Models for a paraconsistent set theory

Thierry Libert

Extended abstract

We know from Russell’s paradox that the first-order axiomatization of Cantor’s

naive set theory is inconsistent in classical logic. More precisely, some peculiar
sets ‘{x | ϕ}’ provided by specific ϕ-instances of the comprehension scheme,
i.e. ‘∀x(x ∈ {x | ϕ} ↔ ϕ)’ , lead to triviality if the underlying logic is classical.
The most popular of those is so called Russell’s set : R = ‘{x | x ∈ / x}’.
Any classical solution to the paradoxes has to get rid of such ‘contradictory
sets’. It appeared besides that these virtual sets are by no means essential
for the foundations of mathematics, hopefully. However, certain logicians have
expressed a desire that such inconsistent objects might be handled and studied
within suitable theories, namely para(in)consistent ones. After all, non-well-
founded sets, as the universal set V = ‘{x | x = x}’, which are no more
indispensable for the foundations of mathematics, have subsequently found
interesting applications in computer science.1
There are many examples in mathematics where the introduction of ‘imagi-
nary/ideal’ objects, though giving some advantage to deal with them, has also
forced us to give up some basic properties or principles.2 Obviously, the price
to be paid here concerns the logic in which the theory is embedded, and its
possible debilitating effects on classical reasoning and mathematical practice.
One could therefore argue that this price seems to outweigh the advantages.3
Nevertheless, we are going to show in this talk that it is possible to construct
‘natural models’ for a paraconsistent set theory.
1
The interested reader should consult Peter Aczel’s book : “Non-well-founded sets”,CSLI
Lecture Notes Number 14, Stanford, 1988 (137 pages).
2
For instance, adding to the reals an ‘imaginary’ number i such that i2 = −1 leads to
sacrifice the structure of ordered field. As well as dealing with non well-founded sets prevents
us from inductive definitions. Etc.
3
Actually, it should be noted here that there exists some standard methodology producing
(para)inconsistent extensions of classical theories, so that nothing classical is lost.

1
What we call a model for a set theory is any ∈-structure satisfying some frag-
ment of the full-comprehension scheme, in a logic to be specified. All models
shall be considered and described within a pre-existing classical universe of
sets (e.g. a model of ZF , plus some reasonable large cardinal assumption if
necessary).4 Note that we will use below a small ‘∈’ to denote and distinguish
the membership relation in the metatheory from the big ‘∈’ in the language of
models for set theories.
The purpose of this extended abstract is to outline some technique common to
the construction of such models in classical and non-classical logics, showing
how these non-classical models can be compared with classical ones. In the
talk, we shall describe in detail and apply that technique to get models of
some paraconsistent set theory. So we will start here with the classical case,
introducing by the way a quite unusual but very natural response to Russell’s
paradox : positive set theory.

The classical case

From the semantical viewpoint, any classical ∈-structure M :≡ hM; ∈M i
can be equally thought of as a non-empty set M together with a function
[·]M : M −→ P(M ) , where P(M ) denotes the set of all subsets of M , by the
way of the following definition :

x ∈M y ⇔ x ∈ [y]M , for any x, y ∈ M .

Thus [y]M shall be called the extension of y in M. In this setting, M is ex-

tensional exactly when [·]M is injective. On the other hand, the semantical
translation of Russell’s paradox, namely Cantor’s theorem, tells us that [·]M
can never be surjective. More precisely, if [M ]M denotes the range of [·]M , then
some ∈-definable subsets of M, e.g. {x ∈ M | x ∈ /M x}, never belong to [M ]M .
Anyway, it is a fact that very interesting models can be constructed by pro-
viding M with some suitable structure and defining [·]M , in relation to that
structure, onto a set of distinguished subsets of M .
A well-known and very simple example is given by the set of natural numbers
N, with its arithmetical structure, where, for each n ∈ N, [n]N is defined to
be the set of those natural numbers which appear as exponent in the binary
expansion of n. Explicitly, [0]N = ∅, [1]N = {0}, . . . , [11]N = {0, 1, 3} , etc.
Thus, [·]N is one-one and [N]N is the set of all finite subsets of N. In this way,
it can be readily proved that hN; ∈N i is a model of Zermelo-Fraenkel set theory
without infinity.
4
At least, we assume that the reader is familiar with ZF .

2
In ZF , the informal argument which shows that Russell’s paradox is blocked
is that the well-founded part of the universe already fulfils the axioms of ZF .5
Indeed, in this cumulative hierarchy, Russell’s set would be nothing but the
universal class V , but this latter cannot be a set, otherwise V ∈ V . Incidentally,
it is worth observing that the universe of the model N described above is
actually well-founded, seeing that m ∈N n implies m < n by definition. As a
matter of fact, the universal class is clearly not a set in the sense of N because
[n]N is finite for any n ∈ N (that is precisely why the axiom of infinity fails to
be satisfied in N ).
There is another route to get around Russell’s paradox in classical logic, which
is based on a radically different idea : just restrict comprehension to those
positive formulae in which classical negation does not occur ‘explicitly’.6 Thus
the universal class V should be a set, and this is a singular departure from the
‘limitation of size’ doctrine of ZF and related set theories.
It was shown that very interesting models for so called positive set theories
can be defined as above by means of a topological space M and a homeo-
morphism [·]M from M onto the set of closed subsets of M , endowed with
the Vietoris topology. These topological models, called hyperuniverses, can
be constructed by projective limit, in a quite similar way as Scott’s models
for the lambda calculus.7 The properties of these models give rise to a very
interesting ‘topological’ set theory, which might be called P F (for Positive
Foundations) and which was axiomatized and deeply investigated by Olivier
Esser in [Ess1], where he showed for instance that P F interprets ZF .8 On
this point, it should be remarked that a pertinent formulation of the axiom of
infinity is actually required in P F in order to recover it in ZF (that is why a
large cardinal assumption is necessary for the construction of hyperuniverses
fulfilling this axiom). Incidentally, Esser also showed in [Ess2] that P F , with
that pertinent axiom of infinity, proves the negation of the axiom of choice,
whatever the formulation of this latter one adopts. As a similar result holds
for Quine’s set theory N F , this seems to be a characteristic feature of a set
theory with an universal set and infinity.9
5
Some authors include the axiom of foundation (regularity) in ZF . Notice that this
axiom is not an instance of the comprehension scheme (nor is the axiom of extensionality).
6
We are deliberately not very explicit on that point. For precise references on ‘positive
comprehension’, the reader should consult [For;Hin] or [Hin;Lib], where a brief historical
account of the subject is given.
7
Hyperuniverses are described in detail in [For;Hon].
8 +
Note that Esser called his theory ‘GP K∞ ’ . Needless to say why we think P F would
have been luckier.
9
A comprehensive bibliography on set theories with a universal set can be found at the
following address : http://math.boisestate.edu/˜holmes/setbiblio.html.

3
In any positive set theory, beside the universal class V , the ‘complement’ of
Russell’s class, that is Rc = ‘{x | x ∈ x}’, should be a set as well. So the next
step is to deal with Russell’s set itself.
In terms of boolean operations, it is rather clear that this further step has to
tamper with our conception of ‘complement’, for if R was a set in some model
M, we would have to choose between R ∈ [R]M ∩ [Rc ]M and R ∈/ [R]M ∪ [Rc ]M .
In other words, from the logical viewpoint, ∈ and ∈ / might be conceived as
weak negation of each other, and the underlying logic of the model might be
paraconsistent or paracomplete.
We shall only focus on the paraconsistent case here, showing how such a weak
negation can be incorporated into positive comprehension.

The paraconsistent case

One might intuitively think of a paraconsistent set S as an ordered pair of
covering parts of the universe : the first part collecting those objects which are
supposed to belong to S, the second gathering those which are supposed not
to belong to S, and where it is now agreed that these two parts might have a
non-empty intersection.
Such a ‘membership ambiguity’ can be easily concocted in the following way.
Consider an universe U which consists of a collection U of objects together
with a topology on it, which might materialize some notion of ‘indiscernibility’
on U or something like that. Then, for x ∈ U and S ⊆ U , we define
½
x ∈U S ⇔ x ∈ S
(where (·) is the closure operator on U ).
x∈/U S ⇔ x ∈ U \S

Note that x ∈ S ⇒ x ∈U S, as well as x ∈/ S ⇒ x ∈/U S. Also, x ∈U U \S ⇔ x ∈/U S.

But it is now possible that both x ∈U S and x ∈ /U S, for some x ∈ U and S ⊆ U ,
and therefore some kind of relative inconsistency may be observed. It is also
clear that the consistent subsets of U are thus exactly the clopen subsets in U.
Of course, one cannot define a model for a (paraconsistent) set theory in this
way, simply because the binary relations ∈U and ∈ /U are subsets of U ×P(U ),
not of U×U . However, the underlying idea is tied up with so called ‘topological
models’ we are going to present.
For any set M , let Pp (M ) denote the set of all ordered pairs (A, B) of subsets
of M such that A ∪ B = M .10 Then, any paraconsistent ∈-structure can be
conceived as a non-empty set M together with a function [·]M : M −→ Pp (M ).
10
In the paracomplete case, this latter condition should be replaced by A ∩ B = ∅.

4
Since Pp (M ) ⊂ P(M )×P(M ), it is convenient here to set [·]M = (d·eM , b·cM )
and thus define, for any x, y ∈ M ,
½
x ∈M y ⇔ x ∈dyeM
x∈ /M y ⇔ x ∈bycM
Then we shall call dyeM the extension of y in M, and bycM its anti-extension.
Naturally, M is said to be extensional if [·]M is injective. Here again, Cantor’s
theorem keeps [·]M from being surjective.
Nevertheless, it is possible to construct by projective limit a suitable topolog-
ical space M together with a bijective function [·]M from M onto the set of all
ordered pairs of covering closed subsets of M . Thus, in such a structure M,
any covering pair of closed subsets of the universe M really becomes identified
with an M-set. To express a straightforward consequence of this, let us adopt
the following definition : ½
dxeM ⊆ dyeM
for any x, y ∈ M , x is said to be less inconsistent than y if .
bxcM ⊆ bycM
It follows that
¯
¯ for each covering pair (A , B) of subsets of M , there exists a minimal
¯
¯ inconsistent M-set y such that A ⊆ dye and B ⊆ byc .
M M

Indeed, it is nothing but the (unique) M-set y such that [y]M = (A, B).
Let us just mention here that the M-sets define a ‘paraconsistent boolean al-
gebra’ (in the sense of [daCos;Bue]), in which the classical part coincides with
the boolean algebra of clopen subsets of M . Furthermore, it can be shown that
the class of hereditarily classical, well-founded M-sets yields a model of ZF ...
In the talk, we shall sketch the construction of such structures and show that
these are extensional models of a ‘maximal’ fragment of the comprehension
scheme in a ‘natural’ three-valued logic.11 Incidentally, we will point out an-
other way of formalizing set theory, namely by using set abstracts in the lan-
guage, as in [Bra;Rou]. To compare this latter with our approach, we will
show the incompatibility of extensionality and set abstracts with equality in
the language. Finally, we shall also introduce and discuss the paraconsistent
counterpart of the theory P F .

References

[Bra;Rou] R.T. Brady and R. Routley, “The non-triviality of extensional

dialectical set theory”, in Paraconsistent Logic (ed. by G. Priest, R. Routley
11
Details can be found in [Hin] & [Lib].

5
and J. Norman), Philosophia Verlag, Munich, 1989, pp.415-436. (Presented:
AAL (Monash), 1979).

[daCos;Bue] N.C.A. da Costa and O.A.S. Bueno, “Paraconsistency : towards

a tentative interpretation”, Theoria, vol 16, 40 (2001), 119-145.

[Ess1] O. Esser, “Interprétations mutuelles entre une théorie positive des

ensembles et une extension de la théorie de Kelley-Morse”, Ph.D. thesis,
Université Libre de Bruxelles (1997), unpublished.
(available at http://homepages.ulb.ac.be/~oesser)

[Ess2] O. Esser, “Inconsistency of the axiom of choice with the positive theory
+
GP K∞ ”, The Journal of Symbolic Logic, 65 (2000), 1911-1916.

[For;Hin] M. Forti and R. Hinnion, “The consistency problem for positive

comprehension principles”, The Journal of Symbolic Logic, 54 (1989), 1401-
1418.

[For;Hon] M. Forti and F. Honsell, “A general construction of hyperuni-

verses”, Quaderni dell’Istitudo di Matematiche Applicate “U. Dini”, Facolta
di Ingegneria - Universita di Pisa (1992).

[Hin] R. Hinnion, “Naive set theory with extensionality in partial logic and in
paradoxical logic”, Notre Dame Journal of Formal Logic, 35 (1994), 15-40.

[Hin;Lib] R. Hinnion and Th. Libert, “Positive abstraction and extensional-

ity”, forthcoming.

[Lib] Th. Libert, “ZF and the axiom of choice in some paraconsistent set
theories”, Logique et Analyse (submitted).

Thierry Libert (tlibert@ulb.ac.be)

Service de Logique Mathématique
Université Libre de Bruxelles
CP211, Boulevard du Triomphe
1050 Brussels
Belgium.