Introduction to Mathematical Logic

Adrien Deloro

Spring 2009
Contents

᾿Εν ἀρχῇ ἦν ὁ λόγος 4

0 Propositional Logic (Allegro) 7

0.1 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
0.1.1 The Language of Propositional Logic . . . . . . . . . . . . 8
0.1.2 Well-Formed Formulas . . . . . . . . . . . . . . . . . . . . 8
0.1.3 Unique Readability . . . . . . . . . . . . . . . . . . . . . . 9
0.2 Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
0.2.1 Truth Assignments . . . . . . . . . . . . . . . . . . . . . . 11
0.2.2 Satisfaction and Semantic Consequence . . . . . . . . . . 12
0.2.3 Simplifying the Language . . . . . . . . . . . . . . . . . . 13
0.3 Natural Deduction (in Classical Logic) . . . . . . . . . . . . . . . 15
0.3.1 Syntactic Consequences and Deductions . . . . . . . . . . 16
0.3.2 Our First Deductions . . . . . . . . . . . . . . . . . . . . 18
0.3.3 Simplifying the Language (and Presentation) . . . . . . . 20
0.3.4 The Soundness Theorem . . . . . . . . . . . . . . . . . . . 27
0.4 The Completeness Theorem . . . . . . . . . . . . . . . . . . . . . 28
0.4.1 Extending the Theory . . . . . . . . . . . . . . . . . . . . 29
0.4.2 Finding a Truth Assignment . . . . . . . . . . . . . . . . 30
0.5 The Compactness Theorem . . . . . . . . . . . . . . . . . . . . . 31
0.5.1 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . 31
0.5.2 Applications to Decidability* . . . . . . . . . . . . . . . . 32
0.5.3 A Topological Proof* . . . . . . . . . . . . . . . . . . . . . 34
0.6 A Little Modal Logic* . . . . . . . . . . . . . . . . . . . . . . . . 35
0.6.1 Semantics* . . . . . . . . . . . . . . . . . . . . . . . . . . 35
0.6.2 Proof Theory* . . . . . . . . . . . . . . . . . . . . . . . . 36
0.6.3 Completeness* . . . . . . . . . . . . . . . . . . . . . . . . 38
0.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

1 First-Order Logic (Largo) 45

1.1 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
1.1.1 First-order Languages . . . . . . . . . . . . . . . . . . . . 46
1.1.2 Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.1.3 Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.2 Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

1
 1.2.1 Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
1.2.2 Parameters and Interpretations . . . . . . . . . . . . . . . 51
1.2.3 Satisfaction and Semantic Consequence . . . . . . . . . . 51
1.3 Substitutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
1.3.1 Substitutability . . . . . . . . . . . . . . . . . . . . . . . . 53
1.3.2 A Renaming Algorithm . . . . . . . . . . . . . . . . . . . 54
1.3.3 Substitutions and Satisfaction . . . . . . . . . . . . . . . . 55
1.4 Deductions and Soundness . . . . . . . . . . . . . . . . . . . . . 57
1.4.1 Deductions . . . . . . . . . . . . . . . . . . . . . . . . . . 57
1.4.2 Simplifying the Language . . . . . . . . . . . . . . . . . . 58
1.4.3 Soundness . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
1.5 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
1.5.0 Strategy and Witnesses . . . . . . . . . . . . . . . . . . . 63
1.5.1 Expanding the Language . . . . . . . . . . . . . . . . . . 64
1.5.2 Extending the Theory . . . . . . . . . . . . . . . . . . . . 66
1.5.3 Finding a Structure (and Assignment) . . . . . . . . . . . 68
1.6 Consequences and Compactness . . . . . . . . . . . . . . . . . . 70
1.6.1 Decidability* . . . . . . . . . . . . . . . . . . . . . . . . . 70
1.6.2 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . 71
1.6.3 Non-Standard Analysis* . . . . . . . . . . . . . . . . . . . 72
1.7 An Alternate Proof of Compactness* . . . . . . . . . . . . . . . 74
1.7.1 Filters and Ultrafilters* . . . . . . . . . . . . . . . . . . . 74
1.7.2 Ultraproducts (The Łoś Structure)* . . . . . . . . . . . . 76
1.7.3 Alternate Proof of Compactness* . . . . . . . . . . . . . . 78
1.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

2 Second-Order Logic (Presto) 85

2.1 Compactness fails . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.2 Peano Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . 87

1’ Some Model Theory (Furioso) 89

1’.0 A Word on Model-Terrorists . . . . . . . . . . . . . . . . . . . . 89
1’.1 Elementary Equivalence; Inclusion and Elementary Inclusion . . 92
1’.1.1 Elementary Equivalence . . . . . . . . . . . . . . . . . . . 92
1’.1.2 Inclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
1’.1.3 Elementary Inclusion . . . . . . . . . . . . . . . . . . . . . 95
1’.2 Morphisms, Elementary Morphisms, Categoricity . . . . . . . . . 96
1’.2.1 Morphisms . . . . . . . . . . . . . . . . . . . . . . . . . . 97
1’.2.2 Elementary Morphisms . . . . . . . . . . . . . . . . . . . 98
1’.2.3 Categoricity . . . . . . . . . . . . . . . . . . . . . . . . . . 99
1’.3 Löwenheim-Skolem Theorems . . . . . . . . . . . . . . . . . . . . 100
1’.3.1 The Substructure Generated by a Set . . . . . . . . . . . 100
1’.3.2 Skolem Functions and the Descending Version . . . . . . . 101
1’.3.3 The General Version and the Łoś-Vaught Criterion . . . . 102
1’.4 Back-and-Forth Methods . . . . . . . . . . . . . . . . . . . . . . 103
1’.4.1 Dense Linear Orderings . . . . . . . . . . . . . . . . . . . 104

2
 1’.4.2 ∞-isomorphisms . . . . . . . . . . . . . . . . . . . . . . . 106
1’.4.3 Finitary Back-and-Forth* . . . . . . . . . . . . . . . . . . 108
1’.5 Quantifier Elimination and ω-Saturation . . . . . . . . . . . . . . 111
1’.5.1 Elimination Sets . . . . . . . . . . . . . . . . . . . . . . . 111
1’.5.2 ω-Saturated Models . . . . . . . . . . . . . . . . . . . . . 114
1’.5.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
1’.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

Indices 128
Index of Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Index of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

List of Lectures 131

3
᾿Εν ἀρχῇ ἦν ὁ λόγος

The Greeks didn’t have a formal language, which in Europe appears during
the Renaissance. Because of our education it now seems very natural to us
that every mathematical statement may be translated into specific signs; yet
this fascinating fact has deeply moved Leibniz’ mind. The adequation of such
a poor alphabet to the manifold hues of mathematics is striking indeed. One
should however not succumb, and their temptations are always present, to the
Sirens of sheer formalism; a mathematician aims at transcribing his intuitions
into symbols, not at assembling symbols and then seeing if they by chance make
any sense. Never does he renounce lucid thinking nor his mother language; and
even less does the logician.
A logician is a mathematician who bears in mind both the mathematical
truths and the way they are stated; able to distinguish the meaning from the
formulation, he can study their interplay. If the mathematician regards language
as a tool to express some interesting properties, the logician thinks the properties
of this tool are themselves worth the interest.

Logic is the study of the relations between things and words

In any particular form of logic, we need both a system of objects and a language.
We must of course make the language expressive enough for the properties of
the objects which appear relevant to us, but not more.
An example will shed some light.
• We are all familiar with the encoding of numbers in the standard, decimal,
number system. Our objects here are numbers; the language is the collec-
tion of representations of numbers. The latter is better described by its
alphabet (the symbols 0 through 9), and the formation rules of what is an
acceptable representation of a number. There is only such rule: writing
from the right to the left, the last digit is not 0.
• Once this is done, we need to be able to translate back and forth. Given
a number, it is easy to compute recursively (using remainders modulo 10
etc.) its decimal representation. Conversely, the number can be retrieved
from its expansion via a certain sum. We have thus seen that our lan-
guage of representations of numbers in base 10 is sufficient to describe all
numbers.

4
 Though the objects and language might seem a little less familiar, we shall
follow essentially the same guidelines in our study of various forms of logic:
• agree on a collection of interesting objects
• give a set of basic symbols (an alphabet)
• give a collection of rules for assembling them (a syntax)
• give a meaning to the language (a semantics).
After that, the study of the relationships between the language and what it
stands for can start.

Logic turns proofs into objects

The expressive power of a language may be measured by the notions of conse-
quence it creates; let us explain why consequence is essentially twofold.
The main feature of contemporay mathematical logic is that not only mathe-
matical statements may be made formal, but also those particular combinations
of mathematical statements which are called proofs. In a sense, modern logic is
able to treat mathematical sentences themselves as words of a language whose
sentences are mathematical deductions. In this case the grammar consists of
the rules explaining which conclusions may be drawn from given assumptions.
(Treating proofs as formal objects is essentially as shocking as treating intuitive
properties as formal statements; in a sense, the XIXth and XXth centuries were
the Renaissance of Mathematical Logic.)
On the semantic side, the concept of logic consequence is that of entail-
ment. Similar to material implication which is only about the truth value of its
components, this semantic approach of consequence is concerned only with one
property never occuring without another one.
Hence there are two notions of consequence:
• consequence at a syntactic level: ϕ follows from Σ if there is a proof, using
Σ, of ϕ. This is written Σ ` ϕ, which may be read “Σ proves ϕ”.
• consequence at a semantic level: ϕ follows from Σ if it is the case that
truth of Σ always causes truth of ϕ. This is written Σ |= ϕ, which may
be read “Σ entails ϕ”.
A syntax remarkably fits a semantics when the two notions ` and |= coincide.
This adequation holds in Propositional Logic and First-Order Logic, but fails
in Second-Order Logic, which is in a sense too expressive. These three forms of
logic are explained hereafter.

The semester
In the course of the semester we shall study three levels of logic, from the less
to the more expressive.

5
 • Propositional Logic concerns itself with compound propositions, using the
five usual connectives. In propositional logic, our basic symbols already
represent propositions; we connect them with the usual connectives. This
form of logic can be dealt with using truth tables: it suffices to specify for
each P is P is true or false. It is not very expressive, but this playground
is an extremely interesting field of investigation for dgging up notions
relevant in deeper logics.
• First-order logic is the natural language for elements. Our basic symbols
are now for elements, and certain (given) relations they can have. One
sees that this is already deeper than propositional logic; one also sees that
one has to introduce quantification on elements. First-order logic is strong
enough to encode much of interesting mathematics, and still enjoys very
nice properties.
• Second-order logic builds on the previous one. Our symbols can now also
be for “predicates‘”, that is for sets, relations. But unlike first-order logic,
second-order logic may quantify on those. So we need new quantifiers
(roughly speaking, we now may quantify on classes, not only on objects).
Second-order logic is extremely expressive, and this is actually the reason
why the beautiful, deep theorems we shall prove in the first two cases will
fail in this somehow too general context.
Looking back at our first steps, we realize that propositional logic is some
kind of zeroeth-order logic. After the investigation of these three forms of logic,
we should believe that the one that best fits our needs is first-order logic. We
shall then move to model theory, which is the mathematical study of first-order
theories. This should explain the rather unorthodox numbering.

6
Chapter 0

Propositional Logic
(Allegro)

Propositional logic, sometimes called Propositional Calculus, is interested in

compound propositions; it is the logic which underlies truth tables. In spite of
its little mathematical interest, it is a perfect field for a first encounter with
important notions.
We shall introduce a formalization of deduction in the setting of propositional
logic, and study (some of) its properties. This will be a brief account of proof
theory.
We shall also introduce the objects of propositional logic, that is truth as-
signments. These are on the side of meanings, called semantics. They bring
their own notion of consequence; the main result of this chapter which serves
as a training area for more expressive logics, is that in propositional logic, both
notions coincide.

In this chapter:
• Describe the language of propositional logic (§0.1)
• Describe the objects of propositional logic (§0.2)
• Formalize the notion of entailment (§0.2.2)

• Formalize the notion of deduction in propositional logic (§0.3)

• Show that entailment and deduction coincide (§0.3.4 and 0.4)
• Derive the compactness theorem and applications (§0.5)

• An incursion to modal logic* (§0.6)

Lecture 1 (Language and Wff’s; Unique Readability)

7
0.1 Syntax
We first deal with the familiar syntax of propositional logic. The expressions are
(finite) sequences of well-known symbols (§0.1.1), but not all such expressions
may be given a meaning. This will lead us to the relevant notion of formula
(§0.1.2). A key, and fairly natural, property of the language is that there is
exactly one way to read a relevant formula (Theorem 0.1.9 of §0.1.3).

0.1.1 The Language of Propositional Logic

As we have said, propositional logic is interested in building coumpound propo-
sitions from elementary propositions.
Definition 0.1.1 (language of propositional logic). The language of proposi-
tional logic consists of:
• two punctuation symbols : the parentheses “(” and “)”
• five connectives ¬, ∧, ∨, →, ↔
• a set A of sentence symbols A1 , . . . , An , . . . .
One should already have an intuition that ¬ and → suffice to encode the
other three connectives. This unformal idea will be given a precise meaning in Exercise 0.1
§0.3.3.
Definition 0.1.2 (expression). An expression is any (finite) sequence of sym-
bols from the alphabet. Its length is its length as a sequence.
Remark 0.1.3. The set E of expressions of propositional logic is countable.
Technically, Definition 0.1.1 defines the alphabet of propositional logic. It is
never perfectly clear what “language” means: does one mean the alphabet, the
set of expressions, or the set of well-formed formulas (Definition 0.1.4 below)?
Very fortunately, the cardinals of the alphabet, the set of expressions, and the
set of well-formed formulas, are the same.

0.1.2 Well-Formed Formulas

As “(→ A1 ¬” is an expression, we must clarify the formation rules.
Definition 0.1.4 (well-formed formula). The collection WFF of well-formed
formulas (wff in short) is the smallest set such that:
• each An is a wff;
• if ϕ and ψ are wff, so are (¬ϕ), (ϕ ∧ ψ), (ϕ ∨ ψ), (ϕ → ψ), and (ϕ ↔ ψ).
Remark 0.1.5. As this is the only relevant notion of formula anyway, we shall
soon omit “well-formed”, and say “formula”, for short.

8
Definition 0.1.6 (theory). A theory is a set of wff’s.
Wff’s are built from the sentence symbols with connectives. As this is a
recursive definition, most of our proofs on formulas will proceed by induction
(notice that the length of a compound proposition is bigger than that of the
propositions which compose it).
Let us be more technical and check that Definition 0.1.4 makes sense.
Notation 0.1.7. We define the following functions.

C¬ : E → E
ϕ →
7 (¬ϕ)

C∧ : E2 → E C∨ : E2 → E
(ϕ, ψ) 7→ (ϕ ∧ ψ) (ϕ, ψ) 7→ (ϕ ∨ ψ)
C→ : E2 → E C↔ : E2 → E
(ϕ, ψ) 7→ (ϕ → ψ) (ϕ, ψ) 7→ (ϕ ↔ ψ)

Notice that these functions actually take WFF (resp. WFF2 ) to WFF.
Notation 0.1.8. Let WFF0 = A, and for n ∈ N,
[ [ [ [
WFFn+1 = WFFn C¬ (WFFn ) C∧ (WFF2n ) · · · C↔ (WFF2n )

It is clear that for all n ∈ N, WFFn is countable. It is also clear that WFF =
∪n∈N WFFn . This proves that WFF is a set, and even a countable set. More
generally, if A is infinite (but perhaps uncountable), then Card WFF = Card A.

0.1.3 Unique Readability

There is still something which needs to be proved about our well-formed formu-
las: that no ambiguity arises when we read one.
Theorem 0.1.9 (unique readability). The restrictions to WFF of the functions
C¬ , C∧ , C∨ , C→ , and C↔ are injective and have disjoint images.

Corollary 0.1.10. Let ϕ be a wff of length > 1. Then there is a unique decom-
position of ϕ among the following possibilities:

(¬ψ1 ), (ψ1 ∧ ψ2 ), (ψ1 ∨ ψ2 ), (ψ1 → ψ2 ), (ψ1 ↔ ψ2 )

(This means that the connective and the composing propositions are unique.)

An initial segment of an expression E is an expression E1 such that there

is an expression E2 with E = E1 E2 (concatenation). A proper initial segment
E1 of E is an initial segment such that E1 6= E. Notice that the empty (blank)
sequence is an expression but not a wff.
The proof of Theorem 0.1.9 will rely on the following key observation.

9
Lemma 0.1.11 (balanced parenthesing).

(i). A wff has as many left parentheses as right parentheses.

(ii). A non-empty proper initial segment of a wff has strictly more left paren-
theses than right parentheses.
(iii). No proper initial segment of a wff is a wff.

Proof .
(i). We show the property by induction on (the length of) ϕ, using the de-
scription of WFF found above (Notation 0.1.8).
If ϕ ∈ WFF0 , then ϕ is a sentence symbol, and this is obvious. Assume
the result is know for elements of WFFn . Then an element of C¬ (WFFn )
is of the form ϕ = (¬ψ) with ψ ∈ WFFn , so by induction ψ has as many
left parentheses as right parentheses, and so does ϕ. The same holds for
elements of C∧ (WFF2n ), C∨ (WFF2n ), C→ (WFF2n ), and C↔ (WFF2n ). So any
element of WFFn+1 has as many left parentheses as right parentheses.
By induction, the property is true on WFF.

(ii). Same method. If ϕ ∈ WFF0 , then ϕ is a sentence symbol, so it has no

non-empty proper initial segments, and the claim is true!
Assume it holds for elements of WFFn . Let ϕ be of the form (¬ψ), for
ψ ∈ WFFn . Consider a non-empty proper initial segment ϕ1 of ϕ: we
show that ϕ1 has more (’s than )’s.

• If ϕ1 is “(” or “(¬”, we are done.

• If ϕ1 is “(¬ψ1 ” where ψ1 is a non-empty proper initial segment of ψ,
then we are done by induction.
• If ϕ1 is “(¬ψ” then we are done as ψ has as many (’s as )’s by (i).
• There are no other cases, as ϕ1 is a proper initial segment of ϕ.

This shows the property for ϕ. One copes with other four connectives in
a similar vein.
(iii). A consequence of (i) and (ii) (the empty expression is not a wff anyway).
Exercise 0.2

Proof of Theorem 0.1.9. Assume that C (ϕ1 , ψ1 ) = C (ϕ2 , ψ2 ), that is:

(ϕ1  ψ1 ) = (ϕ2 ψ2 )

for well-formed formulas ϕ1 , ψ1 , ϕ2 , ψ2 and connectives  and .

We may delete the first “(”, getting

ϕ1  ψ1 ) = ϕ2 ψ2 )

10
 Then ϕ1 or ϕ2 must be an initial segment of the other; without loss of
generality ϕ1 is an initial segment of ϕ2 . If ϕ1 is a proper initial segment of
the wff ϕ2 , then from balanced parenthesing (Lemma 0.1.11 (iii)), ϕ1 can’t be
a wff, a contradiction.
So ϕ1 = ϕ2 , and deleting it we reach

ψ1 ) = ψ2 )

At this point it is clear that  = ; deleting it, and the final right parentheses,
we derive
ψ1 = ψ2
Of course the case of C¬ is even easier than that of binary connectives, so
everything is proved.
Such a quick proof of the Unique Readability Theorem is the reason why we
have introduced so many parentheses, in spite of the disgusting redundancies
they create. From now on we freely omit parentheses when there is no ambiguity,
i.e. we write ¬ϕ ∧ ψ instead of ((¬ϕ) ∧ ψ).

End of Lecture 1.

Lecture 2 (The Semantics of Propositional Logic)

0.2 Semantics
We briefly escape the claws of syntax, and give sense to our symbols. So far
they bore no meaning; at no point yet did we agree on what they stand for.

0.2.1 Truth Assignments

Let us establish the connection between the language and the objects it is about.
Due to the relative poverty of propositional logic, not much can be said: ev-
erything is described once each sentence symbol of A will have been assigned a
truth value.
Definition 0.2.1 (truth values). There are two truth values: True (T ) and
False (F , also denoted ⊥).
Notice that everything from now on could be defined with more truth values;
for instance continuous logic investigates the case where the set of truth values
is [0, 1]. But the important theorems we want might no longer hold.
Definition 0.2.2 (truth assignment). A truth assignment is a function v : A →
{F, T }.

Once the sentence symbols have been given a truth value, the “natural”
truth value of every wff is known.

11
Proposition 0.2.3. Let v be a truth assignment. Then there is a unique ex-
tension v : WFF → {F, T } such that:
• v(An ) = v(An ) for each n ∈ N;
• v(¬ψ) = T iff v(ψ) = F ;
• v(ψ1 ∧ ψ2 ) = T iff v(ψ1 ) = T and v(ψ2 ) = T ;
• v(ψ1 ∨ ψ2 ) = T iff v(ψ1 ) = T or v(ψ2 ) = T ;
• v(ψ1 → ψ2 ) = T iff v(ψ1 ) = F or v(ψ2 ) = T ;
• v(ψ1 ↔ ψ2 ) = T iff v(ψ1 ) = v(ψ2 ).
Proof . This extension v is well-defined by Unique Readability (Theorem 0.1.9).
A quick induction shows that it is unique.
Example 0.2.4. If v(A1 ) = v(A2 ) = T and v(A3 ) = v(A4 ) = F , then one has
v((A1 → ¬A2 ) ↔ (A3 ∨ A4 )) = T .
Only a little later (at the end of §0.2.3) shall we identify v to is extension v̄.

0.2.2 Satisfaction and Semantic Consequence

Definition 0.2.5 (satisfaction). Let Σ be a theory and v be a truth assignment.
v satisfies Σ if for all ϕ ∈ Σ, v(ϕ) = T .
Example 0.2.6.
• v(A1 ) = F , v(A2 ) = T satisfies (A1 → A2 ).
• v(An ) = T for all n ∈ N satisfies {A1 , A1 ∧ A2 , . . . , A1 ∧ · · · ∧ An , . . .}.
Definition 0.2.7 (satisfiability). Let Σ be a theory. Σ is satisfiable if there is
a truth assignment satisfying it.
Example 0.2.8.
• {¬A1 , A1 → A2 , ¬A1 → A3 } is satisfiable (try for instance the truth
assignment v(A1 ) = v(A2 ) = F , v(A3 ) = T ).
• {A1 → ¬A1 } is not satisfiable.
• {A1 , ¬A2 , A1 ↔ A2 } is not satisfiable. Exercise 0.3

A very natural question occurs: when is a theory satisfiable? it will have an

answer with the Compactness Theorem, Theorem 0.5.2 below.
The most important notion of logic is the following.
Definition 0.2.9 (semantic consequence). Let Σ be a theory and ϕ a wff. Σ
entails ϕ, denoted Σ |= ϕ, if every truth assignment satisfying Σ satisfies ϕ.

12
 One also says that Σ tautologically implies ϕ. If ∅ |= ϕ, one may then say
that ϕ is a tautology, or that ϕ is valid. We do not insist on classical terminology.
Lemma 0.2.10. Σ ∪ {ϕ} is satisfiable iff Σ 6|= ¬ϕ.
Proof . Suppose that Σ ∪ {ϕ} is satisfiable. Then there is a truth assignment
v satisfying Σ and ϕ; by definition (Proposition 0.2.3), v does not satisfy ¬ϕ.
This means that Σ does not entail ¬ϕ.
Conversely suppose that Σ 6|= ¬ϕ. There is therefore a truth assignement v
satisfying Σ, but not satisfying ¬ϕ. It follows v(ϕ) = T , so v actually satisfies
Σ ∪ {ϕ}.
Theorem 0.2.11. Let Σ0 be a finite theory and ϕ a wff. Then there is an
algorithm which decides whether Σ entails ϕ or not.
Proof . In Σ0 ∪ {ϕ}, only finitely many sentence symbols occur. List all (but
finitely many) truth tables involving them. If all truth tables satisfying Σ0 also
satisfy ϕ, the answer is “yes”; otherwise answer “no” at first flaw.
Question. And what about an infinite set?
This question will find an answer when we can reduce the infinite to the finite
(Compactness, Theorem 0.5.2). The answer is exactly half as good (Corollary
0.5.14). This is far from obvious, as Σ might now involve infinitely many sen-
tence symbols, giving rise to continuum many truth tables. However ϕ is a
single formula, so one has the feeling that some finite subset Σ0 of Σ would
suffice to entail ϕ, in case Σ does.
There is however no bound a priori on how many sentence symbols are
involved in Σ0 , as shows the example ϕ = A1 , Σ = {A2 , A2 , → A3 , . . . , An−1 →
An , An → A1 }. This is actually the reason why the algorithm for infinite Σ
given in Corollary 0.5.14 cannot be implemented to answer “no”: if it hasn’t
answered “yes” after a long waiting time, it could be
• either because Σ |= ϕ, but we haven’t waited long enough to be positive
about it;
• or simply because Σ 6|= ϕ, but we’ll never know.
These questions will be made more precise in §0.5.2.

0.2.3 Simplifying the Language

Five connectives are too much: they made inductive definitions and proofs
extremely clumsy. Moreover we have a feeling that they are redundant. We
shall cut down on symbols and rules.
In Proposition 0.2.3, we have treated ∧, ∨, ↔ as if they were elementary
connectives, and not abbreviations. We show that considering them as abbre-
viations is harmless. This involves rewriting the whole theory with only two
connectives, and checking that this new setting is equivalent to the previous
one.

13
Notation 0.2.12. The collection WFF0 is the smallest set such that:
• each An is a wff;
• if ϕ and ψ are wff, so are ¬ϕ and ϕ → ψ.

The set WFF0 would require a version of the Unique Readability Theorem
(Theorem 0.1.9), which may be regarded as a special case of the previous one,
since WFF0 ⊆ WFF. In any case, we keep working with few parentheses.
There is of course a natural way to translate a “standard” wff ϕ (using five
connectives) into a wff’ ϕ0 using only ¬ and →.

Notation 0.2.13. We inductively translate any wff ϕ into a formula ϕ0 which

uses only the two connectives ¬ and →:
• A0n is An for each n ∈ N;
• (¬ϕ)0 is (¬ϕ0 );

• (ϕ ∧ ψ)0 is ¬(ϕ0 → ¬ψ 0 );
• (ϕ ∨ ψ)0 is (¬ϕ0 → ψ 0 );
• (ϕ → ψ)0 is (ϕ0 → ψ 0 );
• (ϕ ↔ ψ)0 is ¬((ϕ0 → ψ 0 ) → ¬(ψ 0 → ϕ0 )).

Hence for any formula ϕ, the translation ϕ0 uses only two connectives, and
intuitively bears the same meaning.
Example 0.2.14. Let ϕ be (A1 ↔ A2 )∧(A3 ∨A1 ). Then ϕ0 is ¬((A1 ↔ A2 )0 →
¬(A3 ∨ A1 )0 ), which is ¬(¬((A1 → A2 ) → ¬(A2 → A1 )) → ¬(¬A3 → A1 )).
The issue is that this new notion of formula gives rise to another extension
of truth assignment, and therefore to a new notion of entailment. We must
check that these new notions do coincide with their original “five connectives”
analogues.
Notation 0.2.15. For a truth assignment v : WFF → {F, T }, let v 0 be the
extension of v defined by:

• v 0 (An ) = v(An ) for each n ∈ N;

• v 0 (¬ψ) = T iff v(ψ) = F ;
• v 0 (ψ1 → ψ2 ) = T iff v(ψ1 ) = F or v(ψ2 ) = T .

Notice that this is a subset of the conditions defining v̄ (Proposition 0.2.3),

and that v 0 (ψ) is defined only for ψ ∈ WFF0 .
Lemma 0.2.16. Let v be a truth assignment, ϕ a wff, and ϕ0 its translation
using only ¬ and →. Then v(ϕ) = v 0 (ϕ0 ).

14
Proof . By induction on ϕ.
• If ϕ is a sentence symbol, then by definition v(ϕ) = v(ϕ) = v 0 (ϕ0 ).
• Suppose that ϕ is ¬ψ. Then v(ϕ) = T is equivalent to v(ψ) = F , which is
by induction equivalent to v 0 (ψ 0 ) = F , which is equivalent to v 0 (ϕ0 ) = T .
• Suppose that ϕ is ψ1 → ψ2 . Then v(ϕ) = T iff v(ψ1 ) = F or v(ψ2 ) = T
iff v 0 (ψ10 ) = F or v 0 (ψ20 ) = T iff v 0 (ψ10 → ψ20 ) = T iff v 0 (ϕ0 ) = T .
• Suppose that ϕ is ψ1 ∧ ψ2 . Then v(ϕ) = T iff v(ψ1 ) = v(ψ2 ) = T iff
v 0 (ψ10 ) = v 0 (ψ20 ) = T iff v 0 (¬(ψ10 → ¬ψ20 )) = T iff v 0 (ϕ0 ) = T .
• We proceed similarly for the remaining two connectives ∨ and ↔.
Any modification, even apparently harmless, of the notion of truth assign-
ment, induces a modfication of the notion of semantic consequence (Definition
0.2.9).
Notation 0.2.17. Let Θ be a consistent subset of WFF0 and ψ ∈ WFF0 . Write
Θ |=0 ψ if whenever v 0 satisfies Θ, then v 0 (ψ) = T .
Corollary 0.2.18. Σ |= ϕ iff Σ0 |=0 ϕ0 .
Proof . Suppose that Σ |= ϕ. We show that Σ0 |=0 ϕ0 . Let v be a truth
assignement such that v 0 satisfies Σ0 . By Lemma 0.2.16, v satisfies Σ. By
assumption, v satisfies ϕ. Using Lemma 0.2.16 again, v 0 satisfies ϕ0 .
Now suppose that Σ0 |=0 ϕ0 . We show that Σ |= ϕ. Let v be a truth assign-
ment such that v satisfies Σ. By Lemma 0.2.16, v 0 satisfies Σ0 . By assumption,
v 0 satisfies ϕ0 . Using Lemma 0.2.16 again, v satisfies ϕ.
From now on, we know that the semantics induced by two connectives is the
same as that induced by five. In particular we happily forget about v 0 and |=0 .
(We are not entirely done with ϕ0 however, since a similar verification will have
to be carried on the syntactic side, §0.3.3.)
We shall also identify any truth assignment v with its extension v.

End of Lecture 2.

Lecture 3 (Natural Deduction in Classical Logic)

0.3 Natural Deduction (in Classical Logic)

In this section we formalize the notions of deduction and syntactic consequence;
this will be done in §0.3.1. We shall work in the framework of Natural De-
duction (a certain notion of deduction), using Classical Logic (certain rules of
deduction).
Natural Deduction was invented by Gerhard Gentzen; there are other notions
of deduction (for instance, Hilbert’s system, which relies on modus ponens), and

15
there are other logics (for instance, intuitionistic logic, which does not allow
contradiction proofs). They all yield different notions of deduction. We made
the choices which best reflect a normal mathematician’s thought.
After a couple of examples (§0.3.2), we shall proceed in §0.3.3 to reducing
the language to two connectives. There are indeed redundancies as we have
observed (Corollary 0.2.18). We shall show that one may restrict the language
and deduction rules to ¬ and → without affecting the power of the notion of
deduction.

0.3.1 Syntactic Consequences and Deductions

What is a consequence, again? In terms of proofs, ϕ follows from assumptions
Σ if there is a way to deduce ϕ using the wff’s of the theory Σ.
Definition 0.3.1 (deduction). A deduction is a finite sequence of basic steps
which are regarded as elementary deduction rules.
The deduction can be drawn as a tree showing the consecutive steps. This
will be more precise when we have given an example of a set of such rules, but
we can already formulate the following important definition.
Definition 0.3.2 (syntactic consequence). Let Σ be a theory and ϕ a wff. ϕ is
a theorem of Σ, denoted Σ ` ϕ, if there is a deduction of ϕ under Σ. One also
says that Σ proves ϕ.

If Σ = ∅, one writes ` ϕ, and says that ϕ is a theorem (of our logic).

Example 0.3.3. At the end of the day, (ϕ∧(ϕ → ψ)) → ψ should be a theorem.
We now explain the basic steps of deduction. Bear in mind that we chose
the most natural setting (natural deduction), but there are many other possible
settings, and not all are equivalent. A characteristic feature of natural deduction
is that connectives may be introduced and eliminated. Describing the deduction
rules amounts to explaining how connectives are “proved and used”. In what
follows, Σ stands for a (possibly empty) theory, and ϕ, ψ, etc. for wff’s.

• There is only one axiom (so far). In natural deduction for propositional
logic, the only way to start a proof is by assuming something.

• Weakening allows us to add unused assumptions to an existing deduction.

• The case of negation is very subtle. We have a way to introduce a nega-

tion provided we deduced an inconsistency (¬i ), but we also have a way
of removing double-negations (¬e ). This is a characteristic feature of clas-
sical logic, which is actually equivalent to admitting reductio ad absur-
dum. Keep in mind that this is only one among the possible treatments of

16
 negation and contradiction, though it seems satisfactory to us as working
mathematicians.
The elimination rule ¬e is necessary to show the excluded middle law
(Theorem 0.3.11 below); if we introduce a weaker rule, excluded middle
need not hold any more. Exercise 0.10
How hard it may be to believe at first, ¬i is not related to reduction to
absurdum. Only ¬e is.

• There are two elimination rules for ∧ as there are two possible ways to
use a conjunction.

• Dually, there are two introduction rules for ∨, as there are two ways to
prove a disjunction. This is in general difficult, since when trying to show
ϕ ∨ ψ, one never knows a priori which is true; contradiction proofs will be
helpful (see the proof of Theorem 0.3.11 below).
On the other hand ∨e may look subtle at first sight, but it merely describes
the process of a case-division.

• ∧ and ∨ may be regarded as convenient abbreviations (this will be made

formal in §0.3.3). But the connective → is essential.

The elimination rule →e is classically called modus ponens; dually the

introduction rule →i is called modus tollens.
• It is clear that ↔ should have the moral status of an abbreviation. Yet in
case it is admitted as a primitive connective, we need deduction rules.

We shall start using quantifiers in the next chapters and introduce adequate
rules; for the moment they won’t appear, as we work in Propositional Logic.
Remark 0.3.4. Notice that in our deductions we may have several assumptions,
but at each step we have only one conclusion. This is another typical feature of
Natural Deduction which supposedly reflects the way we mathematicians work.
But there are other, more complex, deduction systems (for instance, Gentzen’s
Sequent Calculus).
A formalization of deduction yields the following essential notion.
Definition 0.3.5 (consistency). A theory Σ is consistent if there is no wff ϕ
such that Σ ` ϕ and Σ ` ¬ϕ.

17
 In other words, a theory is consistent if it is not possible to deduce a contra-
diction from it. There is a symbol for inconsistency (or contradiction), which
we shall not use. Exercise 0.9

Lemma 0.3.6. Σ is inconsistent iff Σ ` ϕ for all wff ϕ.

Proof . Suppose Σ inconsistent. Then there is ψ such that Σ ` ψ and Σ ` ¬ψ.
Hence
Σ`ψ Σ ` ¬ψ
Wk Wk
Σ ∪ {¬ϕ} ` ψ Σ ∪ {¬ϕ} ` ¬ψ
¬i
Σ ` ¬¬ϕ
¬e
Σ`ϕ
The converse is obvious.

0.3.2 Our First Deductions

We begin with an easy example.
Example 0.3.7. For wff’s ϕ, ψ, one has ` (ϕ ↔ ψ) ↔ ((ϕ → ψ) ∧ (ψ → ϕ))
Verification: On the one hand,
Ax Ax
ϕ↔ψ`ϕ↔ψ ϕ↔ψ`ϕ↔ψ
↔e ↔e
ϕ↔ψ`ϕ→ψ ϕ↔ψ`ψ→ϕ
∧i
ϕ ↔ ψ ` (ϕ → ψ) ∧ (ψ → ϕ)
→i
` (ϕ ↔ ψ) → ((ϕ → ψ) ∧ (ψ → ϕ))
On the other hand,
Ax Ax
(ϕ → ψ) ∧ (ψ → ϕ) ` ((ϕ → ψ) ∧ (ψ → ϕ)) (∗1 )
∧e ∧e
(ϕ → ψ) ∧ (ψ → ϕ) ` ϕ → ψ (∗2 )
↔i
((ϕ → ψ) ∧ (ψ → ϕ)) ` ϕ ↔ ψ
→i
` ((ϕ → ψ) ∧ (ψ → ϕ)) → (ϕ ↔ ψ)
where (∗1 ) and (∗2 ) are the symmetric deductions obtained by exchanging ϕ
and ψ.
Combining both trees with ↔i , we find
` (ϕ ↔ ψ) ↔ ((ϕ → ψ) ∧ (ψ → ϕ)) ♦
Theorem 0.3.8 (Contraposition). Σ ` ϕ → ψ iff Σ ` ¬ψ → ¬ϕ.
Proof . Suppose Σ ` ϕ → ψ. Then
Ax
Σ`ϕ→ψ {ϕ} ` ϕ
Wk Wk Ax
Σ ∪ {¬ψ, ϕ} ` ϕ → ψ Σ ∪ {¬ψ, ϕ} ` ϕ {¬ψ} ` ¬ψ
→e Wk
Σ ∪ {¬ψ, ϕ} ` ψ Σ ∪ {¬ψ, ϕ} ` ¬ψ
¬i
Σ ∪ {¬ψ} ` ¬ϕ
→i
Σ ` ¬ψ → ¬ϕ

18
 Suppose Σ ` ¬ψ → ¬ϕ. What we just proved implies Σ ` ¬¬ϕ → ¬¬ψ.
Therefore
Ax Ax
{ϕ} ` ϕ {¬ϕ} ` ¬ϕ
Wk Wk
{ϕ, ¬ϕ} ` ϕ {ϕ, ¬ϕ} ` ¬ϕ
¬i
{ϕ} ` ¬¬ϕ Σ ` ¬¬ϕ → ¬¬ψ
Wk WK
Σ ∪ {ϕ} ` ¬¬ϕ Σ ∪ {ϕ} ` ¬¬ϕ → ¬¬ψ
→e
Σ ∪ {ϕ} ` ¬¬ψ
¬e
Σ ∪ {ϕ} ` ψ
→i
Σ`ϕ→ψ

Parallel to the relationship between entailment and satisfiability (Lemma

0.2.10) runs the following.

Lemma 0.3.9 (see Lemma 0.2.10). Σ ∪ {ϕ} is not inconsistent iff Σ 6` ¬ϕ.
Proof . Suppose Σ ∪ {ϕ} is not inconsistent and Σ ` ¬ϕ. Then Σ ∪ {ϕ} ` ϕ
(by axiom and weakening), and Σ ∪ {ϕ} ` ¬ϕ (by assumption). It follows that
Σ ∪ {ϕ} is inconsistent, against the assumption. So Σ 6` ¬ϕ.
Suppose that Σ 6` ¬ϕ, and Σ ∪ {ϕ} is inconsistent. By inconsistency , there
is ψ be such that Σ ∪ {ϕ} ` ψ and Σ ∪ {ϕ} ` ¬ψ. Then by ¬i , Σ ` ¬ϕ, against
the assumption. So Σ ∪ {ϕ} is not inconsistent.
Notice that the latter proof did not involve a semantic analog of the elimi-
nation of double negation (¬e ), but only ¬i ; whence the odd statement. As you
believe in ¬e , you may say “is consistent” for “is not inconsistent”.

Theorem 0.3.10 (Reductio ad Absurdum). If Σ ∪ {¬ϕ} is inconsistent then

Σ ` ϕ.
Proof . Suppose Σ ∪ {¬ϕ} inconsistent. By Lemma 0.3.9, Σ ` ¬¬ϕ. By (¬e ),
Σ ` ϕ.
Still not impressed?

Theorem 0.3.11 (Excluded Middle holds in Classical Logic). Excluded middle

is a theorem of Classical Logic.
That is: for any wff ϕ, one has ` ϕ ∨ ¬ϕ. Notice that one does not know a
priori which of ϕ or ¬ϕ to prove, so there is little chance to find a quick deduction
relying entirely on ∨i . We shall work our way through a contradiction proof
involving ¬e .

19
Proof . Due to pagesetting limitations, I must pause and resume.
Ax
{ϕ} ` ϕ
Wk Ax
{¬(ϕ ∨ ¬ϕ), ϕ} ` ϕ {¬(ϕ ∨ ¬ϕ)} ` ¬(ϕ ∨ ¬ϕ)
∨i Wk
{¬(ϕ ∨ ¬ϕ), ϕ} ` ϕ ∨ ¬ϕ {¬(ϕ ∨ ¬ϕ), ϕ} ` ¬(ϕ ∨ ¬ϕ)
¬(ϕ ∨ ¬ϕ) ` ¬ϕ
∨i
¬(ϕ ∨ ¬ϕ) ` ϕ ∨ ¬ϕ
| {z }
(∗)

(∗)
Ax
¬(ϕ ∨ ¬ϕ) ` ϕ ∨ ¬ϕ ¬(ϕ ∨ ¬ϕ) ` ¬(ϕ(∨¬ϕ)
¬i
` ¬¬(ϕ ∨ ¬ϕ)
¬e
` ϕ ∨ ¬ϕ

The proof of Theorem 0.3.11 shows that contradiction deductions (typically

excluded middle) are always a little counter-intuitive to formalize. Notice that
in some weaker deduction systems, excluded middle cannot be deduced!
End of Lecture 3.

Lecture 4 (Reduction to Two Connectives (1/2))

0.3.3 Simplifying the Language (and Presentation)

We have introduced a battery of deduction rules, though we had concluded in
§0.2.3 that two connectives were enough to convey all the desired meaning. We
now must check that working with only two connectives would preserve the
deductive strength, i.e. check that all deductions can be imitated even when
one restricts connectives and rules to the (¬, →) setting.
Recall what ϕ0 stands for (Notation 0.2.13). It is intuitively clear that prov-
ing ϕ and proving ϕ0 should be the same.
Lemma 0.3.12 (see Lemma 0.2.16). {ϕ} ` ϕ0 and {ϕ0 } ` ϕ.
Proof . By induction on ϕ. Provided the result has been proved for a formula
ψ, we shall freely write:
Σ`ψ
induction
Σ ` ψ0
though induction is of course not a deduction rule. This is actually short-hand
for the following:
..
.. induction
{ψ} ` ψ 0
→i
` ψ → ψ0
Wk
Σ ` ψ Σ ` ψ → ψ0
→e
Σ ` ψ0

20
Of course
Σ ` ψ0
induction
Σ`ψ
is a similar shortcut, and is licit for a similar reason.
• Suppose that ϕ is a sentence symbol. Then ϕ0 = ϕ, and the claim is
obvious.
• Suppose that ϕ is ¬ψ. Then ϕ0 = ¬(ψ 0 ). Now
Ax
{ψ 0 } ` ψ 0
Ax Wk
{ϕ} ` ϕ {ϕ, ψ 0 } ` ψ 0
Wk induction
{ϕ, ψ 0 } ` ϕ {ϕ, ψ 0 } ` ψ
¬i
{ϕ} ` ¬ψ 0
|{z}
=ϕ0

and
Ax
{ψ} ` ψ
0 0 Ax Wk
{ϕ } ` ϕ {ϕ0 , ψ} ` ψ
Wk induction
{ϕ0 , ψ} ` ϕ0 {ϕ0 , ψ} ` ψ 0
¬i
{ϕ} ` ¬ψ
|{z}
=ϕ0

• Suppose that ϕ is (ψ1 → ψ2 ). Then

Ax
{ψ10 } ` ψ10
Ax Wk
{ϕ} ` ϕ {ϕ, ψ10 } ` ψ10
Wk induction
{ϕ, ψ10 } ` ϕ {ϕ, ψ10 } ` ψ1
→e
{ϕ, ψ10 } ` ψ2
induction
{ϕ, ψ10 } ` ψ20
→i
{ϕ} ` ϕ0

On the other hand

Ax
{ψ1 } ` ψ1
Ax Wk
{ϕ0 } ` ϕ0 {ϕ0 , ψ1 } ` ψ1
Wk induction
{ϕ0 , ψ1 } ` ϕ0 {ϕ0 , ψ1 } ` ψ10
→e
{ϕ0 , ψ1 } ` ψ20
induction
{ϕ0 , ψ1 } ` ψ2
→i
{ϕ0 } ` ϕ

Of course the case of ¬ and → was trivial. The real proof starts here.
• Suppose that ϕ = ψ1 ∧ ψ2 , so that ϕ0 = (¬(ψ10 → ¬ψ20 )).

21
– One has
Ax
{ϕ} ` ϕ
Wk
{ϕ, ψ10 → ¬ψ20 } ` ϕ
∧e Ax
{ϕ, ψ10 → ¬ψ20 } ` ψ1 {ψ10 → ¬ψ20 } ` ψ10 → ¬ψ20
induction Wk
{ϕ, ψ10 → ¬ψ20 } ` ψ10 {ϕ, ψ10 → ¬ψ20 } ` ψ10 → ¬ψ20
→e
{ϕ, ψ10 → ¬ψ20 } ` ¬ψ20
| {z }
(∗)

Hence
Ax
{ϕ} ` ϕ
Wk
{ϕ, ψ10 → ¬ψ20 } ` ϕ
∧e
(∗) {ϕ, ψ10 → ¬ψ20 } ` ψ2
induction
{ϕ, ψ10 → ¬ψ20 } ` ¬ψ20 {ϕ, ψ10 → ¬ψ20 } ` ψ20
¬i
{ϕ} ` ¬(ψ10 → ¬ψ20 )

– On the other hand

Ax
{ψ10 } ` ψ10
Ax Wk
{¬ψ1 } ` ¬ψ1 {ϕ0 , ¬ψ1 , ψ10 , ψ20 } ` ψ10
Wk induction
{ϕ0 , ¬ψ1 , ψ10 , ψ20 } ` ¬ψ1 {ϕ0 , ¬ψ1 , ψ10 , ψ20 } ` ψ1
¬i
{ϕ , ¬ψ1 , ψ1 } ` ¬ψ20
0 0
→i
{ϕ0 , ¬ψ1 } ` ψ10 → ¬ψ20
| {z }
(†)

Therefore
Ax
(†) {ϕ0 } ` ¬(ψ10 → ¬ψ20 )
Wk
{ϕ0 , ¬ψ1 } ` ψ10 → ¬ψ2 {ϕ0 , ¬ψ1 } ` ¬(ψ10 → ¬ψ20 )
¬i
{ϕ0 } ` ¬¬ψ1
¬e
{ϕ0 } ` ψ1

Moreover
Ax
{¬ψ20 } ` ¬ψ20
Wk Ax
{ϕ0 , ψ10 , ¬ψ20 } ` ¬ψ20 {ϕ0 } ` ϕ0
→i Wk
{ϕ0 , ¬ψ20 } ` ψ10 → ¬ψ20 {ϕ0 , ¬ψ20 } ` ϕ0
¬i
{ϕ0 } ` ¬¬ψ20
¬e
{ϕ0 } ` ψ20
induction
{ϕ0 } ` ψ2

Combining both threes, we find {ϕ0 } ` ϕ.

22
 • Suppose that ϕ = ψ1 ∨ ψ2 , so that ϕ0 = (¬ψ10 → ψ20 ).
– {ψ1 , ¬ψ10 } is clearly inconsistent by induction. By Lemma 0.3.6,
{ψ1 , ¬ψ10 } ` ψ20 and {ψ1 } ` ϕ0 .
On the other hand, {ψ2 , ¬ψ10 } ` ψ20 by induction, so {ψ2 } ` ϕ0 .
As ϕ = ψ1 ∨ ψ2 , combining both deduction trees and eliminating ∨,
we find {ϕ} ` ϕ0 . This was the easy part.
– For the converse we need an easy remark:
Ax
{ψ10 } ` ψ10
induction
{ψ10 } ` ψ1
Wk Ax
{¬ϕ, ψ10 } ` ψ1 {¬ϕ} ` ¬ϕ
∨i Wk
{¬ϕ, ψ10 } ` ψ1 ∨ ψ2 {¬ϕ, ψ10 } ` ¬ϕ
¬i
{¬ϕ} ` ¬ψ10
| {z }
(∗1 )

Of course one has similarly {¬ϕ} ` ¬ψ20 (∗2 ).

Therefore
(∗1 )
Ax
{¬ϕ} ` ¬ψ10 {ϕ0 } ` ¬ψ10 → ψ20 (∗2 )
0 0 Wk 0 0 0 Wk
{ϕ , ¬ϕ} ` ¬ψ1 {ϕ , ¬ϕ} ` ¬ψ1 → ψ2 {¬ϕ} ` ¬ψ20
→e Wk
0 0
{ϕ , ¬ϕ} ` ψ2 {ϕ0 , ¬ϕ} ` ¬ψ20
¬i
{ϕ0 } ` ¬¬ϕ
¬e
{ϕ0 } ` ϕ

• The case of ↔ is left as an exercise, but essentially easier than ∨.

Hence ϕ and its translation ϕ0 prove each other. Of course they do so only
in the setting involving all connectives (since ϕ may involve ∧); Lemma 0.3.12
is but a first step as the notion of deduction ` itself has changed.
Notation 0.3.13. Let Σ be a theory and ϕ a wff. Write Σ `0 ϕ if there is
a deduction of ϕ from Σ using only axiom, weakening and the deduction rules
associated to ¬ and →.
Preservation of the notion of deduction under translation is therefore the
following result.
Corollary 0.3.14 (see Corollary 0.2.18). Let Σ be a theory and ϕ a wff. Then
Σ ` ϕ iff Σ0 `0 ϕ0 .
Proof . For the moment we prove only that Σ0 `0 ϕ0 implies Σ ` ϕ. Sup-
pose Σ0 `0 ϕ0 . By finiteness of the notion of a deduction, there are formulas
ψ1 , . . . , ψn ∈ Σ such that {ψ10 , . . . , ψn0 } `0 ϕ0 . Using →i repeatedly, one finds

`0 ψ10 → (ψ20 → . . . (ψn0 → ϕ0 ). . . )

23
As all the rules in `0 are in `, one even has
` ψ10 → (ψ20 → . . . (ψn0 → ϕ0 ). . . )
Now one sees that
0
(ψ1 → (ψ2 → . . . (ψn → ϕ). . . )) is ψ10 → (ψ20 → . . . (ψn0 → ϕ0 ). . . )
So by Lemma 0.3.12, it follows
` ψ1 → (ψ2 → . . . (ψn → ϕ). . . )
In particular, {ψ1 , . . . , ψn } ` ϕ, and Σ ` ϕ.
We shall prove the converse implication next time.
End of Lecture 4.
Lecture 5 (Reduction to Two Connectives (2/2); Soundness)
We finish the proof of Corollary 0.3.14.
Proof of Corollary 0.3.14, completed. The remaining proof that Σ ` ϕ im-
plies Σ0 `0 ϕ0 is by induction on the length of the deduction.
• The cases of axiom and weakening are clear.
• Suppose the last step is for instance ¬i :
Σ ∪ {ψ} ` χ Σ ∪ {ψ} ` ¬χ
¬i
Σ ` ¬ψ
By induction, one has Σ0 ∪ {ψ 0 } `0 χ0 and Σ0 ∪ {ψ 0 } `0 ¬χ0 . As ¬i is in
`0 , it follows
Σ0 ∪ {ψ 0 } `0 χ0 Σ0 ∪ {ψ 0 } `0 ¬χ0
¬i
Σ0 `0 ¬ψ 0
And therefore Σ0 `0 ¬ψ 0 .
• It is clear that the cases of ¬e , →i , and →e are as trivial (since these rules
are in `0 ). We now turn to the interesting configurations.
• Suppose the last step is ∧i , that is
Σ ` ψ1 Σ ` ψ2
∧i
Σ ` ψ1 ∧ ψ2
We aim at proving Σ0 `0 (ψ1 ∧ ψ2 )0 ; recall that (ψ1 ∧ ψ2 )0 is ¬(ψ10 → ¬ψ20 ).
By induction, we know Σ0 `0 ψ10 and Σ0 ` ψ20 . Therefore
..
.. induction
Ax
Σ0 `0 ψ10 {ψ10 → ¬ψ20 } `0 ψ10 → ¬ψ20
Wk
Σ0 ∪ {ψ10 → ¬ψ20 } `0 ψ10 Σ0 ∪ {ψ10 → ¬ψ20 } `0 ψ10 → ¬ψ20
→e
Σ0 ∪ {ψ10 → ¬ψ20 } `0 ¬ψ20
| {z }
(∗)

24
 and
..
.. induction
(∗) Σ0 `0 ψ20
Wk
Σ0 ∪ {ψ10 → ¬ψ20 } `0 ¬ψ20 Σ0 ∪ {ψ10 → ¬ψ20 } `0 ψ20
¬i
Σ0 `0 ¬(ψ10 → ¬ψ20 )

• Suppose the last step is ∧e1 :

Σ ` ψ1 ∧ ψ2
∧e1
Σ ` ψ1

By induction, Σ0 `0 ¬(ψ10 → ¬ψ20 ). Clearly Σ0 ∪ {¬ψ10 } `0 ψ10 → ¬ψ20 (∗).

Hence
..
.. induction
0 0 0 0
(∗) Σ ` ¬(ψ 1 → ¬ψ2 )
Wk
0 0 0 0 0 0 0 0
Σ ∪ {¬ψ1 } ` ψ1 → ¬ψ2 Σ ∪ {¬ψ1 } ` ¬(ψ1 → ¬ψ20 ) 0
¬i
Σ0 `0 ¬¬ψ10
¬e
Σ0 `0 ψ10

Suppose on the other hand that the last step is ∧e2 :

Σ ` ψ1 ∧ ψ2
∧e2
Σ ` ψ2

By induction, Σ0 `0 ¬(ψ10 → ¬ψ20 ). Of course Σ0 ∪ {¬ψ20 } `0 ψ10 → ¬ψ20 (†).

Thus
..
.. induction
0 0 0 0
(†) Σ ` ¬(ψ 1 → ¬ψ2 )
Wk
Σ ∪ {¬ψ2 } ` ψ1 → ¬ψ2 Σ ∪ {¬ψ2 } ` ¬(ψ1 → ¬ψ20 )
0 0 0 0 0 0 0 0 0
¬i
Σ0 `0 ¬¬ψ20
¬e
Σ0 `0 ψ20

• Suppose that the last step is ∨i1 :

Σ ` ψ1
∨i1
Σ ` ψ1 ∨ ψ2

By induction, we know Σ0 `0 ψ10 . From Lemma 0.3.9 we know that Σ0 ∪

{¬ψ10 } is inconsistent, so by Lemma 0.3.6 Σ0 ∪ {¬ψ10 } `0 ψ20 . Introducing
→, we find Σ0 `0 ¬ψ10 → ψ20 , that is Σ0 `0 (ψ1 ∨ ψ2 )0 .
∨i2 is dealt with similarly.

25
 • Suppose the last step of the form:
Σ ` ψ1 ∨ ψ2 Σ ∪ {ψ1 } ` χ Σ ∪ {ψ2 } ` χ
∨e
Σ`χ
By induction, Σ0 `0 ¬ψ10 → ψ20 , Σ0 ∪ {ψ10 } `0 χ0 , and Σ0 ∪ {ψ20 } `0 χ0 . By
contraposition (Theorem 0.3.8, which still holds for `0 ), we easily find:
Σ0 ∪ {¬χ0 } `0 ¬ψ10 (∗1 ) and Σ0 ∪ {¬χ0 } `0 ¬ψ20 (∗2 )
Therefore
..
.. induction
Σ ` ¬ψ10 → ψ20
0 0
(∗1 )
Σ ∪ {¬χ0 } `0 ¬ψ10 → ψ20
0
Σ0 ∪ {¬χ0 } `0 ¬ψ10 (∗2 )
→e
0 0 0 0
Σ ∪ {¬χ } ` ψ2 Σ ∪ {¬χ0 } `0 ¬ψ20
0
¬i
Σ0 `0 ¬¬χ0
¬e
Σ0 `0 χ0

• The ↔ rules are easy to deal with.

All is well that ends well and we may forget about `0 and ϕ0 . From now on
there are two connectives: ¬ and →. The other three are abbreviations. Have I
said that before? It is very important to say it again, and above all, it is crucial
that a redundant connective should be regarded a an abbreviation of the same
expression in either setting (syntactic or semantic).
However we shall be extremely hypocritical about connectives. In our defi-
nitions and inductive proofs, we need only to consider ¬ and →, but whenever
necessary we shall feel free to use ∧, ∨, and ↔ and the related rules. We know
that these are harmless shortcuts, which preserve the notions of truth and of
deduction.
We now aim at simplifying the presentation of deductions.
Notation 0.3.15. When we write a deduction under a set of assumptions Σ, we
write Σ on top (“deduction under Σ”). Whenever an assumption ϕ is eliminated
(via ¬i or →i ), we cross it out.
Let us give an example. Suppose we want to show Σ ` ϕ → ψ. Instead of
Σ ∪ {ϕ} ` . . .
..
.
Σ ∪ {ϕ} ` ψ
→i
Σ`ϕ→ψ
we now write:
Σ ϕ Σ 6ϕ
.. and then ..
. .
ψ ψ
ϕ→ψ

26
As Σ remains uncrossed, one has Σ ` ϕ → ψ.
Similarly,
Σ ∪ {ϕ} ` ψ Σ ∪ {ϕ} ` ¬ψ
¬i
Σ ` ¬ϕ
becomes:
Σ ϕ Σ ϕ Σ 6ϕ Σ 6ϕ
then
ψ ¬ψ ψ ¬ψ
¬ϕ

which means Σ ` ¬ϕ. Do not forget to cross out every relevant occurence of ϕ!
Example 0.3.16. As an example, we prove Excluded Middle (Theorem 0.3.11)
with this notation.
¬(ϕ ∨ ¬ϕ) ϕ
¬(ϕ ∨ ¬ϕ) ϕ ∨i
1. ϕ ∨ ¬ϕ 2.

¬(ϕ ∨ ¬ϕ) 6 ϕ
¬(ϕ ∨ ¬ϕ) 6 ϕ ∨i
∨i ϕ ∨ ¬ϕ
ϕ ∨ ¬ϕ ¬ϕ
¬i
¬i
¬ϕ ∨i
3. ϕ ∨ ¬ϕ 4.

¬(ϕ 6 ∨¬ϕ) 6 ϕ
¬(ϕ 6 ∨¬ϕ) 6 ϕ ∨i
∨i ϕ ∨ ¬ϕ
ϕ ∨ ¬ϕ ¬ϕ
¬i
¬i
¬ϕ ∨i
∨i ϕ ∨ ¬ϕ
ϕ ∨ ¬ϕ ¬i
¬i ¬¬(ϕ ∨ ¬ϕ)
¬¬(ϕ ∨ ¬ϕ) ¬e
5. ϕ ∨ ¬ϕ 6.

0.3.4 The Soundness Theorem

Recall that we have two notions of consequence: semantic (|=, Definition 0.2.9)
and syntactic (`, Definition 0.3.2). Once we have proved the Soundness and
Completeness Theorems (Theorems 0.3.17 and 0.4.1), all will be well that ends
well: they do coincide. The easier part comes first.
Theorem 0.3.17 (soundness of propositional logic). Let Σ be a theory and ϕ
be a wff. If Σ ` ϕ, then Σ |= ϕ.
Proof . Induction on the length of the deduction. We use Corollaries 0.2.18 and
0.3.14, which enable us to treat only the rules associated to ¬ and →.
• If the deduction is just an axiom, it is of the form
Ax
{ϕ} ` ϕ
Clearly {ϕ} |= ϕ.

27
 • The case of weakening is absolutely trivial.
• Suppose that the last step is by ¬i ; that is, Σ ` ¬ϕ because Σ ∪ {ϕ}
is inconsistent. By induction, there are no truth assignments satisfying
Σ ∪ {ϕ}. So any truth assignment satisfying Σ (if any) satisfies ¬ϕ. Thus
Σ |= ¬ϕ (notice that we did not claim satisfiability of Σ).
• Suppose that the last step is
Σ ` ¬¬ϕ
¬e
Σ`ϕ

By induction, we know that Σ |= ¬¬ϕ. Let v be a truth assignment

satisfying Σ. Then v(ϕ) = v(¬¬ϕ) = T , we are done.
• Suppose that the last step of the deduction is the modus ponens:
Σ`ϕ→ψ Σ`ϕ
→e
Σ`ψ

Hence Σ ` ϕ → ψ and Σ ` ϕ hold; by induction, Σ |= ϕ → ψ and Σ |= ϕ.

Let v be a truth assignment satisfying Σ. Then v(ϕ → ψ) = v(ϕ) = T , so
v(ψ) = T . This shows Σ |= ψ.
• Suppose that the last step of the deduction is the modus tollens:
Σ ∪ {ϕ} ` ψ
→i
Σ`ϕ→ψ

By induction, Σ ∪ {ϕ} |= ψ; we aim at showing Σ |= ϕ → ψ. Let v be

a truth assignment satisfying Σ. If v does not satisfy ϕ then v satisfies
ϕ → ψ by definition. If v does satisfy ϕ then v satisfies Σ ∪ {ϕ}; we then
know that v satisfies ψ, and therefore v satisfies ϕ → ψ again. In any case
v satisfies ϕ → ψ. This implies that Σ |= ϕ → ψ.
Even with five connectives the case of ∧ etc. wouldn’t be very complicated.
Induction completes the proof. Exercise 0.11
Exercise 0.12

End of Lecture 5.

Lecture 6 (Completeness of Propositional Logic)

0.4 The Completeness Theorem

Theorem 0.3.17 states that every syntactic consequence is a semantic conse-
quence; roughly speaking, “if you can prove it, it is true”. Very remarkably
(and fortunately), the converse holds.

28
Theorem 0.4.1 (completeness of propositional logic). Let Σ be a theory and
ϕ be a wff. If Σ |= ϕ, then Σ ` ϕ.
Lemma 0.4.2. The Completeness Theorem is equivalent to:
every consistent theory is satisfiable.
Proof . Suppose the Completeness Theorem. Let Σ be a consistent theory.
As Σ is consistent (Definition 0.3.5), there is a wff ϕ such that Σ 6` ϕ. By
completeness, this implies Σ 6|= ϕ. It follows from Lemma 0.2.10 that Σ ∪ {ϕ}
is satisfiable; let v be a truth assignment satisfying Σ ∪ {ϕ}. In particular, v
satisfies Σ!
Suppose that every consistent theory is satisfiable. Let Σ 6` ϕ. Then Σ∪{¬ϕ}
is consistent by Lemma 0.3.9. In particular, Σ ∪ {¬ϕ} is satisfiable. So there is
a truth assignment v satisfying Σ ∪ {¬ϕ}: v satisfies Σ but does not satisfy ϕ,
so Σ 6|= ϕ.

0.4.1 Extending the Theory

We shall prove the Completeness Theorem. In view of Lemma 0.4.2, it suffices
to show that every consistent theory is satisfiable. So, given a consistent set,
we must find a truth assignment satisfying it.
As often in mathematics, choice is embarrassing. We shall force ourself to
have a unique suitable truth assignment. This amounts to maximizing, in a
consistent way, the set of requirements.
Definition 0.4.3 (complete). A consistent theory Σ̂ is complete if for any wff
ϕ, either ϕ or ¬ϕ is in Σ̂.
The following remark will be implicitely used in the proof.
Remark 0.4.4. If Σ̂ is complete, then Σ̂ ` ϕ iff ϕ ∈ Σ̂.
Verification: Suppose Σ̂ ` ϕ, but ϕ 6∈ Σ̂. By completeness, ¬ϕ ∈ Σ̂. So Σ̂ ` ¬ϕ
and Σ̂ ` ϕ, against consistency. ♦
We shall extend any consistent theory to a complete one. As we have said,
the problem of satisfying a complete consistent set is actually easier, as there is
no ambiguity on what truth assignment to chose.
Here is the inductive step.
Lemma 0.4.5. If Σ is a consistent theory and ϕ is a wff, then Σ ∪ {ϕ} or
Σ ∪ {¬ϕ} is consistent.
Proof . Assume that Θ = Σ ∪ {ϕ} is not consistent. By assumption, there is a
wff ψ such that Θ ` ψ and Θ ` ¬ψ. By ¬i , one has Σ ` ¬ϕ. As Σ is consistent,
one cannot have Σ ` ϕ too, so Σ 6` ϕ. So Σ ∪ {¬ϕ} is consistent by Theorem
0.3.10.
Lemma 0.4.6. Let Σ be a consistent theory. Then there is a complete consistent
theory Σ̂ containing Σ.

29
Proof . Recall that WFF is countable (Remark 0.1.3, for instance); enumerate
WFF = {ϕn : n ∈ N}. We shall extend Σ recursively. Let Σ0 = Σ. Assume Σn
has been constructed. By Lemma 0.4.5, Σn ∪ {ϕn } or Σn ∪ {¬ϕn } is consistent.
If Σn ∪ {ϕn } is consistent, let Σn+1 be it; otherwise let Σn+1 = Σn ∪ {¬ϕn }. So
the nth wff ϕn has been taken into account, and Σn+1 is still consistent.
Let Σ̂ = ∪n∈N Σn . Clearly Σ̂ is consistent, and for any wff ϕ, ϕ = ϕn for
some n, so ϕn or its negation is in Σn+1 ⊆ Σ̂. (It is also clear by consistency
that ϕ and ¬ϕ are not both in Σ̂.)

0.4.2 Finding a Truth Assignment

A complete consistent theory induces a unique truth assignment, as for any
sentence symbol An , either An or ¬An is in Σ̂. The following lemma checks
that this happens in a consistent way.
Lemma 0.4.7 (truth lemma). Let Σ̂ be a complete consistent theory. Consider
the truth assignment v : A → {F, T } such that v(An ) = T iff An ∈ Σ̂. Then for
any wff ϕ, v(ϕ) = T iff ϕ ∈ Σ̂.

Proof . By induction. The result is clear if ϕ is a sentence symbol! By com-

pleteness of Σ̂, the case ϕ = ¬ψ is immediate. Now suppose that ϕ is ψ1 → ψ2 .
• Assume v(ϕ) = T . We show ϕ ∈ Σ̂.
If v(ψ1 ) = F , then by induction ψ1 6∈ Σ̂. By completeness of Σ̂, ¬ψ1 ∈ Σ̂.
Hence Σ̂ ∪ {ψ1 } is not consistent; it follows Σ̂ ∪ {ψ1 } ` ψ2 , and therefore
Σ̂ ` ψ1 → ψ2 . By completeness, (ψ1 → ψ2 ) ∈ Σ̂.
If v(ψ2 ) = T , then by induction ψ2 ∈ Σ̂. Therefore Σ̂ ` ψ1 → ψ2 ; by
completeness (ψ1 → ψ2 ) ∈ Σ̂.
• Assume ϕ ∈ Σ̂. We show v(ϕ) = T .
Otherwise, v(ψ1 ) = T and v(ψ2 ) = F . By induction (and completeness
of Σ̂), this shows ψ1 ∈ Σ̂ and ¬ψ2 ∈ Σ̂. If Σ̂ ` ψ1 → ψ2 , then by modus
ponens, Σ̂ ` ψ2 , which contradicts consistency. So Σ̂ 6` ψ1 → ψ2 . This
implies ϕ 6∈ Σ̂.
Induction completes the proof. Notice however that proving the theorem with
five connectives would be essentially as hard, though longer.

Proof of the Completeness Theorem (Theorem 0.4.1). Due to Lemma

0.4.2, it suffices to show that a consistent theory is satisfiable. Let Σ be a
consistent theory. By Lemma 0.4.6, we can extend Σ to a complete consistent
set Σ̂. We consider the truth assignment v induced by Σ̂. By Lemma 0.4.7, v
satisfies Σ̂. In particular v satisfies Σ; Σ is satisfiable.

End of Lecture 6.

Lecture 7 (Compactness of Propositional Logic)

30
0.5 The Compactness Theorem
0.5.1 Compactness
Compactness is the general phenomenon of reduction of the infinite to the finite.
Definition 0.5.1 (finite satisfiability). Let Σ be a theory. Σ is finitely satisfiable
if for all finite Σ0 ⊆ Σ, Σ0 is satisfiable.
Theorem 0.5.2 (compactness of propositional logic). A theory is satisfiable iff
it is finitely satisfiable.
Proof . An implication is trivial: if Σ is satisfiable, then a truth assignment
satisfying it will also satisfy any finite subset.
Let Σ be a finitely satisfiable theory. Any finite fragment of Σ is satisfiable,
hence consistent by soundness (Theorem 0.3.17). So Σ is finitely consistent.
But consistency and finite consistency coincide, as consistency is a finite notion
(our deductions all have finite length). So Σ is consistent, therefore satisfiable
by completeness (Theorem 0.4.1 and Lemma 0.4.2).
Corollary 0.5.3. If Σ |= ϕ, then there is a finite Σ0 ⊆ Σ such that Σ0 |= ϕ.
Proof . By Lemma 0.2.10, Σ∪{¬ϕ} is not satisfiable; by compactness (Theorem
0.5.2) there is a finite subset Σ0 ⊆ Σ such that Σ0 ∪{¬ϕ} is not satisfiable. With
Lemma 0.2.10 again, this means Σ0 |= ϕ.
We now give an example of application.
Definition 0.5.4. A graph is k-colorable if there is a function c from the set
of vertices to {1, . . . , k} such that if {x, y} is an edge, then c(x) 6= c(y).
Recall that a subgraph of a graph is a subset of the set of vertices equipped
with all possible edges.
Theorem 0.5.5 (Erdös). A countable graph is k-colorable iff all of its finite
subgraphs are k-colorable.
Proof . Let V = {vn : n ∈ N} be an enumeration of the vertices. We add
sentence symbols Cv,i for v ∈ V and i ∈ {1, . . . , k} (their meanings will be:
c(v) = i, that is v bears the ith color).
Consider the theory made of:
• Cv,1 ∨ · · · ∨ Cv,k for each v ∈ V (meaning: each vertex has a color)
• ∧i6=j 6= Cv,i ∧ Cv,j for each v ∈ V (meaning: no vertex has two colors)
• ∧ki=1 6= Cv,i ∧ Cv0 ,i for adjacent vertices v, v 0 (meaning: no two adjacent
vertices bear the same color)
By assumption, the theory is finitely satisfiable. By compactness (Theorem
0.5.2), it is satisfiable: there exists a global coloring.
Still not impressed? Wait and see what we shall do with first-order logic. Ex. 0.14, 0.17

31
0.5.2 Applications to Decidability*
Compactness is very useful is the context of decidability, where one wants to
decide whether or not a given wff is a consequence of a theory. We consider this
problem. For the question to make sense, we need of course to have the data in
a form readable by a machine. This is approached by the following definition.
Definition 0.5.6 (decidable). A set is decidable if there is an algorithm which
can decide whether or not a given object belongs to it.
Example 0.5.7. The set of wff’s is decidable, as there is an algorithm which
determines whether a given formula is decidable or not.
Actually if one just wants to list all possible elements of the set, decidability
is a little too strong, and could be weakened as follow.
Definition 0.5.8 (semi-decidable). A set is semi-decidable if there is an algo-
rithmic way to write a list such that sooner or later, every element of the set
will appear on the list.
One also says effectively enumerable for semi-decidable. This does not assert
that there is an algorithm deciding whether or not an object is in the set; all you
can get is an algorithm which will answer “yes” if the object lies in it. Otherwise,
the algorithm doesn’t stop, as it scans through an infinite list without finding
an answer.
Example 0.5.9. Suppose that you are looking for Jorge Luis Borges in an
infinite phone directory.
• If the phone book is in alphabetical order, you start reading. If you reach
Byron without passing through Borges, then Jorge Luis Borges is not
in the phone book. If you reach Borges, Julio without passing through
Borges, Jorge, then Jorge Luis Borges is not in the phone book. Other-
wise you have reached Jorge Luis Borges. The phone book is decidable.
• But now suppose that the phone book is not in alphabetical order. The
only way to find out if Borges is in the phone book is to read through. If
you reach the name Borges, Jorge Luis, then he is in the phone book.
Otherwise, you cannot know whether he is not in the directory, or he is
but you haven’t found his name yet. The phone book is semi-decidable.
Remark 0.5.10. A set is decidable iff both it and its complement are semi-
decidable.
We now are ready to explain what a “readable” theory is.
Definition 0.5.11 (axiomatization). An axiomatization of a set of wff’s Σ is a
set Σ0 such that Σ0 |= Σ and Σ |= Σ0 .
Definition 0.5.12 (finitely, decidably, semi-decidably axiomatizable theory).
If there exists a finite, decidable, or semi-decidable axiomatization, the theory
is said finitely, decidably, semi-decidably axiomatizable, accordingly.

32
 A typical use of Compactness is the following result.
Proposition 0.5.13. If Σ is finitely axiomatizable, then every axiomatization
of Σ contains a finite axiomatization.
Proof . Let Σ0 be a finite axiomatization of Σ; taking the conjunction, we may
assume that Σ0 = ϕ is a single formula.
Let Σ0 be any axiomatization of Σ. Then Σ0 ∪ {¬ϕ} is not consistent, as
Σ |= Σ |= ϕ. So Σ0 |= ϕ. By compactness, there is a finite subset of Σ0 , which
0

taking conjunction we may assume to be a formula ψ, such that ψ |= ϕ |= Σ.

We now state the results which interest us.

Corollary 0.5.14. Let Σ be a semi-decidably axiomatisable theory and ϕ a wff.

Then there is an algorithm which will answer “yes” if Σ |= ϕ.
Proof . If Σ |= ϕ, then by Compactness (Theorem 0.5.2, or rather Corollary
0.5.3), there is a finite Σ0 ⊆ Σ such that Σ0 |= ϕ. We therefore list all finite
fragments Σ0 of Σ and apply the finite case algorithm of Theorem 0.2.11 to
each. If Σ |= ϕ, then we shall be answered “yes” in a finite time.
In other words, if Σ admits a semi-decidable axiomatization, then Σ itself,
as a set of wff’s, is semi-decidable.
Remark 0.5.15. The algorithm cannot be implemented to say “no”, even if
Σ is decidably axiomatizable. If after a very long time the algorithm hasn’t
answered yes, it might be for two reasons:
• you haven’t waited long enough yet;
• Σ 6|= ϕ (in which case the algorithm will never stop).
Hence, in general, the set of semantic consequences of a semi-decidable Σ is
only semi-decidable (as opposed to the finite case, Theorem 0.2.11, where it is
decidable). However, if the theory is maximal in some sense, there is more.
Corollary 0.5.16. Let Σ be a semi-decidable theory in a decidable language.
Assume that for any sentence ϕ, either Σ |= ϕ or Σ |= ¬ϕ. Then there is an
algorithm which decides whether Σ |= ϕ or not.

Proof . Run two instances of the algorithm of Corollary 0.5.14: one for ϕ and
one for ¬ϕ. One of them will sooner or later answer “yes”.
We do not insist further on decidability; our notion of algorithm is to remain
informal. Decidability and Computability are other branches of Mathematical
Logic, with very interesting results.

33
0.5.3 A Topological Proof*
Compactness phenomena first arose in the context of topology. Informally
speaking, a space is compact if whenever one can always find an element meet-
ing a finite number of requirements (lying in the intersection of finitely many
members), then there is an element meeting all requirements simultaneously.
And this is exactly what Theorem 0.5.2 says; it actually states the compact-
ness of a certain topological space. We therefore give an alternate proof of the
compactness theorem.
Definition 0.5.17 (finite intersection property). A family F has the finite
intersection property if the intersection of any finite number of sets in F is
nonempty.
Definition 0.5.18 (compactness). A topological space X is compact if for all
family of closed sets F with the finite intersection property, ∩F 6= ∅.
For instance, a finite space is always compact.

Theorem 0.5.19 (Tychonoff). A product of compact spaces is compact for the

product topology.
Corollary 0.5.20. {0, 1}N is compact for the product topology.
We now introduce a topology on the set of truth assignments.

Notation 0.5.21.
• Let S = {F, T }N be the set of truth assignments.
• For every wff ϕ, let Oϕ = {v ∈ S : v(ϕ) = T }.
• Let τ be the topology generated by the Oϕ ’s.

Proposition 0.5.22. τ is the product topology on S. S is Hausdorff, compact,

and totally disconnected. The Oϕ ’s are exactly the clopen sets of S; they form
a Boole algebra.
In particular, there is a correspondence between sets of wff’s and closed sets:
\
Σ 7→ CΣ = Oϕ
ϕ∈Σ

Remark 0.5.23. This correspondence is not injective (as {ϕ} and {¬¬ϕ} define
the same closed set), but it is surjective since any closed set is an intersection
of clopen subsets.
Lemma 0.5.24. Let Σ be a theory and v be a truth assignment. Then v ∈ CΣ
iff v satisfies Σ.

34
Proof .
v ∈ CΣ iff for all ϕ ∈ Σ, v ∈ Oϕ
iff for all ϕ ∈ Σ, v(ϕ) = T
iff v satisfies Σ.

Corollary 0.5.25. Σ is satisfiable iff CΣ 6= ∅.

Proof of the Compactness Theorem, Theorem 0.5.2. Let Σ be a finitely

satisfiable set of wff’s. Consider F = {Oϕ : ϕ ∈ Σ}, a family of closed subsetes
of S. As Σ is finitely satisfiable, F has the finite intersection property. By
compactness of S, ∩F 6= ∅. So CΣ 6= ∅ and Σ is satisfiable.
Remark 0.5.26. The only problem with this proof is that it is not effective,
as opposed to the one relying on the completeness theorem. But on the other
hand, it is free of references to a proof theory.

End of Lecture 7.

Lecture 8 (Modal Logic*)

0.6 A Little Modal Logic*

We say a word of modal logic. Modal logic is an extension of propositional logic
taking necessity into account. (One could also formalize belief, provability, etc.)
Syntactically speaking, this is done by adding adverbs, which will be unary
connectives. Semantically this is a little more subtle. We need to deal with
several “worlds”, which will describe the possibilities. In particular a statement
will be necessary if true in all possible worlds.
This notion of possibility requires a treatment of imaginability: some worlds
(not necessarily all) are conceivable from others. The notion of a Kripke model
will be detailed in Definition 0.6.2; for the moment we precise the language.
Definition 0.6.1 (language of modal logic). To the language we add two unary
connectives:  and . They denote necessity and possibility, respectively. We
close WFF under  and , getting the set WFFmod .

For instance (A1 → A2 ) is a wff of modal logic, which reads “It is
necessary that the possibility of A1 implies the necessity of A2 ”. Similarly, ¬ϕ
reads “ϕ is contingent”; ¬  ϕ reads “ϕ is impossible”.

0.6.1 Semantics*
The semantics is of course a little more complex than before. As mentioned, we
need a whole collection of worlds, accessible by imagination from each other.
Definition 0.6.2 (Kripke model). A Kripke model is a triple (W, R, v) where

35
 • W is a non-empty collection of worlds;
• R is a binary relation on W ; w1 Rw2 means that w2 is conceivable in w1 ;
• v is a function from A × W to {T, F }.

Caution! Truth values depend on the world. v(ϕ) is now meaningless.

Remark 0.6.3. In Definition 0.6.2, we could require specific axioms on the
accessibility relation R in order to formalize a description of what our intuition
suggests that conceivability means.
Notation 0.6.4. We extend v inductively to WFFmod ×W :
• v(An , w) = v(An , w) for each n ∈ N and each world w

• v(¬ϕ, w) = T if v(ϕ, w) = F
• v(ϕ → ψ, w) = T if v(ϕ, w) = F or v(ψ, w) = T
• v(ϕ, w) = T if for all world w0 such that wRw0 , one has v(ϕ, w0 ) = T
• v(ϕ, w) = T if there is a world w0 such that wRw0 and v(ϕ, w0 ) = T .

The meaning is clear: ϕ (“ϕ is necessary”) holds in the world w if for any
world accessible from w, ϕ holds. (An observer from w will then believe that ϕ
cannot ever fail.) On the other hand, ϕ (“ϕ is possible”) holds in w if there
is a conceivable world w0 in which ϕ holds: the observer of w will think that ϕ
could hold somewhere.
One can of course reduce the language to use the sole modal connective ;
clearly  stands for ¬¬; we implicitely do so and forget about .
Definition 0.6.5 (satisfaction in modal logic). Σ |= ϕ if for all models (W, R, v)
such that Σ is satisfied in all worlds of W , ϕ is also satisfied in all worlds of W .
Example 0.6.6. One has {ϕ} |= ϕ.

This is somehow problematic, as one should not allow anything similar on

the proof-theoretic side. So we shall have to restrict the completeness theorem
(Theorem 0.6.10 below).
The safest is to consider that Example 0.6.6 is not meaningful. We then
agree never to use Σ |= ϕ, but to focus on the case |= ϕ.

0.6.2 Proof Theory*

The proof theory is very much alike. We add two rules.

• Kripke’s Rule:
Σ ` (ϕ → ψ)
K
Σ ` ϕ → ψ

36
 • Necessitation Rule:
`ϕ
N
` ϕ

Caution! There are no assumptions in rule N , i.e. Σ must be the empty set.
Otherwise one finds
{ϕ} ` ϕ
N
{ϕ} ` ϕ
→i
` ϕ → ϕ
in other words, “everything which holds is necessary”, a clear insanity.
Recall however that {ϕ} |= ϕ; in particular, we shall not have “full” com-
pleteness, in the sense that Σ |= ϕ will be stronger than Σ ` ϕ. Completeness
will hold only for Σ = ∅.
Remark 0.6.7. Notice that if we have made specific requirements on the “ac-
cessibility” relation R (see Remark 0.6.3), we might need to add a couple of
extra rules.
Example 0.6.8. In Definition 0.6.2, we impose on the accessibility relation to
be reflexive. Then |= (ϕ) → ϕ. As there is no way to prove it, we admit it as
an axiom, or equivalently we add the following rule:
Σ ` ϕ
Σ`ϕ
Theorem 0.6.9 (soundness of modal logic). If Σ ` ϕ, then Σ |= ϕ.
Proof . A quick induction.
• The case of the rules of Propositional Logic is essentially as in Theorem
0.3.17. For
Σ ∪ {ϕ} ` ψ
→i
Σ`ϕ→ψ
you might need to introduce W1 = {w in W : v(ϕ, w) = T } and W2 =
{w in W : v(ϕ, w) = F }, and prove the claim piecewise.
• Suppose that the last step is
`ϕ
N
` ϕ
(Recall that the Necessitation rule has no assumptions).
Let (W, R, v) be any Kripke model. We show that for any world w of W ,
we have v(ϕ, w) = T .
By induction, |= ϕ, so our Kripke model satisfies ϕ: that is, for any w0 in
W , one has v(ϕ, w0 ) = T . In particular, for any w0 such that wRw0 , one
has v(ϕ, w0 ) = T . Hence v(ϕ, w) = T .
As w is arbitrary, the Kripke model satisfies ϕ. Since we have taken any
Kripke model, we deduce that |= ϕ.

37
 • Now suppose that the last step is

Σ ` (ϕ → ψ)
K
Σ ` ϕ → ψ

We show Σ |= ϕ → ψ. So let (W, R, v) be a Kripke model satisfying Σ;

we show that it satisfies ϕ → ψ.
Fiw a world w. We want to show that v(ϕ → ψ, w) = T . We may
suppose v(ϕ, w) = T (as otherwise we are done). By definition, for all w0
such that wRw0 , one has v(ϕ, w0 ) = T . We aim at showing v(ψ, w) = T ,
that is v(ϕ, w0 ) = T for any world w0 accessible from w. Let w0 be such
that wRw0 .
By induction, we know Σ ` (ϕ → ψ). In particular, for any world w0
with wRw0 , one has v(ϕ → ψ, w0 ) = T . As v(ϕ, w0 ) = T , it follows
v(ψ, w0 ) = T . Since this is true for all w’ accessible from w, one has
v(ψ, w) = T .
One concludes that Σ |= ϕ → ψ.

0.6.3 Completeness*
Theorem 0.6.10 (completeness of modal logic). If |= ϕ, then ` ϕ.
Counter-example 0.6.11. This is not true with a non-empty set of assump-
tions, as {ϕ} |= ϕ, but {ϕ} `
6 ϕ.

Proof of Theorem 0.6.10. The proof will rely on the construction of a uni-
versal Kripke model (Wun , Run , vun ).
• Let Wun be the collection of all maximal consistent modal theories.
• Write wRun w0 iff {ϕ : ϕ ∈ w} ⊆ w0 .

• Let vun : A × Wun → {T, F } be such that vun (An , w) = T iff w ∈ An .

Notice that Wun is not empty by Lemma 0.4.6, or something similar in WFFmod .
Claim (truth lemma; see Lemma 0.4.7). For any ϕ ∈ WFFmod and any maxi-
mal consistent modal theory w, vun (ϕ, w) = T iff ϕ ∈ w.
Verification: By induction. The only non-trivial case is ϕ = ψ.
• Suppose ϕ ∈ w. Let w0 be any world such that wRun w0 . We assumed ψ ∈
w, so by definition of Run , one has ψ ∈ w0 . By induction, vun (ψ, w0 ) = T .
This being true for any w’ with wRun w0 , we deduce vun (ψ, w) = T , that
is vun (ϕ, w) = T .

38
 • We now suppose vun (ϕ, w) = T ; we show that ϕ ∈ w. Consider Nw =
{χ : χ ∈ w}. We form the modal theory Θ = Nw ∪ {¬ψ}.
If Θ is consistent, then it can be extended to a maximal consistent modal
theory w0 . Hence Nw ⊆ w0 and by definition of the relation, wRun w0 . By
assumption, vun (ψ, w) = T , so vun (ψ, w0 ) = T , against ¬ψ ∈ w0 .
So Θ is not consistent. It follows Nw ` ψ. There are therefore χ1 , . . . , χn ∈
Nw such that {χ1 , . . . , χn } ` ψ. We find

{χ1 , . . . , χn } ` ψ
→i
` χ1 → (χ2 → (. . . (χn → ψ)...)
N
`  (χ1 → (χ2 → (. . . (χn → ψ)...))
K
` χ1 → (χ2 → (. . . (χn → ψ)...)

As χ1 , . . . , χn are in w and w is maximal consistent, we see ψ ∈ w. ♦

We now prove completeness of modal logic. Suppose 6` ϕ. So there is a
maximal consistent modal theory w containing ¬ϕ. In particular, vun (¬ϕ, w) =
T ; the universal model Wun does not satisfy ϕ. Hence 6|= ϕ.

Remark 0.6.12.
• Of course if we have made restrictions the accessibility relations (and clev-
erly added the corresponding deduction rules), we need to prove complete-
ness accordingly, that is we must show that the accessibility relation Run
on the universal model Wun has the desired properties.

• Notice further that we cannot show that if Σ 6` ϕ, then Σ 6|= ϕ (see

Counter-example 0.6.11). The proof fails for the following reason. We
naturally restrict our universal model to the set WΣ of maximal consistent
extensions of Σ; of course WΣ ⊆ Wun , and we take the induced accessibil-
ity relation and truth assignment. When trying to prove the truth lemma
for WΣ , we shall form Θ = Σ ∪ Nw ∪ {¬ψ}; it is not consistent; we think
we are very clever.
Hence we find Σ ∪ {χ1 , . . . , χn } ` ψ. But now even if we detach a formula
σ ∈ Σ such that {σ, χ1 , . . . , χn } ` ψ, we will end up with σ → (χ1 →
. . . (χn → ψ)...), and there is no reason for σ to be in Σ!

This discussion reveals that the universal model WΣ satisfies the truth lemma
only if Σ is closed under . This yields the following generalization of Theorem
0.6.10.
Theorem 0.6.13 (completeness of modal logic revisited). Let Σ be a modal
theory which is closed under . If Σ |= ϕ, then Σ ` ϕ.

In particular, by a method similar to that of Lemma 0.4.2, we find the dual

statement.

39
Theorem 0.6.14 (completeness of modal logic revisited). Let Σ be a modal
theory which is closed under . If Σ is consistent, then Σ is satisfiable.
Notice that the resulting compactness phenomenon does not require closure
under .
Corollary 0.6.15 (compactness of modal logic). Let Σ be a modal theory. Then
Σ is satisfiable iff it is finitely satisfiable.
Proof . Suppose Σ is finitely satisfiable; we prove that it is satisfiable (the other
implication being trivial).
Let Σ0 be the closure of Σ under : Σ0 is the smallest subset of WFFmod
which contains Σ and such that ϕ ∈ Σ0 =⇒ ϕ ∈ Σ0 . As for any ϕ ∈ WFFmod ,
{ϕ} ` ϕ, it is clear that Σ0 is finitely satisfiable.
By soundness of modal logic (Theorem 0.6.9), Σ0 is consistent. By complete-
ness (Theorem 0.6.14), Σ0 is satisfiable. As Σ ⊆ Σ0 , so is Σ.
However interesting epistemologically, one sees on the technical side that
modal logic is but an easy extension of propositional logic. We shall return to
more mathematical views.
End of Lecture 8.
Lecture 9 (ME1)
End of Lecture 9.

0.7 Exercises
When not otherwise specified, we work in classical logic (¬e allowed).
Exercise 0.1. Rewrite the following using only ¬ and →:

2. ϕ ∨ ψ 3. ϕ ↔ ψ
1. ϕ ∧ ψ

The idea is to show that our connectives are (intuitively speaking) redundant.
We shall show that they are formally redundant.
Exercise 0.2. The language has four symbols [, ], {, and }. The collection of
nice formulas is the smallest set of expressions such that:
• the empty/blank expression “” is a nice formula;
• if A, B are nice formulas, [A]{B} is a nice formula.

1. An expression is bracket-balanced if it has as many [’s as ]’s. We define

brace-balanced similarly. Show that every nice formula is bracket-balanced
and brace-balanced. (Notice that there are expressions which are bracket-
balanced and brace-balanced but not nice.)

40
 2. An expression is globally balanced if it is both bracket-balanced and brace-
balanced, and has as many [’s as {’s. Show that every nice formula is
globally balanced.
3. Show that no proper initial segment of a nice formula is globally balanced.
Deduce that no proper initial segment of a nice formula is a nice formula.

Exercise 0.3. Are the following theories satisfiable?

1. Σ1 = {A1 , A1 → A2 , A2 → A3 , A1 ∧ A2 , A2 ∨ ¬A3 }
2. Σ2 = {(A1 → A2 ) → A1 , ¬A1 }
3. Σ3 = {A3i ∧ A3i+1 → ¬A3i+2 : i ∈ N} ∪ {A2i+1 → ¬A2i+2 : i ∈ N}
Exercise 0.4. Let ϕ be a wff formula which uses as connectives only ∧ and
∨. Let v be the truth assignment such that v(An ) = T for all n. Show that
v(ϕ) = T .
Exercise 0.5.
1. Show that Σ ` ϕ → ψ iff Σ ∪ {ϕ} ` ψ.

2. Show that Σ ` ϕ ↔ ψ iff Σ ∪ {ϕ} ` ψ and Σ ∪ {ψ} ` ϕ.

Exercise 0.6. Show the following theorems of propositional logic:
1. ` ¬(ϕ ∨ ψ) ↔ (¬ϕ ∧ ¬ψ)
2. ` (¬ϕ ∨ ¬ψ) → ¬(ϕ ∧ ψ)

3. ` ¬(ϕ ∧ ψ) → (¬ϕ ∨ ¬ψ)

You may use excluded middle only for 3.
Exercise 0.7 (Peirce’s law). Show the following theorem of classical logic:
` ((ϕ → ψ) → ϕ) → ϕ.

Exercise 0.8.
1. Write a deduction of ` ϕ → ¬¬ϕ.
2. Show that ` ¬¬(ϕ ∨ ¬ϕ) (without using ¬e nor excluded middle).

Exercise 0.9 (⊥). Let ⊥ (“perp” or “bottom”) denote contradiction. We in-

troduce related deduction rules:
Σ ` ϕ Σ ` ¬ϕ Σ `⊥
⊥i ⊥e
Σ `⊥ Σ`ϕ

1. Show that Σ is inconsistent iff Σ `⊥.

2. We inductively remove ¬ from wff’s:

41
 • A0n = An for each sentence symbol;
• (¬ϕ)0 = (ϕ0 →⊥);
• (ϕ → ψ)0 = (ϕ0 → ψ 0 ).
We extend the notion of deduction ` to also include the ⊥ rules. Show
that Σ ` ϕ iff Σ0 ` ϕ0 .
As we know that other connectives may be coded in terms of ¬ and →, one sees
that all connectives can be coded by → and ⊥. A purely ⊥, → formulation of
¬e is
` ((ϕ →⊥) →⊥) → ϕ
Exercise 0.10 (Double negation, Excluded middle and Reductio ad Absur-
dum). Consider the rules:

Σ ` ¬¬ϕ Σ ∪ {¬ϕ} `⊥
¬e RAA EM
Σ`ϕ Σ`ϕ ` ϕ ∨ ¬ϕ

Show that all three are equivalent, i.e. taking one as a deduction rule allows to
establish the other two.
Exercise 0.11. Prove the soundness theorem by induction with all five con-
nectives.
Exercise 0.12. Let CONS be the property:
Every satisfiable theory is consistent.
Show that soundness is equivalent to CONS.
Exercise 0.13. Let FSAT be the property:
Every finite, consistent theory is satisfiable.
Show that completeness is equivalent to compactness and FSAT.
Exercise 0.14. Let X be a countable set and  a partial ordering on X ((X, )
is a poset):
• ∀x xx
• ∀x ∀y xy∧y x→x=y
• ∀x ∀y ∀z x  y ∧ y  z → x  z.
Show that there is a linear ordering ≤ on X extending , that is:
• ≤ is an ordering on X
• ≤ is linear: ∀x ∀y x≤y∨y ≤x
• ≤ extends : ∀x ∀y x  y → x ≤ y.

42
Hint: When X is finite, an easy induction. When X is countable, use compact-
ness of propositional logic.
Exercise 0.15. A map is a countable partition of the plane satisfying certain
intuitive properties we do not wish to formalize. It is 4-colourable if there is
a function which assigns to each region a colour, such that no two adajacent
regions bear the same colour. Show that a countable map is 4-colourable iff
each of its finite submaps is.
Exercise 0.16. A graph is k-colorable if there is a function c from the set of
vertices to {1, . . . , k} such that if {x, y} is an edge, then c(x) 6= c(y). Show that
a countable graph is k-colorable iff all of its finite subgraphs are k-colorable.

Exercise 0.17.
1. Let {Sn : n ∈ N} be a collection of finite subsets of N with the following
property: for each finite F0 ⊆ N there exists A0 ⊆ N such that for all
n ∈ F0 , Card(A0 ∩ Sn ) = 1. Prove that there exists A ⊆ N such that for
all n in N, Card(A ∩ Sn ) = 1.

2. Give a counterexample if the Sn ’s are not required to be finite.

Exercise 0.18. The purpose of this exercise is to discover intuitionism. We
say that Σ `IL ϕ if there is a deduction using ¬i , ⊥i , ⊥e , but not ¬e nor its
equivalents RAA and EM.

1. Show `IL ϕ → (¬¬ϕ). In general the converse does not hold.

2. Show that `IL (¬¬ ⊥) →⊥.
3. Show `IL (¬¬¬ϕ) → (¬ϕ).

Hence the ¬ operation is not involutive in IL, but quickly becomes involutive.
Exercise 0.19 (Gödel’s translation). We work only with → and ⊥, as we know
(Exercise 0.9) that they are sufficient to encode all connectives.
We define inductively the Gödel translation of a wff
• A∗n = ¬¬An , An a sentence symbol (including ⊥)

• (ϕ → ψ)∗ = ϕ∗ → ψ ∗

1. Show that {ϕ} `CL ϕ∗ and {ϕ∗ } `CL ϕ.

2. Show that if `IL ϕ∗ , then `CL ϕ.

3. Show that if `CL ϕ, then `IL ϕ∗ .

It follows that Intuitionnistic Logic is more subtle than Classical Logic: it has
the ability to encode all of classical logic, and it perceives differences that CL
doesn’t. (It is weaker only when viewed from CL.)

43
Exercise 0.20 (some modal logic). We investigate further the accessibility
relation. A logic (|=, `) is tautology-complete if whenever |= ϕ, then ` ϕ.
1. Suppose that in any frame, the accessibility relation is reflexive (for w in
W , wRw), and that modal logic remains sound and tautology-complete.
Show that the following deduction is now valid:
` ϕ
`ϕ

2. Suppose that the following is now a deduction rule:

` ϕ
`ϕ

Show that in the canonical model (constructed in the proof of the com-
pleteness theorem), the accessibility relation is now reflexive.
3. Deduce that if one considers only relfexive accessibility relations, and al-
lows the above rule, then modal logic remains tautology-complete.

4. Same questions with the rule

` ϕ
` ϕ

and the property of transitivity (wRw0 and w0 Rw00 entail wRw00 ).

5. Find the proof-theoretic dual of symmetry of (wRw0 entail w0 Rw).

44
Chapter 1

First-Order Logic (Largo)

First-order Logic, also called Predicate Logic, is more expressive than Propo-
sitional Logic: it can mention elements. This requires extending the language;
in particular we shall need quantifiers. Up to an abbreviation, we use only one.
But this quantifier deeply affects the notion of deduction; we need new rules.
On the other hand, the meaning of first-order languages is also richer. We shall
introduce structures, which provide their semantic notion of entailment.
If soundness remains easy to show, completeness of first-order logic will be
a major and non-trivial result. Hence, though quite expressive, first-order logic
enjoys excellent properties.

In this chapter:

• Define first-order languages (§1.1)

• Introduce the structures associated to first-order languages (§1.2)
• An unpleasant journey to the theory of substitutions (§1.3)

• Extend the notion of deduction to first-order logic (§1.4)

• Prove a soundness theorem for first-order logic (§1.4.3)
• Prove Gödel’s completeness theorem for first-order logic (§1.5)
• Derive compactness (§1.6.2) and find applications (1.6.2)

Lecture 10 (First-Order Languages; Terms and Formulas)

1.1 Syntax
First-order is essentially more expressive than Propositional Logic: we even-
tually introduce elements. So one will need quantification, but also specific

45
elements (constants), properties of elements (relations), and a way to build on
elements (functions).

1.1.1 First-order Languages

We shall from the start cut down on the number of connectives used; this will
make inductive definitions and proofs way shorter, and is legitimate in view of
the results of §0.2.3 and 0.3.3.
Definition 1.1.1 (first-order language). A first-order language L consists of
the following symbols:
• Logical symbols, which are common to all languages:

– “(” and “)”

– the connectives ¬ and →
– a set V of variables v1 , v2 , . . . , vn , . . .
– the quantifiers ∀ and ∃
– the symbol “=”, which is a binary relation symbol
• Specific symbols, which depend on the language (these form what is called
the signature of the language):
– a set C of constant symbols
– for each n ≥ 1, a set R of n-ary relation symbols
– for each n ≥ 1, a set F of n-ary function symbols
Of course one defines the set of expressions of L.
Remark 1.1.2. When giving a first-order language, one does not specify logical
symbols. In particular, even if it is not mentioned, the symbol “=” is always
part of the language.
Remark 1.1.3. It is technically possible to have only relation symbols in the
signature: one replaces constant symbols by unary relation symbols (which
stand for singletons), and n-ary function symbols by n + 1-ary relations (the
graphs of the functions). It is more convenient to allow constant and function
symbols, but bear in mind that we would be able to work with purely relational
languages throughout.
Notation 1.1.4. One immediately allows the following abbreviations (E1 , E2
denote arbitrary expressions):

• E1 ∧ E2 stands for ¬(E1 → ¬E2 );

• E1 ∨ E2 stands for (¬E1 ) → E2 ;
• E1 ↔ E2 stands for ¬((E1 → E2 ) → ¬(E2 → E1 ));

46
 • E1 6= E2 stands for ¬(= (E1 , E2 )).
This should save us precious amounts of time later, though one could in-
troduce five connectives (as we clumsily did for Propositional Logic before the
verifications of §0.2.3 and 0.3.3).
Remark 1.1.5. On the other hand, we have two quantifiers. By arguments
ressembling those of §0.2.3 and 0.3.3 (see Corollaries 0.2.18 and 0.3.14), we will
get rid of ∃, showing that it may be considered as a mere abbreviation. This
will be done in §1.4.2 (Corollaries 1.4.8 and 1.4.11).
Example 1.1.6.
• The empty language, or language of pure sets, contains the equality as its
only relation symbol.
• The language of orderings is Lord = {<}, with < a binary relation symbol
(we omit to mention =).
• The language of groups is Lgrps = {1, ·,−1 } where 1 is a constant symbol,
· is a binary function symbol, −1 is a unary function symbol.
• The language of rings is Lrings = {0, 1, +, −, ·} where 0 and 1 are constant
symbols, +, −, · are binary function symbols.

1.1.2 Terms
We have defined expressions, but only well-formed formulas are of interest. Yet
before we can define them, we need to explain which combinations of elements
are meaningful.
Definition 1.1.7 (term). The collection of L-terms is the smallest set such
that:
• every variable is a term;
• every constant symbol is a term;
• if t1 , . . . , tn are terms and f is a n-ary function symbol, then f (t1 , . . . , tn )
is a term.
Exercise 1.1

Notation 1.1.8 (variables occuring in a term). For an L-term t, we define the

set Var(t) of variables occuring in t:
• if t is a constant symbol c, then Var(t) = ∅;
• if t is a variable x, then Var(t) = {x};
• if t is f (t1 , . . . , tn ), where t1 , . . . , tn are terms and f is a n-ary function
symbol, then Var(t) = Var(t1 ) ∪ . . . . . . Var(tn ).

47
 It is quite clear that a variable x occurs in a term t iff the symbol x appears
in the expression t.
Example 1.1.9.
• ·(1, ·(v1 , 1)) is a term of Lgrps , which is more conveniently written 1·(v1 ·1).
Only v1 occurs in it.
• 2 · v12 − 3 · v2 is a term of Lrings , as we have implicitely made the following
abbreviations: 2 stands for +(1, 1), 3 for +(1, +(1, 1)), and v12 for ·(v1 , v1 ).
The variables v1 and v2 occur in the term.
No meaning (and not much interest) so far! We start building on elements
to create well-formed formulas.

1.1.3 Formulas
Definition 1.1.10 (atomic formula). An atomic formula of L is an expression
of the form R(t1 , . . . , tn ), where R is an n-ary relation symbol of L and t1 , . . . , tn
are L-terms.
Example 1.1.11.
• v1 = v2 is an atomic formula of any first-order language (as equality is
always part of the language).

• x2 + 1 = 0 is an atomic formula of Lrings .

Definition 1.1.12 (well-formed formula). The collection WFFL of well-formed
formulas of L is the smallest set such that:
• every atomic formula is a wff;

• if ϕ and ψ are wff’s, so are ¬ϕ and ϕ → ψ;

• if ϕ is a wff and x is a variable, then ∀x ϕ and ∃x ϕ are wff’s.
From now on, “formula” will implicitely mean “well-formed formula”.
Example 1.1.13.

• ∀v1 ∃v2 (v1 · v2 = 1 ∧ v2 · v1 = 1) is (short-hand for) a wff of Lgrps .

• ∀v1 1 = 0 is a wff of Lrings .
• ∀v1 ∀v1 v1 = v1 is a wff of any first-order language.

Theorem 1.1.14 (unique readability). There is only one way to read a wff.
Proof . Proceed exactly like for propositional logic, Theorem 0.1.9. This re-
quires a version of the balanced parenthesing lemma, Lemma 0.1.11.

48
Notation 1.1.15 (variables occuring in a formula). For an L-wff ϕ, we define
the set Var(ϕ) of variables occuring in ϕ:
• if ϕ is an atomic formula R(t1 , . . . , tn ), then Var(ϕ) = Var(t1 ) ∪ · · · ∪
Var(tn );
• if ϕ is ¬ψ, then Var(ϕ) = Var(ψ);
• if ϕ is ψ1 → ψ2 , then Var(ϕ) = Var(ψ1 ) ∪ Var(ψ2 );
• if ϕ is ∀x ψ or ∃x ψ, with x ∈ V, then Var(ϕ) = Var(ψ) ∪ {x}
Again, it is clear that a variable x occurs in a formula ϕ iff the symbol x
appears in the expression ϕ.
It is clear from Example 1.1.13 that when working syntactically we will have
to be extremely cautious with variables. This will give rise to several painful
technicalities, such as the notion of substitutability, to which the entire §1.3 is
unfortunately devoted.
Notice however that no one is stupid enough to make such mistakes at the
semantic level. So once we are done with the syntactic study of first-order
language, we shall happily forget the technical details.
Definition 1.1.16 (free or bound variable). Let ϕ be a wff and x a variable.
• if ϕ is atomic, then x occurs free in ϕ if it occurs in ϕ
• if ϕ is ¬ψ, then x occurs free in ϕ if it occurs free in ψ
• if ϕ is ψ1 → ψ2 , x occurs free in ϕ if it occurs free in ψ1 or in ψ2
• if ϕ is ∀y ψ or ∃y ψ, x occurs free in ϕ if x occurs free in ψ and x 6= y.
If x occurs in ϕ but is not free, it is said to occur bound in ϕ.
Example 1.1.17.
• In ∀v1 v1 = v2 , v1 is bound and v2 is free.
• The same holds of ∀v1 ∀v1 v1 = v2 (which is unfortunately a wff).
• In (∀v1 v2 = v3 ) → (v1 = v2 ), v1 , v2 and v3 occur free.
Hence free means: “we’re still missing something to know what it is about”.
Notation 1.1.18. For a formula ϕ, we let FreeVar(ϕ) denote the set of free
variables in ϕ.
Definition 1.1.19 (sentence). A sentence is a wff with no free variable.
Remark 1.1.20. In a sense, in propositional logic all (well-formed) formulas
are sentences, as there are no variables.
Definition 1.1.21 (theory). A theory is a set of sentences.

49
 This definition is the most natural one (we do not want “unexplained vari-
ables”), but we shall need to be very careful.
• Some subtleties around the notion of consistency will arise in §1.4.
• We shall define Σ |= ϕ and Σ ` ϕ for sets of formulas, not of sentences, as
this will be useful later.
Example 1.1.22.
• The theory of infinite sets (in Lsets ) is

{∃v1 . . . ∃vn ∧i6=j vi 6= vj : n ∈ N}

• The theory of orderings (in Lord ) is

 
∀v1 ¬(v1 < v1 ), ∀v1 ∀v2 (v1 < v2 ∨ v1 = v2 ∨ v2 < v1 ),
∀v1 ∀v2 ∀v3 (v1 < v2 ∧ v2 < v3 → v1 < v3 )

• The theory of groups (in Lgrps ) is given by the sentences:

– ∀v1 ∀v2 ∀v3 v1 · (v2 · v3 ) = (v1 · v2 ) · v3
– ∀v1 (v1 · 1 = v1 ∧ 1 · v1 = 1)
– ∀v1 (v1 · v1−1 = 1 ∧ v1−1 · v1 = 1)
• The theory of rings is the Lrings -theory given by the expected axioms.

End of Lecture 10.

Lecture 11 (The Semantics of First-Order Logic)

1.2 Semantics
Let us turn our attention to the meaning a first-order formula should convey.
This is a matter of interpreting the non-logical symbols (the signature of the
language), but also if necessary the free variables. This section is less clumsy
but more important than §1.1.

1.2.1 Structures
Definition 1.2.1 (L-structure). An L-structure M consists of:
• a non-empty base set M (called the base set or underlying set);
• for each constant symbol c of L, a specific element cM of M ;
• for each n-ary relation symbol R of L, a subset RM of M n ;
we require that =M must be the equality on M

50
 • for each n-ary relation function f of L, a function f M from M n to M .
One also refers to M as the universe of the structure.
We have thus given a meaning to all symbols of L. All? No! We have
forgotten the meaning of the variables, which prevent the terms from having a
meaning, which prevent the formulas from having a meaning! So one must also
specify the value of the variables.

1.2.2 Parameters and Interpretations

Definition 1.2.2 (assignment of the variables). Let M be an L-structure. An
assignment of the variables, or “assignment” for short, is a function s from V to
M.
An assignment is merely a choice of parameters (it specifies what the vari-
ables stand for).
Definition 1.2.3 (interpretation of a term with parameters). Let M be an
L-structure and s : V → M be an assignment of the variables. We define the
interpretation in M of a term t with parameters s, denoted s(t).
• s(c) = cM for a constant symbol c;
• s(x) already makes sense for a variable x;

• if t is f (t1 , . . . , tn ), then s(t) = f M (s(t1 ), . . . , s(tn )).

Lemma 1.2.4 (univocity of interpretation). Let M be an L-structure, t an L-
term and s, s0 : V → M be two assignments such that s| Var(t) = s0| Var(t) . Then
s(t) = s0 (t).

Proof . Clear by induction.

1.2.3 Satisfaction and Semantic Consequence

Definition 1.2.5 (satisfaction). Let M be an L-structure and s : V → M be
an assignment. We define satisfaction in M of a formula ϕ with parameters s,
written M |= ϕ[s].
• if ϕ is atomic, say R(t1 , . . . , tn ), then M |= ϕ[s] if (s(t1 ), . . . , s(tn )) ∈ RM
(Recall that the symbol = is always interpreted as the equality on M ).
• if ϕ is ¬ψ, then M |= ϕ[s] if M 6|= ψ[s].

• if ϕ is ψ → χ, then M |= ϕ[s] if M 6|= ψ[s] or M |= χ[s].

• if ϕ is ∃x ψ, then M |= ϕ[s] if there is an assignment s0 such that s and
s0 agree on V \ {x} and M |= ψ[s0 ].

51
 • if ϕ is ∀x ψ, then M |= ϕ[s] if it is the case that for any assignment s0
such that s and s0 agree on V \ {x}, M |= ψ[s0 ].
The idea is clear. The ∃ clause means that there is a value for x such that
ψ holds, the other variables being fixed as in s. The ∀ clause means that if one
freezes all variables y 6= x to s(y) but x assumes all possible values m ∈ M , ψ
holds in M.
Lemma 1.2.6 (univocity of satisfaction). Let M be an L-structure, ϕ an L-
formula, and s, s0 : V → M two assignments such that s| FreeVar ϕ = s0| FreeVar ϕ .
Then M |= ϕ[s] iff M |= ϕ[s0 ].
In particular, if ϕ is a sentence, then M |= ϕ[s] does not depend on s.
Proof . A clear induction.
We naturally write M |= Σ[s] if Σ is a set of wff’s such that for all ϕ ∈ Σ,
M |= ϕ[s].
If Σ is a set of sentences, then M |= Σ[s] does not depend on s; in this case
one says that M is a model of Σ. We do not insist on this terminology now, as
we shall be working mostly with sets of formulas instead of theories.
Definition 1.2.7 (satisfiability). Let Σ be a set of L-wff’s. Σ is satisfiable if
there is an L-structure M and an assignment s : V → M such that M |= Σ[s].
Pay attention to the following. We want our notion of consequence to deal
not only with sentences (which have no free variables), but also with formulas
with free variables. For instance we want to say that x = y does imply x + 1 =
y + 1, even though we have no idea of what x and y are. So we do need to
operate on all wff’s.
Definition 1.2.8 (semantic consequence). Let Σ be a set of formulas, ϕ a
formula. Σ |= ϕ if for any L-structure M and any assignment s : V → M , if
M |= Σ[s] then M |= ϕ[s].
Example 1.2.9.
• Let Tgrp be the theory of groups, ϕ the sentence ∀v1 v1 · v1 = 1, and ψ
the sentence ∀v1 ∀v2 v1 · v2 = v2 · v1 . Then Tgrp ∪ {ϕ} |= ψ.
• For n ∈ N, let ϕn be the sentence ∃v1 . . . ∃vn ∧i6=j ¬(vi = vj ). For each
n ≥ m, one has ϕn |= ϕm .

End of Lecture 11.

Lecture 12 (Substitutions)

1.3 Substitutions
We shall soon give deduction rules extending those of propositional logic. Before
that, we need more notions about the interplay of terms and formulas.

52
1.3.1 Substitutability
Given a formula ϕ and a variable x ∈ Var ϕ, we may want to replace x by a
term t. This will typically be the case in our proof theory, when we want to
deduce ϕ(t) from ∀x ϕ(x). But one must be extremely careful as the following
examples show.
Example 1.3.1.
• Consider the formula v1 = 1 and the term f (v1 ). Replacing, one finds
f (v1 ) = 1, and the replacement procedure stops after one iteration.
• Consider the formula ∀v1 v1 = v2 . If we replace v1 by a term which is not
a variable, the result no longer is a wff! (“∀1 1 = v1 ”)
We shall therefore substitute only free occurences. The problem is that
we cannot naively susbtitute all free occurences.
• If in ∃v1 v1 = v2 we replace v2 by v1 , we find ∃v1 v1 = v1 , which clearly
should mean something else.
• Similarly, if in ∃v1 v1 = v2 we replace v2 by any term involving v1 , we
alter the meaning.
These examples suggest that we may replace only certain free variables, but
not all. A problem arises when the replacement bounds a variable of t that was
free in ϕ before.
Definition 1.3.2 (term substitutable for a variable). A term t is substitutable
for a variable x in a formula ϕ if no variable of Var t becomes bound in ϕ when
one replaces every free occurence of x by t.
Example 1.3.3.
• Let ϕ1 be v1 = v2 and t be any term. Then t is substitutable for v1 in ϕ1 ;
it is also substitutable for v2 , and substitutable for v3 .
• Let ϕ2 be ∀v1 v1 = v1 . Then f (v1 ) is substitutable for v1 in ϕ2 , as there
are no free occurences of v1 . It is also substitutable for v2 , etc.
• Let ϕ3 be ∀v1 v1 = v2 . Then any term t is substitutable for v1 in ϕ3 . On
the other hand, t is substitutable for v2 in ϕ3 iff v1 does not occur in t.
For instance, if t = f (c, v1 ), then writing ∀v1 v1 = t would deeply alter
the structure of ϕ3 . If t = f (c, v2 ), or t = f (c, v3 ), then replacing v2 by t
is perfectly licit.
• Let ϕ4 be (∀v1 v1 = v2 ) → (v2 = v1 ) and t be a term. Then t is substi-
tutable for v1 in ϕ4 . On the other hand, t is substitutable for v2 in ϕ4 iff
v1 does not occur in t.
Notation 1.3.4. If t is substitutable for x in ϕ, one writes ϕ[t/x] for the
expression where every free occurence of x is replaced by t. (“/” is read “for”).

53
 We refrain from writing ϕ[t/x] if t is not substitutable for x in ϕ.
Example 1.3.5. Consider Example 1.3.3 again.
• ϕ1 [t/v1 ] is t = v2 , whereas ϕ1 [t/v3 ] is of course v1 = v2 .
• ϕ2 [f (v1 )/v1 ] is ∀v1 v1 = v1 . So is ϕ2 [f (v1 )/v2 ].
• ϕ3 [f (c, v1 )/v1 ] is ∀v1 v1 = v2 .
ϕ3 [f (c, v3 )/v2 ] is ∀v1 v1 = f (c, v3 ).
“ϕ3 [f (c, v1 )/v2 ]” is of course not allowed (not substitutable), as the v1
from t would become bound.
• ϕ4 [f (c, v1 )/v1 ] is (∀v1 v1 = v2 ) → (v2 = f (c, v1 )).
ϕ4 [f (c, v3 )/v2 ] is (∀v1 v1 = f (c, v3 )) → (f (c, v3 ) = v1 ).
“ϕ4 [f (c, v1 )/v2 ]” is not allowed.
Bear in mind that we replace only free occurences!

1.3.2 A Renaming Algorithm

As we have said, one must be extraordinarily stupid (or wicked) to construct
misleading formulas; intuitively, we always avoid these dangers. Here is why.
Theorem 1.3.6 (renaming). Let ϕ be a wff, t a term, and x a variable. Then
there is ϕ̃ which differs from ϕ only in the names of bound variables, and such
that t is substitutable for x in ϕ̃, and {ϕ} ` ϕ̃ and {ϕ̃} ` ϕ.
We do not wish to define the notion ` right now (this will be done in §1.4.1),
but it will be clear in due time that the renaming procedure described below
makes ϕ and ϕ̃ “mutually provable”.
Caution! ϕ̃ depends on ϕ, on x, and on t.
Proof . We construct a suitable ϕ̃ inductively.
• If ϕ is atomic, then there is nothing to do: ϕ̃ can be ϕ. Clearly t is
substitutable for x in ϕ̃.
• If ϕ is ¬ψ, we apply the algorithm to smaller ψ, and take ϕ̃ to be ¬ψ̃.
Again, t is substitutable for x in ϕ̃.
• If ϕ is ψ1 → ψ2 , we take ϕ̃ to be ψ̃1 → ψ̃2 . ϕ̃ meets the requirements.
• Suppose that ϕ is ∀y ψ, with y ∈ V.
If x = y, that is if ϕ is ∀x ψ, we may take ϕ̃ to be ϕ itself. There are no
free occurences of x in ϕ, so t was trivially substitutable for x in ϕ.
If x 6= y and y 6∈ Var(t), then no problem will occur from the quantification
on y. We let ϕ̃ be ∀y ψ̃; clearly t is substitutable for x in ϕ̃.

54
 If x 6= y but y ∈ Var(t), then something must be done. We let z be any
(or better: the first!) variable not in Var(ψ̃) ∪ Var(t) ∪ {x}. Notice that
z is then substitutable for y in ψ̃ (as it doesn’t appear in it); so ψ̃[z/y]
makes sense. Let ϕ̃ be ∀z ψ̃[z/y]. We claim that t is substitutable for x
in ϕ̃: this is clearly the case, as it is true in ψ̃, and z does not occur in t.
• We use the same method to deal with ∃y ψ.
Notice that if V has been effectively enumerated, then taking the first vari-
able not occuring in Var(ψ̃) ∪ Var(t) ∪ {x} makes ϕ̃ well-defined, and we have
an effective algorithm.
Example 1.3.7. Example 1.3.3 again.
• The easiest is to take ϕ̃1 for ϕ1 . One can substitute any term for any
variable in ϕ1 .
• Similarly, ϕ̃2 can be ϕ2 , as any term is substitutable for any variable in
ϕ2 .
• t is substitutable for v1 in ϕ3 , so with respect to v1 , ϕ̃3 could be ϕ3 .
If v1 6∈ Var(t), then t is substitutable for v2 in ϕ3 ; ϕ̃3 can still be ϕ3 .
Now if v1 ∈ Var(t) and we want to susbtitute t for v1 in ϕ3 , something
must be done. If for instance t is f (v1 ), then “∀v3 v3 = v2 ” will do as
ϕ̃3 . (In this case, ϕ̃3 [t/v1 ] is “∀v3 v3 = f (v1 )”.) On the other hand if t is
g(v1 , v2 , v3 ), then the ϕ̃3 given by our algorithm will be ∀v4 v4 = v2 (and
ϕ̃3 [t/v1 ] will now be ∀v4 v4 = g(v1 , v2 , v3 )).
Notice that ϕ̃3 does depend on x and t!
• We want to be able to substitute t for v2 in ϕ4 ; a problem occurs only if
v1 ∈ Var(t), which we assume. Let z be a variable not occuring in t, nor
equal to v2 . Then “(∀z z = v2 ) → (v2 = v1 )” will do as ϕ̃4 . (Of course
this ϕ̃4 only fits a certain desired substitution, as z depends on t.)
If we were not interested in the proof theory of first-order logic, we would
work permanently “up to renaming”. Unfortunately we will have to be a little
careful.

1.3.3 Substitutions and Satisfaction

What happens if we substitute a term to a variable, and then wonder about
satisfaction? Fortunately, everything goes as expected. In what follows, we
shall consider an assignement s and a term t which we shall substitute to a
variable x. This merely has the effect of changing the value of s at variable x -
the new value at x being the interpretation of t.
Lemma 1.3.8. Let t, τ be L-terms, s an assignment, x a variable. Let š agree
with s on V \ {x}, and such that š(x) = s(t). Then š(τ ) = s(τ [t/x]).

55
Proof . Induction on τ . For clarity, let τ̌ denote τ [t/x]. We want to prove that
š(τ ) = s(τ̌ ).
• This is clear if τ is a constant, or a variable not equal to x.
• If τ is x, then τ̌ is t, and š(τ ) = š(x) = s(t) = s(τ̌ ).
• If τ is f (τ1 , . . . , τn ), then š(τ ) = š(f (τ1 , . . . , τn )) = f M (š(τ1 ), . . . , š(τn )) =
f M (s(τ̌1 ), . . . , s(τ̌n )) = s(f (τ̌1 , . . . , τ̌n )) = s(τ̌ ).
Proposition 1.3.9. Let ϕ be a formula and t a term substitutable for a variable
x in ϕ. Let M be an L-structure and s : V → M an assignment. Let š be
the assignement which agrees with s on V \ {x} but on š(x) = s(t). Then
M |= (ϕ[t/x])[s] iff M |= ϕ[š].
Proof . Induction on ϕ. For clarity we let ϕ̌ denote ϕ[t/x]. Hence we want to
prove M |= ϕ̌[s] iff M |= ϕ[š].
• Suppose that ϕ is atomic, say R(τ1 , . . . , τn ). Then ϕ̌ is the formula
R(τ1 [t/x], . . . , τn [t/x]), which we naturally denote R(τ̌1 , . . . , τ̌n ). Now us-
ing Lemma 1.3.8,
M |= ϕ̌[s] iff (s(τ̌1 ), . . . , s(τ̌n )) ∈ RM
iff (š(τ1 ), . . . , š(τn )) ∈ RM
iff M |= ϕ[š]

• The case where ϕ is obtained via connectives is trivial.

• We suppose ϕ of the form ∀y ψ.
First, if y = x, then ϕ̌ is exactly ϕ (as x does not occur free in ϕ). Then
using Lemma 1.2.6,
M |= ϕ̌[s] iff M |= ϕ[s]
iff M |= ϕ[š]

So the real case is when y 6= x, which we now suppose. By assumption, t

is is substitutable for x in ϕ, and this implies y 6∈ Var(t). Furthermore, t
is substitutable for x in ψ and ϕ̌ is ∀y ψ̌.
– Suppose that M |= ϕ̌[s]; we show that M |= ϕ[š], that is M |=
(∀y ψ)[š]. Let š0 be any assignment agreeing with š except on y. We
need to show that M |= ψ[š0 ].
By assumption, M |= ϕ̌[s], which means M |= (∀y ψ̌)[s]. Let s0 agree
with s except on y, and such that s0 (y) = š0 (y). By assumption,
M |= ψ̌[s0 ].
We note that s0 and š0 agree but on x, and s(t) = š(x) = š0 (x). Now
y does not appear in t, and s and s0 agree except on y, so by Lemma
1.3.8 s0 (t) = s(t) = š0 (x). As M |= ψ̌[s0 ], it follows by induction,
M |= ψ[š0 ]. As š0 is arbitrary on y but coincides with š everywhere
else, this means M |= (∀y ψ)[š], that is M |= ϕ[š].

56
 – Now suppose M |= ϕ[š]. We want to show that M |= ϕ̌[s], that is
M |= (∀y ψ̌)[s]. So let s0 be any assignement agreeing with s except
on y. Let š0 agree with š except on y, and such that š0 (y) = s0 (y).
Notice that š0 (x) = š(x) = s(t) = s0 (t), as y 6∈ Var(t), and s and s0
agree except on y.
We know M |= ϕ[š], so M |= (∀y ψ)[š]. In particular, M |= ψ[š0 ].
But š0 and s0 agree everywhere but on x, and š0 (x) = s0 (t). So
by induction, M |= ψ̌[s0 ]. As s0 was arbitrarily obtained from s
by changing its value on y, we find M |= (∀y ψ̌)[s], and therefore
M |= ϕ̌[s].
• The case of ∃y ψ is similar.
Counter-example 1.3.10. It is extremely important to have t substitutable
for x in ϕ. Consider the case where ϕ is ∀v2 v1 = v2 and t is v2 . Clearly t is not
substitutable for v1 in ϕ. We let ϕ̌ denote the sentence ∀v2 v2 = v2 , resulting
from an illegitimate replacement.
It is the case that for any L-structure M and assignment s : V → M one will
have M |= ϕ̌[s]. But it will not be the case in general that M |= ϕ[š]. Reasoning
as in the proof of Proposition 1.3.9, one does have š0 (x) = š(x) = s(t), but now
y occurs in t, and there is no reason why s(t) = s(y) and s0 (y) should agree!
End of Lecture 12.
Lecture 13 (Deductions; Simplifying the Language)

1.4 Deductions and Soundness

1.4.1 Deductions
We already know reasonable deduction rules for ¬ and →. In order to extend
natural deduction to first-order logic, we must explain the roles of =, ∀, and ∃.
• Equality requires new axioms, as we want the following intuitively obvious
properties to be deducible.
= =
t=t t1 = t2 → (t[t1 /x] = t[t2 /x])
=
t1 = t2 → (ϕ[t1 /x] ↔ ϕ[t2 /x])
where x is a variable, ϕ is a wff, t, t1 , t2 are terms with t1 and t2 sub-
stitutable for x in ϕ. (This class of axioms could be made considerably
smaller, but it is not our goal.) Ex. 1.10, 1.11

• ∀ adds deduction rules. One may infer ∀x ϕ from ϕ if x does not appear in
th assumption. On the other hand, using a universal quantifier amounts
to substituting to a special case.

57
 • ∃ adds dual deduction rules. One shows an existential by providing an
example, and uses one by introducing a witness.
Exercise 1.9
Of course we shall prove that these rules may be regarded as short-cuts,
provided ∃ is considered as an abbreviation for ¬∀¬ (this will be done in
§1.4.2).

One could now finish the proof of the renaming algorithm (Theorem 1.3.6)
by showing that {ϕ} ` ϕ̃ and {ϕ̃} ` ϕ. Exercise 1.14

Remark 1.4.1. We must slightly update the definition of consistency. Why?

Consider {¬(x = x)}. If no reference to equality is made, then this set of
formulas is consistent, but yields inconsistency as soon as one adds the equality
axioms (which we won’t renounce anyway).
So we redefine consistency as consistency when one adds the axioms of equal-
ity (in practice no ambiguity arises).
Proposition 1.4.2.
• ` ∀x ∀y (x = y → y = x)
• ` ∀x ∀y ∀z (x = y ∧ y = z → x = z)
Proof . Let ϕ be the formula z = x, in which we shall replace z by x, or y.
Notice that ϕ[x/z] is x = x, but ϕ[y/z] is y = x.
We find the deduction
x=y x = y → ϕ[x/z] → ϕ[y/z]
x=x→y=x x=x
y=x

Introducing the implication and quantifying, we do get

` ∀x ∀y (x = y → y = x)

Transitivity is an exercise. Exercise 1.13

1.4.2 Simplifying the Language

We now explain briefly why two quantifiers are obviously redundant (as far as
the semantics is concerned). This runs parallel to §0.2.3 and 0.3.3.
Notation 1.4.3 (see Notation 0.2.12). Let WFF0 be the set of wff’s that do
not use ∃.
Notation 1.4.4 (see Notation 0.2.13). We inductively translate any wff ϕ into
a formula ϕ0 of WFF0 :

58
 • for atomic ϕ, ϕ0 is ϕ;
• (¬ϕ)0 is (¬ϕ0 );
• (ϕ → ψ)0 is (ϕ0 → ψ 0 );

• (∀x ϕ)0 is(∀x ϕ0 )

• (∃x ϕ)0 = ¬(∀x ¬ϕ0 )
Of course ϕ and ϕ0 bear the same meaning; we check that viewing ∃ as an
abbreviation doesn’t alter our notions of consequence.

Notation 1.4.5 (see Notation 0.2.15). Let M be an L-structure, s : V → M a

truth assignment, and ϕ a wff. We define M |=0 ϕ[s] inductively.
• if ϕ is atomic, say R(t1 , . . . , tn ), then M |= ϕ[s] if (s(t1 ), . . . , s(tn )) ∈ RM
• if ϕ is ¬ψ, then M |= ϕ[s] if M 6|= ψ[s].

• if ϕ is ψ1 → ψ2 , then M |= ϕ[s] if M 6|= ψ1 [s] or M |= ψ2 [s].

• if ϕ is ∀x ψ, then M |= ϕ[s] if it is the case that for any assignment s0
such that s and s0 agree on V \ {x}, M |= ψ[s0 ].
Notice that this is a subdefinition of Definition 1.2.5, but that |=0 is defined
only on formulas of WFF0 .
Notation 1.4.6. Let Θ be a consistent subset of WFF0 and ψ ∈ WFF0 . We
write Θ |=0 ψ if for any L-structure M and assignment s : V → M such that
M |=0 Θ[s], one has M |=0 ψ[s].
Lemma 1.4.7 (see Lemma 0.2.16). Let M be an L-structure, s an assignment,
and ϕ an L-formula. Then M |= ϕ[s] iff M |=0 ϕ0 [s].
Proof . By induction on ϕ, the only non-immediate case being when ϕ is ∃x ψ.
But this case is very easy.
Corollary 1.4.8 (see Corollary 0.2.18). Σ |= ϕ iff Σ0 |=0 ϕ0 .
Proof . Obvious from Lemma 1.4.7.

The syntactic side is hardly more subtle.

Lemma 1.4.9 (see Lemma 0.3.12). {ϕ} ` ϕ0 and {ϕ0 } ` ϕ.
Proof . Induction again. It clearly boils down to treating the case where ϕ is
∃x ψ, so that ϕ0 is ¬∀x ¬ψ 0 .

59
 • Notice that x is substitutable for x in ψ 0 . Using induction, one finds
Ax,Wk
{∃x ψ, ψ, ∀x ¬ψ 0 } ` ψ
.. Ax,Wk
.. induction {∃x ψ, ψ, ∀x ¬ψ 0 } ` ∀x ¬ψ 0
∀e
{∃x ψ, ψ, ∀x ¬ψ 0 } ` ψ 0 {∃x ψ, ψ, ∀x ¬ψ 0 } ` ¬ψ 0
¬i
{∃x ψ, ψ} ` ¬∀x ¬ψ 0
| {z }
(∗)

so
{∃x ψ} ` ∃x ψ (∗)
∃e
{ϕ} ` ϕ0

• We need Reductio ad Absurdum (Theorem 0.3.10) for the converse. We

show that {ϕ0 , ¬ϕ} is inconsistent.

Ax,Wk
{¬∃x ψ, ψ 0 } ` ψ 0
..
.. induction
{¬∃x ψ, ψ 0 } ` ψ
∃i Ax,Wk
{¬∃x ψ, ψ 0 } ` ∃x ψ {¬∃x ψ, ψ 0 } ` ¬∃x ψ
¬i
{¬∃x ψ} ` ¬ψ 0
| {z }
(∗)

Now as x does not appear free in the assumptions,

{¬∃x ψ} ` ¬ψ 0
∀i
{¬∃x ψ} ` ∀x ¬ψ 0

It follows that {¬ϕ, ϕ0 } is inconsistent, whence {ϕ0 } ` ϕ.

Of course the real problem is to show that one may drop the ∃ rules without
affecting the deductive strength.
Notation 1.4.10 (see Notation 0.3.13). Write Σ `0 ϕ if there is a deduction
which does not use the ∃ rules.
Corollary 1.4.11 (see Corollary 0.3.14). Let Σ be a theory and ϕ a wff. Then
Σ ` ϕ iff Σ0 `0 ϕ0 .
Proof . The implication Σ0 `0 ϕ0 ⇒ Σ ` ϕ is very comparable to the proof of
Corollary 0.2.18, thanks to Lemma 1.4.9.
We now show that Σ ` ϕ implies Σ0 `0 ϕ0 by induction on the deduction.
The only interesting cases are when the last step is a ∃-rule. There are two
possibilities.

60
 • Suppose that the last step is:
Σ ` ψ[t/x]
∃i
Σ ` ∃x ψ
where t is substitutable for x in ψ. Then
.. Ax
.. induction {∀x ¬ψ 0 } `0 ∀x ¬ψ 0
Wk
Σ0 `0 ψ 0 [t/x] Σ0 ∪ {∀x ¬ψ 0 } `0 ∀x ¬ψ 0
Wk ∀e
Σ0 ∪ {∀x ¬ψ 0 } `0 ψ 0 [t/x] Σ0 ∪ {∀x ¬ψ 0 } `0 ¬ψ 0 [t/x]
¬i
Σ0 `0 ¬∀x ¬ψ 0
| {z }
=ϕ0

so Σ0 `0 ϕ0 .
• Now suppose that the last step is:
Σ ` ∃x ψ Σ ∪ {ψ} ` ϕ
∃e
Σ`ϕ
where x does not occur free in Σ ∪ {ϕ}. By induction, Σ0 ∪ {ψ 0 } `0 ϕ0 ;
contraposing (Theorem 0.3.8), Σ0 ∪ {¬ϕ0 } `0 ¬ψ 0 .
Now
Σ0 ∪ {¬ϕ0 } `0 ¬ψ 0
∀i
Σ0 ∪ {¬ϕ0 } `0 ∀x ¬ψ 0
since x does not occur free in Σ0 ∪ {ϕ0 }.
By induction again, Σ0 `0 ¬∀x ¬ψ 0 . So Σ0 ∪ {¬ϕ0 } is inconsistent, and
Σ0 `0 ϕ0 by Reductio ad Absurdum (Theorem 0.3.10).
Hence for us, an existential statement may be proved by showing that a
universal statement does not hold. Intuitively, this is disputable: one has never
given an example! For instance, in intuitionistic logic where one does not allow
¬e as a deduction rule, ∃ is not equivalent to ¬∀¬. Exercise 1.15

End of Lecture 13.

Lecture 14 (Soundness; Completeness (1/2))

1.4.3 Soundness
Theorem 1.4.12 (soundness of first-order logic). If Σ ` ϕ, then Σ |= ϕ.
Proof . By induction on the deduction. Notice that we are extending propo-
sitional logic, so weakening, negation, and implication have been dealt with
in the proof of the propositional logic version (Theorem 0.3.17). Axioms of
equality are easily dealt with. It remains to consider the case of quantification.
By Corollaries 1.4.8 and 1.4.11, only ∀ need be examined: ∃ is now a mere
abbreviation.

61
 • Suppose that the last step of the deduction was:
Σ`ϕ
∀i
Σ ` ∀x ϕ

where x does not occur free (Definition 1.1.16) in Σ. Then Σ ` ϕ and by

induction, Σ |= ϕ. We show Σ |= ∀x ϕ.
Let M be a structure and s : V → M an assignment. We assume M |=
Σ[s] and aim at showing M |= (∀x ϕ)[s].
Let s0 be any other assignment agreeing with s on V \ {x}. We must show
M |= ϕ[s0 ]. As x does not occur free in Σ, M |= Σ[s0 ] (Lemma 1.2.6).
As Σ |= ϕ, this implies M |= ϕ[s0 ]. So M |= ϕ[s0 ] regardless of the value
assigned to x by s0 . So M |= (∀x ϕ)[s]. As a conclusion, Σ |= ∀x ϕ.
• Now suppose that the last step was:
Σ ` ∀x ϕ
∀e
Σ ` ϕ[t/x]

where t is substitutable (Definition 1.3.2) for x in ϕ. We shall show Σ |=

ϕ[t/x]. For clarity, let ϕ̌ denote ϕ[t/x].
Let M be an L-structure, s : V → M any assignment, and suppose
M |= Σ[s]. We have to prove M |= ϕ̌[s].
We know Σ ` ∀x ϕ, and by induction Σ |= ∀x ϕ. As M |= Σ[s], it follows
M |= (∀x ϕ)[s]. So for any s0 which agrees with s on V \ {x}, we have
M |= ϕ[s0 ]. Let š map x to s(t) (and coincide with s elsewhere); again
M |= ϕ[š]. By Proposition 1.3.9, M |= ϕ̌[s]. Hence for any s such that
M |= Σ[s], we have M |= ϕ̌[s]; we are done.
Of course even allowing ∃ as a basic quantifier, the proof wouldn’t be any
harder. Exercise 1.16

1.5 Completeness
Soundness (Theorem 1.4.12) was the easy part. We now show the analog of
Theorem 0.4.1 for first-order logic; informally speaking, it expresses that “if it
is true, then there is a deduction of it" (in first-order classical logic for natural
deduction).
This theorem is due to Kurt Gödel (in his PhD).
Theorem 1.5.1 (completeness of first-order logic). If Σ |= ϕ, then Σ ` ϕ.
We adopt the same strategy as for the case of propositional logic (Theorem
0.4.1). It suffices to show the following:
Theorem 1.5.2 (completeness of first-order logic, equivalent version). If Σ |=
ϕ, then Σ ` ϕ.

62
 Theorems 1.5.1 and 1.5.2 are equivalent (see Lemma 0.4.2).
Remark 1.5.3. Throughout we shall be working with sets of formulas instead
of theories (which are sets of sentences: Definition 1.1.21). Even if you have
a feeling that it would suffice to show Theorem 1.5.2 for theories, there will
be a point in the proof (the fairly technical Lemma 1.5.28 below) in which all
formulas must be taken into account.
Such a subtlety was impossible to imagine in propositional logic, where in a
sense all formulas are sentences!
Remark 1.5.4. The consistent set we start with may be assumed to contain
the axioms of equality. This remark is essential; see Remark 1.4.1.

1.5.0 Strategy and Witnesses

Starting with a consistent set of first-order wff’s, we shall find an L-structure
and an assignment of the variables satisfying it. As in the case of propositional
logic, this will first require extending to a maximal set of formulas (Lemma
1.5.18 of §1.5.2), and then finding an L-structure (Lemma 1.5.28 of §1.5.3).
It is not unrealistic to think that the collection of terms will provide the
base set. (Actually we shall need to “factor out” an equivalence relation, that is
work with a set of equivalence classes; pay no attention now to this slight detail
which will be explained in §1.5.3.) We shall then define an ad hoc structure on
the base set, ensuring that relations and functions behave as required.
For instance, if in the language there are a constant symbol c and a unary
function symbol f , we shall have terms c, f (c), f (f (c)), etc., and we know how
to interpret c and f . But the issue is that we also have quantifiers; for example,
the sentence ∃v1 f (v1 ) = c, which expresses that f (v1 ) = c has a solution, might
be in the theory. The brilliant idea is to add a pointer to such a solution.
First we should simplify the language, following Corollaries 1.4.8 and 1.4.11
of §s:firstorderreducinglanguage. We systematically replace ∃ by ¬∀¬.
For each formula ∀x ϕ, if ∀x ϕ fails, we shall “name the flaw”, i.e. introduce
a new constant cx,ϕ such that

(¬∀x ϕ) → ¬ϕ[cx,ϕ /x]

Such a formula means: “If ϕ is not always true, it is because of cx,ϕ precisely”.
But this will require extending the language with new constants (§1.5.1),
and extending the collection of formulas to force this phenomenon (§1.5.2).
This strategy of adding witnesses is due to Leon Henkin.
Remark 1.5.5. We shall only treat the countable case, but some remarks will
indicate how to generalize to uncountable languages.
Let Σ be a consistent set of L-formulas containing the axioms of equality.
The goal is to find a nice extension of Σ in a language which provides witnesses.

63
1.5.1 Expanding the Language
We want to introduce witnesses for formulas of the form ¬∀x ¬ϕ. This raises
two (minor) issues.
• When we add new constants, we expand the language, so we also create
new formulas, and we need even more witnesses!
• When we add new constants, we expand the class of terms, so we also
expand the equality axioms and the ∀ rules. Will Σ still be consistent?
We deal with the first issue. The intuitive idea is that for any formula of
the form ∀x ϕ, we add a new constant symbol cx,ϕ . Of course when we expand
the language, we add new formulas! It thus looks like we need to repeat the
construction; fortunately the cardinal remains the same.
Lemma 1.5.6. There is a countable extension L0 of L by a set C 0 of new
constant symbols cx,ϕ , such that C 0 is in bijection with the set of pairs (x, ϕ)
where x ∈ V and ϕ is a L0 -formula.
Proof . Let L0 = L and WFF0 = WFF(L0 ). For each pair (x, ϕ) ∈ V × WFF0 ,
we add a new constant symbol cx,ϕ , getting the language L1 .
We repeat the construction with formulas of WFF1 = WFF(L1 ), and intro-
duce new constant symbols, which yields L2 . Let L0 = ∪n∈N Ln . Clearly L0 has
the desired property.
Notice that at each step Card Ln remains countable, so Card L0 = Card L.

We now deal with the second possible issue - why does Σ remain consistent in
L0 ? The intuitive reason is clear: if from Σ one can deduce a contradiction in L0 ,
then the deduction has used new constant symbols; L says nothing about these,
so they behave like free variables. This is expressed by the following theorem.
In the statement below, ϕ[z/c] denotes brute replacement of all occurences of
the constant c by the variable z (this creates no confusion, as there was certainly
no quantification over the constant c).
Theorem 1.5.7 (generalized generalisation). Let c be a constant symbol occur-
ing in ϕ, and suppose Σ ` ϕ where c does not occur in Σ. Then there is a
variable z not occuring in ϕ such that Σ ` ∀z ϕ[z/c].
Example 1.5.8. If Σ ` c = 0 and Σ does not mention c, then Σ ` ∀x x = 0.
Proof of Theorem 1.5.7. We write D : Σ ` ϕ to say that D is a deduction
(our usual deductive tree). Notice that there is a finite subset Σ0 ⊆ Σ such that
D : Σ0 ` ϕ.
Let z be a variable not occuring in D nor Σ0 (which are finite, but we have
countably many variables!) We replace c by z in each node of D, forming a tree
D[z/c]. We show that D[z/c] : Σ0 ` ϕ[z/c].
• If we use an axiom of equality, then substituting z for c remains an axiom
of equality!

64
 • If the step is for instance of the form:
Σ0 ` ψ Σ0 ` ψ → χ
→e
Σ`χ

then by induction Σ0 ` ψ[z/c] and Σ0 ` (ψ → χ)[z/c] which is Σ0 `

ψ[z/c] → χ[z/c]. Hence

Σ0 ` ψ[z/c] Σ0 ` (ψ → χ)[z/c]
→e
Σ0 ` χ[z/c]

which completes the step. The ¬i , ¬e , →i rules are as easy.

• Suppose that the last step was of the form:
Σ0 ` ψ
∀i
Σ0 ` ∀x ψ

where x does not occur free in Σ0 . Notice that by choice of z, (∀x ψ)[z/c]
is ∀x (ψ[z/c]). It follows
..
.. induction
Σ0 ` ψ[z/c]
∀i
Σ0 ` (∀x ψ)[z/c]

• Suppose that the last step was of the form:

Σ0 ` ∀x ψ
∀e
Σ0 ` ψ[t/x]

where t is substitutable for x in ψ. By choice of z, (∀x ψ)[z/c] is ∀x (ψ[z/c])

and (ψ[t/x])[z/c] is (ψ[z/c])[t/x]. It follows
..
.. induction
Σ0 ` ∀x ψ[z/c]
∀e
Σ0 ` (ψ[t/x])[z/c]

Hence Σ0 ` ϕ[z/c]. As z does not occur free in Σ0 , one has Σ0 ` ∀z ϕ[z/c]. So

Σ ` ∀z ϕ[z/c] too.

Remark 1.5.9. Σ0 appears just to make sure that we can provide a new variable
z: perhaps Σ did mention all variable names... but we used only finitely many
in the deduction anyway.
Corollary 1.5.10. Σ with the new equality axioms remains consistent in L0 .

65
Proof . We shall prove a little more: write Σ `0 ϕ if there is a proof using
deduction rules associated to L0 , and Σ ` ϕ if the deduction is carried in L. We
claim that if Σ, ϕ do not use new constant symbols, then Σ `0 ϕ implies Σ ` ϕ.
This will clearly establish L0 -consistency of Σ.
So we suppose Σ `0 ϕ, and we assume Σ finite. Working inductively, for
every new constant symbol ci ∈ C 0 which appears in the deduction, there is
by Theorem 1.5.7 a variable xi such that Σ ` ∀xi ϕ[xi /ci ]; in particular Σ `
ϕ[xi /ci ]. As the ci ’s don’t appear in ϕ, we thus have Σ ` ϕ.
Yes! We have found a reasonable language to work in: it has enough wit-
nesses, and Σ remains consistent.
Remark 1.5.11. The construction of L0 immediately generalizes to uncount-
able L (the construction remains indexed by N; at each stage, Card Ln =
Card L). Corollary 1.5.10 is still true.

End of Lecture 14.

Lecture 15 (Completeness (2/2))

1.5.2 Extending the Theory

From now on we consider Σ as a consistent set of L0 -formulas which contains
the L0 -axioms of equality.
We shall next prescribe the role of the set of witnesses C 0 . Let {(xn , ϕn : n ∈
N} be an enumeration of the set of pairs (x, ϕ), where x ∈ V is a variable and
ϕ is an L0 -formula.
Notation 1.5.12 (Henkin axioms). Let ηn be the formula:

(¬∀xn ϕn ) → ¬ϕn [cxn ,ϕn /xn ]

where cxn ,ϕn is the first element of C 0 which does not occur in {ϕ0 , . . . , ϕn } ∪
{η0 , . . . , ηn−1 }. For simplicity, we write cn instead of cxn ,ϕn (there is no risk of
confusion with the original constants).
ηn means: “if ϕn is not true for all xn ’s, it is because of...” and points to a
specific constant symbol, which has never been mentioned before.
Notation 1.5.13. Let Σ0 = Σ ∪ {ηn : n ∈ N}.
We have expanded our consistent set. Is it still consistent?
Lemma 1.5.14. Σ0 is a consistent set of L0 -formulas.
Proof . Let Σn = Σ ∪ {ηk : k ≤ n}. We know that Σ0 = Σ is consistent.
Suppose that Σn is; we show that Σn+1 is consistent again.
Otherwise, Σn ` ¬ηn+1 . Recall that ηn+1 stands for

(¬∀xn+1 ϕn+1 ) → ¬ϕn+1 [cn+1 /xn+1 ]

66
 Minimal work shows Σn ` ¬∀xn+1 ϕn+1 and Σn ` ϕn+1 [cn+1 /xn+1 ]. But
by construction (Notation 1.5.12), cn+1 does not occur in Σn . So generalizing
on the constant (Theorem 1.5.7), we may replace cn+1 by xn+1 and quantify.
We find Σn ` ∀xn+1 ϕn+1 , which is a contradiction to its consistency.
Σ0 is the basis of our inductive process; it is a consistent set of L0 -formulas.
We now maximize the set of requirements (see §0.4.1). Unfortunately, as ob-
served at the beginning of the proof of completeness (see Remark 1.5.3), we
cannot limit ourselves to sentences, but must take all possible formulas into
account.
Definition 1.5.15 (V-complete; see Definition 0.4.3). A consistent set of first-
order formulas Σ̂ is V-complete if for any formula ϕ, either ϕ or ¬ϕ is in Σ̂.

Caution! This is way stronger than what complete ought to be. A V-complete
(for: variable-complete) set of formulas even prescribes the behaviour of the
variables (v1 = v2 is in it, or v1 6= v2 is in it). The good notion, when working
with theories, is that of completeness (Definition 1.6.2 below), which is only
about sentences.
Remark 1.5.16 (see Remark 0.4.4). If Σ̂ is V-complete, then for any wff ϕ one
has Σ̂ ` ϕ iff ϕ ∈ Σ̂.
We now make our way to a V-complete set of formulas containing the Henkin
axioms.
Lemma 1.5.17 (see Lemma 0.4.5). If Σ is consistent and ϕ is a wff, then
Σ ∪ {ϕ} or Σ ∪ {¬ϕ} is consistent.
Proof . Very much like Lemma 0.4.5.
Lemma 1.5.18 (see Lemma 0.4.6). Let Σ be a consistent set of wff’s. Then
there is a V-complete set Σ̂ containing Σ and the Henkin axioms.
Proof . Like Lemma 0.4.6.

Remark 1.5.19. If L0 is not countable, one has to do a little bit of ordinal

induction. If you know about ordinals, you should see that “taking unions at
limit stages” preserves consistency. The construction is exactly the same.
We have reached a maximal set of L0 -formulas which contains all axioms and
describes the role of witnesses (notice that Σ̂ also describes the relations between
variables, as in Definition 1.5.15 we took all formulas, not only sentences). It is
time to construct a structure.

67
1.5.3 Finding a Structure (and Assignment)
Recall that from a set Σ, we built a maximal, consistent set of L0 -formulas
Σ̂ which contains all axioms of L0 , for some expansion L0 of L by new con-
stant symbols, and the Henkin axioms (Notation 1.5.12), which explain their
behaviours.
We are very close to satisfying Σ̂.
Notation 1.5.20. Let T 0 be the set of terms of L0 .
This looks like a very clever base set, but here is a minor complication. If
in the language there are a constant symbol c and a unary function symbol f ,
then c, f (c), etc. are terms. But perhaps the formula f (f (c)) = c is in Σ̂, in
which case the terms f (f (c)) and c are required to code the same element.
To force this, we identify pairs of terms which are believed to be equal by
the set of conditions.
Notation 1.5.21. Let ∼ be the relation on T 0 :

t ∼ t0 if Σ̂ ` t = t0

Lemma 1.5.22. ∼ is an equivalence relation.

Notation 1.5.23. Let M = T 0 / ∼ be the quotient set of T modulo ∼. We use
brackets [.] to denote equivalence classes.

M will be our universe. We have to interpret symbols of L0 .

Notation 1.5.24. We turn M into an L0 -structure M with the following in-
terpretation:
• For c ∈ L0 , let cM = [c].

• For an n-ary relation symbol R of L0 , let

RM = {([t1 ], . . . , [tn ]) ∈ M n : Σ̂ ` R(t1 , . . . , tn )}

• For an n-ary relation symbol R of L0 , let

f M ([t1 ], . . . , [tn ]) = [f (t1 ), . . . , f (tn )]

Lemma 1.5.25. This is well-defined.

Proof . There is no ambiguity for the constants. We do the case of a relation
symbol (a function symbol is handled similarly).
Suppose t1 ∼ t01 , . . . , tn ∼ t0n and Σ̂ ` R(t1 , . . . , tn ); we show that Σ̂ `
R(t01 , . . . , t0n ). By definition, the formulas “t1 = t01 ”, . . . , “tn = t0n ” are in Σ̂. As
it contains the axioms of equality, we have Σ̂ ` R(t1 , . . . , tn ) → R(t01 , . . . , t0n ).
By assumption, Σ̂ ` R(t1 , . . . , tn ). Detaching, we find Σ̂ ` R(t01 , . . . , t0n ).

68
 We have an L-structure, and now need an assignment of the variables; but
as they are terms, they already have been taken care of.
Notation 1.5.26. Let s : V → M map x to [x].
Lemma 1.5.27. For any L0 -term t ∈ T 0 , s(t) = [t].
Proof . Clear for a constant and a variable; clear by induction.
Lemma 1.5.28 (truth lemma; see Lemma 0.4.7). For any L0 -formula ϕ, one
has M |= ϕ[s] iff ϕ ∈ Σ̂.
Proof . Induction on ϕ. Bear in mind that by V-completeness of Σ̂, for any
formula, ϕ ∈ Σ̂ iff Σ̂ ` ϕ!
• Suppose ϕ is t1 = t2 . Then

M |= ϕ[s] iff s(t1 ) = s(t2 )

iff [t1 ] = [t2 ]
iff Σ̂ ` t1 = t2

• Suppose ϕ is R(t1 , . . . , tn ). Then

M |= ϕ[s] iff (s(t1 ), . . . , s(tn )) ∈ RM

iff ([t1 ], . . . , [tn ]) ∈ RM
iff Σ̂ ` R(t1 , . . . , tn )

• Suppose ϕ is ¬ψ. Then

M |= ϕ[s] iff M 6|= ψ[s]

iff ψ 6∈ Σ̂
iff ¬ψ ∈ Σ̂

• Suppose ϕ is ψ → χ. Then

M |= ϕ[s] iff M 6|= ψ[s] or M |= χ[s]

iff ψ 6∈ Σ̂ or χ ∈ Σ̂
iff (ψ → χ) ∈ Σ̂

(We used consistency and V-completeness of Σ̂.)

• Suppose ϕ is ∀x ψ. The witnesses play their role here. (x, ψ) is some pair
(xn , ϕn ), and therefore ϕ is the formula ∀xn ϕn . Recall (Notation 1.5.12)
that cn denotes the associated witness.
– Assume M |= ϕ[s]; we show that ϕ ∈ Σ̂. By assumption, M |=
(∀xn ϕn )[s]; let s0 agree with s except on xn and such that s0 (xn ) =
cn . By definition of satisfaction, M |= ϕn [s0 ]. Proposition 1.3.9 then
says M |= (ϕn [cn /xn ])[s]. By induction, ϕn [cn /xn ] ∈ Σ̂. As ηn ∈ Σ̂
and Σ̂ is consistent, one has Σ̂ ` ∀xn ϕn . So ϕ ∈ Σ̂.

69
 – Assume M 6|= ϕ[s]; we want to show that ϕ 6∈ Σ̂. By assumption,
there is an assignment s0 which agrees with s except on xn and such
that M |= (¬ϕn )[s0 ].
By construction of M , s0 (xn ) is some class [t] for a term t. We would
like to use Proposition 1.3.9 and say that “M |= (¬ϕn [t/xn ])[s]”; the
issue is that perhaps t is not sustitutable for xn in ϕn , so the above
does not make sense. We need to rename a bit.
Let ϕ̃n be a formula in which t is substitutable for xn and such that
{ϕn } ` ϕ̃n and {ϕ̃n } ` ϕn (Theorem 1.3.6). Then M |= (¬ϕ̃n )[s0 ],
s0 (xn ) = [t] = s(t) (Lemma 1.5.27), and t is substitutable for xn
in ϕ̃n . The use of Proposition 1.3.9 is now legitimate, and we find
M |= (¬ϕ̃n [t/xn ])[s].
By induction, ϕ̃n [t/xn ] 6∈ Σ̂. By consistency (and as t is substituable
for xn in ϕ̃n ), ∀xn ϕ̃n 6∈ Σ̂ either. We are done with substitutions;
as ϕ̃n and ϕn prove each other, we find ∀xn ϕn 6∈ Σ̂. So ϕ 6∈ Σ̂.
The reader should have another look at remark 1.5.3; it is now clear that we
had to work with all formulas, and not only sentences.
Proof of Gödel’s Completeness Theorem for First-Order Logic. Fix a
first-order language L and a consistent set of L-formulas Σ. We find an L-
structure and an assignment satisfying Σ.
We first expand L to a language L0 having many witnesses (Lemma 1.5.6).
We then extend Σ to a V-complete set of L0 -formulas Σ̂, which describes the
behavior of the witnesses (Lemma 1.5.18).
We construct an L0 -structure M as in Notation 1.5.24 and an assignment s
as in Notation 1.5.26. By Lemma 1.5.28, M |= ϕ[s] for all ϕ ∈ Σ0 . In particular,
M |= Σ[s]: Σ is satisfiable.
End of Lecture 15.

Lecture 16 (Consequences; Compactness; Non-Standard Analysis)

1.6 Consequences and Compactness

1.6.1 Decidability*
The definitions of §0.5.2 immediately generalize to the case of first-order logic.
Corollary 1.6.1 (see Corollary 0.5.14). Let L be a first-order language (pre-
sented in an effective way), Σ a set of wff’s (presented in an effective way) and
ϕ a wff. Then there is an algorithm which will answer “yes” if Σ |= ϕ.
Proof . The set of axioms of equality is decidable; so we may assume that Σ
contains all the axioms. There is an algorithm which produces all theorems of Σ
(try all deductions of length n, then all of length n + 1, etc.). So far we have an
algorithm which answers “yes” if Σ ` ϕ. But by soundness and completeness,
this is equivalent to Σ |= ϕ.

70
 We want to generalize Corollary 0.5.16 to first-order logic; this requires defin-
ing some maximality condition for a theory. V-completeness (Definition 1.5.15,
an ad hoc notion for the proof of the completeness theorem) is obviously too
strong, as one should care only for sentences. The good notion is the following.
Definition 1.6.2 (complete). A first-order theory Σ is complete if for any
sentence ϕ, either Σ ` ϕ or Σ ` ¬ϕ.
Notice that Σ is complete iff maximal as a consistent theory. This is more
understandable if one closes Σ under consequence, in which case this definition
is the same as the one we gave when proving completeness of propositional logic,
Definition 0.4.3.
Caution! Let us insist that as opposed to V-completeness, only sentences are
taken into account. Completeness is much weaker than V-completeness, but is
the only relevant notion when working with theories.

Corollary 1.6.3 (see Corollary 0.5.16). Let Σ be a semi-decidable theory in

a decidable language. Assume that Σ is complete. Then there is an algorithm
which decides whether Σ |= ϕ or not.
Hence, “semi-decidable and complete is decidable”.

1.6.2 Compactness
Theorem 1.6.4 (compactness of first-order logic). Let Σ be a set of first-order
formulas. Then Σ is satisfiable iff Σ is finitely satisfiable.
Proof . If Σ is finitely satisfiable, it is consistent. As it is consistent, it is
satisfiable by the equivalent form of completeness.

Caution! Even if you have a clear idea where to satisfy finite fragments of Σ,
you may be surprised by a structure satisfying Σ!
Example 1.6.5. Consider an infinite set A, and let the language contain con-
stants ca (a ∈ A). The set of formulas Σ = {v1 6= ca : a ∈ A} is clearly finitely
satisfiable in A, as it suffices to take an assignment with s(v1 ) not one of the
finitely many mentioned a’s. But of course no assignment to A will satisfy all
of Σ simultaneously.
Corollary 1.6.6. If Σ |= ϕ, there is a finite subset Σ0 ⊆ Σ such that Σ0 |= ϕ.
Proof . Like Corollary 0.5.3.
One may want to have another look at the notion of finite axiomatizability
(Definition 0.5.12).
Example 1.6.7. The theory of an infinite set is not finitely axiomatisable.

71
Verification: Suppose it is. Then by Proposition 0.5.13 the natural axiomati-
zation {∃x1 . . . ∃xn ∧i6=j xi 6= xj : n ∈ N} too contains a finite axiomatisation,
which reduces to a single sentence {∃x1 . . . ∃xn ∧i6=j xi 6= xj }. The latter is
satisfied by any possibly finite set with at least n elements, a contradiction. ♦
Before moving to an application, let us comment on a historical misunder-
standing.

Ce malheureux Théorème de Compacité est entré par la petite

porte, et on dirait que cette modestie originelle lui cause encore du
tort dans les manuels de Logique. C’est à mon avis un résultat beau-
coup plus essentiel, primordial (et donc aussi moins sophistiqué),
que le Théorème de Complétude de Gödel, qui affirme qu’on peut for-
maliser la déduction d’une certaine façon en Arithmétique, et c’est
une erreur de méthode que de l’en déduire.
The poor Compactness Theorem got in through the back door, and
it seems that this initial modesty still mars it in Logic textbooks. It is
in my opinion a result way more essential, primordial (and thus also less
sophisticated) than Gödel’s Completeness Theorem, which states that one
can formalize deduction in a certain way in Number Theory, and it is a
methodological error to deduce the former from the latter.

Bruno Poizat, Cours de Théorie des Modèles

The truth is that the compactness theorem is a purely semantic property,

and has nothing to do with the notion of deduction. It first appeared in days
where deductions were a logician’s main preoccupation; this explains its initial
modesty.
There is an alternate proof of compactness in §1.7.

1.6.3 Non-Standard Analysis*

The genesis of calculus is a complicated story. However the notion of infinitesi-
mals was one of the ways to start a scientific quarrel (and perhaps have a duel)
in the xviiith century. In short, are there numbers which are smaller than all
positive reals? (Of course such numbers could not be real numbers themselves.)
Can one introduce such convenient fictions, and work with them? Is what we
show about real numbers using infinitesimals true or do we create inconcisten-
cies? We provide a xxth century answer to the controversy.
Our goal is to extend R by infinitesimals. First, we agree that R as a field
interests us; that is, R as a ring-structure (as it is customary to treat fields as
ring-structures). We even equip it with the ordering < and the absolute value |.|
(these are actually definable in the ring R, but we may put them in the language
freely). This is the language we are interested in.
But we want a little more. First, we should introduce a new constant for
an infinitesimal. Second, we’re not interested in finding an abstract structure;
we want one which behaves like R, or even better, which contains R in a nice

72
manner. We shall form an ad hoc theory, which will force the structure to
behave.
Consider the language L = {0, 1, <, +, −, ·, |.|} and the ordered field R. We
add new constants {cr : r ∈ R} (which will stand for elements of R), finding the
language LR , and one more new constant α; let L0 = LR ∪ {α} be the resulting
first-order language. Let Σ be the set containing the following:
• all LR -sentences which hold true in R;
• the formulas {0 < α < cr : r ∈ R>0 }
(One cannot formally quantify over r in the latter: the cr ’s are elements
of the language, but their collection is not expressible in the language!)

The theory Σ thus states that the cr ’s have the same relations as the real
numbers they stand for (in particular, some sentences express that we are talking
about an ordered field); the extra formulas say that α is a positive element
smaller than any positive real number.

Lemma 1.6.8. Σ is satisfiable.

Proof . A finite fragment of Σ mentions only finitely many cr ’s, which we inter-
pret in R as the reals they stand for; only finitely many conditions α < cr are
mentioned, and we interpret α as a positive real number smaller than the cr ’s
in question. This discussion shows that Σ is finitely satisfiable; actually that
any finite fragment of Σ is satisfiable in R.
By compactness, Σ is satisfiable (but of course, not in R, because R has no
infinitesimals; same phenomenon as in Example 1.6.5).
Notation 1.6.9 (non-standard reals). Let R∗ be an L0 -structure satisfying Σ.
R∗ contains an interpretation of α, which is an infinitesimal, and an inter-
pretation of the cr ’s, which behave like the real numbers they stand for.
Lemma 1.6.10. R injects canonically into R∗ .
∗
∗
Proof . Map r to cR r (the interpretation of cr in R ). This is natural enough,
and injective alright: if r 6= s ∈ R, then Σ |= cr 6= cs ; hence R∗ |= cr 6= cs .

It goes without saying that the mapping is actually a ring homomorphism.

We now freely identify R with its image in R∗ , and consider that R ⊂ R∗ . The
∗
inclusion is of course proper, as αR 6∈ R: it is an infinitesimal!
We now work on retrieving R. We have added infinitesimals, so in a sense
the non-standard line R∗ is thicker; we have also added their inverses, so it is
also longer than R∗ .

Notation 1.6.11 (bounded elements and infinitesimals). Consider the subsets

of R∗ :
• b = {x ∈ R∗ : there is r ∈ R>0 : |x| < r};

73
 • o = {x ∈ R∗ : for all r ∈ R>0 : |x| < r}.
b is the set of bounded elements (“not infinitely large”); o is the set of
infinitesimals (“infinitely close to 0”). Of course o ∩ R = {0}. If you want to
have a better picture of b and o, you may consider that the real line R embeds
into the non-standard line R∗ . b is the portion of the non-standard line which
is not incommensurably longer than R. o is the thickness around ◦ .
It goes without saying that though nicely defined, these subsets of R∗ are
not (first-order)-definable: the definitions are not expressible in our language
L0 .
Lemma 1.6.12. b is a ring, and o is an ideal of b.
Proof . Easy computations.
We shall now define an isomorphism b/o ' R. The construction is fairly
interesting. Following the same mental picture as before, when one consider
only the commensurable part and factors out the thickness, one retrieves the
standard line R.
Definition 1.6.13 (standard part). For any non-negative x ∈ b, let st x be the
least upper bound of {r ∈ R : r < x}. (We define st x for negative x ∈ b by
st x = − st(−x)).
This makes sense, as R is Dedekind-complete. Let x ∈ b be non-negative;
then the subset A = {r ∈ R : r < x} of R is non-empty and bounded above
since x ∈ b. So A has a least upper bound in R; st is well-defined. A little more
basic algebra reveals:
Lemma 1.6.14. st is a ring-homomorphism with kernel o.
Now as st is clearly onto R (it is the identity on R), we deduce that st
induces an isomorphism b/o ' R. In particular, any element x ∈ b can be
written in a unique way as the sum of a real number st x (its standard part),
and an infinitesimal ε ∈ o. (This is of course not true for elements of R∗ \ b,
which are infinitely big.)
One can then start doing analysis in a very elegant fashion, whith arguments
such as: f is continuous at real a ∈ R iff for any infinitesimal ε, f (a + ε) − f (a)
is infinitesimal.

End of Lecture 16.

Lecture 17 (Proof of Compactness by Ultraproducts*)

1.7 An Alternate Proof of Compactness*

1.7.1 Filters and Ultrafilters*
Definition 1.7.1 (filter). Let X 6= ∅ be a set. F ⊆ P (X) is a filter on X if:

74
 • ∅ 6∈ F;
• if A ∈ F and A ⊆ B, then B ∈ F;
• if A, B ∈ F, then A ∩ B ∈ F.
Example 1.7.2. In a topological space, the collection of neighborhoods of a
given point forms a filter.
Remark 1.7.3. Let B be a family of subsets of X having the finite intersection
property (Definition 0.5.17). Then there is a filter F such that B ⊆ F.
Verification: Let F = {Y ⊆ X : there exist B1 , . . . , Bn in B such that B1 ∩
· · · ∩ Bn ⊆ Y }, i.e. F is the smallest family of sets containing every finite
intersection of members of B. This clearly meets the conditions defining a filter.
Notice that ∅ 6∈ F by the finite intersection property. ♦
Recall that Y cofinite in X means that X \ Y is finite.

Example 1.7.4. Assume X is infinite. The Fréchet filter is F ré = {Y ⊆ X :

Y is cofinite in X}.
Remark 1.7.5. Let F be a filter on X. Consider IF = {X \ Y : Y ∈ F}. Then
IF is an ideal of the ring (P (X), ∅, X, M, ∩). Dually, every ideal of P (X) gives
rise to a filter: there is a bijection between filters on X and ideals of P (X).

The collection of filters on X may be (partially) ordered by inclusion. The

bijection of Remark 1.7.5 is order-preserving.
Definition 1.7.6 (ultrafilter). A maximal filter on X is called an ultrafilter.
Lemma 1.7.7. Let U be a filter on X. Then U is an ultrafilter iff for all A in
P (X), A ∈ U or X \ A ∈ U.
Proof . Let U be a filter, and bear in mind that U is closed under finite inter-
sections.
Suppose that U is an ultrafilter, and let A ∈ P (X). There are three cases.
If for all Y ∈ U, A ∩ Y 6= ∅, then U ∪ {A} is a family as in Remark 1.7.3, so it is
contained in a filter. By maximality of U as a filter, it follows A ∈ U. If for all
Y ∈ U, (X \ A) ∩ Y 6= ∅, then X \ A ∈ U similarly. So it remains the case where
there are Y1 , Y2 ∈ U such that A ∩ Y1 = (X \ A) ∩ Y2 = ∅. But then Y1 ∩ Y2 = ∅,
a contradiction.
Suppose that for all A ∈ P (X), A ∈ U or X \ A ∈ U. Let Û be a filter
extending U: we show U = Û. Let Y ∈ Û. If Y ∈ U we are done. Otherwise
X \ Y ∈ U ⊆ Û, so ∅ = Y ∩ (X \ Y ) ∈ Û, a contradiction.
Ultrafilters trivially exist.
Definition 1.7.8 (principal ultrafilter). Let a ∈ X. The principal ultrafilter
on a is Pa = {Y ⊆ X : a ∈ Y }.

75
Remark 1.7.9. If X is finite, then every ultrafilter on X is principal.
In the interesting case (X infinite), do non-principal ultrafilters exist? As-
suming AC (or “Zorn’s Lemma”), they do; AC is actually slightly stronger than
necessary to show this. In fact, there is an Ultrafilter Axiom which is inde-
pendent of ZF (and weaker than AC) saying that every filter is included in an
ultrafilter. Applying this to the Fréchet filter on X (Example 1.7.4), we find
non-principal ultrafilters. This is however highly non-constructive.
Lemma 1.7.10 (AC, or “Ultrafilter Axiom”). Every filter is included in an
ultrafilter.

Remark 1.7.11. F is an ultrafilter iff IF is a maximal ideal. In particular,

Lemma 1.7.10 is equivalent to Krull’s Theorem in Ring Theory.
Verification: Clear, as the bijection of Remark 1.7.5 is inclusion-preserving. ♦
Remark 1.7.12. Alternately, an ultrafilter can be viewed as a {0, 1} measure
on X (where a set has measure 1 if it is in the ultrafilter, 0 otherwise). This
measure is however not σ-additive since for instance if we look at the Fréchet
filter on N each singleton has measure 0 but their union (equal to N) has measure
1.
Lemma 1.7.13 (AC). If B has the finite intersection property, there is an
ultrafilter U such that B ⊆ U.

Proof . As B has the finite intersection property, it is contained in a filter F by

Remark 1.7.3. Then Lemma 1.7.10 gives an ultrafilter U extending F.
Card X
Remark 1.7.14. There are 22 ultrafilters on X. (This theorem, due to
Hausdorff, is non-trivial.)

1.7.2 Ultraproducts (The Łoś Structure)*

Definition 1.7.15 (ultraproduct). Let I 6= ∅ be a set, U be an ultrafilter on I,
and for each i ∈ I let Mi be an L-structure.
Q We define the ultraproduct of the
Mi ’s (with respect to U), denoted I Mi /U . For simplicity, we refer to it as
M∗ .
The base of M∗ is the set I Mi modulo the equivalence relation
Q

(mi ) ∼ (ni ) if {i ∈ I : mi = ni } ∈ U

We interpret constants in L by:

∗
cM = (cMi )i∈I
 

We interpret relations by:

∗ 
RM (m1i ) , . . . , (mki ) i ∈ I : RMi (m1i , . . . , mki ) ∈ U
  

if

76
 Similarly for functions:
∗ 
fM (m1i ) , . . . , (mki ) = (mk+1
  
 
i )

if
Mi
m1i , . . . , mki = mk+1


i∈I:f i ∈U
Checking that these are well-defined is very similar to showing the transitiv-
ity of the equivalence relation: if we have two equivalent sequences, then they
must agree on a set in U, which means that the functions or relations must agree
on a set containing a set in U, hence lying in U.
Theorem 1.7.16 (Łoś). Let ϕ be a wff. Let s∗ : V → M ∗ be an assignment,
say s∗ (x) = [(mi,x )i∈I ], and let si (x) = mi,x ; each si is an assignment V → Mi .
Then M∗ |= ϕ[s∗ ] iff {i ∈ I : Mi |= ϕ[si ]} ∈ U.
Proof . By induction on ϕ.
• If ϕ is atomic then it is clear from our definition of M∗ .
• Now suppose that ϕ is ¬ψ. If M∗ |= ¬ψ[s∗ ], then M∗ 6|= ψ[s∗ ]. By
induction, this means that {i ∈ I : Mi |= ψ[si ]} 6∈ U, and since U is an
ultrafilter (so it always contains a set or its complement), this means that
{i ∈ I : Mi 6|= ψ[si ]} ∈ U, i.e. that {i ∈ I : Mi |= ¬ψ[si ]} ∈ U. Hence
M∗ |= ϕ[s∗ ]. All of these steps are reversible, so the iff holds.
• Next assume that ϕ is ψ ∧ χ. Then:

M∗ |= (ψ ∧ χ)[s∗ ]
iff M∗ |= ψ[s∗ ] and M∗ |= χ[s∗ ]
iff {i ∈ I : Mi |= ψ[si ]}, {i ∈ I : Mi |= χ[si ]} ∈ U
iff {i ∈ I : Mi |= ψ[si ]} ∩ {i ∈ I : Mi |= χ[si ]} ∈ U
iff {i ∈ I : Mi |= (ψ ∧ χ)[si ]} ∈ U

(We heavily used that U is an ultrafilter here.)

• Finally, assume that ϕ is ∃x ψ.
– If M∗ |= (∃x ψ)[s∗ ], there is an assignment (s∗ )0 which agrees with s∗
except on x, and such that M∗ |= ψ[(s∗ )0 ]. Say (s∗ )0 (x) = [(ni )] for
a family (ni ) ∈ Πi Mi , and for each i ∈ I let s0i agree with si except
on x, where s0i (x) = ni . By induction {i ∈ I : Mi |= ψ[s0i ]} ∈ U,
which implies {i ∈ I : Mi |= (∃x ψ)[si ]} ∈ U.
– Now suppose that J = {i ∈ I : Mi |= (∃x ψ)[si ]} is in U. For
i ∈ J, let ni ∈ Mi and s0i agree with si on V \ {x} where s0i (x) = ni
be such that Mi |= ψ[s0i ]. For i ∈ I \ J, let s0i = si . We find a
resulting assignment (s0 )∗ : V → M ∗ which differs from s∗ only on
x. As J ⊆ {i ∈ I : Mi |= ψ[s0i ]} is in U, we find by induction that
M∗ |= ψ[(s0 )∗ ]. Hence M∗ |= ϕ[s∗ ].

77
1.7.3 Alternate Proof of Compactness*
Proof of compactness of first-order logic, Theorem 1.6.4. Fix a finitely
satisfiable first-order theory Σ. Let I be the collection of finite subsets of Σ.
For each i ∈ I, there is a structure Mi satisfying Σi by the definition of finite
satisfiability.
For each ϕ ∈ Σ, let Aϕ = {i ∈ I : Mi |= ϕ}. Then the family {Aϕ :
ϕ ∈ Σ} has the finite intersection property, since the intersection of a finite
family corresponds to Aϕ1 ∧···∧ϕn , which must be nonempty since Σ is finitely
satisfiable. By QLemma 1.7.13, there is an ultrafilter U extending {Aϕ : ϕ ∈ Σ}.
We let M∗ = Mi /U (i.e. M∗ is the ultraproduct of the Mi ’s with respect
to U).
Let ϕ ∈ Σ. We have (using Łoś’ Theorem and our definitions):

M∗ |= ϕ ⇔ {i ∈ I : Mi |= ϕ} ∈ U ⇔ Aϕ ∈ U,

and the latter is true by construction, so M∗ |= Σ.

Remark 1.7.17. This modern and elegant proof of first-order compactness
has one drawback: it is highly non-constructive, as it relies on the existence of
ultrafilters (Lemma 1.7.13). On the other hand, it is purely semantic and does
not rely on a hidden completeness argument!
Ultraproducts are very important in model theory, if one is not interested
in effectiveness. They appear in important constructions and deep results. For
instance, one could have realized the non-standard reals asQan ultrapower of
R: fix a non-principal ultrafilter U on N, and let R∗ = N R/U. By Łoś’
Theorem, R∗ is an ordered field. If you want to feel an infinitesimal, consider
the equivalence class, modulo U, of the sequence ( n1 )n∈N : it is positive, and
ultimately smaller than any fixed positive real number!

End of Lecture 17.

1.8 Exercises
Exercise 1.1. State and prove a version of the Unique Readability Theorem
for terms of a first-order language.
Exercise 1.2. Let L be a purely relational first-order language (i.e., the signa-
ture consists only of relation symbols). State and prove the Unique Readability
Theorem for L-first-order wff’s.
Exercise 1.3. Let ϕ be the wff (∀v1 v1 = v2 ) → (∀v2 v2 = v1 ).
1. Is v3 substitutable for v1 in ϕ? If yes, give ϕ[v3 /v1 ].
2. Is v1 substitutable for v1 in ϕ? If yes, give ϕ[v1 /v1 ].
3. Is v2 substitutable for v1 in ϕ? If yes, give ϕ[v2 /v1 ].

78
 4. Let t = f (v1 , v2 ). Find a suitable ϕ̃ given by Theorem 1.3.6, when one
wishes to replace v1 by t. Then give ϕ̃[t/v1 ].
Exercise 1.4. Let L be the empty language (pure equality).
1. Write an axiom saying that there at least n elements.

2. Write an axiom ϕn saying that there exactly n elements.

3. Write a theory T saying that there are infinitely many elements.
4. Let ψ be a sentence. Show that there is a structure satisfying ψ but not
T.

Exercise 1.5. Let L = {s}, s a unary function symbol.

1. Write an axiom ϕ saying that s is injective.
2. Write an axiom ψ saying that s is surjective.

3. Classify structures satisfying of {ϕ, ψ} (i.e., describe them all).

Exercise 1.6. Let L be any language, and M be any structure.
1. Assume M finite. Show that there is a sentence ϕ true in M such that
for any L-structure N , N |= ϕ iff Card M = CardN .

2. Show that this is not true for infinite M: find two infinite L-structures M
and N which satisfy the same sentences, but with different cardinalities.
Exercise 1.7. Let E be a binary relation symbol and consider the first-order
language L = {E}.
1. Write a (finite!) set of axioms T expressing that E is an equivalence
relation.
From now on we’ll implicitely be working with extensions of T (theories
extending T ). So it makes sense, for us, to speak about equivalence classes
(we say classes, for short).
2. Write an axiom saying that every class is reduced to a single element.

3. Write an axiom saying that there is a class with at least n elements.

4. Write an axiom saying that every class has at least n elements.
5. Write an axiom saying that there is exactly one class with at least n
elements.

6. Write an axiom saying that there is a class with exactly n elements.

7. Write an axiom saying that there is exactly one class with exactly n ele-
ments.

79
 8. Write an axiom saying that there are exactly m classes with exactly n
elements.
9. Write a theory saying that every class is infinite.
10. Is there a theory satisfied exactly by (sets equipped with) an equivalence
relation with an infinite class?
No justification required.
Exercise 1.8. Let 0 be a constant symbol, s a unary function symbol and
consider the following theory:

T = {∀v1 ∀v2 v1 6= v2 → s(v1 ) 6= s(v2 )} ∪ {∀v1 s(v1 ) 6= 0}

∪{∀v1 ∃v2 v1 6= 0 → v1 = s(v2 )}
· · ◦ s}(v1 ) 6= v1 : n ∈ N}
∪{∀v1 |s ◦ ·{z
n times
Describe all structures satisfying T .
Exercise 1.9. The language has a unary relation S.
1. Let M be a non-empty set. Prove (in English):

∃x ∈ M ∀y ∈ M S(x) → S(y)

2. Write a deduction of:

∃x ∀y (S(x) → S(y))

Exercise 1.10. Suppose the ∀ rules and the axiom v1 = v1 .

1. Show that for any variable x, ` x = x.
2. Show that for any L-term t, ` t = t.
Exercise 1.11. Suppose the ∀ rules, the axiom v1 = v1 , and for any function
symbol f the axiom

(v1 = vn+1 ∧ · · · ∧ vn = v2n ) → f (v1 , . . . , vn ) = f (vn+1 , . . . , v2n )

If t is a term and (si ) is a family of terms indexed by i ∈ N, we write t[si /vi ]

for the term obtained from t where for each i, every occurence of vi is replaced
by si . (There is also an obvious notion of substitutability.)
Let (si ) and (ri ) be two families of terms. Proceeding as follows, show
` (r1 = s1 ∧ · · · ∧ rn = sn ) → t[ri /vi ] = t[si /vi ]:
1. Show that if t is a constant, then the claim holds.
2. Show the same if t is a variable.
3. Show the result for any t.

80
Exercise 1.12. Suppose the ∀ rules, the axiom v1 = v1 , for any function symbol
f the axiom

(v1 = vn+1 ∧ . . . vn = v2n ) → f (v1 , . . . , vn ) = f (vn+1 , . . . , v2n )

and for any relation symbol R the axiom

(v1 = vn+1 ∧ · · · ∧ vn = v2n ) → (R(v1 , . . . , vn ) → R(vn+1 , . . . , v2n )

Sketch a proof (but no deductions, it would be too long!) of the fact that
for any families of terms (si ) and (ri ) and any formula ϕ in which they are
substituable for the vi ’s, ` (r1 = s1 ∧ · · · ∧ rn = sn ) → (ϕ[ri /vi ] ↔ ϕ[si /vi ]).
Just decompose the proof into steps, but do not give any details!
Exercise 1.13. Finish the proof of Proposition 1.4.2 by showing that = is
transitive, that is show ` ∀x ∀y ∀z (x = y ∧ y = z → x = z).
Exercise 1.14. Let ϕ be a wff, t a term, x a variable, and ϕ̃ given by the
renaming algorithm (Theorem 1.3.6) such that t is substitutable for x in ϕ̃.
Show (we haven’t done it yet) that {ϕ} ` ϕ̃ and {ϕ̃} ` ϕ.
Exercise 1.15. Formalize first-order logic using ∃, and consider ∀ as an abbre-
viation. Show that this is consistent syntactically and semantically.
Exercise 1.16. Prove the soundness theorem by induction with both quanti-
fiers.
Exercise 1.17 (Gödel translation, continued). We pursue in the vein of Exer-
cise 0.19. Define:
• ϕ∗0 = ϕ0 for ϕ0 an atomic formula;
• (∀x ϕ)∗ = ∀x ϕ∗
Prove the same conclusions as in Exercise 0.19.
Exercise 1.18. For n ∈ N, let ϕn be the formula ∃v1 . . . ∃vn ∧i6=j vi 6= vj . Let
Σ = {ϕn }. Show that Σ 6` ϕn+1 .
Exercise 1.19. The language is the language of groups.
1. Give a finite theory whose models are all groups.
2. Give a theory whose models are exactly all infinite groups.
3. Find a finite theory whose models are infinite groups (requires some cul-
ture).
4. Show that there is no theory whose models are exactly all finite groups.
Exercise 1.20. The language has a binary relation. Prove that there is no
theory whose models are all connected graphs.

81
Exercise 1.21 (definable sets). Let M be a first-order structure in a first-order
language L.
1. A subset X ⊆ M is definable if there is a formula ϕ with FreeVar ϕ = {v0 }
such that for all assignment s: s(v0 ) ∈ X iff M |= ϕ[s].
(Intuitive meaning: m is in X iff ϕ is true of m.)
Show that {0} is definable in (N, <). Show that {1} is definable in (N, <).
Show that the set of prime numbers is definable in (N, 0, 1, +, ·).
2. A subset X ⊆ M is definable with parameters µ = (µ1 , . . . , µk ) ∈ M k
if there is a formula ϕ with FreeVar ϕ = {v0 , . . . , vk } such that for all
assignment of the variables s mapping vi to mi (i = 1 . . . k): s(v0 ) ∈ X iff
M |= ϕ[s].
(Intuitive meaning: m is in X iff ϕ(m, µ1 , . . . , µk ) is true.)
√
Show that { π} is definable with parameters π in (R, 0, 1, +, −, ·).
Show that R> 0 is definable in (R, 0, 1, +, −, ·).
3. A subset X ⊆ M n is definable with parameters µ ∈ M k if there is a for-
mula ϕ with FreeVar ϕ = {x1 , . . . , xn , y1 , . . . , yk } such that for all assign-
ment of the variables s mapping yi to µi (i = 1 . . . k): (s(v1 ), . . . , s(vn )) ∈
X iff M |= ϕ[s].
Show that the function ÷ is definable in (R, 0, 1, ·). (i.e., show that its
graph is definable.)
4. Bonus questions: R is not definable in (C, 0, 1, +, ·). Z is not definable
in (R, 0, 1, +, ·). All definable subsets of C in (C, 0, 1, +, ·) are finite or
cofinite.
Exercise 1.22.
1. Show that {k ∈ Z : k is odd } is definable in (Z, +, ·).
2. Show that each element of Q is definable in (R, 0, 1, +, −, ·).
3. Show that {i, −i} is definable in (C, 0, 1, +, −, ·). Show that {i} is not.
4. Show that R is not definable in (C, 0, 1, +, −, ·).
Exercise 1.23 (non-standard models of (N, s)). Let L = {0, s} where 0 is a
constant symbol and s a unary function; let c be another constant symbol and
L0 = L ∪ {c}. Let T consist of all sentences true in (N, 0, s) where s is the
successor function.
1. Let T 0 = T ∪ {c 6= s| ◦ ·{z
· · ◦ s} : n ∈ N}. Show that T 0 is consistent.
n times
2. Let N0 be a model of T 0 . Construct a canonical injective function N ,→ N0 .
Is it surjective?

82
 3. Let X ⊆ N0 be a definable, non-empty subset (definable by a formula with
no parameters). Show that X has a least element.
Exercise 1.24. Let M be a finite first-order structure. Let T be the collection
of all L-sentences true in M. Show that all models of T have same cardinal as
M. (They are actually all isomorphic to M, whatever that means.) Show that
this fails if M is infinite.
Exercise 1.25.
1. Let T be a complete theory. Suppose that T has an infinite model. Show
that all models of T are infinite.

2. Show that this might fail if T is not complete.

3. Give, in a suitable language, an incomplete theory whose models are infi-
nite.
Exercise 1.26 (the collection of definable sets). Let M be a first-order struc-
ture. Let Dn be the collection of definable subsets of M n and D the union of
the Dn ’s.
1. Show that each Dn is closed under complement, finite union, finite inter-
section.
2. Show that if X, Y ∈ D, then X × Y ∈ D.

3. Let π1 : M 2 → M be the projection on the first coordinates. Show that if

X ∈ D2 , then π1 (X) ∈ D1 . Show that if X ∈ D1 , then π1−1 (X) ∈ D2 .
Exercise 1.27. Give an example of a language L, an L-structure M and a wff
ϕ with FreeVar ϕ = {v1 } such that M |= ∃v1 ϕ but there is no L-term t such
that M |= ϕ[t/v1 ].
Exercise 1.28. Suppose that Σ ` ϕ and that P is a predicate symbol which
appears in neither Σ nor ϕ. Show that there exists a deduction of ϕ from Σ
which doesn’t involve P . (Hint: use the completeness theorem.)
Exercise 1.29. Let L be the language of groups, and T be the theory of Z in
this language, that is T = {ϕ an L-sentence such thatZ |= ϕ}.
Let L0 = L ∪ {c}, where c is a new constant symbol.
1. Let n ∈ Z. Write an L0 -sentence saying that c is a multiple of n.
2. Let I be a set of prime numbers. Write an L0 -theory extending T such
that only the primes in I divide c.
3. Deduce that there are uncountably many distinct complete L0 -theories
extending T .

83
Exercise 1.30 (universal closures). Let ϕ be a formula in a first-order language
L, and let FreeVar ϕ = {x1 , . . . , xn }. We define the universal closure of ϕ to be

UnCl ϕ : ∀xn . . . ∀x1 ϕ

1. Give the universal closure of (∀v1 v1 = v2 ) → (v1 = v2 ).

2. We have ordered the free variables by the natural ordering. To con-

vince yourself that this is not essential, show that for any formula ψ wth
FreeVar ψ = {x, y}, one has

{∀x ∀y ψ} ` ∀y ∀x ψ and {∀y ∀x ψ} ` ∀x ∀y ψ

3. Show that {UnCl ϕ} ` ϕ.

4. Is the converse true?
Exercise 1.31. Let L have the following signature:

• a binary predicate symbol <;

• two constant symbol a and b;
• a unary function symbol f .
Let T be the L-theory with the following axioms:

• ∀v1 ¬(v1 < v1 )

• ∀v1 ∀v2 (v1 < v2 ∨ v2 < v1 ∨ v1 = v2 )
• ∀v1 ∀v2 ∀v3 (v1 < v2 ∧ v2 < v3 ) → (v1 < v3 )

• ∀v1 ∀v2 (v1 < v2 ) → (∃v3 v1 < v3 ∧ v3 < v2 )

• ∀v1 ∃v2 ∃v3 v2 < v1 ∧ v1 < v3
• a<b
• ∀v1 ∀v2 (v1 < v2 ) → (f (v1 ) < f (v2 ))

Let σ be the sentence : f (a) < b. Prove that T 6` σ and that T 6` ¬σ.

84
Chapter 2

Second-Order Logic
(Presto)

In this short chapter we briefly describe second-order logic, in which one may
quantify on collections of elements. This requires new symbols ∀2 and ∃2 . The
logic is amzazingly expressive, too expressive to have interesting results.

In this chapter:
• Explain what second-order means.

• Show that it is essentially more expressive than first-order.

• Show that it is essentially too expressive for logic.

Lecture 18 (Second-Order Logic)

Definition 2.0.1 (second-order language). We now expand our existing collec-
tion of symbols by the following:

• for each k ≥ 1, a set of k-ary predicate variables X1k , X2k , . . . , Xnk , . . .

• the quantifiers ∃2 and ∀2
∀2 now quantifies over collections/relations/functions (a relation is a set
of n-uples; a function is a certain relation). For instance, X 1 (a) means that
the element a is in the predicate X. Or, the formula Y 2 (a, b) means that the
elements a and b are in relation according to the binary predicate Y . In a word,
one may now quantify over predicates.
Caution! The superscript of a variable denotes its arity, that is the Cartesian
power of the base set in which it lives. But the superscript in ∀2 and ∃2 indicates
the order of the logic.

85
 One should define terms and wff’s, but we do not wish to repeat these
discussions, as we shall not try to describe a proof theory. There are two reasons.
• First, we would have to determine suitable axioms for the relationship
between predicates and elements, as we determined reasonable axioms
for equality in the case of first-order logic. This amounts to agreeing on
axioms for membership, that is agreeing on a theory of sets, which is not
the spirit of this course.
• Second, and worse, this appears a little pointless to us, as compactness
will fail anyhow: hence so would completeness.
We therefore entirely forget about proof theory.

2.1 Compactness fails

Proposition 2.1.1. There is a second-order formula expressing that a set is
infinite.
Proof . A set is infinite iff it injects into a proper subset. There is a formula
saying that Y 1 is a subset of X 1 :

subset(Y 1 , X 1 ) : ∀x Y 1 (x) → X 1 (x)

There is a formula saying that Y 1 is a proper subset of X 1 :

propersubset(Y 1 , X 1 ) : subset(X 1 , Y 1 ) ∧ (∃x X 1 (x) ∧ ¬Y 1 (x))

There is a formula saying that Γ2 is the graph of a functional relation:

functional(Γ2 ) : ∀a∀b1 ∀b2 Γ2 (a, b1 ) ∧ Γ2 (a, b2 ) → b1 = b2

There is a formula saying that the relation Γ2 is injective:

injective(Γ2 ) : ∀a1 ∀a2 ∀b Γ2 (a1 , b) ∧ Γ2 (a2 , b) → a1 = a2

There is a formula saying that Γ2 defines a function from X 1 to Y 1 :

function(Γ2 , X 1 , Y 1 ) : functional(Γ2 ) ∧ ∀x ∃y X 1 (x) → (Y 1 (y) ∧ Γ2 (x, y))

There is a formula saying that Γ2 defines an injective function from X 1 to Y 1 :

injection(Γ2 , X 1 , Y 1 ) : injective(Γ2 ) ∧ function(Γ2 , X 1 , Y 1 )

There is a formula saying that X 1 injects into Y :

injects(X 1 , Y 1 ) : ∃2 Γ2 injection(Γ2 , X 1 , Y 1 )

And eventually, the following formula expresses that X 1 is infinite:

infinite(X 1 ) : ∃2 Y 1 propersubset(Y 1 , X 1 ) ∧ injects(X 1 , Y 1 )

86
 One should compare this to first-order logic, in which “infinite” is not a single
formula, but an infinite theory not finitely axiomatizable (Example 1.6.7). Exercise 2.1

Corollary 2.1.2. No compactness theorem.

Proof . For n ∈ N, let ϕ be the (first-order) sentence: ∃v1 . . . ∃vn ∧i6=j vi 6= vj ,
which expresses that the underlying set has at least n elements. Let ψ be the
formula constructed in Proposition 2.1.1 and consider the following second-order
theory:
T = {ϕ: n ∈ N} ∪ {∀2 X 1 ¬ψ(X 1 )}
Then T is clearly not satisfiable, but is finitely satisfiable!
In particular, regardless of what our proof theory would have been, com-
pleteness would have failed.

2.2 Peano Arithmetic

We now deal with a remarkable property of the integers.
Definition 2.2.1 (Peano Arithmetic). Let 0 be a constant symbol and s a
unary function symbol. Let PA be the second-order theory in the language
{0, s} with the following axioms:

• s is an injective function
• every element not 0 is in the image of s
• there are no cycles, i.e. for each n the axiom:

∀x s| ◦ ·{z
· · ◦ s}(x)¬x
n

• Every subset which contains 0 and is closed under s is the entire base set:

∀2 Y 1 Y 1 (0) ∧ (∀y Y 1 (y) → Y 1 (s(y)) → ∀x Y 1 (x)

Theorem 2.2.2 (absolute categoricity of PA). All models of Peano Arithmetic

are like N (with the natural interpretations of 0 and s as the successor function).
“Like” stands for isomorphic, which we do not wish to define now (this will
be done in model theory, Definition 1’.2.6). It just means that the bijection
preserves the language; that up to names, the objects are exactly the same.
This result will be referred to as absolute categoricity of PA, a terminology
which will become clearer after the definition of categoricity in model theory
(Definition 1’.2.14).

87
Proof . Let M |= PA. We find a bijection between N and M which maps 0 to
0M and which is compatible with succession. For the sake of clarity, we write
n + 1 for sN (n) (succession in N happens to coincide with adding 1, though
addition is not part of the language).
Let f (0) = 0M . Inductively, we now define f (n + 1) = sM (f (n)). This does
define a function N → M , which is by definition compatible with s. Clearly f
is injective (induction in N; s has no cycles in M).
It remains to show that f is surjective. Consider im f ⊆ M . This subset
contains 0M and is closed under sM ; as M satisfies the axioms of PA, one has
im f = M . So f is surjective; it is a successor-preserving bijection.

Caution! The temptation is strong to wonder whether or not N is a model of

PA. But the integers existed before we started doing logic, and the fact that we
have found a nice description of their behaviour does not change it!

There is no reason to stop. One may quantify over collections of collections

of elements and introduce ∀3 and so on. Goes without saying that nth -order
logic remains sound, but does not enjoy compactness.

End of Lecture 18.

Lecture 19 (ME2)

End of Lecture 19.

Exercise 2.1. Find second-order formulas expressing that a set A is:

1. countable
2. uncountable

3. of cardinality ℵ1
4. of cardinality ℵn
5. of cardinality continuum (that is 2ℵ0 , like R).

Hint: for the last one, use the following algebraic fact. There is up to isomor-
phism a unique ordered field which is complete.

88
Chapter 1’

Some Model Theory

(Furioso)

After the failure of compactness in higher logic, we move back to first-order

logic, which is obviously the only logic worth considering. Model theory is the
mathematical study of first-order theories; syntactical considerations disappear,
if not entirely, at least in spirit. One will still argue by induction on the length
of formulas, which is a syntactic notion, but there also exists a purely semantic
approach to model-theory, which emphasizes on back-and-forth methods. For
these reasons, model theory has been called the semantics of predicate calculus.

In this chapter: a first account of model theory

• Elementary equivalence, elementary inclusion (§1’.1)

• Elementary morphisms (§1’.2)

• Löwenheim-Skolem theorems (§1’.3)
• Back-and-forth methods (§1’.4)
• Quantifier elimination (§1’.5.1)

• ω-saturated models and applications (§1’.5.2 and 1’.5.3)

Lecture 20 (Models and Theories; Elementary Equivalence)

1’.0 A Word on Model-Terrorists

We are through with syntax, and will start doing some mathematics. Our
treatment of terms and formulas has been remarkably accurate as long as we

89
have been interested in proof theory; model theory is on the semantic side, so we
shall make notations more convenient, and disregard the risk of an ambiguity.

1’.0.1 Formulas and Parameters

Notation 1’.0.1 (tuple notation). A finite sequence of objects (a1 , . . . , an ) is
more conveniently written a (n is then the length of the tuple a).
We shall often abuse notation of Cartesian powers in order to avoid mention-
ing lengths. For instance, if a1 , . . . , an ∈ M , we shall feel free to write a ∈ M
instead of a ∈ M n .
From now on we assume that every formula we consider has been suitably
constructed or renamed (for instance by the algorithm of Theorem 1.3.6); that
is, we take for granted that no (∀v1 v1 = v2 ) → (v1 = v2 ) will play us tricks.
Furthermore, we take x, y, z, etc. but also x1 , x2 , etc. to be variables them-
selves (and not names for general variables). In particular x = x1 , . . . , xn is a
tuple of variables.
Notation 1’.0.2 (formula with free variables). We write ϕ(x) to indicate that
the only free variables of the formula ϕ are among the tuple of variables x.
Hence ϕ(x) means FreeVar ϕ ⊆ x; but we may write ϕ(x) even if some
variables of the tuple x are not used. In any case, ϕ (with no x) denotes a
sentence. We now get rid of assignments of the variables.
Notation 1’.0.3 (formula with parameters). Let ϕ(x) be a formula with free
variables among x, and M an L-structure. Let m = (m1 , . . . , mn ) ∈ M n have
same length as x. We write M |= ϕ(m) if M |= ϕ[s] with s : xi 7→ mi .
It is customary to expand the language via constants.
Notation 1’.0.4 (expansion by constants; LA ). Let L be a first-order language,
and A a set. LA denotes the language obtained by adding to L new constant
symbols ca for elements a of A.
In this case, we shall always interpret ca by a. Under this convention, M |=
ϕ(a) iff M |= ϕ(ca1 , . . . , can ).

1’.0.2 Quantifications
We rehabilitate poor ∃ and make it a quantifier again.
Definition 1’.0.5 (quantifier-free formula). A formula is quantifier-free if it has
no quantifiers.
Definition 1’.0.6 (universal formula). A formula is universal if it is of the form
∀v1 . . . ∀vn ϕ0 , where ϕ0 is quantifier-free.
Definition 1’.0.7 (existential formula). A formula is existential if it is of the
form ∃v1 . . . ∃vn ϕ0 , where ϕ0 is quantifier-free.

90
 Be very careful that ∀v1 ∃v2 . . . is neither universal nor existential.

Definition 1’.0.8 (quantification rank). We define the quantification rank of a

formula ϕ.
• if ϕ is atomic, then qrk ϕ = 0;
• if ϕ is ¬ψ, then qrk ϕ = qrk ψ;

• if ϕ is ψ1 → ψ2 , then qrk ϕ = max(qrk ψ1 , qrk ψ2 );

• if ϕ is ∀x ψ, then qrk ϕ = qrk ψ + 1.
• if ϕ is ∃x ψ, then qrk ϕ = qrk ψ + 1.
We do not capture the alternation of ∀’s and ∃’s, though a recursion theorist
would feel at some point interested.

1’.0.3 Models and Theories

We shall work only with theories (sets of sentences) instead of arbitrary sets of
formulas. Indeed, let us introduce a new constant cx for each variable x, and let
L0 = L ∪ {cx : x a variable}. To every (possibly open) L-formula ϕ we associate
an L0 -sentence ϕ0 by replacing every free occurence of every free variable by the
corresponding constant; notice that subsitutability is not an issue. And a set
of L-formulas Σ is satisfiable iff the L0 -theory Σ0 = {ϕ0 : ϕ ∈ Σ} is satisfiable.
The advantage being that since Σ0 is a set of sentences, it suffices to provide an
L0 -structure: assignments of the variables are now useless.
As we shall never use truth values again, T denotes a first-order theory.
Definition 1’.0.9 (model). Let M be an L-structure and T be an L-theory. If
M |= T one says that M is a model of T .
Model-theorists often abuse terminology and say that T is “consistent” for
“satisfiable”; by Gödel’s Completeness Theorem (Theorem 1.5.1) this is harm-
less anyway. But one should not think that model-theorists are interested in
syntax; all they care about is structures. We even abuse further, by sometimes
implicitely defining a theory to be consistent. This leads to questions like: “is
this theory a theory?” (which means: “is this theory consistent?”). No ambi-
guity arises in practice.
Also, we freely identify any theory with the set of its consequences (since
there is no ambiguity on the notion of consequence). This bridges the gap
between two different definitions of completeness (Definitions 0.4.3 and 1.6.2),
which now perfectly agree. Ex. 1’.2, 1’.6

91
1’.1 Elementary Equivalence; Inclusion and El-
ementary Inclusion
Remark 1’.1.1. “Elementary” is the standard adjective used in model theory;
one must be careful with the definitions (elementary equivalence, Definition
1’.1.5, and elementary inclusion, Definition 1’.1.13), as the role of parameters
is not necessarily obvious. Terminology tends to be confusing; “elementary”
captures a way of thought, not a notion.

1’.1.1 Elementary Equivalence

Definition 1’.1.2 (theory of a structure). Let M be an L-structure. The first-
order L-theory of M, or the theory of M for short, is Th M = {ϕ : M |= ϕ}.
This amounts to taking only L-sentences which are true in M; if one wants
to allow parameters, we must generalize a little. Recall that LA is the expansion
of L by new constant symbols for elements of A (Notation 1’.0.4).

Definition 1’.1.3 (theory of a structure, with parameters). Let M be an L-

structure and A ⊆ M a set of parameters.
• The theory of M over A is Th(M, A) = {ϕ(a) ∈ LA : M |= ϕ(a)}.
• The quantifier-free theory of M over A is Th0 (M, A) = {ϕ(a) ∈ LA :
M |= ϕ(a) and ϕ is quantifier-free}.

In particular, Th M = Th(M, ∅). One may also notice that Th0 (M, A) =
Th(M, A) ∩ {ϕ ∈ LA : qrk(ϕ) = 0}.
Remark 1’.1.4. If M is an L-structure and A ⊆ M , then Th(M, A) is a
complete LA -theory.

Verification: In M, ϕ(a) is either true or false! ♦

Model theory is interested in the theories of structures. As far as first-order
properties are concerned, “being the same” amounts to satisfying the same
sentences. This is captured by the following, central notion.
Definition 1’.1.5 (elementary equivalence). Let M, N be L-structures. M
and N are elementarily equivalent, denoted M ≡ N , if Th M = Th N .
This behaves very much like an equivalence relation (reflexive, symmetric,
transitive), except that the class of all L-structures is not, set-theoretically
speaking, a set.

Lemma 1’.1.6. A theory T is complete iff for all models M, N of T , one has
M ≡ N.

92
Proof . Recall that we have identified T with the set of its consequences, and
that a theory is complete iff it is maximal as a consistent set of sentences.
Suppose T complete. If M |= T , then T ⊆ Th M. As T is complete, one
has T = Th M; since T = Th N similarly, it follows M ≡ N .
Suppose that any two models of T are elementarily equivalent. Let ϕ be a
sentence such that neither ϕ not ¬ϕ is in T . Then there are models M |= T ∪{ϕ}
and N |= T ∪ {¬ϕ}. In particular, M ≡ N , and this is a contradiction. So
either ϕ or ¬ϕ is in T .
So the running question: is this theory complete? Has a partial answer: can
one show that any two of its models are elementarily equivalent?

1’.1.2 Inclusion
We now turn our attention to substructures; the naive notion (Definition 1’.1.7)
is not very useful, the interesting form will be elementary inclusion (Definition
1’.1.13). As noted in Remark 1’.1.1, terminology might be a bit confusing, as “el-
ementary” will not exactly mean the same thing as in “elementary equivalence”
(Definition 1’.1.5).
But let us begin with the naive, not-so-useful, notion of inclusion.
Definition 1’.1.7 (substructure). Let M, N be L-structures. Then M is a
substructure of N , written M ⊆ N , if the following hold:
• M ⊆ N;

• cM = cN for all constant symbols c ∈ L;

• RM = RN ∩ M n for all n-ary relation symbols R ∈ L;
• f M = f N ∩ M n+1 for all n-ary function symbols R ∈ L.

|M
(f N ∩M n+1 may also be written f N |M n , the restriction and corestriction
of f N to M )
In a word, M bears the L-structure induced by restriction of the structure
of N to the base M . (In general, if you found two structures M and N in the
same language, with M ⊆ N but it is not the case that M ⊆ N as L-structures,
then presumably there is something wrong.)
Example 1’.1.8.
• A subgroup of a group is a substructure as a group structure.

• A subring of a ring is a subbstructure as a ring structure.

Remark 1’.1.9. Let N be an L-structure and M ⊆ N be a non-empty subset.
Then M can be equipped with an L-structure M ⊆ N iff the following hold:
• for each constant symbol c, cN ∈ M

93
 • for each function symbol f , M is closed under f N .
The structure M is then uniquely determined.
When studying Löwenheim-Skolem theorems we shall say a little more about
generating substructures (§1’.3.2).
The following is a very healthy exercise.
Remark 1’.1.10. Let I be a non-empty set and (Mi )i∈I be an increasing chain
of L-substructures, that is i ≤ j ⇒ Mi ⊆ Mj . Then ∪i∈I Mi naturally bears
an L-structure M∗ such that for all i ∈ I, Mi ⊆ M∗ .

Lemma 1’.1.11 (∃’s go up, ∀’s go down). Let M ⊆ N be an L-structure. Let

ϕ(m) be a formula with parameters in M .
(i). If ϕ is quantifier-free, then M |= ϕ(m) ⇔ N |= ϕ(m).
(ii). If ϕ is existential, then M |= ϕ(m) ⇒ N |= ϕ(m).

(iii). If ϕ is universal, then N |= ϕ(m) ⇒ M |= ϕ(m).

Proof .
(i). The case of an atomic formula is by definition. Then a quick induction on
the complexity.

(ii). Induction again. If there are no quantifiers, then this is known. Suppose
that ϕ is ∃x ψ(x, m); ψ is existential again. If M |= ∃x ψ(x, m), then
there is µ ∈ M such that M |= ψ(µ, m). By induction, N |= ψ(µ, m); so
N |= ϕ(m).

(iii). Might be obtained via (ii), or by similar methods. Exercise 1’.3

In general, “existential witnesses need not go down” (dually, “universal prop-

erties need not go up”).
Counter-example 1’.1.12.

• Z ⊆ Q as ring-structures, but Q |= ∃x x + x = 1 though Z doesn’t.

• Q ⊆ R as ring-structures, but R |= ∃x x2 = 1 + 1 though Q doesn’t.
• R ⊆ C as ring-structures, but C |= ∃x x2 = 1 though R doesn’t.
One sees that this notion of inclusion is therefore not very interesting to us.
Elementary inclusions (Definition 1’.1.13) will be more relevant. Exercise 1’.7

End of Lecture 20.

Lecture 21 (Elementary Inclusion; Morphisms)

94
1’.1.3 Elementary Inclusion
We now introduce a better-behaved notion of inclusion (and substructure). The
intuitive idea is that both structures should agree on as many properties as
possible, allowing parameters. Of course in order for the question to make
sense, only elements of the small structure can be taken into account.
Definition 1’.1.13 (elementary substructure). Let M ⊆ N be an inclusion
of L-structures. M is an elementary substructure of N , written M  N , if
Th(M, M ) = Th(N , M ).
Example 1’.1.14.
• (N, <) 6 (Z, <), as N |= ∀x x ≥ 0 but Z |= ∃x x < 0. (This formula uses
0 as a parameter.)
• (Q, <)  (R, <) (not necessarily obvious yet).
• Let M ⊆ N be infinite sets, regarded as pure sets. Then M  N .
• We have observed in Counter-example 1’.1.12 that none of the inclusions
of rings Z ⊆ Q ⊆ R ⊆ C is elementary.
• As rings, Q  C (not necessarily obvious).
• When we constructed non-standard reals in §1.6.3, we considered a the-
ory extending Th(R, R), in order to force the inclusion R ⊆ R∗ to be
elementary.
In other words, M  N iff M and N have the same LM -theory, iff N |=
Th(M, M ). This is also equivalent to: for any LM formula ϕ(m) with parame-
ters in M , M |= ϕ(m) iff N |= ϕ(n).
In particular M  N implies M ≡ N , but the converse need not hold.
Counter-example 1’.1.15. Consider N = (Z, <) and M = (2Z, <) ⊆ N . As
they are clearly isomorphic (Definition 1’.2.6 below), it is clear that they are
essentially the same object, and M ≡ N (this will be more convincing later).
Yet M |= ∀x (x ≤ 0 ∨ x ≥ 2), and N |= ∃x (0 < x < 2). Hence M and N can
be distinguished by the tuple (0, 2).
Lemma 1’.1.16. Let I be a non-empty set and (Mi )i∈I be an increasing el-
ementary chain of L-substructures, that is i ≤ j ⇒ Mi  Mj . Let M∗ =
∪i∈I Mi . Then for all i ∈ I, Mi  M.
Proof . A very healthy exercise. Exercise 1’.1
Ex 1’.4, 1’.5
The following gives a criterion for elementarity of inclusions; it should be
used mostly in order to go down, that is, when one already has perfectly under-
stood the notion of satisfaction in N , and tries to see if M ⊆ N is an elementary
substructure. (This will typically be the case in §1’.3.2). The statement is more
subtle than it first seems.

95
Theorem 1’.1.17 (Tarksi’s test). Let M ⊆ N be an inclusion of L-structures.
Then M  N iff for all formula ϕ(m) with parameters in M ,

if N |= ∃x ϕ(x, m) then there is µ ∈ M s.t. N |= ϕ(µ, m)

Proof . ⇒ is clear by definition. We show ⇐ by induction. Let ϕ(m) be a

formula with parameters m ∈ M . We show M |= ϕ(m) ⇔ N |= ϕ(m).
• Suppose ϕ atomic. Clearly, M |= ϕ(m) iff N |= ϕ(m).
• Suppose that ϕ is built from shorter formulas using connectives. There is
not much to prove.
• Suppose that ϕ is ∃x ψ.
– Assume that M |= ∃x ψ(x, m). Then there is µ ∈ M such that
M |= ψ(µ, m). By induction, N |= ψ(µ, m), so N |= ϕ(m). Notice
that we did not use the assumption yet.
– Now assume that N |= ∃x ψ(x, m). By assumption, there is µ ∈ M
such that N |= ψ(µ, m). By induction, M |= ψ(µ, m), and therefore
M |= ϕ(m).
In short, the inclusion is elementary iff all ∃’s go down; yet it is not a good
idea to try to put Tarski’s test in a nutshell, as there are some subtleties.
Remark 1’.1.18.
• Elementary equivalence is a notion involving satisfaction in two structures
(M and N ). However, Tarski’s Test reduces to only one notion of sat-
isfaction: in N . Notice indeed that the condition is only about truth
in N . This will make Tarski’s test especially valuable when constructing
elementary substructures (§1’.3.2).
• One should be very careful that in Tarski’s test (Theorem 1’.1.17), one con-
siders all formulas of the form ∃x ϕ, where ϕ too might have quantifiers,
and not only existential formulas (read Counter-example below).
Counter-example 1’.1.19. Consider (N, <) and (N ∪ {∞}, <), where ∞ de-
notes an extra element greater than all natural numbers. Then N ∪ {∞}
has a greatest element (this is a first-order sentence), but N doesn’t: clearly
N 6 N ∪ {∞}. But for a purely existential formula ϕ, one has N |= ϕ(n) iff
N ∪ {∞} |= ϕ(n). (This need not be entirely clear yet, as a proof would rely on
back-and-forth methods.)

1’.2 Morphisms, Elementary Morphisms, Cate-

goricity
We now have the objects of model-theory; we want to find the suitable notion
of morphisms between them.

96
1’.2.1 Morphisms
Notation 1’.2.1. If m = (m1 , . . . , mn ) is a tuple and σ : M → N is a function,
we write σ(m) for the tuple (σ(m1 ), . . . , σ(mn )).
Definition 1’.2.2 (L-morphism). Let M, N be L-structures and σ : M → N
be a function. σ is an L-morphism if:
• σ(cM ) = cN for any constant symbol c;
• m ∈ RM iff σ(m) ∈ RN for any relation symbol R;
• σ(f M (m)) = f N (σ(m)) for any function symbol f .
In short, an L-morphism is a function which is compatible with the interpreta-
tions of L in M and N .
In particular, as equality is always among the relations, one sees that an L-
morphism is always injective. This is the reason why one also says L-embedding.
Remark 1’.2.3. The composition of two L-morphisms is an L-morphism; IdM
is always an L-morphism.
Example 1’.2.4.
• Let M and N be orderings. Then σ : M → N is an {<}-morphism iff it
is an order-preserving (injective) function.
• An injective group homomorphism is an Lgrp -morphism.
• Any ring morphism between fields is an Lrng -morphism.
Remark 1’.2.5.
• If M ⊆ N , then the inclusion map M → N is an L-morphism.
• More precisely, let M, N be two structures with M ⊆ N (the under-
lying sets). Then M ⊆ N as L-structures iff the inclusion map is an
L-morphism.
• Let M, N be L-structures. Then there is an L-embedding M → N iff N
can be turned into a model of Th0 (M, M ) (theory of M with no quanti-
fiers, parameters in M ).
One may rush to the definition of an isomorphism.
Definition 1’.2.6 (L-isomorphism).
• A L-morphism is an L-isomorphism if it is surjective.
• M and N are L-isomorphic, written M ' N (L being implicit) If there
is an L-isomorphism between them.
(Recall that an L-morphism is always injective!)

97
Remark 1’.2.7. The composition of two L-isomorphisms, the reciprocal map-
ping of an L-isomorphism, are L-isomorphisms.
Example 1’.2.8. (Z, <) and (2Z, <) are clearly isomorphic.
Remark 1’.2.5 however suggested that we did not take the good definition in
the first place: morphisms reduce to inclusions, which in general have but little
interest. So we look for something more reminiscent of the notion of elementary
inclusion.

1’.2.2 Elementary Morphisms

Definition 1’.2.9 (elementary morphism). An L-morphism σ : M → N is
elementary if for all formulas ϕ(m) with parameters in M :

M |= ϕ(m) iff N |= ϕ(σ(m))

Stronger than ordinary morphisms which only preserve interpretation, one

could say that elementary morphisms preserve satisfaction.
Remark 1’.2.10. The composition of elementary morphisms is elementary;
IdM is always elementary.
The following must be compared with Remark 1’.2.5.
Remark 1’.2.11.
• If M ⊆ N , then the inclusion map M → N is elementary.
• Let M ⊆ N be two L-structures. Then M  N iff the inclusion map is
elementary.
• More generally, there is an elementary embedding M → N iff there is
M0  N such that M ' M0 iff N can be turned into a model of
Th(M, M ). Exercise 1’.8

As for L-isomorphisms, we had already reached a very nice notion.

Proposition 1’.2.12. An L-isomorphism is elementary.
Proof . Let σ : M → N be a bijective L-morphism. We show that σ is ele-
mentary. By induction on a formula ϕ, we show that for any tuple m ∈ M ,
M |= ϕ(m) iff N |= ϕ(σ(m)).
If ϕ is atomic, this is clear by definition of a morphism. The case of connec-
tives is not hard to deal with.
Suppose ϕ = ∃x ψ. If M |= ϕ(m), then there is µ ∈ M such that M |=
ψ(µ, m). By induction, N |= ψ(σ(µ), σ(m)); therefore N |= ϕ(σ(m)). Con-
versely, if N |= ϕ(σ(m)), then there is ν ∈ N such that N |= ψ(ν, σ(m)). Since
σ is surjective, there is µ ∈ M such that σ(µ) = ν. Hence N |= ψ(σ(µ), σ(m));
by induction, M |= ψ(µ, m), so M |= ϕ(m).

98
Corollary 1’.2.13. If M ' N , then M ≡ N .
Proof . Th(M) is the set of properties of the empty tuple ∅ in M. As there is
an isomorphism σ, one has that for any sentence, M |= ϕ iff N |= ϕ.

1’.2.3 Categoricity
Definition 1’.2.14 (categoricity). Let T be a first-order theory with infinite
models and κ an infinite cardinal. T is κ-categorical if it has a model of cardi-
nality κ and any two models of T of cardinal κ are isomorphic.

Caution! This should be opposed to “absolute categoricity” of second-order

Peano arithmetic (Theorem 2.2.2), which is about any two models of PA. First-
order theories cannot capture cardinalities as second-order logic does (this will
become clear with the Löwenheim-Skolem theorems, in §1’.3); absolute cate-
goricity is impossible for a first-order theory with infinite models. One therefore
compares only models of the same cardinality.
Example 1’.2.15. Here are some categorical theories (we do not prove these
facts):
• The theory of dense linear orderings is ℵ0 -categorical (Theorem 1’.4.4 be-
low), but not κ-categorical for any κ ≥ ℵ1 (harder).
• The theory of vector spaces over a finite field is categorical in any cardinal.
• The theory of vector spaces over Q is not ℵ0 -categorical, but it is κ-
categorical for any κ ≥ ℵ1 (generalize this statement to bigger base fields!).

• The theory of algebraically closed fields of a given characteristic is not

ℵ0 -categorical, but it is κ-categorical for any κ ≥ ℵ1 (hint: transcendence
bases).
Categoricity will provide a criterion for completeness of a theory (Corollary
1’.3.14). But let us briefly leave the syllabus.

Theorem 1’.2.16 (Morley’s categoricity theorem). A theory categorical in some

κ ≥ L is categorical in any κ ≥ L.
Micahel Morley actually proved the countable language case in 1965; Saharon
Shelah extended it to any language in 1974.

End of Lecture 21.

Lecture 22 (Löwenheim-Skolem Theorems)

99
1’.3 Löwenheim-Skolem Theorems
1’.3.1 The Substructure Generated by a Set
Notation 1’.3.1. If L is finite, we write Card L = ℵ0 anyway.

Notice that there is no risk of confusion if Card L ≥ ℵ1 . You may think that
in any case, Card L actually denotes Card WFFL .
Let us start with an L-structure M and a subset A of M . One can easily
find the smallest substructure of M containing A: it suffices to close A (union
the constants) under the functions of the language; the resulting substructure
has cardinality at most Card A + Card L.

Example 1’.3.2. Let G be a group and S. Then there is a smallest subgroup

of G containing S; it is precisely the closure of S under the group multiplication
and inversion. If S is countable, then so is hSi.
Lemma 1’.3.3. Let M be a first-order structure, A ⊆ M . Then there is a
smallest L-structure hAi of M containing A. Moreover, CardhAi ≤ Card A +
Card L.
Proof . We first deal with the constants by considering A0 = A ∪ {cM : c ∈
L a constant symbol}. `
We now write the set of function symbols of the language F = n∈N Fn ,
where Fn is the set of function symbols of arity exactly n. Let
[ [
A1 = A0 ∪ f M (An0 )
n∈N f ∈Fn

(The idea is clear. One wishes to add the images of elements of A0 through
functions of L. Unfortunately f (A0 ) does not necessarily make sense, as the
arity of the function may vary. The union indexed over N deals with this minor,
strictly notational, issue.)
From A1 we build A2 in a similar fashion, and keep going. Eventually the
set hAi = ∪n∈N contains A, all (interpretations of) constants, is closed under
(the interpretations of) all functions in the language. We interpret relation
symbols on hAi by the induced interpretation (i.e., as a subset of M ), getting
an L-structure which is a substructure of M.
Notice that A0 has cardinality at most Card A + Card L; at each stage, we
add at most Card L elements; there are ℵ0 ≤ Card L stages, and this proves
CardhAi ≤ Card A + Card L.
Example 1’.3.4. Consider the case of a subset S of a group G again. Then
S0 = S ∪ {1}, S1 = S0 ∪ S0−1 ∪ S0 · S0 , etc. One does find hSi = ∪n∈N Sn .
Definition 1’.3.5 (substructure generated by a set). Let M be an L-structure
and A ⊆ M be any set. The substructure generated by A in M is the smallest
substructure of M containing A. It coincides with hAi from Lemma 1’.3.3.

100
 One can prove that an arbitrary intersection of substructures of M is a
substructure of M again; we have just been generalizing to L a construction
which is well-known for groups. It follows that hAi is the intersection of all
substructures of M containing A. Notice in any case that hAi is canonical, and
that we could construct it explicitely.
All this is very nice, but in general hAi has no reason to be an elementary
substructure of M (Definition 1’.1.13), and this is precisely what we would be
interested in. So Lemma 1’.3.3 and Definition 1’.3.5 are not that useful.

1’.3.2 Skolem Functions and the Descending Version

If we want to find an elementary substructure containing A, then by Tarski’s
Test (Theorem 1’.1.17) we ought to add existential witnesses for all formulas.
Recall how existential witnesses played an essential role in the proof of the
completeness of first-order logic. We now deal with them in a more semantic
fashion, adding to the language functions which choose them (this construction
is of course not canonical, and heavily relies on the axiom of choice).
Caution! The substructure we shall construct in Theorem 1’.3.9 will not be
canonical.
Definition 1’.3.6 (Skolem functions). Let L be a first-order language. For any
formula ϕ(x, y), where y is an n-uple of variables, we add a new function symbol
fϕ , the Skolem function associated to ϕ.
We let LSk = L ∪ {fϕ : ϕ(x, y) an L-formula}.
Definition 1’.3.7 (Skolemization). Let T be an L-theory. The Skolemization
of T is the LSk -theory TSk which is the union of T and the axioms


∀y ∃x ϕ(x, y) → ϕ(fϕ (y), y)

The meaning is clear: our new axioms say that witnesses are explicitely given
by our new function symbols.
Lemma 1’.3.8. Let M be a model of T . Then there is an interpretation of the
Skolem functions making M a model of TSk .
Proof . We must give a meaning to the Skolem function symbols fϕ . Let ϕ(x, y)
be an L-formula, and m ∈ M have same length as y. If M |= ∃x ϕ(x, m), then
there is µ ∈ M such that M |= ϕ(µ, m); we choose such a µ and set fϕM (m) = µ.
If there is no suitable µ, we map m to any element in M .
With heavy use of the axiom of choice (we perform infinitely many choices),
we thus define functions fϕM , which are the interpretations of the Skolem func-


tion symbols. By construction, M |= ∀y ∃x ϕ(x, y) → ϕ(fϕ (y), y) .
Let us return to our problem. If we now close a subset A ⊆ M under all
functions including the Skolem functions, we are in good position to find an
elementary inclusion, as all necessary witnesses have been plugged in.

101
Theorem 1’.3.9 (descending Löwenheim-Skolem theorem). Let M be an infi-
nite L-structure and A ⊆ M . Let κ be a cardinal with Card A + Card L ≤ κ ≤
Card M . Then there is M0  M such that A ⊆ M0 and Card M0 = κ.
Proof . We may assume that Card A = κ. We turn M into an LSk -structure
which satisfies TSk . In the language LSk , we now consider the substructure of
M generated by A (Lemma 1’.3.3). Forgetting the interpretation of the Skolem
functions, this is a fortiori an L-structure M0 . Moreover Card M0 = κ.
It remains to prove that the inclusion is elementary. We use Tarski’s Test
(Theorem 1’.1.17). Let ϕ(m0 ) be a formula with parameters in M0 ; assume
M |= ∃x ϕ(x, m0 ). As M |= TSk , one has M |= ϕ(fϕM (m0 ), m0 ). But by
construction, the element fϕM (m0 ) lies in M0 . This is what we need in order to
apply Tarski’s criterion.

Remark 1’.3.10.
• M0  M as L-structures, not necessarily as LSk -structures. This is
because when we expand the language to LSk , we create new formulas. If
we did want to ensure M0  M as LSk -structures, we would need to add
new Skolem functions for the new formulas, getting LSk,Sk , etc.
However an inclusion of L-structures is what we want; the expanded lan-
guage LSk is a technical device we can forget about.
• The construction is highly non-canonical (it relies on the axiom of choice).
In general, there is no smallest elementary substructure containing A:
everything depends on the way we interpreted the Skolem functions, which
is arbitrary. One should not use the notation hAi for “Skolem hulls”.
As a final interesting word, we prove inconsistency of set theory.
Corollary 1’.3.11 (Skolem’s paradox). (If set theory is satisfiable) there is a
countable model of set theory.

Proof . Write down ZFC as a first-order theory in the language {∈}. Start with
a model and apply Theorem 1’.3.9.
However in this model U0 of set theory we may construct N and P (N),
which is notoriously uncountable, and yet a subset of the countable underlying
universe. . .
[Catch: “countable” does not bear the same meaning in the first model, say
U, and the countable one U0 . All that we have proved is that U0 is so small that
it does not “see” (i.e., perceive as a set) a bijection between P (N) and N.]

1’.3.3 The General Version and the Łoś-Vaught Criterion

We now try to “go up”; that is, provide very large elementary extensions of
existing models.

102
Theorem 1’.3.12 (ascending Löwenheim-Skolem theorem). Let M be an infi-
nite L-structure and κ ≥ Card M be a cardinal. Then there is M∗  M such
that Card M ∗ ≥ κ.
Proof . Let C 0 = {ci : i < κ} be new constants, and consider the language
L0 = LM ∪ C 0 . Form the L0 -theory T 0 = Th(M, M ) ∪ {ci 6= cj : i 6= j}. This
theory is consistent: a finite fragment mentions a finite number of properties
ϕ(m), which are trivially satisfiable in M, and a finite number of formulas
ci 6= cj , meaning that there are at least n distinct elements: this is satisfiable
in M again.
Hence any finite fragment of T 0 is satisfiable in M; by compactness, there is a
a model M∗ of T 0 (notice that M∗ is not likely to be M). As M∗ |= Th(M, M ),
one has M  M∗ ; on the other hand, the constants of C 0 say that M∗ has at
least κ elements.
Combining Theorems 1’.3.9 and 1’.3.12, one derives a sharper version.
Theorem 1’.3.13 (Löwenheim-Skolem theorem). Let M be an L-structure and
A ⊆ M . Let κ be a cardinal with Card A + Card L ≤ κ. Then there is M0 such
that A ⊆ M 0 , Card M 0 = κ, and M0  M or M0  M.
Proof . Go up (Theorem 1’.3.12), then down (Theorem 1’.3.9).
Corollary 1’.3.14 (Łoś-Vaught test). Let T be a first-order theory in a language
L. If there is κ ≥ Card L such that T is κ-categorical, then T is complete.
Proof . Let M, N |= T . We aim at showing M ≡ N .
By the Löwenheim-Skolem theorem (Theorem 1’.3.13), there are M0 and N 0
of cardinal κ such that M  M0 or M  M0 , and similarly for N and N 0 . But
in any case M ≡ M0 and N ≡ N 0 . In particular, M0 , N 0 |= T .
As M0 and N 0 are models of T of cardinal κ, they are isomorphic by cate-
goricity; it follows M0 ≡ N 0 . As a conclusion, M ≡ M0 ≡ N 0 ≡ N .
Example 1’.3.15. We shall prove later (Theorem 1’.4.4) that DLO is ℵ0 -
categorical. In particular it will be complete!

End of Lecture 22.

Lecture 23 (Back-and-Forth Methods)

1’.4 Back-and-Forth Methods

It is not very common in practice to have an isomorphism; a very good rea-
son is that two structures may have the same theory without having the same
cardinality! So one should aim at weaker correspondences.
And precisely, we shall work with very partial functions, functions which are
defined only on a finite set. But these turn out to be enough (as our formulas
are finite too!) Model theory has a wide variety of so-called back-and-forth

103
methods which enable to inductively extend partial functions. One should bear
in mind the image of a bootstrap.
Disclaimer: back-and-forth methods are easier to work out in the case of
relational languages. With unary functions they are still manageable; if there
is a binary function around, it is not very likely that one will understand any
kind of back-and-forth.

1’.4.1 Dense Linear Orderings

We start with a key example.
Definition 1’.4.1 (DLO). A dense linear ordering (“with no endpoints” be-
ing always implicit) is a model of the following theory in the language {<} of
orderings:
• ∀x ¬(x < x)
• ∀x∀y (x < y ∨ x = y ∨ y < x)
• ∀x∀y∀z (x < y ∧ y < z → x < z)
• ∀x∀y∃z x < y → (x < z ∧ z < x)
• ∀x∃y∃z y < x ∧ x < z
DLO denotes either this theory, or a model of the theory (this is harmless as
the theory will turn out to be complete).
Remark 1’.4.2. The first three axioms express that < is a linear ordering; the
fourth (density) says that between any two points there is another point; the
fifth (no endpoints) states that there is no least, nor largest element). In spite
of the name, one should never forget the last one!
`
Example 1’.4.3. Recall that given two orderings A and B, A < B denotes
the extension of the orderings such that any element of A lies below any element
of B (“B is pasted after A”).
• (Q, <) is a countable DLO.
` ` `
• (Q < Q, <) and (Q < {∞} < Q, <) are other countable DLO’s.
• (R, <) is a DLO of cardinality continuum.
`
A close inspection will reveal that Q < Q and Q are actually isomorphic.
Back-and-forth methods start here.
Theorem 1’.4.4 (Cantor). Any two countable DLO’s are isomorphic.
Proof . Let M, N be countable DLO’s. We shall inductively construct an iso-
morphism. To this extent, we enumerate the underlying sets M = {mk : k ∈ N}
and N = {nk : k ∈ N}. We want to construct an isomorphism; it suffices to
construct an order-preserving function σ defined on all of M , and which is onto.
This will be done inductively by considering two kinds of steps:

104
 • at odd stages, we shall make sure that σ is defined everywhere on M ;
• at even stages, we shall make sure that σ reaches every element of N .
We thus want to find an increasing sequence of partial maps which preserve
the ordering. Let us start by defining σ0 (m0 ) = n0 . So far, so good, as all we
can say is m0 = m0 , which certainly holds of n0 .
• (stage 2k+1) Suppose that σ2k is a partial map with finite support, and
such that a = dom σ2k ∈ M and b = im σ2k ∈ N satisfy the same relations.
(You may think that σ2k “preserves the configuration of the tuple”, that
it is a partial L-morphism in the sense of Definition 1’.2.2.) We want to
extend σ2k to an order-preserving map σ2k+1 with mk+1 ∈ dom σ2k+1 .
mk+1 is in a certain configuration with respect to the finite tuple a. If
mk+1 lies above any element of a, then as there are no endpoints in N ,
there is certainly β ∈ N which lies above any element of b. If on the other
hand, mk+1 lies between two elements of the tuple a, then as N is dense,
the same pattern can be reproduced in N . In any case there is β ∈ N such
that the extended tuples a, mk+1 and b, β still satisfy the same relations.
We let σ2k+1 extend σ2k by σ2k+1 (mk+1 ) = β.
• (stage 2k + 2) Suppose that σ2k+1 has been constructed as a finite partial
map which preserves relation between a0 = dom σ2k+1 and b0 = im σ2k+1 .
We aim at extending σ2k+1 in such a way that nk+1 will be in the image.
And we argue exactly like above, getting in any case an element α ∈ M
such that α behaves with respect to a0 exactly the way nk+1 does for b0 .
We let σ2k+2 (α) = nk+1 .
At the limit, we let σ = ∪k∈N σk . This is still an order-preserving bijection.
Because of our construction at odd stages, every element of M will appear in the
domain, so dom σ = M . Because of our construction at even stages, im σ = N .
It follows that σ : M → N is an isomorphism.
Bearing in mind Definition 1’.2.14, one sees that Theorem 1’.4.4 says that
DLO is ℵ0 -categorical. As a consequence, one can deduce from Corollary 1’.3.14
that DLO is complete.

Remark 1’.4.5. This superb proof relies on two essential ideas.

• If you can always extend a partial function, then at the limit something
nice will happen. Proposition 1’.4.14 below will generalize this idea.
It is however difficult to understand at first in what measure exactly count-
ability is needed. Many constructions in model theory rely on induction
beyond ω. But here we know how to deal only with finite configurations,
and this is why the limiting process can’t go further than the first infi-
nite. For instance, there are two non-isomorphic DLO’s of cardinality ℵ1
(Counter-example 1’.4.15 below).

105
 • Above all, the proof relies on this bootstrap construction, arguing that
any finite configuration on one hand can be reflected on the other hand.
This is worth a general definition (Definition 1’.4.10 hereafter).

1’.4.2 ∞-isomorphisms
Definition 1’.4.6 (local isomorphism, 0-isomorphism).
• A local isomorphism, or 0-isomorphism, σ : M → N is an injective partial
function M → N with finite domain which preserves the constants, func-
tions, and relations. Hence σ is a local isomorphism iff dom σ and im σ
satisfy the same atomic formulas (in M, N respectively).
• Two tuples a ∈ M and b ∈ N are locally, or 0-isomorphic, written a '0 b,
if there is a local isomorphism σ mapping a to b.
• Two structures M and N are locally, or 0-isomorphic, written M '0 N ,
if the empty function ∅ 7→ ∅ is a 0-isomorphism.
“Local” is of course, opposed to “global”: we consider only functions with
finite support. So a local isomorphism is a small fragment of a morphism. When
one wants to insist on the poor quality of σ, one calls it a 0-isomorphism: it
preserves only formulas of quantification rank 0. This is very far from being
elementary!
Remark 1’.4.7. A beginner’s mistake is to think that the empty tuple of M
is always 0-isomorphic to the empty tuple from N . This is not the case, as
there might be constants in the language. For instance, the ring Z/2Z satifies
1 + 1 = 0, but the ring Z/3Z doesn’t. So even the empty function is not
compatible with the interpretation!
A little thinking will reveal that M '0 N iff Th0 (M) = Th0 (N ) (quantifier-
free theories, no parameters: this boils down to atomic formulas using terms
built from constants).
Example 1’.4.8.
`
• Consider the orderings (Z, <) and`(Z < Z, <); we denote by 0 the 0 of
Z; there are two copies of 0 in Z < Z, which we denote respectively by
01 and 02 . Clearly ∅ '0 ∅. Clearly too, 0 '0 01 . Also, (0, 2) '0 (01 , 02 )
(as 0 < 2 and 01 < 02 ).
`
• Consider the successor function and the structures (Z, s) and (Z Z, s).
Clearly ∅ '0 ∅ and 0 '0 01 . But it is no longer the case that (0, 2) '0
(01 , 02 ). Indeed, M |= 2 = s(s(0)), whereas N |= 02 6= s(s(01 )).
Example 1’.4.9. A moment’s thought and one realizes that in the language of
orderings, local isomorphisms are exactly order-preserving bijections with finite
support.
We can now give the key definition. Bear in mind the image of a bootstrap.

106
Definition 1’.4.10 (∞-isomorphism, Karp family). Let M and N be two L-
structures. M and N are ∞-isomorphic, written M '∞ N , if there is a non-
empty family F of local isomorphisms M → N such that:
forth: for all σ ∈ F and α ∈ M , there is τ ∈ F extending σ such that α ∈ dom τ
back: for all σ ∈ F and β ∈ M , there is τ ∈ F extending σ such that β ∈ im τ
F is called a Karp family, and its elements are called ∞-isomorphisms.
Remark 1’.4.11.
• If there is a Karp family, there is a largest one (take the union of all).
• M ' N implies M '∞ N (take all restrictions of a global isomorphism),
but the converse need not hold in general: perhaps M and N don’t even
have the same cardinality (Counter-example 1’.4.13 below).
Proposition 1’.4.12. Any two DLO’s are ∞-isomorphic. More specifically, the
family of local isomorphisms is a Karp family.
Proof . Local isomorphisms between orderings are just order-preserving, finite
support bijections (which do exist: the family is not empty!). The back-and-
forth property is the core of the proof of Theorem 1’.4.4.
We emphasize that ∞-isomorphic is much weaker than isomorphic.
Counter-example 1’.4.13. Consider the orderings (Q, <) and (R, <). They
are ∞-isomorphic, but of course Q 6' R.
There is however a partial correction.
Proposition 1’.4.14. If M '∞ N and both are countable, then M ' N .
Proof . This is an abstract version of the bootstrap construction used in The-
orem 1’.4.4.
Propositions 1’.4.12 and 1’.4.14 together yield the proof of Theorem 1’.4.4.
Counter-example 1’.4.15. The notions of ∞-isomorphism and of a Karp
family (Definition 1’.4.10) are about arbitrarily large finite configurations; one
should therefore not expect to generalize Proposition 1’.4.14 past ℵ0 , even if M
and N have the same cardinality. Here
` is a counter-example.
Consider
` an ordered sum M = ℵ1 Q (copies of Q copied after each other),
and N = ℵ1 (Q ∪ {∞}) (Q followed by a point). Both are dense linear order-
ings without endpoints, of cardinality ℵ1 . They are ∞-isomorphic (Proposition
1’.4.12), and yet they are not isomorphic. The latter is non-trivial.
Back-and forth arguments are especially powerful in ω-saturated models
(defined below, Definition 1’.5.12). For instance, it is not true that any two
algebraically closed fields are ∞-isomorphic, but any two ω-saturated models
(defined below, §1’.5.2) are. These are algebraically closed fields of infinite tran-
scendence degree; you may already try to establish a back-and-forth between
them. This will be done in §1’.5.3.

107
1’.4.3 Finitary Back-and-Forth*
We build on local isomorphisms to find weaker versions of back-and-forth.
Definition 1’.4.16 (n + 1-isomorphism).
• A local isomorphism σ : a → b is an (n + 1)-isomorphism if:

for all α there is β and an n-isomorphism τ : aα 'n bβ extending σ

for all β there is α and an n-isomorphism τ : aα 'n bβ extending σ
• a and b are (n + 1)-isomorphic, written a 'n+1 b, if there exists an (n + 1)-
isomorphism between them.

• Two structures M and N are (n + 1)-isomorphic, written M 'n+1 N , if

the empty function ∅ 7→ ∅ is an (n + 1)-isomorphism.
Remark 1’.4.17. Notice that this is an inductive definition, as opposed to
the notion of ∞-isomorphism (Definition 1’.4.10), which is defined for a class,
simultaneously.

Here begin the funny games: given two structures, how isomorphic are they?
Remark 1’.4.18. Back-and-forth may be viewed as a game opposing two play-
ers, having two different boards M and N . A round consists of the following
two steps:

• player 1 puts a pawn on either board;

• player 2 puts a pawn on the other board, in order to keep both configura-
tions similar.
M and N are n-isomorphic if (a clever) player 2 is sure to last at least n rounds.

Example 1’.4.19. Consider Example 1’.4.8 again.

`
• Consider M = (Z, s) and N = (Z Z, s). They are 0-isomorphic (as the
language has no constants!). They are 1-isomorphic, as picking an element
in one structure can certainly be reflected on the other one.
They are however not 2-isomorphic. Consider the element 01 ∈ N ; it will
be reflected in M by some element which we may assume to be 0; 0 '0 01 .
But now player 1 chooses 02 , which cannot be reflected in M, as there is
only one orbit of s in M, and 01 and 02 lie in distinct orbits of N .
Interestingly, these structures turn out to be elementary equivalent (this
might be a little unclear yet but §1’.5.2 and §1’.5.3 will shed some light).
` `
• Consider (Z, <) and (Z < Z, <). Let k ∈ N; I say that Z and Z < Z are
k-isomorphic. For clarity, let us call Z the Z which comes alones; and let
Z1 and Z2 be the left and right Z’s of the pasted structure.

108
 Player 2 will identify Z with Z1 : whenever player 1 makes a move in Z,
player 2 does the same in Z1 ; whenever player 1 makes a move in Z1 ,
player 2 does the same in Z. Now, when player 1 plays in Z2 , player 1
picks a similar configuration in very big elements of Z, of size around 2k .
Player 1 will win, but it will take him at least k rounds!

There is a name for the second phenomenon of Example 1’.4.19.

Definition 1’.4.20 (ω-isomorphism).
• σ : a → b is an ω-isomorphism if for all n it is an n-isomorphism.
• a et b are ω-isomorphic if there is an ω-isomorphism between them.

• Two structures are ω-isomorphic if ∅ is an ω-isomorphism.

Remark 1’.4.21.
• A beginner’s mistake: ω denotes the first infinite ordinal, which is very
small for set theory. In particular, ω-isomorphism is way weaker than
∞-isomorphism;
` here is why. In Example 1’.4.19 we have seen that (Z, <)
and (Z < Z, <) are ω-isomorphic. However they are not ∞-isomorphic,
as otherwise, being countable, they would be isomorphic by Proposition
1’.4.14, which is clearly not the case.
• ω-isomorphisms are a little subtle. They mean that player 2 always has a
strategy that will keep him playing for at least k rounds, but the choice of
the strategy might depend on k (typically Example 1’.4.19). If there were
a uniform strategy, which does not depend on how long player 2 tries to
keep playing, one would have ∞-isomorphism (Definition 1’.4.10 above).
ω-isomorphisms are especially interesting because of our notion of a lan-
guage. Going down through the recursive definition of satisfaction, one sees
that our formulas are always about finite tuples, of arbitrary length.
Proposition 1’.4.22. If M 'ω N , then M ≡ N .
In order to prove this proposition, let us introduce useful terminology.

Definition 1’.4.23 (type of a tuple). Let M be an L-structure and a ∈ M

be a tuple (abusing, as usual, Cartesian powers). The type of a in M is
tpM (a) = {ϕ(x) : M |= ϕ(a)}. When there is no risk of confusion, one drops
the superscript M .
Bearing in mind the definition of the quantification rank of a formula (Defini-
tion 1’.0.8), one can even define tpM
n (a) = {ϕ(x) : qrk(ϕ) ≤ n and M |= ϕ(a)}.

Lemma 1’.4.24. Let M, N be L-structures and a, b be tuples extracted from

M, N respectively. If a 'n b, then tpn (a) = tpn (b).

109
Proof . By induction. First notice that by definition (Definition 1’.4.6), if a '0
b, then a and b satisfy the same relations, hence the same atomic formulas; the
claim is proved.
Suppose that the property is known for n; we show it for n + 1. So let
a 'n+1 b be two n + 1-isomorphic tuples. We show tpn+1 (a) = tpn+1 (b). Let ϕ
be a formula of tpn+1 (a). We show that N |= ϕ(b).
A quick induction on the structure of ϕ reveals that the only interesting case
is that of ϕ(a) being ∃x ψ(a, x), for a formula ψ now of quantification rank ≤ n.
We have assumed M |= ϕ(a): so there exists α ∈ M such that M |= ψ(a, α).
In particular, ψ ∈ tpn (a, α). As a and b are n + 1-isomorphic, there is by
definition (Definition 1’.4.16) β ∈ N such that a, α 'n b, β. By induction,
ψ ∈ tpn (b, β); that is, N |= ψ(b, β). It follows that N |= ϕ(b): we are done.

Remark 1’.4.25. We insist that if a and b are locally isomorphic, and that
there is a Karp family of extensions of a '0 b (one may then write a '∞ b), a
and b satisfy the same formulas with quantifiers.
Proof of Proposition 1’.4.22. Suppose M 'ω N ; we show M ≡ N . Our
assumption means that for each n, ∅M 'n ∅N . By Lemma 1’.4.24, tpM n (∅) =
tpN
n (∅). But tp M
n (∅) = Th n (M) (the theory of M restricted to formulas of
qrk ≤ n). As this is true for any n, one gets Th(M) = Th(N ); so M ≡ N .
Proposition 1’.4.22 does not have a converse, because when n > 0, Lemma
1’.4.24 doesn’t.
`
Counter-example 1’.4.26. Consider the case of (Z, s) and (Z Z, s) again.
Then 0 and 01 clearly satisfy the same formulas of qrk ≤ 1, but as observed,
they are not 1-isomorphic.
Remark 1’.4.27. The choice of Counter-example 1’.4.26 is typical. Actually, in
a language consisting of finitely many relations, the converse to Lemma 1’.4.24
does hold, and in particular M 'ω N iff M ≡ N !
We sum up the isomorphism strength as follows (n < m)

' ⇒ '∞ ⇒ 'ω ⇒ 'm ⇒ 'n

⇓
≡

Counter-example 1’.4.28.
• (R, <) '∞ (Q, <) but they are not isomorphic (Counter-example 1’.4.13)
`
• (Z, <) 'ω (Z < Z, <) but they are not ∞-isomorphic (Remark 1’.4.21).
`
• (Z, s) ≡ (Z Z, s) but they are not ω-isomorphic (Example 1’.4.19).
`
• (Z, s) '1 (Z Z, s) but they are not 2-isomorphic (Example 1’.4.19).

In special cases, converses do hold:

110
 • If M '∞ N are countable, then M ' N (Proposition 1’.4.14).
• There always is an ordinal such that M 'α N implies M '∞ N .
• In a finite, purely relational language, M ≡ N does imply N 'ω N .

End of Lecture 23.

Lecture 24 (Elimination Theorem)

1’.5 Quantifier Elimination and ω-Saturation

Sometimes, every formula (with free variables!) boils down to a formula with
no quantifiers. This was typically the case when in high school, you realized
that ∃x ax2 + bx + c = 0 iff b2 − 4ac ≥ 0 (over Th(R) as an ordered ring). We
give a name to this general phenomenon.
Definition 1’.5.1 (quantifier elimination). A first-order theory T eliminates
quantifiers (“has QE”) if for any formula ϕ(x) there is a quantifier-free formula
ψ(x) such that T |= ∀x ϕ(x) ↔ ψ(x).
(One says that ϕ(x) and ψ(x) are equivalent modulo T ; this amounts to
having, in L ∪ {x}, T ∪ {ϕ(x)} |= ψ(x) and T ∪ {ψ(x)} |= ϕ(x).)
Example 1’.5.2.
• In R as an ordered ring, the formula ∃x ax2 + bx + c = 0 is equivalent to
(b2 − 4ac ≥ 0) ∨ (a = 0 ∧ (b 6= 0 ∨ c = 0)). Actually the theory RCF of R
as an ordered ring eliminates quantifiers.
• In C as a ring, the formula ∃x an xn + · · · + a1 x + a0 is equivalent to
¬(∧ni=1 ai = 0 ∧ a0 6= 0) (this is the D’Alembert-Gauß theorem). Actually
the theory ACF0 of C as a ring eliminates quantifiers.
We shall give a semantic criterion for a theory to admit QE (Corollary 1’.5.6
below). As it can be stated in a more general form, we extend the definition to
arbitrary elimination.

1’.5.1 Elimination Sets

Definition 1’.5.3 (elimination set). Let T be a theory. Let E be a set of
formulas. E is an elimination set modulo T if every formula ϕ of L is equivalent
modulo T to some Boolean combination of formulas from E.
Theorem 1’.5.4 (elimination theorem). Let T be a theory and E a set of for-
mulas. Then E is an elimination set if and only if, for all models M and N of
T and all tuples m and n extracted from M and N respectively,
if m and n satisfy the same formulas from E, then they satisfy the
same formulas.

111
Proof . One implication is trivial. For the other one, we may clearly assume
that E is closed under Boolean operations. (In particular, E contains a con-
tradiction.) Let ϕ(x) be an L-formula. We shall prove that ϕ(x) is equivalent
(modulo T ) to a formula of E.
Step 1. There is a finite subset Ψ ⊆ E such that the theory T ∪ {ψ(c) ↔ ψ(d) :
ψ ∈ Ψ} ∪ {ϕ(c) ∧ ¬ϕ(d)} is inconsistent.
Verification: Let c, d be tuples of new constant symbols (of the same length as
x) and L0 = L ∪ {c, d}. Let

T 0 = T ∪ {ϕ(c) ∧ ¬ϕ(d)} ∪ {ψ(c) ↔ ψ(d) : ψ ∈ E}

The L0 -theory T is not consistent; supppose it is. Then there is a model M |= T 0

with interpretations m = cM and n = dM . Notice that m and n satisfy the
same formulas from E, but disagree on ϕ: against the assumption. So T 0 is not
consistent. By compactness, there is a finite, inconsistent fragment. ♦
Step 2. For every M |= T , there is a formula ψM (x) ∈ E equivalent to ϕ(x).
Verification: Let M |= T , and X = {m ∈ M : M |= ϕ(m)} (the subset of
M n defined by ϕ). If X is empty, then bearing in mind that E contains a
contradiction, we are done. So we may suppose that X 6= ∅.
Foreach m ∈ X, let Ψm = {ψ ∈ Ψ : M |= ψ(m)} and Θm = Ψ \ Ψm . Let
V
ψm = ∧Ψm ψ ∧Θm ¬ψ (hence ψm describes exactly which formulas of Ψ
a tuple m ∈ X satisfies).
Notice that each ψm is in E, and that only finitely many ψm ’s are possible
by finiteness of Ψ (even though X might be infinite). Let ψM = ∨X ψm : this
should be another definition of X! Notice that ψM ∈ E. We now claim that
M |= ∀x (ϕ(x) ↔ ψM (x)).
Let n ∈ M . If M |= ϕ(n), then by definition n ∈ X; hence M |= ψn (n)
and M |= ψM (n). On the other hand, assume M |= ψM (n). By construction
of ψM , there is m ∈ X such that M |= ψm (n); by construction of ψm , one has
M |= ψm (m). It follows that m and n satisfy exactly the same formulas from
Ψ. By Step 1, m and n must agree on ϕ too. As m ∈ X, one has M |= ϕ(m);
it follows M |= ϕ(n). ♦
The issue is that, so far, the formula ψM is equivalent to ϕ only in M (if the
theory T were complete we would however be done). We solve this problem.
Step 3. Let M |= T . There is a sentence θM ∈ E true in M and such that
T ∪ {θM } |= ∀x (ϕ(x) ↔ ψM (x)).
Verification: Let ΘM = {χ ∈ E : M |= χ} (this is E ∩ Th(M)) and

T 0 = {χ : χ ∈ ΘM } ∪ {¬χ : χ 6∈ ΘM } ∪ {¬∀x (ϕ(x) ↔ ψM (x))}

If T 0 were consistent, we would have a model N of T 0 . The empty tuple ∅ ∈

N satisfies exactly the same formulas from E as ∅ ∈ M ; in particular, by

112
assumption, ∅ satisfies the same formulas in M and N , that is M ≡ N . Yet
M |= ∀x (ϕ(x) ↔ ψM (x)), a contradiction. So T 0 is not consistent.
By compactness, there is a finite fragment of ΘM , which we can take to be
a single sentence θM , such that T ∪ {θM } |= ∀x (ϕ(x) ↔ ψM (x)). ♦
Step 4. Finitely many θM ’s suffice for all models of T .
Verification: Let T 0 = T ∪ {¬θM : M |= T } (this makes sense: the collection
of models is a proper class, but the collection of θM ’s is a set alright, as a
subcollection of the set of L-sentences). If T 0 were consistent, we would have a
model N |= T in which none of the θM ’s holds, which is impossible!
By compactness, T 0 is finitely inconsistent, and this means that there is a
finite set Θ of θM ’s such that T |= ∨Θ θM . ♦
Step 5. There is ψ(x) ∈ E such that T |= ∀x ϕ(x) ↔ ψ(x).

Verification: Let Θ = {θM : M |= T }; we may assume that Θ is finite. Consider

_
ψ(x) = (θM ∧ ψM (x))
Θ

Notice that ψ ∈ E. Now let M |= T , and m ∈ M : we show M |= ϕ(m) ↔ ψ(m).

Suppose M |= ϕ(m). By construction, M |= θM . In particular, M |=
∀x (ϕ(x) ↔ ψM (x)), whence M |= ϕ(m) ↔ ψM (m), and M |= ψM (m),
whence M |= ψ(m).
Suppose now M |= ψ(m). Then there is a model N such that M |= θN ∧
ψN (m). As T ∪ {θN } |= ϕ(c) ↔ ψN (c), we see that M |= ϕ(m). ♦
This concludes the proof of Theorem 1’.5.4.

Remark 1’.5.5. This proof is a little long and clumsy. We have just been
covering a normal, compact space; a topological argument is possible, short,
and elegant, but it requires introducing spaces of types, which is a little beyond
the scope of a first introduction.
Exercise 1’.9

Corollary 1’.5.6. A theory T eliminates quantifiers if and only if, for all models
M and N of T and all tuples m and n extracted from M and N respectively,
if m and n satisfy the same atomic formulas, then they satisfy the
same formulas.

The question is now: how do we prove that two tuples do satisfy the same
formulas? In practice, this often results from a back-and-forth construction.
Corollary 1’.5.7. Let T be a first-order theory such that for any two models
M, N |= T , local isomorphisms between M and N form a Karp family. Then
T eliminates quantifiers.

113
Proof . We use Corollary 1’.5.6. Let m, n be drawn from M, N respectively
and satisfy the same quantifier-free formulas; we show that they satisfy the
same formulas. As m and n satisfy the same quantifier-free formulas, they are
locally isomorphic. By assumption, they are ∞-isomorphic. So m '∞ n, and
it follows tpM (m) = tpN (n) (see Remark 1’.4.25). Hence m and n satisfy the
same formulas; T eliminates quantifiers.

We shall later generalize Corollary 1’.5.7 to a weaker assumption (Theorem

1’.5.19 below).
Remark 1’.5.8. QE is strongly language-dependent; actually for any L-theory
T , there is a language L ⊆ L0 and an L0 -theory T ⊆ T 0 such that T 0 eliminates
quantifiers.
Verification: We first get rid of ∀’s, replacing them by ¬∃¬. For any formula
∃y ϕ(x, y) we introduce a new relation symbol Rϕ (x) and we add to the theory
∀x (∃y ϕ(x, y)) ↔ Rϕ (x).
Of course the resulting theory T1 in the resulting language L1 need not
eliminate quantifiers, as we have created new formulas. But we repeat the
construction. After ω steps, any formula of L0 = ∪n Ln is equivalent modulo
T 0 = ∪n Tn to a quantifier-free formula of L0 . ♦
One can actually prove compactness that way, reducing to compactness of
propositional logic.

End of Lecture 24.

Lecture 25 (ω-Saturated Models)

1’.5.2 ω-Saturated Models

Theorem 1’.5.4 suggests a method to prove quantifier elimination: just show that
two locally isomorphic tuples satisfy the same formulas. This can be done by
establishing ∞-back-and-forth (ω-back-and-forth is weaker and sufficient but
more subtle: hope for the strongest.) So let us go back to back-and-forth
constructions.
For instance, if one tries to show that Th(C) eliminates quantifiers, one
should start with algebraically closed fields K, L; one would take two locally
isomorphic tuples a '0 b, and try to check that we can always extend. This
fails if, for instance, K has an element which is transcendental over a, but L is
algebraic over b.
The conclusion of this algebraic discussion is that some models are “richer”
than others, and that presumably back-and-forth is easier to manage in rich
models. A formal notion of wealth is ω-saturation (Definition 1’.5.12 below).
This requires the following definition, which generalizes the notion of a system
of equations in n variables. To understand it, fix an integer n and a tuple of
variables x = (x1 , . . . , xn ) of length n.

114
Definition 1’.5.9 (type). Let L be a first-order language, M an L-structure,
A ⊆ M a set of parameters. An n-type over A is a collection p(x) of LA -formulas
ϕ(x, a) (a ∈ A) which is consistent with Th(M, A).
Treating the variables as constants, one sees that an n-type is merely a
theory in LA (x) extending Th(M, A).
This generalizes the case of the type of a tuple (Definition 1’.4.23), which
is called realized. But any type is the type of a tuple which “hasn’t appeared
yet”: every type is realized somewhere.
Remark 1’.5.10. If p(x) is a type over A ⊆ M , then there is N  M and
n ∈ N such that n realizes p. (Of course N will tend to be a proper extension
of M.)
Types are amazingly important and interesting (they come with nice topo-
logical methods), and students should refrain the teacher from going to deep
into this matter.
Example 1’.5.11.
• If A ⊆ M and b ∈ M , then tp(b/A) = {ϕ(x, a) : ϕ ∈ LA : M |= ϕ(b, a)} is
an n-type over A.
• Consider (N, <). Then {x > n : n ∈ N} is a 1-type over N. Notice that it
is not realized in N, but in some elementary extension.
• When we constructed the non-standard reals in §1.6.3, we simply con-
structed a 1-type over R, namely p(x) = {0 < x < r : r ∈ R>0 }, and then
went to an elementary extension R∗  R satisfying it.
• Consider C as a ring, and study maximal 1-types over Z. There are two
kinds:
– algebraic types, types which contain a formula P (x) = 0 for a poly-
nomial P with coefficients in Z.
– the transcendental type, which contains {P (x) 6= 0 : P ∈ Z[X]}.
• We build on the previous example to study (2-)types of pairs of complex
numbers p(x, y) over Z. There are five kinds:
– both x and y are algebraic over Z: the type boils down to {P1 (x) =
0, P2 (y) = 0} for some polynomials P1 and P2 .
– x is algebraic over Z, but y is transcendental: the type reduces to
{P1 (x) = 0} ∪ {P (y) 6= 0 : P ∈ Z[X]} for a certain polynomial P1 .
– x is transcendental, y is algebraic: p(x, y) = {P (x) 6= 0 : P ∈ Z[X]}∪
{P2 (y) = 0} for some P2 .
– x and y are transcendental but the family (x, y) is algebraically de-
pendent: the type contains {P (x) 6= 0 : P ∈ Z[X]} ∪ {P (y) 6= 0 : P ∈
Z[X]} ∪ {Q1 (x, y) = 0} for some Q1 ∈ Z[X, Y ].

115
 – The family (x, y) is algebraically independent: the type is p(x, y) =
{Q(x, y) 6= 0 : Q ∈ Z[X, Y ]}. This is the type of any pair of tran-
scendence degree 2.
One measures richness of a structure by the amount of types it realizes.
Definition 1’.5.12 (ω-saturated). An L-structure M is ω-saturated if any 1-
type over finitely many parameters is realized in M.

Example 1’.5.13.
• A finite structure is always ω-saturated, but this is a trivial example!
• Any DLO is ω-saturated.
• (N, s) is not ω-saturated. Let ϕn (x) be the formula x 6= sn (0). Then the
1-type p(x) = {ϕn (x) : n ∈ N} uses 0 as its only parameter, and is not
realized in N.
• (N, <) is not ω-saturated. Let ψn (x) be the formula ∃x1 . . . ∃xn x1 < · · · <
xn < x. Then the 1-type {ψn (x) : n ∈ N} uses no parameters and is not
realized in N.

• The ring Q is not ω-saturated. Indeed, any element of Z can be written as

a sum of 1’s. So the transcendental type p = {P (x) 6= 0 : P (X) ∈ Z[X]}
uses no parameters. It is however not realized in Q.
Theorem 1’.5.14. Let M be an infinite L-structure. Then there is an elemen-
tary extension M  M∗ which is ω-saturated.

Proof . This proof relies on ordinal induction, and may be omitted freely; the
result may not.
We enumerate all finite tuples a of elements of M , and all 1-types over a,
finding an enumeration (by some ordinal λ perhaps bigger than ω) of all 1-
types over all finite tuples of M : {pi (x) : i < λ}. The reader not familiar with
ordinals ought to suppose λ = ω for simplicity; in this case the construction is
very understandable.
Let M0 = M. For each i, pi (x) is a 1-type over Mi , so it is realized in some
elementary extension Mi+1 of Mi (a quick look at Remark 1’.5.10 if necessary).
Take unions at limit ordinals (this is required only if λ > ω). Let M1 = Mλ .
As we have taken the increasing union of an elementary chain, Lemma 1’.1.16
says that M  M1 . By construction, all 1-types over finite tuples of M are
realized in M1 .
M1 need not be ω-saturated, as we have added new elements, so we might
have created new types. But we resume the construction, getting M2 . After
countably many steps, the union M∗ is ω-saturated, and an elementary exten-
sion of M by Lemma 1’.1.16 again. e:ultraproductssaturated

116
Remark 1’.5.15. One has no control on the cardinality of an ω-saturated ele-
mentary extension M∗ , even starting with a countable base model M. Indeed,
at each step the number of types might be huge.
For instance, if one considers Z as a ring, then there are continuum many
different 1-types: for each subset I of the primes, consider pI = {x 6= 0} ∪
{∃y x = p.y : p ∈ I} ∪ {∀y x 6= p.y : p 6∈ I} which expresses that the only primes
dividing x are the elements of I.
So there is no countable, ω-saturated elementary extension of the ring Z.
Dealing with 1-types or n-types is the same, as a quick induction reveals.
Lemma 1’.5.16. M is ω-saturated iff any n-type over finitely many parameters
is realized in M.
Proof . Induction. Suppose any n-type with finitely many parameters is real-
ized in M, and let p(x, y) be an n + 1-type with finitely many parameters a (x
still stands for (x1 , . . . , xn ); the n+1th variable is y).
Let q(x) = {∃y ϕ(x, y, a) : ϕ(x, y, a) ∈ p}. This is clearly an n-type over a;
by induction, it is realized by an n-uple b ∈ M .
We now let r(y) = {ϕ(b, y, a) : ϕ(x, y, a) ∈ p}. This is clearly a 1-type over
b, a, which is again a finite tuple. So r is realized in M: there is c ∈ M realizing
it. It is clear that b, c realizes p.
Remark 1’.5.17. Observe that one couldn’t have taken both projections of p,
constructing and realizing first q, then s(y) = {∃x ϕ(x, y, a) : ϕ(x, y, a) ∈ p}.
Indeed, even though c is a realization of s, (a, c) has no reason to realize p!
Again, why are ω-saturated models so important? Because when one tries to
check completeness of a theory, one may focus on them, which are presumably
nicer than the general models. Here is a generalization of Lemma 1’.1.6 to
weaker assumptions.

Lemma 1’.5.18. T is complete iff any two ω-saturated models are elementary
equivalent.
Proof . Let M, N be models of T ; we show that they are elementarily equiva-
lent. By Theorem 1’.5.14, there are M∗  M, N ∗  N which are ω-saturated.
As M ≡ M∗ and N ≡ N ∗ , M∗ and N ∗ are models of T . By assumption, they
are elementary equivalent. It follows M ≡ M∗ ≡ N ∗ ≡ N ; we are done.
Here is our generalization of Corollary 1’.5.7.
Theorem 1’.5.19. Let T be a first-order theory such that for any ω-saturated
models M, N |= T , local isomorphisms between M and N form a Karp family.
Then T eliminates quantifiers.
Proof . We use the elimination theorem, Theorem 1’.5.4. Let M, N |= T ; we
let m ∈ M , n ∈ N satisfy same quantifier-free formulas; we want to show that
they satisfy the same formulas.

117
 First, we may assume that M and N are ω-saturated. By Theorem 1’.5.14,
there are ω-saturated elementary extensions M  M∗ and N  N ∗ . As m
satisfies the same formulas in M and M∗ , it actually suffices to work in M∗
and N ∗ , that is we may assume M and N are ω-saturated.
As m and n satisfy the same quantifier-free formulas, they are locally iso-
morphic, so by assumption they are ∞-isomorphic. They must therefore satisfy
the same formulas. Exercise 1’.10

The conclusion of this abstract discussion is that ω-saturated elementary

extensions always exist, and form a good ground on which to study a theory
(the details are not too important).

End of Lecture 25.

Lecture 26 (Examples)

1’.5.3 Examples
We eventually shed light on all notions by studying examples.

DLO’s
The theory of dense linear ordering has already been defined and studied in
§1’.4.1. We put the existing pieces together.
Proposition 1’.5.20 (DLO’s). DLO is complete and decidable; it eliminates
quantifiers. The theory is ℵ0 -categorical but not ℵ1 -categorical. Moreover, any
model is ω-saturated and local isomorphisms are ∞-isomorphisms.
Proof . We know everything, but here is how a study could go. Past the two
first, many steps are permutable.
(i). The first step is to realize that any DLO is indeed ω-saturated.
(ii). Then, prove that any local isomorphism is an ∞-isomorphism.
(iii). Moreover, as there are no constants, ∅ is a local isomorphism. It follows
from (ii) that any two models are ∞-isomorphic.
(iv). As a consequence of (iii) and with Proposition 1’.4.14, any two countable
models are isomorphic; the theory is ℵ0 -categorical.
(v). One may uses the Łoś-Vaught criterion (Corollary 1’.3.14 to argue that
the theory is complete.
(vi). It would be as simple to say that any two models are elementary equivalent
by (iii); so the theory is complete, again.
(vii). As the theory is clearly effectively presented (finitely many axioms!), and
complete, it is decidable (Corollary 1.6.3).

118
(viii). Turning back to categoricity questions, one would see that the theory
is not ℵ1 -categorical (Counter-example 1’.4.15). It follows from Morley’s
theorem (Theorem 1’.2.16) that it is not κ-categorical in any κ ≥ ℵ1 .
(ix). It follows from (ii) that DLO eliminates quantifiers (Theorem 1’.5.19).

The Successor Function

Proposition 1’.5.21 (N, 0, s). Let T have the following axioms:

• ∀x∀x0 s(x) = s(x0 ) → x = x0 ;

• ∀x s(x) 6= 0;
• ∀x∃y x 6= 0 → s(y) = x;
• ∀x sn (x) 6= x for each n ∈ N.

The theory is complete and decidable; it eliminates quantifiers. It is not ℵ0 -

categorical but is κ-categorical in any κ ≥ ℵ1 . Moreover N embeds elementarily
into every model.
Proof . This is a new study. We must understand the ω-saturated models. In
order to do this we must develop some understanding of the general model.

(i). A model is a set equipped with a function s which is onto except on 0, and
which has no cycles. There are therefore two kinds of orbits: the orbit of
0, which starts and never ends, and other orbits, which have no starting
nor ending point: they are like (Z, s).
``
We see that the general model of T has the form Mi = N I Z, where
I is any set. There is no way to order Z-like orbits.
(ii). Local isomorphisms are easily understood; one would however fail to es-
tablish
` back-and-forth between any two models. It is clear that N and
N Z are not ∞-isomorphic (being countable they would be isomorphic
by Proposition 1’.4.14).

(iii). In particular, the theory is not ℵ0 -categorical. On the other hand since
isomorphism classes are determined by the sole set I, which bears no
structure, T is obviously κ-categorical in any uncountable κ.
(iv). One may apply the Łoś-Vaught criterion to derive completeness of T (and
its decidability, as it is effectively axiomatized).
However we insist that the sooner one understands the ω-saturated models,
the better.
(v). The 1-type p(x) = {sn (0) 6= x : n ∈ N} (no parameters! ` N appears as
indexing set!) is clearly not realized in N, but it is in N Z. On the other
hand, if we fix a ∈ Z, then the 1-type q(x) = {sn (a) 6= x : n ∈ N} over a is

119
 ` ` ` 0
not realized in N Z. If we want to realize it, we must go to N Z Z.
And so on.
``
We then understand that a model is ω-saturated iff of the form N I Z
for an infinite set I.

(vi). One then sees that local isomorphisms between ω-saturated models do
form a Karp family: suppose that a '0 b are drawn from ω-saturated
models, and let α be on the same side as a. We want to find β such that
a, α '0 b, β.
• If α lies in the orbit of some ai , say α = sn (ai ) for n ∈ N, we let
β = sn (bi ). As a '0 b, this creates no confusion!
• If on the other hand α lies in a new orbit, we pick β in a new orbit
too: this is possible, as I is infinite!
(vii). Any two models are clearly locally isomorphic (the orbit of 0 is similar
in either model). It follows from (vi) that any two ω-saturated models
are ∞-isomorphic. It follows that the theory is complete and eliminates
quantifiers.
(viii). In particular, any inclusion of structures is elementary. As N embeds into
any model, it embeds elementarily into any model.

The Ordering
Proposition 1’.5.22 (N, 0, <). Let T be the theory:
• ∀x ¬(x < x)

• ∀x∀y (x < y ∨ x = y ∨ y < x)

• ∀x∀y∀z x < y ∧ y < z → x < z
• (∀y 0 ≤ y) ∧ (∀x∀y x ≤ y → x = 0)
• ∀x∃y∀z y > x ∧ (x < z → y ≤ z)
(There is an element lying immediately above x)
• ∀x∃y∀z x 6= 0 → (y < x ∧ (z < x → z ≤ y))
(If x is not 0, then there is an element lying immediately below x).
The theory is complete and N embeds elementarily into any model. It does not
admit QE, but it does eliminate quantifiers in the language (0, <, {dn : n ∈ N}),
where dn is a new binary predicate symbol standing for

dn (x, y) : ∃x1 . . . ∃xn−1 x < x1 < · · · < xn−1 < y

(”the algebraic distance from x to y is at least n”)

120
Proof . Things are a little more subtle here.
(i). A key observation is that this theory interprets (N, 0, s): one can define (by
` ` function s. In particular, any model
a first-order formula) the successor
of T has the form MI = N < (I,<) Z, where I is now an ordered set.

(ii). One must be very careful that local isomorphisms do not form a Karp
family between ω-saturated models. Indeed, pairs (x, x+ ) and (x, x++ )
are locally isomorphic, but back-and-forth between them is not possible!
We already have a feeling that distance comes into play. But we reserve
language-related issues for the end.
(iii). If one has studied ω-isomorphisms (§1’.4.3), one sees that any two models
are ω-isomorphic, so the theory is complete. There is another proof in (v)
but this one is recommended, being simple and elegant.
(iv). For a model to be ω-saturated, I must be infinite. This is however
not enough. Fix a in some copy of Z, and consider the 1-type p(x) =
{∃x1 . . . ∃xn−1 a < x1 · · · < xn−1 < x : n ∈ N} over a. It means that x
is infinitely bigger than a, that x lies in a copy of Z lying above that in
which a lives.
In particular, we understand that for the model to be ω-saturated, I must
have no largest element. And similarly, it must have no least element: it
has no endpoints. This is still not enough.
Fix a < b in I, and consider the 1-type q(x) = {∃x1 . . . ∃xn−1 a < x1 · · · <
xn−1 < x : n ∈ N} ∪ {∃x1 . . . ∃xn−1 x < x1 · · · < xn−1 < b : n ∈ N} over
(a, b): it expresses that the orbit of x contains neither a nor b, that x lives
in a copy of Z which lies between that of a and that of b. This is density
of the ordering I!
And we now see that MI is ω-saturated iff I is a DLO.
(v). One could reprove completeness: the theory is of course not ℵ0 -categorical
(neither is it ℵ1 -categorical, since DLO isn’t), but any two countable, ω-
saturated models are isomorphic.
Indeed, let MI and MJ be such models. I and J are countable DLO’s,
whence isomorphic. We then construct an isomorphism MI ' MJ by
identifying copies of Z.
In particular, any two countable and ω-saturated models are elementary
equivalent. This is however not enough. Let N be any model; by the
Löwenheim-Skolem theorem (Theorem 1’.3.13), it contains a countable
elementary substructure N0 = MI , where I is countable. Now I embeds
into a countable DLO I 0 , so N0 embeds elementarily into N1 = MI 0 . It
follows that N ' N1 , and completeness is proved.
This method is however not recommended, as it relies on the existence of
countable, ω-saturated models; one should prefer the proof of (iii).

121
(vi). We let L = {0, <} and L0 = L ∪ {dn : n ∈ N} be the expanded language.
Notice that WFF(L) = WFF(L0 ); we don’t change the expressive strength
of the language, but merely force certain quantified formulas to now be
quantfier-free.
In particular, a model of T is ω-saturated in L iff ω-saturated in L0 : our
description of ω-saturation does not change.
(vii). One now sees that L0 -local isomorphisms between ω-saturated models are
L0 -∞-isomorphisms.
From there, quantifier elimination in L0 is immediate. As N ⊆ MI is also
an inclusion of L0 -structures, one finds N  MI .

Algebraically Closed Fields*

Definition 1’.5.23 (ACF). An algebraically closed field is a model of the fol-
lowing theory in the language {0, 1, +, −, ·} of rings:
• the structure is a field;
• for each n ∈ N, the axiom ∀an . . . ∀a0 ∃x an 6= 0 → an · xn + · · · + a0 = 0.
ACF denotes either this theory, or a model of the theory.

We let P denote the set of prime numbers.

Definition 1’.5.24 (ACFq ).
• For p ∈ P, an algebraically closed field of characteristic p is a model of
the theory ACF ∪{1 + · · · + 1 = 0}
| {z }
p times

• An algebraically closed field of characteristic 0 is a model of the theory

ACF ∪{1 + · · · + 1 6= 0 : p ∈ P}.

Proposition 1’.5.25. ACF eliminates quantifiers. Moreover, for any q ∈ P ∪

{0}, ACFq is complete. It is not ℵ0 -categorical, but is κ-categorical in any
uncountable κ.
Proof . We begin with simple remarks: given algebraically closed fields K, L,
one has K '0 L iff they have the same characteristic. Moreover, an ACF is
ω-saturated iff it has infinite transcendence degree over its prime field.
It is then very easy to see that if K, L are ω-saturated ACF’s of the same
characteristic, then local isomorphisms form a Karp family. In particular, K ≡
L. ACFq is complete! Now the characteristic is in the quantifier-free type of
any tuple, so ACF eliminates quantifiers.
As for categoricity, Q and Q(X) are clearly non-isomorphic; but playing
with transcendence bases, one easily finds that ACFq is κ-categorical in any
uncountable κ.

122
Corollary 1’.5.26 (Hilbert’s Nullstellensatz). Let K be an algebraically closed
field, X = (X1 , . . . , Xn ) a tuple of indeterminates, and I C K[X] an ideal of
K[X]. Then I has a solution in Kn .
Proof . We may assume that I is a maximal ideal m; notice that by Noetherian-
ity, m is finitely generated. In particular, existence of a solution is a first-order
formula with parameters in K.
We let L be the algebraic closure of the field K[X]/m; clearly K ⊆ L. As
both are models of ACF and the theory eliminates quantifiers, one actually has
K  L. Now there is a solution to m in Ln (the class modulo m of the tuple X);
since this is first-order, there is a solution in Kn too.
More interesting results hold, but one unfortunately has to stop some day.

End of Lecture 26.

1’.6 Exercises
Exercise 1’.1. Show Lemma 1’.1.16, that is show that given an increasing
elementary chain (Mi )i∈I of L-structures, the union M∗ is an elementary ex-
tension of each Mi .
Exercise 1’.2. Let T be a theory which has models of cardinality n for arbi-
trarily large n ∈ N. Show that T has an infinite model.
Exercise 1’.3. Finish the proof of Lemma 1’.1.11 by proving directly that when-
ever M ⊆ N and ϕ(m) is a universal formula with parameters m in M , one has
N |= ϕ(m) ⇒ M |= ϕ(m).
Exercise 1’.4. Let L = {R}, a binary relation, and T express that R is an
equivalence relation.
1. Let M, N |= T . Find a necessary and sufficient condition for M |= N .
2. Let M ⊆ N |= T . Find a necessary and sufficient condition for M  N .
No justification is required.
Exercise 1’.5. Let H, G be groups with H  G (the language contains Lgrp =
(e,−1 , ·), but could be larger).
1. Prove that if G is simple, so is H.
2. Prove that the converse may fail.
Exercise 1’.6. A theory T is universal if it the set of consquences of a set of
universal sentences.
Also, if T is a theory, let T∀ be the set of all its consequences that are
universal; that is

T∀ = {ϕ an L-sentence such that T |= ϕ}

123
 1. The language is {·}. Is the theory of groups universal? (no justification
required)
2. The language is {e,−1 , ·}. Is the theory of groups universal? (no justifi-
cation required)
3. Show that T is universal iff T∀ |= T .
Exercise 1’.7. Two theories T and T 0 are companions if every model of T is
included (not necessarily elementarily!) in a model of T 0 , and vice-versa.
1. Find an interesting companion of the theory of orderings. Find an inter-
esting companion of the theory of rings.
2. Prove that T and T∀ are companions.
3. Deduce that T and T 0 are companions if and only if T∀ and T∀0 are com-
panions.
4. Prove that T is universal if and only if every substructure of a model of
T is a model of T .
5. Prove that T∀ ⊆ T∀0 if and only if every model of T 0 is included (not
necessarily elementarily, well-understood) in a model of T .
Exercise 1’.8.
1. Suppose M1 ≡ M2 . Show that there is M∗ such that M1 and M2 embed
into M∗ elementarily.
2. Let N be a model which embeds elementarily into Mi for each i ∈ I.
Show that there is a common elementary extension M∗  Mi in which
all copies of N are glued togehter.
Exercise 1’.9. Let T be a theory and ϕ(x) be a formula. Prove that ϕ is
equivalent modulo T to an existential formula if and only if:
for all models M and N of T with M ⊆ N and a ∈ M ,
one has M |= ϕ(a) ⇒ N |= ϕ(a).
[Hint: one implication is trivial. For the other, study the set Ψ := {ψ(c) an
existential formula such that T |= ψ(c) → ϕ(c)}. Prove that T1 := T ∪ {¬ψ(c) :
ψ ∈ Ψ} ∪ {ϕ(c)} is not consistent.]
Exercise 1’.10. Show the converse of Theorem 1’.5.19, that is show: if T elim-
inates quantifiers and M, N are ω-saturated models of T , then local isomor-
phisms form an ∞-isomorphism family.
Exercise 1’.11. The language consists of a single binary relation symbol −.
The theory of graphs is given by the axioms:
• ∀x ¬(x − x)

124
 • ∀x∀y (x − y → y − x).
A graph G is connected if for any x, y ∈ G, there is a path from x to y. (This
definition does not sound first-order: we shall prove it.) Show that there is no
theory whose models are exactly all connected graphs.

Exercise 1’.12. Consider the two groups Q and Q ⊕ Q. How isomorphic are
they (as groups)? Same question when they are considered as rings.
Exercise 1’.13. We consider the following <-structures:
a a a
A = ... Z ... and B = A B
< < <
| {z }
Z
`
where X < Y means that the disjoint union is ordered in such a way that every
y ∈ Y lies above every x ∈ X.
How isomorphic are A and B?

Exercise 1’.14. Let T be the {R}-theory asserting that R is an equivalence

relation. For any model M, define the following numbers:
for each k ∈ N ∪ {∞}, mk ∈ N ∪ {∞} is the number of R-classes
with exactly k elements.
Let M and N be models with numbers (mk ) and (nk ) respectively. Determine
how isomorphic M and N are.
Exercise 1’.15. Let T be a complete theory with infinite models, and ϕ(x) be
a formula. Suppose that there is a cardinal κ such that for any model M, there
are at most κ elements of M satisfying ϕ(x). Show that there is a number n
such that in any model, ϕ(x) has exactly n elements.

Exercise 1’.16. Let M be an L-structure. Prove that a definable subset that

is invariant under every automorphism of M is definable without parameters.
[Hint: let X = ϕ(x, a) be the definable set and consider Ψ := {ψ(x) : M |=
∀x ψ(x) → ϕ(x, a)}. Form an La (x)-theory and prove it inconsistent...]
Exercise 1’.17. Recall that a theory is said to be universal if it has an axiom-
atization by universal sentences. Also, if T is a theory, let T∀ be the set of all
its consequences that are universal.
Two theories T and T 0 are companions if every model of T is included (not
necessarily elementarily!) in a model of T 0 , and vice-versa.
1. Find interesting examples.

2. Prove that T and T∀ are companions.

3. Deduce that T and T 0 are companions if and only if T∀ and T∀0 are com-
panions.

125
 4. Prove that T is universal if and only if every sub-structure of a model of
T is a model of T .
5. Prove that T∀ ⊆ T∀0 if and only if every model of T 0 is included (not
necessarily elementarily, well-understood) in a model of T .

Exercise 1’.18. Show that a non-principal ultraproduct is ω-saturated (it is

even ω1 -saturated).
Exercise 1’.19. Let L = {<} and T be the following theory:
• ∀x ¬(x < x)

• ∀x∀y (x < y ∨ x = y ∨ y < x)

• ∀x∀y∀z (x < y ∧ y < z) → x < z
• ∀x∃y∀z (z > x) → (z > y ∨ z = y)
• ∀x∃y∃z (y < x ∧ x < z)

1. Explain what the theory says (it has a fairly standard name).
2. Describe countable models of T (no justification required).

3. Describe ω-saturated models of T (no justification required).

4. Prove that any two ω-saturated models of T are ∞-isomorphic.
5. Show that it is not true that local isomorphisms between ω-saturated
models of T are ∞-isomorphisms.

6. Show that T is complete.

7. Does T eliminate quantifiers?
8. Is T ℵ0 -categorical? Is it ℵ1 -categorical?

9. Show that any model of T embeds into a model of DLO.

10. Show that any model of DLO embeds into a model of T .

Exercise 1’.20. Let L = {B, R, <} where B and R (“blue” and “red”) are unary
predicate symbols, and < is a binary relation symbol. Consider the following
conditions:
• every point is either blue or red, but not both

• < is an order relation

• above and below any blue point, there is a red point
• above and below any red point, there is a blue point

126
• between any two points, there is a point
• between any two blue points, there is a red point
• between any two red points, there is a blue point

1. Write down L-sentences expressing the above axioms. Let T be the re-
sulting L-theory.
2. Show that an order-preserving function need not be a local isomorphism.

3. Prove that any two models of T are ∞-isomorphic.

4. Show that T is complete.
5. Does T eliminate quantifiers?

6. Is T ℵ0 -categorical? Is it ℵ1 -categorical?
7. What happens if we drop the clause that between any two points there is
a point?

127
Index of Notions

Algebraically Closed Field*, 122 Finite Intersection Property*, 34

ACFq , 122 Formula
Assignment Atomic Formula, 48
Truth Assignment, 11 Existential Formula, 90
of the Variables, 51 Quantifier-Free Formula, 90
Atomic Formula, see Formula Universal Formula, 90
Axiom Turning one into a Sentence, 91
Axioms of Equality, 57
Deduction Rule, 16 Inclusion, 93
Axiomatization of a Theory, 32 Elementary Inclusion, 95
Interpretation, 51
Back-and-Forth Isomorphism, 97
Finitary*, 108 0-Isomorphism, 106
of Height ω*, 109 n-Isomorphism*, 108
Infinite, 107 ω-Isomorphism*, 109
∞-Isomorphism, 107
Categoricity, 99
Absolute Categoricity, 87 Karp Family, 107
Compactness Kripke Model*, 35
in Topology*, 34
Complete Theory, 29, 71 Language
V-Complete, 67 of Propositional Logic, 8
Consequence of Modal Logic*, 35
Semantic Consequence, 12, 52 First-Order Language, 46
Syntactic Consequence, 16, 57 Second-Order Language, 85
Consistency, 17, 58 Local Isomorphism, see Isom.

Decidable, 32 Model, 52, 91

Semi-Decidable, 32 in Modal Logic*, see Kripke
Deduction, 16, 57 Modus Ponens, 17
D. Rules Modus Tollens, 17
for Connectives, 16 Morphism, 97
for ∀ and ∃, 57 Elementary Morphism, 98
Dense Linear Ordering, 104
Non-Standard Reals*, 73
Elementary
Equivalence, 92 Peano Arithmetic, 87
Inclusion, see Inclusion
Morphism, see Morphism Quantification Rank, 91
Substructure, see Substructure Quantifier Elimination, 111
Elimination Set, 111
Satisfaction
Equality Axioms, see Axiom
in Propositional Logic, 12
Filter*, 74 in Modal Logic*, 36

128
 in First-Order Logic, 51
Satisfiability
in Propositional Logic, 12
in First-Order Logic, 52
ω-Saturation, 116
Sentence, 49
Signature, 46
Skolem Function, 101
Standard Part*, 74
Structure, 50
Substitutable Term, see Term
Substructure, 93
Elementary Substructure, 95

Term, 47
Substitutable Term, 53
Theory
in Propositional Logic, 9
in First-Order Logic, 49
of a First-Order Structure, 92
Truth Assignment, see Assignment
Type, 115
of a Tuple, 109

Ultrafilter*, 75
Ultraproduct*, 76

Variable, Free or Bound, 49

Weakening, 16
Well-Formed Formula
in Propositional Logic, 8
in First-Order Logic, 48
Witness, 63

129
Index of Results

Categoricity of Th(N, 0, s), 119

Absolute C. of PA, 87 of ACF*, 122
Morley’s C. Theorem*, 99
ℵ1 -C. of ACF*, 122 Renaming Algorithm, 54
ℵ0 -C. of DLO, 104, 118
ℵ1 -C. of Th(N, 0, s), 119 Skolem’s Paradox, 102
Compactness Soundness
of Propositional Logic, 31 of Propositional Logic, 27
of Modal Logic*, 40 of Modal Logic*, 37
of First-Order Logic, 71 of First-Order Logic, 61
fails in Higher Logic, 87
Tarski’s Test for , 96
Completeness
of Propositional Logic, 29 Unique Readability, 9, 48
of Modal Logic*, 38–40
of First-Order Logic, 62
Completeness of a Theory
Criteria, 92, 103, 117
of DLO, 105, 118
of the Ordering on N, 120
of the Successor Function, 119
of ACFq *, 122

Decidability*
Semi-D. of a Theory, 33, 70
of a Complete Theory, 33, 71

Elimination Theorem, 111

Excluded Middle, 19

Generalisation Theorem, 64

Łoś’ Theorem*, 77
Łoś-Vaught Completeness Test, 103
Löwenheim-Skolem Theorem
Descending Version, 102
Ascending Version, 103
Final Version, 103

ω-Saturated Models Do Exist, 116

QE
Criteria, 113, 117
of DLO, 118
of Th(N, 0, <), 120

130
List of Lectures

Lecture 1 (Language and Wff’s; Unique Readability) . . . . . . . . . . . . 7

Lecture 2 (The Semantics of Propositional Logic) . . . . . . . . . . . . . . 11
Lecture 3 (Natural Deduction in Classical Logic) . . . . . . . . . . . . . . 15
Lecture 4 (Reduction to Two Connectives (1/2)) . . . . . . . . . . . . . . 20
Lecture 5 (Reduction to Two Connectives (2/2); Soundness) . . . . . . . . 24
Lecture 6 (Completeness of Propositional Logic) . . . . . . . . . . . . . . 28
Lecture 7 (Compactness of Propositional Logic) . . . . . . . . . . . . . . . 30
Lecture 8 (Modal Logic*) . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Lecture 9 (ME1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Lecture 10 (First-Order Languages; Terms and Formulas) . . . . . . . . . 45
Lecture 11 (The Semantics of First-Order Logic) . . . . . . . . . . . . . . 50
Lecture 12 (Substitutions) . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Lecture 13 (Deductions; Simplifying the Language) . . . . . . . . . . . . . 57
Lecture 14 (Soundness; Completeness (1/2)) . . . . . . . . . . . . . . . . . 61
Lecture 15 (Completeness (2/2)) . . . . . . . . . . . . . . . . . . . . . . . 66
Lecture 16 (Consequences; Compactness; Non-Standard Analysis) . . . . 70
Lecture 17 (Proof of Compactness by Ultraproducts*) . . . . . . . . . . . 74
Lecture 18 (Second-Order Logic) . . . . . . . . . . . . . . . . . . . . . . . 85
Lecture 19 (ME2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Lecture 20 (Models and Theories; Elementary Equivalence) . . . . . . . . 89
Lecture 21 (Elementary Inclusion; Morphisms) . . . . . . . . . . . . . . . 94
Lecture 22 (Löwenheim-Skolem Theorems) . . . . . . . . . . . . . . . . . 99
Lecture 23 (Back-and-Forth Methods) . . . . . . . . . . . . . . . . . . . . 103
Lecture 24 (Elimination Theorem) . . . . . . . . . . . . . . . . . . . . . . 111
Lecture 25 (ω-Saturated Models) . . . . . . . . . . . . . . . . . . . . . . . 114
Lecture 26 (Examples) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

131