Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Ian Rumfitt
This book has been a long time in the making. Its topichow to adjudicate
between rival logical systemshas intrigued me since Iwas an undergraduate,
but it was only around the turn of the millennium that Ibegan to glimpse the
approach to the subject that Ifollow here. In 2001, Ihad the good fortune to be
awarded a Philip Leverhulme Prize which, by relieving me for two years of my
tutorial duties as a Fellow of University College, Oxford, enabled me to start seri-
ous work on the project; by the summer of 2003 Ihad early versions of the first five
chapters. Since it is now more than ten years since their munificence came to an
end, the Leverhulme Trustees may well have forgotten that they ever awarded me
the Prize, but Iremain deeply grateful for it. This book would certainly not exist
without it, and Ihope the Trustees will regard the present volume as a case of sero
sed serio in lucem editum.
Drafts of those chapters, and later of the rest of the book, formed the bases
of seminars that I gave between 2003 and 2011 in the Universities of Oxford
and of London, where Ihad moved in 2005 to take up the post of Professor of
Philosophy at Birkbeck College. It was only during a sabbatical year in 201112,
however, that Ifound the time to make the revisions that the seminar discus-
sions had shown were needed. Iam most grateful to the College for granting me
this period of leave, as Iam to the three institutions which were kind enough
to offer me hospitality during that year. In the autumn of 2011, Ispent a week
in Munich presenting some of the material to the members of the Centre for
Mathematical Philosophy at the Ludwig-Maximilian University. In the spring
of 2012, Iaired parts of Chapters6 and 7 in a seminar at NewYork University,
where Paul Horwich was an especially generous commentator. Finally, Ihad the
pleasure of spending the Trinity Term of 2012 back in Oxford as a Visiting Fellow
of All Souls College, there enjoying conditions ideal for serious study.
Large as my institutional debts are, they are dwarfed by the personal obliga-
tions Ihave incurred. The greatest are to my familymy wife, Lucy, and our
children, Sebastian and Isabelwho have had to cope with a somewhat dis-
tracted husband and father for some time. I am also indebted to the many
philosophers who have helped me with comments on, criticisms of, and sug-
gestions about the project over the years, often in discussion following talks
Igave at their universities or at conferences. Iam uncomfortably aware that a
list is unlikely to be complete, but Ican remember receiving help of this kind
viii Preface
from Arif Ahmed, Mark van Atten, Thomas Baldwin, George Bealer, J.C. Beall,
Nuel Belnap, Simon Blackburn, Susanne Bobzien, Paul Boghossian, Alexis
Burgess, John Campbell, David Charles, Bill Child, Justin Clarke-Doane, Mark
Crimmins, Neil Dawson, Imogen Dickie, John Divers, Thomas Donaldson,
Cian Dorr, Dorothy Edgington, Stephen Everson, Solomon Feferman, Kit
Fine, Salvatore Florio, Ken Gemes, Marcus Giaquinto, Volker Halbach, Sally
Haslanger, David Hills, Ole Hjortland, Richard Holton, Jennifer Hornsby,
Paul Horwich, Keith Hossack, Luca Incurvati, Daniel Isaacson, Nicholas
Jones, Rosanna Keefe, Jeffrey Ketland, Arnold Koslow, Wolfgang Knne, Rae
Langton, Krista Lawlor, Hannes Leitgeb, ystein Linnebo, Vann McGee,
Edwin Mares, Per Martin-Lf, Hugh Mellor, Peter Milne, the late Grigori
Mints, Benjamin Morison, Julien Murzi, Thomas Nagel, David Papineau, Jan
von Plato, Graham Priest, Agustn Rayo, Greg Restall, Daniel Rothschild, Mark
Sainsbury, Christopher Scambler, Stephen Schiffer, the late Stephen Schulz,
Stewart Shapiro, Peter Simons, Timothy Smiley, Barry Smith, Robert Stalnaker,
Jason Stanley, Florian Steinberger, the late Sir Peter Strawson, Scott Sturgeon,
Gran Sundholm, Zoltn Gendler Szab, Mark Textor, Gabriel Uzquiano,
Mark Vandevelde, David Velleman, James Walmsley, Sean Walsh, Lee Walters,
Ralph Wedgwood, Alan Weir, Bruno Whittle, Timothy Williamson, Mark
Wilson, Stephen Yablo, and Elia Zardini. Thanks to one and all.
Others need special mention. Jonathan Barnes has been reading my work
since Iwas an undergraduate; it continues to profit from his quizzical gaze. Bob
Hale and I found that our approaches to modality shared many features; my
most pleasant memories of the whole project are of the days we spent compar-
ing our ideas while Ienjoyed his and Maggies hospitality at their Glasgow home.
In London, Christopher Peacocke and David Wiggins have been generous in
providing detailed comments on all the chapters, sometimes in more than one
draft. In December 2009, Crispin Wright helped sustain the project by inviting
me to spend a week discussing draft chapters with him and his colleagues at the
Northern Institute of Philosophy in Aberdeen. Peter Momtchiloff has been a
marvellously forbearing editor, patiently encouraging me but never harrying. I
am also most grateful to two anonymous readers for the Press, whose comments
were a finely judged mixture of criticism, suggestions, and encouragement. I
warmly thank Christopher Scambler for his help with the index.
My greatest debt, though, is to someone who is no longer with us. It was
hearing Michael Dummett lecture on this topic when I was an undergradu-
ate that first got me interested in it. Inow realize that Imust have understood
even less of his lectures than Ithought at the time, but my memories of those
occasions, in which Dummett allowed an audience to see him wrestling with
Preface ix
Earlier versions of parts of this book have already appeared in some of my pub-
lished essays. Parts of 2.2, 2.3, 2.5, and 2.6 may be found in Inference, deduction,
logic, in John Bengson and Marc Moffett, eds., Knowing How (NewYork:Oxford
University Press, 2011), pp.33459. 3.2, 3.4, and 3.5 overlap with Logical neces-
sity, in Bob Hale and Aviv Hoffmann, eds., Modality:Metaphysics, Logic, and
Epistemology (Oxford: Clarendon Press, 2010), pp. 3564, although (as n.17 to
Chapter3 explains) my views have evolved considerably since Ipublished that
essay. 4.14.3 and 4.6 descend from Asserting and excluding: steps towards
an anti-realist account of classical consequence, in Randall E. Auxier and
Lewis Edwin Hahn, eds., The Library of Living Philosophers Volume XXXI:The
Philosophy of Michael Dummett (Chicago: Open Court, 2007), pp. 63993.
4.44.5 draw upon Ramsey on truth and meaning, in Benjamin Morison
and Katerina Ierodiakonou, eds., Episteme, etc.:Essays in Honour of Jonathan
Barnes (Oxford:Clarendon Press, 2011), pp.21345. 5.2 contains material from
Sense and evidence, The Monist vol. 96, no.2 (2013), pp.177204. Early versions
of 6.16.4 and 7.3 formed part of my inaugural lecture at London, which was
delivered on 25 June 2008 and published as a pamphlet by Birkbeck College later
that year. 5.5 and 7.2 descend from On a neglected path to intuitionism, Topoi
vol. 31, no.1 (2012), pp.1019. 7.3 incorporates a few paragraphs from Brouwer
versus Wittgenstein on the infinite and the law of excluded middle, Grazer
Philosophische Studien 89 (2014), pp.95110.
I am grateful to the editors and publishers listed above for permission to re-use
material. Ishould say, though, that the process of weaving the various strands
of argument into a book has led me to revise, often radically, the material from
which it derives. Iapologize to readers of my previous publications if they have
less of a head start in getting to grips with this work than they think they ought
to have.
Contents
1. Introduction 1
1.1 Disputes Over Logical Laws 1
1.2 The Scope of This Book, and the Nature of Disputes Over
Basic Logical Laws 14
1.3 The Argument of the Present Book 21
References 321
Index 337
So findet sich auch, wenn die Erkenntnis gleichsam durch ein
Unendliches gegangen ist, die Grazie wieder ein.
Heinrich von Kleist, ber das Marionettentheater
1
Introduction
Thus wrote Gottlob Frege in the foreword to his opus magnum, Grundgesetze der
Arithmetik. Freges laws of truth are laws of logic, and the logical laws that he
took to constitute these immovable boundary stones are the principles of what
we would now call full, second-order classical logic. In the century or more since
Grundgesetze appeared, however, a number of philosophers have tried to displace
classical logic from the status that Frege accorded to it. My chief aim in this book
is to scrutinize five of the most philosophically interesting attempts to do this and
assess whether they succeed.
Frege himself would have had little time for this enterprise. Just after
the passage Ihave quoted, he asks what we should say if beings were found
whose laws of thought directly contradicted ours and thus led frequently
to contrary results even in practice. The psychological logician could only
simply acknowledge the fact and say:those laws hold for them, these laws
hold for us. Ishould say:we have here a hitherto unknown form of madness
(Frege 1893, xvi). Frege is right to say that we cannot follow the psychologi-
cal logicians in resting content with an insouciant acknowledgement of the
fact that different people conform to different logical laws. At least, no ana-
lytic philosopher can rest content with that. Philosophical understanding
does not advance in the way that Descartes and Spinoza hoped that it would,
namely, by deducing theorems from self-evident axioms and definitions.
2 Introduction
Yet scarcely a day goes by when the working philosopher does not need to
judge whether a putative consequence of some claims or hypotheses follows
logically from them. Since logic provides the standards for making those
judgements, differences over logical laws impede the assessment of philo-
sophical arguments. On the other hand, Frege is surely wrong to describe
anyone who deviates from the logic he set forth as insane. One class of coun-
terexamples (among many) is provided by intuitionist logicians. Although
intuitionistic propositional and first-order logics are sub-systems of the cor-
responding classical systems, intuitionistic second-order logic affirms the
negations of some classical theorems; so intuitionists directly contradict
some classical laws of thought. Moreover, the intuitionists deviations even
from first-order classical logic have led to contrary results in mathematical
practice. However, not all intuitionists have been mad. Indeedto make a
claim that they will regard as stronger than that made in the last sentence
some intuitionists have not been mad. On the contrary, some of them have
advanced interesting arguments for preferring their logic to the classical sys-
tem that Frege propounded. Those arguments deserve to be heard; they are
not to be dismissed summarily on the basis of an ungrounded imputation of
insanity.
It is, all the same, far from obvious how a rational discussion between
adherents of rival logical schools could proceed. The basic difficulty is that
attempts to justify basic logical laws are liable to exhibit a particular, and
apparently vicious, form of circularity. Michael Dummett has well explained
the problem. He supposesa supposition that the next chapter will vindi-
catethat a logic is best formalized as a system of rules of inference and rules
of proof. Now in attempting to show that a primitive logical rule is sound, he
observes,
we should be bound to employ deductive argument; and, in doing so, we should
probably make use either of those very forms of inference which we were supposed
to be justifying, or else of ones which we had already justified by reduction to our
primitive rules. And, even if we did neither of those things, so that our proof was not
strictly speaking circular, we should have used some principles of inference or other,
and the question could be raised what justified them: we should therefore either
eventually be involved in circularity, or have embarked upon an infinite regress.
(Dummett 1973, 2912)
In a more recent idiom, proofs of the soundness of primitive logical rules are lia-
ble to be rule-circular:the proof of Rs soundness involves an application (in the
metalogic) of rule R itself.
Introduction 3
Let us suppose for a moment that the sense of a statementby which Imean
its logically relevant contentconsists in its truth-conditions. Let us also sup-
pose that the sense of a wordthat is, the contribution the word makes to the
logically relevant content of a statement containing itis given by an axiom in
a homophonic or modest truth theory. Then it is certainly difficult to see how a
proof of the soundness of ones preferred rules could be other than rule-circular.
Given our two suppositions, the soundness of a rule of inference will consist in
this:that whenever the premisses of an application of the rule are true, its con-
clusion will also be true. And the sense of (for example) the connective it is not
the case that (for short, not) will be given by the principle (N):
(N)For any statement A, not A is true if and only if it is not the case that A
is true.
Now consider how one might prove the soundness of Double Negation
Elimination:
(DNE)From not not A , infer A.
This is a rule of inference which the classicist accepts without restriction but
which the intuitionist accepts only for decidable A. The natural attempt at a proof
of the rules soundness runs as follows:
(1) not not A is true supposition
(2) not not A is true only if it is not the case consequence of (N)
that not A is true
(3) It is not the case that not A is true (1), (2), modus ponens
(4) not A is true if it is not the case that A consequence of (N)
is true
(5) It is not the case that not A is true only (4), contraposition
if it is not the case that it is not the case
that A is true
(6) It is not the case that it is not the case that (3), (5), modus ponens
A is true
(7) A is true
(8) If not not A is true, then A is true (1), (7), conditional proof,
discharging supposition (1)
(9) For any statement A, if not not A is (8), universal generalization
true, then A is true
The conclusion of this derivation expresses the soundness of Double Negation
Elimination. The problematic step, though, is that from (6)to (7). The step is an instance
of Double Negation Elimination, the very rule whose soundness is to be proved.
4 Introduction
1
Boghossian (2000) argues that it can, Dogramaci (2010) that it cannot.
Introduction 5
2
For these explanations of harmony and stability, see Tennant 1978, 745.
3
For suppose that A and B jointly yield a contradiction. Then, by the introduction rule for not,
B entails not A . So, given the introduction rule, not A is a maximally weak statement that can
stand as major premiss of the elimination rule.
6 Introduction
In other writings, Ihave analysed in some detail this approach to the problem
of deciding between rival logical systems, so Ihope Imay be forgiven for say-
ing only briefly here why Ido not regard it as promising. In the first place, what
harmony and stability require is highly sensitive to the way in which the logic is
formalized. Dummett assumes that we shall wish to formalize logic in the usual
affirmative style, in which each line in a deduction is either asserted or is pre-
sented as the consequence of affirmative hypothesesthat is, of suppositions to
the effect that certain questions are correctly answered yes. In such a system, the
usual classical rules for negation are unstable, but we are not obliged to formalize
logic in that way. We may instead adopt a bilateral formalization, in which some
lines in deductions are denied, or are presented as the consequences of negative
hypothesesthat is, of suppositions to the effect that certain question are cor-
rectly answered no. If we adopt bilateralism, it is the classical rules for negation
that exhibit the correlates of harmony and stability. At best, then, considerations
of harmony and stability push the issue back: the choice of logic depends on
4
which style of formalization is preferred.
Second, and more importantly, the philosophical arguments for requiring
that a connectives introduction and elimination rules must be harmonious and
stable are weak. The main argument for imposing these requirementsand the
5
argument on which Dummett chiefly reliesis due to Dag Prawitz. According
to Prawitz, the meaning of any statement is given by certain canonical con-
ditions for affirming it, conditions that we may call the statements warrants.
Prawitz further supposes that the introduction rule for a connective will specify
warrants for those complex statements whose main operator is the connective
in question. Thus, given warrants U and V for the statements A and B, the war-
rant for the conjunction A and B will be a composite, W, comprising U, V, and
an application of and-introduction, the whole of which may be represented
asbelow:
U V
A
_______________B
A and B.
The very existence of deductive argument presents a prima facie problem for this
theory of meaning for, even when the conclusion of a deduction is a complex
4
For this objection to Dummett, see Rumfitt 2000. The debate continued in Dummett 2002,
Gibbard 2002, Rumfitt 2002, Ferreira 2008, Rumfitt 2008b, and Dickie 2010. See also Price 1983,
1990, Smiley 1996, and Humberstone 2000.
5
See especially Prawtiz 1974 and Dummett 1991, chaps. 9 and 11.
Introduction 7
statement, its last step need not be an application of the introduction rule for the
statements main connective. That is, a deduction may be indirect. So, for exam-
ple, the conjunction A and B may be deduced, not from the premisses A and B,
but from C together with If C, then A and B . Since the meaning of the statement
A and B is supposed to be given by the complex warrant displayed above, there
is then a serious question whether the indirect deduction is faithful to the mean-
ing of the statement whose truth it purports to establish. Prawitz and (following
him) Dummett argue that an indirect deduction may be faithful to the meanings
of the statements in it so long as it can be shown that, whenever we have a valid
indirect deduction, warrants for all the premisses may be transformed so as to
form a warrant for the conclusion. They further claim that harmony and stability
between the introduction and elimination rules for the connectives are needed to
ensure that this requirement for the validity of indirect argument is met.
Even granted all the assumptions of this argument, it is doubtful whether the
eventual conclusion follows. The requirement that warrants for the premisses of
a valid deductive argument should always be capable of being transformed into
warrants for the arguments conclusion is vague, but it would seem to be met if the
logic in question admits of strong normalization. That is to say, the requirement
would seem to be satisfied if there is an effective method for converting any valid
deduction into one in normal formthat is, into a deduction in which the elim-
ination rules for the connectives are first applied to break down any complex
premisses, and in which the introduction rules are then applied to the products
of this process of breaking down so as to construct the eventual conclusion. Yet
strong normalization theorems have been proved for certain formalizations of
6
full classical logic, even though the classical rules for negation are, in Dummetts
sense, unstable.
For present purposes, though, we need not delve into these proof-theoretic
niceties, for the meaning-theoretic premisses of Prawitzs argument are in any
case implausible. In particular, there is no reason to suppose that the meaning
of every statement is given by certain canonical conditions for affirming it. This
premiss has some plausibility for elementary mathematical statements. One
might hold that a speaker does not fully grasp the meaning of a simple sum such
as 36 + 47=83 unless he knows that it can be established by performing a certain
6
Some of these formalizations are rather artificial, as in Stlmarck 1991 which proves nor-
malization for a classical system in which the elimination rules for disjunction and the existen-
tial quantifier are restricted to atomic conclusions. However, von Plato and Siders (2012) prove a
normalization theorem for a formalization of classical logic that has as good a claim as any to be
natural. For relevant background, see also Statman 1974, Seldin 1986, and Negri and von Plato 2001,
especially 8.6.
8 Introduction
elementary calculation, which conforms to the rules of adding within the decimal
system of numerals. On this basis, one might go on to claim that the successful
execution of that calculation qualifies as a canonical condition for asserting 36 +
47=83. Even within mathematics, though, the claim that a statements meaning
is given by some canonical method of verifying it is plausible only for those parts
of the subject (such as the elementary theory of addition) which are already well
systematized and for which there are, in consequence, generally accepted basic
criteria for affirming the relevant atomic statements. In those parts of mathemat-
ics which still await a generally accepted systematization, by contrast, it may be
quite unclear which conditions for affirming a statement are criterial for under-
standing it. In these cases, it would be at best premature to claim that certain
grounds for affirming the statement have canonical status.
A more important problem with the theory of meaning on which Prawitzs
argument for requiring harmony and stability rests is that it does not extend to
empirical discourse, for in the case of many empirical statements the distinction
between canonical and non-canonical justifications lacks clear application. The
statement There are Higgs bosons surely has a sense:most physicists understand
it, and many of them are capable of evaluating evidence for and against its truth.
All the same, it makes little sense to divide the possible evidence for the state-
ment into canonical and non-canonical grounds. For one thing, what evidence
confirms, or disconfirms, is a whole theorythe so-called standard model of
subatomic interactive forcesof which the statement There are Higgs bosons
is but one salient part. Moreover, the status of data as confirming or disconfirm-
ing that theory will itself depend on many background assumptions and pos-
itsabout the way particle accelerators work, for example. Dummett and Prawitz
areor wish to bemolecularists about meaning. That is, they suppose that a
statements meaning is determined by its mode of composition and the mean-
ing of its parts. However, cases such as this bring out the difficulty of combining
molecularism and verificationism givenwhat is widely acceptedthe holism of
empirical confirmation. (See Chapter5 for a fuller consideration of the problems
confronting the theory of meaning on which Prawitzs case for requiring har-
mony and stability rests.)
Unless we appeal to harmony considerations, though, are not attempts to adju-
dicate between rival logical systems doomed to futile circularity? Not necessarily.
Many philosophers believe that the best we can do is to apply rather general con-
siderations, familiar from elsewhere in science, of simplicity and strength, per-
haps in tandem with a pragmatic assessment of relative costs and benefits. Amain
object of this book, though, is to show how we can do better. The proof given ear-
lier of the soundness of Double Negation Elimination was futile for our purposes,
Introduction 9
but it was futile at least partly because it took the sense of the negation-sign to be
given by the homophonic truth-theoretic principle (N). One tug on the string of
(N), and the apparent knot involving true in the condition for soundness unrav-
els, taking one straight back to the acceptability or otherwise of the contested
rule. Even if (N) is true, its truth, and the truth of homophonic principles like it,
does not preclude there being another sort of semantic theory which can provide
greater traction on disputes about basic logical laws, and perhaps some pointers
to their resolution.
This, indeed, is the approach to the problem of rationally adjudicating between
rival logical systems that Ishall be exploring. Here, too, Dummett pointed the
way. In an essay written in the 1980s, heasked:
How can the classical logician and the non-standard logician come to understand one
another? Not, obviously, by defining the logical constants. They have to give a semantic
theory; and they need one as stable as possible under changes in the underlying logic of
the metalanguage. The homophonic semantics ... is as unstable as possible:the laws
that can be proved to be valid in the object-language will be precisely those that are taken
to govern the metalanguage. But if the intuitionist gives a semantics in terms of Beth
trees, or the quantum logician one in terms of ortholattices, it will, at least for sentential
logic, be stable under the substitution of classical for intuitionistic or quantum logic as
that governing the metalanguage:assuming the law of excluded middle or the distribu-
tive law for the metalanguage will not affect the ability of the semantic theory to yield a
demonstration of their invalidity for the object-language. The classical logician may thus
continue to reason as he has always done, without impairing his capacity to understand
why the non-classical logicians take these laws to fail; for he has been given an account of
the meanings they attach to the logical constants in a manner that does not presuppose
that he already understands them. (Dummett 1987, 254)
would not be rivals. However, the relevant consequences can be the same. In a
discussion about the validity of certain contested logical rules, the relevant con-
sequences will comprise all those statements saying that this-or-that contested
rule is sound or is unsound. Any discussion about anything requires some com-
mon ground. The question of the choice of logic is not begged if statements that
such-and-such a semantic theory validates, or invalidates, a contested logical rule
are part of that common ground.
How, in more detail, might the use of non-homophonic semantic theories
advance a rational discussion of which logic to go by? As the present book will
show, there are many ways in which this can happen. However, the passage just
quoted from Dummett already suggests some simple ways which are worth intro-
ducing now:
(1) Let us suppose that a thinker adheres to the rules of classical logic when
he goes in for deductive reasoning. Let us also suppose that we give him rea-
sons for thinking that the senses of the connectives and quantifiers (at least
as they are used in a given area of discourse) are faithfully specified by the
Beth-tree semantics, and that Beth also gave a correct account of logical con-
7
sequence. Let us suppose, finally, that our thinker is persuaded by those
reasons. Then he may be rationally persuaded to revise his adherence to the
classical rules. As Dummett remarks, we can deploy the premisses that he
has been brought to accept in a classically valid argument to show that cer-
tain classical rules are not valid. In particular, assuming the law of excluded
middle ... for the metalanguage will not affect the ability of the semantic
theory to yield a demonstration of [its] invalidity for the object-language.
Reasons for preferring Beths account of the senses of the logical words, and
his account of logical consequence, can, then, be parlayed into a non-circular
reason to depart from full adherence to classical logic. It is worth not-
ing that reasons for preferring Beths account of the senses of the connec-
tivesor some other semantic theory that yields a non-classical logic, such as
8
Heytings need not be reasons for supposing that the classical understand-
ing of the connectives is, in Dummetts word, spurious. The semantic case
for an alternative to classical logic does not require that the classical use of the
connectives is unintelligible.
(2) Rational persuasion may work in the opposite direction too. Let us call
classical semantics (for the propositional fragment of a natural language) that
7
For Beth trees, see Beth 1959, 145, Dummett 2000, 5.4 and 7.4, and 6.4 below.
8
For details of this semantic theory, see 5.2.
Introduction 11
semantic theory whose substantial axioms are: (1) the principle that a valid
argument preserves truth; (2)the principle that no statement is both true and
false (the Principle of Non-Contradiction); (3)the principle that each statement
is either true or false (the Principle of Bivalence); and (4)the standard axioms
(recorded in the familiar truth-tables) that relate the truth or falsity of complex
statements in the language to the truth or falsity of their components. In this
semantic theory, the axiom for negation will not be the homophonic (N), but the
distinct claim(M):
(M) not A is true if and only if A is false and is false if and only if A is true.
From these principles, we may give an intuitionistically valid proof of the sound-
ness of Double Negation Elimination as follows:
(1) not not A is true supposition
(2) not not A is true only if not A is false consequence of (M)
(3) not A is false (1), (2), modus ponens
(4) A is false supposition
(5) not A is true if A is false consequence of (M)
(6) not A is true (4), (5), modus ponens
(7) not A is true and not A is false (3), (6), and-introduction
(8) It is not the case that ( not A is true and instance of
not A is false) Non-Contradiction
(9) A is not false (4), (7), (8), reductio,
discharging supposition (4)
(10) Either A is true or A is false instance of Bivalence
(11) A is true (9), (10), disjunctive syllogism
(12) If not not A is true, then A is true (1), (11), conditional proof,
discharging supposition (1)
(13) For any statement A, if not not A is (12), universal generalization
true, then A is true
Unlike the derivation that relied on (N), this argument for the soundness of
Double Negation Elimination is not rule-circular: at no stage in it is the rule
whose soundness it purports to establish applied. Since it is an intuitionisti-
cally correct argument, it may be used rationally to persuade an intuitionist who
comes to accept classical semanticsincluding the unrestricted Principle of
Bivalencealso to accept full classical logic.
(3) Whether a semantic theory is stable (in this new sense) under changes
in the metalogic depends on what those changes are:a semantic theory that
is stable under the change from classical to intuitionistic logic may not be
stable under a change from classical logic to quantum logic, for example.
Indeed, classical semantics is demonstrably not stable under the latter
12 Introduction
9
See also the mention of Robert Goldblatts recent cover semantics for the relevance logic R in
Chapter7, n.3.
14 Introduction
10
See Church 1956, 10910. Churchs proof of Post-completeness assumes the soundness of
Modus Ponens, and there are in fact non-standard formalizations of the theorems of the classical
propositional calculus which lack Modus Ponens as either a primitive or a derived rule and which
consequently are not Post-complete; see e.g. Hi 1959. In Chapter2, though, Iargue that a logical
system is constituted by its rules of inference and rules of proof, not by its theorems alone; on this
view, His system is not a formalization of classical logic.
11
Consider the two formulae xPx yPy and xPx yPy. The first is not intuitionisti-
cally valid, while the second is. Their Skolemized forms, however, are the same.
Introduction 15
Striking as these features of the classical formal systems may be, another sort
of ground for according default status to the classical rules is in the end more
important. God has not been so sparing to Men to make them barely two-legged
creatures, and left it to Aristotle to make them Rational ... He has given them a
12
Mind that can reason without being instructed in Methods of Syllogizing. He
has also, we might add, given us minds that can reason without being instructed
in the non-syllogistic methods that modern logic textbooks expound. No doubt
a child needs to find itself in the right sort of environment for its natural capac-
ity for deductive reasoning to develop. Locke is right to maintain, though, that it
can develop that capacity to a high degree without the help of logic books. This
means that, when we first encounter such books, we can test the logical systems
we find in them against our pre-theoretical but still developed sense of what fol-
lows from what.
This, it seems to me, provides the strongest reason for according default status
to classical logic, forwith only one class of exceptionsclassically valid argu-
ments conform to our intuitive sense of deductions whose conclusions follow
from their premisses. Setting aside the exceptions, the classical natural deduction
rules seem, when we first meet them, to codify norms of deductive reasoning that
we have implicitly followed for years. Thus the tyro logician accepts the Law of
13
Dilemma as codifying the standards he has already been using when arguing
14
by cases. Similarly, he accepts Simple Reductio as codifying a ground he has
long relied upon for asserting negated statements. Of course, the classical rules
have consequencessuch as the theorem that any statement whatever follows
from a contradictionthat are initially puzzling. But the reason why these conse-
quences are puzzling is precisely that we can be led to them by way of a sequence
of argumentative steps, each of which seems to be entirely compelling. So these
consequences do not by themselves cast doubt on the proposal to accord default
status to classical logic.
The exceptional areathe area where the conclusions of some classically cor-
rect sequents intuitively seem not to follow from their premissesis the logic of
the conditional. One classically valid argument schema is:(A B); therefore
A. Few people, though, have any pre-theoretical inclination to think that There
is a god follows from It is not the case that if there is a god then the prayers of
the wicked will be answered (Grices example). Another classically valid schema
12
Locke, An Essay Concerning Human Understanding, Book IV, chapter xvii, 4.
13
The law, Imean, that if a conclusion follows from each of two statements severally it also fol-
lows from their disjunction.
14
As above, this is the law that premisses with which a statement is inconsistent entail that state-
ments negation.
16 Introduction
15
See Stevenson 1970 and Priest 2001, 13.
16
This example has a curious history. In conversation with Dummett, Elizabeth Anscombe pre-
sented the case of If Henry likes any fish, then it is true that he likes any fish as a counterexample to
the minimalist or redundancy theory of truth (see Dummett 2004, 20). The case is a counterex-
ample to the claim that absolutely any instance of the English schema If..., then it is true that... is
itself true on its most natural reading. However, Ido not think that this observation should disturb
a redundancy theorist. As we have seen, when a logician asserts the validity of a schema in a formal
language, he is not committed to the truth of all the instances of any natural-language schema; nei-
ther, Ithink, need a redundancy theorist of truth be so committed. What a redundancy theorist is
committed to is the truth of any biconditional, one of whose limbs is a statement that ascribes truth
to the statement that constitutes the other limb. Because of the scope rule for any, It is true that
Henry likes any fish if and only if Henry likes any fish does not qualify as such a biconditional. So
Anscombes example poses no problem for a properly formulated redundancy theory.
Introduction 17
17
For my analysis of the counterexample to Modus Ponens put forward by Vann McGee (1985),
see Rumfitt 2013, 11.
18
Ido not mean to imply that the right approach to finding the logic of the natural-language
conditional is simply to seek an appropriate restriction on the rule of Conditional Proof. As David
Lewis (1975) and William Lycan (2001) have in different ways suggested, it may be a mistake to treat
the vernacular conditional as a dyadic sentential connective at all. If they are right, then the logic of
the conditional will differ even more radically from the classical model.
18 Introduction
Freges own logical achievement rested on an innovation of just this kind. In ana-
lysing statements in terms of function and argument, rather than the traditional
logical categories of subject and predicate, he was precisely extending the applica-
tion of those notions beyond their customary use in mathematics with a view to
19
achieving a compendious formulation of logical laws.
Subsequent logicians have in the main followed him in this style of analysis,
although a rearguard, led by Fred Sommers, continues to champion the tradi-
tional categories that Frege had rejected as importing psychological and linguis-
tic impurities into logic (see Sommers 1982). Ihave little sympathy with the old
believers and will not discuss their views here; but even those of us who are con-
fident that a return to traditional logic would be a large step backwards should
not be unshakably attached to Freges way of analysing statements. From a logical
point of view, analysing statements in terms of function and argument is a great
improvement on the traditional theory, but we cannot exclude the possibility of
achieving a yet more powerful formulation of logical laws using a quite different
set of fundamental notions.
At any rate, we need to bear in mind the difference between a codification of
logical laws and the evaluations of deductive argument, that those laws codify.
Sommerss objections to Freges logical laws are conservative:they challenge the
claim that those laws comprise the best way of codifying our ordinary standards
of deductive validity, but they do not challenge the standards themselves. By con-
trast, the arguments Ishall be considering in this book attack classical logic in
this second, more radical way. They typically concede that the classical laws for
19
Thus the fragment What may Iregard as the result of my work?, in which Frege took stock of
his achievement in August 1906, begins:It is almost all tied up with the Begriffsschrift. Aconcept
conceived as a function. Arelation as a function of two arguments (Frege 1906=Frege 1969, 200/
Frege 1979, 184).
Introduction 19
because it is impossible to deny either that the meanings of the logical constants deter-
mine the manner in which the truth of a complex sentence depends on its constituents,
or that the validity of a form of argument depends on whether it is so constructed that
the truth of the premisses guarantees the truth of the conclusion. Hence, if we come to
view as invalid a form of argument we had formerly considered valid, although there was
no mistake that could have been pointed out by appeal to existing linguistic practice,
we must have changed the way in which we take the truth-values of the premisses and
conclusions to be determined in accordance with their structure; and this entails that we
have changed the meanings of the logical constants. (Dummett 1991, 3023)
Powerful as this argument may seem, it contains a subtle flaw which the analysis
below will expose (see 7.5).
I have been writing of consequence as a relation among statements; it will
help to be more explicit about my use of this term. Let us consider those ordered
pairs whose first element is a meaningful, indeed disambiguated, declarative
type-sentence, and whose second element is a possible context of utterance; by
a possible context of utterance, Imean a determination of all the contextual fea-
tures which can bear upon the truth or falsity of a declarative utterance. Some of
these ordered pairs will be such that, were the declarative type-sentence that is
the first element uttered in the context that is the second element, a single com-
plete thought would then be expressed:the resulting utterance would say that
such-and-such is the case. As Ishall use the term, a statement is an ordered pair
that meets this condition. Not every ordered pair of declarative type-sentence
and possible context of utterance will qualify as a statement in this sense. For
example, an ordered pair of sentence and context whose first member is You are
ill will not count as a statement unless the context supplies an addressee. On this
way of using the term, each statement belongs to a language, namely, the language
of the sentence that is its first element. Furthermore, each statement possesses a
sense or propositional content:this is what would be expressed by uttering the
statements first element (the declarative sentence) in the context that comprises
its second element. It then makes sense to classify a statement as true or false
simpliciter. When a statement expresses the thought that such-and-such is the
Introduction 21
caseor more briefly, says that such-and-suchthe statement is true if and only
if such-and-such really is the case, and false if and only if such-and-such is not the
20
case.
It is important to note that the verb say here is used to mean express the
thought and not assert. Thus utterances made within the scope of express sup-
positions are instances of statements, as well as assertions. Utterances of either
kind will instantiate a given statement when the utterance is of the declarative
sentence that is the statements first member and is made in the context that is the
statements second member. It is, Iadmit, somewhat infelicitous to have unas-
serted statements, for states often means asserts; however, alternative terms are
more likely to mislead. The term proposition, for instance, would have been bet-
ter (one may propound without asserting) had not so many philosophers already
21
appropriated it to stand for what a complete declarative utterance expresses.
20
This account of truth presupposes that there is a unique thought that a statement expresses.
See 10.1 for the refinements needed when this presupposition fails.
21
Peter Geach urged philosophers not to abandon the old use of proposition, meaning a form of
words in which something is propounded (not necessarily asserted), in favour of the modern use
standing for a sort of abstract intentional objects, whose principle of individuation has thus far
eluded capture in any clearly formulable criterion (Geach 1965, 449; see also Geach 1980, 512). But
this cause, although noble, is, Ifear, lost.
22 Introduction
22
As regards the Liar, Ifollow Prior (1958, 1961, 1971), Mackie (1973), and Smiley (1993) in hold-
ing that Liar-like utterances fail to qualify as statements: they do not succeed in saying that
such-and-such is the case. As such, they fall outside the purview of logic, although there remains
the residual problem of formulating deduction rules for a language in which well-formed formulae
may fail to say anything. For a defence of this view of the Liar, and an approach to the residual prob-
lem, see Rumfitt 2014a.
Introduction 23
non datur is knowable while Bivalence is not, then classical logic is not applicable
to semantic discourse. A fortiori, it is not applicable universally. Iargue that this
attack fails:it misdescribes the relationship between truth and falsehood. Much
can be learned from it, however. It rests on a theory of contentthe exclusion-
ary accountwhereby a statements sense, or logically relevant content, is given
by the possibilities that it excludes. The exclusionary account of content contains
important insights which Iapply later in the book.
In Chapter 5, I turn to the semantic arguments against classical logic that
Dummett deployed in the 1960s and 1970s. Here, too, one needs to discriminate.
Istart by considering his strong verificationist attack on classicism. According
to the strong verificationist, to know what a statement means is to know in
what conditions it would be verified. Dummett contends that a compositional
semantic theory that respects the tenets of strong verificationism will generalize
the semantics that Arend Heyting laid down for the language of intuitionistic
mathematics, and on that basis he argues that some classical validities cannot be
accepted as such. Icounter, though, that his argument for strong verificationism
is self-defeating, that strong verificationism is inherently implausible, and that
even if it were to be accepted, Heyting semantics is a poor basis for a composi-
tional theory of meaning. So the strong verificationist attack on classical logic
gets nowhere.
Also in the 1970s, however, John McDowell suggested a different, and more
powerful, line of attack using some of Dummetts premisses (see McDowell 1976).
This attack abjures the verificationist theory of meaning. Indeed, for McDowell,
understanding the connectives is a matter of knowing the semantic relationships
recorded in the familiar classical truth-tables. But that knowledge, while it is tan-
tamount to knowing that certain basic sequents are valid, does not yield knowl-
edge that all the sequents of the classical propositional calculus are correct. We
need an additional assumptionthat of Bivalencein order to reach that con-
clusion. McDowell argues that the assumption of Bivalence is insufficiently sup-
ported:while the Principle may be true, we do not know that every statement is
either true or false. On his view, then, we cannot know that the laws of classical
logic are sound. Icall this the dubious-grounds challenge to classicism.
Assessing this challenge involves developing an account of consequence that
respects the semantic relationships implicit in the familiar truth-tables with-
out presupposing Bivalence. Given the conclusions of Chapter 3, that account
must contain a modal element. However, the usual possible-worlds treatment of
modality presupposes Bivalence and so cannot be used in the present dialectical
context. Chapter6 develops an alternative modal semantics. Icapture the modal
aspect of consequence by saying that, when a conclusion follows from some
24 Introduction
premisses, there is no possibility at which the premisses are all true but where the
conclusion is not true. In this formula, a possibility is a way (which need not be
fully determinate) in which thingssome things, anywaycould be or could be
conceived to be. The notion of a statements being true at a possibility is under-
stood as follows:where the statement A says that P, A is true at the possibility x
means Necessarily, had x obtained, it would have been the case that P.
This way of understanding the notion of a possibility is natural, but it presents
problems for the construction of a compositional semantic theory of truth at a
possibility. The case of disjunction illustrates the difficulty. One might be tempted
to say that a disjunctive statement is true at a possibility if and only if one of the
disjuncts is true there, but while the if claim is acceptable, the only if claim is
not. The disjunctive statement Either a boy or a girl is at home is true at the pos-
sibility of a childs being at home, but neither of the disjuncts is true at that pos-
sibility. It is not necessary that, had a child been at home, a boy would have been,
for there might only have been girls at home. Equally, it is not necessary that, had
a child been at home, a girl would have been, for there might only have been boys
there.
Snags of this kind have encouraged theorists of modality to work with pos-
sible worlds, despite the contestable presuppositions they bring with them, but
Ipropose a way around the difficulty. The key notion is that of the closure of a set
of possibilities. Apossibility belongs to the closure of U just in case any statement
that is true at every member of U is true at it. Trivially, then, a set is always a sub-
set of its closure. When the converse inclusion obtains, so that a set is identical
with its closure, we call the set closed. One may prove that the closure of U is the
smallest closed set containing U, and Iargue for the following general axiom of a
possibility-based semantics:
This shows what form a semantic theory cast in terms of possibilities will take. The
theory will associate with each atomic statement a closed set of possibilities as the
statements truth-grounds. The theorys compositional principles will then asso-
ciate a closed set of possibilities with each molecular statement. (R) thus points
the way to the correct treatment of disjunction. Adisjunction may be true at pos-
sibilities where neither disjunct is true. It is, though, a maximally strong statement
that is entailed by both disjuncts:both disjuncts entail the disjunction, and the
disjunction entails any statement that both disjuncts entail. So the truth-grounds
of a disjunction will be the smallest closed set of possibilities at all of which one
or other of the disjuncts is true. That is, the truth-grounds of a disjunction are the
closure of the union of the truth-grounds of its disjuncts.
Introduction 25
This semantic axiom for disjunction allows for failures of the Law of
Distribution, so Iconclude Chapter6 by analysing the best-known attack on that
Law, namely, Hilary Putnams argument that it precludes a realist interpretation
of the quantum theory. My analysis confirms the widely held belief that Putnams
argument fails, although my account of the flaw in it differs from the well-known
diagnoses of Dummett and of Putnam himself in his later writings.
In Chapter7, Ireturn to the main line of argument, and apply truth-grounds
semantics in assessing the dubious-grounds challenge to classical logic. The com-
positional principles proposed in Chapters6 and 7 are acceptable to adherents of
many of the logical systems that are serious candidates to serve as codifications of
the norms of deductive inference; they are also highly stable under changes in the
background logic of the theory. However, they validate different logics depending
on the assumptions made about the structure of the underlying space of logical
possibilities. If we assume only that there is a well-defined closure operation on
the space andas per (R)that the truth-grounds of any statement form a closed
set of possibilities, then it is the rules of the intuitionistic propositional calculus
that turn out to be sound and complete. This result holds whether the metalogic
is classical or intuitionistic. If, however, we accept a stronger postulate about the
structure of that space, then we can validate the full classical propositional calcu-
lus. The postulate in question involves a new notion. Let us say that a statement
has a back if its truth-grounds are precisely those which are incompatible with
some set of possibilities. If a statement has a back, an assertion of it amounts pre-
cisely to the rejection or exclusion of all the possibilities in that set. Many philoso-
phersnotably the young Wittgenstein with his notion that any statement must
be bipolarhave postulated
I show that if we assume (B) as well as (R), then our semantic theory validates all
the rules of the classical propositional calculus, even in an intuitionistic met-
alogic. This shows how disagreement over logical laws can arise, not because
the parties attach different senses to a connective or quantifier, but because
they accept different postulates about the structure of the space of logical
possibilities.
I maintain, indeed, that some of the deepest challenges to classical logic are best
viewed as challenges to (B). Thus Brouwers strictures on the way we talk about
the infinite appear to entail that some statements about infinite mathematical
structures cannot be assumed to have backs. Iargue, though, that a classical logi-
cian can defend his position even if he accepts those strictures. Amore problem-
atic case for the classicist is the treatment of infinitesimal quantities in Smooth
26 Introduction
Our task is to examine certain attacks on the laws of classical logic. But what is a
law of logic? What, indeed, is logic?
2.1Consequence
Textbooks typically characterize the subject as the science of consequence. Thus
in an early section of his estimable primera section entitled What logic is
aboutwe find Benson Mates explainingthat
logic investigates the relation of consequence that holds between the premisses and the
conclusion of a sound argument. An argument is said to be sound (correct, valid) if its
1
conclusion follows from or is a consequence of its premisses; otherwise it is unsound.
1
Mates 1965, 2.Mates uses the terms sound and valid as synonyms; others take a sound argu-
ment to be one that is valid and has true premisses. Ishall follow Mates in using sound to be apply
to any argument whose conclusion follows from its premisses, whether or not those premisses are
true. Atheme of this chapter, however, is that there are many different notions of consequence and
correspondingly many notions of argumentative soundness. Ireserve the term valid for those
arguments whose conclusions follow logically from their premisses. Agoal of this chapter is to
articulate the sense of logically.
32 The Nature of Logic
others. The laws of logic are then taken to say, in general terms, which things
stand in that relationship. Thus a logical law which classical logicians accept, but
which intuitionist logicians do not accept without restriction, says that a proposi-
tion follows from the negation of its negation. Many other passages could be cited
which express this view of logic.
There is, to be sure, one important matter over which different adherents of the
view diverge:the nature of the relata of the consequence relation. Lemmon writes
of one statements following from some others. By a statement he means a Fregean
thought:something which is stated, or which could be stated, by the utterance of
a declarative sentence on a given occasion of use (1965, 6). This way of speaking
is undeniably natural. It comes easily to say (as it might be) The statement that
every set can be well-ordered follows from the statement that there is a choice
2
function on every non-empty set. There are, though, problems with this usage,
at least if it is deployed early in an investigation into the nature of logic. Before we
can fully apprehend a relation among things of a kind, we need to know under
what conditions such things are identical. When the things in question are state-
ments in Lemmons sense, any account of those conditions will invoke a number
of logical laws. If possible, then, we should postpone giving such an account until
after we have made some progress in elucidating the nature of those laws.
Mates takes a very different approach. Statements in Lemmons sense are
among putative relata of consequence which, for Mates, appear on sober con-
sideration to share a rather serious drawback, which, to put it in the most severe
manner, is this:they do not exist (1965, 8). For this reason, Mates takes conse-
quence to hold among declarative sentences, rather than among the thoughts
that utterances of such sentences might express. Because, though, he treats the
relata of consequence as true or false simpliciter, he is compelled to confine his
analysis to relatively unambiguous sentences not containing egocentric [sc.,
context-sensitive] words (1965, 11). Since few ordinary utterances are of such sen-
tences, this places a severe restriction on the scope of logical appraisal.
As well being severe, the restriction is unnecessary. As advertised in the
Introduction, I use the term statement to mean an ordered pair whose first
element is a meaningful, disambiguated, declarative type sentence and whose
second element is a possible context of utterance. To qualify as a statement it is
required that a complete thought would be expressed, were the declarative sen-
tence that is the statements first element to be uttered in the context that is its sec-
ond element. Statements in this sense (which is quite different from Lemmons)
2
Achoice function on a set A is a function F whose domain is the set of non-empty subsets of A,
and which is such that F(B) B for every non-empty B A.
Logical Laws 33
3
Contextual factors may be expected to influence the interpretation of an ambiguous
[sic: context-sensitive would have been better] expression uniformly wherever the expression
recurs in the course of the argument. This is why words of ambiguous reference such as I, you,
here, Smith, and Elm Street are ordinarily allowable in logical arguments without qualifica-
tion; their interpretation is indifferent to the logical soundness of an argument, provided merely
that it stays the same throughout the space of the argument (Quine 1982, 56).
34 The Nature of Logic
the logicians attention. Similarly, Lemmon takes it for granted that we shall
know what he means by follow when he says that logicians are concerned with
whether a conclusion does or does not follow from the given premisses. But do we
know this? The glosses philosophers and logicians have placed on follows from
are of limited help in identifying the intended sense. In the famous passage in
which he appropriated the word entails from the lawyers, G.E. Moore laid it
down that we shall be able to say truly that p entails q when and only when we
are able to say truly that q follows from p, ... in the sense in which the conclusion
of a syllogism in Barbara follows from the two premises, taken as one conjunctive
proposition; or in which the proposition This is coloured follows from This is
red (Moore 1922, 291). Even after studying logic, however, one may be forgiven
for doubting that there is a single such sense, or a single relation of properly logi-
cal entailment. Certainly, no logical system that Iknow comes close to providing
a complete characterization of whatever relation Moore had in mind.
I shall soon suggest a way of identifying a relation that deserves the title
of logical consequence, and a correspondingly favoured sense of follow.
However, our grasp of such a sense is surely tenuous in advance of theorizing
about logic. We need to elucidate the notion of logical consequence, not take for
granted that we all understand it. In this work of elucidation, it helps to bring
in some elements from an alternative approach to the problem of saying what
logic is.
4
For reasons that will emerge, though, Ido not think that logical rules are the only rules that
regulate the activity of deducing.
36 The Nature of Logic
5
silent. This contrast should occasion no surprise. One often infers B from A
because B provides the best explanation of A. Thus Whites inference is a good
one if his colleagues silence is best explained by the hypothesis that they have
no objections to his proposal. But that hypothesis explains their silence in part
because one of its implications (in tandem with background facts about Whites
colleagues) is that they will remain silent.
Can we say anything positive about the relationship between inference
and deduction (in the senses specified)? Many philosophers write as though
deduction is a species of inference, but on the present understanding of the
terms, that must be wrong. Since dog is a species of mammal, every dog is a
mammal, but not every deduction is an inference. Indeed, given that every
deduction is a performance while no inference is, no deduction is an infer-
ence, although some deductions result in one. More interestingly, some
deductions do not even issue in an inference. To infer B from A, we said, is
to take up, to accept, B as a result of reflecting on A. But in drawing out the
implications of A one may reach B without accepting itand, a fortiori, with-
out accepting it as a result of reflecting on A. Sometimes a thinker accepts
A and deduces B from it. His acceptance of B is then grounded in, or based
upon, his acceptance of A, and we may describe him as having deductively
inferred B from A. But the deduction of B from A may not issue in this infer-
ence. If B is known to be false, it may instead issue in the thinkers accepting
the negation of A on the basis of the negation of B. Equally, though, it may not
issue in any inference at all. Athinkers deducing B from A may make him
aware of an implicative relationship between A and B without leading him to
accept, or to reject, either A or B.
One thing these cases bring out is that deduction can play the role we expect
it to play in our intellectual economy only if it is applicable in drawing out the
implications of false premisses. Ryle entirely overlooks this important point.
Possessing a capacity for deduction, he tells us, is knowing how to move from
acknowledging some facts to acknowledging others (1945, 227; emphasis added).
Sometimes he goes further, and writes as though deduction were always a matter
of drawing out the implications of premisses that we actually know:As a person
can have a ticket [for a railway journey from London to Oxford] without actu-
ally travelling with it and without ever being in London or getting to Oxford, so
a person can have an inference warrant without actually making any inferences
5
White 1971, 292. White holds that ordinary English speakers respect this distinction between
infer and deduce, a claim which seems to me to be far-fetched. Iclaim only that a good philosophy
of logic will have some way of marking the difference.
Logical Laws 37
and even without ever acquiring the premisses from which to make them (Ryle
1950, 250). Whilst Ryle is not entirely explicit, it appears from the context that
acquiring a premiss means coming to know it, so that inference warrants are
applicable only to what one knows. He was, alas, far from alone in restricting the
scope of logic to the drawing of conclusions from known premisses. For both Mill
and Russell, logic is science of inference, where to infer is to come to know the
conclusions truth on the basis of prior knowledge of the premisses. For Frege, the
premisses of an inference must be, if not known, then at least asserted.
Even those philosophers who recognize that we deduce things from false
premisses sometimes fail to press the observation as far as it should be pressed.
According to Aristotle, whether we are engaged in demonstration (i.e. in drawing
out the implications of what we know) or in dialectic (an enquiry directed towards
deciding between a pair of contradictories) makes no difference to the production
of a deduction ... for both the demonstrator and the dialectician argue deductively
after assuming that something does or does not belong to somethingi.e. after
assuming that such-and-such is the case (Prior Analytics I, 24 a 2527). That is right
as far it goes, and deduction often assists dialectic by drawing out an obviously
false implication from one of the pair of contradictories between which the dia-
lectician is trying to decide, as when Socrates (in the Theaetetus) draws out absurd
implications of the hypothesis that knowledge is perception. However, we can also
deduce things from premisses that we already know to be false, as when an aged
dominie teaching Euclids proof that there is no greatest prime number for the for-
tieth year running begins his latest exposition of it by saying Suppose there were a
greatest prime number, N.
A thinker may, indeed, deduce implications from some premisses whatever his
epistemic attitude to the premisses. In order to infer B from A, a thinker must
accept both A and B, but he may deduce B from A regardless of whether he knows,
believes, or disbelieves A. This partly explains why the basic criterion of success
in deduction is the preservation of truth from premisses to conclusion, rather
than the preservation of knowledge, or knowability, or assertibility. Consider the
argument Suppose Mrs Thatcher was a KGB agent. In that case, she would have
taken great care to destroy all the evidence of her treachery. So no one will ever
6
know that she was a Russian agent. In an appropriate context, that might be a
perfectly good deduction, but the conclusion would make no sense if the argu-
ment were understood to be elaborating the hypothesis that someone knows that
Mrs Thatcher was a KGB agent. In the course of the deduction, we are drawing
out the implications of the truth of the initial supposition, not the implications of
6
Frank Jackson and John Skorupski have made cognate points about related conditionals.
38 The Nature of Logic
This is a sound deduction, and it owes its soundness partly to its dilemmatic form.
In 2.5, Ishall consider how that formal feature helps to account for the deduc-
tions soundness. For the present, though, the relevant point is that the deduc-
tive capacity being exercised is specific to the theory of electrical circuits. It is the
capacity to deduce which appliances will be on or off, given a circuit diagram.
This is not a purely logical capacity although, as our case shows, logical capabili-
ties are involved in it.
Pari passu, the relation of implication that sets the standard for successful
exercises of the present deductive capacity is not that of logical consequence. The
pertinent implication relation, as we may call iti.e. the relation to which the
deduction answersalso pertains specifically to electrical circuits. Apremiss
such as that switch S is openwill stand in this relation to a conclusionsuch as
that lamp L is offif L is off whenever S is open, given that the laws of electrical
7
Some logic textbookse.g. Lemmon 1965, 8draw a distinction between premisses and assump-
tions. Assumptions can be made, or introduced, at any stage in the deduction, whilst the premisses
are given at the start. But whilst the distinction may help to clarify the way deductions are used in
inferences, it is of no relevance to their soundness or validity. The logical rules are applied in just the
same way to draw out implications of premisses and assumptions.
Logical Laws 39
R L
B C
A D
S
assessments rest on our ability to latch onto the implication relation that is rel-
evant in the context of argument. So the enthymematic strategy will only work
if any implication relation that we can readily latch onto factors into a relation
of properly logical consequence, along with the hypothesis that such-and-such
statements are available to serve as suppressed, or unexpressed, premisses.
Even our present, simple case poses a problem for the strategy. To supply the
tacit premiss [or premisses] that the speaker intended but did not express would
be to articulate, in the form of statements, the rules implicitly followed by some-
one who is capable of tracing out the way electrical current flows around a circuit.
Since that capability relates to the diagrammatic presentation of the circuit, how-
ever, formulating it in words is no easy matter. Certainly, someone could be good
at deducing, from a circuit diagram, where the current flows while being unable
to articulate verbal principles from which that fact follows. As our case brings out,
being able to articulate such principles is not required for logical laws to be applied
in a deduction:the reasoner uses the Law of Dilemma to splice together two sub-
sidiary deductions to compose a third. The deductive capacity exercised in each
subsidiary deduction is the ability to work out which appliances are on in a circuit
that is presented diagrammatically. It follows that the very same ability is exercised
in the deduction that the subsidiaries compose. Despite thatand despite the fact
that all three deductions answer to an implication relation to which only electrical
possibilities are relevantthe whole deduction remains an application of the Law
of Dilemma. An account of logical laws must respect this fact.
The enthymematic strategy faces other difficulties. Pace Copi, if there is an
underlying relation of properly logical consequence, it must surely be that of
some variety of higher-order logic. As Stewart Shapiro (1991) has persuasively
argued, and as later chapters in this book will confirm, there are many central
mathematical theories whose underlying logic is higher than first-order. We
cannot replicate a higher-order consequence relation using first-order logic plus
additional premisses, even if those premisses are supplied as the infinitely many
8
instances of an axiom schema. Thus the consequences of second-order Peano
Arithmetic include the theorem that every natural number is separated from zero
by a finite number of applications of the successor operation; by contrast, this
theorem is not a first-order consequence of Peano Arithmetic, even though that
theory is understood to include every instance of the induction schema. If the
8
Smiley has objected to the enthymematic strategy that there may be no way of capturing the
generality of an extra-logical rule in any finite collection of additional premisses (Smiley 1995, 732).
This observation is correct, but we ought also to consider a liberalized enthymematic strategy in
which the tacit premisses may instead be supplied as the instances of an axiom schema.
Logical Laws 41
9
Acognate point is made by Quine, as he argues for the emptiness of the positivist doctrine
that a logically true statement is true by virtue of its meaning alone; see for example Quine 1960,
113. Despite his occasional descriptions of logic as the science of necessary inference, Iread Quine
as a fellow sceptic about the claim that we have a clear pre-theoretical notion of logical conse-
quence:There are philosophers of ordinary language who have grown so inured to the philosophi-
cal terms entails and inconsistent as to look upon them, perhaps, as ordinary language. But the
reader without such benefits of use and custom is apt to feel, even after Mr Strawsons painstaking
discussions of the notions of inconsistency and entailment, somewhat the kind of insecurity over
these notions that many engineers must have felt, when callow, over derivatives and differentials.
At the risk of seeming unteachable, Igo on record as one such reader (Quine 1953, 138).
42 The Nature of Logic
Bill is older than Charlie, while failing to accept that Al is older than Charlie gives
a sign that he does not understand comparative adjectives. As for the idea that
logical validity is special because, being independent of external circumstances,
it can be assessed simply by looking at the words and sentences involved and the
10
way they relate to one another, it fails to draw the distinction in the intended
place:someone can assess our argument about Al, Bill, and Charlie simply by
looking at the words, and without discovering the mens ages. Perhaps, indeed,
the idea fails to draw a distinction at all. The postulated distinction between the
conceptual and the empirical was a target of Quines assault on the dogmas of
empiricism (Quine 1951). Athinker with a sufficient grasp of the words force
and acceleration may be able to recognize the soundness of This body is accel-
erating; so a force is acting on it, but that hardly shows that the soundness of
this argument is independent of the external circumstances. To the contrary:it
depends on whether those circumstances conform to Newtons laws of motion.
We should acknowledge, then, that in exercising our deductive capacities we
trace out a variety of implication relations. However, if we do think about the
matter in this way, we shall need to say what is characteristic of those relations.
I am going to argue that implication relations have the three Tarskian struc-
tural features:they are reflexive, monotonic, and manifest the form of transitiv-
11
ity that is captured in the Cut Law. That is to say, where is any implication
relation, where A and B are individual statements and where X and Y are sets or
pluralities of statements, wehave:
Reflexivity: A A
Montonicity:If X B then X, A B
If X B for all B in Y, and Y A, then X A.
Cut:
10
Smiley 1995, 733. Smiley does not himself accept the idea.
11
Actually, the three features might better be called Hertzian for, as Tarski glancingly acknowl-
edged (Tarski 1930b, 62, n.1), they had been articulated in earlier publications by Paul Hertz. See
Hertz 1922, 1923, and (especially) 1929. For a rather differentbut, Ithink, complementaryphilo-
sophical defence of these three features of implication relations, see Cartwright 1987.
Logical Laws 43
12
Philo [of Megara] says that a sound conditional is one that does not begin with a truth and end
with a falsehood, e.g. when it is day and Iam conversing, the statement If it is day, Iam convers-
ing (Sextus, Pyrrhoneiae Hypotyposes ii 110, as translated in Kneale and Kneale 1962, 128). Sound
conditional translates the Greek dialecticians term for what we would now call a correct sequent.
13
Reflexivity is not in general satisfied if the metalogic is intuitionistic. In such a context,
one might think of adopting an alternative definition (closer to Sextuss original formula)
whereby X Philo B if and only if it is not the case that every member of X is true and B is not
true. Under this definition, and given an intuitionistic metalogic, Philonian consequence satis-
fies Reflexivity, Monotonicity, and Cut. Given that metalogic, however, the relation is not always
truth-preserving:where A Philo B, all that follows from the truth of A is that B is not not true.
Logical Laws 45
If some member of X is not true, then some member of X {A} is not true. So by
applying proof by cases we can move from Either some member of X is not true
or B is true to Either some member of X {A} is not true or B is true. That is to
say, if X Philo B then X, A Philo B, i.e. Monotonicity holds. Finally, suppose that
X Philo B for all B in Y, and that Y Philo A. By the first supposition either some
member of X is not true or all the members of Y are true. By the second supposi-
tion, either some member of Y is not true or A is true. In a classical metalogic,
A B and B C jointly entail A C , so we may conclude that either
some member of X is not true or A is true, i.e. X Philo A. That is to say, Philo
satisfies the Cut Law.
In a classical metalogic, then, we can prove that Philonian consequence is an
implicative relation and, in particular, that it obeys the Cut Law. Why is this a
problem? It is so because these results suffice to generate a version of the Paradox
of the Heap. Let us suppose that we have a sequence of a hundred transparent
tubes of paint, a1,...,a100, with the following properties:tube a1 is clearly red; tube
a100 is clearly orange and hence clearly not red; but for each n, tube an+1 is only
marginally more orange (and hence only marginally less red) than its predecessor
an. Let An be the statement Tube an is red. Since tube a1 is clearly red, statement
A1 is true, and since tube a100 is clearly not red, statement A100 is not true. Let us
now consider an arbitrary pair of adjacent statements, An and An+1. Since the tube
an+1 is only marginally less red than an, there is strong inclination to assert that it
is not the case that an is red whilst an+1 is not red. For if an were red whilst an+1 were
not, we would have a sharp cut-off in the correct application of the predicate red
between the tubes an and an+1, even though these tubes differ only marginally in
colour. There is a correspondingly strong inclination to assert that it is not the
case that the statement An is true whilst An+1 is not true. Now in classical logic, that
assertion is equivalent to an assertion that either An is not true or An+1 is true, i.e. to
an assertion of An Philo An+1. Thus our natural reluctance to contemplate a sharp
cut-off for redness in the sequence of tubes carries with it an acceptance of each
instance of An Philo An+1, for 1 n 99.
It is here, though, that the Cut Law causes problems. We have, in particu-
lar, A1 Philo A 2 and A 2 Philo A3. By Cut, these two statements yield A 1 Philo A3.
Ninety-seven further applications of the same inferential sub-routine eventually
bring us to A1 Philo A 100. This conclusion, however, contradicts our premisses. The
conclusion says that either A 1 is not true or A100 is true. Ex hypothesi, though, tube
a1 is red, so that statement A1 is true. Also, tube a100 is not red, so that statement
A100 is not true. In the light of this contradiction, we seem to be forced to do one of
two things. We either allow that there is a sharp cut-off in the correct application
of the predicate red, so that for some n, tube an is red and tube an+1 is not red; we
46 The Nature of Logic
will have to allow a corresponding sharp cut-off in the correct application of the
predicate true. Or we restrict the Cut Law as it applies to the relation of Philonian
consequence. This will involve restricting the application of the classical laws in
the metalogic.
The present version of the Sorites Paradox is interesting in that it makes no
assumptions whatsoever about the behaviour of the logical particles in the object
language. Indeed, it has been formulated for an object language which may not
have any connectives or quantifiers:the only statements involved in it are atoms
in which a single predicate (red) is combined with various names for coloured
14
tubes. For this reason, Icall the argument the structural version of the Sorites.
Our analysis of vague discourse in Chapter8 will open up a middle way that
allows us to retain the Cut Law without postulating a sharp cut-off in the correct
application of red. So Ishall adopt the working hypothesis that all implication
relations do conform to the Cut Law. Even at this early stage of the discussion,
however, it is worth having in view the problem that the structural version of the
Sorites presents. Even when we have an apparently compelling proof of a logical
principle, it needs to be tested against hard cases.
(I)For all statements A 1,...,An and B, A1,...,An B if and only if, for any
possibility x in , if A1,...,An are all true at x then B is true at x too.
It is easy to verify that a relation defined according to (I) will be reflexive and
monotonic, and will obey the Cut Law. Given also that the space of possibilities
includes the actual circumstancesi.e. includes the way things actually arethe
corresponding relation will be truth-preserving.
We often have an antecedent apprehension of a space of possibilities, , which
via (I) gives us a grip on the corresponding implication relation. However, there
14
I first presented this form of the paradox in the context of a multiple-conclusion logic in
Rumfitt 2008a, 823.
Logical Laws 47
is also a general result which shows how things can work the other way round.
That is to say:given any implicative relation i.e. a relation that is reflexive
and monotonic and obeys the Cut Lawthere will exist a space of possibilities,
preservation of truth at every member of which is equivalent to -relatedness.
Furthermore, if is truth-preserving, then the actual circumstancesthat is,
the way things actually arewill be a member of this space.
What guarantees this converse connection is a theorem which Dana Scott
attributed to Adolf Lindenbaum, but which posterity has insisted upon calling
15
the Lindenbaum-Scott theorem. In order to state and prove the theorem, we
need to introduce some terminology. Suppose we are given a set S of statements.
Abisection of S is a pair T=<K, L> of non-empty subsets of S such that K and L
are disjoint and their union is the whole of S. Given a bisection T, we define the
associated bisection implication T on S as follows:
15
See proposition 1.3 of Scott 1974. For the version of the theorem presented here, see Koslow
1992, 501.
16
Reflexivity:because K L=S, either A K or A L; so A T A. Monotonicity is trivial.
Cut:adapt the proof given in the previous section that Philo satisfies the Cut Law.
17
Suppose that A i L for 1 i n and that A 1,...,A n T B. We need to show that B L. Since
A 1,...,An T B, either some Ai K or B L. Since all the Ai L and K and L are disjoint, we have
B L, as required.
48 The Nature of Logic
Let be any implicative relation on S such that not all members of S are
equivalent under . Then A 1,...,An B if and only if A1,...,An T B for all
bisections T of .
are many other implicative relations to which the Theorem applies even if the
background logic is taken to be intuitionistic.
In order to see the Theorems significance, it helps to consider what the various
bisectionssc., the various bisective extensionsof a given implicative relation
are. By way of an example, let us take to be the smallest relation on a class of
statements concerning the physical attributes of bodies that respects Reflexivity,
Monotonicity, and Cut and for which
Body b is iron; Body b is in a magnetic field Aforce acts on body b.
So far as these three statements are concerned, the bisective extensions of then
include:
This simple example illustrates the way each bisective extension describes, as
fully as the relevant class of statements allows, a possibility admitted by the rel-
evant implicative relation. The possibilities admitted by an implicative relation
will not always be ones that could really obtain; it remains the case, though, that
the bisective extensions of such a relation capture a class of possibilities that is
cognate to the form of deductive reasoning that answers to the relation. Thus, in
the present case, the relevant possibilities are physical possibilities. For example,
item 2 on the list of bisective extensions above describes a physical possibility in
which the body is not made of iron, is not in a magnetic field, but in which some
other force acts upon it. Item 8, by contrast, describes a situation in which the
body is iron, is in a magnetic field, but in which no force acts on it. That situation
is physically impossible. The Theorem, then, tells us that our implicative relation
may be represented as preservation of truth at each of these physical possibilities.
More generally, given any implicative relation , a bisective extension of
describes, as fully as the relevant language permits, a circumstance that is
-possible; a statement will belong to the second of the two classes defined by
50 The Nature of Logic
the bisection just in case it is true at that possible circumstance. The totality of
such bisective extensions then defines the totality of possibilities that respect
the underlying relation . So the Lindenbaum-Scott Theorem says that some
premisses stand in to a conclusion if and only if the conclusion is true at every
one of these possibilities at which all the premisses are true. In other words, it
guarantees the existence of a space of possibilities for which the equivalence
(I) holds.
Given also that is truth-preserving, it follows that one of the possible cir-
cumstances in will be actual. That is to say, for this bisection <K, L>, all and
only the statements in L are in fact true.
In formulating the Theorem, Scott worked in the context of a
multiple-conclusion logic. It is worth explaining why I have not followed
him in this. Scott writes of deductions issuing in conditional assertions, and
some such notion is apposite here. On the strength of a deduction of B from
the assumptions A 1,...,A n, we may say:Given all of A 1,...,A n , we have B or
B, on the assumptions A 1,...,A n. Some have found the notion of a conditional
assertion obscure, but we may understand it by way of a natural generaliza-
tion of the norms for the speech act of outright assertion. An outright asser-
tion is governed (at least) by the norm of truth:one should not assert B when
B is not true. In making an assertion, we present ourselves as conforming to
this norm, even if we breach it. Aconditional assertion is governed by the
corresponding conditional norm: one should not assert B, on the assump-
tions A 1,...,A n, when all of A 1,...,A n are true and B is not true. This norm is
appropriate for the speech act in which a deduction issues, for the implication
relation that the deduction traces out is assumed to be truth-preserving, so a
sound deduction does indeed exclude the case where all the A i s are true and
B is not.
Scott took these conditional assertions to be what was expressed by Gentzens
sequents or Sequenzen. Ishall not try to decide how closely this interpretation
captures Gentzens intentions, but let us use Gentzens notation and write B, on
the assumptions A 1,...,Anas
A1 , , An : B .
Pace Scott, the colon here is not a sign for a relation. Switch S is open:lamp L is
off means L is off, on the assumption that S is open; it is a conditional assertion
that L is off. It does not mean The conclusion L is off follows from (or is deduc-
ible from) the premiss Switch S is open, which is an outright, unconditional
assertion that the conclusion stands in a certain relation to the premiss. All the
Logical Laws 51
same, when using the colon, some implication relation is to be taken as under-
stoodfrom the conversational context or backgroundas setting the standard
for the deduction that issues in the conditional assertion; this relation will meet
the conditions specified earlier.
It is, though, precisely Scotts metalogical reading of the colon that permits
him to follow Gentzen further and allow more than one formula to appear
after the colon in the succedent of the sequent. Thus, for Scott, the general
form of a conditional assertion is A 1,...,A n :B 1,...,Bm. He understands this to
mean:whenever all the statements [in A 1,...,A n] are true under a consistent
valuation, then at least one [statement in B 1,...,Bm] must be true also (Scott
1974, 417). Scotts explanation of A 1,...,A n :B 1,...,Bm is fine in its own terms.
However, it endows the explanandum with the sense of a metalogical state-
ment, one which says that a certain relation obtains between the set (or plural-
ity) of premisses A 1,...,A n and the set (or plurality) of conclusions B 1,...,Bm.
To assign such a sense is to change the subject from deductions. In making a
deduction, we do not merely identify a finite set of conclusions, one or more
of which must be true if all the premisses are true. Rather, we elaborate those
premissesthe deductions initial assumptionsby making specific further
assertions within their scope. (While some deductions terminate in a disjunc-
tive conclusion, such a piece of reasoning is adequately represented in the form
A 1,...,A n:B 1 ... Bm.) For this reason, Ishall confine my analysis to condi-
tional assertions with a single statement as succedent. Although Scotts own
proof of the Lindenbaum-Scott theorem was for the multiple-conclusion case,
it is (as we have seen) straightforwardly adapted to the single-conclusion case.
So we lose no formal power, but maintain the connection with our topic of
18
deduction, by restricting ourselves to succedents with only one member.
18
Greg Restall (2005) has suggested reading X :Y (where X and Y are both sets of state-
ments) as meaning It would be incoherent (or self-defeating) to accept all the statements in X
while rejecting all the statements in Y. Care must be taken to gloss incoherent if this read-
ing is not to generate versions of Moores paradox; we need incoherent to mean something
like logically contradictory and reject to mean reject as false rather than as impolite, or
ungrounded. (See Rumfitt 2008a, 80 and, for further criticism of the reading, Steinberger
2011.) But in any case, the reading is unhelpful for our purposes. We ordinarily assume that an
intuitionist logician will not accept the sequent :A A as valid, for arbitrary A, whereas
the classical logician will accept this sequent. On Restalls reading, however, logicians in both
schools will accept any such sequent as valid. For, on his reading, accepting :A A as
valid amounts only to deeming it incoherent to reject A A, and the intuitionist does accept
this:in his logic, (A A) yields a contradiction. It would confuse matters to work with a
consequence relation which, in this way, generates the appearance of classical logic without its
substance.
52 The Nature of Logic
single-member succedents. Thus the rule that is applied in both of our dilem-
matic arguments may be schematized as follows:
(1) X, A :C Y, B :C
___________________
X, Y, A B :C
Here, A, B, and C are arbitrary single formulae, X and Y are arbitrary sets of for-
mulae, and the horizontal line is read as so or therefore. Again, Ido not claim
that Gentzen had this interpretation in mind when he showed how to formal-
ize classical and intuitionistic logic as sequent calculi (see Gentzen 1935). All the
same, a classical logician will accept the classical sequent rules as sound when
they are interpreted as general rules for constructing new deductions, or mak-
ing new conditional assertions, from old ones, so long as the implication relation
corresponding to the relevant deductive capacity is held constant throughout the
19
derivation, and so long as it meets our conditions on implication relations. In
particular, nothing in this way of formalizing logic requires that the colon should
be taken to signify a notion of specifically logical deduction. On this conception,
logical rules are generally applicable rules for forming new deductions from old;
they are not initially identified as rules that regulate the activity of specifically
logical deduction. This seems to me to be a significant advantage of the account,
for (as Iargued in 2.1) it is not obvious in advance of theory what the activity of
specifically logical deduction is supposed to be. It is correspondingly unclear in
advance of theory to which implication relation specifically logical deduction is
supposed to answer.
Our conception of logical rules makes it easy to see why logic is useful. Being
able to deduce conclusions from premisses is clearly useful, if only because it
often shows that one or other of those premisses is false. So any thinker will
benefit from mastering generally applicable techniques for extending his
deductive capacities. On the recommended conception, mastery of the logi-
cal rules provides such techniques. In learning to reason about physics, say, a
thinker may start with a rather limited deductive capacity. We may pretend,
19
Aclassical logician may, though, worry about how completeness is to be secured, given that
we are eschewing many-membered succedents. For, in Gentzens sequent calculus, the opera-
tional rules yield intuitionistic logic when the system is restricted to single-member succedents;
he obtains full classical logic by allowing succedents containing more than one statement.
However, we can obtain classical logic with single-member succedents if we take the relata of
implication relations to comprise rejections of statements as false, as well as acceptances of them
as true; this is the style of formalization that Irecommended to classical logicians in Rumfitt
2000.
54 The Nature of Logic
just for simplicity, that his competence in this area is confined to deductions
in the form:A resultant force is acting on body a; so a is accelerating. If, how-
ever, the thinker can reliably contrapose, then his competence will extend to
that wider deductive capacity that takes one from the premiss Body a is not
accelerating to the conclusion No resultant force is acting on body a. What
is more, mastery of contraposition, and of other logical rules, will also expand
his deductive capacities in other fieldsindeed, in all other fields, given that
the logical particles such as not and all are ubiquitous. The theorems of logic
may convey no substantive information, but mastery of logical rules expands
all a thinkers deductive capacities. Techniques are no less valuable for being
applicable only indirectly.
Formula (1)above is a rule, not a statement, and so cannot itself be assessed as
true or as false. However, its correctness presupposes the truth of a logical law. For
rule (1)will be generally applicable in producing new deductions from old only if
the following law istrue:
Formula (2)expresses the logical Law of Dilemma, and it illustrates a general the-
sis:at least in the first instance, logical laws do not characterize some more-or-
less elusive relation of specifically logical consequence. Rather, they are general
laws governing all implication relations. What is transcendent about the Law of
Dilemma is not that it specially concerns some favoured relation of logical entail-
ment, although, if there is such a relation, the Law will apply to it a fortiori. Rather,
its transcendence lies in its concerning any implication relation, whether it be
implication in the theory of circuits, implication in mathematics, or implication
in anything else. Of course, we are not entitled to assert a general law such as
(2)simply on the strength of a couple of favourable cases; apparent counterexam-
ples need to be considered too. In recent discussions, cases involving vagueness
and quantum mechanical indeterminacy have been pressed against (2). Later in
this book Iwill try to show how the pressure to restrict the Law of Dilemma in
20
these cases may be resisted.
On the conception of the subject that Iam recommending, the basic logical
laws will be highly general. Thus, in a sequent calculus in which one and only one
formula appears on the right of the colon, the standard rule for introducing the
20
See 5.3, 6.5 and Chapter8. See also Rumfitt 2010, which defends Dilemma against a rather
different sort of challenge, due to Colin Radford (Radford 1985).
Logical Laws 55
conditional on the left of the colon (i.e., for constructing deductions with a condi-
tional premiss)is:
X, B :C Y :A
(3) ___________________
X, Y, A B :C
Now in the special case where X is empty, and Y is a singleton whose only member
is A, and where C is identical with B, rule (3)reducesto
B :B
(5) ___________________A :A
A, A B :B
Given that every implication relation is reflexive, the conditions above the line
will be fulfilled no matter what sort of deduction the colon may signify, so the
special case reduces furtherto:
(6) ___________________
A, A B :B
Rule (6)is Modus Ponens and it presupposes the truth of the followinglaw:
Note that (7)the traditional logical law of detachmentfollows from the more
general law (4).
Although Ihave emphasized the variety of implication relations to which our
ordinary deductions answer, the last paragraph points the way to a principled
21
identification of a relation of specifically logical consequence. Law (4)tells us that
if certain deductions are sound (by the standards laid down by a given implication
relation) then a related deduction will also be sound (when assessed by the same
21
The relation identified here, though, is narrow logical consequence, not the broader notion
that Moore invokes in the passage quoted in 2.1. Isay more about the difference in the next chapter.
56 The Nature of Logic
22
Thus Michael Dummett:The first to correct this distorted perspective [in which a logic is con-
ceived primarily as a collection of logical truths], and to abandon the analogy between a formaliza-
tion of logic and an axiomatic theory, was Gentzen ... In a sequent calculus or natural deduction
formalization of logic, the recognition of statements as logically true does not occupy a central place
... The generation of logical truths is thus reduced to its proper, subsidiary, role, as a by-product, not
the core, of logic (Dummett 1981a, 4334).
23
While my theory accords a central role to the plurality of implication relations, Itake the laws
of logic to be absolute:a logical law is a truth about every truth-preserving implication relation. My
position is, then, very different from the logical pluralism of Beall and Restall (2006). For compel-
ling objections to their view, see Priest 2006, chapter12.
Logical Laws 57
coming to a knowledge of something which we did not know before. (Mill 1891, Book II,
chap. iii, 1)
The difficulty he discerns in an affirmative answer does not arise only for syllogis-
tic reasoning in the strict sense, but for watertight or conclusive deductions more
generally:
Logicians have been remarkably unanimous in their mode of answering this question. It
is universally allowed that a syllogism is vicious if there is anything more in the conclu-
sion than was assumed in the premises. But this is, in fact, to say that nothing ever was,
or can be, proved by syllogism which was not known, or assumed to be known, before.
(II iii 1)
24
Ido not claim that the present stipulation gives an adequate account of what contemporary
philosophers mean by the term epistemic possibility, only that it helps identify the fallacy in Mills
argument.
58 The Nature of Logic
and the conclusion is not true, we need the additional assumption that every
epistemic possibility is a logical possibility. That is, we need the assumption
that whenever A does not know that not P, it is logically possible that P. Again
assuming that the background logic is classical, this assumption is equivalent
to its contrapositive form: Whenever it is logically necessary that not P, A
knows that not P. Because there is nothing special about negated statements
in this regard, the contrapositive form is justified only if one is justified in
assuming, quite generally, that whenever it is logically necessary that Q, A
knows that Q. It is now clear, though, that it is Mill, not the deductive reasoner,
who is engaged in petitio principii. His argument requires the assumption that
the relevant thinker already knows every logically necessary statement. For all
his argument shows, then, a thinker may gain knowledge through deductive
argument. Indeed, for all it shows, a thinker may thereby gain knowledge that
he could not attain by any other method.
There is, however, another problem which arises even if we allow that a deduc-
tive capacity enables a thinker to gain knowledge that he could not otherwise
attain. The problem may be explained in relation to a simple example. Suppose
Iam strapped to the chair in my study. From that chair, Icannot see the street
below. I do, however, see that it is raining, and thus know that it is raining.
Moreover Iknow, ultimately on inductive grounds, that if it is raining the street is
wet. Accordingly, Ireason as follows:
1. It is raining
2. If it is raining, the street is wet
So
25
3. The street is wet.
In this case, exercising my deductive capacity has brought me knowledge that (in
my current position) Icould not otherwise have attained. In making the deduc-
tion, Icome to know that the street is wet. Ex hypothesi, though, Icannot see the
street, so Icannot come to know the conclusion simply by exercising my percep-
tual capacities, which is how Icame to know the first premiss. Similarly, Ican-
not come to know the conclusion on general inductive grounds, which is how
Icame to know the second premiss. Even in England, so pessimistic a view of the
weather (or of the wastefulness of the water companies) would not yield knowl-
edge. By exercising my deductive capacity on the knowledge already delivered by
25
We need not worry what the pertinent implication relation is. For if the rule for introducing
on the left is accepted as regulating the deductive employment of the English conditional, then
arguments by Modus Ponens will be sound no matter what the contextually relevant implication
relation may be.
Logical Laws 59
perception and induction, however, Ican come to know something that Icould
not know on either of those bases severally.
The question such cases raise is this. In this example and in others like it, exer-
cising a deductive capacity certainly yields a belief. But under what conditions
does belief in a deductions conclusion qualify as knowledge?
A natural first shot at stating those conditionsa shot, Ishall argue, that is
rather better than many philosophers now supposeis what we may call the
Deduction Principle:
We clearly need a clause requiring that the thinker should continue to know the
premisses:if his knowledge of the premisses were to be destroyed by misleading
counter-evidence acquired in the course of making the deduction, then we should
not count his belief in the conclusion as knowledge. We have, Ithink, enough of
a grip on the notion of deductive competence for the Deduction Principle to be
more than a tautology. Whatever implication relation may set the standard for
assessing a deduction as sound, some people will be reliable in making deductions
only when the premisses really stand in that relation to the conclusion, while oth-
ers will not be reliable. This division gives us our grip on the notion of deductive
competence. In fact, in discussing the worries about the Deduction Principle that
Iwish to address, it will help to focus on the special case of the Principle where the
sort of deduction under consideration is specifically logical deduction; any logic
teacher certainly has a grip on the notion of logical deductive competence. The
implication relation that correctly executed logical deduction traces out will of
course preserve truth.
Do we need to add any further conditions to the Deduction Principle to ensure
that belief in the conclusion qualifies as knowledge? Perhaps we do. Some epis-
temologists will say that the thinker must not only be deductively competent,
but must know, or at least believe, that he is if his conclusive belief is to qualify
as knowledge. Others saymore cautiouslythat a thinker who gains knowl-
edge in this way cannot believe that he is deductively incompetent. Whether
one imposes these requirements will depend on ones general epistemological
predilections, but two sorts of case have been thought to cast doubt even on the
barebones Principle that has been stated. First, there are the so-called Dretske
26
Compare the formulation of Multi-Premiss Closure in Hawthorne 2004, 33.
60 The Nature of Logic
cases of which the following is the most famous (see Dretske 1970). At the zoo
one day, you glance into a pen labelled zebras and see a black-and-white striped
horse-like mammal. The animal is, indeed, a zebra, so you know, it seems, that the
animal in the pen is a zebra. That premiss logically entails that the animal in the
pen is not a non-zebra carefully disguised to look like a zebra. So, by competently
making a logically valid deduction, you come to believe the conclusionthe true
conclusionthat the animal is not a non-zebra carefully disguised to look like
a zebra. Many philosophers, though, share Dretskes view that you do not know
that conclusion. In order to know it, you would need evidence that excluded the
possibility of the animals being a non-zebra disguised as a zebra, and your inex-
pert glance into the pen fails to provide such evidence.
A second sort of case involves the accumulation of epistemic risks. Aver-
sion of the Paradox of the Preface provides a simple example. Suppose you have
composed a book comprising only true statements. Suppose too that you know
each statement in the book to be true. Now a plausible necessary condition for
knowing a statement to be true is that there should be very little risk, given your
evidence, that the statement is false. Ex hypothesi, then, you meet this condi-
tion in respect of each individual statement in the book. Now suppose, how-
ever, that you apply the rule of and-introduction to all the statements in it,
thereby reaching a conclusion that is the conjunction of all these individual
statements. This seems to be a case of coming to believe a conclusion by com-
petently deducing it from premisses you know, so the Deduction Principle tells
us that your belief in the conjunction will have the status of knowledge. That
claim, though, seems to be inconsistent with the postulated necessary condi-
tion for knowledge. For even when the risk of each conjuncts being false is low,
the risk of the conjunctions being false will be higher, and if the book contains
sufficiently many statements the latter risk can be high enough to disqualify
you from knowing the truth of the conjunction, even though the conjunction is
true and you believe it.
What lies at the root of this latter objection is a probabilistic conception of
epistemic risk. Some philosophers, anxious to ensure that fallible thinkers can
acquire knowledge, will wish to say that Ican know that it is raining by looking
out of the window, even when Iam susceptible to occasional hallucinations of
rain, so long as the chance of my hallucinating rain is small. Suppose then that
Iam prone to occasional brainstorms which can do any of three things:they can
make me hallucinate rain; they can make me reach inductive conclusions that
are not supported by my evidence; and they can make me deduce conclusions
from premisses which do not (in the contextually relevant sense) imply them. On
the present view, my being susceptible in this way need not preclude me from
Logical Laws 61
(see Sainsbury 1997, although he calls the principle the Reliability Conditional).
The modality here relates to the knowing, rather than to what is known. To put
the point in terms of possible worlds, if a thinker knows that P, there is no nearby
world (no world which could easily have been actual) in which he falsely believes
that P using the same method that he uses in forming his actual belief that P. Thus,
given that Iknow that it is raining, there is no nearby possible world in which
Iapply M1 and come falsely to believe that it is raining. Again, given that Iknow
that the street is wet if it is raining, there is no nearby possible world in which
Iapply M2 and come falsely to believe that conditional. Although Sainsbury pro-
posed SP only as a necessary condition for knowledge, those with reliabilist sym-
pathies in epistemology will be tempted to take it to be sufficient as well:on this
view, a true belief that is formed in such a way that it could not easily be wrong
will have the status of knowledge.
How does the Safety Principle bear on knowledge acquired by deduction?
To see how it does, let us return to our original deduction. We can think of the
method Iused to reach its conclusion as a composite, M*, in which my logical
capacity is deployed to splice together M1 and M2. Now let us ask how a belief that
the street is wet that is formed using M* could have been false. Since the premisses
of the deduction logically imply its conclusion, if that conclusion is false then one
or other of the premisses must be false. Moreover, M* includes M1 as the basis
for the first premiss and M2 as the basis for the second. So in any nearby world in
which my belief that the street was wet is formed using M* but is false, at least one
of the following three conditions must bemet:
(1) my belief that it is raining has its actual basis, but is false;
or
(2) my belief that the street is wet if it is raining has its actual basis, but is false;
or
(3) Ideduce the conclusion from my premisses, but in fact my premisses do
not imply my conclusion.
Now we are supposing that I know the first premiss of my argument. By the
Safety Principle, then, there is no nearby world in which possibility (1)obtains.
Similarly, given that Iknow the second premiss, there is no nearby world in which
possibility (2)obtains. Finally, it is a mark of deductive competence that when a
thinker deduces a conclusion from some premisses, it could not easily have been
the case that his premisses fail to imply his conclusion. Accordingly, given that
Iam deductively competent, and that all nearby worlds belong to the space of pos-
sibilities associated with the relevant implication relation, there is no nearby world
64 The Nature of Logic
in which possibility (3)obtains either. There will be cases where the italicised con-
dition does not obtain. However, when we are dealing with logical deduction, as
in the special case of DP that we are considering, the condition will obtain:for
any way in which things could easily have been is logically possible. But we said
that, in any nearby world in which my belief that the street was wet has its actual
basis but is false, at least one of conditions (1) to (3) must be met. And we have
just argued that, when the conclusion is competently deduced from known prem-
isses, none of these conditions is met in any nearby world. Accordingly, there is
no nearby world in which the conclusive belief has its actual basisnamely, M*
but is false. So, if we accept the converse of SP, that conclusive belief will qualify
as knowledge. Thus SP in tandem with its converse vindicates the Deduction
Principle.
What, though, of the apparent counterexamples to the Deduction Principle?
Our analysis gives us the resources, Ithink, to resist both of them. In the Dretske
example, the key point is that whether the application of a method yields knowl-
edge may depend on the circumstances in which it is applied. The method that
Dretskes subject applies is that of classifying an animal on the basis of its visual
appearance. There are two cases to consider. In the first case, there is a joker at the
zoo who is disposed to disguise non-zebras as zebras; here, the subjects belief that
the animal in the pen is a zebra is formed in such a way that it could easily have
been wrong, so that the subject does not know the deductions premiss. In the sec-
ond case, there is no such joker. In this latter scenario, the subject does know the
premiss of the deduction. However, his belief that the animal in the pen is not a
non-zebra disguised to look like a zebra is formed in such a way that it, too, could
not easily have been wrong in these circumstances; so we may allow that the sub-
ject also knows the conclusion of the deduction. Either way, then, Dretskes case
poses no threat to the Deduction Principle.
Matters are similar with the Paradox of the Preface. At any possible world
where the long conjunction is false, at least one conjunct is false. Now if the long
conjunction could easily have been false, there is a nearby possible world at which
it is false. At that nearby world, though, at least one of the conjuncts will be false,
showing that one of the conjuncts could easily have been false. That is to say, if
the long conjunction could easily have been false, then the author does not, after
all, know every statement in his book. So either he does know every statement
in the book, in which case the long conjunction could not easily have been false,
so that we may credit him with knowledge of the conjunction; or his book con-
tains a statement that he does not know, in which case the Deduction Principle is
inapplicable. Properly analysed, then, the Paradox of the Preface also provides no
counterexample to the Principle.
Logical Laws 65
X, B :CY :A
___________________
X, Y, A B :C
Logical Necessity 67
where each of A, B, and C is an arbitrary formula and where X and Y are arbitrary
sets of formulae. In the special case where X is empty, where Y is a singleton whose
only member is A, and where C is identical with B, this reducesto
B :B A :A
____________________
A, A B :B
Since every implication relation is reflexive, the conditions above the line will be
fulfilled no matter what implication relation sets the standard for the conditional
assertions signified by the colon. Accordingly, the special case yields
A, A B : B.
1
If there should be cases for which it is unclear whether this condition is met, we may follow
Tarski in making a definite, although in greater or lesser degree arbitrary, division of terms into
logical and extra-logical (Tarski 1936, 420).
68 The Nature of Logic
say they are, then things are as the conclusion says they are. Ishall call this doc-
trine Aristotles Thesis, after his famous claim that a deduction is a discourse in
which, certain things being stated, something other than what is stated follows
b
of necessity from their being sothat is, from their being the case (An. Pr., 24
1819).
It is worth noting one feature of the formulation of the Thesis. It says that when-
ever a conclusion B follows logically from some premisses A1,...,An and where
B expresses the thought that Q and each Ai expresses the thought that Pi, it is
logically necessary that Q if P1 and...and Pn. In this formulation, no attribution
of truth to A1,...,An or B falls within the scope of the modal operator. This point
is important in avoiding specious objections to the Thesis. As a counterexam-
ple to the claim that logically valid arguments necessarily preserve truth, Jean
Buridan proposed Every statement is affirmative; so no statement is negative
(Sophismata, c hapter8, sophism 1). Given that negative means not affirmative,
Buridans argument is logically valid, but its premiss can be true while its con-
clusion cannot be true. His case, however, is no counterexample to the present
formulation of Aristotles Thesis. Buridans premiss says that every statement
is affirmative; his conclusion says that no statement is negative; and it is indeed
necessary that no statement is negative if every statement is affirmative. What is
peculiar about the case is simply that it is impossible to express a truth by uttering
the English sentence No statement is negative. The case shows, to be sure, that
talk of a valid arguments necessarily preserving truth is somewhat telegraphic.
It would be more accurate to say that, in a valid argument, the facts cannot be as
the premisses say they are unless they are also as the conclusion says they are (I
quote G.E. Hughess translation of a moral that Buridan draws from his case; see
Hughes 1982, 37). The more accurate formula, though, is a mouthful, and Ishall
write of necessary truth-preservation while asking readers to understand me in
the way just explained.
Is Aristotles Thesis correct? If so, how does the sort of necessity that connects
the truth of premisses to the truth of a conclusion that follows logically from them
relate to other kinds of necessity? We may be encouraged to accept the Thesis
by our tendency to use modal verbs in glossing the concept of consequence. If
Thomas is material follows from Everything is material, then it must be the
case that Thomas is material if everything is; Fido barks does not follow from
Most dogs bark and Fido is a dog because it is possible for most dogs to bark
and for Fido to be a dog without Fidos barking. However, the Thesis needs more
support than a linguistic tendency can provide and, in the present case, the ten-
dency is double-edged. For the word must often indicates that an inference has
been made, even in cases where no interesting modality relates premisses to
70 The Nature of Logic
conclusion. The Smiths must be away, we say, noticing the mail on their door-
mat, even though there is no interesting sense in which mail on a doormat neces-
sitates the residents absence.
At any rate, Aristotle had an argument for his Thesis. As Timothy Smiley
explains it, the ingredient of necessity [in Aristotles account of consequence] is
required by his demand that proof should produce understanding (episteme),
coupled with his claim that understanding something involves seeing that it can-
not be otherwise. Hence a proof needs to establish the necessity as well as the truth
of its conclusion, which means not only starting from necessarily true axioms but
proceeding by steps that preserve necessity as well as truth (Smiley 1998, 599).
In due course, Ishall present a related but, I hope, more powerful argument
for Aristotles Thesis. First, though, let us remark that many philosophers have
rejected the Thesis. As we saw in the previous chapter, Philo of Megara held that a
conclusion follows from some premisses if, for whatever reason, it is not the case
that all the premisses are true and the conclusion is not true. Given classical logic,
this is equivalent to the condition that either one of the premisses is not true or the
conclusion is true. Philos view may seem counterintuitivewe do not ordinarily
think that Air conducts electricity follows from The Moon is made of ironbut
it has had distinguished defenders. According to Whitehead and Russell, Philos
account captures the essential property of consequencenamely, that what-
2
ever follows from true premisses is true, and in his Introduction to Mathematical
Philosophy, Russell tried to explain away our tendency to suppose that conse-
quence involves a stronger tie between premisses and conclusion:
In order that it be valid to infer q from p, it is only necessary that p should be true and that
the proposition not-p or q should be true. Whenever this is the case, it is clear that q must
[sic] be true. But inference will only in fact take place when the proposition not-p or q
is known otherwise than through knowledge of not-p or knowledge of q. Whenever p is
false, not-p or q is true, but is useless for inference, which requires that p should be true.
Whenever q is already known to be true, not-p or q is of course also known to be true, but
is again useless for inference, since q is already known, and therefore does not need to be
inferred. In fact, inference only arises when not-p or q can be known without our know-
ing already which of the two alternatives it is that makes the disjunction true. Now, the
circumstances under which this occurs are those in which certain relations of form exist
between p and q ... But this formal relation is only required in order that we may be able to
know that either the premiss is false or the conclusion is true. It is the truth of not-p or q
that is required for the validity of the inference; what is required further is only required
for the practical feasibility of the inference. (Russell 1919, 153)
2
The essential property that we require of implication is this:What is implied by a true propo-
sition is true. It is in virtue of this property that implication yields proofs (Whitehead and Russell
1925, 94).
Logical Necessity 71
to justify the rule of Modus Ponens, one has to show that B is true whenever A
and A B are true. In arguing for this conclusion, Frege appeals to the stipu-
lation by which he gave sense to the operator ; part of that stipulation says that
A B is false if A is true and B is not true. He then reasons as follows:if B were
not the True, then since A is the True, A B would be the False (Frege 1893,
14, 25; Ihave updated the logical symbolism). In a case where Modus Ponens
is applied in a Millian inference, the conclusion B will be true. In such a case,
then, the supposition from which Freges reasoning startsnamely, that B is not
trueis a false supposition. His argument is none the worse for that, but it shows
how even an adherent of a geometrical formalization of logic is driven to rely
on our ability to reason from false suppositions in order to ground one of the
formalizations primitive rules. (Similar points go for the axioms.) Hilbert-style
formalizations of logical systems remain important tools in proof theory, but
from a philosophical point of view they are much inferior to natural-deduction
or sequent-calculus formalizations when it comes to showing logical truths for
what they are:namely, the by-products of rules of inference that may be applied
to suppositions which may be falseby-products that arise when all the suppo-
3
sitions on which a conclusion rests have been discharged.
Even if the Philonian account is rejected, it does not follow that the notion of
logical consequence is implicitly modal. In the passage quoted from Introduction
to Mathematical Philosophy, Russell accords only epistemological significance to
the formal relations that obtain between premisses and conclusion in canonical
instances of consequence, but one might instead take the obtaining of such a rela-
tion to be partly constitutive of logical consequence. The account of consequence
that now prevails among mathematical logicians has this feature. In his great paper
on truth, Alfred Tarski solved the problem of giving a recursive theory of truth for
a language containing quantified statements by invoking the notion of a sentential
function, in which one or more components of a complete formula are replaced by
variables of the appropriate category (see Tarski 1935, 177, definition 10). He then
introduced the notion of a sentential functions being satisfied by an assignment of
objects (or things of other categories) to variables (Tarski 1935, 189 ff.). This notion
generalizes that of a predicates being true of an object:an assignment satisfies a
sentential function if the function is true of the things assigned to the correspond-
ing variables. Thus an assignment in which an object, a, is assigned to the variable
x satisfies the function x is white if and only if a is white. One can then say that
the universally quantified formula xFx is true if and only if every assignment of
objects to the variable x satisfies the function Fx.
3
See again the quotation from Dummett in n.22 of Chapter2.
Logical Necessity 73
consequence meets the conditions for being an implicative relation which pre-
serves truth (see the proof in 2.3). All the same, Ithink that the question can be
answered:we can explain why the modal aspect of properly logical consequence
is not only useful but, in a sense, unavoidable.
implicative relation that is, any relation satisfying the structural conditions
of Reflexivity, Monotonicity, and Cutobtains between some premisses and a
conclusion if and only if truth is preserved at each member of a related space .
(That is, at each member of , either the conclusion is true or some premiss is
not true.) We may think of the members of as possibilities. Admittedly, for
many implicative relations, the members of the space will not correspond to any
real possibility; but for our purposes that does not matter. What does matter is
that a whole range of philosophically significant species of possibility may be
generated from an implicative relation in this way. Given any implicative rela-
tion , we can begin to introduce a corresponding necessity operator by
way of the principle that A is true if and only if A is true at every member
the Big Bang, for examplethat could, physically, have been otherwise. The need
not to beg such questions shows that physical necessity PA cannot be defined as
(P A) where expresses absolute necessity, and P conjoins all the physical
laws. If there is such a notion as absolute necessity, its logic is surely S5. But if PA
is defined in the way suggested, then its logic must also be S5, an issue that the
4
definition of physical necessity ought not to prejudge.
At any rate, by the Lindenbaum-Scott Theorem, wherever we have an implica-
tive relation, we have a cognate space of possibilities. The structure of that space
must be determined if we are to attain a clear apprehension of the relevant notion
of necessity, but the Theorem tells us that wherever we have an implicative rela-
tion, some notion of necessity is at least a target of investigation. Since logical con-
sequence is an implicative relation, the cognate notion of logical necessity is, in
particular, such a target. In fact, a famous argument of Carnaps (1946) shows that
where the relevant notion of logical consequence is (a)restricted to propositional
logic and (b)given by the classical truth-tables (including the assumption of biva-
lence), the logic of the cognate notion of logical necessity is S5. That is, where L
signifies the species of logical necessity cognate to classical truth-tabular conse-
quence, and L signifies the dual notion of logical possibility, all instances of the
schemata L A A, A L L A, and L A L L A hold true.
5
Parsons also observes, though, the modal logic of this L is not axiomatizable.
The first-order formulae that are not valid are not recursively enumerable, so
there can be no proof procedure for logical validity in the extended language
that contains L. Since my concern in this book is with the propositional cal-
culus, not first-order logic, Ishall not try to assess how serious a problem the
6
lack of an axiomatization is.
The method just sketched for attaching sense to L A , when A is a
well-formed formula of the language of first-order logic, does not yet extend
4
For this point, see van Fraassen 1977. Van Fraassen shows how the relative necessities may, in a
sense, be reduced to absolute necessity in a two-dimensional modal framework.
5
For a lucid exposition of Carnaps proof, see Burgess 1999, 3.
6
See Parsons 2008, 87ff. for an interesting discussion.
Logical Necessity 77
formal properties. Again, though, we need not delve into these niceties for pre-
sent purposes.
As advertised in my informal sketch at the start of this section, our analysis
enables us to say something about the relationship between logical necessity and
any species of necessity that is generated by an implicative relation. Any necessity
operator that is obtained from an implicative relation in the indicated way
will be normal with respect to . By this Imean that whenever A 1,...,An B,
A1,...,An B. This result follows directly from the Lindenbaum-Scott
Theorem. By the Theorem, whenever A 1,...,An B, B is true at every possibility
in at which all the Ai are true. So B must be true at every possibility in at
which all the Ai are true. In particular, then, whenever Bthat is, whenever
the null set of premisses stands in to Bwe have B. In other words, a thesis
of the implicative relation will exhibit the cognate form of necessity.
As before, let us use L to signify logical necessity; let us also use L to signify
logical consequence (in the classical propositional calculus). Astatement is true at
every logical possibility just in case it follows logically from the null set of prem-
isses, so we have that L A is true if and only if L A. Now when a conclusion
follows logically from some premisses, those premisses stand to the conclusion
in an arbitrary implicative relation. In particular, then, L A implies A for an
arbitrary implicative relation . As we saw in the previous paragraph, though,
whenever A we have A. Putting these results together, then, we reach the
following conclusion:whenever L A is true, A is true. In other words,
the bare truth of their conclusions. As we have seen, however, we apply many of
our deductive capacities to suppositions, and the implicative relations to which
such exercise of those capacities answer will not be Philonian; these relations
7
will generate notions of necessity stronger than simple truth. Now by virtue of
normality, any implicative relation will preserve its cognate notion of necessity.
Thus a conclusion deduced using geometrically correct rules from geometrically
necessary premisses will itself be geometrically necessary; a conclusion deduced
via physically correct rules from physically necessary premisses will itself be
physically necessary; etc. On the conception of logic recommended here, logical
laws are higher-order laws that can be applied to expand the range of any deduc-
tive principles. However, the expanded deductive principles that result from
their application must retain the property of preserving the requisite species of
necessity:even after our geometrical rules have been expanded by an admixture
of logic, we want it to remain the case that the result of applying them to geomet-
rically necessary premisses will itself be geometrically necessary, and so forth.
Logical laws will possess this desirable feature, however, only if logical neces-
sityviz., the notion of necessity that is cognate to themis at least as strong as
all the notions of necessity that correspond to the subject-specific implications.
Logical necessity, then, must be at least as strong as geometrical necessity, physi-
cal necessity, and all the other notions of necessity that arise from implicative
relations. That is, it must be a maximally strong member of the class of implica-
tive notions of necessity. As characterized here, logical necessity has that desir-
able feature. What is more, our account of logic as comprising the laws of the
laws of implication enables us to see why it helps to have a notion of necessity that
possesses this feature. In this way we have explained for what purposes neces-
sity is an essential ingredient of consequence, and thereby vindicated Aristotles
Thesis. It is vindicated, moreover, in precisely the version in which we formu-
lated it:whenever a conclusion B follows logically from some premisses A 1,...,An
and where B expresses the thought that Q and each Ai expresses the thought that
8
Pi, it is logically necessary that Q if P1 and ... and Pn.
7
Stronger, that is, if the relevant implicative relation preserves truth. Only if it does so will the
corresponding notion of necessity imply truth.
8
Hartry Field has recently tried to cast doubt on Aristotles Thesis. He observes that the obvious
attempt to prove the Thesis involves two logical principlesthe rules for introducing and eliminat-
ing the conditionaland two semantic principlesthe rules for introducing and eliminating the
truth-predicate. The Curry Paradox, he then points out, shows these four rules to be jointly incon-
sistent (see e.g. Field 2008, 19.2). The argument for Aristotles Thesis that Field seeks to undermine
is different from mine, so Ineed not address his concerns here. My view, though, is that the Curry
Paradox does not show that the four rules are jointly inconsistent. Logical rulesand, for that mat-
ter, semantical rules for introducing and eliminating the truth-predicatepertain to statements,
Logical Necessity 79
By the same token, belief that a truth is logically necessary amounts to belief that
it is true, no matter what is supposed to be the case. As McFetridge remarks, the
underlying idea may be found in Mill, according to whom that which is neces-
sary, that which must be, means that which will be, whatever supposition we
9
make with regard to other things.
In assessing McFetridges proposal, we need to recognize that there are two
rather different kinds of supposition. Ernest Adams (1975) observed that an indic-
ative conditional often diverges in truth value (or in assertibility) from the cor-
responding counterfactual. If Oswald did not shoot Kennedy, someone else did
is almost certainly true. If Oswald had not shot Kennedy, someone else would
have is probably false. There are, analogously, two kinds of supposition. Suppose
that Oswald did not shoot Kennedy sends us off in one direction:in that case, the
Warren Commission got things hopelessly wrong. Suppose that Oswald had not
shot Kennedy sends us off (more speculatively) in another:in that case, America
would not have got embroiled in Vietnam. Afull description of the differences
between the two kinds would be complicated but, to a first approximation, the
crux is that, in elaborating a supposition of the first kind, we may invoke any fact
that is taken to be established in the context of discussion (such as the content of
the Warren Commission report); we are, by contrast, restricted in which estab-
lished facts may be drawn upon in tracing out the implications of a supposition of
the second kind. (The difficulty of saying which established facts are in this way
i.e., to declarative sentences that (in the relevant context of utterance) express complete thoughts.
Ihold that paradoxical utterances do not succeed in expressing thoughts and hence fail to qualify
as statements. For a defence of this view, albeit for the Liar rather than the Curry, see Rumfitt 2014a
and the articles cited in Chapter1, n.22. For a parallel solution, whereby paradoxical utterances are
taken to express multiple thoughts rather than no thoughts, see Read 2010. Read shows that, on his
assumptions, Curry-like paradoxes demand no revisions to the laws of classical logic.
9
Mill 1891, Book III, chap. v, 6. Mill was characterizing causal necessity.
80 The Nature of Logic
co-tenable with a given antecedent lies at the heart of the problem of specifying
truth-conditions for counterfactual conditionals.)
When McFetridge writes that to count a mode of inference as logically neces-
sarily truth-preserving is to be prepared to employ it in reasoning from any set
of suppositions, which sort of supposition does he have in mind? Both. In a pas-
sage of argumentation to which we shall soon return, he says that on the basis
of a deduction of Q from P one is entitled to assert the conditional, indicative
or subjunctive, if P then Q (1990a, 138; his emphasis). It is certainly plausible to
contend that a thinker will be disposed to use a mode of inference that he counts
as logically valid in tracing out the implications of either sort of supposition. If
Iaccept the rule of or-introduction as logically valid, then Ishall be prepared to
reason in both of the following ways:Suppose that P is the case; on that supposi-
tion, it is also the case that either P or Q and Suppose that P had been the case;
on that supposition, it would also have been the case that either P or Q. Prima
facie plausible as both contentions are, though, they face different challenges, and
it is the claim that logically valid rules may always be applied in tracing out the
implications of counterfactual suppositions that is directly relevant to the pre-
sent discussion. Ihave been treating logical necessity as a modality, expressed by
a statement-forming operator it is logically necessary that or, in formal dress,
L. There is a close correlation between any necessity operator and a correspond-
ing species of counterfactual conditionals:where the modal operator and the
counterfactual conditional pertain to the same space of possibilities, we
can show on weak logical assumptions that A is true if and only if B A
is true for all antecedents B (see e.g. Williamson 2007, 159). In particular, then,
that excludes as impossible all circumstances that conflict with Newtons laws
of motion. Just for that reason, though, the argument Suppose a force is acting
on a body; suppose also that Newtons laws of motion are false; in that case, the
body will be accelerating cannot sensibly be assessed against R. The introduction
of the new suppositional premiss creates a new argumentative context in which
deductions must be assessed against a different implication relation. One mark of
properly logical consequence, by contrast, is that introducing new suppositions
cannot dislodge it as the standard for assessing deductions. This applies even
when the supposition in question is contrary to an accepted logical law. Aclas-
sical logician who reduces to absurdity the supposition that (P P) need not
consider how his logic might change if Excluded Middle had a false instance. In
his reductio, he will apply the normal classical inference rules without demur. Of
course Ido not mean to imply that Excluded Middle cannot be challenged:Ishall
examine a number of challenges to this law in the course of Part II. The point,
though, is that a thinker who accepts classical logic will apply its constitutive
rules to all suppositions whatever, even to suppositions that are contrary to clas-
sical logical laws. That is what it means to accept the classical rules as ones logic.
In denying that logical necessity entails knowability a priori, Iam at odds with
Dorothy Edgington. Pre-Kripkean discussions of validity, she remarks, madeus
familiar with two thoughts:first, an argument is valid if and only if it is necessary that the
conclusion is true if the premisses are true; and second, if an argument is valid, and you
accept that the premisses are true, you need no further empirical information to enable
you to recognize that the conclusion is true ... Given Kripkes work, and taking necessary
in its metaphysical sense, these two thoughts are not equivalent. (Edgington 2004, 9)
So we have to choose what the criterion for validity is to be. It is, she claims, the
least departure from traditional, pre-Kripkean thinking, and more consonant
with the point of distinguishing valid from invalid arguments, to take validity to
be governed by epistemic necessity, i.e., an argument is valid if and only if there is
an a priori route from premisses to conclusion (2004, 10).
Edgington tells us that she does not mean only formally valid argu-
ments:Iinclude cases such as Its round; so its not square (2004, 89). Even
taking this point, however, Ido not think that her positive account of validity is
adequate. First, there are statements which we can know a priori but which we are
reluctant to classify as logically true even when we do not require logical truths
to be true in virtue of their form. Some people now know a priori that, when
the index n is greater than two, there are no integral solutions of the equation
xn + yn=zn. All the same, it is not simple prejudice to resist the claim that Fermats
Last Theorem is logically true. The ground for resistance is not the complexity of
the proof:there are long and complex logical deductions. Rather, it is the heavy
ontological and ideological commitments of the mathematical theories on which
the proof depends. There are good reasons for postulating the truth of those the-
ories:if there were not, then we should not have a proof of the theorem. Those
theories, however, go far beyond anything that it required for the regulation of
deduction. For that reason, we should classify Fermats Last Theorem as a math-
ematical truth but not as a logical truth.
Perhaps Edgington would respond to this point by emending her position,
and proposing that an argument is logically valid (in her broad sense) if and only
if there is a route from its premisses to its conclusion that a thinker may trav-
erse purely by exercising those strictly deductive capacities that do not depend
on obtaining particular empirical information about the world; strictly deduc-
tive capacities is taken to exclude the ability to construct complex mathematical
theories. One might propose in the same spirit that an argument is formally valid
if and only if a thinker may move from premisses to conclusion by exercising
only his ability to reason with logical notions such as conjunction, disjunction,
and negation. However, this emended account faces a second objection. On the
Logical Necessity 83
10
For an elaboration and defence of this account of metaphysical modality, see Fine 2002.
84 The Nature of Logic
analysis of the previous section. We saw there that logical necessity is at least as
strong as any notion of necessity that is generated from an implicative relation,
but it is not obvious that metaphysical necessity is such a notion. So we have to
address the question on its merits.
As Iremarked, the recommended account of logical necessity is a close rela-
tive to that of Ian McFetridge, who advanced an interesting general argument
purporting to show that logical necessity is the strongest form of non-epistemic
11
necessity. It makes sense to ask, then, whether his argument can be adapted
to the present account of logical necessity. McFetridges argument rests on two
assumptions. First, that adding extra premisses to a [logically] valid argument
cannot destroy its validity ... If the argument P; so Q is valid then so is the
argument P, R; so Q for any R. Second, that there is this connection between
deducing Q from P and asserting a conditional:that on the basis of a deduction
of Q from P one is entitled to assert the conditional, indicative or subjunctive, if
P then Q (McFetridge 1990a, 138; emphasis in the original). The argument then
runs as follows. Suppose it is logically necessary that if P then Q. Suppose also, for
reductio, that in some other sense of necessary, it is not necessary that if P then Q.
Then, in the sense of possible that corresponds to this other sense of necessary,
it is possible that P and not Q.But
if that is a possibility, we ought to be able to describe the circumstances in which it would
be realized:let them be described by R. Consider now the argument P and R; so Q. By the
first assumption if P; so Q is valid, so is P and R; so Q. But then, by the second assump-
tion, we should be entitled to assert:if P and R were the case then Q would be the case. But
how can this be assertible? For R was chosen to describe possible circumstances in which
P and not Q. Ithink we should conclude that we cannot allow, where there is such an R,
12
that an argument is valid.
11
Thus he expressly excludes the notion of necessity that corresponds to mere time- and
person-relative epistemic possibility (1990a, 137). Ifollow him in setting this notion aside.
12
McFetridge 1990a, 1389. Like McFetridge, Ishall not mark use versus mention where there is
no danger of confusion.
Logical Necessity 85
the introduction of a new premiss or supposition can sometimes force a shift from
one implicative relation to another, and such a shift can affect the assertibility of
a conditional. In a context where only Euclidean possibilities are contemplated,
we shall be able to assert If T is right-angled, then the square on its hypotenuse is
equal to the sum of the squares on its other two sides, but it does not follow that
we shall be able to assert If T is right-angled and its internal angles sum to less
than two right angles, then the square on its hypotenuse is equal to the sum of the
squares on its other two sides:the antecedents second conjunct forces us to con-
template non-Euclidean possibilities. Given our account of logical consequence,
however, no such shifting is possible in that special case. We can therefore accept
McFetridges conclusion that, if the argument P; so Q is logically valid, so is the
argument P, R; so Q.
As for his second assumption, it is highly plausible to claim that we may assert the
13
indicative conditional If P then Q on the strength of a deduction of Q from P. The
corresponding claim for subjunctives is again very plausible. As it concerns sub-
junctives, the second assumption amounts to this:that we can apply our capacity
for logical deduction in elaborating a counterfactual supposition. The worry about
not accepting this assumption is that if we were not able to apply that capacity, we
should be quite unable to elaborate counterfactual suppositions at all. Of course, in
elaborating a given counterfactual supposition, some of our deductive capacities
will not be applicable. Acapacity for deducing the consequences of suppositions
according to the principles of classical physics, for example, is quite inapplicable
in elaborating the counterfactual supposition Suppose that the gravitational force
between two bodies had varied with the inverse cube of the distance between them.
Just for that reason, though, we badly need some rules which are guaranteed to
yield consequent elaborations of our counterfactual suppositions. Since logic is tra-
ditionally supposed to apply to anything that is so much as thinkable, one would
expect logical rules (even rules that are logically valid in the broad sense) to provide
14
what we need. If they do not, it is wholly unclear what else does.
13
Note that the rule to which implicit appeal is made here is weaker than that of Conditional
Proof. The general form of Conditional Proof says that, if we have deduced Q from P together with
side premisses X, then we may deduce the conditional If P then Q from the side premisses X alone.
This rule does not apply to subjunctive conditionals. If it did, we could derive from the premiss I
cycled to work this morning the conclusion If Ihad broken my leg last night, Iwould have cycled
to work this morning. But all McFetridges argument needs is the restricted version of Conditional
Proof without side premisses, and it is plausible to maintain that our reasoning with subjunctive
conditionals conforms to that weaker rule.
14
Those concerned to formalize the logic of counterfactual conditionals have, indeed, proposed
rules which subsume the counterfactual part of McFetridges second assumption. Thus it is a special
case of both David Lewiss rule of Deduction within Conditionals (Lewis 1986a, 132)and Timothy
Williamsons rule of Closure (Williamson 2007, 293).
86 The Nature of Logic
All the same, the two assumptions combine to yield consequences which some
commentators have found unpalatable. In support of her account of (broad) logi-
cal validity, Edgington elaborates Kripkes example of Leverrier, who postulated a
nearby, hitherto unobserved planet as the cause of certain observed perturbations
in the orbit of Uranus, and who introduced the name Neptune as a term which was
to stand (rigidly) for such a planet, if indeed there was one (Kripke 1980, 79). About
this case, she remarks that it is epistemically possible that [Leverriers] hypothesis
was wrongthat there is no such planet. But if his hypothesis is rightif Neptune
existsit is the planet causing these perturbations. And this conditional is known
a prioriat least by Leverrier: it follows from his stipulation about the use of
Neptune (Edgington 2004, 7). Now consider the following argument, as it comes
from Leverriersmouth:
(A)Neptune exists
Therefore:Some planet is the cause of the observed perturbations.
Although I have denied that an arguments validity follows from the exist-
ence of an a priori route from its premisses to its conclusion, we may still be
tempted to follow Edgington in classifying (A) as (broadly) valid, and corre-
spondingly tempted to ascribe (broad) logical necessity to the conditional If
Neptune exists, then some planet is the cause of the observed perturbations.
For although (A) is not formally valid, Edgington seems to be right in describ-
ing it as trivial (2004, 7), and it is tempting to assume that a trivially cor-
rect argument must be valid in the broad sense that Edgington is trying to
elucidate.
On McFetridges principles, though, it seems that (A) cannot be valid, even in
the broad sense. Certainly, he took himself to be committed to denying its valid-
ity. Discussing an ancestor of the paper of Edgingtons from which Ihave been
quoting, he writes:
Following Kripke and Evans Edgington claims, and Iagree, that [Leverrier] knows
a priori that if Neptune exists it is a planet causing such and such perturbations.
Thus, on her account, the argument:Neptune exists, so Neptune causes such and
such perturbations is deductively valid. But there certainly is a timeless meta-
physical possibility that the premiss should have been true and the conclusion
false:suppose Neptune had been knocked off course a million years ago. What then,
of the argument:Neptune exists and was knocked off course a million years ago, so
Neptune is the cause of these perturbations? If the original argument is valid so is
this one (by the first assumption). But if it is we ought (by the second assumption)
to be entitled to assert:if Neptune had existed and been knocked off its course a
million years ago then it would have been the cause of these perturbations. But of
Logical Necessity 87
course we are not entitled to assert that:had the antecedent been true the conse-
15
quent would have been false.
McFetridges two assumptions, then, reduce to absurdity the claim that argument
(A) is broadly logically valid. Given the way the name Neptune was introduced,
however, the inference in (A) is trivial, and it seems hard to deny broad logical
validity to a trivial inference. Something in the analysis seems to have gone seri-
ously wrong. But what could it be?
The problem is not confined to this one example. As Edgington remarks, parallel
cases may be constructed whenever we have what Gareth Evans called a descrip-
tive name. Evanss own example was Julius, which he introduced as a descriptive
name that rigidly designates the person (if there was one) who actually invented the
zip fastener (see Evans 1979). Thus Edgington invites us to consider the argument:
As before, the triviality of the inference here makes it tempting to classify (B) as
valid in the broad sense. Yet there are metaphysically possible situations in which
the premisses are true and the conclusion false, namely, ones in which Julius, the
actual inventor of the zip fastener, did not do so and someone else, who emigrated
to Tahiti, did, and no mathematician emigrated to Tahiti, and Julius was a math-
ematician (Edgington 2004, 910). Contrary to McFetridges master thesis, then,
we seem to have a case of a logically necessary statement that is not metaphysi-
cally necessary.
Edgington tries to resolve the difficulty by rejecting the subjunctive part of
McFetridges second assumption. We are familiar with the fact that an indicative
and a subjunctive If A, B can disagree, shesays.
However, his second assumption does not require a unitary space of possibilities.
All it requires is that broadly logical deduction should be applicable in elaborat-
ing both epistemic possibilities and the once real possibilities that Edgington
15
McFetridge 1990a, 139. See also McFetridge 1990b for an interesting explication of a priori
knowledge that vindicates the claim that Edgingtons conditional is knowable a priori.
88 The Nature of Logic
takes subjunctives to present. Edgington does not explain why broadly logical
deduction should be inapplicable to the latter cases, nor does she tell us what
deductive principles we can rely on if it is not applicable. In the absence of such
an explanation, we are left with an aporia, perhaps even with a paradox.
the zip if there was such a person and which otherwise does not stand for any-
thing at all.
We are to take it as given, then, that the name Neptune stands for the planet
that causes the perturbations if there is a unique such planet. This means that any-
one who, like the proponent of (A), uses the name presupposes that, if Neptune
exists, it is the planetary cause of the perturbations. Accordingly, any implication
relation against which (A) is to be assessed will relate Neptune exists to Neptune
is the planet that is the cause of the observed perturbations. Now this second rela-
tum logically entails Some planet is the cause of the observed perturbations and,
as we have seen, the laws of logic may be applied to extend any implication rela-
tion. So the relation against which (A) is to be assessed will relate Neptune exists
to Some planet is the cause of the observed perturbations. That is to say, the rela-
tion against which (A) is to be assessed will relate the premiss of that argument to
its conclusion. Since all these points will be evident to anyone who understands
the name Neptune in the way we are to suppose that it is understood, we can
account for the immediate evidence of (A)s correctness without postulating any
logical relationship between its premiss and its conclusion, even in the broadest
legitimate sense of logical. For we can account adequately for (A)s trivial cor-
rectness by reference to the argumentative context in which we are to imagine it
being propounded.
It may help to reformulate the explanation just given in terms of possibilities
rather than implication relations. In any context in which (A) is trivially correct,
use of the name Neptune presupposes that, if Neptune exists, it is the planetary
cause of the perturbations. So, in assessing (A) in such a context, we shall con-
fine ourselves to a space of possibilities that excludes all those in which Neptune
exists without causing the observed perturbations. In particular, then, the relevant
space excludes the possibility that McFetridge and Edgington mentionin which
Neptune exists but in which it was knocked off course millions of years before
it had the chance to cause any of the perturbations that Leverrier observed. The
point, though, is not that this is an example of a metaphysical possibility that is
not a logical possibility. It clearly is a logical possibility:no logical lawindeed, no
conceptual truthexcludes Neptunes having been knocked off course millions
of years go. The point, rather, is that the metaphysical and logical possibility that
Neptune was knocked off course falls outside the space of possibilities that is rel-
evant to the assessment of (A). It falls outside that space because it is excluded by a
presupposition of one of the terms used in (A)viz., the name Neptune. Properly
analysed, then, argument (A) offers no support whatever for the thesis that there
are metaphysically possible states of affairs which are not logically possible. It is
just one more illustration of the way in which the implication relation against
90 The Nature of Logic
which a deduction is assessed varies with the deductions context. Accordingly, the
apparent problem that (A) presents for McFetridges thesis dissolves.
A similar explanation accounts for the trivial correctness of argument (B).
There are metaphysical and logical possibilities in which Julius exists but did
not invent the zip. Some of these are possibilities at which both premisses of
(B) are true but its conclusion is false. Accordingly, the premisses of (B) neces-
sitate its conclusion neither metaphysically nor logically. All the same, we can
account for the fact that (B) is a trivially correct deduction. Aspeaker who
propounds (B) must be using the descriptive name Julius; his use of that term
(in the relevant sense) presupposes that, if Julius exists, it was he who invented
the zip. So, although there are metaphysical and logical possibilities in which
Julius existed without inventing the zip, they are excludedand are imme-
diately known to be excludedfrom the space of possibilities that is relevant
to assessing (B). Again, then, we may account for the trivial correctness of (B)
without supposing that that deduction is logically valid, even in the broadest
sense.
Edgingtons examples, then, fail to show that some metaphysical possibili-
ties are logically impossible, but they have an interesting moral. Philosophers
have long been familiar with the claim that ordinary proper names are
object-involving. Use of the name Bertrand Russell, for example, presup-
poses the existence of Bertrand Russell. Assuming that Russell existed only
contingently, the claim entails that use of the name presupposes a logical and
metaphysical contingency. David Wiggins and others have claimed that some
natural kind terms are in a similar way kind-involving or species-involving (see
Wiggins 1993). Use of the term gold, for example, presupposes the existence of
samples of gold. Assuming that such samples exist only contingently, this latter
claim entails that use of the natural kind term also presupposes a contingency.
Edgingtons cases show that subtler examples of this phenomenon may be
found even with expressions that are not object-involving or species-involving.
Leverriers use of the name Neptune did not presuppose that there was such
a planet as Neptune:the stipulation by which he (is imagined to have) intro-
duced the name was hedged to allow for the possibility that the perturbations
he observed had some non-planetary cause. His use of that name, though, did
presuppose that, if Neptune exists, it is the planet causing the perturbations;
that presupposition is also a logical and metaphysical contingency.
Examples of this phenomenon are not confined to ordinary proper names,
terms for natural kinds, or names introduced by stipulation such as Neptune
and Julius. Consider the argument (C):
Logical Necessity 91
On one common way of understanding red, argument (C) is also trivially cor-
rect:anyone who understands the terms in it will accept that its premiss implies
its conclusion. The explanation is similar to that for (A) and (B):those speakers
who find (C) trivial presuppose that red things look red to normally sighted view-
ers in optimal viewing conditions. As before, though, that presupposition is a
contingency. Human beings could have been constituted so as to see red things
as violet, and violet things as red, in which case red things would not have looked
16
red, even in optimal viewing conditions.
The resolution of the paradox that I am proposing may be summarized as
follows:
(1) We should accept McFetridges thesis that logical necessity is the strongest
form of non-epistemic necessity. Any non-epistemic possibility is a logical
possibility. Moreover, McFetridges argument for his thesis is essentially
correct.
(2) When properly analysed, the problematical arguments (A), (B), and (C) do
not threaten this thesis. Edgington and others are right to hold it to be meta
physically possible that Neptune should have existed without causing the
perturbations that Leverrier observed. But this is no threat to McFetridges
thesis, for it is also logically possible that Neptune should have existed with-
out causing those perturbations. This means that argument (A) is not logically
valid, even in a broad sense. We shall be tempted to think otherwise only if we
overlook the way in which the standards for assessing a deductions soundness
17
adjust to accommodate the presuppositions of the terms used in it.
It is important for the project of this book to settle whether Aristotles
Thesis is true. We cannot assess arguments for or against rival logical systems
without a clear conception of what logical consequence comes to. Moreover,
16
Other problem cases involve the word actually. Icannot discuss these cases here, but for an
account of the meaning of the word with which Iam sympathetic, and which shows that these cases
too are only apparent problems for the traditional assumption, see Bostock 1988.
17
In a previous essay on logical necessity (Rumfitt 2010), I gave a different account of why
Edgingtons examples do not refute McFetridges claim that any metaphysical possibility is a logical
possibility. When descriptive names such as Neptune and Julius are in play, Iargued, we have to
distinguish between two modal relativizations of truth:truth at a possibility and truth in a pos-
sibility. This generates in turn a distinction between two notions of validity. There are logical and
metaphysical possibilities at which (for example) the premiss of argument (A) is true while its con-
clusion is false, so (A) is logically invalid in the sense that truth is not preserved at every logical pos-
sibility. However, the same deduction is logically valid in the sense that truth is preserved in every
92 The Nature of Logic
acceptance of the Thesis shapes ones approach to the other central problem
in adjudicating between such systemsnamely, that of specifying the senses
of the connectives. The sense of a word is the contribution it makes to the log-
ically relevant part of the content of a statement in which it occurs. Given that
logical modalities are implicit in the notion of logical consequence, an ade-
quate specification of a connectives sense must say how it helps to determine
at which logical possibilities statements in which it figures are true. With
so much clear, I turn to the main business of the book and scrutinize five
attempts to displace the laws of classical logic from their position as boundary
stones of thought.
logically possibility:there is no logical (or metaphysical) possibility in which its premiss is true
while its conclusion is not true.
I no longer like this diagnosis of the fallacy in Edgingtons objection to McFetridge. While more
intricate deductions may well call for a distinction between validity as preservation of truth at all
possibilities and validity as preservation of truth in all possibilities, it now strikes me as implausible
that such a distinction is needed in evaluating simple arguments like (A) and (B). Accordingly,
Itake it to be a merit of the present diagnosis of the fallacy that it does not posit an ambiguity in the
term valid. All the same, Icontinue to accept the distinction between truth at a possibility and
truth in a possibility, and Istill believe that this affords us the best account of Evanss distinction
between the superficial and deep forms of (each variety of) necessity. (See Evans 1979 and 2004,
Davies and Humberstone 1980.) Ialso maintain that the distinction enables us to identify the fal-
lacy in Dummetts counterexample (in his 1993a) to the thesis that S5 is the logic of metaphysical
necessity. On these matters, see 4 and 5 of Rumfitt 2010.
PA RT I I
statement is either true or false; that is, Tertium non datur classically entails the Principle
of Bivalence. So, if classical logic could be applied to semantic discourse, a simple deduc-
tion would enable us to move from our knowledge of Tertium non datur to knowledge of
Bivalence. Dummett argues, however, that other conceptual constraints on the notions
of truth and falsity mean that we cannot know the Principle of Bivalence:
We are entitled to say that a statement P must be either true or false ... only when P is a state-
ment of such a kind that we could in a finite time bring ourselves into a position in which we
were justified either in asserting or in denying P; that is, when P is an effectively decidable
statement. This limitation is not trivial:there is an immense range of statements which, like
Jones was brave [said of a man, now dead, who never encountered danger in his life], are
1
concealed conditionals, or which, like A city will never be built here, containexplicitly or
implicitlyan unlimited generality, and which therefore fail the test. (Dummett 1959, 1617)
But if Tertium non datur is knowable while Bivalence is not, then classical logic is
not applicable to semantic discourse. A fortiori, it is not applicable universally. The
argument also suggests that within semantics, at least, it ought to be replaced by
a logic cast from an intuitionistic mould, for in intuitionistic logic the deduction
fails. Using the obvious symbolization, and with the domain restricted to state-
ments, an intuitionist may move from x(Tx Fx) to x (Tx Fx) .
However, the rule for eliminating double negations can be applied only to an effec-
tively decidable statement, and not all of the matrix instances Ta Fa qualify as
such. So the intuitionist cannot complete the deduction and reach the conclusion
x(Tx Fx) . That is, he cannot complete the inference to Bivalence.
In this argument, the statements Jones was brave and A city will never be
built here (uttered in some rural location) are not advanced as counterexam-
ples to Bivalence. Acounterexample would be a statement which was neither
true nor false, and the existence of such a thing is precisely what Tertium non
datur excludes. Rather, they are put forward as cases in which we are not enti-
tled to say that a statement must be either true or false. Since we are entitled (in
the relevant sense) to say anything that we know, Itake it that they are also put
2
forward as statements which we do not know to be either true or false. Even if
1
According to Dummett, Jones was brave is best understood as abbreviating the counterfac-
tual conditional Had Jones faced danger, he would have acted bravely.
2
John McDowell stresses this consequence of Dummetts position; it implies not that the prin-
ciple of bivalence is false but, at most, that it is not known to be true (McDowell 1976, 59).
It is clearly essential to his argument that Dummetts examples should be statements in the
present, restricted sense; that is, they should be neither ambiguous nor vague. The terms brave
and city are vague. As we shall see, though, his argument for the claim that we are not entitled to
assert the bivalence of certain statements containing these terms does not rest on their vagueness.
We may think, then, of Dummetts examples as proxies for statements from which vagueness and
ambiguity have been eliminated, perhaps by stipulation.
The Argument of Dummetts Truth 97
it could be sustained, this conclusion may not be enough to refute the universal
validity of classical logic:my defence (in 2.6) of the thesis that it should always
be possible to apply logic in extending our knowledge through deduction was
tentative. But Dummetts conclusion, if correct, would certainly restrict the
application of classical logic in an areanamely, semantic theorythat is of
great logico-philosophical interest, so his argument is well worth scrutinizing
closely.
How does Dummett argue for Tertium non datur? A statement, he explains,
so long as it is not ambiguous or vague, divides all possible states of affairs into just two
classes. For a given state of affairs, either the statement is used in such a way that a man
who asserted it but envisaged that state of affairs as a possibility would be held to have
spoken misleadingly, or the assertion of the statement would not be taken as expressing
the speakers exclusion of that possibility. If a state of affairs of the first kind obtains, the
statement is false; if all actual states of affairs are of the second kind, it is true. It is thus
prima facie senseless to say of any statement that in such-and-such a state of affairs it
would be neither true nor false. (Dummett 1959, 8)
The argument rests, then, on what we may call an exclusionary account of con-
tent:a statements sensethat is, the part or aspect of its content that is relevant
to logicis given by the possibilities that it excludes. For reasons to be given in
4.2, this view of a statements sense has many merits. Given the exclusionary
conception as backdrop, the arguments premisses then concern the relationship
between the possibilities that a statement excludes, or leaves unexcluded, and its
falsity or truth.
In assessing Dummetts argument, it helps to formalize it, so let us introduce
the notation S to mean is a statement, O to mean is a state of affairs which
actually obtains, E to mean The statement excludes the possibility , T to
mean is true, and F to mean is false. In this notation, the aim of the argu-
ment is establish Tertium non datur in the form
x (Sx (Tx Fx ))
The method is to reduce to absurdity the contrary supposition that some state-
ment is neither true nor false.
The argument has two substantial premisses. The first says that if a state of
affairs obtains which a statement excludes, then that statement is false:
) x Sx (y (Oy Exy ) Fx ) .
98 Five Attacks on Classical Logic
The second premiss says that if no state of affairs that a statement excludes actu-
ally obtains, then that statement is true:
) x Sx (y (Oy Exy ) Tx ) .
3
The De Morgan law used at line 5 is acceptable to an intuitionist, as are the form of contra-
position used at line 10 and the conversion principle used at line 11. As for the principle used to
reach line 13, perhaps the easiest way to see its intuitionistic acceptability is to apply the famil-
iar equivalence between A and A . By -introduction, we have A, B A B, so that
A, B, (A B) , by -elimination. One application of -introduction then yields
A, (A B) B , and another yields (A B) A (B ), which by the equivalence
is (A B) A B, as required.
The Argument of Dummetts Truth 99
These reasons for adopting the exclusionary theory are less than conclusive,
however. Outside rather special forums, such as parliaments or courts of law,
in which an adjudicating authority has effective control over what people say,
it is strained to claim that a speaker must withdraw an assertion that has
been shown to be false. On the contrary, it is a depressing feature of ordi-
nary, unregulated conversation that some people will go on asserting things
that have been conclusively shown to be wrong. As for the example of con-
ditionals, it is indeed striking that our understanding of them is not greatly
impeded by widespread disagreement about their truth-conditionsinclud-
ing disagreement over whether they have truth-conditions. However, even if
that understanding consists in our all knowing that a speaker who asserts a
conditional means to rule out all those possibilities in which the antecedent
4
is true and the consequent is false, a single favourable case can hardly show
that the senses of statements in general are given by the possibilities that they
4
Which Idoubt:see the remarks on conditionals at the end of 4.5.
100 Five Attacks on Classical Logic
exclude, especially when the favourable case may not even be one in which a
single propositional content is propounded. For, despite Dummetts attack
on the view, many philosophers maintain that in affirming a conditional, a
speaker does not assert any propositional content outright, but instead con-
ditionally asserts the consequent within the scope of a supposition that the
5
antecedent is true.
In a later paper, What is a Theory of Meaning? (II), Dummett advances a
different, and more general, argument for an exclusionary account of content.
According to Dummett, a statements free-standing content is always the content
of a potential assertion of it. An assertion, he proceeds to remark,
is not, normally, like an answer in a quiz programme; the speaker gets no prize for
being right. It is, primarily, a guide to action on the part of the hearers (an interior
judgment being a guide to action on the part of the thinker); a guide which operates
by inducing in them certain expectations. And the content of an expectation is deter-
mined by what will surprise us; that is, by what it is that is not in accord with the expec-
tation rather than by what corroborates it. The expectation formed by someone who
accepts an assertion is not, in the first place, characterized by his supposing that one
of those recognizable states of affairs which render the assertion correct will come to
obtain; for in the general case there is no bound upon the length of time which may
elapse before the assertion is shown to have been correct, and then such a supposition
will have, by itself, no substance. It is, rather, to be characterized by his not allowing
for the occurrence of any state of affairs which would show the assertion to have been
incorrect; a negative expectation of this kind has substance, for it can be disappointed.
The fundamental notion for an account of the linguistic act of assertion is, thus, that
of the incorrectness of an assertion:the notion of its correctness is derivative from that
of its incorrectness, in that an assertion is to be judged correct whenever something
happens which precludes the occurrence of a state of affairs showing it to be incorrect.
(Dummett 1976a, 124)
5
For Dummetts attack on this view, see Dummett 1959, 1011; for a reply, see Edgington 1995,
28991.
The Argument of Dummetts Truth 101
claim. Nothing can show the disjunctive assertion Either it is raining, or it has been
shown that it is not raining to be false:any state of affairs that shows the first disjunct
to be false is one in which the second disjunct is true. Accordingly, if we suppose that
an assertions content is given by the states of affairs that would show it to be false,
then this disjunction has the content of a tautology:no possible state of affairs would
show the disjunction to be false. That, however, is clearly wrong. The content of this
disjunctive assertionwhat it saysis not tautologous:the disjunction is false in a
situation where it is not raining but it has not been shown that it is not raining. The
difference between our disjunction and a genuine tautology comes out clearly if one
considers acts of express supposition, rather than acts of assertion, directed at their
negations. No one can coherently suppose that it is raining and not raining, but one
can coherently suppose that it is not raining and it has not been shown that it is not
raining. Dummetts version of the exclusionary theory, then, gets its assignments of
content wrong:it classifies our disjunction as having a tautologous content when it
does not.
All the same, there is a powerful argument in favour of a version of the
exclusionary conception of content that is not subject to these problems. As we
have seen, in the passages where he tries to justify the conception, Dummett
appeals to a pragmatist theory of meaning. In the Postscript to Truth, he
focuses on the consequences of making a false assertion; in What is a Theory
of Meaning? (II), he explains the notion of an assertions content in terms
of the way it guides the actions of someone who understands and accepts it.
Now the states which possess propositional contents and bear most directly
on action are beliefs:how we actor, at least, how we act rationallyis largely
settled by what we believe, along with what we desire. For this reason, prag-
matists typically take the primary bearers of propositional content to be the
particular states of belief that various thinkers are in at various times, and
they explicate the contents of these states by reference to their potential bear-
ing on action. To believe that P, they typically say, is to be disposed to act as if
6
P (at least when it is a contingent matter whether P). Dummetts concern is
with the contents of assertions, rather than states of belief, but since sincere
assertions express beliefs, and since accepting an assertion generates or rein-
forces a belief, we may adapt the formula so that it applies to a statements con-
tent. The content of a contingent statement, it may be proposed, is determined
6
The first clean statement of this idea that Iknow is to be found in Richard Braithwaites The
Nature of Believing (1933, 1323), although Braithwaite acknowledges drawing on similar ideas
in Alexander Bains Mental and Moral Science (see Bain 1868, 372). The formula remains popular
among pragmatists:for a recent reaffirmation of it, see Mellor 2012, 61.
102 Five Attacks on Classical Logic
by the range of actions its acceptance disposes one to perform. In this way, we
reach principle (P):
A contingent statement A expresses the content that Por, for brevity, says that Pby
virtue of the fact that someone who understands and accepts A is disposed to act as if P.
Where the relevant notion of contingency is the logical one, non-contingent state-
ments will be logically complex:their contents may then be constructed from the
contents of their contingent components, along with the semantic contributions
of the connectives.
Principle (P) provides a ground for an exclusionary conception of content.
The key connection between a disposition to act and the exclusion of possibili-
ties was already perceived by Ramsey in Facts and Propositions. In that paper,
Ramsey propounds a pragmatist theory of content:the meaning of a sentence
is to be defined by reference to the actions to which asserting it would lead
7
(Ramsey 1927, 51). In our terms, the content of a statement is given by the effect
that its acceptance would have upon a thinker who understands it. As Ramsey
8
saw, it is hopeless to tie a statements content too directly to the consequences
of accepting it, for those consequences depend upon the thinkers desires and
9
his other beliefs. My accepting the statement That train goes to Oxford may
lead to my boarding the train, but only if Iwant to go to Oxford; moreover,
even if Ido want to go there, Iwill not board that train if Ialso believe that a
faster service to the same destination leaves in five minutes. Ramsey, however,
puts his finger on an aspect of a statement, or a belief, which bears directly on
the actions (or plans for action) of one who accepts that statement, or who has
that belief, but which is nevertheless constant across subjects with different
aims or desires and with different background beliefs:
To say that feeling belief towards a sentence [sc., accepting a statement] expresses such
an attitude [sc., expresses a belief that P] is to say that it [sc., accepting the statement] has
certain causal properties which vary with the attitude, i.e. with which possibilities are
knocked out and which, so to speak, are still left in. Very roughly, the thinker will act in
disregard of the possibilities rejected. (Ramsey 1927, 46)
7
See also Ramsey 1991, 45, where he propounds the cognate view of the contents of beliefs:we
have to analyse the meaning of saying that a man has a belief that such and such is the case, for
instance, that the earth is flat. This we have seen to be partly an assertion about what he would think
or say and partly... one about how he would behave.
8
No particular action can be supposed to be determined by <one> belief alone; <an agents>
actions result from his desires and the whole system of his beliefs, roughly according to the rule that
he performs those actions which, if his beliefs were true, would have the most satisfactory conse-
quences (Ramsey 1991, 45).
9
For elaborations of this point, see the classic attacks by Roderick Chisholm (1957) and Peter
Geach (1957) on logical behaviourism.
The Argument of Dummetts Truth 103
disposition relate to his being disposed to act as if Black is the murderer, and to
his being disposed to act as if Brown is the murderer? The exclusionary elabora-
tion of (P) offers a clear and simple answer. In being disposed to act as if Black
is the murderer, Morse sets aside four of the five hitherto open possibilities:he
excludes Brown, Green, Scarlet, and White. In being disposed to act as if Brown
is the murderer, Morse sets aside a distinct but overlapping foursome:he excludes
Black, Green, Scarlet, and White. The intersection of these sets comprises Green,
Scarlet, and White, precisely the possible murderers that are set aside when Morse
is disposed to act as if either Black or Brown is the murderer. As we shall see in
4.5, given only weak logical assumptions, this argument extends from possible
murderers to possibilities generally so that the set of possibilities excluded by
A or B is the intersection of those excluded by A with those excluded by B. The
exclusionary version of (P), then, opens the way to a solution to the problem of
10
specifying the sense of or within a pragmatist theory of meaning.
In assessing the exclusionary theory, it is important to be clear what it is and is not
saying. An objector may be tempted to protest that it is only because we know what a
statement says that we know which possibilities it excludes:our knowledge of what a
statement excludes rests upon a prior knowledge of what it says. The objectors claim
may well be true, but it does not gainsay the proposal. The exclusionary theory does
not aim to specify what is involved in knowing what a statement says. Rather, it aims
to say what it is for a statement to possess the content that it has. There is no rea-
son to expect the answer to that constitutive, or metaphysical, question to reflect
dependencies among the various aspects of our knowledge of language.
In fact, many thinkers in a variety of fields have found the exclusionary theory
attractive, even if they do not accept the pragmatist principle which provides my
justification for it. The mathematical theory of information incorporates a ver-
sion of it:information theory identifies the amount of information associated
with, or generated by, the occurrence of an event (or the realization of a state of
affairs) with the reduction in uncertainty, the elimination of possibilities, repre-
sented by that event or state of affairs (Dretske 1981, 4; emphasis added). Thus in
a message saying which of eight starters won a race, the amount of information
associated with the result is a function of how many possibilities were eliminated
11
in reaching that result (Dretske 1981, 4). The exclusionary theory is also implicit
10
The reader may worry that this simple semantic principle for disjunction is bought at the price
of rendering intractable the corresponding problem for conjunction. 4.5 will allay that worry.
11
More exactly, the amount of information (in bits) in a message is defined as the logarithm
(to base 2)of the ratio of the number of possibilities open before the message is received to the
number of possibilities open afterwards (it is assumed that the message is correct). Thus a mes-
sage saying which of eight starters won a race contains log 2 8=3 bits of information.
The Argument of Dummetts Truth 105
12
See, for example, the discussion of the Barcan Formula in 6.2.
106 Five Attacks on Classical Logic
recalled, is one which a speaker who asserts the statement is thereby committed
to ruling out from among those that he takes to be live. The statement in question
could have no coherent sense if such a possibility were one at which the statement
was true.
While a statement cannot be true if an excluded possibility obtains, some
philosophers will insist that it may yet be neither true nor false in such a cir-
cumstance. Aspeaker who asserts The present king of France is wise certainly
excludes any state of affairs in which there is no king of France; but Strawson
and others have held that in such a state of affairs the statement is neither true
13
nor false. In Truth, however, Dummett persuasively explains why Strawsons
analysis is wrong (see also Dummett 1960). If we consider the statement on its
own, there is really no doubt that it is false:it represents things as being a certain
way, and things are not that way. It is, no doubt, an example of a rather particu-
lar kind of falsity; but it is still false. There may be reasons for distinguishing
this kind of falsity from the kind exemplified when there is a reigning, but fool-
ish, king of France, but the point of such distinctions does not lie in anything
to do with the sense of the statement itself, but has to do with the way in which
it enters into complex statements (Dummett 1959, 12). In the terminology that
Dummett was later to adopt, such distinctions lie at the level of a statements
ingredient sense, rather than its free-standing assertoric content. This analy-
sis is persuasive when applied to the present example. Any reluctance we may
have to classify The king of France is wise as false in the current circumstances
stems from a wish to regard The king of France is not wise as a negation of the
original statement, together with the recognition that The king of France is not
wise also excludes any state of affairs in which there is no king of France. That
wish can be gratified if we deem simple statements involving empty descrip-
tions to possess a neuter alethic status, and allow a form of negation whereby
neuter statements have neuter negations. All the same, on such a treatment we
may, and should, insist that this neuter status is a way of being false. For when
there is no king of France, the statements The king of France is wise and The
king of France is not wise both misrepresent the way things are, and so are cor-
rectly classified as false.
We may, then, accept Dummetts premiss . What, though, of premiss , which
says that, if no possible state of affairs that a statement excludes actually obtains,
then the statement is true?
13
What Strawson actually said was that the question of whether [the] statement is true or false
simply does not arise (Strawson 1950, 12). But if, nevertheless, the question were raised, he would
have thought it wrong to return either of the expected answers.
The Argument of Dummetts Truth 107
also seems correct. Dummetts requirement of practical relevance does not stop
his theory from delivering the desired results in this case.
When applied to other examples, however, principle is far less plausible.
Consider the statement that is made when somebody says in One day, a city
will be built here, while leaving the time within which his prediction is to come
true completely unbounded. The possibilities this statement excludes comprise
those whose obtaining would permanently prevent the building of a city:that a
nearby dike is breached and the spot irretrievably flooded; that a nuclear bomb
is detonated so that the place is permanently irradiated; and so forth. We can
certainly suppose that no possibility of this kind actually obtains, but it is implau-
sible to claim on this basis that the statement is true. If no such state of affairs
obtains, then the possibility of building a city at the relevant place will remain
open; but that is not enough to render One day, a city will be built here true. It
may be replied on Dummetts behalf that our inventory of excluded possibilities
leaves out the possibility that a city is never built here; and if that possibility is not
actualized, then the statement in question will be true. That reply, though, wins
the skirmish at the cost of the battle:it is essential to Dummetts case against the
108 Five Attacks on Classical Logic
Principle of Bivalence that there should not be a possible state of affairs that a city
will never be built here (see the discussion of principle below).
All the same, Dummett gives an argument for his premiss . Any situation,
he writes in Truth, in which nothing obtains which is taken as a case of [a state-
ments] being false may be regarded as a case of its being true, just as someone
who behaves so as not to disobey a command may be regarded as having obeyed
it (Dummett 1959, 1011). There is a whiff of stipulation about the words may be
regarded:one would like to know what someone would be losing if he chose not
to regard the absence of any excluded state of affairs as sufficient for truth. But
in any case, the argument trades upon an ambiguity in the phrase so as not to
disobey. X behaves so as not to disobey command C could mean either (a)The
behaviour of X does not, as a matter of fact, include any action that would con-
stitute disobeying C or (b)X comports himself, successfully, with an eye to not
disobeying C. When X behaves so as not to disobey command C is taken in
sense (b), it is plausible to maintain that the formula entails X obeys C. When
taken in sense (a), however, the formula does not entail X obeys C. For in this
sense, the formula may be true of X even though he has never heard of the com-
mand C. Dummetts argument, however, needs the truth of the formula taken in
sense (a)to suffice for Xs obeying C. He needs, that is, the mere non-occurrence
of any disobedient action to suffice for obedience. For only then shall we have
the desired analogy with the way in which the mere non-obtaining of any state
of affairs excluded by a statement suffices for its truth.
Far more importantly for our purposes, though, premiss does not cohere with
what Dummett says about truth later in his article. Our aim is to evaluate the argu-
ment in Truth for deeming classical logic not to be universally applicable. The
nerve of the argument is that, while we know Tertium non datur, we do not know the
Principle of Bivalence. This combination of knowledge and ignorance is impossible
if following a classical deduction from known premisses always issues in knowl-
edge of the conclusion, for we have a simple classical deduction of Bivalence from
Tertium non datur. Now the mainstay of Dummetts argument that we do not know
Bivalence is a necessary condition for truth that Ishall label principle :a state-
14
ment cannot be true unless it is in principle capable of being known to be true.
More briefly, a statement is true only if we can recognize it as true. Assuming that a
statements falsity is equivalent to the truth of its negation, this will entail a cognate
14
Itake principle from the Postscript to Truth (Dummett 1972, 234), where it replaces the
unsatisfactory formula adopted in the main textnamely, that if a statement is true, it must be true
in virtue of the sort of fact we have been taught to regard as justifying us in asserting it (Dummett
1959, 16). The original formula is certainly unsatisfactory. There is no good reason to confine our
modes of justification to those we have been taughtas opposed, for example, to those that we
have discovered for ourselves. It is not clear, however, that principle is strictly weaker than its
The Argument of Dummetts Truth 109
principle for falsehood:a statement is false only if we can recognize it as false. The
argument against our knowing Bivalence may then be spelled out as follows:
Let A be the statement A city will never be built here, uttered at a place where (i)no
city has been built by the time of utterance and there are no plans to build one,
but which (ii) is indistinguishable from the locations of actually thriving cities in
respect of climate, water provision, communications, and all other respects which
best geographical theory suggests are relevant to the viability of a city. Suppose, for
reductio, that we do know the Principle of Bivalence, i.e., we know that every state-
ment is either true or false. Then, in particular, we know that statement A is either
true or false. Now by principle , statement A is true only if we can recognize it as
true. By the same principle, statement A is false only if we can recognize it as false.
Furthermore, we know both of these conditionals. Hence, we know that either
we can recognize statement A as true or we can recognize statement A as false.
Given (ii), however, we may never be in a position to recognize A as true. While
future geographic theory may identify a respect in which the place of utterance is
inimical to the viability of a city, we cannot know that it will. So, for all we know,
it may be that we shall never be able to recognize A as true. Given (i), however, we
also cannot exclude the possibility that we shall never be in a position to recognize
A as false:for all we know, it may be that we shall never be able to recognize A as
false. For all we know, then, it may be that we can neither recognize statement A
as true nor recognize it as false. So we do not after all know that either we can rec-
ognize statement A as true or we can recognize statement A as false. This contra-
diction reduces to absurdity the initial supposition that we know the Principle of
15
Bivalence. We should conclude, then, that we do not know that principle.
predecessor, as Dummett claims (Postscript, 23). Consider the statement (made now) There was
an odd number of geese on the Capitol at the moment of Julius Caesars death. Assuming that
nobody present thought to count them at the time, and that travelling back in time is impossible in
principle, the statement is, now, in principle incapable of being known to be true, and so by princi-
ple cannot be true. All the same, it may be said to be true in virtue of the sort of fact we have been
taught to regard as justifying us in asserting itnamely (let us suppose) that there were precisely
seventeen geese there then. Somebody who thinks that we do know that such a statement is either
true or false, even when we cannot know which it is, may wish to exploit the point that it can be true,
or false, in virtue of a fact which belongs to a sort, many of whose instances are knowable, even if (as
a result of historical contingencies) the particular instance relevant to the statement is not know-
able. See further McDowell 1978, esp. 9.
15
Dummett holds that we are entitled to say that a statement P must be either true or false ... only
when P is an effectively decidable statement (1959, 1617). On my reconstruction of his argument
concerning A city will never be built here, we shall be entitled to say that a statement must be either
true or false only when we know that either we can recognize it to be true or we can recognize it to be
false. The reconstruction suggests, then, a gloss on Dummetts use of the term effectively decidable
whereby it applies to a statement if and only if we can know that either we can recognize it to be true
or we can recognize it to be false. Sanford Shieh has argued for exactly this gloss on Dummetts use of
the term decidable on more general grounds; see his 1998.
110 Five Attacks on Classical Logic
had been to think classically:Ihad understood the quantification all actual states of
affairs classically, and had meant it to be understood classically. The thesis that Iought to
have propounded may be symbolized by the <same> formula
[but with] the logical constants interpreted intuitionistically. So interpreted, the hypoth-
esis y (Oy Exy) may be asserted only if we have a general method of deriving, for
any y, ys not falsifying the statement x from ys being an actually occurring state of affairs.
So understood, () is plausible and consonant with my subsequent remarks about truth.
It is plausible that, if we know a systematic means of rebutting any claim to falsify x, then
x may be safely asserted and is true. (Dummett 2007b, 6945, with incidental changes in
notation)
The Argument of Dummetts Truth 111
Dummett is quite right to say that is more plausible when its connectives are
understood intuitionistically than it is when they are read classically. So con-
strued, however, his argument in Truth affords no reason to abandon classi-
cal logic. We were interested in that argument because it purported to offer a
classically persuasive demonstration of Tertium non datur while showing that
we cannot know the Principle of Bivalence. Had it succeeded, such a demon-
stration would have forced us at least to consider revising classical logic for, in
that logic, any demonstration of Tertium non datur may easily be extended to a
demonstration of Bivalence. Dummetts revised argument, however, puts the
classical logician under no pressure whatever. The classicist will not read the
connectives in intuitionistically, and Dummetts argument now gives him
no reason to do so, for he is well placed to reject the argument as unsound:as
Dummett concedes, from a classical standpoint, premiss is simply false.
16
For persuasive arguments for this claim, see McGinn 1982 and Moore 1999.
The Argument of Dummetts Truth 113
One of the merits of representing this duality using two spaces of possibili-
ties (one epistemic, the other metaphysical) is that we may yet give a uniform
account of the semantic contribution of a compositional device (such as a sen-
tential connective) in terms of an operation that may be applied to both spaces
of possibilities.
Again following Ramsey and Dummett, who identify possibilities with pos-
sible states of affairs, Itake both kinds of possibility to be extra-linguistic enti-
ties. We use clauses to refer to them, as when we speak of the possibility of my
being in London this afternoon, but the possibility is not itself a clause, nor an
equivalence class of clauses. This marks a fundamental difference between the
exclusionary semantics Iam about to present and the incompatibility seman-
tics that Robert Brandom developedalso with a view to elaborating some
pragmatist insightsin his John Locke Lectures. Brandoms incompatibility
semantics aims to codify the so-to-speak horizontal dimension of semantic
content, the one that is articulated by the relations of sentences to each other,
rather than the vertical dimension, which consists in their relations to things
that are not themselves sentences (Brandom 2008, 134). Since possibilities are
things that are not themselves sentences, my semantics is precisely aimed at
characterizing the vertical dimension of a statements contentor better, those
aspects of the vertical dimension that bear most directly on the statements
17
implications for action.
Possibilities, in the Ramseyan sense Iam adopting, need not be fully determi-
nate or complete. All the same, for each of our two spaces of possibilities we have a
clear relative notion of one possibilitys being a determination ofor simply deter-
mininganother. Thus the possibility that a city will be built here next year is a
determination ofa further specification ofthe possibility that one will be built
within the next decade; and the possibility that Jones is both brave and astute is a
determination of the possibility that he is brave. Quite generally, for each space of
possibilities, let us say that a possibility y determines a possibility x when it is logi-
cally necessary that x obtains if y does. (Logically necessary is to be taken in the
sense articulated in the previous chapter.) This relation of determination endows
each of our two spaces of possibilities (the epistemic, the metaphysical) with a
natural and simply describable structure. In fact, the only postulate concern-
ing the relation that we shall need is the evident principle that determination is a
pre-orderi.e., is both reflexive and transitive. When y determines x, Ishall write
x y. Pace the mathematical use of , Ishall not assume that determination is
17
There are other differences between my theory and Brandoms, some consequential upon this
one, others not. Amore detailed comparison would be instructive, but it must wait for another
occasion.
114 Five Attacks on Classical Logic
That is:a possibility belongs to the interior of a set just in case all of its determina-
tions are in the set. Since determination is a pre-order, it is easy to verify that the
operation that maps a set of possibilities to its interior meets the conditions for
being the interior operation of a topology (it is the so-called QO-space topology
generated by ; see e.g. Dummett 2000, 130). That is to say, we have:(1)Int ()=
and Int (U)=U; (2)Int (X) X; (3)Int Int (X)=Int (X); and (4)Int (X Y)=
Int (X) Int (Y). As usual, a set of possibilities is said to be open if it is identical
with its own interior, so that X is open if and only if any determination of any
member of X itself belongs to X:
X is open if and x y (x X x y y X ).
Now, where f A is the set of possibilities excluded by the statement A, our Principle
says that
x y ( x f A x y y f A ).
18
There are surely contexts where the live or open possibilities are too many to form a set.
Techniques from the logic of plurals enable one to extend the present theory to cover these cases,
but Imust leave that extension for a future discussion.
The Argument of Dummetts Truth 115
Thus our Principle requires precisely that the set of possibilities excluded by an
arbitrary statement should be an open set in the QO-space topology that is gener-
ated by the determination relation on the relevant space. This shows what form
an exclusionary semantic theory must take. Such a theory will associate with
each atomic statement that could be made in the relevant language an open set
of epistemic possibilities that the statement excludes, and another open set of
metaphysical possibilities that it excludes. The theory will also contain semantic
principles for the connectives which further associate with each of the languages
complex statements sets of possibilities (of both kinds) that the complex state-
ment excludes. The association will be systematic in the sense of being recursively
determined by the construction of the statement from its parts. The spaces of pos-
sibilities with respect to which a statements content is determined will vary from
context to context, for they will comprise those possibilities that are open or live
in the context. But the argument to show that the possibilities excluded by a given
statement form an open set goes through whatever the underlying spaces might
be, so we know what shape an exclusionary semantic theory will have.
With so much settled, we may return to our question about truth. In order to
pose this more precisely, let us call a possibility a truth-ground of a statement just in
case it is logically necessary that, if the possibility obtains, things will be as the state-
ment says they are. When x is a truth-ground of A, Ishall say that A is true at x. Our
question may then be formulated as follows:when a statement excludes a given set
of possibilities, which possibilities are its truth-grounds?
Assuming that no statement can be both true and false, none of a statements
truth-grounds is excluded by it. So the set of As truth-grounds is disjoint from the
set f A of possibilities that it excludes. However, not every possibility that a state-
ment leaves unexcluded will be a truth-ground of it. The statement This ball is red
all over does not exclude the possibility that the ball weighs 5 oz. All the same,
that last possibility is not a truth-ground of the statement:it is not logically neces-
sary that, if the ball weighs 5 oz, it is red all over.
Which possibilities, then, are truth-grounds of this statement? Arepresenta-
tive example is the possibility that the ball is entirely scarlet:it is logically neces-
sary that if this possibility obtains then things are as the statement says they are.
By virtue of what is this possibility one of the statements truth-grounds? How
does it differ from the possibility that the ball weighs 5 oz, when neither is among
the possibilities that the statement excludes? Although the possibility that the ball
weighs 5 oz is not excluded by our statement, there are further determinations
or specifications of it that are excluded:the possibility that the ball weighs 5 oz
and is green all over, for example. That is why it is logically possible for the ball
to weigh 5 oz without things being as the statement says they are. By contrast, no
116 Five Attacks on Classical Logic
matter how we further determine the possibility that the ball is entirely scarlet,
we shall never reach a possibility that our statement excludes. Quite generally, we
can say that a possibility is a truth-ground of a statement just in case the statement
excludes none of the possibilitys determinations:
Finally, we say that A is true simpliciter if and only if one of its truth-grounds
actually obtains.
Let us define the exterior, X , of a set X to be the set of possibilities, none of
whose determinations belongs to X:
x X if and only if y (x y y X ).
19
This proof turns on the identity of a set and the complement of its complement:that is, it
assumes that the metalogic validates the equivalence of a predicate and its double negation. For the
significance of this, see 4.6.
The Argument of Dummetts Truth 117
Finally, since f ' (f g)', Int (f ') Int ((f g)'), i.e., f (f g); by parallel
reasoning, g (f g), so f g (f g).
(D ) f A B = f A fB .
That is, the possibilities excluded by a disjunction are precisely those excluded by
both disjuncts. The logical law used to justify this principle is common ground
between classical, intuitionist, and even quantum logicians:logicians of all these
schools accept that A B entails C whenever both A and B do. They should all
accept, then, this argument for (D).
What, though, is the corresponding axiom for conjunction? Whenever A
entails C, A B entails C, so any possibility that A excludes is also excluded
by A B:f A f A B. Similarly, we have f B f A B , so f A f B f A B . That is, the pos-
sibilities that A B excludes include those excluded by A and those excluded
by B. This time, though, the converse inclusion does not hold. From the prem-
iss that A B entails C, we cannot infer that either A entails C or B entails
C. Indeed, on classical and intuitionist logical assumptions, the claim that
20
f A B=f A f B is false. Logicians of both these schools accept Ex Contradictione
Quodlibet, so acceptance of a statement in the form A A would commit
one to exclude any possibility whatever. However, when A is a contingent state-
ment, there are many possibilities that are excluded neither by A itself nor by its
negation. This is a consequence of possibilities not being required to be fully
specific or complete. The point holds good even if we accept both the Law of
Excluded Middle and the Principle of Bivalence.
Our account of truth at a possibility points the way to the correct semantic prin-
ciple for conjunction. Aconjunction excludes any possibility that precludes the
truth of both conjuncts. Since a possibility precludes the obtaining of any mem-
ber of a set just when it belongs to the sets exterior, and since the truth-grounds
20
Imade the mistake of accepting this claim in an earlier attempt to develop an exclusionary
semantic theory (Rumfitt 2007).
The Argument of Dummetts Truth 119
of a statement precisely preclude any possibility that the statement excludes, this
yields:
(C ) f A B = ( f A f B ) .
(N ) f A = f A .
That is:the possibilities excluded by A are precisely the truth-grounds of A.
Let us write ver (A) for the set of truth-grounds of A, so that ver (A)=f A. Then
ver (A B)=f A B=(f A f B). Now where f and g are open sets, (fg)=
f g (for proof see e.g. Rumfitt 2007, 667, lemma 4). So (f A f B) =
f A f B. We showed above that f =f whenever f is open. So
(f A f B)=f A f B=ver (A) ver (B). That is, ver (A B)=ver (A) ver (B).
Also ver (A B)=f A B=(f A f B)=(f A f B) since, whenever f and g are
open, (fg) = (f g ) (for proof see Rumfitt 2007, 668, lemma 5). Thus
ver (A B) = (ver (A) ver (B)) = Int Cl (ver (A) ver (B)). Finally,
ver (A)= (ver (A))=Int (ver (A) ).
We may, then, reformulate our exclusionary semantic axioms for , , and
as follows:
Either a boy is at home or a girl is at home is true at the possibility of a childs being
at home, but neither disjunct is true at that possibility:it is not logically necessary
that if the possibility obtains, then things are as A boy is at home says they are, and
similarly for the second disjunct (see further 6.1). In fact, the quasi-intuitionistic
semantic clauses that we have obtained are exactly those that Lloyd Humberstone
proposed for truth at a (perhaps incomplete) possibility in his article From Worlds
to Possibilities (1981). The clause for disjunction, he had to concede (1981, 322), is
on its face unnatural. What we now see, though, is that each of these clauses is the
product of the natural exclusionary semantic principle for the relevant connective
when this is mated with the natural exclusionary account of truth at a possibility.
The lopsided treatment of conjunction and disjunction may seem to jeopardize
the classical logical relationship between these notions, but in fact the threat is illu-
sory. Since ver (A) is fA where fA open, and since f =f whenever f is open, we
always have ver (A)=ver (A), i.e., ver (A)=Int Cl (ver (A)). Topologists call a set reg-
ular open when it is identical with the interior of its closure; so we have just shown
that the set of a statements truth-grounds is always regular open. Now the family
of regular open sets in any non-empty topological space forms a Boolean algebra
when the empty set is the algebras zero, the whole space is its unit, and when meet,
join, and complement are the operations X Y, Int Cl (X Y), and Int ((X)) respec-
21
tively. Furthermore, for a language whose only logical constants are , , and ,
the natural definition of logical consequence in the current framework saysthat
B follows logically from A1,...,An just when, for any set of possibilities U, and
any pre-order on U, and any assignment of subsets of U to the totality of state-
ments such that
(1)each atomic statement is assigned a set of possibilities that is open in the
QO-space topology generated by ; and
(2)the assignment respects (C), (D), and (N),
ver (A1) ... ver (An) ver (B).
It follows that when logical consequence is defined in this way, the valid sequents are
precisely those of the classical propositional calculus of , , and (see Rumfitt 2007,
VII and VIII). So the present semantics validates all the sequents of that calculus.
What about the conditional? As remarked in 4.2, Dummett holds that a
speaker who understands and accepts an indicative conditional is committed to
excluding a possibility just in case it is at once a truth-ground of the antecedent
21
Asimple proof of this theorem is nicely set out in 4 of Halmos 1963. (In fact (Halmos 1963,
25)the Boolean algebra of the regular open sets of any topological space is complete.) Tarski (1938,
449 n.1) claimed credit for this result, on the strength of Theorem B of Tarski 1929.
The Argument of Dummetts Truth 121
and something excluded by the consequent. In other words, he takes the semantic
axiom regulating the conditional tobe
(I ) f A B = f A f B .
Indeed, he cites the simplicity of (I), as a semantic axiom for the conditional, as a
consideration in favour of the exclusionary theory of content (see again Dummett
1972, 22, point (5)). By comparing (I) with (D) and (N), however, we can see that (I)
renders A B equivalent to the material conditional A B , an implausibly
weak account of the truth (or assertability) conditions of indicative conditionals
in natural language.
Suppose, though, that we can make sense of the following three-place rela-
tion among (perhaps incomplete) possibilities:given that x obtains, y is no more
remote a possibility than z is. That is:given that x obtains, it is no less credible
that y should obtain than that z should obtain. Then we can deploy our (defined)
notion of truth at a possibility, and say that A B is true (or assertable) at a
possibility x if and only if B is true at any minimally remote possibility at which
22
A is true, given that x obtains. If we postulate that the relation of ys being no
more remote a possibility than z is reflexive and transitive (with x held fixed),
then this semantics validates Modus Ponens but invalidates hypothetical syllo-
gism, antecedent strengthening, and contraposition. It yields, then, a far more
plausible regimentation of our actual deductive practice with conditionals than
the logic of the material conditional. In fact, when attention is confined to simple
(i.e., non-embedded) conditionals, the resulting logic is the same as that yielded
by the probabilistic account of conditionals, whereby a speaker may assert
A B just when he assigns a high conditional probability to B, given A (for
the proof of equivalence, see Burgess 2009, chap.4). This probabilistic theory of
conditionals is popular, and it may be seen as elaborating some of Ramseys own
23
remarks on the topic. It is encouraging that the exclusionary theory of content
22
For truth-conditions along these lines, see Lewis 1986a, 16. Lewis proposes truth-conditions
for counterfactual conditionals, rather than indicatives; and he supposes that the relata of the
comparative remoteness relation will be fully determinate worlds, rather than possibly indetermi-
nate possibilities. However, if we understand remoteness to be the distance of epistemic possibili-
ties from credibility, rather than the distance of metaphysical possibilities from actuality, we can
apply his ideas to get a theory of truth-conditions for indicative conditionals (see Burgess 2009,
esp.94). This application forces one to treat of possibilities generally, rather than possible worlds in
particular.
23
Besides definite answers If p, q will result, we often get ones If p, q might result or q would
probably result. Here the degree of probability is clearly not a degree of belief in Not-p or q, but a
degree of belief in q given p (Ramsey 1929, 154). See, though, my 2013 for doubts about the currently
prevalent way of elaborating this idea of Ramseys.
122 Five Attacks on Classical Logic
24
Of course, the implicit logic of the metalanguage is a predicate logicbut natural generali-
zations of the proposed axioms for conjunction and disjunction serve as plausible exclusionary
semantic postulates for the universal and existential quantifiers.
The Argument of Dummetts Truth 123
if he allows himself to employ those very rules when attempting to provide that
account.
Intuitionistic logic, by contrast, does not cohere with the proposed definition
of consequence. To see this, observe first that, even if we restrict ourselves to intu-
itionistic logic in the metalanguage, the inference from f g to g f and thus to
Int (g ) Int (f ) still goes through. Even under that restriction, then, f g implies
g f . Similarly, we can show intuitionistically that f f . For suppose that
x f, that x y, and that z(y z z f ). Since f is open, it follows from the first
two suppositions that y f. If we use y to instantiate the variable z in the third
supposition, however, we obtain (y f ). The resulting contradiction means that
x f intuitionistically entails y (x y z (y z z f )), i.e., x f . These
two lemmas together intuitionistically entail that f =f . Given the explana-
tion of negation whereby f A=f A , we shall have, then, that f A =f A =f A
and hence that ver (A)=ver (A). Given the present definition of consequence,
then, we shall have A A even with intuitionistic logic as the metalogic.
Where I signifies intuitionistic deducibility, however, we certainly do not have
A I A. Hence, intuitionistic logic does not cohere with the present definition
of consequence.
It is, moreover, hard to envisage any alternative exclusionary specification of
the meaning of the negation operator, or any alternative exclusionary definition
of consequence, which will possess the slightest intuitive plausibility and which
will combine to generate the characteristically intuitionistic pattern whereby
A A is sound while A A is in general unsound. That is to say, it is hard
to see how the exclusionary account could be parlayed into a definition of con-
sequence with which intuitionistic logic might cohere. In What is a Theory of
Meaning? (II), Dummett contemplated a definition of consequence whereby B
follows from A 1,...,An if and only if f B f A1 ... f An (see Dummett 1976a, 126).
As he later acknowledged, though (Dummett 2007b, 696), this is not a correct
account of consequence from an exclusionary point of view. From that point of
view, the truth-grounds of A are the possibilities in f A , and while f B f A1 ... f An
implies f A1 ... f An f B, the converse does not hold. So there are genuine
instances of consequence that Dummetts definition misses. But in any event
that definition does not help the intuitionist:it validates A A whilst inval-
idating A A (see again Dummett 1976a, 126).
What conclusions should we draw about Dummetts attack on classical logic in
Truth? Our analysis shows it to be a failure. We saw in 4.3 how the direct attack
may be repulsed: the classical logician has every reason to reject Dummetts
premiss . The implied argument for intuitionistic logic fares even worse. The
124 Five Attacks on Classical Logic
Now knowledge of what a statement says is, in general, implicit:in a few instances,
we know what one statement says by virtue of knowing (explicitly) that it says the
same as another, but that cannot be the general case, on pain of regress. Moreover,
our knowledge of what some undecidable statements say must be implicit: an
accurate translation of an undecidable statement will itself be undecidable, so our
understanding of such statements must ultimately rest on implicit knowledge of
what some of them say. Dummett holds, though, that any ascription of implicit
knowledge is vacuous unless the knowledge ascribed is capable of being fully
manifest in the way we behave. Accordingly, he deems to be vacuous ascriptions of
1
Dummett 1981a, 467. To match the terminology of this book, Ihave substituted statement for
Dummetts sentence throughout the quotation.
126 Five Attacks on Classical Logic
for which we have no guarantee of decidability:we do not know that it is feasible, for each
of its statements P, to come to know P or to come to know not-P. Thus this principle holds
(NKD) K(P ) (Feas K(P ) Feas K(P )).
Then given that we also accept
(EC) P Feas K (P )
any truth of the discourse may be feasibly knownwe get into difficulty if we also allow
as valid
(LEM) P P .
For LEM and EC sustain simple reasoning to the conclusion that any P is such that either
it or its negation may feasibly be known. (Wright 2001, 656; the simple reasoning he
refers to is set out in his n.24)
So, given that we are not in a position to assert, of an arbitrary statement, that either
it or its negation may feasibly be known, we are notcontra the classical logician
in a position to assert arbitrary instances of the Law of Excluded Middle.
The Basic Revisionary Argument is Wrights, not Dummetts. The most
Dummetts epistemic constraint on truth yields is that, whenever a statement is
true, it is feasible to know that it is true. In order to advance from this to Wrights
(EC), we would need the further, schematic principle that, if P, then a state-
ment that P is true. Dummett calls this schema and its converse the principles
of semantic shift (Dummett 2004, 12ff.). Most philosophers accept the two shift
principles; but then, most philosophers are not anti-realists, and Dummett is
clear that his anti-realist arguments throw the shift principles into doubt. We
have no general reason, independent of the particular theory of meaning that
we favour, for regarding either of the two [shift] principles ... as holding good.
Whether they hold good in any particular case depends on our theory of meaning
and the conception of truth appropriate to it, as well as on the particular state-
ments to which they are being applied. The widespread obsession with the princi-
ples of semantic shift as known to be inviolable in advance of any further inquiry
The Verificationist Attack on Classical Logic 127
into the concept of truth is utterly misplaced (Dummett 2004, 37). This does not
refute Wrights Basic Revisionary Argument, of course, but it shows that it is no
2
part of Dummetts assault on classical logic.
Dummetts own path to logical revision from his argument about manifesta-
tion is less direct. The first moral he draws from the argument is that knowing
what a statement saysor at least, knowing what an undecidable statement says
does not consist in knowing its realist truth-conditions. In what, then, does it con-
sist? Dummett entertains two answers (cfr. Dummett 1978, Preface, xxiixxiii).
According to the first, the argument shows that we need to adopt non-realist
notions of truth and falsity:the truth- and falsity-conditions for any statement
should ... be taken as ones which we are capable of recognizing effectively when-
ever they obtain (Dummett 1981a, 467). On this conception of the matter, a state-
ments truth is no more than a product of its possible verification, and its falsehood
no more than a product of its possible falsification. If truth is understood in this
way, the objection to the truth-conditional account of understanding a statement
lapses, and we can again say that knowing what a statement says is a matter of
knowing under what conditions it is true. According to the second answer, know-
ing what a statement says is a matter of knowing in what conditions it would be
verified, and perhaps also in what conditions it would be falsified:on this view, we
replace ... the notions of truth and falsity, as the central notions for the theory of
meaning, by those of verification and falsification (Dummett 1981a, 468; emphasis
added). Since, under the first answer, a statements truth is a product of its verifica-
tion, both answers lead to a semantic theory whose kernel is a systematic specifica-
tion of the conditions in which a statement is verified, and perhaps also those in
which it is falsified, but there remains good reason to prefer the second answer. The
realist principle that a statement may be true even though no one is able to recog-
nize its truth is so deeply embedded in our ordinary conception of truth that any
3
account that flouts it is liable to engender confusion.
The argument as Ihave so far expounded it attacks the thesis that knowing what
a statement says involves knowing its (realist) truth-conditions. It offers some sup-
port for a contrary thesis which Ishall follow John McDowell in labelling strong
4
verificationism . According to this thesis, one knows what a statement says when,
2
Ihave not the space to discuss Wrights argument in this book. For a critical analysis of it, see
Incurvati and Murzi 2008.
3
Moreover, Dummetts attempts to give an anti-realist account of truth are hesitant and far from
clear. For a critical analysis of some of them, see Raatikainen 2004, 3.
4
The strong verificationist ... insists, not that mastery of a language should be represented as
not independent of sensitivity to evidence, but that it should be represented as consisting solely in
sensitivity to evidence (McDowell 1976, 63).
128 Five Attacks on Classical Logic
and only when, one knows in what conditions it would be verified or falsified. Given
strong verificationism, the attack on classical logic proceeds as follows. First, one
replaces the discredited truth-conditional semantic theory with a verification-
ist theory of meaning. Dummett assumes that such a theory will generalize the
semantics that Arend Heyting laid down for the language of intuitionistic math-
5
ematics. As Dummett explains that theory,
the meaning of each [logical] constant is to be given by specifying, for any sentence in
which that constant is the main operator, what is to count as a proof of that sentence, it
being assumed that we already know what is to count as a proof of any of the constituents.
The explanation of each constant must be faithful to the principle that, for any construc-
tion that is presented to us, we shall always be able to recognize effectively whether or not
it is a proof of any given statement. For simplicity of exposition, we shall assume that we
are dealing with arithmetical statements ...
The logical constants fall into two groups. First are , and . Aproof of AB is
anything that is a proof of A and of B. Aproof of AB is anything that is a proof either
of A or of B. Aproof of xA(x) is anything that is a proof, for some n, of the statement
A( n ).
The second group is composed of , , and . Aproof of xA(x) is a construction of
which we can recognize that, when applied to any number n, it yields a proof of A( n ).Such
a proof is therefore an operation that carries natural numbers into proofs. Aproofof AB is
a construction of which we can recognize that, applied to any proof of A, it yields a proof
of B. Such a proof is therefore an operation carrying proofs into proofs ... Aproof of A
is usually characterized as a construction of which we can recognize that, applied to any
proof of A, it will yield a proof of a contradiction. (Dummett 2000, 8)
Heytings semantic theory does enforce a departure from classical logic, even if
the metalogic is classical. The theory counts a statement as logically valid if the
semantic principles guarantee it to be provable no matter which atomic statements
are provable. So a statement of the form A A will be valid only if either A is
provable or A is provable (i.e., it is provable that A is unprovable). Since it can-
not be assumed of an arbitrary statement that either it or its negation is provable,
it cannot be assumed that an arbitrary instance of A A is valid:accordingly,
Excluded Middle is not unrestrictedly valid. The reasoning just given could be
formalized as a valid argument in either a classical or an intuitionistic metalogic.
So even if one starts out accepting all the principles of classical logic, adopting
5
See Heyting 1934. Many writers on intuitionism (although not Dummett) refer to the
Brouwer-Heyting-Kolmogorov (BHK) semantics, but the triple-barrelled tag is misleading.
Brouwer did not have a semantic theory at allmerely the idea that mathematical truth amounts
to provabilityand Heyting and Kolmogorov were engaged in very different semantic projects.
Heyting wanted to specify the meanings of the connectives using notions of construction and proof
that intuitionists could accept, whereas Kolmogorov (1932) wanted to identify the intuitionistically
acceptable sequents in a way that a classical mathematician could understand. (For a recent contri-
bution to Kolmogorovs project, see Artemov 2001.) Dummetts work lies squarely in the tradition
initiated by Heyting.
The Verificationist Attack on Classical Logic 129
the Heyting semantics would force one to cease doing so. In this way, Dummetts
argument for the superiority of a semantic theory along the lines of Heytings
extends to yield an argument against classical logic. Following McDowell, Ishall
call this the strong verificationist attack on classical logic.
Although this attack has already been subject to a great deal of analysis and
discussion, it still deserves scrutiny. Many responses to it focus on avoidable mis-
takes in Dummetts elaboration of the strong verificationist case; once the mis-
takes are corrected, the responses miss the target. In 5.2 and 5.3, Ishall correct
what Itake to be the most important mistake in Dummetts presentation of the
argument and then show how classical logic remains threatened even after that
mistake has been rectified. In 5.4, though, Iidentify the basic weakness of the
strong verificationist argument. The analysis of 5.2 and 5.3 will not have been
wasted, however. First, the analysis brings out the differences between the strong
verificationist attack on classical logic and a more subtle and promising critique
with which it is often confused. Second, some of the formal techniques used in
elaborating the strong verificationist attack will be redeployed in developing the
subtler critique, which will be the topic of the following two chapters.
The if part of (DH) is correct:if some evidence warrants the assertion of A, then
it also warrants the weaker assertion of A B , for arbitrary B. For familiar
Gricean reasons, the weaker assertion might well mislead a hearer by suggesting
that the speaker was not in a position to assert A tout court. All the same, evidence
that entitles a speaker to assert A is ipso facto evidence that entitles him to assert
A B . The only if part of the claim, on the other hand, seems plain wrong.
Inspector Morse might have evidence that entitles him to assert Either Black or
Brown is the murderer even though his evidence does not entitle him to assert
either Black is the murderer or Brown is the murderer.
In attempting to deal with apparent counterexamples to Heytings original
axiom for disjunction, Dummett invoked a distinction between canonical and
non-canonical proofs (see Dummett 2000, 7.4). While there may be proofs of
A B that are neither proofs of A nor proofs of B, any canonical proof of a dis-
junction will be a proof of one of its disjuncts. On this view, Heytings axiom can be
defended as a contribution to the project of saying which constructions constitute
canonical proofs of their conclusions; the relation between canonical proofs and
proofs more generally is left for subsequent treatment. With this strategy in mind,
6
One may be epistemically entitled to make an assertion without being entitled tout court:Imay
know that A is true but be precluded from asserting it because doing so would break a confidence,
for example.
Many philosophers hold that warrants come in degrees. For a way of accommodating this view,
see the end of this section.
The Verificationist Attack on Classical Logic 131
7
Sketch of proof (Kreisel): For each n, Pr (n, 0 = 1 ) is true, and (since Pr is primitive
recursive) provable in HA. Hence, for any n, HA (n) y Pr (y, 0 = 1 ). By logic,
x (x) [y Pr (y, 0=1 ) y Pr (y, 0=1 )]. Now suppose for a contradiction that
T (n). Then, by the definition of T, HA, x (x) (n), whence HA x (x) (n). By the
results above, this yields HA [y Pr (y, 0=1 ) y Pr (y, 0=1 )] y Pr (y, 0=1 ),
hence HA y Pr (y, 0=1 ) y Pr (y, 0=1 ), whence HA y Pr (y, 0=1 ). But then
HA proves its own consistency, contrary to Gdels Second Incompleteness Theorem. This con-
tradiction shows that for no n do we have T (n), as required.
The Verificationist Attack on Classical Logic 133
of pieces of evidence that a thinker might come bya set of possible pieces of evi-
dence, for shortthat can serve as a domain of quantification in a given context.
Given that assumption, we will have a set |A| that comprises all and only those
possible pieces of evidence in the domain that would warrant the assertion of a
statement, A.
With this framework in place, there is a natural candidate to serve as an evi-
dentialist semantic axiom for conjunction. It is initially plausible to hold that any
piece of evidence that one might come by warrants a conjunction when and only
when it warrants both conjuncts. Thus, where x ranges over the possible pieces of
evidence in any domain, wehave:
8
(C) x warrants A B if and only if x warrants A and x warrants B.
Equivalently, the possible warrants for a conjunction are simply the intersection
of the warrants of each conjunct:
(C ) |A B| = |A| |B|.
What, though, about disjunction? As we saw, (DH), the analogue of (C), will
not do. Inspector Morse might have evidencecall it xthat entitles him to
assert Either Black or Brown is the murderer (= A B ) even though it does not
entitle him either to assert Black is the murderer (A) or to assert Brown is the
murderer (B). We have yet to find necessary conditions for being a warrant for a
disjunctive statement.
How can we identify these conditions? Since Morses evidence, x, does not
warrant A and does not warrant B, it does not belong to the union |A| |B|.
It does, however, relate to that union in a more subtle way. To see how, let us
return to the exclusionary analysis of Morses predicament that was deployed
in 4.2. Morse, it may be recalled, began his inquiry knowing that the mur-
derer was one of (1)Black, (2)Brown, (3)Green, (4)Scarlet, or (5)White. Ex
hypothesi, x warrants the disjunction A B . As such, it rules out three of
the original five possible culprits:(3), (4), and (5). Now any warrant for A, i.e.,
any member of |A|, rules out four of those possible culprits:(2), (3), (4), and
8
Iam not at all sure that the initial plausibility of (C) survives reflection on the Paradox of
the Preface. Let x be the total evidence of the author of a particular book, b. It seems that we can
consistently suppose that x has three properties:(1)for each individual statement A i in b, x war-
rants Ai; (2)x warrants the claim that any book of the same length as b will contain at least one
false statement so that (3)x does not warrant the conjunction of all the Ais. The conjunction of
(1)and (3), though, is incompatible with (C). The matter is delicate:in particular, the inference
from (2)to (3)might be questioned. However, in view of the other difficulties facing the strong
verificationist Ishall not explore this potential problem for his position.
The Verificationist Attack on Classical Logic 135
(5). Similarly, any warrant for B, i.e., any member of |B|, rules out a distinct
but overlapping foursome:(1), (3), (4), and (5). In ruling out (3), (4), and (5),
then, x rules out all the possible culprits that are ruled out by any warrant for
A and by any warrant for B. That is, x rules out all the possible culprits that are
ruled out by all the members of |A| |B|. This argument does not depend on
the precise nature of x:ruling out all the possible culprits ruled out by all the
members of |A| |B| will be characteristic of any evidence that warrants the
disjunction A B .
How might we generalize this observation? We have the notion of an epis-
temic possibility:something that might be the case, for all we know. The notion
9
has been the subject of much attention in recent philosophy and it certainly
needs analysis; for one thing, there is much to be said about how, in a given con-
text of use, it is determined who we are. But, however the notion is made more
precise, evidence has its status as such by virtue of ruling out some hitherto
open epistemic possibilities. Given a background space of epistemic possibili-
ties, moreover, and a background set of possible pieces of evidence, we can take
it to be a determinate matter which epistemic possibilities a given piece of evi-
dence rules out. We have, then, the following condition for evidence to warrant
a disjunction:
x warrants A B if and only if x rules out any possibility that is ruled
(D)
out by every warrant for A and by every warrant for B.
In tracing out the implications of (D), it helps to reformulate it. Where U is any
set of possible pieces of evidence we may define the closure of U, Cl (U), by the
condition
x Cl (U) if and only if x rules out any epistemic possibility that is ruled out by
all the members of U.
INCREASING U Cl (U)
IDEMPOTENT Cl Cl (U)=Cl (U)
9
See, notably, the essays in Egan and Weatherson 2011.
136 Five Attacks on Classical Logic
and
10
MONOTONEIf U V then Cl (U) Cl (V).
Let us call a set closed when it is identical with its own closure. By idempotence,
the closure of any set is closed, and by the increasing property and monotonicity
11
the closure of U is the smallest closed set containing U.
The foundation of evidentialism is the followingaxiom:
(R) The possible warrants for any statement form a closed set.
Let A be an arbitrarily chosen statement. In arguing for (R), there are two cases to
consider:(1)that in which there is an epistemic possibility of As being untrue and
(2)a degenerate case in which there is no such possibility. For case (1), the argument
runs as follows. Let U be the set of possible warrants for A, and consider an arbitrary
member, x, of the closure of U. We require to show that x belongs to U. Suppose, for
a contradiction, that it does not. Then x is not a warrant for A, i.e., x does not rule
out the epistemic possibility that A is untrue. Let us call this epistemic possibility p,
so that x does not rule out p. By definition, every member of U is a warrant for A, so
every member of U rules out p. But x belongs to the closure of U, so x rules out any
epistemic possibility that every member of U rules out. In particular, then, x rules
out the possibility p. But that is a contradiction:x both rules out and does not rule
out p. So x must belong to U, showing that U is a closed set, as required. As for case
(2), if there is no epistemic possibility that A is untrue, then any speaker will be epis-
temically entitled to assert A, whatever his evidence. We may, then, count the entire
domain of pieces of evidence as warrants for A; and the whole domain is trivially
closed.
Axiom (R) reveals the shape of an evidentialist semantic theory. The theory
associates with each statement in the relevant language a closed set of possible
pieces of evidence; these will be the possible warrants for the statement. The
theorys compositional principles say how the possible warrants for a complex
10
Proofs. INCREASING:immediate from the definition of closure. IDEMPOTENT: since
closure is INCREASING, it suffices to show that Cl Cl (U) Cl (U). Suppose then x Cl Cl (U).
Then x rules out any epistemic possibility that is ruled out by every member of Cl (U). Consider
an arbitrary possibility p that is ruled out by every member of U. By definition, every member
of Cl (U) will rule out p. Hence x rules out p. But since p was chosen arbitrarily, that shows that
x rules out any possibility that is ruled out by every member of U, so that x Cl (U), as required.
MONOTONE:suppose that x Cl (U) and that U V. Since x Cl (U), x rules out any possibil-
ity that is ruled out by every member of U. Since U V, it follows that x also rules out any pos-
sibility that is ruled out by every member of V. That is, x Cl (V), as required.
11
That is to say:Cl (U) is a closed set containing U and, whenever W is such a set, W contains
Cl (U).
The Verificationist Attack on Classical Logic 137
statement relate to the possible warrants for its components. The intersection
12
of any two closed sets is closed, so our semantic postulate for conjunction, (C),
respects the general principle, (R), that the warrants for any statement should
13
form a closed set. So does (D), for the closure of any set is closed.
How should an evidentialist semantic theory treat negation? Evidence consti-
tutes a warrant for A when its apprehension epistemically entitles a speaker to
assert Aor to answer yes to the question whether A. In a similar spirit, let us
say that evidence constitutes an anti-warrant for A when its apprehension epis-
temically entitles a speaker to deny Aor to answer no to the question whether
A. With the notion of anti-warrant in play, we may characterize the semantic con-
tribution of a sentential negation operator bysaying
(N) x warrants A if and only if x is an anti-warrant for A.
As it stands, (N) is unspecific. We have not laid down any principles identifying
anti-warrants for conjunctive or disjunctive statements, so (N) does not tell us
(for example) what warrants the negation of a conjunction. Moreover, we have
as yet no guarantee that (N) respects (R). To solve these problems, we need to
supplement (N) with principles saying how warrants for some statements relate
to anti-warrants for others. These principles may well vary with the topics of the
statements in question. An anti-warrant for a mathematical statement will be a
refutation of it, whilst an anti-warrant for a statement in physics will be empirical
evidence that disconfirms it. In advance of inquiry, there is no reason to suppose
that proofs relate to refutations in precisely the way that confirming empirical
evidence relates to disconfirming evidence. Whilst (N), then, is the strongest gen-
eral principle one can find that specifies warrants for negation, it leaves the logic
of negation undetermined.
Evidence often provides some degree of support for a statement even though
it falls short of putting someone in a position to know it. So an evidentialist
semantic theory ought to include principles that determine the place of com-
plex statements in a network of partial evidential support. In fact, our semantic
postulates yield principles of this latter kind quite directly, once the postulates
12
Proof. Suppose that U=Cl (U) and that V=Cl (V). We need to show that U V=Cl (U V).
Since INCREASING already yields U V Cl (U V), it suffices to show that Cl (U V)
U V. Now U V U, whence by MONOTONE Cl (U V) Cl (U)=U. Similarly, Cl (U V)
Cl (V)=V. Together, these inclusions yield Cl (U V) U V, as required.
13
Although Icannot argue for the claim here, Ibelieve that natural generalizations of (C) and (D)
serve as semantic principles that specify warrants for universally and existentially quantified state-
ments. For the generalizations that Ienvisage, see Mares 2010, 10. Mares works with a notion of
objective information rather than warrants, but modulo differences consequential upon that, his
treatment of disjunction is equivalent to that proposed here.
138 Five Attacks on Classical Logic
In these axioms, t is a known logical truth of the relevant logic, and f is a known
logical falsehood. Where C cannot obtain, we take it to warrant any statement;
thus the second clause of axiom Isays that any evidence that can obtain is an
anti-warrant of a known logical falsehood. Given these axioms, a theorem of van
Fraassens shows that our semantic postulates entail further plausible principles
that specify the relationship between the degrees to which evidence supports
15
atomic statements and the degrees to which it supports complex statements.
There is a strong case, then, for saying that our semantic postulates specify the
contribution that and, or, and not make to the place that statements contain-
ing them occupy in a network of partial evidential support.
14
On this way of understanding the notion, P(A/B) can equal 1 even when apprehension of B does
not render A subjectively certain. See Williamson 2000, 213ff., for elaboration of this point.
15
See propositions (27) and (31) of van Fraassen 1981b (503, 505). Van Fraassen has a rather
different way of understanding P(A/B) (see his 1981a), but the difference in interpretation does not
affect his formal proofs.
The Verificationist Attack on Classical Logic 139
consequence, not in terms of the preservation of truth, but in terms of the preser-
vation of warranted assertability. Specifically, it might be proposed that a premiss
A logically entails a conclusion B whenever |A| |B| in all suitably structured
domains of possible pieces of evidence; a suitably structured domain will con-
form to our compositional principles (C), (D), and (N). More generally, we might
say that some premisses A1,...,An logically entail a conclusion B if and only if
|A1| ... |An| |B| in all suitably structured domains of possible pieces of evi-
dence. Given (C), some premisses (plural) entail a conclusion if and only if their
conjunction (singular) does.
The semantic principle (N), we noted, leaves the logic of negation undeter-
mined. Similarly, (D) leaves the logic of disjunction unsettled. (D) validates
the restricted form of the rule for eliminating or, in which side premisses are
not permitted. That is to say, it validates the rule of proof whereby, if A logically
16
entails C and B entails C, then A B entails C. As it stands, though, it does not
validate the unrestricted elimination rule or the Law of Distributionthe princi-
ple that A (B C) logically entails (A B) (A C) or, more generally, that
A (B1 ... Bn) entails (A B1) ... (A Bn) . We could, of course, vali-
date these stronger rules (and such classical laws of negation as Double Negation
Elimination) by laying down further postulates that constrain the structure of the
domain of possible pieces of evidence. However, strong verificationism threatens
classical logic because the needed postulates are not only unjustified but appear
to conflict with truths about the relevant sorts of evidence. There are, in other
words, cases where Double Negation Elimination and Distribution appear to be
invalid. Itake these in turn.
16
Given that A entails C, |A| |C|. Given that B entails C, |B| |C|. So |A| |B| |C|. By MONOTONE,
Cl (|A| |B|) Cl (|C|). By (D), Cl (|A| |B|) = |A B| and, by (R), Cl (|C|) = |C|. So |A B| |C|, as required.
140 Five Attacks on Classical Logic
m1 or ... or mn. That is to say, quantum theory allows that there is a measurement, x,
which warrants the statement A (B1 ... Bn) . Measurement, x, however, does
not warrant the statement (A B1) ... (A Bn) . Indeed, no measurement war-
while x warrants A (B1 ... Bn) , it does not warrant (A B1) ... (A Bn) .
17
To justify the last step in this argument, recall that x Cl (U) if and only if x rules out any
epistemic possibility that is ruled out by all the members of U. When U is empty, every possibility is
ruled out by all the members of U (vacuously), so x Cl () if and only if x rules out every epistemic
possibility tout court. Since no piece of evidence rules out every epistemic possibility, Cl ()=.
142 Five Attacks on Classical Logic
particles and hence are not subject to the Uncertainty Principle: statements
about them may be assumed to conform to the Law of Distribution and to the
rule for introducing negation. There need be no equivocation, or clashing of
logical gears, in developing a semantic theory for quantum logic in a classical
metalogic.
came down to us; in the turmoil of the assassination, it is unlikely in the extreme
that anyone did even the first of these things. In other cases, admittedly, it is harder
to explain how a statement may be true even though no one can recognize its truth.
Even in advance of finding such an explanation, though, most of us cleave to a realist
conception of truth, and will depart from it only given strong reason.
How strong is Dummetts argument against the conception? It rests crucially
on the idea that an attribution of implicit knowledgeand, in particular, an
attribution to a speaker of implicit knowledge of a statements meaningis vacu-
ous unless the attribution can be fully cashed out in a way that is manifest in the
speakers behaviour. Iagree with Williamson that the demand that such attribu-
tions should be capable of being cashed out in this way is not properly grounded. It
does, indeed, appear to be motivated more by preconceived philosophical reduc-
tionism than by the actual needs of empirical linguistics (Williamson 2007, 282).
What is more, the requirement that implicit knowledge of meaning should be
fully manifest in behaviour is so stringent that Dummetts own account of that
knowledge does not meet it. On that account, knowing what a statement says
is a matter of knowing the conditions for it to be verified. More precisely, it is a
matter of knowing, for any body of evidence that may be presented, whether or
not that evidence warrants the statements assertion. Now this knowledgelike
knowledge of truth-conditionsmust in general be implicit. While we may, in
some instances, know in what conditions A is warranted by virtue of knowing
(explicitly) that it is warranted by precisely the same evidence as B, this cannot
be the general case, again on pain of regress. Accordingly, on Dummetts princi-
ples, any attribution of such knowledge must be capable of being fully manifest
in the speakers behaviour. Now the behaviour in which a speaker manifests his
knowledge of the conditions in which there is warrant to assert A will be his judg-
ing correctly, when presented with various data or pieces of evidence, whether
that evidence warrants the assertion of A. By Dummetts requirement of full
manifestation, then, a thinker who understands A must be able to decide, for any
body of evidence e, whether e really warrants an assertive use of the statement in
18
question.
18
It is well known that Dummett appeals to this implication of the manifestation requirement
to rule out semantic theories according to which knowing a statements sense is a matter of know-
ing under what conditions the statement is true. In understanding Dummetts thought, however,
it is also important to appreciate that he uses the very same principle to rule out some popular
semantic theories that yield intuitionistic logic. In 7.4 of Elements of Intuitionism, he addresses
the question, whether the intended meanings of the logical constants are faithfully represented on
Beth trees. His presentation of Beths account of the meanings of the connectives and quantifiers is
sensitive and, initially, highly sympathetic. However, his answer to the question is in the end nega-
tive. The main reason he gives for this verdict is that it is not in general a decidable matter whether
a formula is verified at a given node on a Beth tree (see Dummett 2000, 282). Thus Dummett wields
The Verificationist Attack on Classical Logic 145
his requirement of decidability to exclude an account of the connectives senses that is otherwise
attractive from an intuitionists perspective.
I think that this was unfortunate, and led to Dummetts mounting a case for intuitionistic logic
that was far less persuasive than it could have been. Indeed, the account of the senses of the connec-
tives that Ishall recommend in the next chapter is a generalization of Beths theory. See the discus-
sion in 6.4.
19
For related criticism of Dummett, see Raatikainen 2004, 139 and 1434.
146 Five Attacks on Classical Logic
Steppe is, Iknow what is meant by the claim that Proto-Indo-European originated
there. According to Dummett, though, because Iknow what the statement means,
Imust be in a position to decide, of any putative evidence e, whether e suffices to
warrant an assertion of the Kurgan Hypothesis. This claim is wholly implausi-
ble. Let us suppose for definiteness that e is the totality of the evidence presented
in Mallory and Adamss recent survey of the field, The Oxford Introduction to
Proto-Indo-European and the Proto-Indo-European World (Mallory and Adams
2006). The evidence recorded in that book is a complex tangle of archaeological,
genetic, and linguistic data. Am Iin a position to decide whether this evidence
warrants an assertion of the Kurgan Hypothesis? Surely not. Iknow that some
authorities in the field take the evidence in question to be sufficient to assert the
hypothesis. Equally, though, Iknow that other leading scholars take the same
evidence to warrant contrary claimsnotably, that Proto-Indo-European origi-
nated in ancient Anatolia (see e.g. Renfrew 1987, 2003). Ialso know that yet other
authorities deem the evidence in question insufficient either to assert the Kurgan
Hypothesis or to deny it: they withhold judgement on the matter. That there
should be disagreement over whether this body of evidence warrants an asser-
tion of the Kurgan Hypothesis is entirely unsurprising. Different authorities will
weigh differently the diverse sorts of data that make up the evidence e. There is no
procedure that could be applied to decide whether e warrants an assertion of the
hypothesis, and, in the absence of such, an amateur like me must admit ignorance
as to whether e warrants the assertion. On Dummetts theory, that entails that
Ido not know what the hypothesis means. But Ido know what it means, so the
theory must be wrong.
Dummett has no good general argument, then, for his conclusion that know-
ing what a statement says is a matter of knowing in what conditions it would be
verified, rather than knowing under what conditions it is true. The principle on
which he relies to exclude the truth-conditional account also excludes the strong
20
verificationist theory that he prefers.
As for the particular cases cited as possible counterexamples to classical logi-
cal laws, they reinforce the importance of cleaving to the conception of conse-
quence as the necessary preservation of truth. This is particularly clear in the
case of Goldbachs Conjecture. On the natural way of identifying warrants and
anti-warrants for quantified statements, there will be mathematical constructions
20
Williamson has also argued that the principle Dummett invokes against the truth-conditional
account of knowledge of meaning equally excludes his own account (Williamson 2000, 4.8).
As he reconstructs Dummetts argument, though, the operative principle is that a statements
truth-conditions (and assertability-conditions) must be luminous rather than decidable.
Williamson argues that only trivial conditions are luminous, but against this see Berker 2008.
The Verificationist Attack on Classical Logic 147
that prove A without proving A. What that case brings out, though, is
the strangeness of the strong verificationist project of laying down what might
warrant a statement without reference to the conditions under which it is true.
Aproof of xFx would show that there is no counterexample to Goldbachs
Conjecture and, in the particular case of that Conjecture, the absence of a coun-
terexample suffices for truth. So a proof of xFx would seem to establish
the Conjectures truth. Given that, it seems that a proof of xFx ought to
count as a proof of xFx , whatever prejudices we start from about what such a
proof should look like. Of course, in bringing truth-conditions back into play, we
need to do so cautiously:we cannot simply assume, without argument, that every
arithmetical statement is bivalent. However, the failure of the strong verification-
ists general arguments, and reflection on the present example, show that we can-
not ignore truth-conditions altogether in assessing consequence.
As for the counterexample to Distribution, it assumes similarly that a realist,
or verification-transcendent, notion of truth is inapplicable to statements about
the quantum realm. In the heyday of positivism, some physicistsnotably Bohr
and Heisenbergdid interpret the quantum theory in this way, and if it makes
no sense to apply a verification-transcendent notion of truth to statements about
subatomic particles, it can make no sense to gloss consequence among such
statements in terms of the preservation of truth. Furthermore, if the only alethic,
or quasi-alethic, notion that is applicable within the quantum realm is that of
possessing a warrant, then preservation of warrant must be the mark of conse-
quence. Few physicists today, though, accept this radically anti-realist interpre-
tation of quantum mechanics, so the present challenge to Distribution rests on
an interpretation of a scientific theory that has few adherents. We may therefore
set it aside.
21
See 1.1. As Iremarked there, in order to validate classical propositional logic, conjunction
must distribute over the or that is used in stating the Principle of Bivalence:only if it does so will
we be able to move from A is either true or false and B is either true or false to Either A and B are
both true, or A is true and B is false, or A is false and B is true, or A and B are both false, which is what
we need to guarantee that the cells of a truth-table exhaust the possibilities. This is why Dummett
formulates Bivalence as the principle that every statement is determinately either true or false; he
uses determinately precisely to signal that the occurrence of or is subject to the unrestricted rule
The Verificationist Attack on Classical Logic 149
This suggests a more insidious threat to classical logic than the frontal, but easily
repulsed, strong verificationist attack. Our understanding of the connectives, it
may plausibly be claimed, yields knowledge of the semantic relationships recorded
in the familiar truth-tables. That knowledge, though, while tantamount to know-
ing that certain basic sequents are valid, does not yield knowledge that all the rules
of classical propositional logic are sound. We need an additional assumption
that of Bivalencein order to reach that conclusion. And then the worry is simply
that the additional assumption is insufficiently supported:we cannot be said to
know that every statement is either true or false. There are, after all, many state-
mentsthose involving vague predicates, mathematical statements such as the
Continuum Hypothesisof which people feel unconfident in asserting that they
must be true or false. Proper epistemic caution, then, should make us hesitate to
assert the Principle of Bivalence. On this account of how our understanding of
the connectives connects with our knowledge of their logical properties, though,
that is enough to create a problem for the classicist. For a theorist who accepts that
account can saythat
if the principle of bivalence is true (as it may be), then the classical theorems are true;
indeed, if the principle of bivalence is true (as it may be), then the senses of the connec-
tives guarantee the truth of the classical theorems. The trouble is that since he does not
know that the antecedent is true, he cannot detach and assert the consequents of those
conditionals. In his view, the classical logical truths are picked out, in the classical model
theory, by a property amounting to no more than this:being such that, for all we know,
the senses of the connectives guarantee their truth. Since that property does not ensure
even the truth of sentences that have it, there is no particular point in a procedure which
generates all and only the sentences that have it:which is what a sound and complete clas-
sical proof-theory would be. (McDowell 1976, 5960)
of or-elimination, (see Dummett 1981b, 436). Together with uncontroversial principles concerning
conjunction, that yields Distribution. The or of an intuitionist metalogic is also determinate in
this sense.
150 Five Attacks on Classical Logic
He then asserts without proof that the rules of the intuitionistic propositional
calculus will be sound on this conception of logical truth.
McDowells claim that it is precisely the intuitionistic rules that are
sound with respect to his conception of logical truth must be moot until
The Verificationist Attack on Classical Logic 151
that conception has been f leshed out with an account of how claims to
know the truth of complex sentences could be justified by knowledge of the
22
truth of simpler sentences. In trying to do this, however, we encounter the
same problem that bedevilled the Heyting postulate for disjunction. For
McDowells intuitionist, as for the classicist, the alethic status of a complex
statement is determined by the alethic status of its parts, insofar as it is deter-
mined at all. However, the epistemic status of a complex statement is not
always determined by the epistemic status of its parts. My glance out of the
window enables me to know the truth of Either Castor or Pollux has just run
past, but Ido not know the truth of either of the disjuncts. Asecond prob-
lem is that it is unclear what interest attaches to McDowells notion of logi-
cal truth:it is far from obvious why an entire scholarly discipline should be
devoted to identifying the truths that are knowable from our understanding
of the connectives alone.
McDowells positive account of logical truth, then, faces problems; further dif-
ficulties would attend any attempt to extend it to an account of consequence. All
the same, he has identified an interesting form of attack on classical logic. The
attack consists in arguing, first, that the soundness of at least some of the clas-
sical rules is not inherent in the meanings of the connectives. This first claim is
highly plausible. The next move is to argue that an additional premiss is needed
to get from our knowledge of what the connectives mean to knowledge of the
logical relations among statements involving them. This claim, too, is plausible.
So we need to inquire (a)whether there are additional premisses weaker than the
Principle of Bivalence that ground classical logic and, if so, (b)whether there is
good reason to believe those premisses.
In fact, pursuing this inquiry will generate a prima facie case for intuitionis-
tic logic, rather than merely casting doubt upon the classical laws. In the follow-
ing two chapters, Ishall construct a framework for semantic theorizing that does
not beg the question against the intuitionist by assuming Bivalence. Within that
framework, Ipropose semantic axioms for the familiar connectives which are
faithful to their ordinary meanings. When consequence is characterized as the
necessary preservation of truth, it is the rules of the intuitionistic propositional
calculus that turn out to be sound and complete with respect to these axioms.
We need additional postulatespostulates that do not appear to be guaranteed
22
Note that the sequents directly recoverable from the truth-tables do not by themselves consti-
tute a complete axiomatization of intuitionistic propositional logic. Both classicists and intuition-
ists accept the following elimination rule for :from A B infer A; and from A B infer B. The
sequents (1)to (4)in the text do not ensure the soundness of this rule.
152 Five Attacks on Classical Logic
by the senses of the logical wordsif the framework is to yield a semantic theory
with respect to which the rules of the classical propositional calculus are sound.
Inasmuch as these postulates are doubtful, we have not only a challenge to clas-
sical logic, but an argument for adopting intuitionistic logic. We can in this way
press the challenge from dubious grounds, and then consider how a classical logi-
cian might reply to it.
6
Possibilities
The framework in which I shall investigate the challenge that classical logic
rests on dubious grounds makes central use of the notion of a logical possibility.
Possibilities are things of the same general character as possible worlds, on one
popular conception of the latter. They differ from worlds, though, in that they are
not required to be fully determinate or complete. In this chapter, Iintroduce pos-
sibilities and develop a semantic theory in which truth is relativized to them. The
main line of argument resumes in Chapter7.
1
Iwrite neutrally of states or properties because the theory of possibilities developed below can
fit various accounts of their metaphysical nature. See n.7 in this chapter.
154 Five Attacks on Classical Logic
happen to inhabit. Not so for the moderate realist. On his view, the statement I
could have been in London on 6 March 2012 is true because there is a possibility
in which Iam then in London. But the existence of that possibility amounts just to
this:there is a way things could have been in which Iam then in London.
In expounding and motivating their theory, moderate modal realists make free
use of the notion of a possibility in the sense specified. However, when they come
to develop a semantic theory for a language containing modal operators, they do
not employ a notion of truth that is relativized to possibilities in general. Rather,
they speak of a statements being true at a possible world, where a possible world
(if there are such things at all) is a very particular kind of possibility. Apossible
world is not any way that thingssome thingscould have been. Rather, it is a
totally determinate way in which absolutely everything could have been. Astate-
ment is said to be true at a possible world if it is logically necessary that, had things
been that way, they would have been as the statement (actually) says they are.
The shift from possibilities to worlds deserves more scrutiny than it usually
gets. The moderate realist starts from the fairly innocent idea that there are
various ways certain things could be or could have been, as well as the way they
actually are. But the semantics brings in the far less innocent notion of a totally
determinate way all things could have been. The latter concept is far from clear,
and it raises interesting metaphysical questions. Some philosophers, after all,
deny even that the way things actually are is fully determinate. There ought to
be a careful argument for casting the semantics in terms of worlds rather than
possibilities.
2
Leibniz had such an argument. He held that each monad mirrors or expresses
every monad with which it is compossible. On that basis, he argued that com-
possibility is an equivalence relation among monads; possible worlds may then
be identified as the corresponding equivalence classes. Modern theorists of pos-
sible worlds, though, do not take themselves to be committed to one of the more
extravagant principles of Leibnizs monadology, and what is striking is how lit-
tle argument they provide in favour of casting a theory of modality in terms of
worlds rather than possibilities.
Thus, in his seminal paper on possible worlds, Stalnaker rightly notes that
Lewiss attempt to domesticate worlds by presenting them as what we refer to
when we speak of ways things could have been does not support conceiving of
them as more things of the same kind as Iand all my surroundings. If possible
worlds are ways things might have been, then the actual world ought to be the way
2
At least, Benson Mates discerned such an argument in the Leibnizian texts. See Mates 1986,
chap. IV, 1.
Possibilities 155
things are rather than I and all my surroundings. The way things are is a property
or state of the world, not the world itself (Stalnaker 1976, 278). The point against
Lewis is well taken, but in fact the ordinary use of the phrase a way things could
have been is just as inhospitable to Stalnakers own conception of worlds. Martin
Davies spotted the crucial disanalogy:I might have had straight hair. That is a
way things might have been different, but it does not specify a way things might
have been, in the sense Lewis needs (Davies 1975, cited in Humberstone 1981, 314).
Quite so, and neither does it specify a way things could have been, in the sense
Stalnaker needs. For the way things would have been if Martin Davies had had
straight hair is not a fully determinate way that every object in the universe could
have been.
Kripke sees that a further argument is needed, but nowhere (in print, at
least) does he provide one. In the introduction to Naming and Necessity (1980),
he notes how school exercises in probability introduce people to possible
states of the worldthe things that I have been calling possibilities and that
he calls mini-worlds (1980, 16). He then remarks that full-blown worlds raise
mind-boggling questions that do not arise if we stick to mini-worlds, and duly
warns the philosopher of possible worlds not to allow his technical apparatus
[to] push him to ask questions whose meaningfulness is not supported by our
original intuitions of possibility that gave the apparatus its point (1980, 18). The
warning is apt, but we should not forget that we have been given no positive rea-
son for venturing down the path that makes it necessary.
All the same, Kripkes reference to the technical apparatus suggests it is
something in the semantic theory that is driving the shift from possibilities to
worlds, and if we take that hint, the motivation is not hard to spot. We may
approach the matter by contrasting the modal semantics of conjunction and
disjunction. We can explain truth at a possibility in just the way we earlier
explained truth at a world:a statement is true at a possibility if, necessarily,
things would have been as the statement (actually) says they are, had the pos-
3
sibility obtained. With the notion so explained, it is easy to say how a con-
junctions truth at a possibility relates to the truth there of its conjuncts. In any
normal modal logic, (C (A B)) is equivalent to (C A) (C B). So
we can specify the modally relevant aspect of the meaning of by saying:
(C)For any possibility x, the conjunction A B is true at x if and only if
A is true at x and B is true at x.
3
The use of the modal operator necessarily in glossing the notion of truth at a possibility pre-
cludes deploying the notion in any attempt to reduce modality to something putatively more primi-
tive, but moderate realists typically disclaim any reductive aspirations; see e.g. Stalnaker 1976, 33.
156 Five Attacks on Classical Logic
What, though, about disjunction? The semantic principle parallel to (C) would
say that a disjunctive statement is true at a possibility if and only if at least one
disjunct is true there. In other words, we shouldhave
(D1)For any possibility x, the disjunction A B is true at x if and onlyif
either A is true at x or B is true at x.
The if part of (D1) is correct. If, for example, the statement A is true at possibility
x, then the disjunction A B is also true there:from the premiss (C A) we
can infer (C (A B)). But the only if part of (D1) is wrong. For consider the
possibility that a child is at home. That is a way things could have been. And, nec-
essarily, if things had been that way, they would have been as the disjunctive state-
ment Either a boy is at home or a girl is at home says they are. Accordingly, that
disjunction is true at the specified possibility. However, neither disjunct is true
there. It is not necessarily the case that, had a child been at home, things would
have been as the statement A boy is at home says they are, for it could have been
the case that there were only girls at home. So the first disjunct is not true at this
possibility. For the parallel reason, it is not necessarily the case that, had a child
been at home, things would have been as the statement A girl is at home says they
are, so the second disjunct is not true at the specified possibility either. Given that
we glossed truth at a possibility using the necessitation of a conditional, the falsity
of the only if part of (D1) should not surprise us: (C (A B)) does not gener-
ally entail (C A) (C B).
There is some temptation to bring in possible worlds at this point, for we can
say that a disjunctive statement is true at a world if and only if one or other of the
disjuncts is true there. As we saw, worlds are assumed to be fully determinate,
and worldly theorists give cash value to that by postulating that, for any statement
4
and any world, the statement is either true at the world or false at it. That is, they
give content to the notion of full determinacy by postulating the necessary truth
of the Principle of Bivalence. In particular, then, each disjunct will be either true
or false at each world, and at a world where the disjunction is true, both disjuncts
cannot be false.
A similar argument for casting a modal semantics in terms of fully fledged
worlds may be run for the case of negation. Because a possibility may leave the
4
Thus, in a recent paper, Stalnaker writes that a possible world would be more accurately
labelled a possible state of the world, or a way that a world might be. It is something like a property
that a total universe might have, and it is a maximal property in the sense that saying that a world
has a particular property of this kind is enough to determine the truth or falsity of every proposition
(Stalnaker 2010, 21; last emphasis added).
Possibilities 157
truth of certain statements undetermined, the axiom (N1) fails to specify the
modal semantic contribution of negation:
(N1)For any possibility x, the negation A is true at x if and only if
A is not true at x.
The absence of any explicit argument for worlds in their writings makes one
suspect it is this consideration that has driven moderate realists to theorize
in terms of worlds rather than possibilities. The obvious semantic principles
for disjunction and negation fail if we relativize truth to possibilities, but they
pass muster if we relativize truth to worlds. So, when we move from exposi-
tory talk to serious theory, we should cast a theory of modality in terms of
worlds.
example of a rational activity:inquiry. Thus, told that only one child is at home,
Iinvestigate whether it is a boy or a girl. It is plausible to describe this inquiry as
one in which Idistinguish between the possibility that the child is a boy and the
possibility that it is a girl, and set to find out which of them actually obtains. So
described, though, the things Iam distinguishing between are possibilities, not
fully fledged worlds. As my inquiry progresses, Imay find myself exploring the
consequences of more determinate possibilities. But at no stage, we may be sure,
will Ientertain a fully determinate possibilityone so specific that it settles the
properties and relations of every object in the universe. For apprehending such a
possibility is beyond the intellectual capacity of any human thinker. In this basic
respect, then, possible-worlds theory seems to misrepresent the structure of our
rational activities.
Stalnaker knows that we cannot apprehend fully determinate ways things
could have been. Why does he still think that such entities are implicit in the
structure of our rational activities? Isuppose he must think, first, that something
is a possibility only if it could be realized in some fully specific way; and second,
that the possibilities we can and do entertain in thought are best conceived as
regions of these fully specific possibilities. Thus the possibility that Ihave red hair
leaves it undetermined whether Ed Miliband will win a General Election. But
there is also the possibility that Ihave red hair while Miliband wins an election,
and the distinct possibility that Ihave red hair while he does not. By iterating
this process, it may be suggested, we shall eventually reach fully determinate pos-
sibilities which do settle the truth or falsity of all statements. These possibilities
will be the points of modal space, and our rational activities are best conceived as
delineating regions or sets of these points.
But are they best conceived in that way? There are at least two reasons to doubt
it. First, the business of making a possibility more determinate seems open-ended.
There are possibilities that the child at home should be a boy, a six-year-old boy,
a six-year-old boy with blue eyes, a six-year-old boy with blue eyes who weighs
3 stone, and so forth. So far from terminating in a fully determinate possibility,
we seem to have an indefinitely long sequence of increasingly determinate pos-
sibilities, any one of which is open to further determination. But then, so far from
conceiving of our rational activities as discriminating between regions of deter-
minate points, we appear to have no clear conception of such a point at all.
A second reason is that some possibilities, or apparent possibilities, resist full
determination. At least, they do so given plausible premisses about modality that
Stalnaker accepts. He holds that there could have been people who do not actu-
ally exist, but he denies that there are things that could have been those people.
That is, he denies the unrestricted validity of the Barcan Formula. This position
160 Five Attacks on Classical Logic
is prima facie coherent, but let us consider the possibility that there are people
who do not actually exist. On the worldly theory, that possibility is a region of
fully determinate possibilities, but what could these fully determinate possibili-
ties be? Given that they are fully determinate, it ought to be determinate which
people each of them contains. That is, for each full determination of the possi-
bility that there are non-actual people, it ought to be determinate which people
there would be, were it to be realized. On the hypothesis that the Barcan Formula
is invalid, though, it is hard to see how this could be determinate. We know that
it is not determined simply by the identities of actual people, for each of the pos-
sibilities we are considering is one in which there are people who do not actually
exist. Ex hypothesi, though, there are no actually existing things which could
have been these people, and it is quite unclear what other features of the cosmos
would determine the identities of the hypothesized non-actual people. (We can
of course further determine the possibility by specifying what these non-actual
people are called, but that is not to specify who they are.) This problem may not
arise for the modal concretist:given that there are concrete people in other uni-
verses, he can say that their identities are determined in whatever ways our iden-
tities are determined. On a moderate realist view, though, we cannot represent
the possibility that there are non-actual people as a region of fully determinate
worlds, for it would appear to be inherently indeterminate which things these
worlds contain. So a moderate realist about modality who denies the Barcan
Formula ought to allow partly indeterminate possibilities alongside fully deter-
minate worlds. So far, then, from our rational activities having a structure which
demands that we think of possibilities as regions of fully specific points, some
possibilities that we can reasonably entertain resist being conceived in that way.
Thus the functional role that Stalnaker marks out for worlds is better played by
5
possibilities.
5
There may well be further benefits of switching from worlds to possibilities. Kit Fine has
recently argued for the advantages of a semantic theory for counterfactuals cast in terms of possi-
ble states of affairs over the more familiar theories that use the framework of possible worlds. Fines
states of affairs are like my possibilities in that they are not assumed to be complete, and this differ-
ence is part of what enables his theory to avoid difficulties that afflict the possible-worlds accounts
of counterfactuals. See Fine 2012a and 2012b.
Possibilities 161
serves as an extended preface to the second (1999) edition of Barwise and Perrys
Situations and Attitudes, John Perry remarks of Stalnakers theory:
If one is going to have objects and propertiesthe world and all the possible ways it could
bewhy not start off on a modest and common-sensical scale, with the properties that
we actually recognize? Thats what we [sc., Barwise and Perry] are trying to do. While
Stalnakers approach to possible worlds seems the most plausible, once one has gotten
that far, the move to studying limited situations and their properties, rather than the total
world and its properties, seems natural and inevitable. (Barwise and Perry 1985, xlv)
With that advertisement, one turns hopefully to the body of Barwise and Perrys
book, only to find that they duck the challenge that disjunctions pose for a seman-
tic theory cast in terms of possibilities or situations. All they say about disjunc-
tion is that a disjunctive statement describes those situations that are described
by either disjunct (Barwise and Perry 1999, 137). This claim may be interpreted
in such a way that it is correctbut not if a statements describing a situation is
equivalent to its being true there. In the absence of an explicit connection, then,
between the relations of describing and being true at, Barwise and Perrys account
of or does not provide what we want. Making that connection explicit, whilst an
interesting project, would take us far afield, so let us set the notion of describing
a possible situation aside, and revert to the problem of relating the truth at a pos-
sibility of disjunctive and negated statements to the truth at a possibility of their
6
components.
In a paper that also recommends possibilities over worlds, Lloyd Humberstone
(1981) addresses the problem for negation. As axiom (C) above shows, there is
no problem giving the modal semantics for conjunction in terms of possibili-
ties, and Humberstone then appeals to De Morgans Laws to define disjunction in
terms of conjunction and negation. Suitable as it may be for his purposes, however,
Humberstones approach fails to satisfy one of our desideratanamely, that of yield-
ing a modal semantic theory that non-classical logicians, notably intuitionists, can
accept. First and most obviously, a logician (such as an intuitionist) who does not
accept one or more of De Morgans Laws will not be able to accept Humberstones
treatment of disjunction. In fact, an intuitionist cannot even accept Humberstones
treatment of negation. Ineed not expound his theory fully here, but one of its pos-
tulates (the Axiom of Refinability) says that any possibility that fails to determine
the truth of a statement possesses one refinement (i.e., a more detailed specification
6
As far as Iknow, the only person working within the framework of situation semantics who
seriously addressed the present problem was my former Ann Arbor student, the late Stephen
Schulz; see his 1993, especially p.176. But his treatment is equivalent to Humberstones theory,
which Idiscuss in the following paragraph.
162 Five Attacks on Classical Logic
of it) at which the statement is true and another refinement at which it is false. In
other words, any statement whose truth is left undetermined must be determina-
ble as true and must also be determinable as false. This postulate is quite unaccep-
table to an intuitionist. For him, certain possibilities fail to determine the truth of
certain instances of Excluded Middle. For an intuitionist, though, there is no pos-
sibility at which any instance of Excluded Middle is false. We cannot, then, employ
Humberstones account here. In its range of application to non-classical logics, it
does not greatly improve on the worldly theory we are seeking to replace.
It follows that the closure of U is the smallest set of possibilities at which every state-
ment that is true throughout U is true.
Closure, in the present sense, is a closure operation in the sense favoured by lattice
8
theorists. That is to say, itis
9
INCREASING U Cl (U)
10
IDEMPOTENT Cl Cl (U)=Cl (U)
7
As remarked earlier (n.1), the present theory of possibilities permits various more precise
accounts of their metaphysical status. The chief constraint on such an account is that the operation
of taking the closure of an arbitrary set of possibilities should be well defined.
8
See for example Davey and Priestley 2002, 145. Following Kuratowski (1958), topolo-
gists demand more of a closure operation. They require in addition that Cl () = and that
Cl (U V)=Cl (U) Cl (V). Ido not assume that closure in the present sense satisfies either of these
additional conditions.
9
Proof:Suppose that x U. Then, for any statement A that is true throughout U, A will be true
at x. So x Cl (U).
10
Proof:To show that Cl Cl (U) Cl (U), suppose that x Cl Cl (U). Then any statement that is
true throughout Cl (U) is true at x. Suppose now that statement A is true throughout U. Then A
will also be true at any y Cl (U), i.e., A will be true throughout Cl (U). Hence A will be true at x.
Possibilities 163
and
11
MONOTONEIf U V then Cl (U) Cl (V).
Let us call a set closed when it is identical with its own closure. By idempotence,
the closure of any set is closed, and by monotonicity the closure of U is the small-
est closed set containing U. These entailments go through in a minimal metalogic
and a fortiori in an intuitionistic or classical metalogic (for minimal logic, see
Johansson 1936).
How does the notion of closure help construct a semantic theory of truth at a
possibility? Some terminology helps in outlining the answer. When a statement
is true at a possibility, let us call that possibility a truth-ground of the statement.
Let us assume that there is a background set comprising all the possibilities of
12
which we need to take account in assessing the statements in question. Given
that assumption, there will be a set |A| comprising all and only the truth-grounds
of an arbitrary statement A. Our axiom for conjunction may then be reformu-
lated as the principle that the set of truth-grounds of a conjunction is the intersec-
tion of the truth-grounds of the conjuncts:
Since this holds for any statement that is true throughout U, we have that x Cl (U), showing that
Cl Cl (U) Cl (U). The converse inclusion follows directly from INCREASING.
11
Proof:Suppose that U V. Then any statement that is true throughout V will be true through-
out U. Now suppose further that x Cl (U). Then any statement that is true throughout U will be
true at x. So any statement that is true throughout V will be true at x. So x Cl (V). Thus if U V
then Cl (U) Cl (V), as required.
12
There are statementsnotably, in set theory itselffor which this assumption fails. But Ishall
follow standard expositions of possible-worlds semantics in suppressing, for the present, the com-
plications that a more general theory would bring in.
164 Five Attacks on Classical Logic
13
Proof:Precisely as in 5.2, n.12.
Possibilities 165
represents a state of information to which we can advance on the day that follows
that represented by a. The meaning of each atomic statement is specified by the set
of nodes at which it is true; such a node will be one whose corresponding state of
information verifies the statement. It is assumed that this specification satisfies
the following two conditions:
(i)If a b and p is true at a, then p is true at b;
(ii)If S is a set of nodes that bars a and, for every b S, p is true at b, then p is
true at a.
A set S is said to bar a node a if there is a natural number n such that, on
every branch of the tree that passes through a, there is a member b of S which
is separated from a by at most n intermediate nodes and for which a b.
Accordingly, condition (ii) says that if a given atomic statement will be rec-
ognized as true within a finite number of days, it is already true. Condition
(i)says that we do not lose information about which atomic statements have
been verified.
Beths condition for a conjunctive statement to be true at a node is as expected:
A B is true at a if and only if A is true at a and B is true at a.
His condition for a disjunction to be true at a node may initially seem surprising:
A B is true at a if and only if for some S T, S bars a and, for every bS,
either A is true at b or B is true at b
(see Dummett 2000, 138), but we can now see its point. Let us say that a Cl (S)
whenever S bars a. Then, where |A| is the set of nodes at which A is true, Beths
axiom for disjunction may be re-writtenas
|A B | = Cl (|A | | B |),
14
Thus a Cl (S) whenever a S because 0 is a natural number. Cl Cl (S) Cl (S) because the sum
of any two natural numbers is again a natural number; etc.
166 Five Attacks on Classical Logic
Disjunction was one of two central logical notions that presented a problem
for a logico-semantic theory cast in terms of possibilities. How should we treat
negation?
It helps to begin by defining an ancillary notion. Let us say that one possibility
is a determination ofor simply determinesanother when any possible state-
ment that is true at the second is true at the first. Thus the possibility that a boy is
at home is a determination of the possibility that a child is at home. When y deter-
mines x, Ishall write x y. It is an immediate consequence of the definitions that
x y if and only if y Cl ({x}). It then follows from the properties of the closure
operator that the determination relation is a pre-order, i.e., is reflexive and tran-
15
sitive. Let us also count possibilities as identical when the same possible state-
ments are true at each. This ruling ensures that determination is anti-symmetric
and hence is a partial order. Determination is what Humberstone calls refine-
ment, but Iprefer my term for its useful resonances. Some philosophers describe
the colour scarlet as a determination of the determinable red; since ways the world
might be are naturally taken to be properties of the world (sc., of the cosmos), it
16
helps to bear this analogy in mind.
Let us say that two possibilities are compossible when there is a possibility that
further determines both of them. Thus x and y are compossible if and only if
z (x z y z). Let us also say that a possibility combines two others when it is a
maximally weak possibility that determines both combinants:z combines x and
y if and onlyif
x z y z w((x w y w ) z w ).
15
Proof:Reflexivity. By MONOTONE, we have {x} Cl ({x}), whence x Cl ({x}), i.e., x x.
Transitivity. Suppose that x y and y z. Then y Cl ({x}) and z Cl ({y}). The first of these yields
{y} Cl ({x}) whence Cl ({y}) Cl Cl ({x}) by INCREASING. So z Cl Cl ({x}). By IDEMPOTENCE,
Cl Cl ({x})=Cl ({x}). Thus z Cl ({x}, i.e., x z.
16
It is important not to confuse the relation of determination with that of relative possibility,
familiar from Kripkean possible-worlds semantics. They are wholly distinct notions. Apossibil-
ity y determines possibility x if it is logically necessary that x obtains if y does. Aworld v is pos-
sible relative to a world w if, given that w obtains, v could (logically) have obtained. Since Iaccept
the widely held thesis that S5 is the logic of logical necessity, Ihave suppressed the relation of
relative possibility in my analysis.
Possibilities 167
compossible, then, there will be such a thing as the combination of x and y, which
Ishall write as x y.
Alongside compossibility, we have the contradictory notion of two possibili-
ties being incompatible:they will be so when there is no logical possibility that
further determines both of them. Following Robert Goldblatt (1974) and Michael
Dunn (see e.g. his 1993), let us use to signify incompatibility:thus x y means
z (x z y z). This definition enables us to prove theses about incompatibil-
ity that Goldblatt and Dunn take to be axiomatic:it follows immediately from the
definition that incompatibility is symmetric; and, since determination is reflex-
ive, incompatibility must be irreflexive.
The orthocomplement, U, of a set of possibilities U comprises precisely those pos-
sibilities that are incompatible with every member of U. Thus x U if and only if
x y for all y U. This notion enables us to formulate the semantic axiom for nega-
tion very simply. Traditional logic called statements contraries when the truth of
one was incompatible with the truth of the other; this is green all over and this is
half blue and half pink are contraries of this is red all over. It is natural to distin-
guish As negation from its other contraries by identifying it as the weakest statement
incompatible with A, and our framework yields a gloss on weakest:absolutely any
possibility which is incompatible with the truth of A is a truth-ground of its nega-
tion. It then follows that the set of truth-grounds of A is precisely the orthocom-
plement of the truth-grounds of A:
(N ) |A| = | A| .
There is a snag with (N):the assumptions made so far do not guarantee that U is
closed whenever U is, so we have no guarantee that negated statements will con-
form to our master axiom (R). Ishall return to this problem in the next chapter, but
first it is in place to consider the validity of a classical law that constrains the relation
between conjunction and disjunction.
17
Putnam himself no longer thinks that classical logic is threatened by this challenge: see
Putnam 1994 and 2012, chaps. 6 and 8.However, his objection to Distribution retains its intuitive
force, and my reasons for resisting the challenge are different from those that now persuade him.
168 Five Attacks on Classical Logic
18
For each i between 1 and n, A, Bi entail A Bi (by -introduction) and hence
(A B1) ... (A Bn) (by -introduction). So, by n application of the classical -elimination
rule, A, B1 ... Bn entail (A B1) ... (A Bn) . Distribution then follows by -elimination
and Cut.
Possibilities 169
not discuss this argument either, because it raises essentially the same issues as
the third. That argument, which will be my focus, is that restricting Distribution
is necessary in order to preserve the possibility of a giving realist interpretation
of languages like L. Or better:a restriction is needed if we are to avoid an extreme
and implausible anti-realism about quantum mechanics.
In detail, the third argument runs as follows. Let S1,...,SQ be statements speci-
fying all the possible positions of our single-particle system S at time t, and let
T1,...,TR be statements specifying all its possible momenta at that time. Atheo-
rem of quantum mechanics is Heisenbergs Uncertainty Principle, from which
Putnam infers that the conjunction Si Tj is inconsistent for all i, j (Putnam 1968,
185, his emphasis). He then reasons as follows:
If Iknow that Sz is true, then Iknow that for each Tj the conjunction Sz Tj is false.
It is natural to conclude (smuggling in classical logic) that Sz (T1 T2 ... TR) is
false, and hence that we must reject (T1 T2 ... TR)i.e. we must say the particle
has no momentum. Then one measures momentum, and one gets a momentumsay,
one finds that TM . Clearly, the particle now has a momentumso the measurement
must have brought it into being. However, the error was in passing from the falsity of
(Sz T1) (Sz T2) ... (Sz TR) to the falsity of Sz (T1 T2 ... TR). This latter
statement is true (assuming Sz); so it is true that the particle has a momentum (even if it
is also true that the position is r3); and the momentum measurement merely finds this
momentum (while disturbing the position); it does not create it, or disturb it in any way.
It is as simple as that. (Putnam 1968, 186)
We may surely assume that any physical possibility is logically possible. Given
that assumption, it suffices, in order to invalidate Distribution, to find statements
A, B1,...,Bn, and a physical possibility x at which A (B1 ... Bn) is true and
at x while the conclusion (Sz T1) ... (Sz TR) is untrue at every physically
possible circumstance, and hence is untrue at x. Putnam, then, does not need to
claim that each conjunction Si Tj is logically inconsistent. All he needs is the
claim that no such conjunction is true at any physically possible circumstance.
In making that claim, Putnam is implicitly rejecting any interpretation of the
quantum theory on which Heisenbergs Principle records a merely epistemic
limitation. Let us call such an interpretation a strong realist interpretation. On
such an interpretation, all the Uncertainty Principle says is that we cannot meas-
ure both a systems position and its momentum at a given time, from which it
follows that we cannot know any conjunction in the form Si Tj . The strong
realist, however, is well placed to block the rebarbative conclusionnamely,
that the measurement of the particles momentum brings that momentum into
beingwithout departing from classical logic. For he can reject the claim that
a conjunction Si Tj is untrue at any physically possible circumstance (not to
mention Putnams further claim that if Iknow that Sz is true, then Iknow that
for each Tj the conjunction Sz Tj is false). The Uncertainty Principle tells us
that it is impossible to know a conjunction in the form Sz Tj :in order to know
such a conjunction, we would have to make a precise determination both of the
particles position at the relevant moment and of its momentum at that time,
which the Principle says is impossible. On the strong realist view, however, that
is all the Principle says, and for any realist, a statement may be true even though
it cannot be known to be true; there is, then, no legitimate inference from No
one can know that Sz Tj is true to Sz Tj is untrue (letalone to If Iknow
that Sz is true, then Iknow that Sz Tj is false). Accordingly, the strong realist
can, and should, reject Putnams claim that the conjunction Sz Tj is untrue
for each j:the conjunction is unknowable but it may well be true. On this con-
ception, the impossibility of coming to know a conjunction in the form Si Tj
provides no basis for claiming that it is untrue, and the epistemic limitations that
Heisenbergs Principle records, deep as they may run, are simply irrelevant to
the question of whether the falsity of (Sz T1) (Sz T2) ... (Sz TR) entails
Whilst this reply is correct as far as it goes, Ido not think that it gets to the
heart of Putnams challenge to Distribution. His argument does implicitly reject
any interpretation of the quantum theory whereby a system has at any instant
a determinate position and a determinate momentum even though it is impos-
sible in principle to discover what that pair of attributes is. There are, however,
good reasons for setting such interpretations aside. They are examples of what
physicists call hidden-variable theories and all face difficulties, notably over
the Kochen-Specker Theorem (see Kochen and Specker 1967). In the light of this,
Possibilities 171
which is the classical formula for the expected position of the particle when x is
2
randomly distributed with probability density |(x)| . Similarly, the dispersion of
A in the state , (A), is definedby
( A) = E ( A2 ) E ( A)2 ,
Possibilities 173
( X ) (P ) h /4 .
We now have enough of the quantum theory on the table to advance an argu-
ment for the conclusion that no conjunction in the form Si Tj can be true
more precisely, for the conclusion that there is no physically possible state at
which such a conjunction is true. First, though, we need to consider what it means
to say that a statement is true at a physical state. As with the case of truth at a
logical possibility, the natural account of this notion involves a modal operator.
Statement B is true at state may be understood tomean
(*)It is physically necessary that, if the system were in state , things would
be as B says they are.
Now suppose that what B says is that the position of the system has some precisely
specified value p. Two conditions must be met in order for (*)to be true. One con-
dition, of course, is that the expectation value of the systems position in should
be p. Asecond requirement, though, is that the dispersion of the position in
should vanish. Only if this second condition is met will (*)be true, for it is incon-
sistent with (*)that there should be some physical possibility of the systems being
in while its position is other than p. The same argument applies to the systems
momentum. If B attributes some precise momentum m to the system, then (*)can
be true only if there is no physical possibility of the systems being in but with
a momentum other than m. So, if B is true at , the dispersion of the momentum
at must vanish.
The argument here generalizes. Suppose that statement B attributes a precise
value to the observable A. The truth of (*)is inconsistent with any physical pos-
sibility of the systems being in the state , but the observables having an unex-
pected value there. So we have the following truth principle:
If a statement B attributes a precise value to an observable A, then B can be true
at a state only if the dispersion of A vanishes at that state.
only at eigenstates of the observable, and only when the value ascribed is the
expectation value of the observable at that state.
In order to see how the truth principle bears on Putnams claim that no con-
junction in the form Si Tj is true at any state, we need a semantic principle say-
ing in what conditions a conjunction is true at a state. Here the analogy with our
previous discussion of truth at a logical possibility is precise. As before, we cannot
assume that a statement is either true at an arbitrary physical state or false there,
but we can stillsay:
(C)The conjunctive statement A B is true at a state if and onlyif
A is true at and B is true at .
(C) is consistent with the explanation, (*), of truth at a physical state. It is physi-
cally necessary is plausibly taken to be a normal modal operator and, whenever
has a normal modal logic, (C (A B)) and (C A) (C B) are
equivalent.
The truth principle and (C) jointly entail that no conjunction in the form
Si Tj is true at any state of our system. For suppose that Si Tj were true at
the state . By the only if part of (C), it would follow that Si is true at and that
Tj is true at . By the truth principle, if Si is true at then the dispersion of the
systems position must vanish at . Again by the truth principle, if Tj is true at
then the dispersion of the systems momentum must vanish at . By Heisenbergs
Principle, however, that is impossible:both dispersions cannot vanish at a single
state. So at no state can Si Tj be true.
There is, then, a powerful argument for Putnams first main contention. Plausible
requirements for the truth of statements that attribute precise positions and
momenta to physical systems entail that no conjunction in the form Si Tj can be
true. By itself, however, this contention does not threaten the Law of Distribution.
The Law says that A (B1 ... Bn) logically entails (A B1) ... (A Bn) , but
so far we do not even know that Putnams long disjunction (Sz T1) ... (Sz TR)
is untrue. We know that none of its component conjuncts is true, but to infer from
that that the disjunction itself is untrue we would need the semantic principle that
a true disjunction must contain at least one true disjunct. In particular, we would
need the semantic principle that if A B is true then either A is true or B istrue.
That principle is, of course, very widely accepted, but the natural attempt to
prove it runs as follows:
1. Tr( A B ) premiss
2. Tr( A B ) A B principle about truth
3. A B 1, 2 modus ponens
Possibilities 175
4. A assumption
5. A Tr( A ) principle about truth
6. Tr( A ) 4, 5 modus ponens
7. Tr( A ) Tr( B ) 6, -introduction
8. B assumption
9. B Tr( B ) principle about truth
10. Tr( B ) 8, 9 modus ponens
11. Tr(A) Tr(B ) 10, -introduction
12. Tr( A ) Tr( B ) 3, 4, 5, 11 -elimination, with discharge of
assumptions 4 and 8
The problem with this proof, in the current dialectical context, is evident:its last
step appeals to the unrestricted form of the -elimination rule. Lines 5 and 9 may
well be true, but they are not logical truths and so must count as side premisses
in an application of that rule. Accordingly, the semantic principle that a true dis-
junction contains at least one true disjunct rests on a logical rule that is moot in
19
the present context of debate.
What we need to pursue the debate further is a semantic principle for disjunc-
tion whose implications for the validity or otherwise of Distribution are sta-
ble under the possible switch from classical logic to a weaker logic without full
-elimination. But what might such a semantic principle be? Ishall briefly can-
vass two familiar proposals before giving the answer that Iprefer.
He then proposed that we should just read the logic off from the Hilbert space
(1968, 179). However, a vital piece of the argument is missing. The operations of
intersection, span, and orthocomplement on the subspaces of a Hilbert space
19
In fact, if we follow Putnams 1968 recommendation and adopt quantum logic in place of clas-
sical logic when reasoning about subatomic particles, there are problems even in formulating lines
(2), (5), and (9)of the argument. Those premisses use a conditional, and there are difficulties intro-
ducing a well-behaved indicative conditional into the language of quantum logic. See Hardegree
1974 and Malinowski 1990.
176 Five Attacks on Classical Logic
(2) Later in the article, Putnam suggests a different way of specifying the
meanings of the connectives. Philosophers of science speak of operational defi-
nitions of geometrical terms:an operational definition of straight line might
be the path of a ray of light in vacuo. Towards the end of Is Logic Empirical?
(192ff.), Putnam asks whether the logical connectives have analogous opera-
tional definitions, and he elaborates sympathetically the affirmative answer
given by David Finkelstein (1968). On Finkelsteins approach, we begin by pos-
tulating that to every physical property there corresponds an experimental test
that a thing passes if and only if it has the property. There is a natural partial
order on the space of possible tests whereby T1 T2 if and only if anything
that would pass T1 would also pass T2 . Putnam contends that the disjunctive
statement Either a is P or it is Q will have an operational meaning only if there
is a test which is passed by all and only the things which have either the prop-
erty P or the property Q. If such a test exists, it will be the least upper bound,
with respect to the partial order , of the tests associated with P and with Q.
For Putnam, then, the operational meaning of or is given by the operation of
taking the least upper bound of two tests; dually, the operational meaning of
and is given by the operation of taking the greatest lower bound. Assuming
that all such tests exist, quantum mechanics entails that the space of tests is a
non-distributive lattice. The operational definitions of the connectives are then
supposed to explain how it is that Distribution is invalid for statements about
subatomic particles.
Dummett identified various technical difficulties in Putnams proposal, espe-
cially in its assumptions about when upper and lower bounds exist (see Dummett
1976b, 27880). The basic problem it faces, though, is philosophical. If quantum
theory is correct, then a certain space of possible tests forms a non-distributive
lattice. That fact threatens the logical Law of Distribution, however, only if opera-
tions on that space give the senses of the connectives. The claim that they do,
though, is highly doubtful. The word or seems to have the same sense in Either a
problem cannot be verified in polynomial time or it can be solved in polynomial
Possibilities 177
time as it has in The particles momentum is either m1 or ... or mR. Yet the sense it
20
bears in the first statement surely does not relate to any space of empirical tests.
if and only if the transition probability between and is zero, i.e.,
2
ifand only if |<|>| =0.
Given that the members of the state space are non-zero vectors, the condition
2
|<|>| =0 will obtain if and only if the inner product <|>=0. When the
inner product of two vectors is zero, quantum theorists call the vectors orthogo-
nal, so our symbol for incompatibility is apt. We have, then, a general semantic
principle, namely (N), saying when a negated statement is true at a state, and a
determination of what this general principle comes to when the states in question
are those of a subatomic physical system.
In fact, our analysis yields more. As Iargued above, a statement attributing a precise
value to an observable can be true only at an eigenstate of the observable. Atheorem of
20
Analogous problems confront John Bells attempt (in his 1986)to specify the senses of the con-
nectives, as they appear in statements about subatomic particles, by reference to proximity spaces.
178 Five Attacks on Classical Logic
quantum mechanics tells us that distinct eigenstates of a given observable are orthog-
onal. Now an atomic statement of L ascribes some precise value, v, to an observable,
B. Any such statement, then, amounts to rejecting the claim that the value of B is an
eigenvalue other than v. Accordingly, where A is any atomic statement of L, the set
|A| of states at which A is true will be identical with U , where U comprises all the
21
other eigenstates of B. Since is a symmetric relation, however, we have (1)U U .
We also have (2)if U V then V U . It follows from (1)and (2)that U=U and
hence that |A|=|A|. Now the operation that maps U to U is a closure operation
22
in the sense of 6.4. Thus, if we take Cl (U) to be U , the states at which an atomic
statement is true form a closed set. This result speaks strongly in favour of taking the
closure of U, where the members of U are the physically possible states of a subatomic
system, to be U , so that our axiom (D) specializes in this caseto
(D) | A B | = (| A | | B |) .
Given that the atomic statements have closed sets of verifying states, (C), (D), and
23
(N) ensure that the molecular statements of L also have closed sets of verifiers.
They also ensure that all De Morgans Laws hold. By themselves, however, they
do not guarantee the validity of Distribution. The inference from a single prem-
iss A to a conclusion B will be valid only if |A| |B|. Given (C) and (D), then,
Distribution will be valid only if |A| (|B| |C|) (|A| (|B| |C|)), when-
ever |A|, |B|, and |C| are closed sets. This condition may fail.
Have we, though, an actual counterexample to Distribution? We saw ear-
lier that none of Putnams conjuncts Si Tj is true at any physically possible
21
Proof:Suppose that x U. We require to show that x U that is, that x y for any y U.
Let y be an arbitrarily selected member of U . By the definition of U, y z for all z U. So, in par-
ticular, y x. By the symmetry of , x y. Since y is arbitrarily chosen from U , we have x U , as
required.
22
Proof:Proposition (1)immediately yields INCREASING. Proposition (2)and U =U yield
U=U, i.e. Cl (U)=Cl Cl (U), showing that closure is IDEMPOTENT. Finally, suppose that
U V. By (2), V U , whence U V by (2)again. Thus if U V then Cl (U) Cl (V), showing
that Cl is MONOTONE.
23
The closure requirement that |A|=|A|, and the clause for disjunction, are equivalent to those
that Goldblatt proposed in his semantics for orthologic and quantum logic; see Goldbatt 1974. My
argument for imposing these requirements, however, is very different from his. Goldblatt takes the
items with respect to which statements are assessed as true or as false to be possible outcomes of
empirical tests, not possible states of a physical system. He then cites Randall and Fouliss papers
on empirical logic as showing that the closure requirement arises from a restriction on what sets
of outcomes may be identified with propositions or events (Goldblatt 1974, 25). Idoubt if the
closure requirement can be justified within Randall and Fouliss framework but, in any event, their
approach to logic fits a sophisticated operationalism in the philosophy of science, not the realism
about quantum mechanics that Putnam was defending when he propounded the argument against
Distribution that we are considering.
Possibilities 179
considerably less straightforward than his exposition might lead one to suppose.
Putnam assures us that nothing is lost if we pretend for now that all physical
magnitudes have finitely many values, instead of continuously many, and in such
a world H(s) [i.e., the Hilbert space comprising all the possible physical states of a
system, S] would be just an ordinary finite dimensional space (1968, 178). This pre-
tence is central to his argument that there are physically possible states at which
Sz (T1 ... TR) is true. He takes the finite list T1,...,TR to exhaust all the pos-
sible momenta of the system, and claims on that basis that T1 ... TR is a valid
statement in quantum logic (184). As with the cognate claim that any Si Tj
is inconsistent, he needs only the less ambitious thesis that T1 ... TR is true
at every physically possible state of the system. For then, assuming that there is a
state at which Sz is true, the conjunction Sz (T1 ... TR) must be true at that
state too.
Is Putnam, though, entitled to that assumption? I do not think he is. The
truth principle tells us that a statement, such as Sz, that attributes a precise
value to an observable quantity, will be true only at eigenstates of that observ-
able. In a finite-dimensional Hilbert space, any self-adjoint operator will pos-
sess eigenstates, so in such a space there will be a state at which Sz is true. In an
infinite-dimensional space, however, even self-adjoint operators need have
no eigenstates at all. To be sure, if A and B are commuting self-adjoint opera-
tors on any Hilbert space H, then the span HA,B of the vectors that are simultane-
ously eigenstates for A and for B is the intersection of the subspace, HA, spanned
by A and that, HB , spanned by B. It follows from this that, when A and B com-
mute, if H admits an orthonormal basis of eigenvectors for A then HA,B=HA.
These theorems, however, apply only to observables that commute, and the
180 Five Attacks on Classical Logic
perplexities with which Putnam is wrestling arise precisely because the position
and momentum operators of quantum mechanics do not commute. Indeed, once
an infinite-dimensional Hilbert space is endowed with observables for position
and momentum, there are no eigenvalues for position:if a state were such, we
would have (X)=0, which would entail that (X) (P)=0, contracting the
Uncertainty Principle. Our argument for classifying (Sz T1) ... (Sz TR) as
untrue, however, rested on the truth principle, whereby each Si and Tj can be true
only at an eigenstate. By that same principle, then, there is no state at which Sz is
true. So Putnams case is one where neither the premiss nor the conclusion of an
instance of Distribution is true at any physical possibility. As such, his case is no
counterexample to the Law.
A natural reply to this criticism of Putnam is to say that the truth principle is
too restrictive:in order for a statement attributing a precise value to an observable
A to be true at , it is not necessary that the dispersion of A should be zero at ; it
is enough that it should be less than some small positive value, . One can select
, Sz, and the Ti in such a way that Sz meets the revised condition on truth whereas
each Sz Ti does not. Indeed, the intuitive force of Putnams counterexample to
Distribution derives precisely from the fact that quantum mechanics makes such
a selection possible. The fact remains, however, that the revised condition is not
a satisfactory constraint on truth. It may be a good constraint to impose on is as
good an approximation to the truth as we can make, but the laws of classical logic
never pretended to preserve that property of statements. Defended in this way,
then, Putnams objection to classical logic simply changes the subject.
Indeed, our analysis suggests that Putnams critique of the application of
classical logic to the quantum realm should really have been more radical.
Although he argues in favour of restricting certain classical laws, he takes it for
granted that, in reasoning about subatomic particles, deductive logic remains
applicable. In particular, he presumes that statements attributing positions and
momenta to small particles may sensibly be assessed for truth or falsehood at
various physically possible states, so that a logician may sensibly seek to identify
general laws that necessarily preserve truth at all such states. The truth princi-
ple, by contrast, suggests that such statements are so rarely true that a quest for
general laws saying when truth is preserved is quixotic. We would do better to
recognize that all the quantum theory gives us are probabilities of making cer-
tain measurements of observable quantities. In the light of that recognition, we
could still explore notions akin to consequence. Thus we might deem an argu-
ment to be quantum-mechanically valid if its conclusion has a high probability
whenever all its premisses have a high probability. More precisely, we might
say that an argument is valid if, for every positive , there is a positive such
Possibilities 181
that the conclusion is assigned a value greater than 1 by every wave function
that assigns a value greater than 1 to each of the premisses. (Compare Ernest
Adamss account of validity in his logic for conditionals; Adams 1975, 57.) It may
be that some simulacrum of classical logic could be vindicated on that basis.
(Thus Adamss logic validates the classical rules for the connectives other than
the conditional.) All the same, what would then be vindicated would only be
a simulacrum, for it would not really be a logic. On the account defended in
Chapter2, logic is constitutively concerned with the necessary preservation of
truth, not with probabilistic relations between statements. The upshot of our
analysis is that, to assess our reasoning about small particles, we need, not a
revised logic, but something that is not a logic. If God really does play dice with
the universe then a deductive system saying what must happen, given the laws
and the initial conditions, has very limited application:what we would instead
need in such a case is a system of rules for calculating the odds. That is, if the
quantum theory is probabilistic all the way down, then what we would need,
and all we could ever expect to find, would be a system of rules for ascribing
probabilities to observable outcomes. Such rules, though, do not constitute a
logic in the sense articulated in Chapter2.
But what is it, exactly, for all the premisses to be true at a given possibility? In
6.4 Idefined a combination of a finite collection of possibilities as a maximally
weak possibility (with respect to the partial order of determination) that deter-
mines all the combinants. This notion provides a natural answer to our question.
Anecessary and sufficient condition for B to follow logically from A 1,...,An is that
any combination of truth-grounds of the premisses should also be a truth-ground
of the conclusion:
| A1 | | An | | B |.
Here, |A| |B| is the set that contains all combinations x y, where x |A| and
y |B|. To be sure, this application of the notion of combination requires an exten-
sion of its range of application. Our earlier definition of the notion presumed that
the combinants are compossible. It cannot be assumed that all the members of
|A1| ... |An| meet this condition, so our theory needs to be supplemented with
an account of what x y means when x and y are not compossible. Ishall deal
with this in the next chapter. Note, though, that since |B| is a closed set of pos-
sibilities for any statement B, our condition for logical consequence is equivalent
to the requirement that any possibility in |A1| ... |An| is a member of Cl (|B|),
24
which is the condition for consequence in Sambin 1995.
The account of consequence is relativized to a background space of possibili-
ties. This is because, in assessing deductions, we latch on to whatever implication
relation is intended in the particular context of argument. The plethora of impli-
cation relations, though, do not preclude there being a privileged one of properly
logical consequence and Iargued in Chapter3 that a mark of logical consequence
is that it is absolute. That is:if a conclusion follows logically from some premisses,
then it follows logically from those premisses together with the assumption that
any other possibility obtains. This tells us something important about the space of
all logical possibilities. Let us call a space regular when, for any members x and y
of the space and any set of members V, x Cl (V) implies x y Cl (V y) for arbi-
trary y (V y is the set obtained by combining each member of V with y). If logical
consequence is absolute, the space of all logical possibilities will be regular. Given
the argument of Chapter3, then, let us lay down the following postulate:
Regularity The space comprising all logical possibilities is regular.
24
Because the truth-grounds of any statement form a closed set, we still have -introduction in
the form:if A is a logical consequence of some premisses X, and B is a logical consequence of X, then
A B is a logical consequence of X. See the proof of soundness for the rule &R in Theorem 4 of
Sambin 1995, at 868.
Possibilities 183
25
Proof:We need to show that Cl (UV) UV when V is closed. Suppose then that x Cl (U V)
and take an arbitrary u U. Given that the space of possibilities is regular, x u Cl (U V u). By
the definition of , it follows that x u Cl (V). Since V is closed, this means that x u V. That is
to say, for any u U, x u V, i.e. x UV, as required.
7
Challenges from the Infinite and
from the Infinitesimal
(N ) | A| = |A| .
That is, the truth-grounds of A are precisely those possibilities that are incom-
patible with any truth-ground of A.
Within the semantic framework of Chapter 6, (N) seems inevitable as the
axiom for negation, but there is a snag with it. Axiom (R) demands that the set
of truth-grounds of any statement should be closed. Supposing that a statement
meets this condition, (N) does not ensure that its negation also meets it. One
could respond to this problem by emending (N) so as to make the truth-grounds
of A the closure of |A|, rather than |A| itself, for the closure of any set is
closed. However, (N) is an intuitively compelling semantic postulate for the sign
for sentential negation, which it would be a shame to sacrifice.
Fortunately, we do not have to sacrifice it. So far, the only substantial con-
straints that we have placed on the space of logical possibilities are (a)that the
closure operation is well-defined on it and (b)that it is regular. If a space of pos-
sibilities is to sustain a well-behaved notion of negation, we must require more.
Different logical schools will have different views about what further require-
ments to impose, but both classical and intuitionist logicians will accept the fol-
lowing principle:
| A1 || An | | B |.
Some readers may find impossible states of affairs rebarbative, and a great deal
of argument would be needed to vindicate their inclusion in ones modal meta-
physics. Such argument would certainly breach the neutrality that Ihave striven
to maintain between different metaphysical conceptions of possibilities. It may
be possible to vindicate impossible states of affairs if one takes states of affairs to
be properties of the cosmos. After all, alongside properties that my table does not
have but could havesuch as that of being red all overthere are properties that
my table could not havesuch as that of being simultaneously red and green all
over. On the strength of the analogy with such impossible properties of tables,
someone who takes states of affairs to be properties of the cosmos may make con-
ceptual house room for impossible states. Clearly, though, the analogy will not
apply to other conceptions of states of affairs.
We need not pursue this matter, however, for Iam not recommending as an
element of a sound modal metaphysics. Rather, Iam suggesting that we should
introduce it as a convenient bookkeeping device that simplifies the algebra of
truth-grounds semantics. An analogy (which Iowe to a reader for the Press) may
help convey the spirit of the proposal. Models of the standard axioms of mereology
are clumsy to deal with, whereas it is easy to work with complete Boolean algebras.
Consequently, so as to enjoy the benefits of simpler mathematics, mereologists
often append a zero element to their theory of parts, without actually supposing
that there is an object that is a part of every other object. It is in the same spirit that
Iappend to our space of logical possibilities. Doing so simplifies the formal the-
ory of truth-ground semantics, and we shall not get into deep metaphysical water
so long as we recall that is nothing more than a bookkeeping device.
Having introduced , the definition of incompatibility needs to be emended;
for now, given any x and y, there will be a member z of the extended space of states
of affairs for which x z and y z. There is, however, a natural emended definition:
1
If x x= then w (x w w=). But x x, so x=.
Challenges from the Infinite and from the Infinitesimal 189
One benefit of adding the impossible state of affairs to our space is that we no
longer need (NC) to ensure that the truth-grounds of A form a closed set:with
2
in the space, the postulate of Regularity ensures that |A| is a closed set. The
crucial point is that is the only point in the space at which every statement
is true, so that Cl ({})={}. We can then argue as follows. Suppose that x
Cl (|A|). To show that |A| is closed, we need to show that x |A|; i.e., that xy=
whenever y |A|. Let y be an arbitrarily chosen member of |A|. Since x Cl (|A|),
Regularity yields x y Cl (|A| y). Since y |A|, Cl (|A| y)=Cl ({})={}. So
x y= whenever y |A|, exactly as required. As before, all these deductions
3
from (Inc) go through whether the metalogic is classical or intuitionist.
What logical laws does this account of negation validate?
In addressing this question, we need to take care. Our semantic principle for
negation makes use of the concept of negation. Apossibility is a truth-ground of
A if it is incompatible with any truth-ground of A, where (on our first account
of the matter) two possibilities are incompatible if it is not the case that some pos-
sibility determines both of them. The switch from this account of incompatibility
to one given in terms of does not affect the general point. is well defined only
if it determines every member of the space, i.e., only if any statement whatever
is true at . My argument to show that this condition is satisfied relied upon a
metalogical application of an inference rule involving negationnamely:From
not A , infer If A then B for arbitrary B.
This situation is surely unavoidable:there is no prospect of specifying the sense
of a negation sign without using a notion of negation. This means that in drawing
consequences from the specification a thinker will inevitably be reasoning with,
inter alia, negation. In the context of a discussion between adherents of classical
and intuitionistic logic, however, that in turn means that we need to take care lest
our validations and invalidations of logical rules become as futile as the proof of
the soundness of DNE discussed in 1.1. For classical logicians and intuitionists
go by different rules for negation.
It will help to have before us a formalization of these rules. Ichoose Prawitzs
famous natural deduction formalizationor, more exactly, the version of it
that takes negation to be a primitive sentence-forming operator rather than one
Note that if is a statement that is true only at , then A is equivalent to A , and the
2
closure of |A| follows from the result about conditionals proved in n.25 of Chapter6.
3
Relevant and dialetheist logicians, who reject Ex Contradictione Quodlibet, will reject this way
of ensuring that negated statements conform to (R). With the sense of specified by (N), however,
there are other ways of ensuring that negated statements meet appropriate closure constraints that
members of these schools will be able to accept. For an example, see the semantics for the relevance
logic R given in Goldblatt 2011, chap.6, noting particularly the axiom for negation on p.222.
190 Five Attacks on Classical Logic
defined in terms of the conditional (see Prawitz 1965, 201). Prawitzs formal lan-
guage includes a well-formed formula that is interpreted as being constantly
false. Within that language, both classical and intuitionists schools adhere to the
same rule for introducing negated statementsviz., Simple Reductio, the rule of
proof which licenses the deduction of A from X, given as premiss a deduc-
tion of from X together with A. But the rival schools have different rules for
eliminating negation. The intuitionists elimination rule is:from A together with
A , infer . Ishall call this Ex Contradictione Falsum (ECF). The classicists
rule is that of Double Negation Elimination (DNE):from A infer A. We also
need (cf. Prawitz 1965)an elimination rule for :from , infer B, for arbitrary B,
a rule Ishall label Ex Falso Quodlibet (EFQ). To avoid futility, we need to show that
the validations, and invalidations, of the contested logical rules that our semantic
4
theory yields are stable whether the metalogic is classical or intuitionist.
Before we can show this, we must extend our theory so that it assigns a
semantic value to the constantly false formula . Under the intended interpre-
tation, is true at no possible state of affairs. That is, the only state of affairs
at which is true is the impossible state of affairs (at which every state-
ment is true). Accordingly, we complete our interpretation of the language by
postulating
(F)|| = {}
Given (F), it is a trivial matter to show that EFQ is sound:the only state of affairs
at which is true is the impossible state of affairs, ; but any statement B is true
at ; so || |B| for any B. This argument could be formalized as a valid proof in
both classical and intuitionistic logic, so we have a validation of EFQ that is stable
under a switch from one of those logics to the other.
Our account of negation also stably validates Simple Reductio, the introduction
rule for negation that is common ground between classicists and intuitionists.
The condition for Simple Reductio to be sound is that A should follow logically
from X whenever the set of statements X {A} entails . Suppose then that the
set X {A} entails . In that case x y= whenever x |X| and y |A|. By trivial
quantifier manipulation, this condition implies that, whenever x |X|, x y=
for all y |A|, i.e., x |A|, i.e., x |A|. By (N), though, |A|=|A|, so whenever
x |X|, x |A|. So, when X {A} entails , |X| |A|, showing that the rule of
Simple Reductio is sound.
4
For this notion of stability, see again the passage from Dummetts Reply to John McDowell
(1987) quoted in 1.1.
Challenges from the Infinite and from the Infinitesimal 191
The trivial metalogical moves in the previous paragraph are again common
ground between classical and intuitionistic logicians, as are the rules implicit in
the definition of incompatibility. Members of both schools, then, will accept the
soundness argument just given. So, in the present dialectical contextnamely,
that of a discussion between adherents of those schoolswe may regard the argu-
ment as demonstrating the soundness of the rule of Simple Reductio.
Matters are similar with Ex Contradictione Falsum, the intuitionists rule for
eliminating negation, whose soundness the classicist also accepts. The condition
for this rule to be sound is that |A||A| ||, for any statement A. Now the only
member of |A||A| is the impossible state of affairs , at which is indeed true.
So the condition for soundness is met. Again, this argument could be formalized
as a correct deduction in a classical or intuitionistic metalogic. So it too may be
deemed to validate Ex Contradictione Falsum on the strength of our postulate for
negation.
In fact, the intuitionistic rules for , , , , and comprise a sound and
complete axiomatization of the present account of consequence. More precisely,
let us consider the semantic theory T1 in which (1)consequence is defined as pres-
ervation of truth at every state of affairs (where the states of affairs include all
the logical possibilities along with ); (2)the truth-grounds of any statement are
required to conform to (R); (3)Regularity holds; and (4)(C), (D), (I), (N), and
(F) specify the senses of , , , , and . Then Prawitzs natural deduction
rules for the intuitionistic propositional calculus are sound and complete with
respect to T1. Aproof of soundness is implicit in our analysis. Completeness may
be proved by constructing a canonical term model in the usual way (see e.g. 5
of Sambin 1995, 8714, especially Lemma 8c and Theorem 10). That completeness
proof could be formalized as a correct deduction in an intuitionistic metalogic,
5
and hence in a classical metalogic too.
What all this shows is that McDowells programme for pressing the intui-
tionist challenge against classical logic can beindeed, has beenexecuted.
According to McDowell (see again 5.5), the intuitionist does not need to adopt
a verificationist theory of meaning:he and the classicist can both accept that a
statements sense (i.e., the logically relevant part of its meaning) is given by its
truth-conditions. Our analysis bears McDowell out:both parties can make sense
of the fundamental semantic notion of a statements being true at a possibility
5
As Sambin observes, the result now is a simple and fully constructive completeness proof for
first-order BL [basic linear logic] and virtually all its extensions, including the usual, or structured,
intuitionistic and classical logic. Such a proof clearly illustrates the fact that stronger set-theoretic
principles and classical metalogic are necessary only when completeness is sought with respect to a
special class of models, such as the usual two-valued models (Sambin 1995, 861).
192 Five Attacks on Classical Logic
(or, more generally, at a possible state of affairs or at ) and they can accept that
a statements sense is given by specifying at which states of affairs it is true. These
states of affairs may be taken to be the statements truth-conditions. McDowell
also says that the particular clauses that specify the connectives senses can be
common ground between the rival logical schools; again, our analysis supports
him. The logico-semantic principles proposed for the connectivesnamely, (C),
(D), (I), (N), and (F)are at once inherently plausible and acceptable to classi-
6
cists and intuitionists alike. The requirements that a statements truth-grounds
must form a closed set, and that the space of logical possibilities must be regular,
are also principles that both classical and intuitionist logicians can and should
accept.
Indeed, we have gone further than McDowell in identifying common ground
between the two logical schools. On McDowells view, while the classicist and
the intuitionist assign the same senses to the connectives (within a common
general conception of sense) they differ over the meaning of logically follows
from. McDowells intuitionist gives a two-stage explanation of logical conse-
quence. The condition for B to follow logically from A 1,...,An is that the condi-
tional (A1 ... An) B is logically true; a statement is then deemed to be
logically true if it can be known to be true solely in virtue of the senses of the log-
ical constants and whatever the epistemic status of [its] components (see again
McDowell 1976, 60). Aclassical logician will certainly not accept this account of
consequence but we found reasons, quite independent of any adherence to clas-
sical logic, to doubt both the interest of the consequence relation that McDowell
here describes, and whether that description can be converted into a logically
rigorous characterization. As McDowell says, his intuitionists account of conse-
quence demands to be made precise ... by a systematic specification of the con-
ditions under which we may claim to know the truth of complex sentences, in
terms of the conditions under which we may claim to know the truth of simpler
sentences (McDowell 1976, 60). Any attempt to give such a specification runs
straight into the problems that afflict Heytings axioms for the connectives (see
again 5.2).
6
Robert Goldblatt has developed a semantic theory with respect to which the rules of Anderson
and Belnaps relevance logic R are sound and complete (see Goldblatt 2011, chap.6). The basic notion
of his theory is that of a state of informations verifying a statement, not that of a statements being
true at a possibility. Modulo that difference, though, his semantic axioms for the connectives are
also (C), (D), (N), and (I). As with classical and intuitionistic logics, the distinctive properties of R
arise from postulates about the structure of the underlying space (in his case, a space of states of
information). Ishall not expound Goldblatts postulates here, but his semantic theory provides fur-
ther support for the thesis that versions of (C), (D), (N), and (I) are faithful to the ordinary mean-
ings of the logical connectives and are common ground between various logical schools.
Challenges from the Infinite and from the Infinitesimal 193
In this respect, the case for intuitionistic logic that rested on the Heyting seman-
tics was more powerful. Since the existence of a method for refuting any claim to
have a refutation of A need not amount to a constructive proof of A, Heyting and
his followers had positive reason to regard the DNE rule as unsound. They were
not restricted to pointing out that we have, as yet, found no reason to suppose that
it is sound.
I now propose a way of breaking the deadlock. Let us say that a statement is
incompatible with a state of affairs x if and only if each of its truth-grounds is
incompatible with x. Thus, in our notation, A is incompatible with x if and only if
|A| x. Let us then say that a statement A has a back if there is a set of possible or
impossible states of affairs whose members are precisely those states with which
A is incompatible. Thus A has a back if and only if |A|= U, for some set U of pos-
sible or impossible states of affairs.
A sufficient condition for a statement to satisfy the DNE condition is that it
should have a back. To show this, it helps to demonstrate first that, for any set of
states of affairs, U, U =U . To prove this lemma, we first establish that U U
for any set U. Suppose, then, that x U and consider an arbitrary y U . By the
definition of U, y z for all z U. In particular, then, y x. Since is sym-
metric, it follows that x y. Since y was arbitrarily chosen from U , we have
x y for all y U, which is to say x U . This shows that U U . This result
immediately gives us that U U . For the converse inclusion, remark that
V U whenever U V (this follows directly from the definition of U ). So
U U also entails that U U , thereby completing the proof of the lemma
that U =U . Suppose now that the statement A has a back. Then there exists
a set U for which |A|=U. By the lemma, though, we have U=U and since
|A|=U it follows that |A|=|A|. In other words, any statement with a back sat-
isfies the DNE condition.
A reformulation of the condition for a statement to have a back may cast fur-
ther light on the notion. Let us assume that denying a statement is logically equiv-
alent to asserting its negation:both classical and intuitionist logicians will grant
this assumption. Then a statement has a back precisely when an assertion of it ipso
facto amounts to a denial of some other statement. Both classical and intuitionist
logicians assume that any statement has a negation. Astatement with a back will
also be a negation, or at least be equivalent to a negation. That is, A has a back if
and only if, for some statement B, A is equivalent to B ; to assert A will be to
deny B. We can now see more clearly why statements with backs satisfy the DNE
condition. For any formula B, the triple negation B is intuitionistically
equivalent to the single negation B . Suppose, then, that A has a back. Then, for
some B, A is equivalent to B , so that A is equivalent to B . By the
Challenges from the Infinite and from the Infinitesimal 195
result about triple negations, this means that, whenever A has a back, A is equiva-
lent to A , even in intuitionistic logic.
Our semantic proof two paragraphs back that a statement with a back satisfies
the DNE condition is also intuitionistically valid. At no stage does it appeal to
Excluded Middle, or to DNE itself, or to any of the classical logical laws that an
intuitionist cannot accept. This shows us where, in the recommended semanti-
cal framework, the issue between the classical and the intuitionist logician really
lies. In the previous section, we were led to consider the semantic theory T1 in
which (1)consequence is defined as preservation of truth at every state of affairs
(where the states of affairs include all the logical possibilities along with ); (2)the
truth-grounds of any statement are required to conform to (R); (3)Regularity
holds; and (4)(C), (D), (I), (N), and (F) specify the senses of the connectives , ,
, , and . As we noted, the rules of the intuitionistic propositional calculus
are sound and complete with respect to T1 whether one works in a classical or
intuitionist metalogic.
We now setalongside T1 a semantic theory T2 which differs from T1 only in its
second clause. Instead of requiring merely that each statement in the language has
a closed set of truth-grounds, which is the content of (R), we require in addition
(B)Each statement in the language has a back, i.e., for any statement A, there
exists a set U for which |A|=U .
T2 then differs from T1 in that clause (2) is replaced by clause (2): the
truth-grounds of each statement are required to conform to both (R) and (B).
What we have shown is that all the laws of classical logic are sound with respect
7
to T2. Moreover, the logical laws we appealed to in showing this are all accept-
able to the intuitionist. The situation is thus the mirror image of that with the
theory T1, which imposes the weaker closure requirement (R). It is the rules
of the intuitionistic propositional calculus that are sound and complete when
consequence is defined by T1; the proof of this result goes through whether the
metalogic is intuitionistic or classical. When consequence is defined by T2, per
contra, DNE and the other classical rules are sound; the proof of this result goes
through even in an intuitionistic metalogic. The assessment of which laws are
sound with respect to T1 and T2, then, is stable under a change between classical
and intuitionist metalogics.
7
In an intuitionist metalogic, we cannot infer from the soundness of a particular instance of
DNE to the validity of the corresponding instance of the Law of Excluded Middle:in intuitionist
logic, A A does not entail A A . However, if every instance of DNE is sound then every
instance of Excluded Middle is valid.
196 Five Attacks on Classical Logic
Within the semantic framework recommended here, then, we may see the dif-
ferences between the two logical schools as arising from different requirements
on the truth-grounds of a bona fide statement. Both parties require that the space
of states of affairs be regular and that the truth-grounds of any statement should
form a closed set:that is, they both accept postulate (R). The classicist, though,
thinks that the closure requirement is by itself too weak:he also requires (B).
In having the choice of logic depend on the choice between our two semantic
theories, T1 and T2, Ithink we can make out the best semantic case for classical
logic and the best case for intuitionistic logic. Amain theme of Chapter5 was
that the constructivist or verificationist semantic theories that are often invoked
to justify the choice of intuitionistic logic sell that logic short. For the reasons
given there, those theories give an implausible account of the way that construc-
tivist mathematicians, letalone speakers discussing empirical matters, actually
use the logical connectives. In a similar way, the classical semantics embodied
in the truth-tables sells classical logic short. Just as intuitionistic logic is used in
areas where constructivist semantics cannot be applied, so classical logic is used
in areas where Bivalence is doubtful. Many set theorists doubt, for example, if
the Generalized Continuum Hypothesis must be either true or false; certainly, its
bivalence is far from obvious. All the same, almost all set theorists use classical
logic in their proofs. Our semantic theories avoid both these drawbacks. The axi-
oms for the connectives that are common to T1 and T2 are plausible descriptions
of the use we make of connectives; and to impose (B) as a closure requirement on
the truth-grounds of any statement is not to postulate Bivalence.
But if we have reached the best account of the choice between our two logi-
cal schools, the crucial question remains:which school is right? Is it a universal
requirement that a statement should have a back? Ishall consider some puta-
tive examples of backless statements in the next two sections, but (B) has great
intuitive plausibility. The mediaeval logicians had a principle, Eadem est scientia
oppositorum:in order to attain a clear conception of what it is for A to be the case,
one needs to attain a conception of what it is for A not to be the case. But then, hav-
ing attained that conception, it is hard, at least prima facie, to see how an assertion
of A could fail to be equivalent to a denial of A . To put the point in terms of our
semantic framework, it is hard to see how a statement could have a determinate
content unless it is determinate which possible states of affairs it excludes. (Both
Ramsey and Dummett, it may be recalled from Chapter4, took the content of a
statement to be given by the states of affairs that it excludes.) The attractiveness of
(B) is particularly clear when we consider the use of statements to express hypoth-
eses. If one focuses merely on assertive uses of statements, then one may perhaps
imagine that a statement could possess a determinate content by virtue of its
Challenges from the Infinite and from the Infinitesimal 197
conception of the infinite should force one to revise classical logic. In particular,
we need to ask why it should lead to restricting the DNE rule and to denying that
every instance of A A is a logical truth.
Dummett tried to explain how the intuitionists conception of the infinite
bears on logic. Discussing quantification over an infinite totality, hewrote:
Given his assumption that the application of a well-defined predicate to each element of
the totality has a determinate value, true or false, the classical mathematician concludes
that its universal closure has an equally determinate value, formed by taking the [logi-
cal] product of the values of its instances, and that the existential closure likewise has
a determinate value, formed by taking the sum of the values of its instances. On such a
conception, the truth-value of a quantified statement is the final outcome of a process
which involves running through the values of all its instances; the assumption that its
truth-value is well-defined and determinate is precisely the assumption that we may
regard an infinite process of this kind as capable of completion. (Dummett 2000, 41)
Dummett shows here that someone who adopts the conception of the infinite that
Brouwer opposes has a general justification for the claim that statements involv-
ing quantification over infinite domains are bivalent. Dummett further contends
that if one accepts Brouwers strictures about the infinite, one will be left without
any general justification for that claim, i.e., one that applies to every statement
involving quantification over any infinite domain:
on an intuitionistic view, neither the truth-value of a statement nor any other mathemati-
cal entity can be given as the final result of an infinite process, since an infinite process is
precisely one that does not have a final result:that is why, when the domain of quantifica-
tion is infinite, an existentially quantified statement cannot be regarded in advance [of
proof] as determinately either true or false, and a universally quantified one cannot be
thought of as being true accidentally, that is independently of there being a proof of it, a
proof which must depend intrinsically upon our grasp of the process whereby the domain
is generated. (Dummett 2000, 41)
Even if Dummett is right in this contention, however, it does not settle our ques-
tion. We wanted to know whether Brouwers strictures on the infinite force any revi-
sions to classical logic, not whether they force one to abandon classical semantics. If
the senses of the connectives are specified by (C), (D), (N), and (I), there is a natural
way to validate classical logicnamely, by requiring bona fide statements to have
backswithout postulating Bivalence. Given those specifications, then, Dummetts
arguments about Bivalence are not to the immediate point.
Some of Brouwers writings, though, deploy his strictures on the infinite to
mount a direct challenge to classical logic. In 2 of his 1923 lecture and paper, On
the Significance of the Principle of Excluded Middle in Mathematics, Brouwer
identifies two fundamental propertiesby which he means propositions that
200 Five Attacks on Classical Logic
In other words, we cannot assert that the species of integers kn is either finite or
infinite.
Brouwers species is surely well defined. This is because, for any integers m and
n, there is a finite procedure that decides whether m=kn. For suppose we wish to
find out whether 538,763=k 2. To do this, it suffices to calculate to the first 538,772
decimal places. If the last 10 digits in the expansion are 0123456789, and if that
segment has occurred precisely once earlier in the expansion, then 538,763=k 2;
otherwise, it is not. ATuring machine could be programmed to apply this test,
and it would report an answer in a finite time. For this reason, it seems clear that
Brouwer has identified a mathematically well-defined species of integers.
Why, though, does Brouwer maintain that we cannot assert that the species is
either finite or infinite? His argument is not fully explicit but a sympathetic elabo-
ration of it runs as follows. The species of kns is finite if and only if there are only
finitely many segments of the form 0123456789 in the decimal expansion of ; it is
infinite if and only if there are infinitely many such segments. Accordingly, if we
were entitled to assert Brouwers species is either finite or infinite, we would also
be entitled to assert Either (1)there are only finitely many segments 0123456789
in the decimal expansion of or (2)there are infinitely many such segments.
Given Brouwers strictures on the meaning of talk about the infinite, however, a
statement about an infinite sequence must be cashed out in terms of the principle
or rule that generates it. Given that, alternative (1)can only mean that the rule
for expanding entails that there are only finitely many segments of the form
0123456789 in the expansion. Pari passu, alternative (2)can only mean that the
rule entails that no bound can be set on the number of such segments. The present
instance of Excluded Middle, then, must mean:either (1)the rule for expanding
entails that there are only finitely many segments of the form 0123456789 in
the expansion, or (2)the rule for expanding entails that no bound can be set
on the number of such segments. At first blush, it certainly seems that we are not
entitled to assert that disjunction in our present state of knowledge. Of course,
our knowledge might expand in such a way that we become entitled to assert it.
Challenges from the Infinite and from the Infinitesimal 201
For example, a mathematician might prove, on the basis of the rule for expanding
, that there could be at most three occurrences of the segment 0123456789 in its
expansion; we would then know that alternative (1)obtains and hence be entitled
to assert the present instance of Excluded Middle. In our present state of knowl-
edge, though, we are not entitled to assert that either (1)or (2)obtains, and so we
cannot assert that Brouwers species is either finite or infinite.
In considering this matter more closely, it helps to work with a slightly simpler
example. At the time of writing, has been calculated to the first ten trillion
13
(10 ) digits. Ido not know whether those ten trillion digits include a segment
0123456789, but let us suppose that they do not. (If they do, one could easily change
the designated segment to one that does not appear in the largest expansion of
that we currently have.) Let us now consider the statement Either Brouwers
species of kns is inhabited or it is not. Given our supposition, it would seem that
we are not entitled to assert this instance of Excluded Middle. Brouwers spe-
cies is inhabited if and only if the segment 0123456789 occurs somewhere in the
decimal expansion of ; it is uninhabited (i.e., empty) if and only if no such seg-
ment occurs. So we would be entitled to assert Either Brouwers species is inhab-
ited or it is not only if we were also entitled to assert Either 0123456789 occurs
somewhere in the expansion of or it does not. Given Brouwers strictures on
statements about the infinite, the latter instance of Excluded Middle can only
mean:Either (1)the rule for expanding entails that the segment 0123456789
occurs somewhere in the expansion, or (2)the rule for expanding entails that
no such segment occurs anywhere. In our current state of knowledge, it seems
that we are not entitled to assert this disjunction. As before, this might change.
In calculating to the first twenty trillion digits, we might find a segment
0123456789; we would then know that alternative (1)obtains. Equally, a math-
ematician might prove that (2) obtains by deducing a contradiction between
the supposition that such a segment occurs and the principles that regulate the
expansion of . In our present state of knowledge, though, we appear to have
no basis for asserting that either (1)or (2)obtains; hence we cannot assert that
Brouwers species is either inhabited or not.
Assuming for a moment that we are not entitled to assert Excluded Middle in
this case, what prevents us from doing so? Let A be the statement The segment
0123456789 occurs somewhere in the expansion of . Given Brouwers view of
infinity, it seems that A has no back. We have a clear conception of circumstances
in which we would be entitled to assert Aif an instance of 0123456789 were to be
found starting at the 500,000th place in the expansion, for example. However, given
Brouwers strictures on the infinite, there is no propositional content whose rejec-
tion is tantamount to accepting A. It might seem as though there is. It might seem as
202 Five Attacks on Classical Logic
8
In part because of a notorious obiter dictum in the transcript of a 1939 lectureIntuitionism
is all boshentirely (Wittgenstein 1975, 237)some scholars maintain that Wittgenstein did
not take Brouwers strictures about the infinite seriously. Ibelieve, to the contrary, that Part V of
Remarks on the Foundations of Mathematics (which dates from 19424) is the product of an intense
engagement with some of Brouwers discussions of the implications of those strictures. Iset out the
textual evidence for this view in Rumfitt 2014b.
Challenges from the Infinite and from the Infinitesimal 203
Moreover, Wittgenstein finds reason to doubt that we do know what the question
means, except when it is posed in rather special circumstances:
For how do Iknow what it means to say:the pattern ... occurs in the expansion? Surely by way
of examples:which show me what it is like for ... [to occur]. But these examples do not show
9
me what it is like for this pattern not to occur in the expansion!
Might one not say:if Ireally had a right to say that these examples tell me what it is like for
the pattern to occur in the expansion, then they would have to show me what the opposite
means (Wittgenstein 1978, 271).
To say of an unending series that it does not contain a particular pattern makes sense only
under quite special conditions.
That is to say:this statement has been given a sense for certain cases.
10
Roughly, for those where it is in the rule for this series, not to contain the pattern...!
If we try to make sense of the hypothesis that 0123456789 appears nowhere in the
expansion of when these special conditions do not obtain, we must entertain
9
The italicized not, although clearly present in Wittgensteins manuscript, is erroneously omit-
ted from both the German and English editions of Remarks on the Foundations of Mathematics.
Iam very grateful to Professor Joachim Schulte for pointing this out to me.
10
Wittgenstein 1978, 2689. See also Philosophical Investigations (Wittgenstein 1953)51617,
where the question Does the sequence 7777 occur in the expansion of ? is presented as one which
we seem to understand, but which leads Wittgenstein to ask whether we can be mistaken in think-
ing that we understand a question.
204 Five Attacks on Classical Logic
Does it make sense to say:While there isnt any rule forbidding the occurrence, as a mat-
ter of fact the pattern does not occur?And if this does not make sense, how can the
opposite make sense, namely, that the pattern does occur? (Wittgenstein 1978, 277)
This last question, Itake it, expects the answer:the opposite statement does not
make sense. We are to preserve (B), then, by holding that some sentences that one
might initially hear as expressing thoughts do not in fact succeed in doing so. In
this way, Wittgenstein offers a way of preserving classical logic while conceding
the correctness of Brouwers strictures about the infinite.
I wish to argue, though, that there is a less radical way of preserving classi-
cal logic while respecting what is right about those strictures. The phrase that
needs unpacking is Wittgensteins it is in the rule for the series, not to contain
the pattern (1978, 269). As we have seen, for any natural number m, there is an
effective procedure for deciding whether the first instance of 0123456789 in the
decimal expansion of begins at the mth place. That is, using Brouwers nota-
tion, there is an effective procedure for deciding whether m=k 1. For a given m
such that m k 1, then, it is in the rule for the series that m k 1. Suppose now
that the expansion of does not contain any instance of 0123456789. To sup-
pose as much is precisely to suppose that m k 1 for every natural number m.
Let us assume for a moment that we can understand an axiom system which
characterizes the natural-number structurei.e., which characterizes the nat-
ural numbers up to isomorphism. (I shall soon return to consider whether this
assumption may be justified.) If the expansion of does not contain the pat-
tern, it will lie in the rule for expanding , along with the characteristic axioms
of number theory, that it does not contain the pattern. In order to understand
the statement The pattern occurs nowhere in the expansion, then, we do not
have to make sense of the thought Wittgenstein claims makes no sense, viz.,
While there isnt any rule forbidding the occurrence, as a matter of fact the pat-
tern does not occur. For if the pattern occurs nowhere in the expansion, it will
lie in the rule for expanding , together with the axioms that characterize the
sequence of natural numbers, that it occurs nowhere. In order to understand the
Challenges from the Infinite and from the Infinitesimal 205
X ( X 0 n ( Xn Xs(n)) nXn).
In this formula, s(n) stands for the successor of n, and X is the full
second-order quantifier:that is, the second-order variable X is taken to range
over arbitrary subsets of natural numbers. An intuitionist will make no sense of
the notion of an arbitrary subset of the natural numbers. Whilst he understands
the idea of an arbitrary subset of a finite set, extending the notion of an arbitrary
selection of elements from a finite to an infinite set involves precisely the sort of
projection that he regards as illegitimate (see further 10.2). So we cannot appeal
to second-order Peano Arithmetic in making sense of the possibility that the laws
of number entail that a given sequence does not appear in the decimal expansion
of , even though it is impossible to prove from those laws that it does not appear.
There are, however, other categorical axiomatizations of number theory that
enable us to make sense of this possibility while respecting intuitionistic scru-
ples. Following Frege (1879), let us say that the ancestral, R*, of a relation R holds
between x and y if there is a finite sequence of objects a0,...,an such that a0=x,
an=y, and for each i between 0 and n1, Raiai+1. Thus, where Rxy means x is a
parent of y, R*xy obtains if and only if x is an ancestor of y. Following John Myhill
11
(1952), we can then introduce ancestral logic. We augment the language of the
first-order predicate calculus (with identity) with an ancestral operator, A:if
is a well-formed formula in which x and y occur free, and if t1 and t2 are singular
terms, then Axy t1 t2 is a well-formed formula in which the variables x and y
are bound. Intuitively, Axy t1 t2 says that the designatum of t1 stands to the
11
In fact, the formalization of ancestral logic sketched here is that given in Shapiro 1991 (227ff.).
Challenges from the Infinite and from the Infinitesimal 207
12
For an interesting discussion of the first sort of doubt, see Field 1998.
208 Five Attacks on Classical Logic
13
If the statement is false, a refutation will exist. For if it is false, there will be a counterexample
to the Conjecture, and in that event it will be possible in principle to identify all the prime numbers
less than the counterexample and verify that no pair of them has the counterexample as its sum.
However, there is at present no reason to assert that the Conjecture must have either a proof or a
refutation. For all we know, there may be no counterexample to it, but at the same time no proof that
every even number greater than two is the sum of two primes.
Challenges from the Infinite and from the Infinitesimal 209
conception of the domain of natural numbers, then, our thinker can assert any
instance of the following schema, where n ranges over the integers:
There is no need at all to appeal to the incoherent idea of running through all
14
the natural numbers in order to justify asserting NOS. Now let Fn mean n is
the sum of two primes if n is an even number greater than two . An instance of
NOSis
14
NOS is the Numerical Omniscience Scheme of semi-constructive theories of arithmetic.
However, while NOS may express the decidability of n(n) on the Heyting interpretation of the
connectives and quantifiers, it expresses the determinacy of the domain of numbers on the inter-
pretation given here. In retaining NOS as a convenient label for the schema, then, we should set
aside the original meaning of the acronym.
210 Five Attacks on Classical Logic
15
Aristotles arguments, indeed, persuaded Leibniz and Kant as well as Poincar, Weyl, Brouwer,
and Ren Thom among recent mathematicians. See the references in Bell 2008, whose exposition of
the SIA theory of infinitesimals Ifollow.
Challenges from the Infinite and from the Infinitesimal 211
Geometry (SDG). As will emerge, this theory does better than Robinsons at cap-
turing the geometrical intuitions that underpinned Newtons and Leibnizs use
of infinitesimals. Focusing on the simple case of a curve in the plane, Lawvere
and Kock first propose an informal Principle of Microstraightness. Let us call a
plane figure that is not identical with a single point non-degenerate. Then the
Principle says that, for any smooth curve C and any point P on C, there is a small
but non-degenerate segment of Ca microsegmentaround P which is straight.
2
By applying the Principle to the smooth curve y=x at the origin, it follows that
2
there is a set of numbers for which =0 that does not reduce to the singleton
{0}. comprises the nilsquare infinitesimals. Anilsquare quantity will be nilpo-
tent for powers higher than2.
Nilsquare infinitesimals open to the way to a very elegant treatment of the defi-
nite integral. It is natural to think of the area under a curve as consisting of vari-
ous rectangles with small bases, plus small residual areas between the top of each
rectangle and the curve. If each rectangle has a base of length , then each residue
2
will be a right-angled triangle whose area is proportional to . Thus, if is a nils-
quare infinitesimal, each residual area will be zero. Hence, when the area under
the curve is divided into rectangles whose bases are nilsquare infinitesimals, that
area is the sum of the rectangles, without residue. In this way, the theory of inte-
gration of smooth functions may be reduced to classical geometry and elemen-
tary algebra. It may, initially, be thought surprising that a very thin sector of a
circle is in fact a triangle, but once one sees the benefits one may come to swallow
this. Asector whose radius is r and whose base is the nilsquare will be a triangle
whose area is r (e.g. the triangle OAB in Figure 7.1).
A B
2
Figure7.1. The SIA proof that the area of a circle with radius r is r .
212 Five Attacks on Classical Logic
Since the circumference is 2r, the whole circle will be composed of 2r/ of these
sectors. In this way, we arrive effortlessly at the area of the circle:this will be
2
2r/ r=r .
This conception inspired the following first-order formal theory, called Smooth
Infinitesimal Analysis (SIA) which characterizes the smooth real number-line,
R. In accordance with Aristotles thesis, we are not to think of R as a set whose
members are particular real numbers. Rather, we are to conceive of R as Euclid
didnamely, as a nest of lines, that is, items with length but no breadth. The axi-
oms of the theory are then as follows.
Axiom R1 says that the structure R has specified members 0 and 1, a unary
map , and binary maps + and . that make it into a non-trivial field. That is, for
any x, y, and z in R,
0+ x = x x + ( x ) = 0 x+ y= y+x 1. x = x x. y = y.x
( x + y ) + z = x + ( y + z ) ( x. y ).z = x.( y.z )
x.( y + z ) = (x. y ) + (x.z )
(0 = 1)
(x = 0) y (x. y = 1)
Axiom R2 says that there is a relation < on R under which it is an ordered field
whose positive elements have square roots. That is, for any x, y, and z in R,
For any function f whose domain is and whose range is R, there is a unique
a R and a unique b R such that for all , f ()=a + b..
This Principle entails the Principle of Microstraightness at the origin, and also says
that the non-degenerate straight line that Microstraightness postulates is unique,
with initial height a and slope b. The second axiom is the Constancy Principle:
For any function f whose domain and range are both R, if (for all x R and all
, f (x + )=f (x)) then (for all x, y R, f (x)=f (y)).
Challenges from the Infinite and from the Infinitesimal 213
Given natural definitions, the Constancy Principle entails the familiar result that
any function whose first derivative is always zero is constant.
2 2 2 2
Since multiplication is associative, (.a) = .a , so that (.a) =0 whenever
2
=0. This means that a whenever . For no can we have 0 < , for
2 2
then 0 < , contradicting =0; so, by R2, < 1 for all . Thus a < 1 for all
and every real number a. Now any real number distinct from zero has a recipro-
cal, so for any non-zero real number a, and any infinitesimal , < 1/a. Given
any regular positive real number r, no matter how small, we can always find a
non-zero real number a for which 1/a < r. So any infinitesimal is smaller than any
regular real number.
What, though, is the underlying logic of SIA? As Lawvere and Kock were well
aware, the logic cannot be classical. To see this, observe first that we can derive
a contradiction in SIA from the supposition that , =0. For consider the
function g()=. If 0 were the only member of , then g()=g(0) and
g()=, so that (g()=g(0) + b.), for both b=0 and b=1, contradicting
the uniqueness of b required by Microaffineness (R1 says that 0 and 1 are distinct).
In both classical and intuitionistic logic, then, we can deduce from the axioms of
SIA that the set of infinitesimals is non-degenerate, i.e., =0. We can
also deduce from those axioms, though, that there is no infinitesimal distinct from
zero. For suppose that there were one, i.e., suppose (=0). Then we should
2
have both =0, since , and (=0), ex hypothesi. Given (=0), we can
2
apply R1 to deduce that there is a y such that .y=1. But then 0=0.y= .y=.(.y)=
.1=, so that the hypothesis (=0) yields a contradiction. In both classical and
intuitionistic logic, then, we can deduce from the axioms of SIA that there is no
infinitesimal distinct from zero, i.e., (=0). In classical logic, however,
the conclusions of our two deductions are contradictory: =0 classically
entails (=0). So SIA is a classically inconsistent theory.
For this reason, the founders of SIA formalized it as an intuitionistic theory. In intui-
tionistic logic, we cannot infer from =0 to (=0). (=0)
is intuitionsitically equivalent to (=0), but again, this does not contradict
=0, for we cannot infer from (=0) to =0. Indeed, SIA
is an intuitionistically consistent theory:Moerdijk and Reyes (1991) construct a topos
model for it (see the Appendix to Bell 2008 for a sketch of their construction).
While of mathematical interest, the construction of a model does not itself vindi-
cate the coherence of SIA. The model shows that there is some interpretation under
which the theory comes out true. That interpretation, though, may not be the intended
one. We need to check that, under the intended interpretation of the theory, the axi-
oms of SIA come out true when the connectives are read intuitionistically. Otherwise
the consistency of the theory will have been bought at the expense of its axioms truth.
214 Five Attacks on Classical Logic
Making sense of the intended interpretation of SIA is not a trivial task. From
a formal point of view, some of its theorems are reminiscent of well-known
theorems of intuitionistic analysis. Thus the argument to show that =0
extends to refute the universal decidability of identity. In other words, among
the theorems of SIAis
x (x = 0 x = 0).
Similarly, SIA refutes the universal Law of Trichotomy; among its theoremsis
x y (x < y x = y y < x ).
However, the surface similarity between these results and the corresponding the-
orems of intuitionistic analysis disguises deep differences in meaning. Consider
the formula x (x=0 x=0). For the intuitionist, this says that there is an
effective method for proving, of any presented choice sequence of rationals, that
it either converges to 0 or that it is refutable that it converges to 0.Accordingly,
x (x=0 x = 0) says that the hypothesis that there is such an effective
method reduces to absurdity. No interpretation along these lines, however, may
be invoked to make sense of the equiform theorem in SIA. The intuitionist thinks
of a real number as being constructed, or given, by a converging sequence of
rationals. As we have seen, however, it is a theorem of SIA that (=0).
Given a construction, or identification, of an infinitesimal distinct from zero, we
could certainly infer (=0). Accordingly, SIA precludes any sort of con-
struction of any non-zero infinitesimal (cf. Hellman 2006).
The intended interpretation of SIA needs to accommodate indeterminacy in
respect of order relations between infinitesimals. We saw above that any infinites-
imal is smaller than any regular positive real number. However, we can reduce to
absurdity the supposition that one nilsquare is less than another, and the supposi-
tion that there are distinct nilsquares. That is, we can prove ( < ) and (=)
for arbitrary , in . As Geoffrey Hellman puts it, we have the curious situation
that non-identity within is determinately false whereas identity remains indeter-
minate, as reflected in x (x=0 x=0) (2006, 636). As he points out, it is hard
to make sense of this without taking (x=y) to mean something other than is
not identical with. He goeson:
Bell in his presentation [of SIA] seems to suggest this, calling two points a, b ... distin-
guishable when they are not identical (), and calling them indistinguishable in case
(a b), so that the latter does not imply that they are identical, a=b, in conformity with
giving up the law of double negation ... This is an attractive reading, but, unfortunately,
it is problematic. For, since the combination is not a new primitive but just abbreviates
Challenges from the Infinite and from the Infinitesimal 215
(r < 0) (r = 0) (0 < r ).
( <0) ( = 0) (0 < ),
2
the first and the third disjuncts would entail > 0, contradicting , so that
disjunctive syllogism (which is intuitionistically valid) would yield =0, showing
that ={0} after all. In other words, infinitesimals do not stand in a determinate
order relation to zero:we cannot say that an infinitesimal is either less than zero,
identical with zero, or greater than zero.
Just as our semantic theory for a purely propositional language allowed
statements to be neither true nor false at some states of affairs, so our semantic
theory for the language of SIA must allow some predicates to be neither true
of, nor false of, some objects. However, our semantic clauses for connectives
linking closed sentences may be extended to characterize connectives that link
predicates or open sentences. Thus we can say that a sequence of objects satis-
fies a conjunction if and only if it satisfies both conjuncts, and that a sequence
satisfies a disjunction if and only if it belongs to the closure of the union of the
satisfiers of the disjuncts. Again, the definition of closure is the natural exten-
sion of the propositional notion:a sequence s belong to the closure of a set of
sequences U if and only if s satisfies every possible predicate that is satisfied by
all the members of U.
216 Five Attacks on Classical Logic
What of negation? We can say that a sequence s satisfies A just in case it
is incompatible with any sequence that satisfies A; but what does incompat-
ible mean when applied to sequences? In the special case of a mathematical
language, the natural answer is this. Let us say that a real number x is distinct
from a real number y if and only if either x < y or y < x. One of the axioms of
SIA is (x=y) (x < y y < x), and the converse conditional follows from
the law of identity and the axiom (x < x); thus the proposed account of dis-
tinctness coheres with the use of distinctness within the theory. Let us then
say that a sequence x is incompatible with a sequence y if, for some natural
number i, the ith member of x is distinct from the ith member of y. In the case
of a one-place predicate A, this reduces to the principle that A is true of a
real number x if and only if, for any real number y of which A is true, either
x < y or y < x.
It is easy to verify that all the axioms of SIA are true when the connectives are
interpreted in this way. Under that interpretation, the logic is intuitionistic, and
we can see why it should be. Consider the predicate (=0). This will be true of a
2
real number x only if (x < 0) (0 < x), i.e., only if x > 0.Now, as proved above, for
any such x and any infinitesimal , < x, so for any , satisfies (=0).
However, satisfies =0 only if =0, and we have already seen that the infer-
ence from is a nilsquare infinitesimal to is identical with zero does not go
through.
What is more, we now have an insight into why DNE fails. The root of the
matter is that the predicate is identical with zero has no back:that is, there is
no property that is ruled out by attributing this predicate to an object. For what
could that property be? Afirst thought may be that the property is that of being
distinct from zero, i.e., being either less than or greater than zero. But that
does not work:a number is distinct from zero if and only if its square is greater
than zero, so in excluding this property we leave it open that the number is a
nilsquare infinitesimal, contrary to the intended meaning of is identical with
zero. Asecond thought is that the property being excluded is the disjunctive
one of either being distinct from zero, or being a nilsquare. But that will not
work either:the square of zero is zero, so in excluding this second property we
exclude zero itself. Athird thought is that the property being excluded is the
disjunctive one of being either distinct from zero or a nilsquare infinitesimal.
The problem here is that SIA does not allow us to distinguish nilsquare infini-
tesimals from zero: infinitesimals are so close to zero as to be theoretically
indiscriminable from it. There is no saying what an attribution of identity with
zero excludes, so the property of being identical with zero is backless. That is
why classical logic cannot be applied within SIA.
Challenges from the Infinite and from the Infinitesimal 217
for saying that the rival parties mean precisely the same by these logical words.
For we also have to take into account how the divergence over whether a state-
ments truth-grounds must respect (R) alone, or (R) together with (B), reflects
deep differences in the general conception of a statements content. Aspace of
possibilities in which (B) holds is differently structured from a space in which
it does not hold. Since (for both parties) a statements logically relevant con-
tent is given by the possibilities at which it is true, that difference in struc-
ture affects the senses of all the statements in the pertinent language. So, even
though our rivals accept the same postulates as specifying the senses of the
connectives, we are not entitled to conclude that they think or express the very
same thoughts.
Far more importantly, our analysis has shown how a debate between the
classicist and intuitionist can avoid futile metalogical appeals to laws that
the opposing side contests. The focus of dispute has become whether the
truth-grounds of a well-formed statement must conform to (B), and in the
debate about that the metalogical strength of the sign for negation is not
crucial. As we saw, the classical logician cannot validate his preferred rules
by abandoning (B), retreating to (R) and appealing to a classical metalogic.
Equally, the intuitionist cannot afford to concede the truth of (B), even if he
insists on using only intuitionist rules in the metalogic. The parties have to
decide what they think about (B).
The claim that (B) should be a crux strikes me as highly plausible. It is,
Ithink, our natural assumption that each statement has a back that accounts
for the great reluctance we feel in deviating from classical logic, even after we
have seen that the Principle of Bivalence is doubtful.
Natural as that assumption is, though, it may be seriously challenged. In
particular, as our analysis of SIA brings out, (B) may fail when our statements
concern objects, such as nilsquare infinitesimals, whose identities are inher-
ently indeterminate. Whether we can really make sense of indeterminate iden-
tity is, of course, a much contested metaphysical question. Adjudicating the
16
issue will clearly not be easy; equally clearly, though, a rational discussion is not
at all precluded. To the contrary:we now have non-question-begging connections
between the choice of logic and another foundational philosophical question.
This brings out a final, and important, methodological difference with
Dummett. On his view, the proof-theoretic harmony that the intuitionist con-
nectives exhibit is a virtue that can be appreciated in advance of any semantic
16
Pace Gareth Evans (1978), it cannot be settled in a single page; see Parsons 2000.
Challenges from the Infinite and from the Infinitesimal 219
theory. The need not to disturb that harmony when we come to do semantics
yields a logical constraint on admissible semantic theories, and thereby on meta-
physical views. Irecognize no such constraint. We need semantic principles for
the connectives which account for their use in perceptual and probabilistic state-
ments, and in inductive and abductive reasoning, as well as their strictly deduc-
tive employment. In the context of a particular dispute about logical laws, we
may then investigate which laws those principles stably validate and invalidate.
In the present case, this led us to focus on the nature of possible states of affairs,
and on a disputed postulate concerning the structure of the space of them. There
is surely no metaphysical basis for logic, but equally there is no logical basis for
metaphysics, if that implies that we can settle the choice of logic in advance of
settling any seriously contested metaphysical-cum-semantic issues. What we are
seeking in the end is the optimalor an optimallogico-metaphysical package.
Metaphysical considerations cannot be extruded from rational decisions between
rival logical systems.
8
The Challenge from Vagueness
1
See, though, 8.6, where Ireturn to the structural version of the Sorites Paradox presented
in 2.3. Some philosophers, notably Crispin Wright (1987, 1992)and Delia Graff Fara (2003),
have constructed versions of the Sorites in which absurd conclusions are classically derived
from what they take to be compelling principles involving a definitely operator. These versions
of the Sorites are interesting, but discussing them would take me too far from the themes of
this book.
The Challenge from Vagueness 221
the correct use of the predicate red. Rendering this claim semi-formally, then,
it seems that we may affirm:
(2) a1 is red
and
Now if we suppose
which directly contradicts (1). Given (1) and (3), then, supposition (4) stands
refuted, so by Simple Reductio we may assert
which contradicts (2). This, then, is the initial paradox. We have strong reason to
accept the trio of postulates (1), (2), and (3), but we also have an apparently valid
deduction showing that the trio is inconsistent. It may be noted that the form of
Reductio that is applied in reaching line (7)and that is re-applied at the corre-
sponding later stepsis acceptable to an intuitionist. So the trio comprising (1),
(2), and (3)is inconsistent in both intuitionistic and classical logic.
222 Five Attacks on Classical Logic
This, however, seems to land us in a yet more acute paradox, one that Crispin
Wright has called the Paradox of Sharp Boundaries. For formula (10) says that at
some point in the sequence a red tube is immediately followed by a non-red tube,
and this seems to ascribe a sharp boundary to the concept red. Some philosophers
take the vagueness of a predicate to consist in there not being a sharp bound-
ary between the objects that satisfy it and the objects that do not. But even if the
absence of a boundary is not constitutive of vagueness, it seems absurd to say that
one tube is red while a tube indiscriminable in colour from it is not red. Yet we
have deduced this apparent absurdity by applying the rules of classical logic to the
indisputably true premisses (2)and (3).
The analysis of this argument will occupy much of the present chapter, but
one initial comment is in order. Although Ihave used an existential quanti-
fier in formulating premiss (1)and some subsequent lines of the argument, its
employment is entirely dispensable. Since the quantifier ranges over the natu-
ral numbers from 1 to 100, we may replace any statement of the form n(n)
by the corresponding disjunction (1) ... (100). Thus a propositional
logic suffices to assess the validity of the argument. Indeed, the only connec-
tives explicit or implicit in it are , , and , so the only logic needed is a
logic of negation, conjunction, and disjunctionprecisely the logic Iam treat-
ing in this book.
The apparent antinomy here has long engendered scepticism about the uni-
versal applicability of classical logic, but it is important to distinguish between
two sorts of sceptic. Some hold that the laws of some other logical system cod-
ify the standards for assessing the validity of deductions involving vague terms,
while others take a more radical position and deny that such assessments admit
of systematic codification at all. Among the radicals was Frege. On his view, state-
ments can stand in law-governed logical relations only if each of their component
expressions possesses a reference or Bedeutung. The reference of a predicate is a
concept or Begriff. Andyet:
The Challenge from Vagueness 223
A concept that is not sharply defined is wrongly termed a concept. Such quasi-conceptual
constructions cannot be recognized as concepts by logic; it is impossible to lay down
precise laws for them. The law of excluded middle is really just another form of the
requirement that the concept should have a precise boundary. (Frege 1902, 56, p.69)
A few years earlier, in 1896, he had elaborated the same message in a letter
toPeano:
Logic can only recognize sharply delimited concepts. Only under this presupposition
can it lay down precise laws ... Just as it would be impossible for geometry to lay down
precise laws if it tried to recognize threads as lines and knots in threads as points, so
logic must demand sharp boundaries of what it will recognize as a concept unless it
wants to renounce all precision and certainty. (Frege 1976, 1823=Frege 1980, 11415)
Indeed, Frege diagnoses the Sorites as arising from the application of precise laws
to quasi-conceptual constructions:The fallacy known as the Sorites [Acervus]
rests on this:that something (for example, heap) is treated as a concept when logic
cannot acknowledge it as such because it is not properly circumscribed (Frege
18978, 168=Frege 1979, 155).
have a simple statement in which a vague predicate is applied to one of its border-
line cases, as it might be Tube a50 is red,
the breakdown in convergence of verdicts leaves us without uncontroverted evidence for
its truth or for the truth of its negation. Since we lack convincing theoretical grounds to
think that one or the other must be true nonethelessbecause such grounds would have
to regard the determinants of truth and falsity as constituted elsewhere than in our lin-
guistic practicewe are left with no compelling reason to regard either [the statement] or
its negation as true. (Wright 2007, 441)
Thus we should refrain both from asserting and from denying Tube a50 is red.
Similarly, Wright thinks, we should refrain both from asserting and from deny-
ing Some red tube in the sequence is immediately followed by a non-red tube
that is, we should refrain both from asserting and from denying formula (10). In
particular, then, we should refrain from denying (10); equivalently, we should
refrain from asserting (1).
What, though, of the second part of the paradoxical argument? As remarked,
(1), (2), and (3)form an (intuitionistically) inconsistent trio, and premisses (2)and
(3)are clearly true. So it seems that (1)is false. But if (1)is false, then its negation
is true. So we are entitled to assertindeed, we are committed to asserting(9):
Wright accepts (9). Because (1)the denial that red has a sharp cut-offcontra-
dicts the evident truths (2)and (3), we must be prepared to deny any denial that
red has a sharp cut-off. Assuming that denying a statement is equivalent to affirm-
ing its negation, it follows that we must be prepared to affirm the doubly negated
formula (9). However, in order to reach (10), the conclusion Wright calls the
Unpalatable Existential, which asserts that there is a sharp cut-off between the
red and the non-red tubes, we would need to eliminate a double negation and this,
Wright argues, we are not entitled to do. Since we cannot assume that statements
involving vague expressions are either true or false, we should ... abstain from
unrestricted use of the law of double negation elimination (Wright 2007, 441).
Such abstention enables us consistently to deny (1), assert (9), but refrain from
asserting or denying (10). This combination of assertions, denials, and absten-
tions from assertion and denial constitutes, Wright proposes, a coherent response
2
both to the initial Sorites Paradox and to the Paradox of Sharp Boundaries.
2
Particularly as presented in Wright 2007, this response to the Sorites may be read as developing
the sketch in Putnam 1985, a reply to Read and Wright 1985, which was itself a critique of Putnam
1983 (esp. pp.2846).
The Challenge from Vagueness 225
We have here a novel and interesting argument for intuitionistic logic. Pace
Frege, there is a logic for assessing deductions involving vague predicates, but the
logic in question is intuitionistic rather than classical. If it works at all, Wrights
argument has very wide application. Many ordinary predicates are such that
nothing in our ordinary practice with them grounds the claim that their exten-
sions have sharp cut-offs. For many of those predicates, indeed, it will be pos-
sible to construct a Sorites sequence. To those predicates, Wrights argument will
apply:on pain of contradicting clear truths, we must deny that there is no sharp
cut-off, whilst resisting the unsupported assertion that there is one. This position
is coherent only if there are restrictions on the elimination of double negations;
so we have an argument in favour of a logic (such as intuitionistic logic) which
imposes such restrictions.
What should we make of this argument? In his reply to Wrights paper Wangs
Paradox, Michael Dummett put on record his admiration for the beauti-
ful solution of the Sorites advocated by Crispin Wright, clouded by a persistent
doubt whether it is correct (Dummett 2007a, 453). Anyone who appreciates
logico-philosophical finesse will share Dummetts admiration for an ingenious
proposal; but there are, indeed, reasons to doubt its correctness.
First, Wright moves too quickly from the observation that we hesitate when
pressed to say whether a borderline red-orange object is red, to what is really
needed to sustain the distinction that the intuitionist wishes to draw between
(9)and (10). The observation is surely correct, and if one further holds that the
only possible grounds for the truth of a vague statement must be overt in our
ordinary practices of asserting and denying it, then it cannot be assumed that
an atomic statement in which a vague predicate is applied to one of its borderline
cases is either true or false. This point, however, is not immediately relevant to the
status of statements (9)and (10). These statements are not atomic, and (on intui-
tionistic principles) we may assert that a complex statement is bivalent even when
3
we cannot assert the bivalence of any of its components. So we need a further
argument to show that we cannot assert the bivalence of (9)and (10).
Wright seems to have in mind a Heyting-style semantics according to which
truth goes with assertibility and bivalence goes with decidability. On that
semantics, the existentially quantified formula (10) may be assumed to be biva-
lent if (a)the relevant domain of quantification is surveyable, and (b)each matrix
instance is decidable. Condition (a)is met; indeed, as noted, the deduction would
go through in essentially the same way if (10) were replaced by a disjunction. So
3
For example, an intuitionist is always entitled to assert the falsity, and hence the bivalence, of
(A A), whatever the status of A.
226 Five Attacks on Classical Logic
Wright must deny that all the matrix instances of (10) are decidable; in particu-
lar, he must deny that we have a general ground for asserting the negation of an
arbitrary formula in the form an is red (an+1 is red). For, in a Heyting-style
semantics, such a ground would render (10) false, so we would have a proof of
(1), as well as of (9), and Wrights way out of paradox would collapse back into
contradiction.
The problem here is that there is a plausible general ground for asserting the
negation of any such conjunctionnamely, Timothy Williamsons thesis that
knowledge of inexact matters demands a margin for error (Williamson 1994, 226
30). According to that thesis, knowledge that an is red requires that an+1 may also
be red, for all we know; this excludes knowledge that an+1 is not red. Since Wright
thinks that we are entitled to assert only what we know, Williamsons thesis pro-
vides a general ground for rejecting any assertion of the form an is red (an+1 is
red):any circumstance in which we are entitled to assert the first conjunct is a
circumstance in which we are not entitled to assert the second conjunct. Under
a realist semantic theory, this is merely an epistemic point:there are no possible
circumstances in which we would be entitled to assert an is red (an+1 is red),
but we have the option of simply refraining from either asserting or denying that
conjunction. Under Heytings semantics, however, matters are more problemati-
cal. For Heyting, a demonstration that we shall never be entitled to assert A is a
basis for asserting A . Under Heytings semantics, then, Williamsons margin
for error principle yields a basis for asserting (an is red (an+1 is red)), for arbi-
trary n. The totality of such assertions is clearly incompatible with the hypothesis
n(an is red (an+1 is red)). Accordingly, the margin for error principle affords
a ground for asserting n(an is red (an+1 is red)), i.e., (1). Wright, however,
wishes to assert n(an is red (an+1 is red)), i.e., (9), which is the negation of
(1). Given the margin for error principle, then, Wright is committed to asserting
a formula and its negation. Hence Wright badly needs to dispel the considerable
attractions of the thesis that inexact knowledge demands margins for error, for
4
that thesis brings his position into ruin.
The objection just laid against Wright assumes that he understands the con-
nectives in conformity with Heytings semantics for the language of intuitionistic
logic; it is not an objection to the use of that logic per se. Inasmuch as any definite
semantic conception animates Wrights discussion, it is Heytings, or something
very close to it. As Dummett remarked, however, it is most implausible to suppose
4
This thesis about knowledge does not entail Williamsons own epistemicism about vague-
nessthe claim that there is a sharp cut-off in the sequence of red things but we cannot in principle
know where it is. Someone who denies, or refrains from accepting, that there is a sharp cut-off may
still hold that knowledge that such-and-such an object is red requires a margin for error.
The Challenge from Vagueness 227
that the Heyting semantics correctly describes the contribution that the connec-
tives make to the meaning of vague statements:
Wright can hardly say that to grasp the meaning of the statement [This curtain is red,
said of a borderline red-orange curtain] is to be able to recognize a demonstration that
the colour of the curtain is red, not orange, or that it is orange, not red. Presumably he
must say that it consists in an ability to judge that the colour of the curtain is neither defi-
nitely red nor definitely orange, but is definitely either one or the other. If so, however, the
theory of meaning for vague statements on which Wright wishes to rest intuitionist logic
as applying to them interprets disjunction quite differently from the intuitionist theory of
meaning for mathematical statements. (Dummett 2007a, 4512)
We cannot assume that the set of actual men who are tall is open in the relevant
topology. Suppose that every man of height greater than 510 dies, except for one
man, Adam, whose height is 61. Adam is tall, but in the circumstance envisaged
there is no neighbourhood of tall men around him. All the same, the view of logic
recommended in this book demands that we take account of things that might
exist when characterizing the semantic values of predicates. I have defended
Aristotles Thesis:when a conclusion follows logically from some premisses, it is
logically necessary that it is true whenever they are all true. For this reason, we
took the semantic value of a complete statement to be the set of all logical possibili-
ties at which it is true. Although we are now concerned with the semantic values of
predicates rather than the values of complete statements, logical relations among
predicates still involve logical necessitation, so it is natural to take the semantic
value of a predicate to be a subset of a domain of all logically possible objects, not
merely the actual ones. Having made this point, however, Ishall bracket for pre-
sent purposes the vexed question of how we might best conceive of that domain.
Why might these ruminations about open sets give aid and comfort to a
theorist who recommends using intuitionistic logic when assessing deductive
arguments involving vague terms? Where S is any topological space, and L is
any propositional language whose connectives are , , , and , let us say
that a map v is an interpretation of L into S if v maps each well-formed formula
of L into an open set in S in such a way that
v( A B) = v( A) v(B)
v( A B) = v( A) v(B)
v( A B) = Int (v(B) (S (v( A)))
and
v(A) = Int (S (v( A)).
(Thus v(A) = v(A ), where v() = .) The interior Int (X) of a set X
is the smallest open set contained in X. Accordingly, the clause for negation
just displayed ensures that a negated statement is mapped into an open set in
S. Let us then define a consequence relationTop as follows:A 1,...,An Top B if
and only if, for every topological space S, and every interpretation v of L into
S, v(A 1) ... v(A n) S v(B). Atheorem due to Tarski then tells us that the
intuitionistic propositional calculus is sound and complete with respect to this
notion of consequence. That is, where Int signifies deducibility by the rules of
5
the intuitionistic propositional calculus, A1,...,AnInt B if and only if A1,...,AnTop B.
5
This result follows from the Second Principal Theorem of Tarski 1938 (see Tarski 1983, 448).
230 Five Attacks on Classical Logic
vague predicate and a two-place relation whose relata may be expected not to dif-
fer in respect of whether the predicate applies. Thus, if x and y differ in height by
only a fraction of a millimetre, we shall expect x to be tall if and only if y is tall, and
we shall expect x to be not tall if and only if y is not tall. Again, if the light reflected
by object x differs in wavelength from that reflected by object y by only a fraction
of a nanometre, then we shall expect x to be red if and only if y is red, and we shall
expect x to be not red if and only if y is not red.
This suggests the following definitions. Let us say that a two-place relation R is a
tolerance relation on a domain U when R is a reflexive and symmetric relation on U.
Atolerance relation is not required to be transitive. Thus the relation between x and
y when they differ in height by less than m millimetres is a tolerance relation; so is
the relation between x and y when x reflects light of a wavelength which differs by
less than n nanometres from that reflected by y. Where x is a member of U and R is
a tolerance relation, the set [x]R comprises all and only those members of U that are
related by R to x. The collection of elementary sets in U (with respect to R) consists of
the sets [x]R for all x belonging to U. Because R is reflexive, the entire domain is cov-
ered by elementary sets, but because R need not be transitive it may not be possible to
partition the domain into elementary sets:a collection of elementary sets that covers
the domain may include sets that overlap each other. Finally, a subset X of U is said
6
to be definable (with respect to R) if it is either empty or a union of elementary sets.
We then introduce the notion of the interior of Xwritten Int (X)as the larg-
est definable set that is wholly contained in X. We immediately have that the Int
operation is
6
Itake these definitions from the theory of rough sets, originated by Zdzisaw Pawlak (1982).
Pawlaks theory was applied to the analysis of vague discourse by Ewa Orowska (1985), who
improved it by not requiring the tolerance relation to be transitive. Pawlak and Orowskas use of
the rough-set framework, though, is very different from that contemplated here; see n. 7.
232 Five Attacks on Classical Logic
Finally, suppose that x Int (X Y). Then x belongs to some elementary set, each of
whose members belongs to X Y. Thus, x belongs to some elementary set, each of
whose members belongs to X, and each of whose members belongs to Y, so that x
Int (X) and x Int (Y). Hence Int (X Y) Int (X) Int (Y). Conversely, sup-
pose that x Int (X) Int (Y). Then x belongs to some elementary set A, each of
whose members belong to X, and to some elementary set B, each of whose mem-
bers belongs to Y. Now the intersection of any two elementary sets is also elemen-
tary. So x belongs to an elementary setnamely, A Beach of whose members
belongs to X Y. That is to say, x Int (X Y). Putting these results together, we
conclude that Int
We have shown, then, that Int meets all of the conditions for being an interior
operation in the full, topologists sense. Let us call a set open if it is identical with
its own interior:X is open if and only if Int (X)=X. It follows that the sets that are
open in this sense form a topological space that Ishall call T. The open sets of T are
precisely the definable subsets of U.
We can now state more precisely the conjecture which we have been trying to
articulate. When a vague predicate is used in a Sorites deduction, a tolerance rela-
tion is operative in the context of the argument. Thus, in the Sorites deduction of
8.1, the contextually operative tolerance relation is that of indiscriminability in
respect of colour. As we have just seen, that tolerance relation generates a topo-
logical space. The conjecture to consider is that the extensions of vague predi-
catesincluding complex predicatesare open sets in this topological space, i.e.,
are definable sets in the specified sense.
If this conjecture were correct, would it support Wrights thesis that intuition-
ist logic is the logic of vagueness? Ithink it would go some distance towards doing
so. Given the conjecture, it is natural to associate the predicate operators and,
or, and not with the operations of intersection, union, and interior of the com-
plement on open sets. For if A and B are open sets in T, then so are A B, A B,
and Int (U A). By Tarskis soundness theorem, it would follow that the rules of
the intuitionistic calculus of monadic predicates are sound when applied to pred-
icates formed using these operators. There is, moreover, some reason to think that
the use of a stronger system of rules would generate formally invalid arguments
involving vague monadic predicates. Given any topological space S, we may
define a relation R that obtains between x and y when x and y belong to the same
open sets in S. R is a tolerance relation and every open set in S is definable with
respect to it. By Tarskis completeness theorem, then, whenever a monadic predi-
cate B in the language L is not intuitionistically deducible from some predicates
The Challenge from Vagueness 233
A1,...,An, there will be a tolerance relation R such that the semantic values of
A1,...,An and B are all definable with respect to R, but some possible objects satisfy
all of A1,...,An without satisfying B. On the assumption, then, that being definable
with respect to an appropriate tolerance relation is the strongest general require-
ment on the extension of a vague predicate, we could conclude that intuitionistic
logic is the strongest logic that may be justified on the basis of general semantic
features of vague predicates.
One attractive feature of the theory is that it explains why classical logic is the
right logic to use in the limiting case where total precision is attained. In this
limit, the tolerance relation will be identity, elementary sets will be singletons,
and the topology will be the discrete topology in which every subset of U is open.
Int (U A) will then be identical with U A, and Excluded Middle will be validated.
Despite these nice features, the proposed vindication of intuitionistic logic
does not, Ithink, succeed. The vindication assumes that the strongest general
requirement on a vague predicate is that its extension is definable with respect to
a tolerance relation. The next section will cast doubt on this assumption:there is a
central class of vague predicates each of whose extensions have stronger topologi-
cal properties, properties that sustain a stronger than intuitionistic logic. There
is, however, a more basic problem with the present attempt to vindicate intuition-
ism. The vindication rests on the conjecture. That is, it rests on the premiss that
the extension of a vague predicate will be open with respect to the appropriate
tolerance relation. That premiss, Inow argue, is highly doubtful.
To see the problem here, it helps to return to the Sorites argument presented
in 8.1. If the conjecture is to vindicate the use of intuitionistic logic in assessing
a Sorites argument, the extensions of the vague predicates involved in the argu-
ment must all be definable with respect to the tolerance relation implicated in
the argument. In the case of the Sorites deduction presented earlier, the vague
predicate was red and the tolerance relation was indiscriminability in colour.
Accordingly, the conjecture requires that the extension of red be definable with
respect to indiscriminabilitythat is, that the extension of red should be a
union of sets that are elementary with respect to (colour) indiscriminability. An
elementary set with respect to colour indiscriminability is a colour shade. The
conjecture requires, then, that the extension of red should be a union of shades.
Should we accept this thesis about the extension of red? Ithink not. Not
only is there no good reason to postulate that the extension of red is defin-
able, there is strong reason to hold that it is not. Let X be the extension of
red. The question is then whether X = Int (X), where Int (X) is the largest
definable set contained in X. Since Int is DECREASING, the crux is whether
X Int (X). Now the closure Cl (X) of X is the complement of the interior of the
234 Five Attacks on Classical Logic
7
In the works cited in n.6, Pawlak and Orowska take the extension of red to be a rough set that
is defined by two approximations. Its lower approximation is its interiorviz., the largest defin-
able setall of whose members are red. Its upper approximation is its closurethe smallest set to
include every object that belongs to an elementary set that also contains a red object. The rough set
itself is taken to be an indeterminate tertium quid that lies between the two approximations but
which is otherwise left unspecified. The theory of rough sets is developed in a classical logic, and
there are various ways of vindicating classical logic for vague predicates on the basis of its semantic
treatment of those predicates. (For one way, see again Orowska 1985; for a survey of the theory, see
Polkowski 2002.) So Ineed not subject that treatment to critical scrutiny here. However, while the
rough set approach may be the best we can do to characterize some vague predicates, Ihope in the
next two sections to say rather more about the set of red objects than that it includes the set of things
that are clearly red and is included in the set of things that are arguably red.
8
Of any concept, we must require that it have a sharp boundary. That means:of any object it
must hold either that it falls under the concept or that it does not. We may not allow a third case in
which it is somehow undecided or indeterminate whether an object falls under a concept (Frege
1914, 248=Frege 1979, 229).
236 Five Attacks on Classical Logic
that are prime (viz., 2, 3, 5, 7,...) and those that are not (viz., 1, 4, 6, 8,...). For other
predicates, though, such as colour terms, the classical account is deeply implausi-
ble. Some objects satisfy red and others do not, but the mere existence of positive
and negative cases does not entail that there is a boundary between them, and
in the present case it seems pretty clear that there is not. Thus Mark Sainsbury
invites usto
consider a very familiar case:the colour spectrum, as displayed, for example, in an illus-
tration in a book on colour. Looking carefully, we can discern no boundaries between
the different colours:they stand out as clearly different, yet there are no sharp divisions.
There are bands, but no bounds. This does nothing to impede the classificatory pro-
cess:the spectrum is a paradigm of classification. (Sainsbury 1990, 258)
The spectrum enables us to attach senses to colour terms not because it shows
boundaries, but because it displays colour paradigms or poles. Sainsbury likens
colour paradigms to magnetic poles exerting various degrees of influence:some
objects cluster firmly to one pole, some to another, and some, though sensitive
to the forces, join no cluster (Sainsbury 1990, 258). Magnetic poles are dipoles,
which repel as well as attract, so Iprefer a simpler analogy which likens para-
digms to gravitational poles, that is, massive bodies. If a small body is sufficiently
close to a gravitational pole, it will be drawn towards it, rather as we are drawn to
classify as red those objects that are sufficiently close in colour to a paradigm, or
pole, of red.
As Sainsbury notes, this analogy permits a natural description of some strik-
ing empirical data. Subjects asked to classify a range of test objects using just
young and old make different assignments to these words from those they
make to them when asked to classify using, in addition, middle-aged. The intro-
duction of a third ... pole can attract some of the things only loosely attached
to two existing poles, without diminishing the forces the existing ones exert
(Sainsbury 1990, 259). Asimilar phenomenon is found when more refined colour
terms supplement the basic ones, as the detailed cross-linguistic investigations of
Berlin and Kay (1969) attest.
Sainsbury draws attention to an important feature of polar classification:
Boundaryless concepts tend to come in systems of contraries:opposed pairs like child/
adult, hot/cold, weak/strong, true/false, and the more complex systems exemplified by
our colour terms. This is a natural upshot of boundarylessness, as we can see by reflecting
on what is involved in grasping a concept.
Such a grasp, it must be agreed on all sides, involves knowing how something would
have to be for the concept to apply to it, and how something would have to be for the
concept not to apply. Adistinctive feature of the classical picture is that it takes this latter
fact as primitive. Grasping what a concept excludes is part of grasping the concept, and is
achieved through the mediation of no other non-logical concept...
The Challenge from Vagueness 237
On the alternative picture, what a concept excludes is graspable in a positive way, medi-
ated by other contrary concepts. Agrasp of red attains grasp of what is not red at a deriva-
tive level, via a grasp of yellow, green, blue, and so on. Asystem of such concepts is grasped
as a whole, as can be seen in the way paradigms are used in learning. There are paradigms
of red, but nothing is non-derivatively classifiable as a paradigm of not-red. Any para-
digm of another colour will serve as a paradigm of how not to be red, but only in virtue of
its positive classification as another colour. (Sainsbury 1990, 258)
Like the classical theorist, then, Sainsbury accepts the principle Eadem est sci-
entia oppositorum:knowing what it is for something to be red involves knowing
what it is for something not to be red. For Sainsbury, however, the latter knowl-
edge requires a grasp of some other colour concepts. Grasping the concept not red,
and hence grasping the concept red, requires the grasp of a basic system of colour
conceptsin the present case, the concepts we call the colours of the rainbow.
This account provides a more plausible description of the semantic func-
tioning of colour terms, and of such words as child and adult, than any
alternative Iknow. The view that there exists a sharp boundary between (for
example) red and non-red things has taken on a new lease of life in recent years
thanks to Timothy Williamsons ingenious attempt to explain how the loca-
tion of such a boundary, although precise, might be unknowable in principle
(see Williamson 1994). However, neither Williamson nor any of his followers
has provided a convincing account of what makes it the case that red is asso-
ciated with one sharp boundary rather than another, so his theory does little
to allay the taint of implausibility that attaches to classical semantics when
it is applied to vague predicates. Moreover, the most familiar alternatives to
classical semantics in this areasupervaluationist theoriesfare little better.
According to Kit Fine, vagueness is a semantic notion. Very roughly, vague-
9
ness is deficiency in meaning. David Lewis, another prominent supervalu-
ationist, spells out in what the deficiency consists:
If Fred is a borderline case of baldness, the sentence Fred is bald may have no deter-
minate truth value. Whether it is true depends on where you draw the line. Relative
to some perfectly reasonable ways of drawing a precise boundary between bald and
not-bald, the sentence is true. Relative to other delineations, no less reasonable, it is
false. Nothing in our use of language makes one of these delineations right and all the
others wrong. We cannot pick a delineation once and for all (not if we are interested
in natural language), but must consider the entire range of reasonable delineations.
(Lewis 1979, 244)
9
Fine 1975, 120. In this paper, Fine appeals to a supervaluationist semantics to vindicate the use
of classical logic in reasoning with vague concepts. There are, however, many ways of developing
the core supervaluationist idea, not all of which sustain classical logic. For a survey, see Varzi 2007.
238 Five Attacks on Classical Logic
But how does the meaning of red relate to the perceived colour pole, r?
Asemantic theory for a vague predicate, as for any predicate, must say under
what conditions the predicate is true of an object, that is, under what condi-
tions an object satisfies the predicate. In formulating such conditions, it is natu-
ral to invoke a notion of perceived closeness in colour, which Iunderstand to be a
three-place comparative similarity relation:x is perceived to be closer in colour to
y than it is to z. With this relation understood, it is natural to postulate that, when
we are classifying objects in relation to colour poles, an object will satisfy red
if and only if it is perceived to be closer in colour to the pole r than to any other
colour pole. More generally, where A is a simple colour predicate associated with a
colour pole q, and where sats abbreviate satisfies, wehave
Thus, when the relevant poles are our seven colours of the rainbow, an object sat-
isfies red if and only if it is perceived to be closer in colour to r than to any of o, y,
g, b, i, or v. We might think of (Sat) as offering a schematic specification of a sim-
ple colour terms sense, with its precise sense (on a given occasion of use) deter-
mined by spelling out the relevant space of colour poles. Certainly the extension
of a simple colour term depends on what the other poles are. An object might be
perceived to be closer in colour to r than to any of the poles o, y, g, b, i, or v, while
10
Exactly how to bring the perceiver into the story is a delicate question. It will not do to say that
an object is red if it looks red to a normally sighted viewer in daylight:that fails to account for the
truth of Betelgeuse is red. Perhaps one could say that an object is red if it looks red to a normally
sighted viewer in the optimal conditions for viewing the object in question. But we need not explore
these niceties for present purposes.
240 Five Attacks on Classical Logic
11
For the purported proofs, see again Evans 1978 and Wiggins 2001, chap.6. Heck (1998) and
Edgington (2002), among others, expose the flaws.
12
Sainsbury seems to have in mind a position of this kind:I believe that the way forward involves
taking the notion of a vague object as basic (Sainsbury 1990, 262).
13
There cannot be an empty class if we take a class to be a collection or totality of individuals, so
that, as the author [Schrder] says, the class consists of individuals, or individuals make up the class. In
the course of this discussion we have once more had it shown to us that this way of talking is logically
useless; that the extension of a concept is constituted in being, not by the individuals but by the concept
itself; i.e., by what is said of an object when it is brought under a concept. There is then no objection to
our talking about the class of objects that are bs even when there are no bs (Frege 1895, 451).
14
It is to be distinguished from the iterative conception of set that some take to be implicit in
Zermelo-Fraenkel set theory. Idiscuss the latter conception in the next chapter.
The Challenge from Vagueness 241
recognition that not all meaningful predicates have extensions (see Geach 1955),
and it allows vagueness in respect of membership. If x belongs to the set of As
means x falls under the concept A, that is, x is A, then vagueness in whether x
is A will engender corresponding vagueness in whether x belongs to the set of
As. But vagueness in respect of membership is consistent with determinacy of
the sets identity, so long as a sets identity is taken to consist, not in its having
such-and-such members, but in its being the extension of the concept A.
Principle (Sat) embodies what one might call a maximalist approach to colour
classification:we deem an object to be red whenever it is perceived to be closer
in colour to r than to any other pole. Some will wish to be stricter:they will only
wish to call an object red when it is much closer to polar red than to any other
pole. Ineed not resolve this issue here, for my analysis will go through if an object
must meet this stricter condition in order to satisfy red. Readers who think the
original condition too lax are invited to make the appropriate changes through-
out. Introducing much makes the revised condition for satisfying red even
more clearly vague than its predecessor.
Both the lax and the strict conditions for an object to satisfy red are akin to a
famous proposal of Quines about naturalkinds:
One may be tempted to picture a kind, suitable to a comparative similarity relation, as any
set which is qualitatively spherical in this sense:it takes in exactly the things that differ
less than so-and-so much from some central norm. If without serious loss of accuracy we
can assume that there are one or more actual things (paradigm cases) that nicely exem-
plify the desired norm, and one or more actual things (foils) that deviate just barely too
much to be counted into the desired kind at all, then our definition is easy:the kind with
paradigm a and foil b is the set of all things to which a is more similar than a is to b. (Quine
1969, 11920)
The last claim is correct, but it is no objection to the proposed account of redness
so long as people are able to make reliable judgements of comparative chromatic
similarity. The empirical evidence suggests forcibly that most of us are able to do
this (see especially Rosch 1975).
242 Five Attacks on Classical Logic
All the same, in the present dialectical context we cannot rest content with Quines
elucidationor better, elucidationsof redness. According to the first of these, a red
thing will be one that differs less than so-and-so much from the norm of red. This
presupposes, not only a comparative similarity relation, but also a measure (less than
so-and-so much) of perceived closeness in colour. Now if there is such a measure, and
it obeys the axioms of a Euclidean metric space (which include the triangle inequality),
and we classify each coloured object alongside the norm or paradigm that is closest
by the measure, then the result will be a partition of coloured objects (the so-called
Voronoi tessellation) into convex colour categories (see Grdenfors 2000, 8792). That
is to say, any object intermediate in colour between two other objects in a given category
will itself belong to that category. So Quines picture of natural kinds as qualitatively
spherical will on these assumptions be vindicated. But the assumptions are really too
strong to be plausible. Even if our judgements of colour implicitly reflect some kind of
measure of perceived closeness in colour, the measure they reflect will be vague, and
there is no reason to expect it to validate the axioms for a Euclidean metric space.
Quines second elucidation does not invoke a measure:an object x will exem-
plify the kind with paradigm a and foil b if and only if a is more similar to x
than it is to b. The problem with this elucidation is that it involves specifying a
foilwhich, in the present case, means specifying a particular object whose
colour deviates from a paradigm of red just barely too much to be counted as
red. Wittgenstein (1974, 240)described as nonsense the command Make me
the smallest heap you would still call a heap. Iam not sure that Wittgensteins
command is always nonsense. If the objects to be heaped are spheres of the same
radius, one could perhaps comply with it by arranging three contiguous spheres
in a triangle and placing a fourth in the indentation between them. However,
when uttered in a well-stocked art materials shop, the command Show me the
colour tube that is closest to red without being red surely is absurd. Even an epis-
temicist about vagueness will deny that we can specify a foil for a vague predicate,
and since a foil for red would mark the boundary of the non-red things, the the-
ory we are considering denies that colour terms have foils. Aproper development
of that theory must specify the senses of colour terms without postulating foils.
y Int ( X ) iff y X p ( M ( p, y ) p X ).
Here, the variable p ranges over the members of a given system of colour poles,
and M (p, y) means Pole p is maximally close (in colour) to the coloured object
y. So defined, the interior of X comprises those members of X to which only poles
lying inside X are maximally close. It is immediate from the definitions that Int is
DECREASING, ZEROED, TOTAL, and MONOTONE. It is, moreover, straight-
forward to prove that Int isalso
15
IDEMPOTENT Int Int (X)=Int (X) and
15
Proof:Because Int is DECREASING, it suffices to prove that Int (X) Int Int (X). Suppose,
then, that y Int (X). That is y X and p (M (p, y) p X). We require to prove that y Int Int (X),
that is, that y Int (X) and p (M (p, y) p Int (X)). The first conjunct follows immediately from
the hypothesis. For the second conjunct, suppose that M (p, y). By hypothesis, it follows that p X.
Further, since p is the only pole that is maximally close to p, q (M (q, p) q X), so that p Int (X).
This completes the proof that Int (X) Int Int (X).
244 Five Attacks on Classical Logic
16
DISTRIBUTES OVER INTERSECTION Int (X Y)=Int (X) Int (Y).
Under this definition too, then, Int is an interior operation in the full topologists
sense. Accordingly, if we define a set as open when it is identical with its own
interior, the collection of open sets forms a topological space, which Ishall call T.
Aset X is open in T if and only if any pole that is maximally close to a member of
X itself belongs to X.
The pole that is associated with a simple colour term A certainly satisfies A,
so (Sat) ensures that the extension of an atomic predicate is open in T. This
time, though, disjunctions and negations of colour predicates also have exten-
sions that are open in T. The poles that are maximally close to objects that
either red or orange are precisely r and o, and each of these poles is either red or
orange. As for negations, the poles that are maximally close to objects that are
not red are o, y, g, b, i, and v, and each of these poles is not red. It is plausible to
conjecture that the extension of any colour predicate will be an open set in the
topology T.
In fact, Ishall argue for a stronger claim, in stating which it helps to introduce
some further terminology. Let us call the interior of a sets complement its exte-
rior:thus the exterior of X, which Ishall write as X , is Int (U X). In T, an object
belongs to the exterior of X if it does not belong to X and no pole that is maximally
close to it belongs to X. Assuming a classical logic, a sets exterior is the comple-
ment of its closure. In the particular case of T,
y Cl ( X ) iff y X p ( p X M ( p, y )).
That is, the closure of X comprises all the members of X together with any objects
to which at least one pole in X is maximally close. Thus the closure of the exten-
sion of red includes borderline red-orange objects as well red ones.
According to our semantic theory, which colour predicates an object satisfies
depends on which poles are maximally close in colour to it. Hence, if A is a simple
or complex colour predicate and X is its extension |A|i.e., |A| = {x:x sats A}the
members of X will be those objects whose colour status is incompatible with
being A. Consequently, the members of X will be those objects whose colour
status is incompatible with any colour status that is incompatible with being A.
Since incompatibility is a symmetric relation, being A is a colour status that is
16
Proof:X Y X. Since Int is MONOTONE, this implies that Int (X Y) Int (X). Similarly,
Int (X Y) Int (Y), whence Int (X Y) Int (X) Int (Y). Conversely, suppose that
z Int (X) Int (Y). Then z Int (X) and z Int (Y). Hence (1)z X and p (M (p, z) p X) and (2)
z Y and p (M (p, z) p Y). It follows from (1) and (2) together that z X Y and
p (M (p, z) p X Y), i.e., z Int (X Y). Thus Int (X) Int (Y) Int (X Y), and we are done.
The Challenge from Vagueness 245
incompatible with any colour status that is incompatible with being A. So, for any
colour predicate, A, we have |A| |A|.
I claim that the converse inclusion holds too. Suppose that B is a colour
status that is incompatible with any colour status that is incompatible with
being A. On Sainsburys picture, a colour has the extension that it has by
virtue of its place in a network of contrary colour classifications. Take A to
be red. The contrary colour statuses are orange, yellow, green, blue, indigo,
and violet. If B is incompatible with any of these, it must a determination of
the colour red. Thus |A| |A|. The argument generalizes:where A is any
colour predicate, |A|=|A| .
In any topological space, for any set X, X =Int (U X )=Int (U Int (U X)).
In any topological space, U Int (Y)=Cl (U Y)). So X =Int Cl (U (U X)).
Assuming a classical logic, U (U X)=X, so X =Int Cl (X). Aset which
17
is identical with the interior of its closure is called regular open. The argu-
ment of the previous paragraph shows, then, that the extension of any colour
predicate is a regular open set in the topology T. This conclusion is reassur-
ing. From a geometric point of view, a regular open set is one with no cracks in
it. We should not expect to find cracks in the extension of a vague predicate.
The argument given two paragraphs back generalizes. The senses of many
vague predicates are understood by reference to a collection of poles that define
a system of contrary terms. (Recall Sainsburys examples of young/middle-aged/
old, hot/cold, weak/strong, true/false, etc.) In classifying objects by reference to
these poles we implicitly apply a comparative similarity relation. Thus someone
is young if they are closer in age to a paradigm of youth than to a paradigm of
middle or old age. Given such a relation, it will be possible to define an interior
operation in just the way we defined the operation Int on sets of coloured objects,
and the corresponding collection of open sets (i.e., the collection of sets that are
identical with their interiors) will form a topology. Because the extension of such
a predicate is determined by its relations to the contrary terms associated with
other poles, our argument may be re-applied to show that the extension is a regu-
lar open set in that topology. The argument only goes through, however, if the
relevant collection of poles is held fixed.
How do the extensions of complex polar predicates relate to those of their com-
ponents? An object satisfies a conjunctive predicate just when it satisfies both
conjuncts:thus, where and links two predicates, x sats A and B if and only if
17
Any regular open set is open. Suppose X=Int Cl (X). Then Int (X)=Int Int Cl (X)=Int Cl (X)
since Int is IDEMPOTENT. Hence Int (X)=X by hypothesis.
246 Five Attacks on Classical Logic
x sats A and x sats B. Where |A| is the extension of the predicate A, we can then
state the semantic axiom for conjunction as follows:
18
Proof:We need to show that X Y=Int Cl (X Y) given that X=Int Cl (X) and Y=Int Cl (Y).
It suffices to show (1)X Y Int Cl (X Y) and (2)Int Cl (X Y) X Y. For (1)we argue as fol-
lows:since X and Y are both open, X Y=Int(X) Int(Y)=Int (X Y) by DISTRIBUTION. Since Cl
is INCREASING, X Y Cl (X Y), whence Int (X Y) Int Cl (X Y) since Int is MONOTONE.
Thus X Y Int Cl (X Y) as required. For (2), X Y X whence Int Cl (X Y) Int Cl (X), since Int
and Cl are MONOTONE. By hypothesis, Int Cl (X)=X, whence Int Cl (X Y) X. Asimilar argu-
ment shows that Int Cl (X Y) Y. Combining these results yields Int Cl (X Y) X Y.
19
Proof:As a first lemma we show that Y Y whenever Y is open. Since Cl is INCREASING,
Y Cl (Y), whence Int (Y) Int Cl (Y) since Int is MONOTONE. When Y is open, Int (Y)=Y, so
Y Int Cl (Y)=Y. Since X =Int (U X), X is open for any set X, so the first lemma gives X X .
As a second lemma, we prove that W Z whenever Z W:suppose Z W; then U W U Z.
Since Int is MONOTONE, this yields Int (U W) Int (U Z), i.e., W Z. By the first lemma,
when X is open X X, whence X X by the second lemma. So, when X is open, X =X.
The Challenge from Vagueness 247
is the logically strongest predicate whose extension includes |A| |B|. Given
that any predicate has a regular open extension, this means that the extension
of A or B must be the smallest regular open set that contains |A| |B|. When
20
X is open, Int Cl (X) is the smallest regular open set that contains X. When
two sets are regular open, their union is certainly open. In this way, we are led
to the following axiom for disjunction:
The results just cited mean that |A or B| is regular open whenever |A| and |B| are.
When applied to the disjunction of two non-neighbouring colour predicates, such
as red or violet, (Dis) tells us that the satisfiers of red or violet will be the union of the
satisfiers of red and the satisfiers of violet. But the result is more interesting when
the principle is applied to a disjunction of neighbouring colour terms, such as red or
orange. By (Sat), a borderline red-orange object, x, to which both the red pole r and
the orange pole o are maximally close, will satisfy neither red nor orange. However,
x does belong to the closure of the union of red objects and orange objects. Asufficient
condition for membership in that closure is that there be a pole in the union which
is maximally close to x. Since both r and o are maximally close to x, this condition is
met. What is more, x belongs to the interior of that closure. The additional condition
for membership in the interior is that p (M (p, x) p Cl (|red| |orange|)).
The only poles that are maximally close to x are r and o. Each of them belongs to
|red| |orange| and a fortiori to Cl (|red| |orange|).
According to (Dis), then, a borderline red-orange object satisfies the disjunc-
tive predicate red or orange, even though it satisfies neither red nor orange.
When applied to adjacent bands of colour, (Dis) sweeps up into the extension of
either red or orange objects that are reddish-orange, even though those objects
satisfy neither disjunct. Some writers on vagueness have been attracted by this
21
result, but our analysis puts it on a firm theoretical basis. It is a consequence of
the general principle that the extensions of polar predicates are regular open sets
in the topology that the poles generate.
20
Proof:For any X, X Cl (X), whence Int (X) Int Cl (X), since Int is MONOTONE. When X is
open, X=Int (X) so in that case Int Cl (X) contains X. To show that Int Cl (X) is itself regular open,
we need to show that Int Cl (X)=Int Cl Int Cl (X), i.e., that X=X . Since X is open, this follows
directly from the result proved in the previous footnote. Finally, suppose that Z is a regular open set
and X Z. Since both Int and Cl are MONOTONE, Int Cl (X) Int Cl (Z), whence Int Cl (X) Z, by
hypothesis. So Int Cl (X) is the smallest regular open set containing X.
21
For example, Dummett. Consider a vague statement, for instance That is orange. If the
object pointed to is definitely orange, then of course the statement will be definitely true; if it
248 Five Attacks on Classical Logic
Quine once derided as a fantasy the idea that a disjunction might be true
without either disjunct being true. He would have said the same about the present
suggestion, that an object might satisfy a disjunctive predicate while satisfying
neither disjunct. But even by Quines lights the idea that satisfaction might fail to
distribute over disjunction is not really fantastical. In his work on natural kinds,
he was one of the first to recognize how our schemes of classification tend towards
the appropriate form of convexity. The present theory, which requires a predi-
cates extension to be regular open, embodies a generalization-cum-liberalization
of that Quinean insight.
Tarski was led to explore the algebra of regular open sets through his inves-
tigations into the geometry of solids, and the geometrical analogy may help
the reader attain a better grasp of the semantic theory that is being proposed.
Imagine, then, two perfect cubes of the same size, aligned on an eastwest
axis, each of whose six faces face north, south, east, west, up, and down. When
the cubes are separated, none of the faces belongs to the interior of any cube.
Suppose, though, that the cubes are joined together to form a double cube.
Then the eastern face of the western cube, which now coincides with the west-
ern face of the eastern cube, does belong to the interior of the double cube.
Thus every solid that is part of the double cube is either a part of the eastern
cube, a part of the western cube, or else it straddles the common face. It is simi-
lar with adjacent colours. There are many colour shades intermediate between
polar red and polar orange, some determinates of red, some determinates of
orange, and some that straddle the border. Those that straddle the border are
determinates of red or orange without being determinates of either red or of
orange.
Although (Dis) leads to this violation of classical semantic theory, it does not
lead to any deviations from classical logic. The result that establishes this is a
theorem that was already applied, in a different context, in Chapter4. Take the
is definitely some other colour, then the statement will be definitely false; but the object may
be a borderline case, and then the statement will be neither definitely true nor definitely false.
But, in this instance at least, it is clear that, if a borderline case, the object will have to be on the
borderline between being orange and being some other colour, say red. The statement That
is red will then likewise be neither definitely true nor definitely false:but, since the object
is on the borderline between being orange and being redthere is no other colour which is a
candidate for being the colour of the objectthe disjunctive statement, That is either orange
or red, will be definitely true, even though neither of its disjuncts is (Dummett 1975, 255). As
already remarked, Dummett held that the only possible meaning we could give to the word
true <in connection with vague statements> is that of definitely true (Dummett 1975,
256). So he was committed to there being cases in which a disjunction is true even though nei-
ther disjunct is true.
The Challenge from Vagueness 249
family of regular open sets of any (non-empty) topological space. Then the fam-
ily forms a Boolean algebra if (1)the zero of the algebra is the empty set; (2)the
unit is the whole space; (3)the meet of two sets is their intersection; (4)the join
of two sets is the interior of the closure of their union; and (5)the complement of
22
a set is its exterior. According to (Con), (Dis), and (Neg), precisely these opera-
tions interpret and, or, and not when these expressions are understood to link
monadic polar predicates. It follows that the logic of these predicate connectives
will be classical when they connect polar predicates.
This argument will not persuade someone who doubts or denies that classical
logic is applicable to vague languages. The metalanguage in which the semantic
analysis has been expounded is as vague as our stipulated object language, and
the analysis has invoked contested classical rules at various points. (For exam-
ple, the argument that X =Int Cl (X) used the premiss that U (U X)=X,
which rests on the rule of Double Negation Elimination.) We can, however, say
this much. If we follow our semantic theory for polar predicates to the only logical
terminus to which it is headingnamely, to classical logicwe attain the sort of
reflective equilibrium we hope to find in the combination of a logic with a seman-
tics. For by reasoning classically in the metalanguage we can prove that the clas-
sical rules (when applied to the object language) are sound and complete with
respect to the semantics.
When the relevant language is interpreted in accordance with that seman-
tics, though, some theorems of classical logic do not say exactly what one
expects them to say. Frege understood the Law of Excluded Middle to express
the requirement that the concept should have a precise boundary. It does not
express that requirement under the present interpretation. An object that is a
borderline instance of a predicate A will have more than one pole maximally
close to it. So it will belong to Cl (|A| |A|) and to Int Cl (|A| |A|), and hence
satisfy A or not A , even though it satisfies neither A nor not A . The Law of
Excluded Middle, then, does nothing to exclude borderline cases or to enforce
sharp boundaries. To do that, we would have to lay it down that each object is
either a clear case of A or a clear case of not A. Our apparatus may help to explain
the relevant notion of being clearly Aperhaps in terms of being much closer to
the pole associated with A than to any other pole. However that may be, border-
line cases are not ruled out by Excluded Middle.
22
For proof of the theorem, see again Halmos 1963, 4. For the previous application of it, see 4.5.
250 Five Attacks on Classical Logic
that seems to assert that the predicate red has a sharp boundary. Now (10) may be
recast as a long disjunction,
where, as we have just seen, disjunctions do not say what one first expects them
to say:an instance of Excluded Middle for predicates does not express the exist-
ence of a sharp boundary between the objects that satisfy the predicate and
those that do not. So we might hope to lay the Sorites to rest, not (as Wright
does) by finding some way to avoid any commitment to (10) or (10), but instead
by reconciling ourselves to (10) and (10) when these formulae are correctly
interpreted.
This thought is an important part of the solution to the Sorites that Irecom-
mend, but further analysis is needed before we can address the argument of
8.1 head on. Our semantic theory validates the classical relationships between
predicate operators:the operators and, or, and not characterized by (Con),
(Dis), and (Neg) all take predicates to form predicates. Imake no apology for
having started with predicate operators rather than with the more familiar sen-
tential connectives. We have been trying to develop a theory of polar predicates
and, in doing so, it is natural to consider complex predicates before consider-
ing complex statements. All the same, in the paradoxes in 8.1, , , and
are all sentential connectives and it is essential that they should be so. Astep
in a Sorites argument cannot be analysed as an inference from premisses say-
ing that a given object has such-and-such a property, to the conclusion that the
same object has some further property:the thrust of the Sorites is to argue for
assigning a given vague property to a newly considered object. So the logic to be
employed in evaluating the Sorites is moot until we have extended our account
of the meanings of and, or, and not to yield an account of the meanings of ,
, and . We do not, however, need to extend our theory further and explain
the meanings of quantified formulae. Instead we can replace existentially
The Challenge from Vagueness 251
the predicate red-red-red if <r, r, r> is the only one of these ternary pole that is
maximally close to <a, b, c>, and similarly for the other atomic classifications.
As before, the relation of maximal closeness generates an interior operation
call it Int3on sets of ordered triples of possible objects. Again as before, the fact
that <r, r, r>, <r, r, o>, <r, o, r>, <r, o, o>, <o, r, r>, <o, r, o>, <o, o, r>, and <o, o, o>
form a system of contrary classifications of ordered triples means that the eight
atomic predicates of ordered triples have extensions that are regular open sets in
the topology defined by Int3. The same is true of the complex predicates generated
from these atoms using and, or, and not. The resulting logic will be classical,
so long as we are considering inferences in relation to a fixed set of poles. Given
that that condition is met, the two-step Sorites inference ai is red; it is not the case
that ai is red while ai+1 is not red; so ai+1 is red; it is not the case that ai+1 is red while
ai+2 is not red; so ai+2 is red may be validated as follows. From the premiss that ai
is red, we infer that the triple <ai, ai+1, ai+2> is either red-red-red, red-red-orange,
red-orange-red, or red-orange-orange. From the second premiss, we infer that
the triple is not red-orange-red or red-orange-orange, so that it must be either
red-red-red or red-red-orange, showing that ai+1 is red. From the third premiss,
we deduce that the triple is not red-red-orange, so that it must be red-red-red,
showing that ai+2 is red.
This suggests, indeed, how our analysis may be extended to cover entailments
between complete statements. Suppose we have a set X of statements and a single
statement B, and suppose that the sequence <a1,...,a N> comprises references for all
the singular terms that appear either in B or in X. Then, for some suitable choice of
poles, the truth of any statement in X {B} will consist in that sequences satisfy-
ing the N-place predicate that is formed by extracting all the singular terms from
the statement in question and adding dummy argument places where necessary.
(Compare Tarskis definition of truth as satisfaction by all countable sequences.)
B may then be said to follow logically from X when the predicate that is formed in
this way from B is a consequence, in our logic of predicates, of the predicates that
are similarly formed from the members of X. In this way, (Con), (Dis), and (Neg)
may be understood to assign senses to the usual sentential connectives, as well as
to predicate-forming operators on predicates.
How does this solve the paradox of 8.1? The paradox lay in the fact that from
the true premisses (2)and (3), we have a classically valid argument to the conclu-
sion (10), where (10) seems to say that red has a sharp boundary. The problem
may be resolved as follows. Given the proposed account of truth, the truth of (10)
consists in the fact that the sequence <a1,...,a100> satisfies the 100-place predicate
The Challenge from Vagueness 253
for some suitable choice of poles for this 100-place predicate. Now the only reason
we have to suppose that (10) is true is that it follows classically from (2)and (3),
both of which are true. So, if we read (10) as being true, that is because we under-
stand it to relate to a choice of poles which validates the argument from (2)and
(3)to (10). When taken in that way, however, (10) does not say that red has a
sharp boundary.
For what will a validating choice of poles look like? Since the only pos-
sible colours for the objects in the sequence are red and orange, the space
100
of possible poles are the 2 poles ranging from red-red-...-red (100 times) to
orange-orange-...-orange. Given the premisses of the Sorites argument, only 98
of these poles are serious contenders for validating the argument:they are the
poles that range from red-...-red (99 times)-orange to red-orange-...-orange
(99 times). But these poles remain poles by reference to which we may classify
whole sequences of objects in respect of their colours; they are not poles by refer-
ence to which we classify individual objects standing alone. When understood
in that way, (10) is entirely innocuous. All it says is that when classifying entire
sequences of coloured objects, whose members are arranged in order of gradu-
ally decreasing redness from something clearly red to something clearly not red,
either the second or the third or ... or the 100th object will be the first object
not classified as red. That claim is obviously true. It does not, however, entail the
existence of a sharp boundary to the concept red, which is a mode of classifying
a single object in respect of its colour. Indeed, it does not say anything directly
about that latter mode of classification at all.
It may be protested that (10) still has implications for the latter, monadic
classification. The formula (10) is a disjunction, and a disjunction cannot be
true unless one of its disjuncts is true. In the present case, at most one disjunct
can be true, so let us suppose that a N is red (a N+1 is red) is the unique true
disjunct. Surely, the protest runs, that means that a N is the last red member
in the sequence. And surely that means that a N marks the boundary of red.
However, our earlier case for denying that satisfaction distributes over dis-
junction extends to show that truth does not distribute over it either. Both the
monadic colour-pole, r, and the monadic colour-pole, o, are maximally close
to a red-orange object. Such an object is perceived to be no nearer the one pole
th
than the other. Suppose then that the N object in the sequence is red-orange.
th
Then both the 100-place colour-pole in which the N object is classified as the
sequences last red member, and the 100-place colour-pole in which that object
is classed as the first orange member, will be maximally close to the whole
254 Five Attacks on Classical Logic
sequence of coloured objects. For the sequence, too, is no nearer one of these
poles than the other. But if we have two poles that are maximally close to a
sequence, then we shall have violations of distributionin this case, of the dis-
tribution of truth. So the argument by which the objector seeks to move from
the truth of (10) to the existence of a boundary for the monadic concept red
breaks down at the first stage.
When introduced to the Sorites, many peoples first thought is to appeal
to the distinction between assigning properties to a whole sequence of objects
and assigning properties to an object considered in isolation. When we classify
entire sequences, they are apt to say, there must be a switch in colour somewhere,
but there is no fact of the matter as to where, and the place we choose to make
the switch does not mark any boundary of the concept red. Our analysis shows
that this nave reaction to the paradox is essentially correct. Or better, the polar
semantics provides a theoretical context within which the nave thought can
alleviate the sense of perplexity. The two cruces are the gradual switching from
monadic to relational poles of increasing adicity as we move along a Sorites series,
and the failure of distribution. It is sometimes futile to ask which limb of a true
disjunction is true, not because the answer is unknowable but because there is
nothing to know.
Essentially the same diagnosis also disposes of the structural version of the
Sorites Paradox that Ipresented in 2.3. Aconclusion B was said there to be a
Philonian consequence of some premisses X if either B is true or some member
of X is not true. Given a classical metalogic, we proved that Philonian conse-
quence obeys the Cut Law. Let A n be the statement that tube an is red (for n
between 1 and 100). If there were a number N such that A N were true but A N+1
were not true, N would mark a sharp cut-off for red in the sequence of tubes.
So it is tempting to maintain that, for any n between 1 and 99, it is not the case
that An is true while A n+1 is not true, i.e., that either An+1 is true or A n is not true,
i.e., that An+1 is a Philonian consequence of An. It is tempting, in other words, to
say that there is a whole sequence of entailments:A 2 is a Philonian consequence
of A 1;...;A 100 is a Philonian consequence of A99. However, the Cut Law then
entails that A 100 is a Philonian consequence of A 1, when it clearly is not:A 1 is true
whereas A 100 is not true.
The poles relevant to this argument are the True and the False, rather than r
and o. Apart from that difference, though, the same analysis applies. The struc-
tural Sorites reasoning shows that either (A1 is true and A 2 is not true) or ... or
(A99 is true and A 100 is not true). However, the truth of this disjunction does not
entail the truth of any of its disjuncts, so we cannot apply distribution to conclude
that there is a sharp boundary to the extension of the predicate true.
The Challenge from Vagueness 255
The solution to the Sorites that Ihave proposed applies only to versions of the
paradox that involve predicates whose meaning is given by reference to para-
digms or poles. Ihave no proof that all soritical predicates are polar, still less that
all vague terms are. It may well be that there are different versions of the Sorites
involving different sorts of predicate, and that these different versions conse-
quently call for different analyses and resolutions. Whether this is so is something
left for future work, but leaving this question open does not deprive our analysis
of interest:it is a mistake to hold that in order to solve any version of a paradox one
must solve all versions. We have, at any rate, vindicated classical logicalthough
not classical semanticsas the appropriate logic to use when assessing deduc-
tions that involve a large and central group of vague predicates.
spaces within which we can find appropriate semantic values for such statements
as a is red and b is hard and Either c is poor or d is bald.
The key technical notion is that of the product of two topological spaces. Let
us represent a topological space T as a pair {A, }, where A is a non-empty set
and is a collection of subsets of A. The space meets the following conditions:
(1)A, ; (2)the intersection of any two sets in is again in ; (3)the union
of any collection of sets in is again in . The members of , then, are the open
sets of the topology . Given two such spaces T1={A1, 1} and T2={A2, 2}, the
product topology for the Cartesian product A 1 A 2 is the topology with basis
B={U1 U2:U1 1, U2 2}. That is to say, the members of are all the sets that
may be formed by taking unions of members of B. The space T1 T2={A1 A 2, }
is called the topological product of T1 and T2. When H is a subset of A1 and K is a
subset of A 2, Int(H K)=Int(H) Int(K) and Cl(H K)=Cl(H) Cl(K). Hence
the Cartesian product of a regular open set in T1 with a regular open set in T2 is a
regular open set in the product space T1 T2. These definitions and results extend
by induction to any finite number of factor spaces.
How does this notion help to account for the validity of heterogeneous infer-
ences? The idea may be conveyed by working through a couple of examples. Let
T1 be the topological space generated by our seven colour poles, and let T2 be the
space generated by a similar system of paradigms of hardness, softness, etc. There
is a regular open set R in T1 such that an arbitrary member, a, of the domain of T1
satisfies red if and only if a R. Now let b be a member of the domain of T2 . Then
the ordered pair <a, b> satisfies x 1 is red if and only if <a, b> R T2; R T2 is a
regular open set in the product topology T1 T2 . There is also a regular open set H
in T2 such that an arbitrary member, b, of the domain of T2 satisfies hard if and
only if b H. The ordered pair <a, b> satisfies x 2 is hard if and only if <a, b>
T1 H; T1 H is another regular open set in the product topology. Our semantics
interprets the conjunction of two predicates as the intersection of the semantic
values of the conjuncts. Accordingly, the semantic value of x 1 is red and x 2 is hard
is (R T2) (T1 H). Quite generally, (A B) (C D)=(A C) (B D), so
(R T2) (T1 H)=(R T1) (T2 H)=(R H). By definition, though, <a, b> R H
if and only if a R and b H, i.e., if and only if a satisfies red and b satisfies
hard. This result accounts for the validity of the inference from the two prem-
isses Tube a49 is red and Tube a50 is hard to the conjunctive conclusion Tube a 49
is red and tube a50 is hard.
Our semantic principles also account for the validity of inferences in the form
Either c is poor or d is bald; c is not poor; so d is bald. The extension of poor will
be a regular open set P in a topological space, T3, generated by a system of poles for
wealth. The extension of bald will be a regular open set B in a distinct topological
The Challenge from Vagueness 257
space, T4, generated by poles for hirsuteness. Apair <c, d> satisfies x 1 is poor if and
only if <c, d> P T4; it satisfies x 2 is bald if and only if <c, d> T3 B. Applying
our semantic principle for disjunction, we conclude that <c, d> satisfies x1 is poor
or x2 is bald if and only if <c, d> Int Cl ((P T4) (T3 B)). Applying our seman-
tic principle for negation, we conclude that <c, d> satisfies x1 is not poor if and
only if <c, d> Int ((T3 T4) (P T4))=Int ((T3 P) T4). Given that <c, d> satisfies
both x1 is poor or x2 is bald and x 2 is not poor, it then follows that <c, d> T3 B,
so that d satisfies bald. In this way, our semantic theory accounts for the validity
of Either Fred is poor or Tom is bald; Fred is not poor; so Tom is bald.
We can account similarly for the validity of every valid sequent in the classi-
cal propositional calculus, even if the sequent is heterogeneous. Admittedly, fur-
ther technical problems arise in extending the semantic theory to the quantifiers.
Ibelieve that those problems can also be solved, but the solution lies beyond the
scope of this book.
The fact that our semantics validates classical sequents even if these contain
heterogeneous predicates may seem to conflict with a key element in the proposed
solution to the Soritesthe claim, namely, that a disjunctive statement may be
true without either disjuncts being true. For there is a classically valid argument
for the conclusion that a true disjunction contains at least one true disjunct which
seems to rely on only innocuous principles about truth. Let us suppose that the
statement u says that either tube a50 is red or tube a50 is orange, that the statement
v says that a50 is red, and that the statement w says that a50 is orange. Then, using
Say (u, P) to formalize u says that P, we have the following premisses:
A parallel derivation may be given no matter what sentences replace Tube a50 is
red and Tube a50 is orange, so we have a classically valid argument, resting upon
apparently compelling principles about truth, for the general conclusion that any
true disjunction contains at least one true disjunct. For our purposes, though,
the particular deduction just set out presents problems enough. According to our
semantic analysis, the disjunction Either tube a50 is red or it is orange is true while
neither disjunct is truewhich is contrary to the conclusion of the deduction.
The flaw in the derivation, Icontend, lies in the rule R2. Of course there is a
close connection between an objects being red and the truth of a statement say-
ing that the object is red. R2, however, misdescribes that connection. What we
can say is that when v is a paradigm case of a statement saying that tube a50 is red,
and when tube a50 is a paradigm case of redness, then v will be a paradigm case of
truth. It will be close to the pole of truth, the True. But the rule R2 commits one
to more than that. In the move from (8)to (9), for example, it is applied to a case
where tube a50 is a borderline case of redness, not a paradigm. In such a case the
inference to v is true is not justified.
It may be replied that, in this application of R2, the conclusion v is true is
no further from the truth than the premiss Say (v, a50 is red) a50 is red. It is
indeterminate whether a50 is red, and hence indeterminate whether v says that
a50 is red, and a50 is red (it is given that v says that a50 is red). So the application
of R2 in moving from (8)to (9)at least takes us no further from the truth than
we already were. In fact, though, even this is wrong. True and false are polar
predicates, and before we can assess v is true, we need to know which poles the
assessment is made against. If the only poles in the system are the True and the
False, it would be correct to assess v as being as close to the True as it is to the
False, just as tube a50 is as close to polar red as it is to polar orange. If, however,
there is a third polelet us call it Indeterminatethen the inference from (8)to
(9)takes us from an indeterminate premiss to a false conclusion:if v is close to
the pole of Indeterminate, then it is false to ascribe the contrary status of truth
to it. Recall Sainsburys remark that the introduction of a third ... pole can
attract some of the things only loosely attached to two existing poles.
The Challenge from Vagueness 259
A lesson of our discussion is, indeed, that this third pole for assessing the
alethic status of statements arises naturally from the polar semantics. For
how did we reach the conclusion that a true disjunction need not contain a
true disjunct? The violation of distribution for truth arose out of the corre-
sponding violation for satisfaction:a red-orange object was found to satisfy
either red or orange even though it satisfies neither red nor orange. This
latter violation arose in turn because a red-orange object does not belong to
the smallest regular open set containing the red pole r, nor to the smallest
regular open set containing orange pole o, but it does belong to the smallest
regular open set containing both these poles. We see here how an interme-
diate semantic pole arises naturally within the theory even though there is
no colour pole intermediate between r and o. There is no red-orange colour
pole:a normal viewer looking at the spectrum continues to perceive the origi-
nal seven colour bands r, o, y, g, b, i, and v. However, a statement in which red
is predicated of an object to which both r and o are maximally close will be
a paradigm case of the semantic pole Indeterminate. Afurther attraction of
the present semantic theory is, indeed, that it explains in this way how our
willingness to use classical logic when reasoning with vague predicates sits
alongside a reluctance to accept classical semantics for those terms (see fur-
ther Chapter10).
The thesis that truth does not distribute over disjunction is akin to the
supervaluationists claim that super-truthi.e., truth under all admissible
precisificationsdoes not so distribute. As already remarked, the philo-
sophical context of the two theories is quite different. Supervaluationists take
vagueness to be semantic deficiency:users of the English word red have not
decided which precise and exhaustive division of possible objects to associ-
ate with the term. On the polar account, exhaustive divisions are beside the
point: users of red are not even trying to draw a boundary, although the
polar sense of the term does not preclude their stipulating one for special
local purposes. The supervaluationist also deems the rule R2 to be inappli-
cable when reasoning from a supposition. He rejects R2 because he takes the
truth of a vague statement to consist in its being true under all admissible
precisifications of its vague terms. This account of truth has been the target
of powerful objections (see notably Williamson 1994, 1624). Those objec-
tions do not touch the present explanation of why R2 fails, which is cast not in
terms of super-truth, but in terms of three poles:the True, the False, and the
Indeterminate.
Our analysis, indeed, enables us to identify the flaw in Williamsons argu-
ment against the possibility of truth-value gaps (Williamson 1994, 18798). The
260 Five Attacks on Classical Logic
argument runs as follows. Let X be the set of statements {u says that a50 is red,
If u says that a50 is red, and a50 is red, then u is true}, let A be the statement a50 is
red, and let B be the statement u is true. X and A jointly entail B. Let Y be the set
{u says that a50 is red, If u says that a50 is red, and a50 is not red, then u is false}, let
C be a50 is not red, and let D be u is false. Y and C jointly entail D. Aclassically
valid rule of proof is the following form of contraposition with side premisses:if
X and A jointly entail B, then X and not B jointly entail not A . Applying
contraposition to the first entailment, we infer that X and u is not true together
entail a50 is not red. Applying contraposition to the second entailment, we infer
that Y and u is not false together entail a50 is not not red. Amalgamating prem-
isses, then, we conclude that Z together with u is not true and u is not false
entails a contradiction, where Z={u says that a50 is red, If u says that a50 is red,
and a50 is red, then u is true, If u says that a50 is red, and a50 is not red, then u
is false}. The first member of Z is true by stipulation and Williamson assumes
that any satisfactory account of truth and falsity will entail Zs second and third
members. Accordingly, anyone who maintains that u is neither true nor false
i.e., that u is both not true and not falseis committed to the contradiction that
a50 is both not red and not not red. Aparallel argument may be used to reduce to
absurdity the hypothesis that any other statement is neither true not false.
Those who wish to allow truth-value gaps have protested that, if such gaps exist,
then one must distinguish between internal and external notions of negation, a dis-
tinction Williamsons argument neglects (see for example Maudlin 2004, 1969).
But that protest will not impress Williamson. The working hypothesis, he will
insist, must be that negation is univocal; on that hypothesis the argument shows
that there are no truth-value gaps, so the purported distinction between internal
and external negations marks no difference.
A standoff looms, which our analysis can avoid. The crux of Williamsons
argument is not the interpretation of not but the number of poles by reference
to which we attach senses to true and false (all parties agree that these terms
are vague). Williamsons argument goes through if the only relevant poles are the
True and the False. If, however, we also admit the Indeterminate, then the two
conditionals which Williamsons argument assumes cannot be accepted. For in
that case, If u says that a50 is red, and a50 is red, then u is true has an indeterminate
antecedent and a false consequent:because a50 is only a borderline case of redness,
u is close to the Indeterminate, so u is true is false. The same goes for the paral-
lel conditional about falsehood:If u says that a50 is red, and a50 is not red, then u
is false also has an indeterminate antecedent and a false consequent. Whatever
ones theory of negation, then, Williamsons argument is question-begging. In
assuming that that any acceptable theory of truth will validate his conditionals,
The Challenge from Vagueness 261
he is already assuming that there are no truth-value gaps. The polar analysis
brings this out clearly.
My admission of truth-value gaps is a basic point of difference with Raffmans
theory of vagueness (see again her 2005). She, too, develops Sainsburys idea
that vague predicates lack boundaries in such a way as to validate classical logic.
Her account, though, also validates classical semantics, including the Principle
of Bivalence. For the reasons given, Ido not regard Bivalence, when applied to
vague statements, as an intuitively compelling principle which we ought to try
to preserve:to the contrary, Ihave explained how our semantic analysis leads
naturally to the recognition of a pole intermediate between the True and the
False. In any event, Raffmans way of preserving Bivalence conflicts with a leit-
motif of my theory. Entirely plausibly, she counts a statement as false when its
negation is true. Less plausibly, though, she also takes the extension of a predi-
cates negation to comprise all the objects that fail to satisfy the predicate. That
is, she takes the extension of a predicates negation to be the complement of the
extension of the predicate itself. In general, the complement of an open set is not
open, so Raffmans treatment of negation conflicts with the principle that both a
predicate, and its negation, have open extensions. The principle that the exten-
sion of a negated predicate is open is far more compelling than the claim that
any statement involving vague terms is either true or false. When x is red, there
is a neighbourhood of possible objects around x all of whose members are red;
equally, when y is not red, there is a neighbourhood of possible objects around y,
all of those members are not red. My semantic theory for polar terms respects this
symmetry; Raffmans does not.
My semantics also has points in common with Dorothy Edgingtons degree
theory (see her 1996). Edgington stresses, as Ihave done, that familiarity with
clear cases (and clear counterfoils) plays a large part in our mastery of a vague
term (1996, 303); her solution to the Sorites also depends on the idea that indi-
vidual deductive steps that are acceptable in isolation cumulatively lead to trou-
ble. However, while her postulation of numerical degrees of closeness to clear
truth or verities may be an acceptable idealization, the key notion of her seman-
tic theory for vague languagesthat of a conditional verityis hard to interpret.
The conditional verity of B given A, she explains, is the value [i.e., the degree of
closeness to clear truth] to be assigned to B on the hypothetical decision to count
A as definitely true (1996, 306). This test is applied when A is not definitely true,
so Edgingtons hypothetical decision involves changing the meaning of a vague
predicate. It is hard to know how far this change of meaning ramifies. Let us revert
to Dummetts red-orange curtain, and let A be the statement This curtain is red.
As we currently use red, A is not definitely true. What degrees of closeness to
262 Five Attacks on Classical Logic
23
If, indeed, there is such a thing; for an interesting discussion of this question, see Sainsbury 1991.
9
On the Use of Classical Logic
in Set Theory
A number of philosophers have argued that special problems attend the use of
classical semantics and classical logic in set theory. Already in his 1963 paper
The Philosophical Significance of Gdels Theorem, Dummett maintained that
the concept of set was indefinitely extensible and that the classical explana-
tion of the quantifiers breaks down when applied to all sets (see Dummett 1963,
199200). In an essay published thirty years later entitled What is Mathematics
About? he went further. Not only does classical semantics give no coherent
account of what all sets means, but statements involving such quantification do
not satisfy the laws of classical logic, but only the weaker laws of intuitionistic
logic (Dummett 1993b, 441).
In his paper Sets and Semantics (1977), Jonathan Lear advanced a different argu-
ment for the same conclusion. Unlike Dummett, Lear is a realist about sets, but his
sort of realist maintains that our understanding of what a set is will always remain
partial <so that> the extension of the term set will always be capable of expansion
(1977, 91). Lear spells out the connection he perceives between these theses:
our comprehensional limitations affect our ability to talk about sets. Given all the proper-
ties that we know are true of sets, the extension of set may encompass all sets with those
properties. But, since we are always in a position to improve our understanding of what a
set is, we will always be able, at some time, to recognize that what appeared, at an earlier
time, to be the universe, actually was elements of a set:e.g., the first inaccessible. (1977, 91)
Each of these arguments has generated a fair amount of analysis and dis-
2
cussion. Iwant to focus, though, on the anti-classical case that was set out by
William Tait in his paper Zermelos Conception of Set Theory and Reflection
Principles (Tait 1998). Parts of Taits case are elements in Dummetts and
Lears, but Taits is the most powerful challenge to the use of classical logic in
this area, largely because his argument engages more directly with the specif-
ics of axiomatic set theory. His case has the further merit of eschewing the
rather dark notion of indefinite extensibility that Dummett took over from
Russell. Despite that difference, Dummett would, Ithink, have accepted all of
Taits premisses, and he would certainly have welcomed his conclusion:the
logic that applies to arbitrary formulas of set theory, when these are inter-
preted in the universe of all sets, should be constructive, not classical logic
(Tait 1998, 478).
I shall contend, though, that Dummett, Lear, and Tait are only half right.
They are correct to hold that the familiar classical semantics, in which the
meanings of the connectives are given by their truth-tables, cannot be applied
to the language of set theory. If it is to validate classical logic, that semantics
needs the Principle of Bivalence, and Iagree with Dummett & co. that we are
not entitled to assert that an arbitrary set-theoretic statement is either true or
false. We need, then, a non-classical semantics for the language of set theory.
The semantics that Irecommend involves a number of elements familiar from
constructivist theories of meaning. However, the logic it validates is in the end
3
classical.
2
On Dummetts argument, see Heck 1993, Velleman 1993, Clark 1998, Oliver 1998, Shapiro and
Wright 2006, and Sullivan 2007. On Lears, see Paseau 2001, 2003, and Incurvati 2008.
3
Dialetheist logicians often cite the set-theoretic paradoxes as cases where their preferred devia-
tions from classical logic come into their own. These anti-classical arguments are interesting and
deserve a detailed reply. The dialetheists arguments, though, raise very different issues from the
constructivist challenges with which Iam concerned here; trying to bundle together different sorts
of anti-classical case only engenders confusion.
On the Use of Classical Logic in Set Theory 265
true, and he further accepted that this statement entails There are prime num-
bers and hence There are numbers; this latter statement, then, is also strictly
and literally true. Dummett, though, was alert to the deep metaphysical differ-
ences between mathematical entities on the one hand and concrete objects on
theother:
In order to confer upon a general term applying to concrete objectsthe term star, for
examplea sense adequate for its use in existential statements and universal generaliza-
tions, we consider it enough that we have a sharp criterion for whether it applies to a given
object, and a sharp criterion for what is to count as one such objectone star, sayand
what as two distinct ones:a criterion of application and a criterion of identity. The same
indeed holds true for a term, like prime number, applying to mathematical objects, but
regarded as defined over an already given domain. It is otherwise, however, for such a
mathematical term as natural number or real number which determines a domain of
quantification. For a term of this sort, we make a further demand:namely, that we should
grasp the domain, that is, the totality of objects to which the term applies, in the sense of
being able to circumscribe it by saying what objects, in general, it compriseswhat natu-
ral numbers, or what real numbers, there are.
The reason for this difference is evident. For any kind of concrete object ... external
reality will determine what objects of that kind there are; but what mathematical objects
there are within a fundamental domain of quantification is supposed to be independent
of how things happen to be in the world, and so, if it is to be determinate, we must deter-
mine it. (Dummett 1993b, 438)
4
We assume that we have a grasp of the totality of natural numbers:but do we? Certainly, we
have a clear grasp of the step from any natural number to its successor:but this is merely the essen-
tial principle of extension. The totality of natural numbers contains what, from our standpoint, are
enormous numbers, and yet others relative to which those are minute, and so on indefinitely; do we
really have a grasp of such a totality? (Dummett 1993b, 443). The ensuing discussion makes it clear
that this question expects the answer no.
Dummetts position in The Philosophical Significance of Gdels Theorem had been less radi-
cal. In that earlier paper he held that there is really no vagueness as to the extension of natural
number (Dummett 1963, 196). However, the concept of natural number is said to exhibit a dif-
ferent form of indefiniteness:it is part of this concept that the natural numbers form a totality
to which induction, with respect to any well-defined property, can always be applied, and ... the
concept of a well-defined property [of a number] is indefinitely extensible (Dummett 1963, 198). In
1963, though, Dummett held that a concept could manifest this latter form of indefiniteness while
still possessing a definite extension:the definiteness of the extension of natural number does not
imply definiteness for the notion of a ground for asserting something to be true of all natural num-
bers (Dummett 1963, 197).
On the Use of Classical Logic in Set Theory 267
5
For example, in formulating the principle of mathematical induction in 60 of Was Sind und
Was Sollen die Zahlen?, Dedekind quantifies over properties of elements of chains with no restric-
tion to the properties that are expressible in a given language.
268 Five Attacks on Classical Logic
for any expression e of L, the semantic value of e with respect to N is the result of
applying f to the value of e with respect to M. This entails that if N then M ,
for any closed sentence of L. Since there is also an isomorphism from N to M, a
parallel argument establishes the converse.
A familiar principle of classical model theory says that, for any model and any
closed sentence, either the sentence or its negation holds in the model. In particu-
lar, then, wehave
(2) For any model M of T, either M or M .
(4)Either T or T .
In other words, given this conception of truth and falsity for arithmetical state-
ments, we may prove that every statement in the language of arithmetic is biva-
lent. This in turn means that, if we take the meanings of the connectives to be
given by the familiar truth-tables, we may vindicate the use of classical logic in
arithmetical reasoning.
I do not suppose for a moment that this argument is going to persuade someone
who is unwilling to reason classically in arithmetic that he ought to do so. At a
number of points, it relies on classical logical principles that have been seriously
contested. The argument does show, though, that someone who is willing to rea-
son classically can tell a coherent story about why the classical rules are the right
ones to go by in arithmetic. By reasoning classically in the meta-theory, we show
that any statement in the language of arithmetic is bivalentjust what we need in
order to show that the classical rules are indeed sound, given that the meanings of
the connectives are given by the usual truth-tables. In the present case, then, clas-
sical semantics and classical model theory combine to yield a coherent account of
the use that a classical logician makes of the language of arithmetic.
On the Use of Classical Logic in Set Theory 269
What is striking is that no similar story can be told for set theory. Even if we
axiomatize set theory using full, classical second-order logic, we do not attain
2
a categorical theory. At least, this is so if we take set theory to be ZF , or for that
2
matter ZFC (I consider stronger theories in the next section). As Zermelo showed
in his paper of 1930, ber Grenzzahlen und Mengenbereiche (On Boundary
2
Numbers and Domains of Sets), ZF is quasi-categorical:given any two models,
one will be isomorphic to an initial sub-model of the other. However, the theory
is not fully categorical:it leaves the height of the set-theoretic universe undeter-
2
mined. More exactly, Zermelo showed that each model of ZF has the form V,
6
where is a strongly inaccessible cardinalwhat he called a boundary number.
Zermelo postulated, as a principle of what he called meta-set theory, the exist-
ence of an unbounded sequence of boundary numbers (1930, 1233). That is, he
postulated that, for any ordinal , there is a strongly inaccessible cardinal greater
than . If this postulate is correct, then categoricity fails radically:there will be
2
an unbounded sequence of essentially different models of ZF of ever increasing
height.
This failure of categoricity means that we cannot vindicate the use of clas-
sical logic in set theory in anything like the way we vindicated its use in arith-
metic. The kernel of that vindication was to deem an arithmetical statement
true (false) when it (its negation) holds in all models of arithmetic, and then to
appeal to categoricity to justify Bivalence. Anyone with structuralist inclinations
will be equally attracted to the corresponding accounts of truth and falsity for
set-theoretic statements:such a statement will be true (false) if it (its negation)
holds in all models of set theory. Without categoricity, however, the subsequent
argument for Bivalence breaks down. Indeed, if we explicate truth and falsity in
2
the suggested way, and if we take ZF to be the theory that articulates our concep-
tion of the domain of sets, then Bivalence certainly fails. For consider the state-
ment A, There is at least one strongly inaccessible cardinal. The first inaccessible
cardinal, 1, is not a member of V1, so A does not hold in V1. However, A does hold
2
in any standard model of ZF greater than or equal to V2 . Since it fails to hold in
2 2
some models of ZF , A is not true. Since it does hold in other models of ZF , it is
7
not false either. In other words, A is not bivalent.
6
The hierarchy of sets V is defined as follows for each sort of ordinal :V0=; V+1=(V) for
each successor ordinal +1; V=< V for each limit ordinal . Cardinal numbers are taken to be a
species of ordinal:the cardinal of a set A is the least ordinal equinumerous with A. Acardinal is strongly
inaccessible if it is neither the power set of a smaller cardinal nor the supremum of a set with heredi-
tarily lesser cardinality.
7
It is widely acknowledged that set theory manifests a kind of indeterminacy that one does not
find in number theory, even though both theories are undecidable. Thus Benacerraf and Putnam:It
is instructive to compare [and contrast] set theory with number theory. In number theory too there
270 Five Attacks on Classical Logic
This means that we cannot apply classical semantics to the language of set
theory. At least, we cannot do so if we hope to vindicate the use of classical logic
in set theory. The Law of Excluded Middle illustrates the difficulty. The first line
of the formulas truth-table tells us that A A is true if A is true. The second
line tells us that A A is true if A is false. The whole truth-table, then, tells
us that A A is true if A is either true or false. But unless we are given the
further premiss that A is either true or false, we cannot detach and infer that
A A is true. So, if the meanings of the connectives are given by their stand-
ard truth-tables, A A is logically true only if every statement in the rel-
evant language is bivalent. Given that some set-theoretical statements are not
bivalent, the truth-tabular semantics fails to sustain classical logic for the lan-
guage of set theory.
are statements that are neither provable nor refutable from the axioms of present-day mathematics.
Intuitionists might <argue> that that this shows (not by itself, of course, but together with other
considerations) that we do not have a clear notion of truth in number theory, and that our notion
of a totality of all integers is not precise. Most mathematicians would reject this conclusion. Yet
most mathematicians feel that the notion of an arbitrary set is somewhat unclear (Benacerraf
and Putnam 1983, 19). Ihope that the account in the main text identifies the key difference between
the two theories.
On the Use of Classical Logic in Set Theory 271
in which the variable f is understood to range over absolutely all the functions
on the domain.
This point does not gainsay Taits type-theoretic account of the second-order
quantifiers. It shows, though, that only full interpretations of those quantifi-
ers will help in the quest for a categorical articulation of our conception of the
domain of sets. In the statements of Replacement, the variable f must be free to
range over all functions on the domain. Similarly, Taits power type (A) must be
understood to comprise all functions from A into {, }. Now the task of appre-
hending the totality of such functions is not the same task as that of apprehending
all the sets whose members are drawn from A, for functions and sets belong to dif-
ferent categories. One might reasonably worry, though, that the tasks, although
distinct, face common difficulties. In particular, one might worry that the totality
of all functions on a given domain is just as indeterminate as the totality of all sets
drawn from that domain, so that the legitimacy of using classical second-order
logic is just as doubtful as that of using classical logic when reasoning about all
sets. Ishall return to this issue.
With this reservation noted, let us revert to our question and ask whether there
2
is any prospect of finding a fully categorical extension of ZF that articulates our
conception of the domain of sets. In his 1930 paper, Zermelo was already clear that
2
there would be many categorical extensions of ZF , but none of them, he argued,
would adequately capture the conception of sets that he was trying to articulate.
Naturally, he wrote, one can always force categoricity artificially by the addition
of new axioms, but always at the cost of generality (1930, 1232). The axioms he
had in mind force categoricity by truncating the set-theoretic universe:the sim-
plest such axiom says that there is no strongly inaccessible cardinal; and, for any
given ordinal , we may formulate a similar axiom saying that there is no strongly
inaccessible cardinal greater than . It is clear that postulating such an axiom
would come at the cost of generality. Consider the theory that is got by adding to
8
Itake this more explicit formulation of Replacement from Shapiro 1991, 85.
On the Use of Classical Logic in Set Theory 273
2
ZF the axiom There is no strongly inaccessible cardinal. The resulting theory is
indeed fully categorical:all its models are isomorphic. But the structure all these
models share is precisely that of the set-theoretic universe up to V1, where 1 is
2
the first strongly inaccessible cardinal. To describe the models of ZF + There is
no strongly inaccessible cardinal, then, we precisely need to postulate the first
strongly inaccessible cardinal, and the parallel problem afflicts any attempt to
achieve categoricity by truncation. Set theory as a science, Zermelo held, must
be developed in the fullest generality (1930, 1232)and categoricity by truncation
frustrates that goal. We can now see, indeed, why Zermelo was driven to postu-
late an unbounded series of essentially different set-theoretic models, in each of
which the whole classical theory is expressed (1930, 1233). This series, he wenton,
reaches no true completion in its unrestricted advance, but possesses only relative
stopping-points, just those boundary numbers [i.e., strong inaccessibles] which separate
the higher model types from the lower. Thus the set-theoretic antinomies, when cor-
rectly understood, do not lead to a cramping and mutilation of mathematical science,
but rather to an, as yet, unsurveyable unfolding and enriching of that science. (1930, 1233)
Just for this reason, Zermelo regarded the non-categoricity of his set theory
not [as] a disadvantage but rather [as] an advantage, for on this very fact rests
the enormous importance and unlimited applicability of set theory (1930, 1232).
For reasons that will emerge, Ithink he was right about this, but since his view
precludes the use of classical semantics for the language of set theory, it is worth
asking if there are other ways of achieving categoricity that are not subject to
his objections. Might we instead achieve categoricity, not by truncation, but by
appending an axiom which ensures that the universe of sets transcends any spec-
ifiable height?
Vann McGee (1997) has proposed a new axiom of this kind. McGee works in
2
a system, ZFU , in which the first-order quantifiers are permitted to range over
Urelementei.e., objects that are not setsas well as over sets. He proposes as a
further set-theoretic axiom the postulate that these Urelemente form a set (call
2
this axiom McG). He then proves that the theory ZFU + McG gives a categorical
characterization of the pure setsi.e., the sets that may be formed from the empty
set using the standard set-theoretic operations. More precisely, McGee shows that
2
there is a theorem of ZFU + McG which, under the intended interpretation of the
2
language of ZFU , says the following:any structure that could, given the axioms
2
of ZFU + McG, serve as the interpretation of pure set and is a member of, will
9
be isomorphic to the intended structure of the pure sets.
9 2
That is, McGee shows that the theory of pure sets in ZFU + McG is internally categorical in
the sense of Walmsley 2002, IV.
274 Five Attacks on Classical Logic
10
On this, see further Muller 2001.
11
Menzel 1986, 413. Of course, if the natural numbers are ordinals, this is not a new applica-
tion of Benacerrafs argument but a reiteration of it. In What Numbers Could Not Be, though,
Benacerraf seemed to think of a natural number as something abstracted from its ordinal and car-
dinal applications.
276 Five Attacks on Classical Logic
their extension by new axioms which assert the existence of still further iterations of the
operation set of ... The simplest of these strong axioms of infinity asserts the existence
of inaccessible numbers (in the weaker or stronger sense) > 0. The latter axiom, roughly
speaking, means nothing else but that the totality of sets obtainable by use of the proce-
dures of formation of sets expressed in the other axioms forms again a set (and, therefore, a
new basis for further applications of these procedures). (Gdel 1964, 264)
The problem with interpreting the language of set theory, then, is not merely that
we have an unbounded sequence of successively larger domains of sets. We are
also doomed to postulate an infinite sequence of successively stronger axiom sys-
tems as we try to spell out what is involved in iterating the power set operation as
far as possible.
Tait argues on this basis that there is no prospect of ever attaining a full axi-
omatization of the iterative conception:
It is on these grounds that Tait challenges the use of classical logic in set theory:
So when we affirm the truth of a proposition about the universe of all sets, we must speak
of it on the basis of those operations for constructing sets that have already been accepted,
while admitting that further principles might also be admitted. For this reason, the logic
that applies to arbitrary formulas of set theory, when these are interpreted in the universe
of all sets, should be constructive, not classical logic. For there is no such well-defined
universe; there is just the possibility of stronger and stronger operations which we might
introduce for constructing sets. So, when x ranges over all sets, there are no grounds for
assuming x(x) x(x) unless it can be derived on the basis of operations that we
have already introduced. Indeed, it is precisely in connection with the absolute infinite, the
universe of all sets, that the ideas of constructive mathematics become compelling, even
for those who reject the restrictions to constructive methods in analysis. (Tait 1998, 478)
We have here a prima facie powerful case against using classical logic in set
theory. In assessing it, we need to focus on three questions:first, whether Tait is
278 Five Attacks on Classical Logic
correct to hold that the notion of iterating the power set operation as far as pos-
sible resists mathematical determination, so that the iterative conception yields
no well-defined universe of sets; second, assuming that there is no well-defined
universe, whether quantification over all sets makes sense; third, assuming that
such quantification makes sense even in the absence of a well-defined universe,
whether the best semantic account of such quantification invalidates some laws
of classical logic, notably x(x) x(x) . Iaddress these issues in turn.
(1) Recent work in set theory supports Taits contention that the iterative con-
ception resists a definitive axiomatization. Under that conception, Tait claims,
the question, given some initial segment of ordinals, of whether it has an upper
bound, so that we may continue the iteration of taking power sets beyond the
ordinals in , should depend only on the properties of or better of the universe
of sets obtained by iterating the power set operation through (Tait 2005b, 134).
This claim is plausible if we think of the universe of sets as being built up from
below. On the strength of this, Tait argues that the axioms that articulate the itera-
tive conception must all be general reflection principles of a particular form:they
must all be what he calls n-reflection principles for some finite n (for details, see
Tait 2005b, definition 6). It does indeed seem that reflection principles of this
form are the strongest that could be justified if ordinals have to be constructed
from below. However, Peter Koellner has shown that these n-principles are in a
sense weak:each of them is consistent with Gdels Axiom of Constructibility,
V=L (see Koellner 2009). More exactly, Koellner shows that Taits hierarchy of
reflection principles divides into those that are consistent with V=L and those
that are internally inconsistent. If as far as possible is taken to mean as far as
possible from below, then, it will not be possible to justify any iterations of the
power set operation that are inconsistent with V=L. Many other set theorists,
however, understand the iterative conception in such a way that it is consistent
with the existence of a measurable cardinal. It follows that there are incompatible
ways of making the iterative conception more precise for, as Dana Scott proved
some time ago, the Axiom of Constructibility is inconsistent with the existence of
a measurable cardinal (see Scott 1961). This bifurcation confirms Taits thesis that
the iterative conception resists full axiomatization. Indeed, it confirms his par-
ticular conjecture (in the passage quoted above from Tait 1998)that some devel-
opments of the iterative idea justify the claim that there are measurable cardinals
whereas others justify the contrary claim.
(2) One response to this difficulty would be to deny that quantification over
absolutely all sets makes sense. This was Russells view in Mathematical Logic as
Based on the Theory of Types (Russell 1908). Russell diagnosed the set-theoretic
On the Use of Classical Logic in Set Theory 279
paradoxes (and some others) as arising from the assumption of a totality such
that, if it were legitimate, it would at once be enlarged by new members defined in
terms of itself (1908, 63). The way out of paradox is to reject that assumption:If,
provided a certain collection had a total, it would have members only definable in
terms of that total, then the said collection has no total (1908, 63). The universe of
sets is a paradigm of a collection that meets Russells criterion for having no total,
but on Russells view, it follows that we cannot sensibly quantify over all members
of the universe:When Isay that a collection has no total, Imean that statements
about all its members are nonsense (1908, 63).
I do not think that one can force a sceptic to concede that statements about all sets
make sense. In quantifying over all sets we put ourselves at risk of falling into para-
dox, and one way to eliminate the risk is to refrain from making such statements. The
fact remains, however, that many set theorists are willing to take the risk; certainly,
vanishingly few of them accept the tight constraints on what can be said that are
imposed by Russells theory of types. So Ipropose to take it as a working hypothesis
that some statements involving quantification over all sets do make sense.
(3) Given that working hypothesis, we need some account of the sense that such
statements make, an account that is best provided by constructing a semantics for
the language of set theory. Tait does not provide such a semantics. If, however, we
parlay his ideas into a semantic theory for that language in the most natural way,
we find confirmation for his claim that x(x) x(x) is not a logical truth
when the variable x ranges over all sets.
On the view that Tait advances, our understanding of set theory relates to an
2
infinite sequence of axiom systems, each of which extends ZF . Let S be the space
of all such systems. Let us say that an axiom system t extends a system swritten
s tif and only if system t entails every proposition that s entails. We may postulate
that the relation of extending is a pre-order, that is, is reflexive and transitive. Moreover,
each axiom system s in the space is associated with a domain D(s) comprising those sets
whose existence is entailed by s; we may further postulate that D(s) D(t) whenever
s t. Finally, we say that an axiom system s forces a formula A of the language of set the-
orywritten s Awhen s semantically entails A. We postulate that, for any atomic
formula A, if s t and s A, then t A. We call an ordered quadruple <S, , , D()>
meeting these conditions a Kripke model for the language of (first-order) set theory if
the forcing of complex sentences is further constrained by the principles K1 toK6:
K5.s xA(x) if and only if for every t such that s t, t A(cz) for every z in D(t)
K6. s xA(x) if and only if for some z D(s), s A(cz).
(In K5 and K6, cz is an arbitrary parameter designating the object z.) Given K1 to
K6, it is straightforward to prove that, for any formula A (not just atoms), if s t
and s A, then t A.
Suppose we use this semantics to define a consequence relation whereby a
conclusion follows from some premisses just in case, for every axiom system s in
every Kripke model, s forces the conclusion if it forces all the premisses. That is,
X A if and only if, in every Kripke model <S, , , D()>, for all s S, s A if s B
for every B in X. Then, as Kripke (1965) proved, first-order intuitionistic predicate
logic is sound and complete with respect to this notion of consequence:a conclu-
sion follows from some premisses in this sense if and only if it is deducible from
the premisses using the rules of the first-order intuitionistic predicate calculus.
In particular, the semantics vindicates Taits claim that x(x) x(x)
is not a logical truth. There may well be an axiom system which does not force
x(x) because it does not entail that some set is not but which equally
does not force x(x) because it has an extension whose domain contains
a set that is not . Given K2, such an axiom system will not force the disjunctive
formula x(x) x(x) .
It may be objected that this deviation from classical logic arises from the
asymmetric semantic treatment of the quantifiers. According to K5, in assessing
whether an axiom system s forces a universally quantified formula, we need to take
into account all possible extensions of s. According to K6, by contrast, in assess-
ing whether s forces an existentially quantified formula, only s itself is relevant.
This asymmetry is, indeed, crucial to the invalidity of x(x) x(x) . On
the present conception of sets, though, the asymmetry is entirely principled. If
we are to avoid having to retract a universally quantified claim as the domain of
sets expands, we must take account of the possible ways in which the set-theoretic
axioms might be strengthened. On the other hand, to say that x(x) is true
when is true of some set in the entire set-theoretic universe would fail to assign
well-defined truth-conditions to existential claims, for on the view we are consid-
12
ering the set-theoretic universe is not a well-defined collection. Nor is there any
mileage in emending K5 so that it matches K6, i.e., so that itreads
12
The emended clause for the existential quantifier is one that Paseau (2003, 392)proposes in his
reply to Lears attack on the use of classical logic in set theory. Paseaus reply is well taken against
Lear, who assumes that there is a well-defined universe of sets. Since, however, Tait rejects that
assumption, the emended clause cannot be prayed in aid against his position.
On the Use of Classical Logic in Set Theory 281
For, under the emended definition, we might have s xA(x), s t, but t xA(x).
Such a failure of monotonicity would be tolerable only if the meaning of set were
to change as we shift from accepting the axiom system s to accepting the system t.
Often, revisions to the axiomatic theory that implicitly defines a theoretical term
will result in a change in the terms sense, but our present problems arise precisely
from the fact that the iterative conception of set cannot be encapsulated in any
single theory. Under that conception, then, the sense of set remains constant as
we adopt successively stronger axioms that make more precise the notion of as far
as possible. It is this combination of constancy in sense with the use of stronger
and stronger axioms that undermines the soundness of the classical rules when
these are applied to reasoning about all sets.
Whether this semantics yields a coherent account of use of the lan-
guage of set theory is doubtful. Ihave been taking the axiom systems in the
2
underlying space to be extensions of ZF . The semantic theory given above
only covers a first-order language and it is far from clear how to extend it
to apply to second-order languages. In any case, assuming it is intuitionis-
tic logic that the semantics validates, we shall need to make that logic the
standard for assessing which theorems the set-theoretic axioms entail, i.e.,
force. As Myhill showed, though, the standard formulation of the Axiom of
Foundation (alias the Axiom of Regularity) intuitionistically entails the Law
of Excluded Middle in the presence of the axioms of Zermelos original system
Z (see Myhill 1973). We could switch to an intuitionistic set theory in which
Foundation is replaced by a principle of transfinite induction on . However,
it is far from clear whether an intuitionist can make sense of the second-order
version of such a theory.
Some delicate problems, then, would need to be cleared up if the logic of set
theory were to be intuitionistic. Ishall return to some of these problems in 9.6.
My main aim, though, is to pursue a different line of inquiry. Given that we can-
not use classical semantics in set theory, might there nevertheless be a way of vin-
dicating the use of classical logic in that field?
that the extending relation should be directed:that is, given any two axiom sys-
tems in the space, there exists a third system that extends both. Alexander Paseau
(2001, 2003)has gone further, and suggested that the relation should be linear. Do
these proposals help?
Not much, Ifear. First, even if we were to impose these constraints on the extend-
ing relation, we would still fall short of classical logic. If extending is required to
be directed, then the Kripke semantics will validate the logic KC, in which the
so-called Weak Law of Excluded Middle, A A , is added to the intuition-
istic rules. If extending is required to be linear, we get the Gdel-Dummett logic
LC, which has among its validities all instances of (A B) (B A) . But even
LC falls short of classical logic, so substantial revisions to set-theorists deductive
practice will still be demanded.
More importantly, under the proposed interpretation of the semantic the-
ory, even the requirement of directedness is implausible. Hazen argues for this
requirement as follows. If directedness were to fail, he says, then one must sup-
pose that incompatible advances in our mathematical knowledge are possi-
ble:that it may, for example, be possible both to establish A as a theorem and to
establish not-A as a theorem, for some proposition A. It is unclear, he continues,
what sort of possibility is at issue here, but on at least some obvious construals,
both classical and intuitionistic, it is absurd to suppose that both the establish-
ment of A and the establishment of not-A are possible extensions of our math-
ematical knowledge (Hazen 1982, 130). No doubt that is right, but it does not speak
to the present case, which concerns different ways of making an inherently inde-
terminate notion more precise, and it is in fact very far from clear that all such
ways are compatible. Quite the contrary, in fact. As we saw in the previous sec-
2
tion, one reasonable way of extending ZF precludes any principle incompatible
with Gdels Axiom of Constructibility; another, also reasonable, postulates the
2
existence of a measurable cardinal. All the same, these two ways of extending ZF
are incompatible. We have good reason, then, to hold that the space of axiom sys-
tems is not directed and a fortiori not linear.
A different strategythis time, one for recovering the whole of classical logic
is suggested by Melvin Fittings monograph, Intuitionistic Logic, Model Theory
and Forcing (1969), where the Kripke semantics for a first-order intuitionis-
tic language is the basis for an elegant recasting of Paul Cohens proof that the
Continuum Hypothesis is independent of the axioms of first-order ZF (see Cohen
1966). Suppose we were to emend K5the semantic clause for the universal quan-
tifierso that x is taken to mean x. In other words, we replace K5by
K5. s xA(x) if and only if for every t such that s t there is a u such that t u
and u A(cz) for every z D(u).
On the Use of Classical Logic in Set Theory 283
The result is the system that Kit Fine called bastard intuitionism (Fine
1975, 127)and it has a nice property:A is a validity of classical first-order logic
if and only if A is a validity of the bastard intuitionistic system. (The cor-
responding result does not hold for the legitimate intuitionistic system:while
x(A(x) A(x)) is classically valid, its double negation x(A(x) A(x))
is not valid in legitimate first-order intuitionistic logic.)
This equivalence result for the bastard system was the basis for Fittings
recasting of Cohens independence proof. Given any first-order formula A,
there is a formula A* that is classically equivalent to A but which contains
no universal quantifiers: to form A*, we simply replace every occurrence of
in A with . Fitting called a Kripke model an intuitionistic ZF model
if classical equivalents of all the axioms of (first-order) ZF, expressed with-
out the use of the universal quantifier, are valid in it. To show the independ-
ence of the Continuum Hypothesis, he then constructed an intuitionistic ZF
model in which CH* is true (as above, CH* is some classical equivalent of
the Continuum Hypothesis, expressed without using the universal quantifier).
Now suppose that CH were deducible, in first-order classical logic, from the
axioms of ZF. Then for some finite collection A 1,...,A n of those axioms, we
should have
A 1,...,A n C CH,
and hence
13
Someone who is able to classify states of affairs into those that render [a declarative sen-
tence] correct and those that render it incorrect, may be said to know the assertoric content of
the sentence. It does not at all follow that he knows enough to determine its contribution to the
assertoric content of complex sentences of which it is a subsentence. What one has to know to
know that may be called its ingredient sense; and that may involve much more than its asser-
toric content. Ingredient sense is what semantic theories are concerned to explain [sc., specify]
(Dummett 1991, 48).
On the Use of Classical Logic in Set Theory 285
we reject the realist picture of set theory as describing a universe which exists
wholly independently of mathematical thinking. So far, our results have been
unremittingly negative. In the case of arithmetic, our inchoate conception of the
natural numbers was refined into a categorical theory; although not axiomatiz-
able, that theory is semantically complete, so we can ensure bivalence for the
language of arithmetic by taking an arithmetical statement to be true if it is a
consequence of the theory and false if its negation is a consequence. Because the
language of arithmetic is bivalent, the familiar truth-tables for the connectives
validate classical logic. Although, then, the classical logician may be unable to
persuade someone who uses a weaker logic in arithmetic to change her ways,
he has a story to tellto tell himself, at leastabout why the classical laws are
justified. For the reasons explained, however, even this degree of comfort is not
available in set theory. To the contrary:the semantic theories that best reflect the
indeterminate height of the set-theoretic universe seem only to validate weaker
than classical logics.
In resuming the quest for justification, it helps to limit our ambition. Ajusti-
fication for the use of classical logic in second-order set theory seems a very tall
order. It is hard enough to justify classical logic when we quantify over all sets; the
challenge is yet greater when we quantify over all properties of sets. For the rest of
this chapter, then, Ishall confine attention to first-order set theory. Higher-order
theories were of interest largely because they held out some promise of categoricity,
and thence of completeness. With that promise unfulfilled, it makes sense to focus
on the question whether the classical laws can be justified in developing first-order
set theory. Such a theory will have non-standard models, but the idea that we start
with an intuitive conception of the set-theoretic universe that can be parlayed into
a categorical or quasi-categorical theory seems hard to sustain in the light of our
discussion.
Just for that reason, indeed, we have to be prepared to adjust our conception
of what set theory includes. The idea that we have a conception of the cumulative
hierarchy in anything like the way we have a conception of the structure of the
natural numbers has been a casualty of the discussion:the attempt to make that
conception explicit brings in so many doubtful elements that the net effect is to
cast doubt on the clarity of the conception. According to Hilbert, taking the prin-
ciple of Excluded Middle away from the mathematician would be the same, say, as
prohibiting the astronomer from using the telescope or the boxer from using his
fists (Hilbert 1928, 476). We should take seriously the possibility that the price of
gratifying set theorists pugilistic instincts may be a weaker set theory than they
would like.
286 Five Attacks on Classical Logic
14
This refinement of Gdels 1933 translation originated in Gentzen 1933, a paper which Gentzen
withdrew from Mathematische Annalen at proof stage when he discovered that its main result had
been anticipated by Gdel. As Gdel acknowledged (1933a, 34=Gdel 1986, 2867), he was not the
first person to have had the idea of interpreting classical logic in intuitionistic logic via a nega-
tive translation. He cites Glivenko 1929 which in turn builds on Kolmogorov 1925 (see also Hao
Wangs introduction to the translation of the latter paper in van Heijenoort 1967). As Gdel notes,
though (in his n.4), the Kolmogorov-Glivenko translation is not adequate for interpreting Peano
Arithmetic in Heyting Arithmetic. For another negative translation, see Kuroda 1951.
290 Five Attacks on Classical Logic
The system of (first-order) set theory whose background logic is first-order intui-
tionistic logic, and whose non-logical principles are the axioms (or schemas) of
Extensionality, Unordered Pair, Union, Infinity, Separation, Replacement and
Power Set, along with the -induction rule, is known as ZFI. ZFI has both the
disjunction and existence properties (Myhill 1973).
Although ZFI has these nice properties, it is, apparently, not strong enough
to validate Replacement in its natural models (see Grayson 1979, 410). Aclosely
related system, though, does not suffer from this drawback. In classical logic, the
Replacement schema is equivalent to the following schema of Collection (in which
a is a parameter ranging over an arbitrary set):
(I x I a)I y ( x , y ) I I b (I x I a)(I y I b) ( x , y ).
15
For a proof, see Grayson 1979, Theorem 2.2. It is not known whether every instance of Collection
+
is a theorem of ZFI. Nicolas Goodman, though, has investigated the properties of the theory ZFI
that is got by adding a single class parameter to the language of set theory and then extending all
the axiom schemas of ZFI to formulae in the new language. He shows that not every instance of
+
Collection in the extended language is a theorem of ZFI (Goodman 1985).
On the Use of Classical Logic in Set Theory 291
16
Infinity, Separation, Collection, or the -induction rule, A N is a theorem of ZF .
Not so, however, when A is an instance of either Extensionality or the Power Set
Axiom (see Friedman 1973, 315). The problem with these two axioms is evident.
Let A be the following instance of Extensionality:
a = b (a c b c)
x y (z(z y z a)) y x.
16
The proofs are routine. Take Unordered Pair, for example. This axiom may be formulated
as x (a x b x), where a and b are distinct parameters. So its Gdel-Gentzen translation is
II I x (II a I x I II b I x). Now an axiom of ZF is I x (a I x I b I x). Since A intuitionistically
entails II A, it follows that II I x (II a I x I II b I x) is a theorem of ZF , exactly as required.
292 Five Attacks on Classical Logic
inner model does not help with our problem. We are concerned to justify the use of
classical logic when reasoning about absolutely all sets. Ajustification of classical
ZF as a description of some particular inner model patently does not contribute to
17
that.
Is the difficulty over Extensionality insuperable, then? I think not. Indeed,
Ithink that the difficulty disappears as soon as we recall the context of the pre-
sent anti-classical argument. Where A is any instance of Extensionality, we can
prove A N very simply in the system that is got by adding Excluded Middle for all
atomic formulae to ZF . For, in this extended system, a =I b is provably equiva-
lent to II a =I b and a I b is provably equivalent to II a I b, so that A N,
which is II a =I b I (II a I c I II b I c), is provably equivalent to
a =I b I (a I c I b I c), which is an instance of Extensionality. Friedman
and Powell do not contemplate this extension of ZF , no doubt because they have
in mind a traditional interpretation of the language of intuitionistic logic under
which A A expresses the decidability of A. The atomic formulae of the lan-
guage of set theory (unlike the atomic formulae of the language of arithmetic) are
not in general decidable, so to add Excluded Middle for all atoms would be simply
illegitimate on the traditional intuitionist interpretation of the connectives. That
interpretation, though, is quite irrelevant to Taits anti-classical argument. Tait
is not arguing that classical logic cannot be applied to undecidable set-theoretic
statements. His claim, rather, is that it cannot be applied to statements exhibit-
ing the particular form of indeterminacy that arises when we quantify over all
setsgiven his master thesis that the universe of sets is not a mathematically
well-defined totality. On this conception of the matter, indeed, we have good rea-
son to postulate Excluded Middle for all atomic statements in the language of set
theory. Aset is a well-defined collection, so that whenever a and b designate
sets, the statements a=b and a b may be assumed, on Taits principles, to be
bivalent. We have, then, good reason to accept the additional logical principle
namely, Excluded Middle for all atomsthat is needed to justify A N, where A is an
arbitrary instance of Extensionality.
What, though, of the Power Set Axiom? Here, Ithink, someone who is sym-
pathetic to the thesis that the totality of sets is not well-defined ought to con-
cede that we have no reason to think that the Axiom is true. Given any set a,
the Power Set Axiom tells us that there is a distinct set b whose members are
precisely the subsets of a. Now a subset of a is a set, each of whose members is a
17
For a general investigation of which parts of classical mathematics can be justified in this way
(albeit under Kolmogorovs original negative translation rather than the Gdel-Gentzen transla-
tion), see Leivant 1985.
On the Use of Classical Logic in Set Theory 293
member of a. Taits master thesis is that the universe of sets is not a mathemati-
cally well-defined collection. But if the universe of sets is not well-defined, then
we have no reason to assume that there is always a well-defined collection of
sets, each of whose members is a subset of a. That collection, however, would
need to be well-defined, if it is to form a set. Accordingly, scepticism about the
determinacy of the universe of sets engenders a corresponding scepticism about
the Power Set Axiom. Indeed, it undermines even Friedmans weaker form of the
Axiom (Friedman 1973, 316). Using y a to abbreviate z (z y z a), the
strong form of the Axiomis
x y ( y a y x )
quantifier is bounded (so any atomic formula is 0, vacuously). The idea underly-
ing Taits challenge to classical logic is that, while formulae involving unrestricted
quantification over sets may be indeterminate, formulae involving only bounded
quantification express determinate conditions. Thus, in place of the unrestricted
Separation schema, we may propose
0 - Separation : x (x = { y a : ( y )})
18
Saul Kripke and Richard Platek independently identified this set theory as an object of interest
in the 1960s; see the abstracts that compose Kripke 1964, and Plateks Ph.D.dissertation at Stanford
(Platek 1966). The theory came to wider notice following the publication of Barwise 1975.
19
i.e., the operation that maps any set x to x {x}.
On the Use of Classical Logic in Set Theory 295
In fact it helps considerably. Suppose we adopt the semantic theory of 9.3 for
20
the language of set theory. Then we shall still need to correct for the fact that
the principles K2 and K6 misdescribe our actual use of and . Ihave sug-
gested that the best way to correct for this is to translate formulae involving
and into an auxiliary language whose semantics is given by K1 to K6. That
way, we get a better account of the actual use of the connectives while respect-
ing the underlying thesis that the universe of sets is not a well-defined totality.
The translation that yields the best account of the senses of the connectives is the
Gdel-Gentzen mapping N. That is why our problem has been to show that the
Gdel-Gentzen translation of each axiom of a classically acceptable set theory
TC is a theorem of some intuitionistically acceptable set theory TI. We initially
supposed that TC would be ZF. Our analysis of the problem, however, has shown
that ZF is too strong, given the master thesis that the universe is not well-defined,
and Ihave been arguing that a far better candidate for TC is KP. In the light of
these developments, then, we need to reconceive our problem. The question now
is whether the Gdel-Gentzen translation of each axiom of KP is a theorem of
an intuitionistically acceptable set theory. We may assume that such a theory will
include all the axioms of IKPthe system whose axioms are precisely those of
KP, but whose logic is intuitionistic rather than classical.
As with ZF, when A is an instance of Unordered Pair, Union, or Infinity,
it is straightforward to verify that A N is a theorem of IKP. Since KP lacks
it, we do not have to worry about the Power Set Axiom. What, though, about
Extensionality, which gave us trouble earlier? Iobserved that the difficulty would
vanish if our system included Excluded Middle for atoms, and our analysis justi-
fies that principle. In fact, it justifies something stronger. As a number of writers
have noted, there is a deep affinity between the operations of set-formation and
of classical quantification. As Stewart Shapiro and Crispin Wright develop the
point, we may think of a set as the value of a many-one function that takes exactly
the elements of the set as arguments ... [and] think of [classical] quantification
in similar terms, as a many-one function that yields a truth-value when given a
range of instances as argument (Shapiro and Wright 2006, 296). This gives us a
justification for using classical logic when reasoning with statements all of whose
quantifiers are bounded. Accordingly, we may lay down as an additional logical
principle the following restricted form of Excluded Middle:
20
Of course, the underlying space S of axiom systems will now comprise extensions of KP, not
extensions of ZF.
On the Use of Classical Logic in Set Theory 297
Since any atomic formula is 0, 0-LEM entails that any atomic formula
is equivalent to its double negation, so that whenever A is an instance of
Extensionality, A N is provable in IKP + 0-LEM. In fact, it is easy to verify
that, for every 0 formula , is provably equivalent to N in IKP + 0-LEM.
This in turn means that, whenever A is an instance of 0-Separation, A N is a
theorem of IKP + 0-LEM. Moreover, the N translation of any instance of the
-induction rule is itself an instance of that very rule, so there is no difficulty
there.
What, though, about the schema of 0-Collection? Let A be
(x a)y (x , y ) b (x a)(y b) (x , y )
where (x, y) is 0. Then A N is
Now since N maps bounded quantifiers to other bounded quantifiers, N (x, y) will
also be a 0 formula. So, while A N is not always a theorem of IKP + 0-LEM, it
is always a theorem of the strengthened system IKP + 0-LEM + 0-MP, where
0-MP is the following restricted form of Markovs Principle:
The same reasoning that justifies 0-LEM also justifies taking 0-MP as part
of our intuitionistically acceptable set theory, and that means we are done:the
Gdel-Gentzen translation of every axiom of KP is a theorem of IKP +
0-LEM + 0-MP.
The observation that KP can be interpreted in IKP + 0-LEM + 0-MP
is due to Solomon Feferman, although (like Friedman etal.) he makes it in the
course of determining the relative consistency of the classical and constructive
versions of KP, rather than as part of a justification of the classical formulation
21
of the theory (see Feferman 2010, 120). In fact, Fefermans preferred set theory
is a stronger system that he calls the semi-constructive theory of sets, SCS. This is
got by reinforcing IKP + 0-LEM + 0-MP with all the instances of two further
schemas. The first is the Bounded Omniscience Schema:
21
Jeremy Avigad (2000, 45) also explores negative translations of classical KP into an intui-
tionistic version of that theory. Following Friedmans discussion of ZF , however, Avigad focuses
on an intensional version of the theory that omits Extensionality.
298 Five Attacks on Classical Logic
where (x) is any formula in the language, not necessarily 0. Our discussion
of Tait has brought out how x(x) x(x) expresses the determinacy of
the domain of quantificationassuming it is determinate whether an arbitrary
member of that domain is . Accordingly, what BOS says is that any set is a deter-
22
minate domain of quantification. The second schema is a form of the Axiom of
Choice, restricted to sets:
Again, (x, y) is an arbitrary formula, not necessarily 0; and Fun (r) and
Dom (r)=z abbreviate formulae in the language of set theory (which are in
fact 0) which say, respectively, that r is a function, and that z is the domain of
r. These two further schemas are highly plausible from a constructive viewpoint
and are helpful in many applications of set theory, but we need not discuss them
here:since KP is interpretable in IKP + 0-LEM + 0-MP, it is a fortiori inter-
pretable in any stronger system, such as SCS.
The fact that the underlying logic is intuitionistic means that IKP + 0-LEM +
0-MP does not contain arbitrary formulae of the form x(x) x(x)
where the variables range over the whole set-theoretic universe. However, the
Gdel-Gentzen translation of any such formulabeing the translation of a
classical validityis a theorem of first-order intuitionistic logic and hence of
IKP + 0-LEM + 0-MP. Once we correct for K2 and K6s misdescription of the
actual use of and , then, we can justify the use of classical logic in the strong-
est set theory that is justified given Taits master thesis, even when reasoning about
absolutely all sets. The upshot of Taits critique, then, is not that classical logic is
unjustified when reasoning about sets. Rather, it is that ZF is too strong a set theory.
When we retreat to the strongest set theory that can be justified on Taits princi-
ples, namely, KP, or perhaps Fefermans SCS, we find that the classical version of
this theory can be justified by way of a negative translation into an intuitionistically
acceptable version of the theory.
I said at the end of 9.4 that the price of rehabilitating the use of Excluded
Middle when quantifying unrestrictedly over sets might be a forced retreat to
a set theory weaker than mathematicians are accustomed to having. So it has
proved. One of the motivations behind Cantors and Dedekinds pioneering
explorations in the field was the ambition to give real analysis a new foundation
22
For this reason, the name Bounded Omniscience Schema is somewhat unfortunate.
As Feferman observes, the schema generalizes the Numerical Omniscience Scheme of
semi-constructive theories of arithmetic. The present context, though, is far removed from ques-
tions of decidability or knowability.
On the Use of Classical Logic in Set Theory 299
statement has in the end to be cashed out by reference to what we can do with
it, and what we do with a piece of mathematics is to apply it. It is on this sort of
ground that we might best hope to rehabilitate the claim that there is a set of real
numbers.
This conception of the meaning of set seems to go with the following sort
of semantic theory. Let S now be a space comprising possible applications
of mathematics: for the present, we adjourn consideration of the interest-
ing metaphysical question, what sort of things applications are. We now take
s A to mean The application s requires the truth of the set-theoretic state-
ment A. The Kripke axiom for , K1, remains plausible on this new inter-
pretation:an application requires the truth of a conjunctive statement just in
case it requires the truth of both conjuncts. K2, though, the axiom for , is
clearly wrong:an application may require the truth of a disjunctive statement
without requiring the truth of either disjunct. Similarly, an application may
require the truth of an existentially quantified statement without requiring
any of its instances.
How should we handle these connectives? We find an answer by adapting some
of the ideas developed earlier in this book, in particular, the pragmatist seman-
tic theory of 4.5. Let us say that an application s belongs to the closure Cl (U)
of a set of applications U if s requiresi.e., requires the truth ofany statement
whose truth is required by all the members of U. It follows immediately from
the definition that the closure operation is INCREASING, IDEMPOTENT, and
MONOTONE. Moreover, since A B is the logically weakest statement to fol-
low from both A and B severally, we have
Similarly we have
Given the properties of the incompatibility relation, , it is easy to show that any
set is a subset of its double orthocomplement:U U. On the present seman-
tics, however, we also have the converse inclusion whenever U comprises all and
only the applications that require a given statement. For let U comprise precisely
the applications that require a given statement A; and suppose that application s
belongs to the double orthocomplement of U. That means that s is incompatible
with its not being the case that A. But that implies that s does require the truth
of A. Whenever U is the set of the applications that require the truth of a given
statement, then, U is identical with its own double orthcomplement. Given K3,
it follows that any statement is equivalent to its own double negation. In fact, the
present semantics validates all the rules of the classical predicate calculus. What
vindicates that logic is the injection of a pragmatist elementin this case, the
applications of mathematicsinto a broadly constructivist semantics. In other
words, the vindication has essentially the same form as that given to justify the
use of classical logic in empirical discourse in Chapter4, where a different prag-
matist elementin that case, the impact of accepting a statement on the accept-
ers intentional behaviourwas injected into an anti-realist semantics to yield a
theory that validates classical logic.
In order to be convincing, this more general vindication of the use of classi-
cal logic in set theory would need a great deal of development. In particular, we
would need to say more fully what an application of a mathematical theory is. The
philosophy of mathematics has yet to produce a satisfactory account of this, and
filling the gap would take a book to itself. It is clear, though, that this justification
of the use of classical logic in set theory, like the more local justification given in
the previous section, will be utterly different from that offered by a strong realist,
who believes that each set-theoretic statement is either rendered true, or rendered
false, by a Platonic realm of sets which exist entirely independently of human cog-
nition. While the parties converge on classical logic, their accounts of why that
logic is the right one to use in this area are deeply antithetical. The pressing philo-
sophical question in this area, it seems, is not whether classical logic is the right
logic to use in set theory, but rather, why it is right. It is in answering this question
that the conflict between a strong realism about sets, and the forms of construc-
tivism and pragmatism that we have been exploring, will have to be fought out.
10
Conclusion
A theme of this book has been to vindicate classical logic without appealing to the
Principle of Bivalence. There is, though, an ancient argument which purports to
show that, however classical logic is defended, adopting it commits one to accept-
ing the Principle. This argument has been, albeit indirectly, a fertile source of doubt
about classical logic:it appears to require only innocuous principles about truth
and falsehood, so doubts about its conclusion transmit directly back to the underly-
ing logic. Statements containing vague predicates, and certain set-theoretic state-
ments, are cases in which the conclusion is indeed doubtful. Where the paint in the
transparent tube a50 is of a borderline red-orange colour, many people are reluctant
to accept that the statement Tube a50 is red is either true or false. Again, as we shall
see, many mathematicians are reluctant to accept the bivalence of, for example,
Cantors Generalized Continuum Hypothesis. It is therefore of some interest that
the analysis of the previous two chapters reveals the flaw in this ancient argument
for Bivalence when it is applied both to vague statements and to statements that
quantify unrestrictedly over sets.
Of one subject we must either affirm or deny any one predicate. This is clear, in the first
place, if we define what the true and the false are. To say of what is that it is not, or of what
is not that it is, is false, while to say of what is that it is, and of what is not that it is not, is
Conclusion 303
true; so that he who says of anything that it is, or that it is not, will say either what is true or
what is false. (Metaphysics 1011b25; cf. Plato, Cratylus 385b2)
In spelling out Aristotles argument, Ishall use u as a variable ranging over state-
ments and P as a schematic letter, to be replaced by a complete statement. As
ever, a statement is an utterance or inscription that expresses a complete thought,
so any statement will say that P, for some suitable specification of P. For any
statement u, then, we have the following premiss:
Given that starting point, we may set out what Ishall call the Simple Argument for
Bivalence:
formalizes the sufficient condition for truth to which Aristotle appeals:to say of
what is that it is, is truei.e., is to speak truly. Similarly, the schema instantiated
at line (10), namely,
formalizes Aristotles sufficient condition for falsity:to say of what is not that it
is, is to speak falsely. The Simple Argument, then, formalizes the reasoning in the
passage quoted from Aristotle.
304 Conclusion
What sustains the two schemata to which the Simple Argument appeals? In a
book On Truth left incomplete at his death, F.P. Ramsey proposed definitions of
truth and falsity which, when applied to statements, come to this:
(T) u(True (u) P(Say (u, P) P))
and
(Ramsey 1991, 9, 15). (T) and (F) are formulated in a first-order formalized lan-
guage that has been enriched with quantifiers whose attendant variables replace
complete well-formed formulae (see further Prior 1971). It is assumed that the
introduction and elimination rules for these sentential quantifiers are precise
analogues of the rules that govern the corresponding quantifiers into name posi-
tion, aside from differences consequential upon the different syntactic categories
of the associated variables. Given that the underlying propositional logic is clas-
sical, the formulae P(P P) and P(P P) will then be theorems of the
extended system. Moreover, the truth of each instance of (TS) then follows from
(T) and the truth of each instance of (FS) follows from (F). We may follow Prior in
reading the existential sentential quantifier P as saying There is a way things
might be said or thought to be. On this reading, (T) says that a statement is true if
and only if there is a way it says that things are and they are thus, and (F) says that
a statement is false if and only if there is a way it says that things are and they are
not thus.
So interpreted, (T) and (F) have great intuitive plausibility, and the theory
whose only non-logical axioms they are is demonstrably consistent. As Timothy
Williamson has observed (1998, 14), we can show this by constructing an unin-
tended model in which formulae are treated as referring to truth-values, in which
the sentential quantifiers range over truth-values, and in which all formulae of the
forms Say (u, P), True (u), and False (u) are counted as false. Many philosophers
regard (T) and (F) as tainted because of the role they play in versions of the Liar
Paradox. The consistency of this theory suggests, to the contrary, that the antin-
omy arises from the assumption that an utterance of (e.g.) This utterance does not
express a truth says somethingin the sense of says that figures in (T) and (F).
On this view, the semantic paradoxes are transformed into sound arguments for
constraints on what can say what in what contexts (Williamson 1998,15).
Although the definitions of truth and falsity to which it appeals are consist-
ent, the Simple Argument is really too simplistic to capture Aristotles consid-
ered view of truth and falsehood. The Principle of Bivalence is the thesis that
Conclusion 305
every statement is either true or false, but the or here is exclusive, not the logi-
cians vel. That is, Aristotles Principle might more explicitly be formulated as
Every statement is either true or false but not both. As they stand, though,
(T) and (F) do nothing to exclude the possibility that a single statement should
be both true and false. Astatement might be true by virtue of its saying that
P, in a circumstance where P, and also false by virtue of its saying that Q, in
a circumstance where not Q. What would exclude that possibility is the fur-
ther assumption that the statement is determinate. Any statement, we said,
expresses a complete thought. If a statement is determinate, it will express just
one such thought:there will be such a thing as the thought it expresses, at least
up to material equivalence. If u is determinate, then, it will meet the following
condition:
Given (T) and (F), and the premiss that u satisfies (Det), it is straightforward to
show that u cannot be both true and false.
How might a statement fail to be determinate? How could an utterance say
something without there being just one thing that it says? We should set aside
a shallow sense in which this is possible. If Isay It is wet and cold, Imay be
correctly reported as having said that it is wet, and as having said that it is cold.
In the intended sense, though, Say (u, P) means What the whole of u says is
that P; a specification of the content of a proper part of the utteranceeven a
proper part that follows logically from the wholeis beside the point. In the
intended sense of Say (u, P), though, there are still cases in which it is prima
facie plausible to claim that (Det) fails. Stephen Read (2009) has revived Thomas
Bradwardines idea that, as well as conveying the content literally expressed by
its component words, some statements express their own truth:over and above
the content the statement expresses au pied de la lettre, it says of itself that it is
true. Bradwardine thought that certain paradoxical utterancese.g. an utter-
ance, , of This utterance does not express a truthsignify multiply in this
way and, on this view, there will be counterexamples to (Det). What literally
expresses is the thought that does not express a truth; but also expresses
the thought that does express a truth. These thoughts are not materially
equivalent.
Another possible counterexample to (Det) takes us back to one of the main
themes of the previous chapter. Suppose one is not convinced by the devices
deployed there to rehabilitate the use of classical logic for statements involv-
ing unrestricted quantification over sets. Then another way to assign a classical
306 Conclusion
Similarly, for a statement to be accounted false, it is not enough that one thought
it expresses should fail to be the case. We shall only count the statement as false if,
however it says that things are, thus they are not, so that (F) needs to be replaced
by (F*):
The theory whose axioms are (T*) and (F*) in a logical system permitting
quantification into sentence position is again consistent. Indeed, this theory is
consistent even if every utterance in the domain of quantification says some-
thing. This may be shown by constructing an unintended interpretation in
which each utterance both says that P and says that not P, so that every formula
P Say (u, P) is true, while every formula in the forms True (u) or False (u) is
false. In this new theory, the semantic paradoxes are transformed into sound
arguments showing that certain utterances express no unique thought. Thus,
when applied to our paradoxical utterance, , the usual reasoning shows that
does not express a truth, but we cannot conclude from this that expresses a
Conclusion 307
truth after all. For while explicitly says that does not express a truthwhich
is the caseit also expresses the thought that does express a truth, which is not
the case, so things are not wholly as says they are. Bradwardines conclusion
that expresses no unique thoughtmay be easier to swallow than the claim
that it says nothing whatever.
Unlike (T), (T*) does not entail the truth of an arbitrary instance of the (TS)
schema. Similarly, (F*) does not guarantee the truth of every instance of (FS).
Given the revised accounts of truth and falsity, then, the Simple Argument
breaks down:it does not even show that an arbitrary statement is true vel false.
Arepair is possible, though. When we took truth and falsity to be defined by (T)
and (F), we needed the assumption that u is determinate in order to show that
u is not both true and false. Under the revised accounts of truth and falsity, the
assumption of determinacy suffices to show that u is true vel false. For, given the
premisses that u is a statement and that is determinate, we can argue as follows:
(I have elided some elementary logical steps at lines (5), (8), and (14).) The principle
about truth that is used at line (10) follows directly from (T*), and that concerning
falsity used at line (16) follows from (F*). The conclusion of this Revised Argument,
308 Conclusion
1
Miroslava Andjelkovi and Timothy Williamson (2000) claim to deduce the thesis that every
statement satisfies (Det) from (T*). If their argument worked, (T*) would entail (T) and (F*) would
entail (F); moreover, the Simple Argument would apply to every statement and we would have no
need for the Revised Argument. See, however, Rumfitt 2015 for critical analysis of their argument.
Conclusion 309
In 8.7, Iidentified a flaw in the natural argument for the thesis that a true dis-
junction contains a true disjunct, and the present argument for us bivalence
involves essentially the same mistake. The problem lies in the conditionals about
truth and falsehood that are taken as premisses at lines (5)and (10). As explained
in Chapter8, if the vague terms true and false are related to a system of poles
that includes the Indeterminate as well as the True and the False, each of these
conditionals has an indeterminate antecedent and a false consequent and is con-
sequently unacceptable. Admittedly, if the only poles in the relevant system are
the True and the False, then Bivalence holds. Both these poles will be maximally
close to a vague statement, so every statement will belong to the interior of the clo-
sure of the union of the extension of true with the extension of false, and hence
(given the recommended semantics) will satisfy is either true or false. However,
to assume that the True and the False are the only poles in the relevant system is to
assume what was to be proved. So the present argument for the bivalence of vague
statements begs the question.
Line (5) follows from Ramseys putative definition of truth, (T), and line
(10) follows from his putative definition of falsehood, (F). What this shows is
that (T) and (F) can serve as definitions of truth and falsehood only in a rather
restricted sense. They succeed in articulating the conditions for a statement
to be a clear case of truth, and a clear case of falsehood, but definiendum and
definiens do not match up when matters are unclear. When a statement is a
clear instance of saying that such-and-such, and when it is clearly the case that
such-and-such, the statement is a clear case of truth. And when a statement is a
clear instance of saying that such-and-such, and when it is clearly not the case
that such-and-such, the statement is a clear case of falsehood. As we have seen,
though, neither of these principles holds when clear is replaced by borderline.
More precisely, neither holds when the poles by reference to which a statements
alethic status is assessed include the Indeterminate as well as the True and the
False.
What about deductions involving unrestricted quantification over sets? Before
venturing a diagnosis of the Simple Arguments failure in these cases, it helps to
stand back and reflect on the sources of resistance to its conclusion. When Cantor
hypothesized There is no set strictly intermediate in cardinality between the
integers and the real numbers (the continuum), it seems he succeeded in express-
ing a thought. So, where u is an inscription of this hypothesishis celebrated
Continuum Hypothesis, CH for shortit seems we may assert the premiss of the
Simple Argument, namely,
2
As remarked in the previous section, Itake seriously the possibility that the indeterminacy of
the set-theoretic universe means that some set-theoretic statements express more than one thought.
In that case, the Revised Argument will apply and its conclusion will not entail the bivalence of the
statements in question. My present contention, though, is that even when it is assumed that a given
set-theoretic statement is determinate, the Simple Argument still does not establish its bivalence.
Conclusion 311
His last published essay on the subject (Gdel 1964), completed after Cohen had
briefed him on his independence proof, is typical. Like most mathematicians,
Gdel regarded the nineteenth-century demonstration of the independence of
Euclids fifth postulate (the axiom of parallels) from the other axioms of classical
geometry as showing that the question of the postulates truth or falsity is mean-
ingless unless some extra-axiomatic sense is attached to the notion of a straight
line. Accordingly, Gdel points to disanalogies between the independence of
the fifth postulate and that of CH (Gdel 1964, 270). He also observes that, even
though CH is not decided by the axioms of ZFC, there remains a chance that the
mathematical community will one day come to accept additional set-theoretic
axioms that decide it (271). Whilst well taken, these observations are evidently
less than conclusive, and what is striking is that Gdel nowhere tries to close the
discussion by appealing to the Simple Argument. The explanation is surely not
that he overlooked it. Rather, the suspicion must be that he would have regarded
the Argument as too cheap to establish its conclusion. The claim that CH has a
truth-value is a deep set-theoretic claim. It is hard to credit that so weighty a con-
clusion can be proved by the jejune logical moves of the Simple Argument.
Even those who accept its conclusion, then, are apt to find something fishy
about the Simple Argument when it is applied to such statements as CH or
GCH. On the other hand, those who are not persuaded by these instances of the
Argument may be hard pressed to say where they go wrong. Gdel is again an
instructive case. The most obvious way to escape from the Simple Argument is
to follow the intuitionists and revise the logic so that one is not entitled to assert
A A in a case where A is undecidable. Gdel, though, was highly critical of
the suggestion that intuitionistic logic codifies the standards for correct deduc-
tive reasoning in set theory:he rejected intuitionism as utterly destructive in its
results (Gdel 1964, 261).
The mainstream intuitionists diagnosis of the flaw in the Simple Argument
certainly forces a radical deviation from accepted mathematics. Mainstream
intuitionists understand the connectives and quantifiers in conformity with
Heytings semantics. According to that semantics, we may assert A A only
when A is decidable. On this view, the attempt to apply the Simple Argument
to show that CH is bivalent does not get started. Since no one today is entitled
to assert that the continuum problem is solvable, no one is entitled to assert
CH CH , so we cannot affirm even line (2)of the Argument. As we have seen,
though, this analysis does not capture the nature of mathematical doubts about
the bivalence of statements like CH. For all we now know, Goldbachs Conjecture
(GC) is also undecidable; so, under the Heyting semantics, we are also not enti-
tled to assert GC GC . Most mathematicians, however, do take themselves to
312 Conclusion
be entitled to assert this; indeed, they take GC to be bivalent (see 7.3 and 9.1). An
intuitionist operating under the Heyting semantics, then, simply misdescribes
the doubts of most mathematicians who are sceptical about CHs bivalence.
According to such an intuitionist, we cannot assert that CH is bivalent because
we cannot assert that CH is decidable. The mathematical majority, though, take
themselves to be entitled to assert that GC is bivalent even though they cannot
assert that GC is decidable. If CH is not bivalent, its being so does not lie simply in
its being undecidable.
It is of some interest, then, that the semantics for the language of set theory pro-
posed in the previous chapter explains why the Simple Argument does not gener-
ally go through for set-theoretic statements, and does this without deviating from
the use of classical logic in set theory and without imposing any restrictions on
the definitions (T) and (F) of truth and falsity as these apply to set-theoretic state-
ments. According to that semantics, arguments involving such statements are to
be assessed by way of a negative translation into an auxiliary intuitionistic sys-
tem. We justified the classical versions of certain set-theoretic axioms by showing
that their negative translations follow from an acceptable set theory formu-
lated in the auxiliary system. The novel elements introduced by our arguments
for bivalence are the notions of saying, of truth, and of falsity; in the case of the
Simple Argument, the axioms that regulate these new notions are (T) and (F). We
can certainly add (T) and (F) to our auxiliary intuitionistic system. The Simple
Argument, though, involves an appeal to Excluded Middle, so it is not as it stands
a valid argument in that system. On the account proposed in Chapter9, its con-
clusion will follow from its premisses only if the auxiliary theory IKP + (T) + (F)
intuitionistically entails the negative translations of both (T) and (F).
What are those translations? Each formula True (u) and False (u) is atomic;
for present purposes, we may also classify as such each statement in the form
Say (u, P). Accordingly, the negative translations of True (u), False (u), and
Say (u, P) are True (u), False (u), and Say (u, P). Thus the negative
translation of axiom (T) willbe
The question we have to decide is whether (T)N and (F)N follow intuitionistically
from IKP + (T) + (F).
Conclusion 313
The crux is the treatment of the new atoms. Where A is any axiom of IKP,
we were able to show that A N follows intuitionistically from IKP by arguing
that the operative conception of sets entitled us to affirm the equivalence of each
atom with its double negation. That is, we were able to affirm the equivalence of
s=t with (s=t) and of s t with (s t). In the present case, a sim-
ilar argument establishes the equivalence of Say (u, P) with Say (u, P). In
applying the Simple Argument at all, we are assuming that any indeterminacy
there may be in the height or the width of the hierarchy of sets does not render
it indeterminate what a set-theoretic statement says. This assumption could cer-
tainly be challenged, but a successful challenge would force us to switch to the
Revised Argument, whose conclusionthat every determinate statement is biva-
lentis entirely consistent with the existence of non-bivalent, but indeterminate,
set-theoretic statements. For present purposes, then, we may assume that all the
utterances over which the variable u ranges conform to the principle
This, of course, entails the equivalence of Say (u, P) with Say (u, P) even
when the underlying logic is intuitionistic.
When it comes to atoms in the form True (u) and False (u), however, matters
are very different. Given the indeterminate height of the hierarchy, there is no rea-
son to affirm, of an arbitrary set-theoretic statement, that it is either true or not,
nor is there reason to affirm that it is either false or not. There is no reason, then,
to hold that True (u) is invariably equivalent to True (u) nor that False (u) is
always equivalent to False (u). We would need these equivalences, however,
if (T)N and (F)N were to follow intuitionistically from IKP + (T) + (F). On the
account of the meaning of set-theoretic statements recommended in Chapter9,
then, the conclusion of a set-theoretic instance of the Simple Argument does not
always follow from the premisses of that instance. The Simple Argument fails to
establish the bivalence of arbitrary statements of set theory.
On the view recommended here, then, set theorists and philosophers of set
theory are right to disdain the Simple Argument as too cheap to establish the
bivalence of statements whose status as such is seriously contested. The case for
this conclusion may be reinforced by comparing the Simple Argument with an
argument for the bivalence of the Continuum Hypothesis that everyone does take
seriously even if they do not in the end accept it. Ihave in mind the argument
put forward by Georg Kreisel in his essay Informal Rigour and Completeness
Proofs (Kreisel 1967). The proofs of Gdel and Cohen show that neither CH nor
CH follows from, or is deducible from, first-order ZFC. Kreisel, though, claimed
314 Conclusion
which is expressed by means of quantifiers over C+2 . As Zermelo pointed out, if we use
the current set-theoretic definition Z (x) of the cumulative hierarchy, in any model of Z,
this formula Z defines a C for a limit ordinal > . Consequently we have
(Z 2 CH) (Z 2 CH). (Kreisel 1967, 878)
3
If Kreisels argument worked, it would confirm Gdels claim that there is a fundamental dis-
analogy between the independence of CH from ZFC and the independence of the fifth postulate
from the other axioms of Euclidean geometry. For there are models of second-order geometry in
which the fifth postulate holds, and other such models in which it does not hold. See Kreisel 1967,
88.
4
Kreisel gives his argument for Zermelos original (1908) axiom system Z, which lacks the
Axiom of Replacement, rather than for the more familiar system ZF, which includes it. (Kreisel
clearly uses Z to stand for Zermelos original system rather than for ZF itself. He observes
(1967, 88)that while CH is determined by the second-order axioms of Z, Replacement is not; this
2
would make no sense if Z meant ZF.) This renders his argument problematical, for Z is not even
2
quasi-categorical. Indeed, in the context of Z , standard formulations of the axiom of infinity turn
2
out to be non-equivalent, and some interpretations of Z + Infinity have non-well-founded models
(see Uzquiano 1999). Acharitable exposition of Kreisels argument cuts through these complexities
by taking it to apply to ZF. As both Fraenkel and Skolem pointed out, Zermelos original axiomati-
zation fails to provide an adequate theory of the ordinals, so set theorists had good reason to switch
from Z to ZF. (In this footnote, Iam indebted to Christopher Scambler.)
Conclusion 315
go through for every closed formula of the language of set theory:as we saw in
Chapter9, there are closed sentences in that language which hold in some of mod-
2
els of ZF but not in others. Kreisels insight, though, was to recognize that CH is
not subject to this variation between models. To put his point in a more up-to-
date notation, CH may be formulated using quantifiers that range only over sets
that lie at levels V+2 and below in the hierarchy. +2 lies far below the first inac-
2
cessible 1, so all the models of ZF , when restricted to those levels, are isomorphic.
The argument that closed sentences hold in the same models goes through, then,
for the particular case of CH, so that wehave
2
(1) For any models M and N of ZF , M2 CH if and only if N2 CH.
A principle of classical model theory says that, for any model, any closed sentence
is either true in the model or false in it; so, in particular,
2
(2) For any model M of ZF , either M2 CH or M2 CH.
As in 9.1, (1)and (2)entail
2 2
(3) Either ZF 2 CH or ZF 2 CH,
which is what Kreisel should have written in the last line quoted from him. On a struc-
turalist view of set theory, a set-theoretic statement is true if it holds in every model of
set theory and false if its negation holds in every such model. Given that, we caninfer
(4) Either CH is true or CH is false,
i.e., that CH is bivalent. Second-order ZF, then, determines the truth or falsity
of CH, even though the incompleteness of the second-order formalization may
leave us unable, even in principle, to discover which truth-value it has.
Kreisels argument rests on particular features of CH; it turns on the fact that
CH quantifies only over sets that lie so low in the set-theoretic hierarchy that every
2
model of ZF will include them. Indeed, Kreisel recalls Gdel having been struck
by the fact that the argument does not extend to the Generalized Continuum
5
Hypothesis. Gdel, in other words, regarded it as an open possibility that, whilst
6
CH is bivalent, GCH is not. One readily sees the point of distinguishing between
5
Or so Daniel Isaacson reports Kreisel as remembering:Kreisel added that [in their conversa-
tions in the 1950s] Gdel had not noted that the second-order axioms [of ZF] do not determine the
generalized continuum hypothesis, which Kreisel pointed out to him, and which he found striking
(Isaacson 2011, 61). The programme Woodin proposed in his 2001 also fails to determine the size
of 2 for greater than zero. Gdels Axiom of Constructibility, by contrast, settles both CH and
GCH. Perhaps it is time to question the orthodoxy in the philosophy of set theory that the Axiom of
Constructibility is false.
6
Or, in the event that CH is false, some qualified version of GCH.
316 Conclusion
the two cases. Because even second-order ZF leaves the height of the set-theoretic
hierarchy undetermined, GCH might be left undetermined even though CH is
bivalent:GCH might hold up to the first strongly inaccessible cardinal and fail
2
thereafter, in which case GCH would be true in the smallest model of ZF but false
7
in larger ones. The Simple Argument, however, is blind to this difference:the
bivalence of GCH rolls off the assembly line like everything else. All the same,
it is not easy to say where the Simple Argument goes wrongespecially if one
takes the background logic to be classical, as Kreisels argument implicitly does,
notably at step (2). Our analysis explains why set theorists are entitled to reason
classically but are not entitled to appeal to the Simple Argument to establish biva-
lence across the board:in the case of a set-theoretic statement that is undecidable
from currently accepted axioms, we shall need a particular argument, tailored to
its specific content, to establish its bivalence. As Isay, this conclusion fits the way
set theorists and philosophers of set theory actually argue.
Kreisels argument for the bivalence of CH was controversial from its inception
and remains so. At the 1965 London conference on the philosophy of mathemat-
ics at which he presented it, it was applauded by Paul Bernays (1967) but attacked
by Lszl Kalmr (1967) and Andrzej Mostowski (1967). More recently, Solomon
Feferman (2009 and forthcoming) has expressed scepticism about the argument,
while Daniel Isaacson (2011) has defended it at length. For present purposes,
Ineed not try to decide if Kreisels reasoning is sound. However, one observation
about it usefully illustrates a theme of Chapters7 and9.
Since the pure hereditarily finite sets are isomorphic to the set of natu-
ral numbers, Kreisels argument assumes that both() and(()) exist as
well-defined sets, over which it is legitimate to quantify classically. Feferman
(2009) has queried whether this assumption is legitimate and our analysis ren-
ders it questionable whether there is even such a set as (). () is supposed
to comprise all the subsets of , but it is doubtful if the subsets of constitute
a mathematically well-defined totality. Certainly, the standard arguments for
the claim that this totality is well defined violate Brouwers strictures on the
infinite. Why are we sure that the power set of {a, b, c} is a well-defined set?
Because we can delimit the members of that set as comprising all the possible
results of making an arbitrary choice, for each member of {a, b, c}, as to whether
to include it in a subset or not. The corresponding justification of the existence
of (), however, would run as follows:() is a well-defined set because we
can delimit its members as comprising all the possible results of making an
7
As Kreisel noted in his review (Kreisel 1977)of Weston 1976, a reference Iowe to Isaacson
2011, 58.
Conclusion 317
Eliots words are not entirely apposite in the present case:most people do not need
a reason to adhere to classical logic. When philosophers try to supply a reason,
though, they are likely to justify that logic by reference to the classical, bivalent
semantics.
It seems to me that this is the wrong reason. Indeed, one does not have to
look far to see the difficulties that ensue when a bivalent semantics is applied to
many areas of discourse in which we are inclined to reason classically. Set theory
provides a particularly clear example. Whilst there are a few constructive and
semi-constructive set theorists, the vast majority of them use classical logic in
their proofs:they apply Excluded Middle or Double Negation Elimination with-
out a second thought. Any attempt to justify that practice by reference to the clas-
sical semantics, though, generates acute philosophical problems. In the paper
that opened the gate to the field we have been exploring, Dummett drew attention
to an important feature of the concept of truth ... that a statement is true only if
there is something in the world in virtue of which it is true (Dummett 1959, 14).
Under a suitably capacious understanding of what the world comprises, this is
318 Conclusion
indeed an aspect of our ordinary thinking about truth, and there is a correspond-
ing feature of the concept of falsity. But then, if a statement is bivalent, there must
either be something in the world in virtue of which it is true, or something in the
world in virtue of which it is false.
As we have seen, there is no guarantee that a set-theoretic statement will find
a basis for its truth or its falsity in the generally accepted axioms of the subject.
When applied to set theory, then, this consequence of Bivalence is a standing
invitation to postulate a non-physical world which contains the things needed to
make each set-theoretic statement either true or false. Explaining how we are able
to come to know things about that world then becomes an acute probleman
insuperable one, Ishould argue. If, however, we reject as philosophical mytholo-
gizing the postulated Platonic realm of truth-makers for set-theoretic statements,
we have to decide what to say about Bivalence.
One way to go would be to continue to adhere to Bivalence, but to adopt minimal-
ist notions of truth and falsity which do not carry the metaphysical implications that
Dummett discerns in our ordinary notions. Perhaps there is a coherent practice of
using the terms true and false as the minimalist wishes to employ them, but for our
purposes this approach leads nowhere. If true and false are used as the minimalist
understands them, then Either A is true or A is false says no more than A or not A,
and we lose any prospect of justifying Excluded Middle (or any other contested logi-
cal law) by reference to the principles of classical semantics:those principles now
emerge simply as re-statements of the logical truths that we are trying to justify. We
would be back in the dialectically futile circles of justification identified in Chapter1,
from which the whole book has been an effort to escape.
My way out of the bind has been to defend classical logic while avoiding any
commitment to classical semantics, in particular, to the Principle of Bivalence.
On this approach, the crucial task is to construct a theory of meaning which vali-
dates classical logic but which provides a more plausible account than classical
semantics of the way we actually use the linguistic devices in question. That is
what the previous chapter tried to provide for the language of set theory.
Our analysis of vague discourse contributes to the same overarching goal. If the
classical logic of and, or, and not codifies our ordinary deductive dispositions, as it
seems to, then it codifies those dispositions as they apply to vague terms, for almost all
the terms used in ordinary argument are vague. The Principle of Bivalence, though,
has little antecedent plausibility for vague statements. Certainly, it is hard to see in
virtue of what a statement like Tube a50 is red could be true, or in virtue of what it
could be false. Again, then, we were led to seek a non-classical semantics that still vali-
dates classical logic. In the case of vague discourse, the importance of this quest has
been widely recognized:the main recommendation for supervaluationist theories of
Conclusion 319
meaning is that they fulfil it. Those theories, though, face problems of their own and
the theory of the meaning of polar predicates sketched in Chapter8 offers an alter-
native way of attaining what is sought. Iam well aware that more work is needed on
that front:for one thing, while many vague predicates are polar, there is no reason
to assume that all of them are. All the same, we have a new modelone very differ-
ent from the supervaluationist accountof how classical logic might be validated for
vague statements that neither assumes nor is otherwise committed to their bivalence.
Other parts of this book also contribute to the goal of liberating classical
logic from classical semantics. Statements about ordinary objects that are pre-
sumed not to contain any vague predicates do not present the same challenges
to Bivalence as do set-theoretic statements or vague statements. All the same, we
shall need in the end an account of what endows such statements with the con-
tents that they have. The exclusionary theory of Chapter4 offers an answer to
this question which, once more, does not presuppose Bivalence. Given Aristotles
Thesis, the modal aspects of those contents are crucial when we come to investi-
gate their logical relations. Chapters6 and 7 duly developed a framework for that
investigation that, again, does not presuppose the Principle.
Much more work needs to be done on many fronts before the defence of clas-
sical logic against the challenges considered in this book could be said to be any-
where near complete. There are, moreover, many other challenges that Ihave not
had the space to consider at all. Between the motion and the act falls the shadow.
Ihope, though, to have said enough to show that many of the most seemingly
potent attacks on classical logic in fact hit a different target. They provide power-
ful challenges to the Principle of Bivalence but not to classical logic proper.
In his mid-twenties, T.S. Eliot spent some years as a graduate student in phi-
losophy at Harvard, the Sorbonne, and Oxford, and he completed (although he
never defended) a doctoral dissertation on the philosophy of F.H. Bradley. In
an essay on Bradleys Appearance and Reality that he wrote for H. H. Joachim at
Oxford during the Michaelmas Term of 1914, Eliot remarkedthat
when a philosopher pretends to emerge with some positive result which can be for-
mulated, which declares triumphantly that reality is this or that, some discovery which
informs us that anything is anything else than what we supposed it to be before we began
to philosophise,then the philosopher is simply pulling out of his pocket what he put
there himself. The token that a philosophy is true is, Ithink, the fact that it brings us to the
exact point from which we started. We shall be enriched, Itrust, by our experience on the
Grand Tour, but we shall not have been allowed to convey any material treasures through
the Custom House. And the wisdom which we shall have acquired will not be part of the
argument which brings us to the conclusion; it is not part of the book, but is written in
pencil on the fly-leaf. For the point to which we return should be the same, but somehow is
not ... But at this point it is wise to stop. (Eliot 2014, 191)
320 Conclusion
Eliot here invokes a long tradition, stretching back to Aristotle if not further,
whereby a central goal of philosophizing is to save the appearances, in part by
removing misconceptions that accrete to disfigure those appearances. Under this
approach, a main task of philosophical analysis will be to refine our sense of what
is to be saved and what is an accretion or an adiaphoron. Ihope the preceding
work may have fortified, or even instilled, the conviction that classical logic can
best thrive without the bivalent semantics. Classical logic is not deemed classical
because of its antiquity:the formalization on which Ihave focused has its origins
in nineteenth-century Jena, not the Athens of the fourth century BC . But it has
a simplicity, power, and indeed gracefulness that make the epithet apposite. By
putting it through the mills of philosophical analysis the dead weight of Bivalence
has been lifted from it. It would be too much to say that the end of this exploration
has been to arrive where we started and know the place for the first time. Ihope,
though, that the present tour has reinforced the conviction that the gasometers of
classical semantics need never be built on the green fields of classical logic. But
to quote Eliot one last timeat this point it is wise to stop.
References
Brouwer, L. E. J. 1923. ber die Bedeutung des Satzes vom ausgeschlossenen Dritten
in der Mathematik, insbesondere in der Funktionentheorie. Journal fr die reine
und angewandte Mathematik 154:17. Page references are to the translation by Stefan
Bauer-Mengelberg and Jean van Heijenoort, entitled On the significance of the
Principle of Excluded Middle in mathematics, especially in function theory, in van
Heijenoort, ed., 1967, pp.33541.
Brouwer, L.E. J. 1927. Intuitionistische Betrachtungen ber den Formalismus. Koninklijke
Akademie van wetenschappen te Amsterdam, Proceedings of the Section of Sciences
31:3749. Page references are to the English translation by Stefan Bauer-Mengelberg,
entitled Intuitionistic reflections on formalism, in van Heijenoort, ed., 1967, pp.4902.
Brouwer, L. E. J. 1928. Mathematik, Wissenschaft und Sprache. Monatshefte fr
Mathematik und Physik 36:15364. Page references are to the English translation by
William Ewald, entitled Mathematics, science, and language, in Ewald, ed., 1996,
pp.117085.
Bub, J. 1982. Quantum logic, conditional probability, and interference. Philosophy of
Science 49:40221.
Burgess, J.P. 1999. Which modal logic is the right one? Notre Dame Journal of Formal
Logic 40:8193.
Burgess, J.P. 2009. Philosophical Logic. Princeton, NJ:Princeton University Press.
Carnap, R. C. 1946. Modalities and quantification. The Journal of Symbolic Logic
11:3364.
Cartwright, R. L. 1987. Implications and entailments. In his Philosophical Essays.
Cambridge, MA:MIT Press, pp. 23756.
Chisholm, R. 1957. Perceiving:APhilosophical Study. Ithaca, NY:Cornell University Press.
Church, A. 1956. Introduction to Mathematical Logic, Volume I. Princeton, NJ:Princeton
University Press.
Clark, P. 1998. Dummetts argument for the indefinite extensibility of set and real num-
ber. Grazer Philosophische Studien 55:5163.
Cohen, P.J. 1966. Set Theory and the Continuum Hypothesis. NewYork:W. A.Benjamin.
Cohen, R.S. and M. W.Wartofsky, eds. 1968. Boston Studies in the Philosophy of Science,
vol. V.Dordrecht and Boston:Reidel.
Copi, I.M. 1973. Symbolic Logic, 4th edition. NewYork:Macmillan.
Davey, B.A. and H. A.Priestley. 2002. Introduction to Lattices and Order, 2nd edition.
Cambridge:Cambridge University Press.
Davies, M. K. 1975. Singular terms, quantification, and modality. B.Phil thesis,
University of Oxford.
Davies, M. K. and I. L. Humberstone. 1980. Two notions of necessity. Philosophical
Studies 38:130.
Dedekind, J.W. R. 1888. Was Sind und Was Sollen die Zahlen? Brunswick:Vieweg. Trans.
W. W.Beman in Ewald, ed., 1996, pp. 787833.
Dickie, I.M. 2010. Negation, anti-realism, and the denial defence. Philosophical Studies
150:16185.
Dogramaci, S. 2010. Knowledge of validity. Nos 44:40332.
Dretske, F.I. 1970. Epistemic operators. The Journal of Philosophy 67:100723.
Dretske, F.I. 1981. Knowledge and the Flow of Information. Cambridge, MA:MIT Press.
References 323
Dummett, M.A. E. 1959. Truth. Proceedings of the Aristotelian Society 59:14162. Page
references are to the reprint in Dummett 1978, pp.119.
Dummett, M.A. E. 1960. Presupposition. The Journal of Symbolic Logic 25:3369. Page
references are to the reprint in Dummett 1978, pp.258.
Dummett, M. A. E. 1963. The philosophical significance of Gdels Theorem. Ratio
5:14055. Page references are to the reprint in Dummett 1978, pp.186201.
Dummett, M.A. E. 1972. Postscript (1972) to Truth. In J. M.E. Moravcsik, ed., Logic
and Philosophy for Linguists:a Book of Readings. The Hague:Mouton, 1974, pp. 2005.
Page references are to the reprint in Dummett 1978, pp.1924.
Dummett, M. A. E. 1973. The justification of deduction. Proceedings of the British
Academy 59:20132. Page references are to the reprint in Dummett 1978, pp.290318.
Dummett, M.A. E. 1975. Wangs Paradox. Synthese 30:30124. Page references are to the
reprint in Dummett 1978, pp.24868.
Dummett, M.A. E. 1976a. What is a theory of meaning? (II). In Evans and McDowell,
eds., 1976, pp. 67137.
Dummett, M.A. E. 1976b. Is logic empirical? In H. D.Lewis, ed., Contemporary British
Philosophy, 4th series. London:Allen & Unwin, 1976, pp. 4568. Page references are to
the reprint in Dummett 1978, pp.26989.
Dummett, M.A. E. 1978. Truth and Other Enigmas. London:Duckworth.
Dummett, M.A. E. 1981a. Frege:Philosophy of Language, 2nd edition. London:Duckworth.
Dummett, M.A. E. 1981b. The Interpretation of Freges Philosophy. London:Duckworth.
Dummett, M.A. E. 1987. Reply to John McDowell. In Taylor, ed., 1987, pp. 25368.
Dummett, M.A. E. 1991. The Logical Basis of Metaphysics. London:Duckworth.
Dummett, M.A. E. 1993a. Could there be unicorns? In Dummett 1993d, pp. 32848.
Dummett, M.A. E. 1993b. What is mathematics about? In Dummett 1993d, pp. 42945.
Dummett, M. A. E. 1993c. Introductory remarks. In J. Czermak, ed., Philosophy of
Mathematics: Proceedings of the 15th International Wittgenstein Symposium Part I.
Vienna:Holder-Pichler-Tempsky, pp. 6976.
Dummett, M.A. E. 1993d. The Seas of Language. Oxford:Clarendon Press.
Dummett, M.A. E. 2000. Elements of Intuitionism, 2nd edition. Oxford:Clarendon Press.
Dummett, M.A. E. 2002. Yes, No, and Cant Say. Mind 111:28995.
Dummett, M.A. E. 2004. Truth and the Past. NewYork:Columbia University Press.
Dummett, M.A. E. 2006. Thought and Reality. Oxford:Clarendon Press.
Dummett, M.A. E. 2007a. Reply to Crispin Wright. In Auxier and Hahn, eds., 2007,
pp. 44554.
Dummett, M. A. E. 2007b. Reply to Ian Rumfitt. In Auxier and Hahn, eds., 2007,
pp. 6948.
Dunn, J.M. 1993. Star and perp:two treatments of negation. Philosophical Perspectives
7:33157.
Edgington, D.M. D. 1995. On conditionals. Mind 104:235329.
Edgington, D. M. D. 1996. Vagueness by degrees. In Keefe and Smith, eds., 1996,
pp. 294316.
Edgington, D.M. D. 2002. Indeterminacy de re. Philosophical Topics 28:2744.
Edgington, D. M. D. 2004. Two kinds of possibility. Proceedings of the Aristotelian
Society Supplementary Volumes 78:122.
324 References
Heijenoort, J.van, ed. 1967. From Frege to Gdel:ASource Book in Mathematical Logic,
18791931. Cambridge, MA:Harvard University Press.
Hellman, G. 2006. Mathematical pluralism:the case of smooth infinitesimal analysis.
Journal of Philosophical Logic 35:62151.
Herbrand, J. 1930. Recherches sur la thorie de la dmonstration. Thesis at the University
of Paris. Reprinted in Herbrand, ed. J.van Heijenoort, crits logiques. Paris:Presses
Universitaires de France, 1968, pp. 35153. Translated as Investigations in proof theory
by B. S.Dreben, W. D.Goldfarb, and J.van Heijenoort in Herbrand, Logical Writings,
ed. W. D.Goldfarb. Cambridge, MA:Harvard University Press, 1971, pp. 44202.
Hertz, P. 1922. ber Axiomensysteme fr beliebige Satzsysteme I. Mathematische
Annalen 87:24669.
Hertz, P. 1923. ber Axiomensysteme fr beliebige Satzsysteme II. Mathematische
Annalen 89:76102.
Hertz, P. 1929. ber Axiomensysteme fr beliebige Satzsysteme. Mathematische
Annalen 101:457514.
Heyting, A. 1934. Mathematische Grundlagenforschung. Intuitionismus. Beweistheorie.
Berlin:Springer Verlag.
Hilbert, D. 1928. Die Grundlagen der Mathematik. Abhandlugen aus dem mathema-
tischen Seminar der Hamburgischen Universitt 6:6585. Page references are to the
translation by S. Bauer-Mengelberg and D. Fllesdal, entitled The foundations of
mathematics, in van Heijenoort, ed., 1967, pp.46479.
Hintikka, K.J. J., ed. 1969. The Philosophy of Mathematics. Oxford:Oxford University
Press.
Hi, H. 1959. Extendible sentential calculus. The Journal of Symbolic Logic 24:193202.
Hughes, G. E. 1982. John Buridan on Self-Reference. Chapter Eight of Buridans
Sophismata. Translated with an Introduction and a Philosophical Commentary.
Cambridge:Cambridge University Press.
Humberstone, I. L. 1981. From worlds to possibilities. Journal of Philosophical Logic
10:31339.
Humberstone, I.L. 2000. The revival of rejective negation. Journal of Philosophical Logic
29:33181.
Incurvati, L. 2008. On adopting Kripke semantics in set theory. The Review of Symbolic
Logic 1:8196.
Incurvati, L. and J. Murzi. 2008. How basic is the Basic Revisionary Argument? Analysis
68:3039.
Isaacson, D.R. 2011. The reality of mathematics and the case of set theory. In Z. Novk
and A. Simonyi, eds., Truth, Reference, and Realism. Budapest: Central European
University Press, 2011, pp. 175.
Johansson, I. 1936. Der Minimalkalkl, ein reduzierter intuitionistischer Formalismus.
Compositio Mathematica 4:11936.
Kalmr, L. 1967. On the role of second-order theories. In Lakatos, ed., 1967, vol. I,
pp. 1045.
Keefe, R. and P. Smith, eds. 1996. Vagueness:AReader. Cambridge, MA:MIT Press.
Kleene, S.C. 1952. Introduction to Metamathematics. Amsterdam:North Holland.
Kneale, W.C. and M. Kneale. 1962. The Development of Logic. Oxford:Clarendon Press.
328 References
Kochen, S. and E. P.Specker. 1967. The problem of hidden variables in quantum mechan-
ics. Journal of Mathematics and Mechanics 17:5987.
Kock, A. 1981. Synthetic Differential Geometry. Cambridge:Cambridge University Press.
Koellner, P. 2009. On reflection principles. Annals of Pure and Applied Logic 157:20619.
Koellner, P. 2010. On the question of absolute undecidablity. In S. Feferman, C. D. Parsons,
and S. G.Simpson, eds., Kurt Gdel:Essays for his Centennial. Cambridge:Cambridge
University Press, pp. 189225.
Kolmogorov, A. N. 1925. tertium non datur.
32: 64667. Page references are to the translation by J. van Heijenoort in
vanHeijenoort, ed., 1967, pp.41637.
Kolmogorov, A. N. 1932. Zur Deutung der intuitionistischen Logik. Mathematische
Zeitschrift 35:5865.
Koslow, A. 1992. A Structuralist Theory of Logic. Cambridge:Cambridge University Press.
Kreisel, G. 1967. Informal rigour and completeness proofs. In Lakatos, ed., 1967, pp. 13871.
Page references are to the partial reprint in Hintikka, ed., 1969, pp.7894.
Kreisel, G. 1977. Review of Weston 1976. Zentralblatt fr Mathematik 341:31.
Kripke, S. A. 1963. Semantical analysis of modal logic. I. Normal modal proposi-
tional calculi. Zeitschrift fr mathematische Logik und Grundlagen der Mathematik
9:6796.
Kripke, S.A. 1964. Transfinite recursions on admissible ordinals, I, II (abstracts). The
Journal of Symbolic Logic 29:1612.
Kripke, S.A. 1965. Semantical analysis of intuitionistic logic. I. In J. N.Crossley and
M. A.E. Dummett, eds., Formal Systems and Recursive Functions. Amsterdam:North
Holland, pp. 92129.
Kripke, S.A. 1980. Naming and Necessity. Oxford:Blackwell.
Kuratowksi, C. 1958. Topologie I, 2nd edition. Warsaw:PWN (Polish Scientific Publishers).
Kuroda, S 1951. Intuitionistische Untersuchungen der formalistischen Logik. Nagoya
Mathematical Journal 2:3547.
Lakatos, I., ed. 1967. Problems in the Philosophy of Mathematics. Proceedings of
the International Colloquium in the Philosophy of Science, London, 1965, 3 vols.
Amsterdam:North Holland.
Lawvere, F. W. 1979. Categorical dynamics. In Topos-Theoretic Methods in Geometry
(Aarhus Mathematical Institute Publications Series, 30), pp. 128.
Lear, J. 1977. Sets and semantics. The Journal of Philosophy 74:86102.
Leivant, D. 1985. Syntactic translations and provably recursive functions. The Journal of
Symbolic Logic 50:6828.
Lemmon, E.J. 1965. Beginning Logic. London:Nelson and Sons.
Levy, A. and R. M.Solovay. 1967. Measurable cardinals and the continuum hypothesis.
Israel Journal of Mathematics 5:23448.
Lewis, D. K. 1970. General semantics. Synthese 22: 1867. Page references are to the
reprint in Lewis 1983, pp.189232.
Lewis, D.K. 1975. Adverbs of quantification. In E. L.Keenan, ed., Formal Semantic of
Natural Language. Cambridge:Cambridge University Press, pp. 315.
Lewis, D. K. 1979. Scorekeeping in a language game. Journal of Philosophical Logic
8:33959. Page references are to the reprint in Lewis 1983, pp.23349.
Lewis, D.K. 1983. Philosophical Papers Volume I. NewYork:Oxford University Press.
References 329
Ramsey, F.P. 1929. General propositions and causality. In Ramsey 1990, pp. 14563.
Ramsey, F. P. 1990. Philosophical Papers, ed. D. H. Mellor. Cambridge: Cambridge
University Press.
Ramsey, F.P. 1991. On Truth, ed. N. Rescher and U. Majer. Dordrecht:Kluwer.
Rasiowa, H. and R. Sikorski. 1953. Algebraic treatment of the notion of satisfiability.
Fundamenta Mathematicae 40:6295.
Read, S. L. 2009. Plural signification and the Liar paradox. Philosophical Studies
145:36375.
Read, S.L. 2010. Fields paradox and its medieval solution. History and Philosophy of
Logic 31:16176.
Read, S.L. and C. J.G. Wright. 1985. Hairier than Putnam thought. Analysis 45:568.
Renfrew, A.C. 1987. Archaeology and Language:The Puzzle of Indo-European Origins.
London:Jonathan Cape.
Renfrew, A.C. 2003. The Emerging Hypothesis:the archaeogenetics of farming/language
dispersals and other spread areas. In P. Bellwood and A. C.Renfrew, eds., Examining the
Farming/Language Dispersal Hypothesis. Cambridge:McDonald Institute, pp. 116.
Restall, G. 2005. Multiple conclusions. In P. Hajek etal., eds., Logic, Methodology, and
Philosophy of Science:Proceedings of the Twelfth International Congress. London:Kings
College Publications, pp. 189205.
Robinson, A. 1966. Nonstandard Analysis. Amsterdam:North Holland.
Rosch, E. 1975. Cognitive representations of the semantic categories. Journal of
Experimental Psychology:General 104:192233.
Rumfitt, I. 2000. Yes and No. Mind 109:781823.
Rumfitt, I. 2002. Unilateralism disarmed: a reply to Dummett and Gibbard. Mind
111:30521.
Rumfitt, I. 2007. Asserting and excluding. In Auxier and Hahn, eds., 2007, pp. 63993.
Rumfitt, I. 2008a. Knowledge by deduction. Grazer Philosophische Studien 77:6184.
Rumfitt, I. 2008b. Coordination principles:a reply. Mind 117:105963.
Rumfitt, I. 2010. Ricky Ponting and the judges. Analysis 70:20510.
Rumfitt, I. 2013. Old Adams Buried. Analytic Philosophy 54:5788.
Rumfitt, I. 2014a. Truth and meaning. Proceedings of the Aristotelian Society,
Supplementary Volumes 88:2155.
Rumfitt, I. 2014b. Brouwer versus Wittgenstein on the infinite and the law of excluded
middle. Grazer Philosophische Studien 89:95110.
Rumfitt, I. 2015. Determinacy and bivalence. In Michael Glanzberg, ed., The Oxford
Handbook of Truth. Oxford:Clarendon Press.
Russell, B. A. W. 1908. Mathematical logic as based on the theory of types. American
Journal of Mathematics 30:22262. Page references are to the reprint in Russell, Logic and
Knowledge:Essays 19011950, ed. R. C.Marsh. London:Allen & Unwin, 1956, pp. 59102.
Russell, B.A. W. 1919. Introduction to Mathematical Philosophy. London:Allen & Unwin.
Ryle, G. 1945. Knowing how and knowing that. Proceedings of the Aristotelian Society
46:116. Page references are to the reprint in Ryle 2009, pp.22235.
Ryle, G. 1946. Why are the calculuses of logic and arithmetic applicable to reality?
Proceedings of the Aristotelian Society, Supplementary Volumes 20:209. Page refer-
ences are to the reprint in Ryle 2009, pp.23643.
References 333
Tarski, A. 1983. Logic, Semantics, Metamathematics: Papers from 1923 to 1938, ed.
J. H.Woodger and J. Corcoran. Indianapolis:Hackett.
Taylor, B.M., ed. 1987. Michael Dummett:Contributions to Philosophy. Dordrecht:Nijhoff.
Tennant, N.W. 1978. Natural Logic. Edinburgh:Edinburgh University Press.
Tharp, L. H. 1971. A quasi-intuitionistic set theory. The Journal of Symbolic Logic
36:45660.
Troelstra, A. S. 1973. Metamathematical Investigation of Intuitionistic Arithmetic and
Analysis. Berlin:Springer.
Uzquiano, G. 1999. Models of Zermelo set theory. The Bulletin of Symbolic Logic
5:289302.
Uzquiano, G. 2002. Categoricity theorems and conceptions of set. Journal of
Philosophical Logic 31:18196.
Van Fraassen, B.C. 1977. The only necessity is verbal necessity. The Journal of Philosophy
74:7185.
Van Fraassen, B.C. 1981a. Probabilistic semantics objectified I:postulates and logics.
Journal of Philosophical Logic 10:37194.
Van Fraassen, B.C. 1981b. Probabilistic semantics objectified II:implication in probabil-
istic model sets. Journal of Philosophical Logic 10:495510.
Varzi, A.C. 2007. Supervaluationism and its logics. Mind 116:63375.
Velleman, D.J. 1993. Constructivism liberalized. The Philosophical Review 102:5981.
Walmsley, J. O. 2002. Categoricity and indefinite extensibility. Proceedings of the
Aristotelian Society 102:21736.
Weston, T. 1976. Kreisel, the Continuum Hypothesis, and second-order set theory.
Journal of Philosophical Logic 5:28198.
Weyl, H.K. H. 1918. Das Kontinuum. Leipzig:Veit.
White, A.R. 1971. Inference. The Philosophical Quarterly 21:289302.
Whitehead, A.N. and B. A.W. Russell. 1925. Principia Mathematica, 2nd edition, vol. I.
Cambridge:Cambridge University Press.
Wiggins, D.R. P. 1993. Putnams doctrine of natural kind words and Freges doctrines of
sense, reference, and extension:can they cohere? In A. W.Moore, ed., Meaning and
Reference. Oxford:Oxford University Press, pp. 192207.
Wiggins, D. R. P. 2001. Sameness and Substance Renewed. Cambridge: Cambridge
University Press.
Williamson, T. 1994. Vagueness. London:Routledge.
Williamson, T. 1998. Indefinite extensibility. Grazer Philosophische Studien 55:124.
Williamson, T. 2000. Knowledge and Its Limits. Oxford:Clarendon Press.
Williamson, T. 2007. The Philosophy of Philosophy. Oxford:Blackwell.
Wittgenstein, L.J. J. 1922. Tractatus Logico-Philosophicus. London:Routledge & Kegan
Paul.
Wittgenstein, L. J. J. 1953. Philosophical Investigations, trans. G. E. M. Anscombe.
Oxford:Blackwell.
Wittgenstein, L.J. J. 1961. Notebooks 191416, ed. G. H.von Wright and G. E.M. Anscombe,
trans. G. E.M. Anscombe. Oxford:Blackwell.
Wittgenstein, L.J. J. 1974. Philosophical Grammar, ed. R. Rhees, trans. A. J.P. Kenny.
Oxford:Blackwell.
336 References
Dummett, M., 6, 16, 22, 23, 57, 72, 92, 121, 165, choice of logic within an exclusionary
196,225 semantics,1224
argument for exclusionary theory of difficulties with disjunction, 1001,1034
content,99101 employing spaces of possibilities,11415
argument for verificationist theory of exclusionary content,
meaning,1257 and assertion,105
assertoric content versus ingredient and information theory,104
sense,284 coarse versus fine-grained,117
attacks on classical logic, 95110,1224 of conditionals,121
on bivalence, 223, 95, 109111 of other sentential connectives,11820
on harmony and stability of introduction existence property (of theories), 1312,2867
and elimination rules,45,8 expectation value (of an observable),172f.
on indefinite extensibility,263 Extensionality, Axiom of, 290, 291, 292, 294,
on infinity, 1989,210 296,297
on justifying logical laws, 23,910 exterior of a set of possibilities,114
on logical truths as by-products of logical
consequence,56,72 (F), semantic principle for falsum, 190
on mathematical versus concrete objects,2646 Feferman, S., 297, 298, 299,316
on meanings of connectives, 1920,217 Fermats Last Theorem,82
on operational interpretations of quantum Ferreira,F.,6
theory,176 Field, H., 78, 83,207
on semantic shift principles,1267 Fine, K., 77, 83, 160, 237, 283,284
on Tertium non datur, 223, 958,10511 Finkelstein, D.,176
on the significance of Gdels First First Isomorphism Theorem (Zermelo),2712
Incompleteness Theorem, 145, 263,266 first-order logic, 14, 76, 205, 253, 267, 285
on truth,317 Fitting, M., 2824,287
on vagueness, 2257, 234, 2478,261 foil (for a vague predicate),2412
on whether logic is empirical, 142,1767 forcing,279
Dunn, J.,167 Foundation, Axiom of, 281,289
Foulis, D.,178
Eadem est scientia oppositorum, principle of, Fraenkel, A.,314
196, 202, 203,237 Frege,G.,
Edgington, D., 823, 867, 8892, 100, 240,261 against the formalists,299
Egan, A.,135 axiomatic formalization of logic,71
elementary set,2313 definition of ancestral,206
eliminationrules, definition of finite cardinal,207
for connectives generally,4 elucidations of basic logical notions,72
for negation,3,5 function/argument analysis of statements,18
Eliot, T.S., 317, 318,319 on functions as unsaturated,271
end-extension,294 on how deductions can be informative,57
entailment (versus deducibility), 83, 2056 on inference,37
enthymemes,3942 on Law of Excluded Middle, 223, 2345,249
Etchemendy,J.,73 on logical laws, 12,18
Euclid, 266,311 on the identity of sets,2401
Evans, G., 86, 87, 92, 218,240 on vagueness, 2223, 225, 234,235
evidential probability,138 Friedman, H., 1312, 2901, 293,297
evidentialist semantics,1308 functional conception of modal
see also challenges to classicallogic entities,15860
Ex Contradictione Falsum, rule of, 190,191
Ex Contradictione Quodlibet, rule of, 118,189 Gallistel, C.,239
Ex Falso Quodlibet, rule of,190 Grdenfors, P.,242
Excluded Middle, Law of, 44, 81, 118, 126, 128, Gardner, M.,168
1489, 162, 195, 1978, 2006, 223, 2345, Geach, P.T., 21, 102,241
249, 250, 270, 281, 285, 290, 292, 296,298 Generalized Continuum Hypothesis, 27, 196,
exclusion, of a possibility by a statement,114f. 302,315
exclusionary account of content, 23, 97, Gentzen, G., 26, 50, 51, 53, 56, 67,289
99105,11124 see also Gdel-Gentzen translation
340 Index
Koellner, P.,278 McDowell, J., 23, 96, 109, 127, 129, 14752,1912
Kolmogorov, A.N., 26, 128, 198, 289,292 McFetridge, I., 7991, 847, 89,912
Koslow, A.,47,75 McGee, V., 17, 73,2735
Kreisel, G.,132 Mc G (McGees axiom in set theory),2735
argument for bivalence of the Continuum McGinn, C.,112
Hypothesis,31316 Mackie, J.L.,22
Kripke, S., 75, 83, 86, 153, 155, 166,294 McKinsey, J.C.C.,20
Kripke semantics for intuitionistic logic, Mac Lane, S.,274
27980, 283,300 Malinowski, J.,175
Kripke-Platek set theory, 294ff. Mallory, J.,146
intended models of,295 Mares, E.,137
Kuratowksi, C.,162 margin for error principle (Williamson),226
Kurgan Hypothesis,1456 Mates, B., 314,154
Kuroda, S.,289 Maudlin, T.,260
maximal closeness (of a colour pole to a
LHpital,210 coloured object),2447
Lakatos, I.,133 measurable cardinals, 277,278
Lawvere, F., 21011,213 Mellor, D.H.,101
Lear, J., 263, 264,280 Menzel, C.,275
Leibniz, G.W., 154,210 mereology, 188
Leivant, D.,292 Microaffineness, Principle of (in SIA), 212,213
Lemmon, E.J., 314,38 Microstraightness, Principle of (in SIA),
Leverrier, U., 86, 88,89,90 211,212
Levy, A.,310 Mill, J.S., 37, 567,71,79
Lewis, C.I.,71 Mills Problem,5665
Lewis, D., 17, 85, 117, 121, 153, 154, 155,237 minimal logic,163
Lindenbaum,A.,47 mini-worlds(Kripke),155
Lindenbaum-Scott Theorem, 4751,746 modal concretism, 153, 155, 158,160
Lbs Theorem,35 moderate modal realism,1535
Locke, J., 15,239 and possibilities versus possible
logical basis of metaphysics, non-existence worlds,15860
of,21719 Modus Ponens, rule of, 17, 55, 56, 58, 67, 68,
logical consequence, 34, 3942, 526, 668,85 72,121
and deductive capacities,3842 Moerdijk, I.,213
narrow versus broad, 55,678 monotonicity of implication,42,43
pre-theoretic obscurityof,34, 41 Moore, G.E.,34,55
Russells account of,714 Moores Paradox, 51,1001
Tarskis account of,723 Moore, J.,112
logical laws, 1, 314,526 Mostowski, A.,316
as general laws of implication relations, 21, Muller, F.,275
546, 58, 668, 74,778 multiple-conclusion consequence, 501,53
challenges to versus alternative Murzi, J.,127
codificationsof,18 Myhill, J., 206, 281,28990
formalization of, 2, 526,667
proof-theoretic justifications of,48 (N), semantic axiom for sentential negation,
semantic justifications of,913 evidentialist form, 137
logical necessity, 21, 68,778 exclusionary form, 119
and apriority,813 in truth-grounds semantics, 167, 185
and metaphysical necessity, 834, 867, (NC), principle of non-contradiction, 186
8892 (Neg), semantic axiom for predicate negation, 246
Ian McFetridge on, 7981,845 necessity, 21,746
logical possibilities, 153, 1627,1813 metaphysical,83f.
logical truths,56,72 relation of logical to non-logical necessity,77
McDowells account of,1501 see also logical necessity
Lwenheim-Skolem Theorems, 205, 208,267 negation,
Lyaletski,A.,14 logic of, 3, 5, 8, 11, 12, 1223, 13941, 1467,
Lycan,W.,17 150,18991
342 Index
Wittgenstein, L., 25, 2024, 207,242 Z (Zermelos original set theory), 281,
and postulate (B), 202-4 289290,314
Woodin, W.H., 145, 310,315 ZF (Zermelo-Frankel set theory), 240, 27686,314
Wright, C., 26, 1267, 220, 222,264 ZFI,28990
argument for intuitionism as the logic of ZF,290
vagueness, 2237,2325 Zermelo, E., 26, 26970, 2713, 276, 281, 314
Basic Revisionary Argument,1267