On Rigorous Definitions Author(s): Nuel Belnap Source: Philosophical Studies: An International Journal for Philosophy in the Analytic Tradition,

Vol. 72, No. 2/3, Definitions (Dec., 1993), pp. 115-146 Published by: Springer Stable URL: http://www.jstor.org/stable/4320448 Accessed: 28/05/2009 14:36
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=springer. 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 organization founded in 1995 to build trusted digital archives for scholarship. We work with the scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

Springer is collaborating with JSTOR to digitize, preserve and extend access to Philosophical Studies: An International Journal for Philosophy in the Analytic Tradition.

http://www.jstor.org

NUEL BELNAP

ON RIGOROUS DEFINITIONS

(Received27 February1993)

Definitions are crucial for every serious discipline.' Here I consider them only in the sense of explanationsof the meaningsof words or as other bits of language.(I use "explanation" a word from common As speech, with no philosophicalencumbrances.) a furtherlimitationI consider definitions only in terms of well-understoodforms of rigor. Prominenton the agendawill be the two standard"criteria" eliminaThere is, alas, bility and conservativeness- and the standard"rules". hardly any literatureon this topic. The discussion will therefore be all preliminary, too elementary, imperfectly and plain.
1. SOME PURPOSES OF DEFINITIONS

There are two especially clear social circumstancesthat call for a meaning-explaining definition,and then manythat are not so clear.The clear ones call either for (1) "dictionary" definitionsor for (2) "stipulative" definitions;one of the less clear circumstancescalls for an (3) "analysis."
1.1. "Dictionary" "Lexical"Definitions or

One mightneed to explainthe existingmeaningof an old word;that is, a word alreadyin use in the community,but unfamiliar the person to the wanting explanation. 1.
EXAMPLE.(Lexical definitions) (1) What is a sibling? A

sibling is a brother or a sister. (2) What does it mean to square a number?One obtains the square of a numberof it multiplying by itself. (3) Whatdo you mean by zero? Zero is the least integer. (4) What is a brother?A brother is a male sibling.
PhilosophicalStudies 72:115-146, 1993. C 1993 KluwerAcademic Publishers. Printedin theNetherlands.

116

NUEL BELNAP

I postpone discussion of (a) the question whether these interchanges should be sprinkled with quotation marks and (b) the circularity threatened combining sibling/brother the by examples. 1.2. "Stipulative" Definitions One might wish to explain a proposed meaningfor a new word. The purpose might be to enrich the languageby makingclear to the communityof users that one intends that the new word be used in accord with the proposed meaning.This case squarelyincludes the puttingto workof an "old" wordwitha new technicalmeaning. 2.
EXAMPLE.(Stipulativedefinitions) (1) Let omega be the first ordinalafter all the finite ordinals.(2) Let a group be a set closed under a binaryoperationsatisfyingthe following principles ... . (3) By a terminological realist I mean a

philosopher who subscribes the followingdoctrines... to The person who introducesa new word mighthave any one of various purposes.Here is a pair:the person mightwant a mere abbreviation, to avoid lengthyrepetition("ByE2 I referto volume 2 of Entailment: The
Logic of Relevance and Necessity"); or the person might take it that he

or she is cuttingat a conceptualjoint (a complete lattice is a partially orderedset in which every subsethas a least upper and a greatestlower bound). 1.3. "Analyses" "Explications" or There are many cases not exhibitingeither of these clear purposes, includingperhapsmost distinctivelyphilosophicalacts of definition;in these cases (Carnapcalls some of them "explications") wants both one to rely on an old, existing meaning and to attach a new, proposed meaning; it seems that one's philosophical purposes would not be servedif one let go of eitherpole. 3. EXAMPLE. (Explicativedefinitions) (1) Let knowledge (in the present technical sense) be justified true belief. (2) We say that A impliesB if B is true on every interpretation on whichA is true.

ON RIGOROUS DEFINITIONS

117

Observethat in these cases the philosopherneitherintendssimplyto be reportingthe existing usage of the community,nor would his or her purposesbe satisfiedby substituting some brandnew word. In some of these cases it would seem that the philosopher'seffort to explain the meaningof a word amountsto a proposal for "a good thing to mean by" the word. I learnedthis phrasefrom Alan Ross Anderson.Part of the implication is that judging philosophical analyses is like judging eggs:Thereareno shortcuts; each has to be held to the candle. 1.4. Invariance Standard of TheoryAcross Criteria Rules and Purposes: The extraordinary thingis this:The applicability the standardtheory of of definitionremainsinvariantacross these purposes - just so long as the purpose is to "explainthe meaning of a word." (The phrase "a word"suggeststhat it is a single word that is being defined,whereasthe standard theory can certainly do better than that. For simplicity, however, I omit discussion of the simultaneousexplanationof several termsby meansof a singleact of definition.) This standardtheory has two parts. In the first place, it offers two criteria for good definitions: the criterion of eliminability (which requires that the defined term be eliminable in favor of previously understoodterms)and the criterionof conservativeness (whichrequires that the definition not only not lead to inconsistency,but not lead to
anything
-

not involving the defined term

that was not obtainable

before). In the second place, it offers some rules for good definitions, ruleswhichif followedwillguarantee the criteria satisfied. that are History. The standard theory of definitions seems to be due to Lesniewski,who modeled his "directives" the work of Frege, but I on cannottell you whereto find a historyof its development. The standard citation seems to be Lesniewski 1931; see also Lesniewski 1981 (Collectedworks).I learnedmost of the theoryfirst from Suppes 1957, who credits Lesniewski (p. 153, note). There should have been minihistoriesin either Church1956 or Curry1963, but I couldn'tfind what I was looking for. The matter was well understoodby Frege (e.g. in Frege 1964), Couturat (see Couturat 1905), Carnap (e.g. in Carnap 1937) and Tarski (see e.g. Tarski 1941 for some well-chosenelemen-

118

NUEL BELNAP

tary words). Tarski himself contributed heavily to the theory, as evidenced in the material translated and reprinted in Tarski 1956. There Tarski gives the dates and circumstancesof his own early contributionsin the 20s and 30s. But no one of these lays out an account of the history of the matter in its beginnings.The standard histories of logic (Bochenski 1956, Kneale and Kneale 1962) do not discuss modem theories of definition.Neither does Kneebone 1963. Neither does Church'sarticleon "definition" Runes 1962. The 207in page book Robinson 1950 neither discusses the technical theory nor refers to its history (though there is some referenceto the history of nontechnicaldiscussions).The definitionarticlein TheEncyclopediaof Philosophy(1967) does not even mention Lesniewski.The only useful general references I happen to know are the definition article in the Dictionaryof Logic, Marciszewski1981, and some penetratingparagraphs and authoritative citations in Luschei 1962 (see especially pp. 36-37 andnn. 34 and 78). The criteriaare like the logician'saccount of semanticconsequence: The latteris proposed as an accountof "goodinference," formeras the an account of "good definition." And the standardrules are like the in logician'srules definingderivability a particular system:If you follow the logician'srules for constructingderivationsin his system, you will derive all and only correct semanticconsequences,i.e., you will make all and only "good inferences."In the same way, if you follow the logician'srules for constructingdefinitions,you will offer all and only definitionsthat satisfythe criteriaof eliminability conservativeness. and You will, that is, offer all and only "gooddefinitions." analogyruns The deep: Just as it is a hard theorem that in first order logic semantic consequences and derivability(in some one formal system) are in so agreement, it is a hardtheoremthat the two criteriaand the standard rules are in agreement.The hard part is Beth's DefinabilityTheorem. There is an additional parallel. We know that agreement between semantic consequence and derivability(in some one formal system) gives out when first order logic is enriched in any one of several respects. So does the agreementbetween the criteriaand the rules of definition. I discuss the two criteriain some length, and then say just a little aboutthe rules.

ON RIGOROUS DEFINITIONS 2. ELIMINABILITY AND CONSERVATIVENESS I

119

The standardcriteriafor good definitionsare those of "eliminability" Where do these criteria come from? Why and "conservativeness." should we pay attentionto them? Indeed, we may and shall raise the presupposedquestion of whether we should pay them attention.It is literature nearlysilent on is that surelyremarkable philosophical-logical these questions. I have only the most elementaryof answersto propose. It seems to me that we can derive a motivationfor these criteriafrom the concept of definitionsas explanations the meaningsof words (or other bits of of language). Under the concept of a definition as explanatory,(1) a definition of a word should explain all the meaningthat a word has, and (2) it should do only this and nothing more. That a definition should (1) explain all the meaningof a word leads to the criterionof eliminability. That a definitionshould (2) only explain the meaningof the word leads to the criterion of conservativeness. Observe that (1) and (2) are not quite analogousin their respectiveplacementsof "all" and "only." This warnsus that, as we shall see, the two criteriaare also not exactlyanalogous. 2.1. Criterion Eliminability of (RoughAccount) One may approachthe criterionof eliminability from the directionof the "use"picture of meaning, with picture-slogan,"meaningis use." Then to explain all its uses, that is, its use in every context. (I intend that this recipe neglect ambiguity).The advance is this: The metaphorical quantifierin "all the meaning"is cashed out in terms of a nonmetaphorical stillimprecise)quantifier (if over contexts. And what is it to explainthe meaningof a word in a single context? That depends on what counts as a context. The traditionhas found it useful to concentrateon (declarative) sententialcontexts.I shalldo so as well, even though these contexts are not enough for an adequatestory about language.So what is it to explain the meaning of a word in a single sententialcontext,say B? The standardmove, and a good one, is to say that it is to explain the meaning of the containingsentence B. The next standardidea is this:To explainthe meaningof a sentence is

120

NUEL BELNAP

to explainits role in inference. Combiningideas yields the doctrinethat to explainall the meaningof a word is to explainthe inferential connections of each containing sentenceB. At this stage there enter two new elements.First, explanationsquite generallyare more prized if they are given in terms that are previously understood. This we may call noncircularity. Second, in favorable situationswe can hope to explainthe inferentialrole of a new sentence by identifyingits role with that of an old sentence.This we may call no inferential enrichment. apologize for the jargon.)The combinationof (I these two elements comes to this: There is in the language another sentence that (a) does not contain the defined term and (b) occupies exactly the same inferential role as that wanted for the containing sentence B. The standard account presupposes that the situation is favorable.Accordingly,the criterion of eliminabilityrequiresthat for each sententialcontext B containingthe defined word, the definition give enough informationto allow formationof an inferentially equivalent piece of language B' that contains only previously understood words. Then, and only then, will we be sure that we have explainedall the meaning of the word to be defined - whether our purpose is to explainan existingmeaningof an old word or to give a new meaningto a new word. In either case, for each context in which the defined word can grammatically occur, we must provideenoughinformation allow to constructionof an equivalentstatementthat contains only previously understoodwords. Under the assumptionthat the defined word is not "previously understood," whatis then requiredis enoughinformation to permit "elimination" the defined word in favor of other words that of arepreviously understood. The above account needs considerablerefinement;containingas it does several obscure phrases, it cannot itself count as a definitionof "satisfiesthe criterion of eliminability" that satisfies the criterion of I eliminability. postpone this necessaryrefinementuntil after a preliminary accountof the criterionof conservativeness, pause to illustrate but some of the obscurity withan example. 4. EXAMPLE. (Holiness) Holiness is what the gods all love (Euthyphro).

Taken as a definition, and read in a standardtwo-valued way, this

ON RIGOROUS DEFINITIONS

121

permits eliminationof "holy"in "extensional" contexts, which is good. As Socrates pointed out, however, it does not permit eliminationof in "holy" "becausal" contexts.Socratesasked Euthyphroto considerthe followingplausiblepremiss. Holy acts areloved by the gods becausethey areholy. If we tried to use Euthyphro'sExample 4 to eliminate the second
occurrence of "holy" - the one in the scope of "because" - we should

findthatwe wereled to the followingimplausible conclusion. Holy acts are loved by the gods because they are loved by the gods This leadsus to see that"context" needs specification. We should keep in mind quotationcontexts, and also psychological contexts involving rapt attention such as "he was turningover in his mind the questionof whetherprosecutionof his fatherwas holy."If we do then we will be more likely to keep remembering that no definition will permit eliminationin absolutely every context. Such remembering will perhaps disinclineus to ask more of a definitionthan it can very well deliver. But perhapsnot. Some people respond by cooking up an account of quotes (for example)on which they do not count as "contexts." This seems to me an unhelpful obfuscation of the general principle. The matterof contexts is easily overlookedand criticallyimportant. Since no definition can sensibly license universal elimination, any sensible act of definition must include a reference to the family of intended contexts. The point is easily overlooked because those who care about the proper use of definitionsfrequentlythink of themselves as alreadyhavingspecifiedsome background familyof contexts,e.g. the firstorderfunctional calculus. The matterof no inferentialenrichment also easily forgettable, is but should not be forgotten.It is easy to describelanguagesin which there are definiteinferentialroles whose statusis changedfrom "unoccupied" to "occupied" stipulativedefinitionsexploitingpreciselythose roles. by Here is an uncomplicated example. 5.
EXAMPLE. (The Absurd) Let the language be entirely

122

NUEL BELNAP

positive, lacking any trace of negation, but be otherwise standard.Stipulatethat to say of anythingthat it is F is to say somethingthat implies everything.Thus each sentence Fa paysthe role of The Absurd. This evidently defines an inferentialrole for each sentence of which each Fa is a part, and thus explainsall the meaningof F. But since no sentence alreadyin the languageimplies every sentence, the definition is inferentiallyenriching and does not satisfy eliminability.(Observe that the failure of eliminabilityis not due to a violation of noncircularity.) 2.2. Criterion Conservativeness of (Rough Account) On the standardview a good definition(in the sense of an explanation of meaning)should not only explain all the meaningof the word, as required by the criterion of eliminability.It should only do this. It should be empty of assertionalcontent beyond its ability to explain meaning.If it were to go beyond the job assigned,say by claimingthat Euthyphro is pious, it might indeed be doing something useful. It would,however,no longer be entirelysatisfactory a definition. this as In perfectly good sense, a definition should be neither true nor false (whetherexplainingan old word or introducinga new one): Whether or not it has a truthvalue qua definition(we mightargueabout that),it shouldmakeno claimsbeyondits explanation meaning. of There is a special case to which I shall not return.Clearlya definition should not take you from consistencyto inconsistency; that would be a dreadfulviolation of conservativeness. The older theoreticiansof definitions in mathematicswere, however, insufficientlysevere when they suggested that consistency is the only requirementon a mathematicaldefinition. Terminologyfor this criterionis a little confusing.Some folks use "conservative" mean somethinglike "does not permit the proof of to anything we couldn't prove before." This accounts for calling the second criterion that of conservativeness: definition satisfyingthe A second criterion is conservativein that very sense. Other folks use "creative"to mean something like "permitsproof of something we

ON RIGOROUS DEFINITIONS

123

couldn't prove before."It is with this sense in mind that the second I criterionis sometimes called "the criterionof noncreativity." opt for the former name merely to avoid the negative particle.Under either name, the criterion demands that the definition not have any consequences (other than those consequencesinvolvingthe defined word itself) that were not obtainablealreadywithoutthe definition.If before the definitionwe could not establishthat Euthyphrois pious, and if the definitionis of neither"Euthyphro" "pious" nor (but perhapsof "zero" or "sibling"), it shouldnot be possible to use the definitionto show then that Euthyphrois pious. Were we able to do so, the definitionwould manifestlycontain more informationthan a mere explanationof meaning:It would not be conservative. Neitherin explaining meaningsof the old words nor in introducingnew words should one use the cover of to definitions smugglein freshassertions. bad manners. It's 6. EXAMPLE. (Noid) Suppose that Wordsmith introduced "noid"by announcingthat all the guys on the other side of Divider Street are noids (a sufficientcondition)and that all noids arebornlosers(a necessarycondition).

Wordsmith's plan appearsinnocent:He wishes to help his gang understand "noid"by squeezing it between two understood ideas. Wordsmith'sintroduction "noid," of however,may or may not be innocent.If the gang and Wordsmithhave already committed themselves to the propositionthat all guys on the other side of Divider Street are born losers, then Wordsmithhas not by this "definition" violated conservativeness.But if Wordsmithhas not done this, if he has not committed himself and the gang to the born-losernessof all guys across Divider, thenhe has violatedconservativeness. This consideration is easily mixed with two others. There is in Example6 an obvious differencein the emotionaland practicalfreighting of the two bits connected via "noid."The one starts out in the unloaded "space language,"as Carnap might say, while the other suggests contempt and a practice of intimidation. What should not be mixedis the divisionbetweensufficientand necessaryconditionson the one hand, and the divisionbetween unfreighted and freightedlanguage on the other. Furthermore,neither should be mixed up with what is relatedto the possibilityof justificatory warrant.

124

NUEL BELNAP

There is a middle case. Suppose Wordsmithis participatingin a practiceaccordingto which his announcementcombines (1) the assertion that all guys across Divider Street are born losers with (2) the conservativeintroductionof "noid."There are two cruciallydifferent subcases. Revealed assertion:The practice is such that Wordsmith's form of words bears on its surface the division into assertion and introduction.Concealedassertion: The practiceis such that the form of words conceals the assertionalcontent or even misleads into the false suppositionthat there isn't any. I shall pursue the differencebetween revealedandconcealedassertionbelow. 7. EXAMPLE. 0 (Emptyset) SupposeJanetsays "Definition: is thatset thathas no members."

If Janet's communityhas already fixed on a theory that implies that there is a unique set with no members,she is all right.But since this "definition" permits her to prove that there is a set with no members (by existential generalization),she violates conservativenessif she couldn't prove it before. She would then be using the "definition" to smugglein an existenceclaimto whichshe is not entitled. Supposethe worst,however,thatJanet'scommunity a Lesniewskihas like theory accordingto which no set has no members.Suppose she goes ahead and introduced 0 in accord with Example 7. What should we say about her? What did she denote by 0? Probablyno one cares, but there are cases like Janet'sabout which philosophersdo seem to care. For exampleLeverrierintroduced"Vulcan" the planet responas sible for certain astronomical anomalies. Later it was learned that relativisticconsiderationssufficedfor the anomalies;there was no such planet.Whatshallwe then say about the meaningof "Vulcan" of all and the discourseusing this term?My own view is this. You should say that the answerto "Isthere a uniqueplanetresponsiblefor the astronomical anomaliesat issue?"is "No."You should say that the answerto "Isthe definitiondefective?"is "Yes."You should say that some of what we prize in science and other activitiesdoes not rest on answersto these eminentlyclear questions,since we often manageto get along very well when presuppositionsfail. You should add that neverthelessthere are no general policies for what to do or what to say in the presence of defectivedefinitions. Just about anythingyou remarkbeyond this will be

ON RIGOROUS DEFINITIONS

125

either artificial or unclear, and is likely only to add to the general babbleof the universe. It is easy to imaginelanguagesthat allow nonconservative creative or definitions in the following sense: In them a single speech act both introduces the new terminologyand makes the assertion needed to In justify that introduction. fact English, even scientificEnglish,is like that. Our practices are lax. I suppose that is one thing people mean whentheyadviseus to abandonthe analytic-synthetic distinction. We should not, however,follow this advice.Those philosopherswho wish to be clear should follow Carnap in maintainingthe analyticsynthetic distinction to the maximumextent possible. That is, when confrontedwith a complex speech act such as that of Leverrier,it is good to idealize it as consisting of two distinct components: the assertionof the existenceand uniquenessof an anomaly-causing planet, and the definitionof Vulcan as this very planet. There is, I think, no other way to continue with a clear conversationabout the matter.For example, I do not think we can otherwise make a clear distinction between revealed and concealed assertions. And we should do so. Speech acts that conceal assertions are (whatever their intentions) We misleading. should try to avoid them wheneverwe are clear enough to be able to do so. This, as we may infer from the historyof even the best science, may not be often. Even when we are muddled,however, we should at least try not to pretendto more claritythan the sad facts warrant. 2.3. JointSufficiency In conclusion of this preliminaryaccount, I add the following. The proposal is that each of the criteriaof eliminability and conservativeness are necessaryfor a good definitionin the sense of an explanation of meaning;and that togetherthey are sufficient(so that their conjunction providesa good definitionof "gooddefinition"). pros and cons The of necessityare easily discussedthroughexamples.In contrast,I do not much know how to defend the claim to sufficiencyother than by a rhetoricalquestion:You've asked for a good definition of this word, and now I've given you a procedure for eliminatingit from every intended context in favor of somethingyou yourself grant is entirely

126

NUEL BELNAP

and equivalent; I've done nothingelse (i.e., I haven'tslipped any further statementspast you); so what more do you want? Until this rhetorical device can be replacedby somethingbetter, there is conspicuouswork to be done.
3. ELIMINABILITY AND CONSERVATIVENESS II

Such a rough account of the criteriaof eliminability and conservativeness calls for an increase in precision.The standardaccount provides exactlythis. 3.1. Limitsof Applicability You cannot buy an increase in precision without paying in the good Here the costs are of threekinds. coin of limitation. First, the discussion applies only to a communityof languageusers whose languagecan profitablybe describedby labelingit an "applied first order language."You should imagine that this language is a fragmentof English,a fragmentthat the users think of as structured in terms of predicate, function, individual, and perhaps sentence constants; truth-functionalconnectives; individual variables; quantifiers; and identity.But you should also imagine that this languageis really used and thereforehas lots of English in it. Think of the languagethat some mathematicianssometimes employ: a mixture of English and notation, but with the "logic"of the matter decided by the first order calculus. For example, the languagewill use English common nouns such as "set"; only in locutions such as "Somesets are finite"that but the usersthinkof as translatable the usualfirstorderway. in I need to say more about what this first limitationmeans. So far I have said that the users take the grammar of their language in the standard first order way. They also think of their proof theoretical notions as given in the standardway:I am thinkingof "axioms," "rules," "theoremhood,""derivabilityfrom premisses," "theory,""equivalent relative to a theory," and so forth. And lastly, they think of their semantic concepts in the standard way: "logical truth," "(semantic) consequence of some premisses,""(semantic) equivalencerelativeto a theory,"and so forth. They know, throughGodel, that there is agree-

ON RIGOROUS DEFINITIONS

127

ment between the appropriateproof-theoretical and semanticideas. I shall refer to these mattersby saying that the first order languagehas
an inferential use, or by saying that its sentences occur in inferential contexts.

In the second place, I consider only definitions (explanationsof meaning)of predicateconstantsand function constantsand individual constants. This limitation is made reasonable (but is certainly not forced) by the previousdecision to consideronly a first order language. There is no consideration, for example, of Russell's treatment of definite descriptions,or of the meaningof a conventioninvolvingthe droppingof outermostparentheses,or of quantifiersof a certainstyle havingsome limitedrange. In the third place, I am going to follow the standard account in consideringonly definitionsthat are themselvessentential,and that are in the very same languagecontainingthe defined and definingterms.In imposingthis limitationI do not intend to be decidingwhetherthe act of defining is "really"imperativalinstead of assertional,or "really" ratherthan not. I do intend to assert, however, that the metalinguistic technicaltheory of definitiongoes very much more easily if in givingit one has to talk about only one language:the languageincludingthe defined term, the definingterms,and the definitionitself. In fact we do need to considertwo languageswith respectto vocabulary: language the with, and the languagewithoutthe newly definedterm;and that is quite enough in the directionof complicationfor such an elementarydiscussion as this. We can therefore see that the policy, exhibited in Example 1, of giving example definitions above in the "material mode" (as Carnap puts it) correctly forecasts this decision. By taking the definition of to is "sibling" be "Anything a siblingif and only if it is a brotheror a sister" instead of "Replace 'sibling' by 'brother or sister' wherever found!",the technical work is greatly simplified,as you will see for yourselfif you try to spell out a technicallyadequatetheory of definitions of (say)the metalinguistic imperative kind. But, you may say, isn't a definition a speech act? How can you pretend adequacy for a theory of definitionsthat does not contain a theory of such acts? Answer: There is no pretense. The standard account is not adequate to the aim you have envisaged. There is,

128

NUEL BELNAP

however, no existingrigorousand helpful explicit theory of definitions in the guise of speech acts carriedout by agents.Such a theory would need to be founded on a general theory of speech acts carriedout by agents.And this in turn would need to be founded on a generaltheory of agency.Perhapsthe Belnap/Perloff"seeingto it that"approximation to agency ("stittheory")is sufficientfor the purpose;it seems plausible thatthis shouldbe so. But the workremainsto be done. Account 3.2. The Terms the Standard of Recall that the focus is on definitions (1) that are of individual constants, predicate constants, and function constants;(2) that are in the "material mode";and (3) that are in (not an arbitrary languagebut) an applied first order language with an inferential use. This focus permitsus to give the criteriaof eliminability and of conservativeness themselvesin the form of definitions,namely,definitionsof "satisfying the criterion of eliminability" and of "satisfying criterion of conthe servativeness." It turnsout that there are four key entitiesinvolvedin explaining the standardideas of eliminabilityand conservativeness; as one might or, well say, each of these ideas is explained as a four-placepredicate.I shall attachto each of these key entitiesa variable,but for intelligibility I shall use for each a spelled-out variablehaving a certain mnemonic force. * Theoryis the backgroundtheory (before the definition)in the context of which the definitionis being entered. Theory is a set of sentences of the usual kind; sometimesit is best to think of it as closed under semantic or proof-theoretic consequence,but my use of the variableis such that often Theorycan be coded as a set of axioms.If the background theory is thoroughlyinterpreted,Theorycan even be the set of truths. Previous-definitions the set of previousdefinitions- a set is of sentences. This element is requiredto permit discussion of circularity(a brother is a male sibling; a sibling is a brotheror sister).

ON RIGOROUS DEFINITIONS

129

Definition will be the definition itself. I remind you that Definition is a sentence on a par with the rest of the sentences in the context of which the act of definition is taking place. One can therefore meaningfullyand without and quibblespeak of its grammatical deductiveand semantic properties and relationships.It is just another first order sentence. Symbolis the new symbolbeing defined.Recallthat Symbol is to be a (new) predicate,operator,individualconstant,or perhaps a sentence constant.The deepest theory would not it requireor even permit Symbol to be a symbol,permitting functionof or even requiringit instead to be a grammatical an appropriate type;but I shall not in this discussionplumb thatparticular abyss.

The above discussionhas mixed intuitiveand rigorousconsiderations. Here I specify the latter,in particular suppressingall verbs of psychologicalattitude. * * * * is Theory a set of sentences. is Previous-definitions a set of sentences. Definitionis a sentence. Symbol is a symbol (a predicate,operator,individualconstant,or perhapssentenceconstant).

3.3. Criterion Eliminability of (RigorousAccount) It is important to have the standard account of eliminabilitymade rigorousso that one can more plainlysee whatis at stake.As promised, the accountis presentedas a definitionof a four-placepredicate.Recall that it is a presuppositionof the discussion that sentences are made exclusively way of truthfunctionsandquantifiers. by 8. Let DEFINITION. Previous-definitions, (Eliminability) Theory, Definition, and Symbol be just as above. Then Definition, qua definition of Symbol, satisfies the criterionof eliminaif bility relativeto Theoryand Previous-definitions and only if: For all (possibly open) sentences B in the languageof

130

NUEL BELNAP

Theory,Previous-definitions, (especially) Symbol,there and is a (possiblyopen) sentenceB' suchthat 1. B' is in the language Theory Previous-definitions, of and 2. Symboldoes not occurin B', and 3. B and B' are EQUIVALENT relativeto Theory,Previand ous-definitions, Definition. Informally: For every sentence with the symbol there is an equivalent sentencewithoutthe symbol- so all its meaningis given. The standard account unpacks the critical notion of EQUIVALENCE in either of two interchangeable or ways: proof-theoretically semantically.

Proof-theoretically: must be able to prove the "material One equivalence"of B and B' from the theorems, previous definitions,and the definition(or their "formal equivalence" if either is open). That is, B B' (or its universalclosure) must be derivable (in a standard calculus) from Theory, Previous-definitions, Definition. and B Semantically: and B' must be semanticallyequivalentin the context of the theorems, previous definitions,and the definition;that is, B and B' must have the same truthvalue in every interpretation the nonlogicalconstants(and free of variables if present) that renders all of Theory,Previousdefinitions, Definitiontrue. and

Equivalenceis at the heart of the eliminability. Because equivalenceis well understoodfor truth-functions-cum-quantifiers languages,we can be sure that for these languagesthe standardaccount makes sense. If, however, a language does not have a well-understood concept of thenwe cannotbe surewhatto makeof eliminability. equivalence, It is plain that the concept of equivalence used in the standard accountis an idealizationof "inferential equivalence." Whatis confusing is that this strong notion of equivalenceis required even though the sentencesthat form the premissesand conclusionsof the inferencesare all assumed to be made from truth functions and quantifiers.In the particular, accountwould go to pieces with respectto the inferential

ON RIGOROUS DEFINITIONS

131

use of sentences if one were to substitute "materialequivalence"in clause 3 above. This is not surprisingsince the gap between material equivalenceand inferentialequivalenceis a pretty severe chasm.Here are two witnesses to the cleft. (1) Clause 3 relativizesequivalenceto Theory,Previous-definitions, Definition;but it makes no sense to and relativize material equivalence. (2) Suppose, contrary to sense, that clause 3 were stated with material equivalence. To make true the
resultant "for every B ... there is a B' such that ... B is materially

equivalentto B"', we would need only two valuesof B': a true one and a false one. Thatwouldbe too easy. 3.4. Contexts Meaning and The discussion of definitionsprobablygoes better if we elaboratethe confusingmatterof "contexts." In the first place, there are explicitcontextsand implicitcontexts.In the exampleat hand, all the explicitcontexts are providedby the usual array of extensional predicates, extensional operators, extensional connectives(truthfunctions),and extensionalquantifiers leadingup to a declarativecontext.The implicitcontext is that of inference(or some close cousin). There is in the standardlanguageno "explicitation," in Brandom'sphrase,of the inferentialcontext.We have to pay attention to it, and so do the languageusers, but they have no device by which explicitlyto speakof inference.Theyjustdo it (so to speak). In the second place, we can therebysee that there are two contrasts, two differentdimensionsalongwhich'context' couldbe enriched. With regardto explicit contexts, one might have added their modal featuresor "metalinguistic" featuresso as to give the users new explicit contexts useful in speaking of inferentialconnections. This addition might be preciselyto explicitatesome contexts so far left implicit.The standardaccount does not deal with these. When such materialsare added, we expect an account of meaningthat is no longer extensional. We expect new predicates,operators,connectives,and quantifiers that are intensionalinstead of extensional.The general outlines of eliminability and conservativenessremain, but the rigorous details must be elaborated. With regardto implicit contexts, the inferentialcontext imaginedin

132

NUEL BELNAP

the standardaccountis thin stuff.For example,it containsonly distant approximations any of evidence, interest,relevance,explanation,or to practical decision. These are cognitive and practical concepts, essentially relativizedto specific persons with specific mental equipmenton specific occasions. The standardaccount does not provide a notion of equivalence that can sustain replacementin such contexts. Such an equivalencewould have to work for intentionalas well as intensional contexts.It wouldneed to provideus withsense equivalence. Becauseboth the inferentialor intensionalcontextsand the cognitive or intentionalcontexts are often implicitratherthan explicit,it is easy to forget this. Even Tarski, the most methodologicallysophisticated definer of all time, once made a mistakeabout context.The mistakeis easy to verifybecause Tarski 1935 stated his methodologyso carefully: For a proposed definitionof "truth" be adequate,it is both sufficient to and necessarythat it be "formally correct"and "materially in adequate" the well knownsenses that Tarskigave to these phrases(p. 188, including note 1). EvidentlyTarski'sidea of formalcorrectnessis just what I am calling the standard account. It is equally evident that Tarski's condition of material adequacy suffices to fix the inferentialrole of "truth," not just its actual extension,as some folks sometimessay. and This is made explicitby his statementin Tarski 1944 that all adequate definitions would have to be "necessarilyequivalent"(p. 354). It is anyhow made implicitly clear by his use of the concept of "consein quence" his statement material of adequacy. The mistake occurs when Tarski suggests an application for his definition. Under the heading of "applicability semantics to the of methodology of the empirical sciences," Tarski first considers the reasonableness the following. of An acceptable theory cannot contain (or imply) any false sentences. His verdict is negative,partlybecause "we do not know, and are very unlikelyto find, any criteriaof truth which enable us to show that no sentenceof an empirical theoryis false"(p. 367). This applicationby Tarskiis admissible. Whatmakesit so is that the occurs in the above sentencein an extensionalcontext. truth-ascription This is the very sort of context for which the techniques on which

ON RIGOROUS DEFINITIONS

133

Tarski relies are apt. "Acceptability" of course a cognitiveidea, but is is the truth-ascription not withinthe scope. Tarski goes on, however, to propose that the following is "an importantpostulate which can be reasonablyimposed on acceptable empirical theoryandwhichinvolvesthe notionof truth": As soon as we succeed in showingthat an empiricaltheory contains(or implies)false sentences,it cannot be any longer consideredacceptable(ibid.). Here Tarskihas made a mistake.The difficultyis that in this postulate does occur withinthe scope of a (not just nonextenthe truth-ascription sional but) intentionalconnective.To show this, let me do as best as I can to put the Tarskipostulateinto middleEnglish. For each time t, for each empiricaltheory T, if (there is a sentence x such that at t we succeed in showingthat ((x E 1) & - (x is true)))then T is not acceptableby us at t. The mistakeis not that the postulateis false;it seems like quite a good thing to say. Nor is it that Tarski'ssemanticdefinitionof truthcannot help elucidate the postulate,though I certainlythink the prosentential theory of truth does a better job of that. The mistakeis about definitions. Tarski's mistake is that his definitional methodology fails to license an eliminationof "x is true"within the scope of an "at t we succeed in showingthat ." A moment'sthoughtsuffices to conclude that inferentialequivalencebetween B and B', which is what Tarski's methods guarantee,could not possibly suffice for replacingB with B' within the scope of this nonextensionaland nonintensionalcognitive connective.The rangeof applications his definitionthat are secured for his methods,thoughgreat,are less thanhe envisaged. by So the standardtheory does not provide for eliminationin intentional (as opposed to intensional)contexts.It cannot since its concept of equivalenceis itself intensional equivalence(or inferentialequivalence) ratherthan intentionalequivalence(or cognitiveequivalence,or sense equivalence).The road forks. (1) Some may argue that we want an enrichmentof the standardaccount of definitionthat would permit eliminationin cognitive contexts. (2) Some may argue that we keep to the intensionality-without-intentionalitythe standardtheory,thereby of

134

NUEL BELNAP

giving up a call for eliminationin intentionalcontexts.For a powerful discussion of the issues with particularreference to Frege, see Horty 1993. Horty gives a compelling argumentfor (2), which is now my
view.

These remarks are relevant to the aim of "extensionaladequacy" sometimes proposed for a definition (but not by Tarski). Having attained such an aim, one has no measure of the inferentialuse of sentences containingthe newly defined Symbol.Extension alone does not suffice to specify a role in inference. Given only extensional adequacy,it thereforemakes doubtfulsense to infer either from or to any sentence with the newly defined Symbol. Furthermore,since the practiceof assertionis unsnippablytied to inference,we should worry equallyaboutassertionon the basisof mereextensional adequacy. One mighttry to base a counterto this line on an idea publishedin the same year by Bressan 1972 and Kripke1972, and muchearlierin a somewhat different form by Marcus 1961. One may take the newly defined symbol to be "rigid," adapt Kripke'sphrase,or to fall under to an "absolute concept,"to adaptBressan's,or to be a "tag," followingthe idea of Marcus.(These ideas are not the same;absolutenessis deepest of these, and the closely relatednotion of "substance concept"in Gupta 1980 is equally deep and more manageable.But any will do for the immediate purpose.) Taking this line makes an indexical out of the newly introducedsymbol,somethinglike "theextensionof the definiens as things actually stand here and now."There is nothing even mildly wrong with such a definition;but it cannot be treated as a continuous part of a language,such as thatwe were envisaging, that is "standard" in being free of indexicals.The only mistakewould be to pretend otherwise. In the standardlanguage,there is no room for such a counter. Even in the richer language one would find the counter to fail, for carefulanalysisrevealsthat you cannotavoid the need for an inferential role, and therefore you cannot avoid awardingan intension.All that happens when a "rigid"term is introduced is that its intension is definitionallydeterminedby its indexicallydeterminedextension. My point, however,is not to make theory.It is only this:If we are to think seriouslyabout such definitions,it should be againstthe background of a carefulandrigorousaccountof them.

ON RIGOROUS DEFINITIONS

135

3.5. Criterion Conservativeness of (Rigorous Account)

9. DEFINITION.(Conservativeness) Let Definition, Symbol,

Theory,and Previous-definitions as above. Then Definibe tion, qua definition of Symbol, satisfies the criterion of relativeto Theoryand Previous-definitions conservativeness if and only if: For all sentences B in the languageof Theory and Previous-definitions(but not containing Symbol), if and Definition,Theory, Previous-definitions togetherIMPLY B, then Theory and Previous-definitions (already, without the necessityof Definition)togetherIMPLYB. for Using the languageof consequenceas a substitute that of implication,as I usually do below, conservativeness comes to this:If a Symbol-freeB is a CONSEQUENCEof Theory and Previous-definitions with the help of Definition,then it is so withoutits help. Informally:Every sentence without the symbol is a consequence withoutthe definitionif it is a consequence at all - so the definition does nothingbut give the meaning. The notion of CONSEQUENCE used above can be unpacked in eitherof two (interchangeable) or ways:proof-theoretically semantically.

* *

Proof-theoretically, consequence means derivability (in a standard calculus). Semantically, B to be a consequenceof some sentences for is for B to be true in every interpretationmaking those sentencestrue.

Just as you cannotsubstitutematerialequivalencefor equivalencein the preceding discussion of eliminability,so here you cannot substitute materialimplicationfor implicationor consequence.These ideas make no sense without an inferentialcontext. For example,suppose Theory and Previous-definitions empty.Then if we were contrasensically are to read "IMPLY" materially,conservativeness would say, I suppose, that for each Symbol-freeB, (DefinitionD B) D B, i.e. (Definition V B). And since some Symbol-freeB's will surelybe false, this comes to just: Definition. thatagain? Say

136

NUEL BELNAP

3.6. Examples 10. EXAMPLE. Take Definition = "forall x, if x # 0, then for all y, Inverse(x) = y iff x X y = 1."

inverse Thus Definitiontakes Inverse(x) as the standardmultiplicative of x, usuallywritten"1/x". And Inverse(x)"is defined"only when x is not zero so that the problem of division by zero is avoided. Such a definitionis often called a "conditional the definition," conditionbeing thatx # 0. Evidently Definitiondoes not satisfyeliminability, since you cannot use it to eliminate all uses of Inverse(0). It does, however, satisfy a "partial eliminability,"since Definition does permit elimination of Inverse(t) for any term t such that Theory and Previous-definitions imply that (t $ 0). More generally,Definitionpermits eliminationof Inverse(x) in any sentence VxB(x) {or 3xB(x)} in which Theoryand Previous-definitions imply that VxB(x) is equivalentto Vx((x $ 0) D B(x)) {or that 3xB(x) is equivalentto 3x((x $ 0) & B(x))}. Since these are known in advance as the only contexts of Inverse(t) or Inverse(x)that we care about, such partialeliminability conceptually is satisfying. We know that Frege thoughtotherwise;he believed in total eliminability. My own view is that Frege's belief (or practice) arose from a certainslightlymisdirectedcompulsiveness the same compulsiveness, however, that gave the world its most glorious standardof uncompromisingrigor.Many of us, perhapsalmost equallycompulsive,avoid conditional definitionsfor what we announce as mere technical convenience. Our alternate device is to use "don't-careclauses," e.g. definingInverse(0)as equal to 0 (or perhapsas equal to 14), and then remarkingthat this bit of the definitionis a don't-care.If one studies our inferential practices,however,one sees thatwe never use the don'tcare clauses, and that we counsel others to avoid relying on them for conceptualinterest.So practicallyspeaking,conditionaldefinitionsand don't-care clausescome to the samething.
11. EXAMPLE.(Solubility) Take Definition = "for all x, if x is

put in water,then x is soluble if and only if x dissolves"as a definitionof "soluble."

ON RIGOROUS DEFINITIONS

137

This is what Carnap called a "reductionsentence."He was surely thinking that Definition gave us a "partial"definition of "soluble." Then clearly Supposethe theory Theorycontains"Samis put in water." in Definitionpermitseliminationof "soluble" the context B = "Samis soluble,"for relative to the theory Theory,B' = "Sam dissolves"is equivalentto B. But we couldn't eliminate the defined symbol from "Maryis soluble"unless her presence in water was a consequenceof our theory. The superficialform of a reductionsentenceis the same as that of a conditionaldefinition.I thinkthis resemblancemisleadsmanywho rely on reductionsentences.But reductionsentencesand conditionaldefinitions are very different.In the case of conditionaldefinitions, do not we want an account of the newly introducedsymbol as applied to arguments not satisfying condition.In the case of the reductionsentence the for solubility,however,the sampleswhose presencein waterwe cannot deduce from the theory are preciselythose sampleswhose solubilitywe care about.For example,it would be ridiculousto provide a don't-care clause for the condition,addingthe followingto Example 11: "andif x is not put in waterthan x is a rotten apple."That this would be foolish is obvious from examinationof the inferentialcontexts in which one findsthe conceptof solubility. I infer that Carnap's terminology represents a serious blunder. Reduction sentences do not give partial "reductions"; unless, on analogy,one wishes to take "counting three"as "partially to countingto infinity." Still considering Definition as in Example 11 above, suppose the following oddity: "t is put in water"is a consequence of Theory,for absolutely every closed term t. This would still not ensure that Definitionabove satisfieseliminability; consider some context B = for
"for every x, x is soluble if and only if x . . . ," where the defined term is

used witha variableinsteadof a constant. Turningnow to the other criterion,the reductionsentenceDefinition of Example 11 is doubtless conservative.Suppose, however, that we enrichDefinitionby addinga conjunct. 12. EXAMPLE. Let Definition' be taken as a definition of solubility,where Definition' = (Definition& VxVy(ifx and

138

NUEL BELNAP

y are of the same naturalkind, then x is soluble if and only if y is soluble)). Adopting Definition' would doubtless enlarge the cases in which we could eliminate "soluble." But notice that we can now conclude that each thing of the same naturalkind as Sam, if put in water,dissolves if and only if Sam does. Unless our Theoryalreadycommittedus to this view, our definition would have smuggledin new assertionalcontent and would not be conservative.A healthy respect for conservation regardssuch sneakingas reprehensible. Thus again,reductionsentences are not much like definitions after all. They do not reduce, or not enoughto count. What, then, do reduction sentences do? It is perfectly all right to between the relations" suggestthat they describe or establish"meaning mysteryvocabularyto which "soluble" belongs and the less mysterious languageof water and the like. The only thing that is wrong is to think the mysterious to the less mysterious, even of them as "reducing" is partially. Onlythe terminology wrong. that early Carnap considered on his way to Here is an example reduction sentences, though I add a moral of which only the later wouldhaveapproved. Carnap
13.
EXAMPLE. (More solubility) Suppose we let

Definition = "forall x, x is soluble if and only if, if x is put in water,then x dissolves."

This definitionof solubilityis bound to satisfyboth criteria,regardless of theorems of previous definitions.Given our understanding that we are readingEnglishas if it were a standardfirst order language,it does, however, have the strange (but still conservative)consequence that whateveris not put in water is soluble, so that as stipulativeit is not a happy choice and as lexical it is doubtlessfalse to our usage. One can see here a motivefor an alternative theoryof "if".The first and best job of following out this insight is to be found in Bressan 1972. In rigorously defining concepts of serious physics, such as the concept of "mass" accordingto the intuitiveideas of Mach and Painlevee,Bressan uses modal concepts precisely because of the failure of extensional concepts to be apt. Carnap,in correspondencewith Bressan, strongly this encouraged work.

ON RIGOROUS DEFINITIONS 4. RULES OF DEFINITION

139

Here, briefly, are the standardrules for producingdefinitionsguaranThey teed to satisfy the criteriaof eliminabilityand conservativeness. are easy. And by Beth's DefinabilityTheorem, they are complete for Why is it, then, the first order languagetruthfunctionswith quantifiers. is that so much philosophyotherwisefaithfulto first-orderism carried out contraryto the policies they enjoin?Answer:Logic books, excepting Suppes 1957, do not give these matters proper discussion. The consequence is that students of philosophy,even those who are thoroughly taught to manipulatequantifiers,are not taught the difference between acceptable and unacceptabledefinitions. Since philosophers spend vastly more time proposing and using definitionsthan they do this quantifiers, is sad. manipulating rules are presentedas definitions.(Surelyit won't As promised,the occurto anyoneto thinkin this regardof circularity.) 14.
DEFINITION. (Standardrule for definingby an equivalence)

Definition is a standarddefinitionof a predicatesymbol R iff relativeto Theoryand Previous-definitions for some n

(the n-arity of R), and some variables vl, . ..
,

vw, and some

sentenceA (the "definiens"), are 1. Theory and Previous-definitions sets of sentences R (theoryand previousdefinitions,respectively), is an nDefinition is a ary relation symbol (to be defined), and definition). sentence(the candidate 2. Definition is an n-times universallyquantifiedbicondiV vn(Rv, **vn A). tionalV vl 3. The variablesv1,. . ., vnare distinct. variables, or 4. The definiensA has no "dangling" "floating" . . ., Vn. otherthanv1, thatis, no free variables 5. Each nonlogicalsymbolin A is drawnfrom amongthose of Theory Previous-definitions. and and 6. R is foreignto Theory to Previous-definitions. R It followsthatin particular, does not occurin A. 15.
DEFINITION. (Standardrule for definingan operatorby an

140

NUEL BELNAP

equivalence)Definitionis a standard definitionof an operator iff symbol 0 relativeto Theoryand Previous-definitions for some n (the n-arityof 0), and some variablesv, . . . V, w, and some sentenceA (the "definiens"), 1. Theory and Previous-definitions are sets of sentences (theoryand previousdefinitions,respectively),0 is an nary operator symbol (to be defined), and Definitionis a sentence(the candidate definition). 2. Definition is an (n + 1)-times universally quantified V biconditional v1... VvnVw((Ov1. . . vn= w) A). 3. The variablesv1,... ., vnand w are distinct. 4. The definiensA has no free variableother than vI,..., v, and w. 5. Each nonlogicalsymbolin A is drawnfrom amongthose of Theory Previous-definitions. and 6. 0 is foreignto Theory to Previous-definitions and 7. The sentence v1 ... vn3yVw[( y = w) A]is a consequence of Theory together with Previous-definitions. That is, it is a consequence of the theory together with previous definitionsthat, for each n-tuple of arguments
V ...
.,

v, there is in fact exactly one value w such that

A; so A is "functional." So 0 does not occurin A.

5. ADJUSTING THE RULES

Even withinthe purviewof the standardaccount,there are other forms of definition with an equal claim to propriety.In the context of the axiom of extensionalityin set theory, for instance, one may properly define set-to-set operators by way of membershipconditions. In the context of higher order axioms establishingan inductivelygenerated structure(such as Peano's axioms),one may properlydefine operators by rehearsingthe mode of inductive generation.Sometimes,in some In contexts, some of these are called "implicit definitions." these cases, means "insuch a way as to satisfythe criteriaof eliminability "properly" and conservativeness." much is a built-ingeneralityof the standard So account.

ON RIGOROUS DEFINITIONS

141

Suppose, however, the inductivematerialis added instead in a first order way, includinga license for new inductivedefinitions. People who wish to avoid relianceon higherorder logic sometimesproceed in this fashion. Expressionsso defined are likely to be inferentiallyenriching This inductivepractice and therebyviolate the criterionof eliminability. is so well understood,however, that it appearsunreasonableto call it Otherwise into question- unless perhapsthe questionis foundational. it maybe betterto weakenthe criteriaandbendthe rules. If the languageat issue is changedor extended,with the implication that eliminability now requiredfor additionalcontexts,then the rules is need adjustingin another sense; that is, following the rules may not suffice for satisfyingthe criteria.For instance,a universallyquantified biconditionalas in Definition 14 will evidentlynot permit elimination of a defined predicatefrom a modal context. Should, then, Definition have the form 0 Vx(Rx A), or should it instead have the form Vxfl (Rx A)? Or will both or neitherof these do? The fundamental point is this:We do not have to resort(much)to intuition.It is a matter of whatis required satisfythe criteria; that'sall. to Relevance logic offers an excellent example of the same phenomenon. It is an interestingexercise to work out how to adjustthe rules for defining new operators in Meyer's relevantarithmetic(see Anderson, Belnap and Dunn 1992). With one eye on the two criteria,you can see with the other preciselywhy and in what sense relevantarithmetic does and shouldprohibitthata divisionoperatorbe defined. A language with explicit tolerance for vagueness unquestionably needs its own matchingstory about definitions.Some of Tappenden's work on "pre-analytic" forthcoming statementsseems to me to count as in the requireddirection.
6. RELAXING THE CRITERIA (OR CHANGING THE RULES)

The standardcriteriaand rules must be counted as logic of the greatest interest. They are, I should say, always to be respected as devices to keep us from addle. On the other hand, one can easily find cases in which good thinkingbids us moderatetheir force. Here is an illustration concerning noncircularity. the Gupta(1988/1989) established following.

142

NUEL BELNAP

1.

2. 3.

4. 5.

Some concepts are essentially circular.(Normal results of are inductivedefinitions,e.g. multiplication, not examplesof circular concepts.) The standard account of definitions says nothing useful aboutcircular concepts(I supposeit deniestheirexistence). One obtains a powerful theory of circular concepts by reworkingthe theory of definitionsto admit circulardefinitions. Truth(in e.g.English)is a circular concept. The ideas of the reworkeddefinitionaltheory,when applied to truth, make fall into place both the ordinary and the extraordinary (pathological, paradoxical) phenomena to whichphilosophers havecalledattention.

These notions are extended and defended in The Revision Theoryof Truth,Guptaand Belnap 1992, which I will referenceas RTT. It is not possible to summarizeeven the main ideas here; I only make a few remarks the relating RTTworkto the standard accountof definitions. RTT fully accepts conservativeness. The norm remains: If it's a definition, then it should not provide camouflage for a substantive I assertion. shan'tsay anything moreon thistopic. RTT abandons eliminability,and in particularnoncircularity. The indispensableidea is that a definitionalscheme cannot illuminatea circularconcept except with the help of a circulardefinition,and that circularityintrinsically prevents eliminability. the standardaccount As is satisfiedwith only partialeliminability conditionaldefinitions,so for the RTT account is satisfiedwith only partialeliminability circular for definitions. For example, the RTT definition of truth permits its eliminationfrom "'Euthyphro pious'is true,"but not from "Thisvery is sentence is true."There is, however, a deep differencebetween how noncircularity relaxedin the presence of standardconditionaldefiniis tions and in the presence of RTT circulardefinitions.In the case of the former, there is an inferentialaccount of when eliminabilityis to be expected, and when not; the very form of a standard conditional definitionsuppliesmaterialfor such an account,as explainedabove. In contrast,on the RTT theory of circulardefinitions,eliminabilitymay depend on contingentfacts. There is no grammatical even inferential or

ON RIGOROUS DEFINITIONS

143

test to distinguisheliminable,innocuous circularityfrom ineliminable, vicious circularity. RTT also abandonsthe other support Furthermore, of eliminability, In namely,no inferentialenrichment. the case of truth, RTT only requiresthat a definitionfix for it an appropriate intensionor inferentialrole. It does not requirethat there alreadybe presentin the languagesomethingelse that occupies that role. RTT thinksof truthas an enrichmentof the languagein that sense, but as "definable" the in sense that its intension is fixed. Adding truth is therefore just like addingThe Absurd, as in Example 5, to a purelypositive language.In both cases the inferentialrole is uniquelydetermined,but unoccupied priorto the definition. RTT changes the grammarof definitions.Definitionsare not given explicitly as sentences, or even as theories, of the language.Instead, "implicit" definitionalideas are introduced.In one scheme, ratherthan requiringDefinition to be a sentence, as in Definition 14, a definition relates a definiendumto a definiens.The system of definistraightway tions is thought of as an implicit part of the languagein just the way that inferential connections are implicit. We, standing outside the language,may clearly describe a part of such a system by writinge.g. "Rx =Df A", just as we may write "A is a consequence of B." Other rulesremainin force,except,crucially, of Definition14. (6) 16. RTT ignores an explicit Tarski biconditional EXAMPLE. such as "'Euthyphro pious'is true Euthyphrois pious," is but it takes very seriously that there be a nonsymmetric definition relation between "'Euthyphrois pious' is true" and"Euthyphro pious." is In an exactlyparallelfashion,where it is assumedthat Liar "Liar is not true", RTT ignores the self-contradictory sentence "Liaris true Liar is not true",but it takes with great seriousness the nonsymmetric definitional relation between"Liaris true"and"Liaris not true".
-

RTT provides a semantic account of circulardefinitions.The Kernel idea is given by the word "revision." You startwith an extensionfor all but the various definienda,and you choose a hypothesis as to what their extensions might be. You now have a hypotheticalextension for

144

NUEL BELNAP

everythingin the language,and thereforefor every definiens.You use the definitionsto reviseyour hypothesis,arrivingat a new hypothetical extension for each definiendum.And so forth - a phrase here indicatingprogressinto the transfinite. The hope is that you can eventually arriveat categorical of resultsindependent your starting hypothesis. An especially fascinatingpart of this story is that RTT furnishes somethingworthwhileto say about circularconcepts such as trutheven in those cases in which eliminabilityfails. It does so by attendingto
their patterns of revision under different hypotheses. In this way forms

of pathologicality be distinguished from the other in a concepcan one tually satisfyingway. You can see the difference between "This very sentence is true" and "This very sentence is false" by watching the of unfolding theiraltogether different patternsof revision. RTT describesproof-theoretical techniquescorresponding closely as as mightbe hoped to the semanticsof revision,thus completingthe trio grammar/semantics/proof theory. Of particular interest is that the proof theoryreflectsa fundamental distinctionbetweenintroducing and a eliminating termby meansof a definition.
7. CODA

Two items remain.The first is to expressregretthat these remarkshave been too coarse to capture the fine structureof even the standard account of definitions.The second is to make explicit that the general lesson has the usual banalform. On the one hand, the standardcriteria and rules are marvelousguides. They have point. They are not to be taken lightly. Every philosopher and indeed I should say everyone engagedin conceptualwork oughtto know and understand them.He or she who unknowinglybreaks with the standardcriteria or the rules risks of folly. On the other hand, there is no single aspect of these criteriaor rulesthatwe shouldconsiderunalterable.
NOTE
1

Thankto A. GuptaandJ. Tappenden many-sided for help.

ON RIGOROUS DEFINITIONS REFERENCES

145

Anderson, A. R., Belnap, N. D. and Dunn, J. M. (1992) Entailment:The Logic of Vol. 2, Princeton, Relevanceand Necessity, PrincetonUniversityPress. Bochenski, J. (1956) Formale Logik, Freiburg,K. Alber. Translatedand edited by I. Thomas,A Historyof FormalLogic, second edition, New York, ChelseaPublishing Company,1970. Bressan, A. (1972) A GeneralInterpreted Modal Calculus,New Haven and London, Yale University Press. libraryof psychology, Carnap,P. (1937) The Logical Syntaxof Language,International philosophy and scientific method, New York, Harcourt, Brace. Translation of LogischeSyntaxderSpracheby A. S. Smeaton(Countessvon Zeppelin). Church, A. (1956) Introduction to MathematicalLogic, Princeton, NJ, Princeton UniversityPress. L'Enseignement mathematique 7, Couturat,L. (1905) "Lesdefinitionsmathematiques," 27-40. Inc. H. Curry, (1963) Foundationof Mathematical Logic,New York,McGraw-Hill, Frege, G. (1964) The Basic Laws of Arithmetic;Exposition of the System,Berkeley, Universityof CaliforniaPress. Translatedand edited, with an introduction,by M. Furth.The translationis of "the introductoryportions of the first volume and an epilogueappendedto the second." in Modal Gupta,A. (1980) The Logic of CommonNouns:An Investigation Quantified Logic,New Haven,Yale UniversityPress. on Gupta,A. (1988/1989) "Remarks Definitionsand the Concept of Truth,"Proceedingsof theAristotelian Society89, 227-246. MA, The Gupta,A. and Belnap,N. (1992) The Revision Theoryof Truth,Cambridge, MITPress.In press. Philosophical Horty,J. (1993) "Fregeon the PsychologicalSignificanceof Definitions," Studies72. 223-263 (this issue). Kneale, W. and Kneale, M (1962) The Developmentof Logic, Oxford,The Clarendon fromcorrectedsheetsof the firstedition, 1966. Press.Reprinted Kneebone, G. T. (1963) MathematicalLogic and the Foundationsof Mathematics, Company. Princeton,NJ,New York,Toronto,London,D. van NostrandPublishing Kripke, S. (1972) "Namingand Necessity."In: D. Davidson and G. Harman (eds.), Semantics of Natural Language,pp. 253-355. Dordrecht, D. Reidel Publishing Company. S. Legniewski, (1931), "Jber Definitionenin der sogenanntenTheorie der Deduktion," Comptesrendus des seances de la Societe des Science et des Lettresde Varsovie, ClasseeIII,24, 289-309. S. and Lesniewski,S. (1981) CollectedWorks. J. Surma,J. T. J. Srzednicki, D. I. Barnett vol. Company. (eds.),Syntheselibrary, 118, Boston,D. ReidelPublishing Luschei, E. C. (1962) The Logical Systemsof Lesniewski,Amsterdam,North-Holland Publishing Company. W. W. Marciszewski, (1981) "Definition." Marciszewski, (ed.), Dictionaryof Logic, In: pp. 89-96. The Hague,Martinus Nijhoff. Synthese13, 303-322. Marcus,R. B. (1961) "Modalities Intensional and Languages," Press. Robinson,R. (1950) Definition,Oxford,The Clarendon Adams. Runes,D. D. (1962) Dictionary Philosophy, of Totowa,NJ, Littlefield, Suppes, P. (1957) Introductionto Logic, PrincetonNJ, New York, Toronto, London, D. vanNostrandPublishing Company. in Tarski, A. (1935) "Der Wahrheitsbegriff den formalisiertenSprachen,"Studia Philosophica 1, 261-405. Germantranslationof a book in Polish, 1933. English

146

NUEL BELNAP

translationby J. H. Woodger,"TheConcept of Truthin FormalizedLanguage," in Logic, Semantics,Metamathetics, papers from 1923 to 1938 by A. Tarski,Oxford, OxfordUniversityPress, 1956, pp. 152-278. Tarski,A. (1941) Introduction Logic and to the Methodology DeductiveSciences. to of Enlargedand revised edition, New York, Oxford UniversityPress. Translatedby Olaf Helmer from a work that appeared in Polish in 1936 and the in German in translation 1937. Tarski,A. (1944) "The SemanticConception of Truth,"Philosophyand PhenomenologicalResearch4, 341-375. Tarski, A. (1956) Logic, Semantics, Metamathematics, London and Oxford, The Clarendon Press.

Department Philosophy of University Pittsburgh of Pittsburgh, 15260 PA USA