Distinguishing one structure from another

Wilfrid Hodges
Queen Mary and Westfield College, London

Johannesburg, December 1999

1. The basic idea of model theory

1
A formula is a grammatical expression φ in
some language (natural or formal) that can
be used to make a true or false statement.
But in fact φ may need interpreting in order
to become either true or false.
So we have

Formula φ
plus yields truth value
interpretation I

We say that φ is true in I,

or that I satisfies φ,
or that I is a model of φ,
or in symbols I |= φ,
if under the interpretation I, φ is true.

FIRST SLOGAN: A formula defines a class of

interpretations, viz. those which make it true.

So model theory is part of the theory of

definition.
2
English example (based on Wason selection
task experiment, 1966)

φ is the sentence

If the letter on one side of the card is

a vowel, the number on the other side
is even.

A suitable card gives an interpretation of ‘the

letter on one side of the card’, and of ‘the
number on the other side’.
The card is a model of φ if it satisfies the rule
expressed in φ.

SECOND SLOGAN: Model-theoretic truth is

ordinary truth. Model-theoretic satisfaction is
ordinary satisfaction.

3
Typical examples of phrases needing
interpretation:

1. Singular definite descriptions

‘the card’

‘the number on the back’

2. Variables

‘X’, ‘G’, ‘g’, ‘g1’ etc. etc.

Model theory concentrates on phrases of

these two kinds.

An interpretation for such phrases says what

they refer to. (Such interpretations are also
known as assignments or valuations.)
4
Sometimes an expression needs interpretation
but gets a default interpretation from the
context (or point of reference) where it is
applied. For example

‘the President of the Russian

Federation’

currently refers to Boris Yeltsin; the default

interpretation is the present one.

If we are studying a particular card, the

default interpretation of

‘the number on the back’

is the number on the back of that card. The

card serves as a point of reference.

The interpretation of this phrase has to be a

certain sort of thing, viz. a number.
Expressions with this feature are called sortal.
5
Quantifiers

These involve interpretation at two levels.

Take for example

‘Everybody loves Madonna.’

1. This is true in a context if in that context

every assignment to the sortal variable
‘xperson’ satisfies the formula

‘xperson loves Madonna.’

2. The context must determine who or what

counts as a ‘person’; we say it has a domain
or universe of persons.

6
Note: In 1 we split the interpretation into a
context and an assignment. This division is
fundamental in model theory.

The context by itself is called a structure.

In logic a formula that needs only a context

to interpret it, and not a further assignment,
is called a sentence. For example

‘Everybody loves Madonna.’

is a sentence, but

‘xperson loves Madonna.’

is only a formula, because it has a free

variable.

A set of sentences is called a theory. A model

of a theory is a model of all the sentences in
the theory.
7
Mathematical example

A partial ordering is defined to be a structure

which is a model of the following theory:
∀x (x 6 x).

∀x∀y (x 6 y ∧ y 6 x → x = y).

∀x∀y∀z (x 6 y ∧ y 6 z → x 6 z).
Here ‘∀x’ means ‘for all x’, ‘∧’ means ‘and’,
‘→’ means ‘if . . . then’. These are logical
symbols with fixed meanings, except that a
partial ordering must have a domain (to
control the possible assignments to variables).

In this formal language an expression ‘aRb’

means: the pair (a, b) are in the relation R.
The relation symbol ‘6’ is short for:

‘the relation called 6’.

Each partial ordering A has a relation called

6. We write this relation as ‘6A’.
8
In model theory since 1950 the usual
situation is that we have a class of structures
and we classify them according to what
sentences are true in them.

We need to make sure that the structures are

suitable for the sentences, i.e. that they give
interpretations for the right symbols.

A signature is a set of symbols needing a

structure to interpret them. The symbols are
called the non-logical symbols of the
signature. A structure interpreting all and
only the symbols in the signature σ is called a
σ-structure.

A language of signature σ is one whose

formulas need for interpretation only a
structure interpreting the symbols in σ (and
assignments to variables).
9
Non-logical symbols normally have one of the
following forms:

• (individual) constants a, b, c etc.,

standing for elements of the domain;

• function symbols F , G, H etc., of fixed

arity, standing for functions over the
domain;

• relation symbols P , Q, R etc., of fixed

arity, standing for relations over the
domain.

One can also have propositional symbols p, q,

r etc., that stand for Truth or Falsehood
depending on the structure.

10
Example: The signature of arithmetic has
the following symbols.

• Individual constant symbol ‘0’.

• No propositional symbol.

• Relation symbol ‘6’ of arity 2.

• Function symbols ‘S’ of arity 1, and ‘+’,

‘.’, both of arity 2.

11
Hilary Putnam says in ‘Models and reality’:

‘Models are not lost noumenal waifs looking

for someone to name them; they are
constructions within our theory itself, and
they have names from birth.’

He seems to mean that no structures exist

except those for which we have names. This
is Putnam’s private ontology; model theory
makes no such restriction.

In particular we don’t require even that

signatures can be named, or that their
symbols can be physically written. The
expression ‘the function named ‘F ’ in the
structure A’ need not be read literally either;
it means the function F A, which is a part of
A.

12
 2. The classical indistinguishability
theorems

These are theorems about first-order model

theory, i.e. where the formulas are built up
from atomic formulas

R(x, y, a),

F (G(x), y) = G(z)
etc., using

• ¬ ‘not’ ;

• ∧ ‘and’, ∨ ‘or’ ;

• ∀ ‘for all elements’,

∃ ‘there is an element such that’.

13
Three useful notions

1. We say that two σ-structures A, B are

elementarily equivalent, in symbols A ≡ B, if
exactly the same first-order sentences are
true in A as in B.

2. We say that two first-order theories T , U

of signature σ are (logically) equivalent if
exactly the same σ-structures are models of T
as of U .

3. Two sentences φ and ψ are equivalent

modulo the theory T if every model of T that
is a model of φ is also a model of ψ and vice
versa.

14
Recall that a structure A of signature σ has a
domain dom(A).
Recall also that if c is an individual constant
in σ then cA must be in dom(A);
and if F is a function symbol in σ, say with
arity n, then for any n elements a1, . . . , an of
dom(A),
F A(a1, . . . , an)
is also an element of dom(A).
We express this by saying that dom(A) is
closed under the constants and function
symbols of σ.

We say that a σ-structure B is a substructure

of A, and that A is an extension of B, if

dom(B) ⊆ dom(A)
and within dom(B), all of cB , RB , F B etc. are
just the restrictions of cA, RA, F A etc.

15
Example.

In the signature with 0, + and −, the

structure 2Z of even integers is a
substructure of the structure Z of integers.
In fact we can ‘generate’ 2Z from 4 and 6,
say, by closing off under 0, + and −.

Generating is an important way of getting

substructures.

By contrast N, the structure of natural

numbers 0, 1, 2, . . ., is not a substructure of Z
in this signature,
since it is not closed under −.
But it is a substructure of Z in the smaller
signature consisting of 0 and +.

16
A first-order sentence is said to be universal if
it consists of a string of universal quantifiers
∀x (or no quantifiers), followed by a formula
with no quantifiers.
A theory is universal if all the sentences in it
are universal.

For example the theory of partial orderings

(previous lecture) is universal.

17
L
à oś-Tarski Theorem. For any first-order
theory T the following are equivalent:

(a) T is logically equivalent to a universal

theory.

(b) If A is any model of T , then every

substructure of A is also a model of T .

Corollary. If A is a model of a universal

theory T and every first-order sentence is
equivalent modulo T to a universal sentence,
then every substructure of A is elementarily
equivalent to A.

18
Augustus De Morgan (1846): ‘If language
were copious enough, [existential]
propositions would seldom occur: and the
idioms of every tongue are probably
influenced by its power of . . . converting
[existentials] into the form of universals.’

Thus if

∀x∃y∀z∃w R(x, y, z, w),

is true in A, it can be rewritten as

∀x∀z R(x, F (x), z, G(x, z)),

where T contains, for example,

∀x∀y∀z∀w (R(x, y, z, w) →

R(x, F (x), z, G(x, z))

which is a universal sentence.
The functions F A, GA (or the symbols F, G)
are called Skolem functions.
19
By the cardinality of a structure A, |A|, we
mean the number of elements of the domain
of A.
By the cardinality of a signature σ, |σ|, we
mean the number of symbols in σ, or ℵ0,
whichever the greater.
(The ℵ0 is to allow for ‘closing off’ in the
proof of the theorem below.)

Putting all these tricks together:

Downward Loewenheim-Skolem
Theorem, Skolem 1920
Let A be a σ-structure of infinite cardinality κ,
and let λ be a cardinal which is < κ but > |σ|.
Then A has a substructure which is
elementarily equivalent to A and has
cardinality λ.

20
Why just go downward?
The following is also true, though its proof is
completely different.

Upward Loewenheim-Skolem Theorem,

Mal’tsev 1938
Let A be a σ-structure of infinite cardinality
κ, and let λ be a cardinal which is > κ and
> |σ|. Then A has an extension which is
elementarily equivalent to A and has
cardinality λ.

Putting these facts together, first-order logic

is absolutely hopeless at distinguishing
between infinite cardinals.

21
The Upward Loewenheim-Skolem Theorem
rests on the following important result.
We say that a first-order theory T is
consistent if it has at least one model.

Compactness Theorem (Goedel,

Mal’tsev, Henkin)
Let T be a first-order theory. If every finite
subset of T is consistent, then T is consistent.

Typical application. Consider the structure

R of real numbers, in any reasonable
signature.
Let T be the set of all first-order sentences
true in R, together with the sentences

c > 0, c < 1, c < 0.1, c < 0.01, . . .

This theory has a model A, by the
compactness theorem. The element cA is
‘infinitesimal’, i.e. greater than 0 but less
than any positive fraction.
22
The Compactness Theorem tells us that
first-order logic is not just bad at
distinguishing infinite cardinals from one
another; it can’t even distinguish finite from
infinite.
(Apply the Compactness Theorem to the set
of sentences expressing:

There is at least one element.

There are at least two elements.
There are at least three elements.
etc.)

Also the theorems of this section tell us

nothing about finite structures. To get useful
information about distinguishing between
finite structures, we must look elsewhere.

23
 3. Distinguishing elements within a
structure

We go back to the first lecture and recall

that interpretations can be split into two
parts: a structure that provides a context,
and an assignment to free variables. So far
we have ignored assignments (though we
would have had to mention them in some
proofs). Now we do the opposite: we fix a
structure and look at assignments.

24
Throughout this lecture, σ is a fixed
signature.
For technical reasons we assume σ has at
least one individual constant.
A is a fixed σ-structure.

We can list the variables of the language as

x1, x2, x3, . . .

An assignment (or tuple) of length n in A is
a list of n elements of the domain of A:

ā = (a1, a2, . . . , an).

It assigns a1 to x1, a2 to x2 and so on.
If φ is a formula whose free variables are all
among x1, . . . , xn, then φ separates the
assignments of length n into those that
satisfy φ and those that don’t.

25
We write φ(An) for the set of assignments of
length n that satisfy φ in A.
φ(An) is an n-ary relation on dom(A);
we say it is a first-order definable relation on
A.

The big question: In any given structure A,

what are the relations of the form φ(An)?

Example: A is R, the field of real numbers;

n = 2.
If φ is the formula

x2
1 + x 2=1
2
then φ(R2) is the unit circle.

26
Fact. The class of first-order definable n-ary
relations on A

• contains the empty set;

• contains the set dom(A)n of all tuples of

length n in A;

• is closed under intersection ∩;

• is closed under union ∪;

• is closed under complement

(dom(A)n \ X).

The last three clauses say that this set is

closed under boolean combinations.
27
In some cases we can describe a simple set of
first-order formulas, called basic formulas,
and prove that every formula defines the
same relation in A as some boolean
combination of basic formulas.
Then the first-order definable relations are
exactly the boolean combinations of the
relations defined by basic formulas.

Arguments along these lines are known as the

method of quantifier elimination. They are
usually not easy, though some relatively
recent adaptations of back-and-forth games
make them easier.

Fortunately the method of quantifier

elimination works (and gives important
information) for many interesting
mathematical structures.

28
The term algebra T

(Martin Davis is responsible for logic

programmers calling T the Herbrand universe,
but he has since regretted this name which is
historically inaccurate.)

The set of terms of the signature σ is defined

inductively by:

• Every individual constant is a term.

• Every variable is a term.

• If F is a function symbol of arity n, and

t1, . . . , tn are terms, then
F (t1, . . . , tn)
is a term.

A closed term is a term with no variables in it.

29
We make the set C of closed terms of σ into
a σ-structure T as follows.

The domain is C.

For each individual constant c, cT is c.

For each n-ary function symbol F , if t1, . . . , tn

are closed terms, then

F A(t1, . . . , tn) = F (t1, . . . , tn).

For each relation symbol R, RT is empty.

Universal algebraists know T as the free

σ-algebra on no generators.

30
Given a closed term

F (G(a), G(F (c, d)))

we shall call G(a) the (1)-th subterm
and G(F (c, d)) the (2)-th subterm.
Likewise we call a the (1,1)-th-subterm,
F (c, d) the (2,1)-th subterm,
d the (2,1,2)-th subterm, etc.
We call the whole term the ()-th subterm.

Quantifier elimination theorem. As the

basic formulas we can take all formulas of the
forms

The (. . . )-th subterm of xi is equal to

the (. . . )-th subterm of xj .

The (. . . )-th subterm of xi begins

with an F (or a c, etc.).

31
Hence none of the following can be expressed
by first-order formulas:

1. x appears in y as a subterm.

2. The symbol ‘F ’ appears in the term x.

3. x and y have a subterm in common.

32
 4. Games for comparing structures

The first-order languages in this lecture have

only finitely many symbols in their signatures,
and none of these symbols are function
symbols.

A typical example is the language of graphs.

It has just one 2-ary relation symbol E;

E(x, y)
means that there is an edge between node x
and node y.

33
In a graph G we say that a list of nodes

(a1, . . . , an)
is a path from a1 to an if there are an edge
from a1 to a2, an edge from a2 to a3, . . . , an
edge from an−1 to an.
The length of this path is n − 1.
The list (a) counts as a path from a to a of
length 0.

We say that nodes a, b in a graph are

connected if there is a path from a to b; their
distance is the length of a shortest such path.
We say that the graph itself is connected if
every pair of nodes of the graph is connected.

Our main task in this lecture is to show that

there is no first-order sentence distinguishing
the connected graphs from the unconnected
ones.
34
We want to compare two σ-structures A and
B.
We shall do it by setting up a game
EFω (A, B) (EF after Ehrenfeucht and
Fraı̈ssé).

The game will use the notion of an atomic

formula.

For any signature σ, the atomic formulas of σ

are the formulas of these two kinds:

• t1 = t2, where t1 and t2 are terms.

• R(t1, . . . , tn), where R is a relation symbol

of arity n and t1, . . . , tn are terms.

So in the signature for graphs, the only

atomic formulas are
x i = xj , E(xi, xj ).

35
There are two players, ∀belard (often known
as Spoiler) and ∃loise (often known as
Duplicator).
Note that Spoiler is male and Duplicator is
female.

Intuitively, ∀belard is trying to show that A

and B are different by finding a feature of one
structure that can’t be matched in the other.
∃loise is trying to show that A and B are the
same, by finding a match in the other
structure for each feature that ∀belard points
to.

36
The game EFω (A, B) takes place in steps,
starting at the 1st.
In each step (say the n-th), first ∀belard
chooses an element from either dom(A) or
dom(B), then ∃loise chooses one from the
other of these sets.
We write an for the element chosen from A
and bn for the element chosen from B.

After infinitely many steps, ∀belard wins if

there is some atomic formula φ which is
satisfied by exactly one of (a1, a2, . . .) and
(b1, b2, . . .); otherwise ∃loise wins.

Note also the game EFn(A, B) where n is a

finite number. This is the same as above,
except that the players stop as soon as they
have chosen an, bn.
The smaller n, the easier it is for ∃loise to
win.
37
EXAMPLE FOR GAME

38
We say that A is n-equivalent to B, in
symbols
A ∼n B,
if ∃loise has a winning strategy for the game
EFn(A, B).
If she has a winning strategy for the infinite
game EFω (A, B), we say that A and B are
back-and-forth equivalent,

A ∼ B.
Note:

A ∼ B ⇒ . . . ⇒ A ∼n B ⇒ A ∼n−1 B ⇒ . . .
because the games get easier for ∃loise to win
as they get shorter.

FACT. All the relations ∼, ∼n are equivalence

relations.

39
The main theorem about these games needs
the notion of the quantifier rank qr(φ) of a
formula φ. This is defined inductively:

• If φ is atomic then qr(φ) = 0.

• qr(¬φ) = qr(φ).

• qr(φ ∧ ψ) = qr(φ ∨ ψ) = max(qr(φ), qr(ψ)).

• qr(∀xφ) = qr(∃xφ) = qr(φ) + 1.

Important fact: Every first-order formula has

a finite quantifier rank.

40
Fraı̈ssé’s Theorem. Let A, B be
σ-structures. Then for each n the following
are equivalent:

(i) A ∼n B.

(ii) If φ is any sentence of quantifier rank at

most n, then A is a model of φ if and only
if B is a model of φ.

Corollary. Let K be a class of σ-structures.

Suppose that for every finite n there are a
structure An in K and a structure Bn not in
K, such that An ∼n Bn. Then there is no
first-order sentence whose models are the
structures in K.

41
Let k, m be positive integers with k > 3.
We write [k, m] for the graph consisting of m
copies of a cycle of length k.

Then [k, m] is connected if and only if m = 1.

Suppose A, B are graphs of the form [k, m]

(possibly different k, m).
Suppose ā = (a1, . . . , ap) are nodes of A and
b̄ = (b1, . . . , bp) are nodes of B.
We say ā and b̄ are h-matched if for all i, j
and every h0 6 h,

distance(ai, aj ) = h0 ⇔ distance(bi, bj ) = h0.

42
One can show: If A = [k, m] and B = [k0, m0]
and ā, b̄ are as above and are
(2n−1 − 1)-matched, then (ā, b̄) is a winning
position for ∃loise in EFp+n(A, B). Then:

Fact. If k, k0 are both > 2n then

[k, m] ∼n [k0, m0].

43
Let n be a positive integer. Then by the Fact
above,
[2n, 1] ∼n [2n, 2].
Since one of these graphs is connected and
the other is not, this shows that no first-order
sentence distinguishes between connected
graphs and unconnected graphs.

Hence there is no first-order formula φ(x, y)

that expresses in graphs that x, y are
connected.
Otherwise the sentence

∀x∀y φ(x, y)
would have distinguished between connected
and unconnected graphs.

44
 5. Ways of cheating

During the 1980s several authors proposed

using first-order logic as either a programming
language or a specification language.
At first glance this is absurd.
We saw in Lecture Two that first-order logic
can’t express that a structure is finite.
We saw in Lecture Three that first-order
logic can’t express that a graph is connected.

45
Example from Prolog

Sterling and Shapiro, The Art of Prolog, give

the following Prolog program to define
connectedness in graphs:

connected(Node,Node).
connected(Node1,Node2) ←
edge(Node1,Link),
connected(Link,Node2).

This is a first-order theory T :

∀x C(x, x).

∀x∀y∀z (E(x, y) ∧ C(y, z) → C(x, z)).

46
 ∀x C(x, x).

∀x∀y∀z (E(x, y) ∧ C(y, z) → C(x, z)).

Note 1: This theory T is not in the language

of graphs, because it has the extra symbol C.

Note 2: If we read ‘E(x, y)’ as ‘there is an

edge from x to y’, and ‘C(x, y)’ as ‘there is a
path from x to y’, then both sentences are
true in any graph.

Note 3: Both sentences are still true in any

graph if we read E as before but take
‘C(x, y)’ to be true for all nodes x and y
(regardless of whether they are connected).

47
So T is hopeless as a straightforward
model-theoretic definition of connectedness.
So why is it called a definition by Prolog
people?

Clue: T is a definite Horn clause theory.

This means that each sentence in T has one
of the forms

∀x1 . . . ∀xn ψ,

∀x1 . . . ∀xn (φ1 ∧ . . . φk → ψ)

where φ1, . . . , φk , ψ are atomic formulas.

48
Suppose G is any graph.
Give each node of G a name, by adding new
individual constants to the signature if
necessary.
There is a smallest set U of atomic sentences
that contains:

(i) All the atomic sentences true in G.

(ii) All the sentences C(a, a) (a an individual

constant).

(iii) The sentence C(a, c) whenever U contains

sentences E(a, b) and C(b, c).

We read the second and third clauses off

from the Horn theory T .

Then U is a complete description of a

structure H which is a model of T .
49
After clause (i) we never added new individual
constants or equalities ‘a = b’ or sentences
beginning with ‘E’.
So H is an exact copy of G, except that it
also has a new binary relation C H .

This relation holds between a and b if and

only if a and b are connected.

There are several ways of saying this.

One (the universal algebraist’s) is that the
only model of T that we are concerned with
is the ‘free model over G’.
Another (the logician’s) is that T is being
used not as a first-order theory but as part of
an inductive definition.

50
Algebraic example

Suppose A is any infinite structure.

Let G(A) be the group of all permutations of
dom(A).
What can we say about A, using first-order
sentences about G(A)?

Let T ransp(x) say:

x2 = 1 and x 6= 1 and for all y there is

a subgroup of order at most 6
containing both x and y −1xy.

Then T ransp(x) expresses that x is a

transposition.

51
Let Overlap(x, y) be

T ransp(x) ∧ T ransp(y) ∧ xy 6= yx.

Then Overlap(x, y) expresses that x, y are
transpositions (a, b), (a, c) with b 6= c; we call a
the pivot of x, y.

One can write a formula Equiv(x, y, z, w)

which expresses that Overlap(x, y),
Overlap(z, w) and x, y have the same pivot as
z, w.

Now to express ‘w moves every element of

dom(A)’ we say:

Whenever Overlap(x, y), it is false that

Equiv(x, y, wx, wy).

52
To express ‘w is cyclic’ we say:

Every x that commutes with w either

moves every element of dom(A)
moved by w, or moves no element of
dom(A) moved by w.

So we can write a first-order sentence φ in

the language of groups, such that for every
infinite structure A, G(A) satisfies φ if and
only if A is countable.
Namely: ‘There is a cyclic permutation that
moves every element.’

53
Question: Why doesn’t this contradict the
Upward Loewenheim-Skolem Theorem?

Answer: Because φ is not about A but about

a structure built up from A using set theory.

One can use first-order sentences to make

other strong statements about A, by talking
instead about other structures built up from
A:
for example the universe of sets with
elements of dom(A) as urelements.

54
References

For Lectures 2–4,

Wilfrid Hodges, A shorter model

theory, Cambridge University Press
1997.

For back-and-forth games on finite structures

there is a fuller account in

Heinz-Dieter Ebbinghaus and Jörg

Flum, Finite model theory, Springer
1999 (2nd edition).

55