You are on page 1of 31

Many-Sorted First-Order Model Theory

Lecture 8

25th June, 2020

1 / 31
Applications of ultraproducts

2 / 31
Fancy models via ultraproducts
Example 1 (Growing chains)
Let Cn be the n-element chain. Let U be a non-principal filter on ω. Then,
Q
n<ω Cn /U is an infinite chain consisting of a copy of ω at the bottom, a
copy of dual ω at the top, and uncountably many copies of Z in between.

Example 2 (Non-standard natural numbers)


Let U be a non-principal ultrafilter on ω. Consider Nω /U. It is
elementarily equivalent to N. Take the element (1, 2, 3, 4, . . . )/U. It is
strictly greater than any standard natural number.

Example 3 (Infinitesimals)
Let U be a non-principal ultrafilter on ω. Consider Rω /U. It is
elementarily equivalent to R. Take the element (1, 12 , 31 , 14 , . . . )/U. It is
strictly greater than 0, yet strictly smaller than any standard real number.

3 / 31
Fancy models via ultraproducts
Example 1 (Growing chains)
Let Cn be the n-element chain. Let U be a non-principal filter on ω. Then,
Q
n<ω Cn /U is an infinite chain consisting of a copy of ω at the bottom, a
copy of dual ω at the top, and uncountably many copies of Z in between.

Example 2 (Non-standard natural numbers)


Let U be a non-principal ultrafilter on ω. Consider Nω /U. It is
elementarily equivalent to N. Take the element (1, 2, 3, 4, . . . )/U. It is
strictly greater than any standard natural number.

Example 3 (Infinitesimals)
Let U be a non-principal ultrafilter on ω. Consider Rω /U. It is
elementarily equivalent to R. Take the element (1, 12 , 31 , 14 , . . . )/U. It is
strictly greater than 0, yet strictly smaller than any standard real number.

4 / 31
Fancy models via ultraproducts
Example 1 (Growing chains)
Let Cn be the n-element chain. Let U be a non-principal filter on ω. Then,
Q
n<ω Cn /U is an infinite chain consisting of a copy of ω at the bottom, a
copy of dual ω at the top, and uncountably many copies of Z in between.

Example 2 (Non-standard natural numbers)


Let U be a non-principal ultrafilter on ω. Consider Nω /U. It is
elementarily equivalent to N. Take the element (1, 2, 3, 4, . . . )/U. It is
strictly greater than any standard natural number.

Example 3 (Infinitesimals)
Let U be a non-principal ultrafilter on ω. Consider Rω /U. It is
elementarily equivalent to R. Take the element (1, 12 , 31 , 14 , . . . )/U. It is
strictly greater than 0, yet strictly smaller than any standard real number.

5 / 31
Elementary classes and ultraproducts
Definition 4
A class C of models is called elementary, if C is the class of models of
some set T of sentences, that is, C = Mod(T ).

Lemma 5
If C is an elementary class, then C is closed under ultraproducts.

I The next theorem was first proved by Keisler, under the assumption
of GCH. Later Shelah proved it with no special assumptions.
I The proof requires sophisticated infinite combinatorics, and is beyond
the scope if the subject.

Theorem 6 (Keisler, Shelah)


Two similar structures A and B are elementarily equivalent if and only if
there is a set I and an ultrafilter U on I such that AI /U ∼
= BI /U.
6 / 31
Elementary classes and ultraproducts
Definition 4
A class C of models is called elementary, if C is the class of models of
some set T of sentences, that is, C = Mod(T ).

Lemma 5
If C is an elementary class, then C is closed under ultraproducts.

I The next theorem was first proved by Keisler, under the assumption
of GCH. Later Shelah proved it with no special assumptions.
I The proof requires sophisticated infinite combinatorics, and is beyond
the scope if the subject.

Theorem 6 (Keisler, Shelah)


Two similar structures A and B are elementarily equivalent if and only if
there is a set I and an ultrafilter U on I such that AI /U ∼
= BI /U.
7 / 31
Elementary classes and ultraproducts
Definition 4
A class C of models is called elementary, if C is the class of models of
some set T of sentences, that is, C = Mod(T ).

Lemma 5
If C is an elementary class, then C is closed under ultraproducts.

I The next theorem was first proved by Keisler, under the assumption
of GCH. Later Shelah proved it with no special assumptions.
I The proof requires sophisticated infinite combinatorics, and is beyond
the scope if the subject.

Theorem 6 (Keisler, Shelah)


Two similar structures A and B are elementarily equivalent if and only if
there is a set I and an ultrafilter U on I such that AI /U ∼
= BI /U.
8 / 31
Elementary classes and ultraproducts
Definition 4
A class C of models is called elementary, if C is the class of models of
some set T of sentences, that is, C = Mod(T ).

Lemma 5
If C is an elementary class, then C is closed under ultraproducts.

I The next theorem was first proved by Keisler, under the assumption
of GCH. Later Shelah proved it with no special assumptions.
I The proof requires sophisticated infinite combinatorics, and is beyond
the scope if the subject.

Theorem 6 (Keisler, Shelah)


Two similar structures A and B are elementarily equivalent if and only if
there is a set I and an ultrafilter U on I such that AI /U ∼
= BI /U.
9 / 31
Elementary classes characterised
Theorem 7
A ≡ B if and only if there is an elementary embedding of A into an
ultrapower of B. More generally, the following is true for any class K of
structures of the same signature Σ:
1. K is the class of all Σ-models of some set of first order sentences.
2. K is closed under elementary embeddings and ultraproducts.
3. K = EPU (K0 ) for some class of Σ-structures K0 (E denotes
elementary embeddings and PU ultraproducts).

Proof.
I (1) ⇒ (2) is immediate, as elementary embeddings and ultraproducts preserve all
sentences (n.b. closure under elementary embeddings implies closure under isomorphism).
I (2) ⇒ (3) by taking K0 to be K.
I Now assume (3) and consider Th(K0 ). If A |= Th(K0 ), then A ≡ B for some model
B ∈ K0 . By Theorem 7, we have AI /U ∼= BI /U for some I and U.
I As A - AI /U, we have A - BI /U; moreover BI /U ∈ PU (K0 ).
I So, A ∈ EPU (K0 ), proving (1).
10 / 31
Elementary classes characterised
Theorem 7
A ≡ B if and only if there is an elementary embedding of A into an
ultrapower of B. More generally, the following is true for any class K of
structures of the same signature Σ:
1. K is the class of all Σ-models of some set of first order sentences.
2. K is closed under elementary embeddings and ultraproducts.
3. K = EPU (K0 ) for some class of Σ-structures K0 (E denotes
elementary embeddings and PU ultraproducts).

Proof.
I (1) ⇒ (2) is immediate, as elementary embeddings and ultraproducts preserve all
sentences (n.b. closure under elementary embeddings implies closure under isomorphism).
I (2) ⇒ (3) by taking K0 to be K.
I Now assume (3) and consider Th(K0 ). If A |= Th(K0 ), then A ≡ B for some model
B ∈ K0 . By Theorem 7, we have AI /U ∼= BI /U for some I and U.
I As A - AI /U, we have A - BI /U; moreover BI /U ∈ PU (K0 ).
I So, A ∈ EPU (K0 ), proving (1).
11 / 31
Digression: finite models
Lemma 8
Let C be a class of finite models of some signature Σ. If C contains
arbitrarily large finite models, then C is not elementary.

Proof.
I Suppose C = Mod(T ) for some theory T .
I Let (Ai : i ∈ ω) be a sequence of models from C with strictly increasing sizes of their
Q
universes. Take A = i∈ω Ai /U, for a non-principal U.
I Then, A |= T , but A is infinite, so C ⊂ Mod(T ), contradicting the assumption.

Finite Model Theory


Lemma 8 can be viewed as dashing all hope for a reasonable model theory of finite structures:
completeness fails. Compactness also fails: say, the set {n < c : n ∈ N} (in the signature of
natural numbers with an additional costant c) is finitely satisfiable in the class of finite models,
but has no finite model. Yet, Finite Model Theory is alive and well. This is mostly because
Ehrenfeucht-Fraı̈ssé games work. We will look at them later.

12 / 31
Digression: finite models
Lemma 8
Let C be a class of finite models of some signature Σ. If C contains
arbitrarily large finite models, then C is not elementary.

Proof.
I Suppose C = Mod(T ) for some theory T .
I Let (Ai : i ∈ ω) be a sequence of models from C with strictly increasing sizes of their
Q
universes. Take A = i∈ω Ai /U, for a non-principal U.
I Then, A |= T , but A is infinite, so C ⊂ Mod(T ), contradicting the assumption.

Finite Model Theory


Lemma 8 can be viewed as dashing all hope for a reasonable model theory of finite structures:
completeness fails. Compactness also fails: say, the set {n < c : n ∈ N} (in the signature of
natural numbers with an additional costant c) is finitely satisfiable in the class of finite models,
but has no finite model. Yet, Finite Model Theory is alive and well. This is mostly because
Ehrenfeucht-Fraı̈ssé games work. We will look at them later.

13 / 31
Digression: finite models
Lemma 8
Let C be a class of finite models of some signature Σ. If C contains
arbitrarily large finite models, then C is not elementary.

Proof.
I Suppose C = Mod(T ) for some theory T .
I Let (Ai : i ∈ ω) be a sequence of models from C with strictly increasing sizes of their
Q
universes. Take A = i∈ω Ai /U, for a non-principal U.
I Then, A |= T , but A is infinite, so C ⊂ Mod(T ), contradicting the assumption.

Finite Model Theory


Lemma 8 can be viewed as dashing all hope for a reasonable model theory of finite structures:
completeness fails. Compactness also fails: say, the set {n < c : n ∈ N} (in the signature of
natural numbers with an additional costant c) is finitely satisfiable in the class of finite models,
but has no finite model. Yet, Finite Model Theory is alive and well. This is mostly because
Ehrenfeucht-Fraı̈ssé games work. We will look at them later.

14 / 31
Ultraproducts and axiomatisability
Lemma 9
Let (ϕn : n < ω) be a sequence of sentences of some signature Σ, and let
T be a Σ-theory. Let C be the subclass of Mod(T ) such that C ∈ C iff
C |= ϕn for some n < ω. Assume that C contains a set {Ci : i < ω} of
models such that for every n < ω the set {i ∈ ω : Ci |= ϕn } is finite.
Then, C is not an elementary class.

Proof.
I Let C = n<ω Cn /U, where U is an ultrafilter containing the Fréchet filter. So, C |= T .
Q

I Moreover, since [[ϕn ]] is finite, we have C |= ¬ϕn for every n.


I Hence C ∈
/ C, so C is not closed under ultraproducts.

Corollary 10
The following classes are not elementary: (1) all torsion groups; (2) all
connected graphs; (3) all fields of non-zero characteristic.

15 / 31
Ultraproducts and axiomatisability
Lemma 9
Let (ϕn : n < ω) be a sequence of sentences of some signature Σ, and let
T be a Σ-theory. Let C be the subclass of Mod(T ) such that C ∈ C iff
C |= ϕn for some n < ω. Assume that C contains a set {Ci : i < ω} of
models such that for every n < ω the set {i ∈ ω : Ci |= ϕn } is finite.
Then, C is not an elementary class.

Proof.
I Let C = n<ω Cn /U, where U is an ultrafilter containing the Fréchet filter. So, C |= T .
Q

I Moreover, since [[ϕn ]] is finite, we have C |= ¬ϕn for every n.


I Hence C ∈
/ C, so C is not closed under ultraproducts.

Corollary 10
The following classes are not elementary: (1) all torsion groups; (2) all
connected graphs; (3) all fields of non-zero characteristic.

16 / 31
Ultraproducts and axiomatisability
Lemma 9
Let (ϕn : n < ω) be a sequence of sentences of some signature Σ, and let
T be a Σ-theory. Let C be the subclass of Mod(T ) such that C ∈ C iff
C |= ϕn for some n < ω. Assume that C contains a set {Ci : i < ω} of
models such that for every n < ω the set {i ∈ ω : Ci |= ϕn } is finite.
Then, C is not an elementary class.

Proof.
I Let C = n<ω Cn /U, where U is an ultrafilter containing the Fréchet filter. So, C |= T .
Q

I Moreover, since [[ϕn ]] is finite, we have C |= ¬ϕn for every n.


I Hence C ∈
/ C, so C is not closed under ultraproducts.

Corollary 10
The following classes are not elementary: (1) all torsion groups; (2) all
connected graphs; (3) all fields of non-zero characteristic.

17 / 31
Failures of finite axiomatisability

Exercise 1
Prove Corollary 10 applying Lemma 9.

A technique for showing failures of finite axiomatisability


Let C = Mod(T ) for some finite set T of sentences.
I By Lemma 5, we have that C is closed under ultraproducts.
I But note that also the complement of C is an elementary class,
V
namely, A ∈/ C iff A |= ¬ T .
I So, C is also closed under ultraproducts.
I Hence, if C is not closed under ultraproducts, then C is not finitely
axiomatised.

Monk used this technique to prove that the class RRA of Representable
Relation Algebras is not finitely axiomatised.

18 / 31
Failures of finite axiomatisability

Exercise 1
Prove Corollary 10 applying Lemma 9.

A technique for showing failures of finite axiomatisability


Let C = Mod(T ) for some finite set T of sentences.
I By Lemma 5, we have that C is closed under ultraproducts.
I But note that also the complement of C is an elementary class,
V
namely, A ∈/ C iff A |= ¬ T .
I So, C is also closed under ultraproducts.
I Hence, if C is not closed under ultraproducts, then C is not finitely
axiomatised.

Monk used this technique to prove that the class RRA of Representable
Relation Algebras is not finitely axiomatised.

19 / 31
Failures of finite axiomatisability

Exercise 1
Prove Corollary 10 applying Lemma 9.

A technique for showing failures of finite axiomatisability


Let C = Mod(T ) for some finite set T of sentences.
I By Lemma 5, we have that C is closed under ultraproducts.
I But note that also the complement of C is an elementary class,
V
namely, A ∈/ C iff A |= ¬ T .
I So, C is also closed under ultraproducts.
I Hence, if C is not closed under ultraproducts, then C is not finitely
axiomatised.

Monk used this technique to prove that the class RRA of Representable
Relation Algebras is not finitely axiomatised.

20 / 31
Example: Algebras of Binary Relations
We view them as two-sorted structures:
I Algebra sort. We will use a, b, c, . . . for its elements. They form a
Boolean algebra with additional structure consisting of a constant 10 ,
a unary operation ˘, and a binary operation ;, which intuitively
represent the identity relation, relational converse, and relational
composition, respectively.
I Base set sort. We will use x, y , z, . . . for its elements. They are
vertices of a complete graph whose edges are labelled with the
elements of the algebra sort. Intuitively, the algebra elements are
names of some binary relations on the base set.
I Labelling. A ternary relation λ(a, x, y ). Intuitively, λ(a, x, y ) says
a(x, y ), or written in infix notation xay .
I Formally an Algebra of Binary Relations is a structure

(A, X , +, ·, ;, −, ˘, 0, 1, 10 , λ)

where λ ∈ A × X × X (all operations are of algebra sort), such that:


21 / 31
Algebras of Binary Relations
A1. (A, +, ·, −, 0, 1) is a Boolean Algebra (+ is join, · is meet, − is complement).
A2. (A, ;, ˘, 10 ) is an involutive monoid.
A3. (a + b) ; c = a ; c + b ; c and a ; (b + c) = a ; b + a ; c.
A4. a ; b · c 6= 0 ⇔ a˘ ; c · b 6= 0 ⇔ b ; c˘ · a˘ 6= 0
R1. λ(a + b, x, y ) ↔ λ(a, x, y ) ∨ λ(b, x, y )
R2. λ(¬a, x, y ) ↔ λ(1, x, y ) ∧ ¬λ(a, x, y )
R3. λ(10 , x, y ) ↔ x = y
R4. λ(a˘, x, y ) ↔ λ(a, y , x)
R5. λ(a ; b, x, y ) ↔ ∃z · λ(a, x, z) ∧ λ(b, z, y )
R6. a 6= 0 → ∃x, y · λ(a, x, y ).

Comments on the axioms


I All apparently free variables are universally quantified.
I All purely algebraic axioms are (equivalent to) equations.

22 / 31
Algebras of Binary Relations
A1. (A, +, ·, −, 0, 1) is a Boolean Algebra (+ is join, · is meet, − is complement).
A2. (A, ;, ˘, 10 ) is an involutive monoid.
A3. (a + b) ; c = a ; c + b ; c and a ; (b + c) = a ; b + a ; c.
A4. a ; b · c 6= 0 ⇔ a˘ ; c · b 6= 0 ⇔ b ; c˘ · a˘ 6= 0
R1. λ(a + b, x, y ) ↔ λ(a, x, y ) ∨ λ(b, x, y )
R2. λ(¬a, x, y ) ↔ λ(1, x, y ) ∧ ¬λ(a, x, y )
R3. λ(10 , x, y ) ↔ x = y
R4. λ(a˘, x, y ) ↔ λ(a, y , x)
R5. λ(a ; b, x, y ) ↔ ∃z · λ(a, x, z) ∧ λ(b, z, y )
R6. a 6= 0 → ∃x, y · λ(a, x, y ).

Comments on the axioms


I All apparently free variables are universally quantified.
I All purely algebraic axioms are (equivalent to) equations.

23 / 31
Algebras of Binary Relations
Definition 11
An algebra (A, +, ·, ;, ¬, ˘, 0, 1, 10 ) is a Relation Algebra (RA) if it satisfies
all equations in A1–A4. An algebra (A, +, ·, ;, ¬, ˘, 0, 1, 10 ) is a
Representable Relation Algebra (RRA) if it is isomorphic to an algebra
reduct of an Algebra of Binary Relations. We will write (A, X , λ) for a
two-sorted algebra of binary relations whose algebra reduct is A.

Historical comments
I RRA is closed under HSP, hence definable by equations (Tarski).
I Not all RAs are RRAs (Lyndon).
I RRA is not finitely axiomatised (Monk).

Exercise 2
Let (A1 , X1 , λ1 ) and (A2 , X2 , λ2 ) be algebras of binary relations. Prove that the
structure (A1 × A2 , X1 ∪˙ X2 , µ), with µ defined by putting µ((a1 , a2 ), x, y ) iff
λi (ai , x, y ) and x, y ∈ Xi (i = 1, 2), is an algebra of binary relations.
24 / 31
Algebras of Binary Relations
Definition 11
An algebra (A, +, ·, ;, ¬, ˘, 0, 1, 10 ) is a Relation Algebra (RA) if it satisfies
all equations in A1–A4. An algebra (A, +, ·, ;, ¬, ˘, 0, 1, 10 ) is a
Representable Relation Algebra (RRA) if it is isomorphic to an algebra
reduct of an Algebra of Binary Relations. We will write (A, X , λ) for a
two-sorted algebra of binary relations whose algebra reduct is A.

Historical comments
I RRA is closed under HSP, hence definable by equations (Tarski).
I Not all RAs are RRAs (Lyndon).
I RRA is not finitely axiomatised (Monk).

Exercise 2
Let (A1 , X1 , λ1 ) and (A2 , X2 , λ2 ) be algebras of binary relations. Prove that the
structure (A1 × A2 , X1 ∪˙ X2 , µ), with µ defined by putting µ((a1 , a2 ), x, y ) iff
λi (ai , x, y ) and x, y ∈ Xi (i = 1, 2), is an algebra of binary relations.
25 / 31
Algebras of Binary Relations
Definition 11
An algebra (A, +, ·, ;, ¬, ˘, 0, 1, 10 ) is a Relation Algebra (RA) if it satisfies
all equations in A1–A4. An algebra (A, +, ·, ;, ¬, ˘, 0, 1, 10 ) is a
Representable Relation Algebra (RRA) if it is isomorphic to an algebra
reduct of an Algebra of Binary Relations. We will write (A, X , λ) for a
two-sorted algebra of binary relations whose algebra reduct is A.

Historical comments
I RRA is closed under HSP, hence definable by equations (Tarski).
I Not all RAs are RRAs (Lyndon).
I RRA is not finitely axiomatised (Monk).

Exercise 2
Let (A1 , X1 , λ1 ) and (A2 , X2 , λ2 ) be algebras of binary relations. Prove that the
structure (A1 × A2 , X1 ∪˙ X2 , µ), with µ defined by putting µ((a1 , a2 ), x, y ) iff
λi (ai , x, y ) and x, y ∈ Xi (i = 1, 2), is an algebra of binary relations.
26 / 31
Monk’s proof
Definition 12
Let A be a complete atomic Boolean algebra, with the set of atoms
{10 } ∪ {ai : i < κ} (where κ a cardinal). Define ai ˘ = ai for every i, and
(
¬(ai + aj ) if i 6= j
ai ; aj =
10 + ai if i = j

All other operations are implicitly defined by this. Then A is a relation


algebra, called a Lyndon algebra. We will write Lyκ for so defined RA.

Lemma 13
(Lyκ , X , λ) is an algebra of binary relations if and only if X is the set of
points of an affine plane. In particular, we have:
I Every infinite Lyndon algebra is representable.
I A finite Lyndon algebra Lyn is representable iff an affine plane of
order n exists.
27 / 31
Monk’s proof
Definition 12
Let A be a complete atomic Boolean algebra, with the set of atoms
{10 } ∪ {ai : i < κ} (where κ a cardinal). Define ai ˘ = ai for every i, and
(
¬(ai + aj ) if i 6= j
ai ; aj =
10 + ai if i = j

All other operations are implicitly defined by this. Then A is a relation


algebra, called a Lyndon algebra. We will write Lyκ for so defined RA.

Lemma 13
(Lyκ , X , λ) is an algebra of binary relations if and only if X is the set of
points of an affine plane. In particular, we have:
I Every infinite Lyndon algebra is representable.
I A finite Lyndon algebra Lyn is representable iff an affine plane of
order n exists.
28 / 31
Monk’s proof
Lemma 14 (Bruck, Ryser, Chowla)
If a finite affine plane of order n exists and n = 1, 2 (mod 4), then n is a
sum of two squares. Thus, affine planes of order n do not exist for
infinitely many n.

Lemma 15
Let I be the set of all theseQn for which there is no affine plane of order n.
Consider the ultraproduct i∈I Lyi /U for some non-principalQ ultraproduct
on I . Then, Lyi is not representable for every i ∈ I , but i∈I Lyi /U is
representable.

Theorem 16
The class RRA is not finitely axiomatised. In particular, RRA has no finite
equational axiomatisation.

Proof.
The complement of RRA is not closed under ultraproducts.
29 / 31
Monk’s proof
Lemma 14 (Bruck, Ryser, Chowla)
If a finite affine plane of order n exists and n = 1, 2 (mod 4), then n is a
sum of two squares. Thus, affine planes of order n do not exist for
infinitely many n.

Lemma 15
Let I be the set of all theseQn for which there is no affine plane of order n.
Consider the ultraproduct i∈I Lyi /U for some non-principalQ ultraproduct
on I . Then, Lyi is not representable for every i ∈ I , but i∈I Lyi /U is
representable.

Theorem 16
The class RRA is not finitely axiomatised. In particular, RRA has no finite
equational axiomatisation.

Proof.
The complement of RRA is not closed under ultraproducts.
30 / 31
Monk’s proof
Lemma 14 (Bruck, Ryser, Chowla)
If a finite affine plane of order n exists and n = 1, 2 (mod 4), then n is a
sum of two squares. Thus, affine planes of order n do not exist for
infinitely many n.

Lemma 15
Let I be the set of all theseQn for which there is no affine plane of order n.
Consider the ultraproduct i∈I Lyi /U for some non-principalQ ultraproduct
on I . Then, Lyi is not representable for every i ∈ I , but i∈I Lyi /U is
representable.

Theorem 16
The class RRA is not finitely axiomatised. In particular, RRA has no finite
equational axiomatisation.

Proof.
The complement of RRA is not closed under ultraproducts.
31 / 31

You might also like