The Existential Quantier in Equideductive Logic

(rst incomplete draft)

Paul Taylor
12 January 2011

1 Ideas of existential quantication

I would appreciate some help with the details behind these historical comments, especially in the
opening paragraphs and the second order (Russell??) denition of from .
The existential quantier

x. p(x) has been given many dierent syntactic and semantic mean-

ings in mathematics, logic and category theory.

constructive

Perhaps the most natural meaning is the

and a proof of

p(a)

in order to assert that

x. p(x)

one, in which we must exhibit a witness

[best ref ? Brouwer?]. At rst sight, this

quantier seems to have very little force, merely concealing the identity of
formalisation of intuitionistic logics in proof theory came the

a.

However, with the

existence theorem

that

x. p(x),
proof of p(a).

from any given explicit proof of

we may derive a term

and a

[Who rst proved this? Gentzen? Gdel?]

The constructive denition stood in opposition to the

classical

one, which said that a mathe-

matical object always exists so long as it is free from contradition. Formally, this seems to mean
that

[Is there a better explanation than this?]

While logicians debate the properties of formulae like

ordinary mathematicians work with an

idiom
x

an actual object. The phrase there exists

x. x

but also that we may proceed to

use

a p

Elments de Mathmatiques

for

p(a).

containing a

bound
x

variable

x,

is apparently

x means not only the formal statement

x that has this property.
which the operator  applied to a predicate p

such that
an object

David Hilbert proposed a calculus [ref ?]

yields a particular witness

x. p(x)

of discourse in which the witness

in

Boubaki subsequently began the rst volume [ref ] of

with this idea, using

instead of

.

On the face of it, the meaning

of the existential quantier therefore relies on the Axiom of Choice.

However, no

choice

permitted to make any

be a

variable.

of witness is actually needed.

analysis

of the value

a.

Notice, in particular, that we are not

So it is not a

term.

The symbol

must therefore

Indeed, the idiom may be reconciled with intuitionistic sequent calculus without making any
additional assumption whatever. The part of the idiomatic argument during which we pretend
that we have a witness

actually provides the formal proof of the premise (top line) of the rule

a,

p(a)

x. p(x)

q.

Then the consequent (bottom line) asserts that the proposition

hypothesis

x. p(x).

follows from the

unwitnessed

There

is

something more here than a proof of

from the witness

of

p(a),

just as there is

something more in a proof by induction than its base case and induction step. However, the extra
ingredient is not the axiom of choice but half of the denition of the existential quantier. The
proof rule above is found in any account of the
known as the

left

rule for

sequent calculus [ref, P+T will do], where it is

natural deduction is called the elimination

The corresponding idea in

rule.
The vernacular of mathematics therefore agrees with standard formal logic. This correspon-

dence is discussed in more detail in Section 5 and [ , 1.6].

The rule is, however, subject to the
of

a.

priviso

Indeed, if theorem with a parameter

that the conclusion

y :Y

must not export the identity

were to do this, it would exactly be asserting

Choice of the form

x: X . y: Y . p(x, y)

f : X Y . x. p(x, f x).

The meaning that emerges from the idiom and rule for
knowing

the

right to pretend

x: X . p(x)

is that

gives us

that we have a witness

in order to prove conclusions

for

p(a),

of a certain form.

In this paper we consider dierent (tighter, more specic) restrictions on the syntactic form
of the conclusion

and thereby obtain a dierent (weaker, more general) existential quantier.

class

This means that the

quantier

x. p(x)

q is being subjected to study.

all such q:


q. x. q(x) q q.

of permissible conclusions

The existential

is dened to have a certain property for

x: X . p(x)

means

This way of dening the quantier using second order logic is due to [whom? Russell? ref ?] and
is also found in [Prawitz?]. Under the CurryHoward correspondence between propositions and
types, it also appears in Girard's System F [ref ? P+T].
The purpose of this paper is to show that this syntactic idea of an existential quantier that is
dened in terms of the universal one agrees with the semantic notion of

epimorphism

in category

theory. Also, we shall not use second order logic or type theory but a new (weaker) predicate
calculus for which the category of sober topological spaces is a model (Section 2 and [equdcl]).
We introduce the categorical ideas by starting with elementary set theory. There the existential
quantier is embodied in the notion of a

e:XY

surjective function, writing

y: Y . x: X . y = ex.

to mean

When we generalise this from functions between sets to morphisms between other kinds of mathematical objects, we nd that the notion of surjectivity splits into several properties that are in
general inequivalent.

e : X  Y is epi if it has the cancellation property that,

g, h : Y Z of morphisms, if g e = h e then already g = h:

The morphism
any pair

X
On the other hand,

e:XY

is

e - Y

regular epi
K

for any object

e - Y
f

..
..
.. h
..
.?

Z
2

and

g - Z.
h

if it is the coequaliser of some pair

p - X
q

p, q : K X :

Being a coequaliser means that the composites

f p = f q

to another object

triangle commute (f

= h e).

ep = eq

are equal and, whenever the composites

are equal, there is a unique map

In algebra, coequalisers are called

h : Y Z

quotients

that makes the

and the algebra

is

often given as a congruence. For groups, rings and vector spaces there are simpler representations
of congruences that are peculiar to these theories, namely normal subgroups, ideals and subspaces.
Plainly we may reverse the arrows in each of these three denitions to obtain dierent notions

of

inclusion.

These dierent forms of surjection and injection do not behave well in

arbitrary

circumstances, for example regular monos do not compose in the category of additive monoids.
But in many mathematically important categories any morphism may be

factorised

as

a regular epi followed by a mono, or

an epi followed by a regular mono.

isomorphism theorems, saying that that image of a morphism is isomorphic to the quotient of its domain
In the settings of particular algebraic theories, results of the rst kind are often known as
by its kernel.
When we analyse this situation using category theory, we nd that the force of these theorems

orthogonality

lies in the

property:

e - I
...
...
...
...
...
...
...
...
...
...

..

?
J-

m - ?
Y
e = m f ) in which e has whichever
m has the injectivity one
ll-in h making the triangles commute

This says that, whenever we have a commutative square (g

surjectivity property (regular or plain epi) we are considering and

(mono or regular mono) then there is a unique diagonal
(h

e=f

and

m h = g ).

Applied to the category of sets, the factorisation theorems, and in particular the orthogonality
property, are directly related to the logical existential quantier:

the
the
the

epi

mono

corresponds to the

introduction rule, p(x) ` x. p(x),

corresponds to the conclusion

diagonal ll-in

to the

above, and

elimination rule.

This connection was established in a more general type-theoretic setting by Martin Hyland and
Andrew Pitts [ref ] and discussed further in [prafm, 9.3].
However, there is still a piece of the categorical description missing. In symbolic logic, we may

substitute

under the quantier:

[b/y] x. p(x, y) x. [b/y] p(x, y) x. p(x, b)

and the proof-theoretic analysis of this is called

commutation.

X0

...................
..
..
..
..
e0 ...
..
..
?
?
Y0

In the categorical setting, the substitution

function

X
e

?
f - ?
Y

[b/y]

is given by

pullback

or

inverse image

along a

(which is the reason for the star in the notation). For this formulation to agree with

the symbolic one, the pullback of an epi

as above must be another epi

e0 .

We would like to take these ideas that came from logic, set theory and category theory and
apply them to general topology. In that subject, all of the notions of injectivity and surjectivity

f : X Y is
mono i it is a 11 function on points;
regular mono i it is mono and X carries the subspace topology inherited from Y ; or
regular mono i it is surjective on points and Y carries the quotient topology inherited from X .

that we have mentioned arise, together with others. A continuous function

epis is more delicate. It depends on the particular category

i.e. the generality of the spaces over which Z above ranges. (This

However, the characterisation of

within which we are working,

is analogous to considering dierent classes of conclusions q.) Classically, a continuous function

f : X Y is epi i
it is surjective on points, in the case where all topological spaces in the traditional sense found
in [Bourbaki] are admitted for Z ;
every point of Y may be expressed as a directed join of images of points of X with respect to
the specialisation order, when all

it

sober topological spaces are allowed;

Hausdor spaces are allowed.

is surjective on points, when just

Since the rst and third cases are the same as in set theory, it is the middle one that is

domain $ of ascending natural numbers. This has a point

that is the directed join of the (nite) natural numbers. Classically, its open subsets are
n $ \ {0, . . . , n 1} (but not {}).

interesting. The basic example is the

called

and

- N

If we want to prove that two continuous functions

is enough to show that

g(n) = h(n)

$
6
6

1
6

.
.

g, h : $ Z are equal, where Z is sober, it

n N, because then f () = g() follows

for all (nite)

automatically. Indeed, most proofs in classical domain theory work by restricting attention to the
nite elements.
Therefore

e:N$

is epi amongst sober topological spaces [ , Lemma II 1.11], even though

it is not surjective on points. Categorically, this means that the pullback above is the empty space
and does not preserve the epi.
This situation is very similar to the vernacular use of there exists in ordinary mathematics
that we described earlier:
In order to show
it is legitimate to
It is manifestly false that

n. = en,

g = h,

pretend

x: $. gx = hx,
x = en for some n N.

that is,

that

so in order to avoid talking utter nonsense, the rules for

manipulating the new quantier must be weaker than those for the ordinary one. Indeed, when we
dene the new quantier and disjunction operation in Section 3, we shall nd that there are severe
restrictions on the variables and contexts that may occur in the proof rules. In order to remind
ourselves of these restrictions whenever we use the new quantier, we shall use a new symbol for
it, writing the property above as

x:$
(The letter

n: N.

x = en.

is an e sound in the Russian alphabet, where

is pronounced ye.)

Section 6 sketches the category that is associated with the logic, for which the equivalence was
proved in the earlier paper. Then

corresponds to a class of epis that is closed under products

but not pullbacks. However, it is still the case that every morphism factorises as an epi followed
by a regular mono and they obey the orthogonality property (Section 7). Sections 89 use these
ideas to show that the category has stable disjoint unions.
The constructive existential quantier with with we began this discussion is stronger than
the classical one because it requires the witness to be identied. Our new quantier, on the other
hand, is much weaker because there need not be any witness at all. However, we have seen that
it neverthless has its roots in the constructive traditions of proof theory and category theory.

2 Equideductive logic
The setting in which we dene this new existential quantier is not second order logic but a
new calculus called
sober

-calculus.

equideductive logic.

It is a predicate calculus whose object language is the

The origins of this logic, its syntax, its topological semantics and the equivalent

category theory are fully explored in the paper

Equideductive Categories and their Logic.

The following remarks were just put here as the targets of cross references in this paper since
it was put together from sections taken out of the one mentioned.

This section will become a

summary of equideductive logic, at whatever level of formality turns out to be appropriate as the
programme develops.

Remark 2.1 We refer to the types of the object language as urtypes in a syntactic setting and
urspaces in a semantic one. In the leading classical model (sober topological spaces), the urspaces
may be either algebraic lattices with the Scott topology or locally compact spaces. The system of
urtypes must admit products and exponentials of the form
the symbols

?, h , i, 0 , 0 , , ev

Remark 2.2

Urtypes are

and

() ,

so its urterms are formed using

focus.

sober . This means that for any term ` P :

, : 3 A

that is

prime ,

P = P x. (. x) ,

we may form

` focus P : A
In particualr, for any

` a : A,

the term

` P . a

Remark 2.3 Any exponential urtype B

Remark 2.4
,

Equideductive

, : A ` (focus P ) = P .

such that

is prime.

is syntactically injective.

predicates are formed from >, equations between urterms of type

conjunctions (&) and quantied implications

~y . ~q(~y ) ==. ~x~y = ~x~y

subject to the

variable binding rule that any variable that appears on the left of =. must be

bound by the quantier

Several

denitional extensions

were made in the earlier paper (and others are added in this

and later ones) that allow predicates instead of just equations on the right of
any predicate is nevertheless equivalent to one in normal form,

i.e.

=..

However,

a conjunction of quantied

implications with equations on the right and similar normal forms on the left.

Remark 2.5 Equality

between urterms is extended from

to a predicate on any urtype

the Leibnizian formula

a=b

: A . a = b.
5

by

In particular, equality for

, : A
,

satises

a`

as well as the usual equality-transmitting and

x: A. x = x

-rules

for the sober

()
-calculus.

Any predciate respects equality in the sense that

, p(a), a = b ` p(b).

Remark 2.6 Any proof of

x : A, y : B,
y : B, z : C,

p(x)

r(z)

q(x, y)

s(y, z)

may be translated into a proof of

x0 . p(x0 ) =. q(x0 , y)

` z. r(z) =. s(y, z)




x0 . p(x0 ) =. q(x0 , y) ==. z. r(z) =. s(y, z) ,

y : B,

and

y:B

or of

vice versa.

3 Disjunction and existential quantication

Now we are ready to dene the new existential quantier

 and disjunction g and show that they

satisfy rules that are similar to those in the sequent calculus, albeit with severe restrictions on the
use of variables.

Notation 3.1 We want the new connectives , g and  to satisfy the rules
~ ~s(~x),
~x : A,

~ ~y : B,
~ ~z : C,
~
~x : A,
s(~z), p(~x, ~y ) ` r(~x, ~z)
=======================================
~ ~z : C,
~
~ p(~x, ~y ) ` r(~x, ~z)
~x : A,
s(~z), ~y : B.

r(~x)

~ ~z : C,
~ s(~z), p(~x) ` r(~x, ~z)
~ ~z : C,
~ s(~z), q(~x) ` r(~x, ~z)
~x : A,
~x : A,
==========================================================
~ ~z : C,
~
~x : A,
s(~z), (p g q)(~x) ` r(~x, ~z)

and

~ and B
~ be ur types and that there be no other predicates that
A
~x or ~y . However, the predicate r(~x, ~z) on the right may depend on additional variables ~z
hypotheses s(~
z ), but the latter must not involve ~x or ~y . It is because of these issues with

We shall see that it is essential that

involve
and on

free variables that we state them explictly in the predicates.

=., we consider just equations on the right rst and develop

r(~x, ~z) from this simple case. We obtain the denitions of the

As we did when we introduced

the results for general predicates

new symbols by applying the technique of Proposition 2.6 to the prospective rules above, with the
equation

~x = ~x

instead of

r(~x, ~z)

on the right.

~ , with no predicates
Notation 3.2 In the context [~x : A]

on

~x,

we dene

: . =
~

(p g q)(~x) : A . (~x0 . p(~x0 )

=. ~x0 = ~x0 )
00
00
& (~x . q(~x ) =. ~x00 = ~x00 ) ==. ~x = ~x
~ p(~x, ~y ) : A~ . (~x0 ~y . p(~x0 , ~y ) =. ~x0 = ~x0 ) ==. ~x = ~x.
~y : B.
6

For clarity, we drop the strings of variables, since they can be recovered using product urtypes.

Lemma 3.3

These connectives obey the following introduction rules:

x : A, p(x) ` (p g q)(x)
x : A,

Proof

For (I ),

y : B,

x : A, q(x) ` (p g q)(x)
`

p(x, y)

y. p(x, y).

I

gives

, : A , x0 y 0 . p(x0 , y 0 ) =. x0 = x0 `


0
0
0
0
Hence . x y. p(x , y) =. x = x ) =
=. x = x, which is y. p(x, y).
Similarly, for (gI0 ), by E we have
x : A, y : B,

p(x, y),

x : A,
Weakening by

gI0 , gI1

x0 . p(x0 ) =. x0 = x0

, : A ,

p(x),

x00 . q(x00 ) =. x00 = x00

and using

x = x.

x = x.

then give the denition of

(p g q)(x).

Lemma 3.4 Conversely, the connectives also obey the simple elimination rules, with just equations
on the right:

x : A, , : A , s0 (, ) ` x = x
x : A, , : A , s0 (, ), p(x) ` x = x

x : A, , : A , s0 (, ), (p g q)(x) ` x = x
x : A, y : B, , : A , s0 (, ), p(x0 , y) ` x0 = x0
0

and

s0 (, ),

x : A, , : A ,

Proof
By

x : A, , : A , s0 (, ), q(x) ` x = x

For

I ,

E ,

use

with

and

y. p(x, y)

x = x

x.

the rst premise of the rule for disjunction gives

, : A ,

s0 (, )

x0 . p(x0 ) ==. x0 = x0 ,

s0 (, ) must not depend on x0 , and the second premise provides the same for q.
E to the denition of (p g q) and these two formulae gives x = x.
Similarly, the premise of the rule for  gives

in which

, : A ,
in which

s0 (, )

x = x,

s0 (, )

must not depend on

x0

`
or

y.

Applying

x0 y. p(x0 , y) ==. x0 = x0 ,
Applying

to the denition of

with this gives

as required.

Theorem 3.5 The connectives , g and  obey the general elimination rules in Notation 3.1.
Proof Suppose rst that r(x, z) is the equation xz = xz . Then the three general rules follow
from the corresponding simple elimination rules with the substitutions

x. xz,

x. xz

and

s0 (, ) s(z).

The case of a quantied implication,

r(x, z) z. s(z) =. xz = xz,

follows from that of an equation by
separately, whilst the case
for

r>

I .

When

is a conjunction we consider the conjuncts

is trivial. By Remark 2.4, these cases exhaust the possibilities

r.
7

4 Properties of the new symbols

These connectives have many familiar properties that can be proved easily from the introduction
and elimination rules.

Lemma 4.1

Disjunction and quantication are covariant:

x : A, p(x) ` p0 (x)

x : A, q(x) ` q0 (x)

x : A, y : B, p(x, y) ` p0 (x, y)

x : A, (p g q)(x) ` (p0 g q0 )(x)

Lemma 4.2

a`

>,

a`

pgp

and

p & (p g q)

a`

p g (p & q)

a`

p g ,

commutative and associative,

p g > a` >
where

Disjunction is idempotent, absorptive, unital,

p
absorbed by

x : A, y. p(x, y) ` y. p0 (x, y)

p g q a` q g p

p g (q g r) a` (p g q) g r

may depend on any variables of urtype but not on any predicates as hypotheses.

&

We nd that the distributive law (for

over

g)

and its analogue for

(which is known

in categorical logic as the Frobenius law) can only be proved when the extra conjunct has no
arguments in common with the disjunction:

Proposition 4.3 The weak distributive

x : A, z : C,
disjoint). For

(p & r) g (q & s) (x, z),


y. p(x, y) & s(z) ,

(p g q)(x) & s(z)

a`


y. p(x, y) & s(z) a`

x : A, z : C,

in which the variables

Frobenius laws hold:

and

and

must be distinct (or, in the case of strings

~x

and

~z

of variables,

there is no diculty regarding variables; it is strict because

(x) & s(y) a` ( & s)(x, y) a` (x, y) a` .

Lemma 4.4

The quantiers satisfy the de Morgan laws

y : B,
y : B,

and

x. p(x) =. r(y)

x. p(x) =. r(y)

a`

a`

(p =. r(y)) & (q =. r(y))

(p g q) =. r(y).

Corollary 4.5 Successive quantiers commute and are the same as multiple ones.
Proof


0 0
0
0 0 0
0
0
zy. p(x, y, z)

a`

. x z . y . p(x , y , z ) =. x = x

=. x = x


. x z . (y . p(x , y , z )) =. x = x =. x = x


z. y. p(x, y, z) .
0 0

Whereas the double quantier is a single one over the product of the urtypes, the nullary
version quanties over

Lemma 4.6

1.

This is a kind of

double negation .

x : A, p(x) a` y: 1. p(x),

which is

. (x0 . p(x0 ) =. x0 = x0 ) ==. x = x.

8



y. p(x) ` p(x) in the case where p(x)
y. q(y) =. xy = xy , in order to see that this can only be done if q does not depend on x.
Instead of y : 1 we could quantify a variable y of any urtype where y does not actually occur
in p. This is the counit of the adjunction between  and weakening:
It would be a useful exercise to prove that

Corollary 4.7
Corollary 4.8

x : A, y: 1. p(x, y) a` y: 1. p(x, ?) a` p> (x, . ?) a` p(x, ?).

, x : A, p(x) a` y: A.

p(y) & (x = y) .




This double negation formula will turn up again all over the place.

In a slightly more

complicated form, it will also be used in the construction of exponential spaces in the following
paper.

Notation 4.9 For

F :

Section 7 explains what

(F ) . (x. p(x) =. x = x) ==. F = F .

p
means in terms of epimono factorisation and in Section 8 we use

algebraic inconsistency () to dene the initial object, and

and

for the coproducts.

5 Idioms of reasoning
This section will become an explanation of how the new existential quantier could be incorporated
into natural dededuction and the idioms of the vernacular of mathematics as described in my book.
At the moment it just consists of a collection of relevant fragments of text that were originally
written in other contexts.
For the purpose of presenting the calculus, it is clearer to use judgements (sequents), in which
the contexts are stated in full in each step. However, when we use the rules to develop ordinary
mathematics, we would like to adopt a Natural Deduction style, in which the contexts remain the
same from one step to the next unless we explicitly make or discharge assumptions and variables.
Since these must obey last-in rst-out scoping rules, one way of formalising them is by means of
boxes that delimit the scopes; this style is described in [prafm, 1.5].
The relationship between the formal rules for the
is explained in [prafm, 1.6]. As with

ordinary quantier and mathematical idiom

the rst step is to translate the sequent style into boxes

that delimit the scope of the witnesses. Then we observe that it doesn't matter when the box is
closed (discharging the hypothetical witness), because everything that follows the end of it can be
brought inside instead. That is, so long as
(a) we do not export the identity of the witness from the box; and
(b) it is closed before the next enclosing one.
The box is therefore redundant, and we may instead pretend that we have a witness, just as
ordinary mathematicians do when they say that there exists something satisfying the predicate.
How can we adapt these ideas to the new quantier and disjunction operator? Whereas there
were restrictions on the
use the

elimination

introduction

rules for

and

of

, focus

and

The necessary additional precautions are that

(c)
(d)

=.,

we now have to be careful about how we

.

substitution for x into y. p(x, y) is not allowed; and

elimination of y. p(x, y) is only allowed when the context contains no other predicates that
have

as a free variable.

Since we cannot deduce anything from

least, in any role for which

y. p(x, y)

if it violates this restriction, it is useless (at

is an appropriate notation).

So we may as well strengthen the

precaution (d) by putting it into the same pattern as those on

(e) we may only

form y. p(x, y)

in a context in which

and

=.,

namely that

is of urtype, and not subject to any

other predicate.
This completes our account of the purely syntactic aspects of the logic.

Section 7 explains

 means in terms of epimono factorisation. In Section 8 we use algebraic inconsistency ()

dene the initial object, and g and  for the coproducts.

what
to

6 Equideductive categories
This section will give a brief sketch of the construction of the classifying category for equideductive
logic.

The full version is in the paper

Equideductive Categories and their Logic.

material in the

Examples

Remark 6.1

Recall that we used the name

Some of the

section of that paper may be moved here.

object language of equideductive logic.

urtype

in the sober

-calculus

that serves as the

The reason for this is that we dene a

type

to be an

urtype together with a predicate,

{x : A | p(x)},
using the familiar comprehension or subset-formation notation from elementary set theory.
Formally, these types behave in essentially the same way as contexts. We may also quantify over
them, writing

x: {x : A | p(x)}. q(x)
x: {x : A | p(x)}. q(x)
These types provide the objects of the
Morphisms are

represented

for
for

x: A. p(x) ==. q(x)

x: A. p(x) & q(x).

category of contexts and substititions for the logic.

by uterms,

- {y : B | q(y)}
{x : A | p(x)} .....
?
?

?
A

f - ?
B

where

x : A ` fx : B

such that

x : A, p(x) ` q(f x).

The same morphism may have many representing urterms, where

f =g

if

x : A, p(x) ` f x = gx : B,

the latter being understood as the Leibnizian equality in Remark 2.5.

However, this simple denition of morphisms is based on the assumption that the urtype

is

injective . For our purposes, it will be enough to take B C , so targets of maps are wlog of the

10

form

{C | r}.

Remark 6.2

partial product , for

orthogonal to partial product inclusions.

In the category, quantied implication is represented by a

which see the main paper. Epis are

Remark 6.3 The functor () : B Bop reects invertibility.

Remark 6.4 The category Sob admits factorisation into epis (which are preserved by products)
and regular monos (inclusions with the subspace topology).

Remark 6.5 Any equaliser is a partial product, so any regmono is in M, and conversely in Sob.
Remark 6.6 However, in the term model the class of regular monos is properly contained in M.
Having a cokernel doesn't help.
We need an example for which the predicate

y. y = y ==. xy = xy
is not equivalent to

y. xy = xy .

7 Factorisation
As our rst application of the new quantier to category theory, we show in this section that it
agrees with the epis in the category, as we claimed in the Introduction.

Epis are preserved by

products and provide one of the classes of a factorisation system.

Notation 7.1 Let E

be the class of morphisms

x : A, p(x) ` q(ex)

e : X {A | p} Y {B | q}

for which

y : B, q(y) ` x. p(x) & y = ex ,

and

where the rst condition is just the denition of a morphism. We shall show that

is the class of

epis.
We also write

for the class of monos that arise from partial product diagrams. These are

isomorphic to canonical inclusions

{A | p}  {A | q}

with

x : A, p(x) ` q(x).

The class

is

therefore closed under composition.

All regular monos are in

Lemma 7.2

but the converse need not be true.

The second condition on

E -maps

y : B, q(y) ` , : B .

Proof

is equivalent to

x: A. p(x) =. (ex) = (ex) ==. y = y.

This is because

xy. p(x) & (y = ex) =. y = y

a`

x: A. p(x) =. (ex) = (ex).

Lemma 7.3 Product with any type Z {C | r} preserves the E -property.

Proof For any type Z {C | r}, Corollary 4.8 says that
z : C, r(z)

z 0 . r(z 0 ) & (z = z 0 ).
11

So, given

eE

as above, we use Proposition 4.3 twice and Corollary 4.5 to deduce

x. p(x) & (y = ex) & z 0 . r(z 0 ) & (z = z 0 )


z 0 . x. p(x) & (y = ex) & r(z 0 ) & (z = z 0 )


z 0 . x. p(x) & (y = ex) & r(z 0 ) & (z = z 0 )


xz 0 . p(x) & r(z 0 ) & hy, zi = hex, z 0 i ,

y : B, z : C, q(y), r(z) `
`
`
`
so the map

(e Z) : X Z  Y Z

Proposition 7.4 A map e : X Y

is also in

E.

belongs to

i it is

that is, it has the cancellation property that, whenever

Hence the class

Proof

epi :

C
- Z { | s}

e- Y {B | q}

X {A | p}

e = e,

already

- C
= .

contains all isomorphisms and is closed under composition.

Suppose that

e:XY

is epi and let

Z .

We may express the cancellation for this

case in equideductive logic as

` : B .
Using

x: A. p(x) =. (ex) = (ex) ==. y: B. q(y) =. y = y .
` to get


x: A. p(x) =. (ex) = (ex) ==. y = y,

(Proposition 2.4) we may move

y : B, q(y) ` : B .
which is the

y: B. q(y) =.

behind the

E -property.

Conversely, suppose that

e E , so e C E

too for any urtype

by the Lemma. By reversing

the previous argument, we have the cancellation property in the diagram

eC
-

X C

Y C

- .

Z C ,

The exponential transpose of this is the epi property in the case

follows from this because any object

has a mono

Z ,

but the general case

Conversely, we may recover the existential quantier from the epis:

Proposition 7.5 Given any predicate y : B,

z : C, s(y, z) ` q(y),

y, B, q(y) ` z. s(y, z)

0 : {y : B, z : C | s(y, z)} {y : B | q(y)}
Corollary 4.8, the condition for 0 E is that (s ` r

holds i the map

Proof

By

y : B, q(y)

y 0 z. s(y 0 , z) & (y 0 = y)

Proposition 7.6 Any map factorises as f

=me

with

eE

is epi.
and)

a`
and

z. s(y, z).

m M.

em
X {A | p} Y {B | r}- - Z {B | q}
?
?
?
A

f
12

?
- B

Proof

Let the intermediate object

image {B | r}, where

be the

y : B ` r(y) x: A. p(x) & (y = ex),

ef :X Y

so we already the second condition for

to be an epi. The rst, that it is a well

dened map, is

x : A, p(x) ` r(f x),

which is

I .

The inclusion

m:Y Z

is also well dened because

y : B, r(y) ` q(y)
by

E ,

from the dening property of

f :X Y,

which was

x : A, p(x) ` q(f x).

The universal property of this factorisation is expressed by

Proposition 7.7 The classes E

and

are

orthogonal , cf. Proposition 6.2.

e

A

eE


X {A | p}

- B
- Y {B
| q}

.
....
....
....
.
.
.
h.........
.
....
.
.
.
..
....
.
.
.
...
....

mM

?
U {C | r}

?
- V {C
- | s}

- ?
?
C ============================================== C
in any commutative square like the one in the middle of this diagram, there is a unique map

h:Y U

Proof

that makes both triangles commute.

Since the inner square commutes and

x : A, p(x) ` r(f x)
x : A, y : B,

so
Since

e E,

and then

m M,

by Lemma 2.5,

f x = g(ex),

and

so

r g(ex) .

(y = ex) ` r(gy).

p(x),

y : B, q(y) ` x. p(x) & (y = ex),

E

y : B, q(y) ` r(gy),

gives

so (the given representative of )

also serves for

h.

Theorem 7.8 Any equideductive category has a factorisation system in which

(a)
(b)

E consists of the epis and is stable under product; and

M is the class of maps that arise from partial products.

Corollary 7.9
which

, g

and

In any particular context

equideductive predicates form a lattice


Sub(),

in

are joins.

This lattice need not be distributive, or even modular. Logically, these symbols allow weakening
but not contraction or cut of variables. They only allow weakening by a hypothesis if it has no

13

variable in common. Topologically, they are preserved by products but not by pullbacks or inverse
images.

Proof

The foregoing results only apply to contexts consisting only of urtypes. A general context

[x : A, p(x)]

is a subspace of an urtype

A,

and

Sub([x : A, p(x)]) = Sub(A) p

is still a lattice.

Warning 7.10 Although our new connective  is very useful, it is not the same as the usual one
extremely careful in using it.

on points, and so we must be

Example 7.11 Let f

: N N

be

f n m. m < n.

Its factorisation

- $ { : N | n. = m. m < n}-

- N

denes the ascending natural number domain when this diagram is interpreted in

Sob.

: $ ` n. p(n, )
where

p(n, ) ( = f n)

and

but
n : N, p(n, ) ` ,
W
m. > =  n m. m < n.

However,

Questions 7.12 (Giuseppe Rosolini)

(a) The variable rule for

is reminiscent of linear (rather, ane) logic; is there a monoidal

structure that would explain this?

(b) Does the counit of this structure (Corollary 4.7) throw light on the need for the variablebinding rule in order to interpret

y : 1?

(c) Can one characterise the maps along which epis,

or the factorisation

can

be pulled back?

The class of such maps may provide the dependent types.

Remark 7.13 There is another existential quantier for maps with dense image:
di y: B. p(x, y) : A .

x0 y 0 . p(x0 , y 0 ) =. (x0 ) = ==. x = .

8 Coproducts
In the second application of the new quantier and disjunction, we now show that the category
has nite coproducts.

Proposition 8.1 The initial object is 0 { | }.

0 { | }
?

Proof

...........................

{C | r}
?

: x 7 c. x -

The dotted map is well formed because, by

E ,

. =

x : ,

?
C

r(c. x).

It is unique because

x, y : ,

14

c. x = c. y.

Notation 8.2
(A B )

H:

: ( ),

0 a

along with subscripts. Then dene

. a
a. (0 a)
q

x. p(x) =. 1 x = 2 x

X +Y

{H | prime(H) & [p, q](H)}

X+Y

1 2

y. q(y) =. 1 y = 2 y

prime(H)

[p, q](H)

. b
b. (1 b)

1 b

1 2

Of course,

a : A, : A , b : B , : B ,

The urtypes of the variables that we use are

. H = H

1 2 1 2 . (1 2 ) & (1 2 ) ==. H1 1 = H2 2 .

is for now just an abbreviation: we have to prove that it has the universal

property of a coproduct, but rst we explore the basic properties of this notation.

Lemma 8.3

This notation satises

>

a`
[>, >] a`

Proof

[p, q](0 x) a`
[p, q](1 y) a`

=
>

p(x)
q(y).

The rst part is Axiom 2.1 and the second follows from this. For the third,

[p, q](0 x)

1 2 1 2 . (1 2 ) & (1 2 ) ==. 0 x1 1 = 0 x2 2

a`

1 2 . (1 2 ) ==. 1 x = 2 x

p> (x)

a`

p(x)


from Lemma 4.6, and the last one is similar.

Lemma 8.4

The inclusion maps

0 : X X + Y

and

1 : Y X + Y

are well dened, mono

and natural.

Proof

H 0 a . a

is prime because, for any

(0 a) (. a) a (0 a).
We have already shown that if

0 a = 0 b

p(x)

then

[p, q](0 x).

The map

a`

. a = . b

is mono because

. a = b

a`

a=b

by Lemma 2.5 and Denition 2.5.

0 -

 1

?
C
D
1
0C

For naturality,

for

(0 a)

?
D

. 0 a(f )(g )

. (f )a

. (f a)

0 (f a).
x {A | p}

The key step in showing that

X +Y

y {B | q}

in an essentially familiar way, but using the new connectives

exhaust

X +Y,

is the coproduct is that the

15

0 x


for

and

1 y
g

and

from Section 3. In traditional categorical language the inclusions

X X +Y Y

jointly epi , so this is the dual of extensionality for products (Axiom 2.1).
Proposition 8.5 For H :

prime(H),

Proof

Let

[p, q](H)

be the context

are

x. p(x) & H = 0 x g y. q(y) & H = 1 y .


H, prime(H), [p, q](H) on the left.
a`

The backwards direction is the easy one. By Lemmas 2.2, 2.5 and 8.3,
A

H :

, x : A, p(x), (H = 0 x)
A

H :

Then

, x. p(x) & (H = 0 x)

A A

H:

and

prime(H) & [p, q](H) .

, RHS

E

gE

Conversely, to prove the disjunction on the right (Notation 3.2), we consider

1 , 2 : 2 (A B )

such that

H 0 . (x. p(x) & H 0 = 0 x) ==. 1 H 0 = 2 H 0

0

H . (y. q(y) & H = 1 y) ==. 1 H = 2 H ,

and

[1 , 2 , , ] for this context. We must

x : A with p(x), H 0 0 x . x satises

so we write
For

x : A,
by

I

(Lemma 3.3), whence

, x : A, p(x)
where

()

p(x)
gives

()

, ` 1 H = 2 H .

x0 . p(x0 ) & (H 0 = 0 x0 )

1 H 0 = 2 H 0

in

1 x 1 x 1 (0 x) 1 H 0 = 2 H 0 2 x,

was dened in Notation 8.2.

` x. p(x) =. 1 x = 2 x, which is 1 2 , and 1 2 is similar. This means that

invoke [p, q](H) for the inner equality below, together with prime(H) for the outer ones:

Hence
we may

deduce that

()

1 H = H(1 , 1 ) H1 1 = H2 2 = 2 H,


which is what was required to prove the disjunction.

Theorem 8.6 X + Y

is the coproduct.

1
0
X + Y .
....... {B | q} Y
...
?
f
h ........?
........
....
g
{C | s} Z
?
?
?
?
A
B

01
A
B
.....
.......
.......
f
..h ..
........
?
........ g
...
C

X {A | p}
?

Proof

Z is a canonical type, i.e.

g : Y Z {C | s}, so

By Proposition 6.1, without loss of generality the test object

subspace of some

with

C A.

Given

f : X Z {C | s}

x : A, p(x) ` s(f x)

and

16

and

y : B, q(y) ` s(gy),

the mediator is necessarily the restriction of some

h0 : (

B )

C ,

by injectivity of

(Lemma 2.1). Indeed,

hH c. H(a. f ac, b. gbc)

makes the diagram commute from

or

to

h(0 a) = c. f ac = f a

because

and

h(1 b) = c. gbc = gb.

h is a well dened map X + Y Z

We must show that this

and that it is the only one that makes

the triangles commute. These are both applications of the general elimination rules for

and

in Section 3.
Since

h(0 x) = f x

and

s(f x),

by Lemma 2.5,

H, x : A,

p(x),

H = 0 x

s(hH)

x. p(x) & (H = 0 x) ` s(hH)

x. p(x) & (H = 0 x) g y. q(y) & (H = 1 y)
H,

H,

H,
by Proposition 8.5, so

prime(H),

H,

H,

q(y)

and

h0 = h

H = 0 x

p(x),

gE

s(hH)

s(hH)

h0 H = f x = hH

given

h0 H = hH

x. p(x) & (H = 0 x)

H = 1 y ,

E

so

x. p(x) & (H = 0 x) g y. q(y) & (H = q y)
H,

Hence

is well dened. The uniqueness argument is similar:

H, x : A,

and similarly with

[p, q](H)

E

prime(H),

[p, q](H)

h0 H = hH

gE

h0 H = hH.


in the sense of Denition 6.1.

9 Extensivity
Finally we show that the coproducts that we constructed in the previous section are stable and disjoint, and the initial object is strict. We do this using the modern categorical notion of

extensivity

[Cockett etc refs]:

Denition 9.1 A category is called extensive

if, in any diagram

- Z 

k
?
1

?
- 2

?
1

the top row is a coproduct i the two squares are pullbacks.

Give a brief introduction to extensivity and distributivity, showing that the former entails that
the initial object is strict and coproducts are stable and disjoint (Corollary

??).

We do, however, have to make a small additional assumption:

Lemma 9.2

If an equideductive category

has disjoint coproducts then

17

has a point.

Proof

012
1 .

If coproducts are disjoint then

respect to

01

provides a point

is an equaliser. Then injectivity of

with

Proposition 9.3 For any urtype A, the type 0 {A | } is a strict initial object, and
`

x:0
for any predicate

p.

p(x)

h C
0 {A | } 
............... { | s} Z
?
?
k
?
h
?
- C
A

............................
k

Proof

x : 0 ` p(x) is x : A, ` p(x), which is E (Notation 3.1). The map

h ac. ?, and it is unique because x : 0 ` hx = h0 x.
If there is a also map k : Z 0 then we already have k h = id0 by uniqueness of maps out
0. But also, z : Z ` (kz) ` h(kz) = z , so Z

= 0.

0Z
of

The judgement

is given by

Notation 9.4

2 1 + 1 { | prime}

with elements

0 .

and

1 . .

Proposition 9.5 These satisfy

(a)
(b)
(c)
(d)
(e)

H : , prime(H), : ` = H ;
H : 2, prime(H) . H = H(0 , 1 ) ` (H = 0) g (H = 1);

the map 2 
is mono;
x: 2. p(x) a` p(0) & p(1); and
x: 2. p(x) a` p(0) g p(1).

Proof

Although [b] is an example of Proposition 8.5, the restriction on the use of

stops us

from substituting into it. This is why we prove [a] rst, using Proposition 8.5, and [b] follows from
this. [c] This map is in

M.

[d,`] By

x : 2, p(0), p(1)

E .

[d,a]

(x = 0 g x = 1) & p(0) & p(1)

(x = 0 & p(0)) g (x = 1 & p(1))

by weak Frobenius (Proposition 4.3), Lemma 2.5 and

denitions of

and

gE .

[e,a] By

I .

p(x)
[e,`] Using (d) and the

g,

x: 2. p(x)
`

. (x: 2. p(x) =. = ) =. =


. (p(0) =. = ) & (p(1) =. = ) =. =

p(0) g p(1).
The consequence of these properties of
of

Z.


2

That is, we may use pullbacks to split

is that any map

Z 2

gives rise to a

partition

into two parts, and recover the original object as

a coproduct.

18

If the two squares in the denition of extensivity above are pullbacks then the top row is a
coproduct.

Lemma 9.6

Z {z : C | r(z)}
= {z : C | r(z) & kz = 0} + {z : C | r(z) & kz = 1}.
Z {C | r}- - C
k

?
-

?
2-

Proof

By sobriety, the two sides of the claimed isomorphism are

C

= {F : | prime &r> }

Z {C | r}
W {C | r & k = 0} + {C | r & k = 1}

C C

{H :

Lemma 6.1

| prime &[r & k = 0, r & k = 1]}.

We must therefore show that there is a bijection given by

F . H
[F 7 H 7 F ]:

H . F z. kz(z)(z) .

and

From Proposition 9.4(b) we have

: , prime(), :
Combining this with the dening property of

gives
which means that

If

k : Z 2  ,
`

prime(kz)

z : C, r(z), : C

z = kz(z)(z),

` z. kz(z)(z).

H = 0 z

= .

z : C, r(z)

prime(F ), r> (F )
[H 7 F 7 H]:

with

r(z)

then

Hence

F = . F z. kz(z)(z) .

F = z

H, z. r(z) & (H = 0 z) & (kz = 0)

with

prime(F )

and

r> (F ),

so by

E ,

. F (z. z) = . z 0 z = H.

prime(H), [r & k = 0, r & k = 1] `

z. r(z) & kz = 0 & H = 0 z g z. r(z) & kz = 1 & H = 1 z .

By Proposition 8.5,

Since we can recover

in each case,

H 7 F 7 H

by

gE .

The converse of this is that the naturality squares for the coproduct inclusions are pullbacks.
If both rows are coproducts then the squares are pullbacks.

Lemma 9.7

The square on the left is a pullback:

A
B
0
X {A-| p} - X + Y { | prime &[p, q]}

?
1

- A B
-

?
- 219

k
?
-

Proof
A

from

The downward maps called

to

take

to

. ,

both take

to

. H(a. , b. ). Then
2

so the quadrilateral commutes. Also,

both maps
is mono by

construction, so the left-hand square commutes too.

We need to dene a unique map
A

{H :

so we let

| prime(H) & [p, q](H) & kH = 0} {A | p},

be the context (object) on the left, in which

kH = 0

a`

. H(a. , b. ) = . .

First we exclude the possibility that

, H = 1 y
Since

0 = kH = 1

p g a` p (Lemma 4.2), Proposition

E , we deduce from this and

a`

. = .

. = .

x. p(x) & (H = 0 x).

8.5 gives

Using

H 0 :
that

, x : A, (H 0 = 0 x)

H 0 = . H 0 (b. ?)

H = . H(b. ?) = . F where F . H(b. ?) and x. F = x.

F is prime, so by sobriety (Axiom 2.2) we may introduce


` a focus F focus . H(b. ?) : A,

Hence

F = a,

with

H = 0 a

and

[p, q](0 a) p(a)

by Lemma 8.3. It is unique by Lemma 8.4.

Theorem 9.8 The category is extensive.

Corollary 9.9

Coproducts are stable and disjoint, the initial object is strict and product dis-

tributes over coproduct: the map

(X Z) + (Y Z)

[h0 0 ,1 i,h1 0 ,1 i]

- (X + Y ) Z


is an isomorphism.

Proposition 9.10 The coproduct of any list of urtypes exists and has an exponential that is itself
an urtype:

0
= 1,

A+B
= A B ,

Ai

Ai .

Proof

Using distributivity, there is a natural bijective correspondence amongst maps

( A) + ( B)
= (A + B)
============================
( A)
( B)
========
==
========
==
A

B
============A
==========
B
and similarly between the unique maps

0
=0

20

and

1 0 .

10 Overt spaces
Depending on whether I decide to publish this paper separately from or alongside the others in the
equideductive programme, it may be appropriate to add a narrative section about how the logic
is developed into one for topology. In particular, an overt space is one for which the quantier
on

predicates

is represented by an

urterm.

Then something could be said about the role of overt

spaces in the connection between general topology and recursion theory.

11 Recursion
This section is just parked here temporarily. It will form the core of another paper about discrete
mathematics in equideductive topology.

Axiom 11.1

primitive recursion

We express

calculus by adding the base (ur)type

` 0:N
The

NE -rule

is the

over the natural numbers in the restricted

and the introduction rules

and

n : N ` n + 1 succ n : N.

NI

recursion scheme , which we formulate with active and passive parameters

as follows. For any urtype

and urterms

` z:B

and

, n : N, b : B ` s(n, b) : B,

there is an urterm

, n : N ` r(n) rec(n, z, s) : B
that has the property that

` r(0) = z : B
The

rec

and

, n : N ` r(n + 1) = s n, r(n) : B.

construction also respects equality (NE=) and the

In order to use

N -rule

is

rec(n, 0, succ) = n.

in equideductive logic, we also need the

Axiom 11.2 The induction scheme for N says that, for any predicate r(n) on N (with no other
parameters),

r(0),

n: N. r(n) ==. r(n + 1)

n: N. r(n).

Remark 11.3 For logical and programming purposes it is more convenient to dene combinatorial
structures using the urtype T of binary trees. These rename 0 as nil , and have a binary constructor
[ | ]

instead of the unary

succ.

The recursion scheme is easily adapted from that for

N,

whilst

the induction scheme,

r( nil ),

xy: T. r(x) & r(y) ==. r([x | y])

z: T. r(z),

obeys the variable-binding rule.

On the other hand, we cannot write

x. y. y x =. r(y) ==. r(x)
to dene a well founded relation

Lemma 11.4

N
=1+N

z. r(z)

because the variable-binding rule for

(this is extensionality).

21

is violated in

y x.

Proof

n : N ` (n = 0) g (m. n = m + 1)

is

: N . (n0 . n0 = 0 =. n0 = n0 ) & (n. (m. n = n + 1) =. n0 = n0 ) ==. n = n,

which is an easy case of the induction scheme.

Theorem 11.5 In the category CnL , N is a parametric natural numbers object.

succ
N 
- ...

Proof

S
Y {B | q} 
?

?
NY
?

?
- B 

?
A N B

?
A

..
..
..
..
.. hid, Ri
..
..
..
.

..
..
..
.. R
..
..
.?

= {A | p}
?

Categorically, the data are as shown above, and we have to nd the map

R.

Symbolically,

we are given

x : A,
x : A, p(x),

and

n : N,

: ,

x : B ,

p(x)

q()

q(x )

DZ
B

Sx (n, ) : ,

q Sx (n, ) .

DS
x and
-calculus,

By syntactic injectivity (Theorem 2.3), without loss of generality, formation of the urterms

Sx (n, )

does not depend on the hypotheses. Therefore, by recursion in the restricted

there is a unique urterm

x : A, n : N
`

x:A

such that

Rx (0) = x : B

RZ

Rx (n + 1) = Sx n, Rx (n) .

x : A, n : N

and

Rx (n) : B

RS

Now consider the equideductive predicate

n : N ` r(n) x: A. p(x) ==. q Rx (n) .
By the hypotheses

DZ

and

DS ,

the two equations

`
and

and

RS

for

Rx (n)

and

I ,

this satises

r(0) x: A. p(x) =. q(x )

r(n) x: A. p(x) =. q Rx (n)

Therefore the induction scheme gives

r(n + 1) x: A. p(x) =. q Sx (n, Rx (n)) .

` n. r(n),

which by

n : N, x : A, p(x)
but this says that the morphism

RZ

R:NY

commute because it did for urterms.

is

q Rx (n) ,

is well dened. It makes the triangle and square

To show that it is unique we consider induction for the

predicate

r(n) x. p(x) ==. Rx0 (n) = Rx (n).

22