Thin Objects

Thin 0 bj ects
An Abstractionist Account
0ystein Linnebo
OXFORD
UNIVERSITY PRESS
For my daughters Alma and Frida
Contents
Preface xi
Part I. Essentials
1. In Search of Thin Objects 3
1.1 Introduction 3
1.2 Coherentist Minimalism s
1.3 Abstractionist Minimalism 7
1.4 The Appeal of Thin Objects 9
1.5 Sufficiency and Mutual Sufficiency 11
1.6 Philosophical Constraints 13
1.7 Two Metaphysical "Pictures" 17
2. Thin Objects via Criteria of Identity 21
2.1 My Strategy in a Nutshell 21
2.2 A Fregean Concept of Object 23
2.3 Reference to Physical Bodies 26
2.4 Reconceptualization 30
2.5 Reference by Abstraction 33
2.6 Some Objections and Challenges 37
2.6.1 The bad company problem 38
2.6.2 Semantics and metasemantics 38
2.6.3 A vicious regress? 39
2.6.4 A clash with Kripke on reference? 40
2.6.5 Internalism about reference 41
2.7 A Candidate for the Job 42
2.8 Thick versus Thin 4S
Appendix 2.A Some Conceptions of Criteria ofldentity 46
Appendix 2.B A Negative Free Logic 48
Appendix 2.C Abstraction on a Partial Equivalence 49
3. Dynamic Abstraction Sl
3.1 Introduction Sl
3.2 Neo-Fregean Abstraction S3
3.3 How to Expand the Domain SS
3.4 Static and Dynamic Abstraction Compared 60
3.5 Iterated Abstraction 61
3.6 Absolute Generality Retrieved 64
3.7 Extensional vs. Intensional Domains 66
viii CONTENTS
Appendix 3.A Further Questions 70

3.A.l The higher-order needs of semantics 70
3.A.2 Abstraction on intensional entities 70
3.A.3 The need for a bimodal logic 71
3.A.4 The correct propositional logic 73
Appendix 3.B Proof of the Mirroring Theorem 74
Part II. Comparisons

4. Abstraction and the Question of Symmetry 77
4.1 Introduction 77
4.2 Identity of Content 79
4.3 Rayo on "Just is" -Statements 81
4.4 Abstraction and Worldly Asymmetry 83
5. Unbearable Lightness ofBeing 87
5.1 Ultra-Thin Conceptions of Objecthood 87
5.2 Logically Acceptable Translations 89
5.3 Semantically Idle Singular Terms 90
5.4 Inexplicable Reference 92
Appendix 5.A Proofs and Another Proposition 94
6. Predicative vs. Impredicative Abstraction 95
6.1 The Quest for Innocent Counterparts 95
6.2 Two Forms of!mpredicativity 96
6.3 Predicative Abstraction 98
6.3.1 Two-sorted languages 98
6.3.2 Defining the translation 100
6.3.3 The input theory 100
6.3.4 The output theory 102
6.4 Impredicative Abstraction 103
Appendix 6.A Proofs 106
7. The Context Principle 107
7.1 Introduction 107
7.2 How Are the Numbers "Given to Us"? 108
7.3 The Context Principle in the Grundlagen 110
7.4 The "Reproduction" of Meaning 114
7.5 The Context Principle in the Grundgesetze 117
7.6 Developing Frege's Explanatory Strategy 123
7.6.1 An ultra-thin conception of reference 123
7.6.2 Semantically constrained content recarving 124
7.6.3 Towards a metasemantic interpretation 127
7.7 Conclusion 129
Appendix 7.A Hale and Fine on Reference by Recarving 129
CONTENTS ix
Part III. Details

8. Reference by Abstraction 135
8.2 The Linguistic Data 137
8.3 Two Competing Interpretations 140
8.4 Why the Non-reductionist Interpretation is Preferable 143
8.4.1 The principle of charity 143
8.4.2 The principle of compositionality 144
8.4.3 Cognitive constraints on an interpretation 146
8.5 Why the Non-reductionist Interpretation is Available 148
8.6 Thin Objects 151
Appendix 8.A The Assertibility Conditions 153
Appendix 8.B Comparing the Two Interpretations 155
Appendix 8.C Internally Representable Abstraction 156
Appendix 8.D Defining a Sufficiency Operator 157
9. The Julius Caesar Problem 159
9.2 What is the Caesar Problem? 160
9.3 Many-sorted Languages 162
9.4 Sortals and Categories 163
9.5 The Uniqueness Thesis 166
9.6 Hale and Wright's Grundgedanke 167
9.7 Abstraction and the Merging of Sorts 169
Appendix 9.A The Assertibility Conditions 171
Appendix 9.B A Non-reductionist Interpretation 173
Appendix 9.C Defining a Sufficiency Operator 174
10. The Natural Numbers 176
10.2 The Individuation of the Natural Numbers 176
10.3 Against the Cardinal Conception 178
10.3.1 The objection from special numbers 179
10.3.2 The objection from the philosophy oflanguage 180
10.3.3 The objection from Jack of directness 181
10.4 Alleged Advantages of the Cardinal Conception 182
10.5 Developing the Ordinal Conception 183
10.6 Justifying the Axioms of Arithmetic 185
11. The Question of Platonism 189
11.1 Platonism in Mathematics 189
11.2 Thin Objects and Indefinite Extensibility 191
11.3 Shallow Nature 192
11.4 The Significance of Shallow Nature 195
11.5 How Beliefs are Responsive to Their Truth 197
11.6 The Epistemology of Mathematics 201
X CONTENTS
12. Dynamic Set Theory 205

12. lIntroduction 205
12.2 Choosing a Modal Logic 206
12.3 Plural Logic with Modality 208
12.4 The Nature of Sets 211
12.4. l The extensionality of sets 211
12.4.2 The priority of elements to their set 212
12.4.3 The extensional definiteness of subsethood 213
12.5 Recovering the Axioms ofZF 214
12.5.l From conditions to sets 214
12.5.2 Basic modal set theory 216
12.5.3 Full modal set theory 217
Appendix 12.A Proofs of Formal Results 219
Appendix 12.B A Harmless Restriction 222
Bibliography 223
Index 233
Preface
This book is about a promising but elusive idea. Are there objects that are "thin" in the
sense that their existence does not make a substantial demand on the world? Frege
famously thought so. He claimed that the equinumerosity of the knives and the forks
on a properly set table suffices for there to be objects such as the number of knives
and the number of forks, and for these objects to be identical. Versions of the idea of
thin objects have been defended by contemporary philosophers as well. For example,
Bob Hale and Crispin Wright assert that
what it takes for "the number of Fs = the number of Gs" to be true is exactly what it takes for
the Fs to be equinumerous with the Gs, no more, no less.[ ... ] There is no gap for metaphysics
to plug. 1
The truth of the equinumerosity claim is said to be "conceptually sufficient" for the
truth of the number identity (ibid.). Or, as Agustin Rayo colorfully puts it, once God
had seen to it that the Fs are equinumerous with the Gs, "there was nothing extra she
had to do" to ensure the existence of the number of F and the number of G, and their
identity (Rayo, 2013, p. 4; emphasis in original).
The idea of thin objects holds great philosophical promise. If the existence of certain
objects does not make a substantial demand on the world, then knowledge of such
objects will be comparatively easy to attain. On the Fregean view, for example, it
suffices for knowledge of the existence and identity of two numbers that an unprob-
lematic fact about knives and forks be known. Indeed, the idea of thin objects may
well be the only way to reconcile the need for an ontology of mathematical objects
with the need for a plausible epistemology. Another attraction of the idea of thin
objects concerns ontology. If little or nothing is required for the existence of objects
of some sort, then no wonder there is an abundance of such objects. The less that
is required for the existence of certain objects, the more such objects there will be.
Thus, if mathematical objects are thin, this will explain the striking fact that math-
ematics operates with an ontology that is far more abundant than that of any other
science.
The idea of thin objects is elusive, however. The characterization just offered is
imprecise and partly metaphorical. What does it really mean to say that the existence
of certain objects "makes no substantial demand on the world"? Indeed, if the truth
of "the number of Fs = the number of Gs" requires no more than that of "the Fs are
1
(Hale and Wright, 2009b, pp. 187 and 193). Both of the passages quoted in this paragraph have been
adapted slightly to fit our present example.
xii PREFACE
equinumerous with the Gs", perhaps the former sentence is just a fafon de parler for
the latter. To be convincing, the idea of thin objects has to be properly explained.
This book attempts to develop the needed explanations by drawing on some
Fregean ideas. I should say straight away, though, that my ambitions are not primarily
exegetical. I use some Fregean ideas that I find interesting in an attempt to answer
some important philosophical questions. By and large, I do not claim that the
arguments and views developed in this book coincide with Frege's. Some of the views
I defend are patently un -Fregean.
My strategy for making sense of thin objects has a simple structure. I begin with
the Fregean idea that an object, in the most general sense of the word, is a possible
referent of a singular term. The question of what objects there are is thus transformed
into the question of what forms of singular reference are possible. This means that
any account that makes singular reference easy to achieve makes it correspondingly
easy for objects to exist. A second Fregean idea is now invoked to argue that singular
reference can indeed be easy to achieve. According to this second idea, there is a
close link between reference and criteria of identity. Roughly speaking, it suffices
for a singular term to refer that the term has been associated with a specification
of the would-be referent, which figures in an appropriate criterion of identity. For
instance, it suffices for a direction term to refer that it has been associated with a
line and is subject to a criterion of identity that takes two lines to specify the same
direction just in case they are parallel. 2 In this way, the second Fregean idea makes easy
reference available. And by means of the first Fregean idea, easy reference ensures easy
being. My strategy for making sense of thin objects can thus be depicted by the upper
two arrows (representing explanatory moves) in the following triangle of interrelated
concepts:
reference
/~
objecthood - - - - - - - - - . identity criteria
(The lower arrow will be explained shortly.)

My concern with criteria of identity leads to an interest in abstraction principles,
which are principles of the form:
(AP) §a = §{3 = a "' f3
2
Admittedly. we would obtain a better fit with our ordinary concept of direction by considering instead
directed lines or line segments and the equivalence relation of "co-orientation': defined as parallelism plus
sameness of orientation. We shall keep this famous example unchanged, however, as the mentioned wrinkle
does not affect anything of philosophical importance.
PREFACE xiii
where a and f3 are variables of some type, § is an operator that applies to such
variables to form singular terms, and "' stands for an equivalence relation on the
kinds of items over which the variables range. An example made famous by Frege is
the aforementioned principle that the directions of two lines are identical just in case
the lines are parallel. My preferred way of understanding an abstraction principle is
simply as a special type of criterion of identity.
How does my proposed route to thin objects compare with others explored in the
literature? My debt to Frege is obvious. I have also profited enormously from the
writings of Michael Dummett and the neo-Fregeans Bob Hale and Crispin Wright.
As soon as one zooms in on the conceptual terrain, however, it becomes clear that the
route to be traveled in this book diverges in important respects from the paths already
explored. Unlike the neo-Fregeans, I have no need for the so-called "syntactic priority
thesis': which ascribes to syntactic categories a certain priority over ontological ones.
And I am critical of the idea of "content recarving': which is central to Frege's project
in the Grundlagen (but not, I argue, in the Grundgesetze) and to the projects of the
neo-Fregeans as well as Rayo.
My view is in some respects closer to Dummett's than to that of the neo-Fregeans.
I share Dummett's preference for a particularly unproblematic form of abstraction,
which I call predicative. On this form of abstraction, any question about the "new"
abstracta can be reduced to a question about the "old" entities on which we abstract.
A paradigm example is the case of directions, where we abstract on lines to obtain
their directions. This abstraction is predicative because any question about the result-
ing directions can be answered on the basis solely of the lines in terms of which
the directions are specified. I argue that predicative abstraction principles can be
laid down with no presuppositions whatsoever. But my argument does not extend
to impredicative principles. This makes predicative abstraction principles uniquely
well suited to serve in an account of thin objects. My approach extends even to the
predicative version of Frege's infamous Basic Law V. This "law" serves as the main
engine of an abstractionist account of sets that I develop and show to justify the strong
but widely accepted set theory ZF.
The restriction to predicative abstraction results in an entirely natural class of
abstraction principles, which has no unacceptable members (or so-called "bad
companions"). My account therefore avoids the "bad company problem': Instead,
I face a complementary challenge. Although predicative abstraction principles are
uniquely unproblematic and free of presuppositions, they are mathematically weak.
My response to this challenge consists of a novel account of "dynamic abstraction':
which is one of the distinctive features of the approach developed in this book. Since
abstraction often results in a larger domain, we can use this extended domain to
provide criteria of identity for yet further objects, which can thus be obtained by
further steps of abstraction. (This observation is represented by the lower arrow in
the above diagram.) The successive "formation" of sets described by the influential
iterative conception of sets is just one instance of the more general phenomenon of
xiv PREFACE
dynamic abstraction. Legitimate abstraction steps are iterated indefinitely to build

up ever larger domains of abstract objects. Dynamic abstraction can be seen as a
development and extension of the famous iterative conception of sets.
A second distinctive feature of my approach is the development of the idea of
thin objects. Suppose we speak a basic language concerned with a certain range of
entities (say, lines). Suppose ~ is an equivalence relation on some of these entities (say,
parallelism). Then it is legitimate to adopt an extended language in which we speak
precisely as if we have successfully abstracted on ~ (say, by speaking also about the
directions of the lines with which we began). I argue we have reason to ascribe to this
extended language a genuine form of reference to abstract objects. Since these objects
need not be in the domain of the original language, we can introduce yet another
language extension, where we talk about yet more objects. In fact, there is no end to
this process of forming ever more expressive languages.
Some words about methodology are in order. I make fairly extensive use oflogical
and mathematical tools. Formal definitions are provided, and theorems proved. I am
under no illusions about what this methodology achieves. As Kripke observes, "There
is no mathematical substitute for philosophy" (Kripke, 1976, p. 416). Definitions and
theorems do not by themselves solve any philosophical problems, at least not of the
sort that will occupy us here. The value of the formal methods to be employed lies
in the precision and rigor that they make possible, not in replacing more traditional
philosophical theorizing. But experience shows that precision matters in the discus-
sions that will concern us. It is therefore scientifically inexcusable not to aspire to a
high level of precision. In fact, much of the material to be discussed lends itself to
a mathematically precise investigation. While the use of formal methods does not by
itself solve any philosophical problems, it imposes an intellectual discipline that makes
it more likely that our philosophical arguments will bear fruit. 3
A quick overview of the book may be helpful. Part I is intended as a self-contained
introduction to the main ideas developed in the book as a whole. Chapter 1 sets the
stage by introducing the idea of thin objects, explaining its attractions as well as some
difficulties. This discussion culminates in a detailed "job description" for the idea of
thin objects. This job description is formulated in terms of a notion of one claim
sufficing for another-although the ontological commitments of the latter exceed
those of the former. By formulating some constraints on the notion of sufficiency,
I provide a precise characterization of what it would take to substantiate the idea
of thin objects. Chapter 2 introduces my own candidate for the job. I explain the
Fregean conception of objecthood and the idea that an appropriate use of criteria of
identity can suffice to constitute relations of reference. Chapter 3 introduces the idea of
dynamic abstraction. The form of abstraction explained in Chapter 2 can be iterated,
3
Compare (Williamson, 2007).
PREFACE XV
resulting in ever larger domains. I argue that this dynamic approach is superior to the
dominant "static" approach, both philosophically and technically.
Part II compares my own approach with some other attempts to develop the idea
of thin objects. I begin, in Chapter 4, by describing and criticizing some symmet-
ric conceptions of abstraction according to which the two sides of an acceptable
abstraction principle provide different "recarvings" of one and the same content.
In Chapter 5, I explain and reject some "ultra-thin'' conceptions of reference and
objecthood, which go much further than my own thin conception. One target is Hale
and Wright's "syntactic priority thesis", which holds that it suffices for an expression to
refer that it behaves syntactically and inferentially just like a singular term and figures
in a true (atomic) sentence. The ultra-thin conceptions make the notion of reference
semantically idle, I argue, and give rise to inexplicable relations of reference. The
important distinction between predicative and impredicative abstraction is explained
in Chapter 6. I argue that the former type of abstraction is superior to the latter, at least
for the purposes of developing the idea of thin objects. Only predicative abstraction
allows us to make sense of the attractive idea of there being no "metaphysical gap"
between the two sides of an abstraction principle. Finally, in Chapter 7, I discuss a
venerable source of motivation for the approach pursued in this book, namely Frege's
context principle, which urges us never to ask for the meaning of an expression in
isolation but only in the context of a complete sentence. Various interpretations of
this influential but somewhat obscure principle are discussed, and its role in Frege's
philosophical project is analyzed.
Part III spells out the ideas introduced in Part I. I begin, in Chapter 8, by developing
in detail an example of how an appropriate use of criteria of identity can ensure
easy reference. Chapter 9 addresses the Julius Caesar problem, which concerns cross-
category identities such as "Caesar = 3''. Although logic leaves us free to resolve
such identities in any way we wish, I observe that our linguistic practices often
embody an implicit choice to regard such identities as false. Chapter 10 examines
the important example of the natural numbers. I defend an ordinal conception of the
natural numbers, rather than the cardinal conception that is generally favored among
thinkers influenced by Frege. The penultimate chapter returns to the question of how
thin objects should be understood. While my view is obviously a form of ontological
realism about abstract objects, this realism is distinguished from more robust forms
of mathematical Platonism. I use this slight retreat from Platonism to explain how
thin objects are epistemologically tractable. The final chapter applies the dynamic
approach to abstraction to the important example of sets. This results in an account
of ordinary ZFC set theory.
The major dependencies among the chapters are depicted by the following diagram.
The via brevissima provided by Part I is indicated in bold.
Xvi PREFACE
11
9
/
12
i
8
i i
'~f/10
I~'
I
Many of the ideas developed in this book have had a long period of gestation.
The central idea of thin objects figured prominently already in my PhD dissertation
(Linnebo, 2002b) and an article (later abandoned) from the same period (Linnebo,
2002a). At first, this idea was developed in a structuralist manner. Later, an abstrac-
tionist development of the idea was explored in (Linnebo, 2005) and continued in
(Linnebo, 2008) and (Linnebo, 2009b). These three articles contain the germs oflarge
parts of this book, but are now entirely superseded by it. The idea of invoking thin
objects to develop a plausible epistemology of mathematics has its roots in the final
section of (Linnebo, 2006a). The second distinctive feature of this book-namely that
of dynamic abstraction-has its origins in (Linnebo, 2006b) and (Linnebo, 2009a)
(which was completed in 2007).
Some of the chapters draw on previously published material. In Part I, the opening
four sections of Chapter 1 are based on (Linnebo, 2012a), which is now superseded
by this chapter. Section 2.3 derives from Section 4 of (Linnebo, 2005), which (as
mentioned) is superseded by this book. The remaining material is mostly new. In
Part II, Sections 4.2 and 4.3 are based on (Linnebo, 2014), and Section 6.2 on (Linnebo,
2016a). These two articles expand on the themes of Chapters 4 and 6, respectively.
Chapter 7 closely follows (Linnebo, forthcoming). In Part III, Chapters 8, 10, and 12
are based on (Linnebo, 2012b), (Linnebo, 2009c), and (Linnebo, 2013), respectively,
but with occasional improvements. Chapter 9 and Section 11.5 make some limited
use of (Linnebo, 2005) and (Linnebo, 2008), respectively, both of which are (as
mentioned) superseded by this book.
There are many people to be thanked. Special thanks to Bob Hale and Agustin
Rayo for our countless discussions and their sterling contribution as referees for
Oxford University Press, as well as to Peter Momtchiloff for his patience and sound
advice. I have benefited enormously from written comments and discussions of ideas
PREFACE xvii
developed in this manuscript; thanks to Solveig Aasen, Bahram Assadian, Neil Barton,
Rob Bassett, Christian Beyer, Susanne Bobzien, Francesca Boccuni, Einar Duenger
B0hn, Roy Cook, Philip Ebert, Matti Eklund, Anthony Everett, Jens Erik Fenstad,
Salvatore Florio, Dagfinn F0llesdal, Peter Fritz, Olav Gjelsvik, Volker Halbach, Mirja
Hartimo, Richard Heck, Simon Hewitt, Leon Horsten, Keith Hossack, Torfinn
Huvenes, Nick Jones, Frode Kjosavik, J6nne Kriener, James Ladyman, Hannes Leitgeb,
Jon Litland, Michele Lubrano, Jonny Mcintosh, David Nicolas, Charles Parsons, Alex
Paseau, Jonathan Payne, Richard Pettigrew, Michael Rescorla, Sam Roberts, Marcus
Rossberg, Ian Rumfitt, Andrea Sereni, Stewart Shapiro, James Studd, Tolgahan Toy,
Rafal Urbaniak, Gabriel Uzquiano, Albert Visser, Sean Walsh, Timothy Williamson,
Crispin Wright, as well as the participants at a large number of conferences and
workshops where this material was presented. Thanks to Hans Robin Solberg for
preparing the index. This project was initiated with the help of an AHRC-funded
research leave (grant AH/E003753/l) and finally brought to its completion during
two terms as a Visiting Fellow at All Souls College, Oxford. I gratefully acknowledge
their support.
PART I
Essentials
1
In Search of Thin Objects
1.1 Introduction
Kant famously argued that all existence claims are synthetic.1 An existence claim
can never be established by conceptual analysis alone but always requires an appeal
to intuition or perception, thus mal<ing the claim synthetic. This view is boldly
rejected in Frege's Foundations of Arithmetic {Frege, 1953), where Frege defends an
account of arithmetic that combines a form of ontological realism with logicism. His
realism consists in taking arithmetic to be about real objects existing independently
of all human or other cognizers. And his logicism consists in tal<ing the truths
of pure arithmetic to rest on just logic and definitions and thus be analytic. Most
philosophers now probably agree with Kant in this debate and deny that the existence
of mathematical objects can be established on the basis of logic and conceptual
analysis alone. This is why George Boolos, only slightly tongue-in-cheek, can offer a
one-line refutation ofFregean logicism: "Arithmetic implies that there are two distinct
numbers" (Boolos, 1997, p. 302), whereas logic and conceptual analysis-Boolos takes
us all to know-cannot underwrite any existence claims (other than perhaps of one
object, so as to streamline logical theory). 2
However, the disagreement between Kant and Frege lives on in a different form.
Even if we concede that there are no analytic existence claims, we may ask whether
there are objects whose existence does not {loosely speal<ing) make a substantial
demand on the world. That is, are there objects that are "thin'' in the sense that their
existence does not (again loosely speal<ing) amount to very much? Presumably, an
analytic truth does not make a substantial demand on the world. 3 But perhaps being
analytic is not the only way to avoid imposing a substantial demand. Instead of asking
Frege's question of whether there are existence claims that are analytic, we can ask the
broader question of whether there are existence claims that are "non-demanding" -in
some sense yet to be clarified.
A number of philosophers have been attracted to this idea. Two classic examples
are found in the philosophy of mathematics. First, there is the view that the existence
1 2
See (Kant, 1997, B622-3). See also (Boolos, 1997, pp. 199 and 214).
3 Analyticity must here be understood in a metaphysical rather than epistemological sense (Boghossian,
1996). I cannot discuss here whether Frege's rationalism led him to depart from a traditional conception of
(metaphysical) analyticity. See (Macfarlane, 2002) for some relevant discussion.
4 IN SEARCH OF THIN OBJECTS
of the objects described by a theory of pure mathematics amounts to nothing more

than the consistency or coherence of this theory. This view has been held by many
leading mathematicians and continues to exert a strong influence on contemporary
philosophers of mathematics. 4 Then, there is the view associated with Frege that the
equinumerosity of two concepts suffices for the existence of a number representing
the cardinality of both concepts. For instance, the fact that the knives and the forks
on a table can be one-to-one correlated is said to suffice for the existence of a number
that represents the cardinality of both the knives and the forks. 5 Agustin Rayo nicely
captures the idea when he writes that a "subtle Platonist" such as Frege
believes that for the number of the Fs to be eight just is for there to be eight planets. So when
God created eight planets she thereby made it the case that the number of the planets was eight.
(Rayo, 2016, p. 203; emphasis in original)
I am not claiming that there is a single, sharply articulated view underlying all these
views, only that they are all attempts to develop the as-yet fuzzy idea that there are
objects whose existence does not make a substantial demand on the world.
We have talked about objects being thin in an absolute sense, namely that their
existence does not make a substantial demand on the world. An object can also be thin
relative to some other objects if, given the existence of these other objects, the existence
of the object in question makes no substantial further demand. Someone attracted to
the view that pure sets are thin in the absolute sense is likely also to be attracted to the
view that an impure set is thin relative to the urelements (i.e. non-sets) that figure in
its transitive closure. The existence of a set of all the books in my office, for example,
requires little or nothing beyond the existence of the books. Moreover, a mereological
sum may be thin relative to its parts. For example, the existence of a mereological sum
of all my books requires little or nothing beyond the existence of these books. 6
I shall refer to any view according to which there are objects that are thin in either
the absolute or the relative sense as a form of metaontological minimalism, or just
minimalism for short. The label requires some explanation. While ontology is the
study of what there is, metaontology is the study of the key concepts of ontology, such
as existence and objecthood. 7 A view is therefore a form of metaontological min-
imalism insofar as it holds that existence and objecthood have a minimal character.
Minimalists need not hold that all objects are thin. Their claim is that our concept of an
object permits thin objects. Additional "thickness" can of course derive from the kind
of object in question. Elementary particles, for example, are thick in the sense that
their existence makes a substantial demand on the world. But their thickness derives
from what it is to be an elementary particle, not from what it is to be an object.
4 See for instance (Parsons, 1990), (Resnik, 1997), and (Shapiro, 1997).
5 See for instance (Wright, 1983) and the essays collected in (Hale and Wright, 200!a).
6
Philosophers attracted to this view include (Lewis, 1991, Section 3.6) and (Sider, 2007).
7 See for instance (Eklund, 2006a).
COHERENTIST MINIMALISM 5
Metaontological minimalism has consequences concerning ontology proper. The

thinner the concept of an object, the more objects there tend to be. Metaontological
minimalism thus tends to support a generous ontology. 8 By contrast, a generous
ontology does not by itself support metaontological minimalism. The universe might
just happen to contain an abundance of objects whose existence makes substantial
demands on the world.
Just as metaontological minimalists are heirs to the Fregean view that there are
analytic existence claims, there are also heirs to the contrasting Kantian view. Hartry
Field has attacked the idea that mathematical objects are thin, sometimes mentioning
the Kantian origin of his criticism.9 And various metaphysicians reject the idea
that mereological sums are thin relative to their parts. 10 Just as with the original
Kantian rejection of analytic existence claims, this contemporary rejection of thin
objects strikes many philosophers as plausible. Metaontological minimalism can
come across as a piece of philosophical magic that aspires to conjure up something
out of nothing-or, in the relative case, to conjure up more out of less.
The chapter is organized as follows. In the next two sections, I outline two influ-
ential approaches to the idea of thin objects that are found in the philosophy of
mathematics and that were mentioned above. Then, I examine the appeal of the idea.
Based on this examination, I formulate some logical and philosophical constraints
that any viable form of metaontological minimalism must satisfy. We thus obtain
a "job description': and the task of the book is to find a suitable candidate for the
job. The chapter ends with an attempt to dramatically reduce the field of acceptable
candidates by rejecting the customary symmetric conception of abstraction in favor
of an asymmetric conception. The left-hand side of an abstraction principle makes
demands on the world that go beyond those of the right-hand side. Thin objects are
nevertheless secured because the former demands do not substantially exceed the
latter. For the truths on the left are grounded in the truths on the right.
1.2 Coherentist Minimalism

One classic example of metaontological minimalism is the view that the coherence
of a mathematical theory suffices for the existence of the objects that the theory
purports to describe. Since it is coherent to supplement the ordinary real number
line IR with two infinite numbers -oo and +oo, for example, the extended real
number line i = IR U {-oo, +oo} exists. And since it is coherent to supplement
IR with the imaginary unit i = .J=T and all the other complex numbers, the complex
field C exists. All that the existence of these new mathematical objects involves,
according to the view in question, is the coherence of the theories that describe the
relevant structures. Let us refer to this as a coherentist approach to thin objects.
8 See (Eklund, 2006b) for a discussion of some extremely abundant ontologies that may arise in this way.
9 10
See (Field, 1989, pp. 5 and 79-80). See for instance (Rosen and Dorr, 2002) .
This approach enjoys widespread support within mathematics itself and is defended
by several prominent mathematicians. In his correspondence with Frege, for example,
David Hilbert wrote:
As long as I have been thinking, writing and lecturing on these things, I have been saying
the exact reverse: if the arbitrarily given axioms do not contradict each other with all their
consequences, then they are true and the things defined by them exist. This is for me the
criterion of truth and existence. 11
As is well known, the word 'criterion' is ambiguous between a metaphysical meaning

(a defining characteristic) and an epistemological one (a mark by which something
can be recognized). Since the context favors the metaphysical reading, the passage
is naturally read as an endorsement of metaontological minimalism, not just of an
extravagant ontology.
A similar view is endorsed by Georg Cantor:
Mathematics is in its development entirely free and only bound in the self-evident respect that
its concepts must both be consistent with each other and also stand in exact relationships,
ordered by definitions, to those concepts which have previously been introduced and are
already at hand and established.12
It may be objected that, while this passage defends an extremely generous ontology,
it is not a defense of metaontological minimalism. In response, we observe that the
passage is concerned with what Cantor calls "immanent reality", which is a matter of
occupying "an entirely determinate place in our understanding". Cantor contrasts this
with "transient reality", which requires that a mathematical object be "an expression
or copy of the events and relationships in the external world which confronts the
intellect" (p. 895). He feels compelled to provide an argument that the former kind of
existence ensures the latter. The most plausible interpretation, I think, is that Cantor
seeks a form of metaontological minimalism with respect to immanent existence but
merely a generous ontology concerning transient existence.
The coherentist approach to thin objects has enjoyed widespread support among
philosophers as well. A structuralist version of the approach has in recent decades
been defended by central philosophers of mathematics such as Charles Parsons,
Michael Resnik, and Stewart Shapiro.13 For instance, Shapiro includes the following
"coherence principle" in his theory of mathematical structures:
Coherence: If <p is a coherent formula in a second-order language, then there is a structure that
satisfies <p. (Shapiro, 1997, p. 95)
11
Letter to Frege of December 29, 1899, in (Frege, 1980). See (Ewald, 1996, p. 1105) for another example
from Hilbert.
12
See (Cantor, 1883), translated in (Ewald, 1996, p. 896).
13 See the works cited in footnote 4 . Also relevant is the "equivalence thesis" of (Putnam, 1967).
ABSTRACTIONIST MINIMALISM 7
It is instructive to compare this principle with Tarski's semantic account of logical

consistency and consequence. On Tarski's analysis, a theory T is said to be seman-
tically consistent (or coherent) just in case there is a mathematical model of T. The
coherence principle can be regarded as a reversal of this analysis: we now attempt
to account for what models or structures there are in terms of what theories are
coherent. 14
Shapiro not only endorses the coherence principle but makes some striking claims
about its philosophical status. He compares the ontologically committed claim that
there is a certain mathematical structure with the (apparently) ontologically innocent
claim that it is possible for there to be instances of this structure. These claims are
"equivalent" (p. 96), he contends, and "[i]n a sense [... ] say the same thing, using
different primitives" (p. 97). Shapiro's view is thus a version of coherentist minimalism,
centered on the claim that
there is a model of T <=> T is coherent
where 1/f means that <p and 1/f "say the same thing''.
The coherentist approach can be extended to objects that are thin only in a
relative sense. Coherence does not suffice for the existence of "thick" objects such as
electrons. But given the existence of certain thick constituents, coherence may suffice
for the existence of further objects that are thin relative to these constituents. Given
the existence of two electrons, for example, their set and mereological sum may exist
simply because the existence of such objects is coherent.
Is coherentist minimalism tenable? I remain neutral on the question. My present
aim is to develop and defend an alternative form of minimalism based on Fregean
abstraction. My pursuit of this aim is unaffected by the success or failure of the
coherentist alternative.
1.3 Abstractionist Minimalism

Another classic example of metaontological minimalism derives from Frege and has
been developed by the neo-Fregeans Hale and Wright. Frege first argues (along lines
that will be outlined in Section 1.4) that there are abstract mathematical objects. He
then pauses to consider a challenge:
How, then, are the numbers to be given to us, if we cannot have any ideas or intuitions of them?
(Frege, 1953,§62)
That is, how can we have epistemic or semantic "access" to numbers, given that their
abstractness precludes any kind of perception of them or experimental detection?
14
This is not to say that we possess a notion of coherence that is independent of mathematics. Our view
on questions of coherence will be informed by and be sensitive to set theory. Here we use some mathematics
to explicate a philosophical notion, which in turn is used to provide a philosophical interpretation of
mathematics. See (Shapiro, 1997, pp. 135-6) for discussion.
Frege's response urges us to transform the question of how linguistic (or mental)
representations succeed in referring to natural numbers into the different question
of how complete sentences (or their mental counterparts) succeed in having their
appropriate arithmetical meanings:
Since it is only in the context of a sentence that words have any meaning, our problem becomes
this: To define the sense of a sentence in which a number word occurs. (Frege, 1953, §62)
This response raises some hard exegetical questions, which are discussed in Chapter 7.
But the argumentative strategy of the Grundlagen is made tolerably clear a few pages
later, where Frege makes a surprising claim about the relation between the parallelism
oflines and the identity of their directions:
The judgement "line a is parallel to line b'; or, using symbols, a II b, can be taken as an identity.
If we do this, we obtain the concept of direction, and say: "the direction ofline a is identical with
the direction of line b". Thus we replace the symbol II by the more generic symbol =, through
removing what is specific in the content of the former and dividing it between a and b. We carve
up the content in a way different from the original way, and this yields us a new concept.
(Frege, 1953,§64)
Consider the criterion of identity for directions:

(Dir)
Frege claims that the content of the right-hand side of this biconditional can be
"recarved" to yield the content of the left-hand side. The idea is that we get epistemic
and semantic access to directions by first establishing a truth about parallelism oflines
and then "recarving" this content so as to yield an identity between directions.
Let <p <=} 1/1 formalize the claim that <p and 1/1 are different "carvings" of the same
content-in a sense yet to be explicated. Then (Dir) can be strengthened to:
(Dir{?)
Inspired by this example, Frege and the neo-Fregeans seek to provide a logical
and philosophical foundation for classical mathematics on the basis of abstraction
principles, which generalize (Dir). These are principles of the form
(AP) §a = §{3 B Ci ~ {3
where a and f3 range over items of some sort, where ~ is an equivalence relation on
such items, and where § is an operator that maps such items to objects. One famous
example is Hume's Principle, which says that the number of Fs (symbolized as #F)
is identical to the number of Gs just in case the Fs and the Gs can be one-to-one
correlated (symbolized as F ~ G):
(HP) #F= #GB F~ G
As Frege discovered, this principle has an amazing mathematical property. When

added to second-order logic along with some natural definitions, we are able to
THE APPEAL OF THIN OBJECTS 9
derive all of ordinary (second-order Dedekind-Peano) arithmetic. 15 As we shall see,

abstraction principles are available not just for directions and numbers but for many
other kinds of abstract object as well, such as geometrical shapes and linguistic types.
What does all this mean? According to Frege and the neo-Fregeans, the left-hand
side of each successful abstraction principle (AP) provides a "recarving" of the content
of the corresponding right-hand side. 16 All that is required for the existence of the
objects §a and §{3 is that the equivalence relation ~ obtain between the entities
a and f3. All that is required for the existence of directions, for example, is the
parallelism of appropriate lines.
The abstractionist approach to minimalism can be extended to objects that are thin
only in the relative sense. It is possible to formulate abstraction principles for sets and
mereological sums, for example, which ensures that the existence of sets and sums
of thick objects does not make any substantial demand beyond the existence of their
thick constituents.
1.4 The Appeal of Thin Objects

Why are so many philosophers and mathematicians attracted to the idea of thin
objects? The most important reason is that metaontological minimalism promises a
way to accept face value readings of discourses whose ontologies would otherwise be
problematic. Arithmetic provides an example. The language of arithmetic contains
proper names which (it seems) purport to refer to certain abstract objects, namely
natural numbers, as well as quantifier phrases which (it seems) purport to range over
all such numbers. Moreover, a great variety of theorems expressed in this language
appear to be true. These theorems are asserted in full earnest by competent laypeople
as well as professional mathematicians. Since the arithmetical competence of these
people is beyond question, there is reason to believe that most of their arithmetical
assertions are true. But if these theorems are true, then their singular terms and
quantifiers must succeed in referring to and ranging over natural numbers. And for
this kind of success to be possible, there must exist abstract mathematical objects.
The argument is certainly valid. Is it sound? The premises can of course be
challenged-like everything else in philosophy. But they have great initial plausibility.
It would be appealing to take the apparent truth of the premises at face value, if
possible. This would save us the difficult task of showing how both laypeople and
experts are wrong about something they take to be obviously true. So the argument
provides at least some reason to believe that there exist mathematical objects such as
numbers.17
15 This result, which is known as Frege's Theorem, is hinted at in (Parsons, 1965) and explicitly stated in
(Wright, 1983). See (Boolos, 1990) for a nice proof.

16 Further evidence that the neo-Fregeans are pursuing the idea of thin objects, as understood here, is
provided in Section 6.4.

17 See (Linnebo, 2017c) for an overview of defenses of the premises, which, if successful, would support
a much stronger conclusion.

On the other hand, the ontology of abstract objects is often found to be

philosophically problematic. The epistemological challenge brought to general
philosophical attention by (Benacerraf, 1973) is well known. Since perception and
all forms of instrumental detection are based on causal processes, these methods
cannot give us access to abstract objects such as the natural numbers. How then can
we acquire knowledge ofthem? 18
Another worry is the perceived extravagance of the huge ontologies postulated by
contemporary mathematics. How can we postulate vast infinities of new objects with
such a light heart? No physicist would so unscrupulously postulate a huge infinity of
new physical objects. Why, then, should mathematicians get away with it? Of course,
philosophers are divided over how serious these worries are. But any successful
account of mathematical objects needs to have some response to the worries, even
if only to explain why they are misguided.
The idea of thin objects suggests a promising strategy for responding to the worries.
The vast ontology of mathematics may well be problematic when understood in a
thick sense. If mathematical objects are understood on the model of, say, elementary
particles, there would indeed be good reason to worry about epistemic access and
ontological extravagance. But this understanding of mathematical objects is not
obligatory. If there are such things as thin objects, then the existence of mathematical
objects need not make much of a demand on the world. It may, for instance, suffice
that the theory purporting to describe the relevant mathematical objects is coherent.
This would greatly simplify the problem of epistemic access. Although our knowledge
of the coherence of mathematical theories is still inadequately understood, it is
at least not a complete mystery in the way that knowledge of thick mathematical
objects would be. More generally, the less of a demand the existence of mathe-
matical objects makes on the world, the easier it will be to know that the demand
is satisfied. 19
Thin objects would help with the worry about ontological extravagance as well. If
mathematical objects are thin, the bar to existence is set very low. So it is only to be
expected that a generous ontology should result. This is just an instance of the general
phenomenon noted above, namely that metaontological minimalism tends to support
generous ontologies. Does this defense of a generous mathematical ontology conflict
with Occam's razor? The answer depends on how the razor is understood. If all the
razor says is that objects must never be postulated "beyond necessity" but must earn
18
An improved version of the challenge is developed in (Field, 1989). In (Linnebo, 2006a) I develop
a further improvement which I argue survives all extant attempts to answer or reject the challenge. Some
ideas about how to answer this improved challenge are found already in that paper but are set out in greater
detail below, especially in Section 11.5.
19
This response to Benacerraf's challenge must be distinguished from that of (Balaguer, 1998). As
I understand it, Balaguer's "full-blooded platonism" is primarily a very generous ontology. My present point,
however, is that metaontological minimalism promises to reduce the explanatory burden by equating the
existence of mathematical objects with some fact to which epistemic access is less problematic.
SUFFICIENCY AND MUTUAL SUFFICIENCY 11
their keep by enabling better explanations than would otherwise be possible, then
the razor may well be compatible with the generous ontology in question. If some
objects are metaphysically "cheap", even a modest contribution to our explanatory
power may suffice to justify their postulation. 20 It is also important to keep in mind
that the explanations in question need not be empirical but can be intra-mathematical.
There are other examples too of how thin objects can be philosophically useful.
Consider the philosophical debate about the existence of mereological sums. We
often speak as if there are various kinds of mereological sums, such as decks of
cards, bunches of grapes, crowds of people. And many of these claims appear to be
true. But some philosophers find mereological sums problematic, often because of
philosophical worries akin to the ones just discussed. If mereological sums are thin
relative to their parts-if, that is, little or nothing is required for their existence other
than the existence of their parts-then we would be in a good position to assuage
these worries.
In sum, if some form of metaontological minimalism can be articulated, its
explanatory potential will be great.
1.5 Sufficiency and Mutual Sufficiency

The principal task confronting any defender of metaontological minimalism is to
explain some locutions that we have so far left loose and intuitive, such as
all that is required for 1/1 is <p

<p is (conceptually) sufficient for 1/1
all that God had to do to ensure that 1/1 was to bring it about that <p
or the symmetrical analogue
<pis a recarving of the content of 1/1

<pand 1/1 make the same demand on the world
for it to be the case that 1/1, the symmetrical
analogue involved in the last three statements. 21 Let us refer to these as sufficiency and
mutual sufficiency statements, respectively. Our task is to provide a proper explanation
of the operator that figures in at least one of these types of statement and to show
how the resulting statements can be used to provide the attractive philosophical
explanations that we discussed in the previous section.
20
See (Schaffer, 2015) for a related observation.
21 I shall use the word 'statement' for a formula relative to some contextually salient assignment to its
free variables, if any.
It will be convenient to allow the sufficiency operator to take a set of formulas on

its left-hand side. Thus, we take r => 1/t to mean that the statements in the set r are
jointly sufficient for 1/t. If our language allows infinite conjunctions, then r => 1/t can
be regarded as just shorthand for (/\aeA <fJa) => 1/t' where r = {<fJa I a E A}. Since we
mostly work with finitary languages, however, the proposed modification does make
a difference.
What logical principles should be adopted to govern the new operator(s)? A
full answer must obviously await a proper explanation of the operator(s). But a
little bit can be said already at this stage. For the idea of thin objects to have any
promise, the operator(s) must be at least as strong as the corresponding material
conditional or biconditional. That is, 1/t must entail <p--+ 1/t; and likewise for
mutual sufficiency and the biconditional. What else? Consider first the sufficiency
operator. It is reasonable to adopt a principle of transitivity. If a set of truths suffice for
another set of truths, and if this second set of truths suffice for a third truth, then the
first set too suffices for the third truth. We formalize this as the following cut rule:
Cut
ri => <p; for each i E I
=> 1/t
{<p;} ie /
U;er r; => 1/t
We shall not at the outset assume any other logical principles governing the sufficiency
operator.
Next, consider the mutual sufficiency operator. Since this is meant to be an
equivalence relation, we assume logical principles to that effect. A trickier question
is for which properties this equivalence is a congruence; that is, for which operators
(')we have:
We make no assumptions at this stage but shall return to the question shortly.
Does it matter which of the two operators we choose as our primitive? Surely, one
might think, it must be possible to define either operator in terms of the other. Suppose
we choose => as our primitive. We can then define mutual sufficiency as two-way
ordinary sufficiency; that is, define 1/t as 1/t /\ 1/t => <p. Ifinstead we choose <=>
as our primitive, we can define 1/t as <p /\ 1/t. (This definition can be motivated
in terms of two natural assumptions: first, that <p /\ 1/t suffices for <p; and second, that
<p suffices for <p /\ 1/t just in case <p suffices for 1/t .) However, these arguments make
strong assumptions about the logic of the two operators. These assumptions have not
been granted. So at least until more has been said, we are not entitled to regard either
operator as definable in terms of the other.
Thus, the choice of one (or even both) of the operators as primitive may well matter.
Indeed, beginning in Section 1.7, I argue that the sufficiency operator is better suited
to deliver what we want than the mutual sufficiency operator.
PHILOSOPHICAL CONSTRAINTS 13
1.6 Philosophical Constraints

I now wish to formulate a "job description" for the desired notion of sufficiency or
mutual sufficiency. I shall concentrate on the case of sufficiency, for the reason just
mentioned, although my discussion is easily adapted to the case of mutual sufficiency.
The "job description'' should be written so as to ensure that any notion that fits the
bill delivers the philosophical benefits discussed in Section 1.4.
As discussed, Frege proposed to define <p ~ 1/t as A(<p ~ 1ft), where A is
an analyticity operator, and where analyticity is understood as truth in virtue of
meaning. 22 How does this proposal fare? Quinean objections to analyticity pose
an obvious threat. Perhaps there is no sharp or theoretically interesting distinction
between the analytic and the synthetic. Any attempt to draw such a distinction would
then be arbitrary. Can the Quinean objections be countered? The question has been
debated at great length. I do not here wish to join this debate. 23
Thankfully, there is no need to gauge the general health of the notion of analyticity
in order to assess the viability of the Fregean proposal. The proposal faces two serious
problems. One problem concerns de re sufficiency statements. Let me explain. Should
the formulas that flank the sufficiency operator be permitted to contain free variables?
The resulting sufficiency statements can be said to be de re, in obvious analogy with the
terminology used in quantified modal logic. This contrasts with de dicta sufficiency
statements, where only sentences (that is, formulas with no free variables) are allowed
to flank the operators. There is, in fact, strong pressure to accept sufficiency statements
that are de re, not merely de dicta. Consider the claim that there are thin objects. To
express this, we need to state that there are objects whose existence is undemanding,
that is, 3x(T =>Ex), where Ex is an existence predicate and Tis some tautology. The
same goes for the claim, discussed above, that there are objects (such as impure sets)
that are thin only in a relative sense, not absolutely.
With this background in place, the problem is easy to state. As (Quine, 1953b)
explained, even if we waive all concerns about the distinction between analytic and
synthetic sentences, this will only licence analyticity statements that are de dicta, not
de re. After all, it is only sentences that are analytic, not open formulas relative to
variable assignments. Analyticity is meant to be an entirely linguistic phenomenon,
whereas variable assignments typically involve non-linguistic objects.
A second problem concerns ontology. To fix the terminology-here and in what
follows-let us adopt the usual Quinean notion of ontological commitment, accord-
ing to which a sentence in the language of first-order logic is ontologically committed
to just such objects as must be assumed to be in the domain in order to make the
22
In terms of (Boghossian, l 996)'s influential distinction, this is a metaphysical rather than an episte-
mological notion of analyticity.
23
For the record, I believe Quine successfully undermines the ambitious notion of analyticity cham-
pioned by the logical positivists, but that a more modest notion may escape his attack. See (Putnam, l 97Sa)
for a similar view. In his later years, Quine too moved in this direction: see (Quine, 1991).
sentence true. A proponent of thin objects is interested in sufficiency statements

where the statement on the right-hand side of the sufficiency operator has ontological
commitments that exceed those of the statement on the left-hand side. A possible
example is the Fregean claim
where the right-hand side is ontologically committed to abstract directions, while

the left-hand side avoids such commitments. Our question is whether there are
analytically true conditionals <p --+ 1/1 where the ontological commitments of 1/1 exceed
those of </J.
Any such conditional would give rise to an analytic existence statement. To see
this, let 3x8 make explicit an ontological commitment that is had by 1/1 but not by <p.
We thus have A(<p--+ 3x8). By moving the quantifier 3x out through the parenthesis
(which is permissible since <p can be assumed not to contain x free), it follows
that A3x(<p--+ 8). As we saw in the opening of this chapter, however, Boolos-and
probably most other contemporary philosophers as well-deny that there are any
analytically true existence statements. 24 It is not hard to see why. An analytic truth
is supposed to be true in virtue of meanings, which are supposed to be cognitively
accessible to us. But the objects in question are supposed to be independent of us and
our linguistic and cognitive activities; for instance, Frege compares the existence of
the natural numbers with that of the North Sea (Frege, 1953, §26). How can meanings
that are so closely tied to our minds ensure the existence of objects external to our
minds and so robustly independent in their existence?
Of course, it was precisely this sort of concern about analytic existence statements
that prompted our investigation of thin objects in the first place. So let us set aside
analyticity and exploit the extra freedom gained by broadening our perspective to
the idea of thin objects, whose existence need not be analytic but nevertheless-in
some sense yet to be pinned down-makes no substantial demand on the world. Our
discussion motivates a constraint that any viable notion of sufficiency must satisfy: 25
Ontological expansiveness constraint
There are true sufficiency statements <p =} 1/1 where the ontological commitments
of 1/1 exceed those of <p.
Indeed, on an abstractionist approach to thin objects, we probably want true suf-

ficiency statements corresponding to the right-to-left direction of all permissible
abstraction principles. This generalizes the Fregean sufficiency claim (Dir=?).
24
We are here setting aside the possible commitment to a non-empty domain, which is often tolerated
in order to streamline our logic. This is permissible because 3xll can be chosen so as to involve a specific
commitment, say to directions, which goes beyond the commitment to a non-empty domain.
25 An analogous constraint would, of course, be needed for the notion of mutual sufficiency. The same
obviously goes for the constraints formulated below. Henceforth, I shall not remind readers of this.
PHILOSOPHICAL CONSTRAINTS 15
Might there be an easy way to obtain the desired sufficiency statements? Perhaps
the right-hand side is merely a fancy-and syntactically misleading-rewording of the
left-hand side. This proposal would certainly ensure that the right-hand side demands
no more of the world than the left-hand side. Whatever its other merits, the proposal
is clearly inadequate for our purposes. If the right-hand side is merely a fa~on de
par/er for the left-hand side, this would at most justify speaking as if there are abstract
objects. But our aim is to make sense of how there in fact are such objects and of how
we come to know about them. For a sufficiency claim to have this philosophical payoff,
both sentences must be taken at face value. Every singular term must function as such
semantically; that is, it must be in the business of referring to an object. And there must
be no singular reference other than what is effected by the singular terms that occur
in the relevant sentence. These considerations motivate the following constraint:
Face value constraint

The formulas involved in a sufficiency statement rp => 1/1 can be taken at face value
in our semantic analysis.
Our most significant discovery so far is that it would be too demanding to define
the sufficiency statement rp => 1/1 as the analytic conditional A(rp-+ 1/J). We need a less
demanding definition. An option that naturally comes to mind is to define 1/1
as the strict conditional D(rp-+ 1/1) (where ' O' stands for metaphysical necessity).
On this analysis, any necessarily existing object counts as requiring nothing for its
existence. The analysis thus collapses the idea of thin objects into the more familiar
idea of necessary existents. This loss of novelty is not itself problematic. The real
problem concerns the promised benefits of thin objects, as discussed in Section 1.4.
Assume, for instance, that there is a necessarily existing god. Let 1f! say so, and let rp be
any tautology. Then we would have <{J => 1f!, although knowledge of 1/1 must
be more demanding than the strict conditional.
Let us therefore try to formulate some further constraints on the sufficiency oper-
ator to ensure that the promised benefits materialize. One of the promised benefits is
a response to the epistemological challenge made famous by (Benacerraf, 1973). How
do we gain knowledge of abstract objects such as directions and numbers, given that
no causal interaction with them is possible? The promised answer is nicely illustrated
in terms of the Fregean example of directions:
(Dir)
Let <p and 1f! be the right- and the left-hand sides of this biconditional, respectively.
Suppose both parties agree that there is no epistemological mystery concerning the
right-hand side rp. But since the left-hand side 1f! refers to abstract directions, 1/1 gives
rise to an epistemological challenge. According to Frege and his followers, the solution
is to observe that <p suffices for 1/1. Therefore, it is possible for us to advance from
knowledge of <p to knowledge of 1/1.
For this response to succeed, the operator::} must have some capacity for transmit-
ting knowledge from left to right. Ideally, we would like the following. Assume <p ::} 1/1
and that <p is known. Then 1/1 too must be known, or at least knowable. We can ensure
this transmission, it turns out, by laying down the following constraint:
Epistemic constraint
If <p::} if!, then it is possible to know <p-+ 1/1; and if additionally <p is known, then
this possible knowledge is compatible with continued knowledge of <p.
To confirm that the constraint would have the desired effect, let the operator JC
represent the epistemic status of being known. It is reasonable to take JC to be at
least potentially closed under modus ponens, in the sense that JC(<p-+ if!) and JC<p
entail OJCif!. Now, suppose that <p::} if! and JC<p. Then the constraint ensures O(JC<p /\
JC( <p -+ 1/1) ). Hence, the mentioned closure property ensures 00 JC if!. 26 Using Axiom
4 of modal logic, this entails 0JC1/1, as desired.
How should we understand the modal operator that figures in the constraint? It
is important that the possibility in question not be wildly idealized. We should not
ascribe to the epistemic agent an ability to carry out supertasks, for example. The
possibility needs to be of the same sort as the possibilities that are realized in cases
of actual mathematical knowledge.
As formulated above, the constraint would give us what we ideally want. 27 But
needless to say, we can't always get what we want! Depending on how the sufficiency
operator is interpreted, the above formulation may be asking too much. So it is reas-
suring to observe that our explanatory project allows the constraint to be weakened. It
is enough if the operator ::} is capable of transmitting knowledge from left to right to
some sufficient extent. To weaken the constraint in this way, we would obviously have
to clarify what counts as a "sufficient extent''. I shall not attempt to do so here. Indeed,
the basic sufficiency statements that I defend satisfy the constraint in its present strong
formulation. 28
A second promised benefit of thin objects is a response to the worry about the
seeming ontological extravagance of modern mathematics and certain other bodies
of knowledge, such as classical mereology. How can these sciences get away with
postulating such an abundance of objects when ontological economy is otherwise
26 Here we use the fact that in any normal modal logic, the environment 'O . . .'is closed under logical
entailment.
27
Or at least something close. We might also want the epistemic constraint to require (possible)
preservation of a priori knowledge or justification. But there is no need to commit to that here.
28
For example, the parallelism (or orthogonality) of two lines suffices for the identity (or orthogonality)
of their directions; see Chapters 8 and 9 for details. Using sufficiency statements such as these, I also
define some less basic forms of sufficiency. First, there are de re sufficiency statements, where the epistemic
constraint is satisfied only modulo choice of co-referring names. Second, the immediate sufficiency
discussed so far is supplemented with a notion of mediate sufficiency, defined as a form of transitive closure
of immediate sufficiency. I do not clam that this mediate notion of sufficiency satisfies the constraint in full
generality.
TWO METAPHYSICAL "PICTURES" 17
regarded as a virtue? Again, the minimalist has an answer, namely that the generous
ontologies in question either make no substantial demand on the world (in the case of
pure abstract objects such as numbers and sets}, or their demands on the world do not
substantially exceed demands that have already been met (in the case of impure sets
or mereological sums). This answer motivates another constraint on=>. Assume that
 1/1. Then any metaphysical explanation of :
Explanatory constraint
If 1/1, then <p -+ 1/1 admits of an acceptable metaphysical explanation.
To see that this constraint has the desired effect, let c 1/1 and C: must be less demanding than analytic implication but more demanding than
strict implication. Moreover, we motivated some philosophical constraints that any
viable notion of sufficiency must satisfy. Is there a candidate that satisfies the resulting
job description?
1. 7 Two Metaphysical "Pictures"

Attempts to satisfy our job description can be divided into two broad families, which
correspond to two different metaphysical "pictures''. I shall now describe the two
pictures and explain why I find one of them more promising than the other.
The first picture takes a highly symmetric view of abstraction. Consider the Fregean
example:
Although one sentence is concerned with directions and the other with lines, the two
sentences are nevertheless said to be made true by precisely the same facts or states of
the world. As Frege put it, the two sentences are merely different ways to "carve up"
a single shared content. Since we are using the operator ¢> to represent sameness of
demands on the world, a mutual sufficiency statement 1/1 ensures that <p and 1/1
are on a par in every respect that has to do solely with how the world is; that is
for every operator CJ that is concerned solely with how the world is. 29 In particular, the
two sides of every legitimate abstraction principle are on a par in every such respect.
29 Of course, a proponent of the symmetric picture can also introduce a non-symmetric operator, where
'rp =} 1/1' is interpreted as 'the demands that rp imposes on the world include those that 1/1 imposes'.
The competing asymmetric picture denies that the two sides of a legitimate
abstraction principle are on a par in every worldly respect. Instead, abstraction is
regarded as an inherently asymmetric matter, where abstraction on "old" entities gives
rise to "new" objects. While the left-hand side of, say, (Dir) demands of the world that
it contain directions, the right-hand side does not. Abstraction therefore involves a
worldly asymmetry. Clearly, this means that the mutual sufficiency statement (Dir¢>)
is inappropriate. Proponents of the asymmetric picture nevertheless insist that the
right-hand side of a legitimate abstraction principle suffices for the left-hand side, for
example:
(Dir=>)
How should this sufficiency statement be understood? It cannot straightforwardly

be understood in terms of demands on the world. For according to the asymmetric
picture, the demands of'd(l1) = d(/2)' exceed those of'/1 II 12'. Of course, a proponent
of thin objects will deny that the demands of the former statement substantially
exceed those of the latter-but this is itself in need of explanation. I find it useful to
understand the desired notion of sufficiency as a species of metaphysical grounding. 30
That is, /1 1112 metaphysically explains d(/1) = d(/2); the directions are identical in
virtue of the lines' being parallel. However, these remarks can at most serve to locate
the desired notion of sufficiency in a broader philosophical landscape, not to define
it. I would resist any identification of the notion of sufficiency with that of grounding,
for two reasons. First, the notion of grounding is highly schematic. It is one thing
to say that <P metaphysically explains 1/1 and quite another to spell out the promised
explanation. 31 My aim-in the cases that concern us-is the latter thing: I want to
understand how 11 II /2, for example, metaphysically explains d(li) = d(/2). This
cannot be done by means of a completely general study of metaphysical grounding but
will require a detailed examination of how abstraction works. Second, it is far from
clear that the general notion of grounding, as understood in the mentioned literature,
satisfies any worthwhile version of the epistemic constraint that was formulated in
Section 1.6. 32
Which of the two metaphysical pictures is correct? I find the asymmetric picture far
more natural and promising. One reason for this view has already been adumbrated
30 Grounding is currently attracting much attention in metaphysics; see e.g. (Fine, 20 I 2a), (Rosen,
2010), and (Schaffer, 2009). In fact, (Rosen, 201 l)'s "qualified realism" about mathematical objects has much
in common with my own conception of such objects as "thin''. Very roughly, where I develop a particular
conception of sufficiency, Rosen proposes that we use the general notion of grounding (but does not develop
this proposal in detail). See also (Rosen, 2016).
31 See (Wilson, 2014) for a related complaint and (Schaffer, 2016) for useful discussion.
32
We might add, as a third reason, that grounding, unlike my notion of sufficiency, is factive: if cp
grounds 1/f, then cp and 1/f. But it is unclear how deep this difference is. A nonfactive notion of grounding
is explored in (Litland, 2017); and if desired, my notion of sufficiency could easily be tweaked to ensure
£activity.
TWO METAPHYSICAL "PICTURES" 19
and will be developed further in Chapter 4. There appears to be a worldly asymmetry

between the two sides of legitimate abstraction principles, for example as concerns
their ontological commitments. Another reason has to do with de re statements of
sufficiency or mutual sufficiency. As already observed, such statements are needed; for
example, in order to express that there are objects whose existence is undemanding,
we need the de re formula 3x(T =>Ex), where Ex is an existence predicate and T is
some tautology, or an analoguous formula involving the mutual sufficiency operator.
Which such de re statements should we accept? And how can we account for the ones
that we do accept? I contend that the asymmetric picture is better equipped to answer
these questions.
To defend this contention, consider any legitimate abstraction principle. Consider
first the symmetric theorist's associated mutual sufficiency statement:
§a = §{3 {:> a ,....., f3
Are we allowed to "export" the function terms ' §a' and ' §{J' so as to derive the
following statement?
3x3y(x = §a /\ y = §{3 /\ (x = y {:> a ,....., fJ))

I believe we should be wary of accepting the resulting statements-whether obtained
by exportation or in some other way. For example, while the parallelism of two
lines suffices for the identity of the corresponding directions, the converse sufficiency
statement is problematic. Why should the self-identity of a certain direction suffice
for the parallelism of two particular lines? While abstraction on any suitably oriented
line yields the relevant directions, there is no way to "retrieve" any particular line from
this direction. Abstraction is a one-way road. What distinguishes one line from any of
its parallels is irretrievably lost in the abstraction that takes us from a line to its direction.
Compare now the asymmetric theorist's sufficiency statement associated with our
abstraction principle, namely:
(AP=}) a ,....., f3 => §a = §{3
Suppose we apply exportation to derive:

3x3y(x = §a /\ y = §{3 /\ (a ,....., f3 => x = y))
In this case, the resulting statements are independently plausible. For example, the
parallelism of two lines suffices for the self-identity (and thus also, as we shall see,
for the existence) of their direction. And in this case, there is no commitment to the
problematic converse sufficiency statement.
In short, on the asymmetric picture, a class of plausible de re sufficiency statements
can be obtained by exportation from sufficiency statements already accepted. By
contrast, on the symmetric picture, the de re statements that would result from
exportation are problematic. This raises tricky questions about which de re statements
of sufficiency or mutual sufficiency are acceptable on this picture and how this class
of acceptable statements should be accounted for.
The next order of business is to introduce my own version of the asymmetric
picture. I do this in Chapter 2, while Chapter 3 shows how the asymmetric abstraction
steps can be iterated. My argument against the symmetric picture is continued in
Chapter4.
2
Thin Objects via Criteria of Identity
2.1 My Strategy in a Nutshell

My strategy for developing a viable form of metaontological minimalism belongs to
the same broad family as the abstractionist approach discussed in Chapter 1. The
main source of inspiration is therefore Frege. But my aim is not primarily exegetical.
I wish to explore and develop some Fregean ideas that may help us make sense of
the idea of thin objects, regardless of whether Frege himself would have agreed with
this development. My argument is structured around the Fregean triangle, which
consists of three concepts central to philosophical logic and metaphysics-object,
reference, and criterion of identity. Frege believed these concepts to be connected in
deep and interesting ways. I develop explanations whose directions are represented
by the arrows in the following diagram:
reference
objecthood - - - - - - - - identity criteria
What is an object? There is a certain audacity to the question. Can we possibly

say anything general and informative about such a fundamental concept? As Frege
observes, the concept is "too simple to admit oflogical analysis" (Frege, 1891, p. 140),
and "what is logically simple cannot have a proper definition" (Frege, 1892, p. 182).
Frege is no doubt right that a "proper definition" of the concept of an object is out
of the question. Nevertheless, I believe an explication of the concept is possible.
Even when a concept cannot be de.fined in more basic terms, it can still be glossed
or characterized, for instance by relating it to other concepts and by explaining its
role in our thought and reasoning. There can be no doubt that an explication of the
concept of an object has the potential to be philosophically valuable. In the absence of
some general account, our thinking about objects will be in danger of being distorted
by intuitions formed in response to paradigm examples of objects, such as ordinary
physical bodies. 1 This may dispose us to think of all objects as thick, that is, as making
1
Compare (Hale, 2013, ch. !).
22 THIN OBJECTS VIA CRITERIA OF IDENTITY
a substantial demand on the world. The explication of the concept of an object to be

developed below pays particular attention to the role that objects play in our semantic
theories.
This leads us to the concept of reference. Singular terms and corresponding mental
representations refer to objects with which they have no intrinsic connection. For
instance, the inscription 'Socrates' refers to an ancient Greek philosopher, although
neither the inscription nor the philosopher depends on or in any way resembles
the other. Likewise, the current configuration of neurons in my brain allows me to
entertain thoughts about the laptop computer directly in front of me. What endows
ink marks on paper and configurations of neurons in brains with this power to refer to
objects with which they have no intrinsic connection? That is, what does the relation of
reference consist in? The problem is particularly serious in the case of abstract objects.
There certainly appears to be such a thing as reference to abstract objects. Even a casual
examination of natural language reveals lots of apparent examples. But it is not at all
clear what such reference might consist in. It certainly cannot be based on any form
of causal interaction.
The final concept in our triangle is that of a criterion of identity, by which I
mean a principle which in a systematic and informative way relates the identity or
distinctness of a certain class of objects to certain other facts. For example, two sets
are identical just in case they have the same elements. This is often taken to provide
metaphysical information about what the identity or distinctness of the relevant
objects "consists in''. 2 Criteria of identity are sometimes thought to do epistemological
work as well, namely by underlying and guiding people's judgments about the identity
and distinctness of objects. Why do I judge that the computer I am now seeing is the
same as the one I saw a minute ago? And why do I identify the object that my fingers
are currently touching with the object that I am currently seeing? A natural answer
is that I am relying-perhaps unconsciously-on a criterion of identity for physical
bodies that informs me that two parcels of matter belong to the same body just in case
they are spatiotemporally connected in some suitable way.
It is far from obvious that the three fundamental concepts that make up the Fregean
triangle should be connected in any deep or interesting ways. But according to Frege,
they are. He connects the first two concepts by characterizing an object as a possible
referent of a singular term. It is important to notice that the singular term in question
need not be available in our current language. The claim is rather that an object is the
sort of entity that could be referred to by a singular term in some language or other.
Next, Frege connects the last two concepts by suggesting that singular reference can be
explained in terms of the concept of a criterion of identity.
This book develops and defends versions of these Fregean links and thus shows
that the three concepts do indeed belong to a tightly integrated triangle. In this way
2
For discussion, see e.g. (Lowe, 2003) as well as Appendix 2.A below.
A FREGEAN CONCEPT OF OBJECT 23
I hope to shed light on the attractive idea of thin objects. My explanatory strategy is
easy to describe now that the stage has been set. By explaining reference in terms
of criteria of identity, we make reference comparatively easy to achieve, including
reference to abstract objects. For example, it suffices for a direction term to refer
that it has been associated with a line and is subject to a criterion of identity that
takes two lines to specify one and the same direction just in case the lines are parallel.
I show how this kind of sufficient condition for reference-based on a specification
of the would-be referent and a criterion of identity operating on such specifications-
can be generalized to cover a vast range of cases. Furthermore, because an object is
characterized as a possible referent of a singular term, easy reference ensures easy
being. The result is that the obtaining of an appropriate criterion of identity becomes
sufficient for the existence of an object governed by this criterion. For example, two
lines' being parallel is sufficient for the existence of a direction that is shared by the
two lines. A truth about parallelism is thus reconceptualized in a way that reveals a
new object, namely a direction, which was not involved in the original truth. 3
Finally, the new objects obtained by reconceptualization may enable us to formulate
criteria of identity for yet more objects. The formation of sets provides an example.
Given some domain, we can reconceptualize so as to "form'' sets of objects from this
domain. This makes available more objects than those with which we began, and
these objects can be used to specify yet further sets. In this way, our explanations can
proceed around the Fregean triangle one more time. This form of iterated abstraction
will be the focus of Chapter 3.
2.2 A Fregean Concept of Object

Let us take a closer look at the Fregean links on which my strategy is based, namely
between objecthood and reference, and between reference and criteria of identity.
To understand the first of these links, it is useful to ask what role objects play in
our semantic theories. 4 Like Frege, however, I am not concerned with the semantics
of natural languages, which tends to be complex and quirky. It is better to focus on
the semantics oflogically regimented languages, such as that of first- or higher-order
logic, which systematically and perspicuously represents a logical core of natural
languages. The concept of an object that we explore is therefore a logico-semantic
one. Thus refined, our question has a clear answer. Objects serve as referents ofsingular
terms and as values of bound first-order variables. Frege's characterization of an object
as a possible referent of a singular term focuses on the former role. Objects are the
3 This reconceptualization must not be conflated with the notion of recarving of content. Where
recarving is tied to the symmetric idea that one and the same worldly fact can be "carved up" as two different
contents, reconceptualization (as I use the term) is tied to an asymmetric conception of abstraction; in
particular, ontological commitments can increase as a result of reconceptualization.
4 Cf. (Dummett, 198la, ch. 14).
kind of entity to which singular terms are suited to refer. 5 By contrast, Quine focuses
on the latter role of objects as values of bound variables, as encapsulated in his famous
slogan that "to be is to be the value of a variable''. 6
It may be objected that both the Fregean and the Quinean explication make the
concept of an object excessively semantic. Even if there were no singular terms or
variables, there would still have been objects. How, then, can the concept of an
object be characterized in terms of the semantics of singular terms and variables?
But the objection is based on a fairly elementary misunderstanding. It is true that
both explications exploit the contingent fact that we have a language-and thus also
singular terms and variables, or at least their natural language equivalents-to expli-
cate one of our fundamental logico-metaphysical concepts. But nothing prevents the
concept thus explicated from being applied to other possible circumstances, including
ones where there is no language. As always, when evaluating counterfactuals, we
keep the interpretation of our own language fixed in order to pronounce on how
things would have been in alternative circumstances. Even if these circumstances
involve there being no language, it is still permissible for us to rely on our language-
contingent though it may be-to pronounce on how things would have been under
these circumstances. In particular, we can ask what entities of the sort to which
singular terms are suited to refer can be found in these circumstances. So it is simply
incorrect that the proposed semantic characterization of the concept of an object
makes all objects language dependent.
A more interesting question concerns the relation between the Fregean and
Quinean explications of the concept of an object. Since the referent of a singular
term can always serve as the value of a bound variable, it follows that an object in
Frege's sense is also an object in Quine's sense. Somewhat surprisingly, however, the
converse conditional turns out to be controversial. A problem with the converse
arises in connection with mathematical structures with non-trivial automorphisms.7
Consider the imaginary units i and -i in complex analysis, which are swapped by the
automorphism of complex conjugation. 8 If there really are such things as sui generis
complex numbers-rather than just set-theoretic simulacra that exhibit the structure
of the complex field-then it is hard to see how a singular term could refer to one of
the imaginary units rather than the other. Nothing internal to the structure could
5 See also (Hale, 2013, ch. 1).

6 By 'variable' Quine here means bound first-order variable, as these are the only variables he thinks can
legitimately be bound. See (Quine, 1986). But the Quinean gloss on the concept of an object is independent
of this: we just need to add that the slogan is concerned with bound first-order variables.
7
An automorphism is a permutation of the objects of the structure which leaves the structure intact.
An automorphism is said to be non-trivial if it is distinct from the identity automorphism. Structures
with non-trivial automorphisms have received much attention in the debate about non-eliminative or
ante rem mathematical structuralism, where the phenomenon is used as the basis for an objection to the
mentioned form of structuralism; cf. (Burgess, 1999) and (Keranen, 2001). My present concern is not with
this objection but with the relation between the two mentioned explications of objecthood.
8
Complex conjugation maps the complex number a + bi to a - bi.
A FREGEAN CONCEPT OF OBJECT 25
ensure that the term refers to one rather than the other. For the existence of non-
trivial automorphisms entails that the two units are structurally indiscernible, in the
sense that every structural property that holds of one also holds of the other. And it is
hard to see how facts external to the structure-whether mathematical or physical-
could be responsible for a reference relation of the sort in question. 9 The upshot is
that the imaginary units appear not to qualify as objects in the Fregean sense: each
unit appears not to be a possible referent of a singular term. 10 By contrast, it seems
possible to quantify over the complex numbers, including the imaginary units, which
would thus qualify as objects in the Quinean sense. This suggests that Quine's concept
of object-if coherent-is strictly more general than Frege's. Quineans will no doubt
regard this as an advantage of their explication.
I see two lines of response to this argument. An uncompromising response would
be to deny that proper sense can be made of quantification over objects to which sin-
gular reference is impossible. The standard truth-condition for a quantified formula
is based on the successive assignment of each object in the domain to the variable
bound by the quantifier. Can we really make sense of such an assignment where the
domain contains objects to which singular reference is impossible? Notice, moreover,
that a negative answer to this question would have no direct effect on the practice of
mathematics. The practice of complex analysis, for instance, does not require that the
complex field be regarded as a sui generis mathematical structure, as was presupposed
in the argument under discussion. It suffices, for all practical purposes, to define the
complex numbers as pairs of reals with operations of addition and multiplication, as
in fact is the official definition in most textbooks. On this approach, we rule out all
non-trivial automorphisms. Indeed, this observation generalizes. For any purported
structure with non-trivial automorphisms, we can find an isomorphic copy based on
pure sets, which permit no such automorphisms. 11
A far more ecumenical response is also possible. Recall that my intention is
to use the Fregean concept of an object as part of an explication and defense of
metaontological minimalism. All that this argumentative strategy requires is that
being a possible referent of a singular term suffices for objecthood. This establishes
one of the links associated with the Fregean triangle. There is no need to insist on the
necessity of this Fregean criterion. Philosophers who prefer the Quinean conception
are invited to develop their own account of lightweight objects. If they succeed, their
account would supplement mine, not threaten it.
I prefer this second response because of its smaller commitments. Thus, although
my arguments in this book rely on the Fregean concept of an object as the kind
of entity to which a singular term is suited to refer, I happily leave the door open
9 Indeed, non-eliminative or ante rem structuralists, who are among the foremost defenders of
sui generis complex numbers, deny that the units have any mathematical properties other than their
structural ones.
10 See (Brandom, 1996) for relevant discussion.
11
See (Parsons, 2004, Section IV) for a discussion of this response.
to developments of the Quinean concept of an object that might supplement my

own account.
2.3 Reference to Physical Bodies

I turn now to the second link, which connects the concepts of reference and criteria
of identity. The idea, we recall, is that criteria of identity provide an easy route to
reference of singular terms. Combined with the link already established, this will
ensure easy existence of objects. How should the second link be developed? Since
the Fregean ideas we are discussing are very abstract and theoretical, I begin with an
example. In later sections, I describe the form of reconceptualization involved in the
second link and explain how our example can be generalized.
The example concerns the simplest and most direct form of reference to ordinary
physical bodies, such as sticks and stones, and tables and chairs. (A more detailed
characterization of the concept of a physical body will be provided shortly.) Let me
state straight away that my aim is a toy model of such reference, not an account that
captures every aspect of this immensely complex problem. What we need is a model
that explains how a criterion of identity can play a role in the constitution of such
reference.
Since our aim is merely a toy model of reference to physical bodies, it is useful to
carry out our investigation in terms of robots that are embedded in, and interacting
with, a physical environment. This allows us simply to stipulate how the robots
function, rather than to engage in speculation about the intricacies of actual human
psychology. 12 We can simplify our task further by focusing on the senses of sight
and touch, and on some very fundamental cognitive processes. Other senses, such as
smell, taste, and hearing, appear to play a less fundamental role in reference to physical
bodies and are therefore set aside in our toy model. Consciousness too-in the sense
of subjective awareness of what it is like to have various sorts of experiences-is put to
one side, as it appears inessential to the core notion of reference. So consider a robot
equipped with senses of sight and touch. Such a robot interacts with its environment
by detecting light reflected by surrounding surfaces and by having a capacity for
touching and grasping things in its vicinity.
Our question is what it takes for such a robot to refer to physical bodies in its
environment. At the very least, the robot must "perceive" the body, in the very
undemanding sense that it receives light from part of its surface or touches some part
of it. The robot thus receives information from some part of the body. These parts
need not have natural boundaries in either space or time; they are simply the sum-
totals of the physical stuff with which the robot causally interacts in this rudimentary
perception-like way. While these simple causal interactions are certainly necessary for
12
My account is inspired, however, by the account of infants' reference to physical bodies defended by
some developmental psychologists; see note 14.
REFERENCE TO PHYSICAL BODIES 27
the robot to refer to bodies, they are not sufficient. The rudimentary perception-like
interactions with parcels of matter enable the robot to develop an internal map of
the distribution of matter in its environment and cif the purely qualitative properties
instantiated by various parcels of matter. But such interactions alone do not enable the
robot to engage with any questions involving identity, such as to determine whether
two parcels of matter belong to one and the same physical body or to count the number
of bodies in its environment. For the robot to be engaged with questions such as these,
it needs to possess some more sophisticated capacities, involving the representation
of space, time, continuous trajectory, and physical cohesiveness. Only then will the
robot be capable of determining whether two parcels of matter belong to one and the
same body.
The characterization of the robot's additional capacity requires some care. We are
surrounded by bodies that are wholly or partially hidden, and that move in and out
of view. A stick can be partially buried, and a stone can be mostly covered by other
stones piled up around it. There are always many different ways of "getting at" one
and the same physical body, both from different spatial points of view and at different
moments of time. All these situations have to be handled by the robot's capacity
for grouping together parcels of matter as belonging to one and the same body.
What ultimately matters is that the parcels of matter with which the robot interacts
through its rudimentary forms of perception are spatiotemporally connected in some
appropriate way. Assume for instance that the robot establishes visual contact with
part of a stick that emerges from the ground and that one of its "arms" simultaneously
probes into the ground nearby and encounters something hard. What should we
"teach" the robot about the conditions under which the two parcels of matter with
which it interacts belong to the same body? Roughly, the kind of connectedness that
matters has to do with solidity and motion. The two parcels of matter must be related
through a continuous stretch of solid stuff, 13 all of which belongs to the same unit
of independent motion-roughly in the sense that, if you wiggle or pull one of the
parcels, the other one follows along or at least has some disposition to do so.
To produce a more precise answer, let us continue our investigation as an exercise
in robotics. I submit that the following fundamental principles are part of an analysis
of the concept of a physical body, and therefore have to be implemented in the robot: 14
(Bl) Bodies are three-dimensional solid objects.
13 I mean "solid" in the ordinary, loose sense in which a stick or a stone is said to be solid. Of course,
physics tells us that even sticks and stones aren't solid in the stricter sense of filling up all space at an
atomic level.
14
These principles are also constitutive of the concept of what psychologists sometimes call "Spelke-
objects''. This concept corresponds closely to the concept of physical body that I am employing here. See
e.g. (Spelke, 1993) and (Xu, 1997). See also (Burge, 2010, pp. 437-71) for a useful discussion and further
references.
Thus, a cloud of gas doesn't qualify as a body in the present sense. This means that not
all spatiotemporal objects are bodies.
(B2) Bodies have natural and relatively well-distinguished spatial boundaries.
For instance, the (undetached) lower half of a rock fails to be a body because it
lacks sufficiently natural boundaries, and a mountain fails because its boundaries are
insufficiently well distinguished. 15
(B3) Bodies are units of independent motion.
Thus, although a book is a body, a pile of papers is not. Bodies need to be cohesive
enough to be disposed to move as a unit.
(B4) Bodies move along continuous paths.
Consider the object that came into being with the birth of Bill Clinton, coincided
with Clinton until the end of his presidency, and thenceforth coincides with George
W Bush. By (B4), this object cannot be a body. Bodies move along continuous paths.
(BS) Bodies have natural and relatively well-distinguished temporal boundaries.
So arbitrary temporal parts of bodies are not themselves bodies.

I believe this relatively simple model captures the core of the most direct, perceptual
way of referring to physical bodies. What matters is that our agents-whether humans
or robots-receive sensory information from parcels of matter and have a capacity
for grouping together such pieces of information just in case these pieces derive
from parcels that are spatiotemporally connected in the way just outlined. Let ,....,
be the relation that obtains between two parcels of matter just in case they are
spatiotemporally connected parts of a cohesive and reasonably well-delimited whole.
This relation is clearly symmetric. For if two parcels of matter u and v specify the
same body, then so do v and u. A similar defense can be given of transitivity. We
should not require that the relation be reflexive, however, since not every parcel
of matter belongs to a coherent and well-delimited whole. For instance, if I point
to some partially solidified mud in a field, I will probably fail to specify a unique
physical body.
Were it not for the possibility of failures of reflexivity, it would have been straight-
forward to formulate a criterion of identity for bodies, namely
(CI-B) b(u) = b(v) *+ u rv v
15 Precisely how well distinguished must the boundaries of a body be? Presumably, a shedding cat is still
a physical body despite all the hairs that are in the process of falling off it. Although I doubt that our question
admits a precise answer, I am hopeful that an approximate answer can be given by empirical investigation
of ordinary people's concept of a body.
REFERENCE TO PHYSICAL BODIES 29
where b is a "body building" function that maps a parcel of matter to the body to
which it belongs.
How should we accommodate the possible reflexivity failures, however? The ques-
tion requires a brief logical digression, which this paragraph will provide. A relation
that is symmetric and transitive but not necessarily reflexive is known as a partial
equivalence relation. Such a relation does to the objects in its field 16 what an equiva-
lence relation does to the entire domain: it partitions the relevant objects into disjoint
and mutually exhaustive equivalence classes. In this way, the partial equivalence
relation "" on parcels of matter determines a partial function b that maps a parcel
u to the physical body, if any, of which u is a part. This suggests a way to tweak
(CI-B). We can accommodate reflexivity failures by letting the "body builder" b be a
partial function and by letting (CI-B) operate in the context of a negative free logic. 17
This allows b(u) = b(u) to be false for any argument u for which the function bis
undefined-which comes to the same thing as an argument u on which "" fails to
be reflexive.
With this understanding of (CI-B) in place, let us return to our main discussion.
I claim that it suffices for our robot to refer to a body that the robot is appropriately
related to some parcel of matter and that it treats two such parcels as specifications
of the same body just in case they are related by the appropriate partial equivalence
relation. There is no more direct way for it to 'get at" a physical body. The most
direct form of reference to bodies is constituted on the basis of a specification and
a partial equivalence relation that provides a criterion of identity. Thus, our toy model
of reference to physical bodies nicely illustrates my general claim that there is a
tight connection between reference and criteria of identity. By operating with an
appropriate criterion of identity, the robot comes to refer to physical bodies in its
environment.
It may be objected that this toy model does not even approximate what is going
on in us humans because bodies are directly given to us in perception in a way that
bypasses the need for any criterion of identity. As a phenomenological point, this is
no doubt correct. In our ordinary perception of bodies, there is no need to actively
and consciously apply a criterion of identity. This does not contradict my claims,
however. A lot of subpersonal processing takes place between the stimulation of our
sense organs and what is phenomenologically given to us in perception. What the toy
model suggests is that these subpersonal processes must involve a criterion of identity
in order for the resulting representation to refer to a physical body. I do not claim that
this process is accessible to our consciousness or ever rises above the subpersonal level. On
the contrary, the lack of conscious or explicit access to the process is something that
16
Recall that x is said to be in the field of a relation R just in case x bears R to, or is borne to R by, some
object or other.
17 This logic is explained in Appendix 2.C, which also provides an elaboration of this and some other
approaches to abstraction on partial equivalence relations.

I welcome. 18 This view represents a major departure from most earlier philosophical
work on criteria of identity. 19
2.4 Reconceptualization
My account of the constitution of reference to physical bodies has a reductionist
character that may be surprising. I set out to explain how things have to be for a robot
to refer to a physical body. But my explanation mentions only parcels of matter with
which the robot causally interacts and a partial equivalence relation that the robot uses
to determine when two such parcels should be regarded as parts of one and the same
body. There is no mention of the body that is the referent! Although bodies figure in
our explanandum-namely, how reference to physical bodies is constituted-bodies
are strangely absent from the proposed explanans. It is as if we were offered an account
of the constitution of marriage which leaves out one of the spouses!
I believe the key to understanding the surprising reductionist character of our
example-and of a vast family of generalizations described in the next section-is the
idea of reconceptualization. The material made available by the explanans is recon-
ceptualized in a way that brings out the existence of bodies. This reconceptualization
is possible because there is not a unique right way to apply the Fregeau concept of an
object to reality. We can always reconceptualize by applying this concept in a new and
different way. Of course, these claims are highly programmatic and need to be spelled
out and defended. The present section makes a start by locating the approach to be
developed in the book as a whole within a broader philosophical landscape.
Our question is how the Fregeau concept of an object is applied to reality. Suppose
we want to introduce new singular terms referring to a range of objects that we have
not yet recognized. What does it take to succeed? A minimal requirement is that truth-
conditions have been assigned to all identity statements and other predications that
involve the new singular terms to be introduced. Moreover, this assignment must be
done in a way that respects the laws of logic. The most important part of logic in
this connection is what we may call the logic of identity, which describes the identity
relation and its interaction with predication. In brief: identity is an equivalence
relation that interacts with predication as described by Leibniz's Law, namely, if x is
identical with y, then x and y have precisely the same properties:
(LL) x = y -+ (<p(x) B- <p(y))
As usual, Leibniz's Law must be restricted to contexts <p( . . .) that are transparent, in
the sense that the context is only concerned with the referents of the terms that fill
18
Indeed, this inaccessibility plays a key role in an argument I develop in Section 8.4.3.
19 See e.g. (Dummett, 198 la, pp. 73-6, 179-80, 545- 6) and (Wiggins, 2001). However, some psycholo-
gists make the same departure; see (Xu, 1997).
RECONCEPTUALIZATION 31
the argument place, not with these terms themselves or any non-referential aspects
of their meaning.
Suppose we have assigned truth-conditions to all identity statements and predi-
cations that involve some new singular terms we are trying to introduce, and that
the assignments are compatible with the logic of identity-or in other words, that we
have satisfied the minimal requirement stated above. In short, suppose an attempted
application of the conceptual apparatus for identification and predication is logically
coherent. Does this attempted application of our conceptual apparatus succeed in
latching on to objects? Two different answers flow from two rival conceptions of
ontology, which I shall now canvass with a deliberately broad brush. I later clarify
and defend some aspects of one of the conceptions.
The rigid conception holds that reality is "carved up" into objects in a unique way
that is independent of the concepts that we bring to bear. This conception introduces
an element of risk into our proposed application of our conceptual apparatus for
identification and predication. Although the attempted application is logically in good
order, reality may fail to cooperate. Reality may simply not contain the sorts of objects
we are trying to "carve out':
The flexible conception, on the other hand, insists that reality is articulated into
objects only through the concepts that we bring to bear. And we often have some
choice in this matter. 20 Of course, our choice is limited. It is not an option for a starving
person to produce more bread by "carving up" reality differently. The choice we have
is of a different character. As Frege observed, reality provides different answers to
ontological questions about what there is depending on which concepts we bring to
bear. 21 If you bring to bear the concept of a loaf of bread, some part of reality (say,
your kitchen) may answer that there are two such objects. If instead you bring to bear
the concept of a molecule of organic material, you will receive a very different answer.
This much should be uncontroversial. The flexible conception of ontology gets its bite
by adding the controversial claim that there is no unique, privileged set of concepts
in terms of which to "carve up" reality, namely the concepts that match some rigid
concept-independent articulation ofreality into objects. This means there is no risk
that reality fails to contain objects answering to some coherent application of our
apparatus for identification and predication. This coherent application "carves out"
the appropriate objects.
The disagreement between the two conceptions hinges on the relative explanatory
priority of the articulation of reality into objects and permissible applications of
our conceptual apparatus for identification and predication. The rigid conception
20 This distinction has much in common with Hilary Putnam's distinction between "metaphysical" and
"internal" realism. See (Putnam, 1987) for a brief introduction. There are important differences, however,
especially as concerns the emphasis that I place on the connection between this ontological debate and the
analysis of the concept of an object.
21 See (Frege, 1953, §46).
contends that first reality provides a range of objects, to which we then apply our
conceptual apparatus. On this view, there is only one correct way to apply language
to reality, namely to ensure that each singular term is assigned one of the objects
that reality provides, and to regard an identity statement as true just in case its
two singular terms have been assigned one and the same object. The view operates
with a Lego block conception of reality on which the world consists of well-defined
objects-"Lego blocks"-that are available independently of any application of our
conceptual apparatus for identification and predication. 22 As a result, the application
of this apparatus to reality is an inherently risky undertaking even when the logical
coherence of the application is beyond doubt. We need to ensure that each singular
term is assigned a unique "Lego block" or sum thereof, and each predicate, a property
(or relation) that may hold of (or obtain between) such blocks or sums.
The flexible conception of ontology rejects the idea of a domain of objects-or
"Lego blocks" -that are available independently of any application of our conceptual
apparatus for identification and predication. It makes no sense to talk about objects
until our conceptual apparatus for identification and predication is already in play.
So we can make no sense of the rigid conception's self-individuating "Lego blocks",
which an attempted application of our apparatus for identification and predication
must latch on to in order to succeed. On the flexible conception, we need an entirely
different explanation of what counts as a permissible application of our apparatus
for identification and predication. This explanation must not presuppose the range
of objects to which this apparatus is applied.
What might the required explanation look like? A radical option is to take the
minimal requirement just mentioned to suffice for successful reference. That is,
provided that truth-conditions are assigned in a way that accords with the laws of
logic and that appropriate sentences come out true, then reference has been achieved.
In particular, there is no need to appeal to criteria of identity. Versions of this radical
option have been defended by the neo-Fregeans Hale and Wright, as well as by Agustin
Rayo. These versions are criticized in Chapter 5 for going too far.
This book defends a less radical version of the flexible conception, which does
attach special significance to criteria of identity. Where we have a criterion of identity,
we can make sense of approaching an object in multiple different ways, of tracking it,
and reidentifying it. These capacities mark a watershed in our individuation of objects.
Physical bodies provide an example. I proposed an account of how reference to bodies
is constituted on the basis of a criterion of identity. As remarked, this account does
not presuppose physical bodies, only the parcels of matter of which the bodies are
composed and the partial equivalence relation discussed above.
Our discourse about books provides another good illustration. Suppose we start
with a practice of identifying and distinguishing physical copies of books-or
22
Thanks to Agustin Rayo for suggesting this image.
REFERENCE BY ABSTRACTION 33
book tokens, as we may call them-and predicating properties of these. We can

meaningfully say, for example, that your token of Anna Karenina weighs more than
mine. But we also talk about books in a more abstract way. Suppose you and I have read
separate tokens of Anna Karenina. We still regard it as true to say that you and I have
read one and the same book. That is, we identify and ascribe properties not only to
book tokens but also to book types, such as the book type that you and I have both read.
Our practice of identifying book types and predicating properties of them is highly
systematic, fully in compliance with the logic of identity, and explicable in a way that
makes recourse only to book tokens and their properties. This illustrates the versatility
and flexibility of our conceptual apparatus for identification and predication.
To avert misunderstanding, let me stress, yet again, that I am not claiming that
criteria of identity are necessary for reference. Since my explanations move counter-
clockwise around the Fregean triangle, I need an appropriate use of criteria of identity
to be sufficient for reference. I do not need the corresponding necessity claim. 23 As far
as this book is concerned, there might be other ways to supplement the minimalist
answer so as to explain how objects come to figure as referents of singular terms or
values of first-order variables. 24
The purpose of this section has (as announced) merely been to locate the account
that I develop in a broader philosophical context. Much work obviously remains.
In particular, I need to explain the central notion of a permissible application
of our apparatus for identification and predication. My next step is to generalize
and elaborate on the form of reconceptualization involved in the example of
book types.
2.5 Reference by Abstraction

Can my account of reference to physical bodies be extended to other kinds of objects,
perhaps even to abstract objects? There is reason to be hopeful. My account is based
on a form of reconceptualization brought about by an appropriate use of a criterion
of identity. And since the notion of a criterion of identity is extremely general, it is
applicable to abstract objects as well as to concrete ones. So if an appropriate use of
criteria of identity can suffice for reference, this holds out the promise of an extremely
general sufficient condition for reference. This approach would liberate the notion of
reference from the requirement of any sort of causal connection between the term and
23
Other thinkers inspired by Frege, however, have sometimes been attracted to some version of the
necessity claim as well. Perhaps all reference is ultimately based on certain fundamental forms of reference
in which criteria of identity play an essential role. For discussion, see (Dummett, 198la, pp. 231-9), and
(Evans, 1982, pp. 109-12).
24
If such alternatives exist, should the objects in question be understood in accordance with the rigid
or the flexible conception of ontology? That is, would these objects too be "carved out" by our concepts, but
in a way that differs from the one developed in this book? For present purposes, I need not answer these
questions-although I confess to a general sympathy with the flexible conception.
the referent-which would be a good thing. For as we saw in Section 1.4, there appears
to be such a thing as reference to abstract objects. This phenomenon of apparent
reference to abstract objects calls for a serious investigation, which our approach
makes possible, unlike approaches that require some form of causal interaction.
One of Frege's favorite examples of a criterion of identity concerns directions.
Suppose we begin with a domain of lines. Parallelism is an equivalence relation on
this domain. So we can adopt the following criterion of identity:
(Dir)
where 'd(l)' stands for the direction of the line /. That is, two lines specify the same
direction just in case they are parallel. On the plausible assumption that lines and
facts about parallelism are epistemically accessible to us, this criterion enables us
to talk about directions being identical or distinct. What about other properties
and relations? Orthogonality provides an example. Two directions are regarded as
orthogonal just in case they are specified by two orthogonal lines. Using ..L and ..L *
as orthogonality predicates for lines and directions, respectively, we formalize this as
follows:
The orthogonality of two directions is, as it were, "inherited" from the orthogonality
of any two lines in terms of which these directions are specified.
Of course, this "inheritance" of properties from lines to their directions presup-
poses that it does not matter which line we choose to specify some given direction.
Thankfully, the presupposition is met. Assume that the direction specified by 11 is
also specified by l~, because 11 II /~. If one of these two specifications is orthogonal
to another line /z, so is the other. For the purposes of assessing orthogonality, any
line is just as good as any of its parallels. Moreover, the example generalizes. For any
formula <p on lines that doesn't distinguish between parallel lines, we can introduce
an associated predicate <p* on directions by letting <p* hold of the directions of some
given lines just in case <p holds of the lines themselves.
What does this account establish? The account indisputably shows how we can
use the resources provided by a domain of lines and the relations of parallelism and
orthogonality of lines to define a way of speaking that is just like the realist's. It is
true to say that there are directions, related to one another in various ways. My goal
is more ambitious, however. I contend that, when properly developed, the account
establishes not merely an acceptable way of speaking but explains how reference to
abstract directions is constituted through a form of reconceptualization.
Before defending this stronger contention, I wish to comment on the scope of
the approach. There is a vast family of examples that are structurally similar to the
one concerning lines and directions. Any form of type-token distinction provides an
example. A syntactic type, for instance, can be specified by means of a syntactic token;
and two such tokens specify the same type just in case they count as equivalent by the
REFERENCE BY ABSTRACTION 35
relevant standards. 25 Geometrical shapes provide another example. A shape can be

specified by means of an object that has the shape in question; and two objects specify
the same shape just in case they are congruent. All of these examples share a common
structure.
To identify this structure, we must distinguish between two forms of criteria of
identity. A one-level criterion of identity says of two objects of some sort F that they
are identical just in case they stand in some relation R: 26
(CI-1) Fx /\ Py-+ (x = y ++ R(x,y))
Criteria of this form operate at just one level, in the sense that the condition for two
objects to be identical is given by a relation on these objects themselves. A good
example is the set-theoretic law of extensionality, which says that two sets are identical
just in case they have precisely the same elements:
(Ext) SET(x) /\ SET(y) -+ (x = y t t Vu(u E x t t u E y))

Clearly, one-level criteria of identity are not connected in any direct way with
abstraction principles.
A two-level criterion of identity relates the identity of objects of one sort to some
condition involving items of another sort. The former sort of objects are given as
the value of some function applied to items of the latter sort. So the criterion takes
the form
(CI-2) f(a) = f(f3) ++a ~ f3

where the variables a and f3 may be of either first or higher order and~ is an equiva-
lence relation on the entities over which the variables range. 27 This philosophical use
of two-level criteria suggests some terminology. Let us call the entities over which the
variables a and f3 range specifications. These entiti~s serve to specify the objects for
which the criterion ofidentity holds. 28 Next, let us refer to the equivalence relation~
as a unity relation.
From a logical point of view, a two-level criterion of identity is just the same
as an abstraction principle. Some famous examples derive from Frege. We have
already discussed the direction principle (Dir). Another important example is Hume's
Principle
25
Ordinary English may not have a good name for this equivalence relation. But the relation can be
grasped without the need for any antecedent knowledge of types. After all, this is how we come to master
discourse about types in the first place. See Chapter 8 for more discussion.
26 This conception of criteria of identity goes back at least to Locke, who ponders what various identities
"consist in"; see his Essay, Book II, Chapter XXVII. More recently this conception has been explicitly
endorsed in (Lowe, 1989) and (Lowe, 1997).
27 An approach based on two-level criteria of identity is found in (Williamson, 1990, ch. 9}.
28
I have previously used the word 'presentation' but have found that this word has phenomenological
and psychological connotations that are avoided by the word 'specification: which accordingly seems
more apt.
(HP) #F=#G~F~ G
where F and G are second-order variables and F ~ G abbreviates the second-order

formalization of the claim that the Fs and the Gs are equinumerous, so as to provide
a criterion of identity for cardinal numbers.
As in the case of directions, we can adopt "inheritance" principles that describe how
certain properties of specifications are "inherited" by the objects that are specified.
Consider a formula .. . , an) ~ <p(f31, . .. , f3n))
For any such formula, we introduce a corresponding predicate <p*, defined on all of the
objects that can be specified in this way and governed by the following "inheritance"
principle:
(Inher)
In this way, we obtain many useful predicates; for example, an orthogonality pred-
icate on directions and a successor predicate S that holds of two cardinal numbers
# F and #G just in case all of the Fs are equinumerous with all but one of the Gs.
The unity relation "" that figures in a two-level criterion is ordinarily required
to be an equivalence relation. But as our discussion of physical bodies revealed, we
have reason to relax this requirement slightly. We wish to allow our specification
variables to include in their range entities that fail to specify an object of the sort
with which we are concerned. Suppose, for example, that we wish to abstract on lines
and other directed items so as to obtain directions, while simultaneously including in
our universe of discourse objects that cannot be said to be parallel with themselves,
such as chairs, electrons, and numbers. We can achieve this by allowing "" to be a
partial equivalence relation, that is, a relation that is symmetric and transitive but not
necessarily reflexive. For instance, in the mentioned example we let ,.._, be defined by
parallelism on any directed items and be false whenever one or both of the arguments
isn't a directed item. When "" is a partial equivalence relation, the function f that
figures in (CI-2) must be allowed to be partial as well, defined on all and only items a
that are in the field of "". 29 As in the case of physical bodies, this generalized two-level
criterion can still be formalized as (CI-2), provided that this criterion operates in the
context of a negative free logic. 30 All in all, the proposed technical adjustments are
minor and do not materially affect the ensuing philosophical discussion.
The distinction between one- and two-level criteria of identity is important, given
our project. In a two-level criterion, the objects whose identity conditions we are ana-
lyzing do not figure on the right-hand side of the biconditional. This promises a way
29
Since ~ is a partial equivalence, this is equivalent to saying that f is defined on all and only items a
such that a ~ a. For suppose a ~ f3 for some f3 . Then symmetry and transitivity ensure a ~ a.
30
See Appendix 2.C for details.
SOME OBJECTIONS AND CHALLENGES 37
to characterize the identity conditions for potentially problematic objects that does
not presuppose such objects. A two-level criterion thus promises a way to single out
objects of one sort in terms of items of another- and presumably less problematic-
sort. Consider the case of physical bodies. Our toy model of reference to physical
bodies explains how such reference is constituted, all in terms that make no explicit
appeal to bodies. This is what gives the model its reductionist character. Abstract
objects provide further examples. How are directions and other abstract objects "given
to us"? Frege's answer is that we relate the identity of two directions to the parallelism
of the two lines in terms of which these directions are specified. Since reference to
lines is less puzzling than reference to directions, this is explanatory progress. By
contrast, a one-level criterion would be oflittle or no use when trying to explain our
most fundamental way of referring to objects of a certain kind. To apply a one-level
criterion, one must already be capable of referring to objects of the sort in question.
The stage is now set for an initial formulation of the link between reference and
criteria of identity that I defend:
Reference by abstraction
Consider a two-level criterion of identity {Cl-2) which purports to provide iden-
tity conditions for some kind F of object in terms that presuppose only some
antecedently accepted ontology but not Fs. Assume that an agent has an appropriate
grasp of the unity relation "' and uses the criterion and property inheritance
principles of the form (Inher) to govern her discourse about Fs. Assume that the
agent stands in an appropriate relation to specifications a and (3, which are in
the field of "', and uses these specifications as if to make claims about f (a) and
f ((3). Then:
(i) the agent is in fact referring to objects f (a) andf ((3),
(ii) a "' f3 suffices forf(a) = j(f3),
(iii) for any instance of (Inher), the right-hand side suffices for the left-hand side.
This initial formulation obviously needs further explication and defense. For one
thing, we need to be told what it is for a criterion of identity not to "presuppose"
the objects for which it is a criterion. 31 For another, we need to be told what it is for
an agent to be appropriately related to a unity relation and some specifications. 32
2.6 Some Objections and Challenges

Let us now consider some objections to the thesis of reference by abstraction, as well
as challenges that require further clarification. A more systematic and detailed defense
of the thesis is provided in Part III of the book.
31
My answer is outlined in Section 2.6.1 and developed in Section 3.3 and Chapter 6.
32 See Section 2.6.3 and, for more details, Chapter 8.
2.6.1 The bad company problem

An obvious objection is that the good instances of two-level criteria of identity are
surrounded by "bad companions" that are inconsistent or otherwise problematic.
A notorious example is Frege's Basic Law V, which states that two concepts share the
same extension just in case they are coextensive. We can formalize this as
(BLV) x.Fx = x.Gx ~ Vx(Fx ~ Gx)
where 'F' and 'G' are second-order variables that range over Fregeau concepts and
where 'x.Fx' stands for the extension of F. As Russell discovered, this "law" gives
rise to the paradox now bearing his name. To see this, it is useful to think of Frege's
extensions as classes. The "law" allows us to consider the Russell class r whose mem-
bers are each and every object that is not a member of itself. We now ask whether r
is a member of itself. It is easy to derive the contradiction that r is a member of itself
just in case it is not.
My response is to dismiss the proposed counterexample on the grounds that one
of the assumptions of the thesis of reference by abstraction is violated. Basic Law V
attempts to provide a criterion of identity for extensions. But it does so in a way
that quantifies over-and thus presupposes-the very extensions that the criterion is
meant to govern. The "law" therefore violates the assumption that the identity condi-
tions for the objects to be introduced by abstraction presuppose only an antecedently
accepted ontology. 33
2.6.2 Semantics and metasemantics
What, exactly, is the kind of "account of reference" that I have in mind and in which
criteria of identity have an important role to play? To answer the question, we need
to distinguish three types of question that can be asked about relations. One type
concerns which objects stand in the given relation. Which bicycle is mine? To whom
am I married? To whom does the name '0I: refer? A second type of question concerns
the causal and historical process by which the objects came to stand in some relation.
When and why did I buy my bicycle? How did I meet my wife, and why did we marry?
Why did my parents choose my name, and why have other speakers followed suit?
A third type of question arises for all relations that aren't metaphysically primitive
(unlike, perhaps, the relation between an elementary particle and its charge or mass).
In virtue of what do I own that bicycle? What makes it the case that I am married
to this woman? And what is involved in the relation of reference obtaining between
this name and that person? Questions of this third type will be called constitutive and
must be carefully distinguished from the extensional and etiological questions of the
first and second types, respectively.
33 A more complete explanation of the bad company problem and my response to it is provided in
Chapters 3 and 6.
In fact, there is already an established name for the subclass of constitutive ques-
tions that concern the relation of reference and other relations studied by semantics.
Such questions are said to belong to metasemantics, as opposed to semantics proper. 34
Semantics is concerned with how the semantic value of a complex expression depends
upon the semantic values of its various simple constituents. We ascribe to each
complex expression a certain semantic structure and explain how its semantic value
is determined by this structure as a function of the semantic values of its simple
constituents, typically in accordance with the principle of compositionality. Meta-
semantics, on the other hand, is concerned with the constitutive question of what
is involved in an expression's having the semantic properties that it happens to have,
such as its semantic structure and its semantic value.
The "account of reference" that I develop is concerned with the metasemantic (and
thus also constitutive) question of what makes it the case that a singular term refers
to a particular object. In particular, Frege's famous question of how numbers are
"given to us" belongs to metasemantics, not to semantics proper. 35 The corresponding
semantic question would be: To what objects do numerals of various sorts refer? For an
ontological realist such as Frege, such questions have easy answers. For instance, both
the ordinary decimal numeral '7' and the Roman numeral 'VII' refer to the natural
number 7. The metasemantic question with which Frege is grappling is much harder:
How is it so much as possible to refer to numbers and other abstract objects, given that
we cannot have any "ideas or intuitions" of them and (we might add) given that no
causal relation to them is possible? The Fregean proposal I am developing holds that
criteria of identity have an important role to play in the answers to this and related
questions.
There is every reason to expect our metasemantic questions to have answers-even
if only complex and messy ones. For it seems deeply implausible that relations of
reference should be fundamental and irreducible. As (Fodor, 1987, p. 97) observes,
when physicists one day draw up a catalogue of "ultimate and irreducible properties
of things", then "the likes of spin, charm, and charge will perhaps appear upon their
list [but] aboutness surely won't''. And the same goes, of course, for reference (or, for
that matter, for ownership and marriage).
2.6.3 A vicious regress?
Some philosophers believe that the kind of account I am developing falls prey to
a vicious regress. 36 The account holds that reference is constituted in part by the
subject's being suitably related to a specification of the referent. Suppose this "suitable
relation" to a specification has to be one of reference. To refer to this specification,
we would then need a further specification and unity relation, which would set off a
34
See (Stalnaker, 1997). This distinction corresponds to Stalnaker's distinction between "descriptive"
and "foundational" semantics. See e.g. (Stalnaker, 2001). See also Section 7.6.3 for further discussion.
35 See (Frege, 1953, §62).
36 See e.g. (Dummett, 199la, pp. 162-3) and (Lowe, 1998, ch. 7).
vicious regress. To refer to an object, we would always need a prior ability to refer to
some more basic object in terms of which the aforementioned object is specified.
A natural response to the alleged regress is to deny that the subject's relation to
a specification needs to be one of reference. Without this supposition, the regress
does not get going. How plausible is the supposition? I grant that there are cases
where the subject's relation to a specification is indeed one of reference. Reference to
natural numbers provides an example. On the account adumbrated in the Preface and
developed in Chapter 10, a natural number is specified by means of a numeral type,
which may in turn be specified by a numeral token. The crucial question, however, is
whether the subject's relation to a specification is always one of reference. The regress
arises only if the answer is "yes". But the correct answer is "no''. There is no reason
to believe that a subject's relation to a specification is always one of reference. Our
discussion of reference to physical bodies provides one example. Here the subject's
relation to the parcel of matter that serves as a specification is a purely causal one. This
stops the regress. Another example is provided by the account of arithmetic based on
Hume's Principle, where a Fregean concept acts as a specification of a number. This
too stops the regress, because the appropriate relation to a concept is one of grasping
or possessing, not one of referring.
While this response is fine as far as it goes, it may be objected that it does not
go far enough. 37 Granted, a subject's relation to a specification need not be one of
reference. But what about us theorists who are trying to explain how this specification
contributes to the constitution of reference? Whatever one says about the subject, it
seems that we theorists are referring to the specification in question. When explaining
the reference to physical bodies, for example, our relation to a parcel of matter that
serves as a specification is not merely a causal one, as we are singling the parcel out
and making various claims about it. And since we theorists are referring to parcels of
matter, we would like to know how this reference is constituted. This reintroduces the
threat of a regress.
The objection shows that our initial response cannot stand on its own but needs
to be supplemented. Thankfully, a supplementary response is ready to hand. Recall
that we are only proposing a sufficient condition for reference, not a necessary one.
For the regress to get going, we need two-level criteria of identity to be necessary for
reference, or at least for certain fundamental forms of it. But I am not committed to
this necessity claim. In particular, I do not claim that we theorists' reference to parcels
of matter comes about through the use of some two-level criterion.
2.6.4 A clash with Kripke on reference?

Suppose reference to an object is achieved by means ofa specification a. Is my account
of reference to this object compatible with the influential view, due to Kripke and
37
Thanks to Tobias Wilsch for pressing this objection.
others, that proper names and demonstratives refer in a direct way? The role of the
specification a calls this into question. Isn't the referent given only indirectly as the
f ofo?
In fact, there is no conflict so long as the thesis of direct reference is properly
understood. The thesis belongs to semantics and holds that the semantic value of
a singular term such as a proper name or a demonstrative is just its referent. This
contrasts with descriptivism about reference, which is a competing semantic thesis
according to which the semantic import of such expressions is much richer and, in
particular, involves descriptive content. The semantic thesis of direct reference is com -
patible with a variety of metasemantic views on the constitution of reference, including
ones according to which this constitution involves some non-trivial structure. Indeed,
when Kripke conjectures that an initial "baptism'' and subsequent causal chains serve
to constitute certain forms of reference, he proposes a metasemantic account with
non-trivial structure. Clearly, Kripke took his conjecture to be compatible with his
emphasis on reference as a semantically direct relation. 38
We need to distinguish the question of what objects stand in some relation from
the entirely different question of what facts make it the case that the objects are so
related. The thesis of direct reference concerns the former type of question. A proper
name bears the reference relation directly to its bearer rather than to other items in
terms of which the bearer is then picked out. By contrast, my sufficient condition for
reference concerns the latter type of question. It is not a brute fact that a name refers
to its bearer but a fact whose proper explanation sometimes involves a further object,
namely a specification of the referent.
2.6.5 Internalism about reference
How will a nominalist respond to my thesis of reference by abstraction? Consider the
account of reference to directions. We may safely assume that the nominalist has no
trouble with lines or parallelism. The lines can be taken to be concrete, and all that is
required concerning parallelism is that the relation be defined. We are not committed
to any particular metaphysics of predication based on universals or reified relations,
which the nominalist might find objectionable. The bone of contention is the claim
that a line and parallelism suffice to specify a direction. The nominalist objects to this
claim on the grounds that there simply are no directions.
The advocate of the Fregean approach will respond that we have specified a
direction, namely the direction of some line (reference to which is not in question).
There is no alternative, more direct or secure way of specifying a direction than what
38
In fact, my view is not even committed to metasemantic descriptivism, which holds that the referent
is fixed by means of a definite description. The specification and unity relation I invoke to explain how the
referent is picked out need not be part of a definite description. As we saw in Section 2.6.3, the relation
to the specification need not be one ofreference; and as emphasized in Section 2.3, the grasp of the unity
relation need not be an explicit one, which is accessible to consciousness.
we have already offered. A direction simply is the sort of thing that is most directly
specified by means of a line and subject to the unity relation of parallelism.
Clearly, we are confronted with a fundamental disagreement about what it takes
to specify a direction. To break the impasse, it is useful to consider a structurally
analogous debate that arises in the case of physical bodies. Here too I claim to have
provided an account of reference. To specify a physical body, it suffices to have causally
interacted with one of its parts and to be operating appropriately with the relevant
unity relation. Assume someone challenges me to demonstrate that there really exists a
physical body associated with some parcel of matter that both parties admit exists and
is in the field of the unity relation. The challenger demands that the alleged referent
be shown to her in a more direct or secure way that she too would find acceptable.
Clearly, there is nothing I can do that would satisfy the challenger. A physical body
just is the sort of thing that is most directly specified by means of a parcel of matter
and is subject to the appropriate unity relation. To demand that a body be specified
or shown in some altogether different way is to demand the impossible.
The comparison with the case of physical bodies brings out some important lessons.
First, the nominalist's challenge is just an instance of a far more general skeptical
challenge concerning what it takes to specify an object. It isn't the abstractness of
the desired object that is fueling the challenge but some very general preconceptions
about what it takes to specify an object. The Fregean response is to reject these
preconceptions as unreasonable.
Second, my Fregean view involves a form of internalism about reference. There
are certain basic forms of reference instances of which cannot be explained without
relying on the very form of reference in question. To have any chance of explaining
reference to a physical body, we need to be willing to engage in this very form of
reference in the metalanguage. And there is nothing special about physical bodies in
this regard. For there to be any hope of explaining reference to a given direction, say,
we must be willing to engage in reference to directions in the metalanguage; otherwise,
we would be demanding the impossible.39 Once this internalism is taken on board, the
Fregean's claim that the line suffices to specify a direction can be properly defended,
or so I claim. My argument is developed in detail in Chapter 8.
2.7 A Candidate for the Job

When the idea of thin objects was first introduced in Chapter 1, I characterized an
object as thin insofar as a comparatively weak claim-not ontologically committed to
39 Chapter 5 distinguishes my internalism about reference from some more radical "ultra-thin" con-
ceptions of reference, and Chapter 8 develops the internalism in more detail. A view related to mine is
found in (Dummett, 1995), where we read : "Frege not only believes in semantics, but constructed a general
framework for it. He held, however, that a semantics for a language must be internal to that language: it
must describe it as from within, not from some external viewpoint" (p. 17; emphasis in original).
A CANDIDATE FOR THE JOB 43
the object in question-nevertheless suffices for the existence of the object. I explained
how the idea of thin objects holds great promise for the philosophy of mathematics
and metaphysics more generally. The crucial question is how to understand the notion
of sufficiency. To make progress, I formulated a "job description" that any notion of
sufficiency must satisfy if the resulting account of thin objects is to deliver on its
promises. Let us now examine whether we have succeeded in identifying a suitable
candidate for the job. The examination can only be preliminary, however, as I have
only outlined my account, leaving for later chapters the task of filling in various details.
Consider the case of directions. Suppose we start with a domain containing lines
(and perhaps other objects) and on which relations of parallelism and orthogonality
are defined (but are deemed false whenever one or both of the relata is not a line or
some other directed item). I explained how our conceptual apparatus for identification
and predication can be used in a novel way such that directions, not lines, are
the objects that are identified or distinguished and that serve as subjects of other
predications. We add to our language vocabulary for talking about directions and
lay down clear and strict rules for how the extended language is to be used. These
rules are carefully chosen so as to ensure that the laws of logic remain valid in the
resulting extension of the language. My central claim is encapsulated in the stated
thesis of reference by abstraction. In essence: when the extended language is used in
the described way, we succeed in referring to directions and expressing meaningful
statements about them.
Let us say that rp suffices for 1/r when rp is a ground for asserting 1/r in a per-
missible language extension of the sort described. In the case discussed above, the
statements 11 1112 and11J_12 suffice for d(/1) = d(l2) and d(l 1 ) _!_ * d(lz), respectively. 40
For abstraction more generally, we obtain the universal closures of the following
sufficiency statements:
a1 ,.,., a2 => f(a1) = f(a2) ....,a, ,.,_, az => -f(a1) = j(a2)

rp(a1i ... , an) => rp*(j(a1), .. . ,f(an)) ....,rp(a1i . .. , an) => ....,rp*(j(a1), ... ,f(an))
Notice that in all these formulas, the statement on the left-hand side of the sufficiency
operator is concerned only with the "old" entities with which we began, which are
accepted by all parties, while the statement on the right-hand side is concerned with
"new" objects that are obtained from the "old" entities by abstraction. 41
40 A general definition is provided in Chapters 8 and 9. Beginning in Chapter 3, we shall iterate the
abstraction-based language extensions just described. This makes it necessary to define a more general
notion of mediate sufficiency (contrasted with the immediate notion just sketched). We say that rp mediately
suffices for i{I just in case there is a series of permissible reconceptualizations that takes us from rp to i{I by
a chain of relations of immediate sufficiency.
41 As discussed in Section 1.7, my notion of sufficiency is closely related to a form of metaphysical
grounding. Let me now be more specific. Suppose at some stage of extending the interpretation of our
language we have established the truth of the left-hand side, but not of the right-hand side, of one of the
displayed sufficiency statements. Then a further language extension establishes the truth of the right-hand
side as well, in such a way that the left-hand side serves as a (strict, full, and immediate) ground for the
I contend that this notion of sufficiency is a good candidate for our job-but hasten
to add that a final assessment will have to await the detailed development of my
account in later chapters. To defend my contention, let us consider the constraints
that make up the job description.
First, we required that there be true sufficiency claims cp ==> 1/1 where the ontological
commitments of 1/1 exceed those of cp. This constraint is clearly satisfied, as shown
by the examples involving directions. 42 Second, we required that the sentences
flanking the sufficiency operator can be taken at face value; in particular, that the
statement on the right-hand side is not just a faron de par/er for the one on the left-
hand side. If my account works at all, this constraint will be satisfied: for the account
was designed specifically to show how reference to the new objects that figure on the
right-hand side of the sufficiency operator is constituted.
We now come to the epistemological and explanatory constraints. Roughly, these
constraints require that for any true sufficiency claim, the corresponding mate-
rial conditional must enjoy-or potentially enjoy-a privileged epistemological and
explanatory status. The material conditional must be within our epistemic reach
and admit of an adequate metaphysical explanation. Are these constraints satisfied?
A proper answer would require a careful analysis of the relevant kinds of epistemic
and explanatory status. I have no such analysis to offer and anyway prefer my account
to be independent of the finer details of any such analysis. For now, I therefore
content myself with observing that we are in a good position to show that the
constraints are satisfied. Consider, once again, the case of directions. I explained how
the criterion of identity for directions can be used to effect a reconceptualization,
where for example the statement 11 II 12 is reconceptualized as d(lt) = d(l2 ). I also
argued that, because of its predicative character, this reconceptualization carries no
metaphysical or epistemological presuppositions. When we use the extended language
with direction vocabulary in the described way, we reconceptualize reality in a novel
way, and there is no sense in which reality might fail to cooperate. Moreover, there is
nothing about this reconceptualization that is deeply hidden from us. By appropriate
reflection, we can therefore come to understand how it works. 43
All that remain are the logical constraints from Section 1.5. These constraints
are shown to be satisfied in Chapters 8 and 9 (in a two- and one-sorted setting,
respectively).
right-hand side. In this way, the sufficiency statements can be seen as recording grounding potentials: if the
left-hand side of such a statement becomes "available" before the right-hand side, then the former provides
a (strict, full, and immediate) ground of the latter.
42
Recall that we are relying on an ordinary Quinean notion of ontological commitment. A non-Quinean
alternative will be discussed, and rejected, in Section 4.4.
43
I have more to say about the epistemology of abstraction in Chapter 11.
THICK VERSUS THIN 45
2.8 Thick versus Thin

In closing, I wish to return to the distinction between thick and thin objects. How,
exactly, does the existence of physical bodies make a more substantial demand on the
world than the existence of natural numbers? We now have the resources to answer
the question.
On the account I have developed, the question of what it takes for an object a to
exist is a matter of what it takes for there to be a specification of a which is in the field
of the relevant unity relation. Some examples will help to convey the idea. Consider a
physical body a, say the table at which I am sitting. Then the candidate specifications
of a are the spatiotemporal parts of the table, and the unity relation is that of belonging
to the same cohesive and naturally bounded whole. Consider any spatiotemporal part
u of a. For this specification to succeed, u must belong to a cohesive and naturally
bounded whole. In contemporary metaphysical jargon, u must be a part of some stuff
arranged "tablewise". The upshot is that the existence of my desk makes a substantial
demand on a particular region of spacetime-just as one would expect.
Other objects are thinner. For a direction to exist, it suffices that there be an
appropriately oriented line. This is far less demanding than what we found in the case
of physical bodies. There is no requirement on any particular region of spacetime,
since the witnessing line could be located anywhere. In fact, I later argue that the
existence of a direction does not require that there must actually exist a line that
instantiates it; it suffices that there possibly exists such a line. 44 In general, abstract
objects are thinner than concrete objects because they do not make demands on any
particular region of spacetime.
Surprisingly, it turns out that not all abstract objects are equally thin. To see why,
let us follow (Parsons, 1980) and say that an object is pure abstract if it lacks both
spatiotemporal location and any kind of intrinsic relation to space or time. The
natural numbers and pure sets are examples. Let us then say that an object is quasi-
concrete if it lacks spatiotemporal location but nevertheless has canonical realizations
in spacetime (and only there). Letters and geometrical figures provide examples, as
these have canonical realizations in the form of tokens and concrete figures with
the shapes in question. The surprising discovery is that quasi-concrete objects are
somewhat thicker than pure abstract objects. The existence of a quasi-concrete object
makes a non-trivial demand on spacetime, however weak and indirect: there must be,
or at least possibly be, concrete realizations of the object somewhere or other in space
and time. There are presumably no quasi-concrete twenty-dimensional geometrical
figures because spacetime is not rich enough to allow realizations of such figures.
Nor are there quasi-concrete Euclidean triangles; for the spacetime that we inhabit is
curved, not Euclidean. (Of course, I do not deny that such figures and triangles exist
44
See Section 11 . 1.
as pure abstract objects.) Far more dramatic conclusions would follow if spacetime
turned out to be granular, as suggested by some recent physical speculations. Then
there would be surprisingly few quasi-concrete geometrical figures.
There is an interesting connection between how thin an object is and the modal
robustness of its existence. The less that the existence of an object demands of the world,
the more modally robust the object is. Since my desk makes a substantial demand
on the world, for example, there are nearby circumstances in which it fails to exist:
the carpenter who made the desk might instead have used the wood to make some
chairs. So my desk is modally fragile. A pure abstract object, by contrast, makes no
demand on the world and therefore exists by necessity. The cardinal number 0 would
have existed however the circumstances had been, as there would still have been
empty collections from which this number could be obtained by abstraction. Quasi-
concrete objects are intermediate as concerns the modal robustness of their existence.
A linguistic type, for example, would still have existed in any circumstances in which
it could be instantiated. Its existence thus depends on the existence of a certain sort
of space, but not on anything specific about the existence or organization of matter in
this space.
Appendices
2.A Some Conceptions of Criteria of Identity
Philosophers have defended some very different conceptions of criteria of identity. I shall now
compare and contrast my own favored conception with two important alternatives.
Some philosophers regard criteria of identity as metaphysical principles, which provide
information about the nature of the objects in question. For instance, the criteria of identity
for sets and directions say something important about the nature of sets and directions,
respectively; and an account of personal identity or the persistence of ordinary material objects
over time would provide valuable information about such objects. Some philosophers even
claim that criteria of identity explain "what grounds the identity and distinctness" of objects. 45
An example frequently used to illustrate this claim are sets, whose identity or distinctness is
said to be "grounded" in accordance with the principle of extensionality.
Another view regards the principal interest of criteria of identity as epistemic. For instance,
(Geach, 1962, p. 39) describes a criterion of identity as "that in accordance with which we judge
as to the identity" of certain objects. How do we find out whether two encounters with an object
are encounters with one and the same object? Assume, for instance, that you and I both extend
our right arms to demonstrate a direction. Then the criterion of identity for directions informs
us that the directions we have demonstrated are identical just in case our arms are "co-directed"
(where two directed items are said to be "co-directed" when they are not just parallel but also
oriented in the same way) .
45
See e.g. (Lowe, 2003).
APPENDICES 47
Finally, the conception defended in this chapter focuses on the role that criteria of identity
play in a metasemantic account of reference. 46 What is it about the world that makes a certain
expression or representation refer to a certain physical body or abstract direction, as opposed
to all of the objects to which it might have referred? I have argued that criteria of identity play
an essential role in the answer.
The three conceptions of criteria of identity differ radically in what they emphasize. My own
view is that, while criteria of identity do have metaphysical and epistemological aspects, it is
advantageous to regard them first and foremost as metasemantic principles. This view allows
us to avoid some difficulties associated with the epistemic and metaphysical conceptions.
Let us start with the epistemic conception. Imagine that you on two different occasions are
presented with an object. How can you find out whether the objects with which you were
presented are identical or distinct? The means by which you can find out depend in part
on facts about reference and the nature of the objects in question-that is, on the matters
that are emphasized by the metasemantic and metaphysical conceptions. If you are referring
to sets rather than to properties, for example, it will be appropriate for considerations of
coextensionality to figure in your deliberations. Or, if you are referring to persons and some
sort of bodily continuity provides the right criterion of personal identity, then this will have
consequences for your ways of answering the question. To a limited extent one may even
proceed in the opposite direction and glean information about the nature of the objects in
question from the means available for answering questions about identity and distinctness. If
we know that coextensionality is such a means, this suggests we are talking about sets rather
than properties. All this is unsurprising. The epistemic conception understands "criterion'' as
a mark by which we identify some phenomenon, or a symptom. And clearly, the marks or
symptoms of any phenomenon will depend in large part on the nature of the phenomenon,
and to a limited extent, may even shed light on the phenomenon itself.
If a phenomenon has enough of a nature to admit of a systematic investigation, however, it
is misguided to focus on the marks rather than on the phenomenon itself. Why focus on the
symptoms rather than on the underlying disease? Doing so can only hamper the investigation.
This general point applies to criteria of identity as well. Our means for finding out about identity
and distinctness are exceedingly complex and messy. Co!lsider our identification of people we
have met. The vast majority of our knowledge of such people is based not on tracing bodies
through space and time (or, for that matter, an investigation of their memories) but on a wealth
of indirect evidence involving looks, utterances, and a vast variety of other cues. Any attempt
to identify a set of canonical ways of finding out about the identity and distinctness of a certain
class of objects will only serve to shift the focus back to where the two competing conceptions
prefer to locate it.
The metaphysical conception of criteria of identity faces difficulties of its own. What distin-
guishes a criterion of identity, on this conception, from other metaphysical truths? A criterion
of identity must presumably have a logical form of the sort represented by (CI-1) or (CI-2)
above. And the criterion must presumably be metaphysically necessary. Let minimalism be the
metaphysical conception of criteria of identity that comprises just these two requirements.47
46 Similar ideas are found in (Dummett, 198la) and (Dummett, 199la).

47 This view is defended in (Horsten, 2010). (Minimalism about criteria of identity must not be confused
with what I have called metaontological minimalism.)
Although the minimalist conception is appealingly simple and well understood, it is insufficient
to distinguish criteria of identity from other necessary truths. The account is unable, for
example, to explain why we should prefer extensionality as the criterion of identity for sets
to the principle that two sets are identical just in case they are elements of the exactly same sets,
or in symbols:
SET(x) /\ SET(y) ~ (x = y ++ 'v'u(x E u ++ y E u))
After all, this principle too is a metaphysical truth concerning the identity of sets. In fact, this
example is just the tip of the iceberg. It turns out that any necessary truth can be made to follow
logically from a criterion of identity in the minimalist sense. Consider a criterion of the form
(CI-1), and assume that there are at least two objects of the relevant kind F. (The case of (CI-2)
is analogous.) Let <p be a necessary truth and consider:
(2.1) Fx /\ Py~ (x = y ++ (Rp(x,y) v -,rp))

It is easy to verify that (2.1) too is a criterion of identity in the minimalist sense-but with the
additional feature that this "criterion'' entails rp! 48
To rule out perverse criteria of identity of this sort, defenders of the metaphysical conception
need to go beyond minimalism. Given a true identity t1 = t2, they need to distinguish the
statements <p that "merely" ensure the truth of the strict biconditional D (t1 = t1 ++ <p) from
those that actually explain or "ground" the identity t1 = tz. Some philosophers happily take on
this explanatory burden (Lowe, 2003) . It is not an easy burden to discharge, however.
I believe my metasemantic conception explains how criteria of identity are distinguished
from other metaphysical truths. Criteria of identity figure at the heart of an account of how
certain fundamental forms of reference are constituted. As a result, the criteria of identity come
to govern the referents in questions. Recall the Fregean triangle that figures so prominently
in this chapter. As this triangle displays, criteria of identity are connected both to reference
(as emphasized by the metasemantic conception) and to objecthood (as emphasized by the
metaphysical conception).
2.B A Negative Free Logic

Throughout this book, we often use a negative free logic, which is so called because any atomic
predication involving a non-referring singular term is deemed false. Thus, 't = t' is true just
in case there is a true atomic predication involving t. This ensures that we can use 't t' as an =
existence predicate.
The possibility of non-referring terms requires some changes to the axiomatization of first-
order logic. To begin with, it is only permissible to instantiate a universally quantified formula
with respect to a term t on the assumption that t refers. We therefore adopt the following
restricted axiom scheme of Universal Instantiation:
48 Clearly, (2.1) has the right logical form. And if (CI- I) and <p are necessary truths, then provably so is
(2.1). It remains to show that (2.1) entails <p. Let x and y be distinct Fs, which by assumption is possible.
Then the biconditional on the right-hand side of(2.l) is true. Since the left-hand side of this biconditional
is then false, so must be its right-hand side, which means that <p must be true. Thus <p follows from (2.1).
This is a trick familiar from the neo-Fregean literature; see e.g. (Heck, Jr., 1992). See also (Leitgeb, 2013)
for a similar argument.
APPENDICES 49
Vvrp(v) /\ A(t) ~ rp (t)
where A(t) is any atomic predication in which t occurs. Since we also accept universally
quantified versions of all the ordinary laws of identity, including Vx(x = x), we can prove
A(t) ~ t = t, with A(t) as above.
The rule of Universal Generalization is restricted such that we are permitted to universally
generalize on a term t only on the assumption that t refers:
Assume I- <p /\ t = t --+ 1/t(t), where tis a singular term that does not occur free
in <p. Then we may infer I- <p--+ \:/v lft(v).
2. C Abstraction on a Partial Equivalence

The literature on abstraction is almost exclusively concerned with ordinary equivalence rela-
tions.49 However, in the main text we observed that there is good reason to permit abstraction
on partial equivalence relations, that is, on relations that are symmetric and transitive but
not necessarily reflexive. I now outline three different approaches to abstraction on a partial
equivalence relation and comment on how they are related. I use abstraction on objects as my
example. The extension to abstraction on various sorts of concepts is straightforward.
One option, to which the main text alluded, is to employ a negative free logic, which permits
non-referring singular terms. (See Appendix 2.B.) In the context of this negative free logic,
abstraction principles can be formulated in the ordinary way, for instance:
(AP) §x = §y ++ x ~ y
Had the logic been classical, the reflexivity of identity would have entailed §x = §x and thus
also the reflexivity of ~. But this inference is blocked in our negative free logic. Instead, we can
prove the desired result that an object specifies an abstract just in case this object is in the field
of~; that is:
(2.2) Vx(x ~ x ++ 3y(y = §x))

There are alternative ways to handle abstraction on a partial equivalence which allow us to
retain classical logic. Let me describe two options. But in each case, the price of classicality is a
reformulation of the abstraction principle. The first alternative is to adopt a conditional version
of the abstraction principle:
(AP,) x ~ x /\ y ~ y ~ (§x = §y ++ x ~ y)
Of course, to maintain classicality, the operator § must be total. So what is the "abstract" of
an object that is not in the field of~? In fact, we may help ourselves to such "abstracts" as an
innocent convenience. In cases where x ~ x is false, §x can be anything whatsoever; we just
need to ensure that the choice never matters. This is precisely what we achieve by giving the
abstraction principle its conditional formulation. In cases where we abstract on objects outside
49
Two noteworthy exceptions are (Macfarlane, 2009) and the response (Hale and Wright, 2009a).
See also (Payne, 2013a), which, in addition to providing a useful overview, defends the importance of
abstraction on partial equivalence relations and shows how this can be achieved by means of negative free
logic.
of the field of~, the antecedent of the abstraction principle is false, thus rendering the relevant
instance of the abstraction principle true regardless of how the unintended cases are handled.
The second alternative is to "factorize" the abstraction principle (AP) into separate criteria
of existence and of identity, namely:
u ~ u-+ 3xAas(u,x)
Aas(u,x) /\ Aas(v,y)-+ (x = y ++ u ~ v)
As we shall see in Chapter 9, this approach is particularly useful when we study systems of
abstraction principles, where we allow the relation ~ on which we abstract to vary.
The different approaches have their respective advantages and disadvantages. Different
choices will therefore be made in different contexts. Thankfully, these choices are ultimately of
little philosophical importance, as the approaches are nothing but notational variants of one and
the same idea. To substantiate this claim, we can show that the approaches are interpretable in
each other and indeed that the translations employed are natural and arguably respect intended
meanings. I now describe the translations but leave the task of verifying that they provide
interpretations as an exercise for the reader.
We begin with a translation from the second approach into the first. The translation maps
an atomic predication A to
where t1, .. . , tn are all the singular terms such that §t; occurs in A . Otherwise, the translation
is trivial, in the sense that it commutes with all the connectives and quantifiers. The reverse
translation maps an atomic predication A to
where t1, .. . , tn are as above. Otherwise the translation is trivial.

Next, I describe a translation from the third approach to the second. The single non-trivial
clause is:
ABS(u,x) x = §u /\ u "' u
In particular, other atomic predications are mapped to themselves. The reverse translation
behaves non-trivially only on identities involving abstraction terms, where we need to allow
the "abstracts" of objects outside the field of~ to behave any way they wish. We ensure this
as follows. For every abstraction term §u involved in the identity, we make the identity claim
conditional on u ~ u by translating as follows:
x= §u u ~ u-+ Aas(u,x)
3
Dynamic Abstraction
3.1 Introduction
Any abstractionist approach to thin objects confronts the threat of paradox. The
most famous manifestation of the threat is Frege's Basic Law V, which states that the
extension of a concept F is identical to that of a cqncept G just in case the concepts
are coextensive. We can formalize this "law" as
(BLV) x.Fx = x.Gx # Vx(Fx # Gx)
where 'F' and 'G' are second-order variables that range over Fregean concepts and
where 'x.Fx' stands for the extension of F. As Frege learnt from (Russell, 1902), the
"law" falls prey to paradox.
Once discovered, Russell's paradox is straightforward. Let us say that xis a member
of y just in case x falls under some concept whose extension is y. That is, we define
x E y as 3F(Fx /\ y = u.Fu). Consider the concept R defined by
(3.1) Vx(Rx # x rf. x).
This definition ensures that the extension of the concept, namely r = u.Ru, is a kind
of Russell class whose members are all and only th~ objects that are not members of
themselves. So we ask whether r is a member of itself. Using Basic Law V, it is easy to
derive a contradiction.
What to do? After Russell's shocking letter, Frege hurriedly proposed a fix. 1 But his
proposal has since proven to be flawed, and Frege (who may or may not have been
aware of the flaw) appears to have given up on abstraction around 1906. 2 This seems
an overreaction, however, given the importance of abstraction to our mathematical
thought. Let us be more courageous and try to do better!
I begin, in the next section, by describing the approach to abstraction that has
dominated in recent decades. This is the neo-Fregean approach developed by Bob
Hale and Crispin Wright in a series of important works, but which has also benefited
from contributions by a large group of other philosophers and logicians (thus making
1
See the appendix to (Frege, 2013).
2 The inconsistency re-emerges provided the domain contains two or more objects. See (Burgess, 2005,
pp. 32-34) for a useful discussion.
52 DYNAMIC ABSTRACTION
this a paradigm of collaborative research in philosophy). 3 A distinctive feature of this

approach is that it is static, in the sense that it operates with a single fixed domain
in which one or more abstraction principles are taken to be true. 4 As in Frege's own
work, this fixed domain is typically taken to comprise all of reality. 5 When abstraction
takes place in some fixed domain, the reasoning that leads to Russell's paradox is
incontrovertible. It follows that Basic Law Vis beyond rescue. But instead of following
Frege into despair, the neo-Fregeans insist that many other abstraction principles are
unaffected by the malaise and may therefore continue to serve as the technical engine
of a philosophical account of mathematics.
This differentiation between acceptable and unacceptable abstraction principles
gives rise to a major challenge for the static approach. On what grounds is the
differentiation to be made? This is known as the bad company problem. 6 The problem
is hard, as we shall see, not least because many "bad companions" are unacceptable
for more subtle reasons than Basic Law V. Ideally, we would like a technical and
philosophical understanding that explains what goes wrong in the bad cases. We
would like to exclude the "bad companions" for a principled reason, not merely as
an ad hoc expediency.
This chapter develops an alternative dynamic approach to abstraction.7 On this
approach, abstraction is not tied to a single fixed domain but may take us from
one domain to a larger one. When directions are obtained by abstracting on lines
under the equivalence of parallelism, for example, there is no reason to assume
that the directions are present already in the domain with which we started. The
dynamic approach to abstraction gives a radically different answer to the bad company
problem: it accepts abstraction on just about any equivalence relation. (The small
qualification will be explained below.) Paradox is avoided, not by being selective about
which abstraction principles we accept but by allowing the abstracta to lie outside of
the domain on which we abstract-much like directions may be taken to lie outside
of the domain oflines. Basic Law V provides a good illustration. (If it can be redeemed,
surely any abstraction can!) Suppose the extensions that result from abstracting on
the concepts on some fixed domain are allowed to lie outside of this domain. Then
3 See (Wright, 1983), (Hale, 1987), and (Hale and Wright, 200la). Other important contributions
include (Boolos, 1998), (Burgess, 2005), (Fine, 2002), (Heck, Jr., 2011), and (Heck, Jr., 2012).
4
An important exception to the predominantly static approach pursued so far is (Fine, 2002), where
various more dynamic ideas are explored. See also Fine's "procedural postulationism" for some related ideas,
e.g. (Fine, 2005b ). Even in the work of Hale and Wright, however, we find traces of more dynamic ideas; for
instance in their discussion of the significance of indefinite extensibility for the possibility of abstraction,
as well as in the construction canvassed in (Wright, l 998a).
5 However, there is no expectation that this all-encompassing domain should form a set-which would
be a potentially problematic universal set. Indeed, there is no immediate need to reify the domain at all,
that is, to regard it as an object. After all, talk about truth in an all-encompassing domain can be replaced
by talk about truth simpliciter.
6 See (Linnebo, 2009d) and (Studd, 2016) for introductions to the problem and further references.
7 For earlier work on this approach, see (Linnebo, 2009a), (Payne, 2013b), and (Studd, 2016), as well as
the works cited in note 4.

NEO-FREGEAN ABSTRACTION 53
the paradoxical reasoning is blocked, as it becomes impermissible to instantiate the

quantifier 'Vx' in (3.1) with respectto the "Russell class" r. So provided thatthe domain
is allowed to expand, there is no paradox (cf. Section 3.3).
In fact, these domain expansions can be iterated. The resulting dynamic approach
to abstraction has important affinities with the influential iterative conception of
sets. 8 I show how modal operators can be used to represent the domain expansions
that are brought about by abstraction and thus also to develop a theory of iterated
abstraction (cf. Section 3.5). The modal operators even enable us to retrieve a form of
absolute generality, where we generalize across all of the possible domain expansions
(cf. Section 3.6). Is the resulting view stable? I argue that it is. When a domain can be
extensionally specified (i.e. specified as a plurality of objects), then it can be extended,
but not so when the domain can only be intensionally specified (i.e. when it does not
admit of specification as a plurality). My form of absolute generality is a case of the
latter (cf. Section 3.7).
3.2 Neo-Fregean Abstraction

To explain the neo-Fregean approach, it is useful to begin by briefly sketching Frege's
account of arithmetic. The account proceeds in two steps.
First, Frege gives an account of the applications and identity conditions of numbers.
He argues that counting involves the ascription of numbers to concepts. For instance,
when we say that there are eight planets, we ascribe the number eight to the concept
" ... is a planet''. Frege's claim is that the number-of operator'#' applies to any concept
expression F to form the expression '#F', which we may read as "the number of Fs''.
Frege argues that the number of Fs is identical to the number of Gs just in case the
Fs and the Gs can be put in a one-to-one correspondence. This is known as Hume's
Principle and is formalized as
(HP) #F= #G~ F ~ G
where F ~ G is a formalization in pure second-order logic of the claim that the Fs

and the Gs are equinumerous.
Second, Frege provides an explicit definition of terms of the form '#F'. In order to
do so, he uses a theory consisting of second-order logic and Basic Law V. He defines
#F as the extension of the concept " .. . is an extension of some concept equinumerous
with F''. 9 It is straightforward to verify that this definition satisfies (HP).
How should we respond to the unfortunate fact that Frege's approach is inconsis-
tent? A simple but radical answer is proposed in (Wright, 1983). Why not simply
abandon the second step of Frege's approach-which introduces the inconsistent
8
See (Boolos, 1971) and (Parsons, 1977) for two classic expositions.
9 More precisely, '#F' is defined as '.X.3G(x = y.Gy /\ F"" G)'.
theory of extensions-and make do with the first step? This proposal has given birth
to the neo-Fregean approach to the philosophy of mathematics.
Wright's proposal relies on two relatively recent technical discoveries. The first
discovery is that (HP), unlike Basic Law V, is consistent. To be precise, let Frege
Arithmetic be the second-order theory with (HP) as its sole non-logical axiom. Frege
Arithmetic can then be shown to be consistent if and only if second-order Dedekind-
Peano Arithmetic is. IO The second discovery is that Frege Arithmetic and some very
natural definitions suffice to derive all the axioms of second-order Dedekind-Peano
Arithmetic. This result is known as Frege's theorem. At least from a technical point of
view, the neo-Fregean approach is therefore a success: it is both consistent and strong
enough to prove all of ordinary arithmetic.
As the case of Basic Law V shows, however, the neo-Fregeans face the bad company
problem. They need to demarcate the acceptable abstraction principles from the
unacceptable ones. This demarcation has to balance two conflicting pressures. On the
one hand, the rot represented by Basic Law V and a wide variety of other unacceptable
abstraction principles has to be excised in a way that is both definitive and well-
motivated. On the other hand, we want to leave in place not only Hume's Principle
but also other abstraction principles that can serve as a foundation for analysis and
set theory, at least in part. This will require a careful balancing act.
The resulting challenge is hard, for at least two reasons. First, abstraction prin-
ciples that are generally regarded as acceptable have close relatives that are plainly
unacceptable. My favorite example is a close, but decidedly evil, cousin of Hume's
Principle. I I Observe that equinumerosity of concepts is just a matter of the concepts'
being isomorphic. So consider the abstraction principle that does to dyadic relations
what Hume's Principle does to concepts. This principle says that the isomorphism types
of two dyadic relations are identical just in case the relations are isomorphic:
tR=tS~R~S
Although this principle is closely related to Hume's Principle, it is inconsistent, as it

allows us to reproduce the Burali-Forti paradox. 12 So good principles, such as Hume,
have very close relatives that are bad.
Second, there are abstraction principles that are bad in more subtle ways than Basic
Law V, which at least has the good grace to wear its badness on its sleeves. These
principles are therefore more sinister than Law V. My favorite example is Wright's
10
Proof sketch: Let the domain D consist of the natural numbers. If a concept F applies to n objects,
let '#F' refer ton+ I. If F applies to infinitely many objects, let '#F' refer to 0.
11
In fact, Hume has other problematic relatives as well. A wonderful example, due to (Cook, 2009},
are "Hume's big brothers~ which do to monadic concepts of orders two and higher what Hume does to
monadic concepts of order one. The satisfiability of these big brothers turns out to be tangled up with the
truth of the generalized continuum hypothesis.
12 See (Hodes, 1984a, p. 138) and (Hazen, 1985, pp. 253-4).
HOW TO EXPAND THE DOMAIN 55
"nuisance principle': 13 Let us say that F and G differ finitely just in case there are finitely
many things that are F-and-not-G or G-and-not-F. This relation can be expressed
in pure second-order logic and is easily seen to be an equivalence. So consider the
associated abstraction principle:
(NP) nuisance(F) = nuisance(G) ++ F and G differ finitely
This principle turns out to be satisfiable on all and only finite domains. 14 While this
is a curious property, it is not inconsistent: there are, of course, finite domains. But
there is a problem. We know that Hume's Principle is satisfiable in all and only infinite
domains. 15 Thus, while the nuisance principle and Hume's Principle are individually
satisfiable, they are not jointly satisfiable. So they cannot both be ruled acceptable. In
fact, this is just the tip of the iceberg. Technical investigations have revealed a plethora
of more complex interactions as well, showing that abstraction principles that appear
acceptable on their own are not jointly acceptable.
How should the neo-Fregeans respond to the bad company problem? It is hard to
deny that a proper response is required. As emphasized by Frege himself, it is part
of the very nature of both mathematics and philosophy to seek general explanations
wherever such are possible. 16 Yet even after decades of work by a number of capable
researchers, there is no consensus in sight; on the contrary, new problems keep
emerging. 17 Most worrisome of all, in my opinion, is the extent to which recent work
on the bad company problem has become a largely technical undertaking, which has
lost touch with the underlying philosophical question of how abstraction might work.
Ideally, we would like our philosophical account of abstraction to motivate, or at least
inform, our answer to the bad company problem.
In short, it is time to try a new tack.
3.3 How to Expand the Domain

The static approach to abstraction is not obligatory. As mentioned, there is no need to
assume that the directions obtained by abstraction on lines under the equivalence of
parallelism belong to the domain on which we abstract. Let us examine the alternative
dynamic approach on which abstraction may result in "new" objects that lie beyond
the "old" domain with which we began.
13 See (Wright, 1997, section VI) . 1his style of bad companion was first observed by (Boolos, 1990), who
gives a slightly more complicated example known as the parity principle.
14
(NP) is obviously satisfiable in all finite domains, as there will then only be one equivalence class.
For the other direction, consider any infinite domain D. 1hen there can be no one-to-one mapping of the
equivalence classes in question to elements of the domain because there are, I claim, as many equivalence
classes as there are subsets. To prove this claim, assume, for contradiction, that there were fewer equivalence
classes than subsets. Since each equivalence class has ID<wl = IDI members, we would then be able to
factor l2D I as the product of two smaller cardinals, which is impossible.
15 16
1he "all" direction relies on the axiom of choice. See for instance (Frege, 1953, §§1-4).
17 See e.g. (Studd, 2016) and (Cook and Linnebo, 2018).
The dynamic approach is very natural in light of the philosophical account of

abstraction canvassed in Chapter 2. The distinctive feature of successful abstraction is,
according to this approach, that every question about the abstracts reduces to-and
thus is "reconceptualized" as-a question about the entities on which we abstracted. 18
For example, the question about the orthogonality of two directions reduces to a
question about the orthogonality of any two lines in terms of which the directions are
specified. We have to be careful, however. How do we know that these "reductions"
always lead to simpler questions, which thus eventually receive answers? The dynamic
approach offers a simple but powerful assurance. Every question about the "new"
abstracts reduces to a question about the "old" ontology on which we abstract. And
every question of the latter form can be assumed to have an answer, or so I contend.
I shall begin with a loose and intuitive sketch of my strategy, before spelling things
out in proper detail. This initial sketch is based on an analogy. Let us use the words
"collection" and "membership" in a deliberately unspecific way such that anything
that can be said to have members counts as a collection. Suppose now that we are
given some collection of web pages and instructed to design a new web page that
links to all and only the members of this collection. For which collections can the
instruction be carried out? In particular, might we design a new web page to link to
precisely the collection of web pages that don't link to themselves? The answer depends
on how the target collection is understood. First, suppose that the target is specified
intensionally, say by means of the concept of a web page that doesn't link to itself. Then
it is logically impossible to design a web page that links to all and only members of
that collection. For the new page would have to link to itself just in case it does not link
to itself. Second, suppose that the target collection is specified in a purely extensional
way, say by means of a comprehensive list of each and every web page that doesn't
link to itself. Then there is no logical or conceptual obstacle to designing a new web
page that links to all and only the items on this list. When we consider counterfactual
circumstances with more web pages, this list stays fixed, unlike an intensionally
specified target, whose extension may shift with the circumstances. In short, the fixity
of an extensionally specified target across counterfactual circumstances ensures that
it is possible to reach the target, while the corresponding shiftiness of an intensionally
specified target leaves us without any such guarantee.
Let us now consider Basic Law Vin a dynamic context, where the domain is allowed
to expand. The crucial question is which collections of objects can be used to form sets.
Consider the collection of objects that are not elements of themselves. Can we form a
set whose elements are the members of this collection? Again, the answer depends on
how the target collection is specified. Suppose the target is specified intensionally, say,
18 This reduction is closely related to the converse of (strict, full, and immediate) grounding. When
1/l
reduces to¢, then¢ has a potential to (strictly, fully, and immediately) ground 1/l. Cf. Section 1.7 and
pp. 43- 4, note 41.
by means of the condition x f/. x. Then the characterization of the desired set is logically
incoherent. We want a set r such that Vx(x E r ~ x ¢ x), but a simple theorem offirst-
order logic tell us that there is no such thing. Suppose instead that the target collection
is specified in a purely extensional way, say, as a (perhaps infinite) list that includes
each and every non-self-membered set in our domain. Then there is no logical or
conceptual obstacle to the formation of a set whose elements are precisely the items
on this list. All that the paradoxical reasoning shows is that the set cannot itself be on
the list and in this sense has to be "new': But this conclusion is unproblematic. There
is no logical or conceptual obstacle to the formation of "new" objects beyond those
that figure on some given list. 19
This approach ensures that every question about the "new" objects reduces to
one about the "old" ontology, as desired. Is this "new" set identical with that? The
answer turns on whether the two lists of"old" objects are coextensive. Likewise, every
question about membership in a "new" set reduces to an unproblematic question
about figuring on a list of "old" objects. This completes the initial sketch of my
account.
Let us now spell things out in more detail. The notion of a collection that can
be extensionally specified is clearly essential to my argument. When the domain is
allowed to expand, we need a way to track collections from the initial domain
into the expanded domain. This tracking is straightforward for collections that are
extensionally specified: these are tracked simply in terms of their members. But a list
is clearly not the only way to specify a collection in a purely extensional way. The linear
order of the list is immaterial. Suppose we wish to keep track of some web pages to
which a new web page is intended to link. Suppose we write each of the target URLs
on a separate piece of paper. It does not matter if the pieces are rearranged. Together,
the pieces still provide a purely extensional specification of the target collection.
A simple but useful way to talk about extensionally specified collections is provided
by the logic of plurals. Let me explain. In ordinary first-order logic we have singular
variables such as x and y, which can be bound by the existential and universal
quantifiers. In plural logic we have in addition plural variables such as xx and yy,
which can also be bound by existential and universal quantifiers to yield so-called
plural quantifiers. 20 '3xx' is read as 'there are some things xx such that .. :, and 'Vxx: as
'given any things xx, .. .' . These quantifiers are subject to inference rules analogous to
those of the ordinary singular quantifiers. There is also a two-place logical predicate
-<,where 'u -< xx' is to be read as 'u is one of xx:
19 This view of set formation traces its roots back to Cantor, whose distinction between "consistent" and
"inconsistent multiplicities" mirrors mine between collections specified in a purely extensional way and
ones that only admit of an intensional specification. See (Linnebo, 2013) for discussion.
20
Plural logic was made popular by George Boolos's classic (Boolos, 1984). For an introduction, see
(Linnebo, 20 l 7a).
Let us now use plural logic to develop our dynamic version of Basic Law V. Consider
the equivalence relation = that obtains between two pluralities xx and yy just in case
they comprise the same objects:
xx= yy ~ Vu(u -< xx ~ u -< yy)

Plural Law V, as I shall call it, abstracts on this relation:
(V) {xx} = {yy} ~ xx= yy
Let us refer to the abstracts {xx} and {yy} as sets. We would like to modify the law by
allowing the domain to expand-and thus also to avoid paradox. I shall now show
how this aim can be achieved by building on the account of abstraction canvassed in
Chapter 2. My explanation proceeds in two steps.21
The first step carries out the desired abstraction in a two-sorted language. We start
with an ordinary plural language that describes a single sort of objects on which we
want to abstract. This is our base language. Let the base domain be the plurality of
objects with which this language is concerned. We define an extended language that
adds a new sort of singular first-order variables and terms to the base language. This
new sort is reserved for sets obtained by abstraction on the base domain. There is
thus a separate register of variables to range over these sets, as well as an abstraction
operator {·} that applies to any plural term of the basic sort to yield a singular
term of the extended sort. For convenience, we let the extended sort have singular
quantifiers only, not plural quantifiers. In this two-sorted extended language, the
desired abstraction (V) is unproblematic. The extended sort is concerned with sets
of objects from the base domain. But these sets are located in a separate extended
domain over which only variables from the extended sort range. This separation of
the "new" sets from the "old" objects on which we abstract ensures that the paradox
is blocked. 22
The second step serves to take us back to a familiar one-sorted setting. The two
domains of the two-sorted language from step one thus need to be "merged" to form
a single domain over which a single sort of object variables can range. How can this
be done? The familiar law of extensionality suggests a simple answer. Consider an
element y of the extended domain, and let xx be its elements. If the base domain
already contains a set z whose elements are precisely xx, we identify y with z. If not, we
refrain from identifying x with any element of the base domain and instead regard x
as a "new" object, distinct from any member of the "old" domain. Clearly, the resulting
"merged" domain contains all the sets of objects from the base domain. As we are
now operating with a single domain, we can lift the type restrictions on the predicates
of the two-sorted language. The identity predicate is interpreted as real identity on
21
The steps receive further development and defense in Chapters 8 and 9, respectively.
22
In particular, it becomes impermissible to instantiate the quantifier ''Ix' in (3.1) with respect to the
Russell set r.
the merged domain. Some predicates of the two-sorted language are most naturally
stipulated to be true of exactly the same objects as before. For example, no new objects
should be allowed to satisfy the predicate 'electron' as a result of the domain expansion.
All that remains is to explain how the set abstraction operator {·} works in the
single-sorted setting. This operator is defined only on pluralities from the base
domain. Suppose we introduce a predicate 'Ow(x): true of all and only the objects
from this domain. We would like to restrict set abstraction to pluralities of "old"
objects. To achieve this, it is useful to disentangle two aspects of plural Law V. First,
the law provides a criterion of identity. When two pluralities form sets, the resulting
sets are subject to the law of extensionality:
(EXT) SET(uu,x) /\ SET(vv,y)-+ (x = y ++ Vz(z-< uu ++ z-< vv))
where SET(xx, y) means that xx form the set y. Second, the law provides a criterion of
existence according to which any plurality can be "collapsed" into a set:
(COLLAPSE) Vuu 3x SET(uu, x)
Together, the two criteria have the same effect as the original law. 23 The desired
restriction on set abstraction can now be obtained by weakening the criterion of
existence as follows:
Vuu(Vx(x-< uu-+ Ow(u))-+ 3xSET(uu,x))
In short, we have identified a theory of set abstraction that is based on the philosoph-
ical account of Chapter 2 and is easily seen to be consistent. 24
What about other forms of abstraction on pluralities? It is straightforward to
generalize the approach just outlined. 25 Consider an abstraction principle
§xx = §yy ++ xx ~ yy
where ~ is an equivalence relation on pluralities. Assume that the equivalence ~ is

unaffected by domain expansions: if aa ~ bb is true (or false) in one domain, then it
remains true (or false) in any extended domain. 26 (Such equivalences are said to be
stable; cf. Section 3.6.) As in the case of sets, we can now "factorize" the abstraction
principle into separate criteria of existence and identity:
23
In technical parlance: each approach can be interpreted in the other.
24
Consider any model M . Let M+ be the result of adding to the first-order domain of Meach set of
objects from said domain. Let 'Ow(u)' be true of all and only objects from the domain of Mand interpret
'SBT(uu, x)' in the intended way.
25 Of course, the purely mathematical value of these generalizations is limited, as we already have
set abstraction and thus also the resources to define equivalence classes, as is standard in contemporary
mathematics.
26
This is the "small qualification" advertised on p. 52. Clearly, this assumption is satified if ~ can be
defined by a formula all of whose quantifiers are restricted to the union of xx and yy, as is possible, for
example, in the case of plural Law V.
Vxx 3u ABs(xx, u)
ABs(xx, u) /\ ABs(yy, v) -+ (u = v ++ xx "' yy)
Finally, we restrict the criterion of existence to pluralities of "old" objects.
3.4 Static and Dynamic Abstraction Compared

How do the static and dynamic approaches to abstraction compare? One advantage
of the dynamic approach has already been mentioned: it solves the bad company
problem, which remains very serious for the static approach. When the domain is
allowed to expand, any form of abstraction on pluralities under a stable equivalence
relation is permissible, including abstraction on coextensionality, as in the case of
plural Law V.
A second advantage of the dynamic approach has to do with the reduction it
enables of any question about the "new" abstracts to a question about the "old"
ontology. This is a pleasing property which ensures that the "new" abstracts have
been properly defined. In particular, these reductions ensure that the structure of the
"new" abstracts is uniquely determined by the structure of the "old" ontology. This
unique determination contrasts favorably with the situation on the static approach,
where even abstraction principles that are typically regarded as acceptable can be
satisfied in a variety of non-isomorphic ways. In short, while dynamic abstraction
uniquely determines the structure of the "new" abstracts, given the structure of the
"old" ontology, static abstraction typically fails to determine a unique structure. 27
Third, the dynamic approach allows abstraction to be iterated, which turns out to be
an extremely powerful tool. The idea is simple. One application of plural Law V takes
us from some initial domain to a larger one. Since this larger domain gives rise to more
pluralities than the initial one, a second application of the law gives rise to even more
objects. We can continue in this way indefinitely. At limit stages, we take the union of
all the objects generated thus far. Since each round yields something new-as Russell's
paradox would otherwise re-emerge-the process never terminates.
In other respects, the dynamic approach appears more restrictive than its static
rival. Hume's Principle provides an example. In its familiar static form, this principle
supports Frege's celebrated "bootstrapping argument" for the existence of infinitely
many numbers, which figures at the heart of the proof of Frege's theorem. This argu-
27 One of the worst offenders is a principle known as New V, which states that two concepts have the
same extension just in case they are coextensive and have fewer instances than there are objects in the entire
universe. This abstraction principle has a plethora ofnon-isomorphic models; cf. (Jane and Uzquiano, 2004).
(Walsh and Eb els- Duggan, 2015) identifies a small family of principles that includes Hume's Principle and
enjoys at least a relative form of categoricity. But even this may not suffice. After all, Hume's Principle has
models of any infinite cardinality. So while this principle uniquely determines the structure of the cardinal
numbers given the size of the universe, it does not by itself determine said size and thus also not which
cardinal numbers there are.
ITERATED ABSTRACTION 61
ment relies on the assumption that the numbers are available in a single fixed domain
of objects. When we have established the existence of the numbers 0 through n, we
have at least n + 1 objects available to count, which enables us to establish that the
number n + 1 too is in the fixed domain. We thus prove, by mathematical induction,
that the fixed domain contains all the natural numbers. Consider now the dynamic
version of Hume's Principle. Suppose we start with a domain containing at least the
numbers from 0 to n. We can then establish the existence of the number n + 1. Since
we are proceeding dynamically, however, this number need not be in the domain with
which we began! Indeed, if we begin with a finite domain, then no finite number of
applications of Hume's Principle can establish the existence of an infinite domain.
Fortunately, this apparent limitation is easily overcome. Although the dynamic
version of Hume does not ensure the existence of a single infinite domain, it does
establish the existence of arbitrarily large finite domains. Given any finite domain,
an application of dynamic Hume yields an even larger such domain. So although
dynamic Hume does not ensure an actual infinity of numbers, it does establish a
potential infinity thereof. We shall shortly explain how this potential infinity can
be put to mathematical use. Moreover, it is possible to iterate the application of
Hume's Principle and other abstraction principles into the transfinite, thus ensuring
the existence of a variety of infinite domains (and thus actual infinities of numbers
and other objects).
There is a general lesson here. On the dynamic approach, each abstraction step
is small but correspondingly well understood. As observed, this modesty has many
philosophical advantages. The modesty means, however, that theories of dynamic
abstraction get most of their strength, not from an individual abstraction step, but
from what the theories say about possible iterations of such steps.
3.5 Iterated Abstraction

To harness the full power of dynamic abstraction, we need a way to represent and
reason about repeated application of abstraction steps. As an illustration, I now
explain how plural Law V can be iterated. The set abstraction operator is defined,
we recall, on all and only pluralities of "old" objects. I showed how the restriction
can be implemented by means of a predicate 'OLD'. But this implementation becomes
impractical when the abstraction step is iterated many times.
A more elegant option uses modal resources to represent the transition from one
interpretation of a language to a more inclusive one. Let us add to our language
the modal operators D and 0. We may think of 'Drp' as meaning 'no matter what
abstraction steps we carry out, it will remain the case that rp', and 'Orp' as 'we
can abstract so as to make it the case that rp'. Obviously, this interpretation of the
modal operators is different from the more familiar one in terms of metaphysical
modality. In the useful terminology of (Fine, 2006), the present interpretation is
"interpretational" rather than "circumstantial"; that is, it is concerned with how the
language is interpreted, not with how reality is. In particular, every interpretational
possibility is compatible with the metaphysically actual world.
How does our plural logic interact with these modalities? A natural answer is that
pluralities are modally rigid. A plurality is tracked from one possible world to another
in terms of its members. Thus, if xis one of some things yy, then necessarily so (at least
assuming the continued existence of the things in question). Likewise, if xis not one
of yy, then that holds of necessity (at least on the mentioned assumption). I believe
this is a plausible analysis of the interaction of plurals and non-epistemic modals in
natural language. 28 Regardless of the workings of natural language, however, I now
simply stipulate that pluralities are to be rigid with respect to the domain expansions
associated with my account of dynamic abstraction. It is clearly possible to track a
collection from one domain into another in terms of its members. I choose the logic
of plurals to do this job.
With our plural and modal resources available, we are ready to formalize claims
about dynamic abstraction. It is useful to base these claims, not directly on plural
Law V, but on our "factorization'' of the law into separate criteria of existence
and identity. Both "factors" can then straightforwardly be transposed to a dynamic
setting. 29 Consider the criterion of existence, (COLLAPSE), which states that every
plurality forms a set. In its current form, this criterion has a static character, because
the set is required to be in the very domain on which we abstract. But it is easy to
formulate a dynamic analogue, which states that necessarily, given any objects, it is
possible that these objects form a set:
(COLLAPSE<>) D'v'xx 03y SET(xx, y)
Unlike its non-modal analogue, this modal principle permits the set in question to be
located in an extended domain. We have thus transformed a principle asserting the
actual collapse of any plurality to a set into one asserting merely the possibility of such
collapse. We also need to transpose the criterion of identity to a dynamic setting. This
results in the statement that two sets are identical just in case they not only happen to
have the same elements but necessarily do so:
SET(uu,x) /\ SET(vv,y)-+ (x = y ++ D'v'z(z ~ uu ++ z ~ vv))
Observe that our transformation of static principles to dynamic ones rests on a

general idea, namely that the two quantifiers 'v' and 3 are replaced by the complex
strings D'v' and 03, respectively. We take a closer look at this idea in Section 3.6.
Have we achieved our aim of avoiding paradox by allowing the domain to expand?
Let us investigate! Consider first the static version of plural Law V. As expected, this
28 See (Rumfitt, 2005, Sect. VII), (Williamson, 2010, pp. 699-700), (Uzquiano, 2011), and (Linnebo,
2016b).
29
This approach is not obligatory. Another option is to introduce modal operators directly into an
abstraction principle to give it a dynamic character; see (Studd, 2016, esp. Sect. 5) but also (Payne, 2013b).
ITERATED ABSTRACTION 63
leads to paradox. We define a membership predicate by letting 'x E y' abbreviate the
claim that there are some objects that include x and whose set is y:
3uu(x-< uu /\ SET(uu,y))
This definition ensures the very reasonable conclusion that a set has the same
members as any plurality from which the set is obtained:
(3.2) SET(uu,y) -+ Yx(x E y ++ x-< uu)
The paradoxical argument is now immediate. By plural comprehension, there are

some objects rr that are all and only the objects that are not elements of themselves:
(3.3) Yu(u -< rr ++ u ¢ u)
By (COLLAPSE), these objects rr form a Russell set r; that is, SET(rr, r). By universal
instantiation, (3.3) entails
(3.4) r-< rr ++ r ¢ r.
Utilizing (3.2), this entails r -< rr ++ r -f:. rr, which is a contradiction.
Pleasingly, the dynamic version of plural Law V escapes the paradox. Given only
the potential collapse of pluralities to sets, it follows only that it is possible for there
to be a Russell set r such that SET(rr, r). Thus, r no longer needs to be in the range of
the quantifier 'Yu' in (3.3). (Indeed, on pain of paradox, r cannot be.) This blocks the
application of universal instantiation that was used to obtain (3.4). Nor can paradox
be obtained in any other way, as there are simple Kripke models that establish the
consistency of our dynamic version of plural Law V. Moreover, there is nothing special
about plural Law V in this regard. Our account extends straightforwardly to any
other form of abstraction on pluralities under a stable equivalence relation. As before,
we begin by "factorizing" the abstraction principle into separate criteria of existence
and of identity. We then transpose these criteria to our dynamic setting by replacing
Y and 3 with DY and 0 3, respectively. The result of this two-step transformation can
be shown to be consistent. 30
Let us take stock. We began with plural Law V, which is a form of naive set theory.
We showed how to transpose the law to a dynamic setting, which removes the paradox
and thus extracts a sound core from the shipwreck of naive set theory. Then, we
introduced a modal language in which we can describe and theorize about iterated
applications of the law. Looking ahead, Chapter 12 shows how the sound core of naive
set theory can be extended in a natural and well-motivated way to a modal set theory
that is strong enough to interpret all of ordinary set theory, ZF. The "principal engine
of set production" in this modal set theory is our intuitively appealing principle of
potential collapse, namely (CoLLAPSE 0 ). 31
30 (Studd, 2016) provides a related development of this form of abstraction.

31
This apt characterization is borrowed from (Yablo, 2006, p. 151), who uses it in a related context.
3.6 Absolute Generality Retrieved

As adumbrated, our modal language provides two different ways to generalize over
objects and other entities. The most straightforward way is by means of the ordinary
quantifiers V and 3. Since the variables of our system range just over the domain of
the relevant possible world, we may think of this as an intraworld form of generality.
There is also a transworld form of generality, which is expressed by the strings DV and
03. Let us refer to these complex strings as modalized quantifiers. These strings use
an additional form of generalization that our account of dynamic abstraction makes
available. By a step of abstraction, we can change the interpretation of our language,
with the result that the language comes to be concerned with a larger domain. The
result is that the modalized quantifiers generalize not just over all entities at the
relevant world but over all entities at all possible worlds. While the generality expressed
by the ordinary quantifiers is merely world-relative, the generality expressed by the
modalized quantifiers is intended to be absolute.
The utility of the modalized quantifiers was illustrated in Section 3.5. Suppose we
are interested in absolutely all sets, not only those that are available at some particular
stage of the process of set abstraction. We can express the desired claims using the
modalized quantifiers, as we did, for example, with the principle of potential collapse,
(COLLAPSE 0 ). Without the modalized quantifiers, we would be unable to generalize
over absolutely all sets (or all abstracts more generally), irrespective of the stage at
which they are "formed''. In particular, the modalized quantifiers are crucial to our
transformation of static principles to dynamic ones. The transformation replaces each
ordinary quantifier with the corresponding modalized quantifier.
Let us spell things out in more detail. For any non-modal language .C, let .c 0 be the
modal language that results from adding the two modal operators D and 0. For each
non-modal formula rp, let its potentialist translation, rp <> , be the result of replacing
each ordinary quantifier with the corresponding modalized quantifier. Let us say that
a formula is fully modalized just in case all of its quantifiers are modalized.
We need a better understanding of the important form of generalization afforded
by the modalized quantifiers. In particular, what are the logical principles that govern
these generalizations? The question receives a satisfying answer in the form of a
theorem stating that, given some plausible background assumptions, the first-order
modalized quantifiers behave logically precisely like ordinary first-order quantifiers,
only that they range over absolutely all objects that can be introduced by abstraction.
Much the same goes for higher-order quantification-though with some interesting
exceptions to be discussed in Section 3.7.
To state the background assumptions, we need some definitions. Let us say that a
formula rp is stable if the necessitations of the universal closures of the following two
conditionals hold:
<P ---+ Drp

ABSOLUTE GENERALITY RETRIEVED 65
Intuitively, a formula is stable just in case it never "changes its mind': in the sense that,
if the formula is true (or false) of certain objects at some world, it remains true (or
false) of these objects at all "later" worlds as well.
Next, recall that the modal logic S4 is sound and complete with respect to Kripke
frames that are reflexive and transitive. This ensures that S4 is appropriate for the
interpretational modality introduced in this chapter. In fact, a slightly stronger modal
logic is appropriate as well, namely a system known as S4.2, which adds to S4 the
following axiom:
(G)
This claim is defended in detail in Section 12.2, although the gist of the argument is
easily explained. Axiom G is sound for all convergent frames, that is, for frames where
any two extensions of a single world themselves have a common extension: 32
wo :::: W1 I\ Wo :::: Wz ~ 3w3(W1 :::: W3 I\ Wz :::: w3)
The property of convergence holds because abstraction is cumulative. Consider two

permissible forms of abstraction. If we choose to exercise only one of the forms, this
has no lasting effects, as we can always exercise the other form later, thus arriving at
the same domain as we would have reached by carrying out the abstractions in the
reverse order.
We can now state the advertised theorem. Let f- be the relation of classical
deducibility in a non-modal first-order language£. Let f- <> be deducibility inc<> by f-,
S4.2, and axioms asserting the stability of all atomic predicates of£,.
Theorem 3.1 (Mirroring) Let <Pi, ... , <Pn and 1" be formulas of£. Then we have:
0
<PI»··•<Pn f- ,/,
'I'
i"ff <P1<> ,. . .,cpn<> f- ,/,
'I'
<>
(See Appendix 3.B for a proof.)
The mirroring theorem, as we shall call it, has a simple moral. Suppose we are
interested in the logical relations between some fully modalized formulas in a context
that satisfies the background assumptions. Then we may delete all the modal operators
and proceed by means of the corresponding non-modal theory logic. In other words,
under the mentioned assumptions, the complex strings D\t' and 03 behave logically
precisely like the ordinary quantifiers-except that they generalize across all possible
worlds rather than within a single world.
Of course, for the theorem to be available in our present context, its assumptions
must hold. The availability of S4.2 was discussed above. What about the stability
32
Such frames are sometimes called "directed''. But strictly speaking, directedness is the stronger
condition that any two worlds-whether or not they are both extensions of some single world- have a
common extension. In our context, however, the difference between convergence and directedness does
not matter much, as we can always assume that the entire collection of worlds has a single minimal element.
axioms? The stability of identity is widely granted. The stability of the plural member-
ship relation 'x -< uu' holds by stipulation, as discussed in Section 3.3. Finally, it is
plausible to require SET(xx, y) and Ass(xx, y) to be stable. When y is specified as the
set (or some other abstract) of xx, it remains so specified regardless of what other
objects we later add to the domain.
3.7 Extensional vs. Intensional Domains

Is my position stable? It might appear not. On the one hand, I argued that every
domain of the ordinary quantifiers can be expanded. It is always possible to shift the
interpretation of our language in such a way that we can refer to an object that is not in
the domain associated with the previous interpretation. On the other hand, I argued
that a form of absolute generality is achieved by means of the modalized quantifiers.
Yet how can absolute generality be possible if every domain can be expanded?
In fact, the apparent tension dissolves once we take a more careful look at the argu-
ment for domain expansion. The argument uses plural Law V and relies essentially
on the fact that a plurality is a purely extensional collection. Given any plurality,
this pure extensionality enables us to make good conceptual and mathematical sense
of a set whose members are given by the plurality-albeit a set that may be located
in some expanded domain. Suppose our language-and thus also our quantifiers
V and 3-has some interpretation. To show that the domain can be expanded,
observe that:
(3.5) 3rr"lu(u -< rr B- u rf_ u)
That is, there is a Russell plurality rr of all the objects that are not elements of
themselves. We now abstract on rr so as to introduce the corresponding set r = {rr}.
Let v+ and 3+ be the quantifiers as interpreted after this abstraction. The familiar
Russellian reasoning shows that r cannot be in the range of the original quantifiers.
So we have
-.3x SET(rr, x) /\ 3+ x SET(rr, x)
This means that the domain has indeed expanded. The post-abstraction domain
contains a set that was not present in the pre-abstraction domain. 33
So far, so good. The worry is that the argument just sketched will extend to the
modalized quantifiers as well, thus showing that their range too is non-absolute,
contrary to what I claimed. So let us investigate the argument analogous to the one
just given but where the modalized quantifiers play the role previously played by the
ordinary quantifiers. This analogous argument needs the modalized analogue of (3.5),
namely
33 Notice, incidentally, that the argument does not require that the domains be reined, which would
be potentially controversial. See e.g. (Cartwright, 1994). While it is heuristically convenient to talk about
domains, the argument does not require us to treat the domain as some sort of entity, let alone an object.
EXTENSIONAL VS. INTENSIONAL DOMAINS 67
03rr D'v'u(u -< rr ++ u ¢ u)
That is, the argument assumes that it is possible for there to be some objects rr such
that no matter what objects are introduced into the domain, rr are all and only the
non-self-membered objects. Is this assumption correct? Consider the two conditions
that flank the biconditional and are said to be necessarily coextensive. Now, every
set is non-self-membered. As more and more sets are introduced, the condition on
the right-hand side will therefore be true of more and more objects. By contrast, a
plurality has an extensional nature and thus includes the very same objects at every
world at which it exists. So whichever objects rr we consider, the condition on the
left-hand side is true of precisely these objects. This condition cannot become true of
more objects as the domain expands, unlike the condition on the right-hand side with
which it is said to be necessarily coextensive. It follows that (3.5° ) is unacceptable.
And without this premise, the argument for domain expansion breaks down. In short,
while the argument works for the ordinary quantifiers, which range over a domain that
is extensionally specified (by means of a plurality of objects), it does not apply to the
modalized quantifiers, whose domain only admits of an intensional specification.
To deepen our understanding of the situation, recall the mirroring theorem, which
states that the modalized first-order quantifiers behave logically precisely like ordin-
ary first-order quantifiers, given some plausible background assumptions. For any
valid first-order argument, the analogous argument that uses modalized quantifiers
in place of ordinary quantifiers is thus valid as well. This is precisely the sort of shift
from ordinary to modalized quantifiers that we have investigated in connection with
the argument for domain expansion-and found to be problematic. So let us explore
to what extent the mirroring theorem extends to plural logic. A crucial part of plural
logic is the so-called plural comprehension scheme:
(P-Comp) 3.xx'v'u(u-< xx++ cp(u))
where as usual cp cannot contain xx free. This axiom scheme states that any condition
cp( u) defines some objects, namely all the objects that satisfy the condition. 34
Notice that this comprehension scheme allows for an empty plurality. Here we
deviate slightly from plural locutions in natural language, where a plurality must
consist of at least one and possibly two objects. This minor deviation can be justified
pragmatically. 35 Allowing an empty plurality is very convenient for the purposes of
formalizing set theory. And even if plural locutions in natural language are in fact not
used in this way, there might have been natural languages where they were.
Let f- be the relation of deducibility in £ based on classical first-order logic, and the
standard introduction and elimination rules for the plural quantifiers, but without
34
The condition may contain additional free variables that are not displayed .
35 Here I follow (Burgess and Rosen, 1997) and (Burgess, 2004).
any plural comprehension axioms, which we set aside for separate treatment. 36 It is
straightforward to extend the mirroring theorem to this restricted plural logic. Any
obstacle to extending the theorem to full plural logic must derive from the plural
comprehension axioms. So let us examine the plausibility of the plural comprehension
axioms-on both of our two understandings of the quantifiers. Consider first the
ordinary quantifiers, which function as devices of intraworld generalization. Under
what conditions does a formula <{J(u) define a plurality xx? The answer is easy and
uncontroversial: always!
Consider next the modalized quantifiers, which function as devices of transworld
generalization. To explicate our question, we apply the potentialist translation to
(P-Comp), which yields:
(Comp 0 ) 03xxD'v'u(u-< xx++ <P 0 (u))
Our question is thus: Which are the acceptable instances of (Comp 0 )? Unlike its
non-modalized analogue, this scheme has false instances, for example, u rf- u. But
a more general answer is possible. An instance of the modalized comprehension
scheme is acceptable just in case there is some stage in the process of dynamic
abstraction at which all possible instances of the comprehension formula <P 0 ( u) have
been introduced.
We can now give a more thorough explanation of why the argument for domain
expansion, which is so effective for the ordinary quantifiers, cannot be extended
to the modalized quantifiers. The argument turns on an instance of plural com-
prehension, namely (3.5), whose modalized analogue is false. Moreover, the fail-
ure of this modalized comprehension axiom is not an isolated case but part of a
general phenomenon. While ordinary plural comprehension is unproblematic, its
modalized analogue conflicts with the extensional nature of pluralities, which has
figured centrally in our discussion throughout this chapter. It is precisely because
of their extensional nature that it is permissible to abstract on pluralities, including
under the relation of coextensionality, as in plural Law V. But this extensional nature
also entails that plural comprehension needs to be restricted when the quantifiers
are modalized. This is a happy situation. What drives one part of the expansion
argument-namely, the extensional nature of the collection on which we abstract-
simultaneously explains why the second part becomes unavailable when the quanti-
fiers are modalized.
I wish to end with some more general remarks. Our discussion in this section
reveals that the application of a precise intension to the world may fail to determine
an extension-or a plurality, which is my canonical way of representing extensions.
36
For instance, the deducibility relation can be given by the logic PFO of Definition 12.2 minus the
plural comprehension scheme.
EXTENSIONAL VS. INTENSIONAL DOMAINS 69
In other words, to determine an extension (or plurality) is not a trivial matter. This
contrasts with much of our ordinary non-mathematical experience. Suppose we have
formulated a perfectly precise notion of a star. For any object whatsoever, this notion
enables a definitive verdict as to whether or not the object is a star. When this precise
intension is applied to the world, reality answers with a determinate extension, namely
the plurality of objects that satisfy the intension. And there is nothing unusual about
stars in this regard. In most ordinary empirical cases, a precise intension determines
an extension when applied to the world. But in mathematical cases, and other cases
involving abstraction, this is no longer so. Here a precise intension often fails to
determine an extension.
This observation suggests an analysis ofDummett's intriguing, but obscure, notion
of indefinite extensibility. 37 Dummett characterizes a concept as indefinitely exten-
sible just in case, whenever
we can form a definite conception of a totality all of whose members fall under that concept,
we can, by reference to that totality characterize a larger totality all of whose members fall
under it. (Dummett, 1993, p. 441)
Let X range over "totalities", and let D(X) mean that X is definite. A concept is then
indefinitely extensible, according to Dummett, just in case:
(IE) VX(D(X) /\ Vu(u E X---+ Fu) ---+ 3y(Fy /\ y ¢ X))
Frustratingly, Dummett never properly explains how he understands the crucial

notion of definiteness. This has hampered later discussions of indefinite extensibility.
This chapter suggests an attractive way to provide this missing definition. Perhaps
a definite totality is simply a collection that admits of an extensional specification by
means of a plurality. This yields the following analysis of indefinite extensibility:
DVxx(Vy(y-< xx---+ Py) ---+ 03yFy /\ y -/. xx)
That is, a concept F is indefinitely extensible just in case: necessarily, given any
instances of the concept, possibly there is another instance that is not among the
mentioned instances. In short, a concept is indefinitely extensible just in case it is
impossible to specify its extension by means of a plurality, because any proposed
extension makes it possible to define yet further instances of the concept.
We have found the concepts of a set and of an object in general to be indefinitely
extensible, as assessed by this definition. A great variety of other concepts too can be
brought into the fold by extending our account of dynamic abstraction to the relevant
type of object, for example, the concepts of an ordinal and a cardinal number.
37
See (Dummett, 1963), (Dummett, 198la, pp. 528-41), and (Dummett, 199la, esp. pp. 312-21) .
Appendices
3.A Further Questions
I wish briefly to discuss some questions that remain to be explored.
3.A.l THE HIGHER - ORDER NEEDS OF SEMANTICS
One of my more distinctive claims is that every plurality potentially forms a set. This means
that there are fewer pluralities than one might naively have thought. The orthodox view is that
there are pluralities that are not exhausted by any given level of the cumulative hierarchy; for
instance, the plurality of every set whatsoever. My view leaves no room for such pluralities, as
it entails that every plurality is exhausted by some possible level of the hierarchy.
Some philosophers will see this as a loss. They will object that my rejection of proper-
class-sized pluralities undercuts some of the most useful applications of plural logic. In
particular, plural logic has been used to analyze talk of proper classes and to develop the
semantics of first-order set theory where the first-order quantifiers are allowed to keep their
intended range over absolutely all sets. 38 These applications rely crucially on our ability to
represent the domain by means of a plurality consisting of absolutely all sets. By disallowing
such pluralities, I reject what some will regard as the most useful and exciting applications of
plural logic.
I grant that such applications are disallowed but deny that this represents any loss. For one
thing, you cannot lose something you never had. And this was never a permissible application
of plural logic. The extensional nature of pluralities makes them unsuited to represent the entire
hierarchy of sets. A plurality consists of a fixed range of objects. But the cumulative hierarchy is
open-ended and thus resists being summed up by a fixed range of objects. For another thing, the
job in question is better and more naturally done by means of second-order logic-understood
as a theory of quantification into predicate position-rather than plural logic. The semantic
contribution of predicates is intensional in a way that makes them well suited to represent the
entire domain of set theory or the semantic value of the membership predicate. It is useful to
think of the semantic contribution of a predicate as a Fregean concept. My proposal is thus that
absolutely infinite collections, such as that of all sets or all ordinals, be represented by means of
Fregean concepts rather than the associated pluralities, whose existence I deny.
3.A.2 ABSTRACTION ON INTENSIONAL ENTITIES
My account of dynamic abstraction has been applied to abstraction on pluralities. The exten-
sional nature of pluralities ensures that each plurality is completed by any stage at which it
is available for abstraction. And this, in turn, enables us to reduce any question about the
"new" abstract to a question about the "old" objects that made up the plurality in question.
Can my account be extended to abstraction on Fregean concepts? Because of their intensional
nature, a Fregean concept cannot be assumed to be completed by a stage at which it is available
for abstraction, or indeed at any stage whatsoever. This makes it far harder to ensure that any
questions about the "new" abstracts can be reduced to questions about the "old" ontology.
38 For the former application, see (Uzquiano, 2003). Concerning the latter, I have in mind the project
initiated by (Boolos, 1985).

APPENDICES 71
The obvious test case is the conceptual version of Basic Law V, which abstracts on Fregean
concepts rather than pluralities, as examined above. Let us apply the same strategy to this
version of the law as we did to the plural version: first, we "factorize" it into distinct criteria of
existence and of identity; then, each criterion is transposed to the dynamic setting. The criterion
of existence gives rise to the principle that necessarily any concept possibly defines a property:
D'v'F 03y PPTY(F,y)
Let us call this potential conceptual collapse. When traµsposed to the dynamic setting, the
criterion of identity becomes:
PPTY(F,x) A PPTY(G,y) --+ (x = y ++ D'v'u(Fu ++ Gu))

A deep difference between the conceptual and the plural case now emerges. Since pluralities
are complete by any stage at which they are available for abstraction, a question about the
identity of two "new" sets reduces to a question about the coextensionality of two pluralities
of "old" objects. A question about the identity of two "new" properties, by contrast, "reduces"
to a question about the necessary coextensionality of two concepts. Yet it is doubtful this is a
genuine reduction. To determine whether the two concepts are necessarily coextensional, we
need to consider not only "old" objects but also all "new" objects that we may go on to introduce.
Fortunately, the problem can be solved. We can restrict the second-order comprehension
scheme so as to ensure that each question about the ensuing concepts, as well as their
properties, reduces to some simpler question. This means we must carefully consider under
what conditions a formula rp(u) should be allowed to define a concept F, in the sense that:
0 3FD'v'u(Fu ++ rp 0 (u))
It is natural to draw on the idea that the individuation of concepts and properties must be
grounded, that is, that it must only make use of entities that are already available. The behavior
of the defining condition rp 0 (u) may only "depend on" entities that are already available. This
idea obviously needs to be spelled out. One way to do so is developed in my (Linnebo, 2009a) .39
This yields an interesting theory of concepts and properties, which can do important work in
semantic theories. No doubt other, probably superior, options are possible as well.
3.A.3 THE NEED FOR A BIMODAL LOGIC
The modality employed in this chapter is not "ordinary" metaphysical modality. It is an

interpretational modality, which is understood in terms of the interpretation shifts-and the
concomitant domain expansions-associated with my account of abstraction. How does this
modality interact with ordinary metaphysical modality? The answer should ideally take the
form of a bimodal logic that governs the two modalities.
While I have no such logic to offer, some of its principles are clear. To avoid confusion, let
us temporarily write ordinary boxes and diamonds for the metaphysical operators and black
boxes and diamonds for the interpretational ones. The two modalities represent two dimensions
that can be varied: the metaphysical circumstances and the interpretation of the language (and
thus also the mathematical ontology). It is useful to picture these two dimensions as laid out
along an x-axis and a y-axis, respectively. Since each dimension can be varied while holding
See also (Fine, 2005a) for a rela~ed account.

39
the other one fixed, two distinct "moves" are available. A horizontal move (along the x-axis)
leaves unchanged the interpretation (and thus also the purely mathematical ontology) while
varying the metaphysical circumstances. As usual, we take any purely mathematical truth
to hold of metaphysical necessity. For any purely mathematical <p, we thus have rp ---+ D<p.
By contrast, a vertical move (along the y-axis) leaves the circumstances unchanged while
varying the interpretation (and thus also the mathematical ontology). Thus, for any statement
rp concerned only with entities already introduced, we have <p ---+ •<p. Most ordinary non-
mathematical truths, for example, cannot be altered by our abstraction-induced shifts of the
interpretation of the language.
The availability of these two seemingly independent moves suggests that the correct bimodal
logic is a simple "product logic': which is characterized by the following commutativity
principles:
Alas, things are not so simple! These principles are threatened by counterexamples, as is
most easily explained by indulging in talk about possible worlds. The first principle says that
any world that can be reached by first making a horizontal move (along the metaphysical
accessibility relation R) and then a vertical one (along the interpretational accessibility relation
~), can also be reached by first making a vertical move and then a horizontal one, and vice
versa. We can represent this principle as follows (where solid arrows represent the given
relations, while dotted arrows represent the relations asserted to exist on the basis of the
given ones):
WQ~WJ
But this principle is problematic. Consider some world w0 . Let w 1 be a metaphysical alternative
to wo at which there are two indiscernible objects a and b that do not exist at wo. So at w1 it is an
interpretational possibility to introduce the singleton of one of these indiscernible objects but
not the other; let this possibility be witnessed by a world w2 . However, it appears impossible to
get from wo to w2 by first taking a vertical step along the interpretational accessibility relation
and then a horizontal step along the metaphysical accessibility relation: for without a and
b available, we do not have the resources needed to specify an interpretation that generates
the singleton of one but not the other. Thus, it is doubtful we are entitled to the simple
product logic.
I leave as a challenge to the reader the task of formulating a more appropriate bimodal logic.40
40
The challenge should not worry us too much, however. For as James Studd pointed out there is no
obvious threat to the product logic if we make the relatively unproblematic maximality assumption that
we always introduce all of the objects that we are entitled to introduce (cf. Section 12.2). This assumption
APPENDICES 73
3.A.4 THE CORRECT PROPOSITIONAL LOGIC

Let us focus, in what follows, on the interpretational modality, setting aside the metaphysical
one. We have assumed that the modal logic governing the former modality is a classical version
of quantified S4.2. The choice of the modal logic S4.2 is defended in Section 12.2. But are
we entitled to let the propositional logic of this modal system be classical, as opposed to, say,
intuitionistic? On reflection, this is not obvious. The modal operators enable us to generalize
over an indefinitely extensible domain. But there is a famous argument to the effect that the
correct logic for such generalizations is intuitionistic, not classical. 41
The argument is not always easy to follow. So let me canvass a relatively free reconstruction.42
Consider a true generalization over some indefinitely extensible domain. What explains, or
grounds, this truth? Since there is no definite or fixed totality of instances to consider, it is not
an option to ground the generalization in the totality of its instances. So what else might explain
a universal generalization, if not its instances? We need some non-instance-based explanation
of true universal generalizations.
The intuitionistic conception of mathematics suggests one answer: the truth of the gener-
alization can be understood in terms of the existence-or perhaps even our possession-of a
proof of it. As is well known, this anti-realist conception validates intuitionistic logic. 43 But
needless to say, the radical anti-realism on which this answer relies is extremely controversial.
The appeal to anti-realism is not obligatory, however, as there turn out to be other ways too
to provide the desired non-instance-based explanations. The following passage from Hermann
Wey! provides some inspiration. Discussing whether there is a natural number that has some
decidable property P, Wey! writes:
Only the finding that has actually occurred of a determinate number with the property P can give
a justification for the answer "Yes;' and-since I cannot run a test through all numbers-only
the insight, that it lies in the essence of number to have the property P, can give a justification
for the answer "No"; Even for God no other ground for decision is available.
(Wey!, 1921, p. 54 (trans. p. 97); emphasis in original)
This is a brilliant suggestion. The truth of a generalization need not be grounded in the totality
of its instances but can also be explained in terms of the essences of the concepts that figure
in the generalization. For example, the universal quantification \fn-,P(n) can also be explained
by its "lying in the essence of" the concept of a natural number that every natural number is a
non-P.
The suggestion is fully compatible with realism. There may well be robust truths about
concepts and their essential connections, which need not be understood in terms of proof.
ensures that the interpretational possibilities are well-ordered, indexed by the number of abstraction steps
that have been carried out. In particular, when (a} is introduced, we must also, by maximality, introduce (b}.
41
See (Dummett, 1963), (Dummett, 199la, ch. 24), and (Dummett, 1993). See also (Santos, 2013) for a
useful overview of such arguments and exposition of their roots in Brouwer.
42 See (Linnebo and Shapiro, 2017) for a fuller discussion.
43
This can be seen, for example, by considering the Brouwer-Heyting-Kolmogorov interpretation of
the logical operators.
Even when understood in a realist manner, however, the suggestion may require a departure
from classical logic. Indeed, a development of the suggestion that I find attractive turns out to
validate only semi-intuitionistic logic, that is, intuitionistic logic where quantification restricted
to any set behaves classically. 44
3.B Proof of the Mirroring Theorem

We begin by proving a form of modal collapse for fully modalized formulas.
Lemma 3.1 Let <p be a fully modalized formula of a modal language .c 0 . Then S4.2 and the
stability axioms for .c 0 prove that O <p, <p, and O<p are equivalent and thus in particular that <p
is stable.
This is a satisfying result. When a formula is fully modalized, it doesn't matter at which world
it is evaluated.
Sketch ofproofofLemma 3.1. Assume <pis fully modalized. Since we are working in an extension
of the modal logic T, it suffices to prove O <p --+ O<p, which we do by induction on the complexity
of <p. If <pis atomic, the stability axioms allow us to prove our target O<p-+ O<p. If <pis -.1/f,
then our target follows straightforwardly from the induction hypothesis applied to 1/f. If <p is
1/11/\1/12, then O<p implies 01/f 1 /\ 01/f2, which by the induction hypothesis implies 01/1 1 /\ 01/f2 ,
which in turn implies O<p. Assume finally that <p is of the form 03x 1/f; the case of the plural
quantifier is analogous. Since we work in an extension of S4, we have O <p,--+ <p for <p of this
form. So it suffices to prove <p--+ O<p. By our induction hypothesis, we know that <pis equivalent
to 03x 01/f. Observe next that the Converse Barcan Formula yields Vx03y(x = y) and thus
also 3x01/f -+ 03x 1/f. The formula 03x 01/f thus implies 003x1/f. By (G), the latter formula
implies 003x 1/1, which is just O<p, as desired. -l
Proof of Theorem 3.1. The proof goes by induction on the proofs. We start with the left-
to-right direction. The only hard cases are the introduction and elimination rules for the
quantifiers. I will outline the case of the first-order universal quantifier; the other cases are
analogous. We begin with the rule UI. Assume we have <{11> • • • , <fin I- Vx 1/1 and conclude by
UI that <{11> •• • , <fin I- 1/l(t) for some suitable term t. By the induction hypothesis we have
 . .. , IPn I- 1/f (t), where t does not occur free
in any of the <p;, and conclude by UG that <{11> • • • , IPn I- Vx 1/1. By the induction hypothesis we
have<pf, .. . ,<p;/ 1- 0 1/f 0 (t),fromwhichUGgivesusip?, ... ,<p;/ 1- 0 Vx1/f 0 . Astandard
trick available in S4 then gives us O<p?, .. . ,O<p;/ 1- 0 OVxl/1 °. Lemma 3.1 then gives us
<p?, .. . , <p;/ 1- 0 OVx 1/f 0 , as desired.
The right-to-left direction is easily established by adopting a Hilbert-style axiomatic

approach to the relevant modal logic. Consider the operation ip 1-+ <p- of deleting all
modal operators. This operation maps every axiom of our modal logic to a theorem of the
corresponding non-modal logic and correlates every inference rule of the former with a
legitimate inference of the latter. The right-to-left direction then follows from the observation
that (<p 0 )- = <p. -l
44
See (Linnebo, 2017b).
PART II
Comparisons
4
Abstraction and the Question
of Symmetry
4.1 Introduction
Consider the biconditional that figures in an abstraction principle:
(AP)
For at least some such principles, Frege held that the two statements flanking the
biconditional are merely different ways to "carve up" one and the same meaning or
content. Here is one expression of this view:1
Now in order to get, for example, from parallelism to the concept of direction, let us try the
following definition: The sentence
"line a is parallel to line b"
is to mean the same as
"the direction ofline a is identical with the direction ofline b" (Frege, 1953, §65)
The distinctive feature of this view is that two sentences which clearly differ in their
logico-syntactic structure are nevertheless said to "mean the same" or have the same
content.
The idea of recarving of content has proved hard to pin down, however. 2 The central
problem is to find a notion of content with just the right fineness of grain. On the
one hand, the content needs to be coarse-grained enough to ensure that appropriate
abstraction principles are devices for the recarving of content. Let 1/1 state that
 a ,...., fJ

Let us call this the requirement of abstractionist recarving. On the other hand, the
idea of recarving of content is meant to enable certain philosophical explanations
that require the content to be fairly fine-grained. Suppose that the two sides of
1 Another example, from §64, was quoted and briefly discussed in Section 1.3.
2
See (Dummett, 199la), (Fine, 2002), and (Hale, 1997) for some useful discussions.
78 ABSTRACTION AND THE QUESTION OF SYMMETRY
an abstraction principle are recarvings of a shared content. We want to use the

comparatively unproblematic right-hand side of the abstraction principle to shed light
on the seemingly problematic left-hand side; in particular, we want the epistemic
and explanatory status of the former to be transmitted to the latter. Let us call
this the transmission requirement. For the transmission requirement to be plausible,
the contents in question must be fairly fine-grained. In short, we face a Goldilocks
problem. Is there a notion of content with granularity that is just right, neither too
coarse nor too fine?
Let us consider some conceptions of content that philosophers have explored. At
the extreme coarse-grained end of the spectrum we find Frege's notion of reference or
Bedeutung. If this notion of content is used, then recarving of content is exceedingly
easy to come by. Any two materially equivalent sentences qualify as recarvings of a
single content; in particular, (AP~) collapses into (AP). But epistemic and explana-
tory status are obviously not transmitted across the relation of material equivalence.
A somewhat less coarse-grained option is content understood as sets of metaphys-
ically possible worlds. Sameness of content is then a matter of necessary equivalence,
which makes it likely that appropriate instances of (AP*>) come out true. Yet this
notion of content is still too coarse-grained to ensure that epistemic and explanatory
status are transmitted along the recarving relation. Necessary equivalence does not
ensure sameness of epistemic or explanatory status (cf. p. 15). So let us consider
some more fine-grained alternatives. Two popular options are Russellian propositions
and structured Fregean senses. On these conceptions of content, each content has an
internal articulation that is isomorphic to the logico-semantic structure of any of its
linguistic representations. It follows that two sentences with different logico-semantic
structure cannot express the same content. In particular, since the two sides of any
abstraction principle differ in logico-semantic structure, these principles cannot effect
a recarving of content.
The neo-Fregeans have attempted to solve the Goldilocks problem. But their
proposals became increasingly baroque and continued to encounter difficulties. 3 And,
more importantly, the neo-Fregeans now favor an approach where content recarving
no longer plays an important role. 4
In this chapter, I consider some fresh approaches to the idea of content recarving.
I begin with a seductive argument that promises an easy solution to our problem
but whose charms must be resisted. Then I discuss a promising recent attempt by
Agustin Rayo to solve the Goldilocks problem by means of a notion of content
concerned with a sentence's "demands on the world''. Using this notion of content,
3 For the proposal. see (Hale, 1997), which is reprinted with an important postscript in (Hale and Wright,
200la). For criticism, see (Potter and Smiley, 2001 ), (Potter and Smiley, 2002), and (Fine, 2002, pp. 39-41).
Some aspects of the debate are presented in Appendix 7.A.
4 See especially (Hale and Wright, 2000).
IDENTITY OF CONTENT 79
Rayo defends a highly symmetric conception of abstraction, according to which the

two sides of appropriate abstraction principles make precisely the same demands on
the world. I end by rejecting this symmetric conception in favor of an asymmetric
alternative. There are important "worldly" differences between the two sides of
abstraction principles: the right-hand side is explanatorily prior to the left-hand side;
the two sides are concerned with different objects; and they differ in their ontological
commitments.
These asymmetries reinforce an argument from Section 1. 7 to the effect that the
asymmetric conception provides a far more direct and appealing account of de re
sufficiency or mutual sufficiency statements. That argument will be set aside in the
present chapter in order to focus on the other objection that we are about to develop.
4.2 Identity of Content

Suppose a notion of content has been defined. As usual, let cp ¢> 1/t express that cp
and 1/t have the very same content. We may think of this as a form of identity:
rp and 1/t are two representations of one and the same content. This is clearly not a
case of ordinary identity, which is a relation defined on all and only objects. It is an
extension of the ordinary notion of identity. Where the ordinary identity predicate
can meaningfully be flanked by all and only singular terms, the operator ¢> can
meaningfully be flanked by all and only formulas. In important recent work, Rayo
endeavors to make sense of such generalizations of identity. He introduces analogues
of the identity sign that can meaningfully be flanked by predicates and sentences and
put to substantive philosophical use in an account of content, modality, ontology, and
their interrelations. 5
The logical properties of ordinary identity are simple and well-known. It is an
equivalence relation that supports Leibniz's Law:
(LL) t = t' ~ (8 ~ 8[t' /t))

where 8[t' /t] is the result of uniformly replacing t with t' in 8. As usual, two
restrictions must be imposed: first, the terms t ~d t' must be purely referential
(unlike, for example, definite descriptions}; and second, the context 8( .. .) must be
transparent (unlike quotational and propositional attitude contexts). 6
5 The most explicit formulation of the task is found in (Rayo, 2016), though the idea of a generalization
of identity plays a significant role also in (Rayo, 2013), e.g. Section 2.4. Rayo reports in personal commu-
nication that the slightly lesser prominence of the idea in the book is a matter of exposition, not a change
in doctrine.
6 A singular term is purely referential if its semantic contribution to linguistic contexts in which it occurs
is exhausted by its referent, or, as Quine {1960) puts it, if the term "is used purely to specify its object, for
the rest of the sentence to say something about" (p. 177). A context is transparent when it is concerned only
with the referents of the terms that fill the argument place, as explained on pp. 30-1 and discussed shortly.
What are the logical properties of the generalized identity symbolized by'¢>'? This
too is an equivalence relation. And Leibniz's Law seems to generalize readily, namely
as the axiom scheme:
{LL-snt)
This generalization of Leibniz's Law appears to be a powerful tool in connection

with the transmission requirement, which says that philosophically interesting prop-
erties are transmitted across the recarving relation. It appears to follow that if <fJ and 1/1
represent one and the same content, then the two statements cannot differ with respect
to their epistemic or explanatory status; that is
(Cong) 1/1 ~ (£' is understood such that <fJ ¢> 1/1 is true just in case <{J and 1/1 express
one and the same content, we appear to have an easy way to show that the transmission
requirement is satisfied. If <P and 1/1 express one and the same content, then anything
that goes for one must also go for the other. (I do not claim that Rayo or anyone else
has endorsed this argument!)
Seductive though it may be, this argument is far too quick. As noted, the ordinary
version of Leibniz's Law is subject to two important restrictions. Perhaps analogous
restrictions are needed on the sentential version of the law. To explore this question,
consider first the restriction of the ordinary law to terms that are purely referential.
Is some analogous restriction needed for the sentential law? Such a restriction would
require a sentence-level analogue of the notion of a purely referential term. Let us label
this desired analogue explicit representation. The idea would be to restrict (LL-snt) to
sentences that explicitly represent their contents. But it is unclear what the notion
of explicit representation might amount to. The difference between a direct and a
descriptive way of singling out an object is clear and compelling. By contrast, all sen-
tences seem to describe. Indeed, in Rayds work there is no suggestion that some sen-
tences represent their contents in a more explicit way than others. I shall therefore not
differentiate between sentences that represent their contents in more and less explicit
ways. Notice that the decision is dialectically permissible as it can only make it easier
for my opponent to extract philosophically interesting conclusions from (LL-snt).
This leaves the second requirement that the context 0( . . .) be transparent. What
is the appropriate notion of transparency? Let us begin with the ordinary version of
Leibniz's Law. A context 0( .. .) open to singular terms is said to be transparent just
in case the truth-value of any sentence of the form O(t) is determined solely by the
7 Recall that an equivalence relation ~ is said to be a congruence with respect to some property F just in
case this property does not differentiate between equivalent objects; that is, just in case Vx"ly(x ~ y -->
(Fx ++ Fy)).
RAYO ON "JUST IS" -STATEMENTS 81
referent of the term t and does not depend on how this referent is presented. What is
the cash value of this notion of transparency? A natural explication of the gloss that
I just provided is (LL) itself; that is, a context is transparent just in case Leibniz's Law
holds for the context! Now we are moving round in a circle. Leibniz's Law holds for
transparent contexts, and a context is transparent just in case Leibniz's Law holds for it!
It would be a mistake, though, to conclude that Leibniz's Law and its possible
generalizations are trivial and devoid of content. The correct lesson is rather that
the law and its generalizations simultaneously constrain two notions, namely the
semantic contents of a certain class of linguistic expressions, and the transparency
of contexts that are open to such expressions: no context that is deemed transparent
may differentiate between expressions deemed to have identical content.
In the case of singular terms, we have a robust antecedent grasp of one of these
notions, namely that of the semantic content-or reference-of the relevant expres-
sions. So in this case, we know what transparency amounts to. In the case of sentences,
by contrast, we do not have a robust antecedent grasp of identity of semantic content.
As we have seen, many notions of content have been proposed. It is not even clear
that we have to choose one of these to the exclusion of all others. (Frege, for example,
famously used both sense and reference.) Only when a notion of content has been
chosen will we have fixed what we mean by identity of content and consequently also
which contexts count as transparent with respect to this notion of content.
Suppose we decide on some notion of semantic content. Then what? It is impor-
tant to observe that the more coarse-grained the notion of content, the fewer the
contexts that count as transparent. Assume, for example, that identity of semantic
content is defined as a shared Fregeau Bedeutung. Then <fJ * 1/1 coincides with
the corresponding material biconditional, with the result that hardly any contexts
count as transparent. The only contexts that count as transparent, given this notion
of content, are those that are insensitive to anything other than the truth-value of
the sentence that fills its gap. On the other hand, if we adopt a very fine-grained
notion of content, then many more contexts count as transparent. The upshot is that
(LL-snt), on its own, is powerless to support congruence claims of the form (Cong)
and thus ensure satisfaction of the transmission requirement. Our entitlement to the
congruence claims does not come for free but needs to be earned by first fixing a
notion of content and then carefully arguing that one's chosen standard of identity of
content is indeed a congruence with respect to the relevant properties.
Our next task is to examine whether Rayo succeeds in pinning down a notion
of content that supports the desired congruence claims, while also satisfying the
requirement of abstractionist recarving.
4.3 Rayo on "Just is" -Statements

Rayo's guiding idea is that two sentences express one and the same content just in case
their truth imposes the same demands on the world. This is a simple and intuitive
idea, which, as we shall see, promises real progress on the questions that we are
discussing.
To convey what he has in mind, Rayo provides a range of examples of what he calls
"just is" -statements. Some of the examples are paradigm cases of analytic truths, for
instance: 8
(1) For Susan to be a sibling just is for her to share a parent with someone else.
Other examples are based on the identity of individuals or of properties:
(2) To be hot just is to have high mean molecular kinetic energy.
The most exciting examples, however, are ones where the phrase "just is" is flanked by
clauses that differ in their ontological commitments (or at least are deemed to do so
by the usual Quinean criterion that we are using):
(3) For Susan to instantiate the property of runningjust is for Susan to run.
(4) For there to be a table just is for there to be some things arranged tablewise.
(5) For the number of the dinosaurs to be zero just is for there to be no dinosaurs.
Admittedly, in these sentences, the phrase "just is" isn't actually flanked by sentences.
But this is for unimportant syntactic reasons. To iron out this minor wrinkle, Rayo
introduces an operator'=: which corresponds to the English phrase "just is" but is to
be flanked by sentences. This corresponds to our operator <::>., which I will continue
to use, even when discussing Rayo's ideas.
What exactly does a "just is" -statement assert? Rayo provides several highly sug-
gestive glosses. The following gloss figures particularly prominently:
What it takes for ((1)] to be true is for there to be no difference between Susan's having a sibling
and Susan's sharing a parent with someone else. [... ].More colorfully: when God created the
world, and made it the case that Susan shared a parent with someone else, there was nothing
extra she had to do, or refrain from doing, in order to ensure that Susan was a sibling.
(Rayo, 2013, p. 4; emphasis in original)
Furthermore, Rayo claims that a "just is" -statement is true just in case the two
sentences flanking the generalized identity operator are "full and accurate descriptions
of the same feature of reality" (p. 5; emphasis in original). Finally, a "just is" -statement
is said to "close a theoretical gap" that would otherwise have to "be plugged by a bit
of theory" (p. 18). For example, when we learn the truth of (2), we realize that there
is simply no substantive metaphysical question to be had about why a gas with high
mean molecular kinetic energy is also hot.
These glosses are meant to support what I regard as the most important
philosophical application of Rayo's view. Consider the two sentences that flank the
generalized identity operator in a true "just is" -statement. According to Rayo, there
8
This and the next examples are found in (Rayo, 2013, p. 3).
ABSTRACTION AND WORLDLY ASYMMETRY 83
is no "theoretical gap" between the claims made by the two sentences, which are just
different descriptions of "the same feature of reality''. We can therefore dismiss as
misguided certain explanatory demands that it would otherwise have been difficult
to meet. Given the truth of one of the sentences, there is simply no space for a
metaphysical explanation of why the other sentence should be true as well. For
example, Rayo finds it senseless to demand a substantive metaphysical explanation
of why there is a table where there are simples arranged tablewise, or of why there
is an abstract cardinal number zero where there are no objects of a certain kind.
These explanatory demands are, he thinks, just as misguided as the demand for a
substantive metaphysical account of why a spaceship that has landed on Phosphorus
has also landed on Hesperus (cf. p. 5). Much the same goes for epistemic status. If rp is
known and there is no "theoretical gap" between rp and 1/1, then non-linguistic reality
poses no obstacle to knowledge of 1/J. The only possible obstacle would be semantic;
that is, to figure out what 1/1 means.
It is instructive to compare these claims with the epistemic and explanatory
constraints that were formulated in Section 1.6. Rayo claims that the "just is" relation
(which we represent with'<=>' rather than Rayds '=')is a congruence with respect
to the properties of having a certain metaphysical explanation and being known
modulo semantic knowledge. That is, rp <¢> 1/1 entails E:rp *+ E:i/I when t: stands for one
of the mentioned properties. These congruence claims are strong enough to ensure
satisfaction of our explanatory constraint, as well as a strong version of our epistemic
constraint.
Why should we believe Rayo's congruence claims? As I understand it, his argu-
ment has a simple structure. First, <¢> is a congruence with respect to all purely
"worldly" aspects of statements. That is, (Cong) holds for all £ that are concerned
with how the world is, as opposed to our epistemic perspective on the world or
linguistic description of it. This premise is meant to follow from his notion of
"just is" -statements and the associated standard for identity of content. Second,
the relevant properties-admitting of a metaphysical explanation and being known
modulo semantic knowledge-are worldly. The resulting argument is clearly valid:
the premises entail the desired congruence claims. But is it sound? For the sake of the
argument, I shall grant the distinction between worldly and non-worldly properties,
as well as the truth of the second premise. But I shall argue that the first premise is
problematic; specifically, the two sides of an abstraction principle need not be on a
par with respect to all worldly properties.
4.4 Abstraction and Worldly Asymmetry

Consider one of the more plausible candidates for an acceptable abstraction principle:
(Dir)
I shall now defend an asymmetric conception of abstraction according to which the

statements that flank the biconditional are not on a par with respect to all worldly
properties. Suppose my argument succeeds. The ensuing asymmetry confronts Rayo
with an awkward dilemma. Either (Dir) and other acceptable abstraction principles
cannot be strengthened to recarving claims, in violation of the requirement of
abstractionist recarving. Or else a recarving claim does not ensure that the statements
flanking the operator '<=>' are on a par with respect to all worldly properties, thus
undermining Rayo's defense of the transmission requirement. Each horn of the
dilemma thus leads to an unacceptable conclusion.
It remains to establish the asymmetry that generates the dilemma. I shall examine
three kinds of property which, I claim, are worldly but not always shared by the two
sides of an abstraction principle.
The first class of examples concerns metaphysical (and thus non-epistemic) explan-
ations. Consider (Dir). It is natural to think that d(/1) = d(/2) is true because 11 II 12,
but not the other way round. Analogous claims are plausible for other forms of
abstraction. 9 For example, the set of xx is identical to the set of yy because xx and yy
are the very same objects. The identity of the sets is explained in terms of the
coincidence of their elements, not the other way round. In fact, there are cases that do
not involve abstraction at all. Many philosophers take the following, for example, to
be an asymmetric truth:
(6) There is a table because there are some things arranged tablewise.
But Rayo endorses the following recarving claim:
(61) There is a table -¢=> there are some things arranged tablewise.
The asymmetry of the former claim threatens to conflict with the worldly symmetry
said to follow from the latter.
How might Rayo respond to these examples? Clearly, he has to deny either that the
proposed asymmetric explanations are genuine or that they are worldly. But neither
option is attractive. The past decade has seen a surge of interest in an asymmetric
notion of metaphysical explanation. And these explanations are regarded as worldly,
not just features of our epistemic or semantic perspective on the world. 10
The second class of examples concerns ontological commitments. As I have empha-
sized, on the usual Quinean notion of ontological commitment, the two sides of an
abstraction principle have different ontological commitments. 11 While the left-hand
side of (Dir) is committed to directions, the right-hand side is committed only to lines.
Again, Rayo has to deny either that the alleged asymmetries are genuine or that the
usual Quinean notion of ontological commitment is an entirely worldly matter. But
in this case, the first option is a non-starter. Some of the statements that Rayo regards
9 See (Rosen, 2010, p. 123), (Schwartzkopff, 2011), and (Donaldson, 2017).

10 11
See e.g. (Fine, 2012a), (Rosen, 2010), and (Schaffer, 2009) . See Section 1.6.
ABSTRACTION AND WORLDLY ASYMMETRY 85
as recarving of one and the same content indisputably differ in Quinean ontological
commitment; (Dir} provides one example.
Rayo therefore chooses the second option. Any proper notion of ontological
commitment, he insists, must give rise to worldly properties, which must accordingly
be shared by the two sides of any recarving claim. 12 Properly understood, the two sides
of(Dir) do indeed share the same ontological commitments. We thus learn, much to
our surprise, that the line statement Ii II 12 is committed to directions after all. An
even more extreme example follows from Rayo's claim that the truths of set theory
impose only a trivial demand on the world, no greater than that of any tautology. 13
Let <p and T be your favorite true large cardinal statement and tautology, respectively.
On the usual Quinean notion, <p has staggering ontological commitments, while T
is as uncommitted as can be. On Rayo's alternative notion, however, the ontological
commitments of T are just as extensive as those of <p.
At this point, I lodge a two-pronged protest. First, a notion of ontological com-
mitment with these mentioned consequences is not worth its salt. Part of the very
purpose of this notion is to differentiate between tautologies and higher set theory.
When a notion of ontological commitment fails to make this sort of differentiation,
it is broken and needs to be replaced. Second, I submit that the ordinary Quinean
notion of ontological commitment-which does make the desired differentiations-
is worldly, Rayo's protestations notwithstanding. The Quinean notion is concerned
with what objects reality must contain for a certain sentence to be true. This has to do
with the world, not just with our epistemic or semantic perspective on it. In particular,
Quinean ontological commitment can be explicated so as to be unaffected by Frege
cases; for example, "there exist two groundhogs" is committed to two woodchucks,
on the grounds that the two nouns are synonymous.
My final class of examples concern the property of being about certain objects.
For example, the left-hand side of (Dir) is about a certain direction, whereas the
right-hand side is not. Moreover, I claim that being about certain objects is a worldly
property of a sentence. To establish this, observe that we can "factor out" the Frege
cases typically deployed to show that some feature of a sentence concerns our
epistemic or semantic perspective on the world, not the world itself. For example,
there is a clear sense in which "Samuel Clemens is a writer" is about Mark Twain, but
not about the Eiffel Tower. The sentence is about the man himself, not merely the man
under some mode of presentation. Aboutness is de re, not de dicto.
In these cases, it is not an option for Rayo to insist on symmetry. On any viable
notion of aboutness, the two sides of, say, (Dir} are about different objects. 14 The only
12 See (Linnebo and Rayo, 2012, p. 291) (where the relevant passage represents AR's view, not 0I:s),
(Rayo, 2013, p. 173), and (Rayo, 2016, p. 214).

13
See (Linnebo and Rayo, 2012, p. 291) (where again the relevant passage represents AR's view) ; cf. also
the argument in (Rayo, 2016, pp. 214-15).
14
Cf. (Hale and Wright, 200la, pp. 164-7).
option is to deny that what a sentence is about is a worldly property. But this option too
has unappealing consequences. Where there is genuinely an object, it must be possible
to make de re statements about this object, not merely make non-worldly statements
about how something is presented to us epistemically or linguistically. By insisting that
all aboutness properties are non-worldly, one casts doubt on the alleged existence of
the object in question, which again appears little more than a fafon de par/er.
I wish to end by briefly explaining how an asymmetric conception of abstraction
avoids the problems just discussed. Suppose we admit that the three kinds of property
discussed above are indeed worldly and that the two sides of acceptable abstraction
principles are asymmetric with respect to these properties; that is, that one side has
the property, while the other does not. On this view, it is not an option to satisfy the
transmission requirement by defending the congruence claims. This need not be a
problem, however, as there can be other ways to satisfy the requirement. Although
abstraction is asymmetric with respect to various worldly properties, the right-hand
side may still suffice for the left-hand side; for example, we may have:
What ultimately matters is that the relevant philosophical properties should be

transmitted-or potentially transmitted-from left to right across the sufficiency
operator. It suffices, for example, that the desirable epistemic and explanatory status
of 11 II 12 should (potentially) transmit to d(/1) = d(/2)-as is indeed the case on
my favored explication of =?. 15 In short, an asymmetric conception of abstraction is
not only more plausible in its own right but also fully compatible with the promised
benefits of thin objects.
The case for asymmetry developed in the present chapter supplements and rein-
forces the arguments from Section 1.5. So the cumulative evidence against the
symmetric conception of abstraction is substantial.
15
Cf. Section 2.7.
5
Unbearable Lightness of Being
5.1 Ultra-Thin Conceptions of Objecthood

On the Fregean conception I defended in Chapter 2, an object is characterized as a
possible referent of a singular term. This connects the concepts of objecthood and
reference. I proceeded to develop an undemanding conception of reference according
to which suitable use of criteria of identity suffices to set up, or constitute, relations
of reference. This connects the concepts of reference and of a criterion of identity.
Together, these two connections give rise to a thin conception of objecthood. The
second connection yields an undemanding conception of reference. When combined
with the first connection, this results in an undemanding, or thin, conception of
objecthood.
The topic of this chapter are some even less demanding conceptions of objecthood.
These "ultra-thin" conceptions arise when the Fregean concept of an object is com-
bined with an exceedingly liberal conception of reference, far less demanding than
my own. A prominent example derives from the "syntactic priority thesis" defended
by Hale and Wright, especially in their earlier work. 1 Here is a frequently quoted
statement of the view:
when it has been established, by the sort of syntactic criteria sketched, that a given class of
terms are functioning as singular terms, and when it has been verified that certain appropriate
sentences containing them are, by ordinary criteria, true, then it follows that those terms do
genuinely refer. And, being singular terms, their reference will be to objects. There is to be no
further, intelligible question whether such terms really have a reference, whether there really
are such objects. (Wright, 1983, p. 14; emphases in original)
In other words, if an expression functions syntactically and inferentially just like a

singular term, then it is a singular term; and if a singular term figures in a true (atomic)
sentence, then there exists an object that the term denotes. It is hard to set the bar to
objecthood much lower than this.
A closely related view has recently been defended by Agustin Rayo under the name
of "compositionalism''. Here is a clear expression of the view:
Let L be an uninterpreted first-order language, and consider an assignment of truth-conditions

to sentences in L. Our assignment must respect logical entailments, but we are otherwise allowed
1 See (Wright, 1983, esp. sections ii, iii, viii). A similar view is espoused in (Dummett, 1956) but is
abandoned in later work by Dummett, most clearly in (Dummett, 199la).

88 UNBEARABLE LIGHTNESS OF BEING
to pick any assignment we like. According to compositionalism, [.. . ] [w]hen the sentences of L
are so interpreted, all it takes for a singular term t of L to refer to an object in the world-all it
takes fort to be non-empty-is for the world to satisfy the truth-conditions that were assigned
to r3x(x = t) ' (or some inferential analogue). (Rayo, 2014, pp. 498-9; emphasis in original)
In a footnote to the quoted passage, Rayo explains what it is for an assignment of truth-
conditions to "respect logical entailments'; namely: "if if! is a logical consequence of
<p, then the truth-conditions assigned to rp must demand at least as much of the world
as the truth-conditions assigned to if! :'2 Overall, the view assures us that once all
sentences have been assigned truth-conditions that respect logical entailments, then
all subsentential expressions obtain their meaning for free; in particular, all singular
terms obtain their reference.3 As Rayo puts it, compositionalism is the view that "our
language only makes contact with the world at the level of sentences" (Rayo, 2014,
p. 500; emphasis in original).
There are important points of agreement between the neo-Fregeans' syntactic
priority thesis and Rayo's compositionalism. Both views take sentences to have a
special explanatory status in semantics. And the two views offer similar sufficient
conditions for an expression to refer. The only semantic assumption involved in the
two proposed sufficient conditions pertains to whole sentences, namely that suitable
atomic sentences should be true. This sole semantic assumption is supplemented,
on both views, with some non-semantic assumptions intended to ensure that the
expression in question functions as a singular term syntactically and inferentially. The
major difference between the two views concerns these non-semantic assumptions.
The neo-Fregeans formulate assumptions intended to apply to natural languages such
as English, while Rayo focuses on formal languages such as that of first-order logic.
Not surprisingly, the latter option makes it easier to characterize the syntactic and
inferential behavior characteristic of a singular term, as witnessed by the clear and
straightforward condition quoted above, contrasted with the neo- Fregeans' more
elaborate and error-prone characterization. 4 I therefore focus on Rayo's proposed
sufficient condition in what follows. But I believe most of what I say carries over to
the syntactic priority thesis, mutatis mutandis.
Are the ultra-thin conceptions of objecthood defensible? One might have thought
that, if you liked the thin conception sketched in Chapter 2, you should love these
ultra-thin alternatives. This thought is badly mistaken, and the purpose of this chapter
is to explain why. My thin conception has advantages, I argue, that are not shared by
its ultra-thin rivals. I should hasten to add that I have no knock-down objections to
the ultra-thin conceptions. If you insist on talking about reference and objecthood in
an ultra-thin way, I cannot stop you. My strategy is to argue that my own approach
2
See also Section 1.3 of (Rayo, 2013).
3 In fact, (Rayo, 2013, p. 14) relaxes this sufficient condition for reference, requiring only that all
sentences "available for use" have been assigned truth-conditions.
4
See essays I and 2 of (Hale and Wright, 2001a).
LOGICALLY ACCEPTABLE TRANSLATIONS 89
yields conceptions of reference and objecthood that are more interesting and worthy
of our attention. I begin, in the next section, by explaining the important notion of a
logically acceptable translation. Using this tool, the rest of the chapter develops two
objections to the ultra-thin conceptions.
5.2 Logically Acceptable Translations

How can we assign truth-conditions to the sentences of an uninterpreted language?
Clearly, we must rely on an interpreted language in which the truth-conditions
to be assigned are expressed. This means that we need to consider two languages
simultaneously. Let .Co be an interpreted language concerned with objects that are
accepted as unproblematic, at least for present purposes; we call this the base language.
Let .C1 be a (wholly or partially) uninterpreted language to whose sentences we wish
to assign truth-conditions; we call this the extended language . .C 1 may or may not
overlap with .Co. If it does overlap, then .C 1 is obviously partially interpreted; though
we shall assume that any vocabulary that .C 1 adds to .Co is uninterpreted. The strategy
is to use the interpreted base language to state the truth-conditions that we assign to
those sentences of the extended language that have not yet been interpreted.
An easy example of Frege's illustrates how the assignment of truth-conditions can
be achieved. We begin with a base language capable of expressing claims about lines
and perhaps other objects, but with no vocabulary for talking about directions. We
now consider an extended language that adds vocabulary suitable for talking about
directions, such as the direction-of operator d and certain predicates concerned
with directions. As Frege no doubt realized, provided that the extended language
is carefully chosen, there is a natural translation back into the base language. For
example, the identity statement 'd(/1) = d(l 2 )' is translated as '1 1 II 12 ', in accordance
with the familiar criterion of identity for directions. This translation can be seen as
mapping each sentence of the extended language to its truth-condition expressed in the
base language.5
We would like a general analysis of the kind of translation just illustrated. Let a
logically acceptable translation be a recursive function r of formulas from the extended
language to formulas of the base language that satisfies the following conditions: 6
(i) For any formula <p of .C 1 and set L of such formulas, we have
if L F <p, then r(L) F r(<p),
where r(X) abbreviates {r(<p) I <p E X}. (That is, r preserves logical entail-
ment.)
5 A precise definition of the languages and the translation, as well as generalizations to other criteria
of identity, are provided in Chapter 6.

6 Condition (i) is a version ofRayo's requirement that logical entailment be respected. Condition (ii) is
independently plausible and also serves to rule out certain unintended interpretations that satisfy condition
(i) , for example, the interpretation that assigns the trivial content to every formula in the language.
(ii) For any formula cp of £i. it is a logical truth that r(-.cp) ++ -.r(cp). (That is, r
commutes with negation modulo logical consequence.)
Though somewhat complex, the definition rests on an intuitive idea. A logically
acceptable translation can be understood as mapping each sentence cp of some con-
troversial extended language to the associated demand on the world, r (cp), expressed
in an uncontroversial base language. In other words, cp and r (cp) make the very
same demand on the world, only expressed in more and less controversial terms,
respectively.7
Two features of our definition merit special comment. First, notice that the defin-
ition applies to open formulas, not just sentences. By contrast, Wright and Rayo talk
only about sentences. But in our Fregean example, it is plausible to allow the terms
'11' and '12' to be variables. When we include open formulas as well, our gloss that cp
and r(cp) make the same demand on the world needs to be relativized to a variable
assignment. If desired, we could nevertheless choose to stay closer to Wright's and
Rayo's characterization by restricting the definition to sentences.
Second, although the definition requires the translation r to preserve logical
entailment, it does not require r to preserve non-entailment. The definition does not,
in other words, require that if r V= ijl, then r(f) V= r(ijl). The reason is obvious.
We want to permit a formula of the extended language that is not a logical truth
to be mapped to a logical truth of the base language. We may for example want to
map the non-logical truth "there is a number" to the logical truth "there is a concept
equinumerous with itself". It is alright to obtain "for free'' the truth of certain sentences
of the extended language. What we need to guard against are translations such that, by
applying logic to the extended language, we obtain consequences in the base language
that were not previously licensed.
5.3 Semantically Idle Singular Terms

The ultra-thin conceptions state that an expression refers provided that it figures in
appropriate true (atomic) sentences and behaves syntactically and inferentially as
a singular term. So long as these conditions are met, it does not matter how the
language obtains its meaning. Is this a reasonable view? In this section and the next,
I develop two objections to the view, both of which rely crucially on the notion of
a logically acceptable translation, which is designed specifically to ensure that the
relevant singular terms behave inferentially just as they should.
7
This gloss is obviously based on a symmetric conception of the translation, where cp and r(cp) are on
a par with respect to their demands on the world. My own favored conception is an asymmetric one; in
particular, I believe the two sides of an abstraction principle make different demands on the world. (See
Section 1.7 and Chapter 4.) For example, only the right-hand side of the direction abstraction principle is
ontologically committed to directions. Fortunately, the intuitive gloss is easily tweaked to fit an asymmetric
conception as well, namely that r:(cp) suffices for cp, for every sentence cp; that is, that r(cp) => cp.
SEMANTICALLY IDLE SINGULAR TERMS 91
Although an erstwhile defender of the view, Dummett has long been critical of the
ultra-thin conception. 8 Consider again the example of direction abstraction. As is
shown in Chapter 6, we can lay down clear and logically coherent rules governing the
use of direction terms of the form 'd(I)'. What is achieved by doing so? According to
Dummett,
[t]his may quite reasonably be taken as a sound method of justifying the use of names of
directions: but it would naturally be construed, not as a demonstration that such names have
reference in the same way as names of concrete objects, but as a way of explaining their use
without ascribing reference to them. (Dummett, 198la, pp. 499-500)
This is a powerful challenge. Perhaps the direction terms serve merely as a novel and
unusual way to express claims about lines. Perhaps we have merely justified a fafon de
par/er, not explained genuine reference to, or quantification over, directions.
Dummett is at least partially right, as can be seen by using our notion of a
logically acceptable translation. Consider how such a translation, say r, might be used.
A simple option is to let r assign a meaning, expressed in the base language, to each
sentence of the extended language. In other words, each sentence <p of the extended
language is stipulated to mean exactly the same as the sentence r (<p) of the interpreted
base language. What would this stipulation involve in practice? Suppose the members
of a language community start out speaking some meaningful base language, but
that they later adopt an extended language, which they learn in a special way. Each
new sentence <p of the extended language is mastered by translating it into the base
language in accordance with r. For example, 'd(/1) = d(/2)' is translated as '/1 1112'. Thus
in particular, to find out whether <p is true or false, the members of the community first
need to determine whether r (<p) is true or false. Since each sentence of the extended
language is assigned a meaning expressed in the unproblematic base language, this
is a reductionist view. Moreover, since each sentence is assigned a meaning only as
a whole, not in virtue of any meanings assigned to its subsentential expressions,
the reductionism can be characterized as holophrastic in the sense of (Quine, 1992),
namely that it works at the level of whole sentences.
When a language is given a holophrastic reductionist interpretation, there is no
direct interaction between the syntactic structure of a sentence and its semantic
interpretation. Each sentence <p receives its semantic interpretation only via the
translation r(rp), not via its subsentential structure. Any singular term that is found
in rp, but not in its <-translation, is therefore "se~antically idle", as Dummett aptly
puts it. The direction terms that occur in 'd(/ 1 ) = d(/2 )' provide an example, as these
disappear when the sentence is translated in the usual Fregean way. This interpretation
completely shortcuts the principle of compositionality, which tells us that the meaning
of rp is obtained in a systematic way from the meanings of its simple constituents.
8 See (Dummett, 1956) for the former and (Dummett, 198la) and (Dummett, 199la) for the latter.
A related criticism of the ultra-thin conception is developed in (Hodes, 1984a).

While holophrastic reductionism justifies a f afon de par/er, it thus fails to ensure

genuine reference. 9
Does this mean that logically acceptable translations are of no use whatsoever in
an account of thin objects? Not necessarily. I have identified a problem with one
particular use of such translations, namely that of holophrastic reductionism. But
there might be other, more fruitful uses. My own view is that such alternative uses
do indeed exist-at least in the special cases where the translation is based on a
criterion of identity. 10 My defense of this view was initiated in Chapter 2 and will be
completed in Chapter 8. The idea is that criteria of identity can be used to constitute
new forms of reference. This view must be distinguished from the holophrastic
reductionism that I just criticized. On the view that I favor, criteria of identity help
to endow singular terms with a reference, which in turn figures in a compositional
semantic theory; while on holophrastic reductionism, the semantic interpretation
takes place only at the level of whole sentences, thus bypassing the principle of
compositionality.
5.4 Inexplicable Reference

My second objection to the ultra-thin conceptions is that they give rise to inexplicable
relations of reference. Consider the relation that obtains between a singular term
t and the object o to which t refers. This relation is contingent and should admit
of explanation, at least in principle. After all, there is a reason why t refers to o,
which typically has to do with how the language is used. I now provide some
examples where the conditions of the ultra-thin conceptions are satisfied but where
the resulting relations of reference would be inexplicable-in the strict sense that it is
in principle impossible to explain why the term t refers to o as opposed to some other
object.
A striking class of examples are based on theories that are complete and decidable.
Let me explain. A theory is said to be complete just in case it is a maximal consistent
set of formulas in its language. A theory is said to be decidable just in case there is
a recursive procedure for determining membership in I;. There are many examples
of theories that are both complete and decidable. A simple example-with features
that prove to be particularly interesting-is the theory of dense linear orders without
endpoints. 11 The language of this theory has a single predicate '<'. The theory says
9 Dummett makes another complaint as well, namely that there are "recognition statements" involving
lines but not directions. I find this complaint problematic, as I explain in Section 11.4.
10
For all I know, there may be other examples as well. As emphasized repeatedly, I am committed to
the sufficiency of criteria of identity (used in a certain way) for reference, but not to the necessity of this
approach.
11
Other examples include the complete theory of any finite model (e.g. a group), Presburger arithmetic,
Tarski's axiomatization of Euclidean geometry, Real Closed Fields, and Algebraically Closed Fields of
characteristic k.
INEXPLICABLE REFERENCE 93
that < is a linear order; that it is dense (in the usual sense that whenever x < y, there
is a third point in between, i.e. 3z(x < z /\ z < y) ); and that it lacks endpoints (in the
sense that, for any point x, there is both a greater point and a lesser point).
I claim that every theory :E that is complete, decidable, and consistent gives rise
to a logically acceptable translation. For example, let :E be the theory of dense linear
orders without endpoints, and let .C1 be its language. Let the base language be that of
propositional logic, but with only two atomic formulas, namely the constants T and ..l,
which are true and false, respectively, on each interpretation. Let T be the translation
that maps a formula <p of .C 1 to Tor ..l according as <pis a theorem of :E or not. As is
proved in Appendix 5.A, we obtain:
Proposition 5.1 The translation r is logically acceptable.
It is easy to modify the example so as to add a constant c to the language .C 1 . The
resulting theory too is complete, decidable, and consistent. 12
This example is problematic for the ultra-thin conceptions. Consider a linguistic
practice where the above translation is used in a holophrastic reductionist manner.
According to the ultra-thin conceptions, this practice endows the new constant c with
a reference. It is hard to see how the practice might succeed in doing so, however. To
which of the infinitely many objects standing in the dense linear ordering < does
c refer? Absolutely nothing distinguishes one of the potential referents from any
of the others. So there is absolutely no reason why c should refer to one of the
objects rather than some other! This would be a brute and inexplicable relation of
reference. Moreover, if the translation endows the singular terms of .C 1 with reference,
it presumably also endows its variables and quantifiers with a range, namely all of the
objects ordered by < . This prompts the question of how many such objects there are.
Since there are dense linear orderings without endpoints of any infinite cardinality,
there is no reason why the domain should have one cardinality rather than another!
The cardinality of the domain over which we quantify would be another brute and
inexplicable fact.
I conclude that it is not sufficient for a term to refer that there is a logically
acceptable translation from a language containing the term to an interpreted language
that does not. The ultra-thin conceptions are unacceptably liberal in their ascription
of reference. My own approach to "easy reference" avoids this objection. On this
approach, begun in Chapter 2 and completed in Chapter 8, the relation between a
singular term and its referent always admits of an explanation. A term refers to a
particular object in virtue of being associated with a specification of that object and
a unity relation, which provides a criterion of identity for the object. So my approach
avoids inexplicable relations of reference.
12
Generalizing further, it is straightforward to add any finite number of constants, so long as we add to
the axiomatization an ordering, say co < C1 < . . . < Cn.
Appendices
5.A Proofs and Another Proposition
Proof of Proposition 5.1. The translation r is recursive because the theory of dense linear
ordering without endpoints is decidable. Next, we observe that, for each <p and 1/1 in £,, the
following two commutation properties are logical truths:
(5.1) r(rp /\ 1/1) B r(rp) /\ r(1/I)

(5.2) r(-.rp) B -.r(rp)
(5.1) holds because a conjunction is a theorem just in case each of its conjuncts is. (5.2) holds
because in a complete and consistent theory, precisely one of rp and -.rp is a theorem.
It remains to show that r preserves logical entailment. Assume r F= rp. By the compactness of
first-orderlogic, there is a finite subtheory ro ~ r such that ro F= rp and thus (by completeness)
also I- ;\ r o --+ <p. This ensures I- r (;\ r o --+ rp) and thus, by the commutation properties just
established, I-(;\ r(fo)--+ r(rp)). This entails r(f) F= r(rp), as desired. ~
I end by establishing a proposition which provides further assurance that our notion of a
logically acceptable translation behaves as it should.
Proposition 5.2 Any logically acceptable translation r commutes with the truth-functional
connectives modulo logical equivalence.
Proof We need only consider conjunction and negation, which suffice to define the other
connectives. The case of negation is ensured by clause (ii) of our definition of a logically
acceptable translation. Consider the case of conjunction. Since rp /\ 1/1 F= rp, clause (i) ensures
r( rp /\ 1/1) F= r( rp). Since r( <p /\ 1/1) F= r( 1/1) follows in an analogous way, we obtain r( <p /\ 1/1) F=
r(rp) /\ r(1/f). For the other direction, observe that rp, 1/1 F= <p /\ 1/1 ensures r(rp), r(1/f) F=
r(rp /\ 1/1) by clause (i) again and thus also r(rp) /\ r(1/f) F= r(rp /\ 1/f) . Taken together, these
two results show that r commutes with conjunction. -j
6
Predicative vs. Impredicative
Abstraction
6.1 The Quest for Innocent Counterparts

Consider an abstraction principle, such as Frege's principle for directions:
(Dir)
The statement on the left-hand side seems philosophically puzzling because of its
commitment to abstract objects. By contrast, the statement on the right-hand side
is unproblematic. This brings us to one of the most attractive ideas in the entire
abstractionist program. Perhaps we can remove the puzzlement that attaches to the
statement on the left by connecting it very tightly with the unproblematic statement
on the right. As Hale and Wright put it, the two statements leave "no gap for
metaphysics to plug" (Hale and Wright, 2009b, p. 193). In other words, there is no
"metaphysical gap" between the two statements. The statement that initially seemed
puzzling is thus shown to be no more problematic than its innocent counterpart on the
right-hand side.
For this attractive idea to be convincing, however, it needs to be properly spelled
out. One obvious question is what kind of "tight connection'' we envisage between a
statement and its innocent counterpart. This question figures centrally in several other
chapters. In the present chapter, however, we shall investigate the more fundamental
question of whether every statement even has an innocent counterpart. Suppose
we start with a base language .Co in which all the innocent counterparts are to be
formulated; in the above example, this would be a language concerned with lines and
other unproblematic objects. Then there is an extended language .C 1 in which the
abstraction principle is formulated and where we find all of the puzzling statements.
Clearly, we need a translation from the extended language to the base language which
maps each statement of the former to its innocent counterpart in the latter. It would
offer only false comfort if some, but not all, of the puzzling statements have such
innocent counterparts. Can an appropriate translation be found?
The aim of this chapter is to show that an appropriate translation can indeed be
found when the abstraction is predicative, in a sense to be explained shortly-but not
when the abstraction is impredicative. The attractive idea about how to demystify talk
96 PREDICATIVE VS. IMPREDICATIVE ABSTRACTION
about abstract objects is therefore only available for predicative abstraction principles,
contrary to what is widely assumed in the literature. 1 This is a major advantage of
predicative abstraction principles.
Before embarking on our argument, we should recall from Section 3.4 that pre-
dicative abstraction principles enjoy several other important advantages as well, which
will not be discussed here. In particular, the recalcitrant bad company problem
receives a simple and definitive solution in that all forms of extensional abstraction
are permissible.
6.2 Two Forms of Impredicativity

The notion of impredicativity originated in the early twentieth-century debate
between Poincare, Russell, and others about the nature and cause of the logical
paradoxes. Both Poincare and Russell took the paradoxes to be the result of some form
of vicious circularity. What goes wrong, they suggested, is that an entity is defined,
or a proposition is formulated, in a way that is unacceptably circular. Sometimes this
circularity is transparent, as in the famous liar paradox of the sentence that says of
itself that it is not true. In other cases, there is no explicit circularity. For example, the
definition of the Russell set as the set whose elements are all and only the non-self-
membered objects makes no explicit reference to the set being defined. The notion of
impredicativity was meant to characterize an implicit form of circularity believed to
be present even in paradoxes such as Russell's. ·
The problem with the Russell set, it was thought, is that its definition quantifies
over a totality to which the defined set (if any) would belong. One of the objects that
needs to be considered for membership in the Russell set is therefore this very set
itself. More generally, by quantifying over a totality we presuppose all the entities that
make up this totality. So if we define an entity by quantifying over a totality to which
the defined entity would belong, we presuppose the entity being defined. And this, it
was thought, is a form of vicious circularity.
This diagnosis of the logical paradoxes is, of course, controversial. But it has given
rise to an important distinction between predicative and impredicative definition.
A definition is said to be impredicative if it quantifies over a totality to which the
entity being defined (if any) belongs. Otherwise the definition is said to be predicative.
Following the influential defenses of impredicative definitions by Ramsey, Bernays,
Godel, and others,2 the received view is now that impredicative definitions are
unproblematic in most of classical mathematics. The distinction between predicative
1 (Dummett, 1991a) provides a rare exception. See (Linnebo, 2016a, §12.5) for discussion and further
references.
2 See for instance (Ramsey, 1931). (Bernays, 1935), and (Godel, 1944).
TWO FORMS OF IMPREDICATIVITY 97
and impredicative definitions is nevertheless widely agreed to mark an important

watershed in the foundations of mathematics.
In line with the received view, most abstractionist approaches to mathematics freely
accept the use of impredicative reasoning. The purpose of this chapter is to argue that
this view needs to be revised, at least in part.
To be more precise, we need to distinguish between two different forms of impre-
dicativity that arise in connection with the abstractionist program. The first and most
familiar form of impredicativity concerns the background second-order logic that we
use. When formalizing second-order logic it is useful to formulate the inference rules
governing the second-order quantifier such that they can be directly applied only
to free predicate variables, not to arbitrary predicate expressions. This formulation
requires us to be explicit about what instances of the second-order variables we accept
as legitimate. We make this explicit by means of so-called second-order comprehension
axioms, which are of the form
(2-Comp) 3Xv'x(Xx -++ rp(x))
where <p(x) is some formula with x free and with no occurrences of X . Let us consider
a simple example. With the above restrictions in place, we cannot use second-order
Universal Instantiation to derive 3xrp(x) directly from VF3xFx. Rather, we must first
use the comprehension axiom 3XVx(Xx-++ <p(x)) to infer that Vx(Xx-++ rp(x)), then
use second-order Universal Instantiation on VF3xFx to obtain 3xXx, whence we can
finally conclude that 3xrp(x) .
A comprehension axiom can be seen as a definition of a second-order entity X by
means of a formula <p(x). A question of predicativity thus arises: Does the formula
<p(x) quantify over a totality to which the defined entity X belongs? The general idea
of predicativity with which this section began motivates the following (standard)
definition. A comprehension axiom is predicative ifthe comprehension formula <p(x)
contains no bound second-order variables, and impredicative otherwise.
The second form of impredicativity concerns the abstraction principles themselves.
Consider an abstraction principle of the form
(AP) §a = §{3 -++ a '""' f3
where § is an operator taking variables of the appropriate type to first-order terms
and '""' abbreviates a formula that defines an equivalence relation. An abstraction
principle can be regarded as implicitly defining the abstraction operator § that figures
on its left-hand side. Again a question of impredicativity arises: Does this implicit
definition quantify over a totality that includes entities in the range of the operator§?
The general idea of predicativity motivates the following definition. An abstraction
principle is impredicative if the terms on its left-hand side denote objects included
in the range of some quantifier occurring on its right-hand side; otherwise the
abstraction principle is predicative.
Let us consider some examples. The principle (Dir) is clearly predicative, since the
variables 11 and 12 range only over lines, not over directions. Here are two other famous
examples:
(V) eF = eG B Yx(Fx B Gx)
(HP) #F=#G BF ~ G
Both of these principles are impredicative because their right-hand sides quantify
over the objects whose identity conditions are specified by their left-hand sides. This
shows that the class of impredicative abstraction principles includes both principles
that are typically regarded as good, such as (HP), and ones universally regarded as
bad, such as (V).
To sum up, two different forms of impredicativity arise in connection with abstrac-
tionism. While there is a common idea underlying them, the two forms of impred-
icativity are conceptually and logically independent of one another and give rise to
different objections and concerns. 3
My focus in the remainder of this chapter is exclusively on the form of impredica-
tivity that pertains to the abstraction principles themselves. I shall freely avail myself
of impredicative higher-order logic. 4 I establish one positive and one negative result
concerning the availability of the desired translations from the extended language to
the base language, which map each statement of the former to its innocent counterpart
in the latter. In Section 6.3, I show that for each abstraction principle that is pre-
dicative in the sense just defined, there is an appropriate translation. In Section 6.4,
I show that impredicative abstraction principles do not always permit the desired kind
of translation and discuss the philosophical significance of this technical fact.
6.3 Predicative Abstraction

How can we ensure that an abstraction principle is predicative? An easy option is
to formulate the principle in a two-sorted language, which has one logical sort for
basic objects and another sort for abstracts. So I begin by describing the relevant
class of such languages. Then I define a translation, which, I argue, delivers what we
want. I focus on the case of abstraction on objects. The case of abstraction on Fregean
concepts or pluralities is analogous.
In fact, predicativity can also be enforced in a one-sorted setting. Since this is less
straightforward, discussion of it is deferred until Chapter 9.
6. 3.1 Two-sorted languages
Consider a base language, .C0 , which can be any interpreted one-sorted language
assumed to be in good philosophical standing. Let ~ abbreviate a formula that
3 It is therefore unfortunate that the two forms of impredicativity are often not clearly distinguished.
For instance, (Dummett, 199la, ch. 17) moves unannounced from one form to the other.
4
The impredicativity that pertains to the higher-order logic is discussed briefly in Appendix 3.A.1
and more extensively in (Linnebo, 2016a).
PREDICATIVE ABSTRACTION 99
defines an equivalence relation, either on objects or on concepts of some sort over

which the base language quantifies. We define a two-sorted extended language, .Cl>
as the result of adding to the base language vocabulary suitable for talking about the
objects obtained by abstraction with respect to""· In particular, .C 1 has an abstraction
operator §, which can combine with any expression of the appropriate sort to form
a complex singular term. For reasons that will shortly become dear, we let .Co have
two sets of variables, xo, XI> . .. and x~, Xi, .... We let .C 1 have one set of variables
xo, x1, ... for the "old" objects with which already .Co is concerned, and another set
yo,y1, ... reserved for "new" abstracta described only by .C 1. By separating in this
way between the "old" and the "new" ontology, it is easy to ensure that an abstraction
principle is predicative: We let any quantifier on the right-hand side range over the
"old" ontology only, while the abstraction terms that figure on the left-hand side refer to
"new" objects.
The formation rules of .C 1 are such that each argument place of each atomic
predicate is open only to terms of one of the two sorts. In the case of directions, for
example, the orthogonality predicate J.. takes two arguments of the basic sort, while
J.. * takes two arguments of the extended sort; any other combination is regarded
as ill-formed. These syntactic restrictions proscribe all "mixed" identities involving
one basic term and one abstract term. Such problematic identities cannot even be
formulated. I am under no illusion that this solves the notorious Caesar problem,
which concerns how such mixed identities are to be handled. All that our use of a
two-sorted language achieves is to postpone the problem. 5
Finally, let us describe the new predicates that are added to the extended language.
There is an identity predicate for the abstracta, which we write as '= 1 ' to indicate
that this identity predicate is distinct from the identity predicate '=' defined on basic
objects. Other new atomic predicates can be added as well. These are derived from
certain open formulas of the base language, much as the direction orthogonality
predicate J.. * is derived from the corresponding predicate J.. for lines. Let us be
more precise. Consider a formula rp(xI> ... ,xn) that does not differentiate between
rv-equivalent objects and that only holds of objects in the field of rv:
(6.1) X1 ""Xi/\ ... /\ Xn x;, ~

rv (rp(x1,. . .,Xn) ++ rp(X~, .. .,X~))
{6.2) rp(XI> ... , Xn) ~ X1 ""X1 /\ . .. /\ Xn rv Xn
(Recall that rv is said to be a congruence with respect to <P when (6.1) holds.) For each
such formula rp, we have the option of introducing into .C 1 a new atomic predicate rp*
all of whose argument places are of the extended SQrt and which is defined by: 6
cp* (§xi> . .. , §Xn) ++ rp(XI> . .. , Xn)
5 As mentioned, the Caesar problem is discussed in Chapter 9.

6
Variants and generalizations are possible. One variant is to let <p* (y1, .. . , Yn) be non-atomic and merely
an abbreviation of3x1 .. . 3xn(y1 =§xi/\ ... /\Yn = §xn /\ rp(x1 •.. . ,Xn)). One generalization is to target
individual free variables of a formula rp that satisfy the analogues of (6.1) and introduce an associated atomic
predicate which takes an abstract term where <p has the basic variable x;.
6.3.2 Defining the translation

As observed, many formulas of .C 1 can be translated in a natural way into .Co. For
example, d(l1) = 1 d(l2) and d(l1) l. * d(l2) are naturally translated as 11 II 12 and
11 l. 12, respectively. We shall now define a systematic translation r, which generalizes
the familiar examples just rehearsed.
First, we define r on singular terms. This requires mapping the two different sorts
of singular term that are available in £ 1 to singular terms of the single sort available
in .Co. We do this as follows:
• r(x;) = x;
• r(y;) = x;
• r(§t) = t
Next, we translate atomic formulas as follows:
• r('P(to, . .. ,tn)') = 'P(r(to), ... ,r(tn))' where P is in Lo

• r('to =1 t1 ') = 'r(to),...., r(t1)'
• r(' rp*(to, .. . , tn) 1 ) = 'rp(r(to), ... , T(tn))'
Finally, we translate complex formulas as follows:
• r commutes with the truth-functional connectives; for example,

r(' .....,rp') = '.....,r(rp)'
• r('Vx;rp 1 ) = ' Vx;r(rp)'
• r('Vy;rp 1 ) = •vx;(x;,...., x;--+
r(rp))'
We can now state the proposal concerning innocent counterparts. For sentences,
the proposal is simple: the innocent counterpart of a sentence rp is r(rp). Suppose we
wish to provide innocent counterparts of open formulas as well. Then we need to
x;
ensure that the values of every pair of free variables y; and that occur in a formula
rp and its translation r(rp), respectively, are appropriately correlated. Thus, we claim
that r (rp) provides an innocent counterpart of rp relative to any variable assignment
that satisfies y; = §.i; for every such pair of variables.
6.3.3 The input theory

Is the translation r logically acceptable? Let us review the analysis oflogical accept-
ability provided in Section 5.2. First, the translation r must be recursive and commute
with negation modulo logical consequence. These requirements are clearly satisfied.
Then, r must preserve logical entailment in the following sense:7
(i) For any formula rp of .C 1 and set :E of such formulas, we have
if :E I== rp, then r(:E) F= r(rp),
where r(X) abbreviates {r(rp) I rp EX}.
7 Here it is important to recall that we are using a negative free logic. For definitions, see Appendix 2.B.
PREDICATIVE ABSTRACTION 101
In particular, any logical truth of the extended language must be mapped to a logical
truth of the base language.
Alas, the requirement that logical entailment be preserved is not satisfied. Consider
the logical truth that identity among "new" objects is symmmetric: §xo =1 §x1 ~
§x1 =1 §xo. This logical truth is mapped by r to the claim that ~ is symmetric:
xo "' x1 ~ x1 ~ xo. Suppose~ stands for parallelism. Then the latter claim-namely
that parallelism is symmetric-is not a logical truth. A related example concerns the
transitivity of identity. Yet another example concerns Leibniz's Law. Consider:
This logical truth is mapped to the statement that~ is a congruence with respect to cp:
While true, this statement is not a logical truth. What to do?

Let us begin by identifying the extent of the problem. Let the input theory I be the
.Co-theory whose axioms are the r-images of all the (free) logical truths of £ 1; that is:
I:= {r(cp): <pE.C 1 and f=cp}
Unless this theory can be presupposed as input to our abstraction, the translation r
may be guilty of mapping logical truths of the extended language to falsehoods of the
base language. Thankfully, the input theory admits of a useful characterization.
Proposition 6.1 The input theory is axiomatized by the formalization in £ 0 of the

following claims:
(i) ~is a partial equivalence relation (that is, symmetric and transitive)
(ii) for each predicate cp* of .C 1 , ~ is a congruence with respect to we have the universal closure of
<p(xo, ... ,Xn) ~ Xo ~ Xo /\ ... Xn ~ Xn
(This ensures that <p holds only of objects on which the relevant form of
abstraction is defined.)
A proof is provided in Appendix 6.A.
Having circumscribed the problem, we now need a solution. A hard-line response

would be to refuse to accept any non-logical presuppositions and therefore require
the input theory to be contained in our background logic. Hume's Principle and other
so-called "logical" abstraction principles would still pass muster. But we would lose
a number of seemingly good abstractions, including an abstractionist approach to
directions, geometrical shapes, and linguistic types.
Thankfully, a more moderate response is available. To see this, recall Rayo's formu-
lation of the requirement that the translation preserve logical consequence:
if 1/1 is a logical consequence of <p, then the truth-conditions assigned to <p must demand at least
as much of the world as the truth-conditions assigned to 1/J. (Rayo, 2013, p. 498, fn. I)
This suggests that we need not require that the input theory belong to logic, only that
it ''demand nothing of the world". Many abstraction principles are plausibly taken to
satisfy this requirement. Consider the case of directions, which is problematic on the
hard-line response. It is part of the nature of parallelism to be an equivalence relation
on directed items and to respect orthogonality. The input theory merely explicates
what lies in the nature of parallelism and is therefore plausibly taken to make no
demand on the world. Similar considerations apply to many other examples, such
as abstraction on geometrical shapes and linguistic types.
The moderate response is not entirely toothless, however, as it excludes any abstrac-
tion principle whose associated input theory makes a substantial demand on the
world. Here is an example. Let ,...., be the relation that holds between two objects just
in case they are painted with the same color of paint. Suppose that the heavy objects
and the light objects are painted red and green, respectively. This means that ,...., is
a congruence with respect to the properties of being heavy and light. But this is a
contingent truth, which makes a substantial demand on the world. It follows that
the moderate response does not allow the properties of being heavy and light to be
inherited from a colored object by the corresponding color abstract. However, this
prohibition seems entirely reasonable.
Let us therefore adopt the moderate response. This means revising clause (i) of
our definition of a logically acceptable translation (which was reproduced above)
so as to require preservation of logical consequence only modulo some suitably non-
demanding input theory I, that is:
(i') if :E F <p, then -r(:E), IF T{q.>).
All that remains is to prove that our translation T satisfies this revised requirement. Let
I be as above. Then the mentioned claim follows from Proposition 6.1. 8 We conclude
that the translation T is indeed logically acceptable-in our slightly revised sense.
6.3.4 The output theory
Any logically acceptable translation -r gives rise to certain truths that are obtained "for
free". Suppose -r maps <p to a theorem of the input theory I. Since <(J and -r(<p) make
the same demand on the world, and since I by assumption makes no demand on the
world, it follows that <p too makes no such demand; in other words, that <(J is obtained
for free. Let the output theory 0 be the set of all formulas of .C 1 that are obtained for
free in this sense-that is, {<(J E .C 1 : I r T( <(J)}.
What is the output theory 0 that results from our translation T relative to the input
theory I? To state the answer, we need a definition. 9
8 Suppose E I= tp. By completeness, E f- tp. So there are al> ... , an E E such that f- /\ ;a; ~ tp. Thus,
by the definition of I , we have If- r (/\ ;a; ~ l{J) and thus also If- /\; r (a;) ~ r (tp) (by the definition ofr).
This ensures r(E),I I= r (tp), as desired.
9 As before, we consider only the case of abstraction on objects. The case of abstraction on concepts is
analogous.
IMPREDICATIVE ABSTRACTION 103
Definition 6.1 Let the elementary theory of abstraction (ETA) be the .Ci -theory whose
axioms are the universal closures of
(ETAl) §xo = §xi ~ xo "" xi
(ETA2} cp* (§xo , ... , §xn) ~ <p(xo, . . . , Xn)
(Of course, (ETA2} is an axiom schema.) We can now answer the question that we set
ourselves.
Proposition 6.2 The output theory 0 can be axiomatized by the union of the axioms
of the input theory I and the elementary theory of abstraction ETA. In symbols,
0= I +ETA.
(A proof is sketched in Appendix 6.A.) In other words, provided we presuppose the
input theory I, our translation r gives us the theory ETA for free.
Let us take stock. We have described a two-sorted predicative form of abstraction
and defined a translation that maps each formulii of the extended language to an
innocent counterpart formulated in the uncontroversial base language. As Hale and
Wright emphasize, there appears to be no "metaphysical gap" between a sentence and
its innocent counterpart. For the abstraction to work, however, the theory I must
be presupposed in the base language. Given this presupposition, the abstraction
outputs a bunch of "free truths" in the extended language, which are axiomatized by
the theory ETA.
6.4 Impredicative Abstraction

It is undeniable that the restriction to predicative abstraction principles is severe.
So let us examine the prospects for the explanatory strategy just pursued when the
restriction is lifted-which, no doubt, was Frege's own intention.
Consider Hume's Principle, which is an impredicative abstraction principle widely
believed to be in good standing. Let .Co be some second-order language that is free of
arithmetical vocabulary; this will be our base language. Let the extended language be
the second-order language .C1 that results from adding to .Co a variable-binding car-
dinality operator# .10 We wish to define a translation r that maps each sentence of the
extended language to an innocent counterpart in the base language. In particular, the
translation should map each formula of the form #F = #G to a formula stating that
the Fs and the Gs are equinumous. Unfortunately, once the values of the free variables
are taken into account, the result of this translation is not guaranteed to be in the base
language. The reason is simple. Suppose that the values of the second-order variables
10
Notice that .C 1 , as currently defined, has a single sort of first -order expressions, unlike the extended
languages investigated in Section 6.3.
F and Gare specified by means of the open formulas cp(x) and 1/J(x), respectively. 11
Then our task is in effect to map the identity statement #x.cp(x) = #x.ijJ(x) to a formula
stating that the cps and the 1/Js are equinumerous. The problem is now apparent.
Since <P and 1/1 may contain occurrences of the cardinality operator'#: the mentioned
equinumerosity statement need not be in the base language.
This problem is a direct consequence of the impredicativity of Hume's Principle,
which includes the cardinal numbers obtained by the abstraction in the domain
of objects that we count. The predicative version of Hume, by contrast, introduces
"new" cardinal numbers to count objects from some antecedently given domain. It
is this predicativity that enabled us to provide the natural and appealing translation
discussed in the previous section.
Might an appropriate translation be defined even in the case of impredicative Hume
if we are only more careful? Regrettably, the answer is negative. It is easy to prove that
no recursive translation from £ 1 to £0 can ensure preservation of the intended truth-
values of the arithmetical sentences of £ 1 . 12 And preservation of truth-value is surely
a minimal condition of adequacy on any translation of the sort that we seek.
How might Fregeans respond to the non-existence of an appropriate translation?
I see three options. One is to invoke additional assumptions to ensure the desired
preservation of truth-values. But this is problematic. An easy modification of my
argument shows that any additional assumptions sufficient for this task would have
to ensure that the domain of the base language is infinite. If this is assumed, then
appropriate translations are indeed available.13 For Frege and his followers, however,
this response would be tantamount to admitting defeat. It is supposed to be a striking
consequence of the Fregean approach that there are infinitely many objects, not a
presupposition. Any such presupposition would provide lethal ammunition to Frege's
opponents, who could argue that this existential assumption is synthetic or at least
imposes a substantial demand on the world. 14
A second option is to deny that the desired translation r needs to be recursive.
But this too would be unpalatable to both Frege and his followers. To allow the
translation to be non-recursive would open the door to all kinds of trivializations of
11
That is, suppose that Vx (Fx # ip(x )) and likewise for the relation between G and 1/t. If the values of
the second -order variables cannot be specified in this way, but instead are specified as subsets of the basic
first-order domain, there is certainly no hope of obtaining a translation into the base language. See (Horsten
and Linnebo, 2016, pp. 5- 6) for further details and discussion.
12
Consider a one-element model for Lo. Truth of Lo-sentences in this model is obviously decidable.
A translation of the envisaged sort would thus provide a decision procedure for arithmetic. Since this is
impossible, there can be no such translation. (A similar but somewhat more involved argument is found in
(Rayo, 2008, fn. 9).)
13 See e.g. (Fine, 2002, II.5) and (Rayo, 2002).
14 What about letting the base language Lo be that of some nominalistic account of arithmetic (e.g. that
of (Hellman, 1989)), and then translating arithmetical sentences of L 1 as suggested by this account?
Technically, this would work, provided the nominalistic account works. But this approach would be
dialectically useless for the Fregean. If the nominalistic account works, why bother with an epicycle
concerned with the tricky idea of content recarving?
IMPREDICATIVE ABSTRACTION 105
Frege's project. For instance, why not simply translate each sentence <P of the language
of arithmetic to a tautology or a contradiction depending on whether or not <P is
true? Moreover, although Frege of course did not possess the concept of a recursive
procedure, there is at least some evidence that he wanted the translation to be given
by an algorithm. 15
The third and final option is to develop a Fregean approach that does not require
a systematic translation at all. Perhaps there is no need to specify an innocent
counterpart of every statement of the extended language. One possibility, explored
by the neo-Fregeans, is to regard the abstraction principle as an implicit definition. 16
To investigate this possibility, they consider the implicit definition of 'Jack the Ripper'
as the perpetrator of certain gruesome murders. This implicit definition functions as
an equation with one unknown, and our task is to solve for the unknown: 'Jack the
Ripper' is to refer to the object x such that xis the unique perpetrator of the specified
murders. For this definition to succeed, the world must cooperate to ensure that there
is a single perpetrator of all the murders. The success of the definition thus rests on
a substantial presupposition, which the world may or may not satisfy. Suppose an
abstraction principle is laid down as an implicit definition. Why shouldn't the success
of this definition too rest on a substantial presupposition? Were it to do so, this would
be very damaging. The neo-Fregeans wish to support our belief in the existence of
infinitely many numbers, not to presuppose it. If we are to use Hume's Principle as an
implicit definition, it must be a very special implicit definition, importantly different
from the example of Jack the Ripper.
In what way, then, are Hume's Principle and other suitable abstraction principles
special? The answer proposed by Hale and Wright turns on the idea that there is no
"metaphysical gap" between matching instances of the two sides of an abstraction
principle. Consider the following passage (also quoted in the preface to this book):
what it takes for 'the number of Fs = the number of Gs' to be true is exactly what it takes for the
Fs to be equinumerous with the Gs, no more, no less. (Hale and Wright, 2009b, p. 187) 17
This lack of "metaphysical gap" is meant to ensure that Hume's Principle and other
permissible abstraction principles avoid the problem that afflicts the implicit defini-
tion of 'Jack the Ripper'. Permissible abstraction principles avoid being "hostage to
conditions of which it's reasonable to demand an independent assurance" (Hale and
Wright, 2009a, p. 465). These principles do not give any "metaphysical hostages" and
"involve no ontological adventure'' (ibid., p. 477). This account of how abstraction
principles are special is certainly attractive. But the account is unavailable in the
15
See Section 7.5. More evidence can be gleaned from Frege's attempt in Grundgesetze I §§29-31 to prove
that every expression of his formal language has a unique reference. Although the proof raises some hard
exegetical questions, it should be uncontroversial that it is based entirely on effectively specified reduction
steps. See (Heck, Jr., 1997b), (Linnebo, 2004a), and (Heck, Jr., 2012) for discussion.
16 See (Hale and Wright, 2000).
17 The text has been adapted slightly to fit our present example.
present dialectical context, since it would reintroduce the need for a systematic
translation, which is precisely what the third option that we are exploring set out
to avoid.
Taking stock, this chapter has analyzed the appealing idea of closing "metaphysical
gaps" by means of an appropriate translation. But regrettably, the needed translation is
only available for predicative abstraction principles, not for impredicative ones, such
as Hume.
It might be objected that predicative abstraction is hopelessly weak. (Potter, 2010)
expresses the concern well: "Predicative abstraction principles [... ] are completely
harmless or, if you prefer, useless: they cannot serve as a foundation for mathematics"
(p. 187). My response is to iterate steps of predicative abstraction, as explained in
Chapter 3. Each step is admittedly small. But by taking a sufficiently large number of
small steps, we can reach as far as we wish.
Appendices
6.A Proofs
Proof sketch for Proposition 6.1. By reasoning as in our discussion immediately preceding the
statement of the proposition, it is easy to show that each of the proposed axioms is the r-image
of a (free) logical truth of £1. So the proposed axioms belong to the input theory.
Conversely, we prove by induction on the length of proofs that every (free) logical truth of
£ 1 is mapped by r to a logical consequence of the proposed axioms. Clearly, every tautology
of £1 is mapped by r to a tautology of £0. And every instance of modus ponens in £1 is
mapped to an instance of this rule in £0. Next, there are all the instances ofLeibniz's Law. Their
translations are all provable from the proposed axioms, as these prove that ~ is a congruence
with respect to the translation of each primitive predicate of £ 1. All that remain are now the
axioms and rules governing the quantifiers. It is straightforward to verify that these too are
mapped by r to theorems or derived rules of the proposed theory. -I
Proof sketch for Proposition 6.2. By considering the definition of 0, it suffices to prove
If- r(ifr) iff 0 +ETA 1-1/f
for each 1/F E £ 1. We do this by induction on the length of proofs. We use the axiomatization of
free first-order logic described in Appendix 2.B. The extension to a free two-sorted first-order
logic is straightforward (cf. e.g. (Enderton, 2001, Section 4.3)).
We begin with the right-to-left implication. Clearly, every non-logical axiom of I+ ETA is
mapped by r to a theorem of I. Next, tautologies are mapped to tautologies. Laws of identity
are mapped either to other such laws or to axioms of I . The axioms and rules governing the
quantifiers are easily seen to be mapped by r to theorems or derived rules of I. Finally, instances
of modus ponens are mapped to other such instances. The proof of the converse implication is
analogous. -I
7
The Context Principle
7.1 Introduction
The most influential strategy for articulating the idea of thin objects is arguably Frege's.
At the heart of his strategy is the context principle, which urges us "never to ask
for the meaning of a word in isolation, but only in the context of a sentence" {Frege,
1953, p. x). A version of this principle is also involved in my own approach to thin
objects. So I wish to discuss Frege's context principle and the explanatory work to
which he puts it.
The context principle poses some very hard interpretive challenges. Not only are the
ideas themselves hard, but relevant parts of Frege's view change in the course of his
career. The principle is without doubt one of the cqrnerstones ofFrege's Grundlagen.
It is announced in the introduction as one of three "fundamental principles" that
Frege has kept in mind throughout his inquiry. The principle plays an essential role
in the ensuing discussion as well, especially that of how numbers and other abstract
objects "are given to us" (Frege, 1953, § 62). By contrast, in Frege's magnum opus, the
Grundgesetze, the context principle figures less prominently. Indeed, there appear to
be weighty theoretical reasons to doubt that there is a role for the context principle
in the Grundgesetze. Here Frege subsumes the category of sentences under that of
proper names. A sentence is now defined as a proper name that refers to a truth-
value. Consequently, it appears no longer to make sense to say that sentential contexts
are privileged when it comes to explaining the meaning of a word. Furthermore, the
context principle appears to clash with the compositional semantics developed in the
Grundgesetze and other works from the same period. While the context principle
seeks to explain the meanings of subsentential expressions in terms of the meaning
of entire sentences, the compositional semantics of Frege's mature period involves the
opposite direction of explanation, where the meaning of any complex expression-
including a sentence-is explained in terms of the meanings of its components.
Finally, the Grundgesetze contains an extensive and vehement criticism of contextual
definitions. It is unclear how the context principle can escape this attack.
The purpose of this chapter is to examine the context principle and its role in
Frege's philosophical project, paying special attention to the question of whether
the principle provides a way to understand the idea of thin mathematical objects.
So while most of this book isn't concerned with Frege scholarship, this particular
108 THE CONTEXT PRINCIPLE
chapter is. After an initial examination of the principle in the Grundlagen, I argue
that-appearances notwithstanding- the context principle retains an important role
in the Grundgesetze. I also argue that reflection on this role sheds important new
light on the context principle. More specifically, the principle is fully compatible with
Frege's compositional semantics; it can and should be disentangled from the idea
of "recarving of content" and an ultra-thin conception of reference; and it suggests
a more lightweight conception of mathematical objects than is typically associated
with Frege.
While it cannot be denied that Frege's discussion of the context principle is under-
developed and partially flawed, I believe it merits our highest admiration. On my
analysis, Frege articulates the important philosophical problem of our linguistic and
cognitive "access" to abstract objects such as numbers and develops a promising new
answer based on the context principle. This contrasts favorably with other thinkers
(such as Dedekind) who indulge in loose talk about "the creation" of mathematical
objects, and with yet others (such as the early Russell and at times Godel) who invoke
a largely unexplained kind of apprehension of an abstract but mind-independent
reality. 1 Frege's use of the context principle is a heroic, if only partially successful,
attempt to face up to a very hard philosophical question, which so many other thinkers
have largely evaded.
Some remarks about terminology are needed before we begin. In the Grundlagen,
Frege never says much about the notion of meaning with which he is concerned but
treats the notion in an informal way. In later works, however, Frege obviously has a lot
to say about meaning. Beginning in the 1890s, he distinguishes sharply between sense
and reference, and develops his enormously influential theory of these notions and
their interrelation. In what follows, I first rely on a single informal notion of meaning,
much as in the Grundlagen. Later it becomes necessary to consider more precise
technical notions, in particular Frege's notions of sense and reference. To distinguish
these different uses, I reserve the words 'sense' and 'reference' for Frege's two technical
notions of Sinn and Bedeutung. The word 'meaning' will be used in part for the
informal notion of the Grundlagen and in part as a placeholder for a more precise
notion yet to be specified.
7.2 How Are the Numbers "Given to Us"?

Frege's context principle has received a bewildering range ofinterpretations. 2 I believe
the best way to understand the principle is by examining its contribution to the
work where it figures most prominently, namely the Grundlagen. Since the main goal
1 It is an interesting question, which I cannot discuss here, whether Frege's own late works, such as
(Frege, 1956), also belong in this group. I have in mind Frege's robust realism about thoughts and his striking
claims about our ability to grasp thoughts using "our power of thinking''.
2 See (Pelletier, 2001) for a wide range of examples.
HOW ARE THE NUMBERS "GIVEN TO us"? 109
of this work is to defend a logicist account of mathematics, I shall be particularly

concerned with the contribution that the context principle is supposed to make to
Frege's logicism.
At Grundlagen §62 (Frege, 1953) raises a question that has dominated much of the
recent philosophy of mathematics:
How, then, are the numbers to be given to us, if we cannot have any ideas or intuitions
of them?
The problem is, of course, that numbers, unlike tables and chairs, cannot be perceived;
nor can they be observed with the help of modern technology, as electrons and DNA
molecules can. How can we then refer to numbers and other kinds of abstract objects,
let alone gain knowledge of them? The question is perfectly reasonable, since the fact
that a singular term or representation succeeds in referring to another object external
to itself can hardly be primitive but must have some explanation.
Frege's next sentence proposes an answer to the above question:
Since it is only in the context of a sentence that words have any meaning, our problem becomes
this: To explain [erklaren] the sense of a sentence in which a number word occurs. 3
Frege here proposes that his context principle has an essential role to play in the
explanation of the meaning of individual words. In particular, since the meaning of
a singular term is closely related to-if not identical with-its reference, the context
principle is relevant to the question of how the numbers are "given to us". The idea
is to transform the hard problem of explaining how an arithmetical term succeeds
in referring into the easier problem of explaining the meanings of certain complete
sentences involving this term.
But is the latter problem really any easier than the former? When we explain the
meaning of a sentence, must we not also explain the reference of every singular term
involved in the sentence? If so, Frege will only have transformed one hard problem
into another that is at least as hard. Frege is fully aware of this worry and responds that,
when we explain the meaning of each of the relevant sentences, "we must reproduce
the content of this sentence in other terms, avoiding the use of" the singular terms
whose reference we are trying to explain (§62). Our task is thus to "reproduce" the
meaning of each sentence containing the problematic singular terms in a way that
avoids use of these terms. Frege provides a useful example of what he has in mind.
He suggests that the identity statement 'the direction of line a is identical with the
direction of line b' -which contains problematic singular terms purporting to refer
3
I have changed the translation of 'Satz' from 'proposition' to 'sentence: which is reasonable given that
Frege talks about words occurring in a Satz.
(Dummett, 1978, p. 38) characterizes this as "probably the most important philosophical statement Frege
ever made''. Later he uses even stronger words, describing it as "arguably the most pregnant philosophical
paragraph ever written" (Dummett, 1991a, p. 111). Although these are bold claims, it is not easy to think
of better examples.
to directions-has a content that can be "reproduced" by means of the sentence

'line a is parallel to line b' -which avoids the problematic direction terms and relies
only on the simpler phenomenon of reference to lines. 4
How might such a "reproduction of meaning" help us explain reference to a
direction? One would have thought that any account of reference to some object
would have to mention this object! But the proposed "reproduction of meaning" is
chosen precisely because it avoids any mention of directions. Of all the questions
surrounding the context principle, this one may well be the hardest-as will become
abundantly clear in what follows.
To sum up, Frege's response to the question of how the numbers are "given to
us" involves two key moves: first, the context principle, which transforms a question
about the reference of certain problematic singular terms into a question about the
meanings of complete sentences containing these terms, then, an answer to the latter
question based on the idea of"reproducing" the meaning of each sentence containing
the problematic singular terms in a way that avoids any use of these terms. The next
two sections consider each move in turn.
7.3 The Context Principle in the Grundlagen

The context principle receives four different statements in the Grund/agen, of which
the first and third (as ordered by their occurrence in the book) have already been
quoted:
CPL never ask for the meaning of a word in isolation, but only in the context of
a sentence. (p. x)
CP2. Only in a sentence have the words really a meaning. [... ] It is enough if
the sentence taken as a whole has sense; in this way its parts too obtain their
content. (§60)
CP3. Since it is only in the context of a sentence that words have any meaning,
our problem becomes this: to define [erklaren] the sense of a sentence in which a
number word occurs. (§62)
CP4. We next laid down the fundamental principle that we must never try to
explain [erklaren] the meaning of a word in isolation, but only as it is used in the
context of a sentence. (§ 106)
How should these passages be interpreted?

Many commentators take the context principle to consist, at least in part, in the
claim that individual words never have any meaning in isolation and that only
4
In the terminology of Section 5.2, Frege appears to be interested in a logically acceptable translation
from the potentially problematic direction language to the unproblematic line language.
THE CONTEXT PRINCIPLE IN THE GRUNDLAGBN 111
complete sentences are capable of possessing an independent meaning. 5 Thus, they

interpret Frege as endorsing the following principle:
A necessary condition for possession of meaning

If a subsentential expression e is meaningful, then e occurs in a meaningful sentence.
This interpretation is supported by the first half of each of CP2 and CP3. Further
evidence is provided by §60, where Frege writes that: "The self-subsistence which I am
claiming for number is not taken to mean that a number word signifies something
when removed from the context of a sentence:'
However, this necessary condition seems unreasonably strong. 6 To see this, let a
unit ofsignificance be the smallest linguistic string that has meaning in isolation. Then
the condition says that the only units of significance are complete sentences, never
individual words.7 Individual words have no meaning whatsoever when they occur
outside of sentential contexts. But this seems to contradict one of the most important
and least controversial principles of linguistics, namely that words have independent
lexical meanings that play an important role in determining the meanings of sentences
in which they occur.8
In fact, throughout most of his career, Frege was firmly committed to the idea that
words have independent meanings. This becomes abundantly clear in Frege's mature
work, where (as we shall see in Section 7.5) he endorses a principle of compositionality
based on exactly this idea. But already in his early work, we find passages sympathetic
to the idea. For example, in the unpublished article "Boole's Logical Calculus and
the Concept-script" from 1880- 1, Frege writes that in his logically perfect language
(Begriffsschrift) the "designations" of properties and relations
never occur on their own, but always in combinations which express contents of possible
judgment. I could compare this with the behavior of the atom: we suppose an atom never to
be found on its own, but only combined with others, moving out of one combination only in
order to enter immediately into another. (Frege, 1979, p. 17)9
This passage suggests a weaker and far more plausible view than that sentences are the
only units of significance. Words have independent "contents" or meanings, much like
5 Quine is an influential exponent of this interpretation; see e.g. (Quine, 1953a, p. 39); see also (Resnik,
1967), (Resnik, 1976) (who on p. 42 makes the even stronger claim that "sentences but not words have
meaning"), and (Janssen, 2001). Yet further examples are discussed in (Pelletier, 2001).
6
Here I follow a long tradition that includes (Dummett, 1956), (Dummett, 1981 a, ch. 1), and (Dummett,
l98lb, ch. 19).
7
Unless the individual word also functions as a sentence, sue~ as the Latin 'cogito' or the Spanish 'pienso'
(cf. (Dummett, 1956, p. 492)). Henceforth, this qualification will be tacitly assumed.
8
Of course, ambiguities need to be resolved, and this will often require information from the broader
context. Henceforth, we shall assume that all ambiguities have already been eliminated (as Frege too
assumed concerning his formal language).
9
Frege can be excused for not knowing that helium and other inert gases falsify his chemical "suppos-
ition·: as helium was isolated only in 1895. The exact empirical facts are anyway irrelevant to Frege's logical
point.
atoms have independent properties. However, these meanings cannot "occur on their
own" but only in the context of a complete judgment, much like atoms (typically)
cannot occur on their own but only bound together in a molecule.
Is it possible to reconcile this apparent acceptance of subsentential units of sig-
nificance with the passages where Frege seems to endorse the necessary condition?
One attractive possibility emerges when we raise an important question that has
so far been ignored. Are the mentioned passages concerned with expression types
or tokens? Our complaint that the condition is unreasonably strong tacitly assumes
that the expressions in question are tokens. An expression token-such as a word
scribbled on a piece of paper- can quite literally be moved in and out of sentential
contexts, while throughout this process contributing its fixed meaning to the resulting
range of sentential meanings. Now, suppose instead that the quoted passages are
concerned with expression types. Thus understood, the necessary condition states that
no subsentential expression type can be meaningful unless it occurs in a meaningful
sentence type. But since every expression type occurs in an infinite range of sentence
types, this is no longer such a demanding requirement! Indeed, provided that a
subsentential expression token is so much as capable of occurring in a meaningful
sentence token, presumably the corresponding expression type in fact occurs in the
corresponding sentence type, which is of course also meaningful. Thus, the proposed
interpretation would reconcile the two apparently conflicting pressures with which
we began. This provides at least some support for the interpretation.10
Regardless of one's view of this interpretation, however, it is clear that the necessary
condition cannot be all there is to the context principle. While some of the quoted pas-
sages suggest the necessary condition, others suggest other principles. In particular,
the second half of each of CP2 and CP3 suggest the following sufficient condition:
A sufficient condition for possession of meaning
If a subsentential expression e occurs in a meaningful sentence, or in some appro-
priate family of meaningful sentences, then e is meaningful.
In fact, this sufficient condition makes far more sense in the dialectical context in
which the context principle occurs. As we saw in the previous section, the principle
is invoked in order to transform a hard problem concerning reference to numbers
and other abstract objects into an easier problem concerning the meanings of certain
complete sentences. This explanatory strategy requires that the meaning or reference
of a singular term can be explained by observations concerning the meaning of certain
complete sentences. This move-from sentential meaning to meaning or reference
of singular terms-can be supported by the sufficient condition. By contrast, the
necessary condition-even on its more plausible type-based interpretation-is plainly
irrelevant to the mentioned explanatory move and the broader dialectical context in
1
° Further support becomes available in the Grundgesetze, as explained in note 22.
THE CONTEXT PRINCIPLE IN THE GRUNDLAGEN 113
which the passages occur. 11 I conclude that there is at least as strong a case for taking
Frege to endorse the sufficient condition as the necessary condition.
In fact, Frege might have taken both conditions to follow from a single underlying
idea. One reason to think so concerns a remarkable feature of CP2 and CP3, which is
rarely noticed. Both passages move unannounced from what appear to be statements
of the necessary condition to a conclusion that requires the sufficient condition.
Consider, for example, CP3. Why should it follow from the necessary condition that
"our problem becomes [... ] to define [erklaren] the sense of a sentence in which a
number word occurs"? If anything, this conclusion requires the sufficient condition,
not the necessary one! So unless Frege is relying on some broader idea of contextuality,
the arguments in CP2 and CP3 would be obvious non sequiturs.
Another reason concerns the two remaining passages, namely CP 1 and CP4, which
in fact express a broader idea of contextuality. 12
Sentence-level explanation of meaning

The meaning of a subsentential expression e can only be explained in terms of its
contribution to the meaning of sentences in which e occurs.
I use the word "explain'' to echo Frege's "erklaren''. While it is somewhat opaque what
kind of explanation Frege had in mind, the basic idea is tolerably clear. A meaningful
expression makes some systematic contribution to the meanings of sentences in which
the expression occurs. But this contribution cannot be fully characterized on its own
but only in terms of its effect on the resulting sentential meanings. It is useful to think
of the meaning of the subsentential expression as a capacity to affect or contribute
to the meaning of sentences in which the expression occurs. This capacity cannot
be fully explained in isolation but only in terms of its effect on sentential meaning.
As Frege suggests in §59, the word 'only' provides a good example. The word clearly
contributes in a systematic way to the meanings of complete sentences. But it makes
no sense to try to explain or characterize its meaning in isolation, without any mention
of complete sentences.
I claim that the explanatory principle coheres well with-and probably even
supports-the necessary and sufficient conditions for possession of meaning. Sup-
pose, with the explanatory principle, that for a subsentential expression to be mean-
ingful is for it to be capable of making a systematic contribution to sentential meaning.
Thus, if a subsentential expression e occurs in a meaningful sentence, or in some
appropriate family of meaningful sentences, then e has the mentioned capacity and
thus is meaningful. This yields the sufficient condition. Next, if the expression e is
meaningful, then its capacity to contribute to sentential meaning must be capable of
11 The same goes for the statements of the context principle that are found in the Grundgesetze, as we
shall see.
12 Cf. (Burge, 2005, p. 15), who distinguishes three interpretations of the context principle that are close
to the ones articulated here.

manifestation. If the expressions in question are types, not tokens, then this ensures
that e in fact figures in meaningful sentences. This yields the necessary condition.
In sum, while Frege's statements of the context principle are not entirely clear,
I have argued that the primary idea is an explanatory principle to the effect that
the meaning of a subsentential expression can be properly explained only in terms
of its contribution to the meaning of complete sentences. The statements of the
context principle also suggest a necessary and (especially) a sufficient condition for
possession of meaning. But reasonable versions of these conditions are supported by
the primary idea.
7.4 The "Reproduction" of Meaning

Let us now take a closer look at Frege's idea of"reproducing" the meaning of a sentence
containing singular terms that purport to stand for numbers (or some other kind of
abstract objects) in a way that avoids using such terms. By §62 of the Grundlagen,
Frege has already argued that number words stand for "self-subsistent objects''. "And
that is enough to give us'', he says, "a class of sentences" whose meanings must be
explained, namely identity statements involving number words (Frege, 1953, §62).
He continues:
In our present case, we have to explain the meaning [Sinn] of the sentence 'the number which
belongs to the concept Fis the same as that which belongs to the concept G'; that is to say, we
must reproduce the content of this sentence in other terms, avoiding the use of the expression
'the number which belongs to the concept F'.
In fact, the "number words" in question have a very specific form. Earlier in the
Grundlagen, Frege argued that numbers are ascribed to concepts; for example, the
statement 'there are eight planets' should be analyzed as an ascription of the number
eight to the concept is a planet. The basic number words are therefore taken to be of
the form 'the number of Fs', where 'F' is a concept term; I shall abbreviate these terms
as '#F'. Thus, Frege takes our initial task to be to explain the meanings of numerical
identity statements of the form '# F = #G'.
What about the meanings of sentences other than identities? Do these too have to
be explained? There is evidence that Frege thought so. For in §65, he comments on
the need to "define" all other assertions about the relevant objects, not just identities.
Moreover, in the "generalized context principle" of the Grundgesetze, to be discussed
shortly, Frege attaches no special significance to identity statements but holds that,
in order to ensure that a singular term refers, all sentences in which the term occurs
must be assigned truth-conditions.
As we have seen, the desired explanation of meaning is one that "reproduces" the
content or meaning of a sentence in other, and less problematic, terms. Let us call this
the elimination requirement. When we "reproduce" the meaning of a basic numerical
identity, for example, we must avoid any use of the number terms that flank the
THE "REPRODUCTION" OF MEANING 115
identity sign. Can the elimination requirement be met? In particular, can we explain
the meaning of basic numerical identities without any use of number terms? Frege
believes that we can. His argument is based on a principle that he attributes to Hume-
and which has consequently become known as Hume's Principle 13 -namely that the
number of Fs is identical to the number of Gs if and only if the Fs and the Gs can be
one-to-one correlated. We can formalize this as follows:
(HP) #F = #G B- F ~ G,
where F ~ G is some second-order formalization of the claim that there is a relation

that one-to-one correlates the Fs and the Gs. 14 Frege's proposal is that instances of
the right-hand side of (HP) can be used to explain the meaning of corresponding
instances of the left-hand side. And this explanation appears to satisfy the elimination
requirement.
In fact, the explanation seems to generalize to other abstraction principles as well,
that is, to principles of the form:
(AP) §a = §{3 B- a ~ fJ,

where a and f3 are variables, § is an operator that takes such variables to singular terms,
and~ is an equivalence relation on the kind of entities that a and f3 range over. Frege's
claim about "reproduction of meaning" seems to generalize as well. An instance of the
right-hand side can be seen as "reproducing" the meaning of the matching instance
of the left-hand side in a way that also appears to satisfy the elimination requirement.
Unfortunately, things are not as simple as they appear. It is true that there is no
explicit use of number terms on the right-hand side of (HP). Even so, there is an
implicit reliance on such terms. Recall that an abstraction principle is said to be
impredicative when its right-hand side quantifies over the objects denoted by its
left-hand side (cf. Section 6.2). For example, (HP) is impredicative because the first-
order quantifiers on its right-hand side have numl>ers in their range. Although the
right-hand side of (fi:P) makes no explicit use of number terms, it uses quantifiers
whose permissible instances include such terms. I argued in Section 6.4 that this
implicit reliance on number terms violates the elimination requirement. 15 Let us
therefore sidestep the complication posed by impredicative abstraction principles and
13
Although the principle was only much later put to systematic use by mathematicians, in particular
Georg Cantor.
14
Hume's Principle is now known to have an amazing mathematical property. Call the second-order
theory with (HP) as its sole non-logical axiom Frege Arithmetic. Then a technical result known as Frege's
Theorem says that Frege Arithmetic, along with some very natural definitions, allows us to derive all of
second-order Dedekind-Peano Arithmetic. This result is hinted at in (Parsons, 1965) and explicitly stated
and discussed in (Wright, 1983). For a nice proof, see (Boolos, 1990).
15
Our present question-whether impredicative abstraction is compatible with the elimination
requirement-must not be confused with the more general question about the permissibility of impredica-
tive abstraction principles. On the latter question, see e.g. (Dummett, !991a) and (Dummett, 1998), who
finds such principles problematic, and (Wright, 1998a) and (Wright, 1998b), who does not.
test the ideas associated with the context principle on predicative principles, where
the complication does not arise.
Our next question is how Frege understands the desired "reproduction'' of mean-
ing. He makes some interesting remarks in §64, in the context of discussing the
example of directions. Since two directions are identical just in case the lines
whose directions they are, are parallel, we get the following abstraction principle for
directions:
(Dir) d(a) = d(b) +* a II b
where a and b range over directions or other directed items. Frege writes: 16
Thus we replace the symbol II by the more generic symbol =, through removing what is specific
in the content of the former and dividing it between a and b. We carve up the content in a way
different from the original way, and this yields us a new concept. (Frege, 1953, §64)
The suggestion is that each instance of the right-hand side of (Dir) "recarves" the
content of the corresponding instance of the left-hand side. That is, the two sides share
a common content or meaning, which they "carve up" in different ways. 17 The idea
of "recarving" raises some difficult questions, as we have seen in Chapter 4 and will
see again in Section 7.6.2. Moreover, there are exegetical questions about the depth
and strength ofFrege's commitment to the idea. As is well known, Frege abandons the
approach based on recarving in §68 of the Grundlagen in favor of an explicit definition
of numbers in terms of extensions. The context principle is nevertheless reiterated
as a "fundamental principle" at the end of the book (§106). This suggests that Frege
already in the Grundlagen is more strongly committed to the context principle than to
the idea of recarving. 18 As we shall see shortly, this differential commitment becomes
indisputable in the Grundgesetze.
Let us take stock of our discussion so far in this chapter. We have obtained a good
understanding of the architecture of Frege's strategy in the Grundlagen for explaining
how numbers are "given to us''. The strategy relies on two main ingredients: the context
principle and the idea of"reproducing" meaning in other, and less problematic, terms.
These ingredients are supposed to work together in tandem. While the overall strategy
is thus fairly clear, the details of its execution are less so. Our next task is to examine
whether the context principle survives into the Grundgesetze and, if so, whether this
can help us attain a better understanding of how Frege's strategy was supposed to
be executed.
16 See also the passage from §65 quoted on p. 77.

17 This "recarving thesis" is also endorsed by Bob Hale and Crispin Wright in (Wright, 1983) and many
of the articles collected in (Hale and Wright, 2001a), esp. essay 4 (originally (Hale, 1997)) but also essay 5,
pp. 149- 50; essay 8, pp. 192-7; and essay 12, pp. 277- 8. However, in more recent work by Hale and Wright,
the recarving thesis plays a less central role; see e.g. (Hale and Wright, 2000) and (Hale, 2007).
18 Thanks to Philip Ebert for this suggestion.
THE CONTEXT PRINCIPLE IN THE GRUNDGESETZE 117
7.5 The Context Principle in the Grundgesetze

Many commentators claim that Frege gave up the context principle at some point
between the Grundlagen and the Grundgesetze. 19 Three arguments for this claim were
identified in the introduction. The purpose of this section is to show that all three
arguments fail, and that, on the contrary, the context principle retains an essential
role in the argument of the Grundgesetze.
The first argument concerns Frege's subsumption of the category of sentences
under that of proper names. In the Grundlagen, Frege still adheres to the more
traditional-and plausible-view that sentences and proper names are two separate
logical categories. The context principle relies on this fact, as it asserts an explanatory
priority of the former category over the latter. By contrast, the Grundgesetze regards a
sentence as merely a complex proper name that denotes a truth-value. This obliterates
the earlier distinction between sentences and proper names on which the earlier
version of the context principle was based. It is thus unclear whether the principle
can even be formulated in the context of the Grundgesetze.
However, a principle very much like the original context principle remains in the
Grundgesetze. For example, in §97 of volume II we read: "One can ask after reference
only where signs are components of propositions [Satze] expressing thoughts:' As
this passage and a number of others illustrate, the subsumption of sentences under
proper names in no way undermines Frege's notion of a thought or his emphasis on
judgment. 20 The most important evidence is found in Grundgesetze I §§29-31, where
Frege makes heavy use of a contextual account of reference, sometimes known as the
"generalized context principle''. This account lays down the following criterion for a
proper name to refer: 21
A proper name has reference if, whenever it fills the argument places of a referential name of
a first-level function with one argument, the resulting proper name has a reference, and ifthe
name of a first-level function with one argument which results from the relevant proper name's
filling the~ -argument-places of a referential name of a first-level function with two argument
places, always has a reference, and if the same also holds for the ~-argument-places.
(Frege, 2013, I, §29; emphasis in original)
19 See (Resnik, 1967) and (Resnik, 1976) for expressions of this view. (Janssen, 2001, p. 133) makes
the slightly weaker claim that after the Grundlagen, "the context principle was not repeated in [Frege's]
published writings" (emphasis in original). See (Kiinne, 2010, p. 595) for a similar claim, as well as (Pelletier,
200 I) for further examples. Some of Dummett's writings may suggest that he agrees; see e.g. (Dummett,
!98!a, pp. 7 and 495). However, the relevant passages are more plausibly read as observing that, as of the
1890s, Frege can no longer regard sentential contexts as privileged because he now thinks that sentences
are a special kind of proper names. Despite this, Dummett claims that Frege "retains the context principle
[... ] as far as can be done without distinguishing between sentences and other complex proper names"
(Dummett, 198!b, p.409). See also (Dummett, 198lb,ch. 19), (Dummett, 1991a, ch. 17), (Dummett, 1991b,
pp. 229-33), and (Dummett, 1995).
20 Rather, what has happened is that the distinctive characteristic of sentences-namely their role in
judgment- has been separated out and located in Frege's judgment stroke, where 'I-- il' indicates the
judgment that il is the True. See (Rumfitt, 2011).
21 See (Heck, Jr., 1997b), (Heck, Jr., 2012, chs. 3 and 5), and (Linnebo, 2004a) for further discussion.
Although the passage is highly compressed, the basic picture is tolerably clear. With
the subsumption of the category of sentences under that of proper names, the context
principle becomes a principle to the effect that a proper name is referential provided
that all larger contexts in which the name occurs are themselves referential-including
contexts that belong to the category of proper names. So the central idea of the
context principle remains unchanged, namely that a proper name refers provided
that it contributes appropriately to the meanings of all larger contexts in which
it occurs. 22
The quoted criterion for a proper name to refer appeals to a notion of a function
name's referring, for which Frege provides the following criterion:
A name of a first-level function with one argument has a reference [ . . . ] if the proper name
which results from this function-name when the argument places are filled by a proper
name always has a reference provided the inserted name refers to something. (ibid.)
Frege also states criteria for names of first-level functions of two arguments to
refer and for names of second- and third-level functions. These criteria are obvious
modifications of the two just quoted and need not be reproduced here. 23
Clearly, to use these criteria of referentiality, we need a way to verify that a complex
name refers that does not depend on a prior verification of the reference of its parts.
How might this be possible? Frege's attempted "proof of referentiality" in §§29-31
reveals what he had in mind. The idea of the proof is to reduce a complex name to
another name that is guaranteed to corefer, if it refers at all, but whose referentiality
can be established more easily. A nice example is provided by Frege's horizontal
function -~, which is defined to map the True to the True and everything else
to the False. A moment's reflection shows that this name is guaranteed to corefer
with the complex name'~ = (~ = ~)'. Thus, if the name of the identity function
refers, so does that of the horizontal function. Indeed, Frege shows that the proof
of referentiality comes down to verifying the referentiality of all identities involving
value-range terms from his formal language. And this is meant to follow from another
reduction of the prescribed sort, this time involving a metalinguistic version of
Basic LawV:
22 In Section 7.3, I asked whether the expressions with which the context principle is concerned should
be understood as types or tokens. I provided some evidence for the former understanding. Further evidence
is provided by the generalized context principle, which becomes unstable unless this understanding is
adopted. Suppose more token names of first-level functions become available as time goes by. This may
cause a proper name that was initially deemed referential to no longer be so.
23 Note the circular dependence of the two quoted criteria: each makes ineliminable reference to the
other. For discussion, see (Linnebo, 2004a), where this problem is referred to as "the old circularity'; as it
also affects the context principle from the Grundlagen. It is relevant, in this connection, that Frege notes
that his criteria "are not to be regarded as explanations of the expressions 'to have a reference' or 'to refer
to something; since their application always presupposes that one has already recognised some names as
referential; but they can serve to widen the circle of such names gradually" (Frege, 2013, I, §30). I believe
this stepwise use of Frege's criteria corresponds to a sequence of expansions of one's linguistic resources of
the sort to be discussed in Section 7.6.3.
THE CONTEXT PRINCIPLE IN THE GRUNDGESETZE 119
I use the words

"the function (~) has the same value-range as the function \jl (~) "
throughout as co-referential with the words

"the functions (~) and ljl(~) always has the same value for the same argument:'
(Frege, 2013, I, §3; emphasis in original)
The importance of this passage is demonstrated by the fact that it opens the section of
the Grundgesetze where value-ranges are introduced. But Frege's proof of referentiality
must be flawed. For by Russell's paradox, Frege's stipulations cannot ensure that every
name of his formal language is assigned a unique reference.
Not all is lost, however. Our investigation it1- Chapter 6 revealed that there
is no problem provided that we restrict ourselves to the predicative version of
Basic Law V. 24
As we have seen, the context principle was part of Frege's attempt in the Grund-
lagen to explain our cognitive "access" to numbers and other abstract objects. The
same philosophical concern arises in the Grundgesetze, as the following passage
makes dear: 25
If there are logical objects at all-and the objects of arithmetic are such-then there must also
be a means to grasp them, to recognise them. (Frege, 2013, II, §147)
Frege's attempted explanation too has the same structure as in the Grundlagen.
Speaking of the "transformation" afforded by the metalinguistic version of Basic Law
V that we quoted above, the above passage continues:
The basic logical law which permits the transformation of the generality of an equality into an
equality serves for this purpose. Without such a means, a scientific justification of arithmetic
would be impossible. For us it serves the purposes that other mathematicians intend to achieve
by the creation of new numbers.
Let us assemble all the pieces. In both of his main works, Frege recognizes the
need to explain how we grasp abstract objects such as numbers. And in both works,
he attempts to meet this challenge by providing a two-step explanation of the sort
we identified in the Grundlagen. First, the context principle is used to transform the
problem of explaining the reference of a class of problematic proper names to that of
explaining the meanings of more complex expressions in which these names occur.
Then, this latter problem is to be solved by showing how the meanings of these more
24
Other fragments can be salvaged as well. (Horsten and Linnebo, 2016) show how a version of
the groundedness idea guiding (Kripke, 1975) can successfully be brought to bear on the impredicative
version ofBasic Law V. Unfortunately, this method works only when the background second-order logic is
predicative. (Dummett, 199 la, ch. 17) is thus half right when he characterizes impredicativity as the serpent
in Frege's paradise. Although each of the two kinds of impredicativity (cf. Section 6.2) can be handled in a
satisfactory way, it is impossible to do so for both kinds simultaneously.
25
A similar passage is found at the very end of the postscript that Frege added to the book, where this
is characterized as the "Urproblem" of arithmetic.
complex expressions can be "reproduced" without any use of the mentioned class of
names. As one would expect, given the developments in Frege's view of semantics,
the exact nature of each step differs between the two works. In the Grundlagen, the
first step amounts to the task of explaining the meanings of sentences in which the
problematic proper names occur, whereas in the Grundgesetze, we face the task of
explaining the reference of more complex names containing the problematic proper
names as parts. Likewise, the "reproduction'' of meaning involved in the second step is
understood differently in the two works. In the Grundlagen, the "reproduction'' is one
of"recarving", that is, of expressing the very same meaning in a different way; whereas
in the Grundgesetze, it is simply a matter of reducing a complex expression to a simpler
yet co-referring one, as illustrated by the passage quoted from Grundgesetze I, §3. The
following table provides a summary (where the arrow..,,,., indicates a reduction of one
problem to another).
Step I: Transformation of the problem

Grundlagen reference of proper names meaning of sentences
Grundgesetze reference of proper names reference of more complex names
Step 2: "Reproduction" of meaning that avoids the problematic names

Grundlagen "recarving" of meaning
Grundgesetze fix the reference of a complex name without going via its parts
I turn now to the second of the three arguments aimed to show that the context
principle is incompatible with Frege's view in the Grundgesetze. This argument is
concerned with the compatibility of the context principle with the principle of
compositionality, which figures prominently in Frege's middle and later works. The
semantic theory developed in this period has at its core a two-place relation of
reference that holds between a linguistic expression and the worldly item that serves
as its Bedeutung or semantic value. 26 One of Frege's great innovations is to provide
a systematic account of how the semantic value of a complex expression is related to
the semantic values of its subexpressions. In particular, he formulates the principle
of compositionality, which says that the semantic value of a complex expression is
determined as a function from the semantic values of its immediate parts. According
to Frege, the syntactic operation of predication corresponds to the semantic operation
of function application. So the semantic value of the simple predication 'John runs' is
obtained by applying the semantic value of'runs' to the semantic value of'John'. Ifwe
write [ell for the semantic value of an expression e, then the analysis of the simple
predication can be formalized as follows:
26
Strictly speaking, we need a family of reference relations, one for each of Frege's fundamental
ontological categories. For example, the relation of reference that obtains between a proper name and the
object to which it refers is not the same as that which obtains between a monadic first -level predicate and
the first-level function of one argument. For ease of exposition, I henceforth suppress this complication.
THE CONTEXT PRINCIPLE JN THE GRUNDGESETZE 121
(7.1) ['John runs'] = ['runs'](['John'])

Frege endorses an even stronger compositionality principle for senses. When one
expression is a syntactic part of another, then the sense of the former is claimed to
be a part of the sense of the latter. 27
We now face a puzzle. How can the context principle be reconciled with the
principle of compositionality? After all, the two principles call for opposite forms of
explanation. The principle of compositionality has a bottom-up character: the seman-
tic properties of complex expressions are to be explained in terms of the semantic
properties of their parts. By contrast, the context principle has a top-down character:
a key semantic property of proper names-namely their reference-is to be explained
in terms of the semantic properties of more complex expressions in which the relevant
names occur.28
In fact, the puzzle has a solution. There would indeed be a conflict if, as some
commentators claim, the cont~t principle denied that individual words can be
units of significance. For when words compose to form a complex expression, then
compositionality, as ordinarily understood, tells us that the meaning of the latter
can be determined on the basis of meanings antecedently possessed by the former.
But the mentioned interpretation of the context principle is not well supported.
As argued above, the purpose of the principle, as it figures in the Grundlagen, is
not to deny that words can be units of significance but rather to insist that their
meaning or significance can be properly explained only in terms of their contribution
to the meanings of complete sentences. Moreover, the mentioned interpretation lacks
any support from the Grundgesetze, where the context principle figures solely as a
sufficient condition for possession of meaning and as an explanatory principle.
Might either the sufficient condition or the explanatory principle conflict with the
principle of compositionality? There is certainly nothing to fear from the sufficient
condition, which in fact goes hand in hand with compositionality. For unless an
expression e has a semantic value, one could not in a compositional way obtain a
semantic value for any complex expression in which e occurs. But let us take a closer
look at the explanatory principle, which, we recall, states that the meaning of certain
expressions can only be explained in terms of their contribution to the meanings of
more complex expressions. There is no direct conflict between this principle and the
principle of compositionality, which states that the meanings of complex expressions
are determined by the meanings of their simpler constituents. While the former
principle concerns the explanation of meaning, the latter concerns the determination of
27 See for instance (Frege, 2013, I, §32) (quoted on p. 126) and (Frege, 1963).
28
The alleged conflict is nicely summarized by (Resnik, 1976, p. 47}, who writes that with composition-
ality, "[t]he meaning of the sentence rather than bestowing meaning upon the words becomes a function
of the meanings of the words.'' See also (Pelletier, 2001} for a wealth of examples of thinkers who believe
there is at least a prima facie tension between the two principles.
meaning. 29 The only worry is that some indirect conflict might emerge once the top-
down explanation promised by the context principle is properly spelled out. Although
this is theoretically possible, no concrete reason for suspicion has been provided-or
emerges when we shortly consider some attempts to spell out the details.
The final argument that Frege cannot have retained the context principle in the
Grundgesetze is concerned with his scathing criticism of creative definitions, espe-
cially in Grundgesetze II, §§138-46. Since the targeted definitions include what have
become known as contextual definitions, it is natural to think that Frege's criticism
must have required him to reject the context principle. Some relevant text occurs at
the end of Frege's discussion of creative definitions, which concludes that it "has thus
become plausible that creating proper is not available to the mathematician" (§146).
In the very next sentence, Frege considers an objection:
Against this, it could be pointed ouHhat in the first volume (§3, §9, §10) we ourselves created
new objects, namely value-ranges. What did we in fact do there? Or to begin with: what did
we not do? We did not list properties and then say: we create a thing that has these properties.
Rather, we said: if one function [ ... ] and a second function are so constituted that both always
have the same value for the same argument, then one may say instead: the value-range of the first
function is the same as the value-range of the second. That [ .. . ] we can convert the generality
of an equality into an equality (identity), must be regarded as a logical basic law. This conversion
is not to be taken as a definition; neither the words "the same'; nor the equality-sign, nor the
expression "value-range'; nor a combination of [value-range signs], nor both at the same time,
are thereby explained. (Frege, 2013, §146)
I believe the "conversion" that Frege has in mind is the metalinguistic version of
Basic Law V. (Had it been the ordinary version of the law, there would have been
no possibility of confusing it with a definition.) This prompts the question of why
the conversion isn't an illegitimate creative definition. The very fact that Frege raises
this question shows that he is aware of the apparent tension between his criticism
of creative definitions on the one hand, and the context principle and its use of the
mentioned conversion on the other. His answer is that the conversion isn't a definition
at all, and that the apparent tension therefore isn't genuine. For our present purposes, it
is immaterial whether the answer is a good one. What matters is that Frege identifies
the question and proposes an answer. This suffices to undermine the thought that
Frege's criticism of creative definitions must have forced him to reject the context
principle. This completes my response to the final argument. 30
29 Compare (Dummett, 1981 a)'s response to the alleged conflict, which is based on a distinction between
"the order of explanation" of meaning and "the order of recognition" (pp. 4-7) .
30 Although not required by my response, it is hard to resist asking what the status of Frege's conversion
is if not a definition. We are told that it "must be regarded as a logical basic law''. But this answer is not very
helpful: only the ordinary version of Basic Law V, not its metalinguistic version, can plausibly be regarded
as a "logical basic law''. We shall return to the question shortly.
DEVELOPING FREGE'S EXPLANATORY STRATEGY 123
To sum up, I have argued that both the context principle and the explanatory
project in which the principle figures survive into the Grundgesetze, subject only to
the changes required by other developments in Frege's view, especially concerning
semantics.
7.6 Developing Frege's Explanatory Strategy

So far, I have focused on the architecture of the explanatory strategy associated
with the context principle in Frege's two main works. It is time to delve into the
nuts and bolts, both to understand how Frege wanted to arrange them, but also to
investigate whether the strategy might in fact succeed. I shall consider three attempts
to develop the explanatory strategy, two inspired by the Grundlagen, and a third by
the Grundgesetze. Since the third strategy is the most promising, this illustrates the
importance of not limiting our investigation of the context principle to its familiar
home in the Grundlagen.
Our investigation takes place in a setting which should by now be familiar. We begin
with an interpreted base language .C0 , which is assumed to be in good philosophical
standing. We then consider an extended language L1> the meanings of whose new
expressions are to be explained by means of the context principle. Our paradigm
example consists of a base language concerned with lines and other directed items, to
which an extended language adds vocabulary appropriate for talking about directions.
As before, we let .C 1 be a two-sorted language, with one sort for lines and another for
directions. This means that problematic "mixed" identities, involving one direction
term and one line term, cannot even be formed. So this simplification postpones the
Julius Caesar problem and allows us to focus entirely on our present task of inter-
preting the context principle. As observed, Frege seeks to "reproduce" the meaning
of each sentence of the extended language in a way that avoids using the vocabulary
whose meaning we are trying to explain. It is plausible to interpret this as an attempt
to "reproduce" the meaning of each sentence of the extended language by means
of a sentence of the base language. In fact, we know from Section 6.3 that there is
a translation r from .C 1 to .Co which satisfies all the logical constraints on content
recarving and has just as good a claim to effecting a "reproduction" of meaning as
Frege's own translation of 'd(/i) = d(/2)' as '11 II 12'.
The crucial difference between the three strategies that I consider concerns their
respective accounts of what the translation r achieves.
7.6.1 An ultra-thin conception of reference
The first explanatory strategy combines the context principle with the "ultra-thin"
conception of reference discussed in Chapter 5. According to this conception, an
expression refers provided that, first, it has all the appropriate syntactic and inferential
characteristics of a singular term, and second, the expression figures in appropriate
true sentences. Since the first requirement is spelled out in entirely non-semantic
terms, the only semantic requirement is the second one, which concerns whole
sentences. The result is an extremely undemanding conception of reference.
Such a conception of reference enables us to put the translation < to good use. We
let < establish relations of synonymy between the sentences of the extended language,
whose meanings we are trying to fix, and the antecedently meaningful sentences of
the base language. We do this by regarding each sentence <P of the former language as
synonymous with the sentence 't" ( <P) of the latter. 31 For example, each sentence of the
direction language is assigned a meaning that is specified by means of an appropriately
chosen sentence of the line language. We now recall that our notion of a logically
acceptable translation was crafted precisely so as to ensure that the requirements
associated with the ultra-thin conception are satisfied. Thus, when the translation
is used as just described, the ultra-thin conception of reference entails that all the
singular terms of the extended language refer. This shows how directions and other
abstract objects are "given to us'; or so the argument goes.
Does this explanatory strategy succeed? Regrettably, the account faces problems, as
explained in Chapter 5. Let me briefly review the most fundamental one, which is of
both exegetical and systematic significance. By the time of the Grundgesetze, Frege had
developed a substantive technical and philosophical theory of sense and reference.
At the heart of this theory is a compositional theory of reference, which ascribes to
each expression e a semantic value, which is its contribution to the determination
of the semantic value of more complex expressions in which e occurs, including, of
course, the truth-values of sentences containing e. As Dummett observes, however,
when the translation is used in the way just outlined, we entirely bypass the ascription
of semantic values to subsentential expressions. Instead, we assign meanings only to
whole sentences. The singular terms of the extended language are thus "semantically
idle", as Dummett aptly puts it. Since compositionality is a serious commitment-
both for Frege and in contemporary semantics-the account under discussion faces
a problem whose significance is both exegetical and systematic.
7.6.2 Semantically constrained content recarving
What goes wrong with the previous strategy is that the syntactic constituents of a
sentence of the extended language are left without any direct semantic significance.
In particular, singular terms and the identity predicate do not make their ordinary
semantic contribution of standing for objects and the relation of identity, respectively.
Only whole sentences have been assigned a meaning.
In fact, there is reason to doubt that Frege had that strategy in mind. Discussing
the idea that one side of an abstraction principle might be taken "to mean the
same as" the other side, he remarks that the relation of identity must be "taken as
already known'' (Frege, 1953, §65). 32 It is reasonable to take this remark to express
31
The resulting interpretation was called "holophrastic reductionism" in Section 5.3.
32 The same view is held in the Grundgesetze; see (Frege, 2013, I, §§3 and 9).
a requirement that is semantic in character, not merely inferential, namely that the
identity predicate be given its ordinary semantic interpretation of standing for the
relation of identity. For this semantic interpretation to work, the singular terms
that flank the identity predicate must obviously be interpreted as referring to objects;
otherwise, the identity relation would have no relata.
We may regard this revised strategy as a semantically constrained form of content
recarving. Consider an abstraction principle. We want to assign the independently
available meaning of the right-hand side to the left-hand side-but subject to a
semantic constraint that was not present in the previous strategy, namely that the
identity predicate be given its ordinary interpretation. The hope is that the assignment
of meaning, subject to this added constraint, will suffice to determine the reference of
the relevant abstraction terms. Metaphorically, the idea is to pour the meaning of each
instance of the right-hand side into a mould provided by the logical form of the left-
hand side, and in this way to fix the reference of the relevant singular terms. I believe
this approach is roughly what Frege had in mind in Grundlagen §64 when he wrote
about "carving up" a content in different ways. 33
Unfortunately, this revised strategy too faces serious problems. Let me begin with
an exegetical point. While the interpretation of the context principle as concerned
with a semantically constrained form of content recarving arguably receives some
support from the Grundlagen , it does not fit the Grundgesetze. 34 This later work aban-
dons the idea of content recarving in favor of transformations that need only preserve
truth-value, while the context principle retains an important role (cf. Section 7.5). It
follows that the mature Frege must have understood the context principle in a way
that was not essentially tied to the idea of content recarving.
Setting exegesis aside, let us now ask whether the strategy might succeed in explain-
ing the reference of the abstraction terms. I believe the answer is negative, because the
approach imposes incompatible requirements on the notion of meaning. On the one
hand, the meanings in question must be coarse-grained enough for two sentences to
share the same meaning although they differ in syntactic-semantic structure. Let us
call this the requirement of recarving. On the other hand, the meanings must be fine-
grained enough to enable the reference of the singular terms occurring in a sentence
to be determined by the sentence's meaning and its syntactic-semantic structure. Let
us call this the requirement of reference determination .
It is easy to find notions of meaning such that one of the two requirements is met.
For example, the requirement of recarving is satisfied when "meaning" is understood
as Fregean reference or as a set of possible worlds. But this partial success is gained
33 Strictly speaking, Frege writes about "splitting up" ("zerspalten") the content of a relation symbol.
(Thanks to Philip Ebert and Bob Hale for discuss ion.) However, this "splitting up" must affect the content
of the entire sentence; otherwise it would not make any sense to talk about "dividing" the content of the
relation symbol between the two objects on which we abstract.
34 See (Ebert, 2016) for some interesting concerns about Frege's commitment to the idea of content
recarving even in the Grundlagen.
by adopting meanings so coarse-grained that we are unable to satisfy the other

requirement. Given only the truth-value of a sentence, there is obviously no way to
retrieve the referents of its subexpressions. Likewise, if the meaning of a sentence is
taken to be just a set of possible worlds, then the meaning of a sentence does not
determine the reference of its subsentential expressions. Conversely, it is easy to find
conceptions of meaning so fine-grained as to satisfy the requirement of reference
determination. A good example are Fregean senses, at least as understood in many
of Frege's later writings: 35
If a name is part of the name of a truth-value [i.e. part of a sentence], then the sense of the
former name is part of the thought expressed by the latter [i.e. by the sentence] .
(Frege,2013,I,§32)
Frege here claims that the sense of a complex expression contains the senses of its
subexpressions as parts.36 Since two wholes cannot be identical unless they share the
same parts, it follows that the thoughts or senses of matching instances of the two
sides of the direction abstraction principle cannot be identical unless the senses of
the direction terms are identical with the senses of the corresponding line terms. The
problem is that this would rule out any non-trivial recarving. Fregean senses therefore
violate the recarving requirement.
Can a notion of meaning be found that satisfies both the requirements of recarving
and reference determination? Such a notion of meaning would have to be intermediate
between Fregean reference and sense in its fineness of grain. In fact, it is likely
that Frege in the Grundlagen had in mind precisely such an intermediate notion
of meaning, namely one hinted at, but never properly developed, in his first book
Begriffsschrift from 1879. Here Frege characterizes the "conceptual content" of a
sentence in terms of its inferential relations to other sentences. When two sentences
or judgments are equivalent in that
the conclusions that can be drawn from one when combined with certain others also always
follow from the second when combined with the same judgements [... ] I call that part of the
content that is the same in both the conceptual content. Since only this has significance for
the Begriffsschrift, no distinction is needed between sentences that have the same conceptual
content. (Frege, 1879, p. 53; emphases in original)
By being intermediate between the later notions of sense and reference,.this notion of
conceptual content stands a better chance of satisfying our two requirements.
The most sophisticated attempt to develop an intermediate notion of meaning
that satisfies both requirements is due to (Hale, 1997), who builds on ideas from
35 Another example are "Russellian propositions': understood as set-theoretic tuples of the referents or
semantic values of the various parts of a sentence.

36 This view plays an important role in Frege's later work, where it is often invoked to explain the so-
called "productivity" of language, that is, our ability to produce and understand a potential infinity of
different sentences. See (Frege, 1963, p. 1) for one example.
the Begriffsschrift. But Hale's attempt has encountered problems, the most serious
of which appears to confirm my principal concern above, namely that no system of
meanings that are coarse-grained enough to satisfy the recarving requirement can
simultaneously be fine-grained enough to satisfy t~e requirement of reference deter-
mination. A version of this argument, due to (Fine, 2002, pp. 39- 41), is summarized
in Appendix 7.A. 37
7.6.3 Towards a metasemantic interpretation
The previous two explanatory strategies take their inspiration from the Grundlagen .
Let us now consider how Frege understood the context principle in the Grundgesetze.
I wish to begin with a problem that lies right at the heart of the context principle as
I understand it. Consider an abstraction term, say 'd(l); where '/' is the name of some
particular line. Suppose that doubt arises as to whether the abstraction term refers.
How can this doubt be assuaged? One option is to look to Frege's semantic theory,
which tells us that 'd(l)' refers to the direction of the line to which '/' refers. But this
specification of a referent involves an analogous abstraction term in the metalanguage,
which is just as vulnerable to the doubt in question. Frege's context principle is meant
to provide a more independent-and therefore more satisfying-way to assuage the
doubt. According to the generalized context principle, it suffices for an expression
to have reference that all appropriate contexts in which the expression occurs also
refer. In order to verify that 'd(I)' satisfies this sufficient condition, we invoke our
translation r, which maps each sentence involving direction terms to a materially
equivalent sentence involving no such terms. Since all the sentences in the range of
r have reference, so do the sentences in its domain. This means that our abstraction
term 'd(l)' satisfies the sufficient condition for havipg reference.
There is something deeply puzzling about this way to defend the referentiality of the
term. The sufficient condition for the term 'd(l)' to refer does not mention any referent
of the term or appear to involve a referent in any other way. By design, the condition is
concerned only with the referentiality of complete sentences not involving the term.
How, then, can the condition suffice for the existence of a referent? Imagine that
someone proposes a sufficient condition for being married that does not mention
or in any other way involve a spouse. 38 We would of course be inclined to reject the
condition. Why should the proposed sufficient condition for having reference be any
better? This is, I think, the hardest and most fundamental of all the problems that
confront the context principle. As far as I can see, no solution is worked out in any of
Frege's writings.
37 Some other problems turn on more specific features of Hale's proposal and might thus, for all we
know, be circumvented by an improved proposal; see (Potter and Smiley, 2001 ) and (Potter and Smiley,
2002). It should also be noted that Hale and Wright have recently moved in a direction that attaches less
significance to the idea of recarving of content; see for instance (Hale and Wright, 2000) and (Hale, 2007) .
38 Cf. Sections 2.4 and 2.6.5.
I therefore ask the reader's indulgence to end with a somewhat speculative sugges-
tion about the kind of solution that Frege might have intended. The context principle
is concerned with a metasemantic question, namely how reference is constituted. It is
not a primitive metaphysical fact that an expression refers to some object. There are
some more basic facts in virtue of which the relation of reference obtains. Perhaps the
more basic facts in virtue of which an abstraction term refers to an abstract object need
not mention or otherwise involve this abstract object. Some other examples suggest that
this is at least a possibility. Consider the relation that obtains between me and a bank
account that I own. It is not a primitive metaphysical fact that this relation obtains.
Instead, the fact obtains in virtue of some complex network of psychological and social
facts. There is no reason why these more basic facts in virtue of which I own the
account should themselves mention or in any obvious way involve the account. What
is constituted by these more basic facts is not only a particular relation of ownership
but also one of the relata.
Of course, there are important differences between this case and that of reference
to mathematical objects. In particular, where a bank account depends on human
thought and actions for its existence, mathematical objects do not. 39 So if this
strategy is to work, mathematical objects would have to be constituted, as it were,
along with their modal profile as necessary existents. If such an account could
be developed, it would provide just what the Grundgesetze needs. There would be
genuine reference to abstract objects, just as there is genuine ownership of bank
accounts. Simultaneously, the sufficient condition for a term to refer would not itself
mention any referent. This would of course give the account a reductionist flavor.
But unlike the holophrastic reductionism criticized in Section 7.6.1, the alternative
form of reductionism currently envisaged would not deny that there is reference to
abstract objects. This alternative would locate the reductionism, not at the syntactic
or semantic level, but at the metasemantic level, as the claim that reference to abstract
objects is constituted by more basic facts that do not themselves involve the relevant
abstract objects. Admittedly, this proposal is highly programmatic. The purpose of
this book is to develop-without any regard for exegetical issues-a view of the sort
that the proposal requires.
Any version of metasemantic reductionism will no doubt strike some readers as a
less robust form of realism than what they associate with Frege. So is my proposed
interpretation compatible with what Frege writes? I believe it is. The discussion
following Frege's criticism of creative definitions is once again relevant. Frege returns
in §147 to the question of whether his procedure for "converting" an instance of the
left-hand side of Basic Law V to the corresponding instance of the right-hand side
should be seen as a form of creation:
39
This independence claim is defended in Section 11.1.
APPENDICES 129
Can our procedure be called a creation? The discussion of this question can easily degenerate
into a quarrel about words. In any case, our creation, if one wishes so to call it, is not
unconstrained and arbitrary, but rather the way of proceeding, and its permissibility, is settled
once and for all. And with this, all the difficulties and concerns that otherwise put into question
the logical possibility of creation vanish; and by means of our value-ranges we may hope to
achieve everything that these other approaches fall short of. (Frege, 2013, II, § 147)
This passage is remarkably concessive. One would have expected Frege to come out
swinging against anything that smacks of creative definition. Instead, he concedes that
his own procedure might, if one so wishes, be called a creation and rather holds up as
the key advantage of his account over those of his creationist opponents that it alone is
systematic and rigorous. This suggests that there may not, after all, be a clash between
Frege's criticism of creative definitions and the interpretation of the context principle
that I have proposed. What matters is that the procedure not be "unconstrained and
arbitrary". And these requirements are satisfied on the view developed throughout
this book, or so I argue.
7.7 Conclusion
While it is widely agreed that the context principle plays a central role in the logicist
project of the Grundlagen, there are arguments that suggest there is no room for the
principle in the Grundgesetze. I have rebutted these arguments and shown that the
principle survives into the Grundgesetze, where it retains a central role in what is
essentially the same explanatory project. Of course, there are changes to both the
context principle and the argument in which it figures. But these changes are brought
about by other developments in Frege's thinking, especially about semantics, and do
not bespeak any dissatisfaction with the context principle per se.
To recognize the longevity of the context principle in Frege's philosophy is not
only of historical interest but also a valuable resource when trying to make sense
of this intriguing but elusive principle. We have learnt that the principle must be
distinguished from the ultra-thin conception of reference and the problematic idea of
content recarving. Instead, I have suggested it might be more fruitful to interpret the
context principle as a metasemantic principle. Thus understood, the principle might
turn out to be both philosophically useful and defensible.
Appendices
7.A Hale and Fine on Reference by Recarving
Let us take a closer look at Hale's attempt to articulate a notion of meaning that is intermediate
between the later notions of sense and reference and thus suited for explicating Frege's idea of
recarving content. As mentioned, Hale's proposal can be regarded as a refinement ofFrege's idea
from the Begriffsschrift, where "the conceptual content" of a sentence was said to be determined
by its inferential relations to other sentences.
Hale proposes that the relevant notion of entailment is a broadly logical one (which includes
analytic entailment) but which is restricted so as to require that the entailment be compact,
roughly in the sense that the premises contain no redundant content that is inessential to the
entailment. The official definition of compactness has undergone some refinement over the
years. The latest version is found at (Hale and Wright, 200la, p. 112) and can be formulated
as follows.
Definition 7.1 (Compact entailment). Assume as given the notion of entailment. Then
A I> . .• , An compactly entail B iff all of the following conditions are met:
(i) A I> .. . , An entail B

(ii) For any non-logical constituent E occurring in At> . .. , An, there is some substitu-
tion E' / E such that the result A;, .. . , A~ of applying this substitution uniformly to
A I• ... , An does not entail B
(iii) For every subformula S of A, there is some formula S' which is modally equivalent to
S-in the sense that both are necessary, both are contingent, or both are impossible40 -
and which is such that the result A;, . . . , A~ of uniform application of the substitution
S' / S to A 1> •• • , An does not entail B
With this definition in place, Hale's guiding idea is the following:
Two sentences have the same [meaning] iff anyone who understands both of them can tell,
without determining their truth-values individually, and by reasoning involving only compact
entailments, that they have the same truth-value. (ibid.)
This guiding idea is then used to motivate a precise definition of sameness of meaning, which
we can formulate as follows.41
Definition 7.2 (Sameness of content). Let'*' stand for mutual compact entailment. Then two
sentences SI and S2 have the same content iff at least one of the following conditions is met:
(i) S1 = S2
(ii) S1 <* S2
(iii) There are sentences A and B such that A <* B and expressions E and E' such that
S1 and S2 result from A and B respectively by uniform replacement of Eby E'
(iv) S1 and S2 stand in the transitive closure of the relation defined by the disjunction of
(i), (ii), and (iii)
Does Hale succeed? Let us grant that Hale has succeeded in articulating a genuine notion
of meaning. (We thus waive the concern that the notion is ad hoc and lacking in theoretical
motivation.) The notion of meaning clearly satisfies the recarving requirement. For the two
sides of an acceptable abstraction principle compactly entail each other. But the requirement of
reference determination remains a concern. Does the meaning of an instance of the right-hand
40
Hale fails to specify how this is to be understood for open formulas. Presumably the notion of modal
equivalence will have to be relativized to an assignment.
41
See (Hale and Wright, 200la, pp. 113-15).
APPENDICES 131
side of an acceptable abstraction principle, along with the logical form of a simple identity
statement, suffice to determine the meanings of the singular terms flanking the identity? It is
hard to see how this reference determination could be accomplished.
Kit Fine has argued that it cannot. 42 Let # be the relation of sameness of meaning. For a
concept F, let p+ be the concept that is true of all and only the Fs and the smallest natural
number that is not an F. Consider two concepts F and G. According to the recarving thesis we
have the following two relations: 43
(7.2) #F= #G#F"'<! G,

(7.3) #F+ = #G+ # p+ ~ c+.
But according to Hale's definition of sameness of meaning, we also have: 44
(7.4)
By transitivity of the relation # it thus follows that:
(7.5) #F = #G # #p+ = #G+.

This means that the meaning with which we began-that of 'F '°"' G' -can be recarved in two
completely different ways: either as #F = #G or as #p+ = #G+. But these latter two sentences
involve reference to distinct pairs of numbers. Hale's notion of meaning is therefore too coarse-
grained to enable the desired determination of reference.
In fact, this objection has force not just against Hale's notion of meaning but much more
broadly. For it is doubtful that any notion of meaning that is coarse enough to validate (7.2)
and (7.3) could fail to validate (7.4) as well. So the root problem remains the same as before.
No notion of meaning that is coarse-grained enough to satisfy the recarving requirement can
also be fine-grained enough to enable the desired determination of reference. I conclude that
the project of explaining reference by an appeal to recarving of meaning should be abandoned.
42
See (Fine, 2002, pp. 39-41).
43
For increased readability I drop quotation marks around the sentences flanking the '<o>'.
44
Proof sketch. The two sides clearly mutually entail each other (in the relevant sense of entailment).
To see that both entailments are compact we first note that every expression on either side can be substituted
for in a way that destroys the entailment of the other side. (Here we hold that p+ and c-+- function as
primitive expressions.) We also note that each side can be replaced by a sentence of the same modal status
so as to destroy the entailment of the other side.
PART III
Details
8
Reference by Abstraction
8.1 Introduction
According to Frege, criteria of identity have an important role to play in the
explanation of reference:
If for us the symbol a is to denote an object, then we must have a criterion which determines
in every case whether b is the same as a, even if it is not always within our power to apply this
criterion. (Frege, 1953, §62)
The idea of a connection between criteria of identity and reference has generated a
lot of enthusiasm. It is not hard to see why. As discussed in Chapter 2, the idea enjoys
considerable intuitive plausibility, and it promises to make available an account of
reference that is general enough to accommodate reference to abstract objects.
Despite this enthusiasm, there has in my opinion not been any satisfactory develop-
ment of the idea.1 The aim of this chapter is to improve on the situation by developing
in some detail a new version of the Fregeau idea. As throughout most of this book, my
aim is systematic, not exegetical. 2 I do not claim that the resulting account is true to
Frege's own intentions. Moreover, I make various simplifying assumptions-for which
no apology should be needed. What is most urgently needed at this stage is a properly
worked-out model of the phenomenon that Frege calls to our attention, not an account
that fully matches our immensely rich and varied practices. If explanatorily successful,
a scientific model can later be tweaked and extended to obtain a better fit with the data.
I focus on a simple example, namely letter types, although further generalizations
are immediate, as explained in earlier chapters.3 Accordingly, my plan is to consider
a community of language users who refer to and quantify over various kinds of
concrete objects, including letter tokens. I then describe a situation in which the
community extends their language by adding some new vocabulary and begins to
1
I am, however, greatly indebted to the important discussions of this Fregean idea, and related ones, in
(Dummett, 198la), (Wright, 1983), (Hale, 1987), (Dummett, 19.9 la), and (Hale and Wright, 200la).
2
The previous chapter is the only major exception to this predominantly systematic orientation.
3 See especially Section 2.5 and Chapter 6. Our simple example has some appealing features. The
tokens are straightforwardly physical objects, unlike, say, the lines in Frege's famous example of directions.
Moreover, since every letter type is instantiated by many tokens (unlike, say, expression types, not all of
which are in fact instantiated), this example allows us to postpone (until Chapter 11) the discussion of
uninstantiated types.
136 REFERENCE BY ABSTRACTION
use their extended language precisely as if they are referring to and quantifying
over letter types. For example, they begin to regard it as correct to utter "this letter
(type) = that letter (type)", pointing at two tokens, just in case the tokens count as
equivalent by some agreed standard. (Here I make the simplifying assumption that
there is an appropriate standard that all the speakers are capable of tracking.) Our
question is whether, after extending their language, the members of the community
are in fact referring to letter types (and thus to abstract objects). There are consider-
ations pulling in each direction. On the one hand, everything is precisely as if they are
so referring. On the other hand, all that has happened is that the speakers have come
to regard it as correct to utter various sentences of the extended language just in case
certain conditions on tokens (and thus concrete objects) are satisfied.
I argue that there are reasons to prefer an interpretation of the extended language
on which the speakers in fact refer to letter types, not only appear to do so. These
abstract referents are obtained by abstraction on tokens. It remains undeniable,
however, that the principles governing the correct use of the extended language have
a reductionistic character. I use this fact to articulate a sense in which the abstract
referents are "thin" objects.
The resulting account of reference differs substantially from extant developments
of the Fregean idea expressed in the quote with which the chapter opened. The
cognoscenti may appreciate an overview. My account avoids some controversial
claims associated with the influential neo-Fregean approach of Hale and Wright.
I make no claim about the priority of syntactic categories to ontological ones. 4
(I would, however, deny that ontological categories can be made sense of indepen-
dently of semantic notions. 5 ) Nor do I invoke a deflationist conception of reference
which equates the referentiality of a singular term t with the truth of the object
language sentence 3x(x = t). 6 And I avoid claims about the ordinary content of
statements being "recarved" or about speakers' ability to translate sentences of the
extended language into the base language.7 Perhaps the most distinctive aspect of
my account is its emphasis on the idea that the abstract objects in question are
"metaphysically lightweight" or thin. This emphasis brings with it another novelty,
namely a restriction to abstraction principles that are predicative, in the sense that
every question about the abstract objects that result can be reduced to (or grounded
in) one about the entities already recognized. 8
4
Contrast (Wright, 1983), (Hale, 1987), and (Hale and Wright, 2001a, chs. 1- 2). The syntactic priority
thesis was criticized in Chapter 5.
5 In this respect I am closer to (Rumfitt, 2003) than to Dummett, Hale, and Wright.
6 See (Dummett, 1956) and (Wright, 1983, p. 83). In the sense of (Hodes, 1990), I aim to show that
certain abstract singular terms have "thick" ontological commitment, not just "thin''.
7 See for instance (Wright, 1983), (Hale, 1987), (Rosen, 1993), (Hale, 1997), and (Rayo, 2016). See also
Chapter 4 for criticism of some recarving-based approaches.

8 See Section 6.2 for a proper definition of this notion of predicativity. While unusual. the last two
features of my account resonate with aspects of Dummett's view. See for instance (Dummett, 1981 a, ch. 14)
and (Dummett, 1991a).
THE LINGUISTIC DATA 137
8.2 The Linguistic Data

Our first task is to make precise the idea that the community uses the extended
language precisely as if they are referring to letter types. This idea has to be cashed
out in the form of a characterization of the linguistic data for which we later attempt
to determine the best semantic interpretation. My general policy is to allow the data
to be characterized in fairly high-level terms-so long as we do not beg any questions
about what is the best semantic interpretation of the extended language. For instance,
I regard the syntax of each language as already settled. 9
The community starts with a base language, Lo, which they use to refer to and
quantify over various concrete objects, including letter tokens. Simplifying substan-
tially, I assume that this language is an ordinary first-order language with identity
and with variables Xi for each natural number i. I further assume that all parties to
the philosophical debate agree on a semantic interpretation I of the base language,
whose domain, Do, consists of objects that are not currently in dispute. While this is a
substantive assumption, it is neutral with regard to our principal question of how the
extended language should be interpreted.
It will be convenient to let the extended language, £1> be a two-sorted first-order
language. 10 The first of its sorts is exactly like the single sort of the base language: the
same variables, individual constants, and predicates. So I call this the base sort. The
second sort, which I call the extended sort, can be thought of as a sort reserved for
talking about letter types. This sort has a distinct collection of individual constants
and variables Yi for each natural number i. There is also an operator § that applies
to any singular term of the base sort to yield a term of the extended sort. The
extended language has one identity predicate for each of its sorts, = and =1, which
may be flanked by two singular terms of the base sort and the extended sort, re-
spectively; "mixed" identity statements involving one term from each sort are con-
sidered ill-formed. 11
Why let the extended language be two-sorted? One reason is to postpone the
so-called Caesar problem, which concerns mixed identity statements, such as "this
token = that type''. Although it requires an answer, this problem is independent of the
main concerns of this chapter. So let us adopt the strategy of the person who lent his
9 Contrast Hale and Wright, who attempt to determine the syntactic category of an expression on the
basis of various grammatical and inferential considerations; cf. note 4. Here I rely on the fact that my notion
of syntactic category is Jess philosophically laden than theirs, as it is no longer claimed to be prior to any
ontological notions.
10 See Section 6.3.l for discussion and (Enderton, 2001, Section 4.3) for a useful introduction to many-
sorted logic.
11 Notice that our languages contain no modal vocabulary. Can the approach be extended to cases
where this shortcoming is remedied? Inspection of some cases suggests an affirmative answer, but I have no
worked-out account to offer. While this obviously calls for more work, remember that what is most badly
needed is a model of how there might be such a thing as reference to abstract objects. Once we have a model
which shows that such reference is intelligible, this model can later be extended. (Thanks to Jon Litland for
discussion.)
name to the problem: divide et impera. We accomplish this by adopting a two-sorted

language where the Caesar problem cannot even be formulated, and where our main
concerns thus come more sharply into focus. 12 Another reason to let the extended
language be two-sorted is to ensure that the abstraction we study is predicative in the
aforementioned sense. The significance of this requirement will become clear as we
go along.
Our task is now to describe what it is for the community to use the extended
language precisely as if they are referring to and quantifying over letter types, obtained
by abstraction on tokens. To carry out this task, we need to assume that the members
of the community have internalized a standard for when two objects count as tokens
of the same type. These standards define a unity relation,...., on the base domain Do.
What are the properties of this relation? If a and b count as tokens of the same type,
then obviously so do b and a. So the relation has to be symmetric. An analogous
argument can be given for its transitivity. However, if a isn't a letter token at all, then
a fortiori a and a are not letter tokens of the same type. So we shouldn't demand that
the relation be reflexive. This means that our unity relation ,...., is a partial equivalence
relation, that is, a relation just like an equivalence except that it is not assumed
to be reflexive. 13
I said that the relation ,...., specifies "when two things are to count as tokens of the
same type''. It is important not to misunderstand this statement. The idea is not that, for
a ,...., b to hold, there must in fact be a type of which both a and b are tokens. This would
be far too demanding. The words "count as" are essential: the relation ,...., specifies the
conditions under which members of the community speak as if two objects are tokens
of the same type. And dearly speakers are capable of tracking this relation without
any antecedent understanding of letter types; indeed, existing physical devices such
as scanners track the relation to a high degree of approximation. 14
With the unity relation ,...., explained, we are now ready to describe the principles
that govern the use of the extended language. The principles specify the condition
under which the speakers regard the assertoric utterance of a formula as correct.
When this condition is met, we say that the formula is assertible in the relevant context.
Thus, to say that a formula is assertible in a context is not to say that it is true, only
that the community regards it as correct. This means that we are not yet engaged in
semantic theorizing, merely in describing what a population regards as correct use of
a language, which will later serve as data for the choice of a semantic interpretation.
Next, I do not require that the assertibility conditions be operationalized in any way,
say by the existence of an effective procedure for determining whether the condition
12 A solution to the Caesar problem is developed in Chapter 9.

13 Abstraction on partial (as opposed to ordinary) equivalence relations receives philosophical discus-
sion in Section 2.5 and technical discussion in Appendix 2.C.
14
Of course, a certain amount of idealization is involved in assuming that people track a partial
equivalence relation. As is well known, vagueness poses a threat to the transitivity of the relation. See
(Shapiro, 2006b, ch. 6) for relevant discussion.
THE LINGUISTIC DATA 139
obtains. 15 Finally, I do not claim that the speakers' mastery of the extended language
is based on translating its formulas into the base language, or that these formulas are
strictly synonymous with the result of such a translation.
I now provide an informal characterization of the assertibility conditions; a precise
characterization can be found in Appendix 8.A. Recall that our extended language
contains the base language as a proper part, and that the base language already has
a semantic interpretation I. So the assertibility conditions had better respect this
interpretation. This means that any formula <p of the base language must be regarded
as assertible of a string of objects just in case <p is true of these objects under the
interpretation I.
The interesting assertibility conditions are the ones governing formulas that are
only available in the extended language. We begin with two simple examples. For any
objects a and b from the base domain Do we have:
• '§x1 =1 §x2' is assertible of a and b iff a"' b
(That is, members of the community regard it as correct to assert 'this letter
(type) = 1 that letter (type); pointing at a and b respectively, just in case a ,. . ., b.)
• 'Vow*(§x 1 )' is assertible of a iff a is a vowel-token
(That is, members of the community regard it as correct to assert 'this letter (type)
is a vowel; pointing at a, just in case a is a vowel-token.)
Notice how the free variables are handled. The assertibility conditions are stated
relative to a string of objects, which function as values of the free variables in question.
Let us now generalize. Each n-place predicate P ~fthe extended sort (or for short,
"extended predicate") is associated with an assertibility condition ctp such that:
P(§x" . .. , §x,.) is assertible of a" . .. , a,. iff ctp(a" . .. , a,.)
We adopt an analogous clause for free variables of the extended sort: P(y1> ... , y,.)
is assertible of a" . .. , a,. just in case ctp(a" . .. , a,.). For these clauses to work as
intended, it is essential that each extended predicate P has been associated with
an assertibility condition ctp which functions as if it is concerned with letter types,
not tokens. For ctp to function in this way, two requirements are needed. First, we
need to require that the assertibility condition ctp treats objects that are equivalent
under ,. . ., as tokens of one and the same type. For example, if the speakers regard it
as correct to assert 'this letter is a vowel; pointing at a token a, they must also regard
it as correct to assert the sentence when pointing at a token b that is equivalent to
a under "". More generally, we require that "' be a congruence with respect to each
assertibility condition ap; that is, that ctp does not distinguish between objects that
are equivalent under"':
(8.1) a1 "'b1 /\ ... /\a,."' b,. ~ (ap(a1> ... ,a,.)~ ctp(b1,. .. , b,.))
15 In this way, our use of assertibility conditions differs from what is often found in discussions of
philosophical anti-realism.
Second, we need to require that, when the speakers take a predicate of the extended
sort to be assertible relative to some objects a1, .. . , an, then these objects are the sort
of thing that is suitable for specifying a letter type; in other words, that each a; is a
letter token, which we can formally express as a;,..., a;. We therefore require that each
assertibility condition exp satisfy:
(8.2)
To complete the characterization of the assertibility conditions, it remains only to

specify how the connectives and quantifiers are handled. Here there are no surprises:
we adopt the obvious compositional clauses.
It should be dear that this assignment of assertibility conditions generalizes readily
from the case of letter types to any other form of predicative abstraction. However,
the method does not generalize to impredicative abstractions. To see this, consider
Hume's Principle, which is impredicative because the numbers that are used to count
objects are supposed to belong to the domain of objects that we can count. 16 Consider
the numerical identity #F = #G. Suppose the concepts assigned to the variables 'F'
and 'G' are defined by the open formulas <p(x) and 1/1 (x), respectively. Relative to
this assignment, we want to count the numerical identity as assertible just in case
it is assertible that the <ps and the 1/Js are equinumerous. However, because of the
mentioned impredicativity, the formulas <p(x) and 1/J(x) may contain occurrences of
the abstraction operator #. This is problematic. To determine whether it is assertible
that the <ps and the 1/Js are equinumerous, the speakers first have to determine the
assertibility of various arithmetical statements, which might well include the very
numerical identity with which we began. So there is no guarantee that the proposed
assertibility condition will be any simpler than the arithmetical statement to which it
is assigned.
We set out to describe what it is for the community to use the extended language
precisely as if they are referring to letter types. I claim that the assertibility condi-
tions formulated above provide the desired description. When a community uses the
extended language in accordance with these assertibility conditions, they speak precisely
as if they are referring to and quantifying over letter types. This claim is supported by the
considerations that we used to motivate the assertibility conditions. Additionally, the
claim will be made formally precise and given a rigorous defense in the next section,
where we give precise content to the phrase "refer to and quantify over letter types''.
8.3 Two Competing Interpretations

We are now in a position to give a more precise statement of a question posed in the
introduction. Assume a community extends their language in the way just described.
16
See Section 6.4 for a fuller discussion, which the remainder of this paragraph summarizes.
TWO COMPETING INTERPRETATIONS 141
What is the best semantic interpretation of their extended language? As adumbrated,

different considerations pull in different directions. On the one hand, the assertibility
conditions that govern its use make no mention of abstract objects such as letter types.
So to ascribe reference to such objects to some of the terms of the extended language,
it seems, is to go beyond what is warranted by the linguistic data. An interpretation
that stays closer to the data seems preferable. On the other hand, the language is
used precisely as if speakers are referring to and quantifying over letter types. So it is
tempting to take appearances at face value and interpret the terms of the extended sort
(or for short, "the extended terms") as actually referring to and ranging over abstract
letter types. The task of this section is to describe two competing interpretations, one
corresponding to each of these two considerations.
The first interpretation is a form of semantic reductionism, which seeks to stay
as close as possible to the data. Indeed, the interpretation is read off directly from
the assertibility conditions. Instead of making the non-semantic claim that speakers
associate a term t of the extended sort with a token a, we make the semantic claim that
t refers to a. And instead of saying that a predicate of the extended sort is assertible
of some tokens, we say that it is true of these tokens. Finally, the connectives and
quantifiers are interpreted in the obvious compositional way. 17 These stipulations
ensure that on the reductionist interpretation, a formula is true relative to a string of
objects just in case the formula is assertible relative to that string. In effect, the "letter
talk" of the extended language is interpreted as merely a peculiar form of "token talk"
where the extended predicate '= i ' is interpreted as ,. . ., .
The second interpretation is a form of semantic non-reductionism, which takes
at face value the fact that the speakers behave precisely as if they are talking about
letter types. An extended term associated with a token a is therefore interpreted
as referring to the type of a. So this interpretation freely avails itself of abstract
letter types. We handle this by means of the following conditional abstraction
principle:
a,....., a /\ b,....., b ~ (§a= §b #a "' b)

Since c "' c just in case c is a letter token, (AP c) ensures that the abstraction operation
maps two tokens a and b to the same type just in case a "' b. If an extended term
is associated with a token a, we let its referent be §a, rather than a itself (as on the
reductionist interpretation). Given its unabashed reliance on abstract letter types,
is the non-reductionist interpretation actually available? An affirmative answer is
defended in Section 8.5.
17
However, because we use free logic to handle abstraction on a partial equivalence relation, the clause
for the universal quantifier needs to be:
R Fu Vu() iff for every u-variant p of er , we have R F p u ~ u --+ ()
We define an extended domain D1 as the range of the abstraction operation, that

is, as the set of letter types with tokens in the base domain Do:
(Bl) u E D1 # 3a(a E Do /\a "" a /\ u = §a)
Just as base terms range over and refer to objects in the base domain Do, so the non-
reductionist holds that extended terms range over and refer to objects in the extended
domain Di.
Next, we consider the interpretation of predicates. Let P be an n-place extended
predicate, which is assertible ofal> .. . , an just in case ap(aL> . .. , an)· As observed, the
semantic reductionist appropriates this assertibility condition for her own semantic
purposes by laying down that:
Pis true of a1, . .. , an on R iff ap(a1, ... , an)
where R is the reductionist interpretation. In keeping with her aims, the reductionist
thus takes the predicate to be defined only on the concrete objects from the base
domain Do. By contrast, the non-reductionist wishes to interpret the predicate as
defined on his extended domain D1 . He achieves this by defining a new condition
aj, on this domain such that:
(B2) Val> ... , an E Do (aj,(§al> ... , §an) # ap(ai, .. . , an))
(Notice that this is well defined because"" is a congruence with respect to ap.) The
non-reductionist uses this new condition aj, to interpret the extended predicate P:
Pis true of b1, ... , bn onN iff a;(bi, . .. , bn)

where N is the non-reductionist interpretation. This treatment of extended predicates
achieves what the non-reductionist wants, namely an interpretation that is defined
on letter types but is compatible with the linguistic data underlying the assertibility
conditions.
Finally, the non-reductionist semantics treats the connectives and quantifiers in the
obvious compositional way. But it is worth recalling that the non-reductionist lets the
variables of the extended sort range over the extended domain D 1 rather than over
the base domain Do, as is done on the reductionist semantics.
The two described interpretations differ in important ways. It is not hard to see
that they differ in the truth-conditions that they ascribe to many sentences, and
therefore also on these sentences' meanings. Consider the sentence 3y 1 Vow*(y 1).
On the reductionist interpretation, this sentence is true just in case there exists a
concrete vowel token, while on the non-reductionist interpretation, it is true just in
case there exists an abstract vowel letter. Moreover, many sentences incur different
ontological commitments on the two interpretations, as is illustrated by the example
just discussed. On the former interpretation, this sentence is not committed to the
existence of anything abstract, while on the latter, it is. In another respect, however,
the two interpretations are closely related. Recall the bridge principles (B 1) and (B2 ),
WHY THE NON-REDUCTIONIST INTERPRETATION IS PREFERABLE 143
which relate some key ingredients of the two interpretations. Modulo these two
principles, the truth-conditions generated by the two interpretations turn out to be
equivalent. 18
This technical result is philosophically important. For one thing, the result enables
me to give precise content to, and proof of, a claim made in the previous section,
namely that to use the extended language in accordance with the assertibility condi-
tions is to speak precisely as if one is referring to and quantifying over letter types. 19
For another, the result sheds light on the nature of the disagreement between the two
interpretations. The interpretations are in complete agreement about which sentences
are true on which occasions of use. Their only disagreement concerns the demarcation
of semantics. The reductionist complains that the non-reductionist semantics involves
an unnecessary and problematic detour via abstract semantic values. The bridge
principles are therefore needed to convert the non-reductionist's truth-conditions to
the reductionist's truth-conditions, which are closer to the linguistic data. By contrast,
the non-reductionist complains that the reductionist mistakenly incorporates various
extra-semantic principles into her semantics. Consider the reductionist's claim that
the truth of an identity flanked by two extended terms is a matter of the two associated
tokens' being equivalent. From the non-reductionist's point of view, this claim is
a confused amalgamation of the semantic principle that the truth of an identity
statement is a matter of the two referents' being identical, and the metaphysical
principle that the identity of letter types is grounded in the equivalence of the
associated tokens. 20
8.4 Why the Non-reductionist Interpretation

is Preferable
I just described two semantic interpretations of the extended language £ 1• Which is
the better interpretation? I now formulate some constraints on the choice of a semantic
interpretation, and based on these, I defend the non-reductionist interpretation.
Throughout the section, I proceed on the assumption that both of the described
interpretations are available, including the non-reductionist one, which is committed
to abstract objects. This assumption is defended in the next section.
8.4.1 The principle of charity
According to the principle of charity, we should, in the absence of countervailing
considerations, interpret a language so as to respect its speakers' dispositions concern-
ing the truth of sentences. The semantic interpretation should be such that sentences
which the speakers firmly regard as true are indeed true.
18
The result is given a precise statement as Proposition 8.2 in Appendix 8.B, where a proof is also
provided.
19 20
Again, see Appendix 8.B for details. I return to this theme in Section 8.6.
Of course, the presumption in favor of a charitable interpretation can be overridden

where there is reason to believe that the speakers are guilty of conceptual or fac-
tual errors. But our speakers are innocent of any such mistakes, or so I argue in
Section 8.5. Regardless, the principle of charity is powerless to differentiate between
the reductionist and non-reductionist interpretations, because both interpretations
satisfy the principle perfectly. On both interpretations, a sentence is true just in case
it is assertible according to the community's conventions.
8.4.2 Ihe principle of compositionality

We generally require that a semantic interpretation be compositional. Let me review
what this means. We assume that each component of a complex expression makes
some definite contribution to the meaning of the complex expression. The contribu-
tion made by an expression e is known as its semantic value and is written [ e]. For
instance, Frege held that the semantic value of a sentence is its truth-value and that
the semantic values of other expressions are their contributions to the truth-values of
sentences in which they occur. In particular, the semantic value of a singular term is
just its referent. The notion of a semantic value thus generalizes the ordinary notion of
reference. Where the notion of reference is intended primarily for singular terms, the
notion of semantic value is meant to be applicable to expressions of all grammatical
categories.
The principle of compositionality states that the semantic value of a complex expres-
sion is functionally determined by the semantic values of its components and their
syntactic mode of combination (for instance, concatenation). 21 For example, accord-
ing to Frege the semantic value of a simple sentence such as 'John runs' is determined
by the equation:
[John runs]= [runs]([John])
That is, the semantic value of the sentence 'John runs' is the result of applying the
function that is the semantic value of the predicate 'runs' to the argument that is the
semantic value of the subject 'John'. More generally, the principle of compositionality
can be formulated as follows:
For any syntactic operation C there is a corresponding semantic operation C* such that the
semantic value of the result of applying C to expressions el> ... , en is identical to the result of
applying C* to these expressions' semantic values:
How do our two competing interpretations fare with respect to compositionality?

It is easy to verify that both interpretations satisfy the constraint perfectly-as far
21 Among the more plausible candidates for exceptions are idioms and propositional attitude contexts.
Since the language with which we are concerned is non-idiomatic and extensional, these possible exceptions
need not worry us.
as the present language is concerned. However, it turns out that the reductionist
interpretation is not at all robust in this regard. Some minor modifications of the
language cannot be handled compositionally by the reductionist interpretation. One
problem arises in connection with generalized quantifiers. 22 The usual semantic
treatment of'most' has '(Most x : rp(x))1/I (x)' come out true just in case more than half
of the objects in the domain that satisfy rp(x) also satisfy 1/t(x). Consider the following
arrangement of tokens:
~
L£_i_J
Then the following sentence is false:
(4) Most letter types are vowels.
For of the three letter types present, only one is a vowel. But by operating with a
domain consisting of tokens rather than types, the semantic reductionist cannot easily
account for the falsity of (4). According to the reductionist, the domain consists of
five tokens, more than half of which are vowels. To retain the ordinary treatment
of generalized quantifiers such as 'most', we seem to be forced to operate at the level
ofletter types rather than tokens.
The reductionist might respond that the extended language has two versions of the
generalized quantifier 'most': one for each of its two sorts. Where the base version
behaves in the ordinary way, the extended version behaves unusually in counting only
"up to equivalence''. This response is unsatisfactory. For one thing, it is unpleasantly ad
hoc. Why should 'most' behave in one way when talking about concrete objects and in
an entirely different way when talking about abstract ones? For another, the argument
against the reductionist can be recast in the context of a one-sorted language, where
the two sorts of our extended language £ 1 have been merged into a single sort.
In the resulting one-sorted language there is no room for two different quantifiers
'most'. Indeed, as we shall see shortly, there are some simple and natural one-sorted
languages (without any generalized quantifiers) that admit no interpretation that is
both reductionist and compositional. 23
To obtain an acceptable account of generalized quantifiers, we need a "coarse
graining" of the objects on which we abstract. The non-reductionist obtains such a
coarse graining from his abstraction operator §, which maps a token to its abstract
letter type. The reductionist has some room to maneuver, however, as she can try to
imitate such a coarse graining. All that the analysis of generalized quantifiers requires
is that the mapping§ must satisfy the principle (AP,). Provided this requirement is
22 This problem is discussed by (Heck, Jr., 2000b) in a somewhat different context. As Heck observes,
similar considerations apply to many other generalized quantifiers as well.

23
See Section 9.7.
met, the mapping can do anything whatsoever. It can, for example, map any token
to some chosen representative from its equivalence class. So perhaps the reductionist
can interpret the mapping in precisely this way and thus modify the non-reductionist
interpretation to suit her own purposes.
Again, the non-reductionist has the resources to strike back. The most definitive
response is that this move is unavailable in the one-sorted setting that we consider
in Section 9.7. But in fact, the proposal is problematic already in our present two-
sorted setting. First, the choice of representatives would be completely arbitrary. 24
Second, as (Rosen, 1993) observes in a related context, the trick of choosing a distinct
representative from each equivalence class is not always available. Assume that the
community extends their language, not by speaking as if every letter token determines
a type, but as if any plurality of objects of the base sort determines a unique set of the
extended sort. This is a perfectly coherent linguistic practice. The abstraction principle
guiding the practice is a two-sorted version of Basic Law V, which unlike its more
famous one-sorted cousin is known to be consistent. But in this case there are as
many equivalence classes as there are pluralities of objects from the base domain. By
a generalization of Cantor's theorem, it is thus out of the question to choose a distinct
representative from this domain for each equivalence class. Our only option would be
to let the pluralities themselves serve as proxies for the associated sets. Since pluralities
are not objects, however, this would result in type clashes. For example, when 'most'
is used in a statement about ordinary objects, the truth-conditions involve cardinality
comparisons among objects, but when the word is used in a statement about sets, we
need to make cardinality comparisons among pluralities.
8.4.3 Cognitive constraints on an interpretation

I now turn to a cognitive constraint on the choice of a semantic interpretation.
According to this constraint, the truth-condition that the interpretation assigns to a
sentence must be properly understood by speakers who understand the sentence. For
instance, the standard truth-condition for 'Snow is white' satisfies the constraint. To
understand this sentence, a speaker must grasp the concepts of snow and whiteness
and master the operation of predication; and together, these cognitive skills suffice to
understand the standard truth-condition. But in other cases, the constraint enables
us to reject attempted semantic analyses which invoke concepts or principles which
competent speakers do not properly understand.
Consider, for example, a community whose language is so impoverished that
it features no genuine reference to properly individuated objects. All that can be
expressed are simple "feature-placing" claims such as "hot here" and "wet there''.
But later, perhaps after many generations, the community extends their language to
one that functions precisely as if referring to and quantifying over physical bodies
24 And of course, it is not an option for the reductionist to let the mapping send each token to its own
equivalence class under~, since this would invoke reference to abstract objects after all.
such as sticks and stones. We discussed in Section 2.3 what this might involve. The
extended language contains demonstratives such as 'this body' and 'that body: On
each occasion of use, a demonstrative of this sort is associated with a parcel of matter.
For example, when a speaker utters 'this body: attending to a stick protruding from
the ground, then the demonstrative is associated with the parcel of matter with which
the speaker perceptually interacts. And the assertibility conditions that govern the
extended language ensure that an identity of the form 'this body = that body' is
assertible, relative to two associated parcels of matter, just in case the parcels of matter
are related through a continuous stretch of solid stuff, all of which belongs to the same
unit of independent motion. Let ,...., be this unity relation on parcels of matter. 25
With these assumptions in place, we ask what is the appropriate semantic interpre-
tation of the extended language. Again, there are two competing options. It suffices
to consider their respective analyses of a simple identity of the form 'this body =
that body', where the demonstratives are associated with two parcels of matter. The
reductionist interpretation takes the sentence to be referring to the two parcels of mat-
ter and asserting that they stand in the relation "'· The competing non-reductionist
interpretation takes the sentence to be referring to the two physical bodies that are
specified by means of the parcels of matter and asserting that these bodies are iden-
tical. The cognitive constraint may help us decide between the two interpretations. It
is plausible to assume that the speakers have no explicit understanding of the unity
relation "'. Although the speakers have a practical ability to tell whether two parcels
of matter stand in the unity relation, this ability relies on subpersonal mechanisms
and low-level practical skills, not on an explicit understanding of this relation. 26
The cognitive constraint therefore provides a reason to favor the non-reductionist
interpretation.
To favor this interpretation is not to deny or belittle the role that the parcels of
matter and the relation "' play in the linguistic practice. All we deny is that these
items figure directly in the best semantic interpretation of the language. The parcels of
matter and the relation "' play an important metasemantic role in what we may call
"the implementation" of reference to bodies. What makes it the case that certain terms
refer to certain physical bodies? The answer, I claim, is that the terms are associated
with parcels of matter and that their use is governed by a principle involving the
relation ,...., . But it would be a mistake to think that every principle that contributes
to the obtaining of the semantic relation of reference is itself a semantic principle. (For
example, according to Kripke some form of baptism often contributes to the obtaining
of relations of reference but plays no role in the semantic theory itself.)
25 Again, see the analysis in Section 2.3. However, my present point is independent of the details of my
analysis of the unity relation ~ and would hold even if readers substitute their own preferred analysis.
26 Indeed, (Burge, 2010, pp. 438-50) argues (citing relevant empirical evidence) that the mentioned
ability is provided by our perceptual system and thus is non-conceptual.

Let us return to the main example of this chapter, where an extended language is
used as if to talk about letter types. Is the cognitive constraint as useful here as it was
in the example of physical bodies? The answer will turn on whether the speakers are
guided by an explicit understanding of the relation of being tokens of the same type
or merely utilize some practical ability to tell whether two objects stand in this unity
relation. The immense difficulty of providing an explicit characterization of the unity
relation suggests the latter. If this is right, the cognitive constraint will again favor the
non-reductionist interpretation.
By denying that speakers need to have an explicit understanding of criteria of
identity, my account differs markedly from much earlier philosophical work on
criteria of identity. 27 Most earlier writers have emphasized the need for an explicit
understanding of-and conscious access to-the relevant criteria of identity.
8.5 Why the Non-reductionist Interpretation

is Available
I have just argued that, if available, the non-reductionist interpretation of the extended
language will be superior to its reductionist rival. One advantage has to do with com-
positionality, and another with a cognitive constraint on the demarcation of seman-
tic principles from non-semantic ones. I will now argue that the non-reductionist
interpretation is indeed available, despite its commitment to abstract letter types. The
argument is based on the idea that the use of a language is prior to semantic theorizing
about it.
Consider again a community that wishes to extend their language by means of pred-
icative abstraction, as described above. When looking at some tokens, for instance, it
would be convenient to be able to ask one another whether most of the letter types
are vowels, rather than having to resort to a more longwinded formulation involving
counting only up to the equivalence ,...., . So the community goes ahead and adopts the
extended language, governed by the described assertibility conditions. Is their new
linguistic practice in good order? It is hard to see why not. The community is not very
sophisticated and takes no interest in theoretical semantics. They have no view on how
the extended language is to be interpreted. What matters to them is simply to be able
to use the extended language in accordance with the assertibility conditions. Given the
unspecific nature of their demand and the fact that each sentence has an assertibility
condition expressed in unproblematic terms that make no use of directions or other
abstract objects and does not rest on any other assumptions that are contested in the
present debate, it is hard to see how the linguistic practice could be flawed. 28
27
See Section 2.3, especially note 19.
28 As we saw in Sections 6.4 and 8.2, the same cannot be said about cases of impredicative abstraction.
WHY THE NON-REDUCTIONIST INTERPRETATION IS AVAILABLE 149
Of course, things would be very different if the assertibility conditions of the

extended language were subject to non-trivial presuppositions. Consider a commu-
nity of pre-Lavoisier chemists who wish to extend their language so as to talk about
phlogiston. In part, they wish to regard it as assertible of any process of combustion
that phlogiston has been released. Furthermore, the community intends this assertion
to be more than just a fancy way of saying that combustion is taking place. They
intend the term 'phlogiston' to pick out a substance that plays a number of roles that
are characterizable already in their base language. The substance is supposed to play
a crucial role in the physical explanation of combustion and to be something that
can be isolated and observed, at least in principle. In short, combustion is supposed
to be evidence for the release of phlogiston, not part of a definition of how the term
is to be used. The success of their new linguistic practice is thus premised on the
existence of a substance that plays at least most of these roles. By contrast, the speakers
of our extended language .C 1 intend nothing other than to use this new language in
accordance with the assertibility conditions, which, as we have seen, are stated in
unproblematic terms and subject to no additional presuppositions. 29
What happens next? As the centuries go by, the speakers hold on to their expressive
resources and add new ones in analogous ways. Gradually they become more sophisti-
cated, and eventually they even begin to take an interest in theoretical semantics. One
day they ask how best to interpret their own extended language .C 1. Our speakers
particularly wonder what objects, if any, the singular terms of .C 1 refer to. When
pondering these matters, they obviously rely on their own language, which thus
functions as their metalanguage. This metalanguage is an extension of .C 1 which
°
we may call .Cf! (with the 'M' for 'meta'). 3 Crucially, this metalanguage functions
precisely as if it is capable of referring to and quantifying over letter types (as
well as other kinds of objects). This enables the community to formulate both of
the competing interpretations of .C 1 that we discussed above. So they compare the
interpretations and favor the non-reductionist one, for the reasons expounded in the
previous section. In short, they decide to take appearances at face value and assert that
the extended terms genuinely refer to abstract letter types.
A skeptic may challenge the community's right to carry out their semantic reason-
ing in a metalanguage which, just like .Ci> functions as if it is capable of referring
to letter types. Should they not be using a more modest metalanguage which, just
like .Co, refrains from using potentially problematic linguistic resources such as terms
intended to refer to abstract objects? If such a modest metalanguage was used, the
community would be unable to formulate the non-reductionist interpretation. For
this interpretation makes crucial use of letter types, which the modest language is
29 My contrast here is essentially the same as that between what (Hale and Wright, 2009b) call the
Lockean model of reference fixation and non-Lockean alternatives.

30
Alternatively, they may use an extension of the one-sorted language £2, described in Appendix 9.A,
that results from merging the two sorts of C 1 .
too impoverished to talk about. According to the skeptic, it is only because the com-
munity helps itself to certain illegitimate referential resources in the metalanguage
.Cff that they appear able to invoke the non-reductionist interpretation of the object
language Ci .
The skeptic's challenge can be answered. If it was permissible, centuries ago, for
the community to adopt the extended language .Ci. and if later extensions were
permissible in an analogous way, why should it not also be permissible for the
community to rely on the resulting language when doing science, in particular, when
doing semantics? The challenge nevertheless points to an interesting phenomenon.
For a semantic interpretation of an object language to be available in a metalanguage,
the metalanguage needs to have sufficient expressive resources. To interpret an object
language as referring to ordinary physical bodies, for example, we need to be able to
engage in precisely this sort of reference in the metalanguage. We cannot state that the
object language refers to physical bodies without ourselves relying on this very form of
reference. This is a manifestation of the internalism about reference first encountered
in Section 2.6.5. And the observation generalizes. We cannot interpret an object
language as referring to Fs without using a metalanguage that is capable of referring
to Fs. By calling into question the use of the relevant referential resources in the
metalanguage, the skeptical challenge can thus be used to undermine any ascription
of reference to some object language. We should therefore reject the challenge.
Let us take stock. We have described a remarkable development. First, an un-
sophisticated community extended their language in an acceptable way so as to speak
precisely as if there are abstract letter types. Second, while still relying on a language
with this feature, the community later raises the question of the correct semantic
analysis of their language. They conclude that they have succeeded in referring to
abstract objects. Their language is in this sense stable under semantic reflection: once
adopted, its speakers can investigate their own language and make good semantic
sense of it. Finally, I argued that this stability under semantic reflection is all that
we could reasonably hope for. Reference to some category of objects can never be
ascribed to an object language without relying on this very form of reference in the
metalanguage.
This three-step argument weaves together considerations developed throughout
the book. Why is the extended language acceptable, as asserted in step one? Because
logically acceptable assertibility conditions can be stated in the unproblematic base
language. And this, in turn, is made possible because of the predicative character
of the abstraction. 31 Second, the best correct semantic analysis of the extended
language involves reference to abstract objects for the reasons given in Section 8.4.
These reasons utilize the criterion of identity for letter types, as well as the principle
of compositionality and our lack of explicit or conscious access to the mentioned
criterion of identity. Finally, step three rests on the internalism about reference.
31
This predicative character was examined in detail in Section 3.3 and Chapter 6.
THIN OBJECTS 151
This internalism extracts a sound core from the ultra-thin conception of reference
associated with the syntactic priority thesis but which was criticized in Chapter 5
for being too unconstrained. In short, given internalism about reference, the use of
predicative criteria of identity ensures easy reference. And when easy reference is
combined with the Fregean conception of an object as a possible referent of a singular
term, we obtain thin objects.
8.6 Thin Objects

The linguistic development that I have described has important ontological conse-
quences. As a result of the described language extension, the community regards it
as correct to assert a variety of sentences which, if taken at face value, are ontologically
committed to abstract objects. For example, they assert '3y1 Vow*(y1)'. And once the
community engages in the semantic reflection that we described, they ascribe truth,
not just correctness, to the sentences in question, and they have reason to favor a
semantic analysis that does take the relevant sentences at face value. The community
thus agrees that many of their sentences are ontologically committed to abstract
objects. Since all of this takes place in an acceptable manner, the community ends
up correctly concluding that there are abstract objects.
This way of introducing objects into one's rational discourse is bound to strike some
philosophers as too easy. Surely, such philosophers will think, the view that there are
abstract objects is a substantive thesis whose truth requires the cooperation of reality,
not just the adoption of some language. I disagree. I have argued that letter types, and
other abstract objects that can be introduced in an analogous way, are thin objects,
which do not require much of reality for their existence. Their existence requires only
the obtaining of some condition which does not mention the objects in question and
which is thus comparatively unproblematic. The existence of a letter type, for example,
requires only the existence of an appropriate token.
My approach has been to explicate the idea of thin objects in terms of the notion
of sufficiency. An object o is thin when there is a statement rp that is not ontologic-
ally committed to o but that nevertheless suffices for o's existence. Of course, this
explication requires a proper account of the sufficiency operator =?. The account that
I favor is based on the idea that rp suffices for 1fr when rp is a ground for asserting
1/r in a permissible language extension of the sort introduced in Chapter 2 and
examined in detail in the present chapter. 32 For example, a1 "'a 2 and Vow(a 1) suffice
for §a1 = §a2 and Vow*(§a1), respectively. We have the resources to give a precise
definition of the sufficiency operator. Since this task is somewhat technical, it is
relegated to Appendix 8.D, where I also show that my account validates (the universal
closures of) the following sufficiency statements: 33
32 See, in particular, Section 2.7.

33
It is noteworthy that no restriction on "exportation" is needed: each abstraction term can be replaced
by a free variable, e.g. in order to infer from the first of these statements that y 1 = §x1 4 (x 1 ~ x 2 =?
y1 = §xi) . As discussed in Section 1.7, it cannot be taken for granted that this "exportation" is permitted.
x1 "" x2 => §x1 = §x2 -.xi "" x2 => -.§xi = §x2

<P(XJ, ... , Xn) => <P*(§X1> ... , §Xn) -.<P(XJ, ... ,Xn) => ....,<P*(§xi, .. • ,§Xn)
I also show that the sufficiency operator satisfies the logical constraints formulated
in Section 1.5.
Next, we need to show that the sufficiency operator satisfies the philosophical
constraints formulated in Section 1.6. A preliminary argument to this effect was
developed in Section 2.7. The very important epistemic constraint-roughly, that
 1/t must ensure that the associated material conditional is within our epistemic
reach-will be defended at greater length in Section 11.5.
I wish to end by comparing my approach to thin objects with some others. There are
two dimensions along which to compare. One concerns the question of symmetry: In
what respects, if any, are the two sides of an abstraction principle on a par? Most extant
approaches to abstraction-from Frege up through the neo-Fregeans and Rayo-
defend a highly symmetric conception. By contrast, I have defended a distinctively
asymmetric conception on which the two sides differ in various worldly respects (that
is, respects that are not just a result of our epistemic and semantic perspective on
the world but concern the world itself). 34 Consider 'this letter = that letter' (pointing
at two tokens a 1 and a 2) and 'a 1 ,...., a2: I claim that the former statement refers to
abstract letter types, while the latter (unsurprisingly) refers only to concrete tokens.
Moreover, the former statement, unlike the latter, is ontologically committed to letter
types. Finally, I claim that, when true, the former statement is grounded in the latter.
The specified letter types are identical because the tokens by means of which they are
specified are equivalent, but not vice versa.
A second dimension of comparison concerns the form of reductionism involved
in the approach. As stressed repeatedly, my approach has a reductionist character:
each sentence of the extended language has an assertibility condition that makes
no mention of letter types or any other abstract objects. Indeed, the reductionistic
character plays an essential role in my argument that the language extension we have
discussed is permissible and rests on no substantive presuppositions. However, the
reductionism involved in my view differs fundamentally from that of the reductionist
approaches to abstraction that have been criticized above. First, there is holophrastic
reductionism, which proceeds by translating each sentence of the extended language
into one of the base language. This approach was criticized in Chapter 5, in part on
the grounds that it renders the extended terms "semantically idle". 35 Then, there is
34
See Chapter 4, especially Section 4.4.
35
Another problem with holophrastic reductionism emerged in Section 6.3.2, namely that the men-
tioned translation is not available for the entire extended language but only a fragment thereof. By contrast,
the assertibility conditions described above (and stated precisely in Appendix 8.A) are available for the
entire extended language.
APPENDICES 153
semantic reductionism, as discussed in Section 8.3. On this approach, the extended

terms are not "semantically idle" but are interpreted as referring to tokens, not types.
The reductionist interpretation R ensures that it can be perfectly true to speak as
if there are letter types. But this apparent reference to letter types is deceptive,
as all reference is really to tokens and other concrete objects. I have rejected this
reductionist interpretation of the extended language in favor of the non-reductionist
interpretation N, which takes extended terms at face value as referring to letter types,
not tokens.
Where, then, is my reductionism located? 36 There is nothing reductionist about
the semantic interpretation N that I favor. The reductionist aspect of my view is
confined entirely to the assertibility conditions that govern the use of the language.
Let me elaborate. The study of the facts in virtue of which linguistic expressions come
to possess their semantic properties is called metasemantics. As already emphasized,
I claim that the correct semantic interpretation of the extended language is the non -
reductionist one. But there is an associated metasemantic question. In virtue of what
is this interpretation the correct one? The answer, I claim, is that the speakers' use of
the language is governed by the assertibility conditions described in Section 8.2. This
answer locates the reductionism entirely at the level of the assertibility conditions.
My account thus combines semantic non-reductionism with a metasemantic form of
reductionism. While I hold that the extended language genuinely refers to abstract
letter types, these truths about reference are grounded in some simpler truths about
the use of the language, which make no mention of letter types or other abstract
objects.
Appendices
8.A The Assertibility Conditions
In this appendix, we work in a metalanguage £{r1, which we assume to be one-sorted. In the
appendices that follow, we work in an extended metalanguage £t1'
that adds a separate sort for
letter types, as well as the usual resources for abstraction on '"".
Let us properly define the assertibility conditions outlined in the main text. Recall that
the base language £ 0 has already been provided with a semantic interpretation 'I, which the
assertibility conditions for the extended language .C 1 must respect.
Predicates. For every n-place predicate P of £1 there is a formula ap in £{r1 with n free
variables such that, for any a i, .. . , an from the base domain D0 , P is assertible of a 1, ... , an
iff ap(at> .. . , an). We also make the following assumptions.
36
Dummett discusses this question, as it arises in connection with an early version of my view (probably
(Linnebo, 2009b)), in (Dummett, 2007, pp. 793-4).
• If Pis in .Co, then exp is the satisfaction condition that 'I assigns to P.
(That is, the assertibility condition of any basic predicate coincides with its truth-condition
on the already given interpretation I.)
• If Pis =1> then exp is~ .
• If the i'th argument place of P is of the extended sort, then we have:
(That is, only objects in the field of ~ are regarded as specifying abstracts.)
• If the i'th argument place of P is of the extended sort, then for any a 1, . .. , an and b; from
Do we have:
a; ~ b; -"* (cxp(a1, . . . , an) ++ cxp(a!> . . . , a,,) [b;/a;J)

where as usual 1/r [e' / e] is the formula that results from 1/r by a uniform substitution of e'
fore.
(That is, argument places of the extended sort do not differentiate between "-'-equivalent
objects.)
Singular terms. We have so far talked about a formula cp being assertible of a finite sequence of
objects. But as usual, it is convenient to state such claims in terms of the notion of a variable
assignment. So let a be an assignment of an element of the base domain Do to each variable of
.C 1. Relative to a, each singular term t of .C 1 is assigned an element a of D0, as follows.
• If t is a constant of the basic sort, then t is assigned its referent on the interpretation 'I.
• If t is a variable, its assignment is provided by a .
• If t is a term of the form §s, then t is assigned the object already assigned to s relative to a.
Formulas. The assertibility conditions for formulas of .C1 are as follows.
• Suppose 1fr is an atomic formula P(t1> ... , tn)· Suppose each singular term t; is assigned
an element a; of Do relative to a. Then 1/r is assertible relative to a iff cxp(a1> . . . , an).
• The clauses for the truth-functional connectives and the quantifiers of the basic sort are
the obvious compositional ones.
• Suppose 1fr is of the form Vve, where v is a variable of the extended sort. Then 1fr is
assertible relative to a iff r v "' v -"* 8' is assertible relative to p for every variable
assignment p that differs from a at most in its assignment to v.
It is a striking feature of the assertibility conditions just defined that they involve only the base
domain Do . Even variables of the extended sort are assigned elements of Do. Let us say that such
variable assignments are based on Do . Pleasingly, however, we can show that the assignments
made to variables of the extended sort matter only up to the equivalence "'.To be more precise,
let us say that two such variable assignments a and a' are equivalent (written a "' a') just in
case a (x;) = a ' (x;) and a (y;) "' a' (y;) for each i E w . Then we have:
Lemma 8.1 Let a and a' be variable assignments based on Do. Let 1fr be any formula of .C 1.
Suppose a"' a ' . Then 1fr is assertible relative to a just in case 1fr is assertible relative to a'.
Proof A straightforward induction on the syntactic complexity of 1fr. -I
APPENDICES 155
8.B Comparing the Two Interpretations

Two semantic interpretations of the extended language £1 were explained in Section 8.3: a
reductionist interpretation R , which is modeled directly on the assertibility conditions for £ 1;
and a non-reductionist interpretation N , which interprets the terms of the extended sort (or,
for short, "the extended terms") as referring to and ranging over letter types, obtained by
abstraction on tokens. I now wish to compare the two interpretations.
This will involve comparing the kinds of variable assignment appropriate for each inter-
pretation. Let us say that a variable assignment p is based on Do and D1 iff p assigns an element
of Do or D1 to each variable of the basic sort or extended sort, respectively. Suppose a is a
variable assignment based on Do. Then we define a variable assignment §a based on Do and
D 1 by letting (§a)(v) = a(v) for each basic variable v and (§a)(y;) = §(a(y;)).
Finally, let us write R l=u 1/r for the claim that 1/r is satisfied by the variable assignment a
on the interpretation n. and likewise for N .
Proposition 8.1 Let 1/r be any formula of£ 1 and a be any variable assignment for£ 1 based on
Do . Then:
1/r is assertible relative to a iff n
l=u 1/r.
Proof An obvious induction on the syntactic complexity of 1fr. -I
Recall the two bridge principles that relate the domains and the interpretations that figure in
the two interpretations:
(Bl) d E D1 ~ 3a(a E Do /\ a "" a /\ d = §a)

(B2) Val> . . . , an E Do (ap(§ai. ... , §an) ~ ap(ai. ... , an))
Proposition 8.2 Let a be a variable assignment for £1 based on Do. Then, for any formula 1/r
of£ 1> we have:
n l=u 1/r iff N h u 1/r
Proof We proceed by induction on the syntactic complexity of 1fr. Suppose first that 1fr is
atomic. Then (B2) ensures that the equivalence holds for 1fr. Suppose the main logical operator
in 1/r is a connective. Then the induction step is trivial. The only non-trivial induction step
concerns the universal quantifier. ('3' is regarded as an abbreviation of'--.¥--.'.) So suppose 1fr
is Vve. We prove only the left-to-right direction of the equivalence, as the other direction is
analogous. Assume that n l=u 1/r. We want to show that N F §u 1/r. It suffices to show that
N Fp e for every variable assignment p based on Do and D 1 that differs from §a at most in
what it assigns to the variable v. Suppose vis of the extended sort; the case of basic variables is
analogous but easier. Let p be just like §a except that it assigns d to v. By (Bl) there is an a E Do
such that a "" a and d = §a. Let a ' be the variable assignment which is just like a except that it
assigns a to v. Then our assumption that n l=u 1/r ensures that n l=u' e, whence the induction
e.
hypothesis yieldsN l=p as desired. -1
The following corollary is immediate from our two last propositions.

Corollary 8.1 Let a be a variable assignment for £1 based on Do. Then, for any formula 1fr
of£ 1> we have:
1fr is assertible relative to a iff N F§u 1/r
This corollary makes good on my claim in Section 8.2 that the formulas that are assertible are
precisely those that are true when the extended sort is interpreted as concerned with abstract
letter types.
8.C Internally Representable Abstraction

Our approach to abstraction has been unusual in that the (partial) equivalence relation on
which we abstract has not been assumed to be represented within the object language £ 1
itself. 37 Let us now examine the special case where £ 1 does contain the expressive resources
needed to represent the abstraction.
Definition 8.1 Let us say that the abstraction is internally representable in £ 1 just in case, for
every predicate P of the extended sort:
(i) Pis of the form cp* for some formula cp of the basic sort which satisfies the requirements
(8.1) and (8.2);
(ii) the assertibility condition associated with P, namely ap, is identical with the truth-
condition that I assigns to <p.
As an illustration of (ii), suppose Pis ..l *. Then ap(l1> /i) just in case Ii and 12 stand in the
relation-presumably just orthogonality-which interprets the basic predicate '.l'.
With these asswnptions in place, it is possible to translate from the extended to the base
language. To be precise, there is a logically acceptable translation r : £1 --+ Lo, as explained in
Section 6.3.2.
We have described three different ways in which a formula 1/1 of £ 1 can be assessed,
relative to an assignment: as assertible, and as true on either of the interpretations R and N .
The reductive translation makes available yet another way to assess 1/1: Is its reductive trans-
lation r(l/!) true on the interpretation 'I? This fourth form of assessment corresponds to the
holophrastic reductionism discussed in Section 5.3. Formulas of £ 1 are not taken at face value
but are first translated by r into the base language .Co and only then subjected to the semantic
interpretation I . Pleasingly, this new form of assessment turns out to agree with the earlier ones.
Proposition 8.3 Let 1/1 be any formula of £1. Let a and p be variable assignments for £ 1 and
Lo, respectively, both based on Do. Assume that a(x;) = p(x;) and a(y;) = p(,i;) for each
i E w. Then:
R F=a 1/1 iff I Fp r(l/!)
In particular, any sentence 1/1 is true on R just in case r (1/1) is true on I.
(The proof is a straightforward induction on the syntactic complexity of 1/1 .)

Taken together, Propositions 8.1, 8.2, and 8.3 establish that all of our four forms of assessment
agree with one another, provided that assignments to any free variables are appropriately
coordinated.
37
Indeed, the fact that £ 1 might lack these expressive resources played an important role in my
argument in Section 8.4.3. However, it is unproblematic to assume that these resources are added at a later
stage when the community of speakers become more sophisticated.
APPENDICES 157
Can anything informative be said about what formulas are assessed as either true or
assertible? Let us recall from Section 6.3.4 that the elementary theory of abstraction is the
L1-theory whose axioms are the universal closures of the following equations:
(ETAl) §xo = §x1 ++ xo ~ x1
(ETA2) ~*(§xo, . . . , §xn) ++ ~(xo, .. . ,xn)
It is easy to verify that each axiom of this theory is assertible and therefore also true on N, R,
and the holophrastic reductionist interpretation.
8.D Defining a Sufficiency Operator

We wish to add a sufficiency operator => to the extended language L1> as first envisaged in
Section 1.5. Let Libe the resulting language. For convenience, let us assume that contains £.i
the extra set of basic variables Xo• Xi •... that are available in Lo. 38 Moreover, to keep matters
as simple as possible, we only allow the sufficiency operator to be flanked by formulas of L 1 .
That is, Li does not permit nested applications of the operator (as in, say,~=> (1/1 => 8)). 39
Our task is to extend the non-reductionist interpretation N of L 1 to an interpretation N+
of Li- Once again, we shall assume that the abstraction is internally representable in L 1
(cf. Definition 8.1). Without this assumption, there is no hope of expressing the relations of
sufficiency within Li.
Our central idea is to let N+ satisfy r(l/I) => 1/f, for every 1/f from L1> relative to every
variable assignment a such that a (yi) = §a C.i;) for any Yi that occurs free in 1/f. Additionally,
we need to ensure that our definition is closed under the operation of replacing an occurrence
of a singular term t with another singular term t1 that receives the same value as t relative to a.
So we lay down that
for any singular terms t and t' that satisfy the mentioned requirement. Finally, we stipulate that
a sufficiency formula is not satisfied relative to an assignment unless it can be shown to be so
by means of the two preceding clauses.
It remains only to observe that our sufficiency operator is well defined and has the desired
properties.
Proposition 8.4
(a) N+ satisfies all i nstances of Leibniz's Law.

(b) If N+ l=o- ~ => 1/f, thenN+ l=o- ~ ~ 1/f.
(c) N+ satisfies the universal closure of every formula of the form:
f\ Yi= §.i; ~ (r(i/f) => 1/1)

y; free in 1/r
38
This assumption can be avoided by using the single set of basic variables xo, x 1 , . .. available in .C 1 in a
way that simulates the availability of both of the sets of basic variables available in .Co. To avoid unnecessary
clutter, we shall not bother.
39
Our sufficiency operator is in this respect like Fine's grounding operators; see e.g. (Fine, 2012b). Of
course, we may eventually wish to lift this restriction. See (Litland, 2017) for an attractive way to do so in
the case of grounding.
In particular, N + satisfies the universal closure of each of the following:
x1 ~ x2 :::} §x1 = §x2 -.xi ~ x2 :::} -.§xi = §x2

<p(X1> . . . ,Xn):::} 1p*(§x1, . .. ,§x n) -.1p(X1> ... ,Xn):::} -.1p* (§X1> . . . , §xn)
Proof sketch. (a) is immediate from the second clause of our definition. For (b) , consider
first the case where N+ Fa T ( 1/1) :::} 1/1 for some variable assignment a for .ct
such that
a (y;) = §a (.xj) for any y; that occurs free in 1/1. Let p be the restriction of a to .Co. Assume that
N Fp T(1f!). Since T(1f!) is in .Co, this means that I Fp T(1f! ) . Hence by Proposition 8.3, we
have R Fa' 1/1 for every variable assignment a ' for .C1 based on Do such that a ' (x;) = p(x;)
and a ' (y;) = p (xj) for each i E w. We now apply Proposition 8.2 to establish N F§a' 1/1. Since
§a ' agrees with a on any y; that occurs free in 1/1, our proof of this particular case is done. The
general claim now follows by observing that N validates Leibniz's Law. Finally, (c) is immediate
from our definition. -l
It is important to understand the significance of clause (a). This clause ensures that it is safe
to quantify into the position occupied by §t in the sufficiency statement so as to form a
corresponding de re sufficiency statement. For example, we may infer as follows (by so-called
"exportation"):
xi ~ xz :::} §x1 = §xz , therefore 3y1 (xi ~ x2 :::} Y1 = §x2)

This makes good on a promise made in Section 1.7.
9
The Julius Caesar Problem
9.1 Introduction
Is the natural number 3 identical with the Roman emperor Julius Caesar? Some
peculiar questions of this sort are raised in Frege's Grundlagen. This is the so-called
Julius Caesar problem. 1
There are two kinds of reaction one may have to the opening question. One reaction
is that such questions are not only pointless but downright meaningless. However
much arithmetic you learn, no answer to the opening question will be forthcoming.
Arithmetic tells you that 3 is the successor of 2 and that it is prime, but not whether
it is identical with Caesar. Why does our opening question not receive an answer?
The reason, it seems, is not that the question is hard but that it simply has not been
"provided for''. The other reaction one may have is that the opening question is to be
answered negatively. Since the number 3 is an abstract object and Caesar is not, it
follows immediately by Leibniz's Law that they are distinct. In light of these two kinds
of reaction, what are we to say about "mixed" identi,ty statements such as the one with
which we began?
An influential response due to Bob Hale and Crispin Wright attempts to do justice
to both of the mentioned reactions:
Within a category, all distinctions between objects are accountable by reference to the criterion
of identity distinctive of it, while across categories, objects are distinguished by just that-the
fact that they belong to different categories. (Hale and Wright, 200la, p. 389)
On this view, our ontology divides into categories, each with its own criterion of
identity for the objects in the category. 2 For instance, the identity or distinctness
of natural numbers is determined by considerations of equinumerosity, whereas
in the case of physical bodies, we appeal to facts about spatiotemporal continuity.
However, since natural numbers and physical bodies belong to different categories,
the statement that 3 is identical with Caesar is not determined as true or false in
1
Actually, Frege never raises this particular question. However, since the first such question he discusses
involves Caesar (Frege, 1953, §56), this Roman emperor has become the stock example and has lent his
name to the problem.
2
Hints of this view are found in (Dummett, !98la) and (Dummett, 199la). (Perry, 2002) and
(Williamson, 1990, ch. 9) contain more explicit anticipations.
160 THE JULIUS CAESAR PROBLEM
any such way. 3 So the first reaction-that there is something amiss with the question
of whether 3 is identical with Caesar-is correct when restricted to the canonical
grounds for identity and distinctness that are provided by a criterion of identity:
no such ground exists for either the identity or the distinctness of 3 and Caesar.
But according to Hale and Wright, objects belonging to different categories are ipso
facto distinct; in particular, 3 is distinct from Julius Caesar. So the second reaction
is correct when all grounds for the identity or distinctness of objects are taken
into account.
The plan for this chapter is as follows. After an initial clarification of the Caesar
problem, I consider and reject the response that all mixed identities can be dismissed
as meaningless. This response fails to do justice to the second of the mentioned
reactions. Then I turn to the response by Hale and Wright outlined above. I have
considerable sympathy for this response and share the broadly Fregean approach to
reference and objecthood on which it is based. But I deny that there is any strong
logical or conceptual support for the claim that every object belongs to a unique
category. 4 I show that it is logically coherent for categories to overlap. It follows
that, when we develop our linguistic practices, we have some degree of choice about
whether or not to allow categories to overlap. To handle mixed identity statements,
we often need conceptual decisions, not just factual discoveries.
This analysis has important consequences for how the Caesar problem should be
approached. My analysis emphasizes the importance of choice and the insufficiency
of always looking to the world for an answer. When our ancestors first confronted
Caesar-style questions, they had a choice which way to go; and this choice played a
role in shaping the concepts that they thereby forged. Today we find ourselves in a
different situation, since many choices are already implicit in the linguistic practices
that we have inherited. Of course, insofar as we are willing to revise these practices, we
still have the same choice as our ancestors had. But we face an important additional
question not encountered by our pioneering ancestors, namely what conceptual
decisions are implicit in our inherited linguistic practices. I shall argue that these
practices have by and large legislated against the overlap of categories. But exceptions
are certainly possible and very likely even actual.
9.2 What is the Caesar Problem?

The Caesar problem arises in the Grundlagen when Frege observes that the direction
abstraction principle (Dir) does not
3 I shall henceforth assume that persons belong to the category of physical bodies. My purpose in doing
so is merely to simplify the discussion, not to make any controversial metaphysical claim. Readers are free
to substitute their own favorite view of persons and adjust tbe argument of this chapter accordingly.
4
In fact, Hale and Wright too express some reservation at tbe end of the article where the quoted passage
occurs.
WHAT IS THE CAESAR PROBLEM? 161
decide for us whether England is the same as the direction of the Earth's axis-if I may be
forgiven an example which looks nonsensical. Naturally no one is going to confuse England
with the direction of the Earth's axis; but that is no thanks to our definition of direction.
(Frege, 1953, §66)
This passage may suggest that the Caesar problem is purely internal to Frege's program
or its abstractionist successors. Of course we all know that England is distinct from the
mentioned direction. But since this cannot be derived from the relevant abstraction
principle, the principle must be deemed inadequate as a "definition of direction': This
reaction is unwarranted. The Caesar problem is both deeper than it initially appears
and more general.
Let us start with its depth. How do we actually come to know that Caesar and the
natural number 3 are distinct? A natural answer is that this knowledge is based on
an application of Leibniz's Law to the prior knowledge that one of the objects has a
property that the other lacks. For instance, Caesar was assassinated in 44 BC, whereas
the number 3 was not. But this is too quick! How do we actually know that the
number 3 was not assassinated and perhaps lives on in a different guise? It is hard to
see how this fact can be known by ordinary empirical means. What kind of knowledge
is this, then? How do we know that the number 3 does not have properties that far
outstrip our conception of it? 5
The Caesar problem is also more general than it initially appears. The problem
arises whenever we have a many-sorted language and wonder whether to "merge its
sorts" by formulating some associated one-sorted language. 6 One example is the two-
sorted language£ 1 from Chapter 8, which has one sort for "basic" objects and another
for the new objects introduced by an abstraction principle. Should these two sorts
be merged, and if so, how? These are questions of fundamental importance to any
abstractionist approach to mathematics.
Another example concerns systems of abstraction principles rather than individual
such principles, which is all we have considered thus far. Let ABs(x, R, a) mean that
x is the R-abstract of a; that is, that x is the result of abstraction on a under the partial
equivalence relation R. Suppose abstraction is permitted on some family of partial
equivalence relations R:
Raa ~ 3xABs(x, R, a)
ABs(x, R, a ) /\ ABs(y, R, f3) ~ (x =y # Raf3)
As a convenient shorthand, we shall sometimes represent such systems of abstraction

principles in the more familiar functional notion:
5 See (MacBride, 2006) for a related observation.

6 See (Enderton, 2001 , pp. 296- 9) for a simple way of doing so, which also corresponds to the approach
to be described in Section 9.6.
It must be borne in mind, however, that each operator §R corresponds to a partial

function whose domain is the field of R. In any event, our official notation is the
relational one displayed in (ABsR3) and (ABSR =).What should we say about mixed
identities of the form §Ru = §sv? We would like either a principled reason to dismiss
such questions as meaningless, or alternatively a systematic answer of the form
for some suitable condition <fJ.
9.3 Many-sorted Languages

Let us begin with the option of regarding mixed identity statements as meaningless.
This suggests a regimentation in a many-sorted language, where the above claim about
meaning receives a syntactical expression in the view that an identity is well-formed
just in case the identity sign is flanked by terms from the same sort. On this approach,
a logical sort is thus used as the syntactic correlate of the ontological notion of a
category.7 A view of this sort has been defended by (Heck, Jr., 1997c) as a neo-Fregean
response to the Caesar problem. 8
It has been objected that this approach to mixed identity statements conflicts with
a realist conception of the objects in question. If a and b exist independently of us,
then the question of whether a is identical with b must be meaningful and have
a determinate answer (although not necessarily one we are able to discover). For
instance, (Field, 1989, p. 22) claims that "if the number 2 is a definite object, and the
set {0, {0}} is a definite object" there has "to be a definite question as to whether the
former is the latter''. 9 However, this objection is unconvincing. How, exactly, is realism
about objects meant to entail that there is only one sort of object? I am unable to
identify any non-question-begging defense of this alleged entailment. A comparison
may help bring out the problem. Frege was a realist about both objects and concepts,
yet vehemently denied that objects and concepts belong to the same category and can
7 Of course, not all many-sorted languages are that strict. One may for instance use separate sorts for
natural numbers and reals, while holding that every natural is also a real and thus admitting mixed identities
as meaningful.
8 It is natural to worry that the use of a many-sorted language will block some of Frege's technical
arguments. In particular, his elegant "bootstrapping argumenC aimed to show that every natural number
has a successor, seems vulnerable, as this turns on counting not just ordinary objects but numbers
themselves. Heck shows how these technical obstacles can be overcome. The key is to allow numbers to
count items from both sorts, that is, both ordinary objects and numbers. (See the article for details.) Notice
that although Heck's versions of Hume's Principle are two-sorted, it is essential to his reconstruction of
the bootstrapping argument that one of them be impredicative (cf. Section 6.4).
9 An analogous objection comes up in the literature on non-eliminative (or ante rem) structuralism,
where it is frequently claimed that cross-structural identity statements are meaningless. For example,
(Shapiro, 1997, p. 79) writes that "it makes no sense to pursue the identity between a place in the natural
number structure and some other object, expecting there to be a fact of the matter. Identity between natural
numbers is determinate; identity between other sorts of objects is not, and neither is identity between
numbers and the positions of other structures:· Confronted with the objection that realism requires
determinate identities, Shapiro is inclined to abandon his earlier indeterminacy claims in favor of the view
that mixed identity statements are determinately false (Shapiro, 2006a, Section 2).
SORTALS AND CATEGORIES 163
thus meaningfully be said to be identical or distinct. If this Fregean combination of

views is coherent, it follows that realism can be compatible with a metaphysics that
calls for regimentation in a many-sorted language. 10
A far better objection, it seems to me, is that it has proved nearly irresistible to
compare and contrast objects from different categories and that such comparisons
are often of great philosophical importance. For example, it is plausible to think that
the number 2 is abstract while Caesar is not, and that 2 is a natural number while
the set {0 , {0}} is not. Indeed, such claims provide important information about the
relevant objects. Someone who does not realize that 2 is abstract and a natural number
is missing some very important truths about this object. To accommodate such truths,
we need to lift the categorical restrictions found in many-sorted languages in order
to bring the relevant objects under a common rubric. Only then can we compare and
contrast objects from different categories.
Thus, while there appears to be no deep logical or conceptual problem with a
metaphysics that is regimented in a many-sorted language, this is not what we
actually do. We are in fact quite fond of cross-categorical comparisons; and in order
to make these, we resort to a one-sorted language. I shall return to this theme in
Section 9.6.
9.4 Sortals and Categories

The alternative option, when confronting the Caesar problem, is to regiment our
language as an ordinary one-sorted language and thus be forced to take a stand on
mixed identities. I shall now consider an influential attempt to do so, namely the one
pursued in (Hale and Wright, 200lb). The strategy is first to characterize the notions
of sortal concept and category, and then to use this characterization to pronounce on
the truth of mixed identity statements. These two steps are discussed in this section
and the next, respectively.
A common informal characterization says that a concept F is a sortal if use of it
in predication calls for the indefinite article. For instance, since we say "That is a cat"
but not "That is a heavy': this characterization implies that the concept cat is a sortal
but that heavy is not. A slightly more precise characterization says that a concept F
is a sortal if it allows for meaningful questions about identity and numerosity.11 •12
10
See (Rayo, 2016) for another defense of the compatibility of realism with regimentation in a many-
sorted language.
11
Definitions of this general sort are found in (Hale and Wright, 2001b), (Lowe, 1997), and (Wright,
1983, p. 2). Deriving from (Strawson, 1959) there is also a more Aristotelian conception of sortalhood,
which requires in addition that a sorta! concept be an answer to the Aristotelian "What is it?"-question and
thus be such that, if it applies to an object, it cannot cease to do so; for instance, a horse allegedly cannot
cease to be a horse. See also (Dummett, 198la, p. 76) and (Wiggins, 2001, pp. 8-11). I make no attempt to
capture this Aristotelian requirement.
12
Arguably, there are concepts F such that it is impossible to assign any determinate number to the
totality of Fs-if such a totality exists at all-despite the fact that grounds have been provided for the
identity or distinctness of any given pair of Fs. For instance, although the principle of extensionality allows
us to identify and distinguish any given pair of sets, one arguably cannot count the number of sets. For
For instance, it makes sense to ask whether this cat is identical with that cat or how
many cats there are in this room. By contrast, if we replace the predicate 'cat' with the
predicate 'heavy', the corresponding questions make no sense. So again it turns out
that cat is a sortal but heavy is not.
Let us attempt to be more precise. We saw in Chapter 2 that criteria of identity come
in two main kinds, which we called one-level and two-level:
(IL) Fx /\ Fy-+ (x = y ++ R(x, y) )
(2L) Raa /\ R/3/3 -+ (§Ra = §R/3 ++ R(a, /3))
In both cases, we say that R is the unity relation associated with the criterion. Of course,
for (IL) to be informative, the unity relation cannot just be identity; and an analogous
restriction applies to (2L). Henceforth, we tacitly restrict ourselves to informative
criteria.
The unity relations must be understood intensionally, not extensionally (say, by
identifying them with appropriate sets). When we grasp a relation, we generally grasp
its intension, not its extension. Moreover, since there is no fixed domain of objects
to which we want to apply a unity relation, the relation must be independent of
any particular domain, and this is only possible when the relation is characterized
intensionally.
We now define what it is for a concept to be associated with-or, for short, to have-
a criterion of identity and to be a sortal.
Definition 9.1 (Sortals) A concept F has a one-level criterion of identity {lL) just in
case the universal closure of (lL) is necessary. 13 A concept F has a two-level criterion
of identity (2L) just in case the following two claims are necessary:
{9.1) Vx(Fx-+ 0 3a(x =§Ra))
(9.2) x =§Ra Ay = §R/3-+ (x = y ++ Raf3)
A concept Fis a sorta[ just in case it has a (one-level or two-level) criterion of identity.
This definition is close to Hale and Wright's, although some possible differences
are discussed in the next section. Moreover, the definition accords well with the
traditional characterizations of sortalhood discussed above. The meaningfulness of
identity contexts is a necessary condition for sortalhood because every sortal concept
has a criterion of identity, which provides grounds for the identity and distinctness of
the objects in question. And the meaningfulness of (most) questions about numeros-
ity is a necessary condition for much the same reason.
Definition 9.1 entails the following useful fact.
this reason, I regard it as more fundamental that a sorta! provides grounds for judgments of identity and
distinctness than judgments about cardinality.
13 In fact, (Hale and Wright, 200lb) sometimes require that the necessity in question be "conceptual".
I omit this qualification as it is not needed for the ensuing discussion.

SORTALS AND CATEGORIES 165
Fact 9 .1 Whenever a concept has a criterion of identity, any sub-concept has the same
criterion.
For example, if the concept human being has a certain criterion of identity, so too
does the concept woman. Notice, however, that our definition does not require that
a sortal have a unique criterion of identity. This observation will be important below.
But it does follow that any two one-level criteria that are had by a single sortal Fare
necessarily equivalent when restricted to Fs.
I turn now to the notion of a category, which is usually characterized as a maximally
inclusive sortal. Plausible examples include concepts such as direction, natural num-
ber, and physical body. A more precise definition, which is essentially that of (Hale
and Wright, 200la, p. 389), can be formulated as follows.
Definition 9.2 (Categories) A concept Fis a category just in case
(a) Fis a sortal that has a criterion of identity, say (*), and
(b) for any x such that -.Fx and any sortal concept G such that Gx, G does not have
a criterion of identity based on the same unity relation as ( *). 14
This definition coheres well with the informal gloss provided above. For instance, the
concept direction is a category because no non-direction is subject to a criterion of
identity based on parallelism as its unity relation.
It should be noted that Definition 9.2 helps itself to a notion of identity of unity
relations, which raises the question of how such relations are individuated. I shall
not attempt to answer the question here. Let me simply remark that necessary
coextensionality provides a fairly plausible option and that none of what follows is
sensitive to the details of the answer.
Definition 9.2 entails another useful fact, namely that the totality of objects that are
subject to some criterion of identity form a category. Or, more precisely:
Fact 9.2
(a) Suppose Fis a sortal that has a one-level criterion of identity based on a unity
relation R. Suppose G is the most inclusive concept whose one-level criterion
of identity is based on R. Then G is a category that subsumes F.
(b) Suppose F is a sortal that has a two-level criterion of identity based on a unity
relation R. Suppose G is the most inclusive concept whose two-level criterion
of identity is based on R. Then G is a category that subsumes F.
14
Hale and Wright's definition includes an additional clause requiring that every sorta! G that is a sub-
concept of Falso have (*) as a criterion of identity. But at least with the definitions adopted here, this clause
is provably redundant, as is seen from Fact 9.1. Furthermore, notice that any difference in our respective
ways of understanding the notion of a sorta! will carry over to anything else defined in terms of this notion,
such as the notion of a category.
9.5 The Uniqueness Thesis

Many philosophers believe every object belongs to a unique category. Does this
uniqueness thesis, as I shall call it, follow from suitable characterizations of the notions
of sortal and category? Hale and Wright appear to think so. Speaking of a notion
of a category that is very similar to mine, they claim that "it is a consequence of
this characterization that a sorta! concept can be subsumed only by one category"
(Hale and Wright, 200la, p. 389). Given our definitions, however, this claim is
incorrect. These definitions are compatible with the overlap of categories, as two
simple examples reveal. 15
Example 9.1 Consider the following two criteria of identity:

(HP) #F = #G ++ F ~ G
(HPo) #oF = #oG ++ (F ~ G) v (.....,FIN(F) /\ -iFJN(G))
where FIN(F) is a formalization of the claim that the Fs are (Dedekind) finite. 16
Clearly, the criteria agree on finite concepts but disagree on infinite ones. It follows
that it is consistent to identify the ranges of # and #o on finite concepts while
simultaneously letting #F =/= #oF whenever Fis infinite.
By Fact 9.1, this example shows that it is consistent to let the concept finite number
have both of the above criteria of identity. This in turn means that it is consistent to
let a sorta! concept be subsumed under two distinct categories. For let C and D be
concepts applicable to all objects in the range of# and #o, respectively. By Fact 9.2,
C and Dare distinct categories that subsume the sorta! concept.finite number.
Example 9 .2 Let Ebe a relation that obtains between two letter tokens just in case they
count as tokens of the same type of the English alphabet. Let S be the corresponding
relation for the Spanish alphabet. We assume (with the usual idealization) that both
E and S are partial equivalence relations. Consider now the following two tokens: n
and fl. These count as tokens of the same letter type of the English alphabet, yet as
tokens of different types of the Spanish alphabet. (These classifications are manifested
for instance in how words are alphabetized in the two languages.) So we have E(n, fl)
but not S(n, fl). It nevertheless seems natural to say that a token of the form 'a presents
the same letter type in the two alphabets. Certainly, doing so is consistent. Moreover,
this is fairly plausible as a characterization of our actual practice. This means that the
letters on which the two alphabets agree are subsumed under two different categories:
that of English letters and that of Spanish letters.
15 As Bob Hale reminded me, it is not clear how deep our disagreement is, as he and Wright too express
reservations about the uniqueness thesis at the end of their article (cf. note 4). See (Fine, 2002, pp. 46-54)
and (Cook and Ebert, 2005) for other discussions of the Caesar problem that express skepticism about the
uniqueness thesis.
16 See (Heck, Jr., !997a) for further discussion.
HALE AND WRIGHT'S GRUNDGEDANKE 167
As mentioned, Hale and Wright hope to derive the uniqueness thesis from suitable
characterizations of the notions of sortal and category. This hope is dashed by our
examples-unless one can show that Hale and Wright's definitions differ from mine.
One might wonder, for example, whether their (intended) definition of a sortal
requires that such concepts have at most one criterion of identity. Suppose this
modification is adopted. Then the desired conclusion follows that a sortal can be
subsumed under at most one category. But the bump in the carpet emerges elsewhere.
To see this, recall Example 9.1, where the concept of a finite cardinal seems to be
subject to two different criteria of identity. The modified definition of a sortal makes
us unable to show that the concept of a finite cardinal is a sortal at all! In short, the
proposed modification ensures that every object belongs to at most one category but
instead threatens the claim that every object belongs to at least one category. So the
envisaged modification of our definitions fails to establish the uniqueness thesis. 17
The moral, I believe, is that we should give up on a strong conceptual support
for the uniqueness thesis. The thesis is simply not forced upon us. This suggests a
different methodology. Rather than attempt to derive the uniqueness thesis from first
principles, we should ask what choices concerning cross-category identity statements
are implicit in our existing linguistic practices and which such choices make sense
when developing new practices. This alternative methodology is pursued in the
next section.
9.6 Hale and Wright's Grundgedanke

According to Hale and Wright, two sortals overlap only if their associated criteria
of identity coincide; indeed, they characterize this as their Grundgedanke (Hale and
Wright, 200lb, p. 389). Restricted to two-level criteria of identity-which are our
primary concern-the Grundgedanke corresponds to the following conditional:
(G)
In fact, since the converse conditional is unproblematic and follows immediately

from the ordinary criterion of identity, (G) might as well be strengthened to the
corresponding biconditional.
What is the relation between the Grundgedanke and the uniqueness thesis? It is not
hard to see that the former entails that the latter holds for all objects that are subject
to some two-level criterion of identity. 18
17
Another possible definitional difference is that Hale and Wright may be operating with the notion
of a "pure sorta!': which imposes an additional requirement that is not part of our definition of a sorta!,
namely that, when an object falls under a sorta!, it do so essentially. As far as I can see, however, this will
have no impact on the validity of the uniqueness thesis.
18
To see this, consider two categories C and D. It follows that C and D must be associated with different
criteria of identity. For by Fact 9.2, a category can be identified with the concept of being an object that
Is the Grundgedanke right? The thesis is attractive in several respects. 19 To begin

with, it enjoys a pleasing minimality. As noted, we know that two specifications
a and f3 subject to the same unity relation R determine the same object just in case
Raf3. The Grundgedanke amounts to adding: that's it. In other words, the thesis
instructs us to admit no identities other than those that the ordinary criteria of
identity require us to admit; all others are to be regarded as false. Furthermore, the
Grundgedanke accords well with our inclination to treat the criterion of identity to
which an object is subject as yielding information about the object itself. Consider
natural numbers and persons, whose criteria of identity turn on completely different
considerations-regardless of which of the usual accounts of personal identity is
correct. We are inclined to regard this fact as revealing an important difference
between the two sorts of object. It says a lot about natural numbers and physical
bodies that the canonical way to identify and distinguish such objects is on the basis
of, say, cardinality comparisons of concepts and considerations involving natural
spatiotemporal continuity, respectively. Since we assume that these two forms of
individuation correspond to incompatible properties of the objects, we conclude that
no natural number can be identical with a physical body.
One might, of course, question why the attractive features just mentioned should
carry any weight. Why should reality be arranged in a way that we humans find
attractive? But recall that our current methodology is to ask what choices concerning
cross-categorical identities are implicit in our conceptual practices or would make
sense when designing new ones. Given this methodology, the mentioned attractions
of the Grundgedanke are obviously significant. The fact that we find certain principles
intuitive and attractive provides evidence that these principles are in fact implicit in
our practices. Even in the absence of any deep logical or conceptual necessity to shore
up the Grundgedanke, our fondness for categorizing objects suggests that we tend
to operate with the principle, even if only tacitly and subject to isolated exceptions
(as suggested by the above examples). 20
It is instructive to pause to recall Frege's objection in Grundlagen §67 to a proposal
that shares some structural similarities with the Grundgedanke. The proposal is that
is subject to the relevant criterion of identity. But the Grundgedanke ensures that concepts associated with
different criteria of identity cannot overlap.
19
Following Hale and Wright, we restrict ourselves to abstraction on equivalence relations. The
Grundgedanke would be far less plausible if extended to abstraction on partial equivalence relations (cf.
Appendix 2.C). To see this, let "" (as before) represent the equivalence relation of two concepts' being
equinumerous. Let ""o be the partial equivalence of two concepts' being finite and equinumerous. Then,
for any finite concept, it is hard to resist identifying the two resulting abstracts.
Suppose we allow such identification. This opens for the possibility of abstracting on only some of
the entities in the field of a given equivalence relation ~. Let ~o be the partial equivalence obtained by
restricting ~ to the entities on which we wish to abstract. On the mentioned supposition, abstraction on
~o is tantamount to ~-abstraction applied only to the desired entities.
20
If our account of reference can be extended to the relevant sorts of object, these considerations can
also be used to answer MacBride's question mentioned in note 5 and the text to which that note is attached.
ABSTRACTION AND THE MERGING OF SORTS 169
the concept of direction can be defined by stipulating that "q is a direction if it was
introduced by means of the definition" that 'd(l1) = d(/2)' is to be true just in case
11 and 12 are parallel. Frege protests that a "definition of an object does not, as such,
really assert anything about the object, but only lays down the meaning of a symbol:'
For this reason, Frege says, we cannot treat the way in which an object is introduced
as a property of the object; in fact, any object can be introduced by any number of
different definitions. Although this objection is surely devastating against its intended
target, it does not threaten the Grundgedanke. For the relation between a symbol and
its referent is very different from the relation that a specification and a unity relation
bear to the referent that they determine. The relation between a symbol and its referent
is entirely contingent. Tradition and convenience aside, any symbol can be used to
denote any object. By contrast, the fact that an object can be determined by a certain
type of specification, subject to a certain unity relation, can coherently be taken to
present information about the object. Given the Grundgedanke, the object could not
have been subject to any other unity relation, which ensures that being subject to a
particular unity relation (and thus also a particular two-level criterion of identity) is
a property of the object itself.
9. 7 Abstraction and the Merging of Sorts

The abstractionist approach to thin objects developed in Chapter 8 is subject to
the major limitation that it only justifies abstraction in a two-sorted language. This
awkwardness needs to be eliminated for my approach to be fully satisfactory. For-
tunately, my approach is easily modified to handle abstraction in an ordinary one-
sorted language, as we shall now see. The conceptual freedom with regard to category
distinctions that has been emphasized throughout the present chapter plays an
important role in this endeavor.
A brief reminder of the abstractionist approach may be useful. We described a
two-sorted extended language £1 and some precise rules that govern its use. These
rules take the form of assertibility conditions which, if observed, ensure that the
language is used precisely as if we have abstracted on some of the objects in the original
domain. My main example involved abstraction on letter tokens to obtain abstract
letter types. I would now like to formulate analogous assertibility conditions for the
one-sorted extended language £2 that is obtained from .C 1 by deleting all type indices
and ignoring all restrictions on well-formedness that have to do with such indices. By
lifting type restrictions in this way, we are obviously forced to take a stand on Caesar-
style questions.
The assertibility conditions that govern the one-sorted extended language are based
on some straightforward ideas. 21 Assume an agent points at a token a. To what
21
A precise characterization is provided in Appendix 9.A.
object does she refer? One possibility is that she points at a in order to refer to
a itself. Another possibility is that she is using a to specify a type, which is her
intended referent. 22 An English speaker might disambiguate between the two possible
interpretations by using a complex demonstrative, such as 'this letter token' or 'that
letter type: But ultimately, the correct interpretation depends on how the agent is
disposed to assess identity statements concerning the referent. If the intended referent
is the token a itself, then a second demonstration, this time pointing at a token b, will
be judged to be a demonstration of the same object just in case a = b. On the other
hand, if the intended referent is the type of a, then the other demonstration (where
the agent points at b) will be judged to be a demonstration of the same object just in
case a ,....., b. In other words, depending on whether the intended referent is a token
or a type, identity statements concerning the referent are taken to turn on whether
the tokens at which the agent points are identical or bear the unity relation to each
other, respectively. Let us assume, for simplicity, that the speaker signals her intended
interpretation by accompanying her pointing with an appropriate predicate so as to
form a complex demonstrative: 'this token' or 'that type'.
What about Caesar-style identity statements such as 'this letter token = that letter
type'? In this particular case, it is natural to invoke the uniqueness thesis and thus
to regard all such mixed identity statements as false. As we have seen, doing so is
logically and conceptually unproblematic. Let us therefore write into the assertibility
conditions that govern the one-sorted language £2 that no letter type is to be identified
with any "old" object that was present already in the domain associated with the base
language Co. In other cases, however, it is far more natural to identify some of the
"new" abstracts with an appropriate "old" object. Recall plural Law V, which was
discussed in Chapter 3:
{xx} = {yy} # 'Vz(z -< xx# z-< yy)
Suppose the domain on which we abstract already contains some sets. Then we have
good reason to identify a "new" set with an "old" set that has precisely the same
elements. 23 Suppose, for example, that the domain of the base language contains the
three objects a, b, and {a, b}. Suppose we carry out set abstraction on the plurality of
a and b. Then we have good reason to identify the "new" set of a and b with the "old"
one. In this way, we ensure that the Law of Extensionality remains valid.
The conceptual freedom emphasized throughout this chapter ensures that we need
not give a uniform answer to the question of whether or not the "new" objects
22 This is what (Quine, 1968) calls "deferred ostension".

23 In fact, this identification is fully compatible with the uniqueness thesis. On the dynamic approach to
set theory developed in Chapter 3, the "old" sets too were obtained by abstraction on pluralities under the
equivalence relation of coextensionality. Thus, "old" and "new" sets alike are subject to coextensionality as
their (two -level) criterion of identity.
APPENDICES 171
introduced by some abstraction should be identified with any of the "old" objects that
were available prior to this abstraction. Different cases will call for different answers.
Two comments are in order before closing. First, our discussion has a very
important consequence concerning the principle of compositionality. We have just
described a language for talking about both letter tokens and letter types. There is no
compositional interpretation of this language which is based solely on the domain of
tokens and which respects the mentioned assertibility conditions. To see this, consider
the identity predicate. Suppose the domain consists just of letter tokens. Then, to
respect the assertibility conditions concerned with letter tokens, the identity predicate
must be satisfied by two tokens just in case they are identical. However, to respect the
assertibility conditions concerned with letter types, the predicate must be satisfied
by two tokens just in case they are equivalent. Clearly, it is impossible to satisfy both
of these constraints simultaneously. This observation reveals an important difference
between the one- and two-sorted settings. A semantic interpretation of some two-
sorted first-order language was defined in Chapter 8 and shown to be compositional,
to be based on the domain of tokens (and in this s.ense reductionist), and to respect
the assertibility conditions that we described. As we have just seen, this kind of
interpretation is no longer possible in the one-sorted setting.
My second comment concerns a phenomenon that arises in cases of abstraction
on pluralities. 24 In the two-sorted setting, there is a basic sort, concerned with the
"old" objects on which we abstract, and an extended sort, concerned solely with
"new" objects obtained by abstraction. While the basic sort has plural resources, the
extended one does not; indeed, it is this expressive limitation that makes it possible to
formulate the assertibility conditions for the extended language in the base language
itself. But when the separate domains for the two sorts are merged, we allow the plural
resources to be defined on the entire domain that results. Is this permissible? I believe
it is. Starting with an interpreted one-sorted language, we have shown how to shift
the interpretation of the language in a way that results in an extended domain. Since
the "new" objects that have been added are in good standing, it is permissible to talk
about them plurally as well as one by one.
Appendices
9.A The Assertibility Conditions
The assertibility conditions for £2 were informally canvassed above. Let us now be precise.
To do so, we shall make use of a domain D1 that is obtained from Do by attaching a label to
each of its elements. We let D1 = Do x {O} US x {1}, where Sis the subset of Do consisting of
tokens. Intuitively, a labeled specification of the form {a, 9) serves to specify the object a itself,
24
The case of abstraction on second-order concepts is analogous.
while a labeled specification of the form (a, 1) serves to specify the abstract letter type of a. Let
us therefore call the two types of specification basic and extended, respectively.
Predicates. For every n-place predicate P there is a formula {3p of the (single-sorted) metalan-
guage LM with n free variables such that, for any C1> .•• , Cn from D1> Pis assertible of ci. ... , Cn
iff {3p(ci. . .. , Cn). We make the following further assumptions.
• Suppose a predicate P, other than'=: is in Lo and that ap is the satisfaction condition for
Pon the interpretation I of L 0 . Then for any {a;, I;) from Dz we have:
(That is, a predicate of the base language, other than '=: is regarded as assertible only of
objects from Do, and as assertible of a string of such objects iff the predicate is true of that
string on the interpretation I.)
• For any elements al> ... , an and b; of Do and any labels lo, ... , In E {O, 1}, we have:
I;= 1 /\a;~ b;--+ (f3p( .. .) ++ {3p( .. .)[(b;, 1)/{a;, l)l)
where' .. .' abbreviates the string '(ai. 11), ... , (an, In)'.
(That is, if P is regarded as assertible of a string of specifications, the i'th of which is
an extended specification, then P is also regarded as assertible of any other string of
specifications that specify the same objects, that is, where the first coordinate of any
extended specification may be replaced by any ~-equivalent token.)
• Suppose f3= is the condition associated with the identity predicate =. Then:
f3=((a1,Ji), (a2,Ji)) ++ ((11=12 = 0 /\ a1 = a1) V (I,= 12 = 1 /\ a1 ~ a1))
(That is, two specifications are regarded as specifying identical objects just in case both
are basic specifications with identical first coordinates, or both are extended specifications
with ~-equivalent first coordinates. 25 )
Singular terms. Let a be an assignment of elements of D1 to each variable of L1. Relative

to a, every singular constant t of L1 is assigned an element c of Di. subject to the following
constraints.
• Suppose tis a constant of Lo. Then c = (a, 0) where a is the referent oft under I.
• Suppose t is a variable. Then t receives its assignment from a.
• Suppose tis a term of the form §sand thats is assigned (a, I) relative to a. Then tis assigned
(a, 1) relative to a. 26
Formulas. The assertibility conditions for formulas of L1 are as follows.
25 This way of handling mixed identities-and thus of resolving the Caesar problem-is plausible in the
case we are considering, as argued in the main text. However, what follows can easily be adapted to any
other formally acceptable way of resolving the Caesar problem.
26 What matters is the case where I = 0. Allowing the operation to be defined where I = 1 is merely a
matter of convenience, corresponding to identifying a "double abstraction" with the corresponding single
abstraction (e.g. the letter type of this letter type= this letter type).
APPENDICES 173
• Suppose rp is an atomic formula P(t1> . .. , tn) and that each singular term t ; is assigned
an element c; of D1 relative to a. Then rp is assertible relative to a iff f3p(ci. . .. , en) .
• The clauses for truth-functional connectives are the obvious compositional ones.
• Suppose rp is of the form Vv 1/f, where v is a variable. Then rp is assertible relative to a iff
for every variable assignment a ' that differs from a at most in its assignment to v, 1/f is
assertible relative to a'.
9.B A Non-reductionist Interpretation

Fix an interpretation I of Lo. I contend that the assertibility conditions for L 2 that we defined
in Appendix 9.A function precisely as if the standard of correctness was truth on some
non-reductionist interpretation that extends I. I wish to define the desired non-reductionist
interpretation N of L2 and to use this definition to prove the mentioned contention.
As in the Chapter 8, we work in a metalanguage that is governed by the same sort of
assertibility conditions and in which we can therefore avail ourselves of the very form of
abstraction in question. Thus, when Do is the domain associated with I, we have the following
abstraction principle in our metalanguage:
(APc) a ~ a /\ b ~ b-+ (§a= §b ~ a ~ b)
where the variables range over Do .

We begin by defining a domain D2 which extends Do by adding §a for each a E Do such that
a ~ a. Moreover, just as in the object language, we assume that all Caesar-style questions are
answered negatively, that is, that a =!= §b for every a, b E Do.
The predicates of L2 are interpreted as follows . The identity predicate is interpreted stand-
ardly on D2. Any other predicate that is present already in Lo retains its old interpretation. Any
predicate P that is present only in L2 is interpreted in accordance with its assertibility condition
f3p , in the following sense. Let us say that b E D2 is the value of a labeled specificiation c = (a, l)
iff either l = 0 and b = a, or l = 1 and b = §a; we symbolize this as [ell = b. Then, for any
string C1> .•• , Cn from D1> we lay down:
Pis trueof[c1Il,. .. , [ cnD onN iff f3p(c1,. . .,en)
Notice that this is well-defined because of the requirements we have imposed on f3p
(cf. Appendix 9.A).
When a is a variable assignment based on Di. we let [a Il be the assignment based on D2 that
is defined by [a Il(x;) = [a(x;)Il. We can now state the desired result, which is easily proved
by induction on the complexity of rp.
Proposition 9.1 Let rp be any formula of L2 and a be any assignment based on D1. Then:
N F[uD rp iff rp is assertible relative to a

This proposition shows that, relative to suitably coordinated variable assignments, a formula
is true on the non-reductionist interpretation just in case the formula is regarded as assertible.
The proposition thus shows that an important feature of the two-sorted language L 1, discussed
in Chapter 8, extends to our single-sorted language L2 as well; see especially Corollary 8.1 .
9. C Defining a Sufficiency Operator

We wish to define a sufficiency operator for the one-sorted language L2. As we recall from
Appendix 8.D, our definition for the two-sorted language L 1 makes crucial use of a translation
from L1 to Lo. No such translation is available for the single-sorted language L2· We shall
therefore restrict ourselves to literals of L2, that is, to atomic formulas and negations thereof.
We shall define a sufficiency operator:=} such that, for every true literal if! of L2, there is a true
formula <{! of Lo such that <{! :=} if! for some suitable assignment (J.
As in the two-sorted setting, we need the assertibility condition of every atomic formula
of L2 to be representable in Lo. We shall therefore assume that the abstraction is internally
representable, essentially in the sense of Definition 8.1, though adapted to our present one-
sorted setting, namely:
Suppose the predicate Pis in L 2 but not in Lo. Then there is an Lo-formula rp which satisfies
requirements (8.1) and (8.2) and is such that Pis assertible of (ai, 1),. .. , (an, l) iff <{!is true of
ai, . .. , an under the interpretation I.
When this condition is met, we often write Pas rp*.

As an illustration, consider the case of direction abstraction, and let P be an orthogonality
predicate on directions (also written as .l *).Then Pis assertible of two direction specifications
(a 1, 1) and (a 2 , 1) just in case the two lines a 1 and a2 stand in the relation-presumably just
orthogonality-which interprets the Lo-predicate '.l'.
As before, the idea is that the assertibility condition of a statement if! suffices for if! itself-
only this idea now needs to be restricted to the case where if! is a literal of L2. We implement the
idea as follows. Consider any singular terms t; and T;, for i = l, . . . , n. Suppose that, relative to
an assignment (J, the value of T; is the §-abstract of the value oft;, for every i. Then we let N+
satisfy the following formulas relative to (J:
ti "" t2 :=? T1 = T2 -.t1 "" t2 :=? -.Ti = T2

:=} <P*(T1> . . . , Tn)
<{!(ti> ... , tn) -.<{!(t1> . . . , tn) :=} -.<P*(Ti, . . . , Tn)
We then obtain the desired result.
Proposition 9.2
(a) N+ satisfies all instances of Leibniz's Law.

(b) If N+ Fa <{! :=} 1/1, then N Fa <{! --+ 1/1.
(c) N+ satisfies the universal closures of the following formulas:
x1 "" Xz :=} §xi = §x2 -.x1 "" x2 :=} -.§xi = §x2

<{!(Xi,. . .,Xn) :=} <P*(§x1,. .. , §xn) -.<{!(Xi, ... ,Xn) :=} -.<P*(§x1, ... , §xn)
So far, we have been concerned with an immediate form of sufficiency. When abstraction is
iterated, as described in Chapter 3, we can also define a mediate form of sufficiency. Consider
an individual abstraction step of the sort that we have studied. We have shown that, for each
literal if! concerned with the extended ontology, there is a statement <P concerned solely with
the previous ontology that immediately suffices for if!. All the literals of the sort just mentioned
fully describe the extended model. We can therefore carry out further abstraction steps, which
result in yet further model extensions. Throughout this process, for each literal if! concerned
APPENDICES 175
with objects introduced at the latest abstraction step, there is a statement <P concerned with the
ontology from the previous step that immediately suffices for 1/f. By stringing together chains
of immediate sufficiency, we can obtain a notion of mediate sufficiency.
Before we can do so, however, it remains to show how literals concerned with a particular
domain suffice for each true logically complex statement concerned with the same domain.
I do not wish to commit to any particular way of doing so. One option is to use brute force
and let the set of all true literals concerned with a certain domain suffice for each true complex
statement concerned with this domain. Less brutal options are possible as well, perhaps by
seeking inspiration from discussions of grounding.27
27
See e.g. (Fine, 2012a).
10
The Natural Numbers
10.1 Introduction
How are the natural numbers individuated? That is, what is our most basic way of
singling out a natural number for reference in language or in thought? In this chapter,
I discuss two competing answers. One answer regards the natural numbers as cardinal
numbers, individuated by the cardinalities of the collections that they number. This
answer is favored by logicists such as Frege and Russell and by their followers. 1
Another answer regards the natural numbers as ordinal numbers, individuated by
their positions in the natural number sequence. This answer is favored by many
constructivists and non-eliminative structuralists. 2
After describing the two alternatives in more detail, I outline some arguments to
the effect that the ordinal conception provides a better fit with our actual practices
of thinking about and referring to natural numbers. 3 I therefore develop the ordinal
conception in more detail and discuss its justification of the axioms of Dedekind-
Peano arithmetic. A subsidiary aim of the chapter is to discuss the phenomenon of
types without any exemplifying tokens. Natural numbers provide a good example.
I argue that it suffices for the existence of an abstract type that it is possible for there
to be exemplifying tokens.
10.2 The Individuation of the Natural Numbers

According to the cardinal conception, the natural numbers are individuated by the
cardinalities of the concepts or the collections that they number. For example, our
most basic way of thinking of the number 5 is as the cardinal measure of quintuply
instantiated concepts or five-membered collections. This view naturally corresponds
to a two-level criterion of identity. According to Frege, a number is specified by
means of a concept which has the number in question, and two concepts F and G
1
See e.g. (Frege, 1953), (Russell, 1919), (Wright, 1983), and (Hale and Wright, 200la).
2
The account is defended by (Dedekind, 1888) and, more recently, by various non-eliminative mathem-
atical structuralists, such as (Parsons, 1990), (Resnik, 1997), and (Shapiro, 1997).
3 I do not claim a perfect fit. Indeed, as Stewart Shapiro reminded me, it is not entirely dear what the
desired account is meant to capture. See (Snyder et al. , 2017) for a useful discussion of this question and
some proposed answers to it.
THE INDIVIDUATION OF THE NATURAL NUMBERS 177
determine the same number just in case the Fs and the Gs can be put in one-to-one
correspondence. Let F ~ G abbreviate the standard second-order formalization of
this requirement. Frege's claim is then that the natural numbers are subject to the
criterion of identity that is now known as Hume's Principle:
(HP) #F= #G ++F ~ G
Similar views have been defended by Russell and the neo-Fregeans Bob Hale and
Crispin Wright. 4
The view that the natural numbers are finite cardinals individuated by (HP) has
some attractive features. Many philosophers find it attractive that the view builds
directly into the identity conditions the application of natural numbers as measures of
cardinality; for example, the number three is individuated as the number that counts
all triples. An impressive feature of this approach is a technical result known as Frege's
theorem. Consider the theory that consists of pure second-order logic and (HP) as
a sole non-logical axiom. Frege's theorem says that this theory and some natural
definitions suffice to derive all the familiar axioms of second-order Dedekind-Peano
Arithmetic and thus all of ordinary arithmetic. 5 It is important to notice, however,
that Frege's theorem is only available when (HP) is understood as an impredicative
abstraction principle, that is, when the number-terms on its left-hand side refer to
objects in the range of the first-order quantifiers that figure on its right-hand side. 6
Without this assumption, we are no longer able to prove that every number has a
successor. 7
According to the competing ordinal conception, the natural numbers are individu-
ated by their ordinal properties, that is, by their position in the natural number
sequence. For example, our most basic way of thinking of the number 5 is as the fifth
element of this sequence. This view too corresponds naturally to a two-level criterion
of identity. A natural number is specified by means of a "counter" or a numeral,
which occupies a unique position in a sequence. Two such counters or numerals
determine the same number just in case they occupy the same position in their
respective orderings. For example, the decimal numeral '5' is equivalent to the Roman
numeral 'V' because both occupy the fifth position in their respective orderings. If we
symbolize this latter relation by "', this can be expressed as the following criterion:
(CI-N) N(m) = N(ii) ++ m "' ii
4
I here gloss over some important differences which are irrelevant to our present concerns. In particular,
Frege and Russell wanted to reduce (HP) to what they took to be more basic class-theoretic notions.
For details, see (Frege, 1953), (Russell, 1919), (Wright, 1983), and (Hale and Wright, 2001a).
5 See (Parsons, 1965) and (Wright, 1983), as well as (Boolos, 1990) for a nice proof.
6 See Section 6.2 for an explanation of the distinction between predicative and impredicative abstraction
principles. Recall that this distinction must not be conflated with the question of whether the background
second-order logic is predicative. See also (Linnebo, 2016a) for an opinionated survey of questions of
predicativity that arise in connection with the neo-Fregean program.
7
See e.g. (Dummett, 1991a, p. 138), (Heck, Jr., 1997c), and (Weir, 1998).
178 THE NATURAL NUMBERS
where the operator N maps a numeral to the associated number. This criterion of
identity will be developed in more detail in Section 10.5.
In light of the preceding chapters, it should come as no surprise that I regard both
the cardinal and the ordinal conceptions as legitimate-at least when developed as
predicative abstraction principles. That is, I believe a predicative version of either
conception could successfully be used by a community in order to refer to objects.
It is nevertheless interesting to ask which of the conceptions provides the most
plausible analysis of our actual practice of referring to the natural numbers. In the
next two sections, I argue that the answer is the ordinal conception. It is important
to realize, however, that the general account of thin objects developed in this book is
not dependent on this conception of the natural numbers but could equally well be
coupled with a cardinal conception.
Does our question presuppose that there are robust and general facts of the matter
about what is responsible for our ability to refer to individual natural numbers? Per-
haps our arithmetical practices involve different but equally legitimate conceptions of
the natural numbers. 8 And perhaps different conceptions of the natural numbers can
be found in cultures with no formal education, among educated laypeople, and among
professional mathematicians. I admit that these outcomes are possible, perhaps even
likely. But I don't think this should discourage us. For one thing, our question might
still allow of an interesting answer when restricted to a particular community. (My
first move would be to restrict the question to educated members of contemporary
Western culture who are not professional mathematicians.) For another, it is generally
good methodology to begin by articulating and exploring various "pure" analyses
of some phenomenon, even while admitting that the phenomenon might turn out
to be too messy to be fully captured by any single pure analysis. When the pure
analyses have been developed and adequately understood, they can later be tweaked
or combined into hybrid analyses.
It is striking that the neo- Fregean program has tended to assume, with little or no
argument, that any abstractionist approach to the natural numbers must be based on
the cardinal conception and some version of Hume's Principle. If nothing else, this
chapter will undermine this assumption by developing an abstractionist version of
the competing ordinal conception.
10.3 Against the Cardinal Conception

As mentioned, I grant that the cardinal conception provides one possible way to
think and talk about the natural numbers. But I deny that this is how we actually
single out the natural numbers for reference in our most basic arithmetical thought
8 See (Moltmann, 2013) and (Snyder, 2017) for some considerations suggesting that there are several
different kinds of use of "number talk" in natural language. See also (Paseau, 2009) for an argument that
our conception of the natural numbers is compatible with a reduction of arithmetic to set theory.
AGAINST THE CARDINAL CONCEPTION 179
and reasoning. I shall now present some objections to the cardinal conception as an
account of our actual arithmetical practice. The objections will be presented roughly
in the order of increasing strength.
Before we begin, I should note that my concern in this section is quite different
from that of some of the leading advocates of the cardinal conception. In particular,
Hale and Wright do not always aim to provide an analysis that matches our actual
arithmetical practice. Their primary aim is to estciblish a possible route to a priori
knowledge of an abstract realm of natural numbers. 9 This is clearly a legitimate and
interesting goal. But in my view the more pressing question is to what extent actual
arithmetical practice provides such knowledge. In particular, how and to what extent
do ordinary, educated laypeople achieve such knowledge? 10
10.3.1 The objection from special numbers
My first objection seeks to show that, if the cardinal conception had been correct,
then certain special numbers would have been obvious and unproblematic in a way
that they are not.
One such number is zero. It takes very little sophistication to know that there are
concepts with no instances. For example, the concept of being a golden coin in my
pocket has no instances. So if our basic conception of natural numbers had the form
# F, then zero would have been a very obvious number. By contrast, the view that our
most basic conception of the natural numbers is as ordinals predicts that zero should
be just as non-obvious and problematic as the negative numbers; for every sequence
of numerals has a first but no zeroth element. As it turns out, it was only at a very
late stage in the history of mathematics that zero was admitted into mathematics as
a number in good standing. 11 This suggests that our most basic conception of the
natural numbers is ordinal-based rather than cardinal-based.
How convincing is this objection? I believe it has force against the view that our
most basic conception of a natural number has the form #F, where Fis a concept.
However, a simple modification of the cardinal-based approach suffices to block the
objection. One may take the most basic conception of a natural number to have the
form #xx, where xx is a plurality of objects. On the reasonable assumption that there is
no empty plurality, this would block the easy route to the number zero. 12 But although
effective against my objection, the response would have some serious drawbacks.
To begin with, the response would be a significant departure from the traditional
9 See e.g. (Hale and Wright, 2000).

10
The importance of this alternative goal is defended in (Heck, Jr., 2000a) and (Heck, Jr.. 2011); cf. also
(Linnebo, 2004b, p. 168).
11
Fibonacci is often credited with the introduction into the European tradition of an explicit numeral
for zero in the early thirteenth century. But apparently even he did not regard zero as a proper number.
According to the three-volume history of mathematics (Kline, 1972), "By 1500 or so, zero was accepted as
a number" (p. 251).
12 It is true that the opposite assumption was made in Section 3.7. But that was motivated entirely as a
technical convenience, not as a matter of fit with natural language.

Frege-Russell view. And more seriously, the response would undermine Frege's
impressive "bootstrapping argument" for the existence of infinitely many natural
numbers, which makes essential use of the number zero. To see this, consider the
numbers up to and including n. When zero is included, we are considering n + 1
objects, which can accordingly be counted so as to yield the successor of n. But clearly,
this elegant move is unavailable when zero is not included. 13
The objection from special numbers can also be developed for various infinite
cardinals. If our basic conception of a natural number had been of the form #F,
then infinite cardinals should have been much more obvious and natural than they
in fact were. For example, not much sophistication is needed to grasp the concept of
being self-identical. And with a sufficient grasp of arithmetic comes a grasp of the
concept of being a natural number. The cardinal conception therefore predicts that
the numbers that apply to these two concepts should have been fairly obvious. But
the history of mathematics reveals that it required the creative genius of Cantor to
take seriously and explore the idea of infinite cardinal numbers. And when he did, he
encountered widespread incomprehension and opposition. This historical evidence
is better explained by the view that our basic conception is an ordinal one based
on numerals. For while there are some very simple concepts with infinitely many
instances, it is far less clear that there are infinitely long sequences of numerals of
the sort that would be required to specify infinite ordinals.
A possible response to this objection would be to modify (HP) so as to assign
numbers only to finite concepts. 14 Let FIN(F) be some formalization of the claim that
the concept F has only finitely many instances. Then let Finite Hume be the following
principle:
(FHP) FIN(F) /\ FIN(G) ---+ (#F = #G ~ F ~ G)
This modification clearly blocks our objection. But as before, this is achieved in a way
that brings with it serious problems. In particular, it is implausible that people with
ordinary arithmetical competence should grasp, however implicitly, the concept of
finitude-or so I argue in Section 10.5.
10. 3.2 The objection from the philosophy of language

Another objection has as its point of departure the observation that in natural lan-
guage, expressions of the form 'the number of Fs' are definite descriptions rather than
purely referential terms. This observation enjoys strong linguistic evidence. Certainly
the surface structure of the expression indicates that it is a definite description rather
than a purely referential term. Moreover, expressions of this form are easily seen to be
non-rigid. For example, the number of bicycles in my possession is 4; so had I bought
13 See (Shapiro and Weir, 2000) for further discussion.

14 See (Heck, Jr., 1997a), who at pp. 590-1 provides a definition (due to Frege) of what it is for a concept
to be finite which does not assume an antecedent conception of the natural numbers.
AGAINST THE CARDINAL CONCEPTION 181
another, the number of bicycles in my possession would have been 5. By contrast,

numerals are rigid designators.
The objection continues by claiming that our most basic conception of an entire
category of objects cannot be based entirely on definite descriptions but must also
involve some more direct form of reference. To understand a definite description, one
must be capable of some more direct way of referring to each of the objects in question,
so as to consider whether it satisfies the description. Consider the sentence 'The tallest
person in this room is male'. To understand this sentence, one needs an ability to
identify people which is prior to and independent of what is provided by definite
descriptions. It is only on the basis of this prior ability that we are able to determine
who (if any) uniquely satisfies the description. More generally, if F is a sorta! and Pa
predicate defined on Fs, then in order to understand 'the Fis P' one needs some more
direct way of referring to Fs than is provided by definite descriptions. 15
Let me be clear about the intended contrast between numerals and descriptions
of the form 'the number of Fs'. The contrast concerns whether the expressions have
an internal semantic articulation. The descriptions cannot serve as a basic mode of
reference to numbers, I claim, because they have an internal semantic articulation
which presupposes some more basic form of reference to numbers. By contrast,
numerals have no internal semantic articulation. This does not mean that there is
nothing we can say about how these terms come to refer; indeed I have argued that
the constitution of their reference involves criteria of identity. But this is an account
of what the reference of a certain semantically simple expression consists in, not an
account of how the reference of a semantically complex expression is determined by
the reference (or semantic values) of its simple constituents. 16
10.3.3 The objection from lack of directness

My third objection aims to show that only the ordinal conception allows the natural
numbers to be given to us with the requisite directness. Let me explain. If there
are five apples on a table, we can think of the number five as the number of apples
on the table. Or, following Frege, we can think of five as the number of (cardinal)
numbers less than or equal to four. But neither of these ways of thinking of the number
five feels particularly direct or explicit. Indeed, many people with basic arithmetical
competence don't even find it immediately obvious that the number of numbers
between 0 and 4 (inclusive) is 5 as opposed to 4.
Rather, the most direct and explicit way to specify a number seems to be by means
of a standard numeral in a system of numerals with which we are familiar. Since
the defining characteristic of a numeral is its position in the numeral sequence, this
suggests that the ordinal conception of the natural numbers is more basic than the
cardinal one. 17
15
For a more developed argument of this sort, see (Evans, 1982), esp. section 4.4.
16 See Section 2.6.2.
17 A view of this sort is defended in (Kripke, 1992).
It may be objected that, since we don't have transparent access to all features of
our thought, we cannot take at face value the kind of phenomenological evidence
that I have just adduced about what kinds of reference feel most direct. This is a
legitimate concern. But although our considerations are particularly powerful when
presented from a first-person point of view, there is no need to present them in that
way. I conjecture that the results would be confirmed by a more objective, third-
personal investigation of cognitive processing of arithmetical claims, for example in
terms of reaction times. 18
10.4 Alleged Advantages of the Cardinal Conception

I now briefly consider two alleged advantages of the cardinal conception.
One alleged advantage is that the applications of the natural numbers are built
directly into their identity conditions. This is just an instance of a more general
requirement sometimes known as Frege's constraint. 19 Should this constraint be
imposed? We obviously need some account of how mathematics is applied. But
why should the account have to build the applications of mathematical objects
directly into their identity conditions?20 Moreover, even if Frege's constraint could
be defended, this would not obviously favor the cardinal conception over the ordinal
one. It is true that the natural numbers are naturally applied as measures of cardinality.
But they are equally naturally applied as measures of ordinality! (For example, "She
finished the race as number 17:') Frege's constraint does not settle which of these kinds
of application should be directly built into the identity conditions of numbers.
Another alleged advantage of the cardinal conception is that it allows for Frege's
famous "bootstrapping argument" for the principle that every number has an imme-
diate successor. Mathematically speaking, this argument is extremely elegant and
interesting. But it is implausible as an account of people's actual arithmetical reasoning
or competence. The argument was developed only in the 1880s and is complicated
enough to require even trained mathematicians to engage in some serious thought.
So this is unlikely to be the source of ordinary people's conviction that every num-
ber has an immediate successor. 21 Moreover, the argument breaks down unless,
first, (HP) is understood as an impredicative abstraction principle, and second, the
background second-order logic is allowed to be impredicative. 22 In particular, this
reliance on impredicative second-order logic in order to justify even the principles of
18
Indeed, most cognitive psychologists appear to think that our capacity for exact representations
of numbers (other than very small ones) is based on our understanding of some system of numerals.
See for example the influential account developed by (Carey, 2009).
19 See (Wright, 2000) for a discussion and partial defense of this constraint.
20
See also (Parsons, 2008, Section 14) for criticism of Frege's constraint.
21 See (Linnebo, 2004b, pp. 168-9).
22
The distinction between these two kinds of impredicativity is explained in Section 6.2. Both claims are
proved in (Linnebo, 2004b ). For the second claim, it is essential that we retain Frege's own definitions of
the primitive expressions of arithmetic. See (Linnebo, 2016a) for discussion and further references.
DEVELOPING THE ORDINAL CONCEPTION 183
predicative Dedekind-Peano arithmetic seems problematic and suggests that some-

thing is awry. 23
10.5 Developing the Ordinal Conception

I now develop the ordinal conception of natural numbers in more detail. 24
Arithmetic teaches us that the natural numbers are notation independent in the
sense that they can be denoted by different systems of numerals. In fact, an awareness
of this notation independence is implicit already in basic arithmetical competence.
Even people with very rudimentary knowledge of arithmetic know that the natural
numbers can be denoted by ordinary decimal numerals, by their counterparts in
written and spoken English and other natural languages, and by sequences of strokes.
Many people also know alternative systems of numerals such as the Roman numerals
and the numerals of position systems with bases other than ten, such as binary and
hexadecimal numerals. To accommodate the notation independence of the natural
numbers, we need an account of acceptable numerals and a condition for two
numerals to denote the same number.
I take the numerals to include any concrete objects that stand in a suitable ordering.
On this very liberal view, a numeral need not even be a syntactic object in any
traditional sense. For example, if a prehistoric shepherd counts his sheep by matching
them with cuts in a stick, then these cuts count as numerals. The purpose of a numeral
is just to mark a place in an ordering. But of course, one and the same object can
inhabit different positions in different orderings. The syntactic string '111' can for
example mean either 3 or 111 depending on the ordering in which it is placed. We
should therefore make explicit the ordering in which a numeral is placed. So I shall
take a numeral to be an ordered pair (u, R}, where u is the numeral proper and R is
some ordering in which u occupies a position.
How should the ordering R be understood? One question is whether R should
be understood in an extensional manner (say, as an ordered set or a list) or in an
intensional manner (say, as an algorithm or procedure). I see two reasons to favor the
latter alternative. First, in order to be learnable and effectively useable, the ordering
must be computable.25 This means that the ordering must be understood as some sort
of procedure rather than as an infinite collection of ordered pairs. Second, there are
only finitely many numeral tokens of any learnable and effectively useable numeral
system. This means that the ordering must be defined not only on actual numerals
but in a way that extends to possible further numerals as well. This is best accounted
for by considering orderings in an intensional sense.
23 This claim is developed and defended in (Linnebo, 20 l 6a).

24
For a broadly similar account, see (Parsons, 1971, section 3).
25 See (Shapiro, 2017).
Another question concerns the order type of the relation R. A minimal requirement
is that R be a discrete linear ordering with an initial object.26 For we want it to be
the case that no matter how far we have counted, there is a unique next numeral-
provided there are further numerals at all. The question is whether we have any reason
to go beyond this minimal requirement.
We could, for example, add the requirement that R be of order type w. 27 I don't
think that would be a good idea. The notion of being of order type w is conceptually
quite demanding. This notion was not clearly grasped until fairly late in the history of
mathematics, at the earliest in the late sixteenth century when mathematical induction
was first explicitly articulated as one of the core principles of arithmetic. In fact,
the problems that mathematical neophytes often have with mathematical induction
suggest that even an implicit grasp of the notion goes beyond what is required for
basic arithmetical competence. This makes it doubtful that the notion should play a
central role in people's most basic conception of the natural numbers.
What about the weaker requirement that R be without an endpoint? This require-
ment too seems undesirable. For it seems plausible to allow numeral systems based
on finite orderings to denote natural numbers. (This may actually be the case on one
construal of the Roman numerals.) Finally, what about the requirements that R be
well-founded or that its order type not exceed w? These requirements too seem too
demanding to be implicit in basic arithmetical competence. The fact that all traditional
numeral systems are well-founded and of order type at most w is better and more
simply explained by observing that these are the kinds of orderings with which we are
familiar from the recursive formation rules of natural language.
To sum up, my view is that no requirement should be imposed on the order type
of R other than the minimal one that R be a discrete linear ordering with an initial
object. On this view, the natural numbers are clearly distinguished from non-standard
numbers only when mathematical induction is introduced as a basic arithmetical
principle.
I turn now to the equivalence relation that must hold between two numerals (u, R)
and (u' , R') for them to specify the same number. This equivalence relation must
clearly be a matter of the two objects u and u' occupying analogous positions in their
respective orderings. More formally, ( u, R) and (u', R' ) are equivalent just in case there
is a relation C which is an order-preserving correlation of initial segments of Rand R'
such that C(u, u'). I write (u, R) ""' (u', R') to symbolize that the two ordered pairs are
equivalent in this sense. 28
26
An order R is linear iff R is reflexive, transitive, and total (in the sense that, for any x and y. either Rxy
or Ryx). An order R is discrete iff, for any x that bears R to some distinct object, there is an R-least object
distinct from x to which x bears R. (That is, whenever Rxy for some y i x, there is some z i x such that
Rxz and ~3w(Rxw /\ Rwz /\ w i x /\ w i z).)
27
Mathematical structuralists are often attracted to this requirement. See the works cited in note 2.
28 Note that ~ is guaranteed to be one-to-one only on finite numerals. This means that the "numbers"
corresponding to any infinite numerals cannot always be identified with ordinal numbers. These "num-
bers" are unintended and pathological objects that are ruled out as soon as one's conception of the natural
numbers is sophisticated enough to include the principle of induction, as described below.
JUSTIFYING THE AXIOMS OF ARITHMETIC 185
A slight refinement of our previous two-level criterion of identity for natural

numbers can now be expressed as follows:
(CI-N') N(u,R} = N(u 1,R1} ++ (u,R},...., (u 1,R1 }
where the variables u and u' range over concrete objects, and R and R' range over
relations of the sort characterized above. Let 'NuM(x)' be a predicate that holds of all
and only the objects that can be specified in this way.
At this point it is appropriate to pause and ask whether the proposed criterion of
identity is even consistent. The criterion seems vulnerable to a version of the Burali-
Forti paradox. 29 To see this, let R be a well-ordering and u an object in its field. Then
the number assigned to the numeral ( u, R} serves as an order type of the well-ordering
that results from restricting R to u and its R-predecessors. The road now appears open
to a version of the familiar Burali-Forti reasoning. 30
Thankfully, the danger can be averted. One option is already implicit in our gloss
on (CI-N'). The variables u and u' have been assumed to range over concrete objects,
whereas the numbers that are the value of our functor N are meant to be abstract. This
blocks the derivation of the Burali-Forti-style paradox. 31 This solution seems to me
largely satisfactory. The requirement that the numerals proper be concrete objects is
not an ad hoc restriction introduced merely to avoid paradox but is independently
motivated by a desire for a system of numerals that we can actually use, say in
computations. To ensure practical usability, the numerals must be epistemically and
semantically more accessible than the numbers themselves.
Even so, this initial response cannot be the end of the matter. The question
inevitably arises whether it is possible to lift the requirement that the numerals proper
be concrete. An affirmative answer can be given, using ideas from Chapter 3, where
we developed a modal approach to abstraction understood as a "dynamic" process.
This theory makes available a notion of definiteness such that adding the plausible
requirement that numerals be tied to definite positions in linear orderings protects
(CI-N') against Burali-Forti and ensures its consistency. 32
10.6 Justifying the Axioms of Arithmetic

I now show how this conception of the natural numbers, supplemented with some nat-
ural and very simple definitions, allows us to justify all the axioms of Dedekind-Peano
Arithmetic (PA), adapted to let the first natural number be 1 rather than 0. I proceed
in three stages. First, I introduce three relations on numerals which correspond to
29
Thanks to Stewart Shapiro for questions forcing me to clarify this threat.
30 To see this, consider all the numbers that can be specified by means of a numeral of the form (u, R),
where Risa well-ordering. These numbers stand in a natural ordering, which we denote :::J. Let::::!* be the
modified ordering obtained by placing the initial element a of::::! at the very end. Consider the number x
specified by (a, ::::!* ). It is easy to see that x cannot be in the field of the ordering :::J. But this contradicts our
assumption that ::::! is an ordering of all the numbers.
31 More precisely, we can deny that there is a number that is specified by (a,::::!*).
32 See Chapter 12, note 7.
the basic arithmetical relations of succession, addition, and multiplication. Then,

I show that these relations on numerals induce corresponding relations on the natural
numbers, thus allowing us to define the non-logical primitives of the language of PA.
Finally, I show how we can justify the axioms of PA on this basis.
We can define a successor relations' on numerals by lettings' ((u, R), (u', R')) just
in case u' is the R' -successor of some v such that (u, R) ,.._, (v, R'). We now wish
to define relations on numerals A# and M# which correspond to the arithmetical
relations of addition and multiplication on natural numbers. Let I# (x) formalize the
claim that the numeral x is first in its ordering. Following the practice of ordinary
counting, we think of such initial numerals as representing the number 1. Then the
two desired relations are partially defined by the following recursion equations (whose
free variables are implicitly understood as ranging over numerals):
(Al) I#(y) ~ (A#(x,y,z) ++ s'l(x,z))
(A2) s'l(y,y') /\A#(x,y,z) ~ (A#(x,y',z') ++ S#(z,z'))
(Ml) fl(y) ~ (M#(x,y,z) ++ x ,.._, z)
(M2) s'l(y,y') /\ M(x,y,z) ~ (M'(x,y',z') ++ A#(z,x,z'))
We can show by induction that these equations uniquely define the relations on
all numerals that are standard (in the sense of being among the first w numerals
in their respective orderings). Since we do not care how the relations behave on
any non-standard numerals, it is unproblematic to add the further requirement that
the relations S#, A#, and M# not distinguish between numerals that are equivalent
under"'·
This congruence requirement entitles us to use the relations to define correspond-
ing relations on the natural numbers themselves:
(Def-S) S(N(x),N(y)) ++ s'l(x,y)
(Def-A) A(N(x),N(y),N(z)) ++ A#(x,y,z)
(Def-M) M(N(x),N(y),N(z)) ++ M(x,y,z)
Following the practice of ordinary counting, we let I be the first number. For example,
1 may be specified as N (' l ', D), where D is the standard ordering of the decimal
numerals.
We can now verify that the various axioms of PA hold (with 'Num' defined as on
p. 185). 33
(PAI) NuM(l)
(PA2) NuM(x) ~ -.S(x, 1)
(PA3) S(x,y) /\ S(x' ,y) ~ x = x'
(PA4) S(x,y) /\ S(x,y') ~ y = y'
33 (PAl) is trivial. To prove (PA2), let x be any number. Then there is a numeral (u, R) such that
x = N( (u, R) ). If S(x, 1), then u would precede the initial object of the relevant ordering, which is impossible.
To prove (PA3), assume S' (u, v), s' (u' , v'), and v ~ v'. We need to show u ~ u'. This follows from our
assumptions and the definition of ~ . (PA4) can be proved in an analogous manner. (PAS) through (PAS)
follow immediately from (Al), (A2), (Ml), and (M2).
JUSTIFYING THE AXIOMS OF ARITHMETIC 187
(PAS) A(x, l, z) # S(x, z)

(PA6) S(y,y') /\ A(x,y, z) --+ (A(x,y', z' ) # S(z, z' ))
(PA7) M(x, 1, z) # x = z
(PAS) S(y,y') /\ M(x,y,z)--+ (M(x,y',z') # A(z,x,z'))
Consider next the axiom of induction. To justify this axiom, we need to specify
some condition of finitude with which to restrict the numbers such that we get all
and only the natural numbers. The needed condition is simply that mathematical
induction be valid of the natural numbers. That is, ~ is a natural number (in symbols:
N(x)) just in case the following open-ended schema holds:
cp(l) /\ VuVv(cp(u) /\ S(u, v) --+ cp(v)) --+ cp(x)
This allows us to justify the axiom scheme of induction:

(PA9) cp(l) /\ VuVv(rp(u) /\ S(u, v)--+ rp(v))--+ Vx(N(x) ~ cp(x))
Prior to the adoption of this scheme, it is doubtful that people had uniquely singled
out the natural number structure as opposed to the objects of some non-standard
model of arithmetic.
I have saved for last the Successor Axiom, namely
(PAlO) Vx(NuM(x)--+ 3yS(x,y))
whose justification is less obvious. It would of course suffice to assume that every
numeral bears s' to some further numeral. But this assumption can be resisted. There
are only finitely many numeral tokens-in the ordinary sense of "numeral token". And
even with our liberal use of "numeral", which includes any concrete object, we don't
want the truth of arithmetic to be hostage to the existence of infinitely many concrete
objects.
Thankfully, we can do better. How does a child learn that every number has a
successor? The standard route to this knowledge, I think, is based on coming to
appreciate that for every numeral, we can produce a successor numeral. (That is
certainly how I taught this arithmetical truth to my children!) Consider therefore the
modal claim that necessarily, for any numeral there could be some further numeral
to which it bears s':
(10.1) DV(u,R)03(u 1,R1 ) Sl((u,R,), (u 1,R1 ))
This claim is extremely plausible. Assume we are given a numeral (u, R). Then it is
possible that there is a concrete object u' distinct from u and all ofits R-predecessors.34
Let R' be the result of appending the pair (u, u' ) to the initial segment of Rending
with u. Then (u', R' ) is as desired.
34
If the numerals are linguistic types (as opposed to the generalized numerals that I permit), then this
possibility claim may well be supported by an appeal to mathematical intuition. For example, (Parsons,
2008, section 29) appeals to mathematical intuition to defend the possibility of extending a sequence of
"Hilbert strokes". What ultimately matters, however, is that the modal claim be defended in some way or
other.
Since the definitions of the predicates NuM and S hold by necessity, ( 10.1) enables
us to prove:
(10.2) D'v'x(NuM(x) --+ 0 3yS(x,y))
We notice that where the Successor Axiom says that every number actually has a
successor, (10.2) says that every number potentially has a successor. So we are close
to our target, but not quite there. One way to advance from (10.2) to the Successor
Axiom is by making the further assumption that a number exists necessarily if at all:
(DE) D'v'x(NuM(x) --+ D3y(y = x))
Provided that our modal logic contains the so-called "Brouwerian Axiom'; namely
cp --+ DOcp, some easy modal logic shows that (10.2) and (DE) entail the Successor
Axiom.
How convincing is this argument? Since the modal operators are here understood
in terms of metaphysical modality, the Brouwerian Axiom is very plausible. (This
modality must be distinguished from the interpretational modality invoked in Chap-
ter 3 and on which the Brouwerian Axiom was observed to fail.) The same goes for
the assumption (DE) of necessary existence, which is very widely shared-and a mere
corollary of the thesis (to be defended in Section 11.1) that a pure abstract object exists
necessarily if at all. Thus, I submit that the argument is sound.
11
The Question of Platonism
11.1 Platonism in Mathematics

This book defends the existence of abstract mathematical objects. Should this be
regarded as a defense of mathematical Platonism? The answer obviously depends
on what is meant by the term. I propose to analyze mathematical Platonism as the
conjunction of the following two theses. 1
Object realism
There exist abstract mathematical objects.
Independence
Mathematical objects, if any, exist and have their properties independently of intelligent agents
and their language, thought, and practices.
The first thesis, which is tolerably clear, was defended in Chapters 2 and 8. The second
thesis is far less clear. So it is not obvious how my view bears on this thesis. The
problem is to understand the crucial notion of independence.
The most straightforward understanding is as a form of counterfactual indepen-
dence. Had there been no intelligent agents, there would still have been mathematical
objects, related to one another just as they in fact are. Is this counterfactual true?
I believe the answer is a resounding "yes''. Pure mathematical objects exist necessarily,
if at all, and their relations to one another obtain by necessity. It follows that, however
things had been, pure mathematical objects would have existed and been related to
one another just as they in fact are. Indeed, the universe of pure mathematical objects
would have remained the same even in circumstances far more remote than the
nearest ones in which there is no intelligent life.
It may not be entirely clear how my view upholds this counterfactual independence
of pure mathematics. I have emphasized the close connection between the existence
of mathematical objects and certain legitimate extensions of our language-and
these extensions are no doubt contingent processes. The key is to recall that, when
evaluating counterfactuals, we always hold fixed the present interpretation of our
language and use the language, thus interpreted, to describe how things would have
1
See (Linnebo, 2017c). What I am here calling "object realism" is the conjunction of two separate theses
in that work.
been had the circumstances been different. Keeping this in mind, we ask what is
required for the existence of different sorts of objects at a possible world. As we have
seen, the existence of physical bodies gives rise to a requirement on a particular region
of spacetime, namely that this region contain a suitably continuous and cohesive
parcel of matter. Since satisfaction of this requirement is contingent, so is the exis-
tence of the physical bodies. Next, consider the natural numbers. I argued that
there is no requirement on any particular region of spacetime; numerals witnessing
the existence of the numbers could exist anywhere. 2 In particular, it suffices that suit-
able numerals be available in the actual language that is used to describe the possible
world in question. This ensures that the existence of any natural number is metaphys-
ically necessary. Analogous arguments can be provided for other pure mathematical
objects obtained by abstraction. The counterfactual independence of pure mathemat-
ical objects from intelligent agents and their doings would no doubt have pleased
Frege.
It is doubtful, however, that this counterfactual independence suffices for my view
to qualify as mathematical Platonism. The independence thesis is often fleshed out
by comparing mathematical objects with physical objects. The two kinds of object are
said to be metaphysically on a par. A central part of Platonism is thus an analogy of
mathematical objects with physical ones. A typical example is (Maddy, 1990), who
characterizes realism or Platonism as the view that
mathematics is the scientific study of objectively existing mathematical entities just as physics
is the study of physical entities. The statements of mathematics are true or false depending on
the properties of those entities, independent of our ability, or lack thereof, to determine which. 3
On the view defended in this book, the analogy between mathematical and physical
objects is not entirely appropriate. It is true that mathematical objects enjoy the same
counterfactual independence from us and our activities as do mountains, rivers, and
trees. But mathematical objects are strikingly different from physical objects in other
respects.
One of these differences has already been extensively discussed, namely that math-
ematical objects are thin, while physical objects are thick. But the thinness of math-
ematical objects has an interesting consequence which requires further discussion,
namely that thin objects give rise to the phenomenon of indefinite extensibility. This
is the topic of the next section. The remaining sections identify a second difference,
which is then put to constructive philosophical use, namely that thin and thick objects
contribute in very different ways to the truth or falsity of predications that are made
of them. Together, these differences erode the analogy between mathematical and
2 See Section 2. 7.
3 Seep. 21. Compare also Crispin Wright's claim, in his pioneering defense of neo-Fregeanism, that
abstraction principles give access to objects "every bit as objective as mountains, rivers and trees" (Wright,
1983, p. 13). See (Linnebo, 201 ?c) for yet further examples.
THIN OBJECTS AND INDEFINITE EXTENSIBILITY 191
physical objects and thus serve to distinguish the object realism defended in this book
from more robust forms of Platonism.
11.2 Thin Objects and Indefinite Extensibility

Suppose we have formulated a perfectly precise definition of the notion of a star. For
every object whatsoever, our definition yields a determinate verdict, either that the
object is a star or that it is not. Given such a definition, there are some objects that are
all and only the stars. When our precise concept is applied to reality, in other words,
reality answers with a determinate extension, namely all the objects that fall under the
concept. Reality and the determinacy of the concept jointly ensure the determinacy
of the extension. And there is nothing unusual about stars in this respect. For most,
if not all, concepts of physical objects, the determinacy of the concept ensures the
determinacy of its extension, given the usual input from reality.
Mathematical concepts differ from physical ones in this regard, or so I have argued.
Consider the concept of a set. On the account developed here, a set is anything
that can be obtained by abstraction on pluralities under the equivalence relation of
coextensionality, that is, anything of the form {xx}, where xx are some objects. We may
also suppose that all Caesar-style questions have been answered. 4 The details of this
conceptual determination are unimportant for our present purposes. What matters is
that we have, or can obtain, a perfectly precise concept of a set. Even so, I claim, there is
not a determinate extension of all and only the sets. Given any plurality of objects that
may serve as the desired extension, we can use this plurality to define yet another set,
thus proving that the plurality did not, after all, include all sets. In short, the concept
of a set is indefinitely extensible: for any plurality of instances, we can use this plurality
to define a new instance of the concept, that is, an instance that is not a member of
the given plurality. And there is nothing unusual about sets in this regard. A variety
of other mathematical concepts too are subject to indefinite extensibility.
On my view, there is thus an important difference between physical objects and
mathematical ones. When we consider physical objects, reality and the determinacy of
a concept jointly ensure the determinacy of the extension, but not so when we consider
mathematical objects. So there is a disanalogy between the two kinds of object. This
is a respect in which my object realism differs from a more robust form of Platonism.
Of course, a robust Platonist may object that, if my view entails this disanalogy
between physical and mathematical objects, then so much the worse for my view.
My full response is this entire book. Instead of summarizing my arguments, I shall
here content myself with a simple observation. What ultimately drives my view is
a liberal conception of what counts as a permissible definition. In particular, given
some objects xx, we can make good mathematical and philosophical sense of the
4
For the past two claims, see Chapters 3 and 9, respectively.
192 THE QUESTION OF PLATONISM
associated set {xx}; so this is a permissible definition. If this claim is granted, then
the indefinite extensibility of the concept of a set follows straightforwardly, and thus
also the mentioned disanalogy and step away from robust Platonism. Since these
conditional claims should be uncontroversial, we face a choice: either to accept the
mentioned disanalogy between physical and mathematical objects, or else to deny the
permissibility of what for all intents and purposes seems to be a permissible definition.
My choice of the latter option is dear, for two reasons. First, my theory of thin
objects shores up the permissibility of the mentioned definition and a vast family
of related ones. 5 Second, even apart from any theory of thin or lightweight objects,
it is good methodology to attach greater weight to the broadly mathematical con-
siderations about permissible definitions than to the metaphysical intuitions that are
typically used to support the Platonistic analogy. A liberal conception of permissible
mathematical definitions underlies much of contemporary mathematics. It is hard
to improve on Cantor's expression of this conception in a passage quoted already in
Chapter 1:
Mathematics is in its development entirely free and only bound in the self-evident respect that
its concepts must both be consistent with each other and also stand in exact relationships,
ordered by definitions, to those concepts which have previously been introduced and are
already at hand and established. 6
In short, to safeguard the freedom of mathematics, the analogy on which Platonism

is based needs to be curtailed.
11.3 Shallow Nature

To introduce our next distinction, let us begin with two simple examples. First,
consider the question of whether a physical body has the property of being a solid
sphere. This question cannot be answered on the basis of any one of the body's proper
parts; rather, information is needed about the entire body. And there is nothing
unusual about this property. Whether a body has some property typically depends
on all, or at least many, of its parts. A body has most of its properties in a way that is
not reflected in the possession of some property by any one of its proper parts. The
physical body thus plays an ineliminable role in the explanation of the truth-values of
predications that are made of it: the truth of such predications cannot be reduced to,
or grounded in, a truth that is solely about any individual specification of the body.
Letter types are very different in this regard. To see this, consider the question
of whether a letter type a has the property of being a vowel. Unlike the case of a
physical body's being a sphere, any token by means of which a can be specified suffices
to answer the question, namely, in accordance with whether or not the token is a
vowel. There is no need to examine other specifications of a, let alone "the type itself",
5 6
See especially Section 3.3. See Cantor 1883 (Ewald, 1996, p. 896).
SHALLOW NATURE 193
unmediated by any specification (whatever that might mean). The question of whether
a letter type a has some graphological property can thus be reduced to a question about
any token by means of which a is specified.
Indulging in some slightly loose talk, we may put our observations as follows.
A predication made of a physical body is typically true or false only because of the way
the body is. To account for the truth or falsity of such predications, we need to proceed
via the subject of the predication, if not explicitly, then at least implicitly (as we do for
instance when we talk about some parts "arranged spherewise"). By contrast, the truth
or falsity of any graphological predication that is made of a letter type is fully explicable
in terms of properties had by any specification of the type, thus making it unnecessary
to consider the proper subject of the predication. Or, in grounding-theoretic terms,
the graphological predication made of a is grounded in a corresponding predication
made of any specification of a. In this sense, a thin object "inherits" its most important
properties from each of its specifications in a way that thick objects do not.
Our discussion has so far focused on two simple examples. Can some more general
lessons be extracted? In particular, does the property "inheritance" phenomenon that
we have found to be present in letter types, but not in physical bodies, extend to
other kinds of abstract object as well? To address these questions, we need some
terminology.
As usual, let us say that a property is intrinsic just in case an object's having the
property is entirely a matter of how the object is and does not depend on how things
stand with the rest of the universe. While this definition is rough and perhaps not
entirely satisfactory, it suffices for present purposes.7 Next, observe that the notion
of intrinsicness generalizes from properties to polyadic relations. Let us say that a
relation R is intrinsic if its holding or not depends entirely on the relata, and extrinsic
if its holding or not also depends on matters external to the relata. For instance, "x has
greater mass than y" is intrinsic (at least in classical physics), whereas "xis less than
two meters from y" is not. 8
Our next task is to spell out the difference we observed between physical bodies and
letter types. Let F be a sorta! that is subject to some (conditional) two-level criterion
of identity:
Consider an intrinsic relation R on Fs. Let us say that R is reducible on F just in case
there is an intrinsic relation R# on the relevant class of specifications such that:
{11.1)
7
See (Weatherson and Marshall, 2012) for discussion.
8 It is important not to confuse this notion of an intrinsic relation with that of an internal relation, that
is, a relation whose obtaining is essential to the relata. For instance, the mass example just mentioned shows
a relation that is intrinsic but not internal.
When R is reducible on Fs, the question of whether R holds of some Fs can thus be
answered by determining whether R# holds of some specifications of these Fs. Finally,
let us say that Fs have a shallow nature just in case all intrinsic relations on Fs are
reducible on F. 9 This captures the intuitive idea that any question that is solely about
Fs has an answer that can be determined on the basis of any given specifications of
these Fs.
It is important to understand why our definitions require the relation R# to be
intrinsic. Suppose this requirement was lifted. Then the incorrect claim would follow
that every relation is reducible. For example, suppose R is the property of being
spherical. Let R# be the property of being part of a body that is spherical. Then ( 11.1)
holds. Yet it would be wrong to say that predications of being spherical reduce to
predications of being part of a body that is spherical! The problem is that R# is not
intrinsic. Consider the difference we discussed between a body being a sphere and
a letter type being a vowel. The property of the letter type, but not that of the body,
is reflected in an intrinsic property of any specification of the object in question. We
therefore add the requirement that the relation R# be intrinsic.
Equipped with these definitions, let us return to the question of how to generalize
from the two examples with which we began. What sorts of object have a shallow
nature? I propose a simple answer. 10
Let F be a sorta! based on a two-level criterion of identity. Assume its unity relation ~ is
intrinsic. Then Fs have a shallow nature. 11
To defend this answer, assume that ,...., is intrinsic. Let R be any intrinsic relation on
Fs, and let the relation R# be defined as in (11.1). I need to show that R# is intrinsic.
My argument is based on two observations.
First, observe that a specification u and an intrinsic unity relation ,...., with u in its
field suffice, all by themselves, to determine a referent. For example, a line and the
unity relation of parallelism suffice, all by themselves, to determine a direction. By
contrast, a parcel of matter and the unity relation for bodies do not suffice to determine
a body. The parcel is capable of belonging to many different bodies. So we cannot
determine to which body, if any, the parcel of matter belongs merely by considering
the parcel itself. We need to look beyond the parcel at the surrounding reality. And
we can generalize. When a unity relation is intrinsic, a specification u suffices to
determine the object §u of which u is a specification, with no further input from reality
required. Second, observe that since R is intrinsic, the objects §u1, ... , §un suffice
9 See (Hale and Wright, 2007) for some closely related ideas. This phenomenon also calls to mind
Wright's apt metaphor of the numbers as mere "shadows of syntax" (Wright, 1992, pp. 181-2) as well as
Stephen Yablo's idea of"preconceived objects" (Yablo, 2010, introduction).
10
Modal properties are set aside; cf. Chapter 8, note 11.
11 Assume, as seems plausible, that identity counts as an intrinsic relation. Then the converse holds as
well. To see this, apply the reducibility of predication to a predication of identity, from which it follows that
the unity relation is intrinsic.
THE SIGNIFICANCE OF SHALLOW NATURE 195
to determine whether R(§u1> ... , §un). Putting these two observations together and
invoking (l l.l), it follows that the specifications U1> ... , Un suffice to determine
whether R#(u 1, ... , Un). This establishes the desired conclusion that R# is intrinsic
and thus also concludes my defense of (what I called) the simple answer. 12
Let us take stock. I observed that letter types have a shallow nature. Any predication
of letter types can be reduced to (or grounded in) a predication made of any letter
tokens in terms of which the types are specified. This prompted the question of
whether other mathematical objects too have a shallow nature. To get a handle on
the question, I set out to analyze the reducibility of predication. My analysis revealed
the importance of intrinsic properties and relations of the objects in question. The
central question is whether some objects' standing in an intrinsic relation can be
reduced to (or grounded in) a matter of some related intrinsic relation's holding of any
specifications of the objects. On the analysis I have defended, the answer is affirmative
whenever the objects in question are associated with a unity relation that is intrinsic.
This analysis naturally prompts a follow-up question: What sorts of object have
intrinsic unity relations? We have seen various examples of mathematical objects that
do. The same goes for other sorts of mathematical objects too, I believe. But I do not
know how to establish this other than in a case-by-case manner.
11.4 The Significance of Shallow Nature

Suppose I am right that abstract objects have a shallow nature. What would be the
philosophical significance of this discovery?
Let me begin by distinguishing the phenomenon of shallow nature from some
other ideas with which it must not be conflated. First, is the shallow nature of some
objects the same as these objects' being non-fundamental, that is, that their existence
and properties are grounded in other facts? 13 When an object is non-fundamental,
it is possible to explain the truth or falsity of any predication of it without involving
the object in question. A ball's being spherical, say, can be explained in terms of the
arrangement of its parts, with no need to involve the ball itself. However, the claim that
an object has a shallow nature is far stronger than the claim that it is non-fundamental.
Only the former claim ensures that any question of whether the object has some
intrinsic property can be reduced to a question of whether any given specification of
the object has some related intrinsic property.
12 The natural numbers, as understood in Chapter 10, require a slight refinement of our analysis. To
determine whether two numerals are equivalent in the sense of occupying matching positions in their
respective numeral sequences, it is not sufficient to consider the two numerals in isolation: we additionally
need to consider all predecessors of each of the two numerals. However, the unity relation is still intrinsic
to an effectively specified and restricted range of numerals, which function as specifications of numbers. So
any predication of an intrinsic property or relation to some natural numbers can be reduced to an associated
predication of an effectively specified and restricted range of nuplerals.
13 Thanks to Michele Lubrano for emphasizing the importance of this distinction.
Second, the shallow nature of some objects must be distinguished from the question
of whether singular terms that purport to stand for such objects are "semantically
idle" in Dummett's sense that these terms have no semantic value that figures in a
compositional semantic theory. 14 On the view I have defended, there are cases of
genuine reference to abstract objects. A singular term can bear the same semantic
relation to an abstract object as such terms bear to concrete objects. Reference to
abstract objects differs from reference to concrete ones, not by being "semantically
idle", but with regard to how the relation of reference is constituted. Criteria of identity
can contribute to the constitution of reference to concrete and abstract objects alike.
Only in the latter case, however, is the unity relation intrinsic. And it is this that gives
rise to objects with a shallow nature.
Third, there is the question of whether it is possible to "confront" objects of a certain
sort. Dummett repeatedly denies there can be a "confrontation" with an abstract object
in anything like the way we confront physical objects. 15 This is wrong, I believe, or at
best misleading. I have identified some deep structural similarities between reference
to the concrete and to the abstract. Both types of objects can be specified by means of
certain parts or aspects, which are subject to a unity relation that determines when two
specifications determine the same object. Each type of object is then "confronted" only
via its specifications. It is true that reference to abstract objects differs in important
ways from reference to concrete objects. But it is not helpful to explain this difference
as a form of "confrontation" that is possible only in cases of reference to concrete
objects. The difference is better understood in terms of the thinness of abstract objects
and their shallow nature, as explained in Sections 2.8 and 11.3, respectively.
I turn now to the significance of the shallow nature of mathematical objects
for some questions of realism. According to truth-value realism, every well-formed
mathematical statement has a unique and objective truth-value that is independent
of whether it can be known by us and whether it follows logically from our current
mathematical theories. Is there any connection between object realism, as defined
above, and truth-value realism? The former is often thought to support the latter.
So long as there are objects for certain statements to be about, the thought goes,
these objects' properties ensure that each statement about the objects is true or
false. For example, since stars and planets really exist, there is a fact of the matter
about the existence of extraterrestrial life, irrespective of our ability to find out. The
shallow nature of mathematical objects shows that this thought cannot be applied to
mathematics. To explain why mathematical objects have their properties, there is no
need to invoke these objects; rather, an explanation can be based on any specifications
of the objects and certain properties of these specifications.
This reversal of the customary explanatory direction is unsurprising, given that my
view is inspired by Frege's context principle. As discussed in Chapter 7, this principle
14
See (Dummett, 199la, pp. 193 and 195), as well as Section 5.3 for discussion.
15 See (Dummett, J98la, ch. 14) and (Dummett, 199la, p. 204).
HOW BELIEFS ARE RESPONSIVE TO THEIR TRUTH 197
regards questions about whole sentences as explanatorily prior to questions about

singular terms.
Next, there is working realism, which asserts the legitimacy of certain methods that
are distinctive of classical mathematics, such as the use of classical logic, impredica-
tive definitions, and reasoning about arbitrary subcollections of infinite domains. 16
A Platonistic conception of mathematics is generally thought to support working
realism. Since it appears legitimate to reason classically about the physical universe,
the Platonistic analogy between the mathematical and the physical suggests that the
same goes for the mathematical universe. It is doubtful, however, that the object
realism I have defended can sustain an argument of this form. For I deny the
analogy on which Platonism is based. Thus, to the extent that the methods of classical
mathematics require a defense, this has to be provided in some other way.
The upshot is that my object realism, which asserts the existence of abstract
mathematical objects but abstains from the Platonist's assimilation of these to concrete
objects, has no direct payoffs concerning mathematical objectivity or permissible
methods. This is a partial vindication of Kreisel's dictum, which makes many appear-
ances in the writings of Michael Dummett, for instance:
As Kreisel remarked in a review of Wittgenstein, "the problem is not the existence of mathemat-
ical objects but the objectivity of mathematical statements". (Dummett, 1978, p. xxxviii) 17
11.5 How Beliefs are Responsive to Their Truth

While the shallow nature of mathematical objects limits their capacity to shore up
mathematical objectivity, this nature has important consequences for the epistemol-
ogy of mathematics. It ensures that mathematics is epistemically tractable, which is far
from obvious. The problem is well known, thanks in large part to (Benacerraf, 1973).
How is knowledge of abstract mathematical objects possible, given that there is no
causal interaction with them? The remainder of this chapter develops an answer.
The epistemological challenge posed by the abstract is easily misunderstood.
Simply to insist that all knowledge be based on a causal connection between the
knower and the known would be sheer philosophical prejudice. Arguably, Benacerraf
is guilty of such a prejudice. But there is no need to insist on a causal connection.
Our mathematical beliefs are generated by means of certain methods and procedures.
What is it about these methods and procedures that makes them conducive to
generating true beliefs? It is not just an accident that we arrive at beliefs that correctly
represent abstract objects and their properties. What is it, then, that explains this
reliability? Granted, we must not prejudge the form of the answer, as we would do if we
16 See (Shapiro, 1997, pp. 21 - 7 and 38-44), as well as (Bernays, 1935).

17 See also (Dummett, 198la, p. 508). The remark of Kreisel's to which Dummett is alluding appears to
be (Kreisel, 1958, p. 138, fn. 1) (which is rather less memorable than Dummett's paraphrase).
required a causal connection between the knower and the known. But it is perfectly
legitimate to insist on some answer or other. 18
To develop this non -prejudiced challenge, let us take a closer look at the reliability of
mathematical beliefs. Mathematicians are highly reliable about mathematical matters.
For most mathematical sentences S, the resulting instance of the schema
If mathematicians accept S, then S is true

is true. Let us call these instances reliability claims. It cannot be just by chance that
mathematicians are reliable in this way. Their reliability requires an explanation. Since
mathematical objects are abstract, however, it is unclear what an explanation could
look like. This contrasts with the reliability of other kinds of belief. Consider our per-
ceptual beliefs about everyday objects in our immediate environment, which are also
fairly reliable. This reliability is a not mystery, as these beliefs depend counterfactually
on the state of our environment. We also know a lot about the physical mechanisms
that underlie this dependence. The case of mathematical beliefs is very different. These
beliefs cannot depend counterfactually on the state of our mathematical environment.
To make sense of such a counterfactual dependence, we would have to make sense of
varying the mathematical facts . But since these facts are necessary, we have no idea
how to do this. 19
How, then, can our mathematical beliefs in any way be responsive to the truth
of these beliefs? The usual channel of influence appears unavailable! The answer, I
shall argue, is that there are two distinct ways in which our beliefs about some subject
matter can be influenced by the truth of these beliefs. So even where one of the ways is
unavailable, the other may still be operative.
Before fleshing out this (so far entirely programmatic) proposal, I need to clarify
my use of the crucial word 'belief'. I am concerned with belief tokens, in the sense
of internal psychological states, considered in abstraction from any propositional
content that these states may have. This is a syntactic, rather than semantic, notion of
belief. A belief, in this sense, is just a sentence tokened in "the belief box" (in Fodor's
famous phrase). Instead of 'belief: I might thus have used 'belief state'.
It is useful to approach my proposal via some examples. Suppose an agent directs
its attention at a near-perfect globe and forms the belief "this body is spherical''. The
truth of this belief depends on two factors: first, that the belief has some particular
proposition as its semantic content; and second, that the world is such as to make
this proposition true. Both factors are needed to explain the truth of the belief.
I shall now argue that both factors are also needed to explain the agent's formation
of the belief.
18 See (Linnebo, 2006a) for a fuller account of how the epistemic challenge to mathematical Platonism
can be formulated in a way that does not beg any philosophical questions.
19 Throughout this section, we indulge in talk about facts as an innocent shorthand which could, if
desired, be eliminated in favor of talk about truths or by use of suitable gerunds.

HOW BELIEFS ARE RESPONSIVE TO THEIR TRUTH 199
Let us begin with the second factor, as this is more familiar. The content of the belief
is the proposition that the globe is spherical. The globe's being spherical therefore
contributes in a direct and obvious way to making the belief true. Had the globe not
been spherical, the belief would not have been true. The globe's being spherical makes
a second kind of contribution as well, namely to the agent's formation of the belief
in question. Had this globe been seriously dented, say, the agent would have noticed
and therefore not formed the belief that it is spherical. So far, so familiar. It is entirely
unsurprising that the worldly facts that directly contribute to making a belief true
often also contribute to our formation of the belief.
Consider now the first and less familiar factor, namely that a belief has a certain
proposition as its semantic content. This factor too contributes to the truth of the
belief, albeit in an indirect (or metasemantic) way. Had the belief had a different
proposition as its content, the belief might not have been true. The crucial question
is now: Does this indirect contribution to a belief's truth also contribute to the agent's
formation ofthe belief? To answer the question, we need to know why the belief has the
particular content that it happens to have; in particular, why its simple constituents
have the particular semantic values that they have. I have developed an account of how
the expression 'this body' comes to refer to a particular body, in this case the globe.20
This account involves facts about the causal transmission of information from the
globe to the agent's sense organs and about this information's being put together in
a way that is sensitive to the natural spatiotemporal connectedness of the parcels of
matter from which the information derives. I claimed that the perceptual information
is put together in accordance with the following criterion of identity:
(CI-B) b(u) = b(v) ++ u,....., v
where b is a "body building" function that maps a parcel of matter to the body to which
it belongs, and where,....., is the unity relation described in Section 2.3. It is important
to recall, however, that I do not claim that we have any explicit or conscious grasp of
this criterion of identity.
The crucial question formulated above is whether all these facts, which contribute
indirectly (or metasemantically) to the truth of the belief, also contribute to the agent's
formation of the belief. I believe the answer is affirmative. Had the agent been in
perceptual contact with another body, or had she put together pieces of perceptual
information in accordance with some principle other than (CI-B), she might not have
formed the belief state in question.
Let me sum up the lessons resulting from our first example. Facts can contribute
to the truth of a belief in two different ways: indirectly (or metasemantically), by
ensuring that the belief has a certain proposition as its semantic content; and directly,
20 See Section 2.3. An analogous account can be given of how 'is spherical' comes to have its semantic
value. This account will crucially involve the fact that our subject takes this predicate to apply to all and
only spherical things.
by ensuring that the world is as this proposition says. 21 Furthermore, facts that
contribute in either of these ways to the truth of a belief can also contribute to the
agent's formation of the belief. This means that a complete account of why the agent
formed the belief in question must appeal to facts that contribute in each of the two
mentioned ways.
The second example is a very simple mathematical one, namely the arithmetical
belief that 3 directly succeeds 2. Here too there are facts in virtue of which the
constituents of the belief state have the semantic values that they happen to have. 22
First, the numerals '2' and '3' occupy the second and third positions of the standard
sequence of decimal numerals. Next, the agent takes two numerals to determine the
same number just in case they stand in the unity relation ,....., . Finally, the agent takes
the successor relation S to hold of two natural numbers just in case the associated
numerals stand in the relation SI:
(Def-S) S(N(x),N(y)) ~ s'(x,y)

All of these facts contribute indirectly (or metasemantically) to the truth of the belief.
Unlike our example of the globe, however, there are no facts whose contribution to the
truth of the belief is direct. This direct form of contribution has vanished completely. 23
Fortunately, the facts that contribute indirectly to the truth of the belief that 3 directly
succeeds 2 also suffice to explain why the agentformed the belief. Because the agent treats
the predicate 'directly succeeds' in accordance with (Def-S), she regards the belief that
3 directly succeeds 2 as true just in case the associated numerals stand in the relation
SI. And because she regards the associated numerals '3' and '2' as ordinary decimal
numerals, she deems that they stand in the relation SI. Consequently, she regards the
belief as true. Had the agent not treated the predicate 'directly succeeds' in accordance
with the definition (Def-S), or had she not taken the numerals '3' and '2' to be ordered
as decimal numerals, she might not have formed the belief.
Let us sum up the lessons we have drawn from our second example. What explains
the truth of the simple arithmetical belief that 3 directly succeeds 2? I argued that the
only facts that contribute to this explanation do so indirectly (or metasemantically),
not directly. I proceeded to argue that the facts that do contribute in this indirect way
21
My distinction between two different kinds of contribution to the truth of a belief is not a distinction
between two different kinds of facts: the metasemantic ones and the rest. No fact is inherently metasemantic
or not. The distinction is between two different ways that a fact can contribute to making a belief state truth,
namely by determining its content or by making the world be as this content says it is.
22
My example relies on the account developed in Chapter 10. But if desired, it would be straightforward
to modify the example to make it rely instead on the competing cardinal account.
23 Does this make the belief in question analytic? The answer depends on how one understands the
notion of analyticity. I deny that the belief is analytic in the traditional sense that anyone who grasps the
proposition can see it to be true by conceptual analysis alone. The metasemantic facts that I have discussed
need n ot be consciously accessible even to people with a perfect grasp of the content of the belief. Similarly,
the metasemantic facts involved in reference to physical bodies need not be consciously accessible even to
people who are fully competent with such reference. See Sections 1.6 and 8.4.3 for further discussion.
THE EPISTEMOLOGY OF MATHEMATICS 201
also suffice to explain why the agent formed the belief. This connection ensures that
the agent's belief is appropriately sensitive to the truth of the belief. The agent formed
the belief, not by accident, but because it is true. This answers the epistemological
challenge as it arises for this very simple arithmetical belief.
11.6 The Epistemology of Mathematics

Can my answer to this very simple instance of the epistemological challenge be gener-
alized? Since other arithmetical relations are definable from the successor relation, my
answer generalizes naturally to more complex arithmetical examples. But we would
like to generalize further.
Consider the mathematical truths established by some form of predicative abstrac-
tion, as developed in this book. As always, we start with a base language, assumed to
be in good philosophical standing, and consider an extended language that is used
precisely as if we abstract on some partial equivalence relation "". This means that the
extended language is used in accordance with a set of assertibility conditions, which
are expressed in the base language and make recourse only to the ontology accepted
in that language. We now come to the heart of the matter. The fact that the assertibility
conditions govern the speakers' use of the extended language has two different effects.
First, it is because the extended language is used in this way that its sentences have the
truth-conditions that they have (namely, the ones assigned by the non-reductionist
interpretation defended in Chapter 8). Second, it is because the language is used in
this way that speakers form those of their mathematical beliefs that are expressed in
this language. In short, the assertibility conditions make a twofold contribution: to the
determination of the semantic content of beliefs (and thus also to the truth of these
beliefs), and to the formation of the beliefs.
This twofold contribution is very important epistemologically. Consider a mathem-
atical truth 1{I expressed in the extended language. The assertibility conditions that
govern the language are responsible for 1{1's meaning what it does and thus also for
its being true. Additionally, when people competently believe 1{1, they are led to do so
by these very assertibility conditions. In this way, the double role of the assertibility
conditions ensures that competent mathematical beliefs are responsive to the truth of
these beliefs. It is because these beliefs are true that competent people believe them. We
can represent this answer to the epistemological challenge by the following diagram:
1{I is true people believe 1/1
constitution cause
assertibility conditions
As represented by the left arrow, the assertibility conditions are constitutively involved
in making 1/t true, namely, by making 1/t mean what it does. Simultaneously, as
represented by the right arrow, the agents' being guided by the assertibility conditions
causally contributes to putting 1/t in their "belief boxes''.
This epistemological account enables me to redeem a premise made earlier in the
book. As part of my account of thin objects, I defined a sufficiency operator ==>
and claimed that this operator satisfies the epistemological constraint articulated in
Section 1.6. The constraint states that rp ==> 1/t must ensure that the corresponding
material conditional rp -+ 1/t is knowable. Let us now verify that this constraint
is satisfied. As we recall, 'rp ==> 1/t' means that either there is a predicative language
extension where <p serves as a ground for 1/t or else there is a series of such extensions
leading from <p to 1/t, with each step serving as a ground for the next. Let us begin
with the former case. So suppose there is a predicative language extension where <p
serves as a ground for 1/t. This means that we can carry out the language extension in
question and obtain the material conditional <p -+ 1/t "for free", that is, as something
that is assertible outright, regardless of how things are. We now invoke the argument
of the preceding paragraphs to show that this conditional is appropriately responsive
to what makes it true. There is good reason, therefore, to think this belief qualifies
as knowledge. 24 What about the latter case, where 1/t is obtained through a series of
predicative language extensions? If the series is finite, there is no particular problem,
provided we allow for our epistemic abilities to be suitably idealized. Where the series
is infinite, however, we require an altogether different form of idealization, which
means that our justification for the resulting statements will be more indirect and
less conclusive.
The section began by asking to what extent my epistemological explanations can
be generalized. Let me now try to answer the question. Consider the most general
mathematical reliability claim:
For nearly all mathematical sentences S, if mathematicians accept S, then S is true.
This claim is far more general than we have the resources to explain. My aim has
only been to explain how abstraction contributes to mathematical knowledge. I am
not committed to all such knowledge resulting in this way. Let us therefore restrict
ourselves to mathematical beliefs arrived at by the form of abstraction that I defend.
Consider an instance of predicative abstraction that takes us from some base language
to an extended language .C. This gives rise to a reliability claim that is more restricted
but still very general:
(R) For nearly all sentences S of .C, if mathematicians accept S, then S is true.
24
Of course, to settle the matter definitively, we would need to explain what is meant by "knowledge'',
something I shall not attempt here.
THE EPISTEMOLOGY OF MATHEMATICS 203
It is useful to "factor" this reliability claim into two components: one concerned
with mathematicians' beliefs, and another, with the truth of these beliefs. As we have
seen, the abstraction yields a number of truths in the extended language "for free".
Let :E be the collection of such "free" truths. Then we can "factorize" as follows:
(RB) For nearly all sentences S of£, if mathematicians accept S, then Sis in :E .
(Rr) For all sentences S of£, if Sis in :E, then Sis true.
Our task is to explain how mathematicians' beliefs, as recorded by (RB), are appro-
priately responsive to the truth of these beliefs, as recorded by (Rr ). Once again, the
explanation is based on the assertibility conditions that guide the agents' use of the
language. The adherence to these assertibility conditions is constitutively involved in
the claim (Rr) by providing a metasemantic account of why the relevant sentences
come to mean what they mean and thus also of why they are true. Additionally, the
adherence to the assertibility conditions is causally responsible for the claim (RB).
The agents form their mathematical beliefs because they are operating with these
assertibility conditions. We can thus generalize our earlier explanation, as represented
by the following diagram:
constitution cause
assertibility conditions
My conclusion is that abstract objects do not pose any insuperable epistemological

problem. It is possible to have knowledge of abstract mathematical objects, provided
that these objects are regarded as thin.
Two final remarks are in order before closing. First, there is an interesting connec-
tion between the explanation just given and the truth predicate that is used in the
reliability claim. My explanation presupposes that this truth predicate is semantic,
not deflationary. In other words, the truth of a mathematical sentence S is regarded as
a matter of both what S says and whether reality is as it says. A deflationary truth
predicate, by contrast, treats 'S is true' as synonymous with S itself. Such a truth
predicate makes it hard to see how the assertibitity conditions can contribute to
explaining the truth of the relevant sentences. Suppose we are asked to explain the
truth of '2 + 2 = 4'. When a deflationary truth predicate is used, this is the very same
thing as to explain 2 + 2 = 4. But this latter statement calls for a purely arithmetical
explanation. Any appeal to assertibility conditions would be a red herring.
Second, I would like to remind readers that some of the hardest problems in the
epistemology of mathematics derive, not from its ontology of abstract objects, but
from the extreme extrapolations and idealizations on which contemporary mathemat-

ics is based, especially in the case of higher set theory. These problems are orthogonal
to the one that Benacerraf placed on our agenda and which has occupied us here.
While Benacerraf's problem is certainly important, it is unfortunate that the great
amount of attention it has received has come at the expense of other, equally important
problems in the epistemology of mathematics. 25
25
The problem posed by the mentioned extrapolations and idealizations receives at least some discus-
sion in Chapters 3 and 12.
12
Dynamic Set Theory
12.1 Introduction
The purpose of this chapter is to develop the dynamic abstractionist approach to set
theory canvassed in Sections 3.5 and 3.6. A brief review of these sections may therefore
be appropriate.
We work in a language of plurallogic with a single non-logical predicate 'SET( xx, y )'
for 'xx form the set y'. We are interested in plural Law V, which states that two
pluralities form the same set just in case they are coextensive. As explained, this law
can be "factorized" into separate criteria of existence and identity:
(COLLAPSE) Vuu 3x SET(uu, x)
(EXT) SET(uu, x) /\ SET( vv, y) ---+ (x = y # Vz(z -< uu # z -< vv))
Next, we added modal operators D and <> to describe how the interpretation of
the language can be shifted-and the domain expanded-as a result of abstraction.
Thus, the intended meaning of 'Dcp' is 'no matter how we abstract and thereby
shift the meaning of the language, cp'; <>cp is the obvious dual. As emphasized, this
is an interpretational modality, which must be distinguished from more familiar
circumstantial modalities such as natural or metaphysical necessity. We can now
express claims about iterated set abstraction. The potentialist translation cp 0 of a non-
modal formula cp is particularly important; this is the result of replacing each ordinary
quantifier with the corresponding modalized quantifiers. Applying this translation,
we obtain dynamic analogues of the two "factors" of plural Law V, namely
0
(CoLLAPsE ) DVxx <>3ySET(xx,y)
(EXT 0 ) SET(uu,x) /\ SET(vv,y)---+ (x = y # D'v'z(z-< uu # z-< vv))
Finally, we showed that, given suitable background assumptions, the modalized

quantifiers 'D'v'x' and '0 3x' behave logically precisely like ordinary first-order
quantifiers. The precise statement of this result is the Mirroring Theorem, which
tells us that inferential relations among some non-modal formulas <Pi. . . . , <Pn and
1/1 are "mirrored" in corresponding inferential relations among their potentialist
translations:
<Pl, . .. , <Pn f- 1/1 iff <P10 ' · · · > <Pn0 f- 0 1/1 0
206 DYNAMIC SET THEORY
where I- 0 is deducibility in the appropriate modal plural logic, given appropriate

background assumptions.
12.2 Choosing a Modal Logic

What is the right modal logic for the interpretational modalities? To answer the
question, let us return to the process of abstraction whereby ever larger domains of
mathematical objects come to be recognized. In this process, entities are introduced
successively through a series of stages. To introduce an entity, all we need to do is
produce some specification of it and provide an appropriate criterion of identity.
To introduce a set, say, it suffices to specify the objects that are to be its elements
and to point out that sets are individuated by the law of extensionality.
Each stage of the process of abstraction can be regarded as a possible world. The
ontology of each possible world consists of the objects and concepts that have been
introduced thus far. The possible world also specifies how these entities are related and
in this way settles all questions about which entities at this world satisfy the various
atomic predicates.
Next, we define an accessibility relation Son the possible worlds. Let w s w' mean
that we can get from w to w' by some permissible introduction of more entities. This
motivates the following principle.
Partial ordering: The accessibility relation s is a partial order.
(Recall that a partial order is a reflexive, anti-symmetric, and transitive relation.) We

also require that the process ofabstraction be well-founded, which yields the following
principle.
Well-Foundedness: The accessibility relations is well-founded.
What if we have a choice about what entities to go on and introduce? Assume we are
at a world wo where we can go on to introduce entities so as to arrive at either w1 or w2.
Now, the license to introduce an entity never goes away as we build up the ontology
but can always be exercised at a later stage. This corresponds to the requirement that
any two worlds w1 and w2 that extend a single world can be extended to a common
world w3. This property of a partial order is called convergence and formalized
as follows:
We therefore adopt the following principle.
Convergence: The accessibility relation S is convergent.
This principle ensures that, whenever we have a choice about which entities to
introduce, the order in which we choose to proceed is irrelevant. Whichever entity
CHOOSING A MODAL LOGIC 207
we choose to introduce first, the others can always be introduced later. Unless :'.S was
convergent, our choice about whether to extend the ontology of wo to that of w1 or
that of w2 would have an enduring effect.
The question of what to do when there are several ways of going on to introduce
entities admits of a stronger answer as well, namely that the license to introduce
an entity must always be exercised immediately. This corresponds to the following
principle.
Maximality: At every stage, all the entities that can be introduced are in fact
introduced. 1
Although this principle too is fairly natural, it goes beyond the minimal conception
of the process of abstraction. Arguments that depend on this principle will therefore
be flagged as such.
The mentioned properties of the accessibility relation :'.S allow us to identify a modal
logic appropriate for studying the process of abstraction. Since :'.S is reflexive and
transitive, the modal logic S4 will be sound with respect to our intended system of
possible worlds. As is well known, the convergence of :'.S ensures the soundness of the
following principle as well:
(G) O D<p ~ DO<p.
The modal propositional logic which results from adding (G) to a complete axioma-
tization of S4 is known as S4.2.
Since the domains of our worlds always increase along the accessibility relation s,
there will be no need for any free logic or existence predicate once quantification is
introduced. Moreover, we can happily accept the fai;;t that our modal logic proves the
Converse Barcan Formula
(CBF) D'v'x<p(x) ~ 'v'xD<p(x),
which requires the domains to increase along the accessibility relation.

Summing up, I have glossed the relevant notion of modality and provided an
informal argument that an appropriate modal logic for this modality is S4.2. To make
things entirely explicit, let the closure of a formula be the result of prefixing its universal
closure with a 'D'. Unless context indicates otherwise, every open formula will
throughout this chapter be understood as short for its closure. With this convention
in place, we adopt the following modal first-order logic.
1
Maximality implies, but is not exhausted by, the requirement that :'.:: be a linear order. As stated,
Maximality is not a precise principle. See (Studd, 2013, section 4) for a way to make its set-theoretic
content sharp.
Definition 12.1 Let MFO (modal first-order logic) consist of classical S4.2, the usual
introduction and elimination rules for identity and the first-order quantifiers, and
the axiom x -::j=. y -+ D(x -::j=. y).z
This system will be our background modal logic in what follows. All the modal
assumptions that are used in the formal arguments described below-as well as
in the proof of the Mirroring Theorem in Appendix 3.B-are available in this
system. The talk about possible worlds is a motivational ladder that can now be
kicked away.
12.3 Plural Logic with Modality

Our next task is to describe the system of plural logic that figures in our dynamic
approach to set abstraction and in particular to formulate some principles concerning
the interaction of plurals and modal operators. The resulting system will be our
background logic throughout the remainder of the chapter.
What is the right logic of plurals? We begin by extending the usual introduction
and elimination rules to the plural quantifiers. Then we adopt the following plural
comprehension scheme:
(P-Comp) 3xx'v'u(u -< xx ~ cp(u)),
where cp ( u) does not contain 'xx' free. Here we deviate slightly from natural language
by allowing an empty plurality (cf. p. 67).
It should also be noted that, since cp(u) may contain bound plural variables, our
plural comprehension scheme is impredicative. Impredicative plural comprehension
can be motivated and justified by what (following (Bernays, 1935)) we may call a
"quasi-combinatorial conception of pluralities''. The idea is to extrapolate from the
finite to the infinite. Just as we can run through a finite plurality, making arbitrary
choices as to which elements are to be included in a subplurality and which are not,
we can idealize and assume this to be possible for any plurality. This is, of course, an
enormous extrapolation. But it is one that has become commonplace in mathematics
and set theory-and very successfully so. We therefore accept the extrapolation and
the comprehension scheme that it motivates.
Our next question is how plural logic interacts with the modal operators. My
guiding thought here is that a plurality comprises exactly the same objects at every
world at which the plurality exists. Or, in more semantic terms, a plural variable is
assigned exactly the same objects as its values at every world at which the variable has
any values at all. I submit that this thought coheres well with the use of plural locutions
2
Although 54.2 proves the necessity of identity, it does not prove the necessity of distinctness. For the
usual proof of th e latter depends on the Brouwerian Axiom cp-+ DOcp, which does not hold in 54.2. (This
obstacle can be overcome by adding further modal operators, as described in note 8.)
PLURAL LOGIC WITH MODALITY 209
in natural language. For instance, Harry is necessarily one of Tom, Dick, and Harry;
and if John is not one of them, then necessarily so. 3
We therefore adopt two axioms to the effect that being one of some objects, and not
being so, are stable from world to world:
u -< xx~ D(u -< xx)

u f xx ~ D(u f xx)
More generally, we say that a formula rp(u) is stable just in case the following two
conditionals hold (where we let variables in boldface abbreviate strings of variables): 4
(STB+-rp) rp(u) ~ Drp(u)
(STB- -rp) -.rp(u) ~ D-.rp(u)
However, the two stability axioms for '-<' do not fully capture our guiding thought
about the modal profile of plurals. Consider the condition that u is one of some given
objects xx. Although this condition cannot "change its mind" about objects existing at
one world as we go to larger worlds, the above axioms nevertheless allow its extension
to grow as we go to larger worlds-provided that this growth involves only objects
that are new at these larger worlds. An example of this phenomenon is provided by
the identity predicate. Although our background modal logic MFO proves that this
predicate is stable, its extension obviously grows as we go to larger worlds.
We would like to express that pluralities are "inextensible" and in this way fully
to capture our guiding thought. This can be done by adopting the following axiom
schema:
(IN EXT--<) 'v'u(u -< xx ~ oe) ~ D'v'u(u -< xx ~ e)
This formalization requires some explanation. Recall that the Barcan Formula
(BF) 'v'uoe ~ D'v'ue
corresponds to the semantic requirement that whenever a world w accesses another

world w', then the domain of w' be included in that of w. The axiom schema works by
relativizing the Barcan Formula to the condition that u is one of some given objects
xx. The axiom schema thus corresponds to the semantic requirement that whenever
a world w accesses another world w', then the objects that are among xx at w' be
included in the objects that are among xx at w. This is exactly the desired property of
inextensibility.
Our discussion motivates the adoption of the following modal plural logic.
Definition 12.2 Let PFO be the system that adds to standard first-order logic
the standard introduction and elimination rules for the plural quantifiers and the
3
Cf. Section 3.3.
4 This corresponds to the notion of"absoluteness" in (Parsons, 1983b). I find my term more suggestive.
comprehension scheme (P-Comp ). Let MPFO be the system that adds to PFO the
modal logic S4.2, the stability of distinctness (i.e. x =/= y ~ D(x =/= y)) , the stability
axioms (STB+ -<), (STB- -<), and all the instances of the inextensibility schema
(INEXT--<).
I wish to conclude this section by explaining how our modal plural logic can be
used to formalize a notion of extensional de.finiteness that will be important in later
sections. 5 It is easy to characterize the notion if we help ourselves to the idiom of
possible worlds: a formula <p is extensionally definite at a world w just in case its
extension remains the same at any later world w' '.'. '.: w. For instance, being one of a
particular plurality is extensionally definite (or so I have argued). But how should the
notion of extensional definiteness be formalized in our official idiom, which eschews
quantification over possible worlds? We can exploit the extensional definiteness that
MPFO ascribes to pluralities to formalize the extensional definiteness of <p . For
simplicity, we consider only the case where rp has a single free variable. 6
(ED-rp) 3xxDVu(u -<xx*"* <p(u))
This notion of extensional definiteness is very useful. 7

The adequacy of our formalization of extensional definiteness is demonstrated by
an easy lemma, which I state without proof.
Lemma 12.1 (ED-rp) entails the stability of <p and all instances (in the relevant
language) of the inextensibility schema for <p:
(STB+ -<p) <p(x) ~ D<p(x)
(STB- -<p) -i<p(x) ~ D-.<p(x)
(INEXT-<p) Vx(<p(x) ~ DO) ~ DV'x(<p(x) ~ 0)
Moreover, if two extensionally definite conditions are coextensive, then they are
necessarily coextensive.
Summing up, our plural logic makes two different kinds of technical contribution,
both of which will be important in what follows. It makes available arbitrary pluralities
from the relevant domain, as enshrined in the unrestricted plural comprehension
scheme. Additionally, it provides an entirely extensional way to track a collection
of objects between possible worlds and thus to formalize the important notion of
extensional definiteness. 8
5 This corresponds to the notion of "rigidity" in (Parsons, l 983b ). My alternative term is meant to
forestall confusion with the more familiar semantic notion of rigidity famous from Kripke.
6 Extensions to formulas with several free variables are immediate in cases (such as set theory) where
we can code for ordered pairs.
7
Recall, for example, the idea from the end of Section 10.S that a permissible (generalized) numeral
has to be tied to a definite position in a linear ordering. This idea can be expressed as the claim that, for
(u, R) to be a numeral (where Risa linear order and u an object in its field), there has to be a fixed plurality
of R-predecessors of u, that is, 3xxDVy(y -< xx ++ Ryu).
8 Alternative formalizations become available if further modal operators are added which enable us to
"look back" at smaller worlds, not just "forward" at larger ones. For instance, we may add a backtracking
THE NATURE OF SETS 211
12.4 The Nature of Sets

Using the logic that has just been motivated and explained, I shall now provide an
account of dynamic abstraction, focusing on our canonical example of sets. I begin by
providing an account of what sets are which is neutral on which sets there are. In the
next section, I use the potential collapse principle (COLLAPSE<> ) to explain how the
familiar sets can be generated and thus how most of the axioms of standard ZFC can
be retrieved.
My account of the nature of sets is based on the simple core idea that sets are
collections that are "constituted" by their elements. One aspect of the core idea is the
extensionality of sets. If a set is a collection that is "constituted" by its elements, then
the nature of a set is exhausted by which elements it has. Once you specify the elements
of a set, you have specified everything that is essential to it. In particular, if two sets
have the same elements, then they are identical. Another aspect of the core idea is the
priority of the elements of a set to the set itself. If a set is a collection "constituted" by
its elements, then these elements will have to be available "before" the set itself can
be formed. For instance, if at some stage an object a is available, we can immediately
form the new set {a} but not the new set {{a}}. For {{a}} is partially constituted by
{a}, which means that the latter must be available "before" the former can be formed.
Notice that our account of the nature of sets says nothing about what sets there are
but is concerned with what sets have to be like if there are any.
Let us now make the ideas just adumbrated technically precise. I begin by defining
the relevant languages. 9
Definition 12.3 Let .Ce be the language of ordinary non-modal set theory, and let
.Cpe be the corresponding plural language. Let .C~ and .C~E be the modal languages
that result by adding the modal operators to the mentioned non-modal languages
respectively.
12.4.1 The extensionality of sets

The extensionality of sets clearly motivates the familiar axiom of Extensionality:
(Ext) x = y ++ 'v'u(u E x ++ u E y)
operator .J. that exempts the formula to which it is applied from the innermost modal operator containing
this occurrence in its scope (Hodes, l 984b ). The extensional definiteness of rp can then be expressed as
follows: D'v'u[rp(u) t t ,(, (rp(u) /\ 3x(x = u))] . Another option, developed in (Studd, 2013), is to adopt a
second set of modal operators whose accessibility relation is the converse of that of the set of "forward-
looking" operators. Without adding "backward-looking" modal operators, however, I know of no first-
order and singular formalization of extensional definiteness. We thus face a choice between adopting a
more complex modal logic and adopting some form of plural or second-order logic. Without denying the
value of the former, I here choose the latter, not least because the relation between pluralities and sets is of
independent interest.
9
Strictly speaking, Le differs from the set-theoretic language used in Chapter 3. Where the former
language has 'e' as its sole non-logical predicate, the latter has 'SET(uu,x)'. But nothing hangs on this, as
the two predicates are straightforwardly interdefinable.
However, there is more to the extensionality of sets than this. Even if (Ext) is a
necessary truth, this only allows us to identify and distinguish sets within each world.
But the idea that sets are constituted by their elements supports a "transworld" version
of the principle of extensionality as well. Intuitively speaking, a set x from one world
w 1 is identical with a set y from another world w2 just in case x's elements at w1 are
precisely the same as y's elements at w2.
This "transworld" principle of extensionality is captured by the "intraworld" prin-
ciple (Ext) together with the claim that being an element of a particular set is
extensionally definite; that is, that a set has precisely the same elements at every
world at which it exists. As we know from Section 12.3, this claim about extensional
definiteness can be formalized as follows:
(ED-E) 3yyD'v'u(u -< yy *+ u E x)
By Lemma 12.l, we know that (ED-E) entails the stability axioms for E, as well as all
instances of the inextensibility schema for E:
(INEXT-E) 'v'x(x E y ---+ De) ---+ D'v'x(x E y ---+ e)
The extensional definiteness of elementhood enables us to prove a counterpart

of the familiar result from ordinary non-modal set theory that bounded first-order
formulas are absolute in transitive structures. To state the result, we first need a
definition.
Definition 12.4 An occurrence 'v'x<p of the universal singular quantifier is said to
be bounded iff <p is of the form x E y ---+ 1/t. An occurrence 3x <p of an existential
quantifier is said to be bounded iff <p is of the form x E y /\ 1/t. We write these formulas
as respectively ('v'x E y)i/! and (3x E y)l/t. A first-order formula is said to be bounded
(or .0.o) iff all of its quantifiers are bounded.
Lemma 12.2 Let . This theory also proves that <p is stable.
A proof of this result-and all the ones to follow-can be found in Appendix 12.A.
12.4.2 The priority of elements to their set

Recall our core idea that a set is a collection constituted by its elements. I have
addressed the first aspect of this idea: the transworld version of the principle of
extensionality. I now turn to the second aspect: the principle that the elements of a
set are prior to the set itself.
A natural first attempt to express this principle is as follows:
x E y ---+ <> (Ex/\ -.By)

THE NATURE OF SETS 213
Unfortunately, this attempt fails because our accessibility relation only allows a world
to access another world if the domain of the former is included in the domain of the
latter. 10
Fortunately, standard ZF set theory suggests an alternative. Recall that one of its
axioms is Foundation, which can be formalized as follows:
(F) Vx(3y(y E x) -+ 3y(y E x /\ 'Vz(z E x -+ z rf. y)))

Instead of searching for the ideal formalization of the priority principle, I propose
that we simply adopt (F) as an axiom. This is permissible because (F) follows from
the priority principle and our assumption that the accessibility relation between
possible worlds is well-founded. We see this by the following informal but compelling
argument. JI Consider a set x. Let C be the class of worlds that contain at least one
element of x. By the well-foundedness of :S, there is a :s-minimal member of this
class, say wo. Let y be an element of x that exists at wo. Assume some element z of xis
an element of y as well. By the priority principle, z would have to be present at some
world w1 < wo. But this contradicts the minimality of wo . So no element of x can be
an element of y as well, which gives us (F). 12
12.4.3 The extensional definiteness of subsethood

The notion of being an element of a particular set is extensionally definite, as argued
in Section 12.4.1. What about the notion of being a subset of a particular set? Is this
too extensionally definite? The question is whether the following is true:
(ED-5;) 3xxD'Vu(u -< xx ++ u 5; a)
The following lemma provides a partial answer.
Lemma 12.3 Consider the .C~E-theory consisting of MPFO and the axioms (Ext),
(ED-E), and (F). This theory proves:
(STB+ -5;) u 5; a-+ D(u 5; a),

(STB- -5;) u %a -+ D(u %a).
10 (Studd, 2013) formulates the well-foundedness of the accessibility relation by means of a version of
the Lob principle D ( Dip 4 qi ) 4 Dip for his backward-looking necessity operator (see note 8). He also
shows how the priority principle can be expressed and how this enables the derivation of (the potentialist
translation of) Foundation as a theorem. This is a satisfying feature of his bimodal approach.
11
The argument cannot be formalized in any of the theories set out in this chapter but could easily be
formalized in a set theory with classes.
12
Attentive readers may have noticed that this argument does not distinguish very carefully between (F)
and its potentialist translation (F 0 ). This is permissible because Lemma 12.2 ensures that (F 0 ) is equivalent
to the result of prefixing (F) with a single necessity operator, which we may anyway do to an axiom.
However, the theory does not prove all instances of

(INEXT-s;) 'v'u(u s; a --+ DO) --+ D'v'u(u s; a --+ ti)
and hence also not (ED-s;). 13

Lemma 12.3 raises the question whether (ED-s;) should be adopted as an axiom.
Is (ED-s;) supported by our core idea that sets are constituted by their elements? The
answer depends on how we understand the process of set formation. Consider a set
a and some subset b s; a that is present at some later world. When a was formed,
all of its elements must already have been available. So a fortiori all the elements
of b must have been available. When a was formed, we therefore had the ability to
form b. But was this ability exercised? According to the principle of Maximality-
which says that we always form all the sets that we are capable of forming-the
answer is yes and (ED-s;) will thus hold. As Lemma 12.3 shows, without Maximality,
(ED-s;) can fail.
I shall distinguish between theories that assume Maximality and those that do
not by attaching a minus sign to the name of each theory that does not assume this
principle. The following definition sums up our discussion of the nature of sets.
Definition 12.5 Let NS be the .C~E -theory that adds to MPFO the following axioms:
(Ext), (ED-E), (F), and (ED-s;). Let NS- be like NS except without the axiom (ED-s;).
In NS we can prove an extension of Lemma 12.2 to formulas with bounded plural

quantifiers. This corresponds to the familiar set-theoretic fact that bounded second-
order formulas are absolute in supertransitive structures, that is, in structures that are
both transitive and contain as an element each subset of each of their elements.
12.5 Recovering the Axioms of ZF

My analysis of the nature of sets has resulted in various conditional claims about
what sets there are. But quite reasonably, it did not yield any categorical claims about
the existence of sets. Our single principle of set existence states that every plurality
potentially forms a set:
D'v'.xx 03y SET(xx, y)
In the apt characterization of (Yablo, 2006, p. 151), this is "the principal engine of set
production''.
12.5.1 From conditions to sets

Consider the naive set comprehension scheme:
(S-Comp) 3x'v'u(u E x ~ <p(u))
13 Nor does (!NEXT -~ ) ensure (INEXT- E). For there is a model ofMPFO, (Ext), the stability axioms for
E, (INEXT-s;}, and the assumption that every object has a singleton in which (INEXT-E) fails.
RECOVERING THE AXIOMS OF ZF 215
Which instances are valid? Transposed to our potentialist setting and slightly gener-
alized, the question is which instances of the following scheme are valid:
(S-Comp 0 ) 0 3x D\t'u(u Ex# <p(u))
Our analysis of the nature of sets provides half of the answer. I argued that any set
x is extensionally definite, in the sense that the objects that comprise x's members
necessarily comprise its members:
3yyD\t'u(u E x # u -< yy)
It follows that, if (S-Comp 0 ) holds for <p(u), then possibly there are some objects that
necessarily are the extension of <p(u):
<>3yyD\t'u(u -< yy # <p(u))
Our analysis of the nature of sets thus implies that, if a condition <p ( u) defines a set-
in the sense that (S-Comp 0 ) holds-then it is possible for <p(u) to be extensionally
definite.
What about the converse? If it is possible for a condition to be extensionally
definite, does the condition define a set-in the sense that (S-Comp 0 ) holds? It is
easily seen that (CoLLAPsE 0 ) not only entails this converse conditional but in fact is
equivalent to it.
Putting the two conditionals together, we obtain the desired answer to our question:
a condition defines a set-in the sense that (S-Comp 0 ) holds for it-just in case it is
possible for the condition to be extensionally definite. This is a pleasing answer. It
is part of the nature of sets to be extensionally definite. So a condition is intrinsically
suited for defining a set just in case it is possible for it to be extensionally definite. On
the present account, a condition therefore defines a set just in case it is intrinsically
suited for doing so. There is nothing mysterious or surprising about the fact that some
conditions fail to define sets. Such conditions have an inexhaustible character that
renders them intrinsically unsuitable for defining sets.
Two remarks about this analysis are in order. First, the above analysis provides a
nice example of how the modal approach to set theory makes available explanations
that are unavailable on the ordinary non-modal approach. We started with a simple
question that is stated without any modal resources, namely which conditions admit
of naive set comprehension (S-Comp ). But the answer we have given makes essential
use of the finer resolution afforded by the modal approach. For the key notion of a
condition's possibly being extensionally definite is not the potentialist translation of
any non-modal formula and is thus expressible only in the richer modal framework.
(Of course, when we adopt this finer resolution, (S-Comp) must be replaced by its
potentialist counterpart (S-Comp 0 ).) In short, the non-modal question of which
conditions define sets admits of a more attractive and natural answer in a modal
framework than in a traditional non-modal one.
Second, the analysis is closely related to Cantor's famous idea that every "consistent
multiplicity" forms a set, whereas "inconsistent" ones do not. 14 A multiplicity is said
to be consistent just in case it can be "completed': such that all of its members "exist
together". It is useful to think of a Cantorian multiplicity in terms of its defining
condition. 15 A consistent multiplicity then corresponds to a condition that is possibly
extensionally definite (and which therefore defines a plurality). An inconsistent
multiplicity corresponds to a condition that cannot be extensionally definite. In fact,
when a condition rp(u) is stable (as any fully modalized condition is by Lemma 12.1),
then the claim that rp(u) cannot be extensionally definite is easily seen to be equivalent
to the claim that it is indefinitely extensible, in the sense that:
(IE-rp) D'v'xx<>3u(rp(u) /\ u -f. xx) 16
12.5.2 Basic modal set theory

The potential collapse principle (COLLAPSE 0 ) plays a key role in our recovery of many
of the familiar axioms of Zermelo-Fraenkel set theory. Indeed, many of these axioms
follow already from (COLLAPSE 0 ) and our theory of the nature of sets. We begin by
reminding ourselves of these axioms.
Definition 12.6 Zermelo-Fraenkel set theory (or ZF) is the Ce-theory whose axioms
are the extensionality principle (Ext), Foundation (F), and the following set existence
claims:
(Empty Set) 3x'v'u (u ¢ x)

(Pairs) 3x'v'u (u Ex# u =av u = b)
(Union) 3x'v'u (u E x # 3v(u E v /\ v E a))
(Separation) 3x'v'u (u Ex# u E a/\ rp(u))
(Power) 3x'v'u (u E x # u ~ a)
(Infinity) 3x(0 Ex /\ 'v'u(u Ex~ {u} Ex))
(Replacement) 'v'u3!v 1/t(u, v) ~ 'v'x3y('v'u E x)(3v E y) 1/t(u, v)
Zermelo set theory (or Z) is ZF without Replacement. We add a minus sign to the name
of a theory to indicate the theory that results from removing the Power set axiom.
14 See Cantor's famous 1897 letter to Hilbert and 1899 letter to Dedekind, both of which are found in
(Ewald, 1996).
15 Compare (Parsons, 1977, section 3), who argues that Cantorian multiplicities should be understood
as intensional entities, specified by means of application conditions.

16
According to (Dummett, 1993, p. 441 ), a concept is indefinitely extensible if, for any definite totality of
objects falling under the concept, there is another object that falls under the concept but is not a member of
the totality. I believe the above characterization of indefinite extensibility is a good explication of Dummett's
notion; cf. also Section 3.7. See (Shapiro and Wright, 2006) for an attempt at a non-modal characterization.
RECOVERING THE AXIOMS OF ZF 217
Theorem 12.1
(a) NS + (COLLAPSE 0 ) proves the potentialist translations of all the axioms of
Z - Infinity.
(b) NS- + (COLLAPSEO) proves the potentialist translations of all the axioms of
z- - Infinity.
(c) Each of the non-modal set theories mentioned in (a) and (b) is interpretable in
the corresponding modal set theory.
Notice that the power set axiom is the only axiom of ZF whose justification requires
(ED-~) and thereby also the Maximality principle (which, we recall, states that at
every stage, all sets that can be formed are formed). This observation shows how
precise content can be attached to the widespread sentiment that the power set axiom
is very strong.
12.5.3 Full modal set theory
To arrive at full ZF, we still need the axioms of Infinity and Replacement. We shall
now see how potentialist translations of these axioms can be derived if some further
plausible assumptions are added to our modal set theory. Just as in Theorem 12.1,
this establishes that larger fragments of ordinary ZF are interpretable in the various
systems of modal set theory.
One natural further assumption is that extensional definiteness is a matter of size. 17
Consider two one-place conditions cp and l/f. If cp is extensionally definite and every
cp is correlated with a unique 1/1, then 1/f too is extensionally definite. This motivates a
replacement principle for the notion of extensional definiteness. Let FuNc( 1/1° (u, v))
abbreviate the following formalization of the claim that 1/1 <> (u, v) is functional:
D'v'u 0 3v D'v'v' ( 1/1 o (u, v') ++ v = v') .
Then we have:
(ED-Repl) FuNc(i/! 0 (u, v)) -+ D'v'xx0 3yy('v'u -< xx) (3v-< yy) 1/1° (u, v)
Theorem 12.2 NS- + (CoLLAPsE 0 ) + (ED-Repl) entails the potentialist translation
of Replacement.
Another natural principle says that truths about the potential hierarchy of sets
are "reflected" in truths about individual possible worlds. This corresponds to the
following reflection principle:
(0 -Refl)
17
(Shapiro and Wright, 2006) make an analogous assumption in their non-modal approach to the
notion of indefinite extensibility.
(Recall that this is short for the closure of the displayed formula.) This principle is best
understood as stating that the truth of a claim in "the model" provided by the potential
hierarchy of sets ensures that the claim is possible. For a claim cp to be true in this
"model" is for cp to be true when all of its quantifiers are understood as ranging over all
possible sets, including ones not yet formed. But for cp to be true when understood in
this way is simply for its potentialist translation cp <> to be true. The principle therefore
says that the truth of cp <> ensures the possibility of cp.
Three remarks about this reflection principle are in order. First, the principle is not
specifically about sets but about the process of forming mathematical objects more
generally. The principle says that we should recognize any situation that is realized in
the potential hierarchy as genuinely possible. ( 0 -Refl) thus complements the modal
principles described in Section 12.2. Where those principles characterize the structure
of the space of possible worlds, ( 0 -Refl) says something about its extent.
Second, it is often claimed that reflection principles have a "top-down" character
that sits poorly with the view of the hierarchy of sets as potential. 18 Indeed, the usual
form of a reflection principle
cp -7 3a cp Va
suggests that the entire hierarchy has to be available in order for cp to be evaluated and
reflected down to an appropriate initial segment Va. The formulation (0-Refl) shows
that this impression is incorrect and illustrates how a reflection principle can enjoy a
more "bottom-up" form of motivation.
Finally, there is one serious worry about (0 -Refl). When we apply the principle to
(COLLAPSE<> ) and do a modus ponens, we get the claim that possibly every plurality
forms a set:
O'lxx3y\lu(u E y ++ u -< xx )
But this claim is, in fact, inconsistent, as is seen by instantiating the quantifier V xx with
respect to the plurality of all non-self-membered sets and reproducing the reasoning
of Russell's paradox.
Fortunately, the motivation I offered for (0 -Refl) allows me to explain what has
gone wrong. The motivation was that every possibility that is realized in the potential
hierarchy of sets should also be realized at some possible world. But when cp contains
plural variables, the possibility witnessed by the truth of cp <> can be seen not quite to
correspond to the possibility described by cp . In the potential hierarchy of sets every
plurality corresponds to a set, as is ensured by (COLLAPSE<> ). So in this hierarchy there
is a one-to-one correspondence between pluralities and sets. The possibility witnessed
by cp <> is thus a possibility in which any plural quantifiers range over precisely such
pluralities as form sets. By contrast, the possibility described by cp is a possibility in
18
See for example (Koellner, 2009} and (Welch and Horsten, 2016).
APPENDICES 219
which the plural quantifiers range over more pluralities than those that form sets.19
Thus, when cp contains plural variables, the possibility witnessed by cp <> is importantly
different from the possibility described by cp. How can we avoid this mismatch and be
faithful to the motivation that was offered for the reflection principle? The simplest
option is to restrict the reflection principle ( <> -Refl) to singular formulas cp, for which
the mismatch does not arise. Moreover, I show in Appendix 12.B that nothing is lost
as a result of this restriction.
With an adequate formulation of the reflection principle now in place, the road is
open to the desired theorem.20
Theorem 12.3 NS- +(COLLAPSE<> )+ (<> -Refl) proves the potentialist translation of
Infinity.
Our final theorem is concerned with the reverse interpretation and thus establishes
a relative consistency result for our modal set theory.
Definition 12.7 Let MS be the modal theory based on NS, (COLLAPSE<> ), (ED-Repl),
and ( <>-Refl).
Theorem 12.4 The modal set theory MS is interpretable in the non-modal theory ZF
and is therefore consistent provided that ZF is.
Appendices
12.A Proofs of Formal Results
Proof of Lemma 12.2. The proof of the first claim goes by induction on the complexity of <p.
The base, conjunction, and negation cases are trivial. So consider the quantifiers. Assume - (Vx E a) l/1 °. Since the induction hypothesis ensures that 1/t and 1/t o are provably
equivalent, this can be strengthened to O(Vx E a)l/t o - (Vx E a)l/t, i.e. cp<> - <p.
19
In fact, (P-Comp) allows us to prove an analogue of Cantor's theorem which says that there are more
pluralities than there are objects, and a fortiori more than there are sets.
20
A stronger reflection principle can be formulated as follows:
(<> -Ref!+) O'v'x(cp 0 (x) ~ cp(x))
where x is a string of singular variables, and where again -Ref!), as is seen by letting x be the empty string and observing that O (cp 0 ~ cp)
implies cp 0 ~ Ocp. In fact, (<>-Ref!+) corresponds to the "complete" non-modal reflection principle
'v'a (3{3 > a )(Vx E Ya )(cp(x) # cp(x) v. ), whereas (0 -Refl) corresponds to the "partial" principle
Vx(cp(x) ~ 3a cp (x) v• ). (<> -Ref!+ ) implies the potentialist translation of all instances of Replacement,
as can be seen by imitating the standard proof of the non-modal analogue of this result ((Levy, 1960, Tum.
6) and (Drake, 1974, p. 102, exercise 2); see also (Studd, 2013, Section 4.6)) . So (<>- Ref!+) holds out the
promise of a unified route to Infinity and Replacement. Unfortunately, I cannot see how the appealing
motivation provided for (<> -Ref!) can be extended to (<>-Ref!+). And the promised unification requires the
full strength of (0 -Refl+): for although the two corresponding non-modal principles are equivalent in a
higher-order setting, (Levy and Vaught, 1961) shows the complete principle to be strictly stronger than the
partial one when only (singular) first-order logic is available.
For the other direction, assume ('v'x E a)ijl, which abbreviates 'v'x(x E a ~ i/J). Observe first
that the induction hypothesis yields that ijl and ijl 0 are provably equivalent. Observe next that
since ijl 0 is fully modalized, Lemma 3. l yields ijl 0 ~ D ijl 0 . Using these two observations,
our assumption implies 'v'x(x E a ~ Di/1 ° ). By Lemma 12.l, we can apply (lNExT-E) to get
D'v'x(x Ea~ i/f 0 ), which abbreviates as D('v'x E a)i/1 ° . This establishes rp ~ rp 0 , as desired.
The second claim- that rp is stable-follows by applying Lemma 3.1 to the first claim. -I
Proof of Lemma 12.3. The two positive claims follow from Lemma 12.2 and the observation
that the formulas u ~ a and u %'; a are bounded. The negative claims can be established by
constructing a simple countermodel. Consider a world with two objects a and b but no sets
containing either of these objects. Then introduce {a, b} first and {a} later. Let() express that
u has two elements. This yields a counterexample to (!NEXT - ~). By Lemma 12.l, this is also a
counterexample to (ED-~). -I
Proof of Theorem 12.1. Observe first that the following conditions are extensionally
definite:
(i) u -I= u
(ii) u = a v u = b
(iii) <>3x(u E x /\ x E a)
(iv) u E a /\ rp(u) where rp(u) is stable
(v) u ~a
As for (i) and (ii), this is trivial. (iii) and (iv) follow easily from the extensional definiteness of
membership. (v) follows from (ED-~). Next we observe that when (COLLAPsE 0 ) is applied to
the pluralities defined by these conditions, we obtain the potentialist translations of the axioms
of Empty Set, Pairing, Union, Separation, and Power. (For Separation we appeal to Lemma 3.1
to ensure the stability of the modal translation of the formula in question. 21 ) The claim about
interpretability is now immediate from (a), (b), and Theorem 3.1. -I
Proof of Theorem 12.2. Assume FuNc(i/f 0 (u, v)) and consider a set x. By Lemma 12.2,
it suffices to show that there may be a set y such that ('v'uEx) (3vey) i/1 ° (u,v). Let xx
be the elements of x and apply (ED-Rep!) to derive <> 3 yy('v'u -< xx)(3v-< yy) i/1 ° (u, v). By
(CoLLAPSEo ) we know that yy possibly form a set y: 0 3yD'v'u(u E y # u-< yy). It follows
that <>3y('v'u E x) (3v E y) i/J o (u, v), as desired. -I
Proof of Theorem 12.3. By Theorem 12.l, NS- + (COLLAPsE 0 ) proves <>3x(x = 0 ) and
D'v'u 03v(v = {u}). Applying (<>- Ref!) to the conjunction of these two formulas yields:
(12.l) <>(3x (x = 0 ) /\ 'v'u3v(v = {u})) .
A routine trick available in S4 allows us to turn a proofof p ~ <>q into a proofof <>p ~ <>q. In
order to prove (12.l) ~ Infinity 0 we may therefore ignore the initial modal operator of (12.l).
21 Without the convergence of :=:, Lemma 3.1 would fail and we would only be able to justify Bounded
Separation. Compare Theorems 1 and 2 in (Parsons, 1983b, p. 337). However, as (Parsons, 1983b, p. 339)
observes, full Separation can be justified when the potentialist set theory is supplemented with a reflection
principle; cf. note 20.
APPENDICES 221
So assume 3x(x = 0) /\ Vu3v(v = {u}). Apply (P-Comp) to a tautology to get a plurality xx

(namely the plurality of all objects at the relevant world) such that
(12.2) 0 -<xx/\ Vu(u -< xx-+ {u} -< xx).
In fact, this is easily strengthened to 0(12.2). Next, by (COLLAPSE 0 ) it is possible for xx to form
a set:
(12.3) 03xOVu(u E x - u -< xx).
We now observe that OVu(u E x - u -< xx) and 0(12.2) imply 0 E x /\ OVu(u E x -+
{u} Ex). This makes it straightforward to advance from (12.3) to Infinity 0 :
0 3x(0 Ex /\ OVu(u Ex-+ {u} E x))
(We use Lemma 12.2 and the fact that '{u} Ex' can be expressed using bounded quantifiers.)
This concludes our proof. -1
Proof of Theorem 12.4. We begin by providing a recursive definition of a translation

cp ~ [cp]Vo from .C~e to .Ce . The translation is trivial on atomic formulas, commutes with
the truth-functional connective, and is otherwise as follows:
(12.4) [u-< xx] Va = u E xx

(12.5) [Vxcp]Va = (Vx E Y,,)[cp]Va
(12.6) [Vxxcp]Va = (Vxx E Ya+1)[cp]Va
(12.7) [Ocp)Va = (Vf3 ~ a)[cp)Vp
where the occurrences of 'xx' on the right-hand sides of (12.4) and (12.6) are understood as
occurrences of an ordinary singular variable written in an unusual way. We now need to verify
that all the axioms of MS are mapped to theorems of the non-modal theory and that logical
relations are preserved.
The axioms ofS4.2 (that is, O(cp -+ 1/1) -+ (Ocp-+ 01/1), Ocp-+ cp, Dcp-+ DDcp, OOcp -+
DOcp) are easily seen to be mapped to truths of first-order logic. Next we consider the axioms
of plural logic. The instances of (P-Comp) translate as formulas of the form:
('v'a)(3xx E Ya+i)(Vu E V,,)(u E xx - [cp]Va),
which are easily seen to be theorems of Z - Infinity. The axioms (STB+ -<), (STB- -<), and
(!NEXT--<) are easily seen to be mapped to theorems of Z - Infinity. The axioms of NS-that
is, (Ext), (ED-E), Foundation, and (INEXT-s;)-are easily seen to be mapped to theorems of
Z - Infinity. Further, (COLLAPSE 0 ) is mapped to the following (trivial but awkward) theorem:
(Va)(Vxx E Ya+1)(3f3 ~ a)(3y E Y,B)(Vy ~ f3)(Vu E Yy)(u E y - u E xx)
Next, (ED-Rep!) is easily seen to translate to the ordinary replacement scheme. Finally, the
translation of (the necessitation of) (0- Refl) is easily seen to be equivalent to:
Va(3f3 ~ a)(cp - [cp)VP).
When cp is non-modal and purely singular, it is not hard to see that this can be simplified to:
(Ref!)
where <p v~ is just the ordinary relativization of <p to Vfl. But this is just the ordinary non-modal
reflection principle, which is known to be a theorem scheme of ZF.
It remains to prove that every step licensed by an inferences rule of the modal theory is
mapped to a step that is licensed by the non-modal theory. In the case of the introduction and
elimination rules for the quantifiers, this is a straightforward but tedious verification. The only
other rule is the rule of necessitation from S4.2: iff- rp, then f- Drp. The non-modal translation
of this rule says that if we can prove [rp] Vo, then we can prove Va [rp] Va. We establish this
claim by induction on proofs: all the axioms have this property, and all the inference rules
preserve it. -l
12.B A Harmless Restriction

Recall from p. 219 the proposal that the reflection principle (0 -Refl) be restricted to singular
formulas rp. One may worry that this restriction is too blunt and fails to capture fully the
motivation provided for the reflection principle. Fortunately, this worry is assuaged by our next
lemma, which shows that every possibility <p 0 to which we may wish to apply the reflection
principle admits of an equivalent singular characterization. The desired instance of reflection
is thus fully captured by (0 -Refl) even when this principle is subjected to the mentioned
restriction.
Lemma 12.4 Let <p be a fully modalized sentence of C~e. Let rp' be the result of replacing every
plural variable uu; of <p with a singular variable x ; (assumed not already to occur in rp') and
replacing every occurrence of'-< ' with an occurrence of'E'. Then NS- + (COLLAPsE 0 ) proves
<p - rp'.
Proof We prove a more general claim where <p may contain free variables. Assume the free
plural variables in <pare uui, .. . , uun. Then I claim that the mentioned theory proves
X1 = {uui} /\ . .. /\ Xn = {uun} ~ (rp *+ rp')
where x = {uu} abbreviates Vv(v Ex*+ v-< uu). This claim is easily established by induction
on the number of (modalized) quantifiers in rp . -l
Bibliography
Balaguer, M. (1998). Platonism and Anti-Platonism in Mathematics. Oxford University Press,

Oxford.
Beaney, M. (1997). The Frege Reader. Blackwell, Oxford.
Benacerraf, P. (1973). Mathematical truth. Journal of Philosophy, 70(19):661-79. Reprinted in
(Benacerraf and Putnam, 1983).
Benacerraf, P. and Putnam, H., editors (1983). Philosophy of Mathematics: Selected Readings,
Cambridge University Press, Cambridge, second edition.
Bernays, P. (1935). On platonism in mathematics. Reprinted in (Benacerraf and Putnam, 1983).
Boghossian, P.A. (1996). Analyticity reconsidered. Nous, 30(3):360-91.
Boolos, G. (1971). The iterative conception of set. Journal of Philosophy, 68(8):215-32.
Reprinted in (Boolos, 1998). '
Boolos, G. (1984). To be is to be a value of a variable (or to be some values of some variables).
Journal of Philosophy, 81 (8):430-49. Reprinted in (Boolos, 1998).
Boolos, G. (1985) . Nominalist platonism. Philosophical Review, 94(3):327-44. Reprinted in
(Boolos, 1998).
Boolos, G. (1990) . The standard of equality of numbers. In Boolos, G., editor, Meaning
and Method: Essays in Honor of Hilary Putnam, pages 261-78. Harvard University Press,
Cambridge, MA. Reprinted in (Boolos, 1998).
Boolos, G. (1997). Is Hume's Principle analytic? In Heck, R., editor, Logic, Language, and
Thought, pages 245-61. Oxford University Press, Oxford. Reprinted in (Boolos, 1998).
Boolos, G. (1998). Logic, Logic, and Logic. Harvard University Press, Cambridge, MA.
Brandom, R. (1996). The significance of complex numbers for Frege's philosophy of mathemat-
ics. Proceedings of the Aristotelian Society, 96(1):293-315.
Burge, T. (2005). Truth, Thought, Reason: Essays on Frege. Oxford University Press, Oxford.
Burge, T. (2010). Origins of Objectivity. Oxford University Press, Oxford.
Burgess, J. P. (1999) . Review of Stewart Shapiro, Philosophy of Mathematics: Structure and
Ontology. Notre Dame Journal of Formal Logic, 40(2):283-91.
Burgess, J. P. (2004). E Pluribus Unum: Plural logic and set theory. Philosophia Mathematica,
12(3):193-221.
Burgess, J. P. (2005). Fixing Frege. Princeton University Press, Princeton, NJ.
Burgess, J. P. and Rosen, G. (1997). A Subject with No Object. Oxford University Press, Oxford.
Cantor, G. (1883). Grundlagen einer allgemeinen Mannigfaltigkeitslehre. Bin mathematisch-
philosophischer Versuch in der Lehre des Unendlichen. B.G. Teubner, Leipzig. English transla-
tion in (Ewald, 1996).
Carey, S. (2009) . Where ournumber concepts come from. Journal ofPhilosophy, 106(4):220-54.
Cartwright, R. L. (1994). Speaking of everything. Nous, 28(1):1-20.
Cook, R. T. (2009). Hume's big brother: Counting concepts and the bad company objection.
Synthese, 170(3):349-69.
Cook, R. T. and Ebert, P.A. (2005). Abstraction and identity. Dialectica, 59(2):121 - 39.
224 BIBLIOGRAPHY
Cook, R. T. and Linnebo, 0. (2018) . Cardinality and acceptable abstraction. Notre Dame Journal
of Formal Logic, 59(1):61-74.
Dedekind, R. (1888) . Was Sind und Was Sol/en die Zahlen? Vieweg, Braunschweig. English
translation in (Ewald, 1996).
Donaldson, T. (2017). The (metaphysical) foundations of arithmetic. Nous, 51(4):775- 801.
Drake, F. R. (1974) . Set Theory: An Introduction to Large Cardinals. North-Holland, Amsterdam.
Dummett, M. (1956). Nominalism. Philosophical Review, LXV(4):491-505. Reprinted in
(Dummett, 1978).
Dummett, M. (1963). The philosophical significance ofGodel's theorem. In Truth and Other
Enigmas (1978), pages 186-214. Harvard University Press, Cambridge, MA.
Dummett, M. (1978). Truth and Other Enigmas. Harvard University Press, Cambridge, MA.
Dummett, M. (198la). Frege: Philosophy of Language. Harvard University Press, Cambridge,
MA, second edition.
Dummett, M. (198lb). The Interpretation of Frege's Philosophy. Harvard University Press,
Cambridge, MA.
Dummett, M. (199la). Frege: Philosophy of Mathematics. Harvard University Press,
Cambridge, MA.
Dummett, M. (199lb). The Logical Basis of Metaphysics. Harvard University Press,
Cambridge, MA.
Dummett, M. (1993). What is mathematics about? In The Seas of Language, pages 429-45.
Oxford University Press, Oxford.
Dummett, M. (1995). The context principle: Centre of Frege's philosophy. In Max, I. and
Stelzner, W., editors, Logik und Mathematik: Frege-Kol/oquium Jena 1993, pages 3-19.
de Gruyter, Berlin.
Dummett, M. (1998) . Neo-Fregeans in bad company? In Schirn, M., editor, Philosophy of
Mathematics Today, pages 369-88. Clarendon, Oxford.
Dummett, M. (2007). Reply to Peter M. Sullivan. In Auxier, R. E. and Hahn, L. E., editors, The
Philosophy of Michael Dummett, volume 31 of Library of Living Philosophers, pages 786-99.
Open Court, Chicago, IL.
Ebert, P.A. (2016) . Frege on sense identity, Basic Law V, and analysis. Philosophia Mathematica,
24(1):9-29.
Eklund, M. (2006a). Metaontology. Philosophy Compass, 1(3):317-34.
Eklund, M. (2006b ). Neo-Fregean ontology. In Hawthorne, J., editor, Philosophical Perspectives,
volume 20: Metaphysics, pages 95-121. Blackwell, Oxford.
Enderton, H.B. (2001). A Mathematical Introduction to Logic. Academic Press, San Diego, CA,
second edition.
Evans, G. (1982) . Varieties of Reference. Oxford University Press, Oxford.
Ewald, W. (1996). From Kant to Hilbert: A Source Book in the Foundations of Mathematics,
volume 2. Oxford University Press, Oxford.
Field, H. (1989) . Realism, Mathematics, and Modality. Blackwell, Oxford.
Fine, K. (2002). The Limits of Abstraction. Oxford University Press, Oxford.
Fine, K. (2005a). Class and membership. Journal of Philosophy, 102(11):547-72.
Fine, K. (2005b). Our knowledge of mathematical objects. In Gendler, T. S. and Hawthorne, J.,
editors, Oxford Studies in Epistemology, volume l, pages 89-109. Oxford University Press,
Oxford.
BIBLIOGRAPHY 225
Fine, K. {2006). Relatively unrestricted quantification. In Rayo, A. and Uzquiano, G., editors,
Absolute Generality, pages 20-44. Oxford University Press, Oxford.
Fine, K. (2012a). Guide to ground. In Correia, F. and Schnieder, B., editors, Metaphysical
Grounding, pages 37-80. Cambridge University Press, Cambridge.
Fine, K. {2012b}. The pure logic of ground. Review of Symbolic Logic, 5(1):1-25.
Fodor, J. A. (1987}. Psychosemantics: The Problem of Meaning in the Philosophy of Mind. MIT
Press, Cambridge, MA.
Frege, G. {1879}. Begriffsschrift: Eine der arithmetischen nachgebildete Formelsprache des
reinen Denkens. Translated and reprinted in (van Heijenoort, 1967}.
Frege, G. (1891). Function and concept. Reprinted in (Beaney, 1997).
Frege, G. {1892}. Concept and object. Reprinted in (Beaney, 1997).
Frege, G. {1953). Foundations of Arithmetic. Blackwell, Oxford. Translated by J. L. Austin.
Frege, G. (1956}. The thought: A logical inquiry. Mind, 65(259):289-311.
Frege, G. {1963}. Compound thoughts. Mind, 72(285):1-17. Originally published in 1923.
Frege, G. {1979). Posthumous Writings. Blackwell, Oxford.
Frege, G. {1980}. Philosophical and Mathematical Correspondence. University of Chicago Press,
Chicago, IL. Edited by G. Gabriel et al., translated by H. Kaai.
Frege, G. {2013). Basic Laws of Arithmetic. Oxford University Press, Oxford. Translated by
P. A. Ebert and M. Rossberg.
Geach, P. T. (1962). Reference and Generality. Cornell University Press, Ithaca, NY.
Godel, K. {1944). Russell's mathematical logic. In (Benacerraf and Putnam, 1983).
Hale, B. {1987}. Abstract Objects. Blackwell, Oxford.
Hale, B. {1997). Grundlagen §64. Proceedings ofthe Aristotelian Society, 97(3):243-61. Reprinted
with a postscript in (Hale and Wright, 200la).
Hale, B. (2007}. Kit Fine on The Limits of Abstraction. Travaux de Logique, 18:103-29.
Hale, B. (2013). Necessary Beings: An Essay on Ontology, Modality, and the Relations between
Them. Oxford University Press, Oxford.
Hale, B. and Wright, C. (2000}. Implicit definition and the a priori. In Boghossian, P. and
Peacocke, C., editors, New Essays on the A Priori, pages 286-319. Oxford University Press,
Oxford. Reprinted in (Hale and Wright, 200la).
Hale, B. and Wright, C. (200la). Reasons Proper Study. Clarendon, Oxford.
Hale, B. and Wright, C. (200lb). To bury Caesar. . . . In (Hale and Wright, 200la), pages 335-96.
Hale, B. and Wright, C. {2007). Abstraction and additional nature. Philosophia Mathematica,
16(2):182-208.
Hale, B. and Wright, C. {2009a). Focus restored: Comments on John Macfarlane. Synthese,
170(3):457-82.
Hale, B. and Wright, C. (2009b ). The metaontology of abstraction. In Chalmers, D., Manley, 0.,
and Wasserman, R., editors, Metametaphysics: New Essays on the Foundations of Ontology,
pages 178-212. Oxford University Press, Oxford.
Hazen, A. {1985}. Review of Crispin Wright, Freges Conception of Numbers as Objects.
Australasian Journal of Philosophy, 63(2):250-4.
Heck, Jr., R. G. (1992). On the consistency of second-order contextual definitions. Nous,
26(4):491-4.
Heck, Jr., R. G. (1997a) . Finitude and Hume's Principle. Journal of Philosophical Logic,
26(6):589-617.
226 BIBLIOGRAPHY
Heck, Jr., R. G. (1997b). Grundgesetze der Arithmetik I §§29-32. Notre Dame Journal of Formal
Logic, 38(3):437-74.
Heck, Jr., R. G. (1997c). The Julius Caesar objection. In Heck Jr., R. G., editor, Language,
Thought, and Logic: Essays in Honour of M. Dummett. Oxford University Press, Oxford.
Heck, Jr., R. G. (2000a) . Cardinality, counting, and equinumerosity. Notre Dame Journal of
Formal Logic, 41(3):187-209.
Heck, Jr., R. G. (2000b). Syntactic reductionism. Philosophia Mathematica, 8(2):124- 49.
Heck, Jr., R. G. (2011). Freges Theorem. Oxford University Press, Oxford.
Heck, Jr., R. G. (2012) . Reading Freges Grundgesetze. Oxford University Press, Oxford.
Hellman, G. (1989) . Mathematics without Numbers. Clarendon, Oxford.
Hodes, H. (1984a). Logicism and the ontological commitments of arithmetic. Journal of
Philosophy, 81(3):123-49.
Hodes, H. (1984b). On modal logics which enrich first-order S. Journal of Philosophical Logic,
13( 4):423-54.
Hodes, H. (1990) . Ontological commitment thick and thin. In Boolos, G., editor, Meaning
and Method: Essays in Honor of Hilary Putnam, pages 235-60. Cambridge University Press,
Cambridge.
Horsten, L. (2010). Impredicative identity criteria. Philosophy and Phenomenological Research,
80(2):411-39.
Horsten, L. and Linnebo, 0 . (2016). Term models for abstraction principles. Journal of Philo-
sophical Logic, 45(1):1-23.
Jane, I. and Uzquiano, G. (2004). Well- and non-well-founded Fregean extensions. Journal of
Philosophical Logic, 33(5):437-65.
Janssen, T. M. V. (2001 ). Frege, contextuality and compositionality. Journal of Logic, Language
and Information, 10(1):115-36.
Kant, I. (1997). Critique ofPure Reason. Cambridge University Press, Cambridge. Translated by
P. Guyer and A. Wood.
Keranen, J. (2001) . The identity problem for realist structuralism. Philosophia Mathematica,
9(3):308-30.
Kline, M. (1972) . Mathematical Thought from Ancient to Modern Times, volume 1. Oxford
University Press, Oxford.
Koellner, P. (2009). On reflection principles. Annals of Pure and Applied Logic, 157(2):206-19.
Kreisel, G. (1958). Review of Wittgenstein's remarks on the foundations of mathematics.
British Journal for the Philosophy of Science, 9(34): 135-58.
Kripke, S. A. (1975). Outline of a theory of truth. Journal of Philosophy, 72(19):690-716.
Kripke, S. A. (1976). Is there a problem about substitutional quantification? In Evans, G. and
McDowell, J., editors, Truth and Meaning, pages 324-419. Oxford University Press, Oxford.
Kripke, S. A. (1992) . Logicism, Wittgenstein, and de re beliefs about numbers. Unpublished
Whitehead Lectures, delivered at Harvard University in May 1992.
Kiinne, W. (2010). Die Philosophische Logik Freges. Klostermann, Frankfurt am Main.
Leitgeb, H. (2013). Criteria of identity: Strong and wrong. British Journal for the Philosophy of
Science, 64(1) :61-8.
Levy, A. (1960) . Axiom schemata of strong infinity in axiomatic set theory. Pacific Journal of
Mathematics, 10(1 ):223-38.
Levy, A. and Vaught, R. (1961). Principles of partial reflection in the set theories of Zermelo
and Ackermann. Pacific Journal of Mathematics, 11(3):1045-62.
BIBLIOGRAPHY 227
Lewis, D. (1991). Parts of Classes. Blackwell, Oxford.

Linnebo, 0. (2002a). Ontology and the concept of an object. Unpublished manuscript. Available
at http://www.oysteinlinnebo.org/oco.pdf.
Linnebo, 0. (2002b). Science with Numbers: A Naturalistic Defense of Mathematical Platonism.
PhD thesis, Harvard University.
Linnebo, 0. (2004a). Frege's proof of referentiality. Notre Dame Journal of Formal Logic,
45(2):73-98.
Linnebo, 0. (2004b). Predicative fragments of Frege Arithmetic. Bulletin of Symbolic Logic,
10(2):153-74.
Linnebo, 0. (2005). To be is to be an F. Dialectica, 59(2):201-22.
Linnebo, 0. (2006a). Epistemological challenges to mathematical platonism. Philosophical
Studies, 129(3):545-74.
Linnebo, 0 . (2006b). Sets, properties, and unrestricted quantification. In Rayo, A. and
Uzquiano, G., editors, Absolute Generality, pages 149-78. Oxford University Press, Oxford.
Linnebo, 0. (2008). The nature of mathematical objects. In Gold, B. and Simons, R., editors,
Proof and Other Dilemmas: Mathematics and Philosophy, pages 205-19. Mathematics Asso-
ciation of America, Washington.
Linnebo, 0. (2009a). Bad company tamed. Synthese, 170(3):371-91.
Linnebo, 0. (2009b ). Frege's context principle and reference to natural numbers. In Lindstrom,
S., editor, Logicism, lntuitionism, and Formalism: What Has Become of1hem?, volume 341 of
Synthese Library, pages 47-68. Springer, Berlin.
Linnebo, 0. (2009c) . The individuation of the natural numbers. In Bueno, 0. and Linnebo, 0.,
editors, New Waves in Philosophy of Mathematics. Palgrave Macmillan, Basingstoke.
Linnebo, 0. (2009d). Introduction [to a special issue on the bad company problem]. Synthese,
170(3):321-9.
Linnebo, 0. (2012a). Metaontological minimalism. Philosophy Compass, 7(2):139-51.
Linnebo, 0. (2012b). Reference by abstraction. Proceedings of the Aristotelian Society,
l 12(lpt1):45-7l.
Linnebo, 0 . (2013). The potential hierarchy of sets. Review of Symbolic Logic, 6(2):205-28.
Linnebo, 0. (2014). 'Just is'-statements as generalized identities. Inquiry, 57(4):466-82.
Linnebo, 0. (2016a). Impredicativity in the neo-Fregean program. In Ebert, P. and
Rossberg, M., editors, Abstractionism: Essays in the Philosophy ofMathematics, pages 247-68.
Oxford University Press, Oxford.
Linnebo, 0. (2016b). Plurals and modals. Canadian Journal of Philosophy, 46(4-5):654-76.
Linnebo, 0. (2017a). Plural quantification. In Zalta, E. N., editor, The Stanford Encyclopedia
of Philosophy. Metaphysics Research Lab, Stanford University, Stanford, CA, summer 2017
edition.
Linnebo, 0. (2017b). Generalization explained: A truth-maker semantics. Unpublished
manuscript.
Linnebo, 0. (2017c). Platonism in the philosophy of mathematics. In Zalta, E. N., editor,
The Stanford Encyclopedia of Philosophy. Metaphysics Research Lab, Stanford University,
Stanford, CA, summer 2017 edition.
Linnebo, 0. (forthcoming). The context principle in Frege's Grundgesetze. In Ebert, P. and
Rossberg, M., editors, Essays on Frege's Basic Laws of Arithmetic. Oxford University Press,
Oxford.
228 BIBLIOGRAPHY
Linnebo, 0. and Rayo, A. (2012) . Hierarchies ontological and ideological. Mind, 121(482):
269-308.
Linnebo, 0. and Shapiro, S. (2017) . Actual and potential infinity. Nous. Available at:
http://onlinelibrary.wiley.com/doi/10.l l l l/nous.12208/full.
Litland, J.E. (2017). Grounding ground. Oxford Studies in Metaphysics, 10, 279-316.
Lowe, E. J. (1989). What is a criterion of identity? Philosophical Quarterly, 39(154):1-21.
Lowe, E. J. (1997). Objects and criteria of identity. In Hale, B. and Wright, C., editors,
Companion to the Philosophy of Language. Blackwell, Oxford.
Lowe, E. J. (1998). The Possibility of Metaphysics: Substance, Identity, and Time. Oxford
Lowe, E. J. (2003). Individuation. In Loux, M. and Zimmerman, D., editors, Oxford Handbook
of Metaphysics, pages 75-95. Oxford University Press, Oxford.
MacBride, F. (2006). The Julius Caesar objection: More problematic than ever. In MacBride, F.,
editor, Identity and Modality, pages 174-202. Oxford University Press, Oxford.
Macfarlane, J. (2002) . Frege, Kant, and the logic in logicism. Philosophical Review, 111 (1 ):
25-65.
Macfarlane, J. (2009). Double vision: Two questions about the neo-Fregean program. Synthese,
170(3):443-56.
Maddy, P. (1990). Realism in Mathematics. Clarendon, Oxford.
Mancosu, P. (1998). From Brouwer to Hilbert: The Debate on the Foundations of Mathematics in
the 1920s. Oxford University Press, Oxford.
Moltmann, F. (2013). Reference to numbers in natural language. Philosophical Studies,
162(3):499-536.
Parsons, C. (1965). Frege's theory of number. In Black, M., editor, Philosophy in America, pages
180-203. Cornell University Press, Ithaca, NY. Reprinted in (Parsons, 1983a).
Parsons, C. (1971). Ontology and mathematics. Philosophical Review, 80(2):151-76. Reprinted
in (Parsons, l 983a).
Parsons, C. (1977). What is the iterative conception of set? In Butts, R. and Hintikka, J.,
editors, Logic, Foundations of Mathematics, and Computability Theory, pages 335-67. Reidel,
Dordrecht. Reprinted in (Benacerraf and Putnam, 1983) and (Parsons, 1983a).
Parsons, C. (1980). Mathematical intuition. Proceedings of the Aristotelian Society, 80(1):
145-68.
Parsons, C. (1983a). Mathematics in Philosophy. Cornell University Press, Ithaca, NY.
Parsons, C. (1983b). Sets and modality. In Mathematics in Philosophy, pages 298-341. Cornell
University Press, Ithaca, NY.
Parsons, C. ( 1990). The structuralist view of mathematical objects. Synthese, 84(3 ):303-46.
Parsons, C. (2004). Structuralism and metaphysics. Philosophical Quarterly, 54(214):56-77.
Parsons, C. (2008) . Mathematical Thought and Its Objects. Cambridge University Press,
Cambridge.
Paseau, A. (2009). Reducing arithmetic to set theory. In Bueno, 0. and Linnebo, 0., editors,
New Waves in Philosophy of Mathematics, pages 35-55. Palgrave, Basingstoke.
Payne, J. (2013a). Abstraction relations need not be reflexive. Thought: A Journal of Philosophy,
2(2):137-47.
Payne, J. (2013b ). Expansionist Abstraction. PhD thesis, University of Sheffield.
Pelletier, F. J. (2001 ). Did Frege believe Frege's principle? Journal of Logic, Language, and
Information, 10(1):87-114.
BIBLIOGRAPHY 229
Perry, J. (2002). The two faces ofidentity. In Identity, Personal Identity, and the Self, pages 64-83 .
Hackett, Indianapolis, IN.
Potter, M. (2010). Abstractionist class theory: Is there any such thing? In Smiley, T. J., Lear, J.,
and Oliver, A., editors, The Force ofArgument: Essays in Honor of Timothy Smiley. Routledge,
Abingdon.
Potter, M. and Smiley, T. (2001 ). Abstraction by recarving. Proceedings ofthe Aristotelian Society,
101(1):327-38.
Potter, M. and Smiley, T. (2002). Recarving content: Hale's final proposal. Proceedings of the
Aristotelian Society, 102(1):301-4.
Putnam, H. (1967). Mathematics without foundations. Journal of Philosophy, LXIV(1) :5-22.
Reprinted in (Benacerraf and Putnam, 1983) and (Putnam, 1975b).
Putnam, H. (1975a). The analytic and synthetic. In Mind, Language and Reality: Philosophical
Papers, volume 2, pages 33-69. Cambridge University Press, Cambridge.
Putnam, H . (1975b). Mathematics, Matter and Method. Cambridge University Press,
Cambridge.
Putnam, H. (1987). Truth and convention: On Davidson's refutation of conceptual relativism.
Dialectica, 41(1-2):69-77.
Quine, W. V. (1953a) . From a Logical Point of View. Harvard University Press, Cambridge, MA.
Quine, W. V. (1953b). Reference and modality. In From a Logical Point of View. Harvard
University Press, Cambridge, MA.
Quine, W. V. (1960). Word and Object. MIT Press, Cambridge, MA.
Quine, W. V. ( 1968). Ontological relativity. Journal of Philosophy, 65(7): 185-212.
Quine, W. V. (1986). Philosophy of Logic. Harvard University Press, Cambridge, MA, second
edition.
Quine, W. V. (1991). Two dogmas in retrospect. Canadian Journal of Philosophy, 21(3):
265-74.
Quine, W. V. (1992) . Pursuit of Truth. Harvard University Press, Cambridge, MA, second
edition.
Ramsey, F. (1931). The foundations of mathematics. In Braithwaite, R., editor, The Foundations
of Mathematics and Other Essays. Routledge & Kegan Paul, London.
Rayo, A. (2002) . Frege's unofficial arithmetic. Journal of Symbolic Logic, 67(4):1623-38.
Rayo, A. (2008). On specifying truth-conditions. Philosophical Review, 117(3):385-443.
Rayo, A. (2013). The Construction of Logical Space. Oxford University Press, Oxford.
Rayo, A. (2014). Reply to critics. Inquiry, 57(4):498-534.
Rayo, A. (2016). Neo-Fregeanism reconsidered. In Ebert, P. and Rossberg, M., editors, Abstrac-
tionism: Essays in the Philosophy of Mathematics, pages 203-21. Oxford University Press,
Oxford.
Resnik, M. (1967) . The context principle in Frege's philosophy. Philosophy and Phenomeno-
logical Research, 27(3):356- 65.
Resnik, M. (1976). Frege's context principle revisited. In Schirn, M., editor, Studien zu Frege III:
Logik und Semantik, pages 35-49. Frommann-Holzboog, Stuttgart.
Resnik, M. (1997). Mathematics as a Science of Patterns. Oxford University Press, Oxford.
Rosen, G. (1993). The refutation of nominalism (?) . Philosophical Topics, 21(2):149-86.
Rosen, G. (2010). Metaphysical dependence: Grounding and reduction. In Hale, B. and
Hoffmann, A., editors, Modality: Metaphysics, Logic, and Epistemology, pages 109-36. Oxford
230 BIBLIOGRAPHY
Rosen, G. (2011). The reality of mathematical objects. In Polkinghorne, J., editor, Meaning in
Mathematics. Oxford University Press, Oxford.
Rosen, G. (2016). Mathematics and metaphysical naturalism. In Clark, K. J., editor, The
Blackwell Companion to Naturalism, pages 277-88. Blackwell, Oxford.
Rosen, G. and Dorr, C. (2002). Composition as fiction. In Gale, R., editor, The Blackwell Guide
to Metaphysics, pages 151-74. Blackwell, Malden, MA.
Rumfitt, I. (2003). Singular terms and arithmetical logicism. Philosophical Books, 44(3):
193-219.
Rumfitt, I. (2005). Plural terms: Another variety of reference. In Bermudez, J. L., editor, Thought,
Reference and Experience, pages 84-123. Clarendon, Oxford.
Rumfitt, I. (2011). Truth and the determination of content: Variations on themes from Frege's
Logische Untersuchungen. Grazer Philosophische Studien, 82(1):3-48.
Russell, B. (1902). Letter to Frege. In (van Heijenoort, 1967).
Russell, B. (1919). Introduction to Mathematical Philosophy. Allen & Unwin, London.
Santos, G. (2013). Numbers and everything. Philosophia Mathematica, 21(3):297-308.
Schaffer, J. (2009). On what grounds what. In Manley, D., Chalmers, D. J., and Wasserman, R.,
editors, Metametaphysics: New Essays on the Foundations of Ontology, pages 347-83. Oxford
Schaffer, J. (2015). What not to multiply without necessity. Australasian Journal of Philosophy,
93( 4):644-64.
Schaffer, J. (2016). Ground rules: Lessons from Wilson. In Aizawa, K. and Gillett, C., edi-
tors, Scientific Composition and Metaphysical Ground, pages 143-69. Palgrave Macmillan,
London.
Schwartzkopff, R. (2011). Numbers as ontologically dependent objects-Hume's Principle
revisited. Grazer Philosophische Studien, 82:353-73.
Shapiro, S. (1997). Philosophy ofMathematics: Structure and Ontology. Oxford University Press,
Oxford.
Shapiro, S. (2006a). Structure and identity. In MacBride, F., editor, Identity and Modality, pages
109-45. Clarendon, Oxford.
Shapiro, S. (2006b). Vagueness in Context. Oxford University Press, Oxford.
Shapiro, S. (2017). Computing with numbers and other non-syntactic things: De re knowledge
of abstract objects. Philosophia Mathematica, 25(2):268-81.
Shapiro, S. and Weir, A. (2000). 'Neo-logicist' logic is not epistemically innocent. Philosophia
Mathematica, 8(2):160-89.
Shapiro, S. and Wright, C. (2006). All things indefinitely extensible. In Rayo, A. and
Uzquiano, G., editors, Absolute Generality, pages 255-304. Oxford University Press, Oxford.
Sider, T. (2007). Parthood. Philosophical Review, 116(1):51-91.
Snyder, E. (2017). Numbers and cardinalities: What's really wrong with the easy argument for
numbers? Linguistics and Philosophy, 40(4):373-400.
Snyder, E., Samuels, R., and Shapiro, S. (2017). Cardinals vs ordinals: Which is more basic?
Unpublished manuscript.
Spelke, E. (1993). Object perception. In Goldman, A., editor, Readings in Philosophy and
Cognitive Science, pages 447-60. MIT Press, Cambridge, MA.
Stalnaker, R. (1997). Reference and necessity. In Hale, B. and Wright, C., editors, Blackwell
Companion to the Philosophy of Language, pages 534-54. Blackwell, Oxford.
BIBLIOGRAPHY 231
Stalnaker, R. (2001). On considering a possible world as actual. Proceedings of the Aristotelian

Society, Suppl. vol. 65:141-56 .
Strawson, P. (1959). Individuals. Methuen, London.
Studd, J. P. (2013). The iterative conception of set: A (bi-)modal axiomatisation. Journal of
Philosophical Logic, 42(5):697-725.
Studd, J. P. (2016). Abstraction reconceived. British Journal for the Philosophy of Science,
67(2):579-615.
Uzquiano, G. (2003) . Plural quantification and classes. Philosophia Mathematica, 11(1):67-81.
Uzquiano, G. (2011 ). Plural quantification and modality. Proceedings of the Aristotelian Society,
l l 1(2pt2):219- 50.
van Heijenoort, J., editor (1967) . From Frege to Godel. Harvard University Press, Cambridge,
MA.
Walsh, S. and Ebels-Duggan, S. (2015). Relative categoricity and abstraction principles. Review
of Symbolic Logic, 8(3):572-606.
Weatherson, B. and Marshall, D. (2012) . Intrinsic vs. extrinsic properties. In Zalta, E. N., editor,
The Stanford Encyclopedia of Philosophy. Stanford, CA, Fall 2014 edition.
Weir, A. (1998) . Dummett on impredicativity. Grazer Philosophische Studien, 55:65-101.
Welch, P. and Horsten, L. (2016) . Reflecting on absolute infinity. Journal of Philosophy,
113(2):89-111.
Wey!, H. (1921). Ober die neue Grundlagenkrise der Mathematik. Mathematische Zeitschrift,
10(1-2):39-79. English translation in (Mancosu, 1998).
Wiggins, D. (2001 ). Sameness and Substance Renewed. Cambridge University Press, Cambridge.
Williamson, T. (1990). Identity and Discrimination. Blackwell, Oxford.
Williamson, T. (2007). The Philosophy of Philosophy. Blackwell, Malden, MA.
Williamson, T. (2010) . Necessitism, contingentism , and plural quantification. Mind,
119(475):657-748.
Wilson, J.M. (2014). No work for a theory of grounding. Inquiry, 57(5-6):535-79.
Wright, C. (1983). Frege's Conception of Numbers as Objects. Aberdeen University Press,
Aberdeen.
Wright, C. (1992) . Truth and Objectivity. Harvard University Press, Cambridge, MA.
Wright, C. (1997). The philosophical significance of Frege's Theorem. In Heck, R., editor,
Language, Thought, and Logic. Essays in Honour of Michael Dummett. Clarendon, Oxford.
Reprinted in (Hale and Wright, 200la) .
Wright, C. (1998a). The harmless impredicativity of N= (Hume's Principle) . In Schirn, M.,
editor, Philosophy of Mathematics Today, pages 339- 68. Clarendon, Oxford.
Wright, C. (1998b). Response to Michael Dummett. In Schirn, M., editor, Philosophy of
Mathematics Today, pages 389-405. Clarendon, Oxfor~.
Wright, C. (2000). Neo-Fregean foundations for analysis: Some reflections. Notre Dame Journal
of Formal Logic, 41(4):317-34.
Xu, F. ( 1997) . From Lot's wife to a pillar of salt: Evidence that physical object is a sorta! concept.
Mind and Language, 12(3&4):365-92.
Yablo, S. (2006). Circularity and paradox. In Bolander, T., Hendricks, V., and Pedersen, S. A.,
editors, Self-Reference, pages 165-83. CSLI Press, Stanford, CA.
Yablo, S. (2010) . Things: Papers on Objects, Events, and Properties. Oxford University Press,
Oxford.
Index
absolute generality 53, 64-6 Burge, Tyler 14nl4, ll3nl2, 147n26

abstraction Burgess, John 24n7, 5ln2, 52n3, 67n35
asymmetric conception of 5, 23n3, 79, 84,
86, 90n7, 152 Caesar problem xiii, 99, 123, 137-8, 159-63,
dynamic xi-xii, 51-3, 55-6, 58, 60-4, 68-71, 166nl5, 172n25
170n23, 185,205,208,211 Cantor, Georg 6, 57nl9, 115nl3, 180,
internally representable 156- 7, 174 192,216
static xiii, 52, 55, 60, 62, 64 Cantorian multiplicity 216
symmetric conception of xiii, 5, 23n3, 79, Cantor's theorem 146, 219nl9
86, 152 cardinality 4, 60n27, 93, 103-4, 146, 164nl2,
abstraction principles x-xi, xiii, 5, 8-9, 14, 168,177, 182- 3
17-19,35,49-50,52,54-5,59-63,77-9, Carey, Susan 182n18
83-4,86,90n7,95,97, 101-2, 105, 115-16, Cartwright, Richard 66n33
124-6, 130-1, 141, 152, 161, 190n3 categoricity 60n27
impredicative 98, 103, ll5, 177, 182 categories xi, 117, 120n26, 136, 144, 159-60,
predicative xi, 96, 98-9, 103, 106, 136, 178 163, 165-7
systems of 50, 161 circularity 96, 118n23
analytic entailment 130 collapse (of pluralities to sets) 59, 62-4, 205,
analytic truths 3, 5, 13-15, 17, 82, 200n23 2ll, 214-22
anti-realism 73, l39nl5 compositionalism 87-8
a priori 16n27, 179 compositionality 39, 91-2, 111, 120-1, 124,
assertibility conditions 138-43, 147-50, 153- 5, 144-6, 148-9, 171
169-73,201-3 comprehension axioms
automorphisms 24-5 modalized 68
naive set 214-15
bad company problem xi, 38, 52, 54-5, 60, 96 plural 63, 67-8, 208, 210
Balaguer, Mark 10n19 predicative 97
Barcan Formula 209 second-order 71, 97
base domain 58-9, 138-9, 142, 146, 153-4 congruence 12, 80n7,83, 99, 101-2, 106, 139,
base language 58, 89-91,93,95, 98-9, 101, 142, 186
103-4, 123-4, 136-7, 139, 149-50, 152-3, congruence claim 80-1, 83, 86
156, 170-2,201-2 consciousness 26, 29, 4ln38
Basic Law V xi, 38, 51-4, 71, 118-19, 122, context principle xiii, 107-29, 196
128, 146 generalized ll4, 117-18, 127
Benacerraf, Paul 10, 15, 197, 204 convergence property 65, 206-7, 220n21
Bernays, Paul 96, 197nl6,208 Converse Barcan Formula 74, 207
bodies 16,21-2,26-30, 32-3,36-7, 40,42,45, Cook, Roy 54nll,55nl7, 166nl5
47, 146-8, 150, 159-60, 168, 190, 193-4, counterfactual conditionals 24, 56, 189
200n23 counterfactual independence 189-90
Boghossian, Paul 3n3, 13n22 criteria of identity x-xiii, 21-3, 26, 29-30,
Boolos, George 3, 9nl5, 14, 52n3, 53n8, 55nl3, 32-3,35-40,46-8, 87,89n5,92, 135, 148,
57n20, 70n38, 115nl4, 177n5 151, 164, 166-8, 181, 196
bootstrapping argument 60, 162n8, 180, 182 one-level 35, 37, 164-5
bounded formulas 212, 214, 219-21 two-level 35-8, 40, 164-5, 167, 169-70,
Brandom, Robert 25nl0 176-7, 185, 193-4
Brouwer, L. E. J. 73n41 cumulative hierarchy of sets 70
Brouwer-Heyting-Kolmogorov cut rule 12
interpretation 73n43
Brouwerian Axiom 188, 208 de dicta 13, 85
Burali-Forti paradox 54, 185 de re 13, 16n28, 19, 79, 85- 6, 158
234 INDEX
Dedekind-Peano arithmetic 176-7, 183, 185 Fregean concepts 38, 40, 51, 70-1 , 98
deferred ostension l 70n22 Fregean triangle 21 - 3, 25, 33, 48
definiteness 69, 185 Frege's constraint 182
extensional 210-13, 217, 220 Frege's theorem 54, 60, 115n14, 177
dense linear order 92-4 fully modalized formula 64, 74, 216, 220
domain expansions 53, 59, 62, 66-8, 71
Donaldson, Thomas 84n9 G (axiom of modal logic) 65, 74, 207
Dorr, Cian 5nl0 Geach, Peter 46
Drake, F. R. 219n20 generalized context principle see context
Dummett, Michael xi, 23n4, 30nl9, 33n23, principle, generalized
39n36,42n39,47n46,69, 73n41, 77n2, generalized quantifiers see quantifiers,
87nl,91-2,96nl,98n3, 109n3, llln6, generalized
llln7, 115nl5, 117nl9, 119n24, 122n29, geometry 35, 45-6, 92nll
124, 135nl, 136n5, 136n6, 136n8, 153n36, Godel, Kurt 96, 108
159n2, 163nll, 177n7, 196-7,216n16 groundedness 119n24
grounding 18, 43n41, 56nl8, 157n39,
Ebels-Duggan, Sean 60n27 175, 193
Ebert, Philip 166n15 Grundgedanke 167-9
Eklund, Matti 4n7, 5n8 Grundgesetze xi, 105nl5, 107-8, l 12nl0,
elementary theory of abstraction (ETA) 103, 113nll, 114, 116-17, 119-25, 127-9
106, 157 Grundlagen xi, 3, 8, 107-10, 114, 116-21, 123,
elimination requirement 114-15 125-7, 129, 159-60, 168
empty plurality 67, 179, 208
Enderton, H.B. 106, 137nl0, 16ln6 Hale, Bob ix, xi, xiii, 4n5, 7, 2lnl, 24n3, 32,
epistemic constraint 16, 18, 83, 152 49n49,51-2, 77n2, 78n3, 78n4,85nl4,
epistemological challenge 10, 15, 197, 201 87-8, 95, 103, 105, 116nl7, 126-7, 129-31,
equivalence relation x-xii, 8-9, 12, 29- 30, 135nl, 136-7, 149n29, 159-60, 163-8,
34-6, 49,52,58-60,63,79-80,97,99, 102, 176nl, 177, 179, 194n9
115, 168nl9, 170n23, 184, 191 Hazen, A. P. 54nl2
partial 29-30, 32, 36, 49, 101, 138, 14lnl7, Heck, Richard 48n48, 52n3,105nl5, 117n21,
156, 161, 166, 168nl9,201 145n22, 162, 166nl6,177n7, 179nl0,
Evans, Gareth 33n23, 18lnl5 180nl4
explanation (metaphysical) 17, 44, 80, 83-4 Hellman, Geoffrey 104nl4
explanatory constraint 17, 44, 83 Hilbert, David 6, 74, 216
exportation inference 19, 15ln33, 158 Hodes, Harold 54nl2, 9ln8, 136n6, 2lln8
extended language xii, 43-4, 58, 89-91, 95, Horsten, Leon 47n47, 104nll, 119n24,
98-9, 101, 103, 105, 123-4, 136-41, 143, 218nl8
145, 147-50, 152-3, 155, 157, 169, 171, Hume's Principle (HP) 8, 35-6, 40, 53-5, 60-1 ,
201-3 98, 101, 103-5, 115, 140, 162n8, 166,
extensional definiteness see definiteness, 177-8, 182
extensional Finite (FHP) 180
extensionality (law of) 35, 58, 170, 206
identity, criteria of see criteria of identity
face value constraint 15 identity, logic of see logic, of identity
fa~on de parler x, 15, 44, 86, 91-2 identity of content 79-81, 83
Field, Hartry 5, 10nl8, 162 impredicativity 96-8, 104, 119, 140,
Fine, Kit 18n30, 52n3, 52n4, 61, 71n39, 77n2, 182n22
78n3,84nl0, 104nl3, 127, 129-30, impredicative abstraction principle
157n39, 166nl5, 175n27 see abstraction principle, impredicative
flexible conception ofreality 31-3 indefinite extensibility 52n4, 69, 190-2,
Fodor, Jerry 39, 198 216nl6,217nl7
Foundation (axiom) 213, 216, 221 infinity
Frege Arithmetic 54, 115nl4 actual vs. potential 61, 126n36
Frege, Gottlob ix-xi, xiii, 3-5, 7-9, 13-15, 17, axiom oflnfinity 216-17, 219-21
21-5, 31,34-5,37,39,42n39,51-3, 55, inheritance principle 36-7
77-8,81,89, 103-5, 107-29, 135, 144, 152, innocent counterparts of statements 95, 100
159-62, 168-9, 176-7, 180-1, 190 input theory 100-3, 106
INDEX 235
internalism about reference 41-2, 150-1 metasemantics 38-9, 153

internally representable abstraction mereological sums 5, 9, 11, 17
see abstraction, internally representable merging
of domains 58-9, 171
Jane, Ignasi 60n27 of sorts 145, 149n30, 161, 169
Janssen, Theo llln5, 117nl9 minimalism
"just is" -statements 4, 11, 81-3 about criteria of identity 47-8
abstractionist 7-9
Kant, Immanuel 3, 5 coherentist 5-7, 10
Keranen, Jukka 24n7 metaontological 4-7, 9-11 , 21, 25
knowledge ix, 10, 15-16, 47, 83, 109, 161, 179, mirroring theorem 65, 67-8, 74, 205, 208
197,202-3 modal logic see logic, modal
knowledge operator 16 modal set theory 63, 216-19
Koellner, Peter 218nl8 modality
Kreisel, Georg 197 circumstantial 61, 205
Kripke models 63 interpretational 61-2, 65, 71-3, 188, 205-6
Kripke, Saul xii, 40-1, 119n24, 147, 18lnl7, modalized quantifiers 64, 66-8, 205, 222
210n5 Moltmann, Friederike l 78n8
Kiinne, Wolfgang 117nl9 multiplicity (consistent vs. inconsistent) 57nl9,
216
language mutual sufficiency statements 11, 17-19, 79
plural 57-8, 211
second- or higher-order 6, 36, 64, 97, 103-4 negative free logic see logic, negative free
two-sorted 58-9, 98-9, 103, 123, 137-8, 169, neo-Fregeanism xi, 7-9, 32, 48n48, 51-5, 78,
173-4 88, 105, 136, 152, 162, 177-8, 190n3
Lego block conception of reality 32 NewV 60n27
Leibniz's Law 30, 79-81, 101, 106, 157-9, non-reductionism (semantic) 141, 153
161, 174 nuisance principle 55
Leitgeb, Hannes 48n48 number
Levy, Azriel 219n20 cardinal 36, 46, 60n27, 69, 83, 104, 176,
Lewis, David 4n6 180-1; see also cardinality
literals 174-5 complex 5, 24-5
Litland, Jon 18n32, 157n39 ordinal 69-70, 176-84
Lob principle 213nl0 numerals 39-40, 177-87, 190, 195, 200,
Locke, John 35n26 210n7
Lockean model of reference l 49n29
logic object
bimodal 71-2 Fregean conception of x, xii, 23-6, 30, 87,
of identity 30-1 , 33 151, 160
intuitionistic 73-4 Quinean conception of 24- 6
modal 13, 16, 73-4, 188, 206-9, 2lln8 ultra-thin conception of 42n39, 87-93, 108,
negative free 29, 36, 48-9 123-4, 129, 151
plural 57-8, 62, 67-8, 70, 205-6, objects
208-10, 221 abstract xii-xiii, 9- 10, 15, 17, 22-3, 33-4,
second-order 8,53,55, 70,97, 119n24, 177, 37,39,45-6,95-6, 107-9, 112, 119, 124,
182, 2lln8 128, 135-7, 141, 150-3,188, 193,
semi-intuitionistic 74 195-7,203
Lowe, E. J. 22n2, 35n26, 39n36, 46n45, 48, pure abstract 17, 45-6, 188
163nl 1 thin ix-xiv, 3-10, 13-16, 18, 21-3, 42-3,
45-6,86,92, 107, 127-9, 136, 151-3,
MacBride, Fraser 16ln5, 168n20 190-5,202
Macfarlane, John 3n3, 49n49 Occam's razor 10
Maddy, Penelope 190 ontological commitment xii, 13- 14, 19, 23n3,
Marshall, Dan 193n7 44,79, 82,84-5, 136n6, 142
maximality principle 207, 214, 217 ontological expansiveness constraint 14
metasemantic reductionism see reductionism, open-endedness 70, 187
metasemantic outputtheory 102-3
236 INDEX
parcels of matter 22, 27-30, 32, 40, 147, 199 reconceptualization 23, 26, 30, 33-4, 43n40, 44
Parsons, Charles 4n4, 6, 9nl5, 25nl l, 45, 53n8, reducibility of predication 192, 194-5
115nl4, 176n2, 177n5, 182n20, 183n24, reductionism
187n34,209n4,210n5,216nl5,220n21 holophrastic 91-3, 124n31, 128, 152, 156-7
Paseau, Alexander 178n8 metasemantic 128, 153
Payne, Jonathan 49n49, 52n7, 62n29 semantic 141, 153
Pelletier, Francis Jeffry 108n2, llln5, 117nl9, reference x, xii-xiii, 15, 21-3, 25-6, 29- 30,
121n28 32-4,37-44,47-8, 78,81,87-9,91-3,
Perry, John 159n2 105nl5, 108-10, 112, 117-21, 123-31,
Platonism xiii, 10n19, 189-92, 197, 198nl8 135-7, 141, 144, 146- 7, 149n29, 150-1,
plural comprehension see comprehension 153, 178, 181-2, 196,200n23
axioms, plural easy x,23,26,93, 151
plural Law V 58-63, 66, 68, 170, 205 inexplicable 92-3
Poincare, Henri 96 reflection principle 217-20, 222
potential collapse (of pluralities to sets) 62-4, complete vs. partial 220n21
205, 211,214-22 reliability claims 198, 202-3
potential hierarchy of sets 217-18 Replacement (axiom scheme) 216-17, 219n20
potential infinity see infinity, actual vs. potential reproduction of meaning 110, 114-16, 120, 123
potentialist translation see translation, Resnik, Michael 4n4, 6, llln5, 117nl9,
potentialist 121n28,176n2
Potter, Michael 78n3, 106, 127n37 rigid conception ofreality 31-2
Power set axiom 216-17 robots 26-30
predicative abstraction principle see abstraction Rosen, Gideon 5nl0, 18n30, 67n35, 84n9,
principle, predicative 84n10, 136n7, 146
predicative second-order logic Rumfitt, Ian 62n28, 117n20, 136n5
see comprehension axioms, predicative Russell, Bertrand 38, 51 , 96, 108, 176-8
priority principle 213 Russellian propositions 78, 126n35
proofofreferentiality 118-19 Russell's paradox 38, 51-3, 58, 60, 62-3, 96,
properties 25n9, 27, 30, 33-4, 36, 39, 47, 71, 119, 218
80-2,94, 102, 111-12, 122, 138, 161, 168,
189-90, 192-3, 196-7 S4 65, 74,207,220
extrinsic 193 S4.2 65, 73-4,207-8,210,221-2
intrinsic 193-6 Santos, Gon<;:alo 73n41
semantic 39, 121, 153 Schaffer, Jonathan lln20, 18n30, 18n31, 84n10
worldly 83- 6 Schwartzkopff, Robert 84n9
psychology 26 second-order logic see logic, second-order
purely referential singular term 79-80, 180 semantic non-reductionism
Putnam, Hilary 6nl3, l3n23, 3ln20 see non-reductionism (semantic)
semantic reductionism see reductionism,
quantifiers semantic
generalized 145 semantic value 39, 41, 70, 120-1, 124, 126n35,
plural see logic, plural 143-4, 181, 196, 199-200
quasi-concrete objects 45-6 semantically idle singular terms 90-1, 124,
Quine, W. V. 13, 24-6, 44n42, 79n6, 91, 1lln5, 152-3, 196
170n22 sense 8, 78,81, 108-10, 113, 121, 124, 126, 129
Separation (axiom scheme) 216, 220
Ramsey, Frank 96 set theory (naive) 63
Rayo, Agustin ix, xi, 4, 32, 78-85, 87-90, 101, sets xi-xii, 4, 9, 13, 17, 22-3, 25, 35, 45-8, 53,
104nl2, 104nl3, 152, 163n10 56,58-9,62-4,67, 70-1,84, 146, 163nl2,
Real Closed Fields 92nll 164, 170, 191,206, 211 - 12,214-20
realism 3, 18n30, 31n20, 73, 128 shallow nature 192, 194-7
object 162, 189, 191, 196-7 Shapiro, Stewart 4n4, 6-7, 73n42, l38nl4,
truth-value 196 162n9, 176n2, 180nl3, 183n25, 197nl6,
working 197 216nl6,217n17
recarving of content xi, 8-9, 11, 23n3, 77-8, 81, Sider, Theodore 4n6
84-5, 104nl4, 108, 116, 120, 123-7, Smiley, Timothy 78n3, 127n37
129-31, l36n7 Snyder, Eric 176n3, 178n8
INDEX 237
sorta! concepts 163-7, 181, 193-4 types 9,33, 35n25, 101-2, 112, 114, 118n22,
Spelke, Elizabeth 27n14 135-43, 145, 148-53, 155-6, 166, 169, 171,
stable formulas 64-6, 74, 209, 212, 216, 220 176, 187n34, 192-3, 195
Stalnaker, Robert 39n34 isomorphism types 54
stars 69, 191, 196
Strawson, P. F. 163nl 1 ultra-thin conception of object see object,
structuralism 24n7, 162n9 ultra-thin conception of
Studd, James 52n6, 55nl7, 62n29, 63n30, uniqueness thesis 166-7, 170
72n40,207nl,21ln8,213nl0,219n20 unit of significance 111
successor axiom 187-8 unity relation 35-7, 39, 4ln38, 42, 45, 93, 138,
sufficiency 147-8, 164- 5, 168- 70, 194-6, 199-200
mediate vs. immediate 16n28, 43n40, 43n41, Uzquiano, Gabriel 60n27, 62n28, 70n38
174-5
operator 11-16, 19, 43-4, 86, 151-2, 157, Walsh, Sean 60n27
174,202 Weatherson, Brian 193n7
statements 11, 13-19, 43-4, 152, 158; web pages 56-7
see also mutual sufficiency statements Weir, Alan 177n7, 180nl3
syntactic priority thesis xi, 87-8, 136, 151 Welch, Philip 218nl8
Wey!, Hermann 73
Tarski, Alfred 7, 92nl 1 Wiggins, David 30nl 9, 163nl l
theory Williamson, Timothy xii, 35n27, 62n28,
complete 65, 92-4, 207 159n2
decidable 92-4 Wilson, Jessica 18n3 l
tokens 33- 4,40, 45, 112, 114, 118n22, 135-43, Wright, Crispin ix, xi, xiii, 4n5, 7, 9nl5, 32,
145-6. 148, 151-3, 155, 166, 169-72, 176, 49n49,51-5, 78n3, 78n4,85nl4,87-8,90,
183, 187, 192-3, 195, 198 95, 103, 105, 115nl4, 115nl5, 116nl7,
translations 50, 95, 98, 101, 103-6, 123, 127, 127n37, 130, 135nl , 136-7, 149n29,
139, 174 159-60, 163-9, 176nl, 177, 179, 182nl9,
logically acceptable 89-94, 100, 102, 110n4, 190n3,194n9,216nl6,217nl 7
124, 150, 156
potentialist 64, 68, 205, 212- 13, 215, Xu, Fei 27nl4, 30nl9
217-20
transparent context 30, 79-81 Yablo, Stephen 63n31, 194n9, 214
truth-conditions 30-2, 87-9, 101. 114, 142-3,
146,201 zero 82-3, 179-80
two-sorted languages 58-9, 98-9, 103, 123, ZF xi, 63, 213, 214-19, 222
137-8, 169, 173-4 ZFC xiii, 211

Thin Objects

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thin Objects

Uploaded by

Copyright:

Available Formats

Thin 0 bj ects

Appendix 3.A Further Questions 70

Part II. Comparisons

Part III. Details

12. Dynamic Set Theory 205

(The lower arrow will be explained shortly.)

(AP) §a = §{3 = a "' f3

dynamic abstraction. Legitimate abstraction steps are iterated indefinitely to build

of the objects described by a theory of pure mathematics amounts to nothing more

Metaontological minimalism has consequences concerning ontology proper. The

1.2 Coherentist Minimalism

As is well known, the word 'criterion' is ambiguous between a metaphysical meaning

It is instructive to compare this principle with Tarski's semantic account of logical

there is a model of T <=> T is coherent

1.3 Abstractionist Minimalism

Consider the criterion of identity for directions:

As Frege discovered, this principle has an amazing mathematical property. When

derive all of ordinary (second-order Dedekind-Peano) arithmetic. 15 As we shall see,

1.4 The Appeal of Thin Objects

(Wright, 1983). See (Boolos, 1990) for a nice proof.

provided in Section 6.4.

a much stronger conclusion.

On the other hand, the ontology of abstract objects is often found to be

1.5 Sufficiency and Mutual Sufficiency

all that is required for 1/1 is <p

or the symmetrical analogue

<pis a recarving of the content of 1/1

It will be convenient to allow the sufficiency operator to take a set of formulas on

1.6 Philosophical Constraints

sentence true. A proponent of thin objects is interested in sufficiency statements

where the right-hand side is ontologically committed to abstract directions, while

Indeed, on an abstractionist approach to thin objects, we probably want true suf-

Face value constraint

1. 7 Two Metaphysical "Pictures"

How should this sufficiency statement be understood? It cannot straightforwardly

and will be developed further in Chapter 4. There appears to be a worldly asymmetry

3x3y(x = §a /\ y = §{3 /\ (x = y {:> a ,....., fJ))

Suppose we apply exportation to derive:

2.1 My Strategy in a Nutshell

objecthood - - - - - - - - identity criteria

What is an object? There is a certain audacity to the question. Can we possibly

a substantial demand on the world. The explication of the concept of an object to be

2.2 A Fregean Concept of Object

5 See also (Hale, 2013, ch. 1).

to developments of the Quinean concept of an object that might supplement my

2.3 Reference to Physical Bodies

(Bl) Bodies are three-dimensional solid objects.

(B2) Bodies have natural and relatively well-distinguished spatial boundaries.

(B3) Bodies are units of independent motion.

(B4) Bodies move along continuous paths.

(BS) Bodies have natural and relatively well-distinguished temporal boundaries.

So arbitrary temporal parts of bodies are not themselves bodies.

approaches to abstraction on partial equivalence relations.

book tokens, as we may call them-and predicating properties of these. We can

2.5 Reference by Abstraction

relevant standards. 25 Geometrical shapes provide another example. A shape can be

(CI-1) Fx /\ Py-+ (x = y ++ R(x,y))

(Ext) SET(x) /\ SET(y) -+ (x = y t t Vu(u E x t t u E y))

(CI-2) f(a) = f(f3) ++a ~ f3

where F and G are second-order variables and F ~ G abbreviates the second-order

2.6 Some Objections and Challenges

2.6.1 The bad company problem

2.6.4 A clash with Kripke on reference?

2.7 A Candidate for the Job