Mind Association

The Structure of the Paradoxes of Self-Reference Author(s): Graham Priest Source: Mind, New Series, Vol. 103, No. 409 (Jan., 1994), pp. 25-34 Published by: Oxford University Press on behalf of the Mind Association Stable URL: http://www.jstor.org/stable/2253956 . Accessed: 31/07/2011 18:57
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at . http://www.jstor.org/action/showPublisher?publisherCode=oup. . Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Oxford University Press and Mind Association are collaborating with JSTOR to digitize, preserve and extend access to Mind.

http://www.jstor.org

The Structure of the Paradoxes of Self-Reference

GRAHAM PRIEST

Introduction
When one meets the paradoxesof self-referencefor the first time, one is struck by the fact that they all appearto be membersof a single family, generatedby a common underlying principle. Indeed, Russell, who inauguratedthe study of these paradoxesthis century,held thatthis is precisely what they are. It is notoriously difficultto pin down the commonunderlyingprinciple,however.Forexample, on closer scrutiny it is not at all clear what the Burali-Forti and Liar Paradoxeshave in common. Russell himself was unableto say whatheld the famremarksconcerning ily of paradoxestogetherbeyond some ratherunsatisfactory vicious circles in Principia Mathematica.(See for example Godel 1944.) It is therefore unsurprisingthat the modern view of the paradoxes is to the effect thattherearetwo distinctfamilies here, which arise from differentsources, and which are to be treatedquite differently.(The view is now so orthodoxthat this claim needs no documentation.)Although one can find something like this view expressed in Peano (1906), the founder of the orthodoxy was Ramsey (1925). thatRussell was right and RamThe only point of this paperis to demonstrate sey was wrong. The paradoxesof self-referencedo have a common underlying involved, andI shall spell out exactly which generatesthe contradiction structure, what that is. Moreover,as we shall see, Russell had been on the righttracksome five years earlier.

1. Ramsey'sdivision
ThoughRamsey's division of the family into two classes is wrong, it will be useful to have it before us. His exact words are not withouthistoricalinterest,so let me simply quote him: It is not sufficiently remarked,and the fact is entirelyneglected in Principia Mathematica,that these contradictions[the paradoxesof self-reference] fall into two fundamentaldistinct groups, which we will call A and B. The best known are divided as follows: A. (1) The class of all classes which are not membersof themselves.
Mind, Vol. 103 . 409. January 1994 ? Oxford University Press 1994

26

GrahamPriest

(2) The relationbetween two relationswhen one does not have itself to the other. of (3) BuraliForti's contradiction the greatestordinal. B. (4) "I am lying." (5) The least integernot nameablein fewer thannineteensyllables. (6) The least indefinableordinal. (7) Richards'scontradiction. (8) Weyl's contradictionabout "heterologische". The principleaccordingto which I have divided them is of fundamental importance.GroupA consists of contradictionswhich, were no provision made against them, would occur in a logical or mathematicalsystem itself. They involve only logical or mathematical terms such as class and number,and show that there must be something wrong with our logic or mathematics. But the contradictions in Group B are not purely logical, and cannot be stated in logical terms alone; for they all contain some referenceto thought,language, or symbolism, which are not formal but empiricalterms. So they may be due not to faulty logic or mathematics,but to faulty ideas concerningthoughtand language.If so, they would not be relevantto mathematicallogic, if by "logic" we mean a symbolic system, though of course they would be relevant to logic in the sense of the analysisto thought. [Footnote:these two meanings of "logic" are frequentlyconfused...](1925, p. 171 of reprint) In hindsight,the unsatisfactorynatureof Ramsey's criterionis clear.For a start, if one wants to draw a fundamentaldistinction,this ought to be done in terms of the structure of the differentparadoxes.Ramsey's distinctiondependson the relatively superficialfact of what vocabularyis used in the paradoxes,and, in particular,whetherthis belongs to mathematicsproperlyso called. But worse, this is a notoriouslyshifting boundary.Ramsey was, of course, writingbefore the heyHad he been writingten years later,it would have been day of meta-mathematics. clear thata numberof items of vocabularyoccurringin paradoxesof GroupB do both syntactical and semanticallinguistic belong to mathematics.In particular, notions became quite integralpartsof mathematics.Indeed, in a sense, the work of Godel andTarskishowed how these notions could be reducedto otherpartsof mathematics(numbertheory and set theory,respectively). Despite this, it behoves anyone who says thatRamsey is wrong to come good with the structure underlyingparadoxesof self-reference.This we will turnto in a moment. But first it will also be convenient to divide Ramsey's GroupB itself into two. In the firstclass, Bi, belong Ramsey's (5), (6) and (7); in the second, Bii, belong, (4) and (8). A rough criterionfor-thedistinctionis that paradoxesin Bi, but not B3ii, make explicit use of the notion of definability.More fundamentally, paradoxes in Bii use diagonalisation,or some similar mechanism, to establish somethingof the form a - a, and hence ct and -a; whereas paradoxesin Bi give independentargumentsfor each of a and -a.

The Structureof the Paradoxes of Self-Reference 27

2. Russell's Schema
To exhibit the structureunderlyingall these paradoxes,let uAs start with what I shall call Russell's Schema, an idea dating from some years before he lighted upon the Vicious Circle Principle.In his own words this is as follows: Given a propertyq and a function 6, such that, if p belongs to all members of u, 6(u) always exists, has the propertyq, and is not a memberof u; thenthe suppositionthatthereis a class w of all termshavingproperty and qp that 6(w) exists leads to the conclusion that 6(x) both has and has not the property p. (1905, p. 142 of reprint) Less tersely, given a propertyq9,and function 6, consider the following conditions:
1) w = {x: qg(x)} exists

2) if x is a subset of w: a) 6(x)ox
and b) 6(x)Ew

For future reference I will call clause (2a) the Transcendence Condition and
clause (2b) the Closure Condition.

Given (1) and (2) we have a contradiction.For when Transcendence Cloand sure are appliedto w, an irresistibleforce meets an immovableobject. The result: b(w)Ewand 6(w)Ew.Hence, any qp 6 thatinstantiateRussell's Schema generand ate a contradiction. Russell did not claim thatthis structure behindall the parwas adoxes. He did, however, demonstrate that it was behind all paradoxes of Ramsey's GroupA. To see how, it is sufficientto examine a couple of examples. In Russell's Paradox,the propertyq(x) is "x?x",so that w is the Russell set R={y; yoy}; and the function 6 is simply the identity function, id. Suppose that it Transcendence; follows thatx xcw; then x ex--xcx. Hence x4x, demonstrating E R, which is Closure.The contradiction thatRERand ROR. is Russell's Paradoxis a stripped-down versionof Cantor'sParadox(see Russell, 1905), so let us move on to Burali-Forti's.q(x) is "x is an ordinal",and so w is the set of all ordinals,On; 6(x) is the least ordinalgreaterthan every memberof x (abbreviation: holds by definition,as does Closure.The log(x)). Transcendence contradictionis that log(On)EOnand log(On)0On. One further example: Mirimanoff's Paradox. In this, qp(x)is "x is well founded",so thatw is the cumulativehierarchy, 6(x) is just the power set of x, V; and Closure hold simply in virtue of x being wellP(x). If xcV, Transcendence in since 6(V) is just V founded.The contradiction this case is thatVeV and VWV, itself. (Since all membersof V are well-foundedP(V)cV. But if xEV then every memberof x is in V. Hence VcP(V).)

28

GrahamPriest

We can tabulatethese observationsas follows: Paradox Russell Burali-Forti Mirimanoff | id(x) log(x) P(x) 3(x) x?x x is an ordinal x is well-founded
9(x)

w R On V

3. Definabilityparadoxes
Let us now move on to the paradoxesin Ramsey's GroupBi, the definabilityparadoxes, and recall that somethingis definableiff there is a (non-indexical)nounphrasethat refersto it. Now consider Konig's Paradoxas an example: there are indefinableordinals,andthe least indefinableordinalhasjust been definedin that very phrase.At first glance, this would appearto fit into Russell's Schema very nicely. q(x) is the predicate"x is a definableordinal",so thatw is the set of definable ordinals,DOn; and ((x) is lon(x) (the least ordinalnot in x). The contradiction is thatthe least indefinableordinallon(DOn)both is and is not definable(that is, a memberof DOn). Unfortunately,things are not so simple. For whilst Transcendence is true, lon(x)ox, Closureneed not be: even if every memberof x is definable,there is no reason to suppose that x itself is definable.And if it is not, there is no reason to suppose thatthe least ordinalnot in x is definable,i.e., that lon(x )e DOn. If, however,x is definable,then we do have thatlon(x)EDOn;andthis suggests a small modificationof Russell's conditions, as follows: For given propertiesw and V, and (possibly partial)function 8: 1) w = {x: q(x)} exists and ip(w) 2) if x is a subset of w such that 4p(x): a) 6(x) ox and b) ((x)Ew If these conditions are satisfied we still have a contradiction.For since 4(w), we have both 6(w)ow and ((w)Ew.I will call this the QualifiedRussell'sSchema.The essential qualificationis that applying ( to an arbitrarysubset of w is not now guaranteedto transcendit; only if the set satisfies t will it do so. Thus ( works on a sub-familyof the power set of w, P(w). Bearing this in mind, it is clear that Russell's Schema properis just a special case of the QualifiedRussell Schema, where the sub-family in question is P(w) itself (i.e., where 4 is the universalproperty, =xx=x). Thus, the QualifiedRussell Schema still encompasses the paradoxesof GroupA. But by choosing 4 approto priately,it also encompasses Konig's Paradox.We just take the property4p(x)

The Structureof the Paradoxes of Self-Reference 29

be "x is definable".DOn is clearly definable, and as we have alreadyobserved, and for definablex, Transcendence Closurehold. all Unsurprisingly, the other paradoxesin GroupBi fit the QualifiedRussell is Schema. Take Berry's Paradox. qp(x) "x is a naturalnumber(= finite ordinal) definable in less than 19 words", so w={n; n is a naturalnumber definable in less than 19 words}, DN19;i(x) is "x is definable in less than 14 words";and 6(x) is just lon(x). Clearly,lon(x)ox, and if x is definable in less than 14 words (Its lon(x)EDN19. definitionis "the least ordinalnot in" plus the definitionof x.) Hence Transcendence and Closure are satisfied. The contradiction is that and lon(DN19)EDN19 lon(DN19)4DN19. One furtherexample:Richard'sParadox.In this, p(x)is "x is a definablereal numberbetween0 and 1",so thatw is the set of such reals, DR; ip(x)is "xis definable" and 6(x) is diag(x), a real, definedby diagonalisationon x, in such a way as to ensurethatb(x)0x.If x is definable,diag(x) is definable.Hence bothTranscendence and Closure are satisfied. The contradiction is that diag(DR)EDR and diag(DR)(DR. We can tabulatethese observationsas follows: Paradox Konig's Berry's
6x

(x) x is definable x is definable in less than 14 words x is definable

9x

w DOn DN19

lon(x) lon(x)

x is a definable ordinal x is a natural numberdefinable in less than 19 words x is a definable real between 0 and 1

Richard's

diag(x)

DR

We have seen that all the paradoxesof GroupsA and Bi fit the QualifiedRussell Schema.In fact, they all fit the Russell Schemaitself. To see this it sufficesto note that (assuming the Axiom of Choice) for any structurethat fits the Qualified Schema there is another,identical in all relevant respects, that fits the regular Schema. Let q, V and 6 be any quantitiessatisfying the QualifiedRussell Schema. As usual, w={x: p(x)}.Let E)be a choice function on P(w). We now define a new function, 6', as follows. Let xcw. 6'(x) = 6(x) ifp(x); 6' (x) = 8 (w-x) otherwise It is easily seen that .pand 6' satisfy the Russell Schema. The only point to note is that 8 (w-x) is undefinedif w=x. But since 'p(w), 6'(x) is simply 6(x) in this case.

30

GrahamPriest

and For the same reason,the contradictions6(w)EwA6(w)qw b'(w)EwA6' (w)qw are identical.

4. The Liar, etc.

We have seen thatRussell's Schema properunderliesthe paradoxesin Ramsey's GroupsA and Bi. It remainsto show thatit underliesthe paradoxesin GroupBii. The Liar Paradoxis a paradigmof this Group.Primafacie this paradoxappears to have nothing to do with the Russell Schema: the Schema concerns totalities the which can be transcended means of the applicationof a suitableoperator; by Liar Paradoxdoes not seem to involve totalities at all. form. In fact, for However,with a bit of care, it can be put into the appropriate reasons that will become clear later, it is safer to show that it fits the Qualified Russell Schema, so let me show this. Assume for the momentthat it is sentences (ratherthan propositionsor some other semantic entities) that are true or false. Let q(x) be "x is true",so that w is the set of true sentences, Tr;let tp(x)be "x is definable".6 is a function, a, definedby some suitabletechniqueof diagonalisation so that if a is any definableset o(a)=a where a=<a4a>. The angle-bracket expression is a sentence expressing the fact that ctis not in the set a. (Hence, a says "This sentence is not in a".) Note that a must be definable, or there would be no guaranteethat there is such a sentence. Now, if a is definableandacTr:
o(a)Ea <a4a>ea
<a4a>eTr

aea
=^o(a)ea

(by the T-schema)

is Hence o(a)ea, andTranscendence satisfied.Moreover,it follows that a4a, and hence by the T-schema<aca>c=Tr, o(a)ETr.Hence, Closure is satisfied too. i.e. The Liar is the sentence c(Tr) and the contradiction is that c(Tr)ETr and o(Tr)4Tr. Anotherparadoxin this family is the KnowerParadox.This is the same as the Liar, except that q(x) is"'x is known to be true",and so w is the set of known things, Kn. Transcendenceis verified as before, since knowledge implies truth; Closure follows, since it has been establishedthat a4a, and so aEKn. The paradox is that o(Kn)EKnand o(Kn)eKn. A varianton these paradoxesis the liar (or knower) chain, where several sentences are inter-related. These fit the pattern,too. To see this, just consider the simplest paradoxof this kind: the Liar pair.This is exactly the same as the Liar, except that a is slightly more complicated.This time, o(a)=a, where a=<fleTr> and fl=<a4a>. To see that a satisfies the appropriateconditions, suppose that acTr and thata is definable.Then:

The Structureof the Paradoxes of Self-Reference 31

o(a)Ea a aca a aETr

<PETr>-eTr =i c-Tr
=

(by the T-schema)

(by the T-schema)

<aqa>eTr a a4a
r (a)qa

Hence, u(a)4a, which is Transcendence. see thatClosureis satisfied,note the To following:

u(a)4a a4a

<a4a>eTr
=ic3Tr

(by the T-schema) (by the T-schema)

<pf3Tr>ETr
aETr a OF (a)-Tr

The contradictionis that o(Tr)ETrand u(Tr)4Tr. A final example from this Groupof paradoxeswill suffice. This is the Heterological Paradox. p(x) is "-x sat x" where x is the satisfaction predicate, so w={y: -y sat y}, Het; t(x) is "x is definable",as before; if a is a definable set, b(a)=<vEa>, where v is any new variable,and so this is an open sentence. Now, suppose that acHet anda is definable.Then:
<vEa>ea

<vEa>elHet -<vEa> sat <vEa >

<vEa>ea (by the SatisfactionSchema) Hence <vEa>ea, i.e., Transcendence.So by the SatisfactionSchema -<vEa> sat <vEa>, i.e., <vEa>EHet, Closure. The paradox is that <vElHet>EHetand <vEHet>eHet. We can recordthese observationsin the following table:

Paradox Liar Knower Liar Chain

(x) a, where a=<a(4x> a, where

a=<a4x>

4(x) x is definable x is definable x is definable

q(x) x is true x is known x is true

w Tr Kn Tr

a, where a=<f3=E-Tr> where

<fl>=<a4x>

Heterological

<vex>

x is definable

-x sat x

Het

32

GrahamPriest

Finally, note that nothing is changed essentially if we take propositionsor some other semanticentities (not sentences) to be truthbearers;we just take <aex> to be the appropriate proposition.Similarly,in the HeterologicalParadox,we take <vex> to be the appropriateproperty.In fact, things are actually simplified by this. For presumably,for any set (not just a definableone), x, there is a proposition, a, such that a=<a0x> (the propositionthatthis propositionis not in x). Similarly, in the Heterologicalparadox,if x is any set there is a propertyof being in x. In this case, one no longer needs to assume that x is definable;and the paradoxes fit the regularRussell Schema directly.

5. ThePrinciple of UniformSolution
We have now seen that all the paradoxesof self-reference,whetherfrom Group A, Bi or Bii, fit the Russell Schema.The structure this describesis, therefore, that the structuire generatesall the paradoxes.Russell was right:there is a single that family here. What one should make of this demands a largerdiscussion than is possible here;but I will at least indicate what seems to me to be the single most importantconclusion. If two paradoxesare of differentkinds, it is reasonableto expect them to have differentkinds of solutions; on the other hand, if two paradoxesare of the same kind, then it is reasonableto expect them to have the same kind of solution. Generalising, it is naturalto expect all the paradoxesof a single family to have a single kind of solution. Any solution that can handle only some members of the family is boundto appearsomewhatone-eyed, andas not having got to gripswith the fundamentalissue. Of course, this observation puts a lot of weight on the notion of kind;to convince ourselves thattwo paradoxesare of the same kind we must convince ourselves (a) thatthere is a certainstructure that producescontradiction and (b) thatthis structure commonto the paradoxes.Still, once kindredis ship has been establishedin this way the point seems undeniable.Let us call it the Principleof Uniform Solution (PUS, sorry):same kind of paradox, same kindof solution. Post Ramsey, logicians have accepted that there are essentially two distinct families of paradoxand, with an intuitive appreciationof the PUS, have, generally speaking,been content with two differentkinds of solution. Typically,solutions to paradoxes of GroupA are of the Zermelo Fraenkel kind and deny the existence of the totality w (Clause 1 of the Russell Schema). There is much less agreementabout solutions to GroupB paradoxes,but, typically, such solutions have denied principles such as the T-schema,or cognate principles,involved in and establishingTranscendence Closure.(All this is documentedin Chs. 1 and 2 of Priest 1987.) Whetheror not each kind of solution is satisfactorywithin its own group(neither is), neitherkind of solution gets a grip on paradoxesfrom the other group. Clearly,semanticprinciplessuch as the T-schemado not occur in the paradoxes

The Structureof the Paradoxes of Self-Reference 33

of GroupA. Totalitiesdo occur in paradoxesof GroupB. However,in most cases, these are totalities of a size whose existence is guaranteed Zermelo Fraenkel by set theory.Typically,they are countablesets (e.g., the set of definableordinalsin Konig's Paradox),but they can even be finite! (as in Berry's).Hence, the PUS, in conjunction with the main result of this paper,is sufficient to sink virtually all orthodoxsolutions to the paradoxes. It is worth noting that some logicians have found Ramsey's categorisation problematicon the ground that the schema 3xVy(Rxy<->'Rxx) underlies paradoxes in both of Ramsey's classes; again with an intuitive appreciationof the PUS, they have proposedfaulting this as a solution to the paradoxes.(See, e.g., Martin1977.) However adequatesuch solutions are, it shouldbe noted that, as a general approach,this fails because there are many paradoxes,such as Berry's, that are simply not of this form. (See e.g. Priest 1983.)

6. Curryparadoxes
Let me conclude with a word on a kind of paradoxthat is relatedto some of the paradoxesI have been discussing, and which also cuts across Ramsey's division: Curryparadoxes.Some of the paradoxesI have discussed proceedby establish-a. ing a sentence of the form a<-> (All the paradoxesin GroupBii do this and some of the paradoxesin GroupA, notablyRussell's, but not the paradoxesin Bi or the other paradoxesin GroupA.) For each paradoxof this kind, we can form a new paradoxby replacing - a uniformlywith a-->3, where/ is an arbitrary formula, or, more simply, with a--I, where I is some logical constant entailing e-verything. Using the Absorptionprinciple(a-*(a->f3)F a-*f3) we can then infer a--I, and hence a, and hence l. (See Priest 1987, 6.2.) Do such paradoxesfit the general scheme given here? Yes and no, depending on what -- is. If it is a materialconditionalthen, in most logics, a--L is logically equivalent to -a, and so the curried version of each paradoxis essentially the same as the uncurried form. If, on the otherhand,-- is a non-material conditional (e.g., a strict conditional),then a->L and -a are quite differentnotions, at least primafacie. (Evaluatingthe truthof the first at a world first requiresa considerationof whatis happeningat otherworlds;evaluatingthe truthof the second does not.) In this case, the curriedversions of the paradoxesbelong to a quite different family. One may thereforeexpect them to be solved in a differentway, most obviously by the rejectionof Absorption.' Departmentof Philosophy The Universityof Queensland Brisbane Queensland 4072 Australia
I A version of

GRAHAM PRIEST

this paperwas given at the meeting of the AustralianAssociation for

34

GrahamPriest

REFERENCES Godel, K. 1944: "Russell's MathematicalLogic" in The Philosophy of Bertrand Russell, ed. P. Schilpp. TudorPublishingCo. Reprintedin P. Benacerrafand H. Putnam(eds.), Philosophy of Mathematics.Oxford:Blackwell, 1964. Martin,R.L. 1977: "On a Puzzling Classical Validity".Philosophical Review, 8, pp. 454-73. Peano, G. 1906: "AdditioneE". Rivista di Mathematica,8, pp. 143-57. Priest,G. 1983: "TheLogical Paradoxesand the Law of ExcludedMiddle".Philosophical Quarterly,33, pp. 160-5. 1987: In Contradiction.The Hague: MartinnsNijhoff. Ramsey, F.P. 1925: "TheFoundationsof Mathematics". Proceedings of the London Mathematics Society, 25, pp. 338-84; reprinted in D. H. Mellor, ed. Foundations: Essays in Philosophy, Logic, Mathematics and Economics. London:Routledge and Kegan Paul, 1978. Russell, B. 1905: "On Some Difficulties in the Theory of TransfiniteNumbers and OrderTypes". Proceedings of the LondonMathematicalSociety, (series 2) 4, pp. 29-53; reprinted D. Lackey, ed. Essays in Analysis.London:Allen in and Unwin, 1973.

Logic in Sydney,August 1991, underthe title "All the Paradoxesof Self-referenceHave a and Common Underlying StructureGeneratingthe Contradiction", abstractedunder that title in the Journal of SymbolicLogic, 57, 1992, pp. 1501. I am gratefulto Greg Restall for his comments.