STML 089

S T U D E N T M AT H E M AT I C A L L I B R A RY
Volume 89
A First Journey
through Logic
Martin Hils
François Loeser
Licensed to Mathematisches Forschungsinstitut. Prepared on Thu Sep 30 04:18:29 EDT 2021for download from IP 188.1.230.74.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
A First Journey
through Logic
10.1090/stml/089
S T U D E N T M AT H E M AT I C A L L I B R A RY
Volume 89
A First Journey
through Logic
Martin Hils
François Loeser
Editorial Board
Satyan L. Devadoss John Stillwell (Chair)
Rosa Orellana Serge Tabachnikov
2010 Mathematics Subject Classification. Primary 03-01.
For additional information and updates on this book, visit

www.ams.org/bookpages/stml-89
Library of Congress Cataloging-in-Publication Data

Names: Hils, Martin, 1973- author. | Loeser, François, author.
Title: A first journey through logic / Martin Hils, François Loeser.
Description: Providence, Rhode Island : American Mathematical Society, [2019] | Series: Student math-
ematical library ; volume 89 | Includes bibliographical references and index.
Identifiers: LCCN 2019014487 | ISBN 9781470452728 (alk. paper)
Subjects: LCSH: Logic, Symbolic and mathematical–Textbooks. | Mathematics–Textbooks. | AMS:
Mathematical logic and foundations – Instructional exposition (textbooks, tutorial papers, etc.). msc
Classification:
LCC QA9 .H52445 2019 | DDC 511.3–dc23
LC record available at https://lccn.loc.gov/2019014487
Copying and reprinting. Individual readers of this publication, and nonprofit libraries acting for
them, are permitted to make fair use of the material, such as to copy select pages for use in teaching or
research. Permission is granted to quote brief passages from this publication in reviews, provided the
customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication is
permitted only under license from the American Mathematical Society. Requests for permission to
reuse portions of AMS publication content are handled by the Copyright Clearance Center. For more
information, please visit www.ams.org/publications/pubpermissions.
Requests for translation rights and licensed reprints should be sent to reprint-
permission@ams.org.
© 2019 by the American Mathematical Society. All rights reserved.

The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.
⃝
∞ The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at https://www.ams.org/
10 9 8 7 6 5 4 3 2 1 24 23 22 21 20 19
Contents
Introduction ix
Chapter 1. Counting to Infinity 1

Introduction 1
§1.1. Naive Set Theory 1
§1.2. The Cantor and Cantor-Bernstein Theorems 2
§1.3. Orders 3
§1.4. Operations on Orders 5
§1.5. Ordinal Numbers 7
§1.6. Ordinal Arithmetic 11
§1.7. The Axiom of Choice 14
§1.8. Cardinal Numbers 14
§1.9. Operations on Cardinals 16
§1.10. Cofinality 19
§1.11. Exercises 22
§1.12. Appendix: Hindman’s Theorem 28
Chapter 2. First-order Logic 33

Introduction 33
vi Contents
§2.1. Languages and Structures 34

§2.2. Terms and Formulas 36
§2.3. Semantics 39
§2.4. Substitution 41
§2.5. Universally Valid Formulas 45
§2.6. Formal Proofs and Gödel’s Completeness Theorem 48
§2.7. Exercises 58
Chapter 3. First Steps in Model Theory 65

Introduction 65
§3.1. Some Fundamental Theorems 66
§3.2. The Diagram Method 70
§3.3. Expansions by Definition 72
§3.4. Quantifier Elimination 75
§3.5. Algebraically Closed Fields 78
§3.6. Ax’s Theorem 81
§3.7. Exercises 83
Chapter 4. Recursive Functions 89

Introduction 89
§4.1. Primitive Recursive Functions 90
§4.2. The Ackermann Function 94
§4.3. Partial Recursive Functions 96
§4.4. Turing Computable Functions 98
§4.5. Universal Functions 107
§4.6. Recursively Enumerable Sets 109
§4.7. Elimination of Recursion 113
Chapter 5. Models of Arithmetic and Limitation Theorems 119

Introduction 119
§5.1. Coding Formulas and Proofs 120
Contents vii
§5.2. Decidable Theories 122

§5.3. Peano Arithmetic 125
§5.4. The Theorems of Tarski and Church 131
§5.5. Gödel’s First Incompleteness Theorem 133
§5.6. Definability of Satisfiability for Σ1 -formulas 134
§5.7. Gödel’s Second Incompleteness Theorem 137
Chapter 6. Axiomatic Set Theory 147

Introduction 147
§6.1. The Framework 147
§6.2. The Zermelo-Fraenkel Axioms 148
§6.3. The Axiom of Choice 155
§6.4. The von Neumann Hierarchy and the Axiom of
Foundation 157
§6.5. Some Results on Incompleteness, Independence and
Relative Consistency 161
§6.6. A Glimpse of Further Independence and Relative
Consistency Results 169
Bibliography 179
Index 181
Introduction
The aim of this book, which originates from a course that we taught suc-
cessively at École Normale Supérieure (FL in 2007-2010 and MH in 2010-
2013), is to present a broad panorama of Mathematical Logic to students
who feel curious about this field but have no intent to specialize in it. As
a consequence we have deliberately chosen not to write another com-
prehensive textbook, of which there already exist quite a few excellent
ones, but instead to deliver a slim text which provides direct routes to
some significant results of general interest.
Our point of view is to treat Logic on an equal footing to any other
topic in the mathematical curriculum. Since one does not have to define
natural numbers when teaching Number Theory, or sets when teaching
Analysis, why should we in a Logic course? For this reason we start the
book with a presentation of naive Set Theory, that is, the theory of sets
that mathematicians use on a daily basis. It is only in the last chapter
that we discuss the Zermelo-Fraenkel axioms, which in fact most math-
ematicians who are not Set Theorists or teaching a logic course are not
so familiar with.
In each chapter we have tried to present at least a few juicy high-
lights, outside Logic whenever possible, either in the main text, or as ex-
ercises or appendices. We consider exercises as an essential component
of the book, and we encourage the reader to work them out thoroughly;
ix
x Introduction
they should be seen not only as a tool to check that the course is correctly
assimilated, but also as a way to provide an opening to additional topics
of interest.
The book is organized as follows. In the first chapter, in addition to
the basic theory of ordinal and cardinal numbers, we cover more exotic
topics like Goodstein sequences, infinite combinatorics (clubs and Solo-
vay’s Theorem) and Hindman’s Theorem (a striking result in additive
combinatorics). In Chapter 2 we introduce First-order Logic and for-
mal proofs. We prove Gödel Completeness via Henkin witnesses. Craig
Interpolation and Beth Definability are treated in exercises. The next
chapter delves deeper inside Model Theory, with detailed coverage of
Quantifier Elimination. In particular we prove Quantifier Elimination
for algebraically closed fields, which allows one to state and prove the
Lefschetz Principle in Algebraic Geometry and Ax’s Theorem on surjec-
tivity of injective polynomial mappings. Chapter 4 is devoted to basic Re-
cursion Theory and culminates with the existence of universal recursive
functions, undecidability of the Halting Problem and Rice’s Theorem.
In Chapter 5 we prove the classical undecidability and incompleteness
results of Tarski and Church and provide a complete proof of Gödel’s Sec-
ond Incompleteness Theorem which we found in Martin Ziegler’s book
[13]. We also present, as an exercise, a theorem of Tennenbaum about
the inexistence of non-standard countable recursive models of Peano.
Finally in Chapter 6 we develop Axiomatic Set Theory, including the
Reflection Principle and some proofs of independence and relative con-
sistency.
This book is intended towards advanced undergraduate students,
graduate students at any stage, or working mathematicians, who seek a
first exposure to core material of mathematical logic and some of its ap-
plications. Prerequisites are minimal: besides familiarity with abstract
reasoning and basic mathematical concepts, some acquaintance with
General Topology and Algebra, especially Field Theory, is required at
various places.
For the interested reader, here are a few suggestions for further read-
ing, providing more comprehensive and advanced material, ordered by
increasing difficulty:
Model Theory: The books by Marker [8], Poizat [9] and Tent-Ziegler [12].
Introduction xi
Set Theory: The books by Krivine [6], Kunen [7] and Jech [5].
Recursion Theory: The books by Cooper [1], Rogers [10] and Soare [11].
Acknowledgements
We heartfully thank the following colleagues and friends who encour-
aged us in the project of transforming our notes into a book and/or helped
us immensely in improving the text: Martin Bays, Antoine Chambert-
Loir, Zoé Chatzidakis, Artem Chernikov, Raf Cluckers, Arthur Forey,
Franziska Jahnke, Silvain Rideau, Pierre Simon. Moreover, we thank
Christian Maurer who provided the drawing for the book cover.
During the preparation of this book the first author was partially
supported by ANR through ValCoMo (ANR-13-BS01-0006) and by DFG
through SFB 878 and the second author was partially supported by ANR
through Défigéo (ANR-15-CE40-0008) and by the Institut Universitaire
de France.
10.1090/stml/089/01
Chapter 1
Counting to Infinity
Introduction
The general purpose of this chapter is to provide tools for comparing
sizes of infinities. To this aim we develop in 1.3-1.6 the notion of ordinals
which constitute a natural infinite extension of natural numbers. Ordi-
nals classify well-ordered sets and are a natural device for using transfi-
nite induction. A fundamental and very useful fact is that ordinals are
endowed with arithmetic operations extending those of natural num-
bers, even though some classical properties do not extend (for instance
addition of ordinals is no longer commutative).
Building on the notion of ordinals, we develop in 1.7-1.10 the con-
cept of cardinals, which is the right notion to compare the size of sets.
Unlike the case of ordinals, one needs to assume the validity of the Ax-
iom of Choice (which we discuss in 1.7) to develop a full fledged the-
ory of cardinals. In the last sections of this chapter, we study cardinal
arithmetic, which appears to have a much richer theory of exponenta-
tion than ordinal arithmetic.
1.1. Naive Set Theory

In this chapter we shall use the notions of set and natural number in
the same way as in any mathematical textbook, that is, in a naive sense,
2 1. Counting to Infinity
without further questioning. It is only in the last chapter that we shall

introduce the classical ZFC system of axioms of Zermelo-Fraenkel (plus
the Axiom of Choice). We shall see that the notions and results of this
first chapter remain valid in the more formal axiomatic Set Theory, using
only the axioms of ZFC. We shall thus develop Cantor’s theory of ordinal
and cardinal numbers from this naive point of view.
When 𝐴 and 𝐵 are sets, we denote by 𝐴∪𝐵, 𝐴∩𝐵 and by 𝐴⧵𝐵 their set-
theoretic union, intersection and difference, respectively. More generally,
if 𝐼 is a set and (𝐴𝑖 )𝑖∈𝐼 is a family of sets indexed by 𝐼, we denote its union
by ⋃𝑖∈𝐼 𝐴𝑖 and its intersection by ⋂𝑖∈𝐼 𝐴𝑖 ; thus we have 𝑥 ∈ ⋃𝑖∈𝐼 𝐴𝑖 if
and only 𝑥 ∈ 𝐴𝑖 for some 𝑖 ∈ 𝐼 and 𝑥 ∈ ⋂𝑖∈𝐼 𝐴𝑖 if and only 𝑥 ∈ 𝐴𝑖 for
every 𝑖 ∈ 𝐼.
We write 𝐴 ⊆ 𝐵 if 𝐴 is a subset of 𝐵, and 𝐴 ⊂ 𝐵 if 𝐴 is a proper subset
of 𝐵. The power set of 𝐴 is denoted by 𝒫(𝐴). It is the set of subsets of 𝐴;
thus 𝐶 ∈ 𝒫(𝐴) if and only if 𝐶 ⊆ 𝐴.
We denote by ℕ the set {0, 1, 2, …} of natural numbers, and we write
ℕ∗ for ℕ ⧵ {0}.
We will make constant use of the extensionality principle according
to which two sets containing the same elements are equal.
We shall also make use of the comprehension principle which states
that given a set 𝐴 and a property 𝑃 of sets, there exists a set whose ele-
ments are exactly those elements of 𝐴 that satisfy property 𝑃. We defer
to Section 6.2 for a more precise formulation.
1.2. The Cantor and Cantor-Bernstein Theorems

The existence of an injective, surjective or bijective function between
two sets may be seen as a way to compare their “size”. We start with two
results going in that direction.
Theorem 1.2.1 (Cantor). Let 𝐴 be a set. There is no surjection 𝐴 → 𝒫(𝐴).
Proof. Let 𝑓 ∶ 𝐴 → 𝒫(𝐴) be a map. Consider the set

𝐵 = {𝑥 ∈ 𝐴 ∣ 𝑥 ∉ 𝑓(𝑥)}.
For every 𝑥 ∈ 𝐴 with 𝑓(𝑥) = 𝐵, we have 𝑥 ∈ 𝐵 if and only if 𝑥 ∉ 𝐵.
Hence 𝐵 does not belong to the image of 𝑓. □
1.3. Orders 3
Theorem 1.2.2 (Cantor-Bernstein). Let 𝐴 and 𝐵 be sets, and let 𝑓 ∶ 𝐴 →

𝐵 and 𝑔 ∶ 𝐵 → 𝐴 be injective maps. Then there exists a bijection ℎ ∶ 𝐵 →
𝐴.
Proof. We may assume 𝐴 is a subset of 𝐵 and 𝑓 is the inclusion map.

Indeed, one may replace 𝐴 by 𝑓(𝐴) and 𝑔 by 𝑓 ∘ 𝑔. Now set 𝐶 = {𝑔𝑛 (𝑥) ∣
𝑛 ∈ ℕ, 𝑥 ∈ 𝐵 ⧵ 𝐴}. Define a map ℎ ∶ 𝐵 → 𝐴 by ℎ(𝑐) = 𝑔(𝑐) when 𝑐 ∈ 𝐶
and ℎ(𝑥) = 𝑥 when 𝑥 ∈ 𝐵 ⧵ 𝐶. The map ℎ is surjective: indeed, any
𝑥 ∈ 𝐴 ∩ 𝐶 is of the form 𝑥 = 𝑔(𝑦) for some 𝑦 ∈ 𝐶 and for any 𝑥 ∈ 𝐴 ⧵ 𝐶,
𝑥 = ℎ(𝑥). It is also clearly injective. □
Definition. Let 𝑋 and 𝑌 be sets. One says that 𝑋 and 𝑌 are equinumer-
ous, and writes 𝑋 ∼ 𝑌 , if there exists a bijection between 𝑋 and 𝑌 ; one
says 𝑋 is subnumerous to 𝑌 , and writes 𝑋 ≼ 𝑌 , if there exists an injection
𝑋 → 𝑌.
Using this terminology, the Cantor-Bernstein Theorem may be re-

stated as: if 𝑋 ≼ 𝑌 and 𝑌 ≼ 𝑋, then 𝑋 ∼ 𝑌 .
1.3. Orders
This section and the next one are devoted to preliminary results on or-
dered sets that are needed to develop the theory of ordinals.
Definition. A partial order < on a set 𝑋 is a binary relation (that is,

given by a subset of 𝑋 × 𝑋) which is transitive (if 𝑥 < 𝑦 and 𝑦 < 𝑧 then
𝑥 < 𝑧) and antireflexive (𝑥 ≮ 𝑥). If furthermore, for every 𝑥, 𝑦 ∈ 𝑋 one
has 𝑥 < 𝑦, 𝑥 = 𝑦 or 𝑦 < 𝑥, one says < is a total order.
One writes 𝑥 ≤ 𝑦 to mean 𝑥 < 𝑦 or 𝑥 = 𝑦, 𝑥 > 𝑦 for 𝑦 < 𝑥, and 𝑥 ≥ 𝑦

for 𝑦 ≤ 𝑥.
If 𝑌 ⊆ 𝑋, 𝑦 ∈ 𝑌 is a smallest element if for every 𝑦′ in 𝑌 , 𝑦 ≤ 𝑦′ . It
is a minimal element if for every 𝑦′ in 𝑌 , 𝑦′ ≮ 𝑦. One similarly defines a
largest element and a maximal element. A lower bound of 𝑌 is an element
of 𝑋 which is ≤ all elements in 𝑌 . An infimum of 𝑌 is a largest element
in the set of lower bounds of 𝑌 . One similarly defines the notion of upper
bound and supremum.
Note that in a partial order, if 𝑎 ≤ 𝑏 and 𝑏 ≤ 𝑎 then 𝑎 = 𝑏. Thus

smallest and largest elements are unique, which is in general not the
case for minimal or maximal elements.
Remark 1.3.1. If < is a partial order, ≤ is a reflexive relation (𝑥 ≤ 𝑥 for
every 𝑥 ∈ 𝑋), transitive and antisymmetric (if 𝑥 ≤ 𝑦 and 𝑦 ≤ 𝑥 then
𝑥 = 𝑦).
Conversely, if ≤ is a binary relation on a set 𝑋 which is reflexive,
transitive and antisymmetric then the relation < defined by 𝑥 < 𝑦 ∶⇔
(𝑥 ≤ 𝑦 and 𝑥 ≠ 𝑦) is a partial order on 𝑋.
Proof. Exercise. □
Definition.
(1) Let < be a partial order on 𝑋. We say < is well-founded if any
non-empty subset of 𝑋 contains a minimal element.
(2) A well-order is a well-founded total order.
Remark 1.3.2. Let < be a partial order on 𝑋.
(1) The map 𝑎 ↦ 𝑋≤𝑎 = {𝑥 ∈ 𝑋 ∣ 𝑥 ≤ 𝑎} identifies (𝑋, <) with
a subset 𝑌 of 𝒫(𝑋) endowed with the partial order induced by
⊂.
(2) < is well-founded if and only if there is no infinite decreasing
sequence in 𝑋.
(3) < is a well-order if and only if every non-empty subset of 𝑋
contains a smallest element.
Proof. The proofs of (1) and (3) are left as exercises.

Let us prove (2). If (𝑋, <) is not well-founded, there exists a subset
∅ ≠ 𝑌 ⊆ 𝑋 without minimal element. By induction on 𝑛 ∈ ℕ one
may construct 𝑦𝑛 ∈ 𝑌 such that 𝑦𝑛+1 < 𝑦𝑛 . Conversely, if (𝑦𝑛 )𝑛∈ℕ is a
decreasing sequence, it is clear that 𝑌 = {𝑦𝑛 ∣ 𝑛 ∈ ℕ} does not contain
a minimal element. □
Example 1.3.3.
(1) The usual order < on ℤ is a non-well-founded total order. Its
restriction to ℕ is a well-order.
1.4. Operations on Orders 5
(2) For every set 𝑋, the relation ⊂ is a partial order on 𝒫(𝑋) which
is well-founded if and only if 𝑋 is finite.1
(3) The restriction of a partial order (resp. total, well-founded) on
𝑋 to 𝑌 ⊆ 𝑋 is a partial order (resp. total, well-founded) on 𝑌 .
1.4. Operations on Orders

Definition. Let 𝑋 and 𝑌 be partially ordered sets.
(1) The ordered sum of 𝑋 and 𝑌 , denoted by 𝑋 + 𝑌 , is the partially
ordered set consisting of pairs (𝑥, 0) with 𝑥 ∈ 𝑋 and (𝑦, 1) with
𝑦 ∈ 𝑌 , the order being defined as follows: (𝑎, 𝑖) < (𝑏, 𝑗) if 𝑖 < 𝑗
or if 𝑖 = 𝑗 and 𝑎 < 𝑏.
(2) The reverse lexicographic product of 𝑋 and 𝑌 is defined by en-
dowing the cartesian product 𝑋 × 𝑌 with the order: (𝑥, 𝑦) <
(𝑥′ , 𝑦′ ) if 𝑦 < 𝑦′ or if 𝑦 = 𝑦′ and 𝑥 < 𝑥′ . It is still denoted by
𝑋 × 𝑌.
By an isomorphism between two partially ordered sets 𝑋 and 𝑌 we

mean a bijection 𝑓 between 𝑋 and 𝑌 such that for any 𝑥, 𝑥′ ∈ 𝑋 one has
𝑥 < 𝑥′ if and only if 𝑓(𝑥) < 𝑓(𝑥′ ).
Lemma 1.4.1.
(1) The ordered sum of total orders (resp. well-founded partial or-
ders) is a total order (resp. well-founded).
(2) The reverse lexicographic product of two total orders (resp. well-
founded partial orders) is a total order (resp. well-founded).
(3) Let 𝑋, 𝑌 and 𝑍 be partially ordered. We have the following canon-
ical isomorphisms of partially ordered sets:
(a) (𝑋 + 𝑌 ) + 𝑍 ≅ 𝑋 + (𝑌 + 𝑍).
(b) (𝑋 × 𝑌 ) × 𝑍 ≅ 𝑋 × (𝑌 × 𝑍).
(c) 𝑋 × (𝑌 + 𝑍) ≅ (𝑋 × 𝑌 ) + (𝑋 × 𝑍).
Proof. The only non-trivial point to check is that the reverse lexico-
graphic product of two well-founded partially ordered sets is well-founded.
1
By a finite set we mean a set into which ℕ does not inject.
Let 𝑋 and 𝑌 be two well-founded partially ordered sets. Let 𝑍 be a non-

empty subset of 𝑋 ×𝑌 . We denote by 𝜋 ∶ 𝑋 ×𝑌 → 𝑌 the projection on the
second factor. The order on 𝑌 being well-founded, there exists a minimal
element 𝑦0 in 𝜋(𝑍) ⊆ 𝑌 . Since the order on 𝑋 is well-founded, there is a
minimal element 𝑥0 in the (non-empty) set 𝑍𝑦0 = {𝑥 ∈ 𝑋 ∣ (𝑥, 𝑦0 ) ∈ 𝑍}.
It is clear that (𝑥0 , 𝑦0 ) is minimal in 𝑍. □
Definition. Let 𝑋 and 𝑌 be totally ordered sets. We assume that 𝑋 ad-

mits a smallest element 0. One defines the partially ordered set 𝑋 (𝑌 ) as
follows. As a set, it is the set of functions from 𝑌 to 𝑋 with finite sup-
port, that is, the subset of the set 𝑋 𝑌 of all functions 𝑌 → 𝑋 consisting
of functions 𝑓 ∶ 𝑌 → 𝑋 such that
supp(𝑓) ∶= {𝑦 ∈ 𝑌 ∣ 𝑓(𝑦) ≠ 0}
is finite. One sets 𝑓 < 𝑔 if there exists 𝑦 ∈ 𝑌 such that 𝑓(𝑦) < 𝑔(𝑦) and
𝑓(𝑦′ ) = 𝑔(𝑦′ ) for every 𝑦′ > 𝑦.
Proposition 1.4.2. Let 𝑋, 𝑌 and 𝑍 be totally ordered sets, and assume

that 𝑋 admits a smallest element 0.
(1) The relation < defines a total order on 𝑋 (𝑌 ) which is well-founded
when the orders on 𝑋 and 𝑌 are both well-founded.
(2) There are canonical isomorphisms of totally ordered sets 𝑋 (𝑌+𝑍) ≅
(𝑍)
𝑋 (𝑌 ) × 𝑋 (𝑍) and 𝑋 (𝑌 ×𝑍) ≅ (𝑋 (𝑌 ) ) .
Proof. The only non-trivial point to check is that if 𝑋 and 𝑌 are well-
ordered, then 𝑋 (𝑌 ) is well-founded. Let 𝑍 be a non-empty subset of 𝑋 (𝑌 ) .
Let us prove that 𝑍 contains a smallest element. If the constant function
with value 0 belongs to 𝑍, there is nothing to prove. Hence, we may
assume supp(𝑓) ≠ ∅ for every 𝑓 ∈ 𝑍. Let
𝑌1 = {𝑠1 (𝑓) ∣ 𝑓 ∈ 𝑍},
where 𝑠1 (𝑓) = max(supp(𝑓)). Let 𝑦1 be the smallest element of 𝑌1 , and
set 𝑍1′ = {𝑓 ∈ 𝑍 ∣ 𝑠1 (𝑓) = 𝑦1 }. The set 𝑍1′ is an initial segment of 𝑍, in
other words 𝑓 < 𝑔 for every 𝑓 ∈ 𝑍1′ and 𝑔 ∈ 𝑍 ⧵𝑍1′ . Let 𝑥1 be the smallest
element of {𝑓(𝑦1 ) ∣ 𝑓 ∈ 𝑍1′ }. We set
𝑍1 = {𝑓 ∈ 𝑍1′ ∣ 𝑓(𝑦1 ) = 𝑥1 }.
1.5. Ordinal Numbers 7
The set 𝑍1 is an initial segment of 𝑍1′ . If 𝑍1 contains the function with

constant value 0 outside {𝑦1 }, we are done. Otherwise, we have supp(𝑓)⧵
{𝑦1 } ≠ ∅ for every 𝑓 ∈ 𝑍1 . Let 𝑌2 = {𝑠2 (𝑓) ∣ 𝑓 ∈ 𝑍1 }, where 𝑠2 (𝑓) =
max(supp(𝑓) ⧵ {𝑦1 }). Let 𝑦2 be the smallest element of 𝑌2 , and 𝑥2 the
smallest element of {𝑓(𝑦2 ) ∣ 𝑓 ∈ 𝑍1 and 𝑦2 = 𝑠2 (𝑓)}. We set 𝑍2 =
{𝑓 ∈ 𝑍1 ∣ 𝑠2 (𝑓) = 𝑦2 and 𝑓(𝑦2 ) = 𝑥2 }. It is an initial segment of 𝑍1 . If 𝑍2
contains the function with constant value 0 outside {𝑦1 , 𝑦2 }, we are done,
otherwise one continues in the same way, constructing 𝑌3 , 𝑦3 , 𝑍3′ , 𝑥3 , 𝑍3
and so on. Since the sequence (𝑦𝑖 ) is strictly decreasing in 𝑌 , this process
stops after a finite number of steps. □
1.5. Ordinal Numbers

A set 𝑋 is said to be transitive if for all 𝑥 ∈ 𝑋 and 𝑦 ∈ 𝑥 one has 𝑦 ∈ 𝑋.
This is equivalent to 𝑥 ∈ 𝑋 ⇒ 𝑥 ⊆ 𝑋.
Definition. A set 𝑋 is an ordinal if it is transitive and if the relation
{(𝑥, 𝑦) ∈ 𝑋 × 𝑋 ∣ 𝑥 ∈ 𝑦} on 𝑋 defines a well-order on 𝑋.
Proposition 1.5.1. Let 𝛼 and 𝛽 be ordinals.
(1) ∅ is an ordinal.
(2) If 𝛼 ≠ ∅, then ∅ ∈ 𝛼.
(3) 𝛼 ∉ 𝛼.
(4) If 𝑥 ∈ 𝛼, then 𝑥 = 𝑆<𝑥 ∶= {𝑦 ∈ 𝛼 ∣ 𝑦 < 𝑥}.
(5) If 𝑥 ∈ 𝛼, then 𝑥 is an ordinal.
(6) 𝛽 ⊆ 𝛼 if and only if 𝛽 ∈ 𝛼 or 𝛽 = 𝛼.
(7) 𝑥 ∶= 𝛼 ∪ {𝛼} is an ordinal, denoted by 𝛼+ .
Proof. (1) is clear. For (2), one considers 𝑥 ∈ 𝛼 minimal. If 𝑦 ∈ 𝑥,

then 𝑦 ∈ 𝛼 by transitivity of 𝛼, and 𝑥 would not be minimal. In (3), by
antireflexivity, we have 𝑥 ∉ 𝑥 for every 𝑥 ∈ 𝛼. Thus 𝛼 ∈ 𝛼 implies
𝛼 ∉ 𝛼. (4) follows from the fact that < is given by ∈. To prove (5), note
that ∈ restricts to a well-order on 𝑥, since 𝑥 ⊆ 𝛼. Furthermore, 𝑥 = 𝑆<𝑥
is transitive, since 𝑧 ∈ 𝑦 ∈ 𝑥 ⇒ 𝑧 < 𝑥 ⇒ 𝑧 ∈ 𝑆<𝑥 .
To prove the ‘only if’ part in (6), let us assume that 𝛽 ⊂ 𝛼. Let 𝑥
be minimal in 𝛼 ⧵ 𝛽. Clearly 𝛽 ⊇ 𝑆<𝑥 by minimality. Furthermore, if
𝑦 ∈ 𝛽, then 𝑦 ∈ 𝑥 since otherwise 𝑥 ∈ 𝑦 and 𝑥 ∈ 𝛽. Hence 𝛽 = 𝑆<𝑥 =

𝑥 ∈ 𝛼. The other implication in (6) is clear, and the verification of (7) is
immediate. □
Proposition 1.5.2. Let 𝑋 be a non-empty set of ordinals. Then ⋂𝛼∈𝑋 𝛼 is
a smallest element of 𝑋.
Proof. The intersection of a family of transitive sets is transitive, and the

restriction of a well-order to a subset is a well-order. Hence 𝛽 = ⋂𝛼∈𝑋 𝛼
is an ordinal. We have 𝛽 ⊆ 𝛼 for every 𝛼 ∈ 𝑋. If 𝛽 ∉ 𝑋, then 𝛽 ∈ 𝛼
for every 𝛼 ∈ 𝑋, by Proposition 1.5.1(6). It follows that 𝛽 ∈ 𝛽, which is
absurd. □
Theorem 1.5.3. Let 𝛼 and 𝛽 be ordinals. Exactly one of the following
properties holds:
(1) 𝛼 ∈ 𝛽, (2) 𝛼 = 𝛽, (3) 𝛽 ∈ 𝛼.
Proof. One sets 𝑋 = {𝛼, 𝛽}, and one applies Proposition 1.5.2. If 𝛼 ∩ 𝛽 =
𝛼, then 𝛼 ⊆ 𝛽, hence 𝛼 = 𝛽 or 𝛼 ∈ 𝛽 by Proposition 1.5.1(6). Similarly,
if 𝛼 ∩ 𝛽 = 𝛽, then 𝛼 = 𝛽 or 𝛽 ∈ 𝛼. The fact that these properties are
mutually exclusive follows from the axioms of a partial order. □
Notation. From now on, we shall write 𝛼 < 𝛽 for 𝛼 ∈ 𝛽, and 𝛼 ≤ 𝛽 for
𝛼 ⊆ 𝛽, when 𝛼 and 𝛽 are ordinals.
Proposition 1.5.4. Let 𝑋 be a set of ordinals. Then 𝑏 = ⋃𝛼∈𝑋 𝛼 is an
ordinal. Furthermore, if 𝛾 is an ordinal with 𝛾 < 𝑏, there exists 𝛼 ∈ 𝑋
such that 𝛾 ∈ 𝛼. We shall also write 𝑏 = sup𝛼∈𝑋 𝛼.
Proof. The set 𝑏 being the union of transitive sets, it is transitive. Fur-
thermore, 𝑏 contains only ordinals. By Theorem 1.5.3, ∈ induces a total
order on 𝑏. If ∅ ≠ 𝑍 ⊆ 𝑏, then ⋂𝛼∈𝑍 𝛼 is a smallest element of 𝑍 by
Proposition 1.5.2. This shows that the order given by ∈ on 𝑏 is well-
founded. □
An ordinal of the form 𝛼+ is called a successor ordinal. It is clear

that 𝛼+ is the smallest ordinal > 𝛼.
Definition. A limit ordinal is a non-empty ordinal which is not a suc-
cessor.
1.5. Ordinal Numbers 9
Proposition 1.5.5. For an ordinal 𝜆 ≠ ∅, the following conditions are

equivalent:
(1) 𝜆 is a limit ordinal;
(2) 𝜆 = ⋃𝛼<𝜆 𝛼.
Proof. (1)⇒(2). Let 𝛽 = ⋃𝛼<𝜆 𝛼 and 𝜆 a limit. It is clear that 𝛽 ⊆ 𝜆.

Conversely, assume 𝛼 < 𝜆. Then 𝛼+ ≤ 𝜆 and it follows that 𝛼+ < 𝜆 since
𝜆 is a limit ordinal. The statement follows, since 𝛼 ∈ 𝛼+ ⊆ 𝛽.
(2)⇒(1). If 𝜆 = 𝛾+ , then ⋃𝛼<𝜆 𝛼 = ⋃𝛼≤𝛾 𝛼 = 𝛾 < 𝜆. □
Example 1.5.6.
(1) One can recover the natural numbers as ordinals as follows.
One sets 0 ∶= ∅, and inductively 𝑛 + 1 ∶= 𝑛+ for 𝑛 ∈ ℕ.
For instance 1 = {∅}, 2 = {0, 1} = {∅, {∅}}, 3 = {0, 1, 2} =
{∅, {∅}, {∅, {∅}}}.
One proves by induction that 𝑛 is an ordinal for every nat-
ural number 𝑛. We shall often identify 𝑛 and 𝑛.
(2) One sets 𝜔 ∶= ⋃𝑛∈ℕ 𝑛. It is an ordinal by Proposition 1.5.4.
Definition. One says that an ordinal is finite if it is not a limit and none
of its elements is a limit.
Proposition 1.5.7.
(1) 𝜔 is the set of finite ordinals.
(2) 𝜔 is the smallest limit ordinal.
Proof. One proves first, by induction on 𝑛 ∈ ℕ, that all elements of 𝜔

are finite ordinals. Furthermore, 𝛼 < 𝜔 implies 𝛼+ < 𝜔. This proves (2).
If 𝛼 ∉ 𝜔, then 𝜔 ≤ 𝛼, so either 𝛼 = 𝜔 or 𝜔 ∈ 𝛼. In both cases, 𝛼 is not
finite. This proves (1). □
Lemma 1.5.8. Let 𝑓 ∶ 𝛼 → 𝛼′ be a strictly increasing map between two
ordinals. Then 𝑓(𝛽) ≥ 𝛽 for every 𝛽 ∈ 𝛼. In particular, 𝛼 ≤ 𝛼′ , and if 𝑓 is
an isomorphism of ordered sets, then 𝛼 = 𝛼′ and 𝑓 is equal to the identity.
Proof. If there exists 𝛽 ∈ 𝛼 with 𝑓(𝛽) < 𝛽, we consider 𝛽0 minimal with

that property. Since 𝑓 is strictly increasing, we have 𝑓(𝑓(𝛽0 )) < 𝑓(𝛽0 ),
which contradicts minimality.
The statement about an isomorphism 𝑓 follows by applying the re-

sult to 𝑓 as well as to 𝑓−1 . □
Theorem 1.5.9 (Classification of well-orders by ordinals). Every well-
ordered set 𝑋 is isomorphic, as an ordered set, to some ordinal. Further-
more, the ordinal and the isomorphism are both unique.
Proof. Uniqueness follows from Lemma 1.5.8. To prove existence, let

us first note that for every 𝑥 ∈ 𝑋, any isomorphism between 𝑆<𝑥 and an
ordinal 𝛼 can be extended to an isomorphism between 𝑆≤𝑥 = 𝑆<𝑥 ∪ {𝑥}
and 𝛼+ . Let
𝑌 = {𝑦 ∈ 𝑋 ∣ there exists 𝑓 ∶ 𝑆≤𝑦 ≅ 𝛼 for some ordinal 𝛼}.
By uniqueness, for 𝑦 ∈ 𝑌 , the ordinal 𝛼 = 𝛼(𝑦) and the isomorphism
𝑓 = 𝑓𝑦 are unique. Let us prove 𝑌 = 𝑋. Otherwise, there would exist
𝑥 ∈ 𝑋 minimal in 𝑋 ⧵ 𝑌 . For 𝑦 < 𝑥 we have an isomorphism 𝑓𝑦 ∶ 𝑆≤𝑦 ≅
𝛼(𝑦). Furthermore, these isomorphisms form a coherent family in the
sense that for every 𝑦′ < 𝑦 < 𝑥 we have 𝑓𝑦 ↾ 𝑆≤𝑦′ = 𝑓𝑦′ . (To see this, note
that an initial segment of an ordinal is an ordinal.) We set
𝛼 = sup 𝛼(𝑦) and 𝑓 ∶ 𝑆<𝑥 → 𝛼, 𝑓(𝑦) ∶= 𝑓𝑦 (𝑦).
𝑦<𝑥
It is clear that 𝑓 is well defined and induces an isomorphism of ordered

sets between 𝑆<𝑥 and 𝛼. By the observation made at the beginning, 𝑓
may be extended to an isomorphism between 𝑆≤𝑥 and 𝛼+ , which leads
to a contradiction. So we have 𝑌 = 𝑋. To conclude, one uses the same
kind of argument, setting 𝛼(𝑋) ∶= sup𝑥∈𝑋 𝛼(𝑥) and 𝑓 ∶ 𝑋 ≅ 𝛼(𝑋),
𝑥 ↦ 𝑓𝑥 (𝑥). □
Remark 1.5.10 (Transfinite induction). Let 𝑃 be a property of ordinals.
One assumes:
• ∅ satisfies 𝑃;
• for every ordinal 𝛼: if 𝛼 satisfies 𝑃, then 𝛼+ satisfies 𝑃;
• for every limit ordinal 𝜆: if every 𝛼 < 𝜆 satisfies 𝑃, then 𝜆 sat-
isfies 𝑃.
Then every ordinal satisfies 𝑃.
1.6. Ordinal Arithmetic 11
1.6. Ordinal Arithmetic

If 𝛼 and 𝛽 are ordinals, by Theorem 1.5.9 there is a unique ordinal iso-
morphic to the ordered sum of 𝛼 and 𝛽, which one denotes by 𝛼+𝛽. One
similarly defines 𝛼𝛽 as the unique ordinal isomorphic to the reverse lex-
icographic product 𝛼 × 𝛽 and 𝛼𝛽 as the unique ordinal isomorphic to the
ordered set 𝛼(𝛽) . Note that 0𝛽 has still to be defined: one sets 00 ∶= 1
and 0𝛽 ∶= 0 for every 𝛽 > 0.
Proposition 1.6.1 (Ordinal addition). Let 𝛼, 𝛽 and 𝛾 be ordinals.

(1) 𝛼 + 0 = 0 + 𝛼 = 𝛼.
(2) 𝛼 + 1 = 𝛼+ .
(3) 𝛼 + (𝛽 + 𝛾) = (𝛼 + 𝛽) + 𝛾, in particular 𝛼 + 𝛽 + = (𝛼 + 𝛽)+ .
(4) 𝛼 < 𝛽 if and only if there exists an ordinal 𝛿 > 0 such that
𝛽 = 𝛼 + 𝛿.
(5) If 𝛽 < 𝛾, then 𝛼 + 𝛽 < 𝛼 + 𝛾 for every 𝛼. In particular, one may
simplify on the left: 𝛼 + 𝛽 = 𝛼 + 𝛾 ⇒ 𝛽 = 𝛾.
(6) If 𝜆 is a limit, then 𝛼 + 𝜆 = sup𝛽<𝜆 (𝛼 + 𝛽) (continuity).
(7) 1 + 𝛼 = 𝛼 + 1 when 𝛼 is finite, otherwise 1 + 𝛼 = 𝛼.
Proof. (1) and (2) are clear, and (3) follows from Lemma 1.4.1. For the
non-trivial implication in (4), one easily checks that the ordinal 𝛿 iso-
morphic to the well-ordered set 𝛽 ⧵ 𝛼 does the job.
(5) If 𝛽 < 𝛾, by (2) and (4) one has 𝛾 = 𝛽+𝛿, hence 𝛼+𝛾 = (𝛼+𝛽)+𝛿,
for some 𝛿 > 0.
(6) 𝛼 + 𝜆 ≥ sup𝛽<𝜆 (𝛼 + 𝛽) follows from (5). Conversely, suppose
𝛼 ≤ 𝜇 < 𝛼 + 𝜆. Then 𝜇 = 𝛼 + 𝛿 for some 𝛿 with 0 ≤ 𝛿 < 𝜆. Since 𝜆 is a
limit, one has 𝛿+ < 𝜆, hence 𝜇 < 𝛼 + 𝛿+ ≤ sup𝛽<𝜆 (𝛼 + 𝛽).
(7) One proves by induction on 𝑛 ∈ ℕ that 1 + 𝑛 = 𝑛 + 1. By (6)
we have 1 + 𝜔 = 𝜔. Finally, 𝛼 ≥ 𝜔 can be written as 𝛼 = 𝜔 + 𝛽, hence
1 + 𝛼 = 1 + 𝜔 + 𝛽 = 𝜔 + 𝛽 = 𝛼. □
From now on, we shall allow the omission of parentheses, using the
convention that exponentiation ties are stronger than multiplication and
that multiplication ties are stronger than addition. For instance, one
should read 𝛼𝛽 + 𝛾 as (𝛼𝛽) + 𝛾, and 𝛾𝛼𝛽 as 𝛾 (𝛼𝛽 ).
Proposition 1.6.2 (Ordinal multiplication). Let 𝛼, 𝛽, 𝛾 be ordinals.
(1) 𝛼0 = 0𝛼 = 0.
(2) 𝛼1 = 1𝛼 = 𝛼.
(3) 𝛼(𝛽𝛾) = (𝛼𝛽)𝛾.
(4) 𝛼(𝛽 + 𝛾) = 𝛼𝛽 + 𝛼𝛾, in particular 𝛼𝛽 + = 𝛼𝛽 + 𝛼.
(5) 2𝜔 = 𝜔 < 𝜔2 = 𝜔 + 𝜔.
(6) Assume 𝛼 ≠ 0. If 𝛽 < 𝛾, then 𝛼𝛽 < 𝛼𝛾. In particular, one may
simplify on the left: 𝛼𝛽 = 𝛼𝛾 ⇒ 𝛽 = 𝛾.
(7) If 𝜆 is a limit ordinal, then 𝛼𝜆 = sup𝛽<𝜆 𝛼𝛽 (continuity).
Proof. (1) and (2) are clear, (3) and (4) follow from Lemma 1.4.1. For
(6), it suffices to note that if 𝛽 < 𝛾 then 𝛾 = 𝛽 + 𝛿 for some 𝛿 > 0, hence
𝛼𝛾 = 𝛼𝛽 + 𝛼𝛿 by (4) from which it follows that 𝛼𝛾 > 𝛼𝛽.
(7) One may assume 𝛼 ≠ 0. Let 𝜆 be a limit ordinal. The inequality
𝛼𝜆 ≥ sup𝛽<𝜆 𝛼𝛽 =∶ 𝛿 follows from (6). Conversely, let 𝛾 < 𝛼𝜆. Eu-
clidean division, proved in the next lemma, provides a pair of ordinals
(𝜌, 𝜇) such that 𝛾 = 𝛼𝜇 + 𝜌, with 𝜌 < 𝛼. Since 𝜇 < 𝜆 by (6), we have
𝜇+ < 𝜆 because 𝜆 is a limit ordinal, hence 𝛾 = 𝛼𝜇 + 𝜌 < 𝛼𝜇 + 𝛼 = 𝛼𝜇+ ≤
𝛿. In (5), 2𝜔 = 𝜔 follows from (7), the other statements being clear. □
Lemma 1.6.3 (Euclidean division). Let 𝛼 and 𝛽 be ordinals, with 𝛼 ≠ 0.
Then there exists a unique pair of ordinals (𝜌, 𝜇) such that 𝜌 < 𝛼 and
𝛽 = 𝛼𝜇 + 𝜌.
Proof. Uniqueness: Assume 𝛼𝜇 + 𝜌 = 𝛼𝜇′ + 𝜌′ with 𝜌, 𝜌′ < 𝛼. If 𝜇 < 𝜇′ ,

then 𝛼𝜇 + 𝜌 < 𝛼𝜇+ ≤ 𝛼𝜇′ ≤ 𝛼𝜇′ + 𝜌′ , which is absurd. Hence 𝜇 = 𝜇′ by
symmetry, and one obtains 𝜌 = 𝜌′ after simplifying.
Existence: When 𝛽 = 0 there is nothing to prove. Assume 𝛽 ≠ 0.
The mapping 𝑓0 ∶ 𝛽 → 𝛼 × 𝛽, 𝑥 ↦ (0, 𝑥) is strictly increasing, hence 𝛽 ≤
𝛼𝛽 by Lemma 1.5.8. If 𝛽 = 𝛼𝛽, one sets 𝜇 = 𝛽 and 𝜌 = 0. Otherwise, we
have 𝛽 ∈ 𝛼𝛽. Let 𝑓 be the unique isomorphism of ordered sets between
𝛼𝛽 and 𝛼×𝛽. One sets (𝜌, 𝜇) = 𝑓(𝛽). Since 𝑆<(𝜌,𝜇) ≅ (𝛼×𝜇)+𝜌 it follows
that 𝛽 = 𝛼𝜇 + 𝜌. □
1.6. Ordinal Arithmetic 13
Note that we have only used properties (1)–(4) and (6) from Propo-
sition 1.6.2 in our proof of Euclidean division, thus avoiding circularity.
Proposition 1.6.4 (Ordinal exponentiation). Let 𝛼, 𝛽, 𝛾 be ordinals.
(1) For every 𝛼, we have 𝛼0 = 1, 𝛼1 = 𝛼 and 1𝛼 = 1. If 𝛼 ≠ 0, then

0𝛼 = 0.
+
(2) 𝛼𝛽+𝛾 = 𝛼𝛽 𝛼𝛾 , in particular 𝛼𝛽 = 𝛼𝛽 𝛼.
𝛾
(3) (𝛼𝛽 ) = 𝛼𝛽𝛾 .
(4) If 𝛼 > 1 and 𝛽 < 𝛾, then 𝛼𝛽 < 𝛼𝛾 .
(5) If 𝜆 is a limit ordinal and 𝛼 ≠ 0, then 𝛼𝜆 = sup𝛽<𝜆 𝛼𝛽 (continu-
ity).
Proof. (1) is checked directly, and statements (2) and (3) follow from
Proposition 1.4.2.
(4) 𝛽 < 𝛾 ⇒ 𝛾 = 𝛽 + 𝛿 for some 𝛿 > 0. Hence 𝛼𝛾 = 𝛼𝛽+𝛿 = 𝛼𝛽 𝛼𝛿 .
But 𝛼𝛿 > 1 since as a set 𝛼(𝛿) contains at least two elements. It follows
that 𝛼𝛾 > 𝛼𝛽 by Proposition 1.6.2(6).
Let us prove the non-trivial inequality in (5). Let 𝑓 ∈ 𝛼(𝜆) . One may
assume 𝑓 is not the constant function with value 0. Then 𝑠1 (𝑓) < 𝜆, and
hence 𝛽 = 𝑠1 (𝑓)+ < 𝜆, which proves there exists a strictly increasing
function 𝑆≤𝑓 → 𝛼(𝛽) . One concludes by Lemma 1.5.8. □
Remark 1.6.5. The following formulas would allow us to define ordinal

addition, multiplication and exponentiation by transfinite induction:
• 𝛼 + 0 = 𝛼, 𝛼 + 𝛽 + = (𝛼 + 𝛽)+ , and 𝛼 + 𝜆 = sup𝛽<𝜆 (𝛼 + 𝛽) for

𝜆 a limit ordinal.
• 𝛼0 = 0, 𝛼𝛽+ = 𝛼𝛽 + 𝛼, and 𝛼𝜆 = sup𝛽<𝜆 (𝛼𝛽) for 𝜆 a limit
ordinal.
+
• Assume 𝛼 ≠ 0. Then one has 𝛼0 = 1, 𝛼𝛽 = 𝛼𝛽 𝛼, and 𝛼𝜆 =
sup𝛽<𝜆 (𝛼𝛽 ) for 𝜆 a limit ordinal.
1.7. The Axiom of Choice

Given a family of sets (𝑋𝑖 )𝑖∈𝐼 , one defines their product as
|
∏ 𝑋𝑖 = {𝑓 ∶ 𝐼 → 𝑋𝑖 || 𝑓(𝑖) ∈ 𝑋𝑖 for all 𝑖 ∈ 𝐼} .
⋃ |
𝑖∈𝐼 𝑖∈𝐼
Definition. The Axiom of Choice (AC) states that the product of a family
of non-empty sets is non-empty: if 𝑋𝑖 ≠ ∅ for all 𝑖 ∈ 𝐼, then ∏𝑖∈𝐼 𝑋𝑖 ≠ ∅.
In the Zermelo-Fraenkel system of axioms ZF, (AC) is equivalent

to Zorn’s Lemma and also to Zermelo’s Theorem. We shall prove these
equivalences in the last chapter of this book, and accept them for the
moment.
Definition. A partially ordered set 𝑋 is inductive if any totally ordered
subset 𝑌 ⊆ 𝑋 admits an upper bound in 𝑋. (In particular, such an 𝑋 is
non-empty).
Zorn’s Lemma. Every inductive partially ordered set admits a maximal
element.
Zermelo’s Theorem (Wohlordnungssatz). Every set can be well-ordered.
1.8. Cardinal Numbers

We now assume, until the end of the penultimate chapter, that the Ax-
iom of Choice holds.
Definition. An ordinal is a cardinal if it is not equinumerous to a smaller
ordinal.
Example 1.8.1.
(1) Any finite ordinal is a cardinal.
(2) The ordinal 𝜔 is a cardinal. When considered as a cardinal it
will be denoted by ℵ0 .
(3) If 𝛼 is an infinite ordinal, then 𝛼+ is not a cardinal. (Indeed,
𝛼+ and 𝛼 are equinumerous.)
Proposition 1.8.2. Any set 𝑋 is equinumerous to a unique cardinal, de-
noted by card(𝑋).
1.8. Cardinal Numbers 15
Proof. By Zermelo’s Theorem and Theorem 1.5.9, 𝑋 is equinumerous

to an ordinal 𝛼. Let 𝛽 ≤ 𝛼 be minimal such that 𝛽 is equinumerous to 𝛼.
Then 𝛽 is a cardinal and is in bijection with 𝑋. Uniqueness is clear. □
Proposition 1.8.3. Let 𝑋 and 𝑌 be sets and assume that 𝑋 is non-empty.
The following statements are equivalent:
(1) card(𝑋) ≤ card(𝑌 ).
(2) There exists an injective map 𝑋 → 𝑌 .
(3) There exists a surjective map 𝑌 → 𝑋.
Proof. (1)⇒(2) is easy.

(2)⇒(3): Let 𝑓 ∶ 𝑋 → 𝑌 be an injective map. As 𝑋 is non-empty,
one may fix 𝑥0 ∈ 𝑋. One defines a surjective map 𝑔 ∶ 𝑌 → 𝑋 by set-
ting 𝑔(𝑦) ∶= 𝑥0 if 𝑦 ∉ im(𝑓) = {𝑓(𝑥) ∣ 𝑥 ∈ 𝑋}, and 𝑔(𝑦) ∶= 𝑓−1 (𝑦)
otherwise.
(3)⇒(1): If there exists a surjective map 𝑌 → 𝑋, then there exists a
surjection 𝑔 ∶ 𝜆 = card(𝑌 ) → 𝜅 = card(𝑋). The map 𝑓 sending 𝛼 ∈ 𝜅
to the minimal 𝛽 ∈ 𝜆 such that 𝑔(𝛽) = 𝛼 provides an injection 𝜅 → 𝜆.
In particular, 𝜅 is in bijection with some ordinal 𝛾 ≤ 𝜆. (One takes 𝛾
as the unique ordinal which is isomorphic to the well-order induced on
im(𝑓).) □
Definition. A set 𝑋 is said to be countable if card(𝑋) ≤ ℵ0 , and finite if
card(𝑋) < ℵ0 .
Proposition 1.8.4. Let 𝑋 be a set of cardinals. Then 𝜆 = sup𝜅∈𝑋 𝜅 is a
cardinal.
Proof. If 𝛼 < 𝜆, then 𝛼 < 𝜅 for some 𝜅 ∈ 𝑋. Since 𝜅 is a cardinal, we

have 𝜅 = card(𝜅) ≤ card(𝜆), and hence 𝛼 < card(𝜆). This proves that 𝜆
is not equinumerous to some smaller ordinal. □
Notation. From now on, 𝜅, 𝜆, etc. will denote cardinals.
There is no largest cardinal. Indeed, if 𝜅 is a cardinal, then 𝜆 ∶=

card(𝒫(𝜅)) > 𝜅 by Cantor’s Theorem. In particular, the set of all cardi-
nals ≤ 𝜆 that are > 𝜅 is non-empty. We denote by 𝜅+ its smallest element,
called the cardinal successor of 𝜅. To avoid confusion, from now on the
ordinal successor of 𝛼 will be denoted by 𝛼 + 1.
Definition. The ℵ-hierarchy assigns to any ordinal a cardinal as fol-

lows:
• ℵ0 ∶= 𝜔.
• ℵ𝛼+1 ∶= ℵ+
𝛼.
• ℵ𝛼 ∶= sup𝛽<𝛼 ℵ𝛽 , if 𝛼 is a limit ordinal.
By transfinite induction, one proves that 𝛼 < 𝛽 ⇒ ℵ𝛼 < ℵ𝛽 . In com-

bination with the next result, it follows that the ℵ-hierarchy provides a
strictly increasing enumeration of the infinite cardinals by the ordinals.
Proposition 1.8.5. Every infinite cardinal is of the form ℵ𝛼 for some 𝛼.
Proof. Let 𝜅 be an infinite cardinal. The function 𝛽 ↦ ℵ𝛽 is strictly

increasing on 𝜅 + 1, and it takes its values in ℵ𝜅+1 . Thus ℵ𝜅 ≥ 𝜅 by
Lemma 1.5.8, and hence ℵ𝜅+1 > 𝜅. Let 𝛼 ≤ 𝜅 + 1 be minimal with
ℵ𝛼 > 𝜅. Since 𝜅 ≥ ℵ0 , we have 𝛼 > 0. If 𝛼 were a limit ordinal, by
definition we would have 𝜅 ∈ ⋃𝛽<𝛼 ℵ𝛽 , and hence 𝜅 ∈ ℵ𝛽 for some
𝛽 < 𝛼, which would contradict the minimality of 𝛼. Thus 𝛼 = 𝛽 + 1 and
also ℵ𝛽 ≤ 𝜅 < ℵ𝛽+1 = ℵ+ +
𝛽 . Since ℵ𝛽 is the cardinal successor of ℵ𝛽 ,
necessarily ℵ𝛽 = 𝜅. □
1.9. Operations on Cardinals

If 𝑋 and 𝑌 are sets, one denotes by 𝑋 + 𝑌 their disjoint union, by 𝑋 × 𝑌
their cartesian product and by 𝑋 𝑌 the set of maps from 𝑌 to 𝑋. If 𝜅 and
𝜆 are cardinals, one denotes by 𝜅 + 𝜆 the cardinal of their disjoint union,
by 𝜅𝜆 the cardinal of their cartesian product and by 𝜅𝜆 the cardinal of the
set of maps from 𝜆 to 𝜅. These operations are respectively called cardinal
addition, cardinal multiplication and cardinal exponentiation.
They should not be confused with the corresponding ordinal oper-
ations. For instance 2𝜔 = 𝜔 = ℵ0 < 2ℵ0 ; also ℵ0 2 = ℵ0 , but 𝜔 < 𝜔2. It
is clear that on finite cardinals all of these operations correspond to the
usual arithmetic operations. Note that card(𝑋 +𝑌 ) = card(𝑋)+card(𝑌 ),
card(𝑋 × 𝑌 ) = card(𝑋)card(𝑌 ) and card(𝑋 𝑌 ) = card(𝑋)card(𝑌 ) .
The proof of the following statements is immediate, using Proposi-
tion 1.8.3.
1.9. Operations on Cardinals 17
Proposition 1.9.1. Let 𝜅, 𝜆 and 𝜇 be cardinals.

(1) Cardinal addition and multiplication are commutative and as-
sociative, multiplication is distributive with respect to addition,
𝜅𝜆+𝜇 = 𝜅𝜆 𝜅𝜇 , (𝜅𝜆 )𝜇 = 𝜅𝜆𝜇 and (𝜅𝜆)𝜇 = 𝜅𝜇 𝜆𝜇 .
(2) If 𝜅 ≤ 𝜆, then 𝜅 + 𝜇 ≤ 𝜆 + 𝜇, 𝜅𝜇 ≤ 𝜆𝜇 and 𝜅𝜇 ≤ 𝜆𝜇 (when 𝜅 > 0)
and 𝜇𝜅 ≤ 𝜇𝜆 (when 𝜇 > 0). □
Proposition 1.9.2. One has card(ℝ) = 2ℵ0 .
Proof. There is an injection ℎ ∶ 2ℵ0 → ℝ sending a sequence (𝑎𝑖 )𝑖∈ℕ

to the sum ∑𝑖 𝑎𝑖 2−𝑖 if the support of the sequence is infinite, and to 2 +
∑𝑖 𝑎𝑖 2−𝑖 otherwise. This proves that 2ℵ0 ≤ card(ℝ). On the other hand,
the image of ℎ contains the interval (0, 1) which is equinumerous to ℝ
(for instance via 𝑥 ↦ 1/𝜋 arctan(𝑥) + 1/2); hence card(ℝ) ≤ 2ℵ0 . □
Proposition 1.9.3 (Hessenberg’s Theorem). For every infinite cardinal

𝜅, one has 𝜅𝜅 = 𝜅.
Proof. By induction on 𝛼, we will prove that ℵ𝛼 ℵ𝛼 = ℵ𝛼 .

For 𝛼 = 0 this is clear. Indeed, the mapping 𝛼2 ∶ ℕ2 → ℕ defined
by 𝛼2 (𝑚, 𝑛) ∶= 1/2(𝑚 + 𝑛 + 1)(𝑚 + 𝑛) + 𝑛 is bijective.
Let us now assume ℵ𝛽 ℵ𝛽 = ℵ𝛽 for every 𝛽 < 𝛼. One endows ℵ𝛼 ×
ℵ𝛼 with the following order:
(𝛽, 𝛾) < (𝛽 ′ , 𝛾′ ) if max(𝛽, 𝛾) < max(𝛽 ′ , 𝛾′ ), or
if max(𝛽, 𝛾) = max(𝛽 ′ , 𝛾′ ) and 𝛽 < 𝛽 ′ , or
if max(𝛽, 𝛾) = max(𝛽 ′ , 𝛾′ ), 𝛽 = 𝛽 ′ and 𝛾 < 𝛾′ .
One checks easily that this is a well-order. Furthermore, for every 𝛿 <
ℵ𝛼 , the set 𝛿 × 𝛿 is an initial segment for <. By Theorem 1.5.9, there is a
unique isomorphism of ordered sets 𝑓 ∶ 𝜀 → ℵ𝛼 × ℵ𝛼 with 𝜀 an ordinal.
Assume 𝜀 > ℵ𝛼 . Then ℵ𝛼 ∈ 𝜀 and 𝑓(ℵ𝛼 ) = (𝛽0 , 𝛾0 ) ∈ ℵ𝛼 × ℵ𝛼 . Set
𝛿0 ∶= max(𝛽0 , 𝛾0 ) + 1. Since no infinite successor ordinal is a cardinal
(by Example 1.8.1), we have 𝛿0 < ℵ𝛼 and the restriction of 𝑓 to ℵ𝛼 is an
injective map from ℵ𝛼 to 𝛿0 × 𝛿0 , a set of cardinality
card(𝛿0 × 𝛿0 ) = card(𝛿0 ) ≤ 𝛿0 < ℵ𝛼
by the induction hypothesis. This is a contradiction, and thus one has

ℵ𝛼 ℵ𝛼 ≤ ℵ 𝛼 .
The inequality in the other direction is clear. □
Example 1.9.4. Let 𝒯 be the set of all open subsets of ℝ. Then card(𝒯) =
2ℵ0 .
Proof. The mapping assigning to a real number 𝑟 ∈ ℝ the open interval

(𝑟, +∞) defines an injection from ℝ to 𝒯, which proves that card(𝒯) ≥
2ℵ0 .
Conversely, note that every open subset of ℝ is a union of intervals
of the form (𝑞, 𝑞 + 𝑞′ ), with 𝑞 ∈ ℚ and 𝑞′ ∈ ℚ>0 . The mapping sending
𝑌 ⊆ ℚ × ℚ>0 to ⋃(𝑞,𝑞′ )∈𝑌 (𝑞, 𝑞 + 𝑞′ ) provides a surjection of 𝒫(ℚ × ℚ>0 )
to 𝒯. Since ℚ × ℚ>0 is countable, one deduces that 2ℵ0 ≥ card(𝒯). □
Proposition 1.9.5.
(1) Let 𝑋 and 𝑌 be non-empty sets and assume that at least one of
them is infinite. Then
card(𝑋 ∪ 𝑌 ) = card(𝑋 × 𝑌 ) = max(card(𝑋), card(𝑌 )).
(2) Let 𝜅 ≥ ℵ0 and 𝜆 > 0 be cardinals. Then 𝜅 + 𝜆 = 𝜅𝜆 =
max(𝜅, 𝜆).
(3) Let (𝑋𝑖 )𝑖∈𝐼 be a family of sets with at least one 𝑋𝑖 infinite. Then
(∗) card ( 𝑋𝑖 ) ≤ sup ({card(𝑋𝑖 ) ∣ 𝑖 ∈ 𝐼} ∪ {card(𝐼)}) .

⋃
𝑖∈𝐼
(In particular, a countable union of countable sets is countable.)

If furthermore the sets 𝑋𝑖 are all non-empty and mutually dis-
joint, then equality holds in (∗).
Proof. (1) Let 𝜅 = max(card(𝑋), card(𝑌 )). We have

𝜅 ≤ card(𝑋 ∪ 𝑌 ) ≤ 𝜅 + 𝜅 = 2𝜅 ≤ 𝜅𝜅
and 𝜅 ≤ card(𝑋 × 𝑌 ) ≤ 𝜅𝜅. One concludes by Hessenberg’s Theorem.
(2) is a special case of (1).
1.10. Cofinality 19
(3) Let 𝑋 = {(𝑥𝑖 , 𝑖) ∣ 𝑥𝑖 ∈ 𝑋𝑖 for some 𝑖 ∈ 𝐼} be the disjoint union of

the sets 𝑋𝑖 . There is a canonical surjection 𝑋 → ⋃𝑖∈𝐼 𝑋𝑖 , hence it suffices
to prove that
card(𝑋) ≤ sup ({card(𝑋𝑖 ) ∣ 𝑖 ∈ 𝐼} ∪ {card(𝐼)}) .
Let 𝜅 = sup{card(𝑋𝑖 ) ∣ 𝑖 ∈ 𝐼}, and let 𝑌𝑖 be the set of injective maps
𝑋𝑖 → 𝜅. Since the sets 𝑌𝑖 are all non-empty, by the Axiom of Choice
there exists some 𝑓 = (𝑓𝑖 )𝑖∈𝐼 ∈ ∏𝑖∈𝐼 𝑌𝑖 . Consider 𝑔 ∶ 𝑋 → 𝜅 × 𝐼, defined
by 𝑔((𝑥𝑖 , 𝑖)) ∶= (𝑓𝑖 (𝑥𝑖 ), 𝑖). The function 𝑔 is injective, hence card(𝑋) ≤
𝜅 card(𝐼) = max(𝜅, card(𝐼)). The equality statement is clear. □
It follows from the preceding proposition that cardinal addition and

multiplication is quite trivial for infinite cardinals. The situation for car-
dinal exponentiation is very different. In fact, the ZFC axioms are far
from completely determining the values of cardinal exponentiation. For
instance they do not allow to settle the continuum hypothesis:
Definition.
• The Continuum Hypothesis (CH) is the statement 2ℵ0 = ℵ1 .
• The Generalized Continuum Hypothesis (GCH) is the state-
ment 2𝜅 = 𝜅+ for every infinite cardinal 𝜅.
If (𝜅𝑖 )𝑖∈𝐼 is a family of cardinals, we shall denote by ∑𝑖∈𝐼 𝜅𝑖 the car-

dinal of the disjoint union of the 𝜅𝑖 , and by ∏𝑖∈𝐼 𝜅𝑖 the cardinal of the
product of the family.
Theorem 1.9.6 (König’s Theorem). Let (𝜅𝑖 )𝑖∈𝐼 and (𝜆𝑖 )𝑖∈𝐼 be families of
cardinals with 𝜅𝑖 < 𝜆𝑖 for every 𝑖. Then ∑𝑖∈𝐼 𝜅𝑖 < ∏𝑖∈𝐼 𝜆𝑖 .
Proof. Let 𝑓 ∶ ∑𝑖∈𝐼 𝜅𝑖 → ∏𝑖∈𝐼 𝜆𝑖 . For every 𝑖, 𝑓 induces a mapping

𝑓𝑖 ∶ 𝜅𝑖 → 𝜆𝑖 given by the 𝑖th component of the restriction of 𝑓 to 𝜅𝑖 .
Since 𝜅𝑖 < 𝜆𝑖 , the set 𝐵𝑖 ∶= 𝜆𝑖 ⧵ im(𝑓𝑖 ) is non-empty for every 𝑖. By (AC)
there exists some 𝑏 ∈ ∏𝑖∈𝐼 𝐵𝑖 ⊆ ∏𝑖∈𝐼 𝜆𝑖 . Clearly 𝑏 ∉ im(𝑓). □
1.10. Cofinality
In this section we shall use the notion of cofinality to prove for instance
that 2ℵ0 ≠ ℵ𝜔 .
Definition.
• Let 𝑋 be a totally ordered set. We say that a subset 𝑌 ⊆ 𝑋 is
cofinal in 𝑋 if 𝑌 is not bounded in 𝑋, that is, if for any 𝑥 ∈ 𝑋
there exists 𝑦 ∈ 𝑌 such that 𝑥 ≤ 𝑦. We say that a function
𝑓 ∶ 𝑍 → 𝑋 is cofinal if im(𝑓) is cofinal in 𝑋.
• The cofinality of an ordinal 𝛼, denoted by cof(𝛼), is the smallest
ordinal 𝛽 such that there exists a cofinal function 𝛽 → 𝛼.
Example 1.10.1.
(1) cof(0) = 0.
(2) cof(𝛼 + 1) = 1 for any ordinal 𝛼.
(3) cof(𝜔) = 𝜔.
Proposition 1.10.2. Let 𝛼 be an ordinal.
(1) cof(𝛼) ≤ 𝛼.
(2) cof(𝛼) is a cardinal.
(3) cof(𝛼) is the smallest ordinal 𝛽 such that there exists a cofinal
and strictly increasing map 𝛽 → 𝛼.
(4) cof(cof(𝛼)) = cof(𝛼).
Proof. (1) is clear, and (2) follows from the fact that any ordinal 𝛽 is in
bijection with card(𝛽) ≤ 𝛽.
(3) It is enough to provide some 𝛽 ≤ cof(𝛼) and a cofinal and strictly
increasing map 𝛽 → 𝛼. By hypothesis, there exists a cofinal map ℎ ∶
cof(𝛼) → 𝛼. Let us define
𝑋 = {𝑥 ∈ cof(𝛼) ∣ ℎ(𝑦) < ℎ(𝑥) for every 𝑦 < 𝑥}.
The set ℎ(𝑋) = {ℎ(𝑥) ∣ 𝑥 ∈ 𝑋} is cofinal in 𝛼. Indeed, let 𝛾 < 𝛼. By
the cofinality of ℎ, there exists 𝑦 ∈ cof(𝛼) such that ℎ(𝑦) ≥ 𝛾. When 𝑦 is
minimal with this property, we have 𝑦 ∈ 𝑋.
Since (𝑋, <) ≅ (𝛽, ∈) for some 𝛽 ≤ cof(𝛼), we are done, because the
restriction of ℎ to 𝑋 is cofinal and strictly increasing.
(4) cof(cof(𝛼)) ≤ cof(𝛼) follows from part (1). For the inequality
in the other direction, let us consider the cofinal and strictly increasing
functions 𝑓 ∶ cof(cof(𝛼)) → cof(𝛼) and 𝑔 ∶ cof(𝛼) → 𝛼, which are
1.10. Cofinality 21
possible by (3). Then the function 𝑔 ∘ 𝑓 ∶ cof(cof(𝛼)) → 𝛼 is cofinal, and

hence cof(𝛼) ≤ cof(cof(𝛼)). □
We shall say that an infinite cardinal 𝜅 is regular if cof(𝜅) = 𝜅, and

singular if cof(𝜅) < 𝜅.
Proposition 1.10.3. Any infinite cardinal which is a successor is regular.

In particular ℵ1 is regular.
Proof. Let 𝜅 = ℵ𝛽+1 = ℵ+ 𝛽 . Note that for a limit ordinal 𝛼, a subset

𝑋 ⊆ 𝛼 is cofinal if and only if 𝛼 = ⋃𝛾∈𝑋 𝛾. (This follows from Propo-
sition 1.5.5.) Consider a function 𝑓 ∶ 𝜆 → 𝜅 for some 𝜆 < 𝜅. Then
𝜆 ≤ ℵ𝛽 and it follows from Proposition 1.9.5(3) that card (⋃𝛽<𝜆 𝑓(𝛽)) ≤
sup ({card (𝑓(𝛽)) ∣ 𝛽 < 𝜆} ∪ {𝜆}) ≤ ℵ𝛽 . Hence 𝑓 is not cofinal. □
Proposition 1.10.4. If 𝜆 is a limit ordinal, then cof(ℵ𝜆 ) = cof(𝜆).
Proof. If 𝑓 ∶ 𝛼 → 𝜆 is cofinal, then 𝑓 ̃ ∶ 𝛼 → ℵ𝜆 , 𝛽 ↦ ℵ𝑓(𝛽) is cofinal

too, since ℵ𝜆 = ⋃𝛾<𝜆 ℵ𝛾 by definition. This proves cof(ℵ𝜆 ) ≤ cof(𝜆).
Conversely, let 𝑔 ∶ 𝛼 → ℵ𝜆 be cofinal. The map 𝑔̃ ∶ 𝛼 → 𝜆, defined by
𝑔(𝛽)
̃ = 0 if 𝑔(𝛽) is finite, and 𝑔(𝛽)
̃ = 𝛾 if card(𝑔(𝛽)) = ℵ𝛾 , is cofinal. □
Proposition 1.10.5. Let 𝜅 ≥ 2 and 𝜆 ≥ ℵ0 be cardinals. Then cof (𝜅𝜆 ) >

𝜆.
Proof. Consider a map 𝑓 ∶ 𝛼 → 𝜅𝜆 , with 𝛼 some ordinal ≤ 𝜆. Since

𝑓(𝛽) < 𝜅𝜆 for every 𝛽 < 𝛼, it follows from König’s Theorem that
card ( 𝑓(𝛽)) ≤ ∑ card(𝑓(𝛽)) < ∏ (𝜅𝜆 ) = 𝜅𝜆⋅card(𝛼) ≤ 𝜅𝜆 .

⋃
𝛽<𝛼 𝛽<𝛼 𝛽<𝛼
Hence 𝑓 is not cofinal. □
Corollary 1.10.6. 2ℵ0 ≠ ℵ𝜔 .
Proof. We have cof(ℵ𝜔 ) = cof(𝜔) = 𝜔 = ℵ0 < cof (2ℵ0 ). □
1.11. Exercises
Exercise 1.11.1.
(1) Prove that an ordinal 𝛼 is a limit if and only if there is an ordinal
𝛽 ≠ 0 such that 𝛼 = 𝜔𝛽.
(2) Prove that 𝜔2 = 𝜔𝜔 is not of the form 𝛿 + 𝜔.
Exercise 1.11.2 (Cantor normal form). Let 𝛼 be an ordinal > 1.
(1) Prove that 𝛼𝛾 ≥ 𝛾 for any ordinal 𝛾. (Is there any example with
𝛼𝛾 = 𝛾?)
(2) Let 𝛽 be an ordinal > 0. Prove that there exists an ordinal 𝛾
+
such that 𝛼𝛾 ≤ 𝛽 < 𝛼𝛾 .
(3) Deduce that any ordinal 𝛽 can be expanded in basis 𝛼: there ex-
ists a finite sequence of ordinals 𝛽1 > … > 𝛽𝑛 ≥ 0 and ordinals
𝑘𝑖 with 0 < 𝑘𝑖 < 𝛼 such that
𝛽 = 𝛼𝛽1 𝑘1 + … + 𝛼𝛽𝑛 𝑘𝑛 .
Furthermore the natural number 𝑛 and the sequences (𝛽𝑖 ) and
(𝑘𝑖 ) are unique.
The expansion into basis 𝜔 is called the Cantor normal
form.
Exercise 1.11.3 (Goodstein sequences). Let 𝑛, 𝑝 be natural numbers,
with 𝑝 ≥ 2. Let us define the iterated expansion of 𝑛 in basis 𝑝 as follows:
first expand 𝑛 in basis 𝑝 (for instance if 𝑛 = 35, 𝑝 = 2, one has 35 =
25 + 2 + 1). Then, expand the exponents in basis 𝑝 and so on, so that all
numbers occurring are < 𝑝. For instance the iterated expansion of 35 in
2
basis 2 is 35 = 22 +1 + 2 + 1.
For 𝑞 ≥ 𝑝 ≥ 2, one defines a function 𝑓𝑝,𝑞 ∶ ℕ → ℕ as follows. Let
𝑛 be a natural number. Then 𝑓𝑝,𝑞 (𝑛) is obtained by replacing all occur-
rences of 𝑝 in the iterated expansion of 𝑛 in basis 𝑝 by 𝑞. One similarly
defines ordinal valued functions 𝑓𝑝,𝜔 by replacing all occurrences of 𝑝 by
𝜔 (when 𝑝 > 2, one writes the coefficients at the right of the 𝑝𝑛 ). For in-
3
stance, 𝑓2,3 (35) = 33 +1 +3+1 = 59053 and 𝑓3,𝜔 (35) = 𝑓3,𝜔 (33 +3⋅2+2) =
𝜔𝜔 + 𝜔 ⋅ 2 + 2.
(1) Prove that 𝑓𝑞,𝑟 ∘ 𝑓𝑝,𝑞 = 𝑓𝑝,𝑟 for any 𝜔 ≥ 𝑟 > 𝑞 > 𝑝 ≥ 2.
1.11. Exercises 23
(2) Prove that the functions 𝑓𝑝,𝜔 are strictly increasing.

For any 𝑎 ∈ ℕ, the Goodstein sequence attached to 𝑎 is the sequence
(𝑔𝑛 (𝑎))𝑛≥2 defined by 𝑔2 (𝑎) ∶= 𝑎 and
𝑓𝑛,𝑛+1 (𝑔𝑛 (𝑎)) − 1 if 𝑔𝑛 (𝑎) ≠ 0,

𝑔𝑛+1 (𝑎) = {
0 otherwise.
For 𝑎 = 5, one has for instance 𝑔2 (5) = 5, 𝑔3 (5) = 33 + 1 − 1 = 27,

𝑔4 (5) = 44 − 1 = 255, and 𝑔5 (5) = 53 ⋅ 3 + 52 ⋅ 3 + 5 ⋅ 3 + 3 − 1 = 467.
(3) Let 𝑎 be a natural number. Study the monotonicity of the se-
quence 𝑓𝑝,𝜔 (𝑔𝑝 (𝑎)).
(4) Prove that for any 𝑎 ∈ ℕ there exists 𝑠(𝑎) ∈ ℕ such that 𝑔𝑛 (𝑎) =
0 when 𝑛 ≥ 𝑠(𝑎).
Exercise 1.11.4 (The order topology). On any totally ordered set 𝑋, one
may define a topology, the order topology, generated by the open inter-
vals, that is, by the sets of the form (−∞, 𝑏) = {𝑥 ∈ 𝑋 ∣ 𝑥 < 𝑏},
(𝑎, 𝑏) = {𝑥 ∈ 𝑋 ∣ 𝑎 < 𝑥 < 𝑏} or (𝑎, ∞), for 𝑎, 𝑏 ∈ 𝑋.
(1) Prove that for any 𝑋, the order topology is Hausdorff.
(2) Let 𝛼 and 𝛽 be ordinals endowed with the order topology. Prove
the following statements:
(a) 𝛼 is discrete if and only if 𝛼 ≤ 𝜔.
(b) 𝛼 is compact if and only if 𝛼 is not a limit ordinal.
(c) Suppose 𝑓 ∶ 𝛼 → 𝛽 is a weakly increasing function, that
is, 𝑥 ≤ 𝑦 ⇒ 𝑓(𝑥) ≤ 𝑓(𝑦). Then 𝑓 is continuous if and only
if, for any limit 𝜆 ∈ 𝛼, 𝑓(𝜆) = sup𝛾<𝜆 𝑓(𝛾).
Exercise 1.11.5 (Ulam’s Theorem). An Ulam matrix is a family of sub-

sets 𝑈𝛼,𝑛 of ℵ1 , 𝛼 < ℵ1 , 𝑛 < 𝜔, such that
• for any 𝑛 < 𝜔 and 𝛼, 𝛽 < ℵ1 with 𝛼 ≠ 𝛽, 𝑈𝛼,𝑛 ∩ 𝑈𝛽,𝑛 = ∅;
• for any 𝛼 < ℵ1 , the set ℵ1 ⧵ ⋃𝑛<𝜔 𝑈𝛼,𝑛 is at most countable.
(1) For any 𝜉 < ℵ1 , let 𝑓𝜉 ∶ 𝜔 → ℵ1 such that 𝜉 ⊆ im(𝑓𝜉 ). Set
𝑈𝛼,𝑛 = {𝜉 < ℵ1 | 𝑓𝜉 (𝑛) = 𝛼}.
Prove that (𝑈𝛼,𝑛 ) is an Ulam matrix.
(2) Let 𝜇 ∶ 𝒫(ℵ1 ) → [0, 1] be a 𝜎-additive measure on ℵ1 , that is,

𝜇(⋃𝑛<𝜔 𝐴𝑛 ) = ∑𝑛<𝜔 𝜇(𝐴𝑛 ) for all families (𝐴𝑛 )𝑛<𝜔 of pairwise
disjoint subsets of ℵ1 . Prove that if 𝜇({𝛼}) = 0 for every 𝛼 < ℵ1 ,
then 𝜇 is identically zero.
Exercise 1.11.6 (Closed unbounded sets). Let 𝜅 > ℵ0 be a regular cardi-

nal. A set 𝐶 ⊆ 𝜅 is said to be a club (closed unbounded) if it is closed and
cofinal, that is, if it is cofinal in 𝜅 and for any non-empty subset 𝐴 ⊆ 𝐶,
sup 𝐴 ∈ 𝐶 ∪ {𝜅}. A subset 𝑆 ⊆ 𝜅 is called stationary if 𝑆 ∩ 𝐶 ≠ ∅ for any
club 𝐶 ⊆ 𝜅.
(1) Let 𝜆 < 𝜅 and (𝐶𝑖 )𝑖<𝜆 be a family of clubs. Prove that ⋂𝑖<𝜆 𝐶𝑖
is a club.
(2) Let (𝐶𝑖 )𝑖<𝜅 be a family of clubs. Prove that the diagonal inter-
section △𝑖<𝜅 𝐶𝑖 ∶= {𝛼 < 𝜅 | 𝛼 ∈ ⋂𝑖<𝛼 𝐶𝑖 } is a club.
(3) (Fodor’s Lemma) Let 𝑆 ⊆ 𝜅 be a stationary set and suppose
𝑓 ∶ 𝑆 → 𝜅 is a regressive map (that is, such that 𝑓(𝛼) < 𝛼 for
any 𝛼 ∈ 𝑆). Prove that there exists a stationary set 𝑇 ⊆ 𝑆 such
that 𝑓 is constant on 𝑇.
(4) (Solovay’s Theorem for ℵ1 ) Let 𝜅 = ℵ1 , and let 𝑆 ⊆ ℵ1 be a
stationary set. Prove that 𝑆 can be written as the union of ℵ1
pairwise disjoint stationary sets.
[Hint: For any limit ordinal 𝛼 ∈ 𝑆, write 𝛼 = sup𝑛<𝜔 𝑎𝛼𝑛 . Use
this to define a regressive function to which one may apply
Fodor’s Lemma.]
Exercise 1.11.7 (Solovay’s Theorem). Let 𝜅 > ℵ0 be a regular cardinal.

The aim of this exercise is to prove Solovay’s Theorem: Any stationary
set 𝑆 ⊆ 𝜅 can be written as the union of 𝜅 pairwise disjoint stationary sets.
(1) For a cardinal 𝜆 < 𝜅, let 𝐸𝜆𝜅 = {𝛼 < 𝜅 | cof(𝛼) = 𝜆}. Prove
Solovay’s Theorem when 𝑆 ⊆ 𝐸𝜔𝜅 .
(2) Prove Solovay’s Theorem when 𝑆 ⊆ {𝛼 < 𝜅 | cof(𝛼) < 𝛼}.
(3) Assume 𝑆 is a set of regular cardinals. Prove that
𝑇 = {𝛼 ∈ 𝑆 | 𝑆 ∩ 𝛼 is not stationary in 𝛼}
is stationary.
1.11. Exercises 25
[Hint: Prove it by contradiction after noticing that if 𝐶 is a club,

the set 𝐶 ′ ⊆ 𝐶 of limits of points of 𝐶 is still a club.]
(4) Prove Solovay’s Theorem in the general case.
[Hint: Prove it is enough to consider the case when 𝑆 is a set
of regular cardinals. For 𝛼 ∈ 𝑇, consider a strictly increasing
sequence (𝑎𝛼𝜉 ∶ 𝜉 < 𝜅) with limit 𝛼 such that 𝑎𝛼𝜉 ∉ 𝑆. Then
adapt the proof of (1).]
Exercise 1.11.8 (Filters and ultrafilters). Let 𝑋 be a non-empty set. A

filter on 𝑋 is a set 𝐹 ⊆ 𝒫(𝑋) such that
(i) 𝑋 ∈ 𝐹, ∅ ∉ 𝐹,
(ii) if 𝐴, 𝐵 ∈ 𝐹, then 𝐴 ∩ 𝐵 ∈ 𝐹, and
(iii) if 𝐴 ∈ 𝐹 and 𝐴 ⊆ 𝐵, then 𝐵 ∈ 𝐹.
A set ℬ of non-empty subsets of 𝑋 is a filter basis on 𝑋 if for any
𝐴, 𝐵 ∈ ℬ there is 𝐶 ∈ ℬ such that 𝐶 ⊆ 𝐴 ∩ 𝐵.
(1) Prove that any filter basis ℬ is contained in a minimal filter
with respect to inclusion, denoted by 𝐹ℬ .
(2) Let 𝐽 be a set and let 𝐼 be the set of finite subsets of 𝐽. For any
𝑖 ∈ 𝐼, set 𝐼𝑖 = {𝑗 ∈ 𝐼 ∣ 𝑖 ⊆ 𝑗}. Prove that the set ℬ whose
elements are the sets 𝐼𝑖 , 𝑖 ∈ 𝐼, is a filter basis on 𝐼.
(3) Prove that when 𝑋 is infinite, the set of subsets of 𝑋 whose
complement in 𝑋 is finite is a filter. This filter is called the
Fréchet filter.
(4) An ultrafilter is a filter which is maximal with respect to inclu-
sion. Prove that a filter 𝐹 on a non-empty set 𝑋 is an ultrafilter
if and only if for any 𝐴 ⊆ 𝑋, either 𝐴 or 𝑋 ⧵ 𝐴 belongs to 𝐹.
(5) Prove that for any 𝑥 ∈ 𝑋, the set 𝑈𝑥 = {𝑌 ⊆ 𝑋 ∣ 𝑥 ∈ 𝑌 } is an
ultrafilter on 𝑋. Such an ultrafilter is called principal.
(6) Prove that an ultrafilter is principal if and only if it contains a
finite set as an element. Deduce that when 𝑋 is finite, any ul-
trafilter on 𝑋 is principal, and when 𝑋 is infinite, an ultrafilter
𝑈 is non-principal if and only if it contains the Fréchet filter as
a subset.
(7) Prove that any filter is contained in some ultrafilter. In partic-

ular, on any infinite set there exists a non-principal ultrafilter.
Exercise 1.11.9 (Hausdorff’s Theorem). The aim of this exercise is to

prove the following result due to Hausdorff: On an infinite set 𝑋 of car-
𝜅
dinality 𝜅, there exist 2(2 ) distinct ultrafilters.
(1) A family 𝐹 of subsets of 𝑋 is said to be free if whenever 𝐴1 , . . . ,

𝐴𝑛 , 𝐵1 , . . . , 𝐵𝑚 are distinct elements of 𝐹, then 𝐴1 ∩ ⋯ ∩ 𝐴𝑛 ∩
(𝑋 ⧵ 𝐵1 ) ∩ ⋯ ∩ (𝑋 ⧵ 𝐵𝑚 ) is non-empty. Prove that Hausdorff’s
Theorem is a consequence of the following statement: there
exists a free family of cardinality 2𝜅 .
(2) Consider the set 𝑌 consisting of pairs (𝐹, (𝑃1 , … , 𝑃𝑛 )) with 𝐹 a
finite subset of 𝑋 and (𝑃1 , … , 𝑃𝑛 ) a finite sequence of of subsets
of 𝐹. What is the cardinality of 𝑌 ?
(3) To any 𝐴 ∈ 𝒫(𝑋) we assign a subset 𝐴′ of 𝑋 ∪ 𝑌 as follows:
• if 𝑥 ∈ 𝑋, then 𝑥 ∈ 𝐴′ if and only if 𝑥 ∈ 𝐴;
• if 𝑦 ∈ 𝑌 and 𝑦 = (𝐹, (𝑃1 , … , 𝑃𝑛 )), then 𝑦 ∈ 𝐴′ if and only if
𝐴 ∩ 𝐹 is one of the 𝑃𝑖 ’s.
Prove that the set of the 𝐴′ for 𝐴 ⊆ 𝑋 is a free family (on 𝑋 ∪ 𝑌 ).
(4) Conclude Hausdorff’s Theorem.
Exercise 1.11.10 (The continuum hypothesis for closed subsets of ℝ).

The aim of this exercise is to prove that the continuum hypothesis holds
for closed subsets of the real line (Cantor’s Theorem), that is, if 𝐹 ⊆ ℝ is
closed, then either 𝐹 is countable or card(𝐹) = 2ℵ0 .
Let 𝐹 ⊆ ℝ be a closed subset such that card(𝐹) > ℵ0 .
(1) Let 𝐶 be the set of 𝑥 ∈ 𝐹 such that there is an open set Ω ⊆ ℝ

which contains 𝑥 and satisfies card(Ω ∩ 𝐹) ≤ ℵ0 . Set 𝐹 ′ ∶=
𝐹 ⧵ 𝐶.
Prove the following properties:
(a) 𝐶 is countable,
(b) 𝐹 ′ is a closed subset of ℝ, and
(c) any non-empty open subset of 𝐹 ′ is uncountable.
(2) Let 𝐼 be an open interval such that 𝐼 ∩ 𝐹 ′ ≠ ∅.
1.11. Exercises 27
Prove that for any 𝜖 > 0 there are open intervals 𝐼0 and 𝐼1
of length at most 𝜖 such that 𝐼𝑖 ∩ 𝐹 ′ ≠ ∅ and 𝐼𝑖 ⊆ 𝐼 for 𝑖 = 0, 1
and such that 𝐼0 ∩ 𝐼1 = ∅. (Here, 𝐽 denotes the closure of 𝐽.)
(3) Prove that there is an injection of 2ℵ0 into 𝐹 ′ .
(4) Prove Cantor’s Theorem.
Exercise 1.11.11. A tree is a partial order (𝑇, <𝑇 ) such that for every
𝑡 ∈ 𝑇, the set
𝑡 ̂ ∶= {𝑠 ∈ 𝑇 | 𝑠 <𝑇 𝑡}
is well-ordered. If 𝑇 has a smallest element, it is called the root of 𝑇.

The height of 𝑡 ∈ 𝑇 is defined as the unique ordinal isomorphic to
(𝑡,̂ <𝑇 ) and denoted by ht(𝑡). Finally, ht(𝑇) ∶= sup{ht(𝑡) + 1 | 𝑡 ∈ 𝑇} is
the height of 𝑇. For an ordinal 𝛼 one sets 𝑇(𝛼) = {𝑡 ∈ 𝑇 ∣ ht(𝑡) = 𝛼}.
A branch in a tree (𝑇, <𝑇 ) is a totally ordered subset which is max-
imal with respect to inclusion. It is called cofinal in 𝑇 if it non-trivially
intersects any level 𝑇(𝛼) for 𝛼 < ht(𝑇).
Prove König’s Lemma: If 𝑇 is a tree of height 𝜔 such that 𝑇(𝑛) is
finite for every 𝑛 ∈ 𝜔, then 𝑇 contains a cofinal branch.
Exercise 1.11.12.
(1) For 𝛼 ∈ ℵ1 , let 𝑌𝛼 be the set of strictly increasing non-cofinal

maps from 𝛼 to ℚ. For 𝛽 < 𝛼, denote by 𝜌𝛽𝛼 ∶ 𝑌𝛼 → 𝑌𝛽 the (sur-
jective) restriction map. Prove that there are countable subsets
𝑋𝛼 ⊆ 𝑌𝛼 such that for any 𝛼 > 𝛽, the following properties hold:
(a) 𝜌𝛽𝛼 (𝑋𝛼 ) = 𝑋𝛽 .
(b) For any 𝑓 ∈ 𝑋𝛽 , any upper bound 𝑞 ∈ ℚ of im(𝑓) and any
𝜖 > 0 there is 𝑔 ∈ 𝑋𝛼 with 𝜌𝛽𝛼 (𝑔) = 𝑓 such that im(𝑔) is
bounded above by 𝑞 + 𝜖.
(2) Let lim 𝑋 be the set of all 𝑓 = (𝑓𝛼 ) ∈ ∏𝛼∈ℵ 𝑋𝛼 such that
←−−𝛼∈ℵ1 𝛼 1
𝛼
𝜌𝛽 (𝑓𝛼 ) = 𝑓𝛽 for all 𝛽 < 𝛼. Prove that lim 𝑋𝛼 = ∅.
←−−𝛼∈ℵ1
(3) Deduce Aronszajn’s Theorem: There is an Aronszajn tree, that
is, a tree 𝑇 of height ℵ1 with 𝑇(𝛼) countable for all 𝛼 < ℵ1 such
that 𝑇 does not admit a cofinal branch.
1.12. Appendix: Hindman’s Theorem

In this appendix we shall freely use the material on ultrafilters treated in
Exercise 1.11.8. Let 𝐸 be a non-empty set and 𝒰(𝐸) the set of all ultra-
filters on 𝐸. When 𝐸 is infinite, there exist non-principal ultrafilters on
𝐸.
For a subset 𝐴 ⊆ 𝐸, set ⟨𝐴⟩ = {𝑈 ∈ 𝒰(𝐸) ∣ 𝐴 ∈ 𝑈}.
Lemma 1.12.1. The sets ⟨𝐴⟩, for ∅ ≠ 𝐴 ⊆ 𝐸, form a basis of open sets for
a compact Hausdorff topology on 𝒰(𝐸).
Proof. We have ⟨𝐴1 ⟩ ∩ ⋯ ∩ ⟨𝐴𝑛 ⟩ = ⟨𝐴1 ∩ ⋯ ∩ 𝐴𝑛 ⟩, which proves that

the sets ⟨𝐴⟩ form a basis of open sets for some topology. Furthermore,
⟨𝐸 ⧵ 𝐴⟩ = 𝒰(𝐸) ⧵ ⟨𝐴⟩, thus ⟨𝐴⟩ is open and closed. In particular, the
topology is Hausdorff.
Consider a covering ⋃𝑖∈𝐼 Ω𝑖 of 𝒰(𝐸) by open subsets. To prove the
existence of 𝐼0 ⊆ 𝐼 finite such that ⋃𝑖∈𝐼 Ω𝑖 = 𝒰(𝐸) we may assume
0
Ω𝑖 = ⟨𝐴𝑖 ⟩ for any 𝑖 ∈ 𝐼. If for some finite 𝐼0 ⊆ 𝐼 one has ⋃𝑖∈𝐼 𝐴𝑖 = 𝐸,
0
we are done, since then ⋃𝑖∈𝐼 ⟨𝐴𝑖 ⟩ = 𝒰(𝐸). Otherwise, the set
0
ℬ={ 𝐸 ⧵ 𝐴𝑖 ∣ 𝐼0 ⊆ 𝐼 finite}
⋂
𝑖∈𝐼0
is a filter basis and is contained in some ultrafilter 𝑈 on 𝐸 by Exercise

1.11.8, which contradicts ⋃𝑖∈𝐼 Ω𝑖 = 𝒰(𝐸). □
Now consider 𝒰(ℕ∗ ). Let 𝑘 ∈ ℕ∗ , 𝐴 ⊆ ℕ∗ and 𝑈 ∈ 𝒰(ℕ∗ ). One sets

𝐴 − 𝑘 ∶= ℕ∗ ∩ {𝑎 − 𝑘 ∣ 𝑎 ∈ 𝐴} and
𝐴𝑈 ∶= {𝑘 ∈ ℕ∗ ∣ 𝐴 − 𝑘 ∈ 𝑈}.
Lemma 1.12.2. For any 𝑈 ∈ 𝒰(ℕ∗ ) and every 𝐴, 𝐵 ⊆ ℕ∗ , one has ℕ∗𝑈 =
ℕ∗ , (𝐴 ∩ 𝐵)𝑈 = 𝐴𝑈 ∩ 𝐵𝑈 and (ℕ∗ ⧵ 𝐴)𝑈 = ℕ∗ ⧵ 𝐴𝑈 .
Proof. These properties are easy consequences of the fact that for 𝑘 ∈
ℕ∗ one has ℕ∗ − 𝑘 = ℕ∗ , (𝐴 − 𝑘) ∩ (𝐵 − 𝑘) = 𝐴 ∩ 𝐵 − 𝑘 and (ℕ∗ ⧵ 𝐴) − 𝑘 =
ℕ∗ ⧵ (𝐴 − 𝑘). The details are left as an exercise. □
For 𝑈, 𝑉 ∈ 𝒰(ℕ∗ ), one sets 𝑈 ⊕ 𝑉 ∶= {𝐴 ⊆ ℕ∗ ∣ 𝐴𝑈 ∈ 𝑉}.
1.12. Appendix: Hindman’s Theorem 29
Lemma 1.12.3.
(1) For any 𝑈, 𝑉 ∈ 𝒰(ℕ∗ ), 𝑈 ⊕ 𝑉 ∈ 𝒰(ℕ∗ ).
(2) ⊕ is associative.
(3) For fixed 𝑈, the function 𝑓𝑈 ∶ 𝒰(ℕ∗ ) → 𝒰(ℕ∗ ), 𝑉 ↦ 𝑈 ⊕ 𝑉, is
continuous.
Proof. (1) It is clear that ℕ∗ ∈ 𝑈 ⊕ 𝑉, and that for any 𝐴 ∈ 𝑈 ⊕ 𝑉 and

𝐴 ⊆ 𝐵 one has 𝐵 ∈ 𝑈 ⊕ 𝑉. Stability of 𝑈 ⊕ 𝑉 under intersection and
the fact that 𝐴 ∈ 𝑈 ⊕ 𝑉 if and only if ℕ∗ ⧵ 𝐴 ∉ 𝑈 ⊕ 𝑉 are consequences
of Lemma 1.12.2.
(2) One has (𝐴 − 𝑘)𝑈 = 𝐴𝑈 − 𝑘 (exercise). From this, one may
deduce that (𝐴𝑈 )𝑉 = 𝐴𝑈⊕𝑉 . Indeed, for 𝑙 ∈ ℕ∗ , one has
𝑙 ∈ (𝐴𝑈 )𝑉 ⇔ 𝐴𝑈 − 𝑙 ∈ 𝑉 ⇔ (𝐴 − 𝑙)𝑈 ∈ 𝑉
⇔ 𝐴 − 𝑙 ∈ 𝑈 ⊕ 𝑉 ⇔ 𝑙 ∈ 𝐴𝑈⊕𝑉 .
∗
If 𝑈, 𝑉, 𝑊 ∈ 𝒰(ℕ ), we deduce that
𝐴 ∈ 𝑈 ⊕ (𝑉 ⊕ 𝑊) ⇔ 𝐴𝑈 ∈ 𝑉 ⊕ 𝑊
⇔ (𝐴𝑈 )𝑉 = 𝐴𝑈⊕𝑉 ∈ 𝑊 ⇔ 𝐴 ∈ (𝑈 ⊕ 𝑉) ⊕ 𝑊,
which proves associativity of ⊕.
(3) Let 𝐴 ⊆ ℕ∗ . Continuity of 𝑓𝑈 follows from the following:
𝑓𝑈−1 (⟨𝐴⟩) = {𝑉 ∈ 𝒰(ℕ∗ ) ∣ 𝐴 ∈ 𝑈 ⊕ 𝑉}
= {𝑉 ∈ 𝒰(ℕ∗ ) ∣ 𝐴𝑈 ∈ 𝑉} = ⟨𝐴𝑈 ⟩. □
We shall say an ultrafilter 𝑈 on ℕ∗ is idempotent if 𝑈 ⊕ 𝑈 = 𝑈.

Proposition 1.12.4. There exists an idempotent ultrafilter in 𝒰(ℕ∗ ).
Remark 1.12.5. For 𝑚, 𝑛 ∈ ℕ∗ , one can easily check that the sum of
the corresponding principal ultrafilters is given by 𝑈𝑚 ⊕ 𝑈𝑛 = 𝑈𝑚+𝑛 . In
particular no principal ultrafilter is idempotent. □
Proof of Proposition 1.12.4. One considers the set

𝔸 ∶= {𝒜 ⊆ 𝒰(ℕ∗ ) ∣ 𝒜 is a non-empty closed set with 𝒜 ⊕ 𝒜 ⊆ 𝒜}.
The set 𝔸 is non-empty since it contains 𝒰(ℕ∗ ). Furthermore the partial
order on 𝔸 given by ⊇ is inductive. This follows from the fact that a
subset of a compact Hausdorff space is closed if and only if it is compact.

Thus, if (𝒜𝑖 )𝑖∈𝐼 is a totally ordered subset of 𝔸, then ⋂𝑖∈𝐼 𝒜𝑖 ∈ 𝔸. By
Zorn’s Lemma, there exists ℬ ∈ 𝔸 which is minimal for inclusion.
Let 𝑈 ∈ ℬ. We shall prove that 𝑈 ⊕ 𝑈 = 𝑈.
First, we consider ℬ′ ∶= 𝑈 ⊕ℬ ⊆ ℬ. Since 𝑓𝑈 is continuous, the set
′
ℬ = 𝑓𝑈 (ℬ) is compact (hence closed). Furthermore, if 𝑉1 , 𝑉2 ∈ ℬ, one
has (𝑈 ⊕𝑉1 )⊕(𝑈 ⊕𝑉2 ) = 𝑈 ⊕𝑊 for some 𝑊 ∈ ℬ, since ℬ⊕ℬ ⊆ ℬ. This
proves that ℬ′ ⊕ ℬ′ ⊆ ℬ′ . Hence we have ℬ′ ∈ 𝔸, and thus ℬ′ = ℬ by
minimality of ℬ. In particular, there exists 𝑉 ′ ∈ ℬ such that 𝑈 ⊕𝑉 ′ = 𝑈.
Set 𝒱 ′ = {𝑉 ′ ∈ ℬ ∣ 𝑈 ⊕ 𝑉 ′ = 𝑈}. By continuity of the map 𝑓𝑈 ,
𝒱 = 𝑓𝑈−1 ({𝑈}) ∩ ℬ is closed, and we just proved it is non-empty. If
′
𝑉 ′ , 𝑉 ″ ∈ 𝒱 ′ , then 𝑈 ⊕ (𝑉 ′ ⊕ 𝑉 ″ ) = (𝑈 ⊕ 𝑉 ′ ) ⊕ 𝑉 ″ = 𝑈 ⊕ 𝑉 ″ = 𝑈.
Thus 𝒱 ′ is closed under ⊕ and it belongs to 𝔸. It follows that 𝒱 ′ = ℬ by
minimality of ℬ. In particular, 𝑈 ∈ 𝒱 ′ , hence 𝑈 ⊕ 𝑈 = 𝑈. □
For a subset 𝐴 ⊆ ℕ∗ , one denotes by
Σ𝐴 ∶= { ∑ 𝑘 | 𝐴0 ⊆ 𝐴 finite }
𝑘∈𝐴0
the set of finite sums of distinct elements of 𝐴.
Theorem 1.12.6 (Hindman’s Theorem). Let 𝜒 ∶ ℕ∗ → {1, … , 𝑛} be any

function. Then there exists 𝐵 ⊆ ℕ∗ infinite such that 𝜒 is constant on Σ𝐵 .
Proof. Let 𝑈 = 𝑈 ⊕ 𝑈 ∈ 𝒰(ℕ∗ ) be an idempotent ultrafilter as pro-

vided by Proposition 1.12.4. There then exists a unique 𝑖0 ≤ 𝑛 such that
𝜒−1 (𝑖0 ) ∈ 𝑈. Set 𝐴 ∶= 𝜒−1 (𝑖0 ). We will prove the existence of an infinite
set 𝐵 ⊆ 𝐴 such that Σ𝐵 ⊆ 𝐴, which will complete the proof.
As 𝑈 = 𝑈 ⊕ 𝑈, for any 𝐶 ∈ 𝑈 one has 𝐶𝑈 ∈ 𝑈, hence 𝐶 ∩ 𝐶𝑈 ∈ 𝑈.
In particular, 𝐶 ∩ 𝐶𝑈 is infinite. Furthermore, for any 𝑘 ∈ 𝐶 ∩ 𝐶𝑈 one
has (𝐶 − 𝑘) ∩ 𝐶 ∈ 𝑈.
By induction on 𝑛 ∈ ℕ, one defines a decreasing sequence (𝐴𝑛 )𝑛∈ℕ
in 𝑈 and a strictly increasing sequence of integers (𝑘𝑛 )𝑛∈ℕ as follows.
One sets 𝐴0 ∶= 𝐴 and one chooses 𝑘0 ∈ 𝐴0 ∩ 𝐴0𝑈 . Assuming 𝐴𝑛 and
𝑘𝑛 already constructed, one sets 𝐴𝑛+1 ∶= (𝐴𝑛 − 𝑘𝑛 ) ∩ 𝐴𝑛 ∈ 𝑈, and one
1.12. Appendix: Hindman’s Theorem 31
chooses some 𝑘𝑛+1 > 𝑘𝑛 such that 𝑘𝑛+1 ∈ 𝐴𝑛+1 ∩ 𝐴𝑛+1

𝑈 . This is possible
since 𝐴𝑛+1 ∩ 𝐴𝑛+1
𝑈 is infinite.
Now 𝐵 = {𝑘𝑛 ∣ 𝑛 ∈ ℕ} is an infinite subset of 𝐴. Let 𝐼 be a finite non-
empty subset of ℕ. We now prove by induction on card(𝐼) that ∑𝑖∈𝐼 𝑘𝑖 ∈
𝐴min 𝐼 . If 𝐼 is a singleton, this is clear. Assume now that card(𝐼) = 𝑚 + 1
for some 𝑚 > 0 and let 𝑖0 < … < 𝑖𝑚 be the elements of 𝐼. By induction
we know that
𝑘𝑖1 + … + 𝑘𝑖𝑚 ∈ 𝐴𝑖1 ⊆ 𝐴𝑖0 +1 = 𝐴𝑖0 ∩ (𝐴𝑖0 − 𝑘𝑖0 ).
It follows that 𝑘𝑖0 + … + 𝑘𝑖𝑚 ∈ 𝐴𝑖0 as claimed. We have thus proved that
Σ𝐵 ⊆ 𝐴. □
10.1090/stml/089/02
Chapter 2
First-order Logic
Introduction
This chapter introduces the basics of first-order logic. First-order logic is
a standard way to formalize mathematics. For instance Peano arithmetic
and Zermelo-Fraenkel set theory formalize respectively number theory
and set theory. In a given formal language, one can define the basic no-
tions of variables, terms and formulas. One can also define formal logical
deduction rules that allow one to deduce new statements from a given
set of axioms via a formal proof. This corresponds to the syntactic side
of formal logic, where constructions are purely symbolic and no specific
interpretation or meaning is given to the symbols.
On the other hand, structures for a given language are sets where it
is possible to interpret terms and formulas. In particular, one can give a
meaning to the validity of a formula with no free variables in a structure.
This is the semantic side.
The main result in this chapter is Gödel’s Completeness Theorem
which states that a formula with no free variables can be formally de-
duced from a given set of axioms if and only if it is valid in every structure
satisfying these axioms. This particularly remarkable result shows that
the syntactic and semantic viewpoints are equivalent. It also provides
an a posteriori justification for our choice of syntax.
33
34 2. First-order Logic
2.1. Languages and Structures

Statements in first-order logic are sequences of symbols describing prop-
erties of certain structures. For instance the statement given by
∀𝑥(𝑥 > 0 → ∃𝑦 𝑦 ⋅ 𝑦 = 𝑥)
is satisfied in some ordered field if and only if every positive element is
a square. It holds in the ordered field of real numbers, that is, in the
structure ℜ = ⟨ℝ; 0, 1, +, −, ⋅, <⟩, and it does not hold in the ordered field
of rational numbers.
We will see that this statement can be expressed in first-order logic.
However, other properties of ordered field, like being complete (any
bounded non-empty subset admits a supremum) or archimedean (for
every 𝑥 > 0 there exists 𝑛 ∈ ℕ such that 𝑛𝑥 > 1), cannot be expressed
as statements in first-order logic.
Definition. A ( first-order) language is a set of symbols ℒ composed of
two disjoint subsets:
(1) The first part (common to all languages) consists of auxiliary
symbols “(” and “)” together with the following logical symbols:
the set of variables 𝒱 = {𝑣𝑛 ∣ 𝑛 ∈ ℕ}
the equality symbol = (“equals”)
connectives ¬ (negation, “not”), and
∧ (conjunction, “and”)
the existential quantifier ∃ (“there exists”)
(2) The second part, called the signature of ℒ and denoted by 𝜎ℒ ,
consists of the non-logical symbols of ℒ. It consists of
• a set of constant symbols 𝒞 ℒ ;
• a sequence of sets ℱ𝑛ℒ , 𝑛 ∈ ℕ∗ , where elements of ℱ𝑛ℒ are
called 𝑛-ary function symbols;
• a sequence of sets ℛ𝑛ℒ , 𝑛 ∈ ℕ∗ , where elements of ℛ𝑛ℒ are
called 𝑛-ary relation symbols (or 𝑛-ary predicates).
The language ℒ is given by the (disjoint) union of these sets of symbols.
Let us note that a language is always infinite. The first part being
common to all languages, we shall make the abuse of notation of identi-
fying ℒ and 𝜎ℒ .
2.1. Languages and Structures 35
Example 2.1.1.
ℒ∅ = ∅ The empty language.
ℒ𝑟𝑖𝑛𝑔 = {0, 1, +, −, ⋅} The ring language (with 1).
ℒ𝑜𝑟𝑑 = {<} The order language.
ℒ𝑜.𝑟𝑖𝑛𝑔 = ℒ𝑟𝑖𝑛𝑔 ∪ ℒ𝑜𝑟𝑑 The ordered ring language.
ℒ𝑎𝑟 = {0, 𝑆, +, ⋅, <} The language of arithmetic.
ℒ𝑠𝑒𝑡 = {∈} The language of set theory.
In these examples, 0 and 1 are constant symbols, − and 𝑆 unary
function symbols, + and ⋅ binary function symbols, and < and ∈ binary
relation symbols.
Definition. Let ℒ be a first-order language. An ℒ-structure 𝔄 consists
of a non-empty set 𝐴 (called the base set of 𝔄) together with an element
𝑐𝔄 ∈ 𝐴 for each 𝑐 ∈ 𝒞 ℒ , a function 𝑓𝔄 ∶ 𝐴𝑛 → 𝐴 for each 𝑓 ∈ ℱ𝑛ℒ and a
subset 𝑅𝔄 ⊆ 𝐴𝑛 for each 𝑅 ∈ ℛ𝑛ℒ . We write 𝔄 = ⟨𝐴; (𝑍 𝔄 )𝑍∈𝜍ℒ ⟩.
𝑍 𝔄 is called the interpretation of the symbol 𝑍 ∈ 𝜎ℒ in the structure

𝔄.
Example 2.1.2.
(1) 𝔑 = ⟨ℕ; 0, 𝑆, +, ⋅, <⟩ is an ℒ𝑎𝑟 -structure, with 𝑆 the successor
function assigning 𝑥 + 1 to 𝑥, and the other symbols having
their usual interpretation.
(2) ℭ = ⟨ℂ; 0, 1, +, −, ⋅⟩, the field of complex numbers, is an ℒ𝑟𝑖𝑛𝑔 -
structure.
(3) ℜ = ⟨ℝ; 0, 1, +, −, ⋅, <⟩, the ordered field of real numbers, is an
ℒ𝑜.𝑟𝑖𝑛𝑔 -structure.
Definition. We say that two ℒ-structures 𝔄 and 𝔅 are isomorphic, 𝔄 ≅
𝔅, if there exists an isomorphism 𝐹 ∶ 𝔄 ≅ 𝔅, that is, a bijection 𝐹 ∶
𝐴 → 𝐵 between the base sets of 𝔄 and 𝔅 which commutes with the
interpretations of the symbols in 𝜎ℒ , that is,
• 𝐹(𝑐𝔄 ) = 𝑐𝔅 for every constant symbol 𝑐 ∈ 𝒞 ℒ ,
• 𝐹(𝑓𝔄 (𝑎1 , … , 𝑎𝑛 )) = 𝑓𝔅 (𝐹(𝑎1 ), … , 𝐹(𝑎𝑛 )) for every function sym-
bol 𝑓 ∈ ℱ𝑛ℒ and every tuple (𝑎1 , … , 𝑎𝑛 ) ∈ 𝐴𝑛 ,
• (𝑎1 , … , 𝑎𝑛 ) ∈ 𝑅𝔄 ⟺ (𝐹(𝑎1 ), … , 𝐹(𝑎𝑛 )) ∈ 𝑅𝔅 for every predi-
cate 𝑅 ∈ ℛ𝑛ℒ and every tuple (𝑎1 , … , 𝑎𝑛 ) ∈ 𝐴𝑛 .
2.2. Terms and Formulas

A word over a set (alphabet) 𝐸 is a finite string 𝑤 = 𝑎0 𝑎1 ⋯ 𝑎𝑘−1 with
𝑎𝑖 ∈ 𝐸 for every 𝑖. We call 𝑘 the length of 𝑤, and we denote by 𝐸 ∗ the set
of words over 𝐸.1
Definition. Let ℒ be a language. The set 𝒯 ℒ of ℒ-terms is the smallest
subset 𝐷 of ℒ∗ containing the variables and the constant symbols such
that if 𝑓 ∈ ℱ𝑛ℒ and 𝑡1 , … , 𝑡𝑛 ∈ 𝐷, then 𝑓𝑡1 ⋯ 𝑡𝑛 ∈ 𝐷.
It is easy to see that 𝒯 ℒ = ⋃𝑛∈ℕ 𝒯𝑛ℒ , where 𝒯0ℒ = 𝒞 ℒ ∪ 𝒱 ℒ and,

inductively,
ℒ
𝒯𝑛+1 = 𝒯𝑛ℒ ∪ {𝑓𝑡1 ⋯ 𝑡𝑘 ∣ 𝑘 ∈ ℕ∗ , 𝑓 ∈ ℱ𝑘ℒ and 𝑡1 , … , 𝑡𝑘 ∈ 𝒯𝑛ℒ }.
Proposition 2.2.1 (Unique reading of terms). Any term 𝑡 ∈ 𝒯 ℒ satisfies
one and only one of the following:
(1) 𝑡 is a variable.
(2) 𝑡 is a constant symbol.
(3) There exists a unique integer 𝑛 ≥ 1, a unique 𝑛-ary function
symbol 𝑓 and a unique sequence (𝑡1 , … , 𝑡𝑛 ) of terms such that
𝑡 = 𝑓𝑡1 ⋯ 𝑡𝑛 .
Proof. Exercise. (One first proves by induction on the length of terms

that no proper initial segment of a term is a term.) □
Notation. For ease of reading, from now on we shall often write
𝑓(𝑡1 , … , 𝑡𝑛 ) instead of 𝑓𝑡1 ⋯ 𝑡𝑛 . When 𝑓 is binary, we sometimes write
𝑡1 𝑓𝑡2 instead of 𝑓𝑡1 𝑡2 . For instance, (𝑥 + 𝑦) ⋅ 𝑧 means ⋅ + 𝑥𝑦𝑧.
The height ht(𝑡) of a term 𝑡 is defined as the least natural number

𝑘 such that 𝑡 ∈ 𝒯𝑘ℒ . It follows from the unique reading property for
terms that ht(𝑓(𝑡1 ⋯ 𝑡𝑛 )) = 1 + max(ht(𝑡𝑖 )). This will allow us to give
definitions by induction on the height of terms.
Definition. An atomic ℒ-formula is
• either a word of the form 𝑡1 = 𝑡2 , where 𝑡1 , 𝑡2 are ℒ-terms,
1
We will continue to write ℕ∗ for the set of non-zero natural numbers. As we will never use ℕ as
an alphabet, this should not cause any ambiguity.
2.2. Terms and Formulas 37
• or a word of the form 𝑅𝑡1 ⋯ 𝑡𝑛 , where 𝑅 ∈ ℛ𝑛ℒ and all the 𝑡𝑖 are
ℒ-terms.
ℒ
The set Fml of ℒ-formulas is the smallest subset 𝐷 of ℒ∗ that contains
all atomic ℒ-formulas and such that if 𝑥 ∈ 𝒱 and 𝜑, 𝜓 ∈ 𝐷, then ¬𝜑,
(𝜑 ∧ 𝜓) and ∃𝑥𝜑 are all in 𝐷.
ℒ ℒ ℒ
It is easy to see that Fml = ⋃𝑛∈ℕ Fml𝑛 , where Fml0 is the set of
atomic formulas, and, inductively,
ℒ ℒ ℒ ℒ
Fml𝑛+1 = Fml𝑛 ∪ {¬𝜑 ∣ 𝜑 ∈ Fml𝑛 } ∪ {(𝜑 ∧ 𝜓) ∣ 𝜑, 𝜓 ∈ Fml𝑛 }
ℒ
∪ {∃𝑥𝜑 ∣ 𝑥 ∈ 𝒱 and 𝜑 ∈ Fml𝑛 }.
Proposition 2.2.2 (Unique reading of formulas). Any ℒ-formula 𝜑 sat-
isfies one and only one of the following:
(1) 𝜑 is an atomic formula.
(2) 𝜑 is equal to ¬𝜓 for some unique ℒ-formula 𝜓.
(3) 𝜑 is equal to (𝜓 ∧ 𝜒) for some unique ℒ-formulas 𝜓 and 𝜒.
(4) 𝜑 is equal to ∃𝑥𝜓 for some unique variable 𝑥 and some unique
ℒ-formula 𝜓.
Proof. Exercise. (One first proves by induction on the length of formu-

las that no proper initial segment of a formula is a formula.) □
If 𝜑 is a formula, its height ht(𝜑) is defined as the least natural num-

ℒ
ber 𝑘 such that 𝜑 ∈ Fml𝑘 . It follows from the unique reading property
for formulas that ht((𝜑 ∧ 𝜓)) = 1 + max(ht(𝜑), ht(𝜓)) and ht(∃𝑥𝜑) =
ht(¬𝜑) = 1 + ht(𝜑). This will allow us to give definitions by induction
on the height of formulas.
Definition.
(1) Let 𝑣𝑘 be a variable. One defines, by induction on the height
of a formula 𝜑, the notion of a free occurrence of 𝑣𝑘 in 𝜑:
• If 𝜑 is atomic, all occurrences of 𝑣𝑘 in 𝜑 are free.
• If 𝜑 is equal to ¬𝜓, the free occurences of 𝑣𝑘 in 𝜑 are those
in 𝜓.
• If 𝜑 is equal to (𝜓 ∧ 𝜒), the free occurrences of 𝑣𝑘 in 𝜑 are
those in 𝜓 and those in 𝜒.
• If 𝜑 is equal to ∃𝑣𝑙 𝜓 and 𝑙 ≠ 𝑘, the free occurrences of 𝑣𝑘

in 𝜑 are those in 𝜓.
• If 𝜑 is equal to ∃𝑣𝑘 𝜓, no occurrence of 𝑣𝑘 in 𝜑 is free.
(2) Occurrences of 𝑣𝑘 in 𝜑 which are not free are called bound.
(3) The free variables of 𝜑 are those having at least one free occur-
rence in 𝜑. One denotes by Free(𝜑) as the set of free variables
of 𝜑.
(4) A sentence is a formula with no free variable.
Example. If 𝜑 is the formula (∃𝑣0 𝑣0 < 𝑣1 ∧ 𝑣0 = 𝑣1 ), the first two

occurrences of 𝑣0 are bound, and the third is free. All occurrences of 𝑣1
in 𝜑 are free, thus Free(𝜑) = {𝑣0 , 𝑣1 }.
Notation. We will use the following abbreviations:

(𝜑 ∨ 𝜓) for ¬(¬𝜑 ∧ ¬𝜓) (disjunction)
(𝜑 → 𝜓) for ¬(𝜑 ∧ ¬𝜓) (implication)
(𝜑 ↔ 𝜓) for ((𝜑 → 𝜓) ∧ (𝜓 → 𝜑)) (equivalence)
∀𝑥𝜑 for ¬∃𝑥¬𝜑 (universal quantifier).
We shall write ∃𝑥1 , … , 𝑥𝑛 instead of ∃𝑥1 ⋯ ∃𝑥𝑛 (similarly for the uni-
versal quantifier), 𝑅(𝑡1 , … , 𝑡𝑛 ) instead of 𝑅𝑡1 ⋯ 𝑡𝑛 , and sometimes 𝑡1 𝑅𝑡2
instead of 𝑅𝑡1 𝑡2 .
𝑛
We shall write (𝜑0 ∧⋯∧𝜑𝑛 ) or sometimes ⋀𝑖=0 𝜑𝑖 instead of (⋯ ((𝜑0 ∧
𝜑1 ) ∧ 𝜑2 ) ∧ … ∧ 𝜑𝑛 ), similarly for ∨ instead of ∧.
Finally, to ease the reading of formulas, we shall sometimes add
parentheses, or we will omit them. In this case, we shall read formulas
according to the convention that symbols from {¬,∃,∀} bind the strongest,
followed by ∧, then ∨, and that symbols from {→, ↔} bind the weakest.
Thus, for instance ∀𝑥𝜑 ∧ 𝜓 → 𝜒 shall mean ((∀𝑥𝜑 ∧ 𝜓) → 𝜒), that
is, ¬((∀𝑥𝜑 ∧ 𝜓) ∧ ¬𝜒), and so finally ¬((¬∃𝑥¬𝜑 ∧ 𝜓) ∧ ¬𝜒).
Example. The field axioms can be expressed as first-order sentences in

ℒ𝑟𝑖𝑛𝑔 :
(1) ∀𝑥 𝑥 + 0 = 𝑥.
(2) ∀𝑥, 𝑦 𝑥 + 𝑦 = 𝑦 + 𝑥.
(3) ∀𝑥 (−𝑥) + 𝑥 = 0.
2.3. Semantics 39
(4) ∀𝑥, 𝑦, 𝑧 (𝑥 + 𝑦) + 𝑧 = 𝑥 + (𝑦 + 𝑧).

(5) ∀𝑥 𝑥 ⋅ 1 = 𝑥.
(6) ∀𝑥, 𝑦 𝑥 ⋅ 𝑦 = 𝑦 ⋅ 𝑥.
(7) ∀𝑥, 𝑦, 𝑧 (𝑥 ⋅ 𝑦) ⋅ 𝑧 = 𝑥 ⋅ (𝑦 ⋅ 𝑧).
(8) ∀𝑥, 𝑦, 𝑧 𝑥 ⋅ (𝑦 + 𝑧) = (𝑥 ⋅ 𝑦) + (𝑥 ⋅ 𝑧).
(9) ∀𝑥 (¬𝑥 = 0 → ∃𝑦 𝑥 ⋅ 𝑦 = 1).
(10) ¬0 = 1.
2.3. Semantics
We now explain how to interpret formulas in a given structure.
Definition. Let 𝔄 be an ℒ-structure.

(1) An assignment (with values in 𝔄) is a function 𝛼 ∶ 𝒱 → 𝐴 from
the set of variables to the base set of 𝔄.
(2) If 𝛼 is an assignment and 𝑡 is an ℒ-term, one defines 𝑡𝔄 [𝛼], by
induction on ht(𝑡), in the following way:
• 𝑣𝑖𝔄 [𝛼] = 𝛼(𝑣𝑖 ) (for 𝑣𝑖 ∈ 𝒱) and 𝑐𝔄 [𝛼] = 𝑐𝔄 (for 𝑐 ∈ 𝒞 ℒ ),
• 𝑓(𝑡1 , … , 𝑡𝑛 )𝔄 [𝛼] = 𝑓𝔄 (𝑡1𝔄 [𝛼], … , 𝑡𝑛𝔄 [𝛼]).
The proof of the following lemma is clear.
Lemma 2.3.1. If two assignments 𝛼 and 𝛽 coincide on all variables having

an occurrence in 𝑡, then 𝑡𝔄 [𝛼] = 𝑡𝔄 [𝛽]. □
Notation. If 𝑡 is a term, we shall sometimes denote it by 𝑡(𝑥1 , … , 𝑥𝑛 ) if

the variables 𝑥𝑖 are distinct and all variables having at least one occur-
rence in 𝑡 belong to the 𝑥𝑖 .
If a term 𝑡(𝑥1 , … , 𝑥𝑛 ) and elements 𝑎1 , … , 𝑎𝑛 ∈ 𝐴 are given, one de-
fines 𝑡𝔄 [𝑎1 , … , 𝑎𝑛 ] by 𝑡𝔄 [𝛼] with 𝛼 an assignment with 𝛼(𝑥𝑖 ) = 𝑎𝑖 for all
𝑖, which is well defined by the previous lemma.
Definition (Satisfaction of a formula). Let 𝔄 be an ℒ-structure. By in-

duction on the height of a formula 𝜑, we now define for any assignment
𝛼 the relation 𝔄 ⊧ 𝜑[𝛼] which will be read “𝜑 is satisfied in 𝔄 by 𝛼”:
𝔄 ⊧ 𝑡1 = 𝑡2 [𝛼] ∶ ⟺ 𝑡1𝔄 [𝛼] = 𝑡2𝔄 [𝛼]

𝔄 ⊧ 𝑅𝑡1 ⋯ 𝑡𝑛 [𝛼] ∶ ⟺ (𝑡1𝔄 [𝛼], … , 𝑡𝑛𝔄 [𝛼]) ∈ 𝑅𝔄
𝔄 ⊧ ¬𝜓[𝛼] ∶ ⟺ 𝔄 ⊭ 𝜓[𝛼] (that is, not 𝔄 ⊧ 𝜓[𝛼])
𝔄 ⊧ (𝜓 ∧ 𝜒)[𝛼] ∶ ⟺ 𝔄 ⊧ 𝜓[𝛼] and 𝔄 ⊧ 𝜒[𝛼]
𝔄 ⊧ ∃𝑥𝜓[𝛼] ∶ ⟺ there exists 𝑎 ∈ 𝐴 with 𝔄 ⊧ 𝜓[𝛼𝑎/𝑥 ].
Here, 𝛼𝑎/𝑥 denotes the assignment defined by 𝛼𝑎/𝑥 (𝑥) = 𝑎 and
𝛼𝑎/𝑥 (𝑦) = 𝛼(𝑦) for 𝑦 ≠ 𝑥.
Proposition 2.3.2. If two assignments 𝛼 and 𝛽 coincide on Free(𝜑), then
one has 𝔄 ⊧ 𝜑[𝛼] if and only if 𝔄 ⊧ 𝜑[𝛽].
Proof. By induction on ht(𝜑). The case of atomic formulas follows from

Lemma 2.3.1. In the induction step, only the case where 𝜑 is equal to
∃𝑥𝜓 deserves an argument. If 𝔄 ⊧ 𝜑[𝛼], there exists 𝑎 ∈ 𝐴 such that
𝔄 ⊧ 𝜓[𝛼𝑎/𝑥 ]. Any variable 𝑦 ≠ 𝑥 that is free in 𝜓 is also free in 𝜑. Hence
𝔄 ⊧ 𝜓[𝛽𝑎/𝑥 ] by the induction hypothesis, and so 𝔄 ⊧ 𝜑[𝛽]. □
Notation. A formula 𝜑 shall sometimes be denoted by 𝜑(𝑥1 , … , 𝑥𝑛 ) if
the variables 𝑥𝑖 are distinct and all free variables in 𝜑 belong to the 𝑥𝑖 .
If a formula 𝜑(𝑥1 , … , 𝑥𝑛 ) and elements 𝑎1 , … , 𝑎𝑛 ∈ 𝐴 are given, one
defines 𝔄 ⊧ 𝜑[𝑎1 , … , 𝑎𝑛 ] by 𝔄 ⊧ 𝜑[𝛼], where 𝛼 is an assignment with
𝛼(𝑥𝑖 ) = 𝑎𝑖 , which is well defined by the previous proposition.
Thus, 𝜑(𝑥1 , … , 𝑥𝑛 ) defines an 𝑛-ary relation on the structure 𝔄, given

by 𝜑[𝔄] ∶= {(𝑎1 , … , 𝑎𝑛 ) ∈ 𝐴𝑛 ∣ 𝔄 ⊧ 𝜑[𝑎1 , … , 𝑎𝑛 ]}.
In particular, when 𝜑 is a sentence, the relation
𝔄⊧𝜑
can be interpreted as “𝜑 is satisfied (or true) in 𝔄” or “𝔄 is a model of 𝜑”.
Example. An ℒ𝑟𝑖𝑛𝑔 -structure ⟨𝐾; 0, 1, +, −, ⋅⟩ is a field if and only if it
satisfies the field axioms (1)–(10) given previously.
Definition. Let 𝔄 be a structure and 𝐷 ⊆ 𝐴𝑛 .
• The set 𝐷 is called ∅-definable in 𝔄 if 𝐷 = 𝜑[𝔄] for some for-
mula 𝜑(𝑥1 , … , 𝑥𝑛 ).
• Let 𝐵 ⊆ 𝐴 be a parameter set. Then 𝐷 is called 𝐵-definable in 𝔄
if there exist a formula 𝜑(𝑥1 , … , 𝑥𝑛 , 𝑦1 , … , 𝑦𝑚 ) and 𝑏 ∈ 𝐵 𝑚 such
2.4. Substitution 41
that 𝐷 is equal to the set

𝜑[𝔄, 𝑏] ∶= {𝑎 ∈ 𝐴𝑛 ∣ 𝔄 ⊧ 𝜑[𝑎1 , … , 𝑎𝑛 , 𝑏1 , … , 𝑏𝑚 ]}.
Exercise 2.3.3.
(1) Let 𝔄 be an ℒ-structure and 𝐵 a non-empy subset of 𝐴 contain-
ing the interpretations 𝑐𝔄 of all constants in ℒ and which is
closed under all operations 𝑓𝔄 . Restricting the interpretations
of all symbols from the set 𝐴 to 𝐵 one obtains an ℒ-structure
𝔅, called substructure of 𝔄. We shall write 𝔅 ⊆ 𝔄 when 𝔅 is a
substructure of 𝔄.
Prove that the intersection of a family of substructures of
𝔄 is either a substructure or empty. Deduce that any non-
empty subset 𝑋 of 𝐴 is contained in a smallest substructure of
𝔄, called the substructure generated by 𝑋 and denoted by ⟨𝑋⟩𝔄 .
Prove that the base set of ⟨𝑋⟩𝔄 is given by
{ 𝑡𝔄 [𝛼] | 𝑡 ∈ 𝒯 ℒ and 𝛼 ∶ 𝒱 → 𝑋} .
(2) Let 𝔄 = ⟨𝐴; 0, 1, +, −, ⋅⟩ be an ℒ𝑟𝑖𝑛𝑔 -structure. One assumes
that 𝔄 is a ring. Prove that the substructure generated by a
subset 𝑋 of 𝐴 is the subring generated by 𝑋.
2.4. Substitution
The aim of this section is to provide a satisfactory notion for the substi-
tution in a formula 𝜑 of a variable 𝑥 by a term 𝑠, in such a way that the
expected semantical properties (cf. Proposition 2.4.2) are satisfied.
One might simply decide to replace every free occurrence of 𝑥 in 𝜑
by 𝑠, but it might then happen that some variable in 𝑠 becomes bound by
a quantifier in the resulting formula. This could have unpleasant con-
sequences, as for the formula 𝜑(𝑣0 ) given by ∃𝑣1 ¬𝑣1 = 𝑣0 , with 𝑥 = 𝑣0
and 𝑠 = 𝑣1 , for which one would get the formula ∃𝑣1 ¬𝑣1 = 𝑣1 which is
satisfied in no structure, while as soon as 𝔄 contains at least two distinct
elements, 𝔄 ⊧ 𝜑[𝑎] for any 𝑎 ∈ 𝐴.
The following somewhat more complicated definition remedies this
issue.
Definition. Let 𝑥0 , … , 𝑥𝑟 be distinct variables and 𝑠0 , … , 𝑠𝑟 be terms.

One defines the similtaneous substitution of the 𝑥𝑖 by the 𝑠𝑖 as follows.
(1) Let 𝑡 be a term. Then 𝑡𝑠0 /𝑥0 ,…,𝑠𝑟 /𝑥𝑟 = 𝑡𝑠/𝑥 is the word obtained
by “simulatenously replacing all occurrences of 𝑥𝑖 in 𝑡 by 𝑠𝑖 ”,
that is, one sets
𝑥 if 𝑥 ≠ 𝑥0 , … , 𝑥 ≠ 𝑥𝑟 ,
• 𝑥𝑠/𝑥 = {
𝑠𝑖 if 𝑥 = 𝑥𝑖 ,
• 𝑐𝑠/𝑥 = 𝑐,
1 𝑛
• [𝑓𝑡1 ⋯ 𝑡𝑛 ]𝑠/𝑥 = 𝑓𝑡𝑠/𝑥 ⋯ 𝑡𝑠/𝑥 , inductively.
(2) By induction on the height of a formula, one sets
′
• [𝑡 = 𝑡′ ]𝑠/𝑥 equal to 𝑡𝑠/𝑥 = 𝑡𝑠/𝑥 ,
1 𝑛 1 𝑛
• [𝑅𝑡 ⋯ 𝑡 ]𝑠/𝑥 equal to 𝑅𝑡𝑠/𝑥 ⋯ 𝑡𝑠/𝑥 ,
• [¬𝜓]𝑠/𝑥 equal to ¬[𝜓]𝑠/𝑥 , and finally
• (𝜓 ∧ 𝜒)𝑠/𝑥 equal to (𝜓𝑠/𝑥 ∧ 𝜒𝑠/𝑥 ).
• Let 𝑥𝑖1 , … , 𝑥𝑖𝑘 (𝑖1 < … < 𝑖𝑘 ) be those variables among 𝑥0 ,
… , 𝑥𝑟 that are free in ∃𝑥𝜓. (In particular one has
𝑥 ≠ 𝑥𝑖1 , … , 𝑥 ≠ 𝑥𝑖𝑘 .)
If 𝑥 has no occurrence in any of 𝑠𝑖1 , … , 𝑠𝑖𝑘 , one sets
[∃𝑥𝜓]𝑠/𝑥 equal to ∃𝑥[𝜓]𝑠𝑖 /𝑥𝑖1 ,…,𝑠𝑖 /𝑥𝑖 .

1 𝑘 𝑘
If 𝑥 has some occurrence in one of 𝑠𝑖1 , … , 𝑠𝑖𝑘 , one sets
(∗) [∃𝑥𝜓]𝑠/𝑥 equal to ∃𝑢[𝜓]𝑠𝑖 /𝑥𝑖1 ,…,𝑠𝑖 /𝑥𝑖 ,ᵆ/𝑥 ,

1 𝑘 𝑘
where 𝑢 is the first variable appearing in the enumera-

tion 𝑣0 , 𝑣1 , 𝑣2 , … which does not occur in any of the words
∃𝑥𝜓, 𝑠𝑖1 , … , 𝑠𝑖𝑘 .
Exercise 2.4.1. Prove the following by induction on height:

(1) If 𝑡 is a term, then 𝑡𝑠/𝑥 is a term.
(2) If 𝜑 is a formula, then 𝜑𝑠/𝑥 is a formula. Moreover, one has
ht(𝜑𝑠/𝑥 ) = ht(𝜑).
Let 𝑥0 , … , 𝑥𝑟 be distinct variables, 𝛼 an assignment with values in 𝔄

and 𝑎0 , … , 𝑎𝑟 elements of 𝐴. One defines the assignment 𝛼𝑎0 /𝑥0 ,…,𝑎𝑟 /𝑥𝑟 =
𝛼𝑎/𝑥 by 𝛼𝑎/𝑥 (𝑥𝑖 ) = 𝑎𝑖 and 𝛼𝑎/𝑥 (𝑦) = 𝛼(𝑦) if 𝑦 ≠ 𝑥𝑖 for every 𝑖.
2.4. Substitution 43
Proposition 2.4.2 (Substitution Lemma). Let 𝑥0 , … , 𝑥𝑟 be distinct vari-

ables, 𝑠0 , … , 𝑠𝑟 terms and 𝛼 an assignment with values in 𝔄.
𝔄
(1) For every term 𝑡 one has 𝑡𝑠/𝑥 [𝛼] = 𝑡𝔄 [𝛼𝑠𝔄0 [𝛼]/𝑥0 ,…,𝑠𝔄𝑟 [𝛼]/𝑥𝑟 ].
(2) For every formula 𝜑 one has
𝔄 ⊧ 𝜑𝑠/𝑥 [𝛼] if and only if 𝔄 ⊧ 𝜑 [𝛼𝑠𝔄0 [𝛼]/𝑥0 ,…,𝑠𝔄𝑟 [𝛼]/𝑥𝑟 ] .
Proof. (1) is easy. One proves (2) by induction on the height of 𝜑, the
only non-trivial case being when 𝜑 is of the form ∃𝑥𝜓. As in the defini-
tion of [∃𝑥𝜓]𝑠/𝑥 , one denotes by 𝑥𝑖1 , … , 𝑥𝑖𝑘 the variables that are exactly
those among the 𝑥𝑖 which are free in ∃𝑥𝜓. Assume first that 𝑥 occurs
in one of the terms 𝑠𝑖1 , … , 𝑠𝑖𝑘 , and let 𝑢 be the variable chosen as in (∗).
Then
𝔄 ⊧ [∃𝑥𝜓]𝑠/𝑥 [𝛼] ⟺ 𝔄 ⊧ ∃𝑢[𝜓]𝑠𝑖 /𝑥𝑖1 ,…,𝑠𝑖 /𝑥𝑖 ,ᵆ/𝑥 [𝛼]

1 𝑘 𝑘
⟺ there exists 𝑎 ∈ 𝐴 such that 𝔄 ⊧ 𝜓𝑠𝑖 /𝑥 ,…,𝑠𝑖 /𝑥𝑖 ,ᵆ/𝑥 [𝛼𝑎/ᵆ ]

1 𝑖1 𝑘 𝑘
⟺ (by the induction hypothesis) there exists 𝑎 ∈ 𝐴 such that

𝔄 ⊧ 𝜓[𝛼𝑠𝔄 [𝛼𝑎/𝑢 ]/𝑥𝑖 ,…,𝑠𝔄 𝔄 ]
𝑖1 1 𝑖 [𝛼𝑎/𝑢 ]/𝑥𝑖 ,ᵆ [𝛼𝑎/𝑢 ]/𝑥
𝑘 𝑘
⟺ there exists 𝑎 ∈ 𝐴 such that 𝔄 ⊧ 𝜓[𝛼𝑠𝔄 [𝛼]/𝑥𝑖 ,…,𝑠𝔄 ]

𝑖1 1 𝑖 [𝛼]/𝑥𝑖 ,𝑎/𝑥
𝑘 𝑘
(by Lemma 2.3.1, since 𝑢 does not occur in any of the 𝑠𝑖𝑗 )
⟺ 𝔄 ⊧ ∃𝑥𝜓[𝛼𝑠𝔄 [𝛼]/𝑥𝑖 ,…,𝑠𝔄 ]
𝑖1 1 𝑖 [𝛼]/𝑥𝑖
𝑘 𝑘
⟺ 𝔄 ⊧ ∃𝑥𝜓[𝛼𝑠𝔄0 [𝛼]/𝑥0 ,…,𝑠𝔄𝑟 [𝛼]/𝑥𝑟 ].
The last equivalence follows from Proposition 2.3.2, since the as-
signments 𝛼𝑠𝔄 [𝛼]/𝑥𝑖 ,…,𝑠𝔄 [𝛼]/𝑥𝑖 and 𝛼𝑠𝔄0 [𝛼]/𝑥0 ,…,𝑠𝔄𝑟 [𝛼]/𝑥𝑟 coincide on the set
𝑖1 1 𝑖𝑘 𝑘
Free(∃𝑥𝜓).
Assume now that the variable 𝑥 has no occurrence in any of the

terms 𝑠𝑖1 , … , 𝑠𝑖𝑘 . Then
𝔄 ⊧ [∃𝑥𝜓]𝑠/𝑥 [𝛼] ⟺ 𝔄 ⊧ ∃𝑥[𝜓]𝑠𝑖 /𝑥𝑖1 ,…,𝑠𝑖 /𝑥𝑖 [𝛼]

1 𝑘 𝑘
⟺ there exists 𝑎 ∈ 𝐴 with 𝔄 ⊧ 𝜓𝑠𝑖 /𝑥𝑖1 ,…,𝑠𝑖 /𝑥𝑖 [𝛼𝑎/𝑥 ]

1 𝑘 𝑘
⟺ (by the induction hypothesis and since 𝑥 ≠ 𝑥𝑖1 , … , 𝑥 ≠ 𝑥𝑖𝑘 )

there exists 𝑎 ∈ 𝐴 with 𝔄 ⊧ 𝜓[𝛼𝑠𝔄 [𝛼𝑎/𝑥 ]/𝑥𝑖 ,…,𝑠𝔄 ]
𝑖1 1 𝑖 [𝛼𝑎/𝑥 ]/𝑥𝑖 ,𝑎/𝑥
𝑘 𝑘
⟺ 𝔄 ⊧ ∃𝑥𝜓[𝛼𝑠𝔄 [𝛼]/𝑥𝑖 ,…,𝑠𝔄 ]

𝑖1 1 𝑖 [𝛼]/𝑥𝑖
𝑘 𝑘
⟺ 𝔄 ⊧ ∃𝑥𝜓[𝛼𝑠𝔄0 [𝛼]/𝑥0 ,…,𝑠𝔄𝑟 [𝛼]/𝑥𝑟 ]. □
Example 2.4.3.
(1) Let 𝜑(𝑣0 ) be the formula ∃𝑣1 ¬𝑣1 = 𝑣0 , and let 𝑠 = 𝑣1 . Then
𝜑𝑠/𝑣0 is equal to ∃𝑣2 [¬𝑣1 = 𝑣0 ]𝑣1 /𝑣0 ,𝑣2 /𝑣1 , that is, to ∃𝑣2 ¬𝑣2 = 𝑣1 .
(2) Let 𝜑 be a formula and 𝑥1 , … , 𝑥𝑛 be distinct variables. One as-
sumes that 𝑠1 , … , 𝑠𝑛 do not contain any variables having an oc-
currence in 𝜑. Then 𝜑𝑠/𝑥 is the word obtained by simultane-
ously replacing every free occurrence of 𝑥𝑖 by 𝑠𝑖 .
We prove this by induction on ht(𝜑), the only non-trivial
case being when 𝜑 is of the form ∃𝑥𝜓. Let 𝑖1 < … < 𝑖𝑘 be
chosen as in the definition of 𝜑𝑠/𝑥 . By assumption, 𝑥 has no
occurrence in any of 𝑠𝑖1 , … , 𝑠𝑖𝑘 , and so by definition [∃𝑥𝜓]𝑠/𝑥
equals ∃𝑥𝜓𝑠𝑖 /𝑥𝑖 ,…,𝑠𝑖 /𝑥𝑖 . As the 𝑠𝑖𝑗 do not contain any vari-
1 1 𝑘 𝑘
ables occurring in 𝜓, 𝜓𝑠𝑖 /𝑥𝑖 ,…,𝑠𝑖 /𝑥𝑖 is inductively given by si-
1 1 𝑘 𝑘
multaneously replacing every free occurrence of 𝑥𝑖𝑗 by 𝑠𝑖𝑗 , for
𝑗 = 1, … , 𝑘. This proves the result.
Lemma 2.4.4. If 𝑦 is a variable with no occurrence in 𝜑, then [𝜑𝑦/𝑥 ]𝑥/𝑦 is

equal to 𝜑.
Proof. The proof is by induction on ht(𝜑), the only non-trivial case be-
ing when 𝜑 is of the form ∃𝑧𝜓. If 𝑥 is not free in 𝜑, then [𝜑𝑦/𝑥 ]𝑥/𝑦 equals
𝜑𝑥/𝑦 by Example 2.4.3(2), and so is equal to 𝜑 by definition, as 𝑦 does
not occur in 𝜑 at all. We may thus assume that 𝑥 is free in ∃𝑧𝜓, so in
particular 𝑧 ≠ 𝑥 and 𝑧 ≠ 𝑦. Hence by definition [[∃𝑧𝜓]𝑦/𝑥 ]𝑥/𝑦 is equal to
2.5. Universally Valid Formulas 45
[∃𝑧[𝜓]𝑦/𝑥 ]𝑥/𝑦 , which is equal to ∃𝑧 [[𝜓]𝑦/𝑥 ]𝑥/𝑦 . But [𝜓𝑦/𝑥 ]𝑥/𝑦 equals 𝜓 by
the induction hypothesis. □
Notation. Let 𝑠1 , … , 𝑠𝑛 be terms. If 𝑡(𝑥1 , … , 𝑥𝑛 ) is a term, we shall often
write 𝑡(𝑠1 , … , 𝑠𝑛 ) for 𝑡𝑠1 /𝑥1 ,…,𝑠𝑛 /𝑥𝑛 , and 𝜑(𝑠1 , … , 𝑠𝑛 ) instead of 𝜑𝑠1 /𝑥1 ,…,𝑠𝑛 /𝑥𝑛
if 𝜑 is a formula of the form 𝜑(𝑥1 , … , 𝑥𝑛 ).
2.5. Universally Valid Formulas

Definition. An ℒ-formula 𝜑 is called universally valid, written ⊧ 𝜑, if it
is satisfied in every ℒ-structure 𝔄 and for any assignment 𝛼 with values
in 𝔄, that is, if 𝔄 ⊧ 𝜑[𝛼] for every 𝔄 and every 𝛼.
Remark 2.5.1. The formula 𝜑(𝑥1 , … , 𝑥𝑛 ) is universally valid if and only
if the sentence ∀𝑥1 , … , 𝑥𝑛 𝜑 is universally valid.
We call ∀𝑥1 , … , 𝑥𝑛 𝜑 a universal closure of 𝜑. We will write 𝔄 ⊧ 𝜑 for

𝔄 ⊧ ∀𝑥1 , … , 𝑥𝑛 𝜑.
Example.
(1) ∃𝑥 𝑥 = 𝑥 is universally valid. (By definition, a structure is non-
empty!)
(2) Let 𝜑 and 𝜓 be ℒ-formulas. Then (𝜑 → (𝜓 → 𝜑)) is universally
valid.
The fact that we do not mention the language ℒ in the notation ⊧ 𝜑

is justified by the next lemma. First some terminology.
′
Definition. Let ℒ ⊆ ℒ′ be two languages. If 𝔄′ = ⟨𝐴; (𝑍 𝔄 )𝑍∈𝜍ℒ′ ⟩ is an
′
ℒ′ -structure, we set 𝔄 ∶= 𝔄′ ↾ ℒ ∶= ⟨𝐴; (𝑍 𝔄 )𝑍∈𝜍ℒ ⟩, the reduct of 𝔄′ to
ℒ ⊆ ℒ′ . The structure 𝔄′ is called an expansion of 𝔄 to ℒ′ .
Lemma 2.5.2. Let 𝜑 be an ℒ-formula and ℒ′ ⊇ ℒ. Then 𝜑 is universally
valid as an ℒ-formula if and only if it is universally valid as an ℒ′ -formula.
Proof. One may identify assignments with values in 𝔄 and those with
values in 𝔄′ , and one has 𝔄′ ⊧ 𝜑[𝛼] ⟺ 𝔄 ⊧ 𝜑[𝛼] for any assignment
𝛼. Thus it suffices to prove that any ℒ-structure 𝔄 has an expansion to
some ℒ′ -structure, which is clear. □
We now specify a framework for formulas like (𝜑 → (𝜓 → 𝜑))

which are always satisfied independently of the truth values of 𝜑 and
of 𝜓, namely propositional calculus.
One fixes a set 𝒫 = {𝑝𝑖 ∣ 𝑖 ∈ ℕ}, where the 𝑝𝑖 are called proposi-
tional variables (they will take only “true” or “false” as values).
Propositional calculus formulas are defined as words over the alpha-
bet 𝒫 ∪ {¬, ∧, (, )}, formed according to the following rules:
• 𝑝𝑖 is a formula for any 𝑖;
• if 𝐹 and 𝐺 are formulas, then ¬𝐹 and (𝐹 ∧ 𝐺) are formulas, too.
Let Fml𝒫 denote the set of propositional calculus formulas. As be-
fore we introduce the abbreviations ∨, → and ↔.
One identifies ℤ/2 = {0, 1} with {“false”, “true”}, by assigning 0 to
“false” and 1 to “true”. An assignment is a function 𝛿 ∶ 𝒫 → {0, 1}.
Any assignment 𝛿 induces a function 𝛿∗ ∶ Fml𝒫 → {0, 1} defined
by 𝛿 (𝑝𝑖 ) = 𝛿(𝑝𝑖 ), 𝛿∗ (¬𝐹) = 1 − 𝛿∗ (𝐹) and 𝛿∗ ((𝐹 ∧ 𝐺)) = 𝛿∗ (𝐹) ⋅ 𝛿∗ (𝐺),
∗
inductively. If 𝛿∗ (𝐹) = 1, we also write 𝛿 ⊧ 𝐹.

One writes 𝐹 = 𝐹(𝑞1 , … , 𝑞𝑛 ) if the 𝑞𝑖 are distinct variables and all
variables occurring in 𝐹 belong to the 𝑞𝑖 .
Exercise 2.5.3. For any function 𝑔 ∶ {0, 1}𝑛 → {0, 1} there exists a propo-
sitional calculus formula 𝐹(𝑝1 , … , 𝑝𝑛 ) such that, for any assignment 𝛿,
one has 𝛿∗ (𝐹) = 𝑔(𝛿(𝑝1 ), … , 𝛿(𝑝𝑛 )).
Definition.
(1) One says that a formula 𝐹 ∈ Fml𝒫 is a tautology for the propo-
sitional calculus if 𝛿 ⊧ 𝐹 for any assignment 𝛿. This can be
rephrased by saying that 𝐹 = 𝐹(𝑞1 , … , 𝑞𝑛 ) is true for any choice
of truth values for the 𝑞𝑖 .
(2) An ℒ-formula 𝜑 is a tautology for the predicate calculus if there
exists 𝐹 = 𝐹(𝑞1 , … , 𝑞𝑛 ) a tautology for the propositional calcu-
lus and ℒ-formulas 𝜓1 , … , 𝜓𝑛 such that 𝜑 equals 𝐹𝜓1 /𝑞1 ,…,𝜓𝑛 /𝑞𝑛
(the word obtained by replacing the variables 𝑞𝑖 by 𝜓𝑖 ).
Lemma 2.5.4. Tautologies for the predicate calculus are universally valid
formulas.
2.5. Universally Valid Formulas 47
Proof. Clear. □
Lemma 2.5.5 (Equality axioms). The following sentences are universally
valid:
• ∀𝑣0 𝑣0 = 𝑣0 (reflexivity, E1)
• ∀𝑣0 , 𝑣1 (𝑣0 = 𝑣1 → 𝑣1 = 𝑣0 ) (symmetry, E2)
• ∀𝑣0 , 𝑣1 , 𝑣2 ((𝑣0 = 𝑣1 ∧ 𝑣1 = 𝑣2 ) → 𝑣0 = 𝑣2 ) (transitivity, E3)
𝑛
• ∀𝑣1 , … , 𝑣2𝑛 (⋀𝑖=1 𝑣𝑖 = 𝑣𝑖+𝑛 → 𝑓𝑣1 ⋯ 𝑣𝑛 = 𝑓𝑣𝑛+1 ⋯ 𝑣2𝑛 )
for any 𝑓 ∈ ℱ𝑛ℒ (congruence for functions, E4)
𝑛
• ∀𝑣1 , … , 𝑣2𝑛 ((⋀𝑖=1 𝑣𝑖 = 𝑣𝑖+𝑛 ∧ 𝑅𝑣1 ⋯ 𝑣𝑛 ) → 𝑅𝑣𝑛+1 ⋯ 𝑣2𝑛 )
for any 𝑅 ∈ ℛ𝑛ℒ (congruence for relations, E5).
Proof. Clear. □
Remark. The sentences (E1-E5) express that = is a congruence relation,
that is, an equivalence relation compatible with the functions and the
relations of the language.
Lemma 2.5.6 (Quantifier axioms). Let 𝜑 and 𝜓 be ℒ-formulas.
(1) For every variable 𝑥 which is not free in 𝜑, the formula
(Q1) ∀𝑥 (𝜑 → 𝜓) → (𝜑 → ∀𝑥 𝜓)
is univerally valid.
(2) For every variable 𝑥 and every term 𝑡, the formula
(Q2) 𝜑𝑡/𝑥 → ∃𝑥 𝜑
is universally valid.
(3) For every variable 𝑥 the formula
(Q3) ∃𝑥 𝜑 ↔ ¬∀𝑥 ¬𝜑
is universally valid.
Proof. (2) follows easily from the Substitution Lemma (Propo-

sition 2.4.2), and (3) is clear.
We will now prove (1). Suppose that 𝑥 ∉ Free(𝜑), and that 𝔄 ⊧
∀𝑥 (𝜑 → 𝜓)[𝛼], that is, 𝔄 ⊧ ¬∃𝑥 (𝜑 ∧ ¬𝜓)[𝛼], from which we infer 𝔄 ⊧
(¬𝜑 ∨ 𝜓)[𝛼𝑎/𝑥 ] for any 𝑎 ∈ 𝐴. We have to prove that 𝔄 ⊧ (𝜑 → ∀𝑥 𝜓)[𝛼],
and for this we may assume that 𝔄 ⊧ 𝜑[𝛼]. But then 𝔄 ⊧ 𝜑[𝛼𝑎/𝑥 ] by
Proposition 2.3.2, as 𝑥 ∉ Free(𝜑), and so 𝔄 ⊧ 𝜓[𝛼𝑎/𝑥 ]. Since 𝑎 ∈ 𝐴 was
arbitrary, 𝔄 ⊧ ∀𝑥 𝜓[𝛼] follows. □
Remark. The restriction on the variable 𝑥 in (Q1) is necessary, as shown
by the following example. Let both 𝜑 and 𝜓 be equal to the formula
𝑥 = 𝑐, where 𝑐 is a constant symbol. Then ⊧ ∀𝑥(𝜑 → 𝜓), but in any
structure 𝔄 with a base set that contains at least 2 elements, one has
𝔄 ⊭ ∀𝑥 (𝜑 → ∀𝑥 𝜓).
Definition. An ℒ-theory is a set of ℒ-sentences.
Let 𝑇 be an ℒ-theory.
• One says that the ℒ-structure 𝔄 is a model of 𝑇, which we de-
note by 𝔄 ⊧ 𝑇, if 𝔄 ⊧ 𝜑 for any 𝜑 ∈ 𝑇.
• A sentence (or more generally a formula) 𝜑 is a logical conse-
quence of 𝑇, denoted by 𝑇 ⊧ 𝜑, if for any ℒ-structure 𝔄 which
is a model of 𝑇 we have 𝔄 ⊧ 𝜑 (that is, in case of a formula
𝜑(𝑥1 , … , 𝑥𝑛 ), we require 𝔄 ⊧ ∀𝑥1 , … , 𝑥𝑛 𝜑).
By the proof of Lemma 2.5.2, the fact that 𝑇 ⊧ 𝜑 is independent of

the language ℒ.
2.6. Formal Proofs and Gödel’s Completeness Theorem

In this section we will prove Gödel’s Completeness Theorem, which
states that any universally valid ℒ-formula may be obtained using a
finite deduction (a formal proof). This fundamental result establishes a
perfect correspondence between semantic truth and syntactic provabil-
ity in first-order logic.
We will start by formalizing the notion of proof.
By the logical axioms we mean:
• predicate calculus tautologies,
• the equality axioms (E1-E5) from Lemma 2.5.5,
• the quantifier axioms (Q1-Q3) from Lemma 2.5.6.
There will be two deduction rules:
2.6. Formal Proofs and Gödel’s Completeness Theorem 49
• (MP) Modus Ponens: From 𝜑 and 𝜑 → 𝜓, one can deduce 𝜓.

• Generalization: From 𝜑, one can deduce ∀𝑥 𝜑 (𝑥 any variable).
Definition.
(1) Let 𝜑 be an ℒ-formula and 𝑇 be an ℒ-theory. A formal proof
of 𝜑 in 𝑇 is a finite sequence of ℒ-formulas (𝜑0 , … , 𝜑𝑛 ) with 𝜑𝑛
equal to 𝜑 such that, for every 𝑖 ≤ 𝑛, either 𝜑𝑖 ∈ 𝑇, or 𝜑𝑖 is a
logical axiom, or it can be deduced by (MP) from some 𝜑𝑗 and
𝜑𝑘 with 𝑗, 𝑘 < 𝑖, or it can be obtained by generalization from a
formula 𝜑𝑗 with 𝑗 < 𝑖.
(2) One says that 𝜑 is provable in 𝑇, denoted by 𝑇 ⊢ℒ 𝜑, if there
exists a formal proof of 𝜑 in 𝑇. We write ⊢ℒ 𝜑 if 𝜑 is provable
in the empty theory.
A priori, ⊢ℒ depends on the language ℒ in which the formula 𝜑

and the theory 𝑇 are considered, since if ℒ′ ⊇ ℒ, there are more ℒ′ -
proofs than ℒ-proofs. However we shall see that 𝑇 ⊧ 𝜑 if and only if
𝑇 ⊢ℒ 𝜑, which shows in particular that the provability of a formula is
independent of the language.
In the sequel, we shall consider proofs (𝜑0 , … , 𝜑𝑛 ) in which formulas
𝜑𝑖 whose provability has been previously established will be considered
as axioms. It is always possible to transform such a proof into a formal
proof; it is enough to replace 𝜑𝑖 by a formal proof of 𝜑𝑖 .
Example 2.6.1.
(1) If 𝜑 and 𝜓 are ℒ-sentences, then {𝜑, 𝜓} ⊢ℒ 𝜑 ∧ 𝜓.
Now let 𝜑 be an ℒ-formula, 𝑥, 𝑦 variables and 𝑡 an ℒ-term. Then we
have the following:
(2) ⊢ℒ ∀𝑥 𝜑 → 𝜑𝑡/𝑥 .
(3) If 𝑦 does not occur in 𝜑, then ⊢ℒ ∀𝑦 𝜑𝑦/𝑥 → ∀𝑥 𝜑.
(4) ⊢ℒ ∀𝑥 𝜑 → 𝜑.
Proof. (1) From 𝜑 and the tautology 𝜑 → (𝜓 → (𝜑 ∧ 𝜓)) one deduces

𝜓 → (𝜑∧𝜓) by (MP), from which one deduces 𝜑∧𝜓, using 𝜓 and applying
(MP) again.
(2) Here is a formal proof of ∀𝑥 𝜑 → 𝜑𝑡/𝑥 (recall that by definition

∀𝑥 𝜑 is the formula ¬∃𝑥¬𝜑):
𝜑0 equals (¬𝜑𝑡/𝑥 → ∃𝑥 ¬𝜑) (Q2)
𝜑1 equals (¬𝜑𝑡/𝑥 → ∃𝑥 ¬𝜑) → (¬∃𝑥 ¬𝜑 → 𝜑𝑡/𝑥 ) (tautology)
𝜑2 equals ∀𝑥 𝜑 → 𝜑𝑡/𝑥 (MP).
(3) The formula ∀𝑦 𝜑𝑦/𝑥 → [𝜑𝑦/𝑥 ]𝑥/𝑦 is an instance of (2). Since
[𝜑𝑦/𝑥 ]𝑥/𝑦 equals 𝜑 by Lemma 2.4.4, we get ∀𝑥 (∀𝑦 𝜑𝑦/𝑥 → 𝜑) by general-
ization, from which we conclude by the quantifier axiom (Q1) and (MP).
(4) is a special case of (2), since 𝜑 is equal to 𝜑𝑥/𝑥 . □
Lemma 2.6.2 (Soundness). Let 𝑇 be an ℒ-theory and 𝜑 an ℒ-formula.
Then
𝑇 ⊢ℒ 𝜑 ⇒ 𝑇 ⊧ 𝜑.
Proof. The logical axioms are universally valid by Lemma 2.5.4, Lemma
2.5.5 and Lemma 2.5.6. Clearly, if 𝑇 ⊧ 𝜑 and 𝑇 ⊧ 𝜑 → 𝜓, then 𝑇 ⊧ 𝜓. It is
also easy to see that 𝑇 ⊧ 𝜑 implies 𝑇 ⊧ ∀𝑥 𝜑. By induction on the length
of a formal proof, we are done. □
Definition. Let 𝑇 be an ℒ-theory.
• 𝑇 is inconsistent if there exists an ℒ-sentence 𝜑 such that 𝑇 ⊢ℒ
𝜑 and 𝑇 ⊢ℒ ¬𝜑.
• 𝑇 is consistent if it is not inconsistent.
• 𝑇 is complete if it is consistent and for any ℒ-sentence 𝜑 either
𝑇 ⊢ℒ 𝜑 or 𝑇 ⊢ℒ ¬𝜑.
Example 2.6.3.
(1) Let 𝔄 be an ℒ-structure. Then the set
Th(𝔄) ∶= {𝜑 is ℒ-sentence ∣ 𝔄 ⊧ 𝜑}
is a theory, called the theory of 𝔄. It is a complete theory (its
consistency follows from Lemma 2.6.2).
(2) The theory of algebraically closed fields is the ℒ𝑟𝑖𝑛𝑔 -theory ACF
which is composed of the field axioms together with a sentence
𝜒𝑛 for any 𝑛 ≥ 1 expressing that any polynomial of degree 𝑛 has
a root. For instance, the formula 𝜒𝑛 given by ∀𝑧0 ,…,𝑧𝑛−1 ∃𝑥(𝑥𝑛 +
𝑧𝑛−1 𝑥𝑛−1 +… + 𝑧0 = 0) works; that formula can be rewritten in

ℒ𝑟𝑖𝑛𝑔 .
Remark 2.6.4.
(1) An ℒ-theory 𝑇 is inconsistent if and only if 𝑇 ⊢ℒ 𝜑 for any
ℒ-formula 𝜑.
(2) If 𝑇 ⊢ℒ 𝜑, there exists 𝑇0 ⊆ 𝑇 finite such that 𝑇0 ⊢ℒ 𝜑.
Corollary 2.6.5.
(1) Let 𝑇 be an ℒ-theory such that all finite subsets of 𝑇 are consis-
tent. Then 𝑇 is consistent, too.
(2) Let (𝑇𝑖 )𝑖∈𝐼 be a family of consistent ℒ-theories, with 𝑇𝑖 ⊆ 𝑇𝑗 or
𝑇𝑗 ⊆ 𝑇𝑖 for any 𝑖, 𝑗 ∈ 𝐼. Then 𝑇 = ⋃𝑖∈𝐼 𝑇𝑖 is consistent, too. □
Lemma 2.6.6 (Deduction Lemma). Let 𝜒 be an ℒ-sentence, 𝑇 an ℒ-
theory and 𝜑 an ℒ-formula. Then
𝑇 ∪ {𝜒} ⊢ℒ 𝜑 if and only if 𝑇 ⊢ℒ 𝜒 → 𝜑.
Proof. It is clear that 𝑇 ⊢ℒ 𝜒 → 𝜑 implies that 𝑇∪{𝜒} ⊢ℒ 𝜑. Conversely,

let (𝜑0 , … , 𝜑𝑛 ) be a formal proof of 𝜑 in 𝑇∪{𝜒}. By induction on 𝑖, we shall
prove that 𝑇 ⊢ℒ (𝜒 → 𝜑𝑖 ). If 𝜑𝑖 equals 𝜒, this is clear, and if 𝑇 ⊢ℒ 𝜑𝑖 , it
follows from (MP) and the fact that (𝜑𝑖 → (𝜒 → 𝜑𝑖 )) is a tautology. This
implies that 𝑇 ⊢ℒ (𝜒 → 𝜑𝑖 ) when 𝜑𝑖 is a logical axiom or an element of
𝑇.
If 𝜑𝑖 is deduced by (MP) applied to 𝜑𝑗 and 𝜑𝑘 for 𝑗, 𝑘 < 𝑖, so 𝜑𝑘
equals 𝜑𝑗 → 𝜑𝑖 , it is enough to use (MP) together with the fact that ((𝜒 →
𝜑𝑗 ) ∧ (𝜒 → (𝜑𝑗 → 𝜑𝑖 )) → (𝜒 → 𝜑𝑖 )) is a tautology.
Finally, if 𝜑𝑖 is obtained by generalization from 𝜑𝑗 for 𝑗 < 𝑖, so 𝜑𝑖
equals ∀𝑥𝜑𝑗 , we have 𝑇 ⊢ℒ 𝜒 → 𝜑𝑗 by induction, from which we deduce
𝑇 ⊢ℒ ∀𝑥 (𝜒 → 𝜑𝑗 ) by generalization. Now, 𝜒 being a sentence, the
formula ∀𝑥 (𝜒 → 𝜑) → (𝜒 → ∀𝑥 𝜑) is a quantifier axiom, so we conclude
that 𝑇 ⊢ℒ 𝜒 → ∀𝑥 𝜑 by (MP). □
Corollary 2.6.7. Let 𝑇 be an ℒ-theory and 𝜑 an ℒ-sentence. Then 𝑇 ⊢ℒ 𝜑
if and only if 𝑇 ∪ {¬𝜑} is inconsistent.
Proof. The ‘only if’ part is clear. Conversely, if 𝑇 ∪ {¬𝜑} is inconsistent,

then 𝑇 ∪ {¬𝜑} ⊢ℒ 𝜑 by Remark 2.6.4(1). It follows from the Deduction
Lemma that 𝑇 ⊢ℒ ¬𝜑 → 𝜑. Since (¬𝜑 → 𝜑) → 𝜑 is a tautology, we
conclude by (MP). □
Lemma 2.6.8 (Simulation of constants by variables). Let 𝜓 be an ℒ-
formula, 𝑇 an ℒ-theory, and let 𝐶 be a set of constant symbols such that
ℒ ∩ 𝐶 = ∅.
(a) Let 𝑥 be a variable and 𝑐 ∈ 𝐶. The following are equivalent:
(1) 𝑇 ⊢ℒ 𝜓,
(2) 𝑇 ⊢ℒ∪{𝑐} 𝜓𝑐/𝑥 ,
(3) 𝑇 ⊢ℒ∪{𝑐} 𝜓.
(b) One has 𝑇 ⊢ℒ 𝜓 if and only if 𝑇 ⊢ℒ∪𝐶 𝜓.
Proof. For ease of notation we shall assume 𝑇 = ∅; the proof for general
𝑇 is just the same. Let us start by proving (a).
(1)⇒(3) is clear since any ℒ-proof is an ℒ ∪ {𝑐}-proof.
(3)⇒(2). If ⊢ℒ∪{𝑐} 𝜓, then generalization yields ⊢ℒ∪{𝑐} ∀𝑥𝜓. Using
Example 2.6.1(2) and (MP), one deduces ⊢ℒ∪{𝑐} 𝜓𝑐/𝑥 .
(2)⇒(1). For 𝜑̃ an ℒ ∪ {𝑐}-formula and 𝑦 a variable not occurring
in 𝜑,̃ we denote by 𝜑 and also by 𝜑𝑦/𝑐 ̃ the word obtained by substituting
every occurrence of 𝑐 in 𝜑̃ by 𝑦.
Let 𝜑,̃ 𝜓 ̃ and 𝜒̃ be ℒ ∪ {𝑐}-formulas with no occurrence of 𝑦. Then
the following statements hold:
(i) 𝜑 is an ℒ-formula.
(ii) 𝜑𝑐/𝑦 is equal to 𝜑.̃
(iii) If 𝜑̃ is an equality axiom (respectively a tautology or a quanti-
fier axiom), then 𝜑 is so, too.
(iv) If 𝜑̃ is obtained using (MP) from 𝜓 ̃ and 𝜒,̃ then 𝜑 is obtained
using (MP) from 𝜓 and 𝜒.
(v) If 𝜑̃ is obtained using generalization from 𝜓,̃ then 𝜑 is obtained
using generalization from 𝜓.
One proves (i) by induction on the height of 𝜑, and (ii) follows from
Example 2.4.3(2). The statements (iv) and (v) are immediate, and also
the case of an equality axiom, a tautology or an instance of (Q1) or (Q3)

in (iii). It remains to check the case of an instance of (Q2) in (iii): If 𝜑̃
̃ → ∃𝑥𝛿),
is of the form (𝛿𝑡/𝑥 ̃ one has to prove that the formula 𝜑, which
̃ ]𝑦/𝑐 → ∃𝑥[𝛿]̃ 𝑦/𝑐 ), is an instance of (Q2). This follows
is given by ([𝛿𝑡/𝑥
̃ ]𝑦/𝑐 and [𝛿𝑦/𝑐
from the fact that [𝛿𝑡/𝑥 ̃ ]𝑡 /𝑥 are equal, which is proved by
𝑦/𝑐
induction on height.
Now, let (𝜑0̃ , … , 𝜑𝑚
̃ ) be a formal proof of 𝜓𝑐/𝑥 in the language ℒ∪{𝑐}.
Choose a variable 𝑦 not occurring in any of 𝜑0̃ , … , 𝜑𝑚 ̃ and such that 𝑦 ≠
𝑥.
One deduces from (i)-(v) that (𝜑0 , … , 𝜑𝑚 ) is a formal proof in the
language ℒ of [𝜓𝑐/𝑥 ]𝑦/𝑐 , which is equal to 𝜓𝑦/𝑥 by Example 2.4.3(2).
By generalization one gets ⊢ℒ ∀𝑦 𝜓𝑦/𝑥 . Using Example 2.6.1(3,4)
and (MP), one gets ⊢ℒ 𝜓, which completes the proof of (2)⇒(1).
Item (b) is a direct consequence of (a). Indeed, as any ℒ ∪ 𝐶-proof
involves only a finite number of elements of 𝐶, one may assume that
𝐶 is finite. One concludes by induction on the cardinality of 𝐶, using
(1)⇔(3). □
Lemma 2.6.9. Let 𝑇 be an ℒ-theory, 𝜑(𝑥) an ℒ-formula and 𝑐 ∈ ℒ a

constant symbol not occurring in 𝑇 ∪ {𝜑(𝑥)}. Assume that 𝑇 is consistent.
Then 𝑇 ∪ {∃𝑥𝜑 → 𝜑𝑐/𝑥 } is a consistent ℒ-theory.
Proof. If the conclusion does not hold, then 𝑇 ⊢ℒ ∃𝑥 𝜑∧¬𝜑𝑐/𝑥 by Corol-

lary 2.6.7. In particular, 𝑇 ⊢ℒ ∃𝑥 𝜑, and, by Lemma 2.6.8(a), 𝑇 ⊢ℒ ¬𝜑,
so 𝑇 ⊢ℒ ∀𝑥 ¬𝜑 by generalization. Using the quantifier axiom (Q3) and
(MP), it follows that 𝑇 is inconsistent. □
Theorem 2.6.10 (Gödel’s Completeness Theorem). Let 𝑇 be an ℒ-theory

and 𝜑 an ℒ-sentence. Then 𝑇 ⊧ 𝜑 if and only if 𝑇 ⊢ℒ 𝜑.
Remark. Since ⊧ does not depend on the language ℒ, a posteriori ⊢ℒ

does not depend on ℒ either. Thus, once Gödel’s Completeness Theorem
is proved, we shall write ⊢ instead of ⊢ℒ .
Proof. 𝑇 ⊢ℒ 𝜑 ⇒ 𝑇 ⊧ 𝜑. This is the easy direction in Gödel’s theorem

and it expresses that our notion of formal proof is sound. It holds by
Lemma 2.6.2.
𝑇 ⊧ 𝜑 ⇒ 𝑇 ⊢ℒ 𝜑. This is the non-trivial direction. It will follow

from the next theorem, which is an existence statement for models. In
fact, as
𝑇 ⊭ 𝜑 ⟺ 𝑇 ∪ {¬𝜑} has a model (by definition) and
𝑇 ⊬ℒ 𝜑 ⟺ 𝑇 ∪ {¬𝜑} is consistent (by Corollary 2.6.7),
this theorem is equivalent to the Completeness Theorem. □
Theorem 2.6.11. A theory has a model if and only if it is consistent.
Proof. The easy direction was already proved. To prove that any con-
sistent theory 𝑇 has a model, we will start by constructing an expansion
𝑇 + of 𝑇 in some language ℒ+ ⊇ ℒ which has better properties. Let us
begin with a definition.
Definition. Let ℒ be a language and 𝐶 a set of constant symbols with
ℒ ∩ 𝐶 = ∅. One says that an ℒ ∪ 𝐶-theory 𝑇 + admits Henkin witnesses
in 𝐶 if for any ℒ ∪ 𝐶-formula 𝜑(𝑥) there exists 𝑐 ∈ 𝐶 such that ∃𝑥𝜑 →
𝜑𝑐/𝑥 ∈ 𝑇 + .
If 𝔄 is an ℒ-structure and 𝐴 = {𝑎𝑐 ∣ 𝑐 ∈ 𝐶} is an enumeration

(possibily non-injective) of its base set by 𝐶, one denotes by 𝔄+ the ℒ ∪
𝐶-structure obtained from 𝔄 by interpreting 𝑐 by 𝑎𝑐 . Then Th(𝔄+ ) is
a complete theory which admits Henkin witnesses in 𝐶. In fact, any
complete theory which admits Henkin witnesses in 𝐶 is of this form:
Proposition 2.6.12. Any complete ℒ∪𝐶-theory 𝑇 + which admits Henkin
witnesses in 𝐶 has a model 𝔄+ consisting of constants of 𝐶, that is, with a
+
base set of the form 𝐴+ = {𝑐𝔄 ∣ 𝑐 ∈ 𝐶}.
Proof of Proposition 2.6.12. We observe that replacing 𝑇 + by the set

of ℒ ∪ 𝐶-sentences 𝜑 such that 𝑇 + ⊢ℒ∪𝐶 𝜑 does not change the as-
sumptions. We may thus assume that 𝑇 + is deductively closed, that is,
𝑇 + ⊢ℒ∪𝐶 𝜑 if and only if 𝜑 ∈ 𝑇 + .
Note that it follows from the equality axioms that the binary relation
𝑐 ∼ 𝑑 ∶ ⟺ 𝑐 = 𝑑 ∈ 𝑇 + on 𝐶 is an equivalence relation. We set
𝑎𝑐 ∶= 𝑐/∼ and 𝐴+ ∶= {𝑎𝑐 ∣ 𝑐 ∈ 𝐶}. Then 𝐴+ ≠ ∅, as clearly 𝐶 ≠ ∅. Let
us define an ℒ ∪ 𝐶-structure 𝔄+ on 𝐴+ as follows:
+
• For 𝑑 a constant symbol, we set 𝑑𝔄 ∶= 𝑎𝑐 if 𝑐 = 𝑑 ∈ 𝑇 + ,
where 𝑐 ∈ 𝐶. Note that such a 𝑐 always exists. Indeed, 𝑑 =
𝑑 → ∃𝑥 𝑥 = 𝑑 ∈ 𝑇 + by (Q2), thus ∃𝑥 𝑥 = 𝑑 ∈ 𝑇 + by (MP);
since 𝑇 + admits Henkin witnesses in 𝐶, there exists 𝑐 ∈ 𝐶 such
that ∃𝑥 𝑥 = 𝑑 → 𝑐 = 𝑑 ∈ 𝑇 + , whence 𝑐 = 𝑑 ∈ 𝑇 + by (MP).
+
Furthermore by the equality axioms 𝑑𝔄 does not depend on
the choice of 𝑐.
• For 𝑅 ∈ ℛ𝑛ℒ we set
+
(𝑎𝑐1 , … , 𝑎𝑐𝑛 ) ∈ 𝑅𝔄 ∶ ⟺ 𝑅𝑐1 ⋯ 𝑐𝑛 ∈ 𝑇 + .
This is well defined by the equality axioms (E5).
• For 𝑓 ∈ ℱ𝑛ℒ we set
+
𝑓𝔄 (𝑎𝑐1 , … , 𝑎𝑐𝑛 ) = 𝑎𝑐0 ∶ ⟺ 𝑓𝑐1 ⋯ 𝑐𝑛 = 𝑐0 ∈ 𝑇 + .
This is well defined by the equality axioms (E4). By an argu-
+
ment similar to the one given for constants one proves that 𝑓𝔄
is defined everywhere as a function.
The following statements then hold:
(I) If 𝑡 is an ℒ ∪ 𝐶-term without variables and 𝑐 ∈ 𝐶, then
+
𝑡 𝔄 = 𝑎𝑐 ⟺ 𝑡 = 𝑐 ∈ 𝑇 + .
(II) Let 𝜓 be an ℒ ∪ 𝐶-sentence. Then 𝔄+ ⊧ 𝜓 ⟺ 𝜓 ∈ 𝑇 + .
One proves (I) by induction on ht(𝑡), using the equality axioms.
To prove (II), first recall that ht(𝜓) = ht(𝜓𝑠/𝑥 ) (cf. Exercise 2.4.1).
One argues by induction on ht(𝜓).
If 𝜓 is of the form 𝑡1 = 𝑡2 , then (II) follows from (I). If 𝜓 is of the form
+
𝑅𝑡1 ⋯ 𝑡𝑛 , one chooses 𝑐𝑖 ∈ 𝐶 such that 𝑡𝑖𝔄 = 𝑎𝑐𝑖 . Since 𝑡𝑖 = 𝑐𝑖 ∈ 𝑇 + , one
has 𝔄+ ⊧ 𝜓 ⟺ 𝑅𝑐1 ⋯ 𝑐𝑛 ∈ 𝑇 + ⟺ 𝜓 ∈ 𝑇 + . The first equivalence
follows from the definition of 𝔄+ , the second from (E5).
If 𝜓 is equal to (𝜑1 ∧ 𝜑2 ) it is clear that (II) for 𝜓 follows from (II) for
both 𝜑1 and 𝜑2 . If 𝜓 is equal to ¬𝜑, one has
𝔄+ ⊧ 𝜓 ⟺ 𝔄 + ⊭ 𝜑 ⟺ 𝜑 ∉ 𝑇 + ⟺ 𝜓 ∈ 𝑇 + ,
since 𝑇 + is complete and deductively closed by hypothesis.
Finally, suppose that 𝜓 is equal to ∃𝑥𝜑. Let us note that 𝜓 ∈ 𝑇 + if

and only if there exists 𝑐 ∈ 𝐶 such that 𝜑𝑐/𝑥 ∈ 𝑇 + . Indeed, since 𝑇 +
admits Henkin witnesses in 𝐶, the ‘only if’ part holds. For the ‘if’ part,
it is enough to use (Q2), (MP) and the fact that 𝑇 + is deductively closed.
One may now conclude as follows:
𝔄+ ⊧ 𝜓 ⟺ 𝔄+ ⊧ 𝜑[𝑎𝑐 ] for some 𝑐 ∈ 𝐶
(Substitution Lemma) ⟺ 𝔄+ ⊧ 𝜑𝑐/𝑥 for some 𝑐 ∈ 𝐶
(Induction Hypothesis) ⟺ 𝜑𝑐/𝑥 ∈ 𝑇 + for some 𝑐 ∈ 𝐶
⟺ 𝜓 ∈ 𝑇 +.
Thus 𝔄+ is a model of 𝑇 + consisting of constants in 𝐶. □
The following lemma will be enough to conclude the proof of Theo-

rem 2.6.11, because if 𝔄+ is a model of the ℒ+ = ℒ ∪ 𝐶-theory 𝑇 + with
𝑇 + ⊇ 𝑇, then the reduct of 𝔄+ to the language ℒ is a model of 𝑇. □
Lemma 2.6.13. For any consistent ℒ-theory 𝑇 there exists a set of constant
symbols 𝐶 disjoint from ℒ such that if ℒ+ denotes the language ℒ∪𝐶, then
there exists a complete ℒ+ -theory 𝑇 + containing 𝑇 that admits Henkin
witnesses in 𝐶.
Proof. To any ℒ-formula with a free variable 𝜑(𝑥) one assigns a new
constant symbol 𝑐𝜑 . Let 𝐶1 be the set of all 𝑐𝜑 . One sets ℒ1 = ℒ ∪ 𝐶1 and
ℒ
𝑇1 ∶= 𝑇˜ ∶= 𝑇 ∪ {∃𝑥𝜑 → 𝜑𝑐𝜑 /𝑥 ∣ 𝜑(𝑥) ∈ Fml }.
Let us prove that the theory 𝑇1 is consistent. Since a theory is consis-
tent if and only if any finite subset is consistent, it suffices to prove that
the theory 𝑇˜′ = 𝑇 ∪ {∃𝑥𝑖 𝜑𝑖 → 𝜑𝑐𝑖 𝑖 /𝑥𝑖 } is consistent for any finite set of
𝜑
formulas {𝜑1 , … , 𝜑𝑛 }. The theory 𝑇 being consistent as an ℒ1 -theory by
Lemma 2.6.8(b), this follows from Lemma 2.6.9, by induction on 𝑛.
We iterate that construction, with 𝐶2 a set of new constant symbols,
ℒ2 ∶= ℒ1 ∪ 𝐶2 and the ℒ2 -theory 𝑇2 ∶= 𝑇˜1 . By the argument just given,
𝑇2 is consistent. By induction, 𝑇𝑛 being constructed, it gives rise to a set
of new constant symbols 𝐶𝑛+1 . One sets ℒ𝑛+1 = ℒ𝑛 ∪ 𝐶𝑛+1 , and the
ℒ𝑛+1 -theory 𝑇𝑛+1 ∶= 𝑇˜𝑛 is consistent. Set 𝐶 ∶= ⋃𝑛∈ℕ 𝐶𝑛 and ℒ+ ∶=
ℒ∪𝐶. Thus (𝑇𝑛 )𝑛∈ℕ is an increasing sequence of consistent ℒ+ -theories.
The ℒ+ -theory 𝑇 ′ ∶= ⋃𝑛∈ℕ 𝑇𝑛 is then consistent by Corollary 2.6.5. By

construction, it admits Henkin witnesses in 𝐶.
Since any ℒ+ -theory 𝑆 ′ containing 𝑇 ′ still admits Henkin witnesses
in 𝐶, it suffices to prove that any consistent ℒ+ -theory is contained in
some complete ℒ+ -theory. This follows from Zorn’s Lemma. Indeed,
the set
𝒮 = {𝑆 ′ ⊇ 𝑇 ′ ∣ 𝑆 ′ is a consistent ℒ+ -theory}
is non-empty and partially ordered under inclusion. As the union of a
chain of consistent theories is still consistent by Corollary 2.6.5, this or-
dering is inductive. Hence, by Zorn’s Lemma, there exists a maximal
element in 𝒮, that is, there exists an ℒ+ -theory 𝑇 + ⊇ 𝑇 ′ which is maxi-
mal consistent. Let 𝜑 be some ℒ+ -sentence. If 𝑇 + ⊬ℒ+ 𝜑, the ℒ+ -theory
𝑇 + ∪ {¬𝜑} is consistent by Corollary 2.6.7. One deduces that ¬𝜑 ∈ 𝑇 +
by maximality, hence 𝑇 + is complete, which finishes the proof.
When ℒ is countable, one can provide a more direct construction
for the theory in the lemma that we shall now sketch. One fixes a set
𝐶 = {𝑐𝑛 ∣ 𝑛 ∈ ℕ} of new constant symbols and one enumerates the
set of ℒ ∪ 𝐶-sentences via 𝜑0 , 𝜑1 , 𝜑2 , …. By induction, one constructs
an increasing sequence (𝑇𝑛 )𝑛∈ℕ of consistent ℒ ∪ 𝐶-theories with the
following properties:
• 𝑇0 = 𝑇, and 𝑇𝑛 ⧵ 𝑇 is finite for any 𝑛 ∈ ℕ;
• 𝜑𝑛 ∈ 𝑇𝑛+1 or ¬𝜑𝑛 ∈ 𝑇𝑛+1 ;
• if 𝜑𝑛 is of the form ∃𝑥𝜓 ∈ 𝑇𝑛+1 , then there exists 𝑐 ∈ 𝐶 such
that 𝜓𝑐/𝑥 ∈ 𝑇𝑛+1 .
Once such a sequence (𝑇𝑛 )𝑛∈ℕ is constructed, it is enough to set 𝑇 + ∶=
⋃𝑛∈ℕ 𝑇𝑛 . Then 𝑇 + is a complete theory by construction, and it admits
Henkin witnesses in 𝐶. To prove the latter, consider some ℒ∪𝐶-formula
𝜓(𝑥). Then there exists some 𝑛 ∈ ℕ such that ∃𝑥𝜓 is equal to 𝜑𝑛 . If
𝜑𝑛 ∈ 𝑇 + , then by construction 𝜓𝑐/𝑥 ∈ 𝑇 + for some 𝑐 ∈ 𝐶, hence ∃𝑥𝜓 →
𝜓𝑐/𝑥 ∈ 𝑇 + . Otherwise, one has ¬∃𝑥𝜓 ∈ 𝑇 + , hence ∃𝑥𝜓 → 𝜓𝑐/𝑥 ∈ 𝑇 +
for any 𝑐.
Assume 𝑇𝑛 is constructed. If 𝑇𝑛 ∪{𝜑𝑛 } is consistent, one sets 𝜓𝑛 equal
to 𝜑𝑛 . Otherwise, 𝜓𝑛 is set equal to ¬𝜑𝑛 . In both cases, the theory 𝑇𝑛 ∪
{𝜓𝑛 } is consistent. If 𝜓𝑛 is not of the form ∃𝑥𝜒, one sets 𝑇𝑛+1 ∶= 𝑇𝑛 ∪{𝜓𝑛 }.
Otherwise, let 𝑚 be minimal such that 𝑐𝑚 has no occurrence in 𝑇𝑛 (such

an 𝑚 exists since 𝑇𝑛 ⧵ 𝑇 is finite). One sets 𝑇𝑛+1 ∶= 𝑇𝑛 ∪ {𝜓𝑛 , 𝜒𝑐𝑚 /𝑥 }. The
theory 𝑇𝑛+1 is consistent by Lemma 2.6.9. □
2.7. Exercises
Exercise 2.7.1 (Compactness in propositional logic). Let Fml𝒫 be the
set of propositional formulas over an infinite set 𝒫 of propositional vari-
ables. A set Σ ⊆ Fml𝒫 is called satisfiable if there is an assignment
𝛿 ∶ 𝒫 → {0, 1} such that 𝛿∗ (𝐹) = 1 for all 𝐹 ∈ Σ; it is called finitely
satisfiable if every finite subset Σ0 ⊆ Σ is satisfiable.
The aim of this exercise is to prove the following result (Compact-
ness Theorem in propositional logic): A set Σ ⊆ Fml𝒫 of propositional
formulas is satisfiable if and only if it is finitely satisfiable.
(1) Prove the theorem in the special case when Σ is complete, that
is, when for every 𝑝 ∈ 𝒫 one has 𝑝 ∈ Σ or ¬𝑝 ∈ Σ.
(2) Prove that every finitely satisfiable set of propositional formu-
las is contained in a finitely satisfiable set which is complete.
(3) Conclude.
Application 1. Let 𝑘 be a natural number. A graph 𝒢 = (𝐺, 𝐸) is said to
be 𝑘-colorable if there is a coloring of its set of vertices 𝐺 with 𝑘 colors
such that any two vertices which are connected by an edge have differ-
ent colors. Prove that a graph 𝒢 is 𝑘-colorable if and only if every finite
subgraph of 𝒢 is 𝑘-colorable.
Application 2. For 𝑛 ∈ ℕ and a set 𝐴 let 𝒫𝑛 (𝐴) be the set of 𝑛-element
subsets of 𝐴.
(a) Prove Ramsey’s Theorem (infinite version): For any natural
numbers 𝑛, 𝑘 and any 𝑓 ∶ 𝒫𝑛 (ℕ) → {0, 1, … , 𝑘 − 1} there is an
infinite subset 𝐴 of ℕ such that the restriction of 𝑓 to 𝒫𝑛 (𝐴) is
constant.
[Hint: Argue by induction on 𝑛 ∈ ℕ. Given some function 𝑓 ∶
𝒫𝑛+1 (ℕ) → {0, 1, … , 𝑘−1} and some natural number 𝑎, consider
the induced map 𝑓𝑎 on 𝒫𝑛 (ℕ⧵{𝑎}) defined by 𝑓𝑎 (𝑆) = 𝑓(𝑆 ∪{𝑎}).
Inductively construct an increasing sequence 0 = 𝑎0 < 𝑎1 <
𝑎2 < ⋯ in ℕ and a sequence of infinite sets ℕ = 𝑋0 ⊋ 𝑋1 ⊋
2.7. Exercises 59
𝑋2 ⊋ ⋯ such that for every 𝑖 one has 𝑎𝑖 ∈ 𝑋𝑖 , 𝑋𝑖+1 ∩ [0, 𝑎𝑖 ] = ∅

and 𝑓𝑎𝑖 is constant on 𝒫𝑛 (𝑋𝑖+1 ).]
(b) Deduce Ramsey’s Theorem (finite version): For any natural
numbers 𝑚, 𝑛, 𝑘 there is a natural number 𝑁 such that for any
function 𝑓 ∶ 𝒫𝑛 ({0, 1, … , 𝑁 − 1}) → {0, 1, … , 𝑘 − 1} there is an
𝑚-element subset 𝐴 ⊆ {0, 1, … , 𝑁 − 1} such that the restriction of
𝑓 to 𝒫𝑛 (𝐴) is constant.
Exercise 2.7.2 (Resolution in propositional logic). We keep the context
of the previous exercise. Let 𝐹, 𝐺 ∈ Fml𝒫 . We say that 𝐹 and 𝐺 are
logically equivalent if (𝐹 ↔ 𝐺) is a tautology. A literal is a formula of
the form 𝑝 or ¬𝑝 for some 𝑝 ∈ 𝒫; a clause is a disjunction of literals.
By convention, the empty clause, denoted by □, is a formula which is
always false.
Observe that any formula 𝐹 is logically equivalent to a conjunction
of clauses.
• If a literal 𝐿 occurs more than once in a clause 𝐶, a clause 𝐷
obtained by deleting one occurrence of 𝐿 in 𝐶 is called a sim-
plification of 𝐶.
• Given clauses 𝐶 = 𝐶1 ∨ 𝑝 ∨ 𝐶2 and 𝐷 = 𝐷1 ∨ ¬𝑝 ∨ 𝐷2 , where
𝑝 ∈ 𝒫, we say that 𝐸 = 𝐶1 ∨ 𝐶2 ∨ 𝐷1 ∨ 𝐷2 is obtained from 𝐶
and 𝐷 by the cut rule.
• A resolution sequence from a set of clauses 𝒞 is a finite sequence
of clauses (𝐶0 , … , 𝐶𝑛 ) such that for any 𝑖 ≤ 𝑛 either 𝐶𝑖 ∈ 𝒞, or
𝐶𝑖 is a simplification of 𝐶𝑗 for some 𝑗 < 𝑖, or there exist 𝑗, 𝑘 < 𝑖
such that 𝐶𝑖 is obtained from 𝐶𝑗 and 𝐶𝑘 by the cut rule.
We say 𝒞 is refutable if there is a resolution sequence from
𝒞 ending in □.
The aim of this exercise is to prove that a set of clauses is satisfiable if
and only if it may not be refuted.
(1) Prove that the resolution method is sound: if a set of clauses
may be refuted, then it is unsatisfiable.
(2) Observe that a clause 𝐶 is a tautology if and only if there is a
propositional variable 𝑝 such that 𝐶 contains the literals 𝑝 and
¬𝑝.
(3) Let 𝒞 be a set of clauses and 𝑝 ∈ 𝒫. If 𝐶 is a clause, let 𝐶𝑝=0

be the clause obtained from 𝐶 by deleting all literals which are
equal to 𝑝. Consider
𝒞𝑝=0 = {𝐶𝑝=0 ∣ 𝐶 does not contain the literal ¬𝑝}.
Similarly define 𝐶𝑝=1 and 𝒞𝑝=1 , interchanging the roles of ¬𝑝

and 𝑝.
(a) Prove that 𝒞 is satisfiable if and only if 𝒞𝑝=0 or 𝒞𝑝=1 is
satisfiable.
(b) Prove that 𝒞𝑝=0 is refutable if and only if there is a reso-
lution sequence from 𝒞 ending in □ or in 𝑝.
(c) Prove that 𝒞 is refutable if and only if both 𝒞𝑝=0 and 𝒞𝑝=1
are refutable.
(4) Prove that the resolution method is complete: if a set of clauses
is not satisfiable, it may be refuted.
[Hint: Prove the result for finite sets of clauses by induction on
the number of variables, and then use compactness (cf. Exer-
cise 2.7.1) to deduce the general case from this.]
Exercise 2.7.3 (Interpolation). Let 𝜑 be an ℒ1 -sentence and let 𝜓 be an

ℒ2 -sentence such that ⊧ 𝜑 → 𝜓. Set ℒ0 ∶= ℒ1 ∩ ℒ2 . The aim of this
exercise is to prove Craig’s Interpolation Theorem: There exists an ℒ0 -
sentence 𝜃 such that ⊧ 𝜑 → 𝜃 and ⊧ 𝜃 → 𝜓.
We may assume that ℒ1 and ℒ2 are countable. Assume for contra-
diction that no such 𝜃 exists. Let 𝐶 be a countably infinite set of new
constant symbols and set ℒ+ +
𝑖 = ℒ𝑖 ∪ 𝐶, for 𝑖 ∈ {0, 1, 2}. For an ℒ1 -
+
theory 𝑇 and an ℒ2 -theory 𝑈 we say that the pair (𝑇, 𝑈) is inseparable
if there is no ℒ+
0 -sentence 𝜃 such that 𝑇 ⊧ 𝜃 and 𝑈 ⊧ ¬𝜃.
(1) Prove that the pair ({𝜑}, {¬𝜓}) is inseparable.

(2) Prove that if (𝑇, 𝑈) is inseparable and 𝜒 is an ℒ+ 1 -sentence,
then either (𝑇 ∪ {𝜒}, 𝑈) or (𝑇 ∪ {¬𝜒}, 𝑈) is inseparable.
(3) Construct an increasing chain of finite ℒ+
1 -theories (𝑇𝑛 )𝑛∈ℕ with
𝑇0 = {𝜑} and an increasing chain of finite ℒ+2 -theories (𝑈𝑛 )𝑛∈ℕ
with 𝑈0 = {¬𝜓} such that, letting 𝑇𝜔 = ⋃𝑛∈ℕ 𝑇𝑛 and 𝑈𝜔 =
⋃𝑛∈ℕ 𝑈𝑛 , the following conditions hold:
2.7. Exercises 61
(i) 𝑇𝜔 is a complete deductively closed ℒ+

1 -theory which ad-
mits Henkin witnesses in 𝐶,
(ii) 𝑈𝜔 is a complete deductively closed ℒ+
2 -theory which ad-
mits Henkin witnesses in 𝐶, and
(iii) the pair (𝑇𝜔 , 𝑈𝜔 ) is inseparable.
Moreover, deduce from (i)-(iii) that 𝑇𝜔 ∩ 𝑈𝜔 is a complete
ℒ+
0 -theory which admits Henkin witnesses in 𝐶.
(4) Prove that 𝑇𝜔 ∪ 𝑈𝜔 has a model, and conclude Craig’s Interpo-
lation Theorem.
(5) Deduce Robinson’s Consistency Theorem: If 𝑇𝑖 are consistent
ℒ𝑖 -theories for 𝑖 = 1, 2 such that 𝑇0 = 𝑇1 ∩ 𝑇2 is a complete
ℒ0 -theory, then 𝑇1 ∪ 𝑇2 is consistent.
Exercise 2.7.4 (Beth Definability). Let ℒ be a first-order language and

𝑅 ∉ ℒ an 𝑛-ary relation symbol. Let 𝑇 be an ℒ ∪ {𝑅}-theory.
• We say 𝑇 implicitly defines 𝑅 if for any ℒ-structure 𝔐 with base
set 𝑀 and any 𝑛-ary relations 𝑅1 , 𝑅2 ⊆ 𝑀 𝑛 such that ⟨𝔐, 𝑅1 ⟩
and ⟨𝔐, 𝑅2 ⟩ are both models of 𝑇, one has 𝑅1 = 𝑅2 .
• We say that 𝑇 explicitly defines 𝑅 if there is an ℒ-formula
𝜑(𝑥1 , … , 𝑥𝑛 ) such that 𝑇 ⊧ ∀𝑥(𝑅(𝑥) ↔ 𝜑(𝑥)).
The aim of this exercise is to prove Beth’s Definability Theorem: The
theory 𝑇 implicitly defines the relation 𝑅 if and only if it explicitly defines
𝑅.
(1) Prove that if 𝑇 explicitly defines 𝑅, then it implicitly defines 𝑅.
(2) Assume that 𝑇 implicitly defines 𝑅. Let 𝑅′ be a new 𝑛-ary rela-
tion symbol and 𝑐1 , … , 𝑐𝑛 new pairwise distinct constant sym-
bols.
(a) Let 𝑇 ′ be the theory obtained from 𝑇 by replacing every
occurrence of 𝑅 by 𝑅′ . Observe that
𝑇 ∪ 𝑇 ′ ⊧ 𝑅(𝑐1 , … , 𝑐𝑛 ) → 𝑅′ (𝑐1 , … , 𝑐𝑛 ).
(b) Prove that there is a finite conjunction 𝜑 of sentences in 𝑇
such that
𝜑 ∧ 𝑅(𝑐1 , … , 𝑐𝑛 ) ⊧ 𝜑′ → 𝑅′ (𝑐1 , … , 𝑐𝑛 ),
where 𝜑′ is the formula obtained from 𝜑 by replacing ev-

ery occurrence of 𝑅 by 𝑅′ .
(c) Prove that 𝑇 explicitly defines 𝑅.
[Hint: Use Craig’s Interpolation Theorem (see Exercise
2.7.3).]
Exercise 2.7.5. If 𝑤 is a word over the alphabet Σ, a subword of 𝑤 is a

word 𝑢 such that 𝑤 = 𝑤1 𝑢𝑤2 for certain words 𝑤1 , 𝑤2 .
By induction on the height of a formula 𝜑, one defines the set Sub(𝜑)
of all subformulas of 𝜑 as a subset of the set of its subwords. If 𝜑 is atomic,
then Sub(𝜑) ∶= {𝜑}; if 𝜑 equals ¬𝜓 or ∃𝑣𝑖 𝜓, then Sub(𝜑) ∶= Sub(𝜓)∪{𝜑};
finally, if 𝜑 equals (𝜓1 ∧ 𝜓2 ), then Sub(𝜑) ∶= Sub(𝜓1 ) ∪ Sub(𝜓2 ) ∪ {𝜑}.
Now let 𝑇 be an ℒ-theory. Two ℒ-formulas 𝜑1 (𝑥) and 𝜑2 (𝑥) are said
to be equivalent in 𝑇 if 𝑇 ⊧ ∀𝑥1 , … , 𝑥𝑛 (𝜑1 ↔ 𝜑2 ). They are called logically
equivalent if they are equivalent in the empty theory.
Assume that 𝜑(𝑥) is an ℒ-formula and that 𝜓(𝑦1 , … , 𝑦𝑚 ) is a subfor-
mula of 𝜑. Let 𝜓′ be equivalent in 𝑇 to 𝜓, and assume that Free(𝜓) =
Free(𝜓′ ) = {𝑦1 , … , 𝑦𝑚 }. Let 𝜑′ (𝑥) be the formula one obtains when one
replaces the subformula 𝜓 in 𝜑 by 𝜓′ . Show that 𝜑(𝑥) and 𝜑′ (𝑥) are then
equivalent in 𝑇.
Exercise 2.7.6 (Herbrand normal form). Let ℒ be a language. A formula

𝜑 is said to be in prenex normal form if it is of the form
𝑄1 𝑥1 … 𝑄𝑚 𝑥𝑚 𝜑0 ,
where 𝑄𝑖 ∈ {∃, ∀} for 1 ≤ 𝑖 ≤ 𝑚 and 𝜑0 is a quantifier-free formula. It is

called existential if it is in prenex normal form and does not contain any
universal quantifier.
(1) Prove that for every ℒ-formula 𝜑(𝑥) there is an ℒ-formula 𝜓(𝑥)
in prenex normal form which is logically equivalent to 𝜑(𝑥),
that is, such that ⊧ ∀𝑥(𝜑 ↔ 𝜓).
The language ℒℋ is obtained from ℒ by adding, for every ℒ-formula
𝜑(𝑥1 , … , 𝑥𝑛 ) with 𝑛 free variables, a function symbol 𝑓𝜑 of arity 𝑛.
If 𝜑(𝑥1 , … , 𝑥𝑛 ) is an ℒ-formula in prenex normal form, one defines its
Herbrand form 𝜑ℋ (𝑥1 , … , 𝑥𝑛 ) by induction on the length of 𝜑:
2.7. Exercises 63
• if 𝜑(𝑥1 , … , 𝑥𝑛 ) is quantifier-free, one sets 𝜑ℋ (𝑥1 , … , 𝑥𝑛 ) equal

to 𝜑(𝑥1 , … , 𝑥𝑛 );
• if 𝜑(𝑥1 , … , 𝑥𝑛 ) is of the form ∃𝑦𝜓(𝑦, 𝑥1 , … , 𝑥𝑛 ), one sets
𝜑ℋ (𝑥1 , … , 𝑥𝑛 ) equal to ∃𝑦𝜓ℋ (𝑦, 𝑥1 , … , 𝑥𝑛 );
• if 𝜑(𝑥1 , … , 𝑥𝑛 ) is of the form ∀𝑦𝜓(𝑦, 𝑥1 , … , 𝑥𝑛 ), one sets
𝜑ℋ (𝑥1 , … , 𝑥𝑛 ) equal to 𝜓ℋ (𝑓𝜑 (𝑥1 , … , 𝑥𝑛 ), 𝑥1 , … , 𝑥𝑛 ).
(2) Let 𝜑 be in prenex normal form. Prove that 𝜑 ⊢ 𝜑ℋ .

(3) Prove that for every ℒ-formula 𝜑(𝑥1 , … , 𝑥𝑛 ) in prenex normal
form and for every ℒ-structure 𝔐, there is an ℒℋ -expansion
𝔐ℋ of 𝔐 such that for every (𝑎1 , … , 𝑎𝑛 ) ∈ 𝑀 𝑛 , if 𝔐 ⊭
𝜑[𝑎1 , … , 𝑎𝑛 ] then 𝔐ℋ ⊭ 𝜑ℋ [𝑎1 , … , 𝑎𝑛 ].
(4) Let 𝜑 be an ℒ-sentence in prenex normal form. Prove that
⊢ 𝜑ℋ if and only if ⊢ 𝜑. In other words, there is an explicit
recipe to construct an existential sentence in an expansion of
the language which is universally valid if and only if the origi-
nal sentence is universally valid.
Exercise 2.7.7. Let ℒ be a first-order language containing at least

one constant symbol, and let 𝜑 be an existential ℒ-sentence, given by
∃𝑥1 , … , 𝑥𝑛 𝜓, with 𝜓(𝑥1 , … , 𝑥𝑛 ) a quantifier-free formula. Prove that ⊢ 𝜑
𝑗
if and only if there are 𝑚 > 0 and ℒ-terms (𝑡𝑖 )1≤𝑖≤𝑛,1≤𝑗≤𝑚 not containing
any variable such that
𝑗 𝑗
⊢ 𝜓(𝑡1 , … , 𝑡𝑛 ).
⋁
1≤𝑗≤𝑚
Exercise 2.7.8 (Omitting Types Theorem). Let 𝑇 be an ℒ-theory and

𝜋 = 𝜋(𝑣1 , … , 𝑣𝑛 ) a set of ℒ-formulas such that the free variables of the
formulas in 𝜋 belong to {𝑣1 , … , 𝑣𝑛 }. One says 𝜋 is a partial 𝑛-type (in
𝑘
𝑇) if for any finite subset {𝜑1 , … , 𝜑𝑘 } of 𝜋, the theory 𝑇 ∪ {∃𝑣 ⋀𝑖=1 𝜑𝑖 } is
consistent.
• One says a model 𝔐 of 𝑇 realizes 𝜋 if there exists 𝑎 ∈ 𝑀 𝑛 such

that 𝔐 ⊧ 𝜑[𝑎] for any 𝜑 ∈ 𝜋; otherwise, one says that 𝔐 omits
𝜋.
• 𝜋 is called isolated if there exists a formula 𝜑(𝑣1 , … , 𝑣𝑛 ) with

𝑇 ∪ {∃𝑣𝜑} consistent such that 𝑇 ⊢ ∀𝑣(𝜑 → 𝜓) for any formula
𝜓 ∈ 𝜋.
The aim of this exercise is to prove the following result, which is
called the Omitting Types Theorem: Let ℒ be countable and 𝑇 a consistent
ℒ-theory. One assumes that, for 𝑗 ∈ ℕ, 𝜋𝑗 = 𝜋(𝑣1 , … , 𝑣𝑛𝑗 ) is a non-isolated
partial type. Then there exists a countable model of 𝑇 that omits all the 𝜋𝑗 .
We consider 𝑇 and (𝜋𝑗 )𝑗∈ℕ as in the statement of the theorem. Let
𝐶 = {𝑐𝑖 ∣ 𝑖 ∈ ℕ} be a set of new pairwise distinct constant symbols and
set ℒ∗ = ℒ ∪ 𝐶.
(1) Let 𝜓(𝑣0 ) be an ℒ∗ -formula, and let 𝑇 ′ ⊇ 𝑇 be a consistent ℒ∗ -
theory such that 𝑇 ′ ⧵ 𝑇 is finite. Prove that there exists 𝑐 ∈ 𝐶
such that 𝑇 ″ = 𝑇 ′ ∪ {∃𝑣0 𝜓 → 𝜓(𝑐)} is consistent.
(2) Let 𝑐 = (𝑐𝑖1 , … , 𝑐𝑖𝑛 ) be some 𝑛𝑗 -tuple of distinct elements of 𝐶,
𝑗
and let 𝑇 ′ ⊇ 𝑇 be a consistent ℒ∗ -theory with 𝑇 ′ ⧵ 𝑇 finite.
Prove that there exists 𝜑(𝑣1 , … , 𝑣𝑛𝑗 ) ∈ 𝜋𝑗 such that 𝑇 ″ = 𝑇 ′ ∪
{¬𝜑(𝑐)} is consistent.
[Hint: Use that, up to logical equivalence, 𝑇 ′ is given by 𝑇 ∪
{𝜒(𝑐, 𝑑)}, where 𝜒(𝑣1 , … , 𝑣𝑛𝑗 , 𝑦) is some ℒ-formula and 𝑑 a tuple
of constant symbols in 𝐶 disjoint from 𝑐.]
(3) Prove that there exists a consistent ℒ∗ -theory 𝑇 ∗ ⊇ 𝑇 satisfying
the following properties:
(H) For every ℒ∗ -formula 𝜓(𝑣0 ) there exists some 𝑐 ∈ 𝐶 not
appearing in 𝜓 such that ∃𝑣0 𝜓 → 𝜓(𝑐) ∈ 𝑇 ∗ .
(O) For any 𝑗 ∈ ℕ and any 𝑛𝑗 -tuple (𝑐𝑖1 , … , 𝑐𝑖𝑛 ) of distinct
𝑗
elements of 𝐶 there exists a formula 𝜑(𝑣) ∈ 𝜋𝑗 such that
¬𝜑(𝑐𝑖1 , … , 𝑐𝑖𝑛 ) ∈ 𝑇 ∗ .
𝑗
[Hint: Using suitable enumerations, construct 𝑇 ∗ as the union

of an increasing chain (𝑇𝑛 )𝑛∈ℕ with 𝑇0 = 𝑇 and 𝑇𝑛 ⧵ 𝑇 finite
for every 𝑛.]
(4) Let 𝑇 ∗ be a consistent ℒ∗ -theory satisfying (H). Prove that for
any 𝑖 ∈ ℕ, the set 𝑋𝑖 = {𝑗 ∈ ℕ ∣ 𝑇 ∗ ⊧ 𝑐𝑖 = 𝑐𝑗 } is infinite.
(5) Conclude the Omitting Types Theorem.
10.1090/stml/089/03
Chapter 3
First Steps in Model

Theory
Introduction
First-order model theory deals with the relationship between theories in
a first-order language and their models. It provides a way to treat prob-
lems from other areas of mathematics (like algebra or combinatorics)
with tools from mathematical logic. First-order logic has a rather lim-
ited expressive power: we will see for instance that an infinite structure
is never determined, up to isomorphism, by its first-order theory (Corol-
lary 3.2.4). In fact this apparent weakness is a strength and a reason for
its efficiency. As we will see it is sometimes quite useful to be able to
switch from one model of a theory to another.
One of the main themes of this chapter is quantifier elimination,
which in concrete examples often yields a particularly simple descrip-
tion of all definable sets in models of the theory under consideration. In
Theorem 3.4.4 we provide a very useful semantical criterion for a for-
mula to be equivalent to a quantifier-free formula in a given theory (a
syntactic condition). This applies in particular to the theory of alge-
braically closed fields which is studied in 3.5, and other examples are
considered in exercises. As a remarkable consequence of this study, we
provide a statement and a proof of the Lefschetz principle from algebraic
65
66 3. First Steps in Model Theory
geometry (Theorem 3.5.5) which allows us to transfer results from alge-

braically closed fields of characteristic zero to algebraically closed fields
of sufficiently large characteristic. A beautiful application is provided
by Ax’s Theorem (Theorem 3.6.3) stating that injective complex polyno-
mial mappings are always surjective, which is proved by reduction to
finite fields.
3.1. Some Fundamental Theorems

Theorem 3.1.1 (Compactness Theorem). Let 𝑇 be a theory such that any
finite subset of 𝑇 has a model. Then 𝑇 has a model.
Proof. The theory 𝑇 is consistent if and only if any finite subset of 𝑇 is

consistent. Thus the statement follows from Theorem 2.6.11. □
Remark. Our proof of the Compactness Theorem uses Gödel’s Com-
pleteness Theorem. For a more direct semantical proof, see Exercise 3.7.5.
Exercise. Fix a language ℒ and denote by 𝑋 the set of ℒ-theories 𝑇
which are complete and closed under deduction. One may define a topol-
ogy on 𝑋 in the following way. For 𝜑 an ℒ-sentence, one sets ⟨𝜑⟩ ∶= {𝑇 ∈
𝑋 ∣ 𝜑 ∈ 𝑇}, and one considers the topology generated by the family of
all ⟨𝜑⟩. Prove:
(1) ⟨𝜑 ∧ 𝜓⟩ = ⟨𝜑⟩ ∩ ⟨𝜓⟩ and ⟨¬𝜑⟩ = 𝑋 ⧵ ⟨𝜑⟩. In particular the sets
⟨𝜑⟩ form a basis of clopen subsets.
(2) The space 𝑋 is compact and Hausdorff.
Definition. Let 𝔐 and 𝔑 be ℒ-structures.
(1) We say that 𝔐 and 𝔑 are elementarily equivalent if they satisfy
the same ℒ-sentences, that is, if Th(𝔐) = Th(𝔑). This is de-
noted by 𝔐 ≡ 𝔑.
(2) Assume that 𝔐 ⊆ 𝔑. We say that 𝔐 is an elementary substruc-
ture of 𝔑 (and 𝔑 is called an elementary extension of 𝔐) if for
every ℒ-formula 𝜑(𝑥1 , … , 𝑥𝑛 ) and every tuple 𝑎 = (𝑎1 , … , 𝑎𝑛 ) ∈
𝑀 𝑛 we have 𝔐 ⊧ 𝜑[𝑎] if and only if 𝔑 ⊧ 𝜑[𝑎]. This is denoted
by 𝔐 ≼ 𝔑.
3.1. Some Fundamental Theorems 67
Remark.
(1) If 𝔐 ≼ 𝔑, then 𝔐 ≡ 𝔑.
(2) If 𝔐 ≅ 𝔑, then 𝔐 ≡ 𝔑. □
The following example shows that if 𝔐 ⊆ 𝔑 and 𝔐 ≡ 𝔑 it is not

always the case that 𝔐 ≼ 𝔑.
Example. One considers the two ℒ𝑜𝑟𝑑 -structures 𝔑 ∶= ⟨ℕ; <⟩ and 𝔐 ∶=
⟨ℕ∗ ; <⟩. One has 𝔐 ⊆ 𝔑 and 𝔐 ≅ 𝔑, in particular 𝔐 ≡ 𝔑. On the other
hand the extension 𝔐 ⊆ 𝔑 is not elementary, since 𝔐 ⊧ ¬∃𝑥 𝑥 < 1 and
𝔑 ⊧ ∃𝑥 𝑥 < 1.
Theorem 3.1.2 (Tarski-Vaught Test). Let 𝔐 and 𝔑 be ℒ-structures with
𝔐 ⊆ 𝔑. Assume that for any ℒ-formula 𝜑(𝑥0 , … , 𝑥𝑛 ) and any tuple 𝑎 =
(𝑎1 , … , 𝑎𝑛 ) ∈ 𝑀 𝑛 , if there exists 𝑏0 ∈ 𝑁 such that 𝔑 ⊧ 𝜑[𝑏0 , 𝑎], then there
exists 𝑎0 ∈ 𝑀 such that 𝔑 ⊧ 𝜑[𝑎0 , 𝑎]. Then 𝔐 ≼ 𝔑.
Proof. By induction on ht(𝜑(𝑥1 , … , 𝑥𝑛 )) one proves that for any 𝑎 ∈ 𝑀 𝑛

one has 𝔐 ⊧ 𝜑[𝑎] ⟺ 𝔑 ⊧ 𝜑[𝑎].
When 𝜑 is atomic, this is clear since 𝔐 is a substructure of 𝔑. The
case of logical connectives is also clear.
Assume now that 𝜑(𝑥1 , … , 𝑥𝑛 ) is equal to ∃𝑥0 𝜓(𝑥0 , … , 𝑥𝑛 ). (As 𝑥0
is not free in 𝜑, we may assume that 𝑥0 ≠ 𝑥𝑖 for 𝑖 = 1, … , 𝑛.) Let 𝑎 =
(𝑎1 , … , 𝑎𝑛 ) ∈ 𝑀 𝑛 . Then
𝔐 ⊧ 𝜑[𝑎] ⟺ there exists 𝑎0 ∈ 𝑀 such that 𝔐 ⊧ 𝜓[𝑎0 , 𝑎]
⟺ there exists 𝑎0 ∈ 𝑀 such that 𝔑 ⊧ 𝜓[𝑎0 , 𝑎]
⟺ there exists 𝑏0 ∈ 𝑁 such that 𝔑 ⊧ 𝜓[𝑏0 , 𝑎]
⟺ 𝔑 ⊧ 𝜑[𝑎],
where the second equivalence follows from the induction hypothesis
and the third one from the hypothesis of the theorem. □
ℒ
Notation. For a language ℒ we set card(ℒ) = card(Fml ).
Note that card(ℒ) ≥ card(𝒯 ℒ ) .

Theorem 3.1.3 (Downward Löwenheim-Skolem Theorem). Let 𝔐 be
an ℒ-structure and 𝐴 ⊆ 𝑀. Assume that card(𝑀) ≥ card(ℒ). Then there
exists an elementary substructure 𝔐0 of 𝔐 containing 𝐴 which is of cardi-
nality sup(card(𝐴), card(ℒ)).
Proof. Up to enlarging the subset 𝐴 if necessary, we may assume that

card(𝐴) ≥ card(ℒ).
In this proof, when ∅ ≠ 𝐵 ⊆ 𝑀, we denote by 𝐵˜ ⊆ 𝑀 the base set of
⟨𝐵⟩𝔐 , the substructure generated by 𝐵. By Exercise 2.3.3 we have
𝐵˜ = {𝑡𝔐 [𝑏1 , … , 𝑏𝑛 ] ∣ 𝑡 = 𝑡(𝑥1 , … , 𝑥𝑛 ) ∈ 𝒯 ℒ , 𝑏1 , … , 𝑏𝑛 ∈ 𝐵}.
In particular, if card(𝐵) ≥ card(ℒ), then card(𝐵) = card(𝐵˜), since there
exists a surjection from 𝒯 ℒ × ⋃𝑛∈ℕ 𝐵 𝑛 onto 𝐵˜.
Let 𝔄0 ∶= ⟨𝐴⟩𝔐 and 𝐴0 ∶= 𝐴. ˜ Assuming 𝐴𝑖 ⊆ 𝑀 is already defined,
one constructs 𝐴𝑖+1 as follows. For any ℒ-formula 𝜑(𝑥0 , … , 𝑥𝑛 ) and any
𝑛-tuple 𝑎 ∈ 𝐴𝑛𝑖 , if 𝔐 ⊧ 𝜑[𝑏0 , 𝑎] for some 𝑏0 ∈ 𝑀, one chooses 𝑐(𝜑, 𝑎) ∈
𝑀 such that 𝔐 ⊧ 𝜑[𝑐(𝜑, 𝑎), 𝑎]. One sets
ℒ
𝐵𝑖+1 ∶= 𝐴𝑖 ∪ {𝑐(𝜑, 𝑎) ∣ 𝜑(𝑥0 , … , 𝑥𝑛 ) ∈ Fml , 𝑎1 , … , 𝑎𝑛 ∈ 𝐴𝑖 }
and 𝐴𝑖+1 ∶= 𝐵˜
𝑖+1 , the base set of 𝔄𝑖+1 = ⟨𝐵𝑖+1 ⟩𝔐 .
Let 𝑀0 ∶= ⋃𝑖∈ℕ 𝐴𝑖 . Then 𝑀0 (is non-empty and) contains 𝑐𝔐 for
any 𝑐 ∈ 𝒞 ℒ and is closed under 𝑓𝔐 for any𝑓 ∈ ℱ ℒ . It is thus the base set
of a substructure 𝔐0 of 𝔐.
Let 𝜑(𝑥0 , … , 𝑥𝑛 ) be a formula, 𝑎 ∈ 𝑀0𝑛 and 𝑏0 ∈ 𝑀 such that 𝔐 ⊧
𝜑[𝑏0 , 𝑎]. There exists 𝑁 ∈ ℕ such that 𝑎1 , … , 𝑎𝑛 ∈ 𝐴𝑁 . By the con-
struction of 𝐴𝑁+1 there exists 𝑐0 ∈ 𝐴𝑁+1 ⊆ 𝑀0 such that 𝔐 ⊧ 𝜑[𝑐0 , 𝑎].
Hence 𝔐0 ≼ 𝔐 by the Tarski-Vaught Test. It is clear that card(𝑀0 ) =
card(𝐴). □
Example 3.1.4. Let ℜ = ⟨ℝ; 0, 1, +, −, ⋅, <⟩ be the ordered field of real
numbers.
(1) There exists ℜ′ ≡ ℜ with ℜ′ non-archimedean, that is, con-
taining an element 𝜀 > 0 such that 𝑛 ⋅ 𝜀 = 𝜀⏟⎵⏟⎵⏟
+ … + 𝜀 < 1 for any
𝑛 times
𝑛 ∈ ℕ. Such an 𝜀 is called infinitesimal.
(2) There exists ℜ′ ≼ ℜ such that ℜ′ is not complete.
Proof. (1) Let 𝑐 be a new constant symbol and ℒ ∶= ℒ𝑜.𝑟𝑖𝑛𝑔 ∪ {𝑐}. Let
𝑐 + … + 𝑐 < 1. One considers the ℒ-theory
𝜑𝑛 be the formula ⏟⎵⏟⎵⏟
𝑛 times
𝑇 ∶= Th(ℜ) ∪ {0 < 𝑐} ∪ {𝜑𝑛 ∣ 𝑛 ∈ ℕ}.
3.1. Some Fundamental Theorems 69
For any finite 𝑇0 ⊆ 𝑇 there exists 𝑁 ∈ ℕ such that 𝜑𝑚 ∉ 𝑇0 for any

𝑚 ≥ 𝑁. The expansion of ℜ to ℒ obtained by interpreting 𝑐 by 1/𝑁 ∈ ℝ
is thus a model of 𝑇0 . By compactness, it follows that 𝑇 has a model ℜ′ .
′
The element 𝜀 = 𝑐ℜ is infinitesimal, hence ℜ′ is non-archimedean.
(2) By the Downward Löwenheim-Skolem Theorem (for 𝐴 = ∅),
there exists ℜ′ ≼ ℜ with a countable base set ℝ′ . Since ℜ′ is a field,
ℚ ⊆ ℝ′ . Take 𝑟 ∈ ℝ ⧵ ℝ′ . Then the set {𝑟′ ∈ ℝ′ ∣ 𝑟′ < 𝑟} has no
supremum in ℝ′ , since ℝ′ is dense in ℝ. □
Remark. One may prove that ⟨ℝ𝑎𝑙𝑔 ; +, −, 0, 1, ⋅, <⟩ ≼ ℜ, where ℝ𝑎𝑙𝑔 =
{𝑟 ∈ ℝ ∣ there exists 0 ≠ 𝑝(𝑋) ∈ ℚ[𝑋] such that 𝑝(𝑟) = 0}.
Remark. Suppose that 𝔐 ≼ 𝔐′ and 𝐷 ⊆ 𝑀 𝑛 is a definable set (with
parameters). Then 𝐷 extends canonically to a definable set 𝐷′ ⊆ 𝑀 ′𝑛 in
𝔐′ , such that 𝐷′ ∩ 𝑀 𝑛 = 𝐷. Indeed, if 𝐷 = 𝜑[𝔐, 𝑏] for some formula
𝜑(𝑥, 𝑦) and tuple 𝑏 from 𝔐, then 𝐷′ = 𝜑[𝔐′ , 𝑏] has the desired property,
and it is independent of the choice of 𝜑 and 𝑏.
Definition. Let 𝒫 be a property that an ℒ-structure may or may not
satisfy. We say that 𝒫 is axiomatizable (resp. finitely axiomatizable) if
there exists an ℒ-theory 𝑇 (resp. an ℒ-sentence 𝜑) such that for any
ℒ-structure 𝔐, the property 𝒫 is satisfied by 𝔐 if and only if 𝔐 ⊧ 𝑇
(resp. 𝔐 ⊧ 𝜑). In this case one says that 𝑇 (resp. 𝜑) axiomatizes the prop-
erty 𝒫.
Proposition 3.1.5. Let 𝒫 be a property that an ℒ-structure may or may
not satisfy. Then 𝒫 is finitely axiomatizable if and only if 𝒫 and its negation
are both axiomatizable.
Proof. If 𝜑 axiomatizes 𝒫, then ¬𝜑 axiomatizes its negation. Conversely,

assume that 𝑇 axiomatizes 𝒫 and that 𝑇 ′ axiomatizes the negation of
𝒫. The theory 𝑇 ∪ 𝑇 ′ being inconsistent, there exist 𝜑1 , … , 𝜑𝑛 ∈ 𝑇,
𝜑1′ , … , 𝜑𝑚
′
∈ 𝑇 ′ such that {𝜑1 , … , 𝜑𝑛 , 𝜑1′ , … , 𝜑𝑚
′
} is inconsistent. Hence
𝑛 𝑚
𝔐⊧𝑇⇒𝔐⊧ 𝜑𝑖 ⇒ 𝔐 ⊭ 𝜑𝑖′ ⇒ 𝔐 ⊭ 𝑇 ′ ⇒ 𝔐 ⊧ 𝑇,
⋀ ⋀
𝑖=1 𝑖=1
𝑛
showing that the sentence ⋀𝑖=1 𝜑𝑖 is a finite axiomatization of 𝒫. □
Example 3.1.6.
(1) For any prime number 𝑝, fields of characteristic 𝑝 are finitely

axiomatizable in the language ℒ𝑟𝑖𝑛𝑔 .
(2) Fields of characteristic 0 are axiomatizable in ℒ𝑟𝑖𝑛𝑔 , but are
not finitely axiomatizable. Indeed, if a sentence 𝜑0 were a fi-
nite axiomatization, it would be a consequence of the theory
𝜑𝑓𝑖𝑒𝑙𝑑 ∪ {¬ 1⏟⎵
+⎵⏟ +⏟1 = 0 ∣ 𝑝 prime}, and then
…⎵⎵
𝑝 times
𝜑𝑓𝑖𝑒𝑙𝑑𝑠 ∪ {¬ 1⏟⎵
+⎵⏟ +⏟1 = 0 ∣ 𝑝 prime and 𝑝 < 𝑁} ⊢ 𝜑0
…⎵⎵
𝑝 times
for some 𝑁 ∈ ℕ. This would lead to a contradiction, since there

exist fields of characteristic 𝑝 > 𝑁.
(3) Archimedean ordered fields are not axiomatizable (cf. Exam-
ple 3.1.4(1)).
(4) Complete ordered fields are not axiomatizable (cf. Exam-
ple 3.1.4(2)).
The last two examples make more precise what we said at the be-
ginning of Chapter 2.
3.2. The Diagram Method

In practice, one often wishes to construct a model of some theory which
contains a given structure as a substructure or as an elementary sub-
structure. In this section, we will see that these properties may be en-
coded in suitable theories.
Let 𝔐 be an ℒ-structure. We consider the language ℒ𝑀 obtained by
adding to ℒ a new constant symbol 𝑐𝑚 for each element 𝑚 of 𝑀. The
structure 𝔐 admits a natural expansion to the language ℒ𝑀 , by inter-
preting 𝑐𝑚 by 𝑚. We denote 𝔐∗ the corresponding ℒ𝑀 -structure.
By abuse of notation, we shall sometimes identify 𝑚 and 𝑐𝑚 in what
follows. This will not cause any trouble, because of the Substitution
Lemma (Proposition 2.4.2).
Definition.
(1) The complete diagram of 𝔐, denoted by 𝐷(𝔐), is defined as
Th(𝔐∗ ). It is the set of ℒ𝑀 -sentences 𝜑(𝑐𝑚1 , … , 𝑐𝑚𝑛 ), where
3.2. The Diagram Method 71
𝜑(𝑥1 , … , 𝑥𝑛 ) is an ℒ-formula and 𝑚 ∈ 𝑀 𝑛 is such that 𝔐 ⊧

𝜑[𝑚1 , … , 𝑚𝑛 ].
(2) The simple diagram of 𝔐, denoted by Δ(𝔐), consists of all ℒ𝑀 -
sentences 𝜑(𝑐𝑚1 , … , 𝑐𝑚𝑛 ), where 𝜑(𝑥1 , … , 𝑥𝑛 ) is a quantifier-free
ℒ-formula and 𝑚 ∈ 𝑀 𝑛 is a tuple such that 𝔐 ⊧ 𝜑[𝑚1 , … , 𝑚𝑛 ].
Proposition 3.2.1. Reducts to the language ℒ of models of 𝐷(𝔐) corre-
spond, up to ℒ-isomorphism, to elementary extensions of 𝔐.
Proof. Let 𝔑′ ⊧ 𝐷(𝔐) and 𝔑 ∶= 𝔑′ ↾ ℒ be its ℒ-reduct. The map

𝔑′
𝑚 ↦ 𝑐𝑚 provides an isomorphism between 𝔐 and an elementary sub-
structure of 𝔑.
Conversely, if 𝔐 ≼ 𝔑, we consider the ℒ𝑀 -expansion 𝔑′ of 𝔑 ob-
tained by interpreting 𝑐𝑚 by 𝑚. We will prove that 𝔑′ ⊧ 𝐷(𝔐). For
this, let 𝜑(𝑐𝑚1 , … , 𝑐𝑚𝑛 ) ∈ 𝐷(𝔐). By definition, this means that 𝔐 ⊧
𝜑[𝑚1 , … , 𝑚𝑛 ], so 𝔑 ⊧ 𝜑[𝑚1 , … , 𝑚𝑛 ] by elementarity, and finally 𝔑′ ⊧
𝜑(𝑐𝑚1 , … , 𝑐𝑚𝑛 ) by the Substitution Lemma. □
Proposition 3.2.2. Reducts to the language ℒ of models of Δ(𝔐) corre-
spond, up to ℒ-isomorphism, to extensions of 𝔐.
Proof. Similar to the previous proof. □

Theorem 3.2.3 (Upward Löwenheim-Skolem Theorem). Let 𝔐 be an
infinite ℒ-structure for some language ℒ, and let 𝜅 be a cardinal such that
𝜅 ≥ sup(card(𝑀), card(ℒ)). Then there exists an elementary extension of
𝔐 which is of cardinality 𝜅.
Proof. It is enough to construct an elementary extension 𝔑 of 𝔐 of car-

dinality ≥ 𝜅, and to apply the Downward Löwenheim-Skolem Theorem
to a subset 𝐴 ⊆ 𝑁 of cardinality 𝜅 containing 𝑀.
For any 𝑖 ∈ 𝜅, let 𝑐𝑖 be a new constant symbol. Consider the lan-
guage ℒ̃ ∶= ℒ𝑀 ∪ {𝑐𝑖 ∣ 𝑖 ∈ 𝜅} and the ℒ̃-theory
𝑇˜ ∶= 𝐷(𝔐) ∪ {¬𝑐𝑖 = 𝑐𝑗 ∣ 𝑖 < 𝑗 < 𝜅}.
Then 𝑇˜ is finitely satisfiable. Indeed, if 𝑇˜0 is a finite subset of 𝑇,
˜ inter-
˜
preting the new constant symbols 𝑐𝑖1 , … , 𝑐𝑖𝑛 occurring in 𝑇0 by distinct
elements of 𝑀 one gets an expansion of 𝔐∗ which is a model of 𝑇˜0 . (This
is possible since 𝑀 is infinite.) By compactness, one gets a model 𝔑 ˜ ⊧ 𝑇˜

whose ℒ-reduct 𝔑 is an elementary extension of 𝔐 (by Proposition 3.2.1)
˜
of cardinality ≥ 𝜅, since it contains the elements 𝑐𝑖𝔑 , for 𝑖 ∈ 𝜅, which are
pairwise distinct. □
Corollary 3.2.4. Let 𝑇 be a theory having an infinite model. Then 𝑇 has
a model of cardinality 𝜅 for any 𝜅 ≥ card(ℒ). □
Exercise 3.2.5. Let ℒ be a finite language (that is, a language the signa-
ture of which is finite), and let 𝔐 be a finite ℒ-structure.
(1) Prove that there exists an ℒ𝑀 -sentence 𝜑 ∈ Δ(𝔐) such that
⊢ 𝜑 → 𝜑′ for any 𝜑′ ∈ Δ(𝔐).
(2) Deduce that if 𝔑 ≡ 𝔐, then 𝔑 ≅ 𝔐.
(3) (More difficult.) Let 𝔐′ be a finite ℒ′ -structure, with ℒ′ not
necessarily finite. Prove that 𝔑′ ≡ 𝔐′ ⇒ 𝔑′ ≅ 𝔐′ .
3.3. Expansions by Definition

It is often useful to enrich the language by symbols for definable rela-
tions, functions or constants. We will describe in this section how this
may be done formally.
Notation. ∃! 𝑥𝜓 (“there exists a unique 𝑥 such that 𝜓”) is an abbrevia-
tion for ∃𝑥(𝜓 ∧ ∀𝑥′ (𝜓𝑥′ /𝑥 → 𝑥 = 𝑥′ )), with 𝑥′ a variable which is distinct
from 𝑥.
Definition. Let 𝑇 be an ℒ-theory and let ℒ′ ⊇ ℒ. Assume that there
are
• for any 𝑛-ary relation symbol 𝑅 ∈ ℒ′ ⧵ ℒ an ℒ-formula
𝜑𝑅 (𝑥1 , … , 𝑥𝑛 ),
• for any 𝑛-ary function symbol 𝑓 ∈ ℒ′ ⧵ ℒ an ℒ-formula
𝜑𝑓 (𝑥0 , 𝑥1 , … , 𝑥𝑛 ) such that 𝑇 ⊧ ∀𝑥1 , … , 𝑥𝑛 ∃! 𝑥0 𝜑𝑓 ,
• for any constant symbol 𝑐 ∈ ℒ′ ⧵ ℒ an ℒ-formula 𝜑𝑐 (𝑥0 ) such
that 𝑇 ⊧ ∃! 𝑥0 𝜑𝑐 .
Then the ℒ′ -theory 𝑇 ′ given by 𝑇 together with
• an axiom ∀𝑥1 , … , 𝑥𝑛 (𝜑𝑅 (𝑥1 , … , 𝑥𝑛 ) ↔ 𝑅𝑥1 ⋯ 𝑥𝑛 ) for any rela-
tion symbol 𝑅 ∈ ℒ′ ⧵ ℒ,
3.3. Expansions by Definition 73
• an axiom ∀𝑥1 , … , 𝑥𝑛 𝜑𝑓 (𝑓(𝑥1 , … , 𝑥𝑛 ), 𝑥1 , … , 𝑥𝑛 ) for any function

symbol 𝑓 ∈ ℒ′ ⧵ ℒ,
• an axiom 𝜑𝑐 (𝑐) for any constant symbol 𝑐 ∈ ℒ′ ⧵ ℒ
is called an expansion by definition of 𝑇.
Recall that two formulas 𝜑1 (𝑥1 , … , 𝑥𝑛 ) and 𝜑2 (𝑥1 , … , 𝑥𝑛 ) are called

equivalent in the theory 𝑇 if 𝑇 ⊧ ∀𝑥1 , … , 𝑥𝑛 (𝜑1 ↔ 𝜑2 ), and that they are
called logically equivalent if they are equivalent in the empty theory (cf.
Exercise 2.7.5).
Lemma 3.3.1. Any formula is logically equivalent to a formula with all
terms of height ≤ 1.
Proof. For 𝜑 a formula, let 𝑀(𝜑) be the maximal height of a term in 𝜑.

The proof will be by induction on (𝑀(𝜑), ht(𝜑)) with lexicographic order.
The only non-trivial case is that of an atomic formula 𝜑, say of the form
𝑅𝑡1 ⋯ 𝑡𝑛 . To simplify notations, we assume that ht(𝑡𝑖 ) > 1 for 𝑖 = 1, … , 𝑚
and ht(𝑡𝑖 ) ≤ 1 for 𝑖 > 𝑚. For 𝑖 = 1, … , 𝑚, let 𝑡𝑖 = 𝑓𝑖 (𝑠𝑖,1 , … , 𝑠𝑖,𝑘𝑖 ). One
chooses new variables 𝑦𝑖,𝑗 , for 𝑖 = 1, … , 𝑚 and 1 ≤ 𝑗 ≤ 𝑘𝑖 . If 𝑦𝑖 ∶=
(𝑦𝑖,1 , … , 𝑦𝑖,𝑘𝑖 ), then 𝜑 is logically equivalent to the formula 𝜓 given by
∃𝑦1 , … , 𝑦𝑚 (𝑅𝑓1 (𝑦1 ) ⋯ 𝑓𝑚 (𝑦𝑚 )𝑡𝑚+1 ⋯ 𝑡𝑛 ∧ 𝑦𝑖,𝑗 = 𝑠𝑖,𝑗 )

⋀
𝑖,𝑗
and which satisfies 𝑀(𝜓) < 𝑀(𝜑). The case where 𝜑 is given by 𝑡1 = 𝑡2
is similar. □
Definition. Let 𝑇 be an ℒ-theory and 𝑇 ′ ⊇ 𝑇 an ℒ′ -theory for some
language ℒ′ ⊇ ℒ. One says that 𝑇 ′ is a conservative expansion of 𝑇 if for
any ℒ-sentence 𝜓 one has 𝑇 ⊧ 𝜓 if and only if 𝑇 ′ ⊧ 𝜓.
Proposition 3.3.2. Let 𝑇 ′ be an expansion by definition of 𝑇. Then the ex-
pansion 𝑇 ′ ⊇ 𝑇 is conservative. Furthermore, any ℒ′ -formula 𝜓′ is equiv-
alent in 𝑇 ′ to an ℒ-formula 𝜓.
Proof. Any model 𝔐 of 𝑇 has an ℒ′ -expansion (in fact unique) to a

model 𝔐′ of 𝑇 ′ . For a relation symbol 𝑅 ∈ ℒ′ ⧵ ℒ, we may and need to
′
define 𝑅𝔐 = 𝜑𝑅 [𝔐] (cf. Exercise 2.7.4(1)). For a new function symbol
′
𝑓, this amounts to defining 𝑓𝔐 (𝑎1 , … , 𝑎𝑛 ) = 𝑎0 if 𝔐 ⊧ 𝜑𝑓 [𝑎0 , 𝑎1 , … , 𝑎𝑛 ].
The case of a new constant symbol is similar. It follows that 𝑇 ′ ⊇ 𝑇 is

conservative.
In order to show that any ℒ′ -formula is equivalent in 𝑇 ′ to an ℒ-
formula, for simplicity we treat the case where ℒ′ = ℒ ∪ {𝑓} for an 𝑛-ary
function symbol 𝑓. The general case is proved in a similar way.
First let 𝜓′ be an atomic ℒ′ -formula containing only terms of height
≤ 1, say 𝜓′ equals 𝑅𝑡1 ⋯ 𝑡𝑚 . Assume for simplicity that for 𝑖 = 1, … , 𝑘 one
has 𝑡𝑖 = 𝑓(𝑠1𝑖 , … , 𝑠𝑛𝑖 ) and that 𝑓 does not appear in 𝑡𝑖 for 𝑖 > 𝑘. The terms
𝑠𝑗𝑖 have height 0, and thus are ℒ-terms. Let 𝑧1 , … , 𝑧𝑘 be new variables.
Then the ℒ-formula
𝑘
∃𝑧1 , … , 𝑧𝑘 (𝑅𝑧1 ⋯ 𝑧𝑘 𝑡𝑘+1 ⋯ 𝑡𝑚 ∧ 𝜑𝑓 (𝑧𝑖 , 𝑠1𝑖 , … , 𝑠𝑛𝑖 ))
⋀
𝑖=1
is equivalent in 𝑇 ′ to 𝜓′ . If 𝜓′ equals 𝑡1 = 𝑡2 , the argument is the same.

Now let 𝜓′ be an arbitrary ℒ′ -formula. By Lemma 3.3.1, we may as-
sume that all terms occurring in 𝜓′ have height at most 1. For any atomic
subformula 𝜒′ of 𝜓′ , one chooses an equivalent ℒ-formula 𝜒. (We just
constructed such a formula.) One then defines 𝜓 as the ℒ-formula ob-
tained by replacing any atomic subformula 𝜒′ of 𝜓′ by the corresponding
ℒ-formula 𝜒. It is a general fact that replacing subformulas by equiva-
lent ones produces an equivalent formula. (This is proved in Exercise
2.7.5, where the notion of a subformula is also defined). □
Example 3.3.3.
(1) Let 𝑇 ′ = Th(ℜ) be the theory of the ordered field of real num-
bers and 𝑇 the theory of the (pure) field of real numbers. Since
in ℝ we have 𝑟 < 𝑠 if and only if there exists 𝑡 ≠ 0 such that
𝑡2 = 𝑠 − 𝑟, 𝑇 ′ is an expansion by definition of 𝑇.
(2) Let 𝑇 be the ℒ𝑜𝑟𝑑 -theory of total orders without endpoints, and
let 𝑇 ′ be the ℒ𝑜.𝑟𝑖𝑛𝑔 -theory of ordered fields. Then 𝑇 ′ is a non-
conservative expansion of 𝑇, since for example 𝑇 ′ ⊧ ∀𝑥, 𝑦(𝑥 <
𝑦 → ∃𝑧(𝑥 < 𝑧 ∧ 𝑧 < 𝑦)), that is, the order in any ordered field is
dense (as any bounded interval has a midpoint), whereas this
is not a logical consequence of 𝑇.
(3) Let 𝑇 = Th(⟨ℵ1 ; <⟩). Then 𝜔 is a definable constant in 𝑇, that
is, there is a formula 𝜑(𝑥0 ) such that 𝑇 ⊧ ∃! 𝑥0 𝜑 and ⟨ℵ1 ; <⟩ ⊧
3.4. Quantifier Elimination 75
𝜑[𝜔]. Indeed, the set of limit ordinals is defined by the formula

∃𝑦 𝑦 < 𝑥 ∧ ∀𝑦∃𝑧(𝑦 < 𝑥 → 𝑦 < 𝑧 ∧ 𝑧 < 𝑥)
which we denote by Lim(𝑥), and it suffices to set 𝜑(𝑥0 ) equal
to the formula Lim(𝑥0 ) ∧ ∀𝑦(Lim(𝑦) → ¬𝑦 < 𝑥0 ).
Exercise 3.3.4. Let 𝑇 ⊆ 𝑇 ′ be theories in languages ℒ and ℒ′ ⊇ ℒ,

respectively. Prove that 𝑇 ′ is equivalent to an expansion by definition of
𝑇 if and only if every model of 𝑇 admits a unique expansion to a model
of 𝑇 ′ .
3.4. Quantifier Elimination

Using induction on height, it is easy to check that every formula 𝜑 is
logically equivalent to a formula 𝜓 which is in prenex form, that is, of
the form 𝑄1 𝑥1 ⋯ 𝑄𝑛 𝑥𝑛 𝜒, with 𝜒 quantifier-free and 𝑄𝑖 ∈ {∃, ∀} for all 𝑖.
In general, the number of alternations of quantifiers ∃ and ∀ provides
a good indication about the complexity of the formula 𝜑 (and of the set
defined by 𝜑 in a given structure). In concrete examples, sets defined by
a quantifier-free formula are usually easier to handle than those defined
by an arbitrary formula. For this reason, quantifier elimination results
are very useful and often open the way for a deeper understanding of the
theory under consideration.
We shall start with the following important technical result.
Theorem 3.4.1. Let 𝑇 be an ℒ-theory, 𝑛 ≥ 1 a natural number and

𝜑(𝑥1 , … , 𝑥𝑛 ) an ℒ-formula. The following properties are equivalent:
(1) There is a quantifier-free ℒ-formula 𝜓(𝑥1 , … , 𝑥𝑛 ) such that 𝜑 and
𝜓 are equivalent in 𝑇.
(2) Let 𝔐 and 𝔑 be models of 𝑇 and let 𝔄 be a common substructure
of 𝔐 and 𝔑. Then for any 𝑎 ∈ 𝐴𝑛 one has 𝔐 ⊧ 𝜑[𝑎] ⟺ 𝔑 ⊧
𝜑[𝑎].
Remark 3.4.2. When 𝑛 = 0 and 𝜑 is a sentence, one may consider 𝜑 as

𝜑(𝑥) and apply the theorem to 𝜑(𝑥) to find a quantifier-free formula 𝜓(𝑥)
which is equivalent to 𝜑(𝑥) in 𝑇. For instance, the theorem ∃𝑦𝑦 = 𝑦 of
𝑇 is equivalent in 𝑇 to the formula 𝑥 = 𝑥.
Note that if the language ℒ does not contain a constant symbol,

there is no quantifier-free ℒ-sentence. In this case, when we assert the
existence of a quantifier-free formula 𝜓 equivalent to a sentence 𝜑 in what
follows, we will allow that 𝜓 has one free variable.
Proof of Theorem 3.4.1. (1)⇒(2). We first observe that if 𝔄 ⊆ 𝔅,

𝜓(𝑥1 , … , 𝑥𝑛 ) is a quantifier-free formula and 𝑎 ∈ 𝐴𝑛 , then one has 𝔄 ⊧
𝜓[𝑎] ⟺ 𝔅 ⊧ 𝜓[𝑎]. Thus, if 𝔐 et 𝔑 are models of 𝑇 having 𝔄 as a com-
mon substructure and if 𝜑(𝑥1 , … , 𝑥𝑛 ) is equivalent in 𝑇 to the quantifier-
free formula 𝜓(𝑥1 , … , 𝑥𝑛 ), then for 𝑎 ∈ 𝐴𝑛 one has
𝔐 ⊧ 𝜑[𝑎] ⟺ 𝔐 ⊧ 𝜓[𝑎] ⟺ 𝔄 ⊧ 𝜓[𝑎]

⟺ 𝔑 ⊧ 𝜓[𝑎] ⟺ 𝔑 ⊧ 𝜑[𝑎].
(2)⇒(1). We consider the set of formulas

Γ(𝑥) ∶= {𝜒(𝑥1 , … , 𝑥𝑛 ) quantifier-free ∣ 𝑇 ⊧ ∀𝑥1 , … , 𝑥𝑛 (𝜑 → 𝜒)}.
We choose new pairwise distinct constants 𝑐1 , … , 𝑐𝑛 , and we consider
the theory Γ(𝑐) ∶= {𝜒(𝑐1 , … , 𝑐𝑛 ) ∣ 𝜒 ∈ Γ(𝑥)} in the augmented language
ℒ′ = ℒ ∪ {𝑐1 , … , 𝑐𝑛 }. We now prove that
(∗) 𝑇 ∪ Γ(𝑐) ⊧ 𝜑(𝑐).
If (∗) did not hold, one could find 𝔐′ ⊧ 𝑇 ∪ Γ(𝑐) ∪ {¬𝜑(𝑐)}. Let 𝔄′ ∶=
′ ′
⟨𝑐1𝔐 , … , 𝑐𝑛𝔐 ⟩𝔐′ = ⟨𝐴; …⟩ be the substructure generated by the elements
𝔐′
𝑐𝑖 in 𝔐′ . Observe that Γ(𝑐) ⊆ Δ(𝔄′ ). Let us prove that
Σ ∶= 𝑇 ∪ Δ(𝔄′ ) ∪ {𝜑(𝑐)}
has a model.
Otherwise, 𝑇 ∪ Δ(𝔄′ ) ⊧ ¬𝜑(𝑐). Since any element of 𝐴 can be writ-
ten as an ℒ′ -term, if one denotes by Δ𝑐 (𝔄′ ) the set of quantifier-free
ℒ′ -sentences in Δ(𝔄′ ), then 𝑇 ∪ Δ(𝔄′ ) is a conservative expansion of
𝑇 ∪ Δ𝑐 (𝔄′ ) by Proposition 3.3.2. In particular,
𝑇 ∪ Δ𝑐 (𝔄′ ) ⊧ ¬𝜑(𝑐).
Hence there exist quantifier-free ℒ-formulas 𝜉1 (𝑥), … , 𝜉𝑘 (𝑥) such that
𝑘 𝑘
𝑇⊧ 𝜉𝑖 (𝑐) → ¬𝜑(𝑐) and Δ(𝔄′ ) ⊧ 𝜉𝑖 (𝑐) =∶ 𝜉(𝑐).
⋀ ⋀
𝑖=1 𝑖=1
3.4. Quantifier Elimination 77
Since the constant symbols 𝑐𝑖 do not appear in 𝑇, or in 𝜑(𝑥) or 𝜉(𝑥), one

deduces (for instance by Lemma 2.6.8) that
𝑇 ⊧ ∀𝑥(𝜉(𝑥) → ¬𝜑(𝑥))
and then 𝑇 ⊧ ∀𝑥(𝜑(𝑥) → ¬𝜉(𝑥)). By definition it follows that ¬𝜉(𝑥) ∈
Γ(𝑥) and ¬𝜉(𝑐) ∈ Γ(𝑐), hence ¬𝜉(𝑐) ∈ Δ(𝔄′ ), which provides a contra-
diction.
Hence Σ has a model 𝔑∗ , and the ℒ-reduct 𝔑 of 𝔑∗ contains an iso-
morphic copy 𝔅′ of 𝔄′ as a substructure by Proposition 3.2.2. Up to iden-
tifying 𝔅′ and 𝔄′ , we have constructed two models 𝔐 = 𝔐′ ↾ ℒ and 𝔑
of 𝑇 containing a common substructure 𝔄 = 𝔄′ ↾ ℒ such that, if one sets
′
𝑎𝑖 = 𝑐𝑖𝔐 , then 𝔑 ⊧ 𝜑[𝑎] and 𝔐 ⊧ ¬𝜑[𝑎], which contradicts (2). We have
thus proved (∗).
By compactness there exist 𝜁1 (𝑐), … , 𝜁𝑚 (𝑐) ∈ Γ(𝑐) such that
𝑚
𝑇⊧ 𝜁𝑖 (𝑐) → 𝜑(𝑐).
⋀
𝑖=1
𝑚
As above, this implies that 𝑇 ⊧ ∀𝑥 (⋀𝑖=1 𝜁𝑖 (𝑥) → 𝜑(𝑥)). Since for all 𝑖
𝑚
we have 𝑇 ⊧ ∀𝑥(𝜑 → 𝜁𝑖 ), we infer that 𝑇 ⊧ ∀𝑥 (⋀𝑖=1 𝜁𝑖 (𝑥) ↔ 𝜑(𝑥)), with
𝑚
⋀𝑖=1 𝜁𝑖 (𝑥) a quantifier-free ℒ-formula. □
Definition. Let 𝑇 be an ℒ-theory. One says that 𝑇 admits quantifier
elimination (in the language ℒ) if every ℒ-formula 𝜑 is equivalent in 𝑇
to a quantifier-free ℒ-formula.
Lemma 3.4.3. Assume that for every quantifier-free formula 𝜑 and any
variable 𝑥 there exists a quantifier-free formula 𝜓 such that ∃𝑥𝜑 and 𝜓 are
equivalent in 𝑇. Then 𝑇 admits quantifier elimination.
Proof. Let 𝜓 and 𝜓′ be two formulas which are equivalent in 𝑇, which

we denote by 𝜓 ∼𝑇 𝜓′ . Since ¬𝜓 ∼𝑇 ¬𝜓′ , ∃𝑥𝜓 ∼𝑇 ∃𝑥𝜓′ and 𝜒 ∧ 𝜓 ∼𝑇
𝜒 ∧ 𝜓′ for any formula 𝜒, we can argue by induction on the height of
the formula, and the statement follows, by considering only formulas in
prenex form and eliminating one quantifier at the time. □
Theorem 3.4.4. Let 𝑇 be an ℒ-theory. One assumes that for any pair of
models 𝔐 and 𝔑 of 𝑇, for any common substructure 𝔄 of 𝔐 and 𝔑 and for
any quantifier-free formula 𝜑(𝑥0 , … , 𝑥𝑛 ), if there exist 𝑎 ∈ 𝐴𝑛 and 𝑏0 ∈ 𝑀

such that 𝔐 ⊧ 𝜑[𝑏0 , 𝑎], then there exists 𝑐0 ∈ 𝑁 such that 𝔑 ⊧ 𝜑[𝑐0 , 𝑎].
Then 𝑇 admits quantifier elimination.
Remark. The converse of this statement is clear: any theory which ad-
mits quantifier elimination satisfies the hypothesis of the theorem.
Proof of Theorem 3.4.4. Let 𝔄 ⊆ 𝔐, 𝔑 be given, with 𝔐, 𝔑 ⊧ 𝑇. Let 𝜑

be a quantifier-free formula and let 𝜒 be ∃𝑥0 𝜑. By hypothesis, we have
𝔐 ⊧ 𝜒[𝑎] ⟺ 𝔑 ⊧ 𝜒[𝑎] for every 𝑎 ∈ 𝐴𝑛 . It follows from Theorem 3.4.1
that 𝜒 is equivalent in 𝑇 to a quantifier-free formula, which is enough to
conclude by Lemma 3.4.3. □
Proposition 3.4.5. Let 𝑇 be a theory which admits quantifier elimination.
(1) Let 𝔐 and 𝔑 be models of 𝑇 with a common substructure. Then
𝔐 ≡ 𝔑.
(2) Let 𝔐 and 𝔑 be models of 𝑇. If 𝔐 ⊆ 𝔑, then 𝔐 ≼ 𝔑.
Proof. (1) This is a special case of the easy implication in Theorem 3.4.1.
Indeed, any sentence 𝜑 is equivalent in 𝑇 to a quantifier-free formula
𝜓(𝑥). Let 𝔄 be a common substructure of 𝔐 and 𝔑. For any 𝑎 ∈ 𝐴, one
then has
𝔐 ⊧ 𝜑 ⟺ 𝔐 ⊧ 𝜓[𝑎] ⟺ 𝔄 ⊧ 𝜓[𝑎]
⟺ 𝔑 ⊧ 𝜓[𝑎] ⟺ 𝔑 ⊧ 𝜑.
(2) This is a direct consequence of Theorem 3.4.1. □
3.5. Algebraically Closed Fields

In this section we treat an important and particularly nice example of
a theory from algebra, namely the ℒ𝑟𝑖𝑛𝑔 -theory ACF of algebraically
closed fields (see Example 2.6.3).
Let 𝐴 be a subring of a field 𝐾. An element of 𝐾 is said to be algebraic
over 𝐴 if it is the root of a non-zero polynomial with coefficients in 𝐴. If 𝐴
is an integral domain, an algebraic closure of 𝐴 is an algebraically closed
field 𝐾 containing 𝐴 such that any element of 𝐾 is algebraic over 𝐴.
3.5. Algebraically Closed Fields 79
In the following fact we list some results from field theory which we
will need in this and the next section.
Fact 3.5.1. Let 𝐴 be an integral domain.
(1) There exists an algebraic closure of 𝐴.
(2) If 𝐾 and 𝐾 ′ are algebraic closures of 𝐴, then there exists an iso-
morphism 𝑓 ∶ 𝐾 ≅ 𝐾 ′ such that 𝑓 ↾ 𝐴 = id𝐴 .
(3) Assume 𝐴 is a subring of an algebraically closed field 𝐿. Then
𝑎𝑙𝑔
the subfield 𝐴𝐿 = {𝑏 ∈ 𝐿 ∣ 𝑏 is algebraic over 𝐴} is an algebraic
closure of 𝐴.
𝑎𝑙𝑔
(4) Let 𝔽𝑝 be an algebraic closure of the field with 𝑝 elements 𝔽𝑝 .
𝑎𝑙𝑔
Then 𝔽𝑝 is an increasing union of finite subfields 𝐹𝑁 , 𝑁 ∈ ℕ.
More precisely, for any integer 𝑘 ≥ 1, the set of roots of the poly-
𝑘 𝑎𝑙𝑔
nomial 𝑋 𝑝 − 𝑋 in 𝔽𝑝 is a subfield 𝔽𝑝𝑘 (with 𝑝𝑘 elements), and
𝑎𝑙𝑔
⋃𝑘∈ℕ 𝔽𝑝𝑘 = 𝔽𝑝 . Since furthermore 𝔽𝑝𝑘 ⊆ 𝔽𝑝𝑙 when 𝑘 ∣ 𝑙, it is
enough to take 𝐹𝑁 ∶= 𝔽𝑝𝑁! .
(5) Any algebraically closed field is infinite.
(6) Let 𝐾 ⊆ 𝐿 be a field extension with 𝐾 an algebraically closed
field, and let 𝑏 ∈ 𝐿 ⧵ 𝐾. Then 𝑏 is not algebraic over 𝐾.
Theorem 3.5.2 (Chevalley-Tarski). The theory ACF admits quantifier
elimination.
Proof. First observe that a substructure of a field in ℒ𝑟𝑖𝑛𝑔 is nothing but

a subring. By Theorem 3.4.4 it is thus enough to prove that if 𝐾 and 𝐿 are
algebraically closed fields, 𝐴 is a common subring, and 𝜑(𝑥0 , … , 𝑥𝑛 ) is a
quantifier-free formula, then for any 𝑎 ∈ 𝐴𝑛 , if there exists 𝑏 ∈ 𝐿 such
that 𝐿 ⊧ 𝜑[𝑏, 𝑎] then there exists 𝑐 ∈ 𝐾 such that 𝐾 ⊧ 𝜑[𝑐, 𝑎].
By Fact 3.5.1, 𝐾 and 𝐿 contain algebraic closures 𝐹𝐾 and 𝐹𝐿 of 𝐴 that
are isomorphic via an isomorphism inducing the identity on 𝐴. Enlarg-
ing 𝐴 if necessary, we may thus assume that 𝐴 is an algebraically closed
field and even that 𝐴 = 𝐾 ⊆ 𝐿.
The formula 𝜑 is logically equivalent to a formula of the form
⋁𝑖 ⋀𝑗 𝜒𝑖,𝑗 , with each 𝜒𝑖,𝑗 (𝑥0 , … , 𝑥𝑛 ) either atomic or the negation
of an atomic formula. If 𝐿 ⊧ 𝜑[𝑏, 𝑎], there exists 𝑖 such that
𝐿 ⊧ ⋀𝑗 𝜒𝑖,𝑗 [𝑏, 𝑎]. It is thus enough to consider the case where 𝜑 is a

conjunction of atomic formulas and negations of atomic formulas. In
the theory of fields, any atomic formula is equivalent to 𝑃(𝑥) = 0 for
some polynomial 𝑃(𝑥) with integer coefficients. We may therefore as-
sume that 𝜑(𝑥) is of the form
𝑛 𝑚
𝑃𝑖 (𝑥) = 0 ∧ ¬𝑄𝑖 (𝑥) = 0.
⋀ ⋀
𝑖=1 𝑖=1
If one of the 𝑃𝑖 (𝑥0 , 𝑎1 , … , 𝑎𝑛 ) ∈ 𝐾[𝑥0 ] is a non-zero polynomial, then 𝑏 is

algebraic over 𝐾, which implies that 𝑏 ∈ 𝐾 and we are done.
𝑚
Thus we may assume that 𝜑 equals ⋀𝑖=1 ¬𝑄𝑖 (𝑥) = 0. By the exis-
tence of 𝑏, each polynomial 𝑄𝑖 (𝑥0 , 𝑎) ∈ 𝐾[𝑥0 ] is non-zero, and hence
has only a finite number of roots. The field 𝐾 is infinite, since it is alge-
braically closed, so there exists 𝑐 ∈ 𝐾 such that 𝐾 ⊧ 𝜑[𝑐, 𝑎]. □
Corollary 3.5.3. In 𝐾 ⊧ ACF, the definable sets (with parameters) are
precisely the constructible sets, that is, sets given by boolean combinations
of polynomial equations with coefficients from 𝐾. □
Let 𝑝 be a prime number or 𝑝 = 0. We denote by ACF𝑝 the theory

of algebraically closed fields of characteristic 𝑝.
Theorem 3.5.4. Let 𝑝 be a prime number or 𝑝 = 0. The theory ACF𝑝 is
complete.
Proof. Any field of characteristic 𝑝 > 0 contains 𝔽𝑝 as a subfield. If

𝐾 and 𝐿 are algebraically closed fields of characteristic 𝑝, then 𝐾 ≡ 𝐿
by Theorem 3.5.2 and Proposition 3.4.5(1), which proves that ACF𝑝 is
complete.
For ACF0 , the argument is the same, replacing 𝔽𝑝 by ℚ. □
Theorem 3.5.5 (Lefschetz Principle). Let 𝜑 be an ℒ𝑟𝑖𝑛𝑔 -sentence. The
following conditions are equivalent:
(1) ℂ ⊧ 𝜑.
(2) There exists an algebraically closed field of characteristic 0 in
which 𝜑 is satisfied.
(3) Any algebraically closed field of characteristic 0 satisfies 𝜑.
3.6. Ax’s Theorem 81
(4) There exists 𝑁 ∈ ℕ such that 𝜑 is satisfied in any algebraically

closed field of characteristic 𝑝 > 𝑁.
(5) There exists an infinite set of prime numbers 𝒫 such that for any
𝑝 ∈ 𝒫 there exists an algebraically closed field of characteristic
𝑝 in which 𝜑 is satisfied.
Proof. (1) ⟺ (2) ⟺ (3) follows from Theorem 3.5.4.

(3)⇒(4). Note that ACF0 is equal to ACF ∪ {𝜒𝑝 ∣ 𝑝 prime}, where
𝜒𝑝 expresses that 𝑝 = 1 + ⋯ + 1 is different from 0. If ACF0 ⊧ 𝜑, by
compactness there exists a finite subset Δ of ACF0 such that Δ ⊧ 𝜑. But
Δ contains only a finite set of sentences 𝜒𝑝 . Thus, there exists 𝑁 ∈ ℕ such
that 𝐾 ⊧ Δ for any algebraically closed field 𝐾 of characteristic 𝑝 > 𝑁.
For such a field 𝐾, one has 𝐾 ⊧ 𝜑.
(4)⇒(5) is clear.
(5)⇒(3). For 𝑝 ∈ 𝒫, let 𝐾𝑝 ⊧ ACF𝑝 such that 𝐾𝑝 ⊧ 𝜑. If ACF0 ⊭ 𝜑,
then ACF0 ⊧ ¬𝜑 by completeness. By the implication (3)⇒(4), there
exists 𝑁 ∈ ℕ such that ¬𝜑 is satisfied in any algebraically closed field of
characteristic 𝑝 > 𝑁. This forces 𝒫 to be finite, a contradiction. □
Theorem 3.5.6 (Hilbert’s Nullstellensatz). Let 𝐾 be an algebraically
closed field and 𝑃1 (𝑥), … , 𝑃𝑚 (𝑥) ∈ 𝐾[𝑥1 , … , 𝑥𝑛 ]. If the system of polyno-
mial equations 𝑃1 (𝑥) = 𝑃2 (𝑥) = … = 𝑃𝑚 (𝑥) = 0 has a solution in some
field 𝐿 ⊇ 𝐾, then it already has a solution in 𝐾.
Proof. Let 𝐿 ⊇ 𝐾 and 𝑎 ∈ 𝐿𝑛 be such that 𝑃1 (𝑎) = … = 𝑃𝑚 (𝑎) = 0.

Up to enlarging 𝐿 if necessary, we may assume that 𝐿 is algebraically
closed. Since ACF admits quantifier elimination, we have 𝐾 ≼ 𝐿 by
Proposition 3.4.5.
We now choose ℒ𝑟𝑖𝑛𝑔 -terms 𝐹𝑖 (𝑥, 𝑧𝑖 ) and tuples 𝑏𝑖 in 𝐾 such that
𝑃𝑖 (𝑥) = 𝐹𝑖 (𝑥, 𝑏𝑖 ). Then 𝐿 ⊧ ∃𝑥 ⋀ 𝐹𝑖 (𝑥, 𝑏𝑖 ) = 0, and therefore 𝐾 ⊧ ∃𝑥 ⋀
𝐹𝑖 (𝑥, 𝑏𝑖 ) = 0, since 𝐾 ≼ 𝐿. □
3.6. Ax’s Theorem

A chain of ℒ-structures is a sequence (𝔐𝑖 )𝑖∈ℕ of ℒ-structures such that
𝔐𝑖 ⊆ 𝔐𝑖+1 for any 𝑖.
If (𝔐𝑖 )𝑖∈ℕ is such a chain, there exists a unique ℒ-structure 𝔐 with

base set 𝑀 = ⋃𝑖∈ℕ 𝑀𝑖 such that 𝔐𝑖 ⊆ 𝔐 for any 𝑖. Indeed, the only way
to interpret the language symbols is to set 𝑐𝔐 = 𝑐𝔐0 , 𝑓𝔐 = ⋃𝑖∈ℕ 𝑓𝔐𝑖
and 𝑅𝔐 = ⋃𝑖∈ℕ 𝑅𝔐𝑖 , which is clearly well defined. The ℒ-structure 𝔐
obtained that way is denoted by ⋃𝑖∈ℕ 𝔐𝑖 .
Definition. A formula of the form ∀𝑥1 , … , 𝑥𝑛 ∃𝑦1 , … , 𝑦𝑚 𝜑, with 𝜑 quan-
tifier-free and 𝑚, 𝑛 ≥ 0, is called a ∀∃-formula.
Lemma 3.6.1 (Preservation of ∀∃-sentences under unions of chains).
Let 𝜓 be a ∀∃-sentence in ℒ and (𝔐𝑖 )𝑖∈ℕ a chain of ℒ-structures such that
𝔐𝑖 ⊧ 𝜓 for any 𝑖. Then 𝔐 = ⋃𝑖∈ℕ 𝔐𝑖 ⊧ 𝜓.
Proof. Let 𝜓 be the sentence ∀𝑥1 , … , 𝑥𝑛 ∃𝑦1 , … , 𝑦𝑚 𝜑(𝑥, 𝑦) with 𝜑 quan-

tifier-free. We have to prove that 𝔐 ⊧ ∃𝑦1 , … , 𝑦𝑚 𝜑[𝑎, 𝑦] for any 𝑎 ∈ 𝑀 𝑛 .
Since the sequence of base sets (𝑀𝑖 )𝑖∈ℕ is increasing, there exists 𝑘 ∈ ℕ
such that 𝑎 ∈ 𝑀𝑘𝑛 . Hence there exist 𝑏1 , … , 𝑏𝑚 ∈ 𝑀𝑘 such that 𝔐𝑘 ⊧
𝜑[𝑎, 𝑏], as 𝔐𝑘 ⊧ 𝜓. One deduces that 𝔐 ⊧ 𝜑[𝑎, 𝑏], since 𝜑 is quantifier-
free and 𝔐𝑘 is a substructure of 𝔐. □
Remark. It will be proved in Exercise 3.7.11 that a sentence is preserved
under unions of chains if and only if it is logically equivalent to a ∀∃-
sentence.
Proposition 3.6.2. Let 𝜑 be a ∀∃-sentence in the language ℒ𝑟𝑖𝑛𝑔 which is
satisfied in every finite field. Then ACF ⊧ 𝜑. In particular 𝜑 is satisfied in
ℂ.
𝑎𝑙𝑔
Proof. As recalled in Fact 3.5.1(4), 𝔽𝑝 is the union of a chain of finite
𝑎𝑙𝑔
fields. So it follows from Lemma 3.6.1 that one has 𝔽𝑝 ⊧ 𝜑 for every
prime 𝑝. The statement is now a consequence of the Lefschetz Principle.
□
Theorem 3.6.3 (Ax’s Theorem). Let 𝑓 ∶ ℂ𝑛 → ℂ𝑛 be a polynomial
mapping, that is, of the form 𝑓 = (𝑓1 , … , 𝑓𝑛 ) with 𝑓𝑖 ∈ ℂ[𝑥1 , … , 𝑥𝑛 ] poly-
nomials. If 𝑓 is injective, then 𝑓 is surjective.
Proof. There exist ℒ𝑟𝑖𝑛𝑔 -terms — which can be interpreted as polyno-

mials with integer coefficients — 𝑃𝑛,𝑑 (𝑧, 𝑥) such that, for any field 𝐾,
any polynomial 𝑔(𝑥) ∈ 𝐾[𝑥1 , … , 𝑥𝑛 ] of degree ≤ 𝑑 may be written as
3.7. Exercises 83
𝑃𝑛,𝑑 (𝑎, 𝑥) for some tuple 𝑎 of elements of 𝐾. The following sentence 𝜓𝑛,𝑑
is ∀∃ and expresses that any injective polynomial function 𝑓 ∶ 𝐾 𝑛 → 𝐾 𝑛 ,
with all polynomials 𝑓𝑖 of degree at most 𝑑, is surjective:
𝑛
′
∀𝑧1 , … , 𝑧𝑛 , 𝑢 ∃𝑥, 𝑥 [( 𝑃𝑛,𝑑 (𝑧𝑖 , 𝑥) = 𝑢𝑖 )
⋀
𝑖=1
𝑛 𝑛
′
∨( 𝑃𝑛,𝑑 (𝑧𝑖 , 𝑥) = 𝑃𝑛,𝑑 (𝑧𝑖 , 𝑥 ) ∧ ¬ 𝑥𝑖 = 𝑥𝑖′ )] .
⋀ ⋀
𝑖=1 𝑖=1
Since 𝜓𝑛,𝑑 is satisfied in every finite field, it follows from Proposi-

tion 3.6.2 that ACF ⊧ 𝜓𝑑,𝑛 , so in particular ℂ ⊧ 𝜓𝑑,𝑛 . □
3.7. Exercises
Exercise 3.7.1. Let 𝔐1 and 𝔐2 be two ℒ-structures. Prove that 𝔐1 ≡
𝔐2 if and only there exist 𝔑1 , 𝔑2 and 𝔓 such that 𝔐𝑖 ≅ 𝔑𝑖 and 𝔑𝑖 ≼ 𝔓
for 𝑖 = 1, 2.
Exercise 3.7.2. Let 𝑇 = Th(𝔑), where 𝔑 denotes the ℒ𝑎𝑟 -structure
⟨ℕ; 𝑆, 0, +, ⋅, <⟩. Prove that there exists 𝔑′ ⊧ 𝑇 such that 𝑁 ′ contains
a non-standard prime number, that is, an element 𝑝′ such that one has
𝔑′ ⊧ ∀𝑥(∃𝑦 𝑥 ⋅ 𝑦 = 𝑝′ → (𝑥 = 1 ∨ 𝑥 = 𝑝′ )) and 𝔑′ ⊧ (𝑆⏟ ⋯ 𝑆 0) < 𝑝′ for
𝑛 times
any 𝑛 ∈ ℕ.
Exercise 3.7.3. Prove that any theory 𝑇 admits an expansion by defini-
tion that admits quantifier elimination.
Exercise 3.7.4 (Vaught’s Criterion). Let ℒ be a first-order language and
𝜅 an infinite cardinal. An ℒ-theory 𝑇 is said to be 𝜅-categorical if all its
models of cardinality 𝜅 are isomorphic.
We assume that 𝜅 is larger than or equal to the cardinality of ℒ.
(1) Let 𝑇 be an ℒ-theory. Prove that any infinite model of 𝑇 is
elementarily equivalent to a model of 𝑇 of cardinality 𝜅.
(2) Prove that if a consistent theory 𝑇 is 𝜅-categorical and has no
finite model then it is complete.
Exercise 3.7.5. Let ℒ be a first-order language, and let 𝐼 be a non-empty

set. Let (𝔐𝑖 )𝑖∈𝐼 be a family of ℒ-structures. Finally, let 𝑈 be an ultrafilter
on 𝐼. (We refer to Exercise 1.11.8 for the notion of an ultrafilter.) If 𝑀𝑖 is
the underlying set of 𝔐𝑖 , we consider the following equivalence relation
on ∏𝐼 𝑀𝑖 : (𝑎𝑖 ) ≡𝑈 (𝑏𝑖 ) if and only if {𝑖 ∈ 𝐼 ∣ 𝑎𝑖 = 𝑏𝑖 } ∈ 𝑈. We denote
by 𝑀𝑈 the set of equivalence classes and by 𝜋 ∶ ∏𝐼 𝑀𝑖 → 𝑀𝑈 the
canonical projection.
(1) Prove that ≡𝑈 is indeed an equivalence relation.
(2) Prove that 𝑀𝑈 admits a unique ℒ-structure 𝔐𝑈 such that
for every atomic ℒ-formula 𝜑(𝑥1 , ⋯ , 𝑥𝑛 ) and every tuple
(𝑎1 , ⋯ , 𝑎𝑛 ) ∈ (∏𝐼 𝑀𝑖 )𝑛 one has
𝔐𝑈 ⊧ 𝜑[𝜋(𝑎1 ), ⋯ , 𝜋(𝑎𝑛 )]
if and only if the set of 𝑖 such that 𝔐𝑖 ⊧ 𝜑[𝑎1𝑖 , ⋯ , 𝑎𝑛𝑖 ] is in 𝑈.
The structure 𝔐𝑈 is called the ultraproduct of the 𝔐𝑖 with re-
spect to 𝑈. It is denoted by ∏𝑈 𝔐𝑖 .
(3) Prove Łoś’s Theorem: For every formula 𝜑(𝑥1 , ⋯ , 𝑥𝑛 ) and every
tuple (𝑎1 , ⋯ , 𝑎𝑛 ) ∈ (∏𝐼 𝑀𝑖 )𝑛 , one has
𝔐𝑈 ⊧ 𝜑[𝜋(𝑎1 ), ⋯ , 𝜋(𝑎𝑛 )]
if and only if {𝑖 ∈ 𝐼 ∣ 𝔐𝑖 ⊧ 𝜑[𝑎1𝑖 , ⋯ , 𝑎𝑛𝑖 ]} ∈ 𝑈.
(4) Use Łoś’s Theorem to give a direct proof of the Compactness
Theorem (Theorem 3.1.1).
[Hint: For a theory 𝑇, let 𝐼 be the set of all finite subsets of 𝑇.
Show that there is an ultrafilter 𝑈 on 𝐼 which contains 𝐴𝑇0 ∶=
{𝑇1 ∈ 𝐼 ∣ 𝑇1 ⊇ 𝑇0 } for every 𝑇0 ∈ 𝐼. Prove that if 𝔐𝑇0 ⊧ 𝑇0 for
every 𝑇0 , then ∏𝑈 𝔐𝑇0 is a model of 𝑇.]
(5) Let 𝑃 be the set of prime numbers, and let 𝑈 be a non-principal
ultrafilter on 𝑃. We work in ℒ𝑟𝑖𝑛𝑔 , and we consider the family
(𝔽𝑝 )𝑝∈𝑃 of fields with 𝑝 elements and their ultraproduct 𝔽𝑈
with respect to 𝑈. Observe that 𝔽𝑈 is a field. Determine its
characteristic.
(6) Let (𝔐𝑖 )𝑖∈ℕ be a family of infinite well-orderings, considered
in ℒ𝑜𝑟𝑑 . Let 𝑈 be a non-principal ultrafilter on ℕ, and let 𝔐𝑈
be the ultraproduct of the 𝔐𝑖 with respect to 𝑈.
3.7. Exercises 85
Prove that there is a strictly decreasing sequence in 𝔐𝑈 of

length ℵ1 . In particular, 𝔐𝑈 is not a well-ordering.
Exercise 3.7.6.
(1) We denote by DLO the ℒ𝑜𝑟𝑑 -theory of dense total orderings

without endpoints.
(a) Let 𝔐, 𝔐′ be models of DLO and let 𝔄 ⊆ 𝔐, 𝔄′ ⊆ 𝔐′ be
finite substructures. Assume that 𝑓 ∶ 𝔄 ≅ 𝔄′ is an ℒ𝑜𝑟𝑑 -
isomorphism and 𝑏 ∈ 𝑀. Prove that 𝑓 can be extended to
an isomorphism with domain 𝐴∪{𝑏} and image contained
in 𝑀 ′ .
(b) Deduce that DLO admits quantifier elimination and is com-
plete.
(c) Prove that DLO is ℵ0 -categorical.
(d) (More difficult.) Prove that for every cardinal 𝜅 > ℵ0 , the
theory DLO is not 𝜅-categorical.
(2) Let ℒ′ = ℒ𝑜𝑟𝑑 ∪ {𝑐𝑖 | 𝑖 ∈ ℕ}, with {𝑐𝑖 ∣ 𝑖 ∈ ℕ} a set of constant
′
symbols. Let DLO be the ℒ′ -theory obtained by adding to DLO
the axioms 𝑐𝑖 < 𝑐𝑗 for any 𝑖, 𝑗 ∈ ℕ with 𝑖 < 𝑗.
′
(a) Prove that DLO is a complete theory.
′
(b) Prove that the theory DLO has exactly three countable
models up to isomorphism.
Exercise 3.7.7. Let 𝑇 be the ℒ𝑜𝑟𝑑 -theory of total discrete orders with-
out endpoints, that is, totally ordered sets such that any element has an
(immediate) predecessor and an (immediate) successor. Observe that
⟨ℤ; <⟩ ⊧ 𝑇.
(1) Prove that 𝑇 does not admit quantifier elimination.

(2) Write an ℒ𝑜𝑟𝑑 -formula which defines the graph of the succes-
sor function in any model of 𝑇. Let 𝑇 ′ be the definitional ex-
pansion of 𝑇 in the language ℒ′ = {<, 𝑆}, where 𝑆 corresponds
to the successor function. Prove that 𝑇 ′ admits quantifier elim-
ination and is complete.
(3) Deduce that 𝑇 is complete.
Exercise 3.7.8. Let 𝑚 ≠ 𝑛 be natural numbers. Prove that there is an

ℒ𝑜𝑟𝑑 -sentence which is true in ⟨𝜔 ⋅ 𝑚, <⟩ and false in ⟨𝜔 ⋅ 𝑛, <⟩. Is there
an ℒ𝑜𝑟𝑑 -sentence which is true in ⟨𝜔𝑚 , <⟩ and false in ⟨𝜔𝑛 , <⟩?
Exercise 3.7.9. In this exercise, we consider the theory of the structure
𝔐0 ∶= ⟨ℚ; 0, +, <⟩. Recall that an ordered abelian group is an abelian
group 𝐴 endowed with a compatible total ordering (𝑎 < 𝑏 implies 𝑎+𝑐 <
𝑏+𝑐 for any 𝑎, 𝑏, 𝑐 ∈ 𝐴). We shall deal with ordered abelian groups in the
language ℒOAG ∶= {0, +, <}, and we denote by OAG the corresponding
theory.
Let DOAG be the theory of non-trivial divisible ordered abelian
groups. It is obtained by adding to OAG the following axioms:
(i) ∃𝑥 𝑥 ≠ 0 and
(ii) an axiom of the form ∀𝑥∃𝑦 𝑦
⏟⎵+⎵⏟⎵
⋯ ⎵⏟
+ 𝑦 = 𝑥 for any integer 𝑛 ≥
𝑛 times
1.
Observe that 𝔐0 ⊧ DOAG.
(1) Prove that if ⟨𝐷; 0, +, <⟩ ⊧ DOAG, then ⟨𝐷; <⟩ ⊧ DLO. (See
Exercise 3.7.6 for the theory DLO.)
(2) Prove that if ⟨𝐴; 0, +, <⟩ ⊧ OAG, then ⟨𝐴; 0, +⟩ is torsion-free,
that is, if 𝑛𝑎 = 0 for 𝑎 ∈ 𝐴 and 𝑛 ∈ ℕ∗ , then 𝑎 = 0.
(3) Let 𝔇 ⊧ DOAG and let 𝔄 = ⟨𝐴; 0, +, <⟩ ⊆ 𝔇. We denote by
div𝔇 (𝔄) the substructure with base set
{𝑑 ∈ 𝐷 ∣ ∃𝑛 ≥ 1 and 𝑎 ∈ 𝐴 such that 𝑎 + 𝑛 ⋅ 𝑑 ∈ 𝐴}.
Prove that if 𝔇, 𝔇′ ⊧ DOAG and 𝔄 ⊆ 𝔇, 𝔄′ ⊆ 𝔇′ , then every
isomorphism 𝑓 ∶ 𝔄 ≅ 𝔄′ extends uniquely to an isomorphism
𝑓 ̃ ∶ div𝔇 (𝔄) ≅ div𝔇′ (𝔄′ ).
(4) Prove that DOAG is complete and admits quantifier elimina-
tion.
(5) Deduce that if 𝔇 ⊧ DOAG and 𝜑(𝑥) is any (ℒOAG )𝐷 -formula,
then the subset of 𝐷 defined by 𝜑(𝑥) is a finite union of open
intervals and singletons.
Exercise 3.7.10 (Embedding Theorem). Given languages ℒ′ ⊆ ℒ and
an ℒ-theory 𝑇, we denote by 𝑇∀′ (or by 𝑇∀ in case ℒ = ℒ′ ) the theory of
3.7. Exercises 87
all universal ℒ′ -sentences (that is, sentences of the form ∀𝑥1 , … , 𝑥𝑛 𝜓 for
𝜓 quantifier-free) which are consequences of 𝑇.
(1) Prove the Embedding Theorem: An ℒ′ -structure 𝔐′ embeds
into the ℒ′ -reduct of a model of 𝑇 if and only if 𝔐′ ⊧ 𝑇∀′ .
[Hint: Use the diagram method to establish the non-trivial im-
plication.]
(2) Application. A group 𝐺 is called left-orderable if there exists
a total order < on 𝐺 such that for all elements 𝑥, 𝑦, 𝑔 from 𝐺,
𝑥 < 𝑦 entails 𝑔 ⋅ 𝑥 < 𝑔 ⋅ 𝑦. Now let ℒ𝑔𝑝 = {𝑒, ⋅, (⋅)−1 } be the
language of groups.
(a) Prove that the class of left-orderable groups may be axiom-
atized by a universal ℒ𝑔𝑝 -theory.
(b) Give an explicit (universal) axiomatzation in the case of
commutative groups.
(3) (a) Prove the Preservation Theorem for universal theories: 𝑇
may be axiomatzed by universal sentences if and only if it is
preserved under substructures, that is, any substructure of
a model of 𝑇 is a model of 𝑇.
(b) Let 𝑇 be the theory of fields in ℒ𝑟𝑖𝑛𝑔 . Determine the mod-
els of 𝑇∀ .
Exercise 3.7.11 (Preservation Theorem for ∀∃-theories). The aim of this
exercise is to prove that the following conditions (i) and (ii) are equiva-
lent for a theory 𝑇 (Chang-Łoś-Suszko Theorem):
(i) 𝑇 is preserved under unions of chains, that is, if (𝔐𝑖 )𝑖∈ℕ is a
chain of ℒ-structures such that 𝔐𝑖 ⊧ 𝑇 for all 𝑖, then ⋃𝑖∈ℕ 𝔐𝑖 ⊧
𝑇.
(ii) 𝑇 may be axiomatzed by ∀∃-sentences.
(1) Observe that (ii)⇒(i).

(2) Let 𝔐 ⊆ 𝔑. We say that 𝔐 is 1-elementary in 𝔑 and we write
𝔐 ≼1 𝔑, if for any existential formula 𝜑(𝑥1 , … , 𝑥𝑛 ) and any 𝑎 ∈
𝑀 𝑛 one has 𝔐 ⊧ 𝜑[𝑎1 , … , 𝑎𝑛 ] if and only if 𝔑 ⊧ 𝜑[𝑎1 , … , 𝑎𝑛 ].
(The same is then true for universal formulas as well. Why?)
Prove that if 𝔐 ≼1 𝔑, then there is 𝔐′ ⊇ 𝔑 such that
𝔐 ≼ 𝔐′ .
[Hint: Use the Embedding Theorem from Exercise 3.7.10.]

(3) Let 𝑇 be a theory which is preserved under unions of chains.
One sets
𝑇∀∃ = {𝜑 ∣ 𝜑 is a ∀∃-sentence in ℒ such that 𝑇 ⊧ 𝜑}.
Prove that every model of 𝑇∀∃ admits a 1-elementary extension
which is a model of 𝑇.
(4) Prove that (i)⇒(ii) and conclude.
(5) Prove that a sentence 𝜑 is preserved by unions of chains if and
only if it is logically equivalent to a ∀∃-sentence.
Exercise 3.7.12 (Indiscernible Sequences). Let 𝔐 be an ℒ-structure, and
let (𝐼, <) be an infinite totally ordered set. The Ehrenfeucht-Mostowski
type of a sequence 𝑎 = (𝑎𝑖 )𝑖∈𝐼 in 𝑀 is given by EM(𝑎) = (EM𝑛 (𝑎))𝑛≥1 ,
where for each 𝑛 ≥ 1, EM𝑛 (𝑎) denotes the set of ℒ-formulas 𝜑(𝑥1 , … , 𝑥𝑛 )
with 𝔐 ⊧ 𝜑[𝑎𝑖1 , … , 𝑎𝑖𝑛 ] for all 𝑖1 < ⋯ < 𝑖𝑛 in 𝐼. We call 𝑎 indiscernible
if for any ℒ-formula 𝜑(𝑥1 , … , 𝑥𝑛 ), either 𝜑 ∈ EM𝑛 (𝑎) or ¬𝜑 ∈ EM𝑛 (𝑎),
that is, if one has
(∗) 𝔐 ⊧ 𝜑[𝑎𝑖1 , … , 𝑎𝑖𝑛 ] if and only if 𝔐 ⊧ 𝜑[𝑎𝑗1 , … , 𝑎𝑗𝑛 ]
for all 𝑖1 < … < 𝑖𝑛 and 𝑗1 < … < 𝑗𝑛 from 𝐼.
Now let (𝐽, <) be an infinite totally ordered set and 𝑏 = (𝑏𝑗 )𝑗∈𝐽 an
arbitrary sequence in 𝑀. Prove that there exists 𝔑 ≽ 𝔐 containing an in-
discernible sequence (𝑎𝑖 )𝑖∈𝐼 such that EM𝑛 (𝑎) ⊇ EM𝑛 (𝑏) for all integers
𝑛 ≥ 1.
[Hint: Use Ramsey’s Theorem (cf. Exercise 2.7.1) to prove that for any
finite set of formulas Φ, there exists an infinite subset 𝐽 ′ of 𝐽 such that the
equivalence (∗) holds for all finite subtuples of 𝑏 indexed by increasing
sequences of elements in 𝐽 ′ and all formulas from Φ.]
10.1090/stml/089/04
Chapter 4
Recursive Functions
Introduction
In this chapter, we will develop a theory of ‘computable’ functions and
sets of natural numbers that will play a fundamental role in the study of
incompleteness of Peano arithmetic that will be undertaken in the next
chapter. The basic building block of computability is provided by the
notion of primitive recursive functions which we introduce in 4.1. Unfor-
tunately this notion is too restrictive since, as the example of the Ack-
ermann function shows (4.2), some obviously computable functions do
not belong to this class. A satisfactory definition of computable func-
tions is provided by the notion of general recursive functions which is
introduced in 4.3.
Another natural candidate for the class of computable functions are
the functions which are computable by a Turing machine that we in-
troduce in 4.4. It is a fundamental result that the two notions coincide:
a function is recursive if and only if it is Turing computable. One ad-
vantage of Turing computability is that it easily yields the existence of
universal recursive functions that play a major role in the theory.
We conclude this chapter with the study of recursively enumerable
sets (4.6). These are sets of natural numbers for which there exists an al-
gorithm that enumerates their members. A set is recursive if and only it
89
90 4. Recursive Functions
is recursively enumerable and its complement is also recursively enu-

merable. However not all recursively enumerable sets are recursive,
which leads to several negative results like the undecidability of the Halt-
ing Problem or Rice’s Theorem.
4.1. Primitive Recursive Functions

Notation. For 𝑛 ∈ ℕ, we set ℱ𝑛 ∶= {𝑓 ∶ ℕ𝑛 → ℕ}. Moreover, we set
ℱ ∶= ⋃𝑛∈ℕ ℱ𝑛 . The function 𝑓 ∈ ℱ𝑛 which to (𝑥1 , … , 𝑥𝑛 ) associates
𝑓(𝑥1 , … , 𝑥𝑛 ) ∈ ℕ will sometimes be denoted by
𝑓 = 𝜆𝑥1 ⋯ 𝑥𝑛 .𝑓(𝑥1 , … , 𝑥𝑛 ).
Definition. The set of primitive recursive functions is the smallest sub-

set 𝐸 of ℱ which satisfies the following properties:
(R0) 𝐸 contains the following basic functions:
• 𝑆 = 𝜆𝑥.𝑥 + 1 (the successor function),
• the 0-ary (constant) function 𝐶00 equal to 0, that is, 𝐶00 =
𝜆.0,
• for any 𝑛 ∈ ℕ∗ and any 1 ≤ 𝑖 ≤ 𝑛 the projection 𝑃𝑖𝑛 on the
𝑖th coordinate, that is, 𝑃𝑖𝑛 = 𝜆𝑥1 ⋯ 𝑥𝑛 .𝑥𝑖 .
(R1) 𝐸 is stable under composition: if 𝑓1 , … , 𝑓𝑛 ∈ ℱ𝑝 ∩ 𝐸 and ℎ ∈
ℱ𝑛 ∩ 𝐸, then 𝑔 = ℎ(𝑓1 , … , 𝑓𝑛 ) ∈ 𝐸.
(R2) 𝐸 is stable under recursion: if 𝑔 ∈ ℱ𝑛 ∩ 𝐸 and ℎ ∈ ℱ𝑛+2 ∩ 𝐸,
then the function 𝑓 defined by
𝑓(𝑥1 , … , 𝑥𝑛 , 0) ∶= 𝑔(𝑥1 , … , 𝑥𝑛 ),
𝑓(𝑥1 , … , 𝑥𝑛 , 𝑦 + 1) ∶= ℎ(𝑥1 , … , 𝑥𝑛 , 𝑦, 𝑓(𝑥1 , … , 𝑥𝑛 , 𝑦))
is also in 𝐸.
We observe that the constant functions 𝐶𝑘𝑛 = 𝜆𝑥1 ⋯ 𝑥𝑛 .𝑘 are primi-
tive recursive for any 𝑘, 𝑛 ∈ ℕ. (Note that the function 𝐶01 is obtained by
recursion from 𝑔 = 𝐶00 and ℎ = 𝑃22 .)
If 𝑋 ⊆ ℕ𝑛 , its characteristic function is denoted by 𝟙𝑋 , that is,
1 if (𝑥1 , … , 𝑥𝑛 ) ∈ 𝑋;
𝟙𝑋 (𝑥1 , … , 𝑥𝑛 ) = {
0 else.
4.1. Primitive Recursive Functions 91
Lemma 4.1.1. The functions 𝜆𝑥𝑦.𝑥+𝑦, 𝜆𝑥𝑦.𝑥⋅𝑦, 𝜆𝑥𝑦.𝑥𝑦 , 𝜆𝑥.𝑥!, 𝜆𝑥𝑦.𝑥−𝑦 ̇

and 𝟙ℕ∗ are primitive recursive. Here, 𝑥−𝑦
̇ ∶= 𝑥 − 𝑦 if 𝑥 ≥ 𝑦, and 𝑥−𝑦
̇ ∶=
0, else.
Proof. The addition 𝑓 = 𝜆𝑥𝑦.𝑥 + 𝑦 is obtained by recursion from 𝑔 = 𝑃11

and ℎ = 𝑆 ∘ 𝑃33 , as 𝑥 + 0 = 𝑥 and 𝑥 + (𝑦 + 1) = 𝑆(𝑥 + 𝑦).
The other functions are easily obtained as well. Let us give the ar-
gument for 𝜆𝑥𝑦.𝑥−𝑦 ̇ and 𝟙ℕ∗ . One first constructs 𝜆𝑦.𝑦−1
̇ by recursion,
via 0−1
̇ = 0, (𝑦 + 1)−1̇ = 𝑦. One may then define 𝑥−0 ̇ = 𝑥, 𝑥−(𝑦
̇ + 1) =
(𝑥−𝑦)
̇ −1.
̇ Finally, 𝟙ℕ∗ (𝑥) = 1−(1
̇ −𝑥).
̇ □
Definition. A subset 𝑋 of ℕ𝑛 is called primitive recursive if its charac-
teristic function 𝟙𝑋 is primitive recursive.
Lemma 4.1.2.
(1) The set of primitive recursive functions is stable under a permu-
tation of variables.
(2) If 𝑋 ⊆ ℕ𝑛 is primitive recursive and if 𝑓1 , … , 𝑓𝑛 ∈ ℱ𝑝 are primitive
recursive, then the set
𝑌 = {(𝑥1 , … , 𝑥𝑝 ) ∈ ℕ𝑝 ∣ (𝑓1 (𝑥), … , 𝑓𝑛 (𝑥)) ∈ 𝑋}
is primitive recursive.
(3) The set of primitive recursive subsets of ℕ𝑛 contains ∅ and ℕ𝑛 ,
and it is stable under ∪, ∩ and passage to the complement (so
under boolean combinations).
(4) The set {(𝑥, 𝑦) ∣ 𝑥 < 𝑦} ⊆ ℕ2 is primitive recursive.
(5) (Definition by cases.) Let ℕ𝑛 = 𝐴1 ∪̇ … ∪𝐴 ̇ 𝑘 be a partition of ℕ𝑛
into finitely many primitive recursive sets 𝐴𝑖 , and let 𝑓1 , … , 𝑓𝑘 ∈
ℱ𝑛 be primitive recursive. Then the function 𝑓, defined by 𝑓(𝑥) =
𝑓𝑖 (𝑥) if 𝑥 ∈ 𝐴𝑖 , is primitive recursive. In particular, the functions
𝜆𝑥1 ⋯ 𝑥𝑛 . max(𝑥𝑖 ) and 𝜆𝑥1 ⋯ 𝑥𝑛 . min(𝑥𝑖 ) are primitive recur-
sive.
(6) (Bounded sums and products.) If 𝑓 ∈ ℱ𝑛+1 is primitive recursive,
so are the following functions:
𝑦 𝑦
𝜆𝑥1 ⋯ 𝑥𝑛 𝑦. ∑ 𝑓(𝑥, 𝑡) and 𝜆𝑥1 ⋯ 𝑥𝑛 𝑦. ∏ 𝑓(𝑥, 𝑡).
𝑡=0 𝑡=0
(7) (Bounded 𝜇-operator.) Let 𝑋 ⊆ ℕ𝑛+1 be primitive recursive. Then

𝑓 ∈ ℱ𝑛+1 , 𝑓(𝑥, 𝑧) = (𝜇𝑡 ≤ 𝑧) ((𝑥, 𝑡) ∈ 𝑋), defined by
⎧0 if there is no 𝑡 ≤ 𝑧 with (𝑥, 𝑡) ∈ 𝑋;

𝑓(𝑥, 𝑧) ∶= 𝑡0 if 𝑡0 is the smallest natural number
⎨
⎩ 𝑡 ≤ 𝑧 such that (𝑥, 𝑡) ∈ 𝑋
is a primitive recursive function.

(8) (Bounded quantification.) If 𝑋 ⊆ ℕ𝑛+1 is a primitive recursive
set, so are the sets
𝑋𝑒 = {(𝑥1 , … , 𝑥𝑛 , 𝑧) ∈ ℕ𝑛+1 ∣ (∃𝑡 ≤ 𝑧) (𝑥, 𝑡) ∈ 𝑋} and

𝑋𝑎 = {(𝑥1 , … , 𝑥𝑛 , 𝑧) ∈ ℕ𝑛+1 ∣ (∀𝑡 ≤ 𝑧) (𝑥, 𝑡) ∈ 𝑋}.
Proof. Part (1) is clear, and (2) follows from 𝟙𝑌 = 𝟙𝑋 (𝑓1 , … , 𝑓𝑛 ).

(3) One has 𝟙ℕ𝑛 ⧵𝑋 = 1−𝟙
̇ 𝑋 and 𝟙𝑋∩𝑌 = 𝟙𝑋 ⋅ 𝟙𝑌 .
(4) One has 𝟙< (𝑥, 𝑦) = 𝟙ℕ∗ (𝑦−𝑥).
̇
(5) One has 𝑓 = 𝟙𝐴1 𝑓1 + … + 𝟙𝐴𝑘 𝑓𝑘 .
(6) is proved by recursion.
(7) Given 𝑋 ⊆ ℕ𝑛+1 , one has 𝑓(𝑥, 0) = 0,
𝑧
⎧𝑓(𝑥, 𝑧) if ∑𝑡=0 𝟙𝑋 (𝑥, 𝑡) ≥ 1;
𝑧
𝑓(𝑥, 𝑧 + 1) = 𝑧 + 1 if ∑𝑡=0 𝟙𝑋 (𝑥, 𝑡) = 0 and (𝑥, 𝑧 + 1) ∈ 𝑋;
⎨
⎩0 else.
Thus 𝑓 is primitive recursive (by recursion and definition by cases).

(8) Passing to the complement, it is enough to treat the first case.
𝑧
One has 𝟙𝑋𝑒 (𝑥, 𝑧) = 1 if ∑𝑡=0 𝟙𝑋 (𝑥, 𝑡) ≥ 1, and 𝟙𝑋𝑒 (𝑥, 𝑧) = 0, else. □
4.1. Primitive Recursive Functions 93
Example 4.1.3.
(1) Let 𝑞 ∶ ℕ2 → ℕ be the function which to (𝑥, 𝑦) associates the
integer part of 𝑥/𝑦 if 𝑦 ≠ 0, and 0 else. Then 𝑞 is primitive
recursive. [Indeed, 𝑞(𝑥, 𝑦) = (𝜇𝑡 ≤ 𝑥) ((𝑡 + 1) ⋅ 𝑦 > 𝑥).]
(2) {(𝑥, 𝑦) ∈ ℕ2 ∶ 𝑦 ∣ 𝑥} is a primitive recursive set. [Indeed, one
has 𝑦 ∣ 𝑥 ⟺ ∃𝑧𝑧 ⋅ 𝑦 = 𝑥 ⟺ 𝑥 = 𝑞(𝑥, 𝑦) ⋅ 𝑦.]
(3) The set 𝑃 of prime numbers is primitive recursive. [Indeed,
𝑥 ∈ 𝑃 ⟺ 𝑥 ≥ 2∧(∀𝑦 ≤ 𝑥) (𝑦 ∣ 𝑥 → (𝑦 = 1 ∨ 𝑦 = 𝑥)). By the
closure properties stated in Lemma 4.1.2, 𝑃 is thus primitive
recursive.]
(4) The function 𝜋, which to 𝑥 associates the (𝑥 + 1)th prime num-
ber, is primitive recursive. [Indeed, one has 𝜋(0) = 2, 𝜋(𝑥 +
1) = (𝜇𝑧 ≤ 𝜋(𝑥)! +1) (𝑧 > 𝜋(𝑥) ∧ 𝑧 ∈ 𝑃).]
(5) A primitive recursive bijection between ℕ2 and ℕ is given by
1
the function 𝛼2 = 𝜆𝑥𝑦. (𝑥 + 𝑦 + 1)(𝑥 + 𝑦) + 𝑥. Moreover,
2
there are primitive recursive functions 𝛽12 , 𝛽22 ∈ ℱ1 such that
𝛼2 (𝛽12 , 𝛽22 ) = idℕ . [Indeed, as 𝛼2 (𝑥, 𝑦) ≥ min(𝑥, 𝑦), one has
𝛽12 (𝑥) = (𝜇𝑧 ≤ 𝑥)(∃𝑡 ≤ 𝑥)(𝛼2 (𝑧, 𝑡) = 𝑥)), and similarly for the
function 𝛽22 .]
By induction on 𝑝 ≥ 2, one defines a primitive recur-
sive bijection 𝛼𝑝 ∶ ℕ𝑝 → ℕ whose inverse has primitive re-
𝑝 𝑝
cursive components 𝛽1 , … , 𝛽𝑝 . For this, it suffices to set
𝛼𝑝+1 (𝑥1 , … , 𝑥𝑝+1 ) = 𝛼𝑝 (𝑥1 , … , 𝑥𝑝−1 , 𝛼2 (𝑥𝑝 , 𝑥𝑝+1 )).
If (𝑥0 , … , 𝑥𝑛−1 ) is a finite sequence of natural numbers, we define its

Gödel number as
⟨𝑥0 , … , 𝑥𝑛−1 ⟩ ∶= 𝜋(0)𝑥0 ⋅ … ⋅ 𝜋(𝑛 − 2)𝑥𝑛−2 ⋅ 𝜋(𝑛 − 1)𝑥𝑛−1 +1 − 1.
Lemma 4.1.4. The map ⟨ ⟩ defines a bijection between the set of finite
sequences of natural numbers and the set of natural numbers ℕ. It satisfies
the following properties:
(1) The binary component function (𝑥)𝑖 , defined by
𝑥𝑖 if 𝑖 < 𝑛,
(⟨𝑥0 , … , 𝑥𝑛−1 ⟩)𝑖 = {
0 else,
is primitve recursive.
(2) The length function, given by lg(⟨𝑥0 , … , 𝑥𝑛−1 ⟩) = 𝑛, is primitive
recursive.
(3) For any 𝑛 ∈ ℕ, ⟨ ⟩ ↾ ℕ𝑛 ∶ ℕ𝑛 → ℕ is primitive recursive.
(4) lg(𝑥) ≤ 𝑥 for any 𝑥, and if 𝑥 > 0, then (𝑥)𝑖 < 𝑥 for any 𝑖.
Proof. Part (4) is clear, and (3) follows from Example 4.1.3.
(2) One has lg(𝑥) = (𝜇𝑦 ≤ 𝑥) [(∀𝑧 ≤ 𝑥) (𝑦 ≤ 𝑧 → 𝜋(𝑧) ∤ (𝑥 + 1))] if
𝑥 > 0 and lg(0) = 0, so lg is primitive recursive by Lemma 4.1.2.
(1) The description
⎧0 if 𝑖 ≥ lg(𝑥),
(𝑥)𝑖 = (𝜇𝑦 ≤ 𝑥) (𝜋(𝑖)𝑦+2 ∤ (𝑥 + 1)) if 𝑖 + 1 = lg(𝑥),
⎨ 𝑦+1
⎩(𝜇𝑦 ≤ 𝑥) (𝜋(𝑖) ∤ (𝑥 + 1)) if 𝑖 + 1 < lg(𝑥)
of the component function yields its primitive recursiveness. □
4.2. The Ackermann Function

We now define the Ackermann function 𝜉 ∈ ℱ2 , a function which is com-
putable in the intuitive sense but which turns out to be non-primitive
recursive:
• 𝜉(0, 𝑥) ∶= 2𝑥 ,
• 𝜉(𝑦, 0) ∶= 1,
• 𝜉(𝑦 + 1, 𝑥 + 1) ∶= 𝜉(𝑦, 𝜉(𝑦 + 1, 𝑥)).
This function is well-defined and may be computed in a finite number
of steps (exercise; it suffices to consider the lexicographic order on ℕ2 ).
We will also write 𝜉𝑛 (𝑥) for 𝜉(𝑛, 𝑥), and 𝜉𝑛𝑘 for ⏟
𝜉𝑛⎵⎵
∘⏟
…⎵∘⎵⏟
𝜉𝑛 .
𝑘 times
Lemma 4.2.1.
(1) 𝜉𝑛 (𝑥) > 𝑥 for all 𝑛, 𝑥 ∈ ℕ.
(2) 𝜉𝑛 is strictly increasing for all 𝑛 ∈ ℕ.
(3) 𝜉𝑛+1 (𝑥) ≥ 𝜉𝑛 (𝑥) for all 𝑛, 𝑥 ∈ ℕ.
(4) 𝜉𝑛𝑘 is strictly increasing for all 𝑘, 𝑛 ∈ ℕ.
4.2. The Ackermann Function 95
(5) 𝜉𝑛𝑘 (𝑥) < 𝜉𝑛𝑘+1 (𝑥) for all 𝑘, 𝑛, 𝑥 ∈ ℕ.

𝑘
(6) 𝜉𝑚 (𝑥) ≤ 𝜉𝑛𝑘 (𝑥) for all 𝑘, 𝑚, 𝑛, 𝑥 ∈ ℕ with 𝑚 ≤ 𝑛.
(7) 𝜉𝑛𝑘 (𝑥) ≤ 𝜉𝑛+1 (𝑥 + 𝑘) for all 𝑘, 𝑛, 𝑥 ∈ ℕ.
Proof. We prove (1) and (2) simultaneously by induction on 𝑛. The case

𝑛 = 0 is clear. Suppose now that (1) and (2) hold for 𝑛. By the induction
hypothesis, one has 𝜉𝑛+1 (𝑥 + 1) = 𝜉𝑛 (𝜉𝑛+1 (𝑥)) > 𝜉𝑛+1 (𝑥), showing that
(2) holds for 𝑛 + 1. As 𝜉𝑛+1 (0) = 1 by definition, (1) follows from (2).
(3) 𝜉𝑛+1 (0) = 𝜉𝑛 (0), and 𝜉𝑛+1 (𝑥 + 1) = 𝜉𝑛 (𝜉⏟⎵⏟⎵
𝑛+1 (𝑥)
⏟) ≥ 𝜉𝑛 (𝑥 + 1) by
≥𝑥+1
(1) and (2).
(4), (5) and (6) are clear.
(7) This is proved by induction on 𝑘, the case 𝑘 = 0 being clear. Now
suppose that the result holds for 𝑘. Using (3), one then has 𝜉𝑛𝑘+1 (𝑥) =
𝜉𝑛 (𝜉𝑛𝑘 (𝑥)) ≤ 𝜉𝑛 (𝜉𝑛+1 (𝑥 + 𝑘)) = 𝜉𝑛+1 (𝑥 + 𝑘 + 1). □
Definition. We say that a function 𝑓 ∈ ℱ1 dominates the function
𝑔 ∈ ℱ𝑛 if there exists 𝑁 ∈ ℕ such that for all 𝑥 ∈ ℕ𝑛 one has 𝑔(𝑥) ≤
𝑓 (max(𝑥1 , … , 𝑥𝑛 , 𝑁)).
Note that if 𝑓 is strictly increasing, then 𝑓 dominates 𝑔 if and only if

𝑔(𝑥) ≤ 𝑓(max(𝑥1 , … , 𝑥𝑛 )) except for a finite number of 𝑛-tuples.
We denote by 𝒞𝑛 the set of functions in ℱ which are dominated by
at least one of the functions 𝜉𝑛𝑘 , where 𝑘 ∈ ℕ.
Lemma 4.2.2.
(1) 𝒞𝑛 ⊆ 𝒞𝑚 for any 𝑛 ≤ 𝑚.
(2) 𝒞0 contains the basic functions (𝑆, 𝐶00 and all the projections 𝑃𝑖𝑛 )
as well as the functions 𝜆𝑥𝑦.𝑥 + 𝑦, 𝜆𝑥.𝑘 ⋅ 𝑥 ( for fixed 𝑘) and
𝜆𝑥1 … 𝑥𝑛 . max(𝑥1 , … , 𝑥𝑛 ).
(3) 𝒞𝑛 is stable under composition.
(4) If 𝑔 ∈ ℱ𝑝 ∩ 𝒞𝑛 and ℎ ∈ ℱ𝑝+2 ∩ 𝒞𝑛 , then the function 𝑓 defined
by recursion from 𝑔 and ℎ is in 𝒞𝑛+1 .
In particular, the set 𝒞 = ⋃𝑛∈ℕ 𝒞𝑛 contains all primitive recursive
functions.
Proof. (1) and (2) are clear.

(3) Suppose that 𝑓1 , … , 𝑓𝑚 ∈ ℱ𝑝 ∩ 𝒞𝑛 and 𝑔 ∈ ℱ𝑚 ∩ 𝒞𝑛 are given.
There exist natural numbers 𝑁, 𝑁1 , … , 𝑁𝑚 , 𝑘, 𝑘1 , … , 𝑘𝑚 such that 𝑔(𝑦) ≤
𝜉𝑛𝑘 (max(𝑦1 , … , 𝑦𝑚 , 𝑁)) for every 𝑦 ∈ ℕ𝑚 and, for any 𝑖 ∈ {1, … , 𝑚} and
𝑘
𝑥 ∈ ℕ𝑝 , 𝑓𝑖 (𝑥) ≤ 𝜉𝑛 𝑖 (max(𝑥1 , … , 𝑥𝑝 , 𝑁𝑖 )). Set 𝑀 = max(𝑁, 𝑁1 , … , 𝑁𝑚 ) and
𝑙 = max(𝑘1 , … , 𝑘𝑚 ). Now let 𝑥 ∈ ℕ𝑝 . Setting 𝑀𝑥 ∶= max(𝑥1 , … , 𝑥𝑝 , 𝑀),
one then has
𝑔(𝑓1 (𝑥), … , 𝑓𝑚 (𝑥)) ≤ 𝜉𝑛𝑘 (𝜉𝑛𝑙 (𝑀𝑥 )) = 𝜉𝑛𝑘+𝑙 (𝑀𝑥 ).
(4) By definition of the recursion, one has 𝑓(𝑥, 0) = 𝑔(𝑥) and

𝑓(𝑥, 𝑦+1) = ℎ(𝑥, 𝑦, 𝑓(𝑥, 𝑦)). By assumption, there exist natural numbers
𝑘
𝑘1 , 𝑁1 , 𝑘2 , 𝑁2 such that 𝑔(𝑥) ≤ 𝜉𝑛 1 (max(𝑥1 , … , 𝑥𝑝 , 𝑁1 )) and ℎ(𝑥, 𝑦, 𝑡) ≤
𝑘
𝜉𝑛 2 (max(𝑥1 , … , 𝑥𝑝 , 𝑦, 𝑡, 𝑁2 )). Using induction on 𝑦, one easily obtains
that
𝑘 +𝑦𝑘2
(∗) 𝑓(𝑥, 𝑦) ≤ 𝜉𝑛 1 (max(𝑥1 , … , 𝑥𝑝 , 𝑦, 𝑁1 , 𝑁2 )).
Applying Lemma 4.2.1(7), one deduces from (∗) that

𝑓(𝑥, 𝑦) ≤ 𝜉⏟⎵⎵⎵⎵⎵⎵⎵⎵⎵⎵⎵⎵⏟⎵⎵⎵⎵⎵⎵⎵⎵⎵⎵⎵⎵⏟
𝑛+1 (max(𝑥1 , … , 𝑥𝑝 , 𝑦, 𝑁1 , 𝑁2 ) + 𝑘1 + 𝑘2 𝑦),
a composition of functions in 𝒞𝑛+1
whence the result by (3). □

Theorem 4.2.3. The Ackermann function is not primitive recursive.
Proof. Suppose, for contradiction, that 𝜉 is primitive recursive. Then

𝜆𝑥.𝜉(𝑥, 2𝑥) is primitive recursive, too, and so, by Lemma 4.2.2, there exist
𝑘, 𝑛, 𝑁 ∈ ℕ such that 𝜉(𝑥, 2𝑥) ≤ 𝜉𝑛𝑘 (𝑥) for all 𝑥 > 𝑁. By Lemma 4.2.1(7)
it follows that 𝜉𝑥 (2𝑥) ≤ 𝜉𝑛+1 (𝑥 + 𝑘) for all 𝑥 > 𝑁. For 𝑥 > max(𝑘, 𝑛 + 1),
this is a contradiction. □
Remark. An easy induction on 𝑛 shows that all the functions 𝜉𝑛 are
primitive recursive.
4.3. Partial Recursive Functions

The existence of the Ackermann function — computable in the intuitive
sense but not primitive recursive — shows that we ought to enlarge the
4.3. Partial Recursive Functions 97
class of functions we are considering. For technical reasons it is more-

over convenient to pass to partial functions.
By definition, a partial function from ℕ𝑛 to ℕ is a pair (𝐴, 𝑓), with
𝐴 ⊆ ℕ𝑛 and 𝑓 ∶ 𝐴 → ℕ. The set 𝐴 is called the domain (of definition)
of 𝑓 and denoted by dom(𝑓). We write ℱ𝑛∗ for the set of such pairs, and
ℱ ∗ = ⋃𝑛∈ℕ ℱ𝑛∗ . Most of the time, we will write 𝑓 instead of (𝐴, 𝑓).
Definition. The set of partial recursive functions is the smallest subset

𝐸 of ℱ ∗ which satisfies the following properties (R0)∗ -(R3)∗ :
(R0)∗ 𝐸 contains the basic functions (𝑆, 𝐶00 and all the projections
𝑃𝑖𝑛 ).
(R1)∗ 𝐸 is stable under composition (of partial functions): Given par-
tial functions 𝑓1 , … , 𝑓𝑛 ∈ ℱ𝑚∗ ∩ 𝐸 and ℎ ∈ ℱ𝑛∗ ∩ 𝐸, then
𝑔 = ℎ(𝑓1 , … , 𝑓𝑛 ) ∈ 𝐸, where 𝑔 is the partial function which
to 𝑥 assigns ℎ(𝑓1 (𝑥), … , 𝑓𝑛 (𝑥)) if 𝑥 ∈ dom(𝑓𝑖 ) for all 𝑖 and
(𝑓1 (𝑥), … , 𝑓𝑛 (𝑥)) ∈ dom(ℎ). Otherwise, 𝑔 is not defined for 𝑥.
(R2)∗ 𝐸 is stable under recursion (of partial functions): Given 𝑔 ∈
∗
ℱ𝑝∗ ∩ 𝐸 and ℎ ∈ ℱ𝑝+2 ∩ 𝐸, then 𝑓 ∈ 𝐸, where
• 𝑓(𝑥, 0) = 𝑔(𝑥) if 𝑥 ∈ dom(𝑔), and otherwise 𝑓 is not de-
fined for (𝑥, 0);
• 𝑓(𝑥, 𝑦 + 1) = ℎ(𝑥, 𝑦, 𝑓(𝑥, 𝑦)) if 𝑓 is defined for (𝑥, 𝑦) (this
property is defined simultaneously by recursion on 𝑦) and
if (𝑥, 𝑦, 𝑓(𝑥, 𝑦)) ∈ dom(ℎ), and otherwise 𝑓 is not defined
for (𝑥, 𝑦 + 1).
∗
(R3)∗ 𝐸 is stable under the 𝜇-operator: Given 𝑓 ∈ ℱ𝑛+1 ∩ 𝐸, then
∗
𝑔 ∈ 𝐸, where 𝑔 ∈ ℱ𝑛 , 𝑔(𝑥) = 𝜇𝑦 (𝑓(𝑥, 𝑦) = 0) is the partial
function defined as follows:
• if there is 𝑧 such that 𝑓(𝑥, 𝑧) = 0 and (𝑥, 𝑧′ ) ∈ dom(𝑓) for
all 𝑧′ ≤ 𝑧, then 𝑔(𝑥) is the minimal such 𝑧;
• otherwise, 𝑔 is not defined for 𝑥.
A partial function 𝑓 ∈ ℱ𝑛∗ is total if dom(𝑓) = ℕ𝑛 . A (total) recursive

function is a total function which is partial recursive.
The following operation is reserved for total functions.
Let 𝑓 ∈ ℱ𝑛+1 be a total function. Assume that 𝑓 satisfies

∀𝑥1 , … , 𝑥𝑛 ∃𝑦𝑓(𝑥, 𝑦) = 0.
Then we say that the (total) function 𝑔 = 𝜆𝑥.𝜇𝑦 (𝑓(𝑥, 𝑦) = 0) is obtained
from 𝑓 by the total 𝜇-operator.
We say that a set of total functions satisfies property (R3) if it is stable
under the total 𝜇-operator.
Exercise 4.3.1. Prove that the Ackermann function is recursive. (See
also Proposition 4.5.4.)
4.4. Turing Computable Functions

Turing machines provide a way to make precise the notion of a function
which is ‘computable by a machine’.
Definition. A Turing machine ℳ is given by the following data:
• A finite number (≥ 1) of tapes 𝐵1 , 𝐵2 , … placed horizontally,
each bounded on the left and unbounded on the right; each
tape is divided into squares which are numbered by ℕ∗ from
left to right. The tapes are arranged in such a way that the
squares with the same number lie on the same vertical line.
• A read-write head which may read, erase and write symbols
on the tape squares (one symbol per square). The head moves
horizontally and is placed, at any moment, over one vertical
column, and it manipulates all squares of this column simul-
taneously.
The set of symbols is given by 𝑆 = {$, ∣, 𝑏} (𝑏 is called the
blank symbol).
Here are the data which are specific to ℳ:
• 𝑛 = 𝑛(ℳ) ∈ ℕ∗ (the number of tapes).
• A finite set 𝑄 of states. There are two (distinct) distinguished
states in 𝑄: 𝑞𝑖 (the initial state) and 𝑞𝑓 (the final state).
• A function 𝑀 ∶ 𝑆 𝑛 × 𝑄 → 𝑆 𝑛 × 𝑄 × {−1, 0, 1}, the transition
function of ℳ.
A description of the way the machine ℳ operates:
4.4. Turing Computable Functions 99
• At every instant (point in time) 𝑡 ∈ ℕ, ℳ is in one of the states

from 𝑄.
• ℳ operates at every instant by changing states, erasing and
writing symbols on the tapes and moving its head.
The way ℳ operates is subject to the following rules:
(1) At the instant 𝑡 = 0, ℳ is in the initial state, and its head is
placed over the squares number 1.
(2) At every instant 𝑡, ℳ reads the symbols (𝑠1 , … , 𝑠𝑛 ) ∈ 𝑆 𝑛 writ-
ten in front of its head, and the transition function 𝑀 describes
what ℳ is supposed to do. If ℳ is in state 𝑞 and reads (𝑠1 , … , 𝑠𝑛 ),
and if 𝑀(𝑠, 𝑞) = (𝑠′ , 𝑞′ , 𝜀), then the head erases 𝑠, writes 𝑠′ ,
moves by 𝜀 horizontally, and ℳ changes its state to 𝑞′ , then
passes to the instant 𝑡 + 1.
(3) ℳ stops to operate if it attains the final state.
Correct input:
• At the instant 𝑡 = 0, $ is written in the tape squares number 1,
and only there.
• On each square a symbol from 𝑆 is written, and only finitely
many symbols are non-blank.
(These conditions then remain satisfied during the whole computa-
tion).
Constraints:
• For any 𝑠 ∈ 𝑆 𝑛 , one has 𝑀(𝑠, 𝑒𝑓 ) = (𝑠, 𝑒𝑓 , 0).
• The head cannot erase the symbol $ (which marks the begin-
ning of the tape) or overwrite a symbol different from $ with
$:
– For any 𝑞 ∈ 𝑄, 𝑀(($, … , $), 𝑞) = (($, … , $), 𝑞′ , 𝜀) for some
𝜀 ∈ {0, 1};
– if 𝑠 ≠ ($, … , $), then 𝑀(𝑠, 𝑞) = (𝑠′ , 𝑞′ , 𝜀) with 𝑠′𝑖 ≠ $ for all
𝑖.
Remark. A Turing machine is deterministic, that is, every step in the
computation follows in a unique way from the preceding one, and one
may thus predict for every instant 𝑡 the position of the head, the state of
the machine and the content of the tape squares.
Definition.
(1) A tape represents (at a given instant) the natural number 𝑚 if
what is written on it equals ($, ∣,
⏟ … , ∣ , 𝑏, … , 𝑏, …).
𝑚 times
(2) A Turing machine ℳ computes 𝑓 ∈ ℱ𝑝∗ if 𝑛(ℳ) ≥ 𝑝 + 1 and if
for all 𝑚 ∈ ℕ𝑝 , when ℳ starts operating with the input where
for 𝑖 = 1, … , 𝑝, the tape 𝐵𝑖 represents 𝑚𝑖 and, for 𝑖 > 𝑝, the tape
𝐵𝑖 represents 0, then
• if 𝑚 ∈ dom(𝑓), then ℳ stops after some finite time,
and its tapes successively represent the natural numbers
(𝑚1 , … , 𝑚𝑝 , 𝑓(𝑚), 0, … , 0) (in this order);
• if 𝑚 ∉ dom(𝑓), ℳ either never stops, that is, 𝑞𝑓 is
never attained, or ℳ stops after some finite time, but
when it stops there is no 𝑛 ∈ ℕ such that the tapes of
the machine ℳ successively represent the natural num-
bers (𝑚1 , … , 𝑚𝑝 , 𝑛, 0, … , 0).
(3) The partial function 𝑓 is Turing computable if there is a Turing
machine ℳ which computes 𝑓.
Remarks.
(1) There are many variants of the model of a Turing machine (one
may for example work with tapes which are unbounded on the
left and on the right), all leading to the same notion of Turing
computability.
(2) Using unary code to represent natural numbers is of course
very inefficient, but it suffices for our purposes, since we are
not concerned with practical feasibility or complexity issues in
this book.
Lemma 4.4.1. The basic functions (𝑆, 𝐶00 and the projections 𝑃𝑖𝑛 ) are Tur-
ing computable.
Proof. 𝐶00 is computed by a Turing machine ℳ with one tape, a set of

states 𝑄 = {𝑞𝑖 , 𝑞𝑓 }, and with a transition function 𝑀($, 𝑞𝑖 ) = ($, 𝑞𝑓 , 0).
(Here, and in what follows, we only give the relevant part of the transi-
tion function 𝑀.)
𝑃𝑖𝑛 is computed by ℳ with 𝑛 + 1 tapes, a set of states 𝑄 = {𝑞𝑖 , 𝑞𝑓 },
and with a transition function 𝑀($, … , $, 𝑞𝑖 ) = ($, … , $, 𝑞𝑖 , +1) and
(𝑠1 , … , 𝑠𝑛 , ∣, 𝑒𝑖 , +1) if 𝑠𝑖 =∣,

𝑀(𝑠1 , … , 𝑠𝑛 , 𝑏, 𝑞𝑖 ) = {
(𝑠1 , … , 𝑠𝑛 , 𝑏, 𝑞𝑓 , 0) if 𝑠𝑖 = 𝑏.
The successor function 𝑆 is computed by a Turing machine ℳ with

two tapes, a set of states 𝑄 = {𝑞𝑖 , 𝑞𝑓 }, and with a transition function
𝑀($, $, 𝑞𝑖 ) = ($, $, 𝑞𝑖 , +1), 𝑀(∣, 𝑏, 𝑞𝑖 ) = (∣, ∣, 𝑞𝑖 , +1) and finally 𝑀(𝑏, 𝑏, 𝑞𝑖 )
= (𝑏, ∣, 𝑞𝑓 , 0). □
Lemma 4.4.2. The set of partial Turing computable functions is stable

under composition.
Proof. Let 𝑓1 , … , 𝑓𝑛 ∈ ℱ𝑝∗ and 𝑔 ∈ ℱ𝑛∗ be Turing computable, and let

ℎ = 𝑔(𝑓1 , … , 𝑓𝑛 ).
By assumption there exist ℳ𝑖 , 1 ≤ 𝑖 ≤ 𝑛, where ℳ𝑖 is a Turing
machine with 𝑝𝑖 ≥ 𝑝 + 1 tapes and a set of states 𝑄𝑖 which computes 𝑓𝑖 ,
as well as a Turing machine ℳ ′ with 𝑛′ ≥ 𝑛 + 1 tapes and a set of states
𝑄′ which computes 𝑔.
𝑛
Let ℳ be a Turing machine with 𝑝+(𝑛′ −𝑛)+∑𝑖=1 (𝑝𝑖 −𝑝) tapes and
̇ ′ ∪̇ ⋃̇ 𝑖 𝑄𝑖 . The initial state of ℳ corresponds
a set of states 𝑄 = {𝑞𝑑 , 𝑞𝑐 }∪𝑄
to that of ℳ1 , and its final state to that of ℳ ′ .
The machine ℳ operates as follows (the precise description of the
transitifion function of ℳ is left as an exercise):
If (𝑚1 , … , 𝑚𝑝 , 0, … , 0) are represented on the tapes, ℳ starts to com-
pute 𝑓1 (𝑚1 , … , 𝑚𝑝 ) as ℳ1 would do, using the states in 𝑄1 ⧵ {𝑞1𝑓 } and
(𝑝1 − 𝑝) auxiliary tapes (all different from 𝐵𝑝+1 ). In state 𝑞1𝑓 , it returns
to the beginning of the tapes, then passes to state 𝑞2𝑖 ∈ 𝑄2 and computes
𝑓2 (𝑚), using the states in 𝑄2 ⧵ {𝑞2𝑓 } and (𝑝2 − 𝑝) auxiliary tapes (all dif-
ferent from 𝐵𝑝+1 and from the previous auxiliary tapes). In this way, ℳ
successively computes 𝑓1 (𝑚), … , 𝑓𝑛 (𝑚). Once 𝑓𝑛 (𝑚) has been computed,
ℳ passes to the initial state of ℳ ′ and computes ℎ(𝑓1 (𝑚), … , 𝑓𝑛 (𝑚)) as
ℳ ′ would do, using 𝑄′ and 𝑛′ −𝑛 auxiliary tapes (distinct from 𝐵𝑝+1 ). For
this computation, it uses the tapes on which are represented the 𝑓𝑖 (𝑚)
as input tapes and 𝐵𝑝+1 as an output tape. Renumbering the tapes, the
transition function of ℳ may thus be established using the transition
functions of ℳ1 ,. . . ,ℳ𝑛 and ℳ ′ .
Once ℎ(𝑓1 (𝑚), … , 𝑓𝑛 (𝑚)) is computed, instead of passing to the final
state of ℳ ′ , ℳ passes to state 𝑞𝑑 which serves to move its head to the
beginning of the tapes. Then, in state 𝑞𝑐 , it ‘clears’ the tapes on which
are represented the 𝑓𝑖 (𝑚), and finally it passes to the final state. □
under the 𝜇-operator.
Proof. The easy proof is left as an exercise. □

under recursion.
Proof. Suppose that 𝑔 ∈ ℱ𝑝∗ is computed by the Turing machine ℳ with

∗
𝑝 + 1 + 𝑘 tapes and a set of states 𝑄, and that ℎ ∈ ℱ𝑝+2 is computed by
′ ′ ′ ∗
ℳ with 𝑝+3+𝑘 tapes and a set of states 𝑄 . Let 𝑓 ∈ ℱ𝑝+1 be the partial
function defined by recursion from 𝑔 and ℎ, that is, 𝑓(𝑥, 0) = 𝑔(𝑥) and
𝑓(𝑥, 𝑦 + 1) = ℎ(𝑥, 𝑦, 𝑓(𝑥, 𝑦)).
We now give an informal description of a Turing machine 𝒩 which
computes 𝑓: 𝒩 has 𝑝 + 4 + 𝑘 + 𝑘′ tapes, and its set of states is given by
̇ ′ plus a couple of auxiliary states, with 𝑞𝒩 ℳ 𝒩 ℳ′
𝑄∪𝑄 𝑖 = 𝑞𝑖 and 𝑞𝑓 = 𝑞𝑓 .
On input (𝑚1 , … , 𝑚𝑝+1 ), 𝒩 operates as follows:
(1) Compute 𝑔(𝑚1 , … , 𝑚𝑝 ) with input tapes 𝐵1 , … , 𝐵𝑝 , an output
tape 𝐵𝑝+2 and auxiliary tapes 𝐵𝑝+5 , … , , 𝐵𝑝+4+𝑘 , operating as ℳ
would do, up to renumbering the tapes.
(2) Compare 𝑚𝑝+1 and the natural number 𝑦 which is represented
on 𝐵𝑝+3 (note that during the whole computation, 𝐵𝑝+3 will
represent a natural number ≤ 𝑚𝑝+1 ):
• if 𝑚𝑝+1 = 𝑦, then go to step (5);
• if 𝑚𝑝+1 > 𝑦, then go to step (3).
(3) Compute ℎ(𝑚1 , … , 𝑚𝑝 , 𝑦, 𝑓(𝑥, 𝑦)) = 𝑓(𝑥, 𝑦 + 1) as ℳ ′ would
do, with input tapes 𝐵1 , … , 𝐵𝑝 , 𝐵𝑝+3 , 𝐵𝑝+2 , an output tape 𝐵𝑝+4
and the last 𝑘′ tapes as auxiliary tapes.
(4) Copy the content of 𝐵𝑝+4 onto tape 𝐵𝑝+2 , then clear tape 𝐵𝑝+4
and increment the content of 𝐵𝑝+3 by one, that is, pass from 𝑦
to 𝑦 + 1. Then go back to step (2).
(5) Clear tape 𝐵𝑝+3 and stop. □
Theorem 4.4.5. Every partial recursive function is Turing computable.
Proof. It is enough to combine the four preceding lemmas. □
In order to establish the converse of Theorem 4.4.5, we ought to code

Turing machines and the way they operate.
Coding of Turing machines. We identify the set 𝑆 = {𝑏, $, ∣} with

{0, 1, 2}, via 0 ↔ 𝑏, 1 ↔ $ and 2 ↔∣. We code a sequence (𝑠𝑖 )𝑖≤𝑛 of sym-
bols in 𝑆, or a sequence (𝑠𝑖 )𝑖∈ℕ with 𝑠𝑖 = 0 for almost all 𝑖, by Γ((𝑠𝑖 )) =
∑𝑖≥0 𝑠𝑖 3𝑖 ∈ ℕ.
Let ℳ be a Turing machine. Then ℳ is given by:
• 𝑛 = 𝑛(ℳ) ≥ 1, the number of tapes of ℳ.
• The set of states 𝑄. We will suppose that 𝑄 = {0, … , 𝑚} (thus,
𝑚 ≥ 1 equals card(𝑄) − 1), with 𝑞𝑖 = 0 and 𝑞𝑓 = 1.
• The transition function 𝑀 ∶ 𝑆 𝑛 × 𝑄 → 𝑆 𝑛 × 𝑄 × {−1, 0, 1}.
To code the transition function 𝑀, if 𝜌 = (𝑠1 , … , 𝑠𝑛 , 𝑞) ∈ 𝑆 𝑛 × 𝑄 and
𝑀(𝜌) = (𝑡1 , … , 𝑡𝑛 , 𝑞′ , 𝜀), we set
𝑟1 (𝜌) = 𝛼2 (Γ(𝑠1 , … , 𝑠𝑛 ), 𝑞)
and
𝑟2 (𝜌) = 𝛼3 (Γ(𝑡1 , … , 𝑡𝑛 ), 𝑞′ , 𝜀 + 1),
and then
⌜𝑀⌝ = ∏ 𝜋(𝑟1 (𝜌))𝑟2 (𝜌) .
𝜌∈𝑆 𝑛 ×𝑄
To decode, we will use the following function 𝛿 ∈ ℱ2 , defined by 𝛿(𝑖, 𝑥) ∶=

(𝜇𝑧 ≤ 𝑥) (𝜋(𝑖)𝑧+1 ∤ 𝑥). We then have
𝛿(𝛼2 (Γ(𝑠1 , … , 𝑠𝑛 ), 𝑞), ⌜𝑀⌝) = 𝛼3 (Γ(𝑡1 , … , 𝑡𝑛 ), 𝑞′ , 𝜀 + 1).
Finally, we define the index of ℳ as ⌜ℳ⌝ = 𝛼3 (𝑛, 𝑚, ⌜𝑀⌝).
Lemma 4.4.6. For any 𝑝 ≥ 0, the set

𝐼𝑝 = {⌜ℳ⌝ ∣ ℳ is a Turing machine with at least 𝑝 + 1 tapes}
Proof. It is easy to see that one may recognize in a primitive recursive

way if the constraints on the transition function are met. The details are
left as an exercise. □
A configuration 𝐶 = 𝐶(𝑡) of ℳ (at an instant 𝑡) is a sequence (𝑠𝑖 )𝑖 ∈

𝑆 ℕ such that 𝑠𝑛𝑣+𝑤 equals the symbol which is written on the (𝑣 + 1)th
square of tape 𝐵𝑤+1 . (We suppose that there is only a finite number of
non-blank symbols and that $ is written at the beginning of the tapes
and only there.)
We code the configuration 𝐶 by Γ(𝐶) ∶= Γ((𝑠𝑖 )). To decode, we will
use the function
𝜂(Γ(𝐶), 𝑢, 𝑣, 𝑛) = 𝑟 (𝑞(Γ(𝐶), 3𝑛(ᵆ−1)+(𝑣
̇ −1)
̇
), 3)
(here, 𝑞 denotes the result of the division with remainder, and 𝑟 denotes
the remainder), which gives the symbol 𝑠 which is written on the 𝑢th
square of tape 𝐵𝑣 . If 𝜎 = (𝑠1 , … , 𝑠𝑛 ) is the sequence of symbols written
on the squares of number 𝑢, then Γ(𝜎) = 𝜀(Γ(𝐶), 𝑢, 𝑛), where 𝜀(𝑥, 𝑦, 𝑧) =
𝑟 (𝑞(𝑥, 3𝑧(𝑦−1)
̇
), 3𝑧 ).
The situation Sit = Sit(𝑡) of ℳ (at an instant 𝑡) is given by (𝑞, 𝑘, 𝐶(𝑡)),
where 𝑞 denotes the state of ℳ at the instant 𝑡, 𝑘 the number of squares in
front of which is placed the head at the instant 𝑡 and 𝐶(𝑡) the configura-
tion at the instant 𝑡. We code the situation via Γ(Sit(𝑡)) = 𝛼3 (𝑞, 𝑘, Γ(𝐶(𝑡))).
Lemma 4.4.7. Let 𝑝 ≥ 0. There exists a primitive recursive function 𝑔𝑝 ∈
ℱ2 such that the following properties hold:
• 𝑔𝑝 (𝑖, 𝑥) = 0 if 𝑖 ∉ 𝐼𝑝 ;
• if 𝑖 = ⌜ℳ⌝ and 𝑥 is the code of the situation of ℳ at the instant 𝑡,
then 𝑔𝑝 (𝑖, 𝑥) is the code of the situation of ℳ at the instant 𝑡 + 1.
Proof. We suppose that 𝑖 = ⌜ℳ⌝ ∈ 𝐼𝑝 . Then

• the current state of ℳ is 𝛽13 (𝑥) = 𝑞,
• the number of squares which are currently observed is 𝛽23 (𝑥) =

𝑘,
• the code of the current configuration 𝐶(𝑡) of the machine ℳ is
𝛽33 (𝑥) = Γ(𝐶(𝑡)),
• the number of tapes of ℳ is given by 𝛽13 (𝑖) = 𝑛,
• the number of states of ℳ minus 1 is given by 𝛽23 (𝑖) = 𝑚,
• the code of the transition function is given by 𝛽33 (𝑖) = ⌜𝑀⌝,
• the code of the sequence of symbols read by the head at the
instant 𝑡 is given by
𝜀 (Γ(𝐶(𝑡)), 𝑘, 𝑛) = 𝜀 (𝛽33 (𝑥), 𝛽23 (𝑥), 𝛽13 (𝑖)) =∶ 𝑐.
If 𝑥 is not the code of a situation, we set 𝑔𝑝 (𝑖, 𝑥) = 0. This is the case
if > 𝑚 or 𝛽23 (𝑥) = 0 or if 𝛽33 (𝑥) is not the code of a configuration
𝛽13 (𝑥)
with $ at the beginning of the tapes and only there (the last condition
may be expressed in a primitive recursive way, using the functions 𝜀 and
𝜂).
Otherwise, let 𝛿 ∶= 𝛿 (𝛼2 (𝑐, 𝑞), ⌜𝑀⌝). Then 𝑐′ = 𝛽13 (𝛿) is the code
of the sequence of symbols written in the place where was previously
written the sequence coded by 𝑐. We may set 𝑔𝑝 (𝑖, 𝑥) = 𝛼3 (𝑞′ , 𝑘′ , Γ(𝐶 ′ )),
where
Γ(𝐶 ′ ) = (Γ(𝐶) + 3𝑛(𝑘−1)
̇
⋅ 𝑐′ ) −̇ (3𝑛(𝑘−1)
̇
⋅ 𝑐)
is the code of the configuration of the machine ℳ at the instant 𝑡 + 1,
𝑞′ = 𝛽23 (𝛿) is the state of ℳ, and where 𝑘′ = (𝛽23 (𝑥) + 𝛽33 (𝛿)) −1
̇ is the
number of squares which are observed by the head of ℳ at the instant
𝑡 + 1. □
𝑝
One defines a function ST ∈ ℱ𝑝+2 as follows:
𝑝
• ST (𝑖, 𝑡, 𝑥) = 0, if 𝑖 ∉ 𝐼𝑝 ;
𝑝
• otherwise, ST (𝑖, 𝑡, 𝑥) is the code Γ(Sit(𝑡)) of the situation at
the instant 𝑡 of the machine ℳ of index 𝑖 which has started to
operate at the instant 𝑡 = 0 with the following configuration:
on 𝐵1 , … , 𝐵𝑝 are represented the natural numbers 𝑥1 , … , 𝑥𝑝 , and
the other tapes represent 0.
𝑝
Theorem 4.4.8. The function ST is primitive recursive for any 𝑝 ≥ 0.
𝑝
Proof. The function 𝜆𝑖𝑥.ST (𝑖, 0, 𝑥) is primitive recursive. (This is easy
to see and is left as an exercise.)
𝑝 𝑝
Moreover, one has ST (𝑖, 𝑡 + 1, 𝑥) = 𝑔𝑝 (𝑖, ST (𝑖, 𝑡, 𝑥)), so the result
follows by Lemma 4.4.7. □
An output configuration for 𝑥 = (𝑥1 , … , 𝑥𝑝 ) ∈ ℕ𝑝 is a configuration

where the tapes 𝐵1 , … , 𝐵𝑝 represent 𝑥1 , … , 𝑥𝑝 , the tape 𝐵𝑝+1 represents a
natural number and where the other tapes represent 0.
For 𝑝 ≥ 0, we define a set 𝐸 𝑝 ⊆ ℕ𝑝+2 as follows:
𝑖 ∈ 𝐼𝑝 and 𝑐 is the code of an
(𝑖, 𝑐, 𝑥) ∈ 𝐸 𝑝 ∶ ⟺ {
output configuration for 𝑥.
It is easy to see that 𝐸 𝑝 is a primitive recursive set. (In order to ex-

press for example that 𝐵𝑝+1 represents a natural number, observe that
one may bound the universal quantifier in the formula ∀𝑧(𝜂(𝑐, 𝑧, 𝑝+1, 𝑛)
= 2 ∧ 𝑧 ≥ 3 → 𝜂(𝑐, 𝑧−1, ̇ 𝑝 + 1, 𝑛) = 2) by 𝑐. The details are left as an
exercise.)
We now define sets 𝐵 𝑝 ⊆ ℕ𝑝+2 and 𝐶 𝑝 ⊆ ℕ𝑝+3 as follows:
𝑝
𝛽 3 (ST (𝑖, 𝑡, 𝑥)) = 1 and
(𝑖, 𝑡, 𝑥) ∈ 𝐵 𝑝 ∶ ⟺ { 1 3 𝑝
(𝑖, 𝛽3 (ST (𝑖, 𝑡, 𝑥)), 𝑥) ∈ 𝐸 𝑝 .
(This expresses that at the instant 𝑡, the machine of index 𝑖 is in the final
state, and its configuration is an output configuration for 𝑥.)
(𝑖, 𝑡, 𝑥) ∈ 𝐵 𝑝 and
(𝑖, 𝑦, 𝑡, 𝑥) ∈ 𝐶 𝑝 ∶ ⟺ {
𝑦 is represented on 𝐵𝑝+1 .
The sets 𝐵𝑝 and 𝐶 𝑝 are primitive recursive.

Now let 𝑓 ∈ ℱ𝑝∗ be a partial Turing computable function. Choose
a Turing machine ℳ which computes 𝑓, and set 𝑖 = ⌜ℳ⌝. One may
define the partial function 𝑇ℳ (which gives the computing time) as
𝑇ℳ (𝑥) ∶= 𝜇𝑡 ((𝑖, 𝑡, 𝑥) ∈ 𝐵 𝑝 ) .
One then gets
(∗) 𝑓(𝑥) = (𝜇𝑦 ≤ 𝑇ℳ (𝑥))((𝑖, 𝑦, 𝑇ℳ (𝑥), 𝑥) ∈ 𝐶 𝑝 ).
The description (∗) is called the Kleene normal form of 𝑓.
4.5. Universal Functions 107
In particular, we have proved the following two results.
Theorem 4.4.9. Every partial Turing computable function is recursive.

□
Proposition 4.4.10.
(1) If 𝑓 is a total Turing computable function which may be com-
puted in a primitive recursive time, then 𝑓 is primitive recursive.
(2) The set of partial recursive functions is the smallest subset of ℱ ∗
which contains the primitive recursive functions and which is sta-
ble under composition and under application of the 𝜇-operator.
(3) The set of total recursive functions is the smallest subset of ℱ
which contains the primitive recursive functions and which is
stable under composition and under application of the total 𝜇-
operator. In other words, every total recursive function may be
obtained by a finite number of applications of the rules (R0)-
(R3). □
Let us observe that our arguments to prove that every Turing com-
putable function is recursive are not specific to Turing machines. This
justifies the following thesis.
Church’s Thesis. Every function which is computable (in the intuitive

sense) is recursive.
4.5. Universal Functions

It suffices to vary the index of the Turing machine in order to obtain a
∗
universal function. Let us first define a function 𝑇 𝑝 ∈ ℱ𝑝+1 as follows:
• if 𝑖 ∉ 𝐼𝑝 , then 𝑇 𝑝 (𝑖, 𝑥) is not defined;
• if 𝑖 ∈ 𝐼𝑝 , then 𝑇 𝑝 (𝑖, 𝑥) = 𝜇𝑡 [(𝑖, 𝑡, 𝑥) ∈ 𝐵 𝑝 ].
∗
We may now define a partial recursive function 𝜑𝑝 ∈ ℱ𝑝+1 , via
𝜑𝑝 (𝑖, 𝑥) = 𝜇𝑦 [(𝑖, 𝑦, 𝑇 𝑝 (𝑖, 𝑥), 𝑥) ∈ 𝐶 𝑝 ] .

𝑝
We set 𝜑𝑖 = 𝜆𝑥.𝜑𝑝 (𝑖, 𝑥).
Proposition 4.5.1. 𝜑𝑝 is a universal partial recursive function, that is,

𝑝
every partial recursive function in 𝑝 variables is of the form 𝜑𝑖 for a suitable
natural number 𝑖. □
Theorem 4.5.2 (s-m-n Theorem). For any pair of natural numbers 𝑚
and 𝑛 there is a primitive recursive function 𝑠𝑚
𝑛 ∈ ℱ𝑛+1 such that for all
𝑖 ∈ ℕ, 𝑥 ∈ ℕ𝑛 and 𝑦 ∈ ℕ𝑚 one has
𝜑𝑖𝑛+𝑚 (𝑥, 𝑦) = 𝜑𝑠𝑚𝑚 (𝑖,𝑥) (𝑦).
𝑛
Proof. Let ℳ be a Turing machine with at least 𝑚 + 𝑛 + 1 tapes. Given

(𝑎1 , … , 𝑎𝑛 ) ∈ ℕ𝑛 , consider the Turing machine ℳ ′ which operates as
follows:
(1) It writes 𝑎1 , … , 𝑎𝑛 on the tapes 𝐵𝑚+2 , … , 𝐵𝑚+𝑛+1 .
(2) It then operates as ℳ, up to permutation of the tapes, and it
writes the result on the tape 𝐵𝑚+1 .
(3) It then clears the tapes 𝐵𝑚+2 , … , 𝐵𝑚+𝑛+1 , and finally it stops.
It is clear that there is a primitive recursive function 𝑔 ∈ ℱ𝑛+1
which computes the code ⌜ℳ ′ ⌝ ∈ 𝐼𝑚 from ⌜ℳ⌝ and 𝑎1 , … , 𝑎𝑛 , that is,
𝑔(⌜ℳ⌝, 𝑎1 , … , 𝑎𝑛 ) = ⌜ℳ ′ ⌝ and 𝑔(𝑖, 𝑎) = 0 if 𝑖 ∉ 𝐼𝑚+𝑛 . We may thus set
𝑠𝑚
𝑛 = 𝑔. □
Theorem 4.5.3 (Kleene’s Fixed Point Theorem). Suppose 𝑚 ∈ ℕ∗ and
𝛼 ∈ ℱ1 is a total recursive function. Then there exists 𝑖 ∈ ℕ such that
𝑚
𝜑𝑖𝑚 = 𝜑𝛼(𝑖) .
Proof. Consider 𝑔 = 𝜆𝑦𝑥1 … 𝑥𝑚 .𝜑𝑚 (𝛼(𝑠𝑚 1 (𝑦, 𝑦)), 𝑥1 , … , 𝑥𝑚 ). By univer-

sality (Proposition 4.5.1), one has 𝑔 = 𝜑𝑎𝑚+1 for some index 𝑎 ∈ ℕ. One
gets
𝑚
𝜑𝛼(𝑠 𝑚+1
𝑚 (𝑦,𝑦)) (𝑥) = 𝜑𝑎 (𝑦, 𝑥) = 𝜑𝑠𝑚𝑚 (𝑎,𝑦) (𝑥).
1 1
Setting 𝑖 ∶= 𝑠𝑚 𝑚 𝑚
1 (𝑎, 𝑎), one obtains 𝜑𝛼(𝑖) = 𝜑𝑖 . □
The Fixed Point Theorem provides an elegant argument to establish

the following result (cf. Exercise 4.3.1).
Proposition 4.5.4. The Ackermann function 𝜉 is recursive.
Proof. Define a partial recursive function 𝜃 ∈ ℱ3∗ as follows:
4.6. Recursively Enumerable Sets 109
• 𝜃(𝑖, 𝑦, 𝑥) = 2𝑥 if 𝑦 = 0;
• 𝜃(𝑖, 𝑦, 𝑥) = 1 if 𝑥 = 0;
• 𝜃(𝑖, 𝑦, 𝑥) = 𝜑2 (𝑖, 𝑦−1,
̇ 𝜑2 (𝑖, 𝑦, 𝑥−1))
̇ otherwise.
By universality, there exists 𝑎 ∈ ℕ such that 𝜃 = 𝜑𝑎3 . By the s-m-n
Theorem, this yields
𝜃(𝑖, 𝑦, 𝑥) = 𝜑𝑠22 (𝑎,𝑖) (𝑦, 𝑥).
1
Set 𝛼 = 𝜆𝑖.𝑠21 (𝑎, 𝑖).

By Kleene’s Fixed Point Theorem (Theorem 4.5.3)
there exists 𝑖0 ∈ ℕ such that 𝜑𝑖20 = 𝜑𝛼(𝑖
2
0)
. The function 𝜑𝑖20 is recur-
sive and satisfies the relations which define the Ackermann function.
Indeed, for 𝑦, 𝑥 ∈ ℕ∗ , one has
𝜑𝑖20 (𝑦, 𝑥) = 𝜃(𝑖0 , 𝑦, 𝑥)
= 𝜑2 (𝑖0 , 𝑦 − 1, 𝜑2 (𝑖0 , 𝑦, 𝑥 − 1)) = 𝜑𝑖20 (𝑦 − 1, 𝜑𝑖20 (𝑦, 𝑥 − 1)).
Finally, note that the totality of the function 𝜑𝑖20 follows by induction as
well. □
4.6. Recursively Enumerable Sets

Definition. A set of natural numbers 𝑋 ⊆ ℕ𝑛 is called recursively enu-
merable if there is a recursive set 𝑌 ⊆ ℕ𝑛+1 such that 𝑋 = 𝜋(𝑌 ), where
𝜋 denotes the projection on the first 𝑛 coordinates.
Remark. Every recursive set is recursively enumerable. □
Proposition 4.6.1. The set of recursively enumerable sets is stable un-
der intersection, union, projection and bounded universal quantification.
Moreover, if 𝑓1 , … , 𝑓𝑛 ∈ ℱ𝑚 are recursive and 𝑋 ⊆ ℕ𝑛 is recursively enu-
merable, then {𝑥 ∈ ℕ𝑚 ∣ (𝑓1 (𝑥), … , 𝑓𝑚 (𝑥)) ∈ 𝑋} is recursively enumer-
able.
Proof. We say that an 𝑛-ary relation 𝑅(𝑥1 , … , 𝑥𝑛 ) on ℕ is recursively

enumerable if there exists an 𝑛 + 1-ary recursive relation 𝑅˜(𝑥, 𝑦) on ℕ
such that 𝑅(𝑥) ⟺ ∃𝑦𝑅 ˜(𝑥, 𝑦) holds in ℕ. In what follows, all equiva-
lences are meant to be in ℕ.
Now assume that 𝑅(𝑥) and 𝑆(𝑥) are 𝑛-ary recursively enumerable re-
˜(𝑥, 𝑦) and 𝑆˜(𝑥, 𝑦) such that 𝑅(𝑥) ⟺
lations. Choose recursive relations 𝑅
∃𝑦𝑅˜(𝑥, 𝑦) and 𝑆(𝑥) ⟺ ∃𝑦𝑆˜(𝑥, 𝑦). We then have the following equiva-
lences:
˜(𝑥, 𝑦) ∨ 𝑆˜(𝑥, 𝑦)) .

𝑅(𝑥) ∨ 𝑆(𝑥) ⟺ ∃𝑦 (𝑅
˜(𝑥, 𝛽12 (𝑦)) ∧ 𝑆˜(𝑥, 𝛽22 (𝑦))) .
𝑅(𝑥) ∧ 𝑆(𝑥) ⟺ ∃𝑦 (𝑅
˜(𝑦, 𝑧, 𝑢) ⟺ ∃𝑠𝑅
∃𝑧𝑅(𝑦, 𝑧) ⟺ ∃𝑧∃𝑢𝑅 ˜(𝑦, 𝛽12 (𝑠), 𝛽22 (𝑠)).
Moreover, (∀𝑧 ≤ 𝑤) 𝑅(𝑥, 𝑧) is equivalent to (∀𝑧 ≤ 𝑤) ∃𝑦 𝑅 ˜(𝑥, 𝑧, 𝑦), which

˜(𝑥, 𝑧, (𝑠)𝑧 ), where (𝑠)𝑧 denotes the
in turn is equivalent to ∃𝑠 (∀𝑧 ≤ 𝑤) 𝑅
component function from Lemma 4.1.4.
Finally, suppose that 𝑓1 , … , 𝑓𝑛 ∈ ℱ𝑚 are recursive functions. Then
˜(𝑓1 (𝑥), … , 𝑓𝑛 (𝑥), 𝑦), which proves
𝑅(𝑓1 (𝑥), … , 𝑓𝑛 (𝑥)) is equivalent to ∃𝑦𝑅
the moreover part. □
Theorem 4.6.2. For 𝑋 ⊆ ℕ𝑛 , the following are equivalent:
(1) 𝑋 is recursively enumerable.

(2) 𝑋 is empty or it is the image of a function 𝑓 = (𝑓1 , … , 𝑓𝑛 ) ∶ ℕ →
ℕ𝑛 , where 𝑓1 , … , 𝑓𝑛 ∈ ℱ1 are (total) recursive.
(3) 𝑋 = dom(𝑔) for a partial recursive function 𝑔 ∈ ℱ𝑛∗ .
(4) There exists a primitive recursive set 𝑌 ⊆ ℕ𝑛+1 such that 𝑋 =
𝜋(𝑌 ).
Proof. We may assume 𝑛 = 1. Indeed, we may identify ℕ𝑛 and ℕ using

the bijection 𝛼𝑛 , as 𝛼𝑛 as well as the component functions 𝛽𝑖𝑛 of 𝛼−1
𝑛 are
(1)⇒(2): Let 𝑟 ∈ 𝑋 and let 𝑌 ⊆ ℕ2 be recursive such that 𝑋 = 𝑃12 (𝑌 ).
Then 𝑋 = im(𝑓), where 𝑓 ∈ ℱ1 is defined as follows:
𝛽2 (𝑧) if (𝛽12 (𝑧), 𝛽22 (𝑧)) ∈ 𝑌 ,

𝑓(𝑧) ∶= { 1
𝑟 else.
(2)⇒(3): If 𝑓 ∈ ℱ1 is a recursive function, we define 𝑔 = 𝜆𝑥.𝜇𝑡(𝑥 =

𝑓(𝑡)). Then 𝑔 is a partial recursive function and dom(𝑔) = im(𝑓). More-
over, ∅ = dom(𝑔) for 𝑔 = 𝜆𝑥.𝜇𝑡(𝑥 > 𝑥).
4.6. Recursively Enumerable Sets 111
(3)⇒(4): Let 𝑔 ∈ ℱ1∗ be a partial recursive function such that 𝑋 =

dom(𝑔). By universality there is 𝑖0 ∈ ℕ with 𝑔 = 𝜑𝑖10 , that is,
𝑋 = {𝑥 ∈ ℕ ∣ ∃𝑡 [(𝑖0 , 𝑡, 𝑥) ∈ 𝐵 1 ]}.
This proves (4), as 𝐵 1 is primitive recursive.
(4)⇒(1): Trivial. □
Remark. Given that (1) and (4) are equivalent, the proof of the impli-
cation (1)⇒(2) actually shows that one may suppose in (2) that the func-
tions 𝑓1 , … , 𝑓𝑛 are primitive recursive.
Corollary 4.6.3. The set 𝑈 = dom(𝜑𝑛 ) ⊆ ℕ𝑛+1 is universal recursively
enumerable in the following sense: 𝑈 is recursively enumerable and every
recursively enumerable set 𝑋 ⊆ ℕ𝑛 is of the form 𝑈𝑒 = {𝑥 ∣ (𝑒, 𝑥) ∈ 𝑈} for
some 𝑒 ∈ ℕ. □
Theorem 4.6.4 (Theorem of the Complement). A set 𝑋 ⊆ ℕ𝑛 is recursive
if and only if both 𝑋 and ℕ𝑛 ⧵ 𝑋 are recursively enumerable.
Proof. “⇒” is clear. To prove “⇐”, assume that 𝑋 ⊆ ℕ𝑛 is such that

there are recursive subsets 𝑌 and 𝑌 ′ of ℕ𝑛+1 with 𝑋 = 𝜋(𝑌 ) and ℕ𝑛 ⧵𝑋 =
𝜋(𝑌 ′ ). Then 𝟙𝑋 (𝑧) = 𝟙𝑌 (𝑧, 𝜇𝑡[(𝑧, 𝑡) ∈ 𝑌 ∪ 𝑌 ′ ]). □
Theorem 4.6.5. The set dom(𝜑1 ) is not recursive. In particular, there
exists a recursively enumerable non-recursive set.
Proof. We will prove that the domain of the function 𝑔 = 𝜆𝑥.𝜑1 (𝑥, 𝑥) is
not recursive. (In other words, one may not decide if a Turing machine
stops when operating on its own code.)
Let 𝐷 = ℕ ⧵ dom(𝑔). If 𝐷 were recursively enumerable, then 𝐷 =
dom(𝜑𝑖10 ) for some 𝑖0 ∈ ℕ by Corollary 4.6.3. In particular, we would get
𝑖0 ∈ 𝐷 if and only if 𝜑1 (𝑖0 , 𝑖0 ) is defined. But this is absurd, since the
definition of 𝑔 implies that 𝑖0 ∈ 𝐷 if and only if 𝜑1 (𝑖0 , 𝑖0 ) is not defined.
□
The Halting Problem HALT is defined as the set

{𝑖 ∈ 𝐼1 ∣ ℳ of index 𝑖 halts when operating on the empty input}.
A slight variation of our argument in the previous proof yields the fol-
lowing.
Theorem 4.6.6 (Undecidability of the Halting Problem). The set HALT

is not recursive.
Proof. Suppose for contradiction that HALT is recursive. It is easy to

see that 𝐼𝑠 = dom (𝜆𝑥.𝜑1 (𝑥, 0)) is then recursive as well, where 𝐼𝑠 is by
definition the set of indices of Turing machines which halt in an out-
put configuration for the input 0 when operating on the empty input.
Indeed, if 𝑖 ∈ HALT, then 𝑖 ∈ 𝐼𝑠 if and only if
1 1
(𝑖, 𝛽33 (ST (𝑖, 𝜇𝑡[𝛽13 (ST (𝑖, 𝑡, 0)) = 1], 0)), 0) ∈ 𝐸 1 .
Now choose a recursively enumerable non-recursive set 𝐴 ⊆ ℕ

(which exists by Theorem 4.6.5). Let ℳ be a Turing machine which com-
putes a function 𝑓 ∈ ℱ2∗ with dom(𝑓) = 𝐴×{0}. Let 𝑖0 be the index of ℳ.
Then 𝐴 = {𝑛 ∈ ℕ ∣ 𝑠11 (𝑖0 , 𝑛) ∈ 𝐼𝑠 }, and so 𝐴 is recursive, contradicting
our assumption. □
Theorem 4.6.7 (Rice’s Theorem). Let 𝒳 be a set of partial recursive func-
tions in one variable. Suppose that 𝒳 is neither empty nor equal to the set of
all partial recursive functions in one variable. Then 𝐼 = {𝑖 ∈ ℕ ∣ 𝜑𝑖1 ∈ 𝒳}
is not recursive.
Proof. Passing to the complement if necessary, we may assume that the

function with empty domain is in 𝒳. Choose 𝑗 ∈ ℕ ⧵ 𝐼, and consider
the function 𝜓(𝑥, 𝑦, 𝑧) = (𝜑1 (𝑗, 𝑧) + 𝜑1 (𝑥, 𝑦)) −𝜑
̇ 1 (𝑥, 𝑦). Letting 𝜓𝑥,𝑦 =
𝜆𝑧.𝜓(𝑥, 𝑦, 𝑧), we get
𝜓𝑥,𝑦 ∈ 𝒳 ⟺ (𝑥, 𝑦) ∉ dom(𝜑1 ) = 𝑈.
We have 𝜓 = 𝜑𝑖30 for a natural number 𝑖0 , so 𝜓𝑥,𝑦 = 𝜑𝑠11 (𝑖 by

2 0 ,𝑥,𝑦)
the s-m-n Theorem. Consider ℎ = 𝜆𝑥𝑦.𝑠12 (𝑖0 , 𝑥, 𝑦), which is a recursive
function. Then
(𝑥, 𝑦) ∈ ℕ2 ⧵ 𝑈 ⟺ ℎ(𝑥, 𝑦) ∈ 𝐼.
As ℕ2 ⧵ 𝑈 is not recursive by Theorem 4.6.4 (and the Theorem of the
Complement), 𝐼 is not recursive either. □
Example 4.6.8.
(1) Let 𝑓 ∈ ℱ1∗ be some partial recursive function. Then the set
{𝑖 ∈ ℕ ∣ 𝜓𝑖1 = 𝑓} is not recursive.
4.7. Elimination of Recursion 113
(2) The set {(𝑖, 𝑗) ∈ ℕ2 ∣ 𝜑𝑖1 = 𝜑𝑗1 } ⊆ ℕ2 is not recursive.

(3) The set of indices of total (recursive) functions is not recursive.
4.7. Elimination of Recursion

The following elementary result is crucial for the next chapter, since it
shows that the recursive functions are definable in the language of arith-
metic ℒ𝑎𝑟 .
Theorem 4.7.1 (Elimination of Recursion). The set of total recursive func-
tions is the smallest subset of ℱ which contains the basic functions 𝐶00 , 𝑆,
𝑃𝑖𝑛 , +, ⋅, 𝟙= and 𝟙< , and which is stable under composition and the total
𝜇-operator.
Let 𝐸 be the smallest subset of ℱ which contains the basic functions

𝐶00 , 𝑆, 𝑃𝑖𝑛 , +, ⋅, 𝟙= and 𝟙< , and which is stable under composition and the
total 𝜇-operator. We will temporarily (until Theorem 4.7.1 is proven) say
that a function is #-recursive if it is in 𝐸, and that a set 𝑋 is #-recursive
if 𝟙𝑋 ∈ 𝐸. We will first prove a lemma.
Lemma 4.7.2 (Gödel’s 𝛽-function).
(1) The function 𝜆𝑥𝑦.𝑥−𝑦
̇ is #-recursive.
(2) The collection of #-recursive sets is stable under boolean combi-
nations and bounded quantification.
(3) The set {(𝑥, 𝑦, 𝑧) ∈ ℕ3 ∣ 𝑥 ≡ 𝑦 mod 𝑧} is #-recursive.
(4) The collection of #-recursive functions is stable under definition
by cases (see Lemma 4.1.2(5)), if both the functions and the sets
of the partition are #-recursive.
(5) There exists a #-recursive function 𝛽 ∈ ℱ3 (called Gödel’s 𝛽-
function) such that for any finite sequence of natural numbers
(𝑐0 , … , 𝑐𝑛−1 ) there are 𝑎, 𝑏 ∈ ℕ with 𝛽(𝑎, 𝑏, 𝑖) = 𝑐𝑖 for 𝑖 = 0, … , 𝑛−
1.
Proof. (1) One has 𝑥−𝑦

̇ = 𝜇𝑧 (𝑥 < (𝑦 + 𝑧) + 1).
(2) One has 𝟙𝑋 𝑐 = 1−𝟙 ̇ 𝑋 and 𝟙𝑋∩𝑌 = 𝟙𝑋 ⋅ 𝟙𝑌 , which establishes
stability under boolean combinations. Now suppose that 𝑋 ⊆ ℕ𝑛+1 is a
#-recursive set. Define a function 𝑓 ∈ ℱ𝑛+1 ,
𝑓(𝑥, 𝑧) ∶= 𝜇𝑦 ((𝑥, 𝑦) ∈ 𝑋 or 𝑦 = 𝑧 + 1) .
As the condition in parentheses may be expressed by a condition of the

form 𝑔(𝑥, 𝑦, 𝑧) = 0 for some #-recursive function 𝑔, it follows that 𝑓 is
#-recursive. Stability under bounded existential quantification follows
from the equivalence (∃𝑦 ≤ 𝑧) (𝑥, 𝑦) ∈ 𝑋 ⟺ 𝑓(𝑥, 𝑧) < 𝑧 + 1. For
∀𝑦 ≤ 𝑧, the same follows by passing to the complement. This finishes
the proof of (2).
(3) One has the following equivalence:
𝑥≡𝑦 mod 𝑧 ⟺ (∃𝑤 ≤ 𝑥 + 𝑦) (𝑥 = 𝑦 + 𝑤 ⋅ 𝑧 ∨ 𝑦 = 𝑥 + 𝑤 ⋅ 𝑧).
(4) Let ℕ𝑛 = 𝐴1 ∪̇ … ∪𝐴 ̇ 𝑘 be a partition with 𝐴𝑖 #-recursive for all

𝑘
𝑖, and let 𝑓1 , … , 𝑓𝑘 ∈ ℱ𝑛 be #-recursive functions. Then ∑𝑖=1 𝟙𝐴𝑖 ⋅ 𝑓𝑖 is
#-recursive.
(5) Set 𝛽(𝑎, 𝑏, 𝑖) ∶= 𝜇𝑧 [𝑧 ≡ 𝑎 mod ((𝑖 + 1)𝑏 + 1)]. By what we have
seen, 𝛽 is #-recursive. Now let 𝑐0 , … , 𝑐𝑛−1 be given, and let 𝑏 ∈ ℕ be such
that 𝑏 is divisible by 𝑛! and 𝑏 > 𝑐𝑖 for all 𝑖. Then 𝑏+1, 2𝑏+1, 3𝑏+1, … , 𝑛𝑏+
1 is a sequence of coprime integers. Indeed, if a prime 𝑝 divides 𝑖𝑏 + 1
and 𝑗𝑏 + 1 for some 1 ≤ 𝑖 < 𝑗 ≤ 𝑛, then 𝑝 ∤ 𝑏 and 𝑝 ∣ 𝑏(𝑖 − 𝑗). Thus
𝑝 ∣ (𝑖 − 𝑗). Since (𝑖 − 𝑗) ∣ 𝑛!, this is absurd.
By the Chinese Remainder Theorem there exists 𝑎 ∈ ℕ such that
for all 𝑖 = 0, … , 𝑛 − 1 one has 𝑎 ≡ 𝑐𝑖 mod ((𝑖 + 1)𝑏 + 1). As 𝑐𝑖 <
(𝑖 + 1)𝑏 + 1 for all 𝑖, 𝑐𝑖 is the smallest natural number 𝑧 such that 𝑧 ≡ 𝑎
mod ((𝑖 + 1)𝑏 + 1). □
Proof of Theorem 4.7.1. By Proposition 4.4.10(3) it is enough to prove

that the collection 𝐸 of #-recursive functions is stable under recursion.
Let 𝑔 ∈ ℱ𝑛 and ℎ ∈ ℱ𝑛+2 be #-recursive functions, and let 𝑓 ∈ ℱ𝑛+1 be
the function obtained by recursion from 𝑔 and ℎ. Consider the set
𝑍 ∶={(𝑥, 𝑦, 𝑎, 𝑏) ∈ ℕ𝑛+3 ∣ 𝛽(𝑎, 𝑏, 0) = 𝑔(𝑥)
and (∀𝑖 ≤ 𝑦) 𝛽(𝑎, 𝑏, 𝑖 + 1) = ℎ(𝑥, 𝑖, 𝛽(𝑎, 𝑏, 𝑖))},
which is #-recursive by Lemma 4.7.2. Moreover, the key property of 𝛽
shows that for any (𝑥, 𝑦) ∈ ℕ𝑛+1 there are 𝑎, 𝑏 ∈ ℕ such that (𝑥, 𝑦, 𝑎, 𝑏) ∈
4.8. Exercises 115
𝑍. Thus, 𝑒 = 𝜆𝑥𝑦.𝜇𝑠 [(∃𝑎 ≤ 𝑠)(∃𝑏 ≤ 𝑠)(𝑥, 𝑦, 𝑎, 𝑏) ∈ 𝑍] is a #-recursive

function. Finally, 𝑓(𝑥, 𝑦) equals
𝜇𝑧 [(∃𝑎 ≤ 𝑒(𝑥, 𝑦))(∃𝑏 ≤ 𝑒(𝑥, 𝑦)) ((𝑥, 𝑦, 𝑎, 𝑏) ∈ 𝑍 and 𝑧 = 𝛽(𝑎, 𝑏, 𝑦))] ,
proving that 𝑓 is #-recursive. □
4.8. Exercises
Exercise 4.8.1. Describe a Turing machine which computes the addi-
tion function 𝜆𝑥𝑦.𝑥 + 𝑦.
Exercise 4.8.2. Let 𝑝 and 𝑞 be prime numbers. Then 𝑞 is called 𝑝-
𝑝𝑛 −1
Mersenne if 𝑞 = for some 𝑛 ∈ ℕ. Prove that the following set is
𝑝−1
primitive recursive:
{𝑁 ∈ ℕ | there is a prime 𝑝 such that 𝑁 is a 𝑝-Mersenne prime}.
Exercise 4.8.3. The Fibonacci function fib ∈ ℱ1 is defined via fib(0) ∶=
0, fib(1) ∶= 1, and fib(𝑛 + 2) ∶= fib(𝑛 + 1) + fib(𝑛) for all 𝑛 ∈ ℕ. Prove
that fib is primitive recursive.
Exercise 4.8.4 (Kalmár elementary functions). The set of elementary
(recursive) functions is the smallest subset 𝐸 of ℱ which satisfies the fol-
lowing properties:
• 𝐸 contains 𝐶00 , the projections 𝑃𝑖𝑛 for all 1 ≤ 𝑖 ≤ 𝑛, the addition,
the multiplication and 𝟙= ;
• if 𝑔 ∈ ℱ𝑘 ∩ 𝐸 and 𝑓1 , … , 𝑓𝑘 ∈ ℱ𝑛 ∩ 𝐸, then 𝑔(𝑓1 , … , 𝑓𝑘 ) ∈ 𝐸;
• if 𝑓 ∈ ℱ𝑛+1 ∩ 𝐸, then the bounded sum
𝑥
(𝑥1 , … , 𝑥𝑛 , 𝑥) ↦ ∑ 𝑓(𝑥1 , … , 𝑥𝑛 , 𝑖)
𝑖=0
and the bounded product

𝑥
(𝑥1 , … , 𝑥𝑛 , 𝑥) ↦ ∏ 𝑓(𝑥1 , … , 𝑥𝑛 , 𝑖)
𝑖=0
are in 𝐸, too.
(1) Prove that for all 𝑛, 𝑘, the constant function 𝐶𝑘𝑛 is elementary.
(2) Call a subset 𝐴 ⊆ ℕ𝑛 elementary if 𝟙𝐴 is elementary. Prove that

{0} ⊆ ℕ is elementary, and that the set of elementary subsets of
ℕ𝑛 is stable under Boolean combinations.
(3) Prove that the exponential function exp = 𝜆𝑥𝑦.𝑥𝑦 is elemen-
tary.
(4) We now define a function 𝑇 ∈ ℱ2 , setting 𝑇(𝑚, 0) = 𝑚 and
𝑇(𝑚, 𝑛 + 1) = exp(2, 𝑇(𝑚, 𝑛)). For a natural number 𝑛, let
𝑇𝑛 = 𝜆𝑥.𝑇(𝑥, 𝑛).
(a) Prove that 𝑇 is primitive recursive.
(b) Prove that 𝑇𝑛 is strictly increasing for all 𝑛 and that, for
fixed 𝑚, the function (𝑚, 𝑛) ↦ 𝑇(𝑚, 𝑛) is strictly increas-
ing in 𝑛.
(c) Prove that for every elementary function 𝑓, there exists
𝑛 ∈ ℕ such that 𝑓 is dominated by 𝑇𝑛 .
(d) Prove that 𝑇 is not elementary.
Exercise 4.8.5.
(1) Let 𝑓 ∈ ℱ1 be an increasing recursive function. Prove that

im(𝑓) is recursive.
(2) Conversely, prove that any infinite recursive subset of ℕ is the
image of a strictly increasing unary recursive function.
(3) Prove that every infinite recursively enumerable set contains
an infinite recursive set.
Exercise 4.8.6 (A primitive recursive bijection whose inverse is not prim-

itive recursive).
(1) Prove that the set of recursive bijections between ℕ and ℕ forms
a group.
(2) Prove that for every Turing machine ℳ which computes a to-
tal function, the graph of the computing time function 𝑇ℳ is
(3) Prove that 𝑓 ∈ ℱ𝑛 is primitive recursive if and only if its graph is
primitive recursive and 𝑓 is bounded from above by a primitive
recursive function.
4.8. Exercises 117
(4) Let 𝑔 ∈ ℱ1 be strictly increasing. Prove that the graph of 𝑔 is

primitive recursive if and only if im(𝑔) is primitive recursive.
(5) Let 𝑓 ∈ ℱ1 be a recursive function which is not primitive re-
cursive, and let ℳ be a Turing machine computing 𝑓.
(a) Let 𝑔0 ∈ ℱ1 be defined by
𝑔0 (𝑥) = sup{𝑇ℳ (𝑦) ∣ 𝑦 ≤ 𝑥} + 2𝑥.
Prove that 𝑔0 is recursive, but not primitive recursive, and
that the graph and the image of 𝑔0 are primitive recursive
sets.
(b) Let 𝑔1 ∈ ℱ1 be a strictly increasing function such that
im(𝑔1 ) = ℕ ⧵ im(𝑔0 ). Consider the function ℎ ∈ ℱ1 given
by ℎ(2𝑥) = 𝑔0 (𝑥), ℎ(2𝑥 + 1) = 𝑔1 (𝑥). Prove that ℎ is bijec-
tive, recursive, but not primitive recursive, and that ℎ−1 is
Exercise 4.8.7 (Existence of recursively inseparable recursively enumer-
able sets).
(1) For a natural number 𝑘 ∈ ℕ, denote by 𝑍𝑘 the set of all 𝑛 ∈
ℕ such that 𝑛 ∈ dom(𝜑𝑛1 ) and 𝜑𝑛1 (𝑛) = 𝑘. Prove that 𝑍𝑘 is
recursively enumerable.
(2) Deduce from this that there exist recursively enumerable sets
𝐴, 𝐵 ⊆ ℕ with 𝐴 ∩ 𝐵 = ∅ such that there is no recursive set
𝐶 ⊆ ℕ with 𝐴 ⊆ 𝐶 and 𝐶 ∩ 𝐵 = ∅.
(3) Prove that there is a partial recursive unary function which
cannot be extended to a total recursive function.
Exercise 4.8.8. Prove that there are primitive recursive functions 𝑠1 , 𝑠2 ∈
ℱ1 such that whenever 𝜑𝑖2 is bijective, the components of its inverse are
given by 𝜑𝑠11 (𝑖) and 𝜑𝑠12 (𝑖) .
10.1090/stml/089/05
Chapter 5
Models of Arithmetic and

Limitation Theorems
Introduction
Our aim in this chapter is to study models of Peano arithmetic. We start
by describing a way to encode formulas and proofs in a finite signature
in a recursive way. This applies in particular to weak Peano arithmetic
PA0 and full Peano arithmetic PA which we introduce in 5.3. A key
result is the Representability Theorem 5.3.5 which asserts that total re-
cursive functions are represented by Σ1 -formulas. Using the diagonal
argument (Proposition 5.4.1), we prove the theorem of Tarski on the non-
definability of truth (Theorem 5.4.3) and the theorem of Church on the
undecidability of arithmetic (Theorem 5.4.5).
In the Section 5.5 we prove Gödel’s First Incompleteness Theorem.
The last two sections are devoted to a complete proof of Gödel’s cele-
brated second incompleteness theorem. This requires full Peano arith-
metic PA contrary to the previous results for which only weak Peano
arithmetic PA0 was needed. A key ingredient is the definability of satis-
fiability for Σ1 -formulas which is proved in Proposition 5.6.1.
119
120 5. Models of Arithmetic and Limitation Theorems
5.1. Coding Formulas and Proofs

Let 𝜎ℒ = {𝜆1 , … , 𝜆𝑙 } be a finite signature. We assign to each symbol
𝑠 of ℒ its Gödel number ⌜𝑠⌝ as follows: to =, ∧, ¬, (, ), ∃ one assigns
⟨0, 0⟩, ⟨0, 1⟩, … , ⟨0, 5⟩, to 𝜆𝑖 one assigns ⟨0, 𝑖 + 5⟩, and to 𝑣𝑖 (𝑖 ∈ ℕ) one
assigns ⟨1, 𝑖⟩.
To a word 𝑚 = 𝑠1 ⋯ 𝑠𝑛 on the alphabet ℒ we assign the Gödel num-
ber
#𝑚 = ⟨⌜𝑠1 ⌝, … , ⌜𝑠𝑛 ⌝⟩.
This coding is clearly injective.
Lemma 5.1.1. The following sets are primitive recursive:
(1) Term = {#𝑡 ∣ 𝑡 is an ℒ-term}.
(2) Form = {#𝜑 ∣ 𝜑 is an ℒ-formula}.
(3) {(#𝑡, 𝑛) ∣ 𝑡 ∈ 𝒯 ℒ and 𝑣𝑛 has an occurrence in 𝑡} as well as
{(#𝑡, 𝑛) ∣ 𝑡 ∈ 𝒯 ℒ and 𝑣𝑛 has no occurrence in 𝑡}.
ℒ
(4) {(#𝜑, 𝑛) ∣ 𝜑 ∈ Fml and 𝑣𝑛 has an occurrence in 𝜑}, and simi-
larly for ‘has no occurrence’, ‘has at least one free occurrence’,
‘has no free occurrence’, ‘has at least one bound occurrence’
and ‘has no bound occurrence’ instead of ‘has an occurrence’.
(5) {#𝜑 ∣ 𝜑 is an ℒ-sentence}.
Proof. We use properties of the coding function ⟨…⟩ for finite sequences
(cf. Lemma 4.1.4) and the primitive recursive decoding function in
two arguments 𝜆𝑥𝑖.(𝑥)𝑖 . In particular, we use the inequality 𝑛 =
lg(⟨𝑠0 , … , 𝑠𝑛−1 ⟩) ≤ ⟨𝑠0 , … , 𝑠𝑛−1 ⟩.
By unique reading properties, we may not only argue by induction
on length, but also on height of terms and formulas. For instance, if 𝑥 is
the code of a word 𝑚 starting with ‘(’, then 𝑥 is the code of a formula if and
only if there exist 𝑦1 , 𝑦2 < 𝑥 such that 𝑦𝑖 = #𝜑𝑖 for some formulas 𝜑1 , 𝜑2
such that 𝑚 is the word (𝜑1 ∧ 𝜑2 ). The details are left as an exercise. □
Recall that we have defined substitution in terms and formulas. If

𝑥1 , … , 𝑥𝑛 are 𝑛 pairwise distinct variables, and 𝑠1 , … , 𝑠𝑛 are terms, we de-
fined 𝑡𝑠/𝑥 and 𝜑𝑠/𝑥 (see Section 2.4). One easily proves that the functions
𝑆𝑡 ∶ ℕ3 → ℕ and 𝑆𝑓 ∶ ℕ3 → ℕ are primitive recursive, where 𝑆𝑡 is the
5.1. Coding Formulas and Proofs 121
function assigning to (𝑎, 𝑏, 𝑐) the integer #𝑡𝑠/𝑣𝑖 ,…,𝑣𝑖 if 𝑎 = ⟨𝑖1 , … , 𝑖𝑛 ⟩ is

1 𝑛
the code of a sequence of pairwise distinct natural numbers, 𝑏 = #𝑡 ∈
Term and 𝑐 = ⟨#𝑠1 , … , #𝑠𝑛 ⟩, with #𝑠𝑖 ∈ Term for any 𝑖, and 0 otherwise.
The function 𝑆𝑓 is defined in a similar way: to
(𝑎, 𝑏, 𝑐) = (⟨𝑖1 , … , 𝑖𝑛 ⟩, #𝜑, ⟨#𝑠1 , … , #𝑠𝑛 ⟩)
one assigns
𝑆𝑓 (𝑎, 𝑏, 𝑐) = #𝜑𝑠/𝑣𝑖 ,…,𝑣𝑖𝑛 .
1
In particular, one obtains the following statement:
Lemma 5.1.2. There exist primitive recursive functions Subst𝑡 and Subst𝑓
in ℱ3 such that, for any 𝑛, if 𝑠 and 𝑡 are terms and 𝜑 is a formula, then
Subst𝑡 (𝑛, #𝑠, #𝑡) = #𝑡𝑠/𝑣𝑛 and Subst𝑓 (𝑛, #𝑠, #𝜑) = #𝜑𝑠/𝑣𝑛 . □
Similarly, one encodes formulas from propositional calculus, via 𝐹 ↦

#𝐹 ∈ ℕ, with 𝐹 ∈ Fml𝒫 .
Lemma 5.1.3. The set Taut of codes of tautologies for the predicate cal-
culus is primitive recursive.
Proof. One checks directly that

• the set Form𝒫 = {#𝐹 ∣ 𝐹 ∈ Fml𝒫 } is primitive recursive;
• the function assigning to (𝑑, 𝑥) the truth value (0 or 1) of the
formula 𝐹 = 𝐹[𝑝0 , … , 𝑝𝑛−1 ] for the distribution of truth values
𝑑 = ⟨𝑑0 , … , 𝑑𝑛−1 ⟩, with 𝑑 the code of a length 𝑛 sequence of
0’s and 1’s, and 𝑥 = #𝐹 the code of 𝐹 ∈ Fml𝒫 containing only
propositional variables 𝑝𝑖 with 𝑖 ≤ 𝑛 − 1, and 0 otherwise, is
primitive recursive;
• the set Taut𝒫 = {#𝐹 ∣ 𝐹 is a tautology} is primitive recursive;
• there exists a primitive recursive function assigning to the
pair (⟨#𝜑0 , … #𝜑𝑛−1 ⟩, #𝐹) the natural number #𝐹𝜑/𝑝 if 𝐹 =
𝐹[𝑝0 , … , 𝑝𝑛−1 ].
This is enough to get the result. □
Similarly one checks that the set of codes of the equality axioms is
primitive recursive and also the set of codes of the quantifier axioms (tak-
ing into account Lemma 5.1.2). We have thus proved the following state-
ment.
Theorem 5.1.4. The set Ax of codes #𝜑 of logical axioms 𝜑 of ℒ is primi-
tive recursive. □
Coding formal proofs. We encode a finite sequence of ℒ-formulas 𝑑 =

(𝜑0 , … , 𝜑𝑛−1 ) by ##𝑑 = ⟨#𝜑0 , … , #𝜑𝑛−1 ⟩.
Lemma 5.1.5. The set
Prf = {(##𝑑, #𝜑) ∣ 𝑑 is a formal proof of 𝜑}
Proof. Given (𝑥, 𝑦), one first checks if 𝑦 ∈ Form, that is, if 𝑦 = #𝜑 for
some ℒ-formula 𝜑. If yes, it suffices to decode and to test whether all
components of 𝑥 are codes of ℒ-formulas 𝜑𝑖 , the last one being equal to
𝜑, and whether for every 𝑖, either 𝜑𝑖 is a logical axiom (which is a prim-
itive recursive property by Theorem 5.1.4) or it can be obtained using
deduction rules from previous formulas. □
Proposition 5.1.6. The set 𝑈 = {#𝜑 ∣ ⊢ 𝜑} is recursively enumerable.
Proof. We have 𝑛 ∈ 𝑈 ⇔ ∃𝑑 (𝑑, 𝑛) ∈ Prf. □
5.2. Decidable Theories

Let 𝑇 be an ℒ-theory. We denote by Thm(𝑇) = {𝜑 ℒ-sentence ∣ 𝑇 ⊢ 𝜑}
the deductive closure of 𝑇.
Definition. Let 𝑇 be an ℒ-theory.
(1) 𝑇 is called recursive if {#𝜑 ∣ 𝜑 ∈ 𝑇} is a recursive set.
(2) 𝑇 is called recursively (or effectively) axiomatizable if there ex-
ists a recursive ℒ-theory 𝑇 ′ such that Thm(𝑇) = Thm(𝑇 ′ ).
(3) 𝑇 is called decidable if the set #Thm(𝑇) = {#𝜑 ∣ 𝑇 ⊢ 𝜑} is
recursive.
5.2. Decidable Theories 123
If 𝑇 is an ℒ-theory, one sets
Prf(𝑇) = {(##𝑑, #𝜑) ∣ 𝑑 is a proof of 𝜑 in 𝑇}.
Lemma 5.2.1. If 𝑇 is recursive, then the set Prf(𝑇) is recursive.
Theorem 5.2.2. If 𝑇 is effectively axiomatizable, then #Thm(𝑇) is recur-

sively enumerable.
Proof. Let 𝑇0 be a recursive theory such that Thm(𝑇) = Thm(𝑇0 ).

We have 𝜑 ∈ Thm(𝑇) if and only if 𝜑 is an ℒ-sentence and there ex-
ists a proof of 𝜑 in 𝑇0 . One concludes by Lemma 5.1.1(5) in combination
with Lemma 5.2.1. □
Remark 5.2.3. The converse of Theorem 5.2.2 holds, too.
Proof. Assume the set #Thm(𝑇) is recursively enumerable. There then

exists a recursive function 𝑓 ∈ ℱ1 such that
#Thm(𝑇) = {𝑓(𝑖) = #𝜑𝑖 ∣ 𝑖 ∈ ℕ}.

𝑛
The function 𝑔 which to 𝑛 assigns # ⋀𝑖=0 𝜑𝑖 is recursive and we have
𝑛
𝑔(𝑛) ≥ 𝑛 for every 𝑛. Its image is therefore a recursive set. Since {⋀𝑖=0 𝜑𝑖 ∣
𝑛 ∈ ℕ} is an axiomatization of Thm(𝑇), this finishes the proof. □
Theorem 5.2.4. Assume 𝑇 is effectively axiomatizable and complete. Then

𝑇 is decidable.
Proof. By Theorem 5.2.2, we already know that #Thm(𝑇) is recursively

enumerable. Its complement ℕ ⧵ #Thm(𝑇) is the union of {#𝜑 ∣ ¬𝜑 ∈
Thm(𝑇)} and {𝑥 ∣ 𝑥 is not the code of an ℒ-sentence}. The former set
is recursively enumerable since #𝜑 ↦ #¬𝜑 is given by a (primitive)
recursive function, while the latter one is primitive recursive by
Lemma 5.1.1(5). One concludes by Theorem 4.6.4. □
Example 5.2.5.
(1) Any finitely axiomatizable theory is effectively axiomatizable.
(2) ACF𝑝 , 𝑝 prime or 0, is decidable.

[Indeed, the axiomatization is clearly recursive, and one con-
cludes by combining Theorem 3.5.4 (ACF𝑝 is complete) with
Theorem 5.2.4.]
(3) ACF is decidable.

[Indeed, as ACF is effectively axiomatizable, it suffices to prove
that 𝑋 = {#𝜑 ∣ 𝜑 is a sentence and ACF ⊭ 𝜑} is a recursively
enumerable set. But one has ACF ⊭ 𝜑 if and only if there exists
𝑝 prime such that ACF𝑝 ⊧ ¬𝜑, as follows from the Lefschetz
principle, which is equivalent to ACF ⊧ 𝜒𝑝 → ¬𝜑, where 𝜒𝑝 is
the formula ⏟1+
⎵⎵⋯
⏟⎵⎵ +⏟1 = 0.
𝑝 times
It follows that #𝜑 ∈ 𝑋 if and only if there exists a prime 𝑝
and 𝑦 such that (𝑦, #(𝜒𝑝 → ¬𝜑)) ∈ Prf(ACF). Since the map-
ping (𝑝, #𝜑) ↦ #(𝜒𝑝 → ¬𝜑) is (primitive) recursive, one con-
cludes that 𝑋 is recursively enumerable.]
(4) The theory of the ordered field of real numbers ℛ is decidable

(this is a result of Tarski).
[Since it is complete, it suffices by Theorem 5.2.4 to give an
effective axiomatization of Th(ℛ). One can prove that any or-
dered field ⟨𝐾; +, −, 0, 1, ⋅, <⟩ satisfying the intermediate value
property for polynomial functions (that is, if 𝑃(𝑋) ∈ 𝐾[𝑋] and
𝑎, 𝑏 ∈ 𝐾 with 𝑎 < 𝑏 and 𝑃(𝑎) ⋅ 𝑃(𝑏) < 0, there exists 𝑐 ∈ 𝐾
such that 𝑎 < 𝑐 < 𝑏 and 𝑃(𝑐) = 0) is elementarily equivalent to
ℛ. To prove this, one shows that the theory we just described
admits quantifier elimination in the language of ordered rings,
for which we refer to [12].]
Exercise 5.2.6. If 𝑇 is decidable and 𝜑0 , … , 𝜑𝑛−1 are ℒ-sentences, then

𝑇 ∪ {𝜑0 , … , 𝜑𝑛−1 } is decidable.
5.3. Peano Arithmetic 125
5.3. Peano Arithmetic

We shall work in the language of arithmetic ℒ𝑎𝑟 = {0, 𝑆, +, ⋅, <}. In any
ℒ𝑎𝑟 -structure, for any 𝑛 ∈ ℕ, one defines by induction 𝑛, setting 0 = 0
and 𝑛 + 1 = 𝑆(𝑛).
Definition.
(1) The set of weak Peano axioms is the finite set PA0 consisting of
the following 8 axioms:
(A1) ∀𝑣0 ¬𝑆𝑣0 = 0.
(A2) ∀𝑣0 ∃𝑣1 (¬𝑣0 = 0 → 𝑆𝑣1 = 𝑣0 ).
(A3) ∀𝑣0 ∀𝑣1 (𝑆𝑣0 = 𝑆𝑣1 → 𝑣0 = 𝑣1 ).
(A4) ∀𝑣0 𝑣0 + 0 = 𝑣0 .
(A5) ∀𝑣0 ∀𝑣1 𝑣0 + 𝑆𝑣1 = 𝑆(𝑣0 + 𝑣1 ).
(A6) ∀𝑣0 𝑣0 ⋅ 0 = 0.
(A7) ∀𝑣0 ∀𝑣1 𝑣0 ⋅ 𝑆𝑣1 = (𝑣0 ⋅ 𝑣1 ) + 𝑣0 .
(A8) ∀𝑣0 ∀𝑣1 (𝑣0 < 𝑣1 ↔ (∃𝑣2 𝑣2 + 𝑣0 = 𝑣1 ∧ ¬𝑣0 = 𝑣1 )).
(2) The set of Peano axioms is the infinite set PA consisting of PA0 ,
and for each ℒ𝑎𝑟 -formula 𝜑(𝑣0 , … , 𝑣𝑛 ) of the following axiom
of induction (where 𝑣 = (𝑣1 , … , 𝑣𝑛 )):
∀𝑣1 ∀𝑣2 ⋯ ∀𝑣𝑛 ([𝜑(0, 𝑣) ∧ ∀𝑣0 (𝜑(𝑣0 , 𝑣) → 𝜑(𝑆𝑣0 , 𝑣))] → ∀𝑣0 𝜑(𝑣0 , 𝑣)).
Remark 5.3.1.
(1) One has 𝔑st ∶= ⟨ℕ; 0, 𝑠𝑢𝑐𝑐, +, ⋅, <⟩ ⊧ PA.
(2) PA0 is an expansion by definition of the ℒ𝑎𝑟 ⧵ {<}-theory (A1)-
(A7). (This is a consequence of (A8).)
(3) PA0 is a quite weak theory. One may construct models in which
addition (or multiplication) is not commutative or associative;
also, there are models in which < does not define a total order.
(4) An element in a model of PA0 is called non-standard if it is
different from the interpretation of 𝑛 for any 𝑛 ∈ ℕ. By com-
pactness, there exist models of PA (and thus of PA0 ) containing
non-standard elements.
Definition. Let 𝔐 ⊆ 𝔐′ be ℒ𝑎𝑟 -structures. One says that 𝔐 is an initial
segment of 𝔐′ if the following two conditions hold:
• For all 𝑎′ ∈ 𝑀 ′ and all 𝑎 ∈ 𝑀 such that 𝔐′ ⊧ 𝑎′ < 𝑎 one has

𝑎′ ∈ 𝑀.
• For all 𝑎′ ∈ 𝑀 ′ ⧵ 𝑀 and all 𝑎 ∈ 𝑀 one has 𝔐′ ⊧ 𝑎 < 𝑎′ .
Lemma 5.3.2. Let 𝔐 ⊧ PA0 . Then 𝑁 = {𝑛𝔐 ∣ 𝑛 ∈ ℕ} is the base set of a

substructure of 𝔐 which is isomorphic to 𝔑st and an initial segment of 𝔐.
Proof. Let 𝔐 ⊧ PA0 . One proves:

(i) For any 𝑚, 𝑛 ∈ ℕ one has PA0 ⊧ 𝑚 + 𝑛 = 𝑚 + 𝑛.
This is proved by (naive) induction on 𝑛 ∈ ℕ, using (A4)
and (A5).
(ii) For any 𝑚, 𝑛 ∈ ℕ one has PA0 ⊧ 𝑚 ⋅ 𝑛 = 𝑚 ⋅ 𝑛.
This is proved by induction on 𝑛, using (A6), (A7) and (i).
(iii) PA0 ⊧ ∀𝑥∀𝑦(𝑥 < 𝑦 ↔ 𝑆𝑥 < 𝑆𝑦).
Indeed, if 𝑎, 𝑏, 𝑐 ∈ 𝑀, one has the following equivalences:
(A3)
𝔐 ⊧ 𝑐 + 𝑎 = 𝑏 ∧ ¬𝑎 = 𝑏 ⟺ 𝔐 ⊧ 𝑆(𝑐 + 𝑎) = 𝑆𝑏 ∧ ¬𝑆𝑎 = 𝑆𝑏
(A5)
⟺ 𝔐 ⊧ 𝑐 + 𝑆𝑎 = 𝑆𝑏 ∧ ¬𝑆𝑎 = 𝑆𝑏.
The statement thus follows from (A8).

𝑛−1
(iv) For any 𝑛 ∈ ℕ one has PA0 ⊧ ∀𝑥(𝑥 < 𝑛 ↔ ⋁𝑖=0 𝑥 = 𝑖).
This is proved by induction on 𝑛. If 𝔐 ⊧ 𝑐 < 0 for some
𝑐 ∈ 𝑀, then (in 𝔐) 𝑎+𝑐 = 0 for some 𝑎 ∈ 𝑀 and 𝑐 ≠ 0 by (A8),
hence 𝑐 = 𝑆𝑑 for some 𝑑 by (A2), and by (A5) one infers that
0 = 𝑎 + 𝑆𝑑 = 𝑆(𝑎 + 𝑑), which contradicts (A1). This proves
the case 𝑛 = 0. For the induction step 𝑛 ↦ 𝑛 + 1, one uses (iii)
together with the fact that 0 < 𝑐 for any 𝑐 ≠ 0 (a consequence
of (A4) and (A8)).
(v) Let 𝑐 ∈ 𝑀 ⧵ 𝑁 and 𝑛 ∈ ℕ. Then 𝔐 ⊧ 𝑛 < 𝑐.
This is proved by induction on 𝑛, the case 𝑛 = 0 being
clear. For the induction step 𝑛 ↦ 𝑛 + 1, note that 𝑐 = 𝑆𝑑 for
some 𝑑 ∉ 𝑁. Hence 𝔐 ⊧ 𝑛 < 𝑑, which proves 𝔐 ⊧ 𝑛 + 1 < 𝑐,
using (iii).
This concludes the proof. □
Up to naming elements of ℕ by constants (which is irrelevant since

any element of ℕ is the interpretation of a term), we have in particular
proved that if 𝔐 ⊧ PA0 , then 𝔐 ⊧ Δ(𝔑st ).
Definition.
(1) The set of Σ1 -formulas is the smallest set of ℒ𝑎𝑟 -formulas con-
taining the quantifier-free formulas and which is stable under
∧, ∨, existential quantification ∃𝑥 and bounded universal quan-
tification (∀𝑥 < 𝑡), with 𝑡 a term not depending on the variable
𝑥. Here, (∀𝑥 < 𝑡) 𝜑 is defined as ∀𝑥(𝑥 < 𝑡 → 𝜑).
(2) The set of strict Σ1 -formulas is the smallest set of ℒ𝑎𝑟 -formulas
containing the formulas 0 = 𝑥, 𝑆𝑥 = 𝑦, 𝑥 + 𝑦 = 𝑧, 𝑥 ⋅ 𝑦 = 𝑧,
𝑥 = 𝑦, ¬𝑥 = 𝑦, 𝑥 < 𝑦, ¬𝑥 < 𝑦 and which is stable under ∧, ∨,
∃𝑥 and, assuming that 𝑥 and 𝑦 are two distinct variables, under
∀𝑥 < 𝑦.
Lemma 5.3.3.
(1) Any Σ1 -formula is equivalent to a strict Σ1 -formula.
(2) If 𝜑 is a Σ1 -formula, then any formula 𝜑𝑠/𝑥 obtained by substitu-
tion is a Σ1 -formula, too.
Proof. Proceeding as in the proof of Lemma 3.3.1 one may eliminate

complex terms by using existential quantifiers. For instance, the formula
𝑥 ⋅ 𝑆𝑦 = 0 is equivalent to ∃𝑢∃𝑣(𝑆𝑦 = 𝑢 ∧ 𝑥 ⋅ 𝑢 = 𝑣 ∧ 0 = 𝑣). The second
statement is clear. □
Proposition 5.3.4. Any Σ1 -sentence satisfied in 𝔑st is a theorem of PA0 .
Proof. By Lemma 5.3.3, it is enough to prove that for any strict Σ1 -

formula 𝜑(𝑥1 , … , 𝑥𝑛 ) and any 𝑚1 , … , 𝑚𝑛 ∈ ℕ we have
𝔑st ⊧ 𝜑[𝑚1 , … , 𝑚𝑛 ] ⇒ PA0 ⊧ 𝜑(𝑚1 , … , 𝑚𝑛 ).
We shall prove this by induction on the height of the formula starting
from the case when 𝜑 is a quantifier-free formula, which follows from
Lemma 5.3.2.
Assume now that the statement holds for 𝜑(𝑥0 , … , 𝑥𝑛 ) and for
𝜓(𝑥0 , … , 𝑥𝑛 ). Clearly it then also holds for 𝜑 ∧ 𝜓 and 𝜑 ∨ 𝜓. We now
prove that it holds for ∃𝑥0 𝜑(𝑥0 , … , 𝑥𝑛 ). If 𝔑st ⊧ ∃𝑥0 𝜑[𝑚1 , … , 𝑚𝑛 ], then
there exists 𝑚0 ∈ ℕ such that 𝔑st ⊧ 𝜑[𝑚0 , … , 𝑚𝑛 ], which implies PA0 ⊧

𝜑(𝑚0 , … , 𝑚𝑛 ) by the induction hypothesis; hence PA0 ⊧ ∃𝑥0 𝜑(𝑚1 , … , 𝑚𝑛 ).
Finally, if 𝔑st ⊧ (∀𝑥0 < 𝑥1 ) 𝜑[𝑚1 , … , 𝑚𝑛 ], then by definition of
∀𝑥0 < 𝑥1 one has 𝔑st ⊧ 𝜑[𝑚0 , 𝑚1 , … , 𝑚𝑛 ] for any 𝑚0 < 𝑚1 . By the
induction hypothesis, we then have PA0 ⊧ 𝜑(𝑚0 , … , 𝑚𝑛 ) for any 𝑚0 <
𝑚1 . By (iv) in the proof of Lemma 5.3.2, one deduces PA0 ⊧ (∀𝑥0 <
𝑥1 )𝜑(𝑚1 , … , 𝑚𝑛 ). □
Definition.
(1) Let 𝑓 ∈ ℱ𝑝 . One says that the formula 𝜑(𝑥1 , … , 𝑥𝑝+1 ) represents
𝑓 if for any 𝑛1 , … , 𝑛𝑝 ∈ ℕ one has
PA0 ⊧ ∀𝑦 (𝜑(𝑛1 , … , 𝑛𝑝 , 𝑦) ↔ 𝑦 = 𝑓(𝑛1 , … , 𝑛𝑝 )) .
(2) Let 𝐴 ⊆ ℕ𝑝 . One says that 𝜑(𝑥1 , … , 𝑥𝑝 ) represents 𝐴 if for any

𝑛1 , … , 𝑛𝑝 ∈ ℕ one has
𝑛 ∈ 𝐴 ⇒ PA0 ⊧ 𝜑(𝑛1 , … , 𝑛𝑝 )
and
𝑛 ∉ 𝐴 ⇒ PA0 ⊧ ¬𝜑(𝑛1 , … , 𝑛𝑝 ).
Remark. If 𝜑(𝑥1 , … , 𝑥𝑝 ) represents 𝐴, then 𝟙𝐴 is represented by the for-

mula (𝜑(𝑥) ∧ 𝑥𝑝+1 = 1) ∨ (¬𝜑(𝑥) ∧ 𝑥𝑝+1 = 0). Conversely, if 𝜓(𝑥, 𝑥𝑝+1 )
represents 𝟙𝐴 , then 𝜓(𝑥, 1) represents 𝐴. □
Theorem 5.3.5 (Representability Theorem). Every total recursive func-

tion is represented by a Σ1 -formula.
Proof. Let 𝒮 be the set of total functions that are represented by a Σ1 -

formula. By Theorem 4.7.1, it is enough to establish the following three
facts:
(i) The basic functions 𝑆, 𝑃𝑖𝑛 , 𝐶00 , +, ⋅, 𝟙= and 𝟙< are in 𝒮.
(ii) 𝒮 is stable under composition.
(iii) 𝒮 is stable under the total 𝜇-operator.
Item (i) is clear, since the basic functions are all represented by quantifier-
free formulas (which are Σ1 by definition).
If 𝑓1 , … , 𝑓𝑝 ∈ ℱ𝑚 are represented by the formulas 𝜑1 , … , 𝜑𝑝 and if
𝑔 ∈ ℱ𝑝 is represented by 𝜓, then ℎ = 𝑔(𝑓1 , … , 𝑓𝑝 ) ∈ ℱ𝑚 is represented by
𝑝
∃𝑦1 , … , ∃𝑦𝑝 𝜑𝑖 (𝑥, 𝑦𝑖 ) ∧ 𝜓(𝑦1 , … , 𝑦𝑝 , 𝑥𝑚+1 ).
⋀
𝑖=1
In particular, this proves (ii).

Finally, let 𝑓 ∈ ℱ𝑝+1 ∈ 𝒮, and let 𝑔 ∈ ℱ𝑝 be the function obtained
from 𝑓 by the total 𝜇-operator, that is, 𝑔(𝑥) = 𝜇𝑦 (𝑓(𝑥, 𝑦) = 0). Let
𝐴 = {(𝑥, 𝑦) ∣ 𝑓(𝑥, 𝑦) = 0} ⊆ ℕ𝑝+1 . Applying (i) and (ii), we get 𝟙𝐴 =
𝜆𝑥𝑦.𝟙= (𝑓(𝑥, 𝑦), 0) ∈ 𝒮. Choose a Σ1 -formula 𝜑(𝑥, 𝑦, 𝑧) representing 𝟙𝐴 .
Then the function 𝑔 is represented by the Σ1 -formula 𝜓(𝑥, 𝑦) given by
𝜑(𝑥, 𝑦, 1) ∧ (∀𝑦′ < 𝑦) 𝜑(𝑥, 𝑦′ , 0). This proves (iii). □
Corollary 5.3.6. A subset 𝐴 ⊆ ℕ𝑛 is recursively enumerable if and only if

there exists a Σ1 -formula defining 𝐴 in the structure 𝔑st .
Proof. Any subset defined by a quantifier-free formula is (primitive) re-

cursive, hence a subset defined by a Σ1 -formula is recursively enumer-
able by the closure properties of recursively enumerable sets, their stabil-
ity under existential quantification (that is, under projection), bounded
universal quantification, intersection and union (Proposition 4.6.1).
Conversely, since any recursively enumerable set is the projection of
a recursive set, it is enough to prove that any recursive set is defined by
a Σ1 -formula, which follows from the Representability Theorem. □
Proposition 5.3.7. Let 𝔐 be a model of PA. Then the following properties

are satisfied in 𝔐:
(1) + and ⋅ are commutative and associative.
(2) ⋅ is distributive with respect to +.
(3) < defines a total order, compatible with + (if 𝑎 < 𝑏, then 𝑎 + 𝑐 <
𝑏 + 𝑐) and with multiplication with a non-zero element (if 𝑎 < 𝑏
and 𝑐 ≠ 0, then 𝑎 ⋅ 𝑐 < 𝑏 ⋅ 𝑐); 0 is the minimal element for this
order, and for any 𝑎, 𝑆𝑎 is the immediate successor of 𝑎.
(4) Any element is regular for addition (𝑎 + 𝑐 = 𝑏 + 𝑐 ⇒ 𝑎 = 𝑏), and

any non-zero element is regular for multiplication (if 𝑎 ⋅ 𝑐 = 𝑏 ⋅ 𝑐
and 𝑐 ≠ 0, then 𝑎 = 𝑏).
Proof. All these properties are easily proved using the induction axioms
of PA. As examples, let us provide the arguments for associativity and
commutativity of addition. The proof of the remaining properties is sim-
ilar.
Associativity of addition.
Let 𝜑(𝑥, 𝑦, 𝑧) be the formula 𝑥 + (𝑦 + 𝑧) = (𝑥 + 𝑦) + 𝑧. In 𝔐, we have
𝑎 + (𝑏 + 0) = 𝑎 + 𝑏 = (𝑎 + 𝑏) + 0 by (A4), and hence 𝔐 ⊧ ∀𝑥, 𝑦 𝜑(𝑥, 𝑦, 0).
If 𝑎, 𝑏, 𝑐 ∈ 𝑀 verify (𝑎 + 𝑏) + 𝑐 = 𝑎 + (𝑏 + 𝑐), then
𝑎 + (𝑏 + 𝑆𝑐) = 𝑎 + 𝑆(𝑏 + 𝑐) = 𝑆(𝑎 + (𝑏 + 𝑐))

= 𝑆((𝑎 + 𝑏) + 𝑐) = (𝑎 + 𝑏) + 𝑆𝑐
by (A5), whence 𝔐 ⊧ ∀𝑥, 𝑦, 𝑧 (𝜑(𝑥, 𝑦, 𝑧) → 𝜑(𝑥, 𝑦, 𝑆𝑧)). The induction

axiom for 𝜑 implies that 𝔐 ⊧ ∀𝑥, 𝑦, 𝑧 𝜑(𝑥, 𝑦, 𝑧).
Commutativity of addition.
One checks the following properties:
(i) 𝔐 ⊧ ∀𝑥 0 + 𝑥 = 𝑥 (by induction, using (A4) and (A5)).

(ii) 𝔐 ⊧ ∀𝑥 1 + 𝑥 = 𝑆𝑥 (by induction, using (i) and (A5)).
(iii) 𝔐 ⊧ ∀𝑥, 𝑦 𝑥 + 𝑦 = 𝑦 + 𝑥 (by induction, using (i), (ii) and asso-
ciativity of addition). □
Remark. By Proposition 5.3.7, PA suffices for the purposes of basic

number theory. However, as we will see later in this chapter, PA is not a
complete theory. Concrete mathematical statements independent from
PA are not so easy to provide, but it is known that for instance Good-
stein’s theorem from Exercise 1.11.3 is such an example.
Lemma 5.3.8 (Overspill). Let 𝔐 be a non-standard model of PA (that

is, a model containing a non-standard element) and let 𝜑(𝑥) be an ℒ𝑎𝑟 -
formula. If 𝔐 ⊧ 𝜑(𝑛) for every 𝑛 ∈ ℕ, then there exists 𝑐 ∈ 𝑀 non-
standard such that 𝔐 ⊧ 𝜑[𝑐].
5.4. The Theorems of Tarski and Church 131
Proof. Otherwise, 𝜑 would be satisfied precisely by the standard ele-

ments, yielding 𝔐 ⊧ 𝜑(0) and 𝔐 ⊧ ∀𝑥(𝜑(𝑥) → 𝜑(𝑆𝑥)). But then 𝔐 ⊧
∀𝑥 𝜑(𝑥), which is a contradiction since 𝔐 contains non-standard ele-
ments by hypothesis. □
5.4. The Theorems of Tarski and Church

We fix a primitive recursive function subst ∈ ℱ2 such that if 𝜑(𝑣)
is any ℒ𝑎𝑟 -formula with one free variable 𝑣 and 𝑛 ∈ ℕ, then one has
subst(#𝜑, 𝑛) = #𝜑(𝑛). Note that such a function subst clearly exists.
By the Representability Theorem, we may choose a Σ1 -formula
𝐺(𝑥, 𝑦, 𝑧) representing subst. Given a formula 𝜑(𝑣) with one free
variable 𝑣, we consider the formula
𝐻𝜑 (𝑥) ∶= ∃𝑧 (𝐺(𝑥, 𝑥, 𝑧) ∧ 𝜑(𝑧)) ,
and we set 𝑛𝜑 ∶= #¬𝐻𝜑 and Δ𝜑 ∶= ¬𝐻𝜑 (𝑛𝜑 ).
Proposition 5.4.1 (The Diagonal Argument). Let 𝜑(𝑣) be a formula with

one free variable, in the language ℒ𝑎𝑟 . Then
PA0 ⊢ 𝜑(#Δ𝜑 ) ↔ ¬Δ𝜑 .
Proof. Let 𝔐 be a model of PA0 . Since by the above definitions

subst(𝑛𝜑 , 𝑛𝜑 ) = subst(#¬𝐻𝜑 , 𝑛𝜑 ) = #Δ𝜑 , we have
(∗) 𝔐 ⊧ ∀𝑧 (𝐺(𝑛𝜑 , 𝑛𝜑 , 𝑧) ↔ 𝑧 = #Δ𝜑 ) ,
and thus
𝔐 ⊧ 𝜑(#Δ𝜑 ) ⇔ 𝔐 ⊧ ∃𝑧 (𝐺(𝑛𝜑 , 𝑛𝜑 , 𝑧) ∧ 𝜑(𝑧)) ⇔ 𝔐 ⊧ ¬Δ𝜑
⏟⎵⎵⎵⎵⎵⎵⏟⎵⎵⎵⎵⎵⎵⏟
𝐻𝜑 (𝑛𝜑 )
by (∗) and by the definition of Δ𝜑 , respectively. □

Corollary 5.4.2 (Fixed Point Theorem). Let 𝜑(𝑣) be an ℒ𝑎𝑟 -formula with
one free variable. Then there exists an ℒ𝑎𝑟 -sentence 𝜓 such that
PA0 ⊢ 𝜑(#𝜓) ↔ 𝜓.
Proof. One may take Δ¬𝜑 as 𝜓 and use the Diagonal Argument. □
Theorem 5.4.3 (Tarski’s Theorem on the Non-definability of Truth). Let

𝔐 ⊧ PA0 . Then there is no ℒ𝑎𝑟 -formula 𝑆(𝑣) with one free variable such
that for any ℒ𝑎𝑟 -sentence 𝜓 one has 𝔐 ⊧ 𝜓 if and only if 𝔐 ⊧ 𝑆(#𝜓).
Proof. Assume 𝑆 is such a formula. By Proposition 5.4.1 one has PA0 ⊢

𝑆(#Δ𝑆 ) ↔ ¬Δ𝑆 . In particular, 𝔐 ⊧ 𝑆(#Δ𝑆 ) ↔ ¬Δ𝑆 , which is absurd in
view of the property satisfied by 𝑆. □
Corollary 5.4.4. There is no ℒ𝑎𝑟 -formula 𝑆𝔑st (𝑣) such that for any ℒ𝑎𝑟 -
sentence 𝜓 one has 𝔑st ⊧ 𝜓 if and only if 𝔑st ⊧ 𝑆𝔑st (#𝜓). In particular,
Th(𝔑st ) is undecidable.
Proof. The first part is a special case of Theorem 5.4.3. Assume now, for
contradiction, that 𝑇 = Th(𝔑st ) is decidable. This means that #Thm(𝑇)
is a recursive set and thus representable by a Σ1 -formula 𝜑(𝑣) by Theo-
rem 5.3.5. For any ℒ𝑎𝑟 -sentence 𝜓 we then have 𝔑st ⊧ 𝜓 if and only if
𝔑st ⊧ 𝜑(#𝜓), contradicting the first part. □
Theorem 5.4.5 (Church’s Theorem). Let 𝑇 ⊇ PA0 be a consistent ℒ𝑎𝑟 -

theory. Then 𝑇 is undecidable.
Proof. Otherwise, #Thm(𝑇) would be a recursive set, and there

would exist, by the Representability Theorem, a formula 𝜏(𝑣) represent-
ing #Thm(𝑇). Applying Proposition 5.4.1 we would get
𝑇 ⊢ Δ𝜏 ⇔ #Δ𝜏 ∈ #Thm(𝑇) ⇔ PA0 ⊢ 𝜏(#Δ𝜏 ) ⇔ PA0 ⊢ ¬Δ𝜏 ,
which is absurd, since by assumption 𝑇 is a consistent theory with 𝑇 ⊇

PA0 . □
Church’s Theorem yields undecidability of the Predicate Calculus:
Corollary 5.4.6 (Church). There exists a finite language ℒ such that 𝑈 =

{#𝜑 ∣ 𝜑 is a universally valid ℒ-formula} is not recursive.
Proof. Let ℒ = ℒ𝑎𝑟 . Assume 𝑈 is recursive. Then, as a formula 𝜑

is universally valid if and only if a universal closure of 𝜑 is, the empty
ℒ𝑎𝑟 -theory is decidable. Since PA0 is a finite theory, PA0 would then
5.5. Gödel’s First Incompleteness Theorem 133
be decidable by Exercise 5.2.6, contradicting the Theorem of Church.

Indeed,
8
𝜑 ∈ Thm(PA0 ) ⇔ PA0 ⊢ 𝜑 ⇔ ⊢ 𝐴 → 𝜑 ⇔ 𝜓𝜑 ∈ Thm(∅). □
⋀ 𝑖
⏟⎵⎵⏟⎵⎵⏟
𝑖=1
𝜓𝜑
Remark. One can prove that the corollary already holds for ℒ consist-
ing only of a binary relation symbol, that is, ℒ𝑠𝑒𝑡 = {∈}. Examples of
simpler languages for which this no longer holds are discussed in Exer-
cise 5.8.3.
5.5. Gödel’s First Incompleteness Theorem

From Church’s Theorem we may now derive Gödel’s First Incomplete-
ness Theorem, which Gödel had proved prior to Church’s work.
Theorem 5.5.1 (Gödel’s First Incompleteness Theorem). Let 𝑇 be a con-
sistent and recursive ℒ𝑎𝑟 -theory such that 𝑇 ⊇ PA0 . Then 𝑇 is not com-
plete.
Proof. If 𝑇 were complete, it would be decidable by Theorem 5.2.4, con-

tradicting Church’s Theorem. □
In the remainder of this section, we shall keep the hypotheses of

Theorem 5.5.1, namely 𝑇 ⊇ PA0 is consistent and recursive. We will
now present an improvement of Theorem 5.5.1, due to Rosser, providing
an explicit sentence 𝜓 such that neither 𝑇 ⊢ 𝜓 nor 𝑇 ⊢ ¬𝜓.
We recall that the set
Prf(𝑇) = {(#𝜑, ##𝑑) ∣ 𝑑 is a formal proof of 𝜑 in 𝑇}
is recursive. By the Representability Theorem, there exists a Σ1 -formula
𝑃𝑇 (𝑥, 𝑦) representing Prf(𝑇) ⊆ ℕ2 .
Let Neg ∶ ℕ → ℕ be a primitive recursive function such that
Neg(#𝜑) = #¬𝜑 and Neg(𝑥) = 0 if 𝑥 is not the code of a formula. Fix a
Σ1 -formula 𝜈(𝑥, 𝑦) representing Neg. We set
𝑃𝑇𝑅 (𝑥, 𝑦) ∶= 𝑃𝑇 (𝑥, 𝑦) ∧ ¬(∃𝑧 ≤ 𝑦)∃𝑢(𝑃𝑇 (𝑢, 𝑧) ∧ 𝜈(𝑥, 𝑢)).
Let ℎ𝑇𝑅 (𝑥) ∶= ∃𝑦𝑃𝑇𝑅 (𝑥, 𝑦), and Δ𝑅𝑇 ∶= Δℎ𝑅 .
𝑇
Theorem 5.5.2 (Rosser’s Variant of the First Incompleteness Theorem).

Let 𝑇 ⊇ PA0 be consistent and recursive. Then 𝑇 ⊬ Δ𝑅𝑇 and 𝑇 ⊬ ¬Δ𝑅𝑇 .
Proof. We set Δ ∶= Δ𝑅𝑇 and 𝑚 ∶= #Δ. By the Diagonal Argument

(Proposition 5.4.1), we have
(†) PA0 ⊢ ∃𝑦𝑃𝑇𝑅 (𝑚, 𝑦) ↔ ¬Δ.
Assume 𝑇 ⊢ Δ. Then there exists a proof of Δ in 𝑇, thus there exists

𝑝 ∈ ℕ such that (𝑚, 𝑝) ∈ Prf(𝑇), whence PA0 ⊢ 𝑃𝑇 (𝑚, 𝑝). Since 𝑇 is
consistent, for any 𝑝′ ∈ ℕ we have PA0 ⊢ ¬∃𝑢 (𝜈(𝑚, 𝑢) ∧ 𝑃𝑇 (𝑢, 𝑝′ )). It
𝑝
is enough to note that PA0 ⊢ (𝑥 ≤ 𝑝 ↔ ⋁𝑖=0 𝑥 = 𝑖) to deduce that
PA0 ⊢ 𝑃𝑇𝑅 (𝑚, 𝑝). By (†), we have PA0 ⊢ ¬Δ, and hence 𝑇 ⊢ ¬Δ (since
𝑇 ⊇ PA0 ). This is a contradiction.
Assume 𝑇 ⊢ ¬Δ. Then there exists a natural number 𝑝 such that
PA0 ⊢ 𝑃𝑇 (#¬Δ, 𝑝). Since 𝑇 is consistent, for any 𝑝′ ∈ ℕ we have PA0 ⊢
¬𝑃𝑇 (𝑚, 𝑝′ ). It follows that PA0 ⊢ ¬∃𝑦𝑃𝑇𝑅 (𝑚, 𝑦), since 𝔑st is an initial
segment of any model of PA0 by Lemma 5.3.2. By (†), we have PA0 ⊢ Δ,
and hence 𝑇 ⊢ Δ, which is a contradiction. □
Remark. Let 𝑇 ⊇ PA0 be a consistant and recursive ℒ𝑎𝑟 -theory. Let

ℎ𝑇 (𝑥) ∶= ∃𝑦𝑃𝑇 (𝑥, 𝑦) and Δ𝑇 ∶= Δℎ𝑇 . Then 𝑇 ⊬ Δ𝑇 .
Proof. The proof is similar to that of Theorem 5.5.2 and left as an exer-
cise. □
5.6. Definability of Satisfiability for Σ1 -formulas

By Tarski’s result (Theorem 5.4.3), there is no formula expressing satis-
fiability for all arithmetical sentences. However, if one restricts to Σ1 -
sentences, one has the following representability result:
Proposition 5.6.1. There exists a Σ1 -formula SatΣ1 (𝑣) such that for any
Σ1 -sentence 𝜑 one has
PA ⊢ 𝜑 ↔ SatΣ1 (#𝜑).
5.6. Definability of Satisfiability for Σ1 -formulas 135
Before we give a proof the proposition, we prove a result which guar-

antees that various arguments involving coding may be carried out in
any model of PA.
Lemma 5.6.2. Let 𝑓 ∈ ℱ𝑛 be a primitive recursive function. Then there is

a Σ1 -formula 𝜒𝑓 (𝑥1 … , 𝑥𝑛 , 𝑦) representing 𝑓 such that
PA ⊢ ∀𝑥1 , … , 𝑥𝑛 ∃! 𝑦𝜒𝑓 .
Proof. We say that function 𝑓 ∈ ℱ is provably total Σ1 if the conclusion

of the lemma holds for 𝑓. We will only sketch the argument and leave
the (rather tedious) details to the reader.
Claim 1. Gödel’s 𝛽-function 𝛽 ∈ ℱ3 is provably total Σ1 .
Recall that 𝛽(𝑥1 , 𝑥2 , 𝑥3 ) = 𝜇𝑧 [𝑧 ≡ 𝑥1 mod (𝑥2 (𝑥3 + 1) + 1)]. One checks
that the formula 𝜒𝛽 (𝑥1 , 𝑥2 , 𝑥3 , 𝑦) given by
(∗) 𝑦 < 𝑥2 (𝑥3 + 1) + 1 ∧ ∃𝑣 𝑥1 = 𝑥0 + 𝑣((𝑥3 + 1) + 1)
is a Σ1 -formula which represents 𝛽 and defines a total function in any

model of PA. This proves Claim 1.
Claim 2. Let 𝔐 be a model of PA, 𝛾(𝑥, 𝑦) an ℒ𝑎𝑟 -formula with parameters
from 𝑀 and 𝑖 ∈ 𝑀 such that 𝔐 ⊧ (∀𝑗 < 𝑖)∃! 𝑦𝛾(𝑗, 𝑦). Then there are
𝑎, 𝑏 ∈ 𝑀 such that 𝔐 ⊧ (∀𝑗 < 𝑖)∀𝑦(𝜒𝛽 (𝑎, 𝑏, 𝑗, 𝑦) ↔ 𝛾(𝑗, 𝑦)), with 𝜒𝛽 as in
(∗).
This holds in 𝔑st by Lemma 4.7.2(5). If it did not hold in 𝔐, there
would be a minimal 𝑖0 where it fails. The proof that if it holds for 𝑖0 −
1, then it holds for 𝑖0 only uses basic number theory and may thus be
carried out in 𝔐 ⊧ PA. Thus, a minimal counter-example cannot exist,
proving Claim 2.
Claim 3. The set of provably total Σ1 -functions contains the basic functions
(𝑆, 𝑃𝑖𝑛 and 𝐶00 ), and it is stable under composition and recursion.
The statements about the basic functions and composition are clear.
To prove stability under recursion, suppose 𝑓 ∈ ℱ𝑛+1 is defined by recur-
sion using 𝑔 ∈ ℱ𝑛 and ℎ ∈ ℱ𝑛+2 , and that 𝜓(𝑥1 , … , 𝑥𝑛 , 𝑦) is a Σ1 -formula
representing 𝑔 which defines a total function in any model of PA, and
similarly 𝜃(𝑥1 , … , 𝑥𝑛+2 , 𝑦) for ℎ. Using Claim 2, one may then construct
a Σ1 -formula 𝜑(𝑥1 , … , 𝑥𝑛+1 , 𝑦) representing 𝑓 which defines a total func-

tion in any model of PA, by coding the computation of 𝑓 as in the proof
of Theorem 4.7.1. □
Remark 5.6.3. Consider the expansion by definitions of PA given by
adding new symbols for the functions defined by the formulas 𝜒𝑓 as in
Lemma 5.6.2. By the proofs of Lemma 3.3.1 and Proposition 3.3.2, not-
ing in addition that the formula ¬𝜒𝑓 (𝑥, 𝑦) is equivalent in PA to the Σ1 -
formula ∃𝑧(𝜒𝑓 (𝑥, 𝑧) ∧ 𝑦 ≠ 𝑧), it follows from Lemma 5.6.2 that every
Σ1 -formula in the expansion is equivalent to a Σ1 -formula in ℒ𝑎𝑟 . Be-
low, we will construct Σ1 -formulas through this process. Abusing nota-
tion, we will use the same letter 𝑓 for the function symbol and for the
corresponding primitive recursive function.
We will in particular use the binary function (𝑥)𝑖 to decode sequences
in 𝔐 ⊧ PA which are of ‘finite’ length in the sense of 𝔐.
Proof of Proposition 5.6.1. By Lemma 5.3.3 any Σ1 -formula 𝜑 is logi-

cally equivalent to a strict Σ1 -formula. It follows from the proof of that
lemma that there exists a primitive recursive function 𝑓 such that, for
any Σ1 -formula 𝜑, 𝑓(#𝜑) is the Gödel number of a strict Σ1 -formula log-
ically equivalent to 𝜑. It is thus enough to prove the existence of a Σ1 -
formula 𝑆 ′ (𝑣) such that for any strict Σ1 -sentence 𝜑, one has PA ⊢ 𝜑 ↔
𝑆 ′ (#𝜑).
By a certificate of a strict Σ1 -formula 𝜑 we shall mean a finite
sequence (𝜑1 , … , 𝜑𝑛 ) of strict Σ1 -formulas with 𝜑𝑛 equal to 𝜑 and a
finite sequence (𝑠1 , … , 𝑠𝑛 ) of finite sequences of natural numbers 𝑠𝑚 =
(𝑠𝑚 𝑚
0 , … , 𝑠𝛼(𝑚)−1 ) such that the following conditions hold:
(i) All free variables of 𝜑𝑚 belong to 𝑣0 , . . . , 𝑣𝛼(𝑚)−1 .

(ii) If 𝜑𝑚 is the formula 0 = 𝑣𝑖 then 0 = 𝑠𝑚
𝑖 .
(iii) If 𝜑𝑚 is the formula 𝑆(𝑣𝑖 ) = 𝑣𝑗 then 𝑠𝑚 𝑚
𝑖 + 1 = 𝑠𝑗 .
(iv) If 𝜑𝑚 is the formula 𝑣𝑖 + 𝑣𝑗 = 𝑣𝑘 then 𝑠𝑚 𝑚 𝑚
𝑖 + 𝑠𝑗 = 𝑠𝑘 .
(v) If 𝜑𝑚 is the formula 𝑣𝑖 ⋅ 𝑣𝑗 = 𝑣𝑘 then 𝑠𝑚 𝑚 𝑚
𝑖 𝑠𝑗 = 𝑠𝑘 .
(vi) If 𝜑𝑚 is the formula 𝑣𝑖 = 𝑣𝑗 then 𝑠𝑚 𝑚
𝑖 = 𝑠𝑗 .
(vii) If 𝜑𝑚 is the formula ¬𝑣𝑖 = 𝑣𝑗 then 𝑠𝑚 𝑚
𝑖 ≠ 𝑠𝑗 .
(viii) If 𝜑𝑚 is the formula 𝑣𝑖 < 𝑣𝑗 then 𝑠𝑚 𝑚
𝑖 < 𝑠𝑗 .
5.7. Gödel’s Second Incompleteness Theorem 137
(ix) If 𝜑𝑚 is the formula ¬𝑣𝑖 < 𝑣𝑗 then 𝑠𝑚 𝑚

𝑖 ≮ 𝑠𝑗 .
(x) If 𝜑𝑚 is the formula 𝜑′ ∧ 𝜑″ then ∃𝑚′ , 𝑚″ < 𝑚 such that 𝜑′
′ ″
equals 𝜑𝑚′ , 𝜑″ equals 𝜑𝑚″ and 𝑠𝑚 = 𝑠𝑚 = 𝑠𝑚 .
(xi) If 𝜑𝑚 is the formula 𝜑′ ∨ 𝜑″ then ∃𝑚′ < 𝑚 such that 𝜑′ equals
′
𝜑𝑚′ and 𝑠𝑚 = 𝑠𝑚 , or ∃𝑚″ < 𝑚 such that 𝜑″ equals 𝜑𝑚″ and
𝑚 𝑚″
𝑠 =𝑠 .
(xii) If 𝜑𝑚 is the formula ∃𝑣𝑖 𝜑′ then ∃𝑚′ < 𝑚 such that 𝜑′ equals
′
𝜑𝑚′ and 𝑠𝑚 𝑚 ′
𝑘 = 𝑠𝑘 , ∀𝑘 < min(𝛼(𝑚), 𝛼(𝑚 )), 𝑘 ≠ 𝑖.
(xiii) If 𝜑𝑚 is the formula (∀𝑣𝑖 < 𝑣𝑗 ) 𝜑′ then ∀𝑎 < 𝑠𝑗𝑚 ∃𝑚′ < 𝑚 such
′ ′
that 𝜑′ equals 𝜑𝑚′ , 𝑠𝑖𝑚 = 𝑎 and such that 𝑠𝑚
𝑘 = 𝑠𝑚
𝑘 , ∀𝑘 <
′
min(𝛼(𝑚), 𝛼(𝑚 )), 𝑘 ≠ 𝑖.
This notion of certificate makes sense in any model 𝔐 of PA and one
has
𝔐 ⊧ “𝜑 has a certificate” if and only if 𝔐 ⊧ 𝜑.
This is checked by induction on #𝜑 in the model 𝔐 (it is crucial here
that 𝔐 is a model of PA and not only of PA0 ).
Furthermore, one may express the existence of a certificate for 𝜑 in
𝔐 by a Σ1 -formula in the Gödel number of 𝜑. More precisely, there exists
a Σ1 -formula 𝑆 ′ (𝑣), such that in any model 𝔐 of PA, for any code #𝜑 of
a ‘strict Σ1 -formula’ (even non-standard), 𝔐 ⊧ 𝑆 ′ (#𝜑) if and only if 𝔐 ⊧
“𝜑 has a certificate”. One constructs 𝑆 ′ (𝑣) by coding conditions (i)-(xiii)
that are clearly expressed by Σ1 -conditions. The equivalence between
𝔐 ⊧ 𝑆 ′ (#𝜑) and 𝔐 ⊧ “𝜑 has a certificate” is checked by induction on #𝜑
in the model 𝔐. Here again it is important that 𝔐 is a model of PA. □
5.7. Gödel’s Second Incompleteness Theorem

Let 𝑇 be a recursive ℒ𝑎𝑟 -theory containing PA. To state and to prove the
Second Incompleteness Theorem, we will explicitly construct a specific
formula expressing that a sentence is provable.
We consider Σ1 -formulas Sen(𝑣), Ax𝑇 (𝑣), MP(𝑣0 , 𝑣1 , 𝑣2 ) as well as
Gen(𝑣0 , 𝑣1 ) representing respectively the set of Gödel numbers of sen-
tences, the union of the sets of Gödel numbers of logical axioms and of
those of 𝑇, the set of triples (#𝜑, #𝜓, #𝛿) with 𝜑, 𝜓 and 𝛿 formulas such
that 𝛿 is obtained from 𝜑 and 𝜓 by modus ponens, and the set of pairs
(#𝜑, #𝜓) with 𝜑 and 𝜓 such that 𝜓 is obtained from 𝜑 by generalization.
We shall also use Gödel’s 𝛽-function to code sequences. We consider
the following formula 𝐵(𝑎, 𝑏, 𝑛):
∀𝑖 < 𝑛[Ax𝑇 (𝛽(𝑎, 𝑏, 𝑖)) ∨ (∃𝑗, 𝑘 < 𝑖) MP(𝛽(𝑎, 𝑏, 𝑗), 𝛽(𝑎, 𝑏, 𝑘), 𝛽(𝑎, 𝑏, 𝑖))
∨ (∃𝑗 < 𝑖) Gen(𝛽(𝑎, 𝑏, 𝑗), 𝛽(𝑎, 𝑏, 𝑖))].
The formula Pr𝑇 (𝑚) given by
Sen(𝑚) ∧ ∃𝑎, 𝑏, 𝑛(𝛽(𝑎, 𝑏, 𝑛) = 𝑚 ∧ 𝐵(𝑎, 𝑏, 𝑛 + 1))
expresses that 𝑚 is the Gödel number of a sentence having a formal proof

in 𝑇. If 𝜑 is a sentence, 𝑇 ⊢ 𝜑 if and only if 𝔑st ⊧ Pr𝑇 (#𝜑). In particular,
if 𝜑 is a Σ1 -sentence, 𝔑st ⊧ 𝜑 ⟹ 𝔑st ⊧ Pr𝑇 (#𝜑). In order for this
property to hold in any model of PA, we are led to consider the following
variants of 𝐵 and Pr𝑇 .
One defines the formula 𝐵˜(𝑎, 𝑏, 𝑛) by
∀𝑖 < 𝑛[Ax𝑇 (𝛽(𝑎, 𝑏, 𝑖)) ∨ (∃𝑗, 𝑘 < 𝑖) MP(𝛽(𝑎, 𝑏, 𝑗), 𝛽(𝑎, 𝑏, 𝑘), 𝛽(𝑎, 𝑏, 𝑖))
∨ (∃𝑗 < 𝑖) Gen(𝛽(𝑎, 𝑏, 𝑗), 𝛽(𝑎, 𝑏, 𝑖)) ∨ SatΣ1 (𝛽(𝑎, 𝑏, 𝑖))]
˜ 𝑇 (𝑚) by
and Pr
Sen(𝑚) ∧ ∃𝑎, 𝑏, 𝑛(𝛽(𝑎, 𝑏, 𝑛) = 𝑚 ∧ 𝐵˜(𝑎, 𝑏, 𝑛 + 1)).
˜ 𝑇 (#𝜑).
Finally, if 𝜑 is a sentence we write □𝑇 𝜑 for Pr
By construction the following properties hold:
Proposition 5.7.1.
(1) If 𝜑 is a Σ1 -sentence,
PA ⊢ 𝜑 → □𝑇 𝜑.
(2) For any sentence 𝜓,
𝔑st ⊧ □𝑇 𝜓 ⟺ 𝑇 ⊢ 𝜓.
5.7. Gödel’s Second Incompleteness Theorem 139
Proof. Let 𝜑 be a Σ1 -sentence and let 𝔐 be a model of PA. By Propo-

sition 5.6.1, if 𝔐 ⊧ 𝜑, it follows that 𝔐 ⊧ SatΣ1 (#𝜑), and hence also
𝔐 ⊧ □𝑇 𝜑. This proves (1).
Note that if 𝜑 is a Σ1 -sentence which is true in 𝔑st , then 𝑇 ⊢ 𝜑. Part
(2) thus follows from the construction of Pr ˜ 𝑇 (𝑚). □
Proposition 5.7.2. Let 𝜑 and 𝜓 be sentences. The operator □𝑇 satisfies
the following properties (Loeb’s axioms):
(L1) 𝑇 ⊢ 𝜑 ⟹ 𝑇 ⊢ □𝑇 𝜑.
(L2) 𝑇 ⊢ (□𝑇 𝜑 ∧ □𝑇 (𝜑 → 𝜓)) → □𝑇 𝜓.
(L3) 𝑇 ⊢ □𝑇 𝜑 → □𝑇 □𝑇 𝜑.
Proof. (L1) By Proposition 5.7.1(2), if 𝑇 ⊢ 𝜑, then 𝔑st ⊧ □𝑇 𝜑. Since Pr ˜𝑇

is Σ1 , it follows that □𝑇 𝜑 is satisfied in any model of PA0 , so in particular
in any model of 𝑇.
(L2) Arguing in an arbitrary model of PA, it follows from modus
ponens that PA ⊢ (□𝑇 𝜑 ∧ □𝑇 (𝜑 → 𝜓)) → □𝑇 𝜓, which implies (L2).
(L3) is a special case of (1) in Proposition 5.7.1 since □𝑇 𝜑 is a Σ1 -
sentence and 𝑇 ⊇ PA. □
Corollary 5.7.3. Let 𝜑 and 𝜓 be sentences. If 𝑇 ⊢ 𝜑 → 𝜓, then 𝑇 ⊢
□𝑇 𝜑 → □𝑇 𝜓.
Proof. Assume 𝑇 ⊢ 𝜑 → 𝜓. Then, by (L1), 𝑇 ⊢ □𝑇 (𝜑 → 𝜓). It follows

that 𝑇 ⊢ □𝑇 𝜑 → (□𝑇 𝜑 ∧ □𝑇 (𝜑 → 𝜓)). By (L2) one deduces that
𝑇 ⊢ □𝑇 𝜑 → □𝑇 𝜓. □
Theorem 5.7.4. Let 𝑇 be a recursive ℒ𝑎𝑟 -theory containing PA. Let 𝜑 be
a sentence. Then
𝑇 ⊢ □𝑇 (□𝑇 𝜑 → 𝜑) → □𝑇 𝜑.
Proof. There exists a formula 𝐹(𝑥) such that for any sentence 𝜃, 𝐹(#𝜃)
is equal to the sentence □𝑇 𝜃 → 𝜑. By Corollary 5.4.2, there exists a
sentence 𝜓 such that
(∗) 𝑇 ⊢ 𝜓 ↔ (□𝑇 𝜓 → 𝜑).
In particular,
𝑇 ⊢ 𝜓 → (□𝑇 𝜓 → 𝜑);
hence, by Corollary 5.7.3,

𝑇 ⊢ □𝑇 𝜓 → □𝑇 (□𝑇 𝜓 → 𝜑).
It follows from (L3) that
𝑇 ⊢ □𝑇 𝜓 → (□𝑇 □𝑇 𝜓 ∧ □𝑇 (□𝑇 𝜓 → 𝜑)).
Since by (L2)
𝑇 ⊢ (□𝑇 □𝑇 𝜓 ∧ □𝑇 (□𝑇 𝜓 → 𝜑)) → □𝑇 𝜑,
one deduces that
(∗∗) 𝑇 ⊢ □𝑇 𝜓 → □𝑇 𝜑.
From (∗∗) one gets that
𝑇 ⊢ (□𝑇 𝜑 → 𝜑) → (□𝑇 𝜓 → 𝜑),
thus by (∗) it follows that
𝑇 ⊢ (□𝑇 𝜑 → 𝜑) → 𝜓.
By Corollary 5.7.3 one infers that
𝑇 ⊢ □𝑇 (□𝑇 𝜑 → 𝜑) → □𝑇 𝜓,
and using (∗∗) one deduces that
𝑇 ⊢ □𝑇 (□𝑇 𝜑 → 𝜑) → □𝑇 𝜑,
which ends the proof. □
Corollary 5.7.5. If 𝑇 ⊢ □𝑇 𝜑 → 𝜑, then 𝑇 ⊢ □𝑇 𝜑.
Proof. Indeed, if 𝑇 ⊢ □𝑇 𝜑 → 𝜑, then it follows from (L1) that 𝑇 ⊢

□𝑇 (□𝑇 𝜑 → 𝜑), so we may conclude by Theorem 5.7.4. □
Since □𝑇 𝜑 → 𝜑 is logically equivalent to ¬□𝑇 𝜑 ∨ 𝜑, we obtain the

following corollary.
Corollary 5.7.6. If 𝑇 ⊢ ¬□𝑇 𝜑, then 𝑇 ⊢ □𝑇 𝜑. □
This corollary implies in particular the following famous result.

Theorem 5.7.7 (Gödel’s Second Incompleteness Theorem). Let 𝑇 be a
recursive ℒ𝑎𝑟 -theory containing PA. Then if 𝑇 is consistent, for any sen-
tence 𝜑, 𝑇 ⊬ ¬□𝑇 𝜑. In particular, let Con𝑇 = ¬□𝑇 (0 = 1), the sentence
expressing the consistency of 𝑇. Then 𝑇 ⊬ Con𝑇 . □
5.8. Exercises 141
5.8. Exercises
Exercise 5.8.1 (Presburger arithmetic). We consider the language
ℒPres = {0, +, <, 1, ≡𝑛 , 𝑛 ≥ 1}, where ≡𝑛 is a binary relation for all 𝑛.
Presburger arithmetic is the ℒPres -theory 𝑇Pres which is given by the fol-
lowing axioms:
• the axioms of ordered abelian groups (cf. Exercise 3.7.9);
• an axiom stating that 1 is the smallest positive element;
• for any 𝑛 ≥ 1, an axiom 𝜑𝑛 of the form
∀𝑥, 𝑦 (𝑥 ≡𝑛 𝑦 ↔ ∃𝑧 𝑥 + 𝑛𝑧 = 𝑦)
and an axiom 𝜓𝑛 of the form
𝑛−1
∀𝑥 𝑥 ≡𝑛 1⏟+
⎵⎵⋯ +⏟1 .
⏟⎵⎵
⋁
𝑖=0 𝑖 times
(1) Observe that 𝒵 = ⟨ℤ; 0, 1, +, <, ≡𝑛 ⟩ ⊧ 𝑇Pres , where all symbols

have their usual interpretation in the integers.
(2) Prove that 𝑇Pres eliminates quantifiers and is complete.
(3) Deduce that 𝑇Pres is decidable.
Exercise 5.8.2.
(1) Let Φ = {#𝜑 ∣ 𝜑 is a satisfiable ℒ𝑎𝑟 -sentence}. Prove that Φ is
not recursively enumerable.
(2) Let Φ𝑚 be the set of codes #𝜑 of ℒ𝑎𝑟 -sentences 𝜑 satisfiable
by an ℒ𝑎𝑟 -structure of domain {0, … , 𝑚 − 1}, with 𝑚 ≥ 1 an
integer. Prove that Φ𝑚 is primitive recursive.
(3) One considers the set Φfin of codes #𝜑 of ℒ𝑎𝑟 -sentences 𝜑 sat-
isfiable by a finite ℒ𝑎𝑟 -structure. Using the previous question
and an appropriate coding, prove that Φfin is recursively enu-
merable.
Exercise 5.8.3. Let ℒ = {𝑃, 𝑐}, with 𝑃 a unary predicate and 𝑐 a constant.
(1) Determine all countable ℒ-structures up to isomorphism.
(2) Deduce that two ℒ-structures 𝔐 and 𝔑 are elementarily equiv-
alent precisely if the following two conditions hold:
• 𝔐 ⊧ 𝑃𝑐 if and only if 𝔑 ⊧ 𝑃𝑐, and

• 𝔐 ⊧ ∃≥𝑘 𝑥 𝑄𝑥 if and only if 𝔑 ⊧ ∃≥𝑘 𝑥 𝑄𝑥, for any 𝑘 ∈ ℕ
and any 𝑄 ∈ {𝑃, ¬𝑃}.
(3) Prove that an ℒ-sentence 𝜑 is universally valid if and only if
𝔐 ⊧ 𝜑 for any finite ℒ-structure. Deduce that the empty ℒ-
theory is decidable.
Exercise 5.8.4. The aim of this exercise is to prove that there is a to-
tal recursive function which is not provably total Σ1 (see p. 135 for the
definition).
(1) Prove that there is a partial recursive function ℎ ∈ ℱ2∗ with the
following properties:
• if 𝑎 = #𝜑 for a Σ1 -formula 𝜑(𝑣0 , 𝑣1 ) and if 𝑛 ∈ ℕ is such
that there exists 𝑚 ∈ ℕ with PA ⊢ 𝜑(𝑛, 𝑚), then PA ⊢
𝜑 (𝑛, ℎ(𝑎, 𝑛));
• if 𝑎 = #𝜑 for a Σ1 -formula 𝜑(𝑣0 , 𝑣1 ) and if 𝑛 ∈ ℕ is such
that there does not exist 𝑚 ∈ ℕ with PA ⊢ 𝜑(𝑛, 𝑚), then
(𝑎, 𝑛) ∉ dom(ℎ);
• otherwise, ℎ(𝑎, 𝑛) = 0.
(2) Choose ℎ ∈ ℱ2∗ as above, and define 𝑔 ∈ ℱ3 as follows:
• if 𝑎 = #𝜑 for a Σ1 -formula 𝜑(𝑣0 , 𝑣1 ) and if 𝑏 = ##𝑑 for a
formal proof 𝑑 of ∀𝑣0 ∃! 𝑣1 𝜑(𝑣0 , 𝑣1 ) in PA, then 𝑔(𝑎, 𝑏, 𝑛) =
ℎ(𝑎, 𝑛);
• otherwise 𝑔(𝑎, 𝑏, 𝑛) = 0.
Prove that 𝑔 is total recursive, and that it is universal prov-
ably total Σ1 in the following sense: a function 𝑓 ∈ ℱ1 is prov-
ably total Σ1 if and only if there are 𝑎, 𝑏 ∈ ℕ such that 𝑓 =
𝜆𝑛.𝑔(𝑎, 𝑏, 𝑛).
(3) Conclude.
Exercise 5.8.5 (End Extensions in Peano Arithmetic). In this exercise,
we shall use the Omitting Types Theorem as proved in Exercise 2.7.8.
The aim is to prove the following result, due to MacDowell and Specker:
Let 𝔐 be a countable model of PA. Then there exists a proper elementary
extension 𝔑 ≽ 𝔐 such that 𝔑 is an end extension of 𝔐: for any 𝑚 ∈ 𝑀
and every 𝑛 ∈ 𝑁 ⧵ 𝑀 we have 𝔑 ⊧ 𝑚 < 𝑛.
5.8. Exercises 143
(1) Let 𝔐 ⊧ PA. Prove that the pigeonhole principle holds in 𝔐:

for any ℒ𝑎𝑟 (𝑀)-formula 𝜃(𝑣, 𝑧) and any 𝑎 ∈ 𝑀 we have
𝔐 ⊧ [∀𝑥(∃𝑧 > 𝑥)(∃𝑣 < 𝑎)𝜃(𝑣, 𝑧)] → (∃𝑣 < 𝑎)∀𝑥(∃𝑧 > 𝑥)𝜃(𝑣, 𝑧).
(2) Let 𝔐 ⊧ PA. Let 𝑐 be a constant symbol which is not in ℒ𝑎𝑟 (𝑀),
and set ℒ = ℒ𝑎𝑟 (𝑀) ∪ {𝑐}. We now consider the ℒ-theory 𝑇 ∶=
𝐷(𝔐) ∪ {𝑐 > 𝑚 ∣ 𝑚 ∈ 𝑀}, where 𝐷(𝔐) is the complete diagram
of 𝔐 (cf. Section 3.2).
(a) Check that 𝑇 is consistent.
(b) Let 𝑎 ∈ 𝑀 and let 𝜃(𝑣, 𝑧) be an ℒ-formula such that 𝑇 ⊢
∀𝑣(𝜃(𝑣, 𝑐) → 𝑣 < 𝑎) and such that 𝑇 ∪ {∃𝑣𝜃(𝑣, 𝑐)} is con-
sistent.
Prove that there exists 𝑚 ∈ 𝑀 with 𝑚 < 𝑎 and such that
𝔐 ⊧ ∀𝑥(∃𝑧 > 𝑥)𝜃(𝑚, 𝑧).
(c) Let 𝑎 ∈ 𝑀 be a non-standard element. Consider the set of
formulas
𝜋𝑎 (𝑣) ∶= {𝑣 < 𝑎} ∪ {𝑣 ≠ 𝑚 ∣ 𝑚 ∈ 𝑀}.
Prove that 𝜋𝑎 is a non-isolated partial 1-type in 𝑇.

(3) Conclude.
Exercise 5.8.6 (Tenenbaum’s Theorem). Let 𝔐 be a non-standard model

of PA and let 𝜂(𝑥, 𝑦) be an ℒ𝑎𝑟 -formula with two free variables. Denote
by 𝑆𝜂 (𝔐) the set of 𝐴 ⊆ ℕ such that there is 𝑎 ∈ 𝑀 with
𝐴 = {𝑛 ∈ ℕ ∣ 𝔐 ⊧ 𝜂(𝑛, 𝑎)}.
Denote by 𝑆(𝔐) the union of all 𝑆𝜂 (𝔐), where 𝜂 runs over the set of
ℒ𝑎𝑟 -formulas with two free variables.
(1) Let 𝜂0 (𝑥, 𝑦) be an ℒ𝑎𝑟 -formula such that for any pair of disjoint
finite subsets 𝐴 and 𝐵 of ℕ, the sentence
∃𝑥 ( 𝜂0 (𝑖, 𝑥) ∧ ¬𝜂 (𝑗, 𝑥))

⋀ ⋀ 0
𝑖∈𝐴 𝑗∈𝐵
is provable in PA. Prove that 𝑆𝜂0 (𝔐) = 𝑆(𝔐).

[Hint: Use overspill in 𝔐.]
(2) Prove that there is a Σ1 -formula 𝜂0 with two free variables such
that for all 𝑛 ∈ ℕ the sentence
𝜂0 (𝑛, 𝑥) ↔ ∃𝑦(𝜋(𝑛) ⋅ 𝑦 = 𝑥)
is provable in PA. Here, 𝜋(𝑛) denotes the (𝑛 + 1)th prime num-
ber. Prove that 𝑆𝜂0 (𝔐) = 𝑆(𝔐).
(3) Let 𝐴 and 𝐵 be two disjoint recursively enumerable subsets of
ℕ.
(a) The set of Δ0 -formulas is defined as the smallest set of
ℒ𝑎𝑟 -formulas containing the atomic formulas and which
is stable under ∧, ¬ and under bounded quantification
(∃𝑥 < 𝑡) and (∀𝑥 < 𝑡), with 𝑡 a term not depending on
the variable 𝑥.
Observe that there are Δ0 -formulas 𝛼(𝑥, 𝑦) and 𝛽(𝑥, 𝑦)
such that in 𝔑st , 𝐴 is defined by ∃𝑦𝛼(𝑥, 𝑦) and 𝐵 by
∃𝑦𝛽(𝑥, 𝑦).
(b) Prove that for any 𝑘 ∈ ℕ,
𝔐 ⊧ (∀𝑥, 𝑦, 𝑧 < 𝑘) ¬(𝛼(𝑥, 𝑦) ∧ 𝛽(𝑥, 𝑧)),
and that there is 𝜁 ∈ 𝑀 non-standard such that
𝔐 ⊧ (∀𝑥, 𝑦, 𝑧 < 𝜁) ¬(𝛼(𝑥, 𝑦) ∧ 𝛽(𝑥, 𝑧)).
(c) Consider 𝐴 and 𝐵 as in Exercise 4.8.7 (that is, infinite and
recursively inseparable) to deduce that 𝑆(𝔐) contains a
non-recursive set.
(4) If 𝑀 is countable and ℎ ∶ ℕ → 𝑀 is a bijection, one may trans-
port the ℒ𝑎𝑟 -structure of 𝔐 via ℎ−1 on ℕ, defining 𝑥 +′ 𝑦 ∶=
ℎ−1 (ℎ(𝑥) + ℎ(𝑦)), 𝑥 ⋅′ 𝑦 ∶= ℎ−1 (ℎ(𝑥) ⋅ ℎ(𝑦)), etc.
We now assume that the structure 𝔐 is recursive, that is,
there exists a bijection ℎ as above such that +′ and ⋅′ are recur-
sive functions.
(a) For any fixed integer 𝑐 ∈ ℕ, prove that the function 𝑓 ∶
ℕ2 → ℕ given by 𝑓(𝑛, 𝑚) = 1 if 𝑚 ⏟⎵+ ′⋯+
⎵⎵⏟⎵ ′ 𝑚 = 𝑐 and
⎵⎵⏟
𝜋(𝑛) times
𝑓(𝑛, 𝑚) = 0 otherwise, is recursive.
(b) Deduce from this that 𝑆(𝔐) does only contain recursive
sets.
5.8. Exercises 145
(c) Deduce Tenenbaum’s Theorem:

There is no recursive non-standard model of PA.
10.1090/stml/089/06
Chapter 6
Axiomatic Set Theory
Introduction
In this final chapter, we will formalize set theory within the first-order
logic framework we have developed in former chapters. The language
of set theory ℒ∈ has only one non-logical symbol, the binary relation
symbol ∈. We start by discussing in detail the Zermelo-Fraenkel axioms
and the Axiom of Choice. In 6.3 we are finally in a position to prove the
equivalence of the Axiom of Choice, of Zorn’s Lemma and the existence
of well-orderings.
Section 6.4 is devoted to the Axiom of Foundation and its connection
with the von Neumann hierarchy. We are then able to prove some in-
dependence and relative consistency results in 6.5, like the relative con-
sistency of the Axiom of Foundation or the relative consistency of the
negation of the existence of inaccessible cardinals. In the final section
we discuss a few famous independence and relative consistency results
whose proofs are outside the scope of this book, like the independence
of the continuum hypothesis.
6.1. The Framework

We will work in the language ℒ∈ , and we will consider ℒ∈ -structures
𝒰 = ⟨𝑈; ∈⟩ (where 𝒰 will be called a universe of sets).
147
148 6. Axiomatic Set Theory
We call a set any element of 𝑈. The universe 𝑈 itself is a set in the

naive sense. The membership relation between two sets is given by the
relation symbol ∈.
In this way we may form the language ℒ∈,𝑈 (a set in the naive sense).
Definition. A naive subset 𝐷 of 𝑈 is a class if there is an ℒ∈,𝑈 -formula
𝜑(𝑥) such that 𝐷 = 𝜑[𝒰].
Note that any set 𝑎 defines a class 𝐶𝑎 , which is given by the formula
𝜑(𝑥) equal to 𝑥 ∈ 𝑎. By the Extensionality Axiom (which is not yet
defined!), two sets 𝑎 and 𝑏 are equal if and only if 𝐶𝑎 = 𝐶𝑏 .
By abuse of notation, we sometimes use ∈ to denote membership
(in the naive sense) in a class, writing 𝑐 ∈ 𝐷 instead of 𝒰 ⊧ 𝜑[𝑐], where
𝐷 = 𝜑[𝒰]. This should not create any confusion.
6.2. The Zermelo-Fraenkel Axioms

Definition.
(1) The Zermelo-Fraenkel axiom system, denoted by ZF, is given by
the following list of axioms: Extensionality, Comprehension
Scheme, Pairing, Union, Power Set, Replacement Scheme, In-
finity, Foundation.
(2) If one adds the the Axiom of Choice (AC) to ZF, one obtains
the axiom system ZFC.
Remark. Some authors do not include the Axiom of Foundation in the
axiom system ZF.
We will now discuss the axioms in detail.

Extensionality Axiom. ∀𝑥, 𝑦 (∀𝑧(𝑧 ∈ 𝑥 ↔ 𝑧 ∈ 𝑦) → 𝑥 = 𝑦).
It expresses that two sets having the same elements are equal.
Notation. We write 𝑥 ⊆ 𝑦 as an abbreviation for ∀𝑧(𝑧 ∈ 𝑥 → 𝑧 ∈ 𝑦).
The Extensionality Axiom is then equivalent to the following sen-

tence: ∀𝑥, 𝑦 (𝑥 ⊆ 𝑦 ∧ 𝑦 ⊆ 𝑥 → 𝑥 = 𝑦).
6.2. The Zermelo-Fraenkel Axioms 149
Comprehension Axiom Scheme. For any formula 𝜑(𝑣0 , … , 𝑣𝑛 ) in the

language ℒ∈ , one puts an axiom of the form
∀𝑣1 , … , 𝑣𝑛+1 ∃𝑣𝑛+2 ∀𝑣0 (𝑣0 ∈ 𝑣𝑛+2 ↔ (𝑣0 ∈ 𝑣𝑛+1 ∧ 𝜑(𝑣0 , … , 𝑣𝑛 ))) .
It expresses that if 𝑎 is a set and 𝐷 is a class, then the class {𝑏 ∈ 𝑎 ∣

𝑏 ∈ 𝐷} is given by a set. We will sometimes denote this set by 𝑎 ∩ 𝐷.
Here are some easy consequences of the axioms stated so far:
• If 𝑎 and 𝑏 are two sets, then one may form their intersection
𝑎 ∩ 𝑏 = {𝑐 ∈ 𝑎 ∣ 𝑐 ∈ 𝑏} and, in a similar way, their difference
𝑎 ⧵ 𝑏 = {𝑐 ∈ 𝑎 ∣ 𝑐 ∉ 𝑏}.
• One obtains the existence of the empty set, denoted by ∅. In-
deed, let 𝑎 be an arbitrary set. Then ∅ = {𝑐 ∈ 𝑎 ∣ ¬𝑐 = 𝑐}. The
uniqueness of ∅ follows from Extensionality.
• If 𝑎 ≠ ∅ is a set, one obtains the existence of ⋂𝑏∈𝑎 𝑏 (as a
set). To see this, choose an arbitrary 𝑏0 ∈ 𝑎, and observe that
⋂𝑏∈𝑎 𝑏 = {𝑐 ∈ 𝑏0 ∣ ∀𝑥(𝑥 ∈ 𝑎 → 𝑐 ∈ 𝑥)}.
Remark 6.2.1. Comprehension implies that there is no set containing
all sets. Indeed, suppose for contradiction that 𝒰 ⊧ ∀𝑧 𝑧 ∈ 𝑎 for some
𝑎 ∈ 𝑈. By Comprehension there is a set 𝑏 such that 𝒰 ⊧ ∀ 𝑧(𝑧 ∈ 𝑏 ↔
𝑧 ∉ 𝑧), so in particular 𝑏 ∈ 𝑏 if and only if 𝑏 ∉ 𝑏. This contradiction is
Russell’s Antinomy, and it shows that 𝑎 cannot exist.
Pairing Axiom. ∀𝑦1 , 𝑦2 ∃𝑥∀𝑧 (𝑧 ∈ 𝑥 ↔ (𝑧 = 𝑦1 ∨ 𝑧 = 𝑦2 )).
It expresses that if 𝑎 and 𝑏 are two sets, then {𝑎, 𝑏} is a set.

Definition. The ordered pair (also called Kuratowski pair) of two sets 𝑎
and 𝑏 is the set (𝑎, 𝑏) ∶= {{𝑎}, {𝑎, 𝑏}}.
Lemma 6.2.2. With the axioms stated so far, we have:
(1) The collection of ordered pairs forms a class. Moreover, there is
an ℒ∈ -formula 𝜑(𝑥, 𝑦) such that 𝒰 ⊧ 𝜑[𝑎, 𝑏] if and only if 𝑎 is
an ordered pair of the form 𝑎 = (𝑏, 𝑐). The analogous fact holds
for the second coordinate.
(2) One has (𝑏, 𝑐) = (𝑏′ , 𝑐′ ) if and only if 𝑏 = 𝑏′ and 𝑐 = 𝑐′ .
Union Axiom. ∀𝑦∃𝑥∀𝑧 (𝑧 ∈ 𝑥 ↔ ∃𝑤(𝑧 ∈ 𝑤 ∧ 𝑤 ∈ 𝑦)).
It expresses that for every set 𝑎, the following class is given by a set:
⋃ 𝑎 ∶= {𝑧 ∣ ∃𝑤(𝑧 ∈ 𝑤 ∧ 𝑤 ∈ 𝑎)}.
Note that, by combining the Pairing Axiom and the Union Axiom,
one gets that the union 𝑎 ∪ 𝑏 of two sets 𝑎 and 𝑏 is given by a set.
Power Set Axiom. ∀𝑦∃𝑥∀𝑧(𝑧 ∈ 𝑥 ↔ 𝑧 ⊆ 𝑦).
It postulates the existence of the set of subsets of a set, that is, for
any set 𝑎 the class 𝒫(𝑎) = {𝑏 ∣ 𝑏 ⊆ 𝑎} is given by a set.
Lemma 6.2.3. The axioms stated so far imply the existence of the carte-
sian product of two sets 𝑎 and 𝑏: 𝑎 × 𝑏 = {(𝑥, 𝑦) ∣ 𝑥 ∈ 𝑎 ∧ 𝑦 ∈ 𝑏} is a
set.
Proof. If 𝑥 ∈ 𝑎 and 𝑦 ∈ 𝑏, then one has {𝑥}, {𝑥, 𝑦} ∈ 𝒫(𝑎 ∪ 𝑏), whence
(𝑥, 𝑦) ∈ 𝒫(𝒫(𝑎 ∪ 𝑏)). One may conclude by Comprehension, using
Lemma 6.2.2. □
One may also define triples (𝑥, 𝑦, 𝑧) ∶= ((𝑥, 𝑦), 𝑧) and more gener-
ally 𝑛-tuples, via (𝑥1 , … , 𝑥𝑛+1 ) ∶= ((𝑥1 , … , 𝑥𝑛 ), 𝑥𝑛+1 ), inductively. One
obtains 𝑎 × 𝑏 × 𝑐, and more generally 𝑎1 × ⋯ × 𝑎𝑛 .
Definition.
• A (binary) relation 𝑅 is a set of ordered pairs. One sets
dom(𝑅) ∶= {𝑥 ∣ ∃𝑦(𝑥, 𝑦) ∈ 𝑅}, called the domain of 𝑅, and
im(𝑅) ∶= {𝑦 ∣ ∃𝑥(𝑥, 𝑦) ∈ 𝑅}, called the image of 𝑅.
• A function 𝑓 is a relation which is unique on the right, that is,
which satisfies ∀𝑥, 𝑦, 𝑧 ((𝑥, 𝑦) ∈ 𝑓 ∧ (𝑥, 𝑧) ∈ 𝑓 → 𝑦 = 𝑧).
Remark.
• Note that if 𝑅 is a relation, then dom(𝑅) and im(𝑅) are sets.
Indeed, for (𝑥, 𝑦) ∈ 𝑅 one has 𝑥, 𝑦 ∈ ⋃ (⋃ 𝑅).
• By definition, a function is identified with its graph.
Notation. When 𝑓 is a function, we usually write 𝑓(𝑥) = 𝑦 instead of
(𝑥, 𝑦) ∈ 𝑓. Sometimes, for 𝑥 ∉ dom(𝑓), we set 𝑓(𝑥) ∶= ∅. If 𝑎 = dom(𝑓)
and im(𝑓) ⊆ 𝑏, we write 𝑓 ∶ 𝑎 → 𝑏.
Lemma 6.2.4. Let 𝑎 and 𝑏 be two sets. With the axioms we have stated so
far, one then gets the following:
(1) {𝑅 ∣ 𝑅 is a relation with dom(𝑅) ⊆ 𝑎 and im(𝑅) ⊆ 𝑏} is a set.
(2) {𝑓 ∣ 𝑓 ∶ 𝑎 → 𝑏} is a set.
(3) The collection of functions forms a class. We denote by Fn(𝑥) an
ℒ∈ -formula which defines this class.
A family of sets, indexed by the set 𝐼, is a function 𝑓 with dom(𝑓) = 𝐼.

One usually writes (𝑎𝑖 )𝑖∈𝐼 for such a family, where 𝑎𝑖 = 𝑓(𝑖).
Remark. If (𝑎𝑖 )𝑖∈𝐼 is a family of sets with non-empty index set 𝐼, then
the class ∏𝑖∈𝐼 𝑎𝑖 ∶= {𝑔 ∶ 𝐼 → ⋃𝑖∈𝐼 𝑎𝑖 ∣ ∀𝑧(𝑧 ∈ 𝐼 → 𝑔(𝑧) ∈ 𝑧)} is given
by a set. □
Definition. We say that 𝐹 ⊆ 𝑈 2 is a functional class if there is an ℒ∈,𝑈 -
formula 𝜑(𝑥, 𝑦) which defines 𝐹 such that
𝒰 ⊧ ∀𝑥, 𝑦1 , 𝑦2 (𝜑(𝑥, 𝑦1 ) ∧ 𝜑(𝑥, 𝑦2 ) → 𝑦1 = 𝑦2 ) .
The class Dom(𝐹) defined by ∃𝑦𝜑 is called the Domain of 𝐹, and the class
Im(𝐹) defined by ∃𝑥𝜑 is called the Image of 𝐹.
Note that a functional class gives rise to a function in the naive sense,
with domain Dom(𝐹). One may also consider functional classes which
correspond to naive functions between a class 𝐷 ⊆ 𝑈 𝑛 and 𝑈 that are
definable in ℒ∈,𝑈 .
Replacement Axiom Scheme. For each formula 𝜑(𝑥, 𝑦, 𝑣1 , … , 𝑣𝑛 ) in
the language ℒ∈ , one puts an axiom of the form
∀𝑣0 , 𝑣1 , … , 𝑣𝑛 [∀𝑥, 𝑦1 , 𝑦2 (𝜑(𝑥, 𝑦1 , 𝑣) ∧ 𝜑(𝑥, 𝑦2 , 𝑣) → 𝑦1 = 𝑦2 )

→ ∃𝑣𝑛+1 ∀𝑦(𝑦 ∈ 𝑣𝑛+1 ↔ ∃𝑥(𝑥 ∈ 𝑣0 ∧ 𝜑(𝑥, 𝑦, 𝑣)))].
It expresses that if 𝐹 ⊆ 𝑈 2 is a functional class and 𝑎 is a set, then

𝐹[𝑎] ∶= {𝑧 ∣ ∃𝑢(𝑢 ∈ 𝑎 ∧ (𝑢, 𝑧) ∈ 𝐹)} is a set. (It is obtained by ‘replacing’
every element of 𝑎 by its image under 𝐹.)
The axioms stated so far are not independent. Indeed:
Lemma 6.2.5.
(1) The Replacement Axiom Scheme implies the Comprehension Ax-

iom Scheme.
(2) The Pairing Axiom is a consequence of the other axioms stated so
far.
Proof. Let 𝜑(𝑥) be given. Let 𝐹(𝑥, 𝑦) be the formula (𝑥 = 𝑦 ∧ 𝜑(𝑥)).

Then {𝑥 ∈ 𝑎 ∣ 𝜑(𝑥)} = 𝐹[𝑎], proving (1).
To prove (2), suppose that 𝑎 and 𝑏 are two sets. Observe that ∅ ∈
𝒫(∅) = {∅}, so in particular ∅ ≠ 𝒫(∅). Let 𝐹(𝑥, 𝑦) be the formula ((𝑥 =
∅ ∧ 𝑦 = 𝑎) ∨ (𝑥 = 𝒫(∅) ∧ 𝑦 = 𝑏)). This is a functional class, and for
𝑐 = 𝒫(𝒫(∅)) = {∅, 𝒫(∅)}, one gets 𝐹[𝑐] = {𝑎, 𝑏}. □
Remark. We may work in an expansion by definition and use for ex-

ample relation symbols 𝑥 ⊆ 𝑦 (binary) and 𝑓 ∶ 𝑥 → 𝑦 (ternary), func-
tion symbols 𝑥 ∪ 𝑦, 𝑥 ∩ 𝑦, 𝑥 ⧵ 𝑦, ⋃ 𝑦, 𝒫(𝑦), {𝑥, 𝑦}, (𝑥, 𝑦), 𝑥 × 𝑦, dom(𝑅),
im(𝑅), the constant symbol ∅, etc., in the Comprehension and Replace-
ment schemes. By Proposition 3.3.2, this does not change anything.
Recall that a set 𝑥 is an ordinal if it is a transitive set such that ∈↾𝑥×𝑥

defines a well-order on 𝑥, that is, a total order which is well-founded.
(Well-foundedness may be expressed by the following formula: ∀𝑦(𝑦 ∈
𝒫(𝑥) ∧ ¬𝑦 = ∅ → ∃𝑧(𝑧 ∈ 𝑦 ∧ 𝑧 ∩ 𝑦 = ∅)).) Let Ord(𝑥) be an ℒ∈ -formula
which defines the class of ordinals (in all models of the axioms stated so
far). We then get a formula
Card(𝑥) ∶= Ord(𝑥) ∧ ∀𝑦(𝑦 ∈ 𝑥 → ‘there is no surjective’ 𝑓 ∶ 𝑦 → 𝑥)
which defines the class of cardinals.
Notation. In what follows, we will use 𝛼, 𝛽, 𝛾, etc., to denote only ordi-

nals. For instance, we will write ∀𝛾𝜑 instead of ∀𝛾(Ord(𝛾) → 𝜑).
Axiom of Infinity (AI). ∃𝑥(∅ ∈ 𝑥 ∧ ∀𝑧(𝑧 ∈ 𝑥 → 𝑧 ∪ {𝑧} ∈ 𝑥)).
It postulates the existence of a set which contains ∅ and is stable

under the ‘successor’ function.
Let 𝑎 be a set as postulated by (AI). Applying Comprehension if nec-

essary, we may assume that 𝑎 is a set of ordinals. Let 𝛼 be the least or-
dinal such that 𝛼 ∉ 𝑎. By construction, ∅ ∈ 𝛼 and 𝛼 is stable under the
successor function, so 𝛼 is a limit ordinal.
We let Lim(𝑥) be a formula which defines the class of all limit ordi-
nals. By what we have just seen, limit ordinals exist. One denotes by 𝜔
the smallest limit ordinal.
Axiom of Foundation (AF). ∀𝑥(¬𝑥 = ∅ → ∃𝑧(𝑧 ∈ 𝑥 ∧ 𝑧 ∩ 𝑥 = ∅)).
Observe that (AF), together with the Pairing Axiom, implies that no
set 𝑎 contains itself as an element, since otherwise {𝑎} would contradict
(AF).
Notation. We denote by ZF− the axioms of ZF without (AF), and by
−
ZFC the axioms of ZFC without (AF).
Remark 6.2.6.
(1) In any model of ZF− , the ordinals satisfy the same properties
as we have seen in Chapter 1. Similarly for the cardinals in
−
models of ZFC .
(2) The class of ordinals is not a set. Indeed, if Ord were given by
the set 𝑎, then 𝑎 would be transitive and well-ordered by ∈,
so an ordinal and thus 𝑎 ∈ 𝑎. But 𝑎 ∉ 𝑎 for all ordinals by
Proposition 1.6.1.
Similarly, Card is not a set. Indeed, if it were given by the
set 𝑎, then Ord would be given by the set ⋃ 𝑎.
Lemma 6.2.7 (Transfinite induction). Let 𝒰 ⊧ ZF− , and let 𝜑(𝑥) be an
ℒ∈,𝑈 -formula. Then 𝒰 satisfies the following induction property:
𝜑(0) ∧ ∀𝛾(𝜑(𝛾) → 𝜑(𝛾 + 1))

∧ ∀𝛾[(Lim(𝛾) ∧ ∀𝛿(𝛿 < 𝛾 → 𝜑(𝛿))) → 𝜑(𝛾)] → ∀𝛾𝜑(𝛾).
Proof. If 𝒰 satisfies
𝜑(0) ∧ ∀𝛾(𝜑(𝛾) → 𝜑(𝛾 + 1)) ∧ ∀𝛾[(Lim(𝛾) ∧ ∀𝛿(𝛿 < 𝛾 → 𝜑(𝛿))) → 𝜑(𝛾)]
and 𝒰 ⊧ ¬𝜑[𝛼] for some ordinal 𝛼, it is enough to choose a minimal such
𝛼 to get a contradiction. □
Theorem 6.2.8 (Definition by transfinite recursion). Let 𝒰 ⊧ ZF− , and

let 𝐺 be a functional class with Dom(𝐺) = 𝑈 𝑛+1 . Then there is a unique
functional class 𝐹 with Dom(𝐹) = 𝑈 𝑛 × Ord such that one has 𝐹(𝑤, 𝛼) =
𝐺(𝑤, 𝐹 ↾{𝑤}×𝛼 ) for every tuple of sets 𝑤 and every ordinal 𝛼.
Proof. The key idea is to approximate 𝐹 by functions. We prove first

that for every 𝑤 and every ordinal 𝛽 there is a unique function 𝑓𝑤,𝛽 with
domain 𝛽 such that 𝑓𝑤,𝛽 (𝛼) = 𝐺(𝑤, {𝑤} × 𝑓𝑤,𝛽 ↾𝛼 ) for all 𝛼 < 𝛽. The
uniqueness is clear by transfinite induction.
Fix 𝑤. To establish existence, let us prove by induction on 𝛽 that
there is a function 𝑓𝑤,𝛽 with domain 𝛽 such that
(∗)𝑤,𝛽 ∀𝛼 (𝛼 < 𝛽 → 𝑓𝑤,𝛽 (𝛼) = 𝐺(𝑤, {𝑤} × 𝑓𝑤,𝛽 ↾𝛼 )) .
If 𝛽 = 0, set 𝑓𝑤,𝛽 = ∅.
If 𝛽 = 𝛽 ′ + 1 and 𝑓𝑤,𝛽′ is a function which satisfies (∗)𝑤,𝛽′ , one may
simply set 𝑓𝑤,𝛽 = 𝑓𝑤,𝛽′ ∪ {(𝛽 ′ , 𝐺(𝑤, {𝑤} × 𝑓𝑤,𝛽′ ))}.
Finally, if 𝛽 is a limit ordinal, we consider the set 𝑋𝑤,𝛽 of all functions
𝑓 such that there exists 𝛽′ < 𝛽 with dom(𝑓) = 𝛽 ′ and 𝑓 satisfies (∗)𝑤,𝛽′ .
By Replacement and uniqueness, 𝑋𝑤,𝛽 is indeed a set. Then, again by
uniqueness, ⋃𝑓′ ∈𝑋 𝑓′ is a function, and it clearly satisfies (∗)𝑤,𝛽 .
𝑤,𝛽
We may now set (𝑤, 𝛽, 𝑦) ∈ 𝐹 ∶⇔ for any function 𝑓 with domain

𝛽 + 1 satisfying (∗)𝑤,𝛽+1 one has 𝑓(𝛽) = 𝑦. □
Example 6.2.9 (Applications of transfinite recursion).
(1) The operations of ordinal arithmetic (ordinal addition, mul-
tiplication and exponentiation) may be defined by transfinite
recursion (see Remark 1.6.5).
(2) The ℵ-hierarchy of infinite cardinals is given by a functional
class ℵ ∶ Ord → Card. Indeed, it is sufficient to apply Theo-
rem 6.2.8 to the functional class 𝐺 ∶ 𝑈 → 𝑈 defined as follows:
• 𝐺(0) = 𝜔.
• If 𝑓 ∶ 𝛼 → 𝛽 for two ordinals 𝛼, 𝛽, then 𝐺(𝑓) equals the
least cardinal 𝜅 such that 𝜅 > 𝑓(𝛼′ ) for every 𝛼′ < 𝛼.
• Finally, if 𝑥 is not a function between two ordinals, then
𝐺(𝑥) = ∅.
6.3. The Axiom of Choice 155
(3) The von Neumann hierarchy. By transfinite recursion, one de-

fines a functional class 𝛼 ↦ 𝑉𝛼 as follows:
• 𝑉0 = ∅.
• 𝑉𝛼+1 = 𝒫(𝑉𝛼 ).
• 𝑉𝜆 = ⋃𝛼<𝜆 𝑉𝛼 for 𝜆 a limit ordinal.
Proposition 6.2.10 (ZF− ). The ordinal 𝜔 endowed with the ordinal op-
erations (addition, multiplication and the successor function), with 0 = ∅
and < given by ∈, is a model of PA.
6.3. The Axiom of Choice

Axiom of Choice (AC).
∀𝑓[(Fn(𝑓) ∧ ∅ ∉ im(𝑓)) → ∃𝑔 (Fn(𝑔) ∧ dom(𝑔) = dom(𝑓)

∧∀𝑥(𝑥 ∈ dom(𝑔) → 𝑔(𝑥) ∈ 𝑓(𝑥)))].
It expresses that the product of a family of non-empty sets is non-

empty.
Definition. Let 𝑎 be a set. A choice function on 𝑎 is a function ℎ ∶
𝒫(𝑎)′ ∶= 𝒫(𝑎) ⧵ {∅} → 𝑎 such that ℎ(𝐴) ∈ 𝐴 for all 𝐴 ∈ 𝒫(𝑎)′ .
Proposition 6.3.1. (AC) is equivalent to the existence of a choice function
on every set 𝑎.
Proof. Suppose (AC). For any set 𝑎, we then have ∏𝐴∈𝒫(𝑎)′ 𝐴 ≠ ∅. Any
element of this product is a choice function on 𝑎.
Conversely, let (𝑎𝑖 )𝑖∈𝐼 be a family of sets with 𝑎𝑖 ≠ ∅ for all 𝑖 ∈ 𝐼.
Let 𝑎 = ⋃𝑖∈𝐼 𝑎𝑖 and ℎ ∶ 𝒫(𝑎)′ → 𝑎 be a choice function on 𝑎. Then
ℎ ∘ 𝑓 ∈ ∏𝑖∈𝐼 𝑎𝑖 , where 𝑓 ∶ 𝐼 → 𝒫(𝑎)′ , 𝑓(𝑖) ∶= 𝑎𝑖 . □
Theorem 6.3.2. The following are equivalent in the theory ZF− :
(1) (AC).
(2) Zorn’s Lemma.
(3) Zermelo’s Theorem (Wohlordnungssatz).
Proof. (1)⇒(2): By way of contradiction, let us suppose that ⟨𝑋, <⟩ is an

inductive partial order without maximal element. We consider the set
𝒯 = {𝑇 ∈ 𝒫(𝑋) ∣ 𝑇 is totally ordered by <}.
As 𝑋 is inductive and there is no maximal element in 𝑋, for any 𝑇 ∈ 𝒯,

the set 𝐵(𝑇) ∶= {𝑥 ∈ 𝑋 ∣ 𝑥 > 𝑡 ∀𝑡 ∈ 𝑇} is non-empty. By (AC), there is
a function 𝑏 ∶ 𝒯 → 𝑋 such that 𝑏(𝑇) ∈ 𝐵(𝑇) for any 𝑇 ∈ 𝒯.
Let 𝐺 be the following functional class:
• if 𝑔 is a function with dom(𝑔) ∈ Ord and im(𝑔) ∈ 𝒯, then
𝐺(𝑔) = 𝑏(im(𝑔));
• 𝐺(𝑔) = ∅, otherwise.
Applying Theorem 6.2.8, we obtain a functional class 𝐹 with
Dom(𝐹) = Ord such that 𝐹(𝛼) = 𝐺(𝐹 ↾ 𝛼 ) for every ordinal 𝛼. One
shows easily by induction that 𝐹(𝛼) ∈ 𝑋 for every 𝛼 and that
𝛼 < 𝛽 ⇒ 𝐹(𝛼) < 𝐹(𝛽).
In particular, 𝐹 is injective as a naive function. The class 𝐹 −1 , defined

by (𝑥, 𝛼) ∈ 𝐹 −1 ∶⇔ (𝛼, 𝑥) ∈ 𝐹, is thus a functional class. We have
𝐹 −1 [𝑋] = Ord, so Ord is a set by the Replacement Axiom Scheme. This
contradicts Remark 6.2.6(2).
(2)⇒(3): Let 𝑎 be a set. We have to show that 𝑎 admits a well-order.
Consider 𝑋 ∶= {(𝑏, 𝑅) ∣ 𝑏 ∈ 𝒫(𝑎) and 𝑅 is a well-order on 𝑏}. It is easy
to see that 𝑋 is a set (exercise).
On 𝑋, we define a partial order as follows:
(𝑏, 𝑅) ≤ (𝑏′ , 𝑅′ ) ∶⇔ 𝑏 is an initial segment of (𝑏′ , 𝑅′ ) and 𝑅′ ↾ 𝑏 = 𝑅.
This partial order is inductive. Indeed, if (𝑏𝑖 , 𝑅𝑖 )𝑖∈𝐼 is a totally ordered

subset of 𝑋, then (𝑏, 𝑅) ∶= (⋃𝑖∈𝐼 𝑏𝑖 , ⋃𝑖∈𝐼 𝑅𝑖 ) is an upper bound of this
subset. By Zorn’s Lemma, there is an element (𝑏,̃ 𝑅)̃ ∈ 𝑋 which is max-
imal for ≤. If 𝑏 ̃ ≠ 𝑎, there exists 𝑦 ∈ 𝑎 ⧵ 𝑏.̃ We set 𝑏′ = 𝑏 ̃ ∪ {𝑦} and
̃ It is then clear that 𝑅′ is a well-order on 𝑏′
𝑅′ ∶= 𝑅̃ ∪ {(𝑥, 𝑦) ∣ 𝑥 ∈ 𝑏}.
which prolongates the one on 𝑏,̃ contradicting the maximality of (𝑏,̃ 𝑅). ̃
(3)⇒(1): Let 𝑎 be a set. By hypothesis, 𝑎 may be well-ordered, say
by <. The function 𝑓 ∶ 𝒫(𝑎)′ → 𝑎 which associates to any non-empty
6.4. The von Neumann Hierarchy 157
subset of 𝑎 its smallest element is a choice function on 𝑎. We conclude

by Proposition 6.3.1. □
Remark 6.3.3 (Skolem’s Paradox). If ZFC is consistent, it has a (neces-
sarily infinite) model. Since ℒ∈ is countable, ZFC thus admits a count-
able model 𝔐 by the Downward Löwenheim-Skolem Theorem. But there
exist uncountable sets in 𝔐, for example ℵ1 or ℝ.
Explanation: The notion of ‘countability’ depends on the model of ZFC
one works in, so it is relative. Being countable in the sense of 𝔐 and
being countable in the sense of the (naive) ground model of set theory
in the metatheory is not the same thing. Moreover, the base set 𝑀 of 𝔐
is a set from the naive point of view, and a proper class from the point of
view of 𝔐.
More generally, the notion of cardinality depends on the model. For
instance, the set of natural numbers ℕ and the set of real numbers ℝ
(both taken in the sense of 𝔐) are both countable from the point of view
of the underlying naive ground model. But the bijection between these
two sets which exists as a naive set is not even represented by a functional
class in 𝔐. Indeed, otherwise, by Replacement, it would be given by a
function in the sense of 𝔐.
6.4. The von Neumann Hierarchy and the Axiom of

Foundation
In this section, we will work in 𝒰 ⊧ ZF− .
Definition. Let 𝑎 be a set. By recursion on 𝑛 ∈ 𝜔, one defines 𝑎0 ∶=
𝑎, 𝑎𝑛+1 ∶= 𝑎𝑛 ∪ ⋃ 𝑎𝑛 . The set tcl(𝑎) ∶= ⋃𝑛∈𝜔 𝑎𝑛 is then called the
transitive closure of 𝑎.
Lemma 6.4.1 (ZF− ). The set tcl(𝑎) is the smallest transitive set containing
𝑎 as a subset. More precisely, the following properties hold:
(1) 𝑎 ⊆ tcl(𝑎).
(2) tcl(𝑎) is a transitive set.
(3) If 𝑎 ⊆ 𝑡 with 𝑡 a transitive set, then tcl(𝑎) ⊆ 𝑡. In particular,
𝑎 ⊆ 𝑏 ⇒ tcl(𝑎) ⊆ tcl(𝑏).
(4) If 𝑎 is transitive, then tcl(𝑎) = 𝑎.
(5) If 𝑏 ∈ 𝑎, then tcl(𝑏) ⊆ tcl(𝑎).

(6) tcl(𝑎) = 𝑎 ∪ ⋃𝑏∈𝑎 tcl(𝑏).
Proof. (1) and (2) are clear. To prove (3), one shows by induction that
𝑎𝑛 ⊆ 𝑡 for all 𝑛 ∈ 𝜔. (4) follows from (1-3).
(5) One has 𝑏 ∈ 𝑎 ⇒ 𝑏 ∈ tcl(𝑎) ⇒ 𝑏 ⊆ tcl(𝑎) ⇒ tcl(𝑏) ⊆ tcl(𝑎),
where the first implication follows from (1), the second from (2), and
the last one from (3).
(6) One has tcl(𝑎) ⊇ 𝑎 ∪ ⋃𝑏∈𝑎 tcl(𝑏) by (1) and (5). To prove the
other inclusion, it suffices by (3) to prove that 𝑎∪⋃𝑏∈𝑎 tcl(𝑏) is transitive,
which is clear. □
Recall the definition of the von Neumann hierarchy: 𝑉0 ∶= ∅, 𝑉𝛼+1 ∶=

𝒫(𝑉𝛼 ), 𝑉𝜆 ∶= ⋃𝛼<𝜆 𝑉𝛼 (for 𝜆 a limit ordinal).
We define the class 𝑉 by the formula ∃𝛼 𝑥 ∈ 𝑉𝛼 , which we will de-
note by 𝑉(𝑥). Informally, one thus has ‘𝑉 = ⋃𝛼∈Ord 𝑉𝛼 ’.
Definition. If 𝑎 is a set, the rank of 𝑎 is defined as
the least 𝛾 such that 𝑎 ∈ 𝑉𝛾+1 , if such a 𝛾 exists;

rk(𝑎) ∶= {
∞, otherwise.
Lemma 6.4.2.
(1) 𝑉𝛼 is a transitive set for any ordinal 𝛼.
(2) 𝛽 ≤ 𝛼 ⇒ 𝑉𝛽 ⊆ 𝑉𝛼 .
(3) 𝑉𝛼 = {𝑥 ∈ 𝑉 ∣ rk(𝑥) < 𝛼}.
(4) If 𝑥 ∈ 𝑉 and 𝑦 ∈ 𝑥, then 𝑦 ∈ 𝑉 and rk(𝑦) < rk(𝑥).
(5) If 𝑥 ∈ 𝑉, then rk(𝑥) = sup{rk(𝑦) + 1 ∣ 𝑦 ∈ 𝑥}.
(6) If 𝑥 ∈ 𝑉 is transitive, then {rk(𝑦) ∣ 𝑦 ∈ 𝑥} is an ordinal 𝛼.
(7) One has rk(𝛼) = 𝛼 for every 𝛼. In particular, 𝛼 ∈ 𝑉 and {𝛽 ∈
𝑉𝛼 ∣ 𝛽 ∈ Ord} = 𝛼.
(8) Let 𝑥 be a set. Then 𝑥 ∈ 𝑉 if and only if 𝑥 ⊆ 𝑉.
(9) Assume (AC). If 𝑥 ∈ 𝑉 is transitive, then 𝑥 ∈ 𝑉card(𝑥)+ .
6.4. The von Neumann Hierarchy 159
Proof. (1) & (2) By transfinite induction on 𝛼, one shows that 𝑉𝛼 is tran-
sitive and that 𝑉𝛽 ⊆ 𝑉𝛼 for all 𝛽 ≤ 𝛼. The cases 𝛼 = 0 and 𝛼 a limit ordinal
are clear. Now suppose 𝛼 = 𝛾+1. Since 𝑉𝛾 is a transitive set by the induc-
tion hypothesis, 𝒫(𝑉𝛾 ) = 𝑉𝛼 is transitive, too. If 𝛽 < 𝛼, then 𝛽 ≤ 𝛾 and so
inductively 𝑉𝛽 ⊆ 𝑉𝛾 , whence 𝑉𝛽 ∈ 𝑉𝛼 and finally 𝑉𝛽 ⊆ 𝑉𝛼 by transitivity.
(3) If 𝑥 ∈ 𝑉, then rk(𝑥) < 𝛼 ⇔ (∃𝛽 < 𝛼)𝑥 ∈ 𝑉𝛽+1 ⇔ 𝑥 ∈ 𝑉𝛼 .
(4) Let 𝛼 = rk(𝑥). Then 𝑥 ∈ 𝑉𝛼+1 = 𝒫(𝑉𝛼 ). For 𝑦 ∈ 𝑥, one obtains
𝑦 ∈ 𝑉𝛼 and hence 𝑦 ∈ 𝑉 and rk(𝑦) < 𝛼 by (3).
(5) For 𝑥 ∈ 𝑉, set 𝛼 = sup{rk(𝑦) + 1 ∣ 𝑦 ∈ 𝑥}. By (4), one has
𝛼 ≤ rk(𝑥). Since rk(𝑦) < 𝛼 for any 𝑦 ∈ 𝑥, it follows from (3) that 𝑥 ⊆ 𝑉𝛼 ,
whence 𝑥 ∈ 𝑉𝛼+1 and finally 𝛼 ≥ rk(𝑥) by definition.
(6) Suppose 𝑥 ∈ 𝑉 is transitive and 𝛽 < rk(𝑥) is given. By (5) there
exists 𝑦 ∈ 𝑥 such that 𝛽 ≤ rk(𝑦). Choose such an element 𝑦 with min-
imal rank. If 𝑧 ∈ 𝑦, then 𝑧 ∈ 𝑥 (by transitivity of 𝑥) and rk(𝑧) < rk(𝑦)
by (4), so rk(𝑧) < 𝛽 by minimality of rk(𝑦). This proves that 𝑦 ⊆ 𝑉𝛽 , and
thus 𝑦 ∈ 𝑉𝛽+1 , and hence rk(𝑦) ≤ 𝛽. It follows that rk(𝑦) = 𝛽.
(7) We prove by transfinite induction that 𝛼 ∈ 𝑉 and rk(𝛼) = 𝛼, the
case 𝛼 = 0 being clear. Suppose that the result holds for all 𝛽 < 𝛼. Then
𝛽 ∈ 𝑉𝛽+1 ⊆ 𝑉𝛼 for all 𝛽 < 𝛼, so 𝛼 ⊆ 𝑉𝛼 and thus 𝛼 ∈ 𝑉𝛼+1 . We get
rk(𝛼) = sup{𝛽 + 1 ∣ 𝛽 < 𝛼} = 𝛼 by (5).
(8) 𝑥 ∈ 𝑉 ⇒ 𝑥 ⊆ 𝑉 follows from (4). Conversely, if 𝑥 is a set such
that 𝑥 ⊆ 𝑉, then {rk(𝑦) + 1 ∣ 𝑦 ∈ 𝑥} is given by a set by Replacement. For
𝛼 = sup{rk(𝑦) + 1 ∣ 𝑦 ∈ 𝑥} we thus infer that 𝑥 ⊆ 𝑉𝛼 , and so 𝑥 ∈ 𝑉𝛼+1 .
(9) By (6), the set {rk(𝑦) ∣ 𝑦 ∈ 𝑥} is equal to an ordinal 𝛼. Since
𝛼 = sup{𝛽 + 1 ∣ 𝛽 < 𝛼}, by (5) we get rk(𝑥) = 𝛼. Thus 𝑥 ∈ 𝑉card(𝑥)+ by
(3), as card(𝛼) ≤ card(𝑥) and hence 𝛼 < card(𝑥)+ . □
Theorem 6.4.3. Let 𝒰 ⊧ ZF− . The following are equivalent:
(1) 𝒰 ⊧ (AF).
(2) 𝒰 ⊧ ∀𝑥𝑉(𝑥).
Proof. (2)⇒(1): Let 𝑥 be a non-empty set. We have to find an element

𝑦 of 𝑥 such that 𝑦 ∩ 𝑥 = ∅. By hypothesis, every set is in 𝑉. We choose
𝑦 ∈ 𝑥 such that rk(𝑦) is minimal. Then 𝑦 ∩ 𝑥 = ∅ by Lemma 6.4.2(4).
(1)⇒(2): Let 𝑥 be a set. Put 𝑦 ∶= tcl(𝑥), and consider the set 𝑧 ∶=

{𝑡 ∈ 𝑦 ∣ 𝑡 ∉ 𝑉}. We have 𝑥 ⊆ 𝑦. By Lemma 6.4.2(8), in order to prove
𝑥 ∈ 𝑉, it suffices to prove that every element of 𝑦 is in 𝑉, in other words
that 𝑧 = ∅. If 𝑧 were not empty, by (AF) there would exist 𝑡 ∈ 𝑧 with
𝑡 ∩ 𝑧 = ∅. Let us consider such a 𝑡. Then, for 𝑢 ∈ 𝑡, we would get 𝑢 ∈ 𝑦
(by transitivity of 𝑦) and thus 𝑢 ∈ 𝑉 (since 𝑡 ∩ 𝑧 = ∅). But then 𝑡 ∈ 𝑉 by
Lemma 6.4.2(8), which is a contradiction. □
Remark. As the elements of 𝑉 are constructed starting from the empty
set (by a transfinite procedure), Theorem 6.4.3 expresses that (AF) is
equivalent to the fact that every set is constructed from the empty set.
−
Remark 6.4.4. In the theory ZFC , the following are equivalent:
(1) (AF).
(2) There is no sequence (𝑎𝑖 )𝑖∈𝜔 with 𝑎𝑖+1 ∈ 𝑎𝑖 for all 𝑖 ∈ 𝜔.
Proof. (1)⇒(2): (This implication does not use (AC).) If (𝑎𝑖 )𝑖∈𝜔 is a se-
quence of sets, consider 𝑎 = {𝑎𝑖 ∣ 𝑖 ∈ 𝜔}. By (AF), there exists 𝑏 ∈ 𝑎
such that 𝑏 ∩ 𝑎 = ∅. Hence for some 𝑛 ∈ 𝜔 one has 𝑎𝑛 ∩ 𝑎 = ∅, so in
particular 𝑎𝑛+1 ∉ 𝑎𝑛 .
(2)⇒(1): Let 𝑎 ≠ ∅ be a set which contradicts (AF). This means that
for every 𝑏 ∈ 𝑎 one has 𝑏 ∩ 𝑎 ≠ ∅. By (AC) there exists a function
𝑓 ∶ 𝑎 → 𝑎 such that 𝑓(𝑏) ∈ 𝑏 for every 𝑏 ∈ 𝑎. Choose 𝑎0 ∈ 𝑎, and
define recursively a sequence (𝑎𝑖 )𝑖∈𝜔 , putting 𝑎𝑛+1 = 𝑓(𝑎𝑛 ). □
Lemma 6.4.5.
(1) If 𝑥 ∈ 𝑉, then ⋃ 𝑥, 𝒫(𝑥) and {𝑥} are in 𝑉. The rank of these sets
is strictly smaller than rk(𝑥) + 𝜔.
(2) If 𝑥, 𝑦 ∈ 𝑉, then 𝑥 × 𝑦, 𝑥 ∪ 𝑦, 𝑥 ∩ 𝑦, {𝑥, 𝑦}, (𝑥, 𝑦) and 𝑥𝑦 are in
𝑉, too. Moreover, the rank of these sets is strictly smaller than
max{rk(𝑥), rk(𝑦)} + 𝜔.
Remark. As the Extensionality Axiom (which expresses in some sense
that there are only sets), the Axiom of Foundation restricts the universe
of sets to the place where the usual mathematics take place:
6.5. Independence and Relative Consistency 161
• The usual objects of mathematics (for instance ℕ and ℝ) are in

𝑉 (cf. Exercise 6.7.1).
−
• If we work in ZFC , then any (first-order) structure has an iso-
morphic copy in 𝑉. (We suppose that the signature 𝜎ℒ0 of the
language we consider is finite, to simplify the presentation.)
Indeed, suppose that 𝔐 = ⟨𝑀; 𝑅𝑖 , 𝑐𝑗 , 𝑓𝑘 ⟩ is an ℒ0 -structure,
and let 𝑔 ∶ 𝑀 ≅ card(𝑀) = 𝜅 be a bijection. There exists a
unique ℒ0 -structure 𝔑 with base set 𝜅 such that 𝑔 is an ℒ0 -
isomorphism between 𝔐 and 𝔑. Then 𝜅 ∈ 𝑉𝜅+1 , and so 𝔑 ∈
𝑉𝜅+𝜔 by Lemma 6.4.5.
• This remains true for structures outside the first-order context.
Consider for example the case of a topological space (𝑋, 𝒯),
with 𝑋 a set and 𝒯 ⊆ 𝒫(𝑋) the set of open subsets of 𝑋. Then
if 𝑔 ∶ 𝑋 ≅ 𝜅 is a bijection, one may argue as before to prove
that (𝑋, 𝒯) admits an isomorphic copy in 𝑉𝜅+𝜔 .
6.5. Some Results on Incompleteness, Independence

and Relative Consistency
In this section, we will suppose that the ℒ∈ -formulas and the proofs are
coded in ZF− . To this aim, we may for example appeal to the coding
we have given in arithmetic and use that if 𝒰 ⊧ ZF− and 𝜔 ∈ 𝑈, then
⟨𝜔; 0, +, ⋅, succ, ∈↾ 𝜔 ⟩ ⊧ PA (see Proposition 6.2.10). We continue to write
#𝜑 for the code of 𝜑.
Gödel’s incompleteness results have their analogues in set theory.
We will state them and leave the proofs to the reader – the arguments
are analogous to those in the case of Peano arithmetic. However, coding
formulas and formal proofs is slightly simpler in set theory. Let us note
−
that ZF− , ZF, ZFC and ZFC are recursive theories.
Theorem 6.5.1 (Gödel’s First Incompleteness Theorem). Let 𝑇 be a re-
cursive and consistent ℒ∈ -theory such that 𝑇 ⊇ ZF− . Then 𝑇 is incom-
plete. □
Remark. If ZF− is consistent, then (AF) is independent of ZF− , that is,
ZF− ⊬ (AF) and ZF− ⊬ ¬(AF). (This is shown in Exercise 6.7.7.) Sim-
ilarly, (AC) is then independent of ZF and (CH) is independent of ZFC.
All of (AF), (AC) and (CH) are mathematically meaningful statements.

In contrast to this, in the case of arithmetic, assuming that PA is con-
sistent, the sentences which we proved to be independent of PA were
rather far fetched and involved the diagonal method.
Theorem 6.5.2 (Gödel’s Second Incompleteness Theorem). Let 𝑇 be a
recursive and consistent ℒ∈ -theory such that 𝑇 ⊇ ZF− . Then 𝑇 ⊬ Con(𝑇).
□
We work in a model 𝒰 of ZF− . One may then code the satisfaction

in 𝒰 as follows. Let ℒ be a (finite) language and 𝔄 = ⟨𝐴; (𝑍 𝔄 )𝑍∈𝜍ℒ ⟩ an ℒ-
structure with 𝔄 ∈ 𝑈. One identifies the set of assignments (with values
in 𝔄) with 𝐴𝜔 , which is an element of 𝑈. The following lemma may be
proved by induction on the height of a formula, which corresponds to
an induction on 𝜔. We leave the details as an exercise.
Lemma 6.5.3. There exists a functional class which to any triple (𝔄, #𝜑, 𝛼)
in 𝒰, with 𝔄 an ℒ-structure, 𝜑 an ℒ-formula and 𝛼 an assignment with
values in 𝔄, associates 1 if 𝔄 ⊧ 𝜑[𝛼], and 0 otherwise.
In particular, if 𝔄 is an ℒ-structure in 𝒰, then the collection # Th(𝔄) =
{#𝜑 ∣ 𝜑 is an ℒ-sentence such that 𝔄 ⊧ 𝜑} is given by a set in 𝒰. □
A relative consistency result has the following form: given two the-
ories 𝑇1 and 𝑇2 , the consistency of 𝑇1 is proved to imply the consistency
of 𝑇2 . The proof method will be to construct a model of 𝑇2 from a model
of 𝑇1 .
Let us start with a result which establishes a relation between set
theory and arithmetic. One verifies without difficulty that the proof of
Proposition 6.2.10 may be done in ZF− . As the structure ⟨𝜔; +, ×, 𝑆, 0, <⟩
is a set in 𝒰, Lemma 6.5.3 entails the following result.
Proposition 6.5.4. One has ZF− ⊢ Con(PA). In particular, if ZF− is
consistent, then Peano arithmetic PA is consistent, too. □
Let 𝒰 ⊧ ZF− . If 𝑋 is a non-empty class in 𝑈, one may consider the

(naive) ℒ∈ -structure ⟨𝑋; ∈↾ 𝑋 ⟩. In the proofs of the relative consistency
results we will present, we will construct classes 𝑋 such that if 𝒰 ⊧ 𝑇1 ,
then ⟨𝑋; ∈↾𝑋 ⟩ ⊧ 𝑇2 , whence the desired relative consistency result.
Notation. We will sometimes write 𝑋 ⊧ 𝜑 instead of ⟨𝑋; ∈↾𝑋 ⟩ ⊧ 𝜑.
Definition (Relativization). Let ℒ be a language, and let 𝐹(𝑣0 ) be an

ℒ-formula. For every ℒ-formula 𝜑 one defines, by induction on ht(𝜑),
an ℒ-formula 𝜑𝐹 , the relativization of 𝜑 to 𝐹:
• 𝜑𝐹 ∶= 𝜑 if 𝜑 is an atomic formula.
• (𝜑 ∧ 𝜓)𝐹 ∶= (𝜑𝐹 ∧ 𝜓𝐹 ) and [¬𝜑]𝐹 ∶= ¬[𝜑𝐹 ].
• [∃𝑥𝜑]𝐹 ∶= ∃𝑥(𝐹(𝑥) ∧ 𝜑𝐹 ).
Proposition 6.5.5. Let 𝒰 ⊧ ZF− , and let 𝑋 ⊆ 𝑈 be a class.
(1) Suppose that 𝑋 = 𝐹[𝒰] = 𝐺[𝒰]. For every ℒ∈ -formula 𝜑, the
formulas 𝜑𝐹 and 𝜑𝐺 are then equivalent in 𝒰. One may thus
write 𝜑𝑋 instead of 𝜑𝐹 , if one is only interested in the formula up
to equivalence.
(2) For every 𝑎1 , … , 𝑎𝑛 ∈ 𝑋 and formula 𝜑(𝑥1 , … , 𝑥𝑛 ), one has 𝒰 ⊧
𝜑𝑋 [𝑎] if and only if 𝑋 ⊧ 𝜑[𝑎].
Proof. (1) The proof is an easy induction on ht(𝜑) and left as an exercise.
(2) The proof is by induction on ht(𝜑), the only non-trivial case being
the case when 𝜑 is of the form ∃𝑥0 𝜓. In this case, one has
𝒰 ⊧ 𝜑𝑋 [𝑎1 , … , 𝑎𝑛 ] ⟺ 𝒰 ⊧ ∃𝑥0 (𝐹(𝑥0 ) ∧ 𝜓𝑋 )[𝑎1 , … , 𝑎𝑛 ]
⟺ there exists 𝑏 ∈ 𝑈 such that 𝒰 ⊧ 𝐹[𝑏] and 𝒰 ⊧ 𝜓𝑋 [𝑏, 𝑎]
⟺ there exists 𝑏 ∈ 𝑋 such that 𝒰 ⊧ 𝜓𝑋 [𝑏, 𝑎]
⟺ there exists 𝑏 ∈ 𝑋 such that 𝑋 ⊧ 𝜓[𝑏, 𝑎] ⟺ 𝑋 ⊧ 𝜑[𝑎]. □
Definition. Let 𝑋 be a class defined by an ℒ∈ -formula 𝐹(𝑣0 ). One
says that an ℒ∈ -formula 𝜑(𝑥1 , … , 𝑥𝑛 ) is absolute for 𝑋 if one has 𝒰 ⊧
𝑛
∀𝑥1 , … , 𝑥𝑛 (⋀𝑖=1 𝐹(𝑥𝑖 ) → (𝜑 ↔ 𝜑𝑋 )).
Definition. The set of Δ0 -formulas is the smallest set of ℒ∈ -formulas
which contains the atomic formulas and which is stable under boolean
combinations and bounded quantification (if 𝜑 is a Δ0 -formula and 𝑥, 𝑦
are two distinct variables, then ∃𝑥(𝑥 ∈ 𝑦 ∧ 𝜑) and ∀𝑥(𝑥 ∈ 𝑦 → 𝜑) are
Δ0 -formulas, too).
Lemma 6.5.6.
(1) Assume that 𝑋 is a non-empty class which is transitive (that is,
𝑧 ∈ 𝑦 ∈ 𝑋 ⇒ 𝑧 ∈ 𝑋). Then all Δ0 -formulas are absolute for 𝑋.
Moreover, the set of formulas which are absolute for 𝑋 is stable

under boolean combinations and bounded quantification.
(2) The following properties may be expressed by Δ0 -formulas: 𝑥 ⊆
𝑦, 𝑥 = ∅, 𝑥 = 𝑦 ∪ {𝑦}, ‘𝑥 is transitive’, 𝑧 = {𝑥, 𝑦}, 𝑦 = ⋃ 𝑥,
𝑧 = 𝑥 × 𝑦.
Proof. (1) Atomic formulas are absolute for every non-empty class.
Moreover, for any class 𝑋, the set of formulas which are absolute for
𝑋 is stable under boolean combinations. We now prove that if 𝑋 is a
transitive (non-empty) class and if 𝜑(𝑥, 𝑦, 𝑣1 , … , 𝑣𝑛 ) is absolute for 𝑋,
then the formula ∃𝑥(𝑥 ∈ 𝑦 ∧ 𝜑) is absolute for 𝑋, too. For elements
𝑏, 𝑐1 , … , 𝑐𝑛 ∈ 𝑋, one has the following equivalences:
𝑋
𝒰 ⊧ (∃𝑥(𝑥 ∈ 𝑦 ∧ 𝜑)) [𝑏, 𝑐]
⟺ 𝒰 ⊧ (∃𝑥(𝐹(𝑥) ∧ 𝑥 ∈ 𝑦 ∧ 𝜑𝑋 )) [𝑏, 𝑐]
⟺ there exists 𝑎 ∈ 𝑏 ∩ 𝑋 such that 𝒰 ⊧ 𝜑𝑋 [𝑎, 𝑏, 𝑐]
⟺ there exists 𝑎 ∈ 𝑏 ∩ 𝑋 such that 𝒰 ⊧ 𝜑[𝑎, 𝑏, 𝑐]
⟺ there exists 𝑎 ∈ 𝑏 such that 𝒰 ⊧ 𝜑[𝑎, 𝑏, 𝑐]
⟺ 𝒰 ⊧ (∃𝑥(𝑥 ∈ 𝑦 ∧ 𝜑)) [𝑏, 𝑐].
The first equivalence holds by definition, the third one follows from
the induction hypothesis and the fourth one from the transitiviy of 𝑋.
Stability under bounded universal quantification follows as well, since
∀𝑥(𝑥 ∈ 𝑦 → 𝜑) is equivalent to ¬∃𝑥(𝑥 ∈ 𝑦 ∧ ¬𝜑).
The proof of (2) is straightforward. For example, the transitivity of
𝑥 is expressed by the Δ0 -formula: ∀𝑦 (𝑦 ∈ 𝑥 → ∀𝑧(𝑧 ∈ 𝑦 → 𝑧 ∈ 𝑥)) . □
Let 𝑋 be a class. We say that a functional class 𝐺(𝑥, 𝑦) with Dom(𝐺) ⊇

𝑋 is an absolute functional class for 𝑋 (in 𝒰) if 𝐺(𝑥, 𝑦) is absolute for 𝑋
and if for any 𝑎 ∈ 𝑋, the unique 𝑏 ∈ 𝑈 such that 𝒰 ⊧ 𝐺(𝑎, 𝑏) is in 𝑋.
The following remark follows easily from the definitions:
Remark 6.5.7. A composition of absolute functional classes is an abso-

lute functional class (for 𝑋). □
Lemma 6.5.8. We work in 𝒰 ⊧ ZF− .
(1) Any non-empty transitive class satisfies the Extensionality Ax-

iom.
(2) Let 𝑋 be a class defined in 𝒰, and let 𝐺 be an absolute functional
class for 𝑋. Then 𝑋 ⊧ ∀𝑥∃! 𝑦𝐺(𝑥, 𝑦).
Moreover, if 𝜑 is absolute for 𝑋, then ∃𝑥(𝑥 ∈ 𝐺(𝑧) ∧ 𝜑) and
∀𝑥(𝑥 ∈ 𝐺(𝑧) → 𝜑) are absolute for 𝑋, where 𝑥, 𝑧 are distinct
variables.
(3) The Union Axiom holds in 𝑉 and in any 𝑉𝛼 with 𝛼 > 0.
(4) The Pairing Axiom holds in 𝑉 and in any 𝑉𝜆 with 𝜆 a limit ordi-
nal.
(5) Suppose that 𝒰 ⊧ (∀𝑥 ∈ 𝑋)(∃𝑦 ∈ 𝑋)𝒫(𝑥) ∩ 𝑋 = 𝑦 for some
non-empty transitive class 𝑋. Then the Power Set Axiom holds
in 𝑋. In particular, the Power Set Axiom holds in 𝑉 and in 𝑉𝜆 for
any limit ordinal 𝜆.
Moreover, 𝑦 = 𝒫(𝑥) is a functional class which is absolute
for 𝑉 and 𝑉𝜆 with 𝜆 a limit ordinal.
Proof. (1) The Extensionality Axiom relativized to 𝑋 is given by the sen-

tence (∀𝑦 ∈ 𝑋)(∀𝑧 ∈ 𝑋) ((∀𝑥 ∈ 𝑋)(𝑥 ∈ 𝑦 ↔ 𝑥 ∈ 𝑧) → 𝑦 = 𝑧). Since 𝑋 is
transitive, if 𝑦, 𝑧 ∈ 𝑋, then 𝑦, 𝑧 ⊆ 𝑋. The result follows.
(2) The first part is clear. Now assume that 𝐺 is an absolute func-
tional class for 𝑋 and that 𝜑(𝑣, 𝑥, 𝑧) is absolute for 𝑋. Given 𝑎, 𝑏 from
𝑋, let 𝑐 = 𝐺(𝑏) (computed in 𝒰 or in 𝑋, which amounts to the same by
the first part of (2)). Then 𝒰 ⊧ ∃𝑥(𝑥 ∈ 𝐺(𝑧) ∧ 𝜑)[𝑎, 𝑏] if and only if
𝒰 ⊧ ∃𝑥(𝑥 ∈ 𝑦 ∧ 𝜑)[𝑎, 𝑐], and similarly for 𝑋 in place of 𝒰. The absolute-
ness of ∃𝑥(𝑥 ∈ 𝐺(𝑧) ∧ 𝜑) then follows from Lemma 6.5.6(1).
(3) The formula 𝑦 = ⋃ 𝑥 is Δ0 by Lemma 6.5.6, so absolute for tran-
sitive classes. As rk(⋃ 𝑥) ≤ rk(𝑥), one has 𝑥 ∈ 𝑉𝛼 ⇒ ⋃ 𝑥 ∈ 𝑉𝛼 for all 𝛼
and in particular 𝑥 ∈ 𝑉 ⇒ ⋃ 𝑥 ∈ 𝑉. One concludes by (2).
(4) One has 𝑥, 𝑦 ∈ 𝑉𝛼 ⇒ {𝑥, 𝑦} ∈ 𝑉𝛼+1 . It follows in particular
that 𝑥, 𝑦 ∈ 𝑉𝜆 ⇒ {𝑥, 𝑦} ∈ 𝑉𝜆 for 𝜆 a limit ordinal. One deduces that
(𝑥, 𝑦) ↦ {𝑥, 𝑦} is given by a functional class which is absolute for 𝑉 and
𝑉𝜆 (for 𝜆 a limit ordinal), and one may conclude by (2).
(5) One has 𝒰 ⊧ ∀𝑥 ∈ 𝑋 ∃𝑦 ∈ 𝑋(𝒫(𝑥) ∩ 𝑋 = 𝑦) if and only if
𝒰 ⊧ ∀𝑥 ∈ 𝑋 ∃𝑦 ∈ 𝑋 ∀𝑧 ∈ 𝑋(𝑧 ⊆ 𝑥 ↔ 𝑧 ∈ 𝑦). The latter sentence is just
the Power Set Axiom relativized to 𝑋, since 𝑧 ⊆ 𝑥 is absolute for 𝑋 by

the transitivity of 𝑋. The proof of the remaining part of (5) is left as an
exercise. □
Lemma 6.5.9. The formulas Ord(𝑥) and Card(𝑥) are absolute for 𝑉 as
well as for 𝑉𝜆 if 𝜆 is a limit ordinal.
Proof. We first consider Ord(𝑥). The transitivity of 𝑥 and the fact that
∈↾ 𝑥 defines a total order are expressed by Δ0 -formulas and are thus ab-
solute in both cases. The items (2) and (5) of Lemma 6.5.8 entail that the
following formula is absolute:
∀𝑧 (𝑧 ∈ 𝒫(𝑥) ∧ 𝑧 ≠ ∅ → ∃𝑢(𝑢 ∈ 𝑧 ∧ ∀𝑤(𝑤 ∈ 𝑧 → 𝑤 ∉ 𝑢))) .
This establishes the absoluteness of well-foundedness.
The formula Ord(𝑥) ∧ ∀𝑦(𝑦 ∈ 𝑥 → ¬∃𝑓 ∈ 𝒫(𝑥 × 𝑦) 𝑓 ∶ 𝑥 ≅ 𝑦)
is equivalent to Card(𝑥). One proves easily that the formula 𝑓 ∶ 𝑥 ≅ 𝑦
(in the three variables 𝑓, 𝑥, 𝑦) is absolute for 𝑉 and for 𝑉𝜆 if 𝜆 is a limit
ordinal, and that the functional class 𝑧 = 𝑥 × 𝑦 is an absolute functional
class in both cases (exercise). This proves the absoluteness of Card(𝑥) in
both cases, using Lemma 6.5.8(2) and Remark 6.5.7. □
Definition. We work in ZF− . Let 𝜅 be a cardinal.
• We say 𝜅 is a strong limit cardinal if for all 𝜇 < 𝜅 one has 2𝜇 < 𝜅.
• We say 𝜅 is inaccessible (strongly) if it is strong limit, regular
and > ℵ0 .
Example. By transfinite recursion, one defines a cardinal hierarchy as
follows: ℶ0 ∶= ℵ0 , ℶ𝛼+1 ∶= 2ℶ𝛼 , and ℶ𝜆 ∶= sup𝛼<𝜆 ℶ𝛼 for 𝜆 a limit
ordinal.
For any limit ordinal 𝜆, ℶ𝜆 is then a strong limit cardinal. But as
cof(ℶ𝜆 ) = cof(𝜆) in this case, ℶ𝜆 is singular in general.
Lemma 6.5.10. Let 𝒰 ⊧ ZF− , and let 𝑋 = 𝑉 or 𝑋 = 𝑉𝜔 , or 𝑋 = 𝑉𝜅
for 𝜅 an inaccessible cardinal. In the last case, we assume in addition that
𝒰 ⊧ (AC). Then 𝑋 satisfies the Replacement Axiom Scheme and the Axiom
of Foundation.
Proof. Choose a formula 𝐹(𝑥) which defines 𝑋. Let 𝐺(𝑣0 , 𝑣1 ) be a for-

mula (with parameters from 𝑋) which defines a functional class in the
structure ⟨𝑋, ∈↾ 𝑋 ⟩. Then the formula 𝐻(𝑣0 , 𝑣1 ) = 𝐹(𝑣0 ) ∧ 𝐹(𝑣1 ) ∧ 𝐺 𝐹

defines a functional class in 𝒰. Set 𝑏 ∶= 𝐻[𝑎] = {𝐻(𝑐) ∣ 𝑐 ∈ 𝑎}, where 𝑎
is a set in 𝑋.
• If 𝑋 = 𝑉, then 𝑏 ∈ 𝑉 by Lemma 6.4.2(8), since 𝑏 ⊆ 𝑉 and 𝑏 is
a set.
• If 𝑋 = 𝑉𝜔 , then, as 𝑎 is finite, 𝑏 is a finite subset of 𝑉𝜔 , so 𝑏 ⊆ 𝑉𝑛
for some 𝑛 ∈ 𝜔 and hence 𝑏 ∈ 𝑉𝜔 .
• If 𝑋 = 𝑉𝜅 for an inaccessible cardinal 𝜅, essentially the same ar-
gument works. One proves by transfinite induction that
card(𝑉𝛼 ) < 𝜅 for all 𝛼 < 𝜅. (Since we assume 𝒰 ⊧ (AC), we
may use cardinalities.) For 𝛼 a successor ordinal, one uses that
𝜅 is strongly limit; for 𝛼 a limit ordinal, one uses the regularity
of 𝜅.
Now, if 𝑎 ∈ 𝑉𝜅 , then 𝑎 ∈ 𝑉𝛼 for some 𝛼 < 𝜅 and thus
𝑎 ⊆ 𝑉𝛼+1 . It follows that card(𝑏) ≤ card(𝑎) < 𝜅. Moreover, one
has 𝑏 ⊆ 𝑉𝜅 . Since 𝜅 is regular, sup{rk(𝑐) + 1 ∣ 𝑐 ∈ 𝑏} < 𝜅 and
hence 𝑏 ∈ 𝑉𝜅 .
This proves the Replacement scheme in all three cases.
In order to prove that (AF) holds, we consider ∅ ≠ 𝑎 ∈ 𝑋. It follows
that 𝑎 ⊆ 𝑋 in all three cases, as 𝑋 is transitive. For any 𝑏 ∈ 𝑎 of minimal
rank, one has 𝑎 ∩ 𝑏 = ∅, and any such 𝑏 is in 𝑋. One concludes, since
the formula 𝑎 ∩ 𝑏 = ∅ is absolute for 𝑋. □
Lemma 6.5.11. One has 𝑉 ⊧ (AI), 𝑉𝜆 ⊧ (AI) for any limit ordinal 𝜆 > 𝜔,
and 𝑉𝜔 ⊧ ¬(AI).
Proof. One has 𝜔 ∈ 𝑉, 𝜔 ∈ 𝑉𝜆 and 𝜔 ∉ 𝑉𝜔 . The details are left to the

reader. □
−
Lemma 6.5.12. If 𝒰 ⊧ ZFC , then 𝑉 ⊧ (AC) and 𝑉𝜅 ⊧ (AC), for 𝜅 an
inaccessible cardinal.
Proof. Let (𝑋𝑖 )𝑖∈𝐼 be an element of 𝑉𝜅 . Then 𝐼 ∈ 𝑉𝜅 and every element

𝑔 ∈ ∏𝑖∈𝐼 𝑋𝑖 is in 𝑉𝜅 . For 𝑉, one concludes by the same argument. □
Theorem 6.5.13 (Relative consistency of (AF)).
(1) If 𝒰 ⊧ ZF− , then 𝑉 ⊧ ZF. In particular, the consistency of ZF−

entails the consistency of ZF.
−
(2) If 𝒰 ⊧ ZFC , then 𝑉 ⊧ ZFC. In particular, the consistency of
−
ZFC entails the consistency of ZFC.
Proof. It suffices to combine the preceding lemmas, taking into account

the fact that the Replacement scheme implies the Comprehension
scheme (cf. Lemma 6.2.5). □
Theorem 6.5.14. If 𝒰 ⊧ ZF− , then 𝑉𝜔 ⊧ ZFC−(AI)+¬(AI). In particular,

ZF− ⊧ Con(ZFC − (AI) + ¬(AI)).
Proof. We have proved everything in the lemmas but 𝑉𝜔 ⊧ (AC). Ob-

serve that every element 𝑎 of 𝑉𝜔 is a finite set and in bijection with an
element of 𝜔. Such a bijection 𝑓 as well as the well-order on 𝑎 induced
by 𝑓 is in 𝑉𝜔 . □
Let (IC) be the statement ‘There exists an inaccessible cardinal’.

−
Theorem 6.5.15. Let 𝒰 ⊧ ZFC and 𝜅 ∈ 𝑈 be an inaccessible cardinal.
−
Then 𝑉𝜅 ⊧ ZFC. In particular, ZFC + (IC) ⊧ Con(ZFC).
−
Proof. All axioms of ZFC hold in 𝑉𝜅 by the preceding lemmas. □
By the Second Incompleteness Theorem, it follows from Theorem

6.5.15 that ZFC ⊬ (IC). The following theorem states the corresponding
relative consistency result. We will give a direct proof which does not
rely on the Second Incompleteness Theorem.
Theorem 6.5.16. If ZFC is consistent, then ZFC + ¬(IC) is consistent, too.
Proof. Let 𝒰 ⊧ ZFC. We may suppose that 𝒰 ⊧ (IC). Let 𝜅 be the

smallest inaccessible cardinal in 𝒰. Then 𝑉𝜅 ⊧ ZFC by Theorem 6.5.15.
To conclude, it suffices to prove that 𝑉𝜅 ⊧ ¬(IC), a fact which follows
from the following observations:
• 𝑦 = 𝒫(𝑥), Ord(𝑥) and Card(𝑥) are absolute formulas for 𝑉𝜅
(cf. Lemma 6.5.8 and Lemma 6.5.9).
6.6. A Glimpse of Further Independence Results 169
• ‘𝜆 is a regular cardinal’ is absolute for 𝑉𝜅 . [Indeed,

it is easy to see that this property may be expressed by the for-
mula ¬(∃𝛼 < 𝜆)(∃𝑓 ∶ 𝛼 → 𝜆 cofinal).]
• ‘𝜆 is a strong limit cardinal’ is absolute for 𝑉𝜅 . [Indeed, this
property may be expressed by the following formula (in 𝜆):
(∀𝛼 < 𝜆)¬(∃𝑓 ∶ 𝒫(𝛼) → 𝜆 surjective).] □
Remark 6.5.17. If ZFC is consistent, then
ZFC ⊭ Con(ZFC) → Con(ZFC + (IC)).
Proof. Assume ZFC ⊧ Con(ZFC) → Con(ZFC+(IC)). Then we have in

−
particular ZFC+(IC) ⊧ Con(ZFC) → Con(ZFC +(IC)), so ZFC+(IC) ⊧
Con(ZFC+(IC)) by Theorem 6.5.15. By the Second Incompleteness The-
orem, this means that ZFC + (IC) is inconsistent, which implies the in-
consistency of ZFC by assumption. □
6.6. A Glimpse of Further Independence and Relative

Consistency Results
At the end of this chapter, we will mention some important further re-
sults on independence and relative consistency in axiomatic set theory.
We will not give any proofs, but we will sketch some key ideas of the
methods behind these results. This part is meant as a motivation for
further reading. The books by Krivine [6] and Kunen [7] are excellent
references.
The results in the following theorem may be obtained, in a
rather elementary way, by the method of Fraenkel-Mostowski. They
will be treated entirely in the exercise section (cf. Exercise 6.7.7 and
Exercise 6.7.8).
Theorem 6.6.1.
−
(1) If ZF is consistent, then so is ZFC + ¬(AF).
(2) If ZF is consistent, then so is ZF− + ¬(AC).
It is possible to replace ZF− + ¬(AC) by ZF + ¬(AC) in the second

part of the preceding theorem, but then the proof gets more involved (see
below).
Recall that the Generalized Continuum Hypothesis (GCH) is given

by the sentence ∀𝛼(2ℵ𝛼 = ℵ𝛼+1 ).
Theorem 6.6.2 (Gödel). If ZF is consistent, then so is the theory ZFC +
(GCH).
Gödel obtains this result using the universe of constructible sets 𝐿.

The construction of 𝐿 is similar to that of 𝑉 inside a model 𝒰 of ZF− .
Here is an outline of the construction.
• One uses the formalization in 𝒰 ⊧ ZF of the syntax and the
satisfaction for structures ⟨𝑎; ∈↾ 𝑎 ⟩, for 𝑎 in 𝒰.
• One proves that there is a functional class 𝒟 which to any set
associates the set of its parameter definable subsets, that is, if
𝑎 is a set, the set 𝒟(𝑎) equals
{𝑏 ∈ 𝒫(𝑎) ∣ there are 𝑛 ∈ 𝜔, an ℒ∈ -formula 𝜑(𝑣0 , … , 𝑣𝑛 ) and

𝑐1 , … , 𝑐𝑛 ∈ 𝑎 with 𝑏 = {𝑐0 ∈ 𝑎 ∣ ⟨𝑎; ∈↾𝑎 ⟩ ⊧ 𝜑[𝑐0 , 𝑐1 , … , 𝑐𝑛 ]}}.
By convention, one sets 𝒟(∅) ∶= 𝒫(∅) = {∅}.
• Using induction, one defines 𝐿0 ∶= ∅, ℒ𝛼+1 ∶= 𝒟(𝐿𝛼 ), and
𝐿𝜆 ∶= ⋃𝛼<𝜆 𝐿𝛼 for 𝜆 a limit ordinal. Finally, one sets ‘𝐿 =
⋃𝛼 𝐿𝛼 ’, that is, 𝐿 is defined by the formula 𝐿(𝑥) = ∃𝛼(Ord(𝛼) ∧
𝑥 ∈ 𝐿𝛼 ).
• As for the von Neumann hierarchy, one may establish certain
basic properties, for instance, 𝛼 ≤ 𝛽 ⇒ 𝐿𝛼 ⊆ 𝐿𝛽 and the tran-
sitivity of the 𝐿𝛼 . In this way, one gets a continuous hierarchy
of transitive sets (𝐿𝛼 )𝛼∈Ord .
• One defines a rank rk𝐿 as in the case of 𝑉, and one may prove
that rk𝐿 (𝛼) = 𝛼, that is, 𝐿𝛼 ∩ Ord = 𝛼. In particular, 𝐿 is a
proper class.
• The proof that 𝐿 ⊧ ZF is very similar to the one for 𝑉. It also
follows that 𝐿 has the same ordinals as 𝒰.
• One proves that in 𝒰, 𝐿𝐿 = 𝐿 holds. In other words, 𝐿 satisfies
the Axiom of Constructibility ∀𝑥 𝐿(𝑥). To prove this, one estab-
lishes that the functional class 𝒟 is absolute for any transitive
class 𝑋 which is a model of ZF with the Power Set Axiom taken
6.6. A Glimpse of Further Independence Results 171
away, and then that the same absoluteness statement holds for
the functional class 𝛼 ↦ 𝐿𝛼 .
• In order to prove that 𝐿 ⊧ (AC), one uses transfinite recursion
to construct a well-ordering on 𝐿𝛼 such that if 𝛼 ≤ 𝛽, then 𝐿𝛼
is an initial segment of 𝐿𝛽 . In the successor step, one uses the
well-ordering 𝐿𝛼 to construct a well-ordering on 𝐿<𝜔 𝛼 (the set
of finite sequences in 𝐿𝛼 ), then on 𝐿𝛼+1 , also enumerating the
(codes) of ℒ∈ -formulas.
• Finally, 𝐿 ⊧ (GCH) is established as follows. One proves first
that, in any model of ZF− , one has card(𝐿𝛼 ) = card(𝛼) for every
infinite ordinal 𝛼. Then, one proves that, in 𝐿, one has 𝐿𝜅 =
{𝑥 ∣ card(tcl(𝑥)) < 𝜅} for every infinite cardinal 𝜅. Thus, 𝐿 ⊧
𝒫(𝜅) ⊆ 𝐿𝜅+ , and so 𝐿 ⊧ (GCH).
Remark. A weaker statement, namely that the consistency of ZF entails

the consistency of ZFC, is the content of Exercise 6.7.6.
Theorem 6.6.3 (Cohen).

(1) If ZF is consistent, then so is ZFC + ¬(CH).
(2) If ZF is consistent, then so is ZF + ¬(AC).
In particular, combining the results by Gödel and Cohen, one gets

the following result concerning the cardinality of the continuum.
Corollary 6.6.4. If ZF is consistent, the Continuum Hypothesis (CH) is

independent of ZFC.
Cantor had formulated (CH) and thus raised the problem of the car-
dinality of the continuum in 1878. It was then put forward by Hilbert
as the first problem of his list of 23 important problems in mathematics
which he presented at the International Congress of Mathematicians in
Paris in 1900.
Contrary to the constructions of models of set theory we have seen
so far (and to the ones which will be treated in the exercises), Cohen’s
construction is not done inside the ground model 𝒰. The method Cohen
introduced in 1964 to prove his results, termed forcing, allows for con-
structions of models which extend a given model.
Weakenings of the Axiom of Choice. The following two variants of

the Axiom of Choice are often considered.
Axiom of Dependent Choice (DC).
∀𝑟∀𝑎∀𝑥0 [(𝑥0 ∈ 𝑎 ∧ 𝑟 ⊆ 𝑎2 ∧ ∀𝑥 ∈ 𝑎 ∃𝑦 ∈ 𝑎 (𝑥, 𝑦) ∈ 𝑟)

→ ∃𝑓(𝑓 ∶ 𝜔 → 𝑎 ∧ 𝑓(0) = 𝑥0 ∧ ∀𝑛 ∈ 𝜔 ((𝑓(𝑛), 𝑓(𝑛 + 1)) ∈ 𝑟))].
Axiom of Countable Choice (CC).
∀𝑓[(Fn(𝑓) ∧ dom(𝑓) = 𝜔 ∧ ∅ ∉ im(𝑓))

→ ∃𝑔 (Fn(𝑔) ∧ dom(𝑔) = 𝜔 ∧ ∀𝑥(𝑥 ∈ dom(𝑔) → 𝑔(𝑥) ∈ 𝑓(𝑥)))].
It expresses that the product of a countable family of non-empty sets

is non-empty.
Proposition 6.6.5. One has (AC) ⇒ (DC) ⇒ (CC).
Proof. (AC)⇒(DC) Let ℎ ∶ 𝒫 ′ (𝑎) → 𝑎 be a choice function on 𝑎. By

induction on 𝜔, define a function 𝑓 ∶ 𝜔 → 𝑎, setting 𝑓(0) ∶= 𝑥0 , 𝑓(𝑛 +
1) ∶= ℎ({𝑦 ∈ 𝑎 ∣ (𝑓(𝑛), 𝑦) ∈ 𝑟}). Clearly, 𝑓 is as required.
(DC)⇒(CC) Let (𝑋𝑛 )𝑛∈𝜔 be a countable family of non-empty sets.
Set 𝑌𝑛 ∶= {𝑛} × 𝑋𝑛 and 𝑎 ∶= ⋃𝑛∈𝜔 𝑌𝑛 , and let
𝑟 ∶= {(𝑥, 𝑦) ∈ 𝑎2 ∣ there exists 𝑛 ∈ 𝜔 such that (𝑥, 𝑦) ∈ 𝑌𝑛 × 𝑌𝑛+1 }.
By (DC) there exists a function 𝑓 ∶ 𝜔 → 𝑎 such that 𝑓(𝑛) ∈ 𝑌𝑛 for
all 𝑛 ∈ 𝜔. Then 𝑔 ∈ ∏𝑛∈𝜔 𝑋𝑛 , where 𝑔(𝑛) ∶= 𝜋((𝑓(𝑛)), with 𝜋 the
projection onto the second coordinate. □
The usual construction of a set of real numbers which is not Lebesgue

measurable may be done in ZFC. It proves the following:
Proposition 6.6.6.
ZFC ⊧ ‘there exists a subset of ℝ which is not Lebesgue measurable’
Let us mention a theorem which is more difficult (its proof uses forc-
ing).
Theorem 6.6.7 (Solovay, 1970). If ZFC+(IC) is consistent, then the theory
ZF + (DC) + ‘every subset of ℝ is Lebesgue measurable’ is consistent, too.
6.7. Exercises 173
6.7. Exercises
Exercise 6.7.1. We work in 𝒰 ⊧ ZF− , and we suppose that the sets
ℕ, ℤ, … are defined in the usual manner.
(1) Prove that the sets ℕ, ℤ, ℝ, ℂ, 𝐶 0 ([0, 1], ℂ) are in 𝑉𝜔⋅2 .
(2) Compute the ranks of the sets in (1).
Exercise 6.7.2 (Mostowski Collapse). We work in 𝒰 ⊧ ZF− . Let 𝑋 be a
class and 𝑅 ⊆ 𝑋 × 𝑋 a definable relation on 𝑋.
We say 𝑅 is set-like if for any 𝑎 in 𝑋, the class pred𝑅 (𝑎) of
𝑅-predecessors of 𝑎, defined by 𝑅(𝑥, 𝑎), is given by a set; 𝑅 is well-founded
if any non-empty subset 𝑎 of 𝑋 contains an 𝑅-minimal element; finally,
𝑅 is extensional if pred𝑅 (𝑎) = pred𝑅 (𝑏) implies 𝑎 = 𝑏.
(1) Assume that 𝑅 ⊆ 𝑋 ×𝑋 is set-like. By induction on 𝛼, for 𝑥 ∈ 𝑋,
define the relation rk𝑅 (𝑥) ≥ 𝛼 as follows:
• rk𝑅 (𝑥) ≥ 0 for all 𝑥 ∈ 𝑋;
• rk𝑅 (𝑥) ≥ 𝛼 + 1 if and only if there is 𝑦 ∈ pred𝑅 (𝑥) such
that rk𝑅 (𝑦) ≥ 𝛼;
• if 𝜆 is a limit, then rk𝑅 (𝑥) ≥ 𝜆 if and only if rk𝑅 (𝑥) ≥ 𝛼 for
all 𝛼 < 𝜆.
Set rk𝑅 (𝑥) ∶= ∞ if rk𝑅 (𝑥) ≥ 𝛼 for all 𝛼; otherwise, rk𝑅 (𝑥)
is defined as the least 𝛼 such that rk𝑅 (𝑥) ≱ 𝛼 + 1. (rk𝑅 is called
the foundation rank.)
Prove that 𝑅 is well-founded if and only if rk𝑅 takes ordinal
values on 𝑋.
(2) Assume now that 𝑅 is well-founded and set-like.
(a) Prove that if 𝐹 is a functional class with Domain 𝑋 and
Image contained in Ord such that 𝑅(𝑎, 𝑏) implies 𝐹(𝑎) <
𝐹(𝑏), then rk𝑅 (𝑥) ≤ 𝐹(𝑥) for all 𝑥 ∈ 𝑋.
(b) Prove that the Image of rk𝑅 is either an ordinal or the
whole of Ord.
(c) Prove that there is a unique functional class 𝜋 with
Dom(𝜋) = 𝑋 such that for every 𝑎 in 𝑋 one has 𝜋(𝑎) =
𝜋[pred𝑅 (𝑎)]. This 𝜋 is called the Mostowski Collapse of
(𝑋, 𝑅). Moreover, prove that Im(𝜋) is a transitive subclass
of 𝑉.
(3) Suppose that 𝑅 is well-founded, set-like and extensional, and

let 𝜋 be the Mostowski Collapse. Prove that 𝜋 induces a naive
isomorphism between (𝑋, 𝑅) and (Im(𝜋), ∈↾Im(𝜋) ).
(4) Suppose that 𝑋 is a proper class and that 𝑅 defines a set-like
well-ordering (that is, a total order which is well-founded) on
𝑋. Observe that 𝑅 is extensional. Prove that Im(𝜋) = Ord, that
is, 𝜋 induces a naive order isomorphism between (𝑋, 𝑅) and
(Ord, <).
<𝜔
Example: Let 𝑋 = Ord be the class of finite sequences of
<𝜔
ordinals. For 𝑠 = (𝑠0 , … , 𝑠𝑛−1 ) and 𝑡 = (𝑡0 , … , 𝑡𝑚−1 ) from Ord
set 𝑅(𝑠, 𝑡) if and only if
• max(𝑠) < max(𝑡), or
• max(𝑠) = max(𝑡) and 𝑛 < 𝑚, or
• max(𝑠) = max(𝑡), 𝑛 = 𝑚, 𝑠 ≠ 𝑡 and 𝑠𝑘 < 𝑡𝑘 , where 𝑘 is the
smallest index where 𝑠 and 𝑡 differ.
Prove that 𝑅 is a set-like well-ordering on 𝑋.
Exercise 6.7.3 (Reflection Principle). We work in some 𝒰 ⊧ ZF− .
(1) A class of ordinals 𝑋 ⊆ Ord is called CLUB if it is closed (that

is, for every subset 𝑥 ⊆ 𝑋, one has sup 𝑥 ∈ 𝑋) and unbounded
in Ord.
Observe that the intersection of two CLUB classes is CLUB.
(2) Let (𝑊𝛼 )𝛼∈Ord be a continuous hierarchy of sets, that is, a func-
tional class 𝒲 ∶ Ord → 𝑈, 𝒲(𝛼) = 𝑊𝛼 , such that
• 𝑊𝛼 ⊆ 𝑊𝛽 for all 𝛼 ≤ 𝛽 and
• 𝑊𝜆 = ⋃𝛼<𝜆 𝑊𝛼 for any limit ordinal 𝜆.
Let 𝑊 be the ‘union’ of the 𝑊𝛼 , that is, the class defined by
the formula ∃𝛼 𝑥 ∈ 𝑊𝛼 . Given an ℒ∈ -formula 𝜑(𝑥1 , … , 𝑥𝑛 ), we
say that 𝑊𝛼 reflects 𝜑 if
𝑛
𝒰 ⊧ ∀𝑥1 , … , 𝑥𝑛 ( 𝑥𝑖 ∈ 𝑊𝛼 → (𝜑𝑊 (𝑥) ↔ 𝜑𝑊𝛼 (𝑥))) .
⋀
𝑖=1
(a) Observe that the ordinals 𝛼 such that 𝑊𝛼 reflects 𝜑 form a

class.
6.7. Exercises 175
(b) Prove the following General Reflection Principle: For ev-

ery formula 𝜑(𝑥), the class of ordinals such that 𝑊𝛼 re-
flects 𝜑 contains a CLUB.
[Hint: Proceed by induction on the height of 𝜑.]
(c) Deduce the usual Reflection Principle: For any formula
𝜑(𝑥), one has
ZF ⊧ ∀𝑦∃𝛼 (𝑦 ∈ 𝑉𝛼 ∧ ∀𝑥 (𝜑(𝑥) ↔ 𝜑𝑉𝛼 (𝑥))) .
Exercise 6.7.4 (Independence of the Replacement Axiom Scheme). The

Zermelo axioms Z are given by the axioms of ZF without the Replacement
scheme. We work in 𝒰 ⊧ ZF− .
(1) Prove that if 𝜆 > 𝜔 is a limit ordinal, then 𝑉𝜆 ⊧ Z.
(2) Prove that 𝑉𝜔2 does not satisfy the Replacement scheme.
[Hint: Consider the functional class sending 𝛼 to 𝜔 + 𝛼.]
(3) Deduce that if Z is consistent, then it does not imply the Re-
placement scheme. Similarly, prove that if Z+(AC) is consis-
tent, then it does not imply the Replacement scheme.
Exercise 6.7.5 (Non-finite axiomatizability of ZF). The aim of this ex-

ercise is to prove that if 𝑇 ⊇ ZF is consistent, then 𝑇 is not finitely ax-
iomatizable.
(1) Let 𝒰 ⊧ ZF, and let 𝑋 ⊆ 𝑈 be a transitive class which is also a
model of ZF. Prove that for any ordinal 𝛽 ∈ 𝑋 one has 𝑉𝛽𝑋 =
𝑉𝛽 ∩ 𝑋, where 𝑉𝛽𝑋 denotes ‘𝑉𝛽 computed in 𝑋’.
(2) Let 𝒰 ⊧ 𝜑, where 𝜑 is a sentence such that 𝜑 ⊢ ZF. Using the
Reflection Principle (cf. Exercise 6.7.3) prove that there is 𝛼 in
𝑈 such that 𝑉𝛼 ⊧ 𝜑.
(3) Conclude.
(4) Prove that if ZF− is consistent, then it is not finitely aximatiz-
able.
(5) Prove the following sharpening: If 𝑇 ⊇ ZF is consistent, then
there is no sentence 𝜑 such that 𝜑 together with the Compre-
hension scheme axiomatizes 𝑇.
Exercise 6.7.6 (Relative consistency of (AC)). We work in 𝒰 ⊧ ZF. In

this exercise, we assume that the set of ℒ∈ -formulas is coded in 𝜔.
(1) The class of ordinal definable sets is given by the ℒ∈ -formula
OD(𝑥) which expresses the following: ‘There exists 𝛼, a
formula 𝜑(𝑧, 𝑦0 , … , 𝑦𝑛−1 ) and a sequence (𝛼0 , … , 𝛼𝑛−1 )
of ordinals < 𝛼 such that 𝑥 is the only element of 𝑉𝛼 with 𝑉𝛼 ⊧
𝜑[𝑥, 𝛼0 , … , 𝛼𝑛−1 ].’
(a) Prove that if there are a (naive) ℒ∈ -formula Φ(𝑥) with pa-
rameters from Ord such that 𝑎 is defined by Φ(𝑥), that is,
Φ[𝒰] = {𝑎}, then 𝑎 is in OD.
[Hint: Use the Reflection Principle (Exercise 6.7.3).]
(b) Prove that there is a (naive) ℒ∈ -formula ΦOD (𝑥, 𝑦) with-
out parameters such that a set 𝑎 is in OD if and only if
there is an ordinal 𝛼 such that 𝑎 is defined by ΦOD (𝑥, 𝛼).
[Hint: Use that any #𝜑 is an ordinal and that any finite
sequence of ordinals may be coded by an ordinal (Exer-
cise 6.7.2).]
(c) Prove that the class of formulas in one free variable
with parameters from Ord may be well-ordered (cf. Exer-
cise 6.7.2) by an ∅-definable set-like relation. Deduce the
same for the class OD.
(2) The class of hereditarily ordinal definable sets is given by the
formula HOD(𝑥) which expresses that every element of tcl({𝑥})
is in OD.
(a) Prove that all ordinals as well as 𝑉𝜔 are in HOD.
(b) Prove that HOD is an ∅-definable transitive class, and that
a set 𝑎 is in HOD if and only if it is in OD and all its ele-
ments are in HOD.
(c) Prove that HOD ⊧ ZF.
(d) Prove that HOD ⊧(AC).
(e) Conclude that if ZF is consistent, then ZFC is consistent.
(3) (a) Prove that 𝑉𝜔HOD = 𝑉𝜔 .
(b) An ℒ∈ -sentence is said to be arithmetical if all its quan-
tifiers are relativized to 𝑉𝜔 . Prove that if an arithmetical
sentence is provable in ZFC, then it is provable in ZF.
6.7. Exercises 177
Exercise 6.7.7 (Fraenkel-Mostowski models I – independence of (AF)).

We work in ⟨𝑈; ∈⟩ = 𝒰 ⊧ ZF− . Let 𝐹 be a functional class which defines
a bijection between 𝑈 and itself. Set 𝑥 ∈′ 𝑦 ∶⇔ 𝑥 ∈ 𝐹(𝑦) and 𝒰 ′ =
⟨𝑈; ∈′ ⟩.
(1) Prove that 𝒰 ′ is a model of ZF− which satisfies (AC) if 𝒰 does.
(2) An atom is a set 𝑎 such that 𝑎 = {𝑎}. Prove that one may choose
𝐹 such that 𝒰 ′ contains an atom. If 𝒰 ⊧(AF), prove that there
is 𝐹 such that the set 𝒜 of atoms in 𝒰 ′ is in bijection with 𝜔.
−
(3) Prove that if ZF− (resp. ZFC ) is consistent, then so is ZF− +
−
¬(AF) (resp. ZFC + ¬(AF)).
Exercise 6.7.8 (Fraenkel-Mostowski models II – relative consistency of
¬(AC)). We work in 𝒰 ⊧ ZF− and assume that 𝒰 contains a non-empty
set of atoms 𝒜.
(1) By transfinite induction, define 𝑊0 = 𝒜, 𝑊𝛼+1 = 𝒫(𝑊𝛼 ), and
𝑊𝜆 = ⋃𝛼<𝜆 𝑊𝛼 for 𝜆 a limit ordinal. Let 𝑊 be the class defined
by the formula ∃𝛼 𝑥 ∈ 𝑊𝛼 .
(a) Prove that all 𝑊𝛼 are transitive sets and that one has 𝛼 ≤
𝛽 ⇒ 𝑊𝛼 ⊆ 𝑊𝛽 .
(b) Prove that 𝑊 ⊧ ZF− .
(c) Prove that an element 𝑎 ∈ 𝑊 is an atom if and only if
𝑎 ∈ 𝑊0 .
[Hint: One may use an appropriate notion of rank in 𝑊.]
(d) Prove that 𝑊 satisfies
∀𝑎(𝑎 ≠ ∅ → ∃𝑏 ∈ 𝑎 (𝑏 = {𝑏} ∨ ∀𝑥 ∈ 𝑎 (𝑥 ∉ 𝑏))).
(e) Prove that 𝒰 ⊧ ∀𝑥(𝑊 𝑊 (𝑥) ↔ 𝑊(𝑥)).
Now assume that the class of atoms of 𝒰 is given by the set 𝒜 which
is in bijection with 𝜔, and that 𝒰 ⊧ ∀𝑥𝑊(𝑥). (Note that if ZF− is consis-
tent, such a model of ZF− exists by the first part, combined with Exer-
cise 6.7.7.)
(2) Let 𝜎0 be a permutation of 𝑊0 = 𝒜.
(a) Prove that for any 𝛼, 𝜎0 extends uniquely to an automor-
phism 𝜎𝛼 of (𝑊𝛼 , ∈↾𝑊𝛼 ). Moreover, prove that if 𝛽 ≤ 𝛼,
then 𝜎𝛼 ↾𝑊𝛽 = 𝜎𝛽 .
(b) Deduce that there is a (unique) functional class 𝜎 which

extends 𝜎0 and induces a (naive) automorphism of 𝒰.
(c) Let Φ(𝑥1 , … , 𝑥𝑛 ) be an ℒ∈ -formula without parameters.
Prove that
𝒰 ⊧ ∀𝑥1 … ∀𝑥𝑛 [Φ(𝑥1 , … , 𝑥𝑛 ) ↔ Φ(𝜎(𝑥1 ), … , 𝜎(𝑥𝑛 ))].
(d) Prove that 𝜎(𝛼) = 𝛼 for every ordinal 𝛼.
(3) The class of ordinal and atom definable sets is given by the
ℒ∈ -formula OAD(𝑥) which expresses the following: ‘There
exists a formula 𝜑(𝑧, 𝑦0 , … , 𝑦𝑛−1 , 𝑧0 , … , 𝑧𝑚−1 ), an ordinal 𝛼,
a sequence (𝛼0 , … , 𝛼𝑛−1 ) of ordinals < 𝛼 and a sequence
(𝑎0 , … , 𝑎𝑚−1 ) of atoms such that 𝑥 is the only element of 𝑊𝛼
with 𝑊𝛼 ⊧ 𝜑[𝑥, 𝛼0 , … , 𝛼𝑛−1 , 𝑎0 , … , 𝑎𝑚−1 ].’
(a) Prove that if there is a (naive) ℒ∈ -formula Φ(𝑥) with pa-
rameters from Ord and 𝒜 which defines the set 𝑐, then 𝑐
is in OAD.
[Hint: For this and what follows, proceed as in Exer-
cise 6.7.6.]
(b) Prove that there is a (naive) ℒ∈ -formula ΦOAD (𝑥, 𝑦, 𝑧)
such that a set 𝑐 is in OAD if and only if there is an ordinal
𝛼 and a finite sequence of atoms 𝑠 such that 𝑐 is defined
by ΦOAD (𝑥, 𝛼, 𝑠).
(c) Let 𝑐 ⊆ 𝒜2 be a total order on 𝒜. Prove that 𝑐 is not in
OAD.
(4) The class of hereditarily ordinal and atom definable sets is given
by the formula HOAD(𝑥) which expresses that every element
of tcl({𝑥}) is in OAD.
(a) Prove that 𝒜 and every ordinal is in HOAD.
(b) Prove that HOAD ⊧ ZF− +¬(AC)+¬(AF).
(c) Conclude that if ZF− is consistent, so is the theory ZF− +
¬(AC)+¬(AF).
Bibliography
[1] S. Barry Cooper, Computability theory, Chapman & Hall/CRC, Boca Raton, FL, 2004. MR2017461
[2] René Cori and Daniel Lascar, Mathematical logic, Part 1, Oxford University Press, 2010.
[3] René Cori and Daniel Lascar, Mathematical logic, Oxford University Press, Oxford, 2001. A course
with exercises. Part II; Recursion theory, Gödel’s theorems, set theory, model theory; Translated
from the 1993 French original by Donald H. Pelletier; With a foreword to the original French edi-
tion by Jean-Louis Krivine and a foreword to the English edition by Wilfrid Hodges. MR1830848
[4] H.-D. Ebbinghaus, J. Flum, and W. Thomas, Mathematical logic, 2nd ed., Undergraduate Texts in
Mathematics, Springer-Verlag, New York, 1994. Translated from the German by Margit Meßmer.
MR1278260
[5] Thomas Jech, Set theory: The third millennium edition, revised and expanded, Springer Monographs
in Mathematics, Springer-Verlag, Berlin, 2003. MR1940513
[6] Jean-Louis Krivine, Théorie des ensembles, Nouvelle bibliothèque mathématique, Cassini, 1998.
[7] Kenneth Kunen, Set theory: An introduction to independence proofs, Studies in Logic and the Foun-
dations of Mathematics, vol. 102, North-Holland Publishing Co., Amsterdam-New York, 1980.
MR597342
[8] David Marker, Model theory: An introduction, Graduate Texts in Mathematics, vol. 217, Springer-
Verlag, New York, 2002. MR1924282
[9] Bruno Poizat, A course in model theory: An introduction to contemporary mathematical logic, Uni-
versitext, Springer-Verlag, New York, 2000. Translated from the French by Moses Klein and revised
by the author. MR1757487
[10] Hartley Rogers Jr., Theory of recursive functions and effective computability, 2nd ed., MIT Press,
Cambridge, MA, 1987. MR886890
[11] Robert I. Soare, Recursively enumerable sets and degrees: A study of computable functions and
computably generated sets, Perspectives in Mathematical Logic, Springer-Verlag, Berlin, 1987.
MR882921
[12] Katrin Tent and Martin Ziegler, A course in model theory, Lecture Notes in Logic, vol. 40, Associa-
tion for Symbolic Logic, La Jolla, CA; Cambridge University Press, Cambridge, 2012. MR2908005
[13] Martin Ziegler, Mathematische Logik (German), Mathematik Kompakt. [Compact Mathematics],
Birkhäuser Verlag, Basel, 2010. MR2683672
179
Index
¬, 34 ⟨𝑋⟩𝔄 , 41
∧, 34 ⟨𝑥0 , … , 𝑥𝑛−1 ⟩, 93
∨, 38 ⟨𝜑⟩, 66
→, 38 (𝑥)𝑖 , 93
↔, 38 𝑥−𝑦,̇ 91
∃𝑥, 34 (A1)-(A8), 125
∀𝑥, 38 (AC), 14, 155
∃!𝑥, 72 (AF), 153
≅, 35 (AI), 152
≡, 66 (CC), 172
𝐴 ⊆ 𝐵, 2 (CH), 19
𝔅 ⊆ 𝔄, 41 (DC), 172
≼, 66 (E1)-(E5), 47
𝔄 ⊧ 𝜑[𝛼], 39 (GCH), 19
𝔄 ⊧ 𝜑[𝑎1 , … , 𝑎𝑛 ], 40 (L1)-(L3), 139
⊧ 𝜑, 45 (Q1)-(Q3), 47
𝛿 ⊧ 𝐹, 46 (R0)∗ -(R3)∗ , 97
𝔄 ⊧ 𝑇, 48 (R0)-(R2), 90
𝑇 ⊧ 𝜑, 48 (R3), 98
𝑇 ⊢ℒ 𝜑, 49 𝟙𝑋 , 90
⊢ℒ 𝜑, 49 ℵ0 , 14
𝑇 ⊢ 𝜑, 53 ℵ𝛼 , 16
∼𝑇 , 77 ℶ𝛼 , 166
□𝑇 𝜑, 138 𝑛, 9
⌜ℳ⌝, 103 𝑡(𝑠1 , … , 𝑠𝑛 ), 45
#𝜑, 120 𝑡𝔄 [𝛼], 39
##𝑑, 122 ℕ, 2
181
182 Index
𝐶𝑚𝑛 , 90 Free(𝜑), 38
𝐷(𝔐), 70 HOAD, 178
𝐸 ∗ , 36 HOD, 176
𝐹[𝑎], 151 IC, 168
𝐼𝑝 , 104 Im(𝐹), 151
𝐿, 170 Lim(𝑥), 153
𝐿𝛼 , 170 MP, 49
𝑃𝑖𝑛 , 90 Neg, 133
𝑆, 90 OAD, 178
𝑇Pres , 141 OD, 176
𝑉, 158 Ord(𝑥), 152
𝑉𝛼 , 155 PA, 125
𝑍 𝔄 , 35 PA0 , 125
𝔄 ↾ ℒ , 45 Prf, 122
ℭ, 35 SatΣ1 (𝑣), 134
𝔑, 35 Taut, 121
𝔑st , 125 Term, 120
ℜ, 35, 68 Thm(𝑇), 122
𝒞 ℒ , 34 Th(𝔄), 50
ℱ, 90 ZFC, 148
ℱ ∗ , 97 ZFC− , 153
ℱ𝑛∗ , 97 ZF, 148
ℱ𝑛 , 90 ZF− , 153
ℱ𝑛ℒ , 34 Z, 175
ℒ𝑟𝑖𝑛𝑔 , 35 card(ℒ), 67
ℒOAG , 86 card(𝑋), 14
ℒPres , 141 cof(𝛼), 20
ℒ𝑎𝑟 , 35 dom(𝑅), 150
ℒ𝑜𝑟𝑑 , 35 dom(𝑓), 97
ℒ𝑜.𝑟𝑖𝑛𝑔 , 35 ht(𝑡), 36
ℒ𝑠𝑒𝑡 , 35 ht(𝜑), 37
𝒫(𝐴), 2 im(𝑅), 150
ℛ𝑛ℒ , 34 lg(𝑛), 94
𝒯 ℒ , 36 rk(𝑎), 158
𝒰, 147 rk𝐿 (𝑎), 170
ACF, 50, 78 rk𝑅 (𝑎), 173
ACF𝑝 , 80 subst, 131
Ax, 122 sup, 8
Card(𝑥), 152 tcl(𝑎), 157
Con𝑇 , 140 𝛼 + 𝛽, 11
DLO, 85 𝛼+ , 7
DOAG, 86 𝛼𝛽 , 11
Dom(𝐹), 151 𝛼𝑝 , 93
Form, 120 𝛼𝑎/𝑥 , 40
ℒ
Fml , 37 𝛼𝛽, 11
𝑝
Fml𝒫 , 46 𝛽𝑖 , 93
Index 183
Δ(𝔐), 71 cardinal multiplication, 16

Δ𝜑 , 131 Chevalley-Tarski Theorem, 79
𝜅 + 𝜆, 16 choice function, 155
𝜅+ , 15 Church’s Theorem, 132
𝜅𝜆 , 16 Church’s Thesis, 107
𝜅𝜆, 16 class, 148
𝜆𝑥1 ⋯ 𝑥𝑛 .𝑓(𝑥1 , … , 𝑥𝑛 ), 90 clause, 59
𝜇𝑦, 97 CLUB, 174
𝜇𝑡 ≤ 𝑧, 92 club, 24
𝜔, 9 cofinal subset, 20
𝜑(𝑥1 , … , 𝑥𝑛 ), 40 cofinality, 20
𝜑(𝑠1 , … , 𝑠𝑛 ), 45 compactness Theorem, 66
𝜑[𝔄, 𝑏], 41 complete diagram, 70
𝜑[𝔄], 40 complete theory, 50
𝜑𝐹 , 163 comprehension axiom scheme, 149
𝜑𝑋 , 163 configuration of a Turing machine, 104
𝜑𝑝 , 107 conservative expansion, 73
𝑝 consistent theory, 50
𝜑𝑖 , 107
𝜑𝑠/𝑥 , 42 constant symbol, 34
𝜎ℒ , 34 Continuum Hypothesis, 19
countable set, 15
absolute functional class, 164 Craig’s Interpolation Theorem, 60
absoluteness, 163
decidable theory, 122
Ackermann function, 94
deduction Lemma, 51
Aronszajn tree, 27
deduction rules, 48
Aronszajn’s Theorem, 27
deductively closed, 54
assignment, 39
Definability of satisfiability for
atomic formula, 36
Σ1 -formulas, 134
Ax’s Theorem, 82
definition by transfinite recursion, 154
axiom of choice, 14, 155
deterministic Turing machine, 99
axiom of countable choice, 172
diagonal argument, 131
axiom of dependent choice, 172
Downward Löwenheim-Skolem
axiomatizable property, 69
Theorem, 67
base set, 35 effectively axiomatizable theory, 122
Beth Definability Theorem, 61 elementary equivalent structures, 66
bound occurrence, 38 elementary extension, 66
bounded 𝜇-operator, 92 elementary substructure, 66
branch, 27 equality axioms, 47
equinumerous, 3
Cantor normal form, 22 equivalent formulas, 62, 73
Cantor’s Theorem, 2 expansion by definition, 73
Cantor-Bernstein Theorem, 3 expansion of a structure, 45
cardinal, 14 extensionality axiom, 148
cardinal addition, 16
cardinal exponentiation, 16 filter, 25
184 Index
filter basis, 25 inductive set, 14

finite ordinal, 9 infimum, 3
finitely axiomatizable property, 69 infinity axiom, 152
first-order language, 34 interpretation of a symbol, 35
fixed point Theorem, 131 isomorphism of ordered sets, 5
Fodor’s Lemma, 24 isomorphism of structures, 35
formal proof, 49
formula, 37 König’s Lemma, 27
Δ0 -formula, 144, 163 König’s Theorem, 19
Σ1 -formula, 127 Kalmár elementary functions, 115
∀∃-formula, 82 Kleene normal form, 106
strict Σ1 -formula, 127 Kleene’s Fixed Point Theorem, 108
foundation axiom, 153
Fraenkel-Mostowski models, 177 language, 34
free occurrence, 37 largest element, 3
free variable, 38 Lefschetz Principle, 80
function represented by a formula, 128 length of a word, 36
function symbol, 34 limit ordinal, 8
literal, 59
generalization, 49 Loeb’s axioms, 139
Generalized Continuum logical axioms, 48
Hypothesis, 19 logical consequence of a theory, 48
Gödel number, 120 logical symbols, 34
Gödel’s 𝛽-function, 113 logically equivalent formulas, 62, 73
Gödel’s Completeness Theorem, 53 Łoś’s Theorem, 84
Gödel’s First Incompleteness Theorem, lower bound, 3
133
Gödel’s Second Incompleteness maximal element, 3
Theorem, 140 minimal element, 3
Goodstein sequence, 22 model of a theory, 48
modus ponens, 49
Halting Problem, 111 Mostowski collapse, 173
Hausdorff’s Theorem, 26
height of a formula, 37 non-principal ultrafilter, 25
height of a term, 36
Henkin witnesses, 54 Omitting Types Theorem, 63
Herbrand normal form, 62 order topology, 23
hereditarily definable set, 176 ordered sum, 5
Hessenberg’s Theorem, 17 ordinal, 7
ℵ-hierarchy, 16, 154 ordinal addition, 11
Hilbert’s Nullstellensatz, 81 ordinal definable set, 176
Hindman’s Theorem, 28 ordinal exponentiation, 13
ordinal multiplication, 12
idempotent ultrafilter, 29 overspill Lemma, 130
inaccessible cardinal, 166
inconsistent theory, 50 pairing axiom, 149
indiscernible sequence, 88 partial function, 97
Index 185
partial order, 3 successor ordinal, 8

partial recursive function, 97 supremum, 3
Peano axioms, 125
power set axiom, 150 Tarski’s Theorem on the
prenex normal form, 62 Non-definability of Truth, 132
Presburger arithmetic, 141 Tarski-Vaught Test, 67
primitive recursive function, 90 tautology, 46
primitive recursive set, 91 Tenenbaum’s Theorem, 143
principal ultrafilter, 25 term, 36
Theorem of the Complement, 111
quantifier axioms, 47 theory, 48
quantifier elimination, 77 total order, 3
transfinite induction, 10, 153
Ramsey’s Theorem, 58 transitive set, 7
rank of a set, 158 tree, 27
recursive function, 97 Turing computable partial function,
#-recursive function, 113 100
recursive theory, 122 Turing machine, 98
recursively axiomatizable theory, 122
recursively enumerable set, 109 Ulam matrix, 23
reduct of a structure, 45 Ulam’s Theorem, 23
reflection principle, 174 ultrafilter, 25
refutable, 59 undecidability of the predicate
regular cardinal, 21 calculus, 132
relation symbol, 34 union axiom, 150
relative consistency, 162 unique reading of formulas, 37
relativization, 163 unique reading of terms, 36
replacement axiom scheme, 151 universal partial recursive function,
representability Theorem, 128 108
reverse lexicographic product, 5 universally valid, 45
Rice’s Theorem, 112 universe, 147
root, 27 upper bound, 3
Upward Löwenheim-Skolem Theorem,
satisfaction of a formula, 39 71
sentence, 38
set represented by a formula, 128 Vaught’s Criterion, 83
signature, 34 von Neumann hierarchy, 155
simple diagram, 71
situation of a Turing machine, 104 well-founded, 4
Skolem’s Paradox, 157 well-order, 4
smallest element, 3 word, 36
Solovay’s Theorem, 24 Zermelo axioms, 175
strong limit cardinal, 166 Zermelo’s Theorem, 14
structure, 35 Zermelo-Fraenkel axioms, 148
subformula, 62 Zorn’s Lemma, 14
subnumerous, 3
substitution, 41
Selected Published Titles in This Series
89 Martin Hils and François Loeser, A First Journey through Logic, 2019
88 M. Ram Murty and Brandon Fodden, Hilbert’s Tenth Problem, 2019
87 Matthew Katz and Jan Reimann, An Introduction to Ramsey Theory,
2018
86 Peter Frankl and Norihide Tokushige, Extremal Problems for Finite
Sets, 2018
85 Joel H. Shapiro, Volterra Adventures, 2018
84 Paul Pollack, A Conversational Introduction to Algebraic Number
Theory, 2017
83 Thomas R. Shemanske, Modern Cryptography and Elliptic Curves, 2017
82 A. R. Wadsworth, Problems in Abstract Algebra, 2017
81 Vaughn Climenhaga and Anatole Katok, From Groups to Geometry
and Back, 2017
80 Matt DeVos and Deborah A. Kent, Game Theory, 2016
79 Kristopher Tapp, Matrix Groups for Undergraduates, Second Edition,
2016
78 Gail S. Nelson, A User-Friendly Introduction to Lebesgue Measure and
Integration, 2015
77 Wolfgang Kühnel, Differential Geometry: Curves — Surfaces —
Manifolds, Third Edition, 2015
76 John Roe, Winding Around, 2015
75 Ida Kantor, Jiřı́ Matoušek, and Robert Šámal, Mathematics++,
2015
74 Mohamed Elhamdadi and Sam Nelson, Quandles, 2015
73 Bruce M. Landman and Aaron Robertson, Ramsey Theory on the
Integers, Second Edition, 2014
72 Mark Kot, A First Course in the Calculus of Variations, 2014
71 Joel Spencer, Asymptopia, 2014
70 Lasse Rempe-Gillen and Rebecca Waldecker, Primality Testing for
Beginners, 2014
69 Mark Levi, Classical Mechanics with Calculus of Variations and Optimal
Control, 2014
68 Samuel S. Wagstaff, Jr., The Joy of Factoring, 2013
67 Emily H. Moore and Harriet S. Pollatsek, Difference Sets, 2013
For a complete list of titles in this series, visit the

AMS Bookstore at www.ams.org/bookstore/stmlseries/.
The aim of this book is to present
mathematical logic to students
who are interested in what this
field is but have no intention of
specializing in it. The point of
view is to treat logic on an equal
footing to any other topic in the
mathematical curriculum. The book starts with a presentation of naive set
theory, the theory of sets that mathematicians use on a daily basis. Each
subsequent chapter presents one of the main areas of mathematical logic:
first order logic and formal proofs, model theory, recursion theory, Gödel’s
incompleteness theorem, and, finally, the axiomatic set theory. Each chapter
includes several interesting highlights — outside of logic when possible —
either in the main text, or as exercises or appendices. Exercises are an essential
component of the book, and a good number of them are designed to provide
an opening to additional topics of interest.
For additional information

and updates on this book, visit
www.ams.org/bookpages/stml-89
STML/89

STML 089

Uploaded by

Copyright:

Available Formats

STML 089

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

STML 089

Uploaded by

Copyright:

Available Formats

S T U D E N T M AT H E M AT I C A L L I B R A RY

2010 Mathematics Subject Classification. Primary 03-01.

For additional information and updates on this book, visit

Library of Congress Cataloging-in-Publication Data

© 2019 by the American Mathematical Society. All rights reserved.

Chapter 1. Counting to Infinity 1

Chapter 2. First-order Logic 33

§2.1. Languages and Structures 34

Chapter 3. First Steps in Model Theory 65

Chapter 4. Recursive Functions 89

Chapter 5. Models of Arithmetic and Limitation Theorems 119

§5.2. Decidable Theories 122

Chapter 6. Axiomatic Set Theory 147

1.1. Naive Set Theory

without further questioning. It is only in the last chapter that we shall

1.2. The Cantor and Cantor-Bernstein Theorems

Proof. Let 𝑓 ∶ 𝐴 → 𝒫(𝐴) be a map. Consider the set

Theorem 1.2.2 (Cantor-Bernstein). Let 𝐴 and 𝐵 be sets, and let 𝑓 ∶ 𝐴 →

Proof. We may assume 𝐴 is a subset of 𝐵 and 𝑓 is the inclusion map.

Using this terminology, the Cantor-Bernstein Theorem may be re-

Definition. A partial order < on a set 𝑋 is a binary relation (that is,

One writes 𝑥 ≤ 𝑦 to mean 𝑥 < 𝑦 or 𝑥 = 𝑦, 𝑥 > 𝑦 for 𝑦 < 𝑥, and 𝑥 ≥ 𝑦

Note that in a partial order, if 𝑎 ≤ 𝑏 and 𝑏 ≤ 𝑎 then 𝑎 = 𝑏. Thus

Proof. The proofs of (1) and (3) are left as exercises.

1.4. Operations on Orders

By an isomorphism between two partially ordered sets 𝑋 and 𝑌 we

Let 𝑋 and 𝑌 be two well-founded partially ordered sets. Let 𝑍 be a non-

Definition. Let 𝑋 and 𝑌 be totally ordered sets. We assume that 𝑋 ad-

Proposition 1.4.2. Let 𝑋, 𝑌 and 𝑍 be totally ordered sets, and assume

The set 𝑍1 is an initial segment of 𝑍1′ . If 𝑍1 contains the function with

1.5. Ordinal Numbers

Proof. (1) is clear. For (2), one considers 𝑥 ∈ 𝛼 minimal. If 𝑦 ∈ 𝑥,

𝑦 ∈ 𝛽, then 𝑦 ∈ 𝑥 since otherwise 𝑥 ∈ 𝑦 and 𝑥 ∈ 𝛽. Hence 𝛽 = 𝑆<𝑥 =

Proof. The intersection of a family of transitive sets is transitive, and the

An ordinal of the form 𝛼+ is called a successor ordinal. It is clear

Proposition 1.5.5. For an ordinal 𝜆 ≠ ∅, the following conditions are

Proof. (1)⇒(2). Let 𝛽 = ⋃𝛼<𝜆 𝛼 and 𝜆 a limit. It is clear that 𝛽 ⊆ 𝜆.

Proof. One proves first, by induction on 𝑛 ∈ ℕ, that all elements of 𝜔

Proof. If there exists 𝛽 ∈ 𝛼 with 𝑓(𝛽) < 𝛽, we consider 𝛽0 minimal with

The statement about an isomorphism 𝑓 follows by applying the re-

Proof. Uniqueness follows from Lemma 1.5.8. To prove existence, let

It is clear that 𝑓 is well defined and induces an isomorphism of ordered

1.6. Ordinal Arithmetic

Proposition 1.6.1 (Ordinal addition). Let 𝛼, 𝛽 and 𝛾 be ordinals.

Proof. Uniqueness: Assume 𝛼𝜇 + 𝜌 = 𝛼𝜇′ + 𝜌′ with 𝜌, 𝜌′ < 𝛼. If 𝜇 < 𝜇′ ,

Proposition 1.6.4 (Ordinal exponentiation). Let 𝛼, 𝛽, 𝛾 be ordinals.

(1) For every 𝛼, we have 𝛼0 = 1, 𝛼1 = 𝛼 and 1𝛼 = 1. If 𝛼 ≠ 0, then

Remark 1.6.5. The following formulas would allow us to define ordinal

• 𝛼 + 0 = 𝛼, 𝛼 + 𝛽 + = (𝛼 + 𝛽)+ , and 𝛼 + 𝜆 = sup𝛽<𝜆 (𝛼 + 𝛽) for

1.7. The Axiom of Choice

In the Zermelo-Fraenkel system of axioms ZF, (AC) is equivalent

1.8. Cardinal Numbers

Proof. By Zermelo’s Theorem and Theorem 1.5.9, 𝑋 is equinumerous

Proof. (1)⇒(2) is easy.

Proof. If 𝛼 < 𝜆, then 𝛼 < 𝜅 for some 𝜅 ∈ 𝑋. Since 𝜅 is a cardinal, we

There is no largest cardinal. Indeed, if 𝜅 is a cardinal, then 𝜆 ∶=

Definition. The ℵ-hierarchy assigns to any ordinal a cardinal as fol-

By transfinite induction, one proves that 𝛼 < 𝛽 ⇒ ℵ𝛼 < ℵ𝛽 . In com-

Proof. Let 𝜅 be an infinite cardinal. The function 𝛽 ↦ ℵ𝛽 is strictly