DRAFT– 31 DEC 2021

Beginning Mathematical Logic

A Study Guide

Peter Smith

LOGIC MATTERS
 DRAFT– 31 DEC 2021
© Peter Smith 2021

All rights reserved. Permission is granted to distribute this PDF as a

complete whole, including this copyright page, for educational
purposes such as classroom use. Otherwise, no part of this publication
may be reproduced, distributed, or transmitted in any form or by any
means, including photocopying or other electronic or mechanical
methods, without the prior written permission of the author, except
in the case of brief quotations embodied in critical reviews and certain
other noncommercial uses permitted by copyright law. For permission
requests, write to peter smith@logicmatters.net.

First published in this format, Logic Matters 2021

This version: December 31, 2021

The latest version of this Guide is always available at logicmatters.net/tyl

This is still a preliminary draft version. Please do send corrections, however

small, and/or suggestions for improvement to peter smith@logicmatters.net,
giving the date of the version you are commenting on!
 DRAFT– 31 DEC 2021

Contents

Preface vii
1 The Guide, and how to use it 1
1.1 Who is the Guide for? 1
1.2 The Guide’s structure 1
1.3 Strategies for self-teaching from logic books 4
1.4 Choices, choices 5
1.5 So what do you need to bring to the party? 6
1.6 Two notational conventions 7
2 A very little informal set theory 8
2.1 Sets: a checklist of some basics 8
2.2 Recommendations on informal basic set theory 11
2.3 Virtual classes, real sets 12
3 First-order logic 14
3.1 FOL: a general overview 14
3.2 A little more about types of proof-system 20
3.3 Basic recommendations for reading on FOL 24
3.4 Some parallel and slightly more advanced reading 26
3.5 A little history (and some philosophy too) 30
3.6 Postscript: Other treatments? 32
4 Second-order logic, quite briefly 37
4.1 A preliminary note on many-sorted logic 37
4.2 Second-order logic: a brief overview 39
4.3 Recommendations on many-sorted and second-order logic 44
4.4 Conceptual issues 45
5 Model theory 46
5.1 Elementary model theory: an overview 46
5.2 Recommendations for beginning first-order model theory 52
5.3 Some parallel and slightly more advanced reading 54
5.4 A little history 56
6 Arithmetic, computability, and incompleteness 58
6.1 Logic and computability 58
6.2 Computable functions: an overview 60
6.3 Formal arithmetic: an overview 62

iii
 DRAFT– 31 DEC 2021
Contents

6.4 Towards Gödelian incompleteness 64

6.5 Main recommendations on arithmetic, etc. 66
6.6 Some parallel/additional reading 69
6.7 A little history 72
7 Set theory, less naively 73
7.1 Elements of set theory: an overview 73
7.2 Main recommendations on set theory 80
7.3 Some parallel/additional reading on standard ZFC 83
7.4 Further conceptual reflection on set theories 85
7.5 A little more history 86
7.6 Postscript: Other treatments? 86
8 Intuitionistic logic 88
8.1 A formal system 88
8.2 Overview: why intuitionistic logic? 89
8.3 Overview: more proof theory, more semantics 92
8.4 Basic recommendations on intuitionistic logic 96
8.5 Some parallel/additional reading 97
8.6 A little more history, a little more philosophy 98
9 Elementary proof theory 100
9.1 Preamble: a very little about Hilbert’s Programme 100
9.2 Deductive systems, normal forms, and cuts: a short overview 101
9.3 Proof theory and the consistency of arithmetic: a short overview 108
9.4 Main recommendations on elementary proof theory 112
9.5 Some parallel/additional reading 115
10 Modal logics 118
10.1 Some basic modal logics 118
10.2 Provability logic 125
10.3 First readings on modal logic 129
10.4 Suggested readings on provability logic 130
10.5 Alternative and further readings on modal logics 131
10.6 Finally, a very little history 133
11 Other logics? 135
11.1 Relevant logics 135
11.2 Readings on relevant logic 141
11.3 Free Logic 142
11.4 Readings on free logic 145
11.5 Plural logic 146
11.6 Readings on plural logic 146
12 Going further 148
12.1 A very little light algebra for logic? 149
12.2 More model theory 150
12.3 More on formal arithmetic and computability 153
12.4 More on mainstream set theory 158
12.5 Choice, and the choice of set theory 161

iv
 DRAFT– 31 DEC 2021
Contents

12.6 More proof theory 166

12.7 Higher-order logic, the lambda calculus, and type theory 167

Index of authors 171

v
DRAFT– 31 DEC 2021
 DRAFT– 31 DEC 2021

Preface

This is not another textbook on mathematical logic: it is a Study Guide, a book

mostly about textbooks on mathematical logic. Its purpose is to enable you to
locate the best resources for teaching yourself various areas of logic, at a fairly
introductory level. Inevitably, given the breadth of its coverage, the Guide is
rather long: but don’t let that scare you off! There is a good deal of signposting
and there are also explanatory overviews to enable you to pick your way through
and choose the parts which are most relevant to you.
Beginning Mathematical Logic is a descendant of my much-downloaded Teach
Yourself Logic. The new title highlights that the Guide focuses mainly on the
core mathematical logic curriculum. It also signals that I do not try to cover
advanced material in any detail.
The first chapter says more about who the Guide is intended for, what it covers,
and how to use it. But let me note straightaway that most of the main reading
recommendations do indeed point to published books. True, there are quite a lot
of on-line lecture-notes that university teachers have made available. Some of
these are excellent. However, they do tend to be terse, and often very terse (as
entirely befits material originally intended to support a lecture course). They
are therefore usually not as helpful as fully-worked-out book-length treatments
for students needing to teach themselves.
So where can you find the titles mentioned here? I suppose I ought to pass
over the issue of downloading books from certain very well-known and extremely
well-stocked copyright-infringing PDF repositories. That’s between you and your
conscience (though almost all the books are indeed available to be sampled
there). Anyway, many do prefer to work from physical books. Most of these
titles should in fact be held by any large-enough university library which has been
trying over the years to maintain core collections in mathematics and philosophy
(and if the local library is too small, books should be borrowable through some
inter-library loans system).
Since I’m not assuming that you will be buying the recommended books,
I have not made cost or being currently in print a significant consideration.
However, I have marked with a star* books that are available new or second-
hand at relatively inexpensively (or at least are unusually good value for the
length and/or importance of the book). When e-copies of books are freely and
legally available, links are provided. Where journal articles or encyclopaedia

vii
 DRAFT– 31 DEC 2021
Preface

entries have been recommended, these can almost always be freely downloaded,
and again I give links.
Before I retired from the University of Cambridge, it was my greatest good
fortune to have secure, decently paid, university posts for forty years in leisurely
times, with almost total freedom to follow my interests wherever they meandered.
Like most of my contemporaries, for much of that time I didn’t really appreciate
how extraordinarily lucky I was. In writing this Study Guide and making it
readily available, I am trying to give a little back by way of heartfelt thanks. I
hope you find it useful.1

1 Many, many thanks are due to all those who commented on versions of Teach Yourself
Logic over more than a decade. Further comments and suggestions for future editions of
this revised Guide will always be gratefully received.

viii
 DRAFT– 31 DEC 2021

1 The Guide, and how to use it

Who is this Study Guide for? What does it cover? At what level? How should
the Guide be used? And what background knowledge do you need, in order to
make use of it? This preliminary chapter explains.

1.1 Who is the Guide for?

It is a depressing phenomenon. Relatively few mathematics departments have
undergraduate courses on mathematical logic. And serious logic is taught less
and less in philosophy departments too.
Yet logic itself is, of course, no less exciting and rewarding a subject than it
ever was. So how is knowledge to be passed on if there are not enough courses, or
indeed if there are none at all? It seems that many will need to teach themselves
from books, either solo or by organizing their own study groups (local or online).
In a way, this is perhaps no real hardship; there are some wonderful books
written by great expositors out there. But what to read and work through? Logic
books can have a very long shelf life, and you shouldn’t at all dismiss older texts
when starting out on some topic area. There’s more than a sixty year span of
publications to select from, which means that there are hundreds of good books
to choose from.
That’s why students – whether mathematicians or philosophers – wanting to
learn some logic by self-study will need a Guide like this if they are to find
their way around the very large literature old and new, with the aim of teaching
themselves enjoyably and effectively. And even those fortunate enough to be
offered courses might very well appreciate advice on entry-level texts which they
can usefully read in preparation or in parallel.
There are other students too who will have interests in areas of logic, e.g.
theoretical linguists and computer scientists. But this Guide isn’t written with
them much in mind.

1.2 The Guide’s structure

There is another preliminary chapter after this one, Chapter 2 on ‘naive’ set
theory, which reviews the concepts and constructions typically taken for granted
in quite elementary mathematical writing (not just in texts about logic). But

1
 DRAFT– 31 DEC 2021
1 The Guide, and how to use it

then we start covering the usual mathematical logic curriculum, at roughly an

upper undergraduate level.
The standard menu of core topics has remained fairly fixed ever since e.g.
Elliott Mendelson’s justly famous Introduction to Mathematical Logic (1st edn.,
1964), and these are explored in Chapters 3 to 7. The following four chapters
then look at other logical topics, still at about the same level. The final chapter
of the Guide glances ahead at more advanced-level readings on the core areas,
and briefly gestures towards one last topic.
(a) In more detail, then,
Chapter 3 discusses classical first-order logic (FOL), which is at the fixed centre
of any mathematical logic course.
The remaining chapters all depend on this crucial one and assume some knowl-
edge of it, as we discuss the use of classical FOL in building formal theories, or
we consider extensions and variants of this logic.
Now, there is one extension worth knowing just a little about straight away
(in order to understand some themes touched on in the next few chapters). So:
Chapter 4 goes beyond first-order logic by briefly looking at second-order logic.
(Second-order languages have more ways of forming general propositions
than first-order ones.)
You can then start work on the topics of the following three key chapters in
whichever order you choose:
Chapter 5 introduces a modest amount of model theory which, roughly speaking,
explores how formal theories relate to the structures they are about.
Chapter 6 looks at one particular kind of formal theory, i.e. formal arithmetics,
and relatedly explores the theory of computable functions. We arrive at
proofs of epochal results such as Gödel’s incompleteness theorems.
Chapter 7 is on set theory proper – starting fairly informally, examining basic
notions of cardinals and ordinals, constructions of number systems in set
theory, the role of the axiom of choice, etc. We then look at the standard
formal axiomatization, i.e. first-order ZFC, and nod towards alternatives.
Now, as well as second-order logic, there is another variant of FOL which
is often mentioned in introductory mathematical logic texts, and that you will
want to know something about at this stage. So
Chapter 8 introduces intuitionistic logic, which drops the classical principle that,
whatever proposition we take, either it or its negation is true. But why
might we want to do that? What differences does it make?
And this topic can’t really be sharply separated from another whole area of logic
which can be under-represented in many textbooks; that is why
Chapter 9 takes a first look at proof theory. OK, this is a pretty unhelpful label
given that most areas of logic deal with proofs! – but it conventionally

2
 DRAFT– 31 DEC 2021
The Guide’s structure

points to a cluster of issues about the structure of proofs and the consis-
tency of theories, etc.
(b) Now, a quick glance at e.g. the entry headings in The Stanford Encyclopedia
of Philosophy reveals that philosophers have been interested in a wide spectrum
of other logics, ranging far beyond classical and intuitionistic versions of FOL
and their second-order extensions. And although this Guide – as its title suggests
– is mainly focussed on core topics in mathematical logic, it is worth pausing to
consider just a few of those variant types of logic.
First, in looking at intuitionist logic, you will already have met a new way of
thinking about the meanings of the logical operators, using so-called ‘possible-
world semantics’. We can now usefully explore this idea further, since it has
many other applications. So:
Chapter 10 discusses modal logics, which deploy possible-world semantics, ini-
tially to deal with various notions of necessity and possibility. In general,
these modal logics are perhaps of most interest to philosophers. However,
there is one particular variety which it is good for any logician to know
about, namely provability logic, which (roughly speaking) explores the logic
of operators like ‘it is provable in formal arithmetic that . . . ’.
Second, standard FOL (classical or intuitionistic) can be criticized in various
ways. For example, (1) it allows certain arguments to count as valid even when
the premisses are irrelevant to the conclusion; (2) it is not as neutral about exis-
tence assumptions as we might suppose a logic ought to be; and (3) it can’t cope
naturally with terms denoting more than one thing like ‘Russell and Whitehead’
and ‘the roots of the quintic equation E’. It is worth saying something about
these supposed shortcomings. So:
Chapter 11 discusses so-called relevant logics (where we impose stronger require-
ments on the relevance of premisses to conclusions for valid arguments),
free logics (i.e. logics free of existence assumptions, where we no longer
presuppose that e.g. names in an interpreted formal language always ac-
tually name something), and plural logics (where we can e.g. cope with
plural terms).
For reasons I’ll explain, these variant logics are indeed mostly of concern to
philosophers. Though any logician interested in the foundations of mathematics
should want to know more about the pros and cons of dealing with talk about
pluralities by using set theory vs second-order logic vs plural logic.
(c) How are these chapters from Chapter 3 onwards structured?
Each starts with one or more overviews of its topic area(s). These overviews
are not full-blown tutorials or mini encylopedia-style essays – they are simply
intended to give helpful introductions, with some rough indications of what
the chapters are about. And I don’t pretend that the level of coverage in the
overviews is uniform. If you already know something of the topic, or if these
necessarily brisk arm-waving descriptions sometimes mystify, feel very free to
skim or skip as much you like.

3
 DRAFT– 31 DEC 2021
1 The Guide, and how to use it

Overviews are then followed by the key sections, giving a list of main recom-
mended texts for the chapter’s topic(s), put into what strikes me as a sensible
reading order.
I next offer some suggestions for alternative/additional reading at about the
same level or only another half a step up in difficulty/sophistication.
And because it can be quite illuminating to know just a little of the background
history of a topic, most chapters end with a few brisk suggestions for reading on
that.
(d) This is primarily a Guide to beginning mathematical logic. So the recom-
mended introductory readings in Chapters 1 to 11 won’t take you very far. But
they should be more than enough to put you in a position from which you can
venture into rather more advanced work under your own steam. Still, I have
added a final chapter which looks ahead:
Chapter 12 offers suggestions for those who want to delve further into the topics
of some earlier core chapters, in particular looking again at model theory,
computability and arithmetic, set theory, and proof theory. Then I add a
final section on a new topic, type theories and the lambda calculus, a focus
of much recent interest.
Very roughly, if the earlier chapters are at advanced undergraduate level (or a
little more), this last one is definitely at graduate level.

1.3 Strategies for self-teaching from logic books

As I said in the Preface, one major reason for the length of this Guide is its
breadth of coverage. But there is another significant reason, connected to a
point which I now really want to highlight:
I very strongly recommend tackling a new area of logic by reading a
variety of texts, ideally a series of books which overlap in level (with
the next one in the series covering some of the same ground and then
pushing on from the previous one).
In fact, I probably can’t stress this bit of advice too much (which, in my experi-
ence, applies equally to getting to grips with any new area of mathematics). This
approach will really help to reinforce and deepen understanding as you encounter
and re-encounter the same material, coming at it from somewhat different angles,
with different emphases.
Exaggerating only a little, there are many instructors who say ‘This is the
textbook we are using/here is my set of notes: take it or leave it’. But you will
always gain from looking at a number of different treatments, perhaps at rather
different levels. The multiple overlaps in coverage in the reading lists in later
chapters, which contribute to making the Guide as long as it is, are therefore
fully intended. They also mean that you should always be able to find the options
that best suit your degree of mathematical competence and your preferences as
to textbook style.
4
 DRAFT– 31 DEC 2021
Choices, choices

To repeat: you will certainly miss a lot if you concentrate on just one text
in a given area, especially at the outset. Yes, do very carefully read one or two
central texts, chosing books that work for you. But do also cultivate the crucial
habit of judiciously skipping and skimming through a number of other works so
that you can build up a good overall picture of an area seen from various angles
and levels of approach.
While we are talking about strategies for self-teaching, I suppose I should add a
quick remark on the question of doing exercises.

Mathematics is, as they say, not merely a spectator sport: so you

should try some of the exercises in the books as you read along,
in order to check and reinforce comprehension. On the other hand,
don’t obsess about this, and do concentrate on the exercises that look
interesting and/or might deepen understanding.

Note that some authors have the irritating(?) habit of burying quite important
results among the exercises, mixed in with routine homework. It is therefore
always a good policy to skim through the exercises in a book even if you don’t
plan to work on answers to very many of them.

1.4 Choices, choices

How have I decided which texts to recommend?
An initial point. If I were choosing a text book around which to shape a lecture
course on this or that area of mathematical logic, I would no doubt be looking at
many of the same books that I mention later; but my preference-rankings could
well be rather different. So, to emphasize, the main recommendations in this
Guide are for books which I think should be particularly good for self-studying
logic, without the benefit of expansive classroom introductions and additional
explanations.
Different people find different expository styles congenial. What is agreeably
discursive for one reader might be irritatingly slow-moving for another. For my-
self, I do particularly like books that are good at explaining the ideas behind the
various formal technicalities while avoiding needless early complications, exces-
sive hacking through routine detail, or misplaced ‘rigour’. So I prefer a treatment
that highlights intuitive motivations and doesn’t rush too fast to become too ab-
stract: this is surely what we particularly want in books to be used for self-study.
(There’s a certain tradition of masochism in older maths writing, of going for
brusque formal abstraction from the outset with little by way of explanatory
chat: this is quite unnecessary in other areas, and just because logic is all about
formal theories, that doesn’t make it any more necessary here.)
The selection of readings in the following chapters reflects these tastes. But
overall, while I have no doubt been opinionated, I don’t think that I have been
very idiosyncratic: indeed, in many respects I have probably been really rather
conservative in my choices. So nearly all the readings I recommend will be very
5
 DRAFT– 31 DEC 2021
1 The Guide, and how to use it

widely agreed to have significant virtues (even if other logicians would have
different favourites).

1.5 So what do you need to bring to the party?

There is no specific knowledge you need before tackling the main recommended
books on FOL. And in fact none of the more introductory books recommended
in other chapters except the last requires very much ‘mathematical maturity’.
So mathematics students from mid-year undergraduates up should be able to
just dive in and explore.
What about philosophy students without any mathematical background? It will
certainly help to have done an introductory logic course based on a book at
the level of my own Introduction to Formal Logic* (2nd edition, CUP, 2020;
now freely downloadable from logicmatters.net/ifl), or Nicholas Smith’s excellent
Logic: The Laws of Truth (Princeton UP 2012). And non-mathematicians could
very usefully broaden their informal proof-writing skills by also looking at this
much-used and much-praised book:
1. Daniel J. Velleman, How to Prove It: A Structured Approach* (CUP, 3rd
edition, 2019).
From the Preface: “Students . . . often have trouble the first time that
they’re asked to work seriously with mathematical proofs, because they
don’t know ‘the rules of the game’. What is expected of you if you are
asked to prove something? What distinguishes a correct proof from an
incorrect one? This book is intended to help students learn the answers
to these questions by spelling out the underlying principles involved in
the construction of proofs.” There are chapters on the propositional con-
nectives and quantifiers, and on key informal proof-strategies for using
them; there are chapters on relations and functions, a chapter on math-
ematical induction, and a final chapter on infinite sets (countable vs.
uncountable sets).
This is a truly excellent student text; at least skip and skim through
the book, taking what you need (perhaps paying especial attention to
the chapter on mathematical induction).
For a much less conventional text than Velleman’s, with a different emphasis,
you might also be both instructed and entertained by
2. Joel David Hamkins, Proof and the Art of Mathematics* (MIT Press,
2020).
From the blurb: “This book offers an introduction to the art and
craft of proof-writing. The author . . . presents a series of engaging and
compelling mathematical statements with interesting elementary proofs.
These proofs capture a wide range of topics . . . The goal is to show
students and aspiring mathematicians how to write [informal!] proofs
with elegance and precision.”
6
 DRAFT– 31 DEC 2021
Two notational conventions

This is attractively written (though it is occasionally uneven in level and tone).

Readers with very little mathematical background could still enjoy dipping into
this, and will learn a good deal, e.g. about proofs by induction. Lots of striking
and memorable examples.

1.6 Two notational conventions

Finally, let me highlight two points of notation.
First, it is helpful to adopt here the following convention for distinguishing
two different uses of letters as variables:

Italic letters, as in A, F , n, x, will always be used just as part of our

informal logicians’ English, typically as place-holders or in making gen-
eralizations. Occasionally, Greek capital letters will also be used equally
informally for sets (in particular, for sets of sentences).
Sans-serif letters by contrast, as in P, F, n, x, are always used as sym-
bols belonging to some particular formal language, an artificial language
cooked-up by logicians.

For example, we might talk in logician’s English about a logical formula being
of the shape (A ∨ B), using the italic letters as place-holders for sentences. And
then (P ∨ Q), a formula from a particular logical language, could be an instance,
with these sans-serif letters being sentences of the relevant language. Similarly,
x + 0 = x might be an equation of ordinary informal arithmetic, while x + 0 = x
will be an expression belonging to a formal theory of arithmetic.
Our second convention, just put into practice, is that we will not in general
be using quotation marks when mentioning symbolic expressions. Logicians can
get very pernickety, and insist on the use of quotation marks in order to make
it extra clear when we are mentioning an expression of, say, formal arithmetic
in order to say something about that expression itself as opposed to using it to
make an arithmetical claim. But in the present context it is unlikely you will be
led astray if we just leave it to context to fix whether a symbolic expression is
being mentioned rather than put to use.

7
 DRAFT– 31 DEC 2021

2 A very little informal set theory

Notation, concepts and constructions from entry-level set theory are very often
presupposed in elementary mathematical texts – including some of the introduc-
tory logic texts mentioned in the following chapters, even before we get round to
officially studying set theory itself. If the absolute basics aren’t already familiar
to you, it is worth pausing to get acquainted at an early stage.
In §2.1, then, I note what you should ideally know about sets here at the
outset. It isn’t a lot! And for now, we proceed ‘naively’ – i.e. we proceed quite
informally, and will just assume that the various constructions we talk about
are permitted, etc. §2.2 gives recommended readings on basic informal set theory
for those who need them. In §2.3 I point out that, while the use of set-talk in
elementary contexts is conventional, it many cases it can in fact be eliminated
without serious loss.

2.1 Sets: a checklist of some basics

(a) So what elementary ideas should you be familiar with, given our limited
current purposes?
Let’s have a quick checklist. This is really for philosophers who haven’t been
well brought up. There shouldn’t be anything here which is not very familiar to
mathematicians, but for the record . . .

(i) A set is a collection of objects, treated as itself a single object. A and B

count as one and the same set if and only if whatever is a member of A is
a member of B and vice versa (that’s the extensionality principle).
(ii) Notation. We use the likes of ‘{a, b, c, d}’ to denote the set whose members
are a, b, c, d. And we use the likes of ‘{x | x is F }’ to denote the set of
things (in some domain) which are F .
Membership is symbolized by ‘∈’; the subset relation is symbolized by
‘⊆’, so A ⊆ B is true just when for all x, if x ∈ A, then x ∈ B.
The membership and subset relations need to be very sharply distin-
guished from each other (the beginning of set-theoretic wisdom!). And
note in particular that the singleton set {a} is to be distinguished from
its sole member a: thus a ∈ {a} and {a} ⊆ {a}, but not a = {a} and not
a ⊆ {a}.

8
 DRAFT– 31 DEC 2021
Sets: a checklist of some basics

(iii) If A, B are sets, then so too are their union, intersection and their power-
sets.
If the intersection A ∩ B is always to exist, then we have to allow a
set which contains no members (since A and B might not overlap). By
extensionality, the empty set ∅ is unique.
The powerset of A, P(A), is the set whose members are all and only the
subsets of A. Note this assumes that sets are indeed things which can be
members of other sets.
(iv) Sets are in themselves unordered. But we often need to work with ordered
pairs, ordered triples, ordered quadruples, . . . , tuples more generally. We
use ‘ha, bi’ – or often simply ‘(a, b)’ – for the ordered pair, first a, then b.
So, while {a, b} = {b, a}, by contrast ha, bi = 6 hb, ai.
We can implement ordered pairs using unordered sets in various ways:
all we need is some definition which ensures that ha, bi = ha0 , b0 i if and only
if a = a0 and b = b0 . The following is standard: ha, bi =def {{a}, {a, b}}.
Once we have ordered pairs available, we can use them to define ordered
triples: ha, b, ci can be defined as first the pair ha, bi, then c, i.e. as hha, bi, ci.
Then the quadruple ha, b, c, di can be defined as hha, b, ci, di. And so it goes.
(v) The Cartesian product A×B of the sets A and B is the set whose members
are all the ordered pairs whose first member is in A and whose second
member is in B. So A × B is {hx, yi | x ∈ A & y ∈ B}. Cartesian products
of n sets are defined as sets of n-tuples, again in the obvious way.
(vi) If R is a binary relation between members of the set A and members of the
set B, then its extension is the set of ordered pairs hx, yi (with x ∈ A and
y ∈ B) such that x is R to y. So the extension of R is a subset of A × B.
Similarly, the extension of an n-place relation is the set of n-tuples of
things which stand in that relation. In the unary case, where P is a property
defined over some set A, then we can simply say that the extension of P
is the set of members of A which are P .
For many mathematical purposes, we treat properties and relations ex-
tensionally; i.e. we regard properties with the same extension as being the
same property, and likewise for relations. Indeed, we can often simply treat
a property (relation) as if it simply is its extension.
(vii) The extension (or graph) of a unary function f which sends members of A
to members of B is the set of ordered pairs hx, yi (with x ∈ A and x ∈ B)
such that f (x) = y. Similarly for n-place functions. For many purposes, we
treat functions extensionally, regarding functions with the same extension
as the same. Again we often treat a function as if it is its extension, i.e.
we identify a function with its graph.
(viii) Relations can, for example, be reflexive, symmetric, transitive; equivalence
relations are all three. Note that if ≡ is an equivalence relation defined
over some set, it partitions that set into equivalence classes (we never
say ‘equivalence sets’ !) of objects standing in that relation. If [x] is the

9
 DRAFT– 31 DEC 2021
2 A very little informal set theory

equivalence class (with respect to ≡) containing x, then [x] = [y] if and

only if x ≡ y.
(ix) Two sets are equinumerous just if we can match up their members one-
to-one, i.e. when there is a one-to-one function, a bijection, between the
sets. A set is countably infinite if and only if it is equinumerous with the
natural numbers.
And here we get to the first exciting claim, a version of Cantor’s Theorem
– there are infinite sets which are not countably infinite. A simple example
is the set of infinite binary strings. Why so? If we take any countably
infinite list of such strings, we can always define another infinite binary
string which differs from the first string on our list in the first place, differs
from the second in the second place, the third in the third place, etc., so
cannot appear anywhere in our given list.
This is just the beginning of a story about how sets can have different
infinite ‘sizes’ or cardinalities. But at this stage you need to know little
more than that bald fact: further elaboration can wait.
(x) There’s one further, rather less elementary, idea that you should also meet
sooner rather than later, so that you recognize any passing references to
it. This is the Axiom of Choice. In one version, this says that, given an
infinite family of sets, there is a choice function – i.e. a function which
‘chooses’ a single member from each set in the family. Bertrand Russell’s
toy example: given an infinite collection of pairs of socks, there is a function
which chooses one sock from each pair.
Note that while other principles for forming new sets (e.g. unions, power-
sets) determine what the members of the new set are, Choice just tells us
that there is a set (the extension of the choice function) which plays a
certain role, without specifying its members.
At this stage you need to know that Choice is a principle which is im-
plicitly or explicitly invoked in many mathematical proofs. But you should
also know that it is independent of other basic set-theoretic principles (and
there are set theories in which it doesn’t hold) – which is why we often
explicitly note when, in more advanced logical theory, a result does indeed
depend on Choice.

(b) An important observation before proceeding.

The set of musketeers {Athos, Porthos, Aramis} is not another musketeer and
so isn’t a member of itself. Likewise, the set of prime numbers isn’t itself a prime
number, so again isn’t a member of itself. We’ll say that a set which is similarly
not a member of itself is normal. Now we ask: is there a set R whose members
are all and only the normal sets?
No. For if there were, it would be a member of itself if and only if wasn’t –
think about it! – which is impossible. The putative set R is, in some sense, ‘too
big’ to exist. Hence, if we overshoot and naively suppose that for any property –
including the property of being a normal set – there is a set which is its extension,
we get into deep trouble (this is the upshot of ‘Russell’s paradox’).
10
 DRAFT– 31 DEC 2021
Recommendations on informal basic set theory

Now, some people use ‘naive set theory’ to mean, quite specifically, a the-
ory which makes that simple but hopeless assumption that any property at all
has a set as its extension. As we’ve just seen, naive set theory in this sense is
inconsistent.
But here we need to avoid getting entangled in one of those rather annoying
terminological divergences. Because, for many others, ‘naive set theory’ just
means set theory developed informally, without rigorous axiomatization, but
guided by unambitious low-level principles. In this different second sense, we
have been proceeding naively in this chapter – yet, fingers crossed, we remain on
track for developing a consistent story! Thus, we were careful in (vi) to assign
extensions just to those properties and relations that are defined over domains
we are already given as sets.
True, our story so far is silent about exactly which putative sets are the kosher
ones – i.e. are not ‘too big’ to be to be problematic. However, important though
it is, we can leave this topic until Chapter 7 when we turn to set theory proper.
Low-level practical uses of sets in ‘ordinary’ mathematics seem remote from
such problematic cases; hopefully, we can continue to proceed naively for now in
elementary contexts.

2.2 Recommendations on informal basic set theory

If you are a mathematics student, then the ideas on our checklist will surely
already be very familiar, e.g. from those introductory chapters or appendices
you so often find in mathematics texts! A particularly good example is
1. James R. Munkres, Topology (Prentice Hall, 2nd edition, 2000). Chapter
1, ‘Set Theory and Logic’. This tells you very clearly about basic set-
theoretic concepts, up to countable vs. uncountable sets and the axiom
of choice (plus a few other things worth knowing about).
But non-mathematicians – or indeed mathematicians who are a bit rusty – might
find one of the following more to their taste:

2. Tim Button, Set Theory: An Open Introduction (Open Logic Project),

Chapters 1–5. Available at tinyurl.com/opensettheory.
Read Chapter 1 for some interesting background. Chapter 2 intro-
duces basic notions like subsets, powersets, unions, intersections, pairs,
tuples, Cartesian products. Chapter 3 is on relations (treated as sets).
Chapter 4 is on functions. Chapter 5 is on the size of sets, countable vs
uncountable sets, Cantor’s Theorem. At this stage in his book, Button
is proceeding naively in our second sense, with the promise that every-
thing he does can be replicated in the rigorously axiomatized theory he
introduces later.
Button writes, here as elsewhere, with very admirable clarity. So this
is warmly recommended.

11
 DRAFT– 31 DEC 2021
2 A very little informal set theory

3. David Makinson, Sets, Logic and Maths for Computing (Springer, 3rd
edn 2020), Chapters 1 to 3.
This is exceptionally clear and very carefully written for students
without much mathematical background. Chapter 1 reviews basic facts
about sets. Chapter 2 is on relations. Chapter 3 is on functions. This too
can be warmly recommended (though you might want to supplement it
by following up his reference to Cantor’s Theorem).
Now, Makinson doesn’t mention the Axiom of Choice at all. While Button
does eventually get round to Choice in his Chapter 16; but the treatment
there depends on the set theory developed in the intervening chapters, so
isn’t appropriate for us just now. Instead, the following two pages should
be enough for the present:

4. Timothy Gowers et al. eds, The Princeton Companion to Mathematics

(Princeton UP, 2008), §III.1: The Axiom of Choice.

2.3 Virtual classes, real sets

An afterword. According to Cantor, a set is a unity, a single thing in itself over
and above its members. But if that is the guiding idea, then it is worth noting that
a great deal of elementary informal set talk in mathematics is really no more than
a façon de parler. Yes, it is a useful and familiar idiom for talking about many
things at once; but in many elementary contexts apparent talk of a set doesn’t
really carry any serious commitment to there being any additional object, a set,
over and above those many things. On the contrary, in such contexts, apparent
talk about a set of F s can very often be paraphrased away into direct talk about
those F s, without any loss of content.
Here is just one example, relevant for us. It is usual to say something like
this: (1) “A set of formulas Γ logically entails the formula A if and only if any
valuation which makes every member of Γ true makes A true too”. Don’t worry
for now about the talk of valuations: just note that the reference to a set of
formulas and it members is doing no work here. It would do just as well to say
(2) “Some formulas G logically entail A if and only if every valuation which
makes those formulas G all true makes A true too”. The set version (1) adds
nothing important to the plural version (2).
When set talk can be paraphrased away like this, we are only dealing with –
as they say – mere virtual classes.
One source for this description is W.V.O. Quine’s famous discussion in the
opening chapter of his Set Theory and its Logic (1963):

Much . . . of what is commonly said of classes with the help of ‘∈’

can be accounted for as a mere manner of speaking, involving no real
reference to classes nor any irreducible use of ‘∈’. . . . [T]his part of
class theory . . . I call the virtual theory of classes.

12
 DRAFT– 31 DEC 2021
Virtual classes, real sets

You will eventually find that this same usage plays an important role in set theory
in some treatments of so-called ‘proper classes’ as distinguished from sets. For
example, in his standard book Set Theory (1980), Kenneth Kunen writes

Formally, proper classes do not exist, and expressions involving them

must be thought of as abbreviations for expressions not involving
them.

The distinction being made here is an old one. Here is Paul Finsler, writing in
1926 (as quoted by Luca Incurvati, in his Conceptions of Set):

It would surely be inconvenient if one always had to speak of many

things in the plural; it is much more convenient to use the singular
and speak of them as a class. . . . A class of things is understood
as being the things themselves, while the set which contains them
as its elements is a single thing, in general distinct from the things
comprising it. . . . Thus a set is a genuine, individual entity. By con-
trast, a class is singular only by virtue of linguistic usage; in actuality,
it almost always signifies a plurality.

Finsler writes ‘almost always’, I take it, because a class term may in fact denote
just one thing, or even – perhaps by misadventure – none.
Nothing hangs on the particular terminology, ‘classes’ vs ‘sets’. What matters
(or will eventually matter) is the distinction between non-committal, eliminable,
talk – talk of merely virtual sets/classes/pluralities (whichever idiom we use) –
and uneliminable talk of sets as entities in their own right.

13
 DRAFT– 31 DEC 2021

3 First-order logic

Now let’s get down to business!

This chapter begins with an overview of classical first-order logic, FOL, mean-
ing standard propositional and predicate logic – which is the starting point for
any mathematical logic course. (Why ‘classical’ ? Why ‘first-order’ ? – all will
eventually be explained!)
At this level, the most obvious difference between various treatments of FOL
is in the choice of proof-system: so I will next comment on two main options.
Then I highlight the main self-study recommendations. These are followed by
some suggestions for parallel and further reading. And after a short historical
section, this chapter ends with a postscript commenting on some other books,
mostly responding to frequently asked questions.1

3.1 FOL: a general overview

FOL deals with deductive reasoning that turns on the use of ‘propositional con-
nectives’ like and, or, if, not, and on the use of ‘quantifiers’ like every, some, no.
But in ordinary language (including the ordinary language of informal mathe-
matics) these logical operators work in surprisingly complex ways, introducing
the kind of obscurities and possible ambiguities we certainly want to avoid in
logically transparent arguments. What to do?
From the time of Aristotle, logicians have used a ‘divide and conquer’ strategy
that involves introducing simplified, tightly-disciplined, languages. For Aristotle,
his regimented language was a fragment of very stilted Greek; for us, our reg-
imented languages are artificial formal constructions. But either way, the plan
is that we tackle a stretch of reasoning by reformulating it in a suitable regi-
mented language with much tidier logical operators, and then we can evaluate
the reasoning once recast into this more well-behaved form. This way, we have

1A note to philosophers. If you have carefully read a substantial introductory logic text for
philosophers such as Nicholas Smith’s, or even my own, you will already be familiar with
(versions of) a fair amount of the material covered in this chapter. However, you will now
begin to see topics being re-presented in the sort of mathematical style and with the sort of
rigorous detail that you will necessarily encounter more and more as you progress in logic.
You do need to start feeling entirely comfortable with this mode of presentation at an early
stage. So it is well worth working through even rather familiar topics again, this time with
more mathematical precision.

14
 DRAFT– 31 DEC 2021
FOL: a general overview

a division of labour. First, we clarify the intended structure of the original ar-
gument by rendering it into an unambiguous simplified/formalized language.
Second, there’s the separate business of assessing the validity of the resulting
regimented argument.
In exploring FOL, then, we will use appropriate formal languages which con-
tain, in particular, tidily-disciplined surrogates for the propositional connectives
and, or, if, not (standardly symbolized ∧, ∨, → and ¬), plus replacements for
the ordinary language quantifiers (roughly, using ∀x for every x is such that . . . ,
and ∃y for some y is such that. . . ).
Although the fun really starts once we have the quantifiers in play, it is very
helpful to develop FOL in two main stages:

(a) Typically, we start by introducing propositional languages whose built-in

logical apparatus comprises just the propositional connectives, and then
discuss the propositional logic of arguments framed in these languages.
This gives us a very manageable setting in which to first encounter a whole
range of logical concepts and strategies.

(b) We then move on to develop the syntax and semantics of richer formal
languages which add the apparatus of so-called first-order quantification,
and explore the logic of arguments rendered into such languages.

So let’s have just a little more detail about stages (a) and (b).
(a.i) We first look at the syntax of propositional languages, defining what count
as the well-formed formulas (wffs) of such languages.
If you have already encountered languages of this kind, you will now get to
know how to actually prove various things about them that seem obvious and
that you perhaps previously took for granted – for example, that ‘bracketing
works’ to block ambiguities like P ∨ Q ∧ R, so every well-formed formula has a
unique unambiguous parsing.
(a.ii) On the semantic side, we need the idea of a valuation for a propositional
language.
We start with an assignment of truth-values, true vs false, to the atomic
formulas, the basic building blocks of our languages. This assignment of values to
atoms then fixes the truth-values of complex sentences involving the connectives.
Here we rely crucially on the ‘truth-functional’ interpretation of the connectives
– so the truth value of a formula like ¬(P ∧ (Q ∨ ¬R)) is entirely fixed as a
function of the truth-values of the atomic formulas P, Q, R.
More generally, then, any wff – whether atomic or complex – is determined to
be either definitely true or definitely false (one or the other, but not both) on any
particular valuation. This core assumption is distinctive of classical two-valued
semantics.
(a.iii) Even at this early point, questions arise. For example, how satisfactory
is the representation of an informal conditional if P then Q by a formula P → Q

15
 DRAFT– 31 DEC 2021
3 First-order logic

which uses a truth-functional arrow connective? And why restrict ourselves to

just a small handful of truth-functional connectives?
You don’t want to get too entangled with the first question! – but you do need
to understand why we represent the conditional in FOL in the way we do. As for
the second question, it’s an early theorem that every truth-function can in fact
be expressed using just a handful of connectives. There are also some related
‘normal form’ results.
(a.iv) Now a crucial pair of definitions:

A wff A from a propositional language is a tautology it is true on any

assignment of values to the relevant atoms.
A set of wffs Γ tautologically entails A when any assignment of values
to the relevant atoms which makes all the sentences in Γ true makes
A true too.

So the notion of tautological entailment aims to regiment the idea of an ar-

gument’s being logical valid in virtue of the distribution of connectives in its
premisses and conclusion.
You will need to explore some of the key properties of this semantic entailment
relation. And note that in this rather special case, we can mechanically determine
whether Γ entails A, e.g. by a ‘truth table test’ (at least so long as there are only
finitely many wffs in Γ, and hence only finitely many relevant atoms to worry
about).
(a.v) Different textbook presentations of stages (a.i) to (a.iv) can go into dif-
ferent levels of detail, but the basic story remains much the same. However, now
the path forks. For the usual next topic will be a formal deductive system in
which we can construct step-by-step derivations of conclusions from premisses
in propositional logic. There is a variety of such systems to choose from, and I’ll
mention five main types in §3.2.
Different proof systems for classical propositional logic will (as you’d expect)
be equivalent – meaning that, given some premisses, we can derive the same
conclusions in each system. However, the systems do differ considerably in their
intuitive appeal and user-friendliness, as well as in some of their more technical
features. Note, though: apart from looking at a few illustrative examples, we
won’t be much interested in producing lots of derivations inside a chosen proof
system; the focus will be more on establishing results about the systems.
In due course, the educated logician will want to learn at least a little about the
various types of proof system – at the minimum, you should eventually get a sense
of how they respectively work, and come to appreciate the interrelations between
them. But in this overview – as is usual when starting out on mathematical logic
– we look at axiomatic and natural deduction systems in particular (I say more
about these in the next section).
(a.vi) At this point, then, we will have two quite different ways of defining what
makes for a deductively good argument in propositional classical logic:

16
 DRAFT– 31 DEC 2021
FOL: a general overview

We said that a set of premisses Γ tautologically entails the conclusion

A if every possible valuation which makes Γ all true makes A true.
(That’s a semantically defined idea.)
We can now also say that Γ yields the conclusion A in your chosen
proof-system S if there is an S-type derivation of the conclusion A
from premisses in Γ. (This is a matter of there being a proof-array
with the right syntactic shape.)

Of course, we want these two approaches to fit together. We want our favoured
proof-system S to be sound – it shouldn’t give false positives. In other words,
if there is an S-derivation of A from Γ, then A really is tautologically entailed
by Γ. We also would like our favoured proof-system S to be complete – we want
it to capture all the correct semantic entailment claims. In other words, if A
is tautologically entailed by the set of premisses Γ, then there is indeed some
S-derivation of A from premisses in Γ.
So, in short, we will want to establish both the soundness and the complete-
ness of our favoured proof-system S for propositional logic (axiomatic, natural
deduction, whatever). Now, these two results will hold no terrors! However, in
establishing soundness and completeness for propositional logics you will en-
counter some useful strategies which can later be beefed-up to give us soundness
and completeness results for stronger logics.
(b.i) Having warmed up with propositional logic, we then turn to full FOL so
we can also deal with arguments whose validity depends on their quantificational
structure (starting with the likes of our old friend ‘Socrates is a man; all men
are mortal; hence Socrates is a mortal’ !).
We need to introduce appropriate formal languages with quantifiers (more
precisely, with first-order quantifiers, running over a fixed domain of objects:
the next chapter explains the contrast with second-order quantifiers). So syntax
first.
Consider the simple ordinary-language sentence (i) ‘Socrates is wise’. And
now note that we can replace the name in (i) with the quantifier expression
‘everyone’ to give us another sentence (ii) ‘Everyone is wise’. Similarly, we can
directly replace the name ‘Juliet’ in (iii) ‘Romeo loves Juliet’ with the quantifier
expression ‘someone’ to get the equally grammatical (iv) ‘Romeo loves someone’.
In FOL, however, while we might render (i) as simply Ws, (ii) will get rendered
by something like ∀xWx (roughly, everyone x is such that x is wise). Similarly
if (iii) is rendered Lrj, then (iv) gets rendered by something like ∃xLrx (roughly,
someone x is such that Romeo loves x). But why?
It is crucial to understand the rationale for this departure from the syntac-
tic patterns of ordinary language and the use of the apparently more complex
‘quantifier/variable’ syntax in expressing generalizations. The headline point is
that in our formal languages we need to avoid the kind of structural ambiguities
that we can get in ordinary language when there is more than one logical oper-
ator involved. Consider for example the ambiguous ‘Everyone has not arrived’.
Does that mean ‘Everyone is such that they have not arrived’ or ‘It is not the
17
 DRAFT– 31 DEC 2021
3 First-order logic

case that everyone has arrived’ ? Our logical notation will distinguish ∀x¬Ax and
¬∀xAx, with the relative ‘scopes’ of the generalization and the negation now
made transparent by the structure of the formulas.
(b.ii) Turning to semantics: the first key idea we need is that of a model struc-
ture, a (non-empty) domain of objects equipped with some properties, relations
and/or functions. And here we treat properties etc. extensionally. In other words,
we can think of a property as a set of objects from the domain, a binary relation
as a set of pairs from the domain, and so on. (Compare our remarks on naive
set theory in §2.1; though, heeding the point of §2.3, we can arguably take the
talk of sets here in a non-committal way.)
Then, crucially, you need to grasp the idea of an interpretation of an FOL
language in such a structure. Names are interpreted as denoting objects in the
domain; a one-place predicate gets assigned a property, i.e. a set of objects from
the domain (its extension – intuitively, the objects it is true of); a two-place
predicate gets assigned a binary relation; and so on.
Such an interpretation of the elements of a first-order language then generates
a valuation (a unique assignment of truth-values) for every sentence of the inter-
preted language. How does it do that? Well, a simple predicate-name sentence
like Ws will be true just if the object denoted by s is in the extension of W; a
sentence like Lrj is true if the ordered pair of the objects denoted by r and j is
in the extension of L; and so on. That’s easy. And the propositional connectives
continue to behave basically as in propositional logic.
But extending the formal semantic story to explain how the interpretation of
a language fixes the valuations of more complex, quantified, sentences requires
a new idea, some variant of the thought that ∀xWx is true just when Wa is true,
whatever ‘a’ picks out when treated as temporary name (compare: ‘everything
is W ’ is true just when ‘that is W ’ is true whatever the demonstrative ‘that’
might pick out in the relevant domain). There are a number of slightly different
ways of developing this story more carefully (for a start, do we take our FOL
languages to have a supply of special symbols available to act as temporary
names? or do we re-use a variable like ‘x’ without a preceding quantifier to then
act as a temporary name?) You need to get your head round the details of one
fully spelt-out story.
(b.iii) We can now introduce the idea of a model for a set of sentences, i.e. an
interpretation in a structure which makes all the sentences true together. And
we can then again define a semantic relation of entailment, this time for FOL
sentences:

A set of FOL sentences Γ semantically entails A when any interpre-

tation in any structure which makes all the sentences in Γ true also
makes the sentence A true – for short, when any model for Γ is also
a model for A.

You’ll again need to know some of the basic properties of this entailment relation.

18
 DRAFT– 31 DEC 2021
FOL: a general overview

For one important example, note that if Γ has no model, then – on our defi-
nition – Γ semantically entails A for any A at all, including any contradiction.
(b.iv) Unlike the case of tautological entailment, this time there is no general
procedure for mechanically testing whether Γ semantically entails A. So the use
of proof systems to warrant entailments now really comes into its own.
You can again encounter five main types of proof system for FOL, with their
varying attractions and drawbacks. And to repeat, you’ll want at some future
point to find out at least something about all these styles of proof. But, as
before, we will principally be looking here at axiomatic systems and at one kind
of natural deduction.
As you will see, whichever form of proof system you take, some care is require
in handling inferences using the quantifiers in order to avoid fallacies. And we
will need extra care if we don’t use special symbols as temporary names but
allow the same variables to occur both ‘bound’ by quantifiers and ‘free’. This
isn’t the place to go into details; but you do need to tread carefully hereabouts!
(b.v) As with propositional logic, we will want to show that our chosen proof
system for FOL is sound and doesn’t overshoot (so giving us false positives) and is
complete and doesn’t undershoot (leaving us unable to derive some semantically
valid entailments).
In other words, if S is our FOL proof system, Γ a set of sentences, and A a
particular sentence, we need to show:

If there is an S-proof of A from premisses in Γ, then Γ does indeed

semantically entail A. (Soundness)
If Γ semantically entails A, then there is an S-proof of A from pre-
misses in Γ. (Completeness)

Now, for future uses, it is important that the completeness theorem actually
comes in two versions. There is a weaker version where Γ is restricted to having
only finitely many members (or indeed is empty). And there is a crucial stronger
version which allows Γ to be infinite.
And it is at this point, proving strong completeness, that the study of FOL
becomes mathematically really interesting.
(b.vi) Later chapters will continue the story along various paths; here though
I should quickly mention just one immediate corollary of completeness.
Proofs in formal systems are always only finitely long; so a proof of A from Γ
can only call on a finite number of premisses in Γ. But the strong completeness
theorem for FOL allows Γ to have an infinite number of members. This com-
bination of facts immediately implies the compactness theorem for sentences of
FOL languages:

(c) If every finite subset of Γ has a model, so does Γ.2

2 That’s
equivalent to the claim that if (i) Γ doesn’t have a model, then there is a finite subset
∆ ⊆ Γ such that (ii) ∆ has no model. Suppose (i). This implies that Γ semantically entails

19
 DRAFT– 31 DEC 2021
3 First-order logic

This compactness theorem, you will discover, has numerous applications in model
theory.

3.2 A little more about types of proof-system

I’ve often been struck, answering queries on an internet forum, by how many
students ask variants of “how do you prove X in first-order logic?”, as if they
have never encountered the idea that there is no single deductive system for FOL!
So I do think it is worth emphasizing here at the outset that there are indeed
various styles of proof-system – and moreover, for each general style, there are
many different particular versions.
This isn’t the place to get into too many details with lots of examples. Still,
some quick headlines could be very helpful for orientation.
(a) Let’s have a mini-example to play with. Consider the argument ‘If Jack
missed his train, he’ll be late; if he’s late, we’ll need to reschedule; so if Jack
missed his train, we’ll need to reschedule’. Inuitively valid, of course. After all,
just suppose for a moment that Jack did miss the train: then he’ll be late; and
hence we’ll need to reschedule. Which shows that if he missed the train, we’ll
need to reschedule.
Using the obvious translation manual to recast the argument in a formal
propositional language, we’ll therefore want to be able to show that – in our
favoured proof system – we can correspondingly argue from the premisses (P → Q)
and (Q → R) to the conclusion (P → R).
(b) You will be familiar with the general idea of an axiomatized theory. We
are given some axioms and some deductive apparatus is presupposed. Then the
theorems of the theory are whatever can be derived from the axioms. Similarly:

In an axiomatic logical system, we adopt some basic logical truths as

axioms. And then we explicitly specify the allowed rules of inference:
usually these are just very simple ones such as the ‘modus ponens’
rule for the conditional: ‘from A and A → B, we can infer B’.
A proof from some given premisses to a conclusion then has the
simplest possible structure. It is just a sequence of wffs – each of
which is either (i) one of the premisses, or (ii) one of the logical ax-
ioms, or (iii) follows from earlier wff in the proof by one of the rules
of inference – with the whole sequence ending with the target con-
clusion.
A logical theorem of the system is then a wff that can be proved
from the logical axioms alone (without appeal to any further pre-
misses).

a contradiction. So by completeness we can derive a contradiction from Γ in your favourite

proof system. That proof will only use a finite collection of premisses ∆ ⊆ Γ. But if ∆
proves a contradiction, then by soundness, ∆ semantically entails a contradiction, which
can only be the case if (ii).

20
 DRAFT– 31 DEC 2021
A little more about types of proof-system

Now, a standard axiomatic system for FOL (such as in Mendelson’s classic

book) will include as axioms all wffs of the following shapes:

Ax1. (A → (B → A))
Ax2. ((A → (B → C)) → ((A → B) → (A → C)))

More carefully, all instances of those two schemas – where we systematically

replace letters like A, B, etc. with wffs (simple or complex) – will count as axioms.
And among the rules of inference for our system will be modus ponens (MP),
i.e. the rule that from A and (A → B) you can infer B. With this apparatus in
place, we can then construct the following formal derivation, arguing as wanted
from (P → Q) and (Q → R) to (P → R).

1. (P → Q) premiss
2. (Q → R) premiss
3. ((Q → R) → (P → (Q → R))) instance of Ax1
4. (P → (Q → R)) from 2, 3 by MP
5. ((P → (Q → R) → ((P → Q) → (P → R))) instance of Ax2
6. ((P → Q) → (P → R)) from 4, 5 by MP
7. (P → R) from 1, 6 by MP

Which wasn’t too difficult!

(c) Informal deductive reasoning, however, is not relentlessly linear like this.
We do not require that each proposition in a proof (other than a given premiss
or a logical axiom) has to follow from what’s gone before. Rather, we often step
sideways (so to speak) to make some new temporary assumption, ‘for the sake
of argument’.
For example, we may say ‘Now suppose that A is true’; we go on to show that,
given what we’ve already established, this extra supposition leads to a contra-
diction; we then drop or ‘discharge’ the temporary supposition and conclude
that not-A. That’s how one sort of reductio ad absurdum argument works. For
another example, we may again say ‘Suppose that A is true’; this time we go
on to show that we can now derive C; we then again discharge the temporary
supposition and conclude that if A, then C. That’s how we often argue for a
conditional proposition: in fact, this is exactly what we did in the informal rea-
soning we gave to warrant the argument about Jack at the beginning of this
section.
That motivates our using a more flexible kind of proof-system:

A natural-deduction system of logic aims to formalize patterns of

reasoning now including those where we can argue by making and
then later discharging temporary assumptions. Hence, for example,
as well as the simple modus ponens (MP) rule for the conditional
‘→’, there will be a conditional proof (CP) rule along the lines of ‘if
we can infer B from the assumption A, we can drop the assumption
A and conclude A → B’.

21
 DRAFT– 31 DEC 2021
3 First-order logic

Now, in a natural-deduction system, we will evidently need some way of keep-

ing track of which temporary assumptions are in play and for how long. Two
particular ways of doing this are commonly used:

(i) A multi-column layout was popularized by Frederick Fitch in his classic

1952 logic text, Symbolic Logic: an Introduction. Here’s a proof in this
style, from the same premisses to the same conclusion as before:

1. (P → Q) premiss
2. (Q → R) premiss
3. P supposition for the sake of argument
4. Q by MP from 3, 1
5. R by MP from 4, 2
6. (P → R) by CP, given the ‘subproof’ 3–5

So the key idea is that the line of proof snakes from column to column,
moving a column to the right (as at line 3) when a new temporary assump-
tion is made, and moving back a column to the left (as at line 6) when the
assumption is dropped or ‘discharged’. This mode of presentation really
comes into its own when multiple temporary assumptions are in play, and
makes such proofs very easy to read and follow. And, compared with the
axiomatic derivation, this regimented line of argument does indeed seem
to warrant being called a ‘natural deduction’ !

(ii) However, the layout for natural deduction proofs favoured for serious work
was first introduced Gerhard Gentzen in his doctoral thesis of 1933. He
sets out the proofs as trees, with premisses or temporary assumptions at
the top of branches and the conclusion at the root of the tree – and he uses
a system for explicitly tagging temporary assumptions and the inference
moves where they get discharged.
Let’s again argue from the same premisses to the same conclusion as
before. We will build up our Gentzen-style proof in two stages. First, then,
take the premisses (P → Q) and (Q → R) and the additional supposition
P, and construct the following proof of R using modus ponens twice:

P (P → Q)
Q (Q → R)
R

The horizontal lines, of course, signal inference moves.

OK: so we’ve shown that, assuming P, we can derive R, by using the
other assumptions. Hence, moving to the second phase of the argument, we
will next discharge the assumption P while keeping the other assumptions
in play, and apply conditional proof (CP), in order to infer (P → R). We’ll
signal that the assumption P is no longer in play by now enclosing it in
square brackets. So applying (CP) turns the previous proof into this:
22
 DRAFT– 31 DEC 2021
A little more about types of proof-system

[P](1) (P → Q)
Q (Q → R)
R (1)
(P → R)

For clarity, we tag both the assumption which is discharged and the cor-
responding inference line where the discharging takes place with matching
labels, in this case ‘(1)’. (We’ll need multiple labels when multiple tempo-
rary assumptions are put into play and then dropped.)
In the second proof, then, just the unbracketed sentences at the tips of
branches are left as ‘live’ assumptions. So this is our Gentzen-style proof
from those remaining premisses (P → Q) and (Q → R) to the conclusion
(P → R).
(d) There is much more to be said of course, but that’s enough by way of some
very introductory remarks about the first half of the following list of commonly
used types of proof system:
1. Old-school axiomatic systems.
2. (i) Natural deduction done Gentzen-style.
(ii) Natural deduction done Fitch-style.
3. ‘Semantic tableaux’ or ‘truth trees’.
4. Sequent calculi.
5. Resolution calculi.
So next, a very brief word about semantic tableaux, which are akin to Gentzen-
style proof trees turned upside down.
The key idea is this. Instead of starting from some premisses Γ and aiming
directly for the desired conclusion A, we begin instead by assuming the premisses
are all true while the conclusion is false. And then we ‘work backwards’ from the
assumed values of these typically complex wffs, aiming to uncover a valuation of
the atoms for the relevant language which indeed makes Γ all true and A false.
If our search gets entangled in contradiction, that tells us that there is no such
valuation: so if Γ are all true, then A indeed has to be true too.
Note however that assuming e.g. that (A ∨ B) is true doesn’t tell us which
of A and B is true too: so as we ‘work backwards’ from the values of more
complex wffs to the values of their components we will typically have to explore
branching options, which are most naturally displayed on a downward-branching
tree. Hence ‘truth trees’.
The details of a truth-tree system for FOL are elegantly simple – which is why
the majority of elementary logic books for philosophers introduce either (2.ii)
Fitch-style natural deduction or (3) truth trees, or both. And indeed, it is well
worth getting to know about tree systems at a fairly early stage because they
can be adapted rather nicely to dealing with logics other than FOL. However,
introductory mathematical logic text books do usually focus on either (1) axiom-
atic systems or (2.i) Gentzen-style proof systems, and those will be our initial
main focus here too.
23
 DRAFT– 31 DEC 2021
3 First-order logic

As for (4) the sequent calculus, in its most interesting form this really comes
into its own in more advanced work in proof theory. While (5) resolution calculi
are perhaps of particular concern to computer scientists interested in automating
theorem proving.
(e) I should add, though, that even once you’ve picked your favoured general
type of proof-system to work with from (1) to (5), there are many more choices
to be made before landing on a specific system of that type. For example, F.
J. Pelletier and Allen Hazen published a survey of logic texts aimed at philoso-
phers which use natural deduction systems (tinyurl.com/pellhazen): they note
that no less than thirty texts use a variety of Fitch-style system (2.ii) – and,
rather remarkably, no two of these have exactly the same system of rules for
FOL!
Moral? Don’t get too hung up on the finest details of a particular textbook’s
proof-system; it is the overall guiding ideas that matter, together with the Big
Ideas underlying proofs about the chosen proof-system (such as the soundness
and completeness theorems).

3.3 Basic recommendations for reading on FOL

A preliminary reference. In my elementary logic book I do carefully explain
the ‘design brief’ for the languages of FOL, spelling out the rationale for the
quantifier-variable notation. For some, this might be helpful parallel reading
when working through your chosen main text(s), at the point when that notation
is introduced:

1. Peter Smith, Introduction to Formal Logic* (2nd edn), Chapters 26–28.

Downloadable from logicmatters.net/ifl.

Unsurprisingly, there is a very long list of texts which cover FOL. But the
whole point of this Guide is to choose. So here are my top recommendations,
starting with one-and-a-third books which, taken together, make an excellent
introduction:

2. Ian Chiswell and Wilfrid Hodges, Mathematical Logic (OUP, 2007).

This is very approachable. It is written by mathematicians primar-
ily for mathematicians, yet it is only one notch up in actual difficulty
from some introductory texts for philosophers like mine or Nick Smith’s.
However – as its title might suggest – it does have a notably more math-
ematical ‘look and feel’. Philosophers can skip over a few of the more
mathematical illustrations; while depending on background, mathemati-
cians should be able to take this book at pace.
The briefest headline news is that authors explore a Gentzen-style
natural deduction system. But by building things up in three stages
– so after propositional logic, they consider an important fragment of

24
 DRAFT– 31 DEC 2021
Basic recommendations for reading on FOL

first-order logic before turning to the full-strength version – they make

e.g. proofs of the completeness theorem for first-order logic unusually
comprehensible. For a more detailed description see my book note on
C&H, tinyurl.com/CHbooknote.
Very warmly recommended, then. For the moment, you only need read
up to and including §7.6 (under two hundred pages). But having got that
far, you might as well read the final few sections and the Postlude too!
The book has brisk solutions to some of the exercises.

Next, you should complement C&H by reading the first third of the following
excellent book:

3. Christopher Leary and Lars Kristiansen’s A Friendly Introduction to

Mathematical Logic* (1st edn by Leary alone, Prentice Hall, 2000; 2nd
edn Milne Library, 2015). Downloadable at tinyurl.com/friendlylogic.
There is a great deal to like about this book. Chs. 1–3, in either edi-
tion, do indeed make a friendly and helpful introduction to FOL. The
authors use an axiomatic system, though this is done in a particularly
smooth way. At this stage you could stop reading after the beginning of
§3.3 on compactness, which means you will be reading just 87 pages.
Unusually, L&K dive straight into a treatment of first-order logic with-
out spending an introductory chapter or two on propositional logic: in
a sense, as you will see, they let propositional logic look after itself (by
just helping themselves to all instances of tautologies as axioms). But
this rather happily means (in the present context) that you won’t feel
that you are labouring through the very beginnings of logic one more
time than is really necessary – this book therefore dovetails very nicely
with C&H.
Again written by mathematicians, some illustrations of ideas can
presuppose a smattering of background mathematical knowledge; but
philosophers will miss very little if they occasionally have to skip an
example (and the curious can always resort to Wikipedia, which is quite
reliable in this area, for explanations of some mathematical terms). The
book ends with extensive answers to exercises.
I like the overall tone of L&K very much indeed, and say more about
this admirable book in another book note, tinyurl.com/LKbooknote.

As an alternative to the C&H/L&K pairing, the following slightly more conven-

tional book is also exceptionally approachable:

4. Derek Goldrei, Propositional and Predicate Calculus: A Model of Argu-

ment (Springer, 2005).
This book explicitly designed for self-study. Read up to the end of
§6.1 (though you could skip §§4.4 and 4.5 for now, leaving them until

25
 DRAFT– 31 DEC 2021
3 First-order logic

you turn to elementary model theory).

While C&H and the first third of L&K together cover overlapping
material twice, Goldrei – in a comparable number of pages – covers
very similar ground once, concentrating on a standard axiomatic proof
system. So this is a relatively gently-paced book, allowing Goldrei to be
more expansive about fundamentals, and to give a lot of examples and
exercises with worked answers to test comprehension along the way. A
great amount of thought has gone into making this text as clear and
helpful as possible. Some may find it occasionally goes a bit too slowly,
though I’d say that this is erring on the right side in an introductory
book for self-teaching: if you want a comfortingly manageable text, you
should find this particularly accessible. As with C&H and L&K, I like
Goldrei’s tone and approach a great deal.
But since Goldrei uses an axiomatic system throughout, do event-
ually supplement his book with at least some reading on a Gentzen-style
natural deduction proof system.

These three main recommended books, by the way, have all had very positive
reports over the years from student users.

3.4 Some parallel and slightly more advanced reading

The material covered in the last section is so very fundamental, and the alter-
native options so very many, that I really do need to say at least something
about a few other books. So in this section I list – in rough order of diffi-
culty/sophistication – a small handful of further texts which could well make for
useful additional or alternative reading. Then in the final section of the chapter,
I will mention some other books I’ve been asked about.
I’ll begin a notch or two down in level from the texts we have looked at so far,
with a book written by a philosopher for philosophers. It should be particularly
accessible to non-mathematicians who haven’t done much formal logic before,
and could help ease the transition to coping with the more mathematical style
of the books recommended in the last section.

5. David Bostock, Intermediate Logic (OUP 1997).

From the preface: “The book is confined to . . . what is called first-
order predicate logic, but it aims to treat this subject in very much more
detail than a standard introductory text. In particular, whereas an in-
troductory text will pursue just one style of semantics, just one method
of proof, and so on, this book aims to create a wider and a deeper un-
derstanding by showing how several alternative approaches are possible,
and by introducing comparisons between them.” So Bostock ranges more
widely than the books I’ve so far mentioned; he does indeed usefully in-
troduce you to semantic tableaux and an Hilbert-style axiomatic proof
system and natural deduction and even a sequent calculus as well. In-

26
 DRAFT– 31 DEC 2021
Some parallel and slightly more advanced reading

deed, though written for non-mathematicians, anyone could profit from

at least a quick browse of his Part II to pick up the headline news about
the various approaches.
Bostock eventually touches on issues of philosophical interest such as
free logic which are not often dealt with in other books at this level.
Still, the discussions mostly remain at much the same level of concep-
tual/mathematical difficulty as e.g. my own introductory book.

To repeat, unlike our main recommendations, Bostock does give a brisk but
very clear presentation of tableaux (‘truth trees’), and he proves completeness
for tableaux in particular, which I always think makes the needed construction
seem particularly natural. If you are a philosopher, you may well have already
encountered these truth trees in your introductory logic course. If not, at some
point you will want to find out about them (see §3.2d). As an alternative to
Bostock,

6. My elementary introduction to truth trees for propositional logic avail-

able at tinyurl.com/proptruthtrees will quickly give you the basic idea in
an accessible way. Then you can dip into my introduction to truth trees
for quantified logic at tinyurl.com/qtruthtrees.

Next, back to the level we want: and even though it is giving a second bite
to an author we’ve already met, I must mention a rather different discussion of
FOL:

7. Wilfrid Hodges, ‘Elementary predicate logic’, in the Handbook of Philo-

sophical Logic, Vol. 1, ed. by D. Gabbay and F. Guenthner, (Kluwer 2nd
edition 2001).
This is a slightly expanded version of the essay in the first edition of
the Handbook (read that earlier version if this one isn’t available), and
is written with Hodges’s usual enviable clarity and verve. As befits an
essay aimed at philosophically minded logicians, it is full of conceptual
insights, historical asides, comparisons of different ways of doing things,
etc., so it very nicely complements the textbook presentations of C&H,
L&K and/or Goldrei.
Read at this stage the very illuminating first twenty short sections.

Next, here’s a much-used text which has gone through multiple editions and
should be in any library; it is a very useful natural deduction based alternative
to C&H. Later chapters of this book are also mentioned later in this Guide as
possible reading on further topics, so it could be worth making early acquaintance
with

8. Dirk van Dalen, Logic and Structure (Springer, 1980; 5th edition 2012).
The early chapters up to and including §3.2 provide an introduction
to FOL via Gentzen-style natural deduction. The treatment is often ap-
proachable and written with a relatively light touch. However, it has to

27
 DRAFT– 31 DEC 2021
3 First-order logic

be said that the book isn’t without its quirks and flaws and inconsisten-
cies of presentation (though perhaps you have to be an alert and rather
pernickety reader to notice and be bothered by them). Still, having said
that, the coverage and general approach is good.
Mathematicians should be able to cope readily. I suspect, however,
that the book would occasionally be tougher going for philosophers if
taken from a standing start – which is another reason why I have re-
commended beginning with C&H instead. For more on this book, see
tinyurl.com/dalenlogic.
As a follow up to C&H, I just recommended L&K’s Friendly Introduction
which uses an axiomatic system. As an alternative to that, here is an older (and,
in its day, much-used) text:
9. Herbert Enderton, A Mathematical Introduction to Logic (Academic Press
1972, 2002).
This also focuses on an axiomatic system, and is often regarded as a
classic of exposition. However, it does strike me as somewhat less ap-
proachable than L&K, so I’m not surprised that students do quite often
report finding this book a bit challenging if used by itself as a first text.
However, this is an admirable and very reliable piece of work which
most readers should be able to cope with well if used as a supplementary
second text, e.g. after you have tackled C&H. And stronger mathemati-
cians might well dive into this as their first preference.
Read up to and including §2.5 or §2.6 at this stage. Later, you can
finish the rest of that chapter to take you a bit further into model theory.
For more about this classic, see tinyurl.com/enderlogicnote.
I should also certainly mention the outputs from the Open Logic Project. This
is an entirely admirable, collaborative, open-source, enterprise inaugurated by
Richard Zach, and continues to be work in progress. You can freely download the
latest full version and various sampled ‘remixes’ from tinyurl.com/openlogic. In
an earlier version of this Guide, I said that “although this is referred to as a text-
book, it is perhaps better regarded as a set of souped-up lecture notes, written
at various degrees of sophistication and with various degrees of more book-like
elaboration.” But things have moved on: the mix of chapters on propositional
and quantificational logic in the following selection has been expanded and de-
veloped considerably, and the result is much more book-like:
10. Richard Zach and others, Sets, Logic, Computation* (Open Logic: down-
loadable at tinyurl.com/slcopen).
There’s a lot to like here (Chapters 5 to 13 are the immediately relevant
ones for the moment). In particular, Chapter 9 could make for very useful
supplementary reading on natural deduction. Chapter 8 tells you about
a sequent calculus (a slightly odd ordering!). And Chapter 10 on the
completeness theorem for FOL should also prove a very useful revision
guide.
28
 DRAFT– 31 DEC 2021
Some parallel and slightly more advanced reading

My sense is that overall these discussions probably will still go some-

what too briskly for some readers to work as a stand-alone introduction
for initial self-study without the benefit of lecture support, which is why
this doesn’t feature as one of my principal recommendations in the pre-
vious section: however, your mileage may vary. And certainly, chapters
from this project could/should be very useful for reinforcing what you
have learnt elsewhere.

So much, then, for reading on FOL running on more or less parallel tracks
to the main recommendations in the preceding section. I’ll finish this section
by recommending two books that push the story on a little. First, an absolute
classic, short but packed with good things:

11. Raymond Smullyan, First-Order Logic* (Springer 1968, Dover Publica-

tions 1995).
This is terse, but those with a taste for mathematical elegance can
certainly try its Parts I and II, just a hundred pages, after the initial
recommended reading in the previous section. This beautiful little book
is the source and inspiration of many modern treatments of logic based on
tree/tableau systems. Not always easy, especially as the book progresses,
but a real delight for the mathematically minded.

And second, taking things in a new direction, don’t be put off by the title of

12. Melvin Fitting, First-Order Logic and Automated Theorem Proving (Spr-
inger, 1990, 2nd end. 1996).
A wonderfully lucid book by a renowned expositor. Yes, at a number of
places in the book there are illustrations of how to implement algorithms
in Prolog. But either you can easily pick up the very small amount of
background knowledge that’s needed to follow everything that is going
on (and that’s quite fun) or you can in fact just skip lightly over those
implementation episodes while still getting the principal logical content
of the book.
As anyone who has tried to work inside an axiomatic system knows,
proof-discovery for such systems is often hard. Which axiom schema
should we instantiate with which wffs at any given stage of a proof?
Natural deduction systems are nicer. But since we can, in effect, make
any new temporary assumption we like at any stage in a proof, again
we still need to keep our wits about us if we are to avoid going off on
useless diversions. By contrast, tableau proofs (a.k.a. tree proofs) can
pretty much write themselves even for quite complex FOL arguments,
which is why I used to introduce formal proofs to students that way
(in teaching tableaux, we can largely separate the business of getting
across the idea of formality from the task of teaching heuristics of proof-
discovery). And because tableau proofs very often write themselves, they
are also good for automated theorem proving. Fitting explores both the

29
 DRAFT– 31 DEC 2021
3 First-order logic

tableau method and the related so-called resolution method which we

mentioned as, yes, a fifth style of proof!
This book’s approach is, then, rather different from most of the other
recommended books. However, I do think that the fresh light thrown on
first-order logic makes the slight detour through this extremely clearly
written book vaut le voyage, as the Michelin guides say. (If you don’t
want to take the full tour, however, there’s a nice introduction to proofs
by resolution in Shawn Hedman, A First Course in Logic (OUP 2004):
§1.8, §§3.4–3.5.)

3.5 A little history (and some philosophy too)

(a) Classical FOL is a powerful and beautiful theory. Its treatment, in one
version or another, is always the first and most basic component of modern
textbooks or lecture courses in mathematical logic. But how did it get this status?
The first system of formalized logic of anything like the contemporary kind –
Frege’s system in his Begriffsschrift of 1879 – allows higher-order quantification
in the sense explained in the next chapter (and Frege doesn’t identity FOL as a
subsystem of distinctive interest). The same is true of Russell and Whitehead’s
logic in their Principia Mathematica of 1910–1913. It is not until Hilbert and
Ackermann in their rather stunning short book Mathematical Logic (original
German edition 1928, English translation 1950 – and still very worth reading)
that FOL is highlighted under the label ‘the restricted predicate calculus’. Those
three books all give axiomatic presentations of logic (though notationally very
different from each other): axiomatic systems similar enough to the third are
still often called ‘Hilbert-style systems’
(b) As an aside, it is worth noting that the axiomatic approach reflects a
broadly shared philosophical stance on the very nature of logic. Thus Frege
thinks of logic as a science, in the sense of a body of truths governing a cer-
tain subject matter (for Frege, they are fundamental truths governing logical
operations such as negation, conditionalization, quantification, identity). And in
Begriffsschrift §13, he extols the general procedure of axiomatizing a science to
reveal how a bunch of laws hang together: ‘we obtain a small number of laws
[the axioms] in which . . . is included, though in embryonic form, the content of
all of them’. So it is not surprising that Frege takes it as appropriate to present
logic axiomatically too.
In a rather different way, Russell also thought of logic as a science; he thought
of it as in the business of systematizing the most general truths about the world.
A special science like chemistry tells us truths about particular kinds of con-
stituents of the world and their properties; for Russell, logic tells us absolutely
general truths. If you think like that, treating logic as (so to speak) the most
general science, then of course you’ll again be inclined to regiment logic as you
do other scientific theories, ideally by laying down a few ‘basic laws’ and then
showing that other general truths follow.

30
 DRAFT– 31 DEC 2021
A little history (and some philosophy too)

Famously, Wittgenstein in the Tractatus reacted radically against Russell’s

conception of logic. For him, logical truths are tautologies. They are not deep
ultimate truths about the most general, logical, structure of the universe; rather
they are empty claims in the sense that they tell us nothing informative about
how the world is: they merely fall out as byproducts of the meanings of the basic
logical particles.
That last idea can be developed in more than one way. But one approach
is Gentzen’s in the 1930s. He thought of the logical connectives as getting their
meanings from how they are used in inference (so grasping their meaning involves
grasping the inference rules governing their use). For example, grasping ‘and’
involves grasping, inter alia, that (i) from A and B you can (of course!) derive
A. Similarly, grasping the conditional involves grasping, inter alia, that (ii) a
derivation of the conclusion C from the temporary supposition A warrants an
assertion of if A then C. But now consider this little two-step derivation:

Suppose for the sake of argument that P and Q; then we can derive
P – by the rule (i) which partly fixes the meaning of ‘and’.
And given that little suppositional inference, the rule (ii) which
partly gives the meaning of ‘if’ entitles us to drop the supposition
and conclude if P and Q, then Q.

Or in a Gentzen-style proof
[P ∧ Q](1)
P (1)
(P ∧ Q) → P)
In short, the inference rules (i) and (ii) enable us to derive that logical truth ‘for
free’ (from no remaining assumptions): it’s a theorem of a formal system with
those rules.
If this is right, and if the point generalizes, then we don’t have to see such
logical truths as reflecting deep facts about the logical structure of the world
(whatever that could mean): logical truths fall out just as byproducts of the
inference rules whose applicability is, in some sense, built into the very meaning
of e.g. the connectives and the quantifiers.
It is a nice question how far we should buy that sort of de-mystifying story
about the nature of logical truth. But whatever your eventual judgement on
that, there surely is something odd about thinking with Frege and Russell that
a systematized logic is primarily aiming to regiment a special class of ultra-
general truths. Isn’t logic at bottom about good and bad reasoning practices,
about what makes for a good proof? Shouldn’t its prime concern be the correct
styles of valid inference? And hence, shouldn’t a formalized logic highlight rules
of valid proof-building (perhaps as in a natural deduction system) rather than
stressing logical truths (as logical axioms)?
(c) Back to the history of the technical development of logic. An obvious start-
ing place is with the clear and judicious

31
 DRAFT– 31 DEC 2021
3 First-order logic

13. William Ewald, ‘The emergence of first-order logic’, The Stanford Ency-
clopaedia, tinyurl.com/emergenceFOL.

If you want rather more, the following is also readable and very helpful:

14. José Ferreirós, ‘The road to modern logic – an interpretation’, Bulletin

of Symbolic Logic 7 (2001): 441–484, tinyurl.com/roadtologic.

And for a longer, though rather bumpier, read – you’ll probably need to skim
and skip! – you could also try dipping into this more wide-ranging piece:

15. Paolo Mancosu, Richard Zach and Calixto Badesa, ‘The development
of mathematical logic from Russell to Tarski: 1900–1935’ in Leila Haa-
paranta, ed., The History of Modern Logic (OUP, 2009, pp. 318–471):
tinyurl.com/developlogic.

3.6 Postscript: Other treatments?

I will end this chapter by responding – often rather brusquely – to a variety of
Frequently Asked Questions raised in response to earlier versions of the Guide
(often questions of the form “But why haven’t you recommended X?”). So, in
what follows,

(a) I quickly mention a handful of books aimed at philosophers (but only one
will be of interest to us at this point).
(b) Next, I consider four deservedly classic books, now more than fifty years
old.
(c) Then I look at eight more recent mathematical logic texts (I again highlight
one in particular).
(d) Finally, for light relief, I look at some fun extras from an author whom we
have already met.

(a) The following five books are very varied in style, level and content, but are
all designed with philosophers particularly in mind.

(a1) Richard Jeffrey, Formal Logic: Its Scope and Limits (McGraw Hill 1967,
2nd edn. 1981).
(a2) Merrie Bergmann, James Moor and Jack Nelson, The Logic Book (McGraw
Hill 1980; 6th edn. 2013).
(a3) John L. Bell, David DeVidi and Graham Solomon, Logical Options: An
Introduction to Classical and Alternative Logics (Broadview Press 2001).
(a4) Theodore Sider, Logic for Philosophy* (OUP, 2010).
(a5) Jan von Plato, Elements of Logical Reasoning* (CUP, 2014).

Quick comments: Sider’s book (a4) falls into two halves, and the second half
is quite good on modal logic; but the first half of the book, the part which is
relevant to us now, is very poor. Only the first two chapters of Logical Options
(a3) are on FOL, and not at the level we really want. Von Plato’s Elements (a5)
32
 DRAFT– 31 DEC 2021
Postscript: Other treatments?

is good but better regarded, I think, as an introduction to proof theory and we

will return to it in Chapter 9.
The Logic Book (a2) is over 550 pages, starting at about the level of my
introductory book, and going as far as results like a full completeness proof for
FOL, so its coverage overlaps considerably with the main recommendations of
§3.3. But while reliable enough, it all strikes me, like some other readers who
have commented, as very dull and laboured, and often rather unnecessarily hard
going. You can certainly do better.
So that leaves Richard Jeffrey’s lovely book. This is relatively short, and the
first half on propositional logic is mostly at a very introductory level, which
is why I haven’t mentioned it before. But if you know a little about trees for
propositional logic — as e.g. explained in the reading reference (6) in §3.4 – then
you could start at Chapter 5 and read the rest of the book with enjoyment and
illumination. For this gives a gentle yet elegant introduction to the undecidability
of FOL and a very nice proof of completeness for trees.
(b) Next, four classic books, again listed in order of publication. All of them are
worth visiting sometime, even if they are not now the first choices for beginners.
(b1) Elliott Mendelson, Introduction to Mathematical Logic (van Nostrand 1964;
Chapman and Hall/CRC, 6th edn. 2015).
(b2) Joseph R. Shoenfield, Mathematical Logic (Addison Wesley, 1967).
(b3) Stephen C. Kleene, Mathematical Logic (John Wiley 1967; Dover Publica-
tions 2002).
(b4) Geoffrey Hunter, Metalogic (Macmillan 1971; University of California Press
1992).
Perhaps the most frequent question I used to get asked in response to early
versions of the Guide was ‘But what about Mendelson, Chs. 1 and 2’ ? Well,
(b1) was I think the first modern textbook of its type (so immense credit to
Mendelson for that), and I no doubt owe my whole career to it – it got me
through tripos when the world was a lot younger!
It seems that some others who learnt using the book are in their turn still
using it to teach from. But let’s not get too sentimental! It has to be said that
the book in its first incarnation was often brisk to the point of unfriendliness,
and the basic look-and-feel of the book hasn’t changed a great deal as it has
run through successive editions. Mendelson’s presentation of axiomatic systems
of logic are quite tough going, and as the book progresses in later chapters
through formal number theory and set theory, things if anything get somewhat
less reader-friendly. Which certainly doesn’t mean the book won’t repay working
through. But quite unsurprisingly, over fifty years on, there are many rather more
accessible and more amiable alternatives for beginning serious logic. Mendelson’s
book is a landmark well worth visiting one day, but I can’t recommend starting
here (especially for self-study). For a little more, see tinyurl.com/mendelsonlogic.
Shoenfield’s (b2) is really aimed at graduate mathematicians, and is not very
reader-friendly. Maybe take a look one day, particularly at the final chapter on
set theory; but not yet! For a little more, see tinyurl.com/schoenlogic.
33
 DRAFT– 31 DEC 2021
3 First-order logic

Kleene’s (b3) – not to be confused with his earlier and hugely influential Intro-
duction to Metamathematics – goes much more gently than Mendelson: it takes
almost twice as long to cover propositional and predicate logic, so Kleene has
much more room for helpful discursive explanations. This was in its time a rightly
much admired text, and still makes excellent and illuminating supplementary
reading.
But if you do want an old-school introduction from the same era, you might
most enjoy the somewhat less renowned book by Hunter, (b4). This is not as
comprehensive as Mendelson: but it was an exceptionally good textbook from
a time when there were few to choose from. Read Parts One to Three at this
stage. And if you are finding it rewarding reading, then do eventually finish the
book: it goes on to consider formal arithmetic and proves the undecidability of
first-order logic, topics we consider in Chapter 6. Unfortunately, the typography
– from pre-LATEX days – isn’t very pretty to look at. But in fact the treatment
of an axiomatic system of logic is extremely clear and accessible.
(c) We now turn to a number of more recent texts in mathematical logic that
have been suggested as candidates for this Guide. As you will see, the most
interesting of them – which almost made the cut to be included in §3.4’s list of
additional readings – is the idiosyncratic book by Kaye.

(c1) H.-D. Ebbinghaus, J. Flum and W. Thomas, Mathematical Logic (Springer,

2nd edn 1994, 3rd edn. 2021).
(c2) René Cori and Daniel Lascar, Mathematical Logic, A Course with Exer-
cises: Part I (OUP, 2000).
(c3) Shawn Hedman, A First Course in Logic (OUP, 2004).
(c4) Peter Hinman, Fundamentals of Mathematical Logic (A. K. Peters, 2005).
(c5) Wolfgang Rautenberg, A Concise Introduction to Mathematical Logic (Sprin-
ger, 2nd edn. 2006).
(c6) Richard Kaye, The Mathematics of Logic (CUP 2007).
(c7) Harrie de Swart, Philosophical and Mathematical Logic (Springer, 2018)
(c8) Martin Hils and François Loeser, A First Journey Through Logic (AMS
Student Mathematical Library, 2019).

I have added the last two to the list in response to queries. But while the relevant
Chapters 2 and 4 of (c7) are quite attractively written, and have some interest,
there also are a number of presentation choices I’d quibble with. You can do
better. While (c8) just isn’t designed to be a conventional mathematical logic
text. It does have a fast-track introduction to FOL, but this is done far too fast
to be of much use to anyone. We can ignore it.
So going back to earlier texts, Ebbinghaus, Flum and Thomas’s (c1) is the
English translation of a book first published in German in 1978, and appears in a
series ‘Undergraduate Texts in Mathematics’, which indicates the intended level.
The book is often warmly praised and is (I believe) quite widely used in Germany.
There is a lot of material here, often covered well. But I can’t find myself wanting
to recommend it as a good place to start. The core material on the syntax

34
 DRAFT– 31 DEC 2021
Postscript: Other treatments?

and semantics of first-order logic in Chs 2 and 3 is presented more accessibly

and more elegantly elsewhere. And the treatment of a sequent calculus Ch. 4
strikes me as poor, with the authors failing to capture the elegance that using
a sequent calculus can bring. You can freely download the old second edition at
tinyurl.com/EFTlogic. For more on this book, see tinyurl.com/EFTbooknote.
Chapters 1 and 3 of Cori and Lascar’s (c2) could appeal to the more math-
ematical reader. Chapter 1 is on semantic aspects of propositional logic, and is
done clearly. Also, an unusually good feature of the book, there are – as with
other chapters – interestingly non-trivial exercises, with expansive answers given
at the end. Chapter 2, I would say, jumps to a significantly more demanding level,
introducing Boolean algebras (and really, you should probably know just a bit of
algebra and topology to fully appreciate what is going on). Chapter 3 gets back
on track with the syntax and semantics of predicate languages, plus a smidgin
of model theory too. Not perhaps, the place to start for a first introduction to
this material, but worth reading. Then Chapter 4, the last in the book, is on
proof systems, but probably not so helpful.
Shawn Hedman’s (c2) is subtitled ‘An Introduction to Model Theory, Proof
Theory, Computability and Complexity’. So there is no lack of ambition in the
coverage! The treatment of basic FOL is patchy, however. It is pretty clear
on semantics, and the book indeed can be recommended to more mathematical
readers for its treatment of more advanced model-theoretic topics (see §5.3 in this
Guide). But Hedman offers a peculiarly ugly not-so-natural deductive system.
By contrast though – as already noted – he is good on so-called resolution
proofs. For more about what does and what doesn’t work in Hedman’s book,
see tinyurl.com/hedmanbook.
Peter Hinman’s (c3) is a massive 878 pages, and as you’d expect covers a
great deal. Hinman is, however, not really focused on deductive systems for logic,
which don’t make an appearance until over two hundred pages into the book (his
concerns are more model-theoretic). And most readers will find this book pretty
tough going. This is certainly not, then, the place to start with FOL. However,
the first three chapters of the book do contain some supplementary material that
could be very interesting once you have got hold of the basics from elsewhere,
and could particularly appeal to mathematicians. For more about what does and
what doesn’t work in Hinman’s book, see tinyurl.com/hinmanbook.
Wolfgang Rautenberg’s (c4) has some nice touches. But I suspect that its first
hundred pages on FOL are rather too concise to serve most readers as an initial
introduction; and its preferred formal system is not a ‘best buy’ either. Can be
recommended as good revision material, though.
Finally, Richard Kaye is the author of a particularly attractively written 1991
classic on models of Peano Arithmetic (we will meet this in §12.3). So I had
high hopes for his later The Mathematics of Logic (c5). “This book”, he writes,
“presents the material usually treated in a first course in logic, but in a way
that should appeal to a suspicious mathematician wanting to see some genuine
mathematical applications. . . . I do not present the main concepts and goals of
first-order logic straight away. Instead, I start by showing what the main math-

35
 DRAFT– 31 DEC 2021
3 First-order logic

ematical idea of ‘a completeness theorem’ is, with some illustrations that have
real mathematical content.” So the reader is taken on a mathematical journey
starting with König’s Lemma (I’m not going to explain that here!), and progress-
ing via order relations, Zorn’s Lemma (an equivalent to the Axiom of Choice),
Boolean algebras, and propositional logic, to completeness and compactness of
first-order logic. Does this very unusual route work as an introduction? I am
not at all convinced. It seems to me that the journey is made too bumpy and
the road taken is far too uneven in level for this to be appealing as an early
trip through first-order logic. However, if you already know a fair amount of this
material from more conventional presentations, the different angle of approach
in this book linking topics together in new ways could well be very interesting
and illuminating.
(d) I have already strongly recommended Raymond Smullyan’s 1968 classic
First-Order Logic. Smullyan went on to write some absolutely classic texts on
Gödel’s theorem and on recursive functions, which we’ll be mentioning later.
But as well as these, he also wrote many ‘puzzle’-based books aimed at a wider
audience, including e.g. the rightly renowned What is the Name of This Book? *
(Dover Publications reprint of 1981 original, 2011) and The Gödelian Puzzle
Book * (Dover Publications, 2013).
Smullyan has also written Logical Labyrinths (A. K. Peters, 2009). From the
blurb: “This book features a unique approach to the teaching of mathematical
logic by putting it in the context of the puzzles and paradoxes of common lan-
guage and rational thought. It serves as a bridge from the author’s puzzle books
to his technical writing in the fascinating field of mathematical logic. Using the
logic of lying and truth-telling, the author introduces the readers to informal
reasoning preparing them for the formal study of symbolic logic, from propo-
sitional logic to first-order logic, . . . The book includes a journey through the
amazing labyrinths of infinity, which have stirred the imagination of mankind as
much, if not more, than any other subject.”
Smullyan starts, then, with puzzles, of this kind: you are visiting an island
where there are Knights (truth-tellers) and Knaves (persistent liars) and then in
various scenarios you have to work out what’s true given what the inhabitants
say about each other and the world. And, without too many big leaps, he ends
with first-order logic (using tableaux), completeness, compactness and more. To
be sure, this is no substitute for standard texts: but – for those with a taste for
being led up to the serious stuff via sequences of puzzles – a very entertaining
and illuminating supplement.
(Smullyan’s later A Beginner’s Guide to Mathematical Logic*, Dover Publi-
cations, 2014, is rather more conventional. The first 170 pages are relevant to
FOL. A rather uneven read, it seems to me; but again an engaging supplement
to the main texts recommended above.)

36
 DRAFT– 31 DEC 2021

4 Second-order logic, quite briefly

Classical first-order logic contrasts along one dimension with various non-classical
logics, and along another dimension with second-order and higher-order logics.
We can leave the exploration of non-classical logics to later chapters, starting
with Chapter 8. I will, however, say a little about second-order logic straight
away, in this chapter. Why?
Theories expressed in first-order languages with a first-order logic turn out to
have their limitations – that’s a theme that will recur when we look at model
theory (Chapter 5), theories of arithmetic (Chapter 6), and set theory (Chap-
ter 7). And you will occasionally find explicit contrasts being drawn with richer
theories expressed in second-order languages with a second-order logic. So, al-
though it’s a judgement call, I think it is worth getting to know just a bit about
second-order logic quite early on, in order to understand the contrasts being
drawn.
But first, . . .

4.1 A preliminary note on many-sorted logic

(a) As you will now have seen from the core readings, FOL is standardly pre-
sented as having a single ‘sort’ of quantifier, in the sense that all the quantifiers
in a given language run over one and the same domain of objects. But this is
artificial, and certainly doesn’t conform to everyday mathematical practice.
To take an example which will be very familiar to mathematicians, consider
the usual practice of using one style of variable for scalars and another for vectors,
as in the rule for scalar multiplication:

(1) a(v1 + v2 ) = av1 + av2 .

If we want to make the generality here explicit, we could very naturally write

(2) ∀a∀v1 ∀v2 (v1 + v2 ) = av1 + av2 ,

with the first quantifier understood as running just over scalars, and with the
other two quantifiers running just over vectors. Or we could explicitly declare
which domain a quantified variable is running over by using a notation like
(∀a : S) to assign a to scalars: mathematicians often do this informally. (And in

37
 DRAFT– 31 DEC 2021
4 Second-order logic, quite briefly

some formal ‘type theories’, this kind of notation becomes the official policy: see
§12.7.)
It might seem strange, then, to insist that, if we want to formalize our theory
of vector spaces, we should follow FOL practice and use only one sort of variable
and therefore have to render the rule for scalar multiplication along the lines of
(3) ∀x∀y∀z((Sx ∧ Vy ∧ Vz) → x(y + z) = xy + xz),
i.e. ‘Take any three things in our [inclusive] domain, if the first is a scalar, the
second is a vector, and the third is a vector, then . . . ’.
(b) In sum, the theory of vector spaces is naturally regimented using a two-
sorted logic, with two sorts of variables running over two different domains. So,
generalizing, why not allow a many-sorted logic – allowing multiple independent
domains of objects, with different sorts of variables restricted to running over
the different domains?
In fact, it isn’t hard to set up such a revised version of FOL (it is first-order,
as the quantifiers are still of the now familiar basic type, running over objects
in the relevant domains). The syntax and semantics of a many-sorted language
can be defined quite easily. Syntactically, we will need to keep a tally of the
sorts assigned to the various names and variables. And we will also need rules
about which sorts of terms can go into which slots in predicates and in function-
expressions (for example, only terms for vectors should be used as inputs to
the vector-addition function). Semantically, we assign a domain for each sort
of variable, and then proceed pretty much as in the one-sorted case. Assuming
that each domain is non-empty (as in standard FOL) the inference rules for a
deductive system will then look entirely familiar. And the resulting logic will
have the same nice technical properties as standard one-sorted FOL; crucially,
you can prove soundness and completeness and compactness theorems in just
the same ways.
(c) As so often in the formalization game, we are now faced with a cost/benefit
trade-off. We can get the benefit of somewhat more natural regimentations of
mathematical practice, at the cost of having to use a slightly more complex many-
sorted logic. Or we can pay the price of having to use less natural regimentations
– we need to translate propositions like (2) by using restricted quantifications
like (3) – but get the benefit of a slightly-simpler-in-practice logic.1
So you pays your money and you takes your choice. For many (most?) pur-
poses, logicians prefer the second option, sticking to standard single-sorted FOL.
That’s because, at the end of the day, we care rather less about elegance when
regimenting this or that theory than about having a simple-but-powerful logical
system.
1 Note though that we do also get some added flexibility on the second option. The use of
a sorted quantifier ∀aFa with the usual logic presupposes that there is at least one thing
in the relevant domain for the variable a. But a corresponding restricted quantification
∀x(Ax → Fx), where the variable x quantifies over some wider domain while A picks out the
relevant sort which a was supposed to run over, leaves open the possibility that there is
nothing of that sort.

38
 DRAFT– 31 DEC 2021
Second-order logic: a brief overview

4.2 Second-order logic: a brief overview

(a) Now we turn from ‘sorts’ to ‘orders’. It will help to fix ideas if we begin with
an easy arithmetical example; so consider the informal principle of induction:

(Ind 1) Take any numerical property X; if (i) zero has X and (ii) any number
which has X passes it on to its successor, then (iii) all numbers must
share property X.

This holds, of course, because every natural number is either zero or is an even-
tual successor of zero (i.e. is either 0 or 00 or 000 or 0000 or . . . , where the prime ‘0 ’
is a standard sign for the function that maps a number to its successor). There
are no stray numbers outside that sequence, so a property that percolates down
the sequence eventually applies to any number at all.
There is no problem about expressing some particular instances of the in-
duction principle in a first-order language. For example, suppose P is a formal
one-place predicate expressing some particular arithmetical property: then we
can express the induction principle for this property by writing

(Ind 2) (P0 ∧ ∀x(Px → Px0 )) → ∀x Px

where the small-‘x’ quantifier runs over the natural numbers and again the prime
expresses the successor function. But how can we state the general principle
of induction in a formal language, the principle that applies to any numerical
property? The natural candidate is something like this:

(Ind 3) ∀X((X0 ∧ ∀x(Xx → Xx0 )) → ∀x Xx).

Here the big-‘X’ quantifier is a new type of quantifier, which unlike the small-
‘x’ quantifier, quantifies ‘into predicate position’. In other words, it quantifies
into the position occupied in (Ind 2) by the predicate ‘P’, and the expressed
generalization is intended to run over all properties of numbers, so that (Ind 3)
indeed formally renders (Ind 1). But this kind of quantification – second-order
quantification – is not available in standard first-order languages of the kind that
you now know and love.
If we do want to stick with a theory framed in a first-order arithmetical lan-
guage L which just quantifies over numbers, the best we can do to render the
induction principle is to use a template or schema and say something like

(Ind 4) For any arithmetical L-predicate A( ), simple or complex, the cor-

responding wff of the form (A(0) ∧ ∀x(A(x) → A(x0 )) → ∀x A(x) is an
axiom.

However (Ind 4) is much weaker than the informal (Ind 1) or the equivalent
formal version (Ind 3) on its intended interpretation. For (Ind 1/3) tells us that
induction holds for any property at all ; while, in effect, (Ind 4) only tells us that
induction holds for those properties that can be expressed by some L-predicate
A( ).

39
 DRAFT– 31 DEC 2021
4 Second-order logic, quite briefly

(b) Another issue to think about. Start with a definition:

Suppose R is a binary relation. Define Rn (for n > 0) to be the
relation that holds between a and b when there are n R-related links
between them – i.e. when there are objects x1 , x2 , . . . xn such that
Rax1 , Rx1 x2 , Rx2 x3 , . . . , Rxn b. And take R0 just to be R.
Then R∗ , the ancestral of R, is the relation that holds between a
and b just when there is some n ≥ 0 such that Rn ab – i.e. just when
there is a finite chain of R-related links between a and b.
Example: if R is the relation ‘is a parent of’, then R∗ is the relation ‘is a direct
ancestor of’. Which explains ‘ancestral’ ! An arithmetical example: if S is the
relation ‘is the successor of’, then S ∗ nm holds when there is a sequence of
successors starting with m and finishing with n. And n is a natural number
just if S ∗ n0.
Now three easy observations:

(i) First note that given a relational predicate R expressing the relation R, we
can of course define complex expressions, which we can abbreviate Rn , to
express the corresponding relations Rn . For example, we just put
R3 ab =def ∃x1 ∃x2 ∃x3 (Rax1 ∧ Rx1 x2 ∧ Rx2 x3 ∧ Rx3 b).
Now suppose we can construct an expression R∗ for the ancestral of the
relation expressed by R. Then consider the infinite set of wffs
{¬Rab, ¬R1 ab, ¬R2 ab, ¬R3 ab, . . . , ¬Rn ab, . . . , R∗ ab}
Then (X) every finite collection of these wffs has a model (let n be the
largest index appearing, and consider the case where a is the ancestor of b
more than n generations removed). But obviously (Y) the whole infinite set
of sentences doesn’t have a model (a can’t be an R-ancestor of b without
there being some n such that Rn ab).
(ii) Now, if we stay first-order, then we know the compactness theorem holds.
That means for first-order wffs we can’t have both (X) and (Y). Hence,
we can’t after all construct an expression R∗ from R and first-order logical
apparatus. In short, we can’t define the ancestral of a relation in first-order
logic.
(iii) On the other hand, a little reflection shows that a stands in the ancestral
of the R-relation to b just in case b inherits ever property that is had by
any R-child of a, and then always preserved by the R relation (why?). And
that’s why Frege could define the ancestral using second-order apparatus
like this:
R∗ ab =def ∀X[(∀x(Rax → Xx) ∧ ∀x∀y(Xx ∧ Rxy → Xy)) → Xb]
And note that, since we can construct a second-order expression R∗ for the
ancestral of the relation expressed by R, then – because (X) and (Y) are
true together – compactness must fail for second-order logic.
40
 DRAFT– 31 DEC 2021
Second-order logic: a brief overview

In sum, we can’t define the ancestral of a relation in first-order logic (and

hence can’t define equivalent notions like transitive closure either), while we can
do so in second-order logic. So we see that – as with induction – allowing quan-
tification into predicate position increases the expressive power of our language
in a mathematically significant way.
(c) And it isn’t difficult to extend the syntax and semantics of first-order lan-
guages to allow for second-order quantification. Start with simple cases.
The required added syntax is unproblematic.

Recall how we can take a formula A(n) containing some occurrence(s)

of the name ‘n’, swap out the name on each occurrence for a partic-
ular (small) variable, and then form a first-order quantified wff like
∀xA(x).
We just need now to add the analogous rule that we can take a
formula A(P) containing some occurrence(s) of the unary predicate
‘P’, swap out the predicate for some (big) variable and then form a
second-order quantified wff of the form ∀XA(X).

Fine print apart, that’s straightforward.

The standard semantics is equally straightforward. We interpret names, predi-
cates and functions just as before, and likewise for the connectives and first-order
quantifiers. And again we model the story about the novel second-order quanti-
fiers on the account of first-order quantifiers. So first fix a domain of quantifica-
tion.

Recall that, roughly, ∀xA(x) is true on a given interpretation of its

language just when A(n) remains true, however we vary the object
in the domain which is assigned to the name ‘n’ as its interpretation.
Similarly then, ∀XA(X) is true on an interpretation just when
A(P) remains true, however we vary the subset of the domain which is
assigned to the unary predicate ‘P’ as its interpretation (i.e. however
we vary ‘P’s extension).

Again, there’s fine print; but you get the general idea.
We’ll now also want to expand the syntactic and semantic stories further
to allow second-order quantification over binary and other relations and over
functions too; but these expansions raise no extra issues.
We can then define the relation of semantic consequence for formulas in our
extended languages including second-order quantifiers in the now familiar way:

Some formulas Γ semantically entail A just in case every interpreta-

tion over a structure that makes all of Γ true makes A true.

(d) So, in bald summary, the situation is this. There are quite a few famil-
iar mathematical claims like the arithmetical induction principle, and familiar
mathematical constructions like forming transitive closures which are naturally

41
 DRAFT– 31 DEC 2021
4 Second-order logic, quite briefly

regimented using quantifications over properties (and/or relations and/or func-

tions). And there is no problem about augmenting the syntax and semantics
of our formal languages to allow such second-order quantifications, and we can
carry over the definition of semantic entailment to cover sentences in the result-
ing second-order languages.
Moreover, theories framed in second-order languages turn out to have nice
properties which are lacked by their first-order counterparts. For example, a
theory of arithmetic with the full second-order induction principle (Ind 3) will
be ‘categorical’, in the sense of having just one kind of structure as a model (a
model built from a zero, its eventual successors, and nothing else). On the other
hand, as you will see in due course, a first-order theory of arithmetic which has
to rely on a limited induction principle like (Ind 4) will have models of quite
different kinds (as well as the intended model with just a zero and its eventual
successors, there will be an infinite number of different ‘non-standard’ models
which have unwanted junk in their domains).
The obvious question which arises from all this, then, is why is the standard
modern practice to privilege FOL? Why not adopt a second-order logic from the
outset as our preferred framework for regimenting mathematical arguments? –
after all, as noted in §3.5, early formal logics like Frege’s allowed more than
first-order quantifiers.
(e) The short answer is: because there can be no sound and complete formal
deductive system for second-order logic.
There can be be sound but partial deductive systems S2 for a language in-
cluding second-order quantifiers. So we can have the one-way conditional that,
whenever there is an S2 -proof from premisses in Γ to the conclusion A, then Γ
semantically entails A. But the converse fails. We can’t have a respectable for-
mal system S2 (where it is decidable what’s a proof, etc.) such that, whenever
Γ semantically entails A, there is an S2 -proof from premisses in Γ to the con-
clusion A. Once second-order sentences (with their standard interpretation) are
in play, we can’t fully capture the relation of semantic entailment in a formal
deductive system.
(f) Let’s pause to contrast the case of a two-sorted first-order language of the
kind we met in the previous section. In that case, the two sorts of quantifier
get interpreted quite independently – fixing the domain of one doesn’t fix the
domain of the other. And it is because each sort of quantifier, as it were, stands
alone, a familiar kind of first-order logic continues to each seperately.
But in second-order logic it is entirely different. For note that on the standard
semantic story, it is now the same domain which fixes the intepretation of both
kinds of quantifier – i.e. one and the same domain both provides the objects for
the first-order quantifiers to range over, and also provides the sets of objects (e.g.
all the subsets of the original domain) for the second-order quantifiers to range
over. The interpretations of the two kinds of quantifier are tightly connected, and
this makes all the difference; it is this which blocks the possibility of a complete
deductive system for second-order logic.

42
 DRAFT– 31 DEC 2021
Second-order logic: a brief overview

(Technical note: If we drop the requirement of standard or ‘full’ semantics

that the second-order big-‘X’ quantifiers run over all the subsets of the domain
of the corresponding first-order small-‘x’ quantifiers, we will arrive what’s called
‘Henkin semantics’ or ‘general semantics’. And on this semantics we can regain
a completeness theorem, but we lose the other nice features that second-order
theories have on their natural ‘standard’ semantics.)
(g) Of course, it’s not supposed to obvious at the outset that we can’t have a
complete deductive system for second-order logic with the standard semantics,
any more than it is obvious at the outset that we can have a complete deductive
system for first-order logic!
However, we have now shown in (b) that compactness fails in the second-
order case, and that is enough to show that we can’t have a strongly complete
deductive system for second-order logic with standard semantics (just recycle
the ideas of §3.1, fn.2). But it takes much more work to show that we can’t
even have a weakly complete proof system: the usual argument relies on Gödel’s
incompleteness theorem.
And it isn’t obvious either what exactly the significance the failure of com-
pleteness might be. In fact, the whole question of the status of second-order logic
leads to some tangled debates.
Let’s briefly touch on one disputed issue. On the usual story, when we give
the semantics of FOL, we interpret one-place predicates by assigning them sets
as extensions. And when we now add second-order quantifiers, we are adding
quantifiers which are correspondingly interpreted as ranging over all these possi-
ble extensions. So, you might well ask, why not frankly rewrite our second-order
induction principle (Ind 3), for example, in the form
(Ind 5) ∀X((0 ∈ X ∧ ∀x(x ∈ X → x0 ∈ X) → ∀x x ∈ X),
making it explicit that the big-‘X’ variable is running over sets? Well, we can do
that. Though if (5) is to replicate the content of (3) on its standard semantics, it
is crucial that the big-‘X’ variable has to run over all the subsets of the domain
of the small-‘x’ variable.
And now some would say that, because (Ind 3) can be rewritten as (Ind 5),
this just goes to show that in using second-order quantifiers we are straying into
the realm of set theory. Others would push the connection in the other direction.
They would start by arguing that the invocation of sets in the explanation of
second-order semantics, while conventional, is actually dispensable (in the spirit
of §2.3; and see the papers by Boolos mentioned below). So this means that
(Ind 5) in fact dresses up the induction principle (Ind 3) – which is not in essence
set-theoretic – in misleadingly fancy clothing.
So we are left with a troublesome question: is second-order logic really just
some “set theory in sheep’s clothing” (as the philosopher W.V.O. Quine famously
quipped)? We can’t pursue this further here (though I give some pointers in the
final section for philosophers who want to tackle the issue). Fortunately, for the
purposes of getting to grips with the logical material of the next few chapters,
you just need to grasp a few basic technical facts about second-order logic.
43
 DRAFT– 31 DEC 2021
4 Second-order logic, quite briefly

4.3 Recommendations on many-sorted and second-order logic

First, for something on the formal details of many-sorted first-order languages
and their logic:

What little you need for present purposes is covered in four clear pages by
1. Herbert Enderton, A Mathematical Introduction to Logic (Academic
Press 1972, 2002), §4.3.

There is, however, a bit more that can be fussed over here, and some might be
interested in looking at e.g. Hans Halvorson’s The Logic in Philosophy of Science
(CUP, 2019), §§5.1–5.3
Turning now to second-order logic:

For a brief review, saying only a little more than my overview remarks, see
2. Richard Zach and others, Sets, Logic, Computation* (Open Logic) §11.3,
excerpted at tinyurl.com/openlogicSOL.
You could then look e.g. at the rest of Chapter 4 of the Enderton (1). Or,
rather more usefully at this stage, read
3. Stewart Shapiro, ‘Higher-order logic’, in S. Shapiro, ed., The Oxford
Handbook of the Philosophy of Mathematics and Logic (OUP, 2005).
You can skip §3.3; but §3.4 touches on Boolos’s ideas and is relevant
to the question of how far second-order logic presupposes set theory.
Shapiro’s §5, ‘Logical choice’, is an interesting discussion of what’s at
stake in adopting a second-order logic. (Don’t worry if some points will
only become really clear once you’ve done some model theory and some
formal arithmetic.)
To nail down some of the technical basics you can then very usefully sup-
plement the explanations in Shapiro with the admirably clear
4. Tim Button and Sean Walsh, Philosophy and Model Theory* (OUP,
2018), Chapter 1.
This chapter reviews, in a particularly helpful way, various ways of
developing the semantics of first-order logical languages; and then it
compares the first-order case with the second-order options, both ‘full’
semantics and ‘Henkin’ semantics.

For alternative very introductory reading you could look at the very clear
5. Theodore Sider, ‘Crash course on higher-order logic’, §§1–3, 5. Available
at tinyurl.com/siderHOL.
While if the initial readings leave you still wanting to fill out the technical story
about second-order logic a little further, you will then want to dive into the
self-recommending
44
 DRAFT– 31 DEC 2021
Conceptual issues

6. Stewart Shapiro, Foundations without Foundationalism: A Case for Second-

Order Logic, Oxford Logic Guides 17 (Clarendon Press, 1991), Chs. 3–5
(with Ch. 6 for enthusiasts).

4.4 Conceptual issues

So much for formal details. Philosophers who have Shapiro’s wonderfully illu-
minating book in their hands, will also be intrigued by the initial philosophi-
cal/methodological discussion in his first two chapters here. This whole book is
a modern classic, and is remarkably accessible.
Shapiro, in both his Handbook essay and in his earlier book, mentions Boo-
los’s arguments against regarding second-order logic as essentially set-theoretical.
Very roughly, the idea is that – instead of interpreting e.g. the second-order quan-
tification in the induction axiom (Ind 3) as in effect quantifying over sets – we
should read it along these lines:

(Ind 30 ) Whatever numbers we take, if 0 is one of them, and if n0 is one of them

if n is, then we have all the numbers.

So the idea is that we don’t need to invoke sets to interpret (3), just a non-
committal use of plurals. For more on this, just because he is so very readable,
let me highlight the thought-provoking

7. George Boolos, ‘On Second Order Logic’ and ‘To Be is to Be a Value of

a Variable (or to Be Some Values of Some Variables)’, both reprinted in
his wonderful collection of essays Logic, Logic, and Logic (Harvard UP,
1998).

You can then follow up some of the critical discussions of Boolos mentioned by
Shapiro.
Note, however, the usual semantics for second-order logic and Boolos’s pro-
posed alternative do share an assumption – in effect, neither treat properties
very seriously! Recall, we started off stating the informal induction principle
(Ind 1) in terms of a generalization over properties of numbers. But in interpret-
ing its second-order regimentation (Ind 3), we’ve only spoken of sets of numbers
(to serve as extensions of properties, the standard story) or spoken even more
economically, just about numbers, plural (Ind 30 , Boolos). Where have the prop-
erties gone? Philosophers, at any rate, might want to resist reducing higher-order
entities (properties, properties of properties) to first-order entities (objectsl, or
sets of objects). Now, this is most certainly not the place to enter into those
debates. But for a nice survey with pointers to relevant discussions, see

8. Lukas Skiba, ‘Higher order metaphysics’, Philosophy Compass (2021),

tinyurl.com/skibameta.

45
 DRAFT– 31 DEC 2021

5 Model theory

The high point of a first serious encounter with FOL is the proof of the complete-
ness theorem. Introductory texts then usually discuss at least a couple of quick
corollaries of the proof – the compactness theorem (which we’ve already met)
and the downward Löwenheim-Skolem theorem. And so we take initial steps into
what we can call Level 1 model theory. Further along the track we will encounter
Level 3 model theory (I am thinking of the sort of topics covered in e.g. the later
chapters of the now classic texts by Wilfrid Hodges and David Marker which
are recommended as advanced reading in §12.2). In between, there is a stretch
of what we can think of as Level 2 theory – still relatively elementary, relatively
accessible without too many hard scrambles, but going somewhat beyond the
very basics.
Putting it like this in terms of ‘levels’ is of course only for the purposes of
rough-and-ready organization: there are no sharp boundaries to be drawn. In a
first foray into mathematical logic, though, you should certainly get your head
around Level 1 model theory. Then tackle as much Level 2 theory as grabs your
interest.
But what topics can we assign to these first two levels?

5.1 Elementary model theory: an overview

(a) Model theory is about mathematical structures and about how to charac-
terize and classify them using formal languages. Put another way, it concerns
the relationship between a mathematical theory (regimented as a collection of
formal sentences) and the structures which ‘realize’ that theory (i.e. the struc-
tures which we can interpret the theory as being true of, i.e. the structures which
provide a model for the theory).
It will help to have in mind a sample range of theories and corresponding struc-
tures. For example, it is good to know just a little about theories of arithmetic,
algebraic theories (like group theory or Boolean algebra), theories of various
kinds of order, etc., and also to know just a little about some of the structures
which provide models for these theories. Mathematicians will already be famil-
iar with informally presented examples: philosophers will probably need to do a
small amount of preparatory homework here (but the first reading recommen-
dation in the next section should provide enough to start you off).

46
 DRAFT– 31 DEC 2021
Elementary model theory: an overview

Here are some initial themes we’ll need to explore:

(1) We’ll be interested in relations between structures. One structure can be

simply a substructure of another, or can extend another. Or we can map
one structure to another in a way that preserves structural information –
so, for example, a structure-preserving map can send one structure to a
copy embedded inside another structure. In particular, we will be interested
in the case where there’s an isomorphism between structures, so that each
is a replica of the other (as far as their structural features are concerned).
We will similarly be interested in relations between languages for de-
scribing structures – we can expand or reduce the non-logical resources
of a language, potentially giving it greater or lesser expressive power. So
we will also want to know something about the interplay between these
expansions/reductions of structures and corresponding languages for them.
(2) How much can a language tell us about a structure? For a toy example, take
the structure (N, <), i.e. the natural numbers equipped with their standard
order relation. And consider the first-order formal language whose sole bit
of non-logical vocabulary is a symbol for the order relation (let’s re-use
< for this, with context making it clear that this now is an expression
belonging to a formal language!). Then, note that we can e.g. define the
successor relation over N in this language, using the formula

x < y ∧ ∀z(x < z → (z = y ∨ y < z))

with the quantifier running over N. For evidently a pair of numbers x, y

satisfies this formula if y comes immediately after x in the ordering. And
given we can define the successor relation, we can now e.g. define 0 as the
number in the structure (N, <) which isn’t a successor of anything.
Now take instead the structure (Z, <), i.e. all the integers, negative and
positive, equipped with their standard order relation. And consider the
corresponding formal language where < gets re-interpreted accordingly.
The same formula as before, but with the quantifier now running over
Z, also suffices to define the successor relation over the integers. But this
time, we obviously can’t define 0 as the integer which isn’t a successor (all
integers are successors!). And in fact no other expression from the formal
language whose sole bit of non-logical vocabulary is the order-predicate <
will define the zero in (Z, <). Rather as you would expect, the ordering
relation gives only the relative position of integers, but doesn’t fix the zero.
OK, those were indeed trivial toy examples! But they illustrate a very
important class of questions of the following form: which objects and rela-
tions in a particular structure can be pinned down, which can be defined,
using expressions from a first-order language for the structure?
(3) Moving from what can be defined by particular expressions to the question
of what gets fixed by a whole theory, we can ask how varied the models
of a given theory can be. In many cases, quite different structures for

47
 DRAFT– 31 DEC 2021
5 Model theory

interpreting a given language can be ‘elementarily equivalent’, meaning

that they satisfy all the same sentences of the language. At the other
extreme, a theory like second-order Peano Arithmetic is categorical – its
models will all ‘look the same’, i.e. are all isomorphic with each other.
Categoricity is good when we can get it: but when is it available? We’ll
return to this in a moment.
(4) Instead of going from a theory to the structures which are its models, we
can go from structures to theories. Given a class of structures, we can ask:
is there a first-order theory for which just these structures are the models?
Or given a particular structure, and a language for it with the right sort of
names, predicates and functional expressions, we can look at the set of all
the sentences in the language which are true of the structure. We can now
ask, when can all those sentences be regimented into a nicely axiomatized
theory? Perhaps we can find a finite collection of axioms which entails all
those truths about the structure: or if a finite set of axioms is too much
to hope for, perhaps we can at least get a set of axioms which are nicely
disciplined in some other way. And when is the theory for a structure (i.e.
the set of sentences true of the structure) decidable, in the sense that a
computer could work out what sentences belong to the theory?

(b) Now, you have already met a pair of fundamental results linking semantic
structures and sets of first-order sentences – the soundness and completeness
theorems. And these lead to a pair of fundamental model-theoretic results. The
first of these we’ve met before, at end of §3.1:

(5) The compactness theorem (a.k.a. the finiteness theorem). If every finite
subset of a set of sentences Γ from a first-order language has a model, so
does Γ.

For our second result, revisit a standard completeness proof for FOL, which
shows that any syntactically consistent set of sentences from a first-order lan-
guage (set of sentences from which you can’t derive a contradiction) has a model.
Look at the details of the proof: it gives an abstract recipe for building the
required model. And assuming that we are dealing with normal first-order lan-
guages (with a countable vocabulary), you’ll find that the recipe delivers a count-
able model – so in effect, our proof shows that a syntactically consistent set of
sentences has a model whose domain is just (some or all) the natural numbers.
From this observation we get

(6) The downward Löwenheim-Skolem theorem. Suppose a bunch of sentences

Γ from a countable first-order language L has a model (however large);
then Γ has a countable model.

Why so? Suppose Γ has a model. Then it is syntactically consistent in your

favoured proof system (for if we could derive absurdity from Γ then, by the

48
 DRAFT– 31 DEC 2021
Elementary model theory: an overview

soundness theorem, Γ would semantically entail absurdity, i.e. would be seman-

tically inconsistent after all and have no model). And since Γ is syntactically
consistent then, by our proof of completeness, Γ has a countable model.
Note: compactness and the L-S theorem are both results about models, and
don’t themselves mention proof-systems. So you’d expect we ought to be able to
prove them directly without appeal to the completeness theorem which mentions
proof-systems. And we can!
(c) An easy argument shows that we can’t consistently have (i) for each n a
sentence ∃n which is says that there are at least n things, (ii) a sentence ∃∞
which is true in all and only infinite domains, and also (iii) compactness.1 In the
second-order case we can have (i) and (ii), so that rules out compactness. In the
first-order case, we have (i) and (iii); hence

(7) There is no first-order sentence ∃∞ which is true in all and only structures
with infinite domains.

That’s a nice mini-result about the limitations of first-order languages. But now
let’s note a second, much more dramatic, such result.
Suppose LA is a formal first-order language for the arithmetic of the natural
numbers. The precise details don’t matter; but to fix ideas, suppose LA ’s built-
in non-logical vocabulary comprises the binary function expressions + and ×
(with their obvious interpretations), the unary function expression 0 (expressing
the successor function), and the constant 0 (denoting zero). So note that LA
then has a sequence of expressions 0, 00 , 000 , 0000 , . . . which can serve as numerals,
denoting 0, 1, 2, 3, . . . .
Now let the theory Ttrue , i.e. true arithmetic, be the set of all true LA sen-
tences. Then we can show the following:

(8) As well as being true of its ‘intended model’ – i.e. the natural numbers
with their distinguished element zero and the successor, addition, and mul-
tiplication functions defined over them – Ttrue is also true of differently-
structured, non-isomorphic, models.

This can be shown again by an easy compactness argument.2

1 Consider the infinite set of sentences

Γ =def {∃1, ∃2, ∃3, ∃4, . . . , ¬∃∞}
Any finite subset ∆ ⊂ Γ has a model (because there will be a maximum number n such
that ∃n is in ∆ – and then all the sentences in ∆, which might include ¬∃∞, will be true
in a structure whose domain contains exactly n objects). Compactness would then imply
that Γ has a model. But that’s impossible. No structure can have a domain which both
does have at least n objects for every n and also doesn’t have infinitely many objects. So
compactness fails.
2 Indulge me! Let me give the proof idea, because it is so very neat. For brevity, write n as
short for 0 followed by n occurrences of the prime 0 : so n denotes n.
OK: let’s add to the language LA the single additional constant ‘c’. And now consider
+
the theory Ttrue formed in the expanded languages, which has as its axioms all of Ttrue plus
the infinite supply of extra axioms 0 6= c, 1 6= c, 2 6= c, 3 6= c, . . ..

49
 DRAFT– 31 DEC 2021
5 Model theory

And this is really rather remarkable! Formal first-order theories are our stan-
dard way of regimenting informal mathematical theories: but now we find that
even Ttrue – the set of all first-order LA truths taken together – still fails to pin
down a unique structure for the natural numbers.
(d) And, turning now to the L-S theorem, we find that things only get worse.
Again let’s take a dramatic example.
Suppose we aim to capture the set-theoretic principles we use as mathemati-
cians, arriving at the gold-standard Zermelo-Fraenkel set theory with the Axiom
of Choice, which we regiment as the first-order theory ZFC. Then:

(9) ZFC, on its intended interpretation, makes lots of infinitary claims about
the existence of sets much bigger than the set of natural numbers. But the
downward Löwenheim-Skolem theorem tells us that, all the same, assum-
ing ZFC is consistent and has a model at all, it has an unintended countable
model (despite the fact that ZFC has a theorem which on the intended in-
terpretation says that there are uncountable sets). In other words, ZFC has
an interpretation in the natural numbers. Hence our standard first-order
formalized set theory certainly fails to uniquely pin down the wildly infini-
tary universe of sets – it doesn’t even manage to pin down an uncountable
universe.

What is emerging then, in these first steps into model theory, are some very
considerable and perhaps unexpected(?) expressive limitations of first-order for-
malized theories. These limitations can be thought of as one of the main themes
of Level 1 model theory.
(e) At Level 2, we can pursue this theme further, starting with the upward
Löwenheim-Skolem theorem which tells us that if a theory has an infinite model
it will also have models of all larger infinite sizes (as you see, then, you’ll need
some basic grip on the idea of the hierarchy of different cardinal sizes to make
full sense of this sort of result). Hence

(10) The upward and downward Löwenheim-Skolem theorems tell us that first-
order theories which have infinite models won’t be categorical – i.e. their

+
Now observe that any finite collection of sentences ∆ ⊂ Ttrue has a model. Because ∆
is finite, there will be a some largest number n such that the axiom n 6= c is in ∆; so just
interpret c as denoting n + 1 and give all the other vocabulary its intended interpretation,
and every sentence in the finite set ∆ will by hypothesis be true on this interpretation.
+ +
Since any finite ∆ ⊂ Ttrue has a model, Ttrue itself has a model, by compactness. That
model, as well as having a zero and its successors, must also have in its domain a non-
standard ‘number’ c to be the denotation of the new name c (where c is distinct from the
denotations of 0, 1, 2, 3, . . .). And note, since the new model must still make true e.g. the
old Ttrue sentence which says that everything in the domain has a successor, there will in
addition be more non-standard numbers to be successor of c, the successor of that, etc.
+
Now take a structure which is a model for Ttrue , with its domain including non-standard
+
numbers. Then in particular it makes true all the sentences of Ttrue which don’t feature the
constant c. But these are just the sentences of the original Ttrue . So this structure will still
make all Ttrue true – even though its domain contains more than a zero and its successors,
and so does not ‘look like’ the original intended model.

50
 DRAFT– 31 DEC 2021
Elementary model theory: an overview

models won’t all look the same because they can have domains of different
infinite sizes. For example, try as we might, a first-order theory of arith-
metic will always have non-standard models which ‘look too big’ to be
the natural numbers with their usual structure, and a first-order theory of
sets will always have non-standard models which ‘look too small’ to be the
universe of sets as we intuitively conceive it.
But if we can’t achieve full categoricity (all models looking the same),
perhaps we can get restricted categoricity results for some theories (telling
us that all models of a certain size look the same) – when is this possible?
An example you’ll find discussed: the theory of dense linear orders is
countably categorical (i.e. all its models of the size of the natural numbers
are isomorphic – a lovely result due to Cantor); but it isn’t categorical at
the next infinite size up. On the other hand, theories of first-order arith-
metic are not even countably categorical (even if we restrict ourselves to
models in the natural numbers, there can be models which give deviant
interpretations of successor, addition and multiplication).
How does that last claim square with the proof you often meet early in a maths
course that a theory usually called ‘Peano Arithmetic’ is categorical? The answer
is straightforward. As already indicated in (3) above, the version of Peano Arith-
metic which is categorical is a second-order theory – i.e. a theory which quantifies
not just over numbers but over numerical properties, and has a second-order in-
duction principle. Going second-order makes all the difference in arithmetic, and
in other theories too like the theory of the real numbers. But why? To understand
what is going on here, you need to understand something about the contrast be-
tween first-order theories and second-order ones. (So see our previous chapter,
and follow up the readings if you didn’t do so before.)
(f) Still at Level 2, there are results about which theories are complete in the
sense of entailing either A or ¬A for each relevant sentence A, and how this
relates to being categorical at a particular size. And there is another related
notion of so-called model-completeness: but let’s not pause over that.
Instead, let’s mention just one more fascinating topic that you will encounter
early in your model theory explorations:
(11) As explained in the last footnote, we can take a standard first-order theory
of the natural numbers and use a compactness argument to show that it
has a non-standard model which has an element c in the domain distinct
from (and indeed greater than) zero or any of its successors. Similarly,
we can take a standard first-order of the real numbers and use another
compactness argument to show that it has a non-standard model with
an element r in the domain such that that 0 < |r| < 1/n for any natural
number n. So in this model, the non-standard real r is non-zero but smaller
than any rational number, so is infinitesimally small. And indeed our model
will have non-standard reals infinitesimally close to any standard real.
In this way, we can build up a model of non-standard analysis with
infinitesimals (where e.g. a differential really can be treated as a ratio of
51
 DRAFT– 31 DEC 2021
5 Model theory

infinitesimally small numbers – in just the sort of way that we all supposed
wasn’t respectable at all). Fascinating!

5.2 Recommendations for beginning first-order model theory

A preliminary point. When exploring model theory you will very quickly en-
counter talk of different infinite cardinalities, and also occasional references to
the Axiom of Choice. You need to be familiar enough with these basic set-
theoretic ideas (perhaps from the readings suggested back in Chapter 2).

Let’s begin with a more expansive and very helpful overview (though you
may not understand everything at this preliminary stage). For a bit more
detail about the initial agenda of model theory, it is hard to beat

1. Wilfrid Hodges, ‘Model theory’, in the The Stanford Encyclopaedia of

Philosophy at tinyurl.com/sepmodel.

Now, a number of the introductions to FOL that I noted in $3.4 have treat-
ments of the Level 1 basics; I’ll be recommending one in a moment, and will
return to some of the others in the next section on parallel reading. Going just
a little beyond, the very first volume in the prestigious and immensely useful
Oxford Logic Guides series is Jane Bridge’s short Beginning Model Theory: The
Completeness Theorem and Some Consequences (Clarendon Press, 1977). This
neatly takes us through some Level 1 and a few Level 2 topics. But the writing,
though very clear, is also rather terse in an old-school way; and the book – not
unusually for that publication date – looks like photo-reproduced typescript,
which is nowadays really off-putting to read. What, then, are the more recent
options?

2. I have already sung the praises of Derek Goldrei’s Propositional and

Predicate Calculus: A Model of Argument (Springer, 2005) for the ac-
cessibility of its treatment of FOL in the first five chapters. You should
now read Goldrei’s §§4.4 and 4.5 (which I previously said you could
skip), and then Chapter 6 ‘On some uses of compactness’.
In a little more detail, §4.4 introduces some axiom systems describing
various mathematical structures (partial orderings, groups, rings, etc.):
this section could be particularly useful to philosophers who haven’t re-
ally met the notions before. Then §4.5 introduces the notions of substruc-
tures and structure-preserving isomorphisms. After proving the com-
pactness theorem in §6.1 (as a corollary of his completeness proof), Gol-
drei proceeds to use it in §§6.2 and 6.3 to show various theories can’t
be finitely axiomatized, or can’t be nicely axiomatized at all. §6.4 intro-
duces the Löwenheim-Skolem theorems and some consequences, and the
following section introduces the notion of ‘diagrams’ and puts it to work.

52
 DRAFT– 31 DEC 2021
Recommendations for beginning first-order model theory

The final section, §6.6 considers issues about categoricity, completeness

and decidability.
All this is done with the same admirable clarity as marked out Gol-
drei’s earlier chapters.
Goldrei goes quite slowly and doesn’t get very far (it is Level 1 model
theory). To take a further step (up to Level 2), here are two suggestions.
Neither is quite ideal, but each has virtues. The first is

3. Marı́a Manzano, Model Theory, Oxford Logic Guides 37 (OUP, 1999).

This book aims to be an introduction at the kind of levels we are
currently concerned with. And standing back from the details, I do like
the way that Manzano structures her book. The sequencing of chapters
makes for a very natural path through her material, and the coverage
seems very appropriate for a book at Levels 1 and 2. After chapters
about structures (and mappings between them) and about first-order
languages, she proves the completeness theorem again, and then has a
sequence of chapters on various core model-theoretic notions and proofs.
Overall, Manzano’s book should all be tolerably accessibly (especially
if not your very first encounter with model theoretic ideas). However,
it seems to me that the discussions at some points would have benefit-
ted from rather more informal commentary, motivating various choices,
and sometimes the symbolism is unncessarily heavy-handed. But over-
all, Manzano’s text could work well enough as a follow-up to Goldrei.
For more details, see tinyurl.com/manzanobook.

Another option is to look at the first two-thirds of the following book,

which is explicitly aimed at undergraduate mathematicians, and is at ap-
proximately the same level of difficulty as Manzano:

4. Jonathan Kirby, An Invitation to Model Theory (CUP, 2019).

As the blurb says, “The highlights of basic model theory are illustrated
through examples from specific structures familiar from undergraduate
mathematics.” Now, one thing that usually isn’t already familiar to most
undergraduate mathematicians is any serious logic: so Kirby’s book is an
introduction to model theory that doesn’t presuppose a previous FOL
course. So he has to start with some rather speedy explanations in Part
I about first-order languages and interpretations in structures.
The book is then nicely arranged. Part II of the book is on ‘Theo-
ries and compactness’, Part III on ‘Changing models’, and Part IV on
‘Characterizing definable sets’. (I’d say that some of the further Parts
of the book, though, go a bit beyond what you need at this stage.)
Kirby writes admirably clearly; but his book goes pretty briskly and
would have been improved – at least for self-study – if he had slowed
down for some more classroom asides. So I can imagine that some readers

53
 DRAFT– 31 DEC 2021
5 Model theory

would struggle with parts of this short book if were treated as a sole
introduction to model theory. However, again if you have read Goldrei,
it should be very helpful as an alternative or complement to Manzano’s
book. For a little more about it, see tinyurl.com/kirbybooknote.

We noted that first-order theories behave differently from second-order the-

ories where we have quantifiers running over all the properties and functions
defined over a domain, as well as over the objects in the domain. For more
on this see the readings on second-order logic suggested in §4.3.

5.3 Some parallel and slightly more advanced reading

I mentioned before that some other introductory texts on FOL apart from Gol-
drei’s have sections or chapters beginning model theory.
Some topics are briefly touched on in §2.6 of Herbert Enderton’s A Mathemat-
ical Introduction to Logic (Academic Press 1972, 2002), and there is discussion
of non-standard analysis in his §2.8: but this is perhaps too little done too fast.
So I think the following suits our needs here better:

5. Dirk van Dalen Logic and Structure (Springer, 1980; 5th edition 2012),
Chapter 3.
This covers rather more model-theoretic material than Enderton and
in greater detail. You could read §3.1 for revision on the completeness
theorem, then tackle §3.2 on compactness, the Löwenheim-Skolem theo-
rems and their implications, before moving on to the action-packed §3.3
which covers more model theory including non-standard analysis again,
and indeed touches on some slightly more advanced topics.

And there is also a nice chapter in another often-recommended text:

6. Richard E. Hodel, An Introduction to Mathematical Logic* (originally

published 1995; Dover reprint 2013).
In Chapter 6, ‘Mathematics and logic’, §6.1 discusses first-order the-
ories, §6.2 treats compactness and the Löwenheim-Skolem theorem, and
§6.3 is on decidable theories. Very clearly done.

For rather more detail, here is a recent book with an enticing title:

7. Roman Kossak, Model Theory for Beginners: 15 Lectures* (College Pub-

lications 2021).
As the title indicates, the fifteen chapters of this short book – just
138 pages – have their origin in introductory lectures, given to graduate
students in CUNY.
After initial chapters on structures and (first-order) languages, Chap-
ters 3 and 4 are on definability and on simple results such as that or-
dering is not definable in the language for the integers with addition,
(Z, +). Chapter 5 introduces the notion of ‘types’, and e.g. gives the
54
 DRAFT– 31 DEC 2021
Some parallel and slightly more advanced reading

back-and-forth proof conventionally attributed to Cantor that countable

dense linearly ordered sets without endpoints are always isomorphic to
the rationals in their natural order, (Q, <). Chapter 6 defines relations
between structures like elementary equivalence and elementary exten-
sion, and establishes the so-called Tarski-Vaught test. Then Chapter 7
proves the compactness theorem, with Chapter 8 using compactness to
establish some results about non-standard models of arithmetic and set
theory.
So there is a somewhat different arrangement of initial topics here,
compared with books whose first steps in model theory are applications
of compactness. The early chapters are indeed nicely done. However,
I don’t think that Kossak’s Chapter 8 will be found an outstandingly
clear first introduction to applications of compactness – it will probably
be best read after e.g. Goldrei’s nice final chapter in his logic text.
Chapter 9 is on categoricity – in particular, countable categoricity.
(Very sensibly, Kossak wants to keep his use of set theory in this book to
a minimum; but he does have a section here looking at κ-categoricity for
larger cardinals κ.) And now the book speeds up, and starts to require
rather more of its reader, and eventually touches on what I think of as
Level 3 topics. Real beginners in model theory without much mathemat-
ical background might begin to struggle after the half-way mark in the
book. But this is very nice addition to the introductory literature.

Thanks to the efforts of the respective authors to write very accessibly, the
suggested main path into the foothills of model theory (from Chiswell & Hodges
→ Leary & Kristiansen → Goldrei → Manzano/Kirby/Kossack) is not at all a
hard road to follow.
Now, we can climb up to the same foothills by routes involving rather tougher
scrambles, taking in some additional side-paths and new views along the way.
Here, then, is a suggestion for the more mathematical reader:

8. Shawn Hedman, A First Course in Logic (OUP, 2004).

This covers a surprising amount of model theory. Ch. 2 tells you about
structures and about relations between structures. Ch. 4 starts with a
nice presentation of a Henkin completeness proof, and then pauses (as
Goldrei does) to fill in some background about infinite cardinals etc.,
before going on to prove the Löwenheim-Skolem theorems and compact-
ness theorems. Then the rest of Ch. 4 and the next chapter covers more
introductory model theory, though already touching on a number of top-
ics beyond the scope of Mansion’s book (we are already at Level 2.5,
perhaps!). Hedman so far could therefore serve as a rather tougher al-
ternative to e.g. Manzano’s treatment.
Then Ch. 6 takes the story on a lot further, beyond what I’d regard
as elementary model theory. For more, see tinyurl.com/hedmanbook.

Last but certainly not least, philosophers (but not just philosophers) will cer-

55
 DRAFT– 31 DEC 2021
5 Model theory

tainly want to tackle (parts of) the following book, which strikes me as a very
impressive achievement:

9. Tim Button and Sean Walsh, Philosophy and Model Theory* (OUP,
2018).
This book both explains technical results in model theory, and also
explores the appeals to model theory in various branches of philosophy,
particularly philosophy of mathematics, but in metaphysics more gener-
ally, the philosophy of science, philosophical logic and more. So that’s a
very scattered literature that is being expounded, brought together, ex-
amined, inter-related, criticized and discussed. Button and Walsh don’t
pretend to be giving the last word on the many and varied topics they
discuss; but they are offering us a very generous helping of first words
and second thoughts. It’s a large book because it is to a significant extent
self-contained: model-theoretic notions get defined as needed, and many
of the more significant results are proved.
The philosophical discussion is done with vigour and a very engaging
style. And the expositions of the needed technical results are usually
exemplary (the authors have a good policy of shuffling some extended
proofs into chapter appendices). They also say more about second-order
logic and second-order theories than is usual.
But I do rather suspect that, despite their best efforts, an amount of
the material is more difficult than the authors fully realize: we soon get
to tangle with some Level 3 model theory, and quite a lot of other tech-
nical background is presupposed. The breadth and depth of knowledge
brought to the enterprise is remarkable: but it does make of a bumpy
ride even for those who already know quite a lot. Philosophical readers
of this Guide will probably find the book challenging, then, but should
find at least the earlier parts fascinating. And indeed, with judicious
skimming/skipping – the signposting in the book is excellent – mathe-
maticians with an interest in some foundational questions should find a
great deal of interest here too.

And that might already be about as far as many philosophers may want or need
to go in this area. Many mathematicians, however, will want go further into
model theory; so we pick up the story again in §12.2.

5.4 A little history

The last book we mentioned has an historical appendix contributed by a now
familiar author:

10. Wilfrid Hodges, ‘A short history of model theory’, in Button and Walsh,
pp. 439–476.

56
 DRAFT– 31 DEC 2021
A little history

Read the first six or so sections. Later sections refer to model theoretic topics a
level up from our current more elementary concerns, so won’t be very accessible
at this stage.
For another piece that focuses on topics from the beginning of model theory,
you could perhaps try R. L. Vaught’s ‘Model theory before 1945’ in L. Henkin et
al, eds, Proceedings of the Tarski Symposium (American Mathematical Society,
1974), pp. 153–172. You’ll probably have to skim parts, but it will also give you
some idea of the early developments. But here’s something which is much more
fun to read. Alfred Tarski was one of the key figures in that early history. And
there is a very enjoyable and well-written biography, which vividly portrays the
man, and gives a wonderful sense of his intellectual world, but also contains
accessible interludes on his logical work:

11. Anita Burdman Feferman and Solomon Feferman, Alfred Tarski, Life
and Logic (CUP, 2004).

57
 DRAFT– 31 DEC 2021

6 Arithmetic, computability, and

incompleteness

The standard mathematical logic curriculum, as well as looking at some elemen-

tary results about formalized theories and their models in general, investigates
two particular instances of non-trivial, rigorously formalized, axiomatic systems.
First, there’s arithmetic (a paradigm theory about finite whatnots); and then
there is set theory (a paradigm theory about infinite whatnots). We consider set
theory in the next chapter. This chapter is about arithmetic and related matters.
More specifically, we consider three inter-connected topics:
1. The elementary theory of numerical computable functions.
2. Formal theories of arithmetic and how they represent computable func-
tions.
3. Gödel’s epoch-making proof of the incompleteness of any sufficiently nice
formal theory that can ‘do’ enough arithmetical computations.
Before turning to some short topic-by-topic overviews, though, it is well worth
pausing for a quick general point about why the idea of computability is of such
very central concern to formal logic.

6.1 Logic and computability

(a) The aim of regimenting informal arguments and informal theories into for-
malized versions is to eliminate ambiguities and to make everything entirely
determinate and transparently clear (even if it doesn’t always seem that way to
beginners!). So, for example, we want it to be entirely clear what is and what
isn’t a formal sentence of a given theory, what is and what isn’t an axiom of
the theory, and what is and what isn’t a formal proof in the theory. We want to
be able to settle these things in a way which leaves absolutely no room left for
doubt or dispute.
(b) As a step towards sharpening this thought, let’s say as an initial rough
characterization:
A property P is effectively decidable if and only if there is an algo-
rithm (a finite set of instructions for a deterministic computation)
58
 DRAFT– 31 DEC 2021
Logic and computability

for settling in a finite number of steps, whether a relevant object has

property P.
Relatedly, the answer to a question Q is effectively decidable if
and only if there is an algorithm which gives the answer, again by a
deterministic computation, in a finite number of steps.

To put it only slightly different words, a property P is effectively decidable just

when there’s a step-by-step mechanical routine for settling whether an object of
the relevant kind has property P, such that a suitably programmed deterministic
computer could in principle implement the routine (idealizing away from practi-
cal constraints of time, etc.). Similarly, the answer to a question Q is effectively
decidable just when a suitably programmed computer could deliver the answer
(in principle, in a finite time).
Two initial examples from propositional logic: we can effectively decide what
is the main connective of a sentence (by bracket counting), and the property of
being a tautology is effectively decidable.
And the point we made at the outset in (a) now comes to this: we will want it
to be effectively decidable e.g. whether a given string of symbols has the property
of being a well-formed formula of a certain formal language, whether a formula
is an axiom of a given formal theory, and whether an array of formulas is a
correctly formed proof of the theory. In other words, we will want to set up a
formal deductive theory so that a computer could, in principle, mindlessly check
e.g. the credentials of a purported proof by deciding whether each step of the
proof is indeed in accordance with the official rules of the theory.
(c) NB: It is one thing to be able to effectively decide whether a purported proof
of P really is a proof in a given formal theory T . It is another thing entirely to
be able to decide in advance whether P actually has a proof in T .
You’ll soon enough find out that, e.g., in a properly set up formal theory of
arithmetic T we can effectively check whether a supposed proof of P indeed
conforms to the rules of the game. But once we are dealing with an even mildly
interesting T , there will be no way of deciding in advance whether a T -proof of
P exists. Such a theory T is said to be undecidable.
It is of course nice when a theory is decidable, i.e. when a computer can tell
us whether a given proposition does or doesn’t follow from the theory. But few
interesting theories are decidable in this sense: so mathematicians aren’t going
to be put out of business!
(d) Now, in our initial rough definition of the notion of effective decidability,
we invoked the idea of what an idealized computer could (in principle) do by
implementing some algorithm. This idea surely needs further elaboration.

1. As a preliminary step, we can narrow our focus and just consider the
decidability of arithmetical properties.
Why? Because we can always represent facts about finite whatnots like
formulas and proofs by using numerical codings. We can then trade in
questions about formulas or proofs for questions about their code numbers.

59
 DRAFT– 31 DEC 2021
6 Arithmetic, computability, and incompleteness

2. And as a second step, we can also trade in questions about the effective
decidability of arithmetical properties for questions about the algorithmic
computability of numerical functions.
Why? Because for any numerical property P we can define a correspond-
ing numerical function (its so-called ‘characteristic function’) cP such that
if n has the property P , cP (n) = 1 and if n doesn’t the have property P ,
cP (n) = 0. Think of ‘1’ as coding for truth, and ‘0’ for falsehood. Then
the question (i) ‘can we effectively decide whether a number has the prop-
erty P ?’ becomes the question (ii) ‘is the numerical function cP effectively
computable by an algorithm?’.

So, by those two steps, we do quickly move from e.g. the question whether it
is effectively decidable whether a string of symbols is a wff to a corresponding
question about whether a certain numerical function is computable.

6.2 Computable functions: an overview

(a) For convenience, we will now use ‘S’ for the function that maps a number
to its successor (where we previously used a prime). Consider, then, the following
pairs of equations:

x+0=x
x + Sy = S(x + y)
x×0=0
x × Sy = (x × y) + x
x0 = S0
xSy = (xy × x)

In some notation or other, these pairs of equations should be very familiar:

they in turn define addition, multiplication and exponentiation for the natural
numbers. It’s useful to spell out the point.
Take the initial pair of equations. The first of them fixes the result of adding
zero to a given number. The second fixes the result of adding the successor of
y in terms of the result of adding y. Hence applying and re-applying the two
equations, they together tell us how to add 0, S0, SS0, SSS0, . . ., i.e. they tell us
how to add any natural number to a given number x. Similarly, the first of the
equations for multiplication fixes the result of multiplying by zero. The second
equation fixes the result of multiplying by Sy in terms of the result of multiplying
by y and doing an addition. Hence the two pairs of equations together tell us
how to multiply a given number x by any of 0, S0, SS0, SSS0, . . .. Similarly of
course for the pair of equations for exponentiation.
And now note that the six equations taken together not only define expo-
nentiation, but they do so by giving us an algorithm for computing xy for any
natural numbers x, y – they tell us how to compute xy by doing repeated mul-
tiplications, which we in turn compute by doing repeated additions, which we
60
 DRAFT– 31 DEC 2021
Computable functions: an overview

compute by repeated applications of the successor function. That is to say, the

chain of equations amounts to a set of instructions for a deterministic step-by-
step computation which will output the value of xy in a finite number of steps.
Hence, exponentiation is indeed an effectively computable function.
(b) In each of our pairs of equations, the second one fixes the value of the
defined function for argument Sy by invoking the value of the same function for
argument y. A procedure where we evaluate a function for one input by calling
the same function for some smaller input(s) is standardly termed ‘recursive’ –
and the particularly simple type of procedure we’ve illustrated three times is
called, more precisely, primitive recursion. Now – arm-waving more than a bit! –
consider any function which can be defined by a chain of equations similar to the
chain of equations giving us a definition of exponentiation. Suppose that, starting
from trivial functions like the successor function, we can build up the function’s
definition by using primitive recursions and plugging one function we already
know about into another. Such a function is said to be primitive recursive.
And generalizing from the case of exponentiation, we have the following ob-
servation:

Any primitive recursive function is similarly effectively computable

(c) So far, so good, However, it is easy to show that

Not all effectively computable functions are primitive recursive.

A neat abstract argument proves the point.1 But this raises an obvious question:
what further ways of defining functions – in addition to primitive recursion –
also give us effectively computable functions?
Here’s a pointer. The definition of (say) xy by primitive recursion in effect
tells us to start from x0 , then loop round applying the recursion equation to
compute x1 , then x2 , then x3 , . . . , keeping going until we reach xy . In all, we
have to loop around y times. In some standard computer languages, implement-
ing this procedure involves using a ‘for’ loop (which tells us to iterate some
procedure, counting as we go, and to do this for cycles numbered 1 to y). In
this case, the number of iterations is given in advance as we enter the loop.
But of course, standard computer languages also have programming structures
which implement unbounded searches – they allow open-ended ‘do until’ loops
(or equivalently, ‘do while’ loops). In other words, they allow some process to
be iterated until a given condition is satisfied, where no prior limit is put on the
number of iterations to be executed.
This suggests that one way of expanding the class of computable functions
beyond the primitive recursive functions will be to allow computations employing
open-ended searches. So let’s suppose we do this (there’s a standard device for

1 Roughly, we can effectively list off the primitive recursive functions by listing their recipes;
so we have an algorithm which gives us fn , the n-th such function. Then define the function
d by putting d(n) = fn (n) + 1. Evidently, d differs from any fn for the value n, so isn’t one
of the primitive recursive functions. But it is computable.

61
 DRAFT– 31 DEC 2021
6 Arithmetic, computability, and incompleteness

this, but let’s not worry about the details now). Functions – more precisely,
total functions that deliver an output for any numerical input – which can be
computed by a chain of applications of primitive recursion and/or open-ended
searches are called (simply) recursive.
(d) Predictably enough, the next question is: have we now got all the effectively
computable functions?
The claim that the recursive functions are indeed just the intuitively com-
putable total functions is Church’s Thesis, and is very widely believed to be
true (or at least, it is taken to be an entirely satisfactory working hypothesis).
Why? For a start, there are quasi-empirical reasons: no one has found a function
which is incontrovertibly computable by a finite-step deterministic algorithmic
procedure but which isn’t recursive. But there are also much more principled
reasons for accepting the Thesis.
Consider, for example, Alan Turing’s approach to the notion of effective com-
putation. He famously aimed to analyse the idea of a step-by-step computation
procedure down to its very basics, which led him to the concept of computation
by a Turing machine (a minimalist computer). And what we can call Turing’s
Thesis is the claim that the effectively computable (total) functions are just the
functions which are computable by some suitably programmed Turing machine.
So do we now have two rival claims, Church’s and Turing’s, about the class of
computable functions? Not at all! For it turns out to be quite easy to prove the
technical result that a function is recursive if and only if is Turing computable.
And so it goes: every other attempt to give an exact characterization of the class
of effectively computable functions turns out to locate just the same class of
functions. That’s remarkable, and this is a key theme you will want to explore
in a first encounter with the theory of computable functions.
(e) It is fun to find out more about Turing machines, and even to learn to write
a few elementary programs (in effect, it is learning to write in a ‘machine code’).
And there is a beautiful early result that you will soon encounter:

There is no mechanical decision procedure which can determine whether

Turing machine number e, fed a given input n, will ever halt its com-
putation (so there is no general decision procedure which can tell
whether Turing machine e in fact computes a total function).

How do we show that? Why does it matter? I leave it to you to read up on the
‘undecidability of the halting problem’, and its many weighty implications.

6.3 Formal arithmetic: an overview

(a) The elementary theory of computation really is a lovely area, where acces-
sible Big Results come thick and fast! But now we must turn to consider formal
theories of arithmetic.
We standardly focus on First-order Peano Arithmetic, PA. It will be no sur-
prise to hear that this theory has a first-order language and logic! It has a built-in
62
 DRAFT– 31 DEC 2021
Formal arithmetic: an overview

constant 0 to denote zero, has symbols for the successor, addition and multipli-
cation functions (to keep things looking nice, we still use a prefix S, and infix +
and ×), and its quantifiers run over the natural numbers. Note, we can form the
sequence of numerals 0, S0, SS0, SSS0, . . . (we will use n to abbreviate the result
of writing n occurrences of S before 0, so n denotes n).
PA has the following three pairs of axioms governing the three built-in func-
tions:

∀x 0 6= Sx
∀x∀y(Sx = Sy → x = y)
∀x x + 0 = x
∀x∀y x + Sy = S(x + y)
∀x x × 0 = 0
∀x∀y x × Sy = (x × y) + x

The first pair of axioms specifies that distinct numbers have distinct successors,
and that the sequence of successors never circles round and ends up with zero
again: so the numerals, as we want, must denote a sequence of distinct numbers,
zero and all its eventual successors. The other two pairs of axioms formalize the
equations defining addition and multiplication which we have met before.
And then, crucially, there is also an arithmetical induction principle. As noted
in §4.2, in a first-order framework we can stipulate that

Any wff of the form ({A(0) ∧ ∀x(A(x) → A(Sx))} → ∀xA(x)) is an

axiom.

Or obviously equivalently, we can formulate the same idea as an inference rule:

If A(x) is a formula with x free, then from A(0) and ∀x(A(x) → A(Sx))
we can infer ∀xA(x).

You need to get at least some elementary familiarity with the workings of the
resulting theory.
(b) But why concentrate on first-order PA? We’ve emphasized in §4.2 that our
informal induction principle is most naturally construed as involving a second-
order generalization – for any arithmetical property P, if zero has P , and if a
number which has P always passes it on to its successor, then every number has
P . And when Richard Dedekind (1888) and Giuseppe Peano (1889) gave their
axioms for what we can call Dedekind-Peano arithmetic, they correspondingly
gave a second-order formulation for their versions of the induction principle.
Put it this way: Dedekind and Peano’s principle quantifies over all properties of
numbers, while in first-order PA our induction principle rather strikingly only
deals with those properties of numbers which can be expressed by open formulas
of its restricted language. Why go for the weaker first-order principle?
Well, we have already addressed this in Chapter 4: first-order logic is much
better behaved than second-order logic. And some would say that second-order
63
 DRAFT– 31 DEC 2021
6 Arithmetic, computability, and incompleteness

logic is really just a bit of set theory in disguise. So, the argument goes, if we
want a theory of pure arithmetic, one whose logic can be formalized, we should
stick to a first-order formulation just quantifying over numbers. Then something
like PA’s induction rule (or the suite of axioms of the form we described) is the
best we can do.
But still, even if we have decided to stick to a first-order theory, why re-
strict ourselves to the impoverished resources of PA, with only three function-
expressions built into its language? Why not have an expression for e.g. the
exponential functions as well, and add to the theory the two defining axioms
for that function? Indeed, why not add expressions for other recursive functions
too, and then also include appropriate axioms for them in our formal theory?
Good question. The answer is to be found in a neat technical observation first
made by Gödel. Once we have successor, addition and multiplication available,
plus the usual first-order logical apparatus, we can in fact already express any
other computable (i.e. recursive) function. To take the simplest sort of case, sup-
pose f is a one-place recursive function: then there will be a two-place expression
of PA’s language which we can abbreviate F(x, y) such that F(m, n) is true if and
only if f (m) = n. Moreover, when f (m) = n, PA can prove F(m, n), and when
f (m) 6= n, PA can prove ¬F(m, n). In this way, PA as it were already ‘knows’
about all the recursive functions and can compute their values. Similarly, PA
can already express any algorithmically decidable relation.
So PA is expressively a lot richer than you might initially suppose. And indeed,
it turns out that even a induction-free subsystem of PA known as Robinson
Arithmetic (often called simply Q) can express the recursive functions.
And this key fact puts you in a position to link up your investigations of PA
with what you know about computability. For example, we quickly get a fairly
straightforward proof that there is no mechanical procedure that a computer
could implement which can decide whether a given arithmetic sentence is a
theorem of PA (or even a theorem of Q).
(c) On the other hand, despite its richness, PA is a first-order theory with
infinite models, so – applying results from elementary model theory (see the
previous chapter) – this first order arithmetic will have non-standard models,
i.e. will have models whose domains contain more than a zero and its successors.
It is worth knowing at an early stage just something about what some of these
non-standard models can look like. And you will also want to further investigate
the contrast with second-order versions of arithmetic which are categorical (i.e.
don’t have non-standard models).

6.4 Towards Gödelian incompleteness

(i) Now for our third related topic: Gödel’s incompleteness theorems.
First-order PA, we said, turns out to be a very rich theory. Is it rich enough
to settle every question that can be raised in its language? No! In 1931, Kurt
Gödel proved that a theory like PA must be negation incomplete – meaning that

64
 DRAFT– 31 DEC 2021
Towards Gödelian incompleteness

we can form a sentence G in its language such that PA proves neither G nor ¬G.
How does he do the trick?
(ii) It’s fun to give an outline sketch, which I hope will intrigue you enough to
leave you wanting to find out more! So:
G1. Gödel introduces a Gödel-numbering scheme for a formal theory like PA,
which is a simple way of coding expressions of PA – and also sequences
of expressions of PA – using natural numbers. The code number for an
expression (or a sequence of expressions) is its unique Gödel number.
G2. We can then define relations like Prf , where Prf (m, n) holds if and only if
m is the Gödel number of a PA-proof of the sentence with code number n.
So Prf is a numerical relation which, so to speak, ‘arithmetizes’ the syn-
tactic relation between a sequence of expressions (proof) and a particular
sentence (conclusion).
G3. There’s a procedure for computing, given numbers m and n, whether
Prf (m, n) holds. Informally, we just decode m (that’s an algorithmic pro-
cedure). Now check whether the resulting sequence of expressions – if there
is one – is a well-constructed PA-proof according to the rules of the game
(proof-checking is another algorithmic procedure). If that sequence is a
proof, check whether it ends with a sentence with the code number n
(that’s another algorithmic procedure).
G4. Since PA can express any algorithmically decidable relation, there will in
particular be a formal expression in the language of PA which we can
abbreviate Prf(x, y) which expresses the effectively decidable relation Prf .
This means that Prf(m, n) is true if and only if m codes for a PA proof of
the sentence with Gödel number n.
G5. Now define Prov(y) to be the expression ∃xPrf(x, y). Then Prov(n), i.e.
∃xPrf(x, n), is true if and only if some number Gödel-numbers a PA-proof
of the wff with Gödel-number n, i.e. is true just if the wff with code num-
ber n is a theorem of PA. Therefore Prov is naturally called a provability
predicate.
G6. Next, with only a little bit of cunning, we construct a Gödel sentence G
in the language of PA with the following property: G is true if and only if
¬Prov(g) is true, where g is the numeral for g, the code number of G.
Don’t worry for the moment about how we do this construction (it in-
volves a so-called ‘diagonalization’ trick which is surprisingly easy). Just
note that G is true on interpretation if and only if the sentence with Gödel
number g is not a PA-theorem, i.e. if and only if G is not a PA-theorem.
In short, G is true if and only if it isn’t a PA-theorem. So, rather stretch-
ing a point, it is rather as if G ‘says’ I am unprovable in PA.
G7. Now, suppose G were provable in PA. Then, since G is true if and only if it
isn’t a PA-theorem, G would be false. So PA would have a false theorem.
65
 DRAFT– 31 DEC 2021
6 Arithmetic, computability, and incompleteness

Hence assuming PA is sound and only has true theorems, then it can’t
prove G. Hence, since it is not provable, G is indeed true. Which means
that ¬G is false. Hence, still assuming PA is sound, it can’t prove ¬G either.
So, in sum, assuming PA is sound, it can’t prove either of G or ¬G. As
announced, PA is negation incomplete.
Wonderful!
(iii) Now the argument generalizes to other nicely axiomatized sound theories
T which can express enough arithmetical truths. We can use the same sort of
cunning construction to find a true GT such that T can prove neither GT nor
¬GT . Let’s be really clear: this doesn’t, repeat doesn’t, say that GT is ‘absolutely
unprovable’, whatever that could mean. It just says that GT and its negation
are unprovable-in-T.
Ok, you might well ask, why don’t we simply ‘repair the gap’ in T by adding
the true sentence GT as a new axiom? Well, consider the theory U = T + GT (to
use an obvious notation). Then (i) U is still sound, since the old T -axioms are
true and the added new axiom is true. (ii) U is still a nicely axiomatized formal
theory given that T is. (iii) U can still express enough arithmetic. So we can find
a sentence GU such that U can prove neither GU nor ¬GU .
And so it goes. Keep throwing more and more additional true axioms at T and
our theory will remain negation-incomplete (unless it stops counting as nicely
axiomatized). So here’s the key take-away message: any sound nicely axiomatized
theory T which can express enough arithmetic will not just be incomplete but
in a good sense T will be incompletable.
(iv) Now, we haven’t quite arrived at what’s usually called the First Incom-
pleteness Theorem. For that, we need an extra step Gödel took, which enables
us to drop the semantic assumption that we are dealing with a sound theory T
for a weaker consistency requirement. But I’ll leave you to explore the (not very
difficult) details, and also to find out about the Second Theorem.
It really is time to start reading!

6.5 Main recommendations on arithmetic, etc.

I hope those arm-waving overviews were enough to pique your interest. But if
you want a more expansive overview of the territory, then you can very usefully
look at one of
1. Robert Rogers, Mathematical Logic and Formalized Theories (North-
Holland, 1971), Chapter VIII, ‘Incompleteness, Undecidability’ (still quite
discursive, very clear).
2. Robert S. Wolf, A Tour Through Mathematical Logic (Mathematical
Association of America, 2005), Chapter 3, ‘Recursion theory and com-
putability’; and Chapter 4, ‘Gödel’s incompleteness theorems’ (more de-
tailed, requiring more of the reader, though some students do really like
this book).
66
 DRAFT– 31 DEC 2021
Main recommendations on arithmetic, etc.

But now turning to textbooks, how to approach the area? Gödel’s 1931 proof
of his incompleteness theorem actually uses only facts about the primitive recur-
sive functions. As we noted, these functions are only a subclass of the effectively
computable numerical functions. A more general treatment of computable func-
tions was developed a few years later (by Gödel, Turing and others), and this in
turn throws more light on the incompleteness phenomenon. So there’s a choice
to be made. Do you look at things in roughly the historical order, first introduc-
ing just the primitive recursive functions, explaining how they get represented
in theories of formal arithmetic, and then learning how to prove initial versions
of Gödel’s incompleteness theorem – and only then move on to deal with the
general theory of computable functions? Or do you explore the general theory
of computation first, only turning to the incompleteness theorems later?
My own Gödel books take the first route. But I also recommend alternatives
taking the second route. First, then, there is
3. Peter Smith, Gödel Without (Too Many) Tears* (Logic Matters, 2020):
freely downloadable from logicmatters.net/igt.
This is a very short book – just 130 pages – which, after some general
introductory chapters, and a little about formal arithmetic, explains the
idea of primitive recursive functions, explains the arithmetization of syn-
tax, and then proves Gödel’s First Theorem pretty much as Gödel did,
with a minimum of fuss. There follow a few chapters on closely related
matters and on the Second Theorem.
GWT is, I hope, very clear and accessible, and it perhaps gives all you need
for a first foray into this area if you don’t want (yet) to tangle with the general
theory of computation. However, you might well prefer to jump straight into one
of the following:

4. Peter Smith, An Introduction to Gödel’s Theorems* (2nd edition CUP,

2013: also now downloadable from logicmatters.net/igt).
Three times the length of GWT and ranging more widely, this starts
by informally exploring various ideas such as effective computability,
and then it proves two correspondingly informal versions of the first
incompleteness theorem. The next part of the book gets down to work
talking about formal arithmetics, developing some of the theory of prim-
itive recursive functions, and explaining the ‘arithmetization of syntax’.
Then it establishes more formal versions of Gödel’s first incompleteness
theorem and goes on discuss the second theorem, all in more detail than
GWT.
The last part of the book then widens out the discussion to explore the
idea of recursive functions more generally, discussing Turing machines
and the Church-Turing thesis, and giving further proofs of incomplete-
ness (e.g. deriving it from the ‘recursive unsolvability’ of the halting
problem for Turing machines).

67
 DRAFT– 31 DEC 2021
6 Arithmetic, computability, and incompleteness

5. Richard Epstein and Walter Carnielli, Computability: Computable Func-

tions, Logic, and the Foundations of Mathematics (Wadsworth 2nd edn.
2000: Advanced Reasoning Forum 3rd edn. 2008).
An excellent introductory book on the standard basics, particularly
clearly and attractively done. Part I, on ‘Fundamentals’, covers some
background material, e.g. on the idea of countable sets (many readers
will be able to speed-read through these initial chapters). Part II, on
‘Computable functions’, comes at them two ways: first via Turing Ma-
chine computability, and then via primitive recursive and then partial
recursive functions, ending with a proof that the two approaches define
the same class of effectively computable functions. Part III, ‘Logic and
arithmetic’, turns to formal theories of arithmetic and the way that the
representable functions in a formal arithmetic like Robinson’s Q turn
out to be the recursive ones. Formal arithmetic is then shown to be
undecidable, and Gödelian incompleteness derived. The shorter Part IV
has a chapter on Church’s Thesis (with more discussion than is often
the case), and finally a chapter on constructive mathematics. There are
many interesting historical asides along the way.
Those two books should be very accessible to those without much math-
ematical background: but even more experienced mathematicians should
appreciate the careful introductory orientation which they provide. Then
next, taking us half-a-step up in mathematical sophistication, we arrive at
a quite delightful book:
6. George Boolos and Richard Jeffrey, Computability and Logic (CUP 3rd
edn. 1990).
A modern classic, wonderfully lucid and engaging, admired by gen-
erations of readers. Indeed, looking at it again in revising this Guide,
I couldn’t resist some re-reading! It starts with a exploration of Turing
machines, ‘abacus computable’ functions, and recursive functions (show-
ing that different definitions of computability end up characterizing the
same class of functions). And then it moves on discuss logic and formal
arithmetic (with interesting discussions ranging beyond what is covered
in my book or E&C).
There are in fact two later editions – heavily revised and considerably
expanded – with John Burgess as a third author. But I know that I
am not the only one to think that these later versions (good though
they are) do lose something of the original book’s famed elegance and
individuality and distinctive flavour. Still, whichever edition comes to
hand, do read it! – you will learn a great deal in an enjoyable way.

One comment: none of these books – including my longer one – gives a full proof
of Gödel’s Second Incompleteness Theorem. The guiding idea is easy enough,
but there is tedious work to be done in implementing it. If you really want more
details, see e.g. the book by Boolos or by Rautenberg mentioned in §10.4.

68
 DRAFT– 31 DEC 2021
Some parallel/additional reading

6.6 Some parallel/additional reading

I should start by mentioning a more elementary book which might well appeal
to some for its debunking of myths about the wider significance of Gödelian
incompleteness:
7. Torkel Franzén, Gödel’s Theorem: An Incomplete Guide to its Use and
Abuse (A. K. Peters, 2005).
John Dawson (who we’ll meet again below) writes “Among the many
expositions of Gödel’s incompleteness theorems written for non-specialists,
this book stands apart. With exceptional clarity, Franzén gives careful,
non-technical explanations both of what those theorems say and, more
importantly, what they do not. No other book aims, as his does, to ad-
dress in detail the misunderstandings and abuses of the incompleteness
theorems that are so rife in popular discussions of their significance. As
an antidote to the many spurious appeals to incompleteness in theologi-
cal, anti-mechanist and post-modernist debates, it is a valuable addition
to the literature.” Invaluable, in fact!
And next, here’s a group of three books at about the same level as those
mentioned in the previous section. First, the Open Logic Project now has a
good volume on our topics:
8. Jeremy Avigad and Richard Zach, Incompleteness and Computability:
An Open Introduction to Gödel’s Theorems*, tinyurl.com/icomp-open.
Chapters 1 to 5 are on computability and Gödel, covering a good deal
in just 120 very sparsely printed pages. Avigad and Zach are admirably
clear as far as they go – though inevitably, given the length, they have to
go pretty briskly. But this could be enough for those who want a short
first introduction. And others could well find this very useful revision
material, highlighting some basic main themes.
But really, you should take a slower tour through more of the sights by follow-
ing the recommendations in the previous section, or by reading the following
excellent book that could well have been an alternative main recommendation:
9. Herbert E. Enderton, Computability Theory: An Introduction to Recu-
sion Theory (Associated Press, 2011).
This is written with attractive zip and lightness of touch (this is a no-
tably more relaxed book than his earlier Logic). The first chapter is on
the informal Computability Concept. There are then chapters on general
recursive functions and on register machines (showing that the register-
computable functions are exactly the recursive ones), and a chapter on
recursive enumerability. Chapter 5 makes ‘Connections to logic’ (includ-
ing proving Tarski’s theorem on the undefinability of arithmetical truth
and a semantic incompleteness theorem). The final two chapters push on
to say something about ‘Degrees of unsolvability’ and ‘Polynomial-time
computability’. This is all very nicely and accessibly done.
69
 DRAFT– 31 DEC 2021
6 Arithmetic, computability, and incompleteness

This book, then, makes an excellent alternative to Epstein & Carnielli in partic-
ular: it is, however, a little more abstract and sophisticated, which why I have
on balance recommended E&C for many readers. The more mathematical might
well prefer Enderton. (By the way, staying with Enderton, I should mention that
Chapter 3 of his earlier A Mathematical Introduction to Logic (Academic Press
1972, 2002) gives a good brisk treatment of different strengths of formal theo-
ries of arithmetic, and then proves the incompleteness theorem first for a formal
arithmetic with exponentiation and then – after touching on other issues – shows
how to use the β-function trick to extend the theorem to apply to arithmetic
without exponentiation. Not the best place to start, but this chapter too could
be very useful revision material.)
Thirdly, I have already warmly recommended the following book for its cov-
erage of first-order logic:

10. Christopher Leary and Lars Kristiansen’s A Friendly Introduction to

Mathematical Logic*, tinyurl.com/friendlylogic.
Chapters 4 to 7 now give a very illuminating double treatment of mat-
ters related to incompleteness (you don’t have to have read the previous
chapters in this book to follow the later ones, other than noting the
arithmetical system N introduced in their §2.8). In headline terms that
you’ll only come fully to understand in retrospect:
a) L&K’s first approach doesn’t go overtly via computability. Instead
of showing that certain syntactic properties are primitive recursive
and showing that all primitive recursive properties can be ‘repre-
sented’ in theories like N (as I do in IGT ), L&K rely on more
directly showing that some key syntactic properties can be rep-
resented. This representation result then leads to, inter alia, the
incompleteness theorem.
b) L&K follow this, however, with a general discussion of computabil-
ity, and then use the introductory results they obtain to prove var-
ious further theorems, including incompleteness again.
This is all presented with the same admirable clarity as the first part of
the book on FOL.

There are, of course, many other more-or-less introductory treatments cover-

ing aspects of computability and/or incompleteness, and we will return to the
topic at a more advanced level in §12.3. For now, I will mention just three further,
and rather more individual, books.
First, of the relevant texts in American Mathematical Society’s ‘Student Math-
ematical Library’, by far the best is

11. A. Shen and N. K. Vereshchagin, Computable Functions, (AMA, 2003).

This is a lovely, elegant, little book, which can be recommended for
giving a differently-structured quick tour through some of the Big Ideas.
Well worth reading as a follow-up to a more conventional text.

70
 DRAFT– 31 DEC 2021
Some parallel/additional reading

Next we come to a stand-out book that you should certainly tackle at some
point (and though this starts from scratch, I rather suspect that many readers
will appreciate it more if they come to it after reading one or more of the main
recommendations in the previous section):

12. Raymond Smullyan, Gödel’s Incompleteness Theorems, Oxford Logic

Guides 19 (Clarendon Press, 1992).
This is delightully short – under 140 pages – proving some rather
beautiful, slightly abstract, versions of the incompleteness theorems. This
is a modern classic which anyone with a taste for mathematical elegance
will find very rewarding.

To introduce the third book, the first thing to say is that it presupposes very
little knowledge about sets, despite the title. If you are familiar with the idea
that the natural numbers can be identified with (implemented as) finite sets in a
standard way, and with a few other low-level ideas, then you can dive in without
further ado to

13. Melvin Fitting’s, Incompleteness in the Land of Sets* (College Publica-

tions, 2007).
This is a very engaging read, approaching the incompleteness theorem
and related results in an unusual but illuminating way. From the book’s
blurb: “Russell’s paradox arises when we consider those sets that do not
belong to themselves. The collection of such sets cannot constitute a set.
Step back a bit. Logical formulas define sets (in a standard model). For-
mulas, being mathematical objects, can be thought of as sets themselves
– mathematics reduces to set theory. Consider those formulas that do
not belong to the set they define. The collection of such formulas is not
definable by a formula, by the same argument that Russell used. This
quickly gives Tarski’s result on the undefinability of truth. Variations on
the same idea yield the famous results of Gödel, Church, Rosser, and
Post.
This book gives a full presentation of the basic incompleteness and
undecidability theorems of mathematical logic in the framework of set
theory. Corresponding results for arithmetic follow easily, and are also
given. Gödel numbering is generally avoided, except when an explicit
connection is made between set theory and arithmetic. The book assumes
little technical background from the reader. One needs mathematical
ability [and] a general familiarity with formal logic ...”

And, finally, if only because I’ve been asked about it such a large number of
times, I suppose I should end by also mentioning the (in)famous

14. Douglas Hofstadter, Gödel, Escher, Bach* (Penguin, 1979).

When students enquire about this, I helpfully say that it is the sort of
book that you will probably really like if you like this kind of book, and
you won’t if you don’t. It is, to say the very least, quirky, idiosyncratic
71
 DRAFT– 31 DEC 2021
6 Arithmetic, computability, and incompleteness

and entirely distinctive. However, as I far as I recall, the parts of the

book which touch on techie logical things are in fact pretty reliable and
won’t lead you astray.

Which is a great deal more than can be said about many popularizing treatments
of Gödel’s theorems!

6.7 A little history

If you haven’t already done so, do read

15. Richard Epstein’s brisk and very helpful 28 page ‘Computability and
undecidability – a timeline’ which is printed at the very end of Epstein
& Carnielli, listed in §6.5.

This will really give you the headline news you initially need. It is then well
worth reading

16. Robin Gandy, ‘The confluence of ideas in 1936’ in R. Herken, ed., The
Universal Turing Machine: A Half-century Survey (OUP 1988). This
seeks to explain why so many of the absolutely key notions all got formed
in the mid-thirties.

And then you might enjoy

17. John Dawson, Logical Dilemmas: The Life and Work of Kurt Gödel
(A. K. Peters, 1997).

Not, perhaps, as lively as the Fefermans’ biography of Tarski which I mentioned

in §5.4 – but then Gödel was such a very different man. Fascinating, though!
(As far as getting any logical insights goes, you can ignore the third-rate book
by Stephen Budiansky Journey to the Edge of Reason: The Life of Kurt Gödel,
OUP 2021).

72
 DRAFT– 31 DEC 2021

7 Set theory, less naively

In Chapter 2, we touched on some elementary concepts and constructions in-

volving sets. We now go further into set theory, though still not beyond the
beginnings that any logician really ought to know about. In §12.4 of the Guide
we will return to cover more advanced topics like ‘large cardinals’, proofs of
the consistency and independence of the Continuum Hypothesis, and a lot more
besides: but here is this chapter we concentrate on some core basics.

7.1 Elements of set theory: an overview

You won’t need to have done much mathematics for there to be little news for
you in the arm-waving remarks in this section: as always with these overviews,
feel very free to skim and skip.
(a) If you have not already done so, you now want to get a really firm grip
on the key facts about the ‘algebra of sets’ (concerning unions, intersections,
complements and how they interact). You also need to know, inter alia, the basics
about powersets, about encoding pairs and other finite tuples using unordered
sets, and about Cartesian products, the extensional treatment of relations and
functions, the idea of equivalence classes, and how to treat infinite sequences as
sets (see Chapter 2).
(b) Moving on, one fundamental early role for set theory was “putting the the-
ory of real numbers, and classical analysis more generally, on a firm foundation”.
But what does this involve?
It only takes a finite amount of data to fully specify a particular natural
number. Similarly for integers and rational numbers. But not so for real numbers.
As is very familiar, a real can rendered e.g. by an infinite sequence of ever-closer
rational approximations, but the sequence need never terminate or repeat. Set
theory gives us a framework for reasoning about such non-finite data. How?
Assume, for the moment, that we already have the rational numbers to hand,
and let’s now define the idea of a sequence of ever-closer rational approximations
more carefully. A Cauchy sequence, then, is an infinite sequence of rationals
s1 , s2 , s3 , . . . which converges – i.e. the differences |sm − sn | are as small as we
want, once we get far enough along the sequence. More carefully, take any  > 0
however small, then for some k, |sm − sn | <  for all m, n > k. Now say that two
Cauchy sequences s1 , s2 , s3 , . . . and s01 , s02 , s03 , . . . are equivalent if their members
73
 DRAFT– 31 DEC 2021
7 Set theory, less naively

eventually get arbitrarily close – i.e. when we take any  > 0 however small,
then for some k, |sn − s0n | <  for all n > k. Cauchy identifies √ real numbers
with equivalence classes of Cauchy sequences. So, for Cauchy, 2 would be the
equivalence class containing any sequence of rationals like 1.4, 1.41, 1.414, 1.4142,
1.41421, . . . , i.e. rationals whose squares approach 2.
Alternatively, dropping the picture of sequential approach, we can identify a
real number with a Dedekind cut, defined as a (proper, non-empty) subset C of
the rationals which (i) is downward closed – i.e. if q ∈ C and q 0 < q then q 0 ∈ C –
and (ii) has no largest member. For example, take the negative rationals together
with the positive ones whose square is less than√two: these form a cut. Dedekind
(more or less) identifies the positive irrational 2 with the cut we just defined.
Assuming some set theory, we can now show that – whether defined as cuts on
the rationals or defined as equivalence classes of Cauchy sequences of rationals
– these real numbers do indeed have the properties assumed in our informal
working theory of real analysis. And given that our set theory is consistent, the
resulting theory of the reals can be shown to be consistent too. Excellent!
We can now go on define functions between real numbers in terms of sets of
ordered tuples of reals, so we can develop a theory of analysis. I won’t spell this
out further here. However, you want to get to know something of how the overall
story goes, and also get some sense of what assumptions about sets are needed
for the story to work to give us a basis for reconstructing classical real analysis.
(You will need a number of levels of sets: sets of rationals, and sets of sets of
rationals, and sets of sets of sets, and up a few more levels depending on the
details.)
(c) Now, as far as construction of the reals and the foundations of analysis
are concerned, we could take the requisite set theory – the apparatus of infi-
nite sets, infinite sequences, equivalence classes and the rest – as describing a
superstructure sitting on top of a given prior basic universe of rational numbers
governed by a prior suite of numerical laws. However, we don’t need to do this.
For we can in fact already construct the rationals and simpler number systems
within set theory itself.
For the naturals, pick any set you like and call it ‘0’. And then consider e.g.
the sequence of sets 0; {0}; {{0}}; {{{0}}}; . . .. Or alternatively, consider the se-
quence 0; {0}; {0, {0}}; {0, {0}, {0, {0}}}; {0, {0}, {0, {0}}, {0, {0}, {0, {0}}}}; . . .
where at each step after the first we extend the sequence by taking the set of all
the sets we have so far. Either sequence then has the structure of the natural-
number series. There is a first member; every member has a unique successor
(which is distinct from it); different members have different successors; the se-
quence never circles around and starts repeating. So such a sequence of sets will
do as a representation, implementation, or model of the natural numbers (call
it what you will).
Let’s not get hung up about the best way to describe the situation; we will
simply say we have constructed a natural number sequence. And elementary
reasoning about sets will show that the familiar arithmetic laws about natural

74
 DRAFT– 31 DEC 2021
Elements of set theory: an overview

numbers apply to numbers as just constructed (including e.g. the principle of

arithmetical induction).
Once we have a natural number sequence we can go on to construct the
integers from it in various ways. Here’s one. Informally, any integer equals m − n
for some natural numbers m, n (to get a negative integer, take n > m). So, first
shot, we can treat an integer as an ordered pair hm, ni of natural numbers. But
since for given m and n, m − n = m0 − n0 for lots of m0 , n0 , choosing a particular
pair of natural numbers to represent an integer involves an arbitrary choice. So,
a neater second shot, we can treat an integer as an equivalence class of ordered
pairs of natural numbers (where the pairs hm, ni and hm0 , n0 i are equivalent
in the relevant way when m + n0 = m0 + n). Again the usual laws of integer
arithmetic can then be proved from basic principles about sets.
Similarly, once we have constructed the integers, we can construct rational
numbers in various ways. Informally, any rational equals p/q for integers p, q,
with q 6= 0. So, first shot, we can treat a rational numbers as a particular ordered
pair of integers. Or to avoid making a choice between equivalent renditions, we
can treat a rational as an equivalence class of ordered pairs of integers.
We again needn’t go further into the details here, though – at least once in
your mathematical life! – you will want to see them worked through in enough
detail to confirm that these can constructions can indeed all be done. The point
we want to emphasize now is simply this: once we have chosen an initial object to
play the role of 0 – the empty set is the conventional choice – and once we have
a set-building operation which we can iterate sufficiently often, and once we can
form equivalence classes from among sets we have already built, we can construct
sets to do the work of natural numbers, integers and rationals in standard ways.
Hence, we don’t need a theory of the rationals prior to set theory before we can
go on to construct the reals: the whole game can be played inside pure set theory.
(d) Another theme. It is an elementary idea that two sets are equinumerous
(have the same cardinality) just if we can match up their members one-to-one,
i.e. when there is a one-to-one correspondence, a bijection, between the sets. It
is easy to show that the set of even natural numbers, the set of primes, the set
of integers, the set of rationals are all countably infinite in the sense of being
equinumerous with the set of natural numbers.
By contrast, as we noted in §2.1, a simple argument shows that the set of
infinite binary strings is not countably infinite. Two corollaries:
1. An infinite binary string can be thought of as representing a set of natural
numbers, namely the set which contains n if and only if the n-th digit in
the string is 1; and different strings represent different sets of naturals.
Hence the powerset of the natural numbers, i.e. the set of subsets of the
naturals, is also not countably infinite.
2. An infinite binary string can equally well be thought of as representing
a real number between 0 and 1 in binary; and different strings represent
different reals. So the set of real numbers between 0 and 1 is not countably
infinite either – hence neither is the set of all the real numbers.
75
 DRAFT– 31 DEC 2021
7 Set theory, less naively

And now a famous question arises – easy to ask, but (it turns out) extraordi-
narily difficult to answer. Take an infinite collection of real numbers. It could
be equinumerous with the set of natural numbers (like, for example, the set of
real numbers 0, 1, 2, . . . ). It could be equinumerous with the set of all the real
numbers (like, for example, the set of irrational numbers). But are there any
infinite sets of reals of intermediate size (so to speak)? – can there be an infinite
subset of real numbers that can’t be put into one-to-one correspondence with
just the natural numbers and can’t be put into one-to-one correspondence with
all the real numbers either? Cantor conjectured that the answer is ‘no’; and this
negative answer is known as the Continuum Hypothesis.
Efforts to confirm or refute the Continuum Hypothesis were a major driver in
early developments of set theory. We now know the problem is a profound one
– the standard axioms of set theory don’t settle the hypothesis one way or the
other. Is there some attractive and natural additional axiom which will settle
the matter? I’ll not give a spoiler here! – but exploration of this question takes
us way beyond the initial basics of set theory.
(e) The argument that the power set of the naturals isn’t equinumerous with
the set of naturals can be generalized. Cantor’s Theorem tells us that a set is
never equinumerous with its powerset.
Note, there is a bijection between the set A and the set of singletons of
elements of A; in other words, there is a bijection between A and part of its
powerset P(A). But we’ve just seen that there is no bijection between A and the
whole of P(A). Intuitively then, A is smaller in size than P(A), which will in
turn be smaller than P(P(A)), etc. We now want to develop this intuitive idea
of one set’s having a smaller cardinal size than another into a competent general
theory about relative cardinal size.
(f) Let’s pause to consider the emerging picture.
Starting perhaps from some given urelements – elements which don’t them-
selves have members – we can form sets of them, and then sets of sets, sets of
sets of sets, and so on and on: and at each new level, we accumulate more and
more sets formed from the urelements and/or the sets formed at earlier levels.
At each level, more and more sets are formed. In particular, once we have an
infinite number of entities at one level, we get an even greater infinity of entities
at the next as we form powersets, and so on up.
Now, for purely mathematical purposes such as reconstructing analysis, it
seems that a we only need a single non-membered base-level entity, and it is
tidy to think of this as the empty set. So for internal mathematical purposes,
we can take the whole universe of sets to contain only ‘pure’ sets (when we look
at the members of members of . . . members of sets, we find nothing other than
more sets). But what if we want to be able to apply set-theoretic apparatus in
talking about e.g. widgets or wombats or (more seriously!) space-time points?
Then it might seem that we will want the base level of non-membered elements
to be populated with those widgets, wombats or space-time points as the case
might be. However, it seems that we can always code for widgets, wombats or

76
 DRAFT– 31 DEC 2021
Elements of set theory: an overview

space-time points using some kind of numbers, and we can treat those numbers
as sets. So our set-theory-for-applications can still involve only pure sets. That’s
why typical introductions to set theory either explicitly restrict themselves to
talking about pure sets, or – after officially allowing the possibility of urelements
– promptly ignore them.
(g) Lots of questions arise. Here are two:

1. First, how far can we iterate the ‘set of’ operation – how high do these levels
upon levels of sets-of-sets-of-sets-of-. . . stack up? Once we have the natural
numbers in play, we only need another dozen or so more levels of sets in
which to reconstruct ‘ordinary’ mathematics: but once we are embarked on
set theory for its own sake, how far can we go up the hierarchy of levels?

2. Second, at a particular level, how many sets do we get at that level? And
indeed, how do we ‘count’ the members of infinite sets?
With finite sets, we not only talk about their relative sizes (larger or
smaller), but actually count them and give their absolute sizes by using
finite cardinal numbers. These finite cardinals are the natural numbers,
which we have learnt can be identified with particular sets. We now want
similarly to have a story about the infinite case; we not only want an
account of relative infinite sizes but also a theory about infinite cardinal
numbers apt for giving the size of infinite collections. Again these infinite
cardinals will be identified with particular sets. But how can this story go?

It turns out that to answer both these questions, we need a new notion, the idea
of infinite ordinal numbers. We can’t say very much about this here, but some
more arm-waving pointers might still be useful.
(h) Let’s start rather naively. Here are the familiar natural numbers, but re-
sequenced with the evens in their usual order before the odds in their usual
order:
0, 2, 4, 6, . . . , 1, 3, 5, 7, . . . .
If we use ‘<’ to symbolize the order-relation here, then m < n just in case either
(i) m is even and n is odd or else (ii) m and n have the same parity and m < n.
Note that < is a well-ordering in the standard sense that it is a linear order and,
for any numbers we take, one will be the <-least.
Now, if we march through the naturals in their new <-ordering, checking
off the first one, the second one, the third one, etc., where does the number 7
come in the order? Plainly, we cannot reach it in any finite number of steps:
it comes, in a word, transfinitely far along the <-sequence. So if we want a
position-counting number (officially, an ordinal number) to tally how far along
our well-ordered sequence the number 7 is located, we will need a transfinite
ordinal. We will have to say something like this: We need to march through
all the even numbers, which here occupy positions arranged exactly like all the
natural numbers in their natural order. And then we have to go on another 4
steps. Let’s use ‘ω’ to indicate the length of the sequence of natural numbers
77
 DRAFT– 31 DEC 2021
7 Set theory, less naively

in their natural order, and we’ll call a sequence structured like the naturals in
their natural order an ω-sequence. The evens in their natural order can be lined
up one-to-one with the naturals in order, so form another ω-sequence. Hence,
to indicate how far along the re-sequenced numbers we find the number 7, it is
then tempting to say that it occurs at ω + 4-th place.
And what about the whole sequence, evens followed by odds? How long is it?
How might we count off the steps along it, starting ‘first, second, third, . . . ’ ?
After marching along as many steps as there are natural numbers in order to
treck through the evens, then – pausing only to draw breath – we have to march
on through the odds, again going through positions arranged like all the natural
numbers in their natural ordering. So, we have two ω-sequences, put end to end.
It is very natural to say that the positions in the whole sequence are tallied by
a transfinite ordinal we can denote ω + ω.
Here’s another example. There are familiar maps for coding ordered pairs of
natural numbers by a single natural: take, for example, the function which maps
m, n to [m, n] = 2m (2n + 1) − 1. And consider the following ordering on these
‘pair-numbers’ [m, n]:

[0, 0], [0, 1], [0, 2], . . . , [1, 0], [1, 1], [1, 2], . . . , [2, 0], [2, 1], [2, 2], . . . , . . .

If we now use ‘≺’ to indicate this order, then [m, n] ≺ [m0 , n0 ] just in case either
(i) m < m0 or else (ii) m = m0 and n < n0 . (This type of ordering is standardly
called lexicographic: in the present case, compare the dictionary ordering of two-
letter words drawn from an infinite alphabet.) Again, ≺ is a well-ordering on the
natural numbers.
Where does [5, 3] come in this sequence? Before we get to this ‘pair’ there
are already five blocks of the form [m, 0], [m, 1], [m, 2], . . . for fixed m, each as
long as the naturals in their usual order, first the block with m = 0, then the
block with m = 1, and three more blocks, each ω long; so the five blocks are in
total ω · 5 long. And then we have to count another four steps along, tallying
off [5, 0], [5, 1], [5, 2], [5, 3]. So it is inviting to say we have to count along to the
ω · 5 + 4-th step in the sequence to get to the ‘pair’ [5, 3].
And what about the whole sequence of ‘pairs’ ? We have blocks ω long, with
the blocks themselves arranged in a sequence ω long. So this time it is tempting to
say that the positions in the whole sequence of ‘pairs’ are tallied by a transfinite
ordinal we can indicate by ω · ω.
We can continue. Suppose we re-arrange the natural numbers into a new
well-ordering like this: take all the numbers of the form 2l · 3m · 5n , ordered by
ordering the triples hl, m, ni lexicographically, followed by the remaining naturals
in their normal order. We tally positions in this sequence by the transfinite
ordinal ω · ω · ω + ω. And so it goes.
Note by the way that we have so far been considering just (re)orderings of
the familiar set of natural numbers – the sequences are equinumerous, and have
the same infinite cardinal size; but the well-orders are tallied by different infinite
ordinal numbers. Or so we want to say.

78
 DRAFT– 31 DEC 2021
Elements of set theory: an overview

But is this sort of naive talk of transfinite ordinals really legitimate? Well, it
was one of Cantor’s great and lasting achievements to show that we can indeed
start to make perfectly good sense of all this.
Now, in Cantor’s work the theory of transfinite ordinals is already entangled
with his nascent set theory. Von Neumann later cemented the marriage by giving
the canonical treatment of ordinals in set theory. And it is via this treatment that
students now typically first encounter the arithmetic of transfinite ordinals, some
way into a full-blown course about set theory. This approach can, unsurprisingly,
give the impression that you have to buy into quite a lot of set theory in order to
understand even the basics about ordinals and their arithmetic. However, not so.
Our little examples so far are of recursive (re)orderings of the natural numbers
– i.e. a computer can decide, given two numbers, which way round they come in
the ordering. There is a whole theory of recursive ordinals which talks about how
to tally the lengths of such (re)orderings of the naturals, which has important
applications e.g. in proof theory. And these tame beginnings of the theory of
transfinite ordinals needn’t entangle us with the kind of rather wildly infinitary
and non-constructive ideas characteristic of modern set theory.
(i) However, here we are concerned with set theory, and so our next topic
will naturally be von Neumann’s very elegant implementation of ordinals in set
theory as the ‘hereditarily transitive sets’. The basic idea is to define a particular
well-ordered sequence of sets – call them the ordinalsvN – and show that any
well-ordered collection of objects, however long the ordering, will have the same
type of ordering as an initial segment of these ordinalsvN . So we can use the
ordinalsvN as a universal measuring scale against which to tally the length of
any well-ordering.
And at this point, I’ll have to leave it to you to explore the details of the
construction of the ordinalsvN in the recommended readings. But once we have
them available, we can say more about the way that the universe of sets is
structured; we can take the levels to be indexed by ordinalsvN (and then assume
that for every ordinal there is a corresponding level of the universe).
We can also now define a scale of cardinal size. We noted that well-orderings
of different ordinal length can be equinumerous; different ordinalsvN can have
the same cardinality. So von Neumann’s next trick is to define a cardinal number
to be the first ordinal (in the well-ordered sequence of ordinals) in a family of
equinumerous ordinals. Again this neat idea we’ll have to leave for the moment
for later exploration. However – and this is an important point – to get this
to all work out as we want, in particular to ensure that we can assign any two
non-equinumerous sets respective cardinalities κ and λ such that either κ < λ
or λ < κ, we will need the Axiom of Choice. (This is something to keep looking
out for when beginning set theory: where do we start to need to appeal to some
Choice principle?)
(j) We are perhaps already rather past the point where scene-setting remarks
at this level of unspecific generality can be very helpful. Time to dive into the
details! But one final important observation before you start.

79
 DRAFT– 31 DEC 2021
7 Set theory, less naively

The themes we have been touching on can and perhaps should initially be
presented in a relatively informal style. But something else that also belongs
here near the beginning of your first forays into set theory is an account of the
development of axiomatic ZFC (Zermelo-Fraenkel set theory with Choice) as the
now standard way of formally regimenting set theory. As you will see, different
books take different approaches to the question of just when it is best to start
getting more rigorously axiomatic, formalizing our set-theoretic ideas.
Now, there’s a historical point worth noting, which explains something about
the shape of the standard axiomatization. You’ll recall from the remarks in
§2.1(b) that a set theory which makes the assumption that every property has
an extension will be inconsistent. So Zermelo set out in an epoch-making 1908
paper to lay down what he thought were the basic assumptions about sets that
mathematicians actually needed, while not overshooting and falling into such
contradictions. His axiomatization was not, it seems, initially guided by a positive
conception of the universe of sets so much as by the desire to keep safe and not
assume too much. But in the 1930s, Zermelo himself and especially Gödel came
to develop the conception of sets as a hierarchy of levels (with new sets always
formed from objects at lower levels, so never containing themselves, and with no
end to the levels where we form more sets from what we have accumulated so
far, so we never get to a paradoxical set of all sets). This cumulative hierarchy
is described and explored in the standard texts. Once this conception is in play,
it does invite a more direct and explicit axiomatization as a story about levels
and sets formed at levels: however, it was only much later that this positively
motivated axiomatization gets spelt out, particularly in what has come to be
called Scott-Potter set theory. Most text books stick for their official axioms
to the Zermelo approach, hence giving what looks to be a rather unmotivated
selection of axioms whose attraction is that they all look reasonably modest and
separately in keeping with the hierarchical picture, so unlikely to get us into
trouble. In particular the initial recommendations below take this conventional
line.

7.2 Main recommendations on set theory

This present chapter is, as advertised, just about the basics of set theory. Even
here, however, there are is a very large number of books to choose from, so an
annotated Guide will (I hope!) be particularly welcome.
But first, if you want a more expansive 35pp. overview of basic set theory,
with considerably more mathematical detail and argument, I think the following
chapter (the best in the book?) works pretty well:

1. Robert S. Wolf, A Tour Through Mathematical Logic (Mathematical As-

sociation of America, 2005), Ch. 2, ‘Axiomatic set theory’.

And let me me mention again an introduction to set-theoretic ideas which I

noted in §2.2, which you may have skipped past then.

80
 DRAFT– 31 DEC 2021
Main recommendations on set theory

2. Cambridge lecture notes by Tim Button have become incorporated into

Set Theory: An Open Introduction* (2019) tinyurl.com/opensettheory.
This short book is one of the most successful outputs from the Open
Logic Project. Its earlier chapters in particular are extremely good, and
are very clear on the conceptual motivation for the iterative conception
of sets and its relation to the standard ZFC axiomatization. However,
things get a bit patchier as the book progresses: later chapters on ordi-
nals, cardinals, and choice, get rather tougher, and might work better (I
think) as parallel readings to the more expansive main recommendations
I’m about to make. But very well worth looking at.

Since Button can’t really get into enough detail into his brisk notes, most readers
will want to look instead at one or other of the first two of the following admirable
‘entry level’ treatments which cover rather more material in rather more depth
but still very accessibly:

3. Derek Goldrei, Classic Set Theory (Chapman & Hall/CRC 1996).

The author taught at the Open University, and wrote specifically for
students engaged in remote learning: his book has the friendly subti-
tle ‘For guided independent study’. The result as you might expect –
especially if you looked at Goldrei’s FOL text mentioned in §3.3 – is
exceptionally clear, and it is indeed admirably well-structured for in-
dependent self-teaching. Moreover, it is rather attractively written (as
set theory books go!). The coverage is very much as as outlined in our
overview. And one particularly nice feature is the way the book (un-
usually?) spends enough time motivating the idea of transfinite ordinal
numbers before turning to their now conventional implementation in set
theory.
4. Herbert B. Enderton’s, The Elements of Set Theory (Academic Press,
1977) forms a trilogy along with the author’s Logic and Computability
which we have already mentioned in earlier chapters.
This book again has exactly the coverage we need at this stage. But
more than that, it is particularly clear in marking off the informal devel-
opment of the theory of sets, cardinals, ordinals etc. (guided by the con-
ception of sets as constructed in a cumulative hierarchy) from the formal
axiomatization of ZFC. It is also particularly good and non-confusing
about what is involved in (apparent) talk of classes which are too big
to be sets – something that can mystify beginners. It is written with a
certain lightness of touch and proofs are often presented in particularly
well-signposted stages. The last couple of chapters perhaps do get a bit
tougher, but overall this really is quite exemplary exposition.

Also starting from scratch, we find two further excellent books which are rather
less conventional in style:

81
 DRAFT– 31 DEC 2021
7 Set theory, less naively

5. Winfried Just and Martin Weese, Discovering Modern Set Theory I: The
Basics (American Mathematical Society, 1996).
Covers similar ground to Goldrei and Enderton, but perhaps more
zestfully and with a little more discussion of conceptually interesting
issues. At some places, it is more challenging – the pace can be a bit
uneven.
I like the style a lot, though, and think it works very well. I don’t mean
the occasional (slightly laboured?) jokes: I mean the in-the-classroom
feel of the way that proofs are explored and motivated, and also the way
that teach-yourself exercises are integrated into the text. The book is ev-
idently written by enthusiastic teachers, and the result is very engaging.
(The story continues in a second volume.)
6. Yiannis Moschovakis, Notes on Set Theory (Springer, 2nd edition 2006).
This also takes a slightly more individual path through the material
than Goldrei and Enderton, with occasional bumpier passages, and with
glimpses ahead. But to my mind, this is very attractively written, and
again nicely complements and reinforces what you’ll learn from the more
conventional books.

Of these two pairs of books, I’d rather strongly advise reading one of the first
pair and then one of the second pair.
I will add two more firm recommendations at this level. The first might come
as a bit of surprise, as it is something of a ‘blast from the past’. But we shouldn’t
ignore old classics – they can have a lot to teach us even when we have read the
more recent books, and this is very illuminating:

7. Abraham Fraenkel, Yehoshua Bar-Hillel and Azriel Levy, Foundations

of Set-Theory (North-Holland, originally 1958; but you want the revised
2nd edition 1973): Chapters 1 and 2 are the immediately relevant ones.
Both philosophers and mathematicians should appreciate the way this
puts the development of our canonical ZFC set theory into some con-
text, and also discusses alternative approaches. Standard textbooks can
present our canonical theory in a way that makes it seem that ZFC has
to be the One True Set Theory, so it is worth understanding more about
how it was arrived at and where some choice points are. This book re-
ally is attractively readable, and should be very largely accessible at this
early stage. I’m not myself an enthusiast for history for history’s sake:
but it is very much worth knowing the stories that unfold here.

Now, as I noted in the initial overview section, one thing that every set-theory
novice now acquires is the picture of the universe of sets as built up in a hierarchy
of stages or levels, each level containing all the sets at previous levels plus new
ones (so the levels are cumulative). It is significant that, as Fraenkel et al. makes
clear, the picture wasn’t firmly in place from the beginning. But the hierarchical
conception of the universe of sets is brought to the foreground in
82
 DRAFT– 31 DEC 2021
Some parallel/additional reading on standard ZFC

8. Michael Potter, Set Theory and Its Philosophy (OUP, 2004).

For philosophers and for mathematicians concerned with foundational
issues this surely is a ‘must read’, a unique blend of mathematical expo-
sition (mostly about the level of Enderton, with a few glimpses beyond)
and extensive conceptual commentary. Potter is presenting not straight
ZFC but a very attractive variant due to Dana Scott whose axioms more
directly encapsulate the idea of the cumulative hierarchy of sets. It has
to be said that there are passages which are harder going, sometimes
because of the philosophical ideas involved, and sometimes because of
occasional expositional compression. However, if you have already read
a set theory text from the main list, you should have no problems.

7.3 Some parallel/additional reading on standard ZFC

There are so many good set theory books with different virtues, many by very
distinguished authors, that I should certainly pause to mention some more.
Let me begin by mentioning a bare-bones, introductory book, a level or so
down in coverage and detail from what we really want here, but which some
might find a helpful preliminary read:

9. Paul Halmos, Naive Set Theory* (1960: republished by Martino Fine

Books, 2011).
The purpose of this famous book, Halmos says in his Preface, is “to tell
the beginning student . . . the basic set-theoretic facts of life, and to do
so with the minimum of philosophical discourse and logical formalism”.
He proceeds pretty naively in the second sense we identified in §2.1(b).
True, he tells us about some official axioms as he goes along, but he
doesn’t explore the development of set theory inside a resulting formal
theory. This is informally written in an unusually conversational style
for a maths book, concentrating on the motivation for various concepts
and constructions. Some might warm to this classic (though perhaps you
ignore the remarks in the Preface about set theory for applications being
‘pretty trivial stuff’ !).

Next, here are four introductory books at the right sort of level, listed in order of
publication; each has many things to recommend it to beginners. Browse through
to see which might suit your interests:

10. D. van Dalen, H.C. Doets and H. de Swart, Sets: Naive, Axiomatic and
Applied (Pergamon, 1978).
The first chapter covers the sort of elementary (semi)-naive set theory
that any mathematician needs to know, up to an account of cardinal
numbers, and then takes a first look at the paradox-avoiding ZF axiom-
atization. This is very attractively and illuminatingly done. (Or at least,
the conceptual presentation is attractive – sadly, and a sign of its time
of publication, the book seems to have been photo-typeset from original
83
 DRAFT– 31 DEC 2021
7 Set theory, less naively

pages produced on electric typewriter, and the result is visually not at-
tractive at all.)
The second chapter carries on the presentation axiomatic set theory,
with a lot about ordinals, and getting as far as talking about higher
infinities, measurable cardinals and the like. The final chapter considers
some applications of various set theoretic notions and principles. Well
worth seeking out, if you don’t find the typography off-putting.
11. Karel Hrbacek and Thomas Jech, Introduction to Set Theory (Marcel
Dekker, 3rd edition 1999).
Eventually this book goes a bit further than Enderton or Goldrei (more
so in the 3rd edition than earlier ones), and you could – on a first reading
– skip some of the later material. Though do look at the final chapter
which gives a remarkably accessible glimpse ahead towards large cardinal
axioms and independence proofs. Recommended if you want to consoli-
date your understanding by reading a second presentation of the basics
and want then to push on just a bit.
Jech is a major author on set theory whom we’ll encounter again,
and Hrbacek once won a AMA prize for maths writing. So, unsurpris-
ingly, this is a very nicely put together book, which could very well have
featured as a main recommendation.
12. Keith Devlin, The Joy of Sets (Springer, 1979: 2nd edn. 1993).
The opening chapters of this book are remarkably lucid and attrac-
tively written. The opening chapter explores ‘naive’ ideas about sets and
some set-theoretic constructions, and the next chapter introduces axioms
for ZFC pretty gently (indeed, non-mathematicians could particularly
like Chs 1 and 2, omitting §2.6). Things then speed up a bit, and by
the end of Ch. 3 – some 100 pages into the book – we are pretty much
up to the coverage of Goldrei’s much longer first six chapters, though
Goldrei says more about (re)constructing classical maths in set theory.
Some will prefer Devlin’s fast-track version. (The rest of the book then
covers non-introductory topics in set theory, of the kind we take up again
in §12.4.)
13. Judith Roitman, Introduction to Modern Set Theory* (Wiley, 1990: a
2011 version is available at tinyurl.com/roitmanset.
Relatively short, and very engagingly written, this book covers quite
a bit of ground – we’ve reached the constructible universe by p. 90 of the
downloadable pdf version, and there’s even room for a concluding chapter
on ‘Semi-advanced set theory’ which says something about large cardi-
nals and infinite combinatorics. A few quibbles aside, this could make
excellent revision material as Roitman is particularly good at highlight-
ing key ideas without getting bogged down in too many details.

Those four books all aim to cover the basics in some detail. The next two books
are much shorter, and are differently focused.

84
 DRAFT– 31 DEC 2021
Further conceptual reflection on set theories

14. A. Shen and N. K. Vereshchagin, Basic Set Theory (American Mathe-

matical Society, 2002).
Just over 100 pages, and mostly about ordinals. But it is very read-
able, with 151 ‘Problems’ as you go along to test your understanding.
Potentially very helpful by way of revision/consolidation.
15. Ernest Schimmerling, A Course on Set Theory (CUP, 2011)
This is perhaps slightly mistitled, if ‘course’ suggests a comprehensive
treatment. This is just 160 pages long, starting off with a brisk intro-
duction to ZFC, ordinals, and cardinals. But then the author explores
applications of set theory to other areas of mathematics such as topology,
analysis, and combinatorics, in a way that will be particularly interesting
to mathematicians. An engaging supplementary read at this level.
Applications of set theory to mathematics are also highlighted in a book in the
LMS Student Text series which is worth mentioning here:
16. Krzysztof Ciesielski, Set Theory for the Working Mathematician (CUP,
1997).
This eventually touches on advanced topics in the set theory. But the
earlier chapters introduce some basic set theory, which is then put to
work in e.g. constructing some strange real functions. So this might well
appeal to mathematicians who know some analysis and want to see set
theory being applied; you could indeed tackle Chs 6 to 8 on the basis of
other introductions.

7.4 Further conceptual reflection on set theories

(a) A preliminary point. Go back to our starting point when we introduced set
theory as giving us a ‘foundation’ for real analysis. But what does that really
mean? As Penelope Maddy notes, “It’s more or less standard orthodoxy these
days that set theory . . . provides a foundation for classical mathematics. Oddly
enough, it’s less clear what ‘providing a foundation’ comes to.” Her opening pages
then give a particularly clear and crisp account of what might be meant by talk
of foundations in this context. It is very well worth reading for orientation:
17. Penelope Maddy, ‘Set-theoretic foundations’, in A. Caicedo et al., eds.,
Foundations of Mathematics (AMS, 2017), available at tinyurl.com/maddy-
found See §1 in particular.
(b) Michael Potter’s Set Theory and Its Philosophy must be the starting point
for further philosophical reflections about set theory. In particular, he gives a
good account of how our standard set theory emerges from a certain hierarchical
conception of the universe of sets as built up in stages. There is also now an
excellent more recent exploration of the conceptual basis of set theory in
18. Luca Incurvati, Conceptions of Set and the Foundations of Mathematics
(CUP, 2020).
85
 DRAFT– 31 DEC 2021
7 Set theory, less naively

Incurvati gives more by way of a careful defence of the hierarchical

conception of sets and also an unusually sympathetic critique of some ri-
val conceptions and the set theories which they motivate. Knowledgeable
and readable.

Rather differently, if you haven’t tackled their book in working on model theory,
you will want to look at

19. Tim Button and Sean Walsh’s Philosophy and Model Theory* (OUP,
2018).
Now see especially §1.B (on first-order vs second-order ZFC), Ch. 8
(on models of set theory), and perhaps Ch. 11 (more on Scott-Potter set
theory).

7.5 A little more history

As already shown in the recommended book by Fraenkel, Bar-Hillel and Levy,
the history of set theory is a long and tangled story, fascinating in its own
right and conceptually illuminating too. José Ferreirós has an impressive book
Labyrinth of Thought: A History of Set Theory and its Role in Modern Mathe-
matics (Birkhäuser 1999). But that’s more than most readers are likely to want.
But you will find some of the headlines here, worth chasing up especially if you
didn’t read the book by Fraenkel et al.:

20. José Ferreirós, ‘The early development of set theory’, The Stanford En-
cyclopaedia of Philosophy, available at tinyurl.com/sep-devset.

This article has references to many more articles, like Kanimori’s fine piece on
‘The mathematical development of set theory from Cantor to Cohen’. But you
might to need to be on top of rather more set theory before getting to grips with
that.

7.6 Postscript: Other treatments?

What else is there? A classic introduction is given by Patrick Suppes, Axiom-
atic Set Theory* (vast Nostrand 1960, republished by Dover 1972). Clear and
straightforward as far as it goes: but there are better alternatives now. There is
also another classic book by Azriel Levy with the inviting title Basic Set Theory*
(Springer 1979, republished by Dover 2002). However, while this is still ‘basic’ in
the sense of not dealing with topics like forcing, this is quite an advanced-level
treatment of the set-theoretic fundamentals. So let’s return to it in §12.4.
András Hajnal and Peter Hamburger have a book Set Theory (CUP, 1999)
which is also in the LMS Student Text series. They nicely bring out how much of
the basic theory of cardinals, ordinals, and transfinite recursion can be developed
in a semi-informal way, before introducing a full-fledged axiomatized set theory.

86
 DRAFT– 31 DEC 2021
Postscript: Other treatments?

But I think Enderton or van Dalen et al. do this better. The second part of this
book is on more advanced topics in combinatorial set theory.
George Tourlakis’s Lectures in Logic and Set Theory, Volume 2: Set Theory
(CUP, 2003) has been recommended to me a number of times. Although this is
the second of two volumes, it is a stand-alone text. You can probably already
skip over the initial chapter on FOL, consulting if/when needed. That still leaves
over 400 pages on basic set theory, with long chapters on the usual axioms, on
the Axiom of Choice, on the natural numbers, on order and ordinals, and on
cardinality. (The final chapter on forcing should be omitted at this stage, and
strikes me as considerably less clear than what precedes it.)
As the title suggests, Tourlakis aims to retain something of the relaxed style
of the lecture room, complete with occasional asides and digressions. And as the
page length suggests, the pace is quite gentle and expansive, with room to pause
over questions of conceptual motivation etc. However, simple constructions and
results take a very long time to arrive. For example, we don’t get to Cantor’s
theorem on the uncountability of P(ω) until p. 455! So while this book might
be worth dipping into for some of the motivational explanations, I can’t myself
recommend it overall.
Finally, I’ll mention another more recent text from the same publisher, Daniel
W. Cunningham’s Set Theory: A First Course (CUP, 2016). But this doesn’t
strike me as a particularly friendly introduction. As the book progresses, it turns
into pages of old-school Definition/Lemma/Theorem/Proof with rather too little
commentary; key ideas seem often to be introduced in a phrase, without much
discursive explanation. (Readers who care about the logical niceties will also
raise their eyebrows at the author’s over-causal way with use and mention, or
e.g. the too-typically hopeless passage about replacing variables with values on
p. 14. And this isn’t just being pernickety: what exactly are we to make of the
claim on p. 31 that a class is “any collection of the form {x : A(x)}”? So not
recommended to logicians of a sensitive disposition!)

87
 DRAFT– 31 DEC 2021

8 Intuitionistic logic

In the briefest headline terms, intuitionistic logic is what you get if you drop the
classical principle that ¬¬A implies A (or equivalently drop the law of excluded
middle which says that A ∨ ¬A always holds). But why would we want to do
that? And what further consequences for our logic does that have?

8.1 A formal system

(a) To fix ideas, it will help to have in front of us a particular natural deduction
system in Gentzen style, initially for propositional logic.
We assume that at least the three binary connectives ∧, ∨, → are built in,
together with the absudity constant ⊥.
The connectives are then governed by pairs of introduction and elimination
rules. Here again, for the record (and for future reference) are the usual intro-
duction rules, presented in the short-hand way you should now be familiar with
from work on standard FOL:
[A]
..
A B A B .
(∧I) (∨I) (→I)
A∧B A∨B A∨B B
A→B
Each elimination rule then in effect just undoes an application of the corre-
sponding introduction rule (putting it roughly, for each binary connective , its
elimination rule allows us to argue onwards from A  B to a conclusion that
we could already have derived from what was required to derive A  B by its
introduction rule):
[A] [B]
.. ..
A∧B A∧B . . A A→B
(∧E) (∨E) (→E)
A B A∨B C C B
C
We next take the absurdity constant to be governed by the rule that given ⊥
we can derive anything – ex falso, quodlibet.
Finally, what about negation? One option is to treat ¬A as simply an ab-
breviation for A → ⊥. The introduction and elimination rules given for the
conditional then immediately yield the following as special cases:

88
 DRAFT– 31 DEC 2021
Overview: why intuitionistic logic?
[A]
..
. A ¬A
(¬I) (¬E)
⊥ ⊥
¬A
Alternatively, we can take these to be the introduction and elimination rules
governing a primitive built-in negation connective. Nothing hangs on this choice.
We then define IPL, intuitionistic propositional logic (in its natural deduction
version), to be the logic governed these rules.
The described rules are of course all rules of classical logic too. However, the
intuitionistic system is strictly weaker in the sense that the following classically
acceptable principles are not derived rules of our intuitionistic logic:
[¬A]
..
¬¬A .
(DN) (LEM) A ∨ ¬A (CR)
A ⊥
A
DN allows us to drop double negations. LEM is the Law of Excluded Middle,
which permits us to infer A ∨ ¬A whenever we want, from no assumptions. CR
is the classical reductio rule. And these three rules are equivalent in the sense
that adding any one of them to intuitionistic propositional logic enables us to
prove all the same conclusions; each way, we get back full classical propositional
logic.
(b) If only for brevity’s sake, we will largely be concentrating on propositional
logic in the two introductory overviews which follow. But we should briefly note
what it takes to get intuitionistic predicate logic in natural deduction form.
Technically, it’s very straightforward. Just as the rules for ∧ and ∨ are the
same in classical and intuitionist logic, the rules for generalized conjunctions and
generalized disjunctions remain the same too. In other words, to get intuition-
istic predicate logic we simply add to IPL the same pair of introduction and
elimination rules for ∀ and ∃ as for classical logic.
But note, because of the different background propositional logic – in particu-
lar, because of the different rules concerning negation – these familiar quantifier
rules no longer have all the same implications in the intuitionistic setting. For
example ∃xA(x) is no longer equivalent to ¬∀x¬A(x). More about this below.

8.2 Overview: why intuitionistic logic?

(a) A little experimentation quickly suggests that we indeed cannot derive an
instance of excluded middle like P ∨ ¬P in IPL. But how can we prove that this
is underivable?
There’s a proof-theoretic argument. We examine the structure of proofs in
IPL, and thereby show that we can only prove A ∨ B as a theorem (i.e. from
no premisses) if there is a proof of A or a proof of B. Since neither P nor ¬P is
a theorem of intuitionistic logic (with P atomic), it follows that P ∨ ¬P isn’t a
theorem either.
89
 DRAFT– 31 DEC 2021
8 Intuitionistic logic

Alternatively, there’s a semantic argument. We find some new, non-classical,

way of interpreting IPL as a formal system, an interpretation on which the
intuitionistic rules of inference are still acceptable, but on which the double
negation rule and its equivalents are clearly not acceptable. It will then follow
that buying into IPL can’t by itself commit us to those classical rules. How might
this new interpretation go?
It is natural to think of a correct assertion as one that corresponds to some
realm of facts (whatever that means exactly). But suppose just for a moment
that we instead think of correctness as a matter of being warranted, where we
understand this in the following strong sense: A is warranted if and only if there is
an informal proof which provides a direct certification for A’s correctness. Then
here is a reasonably natural story about how to characterize the connectives in
this new framework (it’s a rough version of what’s called the BHK – Brouwer-
Heyting-Kolmorgorov – interpretation):

(i) (A ∧ B) is warranted iff (if and only if) A and B are both warranted.
(ii) While there may be other ways of arriving at a disjunction, the direct
and ideally informative way of certifying a disjunction’s correctness is by
establishing one or other disjunct. So we will count (A ∨ B) as warranted
iff at least one disjunct is certified to be correct, i.e. iff there is a warrant
for A or a warrant for B.
(iii) A warranted conditional (A → B) must be one that, together with the
warranted assertion A, will enable us to derive another warranted assertion
B by using modus ponens. Hence (A → B) is directly warranted iff there
is a way of converting any warrant for A into a warrant for B.
(iv) ¬A is warranted iff we have a warrant for ruling out A because it leads to
something absurd (given what else is warranted).
(v) ⊥ is never warranted.

Then, in keeping with this approach, we will think of a reliable inference as one
that takes us from warranted premisses to a warranted conclusion.
Now, in this framework, the familiar introduction rules for the connectives
will still be acceptable, for they will evidently be warrant-preserving (given our
interpretation of the connectives). But as we said, the various elimination rules
in effect just ‘undo’ the effects of the introduction rules: so they should come for
free along with the introduction rules. Finally, we can still endorse EFQ – the
plausible thought is that if, per impossible, the absurd is warrantedly assertible,
then all hell breaks loose, and anything goes.
Hence, regarded now as warrant-preserving rules, all our IPL rules can remain
in place. However:

1. DN will not be acceptable in this framework. We might have a warrant for

ruling out being able actually to rule out A, so we can warrantedly assert
¬¬A. But that doesn’t put us in a position to warrantedly assert A. We
might just have to remain neutral about A

90
 DRAFT– 31 DEC 2021
Overview: why intuitionistic logic?

2. Likewise LEM will not be acceptable. On the present understanding of the

connectives, (A∨¬A) would be correct, i.e. directly warranted, just if there
is a warrant for A or a warrant for ruling out A. But again, must there
always be a way of justifiably deciding a conjecture A in the relevant area
of inquiry one way or the other? Some things may be beyond our ken.

Again, for similar reasons, CR is not acceptable either in this framework: but I
won’t keep mentioning this third rule.
In sum, then, if we want a propositional logic suitable as a framework for
regimenting arguments which preserve warranted assertability, we should stick
with the core rules of IPL – and shouldn’t endorse those further distinctively
classical laws.
But be very careful here! It is one thing to stand back from endorsing the law
of excluded middle. It would be something else entirely actually to deny some
instance of the law. In fact, it is an easy exercise to show that, even in IPL, any
outright negation of an instance – i.e. any sentence of the form ¬(A ∨ ¬A) –
entails absurdity!
(b) The double negation rule DN of classical logic is an outlier, not belonging
to one of the matched pairs introduction/elimination rules. Now we see the
significance of this. Its special status leaves room for an interpretation on which
the remaining rules – the rules of IPL – hold good, but DN doesn’t. Hence, as
we wanted to show, DN is not derivable as a rule of intuitionistic propositional
logic. Nor is LEM.
True, our version of the semantic argument as presented so far might seem
all a bit too arm-waving for comfort; after all, the notion of warrant as we
characterized it so can hardly be said to be ideally clear! But let’s not fuss
about details now. We’ll soon meet a rigorous story partially inspired by this
notion which gives us an entirely uncontroversial, technically kosher, proof that
DN and its equivalents are, as claimed, independent of the rules of IPL.
Things do get controversial, though, when it is claimed that DN and LEM
really don’t apply in some particular domain of inquiry, because in this do-
main there can indeed be no more to correctness than having a warrant in the
form of a direct informal proof. Now, so-called intuitionists do indeed hold that
mathematics is a case in point. Mathematical truth, they say, doesn’t consist
in correspondence with facts about abstract objects laid out in some Platonic
heaven (after all, there are familiar worries: what kind of objects could these ideal
mathematical entities be? how could we possibly know about them?). Rather,
the story goes, the mathematical world is in some sense our construction, and
being mathematically correct can be no more than a matter of being assertible on
the basis of a proof elaborating our constructions – meaning not a proof in this
or that formal system but a chain of reasoning satisfying informal mathematical
standards for being a direct proof.
Consider, for example, the following argument, intended to show that (C),
there is a pair of irrational numbers a and b such that ab is rational:

91
 DRAFT– 31 DEC 2021
8 Intuitionistic logic
√ √2
Either (i) 2 is rational, or √ (ii) it isn’t. In case (i) we are done:
we can simply put a = b = 2, and hence (C) then holds. In case
√ √2 √
(ii) put a = 2 , b = 2. Then a is irrational by assumption, b is
b
√ √2 √2 √ 2
irrational, while a = ( 2 ) = 2 = 2 and hence is rational, so
(C) again holds. Either way, (C).

It will be agreed on all sides that this argument isn’t ideally satisfying. But the
intuitionist goes further, and claims that this argument actually fails to estab-
lish (C), because we haven’t yet constructed a specific a and b to warrant (C).
The cited argument assumes that either (i) or (ii) holds, and – the intuitionist
complains – we are not entitled to assume this when we are given no reason to
suppose that one or other disjunct can be warranted by a construction.
(c) For an intuitionist, then, the appropriate logic is not full classical two-
valued logic but rather our cut-down intuitionistic logic (hence the name!),
because this is the right logic for correctness-as-informal-direct-provability.
Or so, roughly, goes the story. Plainly, we can’t even begin to discuss here the
highly contentious issues about the nature of truth and provability in mathemat-
ics which first led to the advocacy of intuitionistic logic (if you want to know a
bit more, there are some initial references in the recommended reading). But no
matter: there are plenty of other reasons too for being interested in intuitionistic
logic, which keeps recurring in various contexts (e.g. in computer science and
in category theory). And as we will see in the next chapter, the fact that its
rules come in matched introduction/elimination pairs makes intuitionistic logic
proof-theoretically particularly neat.
For now, though, let’s just say a bit more about what can and can’t be proved
in IPL and its extension by the quantifier rules, and also introduce one of the
more formal ways of semantically modelling it.

8.3 Overview: more proof theory, more semantics

(a) We use ‘`c ’ to symbolize classical derivability, and ‘`i ’ to symbolize deriv-
ability in intuitionistic logic. Then:

(i) The familiar classical laws governing just conjunctions and disjunctions
stay the same: so, for example, we still have A ∧ (B ∧ C) `i (A ∧ B) ∧ C
and A ∨ (B ∧ C) `i (A ∨ B) ∧ (A ∨ C). However, although the conditional
rules of inference are the same in classical and intuitionist logic, the laws
governing the conditional are not the same. Classically, we have Peirce’s
Law, (A → B) → A `c A; but we do not have (A → B) → A `i A.
(ii) Classically, the binary connectives are interdefinable using negation. Not
so in IPL. We do have for example (A∨B) `i ¬(¬A∧¬B). But the converse
doesn’t hold – a good rule of thumb is that IPL makes disjunctions harder
to prove. However, ¬(¬A ∧ ¬B) `i ¬¬(A ∨ B).

92
 DRAFT– 31 DEC 2021
Overview: more proof theory, more semantics

Likewise, we do have (¬A ∨ B) `i (A → B). But the converse doesn’t

hold – though (A → B) `i ¬¬(¬A ∨ B).
(iii) The connectives in IPL are not truth-functional. But their behaviour in a
sense still tracks the classical truth-tables.
Take, for example, the classical table for the material conditional. We
can read that as telling us that when A and B holds so does A → B;
when A holds and B doesn’t (so ¬B does), then A → B doesn’t hold (so
¬(A → B) does); while when ¬A holds, so does (A → B) (whether we also
have B or ¬B).
Correspondingly, in intuitionistic logic, we still have A, B `i (A → B);
A, ¬B `i ¬(A → B); and ¬A `i (A → B). The intuitionistic conditional
therefore shares some of the same unwelcome(?) features as the classical
material conditional.
(iv) Glivenko’s theorem: if A is a propositional formula, `c A just when `i ¬¬A.
Note, though, that this doesn’t apply in general to quantified formulas.
(v) The so-called disjunction property applies in IPL, I.e. if Γ `i (A ∨ B) then
either Γ `i A or Γ `i B. And, moving to quantified intuitionistic logic, we
have the following analogue: sentences alike: we only have Γ `i ∃xAx if we
can provide a witness for the existentially quantified sentence, i.e. for some
term t, Γ `i At.
(vi) Just as conjunction and disjunction are not intuitionistically interdefinable
using negation, so too for the universal and existential quantifiers. Thus
while ∃xA `i ¬∀x¬A, the converse doesn’t hold – though, inserting a
double negation, we do have ¬∀x¬A `i ¬¬∃xA. Likewise, ∀xA `i ¬∃x¬A.
But again the converse doesn’t hold – though ¬∃x¬A `i ∀x¬¬A.
(vii) A theme is emerging! While some classical results fail in intuitionistic logic,
inserting some double negations will give corresponding intuitionistic re-
sults. This theme can be made more precise, in various ways. Consider,
for example, the following translation scheme T for mapping classical to
intuitionistic sentences – a double-negation translation:
a) AT := ¬¬A, for atomic wffs A; ⊥T := ⊥
b) (A ∧ B)T := AT ∧ B T
c) (A ∨ B)T := ¬¬(AT ∨ B T )
d) (A → B)T := AT → B T
e) (¬A)T := ¬AT
f) (∀xA)T := ∀xAT
g) (∃xA)T := ¬¬∃xAT
Suppose ΓT comprises the double-negation translations of the sentences in
the set Γ. Then we have the following key theorem due (independently) to
Gödel and Gentzen:
Γ `c A if and only if ΓT `i AT .

93
 DRAFT– 31 DEC 2021
8 Intuitionistic logic

(b) Two comments on the Gödel/Gentzen theorem. First, it shows that for
every classical result, there is already a corresponding intuitionistic one which
has additional double negation signs in the right places. So we can think of
classical logic not so much as what you get by adding to intuitionist logic but
rather as what you get by ignoring a distinction that the intuitionist thinks is
of central importance, namely the distinction between A and ¬¬A.
Second, note this particular consequence of the theorem: Γ `c ⊥ if and only if
ΓT `i ⊥. So if the classical theory Γ is inconsistent by classical standards, then its
intuitionistic translation ΓT is already inconsistent by intuitionistic standards.
Roughly speaking, then, if we have worries about the consistency of a classical
theory, retreating to an intuitionistic version isn’t going to help. As you’ll see
from the readings, this observation had significant historical impact in debates
in the foundations of mathematics.
(c) Let’s now return to those earlier arm-waving semantic remarks in §8.2(a).
They can be sharpened up in various ways, but here I’ll just briefly consider (a
version of) Saul Kripke’s semantics for IPL. I’ll leave it to you to find out how
the story can be extended to cover quantified intuitionistic logic.
Take things in stages. First, imagine an enquirer, starting from a ground state
of knowledge g; she then proceeds to expand her knowledge, through a sequence
of possible further states K. Different routes forward can be possible, so we can
think of these states as situated on a branching array of possibilities rooted at g
(not strictly a ‘tree’ though, as we can allow branches to later rejoin, reflecting
the fact that our enquirer can arrive at the same knowledge state by different
routes). If she can get from state k ∈ K to the state k 0 ∈ K by zero or more
steps, then we’ll write k ≤ k 0 . So, to model the situation a bit more abstractly,
let’s say that
An intuitionistic model structure is a triple (g, K, ≤), where K is a
set, ≤ is a partial order defined over K, and g is its minimum (so
g ≤ k for all k ∈ K).
As our enquirer investigates the truth of the various sentences of her proposi-
tional language, at any stage k a sentence A is either established to be true or not
[yet] established. We can symbolize those alternatives by k A and k 1 A; it is
quite common, for reasons that needn’t now detain us, to read ‘ ’ as forces. And,
as far as atomic sentences are concerned, the only constraint on a forcing relation
is this: once P is established in the knowledge state k, it stays established in any
expansion on that state of knowledge, i.e. at any k 0 such that k ≤ k 0 . Knowledge
persists. Hence, again to put the point more abstractly, we require get forcing
relation to satisfy this persistence condition:
For any atomic sentence P and k ∈ K, if k P , then k 0 P , for all k 0 ∈ K
such that k ≤ k 0 .
And now, next stage, let’s expand a forcing relation defined for a suite of atoms
so that it now covers all wffs built up from those atoms by the connectives. So,
for all k, k 0 ∈ K, and all relevant sentences A, B, we will require
94
 DRAFT– 31 DEC 2021
Overview: more proof theory, more semantics

(i) k 1 ⊥.
(ii) k A ∧ B iff k A and k B.
(iii) k A ∨ B iff k A or k B.
(iv) k A → B iff, for any k 0 such that k ≤ k 0 , if k 0 A then k 0 B.
(v) k ¬A iff, for any k 0 such that k ≤ k 0 , k 0 1 A.

It’s a simple consequence of these conditions on a forcing relation that for any
A, whether atomic or molecular,

(∗) If k A, then k 0 A, for all k 0 such that k ≤ k 0 .

This formally reflects the idea that once A is established it stays established,
whether or not it is an atom.
But what motivates those clauses (i) to (v) in our characterization of ? (i)
The absurd is never established as true, in any state of knowledge. And (ii)
establishing a conjunction is equivalent to establishing each conjunct, on any
sensible story. So we needn’t pause over these first two.
But (iii) reveals our enquirer’s intuitionist/constructivist commitments! – as
per the BHK interpretation, she is taking establishing a disjunction in an accept-
ably direct way to require establishing one of the disjuncts. For (iv) the thought
is that establishing A → B is tantamount to giving you an inference-ticket: with
the conditional established, if you (eventually) get to also establish A, then you
will then be entitled to B too. Finally, (v) falls out from the definition of ¬A as
A → ⊥ and the evaluation rules for → and ⊥. Or more directly, the idea is that
to establish ¬A is to rule out, once and for all, A turning out to be correct as
we later expand our knowledge.
With these pieces in place, we can – next stage! – define a formula of a
propositional language to be intuitionistically valid in a natural way. Classically,
a propositional formula is valid (is a tautology) if it is true however things
turn out with respect to the values of the relevant atoms. Now we say that a
propositional formula A is intuitionistically valid if it can be established in the
ground state of knowledge, however things later turn out with respect to the
truth of relevant atoms as our knowledge expands. Putting that more formally,

A is intuitionistically valid iff g A, whatever the model structure

(g, K, ≤) and whatever forcing relation is defined over the relevant
atoms.1

And now for the big reveal! Kripke proved in 1965 the following soundness and
completeness result:
1 Fine print, just to link up with other presentations you will meet. First, given (∗), g A
holds iff k A for all k. So we can redefine validity by saying A is valid just when k A
for all k. But then, second, we can in fact let g drop right out of the picture. For it is quite
easy to see that it will make no difference whether we require the partial order ≤ to have a
minimum or not: the same sentences will come out valid either way. Indeed, third, we don’t
even require the relation we symbolized ≤ to be a true partial order: again, if we allow any
reflexive, transitive relation over K in its place, it will make no different to what comes out
as valid.

95
 DRAFT– 31 DEC 2021
8 Intuitionistic logic

A formula is a theorem of IPL (can be derived from no premisses) if

and only if it is intuitionistically valid.

Neither direction of the biconditional is particularly hard.

Expanding the idea of valuations over an intuitionistic model structure to
accommodate quantified formulas and then proving soundness and completeness
for quantified intuitionistic logic is, however, rather more involved.
(d) Let’s finish by briefly showing that – given Kripke’s soundness result that
every IPL theorem is intuitionistically valid on his semantic story – it is imme-
diate that the law of excluded middle fails for IPL.
It couldn’t be easier. Consider a propositional language with just a single
atom P; and take the model structure which has just two states g, k such that
g ≤ k. And now suppose that P is not yet established at g but is established at
k, hence g 1 P while k P. By the rule for negation, g 1 ¬P. So g 1 (P ∨ ¬P).
Hence P ∨ ¬P is not valid. Hence, by the soundness result, P ∨ ¬P can’t be an
IPL theorem.

8.4 Basic recommendations on intuitionistic logic

So much for some quick introductory remarks – enough, I hope, to spark interest
in the topic! There is room, then, for a short introductory book which would
develop these and related themes at the kind of accessible level we currently want.
And Grigori Mints’s A Short Introduction to Intuitionistic Logic (Springer, 2000)
is indeed brief enough; however, it soon becomes entangled with more advanced
topics in a way that will too quickly mystify beginners. So we will have to patch
together readings from a few different sources.

We will cherry-pick from the following:

1. Joan Moschovakis, ‘Intuitionistic logic’, in The Stanford Encyclopaedia

of Philosophy, §§1–3, §4.1, §5.1. Available at tinyurl.com/sep-intuit.
2. Dirk van Dalen, Logic and Structure (Springer, 1980; 5th edition 2012),
Chapter 5, ‘Intuitionistic logic’.
3. A.S. Troelstra and Dirk van Dalen, Constructivism in Mathematics, An
Introduction: Vol. I (North-Holland, 1988), Chapter 2, ‘Logic’, §1, §3
(up to Prop 3.8), §4?, §5, §6?.

You could read these in the order given, initially skimming/skipping over
passages that aren’t immediately clear.
Or perhaps better, start with (1)’s §1, ‘Rejection of Tertium Non Datur’,
and then (2)’s §5.1, ‘Constructive reasoning’ which introduces the BHK
interpretation of the logical operators.
Then look at a presentation of a natural deduction system for intuition-
istic logic (as sketched in our overview): this is briskly covered in (2) in the

96
 DRAFT– 31 DEC 2021
Some parallel/additional reading

first half of §5.2. But in fact the discussion in (3) – though this is not an
introductory textbook – is notably more relaxed and clearer: see §1 of the
chapter.
Next, read up on the double-negation translation between classical and
intuitionistic logic. This is described in (1) §4.1, and explored a bit more in
the second half of (2) §5.2. But again, a more relaxed presentation can be
found in (3), §3 (up to Prop. 3.8).
Now you want to find out more about Kripke semantics, which is also
covered in all three resources. (1) §5.1 gives the brisk headline news. (2)
gives a compressed account in the first half of §5.3. But again (3) is best:
Troelstra and Van Dalen give a much more expansive and helpful account
in their Ch. 2 §5 – which sensibly treats propositional logic first before
expanding the story to cover full quantified intutionistic logic.
I would suggest, though, leaving detailed soundness and completeness
proofs for Kripke semantics – covered in (2) §5.3 or (3) §6 – for later (if
indeed tackled at all, at this stage.)
For a few more facts about intuitionistic logic, such as the disjunction
property, see also the first couple of pages of (2) §5.4 (the rest of that section
is interesting but not really needed at this stage).
Return to (1) to look at §2.1 (an axiomatic version of intuitionistic logic),
and the first half of §3 (on Heyting’s intuitionistic arithmetic). Then finally,
for more on Heyting Arithmetic and a spelt-out proof that it is consistent
if and only if classical Peano Arithmetic is consistent, you could dip into

4. Paolo Mancosu, Sergio Galvan, and Richard Zach, An Introduction to

Proof Theory, (OUP, 2021). Their §2.15 on ‘Intuitionistic and classical
arithmetic’ can be read as an approachable stand-alone treatment.

8.5 Some parallel/additional reading

Kripke semantics for intuitionistic logic involves evaluating formulas not once
and for all but at different points in a relational structure. We informally talked
about these points as various ‘states of knowledge’; in a different idiom we could
have talked about various ‘possible worlds’. Now, the use of this kind of relational
semantics is characteristic of modal logics – the simplest modal logics being
logics of necessity and possibility, with their semantics modelling the idea that
being necessarily true is being true at all suitably related possible worlds. So
another way of approaching intuitionistic logic is by first discussing modal logics
more generally, before looking at intuitionistic logic in particular. If you want to
explore this route, you can jump to this Guide’s Chapter 10. In particular, you
could perhaps look at Graham Priest’s terrific An Introduction to Non-Classical
Logic mentioned there, which gives tableaux systems first for modal logic and
then for intuitionistic logic.
There is also a different way using tableaux for intuitionistic logic (which
doesn’t rely on first treating modal logic), which is quite nicely explored by
97
 DRAFT– 31 DEC 2021
8 Intuitionistic logic

5. Harrie de Swart, Philosophical and Mathematical Logic (Springer, 2018),

Chapter 18.

However, I prefer the treatment of the same tableau approach in an earlier

excellent book:

6. Melvin Fitting, Intuitionistic Logic, Model Theory, and Forcing (North

Holland, 1969), Part I.
Ignore the scary title: it is only the beautifully clear but sophisticated
first part of the book which concerns us now! It should particularly appeal
to those who appreciate mathematical elegance.

For a bit more on natural deduction, the sequent calculus and semantics for
intuitionistic logic, you should look at two chapters from a modern classic:

7. Michael Dummett, Elements of Intuitionism (OUP, 2nd ed. 2000), Chap-

ters 4 and 5.

In fact, you could well want to read the opening two chapters and the final
one as well! There are then many more pointers to technical discussions in
Moschovakis’s section of ‘Recommended reading’.

8.6 A little more history, a little more philosophy

A number of the readings mentioned so far include brief remarks about the
history of intuitionism (and constructivism more generally). For something more
substantial, look at

8. A.S. Troelstra and Dirk van Dalen, Constructivism in Mathematics, An

Introduction: Vol. I (North-Holland, 1988), Chapter 1,

which gives a brief characterization of various forms of constructivism (not all

of them motivate the adoption of a non-classical logic like intuitionistic logic).
The early days of intuitionism were wild! To get a sense of how wild Brouwer’s
ideas were, you could take a look at

9. Mark van Atten, On Brouwer (Wadsworth, 2004), Chapters 1 and 2.

The same author has a The Stanford Encyclopedia article on ‘The Development
of Intuitionistic Logic’ at tinyurl.com/dev-intuit; but that’s much more detailed
than you are likely to want.
Turning to more philosophical discussions – and it is a bit difficult to separate
thinking about intuitionism as a philosophy of mathematics from thinking about
intuitionistic logic more specifically – one key article that you will want to read
(which was hugely influential in reviving interest in a ‘tamer’ intuitionism among
philosophers) is

10. Michael Dummett, ‘The philosophical basis of intuitionistic logic’ (origi-

nally 1973, reprinted in Dummett’s Truth and Other Engimas).
98
 DRAFT– 31 DEC 2021
A little more history, a little more philosophy

Then, for more recent discussions, here’s a trio of articles:

11. Carl Posy, ‘Intuitionism and philosophy’; D. C. McCarty, ‘Intuitionism in

mathematics’; and Roy Cook, ‘Intuitionism reconsidered’, all in S. Shapiro,
ed., The Oxford Handbook of the Philosophy of Mathematics and Logic
(OUP, 2005).

99
 DRAFT– 31 DEC 2021

9 Elementary proof theory

The story of proof theory starts with David Hilbert and what has come to be
known as ‘Hilbert’s Programme’, which inspired the profoundly original work of
Gerhard Gentzen in the 1930s.
Two themes from Gentzen are within easy reach for beginners in mathemat-
ical logic: (A) the idea of normalization for natural deduction proofs, (B) the
move from natural deduction to sequent calculi, and cut-elimination results for
these calculi. But the most interesting later developments in proof theory – in
particular, in so-called ordinal proof theory – quickly become mathematically
rather sophisticated. Still, at this stage it is at least worth making a first pass
at (C) Gentzen’s proof of the consistency of arithmetic using a cut-elimination
proof which invokes induction over some small countable ordinals. So these three
themes from elementary proof theory will be the focus of this chapter.

9.1 Preamble: a very little about Hilbert’s Programme

Set theory, for example, is about – or at least, is supposed to be about – an
extraordinarily rich domain of (mostly) infinite objects. How can we know that
such a theory really does make good sense? Indeed, how can we know that it
even gets to the starting line of being internally consistent?
David Hilbert had a wonderful insight. While the topic of a mathematical
theory T such as set theory might be wildly infinitary, the theory T itself is
built from thoroughly finite objects – namely sentences, and the finite arrays of
sentences that are proofs. So perhaps we can use some very tame assumptions
(assumptions that don’t tangle with the infinite) to reason about T when it is
thought of as a suite of finite objects. And in particular, perhaps we can use
tame assumptions to prove T ’s consistency, without needing to worry about T ’s
purported infinitary subject matter.
To make any progress with this idea, we’ll need to fully pin down T ’s basic
assumptions and to regiment the principles of reasoning that T can deploy –
we’ll need, in other words, to have a nice axiomatic formalization of T on the
table. This formalization of the theory T (whether it’s about sets, widgets, or
whatnots) then gives us some definite, mathematically precise, new objects to
reason about (beyond the sets, widgets, or whatnots), namely the T -wffs and
T -proofs that make up the theory. And now, as Hilbert saw, we can set off to

100
 DRAFT– 31 DEC 2021
Deductive systems, normal forms, and cuts: a short overview

mathematically investigate these, developing a Beweistheorie (a theory about

proofs).
We’ll return in §9.3 to say something more about the resulting Programme
of aiming to use entirely ‘safe’, merely finitary, reasoning about a theory T in
order to prove its consistency (though you should already know that Gödel’s
Second Incompleteness Theorem is going to cause some trouble). But, for the
moment, the point we want is simply this: the Programme presupposes that we
can indeed regiment the theory that concerns us into a tidily disciplined formal
shape – and in particular, we can regiment its required principles of reasoning
into a formal deductive logic. Hence the central importance for Hilbert and his
associates of constructing suitable formal systems for logic.

9.2 Deductive systems, normal forms, and cuts: a short overview

(a) The logical systems developed by Hilbert and Bernays1 were axiomatic in
style, and at some remove from the forms of deduction used in practice in mathe-
matical proofs. It was Bernays’ student Gerhard Gentzen who first introduced a
style of deductive system which explicitly aimed to come, as he put it, “as close
as possible to actual reasoning.” The result was Gentzen’s natural deduction
calculi for intuitionistic and classical predicate logic.
Now, these calculi – which I’ll take to be familiar from work on earlier topics
in this Guide – have some lovely features: and as advertised, they do allow
us to formally track natural lines of reasoning. But they also still allow us to
construct some perversely unnatural proofs! For example, consider the following
two derivations to show that from P ∧ Q we can infer P ∨ Q:

P∧Q [R ∧ Q](1)
P∧Q
P [P](1) Q
(i) P (ii)
P∨Q P ∨ (R ∧ Q) P∨Q P∨Q
(1)
P∨Q

(i) is an entirely natural mini-proof. But (ii) takes us on a pointless detour: on

the leftmost branch, the ‘wrong’ disjunction is introduced on the left branch
which involves the quite irrelevant R, before we use a disjunction-elimination
inference at (1) to finally get the proof back on track.
The detour in (ii) is not just inelegant; there is also a sense in which it makes
the proof non-explanatory. After all, if a premiss A logically entails a conclusion
C, this – we suppose – results from the conceptual content of A and C. So we
want a proof to explain how the contents of A and C generate the entailment.
A derivation like (ii), which introduces irrelevant content that is quite unrelated
to either the premiss or conclusion, can’t do that.
So, generalizing on the example of (ii), let’s now define a detour as consist-
ing in the use of the introduction rule for a logical operator (a connective or a
1 PaulBernays was nominally Hilbert’s assistant, but in fact was an absolutely key figure in
his own right.

101
 DRAFT– 31 DEC 2021
9 Elementary proof theory

quantifier) followed by the application of the corresponding elimination rule to

this introduced operator. Then, as just noted, it is not merely to avoid inelegan-
cies that we will want detour-free proofs.
Now, simple detours in a Gentzen-style natural deduction proof can easily be
removed. For example, a detour which involves introducing a conditional (by
conditional proof) and then eliminating it (by modus ponens), as on the left,
can be simply smoothed away or reduced, as on the right:

[A](1) ..
o .
.. A
. B (1) o
A A→B
B B

For another example, going back to the case of introducing and then eliminating
a disjunction, a proof of the shape on the left can be reduced to a proof with
the shape on the right:
.. [A](1) [B](1) ..
. .
A o oo A
A∨B C C (1) o
C C
And similarly for other simple detours involving other connectives and the quan-
tifiers. However, what about the case where a detour gets entangled with the
application of other rules in more complicated ways? Can detours always be
removed?
Gentzen was able to show that – at least for his system of intuitionistic logic
– if a conclusion can be derived from premisses at all, then there will indeed be
a normal, i.e. detour-free, proof of the conclusion from the premisses. And he
did this by giving a normalization procedure – i.e. instructions for systematically
removing detours until we are left with a normal proof. The resulting detour-free
proofs will then have particularly nice features such as the so-called subformula
property: every formula that occurs in a proof will either be a subformula of one
of the premisses or a subformula of the conclusion (as usual, counting instances
of quantified wffs as subformulas of them). There won’t be irrelevancies as in
our silly proof (ii) above.
And now note that, as a corollary, we can immediately conclude that intu-
itionistic logic is consistent: we can’t have a proof with the subformula property
from no premisses to ⊥. Which raises a hopeful prospect: can other normal-
ization proofs be used to establish the sort of consistency results that Hilbert
wanted?
(b) But now the story gets complicated. For a start, Gentzen himself couldn’t
find a normalization proof for his natural deduction system of classical logic
(you can see why there might be a problem – a classical proof might, at least
on the face of it, need to rely on an instance of excluded middle which isn’t a

102
 DRAFT– 31 DEC 2021
Deductive systems, normal forms, and cuts: a short overview

subformula of either the premisses or the conclusion). In order to get a classical

system for which he could prove an appropriate normalization theorem, Gentzen
therefore introduced his sequent calculi, about which more in moment. And
his normalization proof for intuitionistic logic then remained unpublished for
seventy years. In the meantime, the proof was independently rediscovered by
Dag Prawitz in his thesis, published as Natural Deduction (1965), which also
presents a normalization proof for Gentzen’s classical natural deduction system
without ∨ and ∃ (which is of course equivalent to the complete system).
Since Prawitz’s work brought Gentzen-style natural deduction back to cen-
tre stage, there has been a whole cottage industry of tinkering with the in-
ference rules, and tinkering with the definition of a normal proof, in order to
produce classical natural deduction systems with nice proof-theoretic features.
But I rather think that the typical beginner in mathematical logic won’t find the
details of these further developments particularly exciting. However, it is well
worth looking at the opening four chapters of Prawitz’s wonderful short book,
and perhaps just noting a few more ideas. This will be enough on our theme
(A), natural deduction and normalization.
(c) How do we read off what depends on what in a natural deduction proof?
By looking at the geometry of the proof, and its annotations.
For example, consider this derivation of P → (Q → R) from (P ∧ Q) → R:

[P](2) [Q](1)
(P ∧ Q) → R P∧Q
R (1)
Q→R
(2)
P → (Q → R)

Then, reading upwards from R, we see that this wff depends on all three of
(P ∧ Q) → R, P, and Q as assumptions (for neither of the last two have yet been
discharged); while Q → R on the next line depends only on (P ∧ Q) → R and P.
That’s clear enough. But we could alternatively record dependencies quite ex-
plicitly, line by line. To do this, we will make use of so-called sequents. We’ll write
a sequent in the form Γ ⇒ A, and read this as saying that A is deducible from
the finitely many (perhaps zero) wffs Γ.2 Since an (undischarged) assumption
depends just on itself, we can then explicitly record the deducibilities revealed
in our last natural deduction proof like this (check that claim!):

P ⇒ P Q ⇒ Q
(P ∧ Q) → R ⇒ (P ∧ Q) → R P, Q ⇒ P ∧ Q
(P ∧ Q) → R, P, Q ⇒ R
(P ∧ Q) → R, P ⇒ Q → R
(P ∧ Q) → R ⇒ P → (Q → R)
2 Forpresent purposes, we can officially think of Γ as given as a set – though in the end we
might prefer to treat Γ as a multi-set where repetitions matter: Gentzen himself treated Γ
as an ordered sequence.

103
 DRAFT– 31 DEC 2021
9 Elementary proof theory

And now, following Gentzen, instead of thinking of this tree of sequents as in

effect just a running commentary on an underlying natural deduction proof, we
can treat it as itself a new sort of proof in its own right – a proof relating whole
sequents rather than individual wffs.
At the tips of branches of this sequent proof about deducibilities we have
‘axioms’ of the form A ⇒ A (since trivially, A is deducible, given A!). And then
the proof is extended downwards by the application of two sorts of rules, rules
governing specific logical operators, and general structural rules.
For the logical rules, we could replace the familiar natural deduction rules for
wffs with corresponding rules for deriving sequents, as in these examples:
A B Γ ⇒ A ∆ ⇒ B
A∧B Γ, ∆ ⇒ A ∧ B
[A]
..
. Γ, A ⇒ B
B Γ ⇒ A→B
A→B
There should be nothing mysterious here. After all, the terse schematic presen-
tation of the natural-deduction introduction rule for ∧ is to be read as saying
that if we have A (deduced perhaps from some other assumptions) and have B
(again perhaps deduced from some other assumptions), we can infer A ∧ B (with
those earlier assumptions all remaining in play). And that’s what the suggested
sequent calculus rule now explicitly says too. Likewise, the natural-deduction
introduction rule for → is to be read as saying that if we derive B from the as-
sumption A (and perhaps from some other assumptions), then we can drop that
assumption A and infer A → B (with those other assumptions kept in play);
and that’s what the sequent calculus rule says too. There will be similar rules
for other connectives and for quantifiers.
As for structural rules, we will mention here two candidates. The first is tradi-
tionally called thinning or weakening (neither of which is perhaps a very helpful
label). The simple idea is that, if a wff is deducible from some assumptions, it
remains deducible if we add in a further unnecessary assumption. So
Γ ⇒ C
Γ, A ⇒ C
Our second structural rule for sequent proofs corresponds to the structural fact
that we can chain natural deduction proofs together into longer proofs. Thus,
schematically, in natural deduction,
Γ.
A ..
Γ. | {z ∆}
We can splice a proof .. with a proof .. to get A
| {z ∆} .
. ..
A B .
B
In sequent calculus terms this corresponds to the following cut rule:

104
 DRAFT– 31 DEC 2021
Deductive systems, normal forms, and cuts: a short overview
Γ ⇒ A ∆, A ⇒ B
Γ, ∆ ⇒ B
This intuitively sound rule allows us to cut out the middle man A.
So far, then, so good – though of course, we’ve left lots of detail to be filled out.
And there is as yet nothing really novel involved in reworking natural deduction
into sequent style like this. But now, however, Gentzen introduces two very
striking new ideas.
(d) To introduce the first idea, let’s think again about the elimination rules for
conjunction. As a first shot, we might expect to transform the pair of natural-
deduction rules into a corresponding pair of sequent-calculus rules like this:
A∧B A∧B Γ ⇒ A∧B Γ ⇒ A∧B
A B Γ ⇒ A Γ ⇒ B
What could be more obvious? But we could alternatively adopt the following
sequent-calculus rule:
Γ, A, B ⇒ C
Γ, A ∧ B ⇒ C
This is obviously valid – if C can be derived from some assumptions Γ plus A
and B, it can obviously be derived from Γ plus the conjunction of A and B. And
we can use this rule introducing ∧ on the left of the sequent sign instead of the
expected pair of rules eliminating ∧ to the right of the sequent sign. For note,
given the new rule, we can restore the first of the elimination rules as a derived
rule, because we can always give a derivation of this shape:
A ⇒ A (Weakening)
A, B ⇒ A
(New rule for ∧)
Γ ⇒ A∧B A ∧ B ⇒ A (Cut)
Γ ⇒ A
Similarly, of course, for the companion elimination rule.
And now the point generalizes. As Gentzen saw, in a sequent calculus for intu-
itionistic logic, we can get all the rules for handling connectives and quantifiers
to introduce a logical operator – either on the right of the sequent sign (corre-
sponding to a natural-deduction introduction rule) or on the left of the sequent
sign (corresponding to a natural-deduction elimination rule).
(e) We can go further. Still working with a sequent calculus for ⇒ read as in-
tuitionistic deducibility, we can in fact eliminate the cut rule. Anything provable
using cut can be proved without it.
This might initially seem pretty surprising. After all, didn’t we just have to
appeal to the cut rule to show that – using our new introduction-on-the-left rule
for ∧ – we can still argue from (1) Γ ⇒ A ∧ B to (2) Γ ⇒ A? How can we
possibly do without cut in this case?
Well, consider how we might actually have arrived at (1). Perhaps it was by
the rule for introducing occurrences of ∧ on the right of a sequent. So perhaps,
to expose more of the proof from (1) to (2), it has the shape of the first proof
below (supposing Γ to result from putting together Γ0 and Γ00 ):
105
 DRAFT– 31 DEC 2021
9 Elementary proof theory
A ⇒ A
Γ0 ⇒ A Γ00 ⇒ B A, B ⇒ A Γ0 ⇒ A (Weakenings)
Γ ⇒ A∧B A∧B ⇒ A (Cut)
Γ ⇒ A
Γ ⇒ A
But if we already have Γ0 ⇒ A, as in the first proof, then we don’t need to
go round the houses on that detour, introducing an occurrence of ∧ to get the
formula A ∧ B, and then cutting out that same formula: we can just get from
Γ0 ⇒ A to Γ ⇒ A by some weakenings (by adding in the wffs from Γ00 ). Here,
then, eliminating the cut is just like normalizing (part of) a natural deduction
proof.
OK: that only shows that in just one rather special sort of case, we can
eliminate a cut. Still, it’s a hopeful start! And in fact, we can always eventually
eliminate cuts from an intuitionistic sequent calculus proof.
But the process can be intricate. For example, take a slight variant of our
previous example and suppose we want to eliminate the following cut (remember,
combining Γ and Γ gives us Γ!):
Γ ⇒ A Γ ⇒ B ∆, A, B ⇒ C
Γ ⇒ A∧B ∆, A ∧ B ⇒ C
(Cut)
Γ, ∆ ⇒ C
Then we can replace this proof-segment with the following:
Γ ⇒ B ∆, A, B ⇒ C
(Cut)
Γ ⇒ A Γ, ∆, A ⇒ C
(Cut)
Γ, ∆ ⇒ C
Again, as in normalizing a natural deduction proof, we have removed a detour
– this time a detour through introducing-∧-on-the-right and introducing-∧-on-
the-left. So we have now lost the cut on the more complex formula A ∧ B, albeit
replacing it with two new cuts. But still, the new cuts are on the simpler formulas
A and B, and we have also pushed one of the cuts higher up the proof. And that’s
typical: looking at the range of possible situations where we can apply the cut
rule – a decidedly tedious hack though all the cases – we find we can indeed keep
reducing the complexity of formulas in cuts and/or pushing cuts up the proof
until all the cuts are completely eliminated.
(f) So we arrive at this result. In a sequent-calculus setting, we can use a cut-free
deductive system for intuitionistic logic where all the rules for the connectives
and quantifiers introduce logical operators, either to the left or to the right of the
sequent sign. Analogously to a normalized natural-deduction proof, there are no
detours. As we go down a branch of the proof, the sequents at each stage are
steadily more complex (we can make the relevant notion of complexity precise
in pretty obvious ways).
This proof-analysis immediately delivers some very nice results.
(i) The subformula property: every formula occurring the derivation of a se-
quent Γ ⇒ C is a subformula of either one of formulas Γ or of C. (By
inspection of the rules!)
106
 DRAFT– 31 DEC 2021
Deductive systems, normal forms, and cuts: a short overview

(ii) There evidently can be no cut-free, ever-more-complex, derivation that

ends with ⇒ ⊥; in other words, absurdity isn’t intuitionistically deducible
from no premisses. Hence intuitionistic logic is internally consistent.

(iii) Equally evidently, the penultimate line of a cut-free, ever-more-complex,

derivation of ⇒ A∨B has to be either ⇒ A or ⇒ B, which establishes
the disjunction property for intuitionistic logic – see §8.3(a).

Note too that, at least for propositional logic, we can take any sequent and
systematically try to work upwards from it to construct a cut-free proof with
ever-simpler-sequents: the resulting success or failure then mechanically decides
whether the sequent is intuitionistically valid.
(g) I said that Gentzen had two very striking new ideas in developing his se-
quent calculi beyond a mere re-write of a natural deduction system in which
dependencies are made explicit. The first idea is to recast all the rules for logical
operators as rules for introducing logical operators, now allowing introduction
to the left as well as introduction to the right of the sequent sign, and to then
show that we can get a cut-free proof (hence, a proof that always goes from less
complex to more complex sequents) for any intuitionistically correct sequent.
But this first idea doesn’t by itself resolve the problem which Gentzen initially
faced. Recall, he ran into trouble trying to find a normalization proof for classical
natural deduction. And plainly, if we stick with a cut-free all-introduction-rules
sequent calculus of the current style, we can’t get a classical logical system at
all. The point is trivial: one key additional classical principle we need to add to
intuitionistic logic is the double negation rule. We need to be able to show, in
other words, that from Γ ⇒ ¬¬A we can derive Γ ⇒ A. But obviously we can’t
do that in a system where we can only move from logically simpler to logically
more complex sequents!
What to do? Well, at this point Gentzen’s second (and quite original) idea
comes into play. We now liberalize the notion of a sequent. Previously, we took
a sequent Γ ⇒ A to relate zero or more wffs on the left to a single wff on the
right. Now we pluralize on both sides of the sequent sign, writing Γ ⇒ ∆; and
we read that as saying that at least one of ∆ is deducible from the wffs Γ. If you
like, you can regard ∆ as delimiting the field within which the truth must lie if
the premisses Γ are granted. (We’ll continue, for our purposes, to treat Γ and ∆
officially as sets, rather than multisets or lists: note that we will allow either or
both to be empty.)
Keeping the idea that we want all our rules for the logical operators to be
rules for introducing operators to the left or right of the sequent sign, how might
these rules now go? There are various options, but the following can work nicely
for conjunction and disjunction:
Γ, A, B ⇒ ∆ Γ ⇒ ∆, A Γ ⇒ ∆, B
(∧L) (∧R)
Γ, A ∧ B ⇒ ∆ Γ ⇒ ∆, A ∧ B
Γ, A ⇒ ∆ Γ, B ⇒ ∆ Γ ⇒ ∆, A, B
(∨L) (∨R)
Γ, A ∨ B ⇒ ∆ Γ ⇒ ∆, A ∨ B
107
 DRAFT– 31 DEC 2021
9 Elementary proof theory

I won’t give the rules for all the other logical operators here, but let’s note the
left and right rules for negation (these can either be built-in rules, if negation is
treated as a primitive built-in connective, or derived rules, if negation is defined
in terms of the conditional and absurdity):
Γ ⇒ ∆, A Γ, A ⇒ ∆
(¬L) (¬R)
Γ, ¬A ⇒ ∆ Γ ⇒ ∆, ¬A
These rules are evidently correct on the classical understanding of the connec-
tives. For the first rule, suppose that given the assumptions Γ, then (at least)
one of ∆ and A follows: then given the same assumptions Γ but now also ruling
out A, we can conclude that (at least) one of ∆ is true. We can argue similarly
for the second rule. But with these negation and disjunction rules in place we
immediately have the following derivation:
A ⇒ A (¬R)
⇒ A, ¬A
(∨R)
⇒ A ∨ ¬A
Out pops the law of excluded middle! – so we know we are dealing with classical
calculus.
(h) What about the structural rules for our classical sequent calculus which
allows multiple alternative conclusions as well as multiple premisses? We can
now allow weakening on both sides of a sequent. And we can generalize the cut
rule to take this form:
Γ ⇒ ∆, A Γ 0 , A ⇒ ∆0
Γ, Γ0 ⇒ ∆, ∆0
(Think why this is a sound rule, given our interpretation of the sequents!) But
then, just as with our sequent calculus for intuitionistic logic, we can proceed to
prove that we can eliminate cuts. If a sequent is derivable in our classical sequent
calculus, it is derivable without using the cut rule.
And as with intuitionist logic, this immediately gives us some nice results. Of
course, we won’t have the disjunction property (think excluded middle!). But we
still have the subformula property in the form that if Γ ⇒ ∆ is derivable, the
every formula in the sequent proof is a subformula of one of Γ, ∆. And again,
simply but crucially, ⇒ ⊥ won’t be derivable in the cut-free classical system,
so it is consistent.
And that’s perhaps enough by way of introduction to our theme (B), in which
we begin to explore various elegant sequent calculi, prove cut-elimination theo-
rems, and draw out their implications.

9.3 Proof theory and the consistency of arithmetic: a short overview

Now for our third theme (C), Gentzen’s famed proof of the consistency of arith-
metic (more precisely, the consistency of first-order Peano Arithmetic). Recall,
Hilbert’s Programme is the project of using tame proof-theoretic reasoning to
prove the consistency of mathematical theories: PA gives us a first test case.
108
 DRAFT– 31 DEC 2021
Proof theory and the consistency of arithmetic: a short overview

(a) You might very well wonder whether there can be any illuminating and
informative ways of proving PA to be consistent. After all, proving consistency
by appealing to a stronger theory like ZF set theory which in effect contains PA
won’t be a very helpful (for doubts about the consistency of PA will presumably
just carry over to become doubts about the stronger theory). And you already
know that Gödel’s Second Incompleteness Theorem shows that it is impossible
to prove PA’s consistency by appealing to a weaker theory tame enough to be
modelled inside PA (not even full PA can prove PA’s consistency).
However, another possibility does remain open. It isn’t ruled out that we can
prove PA’s consistency by appeal to an attractive theory which is weaker than
PA in some respects but stronger in others. And this is what Gentzen aims to
give us in his consistency proof for arithmetic.3
(b) Here then is an outline sketch of the key proof idea, in Gentzen’s own
words.
We start with a formulation of PA using for its logic a classical sequent calculus
including the cut rule. (We will initially want the cut rule in making use of PA’s
axioms, and we can’t assume straight off the bat that we can still eliminate cuts
once we have more complex proofs appealing to non-logical axioms). Then,

The ‘correctness’ of a proof depends on the correctness of certain

other simpler proofs contained in it as special cases or constituent
parts. This fact motivates the arrangement of proofs in linear order
in such a way that those proofs on whose correctness the correctness
of another proof depends precede the latter proof in the sequence.
This arrangement of the proofs is brought about by correlating with
each proof a certain transfinite ordinal number.

The idea, then, is that the various sequent proof-trees in this version of PA can
be put into an ordering by a kind of dependency relation, with more complex
proof trees (on a suitable measure of complexity) coming after simpler proofs.
And this can be a well-ordering, so that the position along the ordering can
indeed be tallied by an ordinal number.
But why is the relevant linear ordering of proofs said to be transfinite (in other
words, why must it allow an item in the ordering to have an infinite number of
predecessors)? Because

[it] may happen that the correctness of a proof depends on the cor-
rectness of infinitely many simpler proofs. An example: Suppose that
in the proof a proposition is proved for all natural numbers by com-
plete induction. In that case the correctness of the proof obviously
depends on the correctness of every single one of the infinitely many
individual proofs obtained by specializing to a particular natural
3 Gentzen in fact gives four different proofs, developed along somewhat different lines. But
the master idea underlying the best known of the proofs is given in a wonderfully clear
way in his wide-ranging lecture on ‘The concept of infinity in mathematics’ reprinted in his
Collected Papers, from which the following quotations come.

109
 DRAFT– 31 DEC 2021
9 Elementary proof theory

number. Here a natural number is insufficient as an ordinal number

for the proof, since each natural number is preceded by only finitely
many other numbers in the natural ordering. We therefore need the
transfinite ordinal numbers in order to represent the natural ordering
of the proofs according to their complexity.

Think of it this way: a proof by induction of the quantified ∀xA(x) leaps beyond
all the proofs of A(0), A(1), A(2), . . . . And the result ∀xA(x) depends for its
correctness on the correctness of the simpler results. So, in the sort of ordering
of proofs which Gentzen has in mind, the proof by induction of ∀xA(x) must
come infinitely far down the list, after all the proofs of the various A(n).
And now Gentzen’s key step is to argue by an induction along this transfinite
ordering of proofs. The very simplest proofs right at the beginning of the ordering
transparently can’t lead to contradiction. Then

once the correctness [and specifically, freedom from contradiction]

of all proofs preceding a particular proof in the sequence has been
established, the proof in question is also correct precisely because
the ordering was chosen in such a way that the correctness of a proof
depends on the correctness of certain earlier proofs. From this we
can now obviously infer the correctness of all proofs by means of
a transfinite induction, and we have thus proved, in particular, the
desired consistency.

Transfinite induction here is just the principle that, if we can show that a proof
has a property P if all its predecessors in the relevant ordering have P , then all
proofs in the ordering have property P .
(c) We can implement this same proof idea the other way around. We show
that if any proof does lead to contradiction, then there must be an earlier proof
in the linear ordering of proofs which also leads to contradiction – so we get
an infinite sequences of proofs of contradiction, ever earlier in the ordering. But
then the ordinals which tally these proofs of contradiction would have to form
an infinite descending sequence. And there can’t be such a sequence of ordinals.
Hence no proof leads to contradiction and PA is consistent.
(d) Two questions arising. First, how do we show that if a proof leads to a
contradiction, then there must be an earlier proof in the linear ordering of proofs
which leads to contradiction? By eliminating cuts using reduction procedures like
those involved in the proof of cut-elimination for a pure logical sequent calculus
– so here’s the key point of contact with ideas we meet in tackling theme (B).
And second, what kind of transfinite ordering is involved here? Gentzen’s
ordering of possible proof-trees in his sequent calculus for PA turns out to have
the order type of the ordinals less than ε0 (what does that mean? – the references
will explain, but these are all the ordinals which are sums of powers of ω). So,
what Gentzen’s proof needs is the assumption that a relatively modest amount
of transfinite induction – induction up to ε0 – is legitimate.

110
 DRAFT– 31 DEC 2021
Proof theory and the consistency of arithmetic: a short overview

Now, the PA proof-trees which we are ordering are themselves all finite ob-
jects; we can code them up using Gödel numbers in the familiar sort of way.
So in ordering the proofs, we are in effect thinking about a whacky ordering of
(ordinary, finite) code numbers. And whether one number precedes another in
the whacky ordering is nothing mysterious; a computation without open-ended
searches can settle the matter.
So what resources does a Gentzen-style argument use, if we want to code it up
and formalize it? The assignment of a place in the ordering to a proof can be han-
dled by primitive recursive functions, and facts about the dependency relations
between proofs at different points in the ordering can be handled by primitive
recursive functions too. A theory in which we can run a formalized version of
Gentzen’s proof will therefore be one in which we can (a) handle primitive recur-
sive functions and (b) handle transfinite induction up to ε0 , maybe via coding
tricks. It turns out to be enough to have all p.r. functions available, together with
a formal version of transfinite induction just for simple quantifier-free wffs con-
taining expressions for these p.r. functions. Such a theory is neither contained in
PA (since it can prove PA’s consistency by formalizing Gentzen’s method, which
PA can’t), nor does it contain PA (since it needn’t be able to prove instances of
the ordinary Induction Schema for arbitrarily complex wffs).
So, in this sense, we can indeed prove the consistency of PA by using a theory
which is weaker than PA in some respects while stronger in others.
(e) Of course, it is a very moot point whether – if you were really worried about
the consistency of PA – a Gentzen-style proof when fully spelt out would help
resolve your doubts. Are the resources it requires ‘tame’ enough to satisfy you?
Well, if you are globally worried about the use of induction in general, then
appealing to an argument which deploys an induction principle won’t help! But
global worries about induction are difficult to motivate, and perhaps your worry
is more specifically that induction over arbitrarily complex wffs might engen-
der trouble. You note that PA’s induction principle applies, inter alia, to wffs
that themselves quantify over all numbers. And you might worry that if (like
Frege) you understand the natural numbers to be what induction applies to, then
there’s a looming circularity here – numbers are understood as what induction
applies to, but understanding some cases of induction involves understanding
quantifying over numbers. If that is your worry, the fact that we can show that
PA is consistent using an induction principle which is only applied to quantifier-
free wffs (even though the induction runs over a novel ordering on the numbers)
could soothe your worries.
Be that as it may: we can’t pursue that kind of philosophical discussion any
further here. The point remains that the Gentzen proof is a fascinating achieve-
ment, containing the seeds of wonderful modern work in proof theory. Perhaps
we haven’t quite executed Hilbert’s Programme of proving consistency by appeal
to entirely tame proof-theoretic reasoning. But in the attempt, we have found
how far along the ordinals we need to run our transfinite induction in order to

111
 DRAFT– 31 DEC 2021
9 Elementary proof theory

prove the consistency of PA.4 And we can now set out to discover how much
transfinite induction is required to prove the consistency of other theories. But
the achievements of that kind of ordinal proof theory will have to be left for you
(eventually) to explore . . .

9.4 Main recommendations on elementary proof theory

Let’s start with a couple of very useful encyclopaedia entries by some notable
proof theorists.

First, the following exemplary historical outline is particularly helpful for

orientation:

1. Jan von Plato, ‘The development of proof theory’, The Stanford Ency-
clopedia of Philosophy. Available at tinyurl.com/sep-devproof.

And then look at the first half of the main entry on proof theory:

2. Michael Rathjen and Wilfrid Sieg, ‘Proof theory’, §§1–3, The Stanford
Encyclopedia of Philosophy. Available at tinyurl.com/sep-prooftheory.

Skip over any passages that are initially unclear, and return to them when
you’ve worked through some of the readings below.

In keeping with our overviews in the previous two sections, I suggest that – in
a first encounter with proof theory – you focus on (A) normalization for natural
deduction and its implications; (B) the sequent calculus, cut-elimination and
its implications; and (C) a Gentzen-style proof of the consistency of arithmetic.
Now, there is book which aims to cover just these topics at the level we want:

3. Paolo Mancosu, Sergio Galvan and Richard Zach, An Introduction to

Proof Theory: Normalization, Cut-Elimination and Consistency Proofs
(OUP, 2021) – henceforth IPL.

However, as the authors say in their Preface, “in order to make the content acces-
sible to readers without much mathematical background, we carry out the details
of proofs in much more detail than is usually done.” And this isn’t anywhere
near as reader-friendly as they intend: expositions too often become wearyingly
laborious. Also the authors stick very closely to Gentzen’s own original papers,
which isn’t always the wisest choice. So, at least on topic areas (A) and (B), I
will be highlighting some alternatives.
(A) You could find that the following Handbook of the History of Logic article
gives some more helpful orientation:

4 Technical remark. There are no worries about using transfinite induction up to any ordinal
less than ε0 ; for this can be handled inside PA. So Gentzen’s proof calls on the least possible
extension to the amount of induction that can be handled inside PA!

112
 DRAFT– 31 DEC 2021
Main recommendations on elementary proof theory

4. F. J. Pelletier and Allen Hazen, ‘Natural deduction’, §3. Available at

tinyurl.com/pellhazen.
It is §3.1 that is most immediately relevant. But do read the rest of
§3. (And indeed, for your general logical education, why not read all this
informative survey paper sometime?)

You could next tackle Chs 3 and 4 of IPL. But there’s a lot to be said for just
diving into the brisk opening chapters of a modern classic:

5. Dag Prawitz, Natural Deduction: A Proof-Theoretic Study* (originally

published 1965, reprinted by Dover Publications 2006), Chapters I to IV.
Ch. I presents the now-standard Gentzen-style natural deduction sys-
tems for intuitionistic and classical logic. The short Ch. II explains the
sense in which elimination rules are inverses to introduction rules. Then
it notes some basic “reduction steps” for eliminating the sort of unnec-
essary detours which result from the application of an introduction rule
being immediately followed by the application of the corresponding elim-
ination rule. Ch. III shows that we can normalize proofs in a classical
ND system – or at least, a cut down version without ∨ and ∃ built in
as primitive – by systematically eliminating detours. Ch. IV extends the
result to a full system of intuitionistic logic.
And that’s perhaps about as much as you need on natural deduction. OK,
you might be left wondering whether we can improve on Prawitz’s Chapter
III result and prove a similar normalization result for a full classical logic
with the ∨ and ∃ rules restored. The answer is ‘yes’. IPL §4.9 shows how it
can be done for Gentzen’s original natural deduction system. But it is more
interesting to look at what happens if you revise Gentzen’s original classical
rules and use so-called ‘general elimination rules’; this makes establishing
normalization rather more straightforward. For something on this, see

6. Jan von Plato, Elements of Logical Reasoning (CUP, 2013). Chapters 3

to 6.

These very accessible chapters on intuitionistic and classical propositional

logic also introduce the theme of proof-search.

Von Plato’s book is, in fact, intended as a first introductory logic text, based
on natural deduction: but it, very unusually, has a strongly proof-theoretic em-
phasis. And non-mathematicians, in particular, could find the whole book very
helpful.
(B) Next, moving on to sequent calculi, you could start with Chs 5 and 6 of
IPL. But the following is very accessibly written, ranges more widely, and is
likely to prove quite a bit more enjoyable:

113
 DRAFT– 31 DEC 2021
9 Elementary proof theory

7. Sara Negri and Jan von Plato, Structural Proof Theory (CUP, 2001).
The first four chapters gives us the basics. Ch. 1 helpfully bridges our
topics, ‘From natural deduction to sequent calculus’. Ch. 2 gives a sequent
calculus for intuitionistic propositional logic and proves the admissibility
of cut. Ch. 3 does the same for classical propositional logic. Ch. 4 adds
the quantifiers.
You might well want to then read on to Ch. 5 which illuminatingly
discusses some variant sequent calculi. Then you can jump to Ch. 8 which
takes us ‘Back to natural deduction’. This relates the sequent calculus to
natural deduction with general elimination rules, shows how to translate
between the two styles of logic, and then derives a normalization theorem
from the cut-elimination theorem: again this is very instructive.
Negri and von Plato note that, as we ‘permute cuts upward’ in a derivation
– in order to eventually arrive at a cut-free proof – the number of cuts
remaining in a proof can increase exponentially as we go along (though
the process eventually terminates). So a cut-free proof can be much bigger
than its original version. Pelletier and Hazen (4) in their §3.8 make some
interesting related comments about sizes of proofs. And you will certainly
want to read this famous short paper:

8. George Boolos, ‘Don’t eliminate cut’, reprinted in his Logic, Logic, and
Logic (Harvard UP, 1998).

And now, if you really want to know more (in particular about how Gentzen
originally arrived at his cut-elimination proof) you can make use of the relevant
IPL chapters, skipping over a lot of the tedious proof-details.
(C) Next, on Gentzen’s proof of the consistency of arithmetic. Von Plato (1)
and Rathjen and Sieg (2) both provide some context for Gentzen’s work. And
here’s a contemporary mathematician’s perspective on why we might be inter-
ested in the proofs of the consistency of PA:
9. Timothy Y. Chow, ‘The consistency of arithmetic’, The Mathematical
Intelligencer 41 (2019), 22–30. Available at tinyurl.com/chow-cons.
Now we have two options, as Rathjen and Sieg (2) makes clear. We can tackle
something like one of Gentzen’s own consistency proofs for PA; but we then have
to tangle with a lot of messy detail as we negotiate the complications caused
by having to deal with the induction axioms. Or alternatively we can box more
cleverly, and prove consistency for a theory PAω which swaps the induction
axioms for an infinitary rule. The proof uses the same overall strategy, but this
time its implementation is a lot less tangled (yet the proof still does the needed
job, since PAω ’s consistency implies PA’s consistency).
There are a number of versions of the second line of proof in the literature.
There is quite a neat but rather terse version here, from which you should be
able to get the general idea (it assumes you know a bit about ordinals):
114
 DRAFT– 31 DEC 2021
Some parallel/additional reading

10. Elliott Mendelson, Introduction to Mathematical Logic, ‘Appendix: A

consistency proof for formal number theory’ (1st edn., 1964; later dropped
but restored in the 6th edn., 2015).
But let’s suppose that you do want something much closer to Gentzen’s original
proof:

There is a rather austere presentation of a Gentzen-style proof in the classic

textbook on proof theory by Takeuti which I will mention in the next section:
this might suit the more mathematical reader. But the following is more
accessible – though with a distracting amount of detail:

3. Mancosu, Galvan and Zach, IPL, Chapters 7–9.

Read Chapter 8 on ordinal notations first. Then the main line of proof is in
Chapters 7 and 9. Now, after an initial dozen pages saying something about
PA, these two chapters together span another sixty-five pages(!), and it is
consequently easy to get lost/bogged down in the details. And it is not as
if the discussion is padded out by e.g. a philosophical discussion about the
warrant for accepting the required amount of ordinal induction; the length
comes from hacking through more details than any sensible reader will want
or need.
However, if you have already tackled a modest amount of other mathe-
matical logic, you should by now have enough nous to be able to read these
chapters pausing over the key ideas and explanations while initially skip-
ping/skimming over much of the detail. You could then quite quickly and
painlessly end up with a very good understanding of at least the general
structure of Gentzen’s proof and of what it is going to take to elaborate it.
So I suggest first skimming through to get the headline ideas, and then do
a second pass to get more feel for the shape of some of the details. You can
then drill down further again to work through as much of the remaining
nitty-gritty that you then feel that you really want/need (which probably
won’t be much!).

9.5 Some parallel/additional reading

Here I will mention (parts of) just three other books for now. All start again from
scratch, but then their varied modes of presentation are in each case perhaps half
a step up in mathematical sophistication from the readings in the last section;
11. Gaisi Takeuti, Proof Theory* (North-Holland 1975, 2nd edn. 1987: re-
printed Dover Publications 2013).
This is a true classic – if only because for a while it was about the
only available book on most of its topics. Later chapters won’t really
be accessible to beginners. But you can certainly tackle Ch. 1 on logic,
§§1–7 (and perhaps the beginnings of §8, pp. 40–45, which is easier than
115
 DRAFT– 31 DEC 2021
9 Elementary proof theory

it looks if you compare how you prove the completeness of a tree system
of logic). Then tackle Ch. 2, §9 on Peano Arithmetic. You can skip the
next section on the incompleteness theorem, and skim §11 on ordinals
(which makes rather heavy weather of what’s really needed, which is the
claim that a decreasing series of ordinals less than ε0 can only be finitely
long: see p. 98 on). The core consistency proof is then given in §12; read
up to at least p. 114. This isn’t exactly plain sailing – but if you skip
and skim over some of the more tedious proof-details you should pick up
a good basic sense of what happens in the consistency proof.
12. Jean-Yves Girard, Proof Theory and Logical Complexity. Vol. I (Bib-
liopolis, 1987). With judicious skipping, which I’ll signpost, this is read-
able and insightful, though some proofs are a bit arm-waving.
So: skip the ‘Foreword’, but do pause to glance over ‘Background and
Notations’ as Girard’s symbolic choices need a little explanation. Then
the long Ch. 1 is by way of an introduction, proving Gödel’s two in-
completeness theorems and explaining ‘The Fall of Hilbert’s Program’:
if you’ve read some of the recommendations on arithmetic, you can prob-
ably skim this fairly quickly, though noting Girard’s highlighting of the
notion of 1-consistency.
Ch. 2 is on the sequent calculus, proving Gentzen’s Hauptsatz, i.e.
the crucial cut-elimination theorem, and then deriving some first conse-
quences (you can probably initially omit the forty pages of annexes to
this chapter). Then also omit Ch. 3 whose content isn’t relied on later.
But Ch. 4 on ‘Applications of the Hauptsatz ’ is crucial (again, however,
at a first pass you can skip almost 60 pages of annexes to the chap-
ter). Take the story up again with the first two sections of Ch. 6, and
then tackle the opening sections of Ch. 7. A rather bumpy ride but very
illuminating.
13. A. S. Troelstra and H. Schwichtenberg, Basic Proof Theory (CUP 2nd
ed. 2000).
This a volume in the series ‘Cambridge Tracts in Computer Science’.
Now, one theme that runs through the book concerns the computer-
science idea of formulas-as-types and invokes the lambda calculus: how-
ever, it is in fact quite possible to skip over those episodes if (as is
probably the case) you aren’t yet familiar with the idea. The book, as
the title indicates, is intended as a first foray into proof theory, and it
is reasonably approachable. However it does spend quite a bit of time
looking at slightly different ways of doing natural deduction and slightly
different ways of doing the sequent calculus, and the differences may
matter more for computer scientists with implementation concerns than
for others.
You can, however, with a bit of skipping, at this stage very usefully
read just Chs. 1–3, the first halves of Chs. 4 and 6, and then Ch. 10 on
arithmetic again.

116
 DRAFT– 31 DEC 2021
Some parallel/additional reading

We will return to consider more advanced texts on proof theory in the final
chapter, §12.6.

117
 DRAFT– 31 DEC 2021

10 Modal logics

A deduction, Aristotle tells us, requires a conclusion which ‘comes about by

necessity’ given some premisses. So it is no surprise that, from the very beginning,
logicians have been interested in the modal notions of necessity and possibility.
Modern modal logics aim, at least in the first place, to regiment reasoning about
such notions. But as we will see, they can be applied much more widely.
Here’s an attractive thought: it is necessarily true that A just if A is not only
true here in the actual world but also obtains in all relevant possible worlds.
Suppose we add to a logical language a symbol , where A is to be read as it
is necessarily true that A. Then, to formally model our attractive thought, we
will take some objects to represent possible worlds, and say that A is true at
‘world’ w in the model just if A is true at all ‘worlds’ w0 suitably related to w.
Compare: in §8.3(c), we described a semantic model for intuitionistic logic
with the following key feature – to determine whether the conditional A → B
holds in a situation k in the model, it isn’t enough to know whether A holds in k
and whether B holds in k; we also need to know whether A and B obtain in other
situations k 0 suitably related to k. So now the idea is to use a similar relational
semantics for the necessity operator, with the truth of A in one situation w
again depending on what happens in other related situations w0 .
In §10.1, then, we explore this key idea by taking a look at some basic modal
logics. These and similar logics will be of interest to quite a few philosophers and
also eventually to some mathematicians and computer scientists who investigate
relational structures. There is, however, one rather distinctive modal logic which
should be of particular interest to anyone beginning mathematical logic, namely
so-called provability logic: we will highlight that in §10.2. Provability logic can
be tackled without a wider background in modal logic; but it certainly doesn’t
hurt to know a little about the wider picture we introduce first.

10.1 Some basic modal logics

(a) Notation first. As just proposed, we are going to add a one-place operator
 to our familiar logical languages (propositional, first-order), governed by the
new syntactic rule that if A is a wff, so is A.
Now, as we said,  is typically going to be interpreted as some sort of necessity
operator. We could also build into our languages a corresponding possibility oper-

118
 DRAFT– 31 DEC 2021
Some basic modal logics

ator 3 (so we read 3A as it is possibly true that A). But, to keep things simple,
we won’t do that, since 3A can equally well be treated as just a definitional
abbreviation for ¬¬A. Reflect: it is possibly true that A iff A is true at some
possible world, iff it isn’t the case that A is false at all possible worlds, iff it isn’t
the case that ¬A is necessary. So the parallel between the equivalences 3/¬¬
and ∃w/¬∀w¬ is not an accident!
A third modal symbol you will come across is J, for what is standardly called
‘strict implication’. But again, we can treat A J B as a definitional abbreviation,
this time for (A → B).
Hence, following quite common practice, we will here take  to be the sole
built-in modal operator in our languages.
(b) The story of modern modal logic begins with C. I. Lewis’s 1918 book A
Survey of Symbolic Logic. Lewis presents postulates for J, motivated by claims
about the proper understanding of the idea of implication, though unfortunately
his claims do seem pretty muddled.1 Later, in C. I. Lewis and C. H. Langford’s
1932 Symbolic Logic, there are further developments: the authors distinguish five
modal logics of increasing strength, which they label S1 to S5. But why multiple
logics?
Let’s take four schemas, and ask whether we should accept all their instances
when the  is interpreted in terms of necessary truth:

K (A → B) → (A → B)

T A → A
S4 A → A
S5 ¬A → ¬A

Well, on any understanding of the idea of necessity, if A → B and A both hold

necessarily, so does B: so we can accept the principle K. And necessary truth
implies plain truth: so we can accept T too. But what about the principles S4
and S5 (which are in fact distinctive of Lewis and Langford’s systems S4 and
S5 )?
It seems that different principles about repeated modalities will be acceptable
depending on how exactly we interpret the necessity involved. Take a couple
of examples. Suppose we interpret A in a mathematical context as meaning
that A necessarily holds in the sense that it is provable that A (i.e. is provable
by ordinary informal standards of proof): then arguably (i) in this case, S4 but
not S5 holds. Alternatively, suppose we interpret  as indicating analyticity in
the old-fashioned philosopher’s sense (where it is analytically true that A if A
is true just in virtue of its conceptual content): then arguably (ii) in this case,
both the S4 and S5 principles hold. But I’m certainly not going to get into the
business of assessing the supposed arguments for (i) and (ii) – the issues are
far too murky. And that’s exactly the point to make here: the early discussions
of systems of modal logic, and the supposed semantic justifications for various

1 The modern reader might well suspect confusion between ideas that we now demarcate by
using the distinguishing notations →, ` and .

119
 DRAFT– 31 DEC 2021
10 Modal logics

suggested principles, were entangled with contentious philosophical arguments.

No wonder then that modal logic initially had a somewhat shady reputation!
(c) The picture radically changed some thirty years after Lewis and Langford,
when Saul Kripke (in particular) developed a sharply characterized framework
for giving semantic models for various modal logics.
Let’s begin with the headline news about some modal propositional logics. In
this subsection we’ll describe a family of semantic models. In the next subsection
we’ll describe a family of deductive modal proof systems. Then the following
subsection makes the Kripkean connections between the two!
So let’s assume we are working in some suitable language L with the absurdity
constant ⊥ built in alongside the other usual propositional connectives, plus the
unary operator . And to define a relational semantics for such a language, we
obviously need to start by introducing relational structures:

1. The basic ingredients we need are some objects W and a relation R defined
over them. For the moment, think of W as a collection of ‘possible worlds’
and then wRw0 will say that the world w0 is possible relative to w (or if
you like, w0 is an accessible possible world, as seen from w).
2. And we will pick out an object w0 from W to serve as the ‘actual world’.

But we need an important further idea:

3. To get different flavours of relational structure (for interpreting different

flavours of modal deductive system) we will want to specify different condi-
tions S that the relation R needs to satisfy. For just one example, we might
be particularly interested in relational structures where R is specified as
being transitive and reflexive.

Let’s say, for short, that a relational structure where the relation R satisfies the
condition S is an S-structure.
Next we define the idea of a valuation of L-sentences on an S-structure. The
story starts unexcitingly!

10 . We initially assign a value, either true or false, to each propositional letter

of L with respect to each world w. Then,
20 . The propositional connectives behave in the now entirely familiar classical
ways. For example, A → B is true at w if and only if either A is false at w
or B is true at w; and so forth.

The only real novelty, as trailed at the outset, is in the treatment of the modal
operator . We stipulate

30 . A is true at a world w if and only if A is true at every world that is

possible relative to w, i.e. A is true at every world w0 such that wRw0 .

Evidently, given (20 ) and (30 ), every valuation ends up assigning a value to each
L-wff A at each world.

120
 DRAFT– 31 DEC 2021
Some basic modal logics

Let’s say that an S-structure together with such a valuation for L-sentences
is an S-model for L. Then, continuing our list of definitions, when A is an L-
sentence,

40 . A is (simply) true in a given S-model for L if and only if A takes the value
true at the actual world w0 in the model.

Finally, and predictably, we say

50 . A is S-valid if and only if it is true in every S-model.

So that sets up the general framework for a relational semantics for a propo-
sitional modal language. But we are now going to be interested in four different
particular versions got by filling out the specification S in different ways, and so
giving us four different notions of validity for propositional modal wffs:

(K) K-validity is defined in terms of K-models which allow any relation R (the
specification condition S is null).
(T) T -validity is defined in terms of T -models which require the relation R to
be reflexive.
(S4) S4 -validity is defined in terms of S4 -models which require the relation R
to be reflexive and transitive.
(S5) S5 -validity is defined in terms of S5 -models which require the relation R
to be reflexive, transitive and symmetric (i.e. R has to be an equivalence
relation).

As we will soon discover, the labels we have chosen are indeed significant!
(d) Let’s look at a couple of very instructive mini-examples. Take first the
following two-world model, with an arrow w −→ w0 depicting that wRw0 , and
with the values of P at each world as indicated:

w0 w1
P := F P := T

Now, in this model, P is true at w0 , since P is true at every world accessible

from w0 , namely w1 . P is also true at w1 , since P is again true at every world
accessible from w1 , namely w1 itself. And so P is true at w0 , since P is true
at every world accessible from w0 .
But note P → P is false at w0 . So in a model like this one where the
accessibility relation is not reflexive, not every instance of the schema T is true.
Conversely, a moment’s reflection shows that in T -models, which require that
the accessibility relation is reflexive, instances of the schema T must always be
true (because if A is true at w0 then A is true at all accessible worlds, which
will include w0 by the reflexiveness of accessibility).

121
 DRAFT– 31 DEC 2021
10 Modal logics

Moral: if 2 is to be interpreted as necessary truth, where instances of the

schema T should always come out true, then we’ll want our semantic models to
be built using a reflexive relation R.
For our second example, take this three-world model:

w0 w1 w1
P := T P := T P := F

Note, this is not only a K model but also a T -model, because the diagrammed
accessibility relation R is reflexive; but it is not an S4 model since R is not
transitive (we have w0 Rw1 and w1 Rw2 but not w0 Rw2 ).
Now, in this model, P is true at w0 (because P is true at both the accessible-
from-w0 worlds, i.e. at w0 and w1 ). But P is false at w1 (because P is false at
the accessible-from-w1 world w2 ). And then since P is false at w1 and w1 is
accessible from w0 , it follows that P is false at w0 . And hence in this model
P → P is false (i.e. false at w0 ). Moral: the S4 principle can fail in models
where the accessibility relation is not transitive.
But we can also show the reverse – in other words, in S4 models where the
accessibility relation is transitive, the S4 principle holds. That follows because
S4 can only fail in a model if the accessibility relation is non-transitive:

Suppose something of the form A → A is false in a given model,

so (i) A is true at w0 while (ii) A is false at w0 . But for (ii) to
hold, there must be a world w1 such that w0 Rw1 and (iii) A is false
at w1 . And for (iii) to hold there must be a world w2 such that w1 Rw2
and (iv) A is false at w2 . But then (iv) w2 must be ‘invisible’ from
w0 , or else (i) couldn’t hold: i.e. we can’t have w0 Rw2 . In sum, for
A → A to fail we need three worlds such that w0 Rw1 , w1 Rw2
but not w0 Rw2 – which requires R to be non-transitive.

So our two mini-examples very nicely make the connection between a structural
condition on models and the obtaining of a general modal principle such as T or
S4. More about this very shortly.
(e) Since our main concern here is with the formalities, we won’t delve into the
arguments about which specification conditions S appropriately reflect which
intuitive notions of necessity (though note that even the condition T can fail if
e.g. we want to model deontic necessities – i.e. necessities of duty: since what
ought to be the case may not in fact be the case!). We can leave it to the
philosophers to fight things out. For now, it might be more useful to pause to
summarize our semantic story in the style of our earlier account of intuitionistic
semantics in §8.3(c).
So, an S-structure is a triple (w0 , W, R) where W is a set, w0 ∈ W , and R is
a relation defined over W which satisfies the conditions S. Then an S-model for

122
 DRAFT– 31 DEC 2021
Some basic modal logics

a modal propositional language L is an S-structure together with a valuation

relation (‘makes true’) between members of W and wffs of L such that
(i) w 1 ⊥.
(ii) w ¬A iff w 1 A.
(iii) w A ∧ B iff w A and w B.
(iv) w A ∨ B iff w A or w B.
(v) w A → B iff w 1 A or w B.
(vi) w A iff, for any w0 such that wRw0 , w0 A.
We say that A is true in a given S-model when w0 A. As before, A is S-valid
when A is true in all S-models. And for the moment the most significant condi-
tions S on the accessibility relation R in a model are K (null), T (reflexivity),
S4 (reflexivity and transitivity), S5 (equivalence).
Finally, I need to link up what I’ve just said with other presentations you’ll
encounter. So note that – although Kripke’s original presentation did involve, as
here, picking out a ‘world’ w0 from W to play the role of the ‘actual’ world – it
is clear that we can drop that step and can equivalently re-define S-validity as
truth at all worlds in an S-model. (Why? Obviously, if A is valid on the revised
definition it is valid on our original definition. While if A is not valid on the
revised definition, A must be false at some world, and so it will be false on the
Kripke model with that world chosen as the ‘actual’ world w0 .)
(f) Now let’s turn to consider some proof systems for propositional modal log-
ics. And, just because it is simplest way to do things, let’s give an old-school
axiomatic presentation (leaving natural deduction and tableaux versions to be
explained in the recommended reading). Here then are four key systems, starting
with the simplest:
(K) The modal axiomatic system K is the theory whose axioms are
(Ax i) All instances of tautologies.
(Ax ii) All instances of the schema K.
And whose rules of inference are
(MP) From A and A → B, infer B.
(Nec) If A is deducible as a theorem, infer A.
To explain briefly: Read (Ax i) as meaning that, given a schema for a classi-
cal tautology, the result of systematically substituting any wffs of our modal
propositional language for schematic letters – even substituting modalized wffs
– will be an axiom of K. So, for example, (A ∧ B) → A is a schema for a clas-
sical tautology. Hence the result of substituting P for A and Q for B, giving
us (P ∧ Q) → P, is an axiom of K. Such instances of tautologies are still,
surely, logical truths.
We’ve already said that instances of (Ax ii) look good on any suitable reading
of the box. And our old friend the modus ponens rule (MP) is uncontentious.
Which leaves the necessitation rule (Nec). This is to be very sharply distin-
guished from what would evidently be the quite unacceptable axiom schema
123
 DRAFT– 31 DEC 2021
10 Modal logics

A → A: obviously, A can be true without being necessarily true. However, the
idea justifying (Nec) is that if A is actually a logical theorem – i.e. is deducible
from logical principles alone – then it will indeed be necessary (on almost any
sensible understanding of ‘necessary’). Here’s an example of the rule (Nec) in
use in a K-proof:

1. ((P ∧ Q) → P) Axiom, by (Ax i)

2. ((P ∧ Q) → P) By (Nec), since 1 is a theorem
3. (((P ∧ Q) → P) → ((P ∧ Q) → P)) Axiom, by (Ax ii)
4. ((P ∧ Q) → P) From 2 and 3 by (MP)
5. ((P ∧ Q) → P) By (Nec), since 4 is a theorem

In sum, then, all the theorems of the weak system K – i.e. all the wffs deducible
from axioms alone – should be logical truths on (almost all) readings of  read
as a kind of necessity.
And now here are three nested ways of strengthening the system K:

(T) T is the axiomatic system K augmented with all instances of the schema
T as axioms.
(S4) S4 is T augmented with all instances of the schema S4 as axioms.
(S5) S5 is S4 augmented with all instances of the schema S5 as axioms.

The readings will give lots of examples of these (or equivalent) proof systems in
action.
(g) So now at last for the big reveal – except of course I’ve entirely sploit any
element of surprise by the parallel labelling of the flavours of modal semantics
and the flavours of axiomatic proof system!
What Kripke famously showed is the following lovely result:

Whether S is K, T , S4, S5, a wff A is an S-theorem if and only if it

is S-valid.

In short, we have soundness and completeness theorems for our proof systems.
And there are some nice immediate implications. Searching for an appropriate
countermodel which shows that a wff is not S-valid is a finite business, so it is
decidable what’s S-valid – and hence it is decidable what’s an S-theorem.2
These soundness and completeness results are not mathematically very dif-
ficult. Perhaps Kripke’s real achievement was the prior one in developing the
general semantic framework and in finding the required simple proof systems –
some of them different from any of the systems proposed by Lewis and Langford
– thereby making his very elegant result possible.

2 Suppose we define in the now obvious ways (i) the idea of a conclusion being an S-valid
consequence of some finite number of premisses, and (ii) the idea of that conclusion being
deducible in system S from those premisses. Then again we have soundness and weak
completeness proofs linking valid consequences with deductions, and we have corresponding
decidability results too. We won’t worry however about strong completeness, which does
indeed fail for some modal logics, e.g. for GL which we meet in the next section.

124
 DRAFT– 31 DEC 2021
Provability logic

(h) And now, with the apparatus of relational semantics available, the flood-
gates really open! After all, the objects in a S-model don’t have to represent
‘possible worlds’ (whatever they are conceived to be); they can stand in for
any points in a relational structure. So perhaps they could represent states of
knowledge, points of a time series, positions in a game, states in the execution of
a program, levels in a hierarchy . . . with different classes of accessibly relations
appropriate for different cases and so with different deductive systems to match.
The resulting applications of propositional modal logics are indeed very many
and various, as you will see.
(i) And what about quantified modal logics, where we add the modal operator
 to a first-order language? Why might we be interested in them?
Well, philosophers make play with questions like this: Does it make sense to
suppose the very same objects can appear in the domains of different possible
worlds? If it does, do all possible worlds contain the same objects (perhaps some
of them actualized, some not)? Does a proper name (formally a constant term)
denote the same thing at any possible world at which it denotes at all? Are
atomic identity statements, if true at all, necessarily true? Questions of this
stripe pile up, and they motivate different ways of tweaking quantified modal
logic in formally modelling and so clarifying the philosophical ideas.
However, the resulting logics don’t seem to be of particular interest to non-
philosophers; the wider logical community has (as yet) been much more inter-
ested in propositional modal logics.
Still, the beginnings of the technical story about first-order modal logics are
pretty accessible. And the suggested readings will enable you to get some head-
line news about different proof systems and their formal semantics, without
getting too entangled in unwanted philosophical debates!

10.2 Provability logic

As just noted, propositional modal logics have a very wide range of applications.
But there is one that stands out as being of pre-eminent relevance to anyone
beginning mathematical logic. And that is provability logic.
(a) Let’s start with some reminders of what you should already know from
tackling Gödel’s incompleteness theorems (see §6.4). So take a theory in which
we can do enough arithmetic – to fix on an example, take first-order Peano
Arithmetic. Choose a sensible system of Gödel-numbering. Then you can con-
struct a relational predicate in the language of arithmetic – one which we can
abbreviate Prf(x, y) – that nicely3 represents the relation which obtains between
two numbers x, y, when x is the Gödel number of a PA proof of the sentence
with Gödel number y. Now define Prov(y) to be the expression ∃xPrf(x, y). Then
Prov(y) represents the property that a number y has if it numbers a theorem of
PA – so Prov is naturally called a provability predicate.
3 ‘Nicely’
waves a hand at some details which are important but which we won’t need to
delay over here!

125
 DRAFT– 31 DEC 2021
10 Modal logics

If A is wff of arithmetic, let pAq be shorthand for A’s Gödel-number, and let
pAq be shorthand for the formal numeral for pAq. Then, given our definitions,
Prov(pAq) says that A is provable in PA.
Now we introduce yet another bit of shorthand: let’s use A as a simple
abbreviation for Prov(pAq).4 With some effort, we can then show that PA proves
(unpacked versions of) all instances of the following familiar-looking schemas

K· (A → B) → ( A → B)
S4· A→ A

And moreover we have an analogue of the modal Necessitation rule:

(Nec·) If A is deducible as a PA theorem, then so is A.

That package of facts about PA is standardly reported by saying that the theory
satisfies the so-called HBL derivability conditions. And appealing to these facts
together with the First Incompleteness Theorem, it is then easy to derive the
Second Theorem that PA cannot prove ¬ ⊥ (i.e. can’t prove that ⊥ isn’t
provable, i.e. can’t prove that PA is consistent).5
(b) The obvious next question might well seem to be: what other modal princi-
ples/rules should our dotted-box-as-a-provability-predicate obey, in addition to
the dotted principles K· and S4·, and the rule (Nec·)? What is its appropriate
modal logic?
But hold on! We are getting ahead of ourselves, because we so far only have
the illusion of modal formulas here. The box as just defined simply doesn’t have
the right grammar to be a modal operator. Look at it this way. In a proper
modal language, the operator  is applied to a wff A to give a complex wff A
in which A appears as a subformula. But in our newly defined usage where A
is short for Prov(pAq), the formula A doesn’t appear as a subformula at all –
what fills the appropriate slot(s) in the predicate Prov is a numeral (the numeral
for the number which happens to code the formula A).
In short, the surface form of our notation A is entirely misleading as to its
logical form. Which is why the logically pernickety might indeed not be very
happy with the notation.
However, it remains the case that our abbreviatory notation is highly sugges-
tive. And what it suggests is starting with a kosher modal propositional language
of the kind now familiar for §10.1, where the box is genuinely a unary opera-
tor applied to wffs. And then we consider arithmetical interpretations which
map sentences A of our modal language to corresponding sentences A∗ of PA,
interpretations which have the following shape:

i. An interpretative map sends an atomic letter A of our modal language to

some arithmetical sentence A∗ , any you like.

4 I’ve dotted the box here – not the usual notation – for clarity’s sake!
5 Formore details, if this is new to you, see for example Chapter 33 of my An Introduction
to Gödel’s Theorems (downloadable from logicmatters.net/igt).

126
 DRAFT– 31 DEC 2021
Provability logic

ii. The map then respects the propositional connectives: for example, it sends
conjunctions in the modal language to conjunctions in the arithmetic lan-
guage, so (A ∧ B)∗ is (A∗ ∧ B ∗ ); it sends the absurdity constant to the
absurdity constant, i.e. ⊥∗ is ⊥; and so on.
iii. The map sends the modal sentence A to A∗ , i.e. to Prov(pA∗ q).

There is now no notational jiggery pokery; we have a respectable modal language

on the one side, and various interpretative mappings from its sentences into a
regular arithmetical language on the other side.
And now we can ask a cogent versions of the misplaced question we wanted
to ask before. In particular, we can ask: what are the modal sentences which
are such that, on any interpretative mapping into PA, their translations are
arithmetical theorems? What, for short, is the correct modal logic for the 
interpreted this way as tracking formal provability in PA?
(c) Here’s a reminder of another result we can get from the HBL conditions,
namely Löb’s Theorem.
Momentarily using again our now somewhat deprecated dotted-box-as-abbreviation
notation, this rather surprising theorem says:

If PA proves A → A, then it proves A.6

We will presumably want to reflect this theorem in a logic for the genuinely
modal  operator interpreted as arithmetical provability: a natural move, then,
is to build into our modal logic the rule that, if A → A is deducible as a
theorem, then we can infer A.
So this putting this thought together with our previous remarks, let’s consider
the following modal logic – the ‘G’ in its name is for Gödel who made some
prescient remarks, and the ‘L’ is for Löb:

(GL) The modal axiomatic system GL is the theory whose axioms are
(Ax i) All instances of tautologies
(Ax ii) All instances of the schema K: (A → B) → (A → B)
(Ax iii) All instances of the schema S4: A → A
And whose rules of inference are
(MP) From A and A → B, infer B
(Nec) If A is deducible as a theorem, infer A
(Löb) If A → A is deducible as a theorem, infer A.

You can immediately see, by the way, that we don’t also want to add all instances
of the T-schema A → A to this modal logic. For a start, doing that would
make ⊥ → ⊥ a theorem and hence ¬⊥ would be a theorem. But that can’t
correspond on arithmetic interpretation to a theorem of PA, since we know that
PA can’t prove ¬ ⊥.

6 See Chapter 34 of An Introduction to Gödel’s Theorems.

127
 DRAFT– 31 DEC 2021
10 Modal logics

And there’s worse: leaving aside the desired interpretation of this logic, if we
add all instances of A → A as axioms, then in the presence of the rule (Löb),
we can derive any A, and the logic is inconsistent.
Now, given our motivational remarks in defining GL, it won’t be a surprise
to learn that it is indeed sound on the provability interpretation. Once we have
done the (non-trivial!) background work required for showing that the HBL
derivability conditions and hence Löb’s theorem hold in PA, it is quite easy to
go on to establish that, on every interpretation of the modal language into the
language of arithmetic, every theorem of GL is a theorem of PA.
And (with more decidedly non-trivial work due to Robert Solovay) it can also
be shown that GL is complete on the provability interpretation. In other words,
if a modal sentence is such that every arithmetic interpretation of it is a PA
theorem, then that sentence is a theorem of the modal logic GL.
Which is all very pleasingly neat!
(d) We should pause to note that there is another way of presenting this prov-
ability logic.
Suppose we drop the Löb inference rule from GL, and replace the instances
of the S4 schema as axioms with instances of the Löb-like schema

L (A → A) → A

It is then quite easy to see that this results in a modal logic with exactly the
same theorems (because GL in our original formulation implies all instances of
L; and conversely we can show that all instances of S4 can be derived in the new
formulation, for which the Löb rule is also a derived rule of inference). Hence
either formulation gives us the provability logic for PA.
(e) Now, we’ve so far been working with arithmetic interpretations of our modal
wffs. But we can also give a more abstract Kripke-style relational semantics for
GL (it is a nice question, though, whether this ‘semantics’ has much to do
with meaning!). We start by defining a GL-model in the usual sort of way as
comprising a valuation with respect to some worlds W with a relation R defined
over them, where R satisfies . . .
Well, what conditions do we in fact need to place on R so that GL-theorems
match with the GL-validities (the truths that hold at every world, for every GL-
model)? Clearly, we mustn’t require R to be reflexive – or else all instances of
the T-schema would come out GL-valid, and we don’t want that. Equally clearly,
we must require R to be transitive – or else instances of the S4-schema could
fail to be GL-valid. But we need more: what further condition on R is required
to make all the instances of the L-schema come out valid?
It turns out that what is needed is that there is no infinite chain of R-related
worlds w0 , w1 , w2 , w3 , . . . such that w0 Rw1 Rw2 Rw3 . . . (and that condition en-
sures that R is irreflexive, for otherwise we would have some infinite chain
wRwRwRw . . .). Call that the finite chain condition. Then define a GL-model as
one where the accessibility relation R is transitive and satisfies the finite chain

128
 DRAFT– 31 DEC 2021
First readings on modal logic

condition. Then a modal sentence is GL theorem if and only if it is GL-valid

(true in all worlds in all GL-models).
This new soundness and completeness theorem has a lovely upshot. As with
the other modal logics we’ve met, there is a systematic way of testing for GL-
validity (by systematically searching for Kripke-style countermodels). So it is
decidable what’s a GL theorem.
(f) That last result, together with the fact that GL is sound and complete
for arithmetical interpretations into theorems of PA, shows something rather
remarkable. Although PA as a whole is an undecidable theory, there is a very in-
teresting part of that theory – roughly, what it can say by applying propositional
logic and its provability predicate to arithmetical wffs – which is decidable.
For example, consider this question: for any arithmetical sentence A, does
PA know – i.e. can it prove? – that, if A is provably equivalent to the claim it
isn’t provable, then A is provably equivalent to saying that PA is consistent? In
symbols, using the dotted-box-as-abbreviation, can PA prove
(A ↔ ¬ A) → (A ↔ ¬ ⊥)
Well it can so long as the corresponding modal wff
(P ↔ ¬P) → (P ↔ ¬⊥)
is a GL theorem – and that’s decidable (in fact, it is a theorem).
This way, we easily get to know a lot more about what PA can prove about it
can prove. And this is just one example of the kind of payoff we get from applying
modal logic to questions of provability in arithmetics. Hence the considerable
interest of provability logic.

10.3 First readings on modal logic

(a) There is, as so often, a good entry in that wondrous resource the Stanford
encyclopaedia, one which should provide more very helpful orientation:
1. James W. Garson, ‘Modal logic’, The Stanford Encyclopedia of Philoso-
phy: read §§1–11 and 15. Available at tinyurl.com/sep-modal.
Now, because of its interest, modal logic is often taught to philosophers without
much logical background, and so there are a number of introductions written
primarily for them. One often recommended example is the very accessible
2. Rod Girle, Modal Logics and Philosophy (Acumen 2000; 2nd edn. 2009).
Part I of this book provides a clear introduction, which in 136 pages
explains the basic syntax and relational semantics, covering both trees
(tableaux) and natural deduction for some propositional modal logics,
and extends as far as the beginnings of quantified modal logic.
Philosophers may well very want to go on to read Part II of this book, on
applications of modal logic.
129
 DRAFT– 31 DEC 2021
10 Modal logics

But there is a clearer and better-organized account in an extraordinarily useful

book by Graham Priest. I’ll highlight this not only because it is crisper on modal
logics, but because we also get an account of intuitionistic logic in the same
tableaux framework:

3. Graham Priest, An Introduction to Non-Classical Logic* (CUP, much

expanded 2nd edition 2008). This treats a whole range of logics system-
atically, concentrating on semantic ideas, and using a tableaux approach.
Chs. 1 and 12 provide quick revision tutorials on tableaux for classical
propositional and predicate logic. Then Chs. 2 and 3 give the basics on
propositional modal logics. You can then either fill in more about modal
logics in Ch 4 or skip to Ch. 6 on propositional intuitionistic logic. Then
Chs 14 and 15 introduce the basics on quantified modal logics. You can
then fill in more about quantified modal logics in Chs 16–18 or can then
skip to Ch. 20 on quantified intuitionistic logic.
This whole book – which we will revisit in our next chapter – is a
terrific achievement and enviably clear and well-organized.

Then, going half-a-step up in sophistication, though still starting from scratch,

we find another excellent book (elegantly done in a way which might appeal
more to mathematicians):

4. Melvin Fitting and Richard L. Mendelsohn, First-Order Modal Logic

(Kluwer 1998). This gives both tableaux and axiomatic systems for var-
ious modal logics, in an approachable style and with lucid discussions of
options at various choice points. Despite its more mathematical flavour,
the book still includes some interesting discussions of the conceptual mo-
tivations for different modal logics.
Read the first half of this book to get a compact but sufficient intro-
duction to propositional modal logics, and also the initial headlines about
quantified modal logics. Philosophers will then want to read on.

And let me also mention:

5. Johan van Bentham, Modal Logic for Open Minds (CSLI Publications,
2010). This ranges widely and is good at highlighting main ideas and
making cross-connections with other areas of logic. Particularly interest-
ing and enjoyable to read in parallel with the main recommendations.

10.4 Suggested readings on provability logic

Provability logic is nicely introduced in:

6. Rineke Verbrugge, ‘Provability logic’ §§1–4 and perhaps §6, The Stanford
Encyclopedia of Philosophy. Available at tinyurl.com/prov-logic.
130
 DRAFT– 31 DEC 2021
Alternative and further readings on modal logics

Or you could dive straight into the very first published book on our topic, which
I think still makes for the most attractive entry-point:

7. George Boolos, The Unprovability of Consistency: An Essay in Modal

Logic (CUP, 1979), particularly Chs 1–12. This fairly short book is a
famous modern classic, yet very approachable. And you don’t need any
prior acquaintance with modal logic in order to tackle it. Boolos has
an engaging presentational style (and the book can be read surprisingly
quickly in order to get the main news if you are happy to initially skip
some of the longer proofs).

However, this seems to be one of the very few distinguished mathematical logic
books which is not readily available online. So I need to also mention

8. George Boolos, The Logic of Provability (CUP, 1993). This is a signif-

icantly expanded and updated version of his earlier book. And so you
could indeed read the first half of this instead, though I do retain a fond-
ness for the somewhat more streamlined presentations in the shorter
version. The main occasion for the update is the presentation of proofs
of major results about quantified provability logic which were discovered
after Boolos wrote his first book: but these results are really more than
you need in a first encounter with provability logic.

And here is another classic introductory book:

9. Craig Smoryński, Self-Reference and Modal Logic (Springer-Verlag, 1985).

This is a lovely alternative or accompaniment to Boolos’s 1979 book. Not
lovely to look at, as it oddly printed in extremely small type emulating
an electric typewriter, which doesn’t make for comfortable reading: but
the content is extremely lucidly and elegantly presented, with a lot of
helpful explanatory/motivating chat alongside the more formal work.
Also highly recommended.

Then, for more pointers towards recent work on related topics you could look at
§5 of Verbrugge’s article and/or at the following interesting overview:

10. Sergei Artemov, ‘Modal logic in mathematics’ §§1–5, in The Handbook

of Modal Logic, edited by P. Blackburn et al. (Elsevier, 2005).

10.5 Alternative and further readings on modal logics

(a) Other introductory readings for philosophers The first part of Theodore
Sider’s Logic for Philosophy* (OUP, 2010) is poor as an introduction to FOL.
However, the second part, which is entirely devoted to modal logic and related
topics like Kripke semantics for intuitionistic logic, is very much better, and

131
 DRAFT– 31 DEC 2021
10 Modal logics

philosophers could indeed find it useful. For example, the chapters on quanti-
fied modal logic (and some of the conceptual issues they raise) are brief and
approachable.
Sider is, however, closely following a particularly clear old classic by G. E.
Hughes and M. J. Cresswell A New Introduction to Modal Logic (Routledge,
1996, updating their much earlier book). This can still be recommended and
may suit some readers, though it does take a rather old-school approach.
If your starting point has been Priest’s book or Fitting/Mendelson, then you
might want at some point to supplement these by looking at a treatment of
natural deduction proof systems for modal logics. One option is to dip into Tony
Roy’s long article ‘Natural derivations for Priest’, in which he provides ND logics
corresponding to the propositional modal logics presented in tree form in Priest’s
book, though this gets much more detailed than you really need: available at
tinyurl.com/roy-modal. But a smoother introduction to ND modal systems is
provided by Chapter 5 of Girle, or by my main alternative recommendation for
philosophers, namely

11. James W. Garson, Modal Logic for Philosophers* (CUP, 2006; 2nd end.
2014). This again is intended as a gentle introductory book: it deals with
both ND and semantic tableaux (trees), and covers quantified modal
logic. It is quite a long book (one reason for preferring the snappier Fit-
ting/Mendelsohn as a first recommendation), with quite a lot on quan-
tified modal logics: and it is indeed pretty accessible.

(b) Modal logics for philosophical applications If you are interested in appli-
cations of propositional modal logics to tense logic, epistemic logic, deontic logic,
etc. then the relevant chapters of Girle’s book give helpful pointers to more read-
ings on these topics. If your interests instead lean to modal metaphysics, then
– once upon a time – a discussion of quantified modal logic at the level of Fit-
ting/Mendelsohn or Garson would have probably sufficed. And for a bit more
on first-order quantified modal logics, see

12. James W. Garson, ‘Quantification in modal logic’ in Handbook of Philo-

sophical Logic, Vol. 3, edited by Dov M. Gabbay and F. Guenther (Rei-
del, 2nd edition 2001).

However, Timothy Williamson’s notable book Modal Logic as Metaphysics (OUP,

2013) calls on rather more, including e.g. second-order modal logics. However,
there doesn’t seem to be general guide/survey of higher-order modal logics at the
right sort of level, with the right sort of coverage to recommend here. There is a
text by Nino B. Cocchiarella and Max A. Freund, Modal Logic: An Introduction
to its Syntax and Semantics (OUP, 2008), whose blurb announces that “a variety
of modal logics at the sentential, first-order, and second-order levels are devel-
oped with clarity, precision and philosophical insight”. However, the treatments
in this book are relentlessly and rebarbatively formal. In its last two chapters,
the book does cover second-order modal logic: but the highly unfriendly mode of

132
 DRAFT– 31 DEC 2021
Finally, a very little history

presentation will probably put the discussion out of reach of most philosophers
who might be interested. You have been warned.
(c) Four more technical books In order of publication, here are three more ad-
vanced and rather more challenging texts I can suggest to sufficiently interested
readers:
13. Sally Popkorn, First Steps in Modal Logic (CUP, 1994). The author is,
at least in this possible world, identical with the late mathematician
Harold Simmons. This book, which entirely on propositional modal log-
ics, is written for computer scientists. The Introduction rather boldly
says ‘There are few books on this subject and even fewer books worth
looking at. None of these give an acceptable mathematically correct ac-
count of the subject. This book is a first attempt to fill that gap.’ This
considerably oversells the case: but the result is illuminating and read-
able.
14. Alexander Chagrov and Michael Zakharyaschev Modal Logic (OUP, 1997).
This is a volume in the Oxford Logic Guides series and again concentrates
on propositional modal logics. Definitely written for the more mathemat-
ically minded reader, it tackles things in an unusual order, starting with
an extended discussion of intuitionistic logic, and is good but rather
demanding.
15. Patrick Blackburn, Maarten de Ricke and Yde Venema’s Modal Logic
(CUP, 2001). This is one of the Cambridge Tracts in Theoretical Com-
puter Science: but don’t let that provenance put you off! This is an
accessibly and agreeably written text on propositional modal logic – cer-
tainly compared with the previous two books in this group – with a lot
of signposting to the reader of possible routes through the book, and
with interesting historical notes. I think it works pretty well, and will
also give philosophers an idea about how non-philosophers can make use
of propositional modal logic.
16. Lloyd Humberstone, Philosophical Applications of Modal Logic* (Col-
lege Publications, 2015). This very large large volume starts with a
book-within-a-book, an advanced 176 page introduction to propositional
modal logics. And then there are extended discussions at a high level of
a wide range of applications of these logics that have been made by
philosophers. A masterly compendium to consult as/when needed.

10.6 Finally, a very little history

Especially for philosophers, it is very well worth getting to know a little about
how mainstream modern modal logic emerged from the to-and-fro between philo-
sophical debate and technical developments. So do read e.g. one of
17. Roberta Ballarin, ‘Modern origins of modal logic’, The Stanford Ency-
clopedia of Philosophy. Available at tinyurl.com/mod-orig.
133
 DRAFT– 31 DEC 2021
10 Modal logics

18. Sten Lindström and Krister Segerberg, ‘Modal logic and philosophy’ §1,
in The Handbook of Modal Logic, edited by P. Blackburn et al. (Elsevier,
2005).

134
 DRAFT– 31 DEC 2021

11 Other logics?

So far we have looked at just three variants or extensions of standard FOL:

i. One limitation of FOL is that we can only quantify over objects, as op-
posed to properties, relations and functions. Yet seemingly, we quantify
over properties etc. in informal mathematical reasoning. In Chapter 4, we
therefore considered adding second-order quantifiers. (This is just a first
step: there is a rich mathematical theory of higher-order logic, a.k.a. type
theory, which you will eventually want to explore – but I deem that to be
a more advanced topic, so we will return to it in the final chapter, §12.7.)

ii. In Chapter 8 we looked at what happens if we drop the classical law of

excluded middle. The resulting intuitionistic logic is mathematically ele-
gant and also widely applicable (in constructive reasoning, in theoretical
computer science, in category theory).

iii. Then in Chapter 10 we explored the use of the kind of relational semantics
we first met in the context of intuitionistic logic, but now in extending
FOL with modal operators. Again, the development on the formal side
is mathematically quite elegant: and some modal logics – in particular,
provability logic – have worthwhile mathematical applications.

And now, what other exhibits from the wild jungle of variants and/or extensions
of standard FOL are equally worth knowing about at this stage, as you begin
studying mathematical logic? What other logics are intrinsically mathematically
interesting, have significant applications to mathematical reasoning, but can be
reasonably regarded as entry-level topics?
A good question. In this chapter, I’ll be looking at three relatively accessible
variant logics that philosophers in particular have discussed, namely relevant
logic, free logic and plural logic. And – spoiler alert! – I’m going to be suggesting
that mathematical logicians can cheerfully pass by the first, should have a fleeting
acquaintance with the second, and might like to pause a bit longer over the third.

11.1 Relevant logics

(a) Let’s concentrate here on one theme. The usual definition of logical con-
sequence makes an inference of the shape A, ¬A ∴ C come out valid, for any
135
 DRAFT– 31 DEC 2021
11 Other logics?

A and for any quite unconnected C – and correspondingly, in proof systems for
FOL, we can argue from the premisses A and ¬A to the arbitrary conclusion C.
But should we really count arguments as valid even when, as in this sort of case,
the premisses are totally irrelevant to the conclusion? Shouldn’t our formal logic
respect the intuitive idea – arguably already in Aristotle – that a conclusion in
a valid deduction must have something to do with the premisses?
Debates about this issue go back at least to medieval times. So let’s ask: what
might a suitable relevance-respecting logic look like? Is it worth the effort to use
such a logic?
(b) When we very first encounter it in Logic 101, the claim that A and ¬A
together entail any arbitrary conclusion C indeed initially seems odd. But we
soon learn that this result follows immediately from seemingly uncontentious
assumptions. Consider, in particular, these two principles:
Disjunctive syllogism is valid. From A ∨ C and ¬A we can infer C.
Entailment is transitive. In the simplest case, if A entails B and B
entails C, then A entails C. More generally, if Γ and ∆ stand in for
zero or more premisses, then if Γ entail B and ∆, B entail C, then
Γ, ∆ entail C
These indeed seem irresistible. Disjunctive syllogism is a principle we use all the
time in informal arguments (everyday ones and mathematical ones too). If we’ve
established that one of two options must hold, and can then rule out the first, this
surely establishes the second. And the transitivity of entailment is what allows
us to chain together shorter valid proofs to make longer valid proofs: reject it,
and it seems that the whole practice of proof in mathematics would collapse.
But now take the following three arguments:
P
P P∨Q ¬P
P∨Q ¬P
P∨Q Q
Q
The first just reflects our understanding of inclusive disjunction. The second is
the simplest of instances of disjunctive syllogism. The third argument chains
together the first two and, since they are valid entailments, this too is valid
according to the transitivity principle. So we have shown that P and ¬P entail
Q. And of course, we can generalize. In the same way, we can get from any pair
of premisses A and ¬A to an arbitrary conclusion C.
We have just three options, then:
1. Reject disjunctive syllogism as a universally valid principle (or at least, re-
ject disjunctive syllogism for the kind of disjunction for which the inference
A so A ∨ C is uncontentiously valid).

2. Reject the unrestricted transitivity of entailment.

3. Bite the bullet, and accept what is often called ‘explosion’, the principle
that from contradictory premisses we can infer anything at all.
136
 DRAFT– 31 DEC 2021
Relevant logics

The large majority of logicians take the first two options to be entirely unpalat-
able. So they conclude that we should indeed, as in standard FOL, learn to live
with explosion. And where’s the harm in that? After all, the explosive inference
can’t actually be used to take us from jointly true premisses to a false conclusion!
Still, before resting content with the explosive nature of FOL, perhaps we
should pause to see if there is any mileage in either option (1) or option (2).
What might a paraconsistent logic – one with a non-explosive entailment relation
– look like?
(c) Logicians are an ingenious bunch. And it isn’t difficult to cook-up a formal
system for e.g. a propositional language equipped with connectives written ∧, ∨
and ¬, for which analogues of disjunctive syllogism and explosion don’t generally
hold.
For example, suppose we adopt a natural deduction system with the usual
introduction and elimination rulers for ∧ and ∨ (as in §8.1). But the additional
rules governing negation are now just De Morgan’s Laws and a double negation
rule (the double inference lines indicate that you can apply the rules both top
to bottom and also the other way up).
¬(A ∧ B) ¬(A ∨ B) ¬¬A
(¬∧) (¬∨) (¬¬)
¬A ∨ ¬B ¬A ∧ ¬B A
The resulting logic is standardly called FDE for reasons that needn’t delay us.
And a little experimentation should convince you that, with only the FDE rules
in place, we can’t warrant either disjunctive syllogism or explosion.
But so what? By itself, the observation that dropping some classical rules stops
you proving some classical results has little interest. Compare the intuitionist
case, for example. There we are given a semantic story (the BHK account of the
meaning of the connectives) which aims to justify dropping the classical double
negation law. Can we similarly give a semantic story here which would again
justify dropping some classical rules and this time only underpin FDE ?
(d) Suppose – just suppose! – we think that there are four truth-related values
a proposition can take. Label these values T, B, N, F. And suppose that, given
an assignment of such values to atomic wffs, we compute the values of complex
wffs using the following tables:
A∧B T B N F A∨B T B N F A ¬A
T T B N F T T T T T T F
B B B F F B T B T B B B
N N F N F N T T N N N N
F F F F F F T B N F F T
These tables are to be read in the obvious way. So, for example, if P takes the
value B, and Q takes the value N, then P ∧ Q takes the value F, P ∨ Q takes the
value T, and ¬P takes the value B.
Suppose in addition that we define a quasi-entailment relation as follows: some
premisses Γ entail∗ a given conclusion C – in symbols Γ ∗ C – just if, on any

137
 DRAFT– 31 DEC 2021
11 Other logics?

valuation which makes each premiss either T or B, the conclusion is also either
T or B.
Then, lo and behold, we can show that FDE is sound and complete for this
semantics – we can derive C from premisses Γ if and only if Γ ∗ C. And
note, as we wanted, the analogue of disjunctive syllogism is not always a correct
entailment∗ : on the same suggested valuations, both P ∨ Q and ¬P are either T
or B, while Q is N, so P ∨ Q, ¬P 2∗ Q. And we don’t always get explosion either,
since both P and ¬P are B while Q is N, it follows that P, ¬P 2∗ Q.
Which is all fine and good in the abstract. But what are these imagined
four truth-related values? Can we actually give some interpretation so that our
tables really do have something to do with truth and falsity, with negation,
conjunction and disjunction, and so that entailment∗ does arguably become a
genuine consequence relation?
Well, suppose – just suppose! – that propositions can not only be plain true
or plain false but can also be both true and false at the same time, or neither
true nor false. Then there will indeed be four truth-related values a proposition
can take – T (true), B (both true and false), N (neither), F (false).
And, interpreting the values like that, the tables we have given arguably re-
spect the intuitive meaning of the connectives. For example, if A is both true
and false, the same should go for ¬A. While if A is both true and false, and B is
neither, then A ∨ B is true because its first disjunct is, but it isn’t also false as
that would require both disjuncts to be false (or so we might argue). Similarly
for the other table entries. Moreover, the intuitive idea of entailment as truth-
preservation is still reflected in the definition of entailment∗, which says that if
the premisses are all true (though maybe some are false as well), the conclusion
is true (though maybe false as well).
(e) What on earth can we make of this supposition that some propositions are
both true and false at the same time? At first sight, this seems simply absurd.
However, a vocal minority of philosophers do famously argue that while, to
be sure, regular sentences are either true or false but not both, there are certain
special cases – e.g. the likes of the paradoxical liar sentence ‘This sentence is
false’ – which are both true or false.
It is fair to say that rather few are persuaded by this extravagant suggestion.
But let’s go along with it just for a moment. And now note that it isn’t immedi-
ately clear that this really helps. For suppose we do countenance the possibility
that certain special sentences have the deviant status of being both true and
false (or being neither). Then we might reasonably propose to add to our formal
logical apparatus an operator ‘!’ to signal that a sentence is not deviant in that
way, an operator governed by the following table:
A !A
T T
B F
N F
F T

138
 DRAFT– 31 DEC 2021
Relevant logics

Why not? But then it is immediate that !P, P, ¬P ∗ Q. And similarly, if (say)
P and Q are the atoms present in A, then !P, !Q, A, ¬A ∗ C always holds.
So, if built out of regular atoms (expressing ordinary non-paradoxical claims),
a contradictory pair entails∗ anything. Yet surely, if we were seriously worried
by the original version of explosion, then this modified form will be no more
acceptable.
(f) We said that most logicians bite the bullet, and accept explosion because
they deem it harmless. But are they right?
It seems fundamental to a conditional connective → that it obeys the principle
of conditional proof. In other words, if the set of premisses Γ plus the temporary
assumption A together entail C, that shows that Γ entails A → C. But then
suppose we do accept the explosive inference from ¬A and A to C. Applying
conditional proof, we will have to agree that given ¬A, it follows that A → C,
for any unrelated consequent C, however irrelevant. And this, some will say, is
just the unacceptable face of the classical (or indeed intuitionistic) conditional:
so we should indeed reject explosion, not just for its prima facie oddity, but also
to get a nice conditional.
Now, if you have learnt to live happily with the standard conditional of classi-
cal or intuitionistic logic as an acceptable regimentation for serious mathematical
purposes, then you won’t be much moved by this argument. But what if you do
want to add a conditional connective where the inference from ¬A to A → C
generally fails?
Within an FDE -like framework, we can play with four-valued tables again,
now for the connective →. But on the more plausible ways of doing this, we
will still have !P, ¬P ∗ P → Q; and more generally, for wffs built out of regular
atoms, the conditional is just the material conditional again. So again, if we were
worried about the material conditional before, we should surely stay worried
about this sort of four-valued replacement.
(g) Let’s very briefly take stock.
We can run up proof systems like FDE which lack disjunctive syllogism and
explosion and where ¬A doesn’t imply A → C. Further, we can give these sys-
tems what looks like a semantics e.g. using four values (or alternatively we could
use Kripke-style valuations over some relational structure). But if this exercise
isn’t just to be an abstract game, then we do need to tell a story about how to
interpret the formal ‘semantics’ in order to link everything up with considera-
tions about truth and falsity and inference. And as we see in the initial case of
FDE, the supposed linkage can embroil us with highly implausible claims (e.g.
some propositions can be true and false – really?). Moreover, while our resulting
logic may not be classical overall, if we are allowed to distinguish regular true-or-
false propositions from those that behave deviantly according to the enhanced
semantic story, then in its application to the regular propositions, the new logic
can simply collapse back into classical logic again (with an entailment relation
and a conditional that don’t respect worries about relevance).
So already the price of avoiding exposition by rejecting disjunctive syllogism

139
 DRAFT– 31 DEC 2021
11 Other logics?

in the manner of FDE is beginning to look as if could be unattractively high

while the real gains remain pretty unclear.
But of course, all this is just an opening skirmish. There is a great deal more
than can be said, and which has been said, as you will find (to repeat, logicians
are an ingenious bunch). Though by my lights things only get worse when we
move on from the relatively simple FDE to fancier relevant logics such as the one
standardly called simply R. In the case of R, for example, the semantic story
is not superficially-clear-but-implausible (as for FDE ) but downright obscure
without any attractive motivation for ordinary logical use. Or so say most of us.
I’ll give readings on these sorts of semantically deviant relevant logics which
you can follow up if you want: but this is a rabbit hole that most mathematical
logicians very sensibly won’t want to disappear down. (I didn’t say that this
Guide would never be opinionated!)
(h) What about avoiding explosion not by rejecting disjunctive syllogism but
by rejecting the unrestricted transitivity of entailment? At first sight, this idea
might seem to be complete non-starter: as Timothy Smiley once put it, “the
whole point of logic as an instrument, and the way in which it brings us new
knowledge, lies in the contrast between the transitivity of ‘entails’ and the non-
transitivity of ‘obviously entails’, and all this is lost if transitivity cannot be
relied on.”
But perhaps, after all, there is wriggle-room here. Yes, in general, it is essential
to maintain the transitivity principle that if Γ entail B and ∆, B entail C, then
Γ, ∆ entail C. But what about the special case where Γ includes A while ∆
includes ¬A: shouldn’t that give us pause before we put Γ and ∆ together as
joint premisses? Rather than combining those explicitly inconsistent premisses
and arguing onwards regardless, shouldn’t we instead – so to speak – raise a red
flag, and declare that Γ, ∆ together are absurd, and only allow the inference from
Γ, ∆ to ⊥? In other words, the suggestion might go, transitivity holds except
when it shouldn’t, i.e. except when we have explicitly contradictory premisses on
the table and we should flag the absurdity. (So we can’t put the inference A to
A ∨ C together with the disjunctive syllogism from A ∨ C and ¬A to C to justify
the explosive entailment from A and ¬A to C: we should restrain ourselves and
stick to the inference from A and ¬A to ⊥.)
Now, compared with the proposal that we should achieve a relevant logic by
semantically deviant and rejecting disjunctive syllogism, this actually seems a
positively attractive suggestion. But can we actually develop the leading idea
into a smoothly workable logical system without its own oddities?
Well, Neil Tennant has long been arguing that we can indeed arrange things
so that we get very recognizable natural deduction rules but only the described
more restricted form of transitivity. In other words, we can get a proof system
in which we can paste proofs together when we ought to be able to, or else we
must combine the proofs to expose that we now can generate a contradiction.
And this, as Tennant emphasizes, looks like an epistemic plus-point, if we are
forced to highlight a contradiction when one is there to be exposed.

140
 DRAFT– 31 DEC 2021
Readings on relevant logic

Tennant advertises his proof system as core logic (actually there are two ver-
sions, one classical and one intuitionistic). His claim is that systems indeed cap-
ture the core of what we need in mathematical and scientific reasoning (classical
or constructive), without some of the unwanted extras. However, to avoid explo-
sion re-appearing, the operations of Tennant’s natural deduction system for his
core logic are inevitably subject to additional constraints on things like vacuous
discharge, as compared with the more free-wheeling proof-structures allowed in
standard systems for classical or intuitionistic systems. See the reading for more
details.
So here’s the obvious next question: is the occasional potential epistemic gain
from requiring proofs to obey the strictures of ‘core logic’ actually worth the
additional effort of strictly following its rules? A judgement call, of course. But
most mathematical logicians are going to return a negative verdict and, despite
Tennant’s energetic advocacy, feel quite comfortable on cost-benefit grounds of
sticking with their familiar ways.

11.2 Readings on relevant logic

A familiar resource once more provides some excellent entry-points:

1. Graham Priest, ‘Paraconsistent logic’, The Stanford Encyclopedia of

Philosophy, tinyurl.com/paracons. As Priest notes, any logical system
counts as paraconsistent as long as it is not explosive; there are a vari-
ety of motivations for a variety of paraconsistent systems. This is a very
clear introduction to some of the options.
2. Edwin Mares, ‘Relevance logic’, The Stanford Encyclopedia of Philoso-
phy, tinyurl.com/rel-logic. This, among other things, very usefully sum-
marizes a number of semantic interpretations that have been proposed
for relevant logics. Some depend on information-theoretic ideas that
might e.g. be of use in computer science: it is much less clear what
their significance for mathematical reasoning might be.

If you just want to know what it takes to get a relevance-respecting logic by the
route of semantic revisionism, these two pieces should suffice. You may well then
quickly decide that you don’t want to pay the price, being happy to accept the
verdict of e.g.
3. John Burgess, ‘No requirement of relevance’, in S. Shapiro, ed., The
Oxford Handbook of the Philosophy of Mathematics and Logic (OUP,
2005). (Initially, you can skip the later pages of §3, on Tennant.)
If, however, you are tempted to explore further, this is a terrific resource,
already familiar from the recommended readings on modal logic:
4. Graham Priest, An Introduction to Non-Classical Logic* (CUP, 2nd edi-
tion 2008). As we said before, this treats a whole range of logics system-
141
 DRAFT– 31 DEC 2021
11 Other logics?

atically, concentrating on semantic ideas, and using a tableaux approach.

Chs. 7–10 discuss some propositional many-valued logics (including ones
with truth-value ‘gaps’ and ‘gluts’), FDE, R, and much else besides: then
Chs. 21–24 discuss their quantificational counterparts.

And, taking a step up in level, here is the same author again vigorously making
the case for taking paraconsistent logics seriously:

5. Graham Priest, ‘Paraconsistent logic’, in the Handbook of Philosophical

Logic, Vol. 6, ed. by D. Gabbay and F. Guenthner, (Kluwer 2nd edition
2001), pp. 287–393.

You could also follow up Mares’s SEP article by taking a look at his book:

6. Edwin Mares, Relevant Logic: A Philosophical Interpretation (CUP 2004).

As the title suggests, this book has very extensive conceptual discussion
alongside the more formal parts elaborating what might be called the
mainstream tradition in relevance logics.

However, I for one am unpersuaded and remain on Burgess’s side of the debate,
at least as far as relevance-via-semantic-revisionism is concerned.
Going now in a very different direction, I mentioned in the previous section
Tennant’s idea of instead buying a certain amount of relevance by restricting
the transitivity of entailment. For a very lucid introductory account, see

7. Neil Tennant, ‘Relevance in reasoning’, in S. Shapiro, ed., The Oxford

Handbook of the Philosophy of Mathematics and Logic (OUP, 2005).

And for a full-blown development of these ideas, see

8. Neil Tennant, Core Logic (OUP, 2017). This tour-de-force is a rich book,
very well worth reading for its many more general proof-theoretic in-
sights, even if at the end of the day you don’t want to buy the relevantist
aspects.

In the final chapter, by the way, Tennant responds to the technical challenges
laid down by Burgess in §3 of his paper.

11.3 Free Logic

It is often said that pure logic should be topic-neutral. But FOL arguably isn’t
entirely topic-neutral. In particular it isn’t neutral about existence assumptions.
(a) Domains of quantification are assumed to be non-empty; (b) names are as-
sumed to have denotations, and ‘definite descriptions’ (constructions of the kind
the x such that Fx ) which might lack a denotation are massaged away; (c)
functions are assumed to be total, i.e. a value exists for any input. Does this
matter? Is it worth the effort to construct a suitable logic free of such existence
assumptions?

142
 DRAFT– 31 DEC 2021
Free Logic

(a) A reminder of a now familiar elementary point: in standard FOL, ∀xFx

entails ∃xFx. And here’s a Gentzen-style natural deduction derivation to prove
the point:
∀xFx
Fa
∃xFx
The first line states our premiss. At the second line, the story goes, we pick an
arbitrary member of the domain and dub it with a temporary name and then
infer . . .
But not so fast! What if the domain is empty? Then there is nothing to pick
out and dub.
So our natural deduction derivation at the second line in effect presupposes
that the domain is non-empty. Which ties in with the usual semantics for an FOL
language, where we stipulate that domains of quantification are always indeed
non-empty.
Deploying standard FOL to regiment a theory about Xs and using quantifiers
which range over Xs, then, makes an ontological assumption – namely, that there
are some Xs (at least one). For example, when we adopt the usual first-order
logical framework for doing formalized set theory, with quantifiers ranging over
sets, we are assuming that some sets exist (at least one) for our quantifiers to
range over.1
So: if we want to drop the existential presumption and allow for the possibility
that our domain of quantification is empty – in which case ∀xFx can be vacuously
true while ∃xFx is false – we’ll have to revise our logical laws. Should we bother?
Here’s a line of argument on one side:
An inference is logically valid just if it is necessarily truth-preserving
in virtue of topic-neutral features of its structure. And formal logic
is the study of logical validity, using regimented languages to enable
us to bring out how arguments of certain forms are valid irrespective
of their subject-matter.
Now, sometimes we want to argue logically about the properties
of things which we already know to exist (electrons, say). Other times
we want to argue in an exploratory way, in ignorance of whether what
we are talking about exists (superstrings, perhaps). While sometimes
we want to argue about things that we believe don’t exist, precisely
in order to try to show that they don’t exist (tachyons, perhaps). And
logic should aim to regiment correct forms of inference which we can
apply topic-neutrally across these different cases, without taking any
stance about how things are in the world.
1 Oliverand Smiley in their Plural Logic – about which more in the next section – have fun
chastising some set theorists for getting sloppy about this. For example, they quote J.R.
Shoenfield saying “we can use the usual axioms of logic to conclude that there is at least
one set”. But this is, strictly speaking, to get things exactly upside down: it is because we
have already presupposed that there is at least one set, that we can deploy the usual axioms
of FOL in doing formalized set theory.

143
 DRAFT– 31 DEC 2021
11 Other logics?

Hence one way our formal logic should be topic-neutral is by al-

lowing empty domains. But standard FOL rules – being incorrect for
empty domains – are not topic-neutral. So they don’t reliably cap-
ture only logical validities and logical truths. Therefore our standard
logic needs revision.

And how might the defender of our standard FOL logic reply?

There is no One True Logic. Choosing a formal logic always involves

weighing up costs and benefits. And the very small benefit of having
a logic whose inferential principles also hold in empty domains is just
not worth the albeit minor cost. After all, when we want to argue
about things that do not/might not exist, we already have sufficient
resources while still using standard logic.
First, a suitably inclusive wider domain is usually easily found
(indeed, will typically be in play when engaged in serious inquiry
rather than concocting artificial classroom examples). For example,
suppose we are arguing about tachyons. Instead of taking the domain
to be tachyons and regimenting the proposition that all tachyons are
really weird as ∀x Wx, we can more naturally take the domain more
inclusively to be, say, physical particles. We can then regiment that
proposition along the lines of ∀x(Tx → Wx) and lose the unwanted
inference that some really weird particles exists, ∃x Wx.
But put that manoeuvre aside. Suppose we want to adopt an
inclusive domain but we have lingering doubts about it legitimacy.
Then we can and do proceed in an exploratory, non-committal, sup-
positional mode. For example, consider mathematical inquiry which
proceeds in the supposedly all-inclusive framework of full-blown set
theory. What if we are sceptical about this world of sets? We can
bracket our set-theoretic investigations with an unspoken ‘Ok, let’s
take it, for the sake of argument, that there is this wildly infinitary
universe that standard set theory talks about . . . ’. And then, within
the scope of that bracketing assumption, we plunge in and quantify
over sets in the usual way, and continue our explorations as if we are
dealing with a suitably populated domain, to see where our investi-
gations get to. (Of course, if we start off assuming in a hypothetical
spirit that there are at least some Xs, our enquiries might lead us in
the end to backtrack and reject that assumption!)
Now, once we have made the supposition for the sake of further
exploration that there are sets (or superstrings or whatever Xs we
might be interested in), we might reasonably want the same logical
laws to apply in each case, topic-neutrally. But there is no need for
this logic we use, once we are working within the scope of the suppos-
ition that we are talking about something, to continue to remain
neutral about whether there is anything in the domain. The topic-

144
 DRAFT– 31 DEC 2021
Readings on free logic

neutrality can be downstream from the fundamental presumption

that we are talking about something rather than nothing.

The debate, all too predictably, will continue. But we have perhaps said enough
to explain why the usual view is that, particularly for the purposes of regimenting
mathematical reasoning, it is quite defensible to stick with a standard logic
(classical or intuitionist) which relies on the presumption that we aren’t talking
about nothing at all. See the reading, though, for how to give an inclusive version
of FOL which allows empty domains, if you do want one.
(b) Introductory subsections on free logic proper to be added!

11.4 Readings on free logic

Philosophers who want a gently paced introduction might appreciate the ap-
proachable treatment in

1. David Bostock, Intermediate Logic (OUP 1997), Ch. 8.

But for a more detailed overview, you want the very helpful

2. John Nolt, ‘Free logic’, The Stanford Encyclopedia of Philosophy, avail-

able at tinyurl.com/free-log.

Or even better:

3. John Nolt, ‘Free logics’, in D. Jacquette, ed., Philosophy of Logic: Hand-

book of the Philosophy of Science, Vol 5 (North-Holland 2007), pp. 1023-
1060. A more expansive essay covering the ground of the same author’s
SEP article.
This is a judicious and even-handed survey of some off the main issues
and options. Nolt writes “Though unsullied by existential commitment,
free logic does not reveal a tidy and compelling realm of logical truth. In
fact, the whole business is disappointingly messy.” But for all that, he
concludes that “In logic, as elsewhere, freedom, though messy, is often
desirable.”

And here’s a similar survey essay:

4. Ermanno Bencivenga, ‘Free Logics’, in D. Gabbay and F. Guenthner,

eds., Handbook of Philosophical Logic, vol. III: Alternatives to Classical
Logic (Reidel, 1986). Reprinted in D. Gabbay and F. Guenthner (eds.),
Handbook of Philosophical Logic, 2nd edition, vol. 5 (Kluwer 2002).

Moving on from general introductions to detailed formal treatments of various

kinds, the following are worth looking at:

145
 DRAFT– 31 DEC 2021
11 Other logics?

5. Neil Tennant, Natural Logic (Edinburgh UP 1978, 1990), §7.10. Available

at tinyurl.com/nat-logic. An early and original presentation of a free logic
in a natural deduction enviroment.
6. Graham Priest, An Introduction to Non-Classical Logic* (CUP, 2nd edi-
tion 2008), Ch. 13. As you would now expect, neatly and briskly pre-
sented tableau systems for various free logics.
7. Alex Oliver and Timothy Smiley, Plural Logic (OUP 2013: revised and
expanded second edition, 2016). Before giving formal systems for plural
logics in later chapters, Ch. 11 gives an original axiomatic free logic with
interesting features.

Finally, let me mention a collection of articles likely still to be of interest to

philosophers, around and about our topic: Karel Lambert, Free Logic: Selected
Essays (CUP 2003).

11.5 Plural logic

Introductory remarks about plural logic to be added

11.6 Readings on plural logic

For a gentle and discursive introduction, see

1. Salvatore Florio and Øystein Linnebo, The Many and the One (OUP 2021),
Chapter 2, ‘Taking plurals at face value’.

Then we have the excellent

2. Øystein Linnebo, ‘Plural Quantification’, The Stanford Encyclopedia of

Philosophy, tinyurl.com/pluralq

This is particularly lucid and helpful (though it would have been good to have,
perhaps as an appendix, a full-on, all-the-bells-and-whistles statement of the
rules for the natural deduction systems PFO and PLO+ , rather than a slightly
hands-off description, together with an axiomatic version too).
From the many papers which Linnebo mentions, if I have to choose two as
worth reading here at the outset, I’d perhaps pick these classics:

3. George Boolos, ‘To be is to be a value of a variable (or to be some values

of some variables)’, Journal of Philosophy (1984) pp. 430–50. Reprinted
in Boolos, Logic, Logic, and Logic (Harvard University Press, 1998).
4. Alex Oliver and Timothy Smiley, ‘Strategies for a logic of plurals’, Philo-
sophical Quarterly (2001) pp. 289–306.

146
 DRAFT– 31 DEC 2021
Readings on plural logic

Boolos’s paper is an influential early defence of the idea that taking plurals
seriously is logically important. Oliver and Smiley reinforce the point that there
is indeed a real topic here: you can’t readily eliminate all plural talk and plural
reasoning in favour e.g. of singular talk and reasoning about sets.
But now where? The book on Plural Predication by Thomas McKay (OUP
2006) is worth reading by philosophers for its discussion of non-distributive pred-
icates, plural descriptions etc. Then for logicians, there is the philosophically
argumentative, occasionally tendentious, and formally very rich tour de force

5. Alex Oliver and Timothy Smiley, Plural Logic (OUP 2013: revised and ex-
panded second edition, 2016).

However, Oliver and Smiley’s eventual logical system in their Chapter 13, ‘Full
plural logic’, will strike many as having (so to speak) unnecessarily many mov-
ing parts, as they aim – all at once – to accommodate empty domains, empty
names, a plural description operator, partial functions, multivalued functions,
even ‘copartial functions’ (which supposedly map nothing to something).
Oliver and Smiley, among others, make quite bold claims for plural logic. For
a critical look at such claims of defenders of plural logic, this is readable and
interesting:

6. Salvatore Florio and Øystein Linnebo, The Many and the One (OUP 2021),
Chapter 3 onwards. According to the blurb, this “provides a systematic anal-
ysis of the relation between this logic and other theoretical frameworks such
as set theory, mereology, higher-order logic, and modal logic. The applications
of plural logic rely on two assumptions, namely that this logic is ontologically
innocent and has great expressive power. These assumptions are shown to be
problematic.”

In particular, the argument – which applies already to simple systems like PLO
– is that the sort of comprehension principle which is built into plural logics is
problematic. Florio and Linnebo propose circumscribing comprehension.
Their book is approachable and argumentative. I in fact think some of Florio
and Linnebo’s arguments are resistible: see my comments on the first two parts
of the book, tinyurl.com/many-one. But well worth reading.

147
 DRAFT– 31 DEC 2021

12 Going further

This has been a Guide to beginning mathematical logic. So far, then, the sug-
gested readings on different areas have been at entry level, or only a step or so
up from that. In this final chapter, by contrast, we take a look at some of the
more advanced literature on a selection of topics, taking us another step or two
further.
If you have been tackling enough of the introductory readings, you should
in fact be able to now follow your interests wherever they lead, without really
needing help from this chapter. For a start, you can explore the many mathemat-
ical logic entries in The Stanford Encyclopedia of Philosophy, which are mostly
excellent and have large bibliographies. The substantial essays in the eighteen(!)
volumes of The Handbook of Philosophical Logic are of varying quality, but there
are some good ones on straight mathematical logic topics, again with large bibli-
ographies. Internet sites like math.stackexchange.com and the upper-level math-
overflow.net can be searched for useful lists of recommended books. And then
there is always Google!
However, those resources do cumulatively point to a rather overwhelming
range of literature to pursue. So perhaps some readers will still appreciate a few
more limited menus of suggestions (even if they are less systematic and more
shaped by my personal interests than in the core Guide).
Of course, the ‘vertical’ divisions between entry-level coverage and the further
explorations in this chapter are pretty arbitrary; and the ‘horizontal’ divisions
into different subfields can in places also be quite blurred. But we do need to
impose some organization! So this chapter is divided up as follows. First, we
make a very brief foray into logic-relevant algebra:

12.1 A very little light algebra for logic?

There follows a series of sections taking up the core topics of Chapters 5–7 and 9
in the same order as before:

12.2 More model theory

12.3 More on formal arithmetic and computability
12.4 More on mainstream set theory
12.5 Choice, and the choice of set theory
12.6 More proof theory.

148
 DRAFT– 31 DEC 2021
A very little light algebra for logic?

Then there is a final section which introduces a further topic area which is the
focus of considerable recent interest:

12.7 Higher-order logic, the lambda calculus, and type theory.

We could continue; but this is more than enough to be going on with . . . !

12.1 A very little light algebra for logic?

Depending on what you have read on classical propositional logic, you may well
have touched on the notion of a Boolean algebra. And depending on what you
have read on intuitionistic logic, you may have also also encountered Heyting
algebras (a.k.a. pseudo-Boolean algebras). It is worth getting to know a bit more
about these algebras, both because of their relevance to classical and intuitionis-
tic logic, but also because Boolean algebra features in independence arguments
in set theory.
For a gentle and clear first introduction (aimed at those with little mathemat-
ical background), see

1. Barbara Hall Partee, Alice G. B. ter Meulen, and Robert Eugene Wall,
Mathematical Methods in Linguistics (1990, Springer). The (short!) Chs.
9 and 10 introduce some basic concepts of algebra (you can omit §10.3);
Ch. 11 is on lattices; Ch. 12 is then on Boolean and Heyting algebras,
and briefly connects Kripke’s relational semantics for intuitionistic logic
to Heyting algebras.

Also very accessible, for adding a little more on Heyting algebras:

2. Morten Heine Sørensen and Pawel Urzyczyn, Lectures on the Curry-

Howard Isomorphism (Elsevier, 2006), Ch. II, ‘Intuitionistic logic’.

Then, for rather more about Boolean algebras, you need very little background
to start tackling the opening chapters of

3. Steven Givant and Paul Halmos, Introduction to Boolean Algebras (Sprin-

ger, 2009). This is an update of a classic book by Halmos, and is very
accessible; any logician will want eventually to know the elementary ma-
terial in the first third of the book.

If you already know a smidgin of algebra and topology, however, then there
is a faster-track introduction to Boolean algebras in

4. René Cori and Daniel Lascar, Mathematical Logic, A Course with Exer-
cises: Part I (OUP, 2000), Chapter 2.

And for a higher-level treatment of intuitionistic logic and Heyting algebras, you
could read Chapter 5 of the book by Dummett mentioned in §8.5, or work up
to Chapter 7 on algebraic semantics in the book on modal logic by Chagrov and
Zakharyaschev mentioned in §10.5.
149
 DRAFT– 31 DEC 2021
12 Going further

Then, if you want to pursue more generally e.g. questions about when propo-
sitional logics do have nice algebraic counterparts (in the sort of way that classi-
cal and intuitionistic logic relate respectively to Boolean and Heyting Algebras),
then you might get something out of Ramon Jansana’s ‘Algebraic propositional
logic’ in The Stanford Enclyclopedia of Philosophy, tinyurl.com/alg-logic. But this
does strike me as too rushed to be particularly useful. So instead, you could make
a start reading

5. Josep Maria Font, Abstract Algebraic Logic: An Introductory Textbook

(College Publications, 2016). This is written in an expansive and acces-
sible style, and well worth diving into.

12.2 More model theory

(a) If you want to explore beyond the entry-level material of Chapter 5 on
model theory, why not start with a quick warm-up, with some reminders of
headlines and some very useful pointers to the road ahead:

1. Wilfrid Hodges and Thomas Scanlon, ‘First-order model theory’, The

Stanford Encyclopedia of Philosophy, tinyurl.com/sep-fo-model.

Now, we noted before in §§3.6(c) and 5.3 that the wide-ranging mathematical
logic texts by Hedman and Hinman cover a substantial amount of model theory.
But why not look at two classic stand-alone treatments of the area which really
choose themselves? In order of both first publication and eventual difficulty:

2. C. Chang and H. J. Keisler, Model Theory* (originally North Holland

1973: the third edition has been inexpensively republished by Dover
Books in 2012). This is the Old Testament, the first systematic text on
model theory. Over 550 pages long, it proceeds at an engagingly leisurely
pace. It is particularly lucid and is extremely nicely constructed with
different chapters on different methods of model-building. A really fine
achievement that still remains a good route in to the serious study of
model theory.
3. Wilfrid Hodges, A Shorter Model Theory (CUP, 1997). The New Testa-
ment is Hodges’s encyclopedic Model Theory (CUP 1993). This shorter
version is half the size but still really full of good things. It does get
tougher as the book progresses, but the earlier chapters of this modern
classic, written with this author’s characteristic lucidity, should certainly
be readily manageable.

My suggestion would be to read the first three long chapters of Chang and
Keisler, and then perhaps pause to make a start on

4. J. L. Bell and A. B. Slomson, Models and Ultraproducts* (North-Holland

1969; Dover reprint 2006). Very elegantly put together: as the title sug-
gests, the book focuses particularly on the ultra-product construction.
150
 DRAFT– 31 DEC 2021
More model theory

At this point read the first five chapters for a particularly clear intro-
duction.

You could then return to Ch. 4 of C&K to look at (some of) their treatment of
the ultra-product construction, before perhaps putting the rest of their book on
hold and turning to Hodges.
(b) A level up again, here are two further books that should definitely be
mentioned. The first has been around long enough to have become regarded as
a modern standard text. The second is a bit more recent but also comes widely
recommended. Their coverage is significantly different – so I suppose that those
wanting to get really seriously into model theory should take a look at both:

5. David Marker, Model Theory: An Introduction (Springer 2002). Despite

its title, this book would surely be hard going if you haven’t already
tackled some model theory (at least read Manzano or Kirby first). But
despite being sometimes a rather bumpy ride, this highly regarded text
will teach you a great deal. Later chapters, however, probably go far over
the horizon for all except those most enthusiastic readers of this Guide
who are beginning to think about specializing in model theory – it isn’t
published in the series ‘Graduate Texts in Mathematics’ for nothing!
6. Katrin Tent and Martin Ziegler, A Course in Model Theory (CUP, 2012).
From the blurb: “This concise introduction to model theory begins with
standard notions and takes the reader through to more advanced topics
such as stability . . . . The authors introduce the classic results, as well
as more recent developments in this vibrant area of mathematical logic.
Concrete mathematical examples are included throughout to make the
concepts easier to follow.” Again, although it starts from the beginning,
it could be a challenge to readers without some mathematical sophistica-
tion and some prior exposure to the elements of model theory – though
I, for one, find it more approachable than Marker’s book.

(c) So much for my principal suggestions. Now for an assortment of addi-

tional/alternative texts. Here are two more books which aim to give general
introductions:

7. Philipp Rothmaler’s Introduction to Model Theory (Taylor and Francis

2000) is, overall, comparable in level of difficulty with, say, the first half
of Hodges. As the blurb puts it: “This text introduces the model theory
of first-order logic, avoiding syntactical issues not too relevant to model
theory. In this spirit, the compactness theorem is proved via the alge-
braically useful ultraproduct technique (rather than via the complete-
ness theorem of first-order logic). This leads fairly quickly to algebraic
applications, ... .” Now, the opening chapters are indeed very clear: but
oddly the introduction of the crucial ultraproduct construction in Ch. 4
is done very briskly (compared, say, with Bell and Slomson). And there-
after it seems to me that there is some unevenness in the accessibility
151
 DRAFT– 31 DEC 2021
12 Going further

of the book. But others have recommended this text more warmly, so I
mention it as a possibility worth checking out.
8. Bruno Poizat’s A Course in Model Theory (English edition, Springer
2000) starts from scratch and the early chapters give an interesting and
helpful account of the model-theoretic basics, and the later chapters
form a rather comprehensive introduction to stability theory. This often-
recommended book is written in a rather distinctive style, with rather
more expansive class-room commentary than usual: so an unusually en-
gaging read at this sort of level.

Another book which is often mentioned in the same breath as Poizat, Marker,
and now Tent and Ziegler is A Guide to Classical and Modern Model Theory, by
Annalisa Marcja and Carlo Toffalori (Kluwer, 2003) which also covers a lot: but
I prefer the previously listed books.
The next two suggestions are of books which are helpful on particular aspects
of model theory:

9. Kees Doets’s short Basic Model Theory* (CSLI 1996) highlights so-called
Ehrenfeucht games. This is enjoyable and very instructive.
10. Chs. 2 and 3 of Alexander Prestel and Charles N. Delzell’s Mathematical
Logic and Model Theory: A Brief Introduction (Springer 1986, 2011)
are brisk but clear, and can be recommended if you wanting a speedy
review of model theoretic basics. The key feature of the book, however, is
the sophisticated final chapter on serious applications to algebra, which
might appeal to mathematicians with interests in that area.

Indeed, as we explore model theory, we quickly get entangled with algebraic

questions. And as well as going (so to speak) in the direction from logic to
algebra, we can make connections the other way about, starting from algebra.
For something on this approach, see the following short, relatively accessible,
and illuminating book:

11. Donald W. Barnes and John M. Mack, An Algebraic Introduction to

Mathematical Logic (Springer, 1975).

(d) As an aside, let me also mention the sub-area of Finite Model Theory which
arises particularly from consideration of problems in the theory of computation
(where, of course, we are interested in finite structures – e.g. finite databases
and finite computations over them). What happens, then, to model theory if we
restrict our attention to finite models? Trakhtenbrot’s theorem, for example, tells
that the class of sentences true in any finite model is not recursively enumerable.
So there is no deductive theory for capturing such finitely valid sentences (that’s
a surprise, given that there’s a complete deductive system for the sentences which
are valid in the usual broader sense!). It turns out, then, that the study of finite
models is surprisingly rich and interesting. So why not dip into one or other of

12. Leonard Libkin, Elements of Finite Model Theory (Springer 2004).

152
 DRAFT– 31 DEC 2021
More on formal arithmetic and computability

13. Heinz-Dieter Ebbinghaus and Jörg Flum, Finite Model Theory (Springer
2nd edn. 1999).

Both are good, though I prefer Libkin.

(e) In §5.3 I warmly recommended that you read at least early chapters of
Philosophy and Model Theory by Button and Walsh. Now you know more model
theory, do revisit that book and read on!
Finally, I should mention John T. Baldwin’s Model Theory and the Philosophy
of Mathematical Practice (CUP, 2018). This presupposes a lot more background
than Button and Walsh. Maybe some philosophers might be able to excavate
more out of Baldwin’s book than I did: but I find this book badly written and
unnecessarily hard work.

12.3 More on formal arithmetic and computability

(a) The readings in §6.5 have introduced you to the canonical first-order theory
of arithmetic, first-order Peano Arithmetic, as well as to some subsystems of PA
(in particular, Robinson Arithmetic) and second-order extensions. So what to
read next on formal arithmetics?
You will know by now that first-order PA has non-standard models: in fact,
it even has uncountably many non-isomorphic models which can be built just
out of natural numbers. It is worth pursuing this theme. For a taster, you could
look at lecture notes by Jaap van Oosten, on ‘Introduction to Peano Arithmetic:
Gödel Incompleteness and Nonstandard Models’, tinyurl.com/oosten-peano. But
better to dive into

1. Richard Kaye’s Models of Peano Arithmetic (Oxford Logic Guides, OUP,

1991), which tells us a great deal about non-standard models of PA. This
reveals more about what PA can and can’t prove, and will also introduce
you to some non-Gödelian examples of incompleteness. This is a terrific
book, and deservedly a modern classic.

As a sort of sequel, there is also another volume in the Oxford Logic Guides series
for enthusiasts with more background in model theory, namely Roman Kossak
and James Schmerl, The Structure of Models of Peano Arithmetic, OUP, 2006.
But this is much tougher going. For a more accessible set of excellent lecture
notes, see

2. Tin Lok Wong, ‘Model theory of arithmetic’, downloadable lecture by

lecture from tinyurl.com/wong-model.

Next, going in a rather different direction, and explaining a lot about arith-
metics weaker than full PA, here’s another modern classic:

3. Petr Hájek and Pavel Pudlák, Metamathematics of First-Order Arith-

metic (Springer 1993). This is pretty encyclopaedic, but at least the first

153
 DRAFT– 31 DEC 2021
12 Going further

three chapters do remain surprisingly accessible for such a work. This

is, eventually, a must-read if you have a serious interest in theories of
arithmetic and incompleteness.

And what about going beyond first-order PA? We know that full second-
order PA (where the second-order quantifiers are constrained to run over all
possible sets of numbers) is unaxiomatizable, because the underlying second-
order logic is unaxiomatiable. But there are axiomatizable subsystems of second
order arithmetic. These are wonderfully investigated in another encyclopaedic
modern classic:

4. Stephen Simpson, Subsystems of Second-Order Logic (Springer 1999; 2nd

edn CUP 2009). The focus of this book is the project of ‘reverse mathe-
matics’ (as it has become known): that is to say, the project of identifying
the weakest theories of numbers-and-sets-of-numbers that are required
for proving various characteristic theorems of classical mathematics.
We know that we can reconstruct classical analysis in pure set theory,
and rather more neatly in set theory with natural numbers as unanal-
ysed ‘urelemente’. But just how much set theory is needed to do the
job, once we have the natural numbers? The answer is: stunningly little.
The project of exploring what’s needed is introduced very clearly and
accessibly in the first chapter, which is a must-read for anyone interested
in the foundations of mathematics. This introduction is freely available
at the book’s website tinyurl.com/2arith.

(b) Next, Gödelian incompleteness again. You could start with a short old
Handbook article which is still well worth reading:

5. Craig Smoryński, ‘The incompleteness theorems’, in J. Barwise, editor,

Handbook of Mathematical Logic, pp. 821–865 (North-Holland, 1977),
which covers a lot very compactly. Available at tinyurl.com/smory.

Now, the further readings on incompleteness suggested in §6.6 finished by

mentioning two wonderful books which could arguably have appeared on our
main list of introductory readings. However – a judgement call – I think that the
more abstract stories they tell can probably only be fully appreciated if you’ve
first met the basics of computability theory and the incompleteness theorems in
a more conventional treatment. But certainly, now is the time to read them, if
you didn’t tackle them before:

6. Raymond Smullyan, Gödel’s Incompleteness Theorems, Oxford Logic

Guides 19 (Clarendon Press, 1992). Proves beautiful, slightly abstract,
versions of the incompleteness theorems. A modern classic.
7. Equally short and equally elegant is Melvin Fitting’s, Incompleteness in
the Land of Sets* (College Publications, 2007). There is a simple cor-
respondence between natural numbers and ‘hereditarily finite sets’ (i.e.

154
 DRAFT– 31 DEC 2021
More on formal arithmetic and computability

sets which have a finite number of members which in turn have a finite
number of members which in turn . . . where all downward membership
chains bottom out with the empty set). Relying on this fact gives us
another route in to proofs of Gödelian incompleteness, and other results
of Church, Rosser and Tarski. Beautifully done.

After these, where should you go if you want to know more about matters
more or less directly to do with the incompleteness theorems?

8. Raymond Smullyan’s Diagonalization and Self-Reference, Oxford Logic

Guides 27 (Clarendon Press 1994) is an investigation-in-depth around
and about the idea of diagonalization that figures so prominently in
proofs of limitative results like the unsolvability of the halting problem,
the arithmetical undefinability of arithmetical truth, and the incomplete-
ness of arithmetic. Read at least Part I.
9. Torkel Franzén, Inexaustibility: A Non-exhaustive Treatment (Associa-
tion for Symbolic Logic/A. K. Peters, 2004). The first two-thirds of the
book gives another take on logic, arithmetic, computability and incom-
pleteness. The last third notes that Gödel’s incompleteness results have
a positive consequence: ‘any system of axioms for mathematics that we
recognize as correct can be properly extended by adding as a new ax-
iom a formal statement expressing that the original system is consistent.
This suggests that our mathematical knowledge is inexhaustible, an es-
sentially philosophical topic to which this book is devoted.’ Not always
easy (you will need to know something about ordinals before you read
this), but very illuminating.
10. Per Lindström, Aspects of Incompleteness (Association for Symbolic
Logic/ A. K. Peters, 2nd edn., 2003). This rather terse book is probably
for enthusiasts. It is not always reader-friendly in its choices of nota-
tion and the brevity of its arguments. However, the more mathematical
reader will find that it again repays the effort.
11. Craig Smoryński, Logical Number Theory I, An Introduction (Springer,
1991). There are three long chapters. Ch. I discusses pairing functions
and numerical codings, primitive recursion, the Ackermann function,
computability, and more. Ch. II concentrates on ‘Hilbert’s tenth prob-
lem’ – showing that we can’t mechanically decide the solubility of certain
equations. Ch. III considers Hilbert’s Programme and contains proofs
of more decidability and undecidability results, leading up to a version
of Gödel’s First Incompleteness Theorem. (The promised Vol. II which
would have discussed the Second Incompleteness Theorem has never ap-
peared.)
The level of difficulty is rather varied, and there are a lot of historical
disgressions and illuminating asides. So this is an idiosyncratic book; but
is still an enjoyable and very instructive read.

155
 DRAFT– 31 DEC 2021
12 Going further

And if you want the bumpier ride of a lecture course with problems assigned as
you go along, this is notable:

12. Tin Lok Wong, ‘The consistency of arithmetic’, downloadable lecture by

lecture from tinyurl.com/wong-consis.

(c) Now let’s turn to books on computability. Among the Big Books on math-
ematical logic, the one with the most useful treatment is probably

13. Peter G. Hinman, Fundamentals of Mathematical Logic (A. K. Peters,

2005). Chs. 4 and 5 on recursive functions, incompleteness etc. strike me
as the best written, most accessible (and hence most successful) chapters
in this very substantial book. The chapters could well be read after
my IGT as somewhat terse revision for mathematicians, and then as
sharpening the story in various ways. Ch. 8 then takes up the story of
recursion theory (the author’s home territory).

However, good those these chapters are, I’d still recommend starting your more
advanced work on computability with

14. Nigel Cutland, Computability: An Introduction to Recursive Function

Theory (CUP 1980). This is a rightly much-reprinted classic and is beau-
tifully lucid and well-organized. This does have the look-and-feel of a
traditional maths text book of its time (so perhaps with fewer of the
classroom asides we find in some modern, more discursive books). How-
ever, if you got through most of e.g. Boolos and Jeffrey without too much
difficulty, you ought certainly to be able to tackle this as the next step.
Very warmly recommended.

And of more recent books covering computability at this level, I also particularly
like

15. S. Barry Cooper, Computability Theory (Chapman & Hall/CRC 2003).

A very nicely done modern textbook. Read at least Part I of the book
(about the same level of sophistication as Cutland, but with some extra
topics), and then you can press on as far as your curiosity takes you, and
get to excitements like the Friedberg-Muchnik theorem.

Of course, the inherited literature on computability is huge. But, being very

selective, let me mention three classics from different generations:

16. Rósza Péter, Recursive Functions (originally published 1950: English

translation Academic Press 1967). This is by one of those logicians who
was ‘there at the beginning’. It has that old-school slow-and-steady un-
flashy lucidity that makes it still a considerable pleasure to read. It re-
mains very worth looking at.
17. Hartley Rogers, Jr., Theory of Recursive Functions and Effective Com-
putability (McGraw-Hill 1967) is a heavy-weight state-of-the-art-then
156
 DRAFT– 31 DEC 2021
More on formal arithmetic and computability

classic, written at the end of the glory days of the initial development of
the logical theory of computation. It quite speedily gets advanced. But
the actin-packed opening chapters are excellent. At least take it out of
the (e)library, read a few chapters, and admire!
18. Piergiorgio Odifreddi, Classical Recursion Theory, Vol. 1 (North Holland,
1989) is well-written and discursive, with numerous interesting asides.
It’s over 650 pages long, so it goes further and deeper than other books
on the main list above (and then there is Vol. 2). But it certainly starts off
quite gently paced and very accessible and can be warmly recommended
for consolidating and then extending your knowledge.
(d) Classical computability theory abstracts away from considerations of prac-
ticality, efficiency, etc. Computer scientists are – surprise, surprise! – interested
in the theory of feasible computation, and any logician should be interested in
finding out at least a little about the topic of computational complexity. Here
are three introductions to the topic, in order of increasing detail:
19. Herbert E. Enderton, Computability Theory: An Introduction to Recu-
sion Theory (Associated Press, 2011). Chapter 7.
20. Shawn Hedman A First Course in Logic (OUP 2004): Ch. 7 on ‘Com-
putability and complexity’ has a nice review of basic computability the-
ory before some lucid sections discussing computational complexity.
21. Michael Sipser, Introduction to the Theory of Computation (Thomson,
2nd edn. 2006) is a standard and very well regarded text on computation
aimed at computer scientists. It aims to be very accessible and to take its
time giving clear explanations of key concepts and proof ideas. I think
this is very successful as a general introduction and I could well have
mentioned the book before. But I’m highlighting the book now because
its last third is on computational complexity.
And for more expansive, stand-alone treatments, here are three more suggestions:
22. I don’t mention many sets of lecture notes in this Guide, as they tend
to be rather too terse for self-study. But Ashley Montanaro has an ex-
cellent and extensive lecture notes on Computational Complexity, lucid
and detailed. Available at tinyurl.com/cocomp.
23. Oded Goldreich, P, NP, and NP-Completeness (CUP, 2010). Short,
clear, and introductory stand-alone treatment.
24. You could also look at the opening chapters of the pretty encyclopaedic
Sanjeev Arora and Boaz Barak Computational Complexity: A Modern
Approach (CUP, 2009). The authors say that ‘[r]equiring essentially no
background apart from mathematical maturity, the book can be used as
a reference for self-study for anyone interested in complexity, including
physicists, mathematicians, and other scientists, as well as a textbook for
a variety of courses and seminars.’ And at least it starts very readably! A
late draft of the book can be freely downloaded from tinyurl.com/arora.
157
 DRAFT– 31 DEC 2021
12 Going further

12.4 More on mainstream set theory

(a) Some of the readings on set theory suggested in Chapter 7 were beginning
to get quite sophisticated: but still, we weren’t tangling with more advanced
topics like ‘large cardinals’ and ‘forcing’. Now we move on.
And one option is immediately to go for broke and dive in to the modern
bible, which is highly impressive not just for its size:

1. Thomas Jech, Set Theory, The Third Millennium Edition (Springer,

2003). The book is in three parts: the first, Jech says, every student
should know; the second part every budding set-theorist should master;
and the third consists of various results reflecting ‘the state of the art of
set theory at the turn of the new millennium’. Start at page 1 and keep
going to page 705 – or until you feel glutted with set theory, whichever
comes first!

This book is indeed a masterly achievement by a great expositor. And if you’ve

happily read e.g. the introductory books by Enderton and then Moschovakis
mentioned earlier in the Guide, then you should be able to cope pretty well with
Part I of the book while it pushes on the story a little with some material on
‘small large cardinals’ and other topics. Part II of the book starts by telling
you about independence proofs. The Axiom of Choice is consistent with ZF and
the Continuum Hypothesis is consistent with ZFC, as proved by Gödel using
the idea of ‘constructible’ sets. And the Axiom of Choice is independent of ZF,
and the Continuum Hypothesis is independent with ZFC, as proved by Cohen
using the much more tricky idea of ‘forcing’. The rest of Part II tells you more
about large cardinals, and about descriptive set theory. Part III is indeed for
enthusiasts.
(b) Now, Jech’s book is wonderful, but let’s face it, the sheer size makes it a
trifle daunting. It goes quite a bit further than many will need, and to get there
it in places speeds along a bit faster than some will feel comfortable with. So
what other options are there for if you want to take things more slowly?
Let’s start with a book which I mentioned in passing in §7.6:

2. Azriel Levy, Basic Set Theory* (Springer 1979, republished by Dover

2002). This is ‘basic’ in the sense of not dealing with topics like forcing.
However it is a quite advanced-level treatment of the set-theoretic fun-
damentals at least in its mathematical style, and even the earlier parts
are I think best tackled once you know some set theory (they could be
very useful, though, as a rigorous treatment consolidating the basics – a
reader comments that Levy’s is his “go to” book when he needs to check
set theoretical facts that don’t involve forcing or large cardinals.). The
last part of the book starts on some more advanced topics.

Levy’s book ends with a discussion of some ‘large cardinals’. However another
much admired older book remains the recommended first treatment of this topic:
158
 DRAFT– 31 DEC 2021
More on mainstream set theory

3. Frank R. Drake, Set Theory: An Introduction to Large Cardinals (North-

Holland, 1974). This overlaps with Part I of Jech’s bible, though at per-
haps a gentler pace. But it also will tell you about Gödel’s Constructible
Universe and then some more about large cardinals. Very lucid.

For some other topics you could also look at the second volume of a book whose
first instalment was a main recommendation in §7.2:

4. Winfried Just and Martin Weese, Discovering Modern Set Theory II:
Set-Theoretic Tools for Every Mathematician (American Mathematical
Society, 1997).
This contains, as the authors put it, “short but rigorous introductions
to various set-theoretic techniques that have found applications outside
of set theory”. Some interesting topics, and can be read independently
of Vol. I.

(c) But now the crucial next step – that perhaps marks the point where set
theory gets really challenging – is to get your head around Cohen’s idea of forcing
used in independence proofs. However, there is not getting away from it, this is
tough. In the admirable

5. Timothy Y. Chow, ‘A beginner’s guide to forcing’, tinyurl.com/chowf

Chow writes:

All mathematicians are familiar with the concept of an open research

problem. I propose the less familiar concept of an open exposition
problem. Solving an open exposition problem means explaining a
mathematical subject in a way that renders it totally perspicuous.
Every step should be motivated and clear; ideally, students should
feel that they could have arrived at the results themselves. The proofs
should be ‘natural’ . . . [i.e., lack] any ad hoc constructions or brillian-
cies. I believe that it is an open exposition problem to explain forcing.

In short: if you find that expositions of forcing – including Chow’s – tend to be

hard going, then join the club.
Here though is a very widely used and much reprinted textbook, which nicely
complements Drake’s book and which has (inter alia) a relatively approachable
introduction to forcing arguments:

6. Kenneth Kunen, Set Theory: An Introduction to Independence Proofs

(North Holland, 1980). If you have read (some of) the introductory set
theory books mentioned in the Guide, you should actually find much of
this text now pretty accessible, and can probably speed through some of
the earlier chapters, slowing down later, until you get to the penultimate
chapter on forcing which you’ll need to take slowly and carefully. This is
a rightly admired classic text.

159
 DRAFT– 31 DEC 2021
12 Going further

Kunen has since published another, totally rewritten, version of this book as
Set Theory* (College Publications, 2011). This later book is quite significantly
longer, covering an amount of more difficult material that has come to promi-
nence since 1980. Not just because of the additional material, my current sense
is that the earlier book may remain the somewhat gentler read.
Now, Kunen’s classic text takes a ‘straight down the middle’ approach, start-
ing with what is basically Cohen’s original treatment of forcing, though he does
relate this to some other approaches. Here are two of them:

7. Raymond Smullyan and Melvin Fitting, Set Theory and the Continuum
Problem (OUP 1996, Dover Publications 2010). This medium-sized book
is divided into three parts. Part I is a nice introduction to axiomatic set
theory (in fact, officially in its NBG version – see §12.5). The shorter
Part II concerns matters round and about Gödel’s consistency proofs via
the idea of constructible sets. Part III gives a different take on forcing.
This is beautifully done, as you might expect from two writers with
a quite enviable knack for wonderfully clear explanations and an eye for
elegance.
8. Keith Devlin, The Joy of Sets (Springer 1979, 2nd edn. 1993) Ch. 6 intro-
duces the idea of Boolean-Valued Models and their use in independence
proofs. The basic idea is fairly easily grasped, but the details perhaps
trickier.
For more on this theme, see John L. Bell’s classic Set Theory: Boolean-
Valued Models and Independence Proofs (Oxford Logic Guides, OUP, 3rd
edn. 2005). The relation between this approach and other approaches to
forcing is discussed e.g. in Chow’s paper and the last chapter of Smullyan
and Fitting.

(d) Here is a selection of another four books with various virtues, in order of
publication:

9. Akihiro Kanamori, The Higher Infinite: Large Cardinals in Set Theory

from Their Beginnings (Springer, 1997, 2nd edn. 2003). This blockbuster
is subtitled ‘Large Cardinals in Set Theory from Their Beginnings’, and is
very clearly put together with a lot of helpful and illuminating historical
asides. A classic.
10. Lorenz J. Halbeisen, Combinatorial Set Theory, With a Gentle Intro-
duction to Forcing (Springer 2011). From the blurb “This book provides
a self-contained introduction to modern set theory and also opens up
some more advanced areas of current research in this field. The first part
offers an overview of classical set theory wherein the focus lies on the
axiom of choice and Ramsey theory. In the second part, the sophisticated
technique of forcing, originally developed by Paul Cohen, is explained in
great detail. With this technique, one can show that certain statements,
like the continuum hypothesis, are neither provable nor disprovable from

160
 DRAFT– 31 DEC 2021
Choice, and the choice of set theory

the axioms of set theory. In the last part, some topics of classical set
theory are revisited and further developed in the light of forcing.”
True, this book gets quite hairy towards the end: but the earlier parts
of the book should be much more accessible. This book has been strongly
recommended for its expositional merits by more reliable judges than me;
but I confess I didn’t find it notably more successful than other accounts
of forcing. A late draft of the book is available: tinyurl.com/halb-set.
11. Nik Weaver, Forcing for Mathematicians (World Scientific, 2014) is less
than 150 pages (and the first applications of the forcing idea appear
after just 40 pages: you don’t have to read the whole book to get the
basics). From the blurb: “Ever since Paul Cohen’s spectacular use of the
forcing concept to prove the independence of the continuum hypothesis
from the standard axioms of set theory, forcing has been seen by the
general mathematical community as a subject of great intrinsic interest
but one that is technically so forbidding that it is only accessible to spe-
cialists ... This is the first book aimed at explaining forcing to general
mathematicians. It simultaneously makes the subject broadly accessible
by explaining it in a clear, simple manner, and surveys advanced appli-
cations of set theory to mainstream topics.” This does strike me as a
helpful attempt to solve Chow’s basic exposition problem, to explain the
Big Ideas very directly.
12. Ralf Schindler, Set Theory: Exploring Independence and Truth (Springer,
2014). The book’s theme is “the interplay of large cardinals, inner mod-
els, forcing, and descriptive set theory”. It doesn’t presume you already
know any set theory, though it does proceed at a cracking pace in a
brisk style. But, if you already have some knowledge of set theory, this
seems a clear and interesting exploration of some themes highly relevant
to current research.

12.5 Choice, and the choice of set theory

But now let’s leave the Higher Infinite and other excitements and get back down
to earth, or at least to less exotic topics! And, to return to the beginning, we
might wonder: is ZFC the ‘right’ set theory? Indeed, how do we choose which
set theory to adopt?
(a) Let’s start by thinking about the Axiom of Choice in particular. It is com-
forting to know from Gödel that AC is consistent with ZF (so adding it doesn’t
lead to contradiction). But we also know from Cohen’s forcing argument that
AC is independent with ZF (so accepting ZF doesn’t commit you to accepting
AC too). So why buy AC? Is it an optional extra?
Quite a few of the readings already mentioned will have touched on the ques-
tion of AC’s status and role. But for a useful overview/revision of some basics,
see

161
 DRAFT– 31 DEC 2021
12 Going further

1. John L. Bell, ‘The axiom of choice’, The Stanford Encyclopedia of Phi-

losophy, tinyurl.com/sep-axch.

And for a short book also explaining some of the consequences of AC (and some
of the results that you need AC to prove), see

2. Horst Herrlich, Axiom of Choice (Springer 2006), which has chapters

really rather tantalizingly entitled ‘Disasters without Choice’, ‘Disasters
with Choice’ and ‘Disasters either way’.

Herrlich perhaps already tells you more than enough about the impact of AC:
but there’s also a famous book by H. Rubin and J.E. Rubin, Equivalents of the
Axiom of Choice (North-Holland 1963; 2nd edn. 1985) worth browsing through:
it gives over two hundred equivalents of AC!
Then next there is the nice short classic
3. Thomas Jech, The Axiom of Choice* (North-Holland 1973, Dover Publi-
cations 2008). This proves the Gödel and Cohen consistency and indepen-
dence results about AC (without bringing into play everything needed
to prove the parallel results about the Continuum Hypothesis). In par-
ticular, there is a nice presentation of the so-called Fraenkel-Mostowski
method of using ‘permutation models’. Then later parts of the book tell
us something about mathematics without choice, and about alternative
axioms that are inconsistent with choice.
And for a more recent short book, taking you into new territories (e.g. making
links with category theory), enthusiasts might enjoy

4. John L. Bell, The Axiom of Choice* (College Publications, 2009).

(b) From earlier reading you should certainly have picked up the idea that,
although ZFC is the canonical modern set theory, there are other theories on
the market. I mention just a selection here (I’m not suggesting you need to follow
up all these pointer – but it is worth stressing again that set theory is not quite
the monolithic edifice that some presentations might suggest).
For a brisk overview, putting many of the various set theories we’ll consider
below into some sort of order, and mentioning yet further alternatives, see

5. M. Randall Holmes, ‘Alternative axiomatic set theories’, The Stanford

Encyclopedia of Philosophy, tinyurl.com/alt-set.

At this stage, you might well find this a bit too brisk and allusive, but it is useful
to give you a preliminary sense of the range of possibilities here. And I should
mention that there is a longer version of this essay which you can return to later:
6. M. Randall Holmes, Thomas Forster and Thierry Libert. ‘Alternative
set theories’. In Dov Gabbay, Akihiro Kanamori, and John Woods, eds.
Handbook of the History of Logic, vol. 6, Sets and Extensions in the
Twentieth Century, pp. 559-632. (Elsevier/North-Holland 2012).
162
 DRAFT– 31 DEC 2021
Choice, and the choice of set theory

(c) It quickly becomes clear that some alternative set theories are more alter-
native than others! So let’s start with the one which is the closest sibling to
standard ZFC, namely NBG. You will have very probably come across mention
of this already (e.g. even in the early pages of Enderton’s set theory book).
We know that the universe of sets in ZFC is not itself a set. But we might
think that this universe is a sort of big collection. Should we explicitly recognize,
then, two sorts of collection, sets and (as they are called in the trade) proper
classes which are too big to be sets? Some standard presentations of ZFC, such
as Kunen’s, do indeed introduce symbolism for classes, but then make it clear
that class-talk is just a useful short-hand that can be translated away. NBG
(named for von Neumann, Bernays, Gödel: some say VBG) takes classes a bit
more seriously. But things are a little delicate: it is a nice question just what
NBG commits us to. An important technical feature is that its principle of class
comprehension is ‘predicative’; i.e. quantified variables in the defining formula
for a class can’t range over proper classes but range only over sets. Because of
this we get a conservative extension of ZFC (nothing in the language of sets can
be proved in NBG which can’t already be proved in ZFC). For more, see:

7. Abraham Fraenkel, Yehoshua Bar-Hillel and Azriel Levy, Foundations of

Set-Theory (North-Holland, 2nd edition 1973). Their Ch. II §7 remains
a classic general discussion of the role of classes in set theory.

And also worth quickly consulting is

8. Michael Potter, Set Theory and Its Philosophy (OUP 2004) Appendix C
is a brisker account of NBG and of other theories with classes as well as
sets.

Then, if you want detailed presentations of set-theory via NBG, you can see
either or both of

9. Elliott Mendelson, Introduction to Mathematical Logic (CRC, 4th edition

1997), Ch.4. is a classic and influential textbook presentation.
10. Raymond Smullyan and Melvin Fitting, Set Theory and the Continuum
Problem (OUP 1996, Dover Publications 2010), Part I is another devel-
opment of set theory in its NBG version.

(d) Recall, earlier in the Guide, we very warmly recommended Michael Potter’s
book which we just mentioned again. This presents a version of an axiomatiza-
tion of set theory due to Dana Scott (hence ‘Scott-Potter set theory’, SP). This
axiomatization is consciously guided by the conception of the set theoretic uni-
verse as built up in levels (the conception that, supposedly, also warrants the
axioms of ZF). What Potter’s book aims to reveal is that we can get a rich hier-
archy of sets, more than enough for mathematical purposes, without committing
ourselves to all of ZFC (whose extreme richness comes from the full Axiom of
Replacement). If you haven’t read Potter’s book before, now is the time to look
at it. Also, for a slightly simplified presentation of SP, see
163
 DRAFT– 31 DEC 2021
12 Going further

11. Tim Button, ‘Level Theory, Part I’, Bulletin of Symbolic Logic, preprint
available at tinyurl.com/level-th.
(e) We now turn to a somewhat more radical departure from standard ZF(C),
namely ZFA (i.e. ZF − AF + AFA)
Here again is the now-familiar hierarchical conception of the set universe: We
start with some non-sets (maybe zero of them in the case of pure set theory). We
collect them into sets (as many different ways as we can). Now we collect what
we’ve already formed into sets (as many as we can). Keep on going, as far as we
can. On this ‘bottom-up’ picture AF, the Axiom of Foundation, is compelling
(any downward chain linked by set-membership will bottom out, and won’t go
round in a circle).
But here’s another alternative conception of the set universe. Think of a set as
a gadget that points you at some some things, its members. And those members,
if sets, point to their members. And so on and so forth. On this ‘top-down’
picture, the Axiom of Foundation is not so compelling. As we follow the pointers,
can’t we for example come back to where we started? It is well known that in
much of the usual development of ZFC the Axiom of Foundation AF does little
work. So what about considering a theory of sets ZFA which drops AF and
instead has an Anti-Foundation Axiom, AFA, which allows self-membered sets?
To explore this idea, see
12. Start with Lawrence S. Moss, ‘Non-wellfounded set theory’, The Stanford
Encyclopedia of Philosophy, tinyurl.com/sep-zfa.
13. Keith Devlin, The Joy of Sets (Springer, 2nd edn. 1993), Ch. 7. The last
chapter of Devlin’s book, added in the second edition of his book, starts
with a very lucid introduction, and develops some of the theory.
14. Peter Aczel, Non-well-founded Sets (CSLI Lecture Notes 1988). This is
a very readable short classic book, available at tinyurl.com/aczel.
15. Luca Incurvati, ‘The graph conception of set’ Journal of Philosophical
Logic (2014) pp. 181-208, or his Conceptions of Set and the Foundations
of Mathematics (CUP, 2020), Ch. 7, very illuminatingly explores the
motivation for such set theories.
(f) Now for a much more radical departure from ZF.
Standard set theory lacks a universal set because, together with other stan-
dard assumptions, the idea that there is a set of all sets leads to contradiction.
But by tinkering with those other assumptions, there are coherent theories with
universal sets, of which Quine’s ‘New Foundations’ is the probably the best
known. For the headline news, see
16. T. F. Forster, ‘Quine’s New Foundations’, The Stanford Encyclopedia of
Philosophy, tinyurl.com/quine-nf.
For a full-blown but very readable presentation concentrating on NFU (‘New
Foundations’ with urelements), and explaining motivations as well as technical
details, see
164
 DRAFT– 31 DEC 2021
Choice, and the choice of set theory

17. M. Randall Holmes, Elementary Set Theory with a Universal Set (Cahiers
du Centre de Logique No. 10, Louvain, 1998). Now freely available at
tinyurl.com/holmesnf.

The following is rather tougher going, though with many interesting ideas:
18. T. F. Forster, Set Theory with a Universal Set Oxford Logic Guides 31
(Clarendon Press, 2nd edn. 1995).
(g) Famously, Zermelo constructed his theory of sets by gathering together
some principles of set-theoretic reasoning that seemed actually to be used by
working mathematicians (engaged in e.g. the rigorization of analysis or the de-
velopment of point set topology), hoping to get a theory strong enough for
mathematical use while weak enough to avoid paradox. The later Axiom of Re-
placement was added in much the same spirit. But does the result overshoot?
We’ve already noted that SP is a weaker theory which may suffice. For a more
radical approach, see this very engaging short piece:

19. Tom Leinster, ‘Rethinking set theory’. Gives an advertising pitch for the
merits of Lawvere’s Elementary Theory of the Category of Sets (ETCS).
tinyurl.com/leinst.

And for more on that, you could see e.g.

20. F. William Lawvere and Robert Rosebrugh, Sets for Mathematicians
(CUP 2003) gives a presentation which in principle doesn’t require that
you have already done any category theory. But I suspect that it won’t be
an easy ride if you know no category theory (and philosophers will find
it conceptually puzzling too – what are these ‘abstract sets’ that we are
supposedly theorizing about?). In my judgement, to really appreciate
what’s going on, you will have to start engaging with more category
theory. Which is a whole new ball game . . .
(h) I’ll finish by briefly mentioning two other directions you could go in!
First, ZF/ZFC has a classical logic: what if we change the logic to intuitionistic
logic? what if we have more general constructivist scruples? The place to start
exploring is

21. Laura Crosilla, ‘Set Theory: Constructive and Intuitionistic ZF’, The
Stanford Encyclopedia of Philosophy, tinyurl.com/crosilla.

Second, you’ll recall from elementary model theory that Abraham Robinson
developed a rigorous formal treatment that takes infinitesimals seriously. Later,
a simpler and arguably more natural approach, based on so-called Internal Set
Theory, was invented by Edward Nelson. He advertises it here:
22. Edward Nelson, ‘Internal Set Theory: a new approach to nonstandard
analysis’, Bulletin of The American Mathematical Society 83 (1977), pp.
1165–1198. tinyurl.com/nelson-ist.
165
 DRAFT– 31 DEC 2021
12 Going further

You can follow that up by looking at the approachable early chapters of Nader
Vakin’s Real Analysis through Modern Infinitesimals (CUP, 2011), a monograph
developing Nelson’s ideas.

12.6 More proof theory

(a) In §9.5, I mentioned three excellent books which are introductory in intent
but which take us a step up from the basic recommendations on proof theory
given earlier in Chapter 9, namely Takeuti’s Proof Theory, Girard’s Proof Theory
and Logical Complexity, and Troelstra and Schwichtenberg’s Basic Proof Theory.
If you didn’t take a look at them before, now is the time to do so!.
Also worth reading is the editor’s own first contribution to

1. Samuel R. Buss, ed., Handbook of Proof Theory (North-Holland, 1998).

Later chapters of this very substantial handbook do get pretty hard-core,
though yhou might want to look at some of them later. But the 78 pp.
opening chapter by Buss himself, a ‘Introduction to Proof Theory’, is
readable, and freely downloadable from tinyurl.com/buss-intro.1

(b) And now the paths through proof theory fork. One path investigates what
happens when we tinker with the structural rules shared by classical and intu-
itionistic logic.
Note for example the inference which takes us from the trivial P ` P by
weakening to P, Q ` P and on, via conditional proof, to P ` Q → P . If we
want a conditional that conforms better to intuitive constraints of relevance,
then we need to block that proof: is ‘weakening’ the culprit? The investigation
of what happens if we vary rules such as weakening belongs to ‘substructural
logic’, whose concerns are outlined in

2. Greg Restall, ‘Substructural logics’, The Stanford Encyclopedia of Phi-

losophy, tinyurl.com/sep-subs

And the place to continue exploring these themes at length is the same author’s

3. Greg Restall, An Introduction to Substructural Logics (Routledge, 2000),

which will also teach you a more about proof theory generally in a very
accessible way. Do try at least the first seven chapters.

(c) Another path forward picks up from Gentzen’s proof of the consistency of
arithmetic. Recall, that depends on transfinite induction along ordinals up to
ε0 ; and the fact that it requires just this much transfinite induction to prove the
consistency of first-order PA is an important characterization of the strength of
the theory.
The project of ‘ordinal analysis’ in proof theory aims to provide comparable
characterizations of other theories in terms of the amount of transfinite induction

1 Warning: there are, I am told, some confusing misprints in the cut-elimination proof.

166
 DRAFT– 31 DEC 2021
Higher-order logic, the lambda calculus, and type theory

that is needed to prove their consistency. Things do get quite hairy quite quickly,
however. But you can start from two very useful sets of notes for mini courses:

4. Michael Rathjen, ‘The realm of ordinal analysis’ and ‘Proof theory: from
arithmetic to set theory’, downloadable from tinyurl.com/rath-art and
tinyurl.com/rath-ast.

(d) Finally, here are a couple more books of notable interest:

5. Wolfram Pohlers, Proof Theory: The First Step into Impredicativity (Spr-
inger 2009). This book officially has introductory ambitions, focusing on
ordinal analysis. However, I would judge that it requires quite an amount
of mathematical sophistication from its reader. From the blurb: “As a
‘warm up’ Gentzen’s classical analysis of pure number theory is presented
in a more modern terminology, followed by an explanation and proof of
the famous result of Feferman and Schütte on the limits of predicativity.”
The first half of the book is probably manageable if (but only if) you
already have done some of the other reading. But then the going indeed
gets pretty tough.
6. H. Schwichtenberg and S. Wainer, Proofs and Computations (Associ-
ation of Symbolic Logic/CUP 2012) “studies fundamental interactions
between proof-theory and computability”. The first four chapters, at any
rate, will be of wide interest, giving another take on some basic mate-
rial and should be manageable given enough background. However, to
my surprise, I found the book to be not particularly well written and I
wonder if it sometimes makes heavier weather of its material than seems
really necessary. Still, worth getting to grips with.

12.7 Higher-order logic, the lambda calculus, and type theory

The logical grammar of first-order logic is very restricted. We assume a domain
of objects that we can quantify over; we can have names for some of these
objects; we can express properties and relations defined over those objects; and
can express (total) functions from one or more objects as inputs to objects
as outputs. In informal mathematics, by contrast, we quantify over properties,
relations and functions too (as in second-order logic). And we also consider
e.g. properties of relations (like being symmetric), relations between functions
(like being asymptotically equal), functions from one function to another (e.g.
differentiation), and more.
Now, as is familiar, we can trade in properties of relations, relations between
functions, functions of functions, etc. for sets. So we can compensate for the
expressive limitations of first-order logic by adopting enough set theory. Still, we
might reasonably look for a more expressive logical framework in which we can
talk directly about more types of things, and quantify over more types of things,
without playing the set-theory card. And exploring such a higher-order logic

167
 DRAFT– 31 DEC 2021
12 Going further

might even offer the prospect of an alternative, non-set-theoretic, foundation for

mathematics.
We looked at a small fragment of higher-order logic in Chapter 4 on second-
order logic. But now we want to explore theories with a richer type-structure.
Such a theory of types goes back at least until Bertrand Russell’s 1908 paper
‘Mathematical logic as based on the theory of types’. Its history since Russell
has been rather chequered. But particularly in the hands of theoretical computer
scientists, type theories have come back into considerable prominence. And in
the recent guise of homotopy type theory, one particular version is advertised as
a new foundation for mathematics. But where to start?
You could first take a quick look at

1. Jouko Väänänen, ‘Second-order and higher-order logic’, The Stanford

Encyclopedia of Philosophy, tinyurl.com/sep-vaan.
2. Thierry Coquand, ‘Type theory’, The Stanford Encyclopedia of Philoso-
phy, tinyurl.com/sep-type.

But the first of these mostly revisits second-order logic at a probably quite
unnecessarily sophisticated level for now, so don’t get bogged down. The second
gives us pointers forward, but is perhaps also rather too rushed.
Still, as you’ll see from Coquand, basic topics to pursue include Simple Type
Theory and the lambda calculus. For a clear and gentle introduction to the latter,
see the first seven chapters of the following welcome short book which doesn’t
assume much mathematical background:

3. Chris Hankin, An Introduction to Lambda Calculus for Computer Scien-

tists* (College Publications 2004).

Next, as a spur to keep going, you might find this advocacy interesting:

4. William M. Farmer, ‘The seven virtues of simple type theory’, Journal

of Applied Logic 6 (2008) 267–286. Available at tinyurl.com/farm-STT.

And then for a bit more on Simple Type Theory/Church’s Type Theory, though
once more this is less than ideal, you could look at

5. Christoph Benzmüller and Peter Andrews, ‘Church’s type theory’, The

Stanford Encyclopedia of Philosophy, tinyurl.com/sep-CTT.

But then where to go next will depend on your interests and on how much more
you want to know. The book we want, Type Theories for Logicians, A Gentle
Introduction, has yet to be written!
And a complicating factor is that a lot of current work on type theory is bound
up with constructivist ideas developing the BHK conception that ties the content
of a proposition to its proofs (for example, an implication A → C corresponds
to a type of function taking a proof A to a proof of C). This correspondence
between propositions and types of functions gets developed into the so-called
Curry-Howard correspondence or isomorphism. See
168
 DRAFT– 31 DEC 2021
Higher-order logic, the lambda calculus, and type theory

6. Peter Dybjer and Erik Palmgren, ‘Intuitionistic type theory’, The Stan-
ford Encyclopedia of Philosophy, tinyurl.com/sep-ITT.
But again, this isn’t easy going.
Without a Gentle Introduction to hand, you will have to make do with ex-
ploring the following initial suggestions they take you! In order of publication
date:
7. Henk P. Barendregt, The Lambda Calculus: Its Syntax and Semantics*
(Originally 1980, reprinted by College Publications 2012). This is the
weighty standard text: but the opening chapters are fairly accessible.
8. Peter Andrews, An Introduction to Mathematical Logic and Type The-
ory: To Truth Through Proof (Academic Press, 1986). Chapter 5, under
50 pages, is a classic introduction to a version of Church’s type the-
ory developed by Andrews. It is often recommended, and worth battling
through; but it is a rather terse bit of old-school exposition.
9. J. Roger Hindley, Basic Simple Type Theory (CUP, 1997). This short
book is another classic, but again it is pretty terse. Worth making a
start, but perhaps, in the end, mostly for those whose main interest is in
computer science applications of type theory in the design of higher-level
programming languages like ML.
10. Benjamin C. Pierce, Types and Programming Languages (MIT Press,
2002). A frequently-recommended text for computer scientists, and read-
able by others if you skip over some parts about implementation in ML.
The first dozen or so shortish chapters are indeed relatively discursive
and accessible.
11. Morten Heine Sørensen and Pawel Urzyczyn, Lectures on the Curry-
Howard Isomorphism (Elsevier, 2006). This engaging book ranges much
more widely than the title might suggest!
12. J. Roger Hindley and Jonathan P. Seldin, Lambda-Calculus and Combi-
nators: An Introduction (CUP 2008). Attractively and clearly written,
aiming to avoid excess technicalities. More of the feel of a modern maths
book. Recommended.
13. Rob Nederpelt and Hedman Geuvers, Type Theory and Formal Proof:
An Introduction (CUP 2014). Focuses, the authors say, “on the use of
types and lambda terms for the complete formalisation of mathematics”,
so promises to be of particular interest to mathematical logicians. Also
attractively and clearly written (as these things go!).
Then, pointing in a different direction, you might also want to follow up
14. Peter Dybjer and Erik Palmgren, ‘Intuitionistic type theory’, The Stan-
ford Encyclopedia of Philosophy, tinyurl.com/sep-ITT.
And finally, I suppose I should finish by mentioning again one particular new
incarnation of type theory:
169
 DRAFT– 31 DEC 2021
12 Going further

15. The Univalent Foundations Program, Homotopy Type Theory: Univalent

Foundations of Mathematics (2013), tinyurl.com/HOTT-book.

I leave it to you to make what you will of that program!

170
 DRAFT– 31 DEC 2021

Index

Ackermann, W., 30 Cook, R., 99

Aczel, P., 164 Cooper, S. B., 156
Andrews, P., 168, 169 Coquand, T., 168
Arora, S., 157 Cori, R., 35, 149
Artemov, S., 131 Creswell, M., 132
Atten, M. van, 98 Crosilla, L., 165
Avigad, J., 69 Cunningham, D., 87
Cutland, N., 156
Badesa, C., 32
Baldwin, J., 153 Dalen, D. van, 27–28, 54, 83, 96, 98
Ballarin, R., 133 Dawson, J., 72
Bar-Hillel, Y., 82, 163 Delzell, C., 152
Barak, B., 157 DeVidi, D., 33
Barendregt, H., 169 Devlin, K., 84, 160, 164
Barnes, D., 152 Doets, H., 83
Bell, J., 33, 150, 162 Doets, K., 152
Bencivenga, E., 145 Drake, F., 159
Bentham, J. van, 130 Dummett, M., 98, 149
Benzmüller, C., 168 Dybjer, P., 169
Blackburn, P., 133
Boolos, G., 45, 68, 114, 131, 146 Ebbinghaus, H-D., 34
Bostock, D., 26–27, 145 Ebbinghaus, H., 153
Bridge, J., 52 Enderton, H., 28, 44, 54, 69–70, 81,
Budiansky, S., 72 157
Burgess, J., 68, 141 Epstein, R., 68, 72
Buss, S., 166 Ewald, W, 32
Button, T., 11, 44, 56, 81, 86, 153
Farmer, W., 168
Carnielli, W., 68 Feferman, A., 57
Chagrov, A., 133, 149 Feferman, S., 57
Chang, C., 150 Ferreirós, J., 32, 86
Chiswell, I., 24–25 Finsler, P., 13
Chow, T., 114, 159 Fitch, F., 22
Ciesielski, K., 85 Fitting, M., 29–30, 71, 98, 130, 154,
Cocchiarella, N., 132 160, 163

171
 DRAFT– 31 DEC 2021
Index

Florio, S., 146, 147 Incurvati, L., 13, 85–86, 164

Flum, J., 34, 153
Font, J., 150 Jansana, R., 150
Forster, T., 162, 164, 165 Jech, T., 84, 158, 162
Fraenkel, A., 82, 163 Jeffrey, R., 33, 68
Franzén, T., 69, 155 Just, W., 82, 159
Frege, G., 30
Kanamori, A., 160
Freund, M., 132
Kaye, R., 35–36, 153
Galvan, S., 97, 112–115 Keisler, H., 150
Gandy, R., 72 Kleene, S., 34
Garson, J., 129, 132 Kossak, K., 153
Girard, J-Y, 166 Kristiansen, L., 25, 70
Girard, J-Y., 116 Kunen, K., 13, 159
Girle, R., 129
Lambert, K., 146
Givant, S., 149
Langford, C., 119
Goldrei, D., 25–26, 52–53, 81
Lascar, D., 35, 149
Goldreich, O., 157
Lawvere, F. W., 165
Gowers, T., 12
Leary, C., 25, 70
Hájek, P., 153 Leinster, T., 165
Hajnal, A., 86 Levy, A., 82, 86, 158, 163
Halbeisen, L., 160 Lewis, C., 119
Halmos, P., 83, 149 Libert, T., 162
Halvorson, H., 44 Libkin, L., 152
Hamburger, P., 86 Lindström, P., 155
Hamkins, J., 6 Lindström, S., 134
Hankin, C., 168 Linnebo, Ø., 146, 147
Hazen, A., 24, 113 Loeser, F., 34
Hedman, S., 30, 35, 55, 150, 157
Mack, J., 152
Herrlich, H., 162
Maddy, P., 85
Hilbert, D., 30
Makinson, D., 12
Hills, M., 34
Mancosu, P., 32, 97, 112–115
Hils, M., 34
Manzano, M., 53
Hindley, J.R., 169
Marcja, A., 152
Hinman, P., 35, 150, 156
Mares, E., 141, 142
Hodel, R., 54
Marker, D., 46, 151
Hodges, W., 24–25, 27, 46, 52, 56,
McCarty, D., 99
150
McKay, T., 147
Hofstadter, D., 71
Mendelsohn, R., 130
Holmes, M. R., 162, 165
Mendelson, E., 2, 21, 33, 115, 163
Hrbacek, K., 84
Meulen, A. ter, 149
Hughes, G., 132
Mints, G., 96
Humberstone, L., 133
Montanaro, A., 157
Hunter, G., 34
Moschovakis, J., 96
172
 DRAFT– 31 DEC 2021
Index

Moschovakis, Y., 82 Schmerl, J., 153

Moss, L., 164 Schwichtenberg, H., 116, 166, 167
Munkres, J., 11 Segerberg, K., 134
Seldin, J., 169
Negri, S., 114 Shapiro, S., 44, 45
Nelson, E., 165 Shen, A., 70, 85
Nolt, J., 145 Shoenfield, J., 33
Sider, T., 32, 131–132
Odifreddi, P., 157
Sieg, W., 112
Oliver, A., 143, 146, 147
Simmons, H., 133
Oosten, J. Van, 153
Simpson, S., 154
Péter, R., 156 Sipser, M., 157
Palmgren, E., 169 Skiba, L., 45
Partee, B., 149 Slomson, A., 150
Pelletier, F., 24, 113 Smiley, T., 143, 146, 147
Pierce, B., 169 Smith, N., 6
Plato, J. van, 112–114 Smith, P., 6, 24, 27, 67
Pohlers, W., 167 Smoryński, C., 131, 154, 155
Poizat, B., 152 Smullyan, R., 29, 36, 71, 154, 155,
Popkorn, S., 133 160, 163
Posy, C., 99 Solomon, G., 33
Potter, M., 83, 163 Suppes, P., 86
Prawitz, D., 113 Swart, H. de, 34, 83, 98
Prestel, A., 152 Sørensen, M., 149, 169
Priest, G., 97, 130, 141, 142, 146 Takeuti, G., 115, 166
Pudlák, P., 153
Tennant, N., 142, 146
Quine, W., 12, 43 Tent, K., 151
Thomas, W., 34
Rathjen, M., 112, 167 Toffalori, C., 152
Rautenberg, W., 35 Tourlakis, G., 87
Restall, G., 166 Troelstra, A., 96, 98, 116, 166
Ricke, M. de, 133
Rogers, H., 156 Urzyczyn, P., 149, 169
Rogers, R., 66
Väänänen, J., 168
Rosebrugh, R., 165
Vakin, N., 166
Rothmaler, P., 151
Vaught, R., 57
Roy, T., 132
Velleman, D., 6
Rubin, H., 162
Venema, Y., 133
Rubin, J., 162
Vereshchagin, N., 70, 85
Russell, B., 30
Wainer, S., 167
Scanlon, T., 150
Wall, R., 149
Schimmerling, E., 85
Walsh, S., 44, 56, 86, 153
Schindler, R., 161
Weaver, N., 161
173
 DRAFT– 31 DEC 2021
Index

Weese, M., 82, 159

Whitehead, A., 30
Williamson, T., 132
Wolf, R., 66, 80
Wong, T., 153, 156

Zach, R., 28–29, 32, 44, 69, 97, 112–

115
Zakharyaschev, M., 133, 149
Ziegler, M., 151

174