Sei sulla pagina 1di 10

COLIN McLARTY

ANTI-FOUNDATION AND SELF-REFERENCE'

ABSTRACT. This note argues against'Barwise and Etchemendy's claim that their semantics for self~reference requires use of Aczel's anti-foundational set theory, AFA, and that any alternative "would involve us in complexities of considerable magnitude, ones irrelevant to the task at hand" (The Liar, p. 35). Switching from ZF to AFA neither adds nor precludes any isomorphism types of sets. So it makes no difference to ordinary mathematics. I argue against the author's claim that a certain kind of 'naturalness' nevertheless makes AFA preferable to ZF for their purposes. I cast their semantics in a natural, isomorphism invariant form with self-reference as a fixed point property for propositional operators. Independent of the particulars of any set theory, this form is somewhat simpler than theirs and easier to adapt to other theories of self-reference.

Barwise and Etchemedy give an elegant semantics for self-reference in The Liar 1 They also claim this semantics requires a new set theory, Aczel's anti-foundational set theory, AF A, and that any alternative "would involve us in complexities of considerable magnitude, ones irrelevant to the task at hand" (p. 35). Gupta has agreed, saying the main obstacle to semantics of self-reference has been "the difficulty of constructing, within the confines of standard mathematics (that is, set theory), a natural account of circular propositions" (parenthesis is Gupta's).' This note will show the foundations of set theory have nothing to do with it. I do not object to AF A. In particular I do not say 2ormeloFraenkel set theory, ZF, has any "clearer and more compelling" a motivation than Aczel has given for AFA' To me, that claim merely expresses familiarity with ZF. I do say the difference between AFA and ZF makes no difference where normal working methods in mathematics are concerned, and those methods suffice as well for Barwise and Etchemendy's semantics. First I point out that extending a ZF universe of sets to an AFA universe adds no new isomorphism types of sets, so it adds no structures at all np to isomorphism.' I briefly argue on general grounds

1993

Journal of Philosophical Logic 22: 19-28, 1993. Kluwer Academic,Publishers. Printed in the Netherlands.

20

COLIN McLARTY

against Barwise and Etchemendy's claim that a certain consideration of 'naturalness' in semantics nevertheless makes AFA preferable to ZF for their purposes. Then I give a natural form of Barwise and Etchemendy's semantics which can be defined in any set theory. A reader who only wants to see the semantics can go directly to Section 3.
1. AFA AND ISOMORPHISM

The theory AFA includes all the ZFaxioms except well-founding so the well-founded sets of any model of AFA form a model of ZF, the familiar inner model of the well-founding axiom. Every ordinal is well-founded. Since AFA includes the axiom of choice every AFA set is isomorphic to an ordinal and thus to a set in the model of ZF. So every structure existing in an AFA universe is isomorphic to one in a ZF universe. Conversely, Aczel's relative consistency proof shows every ZF model embeds into an AFA model so every structure in a ZF model exists in an AFA model (see The Liar, Ch. 3). Passing from ZF to AFA neither adds nor precludes any isomorphism types of sets or of other structures, and so neither adds nor removes any theorems stated in isomorphism invariant form. This is not the sense in which the authors admit "AFA does not give rise to any new mathematical structures" (p. 47). They mean any statement about AFA sets can be interpreted as a statement about equivalence classes of graphs in ZF. Via this interpretation the theorems on sets in AFA appear as theorems on equivalence classes of graphs in ZF. The authors justly compare this to interpreting statements about the complex numbers as statements about pairs of real numbers: It obscures the natural sense of the structure (AFA sets, or complex numbers, as the case may be). If that were the only sense in which AFA gave no new structures I would agree with them on using AFA. My point is that so long as you keep to isomorphism invariant statements, as is usual in mathematics, the theorems and proofs of AFA already are theorems and proofs of ZF and vice versa.
2. 'NATURALNESS' IN SEMANTICS

Barwise and Etchemendy know every non-well-founded set is isomorphic to a well-founded One but they claim something more than

ANTI-FOUNDATION AND SELF-REFERENCE

21

description up to isomorphism is needed for semantics. They say "By far the most natural way to model a proposition about a given object is to use some set-theoretic construct containing that object (or its representative) as a constituent, that is, where the object appears in the construct's hereditary membership relation" (p. 34). So they require that the relation "x is an object that the proposition y is about" must be modeled not by just any set theoretic relation with the right structure, but specifically by some fragment of the hereditary membership relation. Of course this is not isomorphism invariant. Isomorphisms of sets do not preserve membership. And ZF can not model a self-referential proposition 'naturally' in this sense since no well-founded set is in its own hereditary membership relation, i.e., in
its own transitive closure.

The general set-theoretic point of this note is to show that the author's concern with membership is an unnecessary complication. If such concern with the specific members of sets was ever current in mathematics, as perhaps under logicist influence, it was chased out by Bourbaki. Bourbaki realized that what matters in mathematics is the structural relations between sets, not the particulars of their members, and so he describes structures axiomatically and thus up to isomorphism without specifying the members. Axiomatic descriptions need consistency proofs, and for that Bourbaki uses models in a set theory rather like ZF and says, in principle, exactly what are the elements of each set in the model. But the particular model is not only glossed over after that, it is entirely irrelevant to proofs from the axioms, and there is no use worrying whether it happened to interpret some given relation in the axioms by membership. Of course the authors know the difference between a theory and a model and they prefer to work with a model. Barwise explains his preference saying "set-theoretic modelling has proven so extraordinarily powerful a method in logic".' But in ordinary practice the powerful uses of set theoretic models are not ones where you take a single model as the object of study, much less ones where you worry about the foundational details of constructing the model. They are ones where you construct and compare various models in respect
of their model-theoretic"structure, which is invariant under isomor-

phism of models. From this point of view we want not one model of

22

COLIN McLARTY

Russellian (or Austinian) propositions but a general theory of models. The account of algebras of prepropositions below is such a theory, with Barwise-Etchemendy algebras as the isomorphism invariant description of the one model the authors give. I will point out that Lawvere furthered Bourbakiste struCturalism by axiomatizing set theory itself so that sets are, quite rigorously, described only up to isomorphism. There is no membership relation between sets in his elementary theory of the category of sets, ETCS 6 I thought in terms of ETCS in writing this, but I will not spell it out or argue for it. The point is that the semantics described here can be formalized in the routine way in ZF, AFA, ETCS or any reasonable set theory.
3. SEMANTICS OF SELF-REFERENCE

I treat Barwise and Etchemendy's Russellian theory of propositions but the technique applies as well to Austinian propositions. I assume familiarity with The Liar. Our version of the semantics makes self-reference a fixed point property. For example, we define [FaJ as a function taking propositions to propositions. Then a liar proposition is a fixed point for [FaJ, a solution to p = [Fap]. Barwise and Etchemendy's concern that self-reference leads to set-theoretic problems is completely implausible when you look at it this way. No one expects a real number x to be in the transitive closure of its sine, sin (x). No one sees a set theoretic problem in the equation 0 = sin (0). So we do not expect p to be in the transitive closure of [Fap] nor see any set theoretic problem in self-referential propositions. Define an algebra ofprepropositions to be a class PrePROP with:
I. Elements [aH c] and [aH c] where a is Claire or Max and cis

a card. 2. A function LBelJ from {Claire, Max} x PrePROP to PrePROP and a similar function LBel J 3. Functions [Tr J and [Tr J both from PrePROP to PrePROP. We use [FaJ to abbreviate [Tr J. 4. Functions [/\ Xl and [v X] taking subsets X of PrePROP to members of PrePROP.

ANTI-FOUNDATION AND SELF-REFERENCE

23

The definition so far is not equivalent to Barwise and Etchemendy's. We can build an algebra of prepropositions by starting with the prepropositions of clause I and inductively applying clauses 2-4, adding new members to PrePROP for each new application of a function to arguments. Barwise and Etchemendy would say we had "built PrePROP up from below, by a standard inductive characterization" (p. 63). We can do this in any reasonable set theory, such as ZF or AFA or ETCS, to get an algebra with no circular reference. In this algebra a preproposition can only refer to prepropositions formed before it was. Call this the free algebra. We could take the free algebra and add a new member fto PrePROP, specify [Fail = fthen build up a new algebra by inductively applying clauses 2-4. This algebra has the one liar f and prepropositions built from it but no other circular reference among prepropositions. Another variant takes the free algebra and adds two new members fandf', specifying [Fail = fand [Faf'] = f'. Then build up a new PrePROP by inductively ilPplying clauses 2-4. This gives an algebra with two distinct liars, f andf'. There can not be distinct liars on Barwise and Etchemendy's approach. Their Theorem;j (p. 72) proves there is a unique solution p to the equation p = [Fap]. Someone might claim the Russellian account in The Liar is basically cOITect, except that each new utterance of "This proposition is false" states a new proposition: My indexical 'this' can not refer to the same proposition as uttered by any of innumerable other people on occasions unknown to me. We can formalize that position as a step in considering the argument, using semantics in an algebra of prepropositions with infinitely many distinct liars. For Barwise and Etchemendy to formalize the position in a 'natural' way by their standard, they would first need to invent another set theory.' We expand on this last claim since it is central to the author's "most natural way to model a proposition" and to their argument against using ZF for their semantics. There are many ways to code propositions, as Barwise and Etchemendy stress several times. Any number of sets p could solve the equation p = [Fap] depending on the coding. But once a coding is chosen we have a unique equation p = [Fap]. The author's argument against ZF is that no 'natural'

24

COLIN McLARTY

coding gives this equation any solutions in ZF. To get one solution with a 'natural' coding in their sense they use AFA, but then there is provably only one. The proof of Barwise and Etchemendy's Theorem 3 (p. 70) would be unaltered by supposing a proposition should
include some "representation of its context or its speaker or whatever.

As long as "This proposition is false." translates into a specific set theoretic equation of the kind used in the author's solution lemma it has a unique solution in AFA. If you want more than one solution you must use some other set theory - or abandon the authors's allegedly natural link between the representation of propositions and set theoretic membership. We get Barwise and Etchemendy's particular semantics in our form by applying the idea of anti-foundation directly to prepropositions. Define the immediate constituents of a preproposition as follows. The immediate constituents of[aHc] are: a, H, and c. Those of[aHc] are: Not, a, H, and c. Those of[aBelp] are:.a, Bel, and p. For [aBelp] add Not. The immediate constituents of [Tr p] and its negation are obvious. For [A X] they are A and the prepropositions in X. Similarly for [v X]. A preproposition in this simple language is uniquely deterntined by its immediate constituents. Define an isomorphism of algebras as an isomorphism between the classes which preserves the distinguished elements and the immediate constituent relation. Define a tagged graph as in The Liar (p. 39) and define a prepropositionally tagged graph as a tagged graph in which every node is of one of these kinds:
I. A card node is a node with three children, each childless and

tagged by: one of Claire or Max, H, and a card. A negated card node is the same but with the addition of a childless child tagged by Not. 2. A belief node is a node with three children, one with children of its own, one tagged by Bel and one tagged by Claire or Max. A negated belief node has an additional childless child tagged by Not. 3. A truth node is a node with one child with children and one childless child tagged by Tr. Negated truth nodes are obvious.

ANTI-FOUNDATION AND SELF-REFERENCE

25

4. A conjunction node has any set of children with children, and one childless child tagged by /\. For a disjunction node use v instead of /\. A decoration in an algebra PrePROP for a tagged graph is a function D defined on the nodes of the graph such that for each node x with no children D(x) = tag (x), whereas if x has children then D(x) is the element of PrePROP whose immediate constituents are all and only the D(y) for y a child of x. An algebra of prepropositions PrePROP is called a BarwiseEtchernendy algebra, or B-E algebra, if and only if: Every prepropositionally tagged graph can be decorated in PrePROP in exactly one way, and every member of PrePROP occurs in SOme decoration. It is easy to see the class of prepropositions defined by Barwise and Etchemendy, with the obvious definition of functions LBelJ and [Tr J and so on, is a B-E algebra in AFA. Take the reasoning Aczel used to show AFA is consistent if ZF is. (See The Liar, Ch. 3.) A simplification of this proves the B-E algebra axioms are consistent if ZF is. We take Claire, Max, H, the fifty two cards, Bel, Not, Tr, /\, and v as urelements. Wherever he says a graph represents a set we say it represents a preproposition. Where he says one AFA set is a member of another we say one preproposition is an immediate constituent of another. We look not at all tagged graphs but only at prepropositionally tagged ones, and we verify not the AFA axioms but the B-E algebra axioms. This simplifies a good bit as the B-E algebra axioms are much simpler than those for AF A. This proof, like Aczel's relative consistency proof for AFA, can be given in any set theory strong enough to describe the same graphs as ZF. The proof of the solution lemma for AFA sets has an analogue for prepropositions in a B-E algebra. Given any prepropositional algebra PrePROP and a set of indeterminates X define PrePROP[X] by extending clause I in the definition of a prepropositional algebra to include members of X along with the elements [aH c] and [a He], and apply the other clauses inductively on this new base. In effect PrePROP[X] consists of PrePROP plus expressions like those in PrePROP but with indeterminates in place of some prepropositions.

I'

26

COLIN McLARTY

Define a prepropositional equation in X to be an expression of the form


x

= a

where x E X and a E PrePROP[X]. The analogy to The Liar, p. 49, should be clear. Our solution lemma says that if PrePROP is a B-E algebra then for every set of indetenninates X every set of prepropositional equations in X has a unique solution. (The reader may prove the converse as an exercise: If every such family has a unique solution then PrePROP is a B-E algebra.) The plausibility argument that Barwise and Etchemendy give for their solution lemma works just as well for ours. The only difference is that we have several ways to form prepropositions out of others while they have only set formation. Each equation in a family gives a graph, and the graphs are then to be hooked up with one another. Aczel's formal proof of the lemma for AFA sets also yields a proof for prepropositions just by substituting prepropositions for sets, immediate constituents for members, and prepropositionally tagged graphs for all tagged graphs. The same change of wording turns any proof that there is a unique class meeting Barwise and Etchemendy's definition of PrePROP into a proof that all B-E algebras are isomorphic. We follow the authors in saying the result "follows easily from the general considerations discussed in Chapter 3", (p. 62). Since we apply those considerations to the constituent relation in any B-E algebra they are sound in any set theory, while the authors apply them only to the membership relation in AFA. To prove a B-E algebra contains a particular prepropositioIJ, such as a uniquefsuch thatf = [Fafl, we can use decorations of tagged graphs. Specifically, f decorates the one untagged node in the tagged graph consisting of: one childless node tagged by Not, one childless tagged by Tr and one more node with each of those and itself as children. This parallels the author's use of tagged graphs to show particular sets exist in AFA (p. 4Iff). Or we could use the solution lemma, interpreted as applying to prepropositions in a B-E algebra as above, the same way that they do.

ANTI-FOUNDATION AND SELF-REFERENCE

27

All the rest of Chapters 4-7 of The Liar can be stated for any B-E algebra and all the results proved in essentially the same way that Barwise and Etchemendy do. The only difference is that we apply their theorems and proofs not to membership in AFA but to the constituent relation in any B-E algebra. I stress that the proofs for B-E algebras differ from those for AFA very little and are often simpler. We do not take Barwise and Etchemendy's proofs and eliminate non-well-founded sets by substituting lengthy considerations of equivalence classes of decorated graphs. We do not eliminate non-well-founded sets at all. We simply bypass issues of set membership by dropping the demand that representatives of constituents of a preproposition be in the transitive closure of the representative of the proposition. The constituent relation between prepropositions in a B-E algebra, which is after all the real point of interest for the semantics, is not well-founded no matter what set theory we use.
4. CONCLUSION

The point is not to oppose variant set theories. It is that immediate constituents of propositions are not especially similar to members of sets. Even if a semantics and a set theory can be paired so the constituent relation and the membership relation have similar formal properties, as Barwise and Etchemendy have done, the formal parallel adds nothing to the semantics. Barwise and Etchemendy themselves insist most details of the representation of propositions are irrelevant:
We use [a He] for that set theoretic object which represents the proposition that a has C, . . . Exactly how these are represented by sets doesn't really matter, and burdening our definitions with the inessential detail would just obscure matters. We will assume as a general feature of our coding, though, that the objects referred to in naming the set theoretic representative are members of its transitive closure. Thus we assume that a is in the transitive closure of [a He]. (p. 62)

Their coding is defined up to 'isomorphism plus a condition on transitive closures', which latter condition serves no purpose but to preclude solving p = [Fap] and other equations for circular reference in ZF' It is a mere distraction. The authors show the graph-theoretic

28

COLIN McLARTY

ideas Aczel used to motivate AFA are valuable in analyzing circular reference. These graphs may serve other ends besides, but so do lots of structures and we would do as well not to worry about enshrining them all in the membership relations of set theories. B-E algebras apply the graph theory directly to semantics.
NOTES

* I thank John Mayberry for discussions of axiomatics which inspired this paper. Jon
Barwise's criticism of an earlier draft improved this one as did an anonymous referee. I J. Barwise and J. Etchemendy, The Liar: An Essay on Truth and Circularity (New York: Oxford University Press, 1987). Hereafter page numbers after quotes of Barwise

and Etchemendy refer to this book.


2

"Jon Barwise and John Etchemendy's The Liar", Philosophy of Science 56 (1989): 697-709. The quote is on p. 697. 3 L. Moss takes this position in his review of The Liar, in Bull. Amer. Math. Soc. 20, no. 2 (1989): 216-225. The quote is from p. 222. 4 When I say two sets are isomorphic I mean there is an invertible function between them, that is a one-one unto function. When I say two structures are isomorphic I mean there is an invertible functio'n between their underlying sets, preserving the operations and relations of the structure, whose inverse also preserves them. 5 J. Barwise, The Situation in Logic, CSLI Lecture Notes no. 17 (Stanford: CSLI. 1989) p. 188. 6 See F. W. Lawvere "An Elementary Theory of the Category of Sets", Proc. Nat. Acad. Sci. USA. 52 (1964): 1506-1511 or W. S. Hatcher, The Logical Foundations of Mathematics (New York: Pergamon Press, 1982). 7 In fact Aczel Non-Well-Founded Sets, CSLI Lecture Notes No. 14 (Stanford: CSLI. 1988), gives a number of variants on AFA. Each has a corresponding 'natural' variant account of self-reference in Barwise and Etchemendy's sense. You could look to see whether one of these variants is philosophically useful but that would be the hard way to do semantics. All these variants are algebras of prepropositions and we have a much simpler and more flexible classification theorem for such algebras than we ever will for the set theories. Algebras of prepropositions are models of an infinitely presented infinitary algebraic theory (treat 1\ and v as classes of finitary and infinitary functions), so they are precisely the quotients of free algebras on some number of generators. 8 Gupta misses the importance of the clause on transitive closures when he seeks to agree with the authors by saying all that matters in a model is "the kind of information coded by the modeling" (op. cit., p. 699, n. 4). B-E algebras are only one of many codings for the same information which violate the condition on transitive closures and defeat the argument against ZF.

Department of Philosophy, Case Western Reserve University, Cleveland, OH 44106, U.S.A.

Potrebbero piacerti anche