Professional Documents
Culture Documents
(Studies in Logic 50) John L. Bell - Intuitionistic Set Theory-College Publications (2014) PDF
(Studies in Logic 50) John L. Bell - Intuitionistic Set Theory-College Publications (2014) PDF
J o h n L. B e l l
Studies in Logic
Volume 50
John L. Bell
© Individual author and College Publications 2014.
All rights reserved.
ISBN 978-1-84890-140-7
College Publications
Scientific Director: Dov Gabbay
Managing Director: Jane Spurr
http://www.collegepublications.co.uk
Introduction 1
Lattices 91
Heyting and Boolean algebras 93
Coverages on partially ordered sets and their associated frames 97
Connections with logic 98
Bibliography 107
Index 111
Preface
While intuitionistic (or constructive) set theory 1ST has received a certain attention
from mathematical logicians, so far as I am aware no book providing a systematic
introduction to the subject has yet been published. This may be the case in part
because, as a form of higher-order intuitionistic logic - the internal logic of a
topos - 1ST has been chiefly developed in a tops-theoretic context. In particular,
proofs of relative consistency with 1ST for mathematical assertions have been
(implicitly) formulated in topos- or sheaf-theoretic terms, rather than in the
framework of Heyting-algebra-valued models, the natural extension to 1ST of the
well-known Boolean-valued models for classical set theory.
In this book I offer a brief but systematic introduction to 1ST which develops the
subject up to and including the use of Heyting-algebra-valued models in relative
consistency proofs. I believe that 1ST, presented as it is in the familiar language of
set theory, will appeal particularly to those logicians, mathematicians and
philosophers who are unacquainted with the methods of topos theory.
The title I originally had in mind for this book was Constructive Set Theory. Then it
occurred to me that the term “constructive” has come to connote not merely the use
of intuitionistic logic, , but also the avoidance of impredicative definitions. This is
the case, for example, with Aczel’s Constructive Set Theory in which the power set
axiom (which permits impredicative definirions of sets) is not postulated. Since the
power set axiom and impredicative definitions are very much a part of 1ST, to avoid
confusion I have decided (with some reluctance) to give the book its present title.
In classical set theory free use is m ade of the logical principle known as the
Law of Excluded Middle (LEM): for any proposition p , either p holds or its
negation -<p holds. But there are a num ber of intriguing mathematical
possibilities which are ruled out within classical set theory solely because of
the presence there of LEM. This suggests the idea of dropping LEM in set-
theoretical arguments, or, more precisely, basing set theory on intuitionistic
logic. Accordingly, let us define Intuitionistic Set Theory (1ST) to be any of
the usual axiomatic set theories (e.g. Zermelo-Fraenkel set theory ZF) based
on intuitionistic - rather than classical - logic.
Call a set countable if it is empty or the range of a function defined on the set
N of natural numbers, subcountable if it is the range of a function defined on
a subset of N, and numerable if it is the dom ain of an injection into N. In
classical set theory all of these notions are equivalent, as the following
argum ent shows. Obviously, every countable set is subcountable. If £ is
subcountable, there is a subset U of N and a surjection / U -*E. Then the
function m: £ —» N defined by m(x) = least n e N for which f(n) = x is injective,
and it follows that £ is numerable. Finally, suppose that E is numerable, and
let m: E >-►N be injective. Then either £ = 0 or E ^ 0 ; in the latter case, fix
e e £ and define / N -» £ by setting, for n e rang e(m), f(n) = unique x g E for
which m(x) = n; and, for n € rang e(m),f(n) = e. T h e n /is surjective, and so £ is
countable.
It should be clear that the validity of this argum ent rests on two
assumptions: LEM and the assertion that N is well-ordered (under its natural
ordering). Accordingly, if we base our reasoning on intuitionistic logic in
which LEM is not affirmed, then we can see that, while subcountability
obviously continues to be inferable from countability, the argum ent deriving
countability from num erability breaks dow n because of its dependence on
LEM. One m ight suppose that the validity of the argum ent deriving
numerability from subcountability survives the passage to intuitionistic
logic, but actually it does not, for it uses the assum ption that N is well-
ordered, and this can be shown to imply LEM. For, given a proposition p,
define U = {x e N: x = 0 a p} u {!}. Notice that 0 e U <-> p. Then U is
nonem pty and so, if N is well-ordered, has a least element n. Since n e U, we
have n = 0 v n ~ 1. If n = 0, then 0 e IT, whence p; if n = 1, then 0 £ U, whence
—>p. Hence p v ~^p, and LEM follows.
Now Cantor would also have accepted the extension of this argum ent to
show that ! cannot be subcountable in the above sense For given U c; N and
a surjection (p: U N*', if we define fi N -> N by (*)
2
then clearly f £ range(cp) and so again <p fails to be surjective. But this
argum ent uses LEM and so is not valid within 1ST. In fact, it is consistent
with 1ST for N" to he subcountable, so making N!i both subcountable and
uncountable!
is refutable. For if (*) held, then we could extend cp to a surjection N -> N' by
assigning the constant value 0 to all n g U. This would make N 1countable
which we have already shown to be an impossibility. Secondly, N will have
an uncountable subset. For the suset U of N cintroduced above annot be
countable, since the composite of (p with any surjection N -» 1/ would be a
surjection N -» Nn, making NM, impossibly, countable. Finally, the
subcountability of NNimplies that there is a function defined on a (proper) subset
of N zuhich cannot be extended to the whole of N. To see this, take U and cp as
above and define fi U -» N by setting, for n e N
/ ( n ) = <p(n)(n) + l .
<PK ) K ) = 9 K ) = f ( n 0) - cp(n0)(n0) + 1.
3
countability does not hold up within 1ST and in fact it is consistent with 1ST
for Nn to be num erable1!
Let us see what happens (within 1ST) w hen NN is replaced by the set
Par(N,N) of all partial functions from N to N. First, we observe that Par(N,N)
cannot be subcountable, for suppose U c N and (p: U -» Par(N,N) is a
surjection. Let r e Par(N,N) be the identity map on dom (r) = {x e II: x £
dom((p(x))} . Then since (p is surjective, r = cp(/t) for some n e U quickly
leading to the contradiction n e dom(r) <-> n £ dom(r). 2
Nor can Par(N,N) be numerable. For suppose cp: Par(N,N) >-* N is injective,
and define u g Par(N,N) to be the identity m ap on
1 The proof of this fact is, unfortunately, beyond the scope of the present book.
2 Clearly this argument continues to hold when N is replaced by an arbitrary set £.
4
r <£ Par*(N,N), in other words, {x e N : x £ dom(cp(x))} is uncountable. In fact,
the countability of Par*(N,N) is consistent with 1ST3.
POWER SETS
Let us turn next to another celebrated theorem of Cantor, namely that, for
any set £, the cardinality of £ is strictly smaller than that of its power set P£.
One way of construing this is the assertion that there can be no surjection
£ -*• PE. Within 1ST this can be proved, as in classical set theory, by
employing the argum ent of Russell's paradox: given (p : £ —> P£ one
defines the associated "Russell set"
R —{x g £: x £ ip(x)}
3 Again, the proof of this fact is beyond the scope of the present book.
5
What if we replace PN by the set P*N of all countable subsets of N? Suppose
that P*N s actually countable, and let cp: N P*N be a surjection. Defining
R = {x e N: x g cp(j)} as before, the argum ent above only leads to a
contradiction when R e P*N, from which we conclude that R £ P*N, that is
R is uncountable. In fact, it follows directly from the consistency of the
subcountability of Nn with 1ST that the subcountability of P N is also
consistent with 1ST4, since the m ap Nn -» P N : f h* ra n g e (/) is surjective.
In classical set theory, P£ is naturally bijective with 2£, the set of all maps5
£ -» 2 = {0, 1}. In 1ST, this is no longer the case. Here, in general, P£ = Q£,
where Q is the object of truth values or propositions, that is, the set PI of all
subsets of {0}. Q is only identical with 2 w hen LEM is aassum ed.6 In fact, in
1ST, 2£ is isomorphic, not to P£, but to its Boolean sublattice C£ consisting of
all detachable subsets of £ (a subset 17 of E is said to be detachable if
V* e £. x e U v x £ II). W hat happens when we replace PE by C£ in the
above arguments? Classically, of course, this makes no difference, but do the
"Russell's paradox" argum ents survive the transition to 1ST? Well, if one
takes the first argum ent, showing that there can be no surjection (p; £ P£,
one finds that, when P£ is replaced by C£, the set R & range((p) is actually
detachable and the argum ent goes through, proving in 1ST that there can be
no surjection £ C£. But the second argument, with PE replaced by C£
(and then £ replaced by a subset U of £) goes through in 1ST only if LI is
detachable. And for the third argum ent to go through in 1ST once P£ is
replaced by C£, it is necessary to show that the set B defined there is
detachable. In fact, as we shall see, these argum ents can break down
completely in 1ST even when £ is the set M of real numbers: it is in fact
consistent with 1ST that R has just the two detachable subsets 0 , R, so that
CR (and so also 2ii?) is isomorphic to 2. A fortiori CR is injectible into R,
showing that the third argum ent fails in 1ST.
4 In fact the countability of PN is consistent with 1ST, but the proof of this fact is
beyond the scope of the present book.
5 Here we write 0 for 0 and 1 for {0}.
6 On the other hand the natural bijection between Q£ and P£ is given, as it is
classically, by/»-*-/-i(0).
6
THE CONTINUUM
Accordingly we are reduced to possibilities (/) and (iv). In case (/), L and R
are disjoint, but since neither contains p, they together fail to cover X; while
in case (iv), L and R together cover X, but since each contains p, they are not
disjoint. This strongly suggests that a (linear) continuum cannot be
separated, or decomposed, into two disjoint parts which together cover it.
Herein lies the germ of the idea of cohesiveness.
Of course, this analysis is quite at variance with the account of the (linear)
continuum provided by classical set theory. There the continuum is takes the
form of the discrete linearly ordered set E of real numbers. "Cutting" E (or
any interval thereof) at a point p amounts to partitioning it into the pairs of
subsets ({x: x < p], {x : p < x}) or ({x: x < p\, {x: p < x}): the first and second of
these correspond, respectively, to cases (ii) and (iii) above. Now in the
discrete case, one cannot appeal to symmetry as before: consider, for
instance, the partitions of the set of natural numbers into the pairs of subsets
({n: n < 1}, {n: 1 < n}) and ({n: n < 1}, {n : 1 < n}). The first of these is ( {0, 1},
{2, 3, ...}) and the second ({0}, {1, 2,...}). Here it is manifest that the symmetry
naturally arising in the continuous case does not apply: in the first partition
1 is evidently a member of its first component and in the second partition, of
its second. In sum, when a discrete linearly ordered set X is "cut", no
7
ambiguity arises as to which segment of the resulting partition the cut point
is to be assigned, so that the segments of the partition can be considered
disjoint while their union still constitutes the whole of X.
But now let us return to our original analysis. This led to the idea that a
continuum cannot be decomposed into disjoint parts. Let us take the bull by
the horns and attem pt to turn this idea into a definition. We shall call a space
S cohesive or indecomposable, or a (genuine) continuum if , for any parts, or
subsets U and V of S, whenever U u V = S and U n V = 0 , then one of U, V
m ust = 0 , or, equivalently, one of U, V m ust = S. Clearly S is cohesive
precisely when its only detachable subsets are 0 and S itself.
We observe that in classical set theory, the only cohesive spaces are the
trivial empty space and one-point spaces. But it turns out that the existence
of nontrivial cohesive spaces is consistent with 1ST. In fact it is consistent
with 1ST that R itself is cohesibe. How does this come about? To get a clue, let
us reformulate our definitions in terms of maps, rather than parts. If we
denote by 2 the two-element discrete space, then connectedness of a space S
is equivalent to the condition that any continuous map S -» 2 is constant, and
cohesiveness of S to the condition that any map S -» 2 whatsoever is constant.
Supposing S to be connected and to possess more than one point, then from
8
LEM it follows that there exist nonconstant —and hence discontinuous —
maps S -» 2. But the situation would be decidedly otherwise if all maps
defined on S were continuous, for then, clearly, the connectedness of S
would immediately yield its cohesiveness. In fact it is consistent with 1ST
that all maps E -> E - and hence all maps R —» 2 - are continuous. It
follows that it is consistent with 1ST that E is cohesive.
Some years ago Paul Cohen published an article in the Scientific American
entitled Non- Cantorian Set Theory. There he describes the set theories in
which Cantor's continuum hypothesis is violated. In terming these set
theories non-Cantorian he was making an analogy with non-Euclidean
geometries in which the parallel postulate is violated. If, in Cohen's
analogy, non-Cantorian set theories correspond to non-Euclidean
geometries, then classical (Zermelo-Fraenkel) set theory corresponds to
neutral or absolute goemetry in which no form of the parallel postulate is laid
down. So we are led to reformulate Cohen's analogy by replacing the
continuum hypothesis with the Law of Excluded Middle; it is then
intuitionistic set theory that corresponds to neutral geometry. Intuitionistic set
theory can thus be seen as a "neutral" set theory, compatible with a num ber
of principles - such as the subcountability of N* and the cohesiveness of the
real line - which are incompatible with classical set theory.
9
cohesiveness of the real line - become clear when their meanings in the
models realizing them are understood.
10
Chapter I
\fx[x G A <-» c p (x )]
is called the defining axiom for the class A. Two classes A, B are defined to be
equal and we write A = B if
\/x(x G A - » X G B).
3u\fx(x G i 4 f > X G l i ) .
Set(A) asserts that the class A is a set. For any set u, it follows from the
defining axiom for the class {x: x g w ) that Set({x: x g u }). We shall identify
11
{x : x e u } with w, so that sets may be considered as (special sorts of) classes
and we may introduce assertions such as u c A , m = A, etc.
• {ul9...,un} = { x : x = u l v ... v x —u n\ .
• \ j A = {x:3y(y g A A X & y ) }
• f)A={x:Vy(yeA->xey}}
• A u B ~ { x : x g A v x g B}
• A n B = {x: x e A a x g
• A - B = {x : x g A a x £B}
• u" = u u {u}
• PA = {x: x e A} (the power class ofA)
• {xgA: (p(x)j = { x: x e A a cp(x)}
• {t(x ) : cp(x)} = {y : 3x. y = £(x) a (p(x)} where t(x) is a term such that
V x Set(£(x))
• V = {x: x = xj
• 0 = {x: x * x}
Union V m S e t([J u
Powerset V m Set(PM)
12
For a set u, Yu is called the power set of u' Until further notice all
propositions, theorems, etc. will be proved in IZ (using the above axioms
and intuitiuonistic logic).
We write 0 for 0 , 1 for {0} and 2 for {0, 1}. 2 carries the natural ordering <
given by 0 < 0, 0 < 1,1 < 1.
<u,v> = [{u],{u,v}}
Clearly we have
Proof. This m ust be proved w ithout using LEM. Suppose that <u, v> -
<a, b>.
Since {u} is an element of<u, v> , it m ust also be an element of the <a, b>, so
that either {u} = {a} or {u} = {a, b}. In both cases u = a.
Since {u, v) is an element of<w, v>. , it m ust also be an element of <a, b> , so
that either {u, z;} = {a} or \u, v\ = [a, b}. In either case v = a or v = b. If v = a
then it - a - v so that
{fa}} = {{w}/ {«/ v}} = <u, v> = <a, b> = {{a}, {a, b}}.
13
It is left to the reader as an exercise to show that, Set(A) and Set(B) implies
Set(A x B)
14
For sets A, B, the exponential BA is defined to be the set of all functions from
A to B. (Exercise: show that this is indeed a set.)
It follows easily from the axioms and definitions of IZ that, for any set A, PA,
under the partial ordering c=, is a frame7 with operations u , n and =>, where
For any set a, we write [a | cp} for {x: x - a a cp}; notice that
From Extensionality we infer that {a | cp} - [a | y} iff ((p <-> y); thus, in
particular, the elements of PI (recall that 1 = {0}) correspond naturally to
truth values, i.e. propositions identified under equivalence. PI is called the
frame of truth values and is denoted by Q. The top element 1 of Q is usually
written true and the bottom element 0 as false.
In IZ, Q plays the role of a subset classifier. That is, for each set A, subsets of A
are correlated bijectively with functions A -» Q. To wit, each subset X c A
is correlated with its characteristic function %x : A -» Q given by %x (x)~
{01x e X}; conversely each function fi A -» Q is correlated with the subset
fil(1) = {xeA: fix) = 1} of A.
LOGICAL PRINCIPLES IN IZ
(i) —> (ii). Assuming (i), we have VxVy (x g {y} v x £ {y}), whence (ii).
(ii) —> (iii) and (iii) -> (iv) are both obvious.
(iv) —> (v). Assuming (iv), for any co e Q we have « = {0} or co * {0}. In the
latter case co = 0 , and (v) follows.
(viii) -> LEM. Assume (vii) and let cp be any formula. Let U be the subset
{0|cp} u {1} of 2. Then 0 ell <-» cp; moreover U is inhabited and so has a least
element a , which m ust be either 0 or 1. If a - 0, then 0 g li, whence cp;
while if a = 1, then 0 € U, whence —.cp. LEM follows. ■
16
Observe that the negation operation on formulas corresponds to the
complementation operation on Q; we use the same symbol -i to denote the
latter. This operation of course satisfies (using co , co' as variables ranging
over Q)
But in IZ, this is far from being the case. In fact we can prove
0 e -co —^ 0 ^ co —^ 0 e —1co,
and from this we infer that -co c= -nco. Since, obviously, co u -co = true, it then
follows that, for any co, co u -iCO = true, which is LEM. ■
17
(i) WLEM
(ii) (2, <) is complete, i.e. every subset of 2 has a <-supremum.
Proof, (i) —> (ii). Assume (i) and let U c 2. Clearly 1 g LI *-» U c {0}. By (i),
we have 1 £ If or -n 1 £ If. In the first case U c: {0} and U then has
suprem um 0. In the second case ~iU ^ {0}, so that —V x g U. x = 0 , which is
equivalent to
V x e U ( x < 0) -> 0 = 1.
Also, obviously,
so that U has suprem um 1. Thus in either case U has a suprem um, and (ii)
follows.
(ii) -> (i). Assume (ii), let cp be any formula and define U = {1: <p} Then U has
a suprem um a and there are two cases: a - 0 or a = 1. In the first case 1 £ If,
so that -i(p. In the second case, if -icp, then If = 0 and so a = 0, which is
impossible. Therefore —i—icp, and (i) follows. ■
18
Proof. Let cp be any formula ; define U = {.re 2: x = 0 v cp} and V =
{xe2: x = 1 v cp}. Suppose given a choice fu n ctio n /o n {li, V). Writing a =/(U),
b =f(V), we then have a e U . b e V, i.e.
Hence
a = 0 a (b= 1 v cp),
whence
(*) a * b v cp.
But
cp —> U —V —^ a
so that
a ^ b —^ —.cp .
We fix some more notation. For each set A we shall write QA for the set of
inhabited subsets of A, that is, of subsets X of A for which 3x (x e X). The
class of functions with domain A will be denoted by Fun(A).
• SLEM a v -id
• Lin (a -» ß) v (ß -» a)
• SWLEM -id V -T-ld
• Ex8 3x[3xcp(x) -» cp (j )]
19
• Un 3x[(p(x) —>Vxcp(x)]
• D is9 Vx[a v (p(x)] -> a v Vxcp(x)
Here a and ß are sentences, and cp (x) is a formula with free variable x. In
intuitionistic logic, Lin and SWLEM are consequences of SLEM; and Un
implies Dis. All of these schemes follow, of course, from LEM, the full Law
of Excluded Middle.
The first two of these are forms of the Axiom of Choice for X; while
classically equivalent, in IZ AC*x implies ACx, but not conversely. The
*
principles DACx and DACXare dual forms of the Axiom of Choice for X:
* *
classically they are both equivalent to ACx and AC^ , but in IZ DAC^
implies DACx, and not conversely.
Here cp(x) , vp(x) are formulas with the free variable x. This principle asserts
that, for any pair of instantiated properties of members of 2, instances may
be assigned to the properties in a m anner that depends just on their
extensions. WESP is a straightforward consequence of AC q2- For taking 0(w,
y) to be y e u in AC q2 yields the existence of a function / w i th dom ain Q2
such that fn e u for every u e Q2. Given formulas cp(x), ip(x), and assuming
the antecedent of WESP, the sets U = {xe2: cp(x)} and V - {xe2: y(x)} are
20
members of Q2, so that a ~ f U e U, and b = JV e V, whence cp(fl) and \\j(h).
Also, if Vxe2[ip(x) <-» \]/(x)], then LI - V, whence a = b; it follows then that the
consequent of WESP holds.
cp(x) = x = 0 v a ip (x) = x - 1 v a.
For the converse, we argue informally. Suppose that SLEM holds. Assuming
the antecedent of WESP, choose a e 2 for which cp(a). Now (using SLEM)
define an element b e 2 as follows. If Vxe2[cp(x) <-» holds, let b = a; if
not, choose b so that ß(fr). It is now easy to see that a and b satisfy cp(fl) a v|/(Z?)
a [Vx e 2[cp(x) <-> i]/(x)] —>« = //]. WESP follows. ■
*
Proposition 7. ACt and Ex are equivalent over IZ.
*
Proof. Assuming AC| , take 0(x,y) = cp(y) in its antecedent. This yields an
f e Fun(l) for which Vy(p(y) -» cp(/D), giving 3i/[3i/cp(i/) -> cp(y)], i.e., Ex.
Conversely, define cp(y) = 0(0,y). Then, assuming Ex, there is b for which
3y(p(y) -> cp(fr), whence Vxel3y0(j,y) -> \/xelQ(x,b). Defining f e Fun(l) by
/ = [{Orb}} gives Vx e 13i/0(x, i/) -> Vxel0(x,/x), and ACj follows. ■
21
Further, while DACi is easily seen to be provable in IZ, we have
k
Proposition 8. DACj and Un are equivalent over IZ.
Proof. Given cp, Define 0(x,y) = ip(y). Then, for / e Fun(l), 3xelQ(x,fx) <->
cp(/D) and E k e lVy0(x,y) <-> Vycp(y). DAC} then gives
3/eFun(l)[<p(/0) -» Vycp(i/)],
Conversely, given cp, define cp(y) = 0(O,y). Then from Un we infer that there
exists b for which cp(fr) -> Vycp(y), i.e. 0(0,b) -> Vy0(O,y). Defining / e Fun(l)
by / = {(0,1b)} then gives 0(0/0) -> 3 x e l Vy0(x,y), whence 3 xel Q(x,fix) ->
3xelVy0(x,y), and Un follows. ■
V/eFun(2)[0(O,/O) v
which, in view of the natural correlation between members of Fun (2) and
ordered pairs, is equivalent to the assertion
22
Vi/Vi/lcp(y) v V(y')l -* Vy(p(y) v VyV(y'),
where y does not occur free in ip, nor y' in cp. And this last is easily seen to be
equivalent to Dis. ■
k
Now consider DAC? . This is quickly seen to be equivalent to the assertion
Now (*) obviously entails Un. Conversely, given Un, there is b for which
\\f(b) ->Vyv|/(y). Hence a v \\i(b) -> a v Vyvp(y), whence (*). So we have proved
•k •k
Proposition 10. Over IZ, DAC2 is equivalent to Un, and hence also to DACj .■
*
Clearly ftC ^ is equivalent to
k
and similarly for w a C A Then we have
23
* *
Proposition 11. Over IZ, B,C[ and w a C j are equivalent, respectively, to Lin and
SWLEM.
*
This yields a c t -> Lin. For the converse, define a = 0(0,0) and ß = 0(0,1) and
reverse the argument.
24
Chapter II
The natural numbers can be defined in IZ and their usual properties proved.
It follows from the axiom of infinity that there exists at least one inductive set A
Define
N = P)(X : X c A a X is inductive}.
Then N is inductive and is clearly the least inductive set, that is, N c; K for
every inductive set K. The members of N are called natural numbers; thus 0, 1, 2
are natural numbers. We shall use letters m, n, p, ... as variables ranging over N.
Proposition 1.
(i) m e n-> m+ c n.
(ii) n en
(iii) m+ = n+—» m = n.
Proof, (i) Let K = | n: \fin(m e n -> m+ c n).To prove (i) it suffices to show
that K is inductive. Clearly O g K Suppose now that n e K. Then (*)
(*) rn g n m+ c n.
25
If m e n+then me n or m=n. In the first case m+ c= n by (*) and in the second
case m+ - n+ . Thus in both cases m+ c: n+ and so n+ e K. Hence K is inductive
and (i) is proved.
We shall sometimes write m < n for m e n and m < n for m c n. It follows from
Proposition 1 that m+ < n <-> m < n.
me n, m = n, n e m.
Proof. Proposition 1 implies that any two of the above assertions are mutually
contradictory. To prove that, for every m, n, one of these assertions holds we
define
We need to show that K(n) = N for every n, and for this it suffices to show that
K(n) is inductive.
The set K(0) is inductive, since K(0) = {0} u{m: 0 g mj and it is obvious that
0 e m —> 0 g //7+.
Now suppose that K(n) is inductive, i.e. N c K(n). We show that K (n+) is also
inductive.
26
Q c K (n+). From the fact that K(0) is inductive, it follows that n+ e N c; K(0),
whence n+e 0 or //+ = 0 or 0 e n+. The first two disjuncts are false, so 0 e n+ and
0 e K (n+).
Corollary. N is discrete. ■
PI VpeA . sp ^ o
P2 VpeA \/qeA. sp = sq -» p = q.
P3 VX c A[[o e X a Vp(p g X -> sp e X)] -> X - A]
(P3) is the Induction Principle for A: it is clearly equivalent to the scheme: for an
formula cp(x),
10 Here “o" (Greek omicron) is not to be confused with the set 0. It is just an arbitrary
member of P.
27
A subset K of A satisfying o e K a \/ x (x gK sx g K) is called an inductive
subset of A: the Induction Principle asserts that the only inductive subset of A is
A itself. Establishing that a subset K is inductive is called a proof by induction.
sp * p
Vp g A[p = o v 3q. p = sq ].
If we define s: N -> N by sn = n+, then it follows from the fact that Prop.l (iii)
and the fact that N is the least inductive set that (N, s, 0) is a model of Peano's
axioms.
DEFINITIONS BY RECURSION
Given any set X, any element a g X, and any function e: X -> X, there exists a unique
fiinction f A -» X such that
Proposition 3 Any model ofPeano's axioms satisfies the Simple Recursion Principle.
Define
28
Let / = p | U . We claim that / satisfies the conditions of the proposition; its
uniqueness is left as an exercise to the reader. To show that / satisfies the
conditions of the proposition it clearly suffices to show th a t/is a map from A to
X.
so that
V x « o ,x ) g / -> x = a ) ,
29
Given p eK satisfying ( p 9x) g f a ( s p ,y ) g / , define
(2) (sq,e{y)) e j
and
Assertion (2) follows from the assum ption that (q9y) e J . As for (3), if sp = sq,
then, by P2, p = y, so from (g,y) g 7 it follows that ( p 9y) g 7 • But we are
assuming that p e K and (p 9x) gj , so we conclude from the defining property
of K that x = y. Hence, certainly, e(x) = e(y), proving (2).
(p g K a (p ,x ) g f a (sp 9y ) g / ) -> g U
whence
- > y = e(x).
Now, we know that V = A, i.e. V p3z.(p,x) e f . it follows from this and (1) that
30
(p eK a (sp,y) e / a {s p ,z ) g / ) -> 3x.(p,x >e / a p e K a (sp,p> e / a <sp,z> e /
-> 3x ( p G K A < p ,x ) e / A (sp,y) G / A ( s p ,z > e / )
-> 3x(iy - e(x) = e(x))
y ~ z.
An isomorphism between two Peano structures A = (A, s, o) and A' = (A', s, o') is
a bijection /: A -» A' such th at/(o) = o' and for all p e A, f(s(p)) = s'(f{p)). Two
Peano structures are isomorphic if there is a an isomorphism betw een them.
Proof. Given two Peano structures A = (A, s, o) and A' = (A', s', o'), by Prop. 3
they both satisfy the Simple Recursion Principle, so that there are maps f A A'
and g: A' -> A such that
31
Let us say that a Peano structure A = (A, s, o) satisfies the Extended Recursion
Principle if:
given any set X, any element a e X, and any function e : X x A —» X , there exists a
unique function fi A —» X such that
Proposition 4. Any Peano structure that satisfies the Simple Recursion Principle also
satisfies the Extended Recursion Principle.
h = (x,p)\~>{g({x,p)),p).
Proposition 5. Any Peano structure that satisfies the Simple Recursion Principle is a
model of Peano's axioms.
32
need to verify that A satisfies PI, P2 and P3. Again vve use p, q as variables
ranging over A.
Then
since 0 * 1, PI follows.
Then
and P2 follows.
P3: the Induction Principle. Suppose K ^ A is inductive . Then p •-> sp: K -> K
and so the Simple Recursion Principle furnishes a m ap/: A -> K such that
33
But the identity map ty: A -> A also satisfies (*), so from the uniqueness
condition in the Simple Recursion Principle we infer that j f = l A. It follows
easily from this that K - A, and the Induction Principle follows. ■
A set A is Dedekind infinite if there exists a monic map f: A -> A and an element
a e A such that a & ran(/). We next note that each Dedekinf infinite set gives rise
to a model of Peano's axioms.
Proposition 6. Each Dedekind infinite set contains the domain of a model of Peano's
axioms.
U = f]{u c : A : a e u a Vx g u . f( x ) g u }.
Corollary. A set A is Dedekind infinite if and only if there exists an injection N -> A.
FINITE SETS
There are a num ber of possible definitions of the concept of finite set in IZ. To
introduce (some of) these, it will be convenient to fix a set £. By an E-family or E-
singleton we shall mean ’’set of subsets of £", or"singleton of £ M, respectively. For
a subset X of £ we define
34
M(X) = X is in every E-family containing 0 g ;>and closed under unions with
disjoint E-singletons, that is, if VXg ,> Vxg E-X(X u {x} g ./]. If M(X) holds, we
shall say that X is a strictly finite subset of £.
Proof. Obvious. ■
Proof. Clearly L(X)-> K(X). To prove the converse, it suffices to show that the
family 7,=z {X^E: L(X)} is closed under unions of pairs. To this end let O(LE) be
the property V X g c\J U u X g 7i It suffices to show \ZU[L(U) —> O (U)]. Clearly
0 (0 ). Assuming 0(1/) and X g 9Xve have U u X g ^ a n d so II u X u {x} g '/
for arbitrary x, whence 0(1/ u{xj). Hence \/U[L(U) -> 0(1/)] arid the result
follows. ■
Proof. Obviously D (0). If D(X) and x £ X, clearly D(X u {x}). The result follows.
Proof. Write O(X) for the condition following the first implication. Clearly 0 (0 ).
Suppose that O(X) and x £ X. If D(X u {x} u {*?}), then D(X u {a}), so, since O(X),
either a e X v a £ X. Since D(X u {x} u{a}), it follows that a = x v a * x. Hence
(a g X a a = x) v (a £ X a a = x) v (a g X a n ^ x) v (a £ X a a * x),
The first three disjuncts each imply a g X u {x}, and the last disjunct means
a £ X u {x}.
Accordingly a g X u {x} v a <£ X u {x}. We conclude that 0(X u {x}) and the
result follows. ■
35
Lemma 5. VXc=£ \L(X) a D(X) -> M(X)].
Proof. We need to show VX[L(X) -» 0(X)J, where <£>(X) is D(X}-> M(X). Clearly
<£>(0). Assume <£>(X) and D(X u {a}). Then D(X), so, since O(X), it follows that
M(X). Since D(X u {a}), Lemma 4 gives fl e X v a € X. In either case we deduce
that M(X u {a}). Hence 0(X u {a}), and the result follows. ■
Proposition 7. For any set E, the families of strictly finite, discrete finite, and discrete
Kuratowski finite subsets coincide. ■
We can now define a set £ to be strictly finite, finite, or Kuratowski finite if it is,
respectively, a strictly finite, finite, or Kuratowski finite subset of itself.
Proposition 8. A set is strictly finite if and only if it is bijective with a natural number.
Proof. Suppose that £ is strictly finite, arid for X c £ let O(X) be the property
"X is bijective with a natural number. We need to show that 0(£), and for this it
suffices to show that, for all X c E, M(X) -» <t>(X). Clearly we have 0 (0 ). If O(X)
and x e E - X, le t/: n -» X be a bijection between some natural num ber n and X.
It is easily checked that the set g = / u {<n,x>} is a bijecfion between n+ and
X u {x}. Hence 0(X u {x}) and the result follows.
36
Finally, a set £ is Dedekiud finite if it is not Dedekind infinite, i.e. if there does not
exist a monic / : £ —»£ and an element a g £ such that a £ ran(/).
It is obvious that 0 is Dedekind finite. To prove (*), suppose that a <£ X and
X u {a) is Dedekind infinite. We show that X is Dedekind infinite. Since X u {a\
is Dedekind infinite, there is a m onic/: X u {ö) -> X u {a} and b e X u {a} such
that b £ ran if). There are two cases: b = a or b e X. In the first case,
/ X u {a} -» X ; the restriction / 1X : X -> X is then monic and
f(a) g X - mn(f\ X), so that X is Dedekind infinite. The second case, b g X, splits
into two subcases: f(n) = a or f(a) g X. In the first subcase, / | X : X -» X is
monic and b g X - ran(/*| X), so that again X is Dedekind infinite. In the second
subcase, f(a) g X, there is a unique xo g X for which /(xo) = so that
xo £ / l[X] and it is then easily shown that X = / 1[X] u {x0}.Now define
9 = f \ r 1[ X] v { ( x 0,f(a)\
Proof. In the proof we shall use the easily established fact that for co,u g Q,
37
We first prove
First we have
Conversely
as required.
By Frege's Theorem is m eant the result, implicit in Frege's Grundlagen, that, for
any set £, if there exists a m ap v from PE to £ satisfying the condition
38
VXV Y[ V(X) = v(Y) <r>X * Y]n ,
then £ has a subset which is the domain of a model of Peano's axioms. We are
going to show that a strengthened version of this result can be proved in IZ.
A Frege structure (£, v) is strict if dom(v) is the family of strictly finite subsets of
E.
We now prove
Frege's Theorem. Let (£, v) be a Frege structure. Then we can define a subset N of E
which is the domain of a model of Peano's axioms.
Thus suppose given a Frege structure (£, v).The proof of Frege's Theorem breaks
down into a sequence of lemmas.
Principle of v- Induction for A For any property O defined on the members ofJV, z/O
is v- inductive, then every member of Alias O.
Proof. Write O(X) for this assertion. To establish the claim it is enough, by the
Principle of v- Induction, to show that O is v-inductive. Clearly 0 (0 ). If O(X)
and v(X) g X, then evidently 0 ( Xt). So O is v-inductive. ■
Proof. Writing O(X) for this assertion, it suffices to show that O is v-inductive.
Clearly 0 (0 ). Now assume O(X) and x g Xt. Then either x g X, in which case,
since O(X) has been assumed, there is Y g N for which x = v(Y) and Y c= X, a
fortiori Y c Xt. Or x = v(X), yielding the same conclusion with Y = X. So we
obtain 0(X +), O is v- inductive, and the Lemma follows. ■
Proof. Assume the premises and let / be a bijection between X u {x} and
Y u {i/}. We produce a bijection f between X and Y. Let i/' be the unique
element of Y u {]/} for which <x, i/’> g /. Then either i/' = y, in which case we
t a k e / = / | X, or y' g Y, in which case the unique element x' g X u {x} for
which <x',y> g / satisfies x ! g X. (For if x' = x then <x,y> g / , in which case y' =
y ^ Y.) So in this case we define
/ = \f n (X x Y)] u \<x',y'>\.
Proof. Write O(X) for the assertion X g JVand VY g *A[v (X) = v(Y) -> X = Y]. It
suffices to show that <t> is v-inductive. 0 ( 0 ) holds because v(0) = v(Y) -> Y ^ 0
40
-> 0 = Y. Now assume that <h(X) and v(X) £ X; we derive <£>(Xt). Suppose that
Y e N and v(Xt) = v(Y). Then X+ Y, and so in particular Y ^ 0 . By Lemma 1,
there is Z e N for which v(Z) £ Z and Y = Zt, so that X+« ZL We deduce, using
v(X) £ X.
Proof. It suffices to show that the property v(X) & X is v-inductive. Obviously
0 has this property. Supposing that X e Af v(X) £ X but v(X+) e Xt, we have
either v{Xt) = v(X) or v(Xt) e X. In the former case X = Xt by Lemma 4, so that
v(X) g X, a contradiction. In the latter case, by Lemma 2, there is Y e «yLsuch that
Y c X and v(Xt) = v(Y). Lemma 4 now applies to yield Xt = Y c X, so again
v(X)e X, a contradiction. Therefore v(X) g X -> v(Xt) ^ Xt, and the Lemma
follows. ■
Notice that it follows immediately from Lemma 5 that A ls closed under f, that is,
X G e\ Xt G A.
Now define o = v(0), N = (v(X): X g A}, and s:N -> N by s(v(X)) = v(Xt) for X
g A. Then s is well defined and monic on N. (For if v(X) = v(Y), then, by Lemma
X « Y, whence v(X) = v(Y).) Clearly, also, o * sx for any x e N. The fact that the
structure (N, s, 0) satisfies the Induction Principle follows immediately from the
Principle of v-induction for N. Accordingly (IV, s, o) is a model of Peano's
axioms, as required.
Proof, (i) Let K = {q: Vp[p ek(q) -> k(p) c k(q)]. Obviously o e K. If q e K, then
p efc(p) -> k(p) e k(q), so that
Proof. Write 4>(p) for Vq[k(p) « k(q) *-» p = p]. Then clearly O(o). If ct>(p) and
42
Since, By (i) of Lemma 7, r <£ k(r) and p £ k(p), Lemma 3 implies that k(r) ^ k(p),
so, since O(p), r = p and q = sr = sp. Hence O(sp), and the result follows by
induction. ■
Proof. We need to show that dom(v) is the least family of subsets of £ which
contains 0 and is closed under unions with disjoint £-singletons: let us call such
a family strictly inductive. First, dom(v) clearly contains 0 . If X e dom(v), then
X % k(p) for some p. If x £ X, then X u [x] zz k(p) u {p} = k(sp), whence
X u jx} c dom(v). So dom(v) is strictly inductive. And dom(v) is the least strictly
inductive family. For suppose that <>is any strictly inductive family. For each p
let .I? p = {X: X « g(p)\- We claim that .W p c F for all p. For obviously
J f0 = {0} c; F. Now suppose that c: 3. If X ^ k(sp), then X zz k(p) u {/?}, so for
some x e X (which may be taken to be the image of p under a bijection between
k(p) u {p} and X), we have X - |x| ^ k(p). It follows that X - {x} e JCyn c //, and
so X - (X - {x}) u {x} e ca The claim now follows by induction; accordingly
dom(v), as the union of all the is included in F. Therefore dom(v) is the least
inductive family and the Lemma is proved. ■
Proof. Suppose that <X, p> e v and <X, q> e v. Then X ^ k(p) and X ^ k(q)
whence k(p) ^ k(p) and so p = q by Lemma 7. The remaining claim is obvious. ■
43
Proof. We have, using the previous Lemma, v(X) = v(Y) <-» k(v(X)) zz k(v(Y)) <->
X « Y. ■
(E,v) is called the strict Frege structure associated with £ and the model A of
Peano's axioms.
Finally, we show that the processes of deriving models of Peano's axioms from
strict Frege structures and vice-versa are m utually inverse.
Suppose that we are given a strict Frege structure (£, p). Recall that the
associated model (A ,s, o) of Peano's axioms is obtained in the following way.
First, the family A is defined as the least subfamily of dom(p) containing 0 and
such that, if X e «4^and p(X) £ X, then X u {p(X)} e */h it having been shown that
p(X) g X for all X e Jf. The associated model (A ,s, o) of Peano's axioms was
then defined by N ~ {p(X): X e df), s(p(X)) = p(X u p(X)}), and o = p(0).
We observe that since (£, p) is strict, for any X e dom(p) there is a (unique)
X* e e^for which X ä X*, and so p(X) = p(X*). To prove this, it suffices to show
that the set of X e dom(p) with this property contains 0 and is closed under
unions with disjoint singletons. The first claim is obvious. If X e dom(p), x € X,
and X rz X* with X* e Jt, then
X u { i } ^ r u { p ( X l ! edf,
since, as observed above, p(X*) ^ X*. This establishes the second claim, and the
observation.
Now let (£, v ) be the strict Frege structure associated with the model (A ,s, o) of
Peano's axioms in turn associated with (E, p). We claim that p = v. To prove this
it suffices to show that
44
(*) X % fc(n(X)) for all X e Jt,
(*) now follows from the definition of A. So our claim that p = v is established.
Conversely, suppose we are given a set £ and a model (A ,s, o) of Peano's axioms
with A c E , Let (E, v) be the associated strict Frege structure. We note first that,
for any p e A, we have v(/c(p)) = p. For by Lemma 9, k(p) ^ k(v(k(p)), so that, by
Lemma 7, p = k{y(p)). Now let (A5" ,5* o*) be the model of Peano's axioms
associated with the Frege structure (£,v). We claim that (A ,s, o) and (A* ,5*, o*)
are identical.
so that s* = s by induction.
Thus we have established that the two processes are m utually inverse.
45
Chapter III
We turn now to the construction of the real num bers in IZ. This is done in
essentially the classical manner: first, the (positive and negative) integers are
constructed, next, the rationals, and finally the reals obtained from the latter as
Dedekind cuts or (equivalence classes of) Cauchy sequences. In the classical
context it is well known that these methods of constructing the reals lead to
isomorphic results. This need not be the case in IZ. Neither is it necessarily the
case that the reals are (conditionally) order-complete, or even discrete.
usual w ay1213. It is shown in the standard m anner that Z may be turned into an
The customary procedure for obtaining the rational field as the ordered field of
quotients of Z 13 now yields, in IZ, the ordered field of rationals (Q , +, *, <) . Note
over Q.
Now we can define the Dedekind real numbers as "cuts" in Q. Thus a Dedekind
12 See, e.g. Mac Lane and Birkhoff [1967], Chapter II, section 4.
13 Mac Lane and Birkhoff (1967), Chapter V section 2.
46
(1) LHR =0
(2) Vp[p gl <->3q e L.p < q]
(3) Vp[p g R <-» 3q g R.q < p]
(4) V p V q [p < q -+ p e L V q Gr ]
We also define a weak real number to be a pair is a pair < l , r > of inhabited
subsets L, R c Q satisfying (1), (2), (3) and the pair of conditions (both weaker
than (4)):
We write Ew for the set of weak real numbers. Clearly Ed c Ew. We shall use
letters r, s as vas variables ranging over EW/ and write r - < L r, R r>.
Note that, for any weak real num ber r, R r is recoverable from L r in that
r <S <-> Lr Cj Ls .
44 The asymmetry in this definition is only apparent since from the recoverability of R
from l , noted above, follows that L.r c; l s <-> Rs c: r ,.
47
Clearly < is a (partial) ordering on Rw (hence also on Rd), and it is
While classically it can be shown that r <s o r < s v r - s, this does not hold
in IZ. What can be proved is
(P e Rr a p e Le) v (q e L. a <2 e R j :
In the classical case one now proceeds to show that Rd is conditionally order-
complete, i.e. every inhabited subset with an upper bound has a least upper
bound. The argum ent for the conditional order-completeness of Rd requires an
application of LEM which is not available in IZ. On the other hand, we shall
prove in IZ that Rwis conditionally order-complete. Moreover, we shall show in
IZ that, if De Morgan's Law —.(cp a vp) -> (—.tp v —<\p) holds, then Rd is
48
Proposition 1. In IZ, Mwis conditionally order-complete.
R = {q :3 p < qVr e X (p e R r )}
It is easy to verify that rx satisfies conditions (1), (2), (3) and the second
condition in (4') above. To verify the first condition in (4'), suppose that
p <q a p £ l . Define
Then p <p' < q < q. From the definition of L it follows that -r-i(p' e R) . But since
R c: R, for all r e X, it m ust be the case that R n Lr = 0 for all r e X, whence
Proof, (i) —> (ii). Assume (i) arid let r e Rw. To show that r e Ed, suppose given
rationals p, q with p < q and let e = Vi(p + q). From the disjointness of L r and R,
together with DML, it follows that e ^ R r v e a r . Gut since p < e the first
49
disjunct implies p e L r and the second similarly implies q e R r It follows that
r e Ed, whence (ii).
(iii) -> (i). Assume (iii) and let cp be any formula. Then the set
S - {q e Q : q = 0 v (q = 1) a cp}
cp -> r = 1 a n d —.cp —> r = 0 , so that r < 1 -> —.cp a n d 0 < r —» —.—icp. By (ii) of
Lemma 1, 0 < r v r < 1, and we deduce that —i—i(p v —«cp. WLEM and hence
DML follows. ■
Addition on E j is defined by
16 In fact the procedure introduced here works equally well for Ew.
50
To define the product of r19r2 e E d , we use the idea that if p x - r;a n d p 2 - r 2
have the same sign then (rx - p^)(r2 ~ p 2) should be positive. Thus we define
Lr
ri rri to be
and r to
l ri o be
Finally, we discuss the Cauchy reals. These are obtained as equivalence classes of
Cauchy sequences of rationals and, in classical set theory, the resulting field can
be shown to be isomorphic to Ed. As we shall see in Chapter IV this is not
51
I f ( n) ~ f (6n) \< l/ n, | f(6n) - g(6n) |< l/ 2 n , | g(6n) - h(6n) |< 1/ 2n,
defining them on C:
and
(f g ) ( n ) = f ( k n ) + g (k n ),
We define a map i: C —» Ed by
(To verify that i(f) is indeed a Dedekind real, we note that for given p < q in
Q we can choose n with 3/ n < q - p, and then we m ust have either p <f(n) - 1 / n
or q >f(n) + 1/n.) Moreover, it is not hard to show that
so that i induces a monic map j : Ec -> Ed. This map is a ring homomorphism.
52
Now j is not necessarily an isom orphism 17. But it is one in the presence of the
countable axiom of choice AC( N )18. For from AC( N ) it can be deduced that every
Dedekind real is the limit of a Cauchy sequence of rationals, since for each
Dedekind real r and each n > 0 we can find a rational p with | r - p | <\/n. This
follows from the fact that, since l , and R r are inhabited, we can find rationals p,
q for which p < r <q; then the interval [p, q] can be divided into finitely many
subintervals of length < 1/ n.
Thus, for example, the ordinals 1, 2, 3, ... as well as the first infinite ordinal co to
be defined below, are all simple. Every subordinal (hence every element) of a
simple ordinal is simple. But, in contrast with classical set theory,
In classical set theory the e-induction scheme is equivalent to the a x i o m o f r e g a l a r i t y , which asserts
that each nonempty set u has a member x which is e - m i n i m a l , that is, for which x n u = 0 . It is easy
to see that this implies LEM: an e-minimal element of the set {0 | tp) u {1} is either 0 or 1; if it is 0, (p
must hold, and if it is 1, (p must fail; thus if foundation held we would get cp v —tip.
54
intuitionistically not every ordinal can be simple, because the simplicity of the
ordinal {0, {0 | cp}} implies cp v -Kp.
a < ß <—
>a gß a < ß <e> a c: ß.
It is now easily shown that successors and suprem a of ordinals are again
ordinals and that
But straightforward argum ents show that any of the following assertions (for
arbitrary ordinals a, ß, y) implies LEM:
( i ) a < ß v a ~ ß v ß < a,
(ii) a < ß v ß < a,
(iii) a < ß - > a < ß v a = ß,
(iv) a < ß -> a +< ß v a = ß,
(v) a < ß < y -> a < y.
55
An ordinal a is a successor if 3ß a = ß+, a weak limit ifV ßea 3 yea ß e y, and a
strong limit if V ßea ß+e a .
Note that both the following assertions imply LEM: (i) every ordinal is zero, a
successor, or a weak limit, (ii) all weak limits are strong limits. For (i) this
follows from the observation that, for any formula cp, if the specified disjunction
applies to the ordinal (0 | cp}, then cp v -—
.cp. As for assertion (ii), define
56
r k . (x ) = } J r k (yf .
y-x
When < is e restricted to an ordinal, it is easy to see that the associated rank
function is the identity.
and extensional if
It is easily shown that the well-orderings are exactly those relations isomorphic
to g restricted to some ordinal. For it follows immediately from the axioms of
g -induction and extensionality that the e -relation well-orders every ordinal.
And conversely, it is easy to prove by induction that the rank assigning function
on any well-ordering is an isomorphism.
D efinition. The rank function p: V -> ORD is defined by recursion on the well-
founded relation e by the equation
p(x) = IJp(iy) .
yex
Kt = U pV
ßcu
57
Proposition 1.
(i) V x p (x )e O R D
(ii) x e y -> p(x) < p(y).
(Hi) a < p ^ c V p.
(iv) x ^ y < = V n - > x e V a .
(v) Va p(a) = a.
(vi) Vx x gV .
([U = W|=T
lu = v l = lv = u l
l u = V II A \\v = w $ < l u = w §
I u = u i A IIU e w t< lu G w 11
lv = w} A lu G v} < |u G w\\.
for u, v, w g S.
Then we define
21 w hen H is the frame C{X) of open sets in a topological space X„ WO is called a spatiial
extension.
59
V {H) ^ { x : 3 a [ x e V j H)}.
VxeVW (p(x).
In particular we have
60
(*) f3x<p(x)S = V flkp(u)lj fVx(p(x)]l= A [[<p(u)|!
u c V aiH) u c-iV jH)
Vx e A3ß3u e V ^ . x - §cp(u)j].
If we let a = [J u , then
Vx g ABu g Vt[H).x = I'(p(u)|]
An argum ent of this general sort will be called a Collection argument; we shall
tacitly employ a num ber of such argum ents in the sequel.
II3x g tup(x)| - v
xrdom(w)
|cp(x)l |Vx g ucp(xJS — A
x^dotnfu)
I<p(*)l •
Note that, given an H-set set u, 3x(x e u) does not necessarily imply that
there is an H-set set v for which N= v g u. An H-set u satisfying this latter
condition is called inhabited, and an H-valued set v satisfying t= v g u is
called a definite element of u.
x A = {< y Aj>: y e x }.
61
It is then easily shown that, for x e V, u g VW,
I u e x A5 = Vlu = y Ai.
yex
It follows that
a e b —> | a A g foA| = t
and
It follows immediately from these that the H-valued set 0 A represents the
natural num ber 0 (i.e. 0 ) in %S(H\
(*) I7nd(u)|l<lnAG u |
| Ind(u) a n A g u j < f n +A g u |
Therefore
62
Note that it follows in particular that the Axiom of Infinity holds in W A
It is not in general true that if 1= 3x(p(x), then ®(H) f cp(z/) for some
u e VW. As we show below, certain conditions on H will ensure that this holds.
Mixing Lemma. Let {ut :i e I j c V(H) and suppose that {at : i e 1} is an antichain in
H. Then, writing u for ^ a . • u x, we have a, < |Ju = u{tfor all ie I.
63
Proof. We have, for given z e /, Ju = uxf) = a a b , where
Hence a, < [u(x) => l x gw J for all x g dom(w), so that a. < a . On the other
hand, if x g dom(i//), then
so that a- < [u-(x) => | x g u - J], whence «/ < b. Hence a. < a a b and the result
follows. ■
64
Proof, (i). We first prove the following:
(*) For any H-set v and any complemented a e H such that a < \\v eu% there is a
definite element we U such that a < \v =
Suppose that v and a satisfy the hypothesis. Choose a definite element z of u and
let w = a - v + a* -z. Then a < \\v - w \ and a* <\\z = w j follow from the
Mixing Lemma. Moreover, we have
that is,
A iq>Mi a [lv e u j < lcp(u)l].
w-:U
Snce H s totally disconnected, to prove this it is enough to show that, for any
complemented element a e H,
65
(ii). Obviously (a) implies (b). Conversely, if (b) holds, then y\ lq)(u;)S =T and it
we l l
A core for an H- set u is a set C c V(H) such that (i) each member of C is a definite
element of u; (ii) for any definite element y of u there is x e C such
thatflx = t/| = T.
By a Collection argum ent there is a set W c= VW such that, for any xe VW, there
is y g tV for which ax - %. It is easily shown that the set
C = {x gw : [fx g u l =t } is a core for u.
Now in IZF it can be shown that is an H-valued model of IZF. It was shown
above that the Axiom of Infinity holds in We further verify the Axioms of
Separation, Collection and g -Induction in , with brief comments on the
verification of (some of) the remaining axioms.
To begin with, we note that, given H-sets u, v the H-set {u, vYH) =
{< u, t >, < v,T>} is easilv shown to validate the Pairing Axiom in V(H).
66
To see that each instance holds in $$(*0, let u e V(H), define v e VW by dom(y) =
dom(//j and, for x e dom(z;), i?(x) = u(x) a |cp(x)i|. Then we have
[| Vx[x GV <r>X GU A Cp(jC)l - fVx G v[x GU A (p(x)]l A [Vx Gu[(p(x) -> XGV\\.
Now
Similarly
I Vx G u[(p(x) -> X G vl =T
such that, for all x g dom(zz), we have ff3zycp(x, t/)U < V fcp(x, z/)Jj . If we now
ycVa'w>
define z; g VW by v = x Jt } , then
67
The e- induction axiom is, we recall, Vx[Vy e xcp(y) -> <p(x)j —> Vx<p(x). To
verify the truth of this in SJW, first put
It now suffices to show that, for any x e V W , a < [cp(x)|. To do this we use the
induction principle for V(H). Assume for y e dom(.t) that a < H<p(y)| Then
as required.
To establish the truth of the Axiom of Union in $MH), given u e VW, define
v g VW by dom(z>) = (J dom(i/) and, for x g dom(z;), v(x) = §3y g u .x g y]\.
yedom(u)
For the Power Set Axiom it can be verified that, in £$(**), the power set of a set
u g VW is given by the set v g VW defined by dom(i?) = Hdom(") and, for
x g dom(u), v(x) = [[x g u J When w is of the form a , v may be taken to be the
Now it is readily shown that LEM holds in £$(**) if and only if H is a Boolean
algebra. Since, as we have se en , in IZ the Axiom of Choice implies LEM, it hold
in in fact, it cannot hold in 3$(H) unless H is a Boolean algebra. This is far
from being the case for Zorn's lemma, however, despite the fact that it is
classically equivalent to AC. Indeed, we will show that, in IZF, Zorn's Lemma
implies its truth in any 3S(f/). We shall take Zorn's lemma in the form: any inhabited
partially ordered set in which every chain has a supremum also has a maximal element.
68
Thus suppose X, < x e V(H) satisfy
W*) < X, <x > is an inhabited partially ordered set in which every chain has a
supremum.
Let X' = {x e V(H) : [jx e X[j ^ t } be a core for X and define the relation < \ on X
by X <Y>x ' <-> | x <x X l =T. It is then easily verified that <X', < x> is an
inhabited partially ordered set in which every chain has a suprem um. So, by
Zorn's lemma, X' has a maximal element c. We claim that
To prove (1) we take any a e V(H) and define Z e V(H) by dom(Z) = dom(X) and
69
(3) and (4) yield |a e X a c<x a] < [a = c\; since this holds for arbitrary a e V(H>,
(1) follows.
From the fact that Zorn's lemma holds in every %$(H) but AC does not we may
infer that, in IZF, the former does not imply the latter. In IZF Zorn's lemma is
thus very weak, indeed so weak as to be entirely compatible with intuitionistic
logic. For more on this see Bell [1997].
We shall exploit this last fact in the following way. Suppose we w ant to show
that a certain sentence a of our set-theoretic language is relatively consistent
with IZF. We produce a certain frame H and show in ZF(C) that a holds in VRO.
The latter is accordingly a model of IZF + a, from which it follows that he
consistency of ZF(C) implies that of IZF + a. Since the consistency of IZF
implies that of ZF(C), we conclude that a is relatively consistent with IZF.
22 See Appendix 1 for definition of regular element of a Heyting algebra and of the
Boolean algebra of regular elements. Regular elements of Q are precisely those elements co
satisfying — «oo = true <-» co = true.
70
FRAME-VALUED MODELS OF IZF DEVELOPED IN ZFC
We shall strengthen the concept of a core for an H-valued set in the following
way. A strong core for an H-valued set u that is a set v c= VW such that
(i) \[x g ujj = t for all x g v; (ii) for any y e V(H) such that | y g u \ =t , there is a
unique x e v such th a t|x = yl=^ T . It is easy to show, assuming AC, that any H-
valued set u has a strong core. Starting with a core v for u , define the
equivalence relation —on v by x - y <-> \ x = y \\ =t , and let w be a set obtained
by selecting one member from each —-equivalence class. Then w is a strong core
for u. Clearly a strong core for an H-valued set is unique up to bijection in the
sense that there is a bijection between any pair of such strong cores.
23 For proofs (in the Boolean-valued case) we refer the reader to Chapters 1 and 8 of Bell
[2011].
24 A formula is restricted if each of its quantifiers occurs in the form V a gi/ or Eagi / (i.e.,
Vx( a g if -» . . . . ) or 3a( a g y a . . . . ) or can be proved equivalent in IZF to such a formula.
71
A refinement of a subset A of H is a subset B of H such that, for any b e B, there
is a e A such that b < a. If B is an antichain, it is called a disjoint refinement ofA.
H s called refinable if every subset of H has a disjoint refinement with the same
join.
We say that satisfies the existence principle if for any formula cp(x) there is an
H-set u for which px<p(x)l = f<p(u)|
Proof. Suppose that H is refinable. Then for any formula cp(x), by a Collection
argum ent there is a subset A of Vdh for which
Since clearly
|cp(u)S<I3x(p(x)l
we are done. ■
Remark. The converse to the Refinable Existence Lemma also holds. For
suppose $$(**) satisfies the existence principle. Given A c H define the H-set v by
72
dom(^) = {a :a g A} and v(a) = a for CL G A . Then there is an H-set u for
which ||3x(x g v)l = j[u g v $. In that case
Proof, (i) -> (ii). Assuming (i) let f e V(W be such that
Define uab - 1/(&A) = bA[] • It is then easily verified that the unb satisfy (ii).
(ii) - * (i). Suppose the uab satisfy the conditions of (ii). Define / g V(H) by dom ( / )
= {< a,b > A : a g A ,b g B} and /(< a ,b > A) = uah. Then W H) f dom(/) c AA, the
first condition of (ii) gives Fun (f) and the second condition
1= ran(/) - BA. ■
73
An element a of a frame H is connected if for any disjoint b, c e H , a = b v c
implies b = a or c = a. H is said to be connected if its top element T is connected.
Equivalently, H is connected if whenever T is the join of an antichain A, then
T e A . H is said to be locally connected if each of its elements is the join of
connected elements; equivalently, if, for any elements a, b oi H, a <b iff, for all
connected elements c, c <a implies c < b.
(ii) -> (ii). Assume (ii) and let a, b be disjoint elements of H such that a v b - T.
Given sets /, /, let us call the frame H ]) distributive if, for any subset
jl-(J,
74
(ü) H is J) distributive.
Proof, (i) -> (ii). Assume (i), and let \ay :i e I , j e J } cz H be such that
{ay : j e J} is an antichain for each i e 1. Define h e V(H) by dom(/i) =
{< i.j > A : i g I , j e J} and h(< i.j > A) - a {J:. It is then easily verified that
Fun(h) a d o m ( h ) c / A a r a n ( h ) c J A.
A V S- - ffdom(h) - / AI - I h g ( J Af J - Jh e ( J ') A1 - V I h = f AI
iel
= V A IN iA) = / ( i ) Al
/eJ; n=I
- v Aa if(iy
feJ* M
This is (ii).
(ii) —> (i). Assume (ii). To obtain (i) it suffices to show that
(* ) »M l= (JAf c : ( J I)A.
For h g let ^$ (iA, j A) e h l and atj = hi; a /\[ h y a hy =>±]. Note that
a;
75
Let us say that H
Then we have
Proof. The equivalence of (i) and (ii) follows immediately from Prop. 3. So it
suffices to prove the equivalence of (ii), (iii) and (iv).
(iii) -» (iv). To begin with, for elements b < a of H, call b complemented in a if there
is c < a such that b a c ~±, bw c ~ a. It is easy to check that, if such c exists, it is
unique; denote it by a - b. Now assume (iii), and let a e H. Write {a, :i e 1} for
the set of elements of H which are complemented in a. For each i e I let am= at,
Ui\ - a - ai. Then we have, for each g e 2l,
n a = A (a io v aii 1= V mo : iA«,
el
ig(i) V A a,mo
76
Accordingly b = c and it follows that b is connected. From (*) we conclude that
each ß g H is the join of connected elements, so that H is locally connected.
(iv) —> (ii). Suppose that H is locally connected. To show that H is completely jl-
distributive, it suffices to show that, for any subset {a(j :i e I, j e J} of H such
that, for each i e /, the set jay : j g J} is an antichain, and any connected
element c g H, we have
So assume the antecedent of this implication. Then for each i e /, c < V ai]f and,
jeJ
because c is connected, it follows that c < aXJ for some unique j e /. Define
g: 1 —>/ to be the function which assigns this j to each /. Then
c < A a w(i)< V and (*) follows. ■
is I f t J 1t el
We now set about constructing (in ZF) a frame-valued model of IZF in which
is subcountable.
Given two sets A and ß, let P = P(A, B) be the set of finite partial functions from
A to ß, partially ordered by 3 . We shall write p, q for elements of P. Let C be the
coverage25 on P defined by
The sieve S on the l.h.s of this equivalence is called the cover of p determined by b.
Note that every S g C(/?) is nonempty.
q3 = p v { ( a 1,b)}> q 4 = p v { ( a 2,b)}.
Proof. Let S e C(p) and suppose that S c \ J U j . We claim that S c U, for some
jej
Remark. It follows from Lemma 2 that H e is connected. For if the top element P
of Hcs is the join of an antichain {lh: i e /} of elements of H e , then by Lemma 2 ,
[J t/( = P. Hence there is i e l for which 0 e.U t ; but then U- = P.
re/
o n u ^ u r w , -
i el j e J f e d 1i e /
beB.
Define Uab = { p : (a,b> e p}. Then Unb is C-closed. For suppose the cover S of p
associated with V e B is included in Unb . We need to show that p e LU . If b' e
ran(p), then p e S c UW If fr' £ ran(p), choose a ' £ dom (p) u {a} and define
p ' = p u K a ' j b } ) . Then p' e S, so that p' e Uab, from which it easily follows that
p e Urf,.
To show finally that V Uah = r , for b e B, it suffices to show that P s the onlv
aeA
C-closed sieve containing ( J Uab, and for this it suffices to show that for each
aeA
79
determined by b. If q e S. then q(a) = b for some ag A, whence q e U ^ i t
a eA
W a mnp A =J- f 0 r P ^ Q
(2) AAWmnp = T
m n p
To see this, let n++ n * be any bijection of N with itself lacking fixed points
and define b mn e H by b mn = a tnmn*. Then w e have
and
V b™ = V a-" = V % =T/
n n p
so that
A V b™. = T -
80
On the other hand
V Ab'mf(m)
To prove this, notice that it follows from (3) that \fa mnf(m) = T , whence
m
m m m m m
and
cp{ < m , n , P > A) = a
p : ^ AxNA~>W A.
and t 9(mA,
But the usual diagonal argument, carried out in H*, shows that
81
» (H)N ( N H) A * ( N a )n \
Proof. We have
I V i g J a 3 x g Ap(;,x)f = A V A(x)Alcp(iA, x ) l
i&I x e d o n (A )
Since H is refinable, we may use the Axiom of Choice (in ZFC) to select, for each
i g I, a disjoint refinement ^ { b if : j g J t} of {A(x) a [f(p(iA,x;)l: x g dom(A)}
with the same join as the latter. Again using the Axiom of Choice, select for each
i e I , j e J i an element cdom (A ) for which b- < A(x {j) a [[cp(za , x tj)j). If we
now define f g V [H) by dom (J) - {< iA, x xj >(H): i g IJ g J i} and
82
®(Hc) 1= AC( / ) follows immediately. ■
Now since H is totally disconnected, each fl/ is the join of a set of complemented
elements, and by the argum ent above, this set may be taken to be countable.
For each i e l let {bul : n e FI} be a (countable) set of complemented elements
such that ax - V bm• Then
V a , - V a t “ V V b M.
i«- l id L it/ ., nd Ki
Let (cn : n e 1J] be an enumeration of the countable set {bm :i e I0,n e N}. Then
(cn : n g IT} is a refinement of {a- :i g 1} with the same join as the latter.
Now define dn e H , for each recursively by do = Co, d„+i = c«+i a (do v ... y d H)*
83
T hen {dn : n e N} is a n a n tic h a in , dn < cn for each n, a n d V drn - V
neh' neN
In , the set of rational num bers may be identified with the H-set Q A,
where Q is the usual set of rational numbers. Since « (H) is a model of IZ, we
can carry out within it the construction from the rationals of the set !Rd of
(Dedekind) real numbers as in Ch. 3. Let 1 H be a strong core for the resulting
H-valued set. The members of are naturally thought of as H-valued real numbers.
More generally, given h e H, an H-valued real number of degree h. is an element
r e 1/(H) for which h<\[r e E d |
Since each Dedekind real is a cut in the rationals, each H-valued real num ber r is
a pair < l , r > e V(H) for which
viz.,
0. Jl u r c Q a |= t
1. V l p AGLSAl qAG R l - T
PW-,
2. |pA e l] a [ pA e r ] = i
3. I p Ae L l = V l q A ei - i
4. I p A e R l = VllgA eR]l
C{<p
Similar conditions may be written dow n for H-valued real numbers of degree h:
these are left to the reader.
84
It is easy to check that, writing R for the genuine" set of real numbers in V, we
"
With these definitions l . H is called the ordered ring ofH-valued real numbers.
Now let X be a topological space. For brevity we shall write $$(X), VW , R x for
V(W), [R</( respectively. Members of i x will be called simply real
numbers over X. Thus a real num ber over X may be identified as an element r =
< L , r > of VW satisfying the following conditions:
Ox. Jl u r c Q1 =X
lx. U [ p Ae lJn [|qA e r 1= X
4x. ! p Ae R l = ( J l g Ae Rj
q<p
5x . | p A e l [| u I qA e R jj = X for p < q.
We now prove
85
Proof. To obtain this isomorphism, start with a real num ber r = < l , r> over
X. For each t e X define
i-t - {p € Q : t g |[ p A G L j} R f = { p e Q : f e | p A e r ]|}
p g l £ <~>t g ! p A e Lj
To show that r* is continuous, it suffices to show that the inverse image under
r* of each open interval (p, ->) in R is open in X. This follows from the
observation that, for t e X, we have
Remark. For each " genuine'' real num ber r e V, (rA)* is the constant function
86
each point of X gas a neighbourhood on which r* is constant. Note that, if X is
connected, then each locally constant function on X is constant.
We claim that
N < L /- R f > e !R d .
. We verify conditions 3x and 5x, leaving the rest to the reader. First note that
ilp ^ G L ^ /^ p ,-* ) ) .
Ad 3x-
IPA e L, | = / *((p, ->)) =
q>p
(q, ->)) = U/ ->» = U
q>p
e L/ 1
We next show that the maps r >->r* and are mutually inverse, i.e. (r*)* = r
and (/)*=/.
For the second assertion, note that ( / #)* (t) =< (L.f )t,{f*f )t >. So if fit) =
<L, R>, then
We claim finally that the map r »-►r* is an isomorphism of Mx with C(X, E). To
The proofs of the remaining assertions in (*) are left to the reader. ■
The upshot of Proposition 6 is that real numbers over X can be regarded as real
num bers varying continuously over X.
f * =g*l = li-f = L 9 J n | R / = R 0 1
= Inp)lpAe l 7 So IpA e Lg InInP)|pA € R/i » l p AeR3i
P P
= Inf|(/ '(p, - » J o g *(p, ->)) n Inf|(/ p) o g’' h p))
P P
= Inf) [(/ 1(p, ->) o g 1(P, ->)) n I/“1(4-, p) <=>g 1(<-, P))l
p
= I n ( f : Vp[(/(t) > p g(t) > p) A (f(t) <p<r> g(t) < p)\
= ln{t:f{t) = g{t)}. U
In a similar w a y ,, one shows that, for real num bers r, s over X, we have
Hr = si = In{t:r*(t) = s *(*)}.
and
jfr < s[] = In{t :r* (t)< s* (£)}.
These argum ents can easily be extended to establish, for any open set U in X, a
natural correspondence, with analogous properties, between real numbers over
U and continuous real-valued functions on U. Thus , writing C(U, R) for the set
of real-valued continuous functions on If, real numbers over LI correspond to
elements of C(U, R). Under this correspondence locally constant functions on LI
are associated with real numbers r over U for which l i e [[re AAf .
X should be correlated with certain operators on C(X, E), that is, maps
<t>: C(X, E) —» C(X, E). We now set about identifying these operators.
or equivalently, if
Clearly any local operator is near-local. In general, the converse is false, but we
shall later show that, for metric spaces, every near-local operator is local.
We next show that real functions over X are correlated with near-local operators
on C(X, E).
90
Now suppose given a local operator A on C(X, R). We define the function
D: X x R -» R by the stipulation:
Let us call a function F: X x R -> 1R localizable if, for some local operator A on
C(X, R), we have F (t, f(t)) = A(f)(t) for arbitrary f g X, / g C(X, R). Then local
We now focus attention on the case in which the space X is the space R of real
numbers.
To begin with, let i, a, v be the R-valued real num bers correlated with the
affirmed in ^ (%).
91
| i < 0 v i = 0 v t > 0 1 = (<-, 0) ln{0} u (0, ->) = (<—,0) u (0, ->) = IS - {0} * E.
and
Hence Qi is invertiblel * X.
for all s e Q*’, jfs Ais a Cauchy sequence of rational numbers converging to x 1 = 0
92
(**) U c= \[sAis a Cauchy sequence of rational numbers converging to il U= 0 .
So assume U connected and that the amtecedent of (**) holds. Then for each n
U cz l s nA e C|Aj], and so s„A corresponds to a locally constant rational-valued
function in C(U, M ). Since U is connected, this latter function is constant on U;
let pn be that constant (rational) value. The sequence {pn : n e N) is then Cauchy;
so it converges to some unique t0 e R . We then have
q n tm n
If we endow N with the discrete topology, the set Nr"endowed with the
product topology will be written and called Baire space. * I is totally
disconnected28 and has a countable base consisting of clopen sets. We proceed to
establish various properties of the spatial extension ®('4). Our principal task will
be to show that, in Brouwer's Principle holds, that is, in $$('0, every function
from reals to reals is continuous.
We now proceed to outline the strategy (due to Scott [1968]) for showing that, in
V(N) , every function from reals to reals is continuous.
Step 3. Infer that real functions over , 4 are correlated with continuous localizable
functions on 4 x 1 .
Step 4. Show that each real function over .1 correlated with a continuous
localizable function on l x M is continuous in \
The topological details for carrying out Steps 1 and 2 - which are somewhat
intricate and are omitted here - may be found in Scott [1968]. Step 3 then
follows accordingly.
Now for S tep 4. First we note that since CF is totally disconnected, the Definite
Elements Lemma applies to \ This should be borne in mind in the course of
that argum ent that follows.
Let (p be a real function over 4 ', correlated with the continuous* localizable
function F: • l x M R . To show that cp is continuous in ®(' % we need to show
that the sentence
( f ) VpVq > pV e > 03 d > OVx e [p, q]\/y e [ p ,q ] [ \x - y \< d - ) \ (p(x) - cp(p) |< e]
94
holds29 in SW \ (Here we have used p, q, e, d as rational num ber variables, and
[p, q] denotes the closed interval in 1 .) For this it suffices to show that, for any
rationals p, q with p < q and any positive rational e,
(*) ** = \\3d>0Vxe[p,q]Vye{p,q][\x-y\<d->\(p(x)-q>(y)\<e}l
Using the Definite Elements Lemma and the correlation between real numbers
over, Tand elements of C(e l; R), proving (*) amounts to showing that the set 5 =
LJln fl Inf* : [/(*)’ 9(0\ e [P> <?]A I f ( 0 - 9(0 I < d -H F{t, f(t)) - F(t, g(t) |< ef
d> 0 f,ge.C(X, j
coincides with,d.
This completes Step 5 and we conclude that, in S$(' 0, every function from reals to
reals is continuous.
Remark. Recall from the Introduction that a set S is cohesive if whenever subsets
whenever U u V = S and U n V = 0 , then If = S or V = S. S is cohesive iff every
Now it follows from the truth of Brouwer's Principle in S$(' \ that, in $$(' \ R is
cohesive. To prove this we show that from ( f ) above it follows (in IZ) that any
function q>: R -» 2 is constant. Thus let p, q be rational numbers with p < q, and
take c = 1 in ( f ). We get a rational d> 0 satisfying
Let d' be a rational such that 0 < d' < d. Let n be the least integer such that q - p
<nd', let p 0,...,p n be defined by p 0 - p , = p 7 + d ' and define
Ki = [Pi,Pi Jr^[p,q].Then
[p,q] =i=0 •
A straightforward inductive argument, using (f), now shows that cp has constant
value cp(p) on each K i and so also on [p , q ]. Since p and q were arbitrary with
p < q, (p is constant on the whole of M .
The Fundamental Theorem of Algebra (FTA) asserts that the field C of complex
num bers is algebraically closed, i.e. that every polynomial over C has a zero in
showing that it is false in the spatial extension 3S(C), where now C is the space
In the same way as for 1 , one shows the complex num bers over any open
subset LI of a topological space X are correlated with continuous functions in
96
C(U, C). This holds in particular when X is C itself. Write i for the complex
Thus 0 £ I p(u) = 0 j|. Since this holds for arbitrary u, it follows that
0 £ |3 x(p(x) = 0 )||. Similarly, for any a e C , a <£[3 x(p(x) + a = 0 ) |. From this
we deduce that |C is algebraically closed 1 = 0 , i.e.
97
Appendix
Heyting Algebras and Frames
LATTICES
A lattice is a (nonempty) partially ordered set L with partial ordering < in which
each two-element subset {x, y } has a suprem um or join - denoted by x y - and
v
X V JL = X, X A T — X,
X VX= X , X A X = X,
x v y = y v x, x a 1/ = y a x ,
x v (y v 2) = (x v y) v 2, x a (y a 2) = (x a y) a 2,
(x V y) A y = y, (x a y) v y = y
x < y iff x v y = y. It is then easily shown that (L, <) is a bounded lattice in
which v and a are, respectively, the join and meet operations, and 1 and 0 the
top and bottom elements. This is the equational characterization of lattices.
Examples, (i) Any linearly ordered set is a lattice; clearly in this case we have x a
y = min(x, y) and x v y = max(x, y).
(ii) For any set A, the power set PA is a lattice under the partial ordering ofset
inclusion. In this lattice X v Y = X U Y and X a Y= X f| V. A sublattice of a
power set lattice is called a lattice of sets.
98
(iii) If X is a topological space, the families t{X) and rC\X) of open sets and
closed sets, respectively, in X each form a lattice under the partial ordering of set
inclusion. In these lattices v and a are the same as in example (ii).
x A (y v z ) = (x A y) v (x a z), x v (y a z) = (x v y) a (x v z ).
A lattice is complete if every subset has an infimum and a suprem um. The meet
and join of the empty subset of a complete lattice are, respectively, its top and
bottom elements. It is a curious fact that, for a lattice to be complete, it suffices
that every subset have a suprem um, or every subset an infimum. For the
suprem um (infimum), if it exists, of the set of lower (upper)30 bounds of a given
subset X is easily seen to be the infimum (supremum) of X.
Examples, (i) The power set lattice PA of a set A is a complete lattice in which
joins and meets coincide with set-theoretic unions arid intersections respectively.
(ii) The lattices ^X ) and (3{X) of open sets and closed sets of a topological space
are both complete. In the join and meet of a subfamily {IT: i e 1} are given
by
99
V ^= U ^ A ^,= Inf|^-
f“ / 1&J it . }
VA = UA AA = DA-
Here InA and A denote the interior and closure, respectively, of a subset A of a
topological space.
A Heyting algebra is a bounded lattice (H, <) such that, for any pair of elements
x, y e H, the set of z e H satisfying z a x < y has a largest element. This element,
which is uniquely determined by x and y, is denoted by x => y: thus x => y is
characterized by the following condition: for all z e H,
The binary operation on a Heyting algebra which sends each pair of elements x,
y to the element x => y is called implication; the operation which sends each
element x to the element x* = x => ± is called pseudocomplementation. We also
define the operation o of equivalence by x <=> y = (x => y) a (y => x). These
operations are easily shown to satisfy:
<-> (z a x ) v (z a y ) = _L
100
<”> Z A I = 1 & 2 A 1 / - 1
Any Heyting algebra is a distributive lattice. To see this, calculate as follows for
arbitrary elements x, y, z, u:
Any linearly ordered set with top and bottom elements is a Heyting
algebra in which
A basic fact about complete Heyting algebras is that the following identity holds
in them:
X a V j /, = V x a yt
iel iel
And conversely, in any complete lattice satisfying (*), defining the operation =>
by x => y = V{2: 2 a x < y) turns it into a Heyting algebra.
101
x a \ / y j < z <-> V y i < x =>z
v=I iel
102
In a Heyting algebra H the pseudocom plem ent a* of an element a is not, in
general, a complement for a. (Consider the Heyting algebra of open sets of a
topological space.) But there is a simple necessary and sufficient condition on a
Heyting algebra for all pseudocomplements to be complements: this is stated in
the following
x v y ~ y v x, x Ay - y ax
X V (ifV z) = (x V y) V 2, X A (1/ A 2) = (x A if) A 2
(x V y) A y = y, (x A y) v y = y
x a (y v z) - (x a y) v (x a z ), x v (y a 2) = (x v y) a (x v 2).
X V X* = T, X A X* = JL.
(x v y)* - x* a y*, (x a y)* = x* v y*
x** = x
103
(V*,)*
it-1
=itA*.*
I
xAVjy-VfxAyj
t, j !(?/
(A*,)* = V*. *
ic l IC-I
x v A m = A(xvj/J
le i i± I
The meet, join, and complementation operations in a Boolean algebra are called
its Boolean operations. A subalgebra of a Boolean algebra B is a nonem pty subset
closed under B's Boolean operations. Clearly a subalgebra of a Boolean algebra B
is itself a Boolean algebra with the same top and bottom elements as those of B.
Examples of Boolean algebras, (i) The linearly ordered set 2 = {0,1} with 0 < 1 is
a complete Boolean algebra, the 2-element algebra.
(ii) The power set lattice PA of any set A is a complete Boolean algebra. A
subalgebra of a power set algebra is called a field of sets.
(iii) Let F(A) consist of all finite subsets and all complements of finite subsets of
a set A. With the partial ordering of inclusion, F(A) is a field of sets called the
finite-cofinite algebra of A.
(iv) Let X be a topological space, and let C(X) be the family of all
simultaneously closed and open ("dopen") subsets of X. With the partial
ordering of inclusion, C(X) is a Boolean algebra called the dopen algebra of X.
Remark: The lattices (i), (ii), (iii) are only Boolean algrbras if the Law of
Excluded Middle is assumed.
104
ordering inherited from H, is a Boolean algebra —the regularization of H - i n
which the operations a and * coincide with those of H, but31 vb = ( v h )**. If H is
complete, so is B; the operation /\ in B coincides with that in H while Vs =
(V h)**.
Let (P, <) be a fixed but arbitrary partially ordered set: we shall use letters p, q, r,
s, t to denote elements of p. A subset S of P is said to be a sharpening of or to
sharpen, a subset T of P, w rittenS < T , if V s g S3t g T(s <t). A sieve in P is a
subset S such that p e S and q <p implies q e S.
/=>J = { p : / n p k J ) ~~J> i p : / n p i = 0 } .
3SeC(p)(S^I)~^peL
We write He for the set of all C-closed sieves in P, partially ordered by inclusion.
Heyting and Boolean algebras have close connections with intuitionistic and
classical logic32, respectively.
Intuitionistic first-order logic has the following axioms and rules of inference.
Axioms
a - > (ß - > a )
[a -> (ß h>y) -» [(a —» ß) —> (a -> y)]
a (ß - » a a ß)
aAß-»a aAß->ß
a - » a v ß ß - > a v ß
[a (ß -> y) -> [(a —> ß) —> (ot —» y)]
(a -» y) - ^ [ ( ß ^ y ) ^ ( a v ß ^ y)]
(a - > ß ) - > [(a - > -iß ) - > - i a ]
—iCX—v (ot —> ß)
a (t) -> 3xa(x) Vxa(x) -> a (y) (x free in a and t free for x in a)
x =x a(x) a x = y -» a(y
32 For accounts of both systems of logic, see, e.g. Bell and Machover [1977].
106
Rules of Inference
a, a -» ß
_ P
ß -» a(x) a(x) -» ß
(x not free in ß)
Classical first-order logic is obtained by adding to the intuitionistic system the rule
of inference
- 1-1a
a
are derivable. However LEM and LDN are intuitionisticallv equivalent and
DML is intuitionisticallv equivalent to the weakened law of excluded middle:
107
class. Now define the relation < on the set H(T) of ^-equivalence classes by
M < [y] if and only if T hep—>y. Then < is a partial ordering of H(T) and the
partially ordered set (H(T),<) is a Heyting algebra in which [cp] => [y] = [cp -> y],
with analogous equalities defining the meet and join operations, 0, and 1. H(T) is
called the the Heyting algebra determined by T. It can be shown that Heyting
algebras of the form H(T) are typical in the sense that, for any Heyting algebra L,
there is a propositional intuitionistic theory T such that L is isomorphic to H(T).
Accordingly Heyting algebras may be identified as the algebras of intuitionistic
logic.
baj = I«]*
A formula a is said to be Heyting valid - written h a —if [a] = t for any such map
/ It can then be shown that a is Heyting valid iff h a in the intuitionistic
propositional calculus, i.e., iff a is provable from the propositional axioms listed
above.
108
Finally, again as regards semantics, complete Heyting and Boolean algebras are
related to intuitionistic, and classical first-order logic, respectively. To be precise,
let W be a first-order language whose sole extralogical symbol is a binary
predicate symbol P. A Heyting algebra-valued yt-structure is a quadruple
M = (M, eq, Q, H), where M is a nonem pty set, H is a complete Heyting algebra
and eq and Q are maps M2 —> M satisfying, for all m, n, m', n' e M,
eq(mr m) = t , eq(m, n) = eq(n, m), eq(m, n) A eq(n, n ') < eq(m, n 7),
Q(m, n) A cr/(m, m ') < Q(tri, ri), Q(m, n) A ) < Q(m, n f ).
For any formula a of .^and any finite sequence x = <x\, ..., x„> of variables of y*
containing all the free variables of a, we define for any Heyting-valued c/~
structure M a map
[a]M*: Mn -> H
recursively as follows:
109
order formula is valid in every Boolean-valued structure iff it is provable in
classical first-order logic.
110
Concluding Observations
enumerable subsets, arid the detachable subsets of N with the recursive subsets.
The object N':i may be considered as the set Rec of (total) recursive functions on
M34. That being the case, if we write cpAfor the partial recursive function on N
with index x, arid U for the set of indices of total recursive functions35, the map
Eff. On the other hand is not numerable in Eff because it can be shown that,
33 Accounts of topos theory may be found in Bell [1988], Goldblatt [1979], Johnstone
[1977] and [2002], Lambek and Scott [1986], Mac Lane arid Moerdijk [1992] and McLarty
[1992].
34 Hence, in Eff, Church's thesis holds in the strong sense that every function A -> M is
recursive.
35 Since N is uncountable, so m ust U be; this corresponds to the result from recursion
theory that the set of indices of total recursive functions is not recursively enumerable.
Ill
VF e N( ' ’V/ e W*3n e NVg e N*[Vm < n(f(m) = g(m)) - » F (/) = F(g)] 36
From this it would follow that, if F were an injection from to N, then, for each
In Eff, both Par*(N,N) and P*N are countable. That Par*(N,N) in is countable in
Eff follows from the fact that it corresponds to the set of partial recursive functions
index x is, in Eff, a surjection from N to Par*(N,N). (That being the case, as we
noted above, the set {x eN: x £ dom((px)} m ust be (in Eff) uncountable. This
corresponds to the fact that this set is not recursively enumerable.) A similar
argum ent - using the fact that P*N in Eff corresponds to the set of recursively
36 In Eff the set N( ] of all maps N -» M:i may be identified with the set of effective
operations, that is, the set of functions F: Rec -» N such that there is a partial recursive
function cpwith the property that for every / e Rec and every index n for f we have
(pfw) = F(J). Brouwer's continuity principle asserts the continuity of every map to N from
Baire space, V.
112
While Nn fails to be num erable in Eff, Andrej Bauer has recently shown it to be
numerable in the related topos Eff! in which maps between objects constructed
from the natural num bers correspond to functions which are infinite time
computable, that is, computable by an infinite time Turing machine. This is a
Turing machine which is allowed to run infinitely long, with the computation
steps counted by ordinals. The power of these machines far exceeds that of
ordinary Turing machines: for example, both the halting problem and the
problem of deciding the equality of two total recursive functions are soluble
using infinite time machines. In Eff!, just as in E ff,, N " is subcountable. But in
Eff! Nn also satisfies the axiom of choice in the form: any total relation defined on
NN contains a function (this cannot be the casein Eff). Putting these two facts
There are a num ber of topos models of Brouwer's Principle (BP) that all real
functions are continuous. As we have essentially shown in Chapter IV, BP
holds in Shv(f I)37. Mac Lame and Moerdijk [1992] present a different kind of
topos model of BP. BP also holds in any of the so-called smooth toposes. Each of
these may be considered to be an enlargement of the category Man of manifolds
(or spaces) and smooth maps to a topos which contains no new maps between
spaces, so that all such maps there - in particular those from E to E are still
smooth, and hence also continuous.
37 In fact BP caii be shown to hold in many other spatial toposes: see Hyland [1979]
113
Historical Notes
Chapter 1. Friedman [1973] seems to have been the first to investigate systems of
intuitionistic set theory. Crayson [1978] undertakes a systematic investigation of
topology and ordinal arithmetic in an intuitionistic setting. That LEM follows
from the axiom of choice was first proved, in a category-theoretic setting, by
Diaconescu [1975]; the logical version was formulated and proved by Goodman
and Myhill [1978]. The investigation of the connection between choice principles
and logical principles is taken from Bell [2006}; see also Bell [2009].
Chapter 3. Much of the discussion of real num bers presented here is based on
Johnstone's [2002] account of real num bers in a topos. Proposition 2 is due to
Johnstone [1979].
114
Appendix 2. The effective topos was first introduced by Hyland [1983]. The
concept of a smooth topos is due to F. W. Lawvere.
115
Bibliography
Bell, J. L. [1988] Toposes and Local Set Theories: An Introduction. Clarendon Press,
Oxford, 1988. Dover reprint 2007.
Bell, J. L. [199ii] Finite sets and Frege structures /. Symbolic Logic, 64, no. 4, 152-
156.
116
Fourman, M. P. and Scptt, D. S. [1979] Sheaves and logic. In Fourman, M. P.,
Mulvey, C. J., and Scott, D. S. (eds.) Applications of Sheaves. Proc. L.M.S. Durham
Symposium 1977. Springer Lecture Notes in Mathematics 753, pp. 302-401.
Grayson, R. J. [1975] A sheaf approach to models of set theory. M.Sc. thesis, Oxford
University.
117
Durham Symposium 1977. Springer Lecture Notes in Mathematics 753, pp. 479-
491.
118
Index
119
ordered pair 13
ordered pairof in 67
Peano structure 28
power set 13
preserves exponentials 77
Principle of Induction on Ordinals
57
pseudocom plem ent (*) 102
sieve 107
Simple Recursion Principle 29
spatial extension 6s
strict Frege structure 40
strictly finite 35
strong core 73
subcountable 1
subquotient 75
subset classifier 15
totally disconnected 66
transitive 58
true 61
truth value 61
Zorn's Lemma 70
120
While intuitionistic Cor constructive] s e t theory IST
has received a certain attention from mathematical
logicians, so fa r as I am aware no book providing
a systematic introduction to the subject has yet
been published. This may be the case in part
because, as a form of higher-order intuitionistic
logic - the internal logic of a topos - 1ST has
been chiefly developed in a tops-theoretic context.
In particular, proofs of relative consistency with
1ST fo r mathematical assertions have been
(implicitly] formulated in topos- o r sheaf-theoretic
term s, ra th e r than in the framework of Heyting-
algebra-valued models, the natural extension to
1ST of the well-known Boolean-valued models fo r
classical s e t theory.