Introduction To Axiomatic Set Theory (1971) - Krivine

INTRODUCTION TO AXIOMATIC SET THEORY
SYNTHESE LIBRARY
MONOGRAPHS ON EPISTEMOLOGY,
LOGIC, METHODOLOGY, PHILOSOPHY OF SCIENCE,
SOCIOLOGY OF SCIENCE AND OF KNOWLEDGE,
AND ON THE MATHEMATICAL METHODS OF
SOCIAL AND BEHAVIORAL SCIENCES
Editors:
DONALD DAVIDSON, Rockefeller University and Princeton University
J AAKKO HINTIKKA, Academy of Finland and Stanford University
GABRIEL NUCHELMANS, University of Leyden
WESLEY C. SALMON, Indiana University
JEAN-LOUIS KRIVINE
INTRODUCTION TO
AXIOMATIC SET THEORY
D. REIDEL PUBLISHING COMPANY / DORDRECHT-HOLLAND
THEORIE AXIOMATIQUE DES ENSEMBLES
First published by Presses Universitaires de France, Paris
Translated/rom the French by David Miller
Library of Congress Catalog Card Number 71-146965
ISBN-13: 978-90-277-0411-5 e-ISBN-13: 978-94-010-3144-8
DOl: 10.1007/978-94-010-3144-8
All Rights Reserved
Copyright 1971 by D. Reidel Publishing Company, Dordrecht, Holland
Softcover reprint of the hardcover 1st edition 1971
No part of this book may be reproduced in any form, by print, photoprint, microfilm,
or any other means, without written permission from the publisher
INTRODUCTION
This book presents the classic relative consistency proofs in set theory
that are obtained by the device of 'inner models'. Three examples of such
models are investigated in Chapters VI, VII, and VIII; the most important
of these, the class of constructible sets, leads to G6del's result that the
axiom of choice and the continuum hypothesis are consistent with the
rest of set theory [1]I.
The text thus constitutes an introduction to the results of P. Cohen
concerning the independence of these axioms [2], and to many other
relative consistency proofs obtained later by Cohen's methods.
Chapters I and II introduce the axioms of set theory, and develop such
parts of the theory as are indispensable for every relative consistency
proof; the method of recursive definition on the ordinals being an import-
ant case in point. Although, more or less deliberately, no proofs have
been omitted, the development here will be found to require of the reader
a certain facility in naive set theory and in the axiomatic method, such
as should be achieved, for example, in first year graduate work (2
e
cycle
de mathernatiques).
The background knowledge supposed in logic is no more advanced;
taken as understood are such elementary ideas of first-order predicate
logic as prenex normal form, model of a system of axioms, and so on. They
first come into play in Chapter IV; and though, there too, all the proofs
(bar that of the reduction of an arbitrary formula to prenex normal form)
are carried out, the treatment is probably too condensed for a reader
previously unacquainted with the subject.
Several leading ideas from model theory, not themselves used in this
book, would nevertheless make the understanding of it simpler; for
example, the distinction between intuitive natural numbers and the
natural numbers of the universe, or between what we call formulas and
what we call expressions, are easier to grasp if something is known about
1 Numbers in brackets refer to items of the Bibliography, which is to be found on
p.98.
VI INTRODUCTION
non-standard models for Peano arithmetic. Similarly, the remarks on p.
44 will be better understood by someone who knows the completeness
theorem for predicate calculus.
All these ideas can be found for example in [4] (Chapters I, II, III)
or [5] (Chapters I, II, III).
The approach of the book may appear a little odd to anyone who thinks
that axiomatic set theory (as opposed to the naive theory, for which, per-
haps, this is true) must be placed at the very beginning of mathematics.
For the reader is by no means asked to forget that he has already learnt
some mathematics; on the contrary we rely on the experience he has
acquired from the study of axiomatic theories to offer him another one:
the theory of binary relations which satisfy the Zermelo/Fraenkel axioms.
As we progress, what distinguishes this particular theory from other
axiomatic theories gradually emerges. For the concepts introduced
naturally in the study of models of this theory are exactly parallel to the
most fundamental mathematical concepts - natural numbers, finite sets,
denumerable sets, and so on. And since standard mathematical vocabulary
fails to provide two different names for each idea, we are obliged to use
the everyday words also when referring to models of the Zermelo/Fraenkel
axioms. The words are thereby used with a completely different meaning,
the classic example of this being the 'Skolem paradox', which comes
from the new sense that the word 'denumerable' takes when interpreted in
a model of set theory.
Eventually it becomes obvious that even the everyday senses of these
words are by no means clear to us, and that we can perhaps try to
sharpen them with the new tools which are developed in the study of set
theory. Had this problem been posed already at the beginning of the
study it would have been tempting to shirk it, by saying that mathematics
is merely the business of manipulating meaningless symbols.
It is an open question how much set theory can do in this field; but
it seems likely that what it can do will be of interest beyond mathematical
logic itself.
Note: This translation incorporates a number of additions and correc-
tions to the original French edition.
CONTENTS
INTRODUCTION
Chapter I: The ZermelojFraenkel Axioms of Set Theory
Chapter II: Ordinals, Cardinals
Chapter III: The Axiom of Foundation
Chapter IV: The Reflection Principle
Chapter V: The Set of Expressions
Chapter VI: Ordinal Definable Sets. Relative Consistency of the
v
1
13
35
48
56
Axiom of Choice 63
Chapter VII: FraenkelfMostowski Models. Relative Consistency
of the Negation of the Axiom of Choice (without the
Axiom of Foundation) 70
Chapter VIII: Constructible Sets. Relative Consistency of the
Generalized Continuum Hypothesis 81
BIBLIOGRAPHY 98
CHAPTER I
THE ZERMELO/FRAENKEL AXIOMS OF SET THEORY
When we formulate the axioms of set theory, we have to rely on an intuitive
understanding of sets, in much the same way that we develop the axioms
for a vector space from commonsense ideas about three-dimensional
space. Despite this, once we have the axioms of set theory written down
we are free to study any other structure in which they hold, just as we
study many vector spaces over and above the Euclidean space R3. Thus
set theory is no different from any other axiomatic theory familiar to the
reader. It is, like the theories of groups, rings, fields, vector spaces,
lattices, and so on, an abstract theory.
A structure satisfying the axioms of set theory is called a universe. A
universe tW is basically a collection of objects called sets. We do not say
'a set of objects', since what we are going to call sets are just those objects
of which tW is a collection. It is a straightforward precaution when defining
vector spaces, for example, not to use the same word both for the vector
space and for a vector in it. The same applies here.
Also involved in tW is a single undefined binary relation, a relation
which links some of the objects of tW with others. It is called the member
ship relation, and is denoted by E. We read the expression 'x E y' as 'x
belongs to y', or 'the set x belongs to the set y', or 'x is a member of y',
or 'x is an element of y', or even 'the set y contains the element x'.
Since E is a relation on the universe tW, it holds only between sets. The
symbol'E' and the words 'belong', 'contain', 'member', and 'element' will
be used only when this particular relation is intended; and should we
ever use any of these words in their everyday senses, we shall state this
explicitly. For example, we will not normally say 'the object xis an element
of the collection tW', but' x is in tW'.
So a universe can be pictured as a graph of the kind shown in figure 1 ;
the arrow indicates that the object at its head is an element ofthe one at
its tail. For example, b E a, C E c.
Nothing said so far imposes any restriction on the binary relation E.
In order to do just this we now begin to list the axioms of set theory.
2 INTRODUCTION TO AXIOMATIC SET THEORY
f
e
Fig. 1.
1. AXIOM OF EXTENSIONALITY
No two distinct sets in r1lt have the same elements. We write this as
'v'x'v'y['v'Z(Z E x _ Z E y) X = y].
This axiom is not satisfied by the binary relation shown in the figure; for
band c are different, yet each has c as its only element.
A useful 'axiom' to put down now, though we will not number it since
it turns out to be a consequence of later ones, is the pairing axiom. Given
two sets a and b, this axiom guarantees the existence of a set c containing
a and b as its only elements. By the axiom of extensionality there is only
one such set c. We write the pairing axiom as
'v'X'v'y3Z'v't[tEz-t=X v t=y].
We write {a, b} for the set c whose only elements are a and b. A special
case is when a and b are the same set; it follows from the axiom that there
is a set {a} containing a, and a alone.
THE ZERMELO!FRAENKEL AXIOMS 3
If a=/=b we call {a, b} a pair. The set {a} is sometimes called a singleton
or unit set.
From two sets a and b three applications of the pairing axiom yield
the set {{a}, {a, b }}, called the ordered pair of a and b and usually denoted
by (a, b). The following theorem justifies this definition.
THEOREM: If(a, b)=(a', b') thena=a' andb=b'.
PROOF: If a = b then (a, b) = {{a}}, and so (a, b) has only one element.
Thus (a', b') is also a singleton, which means that a' =b'. It follows that
{{a}} = {{a' }}, and so a=a' ; thus b=b' too.
When a=/=b, (a, b) is a pair, and so (a', b
'
) must also be one. This can
only be true if a' =/= b' .
But as {{a}, {a, b}} = {{a'l, {a', b'}}, either
{a} = {a', b' } and {a, b} = {a'l
or
{a} = {a'l and {a, b} = {a', b' }.
The former possibility is ruled out simply because {a} is a singleton whilst
{a', b'} is a pair. Thus {a} = {a'l -and so a=a' -and {a, b}={a', b' } -and
sob=b'1
If a, b, c are sets, the set (a, (b, c is called the ordered triple of a, b, c,
and is denoted by (a, b, c).
THEOREM: If(a, b, c) = (a', b', c') thena=a', b=b', c=c'.
PROOF: From the identity of (a, (b, c and (a', (b', c' we get a=a'
and (b, c)=(b', c'). Thus also b=b' and c=c'.1
In the same way we can define the ordered quadruple (a, b, c, d) by
(a, b, c, d) = (a, (b, c, d,
and for any n>O the ordered n-tuple (ao, ai' ... , an-i) by
(ao, a1, ... , an-i) = (ao, (a1' ... , an-1
The previous theorems generalize.
THEOREM:.if (ao, a1' ... ' an-i) = ... , then ao a1 ... ,
a,,-l
PROOF: Obvious by induction on n.1
4 INTRODUCTION TO AXIOMATIC SET THEOR Y
It should be noted that, given three distinct sets a, h, c in the universe <11,
we do not yet have any way of proving the existence of a set d containing
just a, h, and c as members. The next axiom fills this gap.
2. UNION AXIOM (OR SUM-SET AXIOM)
According to this axiom, to every set a there corresponds a set h whose
members are precisely the members of the members of a. We write this
formally as
Vx3yVz [z E y +-dt(t E x A Z E t)].
The set b, called the union of the members of a (or, more briefly, the union
of a) and denoted by U x (or Ua), is unique; for any set h' having the
xea
same property would necessarily have the same elements as, and so be
identical with, b.
This axiom solves the problem posed above. For if a, h, and c are sets,
the set d = U {{a, h}, {c}} contains a, h, and c, and nothing more besides.
We write the set d as {a, b, c}.
More generally, any finite number of sets ao, a
1
, ... , a
ll
-
1
can be
collected together into a set {ao, a
1
, ... , a
n
-
1
} containing them and them
only. This is easily proved by induction on n; we need only observe that
U {{a
O
,a
1
, ... ,a,,-2}, {a
ll
-
1
}} has the desired property for n if
{a
o
, a
1
, ... , a
ll
-
2
} has it for n-l.
If a and b are sets, then the union of {a, h} is called the union of a and h,
and written a u b.1t is trivial that a u (b u c)=(a u b) u c=(h u a) u c=
= the union of the set {a, b, c}.
The same applies for any finite number of sets ao, a1'"'' a
ll
-
1
; the
union of the members of {ao, a1' ... , a
ll
-1} is called the union of ao, a1>
and ... , and a,,-l and written a
o
u a
1
u ... u a,,-l .
3. POWER-SET AXIOM
Let a and b be sets in <11. The statement Vx(x E a -+ x E b) is abbreviated
a c: b, and read 'a is included in b' or 'a is a subset of b'.
The power-set axiom tells us that, given a set a, there exists a set b
whose members are just a's subsets; in symbols it becomes
Vx3y'f/z [z E y -Ho Z c: x].
THE ZERMELO/FRAENKEL AXIOMS 5
By the axiom of extensionality only one set b can have this property. It is
called the power-set of a and written gJ(a).
It is worth emphasizing at this point that we shall now use the words
'subset' and 'include' only in the sense given them above, to express a
relation between two objects of the universe. This sense is quite different
from the usual one (when we say, for instance, that a class is included
in the universe). We call it the formal sense of the word 'include' (or
'subset'), and the usual one will be called the intuitive sense. To avoid
from the start any possible confusion that might arise, let us agree that
when we use the words 'subset' and 'include' without qualification, we
are using them in their formal senses; any use of these words in their
intuitive senses will be noted as such explicitly.
For instance: each set a defines a subset (in the intuitive sense) of the
universe <?t, namely the part A composed of the elements of a. Now if b
is a set included in a, then the corresponding B is included (in the intuitive
sense) in A. But there may be - and indeed often are - subsets (in the
intuitive sense) of A to which no object of the universe corresponds; to
which, that is, there is no corresponding subset of a.
Before giving the remaining axioms of set theory we must discuss how
relations, over and above those we already have, can be defined on <?to At
present we have only the membership relation x E y, and the equality
relation X=y (which is satisfied by a and b if and only if a and b are the
same object). The following rules allow us to define others.
I. Given a relation (of three arguments, for instance) R(x, y, z), and
an object a in <?t, then we can define a binary relation R(a, y, z), to be
satisfied by the objects band ciff R(a, b, c) holds.
II. Given a relation (three arguments, say) R(x, y, z), then we can define
a binary relation R(x, x, z); this is satisfied by a and b iff R(a, a, b) holds.
III. Given any relation (say of two arguments) R(x, y), we can define
.R(x, y), a relation satisfied by a and b iff R(a, b) does not hold.
Likewise from two relations (of three arguments again) R(x, y, z) and
S( u, x, v) we can define a relation R( x, y, z) v S( u, x, v) offive arguments
satisfied by a, b, c, d, e iff one or other of R(a, b, c) and Sed, a, e) holds.
IV. Again, from a ternary relation R(x, y, z) we can define a binary
relation 3yR(x, y, z) which is satisfied by a and c iff R(a, b, c) holds for
somebin 'W.
Starting from the two binary relations x E y and x = y we can, by repeated
applications of these rules, obtain relations of any number of argu-
ments.
Suppose that R(x, y) and 8(y, z) are two binary relations; then the
relation ..., R v 8 is written R -+ 8. The relation ..., (..., R v ..., 8) is
written R A 8; it is satisfied by a, b, and c iff R(a, b) and 8(b, c) both
hold. Similarly, R +-+ 8 abbreviates (R -+ 8) A (8 -+ R), and is satisfied
by a, b, c iff R(a, b) and 8(b, c) either both hold or both fai1.
Likewise, the singulary relation ...,3x...,R(x, y), written VxR(x, y) is
satisfied by b iff R(a, b) holds for every object ain 0/1.
Relations constructed from x E y and x=y by rules I-IV are thus
defined by formulas composed (though not in an arbitrary fashion) from
the symbols =, E, ..." v, 3, the variables x, y, z, u, v, ... , and objects in Cl/I.
It is clear that further relations on 0/1 can be defined by other methods.
But we shall not consider them in this book.
A relation R(x) of a single argument is usually called a class. A class is
a subcollection (intuitively speaking) of the universe Cl/I. For example, the
formula
Vu[uEx-+3v[VEX A Vt(tEV+-+t=U v tEU)]]
defines a class; the formula
uEx-+3v[VEX A Vt(tEV+-+t=U v tEU)]
defines a binary relation R(u, x). So if a is in 0/1, the formula R(a, x),
namely
aEx-+3v[VEX A Vt(tEV+-+t=a v tEa)],
defines a class.
Objects of 0/1 which appear in a formula E are called the parameters
of E. R(u, x), for example, has no parameters; R(a, x) has the object a as
its sole parameter.
A formula of no arguments at all, a closed formula or sentence as it is
called, is either true or false in the universe. For example, the formula
3x3y[Vz(z y) AyE X A VuR(u, x)] (where the relation R is defined
above) is a sentence without parameters (it will later turn up in a slightly
different form as the axiom of infinity).
The theorems of set theory (and the axioms in particular) are all closed
formulas without parameters.
THE ZERMELO!FRAENKEL AXIOMS 7
Equivalence Relations: A binary relation R(x, y) is an equivalence relation
if, whatever objects a, b, c might be,
R(a, b) R(a, a) A R(b, b);
R(a, b) R(b, a);
R(a, b) A R(b, c) R(a, c).
The class R(x, x) is called the domain of the equivalence relation. R(a, b)
is also written a",b (mod. R). For any set a in 0/1, the class R(a, y) is
called the equivalence class of a.
Ordering Relations (in the weak sense): A binary relation R(x, y) is an
ordering relation (or, briefly, an ordering) in the weak sense if for every
a,b,andc
R(a, b) R(a, a) A R(b, b);
R(a, b) A
R(a, b) A R(b, c) R(a, c).
The class R(x, x) is the domain of the ordering. R(a, b) is also written
(mod. R); R(a, b) A b=a may similarly be abbreviated as a<b
(mod.R).
R is a (weak) linear (or simple, or total) ordering if, over and above all
this, one or other of R(a, b) and R(b, a) holds for any two objects a, b
in R's domain.
Ordering Relations (in the strict sense): A binary relation R(x, y) defines
astrict ordering on a class D(x) if for all a, b, c,
R(a, b) --+ D(a) A D(b);
-, (R(a, b) A R(b, a);
RCa, b) A R(b, c) R(a, c).
R(a, b) may again be written a<b (mod. R). It is obvious that if R(x, y)
is a strict ordering, then the relation R(x, y) v [D(x) A D(y) A X= y]
is a weak one, with domain D.
A strict ordering R on D is linear if for any objects a and b in the class
D, either R(a, b) or R(b, a) or a= b.
Functional Relations: Consider by way of illustration a ternary relation
R(x, y, z). We will call it afunctional relation of two arguments if
VxVyVzVz' [R(x, y, z) A R(x, y, z') --+ z = z'].
The binary relation 3zR(x, y, z) is then called the domain of the func-
tional relation R. The class 3x3yR(x, y, z) is called the range of the
functional relation R.
For any (for example binary) functional relation defined everywhere
(in other words, a functional relation with domain x = x A Y = y) we can
introduce a new symbol <P, and write the functional relation as z= <P(x, y).
Then, for any formula E(u, v, w), the formulas
3z[R(x, y, z) A E(z, v, w)]
and
'v'z[R(x, y, z) -+ E(z, v, w)]
(which are equivalent to one another) can be written simply as
E[ <P(x, y), v, w].
For example, the functional relation
'v't[tezoH-t=x v t=y]
is written z = {x, y}; thus the equivalent formulas
3z['v't(tezoH-t=x v t=y) A zea]
'v'z [Wet e z oH- t = x v t = y) -+ z e a]
can both be abbreviated by {x, y} Ea.
We can now go on to state the other axioms of set theory.
4. AXIOM SCHEME OF REPLACEMENT (OR SUBSTITUTION)
Suppose a formula E(x, y, ao, ... , a
k
-
1
) with parameters a
o
, ... , ak-l
defines a singulary functional relation; and let a be any set. Then we shall
postulate that the universe I7It contains a set b whose elements are just the
images under this functional relation of those elements of a within its
domain. The demand on I7It that it satisfy this condition for any singulary
functional relation is known as the axiom scheme of replacement. The
scheme is rendered symbolically by the following infinite list of sentences:
'v'xo ... 'v'Xk-l {'v'x'v'y'v'y' [E(x, y, xo, ,." Xk-l)
A E(x, y', xo, ... , Xk-l) -+ y = y']
-+'v't3w'v'v[vewoH-3u[uet A E(u,v,xO, ... ,Xk-l)]]}'
Here E(x, y, xo, .. " Xk-l) may be any parameter-free formula with at
least two free variables x and y.
The axioms 1,2,3, the scheme 4, and the axiom of infinity (introduced
on p. 29 below) make up the set theory of Zermelo and Fraenkel, usually
abbreviated ZF.
Easily derived from the scheme of replacement is the following scheme.
Scheme o/Comprehension: If ais a set, and A (x, a
o
, ... , ak-l) any formula
of one free variable (and parameters ao, ... , ak-l)' we can establish that
there is a set whose elements are exactly those elements of a for which A
holds. This is the content of the comprehension scheme, which consists
of the following infinite list of sentences:
'Vxo ... 'Vxt_
1
'Vx3y'Vz [z e y +-+- (z e x A A(z, xo, ... , Xk-l))]'
Here A(x, xo, ... , Xk-l) is any parameter-free formula of at least one free
variable x. To derive any instance of the scheme from the scheme of
replacementitisenoughtonotethattheformulay=x A A (x, ao, ... , a
k
-
l
)
defines a singulary functional relation F whose domain is the class
A(x, ao, ... , ak-l)' By replacement then, there is a set b whose elements
are just the images under F of those members of a in F's domain; it is
immediate that b is the set whose existence we set out to prove.
The set b is usefully represented by the notation
{x e a I A (x, ao, ... , ak-l)}'
THEOREM: There is one and only one set without elements.
PROOF: Let a be any set at all. We apply the comprehension scheme to
the set a and the formula x::/:x to get a set b={x e a I x::/:x}, which is
obviously without elements. That there cannot be more than one such
set follows at once from the axiom of extensionality .
The set just defined is called the empty set and is denoted by 0.
Let us give a derivation of the pairing axiom from axiom 3 and the
scheme 4. First of all, every subset of 0 is empty, so that &(0) has 0 as
its only element; thus {0} exists. This latter set is a singleton, so if a c: {0}
then a=0 or a= {0}. So &({0}) has just two distinct elements, 0 and {0};
thus {0, {0}} exists. Now let a and b be any sets you like. It is apparent
that the relation (x=0 A y=a) v (x={0} A y=b) is a singulary func-
tional relation, so we can apply the replacement scheme to it and consider
the image of the set {0, {0}} thereunder. This turns out to be the set
{a, b}, and thus the pairing axiom is proved.
We say that a class A (x) corresponds to a set (or even, twisting language a
bit, is a set) if there is a set a such that Vx(x E a +-* A (x)). More generally,
a relation (of three arguments, for example) A (x, y, z) corresponds to a
set, or even is a set, if there is a set a for which we have VxVyVz[ <x, y, z) E
E a +-* A (x, y, z)J.
It is not true that every class corresponds to a set; for example, the
class x x does not. If it did, indeed, we would for some a have that
Vx(x E a +-* x x); so, in particular, a E a +-* a if; a, which is evidently
false. (This is Russell's paradox.)
The class X=X (the universe 0Zt, in other words) also fails to correspond
to a set. For were there any set a such that Vx(x E a) were true, we could,
by using the comprehension scheme, produce a set b such that Vx(x E b +-*
+-* X E a 1\ x x); that is, Vx(x E b +-* X x). So Russell's paradox would
turn up all over again.
Cartesian Product of two sets: Consider the formula X(z) abbreviated by
3x3y[z=<x,y) 1\ xea 1\ yeb].
To write this out in full we would first have to use the definition of the
functional relation Z= <x, y), obtaining
3x3y3u3v[z = {u, v} 1\ U = {x} 1\ V = {x, y}
A xEa A yeb].
The braces could then be eliminated with the definition of the functional
relation Z= {u, v}, leaving
3x3y3u3v[Vt(tez+-*t=u v t=v)
A V t
f
(tf e u +-* tf = x)
A V tff (tff e v +-* tff = X V tff = y)
AxeaAyeb].
So X(z) is just this last formula, and it clearly is the class of all ordered
pairs <x, y) such that x E a and Y E b. (We will not bother to carry
out this sort of check in the future; but it is suggested that the reader try
a few, to get himself used to the business of manipulating formulas.)
Now this class X is in fact a set; for if x E a and y e b then (by the defini-
tion of (x,y) (x,Y)E&(&(aub)). But then X(z) is equivalent to
X(z) 1\ z e & (&(a u b), so that by the comprehension scheme X is
a set.
This set X is called the Cartesian (or direct, or cross) product ofthe sets
a, b, and it is written a x b. The Cartesian powers a x a, a x (a x a), and so
on, are sometimes abbreviated a
2
, a
3
,
A singulary functional relation R(x, y) whose domain is a set is itself
a set. For suppose that the domain is a; then by the replacement scheme
there is a set b consisting of the images of a under R. Consequently
R(x, y) is equivalent to R(x, y) 1\ (x, y) E a x b, and this class is a set
f by the comprehension scheme. Such a set f is called a function defined
on a, with values in b, or a map from a to b, or afamily of sets indexed by a.
If a is a map we will sometimes write dom(a), rng(a) for its domain and
range respectively.
The formula 'fis a map from a into b', written more compactly f: a -+ b
is therefore just the following formula:
f c a x b 1\ 'rfx'rfy'rfy' [(x, y) E f 1\ (x, y') E f -+ y = y']
1\ 'rfx[xea-+3y(yeb 1\ (x,y)Ef)].
We leave to the reader the task of eliminating the defined symbols c, x,
and ( ).
If a and b are sets the class X of all maps from a into b is again a set. For
a map from a to b is a subset ofax b, and therefore a member of &(a x b).
SO XU) is equivalent to XU) 1\ fE &(a x b). Comprehension yields
the desired result.
This set of maps is written abo
Union, Intersection, and Cartesian Product of a family of sets: Let a be a
family of sets indexed by I; a is of course just a map whose domain is I,
and to indicate this we will usually write a family as (ai)iEI'
The union ofthefamity (ai)iEl> written Ua
i
, is the union of the range
iEI
of the map a. By the replacement scheme and union axiom it is a set b,
and we have 'rfx[ x E b 3i(i E I 1\ X E ai)]'
Likewise the intersection of the family (aJiEl is the class X(x) defined
by 'rfi(i E I -+ X E a;). If 1=0, then X is the class of all sets, and so not a
set itself. Otherwise, take any io E I; then X(x) is equivalent to x E aio 1\
1\ Vi(i E I - X E ai), and X is seen to be a set, by comprehension. We
write it n a
i

ieT
Now take the class X of map sf: 1- Ua
i
such thatf(i) E a
i
for every
ieI
i E I. Such a map belongs to the set I( U aJ, so XU) is equivalent to
ieI
XU) 1\ f E I ( U a) X is thus a set, called the Cartesian (or direct, or
ieI
cross) product of the family (ai)ieb and written X ai (or sometimes IT aJ
ieT ieT
Note that for the rest of the book we shall be using the words 'map' and
'function' in the sense defined above, which is by no means their every-
day mathematical sense (since here, for example, all maps f: a -4 b
are in the universe). As we proliferate definitions the same thing will
happen to nearly all the familiar words of mathematical language. If we
ever use the words in their normal senses (and this will not happen often),
we will, as we decided before, say so explicitly. For example: let a and b
be two sets, and A, B be the corresponding parts of the universe. If f
maps a into b there will correspondingly be a map (in the intuitive sense)
from A to B. But there may well be, intuitively speaking, maps from A
to B to which there are no corresponding maps from a into b.
CHAPTER II
ORDINALS, CARDINALS
Well-ordering Relations: Let a be a set, all of whose elements lie in the
domain of some linear ordering R(x, y). We say that a is well ordered
by R if every non-empty subset of a has a smallest element (mod. R). It
is immediate that if a is well ordered by R then every subset of a is also
well ordered by R.
If rca x a then the pair (a, r) is said to be a well-ordered system if a
is well ordered by the relation (x, y) E r.
Suppose that the set a is well ordered by the relation R. A subset s of a
is said to be an initial segment of a if XES and y ~ x (mod. R) imply YES,
for any x, yEa. For each Xo in a we write Sxo(a, R) (or Sxo(a) if there is
no danger of ambiguity) for the set {x E a I x<xo (mod. R)}, and this
set is evidently an initial segment of a.
Indeed, a subset sea is an initial segment of a iff s = a or s = Sxo (a) for
some Xo Ea.
This is easily proved by noting that if s is an initial segment of a, but
not a itself, then the set a""'-s = {x E a I x rf: s} is non-empty. Consequently,
it has a least element xo; so if x<x
o
, x must be in s; and conversely, if
~ xo, we cannot have XES (for that implies that Xo E s).
An initial segment of a which is not just a itself is called a proper initial
segment of a.
If R is a linear ordering, and x is in the domain of R, let Sx(R) stand
for the class of y's in the domain of R which are <x (mod. R). Then R
is called a well-ordering (relation) if, for any x in its domain, the class
Sx(R) is a set, and is also well ordered by R.
If R is a well-ordering with domain D, and T any non-empty subclass of
D (that is, Vx( T(x) D(x) /\ 3xT(x) , then Thas a least element (mod. R).
For Tis non-empty, so take Xo in T. Either Xo is the least element of T,
or else the class T /\ Sxo(R) is non-empty. This class is however a set
(since Sxo(R) is a set), and indeed a non-empty subset of Sxo(R), which
is well ordered. It therefore has a least element, which is obviously also
the least element of T.
(Our talk here of elements of the class T is to be understood as using
'element' in an informal way only.)
The Class of Ordinals: A set a is called transitive if every element of a is
also a subset of ct; that is, if\fx(x E a --+ x c a). A set ct is called an ordinal
if, in addition to being transitive, it is strictly well ordered by the member-
ship relation x E y.
The class 'a is an ordinal' is written On(a); spelt out more or less in
full this becomes
\fx\fy [x E a AyE ct --+ X ~ Y v Y ~ x]
A \fx\fy\fZ[XEct A YEa A ZEa A xEy A YEZ --+ XEZ]
A \fz[z c: ct A Z #0 --+ 3X(XEZ
A \fY(YEZ --+ XEy V x=y))]
A \fx\fy [x E ct AyE X --+ YEa] .
(In this extensive formula there is no explicit mention of the fact that the
ordering x E y is linear; however, no such clause is needed, since we have
stated that every non-empty subset of a contains a smallest element,
and a fortiori this applies to subsets of two elements.)
Examples of ordinals that are easily shown to be such are 0, {0},
{0, {0}}.
Let ct be an ordinal. Then the initial segments of ct are a itself, and the
elements of ct.
For a proper initial segment of ct is S ~ a ) for some e E a; and S ~ a ) =
= {11 E ct 111 < e} = {11 E ct 111 E e} = e n ct. But since e c a, this last term
reduces to e, and the result is proved.
Every element of an ordinal is an ordinal.
Take e E a; then e c ct, so membership well-orders e. Furthermore,
if y E X and x E e then x E a (since e c ct); so x c ct, so yEa. Now
membership strictly orders the elements of ct, so from y E X and x E e we
get y E e. Thus x E e implies that x c e, and the two conditions for e
to be an ordinal are established.
For every ordinal ct, ct ~ ct.
If e is any element of a, then e e, since E strictly orders ct. So if a is
an element of a we must have a ~ ct. Thus ct ~ ct.
For every two ordinals a, p, either ct = p, or p E ct, or ct E p,' moreover, the
three cases are mutually exclusive.
ORDINALS, CARDINALS 15
Put e=(X n p. Then e consists of all members ohthat are <P (mod. e),
and so is an initial segment of (x. Thus e = (X or e e (x. Likewise e is an initial
segment of p, so e = P or e e p. All in all then, there are four possibilities:
e = (X and e = p ; then (X = P .
e = (X and e e p ; then rx e p .
e = p and e e rx ; then p E rx .
e e (X and e e p ; then e e (X n p, so by the definition
of e, e e e. But this is a contradiction, since e is an ordinal.
Furthermore, (X = P and rx E p together would imply that PEP; similarly,
if (X e p, then rx c p, so if P E (X were also true, we would have PEP again.
But p p. Thus the three cases are mutually exclusive.
The membership relation on the class On (that is, the relation x E y 1\
1\ On(x) 1\ On (y) isawell-ordering.
We have just shown the relation in question to be a strict linear ordering.
So we need only note that since e <rx +-+ e e rx, S,,(E)=rx for every ordinal
rx, and rx is well ordered bye.
Observe that if rx, P f}re ordinals then rx ~ p +-+ rx c p.
On is not a set.
Suppose On is a set a. Then a is well ordered by E. Moreover, every
element of a is an ordinal, therefore a set of ordinals, therefore a subset
of a. But then a qualifies as an ordinal; so a e a. But this is impossible
for ordinals.
If rx is an ordinal the least ordinal greater than rx is rx u {rx}. This is called
the successor of rx, and written rx + 1.
It is easily checked that if rx is an ordinal then so is P=rx u {rx}. Now
sincerx e p, wehaverx<p; andifa.<ythenrxey, sorx c y, sorx u {(X} c y.
Thus ~ y
Since 0 c rx for any (x, the ordinal 0 is the least ordinal of all. We call
it O. Its successor {0} is called 1; its successor 1 u {I} (that is, {0, {0}}) is
called 2; and so on.
The union of any set of ordinals is the least upper bound of the set.
Let a be any set of ordinals, and P = U rx. If x is a non-empty subset
"ea
of p, then for at least one rxo E a we will have x n (xo non-empty. Since
rxo is well ordered, its subset x n rxo will therefore have a least element,
which is clearly also the least element of x. Since x was an arbitrary
non-empty subset of p, it follows that P is well ordered bye. Now suppose
that y E X and x E p. Then x E rx for some rx E a, so y E rx since rx is an
ordinal. We conclude that YEP, and this establishes that P is transitive.
It is thus an ordinal. Moreover rx P for any IX E a, so P is an upper
bound of a. And if IX Y for any IX E a, then IX c y for any IX E a, so P c y,
so P y. Thus P is the least upper bound of a.
THEOREM: Let rx, P be ordinals, and / an order-preserving isomorphism
from rx on to p. Then rx=P, and/is the identity map.
PROOF: Suppose / is not the identity, and let be the least member
of rx for which Then 11 E -+ /(11)=1], so c p. Being an
ordinal, is an initial segment of p; but it is not P itself, for in that case
/would map a proper initial segment of rx on to the whole of p, which is
not possible if/is an isomorphism. Thus E p.
Since / is order-preserving, 11 <e -+ /(1]) by the definition of
therefore, 11</(e) for every I] E e. Thus e or, equivalently,
e Equality is ruled out, so e </ (e). We now ask which element of
IX is mapped by Ito (which is in p, as proved above). It is easily seen
that there cannot be one, since x?:e and x<e
thus/is not an isomorphism from IX on to P.I
THEOREM: For each well-ordered system u=<a, r) there exists exactly one
(order-preserving) isomorphism/rom a on to an ordinal.
PROOF: (Uniqueness) Suppose/maps a one-one on to IX, and g does the
same on to p. Then go/ -1 is an isomorphism from IX on to p, and it can
only be the identity. So/ = g.
(Existence) Put b = {x E a I S.,(a) is isomorphic to an ordinal}. By
uniqueness no such Sx(a) can be isomorphic to more than one ordinal,
so let P(x) be the ordinal that Sx(a) is isomorphic to.
If YEb and x<y then x must be in b, and, indeed, P(x) <P(y); for
under the isomorphism from Sy(a) on to P(y), the proper initial segment
S,,(a) of Sy(a) becomes a proper initial segment of P(y), and therefore
an ordinal less than P(y), therefore P(x). Thus b is an initial segment of a.
By replacement we can form the set P of those ordinals P(x) where
x E b. Now P(x) implies that e is isomorphic to an initial segment of
S,,(a), say Sy(a) for some y x (so that = P(y )). It follows that P is an
initial segment of the ordinals, and therefore itself an ordinal.
Thus the map sending x to P(x) is an isomorphism from b on to p.
Our existence proof will be complete if we can show that b=a. And
this is simply done. For suppose b=l=a; since b is an initial segment of a,
it must be true for some Xo E a that b=Sxo{a). We have just proved that
b is isomorphic to an ordinal, so it follows from the definition of b that
Xo E b; but it is impossible that Xo E Sxo(a), by the very definition of
initial segment. Thus b = a after all. I
Consider now any well-ordering R(x, y) whose domain D is not a set;
we shall define a functional relation J which establishes an isomorphism
between D and On.
The functional relation J{x)=rt is defined by the formula 'rt is an
ordinalisomorphic to the segment Sx(R)'. This is a functional relation all
right, and its domain is D; for whatever x we pick in D there is one and
only one ordinal isomorphic to Sx{R). It is also an order-preserving
isomorphism. For if x <y (mod. R) then SxCR) is a proper initial segment
of S,,(R). But Sy(R) can be isomorphically mapped on to J{y), and the
image of Sx(R) under this mapping will form a proper initial segment
of J(y),which is thus an ordinal J(x) <J(y).
Furthermore, the range of J is an initial segment of On; for if P
J(x) then when J(x) is mapped on to Sx(R), P is mapped on to an initial
segment of the latter, S,,(R) say, for some ~ x Thus P=J(y). Since Jis
one-one it has an inverse J-
1
which maps the range of Jinto D. But D is
not a set, so by replacement the domain of J-
1
is not a set. Thus the
range of J is some initial segment of On; but not a set.
Under these circumstances it can only be On itself, which establishes
that J is a surjection.
Note that no other functional relation J' could be used to establish an
isomorphism between D and On. For whatever x in D we choose, J'(x)
is an ordinal isomorphic to Sx(R); only one ordinal has this property, and
this is J(x). Thus J'{x)=J(x) at every point xin D.
Proof by Induction on the Ordinals: Let E(x) be some formula containing
only the variable x free. If E( 01:) is to hold for every ordinal 01:, it is necessary
and sufficient that E(OI:) holds whenever E(P) holds for every P <oc; more
succinctly, that
VOl: [VP{P < rt -+ E(P) -+ E(rt)]
holds generally.
18 INTRODUCTION TO AXIOMA TIC SET THEORY
The proof falls out at once. If E(a.) is false for any ordinal at all, we
let a be the least such ordinal. Then E({3) is true whenever (3 <a, and so
by hypothesis E( a) is true, which is a contradiction.
This method of proof of V aE( a) is called proof by induction.
Definition by Recursion on the Ordinals: Suppose that F is a functional
relation and T a subclass of F's domain. That functional relation y=
=F(x) 1\ T(x) which restricts F to T is written F t T; in particular,
if a is a subset of the domain of F, then F ta represents a map, the re
striction of Fto a.
THEOREM: Let a be an ordinal,' A a class,' M the class of all maps defined
on ordinals less than a, and taking values in A; and H a functional relation
with domain M, and taking values in A. Then there exists one and only one
map f defined on afor which f ({3) = H(J t (3)Jor every (3 E a.
PROOF: (Uniqueness) This is routine. Suppose distinct functions f and 9
both did the job required, and let (3 be the least ordinal in a where
f({3)=f:g({3). Then f(y)=g(y) for y<{3, so f t {3=g t {3. But this means
thatf({3)= H(J t (3)=H(g t (3)=g(3), which is impossible.
(Existence) Let't' be the set of all {3 <a on which there is defined a map
fp satisfying fp(Y)=H(Jp t y)for every y<{3.
It is clear that 't' is an initial segment of a. By the uniqueness part of
the theorem, indeed, we actually have /P=fp' t {3 for (3<{3' E't'. Thus
1: is an ordinal ~ a and we can define a map/" on 't' as follows:/"({3)=
=H(/P), for every {3 E 1:.
Suppose y E (3 E 1:; then!.(y) = H(!;) = H(Jp t y)=fp(y). Thus/p=/. t {3.
Consequently,!.({3)=H(J. t (3) for all (3 E 1:.
But this means that if 1: < a then it satisfies the defining property of 1: ;
in other words that 't' E T, which is impossible. Thus T, which is an initial
segment of a, is actually equal to a, and so/. is the function we are looking
for
Let A be a class; M the class of all maps defined on ordinals, and taking
values in A; and H afunctional relation with domain M, and taking values
in A. We can now define a unique functional relation F with domain On,
and taking values in A, which for every ordinal a satisfies F(a) = H(F t a).
The functional relation y=F(a) is supplied by the formula 'there is a
map f,. defined on IX such that, for all fJ <IX, f,.(fJ) = H(j,. t fJ), and
y=H(f,.)'.
We have just proved that for each ordinal IX there is a unique map f,.
of this sort, and this establishes that the relation just defined is a functional
relation with domain On.
To prove that F has the right values, take fJ <IX. Thenfp = f,. t fJ simply
because the latter map has all the properties definingfp' Thus
F (fJ) = H (lp) = H (I,. t fJ) = I,. (fJ)
for all fJ <IX. Manifestly then, f,.=F t IX; but F(IX)=H(f,.), so F(IX)=
=H(F t IX).
That F is the only such functional relation is proved in the same way
as before. Suppose G is another, and let IX be the least ordinal where
F and G differ (if there is one). Since F(fJ) = G(fJ) for all fJ < IX we must
have F t IX=G t IX, so H(F t IX)=H(G t IX). Thus F(IX)=G(IX), a contra-
diction. It follows that F and G are equal for every ordinal IX.
THE AXIOM OF CHOICE
The axiom of choice (AC for short) is the following assertion:
If a is a set 01 pairwise disjoint but non-empty sets, then there is a set
whose intersections with elements of a are always singletons.
Spelt out more fully this becomes
'v'a{['v'x(xea -+ x #:0)
A 'v'x'v'y(x e a Aye a -+ (x = y v x n y = 0)]
-+ 3b'v'x3u(x e a -+ b n x = {u})}.
Given the other axioms the following statements are equivalent to the
axiom of choice.
ACt: For any set a there is afunction h mapping the set of all non-empty
subsets of a into a in such a way that hex) e x for all non-empty
xc a.
A C": The Cartesian product of a family of non-empty sets is itself non-
empty.
ACt -+ AC: Let b= U a. We apply ACt to b. Every element x of a,
being non-empty, is a non-empty subset of b. So take the function h that
ACt guarantees, and consider {hex) I x e a}. This set has exactly one ele-
ment from each x e a, and is the set needed to establish A C.
AC --+ AC": Let (ai)iEI be a family of non-empty sets, and putb
i
= {i} x ai.
Then the family (bi)iEI consists throughout of pairwise disjoint non-empty
sets. Let be a set with just one element in each bi(i E I). Then E X ai'
isI
by the definition ofX.
AC" --+ AC': Take any set a and form the product X x. By AC" this
xCa
x",0
is not empty, so choose an element <po It is easily checked that <p(x) E x
for every non-empty subset x of a.
A further equivalent of the axiom of choice is the following statement.
ZERMELO'S THEOREM: Every set can be well ordered.
PROOF: AC' follows at once from this. For take some r c a
2
that well-
orders a and define a map h: &(a)".{0} --+ a by putting h(x)= the least
element of x (mod. r).
Conversely, suppose that AC' is true and that the map
h:&(a)". {0} --+ a
satisfies the requirement that hex) E x for every non-empty x c a. Define
a second map
g:&(a)". {a} --+ a
by putting g(x)=h(a".x). Then this satisfies the requirement that
g(x):xforanyproperx c a.
Let () be some object that is not an element of a (since the class of all
sets is not itself a set, there is no difficulty about finding such a 0). We
define by recursion a functional relation F: On --+ a u {O} by decreeing
I
g({F(f3) I f3 < IX}) if {F(f3) I f3 < IX} C a
F(IX) = (that is, if it is in the domain of g); '"
o otherwise.
Suppose that Fnever takes the value 0; in other words, that F(IX) E a for
every ordinal IX. Then {F(f3) I f3<IX} is in g's domain, whatever ordinal
IX is. But F(IX) cannot equal any F(f3) for f3 <IX, because
F (X) = g ( { F (f3) I 13 < (X}) {F (13) I 13 < (X}
by the definition of g. Thus F is an injection; On is its domain, and all
its values are in a. Inverting F, therefore, and using replacement, we can
infer that On is a set. Since this is false we must conclude that F(IX) is not
alwaysina.
Thus F(IX)=O for some ordinal IX, and we may suppose IXo to be the
least such. In this case F(P) e a for every P <IXo, and so {F(fJ) I fJ <IXo} is
a subset of a; yet it is not in the domain of g (since F( IX 0) = 0). By the
definition of g, {F(P) I fJ <IXo} can then only be a itself; so the range of
F t OC
o
is a.
But F t IXo is also an injection; for if P <oco, F(fJ)=g({F(y) I y <P})
{F(y) I y <fJ}, so that (as before) F(fJ):f:F(y) when y <po
Thus F t IXo is a bijection from the ordinal IXo on to a, and conse-
quently a can be well ordered .
Note that what we have just proved can be stated as follows: whether
or not AC holds, a set can be well ordered if and only if there is a map h:
BP(a)"'-{0} -+ afor which hex) E xfor every non-empty x c a.
DEFINITION: If R is an ordering and T a subclass of the domain D of R,
then an object x of D is called an upper bound (respectively, a strict upper
bound) of Tif, for every yeT, we have ~ y (respectively, x>y). If Thas
a strict upper bound we call it dominated. If the class of upper bounds of
T has a smallest element under R, then this element is called the least
upper bound of T.
An object x in D is called a maximal element of D if there exists no
greater element of D (under the ordering R).
We now prove yet another equivalent of the axiom of choice.
ZoRN'S LEMMA: Let u be an ordered system, every well-ordered subset of
which has an upper bound. Then u has a maximal element.
PROOF: Put u=(a, r) where r c a
Z
is an ordering. Assuming the axiom
of choice in the form AC' we have a map h: BP(a)"'-{0} -+ a for which
h (x) E X for every non-empty x c a.
Let c be the set of all dominated subsets of a (subsets with strict upper
bounds). We define amapmfrom cinto a by putting
m (x) = h (the set of strict upper bounds of x) .
Then for x E c, m(x) itselfis a strict upper bound of x, and so m(x) cannot
be a member of x.
Choose again some 0 which is not a member of a. As in the proof of
Zermelo's theorem, we can define a functional relation F: On a u {O}
as follows:
1
m {F (P) I fi < a} if { F (P) I P < a} E C ;
F(a) = 0 otherwise (that is, if {F(P) I P < a}
has no strict upper bound or is not included in a).
Again we show that F must take the value 0 at least once. For, were it
otherwise, andF(a) E a for every ordinal a, we would have {F(P) I P <a} EC
for every ordinal a. Moreover, since F(a) dominates {F(P) I fi<a}, it
cannot belong to this set, and so for p<a, we get FCP)#F(a). Thus Fis
one-one; once more the scheme of replacement applied to F-
1
would
deliver On as a set, and so we have a contradiction.
So let a
o
be the least ordinal for which F(ao)=O. Any smaller ordinal
a will satisfy F(a) E a; and so {F(P) I P <a} E c; thus F(a) will be a strict
upper bound to {F(P) I p<a}, and p<a<ao F(p)<F(a) (mod.r). It
follows that F tao, which maps a
o
one-one into a, is also order-
preserving. Consequently {F(P) I P <ao} is going to be a well-ordered
subset of a, and so by hypothesis it has an upper bound, d say. But it has
no strict upper bound, for the existence of such an object would put
{FCP) I P <ao} in c; and this would mean that F(ao) E a, contrary to
supposition.
Thus there is no element of a that is greater than d (under the ordering
r); and so d is the desired maximal element of a.1
Conversely, we can show that a seemingly weaker statement than Zorn's
lemma is sufficient to entail the axiom of choice.
CONVERSE: Suppose that there is a maximal element in every ordered system
whose linearly ordered subsets all have least upper bounds. Then the axiom
of choice is true.
PROOF: Let a be a set of non-empty pairwise disjoint sets; and let b = U a.
Put X equal to the set of all subsets of b which do not have more than one
element in common with any x E a; then X is ordered by set inclusion.
Suppose Y is any linearly ordered subset of X. We shall show that Y
has a least upper bound in X. Now the least upper bound of Y (under
the c-ordering) is U y: so we will show that U y E X.
yeY yeY
For any x in a, then, the set x n U y must be proved to have fewer
yeY
than two elements. There are two cases to be considered. In the first,
x n y = 0 for every ye Y; here we have at once x n U y=0, and nothing
yeY
remains to be proved. In the second, x n y is a singleton for one or more
y e Y. But however many y e Y yield a singleton when intersected with x,
it will always be the same singleton. For suppose x n Yo ={u} and x n Y1 =
={v}. Then since Y is linearly ordered, Yo u Yl is one or the other of
Yo, Yb which means that x n (Yo u Yl) = {u} u {v} is a singleton; and this
means that u = v. It follows at once from all this that, in this second case,
x n U y is itself a singleton, so that U Y e X, as was required.
yeY yeY
By supposition, then, X has a maximal element Yo. We claim that
Yo n x has exactly one element for each x e a. Were it otherwise, indeed,
Yo n x would have to have no elements for some x e a (since Yo e X). In
this case we could take any element e of x and form the set Yo u {e}. This,
however, would equally be in X, and would destroy Yo's maximality. I
CARDINALS
(In this section the axiom of choice will be used.)
Two sets a and b are termed equipollent (or equinumerous, or sometimes
just equivalent) if there is a one-one map (a bijection) from a on to b.
It is to be observed that two sets a, b can be equipollent in an intuitive
sense without being so according to this definition; for if A, B are the
parts of the universe corresponding to a, b, there may be a bijection, in-
tuitively speaking, from A on to B. But unless this bijection itself corre-
sponds to an object of the universe there is no reason to suppose that a
and b are therefore equipollent.
It is obvious enough that the relation 'x is equipollent to y' is an
equivalence relation whose domain is the class of all sets. With the axiom
of choice, however, we can say more than this; namely that every set is
equipollent to an ordinal. For if a set can be well ordered, it is certainly
equipollent to the ordinal of its well-ordering.
The least ordinal equipollent to a set a, written ii, or sometimes
card(a), is called the cardinal of the set a. It is trivial that a and b are
equipollent if and only if ii = r;.
We write Card for the class of all cardinals. It is defined by the formula
Card(rt.): 'rt. is an ordinal not equipollent to any smaller ordinal'.
LEMMA: Let rt. be an ordinal and ~ a subset of it. Then the ordinal of the
well-ordering induced on ~ by the ordering of rt. is less than or equal to rt..
PROOF: Let P be the ordinal in question, and f: ~ -+ P the isomorphism
between ~ and p. We show that for y ~ f(y)t;;;;,y, which immediately
gives pert., and so p t;;;;, rt. as required.
Suppose, on the contrary, thatf(yy for some y ~ . There will be
a least such, Yo. For y E ~ we have then
y < Yo -+ fey) t;;;;, y < Yo;
and also
y ~ Yo -+ fey) ~ f(yo) > Yo
As in an earlier theorem,! clearly misses out on Yo; but since it does take
greater ordinals as values (for example, f(yoYo), it cannot after all
have an ordinal for its range; and this contradicts the hypothesis .
THEOREM: Let a and b be non-empty sets. Then thefollowing conditions are
equivalent.
(1) There is a one-one map from a into b.
(2) There is a map from b on to a.
(3) Zit;;;;,b.
PROOF: (1) -+ (2 Let j be an injection from a into b. We define a
surjection s in the opposite direction as follows. Take any Xo in a, and
let s(y) be Xo for any y in b that is not in the range ofj; and for elements
y of b that are in the range ofj we can take s(y) as the inverse image of
y underj, sincejis one-one on its range.
(2) -+ (1 Conversely, suppose s: b -+ a is a surjection. AC gives us
a map h: .9(b)""-{0} -+ b such that hey) E Yfor every non-empty Y c b.
So for x E awe putj(x)=h({y E b I s(y)=x}). This is manifestly one-one.
(3) -+ (1 IfZit;;;;,b we have two bijectionsf: a -+ a and g: b -+ b; and,
sincea c b, anidentityinjectioni:a -+ b. Composing these into g-l 0 i oJ,
we produce what is clearly an injection from a into b.
(1) -+ (3 Supposej: a -+ b is an injection. Then, withfand g defined
as above, the map g 0 j 0 f -1 is an injection k from a into b. The range
of k is some subset e c jj which, according to the previous lemma, is
ORDINALS, CARDINALS
25
isomorphic to some ordinal Thus a is equipollent with p, and since
it is a cardinal, so a
CoROLLARY (CANTOR/BERNSTEIN THEOREM): For two sets a, b to be
equipollent it is necessary and sufficient that there be injections from each
one into the other.
PROOF: For given the two injections we get at once a band b a, and
thusa=b.1
(Although the axiom of choice was used in this theorem, and thereby
in the proof of the corollary, it could, in the latter case, have been
dispensed with.)
THEoREM (CANTOR): Foreveryseta,a<flJ(a).
PROOF: Suppose that then there exists a map h from a on to
flJ(a). Let b= {x E a I x 1: hex)}. Then b c: a, so there is ace a such that
h(c)=b. But c E b is equivalent to c : h(c), and this is equivalent to c : b;
and this is a contradiction.1
Note that in proving this we have also proved the following theorem.
For neither it nor the next one do we need rely on AC.
THEOREM: For no set a does there exist a map from a on to flJ(a); and, a
fortiori, no injection offlJ(a) into a.
THEOREM: The class of cardinals is not a set.
PROOF: Suppose it were, x for example. Then x would be a set of ordinals
with a least upper bound A = U IX. Since every ordinal is equipollent
"'ex
with some cardinal or another, every ordinal is equipollent with some
subset of A. Now let Ybe the class of all well-orderings defined on subsets
of A. Every r E Y is a subset of A
2
; so Y( r) is equivalent to the formula
r E flJ(12) A Y( r), which establishes that Y is a set (by the scheme of
comprehension).
But every r in Y is associated with a unique ordinal (namely, the ordinal
of the well-ordering r). Consequently, we can write down a functional
relation performing this association, and the range of this functional
relation is On. Since Y is a set, this contradicts the replacement scheme. I
Let a, b be two disjoint sets, and a', h' two more, respectively equipollent
with a and b. Then a u b and a' u b' are equipollent too; for from the
first to the second there is a bijection whose restrictions to a and bare
just the given bijections from a to a' and from b to b'.
So let ct, /3 be cardinals. There certainly exist disjoint sets a, b with these
cardinalities, for we can, if we like, take ct x {O} for a and /3 x {I} for b.
Whatever disjoint a and b we actually choose, we define the cardinal sum
ct + /3 as the cardinal of a u b; it is easily seen that ct + /3 is quite independ-
ent ofthe choice of a and b.
Amongst the properties easily proved to hold for cardinal addition are
commutativity (ct + /3 = /3 + ct) and associativity (ct + (/3 + ')') = (ct + /3) + I' ).
To get at cardinal multiplication we can drop the disjointness conditions
on a, band a', b' that were imposed above. Iff: a -+ a' and g: b -+ b' are
some appropriate bijections, then we can define a bijection h: a x b -+ a' x b'
by writing hx,y=(f(x),g(y for all xea and yeb; thus these
Cartesian products are equipollent.
So if ct, /3 are cardinals we define the (cardinal) product of ct and /3,
written for brevity ct/3, or ct /3, as the cardinal of ct x /3. It should be clear
that if a has cardinal ct and b has cardinal /3, then a x b has cardinal ct/3.
If ct, /3, I' are all cardinals we can easily check that the commutative
(ct/3 = /3ct) and associative (ct (/3')' ) = (ct/3) I' ) laws of multiplication hold; there
is no more difficulty involved in checking the distributive law (ct(/3 + ')') =
=ct/3+ct,),); we simply pick three sets, a, b, c, the latter two disjoint, and
observethatax(b u c)=(axb) u (axe).
More generally, let (ai)ieI> (bi)ieI be families of sets each indexed by I.
Suppose that for every i e I we have lii = 1)1' Then X ai and X b
i
are
ieI ieI
equipollent. This is proved by noting that for any i e I the set Bi of
bijections from a
i
on to b
i
is non-empty; by AC", therefore, the product
of the family (Bi)ieI is not empty. So there is a family (CPi)ieI of maps, each
CPi being a bijection from ai on to bi' We now define the required bijection
fromXaj on to Xb
i
by associating each element (Xi)ieI of the former set
ieI ieI
with (CPi(Xi)ieI in the latter.
If (ct
i
); e I is a family of cardinals we call the cardinal of the set X cti the
ieI
product of the family, and write it n ct
i
. It is easily checked that this cardinal
ie I
is the cardinal of the Cartesian product of any family (ai)ieI indexed by I
and satisfying lii = cti for all i E I.
Similar considerations for addition lead us to the following. Let (al)ieI
and (bi)ieI be families as above which satisfy the further condition that
i:-j ai II aJ=b
i
II b
J
=0. Then Uai and Ubi are equipollent. As
ieI ieI
before, we prove this with the axiom of choice, which supplies us with
a family (lPi)leI of maps, each lPi being a bijection of a
l
on to bi' Then if
x e U a
i
it can only be in one of the a" say ai (x), and so we can couple it
ieI
with lPi (x) (x) in U b" thus establishing the desired correspondence between
leI
Ualand Ubi'
tel ieI
Given any family of cardinals (lXi)ieI we can construct a family of
disjoint sets (ai)ieI which are equipollent to them for every i e I, by
writing ai = IX; x {i}. The cardinal of Ua" which is quite independent of
leI
the choice of the disjoint family, is called the sum of the family (lXi)lel>
and written LOCI'
leI _
To define exponentiation of cardinals, let a = {l and b = 5'. Then
ba (the set of maps from b into a) is equipollent with b'a'. For supposef:
a a' and g: b b' are appropriate bijections. Then to each IP e ba we
can associate the map l/I e b' a' defined by
l/I = f 0 IP 0 g-l .
This clearly establishes a bijection between the two sets of maps.
So if 0(, P are cardinals, the cardinality of the set Poc is called the pth
power of 0(, and written oc
p
If a, b have cardinalities oc, p respectively, it is
easily checked that ba has cardinality oc
p

The usual laws of exponentiation for cardinals, that O(P+Y=oc
p
O(Y and
that (O(P) Y = O(Py are easily established from this definition.
FINITE ORDINALS
(This section does not use the axiom of choice.)
We recall that the next greatest ordinal after a given ordinal 0( is its
successor 0( U {O(} (or IX + 1 for short; but this + is different from the + of
cardinal addition above). If p is the successor of 0( we say that 0( is the
predecessor of p.
An ordinal oc is called finite if every p 0( (apart from 0) has a pre
decessor. The formula '0( is a finite ordinal' is easily expanded into
On(a) A Vp[On(f3) A 13 c: a A 13#0 3y(p=y u {y})J. Finite ordinals
are also known as natural numbers. It should be obvious that if a is a
finite ordinal and 13 a, then 13 is finite; and that if a is finite, so is a + 1.
Mathematical Induction: Let P be a class such that P(O) is true, and for
every finite ordinal a, Pea) P(a+ 1) is also true. Then Pea) is true for
every finite ordinal a.
For if not, take the smallest finite ordinal ao not in P. Since 0 is in P,
ao # O. Thus it has a predecessor 13 o. As ao was the smallest finite ordinal
not in P, we must have P(Po). We are given that P(Po) P(Po+ 1). So
P( ao), which is impossible.
It is worth pointing out again that the definitions above of the words
'finite' and 'natural number' are not intended to provide for them their
usual senses. In future, therefore, everyday uses of the words will be
explicitly recorded as such.
It is easily seen that what we would intuitively say was an ordinal a
with a finite number of elements is indeed finite according to the
above definition. But it could happen that a finite ordinal 13 had, intui-
tively speaking, infinitely many elements; in this case since 13 c: a is ruled
out, 13 would be greater than a. In addition, the part of the universe made
up of these finite ordinals (in an intuitive sense) could not then be a class;
for supposing P(x) to define the class, we would have -,P(f3), and so
could prove the existence of a least ordinal Po for which -,P(f3o). Then
if Po were equal to Yo + 1, we would have P(Yo), showing that, intuitively
speaking, Yo had finitely many elements; but with f3o=Yo u {Yo} this
would hardly be possible.
THEOREM: Every finite ordinal is a cardinal.
PROOF: This is proved by induction. 0 is obviously a cardinal. So suppose
that a is a finite ordinal which is a cardinal, whilst a + 1 = a u {a} is not.
Then there must exist an ordinal y < a + 1 and a bijection f from a + 1 on
to y. As a+ 1 #0, y cannot be 0; thus it is 13 u {f3} for some 13, and since
y<a+l, wegetp<a.
Weref(a) equal to pJ t a would map a one-one on to 13, which would
contradict the fact that a is a cardinal. Sof(a)= # 13, and asfis bijective
f(17)=p for some 17 E a+ 1; 17#a, so 1J E a. We now define a map g which
does map Ct one-one on to p, by setting g(x)=f(x) for all x in Ct except
for 1'[; and Since Ct is a cardinal, we have a contradiction .
We call a cardinal finite if it is a finite ordinal. Let us note the following
consequence of the theorem above. If Ct is a finite ordinal, then its suc-
cessor Ct + I is indeed the cardinal sum of oc and I. Thus within the domain
offinite ordinals our notation is unambiguous.
THEOREM: ljCt, P are finite cardinals, so too are oc+ p, ocP, oc
p

PROOF: Each proof is done by induction on p. It is trivial that if oc is finite
and Pis 0 then oc+P is finite. Moreover, since oc+(P+1)=(oc+P)+ I, a
finite oc + P means a finite Ct + (P + I). This proves the first part.
The proof for multiplication is similar. Since oc(P + 1) = ocp + oc, the
assumption that ocp is finite when oc, P are, together with the previous
result, yields at once that oc (P + 1) is finite. Again, the identity oc
P
+ 1 = oc
P
oc
supplies all that is needed to establish that oc
P
is finite .
We are now in a position to state the one remaining axiom of Zermelo/
Fraenkel set theory.
5. AXIOM OF INFINITY
This axiom says, in effect, that there is an ordinal which is notfinite.
On the strength of the axiom we write 0) (or, in some cases, N) for the
first ordinal which is not finite. Thus 0) is the set of finite ordinals. For
if oc is a finite ordinal we do not have 0):::;; oc (which would make 0) finite);
so, necessarily, oc <0) or, equivalently, oc E 0). Conversely, if YEO) we
have Y < 0), and so y is finite, by the definition of 0).
Thus the axiom of infinity can be stated in the form: the class of finite
ordinals is a set. For the class in question is an initial segment of On and
so, if a set, is an ordinal 0); but it is not a finite one, since no ordinal
belongs to itself, and in particular 0) w.
An ordinal oc#O without a predecessor is called a limit ordinal; it is
thus a non-zero ordinal for which P <oc -+ P + 1 <oc. Alternatively, a
non-zero ordinal oc is a limit if and only if Ct = sup P = U p.
P<rz P<rz
For if oc is not a limit it is y+ 1, which means that sup P=y. If it is a
P<rt
limit, put sup p=y; since y:::;;oc anyway, suppose y<Ct. By the definition
P<rz
of limit, then, y + 1 < ex, and so y is not the supposed supremum after all,
since y+ 1 :$')'.
So another way of formulating the axiom of infinity is: there is a limit
ordinal.
For, given the axiom, 00 could only have a predecessor if that predecessor
were finite; which would make OJ finite too, directly contrary to its defini-
tion. Since a limit ordinal is necessarily not finite, the converse holds too.
A fourth way of expressing the axiom of infinity is by
3x[Oex A'v'y(yex-+yu{y}ex)].
Given the axiom we can take 00 itselffor x. Conversely, suppose the above
sentence holds, and choose any a satisfying it. Let ex be the least ordinal
not in a. It cannot be 0, and if ex=P u {P} we get P <a; thus P E ex, so by
the characterization of a, p u {P} E a, a contradiction. Thus ex is not a
successor either, so it must be a limit, which is all we need to prove.
INFINITE SETS AND INFINITE CARDINALS
(Unless otherwise stated, the axiom of choice will be
allowed in this section.)
A set a is called finite if its cardinal is finite, infinite otherwise; if a OJ,
we say that a is denumerable.
Every infinite set includes an infinite denumerable subset (one, that is,
equipollent with OJ)
For if a is infinite and / a bijection of a on to a, the range of/ trois
a subset of a equipollent to OJ.
THEOREM: A set a is infinite iff it is equipollent with one 0/ its proper subsets.
PROOF: If a is finite, and b a proper subset, take x E a"-.b. The cardinality
of a is certainly not zero, but it is finite, so of the form ex+ 1 =ex u {ex}.
There is no difficulty in finding some bijection / from a on to ex u {ex}
which maps x to ex; in this case/ t b maps b one-one into ex. Consequently
1j ex, so 1j < a, and b is not equipollent with a.
If a is infinite, a-;:,OJ. Writing/for some bijection from a on to a, we
define an injection g from a into itself by setting
g(x) = x, if f(x) -;:. 00;
g (x) = y, if f(x) < OJ and f(y) = f(x) + 1.
The range of g is a proper subset of a since it does not contain f -1 (0);
thus g is a bijection from a on to a proper subset .
We will write Card' for the class of infinite cardinals; it is a subclass of
On, but not a set (for then Card would be too, since OJ, the class of finite
cardinals, is a set). So Card' is just a well-ordered class, and therefore
there must be a functional relation y = establishing an isomorphism
between On and Card'. For simplicity we write the infinite cardinal
as the relation y= then abbreviates 'y is an infinite cardinal,
and the set of infinite cardinals that are less than y is isomorphic, as a
well-ordered set, to IX'.
We have =OJ; and, for each IX, is the first cardinal >
IflXisalimitordinal, U
1/<1%
To show that U is a cardinal, suppose that y < U so
1/<11. I/<rz
y E U so Y E for some {3 <IX. Thus Y c giving
P<rz
C U
1/<11.
since IX is a limit ordinal.
This proves that U > y and that there is no bijection from U
I/<rz 1/<11.
on to any smaller ordinal y.
Since this cardinal U for all {3 <ct, it is also But con-
1/<11.
versely c for all {3 <IX, so U c and this proves the result.
I/<rz
By Cantor's theorem we have since is the cardinal of

The continuum hypothesis (CH) is the sentence alternatively,
&( OJ) can be so well ordered that every strict initial segment is denumerable.
To show that this second formulation entails the first (the converse is
clear), note that the ordinal of such a well-ordering could not in any case
be denumerable, since, by Cantor's theorem, &(OJ) is not. By the speci-
fication of the well-ordering, it follows that the ordinal in question is
the first non-denumerable ordinal, that is
The generalized continuum hypothesis (GCH) is the sentence: 2
N
= 1
for every ordinal ex.
Write On
2
for the class of all ordered pairs of ordinals (ex, P). On this
class we can define a well-ordering relation R as follows:
(ex, P):;;;' (y, ( (mod. R) if and only if
sup (ex, P) < sup(y, (5)
v sup (ex, P) = sup(y, (5) 1\ ex < y
v sup (ex, P) = sup(y, (5) 1\ ex = y 1\ P :;;;, ().
Every segment is seen to be a set. For if is an initial segment,
and e = (ao, Po), and (ex, P) < (ao, Po) (mod. R), then
sup (a, P):;;;, sup (exo, Po).
Putting Yo = sup (exo, Po), we note that every pair (a, P) in is a
member of the set (Yo+1)2; so is a subclass of that set, and thus
a set itself.
On the other hand, take a non-empty set X all of whose elements are
in On
2
; the set of all sup (a, P) for (a, P) E X has a smallest element Yo;
the set of all ex such that for some P we have both (ex, P) E X and
sup(a, P)=Yo has a smallest element exo; and the set of P for which
sup(ex
o
, P)=Yo and (exo, P) E Xhas a least element Po.
From all this it drops out that (ao, Po) is the least element of X (mod.
R); and so R is a well-ordering.
Since On
2
is therefore a well-ordered class, but not a set, a relation
y=J(a, P) can be found which establishes an isomorphism between On
2
and On. In the definition of J, be it noted, there is no need for the axiom
of choice.
THEOREM: For each ordinal ex, = ...
PROOF: If the theorem is false, it fails first at some ordinal p. To obtain
a contradiction we need only show that since there is no
trouble with Restricting J to x we get a one-one map, and
so the proof reduces to showing that for a, P E J(a, P) E
Now J(a, P) E +-+ J(a, P) so, being a cardinal,
J(ex, P) E +-+ I (a, P) <
So if we prove that Jea, a, P E we are finished.
Since J(a, P) is the set of ordinals less than J(a, P), it is the range of J
restricted to the set of all (ex', P') less than (a, P) (mod. R). Writing y
for sup ( ct, P), we easily see that this latter set is included in (y + 1)2 (since
(ct' , P') ct, P) ct' <y+ 1 1\ P' <y+ 1). Thus its cardinal cannot
exceed the cardinal of(y+l)2; but as so is y+l, and therefore
by the definition of p there are just two cases: y + 1 is finite, or
(y+ 1)2=y+ 1
EachcaseseparatelyyieldsJ(ct, and so the theorem is proved .
COROLLARY: every natural number n 1, and every ordinal ct.
PROOF: Immediate, by induction on n .
COROLLARY: ..
PROOF: then clearly
= =
and so = as required .
COROLLARY: Let (K1)leI be a family of cardinals, all non-zero. If i, or any
Kj , is infinite, then l: K1=SUp(i, SUPKJ
leI leI
PROOF; Recall that L KI is defined as the cardinal of the set X = U K/ X {i}.
leI leI
Since each we have that X:::> U {OJ x {i}, that is that X includes
leI
{OJ X I, which is equipollent with I. Thus X i. It is clear that X Kj for
every i E I, and SUpKj ).
ieI
For the reverse inequality, put K = sup K
i
Then X c: U K x {i} = K x I.
iel ieI
Thus X j which, since either K or i is infinite, is equal to sup(i, K)
An isomorphism between the class of finite sequences of ordinals and On:
(To establish this we do not need the axiom of choice.)
For a class C, afinite sequence s of objects of C is, by definition, a map
whose domain is a natural number, and whose values are in C. This
natural number, the domain, is otherwise called the length of the sequence,
and is written I(s). We write (i(C) for the class of all finite sequences of
objects of C.
An ordering R can be defined on (i(On), the class of finite sequences of
ordinals, by setting
s < t (mod. R) if and only if
sup{s) < sup{t) v
sup{s) = sup{t) A I{s) < let) v
sup{s) = sup{t) A l{s) = l{t) A s:f= t A
sen) < ten) for the first n where
s(n):f= ten).
(Here sup(s) stands for the greatest value taken by the function s.)
We show that R, defined in this way, is a well-ordering. This involves, in
the first place, that the class S,(R) of sequences <t (mod. R) be a set; but
since everything in this class is also in the set q(y+ 1) for y=sup(t), this is
easily proved.
So let X be a non-empty set of finite sequences of ordinals. As s ranges
over X, sup(s) takes on various ordinals as values, and of these there
must be a least, y say. Amongst the sequences s in X for which sup(s)=y,
there will be at least one of minimal length, of length n say. Now define
recursively a map I with domain n as follows: if i E n, I (i) = the smallest
ordinal of the form sCi) for SEX satisfying sup(s)=y, l(s)=n, and
s(j) = I (j) for every j < i.
Now f, as thus defined, is a member of X; and, further, sup(/)=y and
l(f)=n. For it is immediate, by induction on i, that for every ~ n
there is an SEX for which sup(s)=y, l(s)=n, and s t i=1 t i; whence,
putting i=n, the result. Moreover,fis the smallest element of X. For take
some s in X, distinct fromj. If sup (s):f= y we must have sup(sy = sup (f),
by the definition of y. Thus s> I. If sup(s)=y, however, and l(s):f=n,
we have in the same way l(sn=l(f); and so s> I. Finally, if sup(s)=y
and l(s)=n, let i <n be the first integer where s(i):f= I(i). By the definition
off, sCi) > I (i), so once more s > I.
But a well-ordered class that is not a set is isomorphic to On; and there
is consequently a functional relation s=J(rx.) which establishes an iso-
morphism between On and q (On).
CHAPTER III
THE AXIOM OF FOUNDATION
The axiom offoundation (AFfor short) is the sentence
A ynx=0)];
in other words, the axiom states that every non-empty set has an element
which is disjoint from it.
From the axiom it follows that there is no infinite sequence (that is,
no map with domain (1)) (un)ne;ro for which U
n
+l E Un for all n E (1); for if x
were the set of terms of such an infinite descending sequence (its range,
in other words), it would fail to contain an element disjoint from itself.
In particular, therefore, the axiom implies that for every x, x ; x; and,
more generally, that there are no E-cycles - finite sequences xo, ... , X
n
-l
thatis,satisfyingxo EXl E ... EX
n
-l E Xo. For, again, the set {xo, ... , X
n
-l}
would contradict AF.
We can take advantage of the axiom to provide a more simple definition
of ordinals than the one given earlier. This is established in the following
theorem.
THEOREM: For X to be an ordinal it is necessary and sufficient that the two
conditions
'v'U'v'v [u E X A V E X (u E V V U = v v V E u)]
and
'v'U[UEX U c X]
be satisfied.
PROOF: The necessity of both conditions is immediate. Conversely, we
note first that E imposes a strict linear order on any X satisfying them.
For by AF we cannot have U E v and v E U together; moreover, when U
v, WE X, and U E v and v E w, the possibility U= W is ruled out (for then
U E v and v E u), and the possibility W E U is ruled out (for then we get the
cycle U EVE W E u); and so by the first condition all that is left is U E W.
To show that this ordering is a well-ordering, we take any non-empty
Y c X; by AF there is some U E Y satisfying un Y=0. But this means
that if v e Y then v u, and therefore that u is the desired least element
(mod. e) of Y.I
Without necessarily supposing the axiom of foundation to hold in the
universe IYI, we can define by recursion a functional relation y= VIZ by
setting
V
o
=0;
VIZ = U &,(Vp).
P<IZ
It is clear enough from this that o t ~ o t ~ VIZ C V"'. Consequently, since
Voc+l = U &,(Vp), we have V,,+l =&,(V,,)jor every ordinal a.
p",,,
On the other hand, for limit a we have p < a. ... p + 1 < a., so
V,,= U &'(Vp)= U Vp+
1
= U Vpforeverylimitordinala..
P<IZ p<" p<"
The class of all elements of some V" or another is called V; it is defined
by the formula V(x): 3a.(On(ot) 1\ xe V,,).
For any x in V we define the rank of x, rn(x), as the least ordinal a for
which x e V". Since for limit a we have x e V" ... x e U Vp, rn(x) is
always a successor, of the form P+l. p<"
LEMMA: A set a is in V iff aU its elements are,' ifrn(a)=a, then every element
of a has rank < a.
PROOF: Suppose a is in V, and its rank is a=p+1. Then ae V
p
+
1
, so
a c Vp. Thus all a's elements are in V; moreover, their ranks are all
~ p , and so <a. Conversely, suppose that every element of x is in V. The
ranks of these elements form a set, and so must be bounded above by
some ordinal a. It follows that a c V"' whence a e V" + 1; thus a is in V. I
THEOREM: Every ordinal a is in V, and the rank of a is a + 1.
PROOF: If there is one, let a. be the first ordinal for which a V,,+ 1; then
pea ... pe V
p
+
1
,andso
a c U Vp+1 = U &'(Vp) = V"'
p<" p<"
from which a e V,,+1' contrary to hypothesis.
Likewise, let a be the first ordinal, if there is one, such that a e V". Then
a e U &,(Vp), so a is in &,(Vp) for some p <a. But then a c Vp, and since
p<"
pea we obtain p e VP' in contradiction to the definition of a.1
THE AXIOM OF FOUNDA TION
37
THEOREM: The axiom of foundation holds if and only if V is the whole
universe; or, briefly, AF +-+ Vx Vex).
PROOF: Supposing VxV(x) to hold, we pick a non-empty set a, and from
its elements pick some set b of minimum rank !Y.. Then b's elements all
have rank <!Y., and so cannot themselves be in a. Thus b () a=0, as
required for AF.
The converse is most easily proved by defining the transitive closure
ofa set as follows. Recall that a set Xis transitive if'v'xVy[ x E X /\ Y E X --+
--+ Y E Xl The transitive closure of X is defined as the most restricted (that
is, smallest under the order c) set ~ X ) which is transitive and includes
X. That this definition is a correct one is established by the theorem which
we interpose here.
THEOREM: For any set X there exists a unique transitive set Y which includes
X and is itself included in every transitive set including X.
PROOF: We define a mapfon ro by recursion:f(O)=X;f(n+ 1)= U x.
x ef<n)
Let Y= U f(n). Then it is obvious that Xc Y; and further, Y is
new
transitive. For if x E Y, then x ef(n) for some nero, so xc f(n+ 1), so
x c Y. Now if Z is a transitive superset of X, we prove by induction that
every f(n)c Z; n=O is immediate; if x e fen) c z, x e Z, so x c Z,
becauseZistransitive;thereforef(n+l)= U xc Zasrequired.
xef(n)
This completes the inductive step, and we can conclude that
Y= U f(n) c Z .
new
Returning to the stranded theorem, suppose that AF holds, but that
some object a is not in V. Let b be a transitive set including a, and b' the
set of elements of b that are not in V.
Then b' is non-empty; for a is not in V, so has an element Xo not in V;
and since a c b, Xo is also in b' .
But for every x E b' we also have b' () x non-empty. For if x eb' , it is
not in V, so has an element y not in V. Since y e x e b, and b is transitive,
y e b; and thus y E b' , by the definition of b' . But then y e b' () x, which
therefore cannot be empty.
Thus b' contradicts AF, and the theorem is established .
Let us note one property oftransitive closures.
THEOREM: u U
yeX
PROOF: certainly includes X. If y E X then y c and as
is included in every transitive set including y, c Thus
Xu U c Conversely, if Z=Xu U then Xc: Z,
yeX yeX
and, moreover, Z is transitive. For if v E u E Z, then either u is in some
(y) with y E X, in which case v E (y); or u E X, in which case
v E U c Either way, v E U so V E Z, as desired. Thus
yeX
Xc Z, and Z is transitive; so 'tf(X) c Z, and the proof is complete. I
Well-foundedRelationsandExtensionaIRelations: Let r be a binary relation
on a set a (so r c a
2
); r is said to be well founded on a if for every non-
empty X c a there exists an x E X such that <y, x) E r holds for no
y E X. It should be clear that if b c a, the restriction r (') b
2
of r to b is
well founded on b.
A map q> with domain a is called collapsing for r if for all x E a we have
q>(X) = {q>(y) lYE a 1\ <y, x) E r}.
Evidently, if q> is collapsing the range of q> is a transitive set, since each
element of q>(x), for x E a, is of the form q>(y).
THEOREM: Ifr is a well-founded relation on a there exists one and only one
map with domain a which collapses for r. Moreover, the range of this map
is in the class V.
PROOF: (Uniqueness) Let q>, tjJ be two such collapsing maps, and write X
for {x E a I q>(x);6tjJ(x)}. If q>;6tjJ, X must be non-empty, so there will
be an Xo E X for which Vy E X( <y, xo) r). Since Xo E X, we have
q>(xo);6tjJ(xo), and so without loss of generality we can assume a u that
is in q>(xo) but not in tjJ(xo). Now q> is collapsing, so there is ayE a for
which q>(y)=u and <y, x
o
> E r. By the specification of xo, y cannot be
in X; so q>(y)=tjJ(y)=u. But tjJ collapses too, so from tjJ(y)=u and
<y, xo> E r we get u E tjJ(xo), which is a contradiction.
(Existence) A subset Z of a is called r-transitive if x E Z 1\ <y, x) E r -
- y E Z. Clearly if q> has domain a and is collapsing for r then q> t Z is
collapsing for r (') Z 2 when Z is an r-transitive subset of a. If Z, Z' are
both r-transitive subsets of a, so too is Z (') Z'; and so if d>, d>' are col-
THE AXIOM OF FOUNDATION 39
lapsingmapswithdomainsZ,Z'respectively,!l> t Z II Z'and!l>' t Z II Z'
are also collapsing maps, both with domain Z () Z'; and so, by the
uniqueness part of the theorem, they are identical.
So let qJ be the union of all the collapsing maps !l> that are defined on
any r-transitive subset of a. From our discussion above it follows that qJ
is also a map, and that its domain Y is also an r-transitive subset of a
(since it is the union of a set of r-transitive subsets).
Now qJ is obviously collapsing; and so Y must be the largest r-transitive
subset of a which is the domain of a collapsing map.
Is Y equal to a itself? If not, take some Xo e a"'.Y such that
Yy e a"'. Y( (y, xo> r); that such an Xo would exist under these con-
ditions follows at once from the well-foundedness of r on a. We now
define a map t/J on Yu {xo} such that t/J and qJ are identical on Yand
t/J(xo)= {qJ(y) lye a 1\ (y, xo> e r} (this definition is a correct one
since (y, xo> e r ye Y).
But now it falls out at once that Yu {xo} is an r-transitive subset of a
and that t/J is collapsing. And this is impossible, since the domain of
t/Jis Yu {xo}, which properly includes Y.
Consequently Y = a; and qJ is a collapsing map with domain a.
It remains to show that rng(qJ) is in V, or, what is the same thing, that
all its elements yare. Suppose otherwise, and put X = {x e a I -, V(qJ(x};
X is then non-empty. As r is well founded there must be an Xo e X such
that YyeXy,xo>r). Thus -,V(qJ(xo, which means that there is
au e qJ(xo) such that -, V(u).
On the other hand, as qJ is collapsing, there exists aye a such that
qJ(Y)=u and (y, xo> e r. Thus y X, in view of the choice of Xo; so
V(qJ(y, so V(u); and this contradicts the conclusion of the previous
paragraph .
A binary relation r on a set a is called extensional if, whatever elements x,
y may be of a, ifYzea[(z,x>er-(z,y>er] then x=y; or, more
succinctly, if the set a, together with the relation r, satisfies the axiom of
extensionality.
A set a is called extensional if the membership relation is extensional
on a; that is, if x () a=y () a x=yforany x,yina.
Observe that every transitive set is extensional; for if a is a transitive
setandxeathenx () Q=X.
THEOREM: Let r be an extensional relation which is well founded on a. Then
there exists a unique isomorphism from a, together with the binary relation
r, on to a transitive set (together with the membership relation); and this
transitive set is in V.
PROOF: Any isomorphism such as that required by the theorem is evidently
a collapsing map; and so by the previous theorem we have uniqueness.
To prove that there is any such isomorphism at all, we consider that one
map qJ defined on a which collapses for r, and show that it fills the bill.
Let b be the range of qJ; we have already seen that b is transitive and in V.
Firstly, qJ is injective. For otherwise there would be an element d of
b of minimal rank for which there were two inverse images x, yEa; that
is to say, x, yEa, x:Y, and qJ(x)=qJ(y)=d. As x, yare distinct and r is
extensional, we can assume, without loss of generality, that we have some
u E a for which (u, x) E rand (u, y) r. The former yields qJ(u) E qJ(x)
since qJ is collapsing; so qJ(u) E qJ(Y) too. Thus for some v E a we have
(v, y) E rand qJ(v)=qJ(u); and as (u, y) r, we also have v:l:u. But
qJ(u)=qJ(v) e d, whence qJ(u), qJ(v) are of rank less than d, in contradiction
to the definition of d.
Secondly, (y, x) e r -+ qJ(Y) e qJ(x), since qJ is collapsing. And lastly,
the converse of this is proved as follows. If qJ(Y) e qJ(x) then there is a z e a
such that (z, x) e rand qJ(z)=qJ(Y) (since qJ is a collapsing map). But
qJ is also injective, so Z= y. Thus (y, x) e r.
This concludes the proof that qJ is indeed an isomorphism .
Note that this theorem is a generalization of the theorem on pp. 16f. For
if r is a strict well-ordering on a it is easily seen to be both extensional and
well founded. What is more, the range set of the collapsing isomorphism
is transitive and well ordered by E; an ordinal, no less.
The following corollary is proved with the help of the axiom offounda-
tion. It will be used in Chapter VIII.
COROLLARY: Every extensional set is isomorphic, under a unique isomor
phism, to a transitive set.
PROOF: Let a be extensional, and r the restriction of the membership rela-
tion to a; thus r= {(x, y) E a
2
1 x e y}. It follows at once from AF that
r is well founded; it is extensional, by hypothesis; and so the theorem
above gives the result .
THE HEREDITARILY FINITE SETS
A set a is called hereditarily finite if it is in V"" or, in other words, if it is
in Vn for some natural number n. Without having to use the axiom of
choice we are able to define a functional relation without parameters that
maps V", one-one on to ill.
LEMMA: lin is a natural number, n{o, I} is equipollent with a unique natural
number.
PROOF: The uniqueness is trivial since no natural number is equipollent
with any ordinal apart from itself. The rest is a simple induction on n.
Given that n{o, I} is equipollent with k, say, n+l{o, I} is equipollent with
k x {O, I}, and thus with 2k. And that is all .
We now define recursively on the natural numbers a sequence of
bijections CPn from Vn on to a natural number V
n
For n=O, CPn=0. So
suppose that we are presented with CPn: Vn V
n
; our object is to define
CPn+l: Vn+1 Vn+1
In terms of CPn we can define a bijection I/In: &'(Vn) Vn{o, I} (that is,
from V
n
+
1
on to Vn{O, I}) by
I/In (X)(i) = 1 if cP;; 1 (i) EX,
I/In (X)(i) = if cP;; 1 (i)= X,
where X c Vn and 0:::; i <vn'
But Vn{O, I} is a set of finite sequences of ordinals, binary sequences in
fact (that is, the only ordinals appearing are 0, 1), oflength VH" It is thus
well ordered by the relation R(x, y) defined on u( On) on pp. 33f. By the
lemma, the ordinal isomorphic to this well-ordering is some natural
number V
n
+
1
Thus we get our bijection CPn+l from V
n
+
1
on to V
n
+
1
by
writing CPn+l=H(CPn), where cp'=H(cp) is the relation 'Where u is the
domain of cP and v the range, cP' is that map from &'(u) which is composed
of the following two maps: the map 1/1: &'(u) V{O, I} developed from
cP, and the unique isomorphism from V{O, I} (as well ordered by R) on to
its ordinal'. This relation is determined by a parameter-free formula, and
so the functional relation n CPn defined on ill is also void of parameters.
So we have a parameter-free functional relation n r n which associates
with each natural number n a well-ordering rn on Vn (rn is the inverse image
under CPn of the well-ordering of vn)' We can now define a well-ordering r
on V to by setting x <y (mod, r) iff
rn(x) < rn(y) v [rn(x) = rn(y) = n A x < y(mod, r
ll
)] ,
Every initial segment under r is isomorphic to a natural number; for the
set of r-predecessors of Xo is a subset of VII where n is rn(xo); and so its
ordinal is finite;
This shows that the well-ordering r on V", is isomorphic to the well-
ordering of en; and the isomorphism K between en (ordered by member-
ship) and the well-ordered system < V"" r> is the functional relation sought,
since it has been defined without parameters,
RELA TIVIZED FORMULAS
Let X be some class, and E(xo, .. " XII-l) some formula in (at most) the
variables xo, .. " XII-l' whose parameters are all objects of X, We write
EX(xo, .. "XII_l) for the relativization of the formula E to X, namely that
formula defined in the following way by (informal) recursion on the length
of E, (Since the formulas are not objects of the universe, there is no
question here of a recursion on the integers of the universe, but one on
intuitive objects,)
(1) E is of the form x E y, X = y, X E a, x = a, a E x, a = x,
a E b, a = b (a, b, objects of X), E is in this case called
atomic, and is its own relativization.
(2) E is -, F ; then EX is -, FX.
(3) E is F v G; then EX is F
X
v G
X

(4) E is 3xF(x, xo, .. ,' X,.-i);
then EX is 3x[X(x) A FX(x,xo, ... ,X,._I)]'
It follows at once from (2) and (4) that if Eis VxF(x, xo, .. " XII-l)- that
is, at length, -,3x-,F(x, xo, .. " XII-I) - then EX is Vx[X(x) - FX(x, xo,
"" XII-l)]'
Note that if the class X qualifies as a set, then the relation X(x) takes
the form X E X,
Consider any n objects ao, .. " a,.-l of the class X, Then it is apparent
that the formula EX(ao, .. " all-l) is true in the universe iff the formula
E(ao, .. " an-I) is true in the class X (augmented by the binary relation
E t X), As an example, a set X is extensional if and only if the formula EX
is true (in the universe), where E itself is one form of the axiom of
extensionality,
'v'x'v'y[x = y+-+ 'v'Z(ZEX +-+ Z E y)].
RELATIVE CONSISTENCY OF THE AXIOM OF FOUNDATION
In this section we shall show that the deployment of the axiom of founda-
tion cannot of itself lead to contradiction within ZF, by assuming a
model for ZF - that is, a universe 0/1 - and from it constructing a model
for ZF + AF; this latter model will be a set (in the intuitive sense),
augmented by a binary relation, for which all the ZFaxioms hold, and
the axiom of foundation too. More precisely, we show that if ZF has a
model, so also has ZF + AF; this is what is meant by relative consistency.
In fact, ifd/t is a universe, then the ZFaxioms and the axiom offoundation
all hold in the class V constructed within d/t.
This amounts to proving that for each axiom A of ZF+AF, A
V
holds
in 0/1.
Axiom of Extensionality: If x, yare in V, so too are all their members;
so if they have the same members in V, they have the same members all
told. Thus x = y, since the axiom of extensionality holds in 0/1.
Union Axiom: If x is in V, anything in Ux is obviously also in V, so by
the lemma on p. 36, Ux itself is in V.
Power-set Axiom: If x is in V and y E 8i'(x), then y c: x, and so all y's
elements are in V; thus y is too, and by the same lemma, 8i'{x) belongs
toY.
Scheme of Replacement: Let a be a set in V, and R{x, y) some formula
(whose parameters are in V) which, within the class V, defines a functional
relation. Clearly this functional relation is defined in d/t by the formula
V{x) A V(y) A R
V
(x, y); since replacement holds in 0/1, we can apply it
to this functional relation, and obtain in 0/1 a set b made up of the images
of a under it. Since everything in b is in V, b itself must be, and the proof is
complete.
Axiom of Infinity: Every ordinal is in V, ill in particular. Since 0 E ill 1\
1\ 'v'x[x E ill ~ X U {x} E illJ, the axiom of infinity holds in V.
Axiom of Foundation: Take a non-empty set a in V. Let b be some element
of a of minimal rank. But every element of b is of lower rank than b,
and therefore cannot appear in a. Thus b n a = 0.
This proof of relative consistency - like all others - can be looked at
from a different angle. Without reference to models, we can treat the
theory ZF as a formal system, in the sense that its axioms and, more
generally, its parameter-free formulas are regarded as mere finite
sequences of symbols. The rules of inference (which need not be spelt
out here, but can be found in [4J for example) are then formal rules
allowing the derivation of certain finite sequences from certain others.
A proof is just a finite sequence of formulas A
o
, ... , An-I' each of which,
if not an axiom, is derivable from earlier formulas in the sequence by
one of the rules of inference. A theorem of the theory is any sentence
which is the last formula of some proof, and the theory is called consistent
( or non-contradictory) if and only if 0 # 0 is not a theorem.
The relative consistency of the axiom of foundation can then be ex-
pressed as: if 0#0 is not a theorem of ZF, then it is not a theorem of
ZF+AF. We can prove this straightforwardly by supposing a proof of
0#0 in ZF+AF, say the sequence Ao, ... , A
n
-
2
, 0#0, and constructing
from it a proof of 0 # 0 in ZF alone.
To do this we note that every parameter-free sentence E which we have
shown to hold in any universe 0If has in effect been proved from the
axioms of ZF by means of the formal rules of inference. We can thus
consider the sequence of relativized formulas A ~ , ... , ~ _ 2 (O#ot. If
Ai is an axiom of ZF, we have just shown that Ar is a theorem of ZF. Thus
this sequence furnishes a proof in ZF of (0 # ot, and therefore one of
(0 # 0) itself.
We give two more simple examples of relative consistency proofs.
CONSISTENCY OF THE NEGATION OF THE AXIOM OF INFINITY
We shall show here that if ZF is consistent then so also is the theory T
obtained by striking out the axiom of infinity from the list of axioms, and
admitting its negation in its place. In fact, if 0/1 is a universe, the set Vw
(together with the membership relation) satisfies every axiom ofT.
Since Vw is a transitive set, the extensionality axiom automatically holds.
To show the union axiom holds, take a E V
w
; for some nEW, then,
a E V", in which case Ua c Vn for sure, so Ua E V
n
+
1
Thus Ua EVa>'
Likewise, every subset of a is in V
n
+
1
, so that PJ1(a) c V
n
+
1
, which gives
PJ1(a) E V
n
+
2
; and this shows that the power-set axiom holds.
Let a be in V
w
, and take an arbitrary formula R(x, y) with parameters
in V
w
, which, interpreted in V
ro
, defines a functional relation. In 0/1, then,
the functional relation can be defined as x E Vw /\ Y E Vw /\ RV",(x, y)
- indeed, since its domain is a set, this functional relation is a map f
with values in V
w
' Let <p: a --+ W assign to each element x of a the rank
of f(x); then since a c Vn for some nEW, a is equipollent with a finite
ordinal, and so the values of <p are all less than some natural number k.
Thus b, the set of images of members of a under J, is a subset of lik, and so
b E V
k
+1 as desired.
But the axiom of infinity fails in V
w
' For suppose that a is a set, that
o E a, and that if x E a so is x u {x}. Then every nEW is in a, so that
(J) c a. But an element of Vw cannot have (J) as a subset since it is
bound to be equipollent to some finite ordinal. Thus no such a is in V",.
CONSISTENCY OF THE AXIOM OF ACCESSIBILITY
In this section we suppose that the axiom of choice holds in the universe 0/1.
A cardinal A is called (strongly) inaccessible if it has the following three
properties:
(1) A>w;
(2) If fJ is a cardinal < A, then 2/1 < A;
(3) If (fJt)ieI is a family of cardinals less than A, indexed by a
cardinal 1<,1., then sup fJ i < A.
leI
A cardinal that is not inaccessible is called accessible. Note that nothing
prevents a cardinal less than an accessible cardinal from being inaccessible;
for example, if A is inaccessible, 2'\ though strictly greater than A, is
accessible.
LEMMA: If A is an inaccessible cardinal then VA = A; a set a is in VA iff a c VA
anda<A.
=
PROOF: Since A. c V;. we certainly have To obtain the reverse
inequality we first show that if 0( <A. then Va <A.. So let 0( be the first
ordinal <A., if there is one, for which We have V,,= U
P<II.
I &(Vp) = sup(Ci, sup 2
Vp
), by the corollary on p. 33. But if tkO(,
p<" P<II.
then Vp <A. by the definition of 0(, and so 2
v
p <A., since A. is inaccessible.
For the same reason, the supremum of the family of cardinals (2
Vp
)p < '"
being indexed by 0( <A., must be less than A.. Thus V .. is not greater in
cardinality than the supremum of two cardinals each strictly <A.; so
V .. <A., which contradicts our assumption about 0(.
So V" <A. for all 0( <A., which means that V;. = sup VII. and so
a<;'
V;. = A..
If a E V;. then a E Va, so a c Va for some OA., so a Va; thus a<A..
Conversely, suppose a c V;. and a <A.. The map x rn(x) defined on a
is a family of cardinals each less than A., indexed by a set a whose cardinal
is also less than A.. Thus the least upper bound of the family is some
cardinal p < A..
So for every x E a, and therefore =y. Thus
a c V
y
But <A.. Consequently a E V;. .
If A. is inaccessible then in V;. hold all the axioms of ZF, together with AF
andAC.
It is easy to check that the axioms of extensionality, union, power-set,
foundation, and infinity are satisfied.
For the axiom of choice, let a E V;. be a family of disjoint non-empty
sets. Then (since AC is here imagined to hold in It) there is a set b c U a
whose intersection with each x E a is a singleton. Since a c VII. for some
OA., b c VII. too, so bE V
d1
, so bE V;..
For the replacement scheme we take a E V;., and pick any formula
R(x, y), with parameters in V;. only, that within V;. defines a functional
relation. Within o/t the functional relation is defined by the formula
x E V;. /I. Y E V;. /I. RV.;>.(x, y). As its domain is a set, it corresponds to
some map f, and the range off is a set b c: VA' Since trivially Ii a < ..t, it
follows that b EVA'
The axiom of accessibility Ace denies the existence of inaccessible
cardinals: All cardinals are accessible.
We can prove the relative consistency of ACC (relative, that is, to
ZF+AC) as follows. Take a universe Cft where AC holds. Either ACC
does too - in which case nothing needs to be proved - or there is an in-
accessible cardinal in 0/1, the first such being n say. Then, by the above,
V" satisfies ZF + AF + A C. We shall see that it satisfies A CC too.
The ordinals of V" are the ordinals < n. Forif ods an ordinal of V"' it is a
transitive element of V" linearly ordered by E. But then it is an ordinal in
0/1; since it must be in V" it is an ordinal less than n.
The cardinals of V" are the cardinals < n. If 0( is a cardinal < n, we have
0( E V,,; and within V" there is no one-one map of 0( on to a smaller ordinal,
since there is not one even in 0/1 itself. Thus 0( is a cardinal of V".
Conversely, if 0( <n is not a cardinal we can map it by some bijectionf
on to a smaller ordinal y. So f c: 0( x y, so f c: 0( x 0(; thusf E V,,+3'
Consequently f E V" and there is in V" a bijection from 0( on to y, which
proves that 0( is not a cardinal of V".
So let 0( be a cardinal <n; by the definition of n it is accessible in 0/1,
and so one of (1), (2), (3) above must fail for 0(. Therefore one of the
following holds.
(1') 0( (j); here 0( must be accessible in V".
(2') There is a cardinal P < 0( for which 2';;?; 0(; then P E V"'
2' E V"' and so 0( is accessible in V".
(3') 0( U Pi> where I is a cardinal less than 0( and every PI is
leI
likewise a cardinal <0(. The map i --+ PI is a subset of Ix 0(
and therefore of 0( x 0(; so it belongs to V". Thus 0( is accessible
in V".
This suffices to establish that the axiom of accessibility holds in V".
CHAPTER IV
THE REFLECTION PRINCIPLE
Let X be a class and E(xo, ... , X
k
-
1
) a formula with all parameters
confined to X. We shall say that X mirrors the formula E (or is a mirror
for E) if, for arbitrary a
o
, ... , ak-l in X, E holds if and only if EX also
holds. That is, 'the class X mirrors the formula E' is none other than
VXo .. VXk-l [X (xo) /\ ... /\ X(Xk-l)
(E(xo, ... , Xk-l) -- EX(xo, ... , Xk-l)]'
Note that if E is quantifier-free then EX is E itself, and so any class
whatever is a mirror.
A formula E is said to be prenex (or in prenex normal form) if it is
N
o
N
1
".N
r
-
1
E', where each Ni either is "1 or is 3 followed by some
variable x, and E' is free of quantifiers. For prenex formulas E, F we say
that F is a truncation of E if it is obtained from the latter by removing
some of the "1 's and 3x's at the beginning. So if E is N
o
N
1
... N
r
-
1
E', for
somep>O the formulaFis of the form N
p
N
p
+1".N
r
-
1
E'.
It is well known, and will not be proved here, that every formula A
is equivalent to a prenex formula E containing exactly the same parameters
and exactly the same free variables (see [4] or [5J).
LEMMA: Let E be a prenexformula without parameters, and let (Xn)nero be
an increasing sequence of sets whose limit (that is, U Xn) is X. If each
new
Xn is a mirror for E and all its truncations, then X too is a mirror for E and
all truncations.
PROOF: We prove this lemma by informal (metamathematical) induction
on the number of "1 's and 3x's at the front of E. When Eis quantifier-free
there is nothing to prove since either E has no truncations or all its
truncations are quantifier-free as well; in either case every set is a
mirror.
If Eis "1 Fsuppose that each Xn mirrors E and all its truncations; then
it obviously mirrors F and all its truncations, and so by the induction
THE REFLECTION PRINCIPLE 49
hypothesis X mirrors P and all truncations. Consequently,
VXo VX
k
-
1
[X (xo) A . A X(Xk-l)
-+ (P(Xo, ... , Xk-l) +-+ pX (Xo, ... , Xk-l))] '
and since E is ..., P, EX is ..., pX, and so X mirrors E.
Likewise, if E is 3xP(x, x
o
, ... , Xk-l) and every Xn mirrors E and its
truncations, then every Xn mirrors P and its truncations; so (induction
hypothesis) X mirrors P and its truncations. It remains to show that X
is a mirror for E.
Pick ao ... ak-l in X. If EX(ao, ... ak-l) is true we have 3x(x E X A
A pX(x. ao ... ak-l)). and so there is ana E X for which pX (a, ao ... , ak-l)'
As X mirrors p. this gives pea. ao ... ak-l) and so 3xP(x. ao ... ak-l)'
But this is E(ao, ... , ak-l), which is thereby established.
For the converse suppose that E(ao, ... ak-l)' By taking n sufficiently
big we can ensure that each of ao ... ak-l is in XII (since they are in X).
As Xnis a mirror for Ewe get EXn(ao ... , ak-l), or
Thus there is an element a of Xn for which pXn(a. ao, ... , ak-l)' Now XII
mirrors P, so it follows from this that F(a. ao ... ak-d. And X also mir
rors p. so pX (a. ao, ... , ak-l) too. But this entails that 3x(x E X A
A pX (x. ao, ... , ak-l)) or. what is the same thing, EX (ao, ... ak-l)'
And this proves the lemma. I
Note that this lemma is in effect a lemma scheme. since for each
formula E it yields the proof of a true formula depending on E.
The reflection principle is another theorem scheme derivable (if AP is
assumed) within the theory ZP. According to the principle. for every
parameter-free formula E(xo ... Xk-l) there exists an arbitrarily large
limit ordinal /3 for which Vp mirrors E. More formally. we state it as follows.
THEOREM: Por every parameter-freeformula E(xo ... , Xk-l) we have
Vrx3/3 > rxVxo ... VXk-l [/3 is a limit ordinal
A (xo E Vp A .. A Xk-l E Vp
-+ (E (xo ... , Xk-l) +-+ E
V
" (xo. "0, Xk-l))] 0
PROOF: We suppose E to be in prenex normal form, and prove by informal
induction on its length that for each ordinal ex there is a greater limit
ordinal P such that Vp mirrors E and all truncations.
If E is quantifier-free we take for P the first limit ordinal >ex, since
every set - and therefore Vp - is a mirror for E.
If E is I Fwe can find a limit P > ex such that Vp mirrors F and all trunca-
tions of F. So for a
o
, ... , ak-1 in Vp we get
F(ao, ... , ak-1) ~ FVP (ao, ... , a
k
-
1
),
and so
thus Vp also mirrors E.
So suppose E is 3xF(x, xo, ... , x
k
-
1
); according to the induction
hypothesis, whatever ordinal ex is there is a larger limit P such that Vp
mirrors F and all its truncations.
We define the functional relation y= 4i (xo, ... , X
k
-1) of k arguments to
mean that 'y is the set of all x of minimal rank for which F(x, Xo,' .. ,
Xk-1) holds, if there are such x, and the empty set otherwise'. It is clear
from the definition that
3xF(x, xo, ... , Xk-1)
~ 3x [x E 4i (xo, ... , Xk-1) 1\ F (x, x
o
, ... , x
k
-
1
)] .
We now proceed to define an ro-sequence of ordinals, (Pn)ne(JJ> as follows.
Po is the first ordinal > ex such that VPo is a mirror for F and all its
truncations.
When Pi is defined for each i.,,;;,2n we put P211+1 equal to the first
ordinal> P211 such that
for every k-tuple ao, ... , ak-1 drawn from V
p2n
(Since
U {4i(ao, ... , ak-1) I (ao, ... , ak-1) E V ~ J
is a set, it is included in some V
y
, and so P2n+1 is well defined.)
P2n+2, on the other hand, is put equal to the first ordinal Y>P2n+1 for
which we get a mirror Vl' for F and all truncations.
The sequence PH is strictly increasing, as is plain enough. So if
P = sup Pm P is a limit ordinal and Vp = U VPn = U V
p2
". But, by
nero II E ro nero
stipulation, each V
P2n
mirrors F and all truncations; and therefore, by
the lemma, Vp is also a mirror for F.
It remains to prove that Vp is a mirror for E. Take, then, any
ao, ... , ak-l E Vp, and fix n large enough to fit all these objects into V
p2n
.
If E(ao, ... , ak-l) is true we have 3xF(x, ao, ... , ak-l) and so
3X(XE cli(a
o
, ... , ak-l) 1\ F(x, a
o
,, ak-l' This gives anaE cli(a
o
, ... , a
k
-
1
)
for which F(a, a
o
, ... , a
k
-
1
). As each of ao, 00., ak-l is itself in V
p2n
, a
qualifies for V
p2n
+ l' so a E Vp. Vp, however, mirrors F, so FVp(a, ao, ... ,
a
k
-
1
), so 3x(x E Vp 1\ FVp(x, ao, 00', a
k
-
1
, so EVp(a
o
, ... , ak-l) as
required.
Conversely, given the latter fact there must exist an a E Vp verifying
FVP(a, ao, 00', a
k
-
1
). Since Vp mirrors F, we have then F(a, ao, 00., ak-l)
for this a, and so E(a
o
, 00', ak-l)' which concludes the proof .
To see the significance of this theorem in a special case, note first that
if E is a closed formula, a sentence, then to say that Vp is a mirror for
it is just to say that it is true in the universe iff it is true in Vp. Thus if E
is a sentence which is true in the universe, there are arbitrary large limit
ordinals f3 such that E is true in Vp.
In later parts of the book we shall make use of a somewhat more general
principle of reflection. Consider afunctional relation y = W" with domain On,
increasing r t . ~ f 3 ~ W" c Wp) and continuous (for limit rt., W,,= U Wp).
p<"
Let W be the union (in the intuitive sense) of all the W" - that is, the class
defined by 3rt.[On(rt.) 1\ x E W,,]. Then, provided E(xo, 00., X
k
-
1
) has no
parameters, there is,for every ordinal rt., a greater limit ordinal f3 such that
'Vxooo.'VX
k
-
1
[XoEWp 1\ ... 1\ Xk-1EWp
~ (EW (xo, 00', X
k
-l) EWP (xo, ... , X
k
-1))] .
No real modification of the proof is required to establish this form of the
principle; for, apart from its continuity, the only fact about the functional
relation y = V" used above was that it was increasing. In this more general
case, the expression 'the set X mirrors the formula E' can be taken as
shorthand for the assertion that every member of X is in the class Wand
'Vxo ... 'VX
k
-l [xo EX 1\ ... 1\ X
k
-l EX
~ (EW (xo, 00., Xk-l) ~ EX (xo, ... , Xk-1]'
It should be observed that the axiom of foundation is no longer needed
for the more general form of the reflection principle; its only use pre-
viously was in guaranteeing that every set belonged to some Va'
COMPARISON OF THE THEORIES OF ZERMELO
AND ZERMELO/FRAENKEL
Let rx be a limit ordinal >w (rx is then not less than w+w). There is no
difficulty in showing that in the set V" (enriched with the membership
relation) all the following axioms hold.
(1) Axiom ofExtensionaIity.
(2) v t=yJ.
(3) Union Axiom.
(4) Power-set Axiom.
(5) Axiom ofInfinity: 3x[0 E x /\ Vy(y E X -t y U {y} E x)J.
(6) Scheme of Comprehension: for every formula A (x, xo, ... , Xk-l)
without parameters, the axiom
/\ A(z,x
o, ... ,Xk-l)]]'
(1) and (3) hold because V" is transitive; (2) and (4) because rx is a limit
ordinal- for if x, y E Va, we get x, y E Vp for some /J<rx, so {x, y} E V
P
+1>
and .9(X)EVp+1> where /J+l<rx; and (5) holds because WEV
a
- we
selected rx above w.
Lastly, to show that all the comprehension axioms (6) hold, suppose
a, a
o
, ... , a
k
-l E V.; they then belong to some Vp for /J<rx, so any
subset of a, for example the set b={z E a I AV,,(z, a
o
, ... , ak-l)} isin V
p
+
1
,
and so in Va as required.
We designate by Z the theory (Zermelo set theory) axiomatized by
(1), (2), (3), (4), (5), (6).
It is not hard to see that, provided Z is consistent, it is strictly weaker
than ZF. For we can find a sentence - for example 'Every well-ordered
system is isomorphic to an ordinal' - which is derivable in ZF but not
inZ.
For suppose given some universe q{; in Vw+ro (w+w is the first limit
ordinal > w), as we have just seen, all the axioms of Z hold. The ordinals
of Vro+ro - the sets, that is, that satisfy the relativization of On(x) to
V
ro
+
ro
- are simply the ordinals <w+w. But .9(w x w) E V
w
+
ro
, so every
well-ordering of m, being a subset of m x m, is in V
w
+
w
' Among these well-
orderings is one whose ordinal is m + m; but this ordinal is not an ordinal
of V
w
+
w
, so the well-ordering in question is isomorphic to no ordinal. Thus
V
w
+
w
fails to satisfy the sentence mentioned, which therefore cannot be
derived in Z.
Note that the foundation axiom also holds in V .. (IX being a limit ordinal
beyond m). We now intend to prove that no consistent extension of
ZF + AF can be obtained by adding finitely many new axioms to Z + AF;
put another way, this means that whatever sentence, consistent with Z + AF,
A may be, we can find aformula B which is a consequence ofZF+AF+A,
but not ofZ+ AF + A.
The formulaB is 'there exists a limit ordinal P>w such that AVp,. This
is certainly a consequence of ZF + AF + A, since A is a consequence of
that theory. We need only apply the reflection principle with w for IX.
Now if IX is a limit ordinal, it is easily checked that V .. is a mirror for
all the following formulas (whose free variables are just e, x, y).
(1) On( e), when it is written out as
V'x(x e e x c: e) A V'xV'y[x e e Aye e
(2) 'eisalimitordinal',thatis,
On(e) A A u {x}].
(3) y=&J(x), that is V'z[z e y z c: xl
xey v x=y V yex].
(4) y= Ux, that is, V'u[u e y +-dz(z e x A u e z)].
(5) x e that is, On(e) A 3f[fis a map with domain e+ 1 such
thatV'l1 e e+l(j(f/)= U &J(f(C))) A xef(e)].
'ell
Taking these as proved we can proceed to the following lemma.
LEMMA: Let E(xo, ... , Xk-l) be a parameter-free formula, and IX a limit
ordinal. Then V .. mirrors EV;(xo, ... , Xk-l) (whose free variables are e,
xo, ... , Xk-l)'
PROOF: We use metamathematical induction on the length of E, again
assumed to be in prenex normal form. If E has no quantifiers there is
nothing to prove, and if Eis -,Fthen -,Fv;;so V .. mirrors EV;iff
it mirrors FVc (which it does, by the induction hypothesis).
So suppose that E(xo, ... , Xk-l) is 3xF(x, xo, ... , X
k
-
1
), F itself being
a formula for which the lemma holds. Then E
V
" is the formula 3x [x E /\
/\ FV,,(x, xo, ... , Xk-l)] and so [Evxo, ... , Xk-l)Y" is 3x[x E V" /\
/\ (x E /\ (FVIi(X, X
o
,"" X
k
-
1
)t"].
We have noted above (in (5 that V" mirrors the formula of two varia
bles x E As a consequence, if x, (E V"' we get (x E +-+ X E
furthermore, since ( is in V" it is less than IX, so x E V" /\ X E +-+ X E
And again, since V" mirrors F
V
" (induction hypothesis), we have for
x, Xo,"" Xk-l E V" that
(x, xo, ... , Xk-l)Y" +-+ FVI;' (x, xo, ... , Xk-l)'
Gathering together all these remarks, we can conclude that for (,
xo, ... , Xk-l E V" we have
[E
V
" (xo, ... , Xk-l)t" +-+ 3x [x E /\ FVIi(X, XO, ... , Xk-l)];
and this latter being EVIi(XO' ... , Xk-l), the result follows .
So let IX be the first limit ordinal greater than w for which V" mirrors the
sentenceA; since A is true in 0lI, IX is the least limit ordinal > wsuch that A
V
".
By the lemma V" mirrors the formula A
V
" (of one free variable only,
().Bisjust3([(isalimitordinal> w /\ A
V
"],soB
v
"is3( [(E V", /\ g
is a limit ordinal >wt" /\ (AV"t"J. Since V", mirrors both A
V
' and '( is
a limit ordinal >w', B
V
" is equivalent to 3([( E V", /\ (is a limit ordinal
>w /\ AVIiJ. However, as IX was chosen to be the least limit above w for
which A
V
", we cannot have, for a limit (, both A
V
" and ( E Vo;; and, as
we can see, B
V
" requires both to be true. Thus B
V
" is false. This proves
that, though V" satisfies all the axioms of Z, AF, and A (by the choice of IX),
the formula B does not hold in it. So B cannot be a consequence of
Z+AF+A.
As a corollary to the above result we may note that if ZF is consistent,
it is not finitely axiomatizable. For the assumption of consistency allows
us to assume ZF+AF consistent; so if ZF were equivalent to a single
sentence A, ZF + AF would be equivalent to Z + AF + A.
Note that even if A is an arithmetical formula (by this we mean one
relativized to V
ro
), B is not. G6del's second incompleteness theorem
provides us with another proof of the above result, but one which associates
with every A consistent with ZF + AF an arithmetical B' following from
ZF+AF+A, butnotfromZ+AF+A.
The following, however, remains an open problem: if ZFis consistent,
does there exist a sentence A, consistent with Z, such that Z + A is
equivalent to ZF + A? From what we have just shown it is clear that such
a sentence A would have to contradict the axiom of foundation; for
otherwise Z + AF + A would be consistent and equivalent to ZF + AF + A.
CHAPTER V
THE SET OF EXPRESSIONS
In the first two chapters we constructed, within each universe 011, a sort
of replica for several of the fundamental ideas of mathematics; the idea
of a mapping, for example, or that of a natural number. And we agreed
thenceforth to use these words in the senses we had given them in 011,
and not at all in their everyday senses.
We are now going to carry out the same 'reconstruction' for the idea
of a formula of set theory. In this case, however, we shall continue to use
the word 'formula' in its intuitive sense only, as we have used it up to
now, and will take advantage of the word 'expression' in order to refer
to the kind of objects that we shall define in the universe.
Choose from 011 any five distinct sets, to be denoted by v, "', 3, e, and
=; these sets might be, for instance, the sets 0, 1, 2, 3,4. Choose also a
denumerable set "Y whose elements, called variables, differ from all five
sets chosen above - "Y might consist of the natural numbers 5. If x
is a variable we shall write the ordered pair <3, x) as 3x.
We define by recursion a map n -+ Iffn with domain co: Iffo is the set
of ordered triples < e, x, y) and < =, x, y) where x, y are in "Y; more
formally,
The elements of Iff 0 are called atomic expressions. Further,
Iffn+1 = Iffn U [{ ""'} x Iffn] U [{v} x Iff;]
U [({3} x "Y) x Iffn] ;
it consists therefore of everything in Iffn' together with ordered pairs and
triples of the form < "', CP), < v, CP, IP), <3x, CP) for CP, IP E Iffn and x E "Y.
Writing Iff for U Iff n we call Iff the set of all expressions.
ne '"
It is easy to check by induction on n that, in virtue of our choices for
v, ""',3,8, =, and "Y, every Iffn c: V"" and so Iff c: V",; every expression
is a hereditarily finite set.
THE SET OF EXPRESSIONS 57
Given an expression (p we shall call the first n for which (p E C
n
the
depth of(P.
LEMMA: For each expression (P one and only one 0/ the following possi-
bilities is realized:
(P is atomic;
(P iso/the/orm (-, P);
(P is o/the/orm (v, P, PI);
(P is o/the/orm (:Ix, P).
Furthermore, P and P' are expressions determined by (P, and strictly lower
in depth than (P itself.
PROOF: Bearing in mind that the ordered triple (a, b, c) is defined as the
ordered pair (a, (b, c, it is clear that each expression is an ordered pair.
Also clear is the fact that the first element of such a pair is bound to be
8, =, -, v, or :Ix (where x E 1); and these objects are distinct from one
another, as that is the way we chose them.
If the first element of a pair is 8 or = we must have (P E Co; if it is -, the
second element of the pair is an expression P. Writing n+ 1 for the depth
of (P we have (-, P) E C
n
+
1
and (-, P) Cn; so by the definition of
C
n
+
1
we get P E Cn; and therefore the depth of P is less than n+ 1. On
the other hand it is obvious that P is determined uniquely by (P. Exactly
the same goes if the first element of the pair is :Ix. If, however, it is v,
then the second element takes the form (P, PI), P and P' being deter-
mined by (P, obviously enough. From (v, P, P') E C
n
+
1
and (v, P,
PI) C
n
it again follows at once that P, P' E C
m
and therefore P and
P' have depth less than n + 1.1
In the sequel the expressions (8, x, y), (=, x, y), (-, (P), ( v, (P, P),
and (:Ix, (P) will be shortened to x B y, x=y, -(P, (p) v (p), and :Ix (P
respectively; the sets 8, =, -, v will be called the membership sign, the
equals sign, the negation sign, and the disjunction sign respectively; and
:I is the existential quantifier.
The expressions (,..,(p) v (p), - -(P) v (- P, -:lx( -(P) will be
still further shortened to (P) (P), (P) A (P), Yx(P) respectively;
and the expression (P P) A (P (P) will be written (P) (P).
We now define recursively on the depth of an expression a map w
from C into the set of all finite subsets of r; w( (P) will be called the set
of free variables of the expression iP.
DEFINITION:
IfiP=xBY or iP=x=y then w(iP) = {x,y};
w(", iP) = w(iP);
w(iP v P) = w(iP) u w(I[');
w(3xiP) = w(iP)"'-.. {x}.
An expression iP is called c1osedifw(iP) =0.
It should be noted that each parameter-free formula A corresponds in an
obvious way to an expression, an expression we shall denote by r AI.
But the converse is false. Were there to exist in ott a natural number n
with, intuitively speaking, infinitely many elements, then there would be
an expression with n as its depth, but no formula could be expected to
correspond to it. In fact, one can readily see that an expression corre-
sponds to a formula if and only if its depth is a natural number with only
a finite number of elements (in the intuitive sense of the words).
We can also define (recursively on the depth of the expression iP) a binary
functional relation Y = Val (iP, X) whose domain is 'iP is an expression
and X is a set'. Y is called the value of the expression iP in the set X, and
it is a subset of w(</l)X.
DEFINITION:
(1) If iP is x B y (respectively, x = y) then
Val(iP, X) = {b E (x,y}X 1 b(X) E bey)}
(respectively {<5 E {x, y) X I <5 (x) = <5 (y)}) ;
(2) If iP = '" P then Val(iP, X) = w(</l) (X) "'-.. Val(P, X);
(3) IfiP = P v P' then Val(iP, X)
= {b E w(</l) X I <5 t w(P) E Val(P, X)
v <5 t w(P') E Val(P', X)};
(4) If iP = 3xp then Val(iP, X)
= {<5 E W(</l) X 13<5' ::> b(<5' E W('P) X 1\ <5' E Val(P, X}
(remember that in this case w(iP) = w(P)"'-.. {x}).
If iP is the expression corresponding to the formula A (xo, ... , Xk-l) it is
simple to see that Val(iP, X) is the set of maps <5 from {xo, ... , Xk-l} into
Xfor which AX (c5(x
o
), ... , c5(Xk-l holds; this is easily proved by informal
metamathematical induction on the length of the formula A.
We shall write an expression <P whose free variables form a finite set
{xo, ... , xk-d (indexed by a natural number k) as <P(xo, ... , Xk-l); such
an expression will, for brevity, be called k-adic.
An expression with parameters <Po is by definition an ordered pair
(<P, 11) where <P is an expression and 11 a map defined on some subset of
weep) (perhaps weep) itself). If the free variables of <P are Xo, ... ,Xk-l>
Yo, ... , YI-l (and they are all different), and the domain of 11 is the subset
{xo,"" Xk-l} of w( <P), and the value of 11 at each Xj (for i <k) is aj, then
the expression <Po with parameters is written <P{a
o
, ... , a
k
-
l
, Yo, ... , YI-l)'
The set {Yo, ... , YI-l} - that is, w{ <P)",-dom(11) - is, by definition, the set
of free variables of the expression <Po with parameters, and is denoted by
w(<Po).
Suppose now that X is a set and <Po = (<P, 11) an expression all of whose
parameters are taken from X. Such an expression will sometimes be called
an X-expression. We extend the earlier definition of Val so that Val { <Po, X)
becomes the set of all maps c5 E w(tPo)X such that c5 u 11 E Val(<P, X) -
here () u 11 is the map defined on w( <P) as equal to c5 on w( <Po) and to 11
on w( <P)"'-w( <Po).
If A(xo, ... , xk-d is any formula whatever with parameters, there
corresponds to it in an obvious fashion an expression with parameters,
r 1 for short. Val(r A', X) will be properly defined if all the parameters
of the formula A are elements of X; in which case Val(r A', X) is the
set of all maps c5 from {xo, ... , Xk-l} into X for which the formula
A
X
(c5(x
o
),"" c5{Xk-l holds; this again follows easily by informal
induction on the length of A.
An expression <P with parameters is said to be closed if w(<P)=0. If <P
is a closed X-expression - so that the set X contains all the parameters of
<P - then Val(<P, X) is a subset of 0X={0}. Accordingly, Val(<P, X)=l
or O. Ifit is 1 we say that <P is satisfied in the set X.
The following theorem, proved with the help of the axiom of choice, will
be used in Chapter VIII.
THEOREM (LOWENHEIMjSKOLEM): Let X be a set, P a subset of it, and d the
set of all closed P-expressions which are satisfied in X. Then there is a
subset Y c: X such that every expression in d is satisfied in Y, Y";;;; P +
andP c: Y.
PROOF: The axiom of choice is needed to ensure a map 9 : f!IJ(X)"'. {0} X
such that for every non-empty U c: X, 9( U) E U.
Using the map () we define recursively an (V-sequence P
n
of subsets of
X: Po is just P itself; and P
n
+ 1 is defined from P
n
in the following manner.
Let @n be the set of monadic Pn-expressions 4>(x, ao, ... , a
r
-1) for which
3x4>(x, a
o
, ... , a
r
-
1
) is satisfied in X. Let U
w
= g E X 14>(e, ao, ... , a
r
-1) is
satisfied in X}. Then P
n
+
1
is defined as the set of all 9(U
w
) for 4> in @n'
Thus P
n
+
1
contains, for each monadic 4> E @n, at least one 'satisfier'
of 4> in X, namely 9(U
w
).
Observe first of all that P
n
+
1
:::> Pn; for if a E P
n
then the expression
x=a belongs to @m and, for this expression, Uw={a}; thus 9(U
w
)=a,
and a E P
n
+
1

Secondly, since the set of expressions with parameters in P
n
has
cardinality P,. + we have @n";;;; P
n
+ But the map sending the ex-
pression 4> to 9(U
w
) is by definition a surjection, a map from @n on to
P
n
+
1
. Thus By induction, therefore, Pn";;;;P+
We set Y= U Pn' ThenP c: Y; and Y";;;; L
nero new
Now let 4>(ao . ... , ak-1) be any closed expression whose parameters are
in Y. We show by induction on its depth that 4> is satisfied in X if and
only if it is satisfied in Y.
F or atomic cI> the result is obvious. If 4> is ,.., 'P (or 'P v 'P'), 4> is satis-
fied in Yiff 'P is not (or one or other of 'P, 'P' is); the induction hypothesis
then tells us that cI> is satisfied in Y iff 'P is not satisfied in X (or either
'P or 'P' is satisfied in X); so that 4> is satisfied in Y iff it is satisfied in X.
If 4> (a
o
, ... , ak-1)= 3x'P(x, a
o
, ... , ak-1) we suppose to start with that
4> is satisfied in X. Then, for a big enough n, we will have ao, ... , ak-1 all
in Pm and therefore 'P(x, a
o
, ... , ak-1) E @n. Then
say, where a E P
n
+1> and so 'P (a, a
o
, ... , ak-1) is satisfied in X (by the
definitions of 9 and U'l')' The depth of 'P(a, a
o
, ... , ak-1), however, is
strictly less than that of 4>, and so, by the induction hypothesis, 'P(a, ao,
.. , ak-I) is satisfied in Y; from this it follows at once that 3xP(x, aQ, . ,
ak-I), whichisjustIP(aQ, .. " ak_I),issatisfiedin Y.
Conversely, if 3xP(x, aQ, .. " ak-I) is satisfied in Y, there must be an
a E Y for which pea, aQ, , .. , ak-I) is satisfied in Y. By the induction
hypothesis pea, aQ, .. " ak-I) is also satisfied in X, so 3xP(x, aQ, .. " ak-I)
is too.
In particular we have proved that every expression in d is satisfied
in Y .
The relativization of an expression to a set: Suppose we are presented with
an expression IP(xQ, , .. , Xk-I, aQ, .. " ai-I) whose parameters aQ,"', a
l
-
I
are all in a. We define, recursively on the depth of IP, another expression
IPa(xQ," " Xk-I, aQ, ... , al-d known as the relativization of IP to the set a.
DEFINITION: If IP is atomic then IP
a
= IP;
If IP = '" P then IP
a
= '" pa;
If IP= PI V P
z
then IP
a
= P ~ v P ~
IfIP=3xP then IP
a
=3x(xea A pa).
Clearly enough the expressions IP, IP
a
have the same free variables. The
parameters of IP
a
, however, are a, aQ, ... , a/-I'
We shall use the following theorem in Chapter VIII.
THEOREM: Suppose that a E b and a c: b. If IP is an expression with para-
meters in a, then Val(lP, a)= w(c;I)a () Val(IP", b).
PROOF: The proof is by induction on the depth of IP. For atomic IP the
result is obvious. For molecular IP, suppose IP= '" P (the proof is more or
less identical, mutatis mutandis, for IP= PI v P 2)' Then
Val ( IP, a)= W(c;I)a"'-Val(P, a)
= W(<Ia"'-Val(p
a
, b)
=w(c;I)a () Val(IP", b),
since, by hypothesis, the theorem holds for P.
Now suppose IP to be 3xP. We have that
Val(lP, a) = {(j E W(c;I) a 13(j' :::> (j(c5' E W(2') (a)
A (j' E Val(P, a)}.
The induction hypothesis gives Val ('1', a) equal to w('P) a r\ Val ('1'", b), so
Val ( 4>, a) = {.5 e w(0) a I 3.5' .5 (.5' e w('P) a A .5' e Val ('1'", b))}
= {.5 e w(0) a I .5 e Val (:Jx (x 8 a A '1'a), b)}
= w(0) a r\ Val(4)", b). 1
CHAPTER VI
ORDINAL DEFINABLE SETS
Relative Consistency of the Axiom of Choice
Notation: Let 4i(x) be a monadic expression whose parameters are all in
some set X; the value of this expression in X is then a subset of {x) X. By the
canonical bijection of {x) X on to X (that which sends {(x, u)} to u, for
every u E X) there corresponds to this subset a subset of X itself, written
val(4i, X), and also called, by a slight abuse of language, the value of
4i inX.
We now consider a universe <ft where the axiom of foundation holds, and
define a class OD, the ordinal definable sets, by the following formula,
o D (x): 'There are an ordinal IX and a monadic expression 4i (y, lXo, , 1X,-1)
whose parameters are ordinals <IX and whose value in the set Va is {x}'.
LEMMA: Let a be a set, and A(x, lXo, . , IXk-l)' whose parameters are
ordinals, be a formula with a single free variable. If a is the only set for
which A holds, then a is ordinal definable.
PROOF: Since we have chosen a universe <ft in which AF holds, we can
make use of the reflection principle to obtain an ordinal IX not only
greater than lXo, ... , IX
k
-
1
but also large enough for a to be in Va, and such
that Va mirrors the formula A (x, lXo, ... , IXk-l)' Then a is the only member
of Va for which the formula AV",(x, lXo, ... , IXk-l) holds, and so the value
of the expression r A(x, lXo, ... , IXk-tY"' in the set Va is {a}. Thus a is ordi-
nal definable .
CONVERSE: Let a be ordinal definable. Then there is a formula A(x, Do)
of one free variable and one parameter (the ordinal Do) which holds for a
and a alone.
PROOF: Consider the parameter-free formulas S=J(IX) and x=K(n)
which establish, respectively, bijections from On on to 0'( On) (the class
of finite sequences of ordinals), and from (J) on to VO)' These formulas have
been discussed above on pp. 33 and 41 respectively.
With the help of J one can easily define a parameter-free functional
relation 17 = J' (0() which is a surjection from On on to the class of functions
17 with a finite domain included in "f'" and with values in On,
Since a is ordinal definable, there are an expression cPo(x, 0(0' "" 0(,-1)
with parameters in On, and an ordinal Yo, such that val(cP
o
, VYo)={a},
We therefore consider the following quaternary parameter-free formula
E(x, n, y, 13): 'n is a natural number, 13, yare ordinals, the ordered pair
(K(n), J' (13 is an expression cP whose parameters are ordinals less than
y, and val(cP, Vy)={x}'.
Writing then no for that natural number such that K(no) is the parameter-
free expression cPo (x, xo, .. " X,-1) and Po for that ordinal such that
J'(Po) is the function 170 ={ (xo, 0(0), .. " (x'-l> O(,-1)}' it is apparent that
the formula E(x, no, Yo, Po) holds for a and for nothing else; this formula,
however, has three parameters, the ordinals no, Yo, and Po'
To reduce the number of parameters to one, we write a
o
for the inverse
image under J of the sequence no, Yo, Po; then the formula A (x, a
o
): 'there
are a natural number n and two ordinals y, 13, such that J(a
o
) is the se-
quence {(O, n), (I, y), (2, p)} and E(x, n, y, 13)' has a
o
as its sole para-
meter, and holds for a and a alone .
A formula like A (x, a
o
) is said to be a definition of a in terms of ordinals,
or an ordinal definition of a.
Note that what we have actually produced above is a parameter-free
formula A (x, y) of two free variables such that if a is ordinal definable
there is an ordinal 8
0
for which A(x, 8
0
} provides a definition of a. In
other words, the following theorem: 'ltx[OD(x) 38(On(a) "
'ltu(A(u, 8) u=x].
The class OD is not necessarily a transitive class, for there may be ordi-
nal definable sets a not all of whose elements are ordinal definable.
However, we can readily isolate a subclass HOD of OD, the hereditarily
ordinal definable sets, which is transitive, by the formula HOD(x):
'Every element of is ordinal definable'. (Remember that is
the transitive closure of a, the smallest transitive set, that is, which includes
a; and {x} u - see p. 38.)
It should be observed that HOD has been defined by a formula without
parameters.
LEMMA: A set a is hereditarily ordinal definable if and only if it is ordinal
definable and all its elements are hereditarily ordinal definable.
ORDINAL DEFINABLE SETS 65
PROOF: The condition is obviously necessary. So suppose that every
element of a is in HOD, and that a is in OD. We must show then that
aisinHOD.
From the properties of transitive closures investigated on p. 38 we
u u au U
u U ({y} u
yea
yea
= {a} u U
yea
By assumption every element of is in OD, whatever element y
may be of a. Since a is in OD, this shows that every element of
is in OD; and so HOD (a). I
Our main task in this chapter will be the proof of the relative con
sistency of the axiom of choice. We do this by constructing from a model
0/1
0
of ZF another model, one for ZF + AF + A C. Moving off from 0/1
0
we
first construct a model 0/1, as we have done before (p. 43), for ZF + AF.
We then show that not only ZF+ AF but also ACholds in the class HOD
constructed within 0/1.
Axiom of Extensionality: Obvious, since if a, b are in HOD, so are all
their elements.
Union Axiom: If a is in HOD, let b= U a. It is clear that every element
of b is in HOD. According to the above lemma, all that remains to be
shown is that b is ordinal definable.
Now a is certainly ordinal definable, so is the only set for which a
certain formulaA(x, oc) holds; as b is the only set for which
B(y, a): \fz(z E y ... du(u E a A Z E u)
holds, we can combine the two into
3x[A(x,oc) A B(y, x)]
which is a formula of one free variable, and one ordinal parameter. It
obviously holds just for b, so by the lemma on p. 64, b is ordinal
definable.
Power-set Axiom: Suppose a to be in HOD, and b the set of all hereditarily
ordinal definable subsets of a. We want to show HOD (b), and since we
have at once HOD (x) for every x in b, we need only check that OD(b).
If A (x, IX) is an ordinal definition of a, since b is the only set for which
B(y, a): Vz [z E Y - HOD(z) /\ z c: a]
holds, it is clear that 3x[ A (x, IX) /\ B(y, x)] holds for b and for nothing
else. Thus OD(b).
Scheme of Replacement: Let a be some object of HOD, and R(x, y, aQ,
... , ak-i) be a binary relation with parameters in HOD. If, interpreted
in HOD, R defines a functional relation, then the formula HOD(x) /\
/\ HOD(Y) /\ RHOD(x, y, aQ,"" ak-i) defining it in the universe o/t also
defines a functional relation, Sex, y, aQ,' .. , ak-i) say, for brevity, Let b
be the set of images under S of elements of a. Since every element of b is
in HOD, we need only prove that aD (b). So let A(x, IX), A(xQ' IXQ), ... ,
A (Xk-l> IXk-l) be ordinal definitions of a, a
Q, , a
k
-
i
respectively.
Then since b is the only object for which the formula
B(y, a, aQ, ... , ak-i):
Vz[zEy-3t(tea /\ S(t,z,aQ, ... ,ak_l]
holds, it enjoys the same distinction with respect to the formula
3x3xQ ... 3X
k
-i [A (x, IX) /\ A (xa, lXa)
/\ ... /\ A (Xk-i, IXk-l) /\ B (y, x, x
a, ... , X
k
-
i
)] ,
whose only parameters are the ordinals IX, lXa, ... , IXk-l' Thus OD(b), as
required.
Axiom of Infinity: If IX is any ordinal it is obviously ordinal defined by
X=IX; so every ordinal is in fact in HOD. HOD(w) in particular, and the
axiom of infinity holds for w in HOD.
Axiom of Foundation: If a non-empty set a is in HOD, and b is one of its
elements bearing minimal rank, then b is in HOD and b n a = 0.
Axiom of Choice: We can define, by a formula without parameters as
follows, an injective functional relation a=e(x) which associates an
ordinal with each monadic expression whose parameters are in On.
Starting from such an expression <P(x, aa, ... , IX'-i) we go first to the
parameter-free expression <P(x, x
a
, ... , X,-l), which is an element of
Vro' Using our bijection K from Q) on to V
ro
, we let n be the natural number
for which K(n) = 4> (x, Xo, .. " X,-I)' In a similar way we let P be the first
ordinal for which J' (p)isthefunction {(xo, cxo), ... , (x,-l> cx,_I)}whereJ'
is the functional relation defined on p. 64. Then cx= 8 (4) (x, cxo, ... , CX,-I)
is taken as that ordinal cx associated with the pair (n, P) by the isomorphism
between On
2
and On.
Granted 8 then, we proceed to define another functional relation
P=D(x), from aD into On, by the formula 'P is the least ordinal to
represent a pair of ordinals (cx, 1') where 1'=8(4)) for some monadic
expression 4> with parameters in cx and value {x} in V,,'.
The formula P=D(x) has no parameters; what is more, it defines an
injection, for x =1= x' -+ D(x) =1= D(x'), as is easily checked.
It is clear now that the relation R(x, y) defined by the parameter-
free formula OD(x) A OD(Y) A is a well-ordering of the
class aD.
So suppose a to be any hereditarily ordinal definable set. Restricting
the well-ordering R to a yields the set
b = {(x, y) e a
2
1 D(x) D(y)} ,
and it is this set which we shall show to be a well-ordering of a, and itself
in HOD.
Taking the latter point first, note that all of b's elements are in HOD,
so we have only to prove that OD(b). But since the formula
B(y, a): A vea A
z = (u, v) A D(u) D(v)]
holds only for b, and we may assume that some ordinal definition of a,
A(x, cx), holds only for a, it is clear once more that b is ordinal defined
by the formula 3x [A (x, cx) A B(y, x)]. So HOD(b).
As far as all is concerned, b certainly well-orders a; that this is also the
case in HOD is apparent from the fact that all non-empty subsets of a,
those lying in HOD in particular, have a smallest element (mod. b).
Thus Zermelo's theorem holds in HOD; so AC too. This concludes the
proof that A C is consistent relative to ZF.
The ordinals of HOD are the same as the ordinals of all; and the natural
numbers of all appear unchanged in HOD too. Note that since AFholds
in HOD we can write On (x) here as
'v'Y(YEX-+Y ex) 1\ 'v'Z[(YEX 1\ ZEX)
-+ (z E Y V Y = Z v Y E Z)] .
Moreover, Vro is in HOD. For the bijection x=K(n) from W on to Vro
provides, for every hereditarily finite set, a definition whose only para-
meter is a natural number. Thus everything in Vro is in aD, and since Vro
is transitive, this means in HOD too. V
ro
, being definable by the parameter-
free formula 'x is the set of hereditarily finite sets', is itself in aD, so too
in HOD.
Continuing on these lines, we see that if f is the map n -+ Vn defined
on w, then HOD (f); for any element of f is a pair <n, V
n
), clearly
hereditarily finite, and so in HOD. And f can be defined by the following
parameter-free formula: 'fis a map defined on W such thatf(O)=O and
f{k+ l)=&(J(k for eachk E w'.
The upshot is that the class HOD is a mirror for the formula x E Vro;
in other words, that (x E Vro)HOD X E Vro' For x E Vro can be written
3f3n [! is a map defined on W such that f(O) = 0,
f(k + 1) = &(!(k) for all k E W, and nEW and x E f(n)].
This fact, coupled with the relative consistency proof of AC just given,
yields the following result.
THEOREM: If an arithmetical formula E (one whose quantifiers are relativized
to Vw that is) is derivable in the theory ZF + AC, it is derivable already
inZF.
PROOF: Since (x E Vro)HOD X E Vro we can see without difficulty that
E
HOD
E, if Ehas been relativized to Vro'
So if E is a theorem of ZF+AC, there is a proof Ao, ... , A
n
-
2
, E of it
in that theory. We have shown, however, for each axiom A of ZF+AC
that A
HOD
is a theorem of ZF. Thus ... , E
HOD
is a proof of
E
HOD
, and so of E, from ZF alone .
The Choice Principle: We say that the choice principle holds for a universe
cp/ if there is a formula A(x, y) of two free variables and no parameters
which defines a well-ordering of the entire universe.
In such a case the axiom of choice is easily seen to hold. We should
note however that, as it stands, the choice principle is not only not an
axiom, it is not even a scheme of axioms; rather, it is something like an
'infinite disjunction' of sentences, of all those sentences, that is, which
say, of a binary relation without parameters, that it well-orders the
universe <11. Nevertheless, we can establish the following result.
THEOREM: If AF holds in <11, then the choice principle holds in <11 if and
only if the axiom YxOD(x) holds in <11.
PROOF: If the choice principle holds, some parameter-free formula A (x, y)
defines a well-ordering of <11. So there will be generated by this well-
ordering an isomorphism x = J( a) from On on to <11, and this isomorphism
will be definable by a formula free of parameters. Thus x=J(a) will
provide an ordinal definition for every element of <11, and therefore
YxOD(x) will hold.
Conversely, given that YXOD(x) holds, we can proceed as above to
write down the formula D(x)=:;;D{y); this is parameter-free, and it
defines a well-ordering on <11. I
In particular, granted AF, the parameter-free formula D{x)=:;;D(y)
has the following property. Ifthere is any parameter-free formula A (x, y)
defining a well-ordering of the universe, then D(x)=:;;D(y) defines a well-
ordering of the universe.
In Chapter VIII we will be able to show the relative consistency of the
axiom YxOD(x).
CHAPTER VII
FRAENKEL/MOSTOWSKI MODELS
Relative Consistency o/the Negation o/the Axiom o/Choice
(Without the Axiom 0/ Foundation)
Let R(x, y) be a relation of two arguments that establishes a bijection
from the universe %' on to itself; that is, a relation which satisfies the
following conditions:
VxVyVy' [R(x, y) A R(x, y') Y = y'] ;
VxVx'Vy[R(x, y) A R(x', y) x = x'];
Vx3yR(x, y) A Vy3xR(x, y).
We shall write this functional relation as y = F( x); and the binary relation
x e F(y) will be written x e' y. Likewise, the formula obtained from the
formula E(xo, ... , Xk-1) by changing e throughout to e' will be written
E'(xo, . , Xk-1)'
We shall show that every ZFaxiom holds in the class of all sets,
enriched with the relation E' (a universe we shall call %,'). This amounts
to proving that if A is an axiom of ZF, then A' holds in %'.
Axiom 0/ Extensionality: We must show that
VxVy [x = y Vz(z E' X Z e' y)].
So let a, b be two sets in %' such that Vz(z E' a z e' b); by the definition
of E' we then have
Vz(z E F(a) Z E F(b).
ThusF(a)=F(b). SinceFis bijective, this means thata=b.
Union Axiom: If a is a set then the class 3y [y E' a A x E' y] is one too.
For this formula is just an abbreviation of 3y[y e F(a) A x E F(y)],
and so is equivalent to x E U F(y). Putting this last set equal to c, we
obtain yeF(a)
a AXE' y)],
whence, letting b stand for F -1 ( c),
Vx [x E' b 3y(y E' a A x E' y)],
which shows that the union axiom holds.
FRAENKEL/MOSTOWSKI MODELS 71
Power-set Axiom: If a is any set, the formula Vy(y E' x Y E' a) is just
Vy[y EF(x) y EF(a)J;
that is, it is equivalent to F(x) c F(a) or, putting c for @J(F(a, to
F(x) E c. Defining b by F(b) = {x I F(x) E c}, and remembering that F is
bijective, we get F(x) E c equivalent to x E F(b). Thus
x E' b - Vy [y E' X Y E' a] ,
proving that the power-set axiom also holds.
Scheme of Replacement: Let a be a set and R(x, y) a formula such that
R' (x, y) defines a functional relation. Let c be the set of images of elements
of F(a) under this functional relation. Then
VY[YEc-3x(xEF(a) i\ R'(x,y];
so by writing again b for F-
1
( c) we have
VY[YE' b_3x(XE' a i\ R'(x,y].
Thus the replacement scheme holds.
Axiom of Infinity; We define by recursion a mapfwith domain ill :f(O)=
=F-l(O); f(71+ 1) is defined by F(J(n+ l=F(J(n U {fen)}.
Writing 1] for the range off, we then set e=F-
1
(1]).
Since Vx(xf F-
1
(0, the set F-
1
(0) must be the empty set 0' ofOll'.
Since 0' E 1], we get 0' E' e.
Moreover, if x E' e, then x E 1], so X= fen) for some nEW. Now
Vz[z E' f(n+ 1) - ZE' fen) v Z= fen)], from the definition of f(n+ 1);
consequently,
fen + 1) = X u' {x}',
and so Vx[x E' e x u' {x}' E' eJ; thus e is a set satisfying the axiom
of infinity in OlI'.
If the axiom of choice holds in 0lI, it holds also in OlI'.
For suppose that a is a set for which, in OlI', the formula 'the elements
of a are non-empty but pairwise disjoint' holds. Then
'v'x [x E' a 3y(y E' x)]
and
VXVy[(xE'a i\ YE'a i\ v zfy)J.
Put a1 for {F(x) I x E F(a)}; then the two sentences above say that the
elements of at are non-empty and disjoint in pairs. So if the axiom of
choice holds in d/t, we can find a set h1 which intersects each member of
at at a single point; that is,
Vx3y[xEal-VZ(Y=Z+-+ZEX /\ zEb
l
)],
or, alternatively,
Vx3y [F(x) E a
1
- Vz(y = Z +-+ Z E F(x) /\ Z E b
I
)].
Now the definition of al implies that F(x) E a
1
iff x E F(a); if and only
if, that is, x E' a. Thus we get
VX3y[xE'a-Vz(y=Z+-+ZE'X /\ zEb
1
)].
Write now b for F-
1
(b
1
); we obtain
Vx3y [x E' a _ Vz(y = Z +-+ Z E' X /\ Z E' b)],
showing that the axiom of choice holds in d/t'.
A set a is called an atom of the universe d/t if a= {a}; or (equivalently) if
Vx(x E a +-+ x=a) holds in d/t. No atoms can exist, it is clear, where the
axiom of foundation holds.
If ZF is consistent so is ZF + 'there is an atom'.
To prove this we define a simple bijection F from d/t on to itself by the
formula (x=o /\ y= I) v (x= I /\ y=o) v (x:;i:O /\ x:;i: I /\ x=y). In
the universe d/t' obtained by applying this bijection to d/t, the empty set
of d/t, namely 0, is an atom. For since F(0)={0}, we have at once
Vx(x E' 0 +-+ x=0).
It is possible, by a suitable choice for the bijection F, to show that several
other sentences, stronger than the negation of AF, are consistent with
ZF; for example, the sentences asserting the existence of E-cyc1es
aOEat E ... Ean-lEaO'
For the rest of this chapter we shall rely on the following example only.
If ZF is consistent, so is ZF + I there exists a set of atoms that is equi-
pollent to w'.
Define F by setting F(n) = {n}, F({n})=n for every natural number
~ I, and F(x)=x everywhere else. This is a bijection all right since n
and {p} must be different if nand p are both positive natural numbers.
Since for every ~ 1 we have Vx(x E' n - x=n) we see that n is an
atom of 0/1'. We must now show that, within 0/1', this set of atoms is
equipollent to w.
Given two objects a, b we write (we have already done this above,
without explanation) {a, b}' for their pair in 0/1'; it is clear that {a, b}' =
=F-1({a, b}). (a, b)' is similarly the ordered pair in 0/1' of a and b, so
that (a, b)' =F-
1
({F-l({a}), F-1({a, b})}).
We can recursively define a function/with domain w by setting
Vx [x E' fen) - 3i < n (x = f(i)] ;
thus I(n+ 1)=F-
1
({I (O),f (1), ... ,f(n)}) and 1(0)=0'. From this
definitionitis clear that when n runs through w, I(n) runs through the set
of natural numbers in 0/1'.
By putting (x, y)' E' g - (x, y) E f, we can then define a map g in 0/1'.
We obtain (x, y)' E F(g) - (x, y) E I, and thus if h is the set of images of
elements of/underthe map (x, y) ~ (x, y)', wehaveg=F-1(h).
After all this, it is clear that in 0/1' g is a bijection from the natural
numbers of 0/1 to the natural numbers of 0/1'. Since all the positive natural
numbers of 0/1 are atoms in 0/1', we have shown that the latter universe
contains a set of atoms equipollent to w.
RELATIVE CONSISTENCY OF THE NEGATION OF
THE AXIOM OF CHOICE
In this section we shall provide a model of ZF in which the axiom of
choice fails. From what we have just proved we can suppose given a
universe 0/1
0
in which there is a set of atoms, X, equipollent to w.
We begin by defining a functional relation y= W" on the ordinals by
Wo=X; W,,= U @J(Wp)(foralllX#O).
p<"
Note that since Xis a set of atoms, Wois transitive. Thus Wo c @J(W
o
)=
= Wi. For 1 P IX, furthermore, it is clear that Wp c W". It follows that
whatever ordinals IX, P are, if P IX then Wp c W".
In the same way as we proved similar results for V" in Chapter III,
we can show that W,,+1 =@J(W,,), and W,,= U Wp when IX is a limit
p<"
ordinal. The union (in the intuitive sense) of all the W" we denote by W;
W(x), then, is the formula 31X[On(lX) 1\ x E W,,].
If a is in W we again call the least ordinal IX for which a E W" the rank
ofa.
LEMMA: An object a is in W if and only if all its elements are. The rank
of any element of a is strictly below that of a itself, provided only that a does
not have rank O. The axioms ofZF hold in the class W.
PROOF: All parts of the lemma are proved in the same way as similar
results were proved for V in Chapter IlL.
The set X is in W since X E W
1
Moreover, every atom in W is an
element of X. For suppose, on the contrary, that there were an atom a
of W whose rank were non-zero. Since no limit ordinal is the rank of
anything, we can suppose a's rank to be 0(+ 1. Thus a E W,,+l, so a c: W,.;
but a E a, so a E W,,' contradicting the definition of rank. Thus a has
rank 0, so is in X.
Again, the bijection f from X on to w is in W; for f consists only of
pairs (a, n) with a E Xandn E w; (a, n)isthereforein W
n
+
3

Although W does not satisfy the axiom of foundation, it does satisfy
the sentence
Yx[x:#:0-+3Y[YEX A (ynx=0 v y={y})]].
For if a non-empty set a is in W, let b be an element of it ofleast rank. If
the rank in question is positive, every element of b is of lower rank, so
b n a = 0; but if it is zero, b is an atom.
Consequently the class W proves the consistency of the axiom system
T given by: ZF + I the class of all atoms (defined by the formula x = {x}J
is a set equipollent to w'+Yx[x:#:0-+3y[yex A (ynx=0 v y=
= {y})]].
We now consider a universe OU
1
satisfying all these axioms; and, writing
X for the set of atoms of 0//
1
we carry out the above construction within
OUt W,.= U L1'(WfI); Wis the class defined by
31X[On{lX) A XE WJ. f/<,.
Every object of ou1 is in W; equivalently, 0//
1
satisfies the axiom
Yx31X[On(lX) A x e W,,]. For suppose the existence of an a not in W;
let b be its transitive closure, and c the set of those elements of b not in W.
Since a is not in W, some element x of a is not in W; and since this x will
be in c, c cannot be empty. So take any y E c; it is not in W, so has an
element z not in W. Since y E band b is transitive, z is also in b. Thus z
qualifies for c, which means that y n c#O, for every y E c. Moreover,
no such y can be an atom either, for in this universe all atoms are in W.
The set c thus contradicts the axiom
'v'x[x#0-+3Y[YEX /\ {ynx=0 v y={y})]].
If u is a bijection from X on to itself (a permutation it is also called),
we can define by recursion a functional relation (1a. such that (1) if IX is an
ordinal (1a. is an E-automorphism of Wa. and (2) (10=(1, and if P' ~
then up is an extension of (1p" For suppose the family ((1p)p<a. has
been defined to meet these requirements. Then for limit IX we have
Wa.= U Wp, so we can define (1a. simply as U up; and the proof that
P<a. P<rx
(1a. has properties (I) and (2) is entirely straightforward.
For IX=P+I we have Wa.=&' (Wp), so can define Ua. by setting (1a.Cu)=
={(1p(x) I x E u} for every U E Wa.' Since up is a permutation of Wp, it is
clear that (1a. is a permutation of &,(Wp). In addition, (1a. is an extension of
(1p; for if U E Wp each x E U belongs to Wp, and since (1p is an E-automor-
phismofWp, we have X EU+-TUP(X) EUP(U), so that UP(U) = {up(x) I x E u}=
= (1a.(u).
Finally, (1 a. is itself an automorphism of Wa., as can be shown by proving
that, for arbitrary x, U E Wa.,
X E U +-T (1",{x) E (1",(u).
But if x E U, we get x E Wp, so up(x) E u",(u) - by the definition of (1a. -, so
Ua.{X)E(1a.{U).
Conversely, if u..(x) E (1a.(u), then ua.{x) E Wp. As Ua. is a permutation of
W", which simply extends the permutation (1 p of W
P
' we have x E Wp, so
(1p(x) E ua.{u), and therefore x E U, by the definition of (1a.'
Starting from U 0 = (1 then, the functional relation IX -+ U rt as recursively
defined above has all the required properties.
From it we can define another functional relation, y = g'.,( x), as the
common extension of all the (1 rt; this functional relation, defined by
31X[On(lX) /\ X E W", /\ Y = U..(X)] ,
has domain 0//
1
and, indeed, is an automorphism of 0//1' so that the
following are all true in 1 :
'v'x'v'x' [.9"a(X) = .9"a(X') -+ x = X'];
'v'y3x[y = .9"a(x)];
'v'x'v'x' [x ex' +-+ .9" a(x) e .9" .,(x')] .
(The proof of these sentences is immediate from the definition of .9".,.)
Thus for each permutation (j of the set X of atoms we have defined a
functional relation y = .9" i x) which is an automorphism of the entire
universe ~ 1 For notational simplicity we shall from here on write
y=(j(x) for y=.9".,(x); since for any atom x we have (j(x)=.9"a(x), this
convention will not precipitate any confusion.
LEMMA: Let E(ao, ... , all-i) be a sentence with parameters ao, ... , a,,-1>'
and let (j be a permutation of X. Then E(ao, ... , all-i) +-+ E(jao, ... , (jail-i)'
PROOF: The proof is by informal induction on the length of the formula
E(ao, ... , an-i); if this formula is atomic, that is, ao e a
1
or ao=a
1
, then
the theorem holds because (j is an e-automorphism of the universe ~ l'
If E(ao,' '" all-i) is ,F(ao" .. , all-i) we have F(ao, ... , all -1)
+-+ F( (ja
o
, ... , (ja
ll
-1) by the induction hypothesis; and therefore the required
equivalence holds also for E. When E(a
o
, ... , all-i) is F(ao, ... , an-i) V
V G( ao, "., a,. -1) the proof is the same.
If E(a
o
, .. " all-i) is 3xF(x, ao, "" all-i), suppose it to be true in ~ i ;
then for some object a we have F(a, ao, .. ,' all-i)' and so F(ja, (jao, .. "
(jail-i) by the induction hypothesis; thus 3xF(x, (jao, ... , (ja
ll
-1)' The
converse is handled likewise .
Note that this lemma is another example of a theorem scheme, this
time in the theory T defined on p. 74 above. What we have shown is the
sentence 'For every permutation (j of the set of atoms, and for every
xo, ... , X,.-l> E(xo, ... , X,.-1) +-+ E(jxo, ... , (jX,.-l)' for an arbitrary para-
meter-free formula E(xo,"" X
n
-i)'
LEMMA: Let A (x, ao, ... , a
k
-1) be a formula of one free variable holding
for just one set, the set a; let (j be a permutation of X. Then (ja is the only
setfor which theformula A (x, (jao, "., (jak-l) holds.
PROOF: By assumption we have 'v'x[x=a +-+ A(x, ao, ... , ak-1)]; so, by
the previous lemma,
'v'x [x = (ja +-+ A (x, (jao, ... , (ja
k
-
1
)]
LEMMA: If lX is an ordinal, and U is a permutation of X, then UlX = lX.
PROOF: Otherwise, suppose lX is the first ordinal for which UlX#=lX; then
u/J=/J for all /JelX, so ulX={u/Jl/JelX}=lX, which contradicts the
supposition. I
Within the universe ~ we can apply the generalized reflection principle
(p. 51) to the functional relation y= W,.. Since the universe is exhausted
in this case by the (intuitive) union of all the W,.'s, we can conclude that
for every formula Ethat is without parameters
\flX3/J > lX\fxo ... \fX
k
-
1
[xo e Wp A A Xk-l e Wp
~ (E(xo, ... , Xk-1) - EWp(xo, ... , Xk-1)].
We can proceed to define the class OD
1
by the formula OD
1
(x), 'There
are an ordinallX and a monadic expression 4>(y, lXo, ... , lX
r
-1, a
o
, ... , as-t)
all of whose parameters are atoms or ordinals < lX, and val ( 4>, W,.) = {x}'.
This class will be referred to as the class of sets definable by ordinals
and atoms. As in Chapter VI we can prove the following result.
LEMMA: Let A (x, lXo, ... , tXr-1' u) be a singulary formula whose parameters
IXo, , IXr-t, U are all ordinals, except for u, which is a finite sequence of
atoms (a map, that is to say,from some natural number s into X). Then if
a is the only set for which the formula A holds, a is definable by ordinals
and atoms.
PROOF: The generalized reflection principle provides us with an ordinal
IX>IXo, , IXr-1 and big enough for u and a both to belong to W,., and
such that W,. mirrors the formula A.
Then the value of the expression r A(x, IXo, . , IXr-l> up in W,. is {a}.
Write 'l'(x, IXo, . , IXr-l> u) for this expression. Since u is a finite sequence
ao, ... , as-l of atoms we have
u = {(O, ao), ... , (8 - 1, as-I)}.
Thus the expression
Elz(Yt[t 8 z<-> t = (0, ao) v ... v
t = (8 - l,as-l)] A P(x,IXo, ... ,lXr-hZ)
gets the value {a} in W" and has only ordinals and atoms for parame-
ters1
The converse can also be obtained.
CONVERSE: If a is a set definable by ordinals and atoms, then there is a
formula A (x, b
o
, u) of one free variable that holds for a and a alone; the
parameters of A, moreover, are just two - the ordinal b
o
and the finite
sequence of atoms u.
PROOF: By the specification of a we have an ordinal 1o and an expression
!l>(x, 0(0, ,0(,-1> ao, ... , as-t) whose parameters are the atoms a
o
, ... , a
8
-1
and the ordinals 0(0, , O(,-t <1o; and the value of 4> in WyO is {a}.
Such an expression specifies uniquely a natural number (the one asso-
ciated with the parameter-free expression !l>(x, xo, ... , X,-1> Yo, .. , Y8-t),
an ordinal (the one associated with the finite sequence of ordinals
0:o, .. ,0('-1)' and a finite sequence u=ao, ... a
8
-
t
of atoms; and therefore
it singles out an ordinal Po and a finite sequence of atoms, u. A formula
with three parameters E(x, 1o. Po. u). namely 'the expression (whose
parameters are ordinals and atoms) associated with the ordinal Po and the
finite atom-sequence u has the value {x} in the set W
yO
' is therefore at
hand. The number of parameters can then be reduced to two by means of
the bijection from On
2
on to On; and this yields a formula A (x, b
o
, u)
holding, as required, for a and nothing else. Such a formula A(x, b
o
, u)
will be said to be a definition of a by ordinals and atoms .
Note that the class OD
t
has been defined by a parameter-free formula.
Also parameter-free is the formula HOD
t
(x) which defines the class of
sets hereditarily definable by ordinals and atoms, HOD
t
. It is simply
'Every element is in OD
t

LEMMA: A set a is in HOD
t
if and only if a is in OD
t
and every element of
aisinHOD
t

PROOF: The proof is the same as that of the corresponding lemma in
Chapter VI .
That the class HOD
t
satisfies the axioms of ZF can also be proved in the
way the same claim for HOD was proved in the last chapter.
We can prove that the set X of atoms is in HOD
t
For every atom a
is definable by the formula x=a, so is in OD
t
As each atom
is in HOD
1
So we need only show that X is in 0 D
t
; but since it is the set
of atoms, it is definable by 'v'z(z e x +-+ z={z}), so certainly qualifies for
ODt
But the axiom of choice is not satisfied in HOD
t
; worse, the set of atoms
cannot even be linearly ordered.
FRAENKELjMOSTOWSKI MODELS 79
For suppose in HOD
1
there were a set v which imposed a strict linear
order on X. Let A (x, 15
o
, u) be some definition of v by ordinals and atoms.
Then since u is only a finite sequence of atoms ao, ... , as-l> and X is
equipollent with co (in 011
1
), there must be two atoms b, c different from
all theao, ... , a
s
-
1
' v linearly orders X, however, so without loss of generality
we can suppose (b, c) E v. Let (1 be that permutation which switches band
c round, but otherwise leaves X unaltered; then nothing happens to any
of the atoms ao, ... , a
s
-
1
under (1, so (1U=U. Moreover, (115
0
=15
0
, since 15
o
is an ordinal. But, by the lemma on p. 76 (1V is defined by the formula
A(x, (115
0
, (1u), so (1V actually equals v. Now (b, c) E v by assumption, so
(1b, uc) E UV, so (c, b) E v. Thus v is not a strict order, which contradicts
our hypothesis.
Let us note two other properties of X which contradict AC and yet are
true of Xin HOD
1

Every subset of X either is finite or has a finite complement with respect to X.
Let Y be some object of HOD
1
which is a subset of X, but not finite
(that is to say, it is not equipollent with any natural number); and
suppose A (x, Yo, u) to be a definition of Yin terms of ordinals and atoms.
As Y is not finite we can certainly find abE Y which is not among the
terms of the finite atom-sequence u. Let c be any other atom ignored by
u, and let (1 be that permutation of X which switches band c round
whilst keeping the other atoms fixed. We have at once that (1U=U and
(1Yo = Yo (since Yo is an ordinal). And as Y is defined by the formula
A(x, Yo, u), we must also have uY= Y. As b E Ywe get ub E uY, so c E Y.
Thus every atom like c, that is every atom not accounted for in u, is an
element of Y. Thus X". Y is finite.
X is not finite, but it has no infinite denumerable subset (subset equipollent
with co).
As X is equipollent with co within 011
1
, it cannot in HOD
1
be equipollent
to a natural number. If Y is a subset of X that is equipollent with co, it
cannot be finite; and so X".Y is finite according to the previous result.
But then X itself is equipollent with co, which means that it can be well
ordered. And this, we know, is false.
Note that the model of ZF+-,AC just constructed does not satisfy
AF, so it does not solve the problem of whether ZF+AF+,AC is
consistent. Indeed, the set here which cannot be well ordered is a patho
logical one, the set of atoms. The results of P. Cohen described in [2] yield
far more interesting relative consistency results; for example that of
ZF + AF + 'g;( ill) cannot be well ordered'.
CHAPTER VIII
CONSTRUCTIBLE SETS
Relative Consistency of the Generalized Continuum Hypothesis
Given a set X and a subset Y of it we say that Y is a subset of X definable
by parameters, if there is a monadic expression 4>(x, ao, ... , ak-I) which
has parametersao, ... , ak-I inXand the value Yin the set X.
We define a functional relation y= rI(x) on the class of all sets by
the formula 'y is the set of subsets of x definable by parameters'.
If the axiom of choice holds, then ifa is an infinite set, rI(a)=a.
For, first of all, every member b of a provides a subset of a, {b}, definable
by parameters by way of the expression x=b; so And the set
of expressions with parameters in a has cardinality (since
any such expression can be specified by a parameter-free expression and a
finite sequence of members of a); so
This proves that for infinite a, rI (a) is a proper subset of f!jJ(a).
Observe however that the functional relation y= rI(x) is not increasing
with x; that, in other words, we can have a c b without having
rI(a) c rI(b). For if a is a subset of b, but not one definable by para-
meters, then a e rI (a) (defined as it is by the expression x=x), but
a rI(b). However, if a is also a member of b, this possibility no longer
obtains.
THEoREM:Ifa c bandaebthenrI(a) c rI(b).
PROOF: If c e rICa) then c=val(4), a) for some monadic expression 4>(x)
whose parameters ao, ... , ak-I are in a. Thus z e c iff z e val(4), a), that
is iff {(x, z)} e Val(4), a). But by the theorem on pp. 61f., Val(4), a)=
= Val (4)'', b) (") {x}a; thus z e ciff {(x, z)} E Val(4)'', b) A {(x, z)} e (x}a.
Consequently,
z e c +-+ z e val (4)'', b) A Z e a,
so c=val( 4>", b) (") a.
It follows at once that c=val[tP"{x) 1\ x 8 a, b], since in general
val(tP, b) is the set of elements of b that satisfy tP. But all the parameters
of the expression tP"{x) 1\ x 8 a, namely a, ao, ... , ak-l are elements of b.
Thus c E TI{b).1
By setting first Lo = 0 we can define recursively a functional relation on
the whole of On by setting thereafter
The class L defined by the formula L{x): 3(X[On{a) A x E L,,], the union
(in the intuitive sense) of all the L", is called the class of constructible sets.
Appropriately enough, a set a is called constructible if it is in L; if, there
fore, it is in L" for some ordinal a.
The axiom of constructibility is the sentence 'Every set is constructible',
or, equivalently, 'v'x3a[On(a) A x E La]. The main business of this
chapter will consist in showing that if ZF is consistent, it remains so on the
addition of the axiom of constructibility. We do this by proving that the
class L with the usual membership relation satisfies all the ZFaxioms and
the axiom of constructibility too.
We first prove some results for the L,,'s very similar to ones proved already
for the VIZ'S. If a' a then L". c L", since La' = U TI (Lp) c U TI (Lp) =
p<,,' p<"
=L".lfp<a, thenLp E L,,; for Lp E TI(Lp) and TI(Lp) c La. By our most
recent theorem then, since Lp c L" and Lp E L", we have TI (Lp) c TI (L,,).
As L",+1 = U TI(Lp) we have La+l = TI(L,,) for every ordinal (x. On
p .. "
the other hand, if a is a limit ordinal,
so that L" = U Lpfor every limit ordinal a.
p<"
Corresponding to the definition of rank, we now provide a definition
of order. The order of a constructible set x is the first ordinal (X for which
x E L". We write it od{x), and note that od(x) is never a limit ordinal;
for if a is a limit and x E L" then x is already in Lp for some smaller p.
CONSTRUCTIBLE SETS 83
LEMMA: If a is constructible, so are all its elements; their orders, moreover,
are strictly less than the order of a.
PROOF: Let a=p+l be the order of a. Then aELp+
lo
so aETI(Lp).
Thus a c: Lp, and every element of a must belong to Lp; these elements,
therefore, have lower order than a. I
It follows that for any ordinal a the set L", is transitive; for if a E L",
and b E a, the order of a is ::;; a, so also the order of b.
THEOREM: Every ordinal 0( is constructible, and the order ofa is a+ 1.
PROOF: If there is an ordinal a for which a E L"" let 0(0 be the least. Then
0(0 E U TI (Lp) and so 0(0 E TI (Lp) for some P E ao; thus ao c: Lp,
pe",o
giving P E Lp, in contradiction with the definition of 0(0'
That much done, it remains to prove that 0( E La+1 for every ordinala.
Again, let ao be the least ordinal, if there is one, such that 0(0 ; L"'o+l'
For every P E 0(0 then, we have P E Lp+
1
, so P E TI(Lp). It follows that
ao c: U TI (Lp); and therefore that 0(0 c: L",o'
P<"'o
Consider now the parameter-free expression !V(x) of one free variable
VUVV[U8X A V8X=>U8V V U = v V V8U]
A VUVV[U8X A V8U=> V8X].
Let the value of this expression in the set L",o be Y, which is thus the set
of elements x of L",o for which the following formula, which is relativized
to L",o' holds.
VuVv[uEL",o AVEL",o-+[(UEX AVEX)
-+ (u E V V U = v v V E u)]]
A VuVv[uEL",o 1\ VEL",o-+(UEX 1\ VEU)-+VEX)].
Seeing that L",o is transitive, an element x of it will satisfy this formula if
and only if
VUVV[UEX 1\ VEX-+UEV V U=V v VEU]
1\ VuVv [u EX 1\ V E U -+ V E x] ;
just, that is, when it is an ordinal.
The value of !V(x) in L",o is therefore the set of ordinals belonging to
L",o' We have proved that ao anyway is not one of these, and therefore
no greater ordinal is (since L"o is transitive). Moreover, as shown above,
Clo c Lrzo'
As a consequence, 4> (x) takes the vaIueClo inL
lZo
' so Clo E n(L"o)=L"o+1>
which contradicts the definition of Clo. I
We can now proceed to check that all the axioms of ZF + AF hold in the
classL.
Axiom of Extensionality: If a is constructible, its elements are too. So
this axiom holds.
Union Axiom: Let a be constructible, and Cl some ordinal for which
a E LIZ' As LIZ is transitive, the set b = U a is a subset of L". Since, more-
over, the value in LIZ of the expression Ely(y 8 a A x 8 y) - which has one
free variable x, and a single parameter a, in L" by assumption - is clearly
b, we can conclude that bE L
IZ
+
1
, and that b is constructible.
Power-set Axiom: Let a be constructible, b the set of constructible subsets
of a. The map x -+ od(x) defined on b has some subset of the ordinals
for its range, and some ordinal Cl will therefore be an upper bound to
this subset. So for ail x E b we will have so that b c LIZ'
Nothing prevents us from setting Cl so high that a is a member of LIZ'
Take the expression Vu[u 8 x => U 8 a). It has one free variable, x, one
parameter a E L". Its value in Lrz is the set of elements x of Lrz which are
subsets of a; b then, by the definition of Cl. Thus bE L+1> and so b is
constructible.
Scheme of Replacement: Let R(x, y, ao, ... , ak-l) be a formula of two free
variables whose parameters ao, ... , ak-1 are in L; and suppose that, when
interpreted in L, it defines a functional relation. Within the universe IJlt
this functional relation is defined by the formula L (x) A L(y) A
A RL(x, y, ao, ... , ak-l); both domain and range are subclasses of L.
Let b stand for the set of images of the elements of some constructible
set a under this functional relation. We must show b to be constructible.
Each element of it certainly is, so pick some ordinal Clo exceeding the
order of any element of b and big enough for a, ao, ... , ak-l to be in
L"o' Then b c LlZo'
We now take the parameter-free k+2-ary functional relation R(x, y,
xo, ... , Xk-l), and apply to it the generalized reflection principle (p. 51),
using the functional relation y = Lit' This provides us with an ordinal
fJ > lXo such that
'v'x'v'y'v'xo ... 'v'Xk-l [x E Lp 1\ YELp
1\ xoELp 1\ . 1\ Xk_1ELp--+(RL(X,y,xo"",Xk_l)
RLp (x, y, xo, ... , Xk-l] .
Since ao, ... , ak-l E Lp we have
(*) 'v'x'v'y [x E Lp 1\ YELp
--+ (RL(X, y, ao, ... , ak-l) RLp(x, y, a
o
, ... , ak-l)]'
Now the expression r R(u, v, ao, ... , ak-lr
l
has two free variables
u, v, and parameters ao, ... , a
k
-
1
E Lp; call it q,(u, v) for short. Thus
Val (4), Lp) is the set of mapsJ: {u, v} --+ Lp such that RLP(J(u),J(v),
ao, ... , ak-l)' By (*), however, this set is the set ofmapsJ: {u, v} --+ Lp for
which RL(J(U), J(v), ao, ... , a
k
-
1
). Consequently, Val (3u(u 8 a 1\
4>(u, v), Lp) is the set of maps J: {v} --+ Lp for which 3x(x E a 1\
RL(X,J (v), a
o
, ... , a
k
-
1
; so val (3u(u 8 a 1\ q,(u, v), Lp) is just the set of
elements of L which are images of members of a under the functional
relation RL(X, y, ao, ... , ak-l); in fact, b, in view of the choice of fJ. Thus
bE L
p
+
1
, and is constructible.
Axiom oj Infinity: Every ordinal is constructible, as we have seen; so (i)
is constructible.
Axiom oj Foundation: Let a be constructible and non-empty, and b be
an element of it of lowest possible order. Every element of b has strictly
lower order than b itself, so no element of b is an element of a. Thus
b () a=0.
We have now proved that all the axioms of ZF + AFhold in L.
Note that the ordinals of L are the ordinals of the universe O/!. For all
ordinals IX are constructible, and clearly they all satisfy OnL(a). Con-
versely, a constructible IX satisfying OnL(IX) is transitive and linearly or-
dered by E; so an ordinal.
THE AXIOM OF CONSTRUCTIBILITY HOLDS IN L
We must now show that the axiom of constructibility holds in L.
Absolute Formulas: A parameter-free formula E(xo, ... , Xk-l) is called
absolute ifit can be obtained by applications ofthe following rilles.
(1) A formula without quantifiers is absolute.
(2) If A and B are absolute, then A v B and A 1\ B are absolute.
(3) If A (x, xo, ... , Xn-t) is absolute then 3xA(x, xo, ... , xn-t)is absolute.
(4) If A (x, y, xo, ... , Xn-t) is absolute, then
'Ix [x E y A (x, y, x
o
, ... , Xn-t)]'
a formula with free variables y, xo, ... , X
n
-l' is absolute.
On the basis of this definition we call a relation absolute if it can be
defined in IfI by an absolute formula.
THEOREM: Let E(xo, ... , Xk-t) be an absolute formula, Wa transitive class
(that is, 'Vx'Vy[W(x) 1\ y E X W(y)]), and W' an extension of
W (that is 'Vx(W(x) W' (x). If ao, ... , ak-1 are objects in W such that
E
W
(ao . ... , ak-1) is true, then EW'(a
o
, ... , ak-1) is also true. Equivalently,
'Vxo ... 'VXk-t [W (xo) 1\ 1\ W (Xk-t)
(EW(xo, ... , Xk-t) EW'(xo, ... , Xk-1]'
PROOF: The proof is an informal induction on the length of the absolute
formula E. For quantifier-free E the result is immediate since E
W
and
E
W
' are the same as E.
If Eis F v G, takeao, ... , ak-1 in W; by the induction hypothesis
F
W
(a
o
, ... , ak-l) F
W
' (ao, ... , ak-1)
and
so
(FW (ao, ... , ak-1) v G
W
(ao, ... , ak-1
(FW' (ao, ... , ak-1) v G
W
' (ao, ... , ak-1,
as required. The same proof works if Eis F 1\ G.
If E(xo, ... , Xk-1) is 3xF(x, xo, ... , Xk-1), take any ao, ... , ak-1 in W
such that E
W
(a
o
, ... ,a
k
_
1
) holds; thus 3 x [W(x) 1\ FW(x,ao ... , ak-1)].
So for some ain Wwe have FW(a, ao, "" ak_l),Bytheinductionhypothesis
therefore F
W
' (a, ao, .. " ak-l)' whence 3x[W' (x) A F
W
' (x, ao, .. " ak-l)]'
This is E
W
' (ao, , .. , ak-l)'
Lastly, suppose that E(y, xo, , .. , Xk-l) is 'Ix [x E Y --+ F(x, y, xo, .. "
Xk-l)], and pick objects b, ao, .. " a
t
-
1
in Wsuch that EW(b, ao, .. " at-I)'
Then
'Ix [W(x) --+ (x E b --+ FW(x, b, ao, , .. , ak-l)]
which is the same thing as
But as b is in W, which is a transitive class, x E b --+ W(x). Moreover, by
the induction hypothesis, (W(x) A F
W
(x, b, ao, , .. , at-I) --+ F
W
' (x, b,
ao, .. ,' at-I)' Thus the last inset formula yields
'Ix [x E b --+ F
W
' (x, b, ao, , .. , at-I)],
and so E
W
' (b, ao, ... , ak-l)'.
Note that the definition of absolute relation given here is equivalent to
that ofP. Cohen in [2]; it is not the same as K. Godel's definition in [1].
If the relation G(y, zo, .. " Z,-I) and the functional relation y= F(xo, .. ,'
Xk-l) are absolute, then the relation G(F(xo, .. " Xt-l), Zo, ... , Z,-I) is
absolute.
The relation in question can be defined by
3y [y = F(xo, ... , Xt-l) A G(y, zo, ... , ZI-I)] '
so can be realized by application of rules (2) and (3) to the absolute
relations given.
In particular, if a is a set for which y=a is equivalent to some absolute
formula R(y), and if G(y, Zo, .. " Z,-I) is absolute, then G(a, Zo, .... Z,-I)
is absolute.
For the rest of this section we shall assume the axiom of foundation to
hold for "II, and our efforts will be directed towards proving that the
functional relation y=La. is absolute, From this it will follow easily that
the axiom of constructibility holds in L.
The (singulary) relations On (x) and x = ware absolute.
Nothing has to be proved for the former, since AFholds,l and so On(x)
can be written
'v'u'v'v [u E x A V E X ~ U E V V U = v v V E u]
A 'v'u [u E X ~ 'Iv (v E U ~ V EX)].
For the latter we simply check the absoluteness of the following relations:
x={), defined by 'v'y(y E X -+ Y if. x).
y=x u {x}, defined by 'v'z(z E Y -+ Z=X v Z E x) A
x=w,definedby 'v'Z(ZEX-+ZEY) A XEy.
On(x) A () E X A 'v'y [y E X ~ Y u {y} EX]
A 'v'y [y E X ~ (y = () v 3z(y = Z U {z}))].
We have shown above that certain compositions of absolute relations are
themselves absolute. Since the functional relation y=L", whose absolute
ness we are after, is defined by recursion, we shall want to show that func
tional relations defined recursively from absolute functional relations are
themselves absolute. And we shall also want to show that the functional
relation y= IT (x) is absolute. These results are the content of the next two
lemmas.
LEMMA: Let y=H(x) be an absolute functional relation whose domain is
the class of all maps defined on the ordinals. Then the functional relation F
recursively definedfor every ordinal IX py
F(IX) = R(F r IX)
is absolute.
PROOF: We check successively the absoluteness of a dozen or more
relations.
z={X,y}:XEZ A yEZ A 'v't(tEZ-+ t=x v t=y).
Z= (x, y) : Z= {{x}, {x, y}} (composition of absolute functional relations).
y c x: 'v'Z[ZEY -+ ZEX].
z=xuy:x c Z Aye Z A v t ( t E Z ~ tEx V tEY).
Z=X n y: Z C X A Z C Y A 'v't[t E X -+ (t E y ~ t E z)].
z=x"",y: Z C x A 'v't[t E x ~ (t if. y ~ t E z)].
Z C xxy: 'v't[tEZ ~ 3u3v[UEX A VEy A t=(u, v)]].
Z::> xxy: 'v'U[U E x ~ 'Iv [V E Y -+ 3t(t E Z A t=(u, v)]].
1 The formula On (a) as written on p. 14 is patently not absolute. What is more,
it can be shown not to be equivalent, in ZF, to any absolute formula.
CONSTRUCTIBLE SETS
Z = x x y : the conjunction of the previous two relations.
Z maps x into y : Z c x x Y /\
'v'u[uEx-dv3t(VEY /\ tEz /\ t=(u,v]
/\ 'v't'v't''v'u'v'v'v'v' [(t E Z /\ t' E Z /\ U EX /\ V E Y /\ V' E Y
/\ t=(u,v) /\ t'=(u,v'-v=v'].
fis a map with domain x: 3y(Jmaps x into y).
fis a map: 3x(Jis a map with domain x).
g =ft x(a ternary relation):
89
3x' [x c x' /\ (J is a map with domain x') /\ (g is a map
with domain x) /\ g c J] .
y=f(x) (a binary functional relation): (fis a map) /\ (x, y) Ef.
We can now write the formula y=F(rx) which we are interested in as
On(rx) /\ 3f[(I is a map with domain rx) /\
'v'fJ(f3 E rx - f(f3) = H(J t fJ) /\ y = H(J)].
If the relation y = H(J) is absolute, so too is this .
We shall need also the following absolute relations.
fis an injection ofx into y (fis one-one):
(J maps x into y) /\ 'v't'v't''v'u'v'u''v'v [(t E f /\ t' E f
/\ U EX /\ U' E X /\ V E Y /\ t = (u, v)
/\ t' = (u', v _ u = u'].
fis a surjection from x on to y: (fmaps x into y) /\
'v'v [v E Y - 3u (u EX /\ (u, v) E j)] .
h= fo g (a binary functional relation):
3x3y3z [(g maps x into y) /\ (J maps y into z)
/\ (h maps x into z)]
/\ 'v't [t E h - 3u3v3w [(u, v) E g /\ (v, w) E f
/\ t = (u, w)]] .
Now the relation y = La., whose absoluteness is in question, is a functional
relation defined by recursion. We shall therefore apply the above lemma.
For the relation y=H(J) we take y= U f](J(x), and this we
xedom(!)
must show to be absolute (dom(f) is the domain off). It can be written,
however, as
'Ix [x E dom (I) IT (I(x) c y]
"Vz[zey -+3x[xEdom(l) " ZEIT(I(X)]].
The following lengthy lemma will therefore complete the
is absolute.
LEMMA: Thefunctional relation y= IT (x) is absolute.
PROOF: As the relations x = and y = x u {x} are absolute, it follows that
the relations X= 1, x=2, ... are absolute (by successive composition). As
"", v, 3, e, =, and "// are respectively 0, 1,2, 3, 4, and the set of natural
numbers ;;':5, the relations x= "", X= v, x=3, X=8, X= =, and x="//
(this latter being written x=m"'-{O, 1, 2, 3, 4}) are absolute. From
this we can derive the absoluteness of several other relations, as follows.
z = <ff 0 (the set of atomic expressions):
z = ({ e} x "// x "//) u ({ =} x "// x "//)
(a composition of absolute functional relations).
Z= <ffk (a functional relation of one argument, k):
3f [(I is a map with domain m) "f(O) = <ffo
" 'In [n E m -+ fen + 1) = M(I(n)] " z = f(k)] ,
where the functional relation y = M( x) is given by the absolute formula
y = x u [{ "'} x x] u [{ v} x x x x] u [({3} x "//) x x].
Z E <ff (the set of parameter-free expressions): 3k(k em" Z E <ff
k
).
z=<ff: Vk[kEm <ffk C z] " Vy[yez -+ ye <ff].
z E <ff " y=w(z) (as before, w(z) is the set of free variables of the
expression z):
z E <ff " 3f [(I is a map with domain <ff)
"VxVy[xe"//"
= {x, y}] " VxVy[x E <ff " ye <ff (1"', x = f(x)
"fv,x,y=f(x)uf(y)] " VXVY[XE"//" YE<ff
-+f3x,y=f(y)"'-{x}] " y=f(z)].
We now have to show that the binary functional relation y=W(F)X is
absolute. Note that the formula y=YX (three variables, y, X, Y) is not
absolute; for if Z is a transitive set, (y=YX)Z means that y is the set of
maps from Y into X that are elements of Z, and clearly there is no reason
for supposing that y is the set of all maps from Y into X. By the theorem
on p. 86 then, the formula y = Y X is not absolute. Only because w(F) is
always finite can we expect to get the result we want.
Some more absolute relations first.
k is a natural number and y = k X (a functional relation of two arguments,
Xandk):
31 [(I is a map with domain (0) A 1(0) = 0
A 'v'n [n e 00 -+ I(n + 1) = N (n, X,J(n))] A y = I(k)] ,
where the ternary functional relation Z=N(n, X, Y) is given by the
absolute formula
nero A 'v' 9 [g e Z -+ 9 maps n + 1 into X]
A 'v'h'v'x[(h e Y A X eX) -+ h u {(n, x)} eZ].
w(F) X c y (a ternary relation with arguments y, X, F):
3ep3n [n e 00 A (ep is a bijection from w(F) on to n)
A'v'I(lenX-+/oepey)].
y = W (F) X (two arguments, X, F):
w(F)X c y A 'v'1 [J e y -+ (J maps w(F) into X)].
y is a closed expression with parameters in X:
3F3ep[Fetf A (epmapsw(F)intoX) A y=(F,ep)].
y = 9 ~ (y is the set of closed expressions with parameters in X):
'v'x [x e y -+ (x is a closed expression with parameters in X)]
A 'v'F[F e tf -+ 'v'ep [ep e w(F)X -+ (F, ep) e y]].
y e t f ~ (y is an expression with parameters in X and a single free variable
x):
3F3ep [F e tfAx e w(F)
A (epmaps w(F)"'-{x} into X) A y=(F,ep)].
y = tfx:'v'z[z e y -+ Z e tfx]
A 'v'F'v'ep [F e tfAx e w(F) A ep e w(F)'{x}X -+ (F, ep) e y].
y e and z is the closed expression obtained by substituting in y the
element a of X for the variable x (five arguments, x, y, z, X, a):
3F3cp3t/1 [F e is' A X e w(F) A (cp maps w(F)"-{x} into X)
A (t/I maps w(F) into X) A cp c t/I
A t/I(X) = a A y=(F,cp) A z=(F,t/I)].
4J is a closed expression with parameters in X and O=Val(4J, X) (this
is a functional relation of two arguments, 4J, X; 0 takes the values 0, 1):
4J e A 3f [(fmaps into {O, I}) A Va'v'b[a e X Abe X
-+ a e b A f(a 8 b) = 1) v (a b A f(a 8 b) = 0]
A VaVb [a e X Abe X -+ a = b A f(a = b) = 1)
v (a oF b A f(a = b) = 0] A 'v'I[' [I[' e -+ f( '" 1[') = 1 - f(I[')]
A VI[''v'I['' [I[' e A 1[" e -+ f(I[' V 1[") = f(I[') U f(I[")]
A VxVI['[x e -r A I[' e -+ ([3a(a e X A f(I['(a = 1)
A f(3xl[') = 1] v [Va (a eX -+ f(I['(a = 0) A f(3xl[') = 0])]
A 8 = f(4J)].
4J e A Y = val ( 4J, X) (y is the subset represented by the monadic
expression 4J in the set X) :
Va[aey-+(aeX A Val(4J(a),X) = 1)]
A 'v'a [a EX -+ (Val(4J(a), X) = 0 v a e y)].
y e IT (X) (two arguments, y, X):
3x34J[x e -r A 4J e A Y = val(4J, X)].
Y = IT (X):Vz[z e y -+ ZEIT (X)]
A 'v'x'v'4J[xe-r A A z=val(4J,X].
In this waytherelationy= IT (X) is shown to be absolute .
THEOREM: The relation y = L is absolute.
It follows at once that the axiom of consfructibility holds in L. For since
ZF + AFholds in the class L, the relation (y = L,,)L is a functional relation.
Moreover, according to the theorem on p. 86 we have
VyVa[L(y) A L(a) A (y = La)L -+ y = La],
since L is transitive and y=L
a
is an absolute formula. So let a be some
object in L. For some 0(0 then, a E Lao; and since OCo is an ordinal of L,
there is a unique constructible set Yo for which (YO=Lao)L holds; as
(yo = Lao)L Yo = Lao' we get Yo = Lao' soa EYo.
Thus (a E Lao)L is true, and so [30((a E La)]L is also true.
The axiom of constructibility asserts that the three classes 011, V, L are
identical (this axiom therefore implies AF). The axiom of constructibility
is usually denoted V=L, which is an abbreviation of the formula Vx[V(x)
+-+ L(x)] (when this notation is used, AFis understood to hold). What we
have shown, therefore, is that if ZFis consistent, then so is ZF + AF + V = L.
The axiom of constructibility has many important consequences. We
shall now show that it entails the axiom of choice, in particular - and
even the choice principle stated on p. 69-, and the generalized continuum
hypothesis. These two sentences are thus satisfied in L, and therefore
consistent relative to the remaining axioms.
V = L IMPLIES AC
Consider a universe 011 which satisfies the axiom of constructibility.
With a parameter-jree formula we shall define on the class of well-
orderings in tI a functional relation v = F( u) having the following property.
If u is a well-ordering of the set X, F(u) is a well-ordering of the set of
monadic expressions with parameters in X.
Let tff be the set of parameter-free expressions; it is a subset of V"" and
we have already defined (on p. 41) an injection, itself without parame-
ters, which maps tff one-one into w. This injection, K-
1
, is the inverse of
K, suitably restricted. Letj be the isomorphism from u on to its ordinal oc.
A monadic expression with parameters in X is a pair (E, 1'/) where
E E tff and 1'/ maps some subset of weE) into X. If u well-orders X, each
element of 1'/, (x, y), where x E "Y and y E X, can be associated one-one
with a pair of ordinals; namely x itself (remember the definition of"Y as
w'-.5) and the 'place-number' of y in the well-ordering u of X (that is,
the ordinal of the well-ordered initial segment Six, u. This pair
of ordinals can itself be associated one-one with an ordinal, by the
isomorphism from On
2
on to On, defined on p. 33. Thus 1'/ can be asso-
ciated one-one with a finite sequence of ordinals - namely the sequence
which takes in their natural order the ordinals associated with its elements.
As a consequence, and using the injection K-
1
noted above, we can find an
injection from the set of monadic expressions with parameters in X into
(j) x U ( On), aclass which is well ordered by the well-ordering already defined
on u( On). In this way we can obtain the well-ordering v of the set of
monadic expressions with parameters in X.
We can now define recursively on the ordinals a parameter-free func-
tional relation y = B( a) which to each ordinal a associates a well-ordering
B(a) of LI1.' such that for p:::;;'a, Lp is an initial segment of LI1. under B{a);
and on Lp, B(a) and B(P) are identical.
For the definition, suppose B{P) suitably defined for all p<a; then
if a is a limit ordinal, LI1.= U Lp and so we write B(a)= U B{P), the
p<11. p<11.
common extension of all the B(P) for P < a. If a = P + I, we use the fact
that B(P) well-orders L
p
, so that vp = F(B{P well-orders the set of mon-
adic expressions whose parameters lie in Lp. For the well-ordering B{a)
of LI1. = TI (Lp) we therefore write
x:::;;,y(mod.B{a-[xeL
p
A yeLp A x::::;; y(mod. B{P]
v [xeLp A yeLI1."Lp]
v [x e LI1. "Lp AyE LI1. "Lp A
the first expression defining x in Lp is earlier, in the well-ordering vp of
the monadic expressions with parameters in L
p
, than the first expression
defining yin Lp].
It is obvious that the ordering B(a) thus defined on LI1. has all the re-
quired properties.
The functional relationy = B(a) is thus completely defined by recursion.
A binary relation R(x, y) without parameters which well-orders the whole
of all is then available from the formula 3a[On(a) A x:::;;'y (mod. B(a].
This shows that the choice principle holds in all; by a result on p. 69,
the axiom 'every set is ordinal definable' holds also in 0/1.
v = L IMPLIES THE GENERALIZED CONTINUUM HYPOTHESIS
The proof of this result relies on a certain property of absolute relations,
dealt with in the next theorem.
THEOREM: Let all be a universe where ZF+AF+AC holds, and E(x, y) an
absolute formula defining in tPJ a functional relation Y=4>(x) of a single
argument. Then 4> (x) (x) + Nofor every x in the domain of 4>.
PROOF: Let a be in the domain of 4>, and <X an ordinal for which a E V"'
4>{a) E V"' and such that
'v'x'v'y [x E V" AyE V" -+ (Ev,,{x, y) +-+ E{x y)].
That there is such an ordinal is guaranteed by the reflection principle for
the formula E(x, y).
Let P stand for the transitive closure of {a}, that is 'if ({a}). Since
{a} c: V" and V" is transitive, we haveP c: V". Moreover, u {a},
so p= 'if (a) + 1.
Let be the set of all closed expressions that are satisfied in V" and
have all their parameters in P. By the LowenheimJSkolem theorem (pp.
59-61) there is a subset X of V"' including P, satisfying all the expressions
in and of cardinality not exceeding P +
Since V" is a transitive set, the axiom of extensionality holds in it; so
the corresponding expression
VxVy [x = y <=> Vz{z 8 X <=> Z 8 y)]
is in and is therefore satisfied in X. Thus X is an extensional set.
By the corollary on p. 40 we can find an isomorphism j of X on to
some transitive set Y; in consequence, Y will satisfy all the expressions
in the set which is produced from by replacing each parameter u
(inP, and so in X) by j{u).
But j(u) = u for every u E P, as can be seen by supposing Uo to be an
element of P for and one ofleast rank at that. For then
j(v) is equal to v for all v E Uo; for is transitive, so any such
v EP. However,j{uo)={j (v) I v E uo}=uo, which contradicts the defini-
tionofuo'
Thus P c: Y, and Y satisfies all the expressions of in addition,

We chose V" to mirror the formula E(x, y), so since a and 4>(a) are
both in V"' and E(a, 4>(a) is true, we have also EV"(a, 4>(a). Thus the
sentence E(a, 4>(a) is true in V"' and so 3yE{a, y) is also true in V". The
expression r3yE{a, y)' is thus satisfied in V"' and has a single parameter
a, in P; so it belongs to and is therefore satisfied in Y. Consequently,
the formula 3yE(a, y)holdsin Y,andforsomebe YwewillhaveEY(a, b).
But as E( x, y) is an absolute formula and Y is a transitive set, we also
have EY(a, b) -+ E(a, b). Thus E(a, b), and so b=iP(a). But b was in Y,
so 4'(a)e Y, and Ybeing transitive, 4'(a) c Y. Thus
and so
As a particular case of this theorem, suppose a is an object of o/t defined
by an absolute formula E(y); in other words, that Vy[y=a +-+ E(y)];

It is enough for our purpose to put x = 0 and to apply the above theorem
to a formula E(y) which does not contain x at all. We can deduce in this
way that, for example, if ZF + AF + AC holds in 0lI, then the formula
y = gJ( w) is not equivalent to any absolute formula.
=
THEOREM: Suppose again that ZF+AF+AC holds in 0lI. Then L(I.=a,for
every ordinal /fa is a constructible set,
PROOF: We have seen already that (X c L(I.' so that To show the
converse inequality for all we proceed by induction. For w we note
that Ln= v" for every new (a trivial induction proves this); thus Lru= V
ru
,
Now La.= U n(Lp), and by the induction hypothesis if
P<(I.
Butforinfinite awe have n(a)=a, so
An alternative proof of this would apply the previous theorem to the
absolute relation y = L(I.' It is in this way that we demonstrate the second
part of the theorem.
Recall that od(a) is the first ordinal (X for which a e La.; the relation
y=od(x) is absolute since it can be written
On(y) A xeL
y
A Vz[zey -+ x Et:L.] ,
and the relation y=Lx is absolute. At once then,
With all this achieved, let d/I be a universe satisfying the axiom of con-
structibility. We have already proved that the axiom of choice holds in
CFI, so the theorems above can be applied. Any set a is constructible, so
if a c: we have
od(a) +
since c: Thus so a E LNp +1' But a is any subset of
----
so c: LNp +
1
' whence Thus
Cantor's theorem gives us the reverse inequality, so we can conclude that
= 1> which is the generalized continuum hypothesis.
As far as arithmetical formulas are concerned, we have an analogue of
the result obtained in Chapter VI.
If E is an arithmetical formula ( aformula whose quantifiers are relativized
to Vro) that is derivable with the help of the axiom of choice and the generali-
zed continuum hypothesis (and, more generally, the axiom of constructi-
bility), then E is already derivable from ZF alone.
For Vro=Lro, as we have seen. A formula E with quantifiers relativized
to Lro is absolute, and so EL - E is a consequence of ZF.
Let, then, Ao, ... , A
n
-
2
, E be a proof of E from ZF+ V=L. If AI is
an axiom, we know that Af is a theorem of ZF. The sequence ... ,
EL is thus a proof of EL from ZF alone. With EL _ E also a
theorem of ZF, we see that E itself is a theorem of this theory.
BIBLIOGRAPHY
The two standard works on this subject are
[1] Kurt GOdel, 1940, 'The Consistency of the Continuum Hypothesis', Annals of
Mathematics Studies 3, Princeton University Press.
[2] Paul J. Cohen, 1966, Set Theory and the Continuum Hypothesis, W. A. Benjamin,
New York.
An exposition of naive set theory can be found in
[3] Paul R. Halmos, 1960, Naive Set Theory, D. Van Nostrand, Princeton.
The fundamental ideas of elementary logic and model theory are discussed in detail in,
for example,
[4] Elliott Mendelson, 1964, Introduction to Mathematical Logic, D. Van Nostrand,
Princeton.
[5] G. Kreisel and J. L. Krivine, 1967, Elements of Mathematical Logic (Model
Theory), North Holland Publishing Company, Amsterdam.
[6] Roger C. Lyndon, 1966, Notes on Logic, D. Van Nostrand, Princeton.
[7] Joseph R. Shoenfield, 1967, Mathematical Logic, Addison-Wesley.
This last book contains (amongst much else) an important chapter on set theory.
On the subject of ordinal definable sets see
[8] John Myhill and Dana Scott, 1967, 'Ordinal Definability', Summer Institute on
Axiomatic Set Theory, Los Angeles.
The proof given in the text of the generalized continuum hypothesis in L is close to
that of
[9] Carol Karp, 1967, 'A Proof of the Relative Consistency of the Continuum Hypo-
thesis', in Sets, Models and Recursion Theory (ed. by John N. Crossley), North-
Holland Publishing Company, Amsterdam, pp. 1-32.
For the topic of absolute expressions see also
[10] Azriel Levy, 1965, 'A Hierarchy of Formulas in Set Theory', Memoirs of the
American Mathematical Society 57.
SYNTHESE LIBRARY
Monographs on Epistemology, Logic, Methodology,
Philosophy of Science, Sociology of Science and of Knowledge, and on the
Mathematical Methods of Social and Behavioral Sciences
Editors:
DONALD DAVIDSON (Rockefeller University and Princeton University)
JAAKKO HnmKKA (Academy of Finland and Stanford University)
GABRIiiL NUCHELMANS (University of Leyden)
WSSLEY C. SALMON (Indiana University)
HILPINEN (ed.), Deontic Logic: Introductory and Systematic Readings. 1971,
VII + 182 pp. Dft.45,-
W. BETH, Aspect of Modern Logic. 1970, XI + 176 pp. Dft.42,-
tPAUL WEINGARTNER and GERHARD ZECHA, (eds.), Induction, Physics, and Ethics.
Proceedings and Discussions of the 1968 Salzburg Colloquium in the Philosophy of
Science. 1970, X + 382 pp. 00. 65,-
A. EBERLE, Nominalistic Systems. 1970, IX + 217 pp. Dft.42,-
PAAKKO HINTIKKA and PATRICK SUPPES, Information and Inference. X + 336 pp.
Dft.60,-
tKAREL LAMBERT, Philosophical Problems in Logic. Some Recent Developments. 1970,
VII +176 pp. Dft. 38,-
tP. V. TAVANEC (ed.), Problems of the Logic of Scientific Knowledge. 1969, XII+
429 pp. Dft. 95,-
S. CoHEN and RAYMOND J. SEEGER (eds.), Boston Studies in the Philosophy of
Science. Volume VI: Ernst Mach: Physicist and Philosopher. 1970, Vm+295 pp.
Dft.38,-
tMARSHALL SWAIN (ed.), Induction, Acceptance, and Rational Belie/. 1970, VII+
232 pp. Dft.4O,-
tNICHOLAS RESCHER et aI. (ed8.), Essays in Honor of Carl G. Hempel. A Tribute on the
Occasion of his Sixty-Fifth Birthday. 1969, VII + 272 pp. Dft. 46,-
tPATRICK SUPPES. Studies in the Methodology and Foundations of Science. Selected
Papers from 1951 to 1969.1969, XII +473 pp. Dft.70,-
tJAAKKO HnmKKA, Models for Modalities. Selected Essays. 1969, IX + 220 pp.
Dft.34,-
tD. DAVIDSON and J. HnmKKA: (eds.), Words and Objections: Essays on the Work
of W. V. Quine. 1969, vm + 366 pp. Dft. 48,-
tJ. W. DAVIS. D. J. HoCKNEY, and W. K. WILSON (eds.), Philosophical Logic. 1969,
vm + 277 pp. 00. 45,-
tROBERT S. CoHEN and MARX W. WARTOFSKY (eds.), Boston Studies in the Philosophy
of Science. Volume V: Proceedings of the Boston Colloquium for the Philosophy of
Science 1966/1968.1969, VllI+482 pp. Dft.58,-
tRoBERT S. COHEN and MARX W. WARTOFSKY (eds.), Boston Studies in the Philosophy
0/ Science. Volume IV: Proceedings 0/ the Boston Colloquium for the Philosophy 0/
Science 1966/1968. 1969, VIII + 537 pp. Dfl. 69,-
tNICHOLAS RESCHER, Topics in Philosophical Logic. 1968, XIV + 347 pp. Dfl. 62,-
tGONTHER PATZIG, Aristotle's Theory o/the Syllogism. A Logical-Philological Study 0/
Book A 0/ the Prior Analytics. 1968, XVII + 215 pp. Dfl. 45,-
tC. D. BROAD, Induction, Probability, and Causation. Selected Papers. 1968, XI + 296 pp.
Dfl.48,-
tRoBERT S. COHEN and MARX W. WARTOFSKY (eds.), Boston Studies in the Philosophy
0/ Science. Volume III: Proceedings 0/ the Boston Colloquium for the Philosophy 0/
Science 1964/1966. 1967, XLIX + 489 pp. Dfl. 65,-
tGUIDO KONG, Ontology and the Logistic Analysis 0/ Language. An Enquiry into the
Contemporary Views on Universals. 1967, XI + 210 pp. Dfl. 38,-
EvERT W. BETH and JEAN PIAGET, Mathematical Epistemology and Psychology. 1966.
XXII + 326 pp. Dfl. 58,-
EVERT W. BETH, Mathematical Thought. An Introduction to the Philosophy 0/ Mathe-
matics. 1965, XII + 208 pp. Dfl. 32,-
tPAUL LORENZEN, Formal Logic. 1965, VIII + 123 pp. Dfl.22,-
tGEORGES GURVITCH, The Spectrum 0/ Social Time. 1964, XXVI + 152 pp. Dfl.20,-
tAo A. ZINOV'EV, Philosophical Problems 0/ Many- Valued Logic. 1963, XIV + 155 pp.
Dfl.28,-
tMARx W. WARTOFSKY (ed.), Boston Studies in the Philosophy 0/ Science. Volume I:
Proceedings o/the Boston Colloquium/or the Philosophy 0/ Science, 1961-1962. 1963,
VIII + 212 pp. Dfl. 22,50
tB. H. KAZEMIER and D. VUYSJE (eds.), Logic and Language. Studies dedicated to Pro-
lessor Rudolf Carnap on the Occasion 0/ his Seventieth Birthday. 1962, VI + 246 pp.
Dfl.32,50
"'EVERT W. BETH, Formal Methods. An Introduction to Symbolic Logic and to the Study
0/ Effective Operations in Arithmetic and Logic. 1962, XIV + 170 pp. Dfl.30,-
HANS FREUDENTHAL (ed.), The Concept and the Role 0/ the Model in Mathematics
and Natural and Social Sciences. Proceedings 0/ a Colloquium held at Utrecht, The
Netherlands, January 1960. 1961, VI + 194 pp. Dfl. 30,-
tP. L. R. GUIRAUD, Probtemes et methodes de la statistique linguistique. 1960,
VI + 146 pp. Dfl.22,50
.J. M. BOCHENSKI, A Precis 0/ Mathematical Logic. 1959, X + 100 pp. Dfl.20,-
Sole Distributors in the U.S.A. and Canada:
GORDON & BREACH, INC., 150 Fifth Avenue, New York, N.Y. 10011
tHUMANITIES PRESS, INC., 303 Park Avenue South, New York, N.Y. 10010

Introduction To Axiomatic Set Theory (1971) - Krivine

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Introduction To Axiomatic Set Theory (1971) - Krivine

Hochgeladen von

Copyright:

Verfügbare Formate

INTRODUCTION TO AXIOMATIC SET THEORY

Das könnte Ihnen auch gefallen