Intuitionistic Set Theory

J o h n L. B e l l
 Studies in Logic
Volume 50

Intuitionistic Set Theory

Volume 39
Non-contradiction
Lawrence H. Powers, with a Foreword by Hans V. Hansen
Volume 40
The Lambda Calculus. Its Syntax and Semantics
Henk P. Barendregt
Volume 41
Symbolic Logic from Leibniz to Husserl
Abel Lassalle Casanave, ed.
Volume 42
Meta-argumentation. An Approach to Logic and Argumentation Theory
Maurice A. Finocchiaro
Volume 43
Logic, Truth and Inquiry
Mark Weinstein
Volume 44
Meta-logical Investigations in Argumentation Networks
Dov M. Gabbay
Volume 45
Errors of Reasoning. Naturalizing the Logic of Inference
John Woods
Volume 46
Questions, Inferences, and Scenarios
Andrzej Wisniewski
Volume 47
Logic Across the University: Foundations and Applications. Proceedings of
the Tsinghua Logic Conference, Beijing, 2013
Johan van Benthem and Fenrong Liu, eds.
Volume 48
Trends in Belief Revision and Argumentation Dynamics
Eduardo L. Ferme, Dov M. Gabbay, and Guillermo R. Simari
Volume 49
Introduction to Propositional Satisfiability
Victor Marek
Volume 50
Intuitionistic Set Theory
John L. Bell

Studies in Logic Series Editor

Dov Gabbay dov.gabbay@kcl.ac.uk
Intuitionistic Set Theory

John L. Bell
© Individual author and College Publications 2014.
All rights reserved.

ISBN 978-1-84890-140-7

College Publications
Scientific Director: Dov Gabbay
Managing Director: Jane Spurr

http://www.collegepublications.co.uk

Original cover design by Orchid Creative www.orchidcreative.co.uk

Printed by Lightning Source, Milton Keynes, UK

All rights reserved. No part of this publication may be reproduced, stored in a

retrieval system or transmitted in any form, or by any means, electronic, mechanical,
photocopying, recording or otherwise without prior permission, in writing, from the
publisher.
To Sandra, who has given me new life
 Table of Contents

Introduction 1

The natural numbers and countability 1

Power sets 4
The Continuum 6

Chapter I. Intuitionistic Zermelo Set Theory 11

Axioms and basic definitions 11

Logical principles in IZ 15
The axiom of choice 18

Chapter II. Natural Numbers and Finite Sets 25

The natural numbers 25

Models of Peano's axioms 27
Definitions by recursion 28
Finite sets 33
Frege’s construction of the natural numbers 37

Chapter III. The Real Numbers 45

Dedekind real numbers and weak real numbers 45

Cauchy real numbers 49

Chapter IV. Intuitionistic Zermelo-Fraenkel Set Theory

and Frame-Valued Models 51

Intuitionistic Zermelo-Fraenkel set theory IZF 51

Frame-valued models of IZF developed in IZF 54
The consistency of ZF and ZFC relative to IZF 65
Frame-valued models of IZF developed in ZFC 66
A frame-valued model in which NNis subcountable 71
The axiom of choice in frame-valued extensions 76
Real numbers and real functions in spatial extensions 77
Properties of the set of real numbers over R 84
Properties of the set of real numbers over Baire space 85
 The independence o f the fundamental theorem o f algebra from IZF 88

Appendix. Heyting Algebras and Frames 91

Lattices 91
Heyting and Boolean algebras 93
Coverages on partially ordered sets and their associated frames 97
Connections with logic 98

Concluding Observations 103

Historical Notes 105

Bibliography 107

Index 111
 Preface

While intuitionistic (or constructive) set theory 1ST has received a certain attention
from mathematical logicians, so far as I am aware no book providing a systematic
introduction to the subject has yet been published. This may be the case in part
because, as a form of higher-order intuitionistic logic - the internal logic of a
topos - 1ST has been chiefly developed in a tops-theoretic context. In particular,
proofs of relative consistency with 1ST for mathematical assertions have been
(implicitly) formulated in topos- or sheaf-theoretic terms, rather than in the
framework of Heyting-algebra-valued models, the natural extension to 1ST of the
well-known Boolean-valued models for classical set theory.

In this book I offer a brief but systematic introduction to 1ST which develops the
subject up to and including the use of Heyting-algebra-valued models in relative
consistency proofs. I believe that 1ST, presented as it is in the familiar language of
set theory, will appeal particularly to those logicians, mathematicians and
philosophers who are unacquainted with the methods of topos theory.

The title I originally had in mind for this book was Constructive Set Theory. Then it
occurred to me that the term “constructive” has come to connote not merely the use
of intuitionistic logic, , but also the avoidance of impredicative definitions. This is
the case, for example, with Aczel’s Constructive Set Theory in which the power set
axiom (which permits impredicative definirions of sets) is not postulated. Since the
power set axiom and impredicative definitions are very much a part of 1ST, to avoid
confusion I have decided (with some reluctance) to give the book its present title.

JLB January 2014

 Introduction

Challenging the Logical Presuppositions of Classical

Set Theory

In classical set theory free use is m ade of the logical principle known as the
Law of Excluded Middle (LEM): for any proposition p , either p holds or its
negation -<p holds. But there are a num ber of intriguing mathematical
possibilities which are ruled out within classical set theory solely because of
the presence there of LEM. This suggests the idea of dropping LEM in set-
theoretical arguments, or, more precisely, basing set theory on intuitionistic
logic. Accordingly, let us define Intuitionistic Set Theory (1ST) to be any of
the usual axiomatic set theories (e.g. Zermelo-Fraenkel set theory ZF) based
on intuitionistic - rather than classical - logic.

Here are some examples of such mathematical possibilities.

THE NATURAL NUMBERS AND COUNTABILITY.

Call a set countable if it is empty or the range of a function defined on the set
N of natural numbers, subcountable if it is the range of a function defined on
a subset of N, and numerable if it is the dom ain of an injection into N. In
classical set theory all of these notions are equivalent, as the following
argum ent shows. Obviously, every countable set is subcountable. If £ is
subcountable, there is a subset U of N and a surjection / U -*E. Then the
function m: £ —» N defined by m(x) = least n e N for which f(n) = x is injective,
and it follows that £ is numerable. Finally, suppose that E is numerable, and
let m: E >-►N be injective. Then either £ = 0 or E ^ 0 ; in the latter case, fix
e e £ and define / N -» £ by setting, for n e rang e(m), f(n) = unique x g E for
which m(x) = n; and, for n € rang e(m),f(n) = e. T h e n /is surjective, and so £ is
countable.

It should be clear that the validity of this argum ent rests on two
assumptions: LEM and the assertion that N is well-ordered (under its natural
ordering). Accordingly, if we base our reasoning on intuitionistic logic in
which LEM is not affirmed, then we can see that, while subcountability
obviously continues to be inferable from countability, the argum ent deriving
countability from num erability breaks dow n because of its dependence on
LEM. One m ight suppose that the validity of the argum ent deriving
numerability from subcountability survives the passage to intuitionistic
logic, but actually it does not, for it uses the assum ption that N is well-
ordered, and this can be shown to imply LEM. For, given a proposition p,
define U = {x e N: x = 0 a p} u {!}. Notice that 0 e U <-> p. Then U is
nonem pty and so, if N is well-ordered, has a least element n. Since n e U, we
have n = 0 v n ~ 1. If n = 0, then 0 e IT, whence p; if n = 1, then 0 £ U, whence
—>p. Hence p v ~^p, and LEM follows.

Now, as we have said, it is obvious that any countable set is subcountable,

and it is easily shown that any numerable set is subcountable . However, in
stinking contrast with classical set theory, it is consistent with 1ST to assume
the existence of sets which are (a) subcountable, but uncountable; (b)
numerable, but uncountable; and (c) countable, but not numerable.

Perhaps Cantor's most celebrated theorem is the uncountability of the set R

of real numbers. Cantor first published a proof of this theorem in 1874, but
much better known is his second proof, published in 1890, in which he
introduces his famous method of "diagonalization" . In essence, Cantor's
argum ent establishes that the set of all maps N -> N is uncountable in the
above sense. For given a map cp: N -> Nh' the m ap fi N —» N defined by

(*) An) =9(«)(«) +1

clearly cannot belong to range(cp), so that (p cannot be surjective. This
argum ent does not use LEM, and is in fact perfectly valid within 1ST.

Now Cantor would also have accepted the extension of this argum ent to
show that ! cannot be subcountable in the above sense For given U c; N and
a surjection (p: U N*', if we define fi N -> N by (*)

(**) f(n) = y(n)(ri) + 1 if n e l l f(n) = 0 if n <£ l l

2
then clearly f £ range(cp) and so again <p fails to be surjective. But this
argum ent uses LEM and so is not valid within 1ST. In fact, it is consistent
with 1ST for N" to he subcountable, so making N!i both subcountable and
uncountable!

When N is subcountable, remarkable things happen. To begin with, LEM

becomes refutable. The simplest way to see this is to note that given U c N
and a surjection cp: U NN, the assertion

(*) Vx e N[x e U v -{x e U)]

is refutable. For if (*) held, then we could extend cp to a surjection N -> N' by
assigning the constant value 0 to all n g U. This would make N 1countable
which we have already shown to be an impossibility. Secondly, N will have
an uncountable subset. For the suset U of N cintroduced above annot be
countable, since the composite of (p with any surjection N -» 1/ would be a
surjection N -» Nn, making NM, impossibly, countable. Finally, the
subcountability of NNimplies that there is a function defined on a (proper) subset
of N zuhich cannot be extended to the whole of N. To see this, take U and cp as
above and define fi U -» N by setting, for n e N

/ ( n ) = <p(n)(n) + l .

Suppose now th a t/c o u ld be extended to a function g : N -> N. Then since (p

is surjective, there is noe U for which g = cp(rz0) , leading to the contradiction

<PK ) K ) = 9 K ) = f ( n 0) - cp(n0)(n0) + 1.

So all of these possibilites are consistent withlST.

Now Cantor would also, presumably, have accepted that cannot be

numerable in the sense just introduced. For if Nr:i were numerable, then it
would (using classical reasoning as above) also have to be countable,
contradicting its uncountability. But the argum ent from numerability to

3
countability does not hold up within 1ST and in fact it is consistent with 1ST
for Nn to be num erable1!

Let us see what happens (within 1ST) w hen NN is replaced by the set
Par(N,N) of all partial functions from N to N. First, we observe that Par(N,N)
cannot be subcountable, for suppose U c N and (p: U -» Par(N,N) is a
surjection. Let r e Par(N,N) be the identity map on dom (r) = {x e II: x £
dom((p(x))} . Then since (p is surjective, r = cp(/t) for some n e U quickly
leading to the contradiction n e dom(r) <-> n £ dom(r). 2

Nor can Par(N,N) be numerable. For suppose cp: Par(N,N) >-* N is injective,
and define u g Par(N,N) to be the identity m ap on

dom (u) = [x e N: 3fe Par(N,N)fo>(/) = x ax e dom (/)}.

Then, writing n = (p(w), we have

n g dom(u) <-» 3/[cp(/) = n a n e dom(/)].

3/[(p(/) = (p(//) A n e dom (/)]
w a n £ dom(/)]
n £ dom(//),

a contradiction. So Par(N,N) is innumerable.

What if we replace Par(N,N) by the set Par*(N,N) of all partial maps on N

with countable domains? Suppose that Par*(N,N) is actually countable, and let
(p: N -» Par*(N,N) be a surjection. If r g Par(N,N) is the identity map on
dom (r) = {x g N : x <£ dom((p(x))}, then the argum ent above only leads to
contradiction when r g Par*(N,N), from which we conclude that

1 The proof of this fact is, unfortunately, beyond the scope of the present book.
2 Clearly this argument continues to hold when N is replaced by an arbitrary set £.

4
r <£ Par*(N,N), in other words, {x e N : x £ dom(cp(x))} is uncountable. In fact,
the countability of Par*(N,N) is consistent with 1ST3.

POWER SETS

Let us turn next to another celebrated theorem of Cantor, namely that, for
any set £, the cardinality of £ is strictly smaller than that of its power set P£.
One way of construing this is the assertion that there can be no surjection
£ -*• PE. Within 1ST this can be proved, as in classical set theory, by
employing the argum ent of Russell's paradox: given (p : £ —> P£ one
defines the associated "Russell set"

R —{x g £: x £ ip(x)}

and then shows in the customary way that R g range(cp). For U c; £, a

similar argument, replacing R above by R n U, shows that there can be no
surjection Ü PE. Thus, in particular, within IST PN is unsubcountable
(and so also uncountable).

Equally, the (classically equivalent, but not automatically intuitionistically

equivalent) form of Cantor's theorem that, for any set £ there is no injection
P E >-> E can also be given a proof within 1ST using the idea of Russell's
paradox. For suppose given an injection m: PE ^ E. Define

B = [x e £: 3X <= PE. x = m(X) a j e X}.

Writing m(B) =b, we have

b e B <-> 3X. b = m(X) A b z X

*-> 3X. m(B) - m(X) a b £ X
<->3X. B = X A b e X
<-» b £ B,

and we have our contradiction. In particular, within IST, PN cannot be

numerable.

3 Again, the proof of this fact is beyond the scope of the present book.

5
What if we replace PN by the set P*N of all countable subsets of N? Suppose

that P*N s actually countable, and let cp: N P*N be a surjection. Defining
R = {x e N: x g cp(j)} as before, the argum ent above only leads to a
contradiction when R e P*N, from which we conclude that R £ P*N, that is
R is uncountable. In fact, it follows directly from the consistency of the
subcountability of Nn with 1ST that the subcountability of P N is also
consistent with 1ST4, since the m ap Nn -» P N : f h* ra n g e (/) is surjective.

In classical set theory, P£ is naturally bijective with 2£, the set of all maps5
£ -» 2 = {0, 1}. In 1ST, this is no longer the case. Here, in general, P£ = Q£,
where Q is the object of truth values or propositions, that is, the set PI of all
subsets of {0}. Q is only identical with 2 w hen LEM is aassum ed.6 In fact, in
1ST, 2£ is isomorphic, not to P£, but to its Boolean sublattice C£ consisting of
all detachable subsets of £ (a subset 17 of E is said to be detachable if
V* e £. x e U v x £ II). W hat happens when we replace PE by C£ in the
above arguments? Classically, of course, this makes no difference, but do the
"Russell's paradox" argum ents survive the transition to 1ST? Well, if one
takes the first argum ent, showing that there can be no surjection (p; £ P£,
one finds that, when P£ is replaced by C£, the set R & range((p) is actually
detachable and the argum ent goes through, proving in 1ST that there can be
no surjection £ C£. But the second argument, with PE replaced by C£
(and then £ replaced by a subset U of £) goes through in 1ST only if LI is
detachable. And for the third argum ent to go through in 1ST once P£ is
replaced by C£, it is necessary to show that the set B defined there is
detachable. In fact, as we shall see, these argum ents can break down
completely in 1ST even when £ is the set M of real numbers: it is in fact
consistent with 1ST that R has just the two detachable subsets 0 , R, so that
CR (and so also 2ii?) is isomorphic to 2. A fortiori CR is injectible into R,
showing that the third argum ent fails in 1ST.

4 In fact the countability of PN is consistent with 1ST, but the proof of this fact is
beyond the scope of the present book.
5 Here we write 0 for 0 and 1 for {0}.
6 On the other hand the natural bijection between Q£ and P£ is given, as it is
classically, by/»-*-/-i(0).

6
 THE CONTINUUM

It is characteristic of a continuum that it is "gapless" or "all of one piece", in

the sense of not being actually separated into two (or more) disjoint nonempty
parts. On the other hand it has been taken for granted from antiquity that
continua are limitlessly divisible, or separable into parts in the sense that any
part of a continuum can be "divided", or "separated" into two or more
proper parts. Now there is a traditional conceptual difficulty in seeing just
how the parts of a continuum obtained by separation —assumed disjoint —
"fit together" exactly so as to reconstitute the original continuum. This
difficulty is simply illustrated by considering the case in which a straight
line X is divided into two segments L, R by cutting it at a point p. What
happens to p when the cut is made? On the face of it, there are four
possibilities (not all mutually exclusive): (i) p is neither in L nor in R; (ii) p
may be identified as the right-hand endpoint pi of L: (iii) p may be identified
as the left-hand endpoint pR of R; (iv) p may be identified as both the right-
hand endpoint of L and the left-hand endpoint of R. Considerations of
symmetry suggest that there is nothing to choose between (ii) and (iii), so
that if either of the two holds, then so does the other.

Accordingly we are reduced to possibilities (/) and (iv). In case (/), L and R
are disjoint, but since neither contains p, they together fail to cover X; while
in case (iv), L and R together cover X, but since each contains p, they are not
disjoint. This strongly suggests that a (linear) continuum cannot be
separated, or decomposed, into two disjoint parts which together cover it.
Herein lies the germ of the idea of cohesiveness.

Of course, this analysis is quite at variance with the account of the (linear)
continuum provided by classical set theory. There the continuum is takes the
form of the discrete linearly ordered set E of real numbers. "Cutting" E (or
any interval thereof) at a point p amounts to partitioning it into the pairs of
subsets ({x: x < p], {x : p < x}) or ({x: x < p\, {x: p < x}): the first and second of
these correspond, respectively, to cases (ii) and (iii) above. Now in the
discrete case, one cannot appeal to symmetry as before: consider, for
instance, the partitions of the set of natural numbers into the pairs of subsets
({n: n < 1}, {n: 1 < n}) and ({n: n < 1}, {n : 1 < n}). The first of these is ( {0, 1},
{2, 3, ...}) and the second ({0}, {1, 2,...}). Here it is manifest that the symmetry
naturally arising in the continuous case does not apply: in the first partition
1 is evidently a member of its first component and in the second partition, of
its second. In sum, when a discrete linearly ordered set X is "cut", no

7
ambiguity arises as to which segment of the resulting partition the cut point
is to be assigned, so that the segments of the partition can be considered
disjoint while their union still constitutes the whole of X.

Acknowledging the fact that the set-theoretic continuum, as a discrete entity,

can be separated into disjoint parts, classical set theory proceeds to capture
the characteristic "gaplessness" of a continuum by restricting the nature of
the parts into which it can be so separated. In set-theoretic topology this is
done by confining " parts" to open (or closed) subsets, leading to the standard
topological concept of connectedness. Thus a space S is defined to be
connected if it cannot be partitioned into two disjoint nonem pty open (or
closed) subsets —or equivalently, given any partition of S into two open (or
closed) subsets, one of the members of the partition m ust be empty. It is a
standard topological theorem that the space R of real numbers and all of its
intervals are connected in this sense.

But now let us return to our original analysis. This led to the idea that a
continuum cannot be decomposed into disjoint parts. Let us take the bull by
the horns and attem pt to turn this idea into a definition. We shall call a space
S cohesive or indecomposable, or a (genuine) continuum if , for any parts, or
subsets U and V of S, whenever U u V = S and U n V = 0 , then one of U, V
m ust = 0 , or, equivalently, one of U, V m ust = S. Clearly S is cohesive
precisely when its only detachable subsets are 0 and S itself.

Cohesiveness can be furnished with various "logical" formulations. Namely,

S is cohesive if and only if, for any property P defined on S, the following
implication holds:

(*) Vx g S[P(x) v -,P(x)} -> [Vjc g S P( x ) v Vx g S-,P( jc)].

We observe that in classical set theory, the only cohesive spaces are the
trivial empty space and one-point spaces. But it turns out that the existence
of nontrivial cohesive spaces is consistent with 1ST. In fact it is consistent
with 1ST that R itself is cohesibe. How does this come about? To get a clue, let
us reformulate our definitions in terms of maps, rather than parts. If we
denote by 2 the two-element discrete space, then connectedness of a space S
is equivalent to the condition that any continuous map S -» 2 is constant, and
cohesiveness of S to the condition that any map S -» 2 whatsoever is constant.
Supposing S to be connected and to possess more than one point, then from

8
LEM it follows that there exist nonconstant —and hence discontinuous —
maps S -» 2. But the situation would be decidedly otherwise if all maps
defined on S were continuous, for then, clearly, the connectedness of S
would immediately yield its cohesiveness. In fact it is consistent with 1ST
that all maps E -> E - and hence all maps R —» 2 - are continuous. It
follows that it is consistent with 1ST that E is cohesive.

The consistency with 1ST of all these possibilities can be established bv

constructing models of 1ST in which they can be shown to hold. In this
book we shall establish both the consistency with 1ST of the subcountability
of N and the cohesiveness of E through the use of Heyting-algebra valued
models, of which the more familiar Boolean-valued models of classical set
theory are a special case.

Some years ago Paul Cohen published an article in the Scientific American
entitled Non- Cantorian Set Theory. There he describes the set theories in
which Cantor's continuum hypothesis is violated. In terming these set
theories non-Cantorian he was making an analogy with non-Euclidean
geometries in which the parallel postulate is violated. If, in Cohen's
analogy, non-Cantorian set theories correspond to non-Euclidean
geometries, then classical (Zermelo-Fraenkel) set theory corresponds to
neutral or absolute goemetry in which no form of the parallel postulate is laid
down. So we are led to reformulate Cohen's analogy by replacing the
continuum hypothesis with the Law of Excluded Middle; it is then
intuitionistic set theory that corresponds to neutral geometry. Intuitionistic set
theory can thus be seen as a "neutral" set theory, compatible with a num ber
of principles - such as the subcountability of N* and the cohesiveness of the
real line - which are incompatible with classical set theory.

Janos Bolyai, one of the inventors/discoverers of non-Euclidean geometry,

was moved to describe it as a "strange new universe" - and indeed it was ,
by the canons of Euclidean geometry. Similarly, certain of the various
extensions of intuitionistic set theory described above may strike one as
"strange new universes" in comparison with the familiar universe of
classical set theory. But, just as geometers became familiar with non-
Euclidean geometry, providing it with models (such as the pseudosphere)
which made it seem "natural", so seemingly curious properties compatible
with intuitionistic set theory - such as the subcountability of N ' or the

9
cohesiveness of the real line - become clear when their meanings in the
models realizing them are understood.

In this book we shall formulate and develop versions of intuitionistic

Zermelo- and Zermelo-Fraenkel set theories - IZ and IZF, respectively, and
construct the Heyting-algebra valued models which will be used to
establish the relative consistency with IZF of some of the assertions we
have discussed above.

10
 Chapter I

Intuitionistic Zermelo Set Theory

AXIOMS AND BASIC DEFINITIONS

Intuitionistic set theory is formulated as a system of axioms in the same first-

order language as its classical counterpart, only based on intuitionistic logic.
The language of set theory is a first-order language C with equality, which
includes a binary symbol e. We write x * y for -i (x = y) and x <£ y for
-i (x g y). Individual variables x, y , 2, ...of C will be understood as ranging
over sets. C will also allow the formation of terms of the form {x: cp(x)}, for
any formula (p containing the free variable x.. Such terms are called classes;
we shall use upper case letters A, B , ... for classes. For each class
A= {x: cp(x)} the formula

\fx[x G A <-» c p (x )]

is called the defining axiom for the class A. Two classes A, B are defined to be
equal and we write A = B if

\/x(x e A <->x e B).

A is a subclass of B, and we write A c= B, if

\/x(x G A - » X G B).

We also write Set(A) for the formula

3u\fx(x G i 4 f > X G l i ) .

Set(A) asserts that the class A is a set. For any set u, it follows from the
defining axiom for the class {x: x g w ) that Set({x: x g u }). We shall identify

11
{x : x e u } with w, so that sets may be considered as (special sorts of) classes
and we may introduce assertions such as u c A , m = A, etc.

We define the following classes:

• {ul9...,un} = { x : x = u l v ... v x —u n\ .
• \ j A = {x:3y(y g A A X & y ) }

• f)A={x:Vy(yeA->xey}}
• A u B ~ { x : x g A v x g B}
• A n B = {x: x e A a x g
• A - B = {x : x g A a x £B}
• u" = u u {u}
• PA = {x: x e A} (the power class ofA)
• {xgA: (p(x)j = { x: x e A a cp(x)}
• {t(x ) : cp(x)} = {y : 3x. y = £(x) a (p(x)} where t(x) is a term such that
V x Set(£(x))
• V = {x: x = xj
• 0 = {x: x * x}

The system IZ of intuitionistic Zermelo set theory is based on the following

axioms:

Extensionality V m V u [Vx(x g u <-> x e i;) -> u = u]

Empty Set Set(0)

Pairing V m W Set({M, i>})

Union V m S e t([J u

Powerset V m Set(PM)

Infinity 3«[0 e a a Vxen(x+e a)]

Separation Vu1 Vfl Set({,r e a : <p(.r, M j un)})

12
For a set u, Yu is called the power set of u' Until further notice all
propositions, theorems, etc. will be proved in IZ (using the above axioms
and intuitiuonistic logic).

Because we are using intuitionistic logic, we m ust distinguish carefully

between the assertions A * 0 (A is nonempty) and 3x. xe A (A is inhabited).
While an inhabited set is nonempty, the converse does not hold in general.

We write 0 for 0 , 1 for {0} and 2 for {0, 1}. 2 carries the natural ordering <
given by 0 < 0, 0 < 1,1 < 1.

The ordered pair of two sets u, v is defined as usual by

<u,v> = [{u],{u,v}}

Clearly we have

\fu\fv Set(<w, v>).

Proposition 1. <u, v> = <a,b> -» u - a a v = b.

Proof. This m ust be proved w ithout using LEM. Suppose that <u, v> -
<a, b>.

Since {u} is an element of<u, v> , it m ust also be an element of the <a, b>, so
that either {u} = {a} or {u} = {a, b}. In both cases u = a.

Since {u, v) is an element of<w, v>. , it m ust also be an element of <a, b> , so
that either {u, z;} = {a} or \u, v\ = [a, b}. In either case v = a or v = b. If v = a
then it - a - v so that

{fa}} = {{w}/ {«/ v}} = <u, v> = <a, b> = {{a}, {a, b}}.

It follows that {a} = {a, b} so that a - b , and so v = b. So in either case v = b, and

the proposition is proved. ■

We define the Cartesian product of two classes A and B by

A x B= {x: 3ae A3be B(x - <a, b>)).

13
It is left to the reader as an exercise to show that, Set(A) and Set(B) implies
Set(A x B)

A (binary) relation between classes A, B is a subset R c: Ax B. We sometimes

write aRb for <a, b> e R. The doman dom(R) and the range ran (R) of R are
defined by

dom(R) = {x: 3y xRy\ ran (R) = {y: 3x xRy}.

It is left as an exercise to the reader to show that, if Set(R), then Set(dom(R))

and Set(ran(R)).

A relation F is a function, or map, written Fun(F) if for each a e dom(F) there

is a unique b for which aRb. This unique b is written F(a) or Fa. We write F:
A -» B for the assertion that F s a function with dom(F) = A and ran(F) = B.
In this case we occasionally write a F(a) for F.

The identity map 1a on A is the map A -» A given by a ■-* a. If X c: A, the map

x ^ x: X -» X is called the insertion map of X into A.

If F: A -> B and X c A, the restriction F | X of F to X s the map X -> A given

by x •-> F(x). If Y c: B, the inverse image of Y under F is the set

FHY] = {xeA: F(x) e Y}.

Given two functions F: A -> B, G: B -» C, we define the composite function

G o F; A -> C to be the function a >-> G(F(a)). If F : A A, we write F2 for
F o F, F^ for F o F o F.

A function F: A -> B is said to be momc if for all x, y e A, F(x) = F(i/) -> x= y,

and epi if for any F e B there is ^ gA for which b = F(a), and bijcctive, or a
bijection, if it is both monic and epi. It is easily shown that F is bijective if and
only if F has an inverse, that is, a map G: B —> A such that F G = 1B and
G F = l A.Two sets X and Y are said to be equipollent ,and we write X « Y, if
there is a bijection between them.

The projection maps m : A x B -» A and a7n: A x B -> B re defined to be the

maps <a, b> a and <a, b> ^ b respectively.

14
For sets A, B, the exponential BA is defined to be the set of all functions from
A to B. (Exercise: show that this is indeed a set.)

It follows easily from the axioms and definitions of IZ that, for any set A, PA,
under the partial ordering c=, is a frame7 with operations u , n and =>, where

U=> V = { x : x e U -> x e V),

Its top and bottom elements are A and 0 respectively.

For any set a, we write [a | cp} for {x: x - a a cp}; notice that

a e {a | cp} <-> cp.

From Extensionality we infer that {a | cp} - [a | y} iff ((p <-> y); thus, in
particular, the elements of PI (recall that 1 = {0}) correspond naturally to
truth values, i.e. propositions identified under equivalence. PI is called the
frame of truth values and is denoted by Q. The top element 1 of Q is usually
written true and the bottom element 0 as false.

In IZ, Q plays the role of a subset classifier. That is, for each set A, subsets of A
are correlated bijectively with functions A -» Q. To wit, each subset X c A
is correlated with its characteristic function %x : A -» Q given by %x (x)~
{01x e X}; conversely each function fi A -» Q is correlated with the subset
fil(1) = {xeA: fix) = 1} of A.

LOGICAL PRINCIPLES IN IZ

Properties of Q correspond to logical principles of the set theory. For instance,

consider the logical principles (where cp, y are any formulas):

LEM (law of excluded middle) cp v —icp

WLEM (weakened law of excluded middle) —icp v —■—,cp..
DML (De Morgan's law) —i(cp Ay) —» (—.cp v -iy)

In intuitionistic logic WLEM and DML are equivalent.

7 For the definition of frame, see the Appendix.

LEM and WLEM correspond respectively to the properties

VcogQ. (o = true v co = false VcogQ. co = false v co * false.

A class A is said to be discrete (or decidable) if Vx g A VyeA (x = y v x * y).

We then have

Proposition 2. In IZ, each of the following is equivalent to LEM:

(i) Membership is decidable: VxVy ( x e y v i ^ y )

(H) V is discrete : Vx Vy (x = y v x * y)
(iii) Every set is discrete
(iv) Q is discrete
(v) 0 = 2
(vi) Vx (Og x v O ^ x)
(vii) (2, <) is well-ordered, i.e. every inhabited subset of 2 has a <-least element.

Proof. That LEM implies (i) is obvious.

(i) —> (ii). Assuming (i), we have VxVy (x g {y} v x £ {y}), whence (ii).

(ii) —> (iii) and (iii) -> (iv) are both obvious.

(iv) —> (v). Assuming (iv), for any co e Q we have « = {0} or co * {0}. In the
latter case co = 0 , and (v) follows.

(v ) -> (vi). Let (p be the formula 0 e x. T h en , assuming (v), we have {0 | (p} =

{0} or {01tp} = 0 . In the first case we get cp and in the second —«cp. Hence (vi).

(vi) —> (vii). Let U be an inhabited subset of 2. Assuming (vi), we have

0 g U or 0 g U. In the first case 0 is the least element of U. In the second case,
since U is inhabited we m ust have 1 g U, so that 1 is the least element of U.
Thus in either case U has a least element., and (vii) follows.

(viii) -> LEM. Assume (vii) and let cp be any formula. Let U be the subset
{0|cp} u {1} of 2. Then 0 ell <-» cp; moreover U is inhabited and so has a least
element a , which m ust be either 0 or 1. If a - 0, then 0 g li, whence cp;
while if a = 1, then 0 € U, whence —.cp. LEM follows. ■

16
Observe that the negation operation on formulas corresponds to the
complementation operation on Q; we use the same symbol -i to denote the
latter. This operation of course satisfies (using co , co' as variables ranging
over Q)

co c= -ice' <-> o) n co' =false.

Classically, also satisfies the dual law, viz.

-id) cz co' <-» co u co' = true.

But in IZ, this is far from being the case. In fact we can prove

Proposition 3. In IZ, LEM is equivalent to the assertion that there exists an

operation Q -> Q satisfying

(*) c= co' <-» co u co' - true

Proof. If LEM holds, then the complementation operation -1 satisfies (*).

Conversely, suppose given an operation Q -> Q for which (*) holds.

Then

- true ci false <-»false u true = true,

so that -true c=false, whence -true - false. Next,

0 e - coa0 eco - » 0 e - coaco = true -» 0 e -true = false.

Since 0 ^ /n/se, it follows that

0 e -co —^ 0 ^ co —^ 0 e —1co,

and from this we infer that -co c= -nco. Since, obviously, co u -co = true, it then
follows that, for any co, co u -iCO = true, which is LEM. ■

The weak law of excluded middle is also equivalent to a certain property of

the ordered set (2, <):

Proposition 4. In IZ, the following are equivalent:

17
 (i) WLEM
(ii) (2, <) is complete, i.e. every subset of 2 has a <-supremum.

Proof, (i) —> (ii). Assume (i) and let U c 2. Clearly 1 g LI *-» U c {0}. By (i),
we have 1 £ If or -n 1 £ If. In the first case U c: {0} and U then has
suprem um 0. In the second case ~iU ^ {0}, so that —V x g U. x = 0 , which is
equivalent to

V x e U ( x < 0) -> 0 = 1.
Also, obviously,

Vx e U(x < 1) —> 1 = 1.

It follows that, for any u e 2,

(Vx g U.x < u) -> u = 1 ,

so that U has suprem um 1. Thus in either case U has a suprem um, and (ii)
follows.

(ii) -> (i). Assume (ii), let cp be any formula and define U = {1: <p} Then U has
a suprem um a and there are two cases: a - 0 or a = 1. In the first case 1 £ If,
so that -i(p. In the second case, if -icp, then If = 0 and so a = 0, which is
impossible. Therefore —i—icp, and (i) follows. ■

THE AXIOM OF CHOICE

A choice function on a set A is a function / with domain A such that f(a) e a

whenever a is inhabited. The axiom of choice AC is the assertion that every
set has a choice function. While the axiom of choice plays a major role in
classical set theory, in an intuitionistic setting it is far too strong, since even
very weak versions of it can be shown to imply LEM. In fact we have

Proposition 5. It is provable in IZ that if each doubleton has a choice function, then

LEM holds (and conversely).

18
Proof. Let cp be any formula ; define U = {.re 2: x = 0 v cp} and V =
{xe2: x = 1 v cp}. Suppose given a choice fu n ctio n /o n {li, V). Writing a =/(U),
b =f(V), we then have a e U . b e V, i.e.

(a = 0 v cp) a (b= 1 v (p).

Hence
a = 0 a (b= 1 v cp),

whence

(*) a * b v cp.

But
cp —> U —V —^ a
so that
a ^ b —^ —.cp .

This, together with (*), gives cp v -—

.cp. ■

As we have seen, in IZ the Law of Excluded Middle is derivable from AC.

We are now going to show that each of a num ber of classically correct, but
intuitionistically invalid logical principles, including the Law of Excluded
Middle for sentences, is, in IZ, equivalent to a suitably weakened version of
AC. Thus each of these logical principles may be viewed as a choice
principle.

We fix some more notation. For each set A we shall write QA for the set of
inhabited subsets of A, that is, of subsets X of A for which 3x (x e X). The
class of functions with domain A will be denoted by Fun(A).

We tabulate the following new logical schemes

• SLEM a v -id
• Lin (a -» ß) v (ß -» a)
• SWLEM -id V -T-ld
• Ex8 3x[3xcp(x) -» cp (j )]

8 In intuitionistic logic Exi s equivalent to the independence of premises rule:

a->3x<s>(x)
(a-> (p(.v)

19
 • Un 3x[(p(x) —>Vxcp(x)]
• D is9 Vx[a v (p(x)] -> a v Vxcp(x)

Here a and ß are sentences, and cp (x) is a formula with free variable x. In
intuitionistic logic, Lin and SWLEM are consequences of SLEM; and Un
implies Dis. All of these schemes follow, of course, from LEM, the full Law
of Excluded Middle.

We formulate the following choice principles —here X is an arbitrary set and

0(x, y) an arbitrary formula with at most the free variables x, y:

• ACx VxeX 3y 0(x,y) —» 3/eFun(X) VxeX 0(x,/x)

*
• AC, 3/eFun(X) [VxeX 3y 0(x,y) -> VxeX 0(x,/x)]
• DACx V/eFun(X) 3xeX 0(x,/x) -» 3xeX Vy 0(x,y)
*
• DAC, 3/eFun(X) [3xeX 0(x,/x) -> 3xeX Vy 0(x,y)]

The first two of these are forms of the Axiom of Choice for X; while
classically equivalent, in IZ AC*x implies ACx, but not conversely. The
*
principles DACx and DACXare dual forms of the Axiom of Choice for X:
* *
classically they are both equivalent to ACx and AC^ , but in IZ DAC^
implies DACx, and not conversely.

We also formulate the weak extensional selection principle:

WESP 3xg2 cp(x) a 3 x g 2 vp (x) ->

3xG23ye2[cp(x) a v|/( i/ ) a [Vxe2[cp(x) <-> vp(x)] -> x = y]].

Here cp(x) , vp(x) are formulas with the free variable x. This principle asserts
that, for any pair of instantiated properties of members of 2, instances may
be assigned to the properties in a m anner that depends just on their
extensions. WESP is a straightforward consequence of AC q2- For taking 0(w,
y) to be y e u in AC q2 yields the existence of a function / w i th dom ain Q2
such that fn e u for every u e Q2. Given formulas cp(x), ip(x), and assuming
the antecedent of WESP, the sets U = {xe2: cp(x)} and V - {xe2: y(x)} are

9 In intuitionistic logic D is is equivalent to the higher dual distributive law

Vx[a(x) v ß(x)] -» 3xa(x) v Vxß(x).

20
members of Q2, so that a ~ f U e U, and b = JV e V, whence cp(fl) and \\j(h).
Also, if Vxe2[ip(x) <-» \]/(x)], then LI - V, whence a = b; it follows then that the
consequent of WESP holds.

We show that each of the logical principles tabulated above is equivalent

(over IZ) to a choice principle. Starting at the top of the list, we have first:

Proposition 6. WESP and SLEM are equivalent over IZ.

Proof. Assume WESP. Let a be any sentence and define

cp(x) = x = 0 v a ip (x) = x - 1 v a.

With these instances of cp and ip the antecedent of WESP is clearly satisfied,

so that there exist members a, b of 2 for which (1) cp(a) a \\i ( b) and
(2) Vx [[Vxe2[cp(x) <-» iff(x)] a ~ b . l t follows from (1) that a v (a ~ 0 a b = 1),
whence (3) a v a * b. And since clearly a —> Vxe2[(p(x) <-» \j/(x)] we deduce
from (2) that a -> a = b, whence a * b -» -.a. Putting this last together with (3)
yields a v —icx, and SLEM follows.

For the converse, we argue informally. Suppose that SLEM holds. Assuming
the antecedent of WESP, choose a e 2 for which cp(a). Now (using SLEM)
define an element b e 2 as follows. If Vxe2[cp(x) <-» holds, let b = a; if
not, choose b so that ß(fr). It is now easy to see that a and b satisfy cp(fl) a v|/(Z?)
a [Vx e 2[cp(x) <-> i]/(x)] —>« = //]. WESP follows. ■

Next, we observe that, while ACi is (trivially) provable in IZ, by contrast we

have

*
Proposition 7. ACt and Ex are equivalent over IZ.

*
Proof. Assuming AC| , take 0(x,y) = cp(y) in its antecedent. This yields an
f e Fun(l) for which Vy(p(y) -» cp(/D), giving 3i/[3i/cp(i/) -> cp(y)], i.e., Ex.

Conversely, define cp(y) = 0(0,y). Then, assuming Ex, there is b for which
3y(p(y) -> cp(fr), whence Vxel3y0(j,y) -> \/xelQ(x,b). Defining f e Fun(l) by
/ = [{Orb}} gives Vx e 13i/0(x, i/) -> Vxel0(x,/x), and ACj follows. ■

21
Further, while DACi is easily seen to be provable in IZ, we have
k
Proposition 8. DACj and Un are equivalent over IZ.

Proof. Given cp, Define 0(x,y) = ip(y). Then, for / e Fun(l), 3xelQ(x,fx) <->
cp(/D) and E k e lVy0(x,y) <-> Vycp(y). DAC} then gives

3/eFun(l)[<p(/0) -» Vycp(i/)],

from which Un follows easily.

Conversely, given cp, define cp(y) = 0(O,y). Then from Un we infer that there
exists b for which cp(fr) -> Vycp(y), i.e. 0(0,b) -> Vy0(O,y). Defining / e Fun(l)
by / = {(0,1b)} then gives 0(0/0) -> 3 x e l Vy0(x,y), whence 3 xel Q(x,fix) ->
3xelVy0(x,y), and Un follows. ■

Next, while AC2is easily proved in IZ, by contrast we have

Proposition 9. DAC2 and Dis are equivalent over IZ.

Proof. The antecedent of DAC2 is equivalent to the assertion

V/eFun(2)[0(O,/O) v

which, in view of the natural correlation between members of Fun (2) and
ordered pairs, is equivalent to the assertion

Vi/V 1/10(0, xj) V 0(1, y')].

The consequent of DAC2 is equivalent to the assertion

Vye Y0(O,y) v Vy'e Y0(l,y')

So DAC2 itself is equivalent to

VyVy 10(0,y) v 0(1,y')] -» Vy0(O,y) v Vy'0(l,i/).

But this is obviously equivalent to

22
 Vi/Vi/lcp(y) v V(y')l -* Vy(p(y) v VyV(y'),
where y does not occur free in ip, nor y' in cp. And this last is easily seen to be
equivalent to Dis. ■

k
Now consider DAC? . This is quickly seen to be equivalent to the assertion

3z3zf[B(0,z) v 0(1,2') -» Vy0(O,y) v Vy'0(l,y'),

i.e. to the assertion, for arbitrary (p(x), y(x), that

3z32'[(p(z) v v|/(z') -> Vy<p(y) v Vy'ß(y')].

This is in turn equivalent to the assertion, for any sentence a,

(*) 3y[a v i|/(y) -> a v Vyip(y)].

Now (*) obviously entails Un. Conversely, given Un, there is b for which
\\f(b) ->Vyv|/(y). Hence a v \\i(b) -> a v Vyvp(y), whence (*). So we have proved

•k •k
Proposition 10. Over IZ, DAC2 is equivalent to Un, and hence also to DACj .■

In order to provide choice schemes equivalent to Lin and Stone we

introduce

ac* 3fe2x [xeX 3 ye 2 Q(x,y) -> VxeX Q(x,fx)]

W ac 3fe 2X [VxeX 3ye 2 0(x,y) -> VxeX 0(x,/x)] provided

Viz Vx[0(x,O) ->-,0( x/1)]

*
Clearly ftC ^ is equivalent to

3/e2x [VxeX[0(A,O) v 9(.r,l)] -> VxeX 0(x,fx)\

k
and similarly for w a C A Then we have

23
 * *
Proposition 11. Over IZ, B,C[ and w a C j are equivalent, respectively, to Lin and
SWLEM.

Proof. Let a and ß be sentences, and define

0(x,y) = x = 0 a [(y = 0 a a) v (y =1 a ß)].

Then a <-> 0(0,0) and ß <->0(0,1), and so

Vxel[0(x,0) v 0(x,l)] <-> 0(0,0) v 0(0,1) o a v ß .

Therefore

3fe2l [Vxel[0(x,0) v ö(x,l)] -> V xel 0 {x,fx)\

<-> 3fe 2A[a v ß -> 0(0,/))]
f > [ a v ß - > 0(0,0)] v [a v ß ->0(0,1)]
<-> [a v ß -> a] v [a v ß -> ß]
o [ß>ava>ß].

*
This yields a c t -> Lin. For the converse, define a = 0(0,0) and ß = 0(0,1) and
reverse the argument.

To establish the second stated equivalence, notice that, when 0 (x,y) is

defined as above, but with ß replaced by -ia, it satisfies the provisions
*
imposed in w a C j . As above, that principle gives (~^a -> a) v (a -> -»a),
it
whence -»a v -r-.a. So SWELM follows from w a C j . Conversely, suppose
it
that 0 meets the condition imposed in W aCj Then from 0(0,0) -> -iö(0,l) we
deduce -i-iö(0,0) -> -10(0,1); now, assuming SWLEM, we have
-10(0,0) v -,-10(0,0), whence -iö(0,0) v —.0(0,1). Since -iö(0,0) ->
[0(0,0) -> 0(0,1)] and -.0(0,1) -> [0(0,1) -> 0(0,0)] we deduce
[0(0,0) -> 0(0,1)] v [0(0,1) -> 0(0,0)]. From the argum ent above it now
follows that 3fe21 [VxEl[cp(x,0) v tp(x,l)] -> V xe l q>(x,fx)\. Accordingly
waci9*- ’ is a consequence of SWLEM.

24
 Chapter II

Natural Numbers and Finite Sets

THE NATURAL NUMBERS

The natural numbers can be defined in IZ and their usual properties proved.

Let us call a set A inductive if

0 GA A \fx(x G A —» x ' G A).

It follows from the axiom of infinity that there exists at least one inductive set A
Define

N = P)(X : X c A a X is inductive}.

Then N is inductive and is clearly the least inductive set, that is, N c; K for
every inductive set K. The members of N are called natural numbers; thus 0, 1, 2
are natural numbers. We shall use letters m, n, p, ... as variables ranging over N.

Proposition 1.

(i) m e n-> m+ c n.
(ii) n en
(iii) m+ = n+—» m = n.

Proof, (i) Let K = | n: \fin(m e n -> m+ c n).To prove (i) it suffices to show
that K is inductive. Clearly O g K Suppose now that n e K. Then (*)

(*) rn g n m+ c n.

25
If m e n+then me n or m=n. In the first case m+ c= n by (*) and in the second
case m+ - n+ . Thus in both cases m+ c: n+ and so n+ e K. Hence K is inductive
and (i) is proved.

(ii) It suffices to show that the set K - {n: n g n} is inductive. Clearly 0 g K. If n

e K and n+ e n+' , then n+ e n or n+= n. In the first case, it follows from (i) that
n++ c: n, and since n e n++, that n e n. The second case also implies n e n. Thus
in both cases n <£ K. So n e K and n+ e n+' together lead to a contradiction ,
whence n e K ->n+ £ n+. i.e. n e K n+ e K . So K is inductive and (ii) follows.

(iii) It follows from m+ = n+ that m e n+; thus m e n or m - n and so by (i)

m c: n. Similarly n c ///. ■

We shall sometimes write m < n for m e n and m < n for m c n. It follows from
Proposition 1 that m+ < n <-> m < n.

Proposition 2. For arbitrary m, n, exactly one of the following holds:

me n, m = n, n e m.

Proof. Proposition 1 implies that any two of the above assertions are mutually
contradictory. To prove that, for every m, n, one of these assertions holds we
define

K(n) = {m: me n v m = nv n e m}.

We need to show that K(n) = N for every n, and for this it suffices to show that
K(n) is inductive.

The set K(0) is inductive, since K(0) = {0} u{m: 0 g mj and it is obvious that
0 e m —> 0 g //7+.

Now suppose that K(n) is inductive, i.e. N c K(n). We show that K (n+) is also
inductive.

26
Q c K (n+). From the fact that K(0) is inductive, it follows that n+ e N c; K(0),
whence n+e 0 or //+ = 0 or 0 e n+. The first two disjuncts are false, so 0 e n+ and
0 e K (n+).

me K (n+) -» m+ e K (n+). Suppose that me K (n+), that is , me n +or m = n* or

e m. In the second and third cases we obviously have n+ e m+ and hence
m+ e K (n+). In the first case either m = n or me n . If m = n then m+= n+ so that
m+ e K (n+). If m e n, then me K (n), and so, since K(n) has been assumed
inductive, m+e K (n). It follows that m+e n or m+= n or n e m+. The third
disjunct is false, since it implies (using Prop 1 (i)) that n e n, contradicting Prop.
l(ii). Accordingly we have only the two possibilities //7+ e n or m+ = n, both of
which, since n c n+, yield m+ e K (n+). The proof is complete. ■

From Props 2 and l(ii) we deduce immediately the

Corollary. N is discrete. ■

MODELS OF PEANO'S AXIOMS

A Peano structure is a triple A = (A, s, o) where A is a set, s: A -> A, and o e A

.w A is called the domain of A and s the successor operation in A. A model of Peano's
axioms is a Peano structure A such that the following axioms are satisfied:

PI VpeA . sp ^ o
P2 VpeA \/qeA. sp = sq -» p = q.
P3 VX c A[[o e X a Vp(p g X -> sp e X)] -> X - A]

(P3) is the Induction Principle for A: it is clearly equivalent to the scheme: for an
formula cp(x),

MO) a [Vp e A (<p(p) -> <p(sp)]j -> Vp e A cp(p)10

10 Here “o" (Greek omicron) is not to be confused with the set 0. It is just an arbitrary
member of P.
27
A subset K of A satisfying o e K a \/ x (x gK sx g K) is called an inductive
subset of A: the Induction Principle asserts that the only inductive subset of A is
A itself. Establishing that a subset K is inductive is called a proof by induction.

The following facts are easily established by induction:

In any model A ofPeano's axioms,

sp * p
Vp g A[p = o v 3q. p = sq ].

If we define s: N -> N by sn = n+, then it follows from the fact that Prop.l (iii)
and the fact that N is the least inductive set that (N, s, 0) is a model of Peano's
axioms.

DEFINITIONS BY RECURSION

Just as in classical set theory, in IZ any model of Peano's axioms admits

functions defined by recursion. Let us say that a Peano structure A = (A, s, o)
satisfies the Simple Recursion Principle if:

Given any set X, any element a g X, and any function e: X -> X, there exists a unique
fiinction f A -» X such that

A °) =z(l Vp e A f(sp) = e(fp).

Proposition 3 Any model ofPeano's axioms satisfies the Simple Recursion Principle.

Proof. Let A - (A, s, o) be a model of Peano's axioms. To simplify notation we

shall use letters p, q to denote variables ranging over A.

Define

U ^ { u g P(A x X ) : (o, a) gu a VpVx((p, x) gu -~>{sp, e(x )) e u }.

28
Let / = p | U . We claim that / satisfies the conditions of the proposition; its
uniqueness is left as an exercise to the reader. To show that / satisfies the
conditions of the proposition it clearly suffices to show th a t/is a map from A to
X.

Clearly / c A x X . Let V ~ { p : 3 x.(p, x) g /} . Then V ^ A and o g V since

(o, a) e / . Moreover

p g V —>3x.(p, x) g f —>3 x ( s p ,e(x )) x / -> s p g V.

So V is inductive, whence V = A, a n d /is defined on A.

It remains to show th a t/is single-valued. To this end define

K ^ { p f VxVi/Kp, x ) g f a <p, zy) e / -> x - zy)}.

We need to show that K is inductive.

First, we show that og K . Define

/ * - {<p,x> g f \p —o > x —ci}.

It is easily verified that /* g (i, w h e n c e /= /f. Therefore

(o, x) Gf -> (o, X) G / * -» X = a,

so that
V x « o ,x ) g / -> x = a ) ,

from which it follows immediately that o g K.

Finally we need to show that p g sp g K. To do this we first establish the

•auxiliary result

( 1) (p g K a < p ,x > g / a < s p ?zy) e / ) zy = e ( x ) .

29
Given p eK satisfying ( p 9x) g f a ( s p ,y ) g / , define

f px = {<<?>£/> e f ' - q = s p - + y = e(x)}.

We claim that f px eU . First, from P I it follows th a t(0 ,a) g / px.Now , if

( q , y ) e f px, then (q,y) g j and q = sp -> y = e(x). We need to show that
(sg, e(i/)) g / px and for this it suffices to show that

(2) (sq,e{y)) e j

and

(3) sp =sq —» e(y) = c(x).

Assertion (2) follows from the assum ption that (q9y) e J . As for (3), if sp = sq,
then, by P2, p = y, so from (g,y) g 7 it follows that ( p 9y) g 7 • But we are
assuming that p e K and (p 9x) gj , so we conclude from the defining property
of K that x = y. Hence, certainly, e(x) = e(y), proving (2).

We conclude that f px g U. From this it follows that f c; / pv, so that

(p g K a (p ,x ) g f a (sp 9y ) g / ) -> g U

whence

{ p e K A ( p , x ) e f a (sp, y) e / ) - > <sp,y> e / px

- > y = e(x).

Thus (1) is proved.

Now, we know that V = A, i.e. V p3z.(p,x) e f . it follows from this and (1) that

30
(p eK a (sp,y) e / a {s p ,z ) g / ) -> 3x.(p,x >e / a p e K a (sp,p> e / a <sp,z> e /
-> 3x ( p G K A < p ,x ) e / A (sp,y) G / A ( s p ,z > e / )
-> 3x(iy - e(x) = e(x))
y ~ z.

Therefore p e K -> sp e K, and it follows that K is inductive, so that K = A and /

is single-valued. This completes the proof of Proposition 3. ■

An isomorphism between two Peano structures A = (A, s, o) and A' = (A', s, o') is
a bijection /: A -» A' such th at/(o) = o' and for all p e A, f(s(p)) = s'(f{p)). Two
Peano structures are isomorphic if there is a an isomorphism betw een them.

Corollary 1. Any two models of Peano's axioms are isomorphic.

Proof. Given two Peano structures A = (A, s, o) and A' = (A', s', o'), by Prop. 3
they both satisfy the Simple Recursion Principle, so that there are maps f A A'
and g: A' -> A such that

/( o) = o ' & for all p g A, f(s(p)) = s'(/(p)).

g(o') = o' & for all q e A', g(s'(q)) = s(g(q)).

We claim t h a t /i s an isomorphism between A and A'. For this to be the case it

suffices to show that g is an inverse to / i.e. g f = l Aand / g = 1A To prove
the first assertion it is enough to show that the set K - {peA: g(f(p)) = p } is
inductive, and this is a straightforward consequence of the definitions o f /a n d g.
The proof of the second assertion is similar. ■

Corollary 2. The domain of any Peano structure is discrete.

Proof. By the Corollary to Prop. 3, N is discrete, and is the domain of a model of

Peano's axioms. If A is the domain of a model of Peano's axioms, it is, by
Corollary 1, bijective with N, and it follows easily from this and the discreteness
of N that A is discrete. ■

31
Let us say that a Peano structure A = (A, s, o) satisfies the Extended Recursion
Principle if:

given any set X, any element a e X, and any function e : X x A —» X , there exists a
unique function fi A —» X such that

ßp) = a a \/p gA. f(sp) = e(<fp, p>).

Proposition 4. Any Peano structure that satisfies the Simple Recursion Principle also
satisfies the Extended Recursion Principle.

Proof. Let A = (A, s, o) be a Peano structure satisfying the Simple Recursion

Principle. We use p as a variable ranging over A. Given a set X, an element
a e X, and a function e : X x A -> X , let Y = X x A and let h: Y -» Y be the
function given by

h = (x,p)\~>{g({x,p)),p).

Applying the Simple Recursion Principle to h, Y and (a, o) yields a unique

k: A -> Y such that

k(o) - (a, o) a Vp. /c(s(p)) - h(k{p)).

It is now easily checked that f = nx k : A —» A is the unique map such that

/(o) —a a Vp g A. f(sp) = e(<fp, p>).

The Extended Recursion Principle follows. ■

We use Proposition 4 to prove the converse of Proposition 3, namely

Proposition 5. Any Peano structure that satisfies the Simple Recursion Principle is a
model of Peano's axioms.

Proof. Let A = {A, s, o) be a Peano structure satisfying the Simple Recursion

Principle. Then by Prop. 4 A also satisfies the Extended Recursion Principle. We

32
need to verify that A satisfies PI, P2 and P3. Again vve use p, q as variables
ranging over A.

PI: Vp . sp ^ o. Recalling that 2 = {0, 1}, define g: 2 -> 2 by y = |<0,1>, <1,1>}.

Using the Simple Recursion Principle, there is fi P -> 2 such that

/(o) = 0 a Vp. /(s p ) = g(fp).

Then

s p = o -> 0 = / ( o) = /(s p ) = q(/p) = 1;

since 0 * 1, PI follows.

P2: VpVp. sp = sq p = q. Consider n2 : A x A A. By the Extended Recursion

Principle there is a map f: A -> A such that

/ ( o) = o A Vp. / ( p ) = n2((fp , p>) = p .

Then

s p = sq -> q = /(s q ) = /(s p ) = p.

and P2 follows.

P3: the Induction Principle. Suppose K ^ A is inductive . Then p •-> sp: K -> K
and so the Simple Recursion Principle furnishes a m ap/: A -> K such that

/(o) = O A Vp. /(s p ) = s(/p).

Writing ; for the insertion map of K into A, we get

(*) U f)(o) = O A Vp. (j o /)(s p ) = s(j /)(P ).

33
But the identity map ty: A -> A also satisfies (*), so from the uniqueness
condition in the Simple Recursion Principle we infer that j f = l A. It follows
easily from this that K - A, and the Induction Principle follows. ■

A set A is Dedekind infinite if there exists a monic map f: A -> A and an element
a e A such that a & ran(/). We next note that each Dedekinf infinite set gives rise
to a model of Peano's axioms.

Proposition 6. Each Dedekind infinite set contains the domain of a model of Peano's
axioms.

Proof. Let A be Dedekind infinite, f: A A monic, and a e A such that

a <£ ran(/). Define

U = f]{u c : A : a e u a Vx g u . f( x ) g u }.

It is then easily shown that (U ,f a) is a model of Peano's axioms. ■

Corollary. A set A is Dedekind infinite if and only if there exists an injection N -> A.

FINITE SETS

There are a num ber of possible definitions of the concept of finite set in IZ. To
introduce (some of) these, it will be convenient to fix a set £. By an E-family or E-
singleton we shall mean ’’set of subsets of £", or"singleton of £ M, respectively. For
a subset X of £ we define

K(X) = X is in every £-family containing 0 , all E-singletons, and dosed under

unions of pairs of its members. If K(X) holds, we shall say that X is a Kuratozuski
finite subset of £.

L(X) = X is in every E-family containing 0 and dosed under unions with £-

singletons. If L(X) holds, we shall say that X is a finite subset of E.

34
M(X) = X is in every E-family containing 0 g ;>and closed under unions with
disjoint E-singletons, that is, if VXg ,> Vxg E-X(X u {x} g ./]. If M(X) holds, we
shall say that X is a strictly finite subset of £.

We shall also write D(X) for " X is discrete".

Lemma 1. VXc=£[M(X) -> L(X)].

Proof. Obvious. ■

Lemma 2. VXc=£ [K(X) <-> L(X)].

Proof. Clearly L(X)-> K(X). To prove the converse, it suffices to show that the
family 7,=z {X^E: L(X)} is closed under unions of pairs. To this end let O(LE) be
the property V X g c\J U u X g 7i It suffices to show \ZU[L(U) —> O (U)]. Clearly
0 (0 ). Assuming 0(1/) and X g 9Xve have U u X g ^ a n d so II u X u {x} g '/
for arbitrary x, whence 0(1/ u{xj). Hence \/U[L(U) -> 0(1/)] arid the result
follows. ■

Lemma 3. VX[M(X) -> D(X)].

Proof. Obviously D (0). If D(X) and x £ X, clearly D(X u {x}). The result follows.

Lemma 4. V X e £ [M(X) -> Va[D(X u {a}) -> (a g X v a £ X)]].

Proof. Write O(X) for the condition following the first implication. Clearly 0 (0 ).
Suppose that O(X) and x £ X. If D(X u {x} u {*?}), then D(X u {a}), so, since O(X),
either a e X v a £ X. Since D(X u {x} u{a}), it follows that a = x v a * x. Hence

(a g X a a = x) v (a £ X a a = x) v (a g X a n ^ x) v (a £ X a a * x),

The first three disjuncts each imply a g X u {x}, and the last disjunct means
a £ X u {x}.

Accordingly a g X u {x} v a <£ X u {x}. We conclude that 0(X u {x}) and the
result follows. ■
35
Lemma 5. VXc=£ \L(X) a D(X) -> M(X)].

Proof. We need to show VX[L(X) -» 0(X)J, where <£>(X) is D(X}-> M(X). Clearly
<£>(0). Assume <£>(X) and D(X u {a}). Then D(X), so, since O(X), it follows that
M(X). Since D(X u {a}), Lemma 4 gives fl e X v a € X. In either case we deduce
that M(X u {a}). Hence 0(X u {a}), and the result follows. ■

From these lemmas we immediately infer

Proposition 7. For any set E, the families of strictly finite, discrete finite, and discrete
Kuratowski finite subsets coincide. ■

We can now define a set £ to be strictly finite, finite, or Kuratowski finite if it is,
respectively, a strictly finite, finite, or Kuratowski finite subset of itself.

Proposition 8. A set is strictly finite if and only if it is bijective with a natural number.

Proof. Suppose that £ is strictly finite, arid for X c £ let O(X) be the property
"X is bijective with a natural number. We need to show that 0(£), and for this it
suffices to show that, for all X c E, M(X) -» <t>(X). Clearly we have 0 (0 ). If O(X)
and x e E - X, le t/: n -» X be a bijection between some natural num ber n and X.
It is easily checked that the set g = / u {<n,x>} is a bijecfion between n+ and
X u {x}. Hence 0(X u {x}) and the result follows.

Conversely, suppose that £ is bijective with a natural number, i.e. 0(£). We

want to show that £ is strictly finite, and for this it suffices to show that the
subset K of N given by

K = {n g I I : V X c; E( X ~ n > X strictly finite)}

is inductive. Clearly 0 e K. Now suppose ne K, and X « n \ L e t/b e a bijection

between n+and X, let a = fin) and let X' = X - {a}. Then the restriction f\ n is a
bijection between n and X' and so, since ne K, it follows that X' is strictly finite.
But then, since a £ X', it follows that X = X' u {a} is also strictly finite. Hence
n+ e K, so that K is inductive. This completes the proof. ■

36
Finally, a set £ is Dedekiud finite if it is not Dedekind infinite, i.e. if there does not
exist a monic / : £ —»£ and an element a g £ such that a £ ran(/).

Proposition 9. Every strictly finite set is Dedekind finite.

Proof. To prove this, it suffices to show that 0 is Dedekind finite and

(*) X Dedekind finite & a <£ X X u {a} Dedekind finite.

It is obvious that 0 is Dedekind finite. To prove (*), suppose that a <£ X and
X u {a) is Dedekind infinite. We show that X is Dedekind infinite. Since X u {a\
is Dedekind infinite, there is a m onic/: X u {ö) -> X u {a} and b e X u {a} such
that b £ ran if). There are two cases: b = a or b e X. In the first case,
/ X u {a} -» X ; the restriction / 1X : X -> X is then monic and
f(a) g X - mn(f\ X), so that X is Dedekind infinite. The second case, b g X, splits
into two subcases: f(n) = a or f(a) g X. In the first subcase, / | X : X -» X is
monic and b g X - ran(/*| X), so that again X is Dedekind infinite. In the second
subcase, f(a) g X, there is a unique xo g X for which /(xo) = so that
xo £ / l[X] and it is then easily shown that X = / 1[X] u {x0}.Now define

9 = f \ r 1[ X] v { ( x 0,f(a)\

It is easily shown that g is a monic map from X to X and b £ ran(g). Thus X is

Dedekind infinite and the result is proved. ■

We conclude this section with

Proposition 10. Q is Dedekind finite.

This is an immediate consequence of

Proposition 11. If fi Q Q is monic, thenfi = 112, so that f is also epi.

Proof. In the proof we shall use the easily established fact that for co,u g Q,

(1) oy=o <-» (co^l <-> u=l).

37
We first prove

(2) /(co) = 1 -> /(© ) = 03.

Assume/(co) = 1. We show t h a t / ( go) = 1 <-» co = 1, from which (2) then follows by

( 1 ).

First we have

co = 1 -» co = /( cd) - > / ( co) = /(co) =1 .

Conversely

/(co) = 1 ->/(co) = 1 = /(co)-> co =/(co) =1,

as required.

Finally we use (2) to prove

(3) /(co) =/o>),

from which we infer /(co) = co, so t h a t / = 1Q.

To prove (3), by (1) it suffices to show that

(4) /(co) — 1 <-*/(co) = 1.

If /(co) = 1, it follows from (2) that/(co) = co, whence/(co) = /(«) = L Conversely,

if/(co) = 1, then /(/(co)) = 1, so by (2) /(co) = /(/(co)) =/(co). It follows that
l=/(co) = /((o). This proves (4), and the Proposition. ■

FREGE'S CONSTRUCTION OF THE NATURAL NUMBERS

By Frege's Theorem is m eant the result, implicit in Frege's Grundlagen, that, for
any set £, if there exists a m ap v from PE to £ satisfying the condition

38
 VXV Y[ V(X) = v(Y) <r>X * Y]n ,

then £ has a subset which is the domain of a model of Peano's axioms. We are
going to show that a strengthened version of this result can be proved in IZ.

Let us call a family of subsets of a set £ strictly inductive if it contains 0 and is

closed under unions with disjoint E- singletons. We define a Frege structure to be
a pair (£, v) with v a map to £ whose domain dom(v) is a strictly inductive
family of subsets of E such that

YXe dom(v)VY e dom(v) [v(X) = v(Y) <-> X % Y].

A Frege structure (£, v) is strict if dom(v) is the family of strictly finite subsets of
E.

We now prove

Frege's Theorem. Let (£, v) be a Frege structure. Then we can define a subset N of E
which is the domain of a model of Peano's axioms.

Thus suppose given a Frege structure (£, v).The proof of Frege's Theorem breaks
down into a sequence of lemmas.

For X e dom(v) write Xt for X u {v(X)}. Call a property O defined on the

members of dom(v) v-inductive if 0 ( 0 ) and, for any Xe dom(v), if <P(X) and v(X)
g X, then 0 ( Xt). Call a subfamily of dom(v) v-inductive if the property of
being a member of is v-inductive. Then dom(v) is v-inductive, as is the
in tersection^of the collection of all v-inductive families. From the fact that A is
the least v -inductive family we infer immediately the

Principle of v- Induction for A For any property O defined on the members ofJV, z/O
is v- inductive, then every member of Alias O.

Lemma 1. For any X e A,

11 Recall that X « Y means: there is abijection between X and Y.

39
 X = 0 or X = Y*for some Y g N such that v(Y) <£ Y.

Proof. Write O(X) for this assertion. To establish the claim it is enough, by the
Principle of v- Induction, to show that O is v-inductive. Clearly 0 (0 ). If O(X)
and v(X) g X, then evidently 0 ( Xt). So O is v-inductive. ■

Lemma 2. For any X g JVand any x g X,

t/icre i s Y e A such that Y c X and x = v(Y).

Proof. Writing O(X) for this assertion, it suffices to show that O is v-inductive.
Clearly 0 (0 ). Now assume O(X) and x g Xt. Then either x g X, in which case,
since O(X) has been assumed, there is Y g N for which x = v(Y) and Y c= X, a
fortiori Y c Xt. Or x = v(X), yielding the same conclusion with Y = X. So we
obtain 0(X +), O is v- inductive, and the Lemma follows. ■

Lemma 3. I f X J c E, x g £ - X, y g £ - Y, and X u {x} « Yu{j/}, then X « Y.

Proof. Assume the premises and let / be a bijection between X u {x} and
Y u {i/}. We produce a bijection f between X and Y. Let i/' be the unique
element of Y u {]/} for which <x, i/’> g /. Then either i/' = y, in which case we
t a k e / = / | X, or y' g Y, in which case the unique element x' g X u {x} for
which <x',y> g / satisfies x ! g X. (For if x' = x then <x,y> g / , in which case y' =
y ^ Y.) So in this case we define

/ = \f n (X x Y)] u \<x',y'>\.

In either case it is easily checked that / is a bijection between X and Y. This

proves the Lemma. ■

Lemma 4. For all X, Y in JV,

v(X) = v(Y) -> X = Y.

Proof. Write O(X) for the assertion X g JVand VY g *A[v (X) = v(Y) -> X = Y]. It
suffices to show that <t> is v-inductive. 0 ( 0 ) holds because v(0) = v(Y) -> Y ^ 0

40
-> 0 = Y. Now assume that <h(X) and v(X) £ X; we derive <£>(Xt). Suppose that
Y e N and v(Xt) = v(Y). Then X+ Y, and so in particular Y ^ 0 . By Lemma 1,
there is Z e N for which v(Z) £ Z and Y = Zt, so that X+« ZL We deduce, using

Lemma 3, that X « Z, so, since we have assumed O(X), X = Z. Hence

Xt = Zt = Y, and (£>(X+) follows. So O is v-inductive and the Lemma proved. ■

Lemma 5. For any X e A,

v(X) £ X.

Proof. It suffices to show that the property v(X) & X is v-inductive. Obviously
0 has this property. Supposing that X e Af v(X) £ X but v(X+) e Xt, we have
either v{Xt) = v(X) or v(Xt) e X. In the former case X = Xt by Lemma 4, so that
v(X) g X, a contradiction. In the latter case, by Lemma 2, there is Y e «yLsuch that
Y c X and v(Xt) = v(Y). Lemma 4 now applies to yield Xt = Y c X, so again
v(X)e X, a contradiction. Therefore v(X) g X -> v(Xt) ^ Xt, and the Lemma
follows. ■

Notice that it follows immediately from Lemma 5 that A ls closed under f, that is,
X G e\ Xt G A.

Now define o = v(0), N = (v(X): X g A}, and s:N -> N by s(v(X)) = v(Xt) for X
g A. Then s is well defined and monic on N. (For if v(X) = v(Y), then, by Lemma

4, X = Y, and so s(v(X)) = v(Xt) - v(Yt) = s(v(Y)). Conversely, if s(v(X)) = s(v(Y)),

then v(Xt) = v(Yt), so that, by Lemma 4, X+ Y+. Lemmas 3 and 5 now imply

X « Y, whence v(X) = v(Y).) Clearly, also, o * sx for any x e N. The fact that the
structure (N, s, 0) satisfies the Induction Principle follows immediately from the
Principle of v-induction for N. Accordingly (IV, s, o) is a model of Peano's
axioms, as required.

The proof of Frege's Theorem is complete.

We next establish a converse to Frege's Theorem, namely, that an}' set

containing the domain of a model of Peano's axioms determines a map which
41
turns the set into a strict Frege structure: And finally, we show that the
procedures leading from strict Frege structures to models of Peano's axioms and
vice-versa are mutually inverse. It follows that the postulation of a (strict) Frege
structure is intuitionistically equivalent to the postulation of a model of Peano's
axioms.

Let A = (A ,s, o) be a model of Peano's axioms; we use letters p ,p, r as variables

ranging over A. Using the Simple Recursion Principle, define k: A -> PA to
satisfy the equations

fc(o) = 0 k(sp) = k(p) u {p}.

Lemma 6. (i) p e k(q) -> k(p) c; k(q).

(ii) p £k(p).

Proof, (i) Let K = {q: Vp[p ek(q) -> k(p) c k(q)]. Obviously o e K. If q e K, then
p efc(p) -> k(p) e k(q), so that

p e k(sq) = k(q) u {q} -> p e k(q) v p = p.-> k(p) c fc(sp).

Thus K is inductive and (i) follows.

(ii). Let K = {p : p £Ä:(p)}. Obviously o e K. Suppose that p e K. So if sp ek(sp),

then sp - p or sp efc(p). The first case is impossible and the second case, using
(i), yields = k(p) u {p} = fc(sp) c k(p) whence p e /c(p), contradicting p e K. Hence
sp 0 fc(sp), i.e. sp e K. Hence K is inductive and (ii) follows. ■

Lemma 7. For all p, q, k(p) ^ k(q) <-» p = q.

Proof. Write 4>(p) for Vq[k(p) « k(q) *-» p = p]. Then clearly O(o). If ct>(p) and

fc(p) « k(sp) ~ k(p) u jp}, th e n q* o so thatp = sr for some r. Hence

/c(r) u {r} = A:(sr) = A:(p) ^ fc(sp) = fc(p) u {p}.

42
Since, By (i) of Lemma 7, r <£ k(r) and p £ k(p), Lemma 3 implies that k(r) ^ k(p),
so, since O(p), r = p and q = sr = sp. Hence O(sp), and the result follows by
induction. ■

Now suppose that £ is a set such that N c E . Define

v = {<X ,p> g P E x A: X ^ g(p)}-

Lemma 8 .. dom(v) is the family of strictly finite subsets ofE.

Proof. We need to show that dom(v) is the least family of subsets of £ which
contains 0 and is closed under unions with disjoint £-singletons: let us call such
a family strictly inductive. First, dom(v) clearly contains 0 . If X e dom(v), then
X % k(p) for some p. If x £ X, then X u [x] zz k(p) u {p} = k(sp), whence
X u jx} c dom(v). So dom(v) is strictly inductive. And dom(v) is the least strictly
inductive family. For suppose that <>is any strictly inductive family. For each p
let .I? p = {X: X « g(p)\- We claim that .W p c F for all p. For obviously
J f0 = {0} c; F. Now suppose that c: 3. If X ^ k(sp), then X zz k(p) u {/?}, so for
some x e X (which may be taken to be the image of p under a bijection between
k(p) u {p} and X), we have X - |x| ^ k(p). It follows that X - {x} e JCyn c //, and
so X - (X - {x}) u {x} e ca The claim now follows by induction; accordingly
dom(v), as the union of all the is included in F. Therefore dom(v) is the least
inductive family and the Lemma is proved. ■

Lemma 9 .. v is a function and X zz k(v(X)) for all X e dom(v).

Proof. Suppose that <X, p> e v and <X, q> e v. Then X ^ k(p) and X ^ k(q)
whence k(p) ^ k(p) and so p = q by Lemma 7. The remaining claim is obvious. ■

Lemma 10. For all X, Y e dom(v), X ^Y <-> v(X) = v(Y).

43
Proof. We have, using the previous Lemma, v(X) = v(Y) <-» k(v(X)) zz k(v(Y)) <->

X « Y. ■

Lemmas 8 and 10 establish

Proposition 12. (E, v) is a strict Frege structure. ■

(E,v) is called the strict Frege structure associated with £ and the model A of
Peano's axioms.

Finally, we show that the processes of deriving models of Peano's axioms from
strict Frege structures and vice-versa are m utually inverse.

Suppose that we are given a strict Frege structure (£, p). Recall that the
associated model (A ,s, o) of Peano's axioms is obtained in the following way.
First, the family A is defined as the least subfamily of dom(p) containing 0 and
such that, if X e «4^and p(X) £ X, then X u {p(X)} e */h it having been shown that
p(X) g X for all X e Jf. The associated model (A ,s, o) of Peano's axioms was
then defined by N ~ {p(X): X e df), s(p(X)) = p(X u p(X)}), and o = p(0).

We observe that since (£, p) is strict, for any X e dom(p) there is a (unique)
X* e e^for which X ä X*, and so p(X) = p(X*). To prove this, it suffices to show
that the set of X e dom(p) with this property contains 0 and is closed under
unions with disjoint singletons. The first claim is obvious. If X e dom(p), x € X,
and X rz X* with X* e Jt, then

X u { i } ^ r u { p ( X l ! edf,

since, as observed above, p(X*) ^ X*. This establishes the second claim, and the
observation.

Now let (£, v ) be the strict Frege structure associated with the model (A ,s, o) of
Peano's axioms in turn associated with (E, p). We claim that p = v. To prove this
it suffices to show that

44
(*) X % fc(n(X)) for all X e Jt,

where is defined as above. For then, by Lemma 9, we will have

/c(v(X)) « X ^ 7c(jli(X)) and so p(X) = v(X) by Lemma 7. This last equality for all
X Gd^in turn yields p(Y) = v(Y) for all Y e dom(p) = dom(v). For then, by our
observation above, p(Y) = p(Y*) = v(Y*) = v(Y).

So it only remains to prove (*). It is clearly satisfied by 0 . If X ä k(p(X)) with

X e Jf, then, since p(X) g X,

X u MX)} « /c(p(X)) u MX) } = k(sp(X)).

(*) now follows from the definition of A. So our claim that p = v is established.

Conversely, suppose we are given a set £ and a model (A ,s, o) of Peano's axioms
with A c E , Let (E, v) be the associated strict Frege structure. We note first that,
for any p e A, we have v(/c(p)) = p. For by Lemma 9, k(p) ^ k(v(k(p)), so that, by
Lemma 7, p = k{y(p)). Now let (A5" ,5* o*) be the model of Peano's axioms
associated with the Frege structure (£,v). We claim that (A ,s, o) and (A* ,5*, o*)
are identical.

First, A* = {v(X): X e Jt}, where J t is the least subfamily of dom(v) containing 0

and such that X <e J t and v(X) £ X implies X u {v(X)| e J t . Using the fact that
v(k(p)) = p for all p e A, it is easily shown that J t - [k(p): p e A}. Thus
A* = }v(X): X e J t} = {v(fcp)); p e A} = {p; p e A} = A. Finally o* = v(o) = v(g(o))
= o and

s*(p) = s*(v(k(p))) = v(fc(p) u {v(k(p))}) = v(k(p) u {/>}) = v(fc(sp)) = sp,

so that s* = s by induction.

Thus we have established that the two processes are m utually inverse.

45
 Chapter III

The Real Numbers

DEDEKIND REAL NUMBERS AND WEAK REAL NUMBERS

We turn now to the construction of the real num bers in IZ. This is done in
essentially the classical manner: first, the (positive and negative) integers are
constructed, next, the rationals, and finally the reals obtained from the latter as
Dedekind cuts or (equivalence classes of) Cauchy sequences. In the classical
context it is well known that these methods of constructing the reals lead to
isomorphic results. This need not be the case in IZ. Neither is it necessarily the
case that the reals are (conditionally) order-complete, or even discrete.

The set Z of positive and negative integers is constructed within IZ in the

usual w ay1213. It is shown in the standard m anner that Z may be turned into an

ordered ring (Z , +, •, <) .

The customary procedure for obtaining the rational field as the ordered field of
quotients of Z 13 now yields, in IZ, the ordered field of rationals (Q , +, *, <) . Note

that Q, like N and Z, is discrete. We shall use letters p, q as variables ranging

over Q.

Now we can define the Dedekind real numbers as "cuts" in Q. Thus a Dedekind

real number is a pair <L, R> of inhabited subsets L, R c Q satisfying

12 See, e.g. Mac Lane and Birkhoff [1967], Chapter II, section 4.
13 Mac Lane and Birkhoff (1967), Chapter V section 2.
46
(1) LHR =0
(2) Vp[p gl <->3q e L.p < q]
(3) Vp[p g R <-» 3q g R.q < p]
(4) V p V q [p < q -+ p e L V q Gr ]

We write Md for the set of Dedekind real numbers.

We also define a weak real number to be a pair is a pair < l , r > of inhabited
subsets L, R c Q satisfying (1), (2), (3) and the pair of conditions (both weaker
than (4)):

VpVq[p < q r\ p £ L —> q g jr]

(4*)
VpVq[p < q A q s ? R - > p e L ]

We write Ew for the set of weak real numbers. Clearly Ed c Ew. We shall use

letters r, s as vas variables ranging over EW/ and write r - < L r, R r>.

Note that, for any weak real num ber r, R r is recoverable from L r in that

R, = { p : 3q < p(q e Lr )}.

The ordering < on E w (and its restriction to Mj) is defined by 14

r <S <-> Lr Cj Ls .

The strong ordering on E w(and its restriction to Ed) is defined by

r < s<r+ 3p (p e Rr A p e Ls).,

44 The asymmetry in this definition is only apparent since from the recoverability of R
from l , noted above, follows that L.r c; l s <-> Rs c: r ,.
47
Clearly < is a (partial) ordering on Rw (hence also on Rd), and it is

straightforward to show that < is irreflexive and transitive.

While classically it can be shown that r <s o r < s v r - s, this does not hold
in IZ. What can be proved is

Lemma 1. (i) In Rw, r < s <- > -n(s < r).

(ii). In Rd,, r < s —» Ve(r < e v e < s ) .15

Proof, (i) From r < s a n d s < r we deduce 3p(p e Rs Ap g u ), contradicting

(1), so r < s —» -1 (s < r ). Conversely, —* (s < r ) is equivalent to
Vp[(-i(p e r s) a—i(p e Lr)]. So from q e Lr it follows from (2) that
3p(p > q a p e L r ), whence. Thus by (4') q e L s .So Ls C L S and r < S.

(ii). If r, s g Rd and r < s, we get rational p < q for which p e Rr a q g . If now

e g Rd, then p e U v q e Re. It follows that

(P e Rr a p e Le) v (q e L. a <2 e R j :

The first disjunct implies r < e and the second e < s. ■

In the classical case one now proceeds to show that Rd is conditionally order-

complete, i.e. every inhabited subset with an upper bound has a least upper
bound. The argum ent for the conditional order-completeness of Rd requires an
application of LEM which is not available in IZ. On the other hand, we shall
prove in IZ that Rwis conditionally order-complete. Moreover, we shall show in

IZ that, if De Morgan's Law —.(cp a vp) -> (—.tp v —<\p) holds, then Rd is

conditionally order-complete, and conversely.15

15 This condition does not necessarily hold in Ru .

48
Proposition 1. In IZ, Mwis conditionally order-complete.

Proof. Let X be a bounded inhabited subset of Rw. We w ant to construct a least

upper bound rx = < l , R> for X. Define R, l by

R = {q :3 p < qVr e X (p e R r )}

L = {p : 3q(p < q /\q & r )}.

It is easy to verify that rx satisfies conditions (1), (2), (3) and the second
condition in (4') above. To verify the first condition in (4'), suppose that
p <q a p £ l . Define

Then p <p' < q < q. From the definition of L it follows that -r-i(p' e R) . But since
R c: R, for all r e X, it m ust be the case that R n Lr = 0 for all r e X, whence

p x£ Lr for all r e X . l t follows that Vr e X(q' e R r ), so that q e R.

Since L, c= L for all r e X, Tx is an upper bound for X. But if s is any upper

bound for X, then R* c= &r for all r e X, and it follows easily from this that
r s c= R . Hence s <rx and we are done. ■

Proposition 2. The following are equivalent in IZ:

(i) DML (or, equivalently, WLEM)

(ii) Md = Kw

(iii) Ed is conditionally order-complete.

Proof, (i) —> (ii). Assume (i) arid let r e Rw. To show that r e Ed, suppose given

rationals p, q with p < q and let e = Vi(p + q). From the disjointness of L r and R,
together with DML, it follows that e ^ R r v e a r . Gut since p < e the first

49
disjunct implies p e L r and the second similarly implies q e R r It follows that
r e Ed, whence (ii).

(ii) -» (iii) is an immediate consequence of Proposition 1.

(iii) -> (i). Assume (iii) and let cp be any formula. Then the set

S - {q e Q : q = 0 v (q = 1) a cp}

is inhabited and bounded in Ed. Let r be the suprem um of S. Then we have

cp -> r = 1 a n d —.cp —> r = 0 , so that r < 1 -> —.cp a n d 0 < r —» —.—icp. By (ii) of
Lemma 1, 0 < r v r < 1, and we deduce that —i—i(p v —«cp. WLEM and hence
DML follows. ■

We next sketch how the operations of addition and multiplication can be

defined on Ed so as to turn it into a communicative ring16.

Addition on E j is defined by

< L jjR j > + < L 2 , R 2 >

= {Pi + Pi : Pi e Li A P 2 e l.2},{a + P 2 : A e r ! ap2 e r 2 }.

The definition of multiplication is more involved, since without LEM we

cannot, as we can classically, divide into cases according to the signs of the
multipliers. In fact we first define the product of a real num ber and a rational,
which can be divided into cases since the usual ordering on the rationals
satisfies the trichotomy law. Thus we define:

< { p - q : q G L r j , { p - q : q G R r}) if p > 0

p-r J <{p q : q e R r},{p q : q e L r }) i f p < 0
J) if p = 0.

16 In fact the procedure introduced here works equally well for Ew.

50
To define the product of r19r2 e E d , we use the idea that if p x - r;a n d p 2 - r 2
have the same sign then (rx - p^)(r2 ~ p 2) should be positive. Thus we define
Lr
ri rri to be

{P ^ Q -^P iP M P i ap 2 eL^Jvf p, 6R, a p 2 e R . j A t p + pj - p 2 e L pjri.ft,J]}

and r to
l ri o be

{peQ-lPlPlHiP! e R r, A A > G L rJ V ( P l AP'2 G R r2 ) A ( P + A 'P 2 e R p2r, , A r2 )]}

A tedious verification shows that these definitions of addition and

multiplication do indeed turn Ed into a commutative ring.

CAUUCHY REAL NUMBERS

Finally, we discuss the Cauchy reals. These are obtained as equivalence classes of
Cauchy sequences of rationals and, in classical set theory, the resulting field can
be shown to be isomorphic to Ed. As we shall see in Chapter IV this is not

necessarily the case in IZ.

A Cauchy sequence is a function / g QN such that

Vmn g N[0 < m < n | f ( m ) - f ( n ) |< 1/m ].

Let C be the set of Cauchy sequences. We define the relation E c C x C o f

"converging to the same limit" by

E = {< / , g > e C x C : V n > 0 [ | /(ft) - g(n) |< 3 / ft}.

It can then be shown that £ is an equivalence relation on C To establish the

transitivity of £, from < f , g >e E,< g ,h >e E and n > 0, one derives

51
I f ( n) ~ f (6n) \< l/ n, | f(6n) - g(6n) |< l/ 2 n , | g(6n) - h(6n) |< 1/ 2n,

| h(6n) - h(n) |< 1 / n, from which the desired inequality follows.

The set Ec of Cauchy real numbers is defined to be the quotient of C by the

equivalence relation £, i.e. the set of E-equivalence classes of C

The operations of addition and multiplication on Ec are introduced by first

defining them on C:

( f + g)(n) = f(2n) + g(2n),

and
(f g ) ( n ) = f ( k n ) + g (k n ),

where ^ “ I /W I + I 9(^) I can tJnen be checked that these operations are

compatible with £ and so induce operations on Mc which give the latter the
structure of a commutative ring.

We define a map i: C —» Ed by

i ( f ) = ({P : 3 n[p </(n) - 1 /n]}, {p : 3 n[p >/(n) +1 /n]J).

(To verify that i(f) is indeed a Dedekind real, we note that for given p < q in
Q we can choose n with 3/ n < q - p, and then we m ust have either p <f(n) - 1 / n
or q >f(n) + 1/n.) Moreover, it is not hard to show that

V/,gr e C[i (f) - i(g) < - x f , g > e E ] ,

so that i induces a monic map j : Ec -> Ed. This map is a ring homomorphism.

52
Now j is not necessarily an isom orphism 17. But it is one in the presence of the
countable axiom of choice AC( N )18. For from AC( N ) it can be deduced that every
Dedekind real is the limit of a Cauchy sequence of rationals, since for each
Dedekind real r and each n > 0 we can find a rational p with | r - p | <\/n. This
follows from the fact that, since l , and R r are inhabited, we can find rationals p,
q for which p < r <q; then the interval [p, q] can be divided into finitely many
subintervals of length < 1/ n.

So AC( N ) im p lies th a t Ec and Ed are isom orphic. Later w e shall co n stru c t

a m o d el of IZF in w hich Ec and Ed are not isomorphic.

17 See Ch. IV.

18 See Ch. IV.
53
 Chapter IV

Intuitionistic Zermelo-Fraenkel Set Theory and Frame-

Valued Models

INTUITIONISTIC ZERMELO-FRAENKEL SET THEORY IZF

Intuitionistic Zermelo-Fraenkel set theory IZF is obtained by adding to IZ the

axioms of collecction and e- induction

Collection Vu[Vx e u3y<$(x,y) —» e u 3 y e vy(x,y)\.

e-Induction19 \fx[\fy e xcp(i/) > cp(x)] -> Vxcp(x).

It is to be expected that the m any classically equivalent definitions of well-

ordering and ordinal become distinct within IZF. The definitions we give here
work reasonably well.

D efinition. A set x is transitive if \/yex y c; x; an ordinal is a transitive set of

transitive sets. The class of ordinals is denoted by ORD and we use letters
a, ß, y,.. as variables ranging over it. A transitive subset of an ordinal is called a
subordinal. An ordinal a is simple if V ßEayea(ß e y v ß = y v y e ß ) .

Thus, for example, the ordinals 1, 2, 3, ... as well as the first infinite ordinal co to
be defined below, are all simple. Every subordinal (hence every element) of a
simple ordinal is simple. But, in contrast with classical set theory,

In classical set theory the e-induction scheme is equivalent to the a x i o m o f r e g a l a r i t y , which asserts
that each nonempty set u has a member x which is e - m i n i m a l , that is, for which x n u = 0 . It is easy
to see that this implies LEM: an e-minimal element of the set {0 | tp) u {1} is either 0 or 1; if it is 0, (p
must hold, and if it is 1, (p must fail; thus if foundation held we would get cp v —tip.

54
intuitionistically not every ordinal can be simple, because the simplicity of the
ordinal {0, {0 | cp}} implies cp v -Kp.

We next state the central properties of ORD.

D efinition. The successor a+of an ordinal a is a u ( a ) ; the supremum of a set A

of ordinals is UA. The usual order relations are introduced on ORD:

a < ß <—
>a gß a < ß <e> a c: ß.

It is now easily shown that successors and suprem a of ordinals are again
ordinals and that

a < ß <-» a + < ß UA < ß <-» V a e A a < ß < y -> a < y.

But straightforward argum ents show that any of the following assertions (for
arbitrary ordinals a, ß, y) implies LEM:

( i ) a < ß v a ~ ß v ß < a,
(ii) a < ß v ß < a,
(iii) a < ß - > a < ß v a = ß,
(iv) a < ß -> a +< ß v a = ß,
(v) a < ß < y -> a < y.

Notice that as a special case of e-induction we have the Principle of Induction on

Ordinals, namely,

VafVß < aq>(ß) - » cp(a)] -> Va(p(a).

55
An ordinal a is a successor if 3ß a = ß+, a weak limit ifV ßea 3 yea ß e y, and a
strong limit if V ßea ß+e a .

Note that both the following assertions imply LEM: (i) every ordinal is zero, a
successor, or a weak limit, (ii) all weak limits are strong limits. For (i) this
follows from the observation that, for any formula cp, if the specified disjunction
applies to the ordinal (0 | cp}, then cp v -—
.cp. As for assertion (ii), define

l (p = {01cp}, 2(p - {0, l (ph ß = {0, 1„ 2*, 2(p+. 2 / ;

Then ß is a weak limit, but a strong one only if cp v —,cp.

As in classical set theory, in IZF a connection can be established between the

class of ordinals and certain natural notions of well-founded or well-ordered
structure. Thus a well-founded relation on a class A is a binary relation -< on A
which is inductive, that is, for each a e A, the class {x e A : x <a} is a set and, for
every class X such that X c A we have

Vx e A \\fy < x(y e X) -» x e X] -» X = A.

A well-founded relation has no infinite descending sequences and so is

irreflexive. Note that the e- induction axiom asserts that e is a well-founded
relation on V.

The usual proof in classical ZF to justify definitions by recursion on a well-

founded relation does not use LEM, and so is valid in IZF. Thus, given a well-
founded relation -< on a class A and a function F : A x V —> V , it is provable in
IZF that there exists a unique function G: A V such that, for any u e A we
have

G(u) = F{(u,G | {x e A : x < u}).

Accordingly we can make the following

D efinition. If K is a well-founded relation on a class A, the associated rank

function r k _ : A -» ORD is the (unique) function such that for each x e A,

56
 r k . (x ) = } J r k (yf .
y-x

When < is e restricted to an ordinal, it is easy to see that the associated rank
function is the identity.

To obtain a characterization of the order-types represented by ordinals we make

the following

Definition. A binary relation -< on a set A is transitive if

Vx g A\fy g A\fz g A[x < y / \ y <z~~> x < z \ .

and extensional if

Vx g A \ /y g A[\fz[z < x <-> z ~< y —>x = * /]]•

A well-ordering is a transitive, extensional well-founded relation.

It is easily shown that the well-orderings are exactly those relations isomorphic
to g restricted to some ordinal. For it follows immediately from the axioms of
g -induction and extensionality that the e -relation well-orders every ordinal.

And conversely, it is easy to prove by induction that the rank assigning function
on any well-ordering is an isomorphism.

D efinition. The rank function p: V -> ORD is defined by recursion on the well-
founded relation e by the equation

p(x) = IJp(iy) .
yex

The cumulative hierarchy Va for a e ORD is defined recursively by

Kt = U pV
ßcu

57
Proposition 1.

(i) V x p (x )e O R D
(ii) x e y -> p(x) < p(y).
(Hi) a < p ^ c V p.
(iv) x ^ y < = V n - > x e V a .
(v) Va p(a) = a.
(vi) Vx x gV .

Proof. The proofs are straightforward inductions. To illustrate, we prove (vi).

Suppose V i / e x . i / e i / (y), .Then x c ( J Vp(j/), c VpM. Hence x e P V ^, = V ^ , M

yex

Notice that it follows from (vi) that V - U

aeORD

FRAME-VALUED MODELS OF IZF DEVELOPED IN IZF

Throughout this section, we argue in IZF.

Let H be a frame20 with top element t and bottom element ±. An H-va lu e d

s tr u c tu r e is a triple @ = (S, ([• = * | 1* g •!>, where S is a class and
1* = *1 I# e *1 are maps S x S -> H satisfying the conditions

([U = W|=T
lu = v l = lv = u l
l u = V II A \\v = w $ < l u = w §
I u = u i A IIU e w t< lu G w 11
lv = w} A lu G v} < |u G w\\.
for u, v, w g S.

20 For the definition of frame, see Appendix 1.

58
Let C (S) be the language obtained from C by adding a name for each element of
S. For convenience we identify each element of S with its name in C (S) and use
the same symbol for both. The maps [• = *J, |* e •) can be extended to a map
1*1 defined on the class of all C (S)-sentences recursively bv:

l a a if - H al a |x i l a v x j = |[a l v fx f l a => r 1- l a l = > |x l M - [fair

|3x<p(x)1 = V lq M l JVx<p(x)J = Al<p(u)I

ue-S’ ue.S

For each sentence a, l a | e H is called the truth value of a in 3 ; a is true , or holds

in 3 t written 3 b a, if | a l = r and it is false in 3 if Ia ||= _l. We also write
$ £ a for Jojl t; this means that while a is not necessarily false in 3 , it
nevertheless fails to be true in 3 . In this event we say that a is not affirmed in 3 .
It is not hard to show that all the axioms of first-order intuitionistic logic with
equality hold in 3 , and all its rules of inference are, in the evident sense, valid in
3 . 3 is a (frame-valued) model of a set 7 of C (S)-sentences if each member of T is
true in 3 . If 3 is a model of 7 , and a is an intuitionistic consequence of 7, then
3 1= a.

Given a frame H, we set about constructing, within IZF, an H-valued structure

called the universe of H-sets or the H-extension of the universe of sets21, which
can be proved, in IZF, to be itself a frame-valued model of IZF. It follows that
any sentence a which is true in some m ust be consistent with IZF.

The class r H) of H- sets is defined as follows. First, we define by recursion the

sets VJH) for each ordinal a:

Va{H) = [ x : F\in(x) a ran(x) c H a < a[dom(x) c V}H)} .

Then we define

21 w hen H is the frame C{X) of open sets in a topological space X„ WO is called a spatiial
extension.
59
 V {H) ^ { x : 3 a [ x e V j H)}.

It is easily seen that an H- set is precisely an H-valued function whose domain

is a set ofH - sets. We write £ ( H) for the language £ (WO).

The basic principle for establishing facts about H- sets is the

Induction Principle for WO. For any formula cp(x), if

Vxe W0[Vy<Edom(x) cp(y) -> cp(x)],

then

VxeVW (p(x).

This is easily proved by induction on rank.

We now proceed to turn W ) into an H-valued structure. This is done by

defining f u = v fH
) and 1u e v fH) by simultaneous recursion as follows:

i u e v t H)= V [v(y)^u = y t H)]

ye dom( v)

l u = v j H] = A [u(x)=>|xei/l<H>]A ye. dom(y)

xedom(u)
A [vfe/)=> |i/e u f" '] .

It can now be shown by e-induction that = < WO, [* e = *fH)> is

an H-valued structure. This structure is called the universe ofH - sets: a structure
of the form is called a frame-valued universe. We assume that i*fH) has been
extended to the class of all £ (*0 - sentences as above: we shall usually omit the
superscript W

In particular we have

I3x(p(x)I= V ffcp(u)I ifVx<p(x)J= A lcp(u)i

ueV{H) ue V'H]

Note that we can always find an ordinal a for which

60
 (*) f3x<p(x)S = V flkp(u)lj fVx(p(x)]l= A [[<p(u)|!
u c V aiH) u c-iV jH)

For let A = {|cp(u)l: u e V (H)}. Then A cz H, and since H is a set, by Separation so

is A. We then have

Vx e A3ß3u e V ^ . x - §cp(u)j].

So, using Collection there, is a set l i e ORD such that

Vx g A3ß g U3u g V ^ . x - |(p(u)]|

If we let a = [J u , then
Vx g ABu g Vt[H).x = I'(p(u)|]

so that A = {||cp(u)|: u g V^H) } , and (*) follows.

An argum ent of this general sort will be called a Collection argument; we shall
tacitly employ a num ber of such argum ents in the sequel.

Of use in calculating truth values in are the rules:

II3x g tup(x)| - v
xrdom(w)
|cp(x)l |Vx g ucp(xJS — A
x^dotnfu)
I<p(*)l •

u(x) < I]x g u\\ for x g dom(w)

Note that, given an H-set set u, 3x(x e u) does not necessarily imply that
there is an H-set set v for which N= v g u. An H-set u satisfying this latter
condition is called inhabited, and an H-valued set v satisfying t= v g u is
called a definite element of u.

There is a natural map A: V -> VW defined by e-recursion as follows:

x A = {< y Aj>: y e x }.

61
It is then easily shown that, for x e V, u g VW,

I u e x A5 = Vlu = y Ai.
yex

It follows that

a e b —> | a A g foA| = t

and

[|3 x e a A (p(x)l = Vl<p(xA)I J V x e a A <p(x)f= Al<p(aA)i •

x^a x^a

It follows immediately from these that the H-valued set 0 A represents the
natural num ber 0 (i.e. 0 ) in %S(H\

Moreover, the H-valued set N A represents the set of natural numbers in

For let Ind(w) be the formula 0 gua Vx g u (x ' eu). It is then easily checked
that $MH) h Ind( N A ). Also, we have, for each n g N

(*) I7nd(u)|l<lnAG u |

This is proved by induction on n. It is clearly satisfied by 0, and obviously

| Ind(u) a n A g u j < f n +A g u |

Therefore

I Jnd(u)I < ln A g u| -> [f/nd(u)l < [ n +A g u ||].

and (*) follows by induction. Hence

lln d ( u )l < A I^ A Gu l = 1Na c u |

neM

So N Ais the least inductive set in

62
Note that it follows in particular that the Axiom of Infinity holds in W A

A useful fact is the

Unique Existence Principle for $MH) . I f ^ H) f 3!xcp(x), then f (p(w) for

some v e W'O.

Proof. This is proved by translating to W 1) the proof in IZF that, if 3! xcp(x),

then the set u defined by u = { y : 3x[cp(x) a j / g x ]} satisfies cp(r/). Thus,
assuming W(H) N 3!x(p(x), we have V M * )l Using a
t = pjeq>(x))l =
xcV{lh
Collection argum ent we obtain an ordinal a for which r= V |fcp(x)|. If we

now define u e VW by dom |u) - V([H), u(y) = |3x[cp(x) A i / e x ] |, then |cp(u)[| ~ t .

■

It is not in general true that if 1= 3x(p(x), then ®(H) f cp(z/) for some
u e VW. As we show below, certain conditions on H will ensure that this holds.

Given subsets {at : i e /J c H , {ut : i £ / } c V {H\ we define the mixture

^ a t • u, to be the H-set u defined by dom(u) = [Jd o m (u ;) and, for x e dom(w),
i el lei

u(x) = V « . a [[x e u, | If / = {0,1}, we write a0 • u0 + for ^ a f • u, .

Two elements of a frame are disjoint if their meet is ± . A subset of a frame

consisting of m utually disjoint elements is called an antichain. If an antichain is
presented as an indexed set {at: i e /}, we shall always assume that
at a av = jl whenever i * i

Now we can prove the

Mixing Lemma. Let {ut :i e I j c V(H) and suppose that {at : i e 1} is an antichain in
H. Then, writing u for ^ a . • u x, we have a, < |Ju = u{tfor all ie I.

63
Proof. We have, for given z e /, Ju = uxf) = a a b , where

a= [u(x) => lx e u,J], £>= [u ,{ x )= > lx e u % .

Xfcdom(u) xedom (u,)

If x e dom (u), then, since {a. : i e 1} is an antichain,

a{ a u(x) = V « , A CLj A f X G Uf J < A f X G U- J < |jx G Ut

jeJ

Hence a, < [u(x) => l x gw J for all x g dom(w), so that a. < a . On the other
hand, if x g dom(i//), then

a,, a u.(x) < a. a [ x e u. | < u(x) < ( x g u . S,

so that a- < [u-(x) => | x g u - J], whence «/ < b. Hence a. < a a b and the result
follows. ■

An element a of a frame H is said to be complemented if a v a * = T . H is totally

disconnected if every element of H s the join of a set of complemented, elements.
Equivalently, H is totally disconnected if, for any elements a, b, a <b iff, for all
complemented elements c, c< a implies c < b. Notice that every Boolean algebra
is totally disconnected, and, for a topological space X, the frame (j{X) of open
sets in X is totally disconnected iff X s totally disconnected in the topological
sense of having a base of d open sets.

We shall need the following fact later on:

Definite Element Lemma Suppose that H is totally disconnected, let u be an

inhabited H-valued set, and write Ufor the class of definite elements ofu. Then, for any
formula (p(j),
(i) [[Vx e mp(x)l = A I<PM1
weU

(ii) the following are equivalent:

(a) 1= Vx g ucp(x)
(b) SIKH) 1= cp(tu) for all u g U.

64
Proof, (i). We first prove the following:

(*) For any H-set v and any complemented a e H such that a < \\v eu% there is a
definite element we U such that a < \v =

Suppose that v and a satisfy the hypothesis. Choose a definite element z of u and
let w = a - v + a* -z. Then a < \\v - w \ and a* <\\z = w j follow from the
Mixing Lemma. Moreover, we have

ff tu e u )l > \\v = w a v e u fj v \\z = w a z e u f| >a v a* -T ,

and so w is a definite element of u. This proves (*)

To prove (i) it suffices to show that

A lcp M I< fV x eu < p (x )l

i.e., A1<pM I < A [lb Gul=>||cp(u)S].

w*=U v e V{H)

Thus we m ust show that, for each VW-set v,

A \M w n < [ \\v e u \\^ M v m

that is,
A iq>Mi a [lv e u j < lcp(u)l].
w-:U

Snce H s totally disconnected, to prove this it is enough to show that, for any
complemented element a e H,

(**) a < A !<p(tu)|| a \\v e u j -» a < f[cp(u)!).

W*U

So assume the antecedent of (**). By (*), there is w e 11 such that a <§v = w%

Hence a < ||u = w$ a |cp(u;)|| < |[(p(u)| This proves (**), and (i) follows.

65
(ii). Obviously (a) implies (b). Conversely, if (b) holds, then y\ lq)(u;)S =T and it
we l l

now follows from (i) that ('Vx e uq>(u) 1 =T. ■

A core for an H- set u is a set C c V(H) such that (i) each member of C is a definite
element of u; (ii) for any definite element y of u there is x e C such
thatflx = t/| = T.

We show that each H- set u has a core. For each x e let

ax = {(z, u(z) a [z = x| : xe dom(u)}.

By a Collection argum ent there is a set W c= VW such that, for any xe VW, there
is y g tV for which ax - %. It is easily shown that the set
C = {x gw : [fx g u l =t } is a core for u.

By abuse of notation, we shall write {x g V (H) : | x g ul =t } to denote a core for a

given H-set u

Now in IZF it can be shown that is an H-valued model of IZF. It was shown
above that the Axiom of Infinity holds in We further verify the Axioms of
Separation, Collection and g -Induction in , with brief comments on the
verification of (some of) the remaining axioms.

To begin with, we note that, given H-sets u, v the H-set {u, vYH) =
{< u, t >, < v,T>} is easilv shown to validate the Pairing Axiom in V(H).

In this connection = {//, uYH may be identified as the singleton of u in

and <u, v >(H) = {{uYw , {u, vY^Y1^ as the ordered pairof u, v in

We recall that the Axiom of Separation is the scheme

\/lBv\/x[x G V <-» X G UA Cp(x)].

66
To see that each instance holds in $$(*0, let u e V(H), define v e VW by dom(y) =
dom(//j and, for x e dom(z;), i?(x) = u(x) a |cp(x)i|. Then we have
[| Vx[x GV <r>X GU A Cp(jC)l - fVx G v[x GU A (p(x)]l A [Vx Gu[(p(x) -> XGV\\.

Now

JVx G Z;[x GU A (p(x)]]f - A

xedom(y)
[[i*(x) A flcp(x)Jj] ==> [x G U A fcp(x) |j]] —T .

Similarly
I Vx G u[(p(x) -> X G vl =T

and so Separation holds in

As for Collection, we recall that this is

Vu[Vx g u3i/cp(x, zy) - > 3iA/x g u3 i/ g zxp(x?y)\.

To verify this in observe that

|V x g u3y(p(x,y)l = A [«M => Pz/cp(x, zy)J]

x&dom(tt)

For each x g dom(u), there is an ordinal ß for which

|3i/(p(x, i/)| = V |cp(x,iy)| . So by a Collection argum ent there is an ordinal a
yeV0{H)

such that, for all x g dom(zz), we have ff3zycp(x, t/)U < V fcp(x, z/)Jj . If we now
ycVa'w>
define z; g VW by v = x Jt } , then

|V x g u3zycp(x?y)]j = A [u(x) => |3y(p(x,zy)l] < A [u(x) => [pzy g ucp(x,z/)[|]

xpdom ju) xcdom(u)

= IIVx e u3y e u(p(x,y)j].

The truth of Collection in follows.

67
The e- induction axiom is, we recall, Vx[Vy e xcp(y) -> <p(x)j —> Vx<p(x). To
verify the truth of this in SJW, first put

a = |[Vx[Vi/ e x<p{y) -> <p(x)] 1.

It now suffices to show that, for any x e V W , a < [cp(x)|. To do this we use the
induction principle for V(H). Assume for y e dom(.t) that a < H<p(y)| Then

a< A tf<p(i/)1^ A Wy)=>l9(y)l] = lfViyex(p(i/)i.

yedom(x) t /e d om (x )

But a < [fVzy g X(p(y)J => |cp(x)]|, so that

a < [IVy € x(p(y)l] => |<p(x)l] a I \fy e x(?(y)J < !cp(x)H,

as required.

To establish the truth of the Axiom of Union in $MH), given u e VW, define
v g VW by dom(z>) = (J dom(i/) and, for x g dom(z;), v(x) = §3y g u .x g y]\.
yedom(u)

It is then easily shown that in v is the union of u.

For the Power Set Axiom it can be verified that, in £$(**), the power set of a set
u g VW is given by the set v g VW defined by dom(i?) = Hdom(") and, for
x g dom(u), v(x) = [[x g u J When w is of the form a , v may be taken to be the

function on H dom{a) with constant value t. We write P(H)(w) for v.

Now it is readily shown that LEM holds in £$(**) if and only if H is a Boolean
algebra. Since, as we have se en , in IZ the Axiom of Choice implies LEM, it hold
in in fact, it cannot hold in 3$(H) unless H is a Boolean algebra. This is far
from being the case for Zorn's lemma, however, despite the fact that it is
classically equivalent to AC. Indeed, we will show that, in IZF, Zorn's Lemma
implies its truth in any 3S(f/). We shall take Zorn's lemma in the form: any inhabited
partially ordered set in which every chain has a supremum also has a maximal element.

68
Thus suppose X, < x e V(H) satisfy

W*) < X, <x > is an inhabited partially ordered set in which every chain has a
supremum.

Let X' = {x e V(H) : [jx e X[j ^ t } be a core for X and define the relation < \ on X
by X <Y>x ' <-> | x <x X l =T. It is then easily verified that <X', < x> is an
inhabited partially ordered set in which every chain has a suprem um. So, by
Zorn's lemma, X' has a maximal element c. We claim that

(1) [c is a maximal element ofX ] = T.

To prove (1) we take any a e V(H) and define Z e V(H) by dom(Z) = dom(X) and

Z(x) = [x = a a j g X a c <x x] v |x = c].

for x e dom(X). It is then readily verified that V(H) 1= Z is a chain in X; and so ,

using the unique existence principle for V<H)' there is v e X' for which

( 2) V(H) z; zs the supremum ofZ .

Since [c e Z] = T it follows that [c <x v\ = T , whence c < x ' V , so that v = c by the

maximality of c. This and (2) now yield \a e Z -> a <x c] = T ; and clearly

|a e V -> c < x x] = T . Therefore

(3) [fl g Z —> a —c| ~ T .

It is easily verified that

(4) [a g X a c < x a\ < [a e Z J.

69
(3) and (4) yield |a e X a c<x a] < [a = c\; since this holds for arbitrary a e V(H>,
(1) follows.

From the fact that Zorn's lemma holds in every %$(H) but AC does not we may
infer that, in IZF, the former does not imply the latter. In IZF Zorn's lemma is
thus very weak, indeed so weak as to be entirely compatible with intuitionistic
logic. For more on this see Bell [1997].

THE CONSISTENCY OF ZF AND ZFC RELATIVE TO IZF

We noted above that, w hen H is a (complete) Boolean algebra, LEM holds in

so that is a model of (classical) ZF. In IZF the simplest complete
Boolean algebra in IZF is not the two element Boolean algebra 2, since as we
have noted above it is complete if and only if WLEM holds. The simplest
complete Boolean algebra is in fact the frame fi* of regular22 elements of Q .
Accordingly in IZF 3S(Q*), is a model of ZF. It follows that, //IZF is consistent, so
is ZF. Since the consistency of ZF implies (as is well-known) the consistency of
ZFC, we conclude that, if IZF is consistent, so is ZFC.

We shall exploit this last fact in the following way. Suppose we w ant to show
that a certain sentence a of our set-theoretic language is relatively consistent
with IZF. We produce a certain frame H and show in ZF(C) that a holds in VRO.
The latter is accordingly a model of IZF + a, from which it follows that he
consistency of ZF(C) implies that of IZF + a. Since the consistency of IZF
implies that of ZF(C), we conclude that a is relatively consistent with IZF.

This idea will be used in the following sections.

22 See Appendix 1 for definition of regular element of a Heyting algebra and of the
Boolean algebra of regular elements. Regular elements of Q are precisely those elements co
satisfying — «oo = true <-» co = true.
70
 FRAME-VALUED MODELS OF IZF DEVELOPED IN ZFC

Henceforth we argue in ZFC.

The following additional basic facts concerning frame-valued models can be

proved in ZF23.

• a e b <-» « ( H) 1= a Ag A a = b <-> «(H) f= aA =b*

• if (p(xi, ..v xu) is a restricted formula24, then
a,i) <-> «(H) 1= <p(fliA, anA).

These facts have certain straightforward consequences which we shall employ

without explicit mention, for example:

• «(«) 1= <a, b>A = <nA, frA>

• %f A ->B a / a : /4a ->Ba A fA(aA) =f(a)A.
• «(H) 1= (Bj4)Ae ( B A)'4A

We shall strengthen the concept of a core for an H-valued set in the following
way. A strong core for an H-valued set u that is a set v c= VW such that
(i) \[x g ujj = t for all x g v; (ii) for any y e V(H) such that | y g u \ =t , there is a
unique x e v such th a t|x = yl=^ T . It is easy to show, assuming AC, that any H-
valued set u has a strong core. Starting with a core v for u , define the
equivalence relation —on v by x - y <-> \ x = y \\ =t , and let w be a set obtained
by selecting one member from each —-equivalence class. Then w is a strong core
for u. Clearly a strong core for an H-valued set is unique up to bijection in the
sense that there is a bijection between any pair of such strong cores.

We shall henceforth write {x g V{H) : {\x g u | =t } to denote a strong core for a

given H-valued set u.

23 For proofs (in the Boolean-valued case) we refer the reader to Chapters 1 and 8 of Bell
[2011].
24 A formula is restricted if each of its quantifiers occurs in the form V a gi/ or Eagi / (i.e.,
Vx( a g if -» . . . . ) or 3a( a g y a . . . . ) or can be proved equivalent in IZF to such a formula.
71
A refinement of a subset A of H is a subset B of H such that, for any b e B, there
is a e A such that b < a. If B is an antichain, it is called a disjoint refinement ofA.
H s called refinable if every subset of H has a disjoint refinement with the same
join.

We say that satisfies the existence principle if for any formula cp(x) there is an
H-set u for which px<p(x)l = f<p(u)|

We can now prove the

Refinable Existence Lemma. IfH is refinable, satisfies the existence principle.

Proof. Suppose that H is refinable. Then for any formula cp(x), by a Collection
argum ent there is a subset A of Vdh for which

13x(p(x)!= V l<p(x)l = Vlcp(x)|

XG\*H) X^A

Since H is refinable, {[[cp(x)l: x e A) has a disjoint refinement {a- : i e 1} with the

same join, i.e. |3x(p(x)| Using the Axiom of Choice, select for each i e l an
element x j e A for which a. < l(p(xz) | Now define u to be the mixture ^ ax .
xe/

Then for each xx we have a, < lu = x- J| a |icp(xt )l < I(p(u)|, so that

J3xcp(x)l = V I<p(x)l = V a ; < I(p(u)|.

x^A iel

Since clearly

|cp(u)S<I3x(p(x)l

we are done. ■

Remark. The converse to the Refinable Existence Lemma also holds. For
suppose $$(**) satisfies the existence principle. Given A c H define the H-set v by
72
dom(^) = {a :a g A} and v(a) = a for CL G A . Then there is an H-set u for
which ||3x(x g v)l = j[u g v $. In that case

\/A = p x ( x g v)l = f u g v1= V a Alu =a S /

asA

so that {a a f u -a \} :a g A} is a disjoint refinement of A with the same join as

the latter.

We shall need the following propositions.

Proposition 1. The following assertions are equivalent:

(i) There is f g V(H) for which %$(H) 1=f is a surjection from a subset of.AA
onto BA.
(ü) There is a subset {uah : a g A ,b g B} <^H such that, for each a g A,
[uab: b g B] is an antichain, and for each b g B, V Uab = T
aeA

Proof, (i) -> (ii). Assuming (i) let f e V(W be such that

%$(H) |= F u n (/) a dom i f ) c: A A a ra n ( f ) = B A.

Define uab - 1/(&A) = bA[] • It is then easily verified that the unb satisfy (ii).

(ii) - * (i). Suppose the uab satisfy the conditions of (ii). Define / g V(H) by dom ( / )
= {< a,b > A : a g A ,b g B} and /(< a ,b > A) = uah. Then W H) f dom(/) c AA, the
first condition of (ii) gives Fun (f) and the second condition
1= ran(/) - BA. ■

Call a set Y a subquotient of a set X if there is a surjection from a subset of X onto

Y. Proposition 1 then has the immediate

Corollary. IfH satisfies condition (ii) of Prop. 1, then

W ll) \^BA is a subquotient of A A. ■

73
An element a of a frame H is connected if for any disjoint b, c e H , a = b v c
implies b = a or c = a. H is said to be connected if its top element T is connected.
Equivalently, H is connected if whenever T is the join of an antichain A, then
T e A . H is said to be locally connected if each of its elements is the join of
connected elements; equivalently, if, for any elements a, b oi H, a <b iff, for all
connected elements c, c <a implies c < b.

If X is a topological space, connectedness of an open subset U of X corresponds

precisely to connectedness of LI as an element of the frame ^(X), and
connectedness (resp. local connectedness) of X to connectedness (resp. local
connectedness) of 6{X).

Proposition 2. The following are equivalent:

(i) H is connected;
(ii) for any set u, and v e V(H\ if t= v e u A, then there is x e v such
that 1= v = x A .

Proof (i) -» (ii). Assume (i) and suppose that 1= v e u A. Then

t = \[v e u AS= Vl!^ = * A! . But since (|u = x AJ : x e u] is an antichain and H is
xeu

connected it follows that \\v = x A| =t for some x e u, which gives (ii).

(ii) -> (ii). Assume (ii) and let a, b be disjoint elements of H such that a v b - T.

Define the element v e V by v = n.0A + b. 1A. Then 1= v e 2 A, so that, by

(ii), either a = fv = 0 Aj| =t or b = = 1A| =T . (i) follows. ■

Given sets /, /, let us call the frame H ]) distributive if, for any subset
jl-(J,

{atj :i g I, j g J} of H such that, for each i e 1, \atj : j e J ) is an antichain, we

have
AV a fJ = V A<v(I).
re/ j e J f e j ' re/

Proposition 3. The following arc equivalent:

(i) ( J A)/A = ( J ') A

74
 (ü) H is J) distributive.

Proof, (i) -> (ii). Assume (i), and let \ay :i e I , j e J } cz H be such that
{ay : j e J} is an antichain for each i e 1. Define h e V(H) by dom(/i) =
{< i.j > A : i g I , j e J} and h(< i.j > A) - a {J:. It is then easily verified that

Fun(h) a d o m ( h ) c / A a r a n ( h ) c J A.

Thus Ih(i) = j i = a i;/. and so

A V S- - ffdom(h) - / AI - I h g ( J Af J - Jh e ( J ') A1 - V I h = f AI
iel
= V A IN iA) = / ( i ) Al
/eJ; n=I
- v Aa if(iy
feJ* M
This is (ii).

(ii) —> (i). Assume (ii). To obtain (i) it suffices to show that

(* ) »M l= (JAf c : ( J I)A.

For h g let ^$ (iA, j A) e h l and atj = hi; a /\[ h y a hy =>±]. Note that
a;

a {. A a ir = ± for all i g I, j j' g J. Then we have

|h g J < IFun(h) a domtf) = I A a r a n ( / ) c J Al

A [^A h,= >yA = /A ,]A A vfi9

it./ i-I j^J
A IV hy A f tv A K U A = 7"
i,J j e J hJ'eJ
= A V a,j = V A a »7 (0 V Ah, V A W A) = / ( i ) Afl
A/ jgJ /W7ie/ f^j: i=/ •■no f e J 1
= V l f t = / AI = l f t e ( J f)Afl
/eJ'

This proves (*). ■

75
Let us say that H

• is completely x-distributive if it is ±-(l, f) distributive for all 1, /,

• is completely ±-2 - distributive if it is _l-(I, 2) distributive for all I„
• preserves exponentials if b ( J a)/ a = ( J 7) A for all /, /.

Then we have

Proposition 4. T7zefollowing are equvalent:

(i) H preserves exponentials

(ii) H is completely ^-distributive
(iii) H is completely 1-2 - distributive
(iv) H is locally connected.

Proof. The equivalence of (i) and (ii) follows immediately from Prop. 3. So it
suffices to prove the equivalence of (ii), (iii) and (iv).

(ii) -> (iii). Obvious.

(iii) -» (iv). To begin with, for elements b < a of H, call b complemented in a if there
is c < a such that b a c ~±, bw c ~ a. It is easy to check that, if such c exists, it is
unique; denote it by a - b. Now assume (iii), and let a e H. Write {a, :i e 1} for
the set of elements of H which are complemented in a. For each i e I let am= at,
Ui\ - a - ai. Then we have, for each g e 2l,

n a = A (a io v aii 1= V mo : iA«,
el
ig(i) V A a,mo­

lt follows that each A aig(l) is complemented in n (with complement

V A We claim that fc = / \ a ^ (J)is also connected. For suppose that

f * 2 l - <g) i e l ic-I

b - c v d with c a d =L and c * r. Then c is complemented in b and it follows

easily that c is complemented in a. Thus c ^ for some i0 e I and so
L^c = a,o < a ^ 1. Hence g(?„) = 0, so that b < a ^ „ o) = = c.

76
Accordingly b = c and it follows that b is connected. From (*) we conclude that
each ß g H is the join of connected elements, so that H is locally connected.

(iv) —> (ii). Suppose that H is locally connected. To show that H is completely jl-
distributive, it suffices to show that, for any subset {a(j :i e I, j e J} of H such
that, for each i e /, the set jay : j g J} is an antichain, and any connected
element c g H, we have

(*) c < A V aj c< V A aif{i

i(-I j e J f , J < is I

So assume the antecedent of this implication. Then for each i e /, c < V ai]f and,
jeJ

because c is connected, it follows that c < aXJ for some unique j e /. Define
g: 1 —>/ to be the function which assigns this j to each /. Then
c < A a w(i)< V and (*) follows. ■
is I f t J 1t el

A FRAME-VALUED MODEL IN WHICH M IS SUBCOUNT ABLE

We now set about constructing (in ZF) a frame-valued model of IZF in which
is subcountable.

Given two sets A and ß, let P = P(A, B) be the set of finite partial functions from
A to ß, partially ordered by 3 . We shall write p, q for elements of P. Let C be the
coverage25 on P defined by

S g C(p) < - > 3 b e B S = {q g P : p c q a ö g ran(q)].

The sieve S on the l.h.s of this equivalence is called the cover of p determined by b.
Note that every S g C(/?) is nonempty.

Lemma 1. For S e C (p), qlyq2 e S, there are q3,q4,q5 ^ S such that

<?3 e <?! ><?4 £ <?2 q, c q 5.

25 For the definition of coverage see Appendix 1

77
Proof. Suppose S is associated with b e B. The without loss of generality we can
assume that b £ ran(p), for otherwise we can take q3 = q4 = q 5 = p . Choose
a {, a 2 e A so that q(a,) = q (a 2 ) = b and define

q3 = p v { ( a 1,b)}> q 4 = p v { ( a 2,b)}.

If Oj - a 2, take q 5 = q3 = q4.If a l ± a 2, take q5 = p u { (a ,, b), a 2,b )}. ■

Lemma 2. If {U} : j e J } i s an antichain o f C-closed sieves in P, then (J U} is C-

jeJ

closed and is thus the join o f {U i : j e J\ in the fram e H c o f C-closed sieves26 in P.

Proof. Let S e C(p) and suppose that S c \ J U j . We claim that S c U, for some
jej

(unique) j e /. Given qx e S , fix j so that q { e Ur For each q2 e S , there are, by

Lemma 1, q3,q4, q5 e S such that q3 c q ,,q 4 c q , and q3 u q 4 c q 5,
Sinceq3 e S there is j' for which q3 e l f .; since Lf;,is a sieve, it follows that
qt e Ur . Hence j = j ’ and so q3 e U j . Hence q5 e Uj and from this, by an
argument similar to that establishing q3 e Uy , it follows that q4 e U] . We
conclude that q 2 e U j . Since q2was an arbitrary member of S, it follows that S
c U, .Since U, is C-closed and S e C(p), we infer that p e U, and the Lemma is
proved. ■

Remark. It follows from Lemma 2 that H e is connected. For if the top element P
of Hcs is the join of an antichain {lh: i e /} of elements of H e , then by Lemma 2 ,
[J t/( = P. Hence there is i e l for which 0 e.U t ; but then U- = P.
re/

Proposition 5. He is completely i-distributive arid so preserves exponentials.

-6 For the relevant definitions see Appendix 1.

78
Proof. Let [UlJ : i e I , j e J} c; H c be such that {L^ : j e J} is an antichain for
each i e /. By Lemma 2, it suffices to show that

o n u ^ u r w , -
i el j e J f e d 1i e /

If p e Pj [ J Lfy , then for each i e 1 there is a unique j e } such that p eU . We

t el j e J

define h e J 1by taking //(/) to be this unique j. Then p e f ^ U ih(l),w hence

teJ
p e U f l^ /( 0 0 follows. ■
f e J 1 ie l

Remark. It follows from Props 4 and 5 that He is locally connected.

Proposition 6 %${Hc) N=BAis a subquotient o fA A.

Proof. By the Corollary to Prop. 1, it suffices to show that there is a subset

{Uab : a e A ,b e B} c; H c such that Uab n Uab, = 0 for ^ b' and V Uab = t for
at A

beB.

Define Uab = { p : (a,b> e p}. Then Unb is C-closed. For suppose the cover S of p
associated with V e B is included in Unb . We need to show that p e LU . If b' e
ran(p), then p e S c UW If fr' £ ran(p), choose a ' £ dom (p) u {a} and define
p ' = p u K a ' j b } ) . Then p' e S, so that p' e Uab, from which it easily follows that
p e Urf,.

Clearly t /^ n Uab, = 0 for b * b\

To show finally that V Uah = r , for b e B, it suffices to show that P s the onlv
aeA
C-closed sieve containing ( J Uab, and for this it suffices to show that for each
aeA

p e P there is S e C(p) for which S c; [J Uab . Given p e P, let S be the cover of p

aeA

79
determined by b. If q e S. then q(a) = b for some ag A, whence q e U ^ i t
a eA

follows that S c ( J Uab . ■

aeA

Now in the preceding take A = W and B = . Then by Propositions 5 and 6

» (Hc) 1=(NAf A= (Nn) A /s a subquotient of N ,

and so

®(ffc) t= NNis subcountable.

The relative consistency with IZF of the subcountability of N n follows.

Remark. Suppose that H is a frame containing a triply-indexed subset

{a,nnp: m, n, p e N } satisfying the conditions:

W a mnp A =J- f 0 r P ^ Q
(2) AAWmnp = T
m n p

(3> A VAa- / (n, = t

febV' rn n

Then H cannot be jl-( N , N ) distributive.

To see this, let n++ n * be any bijection of N with itself lacking fixed points
and define b mn e H by b mn = a tnmn*. Then w e have

hinn a brun — ± for p

and

V b™ = V a-" = V % =T/
n n p

so that
A V b™. = T -

80
On the other hand

V Ab'mf(m)

To prove this, notice that it follows from (3) that \fa mnf(m) = T , whence
m

m m m m m

It follows that H is not jl-( N , N ) distributive.

This argum ent has an analogue in . Define cp e V ^ by

dom(cp) = {< m ,n ,p > A : m ,n ,p e N}

and
cp{ < m , n , P > A) = a

Then conditions (1) and (2) abve imply that

p : ^ AxNA~>W A.

and t 9(mA,

Now, in let y : MA —>(Na)na be defined so that \\i(mA)(nA) ^ <j)(mA,n A). It

then follows from (3) that

(*) %$(H) t=y is a surjection of T4A onto (NT) A.

But the usual diagonal argument, carried out in H*, shows that

%${H) ^there is no surjection of N Aonto(NA)u*

and hence, using (*),

81
 » (H)N ( N H) A * ( N a )n \

It now follows from Proposition 3 that H is not _l-( N , N ) distributive.

THE AXIOM OF CHOICE IN FRAME-VALUED EXTENSIONS

If I is a set, the Axiom of Choice for I is the assertion:

AC (I) for any formula cp and any set A

\fi g I3x g A c p ( i , x ) —> 3f g A*\fi g I c p ( z , / ( i ) ) .

AC( N ) is known as the Countable Axiom of Choice.

Proposition 7. If H is refinable, then $S(H) AC( J A ) for even/ set 1. In particular,

%S(H) AC( N A), so that the Countable Axiom of Choice holds .

Proof. We have

I V i g J a 3 x g Ap(;,x)f = A V A(x)Alcp(iA, x ) l
i&I x e d o n (A )

Since H is refinable, we may use the Axiom of Choice (in ZFC) to select, for each
i g I, a disjoint refinement ^ { b if : j g J t} of {A(x) a [f(p(iA,x;)l: x g dom(A)}
with the same join as the latter. Again using the Axiom of Choice, select for each
i e I , j e J i an element cdom (A ) for which b- < A(x {j) a [[cp(za , x tj)j). If we
now define f g V [H) by dom (J) - {< iA, x xj >(H): i g IJ g J i} and

f{< i.Xy >(H]) = b , a tedious but straightforward calculation shows that

A V A(x) a [fcp(iA,x)J = A V b. < I f : JA -> A a V i e / A <p(i,/(i))l

i el x e d o n (A ) l ei j e J t

82
®(Hc) 1= AC( / ) follows immediately. ■

A frame H is countably generated if it has a countable subset S such that every

element of H is the join of elements of S. If this is the case, S is called a countable
set of generators for H.

Proposition 8. IfH is countably generated and totally disconnected, it is refinable.

Proof. Let {a, : i e J}be an arbitrary subset of a countably generated, totally

disconnected frame We first show that there is a countable subset I0 ci / such
that V a f - V a - .
1(eI IgIq

Let S be a countable set of generators for H, and for each i e / choose S- c S s o

that V S f Then T = is, as a subset of S, countable, and so can be
16/

presented as {tn : n e N}. Moreover V tn = VT - \ J at . For each n e N there is

nsN icJ
in g / such that tn < al . Then Jo = {in : n e N} is a countable subset of / and
V at = V •
ie/ ieJ0

Now since H is totally disconnected, each fl/ is the join of a set of complemented
elements, and by the argum ent above, this set may be taken to be countable.
For each i e l let {bul : n e FI} be a (countable) set of complemented elements
such that ax - V bm• Then

V a , - V a t “ V V b M.
i«- l id L it/ ., nd Ki

Let (cn : n e 1J] be an enumeration of the countable set {bm :i e I0,n e N}. Then
(cn : n g IT} is a refinement of {a- :i g 1} with the same join as the latter.

Now define dn e H , for each recursively by do = Co, d„+i = c«+i a (do v ... y d H)*

83
T hen {dn : n e N} is a n a n tic h a in , dn < cn for each n, a n d V drn - V
neh' neN

It follows th a t {dn : n g N} is a disjo in t refinem ent of {a. :i e l } w ith th e

sam e jo in a s th e latte r. The refinability of H follows. ■

REAL NUMBERS AND REAL FUNCTIONS IN SPATIAL EXTENSIONS

In , the set of rational num bers may be identified with the H-set Q A,
where Q is the usual set of rational numbers. Since « (H) is a model of IZ, we
can carry out within it the construction from the rationals of the set !Rd of
(Dedekind) real numbers as in Ch. 3. Let 1 H be a strong core for the resulting
H-valued set. The members of are naturally thought of as H-valued real numbers.
More generally, given h e H, an H-valued real number of degree h. is an element
r e 1/(H) for which h<\[r e E d |

Since each Dedekind real is a cut in the rationals, each H-valued real num ber r is
a pair < l , r > e V(H) for which

|< l ,r > is a cut in = r.

This condition translates into conditions on the truth values |p Ag l | |q A g r J

viz.,
0. Jl u r c Q a |= t

1. V l p AGLSAl qAG R l - T
PW-,
2. |pA e l] a [ pA e r ] = i

3. I p Ae L l = V l q A ei - i

4. I p A e R l = VllgA eR]l
C{<p

5. | p A g Lj v | q A g R 11—T for p < q.

Similar conditions may be written dow n for H-valued real numbers of degree h:
these are left to the reader.

84
It is easy to check that, writing R for the genuine" set of real numbers in V, we
"

have S?(H) I c Rd . In particular, for each r e V we have %${H) I- rA g A'.d .

Hence we may assume that r A e R H .

can be turned into an ordered ring by defining +, *, < as follows: for

r, s e mH,
r © s = unique element u o f nH such that [u = r + s i =t
r0s= unique element u o f such that {{u^r.s
r < s <-> f r < s j] =t

With these definitions l . H is called the ordered ring ofH-valued real numbers.

Now let X be a topological space. For brevity we shall write $$(X), VW , R x for
V(W), [R</( respectively. Members of i x will be called simply real
numbers over X. Thus a real num ber over X may be identified as an element r =
< L , r > of VW satisfying the following conditions:

Ox. Jl u r c Q1 =X
lx. U [ p Ae lJn [|qA e r 1= X

2X. |[pA e L.J1n | | p A e Rjj =0

3x. l p A e l| = | J l q Ae lJ
q>p

4x. ! p Ae R l = ( J l g Ae Rj
q<p

5x . | p A e l [| u I qA e R jj = X for p < q.

If U e ^X ) is an open subset of X, an X) - valued real num ber of degree U

will be called a real number over U. Conditions analogous to Ox - 5x above can be
formulated for real num bers over U.

We now prove

Proposition 6. The ordered ring R x of real numbers over X is isomorphic to the

ordered ring C(X, R) of continuous real-valued functions on X.

85
Proof. To obtain this isomorphism, start with a real num ber r = < l , r> over
X. For each t e X define

i-t - {p € Q : t g |[ p A G L j} R f = { p e Q : f e | p A e r ]|}

and rt = (L t , R t ) . Then rt e R and the m ap r*: t ^ rt is continuous and hence an

element of C(X, R).

To show that rt e R (i.e. rt is a Dedekind real), we check, for example, the

condition
V p ( p e L ( <-> 3q e L t .g > p ) .

Using condition 3x, we have for t e X

p g l £ <~>t g ! p A e Lj

<-> t g ( J § q A g L i <-> > p . t G |q A g l ]J <-» 3q > p. q g l£ <r+3qe i-tq > p .

q> p

The other conditions are checked similarly.

To show that r* is continuous, it suffices to show that the inverse image under
r* of each open interval (p, ->) in R is open in X. This follows from the
observation that, for t e X, we have

rt —(Lf, Rf) g (p, >) <-> p g L£ <->t g 1p a g l |

Accordingly the inverse image under r* of (p, ->) is |p A g l | , which is open in

X.

The function r* is said to be correlated with r.

Remark. For each " genuine'' real num ber r e V, (rA)* is the constant function

) b r e E , then r* is locally constant, that is,

on X with value r. In general, if 3$(H

86
each point of X gas a neighbourhood on which r* is constant. Note that, if X is
connected, then each locally constant function on X is constant.

Conversely, g iv e n / g C(X, E), define l /, R / g VW by

dom (L/ ) - dom(R/ ) - {pA : p g Q},

with l ,( P a ) = / ‘((p,->))> <V(Pa ) = / '((<-»P))-

We claim that

N < L /- R f > e !R d .

. We verify conditions 3x and 5x, leaving the rest to the reader. First note that

ilp ^ G L ^ /^ p ,-* ) ) .

Ad 3x-
IPA e L, | = / *((p, ->)) =
q>p
(q, ->)) = U/ ->» = U
q>p
e L/ 1

Ad 5x: For p < q, we have

I P As l, iM jq A s R , 1 = / " ' (p, —»)u / “1(<—, q) = / 1( Ä) = * •

We d e fin e / to be the unique element r of for which 1r = <Ly, Ry) (] = X. f

is called the real number over X correlated w ith /

I f/is locally constant, it is easy to check that %${H) N=/ e M A .

We next show that the maps r >->r* and are mutually inverse, i.e. (r*)* = r
and (/)*=/.

For the first assertion, we note that, for t g X,

87
 t e i p A GLrJ f > t G (r*)_1(p, ->)<-> r * (t) g (p,-~>)
O r f > p f > p E L f f > t G [ p A G Lj

It follows that l p A GLr*| = |fpA g l |, whence (|Lr* = L_jj - X . Similarly

§Rr* = R | - X , whence |(r*)# = rf = X , so that (r*)# = r .

For the second assertion, note that ( / #)* (t) =< (L.f )t,{f*f )t >. So if fit) =
<L, R>, then

P G(L ^ <-> t G|p A G L; S <-> t G/ _1( ( p , - > ) ) <-> / ( t ) > p p G L.

Accordingly (l^), = l . Similarly (R/ )f = r , whence (f)*(t) = f(t). Since this

holds for arbitrary t e X, it follows that (/*)*=/.

We claim finally that the map r »-►r* is an isomorphism of Mx with C(X, E). To

establish this it suffices to show that

(*) r~<s<->r*<s* (r®s)* = r * + s * (r O s)* = r * -s *

The first of these assertions is an immediate consequence of the fact that

{\r < s l = { t : r * (t) < s * (t)} . This latter is proved as follows: let r = < l, r >,
s — < l/, r '>. Then we have

£ G f r < s f <-> t G P p [ p G R A p G L ' J o t G | J l P A G R A P A G L_'J

P

<-> 3 p [f G [ p A G R l n | p A G l '1] <-> 3 p [p gr( a p g r £]

<-> rt < s t *-» r * (t) < s * (t).

The proofs of the remaining assertions in (*) are left to the reader. ■

The upshot of Proposition 6 is that real numbers over X can be regarded as real
num bers varying continuously over X.

For the record, we also note:

88
 If* = g * l= l n { t :f (t ) = g(t)}*

This follows from:

f * =g*l = li-f = L 9 J n | R / = R 0 1
= Inp)lpAe l 7 So IpA e Lg InInP)|pA € R/i » l p AeR3i
P P
= Inf|(/ '(p, - » J o g *(p, ->)) n Inf|(/ p) o g’' h p))
P P
= Inf) [(/ 1(p, ->) o g 1(P, ->)) n I/“1(4-, p) <=>g 1(<-, P))l
p
= I n ( f : Vp[(/(t) > p g(t) > p) A (f(t) <p<r> g(t) < p)\
= ln{t:f{t) = g{t)}. U

In a similar w a y ,, one shows that, for real num bers r, s over X, we have

Hr = si = In{t:r*(t) = s *(*)}.

and
jfr < s[] = In{t :r* (t)< s* (£)}.

These argum ents can easily be extended to establish, for any open set U in X, a
natural correspondence, with analogous properties, between real numbers over
U and continuous real-valued functions on U. Thus , writing C(U, R) for the set
of real-valued continuous functions on If, real numbers over LI correspond to
elements of C(U, R). Under this correspondence locally constant functions on LI

are associated with real numbers r over U for which l i e [[re AAf .

A real function over X is an element cp e VW such that

W x) 1= Fun(cp) a dom(cp) = Rd a ran(cp) c= iSd

27 Here In denotes the interior operation in a topological space.

89
Since real numbers over X correspond to elements of C(X, E), real fimctions over

X should be correlated with certain operators on C(X, E), that is, maps

<t>: C(X, E) —» C(X, E). We now set about identifying these operators.

An operator (£> on C(X, E) is said to be

• near- local if, for any / g e C(X, E),

In{f:/(0 =£(*)}£ {ftO(/)(f) = O(g)(0},

or equivalently, if

ln{t:f{t) =g(t)} c In {f: 0>(/)(f) = ®(g)(f) },

• local if, for any f g e C(X, E),

=g(01 £ U: ^(/XO = <%)(0 I-

Clearly any local operator is near-local. In general, the converse is false, but we
shall later show that, for metric spaces, every near-local operator is local.

We next show that real functions over X are correlated with near-local operators
on C(X, E).

Given a real function cp over X, define the operator O on C(X, E) by <£>(/) =

for / e C(X, E). <£>is the operator correlated with <p.

We claim that O is near-local. To establish this, note that

In{t: f(t) = flf(t)} H I / # = g* 1 c fcp(/#) - cp(g#)» = In{t: cp(/#) * (0

= cp(g#) * W } - I n ! t : 0 ( / ) ( t ) = ct>(g)(t)}.

90
Now suppose given a local operator A on C(X, R). We define the function

D: X x R -» R by the stipulation:

D(t, a) - b ifffor sonief e C(X, R),/(f) - a and A(f)(t) = b,

D is called the function o n X x l correlated with A. Clearly D satisfies

for arbitrary t g X , / g C(X, R).

Let us call a function F: X x R -> 1R localizable if, for some local operator A on
C(X, R), we have F (t, f(t)) = A(f)(t) for arbitrary f g X, / g C(X, R). Then local

operators on C(X M) are correlated with localizable functions on X x .

PROPERTIES OF THE SET OF REAL NUMBERS OVER R

We now focus attention on the case in which the space X is the space R of real
numbers.

To begin with, let i, a, v be the R-valued real num bers correlated with the

identity function, and the absolute value function, respectively, on R. Then we

have

• i < 0 v i —0 v i > 0. That is, the law of trichotomy for Rd is not

affirmed in ^ (%).

This follows from the observation that

91
 | i < 0 v i = 0 v t > 0 1 = (<-, 0) ln{0} u (0, ->) = (<—,0) u (0, ->) = IS - {0} * E.

Similarly, one shows that

• I * 1 11* i < 0 v 0 < i.

• «<*'* i < a f > ( i < a v i = a)

and

• $B( & l = a v i ^ a , so that, <]& Md is discrete.

• While ® (x)l= i * 0, %${%)& i is invertible . Thus M disci field.

To see this, first observe that that ||i - 0§ = ln{0} - 0 . Also

[[i is invertiblel = |3 x (ix = 1)J] = [ J (J In { t: tf(t) - 1} 2 0

Uer>( x) f eC{U: *)

Hence Qi is invertiblel * X.

i is not the limit of a Cauchy sequence ofrationals . It follows that

both Rc = Ed and AC( N ) are false in $S(").

We sketch a proof of this. It is required to show that

(*) p u e Q [u isa Cauchy sequence of rational numbers converging to i] |] = 0

Since K is locally connected, }N (Qa)!' a = (Qh) A, so (*) is equivalent to

for all s e Q*’, jfs Ais a Cauchy sequence of rational numbers converging to x 1 = 0

Accordingly it will enough to show that, for any U e $( R ), s e i f "

92
(**) U c= \[sAis a Cauchy sequence of rational numbers converging to il U= 0 .

Again, because R is locally connected, it suffices to prove (**) for connected LL

So assume U connected and that the amtecedent of (**) holds. Then for each n
U cz l s nA e C|Aj], and so s„A corresponds to a locally constant rational-valued
function in C(U, M ). Since U is connected, this latter function is constant on U;
let pn be that constant (rational) value. The sequence {pn : n e N) is then Cauchy;
so it converges to some unique t0 e R . We then have

U c; \\s Ais a Cauchy sequence of rational numbers converging to 1 j]

= | V q ^ Q A 3 r i e N A V m > r i( |s n A - i|< q ) l

q n tm n

c: In{f: <pn> converges to t)

- Injfo}
= 0 .

Accordingly Ü = 0 and (**) follows.

PROPERTIES OF THE SET OF REAL NUMBERS OVER BAIRE SPACE

If we endow N with the discrete topology, the set Nr"endowed with the
product topology will be written and called Baire space. * I is totally
disconnected28 and has a countable base consisting of clopen sets. We proceed to
establish various properties of the spatial extension ®('4). Our principal task will
be to show that, in Brouwer's Principle holds, that is, in $$('0, every function
from reals to reals is continuous.

We first note that, since is countably generated and totally disconnected, bv

Proposition 8 it is refinable, and hence, by Proposition 7, the Countable Axiom
of Choice holds in $$0. Thus, in every Dedekind real is the limit of a

28.1 can be shown to be homeomorphie to the space of irrational numbers.

93
Cauchy sequence of rationals, and it follows that, in l), the Cauchy reals and the
Dedekind reals coincide.

We now proceed to outline the strategy (due to Scott [1968]) for showing that, in
V(N) , every function from reals to reals is continuous.

Step 1. Show that, w hen X is a metric space, every near-local operator on

C(X, E) is local.

Step 2. Show that every localizable function on . T x R is continuous.

Step 3. Infer that real functions over , 4 are correlated with continuous localizable
functions on 4 x 1 .

Step 4. Show that each real function over .1 correlated with a continuous
localizable function on l x M is continuous in \

Step 5. Conclude that every real function over ,4 is continuous in 0.

The topological details for carrying out Steps 1 and 2 - which are somewhat
intricate and are omitted here - may be found in Scott [1968]. Step 3 then
follows accordingly.

Now for S tep 4. First we note that since CF is totally disconnected, the Definite
Elements Lemma applies to \ This should be borne in mind in the course of
that argum ent that follows.

Let (p be a real function over 4 ', correlated with the continuous* localizable
function F: • l x M R . To show that cp is continuous in ®(' % we need to show
that the sentence

( f ) VpVq > pV e > 03 d > OVx e [p, q]\/y e [ p ,q ] [ \x - y \< d - ) \ (p(x) - cp(p) |< e]

94
holds29 in SW \ (Here we have used p, q, e, d as rational num ber variables, and
[p, q] denotes the closed interval in 1 .) For this it suffices to show that, for any
rationals p, q with p < q and any positive rational e,

(*) ** = \\3d>0Vxe[p,q]Vye{p,q][\x-y\<d->\(p(x)-q>(y)\<e}l

Using the Definite Elements Lemma and the correlation between real numbers
over, Tand elements of C(e l; R), proving (*) amounts to showing that the set 5 =

LJln fl Inf* : [/(*)’ 9(0\ e [P> <?]A I f ( 0 - 9(0 I < d -H F{t, f(t)) - F(t, g(t) |< ef
d> 0 f,ge.C(X, j

coincides with,d.

This is proved by introducing a function e: , l x (0, ->) ->E defined by

e(t, d) = sup{ | F(t, x) ~~F(t, y)\: x ,y < z[p ,q ],\x-y \<d}.

The function | F(t, x) - F(t, y) | is continuous in the variables t, x, y and the

suprem um is taken over as compact subset of R x R . It follows that, for fixed d ,
e(t, d) is a well-defined, continuous function of t. Now for fixed t the real
function F(t, x) is uniformly continuous for x e [p, q], and so e(t d) -> 0 as d -> 0.
Accordingly, given e > 0 and to e we may choose d > 0 so that e(to, d) < e.
Since e is continuous there is a neighbourhood U of to in, I such that e(t, d) < e
for all t e U. It is now easily seen that U is included in S, so that S = ,T, and we
are done.

This completes Step 5 and we conclude that, in S$(' 0, every function from reals to
reals is continuous.

Remark. Recall from the Introduction that a set S is cohesive if whenever subsets
whenever U u V = S and U n V = 0 , then If = S or V = S. S is cohesive iff every

29 In fact this sentence asserts that cp is uniformly continuous on closed interval.

95
2-valued function on S is constant, and it is not hard to show that this is
equivalent to the condition that every N -valued function is constant.

Now it follows from the truth of Brouwer's Principle in S$(' \ that, in $$(' \ R is
cohesive. To prove this we show that from ( f ) above it follows (in IZ) that any
function q>: R -» 2 is constant. Thus let p, q be rational numbers with p < q, and
take c = 1 in ( f ). We get a rational d> 0 satisfying

(t V xy e [ p , q ] [ \ x - y \ < d ^ \ f ( x) - f ( y ) |< 1 -> f ( x) = f{y)].

Let d' be a rational such that 0 < d' < d. Let n be the least integer such that q - p
<nd', let p 0,...,p n be defined by p 0 - p , = p 7 + d ' and define
Ki = [Pi,Pi Jr^[p,q].Then

[p,q] =i=0 •

A straightforward inductive argument, using (f), now shows that cp has constant
value cp(p) on each K i and so also on [p , q ]. Since p and q were arbitrary with
p < q, (p is constant on the whole of M .

THE INDEPENDENCE OF THE FUNDAMENTAL THEOREM OF ALGEBRA FROM IZF

The Fundamental Theorem of Algebra (FTA) asserts that the field C of complex

num bers is algebraically closed, i.e. that every polynomial over C has a zero in

C. While FT A is provable in ZF, we shall establish its unprovability in IZF by

showing that it is false in the spatial extension 3S(C), where now C is the space

of complex numbers with the usual topology.

In the same way as for 1 , one shows the complex num bers over any open
subset LI of a topological space X are correlated with continuous functions in

96
C(U, C). This holds in particular when X is C itself. Write i for the complex

num ber over € correlated with the identity function on C.

Now consider the polynomial p(x) = x 2 - i in

For u e V (~}, if |u <e Cf e ^(C ), then u is correlated with a continuous function

/ : U —» C . In that case |p (u ) = 0(1 = | u 2 - i|] = In{t e U : / ( t ) 2 = t } .

If 0 g lp(u) - 0 ], then there is a neighbourhood l7 of 0 such that

V ci In{t gU : f( t) 2 = t} . Let q > 0 be a rational num ber such that the circles C
C about the origin with radii q, q2 are both contained in V. Then since/(f)2 = t in
V, the restriction o f / t o C is a section of the squaring function t e-* t 2 : C —» C .
But then this restriction would have to be a homeomorphism of C' to half of C.
But this is impossible since any circle, but no half-circle, remains connected
when a single point is removed .

Thus 0 £ I p(u) = 0 j|. Since this holds for arbitrary u, it follows that
0 £ |3 x(p(x) = 0 )||. Similarly, for any a e C , a <£[3 x(p(x) + a = 0 ) |. From this
we deduce that |C is algebraically closed 1 = 0 , i.e.

—iC is algebraically closed.

97
 Appendix
Heyting Algebras and Frames

LATTICES

A lattice is a (nonempty) partially ordered set L with partial ordering < in which
each two-element subset {x, y } has a suprem um or join - denoted by x y - and
v

an infimum or meet —denoted by x v y. A top (bottom) element of a lattice L is an

element, denoted by t (_l) such that x < t (_l < x) for all x e L A lattice with
top and bottom elements is called bounded. A lattice is trivial if it contains just
one element, or equivalently, if in it jl = t . A sublattice of a bounded lattice L is a
subset of L containing T and _l and closed under L's meet and join operations.
It is easy to show that the following hold in any bounded lattice:

X V JL = X, X A T — X,

X VX= X , X A X = X,
x v y = y v x, x a 1/ = y a x ,
x v (y v 2) = (x v y) v 2, x a (y a 2) = (x a y) a 2,
(x V y) A y = y, (x a y) v y = y

Conversely, suppose that (L, v, a , jl, t ) is an algebraic structure, with v, a binary

operations, in which the above equations hold, and define the relation < on L by

x < y iff x v y = y. It is then easily shown that (L, <) is a bounded lattice in
which v and a are, respectively, the join and meet operations, and 1 and 0 the
top and bottom elements. This is the equational characterization of lattices.

Examples, (i) Any linearly ordered set is a lattice; clearly in this case we have x a
y = min(x, y) and x v y = max(x, y).

(ii) For any set A, the power set PA is a lattice under the partial ordering ofset
inclusion. In this lattice X v Y = X U Y and X a Y= X f| V. A sublattice of a
power set lattice is called a lattice of sets.

98
(iii) If X is a topological space, the families t{X) and rC\X) of open sets and
closed sets, respectively, in X each form a lattice under the partial ordering of set
inclusion. In these lattices v and a are the same as in example (ii).

A lattice is said to be distributive if the following identities are satisfied:

x A (y v z ) = (x A y) v (x a z), x v (y a z) = (x v y) a (x v z ).

In the sequel by the term "distributive lattice" we shall understand "bounded

distributive lattice." An easy inductive argum ent shows that any nonempty
finite subset {xi, ..., xn} of a lattice has a suprem um, or join, and an infimum, or
meet: these are denoted respectively by x\ v . . . v xn, x\ a . . . a x „. An arbitrary
subset of a lattice need not have an infimum or a supremum: for example, the
set of even integers in the totally ordered lattice of integers has neither. If a
subset X of a given lattice does possess an infimum, or meet, it is denoted by
y\X; if the subset possesses a supremum, or join, it is denoted by \JX. When X is
given in the form of an indexed set {xz: i e /}, its join and meet, if they exist, are
written respectively Sf x t and / \ x t .
i>:/ t'.el

A lattice is complete if every subset has an infimum and a suprem um. The meet
and join of the empty subset of a complete lattice are, respectively, its top and
bottom elements. It is a curious fact that, for a lattice to be complete, it suffices
that every subset have a suprem um, or every subset an infimum. For the
suprem um (infimum), if it exists, of the set of lower (upper)30 bounds of a given
subset X is easily seen to be the infimum (supremum) of X.

Examples, (i) The power set lattice PA of a set A is a complete lattice in which
joins and meets coincide with set-theoretic unions arid intersections respectively.

(ii) The lattices ^X ) and (3{X) of open sets and closed sets of a topological space
are both complete. In the join and meet of a subfamily {IT: i e 1} are given
by

30 Here by a l o w e r (u p p er) b o u n d of a subset X of a partially ordered set P we mean an element p e P

for which a < x (x < a ) for every a e X.

99
 V ^= U ^ A ^,= Inf|^-
f“ / 1&J it . }

In ^ X ) the join and meet of a subfamily{A,: i e 1} are given by

VA = UA AA = DA-

Here InA and A denote the interior and closure, respectively, of a subset A of a
topological space.

HEYTING AND BOOLEAN ALGEBRAS

A Heyting algebra is a bounded lattice (H, <) such that, for any pair of elements
x, y e H, the set of z e H satisfying z a x < y has a largest element. This element,
which is uniquely determined by x and y, is denoted by x => y: thus x => y is
characterized by the following condition: for all z e H,

z < x => y if and only if z a x < y.

The binary operation on a Heyting algebra which sends each pair of elements x,
y to the element x => y is called implication; the operation which sends each
element x to the element x* = x => ± is called pseudocomplementation. We also
define the operation o of equivalence by x <=> y = (x => y) a (y => x). These
operations are easily shown to satisfy:

x => (y => z) = (x a y) => z, x => y - t <-> x < y, x <=>y = t <-» x = y,

y < z -> (x => y) < (x => z ), x a (x => i/) < 1/
y < x* <-> 1/ a x = 1 <-» x < y* , x < x** , x*** = x* , (x v 1/)* = x* a y*.

To establish the last of these, observe that

2 < (x v y)* <-> z a (x v y ) = jl

<-> (z a x ) v (z a y ) = _L

100
 <”> Z A I = 1 & 2 A 1 / - 1

<-> 2 < X* & 2 < if*

<-> 2 < X* A 1/*.

Any Heyting algebra is a distributive lattice. To see this, calculate as follows for
arbitrary elements x, y, z, u:

x a (y v z) < u <-» y v 2 < x => u

<-> y < (x => U) & 2 < (X => W)

<-» X Al/ < U &XA Z < U

<-> (x V l/) A (x V 2) < IL

Any linearly ordered set with top and bottom elements is a Heyting
algebra in which

x => y = t if x < y x => y = 1/ if y < x.

A basic fact about complete Heyting algebras is that the following identity holds
in them:

X a V j /, = V x a yt
iel iel

And conversely, in any complete lattice satisfying (*), defining the operation =>
by x => y = V{2: 2 a x < y) turns it into a Heyting algebra.

To prove this, we observe that in any complete Heyting algebra,

101
 x a \ / y j < z <-> V y i < x =>z
v=I iel

<-» y i < x =>z, all i

<-> y { a x < z, all i
<-» y x a y t < z.
iel

Conversely, if (*) is satisfied and x=> y is defined as above, then

(x => y) a x < V{z:z a x < y) a x =V {z a x : za x < y) < y .

So z < x => y —> za x < (x => y) a x < y. The reverse inequality is an

immediate consequence of the definition.

In view of this result a complete Heyting algebra may also be defined to be a

complete lattice satisfying (*). Complete Heyting algebras are also known as
frames.

If X is a topological space, then the complete lattice ^X ) of open sets in X is a

Heyting algebra. In tf(X) meet and join are just set-theoretic intersection and
union, while the implication and pseudocomplementation operations are given
by U => V = In((X - li) u V) and U* = In(X - U ) .

Let L be a bounded lattice. A complement for an element a e L is an element b g L

satisfying a v b = t and a a b - ±. In general, an element of a lattice may have
more than one complement, or none at all. However, in a distributive lattice an
element can have at most one complement. For if b, V are complements of an
element a of a distributive lattice, then a v b - a v b ' = t and a a b = a a b' = _l.

From this we deduce

b = b v _l = b v (a a b') = (b v a) a (b v bf) = r v (lb v lb') = b v V.

Similarly bf = b v V so that b - b'.

102
In a Heyting algebra H the pseudocom plem ent a* of an element a is not, in
general, a complement for a. (Consider the Heyting algebra of open sets of a
topological space.) But there is a simple necessary and sufficient condition on a
Heyting algebra for all pseudocomplements to be complements: this is stated in
the following

Proposition. The following conditions on a Heyting algebra H are equivalent:

(i) pseudocomplements are complements, i.e. x v x* = t for all x e H;
(ii) pseudocomplementation is of order 2, i.e. x** = x * for all x e H.

Proof, (i) -* (ii). Assuming (i), we have

X ** = X ** A T = X** A (x V X*) = (x** A x ) V (x** A X*) = (X** A x ) V JL = (X** A x ).

Therefore x** < x whence x** = x.

(ii)-> (i). We have (x v x*)* = x* a x** ~ ±, so assuming (ii) gives x v x* = (x v x*)™

= JL* = T. ■

We now define a Boolean algebra to be a Heyting algebra satisfying either of the

equivalent conditions (i) or (ii). The following identities accordingly hold in any
Boolean algebra:

x v y ~ y v x, x Ay - y ax
X V (ifV z) = (x V y) V 2, X A (1/ A 2) = (x A if) A 2
(x V y) A y = y, (x A y) v y = y
x a (y v z) - (x a y) v (x a z ), x v (y a 2) = (x v y) a (x v 2).
X V X* = T, X A X* = JL.
(x v y)* - x* a y*, (x a y)* = x* v y*
x** = x

It is easy to show that in any Boolean algebra x => y - x* v y. In a complete

Boolean algebra we have the following identities:

103
 (V*,)*
it-1
=itA*.*
I
xAVjy-VfxAyj
t, j !(?/

(A*,)* = V*. *
ic l IC-I
x v A m = A(xvj/J
le i i± I

Calling a lattice complemented if it is bounded and each of its elements has a

complement, we can characterize Boolean algebras alternatively as complemented
distributive lattices. For we have already shown that every Boolean algebra is
distributive and complemented. Conversely, given a complemented distributive
lattice L, write ac for the (unique) complement of an element a; it is then easily
shown that defining implication by x => y = xc v y turns L into a Heyting algebra
in which x* coincides with xr, so that L is Boolean.

The meet, join, and complementation operations in a Boolean algebra are called
its Boolean operations. A subalgebra of a Boolean algebra B is a nonem pty subset
closed under B's Boolean operations. Clearly a subalgebra of a Boolean algebra B
is itself a Boolean algebra with the same top and bottom elements as those of B.

Examples of Boolean algebras, (i) The linearly ordered set 2 = {0,1} with 0 < 1 is
a complete Boolean algebra, the 2-element algebra.

(ii) The power set lattice PA of any set A is a complete Boolean algebra. A
subalgebra of a power set algebra is called a field of sets.

(iii) Let F(A) consist of all finite subsets and all complements of finite subsets of
a set A. With the partial ordering of inclusion, F(A) is a field of sets called the
finite-cofinite algebra of A.

(iv) Let X be a topological space, and let C(X) be the family of all
simultaneously closed and open ("dopen") subsets of X. With the partial
ordering of inclusion, C(X) is a Boolean algebra called the dopen algebra of X.

Remark: The lattices (i), (ii), (iii) are only Boolean algrbras if the Law of
Excluded Middle is assumed.

An element a of a Heyting algebra H is said to be regular if a - a**. Clearly a

Hevting algebra is a Boolean algebra if and only if each of its elements is regular.
Let B be the set of regular elements of H; it can be shown that B, with the partial

104
ordering inherited from H, is a Boolean algebra —the regularization of H - i n
which the operations a and * coincide with those of H, but31 vb = ( v h )**. If H is
complete, so is B; the operation /\ in B coincides with that in H while Vs =
(V h)**.

COVERAGES ON PARTIALLY ORDERED SETS AND THEIR ASSOCIATED FRAMES

Let (P, <) be a fixed but arbitrary partially ordered set: we shall use letters p, q, r,
s, t to denote elements of p. A subset S of P is said to be a sharpening of or to
sharpen, a subset T of P, w rittenS < T , if V s g S3t g T(s <t). A sieve in P is a
subset S such that p e S and q <p implies q e S.

A coveage on P is a map C assigning to each p e P a family C (p) of subsets of

p i = {q: q < p\, called (C-)covers of p, such that, if q < p, any cover of p can be
sharpened to a cover of q, i.e.,

(*) S g C (p )& q < p 3T g C (q)[\/t g T3 s g S(t < s)].

Now we associate a frame with each coverage C on P. First, we define F to be

the set of sieves in P partially ordered by inclusion: F is then a fram e—the
completion of P ~ in which joins and meets are just set-theoretic unions and
intersections, and in which the operations => and are given by

/=>J = { p : / n p k J ) ~~J> i p : / n p i = 0 } .

Given a coverage C on P, a sieve / in P is said to be Q-closed if

3SeC(p)(S^I)~^peL

We write He for the set of all C-closed sieves in P, partially ordered by inclusion.

31 Here we write vp, vh for the join operations in B and H .

105
Lemma. If I e H e, / e H e, then I =>} e He-

Proof. Suppose that I e F , J e He and S c f = > / with S e C (p). Define U - {q e

I: 3seS. q < s}. Then U c J . If q e i n p i , then there is T e C (q) for which T ~<S.
Then for any t e T, there is s e S for which t < s, whence t e U. Accordingly
T c U c= /. Since / is a C-closed, it follows that q e f. We conclude that I n p i n],
whence p e p i c= I => J. Therefore I =>J is C-closed. ■

It follows from the lemma that He is a frame. For clearly an arbitrary

intersection of C-closed sieves is C-closed. So He is a complete lattice. In view of
the lemma the implication operation in F restricts to one in, He making He a
Heyting algebra, and so a frame. He is called the frame associated with C.

CONNECTIONS WITH LOGIC.

Heyting and Boolean algebras have close connections with intuitionistic and
classical logic32, respectively.

Intuitionistic first-order logic has the following axioms and rules of inference.

Axioms
a - > (ß - > a )
[a -> (ß h>y) -» [(a —» ß) —> (a -> y)]
a (ß - » a a ß)
aAß-»a aAß->ß
a - » a v ß ß - > a v ß
[a (ß -> y) -> [(a —> ß) —> (ot —» y)]
(a -» y) - ^ [ ( ß ^ y ) ^ ( a v ß ^ y)]
(a - > ß ) - > [(a - > -iß ) - > - i a ]
—iCX—v (ot —> ß)
a (t) -> 3xa(x) Vxa(x) -> a (y) (x free in a and t free for x in a)

x =x a(x) a x = y -» a(y

32 For accounts of both systems of logic, see, e.g. Bell and Machover [1977].

106
Rules of Inference
a, a -» ß
_ P

ß -» a(x) a(x) -» ß

(x not free in ß)

Classical first-order logic is obtained by adding to the intuitionistic system the rule
of inference
- 1-1a
a

In intuitionistic logic none of the classically valid logical schemes

LEM (law of excluded middle) a v -,a

LDN (law of double negation) -i-ia -» a
DML (de M organ's law) -i(a A ß ) - » -.a v -iß

are derivable. However LEM and LDN are intuitionisticallv equivalent and
DML is intuitionisticallv equivalent to the weakened law of excluded middle:

WLEM —id v -i-ia.

Also the weakened form of LDN for negated statements,

WLDN —l—i—id —» —iCX

is intuitionistically derivable. It follows that ^ny formula intuitionisticallv

equivalent to a negated formula satisfies LDN.

Heyting algebras are associated with theories in intuitionistic logic in the

following way. Given a consistent theory T in an intuitionistic propositional or
first-order language <&
, define the equivalence relation « on the set of formulas
of <0by cp^\|/ifThcp<-»v}/. For each formula cp write [cp] for its ^-equivalence

107
class. Now define the relation < on the set H(T) of ^-equivalence classes by

M < [y] if and only if T hep—>y. Then < is a partial ordering of H(T) and the

partially ordered set (H(T),<) is a Heyting algebra in which [cp] => [y] = [cp -> y],
with analogous equalities defining the meet and join operations, 0, and 1. H(T) is
called the the Heyting algebra determined by T. It can be shown that Heyting
algebras of the form H(T) are typical in the sense that, for any Heyting algebra L,
there is a propositional intuitionistic theory T such that L is isomorphic to H(T).
Accordingly Heyting algebras may be identified as the algebras of intuitionistic
logic.

Similarly, starting with a consistent theory T in a classical propositional or first-

order language, the associated algebra B(T) is a Boolean algebra known as the
Lindenbaum algebra of T. Again, it can be shown that any Boolean algebra is
isomorphic to B(T) for a suitable classical theory T .

As regards semantics, Heyting algebras and Boolean algebras have

corresponding relationships with intuitionistic, and classical, propositional
logic, respectively. Thus, suppose given a propositional language; let <£be its set
of propositional variables. Given a map /: H to a Heyting algebra H, we
e x te n d / to a map a 1—> |a j of the set of formulas of £
c to H by:

[a a ß] = [a] a |ß] [a v ß] = [aj v |ß] |a => ßj = [a] => |ß |

baj = I«]*

A formula a is said to be Heyting valid - written h a —if [a] = t for any such map
/ It can then be shown that a is Heyting valid iff h a in the intuitionistic
propositional calculus, i.e., iff a is provable from the propositional axioms listed
above.

Similarly, if we define the notion of Boolean validity by restricting the definition

of Heyting validity to maps into Boolean algebras, then it can be shown that a
formula is Boolean valid iff it is provable in the classical propositional calculus.

108
Finally, again as regards semantics, complete Heyting and Boolean algebras are
related to intuitionistic, and classical first-order logic, respectively. To be precise,
let W be a first-order language whose sole extralogical symbol is a binary
predicate symbol P. A Heyting algebra-valued yt-structure is a quadruple
M = (M, eq, Q, H), where M is a nonem pty set, H is a complete Heyting algebra
and eq and Q are maps M2 —> M satisfying, for all m, n, m', n' e M,

eq(mr m) = t , eq(m, n) = eq(n, m), eq(m, n) A eq(n, n ') < eq(m, n 7),
Q(m, n) A cr/(m, m ') < Q(tri, ri), Q(m, n) A ) < Q(m, n f ).

For any formula a of .^and any finite sequence x = <x\, ..., x„> of variables of y*
containing all the free variables of a, we define for any Heyting-valued c/~
structure M a map

[a]M*: Mn -> H

recursively as follows:

[xp = xqlMx = <rn\ mn> ^ eq(mP, mq),

[PxPXqlmx = <m ] mn> Q(mP, mq),
[a a ßJMx = [aJM* a [ßJMx, and similar clauses for the other
connectives,
\3y aJM*= <mi mn> »-+ V [a (*//u)]Mux(m,mi mn)
rneM
| V i / a | Mx = < m i m rt> ^ A Ia •••> mn)
meM

Call a M-ua/zd if [a]M* is identically t, where x is the sequence of all free

variables of a. Then it can be shown that a is M-valid for all M iff a is provable in
intuitionistic first-order logic. This is the algebraic completeness theorem for
intuitionistic first-order logic.

Similarly, if we carry out the same procedure, replacing complete Heyting

algebras with complete Boolean algebras, one can prove the corresponding
algebraic completeness theorem for classical first-order logic, namely, a first-

109
order formula is valid in every Boolean-valued structure iff it is provable in
classical first-order logic.

110
 Concluding Observations

In this book we have used frame-valued universes in proving the consistency of

set-theoretic assertions with 1ST. In fact, what has become the standard
procedure for proving the consistency with 1ST of a given assertion A is to
construct a certain sort of category - a topos33 - in which A holds in a "natural"
sense.

For example, the subcountability of , as well as many of the other assertions

concerning N mentioned in the Introduction, can be shown to hold in the so-
called effective topos Eff In Eff maps between objects constructed from the
natural numbers correspond to (partial) recursive functions between them. In

particular the countable subsets of N may be identified with the recursively

enumerable subsets, arid the detachable subsets of N with the recursive subsets.

The object N':i may be considered as the set Rec of (total) recursive functions on

M34. That being the case, if we write cpAfor the partial recursive function on N

with index x, arid U for the set of indices of total recursive functions35, the map

x (px for x g U is, in Eff, a surjection from U to NN, making N subcountable in

Eff. On the other hand is not numerable in Eff because it can be shown that,

in Eff, Brouwer's Continuity Principle holds. This is the assertion

33 Accounts of topos theory may be found in Bell [1988], Goldblatt [1979], Johnstone
[1977] and [2002], Lambek and Scott [1986], Mac Lane arid Moerdijk [1992] and McLarty
[1992].
34 Hence, in Eff, Church's thesis holds in the strong sense that every function A -> M is
recursive.
35 Since N is uncountable, so m ust U be; this corresponds to the result from recursion
theory that the set of indices of total recursive functions is not recursively enumerable.
Ill
 VF e N( ' ’V/ e W*3n e NVg e N*[Vm < n(f(m) = g(m)) - » F (/) = F(g)] 36

From this it would follow that, if F were an injection from to N, then, for each

/ e N äthere would exist an n e N for which

VgfENNVm<n(/(m) = g(m))-»/= g),

which is clearly impossible. Thus MN is not numerable in Eff, and hence so is

Par*(N,N), since it contains NN.

In Eff, both Par*(N,N) and P*N are countable. That Par*(N,N) in is countable in

Eff follows from the fact that it corresponds to the set of partial recursive functions

on N. The map x ^ q>xassigning to each ig N the partial recursive function with

index x is, in Eff, a surjection from N to Par*(N,N). (That being the case, as we

noted above, the set {x eN: x £ dom((px)} m ust be (in Eff) uncountable. This

corresponds to the fact that this set is not recursively enumerable.) A similar

argum ent - using the fact that P*N in Eff corresponds to the set of recursively

enumerable subsets of N - shows that P*N is also countable in Eff.36

36 In Eff the set N( ] of all maps N -» M:i may be identified with the set of effective
operations, that is, the set of functions F: Rec -» N such that there is a partial recursive
function cpwith the property that for every / e Rec and every index n for f we have
(pfw) = F(J). Brouwer's continuity principle asserts the continuity of every map to N from
Baire space, V.
112
While Nn fails to be num erable in Eff, Andrej Bauer has recently shown it to be

numerable in the related topos Eff! in which maps between objects constructed
from the natural num bers correspond to functions which are infinite time
computable, that is, computable by an infinite time Turing machine. This is a
Turing machine which is allowed to run infinitely long, with the computation
steps counted by ordinals. The power of these machines far exceeds that of
ordinary Turing machines: for example, both the halting problem and the
problem of deciding the equality of two total recursive functions are soluble

using infinite time machines. In Eff!, just as in E ff,, N " is subcountable. But in

Eff! Nn also satisfies the axiom of choice in the form: any total relation defined on

NN contains a function (this cannot be the casein Eff). Putting these two facts

together quickly yields an injection of N ! into N\ We conclude that the

numerability of NNis consistent with 1ST.

There are a num ber of topos models of Brouwer's Principle (BP) that all real
functions are continuous. As we have essentially shown in Chapter IV, BP
holds in Shv(f I)37. Mac Lame and Moerdijk [1992] present a different kind of
topos model of BP. BP also holds in any of the so-called smooth toposes. Each of
these may be considered to be an enlargement of the category Man of manifolds
(or spaces) and smooth maps to a topos which contains no new maps between
spaces, so that all such maps there - in particular those from E to E are still
smooth, and hence also continuous.

37 In fact BP caii be shown to hold in many other spatial toposes: see Hyland [1979]

113
 Historical Notes

Chapter 1. Friedman [1973] seems to have been the first to investigate systems of
intuitionistic set theory. Crayson [1978] undertakes a systematic investigation of
topology and ordinal arithmetic in an intuitionistic setting. That LEM follows
from the axiom of choice was first proved, in a category-theoretic setting, by
Diaconescu [1975]; the logical version was formulated and proved by Goodman
and Myhill [1978]. The investigation of the connection between choice principles
and logical principles is taken from Bell [2006}; see also Bell [2009].

Chapter 2. The characterization of N in terms of the simple recursion principle

(Proposition 5) is due to F. W. Lawvere in a category-theoretic setting.
Propostion 11 concerning monies on Q is due to Denis Higgs. Work on finite
sets in an intuitionistic setting (or their equivalents, finite objects in a topos), has
been extensive: for a complete bibliography see Johnstone [2002]. The section on
Frege's theorem is taken from Bell [1999i] and [1991Ü].

Chapter 3. Much of the discussion of real num bers presented here is based on
Johnstone's [2002] account of real num bers in a topos. Proposition 2 is due to
Johnstone [1979].

Chapter 4. Frame-valued models were first investigated by Grayson [1975] and

[1979], where it is also shown that Zorn's lemma is consistent with IZF (see also
Bell [1997]). The consistency of ZF relative to IZF was first proved by Friedman
[1973] and Powell [1975]; the proof given here is due to Grayson [1979]. A topos
model in which is subcountable was first produced by A. Joyal (see Fourman
and Hyland [1979], Johnstone [2002]). The model of the subcountability of
given in the text is a frame-valued version of the topos presented in Example
D.4.1.9. of Johnstone [2002]. The representation of real numbers (in a sheaf
topos) is due to M. Tierney. The proof that Brouwer's Principle holds for the real
numbers over Baire space is due to Scott [1970]. The status of Brouwer's
Principle in spatial toposes has been investigated by Hyland [1979]. The failure
of FTA in the sheaf topos over the complex numbers was first noted in Fourman
and Hyland [1979].

114
Appendix 2. The effective topos was first introduced by Hyland [1983]. The
concept of a smooth topos is due to F. W. Lawvere.

115
 Bibliography

Bell, J. L. [1988] Toposes and Local Set Theories: An Introduction. Clarendon Press,
Oxford, 1988. Dover reprint 2007.

Bell, J. L. [1997]. Zorn's lemma and complete Boolean algebras in intuitionistic

type theories. J. Symbolic Logic. 62,1265-1279.

Bell, J. L. [199i] Frege's theorem in a constructive setting. J. Symbolic Logic 64,

no. 2, 486-488.

Bell, J. L. [199ii] Finite sets and Frege structures /. Symbolic Logic, 64, no. 4, 152-
156.

Bell, J. L. [2005]. Set Theory: Boolean-valued Models and Independence Proofs.

Clarendon Press, Oxford.

Bell, J. L. [2006]. Choice principles in intuitionistic set theory. In A Logical

Approach to Philosophy. Springer. Heidelberg-London, New York.

Bell, J. L. [2008]. A Primer of Infinitesimal Analysis, 2nd. edition. Cambridge

University Press.

Bell, J. L. [2009]. The Axiom of Choice. College Publications, London.

Bell, J. L. and Machover, M. [1977]. A Course in Mathematical Logic. North-

Holland, Amsterdam

Diaconescu, R. [1975] Axiom of choice and complementation. Proc. Amer. Math.

Soc. 51,176-8.

Fourman, M. P. and Hyland, J. ML E. [1979]. Sheaf models for analysis. In

Fourman, M. P., Mulvey, C. J., and Scott, D. S. (eds.) Applications of Sheaves.
Proc. L.M.S. Durham Symposium 1977. Springer Lecture Notes in Mathematics
753, Pp. 280 - 301..

116
Fourman, M. P. and Scptt, D. S. [1979] Sheaves and logic. In Fourman, M. P.,
Mulvey, C. J., and Scott, D. S. (eds.) Applications of Sheaves. Proc. L.M.S. Durham
Symposium 1977. Springer Lecture Notes in Mathematics 753, pp. 302-401.

Goldblatt, R. [1979] Topoi: The Categorial Analysis of Logic. North-Holland,

Amsterdam.

Goodman, N. and Myhill, J. [1978] Choice implies excluded middle. Z. Math

Logik Grundlag. Math 24, no. 5, 461.

Grayson, R. J. [1975] A sheaf approach to models of set theory. M.Sc. thesis, Oxford
University.

Grayson, R. J. [1978]. Intuitionistic Set Theory. D.Phil. Thesis, Oxford University.

Grayson, R. J. [1979] Hey ting-valued models for intuitionistic set theory. In

Fourman, M. R, Mulvey, C. J., and Scott, D. S. (eds.) Applications of Sheaves. Proc.
L.M.S. Durham Symposium 1977. Springer Lecture Notes in Mathematics 753, pp.
402-414.

Haimos, P. R. [1963] Lectures on Boolean Algebras. Van Nostrand, New York.

Higgs, D. [1973]. A category approach to Boolean-valued set theory. Lecture

Notes, University of Waterloo.

Hyland, J. M. E. [1979]. Continuity in spatial toposes. In Fourman, M. R,

Mulvey, C. J., and Scott, D. S. (eds.) Applications of Sheaves. Proc. L.M.S. Durham
Symposium 1977. Springer Lecture Notes in Mathematics 753, pp. 442-465.

Hyland, J. M. E. [1983]. The effective topos. In The L. E. J. Brouwer Centenary

Symposium, Studies in Logic and the Foundations of Math., vol. 110, North
Holland, Amsterdam , pp. 165-216.

Johnstone, P. T. [1977]. Topos Theory. Academic Press, London.

Johnstone, P. T. [1979] Conditions related to De M organ's law. In Fourman, M.

R, Mulvey, C. J., and Scott, D. S. (eds.) Applications of Sheaves. Proc. L.M.S.

117
Durham Symposium 1977. Springer Lecture Notes in Mathematics 753, pp. 479-
491.

Johnstone, P. T. [2002]. Sketches of an Elephant: A Topos Theory Compendium, vols.

I and II. Oxford Logic Guides vols. 43 and 44. Clarendon Press, Oxford.

Lambek, J. and Scott, P. J. [1986]. Introduction to Higher-Order Categorical Logic.

Cambridge University Press.

Mac Lane, S. and Birkhoff, G. [1967] Algebra. Macmillan, New York.

Mac Lane, S. and Moerdijk, I. [1992]Sheaves in Geometry and Logic: A First

Introduction to Topos Theory. Springer-Verlag, Berlin.

McLarty, C. [1992]. Elementary Categories, Elementary Toposes. Clarendon Press,

Oxford.

Scott, D. S. [1970]. Extending the topological interpretation to intuitionistic

analysis, I I .. In Myhill, J. Kino, A. and Vesley, R. E ., e d s ., Intuitionism and Proof
Theory, North-Holland, Amsterdam, pp. 235-255.

118
 Index

associated frame 108 Existence Principle 73

Axiom of Choice for I 83 Extended Recursion Principle 33
Axioms of IZ 12 extensional 58
Axioms of IZF 55
false 61
Baire space 95 finite 35
Brouwer's Principle 95 frame 104
frame of truth values 15
Cantor's Theorem 2, 5 frame-valued model 61
Cauchy real num ber 54 frame-valued universe 62
Cauchy sequence 53 Frege structure 40
cohesive 8 Frege's Theorem 40
Collection argument 63
complemented 66 H-extension 61
completely (jl, 2)~ distributive 77 H-valued structure 60
completely JL-distributive 77
conditionally order-complete 50 indecomposable 8
connected 73 Induction Principle for VW 62
Contable Axiom of Choice 83 inductive 26
continuum 7 inhabited 13
core 68 inhabited 63
countable 1 intuitionistic set theory 1
countably generated 84
cover 107 Kuratowski finite 35
coverage 107
cumulative hierarchy 58 LEM 16
local 92
Dedekind finite 35 localizable 93
Dedekind infinite 35 locally connected 73
Dedekind real num ber 48
definite element 63 Mixing Lemma 65
Definite Element Lemma 66
detachable 6 natural num bers 26
discrete 16 near-local 92
disjoint refinement 73 nu-inductive 40
distributive, _l-(I, /) 76 numerable 1
DML 16

119
ordered pair 13
ordered pairof in 67

Peano structure 28
power set 13
preserves exponentials 77
Principle of Induction on Ordinals
57
pseudocom plem ent (*) 102

rank function 57, 58

real function ove X 92
refinable 73
Refinable Existence Lemma 74
refinement 73
regular 106

sieve 107
Simple Recursion Principle 29
spatial extension 6s
strict Frege structure 40
strictly finite 35
strong core 73
subcountable 1
subquotient 75
subset classifier 15

totally disconnected 66
transitive 58
true 61
truth value 61

Unique Existence Principle for VW

65
universe of H -sets 62

weak real num ber 48

well-founded 57
well-ordering 58
WLEM 16

Zorn's Lemma 70

120
While intuitionistic Cor constructive] s e t theory IST
has received a certain attention from mathematical
logicians, so fa r as I am aware no book providing
a systematic introduction to the subject has yet
been published. This may be the case in part
because, as a form of higher-order intuitionistic
logic - the internal logic of a topos - 1ST has
been chiefly developed in a tops-theoretic context.
In particular, proofs of relative consistency with
1ST fo r mathematical assertions have been
(implicitly] formulated in topos- o r sheaf-theoretic
term s, ra th e r than in the framework of Heyting-
algebra-valued models, the natural extension to
1ST of the well-known Boolean-valued models fo r
classical s e t theory.

In this book I o ffe r a brief but systematic

introduction to 1ST which develops the subject
up to and including the use of Heyting-algebra-
valued models in relative consistency proofs.
I believe th a t 1ST, presented as it is in the familiar
language of s e t theory, will appeal particularly to
those logicians, mathematicians and philosophers
who are unacquainted with the methods of
topos theory.