The Boundary Stones of Thought

The Boundary Stones

of Thought
An Essay in the Philosophy of Logic

Ian Rumfitt

CLAR ENDONPR ESS

3
Great Clarendon Street, Oxford, OX 2 6DP,
United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the Universitys objective of excellence in research, scholarship,
and education by publishing worldwide. Oxford is a registered trade mark of
Oxford University Press in the UK and in certain other countries
Ian Rumfitt 2015
The moral rights of the author have been asserted
First Edition published in 2015
Impression:1
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
198 Madison Avenue, NewYork, NY 10016, United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number:2014953348
ISBN 9780198733638
Printed and bound by
CPI Group (UK) Ltd, Croydon, CR0 4YY
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.
To my mother and to the memory of my father
Preface

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

philosophical problems that even he found difficult, remain an inspiration.

Although Dummett supervised my doctoral work at Oxford, that was in a dif-
ferent area:Iknew that at that stage Ilacked the philosophical maturity needed
to get to grips with conflict of logical laws. However, after Ireturned to Oxford
from America in 1998, Dummett was kind enough to discuss my ideas with me,
conversations that continued until 2007, when he no longer felt strong enough to
talk about philosophical questions. Even though he is in many ways one of the
principal targets of this work, he was always encouraging, and Iam extremely
sorry that his death, in December 2011, has prevented me from presenting him
with a copy. Itake comfort, though, from the knowledge that he strongly disap-
proved of premature publication:the principle he inculcated in his students was
that no one should put anything into print until he no longer sees how to make
it any better (Frege:Philosophy of Mathematics, p.x). Whilst Icannot see how
to make this book any better, Iam sure that others can, so the time has come to
send it on its way. Ihope that some readers will find it interesting.
Acknowledgements

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

Part I. The Nature of Logic

2. Logical Laws 31
2.1 Consequence 31
2.2 Inference and Deduction 34
2.3 The Varieties of Deduction and of Implication Relations 38
2.4 Implications and Possibilities 46
2.5 The Role of Logic 52
2.6 Knowledge by Deduction 56
3. Logical Necessity 66
3.1 Logical Consequence Redux 66
3.2 The Controversy Over Logical Necessity 68
3.3 Notions of Necessity 74
3.4 Logical Necessity versus Apriority and Metaphysical Necessity 81
3.5 Logical and Metaphysical Necessity:The Paradox Resolved 88

Part II. Five Attacks on Classical Logic

4. The Argument of Dummetts Truth 95
4.1 The Argument against Classical Logic in Dummetts Truth 95
4.2 The Exclusionary Theory of Conceptual Content 99
4.3 Where the Argument of Truth Fails 105
4.4 Exclusion and Truth 111
4.5 An Exclusionary Semantics for the Language
of the Propositional Calculus 117
4.6 The Choice of Logic within an Exclusionary Semantics 122
5. The Verificationist Attack on Classical Logic 125
5.1 The Strong Verificationist Attack on Classical Logic 125
xiv Contents

5.2 How to Be a Strong Verificationist 129

5.3 A Renewed Threat to Classical Logic 138
5.4 Why One Should Not Be a Strong Verificationist 143
5.5 Dubious Grounds:McDowells Challenge to Classical Logic 147
6. Possibilities 153
6.1 Moderate Modal Realism and Possible Worlds 153
6.2 Reasons for Seeking an Unworldly Theory 157
6.3 Previous Unworldly Theories 160
6.4 The Structure of the Space of Possibilities; Truth-Ground Semantics 162
6.5 Distribution and Quantum Mechanics 167
6.6 Distribution and Regularity 181
7. Challenges from the Infinite and from the Infinitesimal 184
7.1 The Semantics and Logic of Negation 185
7.2 Statements With and Without Backs 193
7.3 The Intuitionists on Infinity 197
7.4 A Consolation Prize for the Intuitionist:Smooth Infinitesimal Analysis 210
7.5 Logic and Metaphysics 217
8. The Challenge from Vagueness 220
8.1 The Paradox of the Heap 220
8.2 Intuitionism as the Logic of Vagueness 223
8.3 A Semantics for Vague Predicates that Validates Intuitionistic Logic 228
8.4 Paradigms and Poles 235
8.5 A Semantics for Polar Predicates that Validates Classical Logic 242
8.6 The Sorites Revisited 250
8.7 Vagueness and Distribution 255
9. On the Use of Classical Logic in Set Theory 263
9.1 What is Mathematics About? 264
9.2 Attempts to Attain Categoricity 270
9.3 The Iterative Conception of Sets, and the Threat to Classical Logic 276
9.4 Attempts to Justify Classical Logic under the Iterative Conception 281
9.5 Classical Logic Justified via a Negative Translation 286
9.6 Classical Logic in Set Theories Weaker than ZF 289
9.7 As Far as Possible versus As Far as Necessary 299
10. Conclusion 302
1 0.1 The Simple and Revised Arguments for Bivalence 302
10.2 Where the Arguments Go Wrong 308
10.3 Classical Logic versus Classical Semantics 317

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

1.1 Disputes Over Logical Laws

I understand by logical laws not psychological laws of takings-to-
be-true, but laws of truth. If it is true that Iam writing this in my room on
13 July 1893, while the wind howls outside, then it remains true even if all
men should subsequently take it to be false. If being true is thus independ-
ent of being acknowledged by someone or other, then the laws of truth
are not psychological laws:they are boundary stones fixed in an eternal
foundation, which our thinking can overflow, but never displace. (Frege
1893, xvi)

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

Philosophers have recently debated whether a rule-circular argument can jus-

1
tify its conclusion. However, even if it can justify, a rule-circular argument for
the soundness of R cannot advance a rational discussion of whether R should be
accepted as a universally applicable logical rule. For suppose that an adherent
of R gives a rule-circular argument for its soundness. How could this advance
the discussion? The argument cannot be expected to persuade an interlocutor
who doubts or denies the universal applicability of rule R. For it will be an open
question for that interlocutor whether the application of R that is made in the
course of the deduction is one of the doubtful or deniable cases. It may be said
that the argument still advances the discussion, even if it cannot conclude it, by
reducing the question of the rules universal applicability to the correctness of a
single, metalogical application of it. In fact, though, any appearance of reduction
here is spurious. The argument just given for the soundness of Double Negation
Elimination shows why. It constitutes a classically valid deduction of the rules
universal soundness only because it involves a step of universal generaliza-
tion from a formula (viz., (8)) that contains a variable (viz., A) that ranges over
all the statements in the object language. So anyone who allows the step from
(6)to (7)must be prepared to allow the elimination of a double negation from
any statement in the object language. The single metalogical step from (6)to (7),
then, is really a portmanteau which carries within it the universal applicability
of Double Negation Elimination to the object language. Accordingly, not only
will the argument fail to persuade anyone not already prepared to eliminate
arbitrary double negations, it will fail to advance the discussion with him. For
after traversing a very small circle, the argument takes us straight back to the
issue originally in contention, namely, whether Double Negation Elimination
is unrestrictedly sound. Even if it can justify its conclusion, then, a rule-circular
argument for the soundness of a rule will be dialectically futile in a discussion in
which the soundness of that rule is seriously contested.
How might we avoid this kind of futility? In The Logical Basis of Metaphysics
(1991), his most sustained attempt to show how disputes between logical schools
can be and should be resolved, Dummett relied heavily on considerations
drawn from proof theorymore particularly, from the proof theory of natural
deduction formalizations of logic. In a natural deduction system, each connec-
tive is associated with two rules. The introduction rule for the connective states
the conditions under which one may deduce a conclusion whose dominant log-
ical operator is the connective. The connectives elimination rule states what
may immediately be deduced from some premisses, one of which, the major

1
Boghossian (2000) argues that it can, Dogramaci (2010) that it cannot.
 Introduction 5

premiss, is dominated by the connective. Dummett supposes that the introduc-

tion and elimination rules for the logical connectives must exhibit a kind of bal-
ance that he calls harmony and stability. That is to say:the elimination rule
must not allow one to deduce more from a premiss than the introduction rule
requires if an assertion of that premiss is to have been justified (harmony); and
what is needed to justify that assertion cannot outrun what is needed to ground
the conclusions that the elimination rule licenses drawing from the premiss
(stability).
More rigorously, let us call A a maximally strong statement with the property
F if A has F and logically implies any statement that has F; and let us call A a
maximally weak statement with property F if A has F and is logically implied by
any statement that has F. Then the requirements of harmony and stability for the
2
rules that regulate a connective C come to this. Given the elimination rule for C,
a statement whose main operator is C must be a maximally strong statement that
can stand as the conclusion of the introduction rule for C (harmony). And, given
the introduction rule for C, such a statement must be a maximally weak state-
ment that can stand as the major premiss of the elimination rule for C (stability).
In both classical and intuitionistic logic, the rule for introducing a negation
is a rule of proof that we may call Simple Reductio:having deduced a contra-
diction from the supposition A, together with the premisses in a set X, we may

infer that X entails not A . Dummett then argues that the elimination rule
which stably pairs off with Simple Reductio is the intuitionistic rule, which says
3
that the premisses A and not A together yield a contradiction. The stronger
elimination rule that the classical logician goes byviz., Double Negation
Eliminationcannot be justified by the requirements of harmony and stability,
and Dummett eventually concludes that it cannot be justified at all. Indeed, he
maintains that the classical deduction rules for negation fail to endow not with
a coherentsense:
Attempted explanations [of classical negation] rely always on the presumption that,
knowing what it is for the condition for some statement to be true to obtain, in general
independently of the possibility of recognising it to obtain, we thereby know what it is for
it not to obtain; and this blatantly presupposes a prior understanding of classical nega-
tion. It almost seems that there is no way of attaining an understanding of the classical
negation-operator if one does not have one already. That is a strong ground for suspicion
that the supposed understanding is spurious. (Dummett 1991, 299)

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)

Although Dummett is asking how logicians in rival schools might come to a

mutual understanding, what he says here is also suggestive for the present project
of seeking a rational resolution of their dispute. There is no prospect whatever
of giving the sense of a logical constant without using that very constant, and
much else besides, in the metalinguistic principle that specifies that sense. Even
the homophonic axiom (N) uses a negation-sign, along with a biconditional and a
truth-predicate, in giving the meaning of the negation-sign. Anon-homophonic
semantic axiom will involve more apparatus, not less. Given that any seman-
tic axiom for not will involve, inter alia, the concept of negation, we shall have
to apply logical rules concerning negation in deducing consequences from the
axiom. We clearly risk slipping into dialectical futility here, but the danger may
be averted if the relevant consequences of the semantic theory are stable under
the contested changes in the metalogic. Of course, not all the consequences of the
semantic theory can be the same; if they were, then the logical systems in question
10 Introduction

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

change:classical semantics no longer validates the Law of Distribution if the

rules of quantum logic regulate deductions in the metalanguage. For let A
and B be arbitrary statements. The Principle of Bivalence entails that A is
either true or false, and that B is either true or false, so by conjunction intro-
duction we may infer
(1) (Either A is true or A is false) and (either B is true or B is false).
In order to show that Distribution is sound, however, we need to show that the
four cells of a truth-table with two variables exhaust all the possibilities. That is,
we need toshow
(2) Either (A is true and B is true) or (A is true and B is false) or
(A is false and B is true) or (A is false and B is false).
However, while (2) follows from (1) in classical logic, it does not follow in
quantum logic:the inference from (1)to (2)requires the Law of Distribution.
Hence a quantum logician can consistently accept (1)and the rest of classi-
cal semanticswhile denying the soundness of Distribution. The case brings
out clearly, then, how we may need to invoke different non-homophonic
semantic theories depending on which dispute between rival logical schools
is under consideration:a semantic theory that advances rational debate when
a classical logician is arguing with one rival may not help at all in his argu-
ment with another.
(4) Since Dummett takes intuitionistic logic to be the chief rival to classical
logic, we can see why so much of his writing on the philosophy of logic focuses on
the relative merits of classical semantics and the sort of verificationist theory of
meaning of which the Beth semantics (as standardly interpreted) and the Heyting
semantics are examples. It is not just that a classical logician can understand why
someone who uses the connectives in accordance with the Beth semantics will
not accept the soundness of all classical laws. In addition, the relevant deliver-
ances of both semantic theories are part of the common ground between the rival
logical schools:classical semantics validates Double Negation Elimination while
the semantic theories of Beth and Heyting invalidate that rule, whether the meta-
logic is classical or intuitionistic. By comparing the merits of these semantic theo-
ries, we may hope to escape from dialectic futility and bring reason, or at least a
wider range of philosophical considerations, to bear on the choice between these
two systems of logic.
Since the challenges to classical logic that Ishall examine in this book also
seem to lead to an intuitionistic alternative, it may appear that I can follow
Dummett and focus on the relative merits of these two sorts of semantic theory.
 Introduction 13

In fact, though, matters are not so straightforward. Each of the challenges

that Ishall discuss casts doubt upon classical semantics, and especially upon
the Principle of Bivalence. However, for reasons to be given in Chapter5, Ido
not regard any verificationist semantic theory as providing a plausible theory
of meaning even for the language of pure mathematics. Furthermore, Ithink
it is a strategic mistake to rest the case for classical logic on the Principle of
Bivalence: the soundness of the classical logical rules is far more compelling
than the truth of Bivalence. Accordingly, assessing these challenges to classi-
cal logic will involve constructing non-homophonic semantic theories which
are more plausible than either classical semantics or verificationist semantics,
but which remain stable under the prospective changeswhether from classical
to intuitionistic, or vice versain the underlying logic. The family of semantic
theories that Irecommend will emerge, in Chapters6 and 7, from a philosophi-
cal analysis of the nature of logic and, more particularly, of the notion of logical
consequence. Partly for this reason, Ibelieve that semantic theories from the
same family may be useful in adjudicating cases of logical rivalry other than the
dispute between classical and intuitionist logicians, and Itouch briefly (5.3,
6.5) on challenges to the Law of Distribution stemming from quantum mechan-
9
ics. However, Ishall not try to delineate the range of logical disputes to whose
rational resolution members of the family may be expected to contribute.
I have selected the five challenges I discuss because each has some prima
facie plausibility and philosophical interest, and because there are connections
between them which make this book more than a collection of independent
papers. Powerful as they first appear to be, though, Ishall conclude that each
of them fails. The book, then, vindicates classical logic against certain lines of
attack. It is not, of course, a complete vindication. There is not the slightest pros-
pect of proving that the rules of classical logic are sound. At least, there is no such
prospect if we require a proof to be cogent in the sense of rationally compelling
someone who is not initially disposed to do so to accept its conclusion. All that
the defender of classical logic can do is scrutinize particular attacks and try to
repel them. Some readers who deviate from classical logic will think that Ihave
failed to consider the most powerful assaults, and there are certainly other lines
of attack (notably those coming from a dialetheist direction) that deserve detailed
consideration. Asingle book, though, cannot be expected to take on all comers.
This is an essay in the philosophy of logic, not a purported summa that might be
subtitled Gottlobus ab omni naevo vindicatus.

9
See also the mention of Robert Goldblatts recent cover semantics for the relevance logic R in
Chapter7, n.3.
14 Introduction

1.2 The Scope of This Book, and the Nature of

Disputes Over Basic Logical Laws
In developing my account of the nature of logic, Ishall strive as far as possible not
to beg questions against non-classical logicians. All the same, in the philosophi-
cal analysis that precedes the adjudication of the challenges to classical logic,
as in any rigorous inquiry into anything, we shall need to exercise our capac-
ity for deduction, and we shall need implicitly to appeal to a logic that provides
the standards for assessing deductions. On occasions, Ishall need to make a pre-
sumption about what that logic is. Ipropose to start from the presumption that
classical logic provides those standards; this is to be the working assumption
until reason is found to amend or abandon it. In other words, Iaccord to classical
logic default status.
Why should we accord this status to classical logic? To address this ques-
tion, we need to be clear what we mean by classical logic. One might mean a
certain kind of formal system, and the classical systems certainly have some
attractive properties. The classical propositional calculus is Post-complete:if a
10
non-theorem is added to it as an axiom, the resulting system is inconsistent.
There is, then, a sense in which the classical propositional logic is the strongest
logic that deals with the notions (conjunction, negation, etc.) of which it treats.
The classical first-order predicate calculus is not Post-complete (see Church
1956, 185), but Herbrands Theorem holds for it:the question of the deducibility
(or validity) of a first-order formula may be reduced to that of the deducibil-
ity (or validity) of a related formula of the classical propositional calculus (see
Herbrand 1930). While partial analogues of this result have been proved for the
intuitionistic first-order calculus (see Lyaletski 2008), the fully general version
of Herbrands Theorem fails for that system:in intuitionistic logic, first-order
formulae cannot in general be pre-processed into prenex normal form, and
11
Skolemization obliterates intuitionistically crucial distinctions. There is,
accordingly, a much simpler relationship between the classical propositional
calculus and its first-order extension than obtains between their intuitionistic
correlates.

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

is:(A B) C; therefore (A C) (B C). Yet we are not inclined to think that

If you close switch x and switch y, the light will go on entails Either the light
will go on if you close switch x, or it will go on if you close switch y (consider an
electrical circuit in which the switches x and y are in series, with both switches
15
open). The classical logic of the conditional, then, validates arguments where the
conclusion appears not to follow from the premisses.
Some care is needed in reaching this conclusion. Most of the logical systems

that philosophers and logicians take seriously validate the schema A A but
it would be a poor objection to this schema to observe that, on its most natural
reading, the English conditional If Henry likes any fish, then Henry likes any
fish could easily be false. (Suppose Henry likes sole but not bass.) All the obser-
vation shows is that not all instances of the English schema If..., then... have

the logical form A A , even when it is understood that the two occurrences
of ... are to be filled in with the same English sentence. When a logician asserts

the validity of the schema A A , he is not claiming that every instance of
the English schema If..., then... is true, nor is he making a similar claim for
a schema in any other natural language. Rather, he is ascribing truth to every
conditional whose antecedent and consequent say the same thing in the same
words, and the point about If Henry likes any fish, then Henry likes any fish is
that it does not answer to that description. When the quantifier any appears in
the grammatical antecedent of an English conditional, it takes wider scope than
the conditional; so we understand the present conditional to say Take any fish
you like:if Henry likes it then he likes any fish. Once the meaning of the condi-
tional is spelled out in this way, it is clear that it is not one whose grammatical
16
antecedent and consequent say the same thing. We cannot, however, similarly
explain away the counterexamples to our two classically valid schemata involv-
ing the conditional. The English sentence If you close switch x and switch y,
the light will go on, for example, is plausibly analysed as a conditional whose

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

consequent is the light will go on and whose antecedent is the conjunction of

the statements you close switch x and you close switch y.
In this book, Ishall set aside, as far as possible, the whole issue of the correct
logic for the conditional. Iam interested in arguments that challenge those logi-
cal laws which codify generally accepted patterns of argument, and the classical
laws for the conditional are notor, at least, are not clearlyin that category. It
will not be possible to bracket the issue entirely. In arguing about the validity of
logical laws, one cannot avoid deductions that essentially involve conditionals.
17
While Ientertain no serious doubts about Modus Ponens, the vernacular condi-
tional seems not to sustain an unrestricted rule of Conditional Proof, so we shall
need to assess carefully the intuitive plausibility of any arguments that proceed
18
by that rule. Conditionals will force themselves into play in another way too. As
the argument of the book develops, we shall find ourselves examining various
sorts of semantic theory. No semantic theory can be accepted if it precludes any
satisfactory treatment of the conditional, and it is a mark in favour of a theory if
it points to a treatment that validates a weaker logic for the conditional than the
classical one. However, while Iwill note such pointers in passing, Iwill not be able
to investigate them systematically. This renders the analysis here provisional and
incomplete:logic is sufficiently integrated that a change in ones view of the cor-
rect laws for the conditional is liable to impinge on judgements about laws con-
cerning other basic notions. There is, however, no room to treat the conditional
systematically in the present compass, and Ican at least refer the reader to an
excellent book (namely, Priest 2001)that is devoted to investigating what the logic
for the conditional really is.
When Ispeak of classical logic, then, Ishall mean the classical logic of negation,
conjunction, disjunction, and quantification. Since the classical deduction rules
for these notions appear to codify our ordinary norms for assessing the validity
of arguments that essentially involve them, there is an initial presumption that
these rules are correct. All the same, this presumption can be overturned. For one
thing, any codification will classify arguments and their component statements
using notions too recherch to figure in ordinary assessments of arguments, and
it is always open to someone to propose a newor, for that matter, oldset of

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

theoretical notions to effect the codification. Such a proposal will challenge at

least the formulation of the classical logical laws, and the challenge may extend
to the law itself. Frege recognized the importance of using the right theoretical
notions in codifying logical laws. In logic, as in other sciences, he wrote in the
Logik of1897,
it is open to us to coin technical terms, regardless of whether the words are used in pre-
cisely that way in everyday speech. It does not matter if the meaning we fix on conforms
exactly to everyday use or accords with the words etymology; what does matter is that
the word is as well suited as possible for expressing laws. The better suited an apparatus of
technical terms is, the more briefly it will be able to render exactly the complete system of
laws. (Frege 1897=Frege 1969, 148/Frege 1979, 1367)

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

conjunction, disjunction, and negation correctly codify our ordinary standards

for evaluating deductive arguments, but argue that those standards, and hence
our disposition to deduce in accordance with them, must be revised.
When two logical schools differ in this more radical way, does it follow that
they mean different things by the connectives? Dummett claims that it does. Our
fundamental logical laws, he writes, are those which it is an essential part of
our practice in speaking the language to observe. The view that revision of them
involves a change in the meanings of the logical constants is unshakable (1991,
302). Although they stand on opposite sides of the first-order dispute about the
respective merits of classical and intuitionistic logic, this is a point of agreement
between Dummett and the Quine of Philosophy of Logic. The intuitionist, Quine
claims there, should not be viewed as controverting us [sc., us classicists] as to the
true laws of certain fixed logical operations, namely, negation and alternation.
He should be viewed rather as opposing our negation and alternation as unsci-
entific ideas, and propounding certain other ideas, somewhat analogous, of his
own (1986, 87). Quine is clear that this view of the character of the dispute does
not preclude a debate. This is not to say that he is wrong in doing so. In repudiat-
ing p or ~ p, he is indeed giving up classical negation, or perhaps alternation, or
both; and he may have his reasons (1986, 85). However, those reasons must either
be grounds for challenging the correctness of the classical logicians description
of the connectives customary senses, or grounds for recommending new senses.
Quine never seriously contemplates the possibility that the classical logician has
misdescribed the customary senses of and, or, or not; accordingly, he supposes
that the intuitionist must be recommending novel senses for these connectives.
Here, evidently, is the deviant logicians predicament:when he tries to deny the
doctrine he only changes the subject (1986, 81).
Quines argument for this conclusion is unimpressive. It amounts to little
more than the observation that if someone were to propound a heterodox logic
in which all the laws which have up to now been taken to govern alternation were
made to govern conjunction instead, and vice versa ... we would regard his devia-
tion merely as notational and phonetic. For obscure reasons, if any, he has taken
to writing and in place of or and vice versa (1986, 81). No doubt, in the case
described, we would simply reinterpret the deviant logicians speech in this way,
but the issue is precisely whether the same goes for other cases of logical deviancy.
The intuitionist, after all, is not obviously misusing the word not. He uses it just
as the classicist does in making statements that express perceptual judgements
(It is not raining) or judgements founded on induction or abduction (His car
is not in the garage, so he is not at home). Even in deductive argument, his devi-
ation only becomes manifest in some fairly recherch instances, and the most
20 Introduction

familiar classical translation of his deviant talkthe provability interpretation,

proposed by Gdel (1933b), developed further by McKinsey and Tarski (1948), and
then extended to first-order languages by Rasiowa and Sikorski (1953)makes
little sense when applied to uses of not (and the other connectives) outside
mathematics.
Dummetts argument for the same conclusion is much more interesting.
Revision of laws involves a change in the meanings of the logical constants, he
thinks,

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.

1.3 The Argument of the Present Book

It may help the reader if Ioutline the argument to be advanced in this book.
Part I, which comprises Chapters2 and 3, proposes an account of what logic
is. Ibegin by presenting a general characterization of logical laws. On my view,
which emerges from an account of the nature and epistemic value of logic, its
laws are the general laws of implication relations, where an implication relation
is one which obtains between some premisses and a conclusion when the truth of
the premisses guarantees (in the contextually appropriate sense) the truth of the
conclusion.
What, though, does guarantee mean here? Against Russell and others, Iargue
in Chapter3 that the notion is implicitly modal:when the truth of the premisses
guarantees the truth of the conclusion, there is no possibility (in the contextually
relevant set of possibilities) at which all the premisses are true but the conclusion
is not true. Ifurther argue that each of these spaces of possibilities is a restriction
of a master space of logical possibilities, and explore the relationships between
this notion of logical possibility and those of metaphysical possibility and know
ability a priori.

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

With this preliminary characterization of the business of logic in place, Iturn

in Part II to examine five attacks on classical logic. Classical logic has been
attacked from many directions; some, notably, have been led to resile from full
acceptance of it by reflecting on paradoxes such as the Liar and the Heap. Whilst
Idiscuss the Heap in Chapter8, this book focuses primarily on attacks that rest
22
upon premisses from the theory of meaning. Precisely because they connect to
theses in the philosophy of language, the attacks Iconsider resonate far beyond
the philosophy of logic. The relations between them also make the present book
more than a collection of independent papers. That said, by presenting the attacks
in close proximity Ihope to make clearer the differences between them. Whilst
the arguments Iconsider may all be labelled as anti-realist attacks on classical
logic, each has its own distinctive character. Many discussions of the anti-realist
critique of classicism fail to discriminate between lines of argument that are
really distinct, and confusion of this kind has surely impeded a proper evaluation
of that critique. Ihope this book may help to remove that impediment.
I start, in Chapter4, by scrutinizing the attack on classical logic in Michael
Dummetts early paper Truth (Dummett 1959). That attack has the follow-
ing structure. By reflecting on the notion of a statements content, and in par-
ticular by examining the relationship between its content and its truth or falsity,
Dummett claims that we may come to know that no statement is neither true
nor false, unless it is ambiguous or vague:thus no statement whose sense is fully
determinate is neither true nor false. In Dummetts terminology, philosophical
reflection enables us to know the semantic principle of Tertium non datur. Now, if
classical logic could be applied to semantic discourse, a simple deduction would
enable us to move from our knowledge of Tertium non datur to knowledge of
the Principle of Bivalencethat is, to the knowledge that every determinate state-
ment is either true or false. Dummett argues, however, that other conceptual con-
straints on the notions of truth and falsity mean that we cannot know the truth of
Bivalence. We can know that a statement is either true or false only when it is of
such a kind that we could in a finite time bring ourselves into a position in which
we were justified either in asserting [it] or in denying [it] ... This limitation is not
trivial:there is an immense range of statements which ... fail the test (1959, 1617;
Ianalyse Dummetts examples of such statements in Chapter4). But if Tertium

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:

(R)The truth-grounds of any statement form a closed set of possibilities.

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

(B) Any statement has a back.

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

Infinitesimal Analysis (SIA). Because SIA contemplates quantities whose iden-

tities are indeterminate, some statements of SIA are backless. As a consolation
prize to the intuitionist, Iallow that his logic may be the strongest that we are enti-
tled to use when engaged in deductive reasoning about indeterminate entities.
In Chapter8 Iturn to common-or-garden vague statements. Ibegin by pre-
senting a version of the ancient Paradox of the Heap and outline Crispin Wrights
recent argument, in his essay Wangs Paradox (2007), that it is best resolved
by switching from classical to intuitionistic logic. Wright neglects to provide a
semantic theory for vague predicates that explains why intuitionistic logic is the
correct logic to use when reasoning with them. Developing some ideas from the
theory of rough sets, Isketch a semantic theory that seems to give Wright what
he needs. Iargue, though, that the assumptions that this theory makes about
vague terms are not, in the end, tenable. As an alternative in the same line of
country, however, Igo on to develop Mark Sainsburys idea that vague predicates
are boundaryless:a speaker who uses such a predicate does not even aim to draw
a line between positive and negative cases; rather, he classifies objects by reference
to positive and negative paradigms or poles. Ishow how this semantic theory vin-
dicates classical logic when reasoning with vague terms; it also exposes the flaw
in the version of the Paradox of the Heap that seemed to motivate deviating from
classical logic in the first place.
Many philosophers have held that special problems attend the use of classi-
cal logic in set theorymore particularly, in deductions involving quantification
over absolutely all sets. In Chapter9, Ipresent the most detailed argument Iknow
that classical logic cannot be applied to such deductions, that given by William
Tait in his paper Zermelos Conception of Set Theory and Reflection Principles
(Tait 1998). Of all the challenges to classical logic considered in this book, Taits
is the hardest to answer. Taits argument shows, Ithink, that we need a radically
non-classical semantic theory in order to characterize the senses of statements
that quantify unrestrictedly over sets. For all that, though, Istill contend that
classical logic is the right logic to use when assessing deductions involving such
statements. The somewhat backhanded justification that Igive of classical logic
in this arena brings to the fore, for the first time in the book, a technique that may
have wider application, namely, the use of the negative translations from clas-
sical languages into intuitionistic ones that were pioneered by Kolmogorov and
then refined by Gdel and Gentzen.
In their various ways, Chapters5 to 9 show how classical logic may be vindi-
cated against anti-realist attacks without appealing to Bivalence. There is, though,
a simple argument, due in essentials to Aristotle, which seeks to lumber a classi-
cal logician with the obligation to defend Bivalence. More exactly, the argument
 Introduction 27

purports to derive Bivalence, using classical logic, from apparently compelling

principles about truth and falsity. This argument poses an indirect challenge to
classical logic. The bivalence of certain statements is highly doubtful. Many phi-
losophers are reluctant to assert, for example, that a statement in which a vague
predicate is applied to one of its borderline cases is either true or false. Again, many
set theorists are reluctant to assert that the Generalized Continuum Hypothesis
is bivalent. Given the apparently uncontentious nature of the Aristotelian prin-
ciples about truth and falsity, reasons for doubting Bivalence seem to be reasons
for resiling from full acceptance of classical logicfor example, by withdrawing
from unrestricted acceptance of the Law of Excluded Middle.
In Chapter10, Iexplain how the analysis of the previous chapters exposes flaws
in the Aristotelian argument for Bivalence, thereby neutralizing these sources of
doubt about classical logic. One may consistently combine adherence to classical
logic with denial of Bivalence. In fact, Iconclude, this is what we ought to do. As
well as being a defence of classical logic, the book brings out how restrictive is
classical semantics. With the threat of logical deviancy allayed, the path is open to
apply some of our non-classical semantic theories to constructions more compli-
cated than the connectives. That the theories treat the connectives satisfactorily,
though, shows that they pass a vital first test.
In writing a book of this character, the author needs to decide how much
mathematical detail to include. Ihave kept the formal substructure to a mini-
mum in order to make as much space as possible for philosophical discussion.
The references, though, should lead an interested reader to full proofs of all the
mathematical claims that my argument needs. Ihave also includedmainly in
footnotesproofs of some of the basic facts about the semantic models that Iuse.
Those models draw on lattice-theoretic and topological results that are not as well
known among philosophers as they ought to be. Many of the proofs are short and
simple, and Ihope thereby to encourage philosophers to explore alternatives to
the familiar, but often too restrictive, possible-worlds semantics.
PA RT I

The Nature of Logic

 2
Logical Laws

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.

In a similar spirit, E.J. Lemmon begins Beginning Logic by writingthat

logics main concern is with the soundness and unsoundness of arguments ... Typically,
an argument consists of certain statements or propositions, called its premisses, from
which a certain other statement or proposition, called its conclusion, is claimed to follow.
We mark, in English, the claim that the conclusion follows from the premisses by using
such words as so and therefore between premisses and conclusion ... Logicians are
concerned with whether a conclusion does or does not follow from the given premisses. If
it does, then the argument in question is said to be sound; otherwise unsound. (Lemmon
1965, 1)

Both of these passages presuppose that we have some pre-theoretical grasp of

the relationship whereby one thing (statement, proposition) follows from some

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

are well suited to serve as relata of a consequence relation. Consequence is usu-

ally understood to involve the preservation of truth, and on this understanding
of the matter, a statement may be classified as true or false simpliciter accord-
ing as the associated thought is true or false. More precisely, if a statement u
expresses the thought that P, then u is true if and only if P, and u is false if and
only if it is not the case that P. The obscurities that attend Lemmons explanation
are, for the moment, decently veiled:whatever may be said about the thoughts
they express, the statements There is a choice function on every non-empty set
and Every set can be well-ordered are evidently distinct, for their first members
are distinct sentences. The quantification over all possible contexts of utterance
ensures that statements are not confined to actual utterances or inscriptions, so
laws concerning all statements exhibit the generality we expect of logical laws.
Admittedly, we shall not have determined the precise range of statements until
we have enumerated all the features of an utterances context that can bear on its
truth or falsity. Any indeterminacy there, however, will not affect the arguments
to follow, so there is no call to list the relevant features on this occasion.
The premisses and conclusion of a single argument may be understood to be
statements with a common second element. We assume, in other words, that the
context is held constant throughout an argument. This seems to be a presupposi-
tion of the logical appraisal of ordinary arguments. In assessing the argument I
am taller than you. So you are shorter than me, a logician is not expected to take
account of the possibility that our relative heights might have changed between
the utterance of the premiss and that of the conclusion. Nor is the argument an
enthymeme because it takes for granted that there has been no shift in reference
3
between the two occurrences of the pronoun you. Since we are assuming that
consequence goes with argumentative soundness, this means we need only con-
sider consequence as it relates statements which share a context of utterance.
Despite their differences over the nature of premisses and conclusions, Mates
and Lemmon are at one in appealing to an antecedent understanding of conse-
quence in characterizing the logicians task. When Mates describes logic as inves-
tigating the relation of consequence that holds between the premisses and the
conclusion of a sound argument, he implies, or presupposes, that there is some
single, uniquely favoured relation of consequence which has a special claim on

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.

2.2 Inference and Deduction

The approach Ihave in mind focuses not on a relationship, consequence, but on
an activity, inferring. Aleading exponent of this alternative was Gilbert Ryle.
According to Ryle, logic centrally comprises formulations of rules of inference or
consistency rules (1946, 236). These rules of inference, like the rules of grammar,
chess, etiquette and military funerals, are performance-rules (1946, 238). That is
to say, they regulate a certain sort of performance:references to them are refer-
ences to criteria according to which performances are characterised as legitimate
or illegitimate, correct or incorrect, suitable or unsuitable, etc. (1946, 238). Ryle
had various terms for the kind of performance which the logicians rules serve
to regulate, but the most common (unsurprisingly) was inference. The relevant
species of legitimacy or correctness is validity:a breach of a rule of logic is a fal-
lacy; an observance of it is a valid inference. To speak of an inference as an obser-
vance or as a breach of a rule of logic is only a condensed way of saying that the
author of the inference has made his inference in conformity with or in breach of
a rule of inference (1946, 238).
 Logical Laws 35

If we take the logicians rules of inference to be performance-rules, it might

seem inevitable that the relevant performances should be inferences. In fact,
though, if the word inference is taken in its primary ordinary sense, the claim
that it is inferences that inference-rules regulate is a mistake. The basic prob-
lem is that an inference is not a performance:unlike the railway journeys that
fascinate Ryle in If, so, and because (Ryle 1950), inferences do not take
time, nor are they subject to intentional control. Moreover, pace Ryles position
in The Concept of Mind (see Ryle 1949, 3023), an inference is not an achievement
of, or an arrival at, a result. An achievement must be something that an agent
can try to attain, but it makes no sense to say Try to infer It is either raining or
snowing from It is raining. Alan White got much nearer the mark when he
wrote:To infer is neither to journey towards, nor to arrive at or be in a certain
position; it is to take up, to accept or to change to a position. Inference is not the
passage from A to B, but the taking of B as a result of reflection on A (White 1971,
291). At any rate, this gloss captures the focal sense of infer, and Ishall use the
term strictly in this sense.
All the same, there is a species of intellectual activity that the logicians rules
can be thought of as regulating. Sometimes, a thinker engages in the task of
tracing out the implications of some premisses. Sometimes, indeed, he does
this step by step, taking special care to move only to conclusions that the prem-
isses really imply. Let us call this activity deduction. Unlike inferences, deduc-
tions do take time, and they are subject to intentional control. They can also
be achievements: an examination question might sensibly instruct Deduce
Gdels Second Incompleteness Theorem from Lbs Theorem, and a candi-
date might sensibly report I tried to do that but failed. This sort of intellectual
activity is rare in everyday life, but it is central to any discipline (such as math-
ematics, the sciences, and indeed philosophy) where it is important to draw
out the implications of hypotheses in a manner that prevents non-implications
from creeping in. Insofar as the term inference-rules suggests that the rules
of logic regulate inferences, it is misleading: deduction-rules would have
4
been better.
On this way of understanding the terms, there are many cases where B is
inferable, but not deducible, from A. Indeed, there are cases where B is inferable
from A (but not conversely) while A is deducible from B (but not conversely).
White again: We can contrast From your silence I infer that you have no
objections with From your lack of objections Ideduce that you will remain

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

someones knowing it. To be sure, we sometimes come to know a conclusion by

deducing it from premisses that we already know, and in 2.6 Ishall try to explain
how we can gain knowledge in this way. However, not all successful deductions
start from known premisses. Deduction is a matter of drawing out the implica-
tions of premisses, whether or not we know those premisses, and whether or not
7
they are true.

2.3 The Varieties of Deduction and of

Implication Relations
The implications that deduction draws out are not always logical consequences of
the premisses. Consider a young girl who is learning about electrical circuits. She
is presented with the circuit diagram in Figure 2.1 and asked to say if the lamp L
is on or off.
She is not told whether the switch S is open or closed, so she reasons as follows:
Either switch S is open or closed. Suppose first that S is open. Then, on that supposition,
no current flows in any part of the circuit, so lamp L will in this case be off. Suppose on
the other hand that S is closed. In that case, the current will flow around the sub-circuit
A, B, C, D. However, L is not part of that sub-circuit, so L will again be off. Either way,
then, L is off.

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

Figure2.1 A simple circuit diagram.

circuits obtain. Or again, a premiss stands in this implication relation to a conclu-

sion if there is no electrical possibility that the premiss should be true without the
conclusions being true. Alogical possibility need not be an electrical possibility.
It is logically possible that lamp L should light up even though it is not connected
to any electrical cell.
Those philosophers who suppose that we have some pre-theoretical grasp
of a favoured relation of properly logical consequence have of course recog-
nized that that relation does not always provide the standard for assessing the
soundness of ordinary arguments. However, they typically try to explain away
our tendency to classify as sound arguments that do not meet their canonical
standards of logical validity by deeming such argument to be enthymemes
whose fully explicit statement would reveal a tacit or suppressed premiss. Thus
I.M.Copi:

Because it is incomplete, an enthymeme must have its suppressed premiss or premisses

taken into account when the question of its validity arises. Where a necessary premiss
is missing, the argument is technically invalid. But where the unexpressed premiss
is easily supplied and obviously true, in all fairness it ought to be included as part of
the argument in any appraisal of it. In such a case one assumes that the maker of the
argument did have more in mind than he stated explicitly. In most cases there is no
difficulty in supplying the tacit premiss that the speaker intended but did not express.
Thus Al is older than Bill. Bill is older than Charlie. Therefore Al is older than Charlie
ought to be counted as valid, since it becomes so when the trivially true proposition
that being older than is a transitive relation, is added as an auxiliary premiss. (Copi
1973, 132)

It is sometimes right to appraise an argument as though it contained an addi-

tional premissone which its proponent omitted to state. However, the strat-
egy of postulating unexpressed premisses does not provide a satisfactory general
explanation of our intuitive assessments of the soundness of arguments. Those
40 The Nature of Logic

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

relation of properly logical consequence were taken to be a higher-order relation,

we could still account for our ability to latch onto the consequence relation of
first-order logic, although of course we would do that by imposing a syntactic
restriction, not by making explicit any suppressed premises:the valid first-order
consequences are the valid higher-order consequences whose premisses and con-
clusions are first-order formulae. All the same, it remains a nice question for the
defender of the enthymematic strategy exactly which higher-order logic gives the
true underlying relation of properly logical consequence.
There is a more philosophical objection to the enthymematic strategy:the lack of
any persuasive motivation for butchering the surface structure of arguments in the
way that it requires. Our ordinary assessments of argumentative soundness rely on
our ability to latch onto the implication relation that is relevant in the arguments con-
text. Having latched onto that, we can appraise many arguments more or less as they
come. No doubt some are best appraised by postulating an additional premiss which
the proponent intended but did not express. Equally, though, there will be many con-
texts in which we do not need to postulate any additional premiss in order to account
for the soundness of Al is older than Bill. Bill is older than Charlie. Therefore Al is
older than Charlie. We need to ask, then, what philosophical principle is supposed
to sustain the claim that this argument is strictly speaking unsound as stated, and
needs to be supplemented by a premiss expressing the transitivity of being older than.
The features of properly logical consequence that are usually cited to mark out
its special status do not by themselves sustain this claim. Some like to say that
the logical consequences of premisses are implicit in them. What they mean is
murky, but on any reasonable explication, Als being older than Charlie is implicit
in his being older than Bill and Bills being older than Charlie. Certainly, these
premisses necessitate Als being older than Charlie. Others will say that failure
to accept at least the obvious logical consequences of a statement is a sign that
one does not properly understand it. But the same goes for any of a statements
obvious consequences, whether or not those consequences are deemed to follow
9
logically. In particular, someone who accepts that Al is older than Bill, and that

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:

Reflexivity needs little justification. The relations we are trying to characterize

are those where the truth of the premisses guarantees that of the conclusion
even if the guarantee is redeemable only when certain conditions are met. In a
case where the conclusion is the premiss, such a guarantee obtains, whatever the
conditions might be. There is little more to say.
Monotonicity needs more discussion. The principle may seem incontestable:if
the truth of X guarantees that of B, then surely the truth of X and of A does too.

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

Indeed, Monotonicity seems to mark the difference between cases in which a

guarantee obtains and those where the premisses merely provide inductive sup-
port for a conclusion. Premisses saying that a large number of pure samples of
bismuth melt at 271C support the generalization that all pure samples of bismuth
melt at that temperature; not so if those premisses are supplemented by another
saying that a further pure sample melts at 261C.
For all that, freestyle deductive reasoning contains many apparent breaches
of Monotonicity. Consider This triangle is right-angled; so the square on its
largest side is the sum of the squares on the other two sides. In many contexts,
that counts as a sound deduction. However, the deduction This triangle is
right-angled; its internal angles sum to more than two right angles; so the square
on its largest side is the sum of the squares on the other two sides is unsound. So
we have an apparent counterexample to Monotonicity. Ithink, though, that the
counterexample is only apparent:what the case really shows is how flexibly the
freestyle reasoner moves from one implication relation to another. If we judge
the first deduction to be sound, that will be because the contextually relevant
implication relation is one in which the relevant possibilities are confined to
those in which the laws of Euclidean geometry hold good. In the second deduc-
tion, however, the second premiss forces a shift to a different implication rela-
tion. In a context where it is not assumed that a triangles internal angles sum
to 180 degrees, the Euclidean implication relation is inapplicable. Accordingly,
in assessing the second deductions soundness, we must do so against another
implication relation which takes account of possibilities not contemplated in
Euclidean geometry. So far from being a counterexample to Monotonicity, then,
the case illustrates the way in which the operative implication relation is sensi-
tive to the context of the argument.
As for the Cut Law, our deductive practice seems to presuppose it. Suppose that
we correctly deduce, from some premisses X, each statement in a set Y. Suppose
further that we correctly deduce a statement A from Y. The common practice of
breaking down arguments into various lemmas that can be assembled to yield
the eventual conclusion presupposes that these deductions may be combined to
form a correct deduction of A from X. However, that presupposition will be ful-
filled only where the Cut Law holds. Where is the relevant implication relation,
the first deduction shows that X B for all B in Y; the second shows that Y A.
Only if X A will the composite deduction of A from X be correct. As with
Monotonicity, analogues of Cut fail for other relations in which premisses pro-
vide some support for a conclusion. The truth of A may make that of B 99% likely,
and the truth of B may make that of C 99% likely, without the truth of A making
that of C 99% likely. The Cut Law, then, helps to distinguish statements that are
44 The Nature of Logic

properly implications of premisses from statements that those premisses make

likely to a given degree.
Let us call a relation among statements implicative if it meets the conditions
of Reflexivity, Monotonicity, and Cut. These structural conditions isolate a theo-
retically tractable group of relations. There is, however, a further, non-structural
property which we also expect any implication relation to possess. Let us call a
relation R among statements truth-preserving if, whenever some premisses stand
in R to a conclusion, and those premisses are true, the conclusion is also true. As
with the structural properties, our deductive practice presupposes that the rela-
tion to which a given deduction answers will be truth-preserving in this sense.
As we saw earlier, not all deductions issue in inferences, but some do:having
correctly deduced a conclusion from some premisses that we accept as true, we
take ourselves to be entitled to affirm the conclusion on the basis of the prem-
isses. This would be wholly unwarranted unless the pertinent relation were
truth-preserving. Because of its rather different logical character, though, Isepa-
rate truth-preservingness from the structural features that are definitive of an
implicative relation.
In fact, some implicative relations may be proven to be truth-preserving, and
this generates a prima facie problem for the Cut Law. Let us say that a conclusion
is a Philonian consequence of some premisses if either the conclusion is true or at
least one of the premisses is nottrue:

X Philo B if and only if either some member of X is not

(Philo)
12
true or B is true.

The relation of Philonian consequence is clearly truth-preserving:if a conclu-

sion is a Philonian consequence of some premisses, and all those premisses are
true, then the case in which some premiss is not true is excluded, so the conclu-
sion must also be true. Furthermore, given a classical metalogic, Philonian con-
sequence qualifies as an implicative relation. By the classical Law of Excluded
13
Middle, either A is not true or A is true, so A Philo A, i.e. Reflexivity is satisfied.

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.

2.4 Implications and Possibilities

In further glossing the notion of an implication relation, it is natural to invoke a
restricted space of possibilities. Thus A stands to B in our electrical implication
relation if, for every electrical possibility, if A is true at that possibility then B is
true there too; A stands to B in the first of our geometric implication relations if,
for every possibility admitted by Euclidean geometry, if A is true at that possibil-
ity then B is also true there; and so forth. Quite generally, to each space of possi-
bilities, , there corresponds an implicative relation as follows:

(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:

For every A1,...,An and B in S, A 1,...,An T B if and only if either

some Ai K or B L.
16
It is easy to verify that T satisfies Reflexivity, Monotonicity, and Cut. Indeed,
the bisection implication associated with T=<K, L> is the relation of Philonian
consequence in which the members of L are taken to be true and the members of
K are taken not to be true. Accordingly, we may think of T=<K, L> as a valuation
of the statements in S in which all and only the members of L are valued as true.
Let us call a subset R of S closed under an implicative relation if, for every
A1,...,An and B in S, B R whenever A 1,...,An B and each Ai R. It is easy to
check that, for any bisection T=<K, L>, the set L is closed under the bisection
17
implication T associated with T. Suppose now that we are given an implicative
relation on S, i.e. a relation which meets the three Tarskian structural condi-
tions. Then we say that T=<K, L> is a bisection of the relation on S if L is closed
under . Any bisection of is an extension of . For suppose that A1,...,An B
and consider an arbitrary bisection T=<K, L> of . Since L is closed under ,
if every Ai L then B L. So either some Ai L or B L. Since T is a bisection, if
Ai L, Ai K. Hence either some Ai K or B L. That is, A 1,...,An T B. Bisections
of , then, may equally well be called bisective extensions of .

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

The Lindenbaum-Scott theorem saysthis:

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 .

Proof. By the argument of the previous paragraph, whenever T is a bisection of

, then is contained in T. So we already have the only if part of the theo-
rem. Conversely, let us assume that A 1,...,An T B for all bisections T of . We
need to show that A 1,...,A n B. Suppose for a contradiction that this were not
so. Let U be the set of all C in S for which A 1,...,An C. By the Reflexivity and
Monotonicity of , each A 1,...,Ai is in U, so U is non-empty. By the reductive
hypothesis, B U, so S U is also non-empty. Thus T*=<S U, U> is a bisec-
tion of S. Also, U is closed under . For suppose that B 1,...,Bm C where each
Bj U. By the definition of U, it follows that A 1,...,A n Bj for each Bj; by applying
Cut n times, we infer that A 1,...,An C, so that C U. Thus T* is a bisection of
. Now A 1,...,A n T B for all bisections T of . So in particular A 1,...,An T*B.
However, A 1,...,A n T*B holds if and only if either some A i SU or B U.
By definition, though, each Ai U, and by the reductive hypothesis B U. This
contradiction means that the reductive hypothesis is refuted. So A 1,...,An B,
as required.
The proof just given proceeds by Classical Reductio: having derived a con-

tradiction from the hypothesis A together with background premisses X,
we deduce A from X. For an intuitionist, assuming Classical Reductio is tanta-
mount to assuming Excluded Middle. In this case, it is tantamount to assuming
that either A 1,...,An B or not A 1,...,An B, for arbitrary statements A 1,...,An
and B. For some implicative relations, an intuitionist will be unable to accept this
assumption. Where is Philonian consequence, for example, a special case of
the assumption is that either Philo B or not Philo B, i.e. that B is either true or
not true. (In fact, as remarked in n.13, the intuitionist will not even accept that
Philo is an implicative relation: the same assumption is needed to show that
Philo is reflexive.) However, there are many interesting implicative relations for
which the intuitionist is able to accept the assumption that is needed to yield the
Lindenbaum-Scott theorem. When P is either true or not, let us say that it is a
determinate matter whether P. An intuitionist will not regard the truth of an arbi-
trary mathematical statement as a determinate matter, but he may regard it as a
determinate matter whether a given conclusion stands to given premisses in the
implication relation Math that provides the standards for assessing a particular
mathematical deduction. We shall then have that either A 1,...,An Math B or not
A1,...,An Math B so that the Lindenbaum-Scott Theorem applies to Math. There
 Logical Laws 49

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:

1{Body b is iron; Body b is in a magnetic field; Aforce acts on body b} K

2{Body b is iron; Body b is in a magnetic field} K, {A force acts on body b} L
3{Body b is iron; Aforce acts on body b} K {Body b is in a magnetic field} L
4{Body b is in a magnetic field; Aforce acts on body b} K, {Body b is iron} L
5{Body b is iron} K, {Body b is in a magnetic field; Aforce acts on body b} L
6{Body b is in a magnetic field} K, {Body b is iron; Aforce acts on body b} L
7{Body b is iron; Body b is in a magnetic field; Aforce acts on body b} L.

But they do not include:

8 {A force acts on body b} K, {Body b is iron; Body b is in a magnetic field} L.

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

2.5 The Role of Logic

How does a thinkers specifically logical capability relate to these various deduc-
tive capacities?
In addressing this question, it helps to compare the argument about electri-
cal circuits given at the start of 2.3 with the following deduction, which shows
that the result of adding an arbitrarily chosen natural number n to its own square
iseven:
Either n is even or n is odd. Suppose first that n is even. Then n=2k, for some natural
2 2
number k. So n + n=n(n + 1)=2k(2k + 1), which is divisible by 2.So in this case n + n is
even. Suppose on the other hand that n is odd. Then n=2k + 1, for some natural number
2 2 2 2 2
k. So n + n=(2k + 1) + (2k + 1)=4k + 4k + 1 + 2k + 1=4k + 6k + 2=2(2k + 3k + 1), which
2 2
is again divisible by 2.So in this case too n + n is even. Either way, then, n + n is even.

This deduction is a mathematical deduction: the reasoning tacitly appeals to

several principles of elementary algebra. The implication relation to which the
reasoning answers is one in which some premisses stand to a conclusion when
there is no mathematical possibility of their being true without its being true. As
before, this relation is not logical entailment:pace the logicists, there are logical
possibilities that are not mathematical possibilities. Rather, it is a relation whose
extension is determined by specifically mathematical laws.
Our two deductions share a dilemmatic form, and owe their soundness in
part to their having that form. But how does a thinkers mastery of dilemmatic
argument help him to produce sound deductions? The naturaland, Ithink,
correctanswer runs as follows. In each of our cases, the thinkers possession
of a certain topic-specific deductive capacity enables him soundly to deduce a
conclusion from each of two premisses. His mastery of dilemmatic argument
then enables him to splice these two deductions together so as to produce a
new sound argument whose premiss is the disjunction of the premisses of its
components. The new, composite argument is in each case as topic-specific as
its parts:in the first case, it is a deduction in the theory of electric circuits; in
the second, it is a mathematical deduction. Athinkers logical competence, one
might say, consists in an ability to splice together deductions in various fields to
produce new, more complex, deductions in those same fields. Logical compe-
tence, on this view, is a higher-order intellectual capacity:its application yields
new deductive capacities from old.
A thinker will possess this higher-order capacity if, in producing new deduc-
tive capacities from old, he reliably conforms to certain rules. If the colon of
the sequent is understood as in 2.4, these rules are well formalized as the rules
of a sequent calculusmore exactly, as the rules of a sequent calculus with
 Logical Laws 53

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:

(2)Whatever implication relation R may be, if X, A stand in R to C, and Y, B

stand in R to C, then X and Y together with any disjunction of A with B
also stand in R to C.

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

The correctness of rule (3)presupposes the followinglaw:

(4)Whatever implication relation R may be, if X together with B stands in R

to C, and Y stands in R to A, then X and Y together with any conditional
whose antecedent is A and whose consequent is B will stand in R to 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:

(7)Whatever implication relation R may be, any pair of statements

comprising A together with the conditional statement whose
antecedent is A and whose consequent is B stands in R to B.

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

standards). Some deductions will be sound, though, whatever implication relation

provides the standard for assessing soundness. The conclusion of such a deduc-
tion may be said to follow logically from its premisses. If each step in a deduction
follows logically from the premisses or assumptions cited to justify the step, then
the whole deduction may be said to be logically valid. From (6), we have that, what-
ever implication relation sets the standard for assessing soundness, a deduction
by Modus Ponens is sound. So the present account yields the reassuring conclu-
sion that in an instance of Modus Ponens the conclusion follows logically from
the premisses. Gentzens way of formalizing logic has accustomed people to the
idea that logical truths are simply the by-products of logical rulesby-products
that arise when all the assumptions on which a conclusion rests have been dis-
22
charged. Our analysis has taken us further in the same direction. On the concep-
tion Iam recommending, the classification of deductions as logically valid is itself
a by-product of yet more general principles which tell us which deductions stand
or fall together when assessed against a given implication relation. Ascriptions of
23
logical validity are just a limiting case of this wider, relational concern.

2.6 Knowledge by Deduction

Ryle was wrong, Iargued earlier, to say that a deduction must start from facts,
letalone from known facts. But we sometimes deduce conclusions from prem-
isses that we know, and we value our deductive capacities in part because we can
gain knowledge by applying them. So we need to consider the role that deduction
plays in expanding our propositional knowledge.
There are at least two problems in this area. One is to explain how the exercise of a
deductive capacity can yield new knowledge. Ishall call this Mills Problem, for the
sense of puzzlement that prompts the demand for explanation finds its most power-
ful articulation in Mills System of Logic. We have now to inquire, he writesthere,
whether the syllogistic process, that of reasoning from generals to particulars, is, or is
not a process of inference [that is,] a process from the known to the unknown:a means of

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)

This leads Mill to conclude that in every syllogism, considered as an argument

to prove the conclusion, there is a petitio principii (II iii 2). The problem is then to
reconcile this apparently irrefragable doctrine with the indisputable fact that we
do sometimes gain knowledge by deduction.
There is much to be said by way of explaining how the exercise of our deduc-
tive capacities can generate knowledge. Freges masterly account, in Booles
rechnende Logik and die Begrifsschrift (Frege 1880/81), of how deductions
involving quantifiers enable us to discern new predicables in statements that
we already understand provides only part of the story (see Dummett 1981b,
chapters15 and 16). All the same, Mills argument that such generation is impos-
sible need not detain us for long:it owes such plausibility as it possesses to an
equivocation between two things that one might mean by there is no more in
the conclusion than was assumed in the premisses. In the next chapter, Ishall
defend the Aristotelian thesis that, when a conclusion follows logically from
some premisses, there is no logical possibility at which the premisses are all
true and the conclusion is not true. This articulates one sense in which a con-
clusion says no more than has already been said, collectively, by the premisses.
However, nothing follows from this about knowledge. We may stipulate that it
is epistemically possible that P (for a thinker A) when A does not know that not
24
P. In these terms, when there is no epistemic possibility that P, the relevant
thinker already knows that not P, so (given classical logic) an argument will
be nugatory if there is no epistemic possibility at which the premisses are all
true and the conclusion is not true. However, in order to move from There is
no logical possibility where the premisses are all true and the conclusion is
not true to There is no epistemic possibility where the premisses are all true

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:

(DP)If a thinker knows some premisses, and comes to believe a conclusion

by competently deducing it from those premisses, while retaining
knowledge of the premisses throughout the deduction, then he knows
26
the conclusion.

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

knowing the premisses (1)and (2)above, or from being deductively competent

in the specified sense, so long as the brainstorms have little chance of happening.
However, certain events can be individually unlikely without its being unlikely
that at least one of them will happen. On this view, then, an additional condition
needs to be met before we can infer that my conclusive beliefmy belief that the
street is wetis knowledge. It is not enough that each individual brainstorm is
unlikely to have occurred. There must also be a very low chance that at least one of
the brainstorms should have occurred.
How should we react to these cases? We could restrict the original Deduction
Principle so as to exclude the apparent counterexamples. If the restriction is along
the lines just suggested, some beliefs deduced from known premisses will still qual-
ify as knowledge. On the other hand, we could resist the claim that the cases lately
described are counterexamples to the Deduction Principle. They certainly put the
Principle under some intuitive pressure, but before accepting them as counterex-
amples, we should weigh the theoretical costs of restricting the Principle against
those of resisting the counterexamples. This in turn suggests that we should con-
sider what the grounds of the Deduction Principle might be. Once identified, those
grounds ought to show what restrictions, if any, the Principle needs.
A first shot at grounding the Deduction Principle might run like this. Suppose we
are in an argumentative context in which a truth-preserving implicative relation,
R, sets the standard for assessing deductions. If a thinker qualifies as deductively
competent in that context, he will be disposed to deduce a conclusion from some
premisses only when the premisses really do stand in R to the conclusion. Now sup-
pose that a deductively competent person knows the premisses of an argument,
and deduces its conclusion from those premisses. Because he knows the premisses,
those premisses are true. Because the premisses are true, and they stand in R to
the conclusion, the conclusion is also true. (It is given that R is truth-preserving.)
Ex hypothesi, our thinker is deductively competent, so he will deduce a conclu-
sion from some premisses only when they R-relate to the conclusion. Accordingly,
when a deductively competent thinker deduces a conclusion from premisses that he
knows, the belief thereby formed will be true, and it will have been produced in a
way that reliably yields true beliefs. Suppose finally that we accept a reliabilist con-
ception of knowledge. On that conception, what endows a true belief with the status
of knowledge is precisely that it has been produced in a way that reliably yields true
beliefs. So, assuming a reliabilist conception of knowledge, belief in a conclusion
that has been competently deduced from known premisses will have the status of
knowledge, so that the Deduction Principle is unrestrictedly true.
As we shall see, the proponent of this argument puts his finger on something
important when he focuses on the connection between deductive competence
62 The Nature of Logic

and implication. However, the argument as it stands is vulnerable to an objection,

even if we accept the reliabilism needed at the last step. The objection is that the
putative explanation of the truth of the Deduction Principle cannot be right, for if
correct it would explain too much. In the explanation, the only use that is made of
the hypothesis that the thinker knows the premisses of his deduction is to derive
the claim that those premisses are true. So the same account would appear to
explain the truth of the Pseudo-Deduction Principle:
(PDP)If a thinker comes to believe a conclusion by competently deducing it
from true premisses that he believes, then he knows the conclusion.

The Pseudo-Deduction Principle, however, is patently absurd. Coming to believe

a conclusion by competently deducing it from some true beliefs reliablyindeed,
infalliblyyields a true belief. But it clearly does not constitute a reliable way of
producing true beliefs in the sense that is needed for a belief thereby produced to
qualify as knowledge. For PDP does not require that the thinker should know the
premisses of his deduction.
Our account of deductive capacities brings out, I think, the crucial dif-
ference between the defensible Deduction Principle and the indefensible
Pseudo-Deduction Principle. According to that account, a deductive capacity is
an ability to splice together two or more methods of forming beliefs to produce a
new method of forming beliefs. Thus, in the simple argument (1)to (3), we exercise
a deductive capacity to produce a composite perceptual-cum-inductive method
which is applied, in the circumstances of the argument, to form the belief that the
streets are weta belief that could not be formed there on the bases of either per-
ception or induction severally. However, if the composite method is to be reliable,
the methods which compose it must also be reliable. Contra (PDP), the mere truth
of the beliefs formed by those methods is not enough.
With this point clear, we can formulate an improved ground for the Deduction
Principle that does not vindicate the Pseudo-Deduction Principle. Let M1 be the
method, or basis, that Iuse to form my belief that it is raining, and let M2 be the
method Iapply to form my belief that if it is raining, the street is wet. Specifying
M1 and M2 precisely are non-trivial problems in the philosophy of perception and
the theory of induction respectively, and Ishall not address these problems here.
Given, though, that my beliefs that it is raining and that the street is wet if it is
raining both have the status of knowledge, we can say something about M1 and
M2. For a plausible general principle about knowledge is the Safety Principle:
(SP)If a belief that P that is formed using a method M amounts to knowledge
that P, then a belief that P that is formed using M could not easily have
been wrong
 Logical Laws 63

(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

My endorsement of the Deduction Principle is tentative. Perhaps there are

other examples which expose a flaw in the justification advanced for it, and show
how the Principle needs to be restricted. Further exploration of the issue must be
left to the epistemologists, but it is interesting that we have a justification of the
Principle that depends only on what Itake to be an attractive general theory of
deduction and the plausible epistemological thesis SP. Surprisingly many con-
temporary epistemologists are willing to reject the Principle. Our analysis at least
reveals the high cost of doing so.
 3
Logical Necessity

In the previous chapter, Irecommended that we conceive of logic as a second-order

discipline:it sets forth the laws of the laws of implication. This conception, Inow
argue, illuminates another topic of perennial controversy in the philosophy of
logic. Since the earliest days of their subject, logicians have argued over whether
logical truths exhibit a characteristic species of necessity. Assuming that they
do exhibit this, there have been further controversies over how logical necessity
relates to other kinds of modality. By extending the analysis of the previous chap-
ter, we shall be able to make some progress on these issues.

3.1 Logical Consequence Redux

Although that analysis emphasized the variety of implication relations to which
our ordinary deductions answer, it also pointed to the way to a principled iden-
tification of a relation of specifically logical consequence. The laws of logic tell
us that if one argument is sound by the standards laid down by a given impli-
cation relationthat is, if the arguments premisses stand in that relation to its
conclusionthen related arguments will also be sound, when assessed by the
same standards. In some cases, though, those laws ensure that an argument is
sound whatever implication relation provides the standard for assessing sound-
ness; the conclusion of such an argument may be said to be a logical consequence
of its premisses.
We have already seen an example of this. As remarked in 2.2, one of the classi-
cal sequent laws for the conditionalis:

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.

That is to say:whatever implication relation sets the standard for assessing an

arguments soundness, a deduction by Modus Ponens is sound. The present
account, then, yields the reassuring conclusion that in an instance of Modus
Ponens the conclusion follows logically from the premisses. It may be worth
reiterating the comparison with the modern conception of logical truth. Since
Gentzen, many philosophers and logicians have come to regard logical truths
as by-products of principles that tell us which arguments are logically valid.
Iam proposing that the classification of arguments as logically valid is itself a
by-product of yet more general principles that tell us which arguments stand or
fall together when assessed against a given implication relation.
The relation of logical consequence just characterized is narrower than on
some accounts. As we have just seen, for any implication relation, the premisses
of an instance of Modus Ponens stand in that relation to the instances conclu-
sion. Moreover, this universal truth about implication relations is itself grounded
in logical lawsspecifically, in a law concerning the conditional together with
postulates about the logical properties of implication relations. What, though, are
logical laws? They fall into two groups. The first comprises structural principles
for example, that any implication relation is monotonic. The second group com-
prises the sequent laws for particular logical notions, such as that given above for
the conditional. Accordingly, my account of logical consequence presumes that
we have some way of identifying the logical notions. Their characteristic mark,
Itake it, is that they figure in serious deductive argument about any topic what-
1
ever, and the conditional surely qualifies as logical in this sense. Accordingly, the

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

soundness of instances of Modus Ponens is a universal truth about implication

relations whose grounds lie in the general properties of those relations along with
laws concerning a logical notion. That is why the conclusion of any such deduc-
tion is a logical consequence of its premisses.
On this view, an instance of logical consequence must be grounded in laws
concerning logical notions. Some philosophers, though, operate with a broader
conception whereby a conclusion follows logically from some premisses when the
implication is grounded in some conceptual truths. Thus it is held to be a concep-
tual truth that no object is simultaneously red and green all over; so This object is
not green all over is said to be a logical consequence of This object is red all over
in this broad sense. Similarly, it is held to be a conceptual truth that no bachelor is
married; so Fred is unmarried is said to be a broad logical consequence of Fred
is a bachelor.
As we shall see, some prominent disputants in the controversy over logical
necessity invoke this broader notion of consequence, but Iam somewhat wary
of it and do not assume that it can be given a satisfactory explanation. Our nar-
row notion of logical consequence presumes that it is possible to circumscribe the
logical notions and to identify the laws governing them. The broader notion, by
contrast, presumes that it is possible to circumscribe the conceptual truthsa
much harder task. In any event, it is important not to confuse the two notions of
logical consequence, or to assume that what is true of one of them is true of the
other. Such an assumption leads to particular problems when discussing modal-
ity. Many of our concepts are applicable only when certain contingencies obtain,
so some of the statements that adherents of the broad notion classify as concep-
tual truths have contingent presuppositions. In this respect, conceptual truths
differ from laws of logic, and this difference means that the notion of broadly logi-
cal necessity (if it is intelligible at all) has a very different modal character from
the narrow notion on which Ishall focus.

3.2 The Controversy Over Logical Necessity

We have identified an implication relation of special interestthat of narrow, or
formal, logical consequence. Having identified it, however, questions arise about
modality. When a conclusion follows logically from some premisses, many phi-
losophers hold that the truth of the premisses necessitates that of the conclusion:if
it is raining then, necessarily, it is either raining or snowing. More precisely, these
philosophers hold that a full account of the notion of logical consequence will
bring in a particular species of necessity:where a conclusion follows logically
from some premisses, it is logically necessary that if things are as the premisses
 Logical Necessity 69

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

Russell duly castigates C. I. Lewis for invoking modal notions in explicating

consequence (Russell 1919, 154). If his argument were sound, it would tell equally
against Aristotle.
What does Russell mean by inference? An inference, he tells us, requires that
the premisses should be true, and that the conclusion should not be known. This
is scarcely comprehensible unless we take him to mean by inference what his
godfather, John Stuart Mill, had meantnamely, an instance of reasoning in
which a thinker comes to know a conclusion by deducing it from premisses that
he knows already. But whatever its merits as an account of the terms ordinary
meaning, logic, and the theory of deduction more generally, cannot confine its
attention to inferences in Mills sense. As noted in 2.2, we exercise our deduc-
tive capacities as much when we trace out implications of things we know to be
false, or when reasoning from suppositions whose truth-value we do not know,
as when we trace out consequences of things we know to be true. Thus someone
might exercise the deductive capacity that is characteristic of expertise in phys-
ics by arguing Suppose a (resultant) force were acting on the body. Thensc.,
in that case, on that suppositionit would be accelerating as well as by making
a Millian inference from the knowledge that a force is acting. Reasoning from
a supposition, however, plainly demands a stronger condition for validity than
Philonian consequence:the bare fact that either the conclusion is true or the
premiss is untrue is insufficient to underwrite the soundness of arguments from
suppositions, for what is supposed to be the case may well be untrue. Let it be
that no force is acting on the relevant body. In that case, any statement whatever
will be a Philonian consequence of A force is acting on the body. Yet the argu-
ment Suppose a force were acting. Then the Moon would be made of iron is not
a sound argument in physics.
Following Frege, Russell formalized his logic more geometrico, that is, as a
system of putatively known axioms and rules of inference by applying which a
thinker can come to know theorems on the basis of those already known axioms.
The possibility of formalizing logic in this way might be thought to show that one
could in principle avoid any reasoning from suppositions. However, the avail-
ability of this style of formalization does not save Russells Philonian account of
consequence. It is indeed possible to formalize logic in the geometrical style (the
Hilbert style, as proof theorists now call it). However, we would then need to ask
how a thinker knows the axioms of the formalization, and how he knows that its
rules of inference are truth-preserving. There is no plausible answer to either of
these questions which does not appeal to the subjects ability to make deductions
from suppositions. Certainly, the elucidations which Frege gave to justify the
axioms and rules of his formalization rely on exactly that ability. Thus, in order
72 The Nature of Logic

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

In a subsequent paper (1936), Tarski applied this apparatus in explicating logi-

cal consequence. He postulated a divisionperhaps effected by stipulation
between the logical and non-logical vocabulary of the relevant language. He then
defined a model of a statement to be an assignment that satisfies the correspond-
ing sentential functioni.e., the function that is obtained from the statement
by replacing all its non-logical vocabulary with variables of appropriate types.
Finally, he said that a conclusion follows logically from some premisses if every
model of the premisses is a model of the conclusion (Tarski 1936, 417). Similarly,
a statement is logically valid if and only if every assignment satisfies the corre-
sponding sentential function. To use the terminology that is now standard, a
statement is logically valid if and only if it is true in every model.
Unlike Russell, then, Tarski took the logical form of statements to be integral to
the explication of the notions of consequence and validity. Like Russell, though,
his account of these notions eschews modal concepts. Precisely this feature, how-
ever, generates problems for Tarskis account. Certainly, it is hard to see how
that account constitutes an analysis of, or a specification of what we mean by,
consequence or validity. If it did, then it would be conceptually necessary that
a statement that is true in every model is valid, but this condition is not met. An
assignment or model is a kind of set, but it is not conceptually necessary that any
set exists, so it is conceptually possible that every statement should be vacuously
true in every model. In a language that contains a sign for negation, however, it is
not logically possible for every statement to be true, and hence it is not logically
(or conceptually) possible for every statement in such a language to be valid. (For
objections to Tarski along these lines see Etchemendy 1990 and McGee 1992.) In
order to give a more convincing account of the notions of logical consequence and
validity, what we seem to need to bring in is precisely a modal element. Suppose,
for example, that one said that a conclusion follows logically from some premisses
if it is necessarily the case that every model of the premisses is a model of the
conclusion. Then we should no longer be embarrassed by its being conceptually
possible for any statement to follow from any other, for it surely is conceptually
possible for sets (and, in particular, models) to exist.
Commenting on these objections to Tarski, Smiley remarks that a debate is
called for, but it will be more fruitful if it asks for what purposes necessity is an
essential ingredient of consequence. For example, someone who does not endorse
Aristotles doctrine of proof and episteme may well be content with proofs that
establish the bare truth of theorems, and it is not obvious that this requires a
modal relation of consequence (Smiley 1998, 602). Smileys questionwhat pur-
poses postulating a modal element in logical consequence might serveis par-
ticularly pressing for my account; for, given classical logic, Philonian or material
74 The Nature of Logic

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.

3.3 Notions of Necessity

The argument for Aristotles Thesis that Iwish to advance rests on the conception
of logical laws recommended in the previous chapter. As argued there, any impli-
cative relation is associated with a space of possibilities and hence with a cog-
nate species of necessity. Thus the implication relation that a competent physicist
traces out in his deductions is associated with a space of physical possibilities and
hence with a notion of physical necessity. The physicists implication relation pre-
serves this variety of necessity:if some premisses are physically necessary, then
so is any conclusion that they physically imply. So any implication relation other
than the Philonian relation will preserve something richer than bare truth.
The relation of logical consequence, I argued, is distinguished from other
implication relations by the fact that logical laws may be applied in extending any
implication relation so that it applies among some complex statements involving
the logical connectives. The extended implication relation, however, must con-
tinue to preserve the relevant variety of necessity. By applying contraposition to
the basic implication relation among atomic statements in physicsa relation in
which A resultant force is acting upon the body stands to The body is accel-
eratingwe extend that relation so that The body is not accelerating implies
No resultant force is acting upon it. The extended relation, though, must still
preserve physical necessity, and a similar point holds for all other implication
relations. These diverse species of necessity will be preserved by their cognate
implication relations, though, only if the sort of necessity associated with prop-
erly logical consequence is at least as strong as the species of necessity associated
with any implication relation. That is to say, logical necessity must be at least as
strong as any necessity generated from implication.
The argument just sketched generalizes Aristotles argument for his Thesis.
In a deduction, we do, in a way, come to see how the conclusion could not have
been otherwise given the truth of the premissesalbeit in a restricted and
topic-specific sense of could not. The tie between premisses and a conclusion that
follows logically from them must encompass all these more specific varieties of
necessity.
In spelling out more fully the argument Ihave just sketched, we may start
from the Lindenbaum-Scott Theorem (see 2.4). The Theorem says that any
 Logical Necessity 75

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

of the space of possibilities . Similarly, we can begin to introduce the cognate

possibility operator by way of the principle that A is true if and only

if A is true at some member of . Thus, where is the implicative relation

implicit in Euclidean geometry, the space will comprise all those possibilities
that are identifiable in the relevant language and that are admissible in classi-
cal, Euclidean geometry. In this case, then, will express classical geometric
necessity.
I say we can begin to introduce a corresponding species of necessity, for impor-
tant properties of the relevant species are left undetermined by the principles that

A is true if and only if A is true at every member of and that A is

true if and only if A is true at some member of . Let us define a binary rela-
tion Rxy on the space as follows:Rxy if and only if A is true at y whenever

A is true at x. Rxy may then be understood as saying y is possible relative

to x or y is accessible from x. The possibilities in are not fully fledged pos-
sible worlds:while each member of is defined by a bisection of all the state-
ments in the relevant language, it is not assumed that such a bisection specifies
a fully determinate way in which the entire cosmos could be or could have been.
All the same, Kripkes celebrated results (see Kripke 1963)connecting the validity
of central modal laws to properties of the accessibility relation go through in the
present framework. So for example, we have A A if and only if R is reflexive;
A A if and only if R is transitive; and A A if and only if R
is symmetric. (For proofs of these results in a framework that eschews possible
worlds, see Koslow 1992, chap.35.) Substantial issues are involved in deciding
which of these properties a given accessibility relation possesses. So, for example,
when R is the accessibility relation on the space of physical possibilities, deciding
whether R is transitive will involve settling whether physical laws are themselves
physically necessary. Settling this will in turn involve deciding whether some
physical laws depend on initial conditionsconditions obtaining shortly after
76 The Nature of Logic

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

What happens when the consequence relation is that of classical first-order

logic? As Charles Parsons has observed, there is no conceptual problem in
extending the account of logical necessity sketched in the previous paragraph
to cover this case (Parsons 2008, 867). Where A contains no modal operator,

L A is true if and only if A is true in all classical first-order models. Where A

contains the L-operator, we proceed by replacing formulae of the form L B

by T or F according as B is or is not valid, thereby obtaining a L-free equiva-

lent of A to which the first criterion for the truth of L A may be applied. As

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

to give a sense to formulae involving quantification into the scope of L. The

most obvious way of attaching a sense to such formulae faces a problem. It is
a theorem of first-order logic that Hesperus is identical with Hesperus, so it
is logically necessary (in the sense cognate to first-order consequence) that
Hesperus is identical with Hesperus. It is not a theorem of first-order logic
that Hesperus is identical with Phosphorus, so it is not logically necessary that
Hesperus is identical with Phosphorus. We have, then, a situation in which
L (a=a), a=b, but not L (a=b). In fact, as Kit Fine (1989) has shown, substitu-
tion failures of this kind need not preclude attaching a coherent sense to Object
a satisfies the predicate L x and thereby attaching sense to such formulae as

x L x although the resulting quantified modal logic has some unusual

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,

logical necessity is at least as strong as any notion of necessity obtained from an

implicative relation in the way indicated.
We have, then, an answer to Smileys question. Given a classical logic,
Philonian consequence is a bona fide implication relation, and the notion of
necessity associated with that relation is bare truth. Moreover, as Smiley rightly
implies, in some cases we are content with deductions that, in his words, establish
78 The Nature of Logic

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

My account of logical necessity is consonant with Ian McFetridges famous dis-

cussion of the topic (McFetridge 1990a). McFetridge does not advance an account
of what logical necessity is; he only undertakes to say what the belief that an infer-
ence is logically necessary truth-preserving comes to. He suggests
that we treat as the manifestation of the belief that a mode of inference is logically
necessarily truth-preserving, the preparedness to employ that mode of inference in
reasoning from any set of suppositions whatsoever. Such a preparedness evinces the
belief that, no matter what else was the case, the inferences would preserve truth ...
Acentral point of interest in having such beliefs about logical necessity is to allow us to
deploy principles of inference across the whole range of suppositions we might make.
(McFetridge 1990a, 153)

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,

LA is true if and only if B L A is true for all B. The truth of B L A,

however, relates to the elaboration of counterfactual suppositions, so it is this part
of the McFetridges thesis that bears on the present discussion.
While our two accounts of logical necessity are consonant, mine is more gen-
eral. For McFetridge, the mark of a logically valid rule is its applicability to any
set of suppositions; presumably, a supposition is a statement. For me, the crucial
mark is its applicability to any set of subject-specific rules of deduction. Any state-
ment A may be regarded as the output of the rule Write down A, but the converse
does not hold. Since we do not assume that every implicative relation is analysable
as the logical consequence of a set of premisses (see 2.3), the greater generality of
my account is an advantage.
All the same, a similar idea underlies the two accounts. Most implication rela-
tions are rendered inapplicable by certain suppositions. When a classical physicist
reasons Suppose a force is acting on this body; in that case, it will be accelerating,
it may be reasonable to assess his reasoning against an implication relation, R,
 Logical Necessity 81

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.

3.4Logical Necessity versus Apriority and

Metaphysical Necessity
The proposed account of logical necessity, then, has a number of merits. Apart
from one class of apparent exceptions, which Ishall analyse below, it validates
the expected classification of paradigm cases, and paradigm non-cases, of such
necessity. Yet more importantly, it explains why we have such a notion and
accounts for its theoretical significance. We need to inquire, though, how logical
necessity relates to other central modal notions.
There seems to be no reason to suppose that every statement that is logically
necessary in the present sense is knowable a priori. Indeed, there is no general
reason to suppose that a logically necessary statement is, even in principle,
knowable at all. We will have such a reason when the logically necessary state-
ment is a validity of a logic for which we have a completeness theorem. We have
found no basis for assuming, though, that any logic has a complete axiomatiza-
tion. Alogically necessary statement will follow logically from any collection of
premisses, but a statements having that property does not imply that someone
could know it. The truth of a logically necessary statement will be grounded in
general facts about consequence relations and facts about the logical constants.
Again, though, there is no reason to suppose that such grounds can always be
recognized.
82 The Nature of 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

emended proposal, the conclusion of a valid argument must be deducible from

its premisses, but while this principle may have been part of traditional thinking
about validity, it is too strong. It excludes notions of validitysuch as validity
in full second-order logicfor which no complete set of deductive rules can be
given. Aconclusion may be a second-order consequence of some premisses even
when it cannot be deduced from those premisses. Inasmuch as old-fashioned
thinking about validity overlooked this point, it was surely wrong. As for
Edgingtons suggestion that her account comports best with the point of distin-
guishing valid from invalid arguments, many philosophers have taken logical
consequence to be a relation to which our logical rules must answer:a rule is to be
rejected as unsound if it enables us to deduce from some premisses a purported
conclusion that does not follow logically from them. However, deductions cannot
answer to consequence if consequence consists in deducibility. For these reasons,
Iconclude, we should not equate logical necessity with being knowable a priori or
even being knowable through deduction.
How does logical necessity relate to Kripkean metaphysical necessity? It
would take me too far from my theme to analyse the latter notion in detail, but
Ishall assume that its central characteristic mark is that a metaphysical possibil-
ity respects the actual identities of thingsin a capacious sense of thing that
encompasses stuffs such as water, and phenomena such as heat and pain, as well as
individual objects such as the planet Venus. This mark seems to be what is needed
to vindicate Kripkes central contentions in Naming and Necessitynotably his
claims that Water is H2O and Heat is the motion of molecules express meta-
physically necessary truths (Kripke 1980, 99 and 12833) and his crucial thesis
10
that if x=y then it is metaphysically necessary that x=y (1980, 35 and 97105).
If this gloss on the notion is along the right lines, there will be metaphysically
necessary truths that are not logically necessary. We may deem identity to be a
logical constant, so that it is logically necessary that Hesperus is identical with
Hesperus (if Hesperus exists) and logically necessary that if Hesperus is iden-
tical with Phosphorus then Phosphorus is identical with Hesperus. However, it
is not logically necessary that Hesperus is identical with Phosphorus:no logical
law grounds the truth of this statement. For all that, Hesperus is identical with
Phosphorus is a paradigm instance of a metaphysically necessary statement;
because Hesperus is in fact Phosphorus, there is no metaphysical possibility in
which they are distinct.
What, though, of the converse? Is every logically necessary statement meta-
physically necessary? An affirmative answer does not follow directly from the

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.

When it is logically necessary that if P then Q, however, the argument P; so Q

will be valid. So in that case there is no such R. So it is in no sense possible that P
and not Q. So it is in every sense necessary that if P then Q. Hence McFetridges
conclusion: Logical necessity, if there is such a thing, is the highest grade of
necessity.
Does McFetridges argument go through on the present account of logical
necessity and logical consequence? On that account, his first assumption is unas-
sailable. One of the marks of an implicative relation was that it should be mono-
tonic. So, for any such relation , if A C, then A, B C. As remarked in 2.3,

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:

(B)Julius was a mathematician

The person who invented the zip fastener emigrated to Tahiti
Therefore: Some mathematician emigrated to Tahiti.

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.

In the indicative, the antecedent presents something as an epistemic possibility, while in

the subjunctive the antecedent typically presents something as not an epistemic possibil-
ity, but as something which was a real possibility. Each kind of conditional goes with a
different kind of possibility. McFetridges second assumption, that there is a unitary sense
of possible that governs both, is not obligatory. (2004, 13)

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.

3.5Logical and Metaphysical Necessity:

The Paradox Resolved
How should we resolve the paradox?
The problem originates in the claim that arguments (A) and (B) are trivially cor-
rect:it is immediately evident in each case that the premisses imply the conclusion.
Iagree with Edgington that the arguments are trivial in this sense, but it would be
mistake to infer directly from this that, in either case, the conclusion follows logically
from the premisses, even in the broad sense of follows logically that she invokes.
Our deductions, Ihave argued, answer to a variety of implication relations between
which reasoners are adept at shiftingsometimes without noticing that they have
shifted. Given that, triviality is no guarantee of logicality. It may be immediately evi-
dent which implication relation sets the standard for assessing a given deduction,
and also evident that the deductions premisses stand in that relation to its conclu-
sion, even though the relation in question is not that of logical consequence.
I wish to argue that this is exactly the case with Edgingtons examples. In the
context in which she imagines it being propounded, argument (A) is indeed
trivial. But it is trivial only because that imagined context is one in which it is
common knowledge that the name Neptune stands for the planet (if there is
one) that is the cause of the perturbations that Leverrier observed. Whilst this
knowledge sustained the earliest astronomical uses of Neptune, very different
knowledge sustains the names use today. Within a few months of Leverriers
having made his conjecture, Neptune had been sighted through telescopes, ena-
bling people to understand the name in ways that did not depend on knowing
how it had originally been introduced. Were the argument (A) to come from
such a personwere it to come, for example, from someone who understands
the name Neptune as standing for the eighth planet from the Sunit would be
very far from trivial, and we would have no inclination at all to classify it as logi-
cally valid. In other words, the phenomenon we are trying to explainnamely,
the trivial correctness of argument (A)only arises in contexts where the com-
ponent occurrence of Neptune is understood to stand for the planet (if there
is one) that is the cause of the observed perturbations. As for argument (B), it is
already explicit in Edgingtons presentation of the case that Julius, as it occurs
there, is to be understood as a descriptive name, which stands for the inventor of
 Logical Necessity 89

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

(C)This object is red

So:This object looks red to normally sighted observers who are viewing
it in optimal viewing conditions.

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

Five Attacks on Classical Logic

 4
The Argument of
Dummetts Truth

Michael Dummetts paper Truth (1959) contains an interesting argument against

the universal applicability of classical logic, one that has not received the scrutiny
it deserves. It has been neglected partly because Dummett later advanced rather
different arguments against classical logic. Ishall discuss some of these in later
chapters, but it is well worth laying out the reasoning in Truth and assessing it
on its own merits. The main semantic argument against classical logic on which
Dummett later relied depends upon the principle that every aspect of a sentences
meaning must be fully manifest in the use that speakers of the relevant language
make of it, a strong principle that many of his critics have rejected. The argument
of Truth, by contrast, needs no such principle and so is unscathed by many of the
ripostes to his later attacks. The earlier argument, Ishall conclude, fails to show
that there is anything wrong with classical logic per se, but it contains insights
into the nature of content that are highly relevant to the other attacks on classical
logic that we shall consider.

4.1The Argument against Classical Logic in

Dummetts Truth
The anti-classical argument in Truth has the following structure. By reflecting on
the notion of a statements content, Dummett claims, and in particular by examin-
ing the relationship between a statements content and its truth or falsity, we may
come to know that no statement is neither true nor false, unless it is ambiguous or
vague. In this chapter, Ishall restrict the use of the term statement so that it apples
only to statements that are neither ambiguous nor vague. In those terms, then, philo-
sophical reflection enables us to know that no statement is neither true nor false, a
semantic principle which Dummett labels Tertium non datur. Now from the premiss
No statement is neither true nor false, a classical logician may validly deduce Every
96 Five Attacks on Classical Logic

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 ) .

The argument then runs as follows:

1. x(Sx (Tx Fx)) assumption, for reductio

2. Sa (Ta Fa) 1, existential instantiation
3. Sa 2, -elimination
4. (Ta Fa) 2, -elimination
5. Ta Fa 4, by the rule From (A B)
deduce A B
6. Ta 5, -elimination
7. Fa 5, -elimination
8. Sa (y(Oy Eay) Fa) , universal instantiation
9. y(Oy Eay) Fa 3, 8, -elimination
10. y(Oy Eay) 7, 9 contraposition
11. y(Oy Eay) 10, using the rule From y Ay
deduce yAy
12. (Ob Eab) 11, universal instantiation with b
parametric
13. Ob Eab 12, using the rule From (A B),
deduce AB
14. y(Oy Eay) 13, universal generalization, since
b was parametric
15. Sa (y(Oy Eay) Ta) , universal instantiation
16. y(Oy Eay) Ta 3, 15, -elimination
17. Ta 14, 16, -elimination
18. x (Sx (Tx Fx)) 6, 17 reductio ad absurdum,
discharging assumption 1

As we should expect of a reconstruction of Dummetts argument, the reason-

ing just given constitutes an intuitionistically correct proof of the conclusion
3
x(Sx (Tx Fx)) from the premisses and . More to the present point
for we are still assuming that the background logic is classicalthe argument is
also a classically valid proof of Tertium non datur from those premisses.

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

4.2The Exclusionary Theory of

Conceptual Content
Dummetts argument assumes that one should characterize the content of an
assertion in terms of what it excludes ... rather than in terms of what estab-
lishes it as correct (Dummett 1972, 22). We need to ask what grounds this
assumption.
In a Postscript he composed for a reprinting of Truth, Dummett tried to
explain why the former approach is better. It is obvious, hewrote,
that one who makes a conditional assertion does not wish to rule out the anteced-
ents being false, and that one who asserts a singular statement does not wish to allow
for the terms lacking a reference:but, if we tried to contrast the two cases in terms of
what established the assertions as correct, we should quickly find ourselves involved in
disputes about when a conditional statement is said to be true. Of course, we can talk
instead about what is required to be the case by an assertion; but this notion relates, once
again, to how we recognize the assertion as incorrect. The reason is similar to what is
said in the text [of Truth] about obedience and disobedience:our notions of right and
wrong, for assertions as for actions, are asymmetrical, and it is the apparently negative
notion which is primary. There is a well-defined consequence of an assertions proving
incorrect, namely that the speaker must withdraw it, just as there is a well-defined con-
sequence of disobedience [namely, that the proponent of the command may punish the
recipient, provided he had the right to issue the command in the first place]; there is not
in the same way a well-defined consequence of an assertions proving correct, or of obe-
dience. (Dummett 1972, 22)

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)

As someone with pragmatist sympathies, Ifind attractive Dummetts focus on the

effects of accepting an assertion; as we shall see, there are also strong reasons to hold
that the effect of an assertion most relevant to its content is that one who accepts it
does not allow for the occurrence of certain states of affairs. However, Dummetts
characterization of those states of affairs is problematical. On his account, the states
of affairs for whose occurrence someone who accepts an assertion does not allow are
those that would show the assertion to have been incorrect; but even if we take incor-
rect to mean false rather than unjustified or ill-founded this claim sits uneasily
with the main thesis that the assertions content is given by the possibilities thereby
set aside. Variants of Moores Paradox present one class of problem cases for the

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

The thought seems to be as follows. In any context of utterance or deliberation, only

certain possibilities are taken to be open or live; for Ramsey, acquiring a new belief
consists precisely in narrowing those possibilities. This narrowing alters ones dis-
position to act, for rational action and rational planning for action involve taking
account of whatever possibilities are open or live. So, to narrow the range of those
possibilities is to change what one takes into account when one acts, or makes plans
for action. From this pragmatist perspective, then, a beliefs content will be given
by the way its acquisition narrows the range of live possibilities; it will be given, in
other words, by the possibilities that acquiring the belief knocks out or excludes.
Similarly, the content of a statement is given by the possibilities that someone who
understands and accepts it is thereby committed to setting aside. In other words, a
statements content is determined by the possibilities that it excludes. When accept-
ing the statement A commits any agent who understands it to setting aside pos-
sibility x, Ishall say that A itself excludes x. In these terms, a contingent statements
content is given by the range of possibilities that it excludes.
It may still seem as though adverting to the possibilities excluded by a state-
ment is inessential to a pragmatist account of its content. Why not stick with the
simple pragmatist principle (P)? In fact, though, the suggested analysis of being
disposed to act as if P in terms of setting aside a range of possibilities renders
soluble a problem for pragmatist theories of meaning that otherwise looks intrac-
table. If (P) is to be parlayed into a theory of content worthy of the name, some
account must be given of how the contents of complex statements relate to those
of their parts. In particular, we need some account of how the content of a dis-
junction relates to the content of its disjuncts. Given (P), the problem is to say how
the disposition to act as if (either P or Q) relates to the disposition to act as if P and
the disposition to act as if Q.
Now in one direction the relationship is clear. An agent who is disposed to act
as if P is disposed to act as if (either P or Q). Similarly for the disposition to act
as if Q. However, the converse does not hold. An agent who believes that (either
P or Q) but has no view on which disjunct obtains will be disposed to act as if
(either P or Q) without being disposed either to act as if P or to act as if Q. So it
would be wrong say that an agent is disposed to act as if (either P or Q) if and
only if he is either disposed to act as if P or disposed to act as if Q.
Our further analysis of a disposition to act as consisting in setting aside a cer-
tain range of possibilities opens the way to a solution to the problem of specify-
ing the contribution of disjunction to pragmatically relevant content. Suppose
that Inspector Morse has narrowed down the field of possible murderers to five
men:Black, Brown, Green, Scarlet, and White. Suppose he now comes to accept
Either Black or Brown is the culprit. Given (P), that amounts to his becom-
ing disposed to act as if either Black or Brown is the murderer. How does this
104 Five Attacks on Classical Logic

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

in Robert Stalnakers influential analysis of the speech act of asserting, an analysis

that has been successfully applied in analysing a variety of linguistic notions and
constructions, notably presupposition and indicative conditionals (see Stalnaker
1999b, chaps. 13). In his terminology, the possibilities that are live or open at a
given stage in a conversation form the context set, and to make an assertion is to
reduce the context set in a particular way, provided there are no objections from
the other participants in the conversation. The particular way in which the con-
text set is reduced is that all of the possible situations incompatible with what is
said are eliminated (Stalnaker 1978, 86). By a possible situation, Stalnaker means
a possible worlda fully determinate way in which the whole cosmos could be
or could have beenwhereas Ido not require a possibility, or a possible situa-
tion, to be fully specific or determinate. That difference will turn to be important
(see Chapter6), but the exclusionary analysis of linguistic phenomena is not com-
promised if the requirement of determinacy is lifted. To the contrary, admitting
possibilities that are in certain respects unspecific allows more for plausible treat-
12
ments of certain cases.
The version of the exclusionary account that Irecommend may seem to gener-
ate a problem in the present dialectical context. We justified the claim that any
possibility excluded by a disjunction is also excluded by one of its disjuncts by
appeal to the principle that to act as if P is to act as if (either P or Q). That principle
takes for granted the soundness of the logical law of or-introduction. Certain
logical laws, then, are presupposed by the exclusionary semantic theory, and
one might wonder how that theory can be common ground between adherents
of rival logical schools. As always, it matters who the rivals are. Alogician who
rejects or-introduction will indeed be unable to accept the proposed axiom for
or. Logicians of many other schools, however, will be able to accept that axiom,
so the theory is of use in analysing the debates between these schools. In develop-
ing an exclusionary semantic theory in detail, Ishall take care to identify the logi-
cal presuppositions of its axioms.

4.3 Where the Argument of Truth Fails

There is, then, much to be said in favour of the conception of sense, or logically
relevant content, that forms the backdrop of Dummetts anti-classical argument.
Given that backdrop, though, are the premisses of the argument true?
Premiss is highly plausible. It says that if a possibility excluded by a statement
actually obtains, then that statement is false. An excluded possibility, it will be

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

In some cases, seems to deliver the correct answers. Let u be a statement of A

city will never be built here, as uttered in a rural place on day d. u excludes every
member of the following set of possible states of affairs:{a city stands at at the end
of day d+1, a city stands at at the end of day d+2,...}. (In the spirit of the proverb
about Rome, Iassume it takes at least a day to build or demolish a city.) Moreover,
all the relevant states of affairs that are excluded by u belong to this set. While u may
be said to exclude the possibility that a city should one day stand at , this possibil-
ity does not constitute a practically relevant state of affairs. In the words of What
is a Theory of Meaning? (II), such a possibility has no substance, for the expecta-
tion that it obtains cannot be disappointed, and hence cannot guide rational action.
Now according to , a statement is true when no excluded state of affairs obtains.
Applied to u, then, we reach

n (no city stands at at the end of day d + n) T (u)

which is intuitively correct. Indeed, the corresponding biconditional

T (u) n (no city stands at at the end of day d + n)

is intuitively correct. Furthermore, the specification of falsehood conditions that

corresponds to , namely,

F (u) n (some city stands at at the end of day d + n),

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

Dummetts argument for our ignorance of Bivalence, then, rests on principle

together with the assumption that there is a possible context of utterance sat-
isfying conditions (i)and (ii). Precisely these premisses, however, collectively
fail to cohere with premiss of the argument for Tertium non datur. Astate
of affairs excluded by an utterance of A city will never be built here must be
one in which a city has been successfully constructed in the place in question.
Accordingly, since the context of statement A satisfies condition (i), nothing
obtains [in that context] which is taken as a case of [As] being false. By prem-
iss , then, statement A is true. By the necessary condition for truth which is
Dummetts postulate , however, the truth of A entails that A is in principle
capable of being known to be true. And this entailment contradicts the sup-
position that the context of A satisfies condition (ii). For nobody can know that
a city will never be built in a given spot unless he has identified a feature of the
place in which it relevantly differs from the locations of thriving cities. The trio
of propositions comprising premiss , principle , and the supposition that a
possible context of utterance meets conditions (i)and (ii) is, then, inconsist-
ent; so at least one of them must be rejected. We must, in other words, reject
either a premiss crucial to Dummetts argument for Tertium non datur (viz., ),
or one of the premisses of his argument for our ignorance of Bivalence (viz.,
and the supposition about contexts). The present argument against classical
logic, however, was that we know Tertium non datur while being ignorant about
Bivalence. So that argument fails.
In response to an earlier presentation of this objection (in Rumfitt 2007),
Dummett conceded that, when understood classically, premiss was both
implausible in itself and incompatible with what I later say in the article
[sc., Truth] about truth. My error, he continued,

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

x [Sx (y (Oy Exy ) Tx )],

[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.

4.4 Exclusion and Truth

This disposes of the immediate threat that the argument of Truth poses to classi-
cal logic, but the matter is worth pursuing further. An exclusionary conception of
content has merits, but a convincing elaboration of it must eventually include an
account of the relationship between a statements content and its truth. Premiss
misdescribes that relationship. What is the correct account?
Dummett himself suggests an alternative in What is a Theory of Meaning? (II).
There, he reaffirms the exclusionary conception of assertoric content that shapes
the argument of Truth:by making an assertion, a speaker rules out certain pos-
sibilities; if the assertion is unambiguous, it must be clear which states of affairs he
is ruling out and which he is not (Dummett 1976a, 124). But he now offers a very
different account of truth. Instead of saying, as in , that a statement is true if no
excluded possibility actually obtains, he now proposes 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). In our terms, a
statement will be true if a possibility obtains which precludes the obtaining of any
possibility that the statement excludes. This proposal coheres much better with
the leading ideas of Truth than does . First, it rehabilitates the analogy between
truth and obedience to a command that the earlier article mishandles. Just as obe-
dience consists in behaving in such a way as to preclude any instance of disobedi-
ence, so truth consists in somethings obtaining that precludes any instance of
what is excluded. Second, it coheres with the articles further claim that a state-
ment is true only if there is something in the world in virtue of which it is true
112 Five Attacks on Classical Logic

(Dummett 1959, 14). According to Dummett, this claim expresses an important

insight about truth for which the correspondence theorists were groping, but it
sits uneasily with the thesis that the mere absence of a falsifier suffices for truth:a
mere absence is not something in the world. The new gloss fits this claim about
truth much better:a possible state of affairs whose obtaining precludes whatever
the statement excludes is something in virtue of which the statement is true.
The new account of truth remains indeterminate, though, until the rele-
vant sense of precludes is specified. Dummett himself does not explain the
latter notion. What follows is my own proposal; it has been designed to give
as fair a wind as possible to the main ideas in Truth and What is a Theory of
Meaning? (II).
It helps to begin by considering more carefully the nature of the things that a state-
ment excludes. Following Ramsey and Dummett, Ihave been calling these things
possibilities. By a possibility, Imean a way in which thingssome things, anyway
might be or might have been. This formula has a useful ambiguity. Read one way, it
describes epistemic possibilitiesways things might be or might have been, for all
we know. Read another way, it characterizes metaphysical possibilitiesways things
really could be, or really could have been, irrespective of our knowledge. These cat-
egories differ in extension. Given that Hesperus is identical with Phosphorus, there
is no metaphysical possibility of their being distinct. However, there was once an
epistemic possibility:for all that stone-age star-gazers knew, Hesperus might have
been distinct from Phosphorus.
When Isay that it is the possibilities that a statement excludes that deter-
16
mine its content, which sort of possibilities do Imean? Both. The account
as developed so far has stressed the epistemic aspect. In planning a course of
action, an agent will take into account things that might, for all he knows,
obtain even if, as a matter of unknown fact, they cannot obtain. Thus Morses
five possible culprits are men who, for all he knows, might have committed the
murder, even if it was in fact impossible for one or more of them to have done
so. Even from a pragmatist perspective, however, the space of metaphysical
possibilities remains relevant:in explaining why a course of action succeeds
or fails, we shall often have to advert to what could have obtained. So a prag-
matist theory of content can, and should, involve metaphysical possibilities
too. For these reasons, we need in the end to adopt a dual theory of content.
The very notion of linguistic content is structurally duplex:it comprises two
distinct components, each component introduced to serve a different purpose
and each to be theorized in conceptually different ways (McGinn 1982, 229).

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

anti-symmetric. While it is often helpful to identify metaphysical possibilities that

necessarily obtain together, this makes no sense for epistemic possibilities:even
when it is logically necessary that two epistemic possibilities should co-obtain, it
may be that, for all we know, one might obtain while the other does not.
How does determination interact with our basic semantic notionthat of a
statements excluding a possibility? The key Principle here says that whenever a
statement excludes a possibility, it also excludes any determination of that possi-
bility. This Principle holds for both spaces of possibility, and the argument for it is
simple. Suppose that statement A excludes possibility x. That is to say:anyone who
understands and accepts A is committed to excluding the possibility that x should
obtain. Suppose further that y determines x. In that case, it is logically necessary
that x obtains if y does. Hence, anyone who understands and accepts A is also com-
mitted to excluding the possibility that y should obtain. That is:A excludes y.
A reformulation of this Principle shows us what form an exclusionary semantic
18
theory will take. Let us postulate a total set of possibilities U. Then, for any subset
X of U, we define the interior of X (with respect to the determination relation )
by the following condition:

x Int ( X )if and only if y (x y y X ).

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:

x isa truth-ground of A if and only if y (x y y f A ).

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 ).

In these terms, a possibility is a truth-ground of a statement if and if only it belongs to

the exterior of the set of possibilities that the statement excludes. Asets exterior is the
interior of its complement:X=Int (X'). Hence, as expected, the truth-grounds of a
statement are disjoint from the possibilities that it excludes:for any set of possibilities
X, X X=. In classical topology, a sets closure is the complement of the interior
of its complement:Cl (X)=(Int (X'))'. Asets exterior is then the complement of its
closure:X=(Cl (X))'. It follows that X=Int ((X)')=Int ((Cl (X))'')=Int (Cl (X)):a
19
sets double exterior is the interior of its closure.
According to Dummett, a statement is true at each possibility which precludes
the obtaining of any state of affairs that it excludes. If we say that y precludes the
obtaining of any member of X if and only if no determination of y belongs to X,
then our account of truth at a possibility precisely matches his. Indeed, we may
prove from our definition that preclusion has the properties Dummett postulates
for it in What is a THeory of Meaning? (II):
Plainly, f f =, f f and, if f g, then g f ; hence f =f . We may
also assume that ( fg) = fg and that fg ( fg). (Dummett
1976a, 126)

We have already seen that f f = . Since f is open, f Cl (f) entails

f Int (Cl (f)), i.e., f f . If f g, then g' f ', whence Int (g') Int (f ' ), i.e., if f g
then g f . The last two results yield f =f . (f g)=Int ((f g)' )= Int (f ' g'))=
Int (f ' ) Int (g')=f g (in any topological space, Int (X Y)=Int (X) Int (Y)).

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).

4.5 An Exclusionary Semantics for the Language

of the Propositional Calculus
This general account of the relationship between a statements truth and the pos-
sibilities it excludes is of little interest unless the exclusionary account of content
can be developed into a systematic semantic theory for a reasonably rich lan-
guage. In this section Ishow how to construct such a theory for the language of
the propositional calculus.
As Iremarked at the end of 4.2, a prominent exclusionist is Robert Stalnaker,
and the construction of a semantic theory (even for a simple language) may allay
a worry that mention of his name is likely to have prompted. If two statements
are logically equivalent, then a speaker who asserts one is committed to setting
aside the very same possibilities as a speaker who asserts the other. So, on the
exclusionary account, it seems that logically equivalent statements must share
their content, contrary to the difference in meaning that we discern between (for
example) Snow is white or it is not and Grass is green or it is not. Stalnaker
embraces this conclusion (see 1991 and 1999a), but the construction of a semantic
theory enables us to temper it. Following David Lewis (1970), we may concede
that logically equivalent statements are alike in their coarse exclusionary content.
However, that does not stop us from identifying finer differences in meaning by
looking to the analysis of a compound into constituents and to the intensions of
the several constituents. For instance Snow is white or it isnt differs finely in
meaning from Grass is green or it isnt because of the difference in intension
between the embedded sentences Snow is white and Grass is green (Lewis
1970, 200). On the present account, the relevant difference in intension will be
that Snow is white and Grass is green exclude different possibilities. So, if we
follow Lewis in identifying fine meanings with semantically interpreted phrase
markers, our theory will ascribe different meanings to the two statements, even
though speakers who assert them are thereby committed to excluding exactly the
same possibilities.
In constructing a semantic theory for the language of the propositional cal-
culus, the key problem is to find appropriate exclusionary axioms for the con-
nectives. That for disjunction was trailed in the example about Inspector Morse
in 4.2: given only weak logical assumptions, we can show that the possibili-

ties excluded by A B are precisely those excluded by A and by B. To say that
118 Five Attacks on Classical Logic

acceptance of A commits one to excluding x is tantamount to saying that A entails

C, where C says that x does not obtain. Now if A B entails C, then certainly

A entails C, so any possibility excluded by A B is excluded by A:f A B fA.

Similarly, any possibility excluded by A B is excluded by B, so any possibility

excluded by A B is excluded by both A and B: f A B fA fB . For the converse
inclusion, suppose that x is excluded by both A and B, so that A entails C and B
entails C. By applying the rule of -elimination in the form without side prem-

isses, we may conclude that A B entails C, so x is excluded by A B . Thus
f A fB f AB . We reach, then, the following exclusionary semantic axiom for
disjunction:

(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 ) .

Our account of a statements truth-grounds also yields an exclusionary semantic

axiom for negation. In Facts and Propositions (1927, 445), Ramsey asserts the
equivalence between belief in the negation of A and disbelief in A. Although some
dialetheist logicians have challenged it, this equivalence is highly plausible, and it
entails that acceptance of As negation rationally commits one to excluding pre-
cisely those possibilities in which A is true. Thus we reach the following axiom for
negation:

(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:

(C) ver (A B)=ver (A) ver (B)

(D) ver (A B)=Int Cl (ver (A) ver (B))
(N) ver (A)=Int (ver (A)).

Thus the conjunction A B is true at a possibility if and only if both A and B

are true there. The disjunction A B is true at a possibility x if and only if every
determination of x has a determination at which either A is true or B is true. And the

negation A is true at x if and only if A is true at no determination of x. The clauses
for or and not are more complicated than the usual clauses for truth at a possi-
ble world, but some extra complexity is the inevitable price of relativizing truth to
possibilities that may not be fully specific or determinate. Thus, we cannot say that

A B is true at a possibility if and only if either A is true there or B is true there.
120 Five Attacks on Classical Logic

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

that we distilled from Ramsey extends to yield a logic of conditionals that he

could have accepted.

4.6The Choice of Logic within

an Exclusionary Semantics
Whilst it is good that the exclusionary theory extends to cover conditionals, our
chief concern is with the logic of conjunction, disjunction, and negation. What
does our analysis show about the choice of logic in this area?
At a number of points in our development of an exclusionary semantic theory,
distinctively classical logical principles have been assumed. One example is the dem-
onstration that a sets double exterior is the interior of its closure:this involved a meta-
logical application of Double Negation Elimination (see n.19 in this chapter). Other
examples are to be found in the proofs that the classical logic of conjunction, disjunc-
tion, and negation is sound and complete with respect to the proposed semantics.
This means that those proofs cannot be used directly to persuade an intuitionist (say)
to adopt classical logic, even if he accepts the exclusionary conception of content:the
proofs will appeal to rules of inference whose unrestricted validity he does not accept.
All the same, our analysis can, in a more subtle way, advance our project of
rationally adjudicating between rival logical systems. The soundness proof shows
that an adherent of the exclusionary semantics who uses classical logic in reason-
ing in the metalanguage has the resources to account for the soundness of the
24
propositional fragment of that logic in the object language. Let us a say that a
logic L coheres with a semantic definition of consequence if the soundness and
completeness of the rules of L is derivable from that definition using the rules
of L themselves. By extending the exclusionary semantics to cover the language
of the classical first-order predicate calculus, it is possible to show that classical
first-order logic coheres with the definition of consequence that is based upon the
exclusionary theory. If one accepts that theory of content, it is clearly good to use
a logic that coheres with the cognate definition of consequence. For the classicist,
then, there is no clashing of logical gears in moving between the object language
and the metalanguage. Aclassicist is not left in the embarrassing position of being
unable to account for the soundness of the logical rules that he employseven

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

exclusionary theory of content that Dummett proposes has much to recommend

it; but it yields a definition of consequence with which intuitionistic logic cannot
cohere. So far from supporting that logic, then, our analysis of Truth shows that
Dummett faces an unpalatable choice. He cannot adhere both to his preferred
account of semantic contentthe exclusionary theoryand to his preferred
intuitionistic logic.
 5
The Verificationist Attack
on Classical Logic

5.1 The Strong Verificationist Attack

on Classical Logic
In the 1960s and 1970s Dummett came to rely upon a semantic objection to clas-
sical logic very different from the one he had put forward in Truth. The later
argument is well summarized in Frege:Philosophy of Language. Languages of any
sophistication, Dummett observes, will contain statements whose truth-value we
have no effective means of deciding. In the case of any such statement,
the state of affairs which has, in general, to obtain for it to be true is, by hypothesis, one
which we are not capable of recognizing as obtaining whenever it obtains. Hence a knowl-
edge of what it is for that statement to be true is a knowledge which cannot be fully mani-
fested by a disposition to accept the statement as established whenever we are capable of
recognizing it as true:it is a knowledge which cannot, in fact, be fully manifested by actual
linguistic practice; and therefore it is a knowledge which could not have been acquired by
1
acquiring a mastery of that practice.

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

knowledge of the conditions under which an undecidable statement is true:such an

ascription can form no part of an adequate account of what we know when we know
a language. At least, this is so if the notions of truth and falsity are understood in a
realist waythat is, if a statement may be true even though we cannot recognize
that it is true, and may be false even though we cannot recognize that it is false.
How is this thesis about content, and its relationship to truth and falsity, sup-
posed to threaten classical logic? Crispin Wright presents the next stage of the
reasoning (which he calls the Basic Revisionary Argument) as follows. There
are, he observes, many areas of inquiry, or discourses,

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.

5.2 How to Be a Strong Verificationist

Dummett supposes that a verificationist theory of meaning will generalize the
Heyting semantics for the language of intuitionistic mathematics. That was a
major mistake. Ageneralized Heyting semantics fails to capture the senses of the
sentential connectives as these expressions are used in ordinary empirical state-
ments. It may just about pass muster as an account of the customary sense of the
word and, but it fails badly as an attempt to capture the senses of or and not,
letalone the conditional. In fact, Heytings original semantic theory limps even
as an account of the use of the connectives and quantifiers within constructive
mathematics. We can find better verificationist semantic theories for both empir-
ical and mathematical discourseand thereby more powerful arguments against
classical logicif we first liberate ourselves from Heytings unhelpful model.
Heytings theory clearly needs to be generalized before it is even a candidate to
be a semantic theory for any kind of empirical discourse:Heyting tells us under
what conditions mathematical statements are proved, but empirical statements
do not admit of anything that are properly called proofs. There is, though, a
natural way of making the requisite generalization. Having a proof of a math-
ematical statement Aor, at least, knowing of the existence of a proofis what
entitles a mathematician to assert A. More generally, then, let us say that some
evidence, x, warrants a statement A when a thinkers apprehending x would give
130 Five Attacks on Classical Logic
6
him epistemic entitlement to assert A. In these terms, the aim is to construct
a semantic theory that says how the possible warrants of a complex statement
relate to the possible warrants of its parts. Because a warrant licenses an assertion,
such a semantic theory systematically specifies the conditions in which a speaker
is entitled to assert the various statements of a language. It is, then, an instance
of an assertability-conditional theory of meaning. Icall the particular theory
that Iwill develop evidentialist semantics. Evidence is to be understood very
broadly:a piece of evidence may be a mathematical proof, an historical document,
or the result of an experiment. The term apprehension is to be taken correspond-
ingly broadly, so as to cover various more specific cognitive relations:apprehend-
ing a mathematical proof is a different sort of intellectual achievement from
apprehending an ancient charter. If we are to find semantic principles that apply
to statements of all sorts, we shall need appropriately general semantic notions.
How should an evidentialist specify the meaning of a sign for disjunction? If
we follow Heyting, we shall do this by laying down the following semanticaxiom:

(DH) x warrants A B if and only if either x warrants A or x warrants B.

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

one might try to rescue a version of (DH) by postulating a distinction between

direct and indirect warrants for a statement:whilst (DH) is false as stated, one

might hope to maintain that any direct warrant for A B is either a warrant for
A or a warrant for B. This manoeuvre, though, does not really help. What distin-
guishes a canonical proof of a mathematical statement is that no other justification
for it is more basic. Since no justifications for empirical claims are more basic than
those provided by perception, perceptual warrants will qualify as direct, but then
there are counterexamples even to the emended version of (DH). Suppose that Ican
easily distinguish either of the heavenly twins, Castor and Pollux, from the other
residents of the neighbourhood. In that case, a fleeting glance out of the window
provides a direct warrant for the disjunctive statement Either Castor or Pollux has
just run past. Since, however, Ineed a longer or closer view to tell which twin it was,
that glance gives no warrant for asserting either disjunct (cf. Dummett 1991, 267).
For reasons of this kind, Dummett eventually conceded that the Heyting inter-
pretation of disjunction is unsatisfactory in empirical contexts, and said he wished
he knew how to give a better one for an empirically applicable justificationist [i.e.,
assertability-conditional] semantics (Dummett 2007b, 698). In the sequel, Ishall
consider whether there is any prospect of Dummetts wish coming true.
Before doing so, however, it is in place to remark other difficulties that confront
Heytings treatment of disjunction. Let us say that a theory T has the disjunction
property if a disjunctive statement is a theorem of T only if one of its disjuncts
is also a theorem. And let us say that T has the existence property if an existen-
tially quantified statement is a theorem only if a witness for it is also a theorem.

(A witness for x(x) is any statement in the form (c) , where c is a singular
term.) Intuitionistic sentential logic has the disjunction property, and first-order
intuitionistic logic has both the disjunction property and the existence property.
Moreover, certain first-order theories of particular interest to constructive mathe-
maticiansnotably Heyting Arithmetic, HAhave both properties. The proper-
ties are in fact closely related. In any first-order intuitionistic theory, the existence
property implies the disjunction property, since in first-order intuitionistic logic

A B is equivalent to x ((x=a A) (x a B)) . Atheorem of Friedmans
(1975) gives a partial but still substantial converse:in all recursively enumerable
extensions of HA, the disjunction property implies the existence property.
Since the Heyting semantics purports to specify exhaustively the circum-
stances in which statements have proofs, its axiom for implies:

(P) A B is provable if and only if either A is provable or B is provable.
(P) will be acceptable to a constructivist only if the meaning of proof is always
determined in such a way that the class of provable statements has the disjunction
132 Five Attacks on Classical Logic

property. If proof is understood to mean proof in pure intuitionistic logic, or

proof in Heyting Arithmetic, it will meet this condition. However, a constructiv-
ist of Dummetts stripe will not equate provability simpliciter with provability in
either of these systems, or indeed with provability in any other recursively axi-
omatizable formal theory. On his view, one moral of Gdels First Incompleteness
Theorem is that no such equation is possible:what is provable outruns what is
provable in any given formal system (see Dummett 1963). Accordingly, (P) will be
acceptable only if, at any stage in the never-ending expansion of acceptable meth-
ods of proof, the class of statements that are provable by those methods has the
disjunction property.
An example of Kreisels, however, shows that this is not always so (see Troelstra

1973, 91). Let Pr be the two-place proof predicate of HA:that is, Pr (n, m) means
that n is the Gdel number of a proof in HA of the formula whose Gdel number is

m. Then define (x) to mean Pr (x, 0=1 ) y Pr (y, 0=1 ) . Assuming that

HA is true on its intended interpretation, y Pr (y, 0=1 ) is true, so x(x)
is true. So T=HA + x(x) is an intuitionistically acceptable theory, and T (intui-
7
tionistically) entails x(x) . For no n, however, does T entail (n) . Hence T
lacks the existence the property. The usual proof that the theorems of an axioma-
tizable theory are recursively enumerable goes through in intuitionistic logic. So,
by Friedmans result, T lacks the disjunction property too. T, however, represents a
possible stage in the expansion of a constructivists methods of proof. Such a theo-
rist might start with Heyting Arithmetica paradigm of a constructively accept-

able mathematical theory. He might then consider the definition of (x) , reflect

that x(x) must be true, and so expand his notion of provability from provability
in HA to provability in T. In doing this, however, he will be moving from a class of
provable statements that possesses the disjunction property to one that does not. So,
at certain stages in a constructivists determination of what is provable, the theory
comprising all provable statements will not possess the disjunction property, con-
trary to (P). It will not help to propose that (P) may yet be true if provable is taken
to mean provable by absolutely any constructively acceptable method of proof. For
the idea that this gloss gives determinate content to the notion of provability will be

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

anathema to any constructivist:he will not take it to be a determinate matter what

those methods are.
Problems in generalizing Heytings semantics have been a fertile source of
scepticism about assertability-conditional semantic theories, and hence about
the prospects of the strong verificationist attack on classical logic. In his provoca-
tive lecture Must Do Better, for example, Timothy Williamson remarked that
assertability-conditional semantics
began with one more or less working paradigm: Heytings intuitionistic account of
the compositional semantics of mathematical language in terms of the condition for
something to be a proof of a given sentence. The obvious and crucial challenge was to
generalize that account to empirical language: as a first step, to develop a working
assertability-conditional semantics for a toy model of some small fragment of empiri-
cal language. But that challenge was shirked. Anti-realists preferred to polish their for-
mulations of the grand programme rather than getting down to the hard and perhaps
disappointing task of trying to carry it out in practice. The suggestion that the pro-
grammes almost total lack of empirical success in the semantics of natural languages
might constitute some evidence that it is mistaken in principle would be dismissed as
crass. (Williamson 2007, 282)

I agree with Williamson that an assertability-conditional semantics worthy

of the name cannot confine itself to the language of mathematics; it must also
apply to empirical discourse. Our analysis, though, suggests that the advertised
paradigm did not really work even for mathematical statements. The challenge
for the strong verificationist, then, is not to generalize Heytings theory but to
find a new form of assertability-conditional semantics. If one could be found, we
might escape Williamsons baleful conclusion that people trying to construct
anti-realist semantic theories proceed as if Imre Lakatos had never promulgated
the concept of a degenerating research programme (2007, 284).
What alternative form might an evidentialist semantic theory take? In
addressing this question, we need to make some decisions about the framework
in which the theory is to be cast. Its key notion is that of evidence warranting
a statement, but how should we theorize about evidence? In English, the word
behaves as a mass term: we speak of someones having little evidence for an
assertion, and of one thinkers having more evidence than another for a claim.
On the other hand, we also speak of pieces of evidence, in the plural. While it
would be possible to use mereology to treat of evidence as a kind of stuff, it is
more convenient to suppose that evidence has somehow been individuated into
pieces over which a first-order objectual variable may range. Such individuation
will involve making decisions which go beyond any that are mandated by the
ordinary use of the term piece of evidence. Ishall also assume that we have a set
134 Five Attacks on Classical Logic

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.

We may then rewrite (D) as

(D) |A B| = Cl (|A| |B|).

The reformulation helps because Cl is a closure operation in the sense favoured

by lattice theorists. That is to say, the operationis

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

are supplemented by plausible axioms about the structure of evidential support.

Following Williamson (2000, chap.10), let us use the familiar dyadic or condi-
tional probability operator P(A/B) as a measure of evidential probability. That is,
let us understand P(A/B) as a measure of the degree to which B supports the truth
of A. Thus, where B is evidence, apprehension of which enables us to know the
truth of A, we take P(A/B) to be unity; and where B is evidence, apprehension
of which enables us to know the falsehood of A, we take P(A/B) to be zero. In
our terminology, then, B is a warrant for A just in case P(A/B)=1, and B is an
14
anti-warrant for A just in case P(A/B)=0. On this way of understanding P(A/B),
the following axioms are highly plausible:

I 0 P(A/B) P(A/A B)=P(t/B)=1

P(f/C)=0 unless P(D/C)=1 for all D
II P(A B/C)=P(B A/C)
III P(A B/C)=P(A/C) P(B/A C)

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.

5.3 ARenewed Threat to Classical Logic

How might an evidentialist semantic theory generate challenges to classical
logic?
On the strong verificationist view that we are considering, a statements mean-
ing is given, not by its truth-conditions, but by the conditions in which there is
warrant to assert it. On that view, then, there is some temptation to characterize

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.

5.3.1 Double Negation Elimination

Let us consider Goldbachs Conjecture: Every even number greater than two

is the sum of two primes. The statement may be formalized as xFx , where
the quantifier ranges over natural numbers and where F means is the sum

of two primes if is even and greater than two. F is decidable:where n is any

numeral, we may decide the truth of Fn by first deciding whether n is an even
number greater than two and, if it is, using the Sieve of Eratosthenes to identify all
the prime numbers less than n and then checking to see whether any pair of them
has n as its sum.
We have yet to lay down an evidential axiom for the universal quantifier. It is,
however, plausible that such an axiom will generalize (C):evidence will warrant

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

a universally quantified statement if and only if it warrants all of its matrix

instances. Applied to the present case, then, we can say that a construction con-

stitutes a proof of xFx if and only if it proves each statement Fn . In fact, this
result is also delivered by the Heyting semantics for the language of arithmetic.
Now it seems coherent to suppose that both of the following propositions are
true of Goldbachs Conjecture. (1)There is no counterexample to it; that is, there is
no even number greater than two which is not the sum of two primes. But (2)there
is no uniform mathematical reason why each natural number has the property F;
indeed, there is no finite partition of the natural numbers with such a reason for
each class in the partition. From assumption (2)it follows that there is no proof of
the Conjecture:a proof would either present a uniform reason why each natural
number n has the property F, or it would identify a finite partition of the natural
numbers and give such a reason for each class in the partition.

Which constructions, though, qualify as proofs of xFx ? To address

this question, we first consider what a proof of xFx would be. By (N), this

would be an anti-warrant for the statement xFx that is, a warrant for deny-
ing Goldbachs Conjecture. We have yet to venture a general compositional the-
ory of anti-warrants, but it is plausible to suppose that an anti-warrant for the
Conjecture would be a construction showing that a particular natural number n
is not F. Now it follows from assumption (1)that any construction that purports to
show this will contain a mistake:ex hypothesi, there is no counterexample to the

Conjecture. An anti-warrant for xFx will duly be a construction that shows
that there must be a mistake in any such purported construction. Such a con-
struction, however, need not provide a reason why each natural number is F; nor
need it give a finite partition of the natural numbers with such a reason for each

class in the partition. By (N), though, an anti-warrant for xFx is a warrant

for xFx . So there can be a warrant for xFx that is not a warrant for

xFx . By the strong verificationists standards, then, xFx does not follow

from xFx and Double Negation Elimination is not a universally valid rule.

5.3.2 Distribution and Negation Introduction

As remarked, (D) does not validate the Law of Distributionthe principle that
A (B1 ... Bn) entails (A B1) ... (A Bn) . What is more, the strong verifica-

tionist theory of meaning appears to generate clear counterexamples to Distribution.

Let A be a statement saying that the position of a given subatomic particle is precisely
p at time t. For each i between 1 and n, let Bi be a statement saying that the same parti-
cles momentum at time t is precisely mi. Now where n is sufficiently large, quantum
theory allows that there is a measurement, x, which establishes that the position of the
particle at a time t is precisely p and also establishes that its momentum at t is either
 The Verificationist Attack on Classical Logic 141

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-

rants that statement. According to Heisenbergs Uncertainty Principle, no measure-

ment can precisely determine both a particles position and its momentum. So no

measurement can warrant A Bi , for any i. According to (D), then, the set of pieces
of evidence that warrant (A B1) ... (A Bn) is Cl ( ... )=Cl ()=. So,
17

while x warrants A (B1 ... Bn) , it does not warrant (A B1) ... (A Bn) .

On the strong verificationist account of logical consequence, then, the Law of

Distribution is not valid.
In fact, that account of consequence invalidates other classical laws too. Aspe-
cial case of the classical (and intuitionistic) introduction rule for negation says

that when A and B are logically inconsistent, A entails B . By Heisenbergs
Principle, no evidence warrants both A and Bi. So, under the strong verificationist
account of consequence, A and Bi are logically inconsistent, for any i. However,

under that same account of consequence, A does not logically entail Bi .
The measurement x above warrants A, but it does not warrant the denial of Bi.
Measuring a particles precise position at a time precludes determining its precise
momentum at that time, but such a measurement does not entitle us to deny that
the particle has a given momentummi, sayat that time.
Given the strong verificationist account of consequence, then, the laws of
logic would appear be no stronger than the principles of quantum logic. It is well
known that Distribution is not valid in quantum logic, and in fact the rule of
-introduction is not valid there either. Afamiliar algebraic semantics for quan-

tum logic (Goldblatt 1974)takes the semantic value of A to be the orthocom-
plement of the semantic value of A (in a non-distributive lattice). This shows why
-introduction is not valid in quantum logic. Given the premiss that two sets are
disjoint, we cannot infer that one of them is a subset of the orthocomplement of
the other.
This argument for quantum logic is very different from that advanced by Hilary
Putnam in his famous paper Is Logic Empirical? (Putnam 1968). According to
Putnams argument, which Ishall discuss in 6.5, we must switch from classical
logic to quantum logic if we wish to preserve the possibility of giving a realist
interpretation to statements about subatomic particles. The present argument, by

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

contrast, assumes an anti-realist, evidentialist account of the meanings of such

statements (and indeed of other statements too). The reasoning under considera-
tion here is suggested by the last two pages of Dummetts reply to Putnam, which
was also entitled Is Logic Empirical? (Dummett 1976b, 2879). Dummett denies
that the choice of logic is straightforwardly an empirical matter:a logical law
cannot be refuted by an experiment. The bearing of an experimental result on a
logical principle is always mediated by the theory of meaning that we apply to the
language of the relevant scientific theory, and the choice of a theory of meaning
perforce involves philosophical argument. If, however, philosophical argument
leads us to opt for a verificationist theory of meaning, then our best account of
the structure of possible methods of verification will bear on the choice of logic.
According to our best current theory of these matters, what we can know about
subatomic particles is constrained by Heisenbergs Principle. This is why the Law
of Distribution comes out invalid under evidentialism.
The present case for quantum logic, indeed, avoids the familiar objection that
quantum logic is not really a logic. Given the Uncertainty Principle, a certain
algebra of experimental tests turns out to be a non-distributive lattice, but unless
we posit a special sense for or in quantum mechanical contextsa sense that
is specified in terms of these testsfacts about that algebra will be irrelevant to
logical relations among statements (see Dummett 1976b). It is indeed implausi-
ble to claim that or bears a special sense in discourse about subatomic parti-
cles. However, the present evidentialist case in favour of quantum logic does not
require that claim. On the evidentialist view, the sense of the disjunction sign
is always given by (D), a specification that classical and quantum logicians can
both accept. The point is that, while the evidence we are accustomed to acquiring
in everyday life is structured in a way that validates Distribution, that gleaned
from measuring the properties of subatomic particles is not. It is for this reason
not any ambiguity in the word orthat the Law of Distribution applies to some
areas of discourse but not to others. It is not a universal logical law, but breaks
down when we reason about subatomic particles.
This view of the matter has the further advantage that the quantum logician
is not embarrassed by his apparently inevitable recourse to classical logic in
developing a meta-theory for his logic:it is hard to get anywhere in meta-theory
without using proof by cases (in the form with side premisses). From an evi-
dentialist perspective, there is no equivocation between the or that connects
quantum-mechanical statements in the object language, and the or that con-
nects semantic statements in the metalanguage:the sense of both sorts of occur-
rence is given by (D). The relevant difference is that the objects with which a
semantic theory dealsconnectives, statements and the likeare not subatomic
 The Verificationist Attack on Classical Logic 143

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.

5.4Why One Should Not Be

a Strong Verificationist
These purported counterexamples to classical laws may seem to be threatening,
but they rest upon a reconceptualization of the notion of consequence that a clas-
sical logician is well placed to resist. In Part I, Idefended the ancient thesis that the
kernel of consequence is the necessary preservation of truth; the counterexam-
ples of the previous section, however, depend upon taking consequence to consist
in the preservation of warrant. The question to press on the strong verificationist
is what is supposed to justify his deviant account of consequence.
So far as Ican see, the only possible justification is thateither generally, or in
the particular fields of inquiry that the counterexamples concernthe notion of
truth is itself to be explained in terms of warrant:a statement is true only when
some evidence would warrant it. So we have to consider what might sustain this
claim about the relationship between truth and warrant. The problem naturally
divides. First, we have to consider if there are any general reasons for making the
claim across the board. If there are not, we have then to ask if there are particular
features of statements about subatomic particles, or of mathematical statements,
that make a verification-transcendent notion of truth unavailable in those cases.
The first question takes us back to the general Dummettian arguments for strong
verificationism that were adumbrated in 5.1. We surely need a powerful argument
to overturn our pre-theoretical conviction that a statement may be true even though
no one can recognize that it is true. Of course, it is impossible to give an example
of a statement that is known to be true even though no one can recognize its truth.
Confronted, though, with a pair of statements like An even number of Roman sena-
tors were standing within 10 metres of Pompeys statue when Brutus (first) stabbed
Caesar and An odd number of senators were standing there then, most of us have
a strong conviction that one (and only one) of the pair is true, even though it is now
impossible to recognize which it is. Moreover, in the present case, it is easy to explain
how it is that one of the statements is true even though no one today can know that it
is true. In order for any of us to know so much, someone present at Caesars stabbing
would have had to make appropriate measurements and effect an appropriate count;
they would then have had to record their results in a form that, directly or indirectly,
144 Five Attacks on Classical Logic

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

Even in the case of mathematical statements, this requirement of decidability is

problematic. What warrants the assertion of a mathematical statement is a proof
of it, and when a particular formal system S has been specified, a thinker will in
principle be able to decide whether a given derivation is a correct proof of a
statement A within S. On Dummetts view, though, whether is a correct deri-
vation of A within a given formal system does not determine whether is a war-
rant to assert A. The mathematical Platonist takes Gdels First Incompleteness
Theorem to show that the true statements of arithmetic outrun those that are
provable. Dummetts intuitionist, by contrast, takes the First Theorem to show
that what is provable simpliciter outruns what is provable in any given formal sys-
tem (see again Dummett 1963). It is provability simpliciter that sets the standard
for making a mathematical assertion, but someone who understands A may not
always be able to decide whether a deduction is a correct proof simpliciter of A.
If Ideduce some set-theoretic statement A from the premiss There are infinitely
many Woodin cardinals by logically correct steps, and if in due course that prem-
iss comes to be accepted as an axiom of set theory, then Imay, retrospectively, be
deemed to have proved A. But it is not (now) decidable whether my deduction is a
proof of A. In order to decide that, we would have to decide whether the premiss is
19
a correct axiom of set theory, and no one today is in a position to do that.
Matters are even worse when we try to apply the requirement of decidability to
the case of contingent statements. Proto-Indo-European is the name linguists
have given to the hypothesized mother language from which all the members of
the Indo-European family are supposed ultimately to derive. Consider then the
statement Proto-Indo-European was originally spoken in the Pontic-Caspian
Steppe (an area of the Russian steppes, north of the Black Sea and east of the River
Dnieper) in the fourth millennium BC. This statement formulates the so-called
Kurgan Hypothesis, one of the most famous, and controversial, hypotheses in
historical linguistics. I understand the hypothesis. I do not know for certain
that the name Proto-Indo-European has a reference, but because Iknow what
the Indo-European linguistic family is, and because Iunderstand the notion of
a mother language, Iknow what it would take for a language to be the referent
of that descriptive name. Moreover, because Iknow where the Pontic-Caspian

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.

5.5 Dubious Grounds:McDowells

Challenge to Classical Logic
The alleged counterexamples to classical laws given in 5.3, then, rest on a recon-
ceptualization of the notion of consequence that is inadequately supported. But
if we reject that reconceptualization, and adhere to the ancient account of con-
sequence as the necessary preservation of truth, we can discern the outline of a
more subtle and more insidious threat to the classical logical laws, suggested by
some writings of John McDowell (see especially his 1976).
To see how this threat arises, let us again focus on the dispute between the classi-
cal logician and his intuitionist rival. One interesting feature of that dispute is that
both parties accept that no statement is neither true nor false:while the intuitionist
148 Five Attacks on Classical Logic

refrains from asserting, of an arbitrary statement, that it is either true or false, he

has just as little time as the classicist for third truth-values or truth-value gaps.
Accordingly, the parties agree that the two statusesbeing true, and being false
that figure in the classical truth-tables for the connectives are the only alethic
statuses with which a metalogical theory needs to deal. Suppose further that the
truth-tables are read strictly as expressing the way that the truth, or falsity, of a
complex statement depends on the truth, or falsity, of its components. Thus, we
read the four lines of the truth-table for as saying:(1)if A and B are both true,

then the conjunction A B is true; (2)if A is true and B is false, then A B

is false; (3)if A is false and B is true, then A B is false; (4)if A and B are both

false, then A B is false. Suppose, finally, that the falsity of A is equivalent to the
truth of As negation. Then, as McDowell also observes, the two logical schools can
both accept the content of these truth-tables. The truth of principles (1)to (4)cor-
responds to the validity of four sequents(1) A, B; so A B; (2)A, B; so (A B);
(3)A, B; so (A B); (4)A, B; so (A B)all of which both logical schools
accept. It is easy to check that both schools also accept all the sequents that are cor-
respondingly implicit in the truth-tables for and .
Now it is plausible to maintain that our understanding of signs for conjunc-
tion, disjunction, and negation puts us in a position to know the semantical rela-
tionships recorded in the truth-tables, and hence puts us in a position to know
the validity of all the sequents implicit in those truth-tables. As McDowell points
out, however, these basic sequents do not constitute a complete axiomatization of
classical propositional logic. In order to justify the further sequents needed for
a complete axiomatization, we require a further semantical premiss. The Law of
Excluded Middle both exemplifies the problem and points towards the natural
additional premiss. From the truth-tables for and , we can infer that if A is

true then so is A A , and that if A is false, then A A is again true. Even in
a classical metalogic, though, these premisses do not entail that every instance of

A A is true. We need an extra premiss. The natural premiss to add, of course,
is the Principle of Bivalence:the principle that every statement is either true or
false. Armed with this Principle, we can prove the soundness of the rules of the
21
classical propositional calculus even in an intuitionist metalogic.

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)

More pithily:Classical proof theory is unsound not because it misrepresents the

senses of the logical constants but because it purports to prove sentences which
cannot be known to be true (McDowell 1976, 601).
McDowell calls this the weak verificationist challenge to classical logic, but the
name is unfortunate. The nerve of the challenge is simply that the senses of the
connectives do not by themselves support the soundness of all the classical rules.
To show that some of those rules are sound, one needs in addition to assume
a doubtful semantic premisssuch as the Principle of Bivalence. The crux of
the challenge, then, is the dubiousness of the Principle, and verificationism
(whether weak or strong) is not directly to the point. Aphilosopher might doubt

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

Bivalence because he is a verificationist, but he might equally well doubt it for

quite different reasons. Someone who is unwilling to assert that vague state-
ments are always either true or false, for example, need not be motivated by any
kind of verificationism:he is more likely to be motivated by the absence of any
plausible account of what could make such a statement true, or make it false.
McDowells challenge, then, is really that the soundness of the classical laws
rests on dubious grounds, and Ishall accordingly call it the challenge of dubious
grounds.
The argument challenges classical logic for propounding theses that we do not
know. But does it support the adoption of some rival logical system, or does it
simply leave us in a state of aporia where we do not know what the laws of logic
are? The challenge appears to leave us quite unable to determine whether certain
rules are truth-preserving. The rule of Double Negation Elimination, for exam-
ple, preserves truth if the Principle of Bivalence holds; otherwise it does not. Yet,
ex hypothesi, we do not know whether Bivalence holds.
All the same, McDowell attempts to extend his argument so that it provides a
justification for the intuitionistic propositional calculus rather than the classical
one. His first move in this direction is to recommend a novel conception of the
truths of logic, not as true solely in virtue of the senses of the logical constants
without assuming the principle of bivalence, there is no telling which sentences
have that statusbut as knowable solely in virtue of the senses of the logical con-
stants. Such a conception, he continues,
could be made precise ... by a systematic specification of the conditions 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. Keeping to the simplifying
assumption that the only logical constants are the sentential connectives, such a speci-
fication may be thought of as constructed by considering how claims to know the truth
of complex sentences could be justified by knowledge of the truth of simpler sentences,
in the light of, first, that systematic dependence of truth values on truth values which,
we continue to assume, constitutes the senses of the connectives; and, second, the veri-
ficationist insistence that we cannot claim to know that every sentence is either true
or false. Any such systematic specification would certify sentences of some forms as
knowable to be true whatever the epistemic status of their components, and, on the new
conception, these would be the truths of logic. Soundness and completeness in a proof
theory are definable, as before, in terms of the notion thus rendered precise. (McDowell
1976, 60)

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.

6.1 Moderate Modal Realism and Possible Worlds

An attractive position in the metaphysics of modality is moderate modal realism.
On this view, associated particularly with Saul Kripke (1980) and Robert Stalnaker
(2003), the different modalities (the necessary, the actual, the possible) reflect the
existence of what we may call different possibilities:different ways thingssome
things, anywaycould be or could have been. The diverse species of possibility
(physical possibility, logical possibility, etc.) correspond to diverse determinations
of the sense of this could:could physically, could logically, and so forth. Some
moderate realists speak of ways things might have been, but Ishall generally use
could. Might often means might for all we know, but epistemic possibilities are
not the present topic.
1
Moderate realists regard a way things could have been as a state or property of
the actual worldthat is, of the actual cosmos. That is where they differ from the
extreme modal realism, or modal concretism, of David Lewis (see especially his
1986b). On Lewiss view, the different modalities reflect the existence of a multi-
plicity of alternative universesobjects of the same kind as us and all our sur-
roundings, huge pieces of real estate in Stalnakers phraseone of which we

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.

However, because each statement is assumed to be either true or false at an arbi-

trary possible world, the worldly analogue (N2) is correct:

(N2)For any possible world w, the negation A is true at w if and onlyif
A is not true at w.

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.

6.2 Reasons for Seeking an Unworldly Theory

However, if this is the argument for favouring worlds over possibilities more gen-
erally, it is clearly weak. The most obvious semantic principles for disjunction and
negation do indeed fail if truth is relativized to possibilities. For all that, it might
be worth looking for less obvious principles before embracing a worldly theory of
modality. Iargue next that the moderate realist should seek alternative semantic
principles. It is not only in motivating moderate realism that possibilities come
into their own. We also need them, rather than fully fledged worlds, for some
important applications of our theories of modality. So the simpler semantic prin-
ciples that worlds permit come at the cost of restricting the applicability of the
theory to which they belong.
An important example of such a restriction lies at the heart of the current enter-
prise. In Chapter3, Idefended Aristotles Thesis:if a conclusion follows from some
premisses, then there is no possibility of their being true without its being true.
As before, the logically relevant aspects of an expressions meaning constitute its
sense. In specifying the sense of a connective or quantifier, we shall certainly have
to say how it contributes to the determination of consequence. Given Aristotles
Thesis, then, in specifying the senses of the connectives we shall have to lay down
principles that relate the truth of complex statements at various possibilities to
the truth of their components at those and other possibilities.
158 Five Attacks on Classical Logic

If, however, we do this by relating the truth of complex statements at various

worlds to the truth of their components at those and other worlds, we perforce
build a strong assumption into the theory. As we have seen, the worldly semantic
principles presuppose that each statement is either true or false at an arbitrary
world. Intuitionists, of course, will not accept that presupposition, so the use of a
possible-worlds framework is incompatible with our project of specifying senses
for the connectives that are common ground between rival logical schools.
There is, by contrast, no incompatibility between our project and Aristotles
Thesis. Many intuitionists, like many classical logicians, recognize the modal ele-
ment in the concept of consequence. Again, there is no incompatibility between
the project and moderate modal realism. (On the contrary: moderate realism
will be an attractive position for an intuitionist to adopt.) So the way forward is
clear:we should seek a modal semantic theory cast in terms of possibilities rather
than worlds. Non-worldly semantic principles for the logical words, if we could
find them, would enable the moderate realist to apply his theory to a non-classical
logic to which the worldly version of the theory is inapplicable.
In fact, even if we accept classical logic and classical semantics, there are still
reasons for eschewing worlds if we can. Stalnaker describes himself as a func-
tionalist about modal entities. Possible worlds, he tells us, are what people dis-
tinguish between in their rational activities. To believe in possible worlds is to
believe only that those activities have a certain structure, the structure which
possible worlds theory helps to bring out (1976, 38). Or again:The concept of a
possibilitya way a world might beis to be explained functionallyas what
one is distinguishing between when one says how things are (2003, 8).
This functionalist conception of modal entities is attractive, for it offers the best
prospect of eventually harmonizing the metaphysics of modality with its episte-
mology, by explaining why our actual standards for making modal judgements
are appropriate to their truth-grounds. We often judge something to be impos-
sible because we discern incoherence in the supposition that it obtains. Thus, we
judge that there is no possible village where the barber shaves all and only those
villagers who do not shave themselves, because the hypothesis that there is such
a village leads to a manifest dead-end. Explaining why this is a reasonable basis
for modal judgements is a notorious problem for concretism:it is unclear why a
suppositions being incoherent should entail that there is no part of the concrete
pluriverse where it obtains. Afunctionalist, on the other hand, can explain this.
In our rational activities, we set aside suppositions that we know to be incoherent
ab initio:they are not things of which a rational thinker needs to take account.
Attractive as functionalism is, though, the entities whose postulation it jus-
tifies seem to be possibilities rather than worlds. Consider Stalnakers favourite
 Possibilities 159

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.

6.3 Previous Unworldly Theories

For these reasons, we should investigate whether there are alternative semantic
principles that can specify the modal aspect of the senses of or and not in terms
of possibilities. But have such principles already been found? In an interview that

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.

6.4The Structure of the Space of Possibilities;

Truth-Ground Semantics
A generalization of Humberstones notion of a refinement, though, is the key to a
better approach to the problem. Where U is any set of possibilities, let us call the
closure of Uwritten Cl (U)that set comprising all and only those possibilities that
7
include what all the members of U have in common. What the members of U have
in common is given by the statements, or possible statements, that are true at every
member of U.Thus we have the following equivalence:

x Cl (U) i f and only if

for every statement or possible statement A, A is true at x
if A is true at every u U.

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:

(C) |A B|=|A| |B|.

We also have the following general axiom of a semantic theory based on

possibilities:

(R) The truth-grounds of any statement form a closed set of possibilities.

The argument for (R) is straightforward. Let A be an arbitrarily chosen statement

and let U be the set of its truth-grounds. Consider an arbitrary possibility, x, that
belongs to the closure of U. Since the closure of U comprises all those possibili-
ties at which every statement that is true throughout U is true, if A is true at every
member of U, then A is true at x. But since every member of U is a truth-ground
of A, A is true at every member of U, so A is true at x, i.e., x is a truth-ground of A.
But then, since U comprises all the truth-grounds of A, x must be a member of U.
That is to say, any possibility belonging to the closure of U must belong to U itself,

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

i.e., Cl (U) U. By INCREASING, U Cl (U) for any U. Thus Cl (U)=U, showing

that U is closed, as required.
This shows what form a logico-semantic theory cast in terms of possibilities
will take. The theory will associate with each atomic statement a closed set of pos-
sibilities as the statements truth-grounds. The theorys compositional principles
will then associate a closed set of possibilities with each molecular statement.
13
Since the intersection of any two closed sets of possibilities is closed, our postu-
late (C) for conjunctions respects our general principle (R).
Principle (R) also points towards the right semantic postulate for disjunction.
As we remarked, some interesting logics do not validate all of De Morgans Laws,
but all the logics that philosophers take seriously as codifications of our deduc-

tive norms validate the rule of -introduction:A entails A B , as does B. When
C entails D, the truth-grounds of D include those of C, so it follows from the
validity of -introduction that a disjunctions truth-grounds will include those
of its disjuncts:|A B| |A| |B|. All the logical schools with which we are
concerned also accept the restricted form of -elimination or proof by cases:if

A entails C and B entails C, then A B also entails C. (This law, in which side
premisses are not allowed, is valid in classical, intuitionist, and even quantum

logics.) This means that A B is the strongest statement whose truth-grounds
include the union |A| |B|. Since the truth-grounds of any statement form a

closed set, it follows that the truth-grounds of A B are the smallest closed set
to include that union. But the smallest closed set to include U is its closure, so
wereach

(D) |A B|=Cl (|A| |B|).

Although somewhat unfamiliar, (D) is intuitive:it says that the truth-grounds

of A B are what the truth-grounds of A, and B, have in common. Since the
closure of any set is closed, postulate (D) respects principle (R). Although Ican-
not argue for the claim now, Ibelieve that natural generalizations of (C) and (D)
serve as plausible modal semantic principles for the universal and existential
quantifiers.
There is in fact a close connection between (C) and (D) and the semantic axi-
oms that E.W. Beth proposed for conjunction and disjunction (see Beth 1959,
145; Dummett 2000, 5.4). Beth explained the meanings of the logical connec-
tives by reference to trees, which consist of a set T of nodes (in general infinite)
under a partial order . Beth thought of each node as representing a subjects state
of information on a given day; when a b but there is no c such that a c b, b

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 |),

just as in (D). Moreover, closure in the present sense is INCREASING,

14
IDEMPOTENT, and MONOTONE. Given (C) and (D), it is easy to show that
conditions (i)and (ii) hold for any complex statements formed using conjunction
and disjunction and hence that the set of nodes at which such a statement is true
is closed. The theory of truth-grounds, then, generalizes Beths semantics for the
language of intuitionistic logic.

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 ).

More generally, we can say that a combination of a finite collection of possibili-

ties is a maximally weak possibility (with respect to the partial order of deter-
mination) that determines all the combinants. Even in a minimal metalogic, we
have that when z1 and z2 combine x and y, each will determine the other, so that
z1=z2, since the determination relation is anti-symmetric. Whenever x and y are

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.

6.5 Distribution and Quantum Mechanics

The semantic principles (C), (D), and (N) are acceptable to adherents of many rival
logical schools; these principles, then, have a good claim to articulate the commonly
understood senses of the sentential connectives. Ishall illustrate this by considering
the challenge, due to Hilary Putnam (1968), to the classical Law of Distribution that
17
arises from the fact that certain operators in quantum mechanics do not commute.

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

In order to expound this challenge, it suffices to consider a simple physical sys-

tem, S, which comprises only a single subatomic particle that can move in only
one dimension. Indeed, we shall chiefly be concerned with the state of the system
only at a single time, t. Ishall say more about the states of this system soon.
We also consider a simple formalized language, L, whose statements attribute
various physical properties to S at t. Specifically, each atomic statement of L either
(a)attributes to S a precise position at t or (b)attributes to S a precise momentum
at t. We may think of Ss position as being specified by a real number that identi-
fies a place along an axis, with the numbers significant figures implicitly defin-
ing the limits of precision; a more elaborate but essentially similar specification
gives Ss momentum. We assume that some atomic statement of L describes each
of Ss physically possible positions and momenta at t. The compositional devices
in L are the connectives , , and ; L contains no conditional connective and
no quantifiers. The well-formed formulae of L are generated from its atoms in the
usual way.
The Law of Distribution says that any formula in the form A (B 1 ... B n)

logically entails the corresponding formula (A B 1) ... (A Bn) . Given

only uncontroversial assumptions about the logic of conjunction, Distribution

may be derived from the classical -elimination rule, which licenses an infer-

ence of C from A B together with a set of side premisses X, whenever C
18
follows both from A together with X and from B together with X. In the spe-
cial case where X is empty, Putnam allows that this rule is part of the sense of
disjunction (1968, 189). He proposes, though, that we should only accept the
elimination rule in this restricted form, without side premisses:the derivation
of Distribution is then blocked.
What is Putnams ground for restricting full -elimination and Distribution
when we are engaged in deductive reasoning about systems of subatomic par-
ticles? He advances three arguments. The first is that the Law yields the wrong
probabilities in the two-slit experiment (Putnam 1968, 1801). Ishall not discuss
this argument:paradoxical as the two-slit experiment may be, it is widely agreed
that restricting Distribution does not help in calculating the correct probabilities
in it (see Gardner 1971, especially 3 and 4, and Bub 1982). The second argument is
that restricting the Law is required to resolve the paradox of (for example) an elec-
trons being able to pass through a potential barrier (Putnam 1968, 1814). Ishall

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 should all agree that it is incoherent to describe a measurement as bringing

into being the quantity that it measures. As Putnam remarks, if a procedure dis-
torts the very thing it seeks to measure, it is peculiar it should be accepted as a good
measurement (1968, 183). The same point holds if a procedure brings into exist-
ence the very thing it purports to measure. So any interpretation of L that forces
one to conceive of measurement in this way is to be rejected. Putnams contention
is that avoiding this rebarbative conclusion requires restricting Distribution.

A premiss of Putnams argument is that every conjunction of the form Si Tj
is inconsistent. He does not need to make so strong a claim. Suppose the Law of
Distribution were valid. Then, for any statements A, B1,...,Bn, and any logical pos-
sibility x, if A (B1 ... Bn) is true at x, (A B1) ... (A Bn) is also true at x.

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

(A B1) ... (A Bn) is untrue. Putnam selects A to be Sz, B1 to Bn to be T1 to

TR, and he takes x to be a physically possible circumstance at which Sz is true. His

argument is then that Sz (T1 ... TR) the premiss of Distributionis true

170 Five Attacks on Classical Logic

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

the falsity of Sz (T1 T2 ... TR) .

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

Putnams rejection of strong realism has much to recommend it. In effect, he is

arguing that, if we reject strong realism, we must restrict Distribution if we are
not to be forced into the strong anti-realism of saying that an act of measurement
brings what it measures into existence. Given that strong anti-realism is absurd,
and that strong realism is empirically problematical, we have here an apparently
powerful challenge to classical logic. It is in this spirit that I shall pursue the
matter.
Setting strong realist interpretations aside, then, let us revert to our original

question, and ask why no conjunction in the form Si Tj can be true at any
physical possibility.
There is, Icontend, a powerful argument for claiming that no such conjunction
can be true. The argument rests on reflections about what it could mean to ascribe
truth to the statements of a language such as L, given the correctness of quantum
mechanics. Those reflections in turn stem from considerations about quantum
theory itself, relevant parts of which Imust now sketch.
The theory treats of the various physically possible states of a system such
as S. It represents these states by wave functionsfunctions that are solutions of
Schrdingers Equation. In the case of S, considered at a fixed time t, each such
function will be a function (x) of a single variable. The theory postulates that the
set of all wave functions representing the various physically possible states of the
system constitute a complex vector space; two vectors (i.e., wave functions) describe
the same physical state if and only if one is a multiple of the other. The theory fur-
ther postulates that there is a well-defined inner product operation on the vector
space. In most applications of the theory, the inner product <|> is the integral
(x). (x) dx. In any event, it is postulated that <|> is always the complex con-
jugate of <|>, that the inner product is always conjugate linear in its first argu-
ment and linear in its second argument, and that the inner product of any non-zero
vector with itself is a positive real number. (The integral (x).(x) dx meets these
conditions.) In fact, the theory also postulates that the vector space is complete and
has a countable dense subseti.e., is a Hilbert space.
I have said that a wave function represents a physical state, but how does it do
that? The generally accepted answer, due to Max Born, is the so-called statisti-
cal interpretation of quantum mechanics, whereby the wave function specifies
the probability that the system has various observable attributes. So for exam-
ple, the probability that the particle described by the wave function (x) lies in
2 2
the subset T of the real line R is the ratio T |(x)| dx | R |(x)| dx. When the
denominator of this ratio is unity, the wave function is said to be normalized;
in that case the probability that the particle described by lies in T is simply
2
T |(x)| dx. Astrong realist will understand the relevant notion of probability
172 Five Attacks on Classical Logic

epistemically:the particle has a position which is either in T or not, but we can-

not in general know whether it is in T or not; all we can do is to lay down laws say-
ing how probable it is that that its real position lies within T. Amore mainstream
view takes the fundamental laws of the theory to be statistical. There are no
further factssuch as the particles precise position and precise momentum
which the theory omits to characterize. Rather, the physical state of the system
consists in various measurable outcomes having such-and-such probabilities.
This brings me to the second principal notion of which quantum mechan-
ics treats. An observable is a physical quantity that we are sometimes able to
measuresuch as a systems position, momentum, or energy. The kernel of the
quantum theory is a set of rules that determine the expected value of an observ-
able when the system is in a given state.
From a mathematical point of view, observables are described by certain
kinds of linear transformations on the vectors in the state space. So, in our sim-
ple one-dimensional case, the position operator X is defined by (X )(x)=x(x)
and the momentum operator P by (P )(x)=h/2i (x), where h is Plancks con-
stant and (x) is the first derivative of (x). The momentum operator is defined
only on functions that are differentiable, so the theory must allow for operators
which are defined only on a subspace of the state space. An operator A is said
to be self-adjoint when <A|>=<|A>. When the inner product is defined by
(x) . (x) dx it is easy to check that the position and momentum operators are
each self-adjoint, and this is taken to be a general characteristic of observables
in quantum mechanics:they are self-adjoint linear transformations on the state
space, or on a suitable subspace of it.
I said that the kernel of the theory provides rules that determine the expected
value of an observable in various states. When the state is described by the nor-
malized function (x), the expectation value of the observable A, E (A), is defined
to be < |A>. To see the rationale for this definition, consider the expectation
value of the position operator X against the integral inner product. In that case,

E (X) = < |X>=R (x) x(x) dx = R x.|(x)|2 dx,

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

which corresponds to the classical definition of the standard deviation of a ran-

dom variable. The notion of dispersion enables us to formulate Heisenbergs
Principle precisely:for any state , the product of the dispersion of position and
the dispersion of momentum is never less than a certain value:

( 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.

A theorem of quantum mechanics says that the dispersion of A vanishes at

only when A=E (A)i.e., only when is an eigenvector of A with eigenvalue
E(A). Thus a statement ascribing a precise value to an observable will be true
174 Five Attacks on Classical Logic

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.

(1) In the earlier part of Is Logic Empirical?, Putnam proposed a prescrip-

tion for coordinating basic physical statements to subspaces of the Hilbert space
for the relevant system whereby the assignments for the atoms were extended to
compounds by the following rules (in which S(A) is the subspace corresponding
to the statement A):

S(A B)=the intersection of the spaces S(A) and S(B),

S(A B)=the span of the spaces S(A) and S(B),
S(A)=the orthocomplement of the space S(A) (Putnam 1968, 178).

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

indeed form a non-distributive lattice, and that non-distributive lattice is useful

in quantum-theoretical calculations. But we have as yet been given no reason to
assign any logical significance to the lattice. To do so, it would need to be argued
that, when the statement A is true at precisely the states in S(A), and when the

statement B is true at precisely the states in S(B), then the disjunction A B
is true at precisely the states in the span of S(A) and S(B). Putnam gives no such
argument, so his claim that the proposed rules have any logical significance is
unsupported.

(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.

(3) We do much better, Ithink, if we apply to language L the semantic frame-

work of 6.4. Our semantic axiom (D) serves to characterize the truth of dis-
junctions at various sorts of (perhaps incompletely specified) possibilities; so in
saying at which physically possible states a disjunctive statements is true, we may
invoke (D) without giving or a sense that is specific to discourse about subatomic
particles. What is particular to that discourse comes out, rather, in the relation-
ship between closure and incompatibility. In 6.4 Itook closure to be a primi-
tive notion of the semantic theory and defined incompatibility in terms of it, but
in the present case it makes sense to adopt the reverse procedure. Two physical
states may be said to be incompatible when one of them precludes the obtaining
of another, and the quantum theory affords a natural account of this relation of
preclusion. The transition probability between two normalized states and is
2
defined to be |<|>| . This probability has two (mutually consistent) physical
interpretations. First, it is the probability that the state of the system, before a
measurement is effected, is , given that the measurement measures the system
as being in state . Second, it is the probability that a system that starts in state
should be in after some time has elapsed. Together, these interpretations tell
us that, when the transition probability is zero, state precludes the obtaining
of and vice versa, for the condition is symmetrical. In the case of quantum
mechanics, then, we can further determine the relation of incompatibility by
saying that

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

state. Assuming that the background space of possibilities is confined to those

that are physically possible, then, we have |Si Tj|=, for all i, j. By (D), the
semantic value of the conclusion of the problematic instance of Distribution,
viz., |(Sz T1) ... (Sz TR)|, will be ( ... )=. What is the set ?
Any state whatever precludes (vacuously) the obtaining of every member of the
empty set. So every state whatever belongs to . Accordingly, if then
must be incompatible with itself, i.e., <|>=0. This implies that is the zero
vector, and since the state space of a physical system is assumed to comprise only
non-zero vectors, we have that =. Our semantic principles, then, tell us
that no physical state verifies the conclusion (Sz T1) ... (Sz TR) of the

problematic instance of Distribution. So Putnam is right to say that that conclu-

sion is untrue.
What, though, of the premiss of that instance of Distribution? Is Putnam right
to say that Sz (T1 ... TR) is true at some physical state? Here, matters are

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.

6.6 Distribution and Regularity

Putnams attempt to refute the Law of Distribution, then, is unconvincing. But is
there positive reason to accept the Law? The present section shows how the seman-
tic framework of 6.4 combines with a main thesis of Chapter3 to produce one.
It helps to begin by considering how the notion of consequencein particular,
that of a conclusions following from more than one premissmight be explained
within that framework. Ihave been defending Aristotles Thesis that there is a
modal element in the idea of consequence:whenever a conclusion follows from
some premisses, it is necessary that the conclusion should be true if the premisses
are. Anatural initial elaboration of the Thesis says this:when a conclusion follows
from some premisses, there is no possibility of the premisses being true with-
out the conclusions being true. We explained truth at a possibility as follows:a
statement is true at a possibility if, necessarily, things would have been as the
statement (actually) says they are, had the possibility obtained. This suggests the
following account of consequence (for a given background space of possibilities,
S):conclusion B is a consequence of premisses A1,...,An if and only if for every
possibility x in S, B is true at x if all of A 1,...,An are true at x. This definition clearly
validates Aristotles Thesis, so shifting from worlds to possibilities more generally
is no impediment to defining consequence in a way that respects the Thesis.
182 Five Attacks on Classical Logic

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

This postulate allows us to introduce into the language a well-behaved logi-

cians conditional. Whatever may be true of the vernacular if...then, what marks
out the logicians conditional is that it conforms to the Law of Implication:

A B follows logically from X if and only if B follows logically from X {A}.
With this in mind, we define a binary operation on sets of possibilities as fol-
lows:x UV if and only if, for any possibility u U, x u V. We may then

identify the truth-grounds of A B as follows:

(I) |A B| = |A| |B|.

When the underlying space of possibilities is regular, this conditional is a

statement-forming operator on statements:for in that case, UV is closed when-
25
ever V is.
Regularity also validates the unrestricted form of -elimination and hence the
Law of Distribution. As our discussion of Putnam brought out, while the restricted
form of -elimination (the form without side premisses) is built into our semantic
postulate for disjunction, (D) does not by itself ensure the validity of the unre-
stricted form of the rule. In tandem with Regularity, though, it does. The proof is
simple. Suppose that C follows from X, A and also from X, B. Then |X||A| |C|
and |X||B| |C| so that |X||A| |X||B| |C| whence |X|(|A| |B|) |C|. We
require to prove that C follows from X, A B, i.e., that |X||A B| |C|. Let y be
an arbitrarily selected member of |A B|. By (D), y Cl (|A| |B|). Now let x
be an arbitrarily selected member of |X|. By Regularity and the commutativity of
, x y Cl (x (|A| |B|)). Since |X|(|A| |B|) |C|, it follows that x y Cl (|C|),
whence x y |C|, since |C| is closed. Since x and y were arbitrarily chosen from
|X| and |A B|, this shows that |X||A B| |C|, as required (cf. Sambin 1995,
especially 864, 865 (lemma 2), and 868 (theorem 4)). On the present account, then,
the general form of proof by cases owes its soundness to two factors:one specific
to the meaning of , namely, its conforming to (D); the other, the absoluteness of
properly logical consequence. This result is far from being unattractive.

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

As we have just seen, a semantic theory in which truth is relativized to pos-

sibilities that may be incomplete casts light upon the dispute over the validity
of the Law of Distribution. It also illuminates, Icontend, the dispute between
classical and intuitionist logicians that was the main focus of Chapters4 and 5.
More particularly, Ishall argue that it can answer the challenge to classical
logic left hanging at the end of Chapter5. That challenge was as follows. If we
specify the senses of the connectives by way of the standard truth-tables, then
we may justify classical logic only by appeal to the Principle of Bivalence. That
Principle is doubtful:it is far from clear that Cantors Continuum Hypothesis,
for example, or a statement ascribing a vague predicate to one of its border-
line cases, is either true or false. Accordingly, if the senses of the connectives
are given by their truth-tables, classical logic rests upon an insecure basis. As
the present chapter will show, if we take the senses of the signs for conjunc-
tion, disjunction, and negation to be given by the axioms (C), (D), and (N)
proposed in 6.4, we may justify the classical logical laws without appealing
to Bivalence. What is interesting, though, is that this way of putting classical
logic on a firmer basis also reveals the real strength of the intuitionistic alter-
native to it. If the senses of the connectives are given by their truth-tables, the
merits of intuitionism are hard to discern:given Bivalence, all the classical
laws are validated, but without it, some intuitionistically acceptable laws (as
well as some classical ones) come out invalid. By contrast, all the intuitionistic
laws are validated by the semantic principles laid down in Chapter6. Indeed,
intuitionistic logic will emerge as the strongest logical system that may be
justified on the basis of those specifications of sense alone. This relationship
between the classical and intuitionistic systems enables us to see more clearly
what is at stake in the choice between them.
 Challenges from the Infinite and from the Infinitesimal 185

7.1 The Semantics and Logic of Negation

Central to that choice, of course, is the treatment of negation.
We may begin by recalling the axiom for negation that was proposed in Chapter6.
Abasic assumption of truth-grounds semantics is that the closure operation Cl is well
defined on the space of logical possibilities. It was in terms of this closure operation
that we defined the relation of determination:one possibility determines another
when the former belongs to the closure of the latters singleton; x y if and only if
y Cl ({x}). We then defined incompatibility between possibilities in terms of deter-
mination:two possibilities are incompatible when no possibility determines both;
x y if and only if z (x z y z). In terms of incompatibility, we defined the
orthocomplement of a set of possibilities:this comprises precisely those possibilities

that are incompatible with all the members of the original set; x U if and only if
x y for all y U. This notion enabled us to state our semantic axiom for negation:the

truth-grounds of A are the orthocomplement of the truth-grounds of A, i.e.,

(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:

(NC)For no set of possibilities U does there exist a possibility which belongs

both to the closure of U and to the closure of U . In other words, given

any set of possibilities U, z (z Cl (U) z Cl (U )).
186 Five Attacks on Classical Logic

As its label suggests, principle (NC) is a form of the Principle of Non-

Contradiction. It says that there is no possible state of affairs which at once
includes what is common to all the states of affairs in U, and also includes what
is common to all the states of affairs that are incompatible with all the members
of U. (NC) directly yields the theorem that any possibility in the closure of U is
incompatible with any possibility in the closure of U:

(1) If x Cl (U ) and y Cl (U ) then x y.

Proof: Suppose for a contradiction that z (x z y z). Let v witness this

existential supposition, so that x v and y v. Since x Cl (U), {x} Cl (U), so
Cl ({x}) Cl Cl (U), for Cl is MONOTONE. Since Cl is also IDEMPOTENT,
Cl ({x}) Cl (U). By supposition x v, whence v Cl ({x}) by the definition of
. Hence v Cl (U). Aprecisely parallel argument shows that v Cl (U). But

then z (zCl (U) z Cl (U )), contrary to (NC). This contradiction reduces to
absurdity the supposition that z (x z y z), so we may conclude by Reductio
that z (x z y z), which by definition is x y. This completes the proof,
which could be formalized as a valid deduction in either classical or intuitionistic
logic.
Theorem (1)in turnyields

(2) For any set of possibilities U, Cl (U ) (Cl (U)).

Proof:Suppose that y Cl (U ). Let x be any member of Cl (U). By theorem (1),
x y, whence y x by the symmetry of . That is, y (Cl (U)), by the definition

of . Since y was an arbitrarily chosen member of Cl (U ), Cl (U ) (Cl (U)), as
required. Again, this deduction is acceptable to an intuitionist and hence also to a
classicist.
Theorem (2) ensures that when the truth-grounds of A form a closed set,

so do those of A . Because Cl is INCREASING, we automatically have
|A| Cl (|A|), so we need only show that Cl (|A|) |A| whenever |A| is
closed. By (N), |A|=|A|, so it suffices to show that Cl (|A|) |A|, whenever
|A| is closed. Suppose, then, that |A| is closed. By Theorem (2), Cl (|A|)
|Cl (A)|=|A|, since |A|=Cl (|A|), and we are done. This whole chain of deduc-
tions from (NC) is acceptable to both intuitionists and classicists.

Principle (NC) is more than is needed to ensure that the truth-grounds of A
form a closed set whenever the truth-grounds of A do. In particular, a dialetheist
logician who wishes to use the current semantic framework may prefer simply to
impose theorem (2)as a constraint on the closure operation. This will give him a
well-behaved negation operator, whose outputs respect (R), while abjuring (NC),
a form of the Principle of Non-Contradiction.
 Challenges from the Infinite and from the Infinitesimal 187

It would be possible to found the account of negation to be developed in this

chapter upon the postulates that Cl is a well-defined operation on the space of
logical possibilities, and that both Regularity and (NC) hold of that space. In
fact, though, Ishall take a slightly different approach in order to tie up two loose
ends from Chapter6. We said there that B follows logically from A 1,...,An when
any combination of truth-grounds of the premisses is a truth-ground of the
conclusion:

| A1 || An | | B |.

However, our earlier definition of combination presumed that combinants are

compossible. Since both classical and intuitionist logicians suppose that it makes
sense to speak of the consequences of inconsistent premisses, our account of con-
sequence needs supplementation. The same goes for the formulation of the notion
of regularity in 6.6:a space is regular if x y Cl (V y) whenever x Cl (V), but
x y has not yet been defined when x and y are incompatible.
One way of filling this lacuna would be to give supplementary accounts of
what it is for a conclusion to follow from inconsistent premisses, and of what reg-
ularity amounts to, for incompossible combinants. From a formal point of view,
however, it is more elegant to extend the definition of the combination operator
so that it becomes a totally defined operation on our underlying space. When
a finite collection of possibilities are compossiblei.e., when some possibility
determines each of themwe said that z combines the collection if z is a maxi-
mally weak possibility (with respect to the partial order of determination) that
determines all the combinants. Since determination is anti-symmetric, such a
combination is unique. When the members of the collection are not compossible,
there will be no possible state of affairs that determines each of them. Suppose,
though, that we supplement our space with an impossible state of affairs. What
properties would such a state of affairs have? We explained the notion of a state-
ments being true at a possibility using a necessitated conditional:A is true at
x if, necessarily, if x obtains then things are as A (actually) says they are. Since
it is necessarily the case that an impossible state of affairs does not obtain, any
statementindeed, any possible statementis true at any impossible state of
affairs. It follows that, if y is any impossible state of affairs, x y for any pos-
sible or impossible state of affairs x. In particular, any impossible state of affairs
determines any other. Since the relation of determination is anti-symmetric,
this means there is only one impossible state of affairs. If we extend the theory
of Chapter6 to accommodate impossible states of affairs, then, we need coun-
tenance only one of them, which Ishall designate (absurdity); we have x ,
whether x is possible or impossible.
188 Five Attacks on Classical Logic

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:

(Inc) x y iff x y = iff w ((x w y w) w = ).

The new definition directly entails that incompatibility is symmetric; it also

entails that the absurd state of affairs is the only state of affairs that is incom-
1
patible with itself. Our operation x y now becomes a total function on the
space of possible or impossible states of affairs, so our definition of what it is for
a space to be regular does not need to be supplemented with any conventions
about interpreting x y when x and y are incompatible; in any such case, x y
designates . Similarly, when A 1,. . .,An are logically inconsistent premisses,
|A 1| ... |An|={}.

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

The account of consequence recommended here, by contrast, makes no appeal

to what we can know from the senses of the connectives alone. Rather, it is a ver-
sion of Aristotles Thesis:a conclusion follows logically from some premisses if
the conclusion is true at every logical possibility at which all the premisses are
true. As such, it too ought to be acceptable to the classical logician. What delivers
a non-classical logic is that the possibilities by reference to which consequence
is characterized are allowed to be less than fully determinate. So long as this
is allowed, we can show that the intuitionistic calculus is sound and complete
with respect to this notion of consequence, whether the metalogic is classical or
intuitionist. We would obtain classical logic if we were to restrict attention to the
subspace of possibilities comprising complete possible worlds. That is, we would
obtain classical logic if we were to postulate the Principle of Bivalence. As we have
seen, though, that postulate is doubtful so, as yet, we have no inkling of what
might justify classical logic.

7.2 Statements With and Without Backs

Whilst our analysis shows that classical and intuitionist logicians can attach the
same sense to the connectives, and to the central metalogical notion of logically
follows from, it threatens to leave the issue between them as a stand-off. In one
respect, our discussion has put the classical logician on the back foot. He uses DNE
to eliminate negations. Given that (N) specifies the sense of the negation sign, the
condition for this rule to be sound is what Ishall call the DNE condition:for any
statement A, any state of affairs which is incompatible with any state of affairs
which is incompatible with any truth-ground of A is itself a truth-ground of A;
symbolically, |A| |A|. The classicist, however, cannot justify the DNE condi-
tion from the sense of the negation sign; he needs some other reason for holding
that the space of logical possibilities meets it. Until he can provide such a reason,
the intuitionist can fairly complain that adopting classical logic would be a leap
of faith.
By the same token, though, the intuitionist has so far done no more than throw
down a challenge to the classicist. He has not yet provided any reason to think that
the DNE rule is actually unsound. Intuitionist logic is the strongest logic that can
be justified by reference to the senses of the connectives, and by the assumptions
about the space of logical possibilities that we have so far seen to be well-founded.
For all that has been said, however, additional postulates may characterize that
spaces structure, postulates that entail the satisfaction of the DNE condition. If
that condition is satisfied, then a thinker who has made the classical leap of faith
will land on Canaans side:his laws will preserve truth at every logical possibility.
194 Five Attacks on Classical Logic

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

being determinate in which circumstances a speaker would be entitled to assert

it; in mathematical discourse, these circumstances will be ones where the speaker
has a proof of the statement. When we widen our gaze to encompass hypothetical
uses of statements, however, this conception of determinacy of sense is plainly
inadequate:there are no epistemic conditions that need to be met before one is
entitled to entertain a hypothesis. When we try to say in general what entertain-
ing a given hypothesis involves, there is no better answer than that one is thereby
setting asidepro tempore and perhaps for the sake of argumentall and only
those possibilities that are incompatible with the hypothesized statement. That
answer presupposes that the statement has a back.
I think our tendency to assume that every statement has a back accounts for the
widespread adherence to classical logic even in cases where Bivalence is doubt-
ful. All the same, there are statements of which we do not seem entitled to assert
that they have backs, and which consequently pose a challenge to classical logic
in the framework of a decision between T1 and T2. The next sections examine
two classes of such statements, the first concerning the infinitely large, the second
concerning the infinitesimally small.

7.3 The Intuitionists on Infinity

Under Dummetts influence, contemporary philosophers are apt to associate
the intuitionist critique of classical mathematics and logic with the adoption of
verificationist semantic theories, in which the meaning of a statement is given by
specifying the conditions in which a speaker would be entitled to assert it, rather
than by specifying the conditions under which it would be true. It is important
to set these associations aside in reading the early intuitionists, for the found-
ing fathers of the school were not verificationists. In a paper of 1923, Brouwer
wrote that a complete empirical corroboration of the inferences drawn [about
the world of perception] is usually materially excluded a priori and there can-
not be any question of even a partial corroboration in the case of (juridical and
other) inferences about the past (Brouwer 1923, 336). Averificationist would con-
clude from that claim that talk about the past is meaningless; Brouwer, though,
expressly holds that it is meaningful. Indeed, he allows that the laws of classical
logic, including Excluded Middle, may validly be applied in reasoning about the
world of perception, so long as we are able to think of the objects and mecha-
nisms of [that] world ... as (possibly partly unknown) finite discrete systems (1923,
336, emphasis in the original). More exactly, it is the possibility of projecting a
finite discrete system upon the objects in question that is the condition of the
applicability of Excluded Middle to judgements concerning those objects. We see
198 Five Attacks on Classical Logic

here a fundamental difference between Brouwer and Dummett. For Dummett,

the basic mistake of the classical mathematicians is that they apply a realist or
truth-conditional semantic theory to the language of mathematics. For Brouwer,
by contrast, their error was to apply distinctively classical logical rules even in
the mathematics of infinite systems, where the rules condition of applicability
does not obtain. A.N. Kolmogorov, another pioneer of intuitionism, agreed with
Brouwer. He understood Brouwers writing to have revealed that it is illegitimate
to use the principle of excluded middle in the domain of transfinite argument
(Kolmogorov 1925, 416).
As Brouwers reference to infinite systems shows, the early intuitionists did
not impugn as unintelligible expressions, such as the sequence of natural num-
bers, that purport to designate infinite mathematical structures. They did, how-
ever, claim that talk about such structures, if it makes sense at all, must be capable
of being cashed out as talk about the mathematical principles that characterize
them. Thus, to say that the natural number sequence has a property is to say that
the property in question is entailed by the laws of Heyting Arithmetic, these laws
being the principles that intuitionists take to characterize that structure. This
marks a fundamental contrast with the finite case. Afinite structure might be
characterized by certain mathematical principles but, even when it is so char-
acterized, it may still have properties that are not entailed by the principles. As
one might put it, in the finite case the extension of certain mathematical princi-
ples will have properties over and above those consequential upon the principles
themselves. According to the intuitionist, this is conceptually impossible in the
infinite case. Afinite initial segment of an infinite sequence may have properties
over and above those entailed by the principles that generate the sequence. When
we speak of the infinite sequence as a whole, however, we must be referring (per-
haps elliptically) to the generating principles themselves. For the intuitionist, one
might say, infinite structures cannot be conceived purely extensionally. So to con-
ceive them is illegitimately to project into the infinite a notion that only makes
sense in the finite case.
This claim of the early intuitionists has great plausibility. It is in accordance
with the ordinary, common-sense notion of infinity as something which does
not come to an end (Dummett 2000, 41). It is also plausible to hold that parts of
classical mathematics violate that common-sense conception by treating infinite
structures as if they could be completed and then surveyed in their totality, in
other words, as if we could be presented with the entire output of an infinite pro-
cess (Dummett 2000, 41). Properly respecting the common-sense notion is cer-
tain to involve revising the corpus of classical mathematics. The central question
for our enterprise, though, is a different one:we need to ask why the intuitionistic
 Challenges from the Infinite and from the Infinitesimal 199

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

are foundational for the current mathematics of infinitywhich follow from

Excluded Middle. The second of these propositions is that every mathematical
species is either finite or infinite. He then constructs an ingenuous case in which,
as he thought, this second proposition cannot be asserted:
Let d be the th digit to the right of the decimal point in the decimal expansion of , and
let m=kn if, as the decimal expansion of is progressively written, it happens at dm for the
nth time that the segment dm dm+1 ... dm+9 of this decimal expansion forms the sequence
0123456789 ... That the second fundamental property is incorrect is seen from the exam-
ple provided by the species of the positive integers kn defined above. (Brouwer 1923, 337)

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

though accepting A is tantamount to rejecting B, where B=The segment 0123456789

does not occur in the expansion of . But to be tantamount to accepting A, B would
have to be interpreted so as to include the case in which 0123456789 merely happens
never to occur in the expansion of , and this interpretation would violate Brouwers
strictures on the infinite. Giving B the interpretation that it needs to bear in order
to serve as As back involves thinking of the decimal expansion of as a vast array
which might happen not to contain any instance of 0123456789. On Brouwers view,
to think in that way is precisely to conceive an infinite collection on the model of a
large finite one.
For Brouwer, A must have a negation:any statement does. However, in attaching
a sense to 0123456789 does not occur in the expansion of , the intuitionist does
not accept Eadem est scientia as a constraint on the relation between a statement
and its negation. Athinker does not always need a conception of what would be the
case if not P in order to have a conception of what would be the case if P. Rather, his
understanding of what would be the case if not P may draw upon a prior understand-
ing of what would be the case if P. It is this possibility that the intuitionist exploits
in making sense of B. Ex hypothesi, we have a conception of what it would be for
0123456789 to occur at some identifiable place in the expansion of identifiable,
that is, by means of a mathematical calculation or (more generally) a construc-
tion. Drawing upon that conception, we can then form the notion of a proof that
establishes that no such construction is possible. Such a proof will be the ground for
asserting 0123456789 does not occur in the series. Heytings account of the meaning
of negated mathematical statementswe now see more clearlyis not the product
of a general preference for a verificationist semantic theory. Rather, it is the result of
his attempt to ensure that the senses attached to those statements do not fall foul of
Brouwers strictures about the infinite.
How might a defender of classical logic respond to this challenge? One radical
answer was developed by Wittgenstein in Part V of his Remarks on the Foundations of
8
Mathematics (Wittgenstein 1978). Wittgenstein had a long-standing and deep-seated
commitment to (B). When we understand a statement (Satz), he wrote in the Notes
on Logic of September 1913, we know what is the case if it is true and what is the case
if it is false (Wittgenstein 1961, 94). Any statement is thus associated with true and
false poles. To accept the true pole is ipso facto to reject the false pole. The negation

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

operator, on Wittgensteins account, simply reverses the poles, so asserting that P

is ipso facto denying that not P, just as (B) has it. Eadem est scientia follows. This is
why the Tractatus (Wittgenstein 1922)makes no room for doubting the equivalence
between a statement and its double negation. These say the same thing (TLP 5.44);
indeed, in a fully perspicuous symbolism, double negations would vanish (TLP 5.254).
On the other hand, Wittgenstein appreciated the force of Brouwers objections to
the completed infinite, and his way of reconciling that appreciation with his com-
mitment to (B) was to reject the underlying assumption that statements such as A
make sense in all circumstances. We look down a computer print-out of the first
one million digits in the expansion of andlo and beholdwe spot a segment
0123456789. So, to the question Does that segment occur somewhere in the expan-
sion of ?, we confidently answer yes. Wittgenstein allows that we would answer the
question affirmatively in such a circumstance, but he insists that this does not show
that the question possesses a determinatesense:
If someone says:But you surely know what this pattern occurs in the expansion means,
namely thisand points to a case of [its] occurring,then Ican only reply that what he
shows me is capable of illustrating a variety of facts. For that reason Icant be said to know
what the statement means just from knowing that he will certainly use it in this case.
(Wittgenstein 1978, 271)

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

the hypothesis that no occurrence of 0123456789 is to be found in the entire expan-

sion, even though such an occurrence is not precluded by the rule for expand-
ing . That hypothesis is incoherent on the view of the infinite that Brouwer and
Wittgenstein share, but Wittgenstein turns Brouwers argument on its head.
According to Wittgenstein, these considerations do not show that 0123456789
appears somewhere in the expansion of is a statement without a back. Rather,
what it shows is that we cannot assume that these words will always constitute a
statement, i.e., express a complete thought:

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

problematical statement, then, we do not have to make an illegitimate projec-

tion into the infinite of a notion that only makes sense in the finite. To the con-
trary. The way just sketched of making sense of The pattern occurs nowhere in
the expansion respects Brouwers principle that talk about infinite structures
must in the end be explained in terms of the mathematical principles that char-
acterize those structures.
But surely, it will be objected, we can make sense of the possibility that
0123456789 does not occur anywhere in the expansion of , even though it is
unprovable that 0123456789 does not occur in the expansion. And if it is unprov-
able that 0123456789 occurs nowhere in the expansion, how can it lie in the rule
for expanding that this sequence of digits does not occur anywhere in it?
We may answer the objection as follows. What it is for P to lie in the rule for
expanding is that the rule in question, together with the axioms that char-
acterize the natural numbers, entail that P. But these principles may entail
that P without its being provable that P. The crucial point is that the underly-
ing logic of the axioms that characterize the natural-number structure must
be stronger than first-order. If the underlying logic were first-order, then the
Upward Lwenheim-Skolem Theorem would apply, and the axioms could not
characterize the natural-number structureor any other infinite structureup
to isomorphism. Indeed, whatever the precise underlying logic of a categorical
characterization of the natural numbers may be, it will not possess a complete
axiomatization:if it possessed such, then the set of true statements of arithme-
tic would be recursively enumerable, contrary to Gdels First Incompleteness
Theorem.
This gap between entailment and deducibility makes room for a position that
respects Brouwers insight that the infinite differs fundamentally from the finite
while maintaining (contra Brouwer and Wittgenstein) that the statement A both
makes sense and conforms to the laws of classical logic. If the pattern 0123456789
occurs somewhere in the expansion of , then the characteristic axioms for the
natural-number structure, together with the rule for expanding , jointly entail
that it does; in this case, it will also be possible in principle to prove that the pat-
tern appears. If, on the other hand, the pattern does not occur anywhere in the
expansion, then the characteristic axioms and the rules for expanding again
entail that it does not; in the latter case, though, it may not be possible to prove

that the pattern does not occur. In asserting A A , then, a classical logician is
committed to the possibility that some mathematical truths may not be provable.
However, he is not committed to the possibility of truths about infinite structures
that are not entailed by appropriate mathematical laws and hence he is not com-
mitted to trying to make sense of the idea that a universal generalization about
206 Five Attacks on Classical Logic

natural numbers may be true by accident. To be sure, quantification over infinite

structures larger than the domain of natural numbers may generate new chal-
lenges to classical logic; in Chapter9, Ishall examine a challenge to the use of
this logic in set theory. But we have, Itrust, an answer to the particular challenge
to classical logic that Brouwer threw down on the basis of his species of kns
an answer that respects his insight that talk about infinite mathematical struc-
tures needs to be cashed out in terms of the principles that characterize those
structures.
This reply to Brouwer assumes that we can apprehend principles that charac-
terize the natural numbers up to isomorphism. Some familiar principles of this
kind will fall foul of intuitionistic strictures. The best-known categorical axioma-
tization of number theory is second-order Peano Arithmetic, in which the princi-
ple of mathematical induction is formulated as a single axiom, namely,

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

designatum of t2 in the ancestral of the relation signified by . More formally, in

any interpretation I of this augmented language, an assignment s of objects to var-

iables satisfies Axy t1 t2 if and only if the denotation of t1 under s stands to the

denotation of t2 under s in the ancestral of the relation that I assigns to (x, y) .
One obtains ancestral logic by holding this interpretation of Axy fixedi.e., by
deeming the ancestral operator to be a logical constant. It is then a straightforward
matter to construct a categorical axiomatization of number theory in ancestral
logic (see e.g. Shapiro 1991, 228). The crux is the axiom that rules out non-standard
numbers:where Sxy if and only if x is the immediate successor of y (i.e., if and only
if x=s(y)), we lay down that n is a natural number if and only if Axy (Sxy) n 0.That
is:n is a natural number if and only if n stands to 0 in the ancestral of the relation
of immediate succession. This is precisely Freges definition of natural number
i.e., of finite cardinal, endliche Anzahlin 83 of Die Grundlagen (see Frege 1884,
p.96). The present formulation excludes non-standard numbers because it is built
into the definition of the ancestral R* that when R*xy there are only finitely many
R-related intermediaries between x and y.
Commenting on ancestral logic and some related systems that are also inter-
mediate in strength between first-order and full second-order logic, Stewart
Shapiro remarks that finitude can be characterized in them, but if anything has
the advantages of theft over toil, this does. The notion of finitude is explicitly
built in to the systems in one way or another (Shapiro 1991, 238). Shapiro is quite
right to say that finitude is built in to ancestral logic. In the present dialectical
context, though, the use of that logic is not so much a matter of theft as of using
a notion whose intelligibility is granted by the opposition. The intuitionist is not
afflicted by any physicalist, or Skolemite, doubts about the clarity, or the abso-
12
luteness, of the distinction between finite and infinite. To the contrary:when
he criticizes the classical mathematician for treating infinite collections as large
finite ones, the intuitionist presupposes that we shall understand the difference
between them. We can be confident, then, that he will understand the explana-
tion of the ancestral operator and consequently understand the characterization
of the natural-number structure in ancestral logic.
All the same, there is something important that the use of that logic makes
clear. To put the point in Wittgensteins terms, what ancestral logic makes clear
is how the absence of 0123456789 from the decimal expansion of might be
in the rule for expanding even though it is impossible to prove that this is
so. For, although weaker than full second-order logic, there can be no sound
and complete axiomatization of ancestral logic either (see Shapiro 1991, 229).

12
For an interesting discussion of the first sort of doubt, see Field 1998.
208 Five Attacks on Classical Logic

In fact, ancestral logic is considerably weaker than second-order logic. While

the Upward Lwenheim-Skolem Theorem fails for it, the Downward Theorem
holds (Shapiro 1991, 238), so we cannot, in ancestral logic, give a categorical
characterization of the real number-line, as we can in full second-order logic.
While most classical mathematicians will regard this result as showing that
ancestral logic is unacceptably weak, it confirms the view that ancestral logic
is part of the common ground they share with intuitionists. Intuitionists find
the classical conception of the continuum unintelligible:the fact that ancestral
logic provides no path to a categorical articulation of that conception supports
the claim that that logic involves nothing that an intuitionist should reject as
un-kosher.
Our analysis, then, provides a reply to Brouwers problem case for the Law
of Excluded Middle: while respecting his strictures about the infinite, we
may assert The species of kns is either inhabited or not. But what about other
instances of Excluded Middle which intuitionists claim we are not entitled to

assert? Under the Heyting semantics, we may assert A A only when either

A or A is provable, i.e., only if the problem of deciding A is solvable. As we
saw in 5.3, on this way of understanding the connectives, we cannot now assert

GC GC , where GC says that every even number greater than two is the
sum of two primes. At the time of writing, no one has a proof or a refutation of
13
this conjecture, or a demonstration that a proof or refutation must exist. What
happens under our way of understanding the connectives? Do we need to run
through all the natural numbers, and say that that process must either have
thrown up a counterexample to the Conjecture or not, in order to be entitled to

assert GC GC ?
In this case too, there is no need to appeal to that incoherent picture in order

to justify asserting GC GC . I have argued that someone who respects
Brouwers strictures on the infinite can still be brought to understand principles
that provide a categorical characterization of the domain of natural numbers.
Such a thinker will then possess a determinate conception of that domain. Now

x(x) x(x) expresses the determinacy of the domain of quantification
over which the variable x rangesassuming it is determinate whether an arbi-
trary member of that domain is . Simply on the strength of his determinate

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:

NOS n((n) (n)) n(n) n(n).

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

(1) n(Fn Fn) nFn nFn.

Since any statement Fn is clearly bivalentindeed, is decidableit is agreed on
all hands (even by intuitionists operating under the Heyting semantics) that we
may assert n(Fn Fn), which is the antecedent of (1). We may, then, detach to
reach

(2) nFn nFn.

The first disjunct of (2) is GC. Furthermore, xx intuitionistically entails

xx , so that the second disjunct intuitionistically entails nFn

and hence nF , since Fn entails Fn for any n. Using proof by cases,

then, we have that (2)entails GC GC even in intuitionistic logic.
There is, then, no need to appeal to the conception of the infinite that
Brouwer rightly rejects in order to join the mathematical majority in asserting

GC GC . Our determinate conception of the domain of natural numbers
gives us title to assert NOS from which modes of reasoning that the intuitionist

can accept take us to GC GC .
Pace the intuitionists, then, we have yet to find a statement A in the language

of arithmetic for which we are unable to assert A A . So far as arithmetical
deductions are concerned, then, we have found no reason to devolve from classical
logics default status. This conclusion will be reinforced in 9.1, where Ishall show
how classical logic combines with an attractive account of arithmetical truth to
yield the conclusion that any statement in the language of arithmetic is either true
or false. What is interesting, however, is that the argument to be given there does
not generalize to establish the corresponding conclusion for statements involving

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

quantification over larger infinite totalities:in particular, it does not generalize

to statements involving quantification over all sets. Dummett was right, Ithink,
to claim that a wholly general argument for Bivalence for statements involving
quantification over infinite totalities must invoke the conception of the infinite
that Brouwer rightly deems incoherent. Our discussion suggests, though, that the
really doubtful cases of Excluded Middle involve quantification over larger, or
less tractable, domains than the natural numbers. Ishall pursue this theme in
Chapter9, when Idiscuss the use of classical logic in set theory.

7.4 AConsolation Prize for the Intuitionist:

Smooth Infinitesimal Analysis
For now, though, it is in place to examine another class of putatively backless
statements and to consider the threat that they pose to the unrestricted applica-
bility of classical logic. The statements Ihave in mind concern not the infinitely
large but the infinitesimally small.
According to Aristotle, no continuum can be composed of indivisibles:e.g. a
line cannot be composed of points, the line being continuous and the points indi-
visibles (Physics VI, 1:231a 2325). Ihave no space to consider the difficulties that
Aristotle identified in the rival conception, whereby a continuum such as a line
15
is made up of points. Those difficulties, though, are prima facie serious, so it is
worth considering how an Aristotelian account of the continuum might be devel-
oped. The basic problem for such an account is to say what the ultimate parts of a
continuum are, if they are not points.
A natural answer was given by LHpital:a curved line may be regarded as
made up out of infinitely small straight line segments. Similarly, one may regard a
bodys motion as composed of infinitely small instances of its moving at constant
speed in a straight line. LHpital captures the conception of continua implicit in
Isaac Barrows terminology of linelets and timelets, in Newtons talk of evanes-
cent quantities, and in Leibnizs of quantits inassignables.
For years, such talk was thought to be incoherent: Berkeley derided infini-
tesimals as ghosts of departed quantities. In the 1960s, Abraham Robinsons
non-standard analysis tried to revive them (see Robinson 1966), but Iam more
interested in a later mathematical developmentLawveres (1979) and Kocks
(1981) attempt to put them on a firm foundation in their Synthetic Differential

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,

(x < y y < z ) x < z (x < x )

x < y x + z < y + z x < y 0 < z x . z < y .z
0 <1 0 < x x <1
0 < x y (x = y. y )
(x = y ) (x < y y < x )
2 2
We then define by the stipulation ={x R:x =0} (x is of course x.x). This
definition enables us to formulate the two distinctive axioms of SIA. The first is
the Principle of Microaffineness:

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

(=), it suggests a special, non-classical meaning of , along constructivist lines. But

we have already seen the problems with attempting to assign constructivist meanings to
the logical symbols of SIA, and it is unclear what anything short of a systematic construc-
tivist interpretation would look like and whether it could be made coherent. (2006, 6367)

What we need, in other words, is a systematic account of the meaning of in

particular, and of the connectives in general, that is not constructivist but that
can still justify the use of intuitionistic logic.
It is at this point that the semantic theory for that logic proposed in 7.1 can
help. Or rather, what can help is an extension of that theory to a first-order
language. In the case of the language of the propositional calculus, the key
notion was that of a statements being true at various incompletely determined
possible states of affairs. In the first-order case, the key notion will be that of
a predicates being true of an entity whose properties cannot be fully deter-
mined. Ordinary real numbers obey the Law of Trichotomy:in particular, for
any real number r, we can say that r is either less than zero, identical with zero,
or greater than zero:

(r < 0) (r = 0) (0 < r ).

We cannot, though, assert this for infinitesimals. For if we were to assert

( <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

7.5 Logic and Metaphysics

I shall consider further challenges to classical logic in the next two chapters. Even
in advance of that discussion, though, we can draw some conclusions about the
character of debates between rival logical schools.
As we saw, both Dummett and the Quine of Philosophy of Logic hold that parti-
sans of classical and of intuitionistic logic are doomed to equivocate over the con-
nectives. We found Quines attempt to ground this thesis patently inadequate and
we are now in a position to identify the flaw in Dummetts more interesting argu-
ment for it (which was expounded in 1.2). The premisses of that argument, it will
be recalled, are that the meanings of the logical constants determine the manner
in which the truth of a complex sentence depends on its constituents, and 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 (Dummett 1991,
302). Achanged assessment of the validity of a form of argument, Dummett then
argues, must reflect a change in the way we take the truth-values of the premisses
and conclusions to be determined in accordance with their structure; this will
amount to a change in the sense of at least one logical constant. Our analysis,
though, shows that Dummett has overlooked another factor that contributes to
an arguments validity or invaliditynamely, the structure of the space of pos-
sible states of affairs by reference to which we unpack what it is for the truth of
some premisses to guarantee the truth of a conclusion. Achange in ones view
of that structurefor example, the change from thinking that the space must
satisfy principles (R) and (B) to thinking that it need only satisfy (R)can lead
to a change in ones assessment of the validity of a form of argument without any
change in the compositional principles that describe how the truth of a complex
statement depends on its constituents.
In rejecting Dummetts argument for the conclusion that members of rival log-
ical schools inevitably attach different senses to the connectives, Iam not claim-
ing that classical and intuitionist logicians mean precisely the same by words
such as not. Rather, our analysis supports the view of the matter implicit
in Quines earlier writings, notably, in Two Dogmas of Empiricism (Quine
1951)namely, that the notion of sameness of meaning is not precise or deter-
minate enough to sustain a claim either way. On the view recommended here,
classicists and intuitionists can and should accept the same logico-semantic
principles for conjunction, disjunction, negation, and the logicians condi-
tional, namely, (C), (D), (N), and (I). Those principles articulate their shared
understanding of the relevant connectives, at least as far as the connectives
logical properties are concerned. But this commonality is an insufficient basis
218 Five Attacks on Classical Logic

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

8.1 The Paradox of the Heap

Vague predicates have long been a source of doubt about the universal applicabil-
ity of classical logic:when applied to arguments involving such predicates, the
classical laws seem to yield absurd results. The most famous examples of this phe-
nomenon go under the names of the Paradox of the Heap or Sorites Paradox.
There are in fact several threads within the tangled skein that has come down
to us under these titles, but Ishall focus on one, chosen because it reveals most
1
starkly the nerve of the problem.
Let us suppose, then, that we have a sequence of one hundred transparent
tubes of paint, a 1,...,a 100 , with the following properties:tube a 1 is clearly red;
tube a 100 is clearly orange and hence clearly not red; but for each n, tube a n+1 is
only marginally more orange (and hence only marginally less red) than its pre-
decessor a n. Indeed, let us suppose that an observer with good eyesight, when
viewing any pair of adjacent tubes an and a n+1 together in white light but with-
out comparing them with other tubes, is unable to perceive any difference in
colour between them. That is, we suppose that any two adjacent members of the
sequence are indiscriminable in colour. In such a case, the claim that nowhere
in the sequence is a red tube immediately followed by a non-red tube seems
highly plausible. For if there were a number N for which a N were red while a N+1
were not, we would have a pair of tubes which are indiscriminable in colour
but where one member is red while the other is not:this seems to conflict with

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:

(1) n(an is red (an +1 is red)). 

In the case described, however, we also have

(2) a1 is red

and

(3) (a100 is red).

Now if we suppose

(4) a99 is red

then the rule of conjunction-introduction applied to (3)and (4)would yield

(5) a99 is red (a100 is red),

which, by -introduction, yields

(6) n(an is red (an+1 is red))

which directly contradicts (1). Given (1) and (3), then, supposition (4) stands
refuted, so by Simple Reductio we may assert

(7) (a99 is red).

By repeating this inferential sub-routine a further 98 times, we reach

(8) (a1 is red)

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

How should we react to this apparent demonstration of inconsistency? Since

the case is one in which (2)and (3)are clearly true, it seems we must take it as
showing that (1)is false. In other words, we would appear to be entitled to make a
further application of Simple Reductio and infer the negation of (1), namely,

(9) n(an is red (an+1 is red)).

Within classical logic, however, (9)is equivalentto

(10) n(an is red (an +1 is red)).

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).

8.2 Intuitionism as the Logic of Vagueness

I shall consider Freges position further in due course. First, though, let us ask
whether an alternative logical system might provide a way out of the difficulty
that the Paradox of Sharp Boundaries presents.
Crispin Wright argues that intuitionistic logic offers a way out (see especially
Wright 2001 and 2007). As noted above, the derivation showing that (1), (2), and
(3)form an inconsistent trio is intuitionistically correct; Wright accepts that deri-
vation. He further accepts that the statements (2)and (3)are true, so that the deri-
vation shows that (1)is false. But he insists that that conclusion is acceptable. We
are tempted to assert (1)because we think we know that there is no sharp cut-off
between the red members of the sequence and the non-red members. But really
we know no such thing. In order to assert (1)i.e., in order to assert that there is
no sharp cut-offwe would require an insight into the semantics of the predicate
redi.e., an understanding of the way it contributes to the determination of the
truth or falsity of statements containing itthat we do not yet possess. Of course,
we are equally unable to assert that there is any sharp cut-off point between the
red and the non-red tubes in our sequence, but lack of entitlement to assert should
not be confused with entitlement to deny. On a correct view, Wright claims, a
predicates vagueness ... does not consist in the absence of sharp cut-offs; rather,
it consists in there being nothing in our practice with the predicate that grounds
the claim that there is a sharp cut-off (Wright 2007, 439). In particular, where we
224 Five Attacks on Classical Logic

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):

(9) n(an is red (an+1 is red)).

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)

As we saw in the previous chapter, Heytings semantics is far from being

the only semantic theory that sustains intuitionistic logic. This reflection,
though, exposes the hole at the heart of Wrights account:his suggestion that
the logic of vague statements is intuitionistic is not sustained by any detailed
semantic analysis. He points to broad analogies between the intuitionists
attitudes to mathematical statements and the attitude that he recommends
taking towards vague statements. In each case, statements truth and falsity
have to be thought of as determined by our very practice, rather than by prin-
ciples which notionally underlie that practice (Wright 2007, 441). If we are
to vindicate the claim that intuitionistic logic is the right logic to use when
assessing deductions involving vague terms, however, we need something far
more detailed than these broad analogies. Dummett puts the point well:It
is not enough to show that the Sorites Paradox can be evaded by the use of
intuitionistic logic; what is needed is a theory of meaning, or at least a seman-
tics, for sentences containing vague expressions that shows why intuitionistic
logic is appropriate for them rather than any other logic (Dummett 2007a,
453). Of course, the point applies generally. Someone who wishes to defend
the claim of classical logic to codify the standards for assessing deductions
involving vague statements must also give a semantic theory that explains
why that logic is appropriate.
In this chapter, I take up Dummetts challenge. In the next section, I shall
sketch a semantic theory for vague terms which offers some hope of vindicating
intuitionism as the logic of vagueness; as Iexplain, though, the theory in question
is not really satisfactory. In 8.4, Idevelop what Itake to be a better semantic the-
ory for a class of vague predicatespolar predicates, as Icall themthat includes
such colour terms as red. Ithen show that the semantics vindicates using classi-
cal logic to assess the validity of deductions involving polar terms (8.5). Igo on to
explain how the semantic theory deals with the Sorites Paradox (8.6) and how it
avoids commitment to the Principle of Bivalence (8.7).
228 Five Attacks on Classical Logic

8.3 ASemantics for Vague Predicates that Validates

Intuitionistic Logic
Topologists have the notion of an open set. The mark of such a set is that, when-
ever an object belongs to it, there is a neighbourhood around the object, all of
whose members also belong to the set.
Suppose we have a domain which comprises all the possible coloured objects,
not merely those that actually exist. Then it is very natural to postulate that the
set of red objects in this domain will be open in a suitably defined topology. For
suppose that a possible object x is red. Then it is natural to postulate that there
is a neighbourhood of possible objects around x that are also red. An objects
perceived colour does not depend only on the wavelength of the light it reflects.
However, if the red object, x, reflects light of wavelength 700nm, then it is plausible
to claim that all the possible objects which have the same texture as x, but which
reflect light of wavelengths between 699 and 701 nm in circumstances where
human colour perception is as it actually is, will also be red. In a suitably defined
topology, then, these possible objects will constitute a neighbourhood around x,
all of whose members are also red. An object that reflects light of 700nm is a
paradigm case of redness, but the argument may be adapted to show that there is
a neighbourhood around any red object. Suppose that y is a red object that reflects
light of 620nm, and so is close in colour to some orange objects. Even assuming
that colour perception is as it actually is, it may well be that some possible objects
that reflect light of wavelengths between 619 and 621nm are not red; however, so
long as y is red, it is plausible to claim that there is some neighbourhood around
yfor example, those possible objects that are like y in texture but which reflect
light of wavelengths between 619.9999 and 620.0001nmall of whose members
are red. This smaller neighbourhood is enough to ensure that the set of red objects
in the domain is open. It is also natural to postulate that the set of possible objects
that are not red will be another open set in the same topology. For if z is a possible
object that is not red, perhaps because it reflects light of only 550nm, it is plausible
to claim that all the possible objects which have the same texture as z, but which
reflect light of wavelengths between 549 and 551nm, are also not red.
Arguments analogous to that just sketched may be advanced for many vague
predicates. Suppose that x is a tall man in the domain of all possible men. Then,
whatever precise height x has, it is plausible to claim that there is a neighbourhood
around x all of whose members are also tall men. If x is close to the borderline for
being tall, the members of the neighbourhood may have a precise height that is
very close to xs. All the same, the set of tall possible men will be open in an appro-
priately defined topology. So will the set of possible men who are not tall.
 The Challenge from Vagueness 229

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

In fact, Tarskis proof of his theorem may be adapted to establish a corre-

sponding result about monadic predicates rather than complete statements. Let
A1,...,An and B be monadic predicates, formed by applying predicate-forming
operators and, or, if...then, and not to a stock of atomic monadic predicates.
Let us consider a deductive system in which the rules of the intuitionistic prop-
ositional calculus are applied to these simple and complex monadic predicates.
The language L of this system, then, contains no singular terms or quantifiers.

Indeed, it contains no statements: A and B is a conjunction of two monadic
predicates, meaning (as it might be) red and spherical. Then, where Int signifies
deducibility by the rules of this system, we shall again have that A1,...,AnInt B
if and only if A1,...,AnTop B. Here, A1,...,AnTop B means:for every topological
space S, and every interpretation v of this language L of monadic predicates into
S, v(A1) ... v(An) S v(B). The constraint on an interpretation v is that v maps
the predicates in L into open sets in S in such a way that v(A and B)=v(A) v(B),
v(A or B)=v(A) v(B), etc.
This emended version of Tarskis result suggests that we might vindicate
Wrights conjecture that intuitionism is the logic of vagueness by developing the
idea that a vague predicate has an open extension in a suitable topology. Clearly,
the idea needs development if it is to vindicate the conjecture. By the emended
version of Tarskis completeness theorem, we know that when a monadic predi-
cate B is not intuitionistically deducible from some premiss predicates A 1,...,An,
there is some topological space S, and some interpretation v from L into S, for
which v(A1) ... v(An) S is not a subset of v(B). As things stand, though, we do
not know that this particular topological space S is one in which a vague predicate
has an open extension. To put the point another way, even if vague terms have
open extensions in some suitably defined topology, it may befor all that has
been said so farthat those extensions have additional topological features that
sustain a stronger than intuitionistic logic.
Another respect in which the initial idea needs developing concerns the rela-
tionship between the logic of monadic predicates and the logic of statements.
Even if we could show that an intuitionistic system was the right logic for deduc-
tive reasoning using vague monadic predicates, this would not vindicate Wrights
solution to the Sorites, which requires a propositional logic, not merely a logic for
monadic predicates. Despite these caveats, the topological semantics for intui-
tionistic logic seems to be the most promising area in which to seek the sort of
semantic justification for the use of that logic when reasoning with vague terms
that Dummett rightly demands.
In which topological spaces may we expect a vague predicate to have an open
extension? The intuitive arguments for openness with which Istarted involve a
 The Challenge from Vagueness 231

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

DECREASING Int (X) X

ZEROED Int ()=
TOTAL Int (U)=Uand
MONOTONE If X Y then Int (X) Int (Y).
Suppose now that x Int Int (X). That is, x belongs to some elementary set, each of
whose members belongs to Int (X). Since Int (X) is a union of elementary sets all
of which are contained in X, it follows that x belongs to some elementary set, each
of whose members belongs to X. This shows that Int Int (X) Int (X). Since Int is
DECREASING, we have that the operation isalso

IDEMPOTENT Int Int (X)=Int (X).

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

DISTRIBUTES OVER INTERSECTION Int (X Y)=Int (X) Int (Y).

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

complement of X. In these terms, X Int (X) only if X is disjoint from Cl (U X),

where U is the total domain of coloured objects. Now a coloured object y
belongs to Cl (U X) if and only if y belongs to an elementary set some of whose
members lie outside X. That is, y belongs to Cl (U X) if and only if y is indis-
criminable from an object z that is not in X. Now y is indiscriminable from
such a z just in case it is possible to argue as follows:z is not in X; y is indiscrim-
inable from z; therefore y is not in X. So y belongs to Cl (U X) just in case y is
arguably not in X. The claim that X Int (X), then, implies that no member of
X is even arguably not in X. That is, the claim implies that no red thing is even
arguably not red. This consequence of the claim is highly implausible. Because
red is vague, we cannot assert that no red thing is arguably not red. The con-
jecture, then, is to be rejected.
If, indeed, we now assume a classical metalogic, we can explain where the con-
jecture goes wrong. Given a classical metalogic, Int (X)=U Cl (U X). That is,
Int (X) comprises just those things that are not even arguably non-members of
X. We may put this by saying:Int (X) comprises just those objects that clearly
belong to X. The claim that the extension of red is open, then, identifies the red
objects with the objects that are clearly red. To put the point in metalinguistic
terms, it identifies the objects of which red is true with the objects of which the
predicate is clearly true. This identification (or something very close to it) has
attracted some philosophers. Thus, in his path-breaking essay Wangs Paradox,
Dummett writes that in connection with vague statements the only possible
meaning we could give to the word true is that of definitely true (Dummett
1975, 256). No doubt he would also say that, in connection with a vague predicate,
the only possible meaning we could give to the expression is true of is that of
is definitely true of. However, Dummett gives no argument for his claim and,
in the absence of any argument, we should stick to the common-sense view that
true and definitely true mean different things when applied to vague state-
ments. Common sense also tells us that it is possible for object to satisfy red
without satisfying clearly red; what it thereby tells us is incompatible with the
conjecture.
It is not in the least surprising that the conjecture should sustain intuitionism
as the logic of vague predicates. Given the conjecture, the extension of a vague
predicate comprises the objects that clearly satisfy it. If or is interpreted as union

and not as the interior of the complement, predicating A or not A of an object
then amounts to asserting that it is either a clear case of A or a clear case of not A.
Under this interpretation, then, Frege was correct:The law of excluded middle is
really just another form of the requirement that the concept should have a precise
boundary. When A is a vague predicate, we shall not, under this interpretation,
 The Challenge from Vagueness 235

want to assert A or not A of every object. This point, though, shows that the pre-
sent interpretation of vague predicates has changed the subject. All will agree that
we are not always entitled to say, of a coloured object, that it is either clearly red or
clearly not red. However, this point of agreement does not address the question
that we set out to answer. What we wished to know is whether each object is either
7
red or not red.

8.4 Paradigms and Poles

The idea from which we started in 8.3 was that the objects which satisfy a vague
predicate may be expected to form an open set in a suitable topology. The par-
ticular development of that idea in the previous section did not work, but the idea
itself remains plausible. In this section and the next, Ishall explore a different way
of developing it. The development will vindicate classical logic as the appropriate
logic for assessing deductions involving a large class of vague terms, including the
colour predicate used in our original Sorites argument.
Any semantic theory must say how the various expressions of a language
combine to produce meaningful statements; that will surely mean assigning sat-
isfaction conditions to all its predicates, and to that degree treating them alike.
However, that likeness is consistent with substantial differences in how those
satisfaction conditions are specified, and it is pretty clear that natural languages
contain predicates of at least two radically different sorts. Classical semantics
associates each predicate with a sharp boundary between the objects that satisfy
8
it and the objects that do not. For some vernacular predicates, this theory is cor-
rect. For example, the word prime (in its arithmetical sense) is associated with
a rule whose application effects a sharp division between the natural numbers

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

On this view, any predicate ought to be associated with a precise boundary

between the objects that satisfy it and those that do not. In the case of bald,
though, we speakers of English have failed to effect a unique such association;
hence the deficiency in the predicates meaning. Sainsbury proposes a radically
different and prima facie more plausible picture. In saying Fred is bald, it is not
that we try to effect a precise delineation and fail. Rather, we succeed at doing
something quite different:we compare Fred with paradigms of baldness.
This picture needs a great deal of filling out. As Sainsbury remarked back in
1990, the logic of vagueness, characterized as boundarylessness, remains to be
described (1990, 262). Unless Ihave missed something, that is still the case today.
The first task must be to fill that gap.
Let us begin by considering a simple language whose seven atomic predicates
are red, yellow, orange, green, blue, indigo, and violet. The language also
contains a stock of variablesx, y, and so onwhich are understood to range
over tubes of paint of every possible colour. The open sentence that results from
attaching a predicate to a variable is understood to say that the paint in the tube
that has been assigned to the variable is of the colour signified by the predicate.
Thus x is red means that the paint in the tube assigned to the variable x is red. It
is assumed that the paint inside a single tube does not vary in colour.
The idea we are trying to develop is that the meaning of these colour predicates
is to be understood in relation to the poles or paradigms provided by a spectrum.
So let us suppose that their meaning is to be understood in relation to a particu-
lar printed spectrum, S, which is found in a book of colour paradigms. We also
postulate the existence of a normally sighted viewer who, while looking at the
spectrum S, finds that colours really do stand out as clearly different. That is to
say, he sees in S bands of colour. This perceived banding has a complex structure.
At the crudest level, he discerns seven bands within S, which he labels r, o, y, g,
b, i, v from left to right across S. But if he focuses on the leftmost band, r, he will
discern bands within it, bands that can serve as paradigms of different varieties
of red, such as scarlet or crimson. We further suppose, however, that when our
viewer looks again at the whole of S he continues to discern the seven main bands
he had already labelled. So the higher-resolution banding does not undermine or
conflict with the lower-resolution banding.
The idea is that the meaning of red lies in its association with the pole or para-
digm of redness provided by band r. Asympathetic development of this idea will
surely take the whole band r as the paradigm, rather than any specific shade within
it. We can classify objects by colour without agreeing whether such-and-such a
shade is, so to say, one redder than which cannot be conceived. In taking the cru-
cial association to be with a perceived paradigm, the idea respects Lockes famous
 The Challenge from Vagueness 239

contention that red signifies a secondary quality:an explanation of the words

sense must advert to our experience of things as coloured, and cannot confine
itself to whatever might be the physical bases of that experience. Modern inves-
tigations of the psychology of colour perception have borne Locke out on this
point. One recent treatise warns us thatthe
lack of correspondence between psychological representation of spectral composition
and spectral composition itself is a source of confusion and misunderstanding in scien-
tific discussions of colour. Scientists persist in referring to the physical characteristics of
the stimulus and to the tuning characteristics of the transducers (the cones [i.e., the reti-
nal cells that are stimulated when light falls on them]) as if psychological [recte:ordinary]
colour terms like red, green, and blue had some straightforward translation into physical
10
reality, when in fact they do not. (Gallistel 1990, 519)

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

x sats A if and onlyif

(Sat)
x is perceived to be closer in colour to the pole q than to any other
colour pole

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

not being perceived to be closer in colour to r than to an intermediate red-orange

pole. (Compare Sainsburys remarks about young, old, and middle-aged.)
Since an object will satisfy red if and only if it is red, indeterminacy in
whether x is red implies indeterminacy in whether x satisfies red. So the
satisfaction of a vague predicate will itself be vague. But so is the proposed
condition for x to satisfy red:even if the relation of perceived closeness in
colour were perfectly precise (which Idoubt), the property of being closer in
perceived colour to r than to any other pole will be indeterminate because
the paradigm r is a colour band which lacks precise boundaries. So vague-
ness provides no objection to our proposed principle about the satisfaction of
colour terms.
All the same, the vagueness of the satisfaction condition laid down in (Sat)
calls for a caveat. In constructing a semantic theory for vague predicates, it is
natural to apply the familiar apparatus and speak as though a predicate has an
extension:this will be the set of those objects that satisfy the predicate. If we
speak of the extension of red, though, we must allow that there are sets, the
condition for membership of which is vague. Philosophers who accept each of
two theses will regard this commitment as unsustainable. The first thesis says
that there can be no vagueness in respect of an objects identity; the second, that
the identity of a set consists in its membership. The matter deserves extended
discussion, but neither thesis seems to me to be well supported. The proofs that
11
purport to show that identity cannot be vague beg the question, so we might
12
take the extension of a boundaryless predicate to be a vague object. But even
if we do not wish to go so far, there remain strong grounds for rejecting the sec-
ond thesis, whereby a set consists of its members. For one thing, the empty set is
a perpetual embarrassment for this latter thesis. Partly for that reason, indeed,
Frege (1895) suggested an alternative conception of a sets identity, whereby a
13
set is essentially the extension of a concept (= the reference of a predicate).
14
This conceptionthe so-called logical conception of a set can survive the

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)

Quine goes on to resist the application of this model to colourterms:

Thus take red. Let us grant that a central shade of red can be picked as norm. The trouble
is that the paradigm cases, objects in just that shade of red, can come in all sorts of shapes,
weights, sizes, and smells. Mere degree of overall similarity to any one such paradigm
case will afford little evidence of degree of redness, since it will depend also on shape,
weight, and the rest. If our assumed relation of comparative similarity were just compara-
tive chromatic similarity, then our paradigm-and-foil definition of kind would indeed
accommodate redkind. What the definition will not do is distil purely chromatic kinds
from mixed similarity. (Quine 1969, 120)

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.

8.5 ASemantics for Polar Predicates that

Validates Classical Logic
How might this be done?
As a first step towards this goal, let us say that a colour pole p is maximally
close (in colour) to a coloured object x if there is no colour pole distinct from p to
which x is perceived to be closer in colour than it is to p. (This definition takes
 The Challenge from Vagueness 243

as understood the three-place comparative similarity relation that was used

in stating (Sat).) For any coloured object, there will be at least one pole that is
maximally close to it. Under the definition, though, when an object is perceived
to be just as close in colour to one colour pole as to another, both of those poles
will be maximally close to it. Thus both the red pole, r, and the orange pole, o,
are maximally close to a borderline red-orange object. Maximal closeness is a
vague relation:it will often be indeterminate which pole or poles are maximally
close to a given object. All the same, we can express in these terms our condi-
tion (Sat) for an object to satisfy a simple colour term, A, that is associated with
a colour pole q:

(Sat) x sats A if and only if,

for any colour pole p, p is maximally close to x iff p=q.

Since maximal closeness is a vague relation, this reformulation of (Sat) continues

to respect the principle that the satisfaction of colour predicates is itself vague.
(Sat) is highly plausible for atomic colour predicates but it does not apply to
complex predicates, notably disjunctions. There is no single colour pole which
is maximally close to each object that is either red or orange, for instance. So we
have to ask how (Sat) might be generalized to apply to complex colour predicates,
especially disjunctions. We may be guided here by the expectation that the exten-
sion of any vague predicate (not just an atomic predicate) will be an open set in a
suitably constructed topology.
With that thought in mind, let us define an operation Int on a set X of coloured
objects as follows:

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:

(Con) |A and B| = |A| |B|.

18
When two sets are regular open, so is their intersection, so (Con) respects the
principle that the extension of any predicate should be regular open. As for
negation, the idea we have been developing is that an object satisfies the nega-
tion of a predicate just when it has some contrary status in the relevant system
of classification. Thus, where not expresses predicate negation, it is natural to
postulate

(Neg ) | not A| = |A| = Int (U |A|).

19
Whenever the set X is open, X =X , so |not A| = |A| = |A| = |not A|=
Int Cl (|not A|). Thus (Neg) also respects the principle that a predicates extension
should be regular open.
What, though, about disjunction? The union of two regular open sets is not
always regular open:regular open sets have no cracks in them, but when two
such sets are separated by a crack, their union will not be regular open. Since
we expect colour predicates to have regular open extensions, we cannot, con-

sequently, take the extension of A or B to be |A| |B|. Aform of argument
that has already been deployed in this book, however, shows what the correct
semantic axiom for disjunction must be. Let us say that a predicate A entails
a predicate B when it is logically necessary that the extension of B includes

that of A. Since A entails A or B , the extension of A or B must include that
of A, i.e., |A| |A or B|. Similarly, |B| |A or B| so that |A| |B| |A or B|.

Furthermore, A or B entails C whenever A entails C and B entails C, so A or B

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:

(Dis) |A or B| = Int Cl (|A| |B|).

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

8.6The Sorites Revisited

With this last point in mind, let us return to the paradoxical argument of 8.1.
Our semantic analysis seems to validate the laws of classical logic in reasoning
with vague concepts. In the course of that argument, though, the application of
classically valid rules took us from apparently incontestable premissesviz., a1 is
red and (a100 is red)to a conclusionviz.,

(10) n (an is red (an +1 is red))

that seems to assert that the predicate red has a sharp boundary. Now (10) may be
recast as a long disjunction,

(10) (a1 is red (a2 is red)) (a99 is red (a100 is red))

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

quantified formulae such as Wrights (10) with disjunctions such as (10). In

what follows, Ishall write as though the deduction of 8.1 has been recast in just
this way, so that the rule of -introduction, not -introduction, validates the
step from (5) to (6).
How, then, might we extend our semantics for and etc. so as to yield a seman-
tics for etc.? The obvious first step is to consider the colour-classification of
ordered pairs of objects. Thus if the possible colours of the individual objects
are red and orange, we shall have four poles by reference to which we can clas-
sify ordered pairs:<r, r>, <r, o>, <o, r>, and <o, o>. The notion of maximal close-
ness may be extended to apply to these new, binary poles; we may then say that
an ordered pair <a, b> satisfies the predicate first red-second red (red-red for
short) if <r, r> is the only pole that is maximally close to <a, b>, and similarly
for the other classifications. As in the case of the colour classification of single
objects, the relation of maximal closeness generates an interior operationcall
it Int2on sets of ordered pairs of possible objects. Again as before, the fact
that <r, r>, <r, o>, <o, r>, and <o, o> form a system of contrary classifications of
ordered pairs means the four atomic predicates of ordered pairs (viz., red-red,
red-orange, orange-red, and orange-orange) have extensions that are regular
open sets in the topology defined by Int2 . The operations of intersection, interior
of the closure of the union, and exterior are well defined on sets of ordered pairs
of possible objects, and when these operations are applied to regular open sets of
pairs they yield further regular open sets of pairs. So if we use (Con), (Dis), and
(Neg) to specify the senses of and, or, and not, as they apply to predicates of
ordered pairs, the resulting complex predicates will also have regular open exten-
sions. As before, then, the resulting logic will be classical, so long as we are con-
sidering inferences in relation to a fixed set of poles. If that condition is met, then,
the Sorites step from ai is red and it is not the case that ai is red while ai+1 is not
red to ai+1 is red will be validated. The step will be analysed as follows. From the
premiss that an object ai is red, we infer that the pair <ai, ai+1> is either red-red or
red-orange. From the premiss that it is not the case that ai is red while ai+1 is not
red, we infer that the pair <ai, ai+1> is not red-orange. We then infer that the pair
<ai, ai+1> is red-red, from which it follows that ai+1 is red.
The method of argument set forth in the previous paragraph generalizes. In
analysing an inference involving the colours of three objects, we consider the
colour-classifications of ordered triples of objects. So, if the possible colours of the
individual objects are again red and orange, we shall have eight poles by reference
to which we can classify ordered triples:<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>. We may say that an ordered triple <a, b, c> satisfies
252 Five Attacks on Classical Logic

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

(1 is red (2 is red)) (99 is red (100 is red))

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.

8.7 Vagueness and Distribution

So far, we have been considering deductions all of whose component predicates
relate to a single system of poleshomogeneous deductions, as we may call
them. Many arguments, though, are heterogeneous in that they involve predi-
cates whose senses are to be explicated by reference to diverse such systems. From
Tube a 49 is red and Tube a50 is hard, we infer Tube a 49 is red and tube a50 is hard.
From Either Fred is poor or Tom is bald and Fred is not poor, we infer Tom is
bald. The problem in validating these inferences is not that it is difficult to con-
struct polar semantic theories for the predicates they contain. Diana Raffmans
paper Borderline Cases and Bivalence (2005) convincingly shows, for example,
how poles help elucidate the senses of predicates of wealth; those poles generate a
topological account of the satisfaction conditions of rich, poor, etc. in just the
way indicated in 8.5, and cognate accounts may be given for hard, bald, and
so forth. The problem, rather, is this. As things stand, we have one topological
space for colour terms, another for predicates of hardness, a third for predicates
of wealth, and so forth. On the view being propounded, the extension of red is a
regular open set in the first of these spaces, that of hard is a regular open set in
the second space, that of poor is a regular open set in the third, and so forth. As
yet, though, we have no way of connecting spaces in order to get truth-conditions
for such statements as Tube a 49 is red and tube a50 is hard and Either Fred is poor
or Tom is bald.
In fact, a solution to this problem is implicit in our analysis of the Sorites. In
dealing with that paradox, we had to move from predicates that classify individ-
ual objects by colour to predicates that effect classifications of N-tuples of objects;
the method used in that transition may be re-applied to generate topological
256 Five Attacks on Classical Logic

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:

(1) Say (u, a50 is red a50 is orange) premiss

(2) Say (v, a50 is red) premiss
(3) Say (w, a50 is orange) premiss
We also suppose
(4) True (u) premiss
I shall say a little about the principles that regulate the concept of truth in
Chapter10, but it is initially plausible that among those principles are rules of
inference that entitle us to move from True (u) Say (u, P) to P (call this rule
R1) and from Say (u, P) P to True (u) (call this R2). Given these rules, one
may construct the following derivation:
(5) Say (u, a50 is red a50 is orange) (1), (4), -introduction
True (u)
(6) a50 is red a50 is orange (5), R1
(7) a50 is red assumption
(8) Say (v, a50 is red) a50 is red (2), (7), -introduction
258 Five Attacks on Classical Logic

(9) True (v) (8), R2

(10) True (v) True (w) (9), -introduction
(11) a50 is orange assumption
(12) Say (w, a50 is orange) a50 is orange (3), (11), -introduction
(13) True (w) (12), R2
(14) True (v) True (w) (13), -introduction
(15) True (v) True (w) (5), (10), (14) -elimination,
with discharge of assumptions
(7)and(11)

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

clear truth are we to assign to other statements if we decide to count A as definitely

true? Does statement B, This curtain is orange, become definitely false, because
in changing the meaning of red we retain the principle that red and orange are
contrary terms? Or is B also definitely true, because we now count any borderline
instance of a colour as a clear case of that colour? What, if anything, follows about
statements concerning a borderline blue-green curtain? Edgington has not, Ifear,
told us as much as we need to be told in order to be confident that conditional ver-
ity is a coherently applicable notion.
The prospects of the present theory are, Ithink, brighter. More work is needed,
23
notably in extending it to deal with higher-order vagueness, but Ihope to have
indicated a promising direction for further exploration. The Father of Philosophy
is said to have had engraved above the
door to his Academy. Geometric methodsmore particularly, topological
oneshave been an under-exploited resource in theorizing about vagueness.

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)

Accordingly, Lear recommends a semantics for the language of set theory in

which reference and truth are relativized to particular stages of our comprehen-
1
sion of that theory. Such a semantics, he argues, does not validate all the classical
logical laws, but only the intuitionistic laws.
1
Lear acknowledges the influence of Leslie Tharp (see his 1971)and William Powell (1976). The
improved understandings of set that Lear has in mind are those which would validate so-called
large cardinal axioms; that is why the improvements lead to the postulation of more sets.
264 Five Attacks on Classical Logic

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.

9.1 What is Mathematics About?

It helps to begin by stepping back and saying something about the philosophical
context in which these questions are to be addressed. Here an excellent point of
departure is Dummetts essay What is Mathematics About?
In that paper, Dummett was concerned with the differences between math-
ematical objectssuch as natural and real numbers, or setsand physical
objects such as stars. Dummett was no nominalist. He accepted that the state-
ment There are two prime numbers between 15 and 20 is strictly and literally

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)

An entire philosophy of mathematics is latent in this passage; articulating it

properly would take a book to itself. It is clear in outline, though, that Dummett
aims to combine what one might call internal Platonism with transcendental
anti-realism. The statement There are numbers is strictly and literally true, just
as the statement There are stars is strictly and literally true. But the former state-
ment owes its truth, as the latter does not, to the fact that thinkers have deter-
mined or circumscribed a domain of natural numbers. External realitythat is,
the part of reality entirely independent of human cognitiondoes not supply the
members of fundamental mathematical domains.
Dummett takes it to be obvious that mathematics does not describe a realm
wholly independent of human thought. No one, he opines, could be so much of a
realist as to hold that it did (Dummett 1993c, 75). Many mathematicians and philoso-
phers have in fact taken the position that Dummett rules out ab initio but, for present
purposes, we may take the incorrectness of this extreme realist position as a premiss.
For the question Iwish to consider is whether the alternative, anti-realist view of the
nature of mathematics that Dummett proposes constitutes a threat either to classical
semantics or to classical logic.
According to that alternative view, the fundamental mathematical domains are
posits of our mathematical conceptions. Thus the domain of natural numbers just
266 Five Attacks on Classical Logic

is a collection of objects which, collectively, satisfy our conception of a simply infi-

nite sequence; similarly, the real numbers just are objects which satisfy our concep-
tion of the continuum; and the sets just are objects which, collectively, satisfy our
conception of the set-theoretic hierarchy. This view is sometimes criticized on the
ground that it entails that the number two (for example) did not exist before there
were thinkers capable of doing elementary arithmetic, but this criticism is wholly
misplaced. Our conception of a simply infinite sequence makes no reference to times
(although particular thinkers, or groups of thinkers, can grasp, or fail to grasp, the
conception at particular times). On the view we are considering, then, on which nat-
ural numbers are posited to fit that conception, no sense has yet been attached to
claims like The number two exists now but did not exist in 5000 BC. Idoubt if sense
can be made of such claims but, even if it can, they are not consequences of the view
under consideration.
On that view, vindicating the existence of a fundamental mathematical domain
will be a matter of articulating our conception of it to the point where we are con-
fident that we really do grasp the domain in question. Whether a given thinker,
or group of thinkers, succeeds in determining or circumscribing a putative math-
ematical domain is not an all-or-nothing matter. Certainly, one thinker might have
a better grasp than another of a given domain. In 1993, Dummett doubted if anyone
4
possesses a fully determinate conception even of the domain of natural numbers.
Ithink he was too pessimistic on that score, but correct to hold that no one has a
determinate conception of the domain of sets. So it will help to contrast the two cases.
Mathematicians have had an implicit and inchoate conception of the domain
of natural numbers for at least two and a half thousand years:some such con-
ception animates the treatment of arithmetic in Books 79 of Euclids Elements.
Only in the late nineteenth century, though, was this captured in an explicit

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

theory. The achievement of making the conception explicit was primarily

Richard Dedekinds in Was Sind und Was Sollen die Zahlen? (Dedekind 1888).
Not only did Dedekind state what we now know as the Peano-Dedekind axioms,
he also proved that his axiomatization was categorical. That is, he proved that any
two models of his theory of number are isomorphic. It is this result that convinces
most of us that Dedekind succeeded in articulating our implicit conception of the
natural numbers:for the categoricity theorem entitles us to speak of the domain
(in the singular, up to isomorphism) of natural numbers.
The logic of Dedekinds theory must be stronger than first-order: the
Lwenheim-Skolem theorems preclude any categorical first-order axiomati-
zation of arithmetic. Dedekinds own reasoning is naturally formalized in full
5
second-order logic. In the present dialectical context, it cannot be assumed
that we understand the notion of an arbitrary subset of an infinite set, or of an
arbitrary sub-concept of a concept under which infinitely many objects fall. As
remarked in Chapter7, though, a categorical axiomatization of arithmetic may be
attained using a weaker logic. Thus, working in Myhills ancestral logic (see 7.3),
one gets such an axiomatization by adding to the first-order Peano-Dedekind
axioms a single postulate stating that any natural number stands to zero in the
ancestral of the relation of immediate succession (where the latter relation may
be defined in the usual way). Ancestral logic requires us to make sense of the gen-
eral distinction between finite and infinite collections, but many philosophers
will understand this distinction who baulk at the notion of an arbitrary subset of
an infinite set. In particular, intuitionists will surely understand it. They inveigh
against classical mathematicians for illegitimately treating infinite collections
as though they were large finite ones. That complaint evidently presupposes an
understanding of the distinction between the finite and the infinite.
The logic of Dedekinds theory of number is also classical, but in the case of
arithmetic there is no evident reason why it should not be. Let T be any categori-
cal theory, formulated in a language L. Then the same closed sentences of L hold
in all the models of T. That is, wehave

(1)For any models of M and N of T, and any closed sentence in L,

M if and only if N .

The proof of (1)is in essence straightforward. Since T is categorical, there is an

isomorphism f from M to N. One then shows by induction on complexity that,

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 .

Now (1)and (2)jointlyentail

(3)Either M for any model M of T, or M for any model M of T.

An abbreviated way of writing (3)is

(4)Either T or T .

Suppose now that we deem a closed sentence in the language of arithmetic to be

true if it is a semantic consequence of a categorical version of the Peano-Dedekind
axioms, and false if its negation is a consequence of such a theory. Such an account
of truth and falsity for arithmetical statements is attractive in that it captures
what seems to be right about structuralism in arithmetic, and it enables us to
re-write (4)as

(5) Either is true or is false.

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.

9.2 Attempts to Attain Categoricity

So far, all we have seen is the failure of one attempt to explain why the classical
logical laws are the right ones to use when reasoning about sets. We have yet to
see a case in which they are inapplicable. Ishall soon explain why Tait thinks
that such cases exist, but our discussion already raises an important question.
2 2
Zermelo showed that ZF was not categorical and his argument extends to ZFC .
2
But might there not be some extension of ZF that is fully categorical? If so, could
we not justify the use of classical logic in set theory by taking this stronger theory
to be the correct articulation of our implicit conception of the domain of sets?
Some philosophers will deem this question to be footling. Even if you were
to find a categorical second-order axiomatization of set theory, they will say,
it would do nothing to show that we have a determinate grasp of the domain
of sets. For second-order logic already presupposes set theory. More exactly, if
a second-order set theory is to be categorical, its second-order quantifiers must
be interpreted as ranging over arbitrary subsets of the domain. So the use of
second-order logic presupposes the very grasp that you are trying to vindicate.
There are two elements to this objection, one easily answered, the other more
challenging.

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 the present dialectical context, we certainly need an explanation of the

second-order quantifiers that does not presuppose set theory: we shall have
made no progress on the problem of vindicating our grasp of the domain of
sets if a second-order formula such as Xy(Xy y) is read as saying There
is a set whose members are precisely the objects that are . This requirement,
however, is easily met:there are now a number of interpretations under which
second-order logic emerges as something other than set theory in sheeps
clothing (Quine 1986, 66). Most pertinently for present purposes, there is Taits
own type-theoretic interpretation (see Tait 1998, 4745 and, for more details,
Tait 1994). Mathematical objects, he holds, are always objects of some type,
where a type is given by effective introduction and elimination rules and in
this way is proposition-like (Tait 1998, 475). Tait postulates that given a type
A, there is always its power type (A), which we may take to be the proposi-
tion A 2, 2 being the two-element type {, } (Tait 1998, 475; see also Tait
1994, 113). Where the first-order variables range over the members of type A, the
monadic second-order variables will range over the members of (A). Taits
interpretation is attractively Fregean. Suppose A is the type of Fregean objects
or Gegenstnde. Then the monadic second-order variables will be understood
to range over functions from objects to truth-values, just as in the system of
Grundgesetze. On this interpretation of the second-order language, Tait is
entirely right to claim that the type-free notion of set which is formalized in
ZF is distinct from the sort of things over which second-order variables range
(Tait 1998, 474). Sets in ZF are Fregean objects, whereas members of (A) will
always be functionsthat is, unsaturated entities in Freges ontology. On the
proposed interpretation, then, it is simply false to say that a second-order quan-
tifier always quantifies over sets.
A related worry, however, is less easily answered. Zermelos quasi-categoricity
proof only goes through in full second-order logic. The kernel of the proof is the
First Isomorphism Theorem (Zermelo 1930, 12289), which asserts that any two
2
models M and N of ZF with the same boundary number and the same number
of Urelemente are isomorphic. In showing that an isomorphism exists between
any such M and N, Zermelo appeals to the Axiom of Replacement, which he states
as follows:If one replaces the elements x of a set m in a unique way by any arbi-
trary elements x of the domain, then the domain also contains a set m which has
all the objects x as elements (Zermelo 1930, 1220). It is indeed vital that an arbi-
trary replacement of elements in one set should yield a further set:the proof of the
First Isomorphism Theorem would fail if, for example, there had to be a defin-
able relation between any x and its replacement x, for there is no guarantee that
2
such a relation will obtain between the elements of any two models of ZF . From a
272 Five Attacks on Classical Logic

logical point view, then, Replacement cannot be formalized as a first-order axiom

schema. Rather, it has to be formalizedas

f (dom ( f ) is a set ran ( f ) is a set)

or, more explicitly,as

8
f xy z (z y w(w x z = fw )),

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

McGees result is striking, and one commentator has welcomed it as opening

the door to a structuralist account of set-theoretic truth on which every sen-
tence of [the language of] pure set theory is assigned a determinate truth value
(Uzquiano 2002, 181). But have we reason to accept McGees axiom? So far from
its being an evident truth, Ithink it is at best highly doubtful.
A notion that looms large in contemporary mathematics is that of a cat-
egory: mathematicians speak of the categories of all sets, of all topological
spaces, of all manifolds, etc. Categories are not sets:there is no set of all sets,
nor a set of all topological spaces; that is to say, a category is an Urelement.
Accordingly, McGees axiomthat there is a set of all Urelementeentails
that there is a set of all mathematical categories.
To see why this is commitment is problematical, let us focus on the cat-
egory, Set, of all sets. There is, of course, a singleton set {Set} that comprises
this category alone; but there are also, Inow argue, closely related categories
that are too numerous to form a set. The domain of the category Set com-
prises all the sets, and its arrows comprise all functions between sets. Now
an object 0 is called an initial element of any category if, for every object a in
the category, there is one and only one arrow from 0 to a in the category (see
Goldblatt 1984, 43). In Set, the empty set is the only initial element. Where
b is any set, let us define the b-variant of the category Set as follows:in the
b-variant of Set, b is the unique initial element, while has the pattern of
arrows that is characteristic of b in Set; the patterns of arrows pertaining to
all other sets are left undisturbed. In other words, in the b-variant of Set,
and b change places. For any set b, the b-variant of Set is a category; indeed, it
is a category isomorphic to Set. Although isomorphic, however, the categories
are distinct. In Set, is the unique initial element; in the b-variant of Set, b is
the unique initial element.
Because each b-variant of Set is a category, any set of all the categories would
have to contain all the b-variants of Set. But there are just as many b-variants of
Set as there are sets, so a set of all categories would be at least as large as a set of all
sets. On the conception whereby sets are subject to a limitation of size, it follows
that there can be no set of all b-variants of Set, and hence no set of all categories.
As we have seen, though, McGees axiom entails that there is such a set. So we
have strong reason to suppose that his axiom is false. Indeed, we have reason to
reject the weaker, but still potent, postulate that there are not more Urelemente
than sets.
A defender of McG may reply that categories can easily be represented
as sets, for example, in the manner proposed by Saunders Mac Lane in his
Categories for the Working Mathematician (Mac Lane 1994). Ishould respond
 On the Use of Classical Logic in Set Theory 275

that it is one thing to represent categories as sets, another to maintain that

they are sets:indeed, a familiar form of argument in the philosophy of math-
ematics supports the claim that they are not sets. If a category is a set, there
has to be an answer to the question, which specific set it is. There are, though,
many different ways of representing categories as sets, and no fact of the
matter as to which is correct. For the reasons expounded by Paul Benacerraf
in What Numbers Could Not Be (Benacerraf 1965), then, categories cannot
10
be sets.
Mention of Benacerraf s famous paper points the way, indeed, to other
doubtful implications of McG that are perhaps less recherch than state-
ments about categories. Christopher Menzel has used Benacerraf s trope to
argue powerfully that ordinals are not sets. An ordinal, Menzel proposes, is
something that isomorphic well-ordered sets have in common; that is, it is
a property of well-ordered sets (Menzel 1986, 43). An ordinal can be repre-
sented as a set, but it is not itself a set:just as there are distinct but equally good
set-theoretic representations of the natural numbers, so there are distinct but
11
equally good set-theoretic representations of the ordinals. An ordinal, then,
is an Urelement, so McG (together with Separation) implies that there is a set
On of all ordinals; the Burali-Forti Paradox now looms (see Menzel 1986, 38).
Certainly, that Paradox does not need the assumption that ordinals are sets.
All we need is that On is well-ordered by the relation < of being strictly less
than:it follows immediately from this theorem that On possesses an ordinal,
. Now any ordinal is the order-type of the set of ordinals strictly less than
itself. So is also the order-type of the set A of all ordinals that are strictly less
than . Since A and On share an order-type, it follows by definition that they
are isomorphic. It is easy to show, though, that no well-ordered set is isomor-
phic to any of its proper initial segments, so that A and On must be identical.
Since On is the set of all ordinals, On, whence A. But then, by the
definition of A, it follows that < . Menzel (2014) has explored ways of avoid-
ing outright contradiction here by restricting some of the other axiom of ZF.
The fact remains, though, that McGs having this implication renders it highly
doubtful. Nice as having a categorical characterization of the universe of sets
might seem to be, we should not accept McGees theory.

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

9.3The Iterative Conception of Sets,

and the Threat to Classical Logic
Let us return to the main line of argument. Our analysis has shown that some
set-theoretic statements are not bivalent. In the light of this, classical semantics will
not validate classical logic. This does not pose any immediate threat to the use of clas-
sical logic in set theory:the two-valued semantics is far from being the only semantic
theory that sustains classical logic. All the same, it poses a problem for anyone who
uses classical logic in set theory. We surely need some account of how the familiar
connectives and quantifiers contribute to the meanings of set-theoretic statements
in which they appear. It is now an open question what that account might be, and a
further open question whether it validates all the rules of classical logic.
A striking feature of Zermelos discussion is his rejection of the thesis that there
is a single, well-defined universe of sets which constitutes a model for the axioms
of set theory. For Zermelo, a model is a set, and while each V, for ordinal , is a
set in every V with strictly less than , there is no super-model comprising the
entire set-theoretic universe. Tait follows Zermelo here. The view, he writes, that
set theory is the theory of the universe of all (pure) sets in the way that Euclidean
geometry is the theory of Euclidean space ... is certainly a commonly expressed
view. But in as much as it implies that the universe of sets is a well-defined total-
ity, it seems to be a difficult view to defend (Tait 1998, 4778). Tait reads Zermelo
as an early proponent of what is now called the iterative conception of sets. On
this view, sets are the objects obtained in domains that arise from iterating the
[power set] operation as far as possible, obtaining bigger and bigger domains
(Tait 1998, 475). Whether or not it animated Zermelos work, the iterative con-
ception remains popular. However, as Tait remarks, it is hard to see how it can
circumscribe any well-defined totality of sets, for at its heart lies a deep source
of indeterminacy:on the conception we are discussing, all sets are obtained by
iterating the power set operation as far as possible. But this is not a mathemati-
cal characterization; rather, it is an informal idea which leads us to successively
stronger precise principles of iteration (Tait 1998, 478). In the present context, as
far as possible means as far as mathematically possible, but we shall need axioms
2
beyond those of ZF to say which iterations of the power set operation are math-
ematically possible.
Tait argues that the iterative conception resists any attempt to determine how
many iterations of the power set operation are mathematically possible. He begins
by quoting Gdel:
The axioms of set theory by no means form a system closed in itself, but, quite on the con-
trary, the very concept of set on which they are based [sc., the iterative conception] suggests
 On the Use of Classical Logic in Set Theory 277

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:

[A]mathematical characterization [of as far as possible] is precluded on the iterative con-

ception by the fact that any mathematically definable totality P of such principles [i.e., of
principles of iteration] would lead to a new one, namely closure under all the principles in P.
Moreover, it is not at all clear that, as we try to develop the idea of iteration as far as possible,
there is only one direction that this would take us. It is conceivable that there is some devel-
opment of this idea that leads to the construction, for example, of measurable cardinals; but
that there would also be grounds for rejecting the construction. So it does not seem possible
to speak of the universe of sets as a well-defined extension, even in the sense of speaking of
those sets which we construct or might have constructed on the basis of the iterative con-
ception. Even if one believed that, by the nature of the human psyche, there is only one way
in which the idea could be developed, this single path would still be incapable of a math-
ematical characterization:we could then only say of it that it is the path we are destined to
takean anthropological characterization, not a mathematical one. (Tait 1998, 478)

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:

K1.s A B if and only if s A and s B

K2. s A B if and only if either s A or s B
K3.s A if and only if for every t such that s t, it is not the case that t A
K4. s A B if and only if for every t such that s t, if t A then t B
280 Five Attacks on Classical Logic

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

s xA(x) if and only if s A(cz) for every z in D(s).

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?

9.4Attempts to Justify Classical Logic

under the Iterative Conception
One way of vindicating a logic stronger than intuitionism would be to show that
the relation of extending on the space of axiom systems has properties over and
above those of a pre-order. In an interesting article, Allen Hazen (1982) argued
282 Five Attacks on Classical Logic

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,

where C signifies deducibility in first-order classical logic. With * defined as

above, it would follow that

A 1* ,..., An* C CH,

and hence

C (A 1* ... An*) CH*).

By the property that relates classical logic to bastard intuitionistic logic, it

follows that

I ((A 1* ... An*) CH*),

where I signifies deducibility in first-order intuitionistic logic. By a familiar

theorem of intuitionistic propositional logic, it follows in turn that

I (A 1* ... An*) CH*

284 Five Attacks on Classical Logic

But that, of course, contradicts the existence of an intuitionistic ZF model in

which CH* holds. This contradiction reduces to absurdity the initial assump-
tion that CH is classically deducible from the axioms of first-order ZF, so we reach
Cohens independence result.
This result has potential relevance for our inquiry. Suppose we were to con-
vince ourselves that the operative notion of truth for set-theoretic statements is
what Fine (1975, 134)calls anticipatory truth:a formula is true if and only if there
is always going to be a way of forcing it under the bastard intuitionistic semantics.
Then, although we would not have Bivalence, we would have classical logic:all the
rules of classical logic would preserve anticipatory truth.
For myself, though, Icannot think of any persuasive reason why we should
accept this account of truth for set-theoretic statements. If a statement has the
property of anticipatory truth, then it is always going to be open to someone to
defend his assertion of it by adopting an axiom system that forces it. There may,
though, be no reason to adopt such an axiom system, so a statement may pos-
sess anticipatory truth even though it is in no way wrong to reject it. For this
reason, anticipatory truth falls short of genuine truth. It amounts to always
being defensible, or never being refutable, rather than to being true. Moreover,
the bastardy remains a problem. The suggested account will have to invoke
Dummetts distinction between a statements free-standing (or assertoric)
content and its ingredient sense, i.e., the meaning it contributes to a statement
13
of which it is a proper part: K1 to K6 recursively specify the contributions that
the connectives and first-order quantifiers make to ingredient sense, but to
get the assertoric content of a complete statement we shall need to preface the
resulting sense with a double negation. All the same, specifications of ingredi-
ent sense ought to capture our ordinary understanding of the connectives and
quantifiers, and while the legitimate K5 passes this test, the bastardized K5
does not.
Let us take stock. The project of this chapter has been to investigate whether
we can justify the use of classical logic within set theoryparticularly, its use
in evaluating deductions involving unrestricted quantification over sets. More
exactly, our question is whether that use of classical logic can be justified if

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

9.5Classical Logic Justified via

a Negative Translation
With these caveats in mind, let us resume the quest for a justification for the use
of classical logic in set theory. In previous chapters, we noted the importance of
characterizing accurately the sense of signs for disjunction. It helps to scrutinize
the treatment of in the semantic theory of 9.3.
According to K2, the semantic principle for , an axiom system forces a dis-
junction if and only if it forces one of the disjuncts. Given what forcing means in

this context, K2 says that a system of set-theoretic axioms entails A B if and
only if it either entails A or entails B. Similarly K6, the semantic principle for the

existential quantifier, says that a system of axioms entails xA(x) if and only

if it entails at least one of the instances A(cz) . It is, however, far from obvious
that the only if part of either of these semantic principles is correct. K2 builds
into the meaning of or the assumption that any satisfactory set theory will pos-
sess the disjunction property. Similarly, K6 builds into the meaning of the exis-
tential quantifier the assumption that any satisfactory set theory will possess the
existence property. As remarked in 5.2, it cannot be assumed that constructiv-
ist mathematical theories will possess these properties and, as we shall see, the
most popular version of intuitionistic set theory is not known to possess the exist-
ence property (see the start of 9.6). Moreover, both properties are anathema to

a classicist. For example, since any theory T classically entails A A for any
A, possession of the disjunction property implies that T will either entail A or its
negation; that is, it implies that T is semantically complete. K2 and K6, then, are
not plausible specifications of the accepted meanings of the sign for disjunction
or of the existential quantifier.
How might we give a more accurate specification of those meanings? When
addressing similar problems earlier in the book, Idefined a suitable closure
operator and took the semantic value of a disjunction to be the closure of the
union of the values of the disjuncts, rather than taking it to be the bare union of
those semantic values. (For various applications of this method, see 4.5, 6.4,
and 8.5.) In the present context, however, this method is inapplicable. Aclo-
sure operator is an operator on sets. It would be unwise, then, to invoke such
an operator here, where the shape of the set theory is precisely what is at issue.
To be sure, we must respect the idea that underlies our earlier treatments of

disjunctionnamely, that we often have warrant to assert A B even when
we lack warrant either to assert A or to assert B. However, in encapsulating
that idea in a semantic axiom, we need to proceed less directly than in earlier
chapters.
 On the Use of Classical Logic in Set Theory 287

What are we saying when we assert A B ? One natural answer is this:we

are ruling out any possibility that A and B should both be false, i.e., that A

and B should both be true, i.e., that the conjunction A B should be
true. This suggests, as a more plausible account of the meaning of than is

provided by K2, the following principle: A B is equivalent to (A B) .

Analogously, the principle that xA(x) is equivalent to A(x) is more
plausible than K6. The resulting treatment of the quantifiers is the obverse of
Fittings. Instead of replacing with , we take the universal quantifier as
basicwith its meaning specified by K5and explain the existential quantifier
in its terms.
This proposal will ring bells with those who remember the early history
of intuitionism. In a paper of 1933, Zur intuitionistischen Arithmetik und
Zahlentheorie (On Intuitionistic Arithmetic and Number Theory), Gdel
considered a formalization of classical (first-order) Peano Arithmetic in a lan-
guage L 1 whose sentential connectives are ~, ., , and and whose sole
quantifier is . He also considered a formalization of Heyting Arithmetic in
a language L 2 whose only connectives are and and whose sole quantifier
is again . (The usual formation rules regulate the construction of formulae
in L 1 and L 2 .) It is understood that speakers of L 2 use , , and in accord-
ance with the rules of intuitionistic logic. Gdel defined a translation from
well-formed formulae of L 1 into well-formed formulae of L 2 . Under this trans

lation, ~A maps to A , A . B to A B , A B to (A B) ,

A B to (A B) , and xA(x) maps to itself. He then proved the follow-
ing theorem:whenever A is a theorem of Peano Arithmetic, its translation A is
a theorem of (the present formalization of) Heyting Arithmetic. The existential
quantifier is not a primitive symbol of L 1; but, if it were, we could extend Gdels

translation by mapping xA(x) to A(x) . Gdels theorem would still
hold under the extended translation.
Commenting on this theorem, Gdel remarked that it shows that the sys-
tem of intuitionistic arithmetic and number theory is only apparently narrower
than the classical one, and in truth contains it, albeit with a somewhat deviant
interpretation (Gdel 1933a, 37=Gdel 1986, 295). The interpretations of and
indeed deviate from the intuitionists account of these symbols meanings.
As we have seen, though, the intuitionists account is implausible as a descrip-
tion of how even constructive mathematicians actually use these signs, and the
deviations have the merit of bringing us much closer to a correct description.
Accordingly, we might think of Gdels theorem as justifying the employment
of classical logic in elementary number theory:the theorem shows how some-
one who accepts Heyting Arithmetic might be rationally persuaded to accept
288 Five Attacks on Classical Logic

the classical Peano Arithmetic as a result of being brought to accept a more

accurate description of the use that he and other mathematicians actually make
of the sign for disjunction and the existential quantifier. Now the proof of the
theorem rests on particular features of elementary number theory so, as Gdel
rightly warns his readers (1933a, 37=1986, 295), it cannot be assumed that an
analogous argument will go through for set theory. It is, however, clearly worth
investigating how far the analogous justification for deploying classical logic in
set theory might take us.
In conducting this investigation, it helps to work with a slightly different trans-
lation from Gdels . Let us suppose that we use the connectives I, I, I, I,
and the quantifiers I and I in strict accordance with the semantic principles
K1 to K6. As we have seen, it is highly doubtful if such a use matches the normal
employment of the corresponding symbols even within constructive mathemat-
ics. All the same, K1 to K6 appear to succeed in conferring determinate senses
on the expressions I, I, etc., so we may use those expressions as meaningful
auxiliaries in characterizing our actual use of the connectives and quantifiers.
In that spirit, let LI be the first-order language whose only atomic predicates are
the two-place identity predicate =I and the two-place membership predicate
I . The well-formed formulae of LI are built up from atoms using I, I, etc.
in the usual way and their senses are given by K1 to K6. We then define a second
first-order language, L, whose well-formed formulae are precisely those of LI save
that all occurrences of the subscript I are deleted. Thus x=y and x y are
atomic formulae of L. We attach a sense indirectly to the formulae of L by specify-
ing a translation, N, that maps each such formula A to a well-formed formula A N
of LI; the sense of A is then the sense that A N bears in LI.
The mapping N is defined as follows:

T0. (s=t)N=II (s =I t) (s t)N=II (s I t)

T1. (A B)N=(A N I BN)
T2. (A B)N=II(A N I BN)
T3. (A)N=I (A N)
T4. (A B)N=(A N IBN)
T5. (xA(x))N=Ix(A(x))N
T6. (xA(x))N=II Ix(A(x))N.

Because II (A I B) is intuitionistically equivalent to I (I A I IB) and

IIIx(A(x)) is intuitionistically equivalent to IA(x) , the transla-
tions of and under N are essentially the same as their translations under
 On the Use of Classical Logic in Set Theory 289
14
Gdels original mapping. However,N, the so-called Gdel-Gentzen translation,
provides a smoother treatment of the conditional. The key results about N are
two:(1)if A is a theorem of the classical predicate calculus, then A N is a theo-

rem of the intuitionistic predicate calculus; (2)for any formula A, II A N I A N
is a theorem of the intuitionistic predicate calculus (for proofs, see Kleene 1952,
lemma 43a and theorem 60d). The Gdel-Gentzen translation may be regarded
as a syntactic correlate of some of the semantic ideas developed earlier in this

book. In the previous chapter, for example, Itook the semantic value of A B

to be (|A||B|) (see 8.5). Since the value of A was there taken to be |A|, this
axiom for yields directly what we now get indirectly as a result of combining
T2, K2, and K3.
The Gdel-Gentzen translation offers some hope of escaping from Taits coun-
terexample to classical logic. Where x is free to range over all sets, Tait argued,

we are not entitled to assert x(x) x(x) . Under the Gdel-Gentzen

translation, though, that formula maps to II (IxN(x) I II Ix N(x)) ,
which is a logical truth when the meanings of I, I, I, and I are given by
K2, K3, K5, and K6. In order for this reply to Tait to go through, however, it is not
enough to show that the meanings of etc. are correctly given by their trans-
lations under N. We also need to check that the Gdel-Gentzen translations
of the axioms of set theory are acceptable as used in our auxiliary language LI .
More precisely, we need to check that, when A is an axiom of classical set theory,
A N is a theorem of an intuitionistically acceptable set theory, formulated in LI .

9.6 Classical Logic in Set Theories Weaker than ZF

In one important respect, the problem just stated is ill-defined, for we have yet to
see which set theories an intuitionist might find acceptable. One first thinks sim-
ply of ZF, formulated with intuitionistic logic as the background logic, but this
nave conception faces two problems.
First, there is Myhills (1973) observation that the standard formulation of the
Axiom of Foundation intuitionistically implies the Law of Excluded Middle given
Zermelos original axioms (see 9.3). Myhill solved this problem by replacing

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

Foundation with the following principle of transfinite induction on , which

is classically equivalent to Foundation but which does not yield LEM when the
background logic is intuitionistic:

-induction rule : For any formula (x ),

from I x [(I y I x )( y ) I (x )] infer I x(x ).

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 ).

Replacement, however, is not intuitionistically equivalent to Collection. Given

intuitionistic logic as the background logic, then, we get a different set theory

once we replace Replacement with Collection. This latter theory is called ZF . ZF

has ZFI as a sub-theory and ZF has been shown to validate all its axioms in the
15
intended models. For this reason, ZF has supplanted ZFI as the most popular
intuitionistic set theory, and Ishall start by considering our problem as it pertains

to ZF . (It is worth noting that, while ZF has the disjunction property, it is not
known whether it has the existence property. So there is reason to doubt if K6
adequately specifies the meaning of the existential quantifier even as this is used
by constructive set theorists.)

Our problem, then, is whether A N is a theorem of ZF whenever A is an axiom
of ZF (with -induction in place of Foundation, and Collection replacing
Replacement). Harvey Friedman addressed this question in 1973, and obtained
an interesting result. When A is (or is an instance of) Unordered Pair, Union,

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)

AN is then I I a =I b I (I I a I c I I I b I c). And while

a =I b I (I I a I c I I I b I c)

is a theorem of ZF , II a =I b I (II a I c I II b I c) is not. For in intui-
tionistic first-order logic we do not in general have II a =I b I a =I b. Similarly,
let B be the following instance of Power Set:

x y (z(z y z a)) y x.

BN is then II I x Iy (Iz (II z I y I II z I a)) I II y I x), which is not

a theorem of of ZF .
What are we to make of the fact that neither Extensionality nor the Power Set
Axiom can be justified in this way? Friedmans primary interest was in the relative
consistency of various set theories. To this end, he introduced a deviant interpreta-

tion of under which the N translation of Extensionality is a theorem of ZF . He
also gave a weak form of the Power Set Axiom (to be stated below) whose translation

under N was again a theorem of ZF . For our purposes, though, Friedmans first
device is of no help. The Axiom of Extensionality, under the intended interpretation
of set membership, is constitutive of sets:it is part of an account of what the identity
of sets consists in. We shall not have a satisfactory justification of the use of classical
logic in set theory if we have to sacrifice Extensionality in order to achieve it.
In his paper Extending Gdels Negative Translation to ZF (Powell 1975),
William Powell defined an inner model of intuitionistic set theory in which the
negative translations of all the classical Zermelo-Fraenkel axioms are true. He
thereby took himself to have justified classical ZF set theory (see the heading of his
3, 1975, 225). Powells translation preserves the standard interpretation of . As he
remarks, since [the defined inner model] is a standard model, our interpretation
allows us to interpret classical statements as special statements about special sets
(1975, 221; emphasis in the original). That very point, though, shows that Powells

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 )

while the weak formis

x y (z x )(z = y a)

(z=y a abbreviates w (w z w y w a)). Given that the universe of

sets is not well-defined, there is no reason to suppose that, for any set a, the collec-
tion of sets obtained by intersecting a with each and every set is itself well-defined.
Asceptic about a well-defined universe of sets, then, should not be asking how the
classical form of the Power Set Axiom (whether in its weaker or stronger version)
might be justified within a set theory whose underlying logic is intuitionistic.
Rather, he should reject both versions of that Axiom as unjustified.
In fact, the same considerations that undermine the Power Set Axiom also
show that the schemas of ZF need restriction. Given an arbitrary set a, and an
arbitrary formula in the language of set theory, the schema of Separation
tells us that there exists a set comprising exactly those members of a that are
; more formally, it says that there exists a set x such that x={y a:(y)}. The
intuitive idea behind Separation is that we may separate those members of a
that satisfy from those that do not. If, however, the condition expressed by
is indeterminate, there may be no well-defined collectionand hence no
setof members of a that satisfy . If the universe of sets is not well-defined,
the condition expressed by may well be indeterminate if involves unre-
stricted quantification over all sets. Accordingly, those principles demand
that we place an appropriate restriction on the schema of Separation. Apar-
allel argument shows that the Collection schema also needs restriction.
What, exactly, might that restriction be? Let us call an occurrence x (x)
of a universal (existential) quantifier bounded if it occurs in a context x a
(x a), where a is a term that designates a set. Let us call a well-formed for-
mula a 0 formula if every occurrence in it of either the universal or the existential
294 Five Attacks on Classical Logic

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 )})

where (y) is any 0 formula.

Similarly, instead of the unrestricted Collection schema, we propose

0 -Collection : (x a)y ( x , y ) b (x a)(y b)(x , y )

where (x, y) is any 0 formula.

The set theory whose axioms are Extensionality, Unordered Pair, Union, instances
of 0-Separation and of 0-Collection, along with the -induction rule, is called
18
Kripke-Platek set theory, KP. If we add the Axiom of Infinity in the precise form that
there is a smallest set which contains the empty set and is closed under the set-theoretic
19
successor operation, we reach KP, that is, Kripke-Platek set theory with Infinity.
My suggestion is that KP (or something close to it) is the strongest set theory that
can be justified given the thesis that the universe of sets is not well-defined.
Kripke-Platek set theory has, indeed, features that commend it to those sympa-
thetic to that thesis. Amodel N of a theory is said to be an extension of a model M if
the domain, A, of M is included in that of N and the interpretations that M assigns
to the theorys vocabulary are simply the restrictions to A of the interpretations
that N assigns. Amodel N of a theory in the language of set theory is said to be an
end-extension of M if, in addition, N adds no new elements to the interpretation in
M of any term that designates a set. Where T is a theory in the language of set the-
ory, a formula is said to be absolute relative to T if, for any model M of T and any
end-extension N of M, gets the same assignment under M as under N. That is to
say, a formula that is absolute, relative to T, does not shift its semantic value as we
move from an arbitrary model of T to one of its end-extensions and back again.
All 0 formulae are absolute relative to any theory in the language of set theory
(for a proof, see Barwise 1975, 335). This result justifies the restricted forms of

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

Separation and Collection that KP proposes. As with the iterative conception

of the cumulative hierarchy of sets in ZF, we may think of sets in KP as being
formed in stages. Let us suppose that the set {y a:(y)} is formed at the stage
s. Assuming that is absolute, {y a:(y)} does not lose or acquire members
at a later stage in the process of set formation. Since any 0 formula is absolute,
0-Separation assures us that any set formed by separation has a constant mem-
bership as the domain of sets expands. If, per contra, were not a 0 formula, we
would have no guarantee that{y a:(y)} would not change in membership. For if
contained an unbounded quantifier, the range of this quantifier would increase
as more sets were generated, and this expansion might well affect which members
of a satisfied . Asimilar argument shows the need for the cognate restriction
on Collection. Let us suppose that, at a particular stage in the generation of the
domain of sets, we have (x a) y (x, y). We want to be able to collect into
a set some objects to which the members of a stand in the -relation. That is, we
want there to be a set b for which (x a) (y b) (x, y). If, however, contained
an unbounded quantifier, the newly constructed set b would fall within the
range of that quantifier, and the generation of b might then falsify the hypothesis
(x a) y (x, y). The assumption that is a 0 formula excludes this possibility,
so that 0-Collection is justified.
There would be no need for these restrictions on Separation and Collection if
there were a well-defined domain of all sets, for the stages could then be thought
of as mere heuristic props that help us to gain a sense of this all-encompassing
domain. On the view we have been developing, though, there is no well-defined
domain of sets. On that view, the versions of the Separation and Collection sche-
mas that are restricted to 0 formulae would seem to be the strongest versions
that can be justified.
Like ZF, KP is not a categorical theory:its intended models are the so-called
admissible sets, and there is an unbounded series of those (see Barwise 1975,
chap. 2). Taits argument may then be applied to show that the universe of
KP sets is not a mathematically well-defined totality. As in the case of ZF,
this forces the construction of a non-classical semantics for the language of
KP. If that semantics is built along the lines proposed in 9.3, the formula

x(x) x(x) will again turn out not to be invariably true when is an
arbitrary formula and when x ranges over all admissible sets. Interesting as
Kripke-Platek set theory may be, then, it may seem that it does not help at all in
the project of habilitating the use of classical logic in deductions that involve
quantifying over all sets.
296 Five Attacks on Classical Logic

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:

0 - LEM For any 0 formula , .

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

( I x I a)I I I y N(x , y ) I I I b( I x I a)I I (I y I b)N(x , y ).

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:

0 - MP For any 0 formula , x x.

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:

B OS (x a)((x ) (x )) (x a)(x ) (x a)(x )

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:

AC set (x a)y (x , y ) r (Fun(r ) Dom(r )=a (x a)(x , r (x ))).

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

in set theoryand hence a foundation independent of geometry. We cannot do

anything comparable in KP where we cannot even form the set of real numbers.
One upshot of our analysis, then, is scepticism about the project of constructing
real analysis within set theory. Such scepticism is, of course, not new. Indeed,
our discussion has brought us, by a new path, to a position close to that occupied
by Hermann Weyl in Das Kontinuum (1918; for interesting discussion, see also
Feferman 2009). Many mathematicians will find that position far too restrictive.
As far as Ican see, though, it is the strongest classical set theory that can be justi-
fied given the anti-realism about sets that has been our premiss.

9.7 As Far as Possible versus As Far as Necessary

Many readers, of course, will be led to reject that premiss. The justification given
for the use of classical logic in set theory, they will say, is just too artful. It works for
KP, but there is no guarantee that it will work for stronger set theories:where A is
a proposed new axiom, there is no guarantee that As Gdel-Gentzen translation,
AN, will be provable in the intuitionistic version of the strengthened theory, i.e., in
IKP + A. We surely want to be able to explore the consequences of a proposed new
axiom, even if in the end we decide not to accept it. If we are to be entitled to use clas-
sical logic in doing so, we would appear to need a justification of classical logic that
fits a variety of set theories, and is not tailored to the specific shape of KP.
I shall conclude this chapter by sketching, in very rough outline, where such a
justification might be foundfound, moreover, without imbibing the Platonist
mythology of a realm of sets that exist wholly independently of mathematical
practice. Different as ZF and KP are, they both embody what one might call
a maximalist conception of the set-theoretic universe:both admit into the uni-
verse of pure sets as many sets as it is possible to admit (they differ, of course, as
to which admissions are possible). There is, however, a contrary tradition in set
theory, according to which the set-theoretic universe comprises only those things
that it is necessary to admitnecessary, that is, for the theory to find application
either directly in empirical science or (more realistically) in branches of math-
ematics that are themselves empirically applicable. This idea lies behind Quines
and Putnams indispensability arguments for the existence of mathematical
entities (see e.g. Quine 1981 and Putnam 1971, esp.34756). It is also implicit in
Freges claim (made against the formalist mathematicians of his day) that it is
applicability alone which elevates arithmetic from a game to the rank of a science
or a body of truths (Wissenschaft) (Frege 1902, 91, 100):Freges argument for this
claim, if it works at all, works for branches of mathematics other than arithmetic.
It is, au fond, a pragmatist idea:the truth-evaluable content of a mathematical
300 Five Attacks on Classical Logic

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

K2. s A B if and only if s Cl {t: either t A or t B}.

Similarly we have

K6. s xAx if and only if s Cl {t: t A(a), for some a P(s)}.

The space of possible applications of mathematics is also naturally pre-ordered:

s t obtains if t requires the truth of any statement whose truth is required by
s. This pre-order generates a symmetric and irreflexive relation of incompat-
ibility between applications:s t if no possible application includes both, i.e.,
u (s u t u). This notion of incompatibility in turn enables us to state the

clause for negation in this semantic theory:an application requires A if it is
incompatible with any possible application that requires A:

K3. s A if and only if s {t: t A }.

 On the Use of Classical Logic in Set Theory 301

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.

10.1The Simple and Revised Arguments

for Bivalence
The argument Ihave in mind is all but explicit in Aristotle. Everybody remembers
Aristotles explanations of truth and falsity in the Metaphysics, but it is sometimes
forgotten that they form part of a case for Bivalence:

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:

(1) Say (u, P).

Given that starting point, we may set out what Ishall call the Simple Argument for
Bivalence:

(2) P P Excluded Middle

(3) P assumption
(4) Say (u, P) P (1), (3), -introduction
(5) (Say (u, P) P) True (u) principle about truth
(6) True (u) (3), (4)modus ponens
(7) True (u) False (u) (6), -introduction
(8) P assumption
(9) Say (u, P) P (1), (7), -introduction
(10) (Say (u, P) P) False (u) principle about falsehood
(11) False (u) (9), (10) modus ponens
(12) True (u) False (u) (11), -introduction
(13) True (u) False (u) (2), (7), (12) -elimination, with discharge of
assumptions (3)and (8)
Since u was an arbitrarily chosen statement, we may generalize from line (11) to
reach the conclusion Every statement is either true or false. The schema instanti-
ated at line (5), namely,

(Ts) (Say (u, P ) P ) True(u),

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,

(Fs) (Say (u, P ) P ) False(u)

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

(F ) u(False(u) P (Say (u, P ) P ))

(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:

(Det ) P Q(Say (u, P ) Say (u, Q)) (P Q ).

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

content to such a statement would be to hold that apparently unrestricted quan-

tification is implicitly restricted to the members of a standard model for set the-
ory. Since there are many such models, the relevant statements would on this
reading be counterexamples to (Det). So, for example, the statement There exists
a strongly inaccessible cardinal would say that there is such a cardinal in V1,
would also say that there is such a cardinal in V2, and so forth. These claims are
not even materially equivalent:as remarked in 9.1, the first inaccessible, 1, is a
member of V2 but not of V1. Iput the word say in scare quotes because it comes
unnaturally to use the term in this way:we want to ask, Well, which does it say?
On the view in question, though, the only available sense for u says that P is that
of An interpretation of u under which it says that P is as legitimate as any other
interpretation. The view in question is then that (a)a legitimate interpretation
must sustain classical logic and (b)there are many equally legitimate ways of
doing this.
For statements which exhibit this kind of indeterminacy, (T) is inadequate as
an account of truth:if an utterance that expresses multiple thoughts is to qualify
as true, it is not enough that one of those thoughts should be the case; all of them
must be. In other words, it must be the case that however the utterance says that
things are, thus they are. In a context where indeterminacy is a live option, then,
we need to replace (T) by (T*):

(T * ) u(True(u) (P Say (u, P ) Q(Say (u, Q) Q))).

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*):

(F * ) u(False(u) (P Say (u, P ) Q(Say (u, Q) Q ))).

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:

(1) P Say (u, P) premiss:u is a statement

(2) Say (u, P) (1), existential instantiation
(3) PQ(Say (u, P) Say (u, Q) (P Q)) premiss:u is determinate
(4) Q(Say (u, P) Say (u, Q) (P Q)) (3), universal instantiation
(5) Q(Say (u, Q) (P Q)) (2), (4)
(6) P P Excluded Middle
(7) P assumption
(8) Q(Say (u, Q) Q) (5), (7)
(9) P Say (u, P) Q(Say (u, Q) Q) (1), (8)-introduction
(10) (P Say (u, P) Q(Say (u, Q) Q)) principle about truth
True(u)
(11) True (u) (9), (10) modus ponens
(12) True (u) False (u) (11), -introduction
(13) P assumption
(14) Q(Say (u, Q) Q) (5), (13)
(15) P Say (u, P) Q(Say (u, Q) Q) (1), (14) -introduction
(16) (P Say (u, P) Q(Say (u, Q) Q)) principle about falsity
False(u)
(17) False (u) (15), (16) modus ponens
(18) True (u) False (u) (17), -introduction
(19) True (u) False (u) (5), (10), (15) -elimination,
with the discharge of
assumptions (6)and(11)

(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

as Icall it, is a restricted version of the Principle of Bivalence:every determinate

statement is either true or false.
The Revised Argument already shows how a classical logician may consistently
reject the Principle of Bivalence. The Revised Argument shows only that every
determinate statement is either true or false, so there will be counterexamples to
Bivalence if some statements fail to satisfy (Det). Thus, on the view on which some
set-theoretic statements are indeterminate, There exists a strongly inaccessible car-
dinal is neither true nor false, as is , on Bradwardines view of paradoxical utter-
ances. Whether there really are indeterminate statements is an important question
1
in the theory of truth but it is tangential to our inquiry. For, as Inow argue, even if
we assume that vague statements and statements involving unrestricted quantifica-
tion over sets are determinate, the arguments for Bivalence still fail.

10.2 Where the Arguments Go Wrong

In making this case, Ishall focus on the Simple Argument. Readers will have no
trouble extending the analysis to the Revised Argument.
Let us start by examining the Simple Argument as it applies to a vague state-
mentin particular, to a statement, u, that expresses the thought that the tube a50
of borderline red-orange paint is red. As it pertains to u, the Argument runs as
follows:
(1) Say (u, a50 is red) premiss
(2) a50 is red (a50 is red) Excluded Middle
(3) a50 is red assumption
(4) Say (u, a50 is red) a50 is red (1), (3), -introduction
(5) (Say (u, a50 is red) a50 is red) True (u) principle about truth
(6) True (u) (3), (4)modus ponens
(7) True (u) False (u) (6), -introduction
(8) (a50 is red) assumption
(9) Say (u, a50 is red) (a50 is red) (1), (7), -introduction
(10) (Say (u, a50 is red) (a50 is red)) False (u) principle about falsehood
(11) False (u) (9), (10) modus ponens
(12) True (u) False (u) (11), -introduction
(13) True (u) False (u) (2), (7), (12) -elimination,
with discharge of
assumptions (3)and(8)

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,

(1) Say (u, P ),

310 Conclusion
2
where the schematic letter P is replaced by a statement of CH. We may then
run through the Simple Argument for this replacement for P. Many philoso-
phers and mathematicians, however, will resist the conclusion of this instance of
the Argumentthe conclusion, namely, that a statement of CH is either true or
false. Gdel (1940) showed that the truth of CH was consistent with the axioms
of Zermelo-Fraenkel set theory including the Axiom of Choice (ZFC), assum-
ing that those axioms are themselves consistent. Cohen (1966) proved that its
falsity was also consistent with the same axioms under the same assumption.
These results do not refute the claim that CH is either true or false, but they do
cast doubt upon it. Many philosophers believe that a true statement must have a
basis:if a statement is true, there must be something in virtue of which it is true,
even if we can never find out whether that basis obtains. They also accept the cor-
responding principle for falsity. Gdel and Cohens results appear to show that, in
the case of CH, there is no basis for its truth or its falsity in ZFC.
Indeed, the point holds for set theories stronger than ZFC. The further axi-
oms that have gained reasonably wide acceptance among set theorists are large
cardinal axioms, which postulate sets at higher levels of the set-theoretic hierar-
chy than ZFC is committed to. However, Levy and Solovay (1967) proved that CH
is consistent with, and independent of, all large cardinal axioms, provided that
those axioms are themselves consistent. At around the turn of the millennium,
Hugh Woodin put forward his Strong Conjecture which, he claimed, decides
CH negatively (see Woodin 2001 and 2005). He tried to justify his Conjecture as
an extension of the Axiom of Projective Determinacy, but it has failed to gain
wide acceptance by set theorists and Woodin now thinks that CH may be true.
There is, then, no current prospect of set theorys developing in such a way as to
yield a generally accepted basis either for CHs truth or for its falsity.
For all that, some set theorists and logicians hold that CH is bivalent. Asignal
example was Gdel himself. Even after he had learned of Cohens result, Gdel
remained confident that the question of the truth or falsity of CH was, as he put
it, meaningful; it will be so only if CH is bivalent. However, whilst he accepted
its conclusion, Gdels position offers no comfort to a proponent of the present
instance of the Simple Argument. Nowhere in the various writings in which he
advances reasons for supposing that CH is bivalent does Gdel appeal to the
Argument, or to anything like it. Rather, he adduces less direct considerations.

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

(T )N u(True(u) P (Say (u, P ) P ))

and that of (F)

(F )N u(False(u) P (Say (u, P ) P )).

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

Say (u, P ) Say (u, P ).

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

that fundamental mathematical theories are most naturally presented in their

second-order formalizations, and he went on to argue that either CH or its nega-
3
tion is determined by the axioms of second-order set theory:
Let Z be Zermelos axiom with the axiom of infinity, and let CH be the (canonical) for-
mulation of the continuum hypothesis in the following form:if C is the collection of
hereditarily finite sets without individuals, C + 1=C (C), C+2=C+1 (C +1), CH
states that
X C +1 (| X | | C | | X | = | C +1 |) ,

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)

The single turnstile 2 is standardly used to signify deducibility in axiomatic

second-order logic, but Kreisel does not (or should not) mean this:Weston (1976,
28990) extends an argument of Tharps to show that neither CH nor CH is a
2
theorem of axiomatic ZFC (so that, a fortiori, neither is a theorem of the weaker
2
system Z ). Because second-order logic is incomplete, though, the Tharp-Weston
2 2
result is consistent with the claim that either ZF 2 CH or ZF 2 CH, where 2
signifies the model-theoretic consequence relation of full second-order logic.
4
This claim is the proper conclusion of Kreisels argument.
That argument is compressed, but it may be spelled out as follows. As was
shown in 9.1, where T is a fully categorical theory formulated in a second-order
language L, the same closed sentences of L hold in any two models of T. That is,
where M and N are any two models of T, and is any closed sentence of L, M2
2
if and only if N2 . Because ZF is only quasi-categorical, this argument does not

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

arbitrary choice, for each member of {1, 2, 3,...}, as to whether to include it in a

subset or not. Talk of making an arbitrary choice, for each member of an infi-
nite set, exactly involves projecting into the infinite a notion that only makes
sense for finite sets. So at least this justification for the existence of () is ille-
gitimate. The illegitimacy of one purported justification of () does not show
that the existence of that set is incapable of being justified. All the justifications
that Iknow, though, suffer from similar flaws, and until we have a convincing
justification, Kreisels argumentalong, of course, with much contemporary
mathematicsis doubtful. Of course, the Simple Argument for the bivalence
of CH does no better in these circumstances. Isuggested in Chapter9 that we

could affirm A A whenever A involves only bounded quantification over
sets. But unless () and (()) exist, CH is no such statement, so we cannot

affirm CH CH , and the Simple Argument does not even get started.

10.3 Classical Logic versus Classical Semantics

The failure of the Simple and Revised Arguments lies at the heart of the present
project. Some may question the value of a treatise that defends classical logic when
the vast majority of philosophers adhere to it anyway. In reply, Iwould quote a
couplet from Murder in the Cathedral:

The last temptation is the greatest treason:

To do the right deed for the wrong reason.

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

Adams, E.W. 1975. The Logic of Conditionals. Dordrecht:Reidel.

Andjelkovi, M. and T. Williamson. 2000. Truth, falsity and borderline cases.
Philosophical Topics 28:21144.
Artemov, S.N. 2001. Explicit provability and constructive semantics. The Bulletin of
Symbolic Logic 7:136.
Auxier, R. E. and L. E. Hahn, eds. 2007. The Library of Living Philosophers Volume
XXXI:The Philosophy of Michael Dummett. Chicago, IL:Open Court.
Avigad, J. 2000. Interpreting classical theories in constructive ones. The Journal of
Symbolic Logic 65:17851812.
Bain, A. 1868. Mental and Moral Science. London:Longman.
Barwise, K.J. 1975. Admissible Sets and Structures:An Approach to Definability Theory.
Berlin:Springer.
Barwise, K.J. and J. R.Perry. 1983. Situations and Attitudes. Cambridge, MA:MIT Press.
Barwise, K.J. and J. R.Perry. 1985. Shifting situations and shaken attitudes. Linguistics
and Philosophy 8: 10561. Page reference to the reprint in Barwise and Perry 1999,
pp. xxxiiilxxxviii.
Barwise, K. J. and J. R. Perry. 1999. Situations and Attitudes, 2nd edition. Stanford,
CA:CSLI Publications.
Beall, J.C. and G. Restall. 2006. Logical Pluralism. Oxford:Clarendon Press.
Bell, J.L. 1986. A new approach to quantum logic. The British Journal for the Philosophy
of Science 37:8399.
Bell, J.L. 2008. A Primer of Infinitesimal Analysis, 2nd edition. Cambridge:Cambridge
University Press.
Benacerraf, P. 1965. What numbers could not be. The Philosophical Review 74:495512.
Benacerraf, P. and H. W. Putnam. 1983. Introduction to P. Benacerraf and H. W. Putnam,
eds., Philosophy of Mathematics:Selected Readings, 2nd edition. Cambridge:Cambridge
University Press, pp. 137.
Berker, S. 2008. Luminosity regained. Philosophers Imprint 8.2:122.
Berlin, B. and P. Kay. 1969. Basic Color Terms:Their Universality and Evolution. Berkeley,
CA:University of California Press.
Bernays, P.I. 1967. What do some recent results in set theory suggest? In Lakatos, ed.,
1967, vol. I, pp. 10917.
Beth, E.W. 1959. The Foundations of Mathematics. Amsterdam:North Holland.
Boghossian, P. 2000. Knowledge of logic. In P. Boghossian and C. A.B. Peacocke, eds.,
New Essays on the APriori. Oxford:Clarendon Press, pp. 22954.
Bostock, D. 1988. Necessary truth and a priori truth. Mind 97:34379.
Braithwaite, R.B. 1933. The nature of believing. Proceedings of the Aristotelian Society
33:12946.
Brandom, R. B. 2008. Between Saying and Doing: Towards an Analytic Pragmatism.
Oxford:Oxford University Press.
322 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

Egan, A. and B. Weatherson, eds. 2011. Epistemic Modality. Oxford:Clarendon Press.

Eliot, T.S. 2014. The validity of artificial distinctions. In The Complete Prose of T.S. Eliot,
Volume I: Apprentice Years, 19051918, ed., J. SpearsBrooker and R. Schuchard. Baltimore,
MD and London: The Johns Hopkins University Press and Faber & Faber, pp. 18792.
Etchemendy, J. 1990. The Concept of Logical Consequence. Cambridge, MA: Harvard
University Press.
Evans, M.G. J. 1978. Can there be vague objects? Analysis 38:208. Reprinted in Evans
1985, pp.1767.
Evans, M.G. J. 1979. Reference and contingency. The Monist 62:16189. Page references
are to the reprint in Evans 1985, pp.178213.
Evans, M.G. J. 1982. The Varieties of Reference. Oxford:Clarendon Press.
Evans, M.G. J. 1985. Collected Papers. Oxford:Clarendon Press.
Evans, M.G. J. 2004. Comments on Two notions of necessity. Philosophical Studies
118:1116.
Evans, M.G. J. and J. H.McDowell, eds. 1976. Truth and Meaning:Essays in Semantics.
Oxford:Clarendon Press.
Ewald, W. B., ed. 1996. From Kant to Hilbert: A Source Book in the Foundations of
Mathematics, Volume II. Oxford:Clarendon Press.
Feferman, S. 2009. Conceptions of the continuum. Intellectica 51:16989.
Feferman, S. 2010. On the strength of some semi-constructive theories. In S. Feferman
and W. Sieg, eds., Proofs, Categories and Computations: Essays in Honor of Grigori
Mints. London:College Publications, pp. 10929.
Feferman, S. Forthcoming. Is the Continuum Hypothesis a definite mathematical problem?
Ferreira, F. 2008. Coordination principles:a problem for bilateralism. Mind 117:10517.
Field, H.H. 1998. Which undecidable mathematical sentences have determinate truth
values? In H. G. Dales and G. Oliveri, eds., Truth in Mathematics. Oxford:Clarendon
Press, pp. 291310.
Field, H.H. 2008. Saving Truth from Paradox. NewYork:Oxford University Press.
Fine, K. 1975. Vagueness, truth, and logic. Synthese 30:265300. Page references are to the
reprint in Keefe and Smith, eds., 1996, pp.11950.
Fine, K. 1989. The problem of de re modality. In J. Almog, J. Perry, and H. Wettstein, eds.,
Themes from Kaplan. NewYork:Oxford University Press, pp. 197272.
Fine, K. 2002. The varieties of necessity. In T. S. Gendler and J. Hawthorne, eds.,
Conceivability and Possibility. Oxford:Clarendon Press, pp. 25382.
Fine, K. 2012a. Counterfactuals without possible worlds. The Journal of Philosophy
109:22146.
Fine, K. 2012b. A difficulty for the possible worlds analysis of counterfactuals. Synthese
189:2957.
Finkelstein, D. 1968. Matter, space, and logic. In Cohen and Wartofsky, eds., 1968, pp. 199215.
Fitting, M.C. 1969. Intuitionistic Logic, Model Theory, and Forcing. Amsterdam:North
Holland.
Frege, F.L. G. 1879. Begriffsschrift. Eine der arithmetischen nachbegildete Formelsprache
des reinen Denkens. Halle:Louis Nebert.
Frege, F.L. G. 1880/81. Booles rechnende Logik und die Begriffsschrift. In Frege 1969,
pp. 952. English translation in Frege 1979, pp.946.
 References 325

Frege, F.L. G. 1884. Die Grundlagen der Arithmetik. Breslau:Wilhelm Koebner.

Frege, F.L. G. 1893. Grundgesetze der Arithmetik, vol. I. Jena:Hermann Pohle.
Frege, F.L. G. 1895. Kritische Beleuchtung einiger Punkte in E.Schrders Vorlesungen
ber die Algebra der Logik. Archiv fr systematische Philosophie 1: 43356. English
translation in Frege 1984, pp.21028.
Frege, F.L. G. 1897. Logik. In Frege 1969, pp.13763. English translation in Frege 1979,
pp.12651.
Frege, F.L. G. 18978. Begrndung meiner strengeren Grundstze des Definierens. In
Frege 1969, pp. 16470. English translation in Frege 1979, pp.1526.
Frege, F.L. G. 1902. Grundgesetze der Arithmetik, vol. II. Jena:Hermann Pohle.
Frege, F.L. G. 1906. Was kann ich als Ergebnis meiner Arbeit ansehen? In Frege 1969,
p.200. English translation in Frege 1979, p.184.
Frege, F.L. G. 1914. Logik in der Mathematik. In Frege 1969, pp. 21970. English transla-
tion in Frege 1979, pp.20350.
Frege, F.L. G. 1969. Nachgelassene Schriften, ed. H. Hermes etal. Hamburg:Felix Meiner.
Frege, F.L. G. 1976. Wissenschaftlicher Briefwechsel, ed. G. Gabriel etal. Hamburg:Felix
Meiner.
Frege, F.L. G. 1979. Posthumous Writings, trans. P. Long and R. White. Oxford:Blackwell.
Frege, F. L. G. 1980. Philosophical and Mathematical Correspondence, trans. H. Kaal.
Oxford:Blackwell.
Frege, F. L. G. 1984. Collected Papers on Mathematics, Logic, and Philosophy, ed.
B. McGuinness. Oxford:Blackwell.
Friedman, H.M. 1973. The consistency of classical set theory relative to a set theory with
intuitionistic logic. The Journal of Symbolic Logic 38:31519.
Friedman, H.M. 1975. The disjunction property implies the numerical existence prop-
erty. Proceedings of the National Academy of Science 72:28778.
Gallistel, C. R. 1990. The Organization of Learning. Cambridge, MA:MIT Press.
Grdenfors, P. 2000. Conceptual Spaces:The Geometry of Thought. Cambridge, MA:MIT
Press.
Gardner, M.R. 1971. Is quantum logic really logic? Philosophy of Science 38:50829.
Geach, P.T. 1955. Class and concept. The Philosophical Review 64:56170.
Geach, P.T. 1957. Mental Acts. London:Routledge & Kegan Paul.
Geach, P.T. 1965. Assertion. The Philosophical Review 74:44965.
Geach, P. T. 1980. Reference and Generality: An Examination of Some Medieval and
Modern Theories, 3rd edition. Ithaca, NY:Cornell University Press.
Gentzen, G. K. E. 1933. ber das Verhltnis zwischen intuitionistischer und class
ischer Arithmetik. Paper accepted for publication in Mathematische Annalen,
but withdrawn by Gentzen at proof stage after he had read Gdel 1933a. Translated
as On the relation between intuitionist and classical arithmetic in Gentzen 1969,
pp. 5367.
Gentzen, G. K. E. 1935. Untersuchungen ber das logische Schliessen. Mathematische
Zeitschrift 39:176210, 40531. Translated as Investigations into logical deduction in
Gentzen 1969, pp.68131.
Gentzen, G. K. E. 1969. The Collected Papers of Gerhard Gentzen, ed. and trans. M. E.
Szabo. Amsterdam:North Holland.
326 References

Gibbard, P. 2002. Price and Rumfitt on rejective negation. Mind 111:297303.

Glivenko, V.I. 1929. Sur quelques points de la logique de M.Brouwer. Acadmie royale de
Belgique:Bulletin de la classe des sciences 15:1838.
Gdel, K. F. 1933a. Zur intuitionistischen Arithmetik und Zahlentheorie. Ergebnisse
eines mathematischen Kolloquiums 4: 348. Reprinted, with a translation by
S.Bauer-Mengelberg and J.van Heijenoort, in Gdel 1986, pp.28695.
Gdel, K.F. 1933b. Eine interpretation des intuitionistischen Aussagenkalkl. Ergebnisse
eines mathematischen Kolloquiums 4:3940. Reprinted, with a translation, in Gdel
1986, pp.3003.
Gdel, K. F. 1940. The Consistency of the Axiom of Choice and of the Generalized
Continuum-Hypothesis with the Axioms of Set Theory. Princeton, NJ: Princeton
University Press.
Gdel, K.F. 1964. What is Cantors continuum problem? (revised and expanded version).
In P. Benacerraf and H. W. Putnam, eds., Philosophy of Mathematics:Selected Readings.
Englewood Cliffs, NJ:Prentice-Hall, pp. 25873. Reprinted (with the original pagina-
tion indicated) in Gdel 1990, pp.25470.
Gdel, K.F. 1986. Collected Works, Volume I, ed. S. Feferman etal. NewYork:Oxford
University Press.
Gdel, K.F. 1990. Collected Works, Volume II, ed. S. Feferman etal. NewYork:Oxford
University Press.
Goldblatt, R. I. 1974. Semantic analysis of orthologic. Journal of Philosophical Logic
3:1935.
Goldblatt, R.I. 1984. Topoi:The Categorial Analysis of Logic. Amsterdam:Elsevier.
Goldblatt, R. I. 2011. Quantifiers, Propositions and Identity: Admissible Semantics
for Quantified Modal and Substructural Logics. NewYork:Cambridge University
Press.
Goodman, N. D. 1985. Replacement and collection in intuitionistic set theory. The
Journal of Symbolic Logic 50:3448.
Graff Fara, D. 2003. Gap principles, penumbral consequence, and infinitely
higher-order vagueness. In J. C.Beall, ed., Liars and Heaps:New Essays on Paradox.
Oxford:Clarendon Press, pp. 195221.
Grayson, R.J. 1979. Heyting-valued models for intuitionistic set theory. In M. P.Fourman,
C. J.Mulvey, and D. S.Scott, eds., Applications of Sheaves:Lecture Notes in Mathematics
Volume 753. Berlin:Springer, pp. 40214.
Hale, R.L. V. and A. Hoffmann, eds. 2010. Modality:Metaphysics, Logic, and Epistemology.
Oxford:Clarendon Press.
Halmos, P.R. 1963. Lectures on Boolean Algebras. Princeton, NJ:Van Nostrand.
Hardegree, G.M. 1974. The conditional in quantum logic. Synthese 29:6380.
Hawthorne, J. 2004. Knowledge and Lotteries. Oxford:Clarendon Press.
Hazen, A.P. 1982. On a possible misinterpretation of Kripkes semantics for intuitionistic
logic. Analysis 42:12833.
Heck, R.G. 1993. Critical notice of M.Dummett, Frege:Philosophy of Mathematics. The
Philosophical Quarterly 43:22333.
Heck, R.G. 1998. That there might be vague objects (so far as concerns logic). The Monist
81:27799.
 References 327

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

Lewis, D.K. 1986a. Counterfactuals, revised edition. Oxford:Blackwell.

Lewis, D.K. 1986b. On the Plurality of Worlds. Oxford:Blackwell.
Lewis, D.K. 1988. Vague identity:Evans misunderstood. Analysis 48:128130.
Lyaletski, A. 2008. Herbrand theorems:the classical and intuitionistic cases. Studies in
Logic, Grammar and Rhetoric 14:10122.
Lycan, W.G. 2001. Real Conditionals. Oxford:Clarendon Press.
McDowell, J.H. 1976. Meaning, bivalence, and verificationism. In Evans and McDowell,
eds., 1976, pp. 4266.
McDowell, J. H. 1978. On The Reality of the Past. In C. Hookway and P. Pettit,
eds., Action and Interpretation: Studies in the Philosophy of the Social Sciences.
Cambridge:Cambridge University Press, pp. 12744.
McDowell, J.H. 1982. Truth-value gaps. In L. J.Cohen etal., eds., Logic, Methodology,
and Philosophy of Science VI. NewYork:North Holland, pp. 299313.
McFetridge, I.G. 1990a. Logical necessity:some issues. In McFetridge 1990c, pp. 13554.
McFetridge, I.G. 1990b. Explicating x knows a priori that p. In McFetridge 1990c,
pp. 21332.
McFetridge, I.G. 1990c. Logical Necessity and Other Essays. London:Aristotelian Society.
McGee, V. 1985. A counterexample to Modus Ponens. The Journal of Philosophy 82:
46271.
McGee, V. 1992. Two problems with Tarskis theory of consequence. Proceedings of the
Aristotelian Society 92:27392.
McGee, V. 1997. How we learn mathematical language. The Philosophical Review
106:3568.
McGinn, C. 1982. The structure of content. In A. Woodfield, ed., Thought and
Object:Essays on Intentionality. Oxford:Clarendon Press, pp. 20758.
Mackie, J. L. 1973. Truth, Probability, and Paradox: Studies in Philosophical Logic.
Oxford:Clarendon Press.
McKinsey, J.C. C. and A. Tarski. 1948. Some theorems about the sentential calculi of
Lewis and Heyting. The Journal of Symbolic Logic 13:115.
Mac Lane, S. 1994. Categories for the Working Mathematician, 2nd edition.
Berlin:Springer.
Malinowski, J. 1990. The deduction theorem for quantum logic:some negative results.
The Journal of Symbolic Logic 55:61525.
Mallory, J.P. and D. Q.Adams. 2006. The Oxford Introduction to Proto-Indo-European
and the Proto-Indo-European World. NewYork:Oxford University Press.
Mares, E.D. 2010. The nature of information:a relevant approach. Synthese 175:11132.
Mates, B. 1965. Elementary Logic. NewYork:Oxford University Press.
Mates, B. 1986. The Philosophy of Leibniz:Metaphysics and Language. NewYork:Oxford
University Press.
Maudlin, T. 2004. Truth and Paradox:Solving the Riddles. NewYork:Oxford University Press.
Mellor, D.H. 2012. Successful semantics. In his Mind, Meaning, and Reality:Essays in
Philosophy. Oxford:Clarendon Press, pp. 6077.
Menzel, C. 1986. On the iterative explanation of the paradoxes. Philosophical Studies
49:3761.
Menzel, C. 2014. Wide sets, ZFCU, and the iterative conception. The Journal of Philosophy
111:5783.
330 References

Mill, J. S. 1891. A System of Logic, Ratiocinative and Inductive, 8th edition.

London:Longman.
Moerdijk, I. and G. E. Reyes. 1991. Models for Smooth Infinitesimal Analysis.
NewYork:Springer.
Moore, G.E. 1922. Philosophical Studies. London:Routledge & Kegan Paul.
Moore, J. 1999. Propositions without identity. Nos 33:129.
Mostowski, A. 1967. Recent results in set theory. In Lakatos, ed., 1967, vol. I, pp. 8296.
Muller, F.A. 2001. Sets, classes, and categories. The British Journal for the Philosophy of
Science 52:53973.
Myhill, J. R. 1952. A derivation of number theory from ancestral theory. The Journal of
Symbolic Logic 17:1927.
Myhill, J. R. 1973. Some properties of intuitionistic Zermelo-Fraenkel set theory. In
A. R. D. Mathias and H. Rogers, eds., Cambridge Summer School in Mathematical
Logic:Lecture Notes in Mathematics Volume 337. Berlin:Springer, pp. 20631.
Negri, S. and J. von Plato. 2001. Structural Proof Theory. Cambridge: Cambridge
University Press.
Oliver, A.D. 1998. Hazy totalities and indefinitely extensible concepts:an exercise in
the interpretation of Dummetts philosophy of mathematics. Grazer Philosophische
Studien 55:2550.
Orowska, E. 1985. Semantics of vague concepts. In G. Dorn and P. Weingartner, eds.,
Foundations of Logic and Linguistics:Problems and their Solutions. NewYork:Plenum
Press, pp. 46582.
Parsons, C. D. 2008. Mathematical Thought and Its Objects. Cambridge: Cambridge
University Press.
Parsons, T. 2000. Indeterminate Identity:Metaphysics and Semantics. Oxford:Clarendon
Press.
Paseau, A. 2001. Should the logic of set theory be intuitionistic? Proceedings of the
Aristotelian Society 101:36978.
Paseau, A. 2003. The open-endedness of the set concept and the semantics of set theory.
Synthese 135:37999.
Pawlak, Z. 1982. Rough sets. International Journal of Computer and Information Sciences
11:34150.
Platek, R. 1966. Foundations of Recursion Theory. Ph.D. dissertation, Stanford
University.
Plato, J.von and A. Siders. 2012. Normal derivability in classical natural deduction. The
Review of Symbolic Logic 5:20511.
Polkowski, L. 2002. Rough Sets: Mathematical Foundations. Heidelberg: Physical-
Verlag.
Powell, W.C. 1975. Extending Gdels negative translation to ZF. The Journal of Symbolic
Logic 40:2219.
Powell, W.C. 1976. A completeness proof for Zermelo-Fraenkel set theory. The Journal of
Symbolic Logic 41:3237.
Prawitz, D. 1965. Natural Deduction:AProof-Theoretical Study. Stockholm:Almqvist and
Wiksell.
Prawitz, D. 1974. On the idea of a general proof theory. Synthese 27:6377.
 References 331

Price, H. 1983. Sense, assertion, Dummett, and denial. Mind 92:16173.

Price, H. 1990. Why not? Mind 99:22138.
Priest, G. 2001. An Introduction to Non-Classical Logic. Cambridge: Cambridge
University Press.
Priest, G. 2006. Doubt Truth to be a Liar. Oxford:Clarendon Press.
Prior, A.N. 1958. Epimenides the Cretan. The Journal of Symbolic Logic 23:2616.
Prior, A.N. 1961. On a family of paradoxes. Notre Dame Journal of Formal Logic 2:1632.
Prior, A.N. 1971. Objects of Thought. Oxford:Clarendon Press.
Putnam, H.W. 1968. Is logic empirical? In R. S.Cohen and M. W.Wartofsky, eds., 1968,
pp. 21641. Page references are to the reprint (which appeared under the title The logic
of quantum mechanics) in Putnam 1979, pp.17497.
Putnam, H.W. 1971. Philosophy of Logic. NewYork:Harper & Row. Page references are to
the reprint in Putnam 1979, pp. 32357.
Putnam, H.W. 1979. Mathematics, Matter and Method:Philosophical Papers Volume 1,
2nd edition. Cambridge:Cambridge University Press.
Putnam, H. W. 1983. Vagueness and alternative logic. In his Realism and
Reason: Philosophical Papers Volume 3. Cambridge: Cambridge University Press,
pp. 27186.
Putnam, H.W. 1985. A quick Read is a wrong Wright. Analysis 45:203.
Putnam, H.W. 1994. Michael Redhead on quantum logic. In P. Clark and R. L.V. Hale,
eds., Reading Putnam. Oxford:Blackwell, pp. 26580.
Putnam, H. W. 2012. Philosophy in an Age of Science: Physics, Mathematics, and
Skepticism. Cambridge, MA:Harvard University Press.
Quine, W.V. 1951. Two dogmas of empiricism. The Philosophical Review 60:2043.
Quine, W.V. 1953. Mr Strawson on logical theory. Mind 62:43351. Page references are to
the reprint in Quine 1976, pp. 13757.
Quine, W.V. 1960. Carnap and logical truth. Synthese 12:35074. Page references are to
the reprint in Quine 1976, pp.10732.
Quine, W. V. 1969. Natural kinds. In his Ontological Relativity and Other Essays.
NewYork:Columbia University Press, pp. 11438.
Quine, W.V. 1976. The Ways of Paradox and Other Essays, revised and enlarged edition.
Cambridge, MA:Harvard University Press.
Quine, W.V. 1981. Success and limits of mathematization. In his Theories and Things.
Cambridge, MA:Harvard University Press, pp. 14855.
Quine, W.V. 1982. Methods of Logic, 4th edition. Cambridge, MA:Harvard University
Press.
Quine, W.V. 1986. Philosophy of Logic, 2nd edition. Cambridge, MA:Harvard University
Press.
Raatikainen, P. 2004. Conceptions of truth in intuitionism. History and Philosophy of
Logic 25:13145.
Radford, C. 1985. The umpires dilemma. Analysis 45:10911.
Raffman, D. 2005. Borderline cases and bivalence. The Philosophical Review 114:131.
Ramsey, F. P. 1927. Facts and propositions. Proceedings of the Aristotelian Society
Supplementary Volumes 7:15370. Page references are to the reprint in Ramsey 1990,
pp.3451.
332 References

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

Ryle, G. 1949. The Concept of Mind. London:Hutchinson.

Ryle, G. 1950. If, so, and because . In Max Black, ed., Philosophical
Analysis:ACollection of Essays. Englewood Cliffs, NJ:Prentice Hall, pp. 30218. Page
references are to the reprint in Ryle 2009, pp.24460.
Ryle, G. 2009. Collected Papers, Volume 2. London:Routledge.
Sainsbury, R.M. 1990. Concepts without boundaries. Inaugural lecture at the University
of London. Page references are to the reprint in Keefe and Smith, eds., 1996, pp.25164.
Sainsbury, R. M. 1991. Is there higher-order vagueness? Philosophical Quarterly
41:16782.
Sainsbury, R. M. 1997. Easy possibilities. Philosophy and Phenomenological Research
57:90719.
Sambin, G. 1995. Pretopologies and completeness proofs. The Journal of Symbolic Logic
60:86178.
Schulz, S.M. 1993. Modal situation theory. In P. Aczel etal., eds., Situation Theory and
its Applications. Stanford, CA:Center for the Study of Language and Information,
pp. 16388.
Scott, D.S. 1961. Measurable cardinals and constructible sets. Bulletin de lAcadmie
Polonaise des Sciences 9:5214.
Scott, D. S. 1974. Completeness and axiomatizability in many-valued logic. In L.
Henkin etal., eds., Proceedings of the Tarski Symposium. Providence, RI:American
Mathematical Society, pp. 41135.
Seldin, J. P. 1986. On the proof theory of the intermediate logic MH. The Journal of
Symbolic Logic 51:62647.
Shapiro, S. 1991. Foundations without Foundationalism:ACase for Second-Order Logic.
Oxford:Clarendon Press.
Shapiro, S. and C. J.G. Wright. 2006. All things indefinitely extensible. In A. Rayo and G.
Uzquiano, eds., Absolute Generality. Oxford:Clarendon Press, pp. 255304.
Shieh, S. 1998. Undecidability in anti-realism. Philosophia Mathematica 6:32433.
Smiley, T.J. 1993. Can contradictions be true? Proceedings of the Aristotelian Society
Supplementary Volumes 67:1733.
Smiley, T.J. 1995. A tale of two tortoises. Mind 104:72536.
Smiley, T.J. 1996. Rejection. Analysis 56:19.
Smiley, T. J. 1998. Conceptions of consequence. In E. J. Craig, ed., The Routledge
Encyclopedia of Philosophy. London:Routledge, vol. II, pp. 599603.
Sommers, F.T. 1982. The Logic of Natural Language. Oxford:Clarendon Press.
Stlmarck, G. 1991. Normalization theorems for full first order classical natural deduc-
tion. The Journal of Symbolic Logic 56:12949.
Stalnaker, R.C. 1976. Possible worlds. Nos 10:6575. Page references are to the reprint in
Stalnaker 2003, pp.2539.
Stalnaker, R.C. 1978. Assertion. Syntax and Semantics 9:31532. Page references are to
the reprint in Stalnaker 1999b, pp.7895.
Stalnaker, R.C. 1991. The problem of logical omniscience, I. Synthese 89:42540. Page
references are to the reprint in Stalnaker 1999b, pp.24154.
Stalnaker, R.C. 1999a. The problem of logical omniscience, II. In Stalnaker 1999b, pp.
25573.
334 References

Stalnaker, R. C. 1999b. Context and Content: Essays on Intentionality in Speech and

Thought. Oxford:Oxford University Press.
Stalnaker, R. C. 2003. Ways a World Might Be: Metaphysical and Anti-Metaphysical
Essays. Oxford:Clarendon Press.
Stalnaker, R.C. 2010. Merely possible propositions. In Hale and Hoffmann, eds., 2010,
pp. 2132.
Statman, R. 1974. Structural complexity of proofs. Ph.D. dissertation, Stanford
University.
Steinberger, F. 2011. Why conclusions should remain single. Journal of Philosophical
Logic 40:33355.
Stevenson, C.L. 1970. If-iculties. Philosophy of Science 37:2749.
Strawson, P. F. 1950. On referring. Mind 59: 32044. Page references are to the
amended reprint in Strawson, Logico-Linguistic Papers. London: Methuen, 1971,
pp. 127.
Sullivan, P.M. 2007. Dummetts case for constructivist logicism. In Auxier and Hahn,
eds., 2007, pp. 75385.
Tait, W.W. 1994. The law of excluded middle and the axiom of choice. In A. George, ed.,
Mathematics and Mind. Oxford:Clarendon Press, pp. 4570. Page references are to the
reprint in Tait 2005a, pp.10532.
Tait, W.W. 1998. Zermelos conception of set theory and ref lection principles. In
M. Schirn, ed., The Philosophy of Mathematics Today. Oxford:Clarendon Press,
pp. 46983.
Tait, W.W. 2005a. The Provenance of Pure Reason:Essays in the Philosophy of Mathematics
and its History. NewYork:Oxford University Press.
Tait, W.W. 2005b. Constructing cardinals from below. In Tait 2005a, pp. 13354.
Tarski, A. 1929. Les fondements de la gometrie des corps. Supplement to Annales de la
Socit Polonaise de Mathmatique (Krakow), pp. 2933. Translated by J. H.Woodger
as Foundations of the geometry of solids, in Tarski 1983, pp.249.
Tarski, A. 1930a. ber einige fundamentale Begriffe der Metamathematik. Comptes
Rendues des sances de la Socit des Sciences et des Lettres de Varsovie 23 iii:229.
Translated by Woodger as On some fundamental concepts of metamathematics, in
Tarski 1983, pp.307.
Tarski, A. 1930b. Fundamentale Begriffe der Methodologie der deduktiven
Wissenschaften I. Monatshefte fr Mathematik und Physik 37:361404. Page references
are to the translation by Woodger, as Fundamental concepts of the methodology of the
deductive sciences, in Tarski 1983, pp.60109.
Tarski, A. 1935. Der Wahrheitsbegriff in den formalisierten Sprachen. Studia Philosophica
I:261405. Page references are to the translation by Woodger, as The concept of truth in
formalized languages, in Tarski 1983, pp.152278.
Tarski, A. 1936. ber den Begriff der logischen Folgerung. Actes du Congrs International
de Philosophie Scientifique 7:111. Page references are to the translation by Woodger, as
On the concept of logical consequence, in Tarski 1983, pp.40920.
Tarski, A. 1938. Der Aussagenkalkl und die Topologie. Fundamenta Mathematicae
31:10334. Translated by Woodger as Sentential calculus and topology in Tarski 1983,
pp.42154.
 References 335

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

Wittgenstein, L. J. J. 1975. Wittgensteins Lectures on the Foundations of Mathematics,

Cambridge, 1939, ed. C. Diamond. Ithaca, NY:Cornell University Press.
Wittgenstein, L. J. J. 1978. Remarks on the Foundations of Mathematics, 3rd edition,
ed. G. H.von Wright etal., trans. G. E.M. Anscombe. Oxford:Blackwell.
Woodin, W.H. 2001. The Continuum Hypothesis, parts Iand II. Notices of the American
Mathematical Society 48:56776 and 68190.
Woodin, W.H. 2005. The continuum hypothesis. In R. Cori etal., eds., Logic Colloquium
2000:Lecture Notes in Logic 19. Urbana, IL:Association for Symbolic Logic, pp. 14397.
Wright, C.J. G. 1987. Further reflections on the sorites paradox. Philosophical Topics
15:22790.
Wright, C.J. G. 1992. Is higher-order vagueness coherent? Analysis 52:12939.
Wright, C.J. G. 2001. On being in a quandary. Mind 110:4598.
Wright, C.J. G. 2007. Wangs Paradox. In Auxier and Hahn, eds., 2007, pp. 41544.
Zermelo, E. F. F. 1908. Untersuchungen ber die Grundlagen der Megenlehre I.
Mathematische Annalen 65: 26181. Translated by Stefan Bauer-Mengelberg in van
Heijenoort 1967, pp.199215.
Zermelo, E.F. F. 1930. ber Grenzzahlen und Mengenbereiche:Neue Untersuchungen
ber die Grundlagen der Megenlehre. Fundamenta Mathematicae 16:2947. Page ref-
erences are to the translation by Michael Hallett, On boundary numbers and domains
of sets: new investigations in the foundations of set theory, in Ewald, ed., 1996,
pp.121933.
Index

a priori knowability, 21, 813,867 bisection implication,4750

absolute formula,294 bisection of a set of statements, 4750,75
absolute necessity, 76,182 Bivalence, Principle of, 11, 12, 13, 22, 23, 267, 956,
absurdity, see impossible state of affairs 10911, 118, 14851, 156, 184, 1967,218, 261,
accessibility relations (between possibilities), 269, 270,30217
75,166 Boghossian,P.,4
Adams, D.,146 Bohr, N.,147
Adams, E., 79,181 Born, M.,171
admissible sets, 295 Bostock,D.,91
ancestral logic,2067, 267 boundaryless predicates,26, 23641, 2439, 25862
Anderson, A.R.,192 Bounded Omniscience Schema (in
Andjelkovi, M.,308 semi-constructive set theory),2978
Anscombe,E.,16 bounded quantifiers,293
anti-realist semantics for the language of set Bradwardine, T., 305,307
theory,27981 Braithwaite, R.,101
anti-warrants for statements,1378 Brandom, R.,113
applications of mathematics,300 Brouwer, L.E. J., 128, 210,316
Aristotle, 37, 70, 71, 73, 210,320 challenge to classical logic when reasoning
argument for Bivalence, 267,3023 about the infinite, 25, 197202,20810
Aristotles Thesis, 57, 6970, 78, 91, 157, 181, 193, Bub, J.,168
229,319 Burali-Forti Paradox,275
arguments for, 70,748 Burgess, J., 76,121
Artemov, S.,128 Buridan,J.,69
Avigad, J.,297
Axiom of Choice, 32, 298,310 (C), semantic axiom for sentential conjunction,
evidentialist form, 134
(B), postulate that any statement has a back, 25, exclusionary form, 119
1957, 21516,21718, in truth-grounds semantics, 155, 163, 174
see also Wittgenstein, L. (Con), semantic axiom for predicate
back (of a statement), 25,194f. conjunction, 246
backless property,21516 canonical assertibility conditions, 6,1301
Bain, A.,101 Cantor, G., 298, 302,309
Barcan Formula, 105,15960 see also Continuum Hypothesis and
Barrow, I.,210 Generalized Continuum Hypothesis
Barwise, J., 161, 294,295 Carnap,R.,76
Basic Revisionary Argument (Wright),1267 Cartwright,R.,42
bastard intuitionism,2834 categorical axiomatizations,
Beall, J.C.,56 and bivalence, 2689
Bell, J., 177, 210,213 of arithmetic, 2068,2678
Belnap, N.,192 of set theory,2705
Benacerraf, P., 26970,275 see also quasi-categoricity
Berkeley, G.,210 categorical theories, 2678,314
Berker, S.,146 category theory,2745
Berlin, B.,236 challenges to classicallogic
Bernays, P.,316 dubious grounds challenge, 23, 25,
Beth, E.W.,910 14752,18493
Beth trees, 10, 1445,1645 Dummetts challenge in Truth, 95111,1224
bilateral formalizations of logical rules,6,53 evidentialist challenge,13847
bipolarity of statements, 25,2023 from Smooth Infinitesimal Analysis,21016
338 Index

challenges to classicallogic (Cont.) 0-LEM,296f.

Putnams challenge,16781 0-MP,296f.
strong verificationist challenge,1259 0-Separation, see Separation
weak verificationist challenge,14950 (D), semantic axiom for sentential
when reasoning about all sets, 270ff. disjunction,
when reasoning about infinite structures, evidentialist form, 135
197210 exclusionary form, 11819
Chisholm, R.,102 in truth-grounds semantics, 164, 178
Church,A.,14 (Det), postulate of the determinacy of sense,
Clark, P.,264 305f.
classicallogic, (Dis), semantic axiom for predicate
default status of,1415 disjunction, 247
see also challenges to classicallogic
Classical Reductio, rule of, see Reductio Davey, B.,162
classical semantics, 1011, 27, 199, 273, 276,31720 Davies, M., 92,155
see also Bivalence, Principleof decimal expansion of ,2005
closed set, 24, 135,163 Dedekind, R., 267,298
closure, deduction,358
of a set of possibilities, 116, 135, 1623,178 use of logic in,8890
of a set of possible warrants,1356 Deduction Principle, 5965,82
Cohen, P., 2823, 310,313 deductive capacities,
coherence (of a logic and a semantic account of and logical capabilities,523
consequence),122 and logical consequence,3842
Collection schema, 290, 291,293 definable set,231
0-Collection, 294, 295,297 De Morgan Laws, 98, 161,178
colon (of sequent calculus),501 DescartesR.,1
combination (of possibilities), 1667, 1823,187 descriptive names,87,90
comparative similarity relations, 239, 241,243 determinate versus indeterminate
compossibility,1667 statements,3057
conceptual truth,68 determination, see possibilities
conditional assertion,501 dialetheism, 13, 186,264
Conditional Proof, ruleof,17 Dickie,I.,6
conditionals, Dilemma, Law of, 15, 38, 40,534
logic of, 1517, 545, 667, 72,183 disjunction,
semantic principles for, 99, 1202, 128, 183, logic of, 12, 524, 139, 1401, 142, 16881,
192, 217,280 1813
see also (I) semantic principles for, 245, 11819, 128, 130,
conjunction, 131, 1345, 142, 148, 156, 1601, 164, 165,
semantic principles for, 11819, 128, 134, 148, 155, 167, 176, 178, 184, 192, 217, 247, 252, 279,
163, 165, 167, 174, 184, 192, 217, 246, 252,279 2867,300
see also (C) and (Con) see also Law of Distribution, (D), and (Dis)
consequence, 314,40 disjunction property (of theories), 1312,
relata of consequence relations,323 2867
versus deducibility,83, 2056 dispersion (of an observable),172f.
see also logical consequence, Philonian Distribution, Law of (alias Distributive Law), 9,
consequence 1112, 25, 139, 1489,184
Constancy Principle (in SIA),21213 attacks on, 1401, 147,16881
Continuum Hypothesis, 149, 184, 2824,30917 justification of,1813
arguments for bivalence of, 30910, 311, 31316 DNE condition,1935
see also Generalized Continuum Dogramaci,S.,4
Hypothesis Double Negation Elimination (DNE), rule of,
Copi, I.M.,39,40 3, 5, 8, 11, 12, 121, 1223, 13940, 1467,
Currys Paradox,78 150,189, 1934, 195, 214,249
Cut Law,426, and vagueness, 224
Dretske, F., 5960, 64,104
0 Formula, 293, 294,295 dubious grounds challenge, see challenges to
0-Collection, see Collection classicallogic

Dummett, M., 6, 16, 22, 23, 57, 72, 92, 121, 165,
196,225
argument for exclusionary theory of
content,99101
argument for verificationist theory of
meaning,1257
assertoric content versus ingredient
sense,284
attacks on classical logic, 95110,1224
on bivalence, 223, 95, 109111
on harmony and stability of introduction
and elimination rules,45,8
on indefinite extensibility,263
on infinity, 1989,210
on justifying logical laws, 23,910
on logical truths as by-products of logical
consequence,56,72
on mathematical versus concrete objects,2646
on meanings of connectives, 1920,217
on operational interpretations of quantum
theory,176
on semantic shift principles,1267
on Tertium non datur, 223, 958,10511
on the significance of Gdels First
Incompleteness Theorem, 145, 263,266
on truth,317
on vagueness, 2257, 234, 2478,261
on whether logic is empirical, 142,1767
Dunn, J.,167
Eadem est scientia oppositorum, principle of,
196, 202, 203,237
Edgington, D., 823, 867, 8892, 100, 240,261
Egan, A.,135
elementary set,2313
eliminationrules,
for connectives generally,4
for negation,3,5
Eliot, T.S., 317, 318,319
end-extension,294
entailment (versus deducibility), 83, 2056
enthymemes,3942
Etchemendy,J.,73
Euclid, 266,311
Evans, G., 86, 87, 92, 218,240
evidential probability,138
evidentialist semantics,1308
see also challenges to classicallogic
Ex Contradictione Falsum, rule of, 190,191
Ex Contradictione Quodlibet, rule of, 118,189
Ex Falso Quodlibet, rule of,190
Excluded Middle, Law of, 44, 81, 118, 126, 128,
1489, 162, 195, 1978, 2006, 223, 2345,
249, 250, 270, 281, 285, 290, 292, 296,298
exclusion, of a possibility by a statement,114f.
exclusionary account of content, 23, 97,
99105,11124
Index 339
choice of logic within an exclusionary
semantics,1224
difficulties with disjunction, 1001,1034
employing spaces of possibilities,11415
exclusionary content,
and assertion,105
and information theory,104
coarse versus fine-grained,117
of conditionals,121
of other sentential connectives,11820
existence property (of theories), 1312,2867
expectation value (of an observable),172f.
Extensionality, Axiom of, 290, 291, 292, 294,
296,297
exterior of a set of possibilities,114
(F), semantic principle for falsum, 190
Feferman, S., 297, 298, 299,316
Fermats Last Theorem,82
Ferreira,F.,6
Field, H., 78, 83,207
Fine, K., 77, 83, 160, 237, 283,284
Finkelstein, D.,176
First Isomorphism Theorem (Zermelo),2712
first-order logic, 14, 76, 205, 253, 267, 285
Fitting, M., 2824,287
foil (for a vague predicate),2412
forcing,279
Foundation, Axiom of, 281,289
Foulis, D.,178
Fraenkel, A.,314
Frege,G.,
against the formalists,299
axiomatic formalization of logic,71
definition of ancestral,206
definition of finite cardinal,207
elucidations of basic logical notions,72
function/argument analysis of statements,18
on functions as unsaturated,271
on how deductions can be informative,57
on inference,37
on Law of Excluded Middle, 223, 2345,249
on logical laws, 12,18
on the identity of sets,2401
on vagueness, 2223, 225, 234,235
Friedman, H., 1312, 2901, 293,297
functional conception of modal
entities,15860
Gallistel, C.,239
Grdenfors, P.,242
Gardner, M.,168
Geach, P.T., 21, 102,241
Generalized Continuum Hypothesis, 27, 196,
302,315
Gentzen, G., 26, 50, 51, 53, 56, 67,289
see also Gdel-Gentzen translation
340 Index

Gibbard,P.,6 implicative relations, 4651, 748,84

Glivenko, V.I.,289 impossible state of affairs (),18792
Gdel, K., 20, 26, 35, 132, 145, 205, 2767, 310, inaccessible cardinals, 269, 2712,277
311,315 incompatibility
Axiom of Constructibility, 278, 282,315 between a statement and a state of affairs,194
First Incompleteness Theorem, 145, 205, of possibilities, 167,177
263,266 incompleteness theorems, seeGdel
negative translation, 2879,2912 Incurvati, L., 127,264
Second Incompleteness Theorem, 35,132 indefinite extensibility, 2634
Gdel-Dummett logic (LC),282 indeterminate identity, 21416,218
Gdel-Gentzen translation, see negative indeterminate truth-value, 25860,309
translations inference,346
Goldbachs Conjecture, 13940, 1467, rules of, 23,345
2089,311 Russells accountof,71
Goldblatt, R., 13, 141, 167, 178, 189, 192,274 infinitesimals, 25, 21016,
Goodman, N.,290 SIA theory of,21216
Graff Fara, D.,220 infinity, 197202, 267,31617
Grice, H.P.,15 Infinity, Axiom of, 290,291
Grayson, R.,290 information theory,104
interior (of a set), 114,231
Halmos, P., 120,249 introductionrules,
Hardegree, G.,175 for connectives generally,4
harmony (between introduction and for negation, 141, 190 (see also Simple
elimination rules), 5, 68,21819 Reductio)
Hawthorne,J.,59 intuitionistic logic, 44, 96, 98, 1234,
Hazen, A.,2812 1501, 1612, 163, 164, 184,
Heck, R.G., 240,264 18993, 1935, 200202, 213,218,
Heijenoort, J.van,289 2237,22834
Heisenberg, W.,147 and senses of connectives, 193
Heisenberg Uncertainty Principle, 1413, 169, and infinity, 197f., 207
170, 173,180 higher-order,2
Hellman, G.,21415 incoherence with exclusionary
Herbrand, J.,147 semantics,123
Herbrands Theorem,14 not necessarily a manifestation of
Hertz,P.,42 insanity,2
Heyting, A., 10, 23,128 prenex normal form and,14
Heyting Arithmetic, 131, 132, 198,287 semantics for, see Beth trees, Heyting
Heyting semantics for intuitionistic logic, semantics, Kripke semantics for
10, 12833, 140, 151, 202, 208, 209, intuitionistic logic, topological
2257,311 semantics for intuitionistic
higher-order logics, 2, 401, 83, 205f., 267 logic
higher-order vagueness, 262 Isaacson, D., 315,316
Hilbert, D., 71, 72,285 iterative conception of set,27685
Hilbert space, 171, 175,17980
Hi,H.,14 Jackson,F.,37
homophonic semantic principles,3,9 Johansson, I.,163
Hughes, G.E.,69 Julius (Evanss example), 87, 90,912
Humberstone, I.L., 6, 92, 120, 155,1612
KP and KP , see Kripke-Platek settheory
(I), semantic axiom for conditional, Kalmr, L.,316
exclusionary form, 121 Kant, I.,210
truth-grounds form, 183 Kay, P.,236
IKP and IKP , see Kripke-Platek settheory Kleene, S.,289
implication relations,21 Kneale, W.andM.,44
and possibilities, 4651,89 knowledge by deduction,5665
structural features of,424 Kochen, S.,170
varieties of, 3846,8892 Kock, A., 21011,213
 Index 341

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

negation (Cont.) platonism, 265, 299, 318

semantic principles for, 3, 11, 119, 128, 137, 148, Poincar, H.,210
157, 1602, 167, 177, 184, 185f., 192, 217, polar predicates, 227,2357
246, 252, 279,300 semantic theory for,2429
see also Double Negation Elimination, (N), poles (alias paradigms), 236,23941
and (Neg) colour poles, 236,238f.
negative translations,26 true/false poles, 254, 2589,309
Gdel, 2879,2912 Polkowski, L.,235
Gdel-Gentzen, 2889, 292,2969 possible worlds,1537
Kolmogorov-Glivenko, 289,292 and bivalence, 193
Kuroda,289 and disputes over logic,158
Negri,S.,7 functional explanation of, 158-9
Neptune (Kripkes example),8692 truth at,1567
Newton, I.,210 possibilities, 112,153f.
Newtons Laws of Motion, 42, 71, 74,801 accessibility relations between, 75,166
nilsquare infinitesimals, see infinitesimals closure of a set of, 24, 116,1623
Non-Contradiction, Principle of, 11, 1856,189 determination of, 11314,166
non-homophonic semantic theories,914 epistemic, 57,1356
non-standard models, exclusion of,11117
of arithmetic, 207 exterior of a set of,114
of set theory, 285 implications and,4950
normalization of proofs,7 incompatibility between, 167,188
Numerical Omniscience Schema (of interior of a set of,114
semi-constructive arithmetic),209 logical, 153, 1627,1813
metaphysical,112-14
observables (in quantum theory), 172f.,180 open sets of, 11415,120
Oliver, A.,264 physical, 49, 74, 76,168f.
open sets of possible objects,22835 regular space of, 182,187
open extension (of a vague predicate),230, 261 spaceof,46
open sets of possible states of affairs, 11415, truth and falsehood at, 115, 155, 162f.,187
120,228 truth of disjunctions at, 156,161
ordinals,275 Post-completeness,14
orthocomplement, 141, 167, 185f.,301 Power Set Axiom, 290, 291, 2923,296
Orowska, E., 231,235 power set of ,31617
power set operation, 276,277
(P) (pragmatist principle), see pragmatism Powell, W., 263,2912
paradigms, see poles pragmatism
Paradox of Sharp Boundaries,222 pragmatist account of mathematics, 299301
Paradox of the Heap, 22, 26, 2201,25055 principle (P) and pragmatist theory of
structural version of, 456,254 content,1024
Paradox of the Liar, 22, 304,305 Prawitz, D., 68,18990
Paradox of the Preface, 601, 64,134 preclusion, of one possibility by another,111f.
Parsons,C.,76 Price,H.,6
Parsons, T.,218 Priest, G., 16,17,56
Paseau, A., 264, 280,282 Priestley, H.,162
Pawlak, Z., 231,235 Prior, A.N., 22,304
Peano, G.,223 proof-theoretic justifications of logical
Peano Arithmetic, 40, 206,288 laws,48
Peano-Dedekind axioms,2678 product, of topological spaces,256
Perry, J.,161 Projective Determinacy, Axiom of,310
petitio principii,578 Proto-Indo European, see Kurgan Hypothesis
Philo of Megara,44,70 Putnam, H., 25, 1412, 224, 26970,299
Philonian consequence, 446, 47, 48, 70, 71, 73, attack on Law of Distribution in quantum
77,254 mechanics,16781
Platek, R.,294
see also Kripke-Platek set theory. QO-space topology,114
Plato, 37, 262,303 quantification over all sets, 26, 210,
Plato, J.von,7 2789,2845, 2923, 295, 30911
 Index 343

quantification over infinite domains, on boundaryless predicates, 2368, 240, 245,

199,20810 258,261
quantum logic, 9, 1112, 1413,164 Sambin, G., 182, 183,191
anti-realist argument for, 1413,147 satisfaction,72
Putnams argument for,17681 satisfaction conditions for colour predicates,
quantum theory,13,25 23940, 241, 243,246f.
verificationist interpretations of, 1423,147 saying (versus asserting),21
quasi-categoricity (of second-order set theory), Scambler, C.,314
269, 2713,285, 31415 Schrder, E.,240
Quine, W.V.,42 Schrdingers Equation,171
contextual factors in arguments,33 Schulte, J.,203
denying the doctrine and changing the Schulz, S.,161
subject, 19,217 Scott, D., 47, 50, 51,278
doubts about clarity of pre-theoretic notion see also Lindenbaum-Scott Theorem
of logical consequence,41 second-order logic, 2, 401, 83, 2068, 267, 269,
on indispensability of mathematics,299 2702, 285,31516
on natural kinds,2412 type-theoretic interpretation of, 2712
on second-order logic,271 Seldin,J.,7
on truth-conditions for disjunctions,248 semi-constructive theory of sets (SCS)
(Feferman),2979
(R), principle that the truth-grounds of any sense of an expression, 3,92,97
statement form a closed set, 136, 1634, sentential quantification,304
185,21718 Separation schema, 290, 291,293,
Raatikainen, P., 127,145 0-Separation, 294, 295,297
Radford,C.,54 set, iterative conception of, 240,27685
Raffman, D., 255,261 set, logical notion of,2401
Ramsey, F.P., 1023, 11213, 1212,196 Sextus Empiricus,44
definitions of truth and falsehood,304 Shapiro, S., 40, 2068,264
Randall, C.,178 Shieh, S.,109
Rasiowa,H.,20 Siders,A.,7
Read, S., 79, 224,305 Sikorski,R.,20
Reductio, rulesof, Simple Argument for Bivalence,3045
Classical,48 Skolem, T., 207,314
Simple, 5, 15,190 Skolemized forms,14
Refinement (Humberstones Axiom of),1612 Skorupski,J.,37
reflexivity of implication,42,44 Smiley, T.J., 6, 22, 40, 42, 70,73,77
regular open sets, 120, 2459,2557 Smooth Infinitesimal Analysis (SIA), 256,
regular space of possibilities, see possibilities 212ff.
Regularity, Postulate of, 1823, 187, 189,191 differences from constructive analysis,
relative necessity,76 214-15
relevance logics,13 semantic theory for, 215-17
relevance logicR,192 Socrates,37
reliabilism,61 Solovay, R.,310
Renfrew, A.,146 Sommers,F.,18
Replacement, Axiom of, 271, 272,290 Sorites, see Paradox of theHeap
Restall, G.,51,56 soundness, 3, 11,1489
Revised Argument for Bivalence,3068 of Double Negation Elimination,195
Reyes, G.,213 of Ex Contradictione Falsum,191
Robinson, A.,210 of Ex Falso Quodlibet,190
Rosch, E.,241 of Simple Reductio,190
rough sets, 231,235 Specker, E.,170
rules of inference and of deduction,348 Spinoza,B.,1
rule-circularity, 24,11 stability (of introduction and elimination
Russell, B., 21, 37, 70, 71, 72, 73,2789 rules),5
Ryle, G.,347 stability (of semantic theories under changes in
metalogic), 913, 18991,1957
Safety Principle,625 Stlmarck,G.,7
Sainsbury, R.M., 26, 63,262 Stalnaker, R., 105, 117, 1535, 158, 159,161
344 Index

state of affairs,187f. truth-preserving relations among statements,

statements,201 44, 46, 47, 50, 5962,69,78
as relata of consequence relations,323 truth principle, for statements of quantum
Statman,R.,7 mechanics, 173, 174, 179,180
Steinberger,F.,51 truth-tables, semantics
Stevenson, C.L.,16 implicit in, 23,1478
Strawson, P.F.,106 type-theoretic interpretation
Strong Conjecture (Woodin), 310 of quantifiers, see second-order logic
strong verificationism, 23,12547
arguments for,1257 Union, Axiom of, 290,294
attack on classical logic, 1259,13843 universe of sets, 276, 277, 278,2845
optimal version of,12938 Unordered Pair, Axiom of, 290, 291,294
problems for,1438 unworldly theories of modality,1607
Sullivan, P.,264 Urelemente,2735
supervaluational theories of vagueness, usefulness of logic,53
2378,319 Uzquiano, G., 274,314
suppositions,
arguments from, 71,7981 vagueness, 26,22062
indicative versus counterfactual, and Distribution,2559
7980,85,87 Dummett on, 2257, 234,2478
Synthetic Differential Geometry,21011 Edgington on,261
Frege on, 2223,235
Tait, W., 26, 264, 271, 272, 2768, 279, 280, 289, higher-order, 262
293,298 in respect of identity,240
Tarski, A., 20, 42, 67, 120,252 Raffman on, 255,261
on logical consequence,723 Sainsbury on, 2368, 240, 258, 261,262
on regular open sets of a topology, 120, supervalational theories of, 2378,259
2489 Williamson on, 226, 237,25960
topological semantics for intuitionistic Wright on,22335
logic,22930 validity, Russells accountof,70
Tarskian properties of implication relations, Van Fraassen, B., 76,138
42f., 47f.,75 Varzi, A.,237
Tennant,N.,5 Velleman, D.,264
Tertium non datur, Principle of, 22, 958, verificationism, see strong verificationism and
10811,1478 weak verificationist challenge
Williamsons argument for,25960 verificationist semantic theories,12,13
Tharp, L.,263 Voronoi tessellation,242
Thom, R.,210
tolerance relation,2313 Walmsley, J.,273
topological semantics for intuitionistic warrants for statements, 6,1298,
logic,22930 preservation of warrant versus preservation
transfinite induction on , 290,291 of truth,1435
Troesltra, A.,132 see also anti-warrants
truth and falsehood,1257 wave function,171f.
anticipatory (Fine),284 Weak Law of Excluded Middle,282
at a possible world,1567 weak verificationist challenge, see challenges
at a possibility, 115, 155, 162f., 181,187 to classicallogic
for mathematical statements,268 Weatherson, B.,135
in a model,73 Weston, T., 314,316
of a statement, 201,33 Weyl, H., 210,299
Ramseys definitions of,304 White, A.,35,36
revised to accommodate indeterminate Whitehead, A.N.,70
statements,306 Wiggins, D., 90,240
under an exclusionary theory of content, Williamson, T., 80, 85, 133, 138, 144, 146, 226,
107,11112 237, 304,308
truth-grounds semantics, 24, 11422,1634 argument for Tertium non datur,25960
 Index 345

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