The Axiom of Choice and the Law of Excluded Middle in Weak Set

Theories

John L. Bell

Department of Philosophy, University of Western Ontario

In constructive mathematics the axiom of choice (AC) has a somewhat ambiguous status. On the
one hand, in intuitionistic set theory, or the local set theory associated with a topos ([2]) it can be
shown to entail the law of excluded middle (LEM) ([ 3 ], [ 5 ]). On the other hand, under the
“propositions-as types” interpretation which lies at the heart of constructive predicative type
theories such as that of Martin-Löf [9], the axiom of choice is actually derivable (see, e.g. [11] ),
and so certainly cannot entail the law of excluded middle. This incongruity has been the subject
of a number of recent investigations, for example [6], [7], [9], [12]. What has emerged is that for
the derivation of LEM from AC to go through it is sufficient that sets (in particular power sets),
or functions, have a degree of extensionality which is, so to speak, built into the usual set theories
but is incompatible with constructive type theories Another condition, independent of
extensionality, ensuring that the derivation goes through is that any equivalence relation
determines a quotient set. LEM can also be shown to follow from a suitably extensionalized
version of AC. The arguments establishing these intriguing results have mostly been formulated
within a type-theoretic framework. It is my purpose here to formulate and derive analogous
results within a comparatively straightforward set-theoretic framework. The core principles of this
framework form a theory – weak set theory WST – which lacks the axiom of extensionality1 and
supports only minimal set-theoretic constructions. WST may be considered a fragment both of
(intuitionistic) ∆0-Zermelo set theory and Aczel’s constructive set theory ([1]). In particular WST
is, like constructive type theories, too weak to allow the derivation of LEM from AC. But we shall
see that , as with constructive type theories, beefing up WST with extensionality principles (even
very moderate ones) or quotient sets enables the derivation to go through.

Let L be the first-order language of (intuitionistic) set theory which, in addition to the
usual identity and membership symbols = and ∈ also contains a binary operation symbol

1
Set theories (with classical logic) lacking the axiom of extensionality seem first to have been extensively studied in [4] and [10].
 2

〈 , 〉 permitting the formation of ordered pairs2. At certain points various additional predicates

and operation symbols will be introduced into L. The restricted quantifiers x∈a and x∈a are
defined as usual, that is, as ∃x (x ∈ a ∧ ...) and ∀x (x ∈ a → ...) respectively. A formula is

restricted if it contains only restricted quantifiers.

Weak set theory WST is the theory in L with the following basic axioms (in which the
free variables are understood to be universally quantified, and similarly below):

Unordered Pair ∃u ∀x [x ∈ u ⇔ x = a ∨ x = b ]

Ordered Pair 〈a , b 〉 = 〈c , d 〉 ⇔ a = c ∧ b = d

Binary Union ∃u ∀x [x ∈ u ⇔ x ∈ a ∨ x ∈ b ]

Cartesian Product ∃u ∀x [x ∈ u ⇔ ∃y ∈ a ∃z ∈ b (x = 〈y , z 〉 )]

Restricted Subsets ∃u ∀x [x ∈ u ⇔ x ∈ a ∧ ϕ]

where in this last axiom ϕ is any restricted formula in which the variable u is not free.
We introduce into L new predicates and operation symbols as indicated below and
adjoin to WST by the following “definitional” axioms:

a ⊆ b ⇔ ∀x [x ∈ a ⇒ x ∈ b ] a ≈ b ⇔ ∀x [x ∈ a ⇔ x ∈ b ] Ext (a ) ⇔ ∀x ∈ a ∀y ∈ a[x ≈ y ⇒ x = y ]
x ∈a ∪ b ⇔ x ∈a ∨ x ∈b x ∈ {a , b } ⇔ x = a ∨ x = b {a } = {a , a } x r y ⇔ 〈 x , y 〉 ∈ r
y ∈ { x ∈ a : ϕ(x )} ⇔ y ∈ a ∧ ϕ(y ) (ϕ restricted)
¬x ∈ 0 1 = {0} 2 = {0,1}
x ∈ a × b ⇔ ∃u ∈ a ∃v ∈ b(x = 〈u,v 〉 ) x ∈ a + b ⇔ ∃u ∈ a ∃v ∈ b[x = 〈u,0〉 ∨ x = 〈v ,1〉 ]
f : a → b ⇔ f ⊆ a × b ∧ ∀x ∈ a ∃y ∈ b(x f y ) ∧ ∀x ∀y∀z [(x f y ∧ x f z ) ⇒ y = z ]
Fun ( f ) ⇔ ∃a ∃b( f : a → b )
f : a → b ∧ x ∈ a ⇒ x f f (x )
f : a → b ∧ g : b → c ⇒ g f : a → c ∧ ∀x ∈ a[(g f )(x ) = g ( f (x ))]
f :a b ⇔ f : a → b ∧ ∀y ∈ b ∃x ∈ a[y = f (x )]
π1 : a + b → a ∪ b ∧ ∀x ∈ a[π1 (〈 x ,0〉 ) = x ] ∧ ∀y ∈ b[π1 (〈y,1〉 ) = y ]
π2 : a + b → 2 ∧ ∀x ∈ a[π2 (〈 x ,0〉 ) = 0] ∧ ∀y ∈ b[π2 (〈y,1〉 ) = 1]
Eq (s, a ) ⇔ s ⊆ a × a ∧ ∀x ∈ a (x s x ) ∧ ∀x ∈ a ∀y ∈ a (x s y ⇒ y s x ) ∧
∀x ∈ a ∀y ∈ a ∀z ∈ a[(x s y ∧ y s z ) ⇒ x s z ]
Comp(r , s ) ⇔ ∀x ∀x ′∀y[(x s x ′ ∧ x ′ r y ) ⇒ x r y ]
Comp(r ) ⇔ ∀x ∀x ′∀y[(x ≈ x ′ ∧ x ′ r y ) ⇒ x r y ]
Extn ( f , s ) ⇔ Fun ( f ) ∧ ∀x ∀x ′[x s x ′ ∧ ∃y ∃y ′(x f y ∧ x ′ f y ′) ⇒ f (x ) = f (x ′)]
Ex ( f ) ⇔ Fun ( f ) ∧ ∀x ∀x ′[x ≈ x ′ ∧ ∃y ∃y ′(x f y ∧ x ′ f y ′) ⇒ f (x ) = f (x ′)]

2
While the ordered pair <u,v> could be defined in the customary way as {{u}, {u,v}}, here it is taken as a primitive operation —as it
is in type theory—both for reasons of simplicity and to emphasize the fact that for our purposes it does not matter how (or indeed
whether) it is defined set-theoretically.
 3

Most of these definitions are standard. The functions π and π are projections of ordered
1 2

pairs onto their 1st and 2nd coordinates respectively: clearly, for u, v ∈ a + b we have
(proj) u = v ⇔ [π1(u ) = π1(v ) ∧ π2 (u ) = π2 (v )].

The relation ≈ is that of extensional equality. Ext(a) expresses the extensionality of the members of
the set a. Eq(s,a) asserts that s is an equivalence relation on a. If r is a relation between a and b, and
s an relation on a, Comp(r,s) expresses the compatibility of r with s, and Comp(r) the compatibility of
r with extensional equality. If f: a → b, and s is an equivalence relation on a, Etxn(f,s) expresses the
idea that f treats the relation s as if it were the identity relation: we shall then say that f is s-
extensional. Ex(f) asserts that f is extensional in the sense of treating extensional equality as if it
were identity.

In addition to the axioms of WST, We formulate the following axioms additional to those
of WST (recalling that in stating axioms that all free variables are universally quantified):
:

Extensionality Ext({a, b})

Extsub(1) a ⊆ 1 ∧ b ⊆ 1 ⇒ Ext({a, b})

Extsub(2) a ⊆ 2 ∧ b ⊆ 2 ⇒ Ext({a, b})

While Extensionality asserts that all extensionally equal sets are identical, Extsub(1) and
Extsub(2) are weak versions confining the assertion just to subsets of 1 or 2. Explicitly, Extsub(1)
and Extsub(2) assert that each doubleton composed of subsets of 1, or of 2, is extensional. Notice
that since 1 ⊆ 2, Extsub(1) is a consequence of Extsub(2).
Following [1], we define a set a to be a base if every relation with domain a includes a
function with domain a, i.e.

Base(a) ∀b∀r [r ⊆ a × b ∧ ∀x ∈ a ∃y ∈ b(x r y ) ⇒ ∃f : a → b∀x ∈ a (x r f (x ))] .

We call a an extensional base if every relation with domain a compatible with extensional equality
includes an extensional function with domain a, i.e.

Extbase(a):
∀b∀r [r ⊆ a × b ∧ Comp(r ) ∧ ∀x ∈ a ∃y ∈ b(x r y ) ⇒ ∃f : a → b[Ex ( f ) ∧ ∀x ∈ a (x r f (x ))]]
 4

We use these notions to state a number of versions of the axiom of choice (again recalling that in
stating axioms that all free variables are universally quantified):

(Intensional) Axiom of Choice AC Base(a)

Weak Axiom of Choice 1 WAC(1) a ⊆ 1 ∧ a’ ⊆ 1 ∧ b ⊆ 1 ∧ b’⊆ 1 ⇒ Base( { 〈a , a ′〉, 〈b, b ′〉 } )

Weak Axiom of Choice 2 WAC(2) a ⊆ 2 ∧ a’ ⊆ 2 ⇒ Base( {a, a ′} )

Universal Extensional Axiom of Choice UEAC

Eq (s, a ) ∧ r ⊆ a × b ∧ Comp(r , s ) ∧ ∀x ∈ a ∃y ∈ b(x r y ) ⇒ ∃f : a → b[Extn ( f , s ) ∧ ∀x ∈ a (x r f (x ))]

Extensional Axiom of Choice EAC Extbase(a)

In asserting that every set is a base, AC means, as usual, that a choice function always exists
under the appropriate conditions on an arbitrarily given relation. WAC(1) and WAC(2) restricts
the existence of such choice functions to relations whose domains are doubletons of a certain
form3. UEAC asserts that, in the presence of an equivalence relation s with which a given relation
r is compatible, the choice function can be taken to be s-extensional. EAC is the special case of
UEAC in which the equivalence relation is that of extensional equality. In view of the fact that
AC can be seen to be the special case of UEAC in which the equivalence relation is the identity
relation, AC is sometimes known as the intensional axiom of choice.
Our next axiom is

Quotients Eq (s, a ) → ∃u ∃f [ f : a u ∧ ∀x ∈ a ∀y ∈ a[ f (x ) = f (y ) ↔ x s y ]]

This axiom asserts that each equivalence relation determines a quotient set. In WST + Quotients,

we introduce operation symbols i , [i]• and adjoin the “definitional” axiom

i

Eq (s, a ) ⇒ [∀x ∈ a ([x ]s ∈ a /r ) ∧ ∀u ∈ a ∃x ∈ a (u = [x ]s )

(Q) s .
∧ ∀x ∈ a ∀y ∈ a[[x ]s = [y ]s ⇔ x s y ]]

3
Using the fact that subsets of 2 are in natural bijective correspondence with pairs of subsets of 1, it can be shown that WAC(1) and
WAC(2) are in fact equivalent. Nevertheless for our purposes it will be useful to have both principles.
 5

Here a is the quotient of a by s and, for x ∈ a, [x]s is the image of x in a/s.

s

Reminding the reader that our background logic is intuitionistic, we finally introduce the
following logical scheme:

Restricted Excluded Middle REM

ϕ ∨ ¬ϕ for any restricted formula ϕ

We are going to prove the following results.

Theorem 1. REM is derivable in: (a) WST + Extsub(1) + WAC(1); (b) WST + Extsub(2) +
WAC(2); and (c) WST + EAC.

Theorem 2. REM is derivable in WST + AC + Quotients

Theorem 3. AC ⇔ UEAC is derivable in WST + Quotients.

Thus, while in the absence of extensionality for doubletons of subsets of 1 or 2, the

intensional axiom of choice does not entail the law of excluded middle, with that degree of
extensionality the law of excluded middle becomes a consequence of very weak versions of the
intensional axiom of choice. (A fortiori REM is derivable in WST + Extensionality + AC.)
Moreover, the extensional axiom of choice entails the law of excluded middle without additional
extensionality assumptions. And finally, when quotients are present the intensional axiom of
choice is no weaker than its universal extensional version and entails the law of excluded middle
without additional extensionality assumptions.

Proof of Theorem 1.
(a) We argue in WST + Extsub(1) + WAC(1). Given an arbitrary restricted formula ϕ, we
define s = {x ∈ {0}: ϕ} and
a = { 〈s, {0}〉, 〈 {0}, s 〉 } .

It is then easily shown that

〈u,v 〉 ∈ a ⇒ ∃x ∈ 2[(x = 0 ⇒ 0 ∈ u ) ∧ (x = 1 ⇒ 0 ∈ v )].

So WAC(1) gives f: a →2 such that , for 〈u,v 〉 ∈ a

 6

(1) f (〈 u , v 〉 ) = 0 → 0 ∈ u

(2) f (〈 u , v 〉 ) = 1 ⇒ 0 ∈ v .

Since f maps to 2, we have

[ f (〈s, {0}〉 ) = 0 ∨ f (〈s, {0}〉 ) = 1] ∧ [ f (〈 {0}, s 〉 ) = 0 ∨ f (〈 {0}, s 〉 ) = 1].

From this, together with (1) and (2), it follows that

[0 ∈ s ∨ f (〈s, {0}〉 ) = 1] ∧ [ f (〈 {0}, s 〉 ) = 0 ∨ 0 ∈ s ],

whence, using the distributive law,

0 ∈ s ∨ [ f (〈s, {0}〉 ) = 0 ∧ f (〈{0}, s 〉 ) = 1].

Writing ψ(s) for the second disjunct in this last formula, it becomes

(3) 0 ∈ s ∨ ψ(s).

Now from s ⊆1 we deduce

0 ∈ s ⇒ s ≈ {0},
whence using Extsub(1),

0 ∈ s ⇒ s = {0}.
Hence
[0 ∈ s ∧ ψ(s )] ⇒ [s = {0} ∧ ψ(s )]
⇒ ψ({0})
⇒ 0 = 1.
Since clearly 0 ≠ 1, we conclude that
ψ(s ) ⇒ ¬0 ∈ s

and (3) then yields

(4) 0 ∈ s ∨ ¬0 ∈ s.
But obviously 0 ∈ s ⇔ ϕ, so (4) gives ϕ ∨ ¬ϕ, as required.

(b) 4 We argue in WST + Extsub(2) + WAC(2). Given a formula ϕ, define

a = { x ∈ 2 : x = 0 ∨ ϕ }, b = { x ∈ 2 : x = 1 ∨ ϕ }.

Since 0 ∈ a and 1 ∈ b, we have

4
Using the observation in the previous footnote that WAC(1) and WAC(2) are equivalent, it can be seen that (b) actually follows
from (a). However the direct proof given here, based on that given in [5], is, in the author’s view, considerably more illuminating.
 7

∀x ∈ {a , b }∃y ∈ 2.y ∈ x ,

and so WAC(2) applied to the relation

r = { 〈 x , y 〉 ∈ {a , b } × 2 : y ∈ x } )
yields a function f: {a, b} → 2 for which ∀x ∈ {a , b }. f (x ) ∈ x . It follows
that f (a ) ∈ a ∧ f (b ) ∈ b, whence
[ f (a ) = 0 ∨ ϕ] ∧ [ f (b ) = 1 ∨ ϕ].

Applying the distributive law, we then get

ϕ ∨ [ f (a ) = 0 ∧ f (b ) = 1].

whence

(1) ϕ ∨ f (a ) ≠ f (b ).

Now clearly ϕ ⇒ a ≈ b , and from this and Extsub(2) we deduce ϕ ⇒ a = b , whence

(2) ϕ ⇒ f (a ) = f (b ).

It follows that f (a ) ≠ f (b ) ⇒ ¬ϕ, and we conclude from (1) that

ϕ ∨ ¬ϕ ,

as required.

(c) Here the argument in WST+ EAC is the same as that given in (b) except that in
deriving (2) above we invoke EAC in place of Extsub(2). To justify this step it suffices to show

that Comp(r), where r is the relation defined in the proof of (b). This, however, is clear. ■

Proof of Theorem 2.

For b ⊆ a, we say that b is detachable (in a) if

∀x ∈ a[x ∈ b ∨ x ∉ b ].

An indicator for b (in a) is a function g: a × 2 → 2 satisfying

∀x ∈ a[x ∈ b ⇔ g (〈 x ,0〉 ) = g (〈 x ,1〉 )] .

It is easy to show that a subset is detachable if and only if it has an indicator. For if b ⊆ a is
detachable, then g: a × 2 → 2 defined by
 8

g (〈 x ,0〉 ) = g (〈 x ,1〉 ) = 0 if x ∈ b
g (〈 x ,0〉 ) = 0 ∧ g (〈 x ,1〉 ) = 1 if x ∉ b

is an indicator for b. Conversely, for any function g: a × 2 → 2, we have

g (〈 x ,0〉 ) = g (〈 x ,1〉 ) ∨ g (〈 x ,0〉 ) ≠ g (〈 x ,1〉 ) , so if g is an indicator for b, we

infer ∀x ∈ a[ x ∈ b ∨ x ∉ b ], and u is detachable.

Now we show in WST + AC + Quotients that every subset of a set has an indicator, and
is hence detachable. Given b ⊆ a, let s be the binary relation on a + a given by:

s = { 〈〈 x , 0),( x , 0〉〉 : x ∈ a } ∪ { 〈〈 x ,1), 〈 x ,1〉〉 : x ∈ a } ∪ { 〈〈 x , 0), 〈 x ,1〉〉 : x ∈ b } ∪ { 〈〈 x ,1), 〈 x , 0〉〉 : x ∈ b }.

It is easily checked that Eq(s, a + a). Also, it is clear that, for z, z’∈ a + a,
(1) z s z ′ ⇒ π1(z ) = π1(z ′)

and, for x ∈ a,
(2) x ∈ b ⇔ 〈 x ,0〉 s 〈 x ,1〉.

Invoking axiom (Q) above, we introduce the quotient (a + a ) of a + a by s and the image
s

[u]s of an element u of a + a in (a + a ) for which we then have

s

(3) ∀z ∈ (a + a ) ∃u ∈ a + a (z = [u ]s )
s
and
(4) ∀u ∈ a + a ∀v ∈ a + a[[u ]s = [v ]s ⇔ u s v ] .

Applying AC to (3) yields a function f : (a + a ) → a + a for which

s
(5) ∀z ∈ (a + a ) z = [ f (z )]s ) .
s
Clearly f is one-one, that is, we have

(6) f(z) = f(z’) ⇔ z = z’.

Next, observe that, for i = 0, 1, and x ∈ a,

(7) π1([ f (〈 x , i 〉 )]s ) = x .

For from (5) we have [〈 x , i 〉 ]s = [ f ([〈 x , i 〉 ]s )]s , whence by (4) 〈 x , i 〉 s f ([〈 x , i 〉 ]s ) . Hence by (1)

π1(〈 x , i 〉 ) = π1( f ([〈 x , i 〉 ]s )) . (7) now follows from this and the fact that π1(〈 x , i 〉 ) = x .

We have also
(8) x ∈ b ⇔ f ([〈 x ,0〉 ]s ) = f ([〈 x ,1〉 ]s ) .

For we have
 9

x ∈ b ⇔ 〈 x ,0〉 s 〈 x ,1〉 using (2)

⇔ [〈 x ,0〉 ]s = [〈 x ,1〉 ]s using (4)

⇔ f ([〈 x ,0〉 ]s ) = f ([〈 x ,1〉 ]s ) using (6).

Now define g: a × 2 → 2 by
g (〈 x , i 〉 ) = π2 ( f ([〈 x , i 〉 ]s )) .
We claim that g is an indicator for b. This can be seen from the following equivalences:
x ∈ b ⇔ f ([〈 x ,0〉 ]s ) = f ([〈 x ,1〉 ]s ) (by (8))

⇔ π1( f ([〈 x ,0〉 ]s )) = π1( f ([〈 x ,1〉 ]s ))

(by (proj))
∧ π2 ( f ([〈 x ,0〉 ]s )) = π2 ( f ([〈 x ,1〉 ]s ))

⇔ π2 ( f ([〈 x ,0〉 ]s )) = π2 ( f ([〈 x ,1〉 ]s )) (using (7)

⇔ g (〈 x ,0〉 ) = g (〈 x ,1〉 ) .

So we have shown that WST + AC + Quotients every subset of a set has an indicator,
and is accordingly detachable. This latter fact easily yields REM. For, given restricted ϕ, for any
a, the set b = {x ∈ a: ϕ} is then a detachable subset of a, from which ∀x ∈ a (ϕ ∨ ¬ϕ) immediately

follows. By taking a = {x} we get ϕ ∨ ¬ϕ . ■

Proof of Theorem 3. It suffices to derive UEAC from AC in WST + Quotients. Assuming

Eq(s, a), we use AC as in the proof of Theorem 2 to obtain a function p : a → a such that
s

u = [ p(u )]s for all u ∈ a . From this we deduce [x ]s = [ p([x ]s )]s , whence
s
(1) x s p([x ]s )

for all x ∈ a.
Assuming the antecedent of UEAC , viz.,
Eq (s, a ) ∧ r ⊆ a × b ∧ Comp(r , s ) ∧ ∀x ∈ a ∃y ∈ b(x r y ) ,

define the relation r ′ ⊆ a × b by

s
u r’ y ⇔ p(u) r y .

Now use AC to obtain a function g : a → b for which ∀u ∈ a (u r ′ g (u )) , i.e.

s s

(2) ∀u ∈ a ( p(u ) r g (u )) .
s
Define f: a → b by

f(x) = g([x]s).
 10

Then by (2)
∀x ∈ a ( p([x ]s ) r g ([x ]s )) .

From this, (1) and Comp(r, s) it follows that ∀x ∈ a (x r g ([x ]s )) , i.e.

(3) ∀x ∈ a (x r f (x )) .

Moreover, for all x, x’ ∈ a, we have

x s x’ ⇒ [x]s = [x’]s ⇒ f(x) = g([x]s) = g([x’]s) = f(x’),

whence Extn(f, s). This , together with (3), establishes the consequent of UEAC. ■

Concluding remarks.
1. The proof of Theorem 1 (a) is an adaptation of the proof of the analogous result
given in [2], pp. 144-146, and that of (b) is based on that given in [5].

2. The proof of Theorem 2 is an adaptation to a set-theoretical context of the

argument in [3] that, in a topos satisfying the axiom of choice, all subobjects are
complemented. By weakening Quotients to the assertion Quotients(1 + 1) that quotient
sets are determined just by equivalence relations on the set 1 + 1, the proof of
Theorem 2 shows that REM is derivable in the theory WST + AC + Quotients(1 + 1).

3. Quotients can be derived within WST augmented by the full extensional power set
axiom
Extpow ∃u[Ext (u ) ∧ ∀x [x ∈ u ⇔ x ⊆ a ]]

So (cf. [7 ]) adding extensional power sets to WST + AC yields REM.

4. Another version of the axiom of choice, easily derivable in WST from AC is:
AC* ∀x ∈ a[x ⊆ b ∧ ∃y(y ∈ x )] ⇒ ∃f : a → b ∀x ∈ a ( f (x ) ∈ x )].

If one adds to WST the nonextensional power set axiom, viz.

Pow ∃u ∀x [x ∈ u ⇔ x ⊆ a ] ,

then AC* becomes derivable from AC. Note that while Extpow entails REM, Pow is
logically “harmless”, that is, it has no nonconstructive logical consequences such as LEM.
The extensional version of AC*, viz.
EAC* ∀x ∈ a[x ⊆ b ∧ ∃y(y ∈ x )] ⇒ ∃f : a → b[Ex ( f ) ∧ ∀x ∈ a ( f (x ) ∈ x )].

is derivable in WST from EAC. In WST + Pow, EAC and EAC* are equivalent.
 11

5. A version of the axiom of choice considered in [9] is what we shall call

Rep(resentatives)

Eq (s, a ) ⇒ ∃f [ f : a → a ∧ ∀x ∈ a (xsf (x )) ∧ ∀x ∈ a ∀y ∈ a[xsy ⇔ f (x ) = f (y )]].

Rep asserts that unique representatives can be chosen from the equivalence classes of any
equivalence relation. Obviously, in WST, Rep implies Quotients. Moreover, the proof of
Theorem 2 is easily adapted to show that, in WST, Rep yields REM. In WST, AC +
Quotients entails Rep, and, in WST + Pow, conversely.

6. Finally, consider the following versions of AC which are closely related to that
introduced by Zermelo [13] (see also [9]), namely

ACZ [∀x ∈ a∃y( y ∈ x ) ∧ ∀x ∈ a∀y ∈ a[∃z( z ∈ x ∧ z ∈ y ) ⇒ x = y ]] ⇒ ∃u∀x ∈ a∃! y( y ∈ x ∧ y ∈ u)

EACZ [∀x ∈ a∃y( y ∈ x ) ∧ ∀x ∈ a∀y ∈ a[∃z( z ∈ x ∧ z ∈ y ) ⇒ x ≈ y ]] ⇒ ∃u∀x ∈ a∃ ! y( y ∈ x ∧ y ∈ u)

These are construals of the assertion that, given any collection of mutually disjoint
nonempty sets, there is a set intersecting each member of the collection in exactly one
element. Clearly EACZ implies ACZ; the former is readily derivable from EAC and the
latter from AC. Since REMS is not a consequence of AC, it cannot, a fortiori, be a
consequence of ACZ. But, like EAC, EACZ can be shown to yield REM. We sketch the
argument, which is similar to the proof of Thm. 2(b).
Given a restricted formula ϕ, define
b = { x ∈ 2 : x = 0 ∨ ϕ }, c = { x ∈ 2 : x = 1 ∨ ϕ }

and a = {b, c}. A straightforward argument shows that a satisfies the antecedent of EACZ.
So, if this last is assumed, its consequent yields a u with exactly one element in common
with b and with c. Writing d and e for these elements, one easily shows that
(*) ϕ ∨ d ≠ e.

Now since it is also easily shown that ϕ ⇒ d = e, it follows that d ≠ e ⇒ ¬ϕ , and this,

together with (*) yields ϕ ∨ ¬ϕ.

 12

Acknowledgements

I am grateful to the (anonymous) referee for making a number of helpful suggestions for
sharpening the paper.

References

[1] Aczel, P. and Rathjen, M. Notes on Constructive Set Theory. Institut Mittag-Leffler Report no. 40,
2000/2001. Available on first author’s website.
[2] Bell, J. L. Toposes and Local Set Theories: An Introduction. Clarendon Press, Oxford, 1988. To be
republished by Dover, 2008.
[3] Diaconescu, R. Axiom of choice and complementation. Proc. American Mathematical Society 51,
1975, pp. 176-178.
[4] Gandy, R.O, On the axiom of extensionality, Part I, Journal of Symbolic Logic 21, 1956, pp. 36-48;
Part II, ibid., 24, 1959, pp. 287-300.
[5] Goodman, N. and Myhill, J. Choice implies excluded middle. Z. Math Logik Grundlag. Math. 24,
no. 5, 1978, p. 461.
[6] Maietti, M.E. About effective quotients in constructive type theory. In Types for Proofs and
Programs, International Workshop “Types 98”, Altenkirch, T., et al., eds., Lecture Notes in
Computer Science 1657, Springer-Verlag, 1999, pp. 164-178.
[7] Maietti, M.E. and Valentini, S. Can you add power-set to Martin-Löf intuitionistic type theory?
Mathematical Logic Quarterly 45, 1999, pp. 521-532.
[8] Martin-Löf, P. Intuitionistic Type Theory. Bibliopolis, Naples, 1984.
[9] Martin-Löf, P. 100 years of Zermelo’s axiom of choice: what was the problem with it? The Computer
Journal 49 (3), 2006, pp. 345-350.
[10] Scott, D. S. More on the axiom of extensionality. In Essays on the Foundations of Mathematics,
Magnes Press, Jerusalem, 1966, pp. 115-131.
[11] Tait, W. W. The law of excluded middle and the axiom of choice. In Mathematics and Mind, A.
George (ed.), pp. 45-70. New York: Oxford University Press, 1994.
[12] Valentini, S. Extensionality versus constructivity. Mathematical logic Quarterly 42 (2), 2002, pp.
179-187.
[13] Zermelo, E. Neuer Beweis für die Möglichkeit einer Wohlordnung, Mathematische Annalen 65 ,
1908, pp. 107-128. Translated in van Heijenoort, From Frege to Gödel: A Source Book in
Mathematical Logic 1879-1931, Harvard University Press, 1967, pp. 183-198.