Russell’s Logical Atomism

Drawing of Bertrand Russell by Augustus John, c. 1913. Image courtesy of William Ready
Division of Archives and Research Collections, McMaster University Library, Hamilton,
Canada # Bridgeman Art Library
Russell’s Logical
Atomism

David Bostock

1
3
Great Clarendon Street, Oxford OX2 6DP
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York
# David Bostock 2012
The moral rights of the author have been asserted
Database right Oxford University Press (maker)
First published 2012
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, 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 book in any other binding or cover
and you must impose the same condition on any acquirer
British Library Cataloguing in Publication Data
Data available
Library of Congress Cataloging in Publication Data
Data available
Typeset by SPI Publisher Services, Pondicherry, India
Printed in Great Britain
on acid-free paper by
MPG Boooks Group

ISBN 978–0–19–965144–3

1 3 5 7 9 10 8 6 4 2
Contents

Preface vi
Abbreviations ix

Part I: Logic and Mathematics

1. Elementary Logic 5
2. The Contradiction (i): The Problem 16
3. On Denoting 31
4. The Contradiction (ii): A Simple Solution 47
5. The Contradiction (iii): A Ramified Solution 74
6. Problems 97

Part II: Knowledge

7. Acquaintance 113
8. Knowledge and its Foundations 135
9. Logical Constructions (i): Matter 151
10. Logical Constructions (ii): Minds 169
11. The Demise of Neutral Monism 190

Part III: Metaphysics

12. Propositions 201
13. Universals 236
14. Facts 252

Epilogue 279
Concordance: Reprints of Journal Articles 284
References 286
Index 297
Preface

Bertrand Russell would sometimes describe his own basic philosophy as ‘the philoso-
phy of logical atomism’. He first used this description in a little-known lecture that he
delivered in Paris in March 1911, entitled ‘Analytic Realism’, which is a different title
that he also accepted at that time as describing his basic position.1 It was a ‘realism’
insofar as it accepted that there are many things which exist quite independently of
human minds, and was ‘analytic’ because it was concerned with the analysis of these
things. The things in question Russell regarded as ‘complexes’ or ‘facts’, and the
analysis aimed to discover the simple ingredients of these facts, which count as
‘atoms’ because they are simple, and not open to any further analysis. They are called
‘logical atoms’ because they are to be reached by a process which Russell always called
‘logical analysis’, though it is not quite clear why it deserves the adjective ‘logical’. As
we might say, it is counted as logical because it is in the first place an analysis of
propositions, and logic is traditionally a study of propositions. But Russell would rather
say that propositions are used to state facts, and the analysis that he is concerned with is
really an analysis of facts, so that the logical atoms that it aims to reveal are the simple
ingredients of facts. That is to say, they are what the world is ultimately made of. That is
a brisk and bald outline of what he has in mind when he speaks of ‘logical atomism’.
This book aims to fill in many of the details.
Russell described his position as ‘the philosophy of logical atomism’ not only in that
lecture of 1911, but also in two better known works of 1914, namely in the program-
matic article ‘On Scientific Method in Philosophy’, and in the preface to his book Our
Knowledge of the External World.2 Then it is used as the title of his set of lectures on The
Philosophy of Logical Atomism of 1918; and finally it occurs once more in his own
contribution to a collection of essays by various contemporary philosophers in 1924,
which he called simply ‘Logical Atomism’. So one might say that Russell saw himself as
believing in ‘logical atomism’ during the years 1911–24. Looking back on his life from
the much later perspective of My Philosophical Development [1959], Russell actually said
There is one major division in my philosophical work: in the years 1899–1900 I adopted the
philosophy of logical atomism and the technique of Peano in mathematical logic. . . . The change
in these years was a revolution; subsequent changes have been of the nature of an evolution.
(p. 11)

Certainly, these years in Russell’s life saw an important change in his philosophical
attitude, and one might well say that they were the beginning of his development

1
Russell [1911]. The phrase ‘logical atomism’ is used on p. 135.
2
Russell [1914d: 111, 1914c: 14].
 P R E FAC E vii

towards what is now regarded as his ‘philosophy of logical atomism’. But the doctrine
did take some time to evolve. Similarly, it did not end abruptly in 1924, and his later
book The Analysis of Matter [1927a] may fairly be counted as yet a further development
of this philosophical attitude. But we may reasonably take the years 1911–24 as central.
Meanwhile, let us go back to the beginnings.
The emphasis on logic stems from Russell’s first brush with modern logic, which was
with Peano in 1900. This subject was of central importance in his analysis of mathe-
matics, and he never lost his enthusiasm for it. But in later years he employed it in other
areas too. His atomism might be traced back even earlier, to his revolt from the Hegelian
‘monism’ that was so widespread when he was an undergraduate at Cambridge.
Strongly influenced by G. E. Moore he came to reject this monism in about 1898,
and his first important publication as a ‘pluralist’ was his book A Critical Exposition of the
Philosophy of Leibniz [1900]. Of course, if a pluralist is simply one who believes that
there really are many things, and not just one, then a pluralist is not as such committed
to any form of atomism. But it is quite a natural development to suppose that the ‘many
things’ are ultimately composed of ‘many atoms’, and then it is natural to seek for
these ‘atoms’. Much of the progress in Russell’s thought, during our period, was in
his successive discoveries that this or that could be analysed, and so did not have to be
accepted as an atom. The position that he ended with was extremely economical, as we
shall see.

The book is divided into three main parts, namely on Russell’s views: I on logic and
mathematics, II on knowledge, and III on metaphysics. These parts are roughly
chronological. Russell worked mainly on logic from 1900 to 1910; he turned his
attention to the theory of knowledge in many writings from 1910 on; it is always
intertwined with his metaphysics, which has its centrepiece in the lectures on The
Philosophy of Logical Atomism of 1918. These lectures bring together many of the themes
developed independently in earlier writings. But there is one development, which
should surely be counted as part of Russell’s logical atomism, but which belongs to the
years after 1918, namely the views that he puts forward in his book The Analysis of Mind
[1921a]. This work is treated at the end of part II, where it fits very naturally, although
the chronological order is thereby disrupted.
The only other philosopher who is ever called a logical atomist is Ludwig Wittgen-
stein, in his early period, which culminated in his Tractatus Logico-Philosophicus of 1921.
Wittgenstein became a pupil of Russell’s in October 1911, and after only a year Russell
was thinking of him as an equal, and hoping that he might be his successor at Cam-
bridge. It is clear that they saw a lot of one another, and talked a great deal. But in
October 1913 Wittgenstein left Cambridge, in order to work alone (in Norway), and
he and Russell did not meet again until after the war, in December 1919. By that time
Wittgenstein had written his Tractatus, and Russell agreed to write an introduction for
it. We shall never know how much of Russell’s thought, subsequent to their meeting
in October 1911, comes from his talks with Wittgenstein. In his writings Russell very
viii PR EFAC E

clearly acknowledges a debt,3 but in my own opinion he exaggerates it. Some, and
perhaps most, of what he credits to Wittgenstein clearly has its seeds in his own
thought, from times before they met. But on several points all that we can say is that
we have no direct evidence, either way, on who influenced whom.
Wittgenstein’s version of logical atomism is clearly rather different from Russell’s,
but that is a topic for another book, and there has not been the space to include it in this
one. I regard it as a possible project for the future.

Some of the material in my Part I, especially in Chapter 4, is taken from an article of

mine entitled ‘Russell on “the” in the plural’. This was delivered at a conference at
McMaster University in 2005, and later published in the collection Russell vs. Meinong:
The Legacy of On Denoting, edited by Nicholas Griffin and Dale Jacquette, and
published by Routledge in 2008. I thank the editors and the publisher for permission
to reuse this material.
I should like to acknowledge help with this book from Peter Sullivan, who gave me
an extended criticism of all of Part I, and from two anonymous readers for Oxford
University Press, who have done the same for the whole book, and especially for
chapters in Parts II and III. The book has surely benefited from all of their suggestions.
In a quite different way I should like to acknowledge help from what I shall always
think of as ‘The Philosophy Library’ of Oxford University. (It is now officially a part of
the overarching Bodleian Library of this University, but for much of the time that
I have known it, it was an independent institution.) The staff of this library have
continually been helpful to me in all kinds of ways, and have very much facilitated my
efforts to write. Finally, I should like to acknowledge help from Merton College,
particularly in the person of Judith Kirby, who has done all my secretarial work for me
for many years. I do not know how I could have managed without her; she has been
absolutely essential.
David Bostock
Merton College, Oxford
July 2011

3
Russell [1914c: 9, 1918b: 177].
Abbreviations

References are generally given in an abbreviated form, by citing just the author and the date of
publication, e.g. ‘Strawson [1950]’. Full details are then given in the list of references at the end of
the book. In some cases I have used writings by Russell which were not published at the time,
and these are given by the date of their composition. But some important works by Russell,
which are frequently cited, are also referred to by their abbreviated titles, as follows:

PoM The Principles of Mathematics [1903]

OD ‘On Denoting’ [1905b]
ML ‘Mathematical Logic as based on the Theory of Types’ [1908]
PM Principia Mathematica, with A. N. Whitehead, vol. 1 [1910d]
KAD ‘Knowledge by Acquaintance and Knowledge by Description’ [1910/11]
PP The Problems of Philosophy [1912a]
RUP ‘On the Relation of Universals and Particulars’ [1911/12]
NA ‘On the Nature of Acquaintance’ [1914a]
RSP ‘The Relation of Sense-Data to Physics’ [1914b]
KEW Our Knowledge of the External World [1914c]
PLA ‘The Philosophy of Logical Atomism’ [1918b]
IMP Introduction to Mathematical Philosophy [1919a]
Propns ‘On Propositions: What they Are and How they Mean’ [1919b]
AMind The Analysis of Mind [1921a]
LA ‘Logical Atomism’ [1924]
AMatter The Analysis of Matter [1927a]
MPD My Philosophical Development [1959]

I add one work by Wittgenstein, which is also frequently cited:

TLP Tractatus Logico-Philosophicus [1921]
This page intentionally left blank
 PART I
Logic and Mathematics

As an undergraduate at Cambridge (1890–4) Russell had studied first mathematics and

then philosophy. As he tells us when looking back on his life in the book My
Philosophical Development [1959], an important reason for his change from the one
subject to the other was his search for certainty (p. 29). Some time earlier he had lost a
belief in the certainty of religious truth, and he then felt that if anything could be
certain it must be mathematics. However it soon became clear to his keen intellect that
the mathematics then taught at Cambridge contained a number of fallacies, and could
by no means be called ‘certain truth’. So he hoped that philosophy would be better.
At first he had embraced the idealist philosophy that was then prevalent at Cambridge,
and had been particularly interested in its application to science and mathematics. (His
doctoral dissertation was on the foundations of geometry.) But by 1898 he and his
younger contemporary G. E. Moore had come to think that idealism was a great
mistake, and had revolted against it. However he retained his previous interest in
mathematics, and in fact had already made a draft of what was to become his Principles of
Mathematics before the important meeting with Peano in 1900. He was once more
hoping that a revised view of mathematics would fulfil the yearning for certainty that
had never left him.
This was the background situation when, in 1900, he attended a mathematical
conference in Paris. There he was extremely impressed by the exactitude that
Giuseppe Peano and his school had brought to the subject, and at once read all of
Peano’s works. This is where he first met what we now think of as modern logic, for
Peano was a pioneer in this area. Russell soon mastered Peano’s logical system, and
made an important contribution to it (i.e. his paper of 1901 on ‘The Logic of
Relations’). He also felt that he had learnt from Peano how to overcome various
problems that had stood in the way of a proper understanding of the concept of
number, for example the problem of how and in what sense there could be such a
thing as a class (a collection?) of no members, and how a class (a collection?) of just
one member could itself be something different from that one member. As these
problems became clearer, he was led to think that the concept of a number could
after all be analysed in purely logical terms, and so to embrace what has since become
2 L O G I C A N D M AT H E M AT I C S

known as the ‘logicist’ view of the nature of mathematics. This can be summed up as
the two claims (a) that the concepts of mathematics can be analysed in terms of the
concepts of pure logic, and (b) that the truths of mathematics then turn out to be no
more than truths of logic. The aim of deducing mathematics lay behind almost all of
Russell’s work in logic, i.e. of the great bulk of all his efforts during the ten years
from his first introduction to Peano’s ideas in 1900 to the publication of Principia
Mathematica in 1910. Part I of this book is entirely concerned with the work that
received its final form in Principia Mathematica.
I add here that I shall regularly speak of Principia Mathematica as if it were all Russell’s
own work. This is in one way unfair, since the book was written in collaboration with
his former tutor, A. N. Whitehead, who is joint author of the final publication, and
whose collaboration dates back almost to the beginning of the project in 1900. But on
the other hand Russell himself says, in describing this work, that ‘broadly speaking,
Whitehead left the philosophical problems to me’,1 and that is what will matter for us.
With very few exceptions, we shall be concerned with the introductory portions of the
work, and it is clear that these were written just by Russell. He did indeed publish all
the important ideas separately, under his own name only.2
Chapter 1 will contain a brief exposition of what is nowadays called ‘elementary
logic’. I expect that most readers will be quite familiar with this topic already, and so
can skip all of the first section of this chapter, but I add that it was not familiar to
Russell, and he may certainly be counted as one of those who helped to develop it.
Chapter 2 turns to non-elementary matters, first giving some details on how Russell
hoped to deduce mathematics from logic, and then turning to the big problem that was
posed for this project by his discovery of what he always called ‘the contradiction’,
namely what we now call ‘Russell’s paradox’. The chapter describes Russell’s first
efforts to resolve this problem, and why he found them unsatisfactory. Chapter 3 is
then something of a digression, for it is concerned with the theory of definite descrip-
tions in his well-known article ‘On Denoting’ [1905], and this is (on the surface) just a
modification in the theory of how elementary logic should be applied to our ordinary
ways of thinking. But it is a modification with some deeper morals, and Russell himself
describes it as opening the way to the true solution to his problem. Certainly it does
have an importance in the development of that solution, even if it is mostly discussed
today in a different aspect. Then Chapters 4 and 5 describe how Russell reached his

1
This is how Russell described the work when looking back on it from the much later perspective of
My Philosophical Development [1959: 74]. There is a useful assessment of Whitehead’s contribution to the work
in Alasdair Urquhart’s introduction to volume 4 of Russell’s Collected Papers [i.e. Russell 1994], particularly
pp. xxxviii–xli.
2
Russell [1910a] is a French version of almost all of chapter 2 of the Introduction to Principia Mathematica,
together with some themes from chapter 3 of that Introduction. Much of the discussion is also anticipated in
Russell [1908].
 L O G I C A N D M AT H E M AT I C S 3

final theory, with Chapter 4 devoted to what is now called his ‘simple’ theory of types,
and Chapter 5 to the more complicated ‘ramified’ theory that he finally recommends.
In each case there is some discussion both of how Russell himself conceived this theory
and of what more recent developments have made of it. Finally Chapter 6 notes the
main problems that arise, especially for the desired deduction of mathematics, and
makes some suggestions on how they might perhaps be met.
This page intentionally left blank
 1
Elementary Logic

Russell’s meeting with Peano in 1900 was the spur to his developing what we now
think of as ‘modern logic’. Peano had introduced much of today’s notation (which
Russell took over from him), and had set out some basic principles, but done little to
systematize them. Russell began to reflect on ways of improving the basis of the system,
and parts of his Principles of Mathematics [1903], henceforth PoM, are concerned with
what is needed in logic by way of fundamental concepts, in terms of which all others
are defined, and with what he always called ‘primitive propositions’, which are the
basic premises or axioms from which all other truths of logic are to be deduced. Early in
1901 he discovered Frege’s work in the subject, which preceded Peano’s and was
certainly much better, though Russell did not at once realize that. (His first reaction to
Frege was like everyone else’s: the symbolism is entirely unfamiliar, and not at all easy
to read, so why bother?1) But when he did pay serious attention to Frege in 1902 he
began to see the value of the work. So this chapter begins with a brief account of the
‘elementary logic’ that is shared between Frege and Peano and Russell. It is a needed
background. Indeed, it is a needed background to almost all of today’s philosophy, and
I imagine that most readers will already be quite familiar with it. In that case, they may
conveniently skip the first section, which will only summarize what they already
know. But the second section may contain some new information, and the third
describes Russell’s own understanding of this elementary logic, which is not quite the
same as what is nowadays usual.

1 A modern account of elementary logic

Elementary logic, as now understood, contains two parts. First there is what is variously
called ‘propositional’ or ‘sentential’ logic, but which I think is better called the logic
of truthfunctors. Next there is what is often called ‘first-order predicate logic’, but which
I think is better called the logic of quantifiers (of the first level). Here is a brief outline
of what are now taken to be the central ideas.
Logic is concerned with truth and falsehood. (These are called the ‘truthvalues’, and
in classical logic—which is what we are here concerned with—it is assumed that these

1
‘I had seen his [i.e. Frege’s] Grundgesetze der Arithmetik, but, owing to the great difficulty of his symbolism,
I had failed to grasp its importance or to understand its contents’ (Russell, preface to PoM [1903: xvi]).
6 L O G I C A N D M AT H E M AT I C S

two are the only truthvalues.) What is true or false is called a ‘proposition’. Propositions
can be thought of as what a sentence or utterance may say, or what a person may
believe, or in other ways. (Russell’s thoughts on what propositions are will occupy us
later, in Chapter 12.) But so far as pure logic is concerned, a proposition is simply
whatever has a truthvalue: every proposition is either true or false, and no proposition
is both true and false. This book will use the letters ‘P ’, ‘Q’, ‘R’, . . . . as schematic letters
to stand in for any proposition.
Some propositions contain others as parts, with such words as ‘not’, ‘and’, ‘or’, ‘if ’
applied to the parts and (in many cases) joining them together. This book will use the
symbols ‘~’, ‘&’, ‘∨’, ‘!’ to abbreviate these words, or the ideas that they stand
for. We assume (in classical logic) that these symbols are truthfunctors, i.e. signs for
truthfunctions, which is to say that they may be explained by saying how they affect
the truthvalues of the propositions that contain them. To give the explanations in
English, they are:
‘~P ’ is true if and only if ‘P ’ is not true
‘P&Q’ is true if and only if both ‘P ’ and ‘Q’ are true
‘P∨Q’ is true if and only if either ‘P ’ or ‘Q’ or both are true
‘P!Q’ is true if and only if either ‘P ’ is false or ‘Q’ is true2
To these we may add a symbol for ‘if and only if ’, which is an ‘if ’ that works both ways,
and so is abbreviated to ‘$’. The definition is

‘P$Q’ is true if and only if both ‘P!Q’ and ‘Q!P ’ are true, i.e. if and
only if ‘P’ and ‘Q’ both have the same truthvalue.
These explanations may conveniently be summed up in ‘truthtables’, which use ‘T’ for
‘true’ and ‘F’ for ‘false’, and surely need no other explanation:

P ~P P Q P &Q P ∨Q P!Q P$Q

T F T T T T T T
F T T F F T F F
F T F T T F
F F F F T T

The five truthfunctors explained here are the ones that will be used in this book, and
are the ones that are now standard. Naturally one may if desired use other truthtables
to introduce other truthfunctors, but for ordinary purposes little is gained thereby.

2
This is the truthfunctional interpretation of ‘!’, i.e. of ‘if ’. The question of how well it represents the
ordinary English ‘if ’ is controversial. Of course the same question may be raised about the ordinary English
‘not’ and ‘and’ and ‘or’: is the truthfunctional interpretation adequate for these?
 E L E M E N TA RY L O G I C 7

I remark that Russell’s notation differs from what is here introduced, but that is a trivial
point of no importance (so long as you do also know how to read his notation).3 This
introduces the schematic language for the logic of truthfunctors. Let us now turn to the
quantifiers.
Some propositions are ‘subject-predicate’ propositions in the modern sense, i.e.
propositions in which some singular subject is referred to. What is said about
that subject then forms the predicate of the proposition.4 This book will use the
schematic letters ‘a’, ‘b’, ‘c’, . . . to stand in for any reference to a subject, and the letters
‘F ’, ‘G’, ‘H ’, . . . to stand in for what is predicated of a subject. By convention, we
write the predicate letter before the subject letter, as in ‘Fa’ or ‘Gb’. Some propositions
contain a reference to two or more subjects, and what is predicated of them is then
a relation of two or more places. We represent these by the schematic formulae ‘Fab’
or ‘Gabc’, and so on.5 These schematic formulae can then be combined into more
complex schemata by using the truthfunctors just explained. We can now introduce
the quantifiers.
These are expressed in English by ‘every’ and ‘some’, for which we use ‘8’ and ‘9’
respectively. The quantifier sign is always followed by what is called a ‘variable’, which
is then repeated in what follows it, in order to show just where the ‘every’ or the ‘some’
is applied. In a first-level logic, which is all that presently concerns us, the variable
will be one that takes the place of a subject letter, and for these we use the letters ‘x’,
‘y’, ‘z’, . . . Thus if we begin with a simple subject-predicate proposition ‘Fa’, we may
then put a variable, say ‘x’ in place of its subject letter ‘a’, and add at the front
a quantifier ‘8x’ or ‘9x’ which is said to ‘bind’ that variable. The result is ‘8xFx’
or ‘9xFx’ and these represent the propositions which claim that ‘Fa’ remains true for
everything in place of ‘a’, or that it is at least true for something in place of ‘a’. To give
a concrete example,
If ‘Fa’ is ‘Socrates is mortal’
then ‘8xFx’ is ‘Everything is mortal’
and ‘9xFx’ is ‘Something is mortal’
But in order to say that every man is mortal we need another predicate, say ‘G . . . ’, to
represent ‘ . . . is a man’. Then the sentence that we want is ‘8x(Gx!Fx)’. In a more

3
In place of ‘P &Q’, ‘P !Q’, and ‘P $Q’ Russell himself uses ‘P.Q’, ‘P ⊃Q’, ‘P Q’. Others have used
–
‘¬P ’ or ‘P ’ in place of ‘~P ’, and ‘P∧Q’ in place of ‘P &Q’. So far as I know, the use of ‘∨’ for ‘or’ is universal.
The symbol stands for the Latin ‘vel’, i.e. for the inclusive rather than the exclusive ‘or’. The symbol ‘∧’ is a
deliberately contrived contrast to ‘∨’. (I do not give Frege’s notation, which is very different.)
4
Contrast the traditional usage, which hails from Aristotle, by which a proposition of the form ‘All A’s are
B’s’ is counted as a subject-predicate proposition, even though ‘All A’s’ evidently does not refer to a singular
subject.
5
In place of ‘Fab’ Russell uses ‘aRb’, with ‘R’ to suggest a relation. So long as only two-place relations are
concerned, this is a very natural symbolism. It is usual to express particular two-place relations in this form,
e.g. ‘x¼y’, ‘x<y’, ‘x2y’. We may abbreviate by placing a slash through the relation-sign, as in ‘x6¼y’, to
express its negation.
8 L O G I C A N D M AT H E M AT I C S

complex schematic formula there may be several quantifiers. For example the formula
‘8x9yFxy’ says that for everything in place of ‘a’ the schema ‘Fab’ comes out true for at
least one thing in place of ‘b’. By contrast, the formula ‘9y8xFxy’ says that there is at
least one thing which you can put in place of ‘b’ that will make ‘Fab’ come out true for
everything whatever in place of ‘a’. You should think about this example for a bit, in
order to see clearly that there is a difference between these two formulae. (The first
does not imply the second, though the second does imply the first.) Generally,
however complicated the schema ‘(—x—)’ may be, ‘8x(—x—)’ may be read as
‘Everything is such that (—it—)’, and ‘9x(—x—)’ may be read as ‘Something is such
that (—it—)’. But where there are several quantifiers we need to know which of them
each occurrence of ‘it’ goes back to.6
I have to assume that this is an adequate explanation of the schematic language to be
used in setting out the logic of the truthfunctors and the quantifiers. I should also say
that in order to come to a good understanding of this language one needs practice in
translating between it and one’s ordinary language. This brings one to appreciate how
much of what we ordinarily say and think can be schematically represented in the
language. But I do not provide the material for such practice. There are any number of
books on elementary logic that do provide it. Instead I move on at once to what
elementary logic can do with this language.
Traditionally, logic is concerned with the validity of arguments, i.e. with what
follows from what. In logic as it is now conceived we are concerned with what follows
formally, where this is understood in terms of the formal language just introduced,
i.e. one which uses ‘P ’, ‘Q’, . . . as schematic letters for any propositions, ‘a’, ‘b’, . . . as
schematic letters for any reference to a singular subject, and ‘F’, ‘G’, . . . as schematic
letters for any predicate. So we first explain validity for such schemata. An interpretation
for the language assigns some particular propositions, or subjects, or predicates to the
schematic letters involved. It also assigns some domain for the quantifiers to range over,
i.e. some definite objects to be those of which ‘all’ or ‘some’ are concerned. Then a
single schematic formula counts as valid if it always comes out true, however its
schematic letters are interpreted, and whatever the domain of quantification is taken
to be.7 A series of such formulae representing an argument, with several formulae
(to the left) representing the premises, and one (to the right) representing the conclu-
sion, counts as a valid sequent if in all interpretations it is truth-preserving, i.e. if all

6
Note that a variable ‘x’ or ‘y’ is not a schematic letter; it is part of the notation that represents a quantifier.
I add that in place of ‘8x(—x—)’ Russell’s notation is ‘(x)(—x—)’. Other notations that have been used in
place of ‘8’ and ‘9’, often for a special purpose, are ‘∧’ and ‘∨’, and ‘П’ and ‘S’.
7
It is usual to exclude the ‘empty’ domain, i.e. the case where there is no object that qualifies as ‘something’, so
that no 9-proposition is interpreted as true. Whether this is a good idea may be debated, and is debated in many
places (including my book [1997: ch 8]). But for present purposes I follow the usual procedure, without further
debate. This suits Russell’s practice. (But in his later Introduction to Mathematical Philosophy [1919a: 203–4], he does
admit that he is no longer satisfied with this practice. Cf. p. 69, n. 42.)
 E L E M E N TA RY L O G I C 9

interpretations which make all the premises true also make the conclusion true. For
schemata, validity is a matter of truth in all interpretations.
We can here make a simplification. So far as the interpretation of propositional
letters is concerned, all that is needed is to assign them a truthvalue, for that is all that is
relevant to the validity of the formulae (or sequents) in which they occur. Similarly, for
the interpretation of a subject letter the only thing that is relevant is which object in the
selected domain it refers to; how it does so does not matter. And again for predicate
letters the only thing that is relevant is which objects in the domain they are to count as
being true of. The ordinary notion of ‘interpreting’ one of these schematic letters would
no doubt take it to be representing an ordinary sentence or subject expression or
predicate expression (as the case may be), and these expressions would have a genuine
meaning. Moreover, it is because of that meaning that the expression would have
whatever truthvalue it does, or would refer to whatever object it does refer to, or
would be true of whatever objects it is true of. But in elementary logic it turns out that
these ‘meanings’ do not matter. To put this in traditional terms, the intensions of what
counts as an interpretation are irrelevant, and what matters is just the extensions assigned,
where extensions are understood as above.8 That is enough to determine whether a
given schematic formula is ‘true under all interpretations’, or whether a given sche-
matic sequent is ‘truth-preserving under all interpretations’, and hence whether it
counts as valid.
I should make it clear that for this purpose one does not think of the truthfunctors
involved as themselves open to various interpretations. They are each to be held to a
constant interpretation, as given by their truthtables. Similarly, in any permitted
interpretation the subject letters must be taken to be referring to an object in the
domain selected, and the predicate letters must be taken to be true or false of those
objects. Finally, the quantifiers 8 and 9 range over just those objects, and are always
taken to mean every or some one of them. That also is held constant. So there are
constraints on what is to count as a ‘permitted’ interpretation, and validity is a matter of
truth (or truth preservation) in all permitted interpretations. But in future I will take this
qualification to go without saying.
This explains what it is for a schematic formula to count as valid, or a sequent of such
schematic formulae. We may now add that an actual proposition counts as ‘formally
valid’ if and only if it has a valid form, i.e. is an instance of some schematic formula that
is valid.
Similarly, an actual argument is ‘formally valid’ if and only if it has a valid form, i.e. it
is an instance of some schematic sequent that is valid. Rather than ‘formally valid’ it
would be more accurate to say ‘valid just in virtue of the truthfunctors and first-level
quantifiers that it contains’. This begs no question about what is to count as the ‘logical

8
For more on ‘intensions’ and ‘extensions’ see pp. 69–71 below.
10 L O G I C A N D M AT H E M AT I C S

form’ of a proposition or an argument, but it does indicate just which ‘forms’ are
considered in elementary logic.
Finally, the task of logic as nowadays conceived is the task of finding explicit rules of
inference which allow one to discover which formulae (or sequents) are the valid ones. In
the case of the truthfunctors this task is extremely simple. Nowadays one might say
that all that is needed is the rule for testing a whole formula (or sequent) by drawing
up a truthtable for it. But there are many other techniques, including the one
pursued by Russell (and by Frege), which I shall illustrate in a moment. Nowadays it
will seem to most people that there is no great philosophical significance in which
particular method is adopted. What is required is just a set of rules which is both (i)
‘sound’ and (ii) ‘complete’, in the sense (i) that the rules prove only formulae (or
sequents) that are valid, and (ii) that they can prove all such formulae (or sequents).
From his earliest consideration of the topic, Russell does provide such a set of rules for
the truthfunctors. He took longer to find a suitable set for the (first-level) quantifiers, but
his treatment in Principia Mathematica [vol. 1, 1910] would certainly qualify.9

2 A Russellian presentation of elementary logic

Both Russell and Frege set out their logical theories by using what is now called ‘the
axiomatic method’. Since that method is likely to be less familiar in these days than it
once was, I shall now give a detailed example. I do not follow exactly the definitions
and axioms that Russell himself adopted, for they have since been shown to be both
inelegant and inconvenient, but I do follow his general approach. (I also add a number
of details which he passes over in silence.) The system for elementary logic that is
presented here will be extended in Chapter 4, section 5, to give us a system for more
advanced logic.
First, for the logic of truthfunctors we may take just ~ and ! as our basic and
undefined notions, and define the other truthfunctors in terms of them. The formulae
of the system are then defined by these three rules of formation:
(a) Any one of the sentence letters ‘P ’, ‘Q’, . . . standing alone is a formula;
(b) If A is any formula, then ~A is also a formula;10

A &B for ~(A!~B),

A∨B for ~A!B,
A$B for (A!B) & (B!A).

9
There is no adequate treatment of the logic of the (first-level) quantifiers in Peano, nor in Russell’s PoM
[1903]. There is an adequate treatment in Frege’s Begriffsschrift [1879], but in 1903 Russell had not yet
assimilated this.
10
It should be obvious that if A is any formula then ~A is the result of writing that formula preceded by a
negation sign ‘~’. Similarly for other uses of A, B, C in this paragraph.
 E L E M E N TA RY L O G I C 11

(c) If A and B are formulae, then (A!B) is also a formula.

The definitions of other truthfunctors are: for any formulae A, B we abbreviate

We now turn from rules of formation to rules of proof, and begin by laying down
three axiom schemas: for any formulae in place of A, B, C, the following formulae are
axioms

Ax 1: ‘ A ! (B!A),
Ax 2: ‘ (A!(B!C)) ! ((A!B) ! (A!C)),
Ax 3: ‘ (~A!~B) ! (B!A).
The sign ‘‘’ says that what follows it may be proved. So this rule is that any axiom may
be proved, and we add a further rule of proof, which is often called Modus Ponens,
but is better called the rule of Detachment, namely

Det: If ‘A and ‘(A!B), then ‘B.

A proof is then a (finite) series of formulae, each one of which is either an axiom or a
consequence by the rule of Detachment of two formulae that are earlier in the series.
A proof is said to be a proof of the last formula of the series. This completes the system
for the logic of the truthfunctors. It is both sound and complete for that logic, in the
sense explained earlier.
Let us now turn to the logic of the (first-level) quantifiers. This uses the letters ‘a’,
‘b’, . . . as subject letters, the letters ‘F’, ‘G ’, . . . as predicate letters, and the letters ‘x’,
‘y’, . . . as (subject) variables. So we start by adding a further rule of formation:
(d) A predicate letter immediately followed by a series of one or more subject letters
is a formula.
(The formulae given by this rule are called the ‘atomic’ formulae of the system.) We
may use 8 as our only quantifier symbol, and define 9 in terms of it, but before we can
formulate its formation rule we shall need some preliminary explanations. The quanti-
fier 8 will be immediately followed by a variable, say v, to form a quantifier prefix 8v.
This will then be followed by an expression A(v) containing the variable v, and
surrounded by brackets, so as to form the whole expression 8v(A(v)). This whole
expression is said to be the scope of its initial quantifier prefix 8v. In a simple case, there
will be no further occurrence of the same prefix 8v, using the same variable v, within
the expression A(v). In that case we say that the occurrences of v within A(v) are free in
A(v), i.e. are free to be bound by a quantifier, and the occurrences of v within 8v(A(v))
are bound in 8v(A(v)), i.e. are bound by its initial quantifier prefix.11 In a more
complex case, one occurrence of the prefix 8v may occur within the scope of another,
and in that case an occurrence of v is bound by the quantifier prefix which has the

11
Note that an occurrence of v within a quantifier prefix 8v counts as bound by that prefix itself.
12 L O G I C A N D M AT H E M AT I C S

smallest scope that includes it. An occurrence of v is free in an expression A(v) if

and only if it is not bound by any quantifier prefix that occurs within the expression.
We can now formulate our final rule of formation:
(e) If A(t) is any formula containing a subject letter t,12 and if A(v) results by
substituting for each occurrence of t in A(t) an occurrence of a variable v that
is free within A(v), then 8v(A(v)) is a formula.
This completes the formation rules.13 That is, the formulae of the system are just those
expressions that can be shown to be formulae by using the five rules (a)–(e) now given.
We add the obvious definition of the quantifier 9, namely (for any variable n)

9v(A(n)) for ¬8v(¬A(v))

Let us now turn to rules of proof for the quantifier 8. Where A(v) contains
occurrences of v that are free in A(v), and A(t) results by substituting a subject letter
t for all such occurrences of v in A(v), we lay down as axioms all formulae

Ax 4: ‘ 8v(A(v)) ! A(t)
Where A(v) is as before, and where B contains no free occurrence of the variable v,
we also lay down as axioms

Ax 5: ‘ 8v(B!A(v)) ! (B!8v(A(v)))
Finally, we add a further rule of inference which is called the rule of Generalization:
if A(v) contains free occurrences of v just where A(t) contains occurrences of t, then

Gen: If ‘ A(t), then ‘ 8v(A(v))

A proof is defined as before, i.e. as a series of formulae, each of which is either one of
the axioms 1–5 or a consequence by Detachment or Generalization of preceding
formulae. This completes the system for elementary logic. Just as the quantifier-free
part of this system is both sound and complete for the logic of truthfunctors, so too the
whole system is both sound and complete for the logic of the quantifiers of first level.
In each case, the soundness of the system is very easily proved, for one only has to
observe that all the axioms are valid and that the two rules of Detachment and
Generalization preserve validity. But the proof of completeness is not trivial, and no
such proof was discovered until Gödel [1930].
I add a couple of comments on this presentation of elementary logic. First the rules
of formation have been framed so that (i) no variable occurs free in what is counted as a
formula, but only in expressions which are parts of formulae but not themselves

12
The letter ‘t’ stands for ‘term’. In elementary logic our only terms are the subject letters. In the more
advanced logic of Chapter 4 there will be further terms.
13
A usual abbreviation allows a quantifier to bind two variables at once. Thus ‘8xy(A(x,y))’ abbreviates
‘8x(8y(A(x,y)))’, and similarly ‘9xy(A(x,y))’.
 E L E M E N TA RY L O G I C 13

formulae; and (ii) no quantifier is vacuous, i.e. every quantifier prefix does bind one or
more occurrences of a variable in what follows it. These features of the system keep it
closer to what in our ordinary language are familiar sentences expressing propositions.
However they may be lacking in other treatments of elementary logic, for rules which
ignore them are in fact simpler. If vacuous quantifiers are permitted, the rules will
ensure that they add nothing, i.e. that if v does not occur free in A then 8v(A) is
everywhere interchangeable with the plain A. If formulae are allowed to contain free
variables, then those variables will function just as schematic letters do in our treatment,
and in fact there will be no need for separate alphabets to distinguish the two. I shall
exploit this point when I come to formulate the extended logical system of Chapter 4.
Second, it is obvious that there can be different choices of which symbols to take as
defined, of which formulae to take as axioms, and of which further rules of inference to
adopt. But all choices should lead to the same results, in the sense that they should
provide proofs of just the same theorems. The axioms that are adopted here have been
chosen mainly because they lead quickly and easily to other methods of proof, which
are actually much more convenient than are the axiomatic proofs that Russell himself
employs. (The crucial result for this purpose is what is called ‘the deduction theorem’,
and this comes without difficulty from axioms 1, 2, and 5.) But I do not here offer any
further development of this logic; I have merely chosen a basis that can quite easily
be developed.14
Let us now turn back from the formal details of elementary logic, as that subject is
understood nowadays, in order to say something of how Russell himself understood
the situation, so many years ago (i.e. in 1910). The main point of difference is that he
never attained our conception of what counts as a valid formula, i.e. one that is true
under all interpretations, so he had nothing against which to check his (perfectly
correct) rules of proof.

3 Russell and validity

Like Frege, Russell never thought of the quantifiers in his logic as being open
to different interpretations by a different choice of domain. While he does of
course recognize that in practice the use of the quantifiers may be restricted to a specific
domain—e.g. all numbers, or all material objects, or whatever—still he thinks that this
is not their ‘proper’ use. For in logic the quantifiers should always be construed as
ranging over everything whatever, and any intended restriction should be explicitly
expressed, e.g. as ‘for every x, if x is a material object, then . . . ’. Indeed, Russell goes
further than Frege on this point. For Frege always thought of his first-level quantifiers
as ranging over the total domain of ‘all objects’, but at the same time he also made use
of second-level quantifiers, governing variables which take the place of predicate letters

14
Such a development may be found in many textbooks, including my [1997: ch. 5].
14 L O G I C A N D M AT H E M AT I C S

rather than subject letters. And he insisted that these quantifiers have a different
interpretation: they range over ‘all concepts’ rather than ‘all objects’, and—in his
view—concepts are not objects. But in his early days (e.g. in his PoM and for some
time thereafter) Russell could not accept this distinction. His position, put in Fregean
language, was that every concept is an object, and so is anything else that we can talk
about. Everything whatever must be what Frege would call an object and what Russell
calls a ‘term’.15 In consequence, the quantifiers used in logic should always have the
same domain, namely the all-inclusive and maximum domain of everything whatever.
Quite apart from this matter of domains for the quantifiers, Russell did not think of
his logical formulae as containing schematic letters, open to any number of different
interpretations. On the contrary, his view is that in logic one makes genuine assertions,
that no genuine assertion can contain what is merely a schematic letter, and conse-
quently that these letters should always be understood as variables bound by the tacit
occurrence of an initial universal quantifier. (In Russell’s own terminology they are
merely ‘apparent’ variables, i.e. variables which are governed by a quantifier.) It is true
that in working out proofs it is convenient to treat them as ‘free’ variables, i.e. variables
that are not bound by any quantifier (which Russell calls ‘real’ variables). That is, they
are treated as if they stood for some definite proposition or subject or predicate, as the
case might be, but one which is left unspecified. However, when they are treated in
this way, the formulae of logic do not make any genuine assertions, which Russell
thinks that they should do. So his ‘language’ for logic is not thought of as a merely
schematic language, using schemata which propositions of a genuine language may or
may not exemplify. What looks like a schema must really be a universally quantified
assertion, with the quantifier left unexpressed.16
Because he does not think of the formulae of logic as open to different interpreta-
tions, he does not have our concept of validity as ‘truth in all interpretations’. He
simply has the concept of what counts as a truth of logic, and his investigations are
designed to provide a ‘foundation’ for such truths by providing the basic axioms and
rules of inference from which they can all be deduced. In modern terms, Russell was
quite familiar with the idea that logic provides proofs (which we now symbolize by ‘‘’),
but had only an intuitive and unworked-out idea of validity (which we symbolize by
‘⊨’). He welcomed Wittgenstein’s claim that the truths of logic should all be ‘tautol-
ogies’, but then found that he did not know how to define a tautology, and this was a
problem that he never resolved.17 His initial attitude to logic was that one could just
‘see’ that its basic rules and axioms are correct, and so could be confident in whatever
could be deduced from them. But, as we shall see, this attitude changed as a result of his

15
See e.g. PoM, }}47–9.
16
This makes it difficult to be clear about what is happening in a proof. Russell’s problems on this point
are discussed in Kemp [1998].
17
I return to it on pp. 69–72 below.
 E L E M E N TA RY L O G I C 15

continued wrestling with ‘the contradiction’, which from now on will be his major
problem in thinking about logic.
Before I come to this, I add just one further remark about the elementary logic that
has been the subject of this chapter. Neither Russell, nor anyone else who has been
interested in the ‘logicist’ claim about mathematics, can believe that the elementary logic
here described is all of logic. There must be more to the subject than this, if there is to
be any chance that it will imply all of mathematics. Frege had assumed a second-level
logic, in which quantifiers are allowed to bind variables which take the place of
predicate expressions, as well as variables which take the place of subject expressions.
He had also made use of the notion of a class, in a way that turned out to be a disaster.
In consequence, Russell was wary of just what more would be needed as logic, beside
what is now generally called ‘elementary’ logic. But he was always aware that more
would be needed.
 2
The Contradiction (i): The Problem

Mathematicians are used to thinking in terms of classes, or sets. At the time that we are
now concerned with, i.e. around 1900, there was no recognition that one might usefully
distinguish classes from sets, because two distinct notions are involved. Roughly
speaking, sets are to be conceived as ‘built out of their members’, along the lines first
suggested by Zermelo,1 whereas classes are conceived rather as the extensions of pre-
dicates, and therefore as given by our independent understanding of those predicates.
This distinction is a post-Russell development, and when we are thinking of Russell’s
problems we must try to set it aside. I shall therefore use the words ‘class’ and ‘set’
interchangeably in this book, and I shall use the usual current notation for sets, i.e.2
{a, b, c, . . . } for the set which has as members just a, b, c, . . .
{x: —x—} for the set which has as members just those
things x such that —x—.
Anyway, as I say, mathematicians are used to thinking in these terms, and at the time
when Russell was beginning to think about how mathematics might be ‘reduced’ to
logic this tendency had quite recently been reinforced. Aside from Peano, and Frege,
two mathematicians who had each been making important contributions were Dede-
kind and Cantor, and both of them thought in terms of classes (or sets). It will be useful
to begin this chapter with a brief description of their work, for Russell was much
influenced by it.

1 The development of mathematics

Dedekind had proposed in his [1872] a way of thinking of the real numbers, i.e. those
numbers that are expressed in our usual decimal notation only by unending decimals.
His basic thought was that a real number is determined by a ‘cut’ in the rational
numbers, i.e. a cut which separates all the rational numbers into two non-empty classes,

1
I shall describe Zermelo’s system, and the way that his thoughts have since been developed, in }5 of this
chapter.
2
Russell’s own notation for the second is ‘^x (—x—)’.
 THE CONTRADICTION ( I ): T H E P RO B L E M 17

those less than it and those greater than it.3 He demonstrated that the usual arithmetic
of the real numbers could be deduced, via this conception, from that of the rational
numbers, and—more importantly—that this conception also allowed one to prove, for
the first time, the existence of suitable real numbers to be the limits of various familiar
series of numbers. Russell took over this idea, but modified it. Instead of saying that
a real number is something (invented by us) that is ‘determined’ by a cut in the rational
numbers, why not just identify the real number and the cut? Or, to simplify things still
further, let us say that the real number is the lower half of the cut, i.e. the class of all the
rational numbers that are ordinarily thought of as less than it. That is: a real number just
is a class of rational numbers such that (i) at least one rational number is a member of it,
and (ii) at least one is not, and (iii) any rational less than some member of the class is also
a member of the class, and (iv) there is no greatest member of the class. Any class of
rational numbers that satisfies these four conditions will be the lower half of a Dedekind
cut, and if we simply identify the real numbers with such classes then a perfectly good
theory of real numbers will result. It was Russell himself who proposed the identifica-
tion, which may at first seem somewhat artificial, but which clearly does suit the logicist
approach. For the real numbers are then ‘constructed’, by purely logical means, from
the rational numbers, which are presumed to be already familiar. But it was Dedekind
who had shown that the resulting theory was entirely satisfactory from the mathemati-
cal point of view.4
For the next step we need the notion of an ordered pair, which we write as ‘<a,b>’
for the pair which has a as first member and b as second member. It was discovered by
Wiener [1912–14] how to represent an ordered pair in the vocabulary of ordinary set
theory, and his idea, as modified by Kuratowski [1920], has been used ever since.5 This
allows one to regard two-place relations simply as sets of ordered pairs. But Russell
took the opposite approach, with relations as primitive, and defined the ordered pair
<a,b> as the particular relation R such that, for all x,y, ‘xRy’ holds if and only if x ¼ a &
y ¼ b. In any case, once ordered pairs are available, it is easy to see how each rational
number n/m is determined by an ordered pair of integers <n,m>, and so may be
identified either with one such pair (e.g. the fraction ‘in its lowest terms’), or with
the set of all such ordered pairs that determine the same rational number. Similarly,
the signed numbers can be identified either with one particular pair of unsigned

3
There may also be a rational number which is neither less than it nor greater than it, but equal to it. This
is a trivial complication, which will very shortly fall away.
4
Of course, there are also other identifications which are equally satisfactory from this point of view. In
particular, I should mention that Cantor had offered an approach to the real numbers that was rather different
from Dedekind’s, but would have provided an equally effective identification. However Russell chose to
work from Dedekind’s basic idea.
5
What is now called the Wiener/Kuratowski definition is
<a,b> for {{a}, {a,b}}.
Many other definitions are available which have the same required consequence, namely
<a,b> ¼ <c,d> $ a ¼ c & b ¼ d
18 L O G I C A N D M AT H E M AT I C S

numbers—say ‘+n’ for ‘<0,n>‘, and ‘–n’ for ‘<n,0>’—or again with a class of them.
So, in general, all the familiar kinds of number can be ‘constructed’ as classes built up
in several stages from the natural numbers.6 What remains, then, is to show how the
natural numbers may in their turn be identified with classes of some other and familiar
things.
Here another work of Dedekind’s, i.e. his [1888], was helpful, but not yet sufficient
by itself. Dedekind had analysed the structure of the series of natural numbers: they
form a series of the kind nowadays called a ‘progression’, which means that there is a
first number, and for each number a next number, and no more than this requires. The
last condition can also be put by saying that each number can be reached by a
succession of steps, starting with the first number and going on step by step, with
each step taking one from the number last reached to the one next after it. Dedekind’s
important discovery was how to put this in purely logical terms, namely: whatever
is true of the first number, and is such that if it is true of any number then it is also true
of the next, must be true of all numbers.7 This is called the principle of mathematical
induction. It allows one to set down a simple set of axioms for the natural numbers, and
Dedekind in effect showed that these axioms form a sufficient basis for the deduction of
all the familiar properties of the natural numbers. Russell first learnt of these axioms
from Peano, and he therefore called them ‘Peano’s postulates’, which is how they have
been known ever since. But in fact Peano took them (with acknowledgement) from
Dedekind’s work. From Russell’s point of view, the important contribution made by
these postulates was that they showed just what conditions a suitable way of ‘con-
structing’ the natural numbers would have to satisfy, if it was to be adequate for all
mathematical purposes. But they did not by themselves show how to find any items
that did in fact satisfy them.
Here is where Cantor’s work was helpful. Cantor’s principal contribution was his
theory of infinite numbers [1895, 1897], which we shall come to shortly, but as a
preliminary he proposed a criterion of identity for all cardinal numbers,8 whether finite
or infinite, namely this: two classes have the same number of members, i.e. are
equinumerous, if and only if there exists a relation which correlates their members
one-to-one. (That is: each member of the first class bears the relation to one and only
one member of the second, and for each member of the second class there is one and
only one member of the first that bears the relation to it.) As with Dedekind on the real
numbers, Cantor did not himself suggest that this criterion provided an identification of
the cardinal numbers with any class of more familiar items, but Russell saw that it could
be used as such. His idea was that two classes share the same number if they share

6
Russell’s chosen way of doing this is most conveniently set out in his Introduction to Mathematical
Philosophy [1919a], henceforth IMP, chapter 7.
7
Frege’s Begriffsschrift [1879] contained the same discovery, so he has priority. But Dedekind’s work was
independent of Frege’s, and at the time was much better known.
8
The cardinal numbers are those that answer the question ‘how many?’ (A little thought will show that not
all the things that we call numbers do answer this question.)
 THE CONTRADICTION ( I ): T H E P RO B L E M 19

membership of the class of all classes with that number, so the number may simply be
identified with the class of all classes that have that number.9 That is to say: a number is
a class of classes, which (i) has at least one class as a member, and (ii) has as its other
members all and only the classes which are equinumerous with that one. Moreover,
equinumerousness can then be explained in terms of one-one correlations, as Cantor
did, and this is an explanation in purely logical terms. We therefore have a construction
of the cardinal numbers which apparently provides just what logicism needs.
The construction so far applies to all cardinal numbers, whether finite or infinite, but
the natural numbers are just the finite cardinal numbers, and so we now need a way of
distinguishing these. Dedekind’s work shows us how to do so: we need only define the
first number 0, and for each number n the next number n+1, and then his axioms will
do the rest. We have shown how to define 0: it is the class of all classes with 0 members,
i.e. with no members, and that is easily described.10 Moreover, since the number n is
the class of all n-membered classes, it is easy to see that the number n+1 will be the class
of all those classes that have one more member than do the classes that are members of
n. And a class a, has one more member than a class b if and only if, upon removing one
member of a, what is left is a class equinumerous with the class b. So we now have
definitions of 0 and of n+1, and we can define the finite cardinal numbers as those that
can be reached from 0 by a succession of steps from n to n+1, which is to say: those such
that whatever is true of 0, and is true of n+1 whenever it is true of n, must be true of
them. Then all that remains to be done is to show that the finite cardinals, so defined,
do satisfy Peano’s postulates. This requires us to prove just two more, quite simple,
theses: (i) that 0 is a number which does not come next after any (cardinal) number, and
(ii) that for every finite cardinal number there is one and only one (cardinal) number
that comes next after it. The first is in fact an immediate consequence of the definition
of a cardinal number, and the second looks as if it will follow from the theory of classes
which has been presupposed in all our thinking.11
This outlines the position as Russell saw it in 1900. It looked as if the theory of
classes would provide, first a construction of the natural numbers, and a demonstration
that they satisfy Peano’s postulates, and then further constructions from them of the
signed integers, the rational numbers, the real numbers, and any other entities of
traditional mathematics. This would establish the logicist claim that all of mathematics
could be reached from a starting point which comprised just pure logic and nothing
else. But of course the whole scheme, as here outlined, depends upon our being able to

9
Looking back on things in his much later work My Philosophical Development [1959], henceforth MPD,
Russell seems to claim on p. 70 to have discovered this definition for himself. But if he did mean to claim this,
then it is suggested by Rodrı́guez Consuegra [1987] that his memory was not entirely accurate. We need not
enter into this question. (As in note 7, Frege [1884] has the priority.)
10
Since classes are the same when their members are the same, there is just one class which has no
members, namely the null class. So 0 is the class which has just this class as a member and no other members.
11
In fact the second of these theses will give trouble. But we shall not come to that until Chapter 6, when
Russell’s final theory has been reached.
20 L O G I C A N D M AT H E M AT I C S

say that the theory of classes is itself to be counted simply as logic. In June 1901 Russell
discovered the paradox which is called after him, and which made this assumption
about the theory of classes extremely doubtful.

2 How the contradiction was discovered

He made his discovery while thinking of Cantor’s theory of infinite cardinal numbers, so
let us say a little more of this. Apparently there are classes with an infinite number of
members, for example the class of all the natural numbers. According to Cantor’s criterion,
any class which can be correlated one-to-one with this class will also have the same infinite
number of members as it does. Such a class is said to be ‘denumerable’, and both the finite
and the denumerable classes are said to be ‘countable’, for their members can be matched
one-to-one with some or all of the natural numbers, which is exactly what one does when
counting. Now it turns out that the class of all pairs of natural numbers is also denumera-
ble, and since this is a slightly surprising result I briefly indicate a proof of it. Consider the
pairs as written out in a two-dimensional and doubly infinite array thus:

<0,0> <0,1> <0,2> <0,3> <0,4> <0,5> ….

<1,0> <1,1> <1,2> <1,3> <1,4> ….

<2,0> <2,1> <2,2> <2,3> <2,4> ….

<3,0> <3,1> <3,2> <3,3> <3,4> ….

<4,0> <4,1> <4,2> <4,3> <4,4> ….

….

….

….

….

….

We reorder this as a singly infinite array, by taking it in the order indicated by the
arrows, namely
<0,0>, <0,1>, <1,0>, <2,0>, <1,1>, <0,2>, <0,3>, <1,2>, <2,1> . . . .

It is obvious that this array can be matched one-to-one with the natural numbers in
their natural order, and is therefore denumerable. It follows that the class of all signed
integers is denumerable, and so—more surprisingly—is the class of all rational numbers.
But the class of all real numbers is not denumerable, and this is a crucial theorem of
Cantor’s theory.
For simplicity, let us just consider the real numbers between 0 and 1, each repre-
sented as an unending decimal. To secure a unique representation, let us discard all
those which end with an infinite string of occurrences of the digit ‘0’. (That is, we
retain ‘0.4999 . . . ’ but discard ‘0.5000 . . . ’, for these are two different ways of writing
the same real number.) Then, if there were a correlation between these real numbers
and the natural numbers, we could think of it in this way. We list on the left each of the
 THE CONTRADICTION ( I ): T H E P RO B L E M 21

natural numbers, in their natural order, and opposite each we list on the right the real
number with which it is correlated, e.g. thus:
1 012345 ....
2 022345 ....
3 023345 ....
4 013445 ....
5. 034556
.. ....
.. .
Now, go down the diagonal of the array of digits on the right, after the decimal point,
and form a new decimal between 0 and 1 that is different from all those listed, e.g.
by writing a ‘2’ in its nth place if the nth digit of the nth real number is odd, and a ‘3’ if
it is even. So, with the present example, we form the decimal
023233 . . . .

This differs from each one in the list, because it differs from the nth one listed at its nth
place. So the supposed list of all the real numbers between 0 and 1 was not a complete
list, for this one was left out, and the argument shows that there could be no such
complete list. That is, there cannot be a one-to-one correlation between the natural
numbers and all the real numbers between 0 and 1.
This ‘diagonal’ method of argument can be generalized in various ways. Cantor
himself used it to prove the more general result that every class has more subclasses than
it has members. We may begin by picturing the proof in the same way as before.
Imagine that there is a correlation between the members of the class and its subclasses,
and think of the members as listed on the left and the correlated subclasses as listed on
the right. Write the subclass by first writing the whole class, and then crossing out the
members that are not in that subclass, e.g. thus:
a {a,6 b, 6 c, d, . . . .}
b {6 a, b, c, d, . . . .}
c {6 a,6 b, 6 c, d, . . . .}
..d ..{a, b, 6 c, 6 d, . . . .}
. .
Then form a new subclass, omitted from the correlation, by once more going down
the diagonal, and putting a member in the new subclass if and only if it is not in the
subclass with which it is correlated. So, in the present illustration, one forms the new
subclass
{6 a, 6 b, c, d, . . . }

This shows the general idea behind Cantor’s argument, but at the cost of supposing that
there is such a thing as a list of all the members of the class, which apparently assumes
that the class is countable. However this is needed only to help us to picture the
22 L O G I C A N D M AT H E M AT I C S

argument, and we can state the argument in quite general terms without any picturing.
In fact we can show that, however many members our class has, it must have more
subclasses than it has members, for the two cannot be correlated one-to-one. Assuming
any correlation between all the members and some of the subclasses, one finds that
there must be a subclass omitted from it, namely the subclass which has as members
all and only those members of the whole class that are not members of the subclass
with which they are correlated. This result, which is given in Cantor [1895] is often
called ‘Cantor’s theorem’. It shows that his infinite numbers never give out, i.e. that
for each one there is another that is greater than it. But it also led Russell to the
discovery of the contradiction that is called after him.
Russell began with the idea that there ought to be a greatest infinite number, namely
the number of all the things that there are. To put this in terms of classes, it would be
the number of the universal class, i.e. the class which has everything whatever as a
member. If there is such a class, then all of its subclasses should be members of it. So
there is a correlation which correlates all the subclasses with some of the members, and
the correlating relation is simply the one-to-one relation of identity. But Cantor’s
argument deduces a contradiction from this supposition, by showing that there is a
subclass which is omitted from the correlation. What is it? Well, the argument shows
that it is the class of all those classes that are not members of themselves. It follows that if
there does exist such a thing as the universal class then this supposed subclass of it
cannot exist. But once the point is noted one sees that there is a very simple argument
that goes directly to this conclusion, without the roundabout route through Cantor’s
diagonal argument. To suppose that there is a class w which contains as members all and
only those classes that are not members of themselves is to suppose12
9 w 8 x(x2w $ x 2
= x)

From this supposition, taking the particular case in which x is w, we deduce

9 w(w 2 w $ w 2
= w)

This is a contradiction. One can only infer that there is no such class as w is supposed to
be. So either there is something wrong with what is predicated of x in the formula
‘x 2
= x’, or this is a predicate for which there is no corresponding class. When Russell
paid serious attention to Frege’s logic in 1902, he saw that Frege’s system did yield this
contradiction, and wrote to him about it.13 At the time both of them thought that
there must be quite a simple correction that would avoid the contradiction but still
allow for almost all of the ‘new logic’ that they were each committed to. However
Russell quite soon came to see that the problem was serious.

12
‘2’ symbolizes the relation of membership, so that ‘x 2 y’ represents ‘x is a member of y’ (and ‘x 2 = y’
represents ‘x is not a member of y’).
13
Russell’s letter, and Frege’s reply, are conveniently translated in van Heijenoort [1967: 124–8]. One can
only assume that Frege himself had not read Cantor [1895] with sufficient attention.
 THE CONTRADICTION ( I ): T H E P RO B L E M 23

3 Some first reactions

His first and rather hasty reaction was a ‘theory of types’ that he sketched in an
appendix to PoM. The basic idea was to take as a starting point the things which are
not classes, which he here calls ‘individuals’. These are entities of type 0. Then there
are classes which have these individuals as members, and no others, and these are classes
of type 1. Then there are classes which have these classes of type 1 as members, and
these are classes of type 2. And so on upwards. This much is already part of his final
theory of types, which we shall come to later. But the early version in PoM also
contains two features which later disappeared. (i) The hierarchy of classes was not in
fact the strict hierarchy just described but rather a cumulative hierarchy, since Russell adds
that any two types have a sum, which is also a type, and which includes the classes of
both of those two types summed together. (ii) The PoM theory includes an infinite
type, above all the finite types (but, in contrast to the ZF theory of sets, which I shall
describe shortly, Russell has only one infinite type). His main reason at the time was the
case of the numbers. For, as we have seen, he wished to construe the number n as the
class of all classes which have just n members, and such classes may be of any type
whatever. So it appears that the number must be a class of a type that includes them all,
and this can only be an infinite type.
Russell quite soon abandoned this theory, though he never tells us why. The
difficulty over an infinite type is clear: although ‘x 2 x’ will never be true in any finite
type, still it may well be true in the infinite type, and so may ‘x 2 = x’, so the original
contradiction breaks out once more. One can block this by refusing to contemplate an
infinite type, and I imagine that Russell quite soon came to realize this, and so
abandoned infinite types for ever after. But then there is a problem. If we confine
our attention to the finite types, then the theory puts a severe restriction on what classes
there are, and (at present) there seems to be no good rationale for such a restriction. For
example, what is wrong with a class of all n-membered classes? It is obvious that the
theory outlined in PoM contains no answer. A further difficulty is that this theory is only
a theory of classes built from individuals, and has nothing to say about other kinds of
classes. But Russell has already seen that there is also an analogous problem over classes
of propositions. His discussion in PoM ends with the thought that for each class of
propositions there should be a further proposition, which states that all the members
of that class are true. But by Cantor’s theorem there must be more classes of proposi-
tions than there are propositions, and so this too is a contradiction.
In his further thought on this contradiction Russell quite soon came to see that it was
not specially concerned with classes. Just as there is a problem over the (supposed) class
of all classes that are not members of themselves, so there is exactly the same problem
over the (supposed) property of being a property which does not apply to itself. The
same problem even affects such ordinary entities as words and phrases. The linguistic
predicate ‘ . . . is not true of itself ’ (often abbreviated to ‘ . . . is heterological’) is easily
argued to be true of itself if and only if it is not true of itself. One might perhaps infer
24 L O G I C A N D M AT H E M AT I C S

that there is no such class as the one just mentioned, and that there is no analogous
property, but is it possible to conclude that the expression ‘ . . . is not true of itself ’ does
not exist? Have we not just shown, by quoting it, that it does exist? Nevertheless,
despite his awareness of these similar cases, Russell continued for a while to think of his
problem as one that concerned classes. That is why his next thoughts are presented as
thoughts on classes. He tells us of these in his article ‘On some difficulties in the Theory
of Transfinite Numbers and Order Types’ [1906a], where he first outlines what he calls
‘the zigzag theory’ and ‘the theory of limitation of size’. These are both theories about
what classes exist, which introduce no type distinctions between classes. But they do
claim that certain predicates do not define a corresponding class.

4 The zigzag theory

We can deal quite briefly with the first, the so-called ‘zigzag’ theory. This depends
upon a generalization which is intended to capture all cases where we know how to
produce a contradiction which shows that some described class cannot exist.14 Russell
suggests that in all of them there is a kind of ‘zigzagging’ feature, which he describes in
this way: we get a contradiction when we suppose that there is a class u such that (i) all
the members of u have a certain property , and (ii) there is a function f whose value for
that class as argument is again an object with the property , but one which is not a
member of u. For then either there will be no such thing as the class of everything which
has the property , or—if there is such a class—then the function f will have no value
for that class as argument (pp. 141–2). The idea is that in the construction of that class
there is a kind of ‘zigzag’ between the members of the class and the values of the
function f. But, to put the idea in much more general terms, it is this: a class will fail to
exist when the predicate which defines it has a certain kind of complexity, and nice
simple predicates will always succeed in defining classes. For example, there will be a
universal class of all objects, defined by a predicate such as ‘ . . . is identical with itself ’ or
‘ . . . is either red or not red’. But, unsurprisingly, Russell has to admit that he has not
found what seems to him to be a satisfactory way of distinguishing between the simple
predicates that do define classes and the ‘zigzaggy’ ones that do not.
However, there is a general objection to the idea that we can distinguish the
apparent classes that really do exist from the others that do not just by considering
the kind of predicates that define them. First, there will surely be infinitely many
(‘simple’) class descriptions that do succeed in defining classes. (For example, start with
any object a that you like and consider the unending series a, {a}, {{a}}, {{{a}}}, . . . .)
Hence, if every class is an individual there will be infinitely many individuals. By
Cantor’s theorem one then expects there to be uncountably many classes of indivi-
duals. But second, in any learnable language there cannot be more than countably

14
The generalization is reached by reflecting on the argument that leads to the Burali-Forti paradox, as
Russell says at [1906a: 141]. I give a brief statement of that paradox on p. 26 below.
 THE CONTRADICTION ( I ): T H E P RO B L E M 25

many predicates of any kind, whether those that do or those that do not define classes,
as Russell himself well knew [1906b: 184–5]. This is because, if the language is to be
learnable, it must start from a finite basis, e.g. of letters, or phonemes, or words, or
what you will. Further, any expression of the language will have only a finite length.
It follows that the expressions of that language cannot be more than countably many.
Hence only countably many of the candidates for being classes can have their case
determined in this way, i.e. by the kind of predicate that purports to define them, and
this leaves uncountably many still undecided. Some supplementary criterion is there-
fore needed, and has yet to be supplied. Russell’s discussion seems to suggest that all
classes without a definition should be ruled out. At any rate he says that every existing
class must be defined by a ‘norm’, i.e. a propositional function, and that a propositional
function is an expression of a certain kind.15 This is a problem that will come up once
more in Chapter 6, for Russell’s final theory seems to be liable to a similar problem.
But, by the time that we reach that theory, the idea of zigzagginess will have been
dropped and replaced by something altogether simpler.
With that we may leave the ‘zigzag’ theory. It is sometimes said that this theory has
some resemblance to the theory which Quine proposed in his ‘New Foundations for
Mathematical Logic’ [1937]. I make no comment on that, except to remark—as is now
well known—that Quine’s theory would seem to be a very peculiar theory, and one
which appears to have no natural model.16 Before I can treat adequately of Russell’s
second suggestion, a theory of ‘limitation of size’, I must introduce more background
information on Cantor’s theory of infinite numbers, and in particular his theory of the
infinite ordinal numbers.

5 Digression: infinite ordinals and ZF set theory

The natural numbers 1, 2, 3, 4, and so on, may be used both for counting and for
numbering off the members of an ordered series as the first, second, third, and so on of
that series. In the first use they are called cardinal numbers, and in the second use ordinal
numbers. A series which matches the series of all natural numbers, in their natural order,
so that each member can be labelled as ‘the nth member’ for some natural number n, is
the shortest kind of infinite series. It is said to have length o. But, given any such series,
we can always form a longer one by adding a further object at the end, and this new
series is said to have length o+1. A series of length o has no last member, but a series of
length o+1 does have a last member, and this is a member with no immediate
predecessor. Then we can add another object to give us a series of length o+2, and
so on. By continuing such additions one forms a series of length o+o, which has a first

15
[1906a: 136]. Russell takes over the word ‘norm’ from the person with whom he is disputing in that
article (i.e. Dr E. W. Hobson). It is not his usual word, and he does not employ it elsewhere.
16
For useful surveys of investigations into Quine’s NF, and its descendant ML, see Quine [1963: ch. 13],
and Hatcher [1968: ch. 7].
26 L O G I C A N D M AT H E M AT I C S

part of length o followed by a second part of length o. An example would be the series
of all the even numbers followed by the series of all the odd numbers. This length is also
called o2. Clearly we can progress in the same way to o3, o4, and so on, and then to
oo, which is also called o2. These numbers which are used to number off the members
of a series, and to measure its length, are the infinite ordinal numbers, and a sketch of
how the series of all infinite ordinal numbers begins is this:
o, o+1, o+2 . . . . o+o (=o2)
o2, o3, o4 . . . . oo (=o2)
o2, o3, o4, . . . . oo
ω
ω
ω ωω
, .... ∈0
17
ωω, ω , ω

All of these give the lengths of countable series, i.e. series with no more members than
there are natural numbers, so examples can be found just by rearranging the natural
numbers in an unfamiliar order. Cantor’s theory also introduces uncountable ordinal
numbers, but we shall not need to go into any detail about these, save to mention one
problem that concerns the series of all the ordinal numbers. This is the paradox
discovered by Burali-Forti [1897], which antedates Russell’s discovery of his paradox,
and was the first indication of the problems lurking in Cantor’s theory.
The series which are measured by the ordinal numbers are called ‘well-ordered’
series. The distinctive character of such series is that although they may contain several
infinite ‘ascents’ they cannot contain an infinite ‘descent’. That is, if you start from any
member of the series and go back from there to an earlier member, and then back again
to a still earlier member, and so on, then your ‘descent’ must end, by reaching the
bottom, after only finitely many steps. Another way of putting the same condition is
this: if some member of a well-ordered series has a certain property then there must be
an earliest member that has that property, i.e. there must be a point in the series where
that property first occurs.18 Now the idea is that any well-ordered series will have a
length that is given by an ordinal number, for it will be possible to match the members
of the series one by one with the ordinal numbers, and then the length of the whole
series will be given by the ordinal that is next greater than all of those used in the
matching. But this idea apparently leads to a contradiction, for the series of all the
ordinal numbers is itself a well-ordered series, and so should have an ordinal number.
But that ordinal will have to be greater than all the ordinals in the series, and so (inter
alia) greater than itself, which is evidently impossible. So one must apparently infer that
there is no ordinal number which measures the length of this series. Why not? Well,
quite a plausible suggestion is that the series of all the ordinal numbers is too long to be

17
20 is the first ordinal number 2 such that o2 = 2.
18
Notice, incidentally, that the familiar series of the signed integers is not in this sense well ordered. Nor is
any stretch of the series of rational numbers.
 THE CONTRADICTION ( I ): T H E P RO B L E M 27

measured by an ordinal number. As we know from his correspondence, Cantor himself

had pondered a somewhat similar suggestion about infinite classes, namely that some of
them might be too large for their size to be measured by a cardinal number. Neither he nor
anyone else had put such a suggestion into print at this time, but it was quite a natural one,
and it is not particularly surprising that Russell should have explored it, as his next
attempt to meet the problem posed by his contradiction. What is perhaps surprising is
that he should have come so near to the basic ideas that we now see as underlying the set
theory that is widely accepted today, namely the ZF theory, which can certainly be
described as a theory of ‘limitation of size’. It will be helpful if I give a brief outline.
The theory was first presented in Zermelo [1908], i.e. two years after Russell’s paper
of 1906, and initially it did not seem to have much to do with the idea of size
limitation. But its important connection with the ordinal numbers became explicit in
Mirimanoff [1917], and the consequence for size was added as a basic axiom by
Fraenkel [1922]. As we now understand it, the theory is best seen as resulting from a
conception of sets as constructed in stages, with the sets at each stage having as members
only the sets (or other entities19) that already existed at earlier stages. The stages are well
ordered, and so can be indexed by the ordinal numbers, and there are to be as many of
them as there are ordinal numbers. The important constraint is that every set has to
belong to some stage of construction, i.e. the stage at which it is first constructed, and
the ordinal of that stage is also called the ‘rank’ of the set. To illustrate, no set is a
member of itself, because the members of a set have to be of lower ranks than is the set
itself. So if one tried to form a set of all those sets that are not members of themselves
this would have to be a universal set, which had all sets as members. But there cannot be
such a set, because it could not have a definite rank, i.e. there could be no stage at
which it was constructed. For every stage is followed by a further stage, which contains
new sets that cannot be members of sets formed at earlier stages. One must add that this
talk of sets being ‘constructed in stages’ is of course a metaphor, and we are not to
suppose that sets are literally created at earlier or later times. But it is a metaphor which
is very helpful when trying to picture the situation.
Finally, let us come explicitly to the idea of limitation of size. In his [1922] Fraenkel
proposed a new axiom, called the axiom of replacement, which directly enforces this
idea. It stipulates that, if the members of a collection can be correlated one-to-one with
some or all of the members of a known set, so that the collection is no bigger than the
known set, then that collection is itself a set. The rationale is that the correlation
between the members of the known set and the proposed one would yield a correla-
tion between the ordinals that are the ranks of those members, and then since the ranks
of the members of the known set are all less than some ordinal, namely the rank of the

19
One may begin with some individuals which are not sets, and then the first stage of set construction has
sets which contain them as members. But we get a simpler theory if we begin with nothing at all, so that at the
first stage we can form only one set, i.e. the null set, which has no members. (This is known as the theory of
pure sets, i.e. sets built up from nothing but sets.)
28 L O G I C A N D M AT H E M AT I C S

set itself, the correlated ordinals will also be less than some ordinal, which will be
the rank of the proposed set. With Fraenkel’s axiom added, we have what is now called
the Zermelo-Fraenkel (i.e. ZF) theory of sets, which is widely accepted today. To
elaborate Fraenkel’s idea further, one may add that von Neumann in his [1925] and his
[1929] made a further proposal, namely that there is just one size that is too big for there
to be a set of that size, namely the size of the collection of all sets altogether. He offered
various arguments (which I shall not discuss) to show that this is a reasonable way of
making Fraenkel’s proposed axiom more precise, and that it conforms very nicely to
the idea that sets are to be thought of as constructed in a well-ordered series of stages.
Let us now return to Russell in 1906.

6 Russell on limitation of size

When Russell comes to describe the idea that some purported classes would be too big to
exist, he at once connects this with the theory of ordinals. As he says, the Burali-Forti
paradox can certainly be taken as suggesting that a class of all ordinal numbers would be
too big to exist, and this is the starting point of the idea. He adds that any other purported
class, which would have as many members as there are ordinal numbers, would also be
too big. Taking on board Zermelo’s axiom of choice,20 and its consequence that every
class can be well ordered, he therefore proposes as the foundation of the theory that a
class exists if and only if its members can be correlated one-to-one with some proper
initial segment of the series of all ordinal numbers, i.e. with the ordinals less than some
given ordinal. This idea is only slightly different from today’s thought that a class exists if
and only if it has some ordinal number as its rank, to mark the stage at which it is first
constructed. In each case it is the ordinal numbers that are being invoked to determine
what is to count as ‘too big’, and in fact the two suggestions are provably equivalent in
the presence of the axiom of choice. So, since Russell’s proposal in this case is so near to
the theory that is generally accepted these days, one asks why he feels able to reject it in
little more than a page. I think that there are two things to be said in explanation, though
neither of them is actually said in Russell’s own discussion.
The only point that he makes himself is that the proposal is not very helpful until we
know
how far up the series of ordinals it is legitimate to go. It might happen that o was already
illegitimate . . . or it might happen that o2 was illegitimate, or oo or o1 or any other ordinal
having no immediate predecessor. We need further axioms before we can tell where the series
begins to be illegitimate (p. 153, my emphasis).21

20
Zermelo [1904]. I shall discuss this axiom, and its significance, in section 2 of Chapter 6.
21
By ‘illegitimate’ Russell presumably means ‘non-existent’. Compare an earlier statement: ‘At present it
is not easy to see where this series [sc. of ordinals] begins to be non-existent, if such a bull may be permitted’
(p. 144).
 THE CONTRADICTION ( I ): T H E P RO B L E M 29

From today’s perspective one replies that that is exactly what the task is, namely to
provide further axioms. Zermelo began this task by introducing axioms of pairing and
union, which provide the finite ordinals, an axiom of infinity which posits the first
infinite ordinal, and a power set axiom that then takes us yet further. Fraenkel added to
these his axiom of replacement, which certainly ‘legitimized’ o2, oo, and many more.
In more recent times we have become familiar with yet more axioms of the same kind,
postulating the existence of hitherto ‘inaccessible’ ordinals. But it would appear that
Russell himself never gave any attention to this task of providing suitable existence
axioms, even though he had himself suggested exactly that project. Why so?
One reason that was probably influential is that he was still pursuing his goal of
showing that all mathematics is really just logic. But logic is by tradition supposed to be
a completely general subject, making no particular assumptions about the existence of
this or that particular kind of thing. As Russell was aware, it cannot altogether avoid all
existence claims. In the third theory sketched in this paper of 1906 it assumes the
existence of propositions, and in Russell’s final theory it assumes the existence of
propositional functions, but these assumptions are supposed to be of an entirely general
nature. However the various special assumptions of set existence that are made in the
ZF axioms are quite different in style, and in fact, as soon as it became accepted that a
good set theory would have to have such axioms, the idea that set theory counted as
a part of ‘logic’ was very generally rejected. Since what Russell wanted was an
acceptable logic, it is not altogether surprising that he did not pursue this line of thought.
I think that there was probably another reason at work, though again it is not
avowed in the article in question. Russell was aware that the contradiction which
affected the naı̈ve view of classes was equally a contradiction for the naı̈ve view of
properties and relations, of propositional functions, and even of predicates taken as
linguistic expressions. Precisely the same contradiction turns up in each case.22 Now
with classes the ‘limitation of size’ approach is in a way attractive, just because we do
quite naturally think of classes as ‘constructed out of ’ their members, and so the idea
that a class cannot exist until a stage ‘after’ all its members exist does have some appeal.
But there is no such appeal in the other cases. Properties are not ‘constructed from’ the
objects that have them; nor are propositional functions ‘constructed from’ those of
their arguments that happen to yield true propositions as values; nor does a linguistic
predicate depend for its existence on the items that it happens to be true of. It strikes us
as weird to hold that such linguistic predicates as ‘ . . . is the same as itself ’ or ‘ . . . is
either red or not red’ cannot exist, because if they did they would be true of too many
things. In these cases the idea of ‘limitation of size’ seems entirely inappropriate. But
Russell always wanted some general solution to these vexing paradoxes, so this is again a
good reason for him not taking this proposal very seriously.

22
This point is made explicitly at several places in his [1906a], namely pp. 137, 140, 145n., 154n.
30 L O G I C A N D M AT H E M AT I C S

This same paper [1906a] also offers a third theory, which Russell calls the ‘no-classes’
theory. He indicates in this paper that it is his preferred option, and in a paper written
soon after [1906b] he offers a fuller explanation and an even more enthusiastic
endorsement. But to understand this we must first digress to consider his well-
known article ‘On Denoting’ [1905b]. For he says himself that it was this article on
denoting that provided the breakthrough for his hitherto frustrated attempts on the
contradiction.23

23
MPD, p. 79. Cf. letter to Jourdain of March 1906, published in Grattan-Guinness [1977: 79].
 3
On Denoting

1 The early theory of denoting

To appreciate the value which Russell himself placed upon this well-known article, we
must first go back to his earlier theory of denoting, which was published as chapter V of
his PoM. The theory concerns what he calls ‘denoting phrases’, and these consist of a
common noun or noun phrase (such as ‘man’ or ‘man who broke the bank at Monte
Carlo’) preceded by what is now called a ‘determiner’. Russell mentions in particular
phrases of the form ‘all A ’s’, ‘every A ’, ‘any A ’, ‘some A ’, ‘an A ’, ‘the A ’; and his
theory is a theory of how these phrases function.1 The main feature of the theory is, as
the name implies, that they function by denoting, but why should he think this?
The discussion in PoM is preceded by this general principle:
In the present chapter, certain questions are to be discussed belonging to what may be called
philosophical grammar. The study of grammar, in my opinion, is capable of throwing far more
light on philosophical questions than is commonly supposed by philosophers. On the whole,
grammar seems to me to bring us much nearer to a correct logic than the current opinions of
philosophers; and in what follows, grammar, though not our master, will yet be taken as our
guide. (PoM, }46)
It seems to me very likely that Russell is being guided by ordinary English grammar
when he thinks that the role of these phrases is to denote. On the usual understanding,
the paradigmatic kind of denoting expression is a proper name. No one would say that
the phrases which Russell calls ‘denoting phrases’ are proper names, but English
grammar does treat them in much the same way. Setting aside some idioms which
clearly have a special explanation, it is very largely true that wherever grammar allows
you to write a proper name you can also write any of these phrases, and vice versa.2

1
I shall continue to speak neutrally of denoting phrases. Russell usually speaks of the denoting concepts
which (in PoM ) he takes these phrases to express.
2
Russell himself later suggests that it is grammar that has led those whom he calls ‘traditional logicians’ to
treat ‘I met Jones’ and ‘I met a man’ as having the same form (IMP, p. 168). A number of apparent exceptions
to the supposed interchangeability of Russell’s denoting phrases with one another, and with proper names,
are given in Oliver [1999], but in most cases they are quite easily explained away. One point that Russell
notes himself is that ‘all A ’s’ is a plural expression, taking a plural verb, whereas his other phrases are (usually)
singular. Nowadays we all think of this as a trivial point of no significance, but in PoM Russell did not.
32 L O G I C A N D M AT H E M AT I C S

That is surely why Russell thinks of them as denoting. But his theory of how they
denote does distinguish them from proper names, since his view (in PoM ) is that a
name denotes directly, so that the object which is named is itself a constituent of the
proposition that the name is used to express, whereas the other ‘denoting phrases’
denote indirectly, by expressing a denoting concept which then denotes some further
object or objects. So in this case the propositions expressed contain as constituents not
the objects denoted but the concepts that denote them.3
What, then, are the objects that these phrases allegedly denote? Russell thinks of ‘all
A ’s’ as denoting a class, construed as a plural object, and I postpone discussion of this
until the next chapter. He thinks of ‘every A ’, ‘any A ’, ‘some A ’, and ‘an A ’ as in a way
denoting all the things that are A, but each in a different way, for it is clear that
substituting one of these phrases for another can make a difference to what is said.
For example, these two sentences do not mean the same as one another:
I did not kiss every girl there
I did not kiss any girl there
In each of them the denoting phrase refers somehow to all the girls there, but there
must be some difference between ‘every’ and ‘any’ which explains why the two
sentences do not mean the same. Russell’s first thought is that each of them denotes
a combination of all the girls—in fact, a combination formed by ‘and’—but it must be a
different combination in each case. His way of distinguishing between these suppo-
sedly different ‘combinations’ is something that we need not investigate here.4
Unfortunately there is an aspect of this early theory in PoM, which is relevant to the
later theory of ‘On Denoting’ [1905b] (hereafter OD ), but is open to serious doubt:
was Russell assuming in PoM that each of his denoting phrases did succeed in denoting
something? Much later, when looking back on his early days in MPD [1959], he
implies that this was his assumption at the time:
If you say that the golden mountain does not exist, it is obvious that there is something that you
are saying does not exist—namely the golden mountain; therefore the golden mountain must
subsist in some shadowy Platonic world of being, for otherwise your statement that the golden
mountain does not exist would have no meaning. I confess that, until I hit upon the theory of
descriptions, this argument seemed to me convincing. (p. 84)

The position described here is in effect Meinong’s position. I note that it is in the first
place a position concerning denoting phrases which begin with the definite article
‘the’, i.e. those which are nowadays called definite descriptions; it is not obvious

3
This is offered as an explanation of how we can understand a proposition which is about infinitely many
objects: we do not need to grasp each of those many objects, but only the one denoting concept that denotes
them all (}60, }72, }141).
4
The detail of Russell’s theory is frequently obscure. It is discussed in my [2009b]. What is said here
represents what Dau [1986] fairly calls the ‘official’ version of Russell’s theory, but there are also other
versions to be found in the PoM account.
 ON DENOTING 33

whether Russell intended it to apply to other denoting phrases too. I add that when
Russell refers to the time when he ‘hit upon the theory of descriptions’, he need not be
intending to refer precisely to the date (October 1905) when the paper OD was
published. We do now have a working paper of his (called ‘On Fundamentals’5 )
which is dated ‘begun 7th June 1905’, and which shows him struggling towards the
theory that was published in OD. It is therefore not at all surprising that in a short
review which he published in July 1905 6 we find several features which anticipate
the OD theory, including both a Fregean treatment of ‘all’ and ‘some’, and a clear
statement that a definite description beginning with ‘the’ may not denote anything—
neither anything that exists nor anything that subsists (i.e. ‘has being’).7 These anticipa-
tions of OD do not seriously conflict with Russell’s statement in MPD that ‘until he
hit upon the theory of descriptions’ he found Meinong’s argument convincing. But
is he misremembering?8
There are several passages in PoM which do seem to show an agreement with
Meinong, for example this one:
Being is that which belongs to every conceivable term, to every possible object of thought—in
short to everything that can possibly occur in any proposition, true or false . . . “A is not” must
always be either false or meaningless. For if A were nothing it could not be said not to be; “A is
not” implies that there is a term A whose being is denied, and hence that A is . . . Numbers, the
Homeric gods, relations, chimeras and four-dimensional spaces all have being, for if they were
not entities of a kind, we could make no propositions about them. (}427. Cf. }}47–8.)

On the other hand the view that denoting phrases may not succeed in denoting
anything is quite compatible with the overall position of PoM, since PoM does assign
a meaning to such phrases, which is put together from the meanings of the noun phrase
in question and the preceding determiner. They would retain this meaning even if they
failed to denote anything. Moreover, a proposition that is expressed by means of a
denoting phrase contains only the meaning of this phrase as a constituent, and not the
object(s) that it denotes. That would apparently allow us to make propositions about
non-existent things without assuming that they must have some kind of being. And we
do also have passages in PoM that seem to endorse this possible lack of denotation, e.g.

5
This is now published in his Collected Papers, vol. 4 [1994: 359–413].
6
[1905a]. This is a review of a paper by H. MacColl, and is entitled ‘The Existential Import of
Propositions’.
7
In Russell’s vocabulary, being is equated with subsistence, which is a more general notion than
existence. Concrete objects, existing in space and time, are said to exist; abstract objects (such as numbers)
do not exist but they do subsist. That is, both kinds of objects have being. It is not clear how the golden
mountain, which would be a concrete object and not an abstract object, can be said to subsist, though it is
clear that it does not in fact exist.
8
The question here at issue was first raised in a helpful way in Hylton [1990: ch. 6]. Hylton holds to the
traditional view that in PoM Russell was still an adherent of Meinong’s doctrine, and that he changed his
mind somewhere between PoM and OD. It has been discussed by others since, notably Griffin [1996], Makin
[2000: ch. 3] and more recently [2009], and Levine [2001]. Makin argues that Russell was never a Meinongian
over denoting phrases, though he was over proper names, until he came to see that—in doubtful cases—they
could be viewed as covert definite descriptions. Stevens [2010] takes issue with Makin.
34 L O G I C A N D M AT H E M AT I C S

It is necessary to realise that a concept may denote although it does not denote anything . . . All
denoting concepts, as we saw, are derived from class-concepts; and A is a class-concept when
‘x is an A ’ is a propositional function. The denoting concepts associated with A will not denote
anything when and only when ‘x is an A ’ is false for all values of x. (}73)

But Russell also recognizes in this section of PoM that there is a problem over the
supposed propositions that can be expressed by means of these empty denoting
concepts. He has begun by saying that all such propositions seem to be false, but at
once qualifies this idea by calling it ‘only a first step’, and not by itself ‘an adequate
explanation’. The ensuing discussion uses the example ‘chimaeras are animals’ and it
assumes (contrary to }427, cited above) that chimaeras neither exist nor subsist. But this
leads it to conclude that ‘it seems most correct to reject the proposition altogether’, and
this apparently means that the sentence (if taken literally, and not rewritten in some
more tractable way) should be regarded as not expressing a proposition at all.9 One can
only conclude that if, as here, the Meinongian position on non-existents is put aside,
still PoM can hardly offer a satisfactory view of sentences which appear to be about
them.10 However, the position of OD on this question is perfectly clear, so let us now
turn to this.

2 The theory of 1905

For the overall development of Russell’s thought the main importance of OD is that it
is where Russell decisively breaks away from his original respect for English surface
grammar. Wittgenstein is evidently thinking of OD when he says in his Tractatus Logico-
Philosophicus [1921], henceforth TLP,
It was Russell who performed the service of showing that the apparent logical form of a
proposition need not be its real one. (TLP 4.0031)

For OD rejects all of Russell’s early theory of denoting in PoM. It is true that the only
case which it discusses in any detail is that of ‘the A ’, but the opening pages show that
Russell is now putting forward a quite different theory of all his other denoting phrases
too. Thus ‘every man’ and ‘all men’ are now identified, and both are taken to be short
for ‘8x (if x is human, then . . . )’. Similarly ‘a man’ and ‘some man’ are now identified,
and both are taken to be short for ‘9x (x is human, and . . . )’. I add that ‘no man’ is

9
An unpublished mss of 1903, i.e. written shortly after PoM was completed, suggests that ‘we shall have
to say that “the present king of France is bald” is neither true nor false’ (Collected Papers, vol. 4 [1994: 286]).
But this mss apparently accepts that it is a proposition, though it lacks a truthvalue, whereas }73 of PoM claims
that it is not a proposition at all.
10
As the passage continues we find ‘Nothing is a denoting concept, which denotes nothing’. Neither the
word ‘nothing’ nor the similar phrases ‘no A ’ were originally listed as denoting phrases. It is surely very odd to
suppose that they do denote certain special combinations, either of all things or of all A ’s, formed presumably
by the combining notion ‘neither . . . nor . . . nor . . . ’. But this is what the theory of chapter V would seem to
require. However it is even worse to say that, since these are denoting phrases which denote nothing, the
sentences that contain them express no propositions at all. This point is made by Oliver [1999: 264].
 ON DENOTING 35

now included as a similar phrase, and this is no longer any kind of problem, as it was
in PoM.11
What one misses in OD is any kind of argument for the new position on these
phrases, and any account of what was wrong with the old one. There is not even an
explanation of how it can make a difference whether we say ‘every’ or ‘any’, though
this was a central feature of the PoM theory. We can be sure that Russell would now
explain the difference as we do, i.e. as a difference in the scopes of what is actually the
same quantifier. That is, on the new theory ‘I did not kiss any girl there’ and ‘I did not
kiss every girl there’ are respectively to be analysed as
8x (x was a girl there ! ~(I kissed x ))
~8x (x was a girl there ! I kissed x )
What changes from one sentence to the other are the relative scopes of ‘8’ and ‘~’, but
there is no difference in what is ‘denoted’. However it must be admitted that in OD
Russell never says this, and never even hints at it. Presumably the explanation is that
this theory of the quantifiers has been so much a part and parcel of Russell’s own
thinking, ever since his detailed study of Frege in 1903, that he has now forgotten that
it still needs to be explained to others. But in any case, whatever the reason is, one has
to say that OD contains no proper account of what the new theory of the quantifiers
now is, or of what was wrong with the old one.
But it is a radically new theory, and more so than Russell’s own terminology
suggests.12 He opens the article by saying that it is a theory of ‘denoting phrases’, and
he explains what these phrases are by giving the same sort of examples as before.
He adds that ‘a phrase is denoting solely in virtue of its form’, and evidently means by
this its grammatical form. One supposes that he intends this to allow for the fact stated
in his next sentence, that some denoting phrases do not denote anything, e.g. ‘the
present king of France’. For he goes on to state that in other cases the phrases do denote.
But this is not what his theory now is. The theory is much better described as the
theory that these so-called denoting phrases never denote, for that is not what their role
is. Unfortunately, he does not say this. What he says instead is that these phrases ‘do not
have meaning in isolation’, and the intended contrast is with a word that genuinely
does denote, such as a proper name. This does ‘have meaning in isolation’, for its
meaning is just the object that it denotes.
Let us now come to the subject with which OD is mainly concerned, namely
definite descriptions of the form ‘the A ’. In PoM he had treated all his so-called
denoting phrases as having the same role, namely to denote. In OD he still treats

11
See the previous note.
12
In PoM Russell is aware that English sentences containing his denoting phrases are equivalent to those
containing unrestricted quantifiers and bound variables that in OD he gives as their analyses. But he there
claims that they do not mean the same, and so the one cannot be analysed as the other. See in particular
PoM, }89.
36 L O G I C A N D M AT H E M AT I C S

all these phrases in the same way, and the new theory is that none of them work
by denoting; instead they are all complex quantifying expressions. The obvious
opponent here is Frege. For Russell has now adopted Frege’s way of handling ‘all’,
‘every’, ‘any’, ‘a’, and ‘some’, but Frege had treated ‘the’ differently. He had regarded
‘the A ’ as a (complex) name, and not a quantifying phrase. Russell’s main claim in
OD is that the same treatment does extend to ‘the’. Just as ‘An A is B ’ is now to be
analysed as
9x(x is A & x is B )

so also ‘The A is B ’ is to be analysed as

9x(x is A & 8y(y is A ! y=x ) & x is B )

The definite article ‘the’ is thus treated in the same way as the indefinite article ‘a’,
except that it contains a uniqueness claim as well as an existential claim. But in this case
Russell does argue quite extensively for his position. Most of his arguments, but not
all, concern the case when the definite description is not (uniquely) satisfied. This
applies to both of his opening arguments, first against Meinong (p. 45), and then against
Frege (pp. 45–7).

3. Negative arguments: Against Meinong and Frege

Meinong had held that all definite descriptions do refer, and refer to an object that
does satisfy the description in question. However, this object may be one that does
not exist, for existing (sein ) is one thing, and satisfying a description (sosein ) is
another. An object may lack the first while still possessing the second. (I make no
attempt to explain why Meinong held this view.13 ) Russell very fairly objects that
this theory infringes the law of non-contradiction. According to Meinong ‘the
golden mountain’ refers to an object that is golden, and is a mountain, but does
not exist. Similarly ‘the round square’ refers to an object that is round and is square,
and so (of necessity) does not exist. But we have only to build existence into the
description and we have an explicit contradiction. For the description ‘the golden
mountain that does exist’ will now refer to an object that is golden, and is a
mountain, and does exist, but also does not exist. Russell protests that a theory
according to which there are things that both do and do not exist should be rejected,
unless there is absolutely no alternative. In my opinion this is a very strong objection,
and it is not worth trying to rescue Meinong from it.14

13
It is not irrelevant that Meinong studied under Brentano, and that Brentano claimed that it was peculiar
to thoughts that they always have an ‘intentionality’. This means that they are always directed towards some
object, though that object may be one that does not exist.
14
Meinong’s theory was declared dead and buried by Ryle [1972: 7], but by that time serious attempts to
resurrect it had already begun, for example in Routley’s work of the early 1960s. But since Routley’s book
 ON DENOTING 37

Frege’s position is more complicated. For mathematical purposes he recommends a

language in which a definite description ‘the A ’ always has a reference, but not of the
Meinongian sort. If there is one and only one thing that is A, then it refers to that thing;
but if there is not then we must supply something else for it to refer to, and Frege’s
suggestion is that we take it to refer to the class of all A ’s. If there are no A ’s, this will be
the empty class; if there are several A ’s then Frege assumes that they will always form a
non-empty class, and Russell does not here question that assumption.15 But he does
very fairly state that this idea is ‘plainly artificial’, and surely Frege would not deny this.
It is not a proposal for how to construe our existing language, but a proposal to move
away from it towards a better language.
As for our present language, Frege thinks that names, predicates, and sentences will
normally have both a sense and a reference (and, as I have noted, he counts a definite
description as a kind of name). Normally, a name will refer to an object, a predicate will
refer to what Frege calls a concept, and a sentence—according to Frege—will refer to
whichever truthvalue it has. But one of Frege’s leading principles is that the reference
of a complex expression depends upon the references of its simpler components.
It follows that if a definite description ‘the A ’ lacks a reference, then any sentence
that contains it will also lack a reference, i.e. it will have no truthvalue. And this is
what he thinks does happen in our ordinary language. Even though all the expressions
in question may have a perfectly good sense, still some of them may turn out to lack
a reference, with the consequence just described.
There is a very obvious and simple objection to this theory, and it is surprising that
Russell does not at once propose it when discussing Frege’s theory. Consider explicitly
existential statements of the forms ‘the A exists’ and ‘the A does not exist’. (In more
idiomatic language these are ‘there is (or: is not) such a thing as the A ’.) When the
description ‘the A ’ is empty, e.g. when nothing is A, then the first of these sentences
should be false and the second should be true. But on Frege’s theory both of them must
lack a truthvalue, and that is surely an absurd result.16 Although Russell does not
actually say this, he does say something more complex which has the same moral.
Consider the sentence ‘If one and only one thing is A, then the A is A ’.17 Russell
claims that this should always be true, both when there is one and only one thing that is
A, and when there is not. He then objects that on Frege’s theory it cannot be true in
the second case, and this objection is similar to the one that I have just given. But (a) the

[1980] was not published until some time later, priority must be assigned to Findlay [1963]. The most recent
extended defence of Meinong that I know is Jacquette [2009]. (But, against him, see also Griffin [2009].)
15
The assumption must appear very doubtful when one sees how it leads straight to Russell’s paradox.
But that point is hardly important here.
16
Perhaps Russell deliberately avoids this objection here because he is reserving the case of explicitly
existential statements for his puzzle (iii), discussed below.
17
For simplicity, this makes a quite unimportant alteration to Russell’s own wording.
38 L O G I C A N D M AT H E M AT I C S

example is more complex than is needed, which leaves it open to a possible response,18
and (b) Russell carelessly mis-states it. He claims that when there is nothing that is A
then the sentence must for Frege be ‘nonsense’, and Frege should reply that that is not
so; on his theory it has a perfectly good sense, but is lacking in truthvalue. Similarly
Russell complains that ‘The King of France is bald’ is in fact ‘plainly false’, though on
the theory that he is considering it ‘ought to be nonsense’. Frege has the same reply.19
But those mis-statements can be corrected, and then the objections to Frege are
entirely effective. I do not see how he could adequately reply, especially to the
objection that was put first, concerning explicitly existential statements. Perhaps the
explanation is that he was always more concerned to improve on our ordinary
language than to explain how it actually functions. But still, his untenable claim that
an empty name must always lead to a lack of truthvalue does stem from a general
principle which he accepts and regards as important. Besides, he does attempt to defend
this claim as a correct account of our ordinary language when discussing the name
‘Odysseus’ [1892: 62–3]. It is not a claim that he can easily abandon, though Russell is
right to say that it must be abandoned.
We may conclude that Russell does have cogent objections both to Meinong’s
theory and to Frege’s theory. But there are, of course, yet other theories which are
rivals to his own and which have not been considered here. A fairly obvious one, that
seems to me to be defensible, is this modification of Frege’s theory: when a definite
description is not (uniquely) satisfied, then the simple (i.e. ‘atomic’) sentences that
contain it are just false, and the truthvalues of more complex sentences that contain
it are then to be calculated from this starting point. (On this theory ‘the A exists’ is
atomic, and so true or false according as there is or is not a unique A; then the
truthvalue of ‘~ (the A exists)’ is calculated from that of ‘the A exists’ in accordance
with the truthtable for ‘~’.) This rival theory is defended in some detail in part III of my
[1997], and that defence is not repeated here. But one can say that it is to a large extent in
agreement with Russell’s own thoughts about what ought to be the truthvalues of
sentences containing empty descriptions. Let us move on.
After the initial objections to Meinong and to Frege, Russell proceeds to state three
puzzles which a theory of denoting ought to be able to resolve (pp. 47–8). Then he
digresses to develop what he first calls ‘a rather curious difficulty’, and later ‘an
inextricable tangle’, for theories which aim to distinguish between the meaning
(or sense) and the denotation (or reference) of a definite description. This includes
both Frege’s theory and his own past theory of PoM, and it is generally agreed that
he is mainly thinking of the latter. The central example is the phrase ‘the first line of
Gray’s Elegy’ and the passage is generally referred to as the ‘Gray’s Elegy’ argument

18
Russell is assuming, as one does in classical logic, that if ‘P ’ is false then ‘P!Q ’ is true. But someone
might reply that that is not how conditional statements work in our ordinary language.
19
As mentioned earlier (note 9) it is Russell’s own previous theory, in PoM, }73, and not Frege’s theory,
that is forced to dismiss such sentences as ‘nonsense’.
 ON DENOTING 39

(pp. 48–50). After this digression he then returns to his three puzzles concerning
denoting, and shows how his own theory does resolve them (pp. 51–5). I will take
this digression first, though Russell himself might well object to calling it a ‘digression’,
because in his eyes it was the most important argument in OD. That is because it is
the most general. But its interpretation is certainly controversial.20 I will not add to the
many accounts that have been offered, save for one quick comment.
The theory in question holds that a definite description will (usually) have both a
meaning (or sense) and a denotation (or reference). It is the meaning that determines
the denotation, and not vice versa, because ‘there is no backward road from denota-
tions to meanings’ (p. 50). But Russell begins by claiming that the relation between
meaning and denotation ‘is not merely linguistic through the phrase: there must be a
logical relation involved, which we express by saying that the meaning denotes the
denotation’ (p. 49). Compare what he had said earlier in PoM:
There is a sense in which we denote, when we point or describe, or employ words as symbols for
concepts; this, however, is not the sense that I wish to discuss. But the fact that description is
possible—that we are able, by the employment of concepts, to designate a thing which is not a
concept—is due to a logical relation between some concepts and some terms, in virtue of which
such concepts inherently and logically denote such terms. It is this sense of denoting which is here
in question. (}56)
Then the argument which follows (in OD ) can only be interpreted as aiming to show
that the supposed meaning (or sense) of the phrase in question cannot be denoted
except ‘purely linguistically through the phrase’. Of course one can denote the meaning
of the phrase ‘the A ’ by using the definite description ‘the meaning of the phrase “the
A” ’. But Russell’s ensuing argument never mentions this possibility, presumably
because he thinks of it as already ruled out on the ground that it is a ‘merely linguistic’
way of denoting that meaning. The argument that he offers is concerned with other
ways of trying to denote the meaning, and it concludes that none of them work. For
the sake of argument21 one might agree that his reasoning on this is correct, and that
a reference to the meaning of the phrase must go via a reference to the phrase itself.
But what would be wrong with that?
Perhaps the most plausible suggestion is one advanced by Kremer [1994] and by
Noonan [1996]. In later writings Russell explicitly affirms his ‘principle of acquain-
tance’, according to which one must be acquainted with every constituent of a
proposition that one understands. (I shall discuss this principle in Chapter 7.) This
idea was certainly present in his thought at the time when he was writing OD, as is clear
from the fact that it is mentioned as highly relevant both near the beginning and near

20
The latest discussion of the passage known to me is that of Salmon [2009]. In his note 2 he cites no less
than twenty-four previous discussions.
21
In fact the reasoning is just muddled. There is a clear and destructive commentary in Urquhart [2005],
who correctly observes that all those who have attempted to find a defensible argument behind what Russell
actually says have had to import further things that he did not say, or to introduce qualifications to what he
did say, or both.
40 L O G I C A N D M AT H E M AT I C S

the end of that article, i.e. in the second paragraph and in the second-last. At this stage
Russell is not attempting any positive argument for it, but he is assuming it to be
correct, and his claim here is that the position which he adopts on denoting phrases in
OD shows how it could be correct. In contrast, the PoM theory of denoting phrases
seems not to be consistent with this principle. For PoM claims that we do understand
the denoting concept, even though we are not acquainted with all the many things that
it denotes. But, on Russell’s way of thinking, understanding the concept should be a
matter of being acquainted with it, and he believes that what one is acquainted with
one can also name directly, as in ‘that thing’. So, if denoting concepts cannot be named
in this way, but can only be referred to by a description which relates them to
something else, namely a denoting phrase, then we cannot be acquainted with them.
And if we cannot be acquainted with them then they cannot fulfil their role of
explaining how we can understand denoting phrases, and so they no longer have any
raison d’être. (Compare Levine [1998].) That is a speculation on what lies behind the
argument that Russell formulates explicitly, but I confess (a) that it is a speculation, and
(b) that it does not appear to improve the very disputed argument that we actually find
in his text, for that argument never mentions the notion of acquaintance. Let us set this
question aside, and return to what very clearly is argued in OD.

4. Positive arguments: The solution of puzzles

Russell has stated three puzzles which he claims that any good theory of denoting
ought to resolve, and now he shows how his own theory does resolve them. These
puzzles concern
(i) the informativeness of identity statements
(ii) the law of excluded middle
(iii) the truth of claims to non-existence
In each case Russell’s discussion draws attention to the fact that, on his theory, a
definite description is really a complex quantifier, and so—like all other quantifiers—it
has an effective scope. To represent this succinctly, let us abbreviate
9x(Fx & 8y(Fy ! y=x) & (—x—)),
which is Russell’s analysis of ‘(—the F—)’, just to22
Ix:Fx(—x—)

22
The symbol ‘I’ is not Russell’s own symbol for a definite description. Following Peano, he uses ‘Øx:Fx ’
for ‘the x such that Fx ’, and adds a scope indicator in square brackets. Thus in PM he abbreviates
9x(Fx & 8y(Fy ! y=x ) & (—x—))

to
[ x:Fx](— ( x:Fx ) —)
Ø Ø
 ON DENOTING 41

We may read this in English as

Concerning the one and only thing that is F, —it—.
I am proposing this simply as a convenient abbreviation. It is convenient in the same
way as it is often convenient to introduce explicitly restricted quantifiers in place of the
analysis offered by Russell and by Frege, by putting
8x:Fx(Gx) for 8x(Fx!Gx)
9x:Fx(Gx) for 9x(Fx&Gx)

This re-expression is convenient because it is closer to the ordinary surface grammar

of the English language, with which we are familiar. But I do not mean to suggest
any deeper moral about the ‘true logical form’ of phrases beginning with ‘the’ or
‘every’ or ‘some’. It is certainly a feature of Russell’s account that all such phrases have
a similar analysis, but we do not need to stickle over just how that analysis is best
expressed.23
In the symbolism that I suggest there is an obvious difference between
Ix:Fx(~Gx ) and ~(Ix:Fx(Gx ))
To use Russell’s English example, made a little more explicit, this is the difference between
The king of France is (something that is) not bald
It is not the case that the king of France is (something that is) bald
He describes this difference by saying that the definite description ‘the king of France’
has primary occurrence in the first case, and secondary occurrence in the second. This
terminology is a little misleading, for it suggests that there will always be just two
ways of assigning the scope of a description, though in a complex statement there may
well be more. But at this date Russell did not have a word corresponding to our word
‘scope’ (and nor did his potential readers), though he certainly understood distinctions
which we now describe in terms of scope.
It is quite clear how this point bears upon Russell’s puzzles (ii) and (iii). Propositions
of the forms
Ix:Fx(Gx ∨ ~Gx )
Ix:Fx(Gx ) ∨ Ix:Fx(~Gx )

So ‘( x:Fx )’ takes the place of a name, and ‘[ x:Fx]’ indicates its scope. For practical purposes he adds
Ø Ø
conventions for omitting the scope indicator, and then he can treat ‘( x:Fx )’ just as if it were a name. This is
Ø
more convenient in practice, but it does rely upon some tacit conventions.
23
This echoes the sentiments of sections 2.5–2.6 of Neale’s helpful book Descriptions [1990]. (Compare his
p. 60, n. 58.) It is not committed to the idea, which lies behind section 5.6 of that book, and which Neale
argues in some detail in his [1993], that restricted quantification is the ‘true logical form’ of the English
language.
42 L O G I C A N D M AT H E M AT I C S

are not instances of the law of excluded middle, and will be false if there is no unique
thing that is F. The proper instances of this law will be
Ix:Fx(Gx ) ∨ ~(Ix:Fx(Gx ))
and of course these are all true. In a similar way, one may start by thinking that a denial
of existence may take either of these forms
Ix:Fx(~(x exists))
~(Ix:Fx(x exists))
But, if we assume24 that ‘x exists’ may be equated with ‘9y(y=x )’, then the first of these
is contradictory, and so it will always be the second that is intended. And that may
perfectly well be true.
This point about scope does not immediately bear upon puzzle (i), which asks
how statements of identity can be informative. Given Russell’s analysis of definite
descriptions, propositions of the forms
a=a
Ix:Fx(x=a )
are certainly different propositions. It may well be said that the first is never informa-
tive, while it is obvious that the second can be, and that is really all that needs to be said.
It explains why George IV could wish to know whether the second is true (with ‘Scott’
for ‘a’, and ‘x wrote Waverley ’ for ‘Fx’), while having no interest in the first. But
Russell complicates his answer by introducing both of
George IV wished to know whether Ix:Fx(x=a )
Ix:Fx (George IV wished to know whether x=a )
The first is the one that we want, and the one that (presumably) is true. But Russell
claims that the second could also be true.
[This] might be expressed by ‘George IV wished to know, concerning the man who in fact wrote
Waverley, whether he was Scott’. This would be true, for example, if George IV had seen Scott at
a distance, and had asked ‘Is that Scott?’ (p. 52)
This is treading on very dangerous ground. For it is assumed in classical logic that when
‘Ix:Fx(x=a )’ is true then every proposition in which ‘Ix:Fx’ has major scope (i.e. has a
primary occurrence) is equivalent to one which says of a what the original said of Ix:Fx.
This is almost what Russell himself had said when introducing his puzzle on p. 47. So
classical logic apparently endorses this inference:
Ix:Fx (George IV wished to know whether x=a )
Ix:Fx(x=a )

24
Frege and Russell and Wittgenstein would all seem to deny this assumption. The question will be
discussed later, in Chapter 14, section 4.
 ON DENOTING 43

∴George IV wished to know whether a=a

We have no way of representing the idea that in the conclusion the first occurrence
of ‘a ’ is outside the scope of ‘George IV wished to know whether . . . ’, while the second
is within it.
Reflecting on this example, and others like it, one soon comes to see that there is a
real problem here. Classical logic is committed to the principle that what matters is
which things are referred to, and not how they are referred to. So if two names ‘a ’ and ‘b ’
do refer to the same thing, then it assumes that any proposition expressed by using
the one name must retain the same truthvalue when the other name is substituted for
it. That is:
a=b ! (( . . . a . . . ) $ ( . . . b . . . ))
But this principle appears to fail when the proposition in question concerns what
someone knows or believes or doubts, and so on. And when it does fail then it is not at
all clear how to construe quantifiers that ‘quantify in’ to such a context, i.e. where the
quantifier itself lies outside that context, but it governs a variable within the context.
I shall later say (in section 5 of Chapter 12) how Russell himself manages to evade this
problem concerning names, namely by being severely restrictive on what is to count
as a name. But this will still leave the ordinary cases of quantifying in, as in this example
of what George IV wished to know, as something of a problem.25
If we set aside this awkward question about quantifying into an intensional context,
which is not really a problem that specially affects definite descriptions, we may fairly
conclude that Russell’s theory does provide answers to the three puzzles that he
mentions. But (a) we may well ask whether some other theory might also provide
suitable answers, for example the modification of Frege’s theory that was earlier
suggested (on p. 38); and (b) whether the solution that Russell provides is not
altogether too parochial. The point here would be that Russell solves his three puzzles
for the case where an apparent referring expression is an explicit definite description,
but apparently the same puzzles arise where it is not, but is genuinely a name. For surely
names can be empty (as in puzzles (ii) and (iii)), and surely identity statements that use
two different names can be informative (as in puzzle (i))? If so, and if Russell’s solutions
work only for definite descriptions and not for names, then those solutions must appear
to be inadequate. As we shall see later, Russell’s answer to this objection will be that any
expression for which these puzzles arise must really be a definite description and not a
genuine name. There are the beginnings of this answer in OD itself. Although he has
begun by saying that a phrase is a denoting phrase ‘solely in virtue of its [grammatical]

25
The problem arises with contexts that are intensional rather than extensional, and I shall explain the
position in more general terms in the next chapter (on pp. 69–71). Meanwhile I should note that this example
of George IV, and Russell’s apparently similar example of ‘I thought your yacht was larger than it is’, are
discussed in an illuminating way by Kripke [2005: 1020–5] and by Salmon [2009: 352–5], but they do not
address the general problem of how to construe quantifiers that reach in to an intensional context.
44 L O G I C A N D M AT H E M AT I C S

form’ (p. 41), still he goes on to say that such names as ‘Apollo’ and ‘Hamlet’ must be
treated as denoting phrases, even though they evidently do not have the relevant form.
This is just because they are (apparent) names which have no denotation (p. 54).26
This thought develops further in his later writings, where every ordinary proper name
is claimed to be a definite description in disguise. But we have not come to that
development yet, and it depends upon his use of ‘the principle of acquaintance’, which
I shall not discuss until Part II. (It is a feature of his theory of knowledge, i.e. of
knowledge of meanings.) As we have noted, acquaintance is briefly mentioned in OD
(p. 41, p. 56), but a proper development of this idea comes later. By then it is clear
that Russell’s interest has become the analysis of an individual’s thoughts, rather than the
analysis of our common language. But in OD itself it seems that he is thinking of his
theory as an analysis of our ordinary language, so I end this section with just
one comment on the theory, as so conceived.

5 Concluding remarks
The sustained attack on Russell’s theory of 1905 had to wait until Strawson’s article
‘On Referring’ of 1950. Until then, the theory had been either ignored (as in earlier
years) or praised as a ‘paradigm of philosophy’ (which was Ramsey’s description27).
Strawson’s attack relied upon the notion of presupposition, and he claimed that there
were presuppositions that were signalized not only by our ordinary use of ‘the’ but also
by such words as ‘all’. For example ‘All John’s children are asleep’ presupposes, but does
not state, that John has children. But the case that engaged a wider interest was his
claim that sentences containing ‘the A ’ presupposed, but did not state, that there is one
and only one thing that is A.
One way of backing up this suggestion is to consider contexts in which ‘the A ’
occurs, but in which nothing appears to be stated at all, e.g. in questions and com-
mands.28 For example ‘Come at 6.00 p.m. and bring your wife with you’ does not state
that you have one and only one wife, and of course it does not command you to have
one and only one. But it does presuppose that you have one and only one. I find this
claim very plausible, as a claim about how our ordinary language works. Should
Russell be distressed? Well, he should be if his theory aims to be an account of how
our ordinary language works, and that was his aim—or, one of his aims—in OD. This
is shown by the fact that he criticizes Frege’s proposal, always to supply a denotation in
cases where it would appear to be missing, as ‘plainly artificial’. Evidently he is not

26
He has already said in ‘The Existential Import of Propositions’ [1905a] that names such as ‘Apollo’ and
‘Priam’ do have a meaning but not a denotation (p. 100).
27
Ramsey [1931: 263n.], or [1990: 1n].
28
This point is briefly mentioned in Strawson’s [1954: 219], but it is more forcefully developed in Searle
[1969: 159–62]. I add that Russell’s example of what George IV wished to know is an example of an indirect
question; George IV was perhaps presupposing that there was one and only one author of Waverley, but it
would not seem that he was either stating or questioning this fact. (This point is made in L. Linsky [1962].)
 ON DENOTING 45

thinking of his own theory as equally ‘artificial’. But as time went on he came to think
of this question as quite unimportant, for his aim came to be to find a better language
than the one that we all use every day.29
There are many other objections that one might wish to make to Russell’s
theory of definite descriptions, when it is taken to be a theory about our
ordinary language. There are also many ways of responding to them. But this
discussion would become interminable if it tried to encompass them all, so I shall
stop here.30 Let us return to the initial problem: what should we do about ‘the
contradiction’?
The relevant lesson that emerges from OD is that the workings of ordinary
English grammar are not sacrosanct. At first, the ordinary grammar had seemed to
suggest that all of the phrases which Russell had originally called ‘denoting phrases’
function just by denoting objects. In OD he rejects that idea, and adopts Frege’s
account of how such words as ‘all’, ‘every’, ‘any’, ‘a’, and ‘some’ are to be construed
as ways of expressing in English exactly the same propositions as are better
expressed in the vocabulary of logic by the symbols ‘8’ and ‘9’. This has in fact
been his opinion for some time before the writing of OD, which explains why he
does not expatiate upon the point in OD itself, though it would have been helpful
if he had done so. But the central concern of OD is the analysis of definite
descriptions, i.e. phrases of the form ‘the A ’, and here Wittgenstein’s comment is
especially apt:
It was Russell who performed the service of showing that the apparent logical form of a
proposition need not be its real one. (TLP 4.0031)
What is at first ‘apparent’ is that definite descriptions have the same grammatical form as
names do. But what OD shows is that, for the purposes of logic, they should be
analysed differently. They do not ‘have meaning in isolation’, as names do, but
contribute in a different way to the meaning of the propositions that contain them.
As we would say: they are not names but complex quantifiers. However, Russell
himself never put his position in that way. He said, rather, that they are ‘incomplete

29
In his [1957] reply to Strawson, Russell says ‘My theory of descriptions was never intended as an analysis
of the state of mind of those who utter sentences containing descriptions’ (p. 179). This is no doubt a fair
comment on his attitude in [1918b] and later, but it clearly does not apply to his [1910/11], and equally we
should not suppose that it describes his position in OD of [1905].
30
As relevant readings I would make these suggestions. One must begin with Frege, ‘On Sense and
Reference’ [1892], to see what Russell is arguing against. Then, as well as OD, one might also look at
Russell’s IMP, ch. 16, which is a later exposition of the theory, and clearer on what the theory is, but less
detailed on the reasons in its favour. The modern debate may then be pursued in Strawson [1950a],
Donnellan [1966], Searle [1969: ch. 7.1], Kripke [1977]. This is a personal selection; alii alia. The debate
up to this point is helpfully summed up in Sainsbury [1979: ch. 4], but it has not by any means ceased yet. An
extended defence of Russell’s theory may be found in Neale [1990], and a summary of more recent
contributions is given in Ostertag [1998: Introduction].
46 L O G I C A N D M AT H E M AT I C S

symbols’ which disappear upon analysis, or, as he also said, they represent ‘logical
fictions’, for they seem to name objects but do not really do so. The significance is that a
similar verdict might turn out to be right not just for definite descriptions, i.e. singular
phrases of the form ‘the A ’, but also for other expressions that Russell was finding
puzzling, in particular for what appear to be class descriptions, i.e. plural phrases of the
form ‘the A ’s’.
 4
The Contradiction (ii): A Simple
Solution

As we have seen, at the time when Russell was thinking of how mathematics might be
shown to be really no more than a continuation of logic, the most important and
innovative mathematicians of the day all made use of the notion of a class (or set).
Moreover, when Russell first discovered the paradox that is called after him, it was as a
paradox that concerned classes. It is no wonder that the topic of classes should be, for
him, a topic of the greatest importance.

1 Types and classes

His first reaction to the paradox had been to propose a theory of types for classes, and
the broad outline of such a theory is this. Every class belongs to some definite type, and
the types form a hierarchy from lower to higher. The members of a class must be items
of a lower type than is the class itself. Such a hierarchy may be either strict or
cumulative. In a strict hierarchy, the members of the class must all be of the type that
is next lower than its type; in a cumulative hierarchy the members may be drawn from
any of the lower types. The conception that we now see as underlying the ZF theory of
sets is a cumulative hierarchy of sets, and instead of speaking of the ‘type’ of a set one
speaks of its ‘rank’, which is thought of as the stage at which that set was first
‘constructed’. Setting the ‘construction’ metaphor aside, one may just say that the
crucial feature is that each set has a rank, and this is the rank that is next higher than all
the ranks of its members. In the ZF version, the ranks have the structure of the ordinal
numbers, including the transfinite ordinals. These ideas, however, were not to become
prominent until much later, and in 1906 they went no further than Russell himself had
reached when contemplating a theory of ‘limitation of size’. By contrast, the theory
which one thinks of as Russell’s simple theory of types has a strict hierarchy of classes,
rather than a cumulative hierarchy, and its types just have the structure of the natural
numbers. That is, there are no infinite types. So it is altogether more restrictive.
In either case, it is the stratification into different types (or ranks) that keeps the
system free of Russell’s paradox, and other similar paradoxes. For the supposed classes
(or sets) that would give trouble would also have no definite place in the hierarchy of
types. That is a reason for wishing to impose some such hierarchical structure, but it
48 L O G I C A N D M AT H E M AT I C S

needs to be backed by a further explanation of why the universe of classes (or sets)
should be so structured. What is there, in the nature of these items, that brings about
this structure? The ZF answer, as we have seen, is given by the metaphorical idea that
sets are ‘constructed’ from their members, and so cannot exist until ‘after’ all their
members do. But Russell’s simple theory has a yet more restrictive structure, which
could not be justified in the same way. So his problem is to say why classes should fall
into the kind of hierarchy that is envisaged. We shall find that his answer is that it is not
really classes that do exhibit that structure; it is, rather, the propositional functions that
define them.1
In PoM Russell had regarded the plural expression ‘All A’s’ as referring to a plural
item, namely all the A’s, which he called a ‘class as many’. But very soon after writing
PoM he came to adopt Frege’s view, that ‘All’ should be regarded as an English version
of the quantifier ‘8’, and it seems probable that he then thought of the plural phrase
‘The A’s’ as performing this role. (It is certainly a better suggestion than the previous
one.2) In PoM he had also thought that to each class-as-many there should correspond a
class-as-one, which was what the singular phrase ‘the class of (all the) A’s’ referred to.
However he had already seen that this was a dangerous assumption, since a supposed
class-as-one might give rise to contradictions which a mere class-as-many would avoid.
So even in PoM he had expressed a doubt over whether, to each class-as-many, there
did exist a corresponding class-as-one. The way that we ordinarily speak does strongly
suggest this, for we do not ordinarily think that it makes any difference whether you
say ‘the A’s’ (plural) or ‘the class of all A’s’ (singular). But that may be why we do not
ordinarily avoid ‘the contradiction’.
One may fairly say that Russell’s later thought begins from this suggestion: perhaps
we do not really need classes-as-one at all, but could make do just with classes-as-
many. But then it goes on to add that even classes-as-many might be dispensed with.
Just as singular expressions ‘the A’, which appear to refer to a particular object, may be
‘re-parsed’ as not really doing so, the same idea might be applied to the plural
expression ‘the A’s’. Perhaps we do not have to suppose that this expression refers
to a plural item, and can explain how it has quite a different role. This, at any rate, is
the main idea behind the third of the approaches that he mentions in ‘On Some
Difficulties . . . ’ [1906a], namely the ‘no-classes’ theory. I add that this theory of
1906 is not only a ‘no-classes’ theory, but also a ‘no-properties’ theory, and indeed a
‘no-predicates’ theory. The only entities that it admits are individuals of an ordinary
kind and propositions, which are themselves taken to be a special sort of individuals.

1
For the time being assume that a propositional function is just a predicate: it takes individuals as
arguments, and delivers propositions as values. The idea will be generalized as we proceed.
2
In PM the defined symbol for a class is ‘x^ (x)’, which Russell says ‘can always be read as “the x’s which
satisfy x^”’ (p. 30), and he frequently follows this plan in explaining his formal developments (notably in
PM*37, which is headed ‘Plural descriptive functions’). The same idea, that plural expressions of the form
‘The A’s’ refer to classes, is also explicit at IMP [1919a: 181].
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 49

Russell calls it ‘the substitutional theory’, because it makes use of the idea of substitut-
ing one individual for another within a given proposition.3

2 The substitutional theory

The theory is very briefly described in ‘On Some Difficulties’ [1906a], but then is
much more fully explored in the succeeding paper ‘On the Substitutional Theory
of Classes and Relations’ [1906b]. In effect, it is the only place in which Russell
states what we have come to call his simple theory of types, for all his subsequent
expositions concern the more complicated ‘ramified’ theory that we will come to later
in Chapter 5. But the simple theory is presented here with a special and peculiar twist
that is not repeated elsewhere. In its usual version the simple theory is also a ‘no-classes’
theory, in which what looks like a reference to a class is paraphrased away in favour of
a generalization over propositional functions, so that we nowhere assume the existence
of such things as classes. It is not classes, then, that exhibit the hierarchical structure of
simple type theory, but the propositional functions that are invoked in their elimina-
tion. This is a real gain, for propositional functions really do fall into the structure
envisaged. The elimination of classes remains Russell’s final theory, and it is not altered
in any important way in subsequent revisions. But in the usual version of the simple
theory of types the eliminating paraphrase does assume the existence of propositional
functions, whereas the first version of this theory, i.e. the ‘substitutional’ version, also
avoids assuming the existence of propositional functions. Apparent reference to them is
similarly explained away as ‘really’ a reference to something else. The motivation for
this is, no doubt, that Russell’s original paradox applies to the naı̈ve view of proposi-
tional functions just as much as to the naı̈ve view of classes. So, if we aim to avoid this
paradox by banishing classes altogether, we had better banish propositional functions as
well. The usual version of the simple theory will be given on pp. 54–9 below, but first
we have the ‘special twist’ of the early version, i.e. its way of eliminating reference both
to classes and to propositional functions. The basic idea is that variables for proposi-
tional functions may (with a little ingenuity) be replaced by variables for the proposi-
tions that contain them.4 This will make it even more obvious that our theory does
exhibit the structure of the hierarchy of simple types. For in this substitutional theory

3
Russell (at this date) supposes that genuine individuals occur in the propositions that make statements
about them. Chapter 12 will show that he later changed his mind on this point. But for present purposes we
may simply suppose that what he means is that names of individuals occur in sentences that express propositions,
and one may substitute one name for another within a given sentence.
4
The idea goes back to PoM, where already Russell is claiming that a propositional function has no
independent existence: it lives only in the propositions that contain it (}85), for it if had an existence of its
own then we could not prevent it from being a subject to itself as predicate. Again, he claims that: ‘a
propositional function is the class of all propositions which arise from the variation of a single term’ (}91). In
[1906a] it is frequently claimed that propositional functions do not themselves count as entities, even before it
is shown how apparent references to them may be eliminated, e.g. pp. 137, 140, 145n., 154n.
50 L O G I C A N D M AT H E M AT I C S

there is simply no way of writing formulae that would contravene the restrictions
imposed by that structure.
Let us begin with a simple case, for example, the propositional function expressed by
‘ . . . is a man’. We are used to representing this by a letter ‘F ’, and then (in a second-
level logic) we treat this letter as a variable which can be bound by quantifiers. But the
‘no-classes’ theory, as Russell first presents it, has no such letter ‘F’. Instead, we take a
whole proposition, such as ‘Socrates is a man’, and we consider the effect of substituting
for the name ‘Socrates’ the names of other individuals. This gives us a range of
propositions, each of which says that something is a man, and what we might have
wanted to say about the function ‘ . . . is a man’ can now be said instead about the
propositions in this range. So quantifiers over propositional functions are replaced by
quantifiers over propositions. Thus, for brevity let ‘p’ be a variable ranging over
propositions, let ‘a’ be a variable ranging over individuals, and let ‘p/a;x’ represent
what ‘p’ becomes when x is substituted for a in p. Then the simple claims that we might
have expected to express as:
9F(F(Socrates))
9F 8x (Fx)
are now to be written as
9p 9a (a occurs in p & p/a; Socrates)
9p 9a (a occurs in p & 8x (p/a;x)
In effect the collection of symbols ‘p/a’ functions as if it were a variable for monadic
propositional functions of first level, though clearly it is something that ‘has no
meaning in isolation’, for it represents what is expressed in English by ‘the result of
replacing a in p by . . . ’.5 It is not to be thought of as naming either a propositional
function or anything else.
To obtain in a similar way an expression that functions as a variable for propositional
functions of the second level, we may begin with this thought: what is needed is a
proposition which contains a function of first level, and then we need to consider all
propositions obtained by substituting other functions of first level for that one. Since
first-level functions are now represented by incomplete symbols of the form ‘q/a’, we
therefore want to be able to speak of the result of replacing such a symbol by another of
the same form, say ‘r/b’. But Russell thinks that we cannot without further explanation
just write

5
Perhaps better ‘the result of replacing a in p by . . . is true’. Russell’s formulae are deliberately ambiguous
between expressions which assert propositions and expressions which name those propositions. Without
harm to the system one may easily resolve this ambiguity (as Landini [1998] does) by introducing an explicit
symbol, say ‘{ . . . }’, so understood that ‘{p}’ names what ‘p’ asserts. A suitable comprehension axiom then
asserts that all propositions can be named in this way. But in this system it is only propositions and the
individuals that occur in them that can be named, for these are the only things that are assumed to exist.
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 51

p/(q/a); (r/b),
because the positions here filled by ‘q/a’ and ‘r/b’ are supposed to be positions for
individuals, and these expressions are not expressions for individuals. However, the
remedy is simple. All that we need to do is to introduce a way of simultaneously
substituting in a proposition p for two individuals that occur in it, and this is quite a
straightforward idea.6 So let us put
p/(a, b); (x, y)
for ‘what results from p upon simultaneously replacing a by x and b by y’. Then,
recalling that in this system propositions are counted as (a special case of) individuals,
we can obtain the desired effect by means of the formula

p/(q, a); (r, b)

So long as q and a do occur in p in the context ‘q/a’, this evidently yields what we want.
Clearly one could proceed to yet higher levels in the same way, though I shall not do
so. It must also be admitted that formulae written in this notation do become
unpleasantly complex,7 but that is not a point of any philosophical significance.
To sum up: a variable for a monadic propositional function of first level is replaced
by the incomplete symbol ‘p/a’; a variable for a monadic propositional function of
second level is replaced by the incomplete symbol ‘p/(q, a)’; and in the same way all
other variables of the simple theory of types can be replaced by such incomplete
symbols. Hence quantification over such variables can always be represented in this
theory simply as quantifications over propositions and the individuals that occur in
them. Moreover, since these incomplete symbols are not construed as naming any-
thing, all the restrictions of the simple theory of types on how symbols may signifi-
cantly be combined are automatically observed in this theory, simply as a matter of
what ordinary grammar requires. Russell notes this point explicitly on pp. 177–8.8 As is
demonstrated in detail by }4 of Landini [2003], whatever can be said in the ordinary
version of the simple theory of types can also be said in this theory, and for the most part
whatever cannot be said in the simple theory of types also cannot be said in this theory.
But there is an important exception, which must now be considered.
The paper ‘On the Substitutional Theory’ which I have been describing, was in fact
never published, because although Russell had submitted it for publication there was a
delay, and by the time that publication was agreed Russell had changed his mind and so

6
The formal definition is complex, and so I do not give it. The reason is that if—as is natural—one thinks
of a simultaneous substitution as just one substitution followed by another, then what is introduced by the
first substitution may affect the second, which is not what is wanted. Russell shows how to avoid this on his
pp. 173–4. Landini [1998] observes that the complication is not really needed (p. 133).
7
Some fairly simple examples are given in Hylton [1980: 21]; more complex examples may be found in
Landini [1998: chs. 4–7 passim] and [2003: }}3–4].
8
He also repeats it in ‘On “Insolubilia”’ [1906c: 201–2].
52 L O G I C A N D M AT H E M AT I C S

withdrew the paper. It is sometimes suggested (for example, in Lackey [1973: 131–2])
that the change of mind was at least partly due to Whitehead’s complaint that the
notation had become too complex. I doubt whether this complaint can have been very
influential, for when it came to practical purposes Russell was quite capable of
simplifying the official notation of a theory. (Indeed, it is never quite clear what the
official notation of Principia Mathematica is, for the notation that is used in practice is a
simplified one, relying everywhere on the device of ‘systematic ambiguity’.9) It is
surely more important that Russell came to see that quantifying over all propositions at
once, without introducing any type distinctions, leads to trouble. He himself com-
ments, at the end of ‘The Substitutional Theory’, that the only serious danger with this
theory is ‘lest some contradiction should be found to result from the assumption that
propositions are entities; but I have not found any such contradiction, and it is very
hard to believe that there are no such things as propositions’ (p. 188). As a matter of fact
he had already found a contradiction about propositions, which he had published in
the final section of PoM (}500). Perhaps he thought that he now had an answer to that
puzzle, though, if so, we do not know what his answer was.10 In any case, he did soon
discover another, which very clearly does arise within the substitutional theory as first
conceived. It shows how to construct within that theory an analogue of Russell’s own
paradox, as applied to propositional functions. As we have said, the theory will end up
by replacing propositional functions by a reference to all the propositions that contain
them, but it will be helpful to understanding if we first allow ourselves to speak of them
explicitly, using the letters , ł, . . . for this purpose.
From any proposition which concerns an individual, say b, we can form the function
 which results from it upon varying b. (Note that b may be an ordinary individual, or
may itself be a proposition, for in this theory propositions are themselves counted as
one kind of individuals.) Conversely, from any such function  we can form the
proposition which applies  to b, i.e. ‘(b)’. But then it follows that we can also form
the proposition ‘((b))’ which takes this proposition as argument. So there should be
a propositional function ł which is true of any proposition x if and only if x is a
proposition of the form ‘~((b))’. That is
8x(ł(x) $ 9(x=(b) & ~(x)))
Then, taking the case in which x is the proposition ‘ł(b)’, we obtain

9
Officially, each variable should be given an index, to show what type of entities it ranges over. In
practice, these indices are everywhere omitted, and a formula such as ‘x’ is understood as representing any
formula in which ‘’ is given a type one higher than ‘x’ is.
10
The puzzle in its original version concerns classes of propositions. It claims that for each such class there
should be a proposition stating that all members of the class are true, but notes that this conflicts with Cantor’s
proof that there must be more classes of propositions than there are propositions. The reference to classes of
propositions can clearly be replaced by a reference to propositional functions which take propositions as
arguments, and this may then be rephrased once more in the terminology of the substitutional theory,
replacing reference to these propositional functions by references to all the propositions which result from a
given proposition by substituting for some proposition that occurs in it.
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 53

ł(ł(b)) $ 9((ł(b)=(b)) & ~(ł(b)))

But the clause ‘ł(b)=(b)’ implies and is implied by ‘ł=’,11 so we deduce that
ł(ł(b)) $ ~ł(ł(b))
This is our familiar contradiction. Moreover, the leading premise can be formulated in
Russell’s substitutional theory as
8x(po/ao;x $ 9p,a((x=p/a;b) & ~(p/a;x)))
And further, Russell’s proposed axioms for his substitutional theory do provide a proof
that there exists a po/ao as thus described. His substitutional system is therefore
inconsistent.12
Russell tried various ways of modifying the theory so as to avoid this contradiction,
but he always found either that the contradiction reappeared in a new form or that the
deduction of arithmetic was frustrated. So in the end he gave up this substitutional
theory. The problem arose from supposing that all propositions could be counted
equally as individuals, so his response was to propose that propositions be segregated
into different types or orders. This is the start of what we now call his ramified theory of
types, which occupies the next chapter. The basic idea is that a proposition which
contains within itself a quantification over propositions must be of a higher order than
the propositions that are there quantified over, and it is introduced as the obvious
response to the paradox of the Liar.13 Actually, the idea is not put in quite this way at its
first occurrence, in his ‘On “Insolubilia” and their Solution by Symbolic Logic’
[1906c]. There Russell prefers to restrict the notion of a proposition to those that are
quantifier-free, but then to allow what he calls ‘statements’ which do quantify over
propositions and over other statements. So officially it is statements and not proposi-
tions that are first distinguished into different types or orders (pp. 204, 207–8). But in a
paper written shortly afterwards, ‘Mathematical Logic as Based on the Theory of
Types’ [1908], this unexpected approach is dropped, and it is now propositions that
are distinguished into various orders, according to the quantifications over propositions
that they contain. In each of these papers Russell retains the idea of the substitutional
theory, that reference to propositional functions may be replaced by reference to the
propositions (obtained by substitution) that are their values. Hence the type and order
of a propositional function is inherited from the type and order of the propositions that
are its values. But in the second of them, although this approach is endorsed in theory,
it is also admitted that in practice it is ‘technically inconvenient’ (p. 77), and Russell

11
Note that for ‘ł=’ one could put ‘8x(ł(x) $ (x))’, and the reasoning would not be affected.
12
For full details of the proof I refer to Landini [1998: ch. 8, esp. pp. 201–6]. He dubs this contradiction
‘the po/ao paradox’.
13
The Liar is perhaps the simplest of the paradoxes affecting propositions. It is discussed on pp. 76–7
below. In [1906c: 204], Russell claims that the substitutional theory, as first given, will resolve all paradoxes
concerning classes and relations, but not those concerning propositions, and that the theory of types must
therefore be extended.
54 L O G I C A N D M AT H E M AT I C S

goes back to using variables that represent propositional functions directly. Each of
these variables ranges only over functions of a single type, and in [1908] Russell offers,
as a justification for this restriction, not that it is dictated by the substitutional theory that
is the official background theory, but that it is ‘an internal limitation’ given by what is
asserted of this variable, for the assertion will make sense only if its subject is of the
required type (pp. 71–2).14 This shows, I think, that he is already moving away from a
serious adoption of the substitutional theory. In subsequent writings (in particular
Principia Mathematica) that theory is never mentioned, though one might suspect
that it lingers on in the (very questionable) thesis that is so emphasized in PM, that a
propositional function ‘presupposes’ the propositions which are its values.15
The official theory of PM is of course the full ramified theory of types, and that is
postponed to the next chapter. But it should be noted that Landini [1996a] very
plausibly argues that in practice PM uses only the simple theory, for (a) its bound
function variables are confined to variables for predicative functions, (b) the axioms
of reducibility provide suitable comprehension principles for predicative functions,
and (c) the predicative functions do form a simple type hierarchy. However one
should not approach the simple theory by this roundabout route, nor via the
substitutional theory with its awkward reliance on quantification over all proposi-
tions. Let us think of it in the way that it is usually thought of these days. The
quantifiers will range just over individuals and predicates (or propositional functions)
of various levels, but not over propositions. The first level of the theory is just the
ordinary and familiar first-level predicate logic that was explained in Chapter 1. The
idea is to extend this theory to encompass predicates (or propositional functions) of
higher levels.

3 Basic ideas of the simple theory of types

To a first approximation the simple hierarchy of monadic predicates can be thought of as
obtained in this way. Begin with a sentence that mentions an individual, for example,
Socrates is a man.
Drop out the reference to that individual, and substitute a gap in its place, as in
. . . is a man.
This is a first-level predicate. Now consider another sentence which contains that same
predicate in a different context, for example, where the reference to a particular
individual has been replaced by a quantified variable ranging over individuals, as in

14
This justification is repeated in PM (p. 4; cf. pp. 47–8).
15
I shall question the thesis in Chapter 5, section 3.
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 55

9x (x is a man)
8x (x is a man ! x is mortal).
We can now drop out that first-level predicate and leave a gap in its place, as in
9x (—x—)
8x (—x— ! x is mortal).
These are second-level predicates. Then again we can put a variable ‘F’ into the gap
that they contain and introduce a quantifier to bind it, as in
8F 9x (Fx)
9F 8x (Fx ! x is mortal).
Once more we can drop out the second-level predicates, leaving a gap in their place, to
form
8F (---F---)
9F (---F---).
These are third-level predicates. Then again we can introduce a suitable variable to fill
these gaps, and bind it with a quantifier, and so on and on indefinitely. But what would
be a suitable variable in this case?
It helps to start with a particular example of a second-level predicate. One can say of
any two-place relation R that it is transitive by saying
8xyz(Rxy & Ryz ! Rxz).
We quite often wish to confine attention to relations that are transitive, and for this
and other purposes it is convenient to have an abbreviation for this notion. As a first
thought one might suggest just
Trans (R).
But this is inadequate for two connected reasons. First, the notation does not make it
clear that ‘R’ here represents a two-place relation. More importantly, it cannot easily
be extended to cases where the relation we wish to talk of is not represented by a single
letter. For example, one might wish to say that if R is transitive then so is the relation
which holds from x to y if and only if Rxy & Gy, but for this purpose the variables ‘x’
and ‘y’ need to be shown explicitly.16 Nor will it do to write simply
Trans (Rxy) ! Trans (Rxy & Gy).
For this expression appears to contain ‘x’ and ‘y’ as free variables, whereas they are of
course bound in both the antecedent and the consequent of the proposition we wish to

16
PM would be tempted to avoid this by adopting the circumlocution:
8xy (Sxy $ Rxy & Gy) ! (Trans (R) ! Trans (S))
56 L O G I C A N D M AT H E M AT I C S

express. Here are three ways of meeting this problem: ‘ . . . is transitive’ may be
represented by

Frege: Transxy (Rxy)

Russell: Trans (Rx^y^)
Church: Trans (ºxy: Rxy).

Each of these notations is supposed to be just an abbreviation for what we started with,
namely

8xyz(Rxy & Ryz ! Rxz).

In practice Church’s notation is the most convenient, and so it has in effect super-
seded both Frege’s and Russell’s. But there is nothing wrong with Frege’s, and so far as
I am aware whatever can be done with Church’s notation can also be done with
Frege’s, and vice versa. (We have only to view Church’s formula as breaking into two
at the colon, rather than the bracket, to obtain a minor variant on Frege’s formula.) As a
matter of fact there is something wrong with Russell’s notation, for in certain cases
the formulae that it yields would be ambiguous.17 But this is of minor importance.
The main point is that the Russellian ‘cap notation’ is trying to do just what Church’s
º-notation does do, and what Frege’s notation also does in a superficially different
way. In this example it explicitly exhibits a first-level predicate ‘R’ as argument to a
second-level predicate ‘Trans’. But it is only an abbreviation, and is not to be thought of
as making any new assumption about the status of the letter ‘R’.
It is often said that in PM propositional functions occur in two roles, both as predicates
and as logical subjects, and that it is the cap notation that accomplishes this, by turning a
predicate into a name.18 One cannot deny that Russell’s informal explanations do rather
strongly suggest just this, for they frequently treat an expression such as ‘Rx^y^’ as if it were
a noun phrase (such as ‘the relation which holds from any x to any y if and only if Rxy’),
which is used to name what ‘Rxy’ predicates of x and y. But he does this because his
explanations are given in English, and in English one inevitably uses a noun phrase to refer
to what one is talking about. (The present example illustrates this very nicely: in English

17
For a simple example, consider 8 and 9 explicitly as second-level predicates. Then the same formula is
rendered in these three ways by our three notations:
Frege: 8x9y(Rxy)
Church: 8(ºx: 9(ºy: Rxy))
Russell: 8(9(Rx^y^))
Clearly the Russellian version is ambiguous, for it cannot show whether the initial ‘8’ governs the capped ‘x’
or the capped ‘y’. (In the formal development of PM the cap notation is very seldom used, since the device of
systematic ambiguity allows one to omit it. Cf. ‘We have found it convenient and possible . . . to keep explicit
use of symbols of the type ‘x^’ . . . almost entirely out of this work’, p. 19).
18
See, for example, Quine [1963: 244], Chihara [1973: 19 and 24], Sainsbury [1979: 290], Hylton
[1980: 27] and [1990: 295–6], Cocchiarella [1980: 98] and [1989: 46], Wahl [1986: 389–90].
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 57

we use the predicate ‘ . . . is transitive’, which has the syntax of a first-level predicate and
so requires a noun phrase as its subject.) But the language of the theory of types is not
English, and it contains no way of turning a predicate expression (of any level) into a
noun phrase of a lower level. In particular the cap notation does not do this, and because
it looks as though it does I think that it is better avoided. Frege’s notation is less misleading
on this score, because it retains within the symbol for a second-level predicate the
indication that the following individual variables are bound. So it is easier to see that
the symbol ‘Rxy’ is still a symbol for a predicate, and not for an associated noun phrase,
both in ‘Transxy(Rxy)’ and in what this abbreviates.
The point of this discussion of ‘Trans’ was to seek for a suitable way of representing a
second-level predicate. The example was of a particular second-level predicate, which
could be spelt out in full without any abbreviation. But the idea was to use it as a guide
to the best way of introducing a variable which takes the place of any arbitrary second-
level predicate, and is used not to abbreviate a given one but to generalize over them
all. Since Frege’s method seems to be the least misleading, we shall do best to follow it.
He introduces a new letter ‘M ’ which functions in this role, and it will therefore occur
in such contexts as

M xFx, M xyRxy, ...

This new variable can then be bound by quantifiers in the usual fashion, and we are
ready to proceed to the next level of the hierarchy. But we still treat predicates as
predicates; they do not become nouns.
Some, such as Quine, will no doubt protest that we are treating predicates as noun‐
phrases when we use quantifiers that bind predicate variables.19 This is again because in
English the quantifiers are expressed by such words as ‘every’, and English grammar
requires that these words be followed by a noun or noun-phrase.20 But the language of
the theory of types is not English, and its quantifiers are not bound by the same restriction.
We may nevertheless quite easily understand them. Just as a quantifier over individuals,
‘8x (—x—)’, may be explained as saying that ‘—x—’ comes out true for all ways of
interpreting ‘x’, so equally ‘8F (—F—)’ may be explained as saying that ‘—F—’ comes
out true for all ways of interpreting ‘F’. In the first case one interprets the letter (on a given
domain) by saying what object (in that domain) it is to refer to, and in the second case by
saying which objects (in that domain) it is to be true of. But this does not treat a predicate
letter as if it named an object, and in the theory of types there is no way of doing this: a
predicate always occurs with the syntax of a predicate, and not that of a name.21

19
See e.g. Quine [1950: }38; 1953: chs. 1 and 6; 1966: 296–7; 1970: 66–8].
20
There are exceptions. In such words as ‘everywhere’, ‘anyhow’, ‘sometimes’ the quantifier is followed
not by a noun but by an adverb.
21
I should make it clear that this account of the meaning of ‘8F’ is mine, and not Russell’s. I shall come
later to Russell’s own account—or rather, his failure to provide an account—in section 6.
58 L O G I C A N D M AT H E M AT I C S

It is a consequence that a predicate of any level can significantly occur only as applied
to an item of the next lower level, or to a variable that stands in place of such an item, as
the strict hierarchy of types requires. The rationale for this is completely straightfor-
ward, and can be illustrated just from our first few levels. At the lowest level are names
of individuals. Next come the first-level predicates, which contain gaps where such a
name may be slotted in to make a sentence, or a variable which takes the place of those
names to make an open sentence.22 Next come the second-level predicates which
contain gaps where a first-level predicate may be slotted in. And so on up. But a gap
that may be filled by the name of an individual (say ‘Socrates’) cannot be filled instead
by an expression of any other type, e.g. by ‘ . . . is a man’ or by ‘9x (—x—)’, for the
result would simply be ungrammatical, for example,
. . . is a man is a man
9x (—x—) is a man.
These are obviously not well-formed sentences, since they still contain a gap, and the
same continues to apply as we move further up the hierarchy.
We have seen that Russell himself gave just this line of argument in ‘The Substitu-
tional Theory’ (pp. 177–8) and in ‘On “Insolubilia”’ (pp. 201–2). He means to repeat it
in }4 of chapter 2 of the Introduction to PM, which is entitled ‘Why a given function
requires arguments of a certain type’.23 My explanation spoke of the ‘gaps’ which
expressions for propositional functions contain, whereas in this section Russell speaks
of the ‘ambiguity’ that is essential to a function. It would be fair to say that what we
each have in mind is that an expression for a function contains a free variable, and so the
function cannot occur in a proposition unless that proposition either contains some
higher-level function (such as a quantifier) which binds this variable, or the variable is
supplanted by a constant of suitable type. In Russell’s terms, the ‘ambiguity’ must be
eliminated if a genuine proposition is to result. Moreover, he clearly implies that in
x^ is a man is a man
the ambiguity is not eliminated, evidently because there is here nothing to bind the
variable ‘x’ (p. 48).24 There are times when he shows quite clearly that his cap notation

22
An open sentence is like a sentence, except for containing one or more free variables. It becomes a
proper sentence when either those variables are replaced by constants of the appropriate type, or they are
bound by prefixing suitable quantifiers. (In PoM Russell seems to make distinction between a propositional
function, which does contain a variable to mark its gaps, and the ‘functional part’ of that function which
contains no such variables but only gaps (cf. }85). The distinction is emphasized by Klement [2004], but it
seems to me to be pointless.)
23
One should regard this section as applying to the simple theory of types. Of course Russell elsewhere
invokes the vicious circle principle to justify his type restrictions, and that is the appropriate procedure for the
ramified theory. But in this section Russell means by ‘type’ just what I mean by ‘level’. Compare Dummett
[1973: 50–2].
24
His explicit remark concerns the pseudo-sentence ‘x^ is a man’, but it is clear that he is imagining some
definite predicate in place of the schematic ‘ . . . ’.
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 59

is not intended to turn a predicate into a name, grammatically suited to function as a

subject expression to a first-level predicate.

4 The elimination of classes

In ordinary English there are many ways of nominalizing what starts as a predicate, but
Russell’s focus is on mathematics, and in the vocabulary of mathematics there is only
one nominalizing technique that is at all widespread, namely that which prefixes ‘the
class of . . . ’. But although Russell’s theory now admits propositional functions, it is still
a ‘no-class’ theory, and contains no such nominalizing device. It does introduce a
notation which looks as if it refers to classes. Using modern symbols, the definition is
that formulae containing class descriptions

(—{x:Fx}—)
are to be taken as short for the associated formulae25

9G (8x(Fx $ Gx) & (—G—))

As with the somewhat similar definition of definite descriptions (p. 36), this ruling
should officially be supplemented by some way of indicating the scope of the class
description, i.e. some indication of how much of the surrounding context is repre-
sented by the dashes displayed here. But since (a) Russell’s class descriptions are always
satisfied, and (b) what he says of a class under one description will always apply equally
to that same class under any other description,26 it will turn out that all ways of
assigning the scope are equivalent to one another. So in fact Russell omits any such
indication, and we may follow him in this.
Setting this point aside, there is a clear similarity between this definition of class
descriptions and the familiar definition of definite individual descriptions. In that case,
what appeared to be a reference to a particular individual was replaced by a quantifier
ranging over individuals, and in the present case what appears to be a reference to a
particular class is replaced by a predicate quantifier. The quantifiers are different, but
the way in which they are used is in each case much the same. However, the difference
in the quantifiers is important, because it alters the ontological moral to be drawn. The
theory of definite descriptions assumes the existence of individuals. Some expressions
which apparently refer to individuals are ‘re-parsed’ as expressions which quantify over

25
In the PM version the variable ‘G ’ is here restricted to range only over what Russell calls ‘predicative’
functions of x. This is a complication required only by the ramified theory of types, which is the subject of the
next chapter.
26
That is to say: the predicates which Russell will apply to classes are all extensional predicates, and his
definition is only intended to fit this case. When non-extensional predicates are included, the definition
yields some odd results, as is pointed out by Carnap [1956: 147–50], and one must take explicit notice of
the scope (as Russell recognizes on PM, p. 84). (Carnap’s point is repeated in Smullyan [1958: 240–1], and
further elaborated in Boër [1973].)
60 L O G I C A N D M AT H E M AT I C S

individuals, but in each case the relevant individuals must exist if what is said is to be
true. However, the superficially similar theory of classes does not in any way assume
the existence of classes, as ordinarily conceived. For classes are ordinarily conceived as
objects, of the same logical type as more ordinary individuals, and usually named by a
predicate nominalization. But PM does not assume the existence of any such objects,
either named or quantified over. It makes use of a notation which looks just like the
usual notation for speaking of classes, but this is interpreted as a somewhat roundabout
way of speaking of something else, namely propositional functions. Both in the case of
ordinary definite descriptions and in the case of class-descriptions it may be said that
these symbols are introduced as incomplete symbols, which have no meaning in
isolation and which disappear upon analysis. But the first analysis has no effect on the
ontology that is being assumed, whereas the second does. For what appears to be a
reference to a class is actually construed as introducing a quite different kind of item.
Or perhaps not entirely different. As we have seen,27 Russell is apt to construe the
classes which he accepts as plural items, and one could say that quantification over
propositional functions is the right way to represent in logic what in English are plural
quantifications over ordinary objects. This is an idea that does have some support, at
least in the simple case of monadic propositional functions of first level, though it
cannot plausibly be extended either to dyadic functions or to functions of higher
levels.28 But Russell does so extend it, and so one has to admit that his quantifications
over propositional functions go some way beyond what can be said in English without
introducing nominalizations. His claim is that one does not need classes as well as
propositional functions, but he does need the propositional functions.
However, it is convenient to introduce a notation that looks just like the usual
notation for classes, because that is already familiar amongst mathematicians, and easy
to work with. Although Russell does not assume an axiom of extensionality for his
propositional functions,29 still his definition does allow him to deduce the usual
principle of extensionality for classes, namely30

8x(Fx $ Gx) $ {x:Fx} = {x:Gx}

He also introduces a symbol which functions like the relation 2 of membership, so
that by definition we have

a 2 {x:Fx} $ Fa

27
Note 2 of page 48.
28
See Boolos [1984] for a persuasive recommendation of this way of construing quantification over first-
level functions. (For a criticism of Boolos, see e.g. Linnebo [2003].)
29
I shall explore the implications of this point in section 6, below.
30
The deduction assumes that identity is defined for predicate letters. Russell’s definition is
F=G for 8M (M xFx $ M xGx)
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 61

Finally, he adds a style of variable, namely ‘a’, ‘b’, . . . , which functions like a variable
used to generalize over classes, though it is explained just as a variant on a predicate
variable, so that

8a(—a—a—)
is just short for

8F(—{x:Fx}—{x:Fx}—)
Thus everything looks just as it does in an expected theory of classes, but it is all done
without assuming the existence of classes, and as a roundabout way of speaking of
propositional functions.
I add that Russell makes provision for a similar theory of two-place relations, which
he calls a theory of ‘relations in extension’. It begins in just the same way, introducing

(—{xy:Fxy}—)
as a short way of rewriting
9G(8xy(Fxy $ Gxy) & (—G—)).
It also continues in the same way, deducing a principle of extensionality for these
relations, and adding both an analogue to the 2 of membership and variables that
appear to generalize over these ‘relations in extension’. It would perhaps be appropriate
to use the similar title ‘(monadic) functions in extension’ for the no-class theory that we
began with: it looks just like an ordinary theory of classes, and it functions like one,
except insofar as its rules of formation follow the restrictions of the simple theory of
types. (For example ‘x 2 x’ is not well-formed, just because it could only abbreviate a
formula containing a clause ‘F(F)’, which would not be well-formed.) But it does not
really assume the existence of classes. It is a theory of propositional functions, treated as
if their extensions were all that mattered, and Russell does the same thing both for
monadic and for dyadic propositional functions.
Let us now look at a detailed presentation of this simple theory of types, giving more
detail than Russell himself ever gives. But although the details do become complex, the
ideas—I think—are no more complex than what we have had already.

5 A formal presentation
The theory of types, as Russell envisages it, includes variables of all finite levels.
We have so far mentioned variables ‘x’, ‘y’, . . . which range over individuals of
level 0, variables ‘F’, ‘G ’, . . . which range over first-level predicates, i.e. predicates
of individuals, and variables ‘M ’, . . . which range over second-level predicates, i.e.
predicates of first-level predicates. It is not practical to proceed further in the same way,
introducing infinitely many further alphabets of variables, with a separate alphabet for
62 L O G I C A N D M AT H E M AT I C S

each level. Though we may usefully think in terms of different alphabets, since the
entities represented are of quite different types from one another, still for practical
purposes it will be simpler to use a single style of variable throughout, say ‘x’, ‘y’, . . . ,
with the different levels represented by different indices superscripted to the variables.
The three original alphabets can then be rewritten as the alphabets ‘x0, y0, . . . ’, and
then ‘x1, y1, . . . ’, and then ‘x2, y2, . . . ’. Clearly we can continue this plan through
all the numerical superscripts, without having to invent a new style of lettering in
each case.
But a proper presentation of the theory must be more complex than this, for the
superscripts just considered would be suitable only for the monadic predicates. But
there are also dyadic predicates at each level, and moreover dyadic predicates of mixed
level, i.e. those that take one argument at one level and the other at another. The same
applies to triadic predicates, and so on. A scheme that is theoretically elegant is to define
the type indices in this way:31
0 is a type index
If n, . . . , m are type indices, so is (n, . . . , m)
On this scheme the types of the first-level predicates are represented by
(0), (0,0), (0,0,0), . . . ,
with the number of 0’s showing the number of arguments to the predicate. The types
of second-level predicates are

Monadic: ((0)), ((0,0)), ((0,0,0)), . . .

Dyadic: ((0),(0)), ((0),(0,0)), ((0,0),(0)), ((0,0),(0,0)), . . .
Triadic: ((0),(0),(0)), . . .

The types of some mixed-level predicates are

(0,(0)), (0,(0,0)), ((0,0),((0))).
This scheme provides a distinct type index for every distinct type that is recognized in
the simple theory. But when we are concerned only with the monadic types, it will be
convenient to return to the genuine numerals as our type indices, by writing ‘1’ for
‘(0)’, ‘2’ for ‘((0))’, and so on.
The formation rules will require that a variable with type index (n, . . . , m) must be
followed by the appropriate string of arguments with type indices, n, . . . , m respec-
tively. But in view of the earlier discussion of the example ‘Trans((0,0))’ we may allow as
arguments not only a further variable with a suitable type index, but also a more
complex term representing an argument of some specified structure (e.g. the
relation which holds from x to y when and only when xRy & Gy). As observed

31
The scheme is due to Carnap [1937: section 27]. To present it, I use ‘n’, ‘m’ as schematic expressions for
any type index, even though the type indices are no longer just the familiar numerals ‘0’, ‘1’, ‘2’, . . . .
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 63

earlier, for this purpose Church’s º-notation is technically convenient, so our rules of
formation will include provision for his º-abstracts as complex terms.32 The basic rules
are, for any variables ‘v’, ‘u’, . . . , for any terms ‘t’, ‘s’, . . . , and for any type indices
‘n, m’, . . . :
v0 is a term t0 with index 0.
If tn, sm, . . . are terms with indices as shown, then
v(n, m, . . . )(tn, sm, . . . ) is a formula.
If A(vn, um, . . . ) is a formula, containing the variables vn, um, . . . free,33 with
indices as shown, then [ºvn um . . . : A(vn, vm, . . . )](n, m, . . . ) is a term t(n, m, . . . ) with
index as shown.
First, note that these rules make no provision for schematic letters as opposed to
variables. This is because all our letters can now be bound by quantifiers, and it is
therefore convenient to use the same letters in each case, whether they occur bound or
free. A formula which, in the symbolism of Chapter 1, would have been written with a
schematic subject letter ‘a’, or a schematic predicate letter ‘F’, is now written instead
with a variable ‘x0’ or ‘x1’ which occurs freely throughout. As noted on p. 13, this does
allow for a much simpler statement of the rules of formation.
Second, with these rules the predicate position in any formula is filled by a
predicate variable, and the following subject positions are filled by a string of one
or more terms, where the terms are either individual variables or º-abstracts. It
follows that the type indices are in a way superfluous, and could be omitted, since
they can be uniquely recovered from the structure of the formula. But in practice it is
convenient to use the indices, for example by equating the simple term [ºx0: y1(x0)]1
with the variable y1, so that we can interchange the two in subject position, and allow
ourselves to write

z2(y1) for z2([ºx0: y1(x0)]1)

More generally, we shall allow ourselves to abbreviate any º-abstract of the simple form
[ºxn ym . . . : z(n, m, . . . )(xn, ym, . . . )](n, m, . . . )

32
As noted earlier, we may if we wish think of the formula
Trans((0,0))[ºx0y0: z(0,0)(x0, y0) & w(0)(y0)](0,0)
as containing its major break, between the second-level predicate and the first-level complex argument, at the
colon, and not at the opening square bracket. But however we prefer to look at it, the º-term should not be
thought of as taking the place of a reference to an individual, so that ‘Trans’ now becomes a predicate of first
level. That is not how Russell’s theory of types should be understood.
33
An occurrence of a variable vn is bound in a formula or a term if and only if the occurrence is inside the
square brackets of a term [º . . . vn . . . : —] which has vn in its º-prefix; and which itself occurs within the given
formula or term. Occurrences which are not bound are free.
64 L O G I C A N D M AT H E M AT I C S

just to its embedded variable34

z(n, m, ...)

At the moment, this applies only to the abbreviation of terms in subject position, for
that is the only position that they occur in. But we shall soon provide for terms in
predicate position too.
The three formation rules stated so far cover only the atomic formulae of the system,
for nothing has yet been said of how to build more complex formulae from these. To
remedy this, I now add formation rules for the truthfunctors and the quantifiers:
If A, B are formulae, then so are
~A, A&B, A∨B, A!B, A$B
If t(n) is a term, then the following are formulae
8t(n), 9t(n)
(Note that according to this last rule, what is normally written as ‘8x(—x—)’ is now
written as 8[ºx0: —x0—](0). If desired, the usual notation can be restored by abbrevia-
tion.) This completes the rules of formation.35
Inference rules for the truthfunctors may now be given in any standard way, and
readers should supply whatever they are familiar with. In case none are, I simply repeat
here the rules given in Chapter 1, namely three axiom schemata
Ax 1: ‘ A!(B!A),
Ax 2: ‘ (A!(B!C)) ! ((A!B) ! (A!C)),
Ax 3: ‘ (~A!~B) ! (B!A);
and the rule of Detachment

Det: If ‘A and ‘A!B, then ‘B.

Inference rules for the quantifiers match those given in Chapter 1, except that they are
now repeated for variables with all type indices. Where every occurrence of vn that is
free in A(vn) is replaced by a term tn to form A(tn), we have

Ax 4: ‘ 8[ºvn: A(vn)](n) ! A(tn);

where vn does not occur free in B we have
Ax 5: ‘ 8[ºvn: (B!A(vn))](n) ! (B !8[ºvn: A(vn))](n);

34
This should be understood as an inductive definition. Assuming that the variables ‘xn’, ‘ym’, . . . are
already understood in subject position, it tells us how to construe ‘z(n, m, . . . )’ in subject position.
35
There is a natural generalization which is now tempting: why not allow 8 and 9 to be prefixed to any
term with a º-prefix, no matter how complex its prefix is? On this understanding ‘9[ºx0 y0: —]’ would
represent what is normally written as ‘9xy(—)’, and this would then be primitive notation, and no longer an
abbreviation for ‘9x9y(—)’. I refrain from this generalization only in order to stay more closely in line with
the rules for elementary logic given in Chapter 1.
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 65

and the rule of generalization says that, wherever vn occurs free in A(vn)
Gen: If ‘ A(vn), then ‘ 8[ºvn: A(vn)](n)
As written here, these three principles adopt the new syntax for 8, but otherwise they
simply generalize to all types the principles proposed for the first-level quantifiers in
Chapter 1.36 However, there is a complication affecting the axiom schema 4 which we
must now deal with.
First the idea is that the term tn be substituted for all occurrences of vn that are free in
A(vn). But, since our terms may now be complex expressions, they may now contain
their own free variables, and the result of the substitution could be that a variable which
is free in tn inadvertently becomes bound by a º-prefix already present in A(vn). To
prevent this, let us add the condition that no variable free in tn also occurs bound in
A(vn). Where the condition is not satisfied, we must first rewrite the relevant bound
variables of A(vn) by using a different letter, which can always be done without loss.37
This is a very simple point, but the next is more complex. In the official notation the
variable vn will always occur in predicate position in the formula A(vn), except in the
case when n=0. But in the official notation the term tn is not allowed to occur in
predicate position, and so the result of simply substituting tn for vn will not usually be a
well-formed formula. However, we may always reduce it to a well-formed formula by
steps of what is called ‘concretion’, which I now describe.
Assuming that n6¼0, our variable will actually have a type index ‘v(n, m . . . )’. Let us just
take the simplest case, where the variable is just ‘v(n)’, and so occurs in contexts of
the form
v(n)(sn).
The term t(n) to be substituted for it will be of the form [ºun: B(un)](n), so the result of
the substitution will be

[ºun: B(un)](n)(sn).
The operation of concretion simply replaces this by

B(sn),
and the result is that the displayed occurrence of a term in predicate position has
disappeared.38 Of course there may still be other occurrences of this term in predicate
position elsewhere in the formula A, and in particular there may be such an occurrence

36
The three principles are easily generalized to allow the quantifier 8 to govern terms with a complex
º-prefix, containing several variables, possibly of different types. See the previous footnote. (This merely
simplifies proofs. It does not add any new theorems.)
37
This is called ‘alphabetic change of bound variable’. The change makes no difference to the truth
conditions of the formula, as is easily proved.
38
It will be obvious how this account may be extended to the more general case of a variable and a term
with a more complex type index (n, m, . . . ). I omit the details.
66 L O G I C A N D M AT H E M AT I C S

within the term sn that has survived this particular step of concretion. But by
continuing with further steps of concretion a well-formed formula must eventually
be reached. I observe that in the simplest case B(sn) just attaches a predicate variable, say
y(n), to the term sn. The step of concretion then shows the interchangeability of the
variable y(n) and the term [ºxn: y(n)(xn)](n) in predicate position. We have already
observed (on pp. 63–4) their interchangeability in subject position.
My own taste is to stick with the official formation rules that were given first,
whereby all variables except those with index 0 always occur in predicate position, and
all terms always occur in subject position. (This accords with Frege’s principles.) In that
case our axiom schema 4 must be taken as saying that when a term is substituted for a
universally quantified variable, and the result is reduced to a well-formed formula by
concretion, then the formula so formed is an axiom. But we have seen that concretion
does allow us to understand terms in predicate position, and we have earlier observed
that variables may also be understood in subject position, so it would certainly be
possible to adopt more liberal formation rules in the first place. So long as suitable rules
of concretion are then added as explicit rules for inferring one formula from another,
we would still have effectively the same system, i.e. just the same formulae would be
either theorems or abbreviations of theorems.
It is obvious that there are yet other ways of varying the rules and axioms given here,
while still preserving what are in effect the same theorems, and any of these may be
taken as a presentation of Russell’s simple theory of types. (As is often noted, Russell’s
own presentation gives far too little detail. He does not mention any scheme for type
indices, he does not say anything of the rules for concretion, and in general he leaves a
lot unsaid which we would now insist ought to be said. But there is no serious doubt
about what the theorems of the system are intended to be.) I mention here one quite
common way of presenting the theory which in fact is not equivalent to the one just
given. Axiom 4 may be restricted so that it merely allows one to replace a universally
quantified variable by a free variable (of the same type), if at the same time we add a
separate rule of substitution, thus
If the variable un occurs free in A(un) wherever vn occurs free in A(vn), we
adopt the axiom schema
Ax 4(-): ‘ 8[ºvn: A(vn)](n) ! A(un)
And we add, under the same conditions as our axiom 4
Sub: If ‘ A(vn), then ‘ A(tn)
In the presence of the rule of generalization, these are together equivalent to axiom 4.
But sometimes the theory is presented with axiom 4 weakened to 4(-) as shown, and no
rule of substitution, but instead what are called comprehension axioms, thus: wherever
the variables u1n, u2m, . . . do occur free in A(u1n, u2m, . . . ), but the variable v(n, m, . . . )
does not,
Comp: ‘ 9v(n, m, ...)
8u1n, u2m . . . (v(n, m, . . . )(u1n, u2m, . . . )) $ A(u1n, u2m, . . . )
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 67

It is easy to see that this can be deduced in the system which I have given, but the
converse deduction is available only if suitable axioms of extensionality (for every level)
are also adopted.39 Without such axioms, one could not prove

9v(n, m, . . . )(v(n, m, . . . ) = t(n, m, . . . ))

However, in the system that I have given, which does include the substitution rule, this
is a simple consequence of the trivial theorem

‘ (u(n, m, . . . ) = u(n, m, . . . ))
The system I have given does not include axioms of extensionality because in PM
Russell takes such axioms to be incorrect. Consequently, I mention this variation only
to set it aside.
We may accept as a presentation of Russell’s simple theory of types any set of rules of
formation and rules of proof that yields all the same theorems as the one just given, for
we must apparently count these theorems as defining what the theory is. Our discus-
sion of elementary logic in Chapter 1 proceeded differently, for there we began with
the distinction between those formulae that are valid and those that are not, but Russell
pays very little attention to this approach. Let us now turn to the problems that this
involves. (And, since we shall no longer need to consider the more complex formulae
of the full theory, we may for simplicity return to the simple Fregean notation that we
began with. That is, we shall revert to unindexed variables x, y, . . . for individuals,
unindexed variables F, G, . . . for predicates of individuals, and the variable M for
predicates of these in turn.)

6 What is missing: Validity and extensionality

In Chapter 1 I explained that validity is a matter of truth under all (permitted)
interpretations, and the permitted interpretations may be stated thus. The sentence
letters (if any)40 are interpreted by assigning to each of them a truthvalue—any
truthvalue—whereas the truthfunctors are to be interpreted in the intended way, as
given by their truthtables. When we come to quantifiers, we begin by picking any
(non-empty) domain of objects for the quantifiers to range over. Then a free individual
variable x is to be interpreted as denoting some member of that domain. A free one-
place predicate variable F is to be interpreted as true of some (or none) of the members
of that domain, and false of the rest. Similarly, a free two-place predicate variable is to
be interpreted as true of some ordered pairs of members of the domain, or of none, and
as false of all the rest; and similarly again, mutatis mutandis for predicates of three or more
places. Proceeding in the same way, a free one-place variable M of the next level up

39
These axioms are given in the next section.
40
In the theory as just given there are none. As Chapter 12 will show, this conforms to Russell’s intentions
in PM.
68 L O G I C A N D M AT H E M AT I C S

would be interpreted by saying which interpretations of F it is to count as true of, and

so on and on. Finally the quantifiers 8 and 9 are to be held to their intended
interpretations (given the choice of domain). Thus 8a(—a—) is to count as true if
and only if (—a—) is already interpreted as true for every way of interpreting its free
variable a on that domain, and 9a(—a—) is to count as true if and only if (—a—) is
already interpreted as true for some way of interpreting a on that domain. As we have
said, different things will count as interpretations for the different variables a of
different types, but—given this—the explanations of 8 and 9 remain the same for
variables of all types. That, one might say, outlines the notion of validity for our simple
theory of types. But we hear almost nothing of this from Russell, for two main reasons.
The first is simple and straightforward: he does not usually think of his formulae as
having to come out true for every choice of a (non-empty) domain. On the contrary, he
normally assumes that there is just one choice of domain that is the intended one,
namely the domain of ‘all individuals’. As for what is to count as an individual, in his
logical writings he quite often gives a merely negative interpretation. For example in
PM we find the definition: ‘we say that x is “individual” if x is neither a proposition nor
a function’ (p. 132; similarly p. 51).41 It is probably true that all that his formal position
needs is just this negative claim, for so long as that is so it will be possible to fulfil
the requirements of his type theory. However even the logical works do also contain
some more positive characterizations, such as this: ‘We shall define “proper names”
as those terms which can only occur as subjects in propositions . . . We shall further
define “individuals” or “particulars” as the objects that can be named by proper names’
(IMP [1919a: 142]). In Chapter 7 we shall look into the very definite account that
Russell gives in other writings of just what is to count as a ‘genuine proper name’, but
that can wait until we come to it. Meanwhile, the point is that Russell usually thinks of
his logic as concerned with one particular domain, namely the domain of
‘all individuals’, and not with what must be true of any arbitrarily chosen domain.
But there are exceptions.
At the end of IMP [1919a] Russell is trying (and failing) to give some precise
characterization of just what deserves to be counted as a truth of logic. Concerning
the axiom that there are infinitely many individuals, he has earlier said that that axiom
may perhaps be true of this world, for it looks to be a consequence of our present
physical theory, which assumes the continuity of space and time. However he adds that
this is not enough:
But this is, as it were, an accident, a fact about the world in which we happen to find ourselves.
Pure logic . . . aims at being true, in Leibnizian phraseology, in all possible worlds, not only in this
higgledy-piggledy job-lot of a world in which chance has imprisoned us. There is a certain

41
Similarly in ML [1908: 76], an individual is defined as ‘something destitute of complexity’, from which
it is inferred that propositions are not individuals. (Contrast the substitutional theory, in which propositions are
taken to be individuals, and recall that ML officially endorses the substitutional theory (p. 77).)
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 69

lordliness which the logician should preserve: he must not condescend to derive arguments from
the things he sees about him. (IMP, p. 192)42

Here it looks as if only a very small push is needed to get him to admit that the truths of
logic should remain true whatever is chosen as the domain of all individuals. This would
certainly be a step towards our notion of validity.
But there is another point at which he will surely want to dig in his heels: his logic is
designed to accommodate propositional functions that are intensional, whereas our
concept of validity is committed to extensional methods throughout. This needs
explanation.
The extension of a name is the object which it denotes; the extension of a predicate
is the objects that it is true of (or, for a two-place predicate, the ordered pairs of objects
which it is true of, and so on for any number of places); the extension of a sentence is its
truthvalue. A linguistic context is said to be an extensional context if expressions in that
context which have the same extension may always be substituted for one another
without changing the truthvalue of the whole. That is, a context ‘C ( . . . )’ for names or
predicates or sentences (as appropriate) is extensional if and only if
a = b ! (C (a) $ C (b))
8x(Fx$Gx) ! (C (F) $ C (G))43
P$Q ! (C (P) $ C (Q))
A context which is not extensional is said to be intensional.
Let us begin with contexts for names, i.e. contexts which are apparently first-level
predicates. In orthodox logic one always assumes what is often called Leibniz’s Law,
namely

a = b ! (Fa$Fb)
Indeed, it is clear that this supposed law must count as valid, according to our standard
account of validity. For if ‘a = b’ is interpreted as true, then ‘a’ and ‘b’ must be
interpreted as denoting the same object, and so if ‘F’ is interpreted as true of a then
it must also be interpreted as true of b, and conversely. However, there are plenty of
apparent exceptions to this supposed law. To use a hackneyed example: even though
Jocasta was Oedipus’s mother, and Oedipus knew that he was marrying Jocasta, it does
not follow that Oedipus knew that he was marrying his mother. So apparently one
must say that the context ‘Oedipus knew that he was marrying . . . ’ is not an extensional
context for names. The same verdict apparently applies to several other contexts which

42
The idea that the truths of logic should be true in all possible worlds has occurred earlier in KEW
[1914c: 189–91]. Using Wittgenstein’s vocabulary, it is rephrased as the idea that they should be tautologies,
but in PLA [1918b: 240], in IMP [1919a: 203–5], and still in the second edition of PoM [1937: xii], Russell is
admitting that he does not know how to define ‘tautology’. It is worth adding that later in IMP (p. 203) he
observes that the logic of PM does assume that there must be at least one individual, but he adds: ‘I now view
this as a defect in logical purity’. His thought is surely moving in the right direction.
43
Similarly for predicates of two or more places.
70 L O G I C A N D M AT H E M AT I C S

concern people’s mental attitudes—e.g. their beliefs, doubts, fears, and so on—and to
some other examples too, principally to those concerning what is necessary or possible
or probable. The solution which is now almost universally accepted is to say that, for
the purposes of elementary logic, we shall not accept such an intensional context as
being one of the first-level predicates which that logic is concerned with. That is, we
simply admit that the logic is not applicable in such a case. As we shall see more fully
later (pp. 223–5) Russell takes a different view. He does accept that Leibniz’s Law must
be counted as correct, but he deals with the apparent exceptions in a different way, i.e.
by being more demanding on what is to count as a name. The basic idea can be seen by
noting that the Oedipus example makes use of what he has claimed in OD to be a
definite description, and therefore not a name, namely ‘Oedipus’s mother’. The
example is similar in this way to Russell’s own example in OD, which applied the
context ‘George IV wished to know whether Scott was . . . ’ to the definite description
‘the author of Waverley’. As a result, Russell does not accept that there are intensional
contexts for genuine names, but he is able to do this only because he has a very narrow
account of what does qualify as a genuine name. We shall come to this in Chapter 7.
He does accept that there are intensional contexts for (first-level) predicates, and he
wishes his logic to include them. The supposed law that would deny this is often called
the axiom of extensionality, namely

8x(Fx$Gx) ! (MxFx$MxGx)
It is the analogue of Leibniz’s Law at one level up, and it must be counted as valid on
the most natural way of extending the concept of validity from a first-level to a second-
level logic. For just as the interpretation of a first-level predicate letter is a specification
of which objects of the domain it is true of, i.e. of which ways of interpreting a name‐
letter it is true of, so one expects that the interpretation of a second-level predicate letter
should just be a specification of which ways of interpreting a first-level predicate
letter it is to count as true of. On this view of what counts as an interpretation, it is
clear that the axiom of extensionality must always be interpreted as true, and so must be
counted as valid. But, again, there are numerous apparent exceptions, mostly drawn
from the same kind of contexts as before, i.e. those concerning mental attitudes or
modalities. (One of Russell’s favourite examples is that, even if in fact it is all and only
the things that are human beings that are rational animals, still someone x may not
believe this. So the context ‘x believes that all rational animals are . . . ’ will be
intensional.44) A common view these days would be that one should accept the
axiom of extensionality, and therefore should not count such intensional contexts as
being second-level predicates for the purposes of our second-level logic. But Russell
takes the opposite view. He does accept these intensional contexts as genuine second-
level predicates, and therefore rejects the axiom of extensionality. However, a conse-

44
See e.g. ML, p. 89; PM, p. 73; IMP, p. 186.
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 71

quence of this is that it becomes quite unclear how to apply the concept of validity to
his preferred logic. For to apply this concept we need to know what it means to say that
something comes out true for all ways of interpreting the letter M, but if intensional
interpretations are permitted then how is this to be understood? The only suggestion
seems to be that an interpretation must be a string of words (or other symbols) that does
in fact express a second-level predicate in some actual language, but this will lead to any
number of problems, as we shall see when this question is resumed in Chapter 6. The
alternative seems to be to follow Russell’s own practice, and not to try to introduce a
suitable concept of validity for his logic. It will be defined in the way which we
explored in the previous section, by an accurate statement of its rules of formation and
rules of proof.
As Russell could not have known, but as we do know now, this is not an adequate
approach, for a higher-level logic has no complete proof procedure. The result applies
as soon as even first-level predicate letters are permitted to be bound by quantifiers.45
The rules that we gave in the last section, as conforming to Russell’s intentions, are
perfectly sound rules: on any reasonable conception of validity, one can prove from
them only what is valid. But they are not a complete set of rules, and no addition could
make them complete. According to any reasonable conception of validity, there will
always be formulae that must be counted as valid but which are not provable from
whatever rules of proof have been specified. This is a point which outlines the
importance of having, for one’s logic, not only some rules of proof but also a workable
concept of validity. One is tempted to suggest that it was largely Russell’s rejection of
the axiom of extensionality that prevented him from seeing his way to such a
concept.46
This leads to a gap in Russell’s account, which is definitely important. Under what
conditions are we to count as true some generalizations about ‘all propositional
functions’ (of a specified logical type)? What propositional functions are there? If
there are only those that can be expressed in some specifiable language, then in
principle one could aim to give what is called a ‘substitutional’ interpretation of the
quantifiers that seem to range over them, in which one actually quantifies over the
expressions.47 Some things which Russell says strongly suggest this idea, though

45
This is a consequence of Gödel’s incompleteness theorem for second-level logic, which was proved in
his [1931].
46
It is true that in the second edition of PM, in 1925, Russell does propose adopting some axioms of
extensionality and dropping a different axiom (namely the axiom of reducibility, which we shall come to in
section 5 of the next chapter). But the 1925 system is not well worked out, and I shall not explore it. (It is
explored in Hazen and Davoren [2000].) See further note 4 to Chapter 6.
47
That is: ‘9a(—a—)’ is counted as true if and only if there is an expression of the appropriate kind,
which, when substituted for the variable ‘a’, makes the sentence ‘—a—’ come out true. To make this
criterion effective, one has to impose some principle such as Russell’s Vicious Circle Principle—to be
discussed in the next chapter—which restricts the kinds of expression permitted. For the general situation
see Kripke [1976]; the idea is applied to Russell in Sainsbury [1979: ch. 8, pp. 283–9], and in full detail in
Hazen and Davoren [2000].
72 L O G I C A N D M AT H E M AT I C S

some are incompatible with it, as we shall see. The problem will continue to occupy
us, both in the next chapter and beyond. Meanwhile I end this chapter with some
further reflections on the nature of Russell’s (simple) theory of types, as he himself
understood it.

7 Concluding remarks
He certainly saw that what keeps this theory free from paradoxes is that it provides no
way of naming what is first expressed by an expression of a different type. For example,
in the lectures on ‘The Philosophy of Logical Atomism’ [1918a], when he is endea-
vouring to explain what it is to understand a predicate, he says: ‘A predicate can never
occur except as a predicate. When it seems to occur as a subject the phrase wants
amplifying and explaining’ (p. 205; similarly for relations, p. 206). The needed
‘amplification’ would be one that shows how the original claim could be rephrased
so that the predicate no longer seemed to occur as a subject.48 In a similar vein he has
earlier said that propositions do not name facts, and added: ‘You must not run away
with the idea that you can name facts in any other way; you cannot. You cannot name
them at all . . . You can never put the sort of thing that makes a proposition to be true or
false in the position of a logical subject’ (p. 188, compare pp. 269–70). In both places he
should be understood as speaking of what can be done in the ‘logically perfect’
language which these lectures concern, a language which has the syntax of PM
(pp. 197–8). In ordinary English you can certainly transform a predicate or a sentence
into a noun phrase, e.g. by using such expressions as ‘the property of being wise’ or ‘the
fact that Socrates was wise’, and in PoM Russell had insisted that this must always be
possible. But it is true that you cannot do that in the language of PM. Since Russell is
writing these lectures not in that language but in English, he is constantly himself doing
these things which he claims ‘cannot be done’ (as in the sentence just quoted from
p. 188). But the thought is that it is just because such linguistic constructions are freely
available in English that we so often find ourselves led into contradictions. He gives just
this diagnosis himself in ‘The Philosophy of Logical Atomism’: ‘The trouble that there
is arises from our inveterate habit of trying to name what cannot be named’ (p. 267).
Naturally, the ‘logically perfect’ language will be one in which such contradictions
cannot arise, so it will contain no means of ‘nominalizing’ an expression that is not
already a noun.49

48
Similarly in the later article ‘Logical Atomism’ [1924]: ‘All propositions in which an attribute or a
relation seems to be the subject are only significant if they can be brought into a form in which the attribute is
attributed or the relation relates. If this were not the case, there would be significant propositions in which an
attribute or a relation would occupy a position appropriate to a substance, which would be contrary to the
doctrine of types, and would produce contradictions.’ (pp. 337–8)
49
It is claimed by Cocchiarella [1980] that these claims about what one cannot say are due to the influence
of Wittgenstein, and therefore should not be read back into Principia, which was composed before Russell
met Wittgenstein (pp. 109–11). But in my opinion many of the views that Russell generously credits to
Wittgenstein were views that he had already reached himself, or were implicit in what he had already
 THE CONTRADICTION ( I I ): A SIMPLE SOLUTION 73

The distinction between ordinary English on the one hand, and the language of the
theory of types on the other, is relevant to a point which is sometimes raised in
objection to the theory, namely that the theory cannot be stated without infringing
itself.50 This is usually understood, in effect, as the point that you cannot state the
theory in English while conforming to its own rules. But that is in no way surprising,
and not in any way a serious objection. The two languages are different languages,
conforming to different rules of grammar, and if this were not so one could not sensibly
claim that one is better than the other. What is perhaps a more pertinent point is that
there is no way of stating, in the language of the theory of types, the rules of that language
itself. To put this in different terms: the language does not contain its own semantics.
So, if we feel that those semantics need an explicit formulation, we have to ‘ascend’ to a
different language (a metalanguage) at a higher level. But (a) why should we think that
any explicit formulation is required? After all, we all learned to speak English without
learning any explicit semantics for it, and there seems to be no reason in principle why
the language of the theory of types could not be mastered in the same way. The learner
picks up the grammar of the language just by being constantly exposed to it. Wittgen-
stein in his TLP [1921] goes further, claiming that no semantics can be stated, for this is
one of the main thoughts behind his dictum: ‘What can be shown, cannot be said’
(4.1212). But (b) it seems better to say that, if an explicit semantics is required, then
Russell gave the right response in his introduction to the TLP (final paragraph), where
he says that metalanguages are surely possible, though of course they must be handled
with care, lest the contradictions break out once more.
I set this point aside. It is true that the language of the simple theory of types cannot
state its own semantics in any straightforward way, but that is not a serious objection.
Much more serious is the point that there are so many other things which cannot be
said in this language, and which we do wish to be able to say. But that is a point for
Chapter 6.

thought, before they ever met. And this particular view, that nominalizing leads to contradictions, was one
which he himself had canvassed way back in the Principles of Mathematics. He surely did not need to learn it
from anyone else.
50
The objection was raised both by Gödel and by Max Black in their contribution to Schilpp [1944]. It
was pressed again by Fitch [1952: 225] and by the Kneales [1962: 670]. There is a discussion in Copi [1971:
71–5], and a recognition that the objection is of no importance in Landini [1996a: }2].
 5
The Contradiction (iii):
A Ramified Solution

The simple theory of types discussed in the last chapter was not Russell’s final theory.
In fact it is a theory that Russell himself never formulated in print. As we have noted,
the early version that is outlined in Appendix B to PoM contained features that were
quite soon dropped (e.g. the acceptance of an infinite type); the version that is given in
his [1906a] and [1906b] as ‘the substitutional theory’ was soon abandoned as being too
permissive, since it adopts an unlimited quantification over all propositions whatever,
and that leads to contradictions; and the fact that the theory that is actually used in PM
is really just the simple theory is obscured, because it occurs there merely as the used
part of what is presented as a much more complicated theory. This more complicated
theory is nowadays known as Russell’s ‘ramified’ theory of types, as opposed to his
‘simple’ theory. But Russell himself never draws this distinction explicitly. It is first
made explicit in a commentary by Frank Ramsey (pp. 20–1 and 24–9 of his [1925a]);
Ramsey accepted the simple theory but rejected the ramified theory. The subject of
this chapter is the ramified theory, which is first stated in Russell’s article ‘Mathematical
Logic as based on the Theory of Types’ [1908], and the discussion will focus mainly on
this article (henceforward ‘ML’). But the somewhat different account that is given later
in PM [1910d] will also be considered.
This theory is based upon what is called the ‘Vicious Circle Principle’ (hereafter
‘VCP’). In fact the principle was introduced not by Russell himself but by an opponent
Henri Poincaré, who used it in his [1906] as an explanation of why the logics
introduced by Frege and the early Russell had run into contradictions. But we may
pass over Poincaré’s account (which is not altogether clear1), for Russell himself then
took over the principle, and his ramified theory of types aims to conform to his version
of it. In fact in his ML Russell states the principle somewhat vaguely, and in varying
terminology, thus:

1
Poincaré ran together his advocacy of the VCP and his rejection of the actual infinite. Russell very
reasonably protests, in his [1906c: 196–8], that the two ideas are quite distinct.
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 75

Whatever involves, or presupposes, or is only definable in terms of, all of a collection cannot itself be
one of the collection (e.g. pp. 63, 75; my emphasis)2

It is the version in terms of definitions which is mainly important for our purposes.
For definitions which infringe this principle are said to be ‘impredicative definitions’,
and the claim is that these must be avoided. We should add here another version of
the VCP, which Russell somewhat oddly regards as the same principle stated ‘con-
versely’:
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. (And he adds in a footnote: when I say
that a collection has no total, I mean that statements about all its members are nonsense.)
(p. 63, Russell’s emphasis)

(I will replace the word ‘nonsense’ by ‘illegitimate’, as Russell himself often does.) In
fact the second version elaborates on the first, and clarifies the supposed circularity in
such definitions. For it says that a quantification is illegitimate if it quantifies over a
totality which contains members that are definable only in terms of that totality itself,
i.e. by means of that quantification itself. The impredicative definitions are then those
that make use of a quantification that is illegitimate in this sense. In ML Russell offers
two reasons for accepting this VCP, first that it seems to be required to solve certain
paradoxes, and second that it has ‘a certain consonance with common sense’ (p. 59).
He adds that he finds the first reason more important than the second, since ‘common
sense is far more fallible than it likes to believe’, though one may well argue that
it should be the other way round. Anyway, let us begin with the first.

1 The VCP and the paradoxes

ML opens by listing a number of paradoxes. Here is a selection of four:3
(i) Russell’s own paradox, i.e. the attempt to define a set w by the definition

(8 sets x)(x 2 w $ x 2
= x)
(ii) Grelling’s paradox, i.e. the attempt to define an adjective ‘heterological’ by the
definition
(8 adjectives x)(‘het’ is true of x $ x is not true of x)

2
The earlier version in his [1906c], which responds directly to Poincaré, is perhaps a little clearer; it says
‘Whatever involves an apparent [bound] variable must not be among the possible values of that variable’ (p.
204).
3
As it happens, Russell’s list in [1908] does not actually include the second stated here, i.e. Grelling’s
paradox. This would seem to be entirely accidental. I include it here in order to exhibit its similarity with the
first.
76 L O G I C A N D M AT H E M AT I C S

(iii) Berry’s paradox: is there or is there not a number named by this name:4

The least integer not nameable in fewer than 19 syllables.

(The paradox arises because this alleged name itself has fewer than 19 syllables.) To
understand Russell’s response it is helpful to paraphrase with an explicit quantification
The least number n such that (8 names x)(x names n ! x has at least 19 syllables)
(iv) The Epimenides, which is the oldest one in the book. Epimenides was
a Cretan, and he said

All Cretans are liars

We assume that he meant by this
(8 propositions p)(A Cretan says that p ! ~p)
To obtain a paradox from this we add the supplementary assumption that everything
else said by any Cretan is indeed false.5
Russell’s diagnosis is that in each case there is an illegitimate quantification, i.e. a
quantification over a collection that ‘has no total’. I give just a couple of examples.
First, discussing a simplified version of the Epimenides, which is usually called the
paradox of the liar, Russell says
When a man says ‘I am lying’ we may interpret his statement as: ‘There is a proposition which
I am affirming and which is false’ . . . in other words, ‘It is not true for all propositions p that if
I affirm p, p is true’. The paradox results from regarding this statement as affirming a proposition,
which must therefore come within the scope of the statement. This, however, makes it evident
that the notion of ‘all propositions’ is illegitimate; for otherwise there must be propositions
(such as the above) which are about all propositions, and yet can not, without contradiction, be
included among the propositions they are about. (pp. 61–2)

Similarly, discussing Berry’s paradox, he says

‘The least integer not nameable in fewer than nineteen syllables’ involves the totality of names . . .
Here we assume, in obtaining the contradiction, that a phrase containing ‘all names’ is itself a

4
Although Russell writes ‘integer’, it is clearly the natural numbers that he is thinking of, and in particular
the principle that if any natural number has some property then there must be a least natural number with that
property.
5
Other paradoxes noted in ML are: Burali-Forti’s paradox (given on p. 26 above), a version of Russell’s
own paradox for sets applied this time to relations, and two that are similar to Berry’s paradox insofar as they
concern definable numbers. (The simpler is due to König [1905] and is this: there are too many ordinal
numbers for them all to be definable. So some must be indefinable. So there must be a least indefinable
ordinal. But we have just shown how this ordinal can be defined, namely as the least that cannot be defined.)
Russell also remarks that the seven paradoxes which he does list are ‘merely selections from an infinite
number’.
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 77

name, though it appears from the contradiction that it cannot be one of the names which were
supposed to be all the names there are. Hence ‘all names’ is an illegitimate notion. (p. 62)

He shows on p. 61 that he is aware of a connection between what is ordinarily called

self-reference and these quantifications which may be called ‘self-quantifications’,
because they allegedly introduce an object by means of a quantification over a
collection which includes that object itself. This does not actually lead him to modify
his statement of the VCP, no doubt because literal self-reference is not available in the
usual logical systems. But we do better to include it, for several of the paradoxes listed
do have an explicitly self-referential version. The best known of these is probably the
simpler version of the Epimenides that is known as ‘the liar’, i.e. the problem of one
who says

This statement is false.

Berry’s paradox can also be put in this form, i.e. by considering the alleged name

The least number not named by this name.

Let us suppose, then, that the VCP is intended to ban self-reference just as much as self-
quantification, as in each case ‘viciously circular’.
Attempts to obey the VCP have always led to hierarchies. The items that we are
concerned with—e.g. classes, adjectives, names, propositions—are each assigned an
order, and these orders are hierarchically arranged. It may not be at once obvious why
this should be, but the approach via the paradoxes will easily furnish an explanation.
For essentially similar paradoxes may easily be generated not by literal self-reference,
but by cases where one item refers to another and that other refers back to the first.
A simple example is
The next statement is true.
The previous statement is false.
This too has its overtly quantified version (which is due to Prior [1961]), as when
The prisoner says ‘Everything the policeman says is true’
The policeman says ‘Everything the prisoner says is false’.
And as it happens neither says anything else. (Or: as it happens everything else that the
policeman says is indeed true, or vice versa.) These are ‘circles’ of just two members.
It is obvious that longer circles can be generated in the same way. The simplest way
to rule out all such vicious circles, in reference or in quantification or in a mixture of
both, is indeed to introduce a hierarchy of orders, to insist that all reference or
quantification must be restricted to items of some one order, and that what ‘involves
or presupposes or is only definable in terms of ’ such a reference or quantification must
itself be of a higher order. Russell claims that this ruling will provide a ‘solution’ to all
the paradoxes that he is concerned with.
78 L O G I C A N D M AT H E M AT I C S

This claim is disputed, but we need not enter that dispute.6 One can at least say that
his theory does very often prevent a known contradiction from arising, and it is
probable that something along these lines could be made to work in all relevant
cases, given a small modification here and there. Instead I offer two different objec-
tions, namely (i) that the supposed solution seems to be extravagant, in that it proscribes
many ways of speaking which appear to be harmless, and (ii) that it is certainly not
the only solution that is available.
In support of (i) I note that there seem to be many quite harmless examples of
self-reference, e.g.

This sentence is in English,

and similarly of self-quantification, e.g.

All English sentences can be translated into French.

Russell might be inclined to reply that these examples are not examples of definitions
and it is really a certain kind of definition that the VCP aims to rule out. I will address
that topic more directly in what follows. Meanwhile, I add a brief consideration of the
objection (ii).
Ramsey [1925a: pp. 24ff.] observed that the simple theory of types is quite enough to
prevent paradoxes that might arise within set theory itself, such as Russell’s own
paradox, or e.g. the Burali-Forti paradox. Indeed, there are also other methods
of handling these, as we have seen, for example the ZF principle of ‘limitation of
size’ (pp. 27–8). What leads Russell into the complications of the ramified theory is
largely his wish to provide a solution to many other paradoxes, which Ramsey
characterizes as ‘semantic’ paradoxes, because they all involve semantic notions such as
being true, or being true of, or denoting, or something similar.7 Ramsey proposed that
the resolution of these paradoxes be left to semantics, a separate area of study which has
no close connection with basic logic, and nowadays this view is widely accepted. There
is no strong reason to suppose that one and the same theory should provide solutions
both to the paradoxes of set theory and to these semantic paradoxes. Since Russell was
writing there has been much further work on the semantic paradoxes, which I shall not
describe here.8 But it is fair to conclude that, even if something like the VCP is helpful in
these cases, that is not by itself a good reason for saying that it is needed in basic logic.

6
For some discussion see e.g. Copi [1971: 89–91 and 107–14], Sainsbury [1979: 320–5], and Giaquinto
[2002: 75–9]. Giaquinto draws a good distinction between two different kinds of ‘semantical paradox’.
7
In his own words, these contradictions ‘are not purely logical, and cannot be stated in logical terms
alone; for they all contain some reference to thought, language, or symbolism, which are not formal but
empirical terms’ [1925a: 20].
8
Church [1976] contains a favourable comparison between Russell’s way of avoiding the semantical
paradoxes and that adopted by Tarski, with his hierarchy of languages. But the similar discussion in Hazen
[1983: section 5], reminds us that there are alternatives to Tarski’s theory, e.g. that proposed by Kripke
[1975]. For a general discussion, very clearly expressed, see e.g. Sainsbury [1995: ch. 5].
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 79

So let us now turn to Russell’s other argument for the VCP, namely that it has ‘a certain
consonance with common sense, which makes it inherently credible’. (In what follows I
will, like Russell himself, ignore self-reference.)

2 The VCP and common sense

We must begin with some further clarifications of how the VCP is to be understood.
First, it applies only to abstract objects. Ramsey observed that if we specify a certain man
as ‘the tallest man in the room’ then we are specifying a man by quantifying over a
totality (namely all the men in the room) of which he is himself a member. But it
would be absurd to infer that such a man cannot exist, or that the quantification is
somehow illegitimate (Ramsey [1925a: 41]). Russell might reply that this is not the
only way of specifying the man in question, so there is here no counter-example to the
VCP, and no doubt we can grant that. But we must also generalize it. Ordinary
concrete objects may always be specified in a variety of ways, and in principle simply by
a demonstrative which involves no quantification at all. So the VCP does not apply to
them, but only to abstract objects, such as sets or adjectives or names or propositions.
A second point is that even such abstract objects may be specified in various ways,
e.g. as one might refer to a set as ‘the set mentioned on p. 22 of this book’. This is not
the kind of specification that is relevant to the VCP, which should be understood as
concerned with what one might call ‘canonical’ specifications of the object, i.e. those
that may well be regarded as definitions. In the case of sets, the canonical specifications
are those that cite a membership condition; in the case of propositions, those that cite a
that-clause which expresses that proposition, and similarly for propositional functions;
in the case of words or sentences, a quotation of that word or sentence, perhaps with an
explanation of its meaning. For simplicity we may mostly consider sets from now on,
but remembering that Russell’s ‘no-classes’ theory reduces talk of sets to talk of
propositional functions. What the VCP claims in this case is that there does not exist
any set such that all its canonical specifications are impredicative. The idea is that the
same set may well have different canonical specifications, i.e. there may well be
different ways of stating its membership condition, and so long as one of these is
predicative all is well. But if none are, then the supposed set does not exist.
Third, there is an ambiguity over what is to count as a predicative specification of a
set. Introducing a set w the specification will take the form

8x(x 2 w $ (—x—)),
and it will be predicative or not depending on the quantifiers involved. But
which quantifiers? Are we to consider only the quantifiers appearing within the clause
‘(—x—)’ on the right-hand side, or do we also include the initial quantifier ‘8x’? The
usual understanding is that it is only the quantifiers within ‘(—x—)’ that matter, and
the constraint is that they must range only over items of an order less than the order of
the set w hereby introduced. For the most part Russell too follows this interpretation.
80 L O G I C A N D M AT H E M AT I C S

But apparently he is committed to saying that the order of the initial quantifier ‘8x’
must also be taken into account, for he claims that the definition of the Russell set, i.e.
8x(x 2 w $ x 2
= x),
is impredicative, and infringes the VCP. But this definition has no quantifiers on the
right-hand side.
This creates a difficulty, for if the variable ‘x’ is restricted to items of some definite
order, say order n, then the attempted specification of the set w tells us only which
items of order n are members of it, and leaves it open whether it also has members of
order n+1. (This is by no means impossible, as we shall see.) Russell’s response might be
that the definition of w will also specify its order, say as n+1, which will imply that it
cannot have members of any order greater than n. But one may note that this
implication is in effect a quantification over items of all orders, and on Russell’s own
principles it is not clear that this should be permitted.9 In my own view, a better
suggestion is that the proposed definition of the Russell class is ruled out by the simple
theory of types, which distinguishes sets into different levels. As we shall soon see, this is
quite different from the distinction into orders which is imposed by the ramified theory.
But before we come to that, let us come back to the present question: why should
Russell think that the VCP has ‘a certain consonance with common sense’?
His own discussion is rather reticent upon this point, though it does contain some
significant clues. When he is trying to sum up his diagnoses of the paradoxes, he very
frequently says things like this
Something is said about all cases of some kind, and from what is said a new case seems to be
generated, which both is and is not of the same kind as the cases of which all were concerned in
what was said (p. 61).

Or again, on the Epimenides

Whatever we suppose to be the totality of all propositions, statements about this totality generate
new propositions which, on pain of contradiction, must lie outside the totality (p. 62).

And, generalizing,
All our contradictions have in common 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 (p. 63).

The point to notice here is that Russell speaks as though these supposedly illegitimate
quantifications would themselves generate new cases, cases which are apparently cov-
ered by the original quantification and yet are also brought into existence by it. This is
apparently the feature that common sense would reject. But how are we to understand

9
The case is relevantly similar to what was noted in Chapter 4 as a problem for the simple theory, namely
that to explain this theory in English one naturally quantifies over items of all different levels. A similar reply is
perhaps appropriate.
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 81

the idea that the existence of the quantification, if it were legitimate, would generate a
new case?
There is one very natural suggestion. Russell is apparently assuming that the defini-
tion or specification of a new case would bring that new case into existence, because he
is assuming quite generally that abstract objects exist because and only because they are
definable. The idea will be that a legitimate definition always does succeed in defining
an abstract object, and hence these particular definitions must be counted as illegiti-
mate, because there are independent reasons for saying that what they purport to define
cannot exist. And he accounts for the illegitimacy by claiming that the quantifications
involved introduce a kind of circularity. This is a kind of conceptualist position on
abstract objects. The traditional conceptualist holds that abstract objects do exist, but
only in the mind: they exist because we think of them, and if there were no thinkers
then there would be no abstract objects. Russell is not committed to this. He is
apparently equating the existence of abstract objects with their definability,10 but he
need not be supposing that definitions exist only in minds. However, what he does say
would be quite unacceptable to someone of a realist turn of mind. The realist will say
that the existence of concrete objects (say, horses) obviously does not depend upon
whether they have suitable definitions, and the same should apply to abstract objects
too. According to him there may well be abstract objects that are not definable, and
there may well be perfectly good definitions which no abstract object satisfies. (Here is
a familiar analogue. Let us define a supremely efficient village barber as a man in the
village who shaves all and only those men in the village who do not shave themselves.
It is easily seen that there is no such barber, but that is not because there is something
wrong with quantifying over all the men in the village. Similarly, there is no set which
satisfies the definition of the Russell set, but our realist will see nothing wrong with the
quantification over all sets that it contains.) Russell apparently is supposing that, with an
abstract object, a legitimate definition will ensure its existence—or, at any rate, he is
attributing this view to ‘common sense’—so he is supposing that common sense does
not take the realist attitude.
Goldfarb [1989], who wishes to reject this interpretation, describes it as the view that
Russell’s ramified theory can be justified only by
a philosophical cast . . . [which] is nonrealist and in a sense constructivist: these [abstract] entities
do not subsist independently of us, but are created or legitimised by our being able to specify
them (p. 25)

This view is, he admits, ‘by now enshrined as the common wisdom’ (ibid.), and in
support one may certainly cite Gödel [1944: 456–9]11 and Quine [1963: 243].

10
This agrees with Poincaré [1906], who claimed that, where there are supposed to be infinitely many
objects of a certain kind, ‘these objects cannot be conceived as existing prior to their definition’ (p. 194).
11
Note Goldfarb’s cautious phrase ‘nonrealist and in a sense constructivist’. Gödel is cautious in the same
way, as is particularly clear in the revised [1972] version of his paper. (See the new footnote on p. 192.) Just
what kind of ‘constructivism’ is involved is unexplained.
82 L O G I C A N D M AT H E M AT I C S

Nevertheless there are quite strong reasons for being hesitant over this point, for in
earlier days Russell certainly had taken a realistic attitude to such abstract objects as
propositions and universals, even if in [1906c] his attitude seems to be somewhat
ambivalent.12 I will take note of this in Chapters 12 and 13. Commentators have
therefore been reluctant to suppose that his adherence to the VCP does involve a
retreat from realism, and in defence they have invoked the rather different treatment
of the VCP that Russell gives us a little later, in chapter II of his introduction to
PM [1910d]. So let us now turn to this.

3 PM and presupposition
In the PM account Russell does open by repeating the claim that his theory ‘has a
certain consonance with common sense which makes it inherently credible’ (p. 37),
but there is no evident point in his discussion where such a ‘consonance’ is invoked. He
then gives an explanation of what the VCP is, which is much the same as the
explanation given previously in ML in terms of ‘illegitimate totalities’. It adds the
example that the supposed proposition ‘all propositions are either true or false’ cannot
be legitimate unless its phrase ‘all propositions’ is understood as referring to some
already definite collection of propositions, ‘which it cannot do if new propositions
are created by statements about “all propositions”’ (p. 37). This again imports the idea
that stating a proposition somehow creates it, but the thought is not elaborated, and
Russell merely draws the same moral that ‘all propositions’ is an illegitimate totality.
However this is now followed by the observation that in the logic to be presented the
relevant items which could give rise to (supposedly) vicious circles are all propositional
functions, and the next section then aims to elucidate the nature of propositional
functions and to explain how the VCP is justified in their case. No other cases are
considered further.
A propositional function is explained as ‘something which contains a variable x, and
expresses a proposition as soon as a value is assigned to x’ (p. 38).13 Russell thinks of the
function as ‘ambiguously denoting’ the various propositions that result by assigning a
value to the variable as argument to the function, and these propositions are the values
of that function for those arguments. In this explanation he cannot be using the letter
‘x’ specifically as a variable that ranges only over individuals, for he must mean to be
speaking of propositional functions at all levels, and ‘x’ is here signifying an argument
place for whatever arguments are appropriate to that level. The same applies to several
subsequent examples.

12
At [1906c: 149] Russell says that he himself takes a definition to discern a new entity, but not to create it.
On the other hand he apparently insists that an infinite set must always have a definition (p. 163).
13
We should add two further conditions, (i) that the value assigned to x should occur in the proposition
that is then expressed, and (ii) that when two different values are assigned to x the two resulting propositions
should each say the same thing of the values in question. We may take these points for granted in what follows.
(The first of them has been made earlier, at PoM, p. 510.)
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 83

The crucial claim on which all of the following discussion is based appears to
be this:
A function is not a well-defined function unless all its values are already well-defined (p. 39).14

The claim is elaborated in this way:

It follows from this that no function can have amongst its values anything which presupposes the
function, for if it had, we could not regard the objects ambiguously denoted by the function as
definite until the function was definite, while conversely, as we have just seen, the function
cannot be definite until its values are definite. This is a particular case, but perhaps the most
fundamental case, of the vicious circle principle. (p. 39)

Apparently, this is what Russell now thinks of as the justification for his (ramified)
theory of types. The theory is all based on the principle that a propositional function
presupposes all of its values, whereas they do not presuppose it.
One may reasonably suspect that this principle is what now remains of Russell’s
earlier ‘substitutional theory’. For in that theory propositional functions did not appear
as such, but were replaced by a whole proposition that contained the function applied
to a certain argument, together with the idea of substituting other arguments in place of
that one, and thereby obtaining a range of different propositions all containing the
same function. So one may say that in this substitutional approach the role of the
function was played by all the propositions that are its values for different arguments. As
we have seen (on pp. 53–4), Russell officially espoused this substitutional theory as late
as ML [1908], though he knew by then that in its original form it could not be
maintained, and he does not seem to have settled upon any revised form that satisfied
him. However in PM this old substitutional theory is never mentioned, unless it does
partially survive in the new principle that is here stated. For the idea seems to be that a
propositional function is, as it were, ‘defined’ by the propositions that are its values, and
that is why it is not ‘well defined’ unless they are all ‘well defined’.
If this principle is to be at all plausible, it cannot be taken as meaning that we do not
understand a function unless we already understand each and every one of its values. As
Russell acknowledges:
A function can be apprehended without its being necessary to apprehend its values severally and
individually. If this were not the case, no function could be apprehended at all, since the number
of values (true and false) of a function is necessarily infinite and there are necessarily possible
arguments with which we are unacquainted. What is necessary is not that the values should
be given individually and extensionally, but that the totality of the values should be given

14
This claim does not appear in ML. We find there only the weaker claim that ‘a proposition containing
an apparent [i.e. bound] variable presupposes others from which it can be obtained by generalisation’ (p. 75).
If we interpret ‘others’ to mean ‘some others’, we get a very plausible claim, which however comes nowhere
near to justifying the VCP. But if we take ‘others’ to mean ‘all others’ then we do come close to the principle
stated in PM.
84 L O G I C A N D M AT H E M AT I C S

intensionally, so that, concerning any assigned object, it is at least theoretically determinate

whether or not the said object is a value of the function. (pp. 39–40)

It really is not clear what Russell is thinking of when he says that ‘the totality of the
values should be given intensionally’, but perhaps we can think of it in this way. We
may begin with one proposition which we understand, and which is a value of the
function for some argument. We then describe the other values as the propositions
that may be obtained by substituting other arguments for that one, and the question
that must be ‘theoretically determinate’ will be the question whether a proposed
substitution is or is not a legitimate one, i.e. whether it does or does not yield a
genuine proposition. And what we have been told is that the proposition must not in
turn ‘presuppose’ the function that is being explained. We need to know, then, how a
proposition may ‘presuppose’ something, and the answer to this is not obvious, because
the idea of one thing presupposing another has not in fact been explained at all. We have
simply been told that the propositional function presupposes the propositions that are
its values, but we have not been told how it might happen that those propositions also
presuppose the function. We shall need to be able to add some assumptions about this if
we are to show how the restrictions of Russell’s ramified theory of types will follow
from the new principle.
I shall take the discussion in Hylton [1990] as a well-known and straightforward
attempt to spell out what seems to be Russell’s reasoning. Hylton proposes a justifica-
tion in two steps, the first aimed at the simple theory of types, and the second at its
ramified complication. He says:
The idea that a propositional function presupposes its values gives rise to distinctions . . . of order.
The basic principles are two. First, a propositional function is of higher order than any object
which it can take as argument; second, a propositional function is of higher order than any object
within the range of a quantifier contained in that propositional function. (Hylton, pp. 300–1)15

The first of these consequences he claims to follow ‘almost immediately’, because he

assumes that a proposition which is a value of a certain propositional function will
presuppose the object which is the argument of that function in that proposition. Thus
the function presupposes its argument and so is of higher order than the argument,
because it presupposes its value, and the value presupposes the argument.
One might very naturally think that just as the proposition ‘Socrates is human’
presupposes the object Socrates, which is one of its constituents, so it also presupposes
the function ‘x is human’, which is the other, for it presupposes both of its constituents.
If so, then the function ‘x is human’ will (inter alia) presuppose itself. But Russell
explicitly denies this. He says:

15
The discussion in Goldfarb [1989] contains only Hylton’s second point, for he is concerned only with
the ramifications. But I cannot see that he offers anything much by way of an argument for this point.
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 85

The values of a function are presupposed by the function, not vice versa. It is sufficiently obvious,
in any particular case, that a value of a function does not presuppose the function. Thus for
example the proposition ‘Socrates is human’ can be perfectly apprehended without regarding it
as a value of the function ‘x is human’. (p. 39)

I remark that, in the case of a complex propositional function, this point can perfectly
well be understood as a point about understanding. One can understand a proposition
that is a value of the function by understanding its simpler parts, and without ever
seeing it as containing the complex function in question. But in the simple case that is
instanced here the point appears to be highly dubious. How could one understand the
proposition ‘Socrates is human’ without understanding the function expressed by
‘ . . . is human’?16 Well, the answer may perhaps be that, despite what Russell himself
says in this passage, his notion of ‘presupposition’ is not really connected with under-
standing. He does not mean that you cannot understand one thing without also (or
‘already’?) understanding another, and the quotation that is given above from his pp.
39–40 confirms this. After all, he never does define the notion of presupposition that
he is using. Other explanations may be suggested.17 But in any case, for the sake of
argument, let us allow Russell his claim that the proposition ‘Socrates is human’
presupposes what is meant by ‘Socrates’ but not what is meant by ‘ . . . is human’.
Generalizing from this kind of example, Hylton infers that a propositional function will
always presuppose its arguments, but will not presuppose itself, which he takes to be
the central claim of the simple theory of types. Russell too accepts that this is a
consequence of his principle, and apparently identifies this consequence with the
principle itself when he later says:
A function, as we saw at an earlier stage, presupposes as part of its meaning the totality of its
values, or, what comes to the same thing, the totality of its possible arguments. (PM, p. 54)

However, there is a crucial objection. To suppose that this follows for functions of
all levels is to generalize rashly from the one kind of case considered so far, namely
propositional functions of first level. For consider again the simple proposition ‘So-
crates is human’. This is both the value of the first-level function ‘x is human’ for the
argument ‘Socrates’ and the value of the second-level function ‘F (Socrates)’ for the
argument ‘x is human’. It cannot be said that this proposition presupposes ‘Socrates’,
but not ‘x is human’, when it is viewed in the first way, and yet presupposes ‘x is human’,
but not ‘—Socrates—’, when it is viewed in the second way. It must either presuppose the
function ‘x is human’ or not presuppose it, and Russell cannot have it both ways. But,
whichever option we choose, we shall falsify the general claim that every function, of

16
Compare Dummett [1973: 27–30].
17
Perhaps Russell thinks at this date, i.e. before he wrote PLA, p. 205 (which is cited below on p. 124), that
what is required is an acquaintance with the universal humanity, and not any notion of a propositional
function. This elucidation is explicitly endorsed by Linsky [1999: 38]. (The relations between universals and
propositional functions will be discussed later, in Chapter 13.)
86 L O G I C A N D M AT H E M AT I C S

any level, presupposes its arguments and does not presuppose itself. The consequence
that Russell desires will not be forthcoming.
I add that, even if it is granted that every propositional function must be of higher
order than its arguments, still the restrictions of the simple theory of types are not
thereby established. For the hierarchy that is characteristic of the simple theory is a strict
hierarchy and not a cumulative hierarchy. When the simple theory is taken as forbidding
nominalizations, this strict hierarchy is just a consequence required by ordinary gram-
mar, as we saw in the previous chapter. But in the present context grammar is being
ignored in favour of the new and unexplained notion of presupposition, which cannot
justify the imposition of a strict rather than a cumulative hierarchy. The same applies,
I think, to any way of explaining the VCP: it will at best justify a cumulative hierarchy
of orders, which would indeed avoid the paradoxes (as in ZF), but which would lack
the purely grammatical justification that the simple theory has.
Anyway, let us now come to the more complex and ramified theory that Russell is
proposing in PM. Hylton proposes to deduce this as his second consequence of the
leading principle that a propositional function presupposes its values, i.e. the conse-
quence that ‘a propositional function is of higher order than any object within the
range of a quantifier contained in that propositional function’. His elucidation is this:
A propositional function which expresses generality (^ x has all the properties of a great general is
Russell’s example) contains another propositional function (f is a property of a great general). So the
first propositional function presupposes the second, which (by the first principle) presupposes the
entities which are its arguments. (p. 301)18

From this he infers that the first-level function ‘x has all the properties of a great
general’ must be of a higher order than the functions which fall within the range of ‘all
the properties of a great general’, i.e. such first-level functions as ‘x is resourceful’, ‘x is
far-sighted’, and so on. The reasoning here may be spelt out in this way.
Our initial function may be written as
8F(8y(y is a great general ! Fy) ! Fx)
By Russell’s first principle, this presupposes the propositions which are its values, i.e.
propositions such as
8F(8y(y is a great general ! Fy) ! F(Napoleon))
By a second principle, this proposition presupposes the second-level function that is the
argument to its initial quantifier ‘8F’, namely the function
(8y(y is a great general ! Fy) ! F(Napoleon))

18
It is not clear why Hylton puts a cap on the variable ‘x’ but not on the variable ‘f’, but that is a detail of
no importance. (Russell gives the example on p. 56 of PM.)
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 87

Then, by the desired thesis that every function (of every level and order) presupposes its
possible arguments, this second-level function presupposes its first-level arguments, i.e.
such functions as
x is resourceful
x is far-sighted
x is extremely rash
etc.
I note that Russell himself comes near to endorsing the second principle that is here
assumed, at least insofar as he says that a quantified proposition ‘involves’ the function
that is the argument to its quantifier. He could surely have said ‘presupposes’ in place of
‘involves’.19 So this looks like a very nice deduction of a conclusion that Russell
certainly wants.
However, there is once more a crucial objection: by invoking the same principles
one can also deduce results that Russell does not want. For example, suppose that we
start with the first-level function
9y9z(x gave y to z)
As before, this presupposes the propositions that are its values, for example

9y9z(Plato gave y to z)
This proposition in turn presupposes the first-level function that is the argument to
its initial quantifier, namely the function

9z(Plato gave y to z)
So our initial function presupposes a proposition which presupposes our second
function. By the same reasoning, then, the initial function should be of a higher
order than the second. Clearly we can repeat the manoeuvre. The second function
presupposes the propositions that are its values, for example

9z(Plato gave this book to z)

This proposition in turn presupposes the first-level function that is the argument to its
initial quantifier, i.e. a third first-level function

Plato gave this book to z

So we have deduced that the first function is of higher order than the second, and the
second is of higher order than the third. But this is a result that Russell rejects. In his

19
He says: ‘We will denote by the symbol ‘(x).çx’ the proposition ‘çx always’, i.e. the proposition which
asserts all the values for ç^x This proposition involves the function ç^
x , not merely an ambiguous value of the
function’ (p. 41). When recalling this claim somewhat later, he does say that ‘(x).çx presupposes the function
ç^x ’ (my emphasis) (p. 63).
88 L O G I C A N D M AT H E M AT I C S

view all three functions should be of the same order, for they are all first-order
functions.
As a matter of fact it would not be too awkward for him to allow that the first two
are of higher order than the third, for the third quantifies over no totality, whereas the
first two both do contain a quantification. But they quantify in each case over just the
same totality, namely the totality of all individuals. And Russell certainly does not think
that this totality is somehow ‘illegitimate’, and needs to be split into a number of
subtotalities. It is only quantification over functions that he thinks might introduce an
illegitimacy, and so needs to be regulated by distinguishing functions into a variety of
different orders. Moreover, he surely thinks that there may be two quantifiers which
both quantify over functions of the same order, though one is inside the scope of the
other, as in ‘8F1 9G1 (—)’. But the principle that we have just been exploring will not
allow this; it will insist that the outer one must be of a higher order than the inner.
It has to be said here that in the Introduction to PM, in section 3 of chapter II,
Russell does claim that different kinds of truth are applicable to propositions of a more
complex or less complex structure, and according to this claim the sheer number of
quantifiers within a proposition make a difference to its ‘kind of truth’. Hence ‘9y9z
(Plato gave y to z)’ has a ‘higher kind of truth’ than does ‘9z(Plato gave this book to z)’,
which in turn has a ‘higher kind’ than ‘Plato gave this book to Aristotle’. It has been
held by some that Russell hoped to make this distinction of different kinds of truth into
a suitable basis for his ramified theory of types.20 But certainly it is not the way that the
different orders of functions are actually distinguished in PM as we have it. For there
the order of a function is a matter of what is quantified over in its definition, and is not
affected by the mere number of the quantifiers involved. For example, first-order
functions are quite clearly defined as functions which take only individuals as argu-
ments, and which contain only quantifiers that range over individuals. They may
contain any number of such quantifiers, for however many they are it will still be the
case that they ‘presuppose no totality except that of individuals’. These functions are
the first-order ones (p. 51).21
We must therefore set aside the supposed justification of the ramified theory of
types in PM, with the comment that it does not seem to be well thought out. Russell
emphasizes just one principle of presupposition, namely that a propositional function
presupposes all the propositions that are its values. The principle is quite obscure,
since the kind of ‘presupposition’ that is involved is left without any worthwhile

20
I shall discuss this view in section 4 of Chapter 12.
21
The definition of ‘first order’ in PM*12 is the same (p. 162), though it is there noted that a different
number of quantified variables will mean that the different propositions containing them are not all ‘of the
same type’, according to the criterion of sameness of type that is given in *9.131. But this criterion is only
partly in harmony with the different ‘kinds of truth’ distinguished in section 3 of chapter II of the
Introduction, and anyway it seems to be ignored in all the rest of PM. I would say that we do best to ignore
this supposed criterion of sameness of type. At any rate, it is not used in what is basic to PM, namely the
definition of the order of a propositional function.
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 89

clarification. (All that is said is the negative point that it is not intended to mean that
one cannot understand the function without understanding all its values.) Besides, this
principle by itself will not yield any of the consequences that Russell wants unless we
add to it some further principles about what a proposition presupposes. Russell does
state or imply two such principles, (i) that a proposition such as ‘Socrates is human’
does presuppose what is meant by ‘Socrates’ but not what is meant by ‘ . . . is human’,
and (ii) that a proposition of the form ‘8x(fx)’ presupposes the embedded function
‘fx’. He appears to need these supplementary principles in order to deduce some
results that he does want, but he must also reject them in order to avoid some results
that he does not want. Perhaps these principles (or others) could be so revised or
qualified that they did yield all and only the results that he desires, but this is at best
doubtful. In terms of presupposition, the relevant principles required by the VCP
concern functions presupposing totalities, and in PM these will be totalities of func-
tions or of individuals, which are not to be identified with any totalities of proposi-
tions. The principles are (i) that any function presupposes all of its arguments, and (ii)
that a function which is specified by using a quantified variable presupposes all the
values of that variable. But we are still left without guidance on what is here meant by
‘presupposition’, or on why these claims should be accepted. I shall therefore return to
the original exposition of the theory in ML, which does not try to build any results on
the supposed principle that a propositional function ‘presupposes’ all the propositions
that are its values. But it does contain some indication of the kind of ‘presupposition’
that Russell has in mind.

4 Return to ML
In PM it is not clear whether Russell should be regarded as rejecting a realist approach
to propositional functions, but in ML he surely does reject it, insofar as he is thinking of
all abstract objects (including propositional functions) as being created by their canoni-
cal specifications. Nor is it particularly surprising that in 1908 he should move towards
such a non-realist position, for as we shall see in Chapter 12 he is at the same time
moving towards the non-realist position on propositions that they exist only as the
sentences (or other combinations of symbols) that express our beliefs, and that they
have no other more ‘abstract’ kind of existence. As we shall see further in Chapter 13,
he would appear to have taken the same view about propositional functions, i.e. that
they too are nothing more than certain combinations of symbols. Admittedly, such a
view leads very quickly to what is certainly an awkward problem for him, as the next
chapter will emphasize. But let us now revert to the non-realist stance of ML: abstract
objects—and in particular, propositional functions—exist only because they are defin-
able; it is the definition that would (if permitted) somehow ensure their existence.
However, our explanation of the VCP cannot stop here. For even if it is our
definitions that bring such objects into existence, we may still ask why an impredicative
definition cannot do this just as well as one that is predicative. To explain this it looks as
90 L O G I C A N D M AT H E M AT I C S

if we must suppose that Russell is also subscribing to the kind of ‘construction’

metaphor that lies behind the standard interpretation of ZF set theory. It is significant
that in this discussion of ML (i.e. pp. 59–64) he constantly speaks in terms of ‘totalities’,
and it is quite easy to think of a totality (i.e. a set?) as ‘built out of ’ its members. So, if
we may rely on the usual temporal metaphor we may say that, just as a house cannot
exist until after its bricks do, similarly a totality cannot exist until after its members do.
We now add that a totality cannot be referred to until it exists, and that an impredi-
cative definition is one that professes to define a member of a totality by reference
to that totality. This does indeed introduce what looks like a genuinely vicious circle.
By the construction principle a totality cannot exist until all its members do; by the
conceptualist principle a member cannot exist until it is defined; so if all its definitions
are impredicative the member cannot exist until after it exists. This clearly is an
impossible situation.
Even so, the argument is not really very convincing. As stated, it includes the claim
that you cannot refer to a totality until it exists, which in turn cannot happen until all of
its members exist. That may sound plausible at a first hearing. But now let us recall that
this ‘reference’ to a totality is, in context, just a quantification over all members of
that totality, i.e. a use of an expression of the form ‘all so-and-so’s’. One may note (a)
that such a use would not normally be thought to presuppose the existence of a singular
item called ‘the totality’ (e.g. set?) of all so-and-so’s. (For example, in modern set
theory one quantifies over all sets without supposing that there is a set of all sets.) Also
(b) in ordinary cases one may certainly quantify over all so-and-so’s without supposing
that they all exist already. (For example, if I say ‘All men, past present and future, are
mortal’, I am surely not supposing that future men somehow exist now.) Why would
not this same point apply when the so-and-so’s are abstract objects? More generally (c)
it may reasonably be said that if I am meaningfully to quantify over all so-and-so’s then
I must understand what it is to be a so-and-so, i.e. (in Frege’s language) I must know
the sense of this expression. But I do not need to know all the so-and-so’s, or even any
so-and-so, or even whether there are any so-and-so’s. So in what sense is it reasonable
to claim that a quantification over all so-and-so’s cannot ‘exist’ until ‘after’ each of the
so-and-so’s exists? This is quite unclear. So it is also left unclear why a propositional
function which is defined by means of such a quantification cannot exist until ‘after’ all
the entities quantified over exist.
The situation is the same even in what seems to be the simplest and best case for
Russell’s VCP, namely the case of a proposition which contains a quantification over
all propositions of a certain kind. It is not at all clear why we should accept the idea that
that proposition is somehow ‘built from’ each of the propositions that are thus
quantified over. And this point apparently still holds even if we add the non-realist
principle that a proposition does not exist until ‘after’ the means of expressing it exist.
For one may perfectly well be able to express the quantification before one has the
means of expressing each of the items quantified over. We may conclude that Russell
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 91

has given no good ground, either for ‘common sense’ or for any other overall position,
to accept the VCP, save that it does promise a way of avoiding certain paradoxes.
Let us now move on to a different topic: what is the effect of trying to conform to
the VCP?

5 The need for axioms of reducibility

The simple theory of types that we considered earlier stratifies propositional functions
into levels. In the ramified theory that concerns us now, propositional functions are
distinguished not only into levels but also into orders, in obedience to the VCP. Russell
explains that a propositional function is of order 1 if its fundamental specification
quantifies (if at all) only over individuals, of order 2 if the specification quantifies over
functions of order 1, and so on. It follows that some functions of order 2 are of level 1,
and some are of level 2. For example, if the variable G1 ranges over first-level functions
of order 1, then there will be a first-level function F2 of order 2 such that

8x(F2(x) $ 8G1(G1(x) $ G1(a)))

(The function F2 is true of the individual a, and—presumably?22—of nothing else.) But
there will also be a second-level function M 2 which is again of order 2, such that23

8G1(M 2(G1) $ 9x(G1(x)))

(The function M 2 is true of all non-empty first-level functions of order 1.) Thus the
classification of propositional functions into levels, and into orders, is simply a cross-
classification, and neither implies the other.
Russell appears to claim that the classification into levels can itself be derived from
the VCP, but he gives no such derivation, and we have observed that there seems to be
no prospect of constructing one. But he is justified in assuming that there cannot be a
function of level n that is of order less than n, and he pays special attention to the
functions of level n that are also of order n, which he calls ‘predicative’ functions. They
play a special role in his theory, as we shall see.24 The general requirement, then, is that
quantification is permitted only where the bound variables are restricted both to some

22
The ‘presumption’ is satisfied if the axiom of reducibility is true. (This axiom will be introduced
shortly.)
23
For clarity I here abbreviate the full version M xGx simply to M (G).
24
At first sight it is not easy to see why he chooses the word ‘predicative’ to pick out these functions, for
there is no proper contrast with his use of the word ‘impredicative’. But there is a historical explanation,
which is given in Quine [1963: 249, n. 2]. My text gives Russell’s first definition of ‘predicative’, as in ML and
PM Introduction (chapter II, section 5). There is a different and more restrictive definition in PM *12 and in the
second edition [1925: xxviii ff.], but this difference leaves the overall role of predicative functions unaltered.
(On the first account, all first-order functions are predicative; on the second account only those that are
quantifier-free are predicative.)
92 L O G I C A N D M AT H E M AT I C S

definite level n and to some definite order m, where mn. Other quantifications are
deemed to be ‘illegitimate’.
That is the basic outline of Russell’s ramified theory of types. Of course this
exposition has been simplified by considering only monadic propositional functions,
and the situation becomes very much more complicated when we also include
functions that are dyadic, triadic, and so on. They now have to be distinguished not
only by the different levels of their various arguments but also by the permitted orders
of these arguments within the different levels. In effect, the type indices that were
needed for the simple theory must now be doubled in complexity in order to
distinguish all possible ramifications. I do not attempt to give a proper specification
of such indices, or to use them in giving a formal presentation of the theory. For in fact
there are matters of detail which are open to dispute, since Russell’s own exposition of
his theory is often unclear, but it is not worth fussing over such details.25 One popular
way of thinking of Russell’s theory is to do what Russell himself never explicitly did,
i.e. to suppose that we begin with the simple theory and then add to it the ‘ramifications’
which the VCP demands. In that case one quite naturally thinks of the orders as always
beginning in each level with order 1. This is a convenient way of thinking, and one
that makes no important difference to the overall results of this theory.26 So let us now
turn to these results.
The main consequence is distinctly awkward: the theory bans impredicative defini-
tions, but in fact impredicative definitions are common, both in ordinary logic and in
ordinary mathematics. A simple example is the notion of identity. In a second-level
logic this is standardly defined so that

a ¼ b $ 8F(Fa $ Fb)
Here one understands the quantifier 8F as ranging over all first-level predicates, so that
the definition ensures that if a ¼ b then either may be substituted for the other in any
context whatever. But in the ramified theory we are not allowed to do this, since the
variable F must now be confined to first-level predicates of some definite order. And
whatever order is chosen, the definition cannot rule out the possibility that while a and
b share all predicates of that order still they differ on predicates of a higher order. But
this must be ruled out if it is genuinely to be identity that is defined.
In this case there is a straightforward remedy, which is to take identity as a primitive
and undefined notion, governed by the axioms that are usual in a first-level logic,
namely

25
There are expositions of the theory which differ from one another in detail in Hatcher [1968: ch. 4],
Copi [1971: ch. 3], Chihara [1973: ch. 1], Church [1976], and more recently Urquhart [2003]. Linsky [1999:
ch. 4] summarizes all of these (except for the more recent Urquhart [2003]), and fairly concludes that the
variations are not at all important from the philosophical point of view.
26
This convenient method is followed in Copi [1971], but one needs to recall my point at the start of this
section.
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 93

a¼a
a ¼ b ! (Fa $ Fb)
Here the letter ‘F ’ is not a bound variable, and we can therefore interpret the second
axiom as an axiom schema, generating infinitely many axioms, one for each first-level
predicate in place of ‘F’. This covers in one blow all first-level predicates of all orders,
and so avoids the problem.
But an essentially similar difficulty keeps cropping up, in one definition after
another, and we cannot meet them all in the same way without introducing a plethora
of new axioms. For example, the usual definition of ‘natural number’ must now be so
modified that it no longer yields the desired result. Suppose that ‘0’ is already defined as
an item of some definite level and order, and suppose that ‘successor’ is also defined as a
function on such items. Let us use ‘x’, ‘y’, . . . as variables for those items (whatever is
their level and order), and ‘y0 ’ for ‘the successor of y’. Then one standardly defines ‘ℕ’
for ‘natural number’ by27

ℕx $ 8F(F0 ^ 8y(Fy ! Fy 0 ) ! Fx)

But once more the ramified theory requires the bound variable F to be confined to
functions of some definite order, and whatever order that is it will then follow from the
definition that mathematical induction holds for the natural numbers for functions of
that order, but it will not follow that it also holds for functions of higher order. That is
evidently not what is needed. Once again, we can remedy this situation by introducing
the needed principle as a new axiom, and abandoning any attempt to define ‘natural
number’. But the same problem keeps cropping up again and again. For instance, one
standardly defines ‘the Fs are as many as the Gs’ as ‘there is a relation which correlates
the Fs and the Gs one-to-one’, and one understands this as quantifying over all
relations of all orders. But according to the ramified theory some definite order must
be specified, and when that is done the consequence is that there may be a correlating
relation of a higher order without it following that the Fs are as many as the Gs. This
again means that the definition does not adequately define the notion intended. Many
further examples could be cited, but perhaps the best known is the problem over the
real numbers. On Russell’s approach, real numbers are to be understood as (resulting
from) Dedekind cuts in the rationals, but the predicates which define such cuts will be
of different orders. It apparently follows that the real numbers which result will
themselves be of different orders, and consequently no variable will have all the real
numbers in its range. This must prevent us from developing the usual theory of the real
numbers.

27
In terms of classes, this defines the class of natural numbers as ‘the least’ class that contains 0 and contains
the successor of anything that it contains. Is this illegitimate? Compare Ramsey’s remark on ‘the tallest man in
the room’ (p. 79).
94 L O G I C A N D M AT H E M AT I C S

Russell’s reaction to these difficulties is to introduce an Axiom of Reducibility. This

states that to any propositional function of any order, within a given level, there
corresponds another which is of the lowest possible order in that level.28 That is,
there corresponds what he calls a ‘predicative’ function of that level.29 The two
‘correspond’ in the sense that they are equivalent, i.e. true and false of exactly the
same things. Hence if we define identity for individuals by quantifying over all
predicative functions of individuals, we shall be able to deduce that identical individuals
may always be substituted for one another in any context that our logic considers, since
the relevant propositional function will always be equivalent to some predicative
function. Similarly we may define ‘natural number’ by quantifying over all predicative
functions of natural numbers, and it will follow that induction holds for all functions of
natural numbers, whether predicative or not. The same applies to all other cases where
one naturally uses a definition that is impredicative.30
In particular, it applies to Russell’s way of introducing the usual notation for classes
while retaining his ‘no-classes’ theory. In the context of the simple theory of types, the
definition for classes of first level was given earlier (p. 59) in this way

(—{x:Fx}—) for 9G(8x(Fx $ Gx) & (—G—))

Stated in this way the definition has now become illegitimate, and the bound variable
G needs to be confined to some definite order. But if we confine it to functions of
order 1, i.e. to the predicative functions of level 1, then the axiom of reducibility
assures us that nothing has been left out. For every function of level 1, of any order
whatever, will be equivalent to some predicative function of level 1, and so will
determine some class in accordance with this definition. That is why Russell so closely
associates the axiom of reducibility with the ordinary notion of a class, for he thinks that
the ordinary notion both supposes that there always is a class for any propositional
function, and that classes are not distinguished from one another by being of different
orders (e.g. [1908: 81–2], [1919a: 191]).31 More generally, the effect of the axiom of
reducibility is that in the actual deductions of PM it is only the simple theory of types

28
This is strictly an axiom schema for each level, e.g. for the first level
9F18x(F1(x) $ (—x—))
for any well-formed formula containing ‘x’ free, but not ‘F1’ free, in place of ‘(—x—)’. (Compare the axioms
of comprehension mentioned earlier on p. 66) In modern terms, Russell’s distinction between ‘all’ and ‘any’,
which he elaborates in ML section II, is best seen as a distinction between a schema for propositions (or axioms)
and an explicit quantification.
29
This assumes the definition of ‘predicative’ that is given in the Introduction to PM. In *12 we find a more
restrictive definition. Cf. note 24.
30
Russell assumes axioms of reducibility for monadic propositional functions of any level, and for dyadic
functions of any level. He does not introduce such axioms for polyadic functions of any higher degree, but
this is only because they are not needed for the purposes of PM.
31
The ordinary notion also supposes that classes are not distinguished from one another by being of
different levels. That is a point which Russell here ignores.
 THE CONTRADICTION ( I I I ): A RAMIFIED SOLUTION 95

that is used. For almost all of the deductions use the notation of classes, and classes are
defined by quantifying only over predicative functions, and if we confine attention to
these functions then we are in effect working within the simple theory. In fact PM does
not use quantifiers over any other functions than the predicative functions.32 So one
naturally asks: what was the point of first introducing the complications of the ramified
theory, and then adopting an axiom of reducibility which allows us simply to ignore
them?33
In Russell’s view the point is this. The axiom allows us to ignore the distinction into
orders when we are doing logic or mathematics, for the propositional functions that
occur in those subjects are all extensional functions, i.e. they give equivalent proposi-
tions as values when supplied with extensionally equivalent arguments. But when we
come to consider the semantic paradoxes the propositional functions involved are not
usually extensional, and so the axiom of reducibility cannot reintroduce the paradox.34
Here is a simple illustration.
Suppose that I take some object a (say, a pebble on the beach at Brighton), and
assert of it
Whatever is at any time asserted of a is false of a.
Suppose that nothing else ever is asserted of a. Then on the face of it we have a
contradiction of the usual type, for what I assert of a is true of a if and only if it is false of
a. But it is generated by the use of a quantification (‘whatever is asserted’) which
Russell will regard as illegitimate, so his solution is that we must specify a definite order
for the propositional functions here generalized over. Let us take them to be first-order
functions. Then what I assert of a must be more narrowly expressed as

Whatever first-order propositional function is at any time

asserted of a is false of a.
And to assert this of a is to assert of a not a first-order but a second-order propositional
function. Hence what I assert of a is vacuously true of a, for by hypothesis no first-order
propositional function ever is asserted of a; and it is not also false of a because it is not
itself a first-order propositional function. Now by the axiom of reducibility this
second-order function is equivalent to some first-order function, so there is a first-
order function which is true of a. But no contradiction follows from this, for we have

32
It is clearly explained in Landini [1996a] that in effect the logic of Principia Mathematica is that of the
simple theory of types. (It is stated on p. 54 and p. 165 of PM that there is no need to quantify over any
functions other than predicative functions. But, in context, this appears to be just a mistake. A counter-
example is given in Hylton [1990: 308–9n]. The need is avoided only by introducing the axioms of
reducibility (which has not yet been done on pp. 54 and 165), and by confining attention to contexts
which are extensional, so that their ‘reductions’ do apply.)
33
The question is clearly raised in Quine [1936]. (He argues that there is no point, but his argument
assumes axioms of extensionality, and Russell does not make this assumption.)
34
But there is a problem over the paradox of heterologicality. See e.g. Copi [1950], Myhill [1979], Potter
[2000: 154–7], and my [2011].
96 L O G I C A N D M AT H E M AT I C S

no reason to suppose that that first-order function which is true of a is also asserted of a.
The two functions are extensionally equivalent, but it does not follow that the
intensional function ‘ . . . is asserted of a’ must be true of the one if it is true of
the other. So in a context such as this, where non-extensional functions are involved,
the VCP and its associated distinction into orders still has an effect which is not
destroyed by the axiom of reducibility.
That is why Russell thinks that the ramified theory of types and the axiom of
reducibility work nicely together to give us all the desired results. The distinction into
orders is needed to solve the semantic paradoxes, and that solution is not affected by the
axiom of reducibility. But the distinction into orders is not needed in ordinary logic
or mathematics—indeed, it upsets much of ordinary mathematics—and in that context
the axiom of reducibility is important, for it allows us to ignore those distinctions
altogether. Russell does not claim that the axiom of reducibility is a truth of logic,
or that it has any of the ‘self-evidence’ that one might expect in a basic axiom, or even
that it has any ‘consonance with common sense’. Quite explicitly, in his [1907] he
argues that axioms should be judged not by their self-evidence but by whether they
yield all the consequences that are desired and none that are not desired, and he
evidently thinks that this condition is satisfied in the present case.35 We now see
why he should think this. But in the next chapter we shall see why he should not
have done.

35
The point is anticipated in his [1906c: 194–5], and is briefly repeated at PM, pp. 59-60. It is elaborated
once more in his [1924: 325–6], and there is a nice discussion of his overall position in Irvine [1989].
 6
Problems

There are problems both with the simple theory of types and with the ramified theory.
The problems are worse in the latter case, so I will begin with this, and then turn later
to the simple theory.

1 The ramified theory of types

As we have seen, in PM Russell apparently wishes to base his ramified theory on the
principle that a propositional function ‘presupposes’ the propositions that are its values.
But this principle is itself obscure, since it is not at all clear how its notion of
presupposition should be understood, and anyway if it is to yield the results that Russell
desires it must be supplemented, but no suitable supplements suggest themselves.
So we may leave this line of argument aside, and return to the way that Russell had
earlier stated his VCP, especially in ML [1908].
This apparently returns to the general idea behind his earlier ‘zigzag’ theory, that
what matters is how our abstract objects are specified. Certain kinds of specification are
ruled out as ‘illegitimate’, originally those that were ‘zigzaggy’, but now those that
involve what are taken to be ‘vicious circles’. In fact these two criteria are more similar
than may at first appear, for when describing the zigzag approach Russell had offered a
gloss which strongly resembles his later gloss on the VCP, namely this:
There are some properties such that, given any class of terms all having such a property, we can
always define a new term also having the property in question. Hence we can never collect all the
terms having the said property into a whole; because, whenever we hope we have them all, the
collection which we have immediately proceeds to generate a new term also having the said
property. [1906a: 144].1

In each case, the idea is that certain collections would, if permitted, ‘generate new
terms’, and these terms would cause a difficulty because they would then have to
belong to the collection which gave them being. There are of course some differ-
ences between the two theories. The zigzag theory of [1906a] claimed that some

1
There is a closer parallel in a working paper of 1904 which introduces what became the ‘zigzag’ theory,
for we there find the claim that if a generalization over all statements about x is itself a statement about x then
‘this leads to a vicious circle in definition’. Russell [1994: 88].
98 L O G I C A N D M AT H E M AT I C S

apparently meaningful predicates failed to determine classes, whereas the VCP of

[1908] prohibits certain predicates as themselves meaningless, on the ground that
they contain illegitimate quantifications. But the predicates in question seem very
much the same in each case, though the [1908] criterion is certainly easier to state
and easier to understand.
There are two ways in which the VCP apparently commits Russell to the view that
there are at most a countable infinity of propositional functions. First, what he says in
ML, if taken at all literally, must imply that abstract objects of all kinds owe their
existence to the fact that they have appropriate definitions. That is why certain
apparent definitions must be ruled out as illegitimate, i.e. because if they were
permitted then the objects that they purport to define would have to exist, and this
would lead to contradictions. But a definition must be formulated in a language, and
(as we have noted before, pp. 24–5) a language which is comprehensible to human
beings can contain only countably many expressions of any kind, whether definitions
or not. Apparently it must follow that there can be at most countably many proposi-
tional functions. One might seek to avoid this argument in a somewhat unrealistic way,
by saying that we need not limit our consideration either to actual languages or even to
languages that human beings could learn. Perhaps there are uncountably many possible
languages, or perhaps there is a possible language with an uncountable basic vocabulary.
In either case one might claim the potential existence of uncountably many definitions
(all of them ‘legitimate’ definitions), and hence the actual existence of uncountably
many propositional functions. But it might seem more reasonable just to say that
there may also be propositional functions which cannot be expressed in any language.
Why not?
This brings us to the second consideration. The way that Russell proposes to meet
the requirements of his VCP is by insisting that propositional functions be distinguished
into different orders, where the order of a function is determined by the quantifiers used
in its expression. (This is explicit in ML, p. 75.) It apparently follows that there must be a
way of expressing the function, for otherwise there would be nothing to determine
what order it belongs to. So we reach the same conclusion again, but by a different
route.
But Russell cannot accept this conclusion. For he aims to deduce classical mathe-
matics, including the classical theory of the real numbers, and Cantor showed that this
theory requires there to be uncountably many real numbers. To reproduce this theory,
then, Russell must assume the existence of uncountably many propositional functions.
He does this by adding to his ramified theory the axioms of reducibility. We have seen
that without such axioms his deduction cannot succeed, and one important reason is
that these axioms are needed to ensure that the propositional functions are uncount-
able. Here is a direct proof of that implication. (It is a very simple adaptation of Cantor’s
argument about the real numbers, as given earlier on pp. 20–1. But it may be helpful to
see that technique in action once more.)
 P RO B L E M S 99

For simplicity, consider just those propositional functions which take natural num-
bers as arguments and yield truthvalued propositions as values, and assume (for reductio
ad absurdum) that they are denumerable. Begin by confining attention to those of them
that are predicative (or first-order) functions of whatever level is needed to take
a natural number as argument.2 We can think of each such function as generating
an infinite string of truthvalues, i.e. a string of T’s and F’s, with the first in the string
giving the truthvalue of the function for the number 1 as argument, and in general
the nth member in each string giving the truthvalue for each function of its value
for the number n as argument. Since the functions are denumerable, they can be
ordered in a denumerable list, and their strings of values set out in a two-dimensional
array thus:
1st function: T, T, T, F, . . . . .
2nd function: T, T, F, T, . . . . .
3rd function: T, F, T, T, . . . . .
4th function: F, T, T, T, . . . . .
⋮
Now, go down the diagonal of this array of T’s and F’s to form a string which is
different from each one in the array. That is: if in the original list the nth place of the
nth string had a T, then put an F in the nth place of the new string, and vice versa.
(So, in this example, the new string begins ‘F, F, F, F, . . . . .’.) This new string differs
from every string in the original array, since it differs from the nth string in its nth place.
Moreover, we can quite easily specify a propositional function which takes numbers
as arguments and yields this new string of truthvalues as its values. For brevity, let
us just write F1n for the nth member of this list of all F1 propositional functions
taking numbers as arguments. Then the function that we need is the propositional
function F2 such that
8n(F2(n) $ 8F1(:F1n(n)))
Note that, as we have explained it, the new function with the new string of values
has to be classified as a function of second order, since its definition quantifies over all
first-order functions. But the axiom of reducibility states that this second-order
function is equivalent to some first-order function, and we have defined it so that it
cannot be. Either, then, the axiom of reducibility is false or the predicative (i.e. first-
order) functions of some suitable level cannot be enumerated. As a matter of fact we
can be more specific. As we shall see in the next section, Russell has to introduce an
axiom stating that there are infinitely many individuals, and therefore there is a
denumerable collection of individuals which can play the role for which we just
now invoked the natural numbers. The argument then shows that it is the first-order

2
As I remark later, they can actually be taken to be first-order functions of level 1, taking individuals as
arguments. For individuals can play the role here assigned to natural numbers.
100 L O G I C A N D M AT H E M AT I C S

functions of first level that must be non-denumerable if the axiom of reducibility is

accepted.3
We may conclude that Russell’s ramified theory of types, as he put it forward, is a
theory which falls between two stools. Its motivation strongly suggests a conceptu-
alist attitude towards abstract objects in general, or anyway towards propositional
functions in particular. Each such function can be defined, for that is what settles its
order, and apparently it owes its existence to the fact that it can be (legitimately)
defined, while presumably it is our mental activities that produce the languages
which make these definitions available. But a consequence of this approach is that
there cannot be more than countably many propositional functions, and therefore
the theory of propositional functions cannot reconstruct classical mathematics. Rus-
sell is not prepared to accept this consequence, and so introduces the axioms of
reducibility in order to avoid it. If his logic in PM had also included suitable axioms
of extensionality, then reducibility and extensionality together would simply have
wiped out the ramifications that were introduced in order to satisfy the VCP, and we
should be back in all respects just to the simple theory of types.4 As it is, Russell in
PM tries to get the best of both worlds. He does not adopt axioms of extensionality,
which (he hopes) will allow him to retain the VCP solution to the semantic
paradoxes. But he does adopt axioms of reducibility, so that in extensional contexts
the VCP has no effect. Since the deduction of mathematics never involves a non-
extensional context, it is therefore undisturbed. All the same, the theory is an
unstable amalgam. As a matter of principle it is difficult to see how it can
be reasonable both to say that what counts as an acceptable quantification must be
restricted by the VCP, and to adopt axioms of reducibility which posit the existence
of functions which allow us to escape the VCP in all mathematical contexts.
Moreover, these functions are posited to be first-order functions, though the order
of a function is supposed to be given by its definition, and we cannot produce any
first-order definitions for them. Are we supposed to believe that such definitions do
nevertheless exist? It is not surprising that Russell himself later came to believe that
the axiom of reducibility was unsatisfactory. He says in his Introduction to the
Second Edition of PM [1925] that it is ‘not the sort of axiom with which we can
rest content’, and he there proposes to drop it altogether (p. xiv). But he admits that
in that case the classical theory of real numbers cannot be obtained, while he still
insists that this classical theory is ‘an integral part of ordinary mathematics, and can
hardly be the object of reasonable doubt’ (ibid.). So apparently his position in [1925]

3
A more general point is indicated here. According to Russell’s way of eliminating classes, every class is
defined by some predicative function, so there must be at least as many such functions as there are classes.
A realist will certainly expect there to be infinitely many classes for which we have no defining function at all,
whether predicative or not (cf. pp. 24–5). (This point is made by Hazen [1983: 367–8].)
4
The point is demonstrated in Quine [1936].
 P RO B L E M S 101

is that he no longer thinks that the original theory of [1910] is satisfactory, and yet he
equally cannot accept his own proposed revision.5
Something that Russell does not consider in [1925] is the possibility of just dropping
the VCP altogether, and therefore the ramifications which it leads to. This, of course,
would obviate any need for axioms of reducibility. It is somewhat surprising that he
does not even consider this, for it is exactly what Ramsey had recommended in his
[1925a], and Russell surely knew of this.6 Moreover, he had always been somewhat
tentative about just what would be the best form for a theory of types. In the Preface to
PM [1910d] we find him saying
It gradually became evident to us that some form of the doctrine of types must be adopted if the
contradictions were to be avoided. The particular form of the doctrine of types advocated in
the present work is not logically indispensable, and there are various other forms equally
compatible with the truth of our deductions . . . Hardly anything in our book would be changed
by the adoption of a different form of the doctrine of types. (p. vii)

Similarly in IMP [1919a] he gives only a sketch of the theory, saying

It is the purpose of this book to avoid those parts of the subject which are still obscure and
controversial . . . [And] the theory of types emphatically does not belong to the finished and
certain part of our subject: much of this theory is still inchoate, confused, and obscure. (p. 135)

We find the same again in ‘Logical Atomism’ [1924]

Principia Mathematica . . . however, still falls short of finality in some fundamental points (notably
the axiom of reducibility). (p. 325)

Since in practice the deductions of PM just use the simple theory of types, and since it is
the ramifications due to the VCP which need to be straightened out by the doubtful
axiom of reducibility, why not consider just dropping the VCP, as Ramsey had
proposed?
I think it very likely that in [1925] Russell was not prepared to consider this because
he still felt it important to deal with the paradox of the liar, and he then saw no

5
The [1925] theory is not well thought out. It proposes dropping the axioms of reducibility, but
introducing some axioms of extensionality, which however are not clearly specified. The two changes do
not by any means cancel one another out. Russell claims that, even without reducibility, one can still deduce
the classical theory of the natural numbers. Gödel [1944] pointed to a gap in his proof, and Myhill [1974]
argued that this gap could not be filled. (Landini [1996b] offered to fill it, but his proof relied upon an
improbable assumption about the underlying logic that Russell had not stated explicitly. See Hazen and
Davoren [2000: 555].) As a matter of fact one can obtain the classical theory of natural numbers while
accepting the VCP, but only with the help of some extra assumptions, which a conceptualist should find
acceptable. The point is noted in my [2009a: 246–9]. But today’s conceptualist will not accept more than
denumerably many abstract objects, and so will not accept the classical theory of real numbers. There is a
review of conceptualist approaches in my [2009a: 244–59].
6
Ramsey had said: ‘As I can neither accept the Axiom of Reducibility nor reject ordinary analysis, I
cannot believe in a theory which presents me with no third possibility’ [1925a: 29]. (Note: ‘analysis’ here
means the theory of the real numbers.)
102 L O G I C A N D M AT H E M AT I C S

alternative to the VCP diagnosis that this paradox arose from an illegitimate quantifica-
tion over all propositions. It was not until later that he saw a different solution in Tarski’s
suggestion that it involved an illegitimate use of the notion of truth, on the ground that
‘true’ can be understood only as an expression in a metalanguage which is used to speak
of the sentences of an object language. Although the seeds of such a view may be found
in his introduction to Wittgenstein’s TLP [1921b: xxii], we do not find him applying
it to the paradox of the liar until his much later book An Inquiry into Meaning and Truth
[1940: 62]. But let us now take over this rather later Russellian thought, and so drop the
VCP as no longer needed. That is, let us just turn back to the simple theory of types, and
consider Russell’s construction of mathematics within that theory.

2 Mathematics and the simple theory of types

Russell was always convinced that his treatment of mathematics was a success. His
Introduction to Mathematical Philosophy [IMP, 1919a] explains in a very readable way
what that construction is, and the concluding chapter opens with a summary of
the position, as Russell saw it.
Mathematics and logic, historically speaking, have been entirely distinct studies. Mathematics has
been connected with science, logic with Greek. But both have developed in modern times: logic
has become more mathematical and mathematics has become more logical. The consequence is
that it has now become wholly impossible to draw a line between the two; in fact, the two are
one. They differ as boy and man: logic is the youth of mathematics and mathematics is the
manhood of logic. (p. 194)

The paragraph ends by offering a challenge:

If there are still those who do not admit the identity of logic and mathematics, we may challenge
them to indicate at what point, in the successive definitions and deductions of Principia Mathe-
matica, they consider that logic ends and mathematics begins. It will then be obvious that any
answer must be quite arbitrary. (pp. 194–5)

Let us take up this challenge.

PM as it stands makes use of three axioms which it hesitates to call truths of logic.
One is the axiom of reducibility, which has now been set aside. The other two are the
axiom of infinity (for individuals), and (a version of) the axiom of choice (for proposi-
tional functions). Both are needed. Let us take them in turn.
The axiom of infinity states or implies that there are infinitely many individuals, and
it is needed because if there are only finitely many individuals then Russell’s way of
construing the natural numbers will not verify Peano’s postulates (pp. 18–19).7 As I said

7
There are many ways of formulating such an axiom. Both Church [1956: 343] and Hatcher [1968: 120]
use this formulation: There exists a first-level relation R such that
(i) 8x(Rxx), (ii) 8x9y(Rxy), (iii) 8xyz(Rxy & Ryz ! Rxz).
 P RO B L E M S 103

in Chapter 2, his original idea was to identify the number n with the class of all
n-membered classes. However his no-class theory claims that talk about classes should
be reconstrued as talk about propositional functions, and so in effect8 a thesis about the
number n is construed as a thesis about the propositional functions (of second level)
which are true of just those propositional functions (of first level) that are true of just n
individuals. As a little reflection will show, these propositional functions are the familiar
numerical quantifiers ‘there are n individuals x such that —x—’. So claims about the
numbers are construed as claims about these quantifiers. The definition of the number
0 therefore becomes the definition of the quantifier ‘there are 0’, which is obvious:
90x(Fx) for ~9x(Fx)
Moreover, the relation of succession between numerical quantifiers may apparently
be defined in this way:
(9n)S(9m) for 8F(9nx(Fx) $ 9x(Fx & 9my(Fy & y6¼x)))
Given this definition, consider what will happen if there are just finitely many
individuals, say 3. It will follow that, for every numerical quantifier ‘9n’ with n>3,
the formula ‘9nx(Fx)’ will be false for all F. Hence every such quantifier will be
equivalent to every other, and so by this definition ‘(9n)S(9m)’ will be true for all
n>3, m3. But then the numerical quantifiers cannot obey Peano’s postulates for
natural numbers. On this construal, then, ordinary arithmetic requires us to assume that
the number of individuals is not finite.
One might hope that this problem could be overcome by changing the definition.
One suggestion that Russell considers is that we need not confine our attention to the
numerical quantifiers of the second level. For there will also be numerical quantifiers at
the next level up, which in the familiar notation will be written as ‘9nF(—F—)’, and
then again at the next level after that, and so on for ever. Moreover, supposing that
there are only 3 individuals, still there will be more than 3 predicates of individuals (in
fact there will be 23, i.e. 8, such predicates that are not equivalent to one another), and
still more predicates of these in turn (in fact 28), and so on. So, however many
individuals we have at our lowest level, for each numerical quantifier ‘9n’ there will
be a level in the hierarchy that is high enough to ensure that at that level the quantifier
does have some true applications. This is quite a suggestive point, which I shall come
back to (on p. 109). But for the present all that one can say is that, given Russell’s
background theory, it will not give us what we need. For we still have no guarantee
that there will be any one level which contains an infinity of distinct numerical
quantifiers. With some ingenuity, one could fix the definitions of the particular state-
ments of elementary arithmetic so that all the right ones came out to be true however

8
In detail, the definition of classes requires a somewhat roundabout mention of functions that are
equivalent to a given function. I ignore this complication, which makes no difference to the point being
argued.
104 L O G I C A N D M AT H E M AT I C S

high their numbers were. But, as Russell himself observes, there would be no way of
moving on from here to the theory of the real numbers or the theory of infinite
numbers. For that move must require us to get the effect of speaking of the class of all
natural numbers, which could not be done if there is no one level at which they all
exist. So, as he concludes in his very clear discussion of this problem in IMP chapter 13,
the axiom of infinity cannot be avoided.
One naturally asks whether we have any good reason to suppose that the axiom is
true, and on this point Russell simply concludes that we do not know. Usually one
assumes that space and time are infinitely divisible, and this could lead to a positive
answer, but Russell comments that even if today’s physical theory would endorse the
assumption still it is not clear that tomorrow’s will do the same. A very different
suggestion would be that one can be sure that our language contains infinitely many
expressions, but here Russell’s reaction would be that such things are not to be counted
as individuals. Perhaps a token expression, in a particular place at a particular time, might
be allowed as an individual,9 but there is no good reason to suppose that there are or
will be infinitely many of these. No doubt one may hold that there are infinitely many
types of expression, but these would be abstract objects, and Russell will not count
them as individuals. This puts him in a curious position. Like the rest of us, he is no
doubt quite convinced that there are infinitely many numbers, and so if numbers are to
be analysed as numerical quantifiers he should be equally convinced that there are
infinitely many of them. But he cannot prove this without calling upon an axiom
which he admits must be counted as doubtful. This strongly suggests that there must be
something wrong with his approach.
One remedy would be to find a way of proving a suitable axiom of infinity. I shall
later show (on p. 109) how this might be done.10 A different suggestion is that we
should alter the definition of succession between numerical quantifiers, so as to ensure
that it cannot hold ‘accidentally’. The natural way to do this is to enlarge the logic to a
modal logic, by adding the operators ‘it is necessary that’ and ‘it is possible that’,
symbolized by ‘□’ and ‘◊’ respectively. Then we could define succession by using a
necessary equivalence, thus:
(9n) S (9m) for □8F(9nx(Fx) $ 9x(Fx & 9my(Fy & y6¼x)))
This would not resolve the logicist’s problem, for to obtain Peano’s postulates for the
numerical quantifiers we should still need a quasi-axiom of infinity, stating that for each
n it is possible that there are n individuals, i.e.
◊9F 9nx(Fx)

9
I shall say more in Chapters 7 and 9 about just what Russell does count as an individual.
10
In earlier days, when Russell believed in propositions as individuals, he had sought to prove that there
are infinitely many propositions (e.g. [1906b: 181], [1906c: 203]). But by the time of PM he no longer
believes in propositions, as Chapter 12 will show.
 P RO B L E M S 105

This would still not be deducible from anything simpler, though it does have the
advantage over Russell’s axiom that it is evidently true. However I shall leave this
course unexplored. For Russell never looked with any favour on the modal logic that
would be needed, and he was apt to claim that it rested on an idea that was simply a
mistake (e.g. PLA, p. 231).
To sum up: Russell does not claim that his axiom of infinity is a truth of logic, and he
is surely right about this. But he also admits that this axiom is needed. So one could
reasonably answer his challenge ‘where does logic end and mathematics begin’ by
saying that it happens as soon as this axiom is invoked. It is somewhat strange that he
himself never seems to have considered this response. But the truth, I suggest, is that the
need for this axiom is a sign of a deeper problem with his construction, which I shall
come back to. Meanwhile, let us move on to consider the other axiom which he
invokes, but does not call a truth of logic, namely what he calls ‘the multiplicative
axiom’. It is better known these days as (a version of) the axiom of choice. He discusses
it in chapter 12 of IMP, and comments that it is ‘convenient, though not indispensable,
in certain portions of mathematics’ (p. 117). A better comment, I think, would be that
the basic idea behind Cantor’s theory of infinite numbers cannot be accepted unless
this axiom is also accepted.11
Cantor claimed that two classes had the same number of members if and only if there
exists a relation which correlates their members one-to-one. The axiom of choice is
needed to assure us that suitable correlating relations do always exist, even when we
cannot define them. In IMP, pp. 126–7, Russell offers a nice illustration of this point,
comparing boots with socks.12 Imagine that we have some pairs of boots, and some
pairs of socks, and a relation which correlates the pairs of boots one-to-one with the
pairs of socks. So there are as many pairs of the one as there are pairs of the other.
It must follow that there are also as many boots as there are socks, but we need the
axiom of choice in order to prove that there is indeed a relation which correlates the
individual boots with the individual socks. With the boots there is no problem, for each
boot is either a left boot or a right boot, and each pair contains just one boot of
each kind. If we could say the same of the socks then we could easily specify a relation
which does the trick. Just correlate the left boot of each pair of boots with the left sock
of the corresponding pair of socks, and similarly the right boot with the right sock, and
we have what we want. But, unlike boots, socks are not in fact distinguished from one
another as left socks and right socks, so we cannot select from the pairs of socks in
this way. Nor can we specify any other way of selecting, which is guaranteed to work.
So here we have to invoke the axiom of choice, which assures us that there is a way
of selecting just one sock from each pair, i.e. that there is a propositional function

11
There are many equivalent ways of formulating this axiom. I shall suggest the most relevant form on the
next page .
12
I have made a small alteration to his example. Russell specifies that there are denumerably many pairs of
boots, and of socks, whereas I have left this number unspecified.
106 L O G I C A N D M AT H E M AT I C S

which is true of just one of the socks from each pair of socks. It then follows that there is
a relation, which (say) correlates each left boot with a selected sock, and each right boot
with an unselected sock, so we have what we want.
In the case in which we have only a finite number of pairs to begin with, the axiom
need not be invoked, for we can prove that the selection exists. In effect, we show that
the required propositional function can be specified just by listing the socks of which it
is true. (We can of course list just one item from a single pair. If we can list just one item
from each of n pairs, then we can do the same for n+1 pairs, just by adding one more
item to an existing list. So by induction a suitable list will exist for any finite number of
pairs.) But a list of items, which are not distinguished by any shared property, except
that of belonging to the list in question, can be given only when the list is finite. And
Cantor’s criterion for sameness of number is obviously correct for the finite case. But
Cantor’s theory extends it to cover the infinite cases as well, and that is where we shall
need an axiom of choice in order to deduce what seem to be evidently correct results
about infinite numbers. The axiom can be formulated in a way which makes this role
particularly salient. We may define the relation ‘⪯’ in terms of one-one correlations in
the obvious way, and then state as an axiom, for all classes a and b of any level:13
a⪯b v b⪯a
Cantor’s own proofs sometimes assume this axiom, but without acknowledging it.
It was brought to public attention by Zermelo [1904], and since then many different
but equivalent formulations have been discovered. (They may be found in any
textbook on set theory.) It is, in my view, an indispensable part of Cantor’s theory of
infinite cardinal numbers, but one might defend Russell’s comment by saying that
Cantor’s theory could itself be dispensed with. If one takes the view that propositional
functions (and relations) exist only when they can be specified, then one must
presumably reject this axiom. But now that we are no longer trying to be sympathetic
to the VCP we need feel no temptation towards this view. On the contrary, given that
validity for the simple theory of types is to be understood as was proposed on pp. 67–8,
the axiom must presumably be counted as valid (and valid at all levels of the theory).
As I have said, there is no way of providing a complete set of axioms for the simple
theory of types, or even just for a logic of second level (p. 71). So it is hardly surprising
that the rules adopted for the basic system in section 5 of Chapter 4 should need to be
supplemented by further axioms, if we are to reach a satisfactory system. (What is
mildly surprising is that, in all of the last century, this one further axiom is the only one
to have found very general acceptance.) Whether the axiom deserves to be called
a truth of logic rather than a truth of mathematics may be debated, and on this occasion

13
In this form the axiom is usually called ‘the trichotomy’, since a little redefinition easily provides the
equivalent version
a≺b v ab v ab
 P RO B L E M S 107

I am inclined to agree with Russell that a decision can only be somewhat arbitrary. But
it is another point where one might wish to claim that his challenge could be met.
I add one P.S. on the logicist claim about mathematics. Gödel showed in his [1931]
that, whatever rules of proof are adopted for our logical system, there will be truths of
elementary arithmetic that are not provable in the system. (We prove them only
by stepping outside the system.) This has been taken to show that mathematics is not
reducible to logic. But it may be replied that, if we have a suitable conception of
validity for our logic, then all arithmetical truths will be valid in the system, even if not
all of them are provable in it. So perhaps mathematics can be ‘reduced’, not to logical
provability, but to logical validity. If so, would this be enough to establish the logicist
thesis?
But without further discussion let us now leave aside the question of logic versus
mathematics, and turn more directly to what Russell clearly viewed as logic, namely
the theory of types itself. For it has one very evident weakness.

3 The weakness of the simple theory

The simple theory is a theory of individuals, and of their predicates, the predicates of
these in turn, and so on upwards. The individuals can be named, but none of the
predicates can. So the theory employs a language which is far more restrictive than the
English language (or any other natural language), since in English we can always form a
name from a predicate—either a name for the predicate itself, or for the property that
it expresses, or for the class of things that it is true of, and so forth. But this is to treat
the predicate as if it stood for an individual, and the theory of types will not allow this.
In consequence it has no grammatical way of generalizing in one breath over the
predicates of different levels; we are simply not given the language to do it with. But
this is to hamstring both logic and mathematics.
The effect on logic is easily seen by reflecting on our earlier discussion of the
predicate ‘ . . . is transitive’ (pp. 55–7). In English we can say of a relation of any level
that it is transitive. We do this by introducing a name for the relation, and putting that
name into the gap of the English predicate ‘ . . . is transitive’. In the language of the
theory of types that cannot be done. Consequently in that language we also cannot
form any general statements that concern all transitive relations of whatever level. For
example, we cannot say that all such relations must be asymmetric if they are irreflex-
ive.14 But this, that we cannot say, is (in our normal way of thinking) quite evidently
true. The same applies to almost all theses that are of interest to the logician.
A similar point applies to Russell’s construction of mathematics, which is based
on the numerical quantifiers ‘there is 1 . . . ’, ‘there are 2 . . . ’, ‘there are 3 . . . ’, and so
on. Just like the universal and existential quantifiers of ordinary logic, these numerical

14
A first-level relation R is asymmetric iff 8xy:(Rxy & Ryx); it is irreflexive iff 8x:Rxx. Of course the
same idea applies to relations of any other level.
108 L O G I C A N D M AT H E M AT I C S

quantifiers can also be applied to predicates of any level. That is, whatever the level of
the predicate f and its subjects a, one can always claim that (e.g.) there are just 4 items
a such that fa. But this means that the quantifiers in question cannot be assigned to
any level, and in that case the theory of types cannot admit their existence. Given a few
very reasonable assumptions, it can speak of all the numerical quantifiers of level 2, and
all those of level 3, and so on for each level. But it cannot recognize that the same
quantifiers occur at each level, though it seems obvious to us that they do.
The best that Russell can do by way of admitting this point is to talk of ‘systematic
ambiguity’. Where some one and the same notion seems to occur at many different
levels of the theory of types, Russell has to say that this is a mere appearance. The same
word may be used, but cannot mean the same at each level, so it has to be said to be
ambiguous. The ambiguity is said to be systematic because there is an evident analogy
between the supposedly different meanings at the different levels. But the theory insists
that they must really be different meanings. Russell is under no illusions about the scale
of this phenomenon. He says in PM that in all the paradoxes he considers
The appearance of contradiction is produced by the presence of some word which has systematic
ambiguity of type, such as truth, falsehood, function, property, class, relation, cardinal, ordinal, name,
definition. (p. 64)

And he goes on to say that

In spite of the contradictions which result from unnoticed typical ambiguity, it is not desirable to
avoid words and symbols which have typical ambiguity. Such words and symbols embrace
practically all the ideas with which mathematics and mathematical logic are concerned. (p. 65)

That is: in logic and mathematics we have to use words which are ambiguous in order to
secure the generality desired. But, according to the theory of types, those words must be
ambiguous, even though we never notice the supposed ambiguity. (And we do not
notice it because, after all, we speak a natural language; we do not speak the language of
the theory of types.)
This position seems to me to be clearly unsatisfactory. In my [1980] I have proposed
a solution which claims that, as well as the predicates of a given level, which type theory
accepts, there are also predicates of a different kind, which can occur at many levels.
These are called ‘type-neutral’ predicates, and I have argued that reflection on what a
predicate is—namely, a certain kind of pattern shared by many sentences—will show
that they deserve to be called predicates, even though they do not fit into the hierarchy
of the theory of types. A first step towards the recognition of such predicates is that we
do actually apply the same quantifier-symbols 8 and 9 at every level of the theory
of types, and we do not think of them as having different meanings at each level.
(The suggested explanation of these quantifiers, namely that ‘8a(—a—)’ counts as true
iff ‘(—a—)’ is true for every permitted interpretation of the variable ‘a’, is one that
applies unchanged at every level, whatever the type of the variable ‘a’.) So 8 and 9
should be regarded as type-neutral predicates. Hence, whatever can be defined just in
 P RO B L E M S 109

terms of 8 and 9 and the truthfunctors should also count as a type-neutral predicate. So,
if we may assume that identity can be defined in these terms—i.e. by defining ‘a¼b’ as
‘8f(fa $ fb)’—then the numerical quantifiers are also included. So too is our
predicate ‘Trans’ discussed earlier, and apparently almost all of the predicates which
Russell has to regard as having ‘systematic ambiguity’. But it is not obvious that this has
to lead to any contradiction. That will depend upon the detail of the logical system
proposed.
In my [1980] various systems are given in detail, systems which increase in strength
from each to the next, and so far as I know none of them gives rise to any contradiction.
Even in the weakest of these systems it is possible to define each (finite) numerical
quantifier in a suitably type-neutral way, and to avoid the need for an axiom of infinity
by showing that each has some true applications. This is simply a consequence of the
earlier observation that for each such quantifier there must be some level of the
orthodox type hierarchy at which it is true of something, and that is all that is needed.
For with this new conception the quantifiers themselves are not located at any one of
the orthodox levels, but are predicates of a novel kind that can be applied to all the old
levels. It is true that one must progress beyond my weakest system in order to obtain a
properly satisfying general theory, both of the finite numerical quantifiers and of other
type-neutral predicates, but I shall not go into any further details here. For it has to be
admitted that the details do become unpleasantly complicated. This is because it is not
easy to harmonize some very natural thoughts about type neutrality with the
overriding need to avoid contradictions. I think it probable that the discussion of my
[1980] could be improved, but I do not here propose any particular improvement. This
is a task that I recommend to others.
There is a further question that arises from Russell’s way of constructing elementary
arithmetic: supposing that we do have a satisfactory theory of the (finite) numerical
quantifiers as type-neutral predicates, how is such a theory related to the ordinary
theory of elementary arithmetic? Arithmetic, after all, treats not of numerical quanti-
fiers but of numbers. Is Russell right to identify the two? If our concern is to give an
account of the ordinary concept of a natural number, as used either by the man in the
street or the practising mathematician, it can certainly be argued that he is not.15 But
this was not what Russell was trying to do. He was concerned not so much to give an
account of the existing concept as to find a replacement for it, a replacement that
would still be adequate for all practical purposes, but would also have the advantages of
being both clearer and less committed to entities of a dubious nature. As we shall see
more fully later, especially in Chapter 9, this was always what he aimed for in his
‘logical constructions’. And when the task is viewed in this way then one can fairly say
that there is no objection to replacing the ordinary theory of the natural numbers by

15
I have given such an argument in my [2009a: 282–9].
110 L O G I C A N D M AT H E M AT I C S

the theory of the associated numerical quantifiers, for all other uses of these numbers
can perfectly well be explained from that basis.
If there is a doubt, it is this more complicated one. Russell wished his construction of
the natural numbers to lead on in a natural way to Cantor’s theory of infinite numbers.
He does in fact identify the natural numbers with the numerical quantifiers, and one
can similarly think of Cantor’s infinite cardinal numbers as infinite numerical quanti-
fiers. But if it is agreed that the finite numerical quantifiers should be construed type
neutrally, then the same should follow for the infinite ones too. However it is not at
all clear (to me) just how far into the Cantorian world of the infinite we can go, while
still retaining the idea of type neutrality. Since the issue soon becomes quite complex,
I shall make no attempt to explore it here, but I merely note that it is an issue.
It is reasonable to ask ‘why bother with such questions?’ Can we not say that today’s
mathematics is adequately served by today’s set theory, and no one expects to find any
contradictions lurking there? 16 Moreover today’s set theory seems to rest upon a basic
idea which appeals to everyone: sets are to be thought of as ‘built from’ their members.
(This is no doubt a metaphor, but it is one that is very persuasive.) That is why sets are
the same if and only if their members are the same, why sets fall into a (cumulative)
hierarchy, why there cannot be unfounded sets, and so on. All this turns out to be very
well suited to today’s mathematics, and mathematicians need feel no pressure to try to
develop what Russell saw as a basis. For his theory is not a theory of sets, but of
propositional functions, i.e. of predicates, or of the properties which those predicates
express. And while mathematicians need not be interested in properties rather than
sets, philosophers have been interested in properties for centuries. Moreover, no one
would think of properties as ‘built from’ the objects that have them. The basic idea is
quite different.
It seems to me that Russell’s (simple) theory of types is a good start on how to think
of such things as properties are supposed to be, but also that it needs to be supple-
mented by something like the idea of a type-neutral predicate, if obvious inadequacies
are to be avoided. But there I shall leave the account of Russell’s logic, and move on
to other matters.

16
I mean: no one expects there to be any contradictions lurking in the basic set theory of Zermelo-
Fraenkel. Of course there are also more speculative extensions of that theory (e.g. adding large cardinals)
where one cannot be so confident of consistency.
 PART II
Knowledge

Russell distinguishes two grammatically different uses of the English verb ‘to know’. It
may be used with a direct object, as when one is said to know a person, or a place, or a
tune, and so on. Or it may be used to govern a that-clause, either explicitly, as when
one is said to know that P, or with a prefix added, as when one is said to know the fact
that P, or the truth that P, and so on. In some other languages different verbs are used in
these two constructions, e.g. in French connaı̂tre for the first and savoir for the second,
and in German kennen for the first and wissen for the second. The first Russell generally
calls ‘knowledge of things’ or ‘knowledge of objects’, and the second ‘knowledge of
facts’ or ‘knowledge of truths’. In what follows, Chapter 7 will concern the first and
Chapters 8 to 11 will concern the second.
Within the general topic ‘knowledge of objects’ Russell makes a further distinction,
between what may be regarded as a ‘direct’ knowledge of an object, which is a matter
of knowing that object ‘by acquaintance’, and a more indirect knowledge, which is
knowledge ‘by description’. The latter turns out to be a special case of knowledge of
truths, for it is simply a matter of knowing that there is one and only one object that
satisfies the description. By contrast, he claims that being acquainted with an object is
logically independent of all knowledge of the truths that concern it. Moreover, the
point of philosophical interest in what he has to say about it is the connection that he
draws between being acquainted with an object and understanding propositions about
it. This is embodied in what is usually called his ‘principle of acquaintance’, namely
Every proposition which we can understand must be composed wholly of constituents with
which we are acquainted.

He introduces and discusses this principle mainly in the article ‘Knowledge by

Acquaintance and Knowledge by Description’ [1910/11] and in the very similar
chapter 5 of his book The Problems of Philosophy [1912a]. These will be the main
concern in Chapter 7, and their topic could equally well be described as ‘Russell’s
views on our knowledge of meanings’, at least in the period prior to his Analysis of Mind
[1921a].
112 K N OW L E D G E

Chapter 8 turns to his account of our knowledge of truths, beginning with what he
has to say on the topic in most of the rest of his book The Problems of Philosophy. This is a
very clear exposition of what is now known as the theory that knowledge has
‘foundations’, and Chapter 8 will give Russell’s account of what the foundations are.
He always retained his foundational view of knowledge, but very soon after writing
The Problems of Philosophy he began to feel that its gestures towards the further
knowledge that could be built from these foundations had been too optimistic. They
had not allowed for the full force of the sceptic’s attack. Consequently we find a
new approach in the article ‘The Relation of Sense-Data to Physics’ [1914b] and in
chapters 3–4 of the book Our Knowledge of the External World [1914c]. This proposes
a ‘construction’ of the material world which aims to give due weight to scepticism, but
without yielding to it. It is the subject of Chapter 9. After some further changes of view
Russell was then led to a similar ‘construction’ of mind which is announced in his
article ‘On Propositions: What they Are and How they Mean’ [1919b] and fully
developed in his book The Analysis of Mind [1921a]. This is the topic of Chapter 10.
The two proposed ‘constructions’, of matter and of mind, combine to yield Russell’s
version of the theory called ‘neutral monism’. Chapter 11, which concludes Part II,
steps outside the period of what can fairly be called Russell’s philosophy of logical
atomism, in order to give a brief indication of how, in his book The Analysis of Matter
[1927a], Russell later felt the need to modify these theories in a way which in fact
abandoned their central feature. This chapter is included only for those who are
curious; it may be omitted without affecting anything else in this book.
 7
Acquaintance

This chapter will be based mainly on the article ‘Knowledge by Acquaintance and
Knowledge by Description’, which I abbreviate to ‘KAD’, and the corresponding
chapter 5 of The Problems of Philosophy, which I abbreviate to ‘PP ’. They introduce and
explain Russell’s principle of acquaintance, namely
Every proposition which we can understand must be composed wholly of constituents with
which we are acquainted (PP, 58; KAD, 219).

He offers no justification for this principle, except one that is easily seen to beg the
question:
It is scarcely conceivable that we can make a judgment, or entertain a supposition, without
knowing what it is that we are judging or supposing about (PP, 58; KAD, 219).

But ‘knowing what one is thinking of’ does not entail ‘being acquainted with what one
is thinking of’, as Russell himself would have to admit. For he also accepts that there is
such a thing as knowledge (of objects) by description, and he wishes to contrast this
with knowledge by acquaintance. So for the moment all that we can say is that Russell
claims that understanding always depends (ultimately) on acquaintance, but we must for
ourselves investigate what exactly is being claimed, and whether the claim is or is not at
all convincing. We may begin with the observation that Russell asserts that there is
acquaintance both with particulars and with universals, and he explains it rather
differently in each case. Sometimes he adds that there is also acquaintance with items
of a more logical nature, but this is a matter on which he never formed a very settled
opinion. We shall look into all these claims, but let us begin with the idea that there
is and must be such a thing as acquaintance with particulars.

1 Acquaintance with particulars: The idea of

direct reference
Already in the second paragraph of On Denoting Russell had introduced the distinction
between being acquainted with some object (e.g. by perceiving it) and knowing of
it only via some description (i.e. as ‘the thing which . . . ’). The theory of definite
descriptions expounded in that article is supposed to show us how, when a proposition
contains a definite description, it is not the object described that is a ‘constituent’ of the
114 K N OW L E D G E

proposition, but rather the ‘constituents’ of the description itself, i.e. the universals that
answer to the predicates contained within it, and (possibly) the particular objects named
by any genuine names contained in it, and (perhaps?) whatever objects of acquaintance
explain our understanding of the logical apparatus involved (i.e. truthfunctors and
quantifiers and identity). Without here going further into the details, let us concede
that I can understand a description of an object by understanding the ‘constituents’ of the
description, and without being acquainted with the object described. But, Russell
claims, this cannot apply to all the propositions concerning particular things that I can
understand. For some of them must include a ‘direct reference’ to a thing which is not
given wholly by a definite description. In such a case we can have a genuine (‘logically
proper’) name of the object in question, and I understand this name only because I am
acquainted with that object, e.g. in perception.1
This is a fair claim, and here is a sketch of an argument for it. Suppose that I have a
thought which I might express (to myself) as ‘That thing, on the table in front of me, is
a cup of tea’. You might offer a paraphrase, which refers to the cup by a description,
e.g. ‘The one and only thing that is a cup, and is on the one and only thing that is a
table in front of me, contains tea’. There may be several objections to this paraphrase,
but for the sake of argument let us accept that it represents my thought well enough.
Still, it has not eliminated all direct references to particular things, for it retains a
reference to the particular thing that is me. Perhaps we could go on to eliminate that, by
finding a description that I and only I satisfy. One that sometimes appealed to Russell
was ‘the owner of this sense-datum’. But all that this does is to exchange what seems to
be a direct reference to me for another direct reference, i.e. to a particular sense-datum.
It does nothing to advance the idea that all direct reference can (in theory) be
eliminated in favour of definite description.
Indeed, that idea commits one to a very improbable version of the Leibnizian
principle that indiscernibles are identical, i.e. that two different objects cannot share
all the same properties. Now, when the notion of a property is widely understood, so
that any one-place predicate is taken to introduce a property, this principle is trivially
true. For if a and b are any two distinct objects then ‘a ¼ a’ is true and ‘b ¼ a’ is false, so a
has a property which b lacks, namely the property of being identical to a. But suppose
that such properties, which themselves include a reference to a particular individual, are
not to count, and (as is traditional in this topic) we restrict attention to what may be
called ‘pure’ properties, containing no such references. Then is it at all probable that
any two things will differ from one another in at least one of their pure properties?2
Even if that happens to be true, it surely does not follow that I myself can distinguish

1
For the present I mean by a ‘direct reference’ just the negative description given here, i.e. it is a reference
to a particular object that is not to be explained as made by the use of a definite description which is satisfied
just by that object and by nothing else. But I take the expression ‘direct reference’ from Kaplan, whose essay
on ‘Demonstratives’ [1989] gives a much more positive account. I shall consider his account in section 7 of
this chapter.
2
Classic discussions are Black [1952] and Ayer [1953].
 A C Q UA I N TA N C E 115

them in this way. Of course, if both are perceived by me, then I can distinguish
them by the way that they appear to me, say as ‘the one that is to the left of me’ and ‘the
one that is to the right of me’. But, as before, this introduces what looks like a direct
reference to me, and I am a particular thing. Could I distinguish myself from everything
else by citing a pure property that I have and that nothing else does? It seems very
improbable.3
If this point is granted, the consequence must be that if I can refer to any particular
thing at all, then there must be some particular thing (even if only myself) that I can refer
to in a ‘direct’ manner, which does not depend upon my ability to formulate a
Russellian definite description, satisfied by it and it alone. In fact Russell’s own
suggestion is more generous: I can refer in this ‘direct’ way to whatever I am at the
time perceiving. I merely have to point (in a mental way, to myself) and to say in my
own language of thought ‘this’. The word ‘this’, so used, is what Russell will call a
‘logical proper name’, and he will allow that I may introduce such a logical proper
name for anything that I am currently perceiving. Whether there are other things that
I can name in this way, is a question that we may set aside for the time being. The case
of an object now being perceived is a good place to start.
So we next ask: ‘and what exactly do I perceive?’ A very traditional answer, given by
Locke and Berkeley and Hume, is that I perceive what they call ‘ideas’, meaning of
course that I perceive ‘my own’ ideas, for I do not perceive yours.4 Then one goes on
to think that ‘ideas’ are things that exist only in minds, and hence that what I perceive
must be in my own mind. Russell rightly objects that this is a fallacy. The act of
perceiving is no doubt something that exists only in a mind, but that need not be true
of the thing perceived. (This is the substance of his criticism of Berkeley in PP, on pages
38–43.) So the way is open, one would suppose, to the claim that our ordinary
everyday way of thinking is correct, and we can perceive such ordinary objects as
tables and chairs, or trees and birds and people, and so on. But in fact Russell denies
this, on very much the same grounds as did Berkeley and Hume. He says
Let us give the name of ‘sense-data’ to the things that are immediately known in sensation: such
things as colours, sounds, smells, hardnesses, roughnesses, and so on.5 We shall give the name
‘sensation’ to the experience of being immediately aware of these things. Thus, whenever we see

3
There are (philosophically motivated) stories of entire parallel worlds, in which my ‘doppelganger’ does
everything that I do, but in his world rather than in mine. How could there be a ‘pure’ property that will
distinguish me from him?
4
In this context Hume spoke not of ‘ideas’ but of ‘impressions’. This (desirable) change of vocabulary
makes no difference to the underlying doctrine.
5
As the context makes clear, Russell means to be talking of particular patches of colour, particular
occurrences of a sound, and similarly in other cases. (Nasim [2008: ch. 4], is probably right in saying that
Russell’s use of the word ‘sense-datum’ is taken from G. E. Moore, who had used the word in this way in
lectures that he had given in 1910. (Russell’s PP acknowledges a debt to Moore’s ‘unpublished writings’ in its
preface. These writings were subsequently published as Moore [1953].) But the word was not new. Its first
occurrence is apparently due to Josiah Royce [1885: 321], and it was afterwards used by William James, e.g. in
his [1890] and his [1912], both of which Russell knew.)
116 K N OW L E D G E

a colour, we have a sensation of the colour, but the colour itself is a sense-datum, not a sensation
(PP, 12).

And he goes on at once to say that what we perceive is always a sense-datum, i.e.
something given to us in sensation, such as a particular patch of colour, or hardness, or
whatever. It is not a physical thing such as a table; that is not given to us in sensation, but
is at best inferred as something that lies behind the sense-data that we are conscious of
(pp. 8–11). There is therefore a real question over whether the table itself exists at all,
and—if it does—what sort of a thing it is ‘in itself’, i.e. apart from its providing these
sense-data for us (p. 12ff.). In any case, he does not accept the ordinary view that I often
perceive such things as tables. On the contrary, the only things that I perceive are my
own sense-data, and this looks very much like Berkeley’s view, even though he does
insist that it is different.
One asks: why? Assuming (as Russell will) that tables do exist, why can we not be
said to perceive them? Or, in Russell’s own language, why do we not count as being
‘acquainted’ with them? The only reason that we have had so far is that it is possible that
they do not exist. But why is that relevant? Russell is clearly influenced by the point
that my perceptions might be indistinguishable from what they now are, even though
there is not actually any table, and I am suffering from a hallucination. That point may
be accepted while we still maintain the ordinary view that, so long as there is in fact a
table there, then I do count as perceiving it. Since hallucinations are possible, I may
think that I am perceiving a table when I am not, but we can admit that without
adopting Russell’s extreme view that I never perceive such things as tables.
One explanation of his position fastens upon the point that he describes acquaintance
as a kind of knowledge. This seems to imply that if I am acquainted with a so-and-so then
I must know that it is a so-and-so that I am acquainted with, which apparently rules out
acquaintance with physical objects such as tables. For one cannot, strictly speaking,
know, just from one’s perceptions, that there are any such things. But this is to ignore the
contrast that Russell draws between knowledge of objects, which is acquaintance, and
knowledge of truths, including truths about the nature of those objects. Acquaintance, he
claims, does not require any knowledge of truths about the object with which one is
acquainted. For example, when he looks at his table, he says
The particular shade of colour that I am seeing may have many things said about it—I may say
that it is brown, that it is rather dark, and so on. But such statements, though they make me know
truths about the colour, do not make me know the colour itself any better than I did before: so far
as concerns knowledge of the colour itself, as opposed to knowledge of truths about it, I know
the colour perfectly and completely when I see it, and no further knowledge of it itself is even
theoretically possible. (PP, pp. 46–7)

What Russell describes as ‘knowing the colour’ is really just a matter of noticing the
sense-datum, so that one can at least refer to it (in one’s thought) as ‘this’. In order to
do so, one does not have to know that it is something coloured, or to know any other
truths about it at all. No doubt one will in fact have beliefs about the thing perceived—
 A C Q UA I N TA N C E 117

e.g. in Russell’s example that it is a visual sense-datum, that it has a colour, and indeed
that its colour is brown—but his point is that one does not need any such beliefs if
one is to be acquainted with the thing. And if one has them still they may all be
false without destroying one’s acquaintance.6 As Russell himself admits there are some
beliefs such that, if one has them in the situation envisaged, then they cannot be
mistaken, e.g. the belief that one is currently acquainted with the thing in question. But
obviously you do not have to believe this in order to be acquainted with it, for you may
simply lack the concept of acquaintance.7 It is a fair complaint that, if this is what
Russell has in mind, then he should not have described acquaintance as ‘knowledge of
objects’, for it bears almost no resemblance to what we mean when we talk of knowing
a person, or a place, or a book, or anything like that. But it is quite clear that this is his
position.
It follows that we cannot explain in this way why he will not allow me to be
acquainted with my table. In fact the only decent explanation that I see is one that takes
a different direction, and emphasizes the connection that he draws between being
acquainted with a thing and being able to name it, i.e. to use for it what he will accept as
a genuine (‘logical’) proper name. To see the point here, let us return to the argument of
On Denoting [1905b]. As we observed in Chapter 3, Russell recommends his theory of
definite descriptions partly because of its ability to solve certain puzzles. These effec-
tively reduce to two: (i) how can ‘a does not exist’ ever be both meaningful and true?
(ii) how can ‘a ¼ b’ ever be both true and informative? One should concede that
Russell’s theory does provide a solution to these puzzles in the case where ‘a’ is a
definite description. But at the same time we observed that the same puzzles seem also
to arise when ‘a’ is not a description but an ordinary name. So what Russell offers as a
solution would seem to be at best a partial solution, for there are many cases which it
does not cover. Nevertheless he wishes to maintain that his solution is the complete
solution, and he does this by subsequently expanding the original idea of what counts
as a definite description, while simultaneously shrinking the idea of what should be
counted as a name. Even in the original article On Denoting there are hints of this
development, for we have seen (on pp. 43–4) that he there claims that what is
grammatically a name, but has to be regarded as naming nothing, should really be
regarded as not a name but a definite description. But we now see that he is generalizing
this thought: whenever it even makes sense to suppose that something might not exist,
that thing must be a thing that cannot be named but can at best be definitely described.
Similarly, to deal with the puzzle over identity, whenever it even seems to be a possibility
that ‘a ¼ b’ is both true and informative, then at least one of ‘a’ and ‘b’ must be not a
name but a description. Hence anything that can present different appearances, whether
at the same time or at different times, cannot be named but can only be described.

6
All of them? This is a question that I shall return to at the end of the chapter. But Russell himself says ‘all’,
and what he says does apply to almost all.
7
Russell himself makes this point in his later treatment ‘On the Nature of Acquaintance’ [1914a: 168].
118 K N OW L E D G E

It may well be said that a sense-datum of mine will satisfy these two constraints: on
Russell’s view sense-data are very short-lived, and therefore the same sense-datum
cannot appear to me at different times; on anyone’s view the same sense-datum cannot
appear differently to me at the same time. Hence, when a and b are both sense-data of
mine, it can never be a real question for me whether a ¼ b. Similarly, it can never be a
real question to me whether a exists, for if I sense it then of course it does exist. That is:
my own sense-data satisfy (for me) the two conditions on what can be named that can
be derived from On Denoting, but it is not at all clear that the same applies to anything
else that I could be said to experience.8 An interesting article by Donnellan [1990]
argues that what I mean by ‘I’ will satisfy the conditions, even if I am not acquainted
with myself. The same might perhaps be said of ‘here’ and ‘now’. We may note that
Russell always had doubts about ‘I’, but I postpone this topic until the start of Chapter
10, when we shall see how he eventually came to reject this view. As for ‘here’ and
‘now’, he will argue that these are short for ‘this place’ and ‘this time’, and therefore
presuppose our ordinary theory of a public space and a public time that is the same for
all of us. But as we shall see in Chapters 9 and 10 he will regard these as at best logical
constructions, and will note that the ordinary theory is anyway put in doubt by the
considerations that lead to Einstein’s theory of relativity. So we may set these sugges-
tions aside without more ado.9 Meanwhile, let us ask what else I might be supposed to
be acquainted with.

2 Digression: A causal theory of direct reference

Along with sense-data, as ordinarily construed, Russell certainly includes the data of
‘inner sense’, such as one’s feelings of pain and hunger, or (e.g.) the feelings that are in
fact produced by an excess of adrenaline in the bloodstream and that accompany an
emotion such as fear. When he was writing PP he also included perceptions that were
not at the time being experienced, but were being remembered (e.g. PP, pp. 48–9,
114–15, 117; KAD, pp. 210). But he soon revised his view of this, on the ground that
what we seem to remember may not have occurred at all, from which it follows that
the existence claim must in this case be admitted to be meaningful. (As he famously said
somewhat later, in The Analysis of Mind [1921a: 159], it is logically possible that the world
sprang into being five minutes ago, with all our apparent memories just as they are, and

8
Hence being acquainted with a is not really independent of all knowledge of truths about a, for it carries
with it the knowledge that a exists. Similarly, being acquainted with both a and b must carry with it either the
knowledge that a¼b or the knowledge that a 6¼ b. But Russell thinks of these items of knowledge as not really
counting. (See further Chapter 14, section 4.)
9
In his later book An Inquiry into Meaning and Truth [1940] Russell suggests that we could take ‘I-now’ as
an object of acquaintance. Then we could explain ‘here’ and ‘now’ as short for the descriptions ‘the public
place which I-now occupy’ and ‘the public time which I-now occupy’. According to Kaplan’s theory, which
I give in section 7 below, to get the right effect these descriptions should be understood simply as reference-
fixing, and perhaps ‘rigidified’ by explicitly adding the word ‘actual’. (See note 30.)
 A C Q UA I N TA N C E 119

none of our current experiences will show conclusively that this hypothesis is false.) In
fact it is now clear that he had given up the idea that I am presently acquainted, by
memory, with what I once experienced in the past, by the time that he was working at
a book on The Theory of Knowledge in May and June of 1913. (The book was never
completed, and never published.10) He retains the idea only for what he calls ‘imme-
diate memory’, which is my present experience of what happened only a moment ago,
and which still counts as part of what I am experiencing now, since it is still part of my
‘specious present’, though I am aware that its original experience was in the past. (On
his account, one needs such a view of ‘the specious present’, as including some past
experiences that still persist, in order to explain how we can tell, from our present
experience, that something is now moving. For it seems that this must involve a present
perception both of its past positions and of its present position, if we can see that they
are not the same.)
If we set aside this special point about ‘immediate’ memory, and ‘the specious
present’, we may say that Russell quite soon came to reject the idea that if I was
once acquainted with x, and still remember it, then I am still acquainted with it.11
Nevertheless it is quite a natural idea, and apparently it will explain why, when I use a
name ‘a’ to refer to x, it is indeed x that I am referring to, and not something else which
may be exactly like it. The explanation is that it was my past contact with x that now
causes my present use of ‘a’, and we have here the seeds of what is now a popular theory
of ‘direct reference’. To put it very roughly the idea is that I can now refer in this
‘direct’ manner, simply by name and without any associated description, to whatever it
is that lies at the causal origin of my present use of the name. The first step in pursuit of
this idea would be that I can refer directly, not only to what I am now experiencing, but
also to what I did once experience, where it is the past experience that causes my
present memory of it. As a further step in the same direction, I can similarly refer not
only to my present and past sense-data, but also to the physical objects that (as we
believe) are or were their causes. Pressing the idea still more, one may come to
something like the present ‘causal theory of reference’ for ordinary proper names of
ordinary objects, past or present, e.g. as proposed in Kripke’s Naming and Necessity
[1972], especially pp. 22–97. The basic idea is that, if we trace the causal origin of a
present thought, we can explain why the thought is a thought about one particular
thing, rather than another which is intrinsically just like it. But I do not here expand on
such a theory, because it is clear that Russell himself does not follow it, and it is also

10
The part of it that he did succeed in drafting is now published as volume 7 of his Collected Papers
[i.e. Russell 1984]. He was apparently dissuaded from continuing with it by the criticisms of Wittgenstein.
But he evidently thought that the first six chapters were worth rescuing, for they were published (perhaps in a
revised form) in The Monist for 1914 and 1915. The first three are reprinted as ‘On the Nature of
Acquaintance’ [1914a] in Russell [1956], and we shall hear more of them later (in Chapter 10).
11
A classic discussion of Russell’s change of view on memory is Pears [1974]. It is supplemented by
Perkins [1976], and all the evidence now available is usefully put together by Miah [1998: 24–8].
120 K N OW L E D G E

clear why he does not.12 For it cannot ensure that the name ‘a’, used as I use it, must
succeed in referring to something. Equally it cannot ensure that two names ‘a’ and ‘b’,
used as I use them, cannot turn out to be both names of the same thing. But Russell is
committed to saying that these points must be ensured if the names in question are to
be genuine names, and not concealed definite descriptions. A definite description may
(by courtesy) be regarded as referring to a particular object, but it refers only indirectly,
and via the universals that occur in the description.

3 Indirect reference to particulars

Of the things which I ordinarily think of myself as perceiving—e.g. tables and chairs,
trees and birds, and of course other people—Russell will allow only that I perceive the
sense-data which (as we think) they cause me to have. But clearly I can think of these
other things, even if, as Russell claims, I do not perceive them. He would agree, for he
thinks of them as known to me ‘by description’, though that is perhaps a little optimistic.
But at any rate I can use definite descriptions to describe them in my thoughts, and
these descriptions will incorporate my beliefs that such things do exist. For example,
I can refer to the table which I think I am seeing as ‘the one and only physical object
which is causing these (table-like) sense-data in me’, and that description may well be
satisfied by one and only one thing which is what I call a table (PP, p. 47).13 At any rate,
that is what I believe. Naturally, you cannot share exactly this belief, for you cannot
refer directly, as I do, to the sense-data that I now have. But you can refer (in thought)
to your own sense-data, and you may also be having (table-like) sense-data somewhat
similar to mine. So what happens is this. In the public language I say ‘this table is
brown’ and perhaps I gesture at the same time. This is the public way of expressing
what (according to Russell) I actually think, namely ‘the table causing these sense-data is
brown’. You hear what I say, and since you too have learnt to speak our common
language, and to interpret our gestures, you know that I am attempting to refer to a
table that I can see. You find in your own sense-data some suitable table-like sense-
data, and you assume that the table I am talking of is the same as the one that is causing
your sense-data. So I describe the table by its relation to my sense-data, and you
describe it by its relation to yours. But, so long as it is in fact the same table that bears
both of these relations, then our mutual description ‘this table’ has worked perfectly.
On Russell’s account neither of us has actually seen the table, for each of us can see
only our own sense-data, but nevertheless we communicate perfectly well with one

12
He does come to accept something very like this theory in his later view of ‘On Propositions’ [1919b]
whereby a proposition consists of mental images. For an image will (often) mean the object that it is an image
of, i.e. the object that caused it. But by this time his requirement on ‘logical proper names’ has evidently been
relaxed. (I shall discuss this later theory in sections 6–7 of Chapter 12.)
13
The article ‘On the Relation of Universals and Particulars’ [1911/12] suggests an even more non-
committal description: the table is ‘the common cause or origin or whatever vaguer word can be found’ of all
the sense-data that I receive from it (p. 110).
 A C Q UA I N TA N C E 121

another by using the public phrase ‘this table’. While it is true that (for Russell) all
singular thoughts are private, still this is no bar to their ‘communication’ as here
described (cf. ‘On the Nature of Acquaintance’ [1914a: 156]).
This illustrates my remark, towards the end of Chapter 3, that Russell soon lost
interest in the common language that we all speak. It is true that in most of On Denoting
he does seem to be concerned with how best to understand our common language,
and how we need to realize that its surface grammatical form may not be the best guide
to the ‘true’ logical form of the facts that it represents. Subsequent discussions of this
well-known article have almost all understood its topic in this way. But Russell himself
moved in a different direction. His initial idea, that the theory of descriptions best
explains the language that we talk, developed into the idea that it best explains the way
in which we think. It is the nature of thought that is the focus of his interest in KAD
[1910/11] and the corresponding chapter 5 of PP [1912a]. But even this is only a
halfway stage in the development of his final position, as we may see by turning
attention to what he says of our use of ordinary proper names.
Russell is usually understood as claiming that, in the common language that we all
speak, such names are really abbreviations for definite descriptions. Then a standard
objection is that we usually cannot specify any particular description that they can be
thought to abbreviate.14 Apparently, the same objection continues to apply when we
shift attention from the common language to the individual’s supposed ‘language of
thought’. Do we each, in practice, understand an ordinary proper name as short for
some definite description? In some few cases the answer may be ‘yes’. For example, in
my own language of thought the name ‘Homer’ is very strongly associated with the
description ‘the (supposed) author of The Iliad and The Odyssey’. Although I do know
about some other things that are commonly attributed to Homer, for me they have
much less weight than this description. (And, so far as I am concerned, Homer may or
may not have been called ‘Homer’; that is a detail of no importance.) One could find a
few other examples of people from the now distant past, who are remembered only for
some few of their many acts, so that it is not too implausible to equate the name with
the description of those acts. But, assuredly, this is not the usual case.
A moral that one might try to extract from Kripke’s important discussion of 1972 is
that there is a description which nearly all of us do in fact associate with any familiar
name ‘X’, namely ‘the person once called “X”, who stands at the origin of a causal chain
of uses of this name “X” that leads up to my present use of it’.15 On Kripke’s own

14
This objection is central in the well-known attack on Russell’s views by Kripke [1972], but of course
Kripke has many other objections too.
15
Or (better?): ‘ . . . that leads up to its present use in my community’. Of course this description needs to
be filled out, in particular by saying more of the kind of ‘causal chain’ in question. I leave this task to Kripke’s
readers. Kripke does not himself regard his own theory as a special case of a description theory, but there is a
useful discussion in Fumerton [1989] of whether the Russellian is entitled to appropriate it in this way. (If he
does, he should make it clear that the description merely ‘fixes the reference’ of the name, but does not ‘give
its meaning’. I note this distinction later, in section 7.)
122 K N OW L E D G E

account, it should apparently be this one description that is decisive in all cases. But Evans
[1973] has pertinently observed that there are some clear exceptions to this theory, and
drawn the moral that the Kripkean description, though always of importance, can be
overridden by others when there is a clear conflict between them. He therefore restores
the somewhat messy situation that was originally offered as an objection to Russell’s
theory: in practice we have a number of different descriptions, all thought to be relevant,
and they could turn out to be in conflict with one another. But then the theory provides
no way of choosing between them.
This is a fair criticism of the theory of KAD, which says
The thought in the mind of a person using a proper name correctly can generally only be
expressed explicitly if we replace the proper name by a description. (p. 216. Cf. PP, p. 54)

But the ensuing discussion does admit that, with the name of a person about whom
there is a good store of common knowledge (e.g. ‘Bismarck’), it may not be at all
obvious just which description should be supplied in a given case. The point is
admitted in the later lectures on ‘The Philosophy of Logical Atomism’ [1918b],
henceforth PLA, which admit to an almost universal vagueness, not only in what we
say, but also in what we think:
It is a rather curious fact in philosophy that the data which are undeniable to start with are always
rather vague and ambiguous . . . . Everything that you are really sure of, right off, is something
that you do not know the meaning of, and the moment you get a precise statement you will not
be sure whether it is true or false, at least right off. (PLA, p. 179)

Russell’s moral is that it must be the task of philosophy to remove this vagueness in our
ordinary talk and thought. One way of doing so is to replace a familiar name by a
description which is more carefully thought out: it mentions what (for you) are the
important facts about this person (or place, or whatever), and leaves out the rest. Or
perhaps, as is suggested by Searle’s theory of proper names, it grades your various
descriptions, counting some as more central and some as less central.16 But in either
case the idea is to make the thought more precise than it was before. That is certainly in
conformity with one of Russell’s ideals.
Much later still, Russell forgot that he had not always thought in terms of this ideal.
When replying to the criticism of On Denoting that was published by Strawson in his
[1950a] he was moved to say
My theory of descriptions was never intended as an analysis of the state of mind of those who
utter sentences containing descriptions . . . I was concerned to find a more accurate and analysed
thought to replace the somewhat confused thoughts which most people at most times have
in their heads. (1957: 179)

16
Searle [1958].
 A C Q UA I N TA N C E 123

This is not a fair description of what OD has to say on the name ‘Apollo’, nor of
the general position of KAD and chapter 5 of PP. But it does suit the desire to
minimize and if possible eliminate vagueness that we shall find in PLA, to be discussed
in Chapter 14.
Let us now turn from Russell’s views on acquaintance with particulars, as contrasted
with descriptions of particulars, to what he claims to be a similar contrast within our
knowledge of universals.

4 Acquaintance with universals

The theory of Locke and Berkeley and Hume is that ideas may be simple or complex.
Complex ideas may be put together by us, from those that are simpler. (For example,
we can make up the idea of a unicorn by putting together the ideas of a horse and a
horn.) The genuinely simple ideas can only be obtained from experience. Concerning
universals, Russell has exactly the same theory, though he phrases it somewhat
differently. In his language it is this: just as we may know a particular by acquaintance
or by description, so also we may know a universal by acquaintance or by description.
When it is known ‘by description’ that means, on Russell’s account, that we can
explain it in terms of other universals that are known independently, and so eventually
in terms of universals that we know by acquaintance.
As for how it is that we come to know universals by acquaintance, Russell is
distinctly unforthcoming. Here is a representative quotation:
Not only are we aware of particular yellows, but if we have seen a sufficient number of yellows,
and have sufficient intelligence, we are aware of the universal yellow. (KAD, p. 212)17

Thus, to understand the predicate ‘ . . . is yellow’ what is required is experiences of

several instances of the universal yellow and ‘sufficient intelligence’, but Russell never
explains, either here or elsewhere, just how this ‘sufficient intelligence’ is to be used.
One must presume that he is referring to something like the process which Locke
called ‘abstraction’, which is supposed to yield the general (or ‘abstract’) idea of being
yellow from a number of particular ideas of particular instances that are each yellow.
Locke has something to say about how this is done—namely, by ‘leaving out’ what the
various instances do not have in common—but both Berkeley and Hume found his
account incomprehensible. Russell simply gives no account, and we are left to guess at
what he might have envisaged. But there are two comments that one may make at
once. Unlike acquaintance with a particular, we are to understand that acquaintance
with a universal is a process, which may be long-drawn-out, involving acquaintance
with many particulars. And it issues in a lasting state of being acquainted with that
universal, which does not evaporate when the process is completed.

17
Similarly in PP : ‘By seeing many white patches we easily learn to abstract the whiteness which they all
have in common, and in learning to do this we are learning to be acquainted with whiteness’ (p. 101).
124 K N OW L E D G E

The process is supposed to give us understanding of a universal, but there is an

important point, which Russell realized from early days, showing that, whatever exactly
this abstraction is supposed to be, it will not by itself be enough to explain what in fact we
understand. For example, suppose that I have seen many examples of one thing x being
above another thing y, and suppose that I have done the trick of abstracting from these
examples the universal relation expressed by ‘above’. Suppose also that I am acquainted
with the two particulars a and b. This is not yet enough to show that I know the difference
between the list ‘a, b, above’ and the proposition ‘a is above b’. Nor is it enough to show
that I know the difference between this proposition and its converse ‘b is above a’.
A mere acquaintance with the constituents of these propositions cannot provide the
needed information, but of course it is information that we do all possess.
As I have said, Russell appears to see this point—or, at least, the first half of it—from
early on. For example, here is a quotation from PoM [1903]:
The twofold nature of the verb, as actual verb and as verbal noun, may be expressed, if all verbs
are held to be relations, as the difference between a relation in itself and a relation actually
relating. Consider, for example, the proposition “A differs from B.” The constituents of
this proposition, if we analyze it, appear to be only A, difference, B. Yet these constituents,
thus placed side by side, do not reconstitute the proposition. (}54)

He proceeds to generalize the point:

A proposition, in fact, is essentially a unity, and when analysis has destroyed the unity, no
enumeration of constituents will restore the proposition. (Ibid.)

Unfortunately, when he is thinking of acquaintance with universals in KAD and in PP

he appears to forget this point about the ‘unity’ of a proposition, but later in PLA
[1918b] he remembers it again:
To understand a name, you must be acquainted with the particular of which it is a name, and you
must know that it is the name of that particular. You do not, that is to say, have any suggestion of
the form of a proposition, whereas in understanding a predicate you do. To understand ‘red’, for
instance, is to understand what is meant by saying that a thing is red. You have to bring in the
form of a proposition. (PLA, p. 205. My emphasis.)

So the doctrine now is that mere acquaintance is not by itself enough. For we require
not only a knowledge (given by acquaintance) of the constituents of the proposition,
but also a knowledge of its form, i.e. of how those constituents are put together to
make a proposition.
To put this in other words, we now have three distinct requirements for under-
standing a predicate such as ‘red’, which is here assumed (for the sake of example) to be
a simple predicate.18 We require

18
According to Russell’s doctrine, it is only simple predicates that have to be understood by acquaintance.
Cf.: ‘Take the word “red”, for example, and suppose—as one always has to do—that “red” stands for a
 A C Q UA I N TA N C E 125

(i) Experience of red things

(ii) Abstraction of the universal redness
(iii) Knowledge of what it means to say that a thing is red.
As Russell later saw, or half-saw, the only thing that is really required is condition (iii),
for so long as this is satisfied then (i) and (ii) are superfluous. It may fairly be said that, as
things are, conditions (i) and (ii) are often useful steps en route to (iii). For one learns
what ‘ . . . is red’ means by observing cases in which it is true to say that a thing is red and
cases in which it is not true. But (iii) is the goal, and if that can be obtained without
going via (i) and (ii) then no more is needed. The point may also be made in this way.
My present understanding of a word such as ‘red’, which Russell regards as expressing a
universal, depends only on my present state of mind, and is not logically dependent
upon any process of abstraction that I may have been through in order to reach that
state. The situation may be made vivid by recalling Russell’s own observation that the
world may have sprung into existence five minutes ago, with all our memories being in
place, just as they are now. Our understanding of these words for universals could also
be just as it is now, even if there had not actually been any such processes of abstraction
as he was thinking of. For my present understanding of the word ‘red’ is manifested by
my present abilities with this word—e.g. by my being able to recognise that a thing is
red from the way it looks, and by the way I use the word in whole sentences—and it is
logically possible that I should have sprung into existence with these abilities already
formed.
It would appear that Russell later came to realize this. At any rate, in the later book
The Analysis of Mind [1921a] he gives roughly this account of what it is to understand a
word such as ‘red’, and he concludes that
Words of which the logical meaning is universal can therefore be employed correctly without
anything that could be called consciousness of a universal. (p. 228. My emphasis.)

It may be that his point here is more restricted than it needs to be. Given the context,
one could suppose that he means to claim only that one can learn the elementary use of
the word ‘red’, as a reaction to the presence of red things, without any such conscious-
ness. While the same does no doubt apply to more sophisticated uses of the word, in
more complex linguistic contexts, that is not what is here being considered. Even so,
this would still be an admission that acquaintance with universals is not as essential as
KAD had claimed, and we may fairly add that the idea should be abandoned altogeth-
er, even if in fact it is not.19

particular shade of colour. You will pardon that assumption, but one never can get on otherwise’ (PLA,
pp. 193–4).
19
Looking back on things from the later perspective of 1959, Russell says in MPD (p. 169) that he has
always maintained his principle of acquaintance. And although he goes on to add that logical words may be an
exception to it, he says nothing similar about predicates. One can only say that he should have done.
126 K N OW L E D G E

On this topic of ‘acquaintance with universals’ there is a further point worth

considering. Russell held that one could (in principle) be acquainted with a particular
without knowing any truths about that particular. Does he wish to say the same of
acquaintance with a universal? For example, might one be acquainted with the
universal expressed by ‘red’ and with the universal expressed by ‘coloured’ and yet
not know that everything red is coloured? On the face of it, there seems to be no
reason why the supposed process of ‘abstracting’ the two universals ‘red’ and ‘coloured’
should yield this knowledge, for the examples used in each case need not even overlap.
Certainly we need not suppose that every example that was used to teach the word
‘red’ was also used to teach the word ‘coloured’, and even if that did happen to be the
case, could it not be regarded by the learner as a mere accident? But this suggestion
strikes most of us as highly improbable. Could one count as understanding both ‘red’
and ‘coloured’, while still thinking that there might be red things that are not
coloured?20 But reflection on this point reveals a problem. In PP, chapter 10, Russell
claims—reasonably enough—that we do have a priori knowledge of relations between
universals, and that all a priori knowledge is of this sort. A usual view of a priori
knowledge is that it stems simply from our understanding of the concepts involved. If
this is admitted, do we have to say that there is more to understanding a concept than
the account so far has either given or gestured at? I shall leave this question open, for it
soon appears that there are complications. There are all kinds of relations between
concepts that could be claimed to be knowable a priori. Some of them are very
complex, and are not in practice known by most ordinary people who would
ordinarily be deemed to possess the concepts in question. So where can we draw a
limit? There is a temptation to say that knowledge of some simpler a priori connections
is required, but at the same time serious doubt over just which connections these are.

5 Universals known by description

One hopes that Russell also came to modify his original account of how other
universals may be known ‘by description’, but there is very little direct evidence on
this. The doctrine of KAD is that these other universals are understood by being
analysed in terms of more familiar universals, and at this stage Russell understands an
analysis to be an explicit definition. Now it is surely right to say that we come to
understand some (general) words because they are explained to us in terms of other
words that are already understood. But often it is very implausible to suppose that the
explanation takes the form of an explicit definition. A simple example, beloved of
logicians, would be the word ‘ancestor’. In practice we explain the word by saying

20
The passage just quoted from The Analysis of Mind immediately adds: ‘Consciousness [of a universal], in
the only sense in which it can be said to exist, is a matter of reflective judgment consisting in the observation
of similarities and differences’ (p. 228). But Russell is speaking of the similarities and differences between the
instances of one single universal, and not of the relations between several different universals.
 A C Q UA I N TA N C E 127

something like this: ‘one’s ancestors are (1) one’s parents, and (2) their parents in turn
(i.e. one’s grandparents), and (3) the parents of these (i.e. one’s great grandparents), and
so on’. In practice this works, because we do in fact understand the ‘and so on’, but
of course it prevents what we say from qualifying as an explicit definition.21 More
interesting examples of words which we understand, but for which we cannot give
an explicit definition, are words for the ‘theoretical entities’ posited by scientific
theories, such as ‘atom’, ‘electron’, ‘proton’, ‘neutrino’, and so on. Obviously we do
not understand these words by experiencing examples to which they apply, and it must
be broadly right to say that their explanation is given verbally, using other more familiar
words. But in practice this explanation will almost never be an explicit definition, in
the traditional sense.
From a philosophical point of view it is more important to notice a host of more
ordinary words which apparently cannot be treated as Russell desires. Locke and
Berkeley and Hume realized that their account of what we can understand leads to
problems. For example, Locke was puzzled over how we could understand the notion
of substance. Berkeley took this puzzle seriously enough to claim that in fact we cannot
understand the notion of material substance, but then was led to see that there was
equally a problem over immaterial substances, i.e. spirits (or minds). For he admitted
that on his own principles we do not have an idea of a spirit, though he wanted to say
instead that we do have a notion of it, and this at first looks like a mere evasion.
However, if one is kind to Berkeley one says that his ‘notions’ answer to Russell’s
‘knowledge by description’, for we know what spirits are only in the sense that we can
describe what they do, i.e. (in Berkeley’s view) they have ideas and they cause ideas.
But Hume then complained that, according to the theory of meaning which they all
shared, this description is itself not intelligible. For we do not perceive anything that has
ideas, and we therefore cannot understand this ‘having’. More centrally, we do not
perceive one thing causing another, and we cannot explain causing in terms of what we
do perceive. So again, here is what seems to be an ordinary and everyday notion that
the theory cannot account for.
Russell is faced with the same problem, and it is of central importance for his
position in PP and in KAD. For he there wishes to say that we are not acquainted
with such external objects as tables but only with the ‘table-ish’ sense-data that (as we
think) they cause in us. But this means that the word ‘table’ is only understood as short
for ‘whatever external thing it is that causes table-ish sense-data in us’ (PP, p. 47), and
to understand this description we must of course understand the notion of an external
cause. But we are not acquainted with such a thing, for we could be acquainted with a
relation of causing only if that relation had instances which held between our sense-
data, which is not the case here. So Russell is committed to saying that the relation can
be defined in terms of other words which are independently understood, yet in PP and

21
The explicit definition that is due to Frege and to Dedekind (which was given on p. 18 above) is
evidently not the explanation of how we come to understand this word.
128 K N OW L E D G E

in KAD he never pays any attention to the question of whether this can be done, and
if so how. He does consider the notion of cause in another paper written at about the
same date, i.e. ‘On the Notion of Cause’ [1912/13], but what he has to say about it
there only alerts one to a further problem. For he suggests that the notion of a cause is
best understood in terms of a regularity theory: to say that an event a caused an event b
is to say that there are types of event A and B such that a is of type A, and b is of type B,
and every event of type A is always followed by an event of type B (or, anyway, that
this happens nearly always).22 Now whether such an analysis of causation is defensible
may certainly be debated, but I postpone the debate to the discussion of Russell’s
metaphysics in Chapter 14. Instead, let us just note a presupposition of this analysis,
namely that we can understand the universal quantifier. But how is that to be explained
if all our understanding is to be governed by Russell’s principle of acquaintance?

6 Acquaintance with the entities of logic

In earlier days Russell was certainly at a loss over how to explain what he called ‘the
variable’, i.e. the variable when used to express generality. Substantial passages of PoM
[1903] are devoted to this topic (e.g. all of chapter VIII), but without success. This
uncertainty remained when he published OD in 1905, for G. E. Moore in his
comment on OD raised exactly the question that we have just raised:
What I should chiefly like explained is this. You say ‘all the constituents of propositions we
apprehend are entities with which we have immediate acquaintance’. Have we, then, immediate
acquaintance with the variable? And what sort of entity is it? (Moore to Russell, 23 October
1905)

Russell replied two days later

I admit that the question you raise about the variable is puzzling, as are all questions about it. The
view I usually incline to is that we have immediate acquaintance with the variable, but it is not an
entity. Then at other times I think it is an entity, but an indeterminate one . . . . I only profess to
reduce the problem of denoting to the problem of the variable. This latter is horribly difficult,
and there seem equally strong objections to all the views I have been able to think of. (Cited in
Russell [1994: xxxv].)

Did Russell later succeed in resolving these perplexities? Unfortunately one cannot tell.
In KAD he still does not know what to say, for he there mentions ‘the problem of the
nature of the variable, i.e., of the meanings of some, any, and all’, and he comments
‘This is a difficult problem, concerning which I do not intend to say anything at
present’ (p. 230).

22
This criterion paraphrases the result which Russell reaches in his discussion on pp. 182–8, though it
leaves aside his question asking what time interval (if any) is implied by the words ‘followed by’.
 A C Q UA I N TA N C E 129

One would expect our understanding of such words as ‘all’ and ‘some’ to be treated
in much the same way as our understanding of the other logical words such as ‘or’ and
‘if’ and ‘not’. At one stage Russell does anticipate treating both cases together, under
the general heading of our understanding of logical form, which he had planned to
discuss in the proposed book on The Theory of Knowledge that he was working on in
1913.23 Under the same heading he puts our understanding of such atomic forms as
that of the subject-predicate proposition ‘Fa’, the two-place relational proposition
‘aRb’, and so on. In the only part of the book that was ever written we do find a
number of remarks about the forms of atomic propositions, and his theory there is
certainly somewhat surprising. At the opening of the book he has said that the objects
of acquaintance include particulars and universals and in addition what he there calls
‘abstract logical facts’ [1984: 5]. We later learn that he regards a logical form as a ‘fact’
(i.e. the fact that what we would call the form is exemplified, p. 114), so this opening
may be meant to prepare us for the claim that we are acquainted with logical forms.24
At any rate, he certainly does claim in this draft book that we are acquainted with
logical forms (e.g. pp. 99, 101). But he also says that logical objects ‘cannot be regarded
as entities’, and so cannot really be included among the terms of a two-place relation of
acquaintance (p. 97). These claims apparently conflict with one another, and it has
seemed to many commentators that there is no way of finding a consistent view in the
manuscript that we have.25 But since Russell never did publish this part of the work, or
anything resembling it, we may reasonably suppose that he never was satisfied with it.
In its place we find only the very sensible claim of the later PLA [1918b], that in order
to understand a predicate (or relation) ‘you have to bring in the form of a proposition’,
which just means that you have to understand what is meant by saying that a thing has
that predicate (p. 205).
The unfinished manuscript of The Theory of Knowledge breaks off before it reaches
any treatment of the question that began this section: how, consistently with his
principle of abstraction, can Russell explain our understanding of such logical words
as ‘or’ and ‘not’, and again ‘all’ and ‘some’? So we do not know what he might have
thought on this question at the time. For a later view one naturally looks to PLA
[1918b], which gives an account of the truthfunctors that is very much in line with
more modern accounts. It says that such words as ‘or’ and ‘not’ do not fall under the

23
I say more of this project in Chapter 12, section 4.
24
There have been somewhat similar undeveloped hints in earlier writings. Back in OD [1905b] Russell
had said: ‘In perception we have acquaintance with the objects of perception, and in thought we have
acquaintance with objects of a more logical character’ (p. 41). The latter surely include ordinary universals,
but perhaps some genuinely ‘logical objects’ as well? For in PP we hear that the universals known by
acquaintance do include ‘certain abstract logical universals’ (p. 109), though there is no further information
on this point.
25
There is a nice discussion of what The Theory of Knowledge has to say about logical forms in Griffin
[1993]. He concludes: ‘The result is a catastrophic misamalgamation of disparate lines of thought . . . logical
forms are monsters: propositions on the one hand, objects of acquaintance on the other, terms of relations,
but not entities’ (p. 177).
130 K N OW L E D G E

principle of acquaintance, because they do not stand for ‘constituents’ of the proposi-
tions that we express with their help. Lecture III of PLA is quite explicit on this, and it
also gives an account of how we do understand these words, namely by understanding
the truth conditions that they express (pp. 196–7). But his discussion of ‘all’ and ‘some’
in lecture V of PLA, says nothing similar about these words, and we are left to
speculate.26 Much later, in the Introduction that he wrote for the second edition of
PoM in 1937, he seems to treat ‘or’ and ‘all’ as on a par with one another. At least, he
says about ‘or’ and ‘not’ that they do not express constituents of propositions, and adds
not even the most ardent Platonist would suppose that the perfect ‘or’ is laid up in heaven, and
that the ‘or’s here on earth are imperfect copies of the celestial archetype. (p. ix)

And he says about ‘all’

It would seem absurd to say that generality is a constituent of a general proposition. (p. xi)

It would follow that the principle of acquaintance need not apply to it. But these
remarks still contain no hint of how we do understand this notion.27

7 Appraisal
Well, whatever Russell did think, what should he have thought? I take it that the
answer is obvious: he should have dropped the notion of acquaintance altogether, for it
is only very seldom that it is helpful in explaining how we understand what our words
mean. The whole idea begins from the thought that a word has meaning by standing
for an object, so to understand the word is to know which object that is, which in turn
is a matter of being able to point to or to focus upon the object in question.28 But we
have seen that (in PLA, p. 205) Russell himself came to see that this model will not do,
even for those simple predicates of first level that are traditionally regarded as standing
for universals. Many people (e.g. Quine) would say that such words do not actually
stand for anything. But we have observed that, even if they do, still that is not all that
there is to their meaning. A very similar point may be made about the first-level
quantifiers. Even if we do wish to supply objects for which these quantifiers stand—e.g.
as, in PoM, ‘all A’s’ is thought of as standing for all the A’s—still there is more to their

26
PLA is clear that a general proposition is expressed by applying a quantifier to a propositional function,
and that there must be general facts corresponding to true general propositions. But Russell also says ‘I do not
profess to know what the right analysis of general facts is’ (pp. 236–7). Presumably he did not know how to
answer our question. But in section 4 of Chapter 12 I shall speculate that he no longer thinks of
‘acquaintance’ as providing an answer.
27
In 1937 he says of all the logical constants that they ‘must be treated as part of the language, not as part of
what the language speaks about’ (PoM2, p. xi). This perhaps looks forward to the account that he will give in
the Inquiry into Meaning and Truth [1940], which clearly falls outside our period. Contrast what he says in PLA
on the existence of negative and general facts, which I discuss in section 5 of Chapter 14.
28
As Russell knew, this is a very ancient theory, for it is what underlies Plato’s theory of forms. (Cf. PP,
ch. 9.)
 A C Q UA I N TA N C E 131

meaning than this. The most obvious point is that supplying such objects will not
explain how different occurrences of the quantifiers can have different scopes. Indeed,
differences of scope affect many other features of our ordinary language, to which
Russell is paying no attention, e.g. our auxiliaries of tense, such as ‘it will be’, or our
deontic auxiliaries, such as ‘it ought to be’, or our modal auxiliaries ‘it could be’.
In none of these cases is it at all plausible to say that you need to find an object which
the word stands for, and which must be grasped if the word is to be understood. And
in all cases you should add that even if there is such an object it cannot be all that the
word has by way of meaning.
Let us finally return to where our discussion began, i.e. to the topic of referring to a
particular object in order to go on to say something about it. Is the idea of acquaintance
helpful for this purpose? In order to be at all realistic we must begin by making a radical
departure from Russell’s account: we almost never attempt to refer to our own sense-
data in the direct way that Russell envisages. This is partly because we almost never
want to talk about these sense-data anyway, and partly because—if we do—then we
can introduce them to others only by using a description which those others under-
stand. A Russellian direct reference would be relevant only when one is talking to
oneself about one’s present experience, and that is hardly the usual case. Let us think,
therefore, of applying his doctrine to the ordinary public objects that we commonly
take ourselves to perceive, e.g. chairs and tables.
The first thing that strikes one is that in almost all such cases (i) it will be possible to
refer—or, at least, to think of oneself as referring—to an object which then turns out
not to exist; and (ii) it will be possible for what appear to be two distinct references to
turn out to have been references to the same object, and conversely for what appears to
be the same reference made on two occasions to turn out to be two distinct references,
each to a different object. There are perhaps a few exceptions, such as the use of the
word ‘I’, which I shall come to later. But in the usual case both the existence and the
identity of what is thus referred to may prove to be doubtful. Russell was led to think
that this must be impossible partly because he drew too stark a contrast between
naming and describing. Whereas a description clearly has a meaning which one may
grasp without knowing which object (if any) it describes, he thought of a name as
having no meaning other than the object which it names. Consequently, if ‘a’ and ‘b’
are genuine names, then ‘a exists’ must be either true or meaningless, and ‘a 6¼ b’ must
be either true or synonymous with the contradiction ‘a 6¼ a’. But this is to ignore a
compromise which is quite nicely brought out in Kaplan’s discussion.29
On Kaplan’s account a ‘direct reference’ is one that simply has the function of
introducing a particular object as the one that one is speaking of. In Russell’s terms, it
introduces that object as a ‘constituent’ of the proposition expressed, and it introduces
no other constituent. But this does not prevent the referring device from making use of

29
Kaplan, ‘Demonstratives’ [1989].
132 K N OW L E D G E

certain properties of the object in order to pick it out. A demonstrative, for example,
will standardly make use of the location or direction of the object, and of how it
appears (to the speaker and to the audience) from that location. This is how it fixes
attention on the object in question, but the way in which it does so forms no part of the
proposition that it is used to express. To put what is essentially the same idea in the
different words used by Kripke [1972], the referring expression may be explained by
being or being associated with a descriptive phrase, but in that case the description will
merely ‘fix the reference’ of that referring expression, and will not ‘give its meaning’.
So the referring expression will be what he calls a ‘rigid designator’, which always
designates the same object, whatever ‘possible world’ we are considering, when asking
whether our proposition would be true in it.30 Kripke applies this account to proper
names, which may be introduced by a baptism, which involves demonstrating the
object in question, or may be introduced by a reference-fixing description. But in
either case the method of introducing the name plays no part in determining the
truthvalue of the propositions that it is used to express. Kaplan applies the same idea to
the demonstrative ‘this’. In any context of utterance, the demonstrative will be
accompanied by an act of demonstration, which will present the object referred to in
a certain fashion (e.g. by pointing to it). This demonstration has what Frege would
apparently recognize as a ‘sense’, at least when he glosses his word ‘sense’ as ‘mode of
presentation’. The sense does ‘fix the reference’, but that is all it does. It does not
contribute to what Kaplan calls the content of the proposition expressed, i.e. to what has
to be taken into account when assessing the truthvalue of that proposition in this or that
possible situation. But he adds that it does make a difference to our understanding of
what is said.
Thus suppose that ‘a’ is a demonstrative, introduced via a particular sense (i.e. mode
of presentation). Then we can well understand how it may turn out, to the surprise of
the speaker, that that particular sense is hallucinatory, and fails to present any such
object as he is supposing. (For example, Macbeth would have been in this situation if
he had believed ‘This is a dagger that I see before me’.) Equally, if ‘a’ and ‘b’ are both
demonstratives, but associated with different senses, we can see how people could
believe ‘a 6¼ b’ without taking themselves to believe a contradiction. (As Simon Black-
burn once said in conversation: the conjurer may explain his sleight of hand by saying
that the audience ‘do not recognize that this is the same as this’, if his first ‘this’ is
accompanied by an action that shows an object one way up, and his second by an
action that shows the same object the other way up.) In this second case it seems best to
say that there is a proposition that the people believe, and it is in fact a contradiction,
though they do not realize this. However in the first case, where what is meant as a

30
Kaplan and Kripke give different accounts of how the designator is to be understood when it is used in
making statements about a situation in which its object does not exist. That is why Kaplan avoids using the
Kripkean terminology, and speaks of ‘direct reference’ rather than ‘rigid designation’. But from our point of
view the discrepancy is only marginal, and may be ignored.
 A C Q UA I N TA N C E 133

reference-fixing device fails to fulfil its purpose, and no reference is actually deter-
mined, it may seem better to say that no proposition is actually expressed. Nevertheless
the fact that demonstrations do have this Frege-like sense shows how such mistaken
beliefs are entirely possible, and this should dispel Russell’s worries.
Two further points are worth making here. First, a demonstrative will often be
accompanied by a common noun giving the kind of object that is being referred to, as
in ‘that man’ or ‘this cherry’ or ‘that boat over there’, and this descriptive element will
guide the audience to the object which the speaker intended. His description may not
be quite correct. For example an utterance of ‘that man’ may be understood as referring
when the object ostended is not actually a man, but—say—a waxwork, or a picture, or
perhaps a gorilla. But it must be something which (in the circumstances) one could take
to be a man, and there are limits on possible mis-takings. One cannot, for example,
mistake a number for a man. Again, one cannot actually be referring to a colour if one
takes it to be a shape, or a weight, or a temperature, or something equally inappropri-
ate. We may generalize and say that if a demonstrative reference is to succeed then both
the speaker and his audience must be aware of the general type or category of the
object referred to. To this extent, I would say that the ability to refer to a particular
object is not wholly independent of all knowledge of truths about it, and Russell’s
account of ‘knowledge by acquaintance’ needs to be modified accordingly. But the
modification is comparatively slight.
A second point worth making is that a demonstrative reference will standardly be
employed only where both the speaker and the audience are ‘acquainted’ with the
object in question. There may be exceptions where the speaker is assuming an
acquaintance in his audience, though he lacks it himself, as when he points and says
‘that so-and-so’ but is not himself looking in the direction in question. There may also
be exceptions the other way round. But standardly the acquaintance will be present on
both sides, and is why this way of referring to an object is so often successful. It is also a
very basic mechanism for the language learner. But it is clear that acquaintance and
demonstration is not the only way of securing what Kaplan calls a ‘direct reference’, for
we have already noted that an explicit definite description may also be used and
understood in this reference-fixing way. The object thus referred to need not be an
object of acquaintance for anyone.31
To sum up: it was not a good idea on Russell’s part to suppose that all parts of any
proposition that I understand must be things that I am acquainted with. This evidently
does not apply to the logical notions involved, and I have argued that it does not apply

31
If there is doubt over whether a description ‘the so-and-so’ should be understood as doing no more
than fixing a reference, then it may be explicitly ‘rigidified’ by rephrasing it as ‘the actual so-and-so’ or (in
Kaplan’s language) ‘dthat so-and-so’. Whether these rephrasings do empty the description of what Kaplan calls
‘content’ may be debated, but they do ensure that any content that remains (over and above the object that
they pick out) has no effect on the assessment of truthvalues.
134 K N OW L E D G E

to the predicates either. Nor does it apply to all ways of referring to particular objects,
as we have seen, but in this case it is a standard feature of one of the most basic methods
of reference.
Let us now turn from what Russell (somewhat oddly) called ‘knowledge of objects’,
and move on to what he (like everyone else) called ‘knowledge of truths’.
 8
Knowledge and its Foundations

1 Why knowledge needs foundations

Russell inherits and adopts a traditional approach to knowledge of truths which in

fact goes back to Plato (in his Meno, 98a), and which has appealed to many others
since: the basic idea is that knowledge is a matter of justified true belief. Russell’s
short introductory book, The Problems of Philosophy [PP, 1912a] is one of the best
expositions of this approach to the topic of knowledge.1 The approach soon forces
one into a distinction between what we may call ‘direct’ or ‘immediate’ knowledge
on the one hand, and ‘derivative’ knowledge on the other. This is because the
justification for a belief will itself cite further beliefs, and if they are to show that the
original belief is justified then they themselves must be justified in turn, and so on.
But in practice the chain of justifications will come to a halt somewhere, and
indeed theory assures us that it must, if we assume (a) that ‘proper’ justifications
cannot go round in a circle, and (b) that we have only finitely many beliefs that
could play a role in such a chain. So justifications will come to an end in beliefs
which are not further justified, but which form the ‘foundation’ for all beliefs that
are justified. It is these foundational beliefs that must be ‘directly’ or ‘immediately’
known. (In order to maintain the original connection between knowledge and
justification they may be called ‘self-justifying’, but this is a merely verbal manoeu-
vre of no real significance.)
There are two ways of trying to avoid this commitment to an unjustified ‘founda-
tion’ of all beliefs that count as knowledge. One is to argue that in some suitable cases
justifications may be allowed to ‘go round in a circle’. For example, the belief P may be
justified from the beliefs Q and R ; the belief Q may be further justified from the beliefs
S and T ; and then the belief S may be justified from the beliefs U and P . Thus P is itself
used in justifying other beliefs, which in turn are used to justify P . If this kind of thing is
permitted, then of course we need not accept the claim that justifications cannot go on
for ever. Moreover, it is an approach which accords well enough with our actual

1
In this chapter quotations from Russell are all from this work, unless otherwise stated.
136 K N OW L E D G E

practice in offering justifications, and it has seriously been defended as a philosophically

acceptable procedure.2
A different way of escape is to stress that what is required is the possibility of a
justification, and that it does not matter if for the time being we cannot actually
produce one. That is, if I truly believe that P, and if there is a true proposition Q
which would be accepted as justifying the belief that P if anyone were to propose it in
that role, then we shall say that (in the Platonic sense) there is a justification for the belief
that P. I therefore count as knowing that P, even though I cannot myself produce the
justification.3 On this view it need not matter if the actual beliefs that we can produce
in justification are only finite in number, for there is always the possibility of another.
So, to whatever length a chain of justifications has been extended so far, we are never
entitled to conclude that it could not be extended by yet a further step, if only we were
sufficiently ingenious. In that case we need not admit that there is anything that is both
known to be true and incapable of any further justification.
One could pursue either of these escape routes in more detail, but I shall not do so,
for Russell himself does not. He accepts the point made initially that, on the approach
to knowledge which he advocates, there will be a set of beliefs that are ‘immediately
known’, and that provide the ‘foundation’ for all other knowledge. Much of PP is then
devoted to exploring which beliefs these are, and I shall shortly outline his conclusions.
But first the general approach needs more defence, for there are many who think that it
received its deathblow from Edward Gettier’s well-known article of 1963, ‘Is justified
true belief knowledge?’
Gettier’s argument is simple. Suppose that I truly believe that P, and suppose that my
justification for this belief is that P is a logical consequence of Q and I believe that Q.
Moreover, suppose that I have what would ordinarily pass as excellent evidence for Q,
so that my belief in Q is appropriately justified. Then since this belief is justified so also
is its logical consequence P, and by hypothesis P is true. Hence, by the definition of
knowledge as justified true belief, I should know that P. But, as Gettier observes, the
conditions just given do not ensure that Q is true, since it may be supported by all the
evidence but still not actually true. In that case everyone will agree that in fact I do not
know that P, since I believe it only because it follows from another proposition that
I believe, and that other proposition is in fact false.
There is an obvious way of meeting this simple and well-known counter-example.
If we have begun from the idea that what is known must be both true and justified,
then we shall naturally want the beliefs cited in justification to be themselves known,

2
The idea of allowing circular claims of justification soon becomes what is called ‘the coherence theory of
justification’. For a defence of this theory, see e.g. BonJour [1985], especially chapters 5–7.
3
Russell himself allows for this when he says that an unconscious move from one belief to another may be
called a ‘psychological inference’, and ‘we shall admit such psychological inference as a means of obtaining
derivative knowledge, provided there is a discoverable logical inference which runs parallel to the psycholog-
ical inference’ (p. 134, my emphasis). But he does not think of this as allowing us to claim that every chain of
justifications could be extended.
 K N OW L E D G E A N D I T S F O U N DAT I O N S 137

and hence both true and justified (or: both true and foundational). But the point of
Gettier’s example is that the justifying belief Q is in this case not true, so we avoid the
example simply by stipulating that it must be. That is: all the beliefs which enter into
the chain of justifications that leads to P must themselves be both true and either
justified or foundational. This is clearly the solution which Russell himself anticipates
when discussing an example very similar to Gettier’s. Suppose that someone has the
true belief (P) that the late Prime Minister’s name began with a B, but he believes it
because he believes (Q) that the late Prime Minister was Mr. Balfour, whereas in fact
the late Prime Minister (at the time) was Sir Henry Campbell Bannerman (p. 131).
Russell comments
The man who believes that Mr. Balfour was the late Prime Minister may proceed to draw valid
deductions from the true premise that the late Prime Minister’s name began with a B, but he
cannot be said to know the conclusions reached by these deductions. Thus we shall have to
amend our definition by saying that knowledge is what is validly deduced from known premisses.
This, however, is a circular definition: it assumes that we already know what is meant by ‘known
premisses’. It can, therefore, at best define one sort of knowledge, the sort we call derivative, as
opposed to intuitive knowledge [i.e. immediate knowledge]. We may say: ‘Derivative knowledge
is what is validly deduced from premisses known intuitively’. In this statement there is no formal
defect, but it leaves the definition of intuitive knowledge still to seek’ (pp. 132–3).

We may enter a couple of quibbles. (a) In order for the conclusion of a derivation to
count as known, it must be believed because it can be established by the derivation in
question, and not for some quite extraneous reason. (b) Russell is simplifying when he
identifies the kind of derivation in question with a ‘valid deduction’, which suggests
that the conclusion should follow from the premisses simply as a matter of logic. For in
fact he wishes also to include reasoning which shows that, given the premisses, the
conclusion must at least be probable, since that is what happens when our justification
relies on induction. However, these are minor points. The important moral is that
Russell here shows why he would not be at all perturbed by what we have come to call
a Gettier example, and it would seem that he is right on this point. If we accept
Russell’s view that knowledge is a body of true beliefs, built up step by step from a
foundation which itself is immediately known, then one can reasonably require that
no belief that is not itself known should play any part in this building up. The reason
why the ‘Gettier example’ still survives in contemporary discussions, and continues to
receive some detailed attention, is that many people nowadays do not accept Russell’s
foundational view.
One reason why one may wish to reject his two-tier system, distinguishing imme-
diate knowledge and derivative knowledge, is that many very ordinary examples of
knowledge do not seem to involve anything that deserves to be called a justification.
For example, I know that I am now sitting in a chair at my desk. On Russell’s account
I know this because (a) I am directly aware that I am perceiving sense-data that are
appropriate to my being in that condition—e.g. kinaesthetic sensations appropriate
138 K N OW L E D G E

to my being in a sitting position, visual and tactual sensations appropriate to there being
a desk and a chair suitably placed, and so on—and (b) because I infer from these beliefs
about my sense-data, together with other beliefs about the normality of the present
situation, to the conclusion that I am indeed so sitting. But, as Russell himself admits, in
the usual case I am not paying any attention to my sense-data, nor consciously inferring
anything from them. At best we have what he calls a ‘psychological inference’, which
leads me unconsciously from my present sense-data to my explicit belief, and which
I could turn into a proper and conscious inference if I wished to.4 But is it really true
that I, or someone who can reason better than I can, must be able to concoct such an
inference if I am to count as knowing? Many have felt that this claim is implausible, and
so have sought for an alternative account of knowledge which avoids it.
One such theory is what is called the causal theory of knowledge, which replaces
Russell’s talk of justifications with considerations about causes.5 Very crudely, the idea
is that my true belief that P counts as knowledge so long as that belief is caused (at least
partly) by the very fact that P. This appears to fit such simple cases as just mentioned;
that is, my belief that I am sitting in a chair is on this occasion caused by the fact that
I am sitting in a chair, and that is why it counts as knowledge. But clearly several
qualifications will be needed if we are to adapt this theory to more complicated cases.
Another approach, which apparently coincides with the causal theory in nice simple
cases, is known as the reliability theory. Its main idea is that a true belief will count as
knowledge if the belief was reached by a general method of forming beliefs that is in
the circumstances a reliable method of forming beliefs. This is explained by different
authors in different ways, but an attractive approach is that offered by Nozick [1981:
ch. 3]: a reliable method is one that ‘tracks the truth’ in a sense that is elucidated by
counter-factual conditionals: the method would not give me the belief that P if it were
not true that P, but would still give me the belief that P if (in relevantly similar
circumstances) it is still true that P. This theory as stated so far allows one to say that
one method of forming beliefs, which is often reliable, is by justifying them from other
beliefs that are already known. But it is not the only method, and evidently it does not
apply to those beliefs which Russell regards as known without justification.
Both of the theories just mentioned were introduced much later than Russell’s own
thinking on this topic, so of course he does not ever consider them. Nor, therefore,
shall I, except to make two suggestions which one might think worth further investi-
gation. When Russell is thinking about knowledge he is always concerned with the
problem of scepticism. Now the sceptic is almost always presented as asking ‘how do
you know?’, and suggesting that you do not really know. But although scepticism is
presented in this way, it is always in fact understood as asking for suitable justifications,
and as answered by and only by providing such justifications. That is, the sceptic
assumes that knowledge is a matter of having justifications, and that is his interest in the

4
See PP, pp. 133–4, which is cited in note 3 above.
5
The theory is mainly due to Harman. See his [1973: ch. 9], and the discussion in Goldman [1967].
 K N OW L E D G E A N D I T S F O U N DAT I O N S 139

subject. For that reason, it is always Russell’s interest too. By contrast, proponents of
the causal theory, or the reliability theory, or of other modern theories of knowledge
are frankly not much concerned with scepticism. Their aim is to tell us about the
ordinary notion of knowledge, as used by ordinary people in ordinary circumstances,
where scepticism is not a serious issue. So, on the whole, they ignore the sceptic.6
Russell may therefore say that they have simply changed the subject, from the one
that he was always concerned with to a different one which is more concerned with
ordinary ways of talking and thinking. He need have no objection if his conception of
knowledge, as depending on justification, is characterized as a somewhat special and
philosophical conception of what knowledge is, or ought to be. For he can still claim
that this is the conception that is philosophically important.
A second consideration is that Russell himself admits that what he thinks of as
‘immediate knowledge’ has no defence against scepticism. One might try to argue that
the postulated immediate knowledge has an indirect justification: if you accept this as
knowledge then you can explain lots of other cases of knowledge which we are
independently agreed on but cannot explain in any other way.7 However Russell
never himself proposes any such indirect justification, and he simply leaves us with the
question: why should we accept as knowledge all these propositions which he himself
describes as ‘immediate knowledge’? For what is there about them that entitles them
to this honorific description? Again, Russell himself seems to have nothing to say.
But perhaps other accounts of knowledge, e.g. in terms of causes or in terms of reliable
methods, can here step in. Perhaps it is because Russell’s ‘immediate knowledge’ does
count as knowledge by these other criteria that he is entitled to make use of it as
he does?
I leave that question open, and instead turn to a survey of what Russell is counting as
immediate knowledge. It divides into knowledge of particular facts, concerning the
world that we experience, and knowledge of general principles, which are needed to
allow us to infer from these initial data to all our other knowledge about the world
that we live in. I take the general principles first, because their treatment can only be
rather brief. Russell does not supply material for more.

2 General principles as foundations

It is hardly surprising that, among the general principles that allow us to make
inferences from our initial data, the first to be mentioned in chapter 7 are the principles
of elementary logic. Since PP is intended as an introductory work, Russell here gives
no detailed inventory of these, but the indications are that he is thinking mainly of

6
Nozick himself is an exception. Section II of his chapter (op. cit.) does directly address the topic of
scepticism, but it is fair to say that his (unexpected) proposals have not found much support.
7
Compare Russell’s way of justifying the axioms that he provides as a ‘foundation’ for the deduction of
logic and mathematics, as given on p. 96 above.
140 K N OW L E D G E

what we now call ‘elementary’ logic, i.e. the logic of the truth-functors and the first-
level quantifiers. Logical principles of higher levels receive no explicit mention. A slight
hint that for the time being they are being set aside is that Russell here gives a separate
treatment to the truths of elementary mathematics, and does not suggest that they
might be reduced to truths of logic. The separation is, of course, historically appropri-
ate, for there have been philosophers who have argued that mathematics is learnt and
justified empirically, whereas this view of logic has never been found appealing.
Russell, of course, rejects the view in both cases, and claims that logic and mathematics
are each knowable a priori, and hold in all possible worlds (p. 78). But he here gives no
hint of how he actually conceives the relationship between them.
He recognizes that deductions which can be certified by logic and mathematics will
not be enough, for we also accept inferences which rely upon induction rather than
deduction as justifications which establish knowledge. Consequently he gives us quite a
long discussion of inductive reasoning in chapter 6 of PP, but it is not worth our
detailed attention, for it is based throughout on a simplifying assumption which he later
saw to be mistaken.8 The assumption is that it should be possible to formulate a general
principle of induction which holds for all features of objects whatever. For example,
his first formulation of what he hopes will be a suitable principle is this
When a thing of a certain sort A has been found to be associated with a thing of a certain other
sort B, and has never been found dissociated from a thing of the sort B, the greater the number of
cases in which A and B have been associated, the greater is the probability that they will be
associated in a fresh case in which one of them is known to be present. (p. 66)

The hope is that this principle will hold whatever we take as our sorts A and B. But that
hope cannot be satisfied, as is shown by what is nowadays called ‘the new riddle of
induction’, or ‘Goodman’s paradox’, since it was Goodman’s version (in his [1955: part
III]) that became famous. That version may be paraphrased in this way. Let to be some
time in the future, and let us call a thing ‘grue’ at any time t if and only if either
t is before to, and the thing is green at t, or t is at or after to, and the thing is blue at t.
Now since to is in the future, all emeralds have been both green and grue at all times t so
far, but clearly they cannot all continue to be both green and grue at times after to. We
do in fact think that it is much more probable that they will stay green than it is that
they will stay grue, but that is something that Russell’s principle (above) cannot
account for, for it treats ‘green’ and ‘grue’ equally: with ‘emeralds’ for A, either of
them may be taken as the value of B, with the same result in each case.
Let us put the point more generally. The things of sort A may divide into those that
are also of sort B1 and those that are also of sort B2. But it may be that all those
examined so far have been both B1 and B2, whereas fresh cases cannot continue to
be both. So a principle of the kind that Russell was seeking for cannot distinguish the

8
Looking back on PP from the perspective of many years later he himself said: ‘What I said about
induction in this little book now seems to me very crude’ (MPD, p. 102).
 K N OW L E D G E A N D I T S F O U N DAT I O N S 141

probability of a fresh case being B1 from the probability of a fresh case being B2, even
though we know that it cannot be both. For each fits the evidence equally well.
It is remarkable that in an article ‘On the Notion of Cause’ [1912/13], written very
soon after his discussion of induction in PP, Russell came to see this problem. For he
there says
Given some formula which fits the facts hitherto—say the law of gravitation—there will be an
infinite number of other formulae, not empirically distinguishable from it in the past, but
diverging from it more and more in the future . . . We cannot say that every law which has held
hitherto must hold in the future, because past facts which obey one law will also obey others,
hitherto indistinguishable but diverging in future. Hence there must, at every moment, be laws
hitherto unbroken which are now broken for the first time. (p. 204)

This contains exactly the observation which is now called ‘the new riddle of induc-
tion’: we cannot suppose that all the regularities that have held so far will continue to
hold, since they will come into conflict with one another. So how do we choose which
of them should be believed to continue unbroken? Here in [1912/13] Russell does not
perhaps see quite how awkward this problem is, for he seems to suppose that it is
adequately answered by saying that (for pragmatic reasons) we choose to project into
the future the simplest of those regularities that we have found so far, and he pays no
further attention to the question of what should count as simplicity in such a problem.
Others have taken the same line, e.g. by trying to argue that ‘green’ is a simpler
predicate than is ‘grue’. But it is over-simple to suppose that this one criterion will give
us the desired answer in all cases. To see this, we have only to revert to Russell’s own
example in PP. Is it simpler to suppose that the man who keeps the chicken will
continue to feed it for all of its natural life, or to suppose that he will kill it when it
becomes large enough to eat? Surely we feel that here one needs to find out more
about the man’s actual motives, and not to speculate on the ‘simplicity’ of various
possible generalizations.
Anyway, let us just set aside Russell’s own discussion of induction in PP, because it is
open to objection from what he himself saw as a problem very soon afterwards, namely
an objection based on what is called ‘the new riddle of induction’.9 But, at the same
time as setting this supposed principle aside, one should also notice a serious gap in
Russell’s account of what is needed by our scientific investigations. We make use of
a principle that is usually called ‘inference to the best explanation’, and is roughly this.
If a certain hypothesis would explain some suitable range of scientific phenomena, and
if we cannot think of any other hypothesis that would explain them just as well or even
better, then we should take that hypothesis to be true. Russell has himself mentioned

9
Russell’s rather brief discussions of induction in other writings of our period (e.g. [1914c: 224–6])
apparently revert to the principle stated in PP, while admitting that it cannot be quite right as it is there stated.
There is a much longer and more thorough discussion of the problem in the later work Human Knowledge: Its
Scope and Limits [1948], but I shall regard this as falling outside our period.
142 K N OW L E D G E

this idea in his article ‘The Regressive Method . . . ’ [1907], which he put forward as a
justification for adopting the axiom of reducibility in his logic (p. 96 above), and in
which he observed that this method is commonly used in science, but may reasonably
be employed in logic too. It is a strange oversight that PP does not mention it. Perhaps
Russell has come to believe that it should be ignored because its predilection for
the ‘best’ explanation is like our predilection for the ‘simplest’ regularity that has
held so far; it has a merely pragmatic justification, and both are couched in terminology
that is deliberately vague and can be bent in various directions (i.e. ‘simplest’ and ‘best’).
But we do need some way of choosing between rival hypotheses, and this is as
sensible as any. Perhaps that is not quite the way that Russell himself saw these basic
principles of all our knowledge, i.e. these principles that he regards as known immedi-
ately and a priori, and as forming part of the ‘foundation’ for all other knowledge.
But that may be because he has failed to distinguish adequately between these examples
of advice on how to proceed when trying to understand the contingencies of
this world that we inhabit, and the genuine necessities that are reflected in his other
principles.
Of these other principles we have so far noted the laws of logic and of mathematics.
Russell adds to these in chapters 9 and 10 of PP, when elaborating the point that we are
acquainted with universals, for he evidently thinks of this as providing us with a priori
knowledge of suitable properties of universals and of relations between them. For
example, he says
It must be taken as a fact, discovered by reflecting on our knowledge, that we have the power
of sometimes perceiving such relations between universals [as that 2+2=4], and therefore
of sometimes knowing general a priori propositions such as those of arithmetic and logic (p. 105)

Another of his examples is our knowledge that two shades of green resemble one
another more than they resemble a shade of red (cf. also Our Knowledge of the External
World, p. 79). One must assume that he means his discussion here to include such
knowledge as that if x is taller than y, and y taller than z, then x is taller than z. It is not
clear how acquaintance with the universal tallness is supposed to provide us with this
information, but Russell tells us no more on how it is done. He also pays no attention
to cases of this kind where the question of a priori knowledge is controversial.
(For example, do we know a priori that if x occurs before y then y does not also
occur before x?10) In any case, he clearly means to include some such truths amongst
the first principles of knowledge, since we do know them and they cannot be deduced
from anything else that we know. He also thinks that the same may apply to some
ethical principles, e.g. ‘that happiness is more desirable than misery, knowledge than
ignorance, goodwill than hatred, and so on’ (p. 76). Whether these are also to be

10
This is a way of saying that time does not ‘go round in a circle’, and hence that time travel does not
occur. Is this known a priori? Indeed, is it known at all?
 K N OW L E D G E A N D I T S F O U N DAT I O N S 143

discovered by reflecting on what can be learnt from acquaintance with the universals in
question is something that is left undiscussed. But he does admit that, though they may
be counted as ‘self-evident’, still they have less ‘self-evidence’ than do some of the other
general principles that have been mentioned (pp. 112, 138–9).
This completes Russell’s list of the general principles that lie at the basis of all our
knowledge. They include principles that would normally be accepted as known a
priori, i.e. the laws of logic and of mathematics, and the results of conceptual analysis.
But they also include principles that are thought to legitimize our scientific investiga-
tions, which are not normally counted as known a priori, and it may surely be asked
whether they should be counted as known in any way. However it is true that they
do have a foundational role, not only in our scientific theorizing but also in the beliefs
that we take to be ‘common sense’.
Let us now turn to Russell’s other kind of basic propositions, those concerning the
particular facts of experience. These are our beliefs about our own sense-data.

3 Particular facts as foundations

Let us begin with some elucidations. These elucidations will frequently cite the article
‘The Relation between Sense-Data and Physics’ [1914b], henceforth RSP, and the
book Our Knowledge of the External World [1914c], henceforth KEW. These works,
written shortly after PP, show Russell beginning to be more cautious about our
knowledge of our own sense-data. As the word implies, a sense-datum is supposed
to be something given to the mind. At times Russell introduces the restriction that
a sense-datum is only a part of what is given, i.e. a part that is sufficiently small for one to
concentrate one’s attention on all of it:
When I speak of a ‘sense-datum’ I do not mean the whole of what is given in sense at one time.
I mean rather such part of the whole as might be singled out by attention: particular patches of
colour, particular noises, and so on. (RSP, p. 147)

This restriction seldom plays any important part in his discussion. A more relevant
point, which deserves more attention than it receives in PP, is the difficulty of deciding
just what is to count as given, in contrast to what the receiving mind adds to it by way of
interpretation.
A nice example of this problem is given in KEW, where he claims that the senses
give us immediate knowledge, not inferred from other pre-existing beliefs, but at once
adds this comment:
Psychologists have made us aware that what is actually given in sense is much less than most
people would naturally suppose, and that much of what at first sight seems to be given is really
inferred. (KEW, p. 75)11

11
A similar comment may be found in ‘The Ultimate Constituents of Matter’ [1915: 126–7].
144 K N OW L E D G E

As an example he offers this point:

For instance, we unconsciously infer the ‘real’ size and shape of a visible object from its apparent
size and shape, according to its distance and our point of view. (Ibid.)

Elaborating this example, one expects him to say that, at least in monocular vision, the
mechanism involved ensures that distance cannot actually be given to us, but in fact he
says the opposite:
The immediate objects of sight, the coloured surfaces which make up the visible world, are
spatially external in the natural meaning of this phrase. We feel them to be ‘there’ as opposed to
‘here’; without making any assumption . . . we can more or less estimate the distance of a
coloured surface. It seems probable that distances, provided they are not too great, are actually
given more or less roughly in sight. (KEW, p. 80)

But how could they be? Must it not really be a matter of the mind, habituated by past
experience, imposing an interpretation on the data actually received? (Notice that we
‘see’ distance in just the same way even when we know that what we are looking at is a
picture painted on a flat surface.) Of course, the interpretation is unconscious, and
perhaps cannot be brought to consciousness, but we know from another example that
this is a possibility. We cannot make ourselves conscious of our ‘blind spot’, i.e. an area
in the field of vision for which we receive no data from outside, but the mechanism of
vision assures us that there must be such an area. There is a problem, then, over
separating what really is given from what merely presents itself as given, but is in fact
due to our own past experience or instinct.
As these examples show, this problem is serious, and it cannot be solved ‘from
within’, in the way that Russell seems to envisage, simply by concentrating attention
on the nature of the apparent sense-datum. Later, he came to realize this. In The
Analysis of Matter [1927a] he is led to say, about our apparent sense-data:
The element of interpretation can only be eliminated by an elaborate theory, so that what
remains—the hypothetical bare “sensation”—is hardly to be called a “datum,” since it is an
inference from what actually occurs (p. 189).

His aim had been to subtract from what we initially take to be the datum all that we can
become conscious of as added by us, due to our previous experience. (For example, the
many expectations that we have as to what further experiences will be available.) But
he now accepts that there may also be features of the supposed ‘datum’ that are added
by our own minds, but that we cannot become aware of as so added.12 That is why the
‘bare sensation’ must now be regarded as something ‘hypothetical’, and reached only as
a result of theory. But presumably it is our beliefs about this hypothetical ‘pure datum’

12
There is a very similar discussion on pp. 211–17 of his book An Outline of Philosophy [1927b], written at
about the same time. This latter discussion gives the example of the blind spot (p. 213). The difficulty has
apparently been recognized earlier in The Analysis of Mind [1921a: 132 and 299], but it is there left
undiscussed.
 K N OW L E D G E A N D I T S F O U N DAT I O N S 145

that are to count as providing the ‘foundation’ for the whole web of our beliefs about
the world that we experience.13
In RSP and KEW Russell’s doubts have not progressed so far. Thus in KEW he
distinguishes what he there calls ‘hard data’ and ‘soft data’, where the hardness
of a suggested datum is given by its ability to ‘resist the solvent influence of critical
reflection’, and the ‘softer’ data are those that become labelled as ‘doubtful’ when
subjected to this Cartesian examination (pp. 76–8). He does say say that such hardness is
a matter of degree, but in practice he seems to accept what PP has called the
foundations of our knowledge as ‘absolutely’ hard:
The hardest of hard data are of two sorts: the particular facts of sense, and the general truths of
logic. The more we reflect upon these, the more we realise exactly what they are, and exactly
what a doubt concerning them really means, the more luminously certain do they become . . .
Real doubt, in these two cases, would, I think, be pathological. (p. 78)

However it is still the case that, in principle, he is accepting that I can form a judgment,
which seems to me as if it merely describes a pure datum of experience, though in fact it
does go beyond what is strictly speaking given. This is one way in which what seems to
me to be a simple description of my experience may be mistaken. But it is not the only
way. For even in PP he recognizes that my report of what I am experiencing may be
mistaken, not by going beyond what is given, but just by misdescribing it: that is, I have
a complex sense-datum, and I offer an analysis of the complex object that I experience,
but my so-called ‘analysis’ need not be correct.
It is true that Russell’s own text sometimes suggests that my analysis of my own sense-
data cannot be mistaken, for example:
In all cases where we know by acquaintance a complex fact consisting of certain terms in a certain
relation, we say that the truth that these terms are so related has the first or absolute kind of self-
evidence, and in these cases the judgment that the terms are so related must be true. Thus this sort
of self-evidence is an absolute guarantee of truth. (p. 137)

This appears to claim an infallibility, but the claim evaporates when we read on:
But although this sort of self-evidence is an absolute guarantee of truth, it does not enable us to
be absolutely certain, in the case of any given judgment, that the judgment in question is true.
Suppose we first perceive the sun shining, which is a complex fact, and thence proceed to make
the judgment ‘the sun is shining’. In passing from the perception to the judgment, it is necessary
to analyse the given complex fact: we have to separate out ‘the sun’ and ‘shining’ as constituents
of the fact. In this process it is possible to commit an error; hence even where a fact has the first or
absolute kind of self-evidence, a judgment believed to correspond to the fact is not absolutely
infallible, because it may not really correspond to the fact. (Ibid.)14

13
Hence, as in mathematics and logic, we do not in practice start with the so-called ‘foundations’, but
reach them only after much theorizing.
14
Note, incidentally, that we must take the expression ‘the sun’ in this passage as an expression for a sense-
datum, and ‘the sun is shining’ as a description of the sense-datum, if the example is to be relevant to what
Russell regards as immediate knowledge.
146 K N OW L E D G E

This clarifies the position, and shows that the first statement needs revision. He does
not really mean that ‘this sort of self-evidence is an absolute guarantee of truth’.15
There is a rather different point elsewhere in PP, which genuinely does claim a certain
(very limited) infallibility:
It would seem that there are two kinds of self-evident truths of perception, though perhaps in the
last analysis the two kinds may coalesce. First, there is the kind which simply asserts the existence of
the sense-datum, without in any way analysing it . . . The other kind arises when the object of
sense is complex, and we subject it to some degree of analysis. (p. 114)

We have noted earlier (pp. 117–18) that ‘this exists’, said of an object with which one is
at the time acquainted, really does seem to be infallible, from Russell’s point of view.
But one can conjecture to a hesitation in the accompanying remark that ‘perhaps in the
last analysis the two kinds [of self-evidence] coincide’. It may be that he is already
thinking of the doctrine that he asserts later,16 that ‘ . . . exists’ can sensibly be applied
only where the subject is given by a description, so that ‘this exists’ is acceptable only
where the ‘this’ does represent an analysis of the complex sense-datum that is given. But
in that case the judgment is not infallible, for we have already seen that any such
analysis may (in principle) be mistaken.
In any case, one must say that while Russell does (in theory) allow for the possibility
of a mistaken analysis of what is given, he offers no examples. This may be, as Pears
suggests,17 because he thinks that such mistakes almost never happen, and that in
practice we will always get it right. I therefore supply a couple of examples where the
judgment reached by ‘analysis’ of one’s experience could well be wrong. No doubt in a
very simple case a mistake is highly improbable, as e.g. when one sees something red and
round and judges that it is red and (roughly) round.18 But in more complex cases error
is quite likely. For example, suppose that you are looking at a table with 23 oranges on
it, in no noticeable pattern, and suppose that you describe your experience as that of
seeing what is apparently a table with what appear to be 22 oranges on it. Well, it is
possible that your eyes are misleading you, and that you are receiving data from only 22 of
the 23 oranges. But it is much more likely that you have miscounted, i.e. that you do see
each of the 23 oranges but have made a mistake in saying that you see only 22. Nobody
would be surprised at that kind of mistake in describing what you are experiencing.
Here is another example, when it is actually quite surprising if one does not make a
mistake. Suppose that you are confronted with the Müller-Lyer diagram

15
Compare ‘On the Nature of Truth and Falsehood’ [1910b: 150], which makes the same claim for
infallibility, and then (in brackets) the same withdrawal.
16
This strange doctrine will be discussed in Chapter 14, section 4.
17
Pears [1979: 64].
18
The philosopher Charlie Martin used to relate this incident. One day he was dozing in a train when
there was a sudden and unexpected disturbance. He awoke to find that his hands had quickly moved in front
of his eyes, protecting them. But what had happened had been not a flash but a bang. So we may infer that it
is possible to mistake an auditory sense-datum for a visual one (and vice versa).
 K N OW L E D G E A N D I T S F O U N DAT I O N S 147

and asked to describe what you see. (We can make it clear that you are asked to describe
what Russell would call your sense-datum, and not what is actually there on the paper.)
Most normal people will say that they see two horizontal lines, with arrowheads
attached, and that the two lines are not equally long; the upper is shorter than the
lower. But you can quite easily convince them that they have misdescribed what they
see, e.g. by covering up the arrowheads, one by one, while leaving all of the horizontal
lines still visible. At some point in this process they will admit that the lines now look
equal, and that they do not appear to have changed in length. The fact is that we often
estimate relative lengths ‘by eye’, and it is not particularly surprising when we get them
wrong. What is unusual about the present diagram is that almost everyone who is not
familiar with this illusion will get it wrong.
Anyway, to revert to the main theme, Russell rightly claims that our beliefs about
what we are currently experiencing may always be mistaken, so we do not have that
reason for saying that they count as ‘immediately known’. Do we have any other? For
example, is it true that in our normal practice we treat peoples’ reports of their
experiences as more trustworthy than their reports of the physical objects that they
perceive? The answer must be ‘no’. We can easily imagine someone who (sincerely)
says: ‘I know that the table has a rectangular top, but from here it looks like a
parallelogram’. His first claim is very likely to be right: the top is rectangular, and he
does know that. But in that case his second claim is wrong, for there is no viewpoint
from which a rectangular area looks to have the shape of a parallelogram. He knows
what it is, but does not know how it looks, and this is very often the situation that we
are in. The point is obvious with children (or adults) who have not yet learnt to draw in
perspective. From the experiences that they have, they have learnt to recognize how
things are, and can describe them perfectly well. But they cannot describe the
experiences themselves.
In a way, Russell himself admits this, for he does accept that our belief in physical
objects does not arise as a result of reasoning from the nature of our experiences:
Of course it is not by argument that we originally come by our belief in an independent external
world. We find this belief ready in ourselves as soon as we begin to reflect: it is what may be
called an instinctive belief. (PP, p. 24)

Some instinct, then, leads us to interpret our experiences as signs of an external world.
And it is the external world that really concerns us, for that is what is more important
for our life, and we have no need to acquire explicit knowledge of the experiences
themselves. There is therefore no reason that has emerged so far for saying that we know
more about our experiences then we do about the external world (which, as we believe,
causes us to have them), or that our beliefs about experiences are more secure than
148 K N OW L E D G E

are our beliefs about physical objects. Why, then, should it be they, and they alone, that
are counted as ‘immediate knowledge’?
Russell quite often suggests that our beliefs about our experiences are ‘more certain’
than beliefs about physical objects, and indeed in KEW (p. 74) he observes that this is
‘proverbial’.19 (So it may be, amongst philosophers since Descartes, but that does not
make it true.) As we have just seen, it has no decent claim to truth if taken as an
observation about what particular people are certain of. I am entirely certain that there
is a desk in front of me, and I could not be more certain of anything else. I am also quite
certain that its top is (nearly) rectangular, but not at all certain about what shape it looks
to have, from the position at which I am sitting. So, if we take ‘certainty’ in this sense,
then Russell’s claim must simply be rejected. But should we take it in some other way?
Could it be maintained (after Descartes) that our beliefs about the physical world are
open to rational doubt in a way in which our beliefs about our experiences are not?
Well, Descartes seems to be right in saying that it is rational to doubt the existence of
the physical world while not doubting the existence of one’s own experiences. (His
story of the malin genie convinces us of that.) But there is something very odd about the
opposite: could it be rational to doubt the existence of one’s own experiences (and
everyone else’s?) while not doubting the existence of the physical world? The oddity
arises because we all think that it is our experiences that tell us about the physical world,
and so if we had no experiences we would have no reason to believe in that world. But,
if we grant both of these general beliefs (i.e. both the existence of experiences, and the
existence of the physical world), it is difficult to see any asymmetry in the certainty of
our particular beliefs about each, for in each case one can invent stories which would
rationalize a doubt. Indeed, if there is an asymmetry, I would propose that it runs in the
opposite direction to what Russell suggests. For the first thing that happens is that we
have experiences. The next thing that happens is that, on the basis of these experiences,
we form beliefs about the physical world. The third thing that may (or may not) happen
is that we form beliefs about the nature of the experiences themselves. So these last
beliefs are by no means the ‘foundation’ for what has preceded them, and may in fact
depend upon them. (For example: ‘I know it must look elliptical, for I know it is a
penny, and pennies are round, and round things do look elliptical from almost any
angle’.) I conclude that if Russell’s claims for the priority of experience beliefs have any
foundation, it must come from the general doubt introduced by Descartes, and not
from particular instances, which could be piled up in either direction.
The general doubt encourages a sceptical question: if, for the sake of argument, we
grant the truth of all (or almost all) of your beliefs about your present experiences, can
you justify from this basis your belief in an external world, which causes those
experiences, and whose nature can be inferred from the experiences that it causes?
This sceptical question is in itself a perfectly good question, and is one that Russell takes

19
The claim is that ‘what does not go beyond our own personal sensible acquaintance must be for us most
certain’ (KEW, p. 74). But Russell’s considered view is that all judgment ‘goes beyond’ acquaintance.
 K N OW L E D G E A N D I T S F O U N DAT I O N S 149

very seriously, as we shall see. I suggest that it is because he takes this question seriously
that he wishes to class our beliefs about our experiences as (immediate) knowledge,
while denying this title to our beliefs about physical objects: for he wants to be allowed
to take the former beliefs as his starting point in the attempt to justify the latter. This is a
permission that we may reasonably grant while still stickling over the notion of
knowledge. The one is not in fact known any better than the other, and (in the
usual sense) neither is in fact reached by argument from the other, but one may still ask:
could the one be justified from the other? Is such a justification possible in principle?20

4 The force of scepticism

Very soon after completing his work on The Problems of Philosophy, and perhaps partly as
a result of his first conversations with Wittgenstein, Russell began a further paper,
which he entitled ‘On Matter’.21 This was intended as a sceptical discussion of its topic,
and it contains arguments against the position that he had adopted in PP. In particular,
it makes the point on simplicity that, although it makes good pragmatic sense to try out
the simplest hypothesis first, still there is no ground for thinking that a simpler
hypothesis is more likely to be true. (As he somewhat tartly observes, mathematicians
may prefer simplicity, but we have no reason to suppose that the universe was created
‘for the purpose of delighting mathematicians’, p. 86.) The paper also argues that, even
if our ordinary and common-sense views on matter may be regarded as a hypothesis
which succeeds in explaining why our experiences should be as they are, still ‘such a
construction is only valid if it is known that there are not likely to be other hypotheses
which also fit the facts’, and ‘in our case this knowledge is absent’ (p. 87). (Russell is
entitled to this last claim because, as we shall see in a moment, the paper does sketch an
alternative hypothesis.) This part of his argument, then, is designed to show that the
position adopted in PP (especially pp. 20–5) does not have an adequate support.
Moreover, he generalizes its attack on our usual habit of accepting the simplest
explanation, to claim that we do not know of any a priori principle22
by which, from sense-data, we can infer the existence of entities of a sort with which we are
not acquainted, but which we know to possess the kind of properties that physics assigns to
matter. (p. 85)

20
Compare a usual position on induction (espoused by Hume and by Russell and by many others). It is
our instinct to suppose that a regularity which has held in the past will continue to hold in the future. As things
are, this instinct does not result from any argument which justifies the practice, though it may perhaps have
been useful for our survival so far. But this still allows us to raise the question: is there perhaps a rational
justification for what is in practice an instinctive habit?
21
He never published this paper, but it is now available on pp. 80–95 of volume 6 of his Collected
Papers [1992]. For some reason to suppose that Wittgenstein’s views may have been influential, see Miah
[1998: 99–109].
22
Presumably any such principle—if we could find one—would have to count for Russell as a priori, just
because it would have to count (like induction) as one of the foundations of all our knowledge.
150 K N OW L E D G E

What is sticking in his gullet is the thought of inferring to entities which are not ‘of a
sort with which we are acquainted’.
So the paper also outlines an alternative approach, according to which a piece
of matter is a kind of combination of familiar sense-data, ‘for example a combination
of visual, tactile, and other sense-data, associated together by some experienced
relation’ (p. 85). This approach, he suggests, will avoid scepticism
by preserving, as far as possible, what is most essential in the views of naive realism, namely the
belief that the existence of the sense-datum does not depend upon our perception. It seems
possible that this view might be preserved by assuming that all that could be a sense-datum to any
possible observer actually exists, and that collections of such actual and possible sense-data are
bound together in ways which enable us to regard them as one ‘thing’. The ‘matter’ of the
physicist and the ‘thing’ of common sense will then be collections of constituents of the nature of
sense-data, some actually perceived, some not. (p. 95)

The piece ends by recommending the further exploration of this idea, as our best
chance of a delivery from an unwelcome scepticism.23
Russell’s further exploration took a little time. Although he read the paper ‘On
Matter’ to a couple of meetings in 1912, he did not think that its idea was yet ready for
publication, but continued to work upon the topic. The result was the ‘logical
construction’ of matter that he published in 1914, and that forms the topic of the
next chapter.

23
Russell credits this idea to an article by T. P. Nunn [1910], entitled ‘Are Secondary Qualities
Independent of Perception?’. He and Samuel Alexander are the leading ‘British New Realists’ of this date.
For a description of their views I refer to Nasim [2008: chapters 2–3]. Nasim’s later chapters give an account
of how Russell’s ‘construction of the external world’ relates to their ideas.
 9
Logical Constructions (i): Matter

As we have seen, the overall structure of Russell’s view of knowledge is a foundation-

alist one: the foundations must be counted as known ‘immediately’, and all other
(genuine) knowledge has to be inferred from this basis. The previous chapter has
considered his view of the supposedly foundational beliefs, which are mainly our
beliefs about our present sense-data. This chapter and the next continue that theme
by considering Russell’s views on what we may legitimately infer from the founda-
tions, and how the inference is to be construed. Seen in this light, they therefore
continue the discussion of Russell’s epistemology. But at the same time they may
equally be regarded as making a start on his metaphysics. For a central area of
metaphysics is ontology, i.e. the question of what there is, and the chapters also have
that as their concern. This overlap is natural, since what is known has to be true, and
what is true must partly depend upon what exists. These two chapters may be regarded,
then, as concerned both with Russell’s epistemology and with his metaphysics. In this
way they form a bridge between what has gone before them and what will come
afterwards.

1 The idea of a Logical Construction

In 1914 Russell announced a new principle for philosophy, which he at once put to
work in his article ‘The Relation of Sense-data to Physics’ [1914b], henceforth RSP,
and in his book Our Knowledge of the External World [1914c], henceforth KEW. RSP
gives it this snappy formulation:1
The supreme maxim in scientific philosophising is this: wherever possible, logical constructions are to be
substituted for inferred entities (p. 155).

He tells us in both places that the idea is due to A. N. Whitehead, who of course had
recently been collaborating with him on the enormous task of writing Principia

1
A rather clearer formulation is found in the later article ‘Logical Atomism’ [1924: 326]: ‘Wherever
possible, substitute constructions out of known entities for inferences to unknown entities’.
152 K N OW L E D G E

Mathematica (RSP, p. 157; KEW, p. 8).2 Indeed, in KEW he adds that there is to be a
fourth volume of Principia, to be written mainly by Whitehead, in which this idea will
be applied to important areas of physics. No such fourth volume ever appeared, and we
have only a somewhat fragmentary account of what Whitehead intended.3 But in both
RSP and KEW Russell himself aims to apply this ‘supreme maxim’ by treating the
external world as a logical construction (from sense-data) rather than an inference.4
Russell had made use of such constructions before, when working out his views on
the nature of mathematics, though he had not then thought of himself as applying a
method which could be used more widely. For example, suppose that we are
concerned with the theory of the cardinal numbers. Then as a first step we ask for a
suitable criterion of identity for these numbers, i.e. ‘when do we have two instances of
the same cardinal number?’ Russell’s (initial) view is that it is primarily classes that have
cardinal numbers—i.e. they have that number of members—and we can give a non-
circular criterion for when they have the same number, as Cantor did: it is when their
members can be correlated with one another in a one-to-one fashion, without leaving
any out. Then the number itself can simply be identified with the class of all classes that
have the same number as some paradigm class with that number. We may employ a
similar approach in the case of the real numbers. Again, we begin by asking ‘when do
we have the same real number?’ and a suitable answer is the one given by Dedekind:
real numbers are the same if and only if they make the same division in the rational
numbers, i.e. just the same rational numbers are greater than each, or less than each.
Then a real number can be identified with such a division (a ‘cut’) of the rationals, or—
for simplicity—just with the class of all the rationals in the lower half of the division.
You will note that in both of these examples things of one kind (cardinal numbers, real
numbers) are identified with classes of things of another kind, and the same is true of
almost all of Russell’s logical constructions, both in mathematics and elsewhere.5 But
an exception is his logical construction of classes themselves, which was given earlier, in
section 4 of Chapter 4. There again one begins by asking ‘when do we have the same
class?’, and the proposed answer is ‘when the propositional functions that define the
classes are equivalent to one another, i.e. true and false of exactly the same arguments’.
This then allows us to eliminate talk of classes by talking instead of some or all
propositional functions that are equivalent to a given function.

2
An interesting discussion in Miah [1998: 99–109], suggests that Whitehead does not quite deserve the
credit that Russell here gives him. For Russell himself had envisaged the ‘construction of matter’ in his
unpublished ‘On Matter’, written early in 1912.
3
This is elaborated and discussed in my [2010].
4
The construction is also outlined in ‘The Ultimate Constituents of Matter’ [1915], which introduces a
few further details, but no radical departures from RSP and KEW.
5
The examples of cardinal numbers and of real numbers are given by Russell himself, in order to explain
the method, in his RSP, pp. 155–6. But he himself does not stress, as I have done, the importance of
beginning with a suitable criterion of identity.
 L O G I C A L C O N S T RU C T I O N S ( I ): M AT T E R 153

It is convenient to speak, as I have just done, of a logical construction as identifying

the entity to be constructed with a class (or whatever) of entities of a more familiar
kind. Assuming that we are anyway committed to these latter entities, then Russell can
fairly refer to his maxim as a case of Occam’s razor, i.e. as a way of not assuming the
existence of more kinds of entity than are needed (cf. KEW, p. 112). But he is clear
that, for this purpose, we do not actually have to claim an identity. It is enough if the
constructed entity can do all the work that the original entity was thought to be needed
for, so that it can replace the original without necessarily being identified with it. Here is
a clear application of this idea to Russell’s construction of cardinal numbers:
When we come to the actual definition of numbers we cannot avoid what must at first sight seem
a paradox, though this impression will soon wear off. We naturally think that the class of couples
(for example) is something different from the number 2. But there is no doubt about the class of
couples: it is indubitable and not difficult to define, whereas the number 2, in any other sense, is a
metaphysical entity about which we can never feel sure that it exists or that we have tracked
it down. It is therefore more prudent to content ourselves with the class of couples, which we
are sure of, then to hunt for a problematical number 2 which must always remain elusive. (IMP,
p. 18).6

(There is a similar comment on this definition at the end of PoM, }111, claiming that it
serves ‘all mathematical purposes’, and again at KEW, pp. 209–10.)7 Let us bear in
mind, then, that a construction need not be what one might regard as the analysis of an
existing concept; it is enough if it provides a suitable replacement for it.
Although Russell often speaks of his ‘logical constructions’ as the result of what he
calls ‘analysis’, or more fully ‘the logical-analytic method’,8 the best explanation of this
point seems to be that it is analysis in the proper sense that must first be directed on the
concept to be replaced, in order to lay bare the important properties which that
concept has, and which must therefore still be properties of its replacement. Analysis
may also tell you that some of these properties follow from others, which shows you
how to shorten this list of desiderata. In other words, it is analysis that shows you what
conditions your construction has to meet. But it is then sheer ingenuity that tells you
how to construct, from materials that are already accepted, something that does meet
them. No doubt what is constructed will usually be a class of some kind, but one has to
think of what members the class needs to have, and how they will provide it with the
desired properties. Perhaps in the mathematical case there was a point in calling the

6
As we have seen on pp. 107–10, there are in fact plenty of problems with Russell’s supposed class of all
couples. But that is irrelevant here.
7
Compare KEW, pp. 132–4. Having sketched a way of ‘defining’ a point of space as a class of extended
spatial regions, Russell again comments that ‘The first impression produced is likely to be one of wild and
wilful paradox’. But his discussion ends by claiming that, since the existence of such things as points, as
ordinarily conceived, is open to doubt, ‘it is more prudent, in order to avoid needless assumptions, to
substitute the class [in their place]’.
8
E.g. KEW, p. 7, p. 72.
154 K N OW L E D G E

construction ‘logical’, but in the cases to be considered in this chapter and the next the
word ‘epistemological’ might be more apt.
With so much by way of a general preamble, let us turn to the topic of this chapter,
which is the construction of the external world, i.e. (in the first place) of those physical
objects that we ordinarily think of ourselves as perceiving, from entities that are
presumed to be less open to doubt, i.e. sense-data.9 In PP such physical objects
were clearly taken to be ‘inferred’ from sense-data rather than ‘constructed’ from
them. The claim was that our sense-data are best explained by the hypothesis that
they are (for the most part) caused by more or less permanent objects existing
independently of us. But now these objects are to be constructed rather than inferred,
and since they must be neutral between all observers, and capable of existing unob-
served, the basis from which they are constructed must be expanded:10 it will include
not only actual sense-data, but also what Russell calls unsensed sensibilia. The existence
of such things is called a ‘hypothesis’ in both RSP (p. 158) and KEW (pp. 104–5), and
Russell describes it as a simplifying assumption which he hopes will turn out to be
stronger than is actually needed. But, as we shall see, he abandoned this hope quite
soon after writing these works, and the ‘hypothesis’ is one that we do have to take
seriously. He took it over from the British New Realists (p. 150 above), and Nasim
[2008] comments that his use of it was in fact ‘more cautious’ than theirs. It is ‘really
a conservative utilisation of what others felt justified and unabashed in employing’
(pp. 87–8). Nevertheless I shall consider it only as proposed by Russell himself.11

2 The Hypothesis of Unsensed Sensibilia

In RSP, and in the first (1914) edition of KEW, Russell supposed that an experience
was the instantiation of a two-place relation ‘S experiences O’, where S is the subject
that has the experience and O is the object that is experienced, i.e. a sense-datum. He
insists, as he had done in PP (pp. 40–3), that the object experienced is not the same
thing as the act of experiencing it, and that while the act may fairly be said to be ‘in the
mind’ the datum is not. But he now goes on (as he had not in PP) to infer from this that
the same sense-datum could in principle be experienced by another mind, or could

9
Many books on Russell have something to say on this construction. I would recommend Ayer [1971:
chapter 3, section B], and Sainsbury [1979: chapter 7, section 4]. Two more recent books that are wholly
devoted to this topic are Miah [1998] and Nasim [2008].
10
Material objects also differ from sense-data in having a more or less permanent existence, and Russell’s
construction will allow for this. Indeed, he takes it to be a feature both of common sense and of more
developed physics that we like to assume that the basic entities are things that do genuinely last for ever (cf.
KEW, pp. 107–12). But he himself has no sympathy with this idea.
11
In Russell’s writings the hypothesis first appears in his [1913a], when he says ‘Let us give the name
‘‘qualities’’ to those things that have all the properties common to all sense-data, with the possible exception
of being given in sense’. This introduces what he will later call not ‘qualities’ but ‘sensibilia’. It is pointed out
by Nasim [2008: 111–14], that he has here taken over the terminology used by T. P. Nunn (p. 150 above),
and has not yet invented his own.
 L O G I C A L C O N S T RU C T I O N S ( I ): M AT T E R 155

exist without being experienced at all. He adds that the first of these possibilities does
not actually occur, apparently on the ground that no two people ever occupy exactly
the same point of view at exactly the same time (RSP, pp. 158–9; KEW, pp. 94–5).12
But the construction that he is introducing will depend upon the second. This leads to
a slight revision of terminology. A sense-datum is to be called a datum only when it is
actually given, i.e. is being experienced by some mind. Otherwise it is called a sensibile,
i.e. something that can be sensed, but perhaps is not actually being sensed. We may
think of Russell’s position in this way: if I raise my finger then there is such a thing as the
view of the table from where my fingertip now is. No one is now having that view, for
there is no one who is looking from that position. If someone were to change position,
and look from there, then he would be getting that view, or—at least—one rather like
it. For he would be seeing what the table looks like then, at the time when he is
looking, which of course need not be just the same as how it looks now. Why do we
need to suppose that there are such things as these unexperienced views? Because
Russell is going to ‘construct’ ordinary physical objects (such as tables) from the views
and other sense-data which we think of them as causing. But there are many physical
objects, e.g. a rock on the far side of the moon, which never have and perhaps never
will cause any sense-data at all. The move to include merely possible sensibilia is an
attempt not to leave them out. But it certainly creates problems.
Russell of course recognizes that in the normal case (i.e. excluding dreams, hallu-
cinations, etc.) the visual appearance of things depends (a) on the object seen, (b) on the
state of the medium between the object and the perceiver (e.g. what light there is,
whether there is a fog or a heat-haze, and so on), and (c) on the state of the perceiver’s
sense-organs, optic nerves, and brain (e.g. whether he or she is near-sighted or far-
sighted, or is colour-blind, or is also experiencing an after-image, and so on). Never-
theless he wishes to hold that there are things like visual appearances even in places
where there are no perceivers. For example, he says:
We have not the means of ascertaining how things appear from places not surrounded by brain
and nerves and sense-organs, because we cannot leave the body; but continuity makes it not
unreasonable to suppose that they present some appearance at such places. Any such appearance
would be included among sensibilia. (RSP, p. 150; cf. KEW, p. 95)

One asks: how can it not be ‘unreasonable’ to suppose that there are ‘appearances’ at
places where nothing is actually appearing to anyone, and where even the conditions
to determine what would appear to a person are not fully specified, since that person’s
perceiving apparatus is not given? The immediate continuation appears to recognize
this point, for Russell goes on:

12
Does he infer from this that their (visual) sense-data cannot be exactly similar? If so, the inference is
clearly a mistake, but one that resembles a mistake that he clearly will make later on, and which I come to on
p. 161 below.
156 K N OW L E D G E

If —per impossibile—there were a complete human body with no mind inside it, all those sensibilia
would exist, in relation to that body, which would be sense-data if there were a mind in the
body. What the mind adds to sensibilia, in fact, is merely awareness: everything else is physical or
physiological. (RSP, p. 150)

The implication of this passage is that a perceiver’s sense-organs are needed for a
sensibile to exist, i.e. for there to be something to which the mind merely has to add
awareness. But it is an implication which Russell continues to ignore in what follows.
For example, a little later on we read once more:
Besides the appearances which a given thing in the room presents to the actual spectators, there
are, we may suppose, other appearances which it would present to other possible spectators. If a
man were to sit down between two others, the appearance which the room would present to
him would be intermediate between the appearances which it presents to the two others: and
although this appearance would not exist as it is without the sense-organs, nerves, and brain of
the newly arrived spectator, still it is not unnatural to suppose that, from the position which
he now occupies, some appearance of the room existed before his arrival. (RSP, pp. 154–5;
cf. KEW, p. 95)13

So there is an appearance at a place where there are no sense-organs to determine what

it looks like. But what kind of a thing is this?
What we can agree about is that at each unoccupied place there actually are a
number of light-waves, of various wavelengths, travelling in various directions. In our
normal way of thinking, these will cause an appearance if they impinge upon a suitable
eye and nerves and brain, but they are not in themselves an appearance, nor anything
like an appearance. It is true that Russell does not actually say in so many words that
sensed and unsensed sensibilia are the same kind of thing. But he does say that they
have ‘the same metaphysical and physical status’ (RSP, p. 148), he does call both of
them ‘appearances’, and he does claim that when we infer from the existence of those
that are sensed to the existence of those that are not then we are inferring to things that
are ‘similar’ to those we know (RSP, p. 157). That surely cannot be a description of
light-waves.14 So I now make a different suggestion, but one which to some extent
relies upon a later development in Russell’s thought.
In RSP it is stressed that sense-data, and hence sensibilia, are to be construed as the
objects which we sense, and not the acts of sensing them. But later Russell changed his
opinion on this, and came to believe that acts of sensing did not actually have either an
independent object that was sensed, or an independent subject (the self) that did the
sensing. In his later view all that really exists is the event that is the sensing, or the

13
The passage continues: ‘This supposition, however, need merely be noticed and not insisted upon’.
Here Russell is hinting at the view which he then held, that unsensed sensibilia ought to be dispensable. I shall
come to this on pp. 159–60 below. It is enough to say here that he quite soon abandoned this view, but did
not abandon unsensed sensibilia.
14
Compare the passage quoted on pp. 149–50 above from ‘On Matter’. Russell thinks that he is not
inferring ‘the existence of entities of a sort with which we are not acquainted’.
 L O G I C A L C O N S T RU C T I O N S ( I ): M AT T E R 157

appearing, itself. This change of view was part of his conversion to the theory of neutral
monism, which I shall discuss in the next chapter. But for the moment let us just note
his change of mind, which he records in a paragraph that he added to the second
edition (1926) of KEW:
According to some authors—among whom I was formerly included—it is necessary to distin-
guish between a sensation, which is a mental event, and its object, which is a patch of colour or a
noise or what not. If this distinction is made, the object of the sensation is called a ‘sense-datum’,
or a ‘sensible object’. Nothing in the problems to be discussed in this book depends upon the
question whether this distinction is valid or not. If it is not valid, the sensation and the sense-
datum are identical. If it is valid, it is the sense-datum which concerns us in this book, not the
sensation. For reasons explained in The Analysis of Mind [1921a] (e.g. p. 141ff.) I have come
to regard the distinction as not valid, and to consider the sense-datum identical with the
sensation. But it will not be necessary to assume the correctness of this view in what follows.
(2nd edn, p. 83)

For the purposes of KEW, he says, it does not matter which view we take. But in
one respect it surely does matter, namely on the question of how to construe the
notion of an unsensed sensibile. For if an act of sensation has no object, then a sensibile
cannot be such an object, even potentially. It can only be, not an actual act of sensing,
but a possible act of sensing. And perhaps that is how unsensed sensibilia should always
have been construed, namely as possibilities rather than actualities? It is true that in RSP
Russell very clearly claimed that a sensibile should be regarded as something that exists
actually, and is not merely a possible sense-datum (p. 157). But after his change of view
on what sense-data are, this construal is surely inevitable. It is fair to say that in other
respects this change makes little difference to the overall lines of his construction of
matter,15 but it does force a new view of sensibilia, which avoids some of the problems.
For instance, we said that an actual appearance is determined partly by the state of
the sense-organs, nerves, and brain of the perceiver, so where there is no perceiver the
supposed ‘appearance’ is undetermined. But now we say: in that case, all the possible
appearances that could be experienced at that place are to be included, e.g. both how
things would appear to the near-sighted and how they would appear to the far-sighted
(and, of course, to those in between). Similarly for other ways in which those same light
rays might make things appear when conjoined with this or that type of sense-organ.
Moreover, there is no good ground for showing a special partiality to the sense-organs
of human beings. For we should also include the ways in which a thing may appear to a
dog or a cat, a sparrow or an eagle, a goldfish or a shark, and so on and on indefinitely.
For all of them are possible ways in which that thing may appear, and all have an equal
title to be included in the collection of possible appearances that will serve in place of
that thing in Russell’s construction.

15
Here is one small difference. If sense-data are to be identified with the events of sensation, then the
reason why you and I cannot experience the same sense-datum is just that no one event can occur in two
different people. (It is irrelevant that we cannot simultaneously share exactly the same viewpoint.)
158 K N OW L E D G E

Moreover, it is not clear that we have yet gone far enough in including possible
appearances. What we have been considering so far is the same physical object, viewed
from a given position at a given time, and in those same external conditions—e.g. the
same lighting—as actually obtains in that place at that time. But shall we not also wish
to consider what it would have looked like if the lighting had been different (e.g. if
there had been some light, rather than total darkness)? To bring out this point, let us
bring in another example, namely a lump of matter which never is perceived, and
which never can be perceived while its external conditions are kept as they are, e.g.
a lump of rock completely surrounded by other matter in the middle of a mountain.
Do we want to say that in that case it does not count as a physical object? Or would
we rather say: ‘well, there are possible ways in which it could be perceived, e.g. if it
were first dug out of its present surroundings’? Apparently we must take one or other
of these views if Russell’s overall plan is to be maintained, and the second is surely more
attractive.

3 The advantage of this hypothesis?

Without pursuing these details any further, let us now ask: what is supposed to be the
advantage of ‘constructing’ physical objects in this way, i.e. of replacing them by a way
of referring to all their possible appearances? After all, this does posit the existence of a
number of items that are not experienced, whereas our previous examples of ‘logical
constructions’ introduced no new items.
The answer that one usually finds in those who comment on this matter makes use
of a handy metaphor distinguishing ‘vertical’ inferences from ‘horizontal’ inferences.16
A vertical inference takes us ‘upward’ from some level of ontological commitment to a
higher level; in this instance it would take us (as in PP ) from the level of sense-data to
items of a quite different sort, that are posited as causing the sense-data. By contrast,
a horizontal inference takes us only from items at a given level to more items at the
same level. In this instance I start from my own sense-data, I infer from there to other
actual sense-data (e.g. in other people and in animals), and then again to merely
possible sense-data, not actually experienced at all. But the inferences move only
from sensibilia to sensibilia. Then the suggestion is that Russell thinks of horizontal
inferences as somehow ‘more secure’ than vertical ones, and hence less open to
sceptical worries. He evidently takes this thought too far when he says or implies
that physics (or common sense) cannot be verified at all if it is construed as positing
objects which are not just functions of sense-data,17 but he surely does think that it is
better verified if it avoids such posits.
This position seems open to doubt. It is of course true that I believe in unsensed
sensibilia insofar as I believe that, if I were to look towards the table from where my

16
The terminology is due to Ayer. See e.g. his [1971: 57].
17
E.g. RSP, pp. 146–7; KEW, pp. 88–9, 106, 117.
 L O G I C A L C O N S T RU C T I O N S ( I ): M AT T E R 159

fingertip now is, then I would get a view of it. However one must also add that I have
that belief because I also believe that the table exists as a more or less permanent and self-
standing object that is open to view from all appropriate angles. That is, I do in fact
make the vertical inference which Russell is avoiding, and apparently it is that vertical
inference which provides me with my ground for the horizontal one. For if I do not
make the usual vertical inference—for instance, if I think that I may be suffering from
some kind of illusion, and that perhaps there is not really a table there at all—then I also
do not make the horizontal inference either. It is therefore not altogether clear that one
who makes the horizontal inference without the vertical one is really better placed to
answer the sceptic, for his inference is now shorn of its usual justification. But the
position may sensibly be debated.
A different rationale for Russell’s theory would be, not that it allays sceptical doubts,
but that it satisfies what Quine has called ‘a taste for desert landscapes’, i.e. the thought
that it would be an attractive feature of the world if it contained entities of as few
different kinds as possible. Russell evidently did find this feature attractive, for it is
surely a large part of the motivation for his later move to so-called ‘neutral monism’.18
Perhaps, then, it is more ontology than epistemology that urges him towards this
‘reconstruction’ of physical objects.
In any case, there is one other point which is worth mentioning here: at the time of
writing RSP and KEW Russell evidently hoped that his appeal to unsensed sensibilia
was only a temporary measure. His idea was that, for the time being, this appeal will
provide a useful way round the problem of physical objects that happen not to be
perceived, but in the longer term it should be possible to eliminate it. This hope is
explicitly expressed in both RSP and KEW, for example:
A complete application of the method which substitutes constructions for inferences would
exhibit matter wholly in terms of sense-data, and even, we may add, of the sense-data of a single
person, since the sense-data of others cannot be known without some element of inference. This,
however, must remain for the present an ideal, to be approached as nearly as possible, but to be
reached, if at all, only after a long preliminary labour of which as yet we can see only the very
beginning. (RSP, p. 157; cf. KEW, p. 117)

However, when looking back on the situation from his later perspective in MPD
[1959] he reports that he quite soon abandoned this goal:
I soon, however, became persuaded that this is an impossible programme and that physical
objects cannot be interpreted as structures composed of elements actually experienced . . .
Accordingly, I gave up the attempt to construct ‘matter’ out of experienced data alone, and
contented myself with a picture of the world which fitted physics and perception harmoniously
into a single whole. (MPD, p. 105)

18
This motivation is avowed at ‘On the Nature of Acquaintance’ [1914a: 145].
160 K N OW L E D G E

Consequently the unsensed sensibilia that KEW had characterized as ‘ideal’ entities
(p. 117), and that RSP had said ‘may be thought monstrous’ (p. 158) never were
dispensed with, and they remained central to his construction. Moreover the construc-
tion itself remained an important part of his thinking about the nature of the physical
world for some time to come. It should not have done. For it is easy to see that the
construction actually given in RSP and KEW is a failure, and plausible to conjecture
that any other similar construction would also fail.

4 Identity at a time
The idea is to ‘construct’ a material object from its appearances, so the first question to
ask is: when do two distinct appearances count as appearances of the same object?
Russell very reasonably divides this question into two others: when do two simultaneous
appearances count as appearances of the same (momentary) object? and when do two
appearances at different times count as appearances of the same (more or less perma-
nent) object? Quite naturally, he chooses to begin with the first of these.
For simplicity, we may confine attention to visual appearances, and the whole set
of visual appearances experienced by a single perceiver at a single time is what Russell
calls a ‘perspective’ on the world—or, as one might more simply say, a ‘view’ of the
world—from one place at one time.19 Then in RSP he says:
We discover that two different perspectives, though they cannot both contain the same
‘sensibilia’, may nevertheless contain very similar ones; and the spatial order of a certain group
of ‘sensibilia’ in a private space of one perspective is found to be identical with, or very similar to
the spatial order of the correlated ‘sensibilia’ in the private space of another perspective. In this
way one ‘sensibile’ in one perspective is correlated with one ‘sensibile’ in another. Such
correlated ‘sensibilia’ will be called ‘appearances of one thing’. (RSP, p. 160)

Notice that all that is being appealed to is the similarity between the two perspectives.
The same applies to what he says in KEW:
It is possible, sometimes, to establish a correlation by similarity between a great many of the
things of one perspective, and a great many of the things of another. In case the similarity is very
great, we say the points of view of the two perspectives are near together in space . . . We can
now define the momentary common-sense ‘thing’, as opposed to its momentary appearances.
By the similarity of neighbouring perspectives, many objects in the one can be correlated with
objects in the other, namely with the similar objects. Given an object in one perspective, form
the system of all the objects correlated with it in all the perspectives; that system may be identified
with the momentary common-sense ‘thing’. (KEW, p 96)

19
In later writings a single ‘perspective’ is not confined to visual sense-data, but includes all the sense-data
experienced by a given subject (or possible subject) at a given time. They are tied together by the relation ‘x is
(or could be) experienced together with (i.e. at the same time as) y’, which itself is a relation that we can
experience instances of. Cf. ‘On the Nature of Acquaintance’ [1914a: 131] and ‘The Ultimate Constituents
of Matter’ [1915: 141]. This explanation allows for the existence of unperceived perspectives.
 L O G I C A L C O N S T RU C T I O N S ( I ): M AT T E R 161

Here we have the addition that perspectives with very similar contents are to be
regarded as views from very nearby places, and hence to be views of (almost all) the
same objects. But still the only criterion offered is similarity.
If this criterion were adequate, we could use it to introduce suitable spatial relations
as holding between perspectives, which would in fact be the spatial relations between
the positions in (our public) space where those perspectives are views from. For
example, three perspectives x, y, and z would be ‘co-linear’ if the sense-data in each
were very similar, and exactly alike in their arrangement, except that all the sense-data
in perspective x were smaller than the corresponding ones in y, and all those in y were
smaller than the corresponding ones in z. For in this case, x, y, and z would be views of
the same objects, taken in the same direction, and x would be further from those
objects than y is, while y is further from them than z is. Clearly, we could locate further
perspectives along the same line as these initial three, and filling up the gaps between
them. Then, by taking into account some other variations in the sense-data contained
by different perspectives, and their mutual arrangement, we could introduce further
spatial ideas, such as the idea that several perspectives were all views from the same
position, but looking in different directions. In general, given the assumption that
similar perspectives are near to one another, and the more similar they are the more
near they are, one can evidently introduce a space in which the perspectives themselves
are located.20 But the assumption is crucial, and it is quite clear that it cannot be
accepted.
To use an example which Russell himself introduces a little later, consider an
ordinary penny on a white tablecloth, and a view of that penny which shows just it,
and its white background, and nothing else. Must another view, which looks very like
this one, be a view of the same penny? Obviously, the answer is ‘no’. There are very
many pennies and white tablecloths, and they all look much the same. Indeed, given a
specified distance and angle of vision, some will look exactly the same. As Russell
himself later remarks (when thinking of his second problem), ‘two different things may
have any degree of likeness up to exact similarity’ (KEW, p. 113). But this point clearly
destroys the first move in his construction.21
To bring out what is involved here, consider what it is for you and me each to be
seeing the same penny. It would be relevant to know whether we are looking from

20
I have simplified here by assuming that the perspectives in question are those which the same perceiver
would get from the various different positions, or anyway perceivers whose visual apparatus is in an exactly
similar physical state, so that they all share the same sharpness of vision, the same sensitivity to different
colours, and so on. Further complications might allow this simplification to be omitted, as section 2 has
suggested that it should be. Another awkwardness that this space would allow us to overcome is that the
experiences we get in dreams and hallucinations may now be set aside as not belonging to perspectives from
any position in the space.
21
This decisive objection is clearly given in Ayer [1971: 62–3], and in Sainsbury [1979: 248]. It is
mentioned in Miah [1998: 182–3], but its importance is downplayed. It is not even mentioned in the account
given by Nasim [2008: chapter 5].
162 K N OW L E D G E

nearby places, for if in fact we are far apart then it is first-off improbable that we are
seeing the same penny, no matter how similar our views are. But it is not impossible, as
we see by considering such things as telescopes (and perhaps also such things as
television). Conversely, it is also possible that we are standing very near to one another,
and that we each have our eyes turned in a similar direction, and yet we are not seeing
the same penny. You can easily think of arrangements with mirrors that would have
this effect. More strongly, it may be that light-rays from the same penny are impinging
both on your retina and on mine, and both of us are having very similar penny
appearances, and yet we are not seeing the same penny. This would happen if, for
example, your eyes, nerves, and brain were operating in the normal way, so that your
penny appearance was caused in the normal way, whereas mine were not operating at
all, so that I was not actually seeing anything, but was getting a hallucination of a penny
which just happened to be very like your genuine view. Reflecting on these points,
and others like them, one is led to see the importance of the causal theory of perception:
one sees x if and only if x is, in an appropriate way,22 the cause of one’s x-like sense-
datum. This brings in an idea which does in a way figure in Russell’s discussion, but not
in the way that is needed.
If I am seeing a penny, then there is a route along which light reflected from the
penny is travelling towards me. All along this route there are (in Russell’s view) a series
of sensibilia, which are views of the penny that are broadly similar,23 except that the
penny takes up less and less of the view as one gets further from it. This series continues
within my head, as the route continues from my eye to my brain. So Russell could
insist that in any case of genuine seeing such a series of sensibilia must exist. But, in
order to ensure that this requirement does the work it is intended to do, we must add
that the sensibilia in question are taken along a spatially continuous route, i.e. that
sensibilia which are near to one another in the series are located at places that are near to
one another in space. Moreover, to put the claim that you and I are seeing the same
penny, we must add that these two continuous routes should start at the same place (or
at places near to one another). That is, we need to be able to specify further conditions
on the places involved. What happens in Russell’s own discussion is that he supposes
that he has already explained when two sensibilia are of the same penny, and he goes
on to use these ideas to explain what is to count as the place of the penny and the places
between it and me. (The place of the penny is the place where all these routes of
suitably related penny appearances intersect.) But the actual position is the other way
round. In order to say which sensibilia are sensibilia of the same penny we must first
understand the idea of the place of a perspective, and the place of the penny, and hence

22
Some causal routes are too devious to count as appropriate, e.g. if the penny is seen in the normal way
by a wicked neurophysiologist, who then somewhat arbitrarily decides to ‘feed’ into my brain an image
which exactly resembles his. The original penny is then a cause of the way things look to him, which—by this
unusual route—becomes a cause of the way things look to me. But I do not count as seeing that penny.
23
For simplicity I here ignore possible distortions due to the intervening medium, e.g. if the light-rays
have to travel through a coloured pane of glass on their journey.
 L O G I C A L C O N S T RU C T I O N S ( I ): M AT T E R 163

of the spatial routes between the one and the other. To put it briefly, if we are given
‘the same object’ then we can no doubt introduce ‘the same place’, and given ‘the same
place’ then we can no doubt introduce ‘the same object’. But Russell wishes both of
these to be ‘constructed’ from some observable relationship between sensibilia, and the
only relation that he can offer is ‘looking similar’. This clearly cannot do the work that
is needed, and so we can only say that the first step of his construction is inadequate.24

5 Identity over time

The second step is rather better. Given that there is a penny at one particular place at
one particular time, in order to explain what counts as that same penny at another time
one may begin with the idea of a continuous ‘track’ that the penny follows through
space and time. For each time (during whatever period is in question) there will be just
one place on this track, and it will be occupied at the time by just one penny and by
nothing else. Moreover, the track is to be a continuous track, not leaping instanta-
neously from one place to another, nor suffering from gaps in time, i.e. from there
being times at which the track fails to exist. Given an ordinary understanding of places
and times, it is quite easy to give a proper definition of this required continuity. In the
simple case of a penny, this is probably the only further idea that we need: a spatio-
temporally continuous track in which each place is filled by a penny will inevitably be a
track in which each place is filled by the same penny. But other cases are more
complicated, and—as Russell sees—more needs to be said.
To illustrate this need, let us briefly consider living things. The spatio-temporal track
of a living thing is not strictly continuous, for living things eat and excrete, and can lose
bits of themselves—e.g. a nail-paring, or the whole fingernail, or a finger, or even a
whole arm—while yet continuing to exist. Moreover, there are other features of a
living thing that can remain, to preserve their identity, despite the change of matter.
Locke hoped to elucidate this by saying that a living thing has to preserve ‘the same
life’, though reflection is apt to suggest that this idea is too vague to be helpful. (For
example: does a caterpillar preserve ‘the same life’ when it metamorphoses into a
chrysalis and then a butterfly?) However we need not pursue this issue, but should turn
to the question that Russell himself considers. His overall aim is to become clear about
the relation between our experience of the external world and what physics has to tell us
about that world, and physicists are not much concerned with the identity of living
things. But they do need the concept of matter, and even in this (fairly simple) case it

24
I add here a minor clarification and a minor modification. The clarification is that the places here in
question are places in the public space, that we all inhabit, and that Russell also calls ‘perspective space’. I shall
not, here or elsewhere, have anything to say about Russell’s so-called ‘private spaces’. (These are supposed to
be given by the spatial relations perceived as holding between the sense-data of a single perspective.) The
modification is to acknowledge that the sensibilia which mark out the path taken by a light-ray are not strictly
simultaneous with one another, for one must allow for the velocity of light. (The point would be more
obvious if we had been considering not light but sound.)
164 K N OW L E D G E

turns out that we need to rely on more than just spatio-temporal continuity in order to
say what is to count as ‘the same matter’ over a period of time. For example, since
Newton’s time it has been an accepted requirement that the same matter should
preserve always the same mass, but mere spatio-temporal continuity will not by itself
ensure this.25 Let us turn, then, to Russell’s ultimate goal in this endeavour, which is
the construction of matter.
We started with the idea that an ordinary material object should be identified with
(the class of )26 all of its appearances, both actual and possible. ( We have seen that there
are important problems with this idea, but let us now set them aside.) Turning from our
everyday notion of a material object to something that is more relevant to physics,
Russell now notes that the way that a thing appears will depend partly upon the
properties of its matter, and partly upon any distortion introduced by the medium that
intervenes between that matter and the appearance. So in order to come close to the
matter of the thing, we need somehow to discount the effects of the medium. This is to
be done by once more taking into account the series of sensibilia that lie on the route
from the appearance to the matter, and then taking the limit to which this series
converges as the sensibilia approach more and more closely to the originating matter.
For as one approaches more closely the effects due to the medium must diminish, until
in the limit they are reduced to zero. It is the limit, then, that best reveals the matter of
the thing; and its matter at a given time may therefore be identified with the class of all
limits of all the series of sensibilia that start from it at that time. ( The sensibilia are in
principle observable, but for the most part calculated and not actually observed; their
limits have to be calculated, and are not observable.)27
This is Russell’s explanation of what matter is, or rather, his explanation of the
substitute notion that should replace the original notion of matter in any future
developments. It is clear that the explanation leans on the idea of matter as a cause,
but still it does not explicitly invoke the notion of a cause, though it does use the idea of
a continuous space-time track. This is as it should be, for at this date Russell does not
accept what one may reasonably call the ordinary notion of a cause; instead he will
grant only that physics is concerned with what are called ‘causal’ laws, but he regards
these simply as stating regular connections, which allow one to infer from a thing or
event of one kind to a thing or event of another, whether earlier or later, nearby (in

25
Russell gives examples to demonstrate this point. Further examples, and a general (but sceptical)
discussion of ‘the same matter’ may be found in my [2001].
26
More strictly: the object at a time is the class of all sensibilia that originate from that object at that time;
the object over time is a temporal series of such classes.
27
The limit of the series is determined by what the sensibilia are like as they become closer and closer to its
place of origin—i.e. by what they are like when they are so close that no living animal can actually sense them
(for all sensation involves a distance within the animal between the sense-organ and the brain). Does this
mean that no limit can actually be calculated?
 L O G I C A L C O N S T RU C T I O N S ( I ): M AT T E R 165

space and time) or quite distant, and so on. In his thinking at this time there is no
connection between causing and continuous space-time tracks.28
So far we have a ‘construction’ of what is to count as the matter of an ordinary
(perceptible)29 thing at a given time. It is this matter that the laws of physics are
concerned with, but they are concerned with how it behaves over time, which brings
us to the question of what is to count as the same matter from one time to another. As
noted earlier, relevant thoughts are that the same matter (so long as it remains all in one
lump) will exhibit spatio-temporal continuity, and will preserve the same mass. There
may well be other constraints, but Russell makes no attempt to explore them in any
detail. Instead, he aims to sum them all up with a general reference to the laws of
physics (and especially dynamics): to obtain a condition for ‘the same matter’, he says
The characteristic required in addition to continuity is conformity with the laws of dynamics.
(RSP, p. 171)

And this leads him to his conclusion

Thus we may lay down the following definition: Physical things are those series of appearances whose
matter obeys the laws of physics. That such series exist is an empirical fact, which constitutes the
verifiability of physics. (RSP, p. 173)

Presumably Russell means that a physical thing is a temporally ordered and continuous
series of classes of nearly simultaneous (actual and possible) experiences. Presumably the
thing’s matter is a similar temporally ordered and continuous series of classes of
simultaneous limits of chains of such (actual and possible) experiences. Presumably
he is thinking of the laws of physics as reformulated so that they become laws
concerning such series of classes of limits (though he has not given any examples of
how this is to be done). Given these rephrasings, I make no further comment on this
definition, save to repeat that it is matter that has to obey the laws of physics, and not all
those more ordinary things that are made of matter. In this respect the definition just
cited from RSP has it right, whereas the corresponding definition in KEW ( pp. 115–
16) omits this important point. (It has ‘which obeys’ in place of ‘whose matter obeys’ in
the cited definition. This is no doubt because it has earlier set aside the distinction
between a thing and its matter as one that may be ignored, p. 113.)30
What is most obviously wrong with Russell’s attempted ‘construction’ of matter
is not this criterion for what counts as ‘the same matter’ over time, but its initial

28
I shall comment later on Russell’s views on causality, i.e. in section 1 of Chapter 11.
29
Note that Russell here ignores the matter of imperceptible things.
30
The two preceding sentences in RSP suffer from the same error, i.e. they speak of physical things in
general, rather than their matter. (‘If it is to be unambiguous whether two appearances belong to the same
thing or not, there must be only one way of grouping appearances so that the resulting things obey the laws of
physics . . . ’ With ‘things’ in general there obviously is such an ambiguity: we cannot say how much a thing
such as a ship may be changed while still remaining ‘the same thing’, i.e. the same ship. But of course this
poses no problem for the laws of physics.)
166 K N OW L E D G E

explanation of when two (simultaneous) appearances are appearances of the same

material object. If this step is granted, then the rest may well be characterized as giving
a strange and unexpected account of matter, but it is not obvious that it will not do as a
replacement for the more usual understanding of matter. So this chapter ends by once
more going back to the first step.

6 Reconsideration of the problem

In later writings Russell does not attempt to revise the details of this ‘construction’ of
the material world, but he does offer a general argument which aims to show that such
a construction must be possible. In his lectures on ‘The Philosophy of Logical
Atomism’ [PLA, 1918b] his example is ‘what is it for the same desk to persist over
time?’, and he has this to say:
The essential point is this: what is the empirical reason that makes you call a number of
appearances, appearances of the same desk? What makes you say on successive occasions, I am
seeing the same desk? The first thing to notice is this, that it does not matter what is the answer, so
long as you have realised that the answer consists in something empirical and not in a recognised
metaphysical identity of substance. There is something given in experience which makes you call
it the same desk, and having once grasped that fact, you can go on and say, it is that something
(whatever it is) that makes you call it the same desk which shall be defined as constituting it the same
desk, and there shall be no assumption of a metaphysical substance which is identical throughout.
(PLA, lecture VIII, p. 273)31

Now we need hold no brief for what Russell seems to have in mind as ‘metaphysical
desk-substance’,32 but nevertheless there is an obvious response to this proposal. What
in practice leads me to say ‘it is the same desk (as was here yesterday)’ will always be
evidence that I myself regard as inconclusive. Usually, it is just ‘well, it looks the same,
and is in the same position’. Sometimes I may add things like ‘I know of no reason why
anyone should have changed it’, and possibly ‘it would not be dusty in just the same
places if it were a different one that had been substituted’, or ‘the door has been locked
all the week, so no one could have made a substitution’, and so on. We almost always
do have some reasons to give in justification for the claim ‘it is the same desk’, but in all
usual cases we recognize that the evidence that we cite is inconclusive. (It is easy to see
how the reasons just mentioned could turn out to be inadequate.) It would therefore
be quite wrong to define such reasons as constituting what is meant by ‘the same desk’.
This PLA passage concerns identity over time, which was not our main cause of
concern, though it does illustrate Russell’s strategy. But in The Analysis of Mind [1921a]

31
A little later (p. 277) he says something very similar about what makes us say ‘these are two experiences
of the same person’.
32
The idea is that there should be some one thing in the desk which remains unaltered so long as we still
have the same desk, and no matter what (other) alterations the desk may undergo. (Cf. KEW, pp. 111–12.)
 L O G I C A L C O N S T RU C T I O N S ( I ): M AT T E R 167

there is a similar passage which does address the initial problem. Russell is here
considering a number of people looking at the same table, and so each getting an
appearance of that table. He imagines someone who asks33
If there is no single existent [i.e. a ‘real’ table] which is the source of all these [appearances], how
are they collected together?

The answer, he says, is simple:

Just as they would be if there were such a single existent. The supposed ‘real’ table underlying its
appearances is, in any case, not itself perceived but inferred, and the question whether such-and-
such a particular is an [appearance] of this table is only to be settled by the connection of the
particular in question with the one or more particulars by which the table is defined. That is to
say, even if we assume a ‘real’ table, the particulars which are its [appearances] have to be
collected together by their relations to each other, not to it, since it is merely inferred from them.
We have only, therefore, to notice how they are collected together, and we can then keep the
collection without assuming any ‘real’ table as distinct from the collection. (pp. 98–9)

However, when we do ‘notice how the appearances are collected together’, then (a)
we see that this collecting involves assumptions which are not available to Russell’s
constructionist, and (b) that it once more depends upon inferences which we all
recognize are not conclusive. The initial answer to why we think that each person is
getting an appearance of the same table is that they are all in the same room, and all
looking in the right direction. But this crucially involves the concept of ‘the same
place’, which we have said is not yet available (and which apparently cannot be
introduced without circularity).34 Moreover, we are simply assuming that each per-
son’s eyesight is working normally, and this is an assumption that cannot be conclu-
sively established. (Perhaps one of us is getting a television picture of a different table? It
is not obvious that that would have to be revealed by his observable behaviour.)
To put the point more generally, it is this. We all share a theory about what there is in
the world, and how it works. At a common-sense level, this involves the existence of
more-or-less permanent objects, independent of our perceptions, and the causal role
that these objects play. Russell is recommending that we replace this theory by the
perceptual evidence that we have for it. But the replacement cannot preserve the
strength of the original theory, because that theory transcends its evidence. That is, any
evidence that we can in practice gather will always be inconclusive: there will be
possible circumstances in which the evidence is misleading, and things are not really as

33
I have written ‘appearance’ where the text has ‘aspect’. This is simply to avoid a needless change of
terminology.
34
It is also being assumed that we can identify other persons, i.e. can ‘construct’ their bodies and infer that
they have minds like ours. But ‘constructing’ a person’s body is presumably no easier than ‘constructing’ a
table. (Russell always admits that he is presupposing other minds, to legitimize the assumption that there are
appearances other than the ones that he himself has.)
168 K N OW L E D G E

it suggests. If this is right, then Russell’s ‘construction’ of the material world cannot
actually fulfil the task that he sets it.
I return, briefly, to my earlier remark that Russell appears to be ignoring a method of
reasoning that we do often employ, both in our common-sense thinking and in our
more sophisticated scientific theorizing, namely what is called ‘inference to the best
explanation’ (pp. 141–2 above). We would like an explanation for why the various
appearances of the table hang together as they do, and our basic explanatory idea is that
there is a ‘real’ table, which is a fairly permanent occupant of the same space and time
(or space-time) that we inhabit, and which is the cause of these various appearances.
With just a little more sophistication we can give an outline of how these appearances
are caused. For example, the visual appearances are caused by light-rays that travel in
straight lines from the different parts of the table to our eyes, and the science of
geometrical optics will explain why the appearances of the same table vary as they
do from one viewpoint to another. The tactual ‘appearances’ obviously have a different
explanation, but it is still one which involves the same ‘real’ table, in the same place at
the same time, though now it is considered as a more-or-less solid object, able to resist
the pressure applied by our fingers. Even common sense has quite an elaborate theory
of what kind of thing the ‘real’ table is, and this theory gives an outline explanation of
why its various appearances are as they are. A more sophisticated physics will give a
more detailed account of just what kind of thing the ‘real’ table is, and its rationale will
still be that it gives further or deeper explanations of how and why the table has
those effects that common sense ascribes to it. But the justification for all these theories,
even while still at the level of common sense, is that they posit the existence of
something that is not itself an appearance but that does explain why the appearances
are as they are.
That was in fact the position that Russell himself had adopted in PP. In RSP and
KEW he abandons it, and no longer tries to explain the appearances, but takes the view
that appearances—both actual and possible—are the only things that exist (apart,
perhaps, from minds; but that is the topic of the next chapter). But this replacing
theory is not an adequate replacement, even for our ordinary and common-sensical
theory of material objects, because it posits a whole welter of so-called ‘appearances’
that do not actually appear to anyone, but refuses to contemplate the ordinary
explanation of why the actual appearances are as they are, and of why others that
would resemble them in suitable ways would also be possible.
 10
Logical Constructions (ii): Minds

This chapter takes up another of Russell’s ‘logical constructions’, turning from matter
to minds. But in this case there is an extra complication. It is not just that Russell is once
more attempting to replace an inferred entity, the mind, by a construction from
materials that are supposedly closer to experience, but there is also an extra feature:
he is hoping to use the same basis in experience for constructing both matter and minds.
This is the theory called ‘neutral monism’. It is a monism, because it aims to assume just
the one kind of ‘stuff ’ that constitutes every existing thing; and it is neutral as between
matter and mind, because it is the same ‘stuff ’ that constitutes both. The broad
distinction is that some kinds of ‘arrangements’ of this stuff form minds, while other
‘arrangements’ form material bodies. (To express this more accurately, the proposal is
that what we say about minds and bodies can be reconstrued as speaking always of this
neutral stuff, arranged either in the one way or in the other.) As I shall indicate, the
theory can be traced back to Hume, but the version which attracted Russell’s attention
was essentially due to the American philosopher William James.1
Russell did not hold this theory in 1914, when he was elaborating his account of
matter as a construction from sensibilia. In fact he argued against James’s view of the
mind in an article ‘On the Nature of Acquaintance’ (henceforth NA) which was
published in 1914,2 and I begin with some account of those arguments. But later he
came to think that they could be countered, and that there was more to be said in
favour of neutral monism than he had at first thought. There is a foretaste of the change
in his 1918 lectures on ‘The Philosophy of Logical Atomism’ (PLA), which reveal that
he has now come to think that minds ought to be logical constructions, as bodies are.
But his new theory is not clearly stated until his article of 1919 ‘On Propositions: What
they Are and How they Mean’ (henceforth Propns). It is then articulated in much
greater detail in his book of 1921, The Analysis of Mind (henceforth AMind ). I shall say
something of each of these treatments.

1
As his chief sources he cites not only James, Essays in Radical Empiricism [1912] but also Ernst Mach,
The Analysis of the Sensations [1886]. As he notes, James approaches the topic from a background in
psychology, while Mach was a physicist.
2
The three essays which make up NA were the first three chapters of the book Theory of Knowledge that he
was working on in 1913. Although he gave up working on that book in June 1913, he evidently thought that
its first six chapters were worth rescuing, and they were published (perhaps after some revision) in successive
issues of The Monist for 1914 and 1915.
170 K N OW L E D G E

In pursuing this subject now—and especially in adding Chapter 11 as a coda on its

final version in The Analysis of Matter [1927a]—I am for the first time clearly stepping
out of historical sequence. For the following three chapters will concern Russell’s
metaphysics, especially as elaborated in PLA [1918b], and this precedes his partial
acceptance of neutral monism. But it is convenient to consider now this further
example of his practice of ‘the supreme maxim in scientific philosophizing’, namely
the elaboration of so-called ‘logical constructions’.

1 Acquaintance with oneself

The theory of neutral monism contains two main claims, which we may take separate-
ly. The first is that the mind should be treated as a logical construction from items that
are closer to experience, and then the second adds that these items should be taken to
be sensations, i.e. the same items as Russell had already used in his construction of
material bodies. When he first comes to consider this theory, and to reject it in NA, it is
his objections to the first claim that are of most interest. By way of background, it will
be useful to begin with a brief reminder of the situation as Hume saw it.
Descartes had espoused a thoroughgoing dualism of minds and bodies. Bodies are
material substances, and the essence of a material substance is to be extended in space.
Material substances do not think. By contrast, minds are immaterial substances, and the
essence of an immaterial substance is to be a thinking thing. Immaterial substances are
not extended in space. Locke did not exactly reject this dualism, but he did introduce a
number of important qualifications. For our purposes, the most significant is his
account of personal identity. He is ready to concede that there may be immaterial
substances, to do the various things that are done by a mind, but he does insist that
the identity of the person does not depend upon the identity of any such substance.3
His position is that, for all we know, it may be that the same person is constituted by
a succession of different immaterial substances, one after another. Equally, it may
perhaps be that the same immaterial substance animates a succession of different people,
one after the other, as the theory of reincarnation may be taken to imply. His point
is that what it is for a person to remain the same person, over a period of time, has
nothing to do with the preservation of any such substance. What matters is that the
series of mental events that constitute that person should be a suitably connected series,
with each one appropriately related to the others. Locke himself then suggested only
one kind of appropriate relation, namely that the later stages of a person should re-
member the earlier stages. As Hume later saw, there are also several other relations that
will hold between the earlier and the later stages of the same person, and that are
not irrelevant to its being the same person all through. But that is a matter of detail,

3
Note that Descartes and Locke and Hume have no hesitation in identifying a person with his mind.
Russell too accepts this identification.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 171

which we need not explore here. The important idea is that mental episodes will count
as occurring in the same mind if and only if they are suitably related to one another.
This theory was modified by Hume, who roundly denied the existence of a
Cartesian immaterial substance, such as Locke had appeared to accept even though
he had denied it an important part of its role. Hume’s main reason was just that he
could not observe any such substance as could be called himself. In a well-known
passage he says:
For my part, when I enter most intimately into what I call myself, I always stumble upon some
particular perception or other, of heat or cold, light or shade, love or hatred, pain or pleasure.
I never can catch myself at any time without a perception, and never can observe anything but the
perception. (Treatise, p. 252)

So he concludes that a person is ‘nothing but a bundle or collection of different

perceptions’ (ibid.), namely perceptions that are suitably related to one another.
Russell is, of course, very well aware of this background when he first comes to
consider the problem in his Problems of Philosophy, and he remains somewhat agnostic:
The question whether we are also acquainted with our bare selves, as opposed to particular
thoughts and feelings, is a very difficult one, upon which it would be rash to speak positively.
When we try to look into ourselves we always seem to come upon some particular thought or
feeling, and not upon the ‘I’ which has the thought or feeling. Nevertheless there are some
reasons for thinking that we are acquainted with the ‘I’, though the acquaintance is hard
to disentangle from other things. (PP, p. 50)4

His reason for doubting Hume’s conclusion is that I am capable of what may be called a
second-level awareness, i.e. an awareness that I am aware of something. In terms of his
theory of acquaintance, this is to say that, when I am acquainted with a certain sense-
datum, then I can also be acquainted with that very fact, i.e. the fact that I am
acquainted with the sense-datum. And then I can apparently analyse that fact as the
holding of a relation between myself and the sense-datum. But how can I do this if
I am not acquainted with myself ? That is the question posed, but not answered, in
PP (pp. 50–1).
In NA Russell proposes a different solution. He has always claimed that acquain-
tance is a two-termed relation, holding between a self (or subject) and an object such
as a sense-datum. But he now adds that my second-level awareness may be regarded as
an acquaintance with the fact that something is acquainted with the relevant sense-
datum, but that (for Hume’s reasons) we have to admit that the nature of that
‘something’ is unknown to me. I cannot, for example, tell whether or not it is the
same thing as is acquainted with other sense-data that I am aware of being acquainted

4
The position adopted in KAD is similar. In a note which he added later, when KAD was reprinted in his
Mysticism and Logic [1918a], he implies that when he was writing KAD he had thought that he was acquainted
with himself, though later he revised that opinion (p. 224n.).
172 K N OW L E D G E

with on other occasions, or even whether it is the same thing as is now acquainted at
the second level with the original first-level fact of acquaintance (pp. 164–7). But
now, if we may bring in a claim about acquaintance that goes back to the argument of
On Denoting, it is an essential feature of acquaintance that I cannot be in doubt about
the identity of what I am acquainted with (pp. 117–18 above), so it would have to
follow that I am not acquainted with the relevant ‘something’, since I cannot be sure
of recognizing that same thing again in another context.5 This particular line of
argument is not actually given in NA, which cites as its ground only Hume’s inability
to perceive himself. But in any case it concludes in the same way: I am not acquainted
with the ‘something’ that I know is acquainted with my sense-datum. Yet, NA still
insists, acquaintance is a relation, so there must be something that is acquainted with
the sense-datum. Russell’s only recourse is therefore to say that I must know this thing
by description, i.e. I know it only as ‘the thing—whatever it may be—that is
acquainted with this present sense-datum’. That is how NA concludes, and it is
perhaps enough to show how one can both accept Hume’s observation and retain
the view that acquaintance is a two-term relation between a self and a sense-datum.
But a resulting problem, which Russell does not address, is this: what ground does one
have for supposing that acquaintance is a two-term relation, if one of its two terms is
never experienced?6
In NA Russell felt that he must retain this view, because of a problem with the
alternative idea that a self is just a series of subjectless events. All attempts to build a self
from such a series have found it important to consider the relations between different
members of the series, for it is these relations that hold the different members of the
series together, to form a substitute for a single persisting substance. But it then follows
that there could not be a mental event which bears no such relation to other mental
events, for in that case there would be nothing about it which qualified it as ‘mental’,
i.e. which made it an occurrence in a mind. This point is particularly relevant to the
version of the mind that neutral monism proposes, for it wishes to claim that the very
same basic ‘stuff ’ constitutes a mind when arranged in one way, and a material body
when arranged in another. So ‘arrangements’ are all-important, and of course that is
just a matter of what relations hold between the items said to be arranged. But the same
point would apply to any other feasible way of trying to ‘construct’ a mind from some

5
Donnellan [1990] understands a genuine name to be one that refers without describing, and takes it to be
the mark of such a name that one who uses it cannot rationally doubt the existence of what it names. He
argues on this ground that ‘I’ is a genuine name, whether or not I count as being acquainted with myself. But
Russell would wish to bring in the further mark that one who can use both ‘a’ and ‘b’ as genuine names
cannot be in doubt over whether ‘a=b’ is true, and on this ground he would disallow ‘I’. However we may
agree with Donnellan that there is no general and compelling reason to suppose that one must be acquainted
with anything that one can genuinely name.
6
I add that Russell does not ask how (or whether) I know that there is only one ‘something’ that is
acquainted with this sense-datum, or whether I am right in assuming (as I do) that it is in fact the same
‘something’ that is acquainted with all the sense-data that I am ever aware of.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 173

series of independently identifiable events (for example Locke’s way, or Hume’s way).
On this approach it must apparently follow that there could not be a mind which
contains just one such event, not related to any other.
But in NA Russell protests that this will rule out what must be admitted to be a
genuine possibility:
I cannot think that the difference between my seeing [a] patch of red, and the patch of red being
there unseen, consists in the presence or absence of relations between the patch of red and other
objects of the same kind. It seem to me possible to imagine a mind existing for only a fraction
of a second, seeing the red, and ceasing to exist before having any other experience. But such
a supposition ought, on [this] theory, to be not merely improbable but meaningless. According
to [it], things become parts of my experience in virtue of certain relations to each other; if there
were not a system of interrelated things experienced by me, there could not be one thing
experienced by me. (p. 148)

Russell calls this the ‘main’ objection to the theory of neutral monism, and he gives it as
his first objection. He also recurs to it at the end of the section:
Between (say) a colour seen and the same colour not seen, there seems to be a difference not
consisting in relations to other colours, or to other objects of experience, or to the nervous
system, but in some way more immediate, more intimate, more intuitively evident. If neutral
monism were true, a mind which had only one experience would be a logical impossibility, since
a thing is only mental in virtue of its external relations. (p. 158)

It is clear that the same objection would apply to all ways of ‘constructing’ a mind from
events which do not, in themselves, already presuppose minds.
The discussion in NA contains several other objections to the neutral monism put
forward by Mach and James and some others, but they apply only to that theory in
particular. According to it, the relevant events from which both minds and bodies are
to be constructed are all sensations, and Russell complains that these by themselves are
not adequate to do the work intended of them. In fact his complaint is even more
narrow, for it is just that the proposals so far put forward by neutral monists, to explain
beliefs as sensations of a certain kind, will not do (pp. 149–58). These are objections of
detail, and in view of future developments we need not linger over them, save perhaps
to note how limited in scope they are. For Russell himself has said:
Experiencing is only one, though perhaps the most characteristic and comprehensive, of the
things that happen in the mental world. Judging, feeling, desiring, willing, though they presup-
pose experience, are themselves different from it. (p. 161)

A full theory of neutral monism built upon the sole basis of sensations, would have to
show how all mental events are to be explained in terms of sensations. I shall come back
to this requirement when we have seen how Russell himself comes to embrace a form
of neutral monism.
Before we come to this development, I add one further remark to this discussion of
NA . The piece ends with a final objection to neutral monism which Russell evidently
174 K N OW L E D G E

regards as important, but which remains obscure to me.7 He has now given his own
explanation of how, when I am aware of a sense-datum, I can also be aware that there is
something which is aware of this sense-datum. The datum can be named ‘this’, which
I am aware of as occurring ‘now’, which gives me a way of explaining ‘I’ as ‘what is
aware of this now’. He therefore feels entitled to ‘retort on neutral monism with the
demand that it should produce an account of “this” and “I” and “now” ’ (p. 169). But
his reason for supposing that neutral monism could not meet this demand remains
opaque to me. One may speculate that his thought is that these words (which Russell
somewhat oddly calls ‘emphatic particulars’) can be understood only by a conscious
being, and that neutral monism cannot explain consciousness. But this speculation goes
some way beyond anything that he actually says, so I leave it unexplored. When
Russell himself comes to embrace neutral monism, he does not say anything which is a
clear response to this supposed difficulty.8

2 A change of view
Some time after NA was published in 1914, Russell began to have doubts. He was still
in a state of doubt when he wrote the final lecture of his PLA in 1918. We have already
noted (on p. 166 above) how he there supports the earlier claim that material bodies
are logical constructions from experiences. He does so by urging that there must be
some experiential content to our talk of the same material body (e.g. ‘the same desk’)
and proposing that we should take this experiential content to be the whole of its
meaning (pp. 272–3). He now goes on to apply the same idea to our talk of minds
(persons) on pp. 276–7. Admittedly, in this case there is not only the experiential
evidence that is available to an onlooker, but also the evidence that is available to the
person himself by introspection, so the experiential content of ‘the same person’ will be
more complex than is that of ‘the same desk’. But in both cases Russell claims that there
must be such content, and he draws the same moral, that in both cases a logical
construction from what we experience must be available. It does not yet follow, as
neutral monism wishes, that the experiential basis will be the same for the two cases,
but Russell makes it clear that he is now looking with more favour on this idea. He still
admits that there are difficulties, in particular over the analysis of belief, but he is now
thinking that they may turn out to be soluble (pp. 278–80).9
But a year later he is apparently confident that he now has a solution, for in a brief
section (pp. 305–7) of his paper Propns, published in 1919, he does accept a version of

7
I add that Sainsbury [1979] discusses this objection on p. 265, and he too finds it obscure. I am not aware
of anyone who has made good sense of it.
8
That ‘emphatic particulars’ create a difficulty for neutral monism is a point that is repeated in the
discussion of PLA at p. 280. It is there described as a problem for ‘the view which does not distinguish
between a particular and experiencing that particular’. I guess that Russell is here assuming that all experience
is conscious experience.
9
The supposed difficulty over ‘emphatic particulars’ is also mentioned. See previous footnote.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 175

neutral monism, but with an addition of his own. In this paper he gives no reason for
rejecting his earlier discussion in NA, but simply reminds us of Hume’s claim that the
supposed subject of a mental act is unobservable:
I have to confess that the theory which analyses a presentation into act and object no longer
satisfies me.10 The act, or subject, is schematically convenient, but not empirically discoverable. It
seems to serve the same sort of purpose as is served by points and instants, by numbers and
particles and the rest of the apparatus of mathematics. All these things have to be constructed, not
postulated: they are not of the stuff of the world, but assemblages which it is convenient to be
able to designate as if they were single things. The same seems to be true of the subject, and I am
at a loss to discover any actual phenomenon which could be called an ‘act’, and could be regarded
as a constituent of a presentation. (p. 305)

That is all that we have, in Propns, to explain why almost all of the discussion of NA is
now abandoned. One may fairly say that Russell here announces a change of view, but
gives no new arguments for it.
There is also a further change of view that is announced but not justified in this
paper. Just as he now thinks that there is no subject to a supposed act of acquaintance,
so he now thinks that there is no object either. That is, there is no such object as a
sense-datum was supposed to be, something presented to me at a certain time, but
capable in theory of also being presented to others, or of not being presented at all. So
all that we now have is an event which is a sensation, but no subject that senses and no
object that is sensed. Again, we have no explanation for the change of view, and are left
to speculate on whether the thought that there are no objects of acquaintance was
either the cause or the effect of the thought that equally there are no subjects. The
effect of this change of view has been discussed in the previous chapter, where we saw
that in practice it makes very little difference to Russell’s attempt to construct material
objects from sensations.
But the construction of minds is a new endeavour, and here Russell departs from the
official theory of neutral monism. For he introduces the idea that mental images should
be distinguished from sensations, and should be included in the material from which
minds are to be constructed. This departure is surely motivated by his original problem
over the analysis of belief, because he now thinks that this can be given if images are
included. His claim is that ‘sensations and images, suitably related, are a sufficient stuff
out of which to compose beliefs’ (p. 307). This claim does not conform to the
requirements of neutral monism, for although sensations may be ‘neutral’ between
minds and bodies, images clearly are not; they are found only in minds. One may
also note, in the other direction, that material bodies too are constructed partly from

10
Apparently Russell is taking it for granted that something can be called an act only if there is a subject
that performs that act. The terminology is perhaps influenced by his earlier work on Meinong, who had
analysed a presentation into act and content and object (cf. Russell [1904: 24]). He has used the word ‘act’ in
this way before, e.g. [1913a: 184]. He has always regarded ‘x is presented to y’ and ‘y is acquainted with x’ as
two ways of saying the same thing. See e.g. KAD, pp. 209–10.
176 K N OW L E D G E

non-neutral materials. For sensations that are sensed have a part to play both in the
construction of minds and in the construction of bodies, but sensations that are not
sensed (i.e. mere sensibilia) do not occur in minds. In RSP and KEW Russell has
claimed that whether sensed or unsensed still all sensibilia are of the same kind as one
another. He will also claim that sensations and mental images are in a different way of
the same kind as one another, for images are directly copied from sensations, or are put
together from such copies, as Hume had claimed.
Let us sum up on the paper Propns, so far as that affects our present topic.11 While it
does tell us that Russell has changed his mind in several ways, it makes no real attempt
to justify these changes, and we can only feel that further elaboration is needed. It also
tells us that Russell now has a theory which analyses belief in terms of sensations and
images, and this theory is very briefly outlined on pp. 307–9, but again we feel that
more detail would be highly desirable.12 The second wish is largely met by Russell’s
next major work, his book The Analysis of Mind, published in 1921, and I now move on
to consider this. But the first wish is largely disappointed. In particular, we do not find
either in Propns or in AMind any clear response to either of the main objections that in
NA Russell had treated as serious obstacles to any logical construction of minds. One is
confident of what he now thinks about the first of these objections, namely the point
that such a construction must make it impossible for there to be a mind which has just
one sensation and no more, or just a few isolated sensations which bear no suitable
relation to one another. He would now agree that this is indeed impossible. For what is
envisaged is the existence of just one sensibile, or just a few isolated sensibilia, which are
not to be regarded as in any way sense-data, i.e. as given, for by hypothesis there is no
suitable arrangement of mental occurrences of which they form a part. There is,
therefore, nothing to which they could be counted as given. But one cannot be
confident of what he now thinks about the alleged objection concerning ‘this’ and
‘now’ and ‘I’, because one remains at a loss over just what this objection was supposed
to be. However, it may be that the account of consciousness that we do eventually
reach at the end of AMind is supposed to contain an answer.

3 The nature of minds

The opening chapter of The Analysis of Mind reiterates Russell’s new view that there is
no such thing as a self. It gives us some more background to his thinking, but does not
really provide any new argument. The extra background is that, following William
James, Russell tends to identify the positing of a (substantial) self with a stress on the
importance of consciousness. In fact on pp. 22–5 he quotes several pages from James

11
Its new account of what propositions are will be considered further in Chapter 12.
12
It would be better to say ‘ . . . in terms of sensings and imagings’. For when the object of an event of
sensing, i.e. the thing sensed, is abolished, then the image can only be a copy of the event itself, and not a copy
of the thing sensed. But I shall follow Russell in generally ignoring this point in what follows.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 177

which make this connection. Evidently James thought of his opponents as taking
consciousness to be the distinguishing mark of a mind, and indeed to be ‘the entity’
that is the mind, and to this he very fairly objects that much of what happens in a mind
is in fact unconscious. This is a point that Russell will also stress in his following
chapters, and we may surely grant it. However the correct moral to draw is only that
mind and consciousness cannot be identified, for although we are sometimes conscious
of our mental activities at other times we are not. But the conclusion that there are
in fact no such things as minds are supposed to be cannot be said to follow.
There is a generalization of this line of thought which Russell makes explicit in the
next chapter (pp. 41–4), and which evidently underlies much of his discussion else-
where: there is no clear boundary between creatures that do have minds and those that
do not. Almost all of us are prepared to attribute some kind of ‘mental life’ to the higher
mammals (e.g. horses, dogs, apes, and so on). For example, they certainly perceive things,
but we also find it very natural to credit them with desires, beliefs, and emotions (such as
fear), since this is the obvious way to explain their behaviour. But as one moves ‘lower’
on the scale of animal life the behaviour that so strongly suggests these mental explana-
tions diminishes, and eventually it just disappears altogether (as, e.g., with shellfish). But
there are many intermediate cases where we do not know whether to regard such mental
terminology as literally correct or as more in the nature of a metaphor.
One may draw from this the moral that there is a vagueness in our concept of a
mind, as (for example) there is a vagueness in the concept of being an adult. Children
become more adult, as they grow and develop, but there is no particular point at which,
for the first time, they have become adults. Similarly, animals of different species have
more or less mentality, but it is a mistake to think that at some point in the scale there
enters a new entity of a quite different kind, which is now responsible for performing
those mental activities which previously had been attributed to the whole animal.
A mind is not that kind of thing, and this perhaps explains why it is a mistake to hope to
catch sight of it, if not in others then at least in oneself. At any rate, this way of trying to
uphold the moral that Hume drew is near to the surface of Russell’s discussion, for he is
constantly drawing attention to how what human beings do is also done by other
animals, even if the point is not explicitly invoked as a justification for his denial of the
mind as an entity in its own right.
In consequence of the denial, the suggestion of the ‘I’ in ‘I think’ is disallowed, and
the claim is rephrased as ‘there is thought in me’, where the word ‘me’ designates the
whole human being and not some special part called the mind. Similarly with
‘I perceive’, ‘I desire’, ‘I believe’, and so on (p. 18). Consequently, the fact that I can
be aware of my own awareness has to be thought of, not as me being aware of some
object that is ‘my mind’ or ‘my self ’ and of what it is aware of, but simply as there being
(in me) an awareness of an awareness.13 However, Russell’s rather unexpected account

13
This responds to a problem raised by Chisholm [1974: 53].
178 K N OW L E D G E

of what this awareness is will not be given until the end of the book, so for the present
I postpone it.
In the earlier paper Propns Russell had coupled together the two claims that in a
sensation there is no (observable) thing that does the sensing and that there is no
(observable) object that is sensed. He does the same in AMind. We are initially
unwilling to accept either claim, but perhaps especially the second, since we so
constantly talk of objects being sensed (or being observed, or perceived, or seen, or
touched, and so on). But the ‘objects’ that we normally have in mind are thought of as
existing independently of being sensed—e.g. a tomato, or simply a red colour—and
this is not what Russell is talking of. If I claim to see a tomato, or a patch of red colour,
then ordinarily I am talking of a public object, available for others to see too. But in
Russell’s view the claim to see such a thing runs together the sensation itself and all
manner of associated expectations about what other sensations will be available, to me
and to others. For the object that is publicly available is a logical construction, built
from all such sensations, and not itself the object of any one such sensation. If we do
confine attention just to what is happening in the one sensation itself, and set aside all
the further expectations that go with it, then his claim is that one cannot discern
anything that is a ‘constituent’ of that act of sensing distinct from the sensing itself. So to
avoid a misleading form of expression we should not talk of sensing an object, e.g. a
tomato-like sense-datum, but perhaps of sensing tomato-ly. The idea is that this
‘adverbial’ account describes the kind of sensing that is going on, but without transfer-
ring the description to a supposed object being sensed. Russell does not himself use this
‘adverbial’ terminology, or anything like it, but it seems to me that he should welcome
it. However the idea certainly has its problems.
For example, if I am seeing a tomato then on the previous theory I will be presented
with a datum that is both red and (approximately) round. It is the same thing sensed that
has both these qualities, of being red and of being round. But how is this point to be
represented if we no longer suppose that there is a ‘thing’ that is being sensed? It is not
good enough to say merely that I am sensing both red-ly and round-ly, for that is
compatible with me sensing two different things, one of which is red while the other is
round. Perhaps Russell could find a way of meeting this and other problems with his
new theory, but in fact he seems not to see that there are such problems. At any rate, he
does not discuss them, and so I too will leave this debate to others.14
As we have noted, Russell couples his denial of an object sensed together with his
denial of a subject that senses, as if the one necessitates the other. He says:
If we are to avoid a perfectly gratuitous assumption, we must dispense with the subject as one of
the actual ingredients of the world. But when we do this, the possibility of distinguishing the

14
As a proponent of the ‘adverbial’ theory one might choose Chisholm [1957: chapters 8–10]. (Some of
the ideas there are developed further in Chisholm [1966: chapter 6].) A number of objections are raised in
Jackson [1977], especially chapters 3–4.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 179

sensation from the sense-datum vanishes; at least I see no way of preserving the distinction.
(AMind, p. 142)

However, there is not really any connection between the two doctrines. For one may
surely attempt to abolish the subject of a sensation while still retaining its object, or one
may equally set out to abolish the object of a sensation while retaining its subject.
In each case the relevant arguments seem quite independent of one another.

4 The construction of minds

Let us grant to Russell his minimalized ontology. First, there are such things as the
events of sensation. Next, there are also events which are (supposedly) just like
sensations except that they are not sensed; these are sensibilia. The material world is
then to be constructed from sensations and sensibilia, in the way explained by RSP and
KEW. (Russell recapitulates the main lines of the construction in chapter 5 of AMind.)
Finally, there are also events which are like sensations except that they have a different
causal origin, namely events of having images. A mind is then a construction from
certain sensings and imagings, namely all those that we credit to ‘the same mind’. And
which are those? Russell is still tempted by the ideas outlined in KEW, i.e. that a
‘perspective’ is a collection of all the sensations occurring at one place at a given time,
and that a ‘biography’ is then a temporal series of perspectives. But now he sees a
problem, and says:
This [idea] would not really help us, since a ‘place’ has not yet been defined. Can we define what
is meant by saying that two aspects [i.e. sensations] are ‘in the same place’, without introducing
anything beyond the laws of perspective and dynamics? I do not feel sure whether it is possible to
frame such a definition or not; accordingly I shall not assume that it is possible, but shall seek
other characteristics by which a perspective or biography may be defined. (p. 127)

We may note that this scepticism over ‘the same place’ also creates a problem for the
construction of material bodies (cf. pp.161–3 above), but we may here set that aside in
order to concentrate upon minds.
Russell proposes to avoid the problem by changing from places to times, with the
rider that what we are talking of is an individual’s ‘private time’. The basic idea is that
sensations and images are experienced as having temporal relations to one another
when and only when they belong to the same mind. A single perspective, then, will be
a set of all sensations and images experienced as simultaneous with one another, and
two perspectives will belong to the same biography if one contains sensations or images
experienced as earlier or later than those of the other.15 But at the same time as
proposing this definition Russell also seems to be proposing another, which no longer
speaks of sensations which are experienced as simultaneous, but simply of those that are

15
This idea has been anticipated in ‘The Ultimate Constituents of Matter’ [1915: 141].
180 K N OW L E D G E

simultaneous. There is the obvious objection that two sensations may be simultaneous
even when they belong to different people (i.e. different perspectives), but Russell’s
response is to deny this. For according to the theory of relativity the truth is that each
individual thing has its own ‘proper time’, and there is no one ‘universal time’ that can
be constructed from all the many different ‘proper times’. He sums up:
The relations of simultaneity and succession are known to us in our own experience; they may be
analysable, but that does not affect their suitability for defining perspectives and biographies.
Such time-relations as can be constructed between events in different biographies are of a
different kind: they are not experienced, and are merely logical, being designed to afford
convenient ways of stating the correlations between different biographies. (pp. 128–9)

What are we to think of this?

We may surely set aside Russell’s appeal to the theory of relativity, for (a) that is a
recently discovered scientific theory which—like all scientific theories—may reason-
ably be doubted; (b) it does not in practice have any effect upon our ordinary under-
standing of temporal relations as holding universally;16 and (c) we cannot suppose that
the notion of ‘one person’ could not be understood before that theory was discovered.
So for present purposes we should concentrate on the other feature of Russell’s
elucidation, namely the idea that the relevant temporal relations are those that can
be ‘directly experienced’.17
Now presumably two sensations of the same mind may be simultaneous without
being experienced as simultaneous. That is, the thought that they are simultaneous need
not occur, especially if one or both of the sensations occur unconsciously, so that the
subject is not aware of them both. (Russell is clearly committed to the possibility of
unconscious sensations.) More obviously, one sensation may be later than another, in
the same mind, without being experienced as later, perhaps because by the time that the
later one occurs the earlier has been quite forgotten, so that there never is any thought
that relates them. But Russell’s idea is that, if the sensations do in fact occur in the same
mind, then they do have a temporal relation that is of the right kind to be directly
experienced, even if it is not in fact so experienced. And it is temporal relations of that
kind that tie together the bundles of separate sensings and imaginings which constitute
individual minds.

16
The theory has a noticeable effect on temporal relations only when we are considering events that are
way beyond the scale of ordinary experience, e.g. events in the biographies of two observers who are moving
with a velocity (relative to one another) that is near to the speed of light.
17
Russell aims to include unperceived ‘perspectives’ and ‘biographies’ for all continuous space-time
tracks, not only those that are occupied by living things. He says that the temporal relations which define
perceived perspectives and biographies do apply also ‘to particulars which are not sensations. They [i.e. these
relations] are actually required for the theory of relativity, if we are to give a philosophical explanation of
what is meant by “local time” in that theory’ (p. 128). If we wish to pursue this line of thought then we must
stress the mere possibility of perceiving the temporal relations between what are, in themselves, only possible
sensations.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 181

One may well be sceptical of this idea that there are different kinds of simultaneity,
or of occurring earlier or later than. The modern treatment of time in the theory of
relativity may perhaps give it a little encouragement, but it is quite absent from our
ordinary ways of thinking. It seems better, then, to replace the idea by what Russell has
said in partial explanation: the temporal relations in question are those that are directly
experienced, or at least could be directly experienced, given sufficient attentiveness. To
spell this out a little: Russell has always supposed that I can be aware that two or more
of my experiences occur simultaneously, i.e. they are ‘experienced together’ (e.g. NA,
pp. 165–6). A single perspective is then a maximal set of all the experiences that are or
could be experienced together. When two experiences of mine are separated by a wide
temporal gap, it is not clear that I can—even in principle—experience their temporal
relation. Indeed, Russell usually holds that direct experience of a relation requires
one to be simultaneously acquainted with both its terms. If so, then one can experience
one event as being earlier than another only when they are so close in time that both
are part of some one specious present (as explained above, on p. 119). So let us say that
one total perspective is ‘directly earlier’ than another, when the one contains an
experience that is or could be directly experienced as earlier than some experience
in the other. But then we can tie together these very short temporal links between one
perspective and another by taking the ancestral of this relationship (as is envisaged at
NA, p. 133). That is, perspectives pi and pj belong to the same biography if and only if
both are members of a (finite) chain of perspectives, p1, p2, . . . pn, each of which is
directly earlier than its successor. But the basic notion has to be the relation between
experiences when one is or could be experienced as earlier than the other. I shall come
back to this point shortly.
Just as Russell allowed for unsensed sensibilia, which belong to unperceived per-
spectives, so he also intends to allow for unperceived biographies. In the construction
of material objects there was a need for such unperceived perspectives, since there are
material objects which are never perceived. There is no similar need for unperceived
biographies in the construction of minds, for we do not suppose that there are minds
which are capable of having sensations but which never do have any. However we do
think that minds exist without sensations for limited periods, during sleep, and
something needs to be said about this. The first observation to make here is that
Russell does not claim that every biography is the biography of some mind, but only
those in which each perspective is from a position within a living and working animal
brain—a brain which (in the normal case) will be connected in the usual way with the
sense-organs of that animal. So the sensations that occur in that biography will be
caused by suitable happenings in the connected sense-organs. More generally, Russell
assumes that there cannot be minds without bodies, and that what happens in a body
will have an effect on what happens in its mind.18 He further assumes that, in the

18
By pressing on the idea of ‘mnemic causation’, which I shall introduce in a moment, one might try to
relax this assumption. But Russell makes no such attempt, and nor shall I.
182 K N OW L E D G E

normal case, the identity of the mind (over time) will be the same as the identity of its
body. But, when a body goes to sleep and then wakes up again, what leads us to say that
it still has the same mind is not just that it is still the same body. It is, rather, the fact that
its mental states after waking have a certain kind of continuity with its mental states
before sleeping. Russell does not say very much about what kind of continuity this is,
though no doubt he is assuming that it will involve various similarities. But the point
that he concentrates upon is a generalization of what Locke had said about memories,
for genuine memories are causally linked to the earlier experiences that they are
memories of. Russell extends this idea to cover any case where a present mental
occurrence is due, at least in part, to a previous experience, and he calls this a case of
‘mnemic causation’. In chapter 4 of AMind he gives a large number of cases of such
causation, showing how very many of our present thoughts and actions do depend
upon previous experiences, and would not have been the same without those experi-
ences. I pass over all thesedetailed examples.19
We may sum up his account of minds in this way. A single mind will be associated
with a single body throughout its existence. It will also be associated with a series of
perspectives, which we say are its experiences. If we insist upon the existence of
unperceived perspectives, we may if we wish say that a single mind has throughout
its existence what Russell defines as a single biography. But in that case the biography
will include periods during which no sensation is conscious, e.g. when the subject is
asleep. (In dreams there will be what Russell counts as sensations, though they are not
caused in the normal way by the external stimulation of the body’s sense-organs.)
However it is no doubt simpler to say that the history of that mind’s experiences is
given by a string of what Russell’s definition must count as different biographies, each
corresponding to its successive periods of consciousness, and with gaps of unconscious-
ness between them.20 During the waking periods, the episodes in a single biography are
taken to be linked to one another by experienced temporal relations. These waking
periods are then linked to one another (a) by certain similarities, but more importantly
(b) by links of mnemic causation. That is an outline of Russell’s account of what a mind
is (and it applies, presumably, not just to human minds but to the minds of any sentient
animals). What are we to think of it?
It is at once apparent that Russell cannot avoid making use of the notion of
possibility. This emerges both in what he says about the experienced temporal relations
that are supposed to unify a single biography, and in his reliance on what he calls
mnemic causation. As I have already noted, the first has to link together elements that

19
Russell is clearly assuming that his mnemic causation does not rely upon any causal chain that strays
outside the body of the animal in question. He is officially agnostic on whether there will always be a causal
chain existing within the body, or whether the initial experience ‘acts at a (temporal) distance’, by causing
subsequent thoughts or behaviour without there being any continuously existing causal chain that links the
two. But he evidently thinks that such ‘action at a distance’ is not very probable.
20
A fuller account than I wish to give would say something here about ‘semi-conscious’ states, between
those that are conscious and those that are unconscious.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 183

could be perceived as simultaneous, or as earlier or later than one another, and the
second must also invoke the possibility of mnemic causation. We can see this clearly if
we first consider just Locke’s example of memory: an experience of mine is something
that I could remember at a later date, though it may well happen that I never do
remember it. That does not prevent it from being one of my experiences. More
generally, my experiences at some time yesterday could affect my thoughts or actions
of today in many different ways, but it may just happen that none of them actually do.
Perhaps we may accept that it is the possibility of links of such a sort that establishes what
is to count as the same mind over time,21 but one is more likely to want to ask: what is
it that grounds these possibilities? If there is an answer here, it can only be one that takes
us in a circle: it is because two events occur in the same mind that we want to say that
their temporal relation is one that could be experienced, and it is for the same reason that
there could be a mnemic relation between the two. But if we have to explain that it is
because the same mind is in question that these relations are possible, then clearly the
account is moving in a circle.
As with bodies, so with minds. We do in fact understand what is meant by saying of
two sensations or sensibilia that they are appearances of the same material body; and we
do in fact understand what is meant by saying of two sensations or images that they
occur in the same mind. So we do in practice understand what Russell wishes to use as
his basic material for a ‘logical construction’ of bodies and of minds. But the objection is
that our ordinary understanding seems to be founded on our ordinary view of bodies
and minds as inferred entities, postulated as the best way of explaining the observed
appearances and the familiar behaviour, whereas Russell’s own aim is to avoid any such
inference or postulation. His problem, then, is to introduce his starting point in a way
that does not depend upon such assumptions. But this means that in the crucial cases he
has to take certain possibilities as ultimate and irreducible. Whether this procedure is
metaphysically respectable is a question that I must now leave for the reader to
consider.

5 The analysis of belief

The Analysis of Mind contributes rather more to Russell’s construction of minds than we
find in his analogous construction of bodies. He never asks how, if a material body is
merely a collection of actual and possible sensations, it can come to have all those
properties that common sense ascribes to it, and that physicists elaborate upon. But he
does ask the corresponding question for minds. If a mind is really nothing but a
collection of events of sensing and imaging, how can it contain such things as beliefs,
desires, emotions, and so on? In AMind Russell does try to give explanations. In all the

21
A problem which is highly relevant, and which I cannot discuss here, is this: we have mentioned many
relations which are expected to hold if and only if it is the same mind that we are talking of. But what if some
of them do hold and some do not? This is a central issue in modern discussions of personal identity.
184 K N OW L E D G E

cases which he discusses, except for ‘pure’ sensations, belief has some role to play, and
its analysis is therefore of central importance. This is given in Chapter 12, but it draws
upon some ideas about meaning which have occupied the two preceding chapters,
so I here give a very brief account of these. The topic will be taken up in more detail
in sections 6–7 of Chapter 12.
In chapter 10 Russell considers the meaning of single words, and he restricts
attention to words for quite familiar objects, such as ‘box’ and ‘dog’. They are words
which, in Russell’s way of thinking, ‘have meaning in isolation’, and so can be used in
isolation from other words. In each case the meaning may be either some particular
box or dog, or boxes or dogs in general.22 When a child utters the word ‘box’, this
could perhaps be spelt out more fully as ‘That is a box’ or ‘Give me the box’ or ‘I want
my box’. The first two are to be thought of as uttered in the presence of a box, and as
learnt reactions to it, but the last as uttered in its absence. So in the last case it may be
tempting to suppose that the child has an image of the box in his mind, and that it is this
image that prompts the word, rather than the box itself. In fact Russell recognizes that
this suggestion may be doubted, since whatever it is that explains the occurrence of
the image in that situation, might instead be used to give a direct explanation of the
occurrence of the word, without going via any image. Nevertheless he does claim that
images have an important part to play in the account of how words have meaning, and
he argues—using other examples in addition—that it is primarily the image that means
something, either a particular thing or a general type of thing. Then the meaning of a
word is to be explained as the meaning of the image with which it is associated. I shall
consider this theory in sections 6-7 of Chapter 12. For the present, let us just note it,
and pass on from words to sentences, i.e. from images to beliefs. Again, Russell
confines attention to very simple cases.
Beliefs are expressed by those combinations of words that are (declarative) sentences.
Russell calls them ‘word propositions’. He also thinks that there are corresponding
‘image propositions’, which are combinations of the corresponding images, and that
these are in fact more basic. For combinations of images can express propositions in a
more simple and straightforward way than do combinations of words. For that reason
Russell thinks that they can and do occur in animals which have no word language.
Again, I postpone to Chapter 12 a discussion of this view of propositions, and especially
of the problems associated with the so-called ‘image propositions’. All that needs to be
said here is that he thinks of an image proposition as like a word proposition in
incorporating a definite claim, so that it can be said to be either true or false. For the
time being we may think of this claim as made by adding, to the image combination in
question, a thought such as ‘This is how things are’. (In the case of a memory the
addition will be ‘This is how things were’, and in the case of an anticipation ‘This is

22
In chapter 11 Russell discusses the difference between what he used to call words for particulars and
words for universals. His basic answer is that it is just a question of how many different kinds of experiences
prompt the use of the same word.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 185

how things will be’.) The claim is the content of the belief. But in the present chapter I
wish to focus on the other aspect of Russell’s analysis, which asks what it is to believe
such a content, as opposed (for example) to doubting it, or denying it, or merely
contemplating it.
First, we should note that Russell is concerned with a way of understanding ‘belief ’
which is certainly not our usual way. He says:
To begin with, we must distinguish belief as a mere disposition from actual active belief. We speak
as if we always believed that Charles I was executed, but that only means that we are always ready
to believe it when the subject comes up. The phenomenon we are concerned to analyse is the
active belief, not the permanent disposition. (pp. 245–6)

This notion of an ‘actual active’ belief is not the familiar way of thinking of belief, but
perhaps we can grant to Russell that it does make sense, and that it is quite natural to
take it as the primary case for analysis. But notice that it is to be construed not as a
permanent state of the person in question, but as something that occurs in him at
definite times. This may help to explain the analysis that Russell gives.
A belief occurs, then, when some proposition is ‘actively’ before the mind, and
Russell asks what else is happening in the mind at the same time, which would explain
the difference between believing the proposition and (say) doubting it or denying it.
His answer is that there is a certain sort of feeling towards the proposition, which he can
only call a ‘belief-feeling’, and which he does not attempt to describe further except by
also calling it a ‘feeling of assent’. But is there really such a ‘feeling’? (It is certainly not
much like the bodily sensation of feeling an ache or a tickle, or of feeling hot or cold.
But no doubt Russell will reply that it is supposed to be a mental feeling rather than a
bodily feeling.23) So as not to dispute about the word ‘feeling’, let us just say this. Russell
classes this feeling as a kind of sensation (p. 251), and presumably he counts himself as
entitled to include it among the ingredients of a mind because he has begun with the
thought that a mind is a construction from sensations and images. But, as we have seen,
he has also argued that the sensations in question do not have objects (e.g. sense data)
that are distinct from the events of sensing, yet—as he insists—a belief-feeling certainly
has an object distinct from itself, namely the proposition that is believed. So, even if this
alleged ‘feeling’ is counted as a sensation, still it is a new kind of sensation, and not one
of those that we originally accepted as suitable material for the construction of a mind.
And certainly it is not a neutral material, that may equally occur in the construction of a
body. This point may be generalized.

23
Russell says ‘I do not wish to commit myself to any specific analysis of the belief-feeling’ (p. 187n.).
One is reminded of how Hume says something rather similar on this topic in his Treatise, Appx, pp. 623–9,
esp. p. 629.
186 K N OW L E D G E

6 Other mental phenomena

Russell’s claim is that minds may be construed as ‘logical constructions’, built from just
two kinds of materials, namely sensations and images. He admits that images play no
part in the construction of material bodies, but he claims that sensations do, for bodies
are just collections of actual and possible sensations. However very many of the
sensations that figure in those collections that are called minds have no role in the
collections that are called bodies.
When Russell tells us how a material body is to be constructed, he almost always
concentrates on the visual appearances that we regard as appearances of that body. The
full theory no doubt includes also the auditory or olfactory or other sensations that we
regard as due to that particular body, but in these cases there is quite often less
temptation to think of some one material body as the origin of the sensation. Consider,
for instance, the sound of a stone falling into water. Does the hearing of this sound
count as an ingredient of the stone, or of the water, or both? More awkward, perhaps,
is the sensation of feeling warm, or feeling cold, simply because that is how the outside
temperature is. Is that to be regarded as the building material for some external object?
Is it in this way ‘neutral’ as between minds and matter?
But the more obvious cases to mention are those sensations which we think of as the
awareness of our own bodily states, such as feelings of thirst or hunger, of aches or
itches, of freshness or weariness, and so on. These are paradigm sensations, but one’s
first thought is that they do not figure in the construction of any material object.
However, they could perhaps be thought of as among the ingredients from which our
own material bodies are to be built up. For example, is my tooth to be ‘constructed’,
partly at least, from the toothache that it gives me? A similar view would take it that
one of the ingredients from which the knife is constructed is the pain that I feel when it
cuts me, and perhaps someone who is embracing Russell’s attempt at neutral monism
would not find that view absurd. But it certainly becomes awkward when we recall
that objects are to be constructed not only from actual sensations but also from
sensibilia. Are we to build into the knife itself all the possible pains that it would cause
if applied in lots of different ways to lots of different living things? It is surely not what
Russell was originally thinking when he wrote RSP and KEW.
More central, however, are things which we do not ordinarily think of as ‘sensations’
in any literal sense, but which are mental states or events that we can be aware of, for
example desires, or emotions such as fear or anger or love, or choices and decisions, or
of course beliefs. Russell does consider belief in some detail, as we have seen, and he
classes it as a kind of sensation, but one which is directed towards a proposition.
Obviously one can mention other ‘propositional attitudes’ which would appear to
need a similar account, e.g. desiring that P, fearing that P, hoping that P, and so on. But
it would seem that Russell means to treat all of them just in terms of belief, as already
discussed, and one or another of the commonly admitted kinds of sensation. At any
rate, this is the way in which he has earlier treated desire, in his chapter 3, and which he
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 187

rather gestures at in the case of the emotions in chapter 14. Let us take up the case
of desire.
Russell begins with the behaviourist account of desire, which is designed to suit
desires in animals. It relies on the idea of a goal-seeking pattern of behaviour, which
tries various means until it finds one that does achieve the goal, and then it ends. There
is initially a discomfort, and this prompts the goal-seeking behaviour, which ends when
it achieves something that relieves that discomfort. A standard example is that hunger
will prompt food-seeking behaviour, and generalizing from this Russell concludes:
‘We may say that what we call a desire in an animal is always displayed in a cycle of
actions having certain fairly well-marked characteristics’ (p. 63). Moreover he is ready
to accept the behaviourist view that there need be nothing ‘mental’ that is thus
displayed, save for the initial discomfort and the choice of actions. However, he of
course accepts that human beings are sometimes conscious of their desires, and his
proposal is that this is merely a matter of their having, in addition to the discomfort, a
correct belief as to what will relieve it24 (e.g. in the case of hunger, food) (p. 72).
Considered as an account of human desire, this has many weaknesses, most obviously
that one may have a desire and yet not engage in any cycle of activity that will lead to its
satisfaction. For example, one may be trying to give up smoking, and desire to smoke,
yet take no steps to do so. More to the point, one may be in some discomfort, and may
have a correct belief about what would bring it to an end, and may indeed be engaged
in a course of action that will bring it to an end. But it does not follow that the course of
action is explained by that belief, for it does not follow that one undertakes the action
in order to assuage the discomfort. There is, perhaps, some other feature of the goal that is
what is wanted, while the discomfort is really neither here nor there. (To take a
dramatic example, suppose that I am thirsty, and that I assuage my thirst by drinking
poison, knowing that the drink will indeed quench my thirst. But the reason why
I drink the poison is not because I want to quench my thirst, but because I want to end
my life. The thirst was quite accidental.) The moral is that desire is not just a belief that
some course of action will relieve a discomfort; it is a matter of wanting to relieve it,
which may presuppose a belief about what will relieve it, but is not to be identified
with that belief, or with any other. It should also be added that one may well want
things which can hardly be described as relieving a present discomfort (e.g. when I
want to have a clear head tomorrow morning, and for that reason abstain from drinking
too much tonight).
We should conclude that wanting is not just a matter of believing. I suggest that we
should say the same of hoping and fearing, loving and hating, and many others. Perhaps
belief is always involved, and in some cases particular bodily feelings may also be
involved, but these two together do not exhaust the mental state or event in ques-
tion. Russell would therefore have to invoke yet further varieties of what he calls

24
Presumably, Russell is assuming that this is a conscious belief. I shall discuss this point shortly.
188 K N OW L E D G E

‘sensations’, as well as the so-called ‘belief-feeling’. But I shall take only one more case
in some detail, and that is the last case that Russell himself does discuss, namely
consciousness, or being aware of something.
In the opening chapter of AMind Russell had given short shrift to the idea that
consciousness is the essence of minds, on the very fair ground that often we are not
conscious of what is going on in our minds. As he has insisted during the course of the
book, there are unconscious sensations, unconscious desires, unconscious beliefs, and
in general: ‘there is, so far as I can see, no class of mental or other occurrences of which
we are always conscious whenever they happen’ (p. 288). But of course it must be
admitted that sometimes we are conscious of these things, and in his final chapter he
offers an account of what this consciousness is.
He begins with the idea that to be conscious of a sensation is to have an image of that
sensation, an image that occurs directly after the sensation itself. But he adds that this
image must be accompanied by the belief that it is a sign, i.e. that it means something
other than itself, in fact that it means the sensation of which it is an image. This will be
achieved, he thinks, if the image is accompanied by a belief which one might express as
‘this has just occurred’, in much the same way as what we recognize as a memory image
is accompanied by the belief ‘this once occurred’. However, this account is at once
open to an obvious objection. Russell believes that sensations can occur unconsciously,
and that this quite often happens. He also believes that images can occur unconsciously:
for example a simple memory may be unconscious, and on his account simple
memories will standardly involve images. He also accepts that beliefs can be uncon-
scious, including the belief ‘this once occurred’ which will be part of an unconscious
memory. So why should we not suppose that everything that he has mentioned—the
sensation, the immediately succeeding image, and the associated belief that this has just
occurred—could all occur unconsciously?25 If so, then he has not mentioned anything
that makes the difference between consciousness and unconsciousness, i.e. the differ-
ence between being aware of what is happening and not being aware of it.
One can of course be conscious of much more that happens in one’s mind than just
one’s sensations. The only other example that Russell discusses is the case of being
conscious of an image, and he gives a similarly unconvincing explanation of that.26 But
he should surely have considered being conscious of a belief, and perhaps also of one’s
deliberations, desires, emotions, and so on. (For example would the consciousness of a
belief involve both an image of the proposition believed and an image of the belief-
feeling directed to it?) But presumably, with the materials that he regards as available to
him, Russell can give no better account of being aware in these more complex cases

25
This might happen if, for example, the person was at the time in a daydream, or greatly distracted by
another—and perhaps very painful—sensation.
26
He balks at the idea that it is to have an image of that image, presumably because a copy of a copy can
hardly be distinguished from the first copy. So he suggests that it is to have a very similar image, but one that is
associated with the word ‘image’, to show that it is a sign of an image, and not of the sensation which gave rise
to it.
 L O G I C A L C O N S T RU C T I O N S ( I I ): MINDS 189

than he did of the simple case first considered. Apparently he needs awareness as a new
primitive notion, which would have to be, for him, yet another kind of ‘sensation’.
We may conclude that the attempt to reduce minds to just two kinds of occurrences,
called ‘sensations’ and ‘images’, does not in the end work out. Minds are more complex
than this allows for, and the notion of a ‘sensation’ has to be most uncomfortably
stretched if this attempt is to have any hope of success. Moreover, it is only the very
basic kinds of sensation, which we think of as yielding perceptions of the external
world, that have any title to be ‘neutral’ entities, entering into the composition of both
minds and material objects. But, as the next chapter will show, Russell later came to see
that even this amount of neutrality cannot really be maintained.
 11
The Demise of Neutral Monism

As we have seen, Russell’s version of neutral monism was never properly ‘neutral’ or
‘monistic’. Most of the ingredients from which minds are constructed do not also occur
in matter, and conversely most of the ingredients from which matter is constructed do
not occur in minds. For most of them are unsensed sensibilia, but it is only actual events
of sensing that occur in actual minds. So it is only these events that he can claim to be
genuinely ‘neutral’. We quite naturally think of such events as relating mind and matter
in a very direct way, but Russell has treated them as ingredients in both. However, even
this amount of neutrality does not persist in his thinking for very long. It is in effect
abandoned in his next major work, The Analysis of Matter (hereafter AMatter), which
was published in 1927.
What leads to the change of view is his recognition of the importance of the causal
theory of perception,1 for this must bring with it a fuller notion of what is to count as a
cause than he has previously accepted. Let us take this point first.

1 Causes in general
In his previous writings, up to and including AMind, Russell had always refused to
admit that there was anything more to the legitimate notion of a cause than simply a
regular connection. He had set his face against the idea that a cause is somehow
‘responsible’ for its effect, or that it ‘produces’ the effect, or ‘makes’ it happen, or
‘compels’ it, or ‘necessitates’ it. With these idioms in mind, he once said that the reason
why physics has ceased to look for causes is that there are in fact no such things, and
The law of causality . . . is a relic of a bygone age, surviving, like the monarchy, only because it is
erroneously supposed to do no harm. (On the Notion of Cause, 1912/13, p. 180)

What he does accept is that the sciences aim to discover and to use laws, which state
that if there are things or events with a certain property then (almost always) there will
also be related things or events with another suitably related property. Such laws allow
one to infer from the first to the second. But he does not accept that any of these laws
have a better claim than any others to be entitled causal laws. (Or, to be more cautious,

1
This theory is outlined on p. 162 above.
 THE DEMISE OF NEUTRAL MONISM 191

if any of them do then that is simply irrelevant to science.) Any kind of law, that states a
regular connection and thereby legitimizes an inference, is equally important to
science.
A nice illustration of this attitude may be found in an example that he mentions in
AMind [1921a]:
For example, men leave a factory for dinner when the hooter sounds at twelve o’clock. You may
say the hooter is the cause of their leaving. But innumerable other hooters in other factories,
which also always sound at twelve o’clock, have just as good a right to be called the cause. Thus
every event has many nearly invariable antecedents, and therefore many antecedents which may
be called its cause. (AMind, p. 97)

This naturally provokes the criticism made by C. D. Broad [1925: 456], that in fact we
do not suppose that the factory hooters sounding in Birmingham cause the workers in
Manchester to leave their factories, so the moral is that there must be more to the
notion of a cause than Russell is allowing. But in 1921 Russell clearly does not accept
the correction. His view is that if the connection in the one case is just as regular as it is
in the other, then each has an equal title to be called a cause.
In AMatter [1927a] Russell has changed his mind. According to the causal theory of
perception, as it is usually understood, the reason why I see a penny on the table in
front of me is, usually, that there is a penny there, and it is this penny that causes me to
have a perception of a penny. But Russell cannot accept this view while he still believes
that the penny just is the set of all (actual and possible) perceptions of it. For a set cannot
be said to cause the existence of its members, as even Russell will agree. (This is because
he takes it to be a matter of logical necessity that a set cannot exist without its members
existing, and he also holds that no causal connection can itself be a logically necessary
connection.2) Moreover, we may mimic the case of the factory hooters in this way.
Perhaps the cause of my present penny-perception is the possible penny-perception
that exactly matches mine, and is to be had from the opposite side of table at the same
distance from it that I am? This suggestion is one that the Russell who wrote AMind
must apparently accept, for there is a perfectly regular and contingent connection
between the two. More generally, given enough penny-perceptions to allow one to
calculate where the penny itself is, one can then use the laws of perspective to predict
from them all other possible perceptions of that penny from all other positions. So, by
analogy with the case of the factory-hooters, any sufficient selection of perceptions of
that penny will cause all the others. But that is clearly incompatible with what we, and

2
AMind does appear to say that the set causes each one of its members, e.g. ‘Every regular appearance [of
a star] is an actual member of the system which is the star, and its causation is entirely internal to that system’
(p. 134). (A ‘regular’ appearance is one that is not in any way distorted by the medium between the place of
the star itself and the place of the appearance, e.g. an appearance in vacuo.) We can only make sense of this by
supposing that as a spherical shell of appearances ‘spreads outwards’ from the place of the star itself, the earlier
members are taken to cause the later ones.
192 K N OW L E D G E

Russell, understand by ‘the causal theory of perception’. This theory, which Russell
endorses in AMatter, requires a more detailed notion of what is to count as a cause.

2 The new causal theory of perception

Russell’s new theory of perception is given in chapter 20 of AMatter. He begins by
assuming a familiarity with the ordinary notion of space, as a framework in which to
interpret what we experience (p. 198; cf. pp. 207–8). He does not tell us how we come
to understand this ordinary notion, but we can see that his previous ideas on this have
now received an important correction. For he now says that ‘we cannot arrive at the
common-sense idea of space without bringing in movement’ (p. 166; cf. pp. 216–17).
Anyway, assuming this idea of space, his discussion in chapter 20 focuses on the
question whether we have reason to suppose that there exists anything other than
what we ourselves perceive, and it opens with the usual admission that no such reason
can be logically compelling (pp. 198–200). Nevertheless Russell thinks that we do have
reasons which are quite good enough to justify the universal belief, which is one that
we cannot in practice set aside. His suggestion is that we start with the belief in other
people, i.e. my belief that there are other bodies much like mine, and that they also
have minds much like mine (pp. 200–3). (The argument here is the traditional
argument by analogy, comparing their behaviour with mine, and inferring a similar
explanation in each case.) This has already accepted that others have perceptions like
mine, and hence that my perceptions are not the only perceptions (pp. 203–6). It also
allows me to rely on their testimony as well as what I perceive myself, and so to group
together a great number of perceptions that are sufficiently similar to one another to
make it reasonable to believe that they have a common origin. On this Russell
comments
The argument here is, I think, not so good as the argument for other people’s percepts. In that
case, we were inferring something very similar to what we know in our own experience,
whereas in this case we are inferring something which can never be experienced, and of
whose nature we can know no more than the inference warrants. Nevertheless, the common-
sense arguments for an external cause of perception are strong. (p. 207)3

He adds the familiar explanation of how this assumed ‘external cause’ can be assigned a
location in our common space (pp. 207–8). I have outlined this explanation on p. 162
above.
He also adds what he calls an argument for ‘the causal theory of perception’, which is
in effect a suggestion about the working of this posited ‘external cause’:

3
Notice how Russell here elides the step of first accepting other people’s bodies, as an external cause of
many of my own perceptions, which is a preliminary to accepting that they too have perceptions. He speaks
as if I went straight from my perceptions to theirs, and only later considered material bodies.
 THE DEMISE OF NEUTRAL MONISM 193

Suppose a gun on a hilltop is fired every day at twelve o’clock: many people both see and hear it
fired, but the further they are from it the longer is the interval between seeing and hearing . . .
[This] makes it natural to adopt a causal view of sound, since the retardation of the sound
depends upon the distance, not upon the number of intermediate percipients. But hitherto our
space was purely ‘ideal’ except where there were percipients; it seems odd, therefore, that it
should have an actual influence. It is much more natural to suppose that the sound travels over
the intervening space, in which case something must be happening even in places where there is
no one with ears to hear. (p. 209)

He comments ‘The argument is perhaps not very strong, but we cannot deny that it has
some force’. What is interesting from our perspective, however, is the fact that he takes
it to be built into the ‘causal’ theory of perception that when what is perceived is a
distant object or event then there is something that travels (on a continuous path) from
that object or event to the perceiver. He proceeds to add further evidence that
something must be travelling, and he adds that it must be much the same thing that
travels to each of us, for that is why the various perceptions that we all get are so similar.
Moreover, this same thing must be capable of causing reactions in cameras or micro-
phones placed on its route, for that is why they can make records which exactly match
our perceptions (pp. 209–10). We may at once infer that what he is talking of cannot
be the sensibilia that were the focus of attention in RSP and KEW. It surely makes no
sense to speak of a sensibile as travelling from one place to another. That is not the kind
of thing that a sense-datum or an event of sensing can do. This point will very soon be
verified. But before we get to that we find that here Russell’s discussion changes tack,
and he asks whether—instead of the causal theory that he has been sketching—we could
accept what he calls the ‘phenomenalist’ view of these matters.
The phenomenalist is described as one who will accept the existence of actual
perceptions (whether mine or someone else’s), but of nothing else. However he also
finds it permissible to speak of ‘ideal’ perceptions,4 which one may think of as occupying
the space surrounding the object perceived, and filling in the gaps between actual
perceptions, to give a continuous variation from one place to another. A number of
such filled-in groups of (actual and ideal) perceptions, occurring at different times, may
then be further collected together, and arranged in a temporal series to form a
continuous ‘biography’. This biography can then play the role of a continuing object,
namely the object which, according to the causal theory, is the cause of all those actual
and ideal perceptions.
‘Ideal’ percepts, groups, and things, in this theory are really a shorthand for stating the laws of
actual percepts, and all empirical evidence has to do with actual percepts. The above account,
therefore, preserves the truth of physics with the bare minimum of hypothesis. (p. 213)

It is obvious that what is here described as the ‘phenomenalist’ position is in fact the
same as, or at least very similar to, the position that Russell had himself put forward as

4
On p. 210 Russell speaks of ideal percipients, but elsewhere of ideal percepts.
194 K N OW L E D G E

his ‘logical construction’ of material objects in RSP and KEW (and which he had then
recapitulated in AMind). That position invoked ‘unsensed sensibilia’, and we found the
texts somewhat ambivalent on just how these should be construed. But the best
interpretation seemed to be that they were possible appearances rather than actual
appearances, and they were also said to be ‘ideal’ elements (KEW, p. 117). One could
also say that their main role was to allow us to simplify the statement of the laws of
physics.
It is significant, then, that here in AMatter Russell goes on to reject this ‘phenome-
nalist’ theory:
The great difficulty in the above theory of ‘ideal’ elements is that it is hard to see how anything
merely imaginary can be essential to the statement of a causal law. (p. 214)5

His immediate point is that, if the only things that (really) exist are actual perceptions,
then we lack an explanation for why unseen cameras and microphones behave as they
do. His more general point is that, if these ‘ideal’ elements do not actually exist, then
the ‘real’ laws of physics cannot rely upon them, and this must make it very difficult to
state those laws. But it seems to us that the main objection to this phenomenalist
approach to the laws of physics is that it deprives us of all explanations. For the laws
could only tell us that if there are such-and-such perceptions at this time and this place,
then there are such-and-such other perceptions at that time and that place, but they
cannot tell us why. We could not discover why perceptions occur at all, nor why some
are naturally connected to others. But it is very commonly thought that our sciences
should provide explanations as well as predictions. However, this does not seem to be a
thought to which Russell paid much attention. He does not appear to associate the idea
of a cause with the idea of explanation. Even so, what he does say is enough to destroy
his previous attempt at neutral monism.
He tells us that for certain purposes the phenomenalist construction ‘remains valid
and important’ (namely ‘as a method of separating perceptual and non-perceptual
elements of physics’), and that he will therefore retain it, except that what it called
‘ideal’ elements will now be recognized as having a real existence (p. 215). They will
exist at the same places as before, namely those where there are no actual perceptions.
But we no longer assume . . . that what is at such places is what we should perceive if we went to
them. We think, e.g., that light consists of waves of a certain kind, but becomes transformed, on
contact with the eye, into a different physical process. Therefore what occurs before the light
reaches an eye is presumably different from what occurs afterwards, and therefore different from
a visual percept. But it is supposed to be causally continuous with the visual percept; and it is
largely for the sake of this causal continuity that a certain reinterpretation of the physical world
seems desirable. (p. 216)

5
It has not been said earlier that what the phenomenalist describes as ‘ideal’ is simply ‘imaginary’, but that
is the plain implication of this sentence.
 THE DEMISE OF NEUTRAL MONISM 195

What were ‘ideal’ elements (or unsensed sensibilia) have therefore become the causes of
perceptions, e.g. such things as light-rays of various wavelengths travelling in various
directions, sound-waves, molecules that affect our sense of smell, temperatures, and so
on. These are not at all the same kinds of things as are the perceptions that they can
cause, if they meet with suitable sense-organs. Apparently Russell still wishes to
maintain the ‘logical construction’ of matter, but it is now to be constructed from
these things that can cause the perceptions, and not the perceptions themselves.
In AMind Russell had said:
It was natural, though to my mind mistaken, to regard the ‘real’ table as the common cause of all
the appearances which the table presents (as we say) to different observers. But why should we
suppose that there is some one common cause of all these appearances? As we have just seen, the
notion of ‘cause’ is not so reliable as to allow us to infer the existence of something that, by its
very nature, can never be observed (pp. 97–8).

In AMatter he says almost the opposite:

Science holds that, when we ‘see the sun’, there is a process starting from the sun, traversing the
space between the sun and the eye, changing its character when it reaches the eye, changing its
character again in the optic nerve and the brain, and finally producing the event which we call
‘seeing the sun’. Our knowledge of the sun thus becomes inferential; our direct knowledge is of
an event which is, in some sense, ‘in us’ (p. 197).

Somewhat oddly he regards this theory as running counter to the ‘common-sense’

view that ‘when we “see the sun” it is the sun that we see’, but still he praises it as a
good theory:
It has all the merits of a good scientific theory—i.e. its verifiable consequences are never found to
be false. Epistemologically, physics might be expected to collapse if perceptions have no external
causes (ibid.).

He then proceeds to outline its virtues, as I have indicated. It clearly does involve
inferring from our perceptions to the existence of other things that cause our percep-
tions, and that are quite unlike the perceptions that they cause. That is, it does just what
AMind has said that one should not do.
In the course of doing so, it abandons the identification on which neutral monism
depended, for it now accepts that perceptions do have physical causes and are not the
same as those causes. So if minds are still to be constructed then they will of course be
constructed from the perceptions, and not their causes, whereas if bodies are still to be
constructed (rather than postulated) they will be constructed from these physical
causes, that are supposed to exist in space whether or not they do cause a perception
in a sentient body, and not from the perceptions.6 In consequence there is no longer

6
In AMatter Russell does still think that bodies should be constructed, namely from small and short-lived
spatio-temporal events, which are the minimum events that we can perceive. But these events are the causes
of our sensations, and not the sensations themselves.
196 K N OW L E D G E

anything which is supposed to be an ingredient of both constructions, and neutral

monism has finally disappeared.
I add that by the time that he came to write AMatter Russell was much more
concerned than he had been with the interpretation of the new physical theories—in
particular relativity theory and quantum theory—that had been introduced early in the
century, but were by then (i.e. in the 1920s) becoming widely discussed among the
physicists themselves. This had led him to drop his earlier preference for ‘horizontal’
rather than ‘vertical’ inferences (pp. 158–9), for it was clear that the physicists showed
no such preference. But he realised that this leaves things very open-ended:
If we have once admitted unperceived events, there is no very obvious reason for picking and
choosing among the events which physics leads us to infer (AMatter, p. 325).

Much of AMatter is in fact concerned with the question of how the latest discoveries of
the physicists are best conceived, and this often involves setting aside the ‘common-
sense’ view of the world that we live in.
I say no more of the development of Russell’s thoughts on this topic, for it clearly
lies outside the interests and concerns of his philosophy of logical atomism. Instead, I
return for a final word on the topic of neutral monism that led us into the account of
perception in chapter 20 of AMatter.

3 Russell’s understanding of the situation

It is not clear whether Russell himself realized just how much his views in this chapter
of AMatter departed from those of his earlier writings. In AMatter itself he does not
admit that his concept of a cause has now altered, but the admission does come much
later, in his last major work on this and allied topics, i.e. his Human Knowledge: Its Scope
and Limits [1948]. He says there that he now ‘cannot accept the view that causation is
merely invariable sequence’ (p. 455). Quite how much he now wishes to add is not
altogether clear, but at least it includes this:
When there is a causal connection between two events which are not contiguous, there must be
intermediate links in the causal chain, such that each is contiguous to the next, or (alternatively)
such that there is a process which is continuous in the mathematical sense. (p. 508)7

This point about causation is indeed being assumed in AMatter, and it figures explicitly
in the summary on p. 217, but it is not there acknowledged as a new departure.8

7
The alternation is due to the thought that we can never verify the strict continuity of any physical
process, and that there are some indications (drawn from quantum theory) that discontinuity might be
expected.
8
Note that it denies what he had been willing to accept as a possibility in AMind, that ‘mnemic causation’
may involve causal ‘action at a distance’ (p. 182n. above).
 THE DEMISE OF NEUTRAL MONISM 197

I add that in an article entitled ‘Physics and Perception’, published in 1922, i.e. some
years after the logical construction of RSP and KEW, and one year after its recapitula-
tion in AMind, Russell claimed that he had never accepted phenomenalism. Indeed, it
would not have been particularly surprising if he had repeated this same claim when
discussing what AMatter describes as phenomenalism. If so, that could only be because
he then thought of himself as having assigned some kind of real existence to unsensed
sensibilia even in RSP and KEW. It is true that we failed to see how this could be
squared with the other claims that he makes in those two works, but perhaps he could
shut his eyes to that. It may be that he even thought of those sensibilia as spreading out
from an origin in such a way that those nearer to the origin caused those further from it,9
but really that suggestion makes little sense.
However, by far the most significant point is that Russell himself seems not to have
realized that AMatter abandons neutral monism. In a volume of essays on his philoso-
phy edited by P. A. Schilpp [1944], there is one by W. T. Stace which claims that
AMatter does abandon the theory. But Russell’s response denies this. Quite generally,
he claims that
In the main there is [in AMatter] a fuller and more careful statement of theories not very different
from those of [AMind]. (p. 707)

In particular, he says that AMatter retains the doctrine of neutral monism, and adds that
this theory is not in any way incompatible with the view that matter causes sensations.
It is tempting to suppose that what lies behind this statement is a passage in AMind that I
have been ignoring. At pp. 99–101 of that book he explains a physical object as the
collection of its ‘appearances’, but then adds
When I speak of ‘appearances’, I do so only for brevity: I do not mean anything that must
‘appear’ to somebody, but only that happening, whatever it may be, which is connected, at the
place in question, with a given physical object—according to the old orthodox theory, it would
be a transverse vibration in the æther [i.e. what we now call a light-wave] (p. 101).

The physical object, according to this passage, is the collection of its physical effects,
such as light-waves, which is indeed close to what AMatter says. When reading AMind
it is not unreasonable to set this passage aside as quite untypical, because for the most
part AMind appears to endorse the construction of physical objects from sensibilia that
we find in RSP and KEW. Moreover, AMind aims to construct minds from sensations
(and images), and clearly it does not construe a sensation as a physical event, such as the
passing of a light-wave. But it does claim that the same sensations are ingredients both
of minds and of bodies, and that is how it conforms to neutral monism. However if we
take seriously this passage just cited from AMind, then we have to say that it too, like
AMatter, does reject neutral monism. Perhaps Russell has not seen that there is a clear
change of view on the construction of matter between RSP and KEW on the one side,

9
Compare note 2 to this chapter.
198 K N OW L E D G E

and AMatter on the other. And perhaps he has failed to see this because he has
incorporated both views side by side in AMind, as if they were entirely consistent
with one another. But they are not.
With that comment I now leave Russell’s various theories on what we can know,
and how we know it, and come back to the period that is always regarded as containing
the heart of his philosophy of logical atomism. But from now on the emphasis will be
on its central area, i.e. metaphysics.
 PART III
Metaphysics
The last three chapters have considered Russell’s views on the nature of material
objects and the nature of persons (and other animated beings). These may fairly be
said to be topics in metaphysics, and Russell himself does so describe them, when he
calls the last of his eight lectures on ‘The Philosophy of Logical Atomism’ an ‘Excursus
into Metaphysics’. The main line of argument, however, was centred on the question
of how we could know about these things. In the present part we shall be concerned
with more abstract topics, namely the natures of propositions and universals and facts,
where Russell is less worried by that question. It is true that one can be puzzled over
how we can know about propositions, especially when conceived as Russell first
conceived them, and how we can know about universals. But for the most part
these entities are not regarded as problematic for this reason, though it is acknowledged
that there are other difficulties which affect them. Or perhaps it is better to say that
while Russell certainly did see other difficulties over propositions, which twice led him
to change his mind on this topic, what calls for comment in his treatment of universals
and of facts is more his failure to see difficulties which naturally strike us. But I say no
more by way of introduction to these topics than just this: we have seen that the
principle known as Occam’s Razor, which claims that entities should not be multiplied
beyond necessity, has an important role in his thinking about matter and about minds.
It continues to be important as we turn to these further topics.
This page intentionally left blank
 12
Propositions

In his early book The Philosophy of Leibniz [1900] Russell had said: ‘That all sound
philosophy should begin with an analysis of propositions is a truth too evident, perhaps
to demand a proof ’ (p. 8). Indeed, his concern with propositions and their analysis is
evident throughout his career. But his views on the subject did not remain constant.
There are three broad periods to be distinguished. The first lasts from his break with
idealism in 1898 until about 1907; during this time he has a realistic attitude towards
the existence of propositions, but increasing doubts about just what kind of things they
are. This gives way to a sceptical period in which he claims that we do not need to
recognize propositions as entities at all. The method of dispensing with them is given in
what is known as his ‘multiple-relation’ theory of judgment, which he proposed in a
paper of 1910, used in Principia Mathematica [1910d], and expounded once more in his
more popular book Problems of Philosophy [1912a]. He was working on a further
elaboration of the theory in 1913, but was apparently persuaded by the young
Wittgenstein that it would not do, and so dropped the project. Finally, the theory is
noted in his lecture on The Philosophy of Logical Atomism of 1918, but only because at
that stage Russell has nothing better to put in its place. He is no longer confident of its
merits. The third period follows at once in the paper ‘On Propositions’ of 1919. In
effect, he once more admits that there are such things as propositions, but now regards
them as mental entities. This attitude continues into The Analysis of Mind [1921a], and
beyond. The final theory has already been mentioned in Chapter 10, but let us begin at
the beginning.

1 The initial realism

Until 1898 Russell was still under the spell of the idealist philosophy, stemming from
Hegel, which was prevalent at Cambridge during his early years there. Propositions
were then thought of as being in some way mental items. But when he and G. E.
Moore broke free of idealism, one of their first moves was to distinguish on the one
hand the mental state or event of believing a proposition, and on the other the
proposition itself which is the object of that belief. In a nutshell, their view was that
the proposition is ‘independent of our acts of judging, or acts of synthesis, or acts of
202 M E TA P H Y S I C S

any kind. Propositions, in short, are conceived of as objective and independent

entities’.1 Indeed, they held a similar view about all mental states or events, e.g.
knowledge, belief, thought, perception, and even imagination: ‘in the case of each of
these mental acts or states their view is that we are in contact with an object that is not
mental. And the object with which we are in contact is, in all such cases, unaffected
by the fact that we are in contact with it.’2 When the object in question is a
proposition it will (usually?) be something that we can express in words, i.e. in a
language which we human beings have created. But we must distinguish the propo-
sition itself, which is not a human creation, from the words that we use for it. As
Russell said in PoM [1903]:
Words all have meaning, in the simple sense that they are symbols which stand for something
other than themselves. But a proposition, unless it happens to be linguistic, does not itself contain
words: it contains the entities indicated by words. Thus meaning, in the sense in which words
have meaning, is irrelevant to logic. (PoM, }51)

Admittedly, he has earlier recognized that it can be useful to note how the proposition
is expressed in language:
The study of grammar, in my opinion, is capable of throwing far more light on philosophical
questions than is commonly supposed by philosophers. Although a grammatical distinction
cannot be uncritically assumed to correspond to a genuine philosophical difference, yet the
one is prima facie evidence of the other, and may often be most usefully employed as a source of
discovery. . . . In what follows, grammar, though not our master, will yet be taken as our guide.
(PoM, }46)

As we have seen again and again in this book, Russell’s ability to meet a difficulty by
proposing new and unexpected theories has rather depended upon his willingness to
set aside the suggestions of English grammar. So it is worth noting that this willingness
is present right at the beginning, for grammar is not ‘our master’, even if in PoM itself it
still seems to have more influence than it should.
In PoM he had offered what is now a familiar analysis of singular, atomic propositions
as containing one or more things that are referred to, which he called the ‘terms’ of the
proposition, and something said of those terms, which he called the ‘assertion’ of the
proposition. Various distinctions are made within these general categories, which need
not concern us here, though we should note three assumptions which are significant
mainly because Russell will later deny them. First, the usual interpretation3 is that
in PoM he will accept more or less any singular expression as introducing a term. In a
well-known passage he says:

1
I quote this summarizing remark from p. 109 of the useful exposition in Hylton [1990: 105–16].
2
Hylton [1990: 110]. (As Hylton notes, in some cases the object of a mental state or event—e.g. an
introspection—may itself be another mental state or event, but this is not the usual case.)
3
For some doubts, see Chapter 3, pp. 32–4.
 P RO P O S I T I O N S 203

Whatever may be an object of thought, or may occur in any true or false proposition, or can be
counted as one, I call a term. This, then, is the widest word in the philosophical vocabulary. . . .
A man, a moment, a number, a class, a relation, a chimaera, or anything else that can be
mentioned, is sure to be a term; and to deny that such and such a thing is a term must always
be false. (PoM, }47)

Second, and partly as a consequence of the first, he insists that any part, or the whole, of
a proposition may itself figure as a term of another proposition. This is just because we
can refer to the proposition itself, and to what is asserted of its terms, as well as referring
to its original terms. Third, in PoM he of course recognizes the truthfunctors that we
symbolize by ‘’, ‘&’, ‘Ú’, and so on; in fact he takes ‘!’ as basic, and analyses the
others in terms of that.4 But he does not have much to say about their metaphysical
status, or about how we understand them. So far as the quantifiers are concerned, he
there offers the strange theory of the words ‘all’, ‘every’, ‘any’, ‘an’, ‘some’, and ‘the’,
which we now call his early theory of denoting. But he considers no other ways of
building up more complex propositions from the atomic ones that we begin with.
In subsequent developments he first came to jettison the early theory of denoting,
and to adopt a more familiar Fregean attitude to the quantifiers. This is very briefly
indicated in the first few pages of the article ‘On Denoting’, which of course is much
better known for its new account of ‘the’ and for its more restrictive attitude to what
counts as a genuine term (or constituent) of a proposition. It allows us to dismiss non-
entities, as no longer needed in this role, and to argue that many other expressions
which appear to introduce terms as the subjects of propositions do not really do so.
Instead these expressions should be construed as overt or concealed definite descrip-
tions, and so analysed away in terms of quantifiers and identity. (This approach is then
pushed to the limit when it is argued that only the momentary objects of direct
acquaintance, i.e. sense-data, can qualify as genuine subject terms.) Finally, the assump-
tion that propositions and assertions may themselves be referred to is rejected by
Russell’s eventual theory of types, as we saw in Chapter 5. For that theory insists
that a predicate can be accepted only when (after analysis) it occurs ‘as a predicate’, and
similarly a proposition only when it occurs ‘as a proposition’, which in each case means
that these things cannot be named as terms. However we shall in this chapter find a
consideration of some other ways of building more complex propositions from our
starting point, in particular by prefixing ‘John believes that’. But all of these points are,
in a way, points of detail. They concern the question of how propositions should be
analysed, and the restrictions thereby imposed upon what propositions there can be.
But let us now set such things aside in order to return to our original question: what
kind of thing is a proposition?

4
This analysis must also make use of some paradigmatically false proposition. Russell employs ‘8p(p)’ in
this role.
204 M E TA P H Y S I C S

It is part of Russell’s revolt from idealism that he comes to claim that propositions
exist, without any dependence on human activities. The same will therefore apply to
the terms and the assertions that compose them, and at this stage Russell supposes that
their terms will usually be the familiar objects of the external world. There is here a
contrast with Frege, who had also regarded propositions as existing independently of
any human activity (even though he called them ‘thoughts’), but who had held that
these propositions are made up of the senses of the words used to express them, and not
the items that those words referred to. Russell disagrees. In December 1904 he wrote
to Frege, using an example that Frege had introduced in a previous letter:
Concerning sense and reference, I see nothing but difficulties which I cannot overcome. I
explained the reasons why I cannot accept your view as a whole in the appendix to my book
[i.e. PoM], and I still agree with what I there wrote. I believe that in spite of all its snowfields
Mont Blanc itself is a component part of what is actually asserted in the statement5 ‘Mont Blanc is
more than 4,000 metres high’. We do not assert the thought, for this is a private psychological
matter: we assert the object of the thought, and this is, to my mind, a certain complex (an
objective statement,5 one might say) in which Mont Blanc is itself a component part. If we do
not admit this, then we get the conclusion that we know nothing at all about Mont Blanc.6

One could wish that at the same time Russell had said what there was in the world that
was signified by the assertion ‘ . . . is more than 4,000 metres high’, and what there was
that corresponded to the whole sentence. The first of these questions is postponed to
the next chapter (for the answer may be ‘a universal’). The second must be pursued
now: are propositions to be thought of as existing in the material world?
The main problem was that Russell was apt to think of a true proposition as simply
being a fact, and to regard a fact as just as much something in the world as its terms are,
yet he could not bring himself to say anything similar about false propositions.
Apparently there must be such things as false propositions, since they are needed as
the objects of false beliefs, and they must presumably have the same ontological status as
do true propositions. But it is also very tempting to say that truths are to be found in the
world while falsehoods are not. Russell felt himself caught by this puzzle. Here, for
example, is how he ends a long review of Meinong’s publications on this and similar
topics:
It may be said—and this is, I believe, the correct view—that there is no problem at all in truth and
falsehood; that some propositions are true and some false, just as some roses are red and some
white; that belief is a certain attitude towards propositions, which is called knowledge when they
are true, error when they are false. (1904: 75)

5
’Statement’ here translates Russell’s German word ‘Satz’. (One might put ‘sentence’ for the first
occurrence, and ‘proposition’ for the second.)
6
Frege [1980: 169]. In ‘On Denoting’ Russell argued that there could not be such things as Fregean
senses. And later he came to think that the genuine terms of a proposition had to be sense data, and so would
not include Mont Blanc. The present letter precedes ‘On Denoting’, but the example persists on its p. 46n.
 P RO P O S I T I O N S 205

Yet in the very next sentence he sows a doubt:

But this theory seems to leave our preference for truth a mere unaccountable prejudice, and in no
way to answer to the feeling of truth and falsehood. (Ibid.)

At first the doubt seems not to be too serious. It merely asks for more, and not for any
revision of what has just been said. But in the next paragraph the real source of worry
begins to emerge:
The fundamental objection may be simply expressed by saying that true propositions express fact,
while false ones do not. This at once raises the problem: What is a fact? And the difficulty of this
problem lies in this, that a fact appears to be merely a true proposition, so that what seemed a
significant assertion becomes a tautology. (Ibid.)

However the continuation shows that he still hankers after the idea that truth requires
the existence of something else, which falsehood does not, or at any rate that this
holds for positive truths, if not for negative ones. Yet the discussion ends by once more
setting this idea aside:
Thus the analogy with red and white roses seems, in the end, to express the matter as nearly as
possible. What is truth, and what falsehood, we must merely apprehend, for both seem incapable
of analysis. (p. 76)

This was Russell’s position in 1904, and he continued to believe in the existence of
propositions (including false propositions) for the next few years. Indeed, he spent
much of 1906 working on a new approach to logic, which he called ‘the substitutional
theory’ of classes and relations, and—as we have seen (pp. 48–53)—this theory very
clearly assumes the existence of propositions. For it banishes not only classes and
relations, as ordinarily conceived, but also propositional functions, and so it is only
whole propositions, and the ordinary individuals that are their terms, that remain as the
material available for logical theory. Although Russell quite soon began to have doubts
about this theory, he never published a proper and reasoned renunciation. In fact he
still refers to it as a possible way of simplifying the ontology of logic even in the paper
ML of 1908, which in effect contains the theory of types that will actually be used two
years later in PM.7 But one should take that suggestion with a pinch of salt, for his
thought was already moving in a different direction. In a paper ‘On the Nature of
Truth’ in 1906/7 he had suggested, but not adopted, a different theory, which would
allow us to claim that propositions are ‘logical fictions’, just as classes are. That is, the
expressions which appear to mention propositions can be seen, on a proper analysis,
not really to mention any such things, but to have a different role. After some
alterations,8 this became the theory which is known as Russell’s ‘multiple-relation’

7
See ML, p. 77.
8
In the version suggested in 1906/7 a belief occurs when a mind contains a complex of interrelated ideas,
as in the later theory of 1919 (to be discussed in sections 6–7 of this chapter). It is not described as the mind
being related to the objects which those ideas stand for, which is what we find in the multiple-relation theory.
206 M E TA P H Y S I C S

theory of judgment, and he always gives as his motivation for the theory the point that
it resolves his previous problem over what kind of thing a false proposition could be.
It resolves this problem by not assuming that there are such things as propositions.9
The new theory is given in a paper ‘On the Nature of Truth and Falsehood’
(henceforth NTF), which was designed to supersede part III of the paper ‘On the
Nature of Truth’ of 1906/7, and was used to replace it in Russell’s collection
Philosophical Essays of 1910. The new theory is also employed in the Introduction to
Principia Mathematica (chapter 2, section 3), and the logic of that work aims to conform
to it, for it does not quantify over propositions10 or suggest that quantification over
propositional functions could be replaced by quantification over propositions. But its
best known version is that of the slightly later book Problems of Philosophy [1912a], and
my discussion will start from there.

2 The multiple-relation theory

Chapter 12 of PP is entitled ‘Truth and Falsehood’, and it begins by insisting that any
account of what truth is must also allow for its opposite, falsehood. Then it goes on to
claim that the only things that are either true or false are beliefs:
It seems fairly evident that if there were no beliefs there could be no falsehood, and no truth
either, in the sense in which truth is correlative to falsehood. If we imagine a world of mere
matter, there would be no room for falsehood in such a world, and although it would contain
what may be called ‘facts’, it would not contain any truths, in the sense in which truths are things
of the same kind as falsehoods. (PP, p. 120; cf. NTF, p. 143)

There is an obvious objection: belief is not the only mental attitude that is relevant
here. What one disbelieves will also be true or false, and so will what one doubts, or
what one simply assumes for the sake of seeing what would follow, and so on. In a later
version of his theory Russell corrected this point, and spoke of understanding rather
than believing, on the ground that any such mental attitude to a proposition must
involve understanding it. But since his theory of understanding exactly matches his
theory of belief, this is not a point that is worth fussing over.11 For simplicity we may

9
Linsky [1999] claims that ‘the view that everything real must be either a simple or a fact was part of
Russell’s views from before the multiple relation theory of judgment and must be compatible with the views
expressed in PM ’ (p. 49). But the truth is that the earlier PoM contains several exceptions to this claim, not
only because it accepts false propositions, but also because of the denoting concepts that it introduces, its
admission of classes, and probably in other ways too. Linsky’s claim fits what Russell was moving towards, in
PLA, but not where he started from.
10
The sole exception is proposition *14 .3. The accompanying comment makes it clear that this is an
exception.
11
The unpublished manuscript on The Theory of Knowledge, which Russell was working on in 1913,
embodies this change from believing to understanding. It also contains some other changes, as we shall see in
due course.
 P RO P O S I T I O N S 207

continue to speak just of belief (or judgment12), while recognizing that exactly the
same theory will apply to any propositional attitude. The more interesting point is that
Russell does mean to restrict attention to mental states or events. He genuinely does
suppose that in ‘a world of pure matter’, lacking all minds, there would be no truths
and no falsehoods. For example it would not be true that 2+2=4, nor false that 2+2=5.
But he offers no kind of reason for this rather surprising claim, so all that we can do
is just to note the point, and move on.13
What he wishes to say about belief is that it should not be regarded as a relation
between the believer, or his mind, and a single proposition, again because of the
difficulty over false propositions:
The necessity of allowing for falsehood makes it impossible to regard belief as a relation of the
mind to a single object, which could be said to be what is believed. . . This may be made clear by
examples. Othello believes falsely that Desdemona loves Cassio. We cannot say that this belief
consists in a relation to a single object, ‘Desdemona’s love for Cassio’, for if there were such an
object the belief would be true. . . . It might be said that his belief is a relation to a different object,
namely ‘that Desdemona loves Cassio’; but it is almost as difficult to suppose that there is such an
object as this, when Desdemona does not love Cassio, as it was to suppose that there is
‘Desdemona’s love for Cassio’. Hence it will be better to seek for a theory of belief which
does not make it consist in a relation of the mind to a single object. (PP, p. 124)

The earlier exposition in his NTF is more definite on insisting that a phrase ‘that so and
so’ does not stand for an object:
The phrase ‘that so and so’ has no complete meaning by itself. . . . We feel that [it] is essentially
incomplete, and only acquires full significance when words are added so as to express a judgment,
e.g. ‘I believe that so and so’, ‘I deny that so and so’, ‘I hope that so and so’. (NTF, p. 145)

This piece of special pleading may be set aside. For while it is true that one does feel a
kind of incompleteness with the bare phrase ‘that so and so’, there is no such
incompleteness in ‘the proposition that so and so’ or in ‘the fact that so and so’. And
of course we do think that when Othello believes that so and so then there is something
that he believes, even if the belief is false. Ordinarily, we are perfectly happy to accept
propositions as objects, and what Russell has said here will hardly convince one
otherwise.
But he does have a quite different reason for not accepting propositions as objects,
namely that doing so can lead to just those types of contradiction that his theory of

12
Russell makes no distinction between belief and judgment, and he speaks sometimes of the one and
sometimes of the other. I shall do the same.
13
The unpublished book on The Theory of Knowledge is emphatic on this point. Russell there admits that
logic seems to require the existence of propositions which no one ever entertains, but he replies that he cannot
accept this requirement. (‘I can only say that, to me personally, no such [non-mental] entities are visible’,
p. 155.) However in the later theory of AMind we meet a liberalization: ‘we may identify propositions in
general with the contents of actual and possible beliefs’ (p. 241, my emphasis). The position adopted in PP and
in TK implies that there are only finitely many propositions.
208 M E TA P H Y S I C S

types is designed to block. That is why he claims that in his ‘logically perfect’ language,
propositions cannot be mentioned or referred to, for of course the perfect language does
obey the theory of types (pp. 252–3 below). This leads us to quite a different motivation
for his theory of belief, and one that is apparently rather better. If, in order to avoid
paradoxes, we have to avoid all locutions that mention propositions (whether true or
false), then we shall naturally want some other way of saying what one would normally
say by making such a reference. That is what Russell’s theory aims to provide.14
As he presents this theory, it deals only with the simplest case, where what is believed
is an atomic proposition, referring to one or more individuals and predicating some-
thing of them. (It is assumed, for the sake of illustration, that ‘Desdemona loves Cassio’
is such a proposition.) Then the idea is that what appears to be a reference to this whole
proposition should instead be replaced by references to its several constituents. That is

Othello believes that Desdemona loves Cassio

should be analysed as asserting a four-termed relation thus

Believes (Othello, Desdemona, loving, Cassio)

It is called a ‘multiple-relation’ theory, because Russell counts any relation of three or
more terms as a ‘multiple’ relation. (See e.g. NTF, pp. 147–8.) It might have been better
to say that the theory treats belief as a ‘multigrade’ relation, meaning that it is a relation
which can take a variable number of terms, i.e. any number from three upwards. (One
term will be the believer; the others will be the constituents of the atomic proposition
believed, which will consist of an n-place relation asserted of n subject terms. This
includes the case n ¼1.) As with any other relation, the order in which the terms of the
belief-relation are given will be significant. For example, if in the above example the
terms ‘Desdemona’ and ‘Cassio’ are interchanged, then this credits Othello with quite a
different belief. We may add that his original belief is true if and only if there is a
complex fact in which loving relates Desdemona to Cassio in the way specified in the
belief-statement, i.e. if and only if Desdemona does in fact love Cassio. In this way truth
(for atomic beliefs) consists, as one would expect, in correspondence with fact.
I have earlier noted that the problems behind Russell’s theory of types provide a
motivation for dispensing with propositions. But we must now note that the theory is
also relevant in a different way, as introducing a complication: the terms of the belief-
relation cannot all be of the same level as one another. On the contrary, one constitu-
ent of the atomic proposition believed, in this case indicated by ‘loves’, must be of a
level one higher than the others, in this case indicated by ‘Desdemona’ and ‘Cassio’.
For it is only in that case that these constituents can combine to form a single atomic

14
The need to avoid the paradoxes which concern propositions is given as a motive when Russell first
suggests this kind of theory in his [1906/7: 46]. But it is not mentioned in his subsequent expositions.
 P RO P O S I T I O N S 209

proposition. But it then follows that the belief-relation, which has this higher-level
constituent as one of its terms, must itself be of a still higher level, for if not then again it
could not combine with all of its terms to yield a proposition. For brevity, let us show
the relevant levels by adding numerical superscripts, with ‘0’ for individuals, ‘1’ for
predicates of individuals, ‘2’ for predicates of these in turn, and so on. It will also be
helpful to write the predicate ‘loves’ before the two terms that it governs. Then our
original example becomes

Believes2 (Othello0, loves1, Desdemona0, Cassio0)

But now notice that the belief-relation can be iterated. Here Othello is said to have a
false belief. But Iago knows this, i.e. he has a true belief about Othello’s belief. So on a
parallel analysis we have

Believes3 (Iago0, believes2, Othello0, loves1, Desdemona0, Cassio0)

Someone else, of course, may have a further belief about Iago’s belief, and so on up.
The belief-relation, it appears, is not only multigrade but also multilevel; it can occur at
any level of the (simple) theory of types from the second upwards. Can this be
accepted?
One way of trying to avoid it is to suppose that, when forming our first belief-
statement, we should lower the type of the predicate ‘ . . . loves . . .’, e.g. by following
Russell’s own terminology and changing it to the verbal noun ‘loving’. However, this
offends against the type theory of the logically perfect language, which demands that a
predicate should always occur as a predicate, and the noun ‘loving’ does not do so. It
would also open the way for an objection which Wittgenstein brought against Russell,
namely that if ‘loving’ is construed as a term on the same level as the term ‘Desdemo-
na’, then it should be possible to substitute for it any other term of that same level (say
‘Iago’) while still preserving significance.15 But this leads to the obvious nonsense

Othello believes that Desdemona Iago Cassio

Russell acknowledged the force of this objection, and admitted that it did apply to the
way he had been thinking. In his lectures on The Philosophy of Logical Atomism he
mentions his earlier theory, but does not endorse it, just because ‘Desdemona’ and
‘loves’ are not on a level:
That is a point in which I think that the theory of judgment which I set forth once in print some
years ago was a little unduly simple, because I did then treat the object verb as if one could put it
as just an object like the terms, as if one could put ‘loves’ on a level with Desdemona and Cassio
as a term for the relation ‘believe’. That is why I have been laying such an emphasis in this lecture
today on the fact that there are two verbs at least. (PLA, p. 226).

15
I come to Wittgenstein’s objection in the next section.
210 M E TA P H Y S I C S

The emphasis on two verbs, i.e. both ‘believes’ and ‘loves’, rather suggests that Russell is
thinking of them as each on a level with the other. Though he has no particular
objection to the general idea that the same verb ‘believes’ may occur at many different
levels in the theory of types, as is clear from the preceding paragraph of PLA, still he
does not point out that it must always be of a type one higher than the type of the main
verb that it governs. However, this does seem to be the simplest course for him to take.
I note here that the multilevel nature of belief is bound to be a feature of Russell’s
theory of types, whether or not propositions are admitted as entities. For if they are
admitted, and if the theory is to contain a solution to the paradoxes that affect
propositions (most simply, the liar), then propositions will have to be distinguished
into orders, as Russell did once envisage. But then once more a proposition of the form
‘x believes that P’ will have to have an order one higher than whatever is the order of
the proposition in place of ‘P ’. So there is no way of getting away from what Russell
would call the ‘systematic ambiguity’ of the expression ‘believes that’, from one type or
order to another, at least so long as the underlying motivation for his ramified type
theory is maintained.

3 Problems with Wittgenstein

After writing The Problems of Philosophy Russell quite soon decided that he wanted to
write ‘a big book’ on The Theory of Knowledge. Its first half was to be devoted to
‘Analysis’, and the second would be on ‘Construction’. In the event, he never finished
the first half, and we have only a few very general headings for the second. But he did
write parts I and II of the first half, though he broke off before beginning on part III. It
is clear that this proposed book, henceforth TK, would make use of the multiple-
relation theory of judgment, though the theory proposed there involves several
departures from the simple version of PP. It is also clear that Russell abandoned his
attempt to write this book because of criticisms that he received from the young
Wittgenstein. What is not clear is just what Wittgenstein’s criticisms were, and why
Russell found them so damaging. This has been the subject of some dispute.
Russell was writing TK in May and June of 1913, and by then Wittgenstein had
been his graduate pupil at Cambridge for a year and a half. During that time Russell had
grown to have a great respect for his opinions, and was already thinking of him as his
possible successor at Cambridge. We know, from letters that Russell was then writing
to Lady Ottoline Morrell, of the sequence of events.16 On 14 May Wittgenstein learnt
that Russell was writing on the theory of knowledge, and on 20 May he came to see
Russell with what Russell describes as ‘a refutation of the theory of judgment which

16
These letters are held at the University of Texas at Austin, and copies are held at the Bertrand Russell
Archives at McMaster University in Ontario, Canada. Detailed references to the letters cited in this paragraph
may be found on pp. xxvii–xxviii of the introduction by E. R. Eames to volume 7 of Russell’s Collected Papers
[1984]. This volume contains all that remains of his TK manuscript.
 P RO P O S I T I O N S 211

I used to hold’. Presumably this was a ‘refutation’ of the theory that had quite recently
been published in PP, and Russell’s reaction was that the objection was correct in itself,
but could quite easily be met by a modification to his theory. On 24 May Russell
began to work on that part of his proposed TK that introduces his theory of judgment,
now in a revised form, and he showed some of what he was writing to Wittgenstein.
We have no record of what it was that he showed to Wittgenstein, but it seems likely
to have been this revised theory of judgment, for it seems likely that Russell took the
revised theory to meet Wittgenstein’s earlier criticism. On 26 May Wittgenstein came
to talk about this piece of writing, and evidently it was a stormy meeting. Apparently
Wittgenstein claimed that Russell’s view was ‘all wrong’, and said that he had himself
tried this view and knew that it would not work. Russell also comments that
Wittgenstein’s reasoning was not easy to understand, since he was ‘very inarticulate’.
In any case, for the next few days Russell continued to work on his manuscript, and on
31 May he wrote optimistically to Lady Ottoline that he thought he could ‘circumvent
Wittgenstein’s problems’. A week later (on 6 June) he had reached the end of part II of
the proposed book, and he took a break before starting on part III. But apparently he
decided during this break that he did not really have an answer to what Wittgenstein
had said (or, anyway, to what he—Russell—had now come to think of as what
Wittgenstein had said). He therefore stopped work on TK, and never resumed it.17
Russell met with Wittgenstein and his mother for a lunch on 18 June, and
presumably they then discussed their disagreement further. At any rate, he wrote to
Lady Ottoline on the next day to say that on the previous day he had felt ‘ready for
suicide’, and he went on:
All that has gone wrong with me lately comes from Wittgenstein’s attack on my work—I have
only just realised this. It was very difficult to be honest about it, as it makes a large part of the
book I meant to write impossible for years to come probably.

Somewhat later he must have written to Wittgenstein to say that he was ‘paralysed’
by Wittgenstein’s criticisms, and Wittgenstein replied (on 22 July) to say that he
was sorry to hear of this ‘paralysis’, and that what was needed was ‘a correct theory
of propositions’. Finally, looking back on this episode from nearly three years after-
wards, Russell wrote to Lady Ottoline in May 1916:
His [Wittgenstein’s] criticism, though I don’t think you realized it at the time, was an event
of first-rate importance in my life, and affected everything I have done since. I saw that he was

17
The initial parts of his manuscript, which do not mention the theory of judgment, were later published
as six articles in the journal Monist. The first three form what we now have as ‘On the Nature of
Acquaintance’ [1914a], and the next three are entitled ‘Definitions and Methodological Principles’, ‘Sensa-
tion and Imagination’, and ‘On the Experience of Time’. The original manuscript version of what became
these six articles was then discarded, so we do not know how much (if at all) Russell may have revised them
for publication. The editor of volume 7 of the Collected Papers judges that much of the fourth was newly
written, but that the others were mostly left unchanged.
212 M E TA P H Y S I C S

right, and I saw that I could not hope ever again to do fundamental work in philosophy. . . .
Wittgenstein persuaded me that what wanted doing in logic was too difficult for me.

He must have been in a low mood when he wrote this, for in the intervening period he
had of course written and published his ‘construction of the external world’ in KEW
and related articles, and this surely counts as ‘fundamental work in philosophy’. But it
is true that he had not returned to the writing of TK, and to what he must have
thought of as the problems ‘of logic’ that made him abandon that book. So this brings
us to the question: what was Russell’s real difficulty? For he never tries to explain it
to Lady Ottoline.
We know how Wittgenstein construed his own objection: he thought that Russell’s
theory of judgment failed to ensure that what is judged is a proposition with a definite
truthvalue. As he wrote to Russell, when arranging the lunch of 18 June:
I can now express my objection to your theory of judgment exactly: I believe it is obvious that,
from the prop[osition] “A judges that (say) a is in the Rel[ation] R to b”, if correctly analysed, the
prop[osition] “aRb Ú aRb” must follow directly without the use of any other premise. This
condition is not fulfilled by your theory.

It may surely be assumed that ‘aRb Ú aRb’ holds if and only if ‘aRb’ is a proposition.
Essentially the same objection can be found in Wittgenstein’s Notes on Logic, written
later in 1913 (p. 103, cf. p. 95)18 and in his TLP (5.5422). But quite why Russell’s
theory fails this condition is not made clear, either in what Wittgenstein has written or
in how Russell reacted to it. The obvious suggestion is the one that I have already
mentioned. In the old days of PoM Russell had assumed that any entity of any kind
could always be mentioned, and thereby be made the subject term of a proposition. His
theory of judgment in PP had apparently assumed that ‘loves’ could be treated as such a
term in the analysis of ‘Othello believes that Desdemona loves Cassio’. But if we take
this to imply that any other entity may be substituted for loving in that analysis, while
still preserving significance, then we license what is evidently nonsense. No one could
possibly believe that Desdemona Iago Cassio. The obvious remedy is to say, as already
suggested, that ‘loves’ should be assigned to a type that is one higher than the type
assigned to ‘Desdemona’ and to ‘Cassio’. So long as such type restrictions are observed
in the analysis of a judgment, then this source of nonsense is easily avoided.
One problem is that this answer seems to be so simple and straightforward. Surely
Russell could have thought of it himself, and not been so upset? Another problem is
that this point is not found in the revised version of his theory that (we assume) Russell
showed to Wittgenstein on or before 26 May, and which he presumably took to
contain the modification that was needed in order to meet Wittgenstein’s earlier
objection. But the revised version does contain several other new departures, most

18
‘Every right theory of judgment must make it impossible for me to judge that this table penholders the
book. Russell’s theory does not satisfy this requirement’ (p. 103). ‘The proper theory of judgment must make
it impossible to judge nonsense’ (p. 95).
 P RO P O S I T I O N S 213

notably that Russell now proposes to include an explicit mention of the form of what is
believed amongst the terms of the belief-relation. Thus a’s belief that b loves c is now
reported by

B (a, f, love, b, c)
where f is the form ‘xRy’ of a dual relationship. The reason for this new addition is not
entirely clear, but Russell does seem to connect it with the point that a belief may be
false. At any rate, he says
In an actual complex, the general form [e.g. ‘xRy’] is not presupposed; but when we are
concerned with a proposition which may be false, and where, therefore, the actual complex is
not given, we have only, as it were, the “idea” or “suggestion” of the terms being united in such
a complex; and this, evidently, requires that the general form of the merely supposed complex
should be given. (TK, p. 116)

So the suggestion now is that Russell hopes that by including the form he has met
Wittgenstein’s objection. In that case, the objection might well have been that the
mere list of constituents—e.g. ‘love’, ‘Desdemona’, ‘Cassio’—is not yet a proposition
with a truthvalue. It does not have that kind of unity that a proposition has.19
Moreover, the point seems particularly relevant where the proposition in question is
in fact false. For what Russell has said in the past is that it is the verb ‘loves’ that creates
the ‘unity’ of the proposition that Desdemona loves Cassio, and it does so by here
being what he calls ‘a relating relation’, as opposed to being merely a subject term in
some proposition about the relation.20 So it is argued that in the proposition ‘Othello
believes that Desdemona loves Cassio’ the word ‘loves’ either signifies a relating
relation or it does not. If it does, then that is to say that Desdemona is related to Cassio
by love, and hence what is believed is true. And if it does not, then there is no unity
in what Othello is said to believe, and hence no truthvalue. Either way, his belief
cannot be false.
Now we may set aside this last point about ‘relating relations’ as arising only from a
failure to understand Russell’s somewhat inappropriate vocabulary for describing the
difference between a proposition that is a ‘unity’ and a mere list of its constituents,
which is not. For clearly he must say that in the sentence ‘Desdemona loves Cassio’ the
verb ‘loves’ does signify a ‘relating relation’ both when the sentence is true and when
the sentence is false. So there is no special problem about a false belief. But one does
have to admit that, when writing TK, he seems to suppose that in the sentence
‘Othello believes that Desdemona loves Cassio’ the verb ‘loves’ is not functioning as
a ‘relating relation’. This is partly because there is another verb in this sentence that
clearly is functioning in this way, namely the verb ‘believes’, and partly because he

19
The suggestion is put most clearly in Hanks [2007]. Cf. also Candlish [2007: ch. 3] and Carey [2007:
ch. 2].
20
PoM, }54.
214 M E TA P H Y S I C S

wishes to deny that this sentence relates Othello to a single (‘unified’) thing, a
proposition, and so thinks it necessary to say that it only relates him to the several
constituents of that proposition. One can well see how Wittgenstein’s initial criticism
could have been directed to this point, and how Russell might have been tempted to
try to meet it by building in an explicit reference to the form, which shows how those
constituents are believed to be combined. But then it may be objected that merely
adding a form to a list of propositional constituents just produces a longer list; we still
do not have the kind of ‘unity’ that is needed to create a genuine belief. Russell was
partly aware of this point. He apparently thinks that merely listing the form is adequate
when either its verb is a one-place predicate, or it is a symmetrical relation, so that there
are not significantly different ways of putting the constituents together in that form—
different, that is to say, in that they create different truth conditions for the belief.21 But
where the relevant verb is a non-symmetrical relation, as with our standard example
‘loves’, he does accept that more needs to be said. Obviously the belief that Desdemona
loves Cassio is not the same as the belief that Cassio loves Desdemona, though in each
case the form of what is believed is the same, and so are the constituents.
This brings me to a new suggestion about just what Russell’s problem was, which is
due to Pincock [2008: 124–9]. Russell had always accepted that his account of belief
had to include an explanation of when the belief was true and when it was false, and his
idea was that this would depend on whether there did or did not exist the complex (or
fact) that ‘corresponded’ to the belief. From the time when he first contemplated a
theory of this kind, i.e. in his [1906/7], he had noted that there would be some
difficulty in explaining just what this ‘correspondence’ was (p. 46n.). The expositions
of his theory in 1910 and in 1912 had rather skated over this foreseen difficulty,
apparently taking the required correspondence to be obvious. But in TK he accepts
that he does have to spell out how to determine, from the account of what the belief is,
what the ‘corresponding fact’ has to be, and this is treated in some detail in chapter 5 of
part II.22 This leads him to build into the account of what Othello’s belief is the explicit
statement that it requires the existence of a loving-complex in which Desdemona
has first position and Cassio has second position. Thus what began as a belief in a simple
atomic proposition has apparently been transformed into belief in an existential
quantification ‘there exists a complex such that . . . ’ governing what Russell calls a
molecular proposition, in this case a conjunction ‘Desdemona is first and Cassio is
second’. He takes note of the second point, which gives him an excuse to postpone
further discussion to his proposed discussion of molecular belief in part III (p. 147).
He fails to take explicit note of the first point, which evidently complicates the

21
For the bulk of chapter 1 of part II, which is presumably what Russell showed to Wittgenstein, he
changes his example to one where the subordinate verb signifies a symmetrical relation, namely ‘is similar to’.
This is so that he can put off the complication that will be needed. (TK, pp. 114–18)
22
A crucial feature of the account has been anticipated earlier, in chapter 7 of part I, at pp. 88–9. This was
apparently written very shortly after his first meeting with Wittgenstein on 20 May. There may or may not be
a connection.
 P RO P O S I T I O N S 215

situation yet further. So part II, which was supposed to complete the analysis of atomic
beliefs, leaves off with a crucial problem unsolved. It would not be at all surprising if,
when Russell began to think about his part III, he realized that there was much more
of a problem here than he had at first envisaged. Pincock’s suggestion is that he did
realize this, that he found himself unable to resolve the problem, and that that is why he
abandoned the attempt to write this book.
This strikes me as quite a plausible suggestion, but if it is right then Russell gave up
prematurely. For he very soon discovered how to avoid this problem. It is done by
accepting that a belief-statement involves ‘two verbs’, in this case both the verb
‘believes’ and the verb ‘loves’. They can both be said to express ‘relating relations’ if
one likes, but this simply means that the sentences of which they are the verbs are
genuine sentences, with a truthvalue, and not mere lists. (Hence, if we are speaking
English, the order in which their terms are mentioned is significant.) As we have seen
(on pp. 209–10) this is just how Russell does present his theory in the later account of
PLA, and he does not suppose that recognizing ‘loves’ as a verb involves recognizing
the sentence ‘Desdemona loves Cassio’ as naming an entity, a proposition, to which
Othello is belief-related. On the contrary, his account in PLA insists that:
There are really two main things that one wants to notice in this matter that I am treating of just
now. The first is the impossibility of treating the proposition believed as an independent entity,
entering as a unit into the occurrence of the belief, and the other is the impossibility of putting the
subordinate verb on a level with its terms as an object term in the belief. (PLA, p. 226)23

He evidently does not suppose that these two claims are inconsistent, and this is surely
correct. Moreover, the problem of the ‘corresponding fact’ is now solved automatical-
ly. For we may say that the analysis is now given as

Believes (Othello, ([(lxy: x loves y] (Desdemona, Cassio)))

and of course this is a true belief if and only if

[lxy: x loves y] (Desdemona, Cassio)

i.e. if and only if

Desdemona loves Cassio

But the point to notice here is that this version of the theory which recognizes two
verbs did not have to wait until Russell wrote PLA in 1918 to be discovered, for he had
reached it much earlier. In PLA he credits to Wittgenstein the discovery of what he
calls a new form of proposition (‘a new beast for our zoo’, p. 226), and it is clearly just
the same point that he credits to Wittgenstein in what we have as our version of
chapter 4 of part I of TK, where he says

23
This passage is immediately followed by the one given above, on p. 209.
216 M E TA P H Y S I C S

It can be shown that a judgment . . . must be a fact of a different logical form from any of the
series: subject-predicate facts, dual relations, triple relations, etc. (TK, p. 46)

(A footnote adds that he has come to know this ‘through unpublished work of my
friend Mr. Ludwig Wittgenstein’.) No doubt this passage did not occur in his original
draft of TK, and presumably he had not reached this view when he wrote despairingly
to Lady Ottoline Morrell on 19 June 1913 (cited above, p. 211). But it had been
reached by the time that he was revising his chapter 4 a year later, for publication in the
Monist of October 1914.24 So why did he not then return to what was originally
planned as ‘a big book’? The answer may be quite mundane, for example a conjunction
of the two facts (a) that much of the material planned for TK had already been
incorporated in the book Our Knowledge of the External World, which was published
in August 1914, and (b) that the First World War broke out in August 1914, and
Russell had resolved to oppose it, which led him into a campaign which absorbed
much of his energy. But there are two further suggestions, both very speculative, of
deeper reasons why Russell might have abandoned his attempt to write TK. Each
of them introduces a problem ‘of logic’ which he continued to feel that he could not
resolve.

4 Two further speculations

The multiple-relation theory is first introduced in chapter II of the Introduction to PM.
Section 1 of that chapter announces the Vicious Circle Principle, section 2 aims to
show how that principle applies to propositional functions, and section 3 is then headed
‘Definition and Systematic Ambiguity of Truth and Falsehood’. It opens with the
observation that there appears to be a proposition about all propositions, saying that
they are all false, namely

8p(p is false)
But this would be a proposition that is itself false, and so would fall within its own
scope, which is something that the Vicious Circle Principle forbids. On the face of it,
the VCP claims that the attempted quantification over all propositions is an illegitimate
quantification, but what Russell says here is rather different. He says that we must
recognize that ‘the word “false” really has many different meanings, appropriate to
propositions of different kinds’ (p. 42). He then proceeds to try to convince us that the

24
Perhaps we can be a little more precise. The text of KEW says: ‘If we knew all atomic facts, and also
knew that there were none except those we knew, we should, theoretically, be able to infer all truths of
whatever form’ (p. 53 first edn, p. 63 second edn). A footnote added in time to be printed in the first edition
says: ‘This perhaps requires modification in order to include such facts as beliefs and wishes, since such facts
apparently contain propositions as components [and hence are not strictly atomic]’. The point is not here
credited to Wittgenstein, but it seems as if it should be. (A footnote added to KAD in 1917, for its reprinting
in Russell’s Mysticism and Logic (p. 207) does credit Wittgenstein with showing that the original theory of
NTF needs a modification.)
 P RO P O S I T I O N S 217

words ‘true’ and ‘false’ do have many different meanings, by sketching an account of
the many different kinds of truth that there are. This begins with truth of the simplest
kind, which is called ‘first truth’, and applies to atomic propositions. It goes on to
explain how truth of higher kinds may be generated from this, in a way that I shall give
in a moment. It then turns back to the atomic case, which arises when the proposition
in question is of the sort that may be a judgment of perception, and gives this
elucidation:
When we judge “a has the relation R to b”, our judgment is said to be true when there is
a complex “a-in-the-relation-R-to-b”, and is said to be false when this is not the case. This
is a definition of truth and falsehood in relation to judgments of this kind. (p. 43)

Notice that this defines the truth of an atomic judgment, in just the way that the
multiple-relation theory standardly does, and the following paragraphs make it quite
clear that Russell is here using that theory, for he comments that, on his account, ‘a
judgment does not have a single object, namely the proposition, but has several
interrelated objects’ (ibid.). So the position apparently is this. Truth for an atomic
proposition is defined in a way which does not treat that proposition as an object, by
saying that the proposition is true if and only if anyone who believes that proposition
will believe truly,25 where true belief is as just explained. Truth for other kinds
of proposition is then to be explained in terms of the relation of these other kinds of
proposition to atomic propositions.
The explanation goes like this. Assuming (for simplicity) that ‘fa’ is an atomic
proposition, Russell continues:
Let us call the sort of truth which is applicable to fa ‘first truth’. . . . Consider now the
proposition 8xfx. If this has truth of the sort appropriate to it, that will mean that every value
fx has ‘first truth’. Thus if we call the sort of truth that is appropriate to 8xfx ‘second truth’, we
may define ‘8xfx has second truth’ as meaning ‘every value for f^x has first truth’, i.e. ‘8x (fx
has first truth)’. (p. 42)26

The case for 9xfx is of course analogous, and he adds that ‘similar considerations will
enable us to deal with “not-p” and with “p or q”’ (p. 43). What he has in mind is
presumably this:

If ‘p’ has first truth or falsehood, then ‘p’ has second truth or falsehood, and ‘p’ is
second-true if and only if ‘p’ is not first-true.

25
Russell is not fully explicit on the relation between the truth of the atomic proposition, and of the
corresponding belief, and I have supplied it. A slightly different account (with ‘someone’ for ‘anyone’) is
given in Landini [1993: 380] and [1998: 288]. Notice that in either case we must assume that every atomic
truth is believed (or at least, in the revised treatment of TK, is understood), for without this assumption we
shall get the wrong results. Cf. note 13, above.
26
Observe that, in view of the preceding note, ‘every value of f^x has first truth’ must be taken as
implying ‘every value of f^x is believed (or at least, is understood)’. (Cf. Weiss [1995: 270].) Could Russell
have intended this? (Contrast the passage from PM, pp. 39–40 that is cited on pp. 83–4 above.)
218 M E TA P H Y S I C S

If both ‘p’ and ‘q’ have first truth or falsehood, then ‘pÚq’ has second truth or falsehood,
and ‘pÚq’ is second-true if and only if either ‘p’ is first-true or ‘q’ is first-true.
Generalizing from ‘first’ and ‘second’ to any kind of truth, we presumably have this:
Assume that ‘aRb’ has first truth or first falsehood.
Then ‘aRb’ has second truth iff 27 not-(‘aRb’ is first-true);
And ‘8y(aRy)’ has third truth iff, for all y, ‘aRy’ is second-true;
And ‘9x8y(xRy)’ has fourth truth iff, for some x, ‘8y(xRy)’ is third-true;
And ‘9R9x8y(xRy)’ has fifth truth iff, for some R, ‘9x8y(xRy)’ is fourth-true.
The kinds of truth differ with each addition of a truthfunctor or a quantifier. So what is
nowadays given as a recursive account of a single notion of truth is here used in an
explanation of how truth has many different kinds.
Moreover, this theory that there are different kinds of truth is evidently connected,
in Russell’s mind, with the way that it provides a solution to the paradoxes affecting
propositions. This is hinted at by his initial interest in the supposed proposition ‘8p(p is
false)’, but it is confirmed a little later when he explicitly applies to the Epimenides
paradox his claim that a quantification never has the same kind of truth as its instances
do. For he points out that it will follow that
Such a proposition as “all the judgments made by Epimenides are true” will only be prima facie
capable of truth if all his judgments are of the same order. If they are of varying orders, of which
the nth is the highest, we may make n assertions of the form “all the judgments of order m made
by Epimenides are true”, where m has all values up to n. But no such judgment can include itself
in its own scope, since such a judgment is always of higher order than the judgments to which it
refers. (p. 46)

Notice that here Russell slips from speaking neutrally in terms of different kinds of truth
into speaking instead of orders. His ramified theory of types does of course distinguish a
hierarchy of different orders (not of truths, nor of propositions but) of propositional
functions.
This has led to the following speculation.28 From this section of the Introduction to
PM we may infer that Russell hoped one day to show how the distinction into
different kinds, or orders, of truth could be made to yield the restrictions of his ramified
theory of types. The theory of types as we have it insists upon distinguishing proposi-
tional functions into different orders, and the definition of these orders is actually given
a little later in the Introduction to PM, in sections 4 and 5, without any reference back
to the earlier section 3 that we have just been noting. But one can guess at how the
earlier section might perhaps be used in justification of the later. Moreover the

27
‘Iff ’ abbreviates ‘if and only if ’.
28
The speculation is due to Sommerville [1980], considered with approval in Griffin [1985], and
endorsed in chapter 4 of Stevens [2005]. Its interpretation of PM is adopted in Landini [1993] and [1998:
chapter 10] independently of any speculations about TK.
 P RO P O S I T I O N S 219

hierarchy of orders of truth that is given in section 3 does rely on the multiple-relation
theory to provide a definition of the starting point, i.e. of truth for atomic propositions.
So the speculation is that that is what Russell wanted the multiple-relation theory for.
This explains (a) why he never develops this theory for any more than the basic case of
atomic judgments, for that is all that he needs, and (b) why he was so upset by
Wittgenstein’s criticisms, and why ‘they made a large part of the book I meant to
write impossible’. For he wanted to derive his theory of types from the different orders
of truth, and he wanted these orders to depend upon his account of truth for atomic
judgments. But Wittgenstein had claimed that an account of judgment must make
it impossible to judge nonsense, and the obvious way to meet this claim was to invoke
the theory of types in the account of what a judgment is. But, if this is admitted,
it seems that Russell’s plan must collapse into circularity. For if he cannot make the
first step without calling upon the theory of types, it will be of no use to go on later to
show why that theory must be correct.
Now it has to be admitted that there is here some confusion between what we call
the ‘simple’ and the ‘ramified’ theory of types. For it is only the simple theory that is
needed to meet Wittgenstein’s objection about judging nonsense, but it is the ramified
theory that one might aim to justify on the basis of the different orders of truth.29 But
one must remember that it is we (principally Ramsey [1925a]) who have made this
distinction between a simple and a ramified theory, and that this is something which
Russell himself never did. So perhaps, even if there is an escape route here, it would not
be too surprising if Russell did not notice it. The more important difficulties for this
speculation are these two. (i) There is no direct evidence that Russell aimed to deduce
the different orders of the ramified theory of types from the different kinds of truth that
he elaborates in section 3. On the contrary, there are many more kinds of truth in
section 3 than the ramified theory would need. For example, it has no need to
distinguish the formulae ‘p’ and ‘p’ as being of different orders.30 More important-
ly, (ii) the speculation is that Russell intended to do this deduction in the unwritten
parts of TK, but what we have by way of headings for these parts does not at all suggest
that the theory of types was to be treated anywhere in that book.31 But we do know
that part III was to treat of belief in molecular propositions (in its chapter 1) and very
probably of belief in general propositions (in its chapter 6). However, on the specula-
tion now in question those treatments would be irrelevant. It seems to me that these
difficulties do deprive the speculation of much of its attractiveness. I turn to another,
which is very different.

29
This point is acknowledged in Stevens [2005: 101–2].
30
The distinctions of section 3 are more closely allied to the treatment of truthfunctors and quantifiers in
PM*9, which is not connected with any theory of types.
31
The headings are given in Appendices A and C (pp. 179–202) of volume 7 of Russell’s Collected Papers
[1984].
220 M E TA P H Y S I C S

Why does Russell decide to include the logical form of the proposition believed
amongst the terms to which the believer is belief-related? The answer may be, as
already suggested, in order to help him specify what fact has to obtain if the believer is
to believe truly. Or it may just be that he has often said in the past that in order to
understand any proposition one must understand (by acquaintance, apparently) both the
constituents of that proposition and its form, i.e. the way in which those constituents
are combined. He is therefore trying to include amongst the terms of the belief-relation
all the things which the believer must understand if he is to have that belief. (If this is
right, it would explain why he also feels the need to expand this belief-relation still
further, in order to spell out the order in which the loving-relation relates its terms.)
But there may perhaps be a further motive: the logical form will be useful when we
come to deal with non-atomic beliefs, for we can put into it the extra complexity that
is needed in order to deal with these cases.32
Let us use our l-abstracts (from pp. 63–6) to represent forms. In this use they should
therefore contain only variables, and logical constants, but no ordinary propositional
constituents. Then Othello’s belief that Desdemona loves Cassio is represented by

B(a, [lFxy:Fxy]([(lzw : z loves w], b, c))

If he had believed that Cassio does not love Desdemona, this would be

B(a, [lFxy:Fxy]([(lzw : z loves w], c, b))

If he had believed that either Desdemona loves Cassio or vice versa, this would be

B(a, [lFxy:Fxy ÚFyx]([(lzw : z loves w], b, c))

And if he had merely believed that someone loves Desdemona, then we should have had

B(a, [lFy: 9x(Fxy)]([(lzw : z loves w], b))

We may add that his belief that someone loves someone is given by

B(a, [lF: 9x9y(Fxy)]([(lzw : z loves w]))

And finally, his belief that someone bears some relation to someone is simply

B(a, [9F9x9y(Fxy)]
In this last case what is believed is a purely logical proposition, i.e. one which contains
no non-logical constants. But it is a proposition and is not itself a form.33 I simply leave
it as a suggestion that Russell’s account might be filled out in this way, and something

32
I take the idea from Boër [2002]. Boër does not in fact suggest that it is what Russell himself might have
said, if he had continued work on TK. But he does suggest that it is a good way of extending Russell’s theory.
33
Russell himself does not (of course) use the l-notation for forms, but nor does he use his corresponding
cap notation, as one might expect. Instead he identifies the form ‘[lFxy:Fxy]’ with the existential quantifi-
cation ‘9F9x9y(Fxy)’. This is an unfortunate move on his part, and is best ignored.
 P RO P O S I T I O N S 221

like this could have been in his mind when he expanded the belief-relation by explicitly
including a form.34 There is no real evidence for this speculation, but it does provoke a
further thought.
In the later version of his theory that is given in PLA, Russell has once more dropped
all mention of the logical form. This may be because, now that he has acknowledged
the ‘two verb’ nature of a belief-statement, he has seen that the form is no longer
needed for the purpose for which he first introduced it. But there may be more to it.
He may have become convinced by his reflection on what he took to be Wittgen-
stein’s criticism that he no longer knew what to say about how we do understand the
logical forms which, in practice, we use all the time. In TK itself he insists that we must
have acquaintance with logical forms (though he adds: ‘possibly in an extended sense of
the word “acquaintance”’, p. 99). This is needed, he thinks, both to explain how we
understand sentences that assert propositions of a given form, e.g. of the form ‘xRy’,
and to explain how we understand such terms of logic as ‘relation’. He admits that this
kind of acquaintance is elusive:
As a matter of introspection, it may often be hard to detect such acquaintance; but there is no
doubt that, especially where very abstract matters are concerned, we often have an acquaintance
which we find it difficult to isolate or to become acquainted with. The introspective difficulty,
therefore, cannot be regarded as fatal. (p. 99)

As he acknowledges, more needs to be said, but he cannot now explore the subject any
further. One simply has to accept that
Acquaintance with logical form, whatever its ultimate analysis may be, is a primitive constituent
of our experience. (Ibid.)

Well, one can imagine what Wittgenstein would say to that! In fact we do not have
to imagine it, for we have what Wittgenstein did say, probably with this claim of
Russell’s in mind. In his Notes on Logic, written later in 1913, we read:
There is no thing which is the form of a proposition, and no name which is the name of a
form. Accordingly we can also not say that a relation which in certain cases holds between
things holds sometimes between forms and things. This goes against Russell’s theory of judg-
ment. [1961: 105]

While Wittgenstein does not here specify what relation he has in mind, it seems likely
that he is thinking of Russell’s relation of acquaintance, which was supposed to hold
between a person and an individual, between a person and a universal, and between
a person and a logical form. Wittgenstein thinks that acquaintance with a logical form
is an impossibility, and surely we will all agree that ‘acquaintance’ is not at all a good
model for the kind of understanding that is here involved.

34
I have omitted a detail of Boër’s account, which he includes for technical reasons that affect a further
development which he pursues but which I am ignoring.
222 M E TA P H Y S I C S

Now we do not know for certain that Wittgenstein said anything about logical
forms in that stormy meeting of 26 May. But it seems very probable that he did, since
the addition of logical forms is the most prominent new departure in the revised
version of the multiple-relation theory that (we presume) Russell sent to him. Even if
he did attack what Russell had said about forms, we do not know how his attack went,
and we do not know what it was about that attack that led Russell some days later to
decide that his planned book TK simply had to be abandoned. We do know that
Wittgenstein took his central point to be that an account of judgment had to ensure
that it is impossible to judge nonsense, but it may well be that Russell had a different
view of what was important in this criticism. And it really is difficult to believe that
Wittgenstein said nothing about Russell’s supposed acquaintance with logical forms.
Moreover, we should take into account how Russell’s thought developed thereafter.
Here the first point is that, although he abandoned TK, he did not abandon
his multiple-relation theory of judgment. He did alter it, by now saying that a
belief statement contains ‘two verbs’, and acknowledging that this new version was
due to what Wittgenstein had said. But the general outline of the theory was
still maintained in PLA. However (i) he made no further attempt to expand the theory
beyond the atomic case, and (ii) he no longer claimed that a logical form is
needed amongst the terms of the belief-relation. These points may be connected,
and I have suggested how. But it is more important that he appears simply to have
dropped the idea that one is acquainted with logical forms. It does not recur in PLA,
where instead Russell sketches a rather different account of how we come to under-
stand what a form is, namely by abstracting from many propositions, independently
understood, that share that form (pp. 237–8). This is accompanied by the beginning
of a different approach to acquaintance with universals, for—as we saw on pp.124–5—
Russell is now realizing that in this case too mere acquaintance is not enough by
itself. In fact he is on the way to the later view of AMind [1921a] that here acquaintance
is not really needed at all. If this is the result of his further reflection on some point that
Wittgenstein had made on 26 May 1913, then that meeting was more fruitful than is
often supposed.
I end this section with a brief further comment on the theory that Russell presents in
PLA. He now says that a belief-statement contains two verbs, and the structure that he
has in mind for it would seem to be this

Believes (Othello, ([lxy: x loves y](Desdemona, Cassio)))

He says that this structure is something new. But it is not new at all, for any proposition
of the form M xFx contains two verbs, one in its M -part and one in its F-part. For
example ‘Desdemona loves just one person’ is taken to have the structure

91x(Desdemona loves x),

 P RO P O S I T I O N S 223

where ‘91’ and ‘loves’ are both what Russell calls ‘verbs’.35 What is new in Russell’s
analysis is not the ‘two-verb’ structure but the fact that the relation ‘believes’, unlike
any concept that figures in the logic of PM, creates an intensional context. This topic
deserves a new section.

5 Problems of intensionality
It has been explained earlier (pp. 69–71) what is meant by saying that a sentential context
for a name or a predicate or an embedded sentence is an extensional context for that
expression. A context which is not extensional is said to be intensional. Contexts
which introduce the notion of belief are prime examples of intensional contexts. For
instance, according to Shakespeare’s play Cassio is in fact the man who will next govern
Cyprus, after Othello, but that is not something that Othello himself knows or believes.
So it appears that, despite the truth of this identity, from the premise that Othello believes
that Desdemona loves Cassio we cannot deduce the conclusion that he believes that
Desdemona loves the next governor of Cyprus. Assuming that this is so, the context
‘Othello believes that Desdemona loves . . . ’ is an intensional context for names. Similarly,
the truth is that Desdemona loves Othello, and only Othello, and hence, for all x,
Desdemona loves x if and only if Othello is x. That is, the two predicates ‘Desdemona
loves . . . ’ and ‘Othello is . . . ’ have the same extension. But despite this we obviously
cannot interchange them in the context ‘Othello believes that—Cassio—’. This context is
therefore an intensional context for a one-place predicate. Finally, it is clear that ‘Othello
believes that . . . ’ is an intensional context for a sentence, for we are given that Othello
believes one falsehood, but it obviously does not follow that he believes every other.
It is the first point, about names or other referring expressions, that is most relevant
to Russell’s account of belief. For it seems that, on his account of what Othello’s belief
is, it relates him directly to the person Cassio. If so, then we should be able to substitute
for the name ‘Cassio’ any other expression that in fact refers to Cassio, without
changing the belief that is ascribed to Othello. For in either case we have still got
a relation to the same person. Now one can actually accept this conclusion, though
one can also reject it (as suggested above), because the relevant English sentence is
ambiguous. The sentence is

Othello believes that Desdemona loves the next governor of Cyprus,

and the question is whether the definite description ‘the next governor of Cyprus’ is to
be understood as giving us how Othello is thinking of the man in question (in which

35
The point may be made more obvious by complicating the initial verb ‘there is just one thing x such
that . . . ’ to what Frege gives as its analysis, namely ‘there are as many members of the class whose only
member is 0 as there are things x such that . . . ’. Using ‘’ for ‘as many as’, this is
([lx:x[ {0}], [lx: Desdemona loves x])
(Incidentally, there are three verbs here, not only ‘’ and ‘loves’, but also ‘[’.)
224 M E TA P H Y S I C S

case the sentence is false), or just as telling us which man Othello is thinking of, but not
how he is thinking of that man. In this case, where we have an explicit definite
description, the ambiguity can be revealed as an ambiguity over the scope of that
description. For brevity, put

‘Fx’ for ‘x governs Cyprus next after Othello does’.

Then, using the quantifier ‘Ix: Fx’ for ‘Concerning the thing x such that Fx’, the scope
ambiguity is that between
(Ix:Fx)(Othello believes that Desdemona loves x)
Othello believes that (Ix:Fx)(Desdemona loves x)
The English sentence is ambiguous between these two. In the situation envisaged, the
first is apparently true (as Russell’s theory seems to require), while the second is false.
Whether the same ambiguity persists when we have an ordinary name, ‘Cassio’, in
place of the explicit definite description, is an issue which may certainly be debated.36
This point has given rise to what is nowadays called the theory of ‘Russellian
thoughts’ (or ‘Russellian beliefs’), as opposed to the ‘Fregean’ version of thoughts.
For in Frege’s view the thought would contain, not the object referred to, but the sense
of the name or description that refers to it. So the thought will usually be different
if different referring expressions are employed. By contrast, the Russellian theory is
supposed to be that the thought contains the actual object referred to, so that it will
remain the same thought so long as it is still the same object that is in question. There is
then an obvious problem for this Russellian theory: what shall we say if it turns out that
the referring expression fails to refer? Will it have to follow that there was then no
thought? Or should we say that a thinker must know that he is thinking something, even
if he may not know how that thought should be analysed? But a little reflection on this
problem shows that the theory in question has been misnamed. It is not actually the
theory which Russell himself held, though it does resemble his theory.37

36
In the theory presented by Quine in his [1960a], the ambiguity is revealed as the distinction between
these two regimented versions:
Othello believes, of the next governor of Cyprus, that Desdemona loves him
Othello believes that Desdemona loves the next governor of Cyprus
Quine thinks that the same ambiguity persists, and has the same diagnosis, when the name ‘Cassio’ replaces
the description. That is, there is a difference between
Othello believes, of Cassio, that Desdemona loves him
Othello believes that Desdemona loves Cassio
His idea is that one can ‘quantify in’ to a position governed by ‘of ’, but not to one that is governed by ‘that’.
But on his account this is not a matter of scopes, for names do not have scopes.
37
The theory is mainly due to Gareth Evans, in his [1982], especially chapters 2 and 4, but many others
have also discussed the idea. (I do not mean to imply that they have been under any misapprehension about
Russell’s own views.)
 P RO P O S I T I O N S 225

Although Russell’s chief example in PP (and in PLA) uses the ordinary names
‘Desdemona’ and ‘Cassio’, this is not how he means to be taken. For he is talking of
a belief held by Othello, and of course he must suppose that Othello understands this
belief. But, according to his principle of acquaintance ‘every proposition which we can
understand must be composed wholly of constituents with which we are acquainted’,
and Othello is not (in Russell’s sense) acquainted with Desdemona or with Cassio.
Rather, he knows them ‘by description’, no doubt descriptions which relate them to
his own sense-data. However Russell does not mean to be discussing the belief in a
complex proposition, containing two definite descriptions, which would need several
quantifiers and bound variables for its proper analysis. He means to be discussing the
simplest kind of case, which is belief in a genuinely atomic proposition. So he is letting
the familiar names ‘Desdemona’ and ‘Cassio’ stand in for what should really be ‘logical
proper names’, which is after all an illustrative technique that he very frequently
employs.38 Now, as we have seen, his understanding of logical proper names ensures
that they cannot be empty, and that they cannot appear in an informative statement of
identity. So in their case the problem over empty names cannot arise, and all belief
contexts for such names must be extensional. For in their case if ‘a ¼ b’ is true then
anyone who understands those names must find this truth a triviality. So of course he
will be ready to use either of them, without differentiation, when telling us of his
beliefs.
It results that, given Russell’s very restrictive account of what qualifies as a name,
belief-statements do not after all create intensional contexts for names. They do create
intensional contexts for predicates and for sentences, but as we have noted (p. 70) he
does indeed accept that there are such contexts. In his logical system one could
consistently add a principle of extensionality for first-level predicates, namely

8x(Fx $ Gx) ! (M xFx $ M xGx),

but in 1910 Russell does not in fact add any such principle, and evidently does not
believe it. It is true that in the second edition of Principia Mathematica (i.e. 1925) he does
consider adding the principle, but he recognizes that it is controversial, and he does not
give any explicit consideration to the question of how many of our ordinary ways of
talking and thinking are thereby ruled out. On the face of it, they are rather numerous.
But this is a topic that I shall take up again, from a more general point of view, in
section 6 of Chapter 14.
We may conclude that there is nothing seriously wrong with the theory of belief
that Russell proposed in 1910, in ‘On the Nature of Truth and Falsehood’, repeated
with only verbal changes in PP, and gave once more in a perhaps revised version in
PLA. What he says may be defended against the usual objections, provided that we are
prepared to find belief-relations at all levels of the theory of types, and to accept his

38
Recall the much-used example ‘Scott is the author of Waverley’. Russell’s discussion always presumes
that the name ‘Scott’ is here standing in for a logical proper name, though of course it is not really one.
226 M E TA P H Y S I C S

restricted view of what to count as a genuine name. It is true that what he says there is
very incomplete, for it considers only the simplest case, where the proposition believed
is an atomic proposition, though most of our ordinary beliefs have to be analysed by
him as non-atomic. An extension that would include more complex beliefs had been
planned for TK, but—for whatever reason—that never did materialize, and the
published theory never got beyond the PLA version. However the theory was
abandoned altogether only one year later, in 1919.

6 Propositions as mental entities

Let us recapitulate the story so far. When Russell broke away from idealism in 1898, he
took up a strongly realistic theory of propositions. They were grasped by the mind, but
did not depend upon anything mental for their existence. Indeed, Russell thought of
them as existing ‘in the world’ in a way which made it difficult to distinguish a (true)
proposition from a fact, and so led to an obvious problem over false propositions. It was
this problem that he invoked, from 1907 on, as the reason for moving to his next
theory of propositions, which aimed to avoid assuming their existence altogether. The
role that they had been expected to play, e.g. as objects of belief, was now to be taken
by what had been the several constituents of a proposition, but which could equally
well be viewed as the several constituents of a fact. This attempt to dispense with
propositions had some partial success, but it was only partial, as we have seen. So we
now come to Russell’s third and final theory of propositions, which once more accepts
them as things that do exist, but returns to the idea that they are mental entities. The
motivation for this new theory of propositions is neutral monism, and it begins when
Russell’s neutral monism begins, i.e. with the article of 1919 entitled ‘On Propositions:
What they Are and How they Mean’ (henceforth Propns).
The previous multiple-relation theory could have introduced appropriate mental
items, e.g. mental images, if Russell had seen any need for it. He could have held
that a belief is in the first place a relation between the subject and a number of such
images, and these are in turn related to the constituents of the relevant fact by
‘meaning’ them.39 He did not do so, partly because he then saw no useful role for
such intermediate entities, but more importantly because he thought that they would
introduce a problem. The mental images apparently introduce some kind of a ‘veil’
between the thinking mind and the objects that it aims to think of, for the objects of
thought are not the mental images themselves, but—on this theory—would be seen
‘through’ them.40 (As we noted, p. 204, Russell was initially inclined to say that we
could think directly of Mont Blanc itself, with all its snowfields, and that the thought
did not have to go via some mental entity that represented Mont Blanc.) But now that

39
In his [1906/7], when he first suggested a multiple-relation theory (but did not adopt it), it was along
these lines.
40
This objection is made explicit at KAD, p. 222.
 P RO P O S I T I O N S 227

neutral monism has led to the abolition of the thinking subject, Russell evidently
supposes that such mental entities do have to be invoked. His reason appears to be that
there must be something that is happening in a person’s mind at the time when he is
(actively) believing something, and that the so-called ‘feeling’ of assent, which he
identifies with the believing, is not enough by itself. For the assent must be directed to
something, and this ‘something’ has to be available to it at the time. So it cannot, for
example, be a complex of items that are now in the distant past. (Cf. AMind, p. 21 and
pp. 233–4.) This line of argument is not altogether convincing. It is better to observe
(as Russell does, at Propns p. 307) that for logical and other purposes we can hardly
avoid talking of propositions, and that if we take a proposition to be something in the
mind when that mind assents to it then there will no longer be any problem over false
propositions. For falsehoods in the mind are quite acceptable; what causes difficulty is
the idea of there being falsehoods in the world. But we do still have the consequence
that a proposition’s only form of existence is its occurring in some mind (either by
being believed, or more generally just by being understood). Russell claims to avoid
this restriction by including possible mental occurrences. He says that ‘we may identify
propositions in general with the contents of actual and possible beliefs’ (AMind, p. 241).
But we should note (a) that this makes use of a modal notion which Russell usually
rejects (p. 270 below), and (b) that there is still a problem over whether it provides
enough propositions for his logic and mathematics (Chapter 13, section 4). For the
present, let us put these doubts aside.
The mental items which compose propositions, e.g. images, must have ‘meanings’.
For example, an image will (usually) ‘mean’ the thing that it is an image of. It is these
‘things meant’ that were previously regarded as constituents of propositions, but now
they are constituents of a fact in the world, namely the fact that makes that proposition
true or false, as the case may be. For Russell now thinks that the mental item which is
the proposition that P, will always ‘refer’ to a fact, either to the fact that P or to the fact
that P, whichever of these happens to be a fact. To put this the other way round,
there will be a single fact in the world that is ‘referred to’ both by the mental
proposition that P and by the mental proposition that P, for each proposition will
contain just the same mental items (e.g. images) that ‘mean’ the constituents of that
fact. 41 Where the proposition that P is an atomic proposition, it will be said to be a
positive proposition, and its negation P will be a negative proposition. Correspond-
ingly, the fact that they both refer to will then be a positive or a negative fact according
as P is true or P is true. It is admitted that, where P is not atomic, the distinction
between negative and positive is not so simple, but in any case this scheme clearly
commits Russell to the existence of negative facts. The existence of negative facts is
controversial, but I postpone discussion of the issue until (in Chapter 14) we come to
Russell’s discussion of facts in PLA.

41
The idea is already present in PLA [1918b: 187–8], where it is credited to Wittgenstein.
228 M E TA P H Y S I C S

Another consequence of the new system is that the old principle of acquaintance
now has no work to do. The old principle was that if a subject understood any
proposition then that same subject had to be acquainted with each constituent of the
proposition. We now have to rephrase this so that it does not assume the existence of a
subject, but that is quite straightforward: if in any mind there occurs a proposition
which is understood, then there also occurs in that mind an acquaintance with each
constituent of the proposition. However, since the proposition is now a complex of
mental items, e.g. images, all that this now requires is that the mind is conscious of its
own images, and that is hardly controversial. But does the mind also have to contain an
acquaintance with those worldly items that the images mean? One might be tempted
to suppose that the answer should be ‘yes’, simply on the familiar grounds that one
cannot understand a proposition without knowing what it means.42 But should one
still take this to imply that the existence and the identity of the thing meant cannot ever
be a matter of doubt? This is not at all clear, for the thing meant may be known only by
description, e.g. as ‘the thing that looks like that’, where one points to the image.43
Such knowledge need not be immune from error in the way that knowledge by
acquaintance was supposed to be. If this is right, then what used to be a constraint on
what could be taken as a constituent of a proposition, and might now be viewed as a
corresponding constraint on the constituents of the fact that that proposition refers to,
has in fact been relaxed.
We naturally think of propositions as expressed in words, and in that case a proposi-
tion as a mental item would consist of mental occurrences of words, either explicitly
pronounced or simply thought. But Russell also wishes to claim that the relevant mental
items may be not words but images, and in fact he thinks of this as the more basic case.
For a proposition made up of words (i.e. a word-proposition) is said to ‘mean’ a
proposition made up of images (i.e. an image-proposition), and it is the latter that
then directly ‘means’ the fact that they both refer to. But word-propositions and the
image-propositions that they ‘mean’ do not necessarily match one another item by
item, for words can say more than images can, though this deficiency in images can to
some extent be remedied by the variety of belief-feelings that there are. Believing is
a matter of having a particular feeling towards a proposition, which in the basic case is a
feeling of assent to the proposition, but other feelings are also possible, most obviously
dissent. This is relevant because, on Russell’s account, there is no image corresponding to
the word ‘not’, and so no image-proposition is a negative proposition. But we can
dissent from a (positive) image-proposition, and this has the same effect as assenting
to the corresponding word-proposition, modified by adding the word ‘not’. In a similar
way, an image-proposition merely depicts a state of affairs, and contains no indication of
pastness or presentness or futurity, whereas a corresponding word-proposition will

42
Cf. KAD, p. 219, noted and commented on earlier on p. 113.
43
As will soon emerge, the particular description suggested here will often be too simple to do by itself all
that is needed.
 P RO P O S I T I O N S 229

normally have an explicit tense. But again, what is put into words in the word-
proposition corresponds to a suitable kind of feeling that may be attached to the
image-proposition, e.g. either a memory or an anticipation, which will involve
the feeling that something like this did occur or will occur. Russell does not go so far
as to imply that there will be a different belief-feeling for each different way of
combining a number of atomic propositions by means of logical connectives, though
he does toy with the idea that there is a special belief-feeling for disjunctive propositions
(Propns, p. 311).44 But it seems best to say that it is only word-propositions, and not
image-propositions, that can occur as subordinate clauses in other longer propositions.
In any case, in Russell’s new theory, it is the image-propositions that are simple and
basic, and word-propositions that are explained in terms of them and that may also be
more complex. But for completeness I should add here that Russell also allows for a
proposition to be made up partly of words and partly of images. He further allows that it
may contain an actual sensation as the subject which it concerns, and which would be
referred to by the word ‘this’ in a word version of the proposition.
So much, then, in exposition of the overall position that Russell adopts in his article
Propns of 1919, and in the book AMind that followed quite soon in 1921. Though
there are some slight differences in the details of these two treatments, they are of no
real importance, and we may continue to focus mainly on the shorter article Propns.
We may also continue to concentrate attention on nice simple examples, which one
may assume to be atomic propositions. Let us begin, then, by asking: why does Russell
place such emphasis on images as more basic than words? Is it a good idea?

7 The role of images

One consideration which influenced him (and which is in fact more clear in AMind
than it is in Propns) is the thought that it is not only human beings that have minds. For
other animals too may have those ‘feelings’ towards propositions that we call belief and
desire. For example, a cat may expect to be fed, a dog may fear that its owner will be
angry, a horse may remember that this road leads to home, and so on. Since such
animals do not use language, their beliefs and desires cannot be ‘feelings’ directed
towards what can only be formulated in language. But they are directed to proposi-
tions, and so we must allow that propositions can also be formulated in some different
way, and then the idea that they can be formulated in images is perhaps the most
obvious suggestion. It is easy to suppose that animals can form images, so they could
also have ‘feelings’ towards propositions formed from them, and this point may be used
to explain their behaviour. But, as Russell well knew when he wrote AMind, there is a
rival ‘behaviourist’ school of thought which sees no need to posit such mental events as
beliefs and desires, construed in this way, in order to explain animal behaviour. The

44
This idea will be developed further in the much later work Inquiry into Meaning and Truth [1940:
chapter 5].
230 M E TA P H Y S I C S

anthropomorphic theory just sketched will surely allow that it is past experience that
causes the animal to have the beliefs and desires that it (supposedly) does have, and that
these in turn then cause its behaviour. But the behaviourist theory just cuts out these
(supposed) beliefs and desires, as an idle intermediary cog, and claims that all that we
require is a direct causal relationship between the animal’s past experience and its
present behaviour. There is no real need to suppose that it somehow ‘understands
propositions’, expressed in any form at all. As I say, Russell was well aware of this rival
theory, for it was popular amongst the American neutral monists that he studied at the
time. He rejected it as a theory of human behaviour, because it simply sets aside all that
we can know of our own minds by introspection. But he could not rely upon the same
reason for rejecting it when applied to animals, and so he was aware that the argument
that we need to posit image-propositions in order to explain animal behaviour was not
a strong one. Let us return to humans.
A standard feature of British Empiricism, as exemplified by Locke and Berkeley and
Hume, is the doctrine that words have meaning by standing for mental items, which
they called ideas. Words may be said to stand for things in the external world, but they
do so only indirectly; in the first place they stand for ideas, and then the ideas may in
turn stand for external objects. It is true that one does not find in the British Empiricists
(or in others of the period) anything much by way of a reason for bringing in ideas in
this way, but it is a tradition with a long history. As Russell notes, the founder of
behaviourism, J. B. Watson, apparently denies the existence of such images (or ‘ideas’)
associated with particular words, but this looks (at first) as if it leads to a lacuna in his
account of how a child initially learns a language. The broad outline of his account is
this. When the child’s attention is focused on some particular object, say his box, the
nurse says the word ‘box’. Frequent repetitions of this conjunction set up in the child’s
mind an association between the object and the word, and he too begins to say ‘box’
when the box is presented to him. This is the beginning of the idea that the word ‘box’
has a meaning for the child, namely that it means that particular box. So far, we have
not needed to suppose that he also has an idea of the box. But the next stage is that
the child begins to say ‘box’ when the box is not in fact presented to him, but he wishes
it to be, and here a doubt may set in. Watson supposes that the previous experience,
establishing an association between the object and the word, is by itself enough to
explain this stage too.45 But an opponent may say that here you need to bring in
something else, namely an image of the box, which has also become associated with the
box, and occurs here although the box itself is not present. Indeed, what makes the
wish a wish for the box, rather than for anything else, is just that it is associated with an
image of the box.

45
Watson, Behavior [1914: 321–34] aims to be a full account of language-learning. The crucial paragraph
is cited in full on p. 292 of Russell’s Propns. My summary abbreviates a great deal, and reproduces only the
essential points.
 P RO P O S I T I O N S 231

Russell’s comment is that the opponent has not proved his case, for we know what
we wish for when we know what will satisfy the wish, and we can know this just as
well by knowing a word for the object as by producing an image for it. Once again the
image, as a supposed intermediary, may simply be cut out as superfluous, and the
explanation will still stand (Propns, pp. 296–9). This comment appears to be correct.
One may say that a simple word such as ‘box’ means a certain object (for the child) so
long as the utterances of that word bear a suitable relation to the object in question.
The suggestion so far is that the word should be prompted by the presence of the
object, and should be used in the absence of the object to express a desire for it.
We may add that when others use the word then the child’s attention is directed to that
object. Other connections of a like sort may be suggested. In general terms, the point is
that the individual comes to associate that word with that object (or those objects), and
that this association governs both his own use of the word and his understanding of its
use by others.46 In all this there is no need to bring in a further association between the
word and an image. Russell does believe that (in suitable cases) the word is associated
with an image, and that an image may be said to have a ‘meaning’ in just the same way
as a word does, namely by being suitably associated with a particular object or event, or
a general type of objects or events. The one obvious difference is that the image will, in
a suitable way, resemble the object(s) or event(s) that it means, as a word will not. But it is
not at once clear that this is a point of any importance.
While Russell accepts that this is the dialectical situation, still it is clear that in his
own opinion images do have a more basic role than words. He expresses his overall
view of the behaviourist position, which denies the relevance of any images that there
may be, in this way:
I do not wish to deny that much action, perhaps most, is physically explicable, but nevertheless it
seems impossible to account for all action without taking account of ‘ideas’, i.e. images of absent
objects. If this view is rejected, it will be necessary to explain away all desire. (Propns, p. 297)

However, the last sentence is given no adequate support (either here or elsewhere).
Even when we know what we desire (which is not always the case), there is no good
reason to suppose that it has to be given by an image, though perhaps it quite often is.
And, to come back to our proper topic, even if images were essential to desires, it
would not follow that they were essential to our use of language. For surely words
may relate to objects directly, and without going via any mental images. I add that
Russell seems to suppose that images play an essential role in more sophisticated uses of
language, for example when giving or receiving a narrative account of some recent

46
We are concerned here with what Russell calls ‘object-words’, which include both words for particular
objects or events and words for universal kinds of objects or events (e.g. ‘rain’). In his earlier way of speaking,
these are words which ‘have meaning in isolation’. The idea is that the understanding of such words does not
yet involve the understanding of the grammatical constructions of whole sentences.
232 M E TA P H Y S I C S

happening (AMind, pp. 201–3), but again one cannot see that he offers any worthwhile
argument.
The case for there being such things as image-propositions is therefore unproven,
and moreover they certainly create problems which Russell seems only partly aware of.
Apparently he thinks that they have a claim to being especially simple, because the
image will resemble what it means, and a word does not do this (Propns, p. 303). But
this ignores several complications, as we may see by considering one of his examples.
He is considering the case of a visual memory-image, and he says:
I call up a picture of a room that I know, and in my picture the window is to the left of the fire.
I give to this picture that sort of belief which we call ‘memory’. When the room was present
to sense, the window was, in fact, to the left of the fire. In this case I have a complex image,
which we may analyse, for our purposes, into (a) the image of the window, (b) the image of
the fire, (c) the relation that (a) is to the left of (b). [In this case] the objective [i.e. the fact referred
to] consists of the window and the fire with the very same relation between them. [That is to say,
in this case the image-proposition is true.] (Propns, pp. 315–16)

We may first ask: what is it about (a), the image of a window, that makes it ‘mean’ that
particular window that it does, rather than any other window in the world which looks
much the same? Part of the answer is that the image is in fact a memory-image, and was
derived from a view which included just that window and no other. But this is not
enough by itself, for the image might be derived from memory and yet not be being
used here in order to recreate that memory, but perhaps to illustrate some claim about
all windows that look like that. So we must add that the imager is in this case intending
the image to ‘mean’ that window in particular, i.e. the one that was the original cause
of this memory-image. As Russell himself has said earlier: ‘The question what a given
image “means” is partly within the control of our will’ (p. 303). He illustrates the point
by observing that the same image can be used to represent a universal or a particular,
and he might have added that it is also the user’s intention that determines which
universal is in question (for the same image will exemplify many different universals),
or which particular (and, if relevant, as seen at what time, or under what conditions, or
from what angle, and so on). Obviously, the same considerations apply to (b), the
image of the fire. Let us turn to (c), the relation between these two images.
Russell’s comments show that the relation which he has in mind is just that the
window in question is to the left of the fire in question. But there will be many other
spatial relations which his image will illustrate, both more specific relations (e.g. that
the window is more than three feet to the left of the fireplace), and less specific relations
(e.g. that there is some space between the window and the fireplace), and relations that
are just different (e.g. that the top of the window is higher than the top of the fireplace).
One cannot tell, just by considering the image, which of these relations it is to be taken
as illustrating, nor therefore what fact has to obtain for the supposed image proposition
to be true. Moreover, we have been given no particular reason to suppose that the
relevant relation is a spatial one. The image could be taken as claiming that the window
 P RO P O S I T I O N S 233

and the fireplace are painted the same colour, or that the window has shutters while
the fireplace does not, or that they are of different periods and do not go well with
one another, and so on and on and on. There is no definite claim which an image-
proposition puts forward, save perhaps the very vague claim that things are or were
in some way like this. To obtain a definite claim one again has to ask the imager
how he understands it. This makes an image-proposition very different from a word-
proposition.
In each case, the words or images which are being used to make a reference to
particular objects will very probably not show by their own nature just which objects
are the ones intended. Given a particular context of utterance or thought, it may be
fairly clear which person called ‘John’ the speaker is speaking of, or which of the many
people who can look like ‘that’ the thinker is thinking of. But this is not always the
case, and whether we are dealing with words or with pictures it may be necessary to
enquire further into the believer’s intentions in order to find out which objects he
means. However in the case of words there is often no further uncertainty over just what
proposition is being put forward, nor what has to be the case if it is to be true. Anyone
who understands the language will thereby know what is being said about those
objects. With pictures or images that is never so, for a picture or an image does not
by itself make any claim at all, and even when it is prefixed by ‘This is how things are’ it
still does not specify in what respect things are claimed to be ‘like this’.
Russell’s ‘image-theory’ of meaning, if one may call it that, was reached at about the
same time as Wittgenstein was elaborating the ‘picture-theory’ of meaning that we find
in his TLP. They each reached their theory independently of the other, though there
are very clearly similarities between them,47 and in each case there is a commitment to
a ‘correspondence’ theory of truth. That is, the simple image-proposition (or picture-
proposition) is true if and only if there is a corresponding fact, consisting of the objects
meant by the images (or pictures) in a relation which is ‘meant’ by the relation between
those images (or pictures). In the case which Russell takes to be central, the relation
between the images is a spatial relation (i.e. ‘to the left of’), and the desired relation
between their objects is the same spatial relation.48 He toys with the idea that a temporal
relation between events, such as ‘precedes’, might perhaps be represented in the same
way by the same temporal relation between their images, but he hesitates to affirm it
(Propns, pp. 318–19). He might perhaps have gone on to suggest that visual images can
equally reproduce the colour relations between their objects, and maybe auditory or
tactual images can mirror other perceived relations in the same way (e.g. temperature
relations?). But he accepts that most relations between objects cannot also hold,

47
Russell had written Propns before he saw Wittgenstein’s TLP, which was in June or July 1919. (For
evidence, see Blackwell [1981: p. 30, n. 67].)
48
A spatial relation between two visual images is presumably a two-dimensional relation. The
corresponding relation between what those images mean is presumably three-dimensional. This is a
discrepancy which need not perturb us.
234 M E TA P H Y S I C S

without alteration, between their images, and he seems to leave it open whether such
relations can be represented in any way by image-propositions. In any case, he accepts
that it is only rather simple propositions that have an image-version, as well as a word-
version.
A word-proposition never employs the same relation between words as it claims
to exist between the objects that those words mean. Indeed Russell says: ‘in the phrase
“A is to the left of B”, even if we treat “is-to-the-left-of ” as one word, we have a fact
consisting of three terms with a triadic relation, not two terms with a dyadic relation’
(Propns, p. 316). But here he might perhaps have welcomed Wittgenstein’s suggestion
that the form of words ‘A is to the left of B’ does exhibit the two words ‘A’ and ‘B’ in a
dual relation, namely the relation that they fill, respectively, the first and second blanks
of the expression ‘ . . . is to the left of . . . ’ (TLP, 3.1432). So again a dual relation
between the words in the proposition is representing a dual relation between the
objects that those words mean.49 But we do not find in Russell’s writings of this date,
i.e. either in Propns or in AMind, any proper discussion of how it is the verb or verb-
phrase of a word-proposition that ‘unifies’ that string of words to make it a proper
sentence with a truthvalue, rather than just a series of words. Nor do we find anything
much by way of explanation of how that verb ‘means’ a relation, which will ‘unify’ the
fact referred to. In the early days of PoM Russell had been at pains to point out that
it is particularly the verb that has this unifying role (e.g. PoM, chapter 4, esp. sections
52–4).50 Later in PLA the corresponding claim is that the verb is, or is part of, the
predicate, and to understand a predicate ‘you have to bring in the form of a proposi-
tion’ (PLA, p. 205). But in Propns we hear nothing about any special role for the verb.
This may be connected with the fact that the image-proposition, as he envisages it, is
just an arrangement of images, and it has no verb. The verb has to be supplied by adding
‘this is how things are’, or something comparable, and though this addition does
produce a genuine claim, still it is a very indefinite claim, by comparison with what
a word-proposition may assert.51 Perhaps Russell is right to say that animals have mental
attitudes, such as belief and desire, towards propositions. But if so, since they have no
language in which to articulate those propositions, but only images, the attitudes must
be (by our standards) very indistinct.52
I postpone to the final chapter a discussion of the idea that what makes an atomic
proposition true is its correspondence with a fact, a correspondence which goes bit-by-

49
Of course, a given string of words exhibits all kinds of facts about the words in it, e.g. in this case the fact
that the word ‘to’ has fewer words to the left of it than it has to the right. What is of interest is those facts
about the string that are relevant to its having the meaning that it does.
50
The point is subsequently repeated in several places. A particularly clear instance is his [1911/12: 108].
51
The point that Russell’s image-propositions contain no verb, and therefore make no definite claim, is
noted in Tully [2003: 353].
52
Davidson [1975] argues that since non-human animals have no language, and therefore cannot
articulate any definite beliefs or desires, they cannot be regarded as having beliefs or desires. The compromise
suggested by Russell’s ‘image-language’ has some appeal.
 P RO P O S I T I O N S 235

bit, i.e. with separate objects corresponding to the separate object-words involved,
and with relations between those objects corresponding to the relations between
those object-words. Russell evidently thinks that this kind of correspondence does
apply in the most basic case, where we are dealing not with words but with images. But
the criticism that I shall make will apply equally in either case.
 13
Universals

This chapter concerns two different distinctions. One is the very traditional distinction
between universals and particulars. During the period considered in this book Russell
accepted this distinction. In his classical article of 1911/12, ‘On the Relation of
Universals and Particulars’ (henceforward RUP) he argued that the traditional distinc-
tion did mark a genuine difference, that both universals and particulars existed, and that
neither was reducible to the other. In fact that article is more concerned to prove the
(irreducible) existence of particulars, and its discussion gives the impression that there is
no real problem over the existence of universals. Russell continues to hold this position
in chapter 9 of The Problems of Philosophy [1912a], and later in The Philosophy of Logical
Atomism [1918b], and apparently he sees no problem with it. Much later, in the Inquiry
into Meaning and Truth [1940], his position changes. Both there and in Human Knowl-
edge: Its Scope and Limits [1948] he revokes his earlier argument and claims that
particulars can, after all, be regarded as no more than complexes of universals. From
the point of view of this book, the change of mind is interesting mainly because of the
weakness that it reveals in his original reasoning of 1912, but I shall not explore the
positive doctrine which replaces it. That evidently falls outside Russell’s period of
logical atomism, and cannot contribute to our understanding of it. Besides, a much
greater interest attaches to the other side of the traditional distinction, i.e. Russell’s
acceptance of universals.
In RUP he had offered an argument for their existence, but the argument contains
a very noticeable gap. His subsequent writings, and particularly his lectures on
The Philosophy of Logical Atomism (i.e. PLA), do make this gap quite clear, and yet he
seems never to have taken it seriously. This is in itself puzzling, and it connects with a
further puzzle. Russell’s logic is all based on what he calls propositional functions.
Considered as a piece of language, a propositional function is an open sentence, i.e. a
sentence which contains ‘gaps’ that are or may be marked by free variables. The open
sentence presumably has the same ontological status as the closed sentence that results
by filling its gaps in a suitable way. Now, on the face of it, the closed sentence ‘means’ a
proposition, and we know that in PM and in PLA Russell denied the existence of
propositions. But equally, on the face of it a (first-level) open sentence ‘means’ a
universal, and neither in PM nor in PLA nor anywhere else did Russell deny the
existence of universals. Yet there are hints that he does wish to say that an open
sentence has no ‘meaning in isolation’, just as he clearly does say at this time that a
 UNIVERSALS 237

closed sentence has no ‘meaning in isolation’. So this creates the second issue of this
chapter: what distinction (if any) did Russell see between a universal and a proposi-
tional function?
But I begin with the argument of RUP, and its support of the traditional distinction
between universals and particulars.

1 The 1912 argument for particulars

The article begins by noting three different ways of approaching the distinction
between particulars and universals. The first concerns the different ways in which we
have access to them: particulars are perceived, whereas universals are conceived. To fill
this out one would need to say more about how perception differs from conception,
but in any case we cannot reasonably suppose that all particulars are perceived, or that
all universals are conceived, so this way of approaching the distinction will not be
sufficiently general. Russell therefore says no more about it. A second approach is more
metaphysical: particulars exist in time, i.e. they are not themselves parts of time, but
they do bear to one another such temporal relations as being before or after or
simultaneous with. By contrast, universals do not exist in time. The same applies to
place: each particular has a unique place, but universals either have no place or exist in
many places at once. The argument that Russell proceeds to give will concentrate
upon this latter idea. The third approach is logical in nature: particulars are the subjects
of propositions (or the terms of relations) and are never themselves either predicates or
relations. By contrast universals are either (one-place) predicates or (many-place)
relations, and Russell makes it clear that in each case he means not the different
expressions but what they signify (p. 108).1 However, when universals are contrasted
with particulars, it is usually those (signified by) one-place predicates that are in
question ‘for universals are generally conceived as common properties of particulars’
(p. 107). Evidently, he looks with favour on this way of drawing the distinction, but he
wishes to rephrase it. In a subject-predicate proposition, the (one-place) predicate may
be expressed by a single verb, but the verb can always be re-expressed as a noun (or
noun phrase) linked to the subject by a word which expresses the relationship between
the two. (Russell’s example is that ‘A exists’ can be rephrased as ‘A has existence’.2) In
these terms we can say that the question is ‘whether there is a specific relation of
predication, or whether what are grammatically subject-predicate propositions are

1
I choose ‘signify’ as a deliberately vague term, not implying one kind of meaning rather than another.
Russell himself speaks of ‘the objects denoted by verbs and . . . substantives’ (pp. 107–8), but he surely intends
‘denote’ in this deliberately vague way.
2
The choice of example is odd, because Russell’s own doctrine (pp. 262–3 below) is that ‘A exists’ is
meaningless when ‘A’ names a particular. But since he thinks of the point to be illustrated as a merely
grammatical point, the example is no doubt acceptable. I note, however, that with more complicated
predications some considerable verbal gymnastics will be needed for the rephrasing, unless we explicitly
introduce some prefix such as ‘the property of being . . . ’ to begin the required noun phrase.
238 M E TA P H Y S I C S

really of many different kinds, no one of which has the characteristics one naturally
associates with subject-predicate propositions’ (p. 108). It is in these terms that Russell
will phrase his final answer to the question.
Let us now come to his argument about things and their places.3 He begins by saying
that he will consider, not the supposed ‘real’ things that occupy a ‘real’ space, which
somehow combines the various perceived spaces of sight, of touch, of smell, and so on,
but a space confined to one particular mode of perception, e.g. vision. He lets it be
understood that his purpose is just to simplify the question at issue, by concentrating
upon things that we really do perceive, and whose nature we can be certain of. For his
view in 1912 is that, of the supposed ‘real’ things in their ‘real’ space, we can know
nothing whatever that is not deduced from their being posited as causes of what we do
perceive. This is a ‘simplification’ that we shall later return to, but for the moment let us
accept it. Then Russell’s claim in RUP is really very straightforward. My visual field
may contain two patches of a white colour, possibly the same in shape and size and all
other visual qualities, but distinguished from one another by occupying distinct posi-
tions in the visual field. He remarks that I can see that they are two just because they
have different places, and it does not matter whether they also differ in other qualities.
He infers that there must be particulars which are not reducible to the sums of their
qualities, for either the two patches are themselves such particulars, or the places that
they occupy are distinct particulars, and that is really all that he has to say. He gives this
as the ‘outline’ of his argument:
It is logically possible for precisely similar things to coexist in two different places, but that things
in different places at the same time cannot be numerically identical, forces us to admit that it is
particulars, i.e. instances of universals, that exist in places, and not universals themselves. (p. 113)

The ensuing elaboration prefers to rephrase the argument, so that it relies on the
premise that such spatial relations as ‘to the left of ’ or ‘above’ or ‘surrounding’ are
irreflexive relations: for example, one place is to the left of the other, but nothing can
be to the left of itself, and therefore the places are different. But the important claim is
just that the places are different, and that is surely undeniable. However, its justification
may be disputed, as we shall see.
Before we come to this dispute, let us just round off the argument of RUP. The
example that is relied on is one that contrasts two perceived patches (in different places
in the visual field) with one conceived property, whiteness, that they both exemplify.
So it illustrates the first idea that particulars are perceived, while universals are con-
ceived—or so Russell here says. (But one might equally well think that it shows how
one can perceive a universal such as whiteness, as is suggested by Russell’s claim
elsewhere—e.g. in KAD [1910/11]—that one can be ‘acquainted’ with universals, as
well as with particulars.) It evidently illustrates the second idea, that particulars do have
places while universals do not, and an obvious modification applies the same point to

3
There are useful comments on this argument in Casullo [1982].
 UNIVERSALS 239

times. Finally, the two patches are particular subjects of predication, whereas their
shared predicate whiteness is a universal, so the third and logical approach is equally
confirmed. Since the subjects do exist in space and time, but their predicate does not—
or so Russell assumes—the subject-predicate relation does indeed hold between
entities of quite different kinds. So this allows Russell to end by formulating the
distinction in the way in which he thinks is most basic: particulars are the subjects of
subject-predicate relations, but never the predicates. By contrast, universals may be
either predicates or subjects.
Let us come back to the argument that is proposed. It relies upon the point that the
two different places in the visual field are indeed different, so we next ask: what is it that
differentiates them? Russell’s answer in RUP is that this cannot be due to their
exemplifying different universals, just because ‘places cannot be supposed to differ as
to qualities’ (p. 113). A little more fully, he says:
If absolute space is admitted, we can of course say that it is the difference of place that makes the
patches two . . . [But] perceived space is certainly not absolute, i.e. absolute positions are not
among objects of perception. (p. 116)

The thought is this: if we take no account of what may occupy a place, or of what may
occupy other places which are spatially related to it, then any one place is exactly like any
other. As applied to ‘real’ places in ‘real’ space, this claim is very plausible. Indeed, it
seems undeniable if the space has no boundaries, and therefore no ‘centre’ or other
distinguished points, and no such property as ‘curvature’ which varies from one place
to another. For there is then no property or relation that distinguishes any one region
of the space from any other that matches it exactly in shape and size. But this is a
consideration which would apply to ‘real’ space (especially as conceived by Newton).
Russell in RUP appears to be thinking that the same holds of space as perceived, and in
particular of different places in one’s visual field, but the situation in this case is not the
same, as we may see by considering his later doctrine.
In the Inquiry into Meaning and Truth (pp. 99–100) he claims that different places,
either in visual space or in real space, can be distinguished by their having differing
coordinates in some suitable coordinate system. In the case of real space, this gives only
half the answer. For the coordinates of a place fix its distance and direction from some
point given as the origin of the system, and that apparently specifies the place by giving
its (universal) relations to some particular point. But it is not at all clear that that
particular point can in turn be specified by its (universal) properties or relations.
However, in the case of a visual space this potential objection fails, because the visual
field is perceived as a finite space, with a discernible centre, which can act as a point of
origin. It also has distinguishable directions, such as upwards and downwards, or to the
left and to the right, so that any two positions in it can be differentiated by their
different relations to the point of origin. (In Human Knowledge Russell introduces
a slight variation on this idea, now claiming that visual space is an ‘absolute’ space,
and that its various positions can be distinguished by having different ‘qualities’,
240 M E TA P H Y S I C S

corresponding to such descriptions as ‘in the upper left quarter of the total field’. This is
a matter of detail.) The important point is the recognition that different places in the
visual field are differentiated from one another by being perceived as having different
qualities, i.e. different universal properties, or different universal relations to other
places in the field. Consequently they do not also have to be individuated by being
different particulars. The way is then open to regarding each of them as being just a
‘bundle’ of such universal qualities.
It is worth adding a quick word on this idea that a particular thing may be identified
with the ‘bundle’ of qualities that apply to that thing. What is to count as a ‘bundle’?
Supposing that the relevant qualities are F, G, H, . . . then the claim that there exists a
‘bundle’ of such qualities is the claim that there exists a particular which has them all,
i.e. that 9 x(Fx & Gx & Hx & . . . ). But we cannot express the claim in this way, i.e. by
using a quantifier that is to be understood as ranging over particulars, if at the same time
we are aiming to eliminate all reference to particulars. To overcome this problem
Russell introduces a relation C(F,G) which is a relation called ‘compresence’ holding
directly between two universal qualities. Informally, it is understood that ‘C(F,G)’
holds if and only if, in the ordinary way of thinking, ‘ 9 x(Fx & Gx)’ holds. Russell can
then provide an analogue of the prolonged quantification ‘ 9 x(Fx & Gx & Hx & . . . )’
defined in terms of compresence, in a way that I do not detail here, and thereby
provide for the desired ‘bundles’ of qualities. So long as there is such a bundle of
qualities for each particular, and a different bundle for each different particular, there
cannot be more particulars than there are bundles. And if in addition the ‘bundling’ has
been explained so as to ensure that each bundle does characterize one and only one
particular, then there will be a one-one correspondence between the particulars and
the bundles, and—given a few verbal changes—whatever holds of the one will also
hold of the other. We may therefore treat these bundles of qualities as ‘logical
constructions’ that replace particulars, and so delete particulars from our ontology
without suffering any real loss.4 That is the outline of Russell’s later position, but
I shall not go into the details, which do become rather complicated.5
Let us come back to the argument of RUP. We have seen that it concerns the visual
places in a visual field, and it claims that these are distinguished from one another only
by being different particulars, and that they do not also differ in any of their purely
universal properties or relations. We have also seen that this claim is mistaken, for since
the visual field is finite its various places can be distinguished from one another by their
relations to the total field. Hence each distinguishable visual sense-datum may be
distinguished from all others in the same visual field by its universal properties and

4
At the same time we shall of course delete from our language the variables which are supposed to range
over particulars. For some information on this see e.g. Quine, ‘Variables Explained Away’ [1960b].
5
To construct the required quality-bundles, each corresponding to one and only one particular, Russell is
re-employing a method that he had first given in KEW [1914c: 121–6], of constructing sets of regions which
will each correspond to one and only one point. The method needs correction in both cases. For the case of
points, see my [2010].
 UNIVERSALS 241

relations. But one is surely more interested not in the sense-data but in the ‘real’ objects
(in the ‘real’ space) that cause those sense-data. Can these also be distinguished from
one another just by their different universal properties and relations? One comes to the
same question by a different route if one says: the visual sense-datum is identified by its
relation to the total visual field in which it occurs, but how is that field itself identified?
Or indeed how are the sense-data of other senses to be identified (for the idea that they
all occupy distinguishable positions in a sense-field would seem to be quite inappro-
priate)? The answer is surely that sense-data are identified by the time at which they
appear and the perceiver to whom they appear. The perceiver in turn may be identified
by his or her spatial position at that time. So what is needed is once more a suitable
criterion of identity for ‘real’ positions in ‘real’ space-time. But in RUP this question is
almost entirely ignored.6
No doubt Russell is in 1912 quite happy to suppose that these places are particulars,
which do not need to be distinguished from one another by exemplifying different
universal properties and relations. His later position will be that they are in fact
distinguishable in this way, i.e. by the different universals that apply to them, for if
one fixes the origin of a coordinate system by reference to a sufficiently large space-
time event, then there will not actually be an exact duplicate of that event elsewhere in
the universe or elsewhen. One notes that this assumes that the principle of the identity
of indiscernibles in its controversial form is true in fact, but not that it is a necessary
truth. It may be debated whether that assumption can reasonably be made.7

2 The assumption of universals

Let us now turn from particulars to universals. Particulars surely do exist in some form,
even if only as bundles of qualities that are never duplicated. But with universals the
mere claim that there are such things is already a bone of contention.
Russell’s own discussion in RUP recognizes only one move in the history of
philosophy that (as he thinks) aims not to admit the existence of universals, and that
is the attack mounted by Berkeley and Hume on Locke’s notion of ‘abstract ideas’.
He discusses this on pp. 111–12. In fact Berkeley and Hume do not really deny
anything that Russell wishes to affirm, though Russell makes it seem as if they do.
Locke’s overall position had been that words have meaning by standing for ideas, and
he had supposed that a general word (such as ‘man’ or ‘triangle’) stood for a suitably
general idea. Such an idea, he thought, was formed by ‘abstraction’, which he
explained in this way: one reviews a number of ideas of particular men, or particular

6
On p. 121 of RUP it is said that if there are (as we suppose) external causes of our two visual white
patches, then there must be two distinct external objects that cause them. But (a) this remark is evidently
careless, for it is easy to think of exceptions, and (b) it anyway does not address the issue. For it does not say
whether the two external objects could share all their purely universal properties and relations.
7
For the identity of indiscernibles see above, pp. 114–15.
242 M E TA P H Y S I C S

triangles, and ‘leaves out’ the features that those ideas do not have in common. So, for
example, the general idea of a man is the idea of a thing which has some colour, but not
the idea of a thing of any particular colour, such as white or black or brown or
whatever. It is also the idea of a thing which has some height, but does not specify
any particular height. Similarly the general idea of a triangle is the idea of something
that has three straight sides, but their lengths are not in any particular ratio to one
another, e.g. neither equal nor unequal. Berkeley, however, claimed that such an idea
was impossible. This was because Berkeley thought of ‘ideas’ as mental pictures, copied
from experiences, and he complained that there could not be a picture of a man which
did not picture him as having any specific height or colour, nor a picture of a triangle
which depicted no particular relationship between the length of its sides. At one point
in his discussion Locke had rather carelessly described the general idea of a triangle as
neither oblique nor rectangle, neither equilateral, equicrural, nor scalenon; but all and none of
these at once . . . It is something imperfect that cannot exist, an idea wherein some parts of several
different and inconsistent ideas are put together. (Essay IV, vii, 9.)

Locke should have said that the idea contains none of these features; he certainly should
not also have said that it contains all of them. Berkeley’s criticism fastens upon what he
did say, but even if we confine attention to what he should have said the criticism has
force. How could there be a picture of a triangle which did not picture it as either
roughly equilateral, or clearly not equilateral? Mental pictures may be vague and
undetermined in certain respects, but surely there could not be a picture of a triangle
which was ‘neutral’ as between these two:

Locke should deny Berkeley’s unspoken premise, that ideas can only be pictures.
But if we grant Berkeley’s premise, and stick to the thought that words have meaning
by standing for ideas, then the solution that Berkeley offers seems to be the only one
that is at all plausible: a general word has meaning, not by standing for some one
supposedly general idea, but by standing for a whole range of more specific ideas.
These will include, in the present example, ideas of triangles with very different shapes,
which we can think of as together representing all triangular shapes. But they will
resemble one another just in the respect that they are all (pictures of) triangles. Russell
fastens upon this last point. He takes it that Berkeley (and Hume too, who follows
Berkeley on this issue) means to analyse the general word ‘triangle’ in terms of
resemblance, so that to say that something is a triangle is to say that it bears the relevant
similarity to these. And he takes it that this is an attempt to avoid admitting that there is
such a thing as the universal triangularity. (Actually, to give him his own example, he
takes it that to analyse ‘is white’ as ‘is coloured similarly to this’ is a way of trying to
 UNIVERSALS 243

avoid the claim that there is such a thing as whiteness.) This is a complete misunder-
standing of what Berkeley and Hume were trying to do, which was to offer some
account of how we understand general words. In fact the account is not very different
from what Russell himself has to say on the matter, either in KAD [1910/11] where he
tries to explain how we can be acquainted with universals, or later in AMind [1921a]
where he tries to explain how a word can be used as a general word. In neither case is
he trying to produce a paraphrase which eliminates that word, and nor were Berkeley
or Hume. As a matter of fact Berkeley and Hume did not believe in the kind of
universals that Russell believed in, which are neither mental entities nor spatio-
temporal particulars, but that is not the point that they are trying to argue when they
take issue with Locke’s account of ‘abstract ideas’. It seems better to say that they simply
had no conception of universals, as Russell understood that idea.
The proposed analyses of ‘triangle’ and ‘white’ are quite easily seen to be mistaken.
For they entail that ‘x is triangular’ and ‘x is white’ each imply the existence of
something else other than x, which is triangular or white, namely the standard example
which x is said to resemble. But the original evidently has no such implication. However
Russell raises a quite different point. We may (for the sake of argument) accept the
proposed analyses, but reply that they have not eliminated all reference to universals.
Perhaps they no longer commit us to the existence of triangularity, or whiteness, but
they do commit us to a different universal, namely similarity (or resemblance) in the
relevant respect. And this is a universal, since it holds of many pairs of particulars, each in
different places, and so itself has no place. How did we reach this conclusion? The
assumption is that if we assume the truth of some sentence which uses the phrase ‘is
similar to’, then we are committed to the existence of an entity called ‘similarity’, and
this entity of course will be a universal. Similarly, if we assume the truth of a sentence
which contains the phrase ‘is white’, then we are committed to the existence of
whiteness; and ‘is triangular’ commits us to the existence of triangularity, and so on.
Plato long ago observed that every sentence contains a verb (Sophist 262). The modern
version is that every simple sentence contains a predicate, of one or more places.
Russell’s assumption seems to be that if you accept the truth of any such sentence
then you thereby accept the existence of an entity that is named by nominalizing its
predicate. That entity will of course be a universal. So to accept any simple sentence as
true is to accept the existence of a universal. No wonder Russell always writes as if the
existence of universals is not—or should not be—a matter of controversy.
Russell was of course aware that in mediaeval times the scholastics had debated the
status of universals, and were divided into three broad schools of thought: nominalism,
conceptualism, and realism. The classic statement of the question is due to Porphyry
much earlier (AD c.232–c.305).8 Do universals exist at all? If so, do they exist outside
the mind, or simply as mental entities? If the former, are they corporeal or incorporeal?

8
The statement comes from his Isagoge. I paraphrase the useful summary in Kenny [2005: 121].
244 M E TA P H Y S I C S

And do they exist in the things that are perceptible by the senses, or are they separate
from such things? The nominalists answered ‘no’ to the first question; they accepted
the existence of words which appeared to be names (nomina) of universal entities, but
held that this appearance was misleading, for there were not really any such entities.
The conceptualists answered ‘yes’ to the first question but ‘no’ to the second; in their view
there are universal entities, but they exist only in minds. The realists, however, claimed that
universals existed quite independently of any human thinking, either supposing (like
Aristotle) that they exist in the various spatio-temporal particulars that are their instances,
or boldly claiming (like Plato) that they exist quite separately from anything in space
or time. In RUP [1911/12], and more explicitly in chapter 9 of PP [1912a], Russell’s
position on universals is that of a Platonic realist. But he never seems to take the alternative
positions at all seriously, and this is somewhat puzzling, as we shall see.
I shall not try to describe the early versions of nominalism, that may be found, for
example, in Roscelin and Abelard in the twelfth century AD, and in some later
thinkers.9 A modern version finds its classical statement in Quine’s article ‘On What
There Is’ of 1948, and we may fix upon that. Quine’s principal claim is that ‘To be is to
be the value of a variable’, i.e. that you commit yourself to the existence of things of a
certain kind by accepting the truth of quantified propositions in which the quantified
variables must range over things of that kind if the proposition is to be true. But since
Quine also thinks that quantified variables are properly employed only when they take
the place of nouns or pronouns, you might well say that—on his approach—a first step
towards existential commitment is to make use of noun phrases that apparently refer to
the items in question. In Quine’s own view this is merely a prima facie commitment,
firstly because one may then offer to paraphrase away the suspected noun phrases as a
mere façon de parler. (For example, one may hold that such noun phrases are used only
‘for the sake of brevity’, and then add by way of illustration that the two noun phrases
used here, namely ‘brevity’ and ‘the sake of brevity’, are obviously not to be taken at
face value, as referring to one abstract object named ‘brevity’ and another called its
‘sake’. For one could have said instead ‘in order to be brief ’, and this paraphrase no
longer appears to mention the two abstract objects in question.) Secondly, even if no
suitable paraphrase seems to be available, still Quine holds that it is not the name-like
expression itself that commits one to the existence of what it apparently names, but the
readiness to infer what results when this expression is replaced by the quantifier
‘something’, or (more accurately) when it is replaced by a variable which is then
bound by an existential quantifier. On Quine’s account, the nominalist is one who
holds that what appear to be names of universals are not really such; they are mere façons
de parler, and do not actually name anything at all. That is shown by the (supposed) fact
that they do not give rise to existential quantification.

9
The best known of these nominalists is probably William of Occam (c.1287–1349).
 UNIVERSALS 245

Russell does have some appreciation of Quine’s point of view. In an unpublished

paper on the paradox of the liar, written in September 1906, he remarks that ‘whatever
can be an apparent [i.e. bound] variable must have some kind of being’, and he
specifically applies this remark to quantification over predicate variables (p. 106 of
the mss).10 But he never published this paper, possibly because he came to disagree
with some of the things said in it, and certainly there is no mention of this idea in his
RUP (or PP). Indeed, there is no mention even of the old nominalist idea that the mere
occurrence of an apparent name does not by itself commit one to anything. On the
contrary, his discussion in RUP appears to assume the opposite, for he emphasizes the
point that a predicate (such as ‘ . . . exists’) can always be rephrased as a noun, provided
that we add a little verb such as ‘has’ (as in the phrase ‘ . . . has existence’), and he
apparently takes it for granted that this noun does refer to an object, which must of
course be a universal (p. 108). Similarly, his argument on universals supposes that if you
use the predicate ‘ . . . is similar to . . .’, then you are committed to the existence of a
universal called ‘similarity’. At first one might think that this is merely a hangover from
his early work PoM [1903], which had assumed that everything could be named, i.e.
that everything was what he called a ‘term’, namely something that could be referred
to and so could function as the subject of a proposition. But that was way back in
1903, and since then he has worked at resolving the paradoxes by his theory of types,
which must reject this approach. For example, we cannot suppose that the predicate
‘ . . . is not true of itself ’ can be regarded as naming an object, true of all and only
those objects that are not true of themselves, for such a supposition can only lead to
a contradiction.
In fact Russell’s (simple) theory of types forbids one to alter the grammatical type of
an expression by turning it into a name. In particular, it insists that ‘a predicate can only
occur as a predicate. When it seems to occur as a subject, the phrase wants amplifying
and explaining’ (PLA, p. 205). The kind of ‘amplification’ or ‘explanation’ that he has in
mind is clearly the one that the nominalist is committed to, i.e. it shows how an apparent
nominalization of the predicate can be paraphrased away, as merely an idiomatic way of
saying what could be said more long-windedly by using the predicate as a predicate. We
have earlier considered how awkward this view can be (pp. 107–8 above), and we shall
return to the point in Chapter 14 (p. 269), but there is no doubt that it is what Russell’s
(simple) theory of types requires. However RUP and PP were written after Russell
had settled on his final theory of types, yet they readily introduce nominalizations
of predicates, and they take it for granted that these apparent names do really
name things called ‘universals’. (PP adds that these things do not exist only in our
thought. Its example is that Edinburgh would still be north of London whatever we
thought, and therefore the noun phrase ‘being to the north of ’ does not name a mere

10
I rely on the citation in Lackey, i.e. in his Essays in Analysis (= Russell [1973: 134]). For volume 5 of
Russell’s Collected Papers has not yet appeared.
246 M E TA P H Y S I C S

thought.11) In these writings Russell makes no mention of his theory of types, but surely
he should have done? His silence evidently leaves us with a problem of interpretation.

3 Universals and propositional functions

The problem of finding a consistent position for Russell to occupy is compounded by
what he has to say about propositions at this time, and what seems to be thereby
implied about propositional functions. For, at least at first glance, a predicate just is
(a simple case of) a propositional function, and a propositional function is explained as
something which has propositions as its values. In PoM [1903] he had said that a
propositional function has no existence outside the propositions that are its values:
According to the theory of propositional functions here advocated, the ç in çx is not a separate
and distinguishable entity: it lives in the propositions of the form çx, and cannot survive analysis.
I am highly doubtful whether such a view does not lead to a contradiction, but it appears to be
forced upon us, and it has the merit of enabling us to avoid a contradiction arising from the
opposite view. If ç were a distinguishable entity, there would be a proposition asserting ç of
itself, which we may denote by ç(ç); there would also be a proposition not-ç(ç), denying
ç(ç) . . . [And this would lead to a contradiction.] (PoM, }85)

In the substitutional theory that he had explored in [1906a–b], Russell had abolished all
mention of classes and of propositional functions, instead proposing a theory which
quantified over propositions and the individuals that were their subjects, but over
nothing else. Moreover, when in [1908] he finds this theory ‘technically inconvenient’,
and reintroduces explicit variables for propositional functions, he still claims that these
variables could in principle be eliminated in the same way as before (ML, p. 77). But in
PM [1910d], there is a clear change, for there is now no longer any talk of how these
variables might be avoided, and in fact there is no quantification over propositions, so
that the avoiding technique is now no longer available. Instead, propositional functions
are apparently taken as fundamental. They are distinguished both into the simple types
that his original substitutional theory would have demanded, and into the ramified
types that an admission of propositions would have demanded. But they are no longer
implied to be parasitic on propositions.
The result is that the logical system of PM does not contain quantified variables for
propositions, and does not assume that there are such things. At the same time the
‘multiple-relation’ theory of judgment claims that locutions which apparently refer to
propositions do not really do so: they are ‘incomplete symbols’, which disappear upon
analysis, and are replaced by a reference to the constituents of those propositions. (It is true
that in PM and in PP we are told only of how this theory applies to atomic propositions,

11
PP, pp. 97–9. But a conceptualist might reply that it is only in our thought that the phrase ‘is north of ’ is
taken as naming an entity, and therefore that entity exists only because it is thought of.
 UNIVERSALS 247

but Russell assumed that it would also apply to more complex propositions, and he had
intended to demonstrate this in TK.) The logic, then, has replaced propositions by
propositional functions, and the accompanying metaphysics claims that propositions
do not exist, though their constituents do. But the constituents of propositions include
universals, and universals are apparently the entities that propositional functions refer to.
So what has happened to the claim of PoM that propositional functions have no
independent existence? Has it simply been overruled by subsequent developments?
The answer is surely more complex. When Russell came to reject propositions he
did not of course reject the sentences which we think of as expressing them. Those
combinations of words exist, even if they should not be understood as standing for
some kind of object. Consequently he came to use the word ‘proposition’ simply as a
word for the linguistic item, the indicative sentence, or for any other combination of
symbols (e.g. images) that may be regarded as having the same function. The identifi-
cation is explicit in works written after PM, e.g. in PLA [1918b]:
A proposition is just a symbol. It is a complex symbol in the sense that it has parts which are also
symbols: a symbol may be defined as complex when it has parts that are symbols. In a sentence
containing several words, the several words are each symbols, and the sentence composing them
is therefore a complex symbol in that sense. (PLA, p. 185)

Similarly a year later in IMP :

We mean by a ‘proposition’ primarily a form of words which expresses what is either true or
false. I say ‘primarily’ because I do not wish to exclude other than verbal symbols, or even mere
thoughts if they have a symbolic character. But I think the word ‘proposition’ should be limited
to what may, in some sense, be called ‘symbols’, and further to such symbols as give expression to
truth and falsehood. (IMP, p. 155)

The doctrine of Propns [1919b] and AMind [1921a] is the same. A proposition there
is either a word-proposition, which is a series of mental occurrences of words, or an
image-proposition, which is a complex of mental images. And what these mental
symbols ‘mean’ is not the ‘proposition’ as originally understood, but rather the ‘fact’
that makes that proposition either true or false. A proposition, then, is just a series of
symbols.
The passage just cited from IMP goes on to say that the same applies to propositional
functions: they too are just sequences of symbols, i.e. sentences which contain free
variables, and which become proper (closed) sentences when the variables are suitably
replaced by constant expressions:
A ‘propositional function’, in fact, is an expression containing one or more undetermined
constituents, such that, when values are assigned to these constituents, the expression becomes
a proposition. (IMP, pp. 155–6)

Thus both propositions and propositional functions are regarded simply as linguistic or
quasi-linguistic sequences of symbols. Admittedly, the passages just cited are from
248 M E TA P H Y S I C S

writings after PM [1910d] and RUP [1911–12], but when he is looking back on things
from a much later perspective (i.e. in My Philosophical Development, 1959) Russell tells us
that he thought in the same way during the time when PM was being composed.
Referring to that time he says ‘Whitehead and I thought of a propositional function as
an expression’ (p. 124).12 So from 1910 onwards Russell regarded both propositions
and propositional functions simply as combinations of symbols, and not as what
(if anything) those combinations of symbols might signify. We need not suppose that
in either case there is any item that they signify.
But now let us come back to the criterion proposed in Quine’s ‘On What There Is’
[1948]: one is committed to the existence of so-and-so’s when one uses quantified
variables that have to be taken as ranging over so-and-so’s if what one says by means of
them is to be true. Russell certainly uses quantified variables that take the place of
propositional functions. But how should those variables be interpreted? I suggested
earlier that a substitutional interpretation seems to be required by the fact that Russell
thinks of propositional functions ‘intensionally’. That is, he thinks that ç and c may be
different functions, even if they have the same extension, i.e. even if they are true of
just the same objects. For this reason one cannot think of ‘ 9 ç( . . . ç . . . )’ as claiming
simply that there is an extension that could be assigned to ‘ç’ and would make this
proposition true. Rather, one has to think in terms of ‘meanings’ for the various
functional expressions that are here replaced by ‘ç’. But recently the point has been
that not even this is required, since the expressions in question need not be thought of
as ‘meaning’ any kind of object. Instead, we just say that the variable stands in for
expressions of a particular kind, and so we are led directly to the substitutional interpre-
tation of the quantifier: ‘ 9 ç( . . . ç . . . )’ is true if and only if there is an expression of the
relevant kind that yields a truth when substituted for ‘ç’ in ‘( . . . ç . . . )’. What universal
object that expression may mean, or what particular objects it is true of, is not directly
relevant; for all that matters is the truth of the whole proposition, and if this can be
secured without assigning any object, or set of objects, to the functional expression,
then that is quite good enough.13
I add two footnotes to this interpretation. (i) Russell himself apparently endorses this
way of understanding his higher-level quantifiers when he says (somewhat later, in
1940) ‘In the language of second-order, variables denote symbols, not what is symbo-
lised’ (Inquiry into Meaning and Truth, p. 202). This thought has appealed to many who

12
Compare pp. 68–9 and p. 82. Similarly in ‘On The Notion of Cause’ [1912/13], written in 1912, he
says in a footnote that ‘a propositional function is an expression containing a variable . . . ’ (p. 182n., my
emphasis).
13
One could equally construe quantification over propositions in this substitutional way, if propositions
are no more than sentences. (Linsky [1999: 58] notes that one can reasonably construe the comments
surrounding PM*14.3 as saying that it would be legitimate to quantify over propositions if we had first
distinguished the different orders of propositions, but PM itself does not do this.)
 UNIVERSALS 249

have commented upon him,14 and it is very difficult to see how it can be avoided. (ii)
Quine himself later admitted that his remarks about ontological commitment presup-
posed an ontological rather than a substitutional interpretation of the quantifiers,15 and
so on this interpretation Russell is not affected by them. He does, as we say, ‘quantify
over propositional functions’, but if this commits him only to the existence of expres-
sions, then that is presumably uncontroversial. He need not claim that any of these
expressions have what he would call ‘meaning in isolation’.

4 The problem
But there is a crucial objection to this interpretation of Russell’s logic, which it is very
difficult to deflect. Russell aims to deduce orthodox mathematics, which includes
Cantor’s theorem that there are non-denumerably many real numbers. His approach
assumes that statements about the real numbers may always be paraphrased as state-
ments about the propositional functions from which he ‘constructs’ them. So this must
presume that there are non-denumerably many propositional functions. But then
propositional functions cannot be merely linguistic expressions, for no (learnable)
language can contain more than denumerably many of these, as Russell well knew
[1906b: 184–5]. So far as I can see, Russell in 1910 has no way out of this contradiction.
A possible way out might be to come back to the notion of universals. Russell’s
theory must be that some propositional functions do stand for universals, and if one asks
‘which?’ then the only answer that has any support from what Russell himself says is
‘those that are simple’.16 A good reason for thinking that propositional functions are
generally derived from propositions, and have no meaning of their own, is that they
may have all the complexity that propositions have. From any proposition, however
complex, we may abstract a number of functions, just by substituting variables for some
or all of the expressions in it that can significantly be varied. But just as, by mastering a
finite vocabulary, we can come to understand any of the infinitely many propositions
of the language, so the same applies to propositional functions. We certainly do not
have to grasp all the propositional functions that occur in a proposition before we can
understand that proposition. If anything, the reverse order is more appropriate. We
come to understand whole propositions first, and then come to see how they may be
viewed as containing propositional functions which also occur in other propositions.
More accurately, the position is this.

14
E.g. Sainsbury [1979: ch. 8, sec. 3]; Landini [1998: ch. 10, sec. 6]; Stevens [2005: ch. 3, pp. 81–9]. On
the other side see e.g. Linsky [1999: pp. 9, 15–20].
15
E.g. Quine [1969: esp. pp. 104–7].
16
This answer is clearly given in Linsky [1999: 29–30]. In his earlier [1988] he had suggested that all of
Russell’s predicative functions might represent universals, but there is no need for this. (His reasoning in
[1999: 106] needs the assumption that every universal is expressed by a predicative function, but not the
converse.)
250 M E TA P H Y S I C S

Let us begin by accepting a simplification that Russell himself quite often makes:
assume that all propositions can be seen as built up from the simple atomic propositions
by truth-functional combination and the introduction of quantifiers. So one’s under-
standing of propositions depends upon one’s understanding of these methods of
building complexes and one’s understanding of the atomic propositions from which
they begin. These atomic propositions use names and simple predicates. To understand
them we need to know what are the particular things that are being named, and what it
is that is being predicated of them. By Russell’s standard doctrine, one understands this
latter by being acquainted with the universals that those predicates signify. So what he
must be supposing is that simple predicates do in some way signify universals, but we
need not extend the same recognition to more complex predicates. They can be
regarded as no more than a series of symbols, which do not have any ‘meaning in
isolation’, and which are understood just by understanding the propositions in which
they occur. But to understand the propositions one must understand the simple
symbols that they are built from, and in the case of predicates that is a matter of
being acquainted with the universals that they stand for. At any rate, that is Russell’s
doctrine at all times before The Analysis of Mind [1921a].
As we have seen (on pp. 124–5, above) in that book he drops his claim that all
understanding of language requires acquaintance with universals. He has been com-
paring the human use of language, as a reaction to environmental stimulus, with the
non-linguistic reactions that other animals display. But he takes it that animals have no
grasp of universals, and this leads him to say that the same may hold of human beings
too, and hence:
Words of which the logical meaning is universal can therefore be employed correctly, without
anything that could be called consciousness of universals. (AMind, p. 228)

It is true that he at once goes on to say

I think a logical argument could be produced to show that universals are part of the structure of
the world, but they are an inferred part, not a part of our data. (Ibid.)

However, he does not say either there or elsewhere in the book what this ‘logical
argument’ would be. One can guess that he is thinking along these lines: a fact consists of
one or more particulars exemplifying a definite property or relation, and they could not
combine to form a fact without this property or relation that unites them into one complex
entity. So properties and relations must occur in the world, and that is to say that universals
must occur, because there must be something that makes the various separate items unite to
form a single fact. However, this is only a speculation, and to understand Russell’s views in
1910 we shall do better to set AMind on one side. Elsewhere his doctrine has been that
universals are needed to explain how we can understand our language. But they are
needed only to explain how we can grasp the simple predicates of the language. We do not
need to suppose that each complex predicate also stands for a universal.
In any particular language there will be only finitely many simple predicates, for
otherwise no finite being could learn the language. But even if it is only simple
 UNIVERSALS 251

predicates that have to stand for universals it does not follow that there are only finitely
many universals. For presumably there are infinitely many different languages that are
possible, and in any case there may also be universals which no predicate of any
language would stand for. We could consistently suppose that there are infinitely
many universals, and perhaps even a non-denumerable infinity of them. Would this
rescue Russell’s position on mathematics? On the face of it, it would not. For the
construction that he gives is one that uses variables which are explained only as standing
in place of propositional functions. They are not explained as variables that range over
universals, and so however many universals there may be that would not rectify the
situation. But perhaps we could add something to Russell’s stated assumptions, in order
to build in the idea that every universal is expressed by a predicative function. Then,
although his official doctrine quantifies over predicative functions, he could be thinking
of it as really quantifying over universals. To evade the main problem, one would still
need a reason for supposing that there were in fact more universals than the proposi-
tional functions that are actually expressed in any language, but I can only leave it to
others to invent such reasons.17
I simply end this chapter with a question: how many universals should one believe in?
My own preferred answer is the nominalist answer: none. As it seems to me, the basic
idea—which goes back to Plato18—is that our language has a meaning. In particular,
the meaning of a whole sentence is built up from the meanings of its individual parts,
and so—amongst others—the parts that we now call ‘predicate expressions’ must have
a meaning. But it seems obvious that the parts that we now call ‘subject expressions’
have their meaning by denoting particular objects. So it is supposed that the other parts
must also have their meaning in a similar way, but by denoting objects of a different
kind, namely universals. However, this supposition is entirely mistaken. Predicate
expressions do not have a meaning in that way. It is at least better to say that in their
case the important notion is that of a truth-condition: the predicate is true of an object
(or pair of objects, etc.) if and only if . . . This gives no reason for supposing that a
predicate has a meaning by standing for a special kind of object. Nevertheless, as I see it,
such a supposition is the root of the whole idea that there are such things as universals.
They are, in the first place, what predicates are thought to stand for, and if you deny
that predicates stand for anything then no worthwhile case remains for the existence of
such things as universals are supposed to be.
This topic is clearly controversial. I cannot here argue further for the view that
I would recommend.19

17
I have argued in my [2009a: chapter 9], that numbers are fictions. It is no doubt convenient to feign that
there are non-denumerably many entities, but I myself see no good reason to suppose that this fiction should
be taken seriously.
18
I have discussed Plato’s views in my [1994] and my [2009c: 160–3]. I do not mean to deny that Plato
had other reasons too, apart from the fact that predicate expressions do have meaning, but these other reasons
may be easily disposed of.
19
For a view that is clearly opposed to mine, see e.g. Armstrong [1989].
 14
Facts

Russell’s lectures on The Philosophy of Logical Atomism begin with an outline sketch of
what he means by ‘logical atomism’. Its basic premise is that the world contains many
different things, and that to find out what these things are we need to practise what he
calls ‘analysis’. The idea is that almost all familiar things are in one or another way
complex, but by analysis we can find out what simpler components these complex
things are put together from, and if we continue this process we should end by reaching
the ultimately simple things from which all else is composed. These are the ‘atoms’, and
they are called ‘logical atoms’ because they are the last residue of ‘logical analysis’,
which is said to be something quite different from physical analysis. They will in fact be
the ultimate particulars and the ultimate universals, that combine together to make the
simplest kind of fact, i.e. the atomic facts. The whole world consists just of these.1 It is
worth remarking that a common view nowadays is that philosophers are concerned
with the analysis of language, and Russell is not objecting to this conception. In fact he
proposes to study the nature of the world by studying the nature of a language which is
designed to be an adequate reflection of it. But the opening paragraphs of these lectures
clearly state that the real goal of the study is the world itself, and not the language in
which we speak of it.
The language which reflects the nature of the world is called by Russell a ‘logically
perfect’ language (p. 197). This will contain names which refer to particulars, n-adic
predicates which express n-adic universals,2 and atomic sentences which express the
atomic facts that have these as their components. The language will also contain
suitable logical symbols, for the truthfunctors and the quantifiers, and suitable variables
to be bound by those quantifiers, but no other vocabulary. There will of course be rules
of formation, explaining how the symbols of the language can be put together to form
sentences, and rules of proof saying how one sentence can be proved from others. This
logical vocabulary, and the rules for it, are to be those used in Principia Mathematica.
I shall comment later on whether this is an adequate basis for a ‘logically perfect’

1
Russell will later ask whether there are also other facts, besides the atomic facts here mentioned. For
simplicity, we may begin by ignoring this possibility.
2
Russell does not speak of dyadic or triadic predicates, as we do, but instead uses ‘relation’ for the general
notion, and includes ‘monadic relations’ (p. 199) to cover our monadic (or one-place) predicates. He
sometimes uses ‘relation’ to mean a relational expression, but more often to mean the quality or relation
expressed. In these lectures he does not use the word ‘universal’.
 FAC T S 253

language, but we may begin with the names and predicates and atomic sentences
which form the basic and non-logical part of the language. Here Russell imposes a
special requirement. The names and predicates to be included will all be simple
symbols, and the idea is that they will be symbols just for the simple particulars in
the world and the simple universals that are predicated of them. But here we must ask
what Russell means by ‘simple’, and why he thinks that this requirement is either
needed or desirable. So this brings us to our first topic, simplicity.

1 Simple particulars
In these lectures Russell gives us no useful definitions of particulars or universals. It is
true that ‘particulars’ are officially defined as ‘terms of relations in atomic facts’ (p. 199),
but since neither ‘relation’ nor ‘atomic fact’ is further defined, this is not very helpful.3
However we surely do understand what he is thinking of. Equally, we get no
definition of ‘simple’ in either case, and here it would certainly have been useful if
he had given some further explanation.
Early in the second lecture we are told that
All the ordinary objects of daily life are apparently complex entities: such things as tables and
chairs, loaves and fishes, persons and principalities and powers—they are all on the face of it
complex entities. All the kinds of things to which we habitually give proper names are on the face
of them complex entities: Socrates, Piccadilly, Rumania, Twelfth Night or anything you like to
think of, to which you give a proper name, they are all apparently complex entities. (p. 190)

Such things as these, says Russell, will turn out to be ‘series of classes of material
entities’, and we know from elsewhere that even a simpler ‘material entity’ will actually
turn out to be a construction from sensibilia. So these are not the simple particulars of
which the world is composed. In fact Russell is still accepting his principle of acquain-
tance, which he now puts in this general form:
The components of a proposition are the symbols we must understand in order to understand the
proposition. (p. 196)

This certainly applies to the names and predicates of atomic propositions, and Russell is
evidently taking it for granted that we do understand the propositions of the logically
perfect language. The application is just what one now expects:

3
In the second edition of PM [1925: xv], we are given this more positive account. Atomic propositions
have the forms ‘R1(x)’, ‘R2(x,y)’, ‘R3(x,y,z)’, ‘R4(x,y,z,w)’, etc. The letters ‘x’, ‘y’, ‘z’, . . . represent
particulars; the letters ‘Ri’ represent relations. The same particular may occur in atomic propositions of
any of these forms; the same relation can occur in only one of these forms. This provides some response to
Ramsey’s well-known attack on the distinction in his [1925b] (but at the cost of denying multigrade
relations, such as belief).
254 M E TA P H Y S I C S

A name, in the narrow logical sense of a word whose meaning is a particular, can only be applied
to a particular with which the speaker is acquainted, because you cannot name anything you are
not acquainted with. (p. 201)

The relevant particulars, then, are sense-data. This explains why he has earlier said that
a logically perfect language would be very largely private to one speaker, for it is only
the person who has the sense-data who is acquainted with them, and can name them.
So far, we are on familiar territory.
But are all sense-data simple? Evidently, Russell does not think so. He has earlier said
When I speak of a ‘sense-datum’, I do not mean the whole of what is given in sense at one time.
I mean rather such a part of the whole as might be singled out by attention: particular patches of
colour, particular noises, and so on. (RSP, p. 147)

But this does not mean that only what is simple will count as a datum, for he goes
on:
An observed complex fact, such as that this patch of red is to the left of that patch of blue, is also
to be regarded as a datum. (Ibid.)

It appears, then, that a simple sense-datum will contain no distinguishable parts that are
sense-data. It will be a minimum visibile (in the case of vision), or a minimum audibile
(in the case of hearing), and so on. Russell does believe in such minima. This is evident
from his discussion of the apparent continuity of motion and of some other changes in
chapter 5 of KEW, esp. p. 156.4 The suggestion that a simple sense-datum has no
distinguishable parts is also confirmed by a moral that he draws in PLA:
Particulars have this peculiarity, among the sort of objects that you have to take account of
in an inventory of the world, that each of them stands entirely alone and is completely self-
subsistent . . . That is to say, each particular that there is in the world does not in any way logically
depend upon any other particular. Each one might happen to be the whole universe; it is a merely
empirical fact that this is not the case. (pp. 201–2)

Particulars are logically independent of one another, but that would not be so if one
particular could be a part of another. For the whole could not exist without its part.
Indeed, Russell has said elsewhere that the only way in which one thing can be logically
dependent upon another is when the other is a part of the one (KEW, p. 81). All the
genuine particulars, then, are sense-data, and the simple particulars are sense-data that
are minimally extended in space and in time, i.e. those that have no smaller parts that
are sense-data. These will be referred to by simple symbols of the perfect language, i.e.
by genuine names. And no doubt we are to assume that in this language no name
names more than one particular, and no particular has more than one name. Other
non-simple particulars will be introduced only by complex symbols, namely definite

4
Cf. AMatter, p. 292.
 FAC T S 255

descriptions, which are to be found by analysing them into their simple parts, and
saying how those parts are related.
That answers our first question about simplicity, i.e.: what is to count as a simple
particular? Let us now turn to universals.

2 Simple universals
The position with universals must be rather different, (a) because universals do not in
any obvious sense have parts, and (b) because in their case our text says nothing about
the logical independence of one from another. We can say, as before, that Russell is
not abandoning his principle of acquaintance, so the logically perfect language will
contain predicates only for universals that one can be acquainted with, i.e. those that
are exemplified in one’s sense-data. But presumably only some of these will count as
simple. Which are they?
The text of PLA appears to contain only one hint that bears upon this question. As a
preliminary, let us note that ‘red’ is one of Russell’s standard examples of a predicate
that stands for a universal applicable to sense-data. He uses it, for instance, as his
illustration of the point that, in order to understand a predicate, ‘you have to bring
in the form of a proposition’, i.e. you have to understand what it means to say of
something that it is red (p. 205). But earlier, when he is speaking of how we understand
an expression for something simple, he has said this:
Take the word ‘red’, for example, and suppose—as one always has to do—that ‘red’ stands
for a particular shade of colour. You will pardon that assumption, but one never can get on
otherwise. You cannot understand the meaning of the word ‘red’ except through seeing red
things . . . (pp. 193–4)

The implication is clear. ‘Red’ is not really simple, because the word does not really
stand for a particular shade of colour. In practice, it stands for a range of colours, all of
which qualify as red. One infers that a ‘simple’ universal must, for Russell, mean one
that is completely specific, already fully determinate, and not a range that is further
determinable. This is a notion that we do understand in some uncomplicated cases,
where we can say that we mean a specific colour, or a specific pitch, or a specific
temperature, and so on. No doubt there is a sense in which to say that something is
both red and round, or both hot and wet, is to say something ‘more specific’ than just
that it is red, or that it is hot. But this is not a more specific account of its colour, or of its
temperature, and so is not what we asked for. And presumably we should admit that
the demand for an absolutely specific account will not always make good sense. (For
example, what would be an absolutely specific tone of voice?) But we certainly have
enough understanding of this idea to be able to generalize beyond the one example
that Russell seems to give, so at least we can progress to our next question: why should
Russell think that if a language is to be ‘logically perfect’ its names and predicates
256 M E TA P H Y S I C S

should stand only for particulars and universals that are in this sense simple? His own
discussion gives us no obvious answer to this question.
A possible suggestion is that he wished the simple symbols of his language to stand
for things that are not definable. For example, one can apparently say that a complex
particular will have various parts arranged in a particular way, and so can be defined by
first listing those parts and then specifying their arrangement. Similarly, a non-specific
universal can presumably be defined by listing its more specific sub-universals, and
thereby specifying the range that it covers. But although complexes can in this way be
defined in terms of simples, the converse definitions seem equally possible. Why should
one not instead define simples in terms of complexes? (I shall give examples in a
moment.) So long as both the simples and the complexes are objects of acquaintance,
they could in each case be understood directly from the acquaintance, and so neither
has to be introduced by a definition in terms of the other. If, for economy, some of
them are to be defined, what is supposed to be the advantage of proceeding in one
direction rather than the other?
It is tempting to suppose that Russell’s observation on the mutual independence of
his simple particulars provides a clue: perhaps he is aiming for a situation in which all of
his atomic propositions are logically independent of one another? This is suggested by
the fact that, when he is introducing the idea of a logically perfect language, and
requiring its names and predicates to stand only for what is simple, he says:
A language of that sort will be completely analytic, and will show at a glance the logical structure
of the facts asserted or denied. (PLA, p. 198)

Is this meant to imply that the logical relations between facts will thereby be shown ‘at a
glance’? If so, then it must at least be claiming that those logical relations will be
provable, and hence in particular that any logical relations between atomic facts will be
provable. But the logical rules of Principia Mathematica will not allow you to prove any
logical relations between atomic facts, and so the inference is that there are none. In
that logical system (as in all others) every atomic sentence is treated as logically
independent of every other. It is obvious that the sentence quoted need not be seen
as carrying this implication, but it is a tempting suggestion.
A further pointer in the same direction may be found in the earlier discussion of
chapter 2 of KEW, which extols the usefulness of logic for philosophical purposes.
Russell there asserts that ‘in all inference, form alone is essential’ (p. 53), and he
evidently means by this that inferences are available only when the propositions in
question contain explicitly logical notions, such as are expressed by the truthfunctors
and quantifiers. That is why he goes on to say that ‘the forms of propositions giving rise
to inferences are not the simplest forms’ (p. 54), which clearly implies that from one or
more atomic propositions nothing can be inferred about a further atomic proposition.
We may add that in Wittgenstein’s version of the theory it is explicit that all atomic
propositions are logically independent of one another (TLP 4.2115), and if Russell held

5
Wittgenstein speaks of ‘elementary’ propositions where Russell speaks of ‘atomic’ propositions.
 FAC TS 257

the same view then that might explain his liking for particular sense-data without parts.
For in his preferred language ‘a is a part of b’ cannot be both true and unprovable. (This
is because if it is true then b is a complex, and is therefore given by a description which
specifies its parts.)
But Wittgenstein’s view would not explain Russell’s liking for fully specific uni-
versals. For it is obvious that fully specific predicates in the same family (e.g. the colour
predicates) logically exclude one another (as ‘red’ excludes ‘green’), and Russell must
surely have been aware of this. In fact one maximizes the logical independence of
atomic facts by choosing as one’s basic predicates in a given family some that are less
specific, and defining the more specific in terms of these. To give an illustration which
is very oversimplified, suppose that there are just four specific colour possibilities for
any object: it may be red or yellow or green or none-of-these. To create these four
possibilities, each excluding the others as a matter of logic, we need to take two non-
specific colour-predicates as our basic predicates, say ‘rellow’ meaning either red or
yellow, and ‘grellow’ meaning either green or yellow. These predicates are logically
independent of one another, i.e. neither of them either entails the other or entails its
negation. Then in terms of them we can define our four specific possibilities:
red = rellow and not grellow green = grellow and not rellow
yellow = both rellow and grellow none-of-these = neither rellow nor grellow
Because of their definitions, these four predicates are now, as a matter of logic,
mutually exclusive and jointly exhaustive.
The same trick can be played with larger numbers of specific possibilities. Thus,
suppose that we have just eight specific colour possibilities, say: white, yellow, green,
brown, red, blue, black, and none-of-these. In this case we shall need three non-
specific colour predicates as our basic predicates, say
A = white or yellow or green or blue
B = red or brown or green or yellow
C = black or blue or green or brown
Then the eight specific possibilities may be defined
green = A and B and C white = A and not B and not C
yellow = A and B and not C red = not A and B and not C
blue = A and not B and C black = not A and not B and C
brown = not A and B and C none-of-these = not A and not B and not C
In the same way we could have four basic and non-specific colour predicates, used to
define sixteen specific predicates, all mutually exclusive, and so on. In general, so long
as the number of the specific possibilities to be distinguished is a power of 2, say 2n,
then we shall be able to generate them as the various ways of compounding n suitably
chosen non-specific predicates. The basic and non-specific ones will be logically
258 M E TA P H Y S I C S

independent of one another, and the defined specific ones will be, as a matter of logic,
mutually exclusive and jointly exhaustive.6
I do not imagine that Russell had thought of this point, but I also think that he
would not have been much impressed by it. For (a) as we increase the number of
specific possibilities to be catered for, the non-specific possibilities to be taken as basic
do become quite unnatural groupings; (b) there is no particular reason to suppose that
the number of specific possibilities should be a (finite) power of 2; and (c) while one
might be able to deal in this way with the logical relations that hold between the
monadic predicates of a particular family, e.g. the colour predicates, I see no similar
trick that can be played with predicates of two or more places, though these will clearly
be needed amongst the atomic predicates. For example, I do not see any definitions
from which you could prove all the simple logical relations that obtain between
predicates of the family ‘larger than’, ‘smaller than’, and ‘equal to’, including (for
instance) the fact that they are all transitive relations and the first two are irreflexive.
The point seems even more obvious with the entailments that stem from the logical
properties of the triadic relation of between-ness. One can perhaps choose what to
count as an atomic fact in such a way that no one atomic fact entails any other (as ‘x is
equal to y’ entails ‘y is equal to x’), but Russell is not strongly committed even to this.7
He surely is not committed to stronger claims about the lack of all logical relations
between atomic facts, and he cannot reasonably have supposed that a judicious choice
of what to count as an atomic fact might secure this result.
Etymologically the word ‘atomic’ means ‘not split-able’. Of course Russell’s atomic
facts can be split, not into further facts, but into a universal and the one or more
particulars that are its terms. But he wishes to say that their ingredients are ‘atomic
particulars’, which cannot be split into further particulars, and ‘atomic universals’,
which again cannot be split into further universals. His thought is that particulars
which have parts could be split into those parts, and so should be ruled out. Similarly,
universals which have more specific sub-universals could be split into those sub-
universals, and so should equally be ruled out. This idea applies in a familiar way to
monadic universals, and Russell does so apply it. It can also be applied to polyadic
universals, and one presumes that Russell would also apply it here. (That is, if ‘xRy’

6
In a somewhat similar way, one can if one wishes take complex and overlapping particulars as one’s basic
particulars, and define their smaller parts as those that are or are not common to the several larger wholes.
(When restating his position in ‘Logical Atomism’ [1924] Russell says that a logical language ‘will not lead to
error’ if it has simple names for complex particulars, but that it will be ‘incapable of dealing with anything
simpler than the objects which it represents by simple symbols’ (p. 337). This apparently overlooks the
possibility of either describing the simpler components by definite descriptions or naming them merely by
adding new names to the language. Similarly with universals.)
7
He says in KEW (p. 62) that ‘perhaps one atomic fact may sometimes be capable of being inferred from
another, though this seems very doubtful’. Earlier he has set the doubt aside, and claimed that ‘in all
inferences, form alone is essential’ (p. 52). Here he clearly means logical form, and is claiming that an inference
is valid only if its validity can be certified by the rules of Principia Mathematica. But the later discussion of
p. 62ff. is more circumspect, and more ready to acknowledge the possibility of exceptions.
 FAC TS 259

entails ‘xSy’, but not conversely, then ‘R’ might be atomic but in that case ‘S’ cannot
be, and vice versa.) There are other considerations that one might put forward as
governing what is to count as an ‘atomic’ relation, but Russell does not explore this
topic, and nor shall I. In any case his views on atomicity are not governed by the
thought that atomic propositions should have no logical relations with one another,
and it is not this that explains his requirement of simplicity. It is, rather, an intuition
into how the world must be structured.

3 More on simplicity
During his PLA lecture Russell does not appear to recognize that there could be any
doubt about the existence of the simple particulars and simple universals that he wishes
his logically perfect language to reflect. But in the discussion that is recorded as
following it we find this interesting exchange:
Mr. Carr: You think there are simple facts that are not complex. Are complexes all composed of
simples? Are not the simples that go into complexes themselves complex?
Mr. Russell: No facts are simple. As to your second question, that is, of course, a question that
might be argued—whether when a thing is complex it is necessary that it should in analysis have
constituents that are simple. I think it is perfectly possible to suppose that complex things are
capable of analysis ad infinitum, and that you never reach the simple. I do not think it is true, but it
is a thing that one might argue, certainly. I do myself think that complexes – I do not like to talk
of complexes—are composed of simples, but I admit that that is a difficult argument, and it might
be that analysis could go on for ever.
Mr. Carr: You do not mean that in calling a thing complex, you have asserted that there really
are simples?
Mr. Russell: No, I do not think that is necessarily implied. (PLA, p. 202)

So one naturally asks: how might it be argued that what Russell in his lecture assumes
to be simple is not really simple?
When we are concerned with simple particulars, the question seems to be one that
could be settled by experiment. As Russell says in AMatter, ‘I do not know what is the
shortest perceptible event, but this is the sort of question which a psychological labora-
tory could answer’ (p. 292). And what he says here about temporal minima for perceived
events he could equally have said about spatial minima. We know that there must be
such minima, both in time and in space, because we know that some happenings are so
short-lived, or so exiguous, that they cannot be perceived without instrumental aids. It
may fairly be said that the relevant minima may perhaps differ from one person to
another, and could be expected to differ from one species of animal to another, but that
does not affect Russell’s position. For he anyway thinks that different people (or animals)
never do perceive the same particular objects, since they perceive only their own sense-
data and no one else’s. So the question whether a given sense-datum is divisible into
260 M E TA P H Y S I C S

smaller parts can only be understood as the question whether it is or is not a minimum for
that perceiver. And this question is, as Russell says, one that empirical investigation could
answer. But in any case, whatever these minima turn out to be, it is evident that there
must be such minima. So there surely are the simple particulars that Russell assumes.8
But with universals the position is rather different, for he does allow that different
perceivers may be acquainted with the same universals, even though the particular
examples of those universals that each perceiver has are different particular sense-data.
So the question of what universals there are is not in the same way relative to each
individual perceiver. The universals that each perceiver can discern, as distinct univer-
sals, is no doubt something that depends upon his or her particular sensibility, but
apparently there may be universals which other people can discern although I cannot.
For example, you may be able to distinguish twenty different shades of red, whereas
I cannot manage more than about five. It is also possible that I should come to discern
more shades than I at first think, by engaging in a little experimentation. It may be that
shades A and B look just the same to me, and that when I look at B and C together
I can also see no difference, yet when I look at A and C together I do detect a
difference. Hence I can after all discern A and B, not when I just compare them directly
with one another, but when I compare each with a third shade C. And this process may
continue indefinitely.9 So how many different shades are there? According to our
present theory of colour, which incorporates an assumption of the continuity of space
and time, there are infinitely many different wavelengths of light between any two
given wavelengths, and so one would expect there to be infinitely many different
shades of colour. In practice, each person’s power of discernment is no doubt finite, but
we have no ground to pick upon any particular number n and say that the number of
discernible shades is limited to be just n. In this way one might ‘argue’ concerning
colours ‘that complex things [in this case colour-ranges] are capable of analysis ad
infinitum, and that you never reach the simple [i.e. the fully specific]’. But notice that
the argument here suggested does of course depend upon empirical premises, in
particular upon the continuity of space and time, and hence of possible wavelengths,
as well as the assumed relation between wavelength and shade of colour. But all of this
is put in doubt by the possible development of quantum theory. Once more, the
question is an empirical question.
On Russell’s phenomenalist approach, we must begin with what we can perceive, for
our words can have meaning only by being related to that. He will allow that we can
also theorize about what is not perceived, but his phenomenalism insists that such
theories should be understood only as positing more things of the same kind as we can

8
Notice that on this account simple particulars may overlap one another. For example, suppose that (at a
certain distance) I can perceive a line that is one millimetre long, but no shorter line, and suppose that (at the
same distance) I am actually looking at a line that is two millimetres long. This line has indefinitely many parts
that are one millimetre long, and presumably I am seeing all of them. So they are all, for me, simple
particulars.
9
Russell is fond of this example. (It occurs at KEW, pp. 148–51, and quite often elsewhere.)
 FAC TS 261

perceive, i.e. not just sense-data but also sensibilia. At the time when he wrote PLA his
view was that physical theories should be construed as concerned only with ‘logical
constructions’ from sensibilia, i.e. with classes of them, or series of such classes, and so
on, but not as introducing any entities of a new kind. Admittedly, it is somewhat
difficult to see how he could have held such a view of the nature of the world, i.e. as
nothing more than a ‘logical construction’ from sense-data, while being so very
interested—as he obviously was—in the scientific theories being propounded and
discussed at this time. But when he wrote PLA he did hold this view, and it was not
until the later work AMatter [1927a] that he felt able to return to the earlier position of
PP [1912a] and to accept that in science (and in common sense) we frequently posit the
existence of things that are not even in principle perceptible. In PLA, then, he is
assuming not only that we begin with what is actually perceived but also that we never
get beyond what could in principle be perceived. So what he is noticing here, in his
reply to Mr. Carr’s question, is that there are no clear limits to what is perceptible, and
that is certainly ‘a question that might be argued’.
He perhaps shows his awareness of this point in his later article ‘Logical Atomism’
[1924], in which he once more tries to sum up his approach to basic questions. For he
there says:
When I speak of ‘simples’ I ought to explain that I am speaking of something not experienced as
such, but known only inferentially as the limit of analysis. It is quite possible that, by greater
logical skill, the need for assuming them could be avoided. (p. 337)10

While Russell is still in his phenomenalistic phase, the kind of ‘analysis’ that he must
be thinking of as allowing us to approach, and perhaps to reach, the relevant
‘simples’ can only be regarded as resulting from the empirical analysis of what in
fact we can perceive. For his view is that everything we can understand must be
built up in some way from that. So one cannot do it just by sitting in one’s armchair
and thinking, which is how ‘analysis’ – and especially ‘logical analysis’ – is usually
conceived. But this remark is intended as an elucidation of Russell’s thought, and
not as a criticism.

4 The existence of particulars

The discussion so far has been concerned only with what Russell in PLA counts as ‘atomic’
facts. I shall shortly move on to address the question whether there are other kinds of facts
that the perfect language will speak of, though they are not counted as ‘atomic’. This
question may be pursued without restricting atomic facts to components that can be
perceived, and I shall pay no further attention to that Russellian requirement. But before

10
It is a speculation that the ‘greater logical skill’ that Russell is thinking of might be a development of his
method of introducing geometrical points as a ‘limit’ of ever more specific geometrical locations. The
method certainly has its problems. I have discussed them elsewhere in my [2010].
262 M E TA P H Y S I C S

I come to that more general issue it will be useful first to deal with a curious point in his
discussion, which may perhaps be connected with his idea that this language will contain
names only for particulars that are simple. Very much to one’s surprise, he claims that,
although such simple particulars do exist, in the logically perfect language one cannot say
so. For, if ‘a’ names a simple particular, then (he says) ‘a exists’ will be nonsense (e.g. pp.
233, 241).11 This can only be regarded as an aberration on his part. But what explains it?
One suggestion is that such a claim will be counted as ‘nonsense’ because its denial
could never be true. If a is a simple particular, known by acquaintance, then of course ‘a
does not exist’ will not be true, and it is tempting to say that there is no possible
situation in which it might have been true. Russell does accept that there are possible
situations in which this particular does not exist, for that is an obvious consequence of
his claim that particulars are logically independent of one another, i.e. ‘each one might
happen to be the whole universe’ (p. 202). Clearly, if a different particular b had been
the whole universe, then a would not have existed. But also, you could not say so. For
then there would have been no name ‘a’, and hence no proposition ‘a does not exist’
with which to describe that situation. This is perhaps Russell’s reason for saying that
‘a does not exist’ is not just false but nonsense.12 At any rate, it is what is suggested by
this comment:
There is no sort of point in a predicate which could not conceivably be false. I mean, it
is perfectly clear that, if there were such a thing as this existence of individuals that we talk
of, it would be absolutely impossible for it not to apply, and that is the characteristic of a mistake.
(PLA, p. 241)

But one must hold that this comment is misleading.

If the comment is to be taken seriously, then one should note that something very
similar should be said about what Russell regards as a simple predicate, say ‘F’. On his
account this predicate can be understood only by someone who is acquainted with
objects to which it applies, and hence the sentence ‘there do exist some things which
are F ’ cannot fail to be true in any situation in which it is understood. So for the same
reason as has just been suggested, Russell would apparently be committed to regarding
it as nonsense. But he never affirms this, and there is no reason to suppose that
he believed it. In consistency one would have to suppose that he also thought that
all logical truths were not really true at all, but nonsense, but he evidently did not
think this.13
What Russell does say makes it clear that he denies sense to ‘a exists’, when ‘a’ is a
simple name, but does not deny sense to ‘there exist things which are F ’, where ‘F ’ is

11
The doctrine has been anticipated, but without explanation at PM *14, pp. 174–5. Here in PLA it is
stressed over and over again on pp. 233–59.
12
If so, then G. E. Moore’s observation in his [1936] is pertinent: it may be admitted that ‘this does not
exist’ can never be true, but ‘this might not have existed’ usually is.
13
One can claim that Wittgenstein’s version of logical atomism does take this idea quite seriously (e.g.
TLP 4·461), but Russell’s does not.
 FAC TS 263

a simple predicate. For his reason is that the notion of existence is one that applies to
a propositional function, when it can be symbolized by ‘ 9 ’, but it is simply not
grammatical to try to apply this notion to a name.14 As he puts it:
When you say ‘unicorns exist’, you are not saying anything about individual things, and the same
applies when you say ‘men exist’. If you say that ‘Men exist, and Socrates is a man, therefore
Socrates exists’, that is exactly the same sort of fallacy as it would be if you said ‘Men are
numerous, Socrates is a man, therefore Socrates is numerous’, because existence is a predicate of a
propositional function. (PLA, p. 233)

His point is that the sentence ‘There are many things which are men’ makes perfectly
good sense, whereas ‘There are many things which Socrates’ does not, for it is not even
grammatical. To put the point more formally, for any predicate ‘F’ it makes perfectly
good sense to say ‘there is at least one thing x such that Fx’, which we abbreviate to
‘ 9 xFx’. This is perfectly grammatical, but ‘ 9 (Socrates)’ is not.
However, this line of argument contains a weakness. Russell himself notes that an
existence claim must make sense when it is applied to a definite description. One could
think that ‘ 9 (the husband of Xanthippe)’ is just as ungrammatical as ‘ 9 (Socrates)’,
and yet Russell allows that something very like it does have a sense. He introduces the
new symbol ‘E!’ so defined that ‘E! (the husband of Xanthippe)’ means ‘there exists
one and only one person who is husband of Xanthippe’. More generally

E!( x:Fx) $ 9 x(Fx & 8y(Fy ! y ¼ x))

ß

So one asks: why could one not also define a predicate like ‘E!’, which would make
sense when applied to a name ‘a’, and would mean ‘a exists’? Well, the simple answer is
that we already have such a predicate. To claim that a exists we need only put

9 x(x ¼ a)
This is well-formed in Russell’s logically perfect language, and is indeed provable, as it
should be, given his conventions for a name symbol ‘a’. Presumably he cannot mean to
be claiming that this formula is nonsense, so one can only interpret him as insisting that
it should not be seen as meaning ‘a exists’. But why on earth not?15
I note, incidentally, that if in this last formula we substitute the description ‘ y:Fy’ for
ß
the name ‘a’, then we obtain

9 x(x ¼ ( y:Fy))
ß

On expanding this according to the standard recipe we get

9 y(Fy & 8z(Fz ! z ¼ y) & 9 x(x ¼ y))

14
Compare Lycan [1993].
15
Wittgenstein’s TLP also claims that ‘a exists’ is nonsense, but at least his preferred language does not
accept ‘ 9 x(x ¼ a)’ as a well-formed sentence, for he does not allow a sign for identity. (TLP 5.53–5.534)
264 M E TA P H Y S I C S

Here we may drop the last clause ‘ 9 x(x ¼ y)’ as logically superfluous, since it is always
provable, and thus reach what Russell gives as the definition of
ß
E!( y :Fy)
It follows that he could have defined ‘E!(a)’, meaning that a exists, as short for
‘ 9 x(x ¼ a)’ in either case, i.e. when ‘a’ is either a simple name or a definite description.
While there are of course differences between names and descriptions which one must
take into account, still they do not lead to any differences in the behaviour of the
existential quantifier ‘ 9 ’, which applies in the same way to both. Russell thought
otherwise, but one can only say that this was an aberration on his part.

5 Other kinds of facts (i): Logical notions

A logically perfect language will contain atomic propositions, expressing atomic facts.
Russell tends to assume that the logically perfect language will be able to express all the
atomic facts that there are, for he thinks of it as capable of giving ‘a complete
description of the world’ (e.g. p. 215, p. 236). Such a language is of course an
idealization, for no person could be in a position even to name all actual (minimal)
sense-data, let alone the many (minimal) sensibilia which no one actually experiences.
But it is an idealization which seems harmless, and we need not fuss over it. In any case,
let us now turn from the atomic propositions to the non-atomic ones. Must we assume
that, when these are true that is always because there is some corresponding non-
atomic fact that makes them true? Russell gives different answers for the different cases.
He first argues that there are no disjunctive facts to correspond to true disjunctive
propositions:
When you take such a proposition as ‘p or q’, . . . you will have two different facts involved in the
truth or falsehood of your proposition [i.e. either the fact that p or the fact that not-p, whichever
of those is a fact, and either the fact that q or the fact that not-q] . . . I do not suppose there is in the
world a single disjunctive fact corresponding to ‘p or q’. It does not look plausible that in the
actual objective world there are facts going about which you could describe as ‘p or q’. (p. 209)

He does also add: ‘I would not lay too much stress on what strikes one as plausible; it is
not a thing you can rely on altogether’. And later on he does introduce an apparent
doubt about this point (p. 237), though it seems to me to be without substance, and I
shall ignore it. He clearly does not wish to believe in disjunctive facts, and quite often
he simply denies their existence. In fact he generalizes his argument, to cover all
‘molecular’ propositions, i.e. those that are truthfunctional compounds of two or
more simpler components. If such a proposition is true, then its truth will be due to
the individual facts which make its components true or false, and we do not have to
suppose that there is some further fact which combines them. This is no doubt a
comprehensible position, for the truth or falsehood of the components will entail the
 FAC TS 265

truth or falsehood of the compound, and this entailment does not need to be supported
by any further fact in the world.
However Russell does not take a similar view about negative facts. Again, he would
prefer not to admit negative facts, and he accepts that the idea is certainly unpopular,
but he feels that it cannot be avoided. As a preliminary, note that we need only
consider the negations of atomic propositions, for in the language of Principia Mathe-
matica it is always possible to exchange the negation of a compound proposition for a
compound of its negated components. Consider, then, that familiar example ‘this is
red’, which we may take to be an atomic proposition and a false proposition, so that its
negation ‘this is not red’ is true. What makes it true? Do we have to suppose that there
is such a thing as the fact that this is not red?16 A natural suggestion might be that we
need not assume so, for it is enough that there is no such fact as the supposed fact that
this is red. That is, a proposition may be true either because what it asserts is a fact or
because what it denies is not a fact, and no further explanation is needed. But Russell
would not agree. He would respond that the absence of a fact is itself a fact, and so
would have to be mentioned in a complete description of the world. It is only another
way of speaking of a negative fact. He says just this in Propns [1919b: 288], written
shortly after PLA, but in PLA itself he considers only a different proposal for avoiding
negative facts that is due to Demos [1917], and which he appears to have misunder-
stood.
He takes it that Demos is proposing an analysis of the meaning of ‘not’, according to
which

‘not-p’ means ‘there is a proposition q which is true and is incompatible with p’.
Note first that, if this is the proposal, then ‘not-p’ is taken to be a general proposition,
asserting the existence of some proposition other than p (and presumably of a positive
proposition other than p). But Russell’s comment ignores this completely, and focuses
only on the word ‘incompatible’. Here there is almost certainly a misunderstanding, for
Demos no doubt understands ‘q is incompatible with p’ in its usual sense, in which it
says that it is not possible for both q and p to be true. But Russell professes not to
understand this use of ‘possible’ applied to a whole proposition (as he makes clear on p.
231), and instead interprets ‘incompatible’ in the way that he has said earlier that he will
understand it, i.e. as a truthfunctional compound of ‘p’ and ‘q’, which is true if and only
if ‘p’ and ‘q’ are not both true (pp. 210–11). That is why he thinks that Demos is

16
Where our proposition contains just a single subject and a monadic predicate (as here), we do have
genuine intuitions on whether the proposition is positive or negative. (Definitions are attempted by Cook
Wilson [1926: vol. I, ch. 12], and Ayer [1952]; they are discussed in some detail in Gale [1976: ch. I, part 1.3].)
But in the polyadic case our intuitions are much less secure. For example, in many contexts ‘<‘, ‘’, and ‘>’
may be interpreted so that ‘a<b ∨ ab ∨ a>b’ is always true. (We may take ‘<’ as any transitive and
asymmetrical relation, e.g. ‘is smaller than’.) Do we have to conclude that one at least of these three relations
must be negative?
266 M E TA P H Y S I C S

introducing what he calls a ‘molecular’ proposition, asserting incompatibility. So his

objection is
We have been trying to avoid both negative facts and molecular facts, and all that this succeeds in
doing is to substitute molecular facts for negative facts, and I do not consider that that is very
successful as a means of avoiding paradox. (p. 214)

But Demos is not proposing to introduce a molecular fact, and if he were then he could
presumably avail himself of Russell’s own previous argument to show that this is not
needed.
What Demos is proposing is surely better understood in this way. What ‘makes true’
the negative proposition ‘this is not red’ need not be a corresponding negative fact, but
can perfectly well be taken as a positive fact, such as (in this case) the fact that this is
white. This is incompatible with ‘this is red’, and so excludes it by entailing its
negation. Just as with Russell’s argument on disjunctive facts, we do not have to
suppose that this entailment needs to be made true by an entailment fact ‘in the world’,
for the entailment holds irrespectively of what is in the world.17 To generalize, what
Demos is proposing is best understood in this way. Where p is any false atomic
proposition, there will be a positive atomic proposition q which is true and which
entails not-p. If so, then we can say that it is the fact that q that makes it true that not-p.
This is a line of thought that Russell’s own discussion simply does not address, though
surely it should have done. The doubtful point is the assumption that, for any atomic
falsehood, there always will be a positive atomic truth that entails its negation. How can
one be sure of this? Well, it appears that we would need an inventory of all the atomic
predicates and their logical relations, and then we simply look through the list to see
whether this is always so. However, it is a list which Russell cannot provide, so the
check cannot be carried out. But the claim seems plausible enough at first sight, and
one might perhaps make it a requirement on one’s choice of what to count as an
atomic predicate that it should verify this proposal.18
Whereas Russell professes some doubt over negative facts, he finds it quite clear that
one must admit that there are general facts. Indeed, he deems it best to keep both the
quantifiers 8 and 9 in his basic language, not defining either of them as the negation of

17
This contradicts the succinct discussion in Molnar [1999], who claims that ‘P entails ~Q’ requires a
truthmaker, though ‘P entails Q’ does not. (Compare Russell’s Propns, p. 288, which claims that ‘it has to
be a fact that what is round is not square’, and adds that this is itself a negative fact. But one might just as
well claim that it has to be a fact that what is round is either red or round, and add that this counts as a
disjunctive fact.)
18
The discussion in Gale [1976: ch.1, part 2], is in practice limited to propositions that could be regarded
as atomic propositions with monadic predicates. He is attracted by the view that, for any one of these
propositions that is false, there is always one that is true and entails its negation. But he admits that he cannot
prove it, and concludes that it is at best an empirical truth. However, a scrutiny of his examples suggests that
‘entails’ should be replaced by ‘leads by (scientific) law to’ if the claim is to remain plausible.
 FAC TS 267

the other, and he affirms the existence of both universal and existential facts.19 On
universal facts he has a good argument, but in the existential case he give us only this
very weak remark:
It is not so difficult to admit what I might call existence-facts—such facts as ‘There are men’,
‘There are sheep’, and so on. Those, I think, you will readily admit as separate and distinct facts
over and above the atomic facts I spoke of before. (p. 236)

However there is no more need to admit existence-facts than there was to admit
disjunctive facts, and in each case the argument is very similar. For if an existential
generalization is true that can only be because some instance of it is true. Moreover, the
instance will entail the generalization, and can clearly be thought of as what makes it
true, just as what makes a disjunction true can only be the truth of one or other of the
disjuncts. Russell has commented earlier that one may know that an existential
generalization is true without knowing any instance of it. (His example is: ‘You
know that there are people in Timbuctoo, but I doubt if any of you could give me
an instance of one’, p. 234.) However the same applies to disjunctions, for one may
know that a disjunction is true without knowing which disjunct is true. But in each
case the proposition cannot be true unless there is a simpler true proposition which
entails it, and the simpler fact is all that we need to explain why the more complex
proposition is true.
However in the case of universal propositions Russell is on much stronger ground.
He says:
I do not think one can doubt that there are general facts. It is perfectly clear, I think, that when
you have enumerated all the atomic facts in the world, it is a further fact about the world that
those are all the atomic facts about the world, and that is just as much an objective fact about the
world as any of them are. It is clear, I think, that you must admit general facts as distinct from and
over and above particular facts. (p. 236)

If all the particular instances of a universal generalization are true, then the generaliza-
tion itself must be true, but still the conjunction of all those instances will not entail it.
For that, you need the further premise that these are all the instances, and that is not
itself an atomic premise.
It may be helpful if I add a word in explanation of the difference that has here
emerged between 8 and 9 . The universal generalization is somewhat like the con-
junction of all its instances, just as the existential generalization is somewhat like the
disjunction of all its instances. But the entailments work in opposite directions. An
existential generalization is entailed by each of the disjuncts (or any disjunction of them),
but there is no entailment in the other direction. By contrast, a universal generalization

19
This corresponds to the treatment in *9 of PM, which is connected with the idea of there being
different ‘kinds of truth’, as mentioned on pp. 216–18 above. In *10 of PM we find the treatment which is
nowadays more usual, in which 9 is defined in terms of 8 and negation.
268 M E TA P H Y S I C S

entails each of the conjuncts (and any conjunction of them), but again there is no
entailment in the reverse direction. What is needed for the converse is in each case the
same extra premise, that the items listed in the continued conjunction or disjunction
are all the items that there are. Then we get a provable equivalence between 8 and the
continued conjunction, and between 9 and the continued disjunction.20 But one does
need the extra premise, and it is a perfectly fair observation that this premise is not
entailed by any collection of singular facts. So if a ‘complete description of the world’ is
supposed to be a description from which every truth about the world can be inferred,
then this must be separately included. So long as it is, then in Russell’s view every truth
about the world will follow just from this and the atomic truths concerning each of
these items, so our list of facts will be complete.
At once, one needs a qualification: we shall have a basis from which one can deduce
all the truths that can be stated in the logically perfect language, as Russell conceives it.
But there will apparently be very many truths which cannot be stated in this language.

6 Other kinds of facts (ii): Intensional notions

As we have noted, there is one way in which the language that Russell is envisaging can
say more than can be expressed in any natural language. For he is imagining a language
in which every atomic fact can be stated, and this requires that every simple particular
can be named, and every simple universal can be expressed. But he also supposes that all
simple particulars are minimal sense-data, or at least sensibilia, and no learnable
language could contain a name for each of these. It is also highly doubtful whether
there could be a language capable of expressing every simple universal that applies to
sense-data, for certainly no actual language does.21 In an opposite way, one may say
that his language is highly restricted, since sensibilia are the only things that it can talk
about, because Russell sees no need to admit the existence of anything else. So the
names are all names of sensibilia, and the predicates are all predicates that apply to
sensibilia. There is no vocabulary for talking of anything else, except that there is also
the purely logical vocabulary, which contains variables not only for names and
predicates, but also for propositional functions of arbitrarily high types. In this logical
vocabulary one can introduce a way of speaking which mimics our usual way of
speaking of classes, as is done in Principia Mathematica, and all talk of objects other than
sense-data can at best be replaced by this way of mimicking talk of classes of sense-data,
and classes of these in turn, and so on. In this language what we think of as talk of
ordinary material objects can only be treated in this way, and the same applies to talk of

20
Strictly, we must assume that the list of all the items that there are is finite, if this list is to be stated in
Russell’s language, and the equivalences proved. However, we have already seen (pp. 102–4) that Russell is
forced to assume an axiom which implies that the list is infinite. Consequently his ‘logically perfect’ language
cannot actually state the one general fact that is needed.
21
No actual language could do so if, as suggested at the end of Chapter 13, Russell must hold that there
are uncountably many simple universals.
 FAC TS 269

more recondite things such as electrons, or fields of force, or other entities that the
physicist is concerned with. For in PLA Russell is firmly in his phenomenalistic phase,
and thinks that all material things are adequately treated just as ‘logical constructions’
from sensibilia. (And only one year later, in 1919, he will also embrace his version of
neutral monism, which aims to treat minds as similar ‘logical constructions’. In PLA he
still hesitates over this.) The language, then, is generous over sensibilia but very
niggardly over everything else. For Russell thinks that there is no need to assume the
existence of anything else, except of course for whatever may be needed to ensure the
truth of the logical part of the language, including its quantification over propositional
functions. (This has been discussed in the previous chapter.)
One obvious way in which the language is limited is that it is intended to obey the
theory of types, and this does severely restrict what one can say. For instance, a predicate
must always occur as a predicate, and it cannot be turned into a name and thereby be
made the logical subject of some other predicate, except where the other predicate is of
higher level. (For example, it can be made the ‘logical subject’ to a quantifier 9 or 8.)
To illustrate: ‘x is honest’ is presumably a predicate of persons, and therefore cannot
occur except as such a predicate. So if you want to get the effect of saying ‘honesty is a
virtue’ you must find a way of doing so which mentions honesty only in this way. An
obvious first attempt would be ‘8x(x is honest ! x is virtuous)’, but then there are also
obvious objections. For one thing, there may be people who are honest but not
virtuous, because although they do have the virtue of honesty they have no other
virtues but very many vices. For another, although it may be true that all honest people
are also virtuous, we wish to rule out the idea that that is just an accident, which might
be so if there is really no connection between the two. This second objection is the
more important one, and it applies to many similar examples, e.g. the proposal to
paraphrase ‘red is a colour’ as ‘8x(x is red ! x is coloured)’. One wishes to add an adverb
such as ‘necessarily’ in order to show that this is no accident, but Russell’s ‘ideal
language’ will not contain such an adverb. So this is a problem, though it is not one
that Russell anywhere addresses.22 But it is perhaps a special case of this more general
problem. With the possible exception of belief contexts—which I shall come to at the
end of this section—the language that Russell envisages contains only extensional
methods of building up more complex propositions from the atomic propositions that
we began with. Although in PM he does not impose any axiom of extensionality, still
one can be added without contradiction, and in the second edition of 1925 he does
propose adding one. He can do this because, with the possible exception of belief
contexts, the preferred language does not contain any way of embedding a name or a
predicate or a sentence in some complex context in which it does not occur extension-
ally. Let us note some examples of what must therefore be left out.

22
Notice that even if we were allowed to turn the predicate ‘ . . . is honest’ into the noun ‘honesty’, or the
predicate ‘ . . . is red’ into the noun ‘red’, it is not at all clear that this by itself would resolve the problem.
Something extra still appears to be needed.
270 M E TA P H Y S I C S

(a) Modality When we talk about logic we are perpetually talking of entailment, and
showing that certain propositions do or do not entail certain others. It is obvious that
entailment is an intensional notion: what entails a given proposition will not also entail
every other proposition of the same truthvalue. Indeed, a usual way of defining
entailment is to say that the propositions P, Q, R, . . . entail a further proposition S if
and only if it is not possible for P, Q, R, . . . all to be true and S false. Russell quite often
makes use of this notion of possibility in other ways too, for example when he says that
the truths of logic should be true ‘in all possible worlds’ (pp. 68–9 above), and when he
says that each simple particular is logically independent of every other, since each could
exist without any other existing (p. 254). It is true that he also claims that such a notion
of possibility is illegitimate, and that it has arisen only as a result of a mistake. He says:
In all traditional philosophy there comes a heading of ‘modality’, which discusses necessary,
possible, and impossible as properties of propositions, whereas in fact they are properties of
propositional functions. Propositions are only true or false. (PLA, p. 231)

He seems not to have noticed that his own talk of logic constantly uses these modal
notions that he here condemns as illegitimate. But in any case we can certainly say that
in the ‘logically perfect’ language these notions do not occur, and you cannot talk of
logic.23
(b) Causality In our ordinary and everyday explanations we frequently say that one
thing happened because of another, but it is clear that a sentence of the form ‘P because
Q’ is not an extensional context for the sentences ‘P ’ and ‘Q’. The same comment
applies to the closely related use of what are called ‘counterfactual’ conditionals, such as
‘If it had been the case that P, then it would have been the case that Q’.24 It is plausible
to say that these constructions imply some sort of connection between what does or
would make P true and what does or would make Q true, and it has certainly been
regarded as a kind of ‘necessary connection’, though not of the same kind as the strict
necessity which is involved in entailment. Since Hume, it has been claimed that this is
an important ingredient in the concept of causation. When Russell discusses this
concept in ‘On the Notion of Cause’ [1912/13], he sets his face against the introduc-
tion of any such non-extensional condition, and propounds a view according to which
all that matters is the regularity of an association between events of different kinds
(pp. 190–1 above). But again we find that he cannot conform to this in his own
thinking. This is particularly clear in what he has to say about ‘mnemic causation’ in
chapter 4 of AMind, i.e. cases where past experience is causally effective in bringing
about a present occurrence. Here is his account of the central case of memory:

23
There are some clear analogies between possibility and probability, and the latter is sometimes counted
as a modal notion. In any case, it certainly does generate intensional contexts.
24
Counterfactuals are usually understood to be subjunctive conditionals in the past tense. But the point is
a more general one, and applies to subjunctive conditionals in any tense.
 FAC TS 271

From time to time we remember things that have happened to us, because something in the
present reminds us of them. Exactly the same present fact would not call up the same memory if
our past experience had been different. Thus our remembering is caused by –
(1) The present stimulus
(2) The past occurrence
It is therefore a mnemic phenomenon according to our definition. (AMind, p. 82, my
emphasis.)

Notice that he infers the existence of a cause from one premise that uses a ‘because’,
and one that is a counterfactual conditional. He does not try to convince us that there is
a regular connection between past experiences of a certain kind and later memories of
that experience, and this seems very wise, for indeed what a person will or will not
remember is something that varies a great deal. And we can all agree that this is a case of
causation before we even start to wonder what kind of regularity might be involved. As
a matter of fact Russell does believe that there are general laws of mnemic causation to
be found, and in AMind he does attempt to offer some, but they are formulated in such
vague and general terms that all kinds of apparent exceptions at once present them-
selves. I shall not discuss them.
I add two footnotes on this topic of causation. (i) We have noted that later in AMatter
Russell does accept that a causal link from a to b must be carried along some continuous
space-time track.25 But when discussing mnemic causation in AMind he explicitly says
that although one may hope and expect to find such a track, in the form of an enduring
physical brain-trace, one is entitled to be sure that there is a causal link without being
sure of that (pp. 85–6, 89–92). (ii) Concerning causation in general, a principle that is
quite widely accepted is that if there is a causal link from a to b then there must be some
regularity that covers it.26 What is not at all widely accepted is a converse principle, that
if there is a suitable ‘regularity’ linking a to b then the one must cause the other. For it is
difficult to see how to explain what will count as a ‘suitable’ regularity in a way that is
non-circular, and if possible given in purely extensional terms, but that rules out the
possibility of a regularity that is purely accidental. However, I cannot enter upon that
question here, for it soon turns out to be a very complex issue.
(c) Values The context ‘It ought to be the case that P’ is evidently not an extensional
context for ‘P ’; nor is ‘It is good that P’. So such things cannot be said in Russell’s
‘perfect’ language. Equally, the context ‘Mermaids are (or would be) pretty, while
ogres are not (would not be)’ is not an extensional context for the predicates ‘x is a
mermaid’ and ‘x is an ogre’. For these two predicates have the same extension, since
they are both true of nothing. It looks as if the point generalizes, so that Russell’s

25
Or, one should say, ‘near-continuous’. For our usual assumption that space-time is mathematically
continuous is only a theory, and may turn out to conflict with quantum mechanics, as Russell well knew.
26
An interesting dissentient is Anscombe [1971].
272 M E TA P H Y S I C S

preferred language cannot speak of any kind of values—moral, or aesthetic, or of any

other sort—but this may be doubted. It must be doubted if affirmations of value can be
analysed in terms of what people believe or desire or are moved by, and so on, and if
the perfect language is allowed to speak of such mental attitudes and emotions. But
again, the question whether any such analysis would be adequate is one that I must set
aside. However we should say something of the mental attitudes, and let us, as usual,
concentrate upon the central case of belief.
(d) Belief As we observed (pp. 223–5), belief contexts are naturally construed as
intensional, but apparently Russell is thinking of including them. In PLA he outlines
his earlier multiple-relation theory of this topic, but he does not elaborate upon it, and
he gives the impression that he is only sketching an account that is not yet fully worked
out. As I have noted (on pp. 208–10), all that he needs to do in order to retain the
original idea is to take the belief-relation to be of a type that is one higher than the type
of its subordinate verb, but one may doubt whether he has seen this point himself. (It
seems an obvious point, but Russell evidently regards the fact that a belief-relation
‘contains two verbs’ as a kind of novelty, which he should not do if he has been
thinking of how that relation would fit into his theory of types as a higher-level
relation.27) It is true that if belief is to be admitted as a primitive relation of higher level,
then the language is much more complicated than we had at first thought. For (a)
belief requires a believer, and in Russell’s view at that time the believer is not a
construct of sensibilia but a new type of particular (and one with which we are not
acquainted). Moreover (b) the belief-relation is multigrade, in that it can take any
number of terms, and also multilevel, in that it recurs at all levels of the theory of types
above that of first-level predicates. Consequently the primitive vocabulary of the
‘logically perfect’ language would apparently require a great variety of higher-level
relations, all of which were basic and unanalysed belief-relations. (And so far (a) we
have only considered atomic beliefs, and (b) there are of course many other ‘proposi-
tional attitudes’ besides belief.) This adds a noticeable complexity to what is otherwise
Russell’s rather nice and simple language for describing the world.
So it is no surprise that very soon after writing PLA Russell dropped this theory, and
thought of belief once more as a two-termed relation between a person and a
proposition, with persons now being constructed like other objects from sensibilia
(or things very like them, namely images), and propositions now being taken to be
complex arrangements of words or images, existing in minds. However, the crucial
point about these propositions is that the words or images of which they are composed
must have meanings, and the way that they are arranged in the mind must also say
something about how those meanings are arranged in the world, so that there is some
definite fact in the world which makes that mental proposition either true or false, as

27
Setting belief aside, Russell’s preferred vocabulary apparently contains no other higher-level predicates,
except for the logical predicates 8 and 9, and whatever can be defined in terms of them.
 FAC TS 273

the case may be. Will the logically perfect language be able to speak of what such words
or images mean, or of what fact has to exist if a mental proposition is to be true? If so,
then it certainly seems that the language must be able to accommodate intensional
idioms, for these are prime examples. (Consequently Russell should not have supposed
in 1925, in his introduction to the second edition of PM, that he could afford to add to
it suitable axioms of extensionality.)28 Or would it be better to say that the logically
perfect language should be taken to be a language which speaks only of the non-
linguistic world, and it does not contain the vocabulary for talking either about itself or
about any other language? This would be in accordance with Tarski’s idea that no
paradox-free language can talk about itself, and that one can only talk of one language
by oneself using a different language that is its ‘meta-language’. As we have noted, this
is an idea that had some appeal for Russell (pp. 101–2),29 and certainly his interest in the
so-called perfect language was in how it related to the non-linguistic world that it was
designed to talk about, excluding both persons and language. So let us finally turn
to this.

7 The structure of language and the structure

of the world
We have seen that Russell’s ‘logically perfect’ language is quite notably restricted, so
one naturally asks: what compensating advantages is it supposed to have? As I noted at
the start of this chapter, the main advantage that Russell sees in it is this. His aim is to
study the nature of the world, and he thinks that this can be done by studying his
perfect language because he thinks that its structure will mirror the structure of the
world. He does not argue this point, but assumes it:
When we come to the principles of symbolism, which I shall deal with in Lecture VII, I shall try
to persuade you that in a logically correct symbolism there will always be a certain fundamental
identity of structure between a fact and the symbol for it; and that the complexity of the symbol
corresponds very closely with the complexity of the facts symbolised by it. Also, as I said before, it
is quite directly evident to inspection that the fact, for example, that two things stand in a certain
relation to one another—e.g., that this is to the left of that—is itself objectively complex, and not

28
His new introduction credits to Wittgenstein the idea that extensionality can be accepted, and it refers
to Wittgenstein’s obscure remarks on belief (TLP 5·541–2) as showing how an extensional account of belief is
possible (p. xiv). But it seems that what Wittgenstein had in mind—at least as interpreted by Russell—was
something very like the Russellian theory just sketched, and that surely cannot be stated in a fully extensional
language.
29
Despite TLP 5·542, and its apparent acceptance of ‘“P ” says P ’, Wittgenstein’s general position in TLP
seems to be that you cannot in language say what the language means, but have to show it. It is Russell’s
[1921b] comment on this (p. xxii) that looks favourably on the idea of a hierarchy of different languages, each
except the lowest capable of talking about the one below it. In these terms, his ‘logically perfect’ language
would be a language at the lowest level. (In later works, e.g. the Inquiry into Meaning and Truth, he does accept
the idea of a ‘lowest level’ of language, on which higher levels may be constructed.)
274 M E TA P H Y S I C S

merely that the apprehension of it is complex . . . . I shall therefore in future assume that there is
an objective complexity in the world, and that it is mirrored by the complexity of propositions.
(PLA, p. 197)

The attempt to persuade us of this point, which is here promised for Lecture VII, does
not seem to materialize in that lecture.30 In fact the lecture is headed ‘The Theory of
Types and Symbolism’, and it is mainly an exposition of the higher levels of the theory
of types. There is no further elaboration of Russell’s point here about the structure of
atomic facts and atomic propositions. So let us consider it without his assistance.31
Take a simple situation, consisting of three round black dots on a rectangular white
ground, thus:

With unusual originality, let us call the dots ‘a’, ‘b’, and ‘c’. (To come closer to what
Russell tells us is really an atomic fact, you may suppose that what actually concerns us
is the sense-datum that you have when you look at this rectangle, and that the dots are,
for you, minima sensibilia.) Now, using his example of a simple relation, i.e. ‘to the left
of’, there are three things that we can at once say about this situation:

a is to the left of b
b is to the left of c
a is to the left of c.
We might aim to use a more specific relation in this way

a is 2 centimetres to the left of b

b is 2 centimetres to the left of c
a is 4 centimetres to the left of c.
Alternatively we might feel that it is moving beyond simple atomicity to bring in such
things as the numbers 2 and 4, and a specified unit of measurement. So instead we may
introduce special cases of some tetradic relations

a is as far from b as b is from c

a is twice as far from c as b is from c.
This description omits an obvious feature of the situation, which is naturally reported
by using two triadic relations

30
The announced topic of the lecture was cut short, because Russell began in the first five pages by
responding to some questions concerning existence, which had been put to him in a letter arising out of his
previous lecture. (See p. 254.) Perhaps the original version of Lecture VII had contained more of the
promised ‘persuasion’.
31
The later article ‘Logical Atomism’ [1924: 338], repeats the point that the structure of a logically
ideal language may be expected to match the structure of the world, but again it gives no argument for this
claim.
 FAC TS 275

a and b and c are co-linear

b is between a and c
Obviously, we could go on to add further descriptions of the spatial relations between
the three dots, which also brought in how they were related to the rectangle. Or, of
course, we could choose quite a different feature to describe, e.g.

a is black
b is black
c is black
a is the same colour as b
b is the same colour as c
a and b and c all have the same colour.
Or in place of their colour, i.e. black, one could have said the same things about their
shape, i.e. round, or their size, and so on.
Perhaps there is a simpler feature of the situation that one should have begun with,
namely

a 6¼ b
b 6¼ c
a 6¼ c
But here Russell might reply that these are no longer candidates for being atomic
propositions, since in the language of PM identity is a defined symbol. Presumably he
would say the same of other propositions which make use of quantifiers as well as
identity in order to describe this simple situation, e.g.

a, b, and c are the only dots in the rectangle,

or again

every part of the rectangle is white, except for those

parts that contain parts of a or b or c.
You may amuse yourself by thinking of many, many, further ways of describing
this very simple situation, for if quantifiers are permitted then there are of course
lots more.
What moral should one draw? It is clear that there are a good many propositions
which all describe this same situation. Even amongst those that name the dots, and so
are candidates for being atomic propositions, there are propositions of several different
logical forms, i.e. those that attribute to the dots a monadic property, a dyadic relation,
a triadic relation, and a tetradic relation. If we may also include propositions containing
quantifiers, then we shall find many more logical forms being employed, all in
quite simple descriptions of this same situation. So what is the ‘logical form’ of the
situation itself? The only possible answer is that it does not have one. The propositions
276 M E TA P H Y S I C S

that describe it have their various logical forms, but since it is not the right kind of
thing to have a logical form, it cannot be said that any of their forms correctly reflect its
form, or share its structure, or mirror its complexity, and so on. More generally,
the kinds of structure that propositions exemplify cannot also be attributed to bits of
the world.32
I have presented this point by using the word ‘situation’, and have avoided speaking
of ‘facts’, as Russell so often does. This is because there is nowadays a very strong
temptation to think that a fact must always be a fact that so-and-so, where what takes
the place of ‘so-and-so’ will be a sentence which expresses a proposition—or, as
Strawson says in his discussion of this issue, a statement.33 On this understanding of
the word ‘fact’, facts do have logical forms, for the fact that p can of course be credited
with the same form as the statement or proposition that p. But, as Strawson remarks, on
this understanding, facts and statements are simply ‘made for each other’, and the claim
that facts are ‘in the world’ becomes highly suspect. Rather, ‘if you prise the statements
off the world, you prise the facts off it too; but the world would be none the poorer’
(p. 168). I have some sympathy with Austin’s response in his [1961] that this is being
‘unfair to facts’, for although the word ‘fact’ is used in this way it also has other uses too.
But in any case, when a fact is thought of as always a ‘fact that p’, then it is already
conceived as having a linguistic structure. However the same does not apply to a word
such as ‘situation’ or ‘state of affairs’, or to what Wittgenstein calls a ‘configuration of
objects’, or to what Russell often has in mind when he speaks just of ‘complexes’.
These expressions are used to refer to a bit of the world, but they do so without already
bringing in a particular sentence which describes it.
This is a distinction that Russell fails to appreciate. He constantly runs together the
words ‘fact’ and ‘complex’, as if they mean the same as one another, and he sometimes
describes these things with noun phrases, which apparently refer to complex objects,
events, or situations, and sometimes with that-clauses. A somewhat startling instance is
how he summarizes his position at the opening of Propns [1919b]:
I mean by a ‘fact’ anything complex. If the world contains no simples, then whatever it contains
is a fact; if it contains any simples, then facts are whatever it contains except simples. When it is
raining, that is a fact; when the sun is shining, that is a fact. The distance from London to
Edinburgh is a fact . . . (Propns, p. 285)

The first two examples are, on a first hearing, acceptable. The third brings one up short.
Which fact is supposed to be the distance from London to Edinburgh? Is it, for example,
just the fact that London is some distance from Edinburgh? Or that London is more

32
This point is perhaps recognized in the much later work An Inquiry into Meaning and Truth, when it is
said that ‘the words that we use never exhaust all that we could say about a sensible experience. What we say
is more abstract than what we see’ (p. 55).
33
See the well-known debate on ‘Truth’ between Austin and Strawson in 1950. Austin had defended the
claim that a statement is true if and only if it ‘corresponds to the facts’, though his defence scarcely uses the
word ‘fact’ but speaks instead of ‘situations’. But Strawson’s response is explicitly all about ‘facts’.
 FAC TS 277

than one mile from Edinburgh? Or that their centres are about 413 miles apart? Or
what? Alternatively, since the distance in question is about 413 miles, is Russell
committed to saying that 413 miles is a fact? When facts are conceived as always to
be specified by that-clauses, then they can fairly be ascribed the same ‘logical structure’
as the propositions that are specified in the same way. But when we count as a fact
absolutely everything that is not either a minimum sensibile or a completely specific
universal that such sensibilia exemplify, then the idea that every fact has a logical
structure is surely absurd.34

8 Concluding remarks
We human beings find it natural to conceive of the world around us in a particular
way, and so we have created languages that encapsulate this way of thinking about the
world. This does involve singling out some particular things as objects of attention, so
that we can go on to say things about them, and surely this is a convenient way of
thinking of the world. (Presumably many of the higher mammals also conceive the
world in a similar way.) There is some arbitrariness in our way of slicing up the world,
and often the precise limits of things are left somewhat vague. (For example, just where
does the river end and the sea begin? Just how much can a house be changed and still
qualify as ‘the same house’? Or—to take something that is now controversial—just
when does a human being begin to exist? And would the same answer apply to e.g. a
chicken, or an apple tree?) But by and large we do see the world as containing
particular things, which are chunks of matter more or less well distinguished from
their surroundings. It is not necessary to think of the world in this way. If one takes
something which has experiences quite different from ours, say an oyster, then if one
imagines it as thinking of the world at all one surely imagines it as having thoughts of a
quite different structure. To take a more realistic example, consider what has happened
in modern physics to this ancient idea of the world as made up of separate and
distinguishable chunks of matter, or even of things (such as electrons were once
conceived to be) which resemble those chunks in having definite spatio-temporal
positions and a continuing identity over time. In today’s fundamental physics this
picture has certainly been put under pressure, and future generations may well find that
it is best abandoned altogether.
Russell is moving in this direction when he abandons enduring objects and takes
short-lived events to be his basic particulars. But in our period he is still a phenome-
nalist, and so takes these events to be of the kind that one can experience, namely
events of sensing or of imaging. He supposes that everything else that we can have

34
It leads into this instructive paradox. From KAD onwards Russell has maintained that there are sense-
data which are complex, and that we can be acquainted with them. So, apparently, we can name them. But
according to the present doctrine whatever is complex is a fact, and in the perfect language facts cannot be
named but must be stated (e.g. PLA, pp. 187–8).
278 M E TA P H Y S I C S

reason to believe in must be built from these. But later in AMatter he will come round
to the idea that the scientist’s conception of a basic space-time event might be closer to
the way that things actually are, and these events need not be at all the kinds of things
that we human beings can experience. Nevertheless in that book he is still wedded to
the idea that the basic events must occupy small regions of space-time, and so can be
thought of as deserving to be called ‘particulars’. It is not obvious that today’s science
would endorse that assumption.
To conclude this account of Russell’s metaphysics: he is no doubt broadly right to
say that our ordinary and everyday language has something like the ‘atomic’ structure
that he envisages, even if the fact that we do not talk only about minimal sense-data
must make some difference to the details. But the ordinary ways of talking, which
reflect our ordinary ways of thinking, embody a conception of the world which may
turn out to be one that is well-suited only to the ordinary purposes of our everyday life.
For the scientific purposes of explaining and predicting what happens in the world, it
may be that a very different conception proves to be more successful, and one that is no
longer ‘atomistic’ in anything like the way that Russell envisages. That is surely a
possibility which he cannot rule out purely on a priori grounds.
Epilogue

In 1924 J. H. Muirhead published two volumes entitled Contemporary British Philosophy,

in which various philosophers each contributed a statement of their current views.
Russell was one of them. He called his piece ‘Logical Atomism’ (henceforth LA), and
we may fairly regard it as his final statement of the philosophy that characterizes this
period of his life. A point which one notices at once is that the article opens with an
emphasis on logic, saying ‘I hold that logic is what is fundamental in philosophy’
(p. 323). He does add almost at once ‘My own logic is atomic’, but the words ‘atom’ or
‘atomic’ never occur again in this article. Russell does make it clear that he assumes the
existence of many things, rather than just the one all-embracing thing, and that his
logic can accommodate this assumption. But the claim that these many things will
eventually be analysed into a number of unsplitable atoms plays no important role in his
discussion. This is as it should be. Russell’s greatest contribution to the development of
philosophy was his emphasis on logic, and the supposed atomism is really neither here
nor there, as he himself sometimes admits.1

By ‘logic’ Russell understands more than we may at first think of. He includes of course
what we nowadays call ‘formal’ or ‘mathematical’ logic at an elementary level, and
which we see as aiming to express very general truths by using schematic letters which
take the place of any proposition, or any name, or any predicate. We should not forget
that although Russell cannot quite be credited with having introduced such a logic,2
still it is his advocacy that has now made this elementary logic into a tool which all who
wish to study philosophy must master. The logic introduces clear and definite mean-
ings for its central notions, such as ‘if ’ and ‘or’ and ‘not’ in the logic of truthfunctors,
and ‘all’ and ‘some’ in the logic of quantifiers, and with these clear meanings it can state
and justify its quite definite rules for what is to be counted as a logical truth. The
current and widespread recognition of this point is due, in no small part, to the
influence of Russell’s thought and writing.
Of course his own ‘formal’ logic does not stop at this elementary level, but is
elaborated into a theory of types that ascends through all finite levels. It is clear that
Russell himself was never very happy with this theory. He had begun in PoM with the
view that all entities were of the same logical type, since every one of them was what he

1
On pp. 336–7 of LA Russell speaks of ‘simples’, and these are what his title has referred to as ‘atoms’. He
notes that any language must have expressions which are simple, in that their significance does not depend
upon the significance of their parts. He thinks that an ‘ideal’ language will use simple expressions only when
what they signify is itself simple, but he adds that it is no great loss if this ideal is not attained. (His later
comment in MPD dismisses the ideal as unimportant, pp. 165–6.)
2
In this work he himself cites Frege and Peano as his forerunners (LA, p. 324).
280 EPILOGUE

called a ‘term’. This meant that it could be named, and thereby made the subject of a
proposition.3 He had at that stage considered and rejected what Frege saw as a
fundamental distinction between concepts (or, more generally, functions) and objects,
a distinction which had led Frege to make the paradoxical claim that the concept horse
is not a concept. (For Frege had held that concepts could not be named, and the
expression ‘the concept horse’ apparently tries to do just this.) Russell in PoM found this
position intolerable. But at the same time he also became aware of what he always
called ‘the contradiction’, which comes about just because we are always willing to
turn any part of speech into a grammatical subject term, in order to speak about what it
signifies. So he was torn between a feature of language which is found (so far as I am
aware) in all natural languages, and the discovery that this feature leads to a contradic-
tion. The only conclusion to draw is that there must be something wrong with our
normal ways of speaking and thinking, and Russell should certainly be credited with
this discovery.
The question of what to do about it is one that vexed him for several years. He was
certainly much attracted by what he called his ‘substitutional’ theory, which did move a
good way towards Frege’s way of thinking. For this theory did not allow you to form
names either for classes or for the propositional functions that would define them, and
it did apparently provide a justification for the type restrictions which had been his first
and somewhat instinctive reaction to the problem. He was evidently very favourably
impressed by how far one could get with such a theory, which affirmed the existence
only of ordinary individuals and of propositions. But in the end it proved to be too
generous over propositions, since it treated them all as nameable, and of the same type
as one another, which once more led to contradictions. So in the end he was forced to
introduce the VCP, and to insist that propositions must be distinguished into different
orders. The order distinctions that are imposed on propositions naturally generate
similar order distinctions among the propositional functions that are abstracted from
them, and this becomes the ramified type theory that we find in PM. For substitution is
now abandoned, so propositions are no longer needed for the deductions, and can be
dropped without loss. The PM theory is just a theory of propositional functions,
distinguished both into the different levels of a simple type theory, and into the
different orders required by the VCP.
But Russell could not remain satisfied with this theory for very long. The distinc-
tions that are required by the VCP wreak havoc with the deductions of mathematics,
and have to be ameliorated by adding suitable axioms of reducibility. In PM (and earlier
in ‘The Regressive Method . . . ’ of [1907]) Russell does his best to argue that these
axioms should be accepted because they are needed, but it is clear from later references
that he did not find his own attempted justification very convincing.4 That is mainly
why I say that he was never very happy with the type theory that he eventually

3
PoM, }}47, 49.
4
E.g. IMP [1919: 193].
 EPILOGUE 281

reached. In my view he would have done better to drop the VCP, and the troubles that
it brings, and to return to the simple theory which is nearer to his original thinking. But
I personally am inclined to go further, and to say that even the simple theory is forced
to put too much weight on the notion of ‘systematic ambiguity’, and some quite
serious modifications are called for if this situation is to be improved. That is no doubt
debatable. But in any case Russell is surely right to say that our natural ways of talking
and thinking must be curbed if contradictions are to be avoided.
When Russell claims that ‘logic is what is fundamental in philosophy’ he does not
mean to confine his attention to the formal system which reaches its final shape in PM.
He certainly wishes to include what he has often referred to as ‘logical constructions’.
Towards the end of the article LA he sums up his overall position in this way: ‘The
business of philosophy, as I conceive it, is essentially that of logical analysis, followed by
logical synthesis’ (p. 341). He goes on to expand this characterization as follows:
Although comprehensive construction is part of the business of philosophy, I do not believe it is
the most important part. The most important part, to my mind, consists in criticizing and
clarifying notions which are apt to be regarded as fundamental and accepted uncritically. As
instances I might mention: mind, matter, consciousness, knowledge, experience, causality, will,
time. I believe all these notions to be inexact and approximate, essentially infected with
vagueness, incapable of forming part of any exact science. Out of the original manifold of events,
logical structures can be built which will have properties sufficiently like those of the above
common notions to account for their prevalence, but sufficiently unlike to allow a great deal of
error to creep in through their acceptance as fundamental. (LA, p. 341)

To paraphrase: logical analysis will aim to uncover what is of central importance in

these vague and inexact notions, and then logical synthesis will in many cases go on to
produce a reconstruction which conforms to Russell’s supreme maxim of scientific
philosophizing, namely: ‘Wherever possible, substitute constructions out of known
entities for inferences to unknown entities’ (LA, p. 326). It is mainly in the course of
these ‘constructions’ that he employs the technical methods that are provided by his
mathematical logic, and it is because he finds these methods so productive that he
frequently insists that a grasp of mathematical logic is essential. At the time when he was
writing LA, i.e. in 1924, the vast majority of professional philosophers did not possess
this familiarity with modern logic. But the situation now is very different.
Some analyses are purely eliminative, and do not provoke a succeeding synthesis.
The star example of this is Russell’s analysis of definite descriptions, which shows how
we do not need to think of them as having a naming role, but may treat them just as
Frege had already treated all the other quantifiers of a natural language. Somewhat
similar is Russell’s analysis of classes, in his ‘no-class’ theory, though it is perhaps better
to describe this as showing that there is no need to assume the existence of classes,
as commonly conceived. For we can do all we want—or so Russell claims—by
introducing instead an extensional way of speaking of propositional functions. That
was also the idea behind his multiple-relation theory of propositions, which aimed
282 EPILOGUE

to abolish talk of propositions as abstract entities, and to show how we could say what
we wanted to say by speaking instead of concrete episodes of judgment. (But this
theory was never fully developed.) In contrast to these purely eliminative analyses, we
have looked in some detail at three of Russell’s major ‘logical constructions’, namely
his constructions of classical mathematics, of material objects, and of minds.5 Again,
they are each rather different.
In the case of mathematics one might fairly say that much of the needed ‘analysis’
had already been done by others in the preceding century, notably by Weierstrass,
Cantor, and Dedekind. Russell had to add an analysis of the basic and underlying logic
(for he had himself shown that Frege’s logic was inadequate), but apart from that his
problem was mainly a work of synthesis, putting together into one coherent whole a
lot of results which were largely due to others. The volume did also contain some new
results of his own (most notably the treatment of ‘relation arithmetic’ in volume 2 of
PM), but I pass over that.
The constructions of matter and of mind were almost all Russell’s own work, for
there was very little that had been introduced by others and that was relevant to these
projects. In each case his approach was to look hard at the empirical evidence which we
have for our ordinary theory of these topics, and then to attempt a reconstruction
which went as little beyond this evidence as it could. It is fair to say that these
constructions do leave some rather noticeable gaps. As we have noted, his account of
the external world makes an assumption about places—i.e. on what counts as the same
place, or a nearby place—which is evidently far too simple, though it is of central
importance to him. As we have also noted (p. 192) he did later come to realize that
more was required, but he never tried to elaborate a revised theory himself, and this
task was left to his successors.6 This was no doubt because his own interests in the
external world quite soon moved on, to take more account of the rather startling views
proposed in the new physical theories of relativity and quantum mechanics.7 In
comparison with these, further work on the empirical evidence that might justify a
merely common-sense view must have seemed rather humdrum. There are some
equally noticeable gaps in Russell’s attempted construction of minds, which arise
because the apparatus which he allows himself to use—i.e. just sensations and images,
and nothing else—really is too restricted for this task. So far as I am aware, no one since
Russell has ever aimed to tackle this project with such limited resources. But on this
topic I would like to draw attention to a different aspect of Russell’s approach, which
was quite significant: he emphasizes the point that animals have minds too. It is true

5
I have discussed his construction of the points and instants of space and time in my [2010]. But the
discussion is complex, and I do not repeat it in this book.
6
The project of providing a phenomenalist construction of the external world was taken up by the logical
positivists. English speakers will be most familiar with the version of their thinking that is given by A. J. Ayer in his
[1936], esp. ch. 3, his [1940], esp. ch. 5, and then further elaborated in his [1947/8], where its difficulties are seen.
7
These topics take up much of AMatter [1927a]. Russell had already published his ABC of Relativity in
[1925].
 EPILOGUE 283

that the behaviourist school of philosophy was already quite familiar with this fact, but
Russell was unusual in bringing it to the attention of mainstream philosophers.
Russell contributed to the analysis of many other topics too, including all those listed
in the passage which I cited on p. 281. It is in large measure his emphasis on the
importance of such ‘analysis’ that has led to the whole of the recent century being
thought of as the age of ‘analytical philosophy’. No doubt it can be said that he himself
had a somewhat elastic understanding of just what kind of philosophical investigation
did count as ‘analysis’, and his successors have been ready to extend the notion yet
further. But we can all agree that the ‘clarification’ which Russell puts at the head of his
account is recognized by all practitioners as a crucial component. Of course, Russell
was highly inventive too, and a prolific source of new and arresting theories, but the
virtue of clarity is to be found in everything that he wrote. In this respect, and in the
range of topics that he tackled, he has left his mark on all subsequent philosophers far
more than anyone else who was active in the subject between 1900 and 1925.