An Even Shorter Model Theory

Justin Bledin
June 4, 2007
These notes were prepared for the UC Berkeley Group in Logic & the Methodology of Science
foundations exam. Material is adapted from Chang and Keislers (1973) Model Theory and Wilfrid
Hodges (1997) A Shorter Model Theory. Sections 1-2 review basic semantic/syntactic concepts;
Section 3 introduces Robinson diagrams and canonical interpretations; Section 4 presents back-
and-forth techniques; Section 5 introduces the method of quantier elimination; Section 6 discusses
Compactness and types; Section 7 examines atomic and saturated models; Section 8 presents the
theorems of Lowenheim-Skolem; Section 9 examines preservation phenomena; Section 10 presents
amalgamations and Frasse limits; Section 11 discusses interpolation; Section 12 discusses model-
completeness; Section 13 covers further miscellaneous topics.
1 Mathematical Structures & Mappings
Denition 1 a structure A is an object with these ingredients:
(1) domain with cardinality [A[ = [dom(A)[;
(2) constants c naming c
A
, elements in A;
(3) n-ary relations R naming R
A
, subsets of (dom(A))
n
;
(4) n-ary functions F naming F
A
, mappings from (dom(A))
n
dom(A).
Examples: Klein 4-group with signature L = 'a, b, c, , e`; ordered naturals with signature
L = 'N, <`.
Denition 2 A is a substructure of B if dom(A) dom(B) and the identity i:dom(A) dom(B) is
an embedding (see below). Then (i) c
A
= c
B
; (ii) R
A
= R
B
(dom(A))
n
; (iii) F
A
= F
B
[(dom(A))
n
.
A = 'Y `
B
is the substructure of B generated by the hull Y , where Y is said to be a set of
generators for A ([A[ [Y [ +[L[; see Hodges, p.8 for proof)
Denition 3 a homomorphism f is a mapping such that:
(1) f(c
A
) = c
B
;
(2) a R
A
fa R
B
; (e.g. x <
A
y f(x) <
B
f(y))
(3) f(F
A
(a)) = F
B
(fa). (e.g. x = y
A
z f(x) = f(y)
B
f(z))
Denition 4 an embedding f has a stronger (2*) a R
A
fa R
B
.
1
2
Denition 5 an isomorphism f is a surjective embedding.
Denition 6 a homomorphism f : A A is an endomorphism of A.
Denition 7 an isomorphism f : A A is an automorphism of A.
e a b c
e e a b c
a a e c b
b b c e a
c c b a e
f

+ 0 x y z
0 0 x y z
x x 0 z y
y y z 0 x
z z y x 0
Figure 1: 'A, , e` and 'B, +, 0` are isomorphic
0 < 1 < 2 < 3 < 4 < 5 < ...

0 1 2 3 4 5 ...
Figure 2: 'N, <` is homomorphic but not isomorphic to 'N, `
Intuitively, if A is homomorphic to B, then the structural properties of A carry over to B. If A
is isomorphic to B (we write A

= B), they are essentially the same structure. Here are some nice
properties of homomorphisms/embeddings:
(i) if f : A B and g : B C are homomorphisms/embeddings, then gf is a homomorphism/
embedding from A C;
(ii) The identity map is an automorphism of A;
(iii) if f is an isomorphism from A B, g = f
1
exists and is an isomorphism from B A;
(iv) if f : A B is a homomorphism and g : B A is a homomorphism with g = f
1
, then f
and g are both isomorphisms;
(v) every homomorphism f : A C can be factored as f = hg for some surjective homomorphism
g : A B and extension h : B C;
(vi) every embedding f : A C can be factored as f = hg where g : A B is an extension and
h : B C is an isomorphism (so we can assume A is a substructure of C).
3
2 Language & Interpretations
A formal language L can be described in stages: rst come the terms; from these, the basic atomic
formulae are built; then after considering the whole array of logical symbols, more complex formulae
can be expressed in L.
Denition 8 a term is a variable, constant or function F(t
1
, ..., t
n
) whose inputs t
i
are terms.
If t(x) is a term in the language L, t
A
(a) is the element in A named by t with x interpreted as a
name of a A.
It is sometimes useful to consider an absolutely free structure in which terms describe themselves.
If L is a language and X a set of variables, the term algebra of L with basis X is the L-structure
A with dom(A)=set of all terms of L whose variables are in X, c
A
= c for constants in L,
and F
A
(t) = F(t) for functions in L and elements t in dom(A) (R
A
is empty). For example,
if L is the signature of additive groups with identity element 0 and X = x, y, z, dom(A) =
0, 0 + x, x + y, (x + y) + z, .... Note that elements of the term algebra A are not elements of a
particular group but rather terms in the language of groups.
Denition 9 atomic formulae are strings of symbols built from terms as follows:
(1) if s, t are terms, s = t is an atomic formulae of L;
(2) if R is an n-ary symbol and t
1
, ..., t
n
are terms, R(t
1
, ..., t
n
) is an atomic formulae of L.
The class of all formulae of L is then generated from the atomic formulae by introducing the
logical connectives. If , are atomic formulae, then , , , y and y are all formulae
of L and so on. Formulae with only bound occurrences of variables (i.e., containing only closed
terms) are called sentences.
Languages are typically denoted with two subscripts L
xy
. The rst x says how many formulae
we can join together by

and

. The second y says how many quantiers we can put together
in a row. A rst-order language L

is one in which all formulae are nite and only nitely many
quantiers appear in each formula.
The cardinality of a language [[L[[ is the least innite cardinal the number of symbols [L[ in the
language (i.e., [[L[[ = [L[). This denition has two virtues: (a) [[L[[ coincides with the number
of non-equivalent (under the relation of being variants) formulae expressible in L (to see this, note
that if [L[ is countable, then # terms is countable # atomic formulae is countable # formulae
is countable; otherwise, if # functions F
i
is uncountable # of terms is uncountable and if #
relations R
i
is uncountable # atomic formulae is uncountable); (b) what we often really care
about is whether a language is countable or not as this will be relevant in certain constructions
where we wish to build countable models (such as in the proof of the Omitting Types theorem). If
[L[ is nite, it is still common to call L a nite signature.
4
In model theory, quantiers are especially important (see Section 5 on quantier elimination and
Section 9 on preservation phenomena) and we will work with a hierarchy of rst-order formulae:
(i)
0
0
and
0
0
formulae have all quantiers bounded where the quantiers (x < y) and (x < y)
are bounded (alternatively, we use the notation
0
and
0
when no distinction is made between
bounded/unbounded quantication so the formulae are quantier-free);
(ii) a
0
k+1
formula is of the form x where is a
0
k
formula (or
k+1
,
k
respectively);
(iii) a
0
k+1
formula is of the form x where is a
0
k
formula (or
k+1
,
k
respectively);
(as x in (ii) and (iii) can be empty, every
0
k
/
0
k
formula is also both a
0
k+1
and
0
k+1
formula)
Through interpretation of its symbols, a language allows us to talk about mathematical struc-
tures and the relations between them. Specically, a formula (a) in L
A
talks about an element a
in the structure A; a sentence in L
A
talks about global structural features of A (e.g., x(x < 0)
says there are no negative elements in a structure). A formula may speak truthfully but it might
also talk rubbish. In the former case, A [= and is true in A; otherwise, A [= and is false
in A (for full truth conditions, see Hodges: 12/24). If A [= T where T is a theory (i.e., set of
sentences), we say A is a model of T.
We write Mod(T) for all models of the theory T in L. If K = Mod(T), we say T axiomatizes
the class K of L-structures. Theories T and T

are equivalent if Mod(T) = Mod(T

). We write
Th(K) for the set of all rst-order sentences true in every structure in K. A theory often aims to
describe a particular structure or class of structures. For example, here are the rst-order Peano
axioms:
(i) x(x + 1 = 0)
(ii) x(x + 0 = x)
(iii) xy(x + 1 = y + 1 x = y)
(iv) xy(x + (y + 1) = (x +y) + 1)
(v) x(x 0 = 0)
(vi) xy(x (y + 1) = x y +x)
(vii) [(0) x((x) (x + 1))] x(x) (induction schema for all denable sets)
The conditions (i) and (iii) dene succession; (ii) and (iv) recursively dene addition; (v) and
(vi) recursively dene multiplication; (vii) is the rst-order induction axiom schema (so (vii) fathers
an innite number of axioms). While the intended model of rst-order Peano arithmetic (PA) is
'N, 0, 1, +, `, we shall see in Section 6 that there are nonstandard models as well (as implied by
Godels proof of the incompleteness of arithmetic).
5
Given a language L, certain relations are denable on a structure A:
Denition 10 a relation R is rst-order denable if R
A
= (A
n
, b) = a : A [= (a, b) for some
rst-order formula in L and parameters b A.
Example: the formula (p) = x(x[p (x = 1 x = p)) denes the prime numbers (where
x[p says x divides p). The ordering x < y can also be explicitly dened in PA with (x, y) =
z(x + (z + 1) = y). More generally, a set of natural numbers is a
0
k
(resp.
0
k
) relation if it is
denable by some
0
k
(
0
k
) formula. R is a
0
k
relation if it is both a
0
k
and
0
k
relation.
Application: a structure A is minimal if A is innite but the only rst-order denable sub-
sets of dom(A) with parameters are either nite or conite. Likewise, a rst-order denable set
X dom(A) is minimal if X is innite and for every other denable set Y , XY or X`Y is nite.
A is said to be O-minimal if L contains < which linearly orders dom(A) in such a way that any rst-
order denable subset is a union of nitely many intervals of the form (a, b), (a), (, b), (a, ) for
a, b A. The following lemma is often helpful in analyzing the denable relations on a structure:
Lemma 1 suppose Y is a denable set on A with parameters from X, then every automorphism
of A which xes X pointwise (i.e., (a) = a for all a X) xes Y setwise (i.e., for all a A,
a Y (a) Y ).
Having a language to talk about structures also provides syntactic characterizations of mappings
between structures (more of this in Section 5):
Theorem 1 f : A B is a homomorphism A [= (a) B [= (fa) for all atomic formulae
L and tuples a A.
Theorem 2 f : A B is an embedding A [= (a) B [= (fa) for all literals L and
tuples a A.
This agrees with our earlier denitions of homomorphism and embedding as formula preserva-
tion between structures structural equivalence. If A is homomorphic to B, anything an atomic
truthfully says about A (and its elements) will also hold of B as the structural properties of A
carry over to B (and conversely if the structures are isomorphic). There is also the purely syn-
tactic condition of elementary equivalence where a rst-order language cannot distinguish between
structures:
Denition 11 an elementary embedding f is a mapping which preserves all rst-order formulae.
Two structures A and B are elementary equivalent, A B, if they agree on all rst-order sentences.
A hierarchy of mappings has now emerged: homomorphisms preserve atomic formulae; embed-
dings preserve literals; elementary embeddings preserve all formulae. Clearly, A

= B A B
but the converse does not hold (unless A, B are nite) as rst-order languages cannot distinguish
between models of diering innite cardinalities (see Section 8).
6
Denition 12 A is an elementary substructure of B, denoted A B, if A B and the inclusion
map i : A B is an elementary embedding (B is said to be an elementary extension of A). We
write A B when A is a proper elementary substructure of B.
Note that A B and A B does not necessarily imply that A B (though the converse
clearly holds). Consider when A = '[1, ), <` and B = '[0, ), <`. Though f(n) = n 1 is an
elementary embedding from A B, the inclusion map i : A B is not an elementary embedding
as A [= y(1 y) while this formula (1) does not hold in B. Thus A B is a stronger notion
than A B as the elements common to A and B must have exactly the same rst-order properties
(with respect to A,B). The following theorems are useful for constructing/working with elementary
substructures:
Theorem 3 (Tarski-Vaught criterion for elementary substructures) A B for every formula
L and a A, B [= y(a, y) B [= (a, d) for some d A.
Theorem 4 (Tarski-Vaught theorem on unions of elementary chains) Let A
0
A
1
... A
n
...
be an elementary chain of L-structures. Then for all A
i
, A
i

i
A
i
.
To form an elementary substructure of A, we can start with some subset X A and close o X
under existential claims true in A. Constructing elementary chains is also a powerful method for
iterating the notion of elementary equivalence into the transnite.
3 From Structures to Language and Back Again
(1) Structures Language (diagrams)
A powerful tool in model theory, a diagram of a structure is formed by naming elements of the
structure by new constants (called parameters). The language obtained from L by adding the con-
stants c is denoted L(c) and L(c) L. We often do not need to name every element in A (though
this would do the trick) as if A = 'a`
A
for some a A, then introducing parameters for a ensures
every element in A is named in L(c) (the structure (A, a) in L(c) is considered an extension of A).
Diagrams can be regarded as generalizations of the multiplication table for a group (see the Klein
4-group table on p.2 above).
Denition 13 the Robinson diagram of A, diag(A), is the set of all closed literals of L(c) which
are true in (A, a) (i.e., the set L(c)[(A, a) [= ). The set of atomic sentence of L(c) true in
(A, a) is called the positive diagram of A and denoted diag
+
(A). The set of all sentences of L(c)
true in (A, a) (i.e., Th(A, a)) is called the elementary diagram of A, eldiag(A).
7
Lemma 2 (Diagram Lemma) Let (A, a) and (B, b) be structures in L(c), then the following are
equivalent:
(a) for every atomic sentence L(c), (A, a) [= (B, b) [= ;
(b) there exists a unique homomorphism f : 'a`
A
B such that fa = b.
(if (a) holds for all literals in L(c), f is an embedding; if (a) holds for all sentences in L(c), f
is an elementary embedding)
Note that when A = 'a`
A
the Diagram Lemma essentially says that A is homomorphic to a
reduct of B if B [= diag
+
(A). Once we know the diagram of a structure, we know the structure
up to isomorphism. As we will see, diagrams are particularly useful in constructive proofs where
we want to nd extensions/reductions of structures with additional properties. More generally, by
converting a structure to language, diagrams give us a way to work with structures in syntactic
arguments.
(2) Language Structures (canonical models)
One of the essential results in logic is the Completeness of rst-order logic; namely, that for every set
of consistent sentences we can always nd a model in which the sentences hold true. Here we explore
the basis of one route to this result (Henkins famous Completeness/Compactness proof; presented
in detail in Section 6), constructing the canonical interpretation of a set of atomic sentences.
Let T be an =-closed set of atomic sentences in L (so t = t T for all closed terms t L and
s = t T implies (s) T (t) T). Dene the equivalence relation by: s t s = t T.
The relation thus creates equivalence classes t` of closed terms in L and for the canonical
model A, we set dom(A) = t

t` . Every element in A is named by a closed term of L.

Constants, relations and functions are then dened as:
(i) c
A
= c

;
(ii) (t

1
, ..., t

n
) R
A
R(t
1
, ..., t
n
) T;
(iii) F
A
(t

1
, ..., t

n
) = (F(t
1
, ..., t
n
))

.
It follows that T is the set of all atomic sentences true in A as we have:
A [= s = t s
A
= t
A
s

= t

s = t T
A [= R(t
1
, ..., t
n
) (t
A
1
, ..., t
A
n
) R
A
(t

1
, ..., t

n
) R
A
R(t
1
, ..., t
n
) T
For example, in the language L
G
= 'G, , e` the only atomic sentences are of the form e = e,
e e = e, (e e) e = e, etc. so in the canonical model A of T = all such atomic formulae in L
G
,
dom(A) = e

, c
A
= e

and e

A
e

= e

. Note that if T is a set of atomic formulae that is

not =-closed, we can always add sentences to T to form its =-closure T

and if A is the canonical

model of T

, we then have A [= T as T T

. We will later see that this construction can also be

generalized to more complex theories (providing certain additional conditions are still met).
8
4 Back and Forth between Structures
Back-and-forth techniques are useful for showing two structures are isomorphic. The origin of this
proof strategy allegedly goes back to Cantor who used back-and-forth methods to show that the
elements of any two countable dense linear orderings without endpoints could be mapped from one
structure to the other and back again. The modern incarnation of the back-and-forth argument is
Frasses notion of back-and-forth equivalence, introduced in the 1950s and sometimes presented,
as in Hodges (1997: 74-81), in terms of Ehrenfeucht-Frasse games.
The players are belard and loise, a twelfth-century Parisian logician and his student/lover, the
niece of a Notre Dame canon. Given two structures A and B, belard wants to prove A dierent
from B while loise wants to prove them identical. They take turns choosing elements a
i
from A
and b
i
from B at each step of the game, belard freely choosing from either A or B and loise
choosing from the opposite structure. At the end of the game with a (countably) innite number of
steps, denoted EF

(A, B), sequences of elements a = a

i
and b = b
i
have been chosen from A
and B respectively with the pair (a, b) known as the nal play. The play (a, b) is a win for loise
if there exists an isomorphism f : 'a`
A
'b`
B
between substructures 'a`
A
A and 'b`
B
B such
that fa = b. In other words, the substructures of A and B generated from the elements a
i
and
b
i
must be structurally equivalent.
In this context, Frasses back-and-forth equivalence can now be dened as follows:
Denition 14 Two structures A and B are back-and-forth equivalent, A

B, if loise can always

win the game EF

(A, B).
When loise knows of an isomorphism f : A B, EF

(A, B) is easily won as she can choose

f(a
i
) from B when belard chooses the corresponding element a
i
from A and f
1
(b
i
) when belard
picks b
i
. The game is more interesting when loise knows of no such isomorphism but fortunately
there is a useful criterion for determining exactly when two structures are back-and-forth equivalent
(and in the countable case, isomorphic):
Denition 15 A back-and-forth system from A to B is a set of functions J satisfying the following
conditions:
(BF1) each f J is an isomorphism f : 'a`
A
'b`
B
(BF2) J is non-empty
(BF3) f J and c A, there is g f such that g J and c dom(g)
(BF4) f J and d B, there is g f such that g J and d im(g)
Lemma 3 A,B are back-and-forth equivalent there is a back-and-forth system from A to B.
Lemma 4 Let A,B be countable structures. Then A,B are isomorphic they are back-and-forth
equivalent.
Two examples are provided below.
9
Theorem 5 If A,B are countable dense linear orderings without endpoints, A,B are isomorphic.
Proof : In the base case, '`
A
= and '`
B
= so (BF2) clearly holds. Now assume that
'a`
A
= a
1
< ... < a
n
, 'b`
B
= b
1
< ... < b
n
and there exists an isomorphism f : 'a`
A
'b`
B
with
f(a) = b. We must show that the forth step holds. Choose c A not in a
1
, ..., a
n
. Either c < a
i
,
a
i
< c a
i
, or a
i
< c < a
j
for some a
i
, a
j
a
1
, ..., a
n
. In the latter case, the density of B ensures
we can nd some d B such that f(a
i
) < d < f(a
j
) - let g = f (c, d). In the previous cases,
the innite extension of B in both directions ensures we can nd the required d. An analogous
argument works for the back step.
Theorem 6 If A,B are countable atomless Boolean algebras, A,B are isomorphic.
Proof : In the base case, where '`
A
and '`
B
are Boolean algebras with only zero and one
elements, let f(0
A
) = 0
B
and f(1
A
) = 1
B
; so (BF2) holds. Now consider 'a`
A
with atoms a
1
, ..., a
n
and 'b`
B
with atoms b
1
, ..., b
n
and assume there exists an isomorphism f : 'a`
A
'b`
B
with
f(a) = b. We must show that the forth step holds. Choose c A not in a
1
, ..., a
n
. Then the
structural identity (what Hodges calls isomorphism type) of c over 'a`
A
is determined by, for each
atom a
i
, whether c a
i
equals a
i
, 0 or neither. c a
i
is the greatest lower bound of c and a
i
so,
intuitively, the split concerns whether c lies above a
i
, beside a
i
or below a
i
. Now as B is atomless,
we can nd some d B such that d f(a
i
) = f(a
i
), 0 or neither c a
i
= a
i
, 0 or neither - let
g = f (c, d). An analogous argument works for the back step.
5 Quantier Elimination
This section takes us back to the early days of model theory, the method of quantier elimination
originating in Tarskis Warsaw seminar in the late 1920s. The idea is simple but useful: show
that relative to a theory T, all formulae in a rst-order language L are equivalent to Boolean
combinations of formulae in an elimination set . If contains only quantier-free formulae, then
all formulae in L are equivalent (mod T) to quantier-free formulae and T is said to have the
property of quantier elimination. It is important to distinguish between the method of quantier
elimination, i.e., nding an elimination set or reducing a formula to a Boolean combination of
formulae in , from the property of quantier elimination. In the special cases where the method of
quantier elimination is successful, the result is a condensed description of all complete extensions
of T, simplifying the study of denable sets on models of T and usually leading to completeness
and decidability proofs (see Hodges 1997: 60-1 for details).
The traditional approach to quantier elimination (used here) is a heavily syntactic approach
(model theorists such as Abraham Robinson later encouraged the use of good structural information
about the models of T, rather than syntax, to show T admits quantier elimination and we will
later see an alternative route to quantier elimination in Section 12). Chang and Kiesler (1997:
49) write: the method may be thought of as a direct attack on a theory. Though these attacks
are not very dicult, they can be quite tedious. Fortunately, the following lemma eases the burden
tremendously:
10
Lemma 5 Suppose that:
(i) every atomic formula L is in ;
(ii) for every = x(x, y) in L with a Boolean combination of formulae in , is equivalent
to a Boolean combination of formulae in with respect to every structure in a class K;
Then is an elimination set for K.
So nding an elimination set reduces to eliminating the quantier x. One thorough example is
provided here from the theory of dense linear orderings without endpoints (DLO). To tighten the
below proof, I take the completeness of DLO for granted at the onset. An alternative route, taken
by Chang and Kiesler (50-4), is to prove DLO admits quantier elimination and have completeness
fall out as an easy consequence.
Theorem 7 DLO admits quantier elimination
Proof : It must be shown that for every formula in L, there exists a quantier-free such
that DLO ( is equivalent to mod DLO). First consider when is a sentence so as DLO
is complete, either DLO or DLO . If DLO , then DLO x
1
= x
1
; if DLO ,
then DLO x
1
< x
1
(note that x
1
= x
1
and x
1
< x
1
could be replaced here by and if our
language allowed it).
So suppose is a formula with free variables x
1
, ..., x
n
. We show the set of atomic formulae
: x
i
= x
j
,x
i
< x
j
forms an elimination set for the class of all models of DLO (i.e., is DLO-
equivalent to a Boolean combination of the formulae in ). Dene an arrangement of x
1
, ..., x
n
to
be the nite conjunction of formulae =
1

2
...
n
where each
i
is of the form x

i
< x

i+1
or
x

i
= x

i+1
for a renumbering of x
1
, ..., x
n
as x

1
, ..., x

n
. Each arrangement gives us an exact ordering
of the n free variables. Now when n = 1, an open formula is built from and DLO x
1
= x
1
and DLO x
1
< x
1
so we must have DLO (in which case DLO x
1
= x
1
) or DLO
(in which case DLO x
1
< x
1
). So consider when n > 1. If DLO is inconsistent,
DLO x
1
< x
1
. If DLO is consistent with open, must be DLO-equivalent to a
disjunction of nitely many arrangements of x
1
, ..., x
n
as this exhausts what we can say about the
variables in in terms of Boolean combinations of the atomic formulae in (i.e., DLO

).
As an example, the formula (x
1
, x
2
, x
3
) = (x
1
< x
3
) (x
2
< x
3
) is equivalent (mod DLO) to

= ((x
1
= x
2
) (x
2
< x
3
)) ((x
1
< x
2
) (x
2
< x
3
)) ((x
2
< x
1
) (x
1
< x
3
)) as only these
three possible arrangements of x
1
, x
2
, x
3
are compatible with .
We can now use this condition to show that every formula in L is DLO-equivalent to an open
formula. Here, it suces to show that if is an open formula, x
n
is DLO-equivalent to an open
formula. By the above, DLO x
n
x
n

so DLO x
n

x
n
. For each arrangment
, let

be the restricted arrangement of x

1
, ..., x
n1
obtained by omitting the literal involving
x
n
. Now the crucial step is this: given the extension axioms ((iv) and (v) above) for DLO, telling
us that an element can be found in any possible interval over the ordering, x
n
is equivalent (mod
DLO) to

. Thus, DLO x
n

where

is a Boolean combination of the formulas in

so DLO admits quantier elimination.
11
In the rst part of this proof, the completeness of T (i.e., that T or T for every
sentence L, or alternatively, that any models of T are elementarily equivalent) allowed us to
swiftly deal with the case where is a sentence. While complete theories do not necessarily have
the property of quantier elimination (e.g., the theory of dense linear orderings with endpoints),
it is generally easy to show that an incomplete theory does not have this property by nding a
sentence with quantiers that is not a consequence of the theory. Slight modications of this proof
(in particular, regarding the elimination of the nal existential quantier) can also be used to show
that other complete theories admit quantier elimination, such as the theory of atomless Boolean
algebras and the theory of an equivalence relation with innitely many innite classes.
6 Into the Innite
6.1 The Compactness Theorem
Earlier in Section 3, we constructed the canonical interpretation of a set of atomic sentences. This
model-existence argument can be generalized to an arbitrary set of sentences satisfying certain
properties:
Denition 16 A Hintikka set for L is a theory with the following properties:
(H1) for every atomic sentence L, T T;
(H2) for every closed term t L, the sentence t = t T;
(H3) for closed terms s, t L and atomic formula , s = t L ((s) T (t) T);
(H4) T T;
(H5)

T T;

T such that T;
(H6)

T such that T;

T T;
(H7) x(x) T (t) T for some closed term t L; x(x) T (t) T for every
closed term t L.
Theorem 8 If T is a Hintikka set for L, then the canonical model A of the set of atomic sentences
in T is a model of T.
The conditions (H1)-(H7) ensure that for every sentence L, A [= T (so A [= T).
(H2) and (H3) are the conditions from Section 3 which ensure T is =-closed and the closed terms in
L can be divided into equivalence classes based on whether they are equal relative to T, allowing us
to ignore redundant closed terms when constructing A. (H1)/(H4) ensure that negation does not
lead to any inconsistencies in T and (H5)/(H6) analogously ensure that conjunction and disjunction
do not lead to problems. (H7) is known as the Henkin property and the closed terms t witness
the existential claims.
12
Theorem 9 Let T be a theory in L such that:
(i) T is consistent;
(ii) for every sentence L, either T or T;
(iii) for every sentence x(x) T there is a witness t L such that (t) T;
Then T is a Hintikka set for L.
While we had that for a Hintikka set T, A [= T where A is the canonical interpretation of the
set of atomic sentences in T, we now have the reverse: Th(A), the complete theory of A which
clearly satises (i)-(iii) above, is a Hintikka set. But we now also have a blueprint for constructing
Hintikka sets as given a consistent set of sentences in L, we must only expand this set to a maximally
consistent set with the Henkin property. Such a construction underlies Henkins (1949) proof of
the Completeness/Compactness of rst-order logic:
Theorem 10 (Completeness Theorem for rst-order logic) Let T be a set of sentences in L.
Then T is consistent T has a model.
Proof : The direction () is trivial. To prove the direction (), we expand T to a Hintikka
set T
+
and consider the canonical interpretation of T
+
. First add new constants c
i
[i < [[L[[ to
L to act as witnesses, forming L
+
, and enumerate the sentences in L
+
as
i
[i < [[L[[. We then
dene an increasing chain of theories T
i
[i < [[L[[ in L
+
as follows:
Set T
0
= T. Then let T

i+1
= T
i

i+1
if this set is consistent; otherwise, set T

i+1
= T
i
.
If
i+1
is of the form x(x), then take the earliest witness c
j
not already used in T

i+1
and set
T
i+1
= T

i+1
(c
j
); otherwise, put T
i+1
= T

i+1
. At limit ordinals , take T

=

i<
T
i
. If we
then take the union T
+
=

i<||L||
T
i
, then by construction T
+
is consistent, maximal and has the
Henkin property so T
+
is a Hintikka set. It follows that the canonical interpretation A of the set
of atomic sentences in T
+
models T
+
so as T T
+
, A[L [= T.
Theorem 11 (Compactness Theorem for rst-order logic) Let T be a set of sentences in L.
Then T is satisable T is nitely satisable.
Proof : If T is nitely satisable, then by Completeness every nite subset of T is consistent so
T is consistent. By Completeness again, T has a model. The other direction is trivial.
Henkins proof is our rst example of constructing models from constants. Similar constructions
are used in the proofs of the Omitting Types theorem (next subsection) and Craig Interpolation
theorem (Section 9). To see that Compactness fails in innite languages, consider the theory
T =

0<i<
c
0
= c
i
, c
0
= c
1
, c
0
= c
2
, c
0
= c
3
, .... While T is nitely satisable, T does not have
a model.
13
We end this subsection with various easy corollaries/applications of Compactness (though we
leave the Upward Lowenheim-Skolem theorem for Section 8):
Theorem 12 If T is a rst-order theory and T for some sentence , then U for some
nite subset U T.
Proof : Suppose U for every nite subset U T. Then T is nitely satisable so
by Compactness is satisable. Thus T .
Theorem 13 If T is a rst-order theory with arbitrary large nite models, then T has an innite
model. If is a formula in L such that for every n < , T has a model A with [(A)[ n, then T
has a model B for which [(B)[ is innite.
Proof : For the rst part, add new constants to L and consider T

= T c
i
= c
j
[i < j < .
T

is nitely satisable so by Compactness is satisable. Thus there exists an innite model A [= T

and A[L [= T. For the second part, consider T

= T (c
i
) (c
j
) c
i
= c
j
[i < j < .
Theorem 14 The class of innite sets is not rst-order denable.
Proof : Assume it was by some sentence L so denes the class of nite sets. Now
consider T =
n
x(x = x)[n which is nitely satisable so by Compactness is
satisable, a contradiction.
Theorem 15 Let T be a theory in the language of elds with models of arbitrarily high nite
characteristics. Then T has a model which is a eld of characteristic 0.
Proof : Let T
f
be the theory of elds and consider T

= T T
f
p1 = 0[all primes p.
6.2 Realizing and Omitting Types
While our discussion in the previous subsection was restricted to theories, or sets of sentences in
L, similar results hold for formulae with free variables. Given a set (x) of open formulae, we can
ask whether there is a model A and tuple a A such that A [= (a). And so, we have types.
Denition 17 Let L be a rst-order language, A an L-structure with B A, b a sequence listing
the elements of B and a A. The complete type of a over B w.r.t A, denoted tp
A
(a/B), is the set
of all formulae (x, b) L(b) such that A [= (a, b). More generally, a set of formulae p(x) is a
complete type w.r.t A if it is the complete type of some tuple w.r.t some elementary extension of A.
When we talk of the complete type of a over X, we refer to everything that can be said about
the particular element a in terms of X. When we talk of a complete type over X, full stop, we
refer to everything we can say about some possible tuple in terms of X, whether the tuple lies in
A or only in some elementary extension of A.
14
Denition 18 A type over X w.r.t A is a subset of a complete type over X. A type is an n-type
if it has just n free variables. The sets of complete n-types over X w.r.t A are denoted S
n
(X; A)
and are known as the Stone spaces of A.
Denition 19 A type of a theory T is a set of formulae such that T x

(x) is consistent
for every nite (if T is complete, T x

(x) for all ). A complete type of T is a

maximal type of T.
For example, in L
N,+,<
the type (x) = 0 < x, 1 < x, x = 1 + 1, x < 1 + 1 + 1 is a 1-type in
A = 'N, +, <` of the number 2 (if were expanded to include all formulae in L
N,+,<
which hold
of 2 in A, then would be a complete type w.r.t A). If a type tp
A
(a/X), then is realized
by the tuple a A. If is not realized by any tuple in A, then A omits . Now consider Peano
arithmetic and let (x) = x = 0, x = 1, x = 1 + 1, x = 1 + 1 + 1, .... Clearly, is not realized in
the standard model N of PA but is realized in other models of PA? To answer this, we have the
following corollary from Compactness:
Theorem 16 is a type w.r.t A is nitely realized in A.
Proof : () Let be a nite subset of . Then as is a type w.r.t A, there is an
elementary extension B A and tuple b B such that B [= (b). But now B [= x(x) so
A [= x(x) and is nitely realized in A.
() Now suppose is an n-type nitely realized in A and consider T = eldiag(A) (c). T is
nitely satisable so by Compactness is satisable by some model B. As B [= eldiag(A), A B[L
by the elementary diagram lemma and the tuple b = (c
B
0
, ..., c
B
n
) realizes . Thus is a type w.r.t
A.
It can also be shown that if is nitely realized in A, then can be extended to a complete
type w.r.t A. But returning to our example, (x) = x = 0, x = 1, x = 1 + 1, x = 1 + 1 + 1, ...
is nitely realized in the standard model N of PA so by the above theorem, there must be some
elementary extension M N in which is realized. The model M is called a nonstandard model
of Peano arithmetic.
Here are some more examples/applications of realizing types:
Theorem 17 There exist nonstandard primes.
Proof : Let (x) be a formula in L
Z
expressing that x is a prime and consider the set of formulae
= (x), x > 0, x > 1, x > 2, .... As is nitely realized in A = 'Z, +, , 0, 1, <`, is a type
w.r.t A and is realized by some element Z in some elementary extension of A.
Example: (Dedekind cuts) A nice historic example, let us describe the complete 1-types of the
ordered rationals, i.e., the space S
1
(Q; A) for A = 'Q, <`. Consider a 1-type p(v) S
1
(Q; A).
p(v) is uniquely identied with a cut in the rationals: p =

qLp
[q < v]

Up
[v < q

] where
15
L
p
= q Q : q < v p and U
p
= q

Q : v < q

p (if v is a rational, we add [v = q

] for
some q

Q). The identication is unique since a cut is completely determined using only atomic
formulae and A has quantier elimination. S
1
(A) then contains the following 1-types:
(i) for all q Q, the unique types p
q
containing v = q;
(ii) p
+
where L
p
= Q and U
p
=
p

where L
p
= and U
p
= Q;
(iii) for all r R`Q, the unique types p
r
where L
p
= q Q : q < r and U
p
= q

Q : r < q

.
As there are as many types p S
1
(A) as there are real numbers, [S
1
(A)[ = 2

.
Omitting types is more dicult. As Professor Leo Harrington remarked in lecture: any idiot
can realize types but it takes a real model theorist to omit them. In analyzing when a theory T
has a model which omits particular types, the central idea will be that of T locally realizing a type.
Denition 20 Given a set of formulae L, T locally realizes i there exists a formula L
such that T x has a model (or if T is complete, T x) and for every formula ,
T x( ). We call a support of over T (or when , we say generates ) and call
a supported type (or principal type when has a generator). A formula is complete (for T) if
it generates a complete type.
The main results here explicate the relationship between local realization of a set of formulae
by T and the realization of in every model of T. One direction is simple (when T is complete)
while the other, characterized in the Omitting Types theorem, requires a construction similar to
that used by Henkin in proving Completeness/Compactness:
Theorem 18 Let T be a complete theory. T locally realizes for every model A [= T, A [= .
Proof : As T is complete and locally realizes , there is some formula L such that T x
and T x( ) for all . Now if A [= T, A [= x so A [= (a) for some a A. But we
also have A [= x( ) so A [= (a) (a) for all . Thus A [= (a) as desired.
Theorem 19 (Countable Omitting Types theorem) Let T be a consistent theory in a countable
language L and let be a set of formulae in L. T locally omits T has a countable model
which omits .
Proof : As in Henkins Completeness proof, add new constants c
i
[i < to L, forming L
+
,
and enumerate the sentences in L
+
as
i
[i < . Similar to the previous construction, we then
dene an increasing chain of theories T T
i
[i < in L
+
as follows:
Set T
0
= so T T
0
= T. As before, let T

i+1
= T
i

i+1
if T T
i

i+1
is consistent;
otherwise, set T

i+1
= T
i
. If
i+1
is of the form x(x), then take the earliest witness c
j
not already
used in T

i+1
and set T
i+1
= T

i+1
(c
j
); otherwise, put T
i+1
= T

i+1
. But there is now an
additional crucial step: given T
i
, write

T
i
as a sentence (c, d) with L and d the distinct
witnesses that occur in T
i
but not in c. Then T is consistent so T xy has a model
16
and as T locally omits , there is some formula such that T is consistent.
Take (c) and put it in T
i+1
.
If we then take the union T
+
= T

i<
T
i
, then by construction T
+
is consistent, maximal
and has the Henkin property so T
+
is a Hintikka set. It follows that the canonical interpretation
A of the set of atomic sentences in T
+
models T
+
so as T T
+
, A[L [= T. Further, for each tuple
of distinct witnesses c, there is a formula such that (c) T
+
. As every tuple in A is
named by these witnesses (consider sentences x(x = t) where t is a closed term), A[L omits
(note also that we constructed A by adding only countably many new constants c
i
to L so A[L
is a countable model).
While the Omitting Types theorem was presented with only one type , it is easily extended
to countably many types
i
[i < n. We simply run the same argument but at each incremen-
tal step, we witness a formula
i
for each
i
. However, the Omitting Types theorem fails
for sets of formulae with innitely many free variables as consider the complete theory DLO and
= x
0
< x
1
, x
1
< x
2
, x
2
< x
3
, .... Though DLO has no model which omits , DLO does locally
omit as if DLO (x
0
, ..., x
n
) is consistent, then DLO (x
0
, ..., x
n
) (x
n+1
< x
n+2
) is
consistent as well. Moreover, the Omitting Types theorem fails for uncountable languages as let
T = a
i
= a
j
[i < j <
1
be a theory in the language L(c, d) with constants c
i
[i <
1
d
i
[i <
and consider = x = d
i
[i < . Though T locally omits , no model of T omits as every
model of T has uncountably many elements.
But back to the case where we are working with countable language L and type with only
nitely many free variables, we can now combine the above theorems to give a necessary and
sucient condition for T to have a model omitting :
Theorem 20 Let T be a consistent theory in a countable language L. T has a countable model
which omits T has a complete extension which locally omits .
And so we now know how to realize and omit types. As we will see in the following section,
this will be particularly useful for systematically investigating the countable models of a complete
theory. For the time being, we will assume that L is a countable language.
17
7 Big and Small Countable Models
7.1 Atomic Models
We begin with the small countable models:
Denition 21 A is an atomic model if for every a A, tp
A
(a) is principal (i.e., there exists some
formula L which generates tp
A
(a) relative to Th(A) and A [= (a)). A is a prime model if A
is elementarily embedded in every model of Th(A).
Denition 22 a theory T is atomic i for every L such that T is consistent, is
generated by a complete formula in T (i.e., lies in some principal type of T).
Examples: nite models, the standard model of PA, DLO.
Theorem 21 (Existence Theorem for Atomic Models) Let T be a complete theory.
T has a countable atomic model T is atomic.
Proof : () Let A [= T be a countable atomic model and consider a formula L such that
T is consistent. As T is complete, T x so A [= (a) for some a A. But A is atomic so
tp
A
(a) is principal and A [= (a) (a) for a complete formula . Thus T so T is atomic.
() Assume T is atomic and consider the sets
n
=
i
(x
1
, ..., x
n
)[i < for n < where

i
are the complete formulae w.r.t T. If some formula (x
1
, ..., x
n
) L is consistent with T,
then as T is atomic, T
i
is consistent for some complete
i
. But then T locally
omits
n
so by the Omitting Types theorem, there is a countable model A [= T which omits
n
.
But then we have that for each a A, A [=
i
(a) for some complete formulae
i
. Thus T has a
countable atomic model A. [note: instead of requiring that T is atomic, we could have here used
the equivalent condition that for every n < , [S
n
(T)[ is at most countable (recall that S
n
(T) is
the set of complete n-types of T) as by the Omitting Types theorem, the at most countably many
non-principal complete types can be omitted]
Theorem 22 (Uniqueness Theorem for Atomic Models)
A, B countable atomic models and A B A

= B.
Proof : We show a back-and-forth system exists from A to B. In the base case, as A B, there
is an isomorphism f : '`
A
'`
B
. Now let 'a`
A
and 'b`
B
be nitely generated substructures
of the countable atomic models A, B respectively and assume that there exists an isomorphism
f : 'a`
A
'b`
B
with f(a) = b. This implies (A, a) (B, b) so a and b realize the same complete
n-type in A, B. We must show the forth step holds so choose c A not in a. As A is atomic,
tp
A
(ac) is principal so has a generator (x, y) L. But now tp
A
(a) = tp
B
(b) and A [= y(a, y)
so B [= y(b, y) and for some d B, B [= (b, d). The tuples ac and bd thus realize the same
complete n + 1-type in A, B so (A, ac) (B, bd) and there is an isomorphism f : 'ac`
A
'bd`
B
with f(ac) = bd as desired. An analogous argument works for the back step so as A

B, A

= B.

18
Theorem 23 A is a countable atomic model A is a prime model.
Proof : () Let A be a countable atomic model and let B [= Th(A). Add new constants to the
language L and list the elements of A as a
i
[i < . Using the argument in the previous proof, we
can nd elements b
i
[i < B such that (A, a
i
) (B, b
i
) for all i < . Then by the elementary
diagram lemma, the mapping f(a
i
) = b
i
is an elementary embedding of A into B.
() Assume A is not atomic so there is some a A such that tp
A
(a) is not principal. By the Omit-
ting Types theorem, there is B [= Th(A) which omits tp
A
(a). But then A cannot be elementarily
embedded in B as for some formula tp
A
(a), A [= (a) for some a A while B [= x(x). Thus
A is not a prime model. [note: a prime model must be countable as it is elementarily embedded in
every countable model of T]
Let us summarize: we are interested in the countable models of a complete theory T and the
rst appearance is made by the small atomic models, structures in which each element satises a
principal complete type. As T is complete, none of these types can be omitted in any countable
model of T so they are here to stay. The results stated above should then come as no surprise as,
intuitively, the unique atomic model of T - the model containing the bare minimum of what must
be included in any model of T - must be elementarily embedded in every countable model of the
theory. In a sense, prime models are like prime numbers: they cannot be factored into pieces but
are rather the raw, indestructible components of all countable models.
7.2 -Saturated Models
We now turn to the big countable models:
Denition 23 A is -saturated if for every nite subset X A, all complete 1-types over X w.r.t
A in L(X) consistent with Th(A) are realized by elements in A [note: this immediately extends to
all complete n-types]. A is a countable universal model if A is countable and every countable model
B A is elementarily embedded in A.
The restriction to nite subsets X A is important here as if X were allowed to be countable,
we have the type = x = a
1
, x = a
2
, x = a
3
, ... where a
i
[i < lists the elements in A. In this
case, the very notion of a countable -saturated model becomes impossible.
Examples: nite models, the ordered rationals. For a better example clearly distinguishing atomic
and -saturated models, consider a language with only constants c
i
[i < and the theory
T = c
i
= c
j
[i < j < . There are (up to isomorphism) countably many models of T as for
each n < , there is a model with exactly n elements that are non-constants. The atomic model
has zero non-constants while the saturated model has non-constants.
19
Theorem 24 (Existence Theorem for -Saturated Models) Let T be a complete theory.
T has a countable -saturated model for each n < , [S
n
(T)[ is countable.
Proof : () Let A [= T be a countable -saturated model so all complete n-types of T in L
are realized in A. As each tuple a A realizes only one complete type and A is countable, [S
n
(T)[
must be at most countable for all n < .
() The argument is similar to Henkins Completeness/Compactness proof. Without getting into
the details, we can construct a chain of theories T = T
0
T
1
T
2
... in an expanded language
L(c) such that the union of the chain T
+
=

i<
T
i
is a Hintikka set and the canonical model
A[L [= T realizes all complete types over nite subsets of A[L (see Chang & Keisler: 98-9). The
construction of a countable canonical model is possible as [S
n
(T)[ is at most countable for each
n < (so if L(d) is the language obtained by adding m < new constants d c, then since
the types
L(d)
(x
1
, ..., x
n
, d
1
, ..., d
m
) of T in L(d) are in one-to-one correspondence with the types

L
(x
1
, ..., x
n+m
) of T in L, there are at most countably many types
L(d)
as well).
Theorem 25 Let T is a complete theory with countably many nonisomorphic countable models.
Then T has a countable -saturated model.
Proof : Each complete n-type of T is realized in a countable model of T (the countability of
L is essential here) and there are only countably many nonisomorphic countable models of T so T
has only countably many complete n-types. By the above theorem, T has a countable -saturated
model.
Theorem 26 (Uniqueness Theorem for -Saturated Models)
A, B countable -saturated models and A B A

= B.
Proof : The proof is similar to the uniqueness proof for countable atomic models. The previous
proof was driven by the result that for A an atomic model and B an L-structure, if (A, a) (B, b)
and c A, there is an element d B such that (A, ac) (B, bd). Analogously, we now have that
for A an -saturated model and B an L-structure, if (A, a) (B, b) and d B, there is an element
c A such that (A, ac) (B, bd). This holds as tp
A
(a) = tp
B
(b) and the complete type tp
B
(d; b)
is also realized in A (though with parameter a) as A is -saturated. We can thus construct a
back-and-forth system from B to A.
Theorem 27 A is a countable -saturated model A is a countable universal model.
The proof is again similar to the earlier proof that all countable atomic models are prime.
However, the converse of this theorem now fails (see Chang & Keisler: 101 for a counterexample).
Intuitively, if A is a countable -saturated model, all countable models B A must be elementarily
embedded in A as A realizes all the complete types we can reasonably expect to be satised in a
countable model. Note that saturation is a very strong condition as not only does A realize all
n-types in the base language L (in which case we say A is weakly saturated) but also all n-types with
parameters dened on nite subsets of A. Also, if a complete theory T has a countable -saturated
model, then T has a countable atomic model, though the converse is not true.
20
7.3 -Categoricity
Now that we have examined atomic and -saturated countable models, we might also ask when
a complete theory T has (up to isomorphism) a unique countable model. In this section, we give
several characterizations for when this is the case. The exact theorem given below is part of the
Theorem of Engeler, Ryll-Nardzewski and Svenonius (see Hodges: 171-2 and Chang & Keisler:
101-3) but ignores the bits on automorphism groups.
Denition 24 A theory T is -categorical if T has exactly one countable model up to isomorphism
[note: such a theory must be complete]. A structure A is -categorical if Th(A) is -categorical.
Theorem 28 Let T be a complete theory in a countable language L. Then the following are
equivalent:
(a) T is -categorical;
(b) T has a countable model which is both atomic and -saturated;
(c) for each n < , every type in S
n
(T) is principal;
(d) for each n < , [S
n
(T)[ is nite;
(e) for each n < , there are only nitely many pairwise non-equivalent formulae (x
1
, ...x
n
) in
the language L mod T;
(f ) all models of T are atomic.
Proof : Following Chang & Keisler, we prove (a) (b) (c) (d) (e) (f) (a).
(a) (b): Let A [= T be the unique countable model of T (up to isomorphism). Then A is trivially
prime so A is atomic. Further, T has only countably many n-types so A is -saturated.
(b) (c): Since A is -saturated, every type in S
n
(T) is realized in A. But A is also atomic so
every type in S
n
(T) is principal.
(c) (d): Consider = (x
1
, ..., x
n
)[ is a complete formula in T for n < . As every
type in S
n
(T) is principal, T is inconsistent, so T is inconsistent for some nite subset
=
j
[1 < j < m by Compactness. Thus T (
1
...
m
) so T
1
...
m
.
But
j
are complete formulae (so generate principal types in T) and every element in every model
of T must satisfy one of them so [S
n
(T)[ m for some nite m.
(d) (e): Let

be the set of all types of T containing the formulas (x

1
, ..., x
n
) and
(x
1
, ..., x
n
) respectively. Then

implies T so as there are at most m <

types in S
n
(T), there are at most 2
m
pairwise non-equivalent formulae with n free variables in L
mod T.
(e) (f): Let A be a model of T and consider tp
A
(a) for some a A. As for each n < , there
are only nite non-equivalent formulae (mod T), the conjunction of the non-equivalent formulae in
tp
A
(a) (i.e.,
a
=
1
(a)
2
(a) ...
k
(a)) is itself a formula in L. But
a
generates tp
A
(a) so A
is atomic.
(f) (a): If all models of T are atomic, then since the countable atomic model of T is unique, T
is -categorical.
21
Examples of -categorical theories include DLO, atomless Boolean algebras, innite Abelian
groups with all elements of prime order, and the theory of an equivalence relation with innitely
many innite classes.
Interestingly, no complete theory has exactly two nonisomorphic countable models:
Theorem 29 (Vaught) Let T be a complete theory. Then either T is -categorical or T has 3
nonisomorphic countable models.
Proof : Assume T has exactly two nonisomorphic countable models A, B so A is atomic and
B is -saturated. Choose some tuple b B that does not realize a principal type in T and let
B

= (B, b). Since B is -saturated and there is a one-to-one correspondence between the types
in L and L(b), B

is also -saturated so letting T

= Th(B

), T

has an -saturated model (B

)
and hence an atomic model (C, b) [= T

. Now consider the reduct C = (C, b)[L [= T. First, C

is not atomic as the tuple in C named by b in L(b) does not realize a principal type. Second,
C is not -saturated: as T is not -categorical, neither is T

(both theories have innitely many

nonequivalent formulae) and since no model of T

is both atomic and -saturated, (C, b) is not

-saturated so neither is C. Thus C [= T is neither atomic nor -saturated so C

= A and C

= B
respectively.
8 The Innite Elevator of Lowenheim-Skolem
In the previous section, we explored what happens in the countable realm, categorizing the models
of a complete theory T with countable elements. We now broaden our analysis to the uncountable
case. Fortunately, here we have the theorems of Lowenheim-Skolem which tell us we can move up
and down with ease through the higher reaches of the innite. The Downward Lowenheim-Skolem
(DLS) theorem is presented rst and involves the construction of models by adding Skolem functions
to the language. The Upward Lowenheim-Skolem (ULS) theorem is then an easy consequence of
DLS and Compactness.
Disliking uncountable structures, Skolem proved that for every innite structure A of countable
signature, there is an elementarily equivalent countable substructure B A. Hence, uncountable
structures are redundant (and sometimes lead to counterintuitive results). An easy way to show
this is to start with some countable subdomain of dom(A) and close it o by adding new constants
so that the Tarski-Vaught criterion for elementary substructures holds. But instead of adding
constants, Skolem added functions:
Denition 25 A Skolemization of T is a theory T
+
T in L
+
L such that:
(a) if A [= T, there exists an expansion A
+
A such that A
+
[= T
+
;
(b) for every (x, y) L
+
, T
+
x(y(x, y) (x, t(x))) for some term t L
+
.
The terms t are called Skolem functions for T
+
.
Denition 26 A theory T is a Skolem theory (or has Skolem functions) if T is a Skolemization of
22
itself. A structure A has Skolem functions if Th(A) is a Skolem theory. If A [= T with T a Skolem
theory, 'X`
A
is called the Skolem hull of X.
Theorem 30 T is a Skolem theory T has the property of quantier elimination.
Proof : follows immediately from denitions (T x(y(x, y) (x, t(x))).
Theorem 31 Let A [= T with T a Skolem theory and X dom(A) such that 'X`
A
is non-empty.
Then 'X`
A
A.
Proof : Let B = 'X`
A
and consider some L and b B such that A [= y(b, y). As
T is a Skolem theory, A [= (b, t(b)) for some term t L and element t
A
(b) A. Thus by the
Tarski-Vaught criterion for elementary substructures, B A as desired.
By this last theorem, Skolem theories have the nice property that every non-empty substructure
of their models is an elementary substructure. But there are few natural Skolem theorems. Like
Hintikka sets (recall Henkins Completeness/Compactness proof), Skolem theories are generally
articially constructed:
Theorem 32 Let L be a rst-order language. Then there are a rst-order language L

L and
set of sentences L

such that:
(a) every L-structure A can be expanded to a model A

[= ;
(b) is a Skolem theory in L

;
(c) [L

[ = [L[.
Proof : For each formula (x, y) L, introduce the new function F
,x
and take
1
to be
the set of all sentences of the form x(y(x, y) (x, F
,x
)) in the new language L
1
. Now
consider an arbitrary non-empty L-structure A. Given (x, y) L and a A, if there is a b A
such that A [= (a, b), put F
A

,x
(a) = b; otherwise, set F
A

,x
(a) = a (it does not matter). The
expanded model A

1
is then a model of
1
. Iterating this expansion, we form the increasing chains
(L
n
: n < ) and (
n
: n < ). Now letting L

and be their unions, (a) and (b) clearly hold.

[L

[ [L[ +[L[ = [L[ so (c) holds as well.

Theorem 33 Every theory T L has a Skolemization T
+
L
+
with [L
+
[ = [L[.
As all the work has already been done in constructing the set , simply let T
+
= T . We
are now ready to prove DLS:
Theorem 34 (Downward Lowenheim-Skolem theorem) Let L be rst-order, A an L-structure,
X Dom(A) and a cardinal such that [L[ + [X[ [A[. Then there exists B with B A,
[B[ = , and X dom(B).
23
Proof : Consider Y X in dom(A) with [Y [ = and expand A to a model A

[= in L

.
Let B = 'Y `
A
[L. Then [B[ [Y [ +[L

[ = +[L[ = = [Y [ [B[. Also, 'Y `

A
A

since is
a Skolem theory so B A.
DLS thus tells us that we can move downward through the transnite hierarchy, bounded from
below by only the cardinality of the language L. It is now easy to show that we can freely move
upward as well:
Theorem 35 (Upward Lowenheim-Skolem theorem) Let L be rst-order, A an innite L-structure,
both of cardinality . Then there exists B with A B and [B[ = .
Proof : Introduce new constants and let T = eldiag(A) c
i
= c
j
[i < j < in L(c). As A is
innite, T is nitely satisable so by Compactness, there is a model B

[= T. Let B = B

[L. Since
B

[= eldiag(A), by the elementary diagram lemma we have an elementary embedding e : A B

so A B. Further, B must have at least elements so we can invoke DLS to ensure [B[ = .
9 Preservation Theorems
We have already seen some examples of preservation in our discussion of dierent mappings in
Section 2: f preserves if A [= (a) B [= (fa) for every a A. In this section, we explore
several more important preservation results, i.e., that certain mappings preserve all formulae with
certain syntactic features. But preservation phenomena (applied to sets of sentences) also work in
the other direction: if a theory is preserved under certain mappings, we can also infer the syntactic
form of the theory (up to equivalence).
Theorem 36 embeddings preserve
1
formulae.
Proof : From before, we have that an embedding f preserves all literals. By induction on the
complexity of formulae, this easily extends to all Boolean combinations of atomic formulae. Now
consider when = x and assume f preserves , i.e., A [= (a) B [= (fa). Now if A [= ,
then A [= (c) for some c so B [= (fc) and B [= as desired.
Theorem 37 a theory T is preserved under submodels T is an
1
theory.
Proof : () Let B be a submodel of A. Then there exists an embedding f : B A so from
the above theorem, B [= A [= where is an
1
formula. But now A [= B [= and
as is an
1
formula,
1
formulae are preserved in substructures.
() Assume T is preserved under submodels and let = set of sentences in L equivalent to
1
sentences. Consider A [= T and B such that A [= B [= where . We must show B [= T
(as then T is equivalent to an
1
theory). Set T

= T diag(B) in L(b) and let

1
(b), ...,
n
(b) be
a nite set of formulae from diag(B). Then B [= x(
1
(x), ...,
n
(x)) so given the contrapositive of
the above conditional, A [= x(
1
(x), ...,
n
(x)). As A [= T, T
1
, ...,
n
is consistent so as T

is
24
nitely satisable, T

is satisable by Compactness. Let C [= T

. Then C [= T and B C[L so as

T is preserved under submodels, B [= T.
Theorem 38 a theory T is preserved in unions of chains T is an
2
theory.
Proof : () Let (A
i
: i < ) be a chain of L-structures (i.e., A
0
... A
n
...) and a A
0
such that A
i
[= yx(y, x, a) for all i < . Put B =

i<
A
i
. We must show B [= yx(y, x, a).
If b B, then b A
i
for some i < so by assumption, A
i
[= x(b, x, a). Now as
1
formulas are
preserved under embeddings and A
i
B, B [= yx(y, x, a) as desired.
() Assume T is preserved in unions of chains and let = set of sentences in L equivalent
to
2
sentences. Consider A [= T and B such that A [= B [= where . Set
T

= Th(A) diag

1
(B) where Th(A) is the complete theory of A in L and diag

1
(B) is the
set of universal sentences of L(b) in eldiag(B). By the contrapositive of the above conditional,
B [= A [= where is an
2
sentence in L so T

is nitely satisable and by Compactness,

there exists A

[= T

(note that A

[L A and B A

[L). Now add new constants and set

T

= diag(A

) eldiag(B). As we also have A

[= diag

1
(B), every existential sentence in L(b) true
in A

is true in Mod(diag(B)) so T

is satisable by Compactness. Let B

[= T

so B B

[L and
A

[L B

[L.
Iterating this construction, we can form the chain (we write A

for A

[L, B

for B

[L, etc.):
B A

... and let C be the union of this chain. Now C is the union of
(A
i
: i < ) and as each A
i
[= T and T is preserved in unions of chains, C [= T. But C is also the
union of the elementary chain B B

... so by the Tarski-Vaught theorem on unions of

elementary chains, B [= T as desired.
The latter part of this proof combines several powerful model-theoretic techniques. First, note
that by deliberately considering particular kinds of theories, we can ensure the models given by
Compactness have specic features (we have already used this technique extensively in Section 6
and in the proof of ULS). Given a nitely consistent set of sentences, Compactness simply tells
us that a model exists in which all the sentences are satised. But by using diagrams (or other
carefully chosen sentences), we can exercise some control. In T

, the component Th(A) ensures that

the model A

[L is elementarily equivalent to A; the component diag

1
(B) ensures that B A

[L
(and a bit more); using eldiag(B) in T

ensures the stronger correspondence B B

[L. Second,
the proof also uses the neat trick of alternating chains - we create an increasing chain of models
in which alternating components have special properties. As these properties (namely, elementary
equivalence and modeling T) are preserved in the union of the chain, transnite induction allows
us to nd a model in which both properties hold.
Other preservation results not proven here include:
(i)
+
1
formulae are preserved under homomorphisms
(an
+
1
formula is an
1
formula in which negation does not occur);
(ii) all positive formulae are preserved under surjective homomorphisms
(a positive formula is one in which negation does not occur);
25
(iii) all formulae are preserved under isomorphisms (surjective embeddings).
Aside from some of our other preservation results in this section, there is now the following preser-
vation hierarchy: homomorphisms preserve
+
1
formulae and if they are surjective, all positive
formulae; embeddings preserve
1
formulae and if they are elementary, all formulae.
10 Amalgamation Theorems
10.1 Merging Models
In many instances, it will be useful to amalgamate, or join, several models together into a larger
structure. There are many ways to do this and many interesting proofs that use such constructions.
We begin with the father theorem:
Theorem 39 (Elementary Amalgamation theorem) Let L be rst-order and let B, C be L-structures.
If there exists a B and c C with (B, a) (C, c), there is an elementary extension D B and
an elementary embedding g : C D such that gc = a.
Proof : Without loss of generality, assume a = c (we consider an isomorphic copy of C if
necessary) and consider T = eldiag(B) eldiag(C). It must be shown that T is nitely satisable.
As in any nite subset T

T there are only nitely many sentences in eldiag(C), we can take

their conjunction (a, d) with L and d elements in C which are not in a. If T

is unsatisable,
eldiag(B) (a, d) so as the elements d are not in B, eldiag(B) x (a, x). But now
(B, a) [= x (a, x) so by assumption, (C, a) [= x (a, x), a contradiction. By Compactness
then, there exists a model D

[= T and by two applications of the elementary diagram lemma, the

reduct D = D

[L is an elementary extension of both B and C.

D
B C
'a`
B
d
d
d
d
d

d
d
d
d
d

g
f
Figure 3: Elementary Amalgamation (Hodges: 135)
The elementary amalgamation theorem is depicted in the above diagram. On the right side,
the condition (B, a) (C, c) ensures, by the diagram lemma, that there is a unique embedding
f : 'a`
B
C with fa = c. The right side of the diagram is essentially still symmetric to the left as
there is a C

= C with 'a`
B
C

and C

D (so there is really not that much extra commotion).

The idea here is that if a substructure is contained in two (or more) distinct elementarily equivalent
structures, the structures can be elementarily embedded into a larger structure under mappings
26
that agree on the common substructure. In the special case where a is empty, the elementary amal-
gamation theorem tells us that elementarily equivalent structures can be elementarily embedded in
a larger structure.
A nice application of the elementary amalgamation theorem is that all of the complete types
with respect to a structure A can be realized in some elementary extension of A:
Theorem 40 Let L be rst-order and let A be an L-structure. There is an elementary extension
B A such that every type over dom(A) w.r.t A is realized in B.
Proof : Consider the set of complete types p
i
: i < over dom(A) w.r.t A. With each
p
k
= tp
A
k
(a
k
/dom(A)) for some elementary extension A
k
A and a
k
A
k
, we can construct an
increasing chain of models B
0
B
1
B
2
... as follows: let B
0
= A and form B
i
by elementarily
amalgamating B
i1
and A
i
. Take unions at limit ordinals and let B =

i<
B
i
.
An amalgam is said to be strong if it has the minimum-overlap property, i.e., the overlap of B
and C

= gC in D is precisely 'a`
B
. It can be shown (see Hodges: 139-140) that a necessary
condition for a strong amalgam is that the common substructure 'a`
B
is algebraically closed in B
(or C by symmetry). Note that if X domB, an element b B is algebraic over X if there is a
rst-order formula (x, y) and a X such that B [= (b, a)
n
x(x, a) for some nite n. Letting
acl
B
(X) represent the set of all elements in B algebraic over X, a strong amalgam is then possible
when 'a`
B
= acl
B
('a`
B
).
We now give two of the many variations on amalgamations:
Theorem 41 (Existential Amalgamation theorem) Let B, C be L-structures. If there exists a B
and a homomorphism f : 'a`
B
C such that (C, fa)
1
(B, a) (i.e., (C, fa) [= (B, a) [=
for all rst-order
1
sentences L), there is an elementary extension D B and an elementary
embedding g : C D such that gfa = a.
Proof : Note that the condition (C, fa)
1
(B, a) implies that f is an embedding so without
loss of generality, assume f is the identity on 'a`
B
. The proof of the elementary amalgamation
theorem now goes through as before as at the point where (B, a) [= x (a, x), this new condition
also implies that (C, a) [= x (a, x).
When a is empty, the existential amalgamation theorem says that if C
1
B, C is embeddable in
some elementary extension of B. Existential amalgamation thus trims the fat from the elementary
case, showing us that the weaker condition
1
, rather than full-blown elementary equivalence,
ensures that the structures B and C can be amalgamated. There are numerous interesting corol-
laries/applications of the existential amalgamation theorem and related amalgamation theorems,
such as preservation results and interpolation theorems (see Hodges 5.4/5.5).
27
Our second amalgamation theorem will be useful in proving the Craig interpolation theorem in
Section 11:
Theorem 42 Let B, C be L
1
, L
2
-structures respectively and let L
0
= L
1
L
2
and L = L
1
L
2
.
If there exists a B, C such that (B[L
0
, a) (C[L
0
, a), there is an L-structure D such that
B D[L
1
and there is an elementary embedding g : C D[L
2
with ga = a.
Proof : Using the elementary amalgamation proof with T = eldiag(B) eldiag(C[L
0
), we
obtain the weaker result that there is an L
1
-structure D B and an elementary embedding
g : C[L
0
D[L
0
. Though this is insucient to prove the theorem at hand, it allows us to build
chains of elementary extensions of B and C such that the L
0
-reducts of the unions are isomorphic,
i.e.,

B
i
[L
0

=

C
i
[L
0
(see Hodges: 147-8 for details). So as the union

B
i
is an L
1
-structure
and

C
i
is an L
2
-structure, we can expand

B
i
(or

C
i
by symmetry) to an L-structure D
[note: expanding a structure involves adding new properties and introducing new names on a xed
domain of elements; this diers from extending a structure where the universe itself is enlarged].
The Tarski-Vaught theorem on unions of elementary substructures does the rest.
10.2 Frasse Limits
One of the more remarkable instances of amalgamation is the Frasse limit. Frasses ingenious
idea was that given a class of nite structures having various properties, we can amalgamate them
together to form a limit structure. For example (and one we will use throughout this subsection),
the limit of the class of nite linear orderings is the ordered rationals 'Q, <`. Formal proofs of the
existence and uniqueness of Frasse limits can be found in Hodges (161-4).
The starting point is a class K of nitely generated structures. K is called the age of some
structure if K is non-empty and has the following properties:
Hereditary property (HP): If A K and B is a nitely generated substructure of A (i.e.,
B = 'a`
A
for nite a), then B is isomorphic to some structure in K
Joint embedding property (JEP): If A,B K, then there exists a C K and embeddings
f : A C and g : B C (an embedding from A C is an isomophism from A to some
substructure of C).
In addition, the class of nite linear orderings has the following important property:
Amalgamation property (AP): If A,B,C K and e : A B and f : A C are embeddings,
then there is a D K and embeddings g : B D and h : C D such that ge = hf
By the AP, the countable limit of the class K of nite orderings must be a dense linear ordering
without endpoints so K tends to the rationals rather than, say, the integers or natural numbers
(though K is still the age of both 'Q, <` and 'Z, <`).
28
Theorem 43 (Frasses Theorem) Let L be a countable language, K a non-empty nite/countable
collection of nitely generated L-structures which has HP, JEP and AP. Then there is a unique
(up to isomorphism) L-structure D of cardinality such that K is the age of D and D is
homogeneous.
Denition 27 A structure D is homogeneous if every isomorphism between nitely generated sub-
structures of D extends to an isomorphism from D D ( i.e., an automorphism of D).
The structure D in Frasses Theorem is what I, following Hodges, have been calling the Frasse
limit (it is also known as the universal homogeneous structure of age K). As already mentioned,
when K is the class of nite orderings, D is the ordered rationals. Frasses Theorem thus says that
the theory of dense linear orderings without endpoints is -categorical. Other interesting examples
of Frasse limits are the countable atomless Boolean algebra (where K is the class of nite Boolean
algebras) and the celebrated random graph (where K is the class of all nite graphs).
But what exactly does it mean to say 'Q, <` is homogeneous ? Well if every isomorphism
between substructures of the ordered rationals extends to an automorphism, then when we take a
peek at some local region of 'Q, <` and it looks the same, it really is. By contrast, consider an
ordering of the natural numbers 'N, <`. As only the zero element is a lower bound of all other
elements and the naturals are not dense (this creates a problem as the substructures A

= 0 < 1
and A

= 0 < 2 are isomorphic but any isomorphism between them can clearly not be extended to
an automorphism of 'N, <`), the ordered naturals are not homogeneous. Homogeneity is discussed
further in Section 13.
11 Syntactic Theorems
This section presents several interesting theorems dealing with language.
Theorem 44 (Craig Interpolation theorem) Let , be sentences in L such that . Then
there is a sentence such that:
(a) and ;
(b) every symbol (excluding identity) that occurs in also occurs in both and .
The sentence is the Craig interpolant of and .
Proof : We give a constructive argument similar to Henkins Completeness/Compactness proof
and the proof of the Omitting Types theorem but the focus is now on an inseparable pair of theories
rather than on a consistent theory. Given T
1
L
1
and T
2
L
2
, a sentence L
0
= L
1
L
2
separates T
1
and T
2
if T
1
while T
2
. If no such exists, T
1
, T
2
are inseparable.
Now assuming there is no Craig interpolant of L
1
and L
2
, we prove by
constructing a model of . Let L
0
= L
1
L
2
and L = L
1
L
2
and add new constants C so
L

i
= L
i
C. We construct a pair of maximally consistent inseparable theories T and U in L

1
, L

2
respectively as follows:
29
Enumerate the sentences in L

1
as
i
with
1
= and the sentences in L

2
as
i
with

1
= . Letting T
1
= and U
1
= , the theories T
1
, U
1
are inseparable as if not, i.e.,
and for some L

0
, the sentence x(x) L
0
is a Craig interpolant of and . Dene
increasing chains of theories T
i
and U
i
where we now only add
n
to T
n
if T
n

n
and U
n
are inseparable (and similarly for
n
and U
n
). As usual, we witness existential claims. Finally, put
T and U as the unions of the respective chains T
i
and U
i
so T, U are inseparable.
T is clearly consistent as if not, everything follows from T so T is separable from U. It can also
be shown that T is maximally consistent and similarly for U (see Chang & Keisler: 86) so consider
the canonical interpretations B [= T and C [= U. We must have B[L
0
C[L
0
as if not, there
is a sentence L
0
such that B [= while C [= , contradicting the separability and maximal
consistency of T and U. So by the syntactic amalgamation theorem in Section 10, there is an
L-structure D such that B D[L
1
and there is an elementary embedding g : C D[L
2
. Thus
D [= T U so D [= as desired.
The notion of inseparable theories provides us with another valuable construction technique. In
such previous arguments as Henkins Completeness proof, we restricted our attention to a single
chain of theories, adding constants until we arrived at a maximally consistent theory (the union
of the chain) with certain built-in properties. In the proof of Craigs theorem, we not only build
maximally consistent theories T and U (as before) but by ensuring that T, U are inseparable, the
construction also ensures that the theories are consistent with each other, i.e., T U is consistent.
When the language L has no constant or function symbols, there is the following strengthening
of Craigs result:
Theorem 45 (Lyndon Interpolation theorem) Let , be sentences in a language L with no
function or constant symbols such that . Then there is a sentence such that:
(a) and ;
(b) every relation symbol (excluding identity) that occurs positively in also occurs positively in
both and ;
(c) every relation symbol (excluding identity) that occurs negatively in also occurs negatively
in both and .
Finally, here is a nice application of Craigs Theorem:
Theorem 46 (Robinson Consistency Theorem) Let L
1
, L
2
be languages and let L = L
1
L
2
. If T
is a complete theory in L and T
1
T, T
2
T are consistent theories in L
1
, L
2
respectively, then
T
1
T
2
is consistent in L
1
L
2
.
Proof : Assume T
1
T
2
is inconsistent. Then by Compactness, there are nite subsets
1
T
1
,

2
T
2
with
1

2
inconsistent. Let
1
,
2
be the conjunctions of
1
,
2
respectively. Then

1

2
so by Craigs theorem, there is a sentence L such that
1
and
2
. Now
since T
1
and T
1
is consistent, T
1
so T . By analogous reasoning with T
2
, T ,
contradicting the completeness of T.
30
12 Model Completeness
In this section, we discuss existentially-closed structures and Abraham Robinsons related notions of
model-completeness and model companions. While model-completeness/companions are important
tools for the model theory of algebra, we will only dene these notions and present some basic
results, staying away from their algebraic applications.
Denition 28 A [= T is existentially closed (e.c.) i for every model B [= T with A B, A
1
B
(i.e., any existential formula x(x, a) which holds in B with a A also holds in A).
Examples of e.c. structures are the algebraically closed elds and Frasse limits. Given a
2
theory T, there is also a general method for constructing an e.c. extension of a model of T:
Theorem 47 let K be the class of all models of
2
theory T and let A [= T. Then there exists an
e.c. structure B K s.t. A B.
Proof : Enumerate as (
i
, a
i
)
i<
all pairs (, a) where is an
1
formula in L and a A. Now
dene a chain of structures A
0
A
1
... where A = A
0
and if there exists C A
i
with C K
and C [=
i
(a
i
), then A
i+1
= C; else A
i+1
= A
i
. Letting A

i<
A
i
, then A

K (since T is

2
, T is preserved in unions of chains) and all
1
formulas in L with parameters from A are realized
in A

. Now dene a second chain A A

(A

... (i.e., repeat the rst chain construction

times). Then B =

n<
A
(n)
K and B is existentially-closed. To see this, suppose is
1
,
b B and C [= (b) for some C K which extends B. Since b is nite, b A
(n)
for some n < .
But then A
(n+1)
[= (b) so as
1
formulae are preserved under embeddings, B [= (b) as required.
Note that if [[L[[ , we can ensure that the e.c. structure B found in the above proof has
cardinality [B[ (in constructing the rst chain, we use DLS to nd extensions C of cardinality
). Also (as we prove below), in cases where a theory T is model-complete (such as when T is a
Skolem theory), all models of T turn out to be e.c. structures.
Denition 29 T is model-complete i for all models A, B [= T, if A B then A B (i.e., all
embeddings between models of T are elementary).
Neither completeness nor model-completeness implies the other. The theory DLO with end-
points and Th(N, <) are both complete but not model-complete. The theory of algebraically closed
elds is model-complete but not complete. Nonetheless, model-completeness combined with some
additional properties of a theory does imply completeness, as in the following theorem:
Theorem 48 Let T be model-complete.
(i) if any two models of T are isomorphically embedded in a third model, then T is complete.
(ii) if T has a model which is isomorphically embedded in every model of T, then T is complete.
31
Proof : both (i) and (ii) follow from the fact that isomorphic embeddings are elementary when
T is model-complete [note: (ii) equivalently says that a model-complete theory with a prime model
is complete].
Theorem 49 If T is model-complete, then T is equivalent to a
2
theory.
Proof : By Tarski-Vaught theorem on unions of chains, T is preserved in unions of chains.
Theorem 50 The following are equivalent:
(a) T is model complete;
(b) for every A [= T, T diag(A) is complete in L
A
;
(c) every model of T is existentially-closed (Robinsons test);
(d) for every
1
, there is
1
s.t. T .
Proof : We prove (a) (b) (c) (d) (a).
(a) (b): T

= T diag(A) has the same models as eldiag(A) so all models of T

are elementarily
equivalent in L
A
.
(b) (c): if A B and B [= T then B [= T

and since T

is complete in L
A
, A
1
B.
(c) (d): consider (x
1
, ..., x
n
)
1
. Add c
1
, ..., c
n
to L and let =
1
sentences in L
c
s.t.
T . Let A [= T and consider T

= T diag(A) in L
A
. If T

is inconsistent,
T for some nite conjunction of formulae diag(A). But then T x(x)
in L
c
so T x(x), a contradiction since x(x) and A [= . Let B [= T

so A B
and B [= T. By assumption, A
1
B so holds in A and since A was arbitrary, T . But
now by Compactness, there is a nite conjunction s.t. T so T as desired.
(d) (a): rst note that (d) implies that for every formula L, T for some
1
.
Now consider A, B [= T with A B and B [= (a) for some a A. Then B [= (a) for some

1
and since
1
formulae are preserved in substructures, A [= (a) so A [= (a) (and similarly
for ).
We can actually get the stronger result: if T has only innite models, [[L[[ and A
1
B for
any A, B [= T of power with A B (i.e., all models of power are e.c.), then T is model-complete.
This gives us a nice criterion for model-completeness:
Theorem 51 (Lindstroms theorem) Let T be a
2
theory in countable L which has only innite
models and is -categorical for some innite . Then T is model-complete.
Proof : starting with some model A [= T of power , we can construct an e.c. model of T
of power as above. But now since T is -categorical, all models of power are e.c. so T is
model-complete.
32
Alternatively, to show T is model-complete we need only prove that T admits quantier elimina-
tion as then condition (d) in Theorem 50 is clearly satised. But the reverse argument also holds.
A particularly useful feature of model-completeness is that it provides an alternative route (recall
Section 5) to showing a theory has quantier elimination:
Theorem 52 T has quantier elimination i T is model complete and T

- the set of all

1
sentences of L s.t. T - has the amalgamation property (see Section 10.2).
Finally, a few words on companionship:
Denition 30 a theory U in L is a model companion of theory T if: (i) U is model-complete, (ii)
every model of T has an extension which is a model of U, (iii) every model of U has an extension
which is a model of T. We say T is companionable.
Theorem 53 Let T be an
2
theory in L.
(i) T is companionable i the class of e.c. models of T is axiomatizable by a theory in L.
(ii) If T is companionable, then up to equivalence of theories, its model companion U is unique
and is the theory of the class of e.c. models of T.
13 Further Topics
13.1 Saturation & Homogeneity
In Section 7.2, we discussed the -saturated models of a theory (models which realize all types over
nite subsets of their domain) and countable universal models (models in which every other count-
able elementarily equivalent model is elementarily embeddable). Here, saturation and universality
are generalized from the countable case and we also introduce the notion of homogeneity.
Denition 31 A is -saturated i for every set X of elements of A, if [X[ < then all complete
1-types over X w.r.t. A are realized in A. A is saturated i A is [A[-saturated.
Denition 32 A is -homogeneous i for every pair of sequences a, b A of length < and
c A, if (A, a) (A, b) then there is some d A s.t. (A, ac) (A, bd). A is homogeneous i A is
[A[-homogeneous.
Denition 33 A is -universal i any structure B of cardinality < with B A is elementarily
embeddable in A.
Our rst two theorems can be proven using back-and-forth constructions akin to those from
Theorems 22/26 above:
Theorem 54 A is countable and atomic A is countable and -homogeneous.
Theorem 55 A is -saturated A is -homogeneous. [note: in fact, it can be shown that
-saturation = -homogeneity + -universality.]
33
As with building e.c. structures in Section 12, it is a fairly simple matter to construct countable
-homogeneous (i.e., countably homogeneous) elementary extensions of countable models:
Theorem 56 every countable model has a countably homogeneous elementary extension.
Proof : Consider a countable model A and for any two nite sequence a
1
, ..., a
n+1
and b
1
, ..., b
n
in A s.t. (A, a
1
...a
n
) (A, b
1
...b
n
), introduce a new constant c
i
and type
i
(b
1
...b
n
, x) in L
b
cor-
responding to the set of formulae s.t. A [=
i
(a
1
...a
n+1
). Let T = eldiag(A)
i<
(b
1
...b
n
, c).
Then if A

[= T (we use a standard Compactness argument here), A

[L A and for all nite

sequences a
1
, ..., a
n+1
and b
1
, ..., b
n
in A s.t. (A, a
1
...a
n
) (A, b
1
...b
n
), there exists a b
n+1
A

s.t. (A

, a
1
...a
n+1
) (A

, b
1
...b
n+1
). Repeating this procedure, we form the elementary chain
A A

... and letting B =

n
A
(n)
, it is easy to verify that B is a countably homoge-
neous elementary extension of A.
Now that we know how to construct them, here are some simple properties of countably homo-
geneous models (the rst two theorems can be proven with a standard back-and-forth argument):
Theorem 57 If A B and A, B are both (countably) homogeneous, of the same cardinality, and
realize the same n-types over for each n < , then A

= B.
Theorem 58 Let A be a countably homogeneous and let a
1
, ..., a
n
and b
1
, ..., b
n
be sequences s.t.
(A, a
1
, ..., a
n
) (A, b
1
, ..., b
n
). Then there is an automorphism f : A A with f(a
i
) = b
i
.
Theorem 59 Let (A
i
: i < ) be an elementary chain of countably homogeneous models.
(i) B =

i<
A
i
is countably homogeneous.
(ii) If A
i

= A
j
for all 1 < i < j < , B

= A
i
for each i.
13.2 Two-Cardinal Theorems
As implied by the Lowenheim-Skolem theorems, a theory T with an innite model in a countable
language L cannot distinguish between innite cardinals as T will have a model in each innite
power. Nonetheless, model theorists have been interested in whether theories can distinguish be-
tween pairs of innite cardinals in another sense: given a theory T, is there a model A [= T with a
one-place relation V dened on A s.t. [A[ = and [V [ = ? If so, we say A is a (, ) model and T
admits the pair of cardinals (, ). Alternatively, a two-cardinal theorem tells us that under certain
conditions, there exists a model A [= T in which the -denable subsets (A),(A) dened by ,
have dierent cardinalities (note: if is x = x and is V , this collapses to our rst denition with
= ).
Theorem 60 Let T be a theory in a countable language and let , , be innite cardinals.
(i) If T admits (, ), then T admits (, ) for all s.t. .
(ii) If T admits (, ), then T admits all (, ).
34
Proof : (i) follows immediately from DLS; (ii) follows from (i) since T admits (, ) and we can
introduce a new function F in L and extend T by a single sentence which says F is a one-to-one
mapping from A V .
Theorem 61 (Vaughts Two-Cardinal theorem) Let L be countable, (x), (x) L and T a theory.
Then the following are equivalent:
(a) T has a model A in which [(A)[ but [(A)[ =
1
;
(b) T has a model A in which [(A)[ < [(A)[ ;
(c) T has models A, B s.t. B A and (A) = (B) but (A) = (B).
Denition 34 a Vaught pair for is a pair of structures A, B s.t. B A, B = A and
[(A)[ = [(B)[ = for innite .
A formula is two-cardinal for T if there is A [= T s.t. [A[ = [(A)[; otherwise it is one-cardinal
for T. A theory T is two-cardinal if there is a two-cardinal formula for T s.t. [(A)[ is innite
in every model of T; otherwise T is one-cardinal. For a countable and complete theory T, the
direction (c) (b) in Vaughts theorem says that if has a Vaught pair of models of T, then is
two-cardinal for T and T is a two-cardinal theory.
13.3 Indiscernibles
Given a set X, let [X]
n
denote the set of all subsets of X with n elements. If X is simply ordered
by <, we can take [X]
n
to be the set of increasing sequences x
1
< ... < x
n
from X. We will need
the following combinatorial result:
Theorem 62 (Ramseys theorem) Let I be an innite set and let n . Suppose that [I]
n
=
A
0
A
1
. Then there is an innite subset J I s.t. either [J]
n
A
0
or [J]
n
A
i
.
In the case where n = 1, Ramseys theorem amounts to the pigeon-hole principle with two holes.
When n = 2, Ramseys theorem tells us that any innite graph whose edges have been colored with
two colors has an innite monochromatic subgraph. And so on.
Denition 35 Let A be a model in L and let X dom(A) be simply ordered by <. Then X is a
set of indiscernibles in A if for all n and nite sequences x
1
< ... < x
n
and y
1
< ... < y
n
from X,
(A, x
1
...x
n
) (A, y
1
...y
n
).
In other words, any nite sequences x
1
< ... < x
n
and y
1
< ... < y
n
from X cannot be distin-
guished by any formula (v
1
, ..., v
n
) L. Here is a sucient condition for a set to be indiscernible:
Theorem 63 Let 'X, <` dom(A) and for any sequences x
1
< ... < x
n
and y
1
< ... < y
n
from
X, suppose there is an automorphism f : A A s.t. f(x
i
) = y
i
for all i. Then X is a set of
indiscernibles in A.
Proof : follows immediately from the elementary diagram lemma.
35
For example, if A is an algebraically closed eld of characteristic zero, then X = algebraically
independent elements in X ordered by < is a set of indiscernibles in A. If A is a Boolean algebra,
then the set of all atoms X dom(A) is a set of indiscernibles. More generally, indiscernibles
of any order type must exist in models of any theory T with innite models, as outlined in the
following two results:
Lemma 6 Let T have innite models and let L

= L c
n
: n .
Then T

= T (c
i
1
, ..., c
in
) (c
j
1
, ..., c
jn
) : L, n , i
1
< ... < i
n
, j
1
< ... < j
n
c
1
= c
2

is consistent in L

.
Proof : let A be an innite model of T and I a countable innite well-ordered subset of dom(A).
We show by induction that for any nite subset T

, there is an innite subset J

I
s.t. (A, j
i
)
i
[= for j
i
J

. Assume this holds for some and consider some new formula
(v
1
, ..., v
m
) L. We divide [J

]
m
into two parts: A
0
= j
1
< ... < j
m
: j
i
J

A [= (j
1
, ..., j
m
)
and A
1
= j
1
< ... < j
m
: j
i
J

A [= (j
1
, ..., j
m
). Then [J

]
m
= A
0
A
1
so by Ramseys
theorem, there is an innite subset K J

s.t. either [K]

m
A
0
or [K]
m
A
1
. In either case,
(A, k
i
)
i
[= (c
i
1
, ..., c
im
) (c
j
1
, ..., c
jm
) so the induction is complete.
Theorem 64 Let T have innite models and let 'X, <` be any simply ordered set. Then there
exists A [= T with X dom(A) s.t. X is a set of indiscernibles in A.
Proof : let L

= L c
x
: x X and T

= T (c
x
1
, ..., c
xn
) (c
y
1
, ..., c
yn
) : L, n ,
x
1
< ... < x
n
, y
1
< ... < y
n
X c
x
1
= c
x
2
: x
1
= x
2
X. Then T

is nitely consistent as
we can ordermorphically embed every nite subset of X into the set K from the above proof. So
by Compactness, let A

[= T

. Then interpreting c
x
as x X, A = A

[L [= T, X A and X is a
set of indiscernibles in A.