expansion of
0
A by the relation defined by φ in A (i.e., RA = φA ), and B 0 is the τ ∪ {R} expansion
of B by the relation defined by φ in B, then f preserves φ if and only if f is a
homomorphism from A0 to B 0 .
Functions that preserve all first-order formulas are called elementary (they are
necessarily embeddings). If B is an extension of A such that the identity map from
A to B is an elementary embedding, we say that B is an elementary extension of A,
and that A is an elementary substructure of B, and write A ≺ B. Note that A ≺ B
if and only if A is a substructure of B and Th(B A ) = Th(AA ).
Example 21. The structure An+1 from Example 20 is not an elementary exten-
sion of An (why?). 4
21
22 3. THE LÖWENHEIM-SKOLEM THEOREM
Note that ≺ is a transitive relation. A chain is called an elementary chain if

Ai ≺ Aj for all i < j.
Lemma 3.1.3 (Tarski-Vaught elementary chain theorem). Let (Ai )i∈I be an ele-
mentary chain of τ -structures. Then Ai ≺ B := limi∈I Ai for each i ∈ I.
Proof. We have to show that for each i ∈ I, every τ -formula φ(x1 , . . . , xk ), and
every ā ∈ (Ai )k we have
B |= φ(ā) iff Ai |= φ(ā)
First consider the case that φ is atomic of the form R(x1 , . . . , xk ) for R ∈ τ . We
then have
B |= R(ā) iff ā ∈ RB
iff ā ∈ RAi
iff Ai |= R(ā)
Claim 1. If t(x1 , . . . , xk ) is a τ -term and ā ∈ (Ai )k , then tB (ā) = tAi (ā). This
can be shown by a straightforward induction on the structure of terms. Consequently,
the statement is true for atomic formulas of the form t = s, for τ -terms t and s.
Claim 2. If the statement is true for φ, then the statement is true for ¬φ:
B |= ¬φ(ā) iff not B |= φ(ā)
iff not Ak |= φ(ā) (by IA)
iff Ak |= ¬φ(ā)
The verification for formulas of the form φ1 ∧ φ2 is similarly straightforward. We
therefore go straight to the existential quantifier.
Claim 3. If the statement is true for φ then the statement is also true for ∃x1 . φ.
B |= ∃x1 . φ(x1 , . . . , xk )(b2 , . . . , bk )
iff there is b1 ∈ B such that B |= φ(b1 , b2 , . . . , bk )
iff there is b1 ∈ Aj such that Aj |= φ(b1 , b2 , . . . , bk ) for some j > i
iff Aj |= ∃x1 . φ(b2 , . . . , bk )
iff Ai |= ∃x1 . φ(b2 , . . . , bk ) since Ai ≺ Aj .
We say that a τ -formula φ(x1 , . . . , xn ) is satisfiable in a τ -structure B if there
are elements b1 , . . . , bn such that B |= φ(b1 , . . . , bn ); in this case we also say that φ is
satisfied by b1 , . . . , bn in B.
Lemma 3.1.4 (Tarski’s Test). Let B be a τ -structure and A ⊆ B. Then A is the
domain of an elementary substructure of B if and only if every (τ ∪ A)-formula φ(x)
which is satisfiable in B A can be satisfied by an element of A.
Proof. First suppose that A ≺ B and that B A |= φ(b) for some b ∈ B. Then
B A |= ∃x.φ(x), hence AA |= ∃x.φ(x). So there exists a ∈ A such that AA |= φ(a).
Thus B A |= φ(a).
Conversely, suppose that the condition of Tarski’s test is satisfied. To verify that
A is the domain of a substructure A of B, let a1 , . . . , ak ∈ A and f ∈ τ . Then
x = f (a1 , . . . , ak ) is satisfiable in B A , and hence there is an a ∈ A with B A |= a =
f (a1 , . . . , ak ). To show that A is an elementary substructure of B it suffices to show
that AA and B A satisfy exactly the same first-order sentences. We prove this by
induction over the length of first-order sentences. The statement is clear for atomic
sentences. The induction steps for ψ = ¬φ and ψ = (φ1 ∧ φ2 ) follow immediately
from the inductive assumption for the shorter sentences φ and φ1 , φ2 . It remains
3.1. CHAINS 23
to consider the case ψ = ∃x.φ(x). If ψ holds in AA , there exists a ∈ A such that

AA |= φ(a). The induction hypothesis yields B A |= φ(a), thus B A |= ψ. For the
converse suppose that B A |= ψ. Then φ(x) is satisfiable in B A and by assumption we
find a ∈ A such that B A |= φ(a). By induction AA |= φ(a), and hece AA |= ψ.
We use Tarski’s Test to construct small elementary substructures.

Corollary 3.1.5. Let S be a subset of a τ -structure B. Then B has an elemen-
tary substructure A containing S and of cardinality at most max(|S|, |τ |, ℵ0 ).
Proof. We construct A as the union of an ascending sequence S = S0 ⊆ S1 ⊆ · · ·
of subsets of B. Suppose that Si is already defined. Choose an element aφ ∈ B
for every (τ ∪ Si )-formula φ(x) which is satisfiable in B Si , and define Si+1 to be
Si together with these aφ . Then Tarski’s Test implies that A is the domain of an
elementary substructure. It remains to prove the bound on |A|. There are at most
max(|τ |, ℵ0 ) many τ -formulas (Corollary A.4.3). Let κ = max(|S|, |τ |, ℵ0 ). There are
κ many (τ ∪ S)-formulas; therefore |S1 | ≤ κ. Inductively it follows for every i that
|Si | ≤ κ. Finally we have |A| ≤ κ · ℵ0 = κ.
Exercises.
(28) Suppose that A is an elementary substructure of B, and B is an elementary
extension of C, and A ⊆ C. Is A an elementary substructure of C?1
(29) Let A, B sets with A ⊆ B and suppose that A is infinite. Show that A :=
(A; =) is an elementary substructure of B := (B; =).
Solution. We have to show that idA : A → B is an elementary embed-
ding of A into B, i.e., preserves all first-order formulas: if φ(x1 , . . . , xn ) is a
first-order formula and a1 , . . . , an ∈ A, then A |= φ(a1 , . . . , an ), if and only if
B |= ψ(a1 , . . . , an ). We show this by induction on the size of φ. If φ is atomic,
it must be of the form x = y, and in this case the statement is clear. The
inductive step is clear if φ is of the form ¬ψ or of the form φ1 ∧φ2 . The inter-
esting case is that φ is of the form ∃y.ψ(y, x1 , . . . , xn ). If A |= φ(a1 , . . . , an )
then clearly B |= φ(a1 , . . . , an ). Conversely, suppose that B |= φ(a1 , . . . , an ).
Then there exists b ∈ B such that B |= ψ(b, a1 , . . . , an ). Since A is infinite
there exists c ∈ A \ {a1 , . . . , an }. Note that (B, a1 , . . . , an ) has an automor-
phism that exchanges b to c and fixes all other elements. The automorphism
shows that A |= ψ(c, a1 , . . . , an ), and hence A |= φ(a1 , . . . , an ). This con-
cludes the induction.
(30) Is (Q; <) an elementary substructure of (R; <)?
(31) Is (Z; <) an elementary substructure of (Q; <)?
(32) (Exercise 2.2.4 in [29]) Let A = (R; 0, <, f A ) where f is a unary function
symbol. Let B be an elementary extension of A. Call b ∈ B an infinitesimal
if − n1 < x < 21 for all n ∈ N. Show that if f A (0) = 0, then f A is continuous
in 0 if and only if for every elementary extension B of A the map f B maps
infinitesimals to infinitesimals.
(33) A formula is called existential if it is built from a quantifier-free formula by
existential quantification (universal quantification is forbidden). Show that
the class of all models of an existential sentence is closed under extensions.
(34) A formula is called ∀∃ if it is built from an existential formula by universal
quantification. Show that the class of all models of a ∀∃ sentence is closed
under unions of chains.
1Thanks to Anna Tölle for the idea for this exercise.

24 3. THE LÖWENHEIM-SKOLEM THEOREM
(35) Let φ be a formula in prenex normal form (see Exercise 11) whose quantifier-
free part is written in conjunctive normal form (see Exercise 11). Then φ
is called a Horn formula if each conjunct has at most one disjunct which
is positive, i.e., an atomic formula. In other words, all but one of the dis-
juncts of each conjunct must be of the form ¬ψ for some atomic formula ψ.
Prove that the class of all models of a Horn sentence is closed under taking
products.
3.2. Proving Löwenheim-Skolem

Theorem 3.2.1 (Löwenheim-Skolem). Let A be an infinite τ -structure, S ⊆ A,
and κ an infinite cardinal.
• (Downwards) If max(|S|, |τ |) ≤ κ ≤ |A| then A has an elementary substruc-
ture of cardinality κ containing S.
• (Upwards) If max(|A|, |τ |) ≤ κ then A has an elementary extension of car-
dinality κ.
Proof. Downwards: choose a set S 0 ⊆ A that contains S and has cardinality κ,
and apply Corollary 3.1.5.
Upwards: we first construct an elementary extension A0 of A of cardinality at
least κ. Choose a set ρ of new constants of cardinality κ. As A is infinite, all finite
subsets of the theory T := Th(AA )∪{¬(c = d) | c, d ∈ ρ, c 6= d} are satisfiable. By the
compactness theorem (Theorem 2.0.1) T has a model A00 , which must have cardinality
at least κ. Clearly, the τ -reduct A0 of A00 is an elementary extension of A. Finally
we apply downwards Löwenheim-Skolem to A0 and S = A and obtain an elementary
substructure of A0 of cardinality exactly κ, which is an elementary extension of A (see
Exercise 28).
Note that in part (1), the assumption κ ≥ max(|S|, |τ |) is certainly necessary in
general.
Corollary 3.2.2. A theory which has an infinite model has a model in every
cardinality κ ≥ max(|τ |, ℵ0 ).
3.3. Vaught’s test

The theorem of Löwenheim-Skolem implies that no theory with an infinite model
can describe this model up to isomorphism. The best we can hope for is a unique
model for a given cardinality.
Definition 3.3.1. A theory T is called κ-categorical if all models of T of cardi-
nality κ are isomorphic.
Note that the theory of a finite structure does not have infinite models, and hence
is κ-categorical for every infinite κ. In Example 15 we have seen that Th(Q; <) is
ℵ0 -categorical. A second important example of an ℵ0 -categorical theory is the theory
of the countable random graph (V; E).
Example 22. The countable random graph is the (simple and undirected) graph
that has the following extension property: for all finite disjoint subsets U, U 0 of V there
exists a vertex v ∈ V \ (U ∪ U 0 ) such that v is adjacent to all vertices in U and to no
vertex in U 0 . The existence of such a graph will be shown in the following chapter. We
claim that this graph has an ω-categorical theory: note that the extension property of
(V; E) given above is a first-order property; a back-and-forth argument similar to the
one given Example 15 shows that every countably infinite graph with this property
is isomorphic to (V; E). 4
3.3. VAUGHT’S TEST 25
Theorem 3.3.2 (Vaught’s Test). A κ-categorical theory T is complete if the fol-

lowing conditions are satisfied:
• T is consistent;
• T has no finite model;
• |τ | ≤ κ.
Proof. We have to show that all models A, B of T have the same theory. Since
A and B must be infinite, by Corollary 3.2.2 the theory Th(A) and Th(B) must have
models A0 and B 0 of cardinality κ. By κ-categoricity of T , we have Th(A0 ) = Th(B 0 ),
hence Th(A) = Th(A0 ) = Th(B 0 ) = Th(B).
Exercises.
(36) Is (Z; <) ℵ0 -categorical?
CHAPTER 4
Fraïssé Amalgamation
A relational1 structure A is called homogeneous (sometimes also called ultra-

homogeneous [19]) if every isomorphism between finite substructures of A can be
extended to an automorphism of A. We have already encountered an example of a
homogeneous structure.
Example 23. The back-and-forth argument we presented in Example 15 shows in
particular that isomorphisms between finite substructures of (Q; <) can be extended
to automorphisms of (Q; <). Thus, (Q; <) is homogeneous. 4
A versatile tool to construct countable homogeneous structures from classes of
finite structures is the amalgamation technique à la Fraïssé. We present it here for
the special case of relational structures; this is all that is needed in the examples we
are going to present. For a stronger version of Fraïssé-amalgamation for classes of
structures that might involve function symbols, see [19].
4.1. The Age of a Structure

In the following, let τ be a countable relational signature. The age of a τ -structure
A is the class of all finite τ -structures that embed into A. A class C has the joint
embedding property (JEP) if for any two structures B 1 , B 2 ∈ C there exists a structure
C ∈ C that embeds both B 1 and B 2 .
Proposition 4.1.1. Let C be a class of τ -structures. Then C is the age of a
(countable) relational structure if and only if C is closed under isomorphisms and
substructures, has the JEP, and contains countably many isomorphism classes.
Proof. Let C 0 , C 1 , . . . be a set of representatives for each isomorphism class in
C. Construct a chain A0 , A1 , . . . as follows. A0 := C 0 . If Ai has been chosen, find
B ∈ C such that Ai ,→ B and C i+1 ,→ B. Let Ai+1 be an extension of Ai isomorphic
to B. Then the age of A = limi∈N Ai equals C.
4.2. Amalgamation Classes

A class C of τ -structures has the amalgamation property (AP) if for all A, B 1 , B 2 ∈
C and embeddings e1 : A ,→ B 1 and e2 : A ,→ B 2 there are a C ∈ C and embeddings
fi : B i ,→ C for i ∈ {1, 2} such that
f1 ◦ e1 = f2 ◦ e2 .
To verify the amalgamation property of classes of τ -structures that are closed
under isomorphism, a slightly different perspective is sometimes convenient. The
union of two relational τ -structures B 1 , B 2 is the τ -structure C with domain B1 ∪ B2
and relations RC := RB 1 ∪ RB 2 for every R ∈ τ . The intersection of B 1 and B 2 is
defined analogously. Let B 1 , B 2 be τ -structures such that A is a substructure of both
1The entire theory can be adapted to general signatures that might also contain function sym-
bols; to keep the exposition simple, we restrict our focus to relational signatures in this section.
27
28 4. FRAÏSSÉ AMALGAMATION
B 1 and B 2 and all common elements of B 1 and B 2 are elements of A; note that in
this case A = B 1 ∩ B 2 . Then we call B 1 ∪ B 2 the free amalgam of B 1 , B 2 (over A).
More generally, a τ -structure C is an amalgam of B 1 and B 2 over A if for i ∈ {1, 2}
there are embeddings fi of B i to C such that f1 (a) = f2 (a) for all a ∈ A. We refer
to (A, B 1 , B 2 ) as an amalgamation diagram.
Definition 4.2.1. An isomorphism-closed class C of τ -structures
• has the free amalgamation property if for any B 1 , B 2 ∈ C the free amalgam
of B 1 and B 2 is contained in C;
• has the amalgamation property if every amalgamation diagram (A, B 1 , B 2 )
has an amalgam C ∈ C;
• is an amalgamation class if C is countable up to isomorphism, has the amal-
gamation property, and is closed under taking substructures.
Note that since we only look at relational structures here (and since we allow
structures to have an empty domain), the amalgamation property of C implies the
joint embedding property.
Example 24. The class of all finite graphs has AP: an amalgam of B 1 and B 2 is
C := B 1 ∪ B 2 (and fi : B i ,→ C the identity). Similarly, for every relational signature
τ , the class of all finite τ -structures has AP. 4
A systematic source of amalgamation classes is the following proposition.
Proposition 4.2.2. Let B be a homogeneous relational structure. Then Age(B)
is an amalgamation class.
Proof. Let A, B 1 , B 2 ∈ Age(B) so that for i ∈ {1, 2} there are embeddings
ei : A ,→ B i . By definition, there are embeddings gi : B i ,→ B. Let A0 be the sub-
structure of B with domain g1 (e1 (A)). Then the restriction of g2 ◦e2 ◦e−1 −1
1 ◦g1 to A is
0
an embedding of A into B, and by the homogeneity of B can be extended to an auto-
morphism a of B. But then the substructure C of B with domain a(g1 (B1 ))∪g2 (B2 )
is Age(B) and a ◦ g1 : B 1 → B and g2 : B 2 → B are embeddings of B 1 and B 2 into C
satisfying a ◦ g1 (e1 (a)) = g2 (e2 (a)) for all a ∈ A, showing AP.
Since we have already seen that (Q; <) is homogeneous, the age of (Q; <), which
is the class of all finite linear orders, has AP; this can also bee seen directly.
4.3. Fraïssé’s theorem

We prove a converse to Proposition 4.2.2.
Theorem 4.3.1 (Fraïssé [12,13]; see [19]). Let τ be a countable relational signa-
ture and let C be an amalgamation class of τ -structures. Then there is a homogeneous
and at most countable τ -structure C whose age equals C. The structure C is unique
up to isomorphism, and called the Fraïssé-limit of C.
Proof. We first prove uniqueness. This can be shown by a back and forth argu-
ment, generalising the argument in Example 15. Let C and D be countable homoge-
neous of the same age. We have to show that C and D are isomorphic, and construct
the isomorphism f : C → D by a back-and-forth argument, similarly to the proof that
(Q; <) is ω-categorical in Example 15. Let C = {c1 , c2 , . . . } and D = {d1 , d2 , . . . }.
Suppose f is already defined on a finite subset F of C.
• Going forth: Let i ∈ N be smallest so that ci ∈ / F . Then there is e : C[F ∪
{ci }] ,→ D. Note that the finite structures D[e(F )] and D[f (F )] are isomor-
phic via f ◦e−1 . By the homogeneity of D, this isomorphism can be extended
4.3. FRAÏSSÉ’S THEOREM 29
to an automorphism a of D. Then a ◦ e = f and we extend f by setting

f (ci ) := a(e(ci )). Clearly, the extension of f thus defined is an embedding
of C[F ∪ {ci }] into D because it is the composition of the embedding e with
the automorphism a.
• Going back: let i ∈ N be smallest so that di ∈ / f [F ]. Analogously, find c ∈ C
so that the extension f (c) := di is an isomorphism.
A structure C is called weakly homogenous if for all B ∈ Age(C), substructure A of
B, and e : A ,→ C there is g : B ,→ C which extends e. Note that in the proof above,
we only needed weak homogeneity of C and D to construct f .
We now prove the existence of the homogeneous structure C from the statement
of the theorem. The structure C can be constructed as a union of a chain (C i )i∈N of
structures in C i ∈ C such that if A, B ∈ C, with A substructure of B and e : A ,→ C i
for some i ∈ N, then there are j ∈ N and g : B ,→ C j which extends e. Note that C
is weakly homogeneous, and hence homogeneous, by the comments above.
Also note that Age(limi C i ) = C. Here, the inclusion ⊆ is clear. For the converse
inclusion, first note that for every A ∈ C there is B ∈ C such that A ,→ B and C 0 ,→ B
by the JEP. So B ,→ C j for some j ∈ N, and hence A ,→ B ,→ limi C i .
Let P be a countable set of representatives for all (A, B) ∈ C 2 such that A is
a substructure of B. Let α : N2 → N be a bijection such that α(i, j) ≥ i for all
i, j. Suppose C k already constructed. Let (Ak,i , B k,i , fk,i )i∈N be a list of all triples
(A, B, f ) where
• (A, B) ∈ P and
• f : A ,→ C k .
Let i, j be such that k = α(i, j). Construct C k+1 as amalgam of C k and B i,j so that
fi,j extends to B i,j ,→ C k+1 .
Example 25. Let C be the class of all finite partially ordered sets. Amalgamation
can be shown by computing the transitive closure: when C is the free amalgam of
B 1 and B 2 over A, then the transitive closure of C gives an amalgam in C. The
Fraïssé-limit of C is called the homogeneous universal partial order. 4
Example 26. Let C be the class of all finite undirected graphs. It is even easier
than in the previous examples to verify that C is an amalgamation class, since here the
free amalgam itself shows the amalgamation property. The Fraïssé-limit of C is also
known as the countable random graph, or also the Rado graph, and will be denoted
by (V; E). 4
Example 27. Let C be the class of all finite triangle-free graphs, that is, all finite
undirected graphs that do not contain K3 as a subgraph. Again, we have the free
amalgamation property. The Fraïssé-limit is up to isomorphism uniquely described
as the triangle-free graph A such that for any finite disjoint S, T ⊂ A such that S
is stable (i.e., induces a graph with no edges; such a vertex subset is sometimes also
called an independent set) there exists v ∈ A \ (S ∪ T ) which is connected to all points
in S, but to no point in T . 4
We now introduce a convenient notation to describe classes of finite τ -structures.
When N is a class of τ -structures, we say that a structure A is N -free if no B ∈ N
embeds into A. The class of all finite N -free structures we denote by Forb(N ).
Example 28. Henson [16] used Fraïssé limits to construct 2ω many homogeneous
directed graphs. A tournament is a directed graph without self-loops such that for
all pairs x, y of distinct vertices exactly one of the pairs (x, y), (y, x) is an arc in the
graph. Note that for all classes N of finite tournaments, Forb(N ) is an amalgamation
class, because if A1 and A2 are directed graphs in Forb(N ) such that A = A1 ∩ A2 is

a substructure of both A1 and A2 , then the free amalgam A1 ∪ A2 is also in Forb(N ).
Henson specified an infinite set T of tournaments T 1 , T 2 , . . . with the property
that T i does not embed into T j if i 6= j. Draw the family on the board. Note that this
property implies that for two distinct subsets N1 and N2 of T the two sets Forb(N1 )
and Forb(N2 ) are distinct as well. Since there are 2ω many subsets of the infinite
set T , there are also that many distinct homogeneous directed graphs; they are often
referred to as Henson digraphs. 4
Exercises.
(37) Which of the following classes of finite structures are amalgamation classes?
(a) The class of all finite directed graphs?
(b) The class of all finite equivalence relations?
(c) The class of all finite left-linear orders? A partial order (P ; ≤) is called
left-linear if for every x the set {y | y ≤ x} is linearly ordered by ≤.
(d) The class of all finite 3-uniform hypergraphs? One way to formally
define 3-uniform hypergraphs is to view them as relational structures
with a single ternary relation R which only contains triples with pair-
wise distinct elements and is fully symmetric, i.e., if (a, b, c) ∈ R then
(π(a), π(b), π(c)) ∈ R for every permutation π of {a, b, c}.
(e) The class of all finite 3-uniform hypergraphs that do not embed a
tetrahedron: a tetrahedron is a 3-uniform hypergraph with vertices
{1, 2, 3, 4} whose edge relation contains all triples of pairwise distinct
elements.
(38) The complement of a graph (V ; E) is the graph
V ; V 2 \ (E ∪ {(a, a) | a ∈ V }) .

Show that the complement of a homogeneous graph is homogeneous.

(39) Let (V; E) be the Rado graph (Example 26). Show that if we remove finitely
many vertices from (V; E), the resulting graph is isomorphic to the Rado
graph. Formally: show that for every finite F ⊆ V, the subgraph of (V; E)
induced by V \ F is isomorphic to (V; E).
(40) Let (V; E) and F be as in the previous exercise. Let (V; E 0 ) be the graph
obtained from (V; E) by flipping edges and non-edges between F and V \ E.
Formally,
E 0 := E \ {(a, b) ∈ E | a ∈ F if and only if b ∈
/ F}
∪ {(a, b) ∈
/ E | a ∈ F if and only if b ∈
/ F }.
Show that (V; E 0 ) is isomorphic to (V; E).
(41) Show the age C of a structure has the amalgamation property if and only if
it has the 1-point amalgamation property, i.e., if for all A, B 1 , B 2 ∈ C and
embeddings e1 : A ,→ B 1 and e2 : A ,→ B 2 such that |B1 | = |B2 | = |A| + 1
there are a C ∈ C and embeddings fi : B i ,→ C for i ∈ {1, 2} such that
f1 ◦ e1 = f2 ◦ e2 .
4.4. Strong Amalgamation

A strong amalgam of B 1 , B 2 over A is an amalgam of B 1 , B 2 over A where
f1 (B1 ) ∩ f2 (B2 ) = f1 (A)(= f2 (A)) (as in Section 4.2). A class C has the strong
amalgamation property if all amalgamation diagrams have a strong amalgam in C.
Examples:
• The class of all finite graphs.
4.4. STRONG AMALGAMATION 31
• The age of (Q; <).

The following simple example shows an amalgamation class that does not have
the strong amalgamation property.
Example 29. Let P be a unary relation symbol, and let C be the class of all finite
{P }-structures where P contains at most one element. Then C is an amalgamation
class, and the Fraïssé-limit is a countably infinite structure where P contains only
one element. But C does not have the strong amalgamation property. 4
For strong amalgamation classes there is a powerful construction to obtain new
strong amalgamation classes from known ones.
Definition 4.4.1. Let C1 and C2 be classes of finite structures with disjoint
relational signatures τ1 and τ2 , respectively. Then the generic superposition of C1 and
C2 , denoted by C1 ∗ C2 , is the class of (τ1 ∪ τ2 )-structures A such that the τi -reduct of
A is in Ci , for i ∈ {1, 2}.
The following lemma has a straightforward proof by combining amalgamation in
C1 with amalgamation in C2 .
Lemma 4.4.2. If C1 and C2 are strong amalgamation classes, then C1 ∗ C2 is also
a strong amalgamation class.
When A1 and A2 are homogeneous structures whose ages have strong amalgama-
tion, then A1 ∗ A2 denotes the (up to isomorphism unique) Fraïssé-limit of the generic
superposition of the age of A1 and the age of A2 .
Proposition 4.4.3. For i = 1 and i = 2, let Ai be a homogeneous τi -structure
whose age has strong amalgamation. Then the τi -reduct of A1 ∗ A2 is isomorphic to
Ai .
Proof. A back-and-forth argument.
Example 30. For i ∈ {1, 2}, let τi = {<i }, let Ci be the class of all finite τi -
structures where <i denotes a linear order, and let Ai be the Fraïssé limit of Ci . Then
A1 ∗ A2 is known as the random permutation (see e.g. [8]). 4
Exercises.
(42) Let D be the tournament obtained from the directed cycle C ~ 3 of length three
by adding a new vertex u, and adding the edges (u, v) for every vertex v of
~ 3 . Let D0 be the tournament obtained from D by flipping the orientation
C
of each edge. Show that Forb({D, D0 }), the class of all finite tournaments
that embeds neither D nor D0 , is an amalgamation class.
(43) Let P be a unary relation symbol. Let D be the class of all finite {P, <}-
structures A such that <A is a linear order.
(a) Show that D is an amalgamation class.
(b) Let B be the Fraïssé-limit of the class D, and define E ⊆ B 2 by (u, v) ∈
E if
• u < v and (u ∈ P ⇔ v ∈ P ), or
• u > v and not (u ∈ P ⇔ v ∈ P ).
Show that (B; E) is a tournament.
(c) Show that the class Age(B; E) equals the class of tournaments that
can be obtained from tournaments T in Age(Q; <) by performing the
following operation: pick u ∈ T and reverse all edges between u and
other elements of T (we ‘switch edges at u’).
(d) Show that (B; E) is homogeneous.
(e) Show that Age(B; E) equals the class C from Exercise 42.
(f) Show that (B; E) is isomorphic to the tournament whose vertices are
a countable dense subset S ⊆ R2 of the unit circle without antipodal
points, and where the edges are oriented in clockwise order, i.e., put
((u1 , u2 ), (v1 , v2 )) ∈ E if and only if u1 v2 − u2 v1 > 0.
(44) Show that there are permutation groups G1 , G2 on a countably infinite set
such that both G1 and G2 are isomorphic (as permutation groups) to (Q; <),
but G1 ∩ G2 = {id}.
(45) Construct a permutation group G on a set X with precisely n! orbits of
n-element subsets.
(46) Show that the random graph can be partitioned into two subsets so that
both parts are isomorphic to the random graph. Show that the same is not
true for all partitions of the random graph into two infinite subsets.
(47) Show that for every partition of the vertices of the Rado graph into two
subsets, one of the two subsets induces a subgraph of the Rado graph which
is isomorphic to the Rado graph.
(48) A permutation group G on a set D is called n-transitive if the componentwise
action of G on Dn has precisely one orbit. Construct for every n ∈ N a
permutation group on D which is n-transitive, but not (n + 1)-transitive.
(49) Give an example of a homogeneous structure with a transitive automorphism
group whose age does not have strong amalgamation.
(50) Let A be a homogeneous structure with finite relational signature. Show
that the following are equivalent.
(a) There exists a structure B with finite relational signature such that
Aut(B) = Aut(A).
(b) There exists a relation R ⊆ An such that Aut(A) = Aut(A, R).
(c) There exists a structure A with finite relational signature such that
G = Aut(A), and all relations in A have pairwise distinct entries.
(d) There exists a relation R ⊆ An such that Aut(B) = Aut(A, R) and R
has pairwise distinct entries.
Which implications between these items require the homogeneity of A?
4.5. Further Examples: C-relations

Let T be a tree with vertex set T and with a distinguished vertex r, the root of
T . For u, v ∈ T , we say that u lies below v if the path from r to u passes through v.
We say that u lies strictly below v if u lies below v and u 6= v. The youngest common
ancestor (yca) of two vertices u, v ∈ T is the node w such that both u and v lies below
w and w has maximal distance from r.
Let T be the class of all finite rooted binary trees T . The leaf structure C of a
tree T ∈ T with leaves L is the relational structure (L; |) where | is a ternary relation
symbol, and ab|c holds in C iff yca(a, b) lies below yca(b, c) in T (recall that yca(a, b)
denotes the youngest common ancestor of a and b in a rooted tree T . We also call T
the underlying tree of C.
Observation 4.5.1. Any T ∈ T can be recovered uniquely from its leaf structure.
Let C be the class of all leaf structures for trees from T .
Proposition 4.5.2. The class C is an amalgamation class.
Proof. Closure under isomorphisms is by definition. Closure under substruc-
tures is easy to verify. For the amalgamation property, let B 1 , B 2 ∈ C be such that
A = B 1 ∩ B 2 is a substructure of both B 1 and B 2 . We want to show that there is an
4.5. FURTHER EXAMPLES: C-RELATIONS 33
amalgam of B 1 and B 2 over A in C. We inductively assume that the statement has

been shown for all triples (A, B 01 , B 02 ) where B10 ∪ B20 is a proper subset of B1 ∪ B2 .
Let T 1 be the rooted binary tree underlying B 1 , and T 2 the rooted binary tree
underlying B 2 . Let B 11 ∈ C be the substructure of B 1 induced by the vertices below
the left child of T 1 , and B 21 ∈ C be the substructure of B 1 induced by the vertices
below the right child of T 1 . The structures B 12 and B 22 are defined analogously for
B 2 instead of B 1 .
B1
(B1)1 (B1)2
u v
u v
(B2)1 (B2)2
B2
Figure 4.1. Illustration for the proof of Proposition 4.5.2.
First consider the case that there is a vertex u that lies in both B 11 and B 12 , and a
vertex v that lies in both B 12 and B 21 (see Figure 4.1 for an illustration). We claim that
in this case no vertex w from B 22 can lie inside B 1 : for otherwise, w is either in B 11 ,
in which case we have uw|v in B 1 , or in B 21 , in which case we have vw|u in B 1 . But
since u, v, w are in A, this is in contradiction to the fact that uv|w holds in B 2 . Let
C 0 ∈ C be the amalgam of B 1 and B 12 over A, which exists by inductive assumption,
and let T 0 ∈ T be its underlying tree. Now let T be the tree with root r and T 0 as
a left subtree, and the underlying tree of B 22 as a right subtree. It is straightforward
to verify that the leaf structure of T is in C, and that it is an amalgam of B 1 and B 2
over A (via the identity embeddings).
Up to symmetry, the only remaining essentially different case we have to consider
is that B11 ∪ B21 and B12 ∪ B22 are disjoint. In this case it is similarly straightforward
to first amalgamate B 11 with B 12 and B 21 with B 22 to obtain the amalgam of B 1 and
B 2 ; the details are left to the reader.
Let B be the Fraïssé-limit of C. The relation | in B is closely related to so-
called C-relations, following the terminology of [1]. C-relations became an important
concept in model theory, see e.g. [15]. They are given axiomatically; the presentation
here follows [1, 4].
A ternary relation C is said to be a C-relation on a set L if for all a, b, c, d ∈ L
the following conditions hold:
C1 C(a; b, c) ⇒ C(a; c, b);
C2 C(a; b, c) ⇒ ¬C(b; a, c);
C3 C(a; b, c) ⇒ C(a; d, c) ∨ C(d; b, c);
C4 a 6= b ⇒ C(a; b, b).
A C-relation is called proper if it satisfies
C5 ∀b, c∃a : C(a; b, c);

C6 ∀a, b a 6= b ⇒ ∃c(b 6= c ∧ C(a; b, c) .
A C-relation is called
C7 dense if it satisfies C(a; b, c) ⇒ ∃e (C(e; b, c) ∧ C(a; b, e)).
C8 binary branching if it satisfies

∀x, y, z (x 6= y ∨ x 6= z ∨ y 6= z) ⇒ (C(x; y, z) ∨ C(y; x, z) ∨ C(z; x, y)) .
Proposition 4.5.3. Every dense binary branching proper C-relation is isomor-
phic to the structure (B; {(a, b, c) ∈ B 3 : a|bc ∨ (a 6= b ∧ b = c)}) (where B is the
Fraïssé-limit of the class C as introduced above).
Proof. A back-and-forth argument.
Exercises.
(51) Show that the class of all finite forests (i.e., undirected graphs without cycles)
is not an amalgamation class.
(52) Prove Observation 4.5.1.
(53) Prove Proposition 4.5.3.
CHAPTER 5
Types
Loosely speaking, a type of a τ -structure M is a set of formulas that is satisfied

by a real or by a ‘virtual’ element of M , that is, an element of some structure that
has the same theory as M .
A (not necessarily finite) set Σ of formulas with free variables x1 , . . . , xn is called
satisfiable over a structure A, or realised in A, if there are elements a1 , . . . , an of A
such that for all sentences φ ∈ Σ we have A |= φ(a1 , . . . , an ). We say that Σ is
satisfiable if there exists a structure A such that Σ is satisfiable over A.
Lemma 5.0.1. A set Σ of formulas with free variables x1 , . . . , xn is satisfiable if
and only if all finite subsets of Σ are satisfiable.
Proof. Introduce new constant symbols c1 , . . . , cn . Then Σ is satisfiable if and
only if Σ(c1 , . . . , cn ) := {φ(c1 , . . . , cn ) | φ(x1 , . . . , xn ) ∈ Σ} is satisfiable. Now apply
the compactness theorem.
For n ≥ 0, an n-type of a theory T is a set p of formulas with free variables
x1 , . . . , xn such that p ∪ T is satisfiable. An n-type of a structure A is an n-type of the
first-order theory of A. Note that an n-type p of a theory T might or might not be
realised in models A of T ; if p is not realised in A, then we also say that p is omitted
in A.
Lemma 5.0.2. Let A be a τ -structure and Σ a set of first-order τ -formulas with
free variables x1 , . . . , xn . Then the following are equivalent.
(1) Σ is an n-type of A;
(2) every finite subset of Σ is realised in A;
(3) A has an elementary extension that realises Σ.
Proof. (3) ⇒ (1): immediate.
(1) ⇒ (2): Let Σ be an n-type of A, i.e., there exists a model B of Th(A) andVb̄ ∈
B n such that B |= Σ(b̄). Hence, if Ψ is a finite subsetVof Σ, then B |= ∃x1 , . . . , xn Ψ
and since Th(A) = Th(B) we have A |= ∃x1 , . . . , xn Ψ, so Ψ is realised in A.
(2) ⇒ (3): suppose that every finite subset Ψ of Σ is realised in A. Then in
particular every finite subset of Σ ∪ Th(AA ) is satisfiable, and hence Σ ∪ Th(AA ) has
a model B by compactness. Then the τ -reduct of B is an elementary extension of A
that realises Σ.
An n-type p of a τ -theory T is maximal, or complete, if T ∪ p ∪ {φ(x1 , . . . , xn )}
is unsatisfiable for any τ -formula φ ∈ / T ∪ p. In other words, for every τ -formula
φ(x1 , . . . , xn ), either φ ∈ p or ¬φ ∈ p. We write Sn (T ) for the set of all complete
n-types of T . The set of all first-order formulas with free variables x1 , . . . , xn satisfied
by an n-tuple ā = (a1 , . . . , an ) in A is a maximal type of A, and called the type of ā
in A.
Example 31. The structure (Q; <) has precisely one complete 1-type and pre-
cisely three complete 2-types. This follows easily from the homogeneity of (Q; <)
(Exercise 55 and Example 23; also see Exercise 55). 4
35
36 5. TYPES
Exercises.
(54) Let κ be an infinite cardinal. Show that a structure A realises all 1-types
over all B ⊆ A with |B| < κ if and only if A realises all n-types over all
B ⊆ A with |B| < κ.
(55) Let A, B be a τ -structures and f an isomorphism between A and B. Then
every tuple ā ∈ An has the same type in A as f (ā) in B.
(56) Prove or disprove:
• The set {x > n | n ∈ N} is a type of (Z, <)N .
• The set {x > 0} ∪ x < n | n ∈ {1, 2, 3, . . . } is a type of (Z, <)N .
• The set {x > n | n ∈ N} is realised (Q, <)N .
• The set {x > 0} ∪ x < 1/n | n ∈ {1, 2, 3, . . . } is a type of (Q, <)Q .
(57) Show that (R; 0, +) has exactly two complete 1-types and ℵ0 many complete
2-types, and that (R; 0, +, <) has exactly three complete 1-types and 2ℵ0
many complete 2-types.
Solution to Exercise 57.

(a) The elements 0 and 1 clearly have a different 1-type in (R; 0; +). We claim
that every complete 1-type of (R; 0, +) that contains the formula x1 > 0 is
realised by 1. Let φ ∈ p. By Lemma 5.0.2 the type {φ(x1 ), x1 > 0} ⊂ p is
realised by some element a of (R; 0, +). Then x 7→ x/a is an automorphism
of (R; 0; +) and takes a to 1, showing that φ also holds in 1.
(b) For any rational number r/s ∈ Q, where r ∈ Z and s ∈ N \ {0}, the pair
(r, s) realises a maximal 2-type pr,s . Note that pr,s contains the formula
(1) x1 + · · · + x1 = x2 + · · · + x2 .
| {z } | {z }
s times r times
If pu,v , for u ∈ Z and v ∈ N\ {0}, also contains this formula, then s·u = r ·v,
so r/s = u/v. This implies that there are at least |Q| = ℵ0 many complete
2-types.
Then there is the complete 2-type of (0, 0) and the complete 2-type of
(0, 1); they are clearly distinct and distinct from all the types of the form
pr,s for some r ∈ Z and s ∈ N \ {0}. I claim that there is exactly one more
complete 2-type of (R; 0; +), namely the type of (1, i) where i is any irrational
number. To see this, view R as a Q-vector space (of dimension 2ℵ0 ); we may
choose a basis B1 that contains the element i. Let p be a complete 2-type
(R; 0; +) which does not contain the formula x1 = 0, does not contain the
formula x2 = 0, and does not contain (1) for any r, s ∈ N. By Lemma 5.0.2,
p is realised by an element a of an elementary extension R of (R; 0; +); by
the theorem of Löwenheim-Skolem we may choose an elementary extension
of cardinality 2ℵ0 . The elementary extension can be viewed as a Q-vector
space of dimension 2ℵ0 ; using the axiom of choice, we may choose a basis B2
that contains a. Then there exists a bijection between the basis elements B1
and B2 that extends to a vector space isomorphism between (R; 0; +) and
R which shows that i satisfies p.
(c) The structure (R; 0, +, <) has the three complete 1-types realised by −1,
0, and 1, and the argument that there are all the complete 1-types can be
shown similarly as in part (a) of this solution.
(d) There are clearly at most 2ℵ0 many 2-types in (R; 0, +, <) because the sig-
nature is countable. For a ∈ R, let pa be the 2-type of the pair (a, 1) in
(R; 0, +, <). I claim that if a, b ∈ R are such that a 6= b, then pa 6= pb . It
then follows that there are |R| = 2ℵ0 many complete 2-types in (R; 0, +, <).
5.2. SATURATED STRUCTURES 37
Without loss of generality, suppose that a < b. Choose a rational number

r/s, for r ∈ Z and s ∈ N \ {0}, such that a < r/s < b. Then
x1 + · · · + x1 < x2 + · · · + x2
| {z } | {z }
s times r times
holds for (a, 1), but not for (b, 1).
5.1. Stone Spaces

There is a natural topology on the set Sn (T ) of complete n-types of a τ -theory
T . For a τ -formula φ(x1 , . . . , xn ), define
[φ] := {p ∈ Sn (T ) | φ ∈ p}.
The Stone topology on Sn (T ) is the topology generated by taking the sets [φ] as basic
open sets (also recall Remark 2.0.2). Note that for complete types p, exactly one of φ
and ¬φ is in p; hence, [φ] = Sn (T ) \ [¬φ] is also closed; we refer to sets that are both
closed and open as clopen.
Lemma 5.1.1. The Stone topology on Sn (T ) is compact and totally disconnected.
Proof. To show compactness, let C = {[φi ] | i ∈ I} be a cover of Sn (T ) by basic
open sets. Thus, if B is a model of T and a ∈ B n we have that B |= φi (a) for some
i ∈ I. This shows that T ∪ {¬φi | i ∈ I} is unsatisfiable. So by compactness of first-
order logic (Theorem 2.0.1) I must have a finite subset F such that T ∪ {¬φi | i ∈ F }
is unsatisfiable. In other words, for every p ∈ Sn (T ) there is an i ∈ F such that
φi ∈ p, which implies that {[φi ] | i ∈ F } is a finite subcover of C.
To show total disconnectivity, let p, q ∈ Sn (T ) be distinct. Then there is a formula
φ ∈ p such that ¬φ ∈ q. Thus, [φ] is a basic clopen set separating p and q.
Exercises.
(58) Let A be a structure. Let T be the topology of A where the basic open sets
are the subsets of A that can be defined in A by a first-order formulas φ(x)
with parameters from A. Show that T is homeomorphic to S1 (Th(A)).
5.2. Saturated Structures

A structure A is saturated if, informally, ‘as many types as possible’ are realized
in A. Recall that AB , for B ⊆ A, denotes an expansion of A by constants for the
elements of B. We refer to the n-types of AB as the n-types of A over B. The set of
A
all maximal n-types of A over B is denoted by Sn (B).
Definition 5.2.1 (Saturation). For an infinite cardinal κ, a structure A is κ-
saturated if A realises all 1-types over B for all B ⊆ A with |B| < κ. We say that an
infinite structure A is saturated if it is |A|-saturated.
Example 32. The structure (Q; <) is saturated: if q1 , . . . , qn ∈ Q, then the
homogeneity of (Q; <) implies that (Q; <, q1 , . . . , qn ) has precisely 2n + 1 complete
1-types (see Example 15), and each of them is realised: we have n types realised by
the elements q1 , . . . , qn , one type realised by an element smaller than all the qi , one
type realised by an element larger than all the qi , and n − 1 types realised by elements
that lie between some of the elements qi . 4
The proof of the next result uses again a back and forth argument that we have
seen already in Example 15.
Theorem 5.2.2. Let A and B be two saturated structures with the same first-order
theory and the same cardinality. Then A and B are isomorphic.
38 5. TYPES
Proof. Let (aα )α<κ be an enumeration of A and (bα )α<κ an enumeration of

B. We inductively construct a sequence (cα )α<κ of elements of B and (dα )α<κ of
elements of A such that for all β < κ
(2) Th(A; (aα )α<β , (dα )α<β ) = Th(B; (cα )α<β , (bα )α<β ).
The base case β = 0 holds by the assumptions of the theorem. Suppose that
(cα )α<β and (dα )α<β have already been constructed. Let p be the 1-type of aβ
in (A; (aα )α<β , (dα )α<β ). By (2) and since B is saturated there exists cβ ∈ B that
realises p. We will prove that
Th(A; (aα )α≤β , (dα )α<β ) = Th(B; (cα )α≤β , (bα )α<β ).
Let φ be a first-order sentence such that (A; (aα )α≤β , (dα )α<β ) |= φ. Let ψ(x1 ) be the
first-order formula obtained from φ by replacing cβ by the new variable x1 . Note that
ψ lies in p, and since cβ realises p we have (B; (cα )α<β , (bα )α<β ) |= ψ(cβ ). Hence,
(B; (cα )α≤β , (bα )α<β ) |= φ. Similarly, we use saturation of A to find dβ such that
Th(A; (aα )α≤β , (dα )α≤β ) = Th(B; (cα )α≤β , (bα )α≤β ).
At the end of the day, the map f : A → B defined by f (aα ) := cα for all α < κ is a
homomorphism from A to B, and the map bα 7→ dα is a homomorphism from B to A
which is the inverse of f .
We now prove general results about the existence of κ-saturated structures. We
start with a lemma about realisation of types.
Lemma 5.2.3. Every structure A has an elementary extension B that realises all
1-types over A.
A
Proof. First proof: Let (pα )α<λ be an enumeration of S1 (A) where λ is an
ordinal1. We construct an elementary chain
AA =: A0 ≺ · · · ≺ Aβ ≺ · · · (β ≤ λ)
such that pα is realised in Aα+1 . Suppose that (Aα0 )α0 <β is already constructed.
• β is limit ordinal. Define Aβ := limα<β Aα . Then (Aα0 )α0 ≤β is an elementary
chain, using Tarski’s chain lemma (Lemma 3.1.3).
• β = α + 1. Every finite subset Ψ of pα is realised in AA , and therefore also
in Aα . By Lemma 5.0.2, Aα has an elementary extension Aα+1 that realises
pα .
Second proof: For each 1-type p of A over A, we introduce a new constant
symbol cp . Let T be the set of all atomic sentences of AA together with all the
formulas φ(cp ) where p is a 1-type of A over A and φ ∈ p. We will show that finite
subsets F of T are satisfiable. Let p be a 1-type of A over A. By definition, Th(AA )∪p
has a model B, which satisfies
^
ψ := ∃x φ(x).
φ(cp )∈F
V
Therefore, AA |= ψ, i.e., AA |= φ(cp )∈F φ(a) for some a ∈ A. Expanding AA by
cp := a for all 1-types p of A over A, we obtain a model of F .
So by compactness, T has a model B. Since T contains the atomic sentences that
hold in AA , the structure B is an elementary extension of A. Also, for each type p of
B over A, the structure B contains an element cp satisfying p, by choice of T .
1By Theorem A.3.4 there is a well-ordering of S A (A), and by Proposition A.3.2 there is an
1
A
ordinal λ and a sequence (pα )α<λ that enumerates the elements of S1 (A).
5.2. SATURATED STRUCTURES 39
Theorem 5.2.4. Let A be a structure and κ an infinite cardinal. Then A has a

κ-saturated elementary extension.
Proof. Build a chain (Aα )α<κ+ of structures inductively as follows:
• A0 := A.
• Aα+1 is an elementary extension of Aα realising all 1-types over Aα . Such
a structure exists by Lemma 5.2.3.
• If β is a limit ordinal then Aβ := limα<β Aα .
It follows by induction on α that Aα is an elementary extension of Aβ for all β < α,
using Tarski’s elementary chain lemma at the limit ordinals. So (Aα )α<κ+ is an
elementary chain of models.
By Tarski’s elementary chain lemma again, B := limα<κ+ Aα is an elementary
extension of A. We show that B is even κ+ -saturated (which implies κ-saturation).
Let S be a subset of B of size less than κ+ . Then S ⊆ Aα for some α ≤ κ: otherwise,
S contains for every γ < κ+ an element from Aγ+1 \ Aγ , so cf(κ+ ) < κ+ , in con-
tradiction to Proposition A.4.7. By construction, Aα+1 realises all 1-types over Aα .
Consequently, B realises all 1-types over S.
By paying attention to the sizes of the structures that we build, one can show the
following.
Theorem 5.2.5. Let τ be a signature and let κ ≥ |τ | an infinite cardinal. Then
every τ -structure of size at most 2κ has a κ+ -saturated elementary extension of size
2κ .
Proof. The statement follows from the following observations:
• In the proof of Lemma 5.0.2, the elementary extension of A in item (3) can
be chosen to have the same cardinality as A, by applying the Löwenheim-
Skolem theorem (Theorem 3.2.1).
• In the first proof of Lemma 5.2.3, the structure B can be build to have
cardinality 2|A| (there are at most 2|A| many subsets of A).
• The union of a chain of length at most κ+ of models of cardinality 2κ has
cardinality 2κ (see Theorem A.4.2).
So if we assume the Generalised Continuum Hypothesis, then there are saturated
models of size κ+ for all cardinals κ. We particularly point out the following special
case.
Corollary 5.2.6. Let τ be an at most countable signature, and T be a satisfiable
τ -theory. Then T has a ℵ1 -saturated model of cardinality 2ℵ0 .
Hence, assuming the continuum hypothesis, every satisfiable theory with a count-
able signature has a saturated model! Note that for general T , some set-theoretic
assumption is necessary for the existence of saturated models: if T has 2ℵ0 many
1-types (take for instance (Q; <) expanded by constants for all elements) then any
ℵ0 -saturated model has size 2ℵ0 . Hence, if ℵ1 < 2ℵ0 , then there is no saturated model
of size ℵ1 . Saturated structures can also be constructed using ultraproducts.
Theorem 5.2.7. Let (Ai )i<ω be a sequence of structures withQa countable sig-
nature τ and let U a non-principal ultrafilter on ω. Then B := i<ω Ai /U is ℵ1 -
saturated.
Proof. Let S be a subset of B of cardinality strictly smaller than ℵ1 , i.e., of
cardinality ℵ0 . Let p be a 1-type over S. Note that p = {φ1 , φ2 , . . . } is countable by
our assumption that |τ | ≤ ℵ0 . For each n ≥ 1, define

Xn := i ∈ {n, n + 1, . . . } | Ai |= ∃x(φ1 (x) ∧ · · · ∧ φn (x)) .
40 5. TYPES
Note that {n, n + 1, . . . } ∈ U since U is non-principal. Also note that B realises

φ1 (x)∧· · ·∧φn (x) by Lemma 5.0.2, and hence {i ∈ N | Ai |= ∃x(φ1 (x)∧· · ·∧φn (x))} ∈
U
T by Theorem 2.4.1. Thus, Xn ∈ U for every n ≥ 1. We have Xn+1 ⊆ Xn and
n≥1 Xn = ∅. So for every Q
i ∈ X1 there exists a maximal n(i) such that i ∈ Xn(i) .
We define an element f ∈ i∈ω Ai as follows. For i ∈ X1 , let f (i) ∈ Ai be such that
Ai |= φ1 (f (i)) ∧ · · · ∧ φn(i) (f (i)); if i ∈ N \ X1 , let f (i)
Qbe any element of Ai (using
the axiom of choice). We claim that [f ]U realises p in i∈ω Ai /U. By Łoś’s theorem
(Theorem 2.4.1) we have to show that for every n
{i ∈ N | Ai |= φn (f (i))} ∈ U.
It suffices to show that Xn ⊆ {i ∈ N | Ai |= φn (f (i))} since Xn ∈ U. Let i ∈ Xn .
Then i ∈ X1 and i ∈ Xn(i) . Thus, Ai |= φn(i) (fi ). This shows the claim and concludes
the proof.
5.3. Omitting Types

Let p be an n-type over T . A formula φ ∈ p isolates p if for every formula ψ ∈ p
we have that T |= ∀x̄(φ(x̄) ⇒ ψ(x̄)). A type p of a theory T is called principal if it is
isolated by some formula. In the terminology of Section 5.1, the principal types are
precisely the isolated points in Sn (T ): if φ isolates p, then [φ] = {p}.
If T is complete, then every principal type p is realised in every model of T . To
see this, let B be a model of T and suppose that the formula φ isolates the type p.
Since p is a type, T ∪ ∃x̄.φ(x̄) is satisfiable, and since T is complete we we have that
∃x̄.φ(x̄) is contained in T . Therefore, B |= ∃x̄.φ(x̄). If a ∈ B is such that B |= φ(ā),
then ā realises p. The omitting types theorem can be viewed as a converse of this
observation.
Theorem 5.3.1 (Omitting types theorem). Let τ be a countable signature, let
T be a satisfiable τ -theory, and let p be a non-principal n-type of T . Then T has a
countable model that omits p.
Proof. Let ρ := {c0 , c1 , . . . } be countably many constant symbols that are not
contained in τ . We will inductively construct a sequence θ0 , θ1 , . . . of (τ ∪ρ)-sentences
such that T ∗ := T ∪ {θ0 , θ1 , . . . } has model B with an elementary substructure A
that omits p. To do this, we alternate between steps that make sure that A omits p,
and steps that make sure that we can apply Tarski’s test (Lemma 3.1.4) to find the
elementary substructure A of B.
Let φ0 , φ1 , . . . be an enumeration of all (τ ∪ ρ)-formulas φ(x). Let d0 , d1 , . . . be
an enumeration of ρn .
• Stage 0. Let θ0 be >.
• Stage s = 2i + 1 (for Tarki’s test). Choose c ∈ ρ which does not occur
in T ∪ {θ0 , . . . , θ2i } and define θ2i+1 := (∃x.φi (x) ⇒ φi (c)). Clearly, T ∪
{θ0 , . . . , θ2i+1 } is satisfiable.
• Stage s = 2i+2 (for omitting p). Let di = (e1 , . . . , en ). Let ψ(x1 , . . . , xn ) be
the τ -formula obtained from θ1 ∧ · · · ∧ θ2i+1 by replacing each occurrence of
ei by xi and then replacing every other symbol c ∈ ρ \ {e1 , . . . , en } occurring
in θ by the variable xc and existentially quantifying over xc . Because p is
non-principal, there is a formula φ(x1 , . . . , xn ) ∈ p such that T ∪ {ψ, ¬φ}
is satisfiable, i.e., there is a model B of T and b ∈ B n such that B |=
ψ(b) ∧ ¬φ(b). Define θ2i+2 := ¬φ(di ). Note that T ∪ {θ0 , . . . , θ2i+2 } is
A
satisfiable: a model A can be obtained as the expansion of B where di = b.
The thus constructed theory T ∗ has a model B by compactness because T ∪
{θ1 , . . . , θs } is satisfiable for each s ∈ N. Let ψ(v) be a (τ ∪ ρ)-formula such that
5.3. OMITTING TYPES 41
T ∗ |= ∃v.ψ(v). Then there is an i ∈ N such that ψ = φi and at stage 2i + 1 we

have added the axiom ∃v.ψ(v) ⇒ ψ(c) for some c ∈ ρ. Hence, by Tarski’s test
(Lemma 3.1.4), {cB | c ∈ ρ} is the domain of an elementary substructure A of B.
Note that A is countable.
B
We claim that A omits p. Let a ∈ An . Then there is an i ∈ N such that a = di .
At stage 2i + 2 we ensure that B |= ¬φi (di ) for some φi ∈ p. Thus, ā does not realise
p. This concludes the proof.
We mention that with the same proof idea one can also construct models of T
that omit finitely many types at once. By cleverly designing countably many stages
in the construction we may even omit countably may types at once.
Theorem 5.3.2 (Countable omitting types theorem). Let τ be a countable sig-
nature, let T be a satisfiable τ -theory, and let p1 , p2 , . . . be non-principal types of T .
Then T has a countable model that omits all of the pi ’s.
We say that a structure B is atomic if for every a ∈ B n , the complete type of a
in B is principal.
Theorem 5.3.3 (Theorem 6.2.2 in [19]). Let T be a complete satisfiable theory
with countably many n-types for every n ∈ N. Then T has a countable atomic model.
Proof. There are only countably many non-principal complete types in T , so
by the countable omitting types theorem (Theorem 5.3.1) there is a countable model
B of T that omits all of them.
Lemma 5.3.4. Let A and B be atomic countable structures with the same theory.
Then A and B are isomorphic.
Proof. Back and forth.
Exercises.
(59) Spell out the proof of Lemma 5.3.4.
(60) (Exercise 13.1.5 in [25]) Show that every countable model M of Peano Arith-
metic has an elementary extension N such that for all a ∈ M and b ∈ N \ M
we have a < b (such extensions are called end extensions).
CHAPTER 6
Countably Categorical Structures
A structure B is called ω-categorical if its first-order theory is ω-categorical. There

are many equivalent characterisations of ω-categoricity; the most important one is in
terms of the automorphism group of B. In the following, let G be a set of permutations
of a set X. We say that G is a permutation group if G contains the identity idX , and
for g, f ∈ G the functions x 7→ g(f (x)) and x 7→ g −1 (x) are also in G . For n ≥ 1 the
orbit of (t1 , . . . , tn ) ∈ X n under G is the set {(α(t1 ), . . . , α(tn )) | α ∈ G }.
Proposition 6.0.1. Let A be an atomic countable structure. If c, d ∈ An have
the same type, then c and d lie in the same orbit of Aut(A).
Proof. We prove that if c, d ∈ An have the same type, then for any a ∈ A we can
pick b ∈ A such that then n + 1-tuples (a, c) and (b, d) have the same type. Since A is
atomic, the n + 1-type p of (a, c) is principal; let ψ(x0 , x1 , . . . , xn ) be a formula that
isolates it. Then ∃x0 .ψ holds on c, and since types are preserved by automorphisms
(Exercise 55) we have that ∃x0 .ψ also holds on d. This gives us an element b ∈ A
such that (b, d) satisfies ψ; since ψ isolates p, it follows that (a, c) and (b, d) have the
same type. Since A is countable, the statement now follows from a back and forth
argument just as in the proof of Theorem 5.2.2.
Definition 6.0.2. A permutation group G over a countably infinite set X is

oligomorphic if G has only finitely many orbits of n-tuples for each n ≥ 1.
Theorem 6.0.3 (Engeler, Ryll-Nardzewski, Svenonius). For a countably infinite
structure B with countable signature, the following are equivalent:
(1) B is ω-categorical;
(2) every type of B is principal;
(3) every model of Th(B) is atomic;
(4) B has finitely many complete n-types, for all n ≥ 1;
(5) for each n ≥ 1, there are finitely many inequivalent formulas with free vari-
ables x1 , . . . , xn over B;
(6) every model of Th(B) is ω-saturated;
(7) every relation that is preserved by Aut(B) is first-order definable in B;
(8) the automorphism group Aut(B) of B is oligomorphic.
Proof. We show the following implications:
(1) ⇒ (2) ⇒ (3) ⇒ (1)
(2) ⇒ (4) ⇒ (5) ⇒ (6) ⇒ (1)

(2) ∧ (3) ∧ (5) ⇒ (7) ⇒ (8) ⇒ (5)
(1) ⇒ (2). Suppose B has a non-principal type p. By the omitting types theorem
(Theorem 5.3.1) a countable model where p is omitted. But Th(B) also a countable
model where p is realised by the theorem of Löwenheim-Skolem (Theorem 3.2.1), so
B is not ω-categorical.
43
44 6. COUNTABLY CATEGORICAL STRUCTURES
(2) ⇒ (3). If all types of B are principal then B is atomic; the same applies to
all models of Th(B).
(3) ⇒ (1). All countable atomic models with the same theory are isomorphic by
Lemma 6.0.1.
(2) ⇒ (4). Suppose that all types are principal and let n ≥ 1. Then there exists
a sequence of formulas (φi )i∈I such that every n-type is isolated by one of those
formulas. Then Th(B) ∪ {¬φi | i ∈ I} is unsatisfiable and hence by the compactness
theorem (Theorem 2.0.1) there exists a finite F ⊆ I such that Th(B) ∪ {¬φi | i ∈ F }
is unsatisfiable. That is, in every model of Th(B), every n-tuple satisfies φi for some
i ∈ F , which shows that there are finitely many complete n-types in B.
(4) ⇒ (5). Every n-type is described by the complete n-types that contain the
n-type, so if there are finitely many complete n-types, there are finitely many n-types
in B. And this provides a finite upper bound for the number of formulas with free
variables x1 , . . . , xn .
(5) ⇒ (6). Let A be a model of Th(B), let a ∈ An , and let p be a com-
plete 1-type of (A, a). If there is a finite number of inequivalent first-order formulas
φ(x1 , x2 , . . . , xn+1 ), the conjunction over all formulas such that φ(x1 , ā) ∈ p isolates
p. So p is realised in (A, a). This shows that A is ω-saturated.
(6) ⇒ (1) follows from the fact that all countable ω-saturated structures with the
same theory are isomorphic (Theorem 5.2.2).
(2)∧(3)∧(5) ⇒ (7). Let R be an n-ary relation that is preserved by Aut(B). The
relation R is a union of orbits of n-tuples of Aut(B). It suffices to show that orbits
are first-order definable: by assumption (5), there are only finitely many inequivalent
first-order formulas, we can then define R by forming a finite disjunction. Since B is
atomic (3), if two n-tuples have the same type, then there is an automorphism that
maps one to the other, by Lemma 6.0.1. So types define orbits of n-tuples. Since
n-types of B are principal (2), it follows that the orbits of n-tuples are first-order
definable in B.
(7) ⇒ (8). Suppose that Aut(B) are infinitely many orbits of n-tuples, for some
n. Then the union of any subset of the set of all orbits of n-tuples is preserved by all
automorphisms of B; but there are only countably many first-order formulas over a
countable language, so not all the invariant sets of n-tuples can be first-order definable
in B.
(8) ⇒ (5) is immediate since automorphisms preserve first-order formulas.
The second condition in Theorem 6.0.3 provides another possibility to verify that
a structure is ω-categorical. We again illustrate this with the structure (Q; <). The
orbit of an n-tuple (t1 , . . . , tn ) from Qn with respect to the automorphism group of
(Q; <) is determined by the weak linear order induced by (t1 , . . . , tn ) in (Q; <). We
write weak linear order, and not linear order, because some of the elements t1 , . . . , tn
might be equal (that is, a weak linear order is a total quasiorder). The number of
weak linear orders on n elements is bounded by nn , and hence the automorphism
group of (Q; <) has a finite number of orbits of n-tuples, for all n ≥ 1.
Two structures A and B with the same domain A = B are called interdefinable
if every relation and operation of A is first-order definable in B, and vice versa.
Corollary 6.0.4. Two ω-categorical structures A and B with the same domain
A = B have the same automorphism group if and only if they are interdefinable.
Exercises.
(61) Is the structure (Z; {(x, y) : |x − y| ≤ 2}) ω-categorical?
6.1. ALGEBRAICITY 45
(62) Show that if A is countable, B is ω-categorical, and Age(A) ⊆ Age(B), then

A embeds into B.
(63) Prove: there exists a countable bipartite graph that embeds all countable
bipartite graphs.
(64) Show that if B is ω-categorical, then the expansion (B, c) of B by the con-
stant c ∈ B is ω-categorical as well.
(65) Which of the implications in Theorem 6.0.3 also work without the assump-
tion that the signature is countable, and which implications fail?
(66) Which of the implications in Theorem 6.0.3 also work without the assump-
tion that the structure B is countable, and which implications fail?
6.1. Algebraicity
The model-theoretic notion of algebraic closure generalises the notion of algebraic
closure in fields (see Chapter B).
Definition 6.1.1. Let A be a structure and let B ⊆ A. We say that a ∈ A is
algebraic over B if there exists a (τ ∪{B})-formula φ(x) such that {b ∈ A | AB |= φ(b)}
is finite and contains a. The algebraic closure of B in A is the set of elements of A
that are algebraic over B, and is denoted by aclA (B). We say that A has algebraicity
if aclA (B) 6= B for some finite B ⊆ A.
Remark 6.1.2. Let A be a countable ω-categorical structure and let B ⊆ A be
finite. Then the algebraic closure of B in A is the set of elements of A that lie in
finite orbits in Aut(A)(B) ; this is a direct consequence of Theorem 6.0.3 (also see
Exercise 64). Also note that for all finite subsets B the set aclA (B) is finite (see
Exercise 69).
Theorem 6.1.3 (See (2.15) in [7]). Let C be a homogeneous structure. Then the
age of C has strong amalgamation if and only if C has no algebraicity.
Proof. We first show that strong amalgamation of Age(C) implies no algebraic-
ity of C. Let A ⊆ C be finite, and u ∈ C \ A. We want to show that the orbit of u
in Aut(C)(A) is infinite. Let n ∈ N, A := C[A], and B := C[A ∪ {u}]. Then there
exists a strong amalgam B 0 ∈ Age(C) of B with B over A. We iterate this, taking
a strong amalgam of B with B 0 over A, showing that, because of homogeneity, there
are n distinct elements in C \ A that lie in the same orbit as u in Aut(C)(A) . Since
n ∈ N and u ∈ C \ A were chosen arbitrarily, the group Aut(C)(A) has no finite orbits
outside A.
For the other direction, we rely on the following lemma of Peter Neumann.
Lemma 6.1.4. Let G be a permutation group on D without finite orbits, and let
A, B ⊂ D be finite. Then there exists a g ∈ G with g(A) ∩ B = ∅.
Proof. The proof here is from Cameron [7], and is a nested induction. The
outer induction is on |A|. We assume the result for any set A0 with |A0 | < |A|. The
induction base A = ∅ is trivial. Suppose for contradiction that no g ∈ G with the
required property exists.
Claim. For any C with |C| ≤ |A|, there are only finitely many translates g(A)
of A that contain C. The proof of the claim is by induction on |A| − |C|. When
|A| − |C| = 0 then the only translate of A that contains C is C, and the statement
holds. So suppose that |C| < |A| and that the claim holds for all C 0 with |C 0 | > |C|.
By the outer induction hypothesis, we may assume that C ∩ B = ∅. By the inner
induction hypothesis, for each of the finitely many points b ∈ B, only finitely many
translates of A contain C ∪ {b}. So only finitely many translates of A contain C and
have non-empty intersection with B. Since we assumed that every translate of A has
non-empty intersection with B, we have shown the claim.
For C = ∅, the claim implies that A has only finitely many translates, a contra-
diction to the assumption that G has no finite orbits.
We now continue with the reverse implication of Theorem 6.1.3. Let A be with-
out algebraicity, and let (A0 , B 1 , B 2 ) be an amalgamation diagram with A0 , B 1 , B 2 ∈
Age(A). By homogeneity of A we can furthermore assume that A0 , B 1 , B 2 are sub-
structures of A; that is, the structure induced by B1 ∪B2 in A is an amalgam, but pos-
sibly not a strong one. Since A has no algebraicity, Aut(A)(A0 ) has no finite orbits out-
side A0 . By the lemma, there exists g ∈ Aut(A)(A0 ) such that (B1 \A0 )∩(B2 \A0 ) = ∅.
Then the structure induced by g(B1 ∪ B2 ) is a strong amalgam of B 1 and B 2 over
A0 .
Exercises.
(67) Find an example of a countable ω-categorical structure with only one 1-type
and algebraicity.
(68) Show that if a A is structure and B ⊆ A is finite such that aclA (B) is infinite,
then A is not ω-categorical.
(69) Show that the converse of the implication in the previous exercise is false.
6.2. Quantifier Elimination

A τ -theory T has quantifier elimination if every τ -formula φ(x1 , . . . , xn ) is modulo
T equivalent to a quantifier-free τ -formula ψ(x1 , . . . , xn ), i.e.,
T |= ∀x1 , . . . , xn (φ ⇔ ψ).
If a theory T has quantifier elimination in a reasonable language then this can be very
useful when working with T .
In this context, our assumption that we allow ⊥ and > as a first-order formula (de-
noting the empty and full 0-ary relation, respectively) becomes relevant; Hodges [18]
does not make this assumption, and therefore has to distinguish between quantifier-
elimination and what he calls quantifier-elimination for non-sentences.
Remark 6.2.1. Note that one can extend any theory T to a theory T 0 with
quantifier elimination by replacing τ by a larger signature that additionally contains
for every first-order formula φ(x1 , . . . , xn ) a relation symbol Rφ , and by adding to T 0
all the axioms
∀x1 , . . . , xn Rφ (x1 , . . . , xn ) ⇔ φ(x1 , . . . , xn ) .
This transformation preserves many model-theoretic property: if T is complete, or
κ-categorical, then so is T 0 .
An interesting source of theories with quantifier-elimination is the following lemma.
Lemma 6.2.2. An ω-categorical structure B has quantifier elimination if and only
if it is homogeneous.
Proof. Suppose first that B has quantifier-elimination. Let ā = (a1 , . . . , ak ) and
b̄ = (b1 , . . . , bk ) be k-tuples of elements of B such that the mapping that sends ai to
bi , for 1 ≤ i ≤ k, is an isomorphism f between the structures induced by {a1 , . . . , ak }
and by {b1 , . . . , bk }. Since B has quantifier-elimination, these two tuples satisfy the
same first-order sentences. By Theorem 6.0.3, the orbit of (a1 , . . . , ak ) is first-order
definable, and hence (b1 , . . . , bk ) lies in the same orbit as (a1 , . . . , ak ). It follows that
f can be extended to an automorphism of B.
6.3. FIRST-ORDER INTERPRETATIONS 47
Now suppose that B is homogeneous, and let φ(x1 , . . . , xk ) be a first-order for-

mula. By the theorem of Ryll-Nardzewski (Theorem 6.0.3), there are finitely many
orbits O1 , . . . , Om of k-tuples that satisfy φ. Clearly, it suffices to show that each of
those orbits can be defined by a quantifier-free formula. Let a ∈ B k be such that
B |= φ(a). We claim that the set of quantifier-free formulas that hold on (a1 , . . . , ak )
defines the orbit of a over B. To see this, let (b1 , . . . , bk ) be another k-tuple that
satisfies the same quantifier-free formulas as (a1 , . . . , ak ). Then the mapping that
sends ai to bi is a partial isomorphism, and by homogeneity can be extended to an
automorphism of B. Since automorphisms preserve first-order formulas, (b1 , . . . , bk )
also satisfies φ, which proves the claim.
Exercises.
(70) Prove that the theory T 0 constructed in Remark 6.2.1 indeed has quantifier
elimination.
(71) Which of the following ω-categorical structures has quantifier elimination:
• the structure (N × {0, 1}; E) where E = ((a, b), (a, c)) | b 6= c .
• the structure (N × {0, 1, 2}; E) where

E = ((a, b), (a, c)) | (b, c) ∈ {(0, 1), (1, 0), (1, 2), (2, 1)} .
• the structure (Q+ +
0 ; <) where Q0 := {x ∈ Q | x ≥ 0}.
• (Q; O) where O := {(x, y, z) ∈ Q3 | x < y ∨ x < z}.
• (Q; B) where B := {(x, y, z) ∈ Q3 | x < y < z ∨ z < y < x}.
6.3. First-Order Interpretations

First-order interpretations are a powerful tool to derive new structures from
known structures.
Definition 6.3.1. Let A and B be structures with the relational signatures τ
and σ and let d ∈ N. A (first-order) interpretation of dimension d of B in A is a
partial surjection I : Ad → B (also called the coordinate map) such that for every
relation R, say of arity k, defined by an atomic σ-formula φ in B, the dk-ary relation
I −1 (R) := (a11 , . . . , a1d , . . . , ak1 , . . . , akd ) | (I(a11 , . . . , a1d ), . . . , I(ak1 , . . . , akd )) ∈ R

has a first-order definition φI in A.

Since x = y is always allowed as an atomic formula, there must in particular
exist a τ -formula =I such that =I (x1,1 , . . . , x1,d , x2,1 , . . . , x2,d ) holds if and only if
(x1,1 , . . . , x1,d ), (x2,1 , . . . , x2,d ) lies in the kernel of I.
In order to specify a σ-structure B with a first-order interpretation in a given
τ -structure A up to isomorphism, it suffices to specify the interpreting formulas for
the atomic σ-formulas of A; in particular, if the signature of A is relational and finite,
then an interpretation has a finite presentation.
We say that B is interpretable in A with finitely many parameters if there are
c1 , . . . , cn ∈ A such that B is interpretable in the expansion of A by the constants ci
for all 1 ≤ i ≤ n.
Example 33. The field of rational numbers (Q; 0, 1, +, ∗) has a 2-dimensional
interpretation I in (Z; 0, 1, +, ∗). The interpretation is now given as follows.
• The formula =I (x1 , x2 , y1 , y2 ) is x2 6= 0 ∧ y2 6= 0 ∧ x1 y2 = y1 x2 ;
• The formula 0I (x1 , x2 ) is x2 6= 0 ∧ x1 = 0, the formula 1I (x1 , x2 ) is x2 6=
0 ∧ x1 = x2 ;
• The formula +I (x1 , x2 , y1 , y2 , z1 , z2 ) is
x2 , y2 , z2 6= 0 ∧ z2 ∗ (x1 ∗ y2 + y1 ∗ x2 ) = z1 ∗ x2 ∗ y2 ;
• The formula ∗I (x1 , x2 , y1 , y2 , z1 , z2 ) is

x2 , y2 , z2 6= 0 ∧ x1 ∗ y1 ∗ z2 = z1 ∗ x2 ∗ y2 . 4
Example 34. Allen’s Interval Algebra is a structure studied in artificial intelli-
gence for temporal reasoning; it is defined via a 2-dimensional first-order interpre-
tation I in (Q; <). The domain formula >I (x, y) is x < y. Hence, the elements of
A can indeed be viewed as non-empty closed bounded intervals [x, y] over Q. The
template A contains for each inequivalent {<}-formula φ with four variablesa binary
relation R such that (a1 , a2 , a3 , a4 ) satisfies φ if and only if (a1 , a2 ), (a3 , a4 ) ∈ R. In
particular, A has relations for equality of intervals, containment of intervals, and so
forth. 4
Example 35. Let G = (V ; E) be an undirected graph (viewed as a symmetric
digraph). Then the line graph LG of G is the (undirected) graph with vertex set

V (LG ) := {u, v} | (u, v) ∈ E
and the edge set

E(LG ) := ({u, v}, {v, w}) | {u, v}, {v, w} ∈ E .
The line graph has the 2-dimensional first-order interpretation I : E → V (LG ) in G
given by I(x, y) := {x, y}:
• I −1 {(u, u) | u ∈ V (LG )} has the first-order definition

E(x1 , x2 ) ∧ E(y1 , y2 ) ∧ (x1 = y1 ∧ x2 = y2 )

∨ (x1 = y2 ∧ x2 = y1 ) .
• I −1 E(LG ) has the first-order definition

E(x1 , x2 ) ∧ E(y1 , y2 ) ∧ (x1 = y1 ∧ x2 6= y2 ) ∨ (x1 = y2 ∧ x2 6= y1 )

∨ (x2 = y1 ∧ x1 6= y2 ) ∨ (x2 = y2 ∧ x1 6= y1 ) . 4
Many ω-categorical structures can be derived from other ω-categorical structures
via first-order interpretations.
Lemma 6.3.2. Let A be an ω-categorical structure. Then every structure B that
is first-order interpretable in A with finitely many parameters is ω-categorical.
Proof. By the theorem of Ryll-Nardzewski (Theorem 6.0.3) it suffices to show
that the number o(n) of orbits of n-tuples under Aut(B) is finite, for every n. If B
is the expansion of A by a constant c, then oB (n) ≤ oA (n + 1) since the map that
sends the orbit of t to the orbit of (c, t) is an injection (see Exercise 64). If B has a
d-dimensional interpretation in A then oB (n) ≤ oA (dn) and hence is finite, too.
Note that in particular all first-order reducts of an ω-categorical structure and
all expansions of an ω-categorical structure by finitely many constants are again ω-
categorical.
Lemma 6.3.3. Let B be a structure with at least two elements. Then every finite
structure has a first-order interpretation in B.
Proof. Let A be a τ -structure with domain {1, . . . , n}. The statement is trivial
if n = 1; so let us assume that n > 1 in the following. Our first-order interpretation
I of A in B is n-dimensional. For k ∈ {1, . . . , n − 1}, define
k
!
^
ρk (x1 , . . . , xn ) := xk 6= xk+1 ∧ x1 = xi
i=1
ρn := (x1 = · · · = xn ) .
The domain formula of our interpretation is true. Equality is interpreted by the

formula
_
=I (x1 , . . . , xn , y1 , . . . , yn ) := ρk (x1 , . . . , xn ) ∧ ρk (y1 , . . . , yn ) .
k<n
Note that the equivalence relation defined by =I on An has exactly n equivalence
classes. If R ∈ τ is m-ary, then the formula R(x1 , . . . , xm )I is a disjunction of con-
junctions with the nm variables x1,1 , . . . , xm,n . For each tuple (t1 , . . . , tm ) from RA
the disjunction contains the conjunct
^
ρti (xi,1 , . . . , xi,n ).
i≤m
Composing interpretations. First-order interpretations can be composed. In

order to conveniently treat these compositions, we first describe how an interpreta-
tion of a σ-structure B gives rise to interpreting formulas ψI for arbitrary σ-formulas
ψ(x1 , . . . , xn ). If ψ(y1 , . . . , yn ) is atomic, then ψI (y1,1 , . . . , y1,d , . . . , yn,1 , . . . , yn,d ) has
already been defined. If ψ is of the form ∃y.ψ 0 (y) by, then ψI is ∃y1 , . . . , yd . ψI0 (y1 , . . . , yd ).
Note that if ψ defines the relation R in B, then ψI defines I −1 (R) in A. For all d-tuples
a1 , . . . , an ∈ I −1 (B)

B |= ψ I(a1 ), . . . , I(an ) ⇔ A |= ψI (a1 , . . . , an ) .
Definition 6.3.4. Let C, B, A be structures with the relational signatures ρ, σ,
and τ . Suppose that
• C has a d-dimensional interpretation I in B, and
• B has an e-dimensional interpretation J in A.
Then C has a natural de-dimensional first-order interpretation I ◦ J in A: the domain
of I ◦ J is the set of all de-tuples t in A such that (t, t) satisfies the τ -formula (=I )J ,
and we define

I ◦ J a1,1 , . . . , a1,e , . . . , ad,1 , . . . , ed,e ) := I(J(a1,1 , . . . , a1,e ), . . . , J(ad,1 , . . . , ed,e ) .
Let φ be a τ -formula which defines a relation R over A. Then the formula (φI )J
defines in A the preimage of R under I ◦ J.
Let I1 and I2 be two interpretations of B in A of dimension d1 and d2 , respectively.
Then I1 and I2 are called homotopic 1 if the relation {(x̄, ȳ) | I1 (x̄) = I2 (ȳ)} of arity
d1 + d2 is first-order definable in A. Note that idC is an interpretation of C in C,
called the identity interpretation of C (in C).
Definition 6.3.5. Two structures A and B with an interpretation I of B in A
and an interpretation J of A in B are called mutually interpretable. If both I ◦ J and
J ◦ I are homotopic to the identity interpretation (of A and of B, respectively), then
we say that A and B are bi-interpretable (via I and J).
Example 36. The directed graph C := (N2 ; M ) where

M := ((u1 , u2 ), (v1 , v2 )) | u2 = v1
and the structure D := (N; =) are bi-interpretable. The interpretation I of C in D is 2-
dimensional, the domain formula is true, and I(u1 , u2 ) = (u1 , u2 ). The interpretation
J of D in C is 1-dimensional, the domain formula is true, and J(x, y) = x. Both
interpretations are clearly .
Then J(I(x, y)) = z is definable by the formula x = z, and hence I◦J is homotopic
to the identity interpretation of D. Moreover, I(J(u), J(v)) = w is definable by

M (w, v) ∧ ∃p M (p, u) ∧ M (p, w) ,
1We follow the terminology from [2].
so J ◦ I is also homotopic to the identity interpretation of C. 4

Example 37. Let I be the set of all non-empty closed bounded intervals over Q
(also see Example 34). Let I be the 2-dimensional interpretation of (I; m) in (Q; <)
with domain formula x < y, mapping (x, y) ∈ Q2 with x < y to the interval [x, y] ∈ I.
The formula (y1 = y2 )I is true, and the formula (m(y1 , y2 ))I has variables x11 , x12 , x21 , x22
and is given by x12 = x21 .
Let J be the 1-dimensional interpretation with the domain formula true and
J([x, y]) := x. The formula (x < y)I is the formula

∃u, v m(u, x) ∧ m(u, v) ∧ m(v, y) .
We show that J ◦ I and J ◦ I are homotopic to the identity interpretation. The
relation (x1 ,x2 , y) | J(I(x1 , x2 )) = y has the definition x1 = y. To see that the
relation R := (u, v, w) | I(J(u), J(v)) = w has a definition in (I; m), first note that
the relation

(u, v) | u = [u1 , u2 ], v = [v1 , v2 ], u1 = v1

has the definition φ1 (u, v) = ∃w m(w, u) ∧ m(w, v) in (I; m). Similarly, {(u, v) | u =
[u1 , u2 ], v = [v1 , v2 ], u2 = v2 } has a definition φ2 (u, v). Then the formula φ1 (u, w) ∧
φ2 (v, w) is equivalent to a primitive positive formula over (I; m), and defines R. 4
Proposition 6.3.6. If A and B are bi-interpretable, then Aut(A) and Aut(B)
are isomorphic as abstract groups.
Proof. Let I be a d-dimensional interpretation of B in A, and let J be an e-
dimensional interpretation of A in B that witness that A and B are bi-interpretable.
Let α ∈ Aut(A); define µ(β) : B → B as follows. For b ∈ B, pick (a1 , . . . , ad ) ∈ Ad
such that I(a1 , . . . , ad ) = b. Define µ(β)(b) := I(α(a1 ), . . . , α(ad )); note that this
value is independent from the choice of (a1 , . . . , ad ), and that µ(β) defines an auto-
morphism of B. Conversely, every automorphism β of B induces an automorphism
ν(β) of A. The assumption that I ◦ J is homotopic to the identity interpretation of A
implies that ν(µ(α)) = α, and the assumption that J ◦ I is homotopic to the identity
interpretation of B implies that µ(ν(β)) = β, which shows that µ is a bijection with
inverse ν. It is straighforward to verify that µ maps idA to idB and that it preserves
composition.

Example 38. The structures C := N2 ; {(x, y), (u, v) | x = u} and D := (N; =)
are mutually interpretable, but not bi-interpretable. There is a interpretation I1
of D in C, and a interpretation of C in D such that I2 ◦ I1 is homotopic to the
identity interpretation. However, the two structures are not bi-interpretable. To see
this, observe that Aut(C) has the non-trivial normal subgroup N := Sym(N)N and
that Aut(C)/N is isomorphic to Aut(D). However, Aut(D) = Sym(N) has exactly
three proper non-trivial normal subgroups [27], none of which is the automorphism
group of a structure. Therefore, Proposition 6.3.6 implies that C and D are not
bi-interpretable. 4
Definition 6.3.7. A structure B has essentially infinite signature if every rela-
tional structure C that is interdefinable with B has an infinite signature.
We show that the property of having essentially infinite signature is preserved by
bi-interpretability.
Proposition 6.3.8. Let B and C be structures that are bi-interpretable. Then B
has essentially infinite signature if and only if C has essentially infinite signature.
Proof. Let τ be the signature of B. Suppose that the interpretation I1 of C

in B is d1 -dimensional, and that the interpretation I2 of B in C is d2 -dimensional.
Let θ(x, y1,1 , . . . , yd1 ,d2 ) be the τ -formula that shows that I2 ◦ I1 is homotopic to the
identity interpretation of B. That is, θ defines in B the (d1 d2 + 1)-ary relation that
contains a tuple (a, b1,1 , . . . , bd1 ,d2 ) iff

a = h2 h1 (b1,1 , . . . , b1,d2 ), . . . , h1 (bd1 ,1 , . . . , bd1 ,d2 ) .
We have to show that if C has a finite signature, then B is interdefinable with
a structure B 0 with a finite signature. Let σ ⊆ τ be the set of all relation symbols
that appear in θ and in all the formulas of the interpretation of C in B. Since the
signature of C is finite, the cardinality of σ is finite as well. We will show that there
is a definition of B in the σ-reduct B 0 of B.
Let φ be an atomic τ -formula with k free variables x1 , . . . , xk . Then the
σ-formula
^
1
∃y1,1 , . . . , ydk1 ,d2 i
θ(xi , y1,1 , . . . , ydi 1 ,d2 )
i≤k
1 k 1 k
, . . . , ydk1 ,d2 )

∧ φI1 I2 (y1,1 , . . . , y1,d 2
, y2,d2
, . . . , y2,d 2
is equivalent to φ(x1 , . . . , xk ) over B 0 . Indeed, by the surjectivity of h2 , for every

element ai of B there are elements ci1 , . . . , cid2 of C such that h2 (ci1 , . . . , cid2 ) = ai , and
by the surjectivity of h1 , for every element cij of C there are elements bi1,j , . . . , bid1 ,j
of B such that h1 (bi1,j , . . . , bid1 ,j ) = cij . Then
B |= R(a1 , . . . , ak ) ⇔ C |= φI2 (c11 , . . . , c1d2 , . . . , ck1 , . . . , ckd2 )
⇔ B 0 |= φI1 I2 (b11,1 , . . . , bk1,d2 , b12,d2 , . . . , bk2,d2 , . . . , bkd1 ,d2 ).
Exercises.
(72) Show the claim from Example 38 that (N2 ; {(x, y), (u, v) | x = u}) and
(N; =) are mutually interpretable.
(73) Show that the structure (N2 ; {((x, y), (y, z)) | x, y, z ∈ N}) is bi-interpretable
with (N; =).
(74) Show that (N2 ; {((x, y), (x, z)) | x, y, z ∈ N}, {((x, y), (z, y)) | x, y, z ∈ N})
and (N; =) are not bi-interpretable.
(75) (*) Show that two finite structures A and B have isomorphic automorphism
groups (as abstract groups) if and only if A and B are bi-interpretable.
CHAPTER 7
Model Completeness and Quantifier-Elimination
Quantifier elimination in ω-categorical structures is particularly easy to check,

as we have seen in Section 6.2. Proving quantifier elimination in more complicated
mathematical structures such as (C; +, ∗, 1) or (R; +, 1, ∗, <) is more difficult. In
this chapter we first introduce a concept which can be seen as half-way to quanti-
fier elimination, namely model completeness. We will also learn some basic lemmas
about quantifier elimination and then apply them to prove that the two interesting
mathematical structures above have quantifier elimination.
7.1. Preservation Theorems

Preservation theorems in model theory establish links between definability in
(a syntactically restricted fragment of) a given logic with certain ‘semantic’ closure
properties.
Definition 7.1.1. When T is a first-order theory and φ(x̄) and ψ(x̄) are formulas,
we say that φ and ψ are equivalent modulo T if T |= ∀x̄(φ(x̄) ⇔ ψ(x̄)).
7.1.1. The Homomorphism Preservation Theorem. A formula is called

existential positive if it is built from conjunction, disjunction, and existential quan-
tification, only (negation and universal quantification are forbidden). Note that every
existential positive τ -formula can be written as a disjunction of primitive positive
formulas, i.e., formulas of the form
∃x1 , . . . , xn (ψ1 ∧ · · · ∧ ψm )
where ψ1 , . . . , ψm are atomic τ -formulas. Such formulas are also called unions of
disjunctive queries in database theory and the most frequently encountered database
queries in practice.
Such formulas also play an important role in geometry. For example, note that
if S1 , . . . , Sk are convex subsets of Rn , and S0 has a primitive positive definition in
(R; S1 , . . . , Sk ), then S0 is convex as well. The homomorphism preservation theo-
rem links existential positive definability modulo T with the fundamental concept of
preservation under homomorphisms between models of T .
Theorem 7.1.2 (Homomorphism Preservation Theorem). Let T be a first-order

theory. A first-order formula φ is equivalent to an existential positive formula modulo
T if and only if φ is preserved by all homomorphisms between models of T .
Note that here the assumption that ⊥ is always part of first-order logic is impor-
tant: the first-order formula ∃x. x 6= x is preserved by all homomorphisms between
models of T , but without ⊥ it might not be equivalent to an existential positive
formula modulo T (for instance when T is the empty theory).
In the proof of Theorem 7.1.2 we need the following property of saturated struc-
tures.
53
54 7. MODEL COMPLETENESS AND QUANTIFIER-ELIMINATION
Lemma 7.1.3. Let A and B be τ -structures, where B is |A|-saturated. Suppose

that every primitive positive sentence that is true in A is also true in B. Then there
is a homomorphism from A to B.
Proof. Let (ai )0≤i<|A| well-order A. We claim that for every µ ≤ |A| there is a
sequence (bi )i<µ of elements of B such that every primitive positive (τ ∪ {ci | i < µ})-
sentence true on (A, (ai )i<µ ) is true on (B, (f (ai ))i<µ ). The proof is by transfinite
induction on µ.
• The base case, µ = 0, follows from the hypothesis of the lemma.
• The inductive step for limit ordinals µ follows from the observation that
a sentence can only mention a finite collection of constants, whose indices
must all be less than some ν < µ.
• For the inductive step for successor ordinals µ = ν + < |A|, set

Σ := φ(x) | φ is a ep-(τ ∪ {ci | i < ν})-formula such that
(A, (ai )i<ν ) |= φ(aµ ) .
By inductive assumption (B, (bi )i<ν ) |= ∃x.φ(x) for every φ ∈ Σ. By com-

pactness, since Σ is closed under conjunction, we have that Σ is a ep-1-type
of (B, (bi )i<ν ). Then Σ is realised by some element bγ ∈ B because B is
ep-|A|-saturated. By construction we maintain that all primitive positive
(τ ∪ {ci | i < ν + })-sentences true on (A, (ai )i<γ + ) are true on (B, (bi )i<γ + ).
The function that maps ai to bi for all i < |A| is a homomorphism from A to B.
Proof of Theorem 7.1.2. It is clear that homomorphisms preserve existential

positive formulas. For the converse, let φ be first-order, with free variables x1 , . . . , xn ,
and preserved by homomorphisms between models of T . Let τ be the signature of T
and φ, and let c̄ = (c1 , . . . , cn ) be a sequence of constant symbols that do not appear
in τ . If T ∪ {φ(c̄)} is unsatisfiable, the statement is clearly true, so assume otherwise.
Let Ψ be the set of all existential positive (τ ∪ {c1 , . . . , cn })-sentences ψ such that
T ∪ {φ(c̄)} |= ψ. Let A be a model of T ∪ Ψ. Let U be the set of all primitive positive
sentences θ such that A |= ¬θ.
We claim that T ∪{¬θ | θ ∈ U }∪φ(c̄) is satisfiable. For otherwise, by compactness,
there would be aWfinite subset U 0 of U such that T ∪{¬θ | θ ∈ U 0 }∪φ(c̄) is unsatisfiable.
But then ψ := θ∈U 0 θ is an existential positive sentence such that T ∪ {φ(c̄)} |= ψ,
and hence ψ ∈ Ψ. This is in contradiction to the assumption that A |= ¬θ for all
θ ∈ U . We conclude that there exists a model B of T ∪ {¬θ | θ ∈ U } ∪ φ(c̄).
By Theorem 5.2.5, A has an elementary extension A0 which is |B|-saturated.
Every primitive positive (τ ∪ {c1 , . . . , cn })-sentence θ that is true in B is also true in
A0 : for if otherwise θ were false in A0 , then it were also false in A, and hence θ ∈ U
in contradiction to the assumption that B |= {¬θ | θ ∈ U }. Hence, by Lemma 7.1.3,
there exists a homomorphism from B to A0 . Since B |= φ(c̄), and φ is preserved by
homomorphisms between models of T , we have A0 |= φ(ā).
We conclude that T ∪Ψ∪{¬φ(c̄)} is unsatisfiable, and again by compactness there
exists a finite subset Ψ0 of Ψ such that T ∪ Ψ0 ∪ {¬φ(c̄)} is unsatisfiable. Then Ψ0
V
is an existential positive sentence; let ψ be the formula obtained from this sentence
by replacing for all i ≤ n all occurrences of ci by xi . Then T |= ∀x̄(ψ(x̄) ⇔ φ(x̄)),
which is what we wanted to show.
The classical theorem of Łos-Tarski for preservation under embeddings of models

of a theory is a direct consequence of the homomorphism preservation theorem.
7.1. PRESERVATION THEOREMS 55
Corollary 7.1.4 (Łos-Tarski; see e.g. Corollary in 5.4.5 of [19]). Let T be a first-
order theory. A first-order formula φ is equivalent to an existential formula modulo
T if and only if φ is preserved by all embeddings between models of T .
Proof. Add for each atomic formula ψ a new relation symbol Nψ to the signature
of T , and add the sentence ∀x̄(Nψ (x̄) ⇔ ¬ψ(x̄)); let T 0 be the resulting theory. Then
every existential formula φ is equivalent to an existential positive formula in T 0 , which
can be obtained from φ by replacing negative literals ¬ψ(x̄) in φ by Nψ (x̄). Similarly,
homomorphisms between models of T 0 must be embeddings. Hence, the statement
follows from Theorem 7.1.2.
7.1.2. The Theorem of Chang-Łos-Suszko. Another important preserva-
tion theorem concerns definability by so-called ∀∃-formulas; theories consisting of
∀∃-sentences are also called inductive for reasons that will soon become clear. Such
theories will play an important role in Section 7.5.
Definition 7.1.5. A formula φ is called ∀∃ if it is of the form ∀y1 , . . . , ym . ψ
where ψ is existential.
Example 39. The property of the random graph which is described in Exam-
ple 26 can be formulated as a set of ∀∃-sentences. Note that this property described
the random graph up to isomorphism. It follows that the theory of the random graph
is equivalent to a ∀∃-theory. 4
Our next goal will be a preservation theorem that characterises whether a formula
is equivalent to a ∀∃ formula modulo a theory T (Theorem 7.1.8 below). We first need
the following simple but important and frequently used lemma.
Lemma 7.1.6. Let A and B be τ -structures and suppose that every existential
sentence true in A is also true in B. Then there exists an elementary extension C of
B and an embedding g : A → C.
Proof. Let T be the set of all quantifier-free sentences in Th(AA ). It suffices to
show that T ∪ Th(B B ) has a model C 0 , since then
0
• the map cAA 7→ cC embeds A into the τ -reduct C of C 0 , and
• C is an elementary extension of B.
If T ∪ Th(B B ) has no model, then by the compactness theorem there is a τ -formula
φ such that φ(c̄) ∈ T and {φ(c̄)} ∪ Th(B B ) has no model, and in particular B |=
¬∃ȳ.φ(ȳ). Since ∃ȳ.φ(ȳ) is existential, the assumptions imply that A |= ¬∃ȳ.φ(ȳ).
This contradicts that AA |= φ(c̄).
Similarly, one can show the following.
Lemma 7.1.7. Let T be a first-order theory, and let A be a model of the ∀∃-
consequences of T . Then A can be extended to a model B of T such that every
existential formula that holds on a tuple ā in B also holds on ā in A.
Proof. Let A0 be an expansion of A by constants such that each element of A0
is denoted by a constant symbol. It suffices to prove that T ∪ Th(A0 )qf ∪ Th(A0 )∀ has
a model B. Suppose for contradiction that it were inconsistent; then by compactness,
there exists a finite subset U of Th(A0 )∀ ∪ Th(A0 )qf such that T ∪ U is inconsistent.
Let φ be the conjunction over U where all new constant symbols are existentially
quantified. Then T ∪ {φ} is inconsistent as well. But ¬φ is equivalent to a ∀∃
formula, and a consequence of T . Hence, A |= ¬φ, a contradiction.
Theorem 7.1.8 (Chang-Łoś-Suszko Theorem). A theory T is equivalent to a ∀∃-
theory if and only if the class of models of T is closed under forming unions of chains.
Proof. Unions of chains clearly preserve ∀∃-sentences. Conversely, suppose that

the models of T are preserved by forming unions of chains. Let S be the set of all
∀∃+ -sentences that are consequences of T . We show that S implies T . It suffices to
show that every model of S is elementary equivalent to a chain (B i )i<ω of models of
T . To construct this chain, we define an elementary chain of models (Ai )i<ω of T
such that there are
• embeddings fi : Ai → B i , with B i |= T , such that for every tuple āi of
elements from Ai and every existential formula θ, if B i |= θ(fi (āi )), then
Ai |= θ(āi ), and
• embeddings gi : B i → Ai+1 , such that gi ◦ fi is the identity on Ai .
Let A0 be a countable model of T . To construct the rest of the sequence, suppose that
Ai has been chosen. Since A0 is an elementary substructure of Ai , in particular all the
∀∃-consequence of S hold in Ai . By Lemma 7.1.7, the structure Ai can be extended to
a model B i of T such that every existential sentence that holds in (B i , āi ) also holds
in (Ai , āi ). By Lemma 7.1.6 there are an elementary extension Ai+1 of AiSand an
embedding gi : B i → Ai+1 such that gi ◦ fi is the identity on Ai . Then C := i<ω Ai
equals limi<ω B i , and by the Tarski-Vaught elementary chain theorem (Lemma 3.1.3)
A0 is an elementary substructure of C. So C is a model of Φ.
7.2. Model Completeness

A theory T is model complete if every embedding between models of T is elemen-
tary, i.e., preserves all first-order formulas. An equivalent characterisation of model
completeness is as follows.
Theorem 7.2.1. Let T be a theory. Then the following are equivalent.
(1) T is model complete.
(2) Every first-order formula is modulo T equivalent to an existential formula.
(3) For every embedding e : A ,→ B between models A and B of T , every tuple
ā of elements of A, and every existential formula φ, if B |= φ(e(ā)) then
A |= φ(ā).
(4) Every existential formula is modulo T equivalent to a universal formula;
(5) Every first-order formula is modulo T equivalent to a universal formula.
Proof. (1) ⇒ (2). Suppose that T is model complete, and let φ be a first-order
formula. Since T is model complete, φ is preserved by all embeddings between models
of T . It follows from Corollary 7.1.4 that φ is equivalent to an existential formula.
(2) ⇒ (3). Let e be an embedding of a model of A of T into a model B of T . Let
ā be a tuple of elements of A, and φ an existential formula such that B |= φ(e(ā)).
By (2), the formula ¬φ is equivalent to an existential formula. Therefore, e preserves
¬φ. Since B |= φ(e(ā)) we therefore must have A |= φ(ā).
(3) ⇒ (4): Let φ be a first-order formula; we have to show that ¬φ is equivalent to
an existential formula. But (3) implies that ¬φ is preserved by embeddings between
models of T , so the statement follows from Corollary 7.1.4.
(4) ⇒ (5). Let φ be a first-order formula, written in prenex normal form
Q1 x1 · · · Qn xn .ψ for ψ quantifier-free. If n = 0 then there is nothing to be shown.
Otherwise, let i ≤ n be smallest so that Qi = · · · = Qn . If i = 1 then φ is already
universal, or equivalent to a universal formula by (4), and we are done. Otherwise,
we distinguish two cases:
• If Qi = · · · = Qn = ∃ then by (4) the formula ∃xi . . . ∃xn .ψ is equivalent
modulo T to a universal formula ψ 0 . Then φ0 := Q1 x1 · · · Qi−1 xi−1 .ψ 0 is
equivalent to φ.
7.3. MODEL COMPANIONS 57
• If Qi = · · · = Qn = ∀ then by (4) the formula ∃xi . . . ∃xn .¬ψ is equiva-

lent modulo T to a universal formula ψ 0 . In this case, the formula φ0 :=
Q1 x1 · · · Qi−1 xi−1 .¬ψ 0 is equivalent to φ.
In both cases, φ0 has fewer quantifier alternations. The claim therefore follows by
induction on the number of quantifier alternations of φ.
(5) ⇒ (1). Follows from the fact that existential formulas are preserved by
embeddings.
Theorem 7.2.1 shows that in particular theories with quantifier-elimination are
model complete. We say that a structure A is model complete if and only if the
first-order theory Th(A) of A is model complete.
Example 40. All finite structures A are model complete: self-embeddings of A
are automorphisms, and hence they are elementary. Every relation that is first-order
definable in a finite structure also has an existential definition. On the other hand,
finite structures might not have quantifier elimination. 4
Example 41. The structure (Q+ +
0 ; <), where Q0 denotes the non-negative rational
numbers, is not model complete, because the self-embedding x 7→ x + 1 of (Q+ 0 ; <)
does not preserve the formula φ(x) = ∀y. ¬(y < x) (which is satisfied only by 0). 4
Example 42. Let T be the first-order theory of (Z; succ) where succ is the binary
relation {(x, y) | y = x + 1}). Then T does not have quantifier elimination, but is
model complete. 4
Proposition 7.2.2. Every model complete theory T is equivalent to a ∀∃-theory.
Proof. Model completeness implies that any chain of models of T is an ele-
mentary chain. So unions of chains are again models by Tarski’s elementary chain
theorem, and the statement following from the preservation theorem of Chang-Łoś-
Suszko (Theorem 7.1.8).
Exercises.
(76) Show that Th(Q; 0, 1, +, ·) is not model complete. You may use the following
facts:
• There exists a first-order formula φ(x) that defines Z in (Q; 0, 1, +, ·)
(the formula can even be chosen to be a ∀∃-formula; Poonen [24]).
• Every recursively enumerable subset of Z has an existential definition
in (Z; 0, 1, +, ·) [23].
• There are recursively enumerable subsets of Z whose complement is not
recursively enumerable (see, e.g., [3]).
7.3. Model Companions

In this section we study conditions that imply that we can pass from a theory T
to a model complete theory T 0 satisfying the same universal sentences. We write T∀
(T∃ ) for the set of all universal (existential) sentences that are implied by T . We say
that two theories T and T 0 are companions if T∀ = T∀0 . First we observe the following
consequence of Lemma 7.1.6.
Corollary 7.3.1. If T∀ ⊆ T∀0 then every model of T can be embedded into a
model of T 0 .
Proof. Let A be a model of T 0 , and let S := Th(A)∃ . We claim that T ∪ S is
satisfiable. If not, then by compactness (Theorem 2.0.1) there is some finite subset
{φ1 , . . . , φk } of S such that T ∪{φ1 , . . . , φk } is unsatisfiable. Note that ¬φ1 ∨· · ·∨¬φk
is equivalent to a universal sentence ψ, and T implies ψ, so by assumption we have

that T 0 implies ψ, and in particular A |= ψ. We have reached a contradiction since
A |= φi for all i ∈ {1, . . . , k}. So indeed T ∪ S has a model B. Then B satisfies the
assumptions from Lemma 7.1.6, so there exists an elementary extension C of B that
embeds A.
Proposition 7.3.2. Two theories T and T 0 are companions if and only if every
model of T can be embedded into a model of T 0 , and vice versa.
Proof. If T and T 0 are companions, then the statement follows from two sym-
metric applications of Corollary 7.3.1.
For the converse direction of the statement, let φ be a universal sentence. If φ
is not implied by T 0 , then there exists a model A of T 0 ∪ ¬φ. By assumption, there
exists an embedding of A into a model B of T . Note that ¬φ is equivalent to an
existential sentence, and hence it is preserved by the embedding, and must hold in B.
Hence, B cannot satisfy φ. This shows that every universal sentence that is implied
by T is also implied by T 0 . The statement is symmetric in T and T 0 , so we obtain
that T∀ = T∀0 .
Definition 7.3.3. A theory T 0 is a model companion of a theory T if
• T and T 0 are companions, and
• T 0 is model complete.
Example 43. We have seen in Example 41 that (Q+ 0 ; <) is not model complete.
However, it has a model companion: the first-order theory of (Q; <). 4
It is known that every ω-categorical theory has a model companion [26]. An
example of a theory without model companion is the theory of groups, as we will see
later (Example 45). We will now prove that the model companion of a theory T , if it
exists, is unique up to logical equivalence. To prove this, we need the concepts from
the next section.
7.4. Existentially Closed Structures

Definition 7.4.1. Let T be a theory. A structure A is called existentially closed
(e.c.) for T (sometimes also called existentially complete) if there is an embedding
from A to a model of T , and if for any embedding h from A into a model B of T , any
tuple ā from A, and any existential formula φ with B |= φ(h(ā)) we have A |= φ(ā).
Example 44. Every field K that is existentially closed for the theory of fields
must be algebraically closed: suppose that f (x) is a polynomial of degree at least one
and coefficients from K. We have to show that f (x) has a root in K. We can suppose
that f is irreducible; then E := K[x]/(f ) is a field which extends K and contains
a root of f (see Proposition B.4.9). So E |= ∃y. f (y) = 0. Since K is existentially
closed, K |= ∃y. f (y) = 0, too, so that f has a root already in K. 4
We have the following lemma about existence of existentially closed models. For
a theory T , we write T∀∃ for the set of ∀∃ sentences that are implied by T .
Lemma 7.4.2. Let T be a theory with a countable signature. Then any model
A of T of infinite cardinality κ embeds into a structure B of cardinality κ which is
existentially complete for T and a model of T∀∃ .
Proof. For simplicity of presentation, we prove the statement only for the count-
able case κ = ω, but the proof can easily be modified for general cardinalities κ. Set
B0 := A. Having constructed a countable Bi let {(φj , āj ) | j ∈ N} be an enumeration
7.5. THE KAISER HULL 59
of all pairs (φ, ā) where φ(x1 , . . . , xn ) is existential and ā is an n-tuple from Bi . We
construct a chain (Bji )j∈N of countable models of T as follows.
Set B0i = Bi−1 , and let j ∈ N. If there is a model C of T and an embedding
e : Bji → C such that C |= φj (e(āj )), then by the theorem of Löwenheim-Skolem
there is also a countable model C 0 of T and an embedding e0 : Bji → C 0 such that
C 0 |= φj (e0 (āj )). Identify the elements of B ji with elements of C 0 along e0 , and set
Bj+1
i := C 0 . Otherwise, if there is no such model C, we set Bj+1 i := Bji .
j
Let B i be limj∈N B i . Clearly, B := limi<ω B i is a countable model of T∀∃ . To
verify that B is T -ec, first note that B embeds into a model of T by Lemma 7.1.6.
Now suppose that g is an embedding from B to a model C of T , and suppose that
there is a k-tuple b̄ over B and an existential formula φ such that C |= φ(g(b̄)). There
is an i ∈ N such that b̄ ∈ Bik . By construction we have that B i+1 |= φ(ā). Thus, we
also have that B |= φ(b̄), which is what we had to show.
Corollary 7.4.3. Let K be a field. Then K has an algebraically closed field
extension.
Proof. Let T be the theory of fields. By Theorem 7.4.2, K embeds into a T -ec
structure F which is a model of T . Note that the theory of fields is ∀∃, so F is a field,
too. The observations in Example 44 show that F is algebraically closed.
7.5. The Kaiser Hull

We will show that the model companion of T , if it exists, is precisely the theory
of all existentially closed structures for T . To prove this, we need the following.
Lemma 7.5.1. Every theory T has a unique largest ∀∃-theory as companion.
Proof. Suppose for contradiction that the set of ∀∃-theories that are companions
of T is not closed under unions. Then by compactness there are ∀∃-theories S and S 0
and an existential sentence φ such that
• S and S 0 are companions of T
• T ∪ {φ} has a model A, and
• S ∪ S 0 ∪ {φ} is unsatisfiable.
By Corollary 7.3.1 there exists an embedding from A to a model A0 of S, and an
embedding from A0 to a model A1 of S 0 . Repeating this step we construct a chain of
structures (Ai )i∈N such that Ai is a model of S for even i and a model of S 0 for odd
i. The union B := limi<ω Ai is a model of S ∪ S 0 . Since there is an embedding from
A to B and A |= φ we have that B |= φ, a contradiction.
The theory constructed in Lemma 7.5.1 will be called the Kaiser hull of T , de-
noted in the following by T KH .
Lemma 7.5.2. T KH equals the set of ∀∃-sentences that hold in all T -ec structures.
Proof. Let T ∗ be the set of all ∀∃-sentences satisfied by all T -ec structures. To
show that T ∗ ⊆ T KH it suffices to show that (T ∗ )∀ = T∀ . By Lemma 7.4.2, every
model of T∀ embeds into a T ∗ -ec structure. Therefore, every universal sentence that
holds in all T ∗ -ec structures must be in T∀ , i.e., (T ∗ )∀ ⊆ T∀ . Conversely, every T -ec
structure embeds into a model of T∀ , and therefore satisfies T∀ , so T∀ ⊆ (T ∗ )∀ .
We now show that T KH ⊆ T ∗ . For this, we have to show that every T -ec structure
A satisfies all φ ∈ T KH . Since T KH is a ∀∃-theory, φ is of the form ∀ȳ.ψ(ȳ) where ψ is
existential. Let ā be a tuple of elements of A. We have to show that A |= ψ(ā). Since
(T ∗ )∀ = T∀ = T∀KH , by Corollary 7.3.1 there is an embedding e from A to a model B
of T∀KH . Since B |= φ we have that B |= ψ(e(ā)). Since T∀KH = T∀ , by Corollary 7.3.1
there is an embedding g from B to a model C of T . Since g preserves ψ we have that

C |= ψ(g(e(ā))). Now A |= ψ(ā) since A is T -ec. We conclude that A |= T KH .
Theorem 7.5.3 (Theorem 3.2.14 in [29]). Let T be a theory. Then the following
are equivalent.
(1) T has a model companion;
(2) all models of the Kaiser hull of T are T -ec;
(3) the class of T -ec structures has a first-order axiomatization.
In particular, if T has a model companion T ∗ , then it is unique up to equivalence of
theories, and T ∗ = T KH is the theory of all T -ec structures.
Proof. (1) ⇒ (2). Let U be the model companion of T . By Proposition 7.2.2,
U is equivalent to a ∀∃-theory. Since U∀ = T∀ we therefore have U ⊆ T KH . So it
suffices to show that every model A of U is T -ec. The structure A embeds into a
model of T since A |= U∀ (= T∀ ) by Proposition 7.3.1. Let e be an embedding from
A to a model B of T , let ā be a tuple from A, and φ an existential formula with
B |= φ(e(ā)). Then B has an embedding g to a model C of U . Since U is a model
complete core theory, the embedding g ◦ h is elementary. Since g preserves existential
formulas, C |= φ(g(e(ā))). Since g ◦ e is elementary, A |= φ(ā).
(2) ⇒ (3). Lemma 7.5.2 implies that T KH is satisfied by all T -ec structures.
Together with (2) this implies that the first-order theory T KH axiomatises the class
of all T -ec structures.
(3) ⇒ (1). Suppose that the class of T -ec structures equals the class of all models
of a first-order theory U . We claim that U is the core companion of T . Every model
of U is in particular a model of T∀ , and every model of T∀ homomorphically maps to
a model of U by Lemma 7.4.2. So we only have to verify that U is model complete.
Let e be an embedding between models A and B of U , let ā be a tuple from A, and
let φ an existential formula. Since B satisfies T∀ and A is T∀ -ec we have that φ(ā),
as desired.
The final statement of the theorem is a clear consequence of the proof above.
Example 45. The theory T of groups does not have a model companion. By
Theorem 7.5.3 it suffices to show that the class of T -e.c. structures does not have a
first-order axiomatisation. In the following, a structure that is T -e.c. is called an
existentially closed group. We need the following group-theoretic fact, which follows
from Cayley’s representation theorem for groups.
Claim. a, b ∈ G have the same order in G if and only if there is a supergroup H
of G where a and b are conjugate, i.e., there exists an h ∈ H such that h−1 ah = b.
It follows from this claim that in existentially closed groups two elements have
the same order if and only if they are conjugate. Clearly, there are groups that
contain elements of arbitrarily large finite order. Every group can be extended to an
existentially closed group, so there is an existentially closed group G which contains
elements of arbitrarily large finite order. Therefore, for each n ∈ N there are elements
of order ≥ n which are not conjugate in G. By compactness, G has an elementary
extension G0 which has two elements of infinite order which are not conjugate in G0 .
Thus, G0 is not existentially closed, and the class of existentially closed groups is not
elementary. 4
Exercises.
(77) Prove that the model companion of the theory of equivalence relations is the
theory of equivalence relations with infinitely many infinite classes.
(78) Prove that the model companion of the theory of loopless undirected graphs
is the theory of the countably infinite random graph.
7.6. QUANTIFIER ELIMINATION 61
(79) Let us call a partial order (P ; ≤) tree-like if for all a, b, c ∈ P such that a ≤ b
and a ≤ c, then b and c must be comparable in (P ; ≤). Show that the theory
of tree-like partial orders has an ω-categorical model companion.
(80) (*) Does the theory of monoids have a model companion?
(81) (*) Does the theory of rings have a model companion?
(82) (*) Does the theory of abelian groups have a model companion?
7.6. Quantifier Elimination

In Section 6.2 we studied quantifier elimination in ω-categorical theories. In this
section we present characterisations of quantifier elimination for general theories.
Lemma 7.6.1. A first-order formula φ(x1 , . . . , xn ) is equivalent to a quantifier-
free formula modulo a theory T if and only if φ is preserved by isomorphisms between
substructures of models of T .
Proof. Let A, B be models of T and a ∈ An . Suppose that h is an isomorphism
between a substructure of A that contains a and a substructure of B. Then clearly,
if φ is quantifier-free and A |= φ(a), then B |= φ(h(a)).
Now suppose that conversely φ is a first-order formula that is preserved by isomor-
phisms between substructures of models of T . Let Ψ be the set of all quantifier-free
formulas ψ(x1 , . . . , xn ) that are implied by φ modulo T . To show that Ψ implies φ
modulo T , let A |= T and a ∈ An be such that A |= Ψ(a). Let Θ be the set of all
quantifier-free formulas θ such that a satisfies ¬θ in A. Then T ∪ {¬θ | θ ∈ Θ} ∪ {φ}
0
is satisfiable, because otherwise by compactness there would a be a finite
0
W subset Θ
of Θ such that T ∪ {¬θ | θ ∈ Θ } ∪ {φ} is unsatisfiable. But then ψ := θ∈Θ0 θ is a
quantifier-free formula such that T ∪ {φ} implies ψ, and hence ψ ∈ Ψ. So a satisfies
ψ in A, in contradiction to the assumption that A |= ¬θ(a) for all θ ∈ Θ. So there
exists a structure B and b ∈ B n which satisfy T ∪ ∪{¬θ(b) | θ ∈ Θ} ∪ {φ(b)}.
Let A0 be the smallest substructure of A that contains the entries of a and let B 0
be the smallest substructure of B that contains the entries of b. Since a and b satisfy
the same quantifier-free formulas, the map that sends bi to ai can be extended to an
isomorphism h between B 0 and A0 (see Exercise 7). By assumption, h preserves φ,
so A |= φ(a), and indeed ψ implies φ modulo T . The compactness theorem implies
that Ψ is equivalent to a single quantifier-free formula modulo T , as in the proof of
Theorem 7.1.2.
Lemma 7.6.2. A theory T has quantifier elimination if and only if every formula
∃x.φ where φ is quantifier-free is equivalent modulo T to a quantifier-free formula.
Proof. For the interesting direction, let φ be a first-order formula. We can
assume that φ is of the form Q1 x1 . . . Qn xn .ψ for Q1 , . . . , Qn ∈ {∃, ∀} and ψ quantifier-
free. We show the statement by induction on n. For n = 0 there is nothing to be
shown. First consider the case that Qn = ∃. By assumption, ∃xn .ψ is equivalent to
a quantifier-free formula ψ 0 , so φ is equivalent to Q1 x1 . . . Qn−1 xn−1 .ψ 0 , which is in
turn equivalent to a quantifier-free formula by induction.
Now consider the case that Qn = ∀. Then ∃xn .¬ψ is by assumption equivalent
to a quantifier-free formula ψ 0 , and φ is equivalent to Q1 x1 . . . Qn−1 xn−1 .¬ψ 0 , which
is equivalent to a quantifier-free formula by induction.
Lemma 7.6.3. A theory T has quantifier elimination if and only if for all models
B 1 and B 2 of T with a common substructure A, for all a ∈ An , and quantifier-free
formula ψ(y, x1 , . . . , xn ), if B 1 |= ∃y.ψ(y, a) then B 2 |= ∃y.ψ(y, a).
Proof. The forward implication is trivial. For the backward implication it suf-
fices to show that every formula φ(x1 , . . . , xn ) of the form ∃y.ψ, for ψ quantifier-free,
is equivalent modulo T to a quantifier-free formula (Lemma 7.6.2). By Lemma 7.6.1,
it suffices to show that φ is preserved by isomorphisms between substructures of mod-
els of T . So let B 1 , B 2 |= T and a ∈ B1n , and let h be an isomorphism between a
substructure that contains a and a substructure of B 2 . We identify the elements
of the substructures along h. Clearly, a quantifier-free formula holds on a in B 1 if
and only if it holds on h(a) in B 2 . The assumption then implies that B 1 |= φ(a)
if and only if B 2 |= φ(h(a)). Hence, φ is equivalent to a quantifier-free formula by
Lemma 7.6.1.
Example 46 (Ferrante and Rackoff [11], Section 3, Theorem 1). Let L be the
structure with domain Q and the signature {+, 1, ≤} where
• + is a binary relation symbols that denotes in L the usual addition over Q;
• 1 is a constant symbol that denotes 1 ∈ Q in L.
• ≤ is a binary relation symbol that denotes the usual linear order of the
rationals.
The structure L has quantifier elimination: TODO. 4
Example 47 (taken from [5]). Let H be the structure with domain Q and the
signature {1, <} ∪ {c·}c∈Q where
• 1 is a constant symbol that denotes 1 ∈ Q in L.
• < is a binary relation symbol that denotes the usual strict linear order of
the rationals.
• ·c, for c ∈ C, is a unary function symbols that denotes in L the function
x 7→ cx (multiplication by c).
The structure L has quantifier elimination: TODO. 4
Exercises.
(83) Find quantifier-free formulas that define the relations of H in L.
(84) Show that (Z; s) has quantifier elimination, where s : Z → Z is given by
x 7→ x + 1.
(85) Show that (Z; <, s) has quantifier elimination.
7.7. Algebraically Closed Fields

The model companion of the theory of fields is the theory ACF of algebraically
closed fields. This follows from the following important theorem.
Theorem 7.7.1 (Tarski). ACF has quantifier elimination.
Proof. We use Lemma 7.6.3. Let K 1 , K 2 |= ACF and R a common substructure;
then R is a ring, and in fact an integral domain (see Appendix B.1). TODO: explain
why. Let ψ(y, x1 , . . . , xn ) be quantifier-free and a ∈ Rn be such that K 1 |= ∃y.ψ(y, a).
We have to show that K 2 |= ∃y.ψ(y, a). For i ∈ {1, 2}, let fi be the unique embed-
ding of the field F of fractions of R into K i which is the identity on R (see Proposi-
tion B.1.2). By Corollary B.4.12, there exists an isomorphism g between the algebraic
closure G1 of f1 (F ) in K 1 and the algebraic closure G2 of f2 (F ) in K 2 which extends
f2 f1−1 .
Choose an element b1 ∈ K1 such that K 1 |= ψ(b1 , a). If b1 ∈ G1 , then K 2 |=
ψ(g(b1 ), a). Otherwise, b1 is transcendental, and there is an isomorphism between the
field extension G1 (b1 ) and the rational function field G1 (x) which fixes G1 (under the
usual identification of polynomials of degree 0 with elements of G1 ) and that maps b1
to x (see Section B.4.1).
7.7. ALGEBRAICALLY CLOSED FIELDS 63
• If K 2 is a proper extension of G2 , arbitrarily choose b2 ∈ K 2 \ G2 for b2 ;

note that b2 is transcendental over G2 , so G2 (b2 ) is isomorphic to G2 (x),
too. We can then extend g to an isomorphism between G1 (b1 ) and G2 (b2 )
which maps b1 to b2 and extends f2 f1−1 . Hence, K 2 |= ψ(g(b1 ), a).
• In the case that K 2 = G2 we take a proper elementary extension K 02 of K 2 ,
which exists by the theorem of Löwenheim-Skolem (Theorem 3.2.1) since
K 2 is infinite. Similarly as above it then holds that K 02 |= ∃y.ψ(y, a), and
therefore K 2 |= ∃y.ψ(y, a).
Quantifier elimination follows by Lemma 7.6.3.
The theory ACF is not complete. For any prime p let
ACFp := ACF ∪{1 + · · · + 1 = 0}
| {z }
p times
and for p = 0 let

ACF0 := ACF ∪{¬ 1 + · · · + 1 = 0 | n ∈ N}
| {z }
n times
be the theory of algebraically closed fields of characteristic p.
Theorem 7.7.2. For p = 0 or p prime, the theory ACFp is κ-categorical for any
κ > ℵ0 .
Proof. Let E be a model of ACFp and let F be the smallest subfield of E (also
called the prime field ; e.g., if p = 0 then F = (Q; +, ∗, 0, 1)). Let G be the algebraic
closure of F . Then any two algebraically closed fields of the same characteristic and
of the same cardinality κ > ℵ0 have transcendence bases over G (see Section B.4.4)
of cardinality κ. Any bijection between these transcendence bases induces an isomor-
phism of the fields.
Note that infinite fields K are never ω-categorical since there are (non-isomorphic!)
countable models of Th(K) of transcendence degree 0, 1, 2, . . . .
Corollary 7.7.3. For p = 0 or p prime, ACFp is a complete theory.
Proof. Follows from Theorem 7.7.2 by Vaught’s test (Theorem 3.3.2).
We also obtain the following ‘cross-characteristic transfer’ result.
Theorem 7.7.4 (Lefschetz principle). Let φ be a first-order sentence over the
signature of rings. Then the following are equivalent.
(1) φ holds in the field of complex numbers.
(2) ACF0 |= φ
(3) ACFp |= φ for all but finitely many prime numbers p.
Proof. (1) ⇔ (2) follows from the fact that the complex numbers are alge-
braically closed, and that ACF0 is complete.
(2) ⇒ (3). Suppose that ACFp |= φ co-finitely often. We claim that T :=
ACF0 ∪{φ} is satisfiable. A finite subset T 0 of T will mention only finitely many
inequalities p 6= 0, so taking q large enough, ACFq |= T 0 . By compactness, T is
satisfiable. Since ACF0 is complete, ACF0 |= φ.
(3) ⇒ (2). Suppose that infinitely often, ACFp 6|= φ. Since ACFp is complete, this
means that ACFp |= ¬φ co-finitely often, and we show as above that ACF0 |= ¬φ.
Corollary 7.7.5 (Hilbert’s Nullstellensatz). Let K be an V algebraically closed
field and let f0 , . . . , fm ∈ K[x1 , . . . , xn ]. Then K |= ∃x1 , . . . , xn i∈{1,...,n} fi = 0 if
and only if I = (f1 , . . . , fm ) is a proper ideal, i.e., 1 ∈/ I.
Proof. It is clear that if a1 , . . . , an ∈ K are such that f1 (a1 , . . . , an ) = · · · =

fm (a1 , . . . , an ), and f ∈ I, then f (a1 , . . . , an ) = 0, so 1 ∈
/ I.
For the converse, an application of Zorn’s lemma to the set S of proper ideals
J in K[x1 , . . . , xn ] containing I gives a maximal ideal J ∈ S. Since J is maximal,
L := K[x1 , . . . , xn ]/J is a field (Proposition B.2.7). We identify K with a subfield of
L by identifying a ∈ K with a + I. For each i ≤ k we have
fi (x1 + J, . . . , xn + J) = fi (x1 , . . . , xn ) + J = J
V
since fi (x1 , . . . , xn ) ∈ I ⊆ J. Hence ∃x1 , . . . , xn i∈{1,...,n} fi = 0 holds in L, and
thus also in its algebraic closure. By quantifier-elimination for ACF, this is also true
for the substructure K |= ACF.
Exercises
(86) Find a quantifier-free formula which is equivalent to ∃x(ax2 + bx + c = 0)
over (C; +, ∗).
(87) Find a quantifier-free formula which is equivalent to the formula of the pre-
vious exercise over (R; +, ∗).
(88) Write down a first-order φ(a, b, c, d) in the language of fields which
formula
a b
states that the matrix is invertible. Find a quantifier-free formula
c d
which is equivalent to φ for any field F .
(89) (Exercise 2.3.3 in [29]) Show that a ∀∃-sentence which holds in all finite
fields is true in all algebraically closed fields.
(90) Show that in every algebraically closed field K, every injective polynomial
map of a definable subset of K n to K is surjective.
CHAPTER 8
Stability
Stability theory grew out of the attempt to understand the possible spectra of
first-order theories (Section 1.3). Today, stability theory might be thought of as a
way to classify definable sets in a structure whenever this is possible. The starting
point is that instead of counting the number of models of a first-order theory T , we
count the number of types in models of T over parameter sets of some cardinality.
This chapter presents some of the basic and fundamental facts of stability theory, and
is not yet finished.
Definition 8.0.1. Let κ be an infinite cardinal. A theory T is κ-stable if for
M
every M |= T and for every A ⊆ M , if |A| ≤ κ then |S1 (A)| ≤ κ. If T is κ-stable
for some κ, then T is called stable, and it is called unstable otherwise. A structure is
called κ-stable if its first-order theory is κ-stable.
Example 48. Let E be an equivalence relation on N with infinitely many infinite
classes. Then the structure (N; E) is stable. 4
Example 49. Let p be a prime or 0. Algebraically closed fields of characteristic p
are κ-stable for all κ. Recall that ACFp is κ-categorical for all κ > ω (Theorem 7.7.2).
We show in Theorem 8.3.2 below that uncountable categoricity implies ω-stability.
Then we show that ω-stability implies κ-stability for all κ in Proposition 8.3.3. Alter-
natively, it is easy to prove κ-stability directly, using quantifier elimination in ACF
(Theorem 7.7.1). 4
The following lemma shows that it does not matter whether we define stabil-
ity with respect to 1-types or with respect to n-types (similarly as with saturation,
Exercise 54).
Lemma 8.0.2. Let κ be an infinite cardinal and let T be κ-stable. Then T is
κ-stable for n-types, for every n ∈ N, i.e., for every model M of T and A ⊆ M of
M
cardinality at most κ we have |Sn (A)| ≤ κ.
Proof. Our proof is by induction on n. For n = 1 the statement holds by
assumption, so suppose that n > 1. Let M be a model of T and A ⊆ M of cardinality
at most κ. By Lemma 5.2.3, M has an elementary extension N that realises all
1-types over A. For an n-type p of M over A, let p0 be the 1-type that contains
all sentences of the form ∃x2 , . . . , xn .φ for φ ∈ p. By assumption, there are only
countably many 1-types of M over A. Therefore, it suffices to show that every n-type
N
p, the set Q : {q ∈ Sn (A) | q 0 = p0 } has cardinality at most κ, because ω · κ = κ.
By assumption, p0 is realised by an element b ∈ N in N . Let p00 be the (n − 1)-
type obtained from p by replacing x1 by the constant b. Then the (n − 1)-types of
M
N over A ∪ {b} are in bijective correspondence to the set of all q ∈ Sn (A) such that
0 0
q = p . By inductive assumption, there are at most κ many (n − 1)-types of N over
A ∪ {b}.
Stability has many equivalent characterisations; one is based on the binary tree
property and another one on the order property; they will be presented in Section 8.1.
65
66 8. STABILITY
Another equivalent characterisation is based on indiscernible sequences, a concept

for convenient use of Ramsey’s theorem from combinatorics (which is recalled in
Appendix C); indiscernible sequences are the topic of Section 8.2. We continue with
the strongest form of stability, namely ω-stability, in Section 8.3, and prove that every
uncountably categorical theory is ω-stable.
Unstable theories have the independence property or the strict order property
(Section 8.4). Many of the good properties of stable theories also hold for theories
that do not have the independence property; such theories are called NIP and they
are currently intensively studied in model theory [28]. They also found applications
in structural graph theory and theoretical computer science (see [6] and the references
therein). We present an interesting connection between NIP theories and the concept
of VC dimension from computational learning theory (Section 9.2).
For further reading on stability theory, I recommend [9].
8.1. The Order Property

Stability has many equivalent characterisations, some of which will be presented
in this section. We start with the observation that if a theory is unstable, then
this is witnessed by a single formula (Proposition 8.1.2). Such witnessing formulas
have many equivalent characterisations (Theorem 8.1.5). Throughout this section, T
denotes a complete theory over the signature τ .
The following definitions about types were essentially already given in Chapter 5.
Let M be a model of T and B ⊆ M . For a τ -formula φ(x̄, ȳ), a (complete) φ-type in
M over B is a maximal set p of formulas of the form φ(x̄, b̄) or of the form ¬φ(x̄, b̄)
for tuples b̄ of elements of B such that p is satisfiable over some elementary extension
M
of M . Define Sφ (B) to be the set of all φ-types in M over B.
Definition 8.1.1. A formula φ is called stable in T if for every infinite model B
B
of T we have |Sφ (B)| ≤ |B|, and unstable otherwise.
The terminology is motivated by the following corollary.
Proposition 8.1.2. T is unstable if there exists a τ -formula φ which is unstable
in T .
Proof. Suppose that φ(x̄, ȳ) is unstable in T , that is, there exists an infinite
B
model B of T such that |Sφ (B)| > |B| It follows that T is not |B|-stable, and hence
unstable.
Now suppose that every τ -formula φ(x̄, ȳ) is stable. Let M be a model of T and
M
let B ⊆ M be of some infinite cardinality κ. Then |Sφ (B)| ≤ |B|. Since every
M M
p ∈ S1 (B) is determined by {q ∈ Sφ (B) | φ a τ -formula, q ⊆ p} we have that
M
|S1 (B)| ≤ κω = κ and hence T is stable.
How can we show that a formula φ is unstable in T ? We have to find for every
model of T and every cardinal κ strictly more than κ φ-types if we are allowed κ many
parameters. The idea how to find so many types is to construct a big tree where the
infinite branches correspond to φ-types.
If A and B are sets, we write AB for the set of all functions from B to A. If
w ∈ An then we define |w| := n and treat w as a word whenever this is convenient;
e.g., 0100 denotes a word w ∈ {0, 1}4 . For an ordinal µ we write A<µ for n<µ An .
S
A binary tree over A is a family of elements of A of the form (as )s∈{0,1}<ω .

Definition 8.1.3. A formula φ(x̄, ȳ) has the binary tree property in T if there
exists a model M of T and a binary tree (b̄w )w∈{0,1}<ω of tuples of elements of M
8.1. THE ORDER PROPERTY 67
such that for every w ∈ {0, 1}<ω

{φ(x̄, b̄w|n ) | n < ω, w(n) = 0} ∪ {¬φ(x̄, b̄w|n ) | n < ω, w(n) = 1}.
is satisfiable in M .
Example 50. If M = (Q; <) then the formula φ(x, y1 , y2 ) given by y1 < x < y2
has the binary tree property in Th(M ). 4
The next property, called the order property, is often more convenient to verify
than the binary tree property, but in fact implies the binary tree property.
Definition 8.1.4. Let φ(x̄, ȳ) be a first-order formula. Then φ has the order
property in T if there exists a model M of T and sequences (āi )i∈ω and (b̄i )i∈ω of
tuples of elements of M such that
M |= φ(āi , b̄j ) ⇔ i < j.
Example 51. If M = (Q; <) then the formula φ(x, y) given by x < y has the
order property in Th(M ). 4
Example 52. If M = (V; E) is the Rado graph then the formula E(x, y) has the
order property in Th(M ): pick sequences (an )n∈ω and (bn )n∈ω of elements of V such
that E(ai , bj ) if and only if i < j; such elements exist by the universality of the Rado
graph. 4
Theorem 8.1.5. Let T be a theory and let φ(x̄; ȳ) be a τ -formula. The following
are equivalent.
B
(1) φ is stable in T , i.e., for every infinite model B of T we have |Sφ (B)| ≤ |B|.
(2) For every model M of T there exists an infinite cardinal λ such that for
M
every B ⊆ M with |B| ≤ λ we have |Sφ (B)| ≤ |B|.
(3) φ(x̄; ȳ) does not have the binary tree property in T .
(4) φ(x̄; ȳ) does not have the order property in T .
Proof. The implication from (1) to (2) is immediate.
(2) ⇒ (3): We prove the contraposition. Let φ(x̄, ȳ) be a formula with the binary
tree property in T and let λ be an infinite cardinal. Choose µ minimal such that
λ < 2µ . Then |{0, 1}<µ | ≤ λ. The binary tree property and compactness imply that
T has a model M and a sequence (b̄ρ )ρ∈{0,1}<µ of tuples of elements of M such that
for every σ ∈ {0, 1}<µ
pσ := {φ(x̄, b̄σ|α ) | α < µ, σ(α) = 0} ∪ {¬φ(x̄, b̄σ|α ) | α ≤ µ, σ(α) = 1}
M
is satisfiable in M . Complete every pσ to an element of Sφ (B) where B is the set of
all entries of all tuples in (b̄ρ )w∈{0,1}<µ . Since all the qσ are pairwise distinct, we have
M
|B| ≤ λ < 2µ ≤ Sφ (B).
(3) ⇒ (4): We again prove the contraposition. Let φ(x̄, ȳ) be a formula with the
order property in T , so there is a model M and sequences (ā)i∈ω and (b̄)i∈ω such that
M |= φ(āi , b̄j ) ⇔ i < j. Let < be the infix ordering on I := {0, 1}ω ∪ {0, 1}<ω , which
is defined for u, v ∈ I by
u<v if v = u1x for some x ∈ I
or u = w0x and v = w1y for some w ∈ {0, 1}<ω and x, y ∈ I.
It suffices to prove that there are sequences (ā0u )u∈I and (b̄0u )u∈I of tuples of M such
that M |= φ(ā0u , b̄0v ) if and only if u < v. The existence of such sequences follows from
the compactness theorem and the existence of (ā)i∈ω and (b̄)i∈ω .
68 8. STABILITY
(4) ⇒ (1). Again we prove the contraposition. Let M be an infinite model of T

M
such that |Sφ (M )| > |M |. Let M 0 be an elementary extension of M that realises
M M
all the φ-types in Sφ (M ) (Lemma 5.2.3). Let |ȳ| = n. Note that every p ∈ Sφ (M )
is uniquely given by the set Sp ⊆ M n of all b̄ ∈ M n such that φ(x̄, b̄) ∈ p. Let
M
S := {Sp | p ∈ Sφ (M )}. Applying the Erdős-Makkai theorem (Theorem C.0.5) to
A := M n and S we obtain a sequence (ai )i∈ω of elements of A and a sequence (Si )i∈ω
of elements of S such that either for all i, j ∈ ω we have ai ∈ Sj ⇔ i < j, or for all
i, j ∈ ω we have ai ∈ Sj ⇔ j < i. In the first case, φ(x̄, ȳ) has the order property in
T . In the second case, note that the definition of the order property is symmetric, so
also in the second case φ has the order property in T .
Stability has another beautiful reformulation, based on item (2) of Theorem 8.1.5.
Theorem 8.1.6. T is unstable if and only if there exists a τ -formula φ(x̄, ȳ), a
model M of T , and a sequence (c̄i )i∈ω such that M |= φ(c̄i , c̄j ) ⇔ i < j.
Proof. Let ψ(x̄, ȳ) be a τ -formula so that there are a model M of T and se-
quences (āi )i∈ω and (b̄i )i∈ω of tuples from M such that M |= φ(āi , b̄j ) if and only if
i < j. Let ψ(x̄1 , x̄2 , ȳ1 , ȳ2 ) be the formula φ(x̄1 , ȳ2 ). For i ∈ ω, define ci := (ai , bi ).
Then M |= ψ(c̄i , c̄j ) if and only if M |= φ(ai , bj ) if and only if i < j. The backwards
direction is immediate, since we may take (bi )i∈ω = (ai )i∈ω = (ci )i∈ω .
8.2. Indiscernible Sequences

Let A be a τ -structure. Let I be a linearly ordered set and (ai )i∈I is a se-
quence of elements of An for some n ∈ N. Then (ai )i∈I is called an order in-
discernible sequence (in A) 1 if for every τ -formula φ(x̄1 , . . . , x̄k ) and all elements
i1 , . . . , ik , j1 , . . . , jk ∈ I with i1 < · · · < ik and j1 < · · · < jk we have that A |=
φ(āi1 , . . . , āik ) ⇔ φ(b̄i1 , . . . , b̄ik ). Note that if āi = āj for distinct elements i, j ∈ I,
then ai = aj for all elements i, j ∈ I. This is why we usually assume that the āi are
pairwise distinct.
Example 53. Any strictly increasing or strictly decreasing sequence of elements
of A := (Q; <) is order indiscernible. 4
Example 54. Let A be a vector space and let X a basis of A. Then any sequence
(ai )i∈I of elements of X is order indiscernible in A. To see this, let φ(x1 , . . . , xk ) be a
τ -formula and let i1 , . . . , ik , j1 , . . . , jk ∈ I be such that i1 < · · · < ik and j1 < · · · < jk .
Since X is a basis of A, there is an automorphism of A which takes ā := (ai1 , . . . , aik )
to b̄ := (bj1 , . . . , bjk ). Hence, A |= φ(ā) if and only if A |= φ(b̄). 4
Remark 8.2.1. The order indiscernible sequences in the previous example are
indiscernible in a very strong sense. We say that (ai )i∈I is totally indiscernible (in
A) if (ai )i∈I is order indiscernible for every linear ordering of I. Example 53 shows
an order indiscernible set which is not totally indiscernible.
Definition 8.2.2. Let A be a τ -structure. Let I be an infinite linearly ordered set
and let (āi )i∈I be a sequence of elements of An for some n ∈ N. Then the Ehrenfeucht-
Mostowski type EM((āi )i∈I ) is the set of τ -formulas φ(x̄1 , . . . , x̄k ), k ∈ N, such that
φ(āi1 , . . . , āik ) holds for all i1 , . . . , ik ∈ I with i1 < · · · < ik in some elementary
extension of A.
1Order indiscernible sequences are sometimes just called indescernible.

8.2. INDISCERNIBLE SEQUENCES 69
Lemma 8.2.3. Let A be an infinite τ -structure. Let I be an infinite linearly ordered

sets and let (āi )i∈I be a sequence of elements of An for some n ∈ N. Then for any
linearly ordered set J there exists a model B of Th(A) and an order indiscernible
sequence (b̄i )i∈J of elements of B n with the same Ehrenfeucht-Mostowski type as
(āi )i∈I .
Proof. By Ramsey’s theorem and compactness. For every j ∈ J and i ∈
{1, . . . , n} let cij be a new constant symbol; we write c̄j for (c1j , . . . , cnj ). Let C :=
{c̄j | j ∈ J}. Define
S := {φ(c̄i1 , . . . , c̄ik ) | k ∈ N, φ(x̄1 , . . . , x̄k ) ∈ EM((āi )i∈I ), i1 , . . . , ik ∈ J, i1 < · · · < ik }
T := {φ(c̄i1 , . . . , c̄ik ) ⇔ φ(c̄j1 , . . . , c̄jk ) | k ∈ N, i1 , . . . , ik , j1 , . . . , jk ∈ J,
i1 < · · · < ik , j1 < · · · < jk , φ a τ -formula}
We have to show that Th(A) ∪ S ∪ T is satisfiable. By compactness, it is enough to
show that for every k ∈ N, every finite set ∆ of τ -formulas φ(x̄1 , . . . , x̄k ), and every
finite subset C0 ⊆ C the theory Th(A) ∪ S 0 ∪ T 0 is satisfiable where
S 0 := {φ(c̄i1 , . . . , c̄ik ) ∈ S | c̄i1 , . . . , c̄ik ∈ C0 }
T 0 := {φ(c̄i1 , . . . , c̄ik ) ⇔ φ(c̄j1 , . . . , c̄jk ) ∈ T | φ ∈ ∆, c̄i1 , . . . , c̄ik , c̄j1 , . . . , c̄jk ∈ C0 }.
Define an equivalence relation ∼ on the set D of k-tuples (c̄i1 , . . . , c̄ik ) ∈ C k such that
i1 < · · · < ik : for ū, v̄ ∈ D, define ū ∼ v̄ if for all φ ∈ ∆
A0 |= φ(ū) ⇔ φ(v̄)
holds in some elementary extension A0 of A. Since this equivalence relation has at
most 2∆ classes, by Ramsey’s theorem there exists an infinite subset C1 ⊆ C such that
all tuples in C1k ∩ D lie in the same class. Let {c̄p1 , . . . , c̄ps } = C0 with p1 < · · · < ps .
Choose c̄q1 , . . . , c̄qs ∈ C1 with q1 < · · · < qs . Then the expansion of the τ -reduct of
A0 by interpreting c̄pi by c̄qi is a model of Th(A) ∪ S 0 ∪ T 0 .
Corollary 8.2.4. Let T be a τ -theory with infinite models. Then for any linearly
ordered set I there exists a model A of T with an order indiscernible sequence (ai )i∈I
of distinct elements of A.
Proof. By the upwards Löwenheim-Skolem theorem (Theorem 3.2.1) there ex-
ists a model of T with a sequence of distinct elements indexed by I. Apply Lemma 8.2.3
to this sequence.
Exercises
(91) Let T be the theory ACF0 of algebraically closed fields of characteristic 0.
Find an order indiscernible sequence (ai )i∈ω of distinct elements in a model
of T .
Indiscernible sequences can also be used to characterise stability.
Theorem 8.2.5. T is stable if and only if every order indiscernible sequence in
every model M of T is totally indiscernible.
Proof. If T is unstable, then by Theorem 8.1.6 there exists a τ -formula φ(x̄, ȳ),
a model M of T , and a sequence (āi )i∈ω of tuples of elements of M such that for all
i, j ∈ ω we have M |= φ(āi , āj ) ⇔ i < j. Then (āi )i∈ω is not totally indiscernible.
Conversely, suppose that T has a model A with an order indiscernible sequence
(āi )i∈ω which is not totally indiscernible. By Lemma 8.2.3 T has a model B with
an order indiscernible sequence (b̄i )i∈Q of the same Ehrenfeucht-Mostowski type as
(ai )i∈ω ; note that (b̄i )i∈Q is in particular not totally indiscernible. This means that
70 8. STABILITY
there exists a formula φ(x̄1 , . . . , x̄k ), some indices i1 , . . . , ik ∈ ω with i1 < · · · < ik and
some permutation σ of {1, . . . , k} such that B |= φ(b̄i1 , . . . , b̄ik )∧¬φ(b̄σ(i1 ) , . . . , b̄σ(ik ) ).
Any permutation on {1, . . . , k} is a product of transpositions of consecutive elements
of {1, . . . , k}, and hence we may assume that σ is a transposition that exchanges j, j +
1 ∈ {1, . . . , k} and fixes all elements of {1, . . . , k}. Then the formula φ(x1 , x2 ) given
by φ(ā1 , . . . , āj , x1 , x2 , āj+2 , . . . , āk ) orders the infinite sequence (āi ){i∈I|ij <i<ij+1 } .
Hence, Theorem 8.1.6 implies that T is unstable.
Lemma 8.2.6. Let τ be a countable signature and let M be a τ -structure generated
by a well-ordered set I of indiscernibles. Then M realises only countably many types
over every countable subset of M .
Proof. First consider the simpler case where the signature is relational. Let
M
S ⊆ M be countable. Suppose that a realises p ∈ S1 (S). Then p is determined
by the position of a with respect to the well-ordered subset S, and there are only
countably many possibilities. The general case with function symbols can be shown
similarly.
We conclude this section by showing that every countable theory has a model
that realises only countably many types over every countable subset.
Definition 8.2.7. Let τ be a signature. A Skolem theory for τ is a theory S in
a bigger signature σ such that
(1) |σ| ≤ max(|τ |, ω).
(2) S is universal.
(3) Every τ -structure can be expanded to a model of S.
(4) S has quantifier elimination.
Theorem 8.2.8. For every signature τ there exists a Skolem theory S.
Proof. We define an ascending sequence of signatures τ = σ0 ⊆ σ1 ⊆ · · · by
introducing for every quantifier-free {σi }-formula φ(x1 , . . . , xn , y) a new n-ary Skolem
function fφ and
S defining σi+1 as the union of σi and the set of these function symbols.
Define σ := i σi . Now define
S := {∀x̄(∃yφ(x̄, y) ⇒ φ(x̄, fφ (x̄)) | φ(x̄, y) quantifier-free σ-formula}.
The first two items are clearly satisfied. For the third item, let A be a τ -structure.
To define the expansion B |= S of A, suppose f ∈ σ is n-ary, and let a ∈ An .
If A |= ∃yφ(a, y) for φ quantifier-free, then choose for f B (a) any b ∈ A such that
A |= φ(a, b). Clearly, B |= S. We leave the proof that S has quantifier elimination as
an exercise.
Lemma 8.2.9. Let T be a countable τ -theory with an infinite model and let κ be
an infinite cardinal. Then T has a model of cardinality κ which realises only countably
many types over every countable subset.
Proof. Let τ be the signature of T , and consider the theory T ∗ := T ∪ S where
S is a Skolem theory for τ with signature σ (see Theorem 8.2.8). Then T ∗ is count-
able, has an infinite model and quantifier elimination. Since S is universal and has
quantifier-elimination, T ∗ is equivalent to a universal theory. By Lemma 8.2.3, T ∗
has a model N ∗ with a set of order indiscernibles I = {ai | i ∈ κ}. Then M ∗ := N ∗ [I]
is a model of T ∗ of cardinality κ. Since T ∗ has quantifier elimination, M ∗ is an ele-
mentary substructure of N ∗ and I is order indiscernible in M ∗ . By Lemma 8.2.6 the
structure M ∗ realises only countably many types over every countable subset of M ,
and the same is true for the τ -reduct M of M ∗ .
8.4. THE CLASSIFICATION PICTURE 71
8.3. ω-stable Theories

An example of an ω-stable structure is (Q; =). Together with the following theo-
rem, it provides many examples of (ω-categorical) ω-stable structures.
Proposition 8.3.1. Let A be an ω-categorical ω-stable structure. Then every
structure B with a first-order interpretation in A is ω-stable.
Proof. Recall that since B has a first-order interpretation in an ω-categorical
structure, it is ω-categorical as well (Lemma 6.3.2). Let n be the dimension of the
interpretation of B in A. Suppose that M is a model of Th(B) and S ⊆ M is countable
M
such that S1 (S) is uncountable. By the Theorem of Löwenheim-Skolem downwards
(Theorem 3.2.1) M has a countable elementary substructure C that contains S. By
the ω-categoricity of B the structure C is isomorphic to B, so we may identity B
and C. For each b ∈ S, fix an n-tuple of elements of A that is mapped to b by the
interpretation; let T be the countable set of entries of all these tuples. Let c, d ∈ B
and let c0 and d0 be n-tuples of elements of A that are mapped to c and to d by the
interpretation. If c, d ∈ B have distinct 1-types of A over S, then c0 and d0 have
distinct n-types over T . Lemma 8.0.2 then implies that A also has uncountably many
1-types over T .
Many more examples of ω-stable theories come from the following theorem.
Theorem 8.3.2. A countable theory T which is categorical in an uncountable
cardinal κ is ω-stable.
M
Proof. Let M be a model and A ⊆ M countable with S1 (A) uncountable. Let
(bi )i∈I be a sequence of ℵ1 many elements with pairwise distinct types over A. We
assume that all types over A are realised in M : if not, we replace M by an elementary
extension of M where all types over A are realised. By Löwenheim-Skolem downwards
(Theorem 3.2.1) there is an elementary substructure M 0 of M of cardinality ℵ1 which
contains A and all bi . By Löwenheim-Skolem upwards (Theorem 3.2.1) there is an
elementary extension N of M 0 of cardinality κ. The model N realises uncountably
many types over A. By Lemma 8.2.9, the theory T has another model in which only
countably many types over A are realised. So T is not κ-categorical.
Proposition 8.3.3. If T is ω-stable, then T is κ-stable for all κ.
Proof. Suppose that T is not κ-stable for some κ, i.e., there is a model M and
M
A ⊆ M such that |A| ≤ κ and |S1 (A)| > κ. Recall our notation from Section 5.1: for
M
a formula φ(x) in the language of M A we write [φ] for the set of all p ∈ S1 (A) that
contain φ. By assumption, |[x = x]| > κ. Suppose now that |[φ]| > κ. Note that every
formula ψ such that |[ψ]| ≤ κ belongs to at most κ types in [φ], and that there are at
most κ formulas. Hence, there must be distinct types p, q such that |[ψ]| > κ for all
formulas ψ in p and in q. Let ψ ∈ p \ q. We then decompose [φ] = [φ ∧ ψ] ∪ [φ ∧ ¬ψ].
Inductively, we can now construct an infinite complete binary tree of formulas. Only
a countably infinite subset B of constants from A are used in these formulas, but
M
S1 (B) = 2ω , showing that T is not ω-stable.
8.4. The Classification Picture

The failure of stability can occur in two different ways.
Definition 8.4.1 (The strict order property). Let T be a theory. A formula
formula φ(x̄, y1 , . . . , ym ) has the strict order property (SOP) in T if there exists A |= T
and a sequence (bi )i∈ω of elements of Am such that φ(x̄, bi )A ( φ(x̄, bj )A for all i < j.
We say that T has the SOP if there exists a formula that has the SOP in T .
72 8. STABILITY
Example 55. A typical theory with the SOP is the theory of (Q; <). 4
Proposition 8.4.2. If a theory T has the SOP, then it is unstable.
Proof. Let φ(x̄, y1 , . . . , ym ) be a formula, let A be a model of T , and let (bi )i∈ω
be a sequence of elements of Am that witness that T has the SOP. Let ψ(x̄, x̄0 ) be the
formula ¬∃ȳ. φ(ȳ, x̄) ∧ ∃ȳ. φ(ȳ, x̄). Then A |= ψ(bi , bj ) if and only if i < j, and hence
Theorem 8.1.6 implies that T is unstable.
Let T be a τ -theory and let φ(x1 , . . . , xn ; y1 , . . . , ym ) be a first-order τ -formula.

If M is a model of T , then A ⊆ M n is shattered by φ if there exists a family {bI ∈
M m | I ⊆ A} such that
M |= φ(a, bI ) if and only if a ∈ I.
By compactness, A is shattered by φ if and only if every finite subset of A is shattered
by φ.
Definition 8.4.3. Let T be a theory. A formula φ(x1 , . . . , xn ; ȳ) has the inde-
pendence property (IP) in T if T has a model A with an infinite subset of An that is
shattered by φ. Otherwise, we say that φ is dependent in T . A theory T has the IP
if there exists a formula that has the IP in T ; otherwise, we say that T has the NIP.
Example 56. The theory of (N; ·) is unstable, witnessed by the formula φ(x, y)
given by
∃k. y · k = x
(i.e., the formula expresses that x divides y). Then φ(x, y)Qshatters the set of prime
numbers: for any subset I of prime numbers define bI := I. Then we have M |=
φ(a, bI ) if and only if a ∈ I. 4
Remark 8.4.4. To prove that φ has the IP, the by compactness of first-order
logic it suffices to find arbitrarily large finite subsets of An that are shattered by φ.
Example 57. A typical example of a theory that has the IP is the theory T of
the Rado graph (V; E) (Example 26). Let S be any finite subset of V. Then for any
U ⊆ S there exists a vertex v ∈ V such that for every x ∈ A we have E(x, v) ⇔ x ∈ U .
Therefore, the formula E(x, y) shatters S, and Remark 8.4.4 implies that T has the
IP. 4
Example 58. If T is the theory of an infinite Boolean algebra (A; 0, 1, ¬, ∨, ∧),
then the formula φ(x; y) := (x ∧ y = x) is independent in T : indeed, let W A be a set
such that x ∧ b = 0 for all distinct
W a, b ∈ A.
W For any I ⊆ N , set b I to be i∈I bi ; then
T |= φ(x; y) if and only if x ∧ i∈I bi = i∈I (x ∧ bi ) = x if and only if x ∈ I. 4
Proposition 8.4.5. If a theory T has the IP, then it is unstable.
Proof. Let M |= T and let A ⊆ M n be infinite and shattered by the formula
φ(x1 , . . . , xn , y1 , . . . , ym ), witnessed by the family {bI ∈ M m | I ⊆ A}. Let a1 , a2 , . . .
be distinct elements of A. Set ci := b{a1 ,...,ai−1 } . Then note that M |= φ(ai , cj ) if
and only if i < j, so φ has the order property. Theorem 8.1.5 then implies that φ is
unstable in T .
Proposition 8.4.6. A formula φ(x1 , . . . , xn , y1 , . . . , ym ) is NIP in T if and only

if for any model A |= T , every indiscernible sequence (bi )i∈I of elements of Am , an
every a ∈ An there are at most n indices i0 < · · · < in−1 such that for all i < n − 1
A |= φ(a, bi ) ⇔ ¬φ(a, bi+1 ).
8.4. THE CLASSIFICATION PICTURE 73
Proof. If φ has the IP, then by Lemma 8.2.3 we can choose an indiscernible
sequence (bi )i∈ω of elements in Am such that {bi | i ∈ ω} is shattered by φ. Let
s ⊆ ω be the set of even integers. Then φ(x̄, bs ) satisfies the statement of the lemma:
TODO.
Theorem 8.4.7. A theory T is unstable if and only if it has the IP or the SOP.
Proof. We have already proved the backwards direction in Proposition 8.4.2
and Proposition 8.4.5.
By Theorem 8.1.6 there exists a formula φ(x̄, ȳ), a model M |= T , and a sequence
(c̄i )i∈ω such that M |= φ(c̄i , c̄j ) ⇔ i < j. By Lemma 8.2.3 we may assume that there
exists a order indiscernible sequence (b̄i )i∈Q such that TODO.
Since T is NIP, Proposition TODO

CHAPTER 9
NIP Theories
UNDER CONSTRUCTION.
9.1. VC Dimension
We have already mentioned in Remark 8.4.4 that if φ(x1 , . . . , xn ; ȳ) is dependent
in T , then by compactness there is some integer k such that no subset A of M n of
size k is shattered by φ. The maximal k for which there is some A of size k that is
shattered by φ is called the VC-dimension of φ (called after Vapnik-Chervonenkis [30]
and discovered in the context of machine learning). If φ is independent, we say that
the VC-dimension of φ is infinite.
Example 59. The formula x ≤ y is dependent in Th(Q; ≤), and has VC-
dimension 1. indeed, A = {0} is shattered by φ because we can choose bA := 0
and b∅ := −1. On the other hand, for A = {0, 1} we cannot find a b1 such that
(Q; ≤) |= ¬(0 ≤ b1 ) ∧ 1 ≤ b1 . 4
Example 60. We have seen in Example 56 that the formula φ(x; y) := ∃z. x·z = y
is independent in T := Th(N; ·), and hence has infinite VC dimension. 4
Example 61. In the countable random graph, every finite subset of vertices is
scattered by the formula E(x, y), so this formulas has infinite VC dimension. 4
If φ(x̄; ȳ) is a formula, we write φ− (ȳ; x̄) for φ(x̄; ȳ) (so it is the same formula,
but we change the role of the variables and parameters).
Lemma 9.1.1. The formula φ(x̄; ȳ) is NIP in T if and only if the formula φ− (ȳ; x̄)
is NIP in T .
Proof. Let M be an uncountable model of T . If φ(x̄; ȳ) is independent then by
compactness, there is an uncountable set A = {aJ | J ⊆ ω} which is shattered by φ,
as witnessed by tuples bI for I ⊆ A. For j ∈ ω, let Ij := {X ⊆ ω | j ∈ X}. We will
show that B := {bIj | j ∈ ω} is shattered by φ(ȳ; x̄). Let K ⊆ ω. Then we have
T |= φ− (bIj , aK ) if and only if T |= φ(aK ; bIj )
if and only if K ∈ Ij
if and only if j ∈ K .
−
Therefore φ (ȳ; x̄) is independent.
Exercises.
(92) Show that the VC dimension of φ(x̄; ȳ) can be different from the VC dimen-
sion of φ− (ȳ, x̄).
Lemma 9.1.2. The formula φ(x; y) is independent in T if and only if there is an
indiscernible sequence (ai | i ∈ ω) and a tuple b such that T |= φ(ai ; b) if and only if
i is even.
75
76 9. NIP THEORIES
Proof. (⇐): (⇒): Assume that φ(x; y) in independent in T . Let A = (ai | i ∈ ω)

be a sequence of |x|-tuples that is shattered by φ(x; y).
Proposition 9.1.3. The formula φ(x; y) is NIP in a theory T if and only if for
any model M of T and. every indiscernible sequence (ai ∈ M | i ∈ ω) and tuple b,
there is some end segment S and ∈ {0, 1} such that φ(ai ; b) holds for any i ∈ S.
Lemma 9.1.4. Let T be a theory. Then a Boolean combination of NIP formulas
is NIP.
Proof. It is clear from the definition that the negation of a NIP formula is NIP.
Let φ(x; y) and ψ(x; y) be two NIP formulas and we want to show that θ(x; y) := φ∧ψ
is NIP. TODO: finish.
A τ -theory T is NIP if all τ -formulas φ are dependent in T . Note that if T is
dependent, then also all formulas with parameters are dependent, since if φ(x; y, d)
has IP, the so does φ(x; y, z).
Proposition 9.1.5. Let T be a τ -theory. Then T is NIP if and only if all τ -
formulas φ(x; y) with |y| = 1 are dependent.
Proof. For the interesting direction of the statement, assume that all formulas
φ(x; y) with |y| = 1 are dependent.
We give examples of NIP theories.
Example 62. Any theory of a linear order is NIP. TODO. 4
Example 63. Any theory of a tree is NIP. TODO. 4
Example 64. The theory Th(Qp ) of the p-adic numbers is NIP (either in the
language of fields, or in the language of valued fields). TODO. 4
Theorem 9.1.6. Every unstable ω-categorical NIP theory interprets (Q; <).
Proof. TODO.
9.2. The Lemma of Sauer-Shelah

Let F = {S1 , S2 , . . . } be a family of sets. Then a set T is shattered by F if for
every subset T 0 ⊆ T there exists an Si ∈ F such that T 0 = Si ∩ T . The VC dimension
of F is the largest cardinality of a set shattered by F.
Theorem 9.2.1 (Sauer-Shelah). Let F be a set of subsets of {1, . . . , n}. If the
VC dimension of F is at most d, then |F| ∈ O(nd ).
We deduce this theorem from a slightly more general theorem which has a very
elegant short proof.
Lemma 9.2.2 (Pajor). Let F be a finite family of sets. Then there are at least
|F| many sets that are shattered by F.
Proof. By induction on |F|. For the induction base, note that every family
containing only one set shatters the empty set.
For the inductive step, assume the lemma is true for all families of size less than
|F| and let F be a family of two or more sets. Let x be an element that belongs
to some but not all of the sets in F. Split F into two subfamilies: the sets that
contain x and the sets that do not contain x. By inductive assumption, these two
subfamilies shatter two collections of sets whose sizes add to at least |F|. None of
these shattered sets contain x since a set that contains x cannot be shattered by a
9.2. THE LEMMA OF SAUER-SHELAH 77
family in which all sets contain x or all sets do not contain x. Some of the shattered
sets may be shattered by both subfamilies. When a set S is shattered by only one
of the two subfamilies then it contributes both to the number of shattered sets of
the subfamily and to the number of shattered sets of F. When a set S is shattered
by both subfamilies, then both S and S ∪ {x} are shattered by F, so S contributes
twice to the number of shattered sets of the two subfamilies and to the number of
shattered sets of F. Therefore, the number of shattered sets of F is at least equal to
the number shattered by the two subfamilies of F, which is at least F.
Pd−1 n
Proof of Theorem 9.2.1. Only i=0 i subsets of {1, . . . , n} ∈ O(nd ) have
cardinality less than d, so if |F| 6∈ O(nd ) then by Lemma 9.2.2 there must be a set of
cardinality at least k that is shattered, which implies that the VC dimension is larger
than d.
Corollary 9.2.3. Let T be ω-categorical. Then T is NIP if and only if (orbit
growth description).
CHAPTER 10
Geometric Model Theory
One goal of geometric model theory (or geometric stability theory) is the classi-
fication of structures in terms of dimension-like quantities that can be axiomatised
using ideas from combinatorial geometry, such as matroids. Key guiding examples
include vector spaces and algebraically closed fields. This chapter is not yet finished.
10.1. Pregeometries
Pregeometries and geometries are the key axiomatic notion of geometric model
theory; geometries are also known as matroids in combinatorics.
Definition 10.1.1. Let X be a set. A closure operator on X is a function
C : 2X → 2X satisfying the following conditions:
(1) C is extensive (or reflexive), i.e., U ⊆ C(U ) for every U ⊆ X;
(2) C is increasing, i.e., U ⊆ V ⇒ C(U ) ⊆ C(V ) for all U, V ⊆ X;
(3) C is idempotent (or transitive), i.e., C(C(U )) = C(U ) for every U ⊆ X.
If C(U ) = U then U ⊆ X is called closed.
We sometimes write C(x1 , . . . , xn ) instead of C({x1 , . . . , xn }).
Proposition 10.1.2. Algebraic closure in a structure A (see Definition 6.1.1) is
a closure operator on A.
Proof. It is obvious that aclA is extensive and increasing. To show that it is
idempotent, suppose that b1 , . . . , bk ∈ aclA (C) and that A |= ψ(a, b1 , . . . , bk ) where
{y ∈ A | A |= ψ(y, b1 , . . . , bk )} has exactly m ∈ N elements. For i ∈ {1, . . . , k}, let
φi (x, c̄i ), for a tuple c̄i of elements from C, be a formula witnessing that bi ∈ aclA (C).
Let φm,ψ (x1 , . . . , xk ) be a formula that says that {y ∈ A | A |= ψ(y, x1 , . . . , xk )} has
at most m elements. Then
k
^
∃x1 , . . . , xk (ψ(y, x1 , . . . , xk ) ∧ φm,ψ (x1 , . . . , xk ) ∧ φi (xi , c̄i ))
i=1
is a formula with parameters in C that witnesses that c ∈ aclA (C), proving idempo-
tency of aclA .
Example 65. In algebraically closed fields K, we have a ∈ aclK (A) precisely if a
is algebraic (in the field-theoretic sense) over the subfield F of K generated by A (see
Appendix B.4.3). This follows from quantifier elimination of ACF (Theorem 7.7.1):
• if a is algebraic in the field-theoretic sense over F , then there exists a poly-
nomial p ∈ F [x] such that p(a) = 0. Since {b ∈ K | K |= p(b) = 0} is finite
and contains a. This shows that a is algebraic in the model-theoretic sense.
• if a ∈ aclK (A), there exists a first-order formula φ(x) such that {b ∈ K |
K |= φ(b)} is finite and contains a. Since ACF has quantifier elimination,
φ can be written as a finite disjunction of finite conjunctions of equalities
p(x) = 0 and inequalities p(x) 6= 0. So a must satisfy at least one of the
disjuncts. This disjunct cannot consist entirely of inequalities since in this
79
80 10. GEOMETRIC MODEL THEORY
case infinitely many elements of K would satisfy the disjunct, contradicting

our assumptions. Otherwise, if this disjunct contains an equality p(x) = 0,
then p(a) = 0 witnesses field-theoretic algebraicity of a. 4
Definition 10.1.3. A pregeometry (or matroid ) on X is a closure operator C on

X satisfying the following additional conditions:
(4) C is finitary, i.e., if A ⊆ X and a ∈ C(A), then there is a finite A0 ⊆ A such
that a ∈ C(A0 ).
(5) C has the exchange property, i.e., if A ⊆ X and a, b ∈ X then
a ∈ C(A ∪ {b}) implies a ∈ C(A) or b ∈ C(A ∪ {a}).
A geometry (or simple matroid ) on X is a pregeometry on X such that C(∅) = ∅ and

C({x}) = {x} for all x ∈ X.
Example 66. Algebraic closure in a structure B (see Definition 6.1.1) is clearly

always finitary. But it might not have the exchange property. Consider for example
the structure A := (A; E, P ) where
• A := Q × Q≥0 ;
• P := {(u, v) ∈ D | v = 0};
• E := {((u, v), (p, q)) ∈ D2 | v = 0, q 6= 0, u = p}.
It is easy to see that this structure is homogeneous and therefore ω-categorical. Note
that (0, 0) is in aclA ({(0, 1)}), but not in aclA (∅) since (0, 0) lies in the infinite orbit
P of Aut(A). But (0, 1) is not in aclA ({(0, 0)}) because there are automorphisms that
fix (0, 0) and map (0, 1) to (0, 2), say, showing that aclA fails exchange. 4
Example 67. Let V be a vector space over a field F . For A ⊆ V , let C(A)
be the linear span of A, i.e., the set of all elements that can be written as linear
combinations u1 a1 + · · · + un an for a1 , . . . , an ∈ A and u1 , . . . , un ∈ F . (The linear
span is the smallest linear subspace of V that contains A.) Then C is an example of
a pregeometry. It is not a geometry since 0 ∈ C(∅). 4
Example 68. Let V be a vector space over a field F . For A ⊆ V , let C(A) be the
affine hull of A, i.e., the set of all elements that can be written as u1 a1 + · · · + un an
where a1 , . . . , an ∈ A and u1 + · · · + un = 1 u1 , . . . , un ∈ F . The affine hull is the
smallest affine subspace of V that contains A. Then C is a geometry. 4
Example 69. Let V be a vector space over a field F , and let X = P (V ) be the
projective space over V , i.e, the set of equivalence classes of the equivalence relation
∼ defined on V \ {0} by
u ∼ v if and only if v = λu for some λ ∈ F \ {0}.

S
Recall that a subspace of X is a subset S ⊆ X such that S ∪ {0} is a linear subspace
of V . For S ⊆ X, define C(S) to be the smallest subspace of X that contains S. Then
C is a geometry. 4
Example 70. Let F be an algebraically closed field. Then aclF (recall that
model-theoretic and field-theoretic algebraic closure coincide, see Example 65) is a
pregeometry on F . To prove the exchange property, let A ⊆ F be a subset, and let K
be the subfield of F with domain aclF (A). Let a, b ∈ F be such that a ∈ aclF ({b}∪A).
So there is a non-zero polynomial p ∈ K[x, y] with p(a, b) = 0. This implies that
b ∈ aclF ({a} ∪ A). The operator aclF is not a geometry since 1 ∈ C(∅). 4
10.1. PREGEOMETRIES 81
10.1.1. Bases and dimension. Let C be a pregeometry on X.

Definition 10.1.4. A subset Y ⊆ X is called
• independent if a ∈/ C(Y \ {a}) for all a ∈ Y ;
• generating if X = C(Y );
• a basis if Y is independent and generating.
Lemma 10.1.5. Let C be a pregeometry on X with generating set E. Any inde-
pendent subset of E can be extended to a basis contained in E. In particular, every
pregeometry has a basis.
Proof. By Zorn’s lemma we can choose a maximal independent subset B of E.
We claim that E ⊆ C(B) and therefore C(B) = X. Suppose otherwise that there is
an x ∈ E \ C(B). By the maximality of B the set B ∪ {x} is not independent, i.e.,
there is a b ∈ B ∪ {x} such that b ∈ C(B ∪ {x} \ {b}). This is clearly impossible for
b = x since x ∈ / C(B). If b ∈ B, then b ∈/ C(B \ {b}) and the exchange property
implies that x ∈ C(B ∪ {b} \ {b}) = C(B), a contradiction.
Since a pregeometry satisfies the Steinitz exchange property all bases are of the
same cardinality.
Lemma 10.1.6. All bases of a pregeometry C have the same cardinality.
Proof. Let I be independent and G a generating subset of X. It suffices to
show that |I| ≤ |G|. Assume first that I is infinite. Then we extend I to a basis B
by Lemma 10.1.5. Choose for every g ∈ G a finite Bg ⊆ B with g ∈ C(BS g ). Since
the union of the Bg is a generating set, and B is independent, we have B = g∈G Bg .
This implies that G is infinite, and that |I| ≤ |B| ≤ |G| since each of the Bg is finite.
Now assume that I is finite. It suffices to show that for every a ∈ I \ G there
is some b ∈ G \ I such that I 0 = {b} ∪ I \ {a} is independent. To show this, first
observe that G cannot be contained in C(I \{a}) because a ∈ C(G), contradicting the
independence of I. Choose b ∈ G \ C(I \ {a}). Then the exchange property implies
that {b} ∪ I \ {a} is independent.
We can therefore define the dimension dim(X) of X as the size of a basis for X.
Example 71. The dimensions of the spaces from Examples 67, 68, 69, and 70
are as follows.
• The linear space of a κ-dimensional vector space has dimension κ.
• The affine space of a κ-dimensional vector space has dimension κ + 1.
• The projective space of a κ-dimensional vector space has dimension κ.
• The dimension of an algebraically closed field F with respect to aclF equals
the transcendence degree of F over the prime field. 4
Exercises.
(93) Let C be a pregeometry over X. Show that the following are equivalent for
all B ⊆ X:
• B is a basis for X.
• B is a maximal independent set.
• B is a minimal set that generates X.
10.1.2. Restriction and Relativisation. Any subset Y ⊆ X of a pregeometry
C on X gives rise to two new pregeometries, the restriction and the relativisation of
C with respect to Y . The restriction of C to Y is the pregeometry C Y on Y defined
by
C Y (S) := C(S) ∩ Y
for all S ⊆ Y . It is immediate that A ⊆ Y is independent with respect to C if and

only if it is independent with respect to C Y , and that it is a basis for Y with respect
to C if and only if it is a basis for Y with respect to C Y .
Definition 10.1.7. Let C be a pregeometry on X. The dimension of Y ⊆ X is
the dimension of the pregeometry C Y on Y .
The relativisation of C to Y is the pregeometry CY on X defined by
CY (S) := C(S ∪ Y )
for all S ⊆ X.
Proposition 10.1.8. Let C be a pregeometry on X and Y ⊆ X. Let A be a basis
of C Y and B a basis of CY . Then A ∪ B is a basis of C.
Proof. We have C Y (A) = Y and CY (B) = X. Therefore, X = CY (B) =
C(B ∪ Y ) = C(B ∪ (C(A) ∩ Y )) ⊆ C(B ∪ A) and hence A ∪ B is generating. Since B
is independent with respect to CY , we have b ∈
/ CY (B \ {b}) = C(A ∪ B \ {b}) for all
b ∈ B. Consider an a ∈ A. We have to show that a ∈ / C(A0 ∪ B) where A0 = A \ {a}.
0 0
As a ∈/ C(A ) we let B be a maximal subset of B with a ∈ / C(A0 ∪ B 0 ). If B 0 6= B
this would imply that a ∈ C(A ∪ B ∪ {b}) for any b ∈ B \ B 0 which would in turn
0 0
imply b ∈ C(A ∪ B 0 ), a contradiction.
10.1.3. The Associated Geometry. There is a standard way of obtaining a

geometry from a pregeometry C on X. First, if C(∅) 6= ∅, then replace X by
X 0 := X \ C(∅),
equipped with the restriction pregeometry C 0 . Then define an equivalence relation ∼
on X 0 by defining u ∼ v if C 0 (x) = C 0 (y), and define a pregeometry C 00 on X 00 := X 0 /∼
by
[
C 00 (S) := [b]∼ | b ∈ C 0 ( S)

for all S ⊆ X 00 . Then C 00 is indeed a geometry, called the geometry associated to C.

This generalises the way in which we obtain projective spaces from vector spaces.
10.1.4. Special Types of Pregeometries. A pregeometry C on X is called

S
• trivial if C(A) = a∈A C({a}).
• modular if for any two closed A, B ⊆ X
dim(A ∪ B) = dim(A) + dim(B) − dim(A ∩ B).
• locally modular if dim(A ∪ B) = dim(A) + dim(B) − dim(A ∩ B) holds for
all closed sets A, B ⊂ X with dim(A ∩ B) > 0.
• locally finite if closures of finite sets are finite.
Example 72. Structures A with no algebraicity (i.e., aclA (S) = S for all S ⊆ A)
are trivial. Less trivial, but still trivial in the sense of the definition above is the
structure that contains an equivalence relation with infinitely many classes of size
k ∈ N, which has algebraicity for k > 1. 4
Example 73. The linear span in vector spaces is a prototypical example of a
modular pregeometry. Affine spaces are not modular. For example, if A and B are
parallel lines in V then dim(A) + dim(B) + dim(A ∩ B) = 2 + 2 + 0 which is strictly
bigger than dim(A ∪ B) = 3. However, affine spaces are locally modular. The linear
span and affine hull in a vector space are locally finite if and only if the underlying
field is finite. 4
10.1. PREGEOMETRIES 83
Lemma 10.1.9. C is locally modular if and only if for all p ∈ X \ C(∅) the
relativised pregeometry C{p} is modular.
Proof. Let p ∈ X \ C(∅). Let A, B ⊆ X be closed with respect to C{p} . Then

A ∩ B contains p, and since p ∈
/ C(∅) the dimension of A ∩ B with respect to C is at
least one. The statement now follows from Proposition 10.1.8.
Example 74. Suppose that K |= ACF has transcendence degree at least five
over the prime field. Then aclK is not locally modular. By Lemma 10.1.9 it suffices
to show that for some p ∈ X \ aclK (∅) the relativised pregeometry C := (aclK ){p}
is not modular. Choose an independent set {p, a, b, c, d}, set x := a−cb−d , and let A :=
C(a, b) and B := C(x, a − b). Then C(A ∪ B) = C(a, b, x) has dimension 3, while
dim(A) = dim(B) = 2. To show that C is not modular it is sufficient to prove that
dim(A ∩ B) 6= 1.
Claim. A ∩ B = C(∅). Let z ∈ A ∩ B. We claim that there exists an α ∈ Aut(K)
that maps (a, b, c, d) to (c, d, a, b) and that fixes B pointwise. First note that any α
that maps (a, b, c, d) to (c, d, a, b) satisfies α(x) = x and
cb − cd − da + dc ab − da − ba + bc
α(a − bx) = c − dx = = = a − bx.
b−d b−d
Hence, α fixes {x, a−bx}, and can be extended to to an automorphism of K that fixes
B = C(x, a − bx) pointwise. In particular, α(z) = z. Furthermore, z ∈ A = C(a, b)
implies z = α(z) ∈ C(c, d). Consequently, z ∈ C(a, b) ∩ C(c, d) = C(∅). 4
The sets A and B are independent over C if for all sets A0 ⊆ A and B 0 ⊆ B, if
each A0 and B 0 is independent over C, then A0 and B 0 are disjoint and their union is
independent over C. Note that this relation is symmetric. The following is then easy
to see.
Lemma 10.1.10. For a pregeometry C over X the following are equivalent.

• C is modular.
• Any two closed A and B are independent over A ∩ B.
• For any two closed sets A and B we have dimB (A) = dimA∩B (A).
Exercises.
(94) Show that every pregeometry is submodular, i.e.,
dim(A ∪ B) + dim(A ∩ B) ≤ dim(A) + dim(B).
(95) The set of all closed subsets of a closure operator forms a lattice, where the
infimum ∧ is intersection and the supremum ∨ of X and Y is C(X ∪ Y ).
Show that a pregeometry is modular if and only if the lattice of closed sets
is modular, i.e., for all closed A, B, C
A ⊆ C ⇒ A ∨ (B ∧ C) = (A ∨ B) ∧ C.
(96) Show that every trivial pregeometry is modular.

(97) Show that a pregeometry C on X is modular if and only if for all c ∈ C(A∪B)
there are a ∈ C(A) and b ∈ C(B) such that c ∈ C({a, b});
(98) Prove that a geometry C on X is modular if and only if whenever a, b ∈ X
and A ⊆ X such that dim({a, b}) = 2 and dimA ({a, b}) ≤ 1 then
C({a, b} ∩ C(A)) \ C(∅) 6= ∅.

10.2. Minimal and Strongly Minimal Sets

Definition 10.2.1 (Minimal sets). Let B be a structure and let D ⊆ B be
definable in B B . Then
• D is minimal in B if the only subsets of D that are definable in B B are
finite or cofinite in D.
• If φ(x) is a formula that defines a minimal set D over B B , then we also say
that φ is minimal in B.
• The structure B is called minimal if if B is minimal in B.
Exercises.
(99) Show that (Q; <) and the random graph are not minimal.
Example 75. Let E be an equivalence relation on a countable set S with infinitely

many two-element classes. Then (S; E) is minimal. 4
In fact, the structure from Example 75 satisfies a strengthened from of minimality,
defined next.
Definition 10.2.2 (Strongly minimal sets). Let B be a structure. Then D ⊆ B is
strongly minimal in B if it is minimal in every elementary extension of B. A theory T
is strongly minimal if for every model B of T the underlying set B is strongly minimal.
A structure B is strongly minimal if its first-order theory is strongly minimal.
The claims for the following examples follow easily from quantifier elimination.
Example 76. Let V be the countably infinite vector space over the two-element
field F2 . Then (V ; +) is strongly minimal. 4
Example 77. Let K be an algebraically closed field. Then K is strongly minimal.
4
Example 78. The structure (N; <) is minimal, but not strongly minimal since
in every elementary extension B there is an element c ∈ B \ N such that (x < c)B
and (x > c)B are both infinite. 4
Theorem 10.2.3 (Baldwin, Lachlan). Let B be a structure and let X ⊂ B be
minimal in B. Then the restriction of aclB to X is a pregeometry.
Proof. Let A ⊆ X and let c ∈ acl(A∪{b})\acl(A) for some b, c ∈ X; we want to
show that b ∈ acl(A∪{c}). By assumption, there is a formula φ(x, y) with parameters
from A such that B |= φ(c, b) and |{x ∈ X | B |= φ(x, b)}| = n for some n ∈ N. Let
ψ(y) be a formula with parameters from A that says that |{x ∈ X | B |= φ(x, y)}| = n.
Then B |= ψ(b). If S := {y ∈ X | B |= φ(c, y) ∧ ψ(y)} is finite, then φ(c, y) ∧ ψ(y)
would witness that b ∈ acl(A ∪ {c}) and we are done. Otherwise, S is cofinite by the
minimality of B, so that |X \ S| = m for some m ∈ N. Let χ(x) be a formula with
parameters from A stating that
|{y ∈ X | B |= ¬(φ(x, y) ∧ ψ(y))}| = m.
If χ(x) defines a finite subset of X, then c ∈ acl(A), a contradiction. Hence, χ(x)
defines a cofinite subset of X. We may therefore choose n + 1 distinct elements
a1 , . . . , an+1 ∈ X such that B |= χ(ai ). This means that
Bi := {u ∈ X | B |= φ(ai , u) ∧ ψ(u)}
is cofinite for i ∈ {1, . . . , n + 1}. Hence i∈{1,...,n+1} Bi contains an element b0 . We
T
have at least n + 1 elements x ∈ X such that φ(x, b0 ) is satisfied, namely a1 , . . . , an+1 .

But this contradicts the fact that ψ(b0 ).
10.2. MINIMAL AND STRONGLY MINIMAL SETS 85
In the following, the notion of dimension in the context of minimal subsets

is always with respect to the pregeometry from algebraic closure (also see Defini-
tion 10.1.7).
Lemma 10.2.4. Let A and B be minimal structures with the same first-order
theory, let A0 ⊆ A and B 0 ⊆ B. If dim(A0 ) = dim(B 0 ), then any bijection between
a basis for A0 and a basis for B 0 can be extended to an isomorphism between the
substructure of A on aclA (A0 ) and the substructure of B on aclB (B 0 ).
Proof. Let f be a bijection between a basis for A0 and a basis for B 0 . We first
show that f preserves all first-order formulas φ(x1 , . . . , xn ) by induction on n. If n = 0
then the statement holds by the assumption that A and B have the same first-order
theory. Let a1 , . . . , an be elements from the basis for A0 such that φ(a1 , . . . , an ). Since
a1 , . . . , an are independent, we have that an ∈ / aclA ({a1 , . . . , an−1 }. Since A is mini-
mal, the formula φ(a1 , . . . , an−1 , xn ) is satisfied by cofinitely many xn in A. In other
words, there exists an ` ∈ N such that A |= ∃x=` xn . ¬φ(a1 , . . . , an−1 , xn ) (see Exer-
cise 8). By the inductive assumption, B |= ∃x=` xn . ¬φ(f (a1 ), . . . , f (an−1 ), xn ). Since
f (a1 ), . . . , f (an ) are independent, there is no formula satisfied by f (an+1 ) and finitely
other elements from the basis of B 0 . We conclude that B |= φ(f (a1 ), . . . , f (an )).
Next, we show how to extend f to an isomorphism between the substructure of A
on A00 := aclA (A0 ) and the substructure of B on B 00 := aclB (B 0 ). Suppose inductively
that we have already constructed a map g : E → B 00 , for E ⊆ A00 with A0 ⊆ E, that
preserves all first-order formulas, and let a ∈ A00 \ A0 . Since a ∈ aclA (A0 ) there is a
formula φ(x, y1 , . . . , yn ) and a1 , . . . , an ∈ E such that A |= ∃=` x.φ(x, a1 , . . . , an ) for
some ` ∈ N. Choose φ and a1 , . . . , an ∈ E such that ` is as small as possible. By
assumption, N |= ∃=` x.φ(x, g(a1 ), . . . , g(an )). In particular there exists b ∈ B 00 such
that N |= φ(b, g(a1 ), . . . , g(an )). Define the extension g 0 of g by setting g 0 (a) := b.
To show that g 0 preserves all first-order formulas, let ψ(x, y1 , . . . , ym ) be a for-
mula and d1 , . . . , dm ∈ E. Suppose that A |= φ(a, d1 , . . . , dm ). By the minimal-
ity of `, we have that A |= ∀x(φ(x, a1 , . . . , an ) ⇒ ψ(x, d1 , . . . , dm )). Since B |=
φ(b, g(a1 ), . . . , g(an )), this implies that B |= ψ(b, g(a1 ), . . . , g(an )).
By transfinite induction, we may therefore obtain an embedding g 0 : A00 → B 00
which extends f . It remains to show that g 0 is surjective. Let c ∈ B 00 = aclB (B 0 ).
Then there is a formula φ(x, c1 , . . . , cn ) for some elements c1 , . . . , cn from the basis of
B 0 witnessing that c ∈ aclB (B 0 ). Let ` ∈ N be such that B |= ∃=` x.φ(x, c1 , . . . , cn ).
For each ci let ai :− f −1 (ci ). Since f preserves all first-order formulas we obtain that
A |= ∃=` x.φ(x, a1 , . . . , an ). The ` distinct elements of B 00 that satisfy φ(x, c1 , . . . , cn )
are mapped by g 0 to the ` distinct elements of A00 that satisfy φ(x, c1 , . . . , cn ). In
particular, there exists a ∈ A00 such that g 0 (a) = c. This proves that g 0 is an isomor-
phism.
Corollary 10.2.5. Let A be a countable minimal structure without algebraicity.
Then A is preserved by all permutations, i.e., A is bi-definable with (N; =).
Exercises.
(100) Show that an ω-categorical structure is strongly minimal if and only if it is
minimal. How about the same statement for ℵ1 -categorical structures?
(101) In which of the following structures B is the algebraic closure operator a
pregeometry on B?
• An equivalence relation on a countably infinite set with infinitely many
classes of size two
• The random graph
• (Z; +)
10.3. Uncountably Categorical Structures

We have already seen that ACFp , for p = 0 or p prime, is κ-categorical for all
uncountable κ. Another example are equivalence relations on a countably infinite
set with infinitely many classes of size two. The following example presents some
structures that are ω-categorical, but not κ-categorical for uncountable κ.
Example 79. Let (D; A) be a structure where D is countably infinite and A is
such that both A and D\A are infinite. Then T := Th(D; A) has models of cardinality
κ > ω where |A| = κ and |D \ A| = ω, and models where |A| = ω and |D \ A| = κ,
and these are not isomorphic, so T is not κ-categorical.
Similarly, one can show that Th(D; E), where E is an equivalence relation with
two infinite classes, is not κ-categorical for κ > ω. 4
Lemma 10.3.1. Let τ be a countable signature and let T be a complete and strongly
minimal τ -theory. Then T is κ-categorical for every uncountably cardinal κ.
Proof. By Lemma 10.2.4, it is enough to show that any two models of T of
cardinality κ have the same dimension. Let A be a model of T of cardinality κ > ω
and let B ⊆ A be a basis of A. Since τ is countable, there are only countably many
τ -formulas. In particular, there are only countably many τ -formulas over parameters
from B that are satisfied by finitely many elements, and it follows that | aclA (B)| ≤
|B|. Since aclA (B) = A we conclude that |B| = |A|.
We mention a few highlights about κ-categorical structures. We have already
seen in Theorem 8.3.2 that κ-categoricity implies ω-stability.
Theorem 10.3.2 (Morley). Let κ be an uncountable cardinal. Then T is κ-
categorical if and only if it is ℵ1 -categorical.
An important step in the proof of Theorem 10.3.2 is the following partial converse
of Lemma 10.3.1.
Lemma 10.3.3. Every uncountably categorical structure contains a strongly min-
imal set.
Definition 10.3.4. Let T be a τ -theory. Two structures A, B |= T are called a
Vaughtian pair if there exists a (τ ∪ A)-formula φ(x) such that
• B is a proper elementary extension of A;
• φA is infinite;
• φA = φB .
Example 80. The theory of the structure (D; E) from Example 79 has the Vaugh-
tian pair (D; A), (D ∪ E; A) where |E| = ℵ1 . 4
Theorem 10.3.5 (Baldwin-Lachlan). Let κ be an uncountable cardinal. A count-
able theory T is κ-categorical if and only if T is ω-stable and has no Vaughtian pairs.
Exercises.
(102) (*) Show that the theory of the random graph is not ℵ1 -categorical.
APPENDIX A
Set Theory
In this appendix chapter we freely follow Appendix A of the book of Tent and
Ziegler [29]. See [20] for a more detailed introduction to set theory.
A.1. Sets and Classes

We often speak of the class of all τ -structures or the class of all ordinals, since
we know that these things cannot be sets (as we will see in the next section, if the
class of all ordinals were a set, we could derive a contradiction). Using first-order
logic, we therefore want to give an axiomatic treatment of set theory that allows for
the distinction between sets and (proper) classes. For this, we may work in Bernays-
Gödel set theory (BG) which is formulated in a two-sorted language. One type of
objects are sets and the other type of objects are classes; sets can be elements of sets,
and sets can be elements of classes, but classes can’t be elements of sets or classes.
Here are the axioms of BG:
(1) (a) Extensionality. Sets containing the same elements are equal:

∀x, y ∈ Sets ∀z(z ∈ x ⇔ z ∈ y) ⇒ x = y
(b) Empty set. There exists an empty set (denoted by ∅):
∃x ∈ Sets ∀y (¬y ∈ x)
(c) Pairing. For all sets a and b there is a set (denoted by {a, b}) which
has exactly the elements a and b:
∀a, b ∈ Sets ∃c ∈ Sets ∀x (x ∈ c ⇔ x = a ∨ x = b)
S
(d) Union. For every set a there is a set (denoted by a) that contains
precisely the elements of the elements of a:
∀a ∈ Sets ∃b ∈ Sets ∀x (x ∈ b ⇔ ∃y ∈ a. x ∈ y)
(e) Power set. For every set a there is a set (denoted by P(a)) that consists
of all subsets of a:
∀a ∈ Sets ∃b ∈ Sets ∀x (x ∈ b ⇔ x ⊆ a)
(f) Infinity. There is an infinite set. One way to express this is to assert
the existence of a set which contains the empty set and is closed under
the successor operation x 7→ x ∪ {x}:

∃a ∈ Sets ∅ ∈ a ∧ ∀x (x ∈ a ⇒ x ∪ {x} ∈ a
(2) (a) Class extensionality: Classes containing the same elements are equal.
(b) Comprehension: If φ(x, y1 , . . . , ym , z1 , . . . , zn ) is a first-order formula in
which only set-variables are quantified, and if b1 , . . . , bm are sets, and
c1 , . . . , cn are classes, then there exists a class, denoted by
{x ∈ Sets | φ(x, b1 , . . . , bm , c1 , . . . , cn }
87
88 A. SET THEORY
containing precisely those sets x that satisfy φ(x, b1 , . . . , bm , c1 , . . . , cn ).

This is in fact an infinite family of axioms (for every first-order formula
φ we have one axiom).
(c) Replacement: If a class c is a function, i.e., if for every set a there is a
unique set b =: c(a) such that (a, b) := {{a}, {a, b}} belongs to c, then
for every set d the image {c(z) | z ∈ d} is a set.
(3) Regularity (or Axiom of Foundation):

∀x ∈ Sets x 6= ∅ ⇒ ∃y ∈ x (y ∩ x = ∅)
Consequently: no set can be an element of itself. Moreover, there is no
infinite sequence (an ) such that ai+1 is an element of ai for all i. Most
parts of mathematics can be practised without this axiom, but assuming
this axiom simplifies some proofs of fundamental properties of ordinals.
(4) For BGC we add Global Choice: There is a function c such that c(a) ∈ a for
every nonempty set a.
We mention that BG is a conservative extension of ZF: any set-theoretical statement
provable in BG (BGC) is also provable in ZF (ZFC). (The substructure of a model of
BG induced by sets is a model of ZF. Conversely, a model M of ZF becomes a model
of BG by taking the first-order definable subsets of M as classes.)
A.2. Zorn’s Lemma

The axiom of choice implies Zorn’s lemma, which will be used several times in
this text.
Theorem A.2.1 (Zorn’s Lemma). Let (P ; ≤) be a partially ordered set with the
property that every chain in P has an upper bound in P . Then P contains at least
one maximal element.
A.3. Ordinals
A well-ordering of a set A is a linear order < on A such that any non-empty subset
of A contains a smallest element with respect to <. The usual ordering of the set N
is an example of a well-ordering, and the usual ordering of Z and the usual ordering
of the non-negative rational numbers are non-examples. For a ∈ A, the predecessors
of a are the elements b ∈ A with b < a.
Definition A.3.1. An ordinal is a well-ordered set in which every element equals
its set of predecessors.
The well-ordering of an ordinal α is given by the relation ∈: if β, γ ∈ α, then
β < α ⇔ β is a predecessor of α
⇔β∈α
Hence, an ordinal is uniquely given by the set of its elements. Suppose that α is an
ordinal, and γ ∈ β ∈ α. Then γ is a predecessor of β, and hence must be in α. That
is, ordinals α are transitive: if β ∈ α then β ⊆ α.
An element β of an ordinal α is again an ordinal number: Since β ⊆ α, we obtain
a well-ordering of β by restricting the well-ordering of α to β. Moreover, if γ, γ 0 ∈ β,
then γ < γ 0 iff γ ∈ γ 0 so β is indeed an ordinal.
Proposition A.3.2. Every structure (A; <) where A is well-ordered by < is iso-
morphic to (α; ∈) where α is an ordinal.
A.4. CARDINALS 89
Proof. Define f : A → α inductively by f (y) := {f (z) | z < y}. The image of

f is an ordinal α such that (α; ∈) is isomorphic to (A; <) via f . (Note that f is the
only isomorphism between (A; <) and (α; ∈).)
We denote the class of all ordinals by On.
Proposition A.3.3. On is a proper class, i.e., a class which is not a set, and it
is well-ordered by ∈.
Proof. Let α and β be different ordinals. We have to show that either α ∈ β
or β ∈ α. If not, then x = α ∩ β would be a proper initial segment of α and β, and
therefore itself an element of α and β, so x ∈ x, a contradiction.
The class of all ordinals is not a set (the Burali-Forti Paradox): if it were a set,
then it were an ordinal itself, and thus a member of itself, contradicting the Axiom
of Foundation.
So every ordinal α is a set having as elements precisely the smaller ordinals:
α = {β ∈ On | β < α} .
In the following, for ordinals α, β we write α < β rather than α ∈ β.
For any ordinal α its successor is defined as α + 1 := α ∪ {α}: it is the smallest
ordinal greater than α. Starting from the smallest ordinal 0 = ∅, its successor is
1 = {0}, then 2 = {0, 1}, and so on, yielding the natural numbers N. When we view
N as an ordinal, we denote it by ω. The next ordinal is ω + 1 = {0, 1, . . . , ω}, etc.
By definition, a successor ordinal β contains a maximal element α (so β = α + 1).
Ordinals greater than 0 which are not successor ordinals S are called limit ordinals.
These are precisely the ordinals γ that can be written as β<γ β. Any ordinal can be
written uniquely as
λ +1 + · · · + 1
| {z }
n times
where λ is a limit ordinal.
Theorem A.3.4 (Well-ordering theorem). Every set has a well-ordering.
Proof. Let A be a set. Fix a set B which does not belong to A and define a
function f from the class of all ordinals to A ∪ {B} as follows:
• if A \ {f (β) | β < α} =
6 ∅ then set f (α) to be an element from this set (here
we use the Axiom of Choice).
• Otherwise, f (α) := B.
Then γ := {α | f (α) 6= B} is an ordinal and f defines a bijection between γ and
A.
In fact, the well-ordering theorem is equivalent to the Axiom of Choice. Note
that the ordinal γ in the construction of the well-ordering of A is not unique, unless
A is finite.
A.4. Cardinals
Two sets A and B have the same cardinality (|A| = |B|) if there exists a bijection
between them. By the well-ordering theorem, every set has the same cardinality as
some ordinal. We call the smallest such ordinal the cardinality |A| of A. Ordinals
occurring in this way are called cardinals. An ordinal α is a cardinal if and only if all
smaller ordinals do not have the same cardinality.
Notes.
• All natural numbers and ω are cardinals.
90 A. SET THEORY
• ω + 1 is the smallest ordinal that is not a cardinal.

• The cardinality of a finite set is a natural number.
• A set of cardinality ω is called countably infinite.
Sums, products, and powers of cardinals are defined as the cardinality of disjoint
unions, Cartesian powers, and sets of functions:
|x| + |y| := |x ] y| where x ∩ y = ∅
|x| · |y| := |x × y|
|x||y| := |xy |
and likewise for infinite sums and products:
X ]
|x| := x
x∈I x∈I
Y Y
|x| := x.
x∈I x∈I
Note that
(κλ )µ = κλ·µ .
Cantor’s well-known diagonalization argument shows that
2κ > κ .
In particular, there is no largest cardinal. Cantor’s result also follows from König’s
theorem below for κi := 1 and λi := 2 for all i ∈ I := ω.
Theorem A.4.1 (König’s theorem). Let (κi )i∈I and (λi )i∈I be sequences of car-
dinals. If κi < λi for all i ∈ I, then
X Y
κi < λi .
i∈I i∈I
P Q
Proof. We first show that i∈I κi ≤ i∈I λi . Choose pairwise disjoint sets
(Ai )i∈I and (Bi )i∈I such thatS|Ai | = κi , Q|Bi | = λi , and Ai ⊂ Bi for all i ∈ I. We
will construct an injection f : i∈I Ai → i∈I Bi .SChoose di ∈ Bi \ Ai for each i ∈ I
(here we use the Axiom of Choice). For x ∈ A := i∈I Ai , define
(
x if x ∈ Ai
f (x) := (ai )i∈I where ai :=
di otherwise.
To show the injectivity of f , let x, y ∈ A be distinct. Let i ∈ I be such that x ∈ Ai .
If y ∈ Ai then f (x)i = x 6= y = f (y)i . If y ∈ / Ai then f (x)i = x 6= di = f (y)i since
x ∈ Ai but di ∈ Bi \ Ai . So in bothP Q6= f (y).
cases, f (x)
Suppose for contradiction that i∈I κi = i∈I λi . Then we can find sets (Xi )i∈I
with |Xi | = κi such that
Y [
B := Bi = Xi .
i∈I i∈I
For each i ∈ I, define
Yi := {ai | a ∈ Xi } .
For every i ∈ I there exists bi ∈ Bi \ Yi because |Yi | ≤ |Xi | = κi < λi = |Bi |. Now
define Y
b := (bi )i∈I ∈ Bi .
i∈I
Let j ∈ I. Then bj ∈
/SYj by the choice of bj , and hence b ∈
/ Xj by the definition of Yj .
This shows that b ∈
/ i∈I Xi , a contradiction.
A.4. CARDINALS 91
We write κ+ for the smallest cardinal greater than κ, the successor cardinal of κ.
Positive cardinals which are not successor cardinals are called limit cardinals. There
is an isomorphism between the class of ordinals and the class of all infinite cardinals,
which is denoted by
α 7→ ℵα
and can be defined inductively by

ω
 if α = 0
ℵα := ℵ+ β if α = β + 1

S
ℵ
β<α β if α is a limit ordinal.
S
Note that if (κi )i∈I is a family of cardinals, then κ := i∈I κi is again a cardinal:
S
• if there is an i ∈ I such that κj ≤ κi for all j ∈ I, then i∈I κi = κi and
the statement is true;
• otherwise, for every i ∈ I there is a j ∈ I with κi < κj . For each ordinal α
with α < κ we have that α ∈ κ and hence α ∈ κi for some i ∈ I. By the
above, there is a j ∈ I such that |α| ≤ κi < κj ≤ |κ|. Thus, every ordinal
smaller than κ has smaller cardinality than κ, and κ is a cardinal.
Theorem A.4.2. Let κ be an infinite cardinal. Then
(1) κ · κ = κ.
(2) κ + λ = max(κ, λ).
(3) κκ = 2κ .
Proof. For ordinals α, β, α0 , β 0 , define (α, β) < (α0 , β 0 ) iff
(max(α, β), α, β) <lex (max(α0 , β 0 ), α0 , β 0 )
where lex is the lexicographical ordering on triples of ordinals. Since this is a well-
ordering, there is a unique order-preserving bijection f between pairs of ordinals and
ordinals by Proposition A.3.2.
Claim. If κ is an infinite cardinal, then f maps κ × κ to κ, and hence κ · κ = κ.
The proof of the claim is by induction on κ. For α, β ∈ κ let Pα,β be the set of
predecessors of (α, β). Note that:
• Pα,β is contained in δ × δ with δ = max(α, β) + 1.
• Since κ is infinite and α, β < κ, the cardinality of δ is smaller than κ.
IA
• By inductive assumption |Pα,β | ≤ |δ × δ| = |δ| · |δ| = |δ| < κ.
Hence, f (α, β) < κ since f is an order isomorphism and thus f (α, β) ∈ κ.
Now (2) and (3) are simple consequences. Let µ := max(κ, λ).
µ≤κ+λ≤µ+µ≤2·µ≤µ·µ=µ
2κ ≤ κκ ≤ (2κ )κ = 2κ·κ = 2κ
Corollary A.4.3. For a non-empty set A, the set A<ω := An has cardi-
S
n∈ω
nality max(|A|, ℵ0 ).
Proof. Clearly, |A| ≤ |A<ω | and ℵ0 ≤ |A<ω |. On the other hand,
X
|A<ω | = |A|n ≤ (sup |A|n ) · ℵ0 = max(|A|, ℵ0 ).
n∈ω
n∈ω
The final equality holds because clearly supn∈ω |A|n = 1 if |A| = 1, supn∈ω |A|n = ℵ0
if 2 ≤ |A| ≤ ℵ0 , and |A| if |A| ≥ ℵ0 by Theorem A.4.2 (1).
92 A. SET THEORY
The Continuum Hypothesis (CH) states that ℵ1 = 2ℵ0 , that is: there is no cardinal
lying strictly between ω and the cardinality |R| of the continuum. The Generalised
Continuum Hypothesis (GCH) states that κ+ = 2κ for all infinite cardinals κ. As
with CH, the GCH is known to be independent of ZFC, that is, there are models of
ZFC where GCH is true , and models of ZFC where GCH is false (assuming that ZFC
is consistent; see [20]).
Let A be a set that is linearly ordered by <. A subset B ⊆ A is called cofinal if for
every a ∈ A there is some b ∈ B with a ≤ b. Any linear order contains a well-ordered
cofinal subset.
Definition A.4.4. The cofinality cf(A) of A is the smallest ordinal that is order-
isomorphic to a well-ordered cofinal subset of A.
Examples:
• If A has a greatest element, then the cofinality is one since the set consisting
only of the greatest element is cofinal and must be contained in any other
cofinal subset of A.
• A subset S of N is cofinal if and only if S is infinite, and thus cf(ω) = ω.
• cf(2ω ) > ω; see Exercise 92.
Lemma A.4.5. For linearly ordered A we have cf(cf(A)) = cf(A).
Proof. If {aα | α < β} is cofinal in A and {α(ν) | ν < µ} is cofinal in β, then
{aα(ν) | ν < µ} is cofinal in A.
Lemma A.4.6. For linearly ordered A we have that cf(A) is a cardinal.
Proof. Suppose for contradiction that cf(A) is not a cardinal. Choose a surjec-
tive map f from | cf(A)| to cf(A). This maps provides a cofinal sequence in cf(A) of
length | cf(A)|, and therefore cf(cf(A)) ≤ | cf(A)| < cf(A). This is in contradiction to
cf(cf(X)) = cf(X) from Lemma A.4.5.
A cardinal κ is regular if cf(κ) = κ, and singular otherwise. As we have seen
above, ℵ0 is an example of a regular cardinal. Lemma A.4.6 and Lemma A.4.5 show
that cf(A) is a regular cardinal. Assuming the axiom of choice, we also have the
following.
Proposition A.4.7. Successor cardinals ℵα+1 are regular.
Proof. Suppose for contradiction that cf(ℵα+1 ) ≤ ℵα . Then ℵα+1 would be
the union of at most ℵα sets of cardinality at most ℵα , contradicting item (1) in
Theorem A.4.2.
Exercises.
(91) Show that ℵω is singular.
(92) Show that cf(2ω ) >Pω.
Hint: write 2ω as ω
ν<µ κν for µ := cf(2 ), and apply König’s theorem
(Theorem A.4.1) with (κν )ν<µ and (λν )ν<µ where λν := 2ω for all ν < µ.
APPENDIX B
Rings and Fields
In this section we collect some standard material from classical algebra that is
typically taught in the second year; we freely follow Lang [21] and Dummit and
Foote [10]. My impression is that Dummit and Foote is the more elementary and
detailed presentation; on the other hand, the proof of Steinitz’ theorem (Corol-
lary B.4.12) is omitted there, but can be found in Lang’s book, which is also the
reference of Tent and Ziegler.
A commutative ring is a structure R over the signature τRing = {+, −, 0, 1, ·}
such that the reduct (R; +, −, 0) of R is an Abelian group and such that the following
axioms hold:
∀x, y, z. (xy)z = x(yz)
∀x. 1x = x
∀x, y. xy = yx
∀x, y, z. x(y + z) = xy + xz
Let R be a ring. A nonzero element a of R is called a zero divisor if there is an
b ∈ R \ {0} such that ab = 0. An element u ∈ R is called a unit if there is some v ∈ R
such that uv = vu = 1. In this terminology, a field is a commutative ring such that
0 6= 1 in which every nonzero element is a unit. Observe that a zero divisor cannot
be a unit.
Lemma B.0.1. Let R be a ring and a, b, c ∈ R be such that a is not a zero divisor.
If ab = ac then a = 0 or b = c.
Proof. If ab = ac then a(b − c) = 0 so a = 0 or b − c = 0.
B.1. Integral Domains

A commutative ring R with more than one element and without nonzero zero
divisors is called an integral domain. An equivalent formulation is that R is a ring
for which the set of nonzero elements is a commutative monoid with respect to multi-
plication. Examples of integral domains are fields, the ring of all integers, the p-adic
integers, and rings of polynomials over integral domains.
A fundamental property of integral domains is that every subring of a field is an
integral domain, and that, conversely, given any integral domain, one may construct a
field that contains it as a subring, e.g., the field of fractions, introduced below. Hence,
a ring is an integral domain if and only if it is isomorphic to a subring of a field.
Let R be an integral domain. Let ∼ be the equivalence relation defined on A × S
by setting (a, s) ∼ (a0 , s0 ) if and only if s0 a = sa0 . We write as for the equivalence
class containing (a, s).
Definition B.1.1. The field of fractions of an integral domain R is the field Q
with the domain Q := (R × R \ {0})/∼ where multiplication is defined by the rule
a a0 aa0
:= .
s s0 ss0
93
94 B. RINGS AND FIELDS
It is easy to verify that this is well defined, and that multiplication has a unit, namely
1
1 , and is associative. Addition is defined by the rule
a a0 s0 a + sa0
+ 0 := .
s s ss0
It is easy to verify that this is well defined, with neutral element 01 , and that addition
and multiplication yield a ring structure.
Sometimes, the term quotient field is used, but we try to avoid this to avoid
confusion with the quotients that will be introduced in the next section.
Note that the map fQ : R → Q given by f (a) := a1 is an (injective) ring homo-
morphism from R to Q. Every element of fQ (R \ {0}) is invertible in Q: the inverse
of 1s is 1s . We often identify the elements of R with elements of Q along fQ , and hence
view R as a subring of Q.
Proposition B.1.2. Let R and B be rings and h : R → B be a homomorphism
such that all elements of h(R) are invertible in B. Let Q be the field of fractions of
R. Then there exists a unique homomorphism u from Q to B such that u ◦ fQ = h.
Example 81. If R is an integral domain, then the ring R[x1 , . . . , xn ] of polyno-
mials over R is also an integral domain: this comes from the fact that if n and m are
the degrees of the polynomials f ∈ R[x] and g ∈ R[x], then the degree of f g ∈ R[x]
is n + m, so there are no nonzero zero divisors. If K is the field of fractions of R,
then the field of fractions of R[x1 , . . . , xn ] is denoted by K(x1 , . . . , xn ); its elements
are called rational functions. A rational function can be written as fg(x̄) (x̄)
where f and
g are polynomials. 4
B.2. Ideals and Quotient Rings

Let R be a commutative ring. A subset I of R is an ideal of R if
• rI = Ir = I for all r ∈ R, and
• (I; +) is a subgroup of (R; +).
(i.e., the ideals of R are the domains of submodules of R). For every ring homo-
morphism h : R → S, the set {r ∈ R | h(r) = 0}, also called the kernel of the ring
homomorphism h, is an ideal. Conversely, any ideal I of R gives rise to a congruence
of R, namely C = {(r1 , r2 ) | r1 − r2 ∈ I}. We also write R/I for the structure
R/C, which will again be a ring since the ring axioms are universal. Note that each
congruence class A can be written as A = u + I for some u ∈ A. The homomorphism
h : R → R/I that sends r to r + I is called the canonical homomorphism from R to
R/I.
Definition B.2.1 (principal ideal). For any a ∈ R, the set Ra = aR is an ideal
of R, denoted by (a), the principal ideal of R generated by a.
Examples of principal ideals are (0) = {0} (called the trivial or zero ideal ) and
(1) = R. Ideals different from R are called proper ideals.
Example 82. Let K be a field and K[x] be the polynomial ring. Then the ideal
(x) is the set of all f ∈ K[x] with zero constant coefficient. Let h be the canonical
homomorphism from K[x] to K[x]/(x) given by h(f (x)) := f (x) + (x). Then the
restriction of h to K is an isomorphism between K and K[x]/(x): every element of
K[x]/(x) can be written as u + (x) where u ∈ K, and h(u) = u + (x). 4
Proposition B.2.2. A commutative ring R is a field if and only if its only ideals
are {0} and R.
B.2. IDEALS AND QUOTIENT RINGS 95
Proof. R is a field if and only if every nonzero element is a unit. If R is a field

then every nonzero ideal I contains a unit u with inverse v. Then for every r ∈ R
r = r(vu) = (rv)u ∈ I
hence R = I. Conversely, if {0} and R are the only ideals of R, then let u ∈ R \ {0}.
By assumption (u) = R and so 1 ∈ (u). Thus, there is some v ∈ R such that 1 = uv,
so u is a unit.
P
Note that for any finite family {Ij }j∈A of ideals of R the sum j∈A Ij is an ideal
P
of R. If A ⊆ R, then a∈A (a) is the smallest ideal of R that contains A. It is called
the ideal generated by A. If A = {a1 , . . . , an } then we write (a1 , . . . , an ) for the ideal
generated by A.
Definition B.2.3. An integral domain R is called a principal ideal domain (PID)
if every ideal of R is principal.
See Figure B.1 for an overview of the relationship with other fundamental concepts
for rings and fields, along with examples that show that there is a chain of strict
inclusions of the shown concepts.
Example 83. Z is a principal ideal domain: for if I ⊆ Z is a non-trivial ideal,
then I = nZ = (n) for some n ∈ N. 4
Example 84. When K is a field, then the ring of polynomials K[x] is a principal
ideal domain. 4
Example 85. Z[x] is not a principal ideal domain: e.g. the ideal (2, x) is not
principal. Assume for contradiction that there exists a(x) ∈ Z[x] such that (2, x) =
(a(x)). Since 2 ∈ (a(x)) there must be some p(x) such that 2 = p(x)a(x). The
degree of p(x)a(x) equals the degree of p(x) plus the degree of a(x), hence both p(x)
and a(x) must be constant polynomials, i.e., integers. Since 2 is prime, a(x), p(x) ∈
{−2, −1, 1, 2}. If a(x) ∈ {−1, 1} then every polynomial would be a multiple of a(x)
contrary to (a(x)) 6= R. If a(x) ∈ {−2, 2} then x ∈ (a(x)) = (2) = (−2) and so
x = 2q(x) for some q ∈ Z[x], clearly impossible. 4
Example 86. Q[x, y] is not a principal ideal domain. This follows from Corol-
lary B.2.10 below, because if it was, then Q[x] would have to be a field, which it is
not. 4
B.2.1. Prime Ideals. An ideal I of R is called prime if I 6= R such that when-

ever xy ∈ I, then x ∈ I or y ∈ I.
Proposition B.2.4. An ideal I of R is prime if and only if R/I is an integral
domain.
Proof. An ideal I of R is prime if and only if I 6= R and if ab ∈ I then a ∈ I or
b ∈ I. This is the case if and only if R/I 6= {I} and if abI = I then aI = I or bI = I,
i.e., if and only if R/I is an integral domain.
Let R be a ring with a 1. Then the map h : Z → R given by
h(n) := 1 + · · · + 1
| {z }
n times
is a ring homomorphism, and its kernel is an ideal (n) = nZ, generated by n ∈ N. If
(n) is a prime ideal then n = 0 or n = p for some prime number p. In the first case, R
has a subring which is isomorphic to Z; in this case, we say that R has characteristic
0. If on the other hand n = p then we say that R has characteristic p, and R has a
Fields:
ℚ, ℂ, ℤp
Euclidean domains:
ℤ, ℚ[x], ℤ[i]
Principal ideal domains:

ℤ[(1/2+(-19)!"#/2]
Unique factorisation domains:
ℤ[x], ℚ[x,y]
Integral domains:
ℤ[2i], ℤ[(-5)!"#]
Commutative rings:
ℤxℤ, ℤ6
Rings:
ℚ#$#
Figure B.1. From rings to fields: algebra reminder.
subring isomorphic to Z/nZ. If K is a field, then K has characteristic 0 or p > 0. In

the first case, it contains an isomorphic copy of Q, and in the second case, it contains
an isomorphic copy of Fq . In either case, this subfield will be called the prime field
(contained in K). The prime field is the smallest subfield of K containing 1, and it
has no non-trivial automorphisms.
Definition B.2.5. Let r be a nonzero non-unit element of an integral domain R.
Then r is called
• irreducible in R if whenever r = ab for a, b ∈ R, then at least one of a or b
must be a unit in R;
• prime in R if the ideal (r) is prime.
Proposition B.2.6. Every prime element in an integral domain is irreducible.
Proof. Suppose that (p) is a nonzero prime ideal and p = ab. Then ab = p ∈ (p),
so by definition of prime ideals one of a or b, say a, is in (p). Thus a = pr for some r.
This implies that p = ab = prb, so rb = 1 (Lemma B.0.1) and b is a unit. This shows
that p is irreducible.
In Proposition B.2.11 we will see a converse of Proposition B.2.6 in principal ideal
domains.
B.2.2. Maximal Ideals. An ideal I of R is called maximal if I 6= R and if there
is no ideal J 6= R strictly containing I.
Proposition B.2.7. An ideal I of a commutative ring R is maximal if and only
if R/I is a field.
Proof. By Proposition B.2.2, R/I is a field if and onlySif {0} is the only proper
ideal of R/I. Note that J is an ideal of R/I if and only if J is an ideal of R that
contains I; so I is maximal if and only if {0} is the only proper ideal of R/I.
Corollary B.2.8. Every maximal ideal of a commutative ring R is prime.
B.2. IDEALS AND QUOTIENT RINGS 97
Proof. If I is a maximal ideal, then R/I is a field by Proposition B.2.7, which

is in particular an integral domain. Proposition B.2.4 the implies that I is prime.
In principal ideal domains, we have a converse for Corollary B.2.8.
Proposition B.2.9. Every nonzero prime ideal in a principal ideal domain R is
maximal.
Proof. Let (p) be a nonzero prime ideal in R and let I = (m) be any ideal
containing (p). We must show that I = (p) or I = R. Since p ∈ I there is an r ∈ R
such that p = rm. Since (p) is prime and p = rm ∈ (p) either r ∈ (p) or m ∈ (p). If
m ∈ (p) then (p) = (m) = I. If r ∈ (p) write r = sp. In this case p = rm = psm.
Since R is in particular an integral domain, p 6= 0 is not a zero divisor, and hence
sm = 1 (Lemma B.0.1). Hence m is a unit, so I = R.
Corollary B.2.10. Let R be a commutative ring such that R[x] is a principal
ideal domain. Then R must be a field.
Proof. Since R[x] is by assumption in particular an integral domain, so is its sub-
ring R. By Proposition B.2.4, the non-zero ideal (x) in R[x] is prime because R[x]/(x)
is isomorphic to the integral domain R (see Example 82). By Proposition B.2.9, (x)
is maximal, hence R[x]/(x) is a field by Proposition B.2.7.
Note that this shows in particular that Q[x, y] is not a principal ideal domain:
we have Q[x, y] = R[x] for R := Q[y], but R is not a field. Now comes the promised
converse of Proposition B.2.6 for principal ideal domains.
Proposition B.2.11. Let R be a principal ideal domain and let I = (p) be a
nonzero ideal. Then the following are equivalent.
(1) p is prime;
(2) p is irreducible;
(3) I is maximal;
(4) I is prime.
Proof. (1) ⇒ (2) follows from Proposition B.2.6.
For (2) ⇒ (3), let p be irreducible. If I is any ideal containing (p) then by
assumption I is principal, i.e., I = (m) for some m ∈ R. Since p ∈ (m) there exists
r ∈ R such that p = rm. But p is irreducible, so by definition either r or m is a unit.
If r is a unit then (p) = (m), and if m is a unit then (m) = R. It follows that I is
maximal.
(3) ⇒ (4) follows from Corollary B.2.8.
(4) ⇒ (1) holds by definition.
Exercises.
(93) A unique factorisation domain is an integral domain in which every non-
zero non-unit element can be written as a product of irreducible elements,
uniquely up to order and units. Show that every principal ideal domain is a
unique factorisation domain. See Figure B.1.
(94) Show that every principal ideal domain is a Dedekind domain, i.e., an integral
domain in which every nonzero proper ideal factors into a product of prime
ideals.
(95) Show that Q[x, y] is an example of a unique factorisation domain which is
not a Dedekind domain.
(96) Show that a Dedekind domain is a unique factorisation domain if and only
if it is a principal ideal domain.
B.3. Polynomial Rings

Let R be a commutative ring. We write R[x] for the ring of polynomials over R.
Let S be a subring of R. If p(x) ∈ S[x], we write pR : R → R for the corresponding
polynomial function. Note that p 7→ pR (0) is a ring homomorphism from S[x] to R.
Let b ∈ R. We write S[b] for the subring of R generated by S ∪ {b} (i.e., the
smallest substructure of R that contains S ∪ {b}); this is also called a ring adjunction.
Note that the elements of S[b] are precisely
{pR (b) | p ∈ S[x]}.
If the map p 7→ pR (b) is an isomorphism between S[x] and S[b] then we say that x is
transcendental over S.
We have already mentioned that for any field K, the ring of polynomials K[x] is
a principal ideal domain. A polynomial is called monic if the leading coefficient is 1.
Proposition B.3.1. Let K be a field and p ∈ K[x] of degree n ≥ 0. Then f has
at most n roots in K, and if a is a root of f in K, then x − a divides f (x).
Proof. By polynomial division and induction.
Proposition B.3.2. Let p ∈ K[x] be irreducible. Then the quotient K[x]/(p(x))

is a field.
Proof. Since p(x) is irreducible, the ideal (p(x)) is maximal by Proposition B.2.11,
so K[x]/(p(x)) is a field by Proposition B.2.7.
B.4. Field Extensions

A field extension E of a field F is a field that extends F in the sense of Sec-
tion 1.1.3. We may view F as a vector space over E, and write [E : F ] for the
dimension of the vector space, called the degree of the extension. We say that E
is a finite extension or an infinite extension of F according to whether the degree
is finite or not. If S ⊆ E then there exists a smallest subfield of E that contains
F and S, denoted by F (S), and called the field generated by S over F (this is not
the same as the notation F [S] from Section B.3, since F (S) must also contain the
multiplicative inverses for non-zero elements). When S = {a1 , . . . , an } is finite one
writes F (a1 , . . . , an ) instead of F ({a1 , . . . , an }). If S consists of a single element a,
then the extension F (a) of F is called simple and a is called a primitive element of
the extension. Note that the elements of F (a) are precisely the set of elements that
can be written as p(a)/q(a) where p, q ∈ F [x] (for q(a) 6= 0).
B.4.1. Algebraic Field Extensions. An element a of a field extension E of

F is called algebraic over F if a is the root of some non-zero polynomial f (x) with
coefficients from F .
Lemma B.4.1. Let E be a field extension of F . Then a ∈ E is algebraic over F
if and only if a is not transcendental over F (see Section B.3).
Proof. If a is not transcendental, then p 7→ pE (a) is not injective, i.e., there are
E E
distinct p1 , p2 ∈ F [x] such that p1 (a) = p2 (a). Hence, a is a zero of the non-zero
polynomial p1 − p2 , and algebraic over F .
Conversely, if a is a zero of the non-zero polynomial p ∈ F [x], then pE (a) =
0, but the map p 7→ pE (a) sends the 0 polynomial in F [x] to 0, too, so a is not
transzendental.
B.4. FIELD EXTENSIONS 99
Lemma B.4.2. If a is transcendental over F , then the isomorphism p 7→ pE (a)

between F [x] and F [a] can be extended to an isomorphism between the field of rational
functions F (x) (Example 81) and F (a).
Proof. If f, g ∈ F [x], and g is non-zero, then g(a) 6= 0, and we map f /g to
f E (a)/g E (a) ∈ F (a). This map is clearly a homomorphism, and it is injective:
suppose that f E (a)/g E (a) = (f 0 )E (a)/(g 0 )E (a) for f 0 , g 0 ∈ F [x], g 0 non-zero, then
f (a)g 0 (a) − f 0 (a)g(a) = 0. By Lemma B.4.1, a is not algebraic, so f g 0 − f 0 g must be
the zero-polynomial, which in turn means that f /g = f 0 /g 0 .
Lemma B.4.3. If E is a field extension of F and a ∈ E is algebraic over F , then

there exists an irreducible p ∈ F [x] such that p(a) = 0. Furthermore, if f ∈ F [x] and
f (a) = 0, then p divides f .
Proof. There exists a polynomial p ∈ F [x] such that p(a) = 0; choose p of
minimal degree. We claim that p is irreducible in F [x]. For if p = tq for t, q ∈ F [x]
then 0 = p(a) = t(a)q(a), and so a would be a root of at least one of t and q, say
of t. Hence, deg(t) = deg(p), since p is of minimal degree among those polynomials
with root a. Hence, q must be a constant polynomial, as required. To show that
p divides every f ∈ F [x] with root a, we write f = pq + r where q, r ∈ F [x] and
deg(r) < deg(p). Then 0 = f (a) = p(a)q(a) + r(a) = r(a), so a is a root of r. Hence
r = 0, and p divides f .
The proof of Lemma B.4.3 shows that there exists a unique monic polynomial in
F [x] that is irreducible and has a as a root, and which is called the minimal polynomial
of a over F .
Proposition B.4.4. Let E be a field extension of F . Then a ∈ E is algebraic
over F if and only if F (a) is a finite extension of F .
Proof. If a is algebraic then the degree of F (a) over F is the degree of the
minimal polynomial for a over F . Conversely, suppose that F (a) has degree n
over F . Then 1, a, a2 , . . . , an cannot be linearly independent over F . So there are
b0 , b1 , . . . , bn ∈ F , not all zero, such that b0 + b1 a + · · · + bn an = 0. Hence, a is the
root of the non-zero polynomial b0 + b1 x + · · · + bn xn .
A field extension E of F is algebraic if every element of E is algebraic over F .

Field extensions that are not algebraic are called transcendental.
Corollary B.4.5. Every finite field extension E of F is algebraic over F .
Proof. If a ∈ E, then F (a) is a subspace of the vector space E over F . Hence
[F (a) : F ] ≤ [E : F ] and so a is algebraic over F by Proposition B.4.4.
A literal converse of Lemma B.4.5 is false: for example the subfield of C consisting
of all algebraic numbers is an infinite algebraic extension of Q. The following lemma
prepares an exact characterisation of finite field extensions (Proposition B.4.7 below).
Lemma B.4.6. Let F be a subfield of K, and K a subfield of L. Then [L : F ] =
[L : K][K : F ].
Proof. Suppose first that [L : K] = m and [K : F ] = n are finite. Let a1 , . . . , am
be a basis for L over K and let b1 , . . . , bn be a basis for K over F . Then every c ∈ L
can be written as a linear combination
u1 a1 + · · · + um am
where u1 , . . . , um ∈ K. Each ui , for i ∈ {1, . . . , m} can be written as a linear combi-

nation
(3) ui = vi,1 b1 + · · · + vi,n bn
where the vi,j ∈ F . Substituting these expressions we obtain the linear combination
X X
c= vi,j ai bj
i∈{1,...,m} j∈{1,...,n}
so the elements ai bj span L as a vector space over F .

Suppose now that
X X
vi,j ai bj = 0
i∈{1,...,m} j∈{1,...,n}
for some coefficients vij ∈ F . Then defining the elements ui ∈ K by equation 3 above
this equation could be rewritten to u1 a1 + · · · + um am = 0, and since a1 , . . . , am form
a basis for L over K this implies that u1 = · · · = um = 0. So we have vi,1 b1 + · · · +
vi,n bn = 0, and since b1 , . . . , bn for a basis for K over F this implies that vi,j = 0 for
all i ≤ m and j ≤ n. Hence the ai bj are linearly independent over F , so they form a
basis for L over F and [L : F ] = mn as claimed.
If one of [L : K] or [K : F ] is infinite, then [L : F ] is infinite, too.
Proposition B.4.7. E is a finite extension of F if and only if E is generated by

a finite number of elements of E that are algebraic over F .
Proof. If [E : F ] = n ∈ N, let a1 , . . . , an be a basis for E as a vector space
over F . By Lemma B.4.6, [F (ai ) : F ] divides [E : F ] = n for i ∈ {1, . . . , n}, so
Proposition B.4.4 implies that each ai is algebraic over F .
For the converse, suppose that E = F (a1 , . . . , an ) where each of a1 , . . . , an is
algebraic over F . We have F (a1 , . . . , an ) = F (a1 ) . . . (an ), and by Proposition B.4.4,
and [F (a1 ) : F ] is finite by Proposition B.4.4. note that [F (a1 )(a2 ) : F (a1 )] is equal to
the degree of the minimal polynomial of a2 over F (a1 ), which is at most the degree of
the minimal polynomial p of a2 over F (the degrees are equal if p remains irreducible
over F (a1 )). By Lemma B.4.6 and induction it follows that [E : F ] is finite.
Lemma B.4.8. Let L be an algebraic extension of K, and K an algebraic extension

of F . Then L is an algebraic extension of F .
Proof. Let a ∈ L. Then a satisfies an equation bn an + · · · + b0 = 0 where
b1 , . . . , bn ∈ K, b0 6= 0. Consider the subfield F (a, b0 , b1 , . . . , bn ) of L. Since K is
algebraic over F , the elements b0 , . . . , bn are algebraic over F , so [F (b0 , . . . , bn ) :
F ] is finite by Proposition B.4.7. Since a satisfies the equation above, we have
that [F (a, b0 , . . . , bn ) : F (b0 , . . . , bn )] ≤ n, and by Lemma B.4.6 it follows that
[F (a, b0 , . . . , bn ) : F )] is finite, which proves that a is algebraic over F .
B.4.2. Algebraically Closed Field Extensions. A field K is called alge-

braically closed if it has no proper algebraic extension, i.e., it contains a root for
every non-constant polynomial with coefficients in K. The typical example of an
algebraically closed field is the field of complex numbers. Also note that no finite
field K is algebraically closed, because if K = {a1 , . . . , an } then the polynomial
(x − a0 ) · · · (x − an ) + 1 has no root in K.
Proposition B.4.9. Let K be a field and p ∈ K[x] of degree at least one. Then
there exists an extension E of K in which p has a root.
L
g
B
E
k(K) k
e(a) K
Figure B.2. Illustration for Theorem B.4.11.
Proof. We may assume that p is irreducible. Then the quotient ring E :=

K[x]/(p(x)) is a field by Proposition B.3.2. Let h : K[x] → K[x]/(p(x)) be the canon-
ical homomorphism given by f (x) 7→ f (x) + (p(x)), and let e be the restriction of h
to K. Then e : K → E is not constant 0 since it maps 1 ∈ K to 1 ∈ F . Since {0}
is the only proper ideal of a field (Proposition B.2.2), e must be an embedding. We
identify K with its image in E via e and view K as a subfield of E. Then
p(e(x)) = e(p(x)) (since e is a homomorphism)
= p(x) + (p(x)) (an element of E)
E
=0
so E does indeed contain a root of the polynomial p(x). Hence, E is a field extension
of K with the desired property.
Recall from Corollary 7.4.3 that every field has an algebraically closed field exten-
sion. Clearly, algebraically closed field extensions of a field are not unique: they might
be of different cardinality. But we will prove that algebraically closed algebraic field
extensions are unique up to isomorphism (Corollary B.4.12). First we prove existence
of such field extensions.
Corollary B.4.10. Every field K has an algebraically closed algebraic field ex-
tension.
Proof. By Corollary 7.4.3 there exists an algebraically closed field extension E
of K. Let A be the union of all subfields of E that are algebraic over K; this is itself
a subfield of E, and certainly algebraic over K. If p ∈ A[x] has degree at least one,
then it has a root a in E, so a is algebraic over A. If a is algebraic over A, then a is
already algebraic over K by Lemma B.4.8, and therefore a ∈ A. We have shown that
A is algebraically closed.
B.4.3. Algebraic Closure.
Theorem B.4.11. Let K be a field, E an algebraic extension of K, and k : K → L
an embedding of K into an algebraically closed field L. Then there exists an extension
of k to g : E ,→ L. If E is algebraically closed and L is algebraic over k(K), then any
such extension of k is an isomorphism between E and L.
Proof. Let S be the set of all pairs (F , f ) where F is a subfield of E containing

K and f is an extension of k to an embedding of F into L. We define the following
partial order on S: if (F , f ), (F 0 , f 0 ) ∈ S define (F , f ) ≤ (F 0 , f 0 ) if F ⊆ F 0 and f

is the restriction of f 0 to F . Note that S 6= ∅ because it contains (K, k). Also S note
that every chain (F i , fi )i∈I of (S, ≤) has an upper bound in S: define F := i∈I Fi
and define f on F to be equal to fi on Fi for each i ∈ I. Using Zorn’s Lemma
(Theorem A.2.1), let (G, g) be a maximal element in S.
Then g is an extension of k, and we claim that G = E. Indeed, if a ∈ E \ G, then
G(a) is algebraic over G because E is algebraic over K. For any b ∈ G(a), let p ∈ G[x]
be the minimal polynomial of b over G. Note that p(x) = t(x, c1 , . . . , cn ) where t is a
term in the language of rings and c1 , . . . , cn are constant symbols for elements of G.
Since L is algebraically closed, there exists a b0 such that L |= (b0 , g(c1 ), . . . , g(cn )),
and we define g(b) := b0 . Then g has an extension to G(a): if b ∈ G(a), we can
write it in the form s(a, d1 , . . . , dm ) for some term s in the language of rings and
d1 , . . . , dm constant symbols for elements of G. Then the map s(a, d1 , . . . , dm ) 7→
s(b, g(d1 ), . . . , g(dm )) is well-defined (independent from the choice of s) and gives a
embedding of G(a) into L that extends g. This is contradicting the maximality of
(G, g), and shows that g is an embedding E ,→ L.
If E is algebraically closed and L is algebraic over g(K), then g(E) is algebraically
closed and L is algebraic over g(E), hence L = g(E).
Corollary B.4.12 (Steinitz). Let K be a field and let E and E 0 be algebraic
extensions of K. Assume that E and E 0 are algebraically closed. Then there exists
an isomorphism between E and E 0 that fixes the elements of K.
Proof. By Theorem B.4.11 the identity map on K can be extended to a map
from E to E 0 . Since E is algebraically closed and E is algebraic over h(K), by
the second part of Theorem B.4.11, the extension is an isomorphism between E and
E0.
Hence, algebraically closed algebraic extensions are uniquely determined up to
isomorphism, and will be called the algebraic closure of K.
Example 87. The subfield of C consisting of the complex numbers that are
algebraic over Q is countable, and the algebraic closure of Q. 4
B.4.4. Transcendence degree. The notion of transcendence degree is analo-
gous to the notion of dimension in linear algebra.
Definition B.4.13. Let E be a field extension of a field F .
• A subset S ⊆ E is called algebraically independent over F if for every p ∈
F [x1 , . . . , xn ] and a1 , . . . , an ∈ S, if p(a1 , . . . , an ) = 0 then p is the zero
polynomial.
• A transcendence base (or transcendence basis) of E over F is an algebraically
independent subset S ⊆ E such that E is algebraic over F (S).
• The transcendence degree of E over F is the cardinality of a transcendence
base.
Note that a ∈ E is transcendent over F if and only if a is algebraically independent
over F if and only if {a} is a transcendence basis of F (a) over F .
√ √
Example 88. { 2, e} has transcendence degree one over Q, since 2 is algebraic
over Q, but e is not. It is not known whether {π, e} are algebraically independent
over Q. The field of complex numbers has transcendence degree 2κ . 4
The following facts can be shown analogously to the corresponding statements
about bases of vector spaces; these facts also follow from more general facts about
pre-geometries that we present in Section 10.1.
Proposition B.4.14. Let E be a field extension of F . Then

• S ⊆ E is a transcendence basis if and only if it is a maximal algebraically
independent subset of E;
• Every field has a transcendence basis (this statement relies on Zorn’s lemma)
• Any two transcendence bases for E over F have the same cardinality.
APPENDIX C
Ramsey’s theorem
We denote the set {0, . . . , n − 1} also by [n].

Subsets of a set of cardinality s will
be called s-subsets in the following. Let M s denote the set of all s-subsets of M .
We also refer to mappings χ : M

s → [c] as a coloring of M (with the colors [c]). In
Ramsey theory, one writes
L → (m)sc
L

if for every χ : s → [c] there exists an M ⊆ L with |M | = m such that χ is constant
on M

s . In the following, ω denotes the cardinality of N. Note the following.
• For all n ∈ N we have [n + 1] → (2)1n : this is the pigeon-hole principle.
• For all c ∈ N we have N → (ω)1c : this is the infinite pigeon-hole principle.
We first state and prove a special case of Ramsey’s theorem.
Theorem C.0.1. N → (ω)22 .
This statement has the following interpretation in terms of undirected graphs:
every countably infinite undirected graph either contains an infinite clique (a complete
subgraph) or an infinite independent set (a subgraph without edges).

Proof. Let χ : N2 → [2] be a 2-colouring of the edges of N2 . We define an
infinite sequence x0 , x1 , . . . of numbers from N and an infinite sequence V0 ⊇ V1 ⊇ · · ·
of infinite subsets of N. Start with V0 := N and x0 = 0. By the infinite pigeon-hole
principle, there is a c0 ∈ [2] such that {v ∈ V0 | χ(x0 , v) = c0 } =: V1 is infinite. We
now repeat this procedure with any x1 ∈ V1 and V1 instead of V0 . Continuing like
this, we obtain sequences (ci )i∈N , (xi )i∈N , (Vi )i∈N .
Again by the infinite pigeon-hole principle, there exists c ∈ [2] such that ci = c
for infinitely many i ∈ N. Then P := {xi | ci = c} has the desired property. To see
this, let i < j be such that xi , xj ∈ P . Then χ({xi , xj }) = ci = c.
We now state Ramsey’s theorem in it’s full strength; the proof is similar to the
proof of Theorem C.0.1 shown above.
Theorem C.0.2 (Ramsey’s theorem). Let s, c ∈ N. Then N → (ω)sc .
A proof of Theorem C.0.2 can be found in [19] (Theorem 5.6.1); for a broader
introduction to Ramsey theory see [14]. Compactness and Theorem C.0.2 imply the
following finite version of Ramsey’s theorem (exercise). In fact, full compactness is
not needed, as we will see in our proof which is based on Kőnig’s tree lemma.
A walk in a graph (V, E) (see Example 1) is a sequence x0 , x1 , . . . , xn ∈ V with the
property that (xi , xi+1 ) ∈ E for all i ∈ {1, . . . , n−1}. A walk is a path if all its vertices
are distinct. A cycle is a walk of length at least three of the form x0 , x1 , . . . , xn = x0
such that x1 , . . . , xn are pairwise distinct. A tree is a connected graph (V, E) (see
Section 1.1.6) without cycles. The degree of a vertex u ∈ V is the number of vertices
v ∈ V such that {u, v} ∈ E.
105
106 C. RAMSEY’S THEOREM
Lemma C.0.3 (Kőnig’s Tree Lemma). Let (V, E) be a tree such that every vertex
in V has finite degree, and let v0 ∈ V . If there are arbitrarily long paths that start in
v0 , then there is an infinitely long path that starts in v0 .
Proof. Since the degree of v0 is finite, there exists a neighbour v1 of v0 such
that arbitrarily long paths start in v0 and continue in v1 (by the infinite pigeon-
hole principle). We now construct the infinitely long path by induction. Suppose
we have already found a sequence v0 , v1 , . . . , vi that can be continued to arbitrarily
long paths in (V, E). Since the degree of vi is finite, vi must have a neighbour vi+1 in
V \{v0 , v1 , . . . , vi } such that v0 , v1 , . . . , vi+1 can be continued to arbitrarily long paths
in (V, E). In this way, we define an infinitely long path v0 , v1 , v2 , . . . in (V, E).
The degree assumption in Lemma C.0.3 is necessary (exercise).

Theorem C.0.4 (Finite version of Ramsey’s theorem). For all c, m, s ∈ N there
is an l ∈ N such that [l] → (m)sc .
Proof. A proof by contradiction: suppose that there are positive integers c, m, s
such that for all l ∈ N there is a χ : [l] → [c] such that (∗)[l] for all m-subsets M of [l]
s
the mapping χ is not constant on M s . We construct a tree as follows. The vertices are
the maps χ : s → [c] that satisfy (∗)[l] . We make the vertex χ : [l]
[l]

s → [c] adjacent
to χ : [l+1] 0

s → [c] if χ is a restriction of χ . Clearly, every vertex in the tree has finite
degree. By assumption, there are arbitrarily long paths that start in the vertex χ0
where χ0 is the map with the empty domain. By Kőnig’s tree lemma the tree contains
an infinite path χ0 , χ1 , . . . We use this to define a map χN : Ns → [c] as follows. For
every x ∈ N, there exists a c0 ∈ [c] and an i0 ∈ N such that χi (x) = c0 for all i ≥ i0 .
Define χN (x) := c0 . Then χN satisfies (∗)N , a contradiction to Theorem C.0.2.
Exercises.
(97) Let (X; <) be a partially ordered set on a countably infinite set X. Show
that (X; <) contains an infinite chain, or an infinite antichain.
(98) Show that an infinite sequence of elements of a totally ordered set contains
one of the following:
• a constant subsequence;
• a strictly increasing subsequence;
• a strictly decreasing subsequence.
Derive the Bolzano-Weierstrass theorem (every bounded sequence in Rn has
a convergent subsequence), using the completeness property of Rn .
Ramsey’s theorem can be used to prove the Erdös-Makkai theorem which we need
in the proof of Theorem 8.1.5 about equivalent characterisations of stability.
Theorem C.0.5 (Erdős-Makkai). Let A be an infinite set and let S be a set of
subsets of A with |A| < |S|. Then there are a sequence (ai )i∈I of elements of A and
a sequence (Si )i∈I of elements of S such that one of the following holds:
(1) for all i, j ∈ ω ai ∈ Sj ⇔ i < j
(2) for all i, j ∈ ω ai ∈ Sj ⇔ j < i.
Proof. There are at most |A| many pairs of finite subsets of A. So we may
choose S 0 ⊆ S with |S 0 | = |A| such that any two finite subsets A0 and A1 of A that
can be separated by a set in S can also be separated by a set in S 0 . By assumption
there is some S ∗ ∈ S which is not a Boolean combination of elements of S 0 , because
there are at most |B| different Boolean combinations of elements of S 0 .
C. RAMSEY’S THEOREM 107
We construct by induction a sequence (a0i )i∈ω of elements of S ∗ , a sequence (a00i )i∈ω

of elements of B \ S ∗ , and a sequence (Si )i∈ω of elements of S 0 such that for all n ∈ ω
• {a00 , . . . , a0n } ⊆ Sn
• {a000 , . . . , a00n } ⊆ A \ Sn ,
• a0n ∈ Si ⇔ a00n ∈ Si for all i < n.
Assume that (a0i )i<n , (a00i )i<n , and (Si )i<n have already been constructed. Since
S ∗ is not a (positive) Boolean combination of S0 , . . . , Sn−1 , there are a0n ∈ S ∗ and
a00n ∈ A \ S ∗ such that for all i < n
a0n ∈ Si ⇔ a00n ∈ Si .
By the choice of S 0 there is some Sn ∈ F 0 with
{a00 , . . . , a0n } ⊆ Sn and {a000 , . . . , a00n } ⊆ A \ Sn .
By Ramsey’s theorem (Theorem C.0.2) we may assume that either
• a0j ∈
/ Si for all i < j ∈ ω, or
• a0j ∈ Si for all i < j ∈ ω.
In the first case we set (ai )i∈ω := (a0i )i∈ω and are in case (1): for all i < j ∈ ω we
have a0i ∈ Sj and a0j ∈ / Si . In the second case, we set (ai )i∈ω := (a00i )i∈ω and are in
case (2): for all i < j ∈ ω we have a00i ∈/ Sj and a00j ∈ Si since a0j ∈ Si .
Bibliography
[1] S. A. Adeleke and P. M. Neumann. Relations related to betweenness: their structure and auto-
morphisms, volume 623 of Memoirs of the AMS. American Mathematical Society, 1998.
[2] G. Ahlbrandt and M. Ziegler. Quasi-finitely axiomatizable totally categorical theories. Annals
of Pure and Applied Logic, 30(1):63–82, 1986.
[3] M. Bodirsky. Introduction to mathematical logic, 2023. Course notes, TU Dresden, https:
//wwwpub.zih.tu-dresden.de/~bodirsky/Logic.pdf.
[4] M. Bodirsky, P. Jonsson, and T. V. Pham. The reducts of the homogeneous binary branching
C-relation. Journal of Symbolic Logic, 81(4):1255–1297, 2016. Preprint arXiv:1408.2554.
[5] M. Bodirsky, M. Mamino, and C. Viola. Piecewise linear valued CSPs solvable by linear pro-
gramming relaxation. ACM Transactions of Computational Logic, 23(1):1–35, 2022. Preprint
arXiv:1912.09298.
[6] S. Braunfeld, A. Dawar, I. Eleftheriadis, and A. Papadopoulos. Monadic NIP in monotone
classes of relational structures. In 50th International Colloquium on Automata, Languages, and
Programming, ICALP 2023, July 10-14, 2023, Paderborn, Germany, pages 119:1–119:18, 2023.
[7] P. J. Cameron. Oligomorphic permutation groups. Cambridge University Press, Cambridge,
1990.
[8] P. J. Cameron. Homogeneous permutations. Electronic Journal of Combinatorics, 9(2), 2002.
[9] A. Chernikov. Lecture notes on stability theory. AMS Open Math Notes, 2019.
[10] Dummit and Foote. Abstract Algebra. Wiley, 2004. Third edition.
[11] J. Ferrante and C. Rackoff. A decision procedure for the first order theory of real addition with
order. SIAM Journal on Computing, 4(1):69–76, 1975.
[12] R. Fraïssé. Sur l’extension aux relations de quelques propriétés des ordres. Annales Scientifiques
de l’École Normale Supérieure, 71:363–388, 1954.
[13] R. Fraïssé. Theory of Relations. Elsevier Science Ltd, North-Holland, Amsterdam, 1986.
[14] R. L. Graham, B. L. Rothschild, and J. H. Spencer. Ramsey theory. Wiley-Interscience Series
in Discrete Mathematics and Optimization. John Wiley & Sons, Inc., New York, 1990. Second
edition.
[15] D. Haskell and D. Macpherson. Cell decompositions of C-minimal structures. Annals of Pure
and Applied Logic, 66:113–162, 1994.
[16] C. W. Henson. Countable homogeneous relational systems and categorical theories. Journal of
Symbolic Logic, 37:494–500, 1972.
[17] C. W. Henson. Model theory. Class Notes for Mathematics 571, University of Illinois, 2010.
[18] W. Hodges. Model theory. Cambridge University Press, Cambridge, 1993.
[19] W. Hodges. A shorter model theory. Cambridge University Press, Cambridge, 1997.
[20] T. Jech. Set theory. Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2003. The
third millennium edition, revised and expanded.
[21] S. Lang. Algebra. Springer, 2002. Revised third edition.
[22] D. Marker. Model Theory: An Introduction. Springer, New York, 2002.
[23] Y. V. Matiyasevich. Hilbert’s Tenth Problem. MIT Press, Cambridge, Massachusetts, 1993.
[24] B. Poonen. Characterizing integers among rational numbers with a universal-existential formula.
American Journal of Mathematics, pages 675–682, 2009.
[25] P. Rothmaler. Introduction to Model Theory. CRC Press, 2000.
[26] D. Saracino. Model companions for ℵ0 -categorical theories. Proceedings of the American Math-
ematical Society, 39:591–598, 1973.
[27] J. Schreier and Stanisław Marcin Ulam. Über die Permutationsgruppe der natürlichen Zahlen-
folge. Studia Mathematica, 4:134–141, 1933.
[28] P. Simon. A Guide to NIP Theories. Lecture Notes in Logic. Cambridge University Press, 2015.
[29] K. Tent and M. Ziegler. A course in model theory. Lecture Notes in Logic. Cambridge University
Press, 2012.
109
110 BIBLIOGRAPHY
[30] V. N. Vapnik and Z. Y. Chervonenkis. On the uniform convergence of relative frequencies of

events to their probabilities. Thoeret. Probl and its Appl., 16(2):264–280, 1971.

Model Theory

Uploaded by

Copyright:

Available Formats

Model Theory

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Model Theory

Uploaded by

Copyright:

Available Formats

Model theory

November 30, 2023

Chapter 1. First-order Logic 5

Chapter 3. The Löwenheim-Skolem Theorem 21

Chapter 4. Fraïssé Amalgamation 27

Chapter 6. Countably Categorical Structures 43

Chapter 7. Model Completeness and Quantifier-Elimination 53

Chapter 9. NIP Theories 75

1.1. First-order Structures

1.1.2. Structures. A τ -structure A is a set A (called the domain, or the base

obtain a bijective correspondence between undirected graphs with vertices V and

1.1.5. Homomorphisms. In the following, let A and B be τ -structures. A

The n-colorability problem can be formulated as a graph homomorphism problem: is

and for f ∈ τ of arity k, we have

(3) Let A and B be τ -structures and suppose that G ⊆ A generates A, i.e.,

1.2. Formulas, Sentences, Theories

• R(t1 , . . . , tk ) where t1 , . . . , tk are τ -terms and R ∈ τ is a k-ary relation

1.2.5. First-order Theories. Let τ be a signature. A (τ -)theory is a set of

Example 11. A famous example of a first-order theory is the set ZF of axioms of

∀x, y, z. (xy)z = x(yz) (associativity)

1.3. Vaught’s Conjecture

Vaught’s Conjecture: if IT (ℵ0 ) is infinite, then IT (ℵ0 ) ∈ {ℵ0 , 2ℵ0 }.

There are several different proofs of the compactness theorem. It is a consequence

if such an n exists, and 0 otherwise. Let T be a theory that contains the

Example 16. For a non-empty subset Y ⊆ X, the family

Are there non-principal ultrafilters?

Lemma 2.2.3 (Ultrafilter Lemma). Every filter F is contained in a ultrafilter.

Proof. Let U be a free ultrafilter on X and let x ∈ X. Either {x} ∈ U or

1Thanks to Lukas Juhrich for the idea for this exercise.

2.4. The Theorem of Łoś

• Suppose that the statement holds for φ and for ψ. Then

Corollary 2.4.2. Let M := AX /U. Then Th(M ) = Th(A).

Example 19. Let U be a free ultrafilter on ω, and consider A := (R; 0, 1, +, ∗, ≤).

2.5. Proof of the Compactness Theorem

2.6. Proving the Ultrafilter Lemma with Compactness

We have M 0 |= T 0 . The compactness theorem asserts the existence of a model M of

The Löwenheim-Skolem Theorem

Note that ≺ is a transitive relation. A chain is called an elementary chain if

to consider the case ψ = ∃x.φ(x). If ψ holds in AA , there exists a ∈ A such that

We use Tarski’s Test to construct small elementary substructures.

1Thanks to Anna Tölle for the idea for this exercise.

3.2. Proving Löwenheim-Skolem

3.3. Vaught’s test

Theorem 3.3.2 (Vaught’s Test). A κ-categorical theory T is complete if the fol-

A relational1 structure A is called homogeneous (sometimes also called ultra-

4.1. The Age of a Structure

4.2. Amalgamation Classes

4.3. Fraïssé’s theorem

to an automorphism a of D. Then a ◦ e = f and we extend f by setting

class, because if A1 and A2 are directed graphs in Forb(N ) such that A = A1 ∩ A2 is

Show that the complement of a homogeneous graph is homogeneous.

4.4. Strong Amalgamation

• The age of (Q; <).

4.5. Further Examples: C-relations

amalgam of B 1 and B 2 over A in C. We inductively assume that the statement has

Figure 4.1. Illustration for the proof of Proposition 4.5.2.

C5 ∀b, c∃a : C(a; b, c);

Loosely speaking, a type of a τ -structure M is a set of formulas that is satisfied

Solution to Exercise 57.